Merge main and adapt multi-API startup to structured configs

Signed-off-by: Sy03 <1370724210@qq.com>

.buildkite/amd/scripts/bootstrap-amd-omni.sh

@@ -1,4 +1,7 @@
 #!/bin/bash
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
+
 # vllm-omni customized version
 # Based on: https://github.com/vllm-project/ci-infra/blob/main/buildkite/bootstrap-amd.sh
 # Last synced: 2025-12-15
@@ -6,8 +9,13 @@
 
 set -euo pipefail
 
+# The bootstrap runs from the repository root; ShellCheck does not follow
+# sourced files unless invoked with -x.
+# shellcheck disable=SC1091
 source .buildkite/common/scripts/resolve_skip_ci.sh
 
+AMD_SUITE_SELECTOR=".buildkite/amd/scripts/select_test_suites.py"
+
 if [[ -z "${RUN_ALL:-}" ]]; then
     RUN_ALL=0
 fi
@@ -24,31 +32,20 @@ if [[ -z "${AMD_MIRROR_HW:-}" ]]; then
     AMD_MIRROR_HW="amdproduction"
 fi
 
-fail_fast() {
-    DISABLE_LABEL="ci-no-fail-fast"
-    # If BUILDKITE_PULL_REQUEST != "false", then we check the PR labels using curl and jq
-    if [ "$BUILDKITE_PULL_REQUEST" != "false" ]; then
-        PR_LABELS=$(curl -s "https://api.github.com/repos/vllm-project/vllm-omni/pulls/$BUILDKITE_PULL_REQUEST" | jq -r '.labels[].name')
-        if [[ $PR_LABELS == *"$DISABLE_LABEL"* ]]; then
-            echo false
-        else
-            echo true
-        fi
-    else
-        echo false  # not a PR or BUILDKITE_PULL_REQUEST not set
+PR_LABELS=""
+# Explicit debug suites do not depend on PR labels.
+if [[ "${BUILDKITE_PULL_REQUEST:-false}" != "false" && -z "${DEBUG_TEST_YAML:-}" ]]; then
+    if ! PR_LABELS=$(curl -fsSL \
+        "https://api.github.com/repos/vllm-project/vllm-omni/pulls/$BUILDKITE_PULL_REQUEST" \
+        | jq -r '.labels[].name'); then
+        echo "ERROR: Could not read PR labels; refusing to select an incorrect AMD test tier." >&2
+        exit 1
     fi
-}
+fi
 
-check_run_all_label() {
-    RUN_ALL_LABEL="ready-run-all-tests"
-    # If BUILDKITE_PULL_REQUEST != "false", then we check the PR labels using curl and jq
-    if [ "$BUILDKITE_PULL_REQUEST" != "false" ]; then
-        PR_LABELS=$(curl -s "https://api.github.com/repos/vllm-project/vllm-omni/pulls/$BUILDKITE_PULL_REQUEST" | jq -r '.labels[].name')
-        if [[ $PR_LABELS == *"$RUN_ALL_LABEL"* ]]; then
-            echo true
-        else
-            echo false
-        fi
+fail_fast() {
+    if [[ "${BUILDKITE_PULL_REQUEST:-false}" != "false" ]]; then
+        echo true
     else
         echo false  # not a PR or BUILDKITE_PULL_REQUEST not set
     fi
@@ -58,11 +55,71 @@ if [[ -z "${COV_ENABLED:-}" ]]; then
     COV_ENABLED=0
 fi
 
+resolve_test_specs() {
+    local selected_specs
+    if ! selected_specs=$(python3 "$AMD_SUITE_SELECTOR" \
+        --branch "$BUILDKITE_BRANCH" \
+        --labels "$PR_LABELS" \
+        --debug-test-yaml "${DEBUG_TEST_YAML:-}" \
+        --nightly "$NIGHTLY"); then
+        exit 1
+    fi
+
+    TEST_SPECS=()
+    while IFS= read -r suite_spec; do
+        [[ -z "$suite_spec" ]] && continue
+        TEST_SPECS+=("$suite_spec")
+    done <<< "$selected_specs"
+}
+
+filter_test_specs_by_skip_ci() {
+    local suite_spec level decision
+    local skip_all=0
+    local -a runnable_specs=()
+
+    for suite_spec in "${TEST_SPECS[@]}"; do
+        case "$suite_spec" in
+            READY_TESTS:*) level="l2" ;;
+            MERGE_TESTS:*) level="l3" ;;
+            NIGHTLY_TESTS:*)
+                # Explicit PR and scheduled nightly selections must survive
+                # L2/L3 diff gating, including docs-only main commits.
+                runnable_specs+=("$suite_spec")
+                continue
+                ;;
+            *)
+                echo "ERROR: unknown AMD test suite spec '$suite_spec'" >&2
+                exit 1
+                ;;
+        esac
+
+        if decision=$(python3 "$SKIP_CI_PY" gate amd "$level"); then
+            echo "Skipping AMD $level suite due to $decision decision."
+            [[ "$decision" == "skip-all" ]] && skip_all=1
+        else
+            runnable_specs+=("$suite_spec")
+        fi
+    done
+
+    TEST_SPECS=("${runnable_specs[@]}")
+    if [[ ${#TEST_SPECS[@]} -eq 0 ]]; then
+        if [[ $skip_all -eq 1 ]]; then
+            buildkite-agent annotate \
+                ":memo: CI skipped — docs or pytest skip-mark changes only" \
+                --style "info" 2>/dev/null || true
+        fi
+        echo "No AMD suites remain after skip-ci filtering."
+        exit 0
+    fi
+}
+
 upload_pipeline() {
     echo "Uploading pipeline..."
     # Install minijinja
     ls .buildkite || buildkite-agent annotate --style error 'Please merge upstream main branch for buildkite CI'
     curl -sSfL https://github.com/mitsuhiko/minijinja/releases/download/2.3.1/minijinja-cli-installer.sh | sh
+    # Installed by the minijinja bootstrap above and only present on the CI agent.
+    # shellcheck disable=SC1091
     source /var/lib/buildkite-agent/.cargo/env
 
     if [[ $BUILDKITE_PIPELINE_SLUG == "fastcheck" ]]; then
@@ -76,7 +133,7 @@ upload_pipeline() {
     # (WIP) Use pipeline generator instead of jinja template
     if [ -e ".buildkite/amd/pipeline_generator/pipeline_generator.py" ]; then
         python -m pip install click pydantic
-        python .buildkite/amd/pipeline_generator/pipeline_generator.py --run_all=$RUN_ALL --list_file_diff="$LIST_FILE_DIFF" --nightly="$NIGHTLY" --mirror_hw="$AMD_MIRROR_HW"
+        python .buildkite/amd/pipeline_generator/pipeline_generator.py --run_all="$RUN_ALL" --list_file_diff="$LIST_FILE_DIFF" --nightly="$NIGHTLY" --mirror_hw="$AMD_MIRROR_HW"
         buildkite-agent pipeline upload .buildkite/amd/pipeline.yaml
         exit 0
     fi
@@ -89,28 +146,46 @@ upload_pipeline() {
 
     cd .buildkite/amd
 
-    # Select test definition file: merge suite for main, ready suite for PRs.
-    # For debugging, DEBUG_TEST_YAML can override the selection — accepts
-    # "merge" or "ready" (case-insensitive).
-    if [[ -n "${DEBUG_TEST_YAML:-}" ]]; then
-        case "${DEBUG_TEST_YAML,,}" in
-            merge)
-                TEST_YAML="test-amd-merge.yml"
-                ;;
-            ready)
-                TEST_YAML="test-amd-ready.yml"
-                ;;
-            *)
-                echo "ERROR: DEBUG_TEST_YAML must be 'merge' or 'ready', got '$DEBUG_TEST_YAML'" >&2
-                exit 1
-                ;;
-        esac
-        echo "DEBUG_TEST_YAML override: using $TEST_YAML"
-    elif [[ $BUILDKITE_BRANCH == "main" ]]; then
-        TEST_YAML="test-amd-merge.yml"
+    # Multiple label-selected or debug-selected suites share one image build.
+    if [[ ${#TEST_SPECS[@]} -eq 1 ]]; then
+        TEST_YAML="${TEST_SPECS[0]#*:}"
     else
-        TEST_YAML="test-amd-ready.yml"
+        TEST_YAML=$(mktemp "${TMPDIR:-/tmp}/amd-selected-tests.XXXXXX.yml")
+        python - "$TEST_YAML" "${TEST_SPECS[@]}" <<'PY'
+import sys
+
+import yaml
+
+
+output_path, *suite_specs = sys.argv[1:]
+combined = {"env": {}, "steps": []}
+
+for suite_spec in suite_specs:
+    group_name, input_path = suite_spec.split(":", 1)
+    with open(input_path, encoding="utf-8") as test_file:
+        suite = yaml.safe_load(test_file)
+
+    for name, value in (suite.get("env") or {}).items():
+        previous = combined["env"].get(name, value)
+        if previous != value:
+            raise ValueError(
+                f"Conflicting environment value for {name}: {previous!r} != {value!r}"
+            )
+        combined["env"][name] = value
+
+    suite_steps = []
+    for entry in suite.get("steps") or []:
+        if "group" in entry:
+            suite_steps.extend(entry.get("steps") or [])
+        else:
+            suite_steps.append(entry)
+    combined["steps"].append({"group": group_name, "steps": suite_steps})
+
+with open(output_path, "w", encoding="utf-8") as output_file:
+    yaml.safe_dump(combined, output_file, sort_keys=False)
+PY
     fi
+    echo "AMD test suites: ${TEST_SPECS[*]}"
 
     (
         set -x
@@ -130,19 +205,20 @@ upload_pipeline() {
             > pipeline.yaml
     )
     cat pipeline.yaml
+    if [[ "$TEST_YAML" == "${TMPDIR:-/tmp}/amd-selected-tests."*.yml ]]; then
+        rm -f -- "$TEST_YAML"
+    fi
     buildkite-agent artifact upload pipeline.yaml
     buildkite-agent pipeline upload pipeline.yaml
     exit 0
 }
 
 get_diff() {
-    $(git add .)
-    echo $(git diff --name-only --diff-filter=ACMDR $(git merge-base origin/main HEAD))
+    git diff --name-only --diff-filter=ACMDR "$(git merge-base origin/main HEAD)"
 }
 
 get_diff_main() {
-    $(git add .)
-    echo $(git diff --name-only --diff-filter=ACMDR HEAD~1)
+    git diff --name-only --diff-filter=ACMDR HEAD~1
 }
 
 file_diff=$(get_diff)
@@ -150,12 +226,10 @@ if [[ $BUILDKITE_BRANCH == "main" ]]; then
     file_diff=$(get_diff_main)
 fi
 
-# Early exit: unified skip-ci (docs / skip marks) and CI-yaml-only level targeting.
-if [[ $BUILDKITE_BRANCH == "main" ]]; then
-    gate_bootstrap_ci amd l3
-else
-    gate_bootstrap_ci amd l2
-fi
+# Resolve PR tier labels before skip-ci so ready/merge suites can be filtered
+# independently while an explicitly selected nightly suite passes through.
+resolve_test_specs
+filter_test_specs_by_skip_ci
 
 patterns=(
     "docker/Dockerfile"
@@ -181,7 +255,7 @@ for file in $file_diff; do
     # First check if file matches any pattern
     matches_pattern=0
     for pattern in "${patterns[@]}"; do
-        if [[ $file == $pattern* ]] || [[ $file == $pattern ]]; then
+        if [[ $file == "$pattern"* ]] || [[ $file == "$pattern" ]]; then
             matches_pattern=1
             break
         fi
@@ -191,7 +265,7 @@ for file in $file_diff; do
     if [[ $matches_pattern -eq 1 ]]; then
         matches_ignore=0
         for ignore in "${ignore_patterns[@]}"; do
-            if [[ $file == $ignore* ]] || [[ $file == $ignore ]]; then
+            if [[ $file == "$ignore"* ]] || [[ $file == "$ignore" ]]; then
                 matches_ignore=1
                 break
             fi
@@ -205,14 +279,6 @@ for file in $file_diff; do
     fi
 done
 
-# Check for ready-run-all-tests label
-LABEL_RUN_ALL=$(check_run_all_label)
-if [[ $LABEL_RUN_ALL == true ]]; then
-    RUN_ALL=1
-    NIGHTLY=1
-    echo "Found 'ready-run-all-tests' label. Running all tests including optional tests."
-fi
-
 # Decide whether to use precompiled wheels
 # Relies on existing patterns array as a basis.
 if [[ -n "${VLLM_USE_PRECOMPILED:-}" ]]; then

.buildkite/amd/scripts/run-amd-test.sh

@@ -1,153 +1,103 @@
 #!/bin/bash
-# vllm-omni customized version
-# Based on: vllm/.buildkite/scripts/hardware_ci/run-amd-test.sh
-# Last synced: 2025-12-15
-# Modifications: docker image name for vllm-omni
 
-# This script runs test inside the corresponding ROCm docker container.
-set -o pipefail
+# Run vLLM-Omni ROCm tests directly in the MI300 Kubernetes pod. The pod's
+# container image is selected by test-template-amd-omni.j2; MI300 has no DinD.
+set -euo pipefail
 
-# Export Python path
-export PYTHONPATH=".."
+: "${PYTHONFAULTHANDLER:=1}"
+: "${HF_HOME:=/home/buildkite-agent/huggingface}"
+: "${HF_HUB_DOWNLOAD_TIMEOUT:=300}"
+: "${HF_HUB_ETAG_TIMEOUT:=60}"
+: "${MIOPEN_DEBUG_CONV_DIRECT:=0}"
+: "${MIOPEN_DEBUG_CONV_GEMM:=0}"
+: "${VLLM_ROCM_USE_AITER:=0}"
+export PYTHONFAULTHANDLER HF_HOME HF_HUB_DOWNLOAD_TIMEOUT HF_HUB_ETAG_TIMEOUT
+export MIOPEN_DEBUG_CONV_DIRECT MIOPEN_DEBUG_CONV_GEMM VLLM_ROCM_USE_AITER
+export PYTORCH_ROCM_ARCH=""
+export PYTHONPATH="${PYTHONPATH:-..}"
+
+if [[ "${VLLM_CI_DOCKER_DISABLED:-0}" != "1" ]]; then
+    echo "Error: MI300 CI must run natively with Docker disabled." >&2
+    exit 1
+fi
+
+if [[ -n "${TEST_COMMANDS:-}" ]]; then
+    commands="${TEST_COMMANDS}"
+else
+    commands="$*"
+fi
+if [[ -z "${commands}" ]]; then
+    echo "Error: No test commands provided." >&2
+    exit 1
+fi
+
+echo "--- Native in-pod ROCm CI"
+
+job_id="${BUILDKITE_JOB_ID:-${BUILDKITE_PARALLEL_JOB:-local}}"
+job_id="${job_id//[^A-Za-z0-9_.-]/_}"
+native_cache_root="/tmp/vllm-omni-native-${job_id}"
+TMPDIR="${native_cache_root}/tmp"
+TORCHINDUCTOR_CACHE_DIR="${native_cache_root}/torchinductor"
+TRITON_CACHE_DIR="${native_cache_root}/triton"
+VLLM_CACHE_ROOT="${native_cache_root}/vllm"
+XDG_CACHE_HOME="${native_cache_root}/xdg"
+HF_DATASETS_CACHE="${native_cache_root}/huggingface/datasets"
+export TMPDIR TORCHINDUCTOR_CACHE_DIR TRITON_CACHE_DIR VLLM_CACHE_ROOT
+export XDG_CACHE_HOME HF_DATASETS_CACHE
+mkdir -p "${TMPDIR}" "${TORCHINDUCTOR_CACHE_DIR}" "${TRITON_CACHE_DIR}" \
+    "${VLLM_CACHE_ROOT}" "${XDG_CACHE_HOME}" "${HF_HOME}" \
+    "${HF_DATASETS_CACHE}"
+
+if [[ "${VLLM_CI_REQUIRE_PERSISTENT_HF_CACHE:-0}" == "1" ]]; then
+    if ! command -v findmnt >/dev/null 2>&1; then
+        echo "Error: findmnt is required to verify the Hugging Face cache mount." >&2
+        exit 1
+    fi
+    hf_mount=$(findmnt -n -T "${HF_HOME}" -o TARGET 2>/dev/null || true)
+    if [[ -z "${hf_mount}" || "${hf_mount}" == "/" ]]; then
+        echo "Error: MI300 CI requires a persistent volume at or above ${HF_HOME}." >&2
+        exit 1
+    fi
+fi
 
-# Print ROCm version
-echo "--- ROCm info"
 rocminfo
 
-# cleanup older docker images
-cleanup_docker() {
-  # Get Docker's root directory
-  docker_root=$(docker info -f '{{.DockerRootDir}}')
-  if [ -z "$docker_root" ]; then
-    echo "Failed to determine Docker root directory."
-    exit 1
-  fi
-  echo "Docker root directory: $docker_root"
-  # Check disk usage of the filesystem where Docker's root directory is located
-  disk_usage=$(df "$docker_root" | tail -1 | awk '{print $5}' | sed 's/%//')
-  # Define the threshold
-  threshold=70
-  if [ "$disk_usage" -gt "$threshold" ]; then
-    echo "Disk usage is above $threshold%. Cleaning up Docker images and volumes..."
-    # Remove dangling images (those that are not tagged and not used by any container)
-    docker image prune -f
-    # Remove unused volumes / force the system prune for old images as well.
-    docker volume prune -f && docker system prune --force --filter "until=72h" --all
-    echo "Docker images and volumes cleanup completed."
-  else
-    echo "Disk usage is below $threshold%. No cleanup needed."
-  fi
-}
+expected_gpus="${VLLM_CI_EXPECTED_GPU_COUNT:-1}"
+python3 - <<'PY'
+import os
 
-# Call the cleanup docker function
-cleanup_docker
+import torch
 
-echo "--- Pulling container"
-## Temporary change to use AMD Docker Hub to store the vllm-omni image
-# to bypass the rate limit issue with ECR Public Gallery.
-# Images are now stored in a separate repository for vllm-omni, instead of vllm-ci.
-# TODO: @tjtanaa point back to ECR Public Gallery
-# once the amd agents are configured to use ECR Public Gallery.
-# image_name="public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:${BUILDKITE_COMMIT}-rocm-omni"
-image_name="rocm/vllm-omni:${BUILDKITE_COMMIT}"
-container_name="rocm_${BUILDKITE_COMMIT}_$(tr -dc A-Za-z0-9 < /dev/urandom | head -c 10; echo)"
+expected = int(os.environ.get("VLLM_CI_EXPECTED_GPU_COUNT", "1"))
+assert torch.version.hip, "PyTorch is not a ROCm build"
+assert torch.cuda.is_available(), "ROCm GPU is not available to PyTorch"
+actual = torch.cuda.device_count()
+assert actual == expected, f"Expected {expected} ROCm GPU(s), found {actual}"
+PY
 
-# TODO: @tjtanaa uncomment this once the amd agents are configured to use ECR Public Gallery.
-# # Install AWS CLI to authenticate to ECR Public Gallery to get higher rate limit for pulling images
-# sudo apt-get update && sudo apt-get install -y awscli
-##  Use safe docker login helper to prevent race conditions
-# source "$(dirname "${BASH_SOURCE[0]}")/../../common/scripts/docker_login_ecr_public.sh"
-# safe_docker_login_ecr_public
-
-## Pull the container from AMD Docker Hub
-
-docker pull "${image_name}"
-
-remove_docker_container() {
-   docker ps -aq --filter "name=${container_name}" | xargs -r docker rm -f || true
-   docker image rm -f "${image_name}" || true
-}
-trap remove_docker_container EXIT
-
-echo "--- Running container"
-
-HF_CACHE="$(realpath ~)/huggingface"
-mkdir -p "${HF_CACHE}"
-HF_MOUNT="/root/.cache/huggingface"
-
-if [[ -n "${TEST_COMMAND:-}" ]]; then
-    commands="$TEST_COMMAND"
-else
-    commands="$@"
-fi
-echo "Commands:$commands"
-
-PARALLEL_JOB_COUNT=8
-MYPYTHONPATH=".."
-
-# Test that we're launching on the machine that has
-# proper access to GPUs
-render_gid=$(getent group render | cut -d: -f3)
-if [[ -z "$render_gid" ]]; then
-  echo "Error: 'render' group not found. This is required for GPU access." >&2
-  exit 1
-fi
-
-# check if the command contains shard flag, we will run all shards in parallel because the host have 8 GPUs.
-# TODO: @tjtanaa reenable to run VLLM_ROCM_USE_AITER=1 when AITER is shipped with prebuilt kernels.
-if [[ $commands == *"--shard-id="* ]]; then
-  # assign job count as the number of shards used
-  commands=$(echo "$commands" | sed -E "s/--num-shards[[:blank:]]*=[[:blank:]]*[0-9]*/--num-shards=${PARALLEL_JOB_COUNT} /g" | sed 's/ \\ / /g')
-  for GPU in $(seq 0 $(($PARALLEL_JOB_COUNT-1))); do
-    # assign shard-id for each shard
-    commands_gpu=$(echo "$commands" | sed -E "s/--shard-id[[:blank:]]*=[[:blank:]]*[0-9]*/--shard-id=${GPU} /g" | sed 's/ \\ / /g')
-    echo "Shard ${GPU} commands:$commands_gpu"
-    echo "Render devices: $BUILDKITE_AGENT_META_DATA_RENDER_DEVICES"
-    docker run \
-        --device /dev/kfd $BUILDKITE_AGENT_META_DATA_RENDER_DEVICES \
-        --network=host \
-        --shm-size=16gb \
-        --group-add "$render_gid" \
-        -e MIOPEN_DEBUG_CONV_DIRECT=0 \
-        -e MIOPEN_DEBUG_CONV_GEMM=0 \
-        -e VLLM_ROCM_USE_AITER=0 \
-        -e HIP_VISIBLE_DEVICES="${GPU}" \
-        -e HF_TOKEN \
-        -v "${HF_CACHE}:${HF_MOUNT}" \
-        -e "HF_HOME=${HF_MOUNT}" \
-        -e "PYTHONPATH=${MYPYTHONPATH}" \
-        --name "${container_name}_${GPU}" \
-        "${image_name}" \
-        /bin/bash -c "${commands_gpu}" \
-        |& while read -r line; do echo ">>Shard $GPU: $line"; done &
-    PIDS+=($!)
-  done
-  #wait for all processes to finish and collect exit codes
-  for pid in "${PIDS[@]}"; do
-    wait "${pid}"
-    STATUS+=($?)
-  done
-  for st in "${STATUS[@]}"; do
-    if [[ ${st} -ne 0 ]]; then
-      echo "One of the processes failed with $st"
-      exit "${st}"
+echo "Commands:${commands}"
+if /bin/bash -o pipefail -c '
+set -E
+test_status=0
+trap '\''
+    command_status=$?
+    # Keep running subsequent commands, but retain a failure for the final exit.
+    # Prefer a test failure over pytest "no tests collected" when both occur.
+    if (( test_status == 0 || (test_status == 5 && command_status != 5) )); then
+        test_status=${command_status}
     fi
-  done
+'\'' ERR
+eval "$1"
+exit "${test_status}"
+' _ "${commands}"; then
+    exit 0
 else
-  echo "Render devices: $BUILDKITE_AGENT_META_DATA_RENDER_DEVICES"
-  docker run \
-          --device /dev/kfd $BUILDKITE_AGENT_META_DATA_RENDER_DEVICES \
-          --network=host \
-          --shm-size=16gb \
-          --group-add "$render_gid" \
-          -e MIOPEN_DEBUG_CONV_DIRECT=0 \
-          -e MIOPEN_DEBUG_CONV_GEMM=0 \
-          -e VLLM_ROCM_USE_AITER=0 \
-          -e HF_TOKEN \
-          -v "${HF_CACHE}:${HF_MOUNT}" \
-          -e "HF_HOME=${HF_MOUNT}" \
-          -e "PYTHONPATH=${MYPYTHONPATH}" \
-          --name "${container_name}" \
-          "${image_name}" \
-          /bin/bash -c "${commands}"
+    exit_code=$?
 fi
+
+if [[ ${exit_code} -eq 5 && "${VLLM_CI_ALLOW_NO_TESTS:-0}" == "1" ]]; then
+    echo "Pytest collected no tests; VLLM_CI_ALLOW_NO_TESTS=1, treating exit code 5 as success."
+    exit 0
+fi
+
+exit "${exit_code}"

.buildkite/amd/scripts/select_test_suites.py

@@ -0,0 +1,100 @@
+#!/usr/bin/env python3
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
+"""Select AMD ready/merge/nightly suites after Buildkite creates a build.
+
+This repository-side selector does not decide which GitHub label events start
+the external ``vllm-omni-amd-ci`` pipeline. Its Buildkite PR build condition
+must admit the relevant label event before this code can inspect the labels.
+"""
+
+from __future__ import annotations
+
+import argparse
+import sys
+from collections.abc import Iterable
+
+SUITE_SPECS = {
+    "ready": "READY_TESTS:test-amd-ready.yml",
+    "merge": "MERGE_TESTS:test-amd-merge.yml",
+    "nightly": "NIGHTLY_TESTS:test-amd-nightly.yml",
+}
+
+
+def _parse_debug_suites(value: str) -> tuple[str, ...]:
+    suites = tuple(item.strip().lower() for item in value.split(",") if item.strip())
+    if not suites:
+        raise ValueError("DEBUG_TEST_YAML did not contain a test suite")
+
+    seen: set[str] = set()
+    for suite in suites:
+        if suite not in SUITE_SPECS:
+            raise ValueError(f"DEBUG_TEST_YAML entries must be 'merge', 'nightly', or 'ready', got {suite!r}")
+        if suite in seen:
+            raise ValueError(f"duplicate DEBUG_TEST_YAML suite {suite!r}")
+        seen.add(suite)
+    return suites
+
+
+def select_amd_test_suites(
+    *,
+    branch: str,
+    labels: Iterable[str],
+    debug_test_yaml: str = "",
+    nightly: bool = False,
+) -> tuple[str, ...]:
+    """Return ordered suite names while preserving legacy unlabeled PRs."""
+
+    if debug_test_yaml:
+        return _parse_debug_suites(debug_test_yaml)
+    if branch == "main":
+        return ("nightly",) if nightly else ("merge",)
+
+    label_set = {label.strip() for label in labels if label.strip()}
+    selected = tuple(
+        suite
+        for label, suite in (
+            ("ready", "ready"),
+            ("merge-test", "merge"),
+            ("nightly-test", "nightly"),
+        )
+        if label in label_set
+    )
+    # AMD historically uploaded the ready suite for every admitted PR build.
+    # Preserve that fallback when no explicit tier label is present.
+    return selected or ("ready",)
+
+
+def main() -> int:
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--branch", required=True)
+    parser.add_argument("--labels", default="")
+    parser.add_argument("--debug-test-yaml", default="")
+    parser.add_argument("--nightly", type=int, choices=(0, 1), default=0)
+    args = parser.parse_args()
+
+    labels = tuple(args.labels.splitlines())
+    try:
+        suites = select_amd_test_suites(
+            branch=args.branch,
+            labels=labels,
+            debug_test_yaml=args.debug_test_yaml,
+            nightly=bool(args.nightly),
+        )
+    except ValueError as exc:
+        parser.error(str(exc))
+
+    label_set = {label.strip() for label in labels if label.strip()}
+    if args.branch != "main" and not args.debug_test_yaml and not ({"ready", "merge-test", "nightly-test"} & label_set):
+        print(
+            "No AMD ready/merge/nightly label found; preserving the legacy ready-suite fallback.",
+            file=sys.stderr,
+        )
+
+    for suite in suites:
+        print(SUITE_SPECS[suite])
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

.buildkite/amd/test-amd-merge.yml

@@ -7,84 +7,113 @@ steps:
 
 - group: ":card_index_dividers: Simple Test"
   steps:
-    - label: "Simple · Diffusion & Model Executor Test (1/2)"
-      agent_pool: mi325_1
+    - label: "Simple · Model Executor Test · Shard %N/%t"
+      agent_pool: mi300_1
+      parallelism: 2
       depends_on: amd-build
-      timeout_in_minutes: 70
       mirror_hardwares: [amdproduction]
       grade: Blocking
       commands:
         - export VLLM_ROCM_USE_AITER=0
         # ignore test_teacache_extractors.py because it use rocm gemm kernel from vLLM
         # that is not supported on CPU
-        # CPU tests on ROCm gpu machines are very slow, so we give them more time.
-        - "timeout 60m pytest -sv tests/diffusion -m 'core_model and cpu' --ignore=tests/diffusion/cache/test_teacache_extractors.py --cov=vllm_omni --cov-report=term-missing:skip-covered --cov-report=xml"
+        - "pytest -sv tests/model_executor -m 'core_model and cpu' --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --shard-id=$$BUILDKITE_PARALLEL_JOB"
 
-    - label: "Simple · Diffusion & Model Executor Test (2/2)"
-      agent_pool: mi325_1
+    - label: "Simple · Diffusion Test · Shard %N/%t"
+      agent_pool: mi300_1
+      parallelism: 4
       depends_on: amd-build
-      timeout_in_minutes: 70
       mirror_hardwares: [amdproduction]
       grade: Blocking
       commands:
         - export VLLM_ROCM_USE_AITER=0
-        # CPU tests on ROCm gpu machines are very slow, so we give them more time.
-        - "timeout 60m pytest -sv tests/model_executor -m 'core_model and cpu' --cov=vllm_omni --cov-report=term-missing:skip-covered --cov-report=xml"
+        # ignore test_teacache_extractors.py because it use rocm gemm kernel from vLLM
+        # that is not supported on CPU
+        - "pytest -sv tests/diffusion -m 'core_model and cpu' --ignore=tests/diffusion/cache/test_teacache_extractors.py --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --shard-id=$$BUILDKITE_PARALLEL_JOB"
+
+    - label: "Simple · Engine&Entrypoints Test"
+      agent_pool: mi300_1
+      depends_on: amd-build
+      mirror_hardwares: [amdproduction]
+      grade: Blocking
+      commands:
+        - export VLLM_ROCM_USE_AITER=0
+        - "pytest -sv tests/entrypoints tests/engine -m 'core_model and cpu'"
 
     - label: "Simple · Other Test"
-      agent_pool: mi325_1
+      agent_pool: mi300_1
       depends_on: amd-build
-      timeout_in_minutes: 50
       mirror_hardwares: [amdproduction]
       grade: Blocking
       commands:
         - export VLLM_ROCM_USE_AITER=0
         # ignore test_teacache_extractors.py because it use rocm gemm kernel from vLLM
         # that is not supported on CPU
-        - "timeout 40m pytest -sv tests/diffusion tests/model_executor -m 'core_model and cpu' --ignore=tests/diffusion/cache/test_teacache_extractors.py --cov=vllm_omni --cov-report=term-missing:skip-covered --cov-report=xml"
+        - "pytest -sv tests/ -m 'core_model and cpu' --ignore=tests/diffusion --ignore=tests/model_executor --ignore=tests/entrypoints --ignore=tests/engine  --ignore=tests/diffusion/cache/test_teacache_extractors.py"
 
 - group: ":card_index_dividers: Diffusion Test"
   steps:
     - label: "Diffusion · LoRA Test"
-      agent_pool: mi325_1
+      agent_pool: mi300_1
       depends_on: amd-build
       mirror_hardwares: [amdproduction]
       grade: Blocking
+      artifact_paths:
+        - lora-server.log
       commands:
-        - timeout 20m pytest -s -v tests/diffusion/lora/test_images_generations_lora.py
+        - pytest -sv tests/diffusion/lora/ -m 'advanced_model and cuda' --run-level "advanced_model" 2>&1 | tee lora-server.log
 
     - label: "Diffusion Model CPU offloading Test"
-      agent_pool: mi325_1
+      agent_pool: mi300_1
       depends_on: amd-build
       mirror_hardwares: [amdproduction]
       grade: Blocking
       commands:
         - export VLLM_LOGGING_LEVEL=DEBUG
         - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-        - |
-          timeout 20m bash -c '
-            set +e
-            pytest -s -v tests/diffusion/offloader/test_diffusion_cpu_offload.py
-            EXIT1=\$?
-            pytest -s -v tests/diffusion/offloader/test_diffusion_layerwise_offload.py
-            EXIT2=\$?
-            exit \$((EXIT1 | EXIT2))
-          '
+        - pytest -s -v tests/diffusion/offloader/test_diffusion_cpu_offload.py
+        - >-
+          pytest -s -v tests/diffusion/offloader/test_diffusion_layerwise_offload.py
+          --deselect='tests/diffusion/offloader/test_diffusion_layerwise_offload.py::test_layerwise_offload_diffusion_model[stabilityai/stable-audio-open-1.0]'
 
-    - label: "Diffusion Sequence Parallelism Test (Need 4 GPUs)"
-      agent_pool: mi325_4
+    # TODO: Diagnose unstable Stable Audio layerwise-offload memory savings on ROCm.
+    - label: "Diffusion Stable Audio Layerwise Offload Test (quarantined)"
+      agent_pool: mi300_1
       depends_on: amd-build
       mirror_hardwares: [amdproduction]
+      grade: NonBlocking
+      commands:
+        - export VLLM_LOGGING_LEVEL=DEBUG
+        - export VLLM_WORKER_MULTIPROC_METHOD=spawn
+        - >-
+          pytest -s -v
+          'tests/diffusion/offloader/test_diffusion_layerwise_offload.py::test_layerwise_offload_diffusion_model[stabilityai/stable-audio-open-1.0]'
+
+    - label: "Diffusion Sequence Parallelism Test (Need 4 GPUs) · Shard 1/2"
+      agent_pool: mi300_4
+      depends_on: amd-build
+      mirror_hardwares: [amdproduction]
+      timeout_in_minutes: 40
       grade: Blocking
       commands:
         - export VLLM_LOGGING_LEVEL=DEBUG
         - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-        - timeout 20m pytest -s -v tests/diffusion/distributed/test_sequence_parallel.py
-        - timeout 20m pytest -s -v tests/diffusion/distributed/test_ulysses_uaa_perf.py
+        - pytest -s -v tests/diffusion/distributed/test_sequence_parallel.py
+
+    - label: "Diffusion Sequence Parallelism Test (Need 4 GPUs) · Shard 2/2"
+      agent_pool: mi300_4
+      depends_on: amd-build
+      mirror_hardwares: [amdproduction]
+      timeout_in_minutes: 40
+      grade: Blocking
+      commands:
+        - export VLLM_LOGGING_LEVEL=DEBUG
+        - export VLLM_WORKER_MULTIPROC_METHOD=spawn
+        - pytest -s -v tests/diffusion/distributed/test_ulysses_uaa_perf.py
 
     # merge-only tests
     - label: "Diffusion Tensor Parallelism Test"
-      agent_pool: mi325_2
+      agent_pool: mi300_2
       depends_on: amd-build
       mirror_hardwares: [amdproduction]
       grade: Blocking
@@ -92,7 +121,7 @@ steps:
         - pytest -s -v tests/diffusion/distributed/test_tensor_parallel.py
 
     - label: "Diffusion Model Test"
-      agent_pool: mi325_1
+      agent_pool: mi300_1
       depends_on: amd-build
       mirror_hardwares: [amdproduction]
       grade: Blocking
@@ -100,7 +129,7 @@ steps:
         - timeout 30m pytest -s -v tests/e2e/offline_inference/test_z_image.py
 
     - label: "Diffusion GPU Worker Test"
-      agent_pool: mi325_2
+      agent_pool: mi300_2
       depends_on: amd-build
       mirror_hardwares: [amdproduction]
       grade: Blocking
@@ -108,7 +137,7 @@ steps:
         - timeout 20m pytest -s -v tests/diffusion/test_diffusion_worker.py
 
     - label: "Diffusion · Wan22 Test"
-      agent_pool: mi325_1
+      agent_pool: mi300_1
       depends_on: amd-build
       mirror_hardwares: [amdproduction]
       grade: Blocking
@@ -121,7 +150,7 @@ steps:
 
 ## ISSUE depends on `diffusers` package: https://github.com/huggingface/diffusers/issues/13274
 # - label: "Audio Generation Model Test"
-#   agent_pool: mi325_1
+#   agent_pool: mi300_1
 #   depends_on: amd-build
 #   mirror_hardwares: [amdproduction]
 #   grade: Blocking
@@ -132,7 +161,7 @@ steps:
 #     - timeout 20m pytest -s -v tests/e2e/offline_inference/test_stable_audio_expansion.py -m "advanced_model and diffusion and L4" --run-level advanced_model
 
 - label: "Engine Test"
-  agent_pool: mi325_1
+  agent_pool: mi300_1
   depends_on: amd-build
   mirror_hardwares: [amdproduction]
   grade: Blocking
@@ -147,7 +176,7 @@ steps:
 - group: ":card_index_dividers: E2E Test"
   steps:
     - label: "Omni Model Test Qwen2-5-Omni"
-      agent_pool: mi325_4
+      agent_pool: mi300_4
       depends_on: amd-build
       mirror_hardwares: [amdproduction]
       commands:
@@ -155,48 +184,70 @@ steps:
         - pytest -s -v tests/e2e/*/test_qwen2_5_omni_expansion.py --run-level "core_model"
 
     - label: "Omni Model Test Qwen3-Omni"
-      agent_pool: mi325_2
+      agent_pool: mi300_2
       depends_on: amd-build
       mirror_hardwares: [amdproduction]
       grade: Blocking
       commands:
         - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-        - pytest -s -v tests/e2e/offline_inference/test_qwen3_omni.py tests/e2e/online_serving/test_qwen3_omni.py -m 'advanced_model and cuda' --run-level 'advanced_model'
+        - pytest -s -v tests/e2e/offline_inference/test_qwen3_omni.py tests/e2e/online_serving/test_qwen3_omni.py tests/e2e/offline_inference/test_qwen3_omni_colocate_async.py -m 'advanced_model and cuda' --run-level 'advanced_model'
 
-    - label: "Qwen3-TTS CustomVoice E2E Test"
-      agent_pool: mi325_1
+    - label: "Qwen3-TTS CustomVoice E2E Test · Shard %N/%t"
+      agent_pool: mi300_1
+      parallelism: 3
       depends_on: amd-build
       mirror_hardwares: [amdproduction]
       grade: Blocking
+      timeout_in_minutes: 80
       commands:
         - |
-          timeout 50m bash -c '
+          timeout --signal=TERM --kill-after=2m 80m bash -c '
             export VLLM_LOGGING_LEVEL=DEBUG
             export VLLM_WORKER_MULTIPROC_METHOD=spawn
             export VLLM_ALLOW_LONG_MAX_MODEL_LEN="1"
-            pytest -s -v tests/e2e/online_serving/test_qwen3_tts_customvoice.py tests/e2e/offline_inference/test_qwen3_tts_customvoice.py -m "advanced_model" --run-level "advanced_model"
+            TESTS=(
+              "tests/e2e/online_serving/test_qwen3_tts_customvoice.py::test_default_cuda_graph_startup[async_chunk]"
+              "tests/e2e/online_serving/test_qwen3_tts_customvoice.py::test_text_to_audio_002[async_chunk]"
+              "tests/e2e/offline_inference/test_qwen3_tts_customvoice.py::test_text_to_audio_001[no_cuda_graph]"
+            )
+            TEST="$${TESTS[$$BUILDKITE_PARALLEL_JOB]}"
+            pytest -s -v "$$TEST" -m "advanced_model" --run-level "advanced_model"
+          '
+
+    # TODO: Restore this case to the blocking sharded step once its long-running
+    # Whisper validation on AMD CI is resolved.
+    - label: "Qwen3-TTS CustomVoice async_chunk Test (quarantined)"
+      agent_pool: mi300_1
+      depends_on: amd-build
+      mirror_hardwares: [amdproduction]
+      grade: NonBlocking
+      timeout_in_minutes: 80
+      commands:
+        - |
+          timeout --signal=TERM --kill-after=2m 75m bash -c '
+            export VLLM_LOGGING_LEVEL=DEBUG
+            export VLLM_WORKER_MULTIPROC_METHOD=spawn
+            export VLLM_ALLOW_LONG_MAX_MODEL_LEN="1"
+            pytest -s -v "tests/e2e/online_serving/test_qwen3_tts_customvoice.py::test_text_to_audio_001[async_chunk]" -m "advanced_model" --run-level "advanced_model"
           '
 
     - label: "Qwen3-TTS Base E2E Test"
-      agent_pool: mi325_1
+      agent_pool: mi300_1
       depends_on: amd-build
       mirror_hardwares: [amdproduction]
       grade: Blocking
       timeout_in_minutes: 50
       commands:
-        - |
-          timeout 40m bash -c '
-            export VLLM_LOGGING_LEVEL=DEBUG
-            export VLLM_WORKER_MULTIPROC_METHOD=spawn
-            export VLLM_ALLOW_LONG_MAX_MODEL_LEN="1"
-            pytest -s -v tests/e2e/online_serving/test_qwen3_tts_base.py tests/e2e/offline_inference/test_qwen3_tts_base.py -m 'advanced_model and cuda' --run-level 'advanced_model'
-          '
+        - export VLLM_LOGGING_LEVEL=DEBUG
+        - export VLLM_WORKER_MULTIPROC_METHOD=spawn
+        - export VLLM_ALLOW_LONG_MAX_MODEL_LEN="1"
+        - timeout 40m pytest -s -v tests/e2e/online_serving/test_qwen3_tts_base.py tests/e2e/offline_inference/test_qwen3_tts_base.py -m "advanced_model and cuda" --run-level "advanced_model"
 
 # TODO: Bagel test on ROCm is very unstable. @tjtanaa
 # Need to debug before reneable numerical changes across large PRs
 # # split Bagel Model Test with H100 (Real Weights) into three tests
 # - label: "Bagel Text2Img Model Test (1/3)"
-#   agent_pool: mi325_1
+#   agent_pool: mi300_1
 #   depends_on: amd-build
 #   mirror_hardwares: [amdproduction]
 #   grade: Blocking
@@ -208,7 +259,7 @@ steps:
 #     - timeout 30m pytest -s -v tests/e2e/offline_inference/test_bagel.py -m "advanced_model" --run-level "advanced_model" -k "shared_memory" -k "rocm"
 
 # - label: "Bagel Img2Img Model Test (2/3)"
-#   agent_pool: mi325_1
+#   agent_pool: mi300_1
 #   depends_on: amd-build
 #   mirror_hardwares: [amdproduction]
 #   grade: Blocking
@@ -220,7 +271,7 @@ steps:
 #     - timeout 30m pytest -s -v tests/e2e/offline_inference/test_bagel.py -m "advanced_model" --run-level "advanced_model" -k "rocm"
 
 # - label: "Bagel Online Serving Test (3/3)"
-#   agent_pool: mi325_1
+#   agent_pool: mi300_1
 #   depends_on: amd-build
 #   mirror_hardwares: [amdproduction]
 #   grade: Blocking

.buildkite/amd/test-amd-nightly.yml

@@ -0,0 +1,66 @@
+# AMD MI300 `full_model` nightly coverage during burn-in.
+# All jobs are non-blocking until the lane has stable runtime evidence.
+# Historical build #11646 was red in Function Expansion and Accuracy (one
+# async_chunk similarity mismatch; Seed-TTS WER 0.376480 > 0.35).
+
+env:
+  VLLM_WORKER_MULTIPROC_METHOD: spawn
+  HF_HUB_DOWNLOAD_TIMEOUT: 300
+  HF_HUB_ETAG_TIMEOUT: 60
+
+steps:
+  # MI325 is the repository's legacy ROCm pytest marker. The AMD template maps
+  # agent_pool to the current MI300 Buildkite queues independently of that mark.
+  - group: ":full_moon: Qwen3-Omni MI300 Nightly"
+    steps:
+      - label: "Qwen3-Omni Function Expansion"
+        agent_pool: mi300_2
+        depends_on: amd-build
+        mirror_hardwares: [amdproduction]
+        grade: NonBlocking
+        timeout_in_minutes: 180
+        commands:
+          - export VLLM_ALLOW_LONG_MAX_MODEL_LEN="1"
+          - pytest -s -v tests/e2e/online_serving/test_qwen3_omni_expansion.py -m "full_model and rocm and MI325 and cards_2" --run-level "full_model"
+
+      - label: "Qwen3-Omni Accuracy"
+        agent_pool: mi300_2
+        depends_on: amd-build
+        mirror_hardwares: [amdproduction]
+        grade: NonBlocking
+        timeout_in_minutes: 180
+        artifact_paths:
+          - tests/e2e/accuracy/qwen3_omni/results/qwen_omni_acc/*.json
+        commands:
+          - export SEED_TTS_WER_EVAL=1
+          # PyTorch ROCm uses the torch.cuda device namespace.
+          - export SEED_TTS_EVAL_DEVICE=cuda:1
+          - pytest -s -v tests/e2e/accuracy/qwen3_omni/test_qwen3_omni.py -m "full_model and rocm and MI325 and cards_2" --run-level "full_model"
+          # Native tests run from the image copy under /app/vllm-omni, while
+          # Buildkite's artifact hook searches the checkout workspace.
+          - |
+            if [[ "$$PWD" != "$${BUILDKITE_BUILD_CHECKOUT_PATH:?}" ]]; then
+              artifact_dir="$${BUILDKITE_BUILD_CHECKOUT_PATH}/tests/e2e/accuracy/qwen3_omni/results/qwen_omni_acc"
+              mkdir -p "$$artifact_dir"
+              cp -v tests/e2e/accuracy/qwen3_omni/results/qwen_omni_acc/*.json "$$artifact_dir/"
+            fi
+
+      - label: "Qwen3-Omni Documentation Examples"
+        agent_pool: mi300_2
+        depends_on: amd-build
+        mirror_hardwares: [amdproduction]
+        grade: NonBlocking
+        timeout_in_minutes: 120
+        commands:
+          - export VLLM_ALLOW_LONG_MAX_MODEL_LEN="1"
+          - pytest -s -v tests/examples/online_serving/test_qwen3_omni.py -m "full_model and rocm and MI325 and cards_2" --run-level "full_model"
+
+      - label: "Qwen3-Omni AITER-on Smoke"
+        agent_pool: mi300_2
+        depends_on: amd-build
+        mirror_hardwares: [amdproduction]
+        grade: NonBlocking
+        timeout_in_minutes: 90
+        commands:
+          - export VLLM_ROCM_USE_AITER=1
+          - pytest -s -v tests/e2e/online_serving/test_qwen3_omni_expansion.py::test_text_to_text_audio_001 -m "full_model and rocm and MI325 and cards_2" --run-level "full_model"

.buildkite/amd/test-amd-ready.yml

@@ -7,8 +7,9 @@ steps:
 
 - group: ":card_index_dividers: Simple Test"
   steps:
-    - label: "Simple · Diffusion & Model Executor Test"
-      agent_pool: mi325_1
+    - label: "Simple · Model Executor Test · Shard %N/%t"
+      agent_pool: mi300_1
+      parallelism: 2
       depends_on: amd-build
       mirror_hardwares: [amdproduction]
       grade: Blocking
@@ -16,10 +17,30 @@ steps:
         - export VLLM_ROCM_USE_AITER=0
         # ignore test_teacache_extractors.py because it use rocm gemm kernel from vLLM
         # that is not supported on CPU
-        - "timeout 40m pytest -sv tests/diffusion tests/model_executor -m 'core_model and cpu' --ignore=tests/diffusion/cache/test_teacache_extractors.py --cov=vllm_omni --cov-report=term-missing:skip-covered --cov-report=xml"
+        - "pytest -sv tests/model_executor -m 'core_model and cpu' --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --shard-id=$$BUILDKITE_PARALLEL_JOB"
+
+    - label: "Simple · Diffusion Test"
+      agent_pool: mi300_1
+      depends_on: amd-build
+      mirror_hardwares: [amdproduction]
+      grade: Blocking
+      commands:
+        - export VLLM_ROCM_USE_AITER=0
+        # ignore test_teacache_extractors.py because it use rocm gemm kernel from vLLM
+        # that is not supported on CPU
+        - "pytest -sv tests/diffusion -m 'core_model and cpu'  --ignore=tests/diffusion/cache/test_teacache_extractors.py"
+
+    - label: "Simple · Engine&Entrypoints Test"
+      agent_pool: mi300_1
+      depends_on: amd-build
+      mirror_hardwares: [amdproduction]
+      grade: Blocking
+      commands:
+        - export VLLM_ROCM_USE_AITER=0
+        - "pytest -sv tests/entrypoints tests/engine -m 'core_model and cpu'"
 
     - label: "Simple · Other Test"
-      agent_pool: mi325_1
+      agent_pool: mi300_1
       depends_on: amd-build
       mirror_hardwares: [amdproduction]
       grade: Blocking
@@ -27,19 +48,19 @@ steps:
         - export VLLM_ROCM_USE_AITER=0
         # ignore test_teacache_extractors.py because it use rocm gemm kernel from vLLM
         # that is not supported on CPU
-        - "timeout 40m pytest -sv tests/diffusion tests/model_executor -m 'core_model and cpu' --ignore=tests/diffusion/cache/test_teacache_extractors.py --cov=vllm_omni --cov-report=term-missing:skip-covered --cov-report=xml"
+        - "pytest -sv tests/ -m 'core_model and cpu' --ignore=tests/diffusion --ignore=tests/model_executor --ignore=tests/entrypoints --ignore=tests/engine  --ignore=tests/diffusion/cache/test_teacache_extractors.py"
 
 - label: "Custom Pipeline Test"
-  agent_pool: mi325_1
+  agent_pool: mi300_1
   depends_on: amd-build
   mirror_hardwares: [amdproduction]
   grade: Blocking
   commands:
-    - timeout 20m pytest -s -v tests/e2e/features/custom_pipeline/ -m "core_model"
+    - pytest -s -v tests/e2e/features/custom_pipeline/ -m "core_model"
 
 
 - label: "Voxtral TTS CUDA Unit Test"
-  agent_pool: mi325_1
+  agent_pool: mi300_1
   depends_on: amd-build
   mirror_hardwares: [amdproduction]
   grade: Blocking
@@ -49,33 +70,25 @@ steps:
 - group: ":card_index_dividers: Diffusion Test"
   steps:
     - label: "Diffusion · Batch Test"
-      agent_pool: mi325_1
+      agent_pool: mi300_1
       depends_on: amd-build
       mirror_hardwares: [amdproduction]
       grade: Blocking
       commands:
-        - timeout 20m pytest -sv tests/diffusion/batching -m 'core_model and cuda' --run-level "core_model"
+        - pytest -sv tests/diffusion/batching -m 'core_model and cuda' --run-level "core_model"
 
     - label: "Diffusion · Offloader Test"
-      agent_pool: mi325_1
+      agent_pool: mi300_1
       depends_on: amd-build
       mirror_hardwares: [amdproduction]
       grade: Blocking
       commands:
         - export VLLM_LOGGING_LEVEL=DEBUG
         - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-        - |
-          timeout 20m bash -c '
-            set +e
-            pytest -s -v tests/diffusion/offloader/test_diffusion_cpu_offload.py
-            EXIT1=\$?
-            pytest -s -v tests/diffusion/offloader/test_diffusion_layerwise_offload.py
-            EXIT2=\$?
-            exit \$((EXIT1 | EXIT2))
-          '
+        - pytest -s -v tests/diffusion/offloader -m "core_model and cuda" --run-level "core_model"
 
     - label: "Diffusion · Cache Test"
-      agent_pool: mi325_1
+      agent_pool: mi300_1
       depends_on: amd-build
       mirror_hardwares: [amdproduction]
       grade: Blocking
@@ -83,10 +96,10 @@ steps:
         - export GPU_ARCHS=gfx942
         - export VLLM_LOGGING_LEVEL=DEBUG
         - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-        - timeout 15m pytest -s -v -m "core_model and cache and diffusion and not distributed_cuda and L4"
+        - pytest -sv tests/diffusion/cache -m 'core_model and cuda' --run-level "core_model"
 
     - label: "Diffusion Sequence Parallelism Test"
-      agent_pool: mi325_2
+      agent_pool: mi300_2
       depends_on: amd-build
       mirror_hardwares: [amdproduction]
       grade: Blocking
@@ -94,9 +107,14 @@ steps:
         - export VLLM_LOGGING_LEVEL=DEBUG
         - export VLLM_WORKER_MULTIPROC_METHOD=spawn
         - timeout 20m pytest -s -v tests/diffusion/distributed/test_sequence_parallel.py -m core_model
+        - >-
+          timeout 20m pytest -s -v
+          tests/diffusion/models/ltx2/test_ltx2_transformer_ulysses.py
+          -m 'core_model and rocm and cards_2'
+          --run-level "core_model"
 
     - label: "Diffusion GPU Worker Test"
-      agent_pool: mi325_2
+      agent_pool: mi300_2
       depends_on: amd-build
       mirror_hardwares: [amdproduction]
       grade: Blocking
@@ -104,15 +122,21 @@ steps:
         - timeout 20m pytest -s -v tests/diffusion/test_diffusion_worker.py
 
     - label: "Diffusion · Model Test"
-      agent_pool: mi325_1
+      agent_pool: mi300_1
       depends_on: amd-build
       mirror_hardwares: [amdproduction]
       grade: Blocking
       commands:
-        - timeout 15m pytest -sv tests/diffusion/models/ -m 'core_model and cuda' --run-level "core_model"
+        # Preserve legacy one-GPU tests without cards_1 while excluding every
+        # registered multi-GPU marker from this one-GPU job.
+        - >-
+          timeout 15m pytest -sv tests/diffusion/models/
+          -m 'core_model and cuda and not
+          (cards_2 or cards_3 or cards_4 or cards_5 or cards_6 or cards_7 or cards_8)'
+          --run-level "core_model"
 
     - label: "Diffusion · Wan22 Test"
-      agent_pool: mi325_1
+      agent_pool: mi300_1
       depends_on: amd-build
       mirror_hardwares: [amdproduction]
       grade: Blocking
@@ -126,7 +150,7 @@ steps:
 
 ## ISSUE depends on `diffusers` package: https://github.com/huggingface/diffusers/issues/13274
 # - label: "Audio Generation Model Test"
-#   agent_pool: mi325_1
+#   agent_pool: mi300_1
 #   depends_on: amd-build
 #   mirror_hardwares: [amdproduction]
 #   grade: Blocking
@@ -136,8 +160,41 @@ steps:
 #     - export VLLM_WORKER_MULTIPROC_METHOD=spawn
 #     - timeout 20m pytest -s -v tests/e2e/offline_inference/test_stable_audio_expansion.py -m "advanced_model and diffusion and L4" --run-level advanced_model
 
+- label: "ROCm · Engine & Model Executor GPU Coverage (R2-01)"
+  agent_pool: mi300_1
+  depends_on: amd-build
+  mirror_hardwares: [amdproduction]
+  grade: NonBlocking
+  timeout_in_minutes: 20
+  artifact_paths:
+    - "artifacts/rocm-r2-01/**/*"
+  commands:
+    # Native AMD tests run from /app/vllm-omni, while Buildkite uploads
+    # artifacts relative to its checkout directory.
+    - export R2_01_ARTIFACT_DIR="$$BUILDKITE_BUILD_CHECKOUT_PATH/artifacts/rocm-r2-01"
+    - mkdir -p "$$R2_01_ARTIFACT_DIR"
+    # The CUDA ready job also covers tests/worker/test_batched_omni_output.py.
+    # Keep the initial ROCm burn-in scoped to the requested directories.
+    # Verbose pytest output and JUnit retain node IDs without a second
+    # collection-only pass over the same GPU-backed test tree.
+    # The CUDA/L4-only CosyVoice3 ragged-flow parity test cold-JITs multiple
+    # AITER kernels on ROCm and is covered by the CUDA ready lane.
+    - >-
+      pytest -s -v -ra
+      tests/engine/ tests/model_executor/
+      -m 'core_model and cuda and not
+      (cards_2 or cards_3 or cards_4 or cards_5 or cards_6 or cards_7 or cards_8)'
+      -k 'not test_code2wav_streaming_batch_matches_ragged_flow_numerics'
+      --run-level "core_model"
+      --durations=0
+      --junitxml="$$R2_01_ARTIFACT_DIR/pytest.xml"
+      2>&1 | tee "$$R2_01_ARTIFACT_DIR/pytest.log"
+    - >-
+      tail -n 100 "$$R2_01_ARTIFACT_DIR/pytest.log" >
+      "$$R2_01_ARTIFACT_DIR/pytest-summary.txt"
+
 - label: "Engine Test"
-  agent_pool: mi325_1
+  agent_pool: mi300_1
   depends_on: amd-build
   mirror_hardwares: [amdproduction]
   grade: Blocking
@@ -148,8 +205,73 @@ steps:
               pytest -s -v tests/engine/test_async_omni_engine_abort.py
       '
 
+- label: "ROCm · Entrypoints GPU Coverage (R2-02)"
+  agent_pool: mi300_1
+  depends_on: amd-build
+  mirror_hardwares: [amdproduction]
+  grade: NonBlocking
+  timeout_in_minutes: 30
+  env:
+    VLLM_WORKER_MULTIPROC_METHOD: spawn
+  artifact_paths:
+    - "artifacts/rocm-r2-02/**/*"
+  commands:
+    # Native AMD tests run from /app/vllm-omni, while Buildkite uploads
+    # artifacts relative to its checkout directory.
+    - export R2_02_ARTIFACT_DIR="$$BUILDKITE_BUILD_CHECKOUT_PATH/artifacts/rocm-r2-02"
+    - mkdir -p "$$R2_02_ARTIFACT_DIR"
+    - >-
+      ps -eo pid=,ppid=,lstart=,stat=,comm= >
+      "$$R2_02_ARTIFACT_DIR/processes-before.txt"
+    # Verbose pytest output and JUnit retain selected node IDs without a
+    # second collection-only pass over the same GPU-backed test tree.
+    # The inner timeout leaves five minutes for teardown verification and
+    # artifact generation before the Buildkite step timeout.
+    - >-
+      timeout --signal=TERM --kill-after=1m 25m
+      pytest -s -v -ra
+      tests/entrypoints/
+      -m 'core_model and cuda and not
+      (cards_2 or cards_3 or cards_4 or cards_5 or cards_6 or cards_7 or cards_8)'
+      --run-level "core_model"
+      --durations=0
+      --junitxml="$$R2_02_ARTIFACT_DIR/pytest.xml"
+      2>&1 | tee "$$R2_02_ARTIFACT_DIR/pytest.log"
+    - >-
+      tail -n 100 "$$R2_02_ARTIFACT_DIR/pytest.log" >
+      "$$R2_02_ARTIFACT_DIR/pytest-summary.txt"
+    - |
+      sleep 5
+      ps -eo pid=,ppid=,lstart=,stat=,comm= > "$$R2_02_ARTIFACT_DIR/processes-after.txt"
+      # Compare PID plus process start time so PID reuse cannot hide a leak.
+      # Double dollars preserve awk field references through Buildkite interpolation.
+      awk '
+        NR == FNR {
+          if ($$9 != "ps") {
+            baseline[$$1 "|" $$3 "|" $$4 "|" $$5 "|" $$6 "|" $$7] = 1
+          }
+          next
+        }
+        $$9 != "ps" {
+          identity = $$1 "|" $$3 "|" $$4 "|" $$5 "|" $$6 "|" $$7
+          if (!(identity in baseline)) {
+            print
+            leaked = 1
+          }
+        }
+        END {
+          if (leaked) {
+            print "status=FAIL leaked_processes_detected=1"
+            exit 1
+          }
+          print "status=PASS leaked_processes=0"
+        }
+      ' "$$R2_02_ARTIFACT_DIR/processes-before.txt" \
+        "$$R2_02_ARTIFACT_DIR/processes-after.txt" \
+        2>&1 | tee "$$R2_02_ARTIFACT_DIR/process-cleanup.txt"
+
 # - label: "Omni Model Test Qwen3-Omni"
-#   agent_pool: mi325_2
+#   agent_pool: mi300_2
 #   depends_on: amd-build
 #   mirror_hardwares: [amdproduction]
 #   grade: Blocking
@@ -162,7 +284,7 @@ steps:
 - group: ":card_index_dividers: E2E Test"
   steps:
     - label: "Qwen3-TTS E2E Test"
-      agent_pool: mi325_1
+      agent_pool: mi300_1
       depends_on: amd-build
       mirror_hardwares: [amdproduction]
       grade: Blocking
@@ -173,20 +295,17 @@ steps:
         - timeout 30m pytest -s -v tests/e2e/online_serving/test_qwen3_tts_customvoice.py -m "core_model" --run-level "core_model"
 
     - label: "CosyVoice3-TTS E2E Test"
-      agent_pool: mi325_1
+      agent_pool: mi300_1
       depends_on: amd-build
       mirror_hardwares: [amdproduction]
       grade: Blocking
       commands:
-        - |
-          timeout 20m bash -c '
-            pytest -s -v tests/e2e/online_serving/test_cosyvoice3_tts_expansion.py -m "slow" --run-level "core_model"
-          '
+        - pytest -s -v tests/e2e/online_serving/test_cosyvoice3_tts_expansion.py -m "slow" --run-level "core_model"
 
 # TODO: Bagel test on ROCm is very unstable. @tjtanaa
 # Need to debug before reneable numerical changes across large PRs
 # - label: "Bagel Text2Img Model Test"
-#   agent_pool: mi325_1
+#   agent_pool: mi300_1
 #   depends_on: amd-build
 #   mirror_hardwares: [amdproduction]
 #   grade: Blocking
@@ -198,7 +317,7 @@ steps:
 #     - timeout 30m pytest -s -v tests/e2e/offline_inference/test_bagel.py -m "core_model" --run-level "core_model" -k "rocm"
 
 # - label: "Bagel Img2Img Model Test"
-#   agent_pool: mi325_1
+#   agent_pool: mi300_1
 #   depends_on: amd-build
 #   mirror_hardwares: [amdproduction]
 #   grade: Blocking
@@ -210,7 +329,7 @@ steps:
 #     - timeout 30m pytest -s -v tests/e2e/offline_inference/test_bagel.py -m "core_model" --run-level "core_model" -k "rocm"
 
 # - label: "Bagel Online Serving Test"
-#   agent_pool: mi325_1
+#   agent_pool: mi300_1
 #   depends_on: amd-build
 #   mirror_hardwares: [amdproduction]
 #   grade: Blocking

.buildkite/amd/test-template-amd-omni.j2

@@ -1,11 +1,65 @@
 {# vllm-omni customized version
    Based on: https://github.com/vllm-project/ci-infra/blob/main/buildkite/test-template-amd.j2
-   Last synced: 2025-12-15
-   Modifications: Removed unused CUDA/NVIDIA logic, keeping only AMD tests
+   Last synced: 2026-08-04
+   Modifications: Use the shared MI300 queues and Omni command-passing contract
 #}
 {% set docker_image_amd = "rocm/vllm-omni:$BUILDKITE_COMMIT" %}
 {% set default_working_dir = "/app/vllm-omni" %}
 
+{# MI300 runs tests natively in Kubernetes; Docker-in-Docker is unavailable. #}
+{% macro amd_native_k8s_plugin(indent, step) -%}
+{{ indent }}plugins:
+{{ indent }}  - kubernetes:
+{{ indent }}      podSpecPatch:
+{{ indent }}        automountServiceAccountToken: false
+{{ indent }}        securityContext:
+{{ indent }}          seccompProfile:
+{{ indent }}            type: RuntimeDefault
+{{ indent }}        imagePullSecrets:
+{{ indent }}          - name: docker-config
+{{ indent }}        containers:
+{{ indent }}          - name: container-0
+{{ indent }}            image: {{ docker_image_amd }}
+{{ indent }}            imagePullPolicy: Always
+{{ indent }}            securityContext:
+{{ indent }}              allowPrivilegeEscalation: false
+{{ indent }}              capabilities:
+{{ indent }}                add:
+{{ indent }}                  - IPC_LOCK
+{{ indent }}            resources:
+{{ indent }}              limits:
+{{ indent }}                amd.com/gpu: "{{ step.agent_pool[-1:] }}"
+{{ indent }}              requests:
+{{ indent }}                amd.com/gpu: "{{ step.agent_pool[-1:] }}"
+{{ indent }}            volumeMounts:
+{{ indent }}              - name: devshm
+{{ indent }}                mountPath: /dev/shm
+{{ indent }}            env:
+{{ indent }}              - name: AMD_CI_RUNTIME
+{{ indent }}                value: native
+{{ indent }}              - name: NATIVE_CI
+{{ indent }}                value: "true"
+{{ indent }}              - name: VLLM_CI_DOCKER_DISABLED
+{{ indent }}                value: "1"
+{{ indent }}              - name: VLLM_CI_EXPECTED_GPU_COUNT
+{{ indent }}                value: "{{ step.agent_pool[-1:] }}"
+{{ indent }}              - name: HF_HOME
+{{ indent }}                value: /home/buildkite-agent/huggingface
+{{ indent }}              - name: MIOPEN_DEBUG_CONV_DIRECT
+{{ indent }}                value: "0"
+{{ indent }}              - name: MIOPEN_DEBUG_CONV_GEMM
+{{ indent }}                value: "0"
+{{ indent }}              - name: VLLM_ROCM_USE_AITER
+{{ indent }}                value: "0"
+{{ indent }}              - name: PYTORCH_ROCM_ARCH
+{{ indent }}                value: ""
+{{ indent }}        volumes:
+{{ indent }}          - name: devshm
+{{ indent }}            emptyDir:
+{{ indent }}              medium: Memory
+{{ indent }}              sizeLimit: 16Gi
+{%- endmacro %}
+
 {% if env %}
 env:
 {% for name in env %}
@@ -53,19 +107,31 @@ steps:
           {% if step.agent_pool %}
           queue: amd_{{ step.agent_pool }}
           {% else %}
-          queue: amd_mi325_1
+          queue: amd_mi300_1
           {% endif %}
+{{ amd_native_k8s_plugin("        ", step) }}
 {% set cmd_body = (step.command or (step.commands | join("\n"))) | trim %}
 {% set indented_cmd = cmd_body | replace("\n", "\n            ") %}
         command: bash .buildkite/amd/scripts/run-amd-test.sh
         env:
-          DOCKER_BUILDKIT: "1"
+          AMD_CI_RUNTIME: "native"
+          NATIVE_CI: "true"
+          VLLM_CI_DOCKER_DISABLED: "1"
+          VLLM_CI_EXPECTED_GPU_COUNT: "{{ step.agent_pool[-1:] }}"
+          VLLM_CI_REQUIRE_PERSISTENT_HF_CACHE: "1"
+          HF_HOME: "/home/buildkite-agent/huggingface"
+          HF_HUB_DOWNLOAD_TIMEOUT: "300"
+          HF_HUB_ETAG_TIMEOUT: "60"
+          MIOPEN_DEBUG_CONV_DIRECT: "0"
+          MIOPEN_DEBUG_CONV_GEMM: "0"
+          VLLM_ROCM_USE_AITER: "0"
+          PYTORCH_ROCM_ARCH: ""
 {% if step.env %}
 {% for name in step.env %}
           {{ name }}: "{{ step.env[name] }}"
 {% endfor %}
 {% endif %}
-          TEST_COMMAND: |-
+          TEST_COMMANDS: |-
             (command rocm-smi || true) && cd {{ (step.working_dir or default_working_dir) | safe }}
 {% if env %}
 {% for name in env %}
@@ -82,6 +148,15 @@ steps:
 {% endif %}
             {{ indented_cmd | safe }}
         priority: 100
+{% if step.parallelism %}
+        parallelism: {{ step.parallelism }}
+{% endif %}
+{% if step.artifact_paths %}
+        artifact_paths:
+{% for path in step.artifact_paths %}
+          - "{{ path }}"
+{% endfor %}
+{% endif %}
         {% if step.timeout_in_minutes %}
         timeout_in_minutes: {{ step.timeout_in_minutes }}
         {% endif %}
@@ -100,19 +175,31 @@ steps:
       {% if step.agent_pool %}
       queue: amd_{{ step.agent_pool }}
       {% else %}
-      queue: amd_mi325_1
+      queue: amd_mi300_1
       {% endif %}
+{{ amd_native_k8s_plugin("    ", step) }}
 {% set cmd_body = (step.command or (step.commands | join("\n"))) | trim %}
 {% set indented_cmd = cmd_body | replace("\n", "\n        ") %}
     command: bash .buildkite/amd/scripts/run-amd-test.sh
     env:
-      DOCKER_BUILDKIT: "1"
+      AMD_CI_RUNTIME: "native"
+      NATIVE_CI: "true"
+      VLLM_CI_DOCKER_DISABLED: "1"
+      VLLM_CI_EXPECTED_GPU_COUNT: "{{ step.agent_pool[-1:] }}"
+      VLLM_CI_REQUIRE_PERSISTENT_HF_CACHE: "1"
+      HF_HOME: "/home/buildkite-agent/huggingface"
+      HF_HUB_DOWNLOAD_TIMEOUT: "300"
+      HF_HUB_ETAG_TIMEOUT: "60"
+      MIOPEN_DEBUG_CONV_DIRECT: "0"
+      MIOPEN_DEBUG_CONV_GEMM: "0"
+      VLLM_ROCM_USE_AITER: "0"
+      PYTORCH_ROCM_ARCH: ""
 {% if step.env %}
 {% for name in step.env %}
       {{ name }}: "{{ step.env[name] }}"
 {% endfor %}
 {% endif %}
-      TEST_COMMAND: |-
+      TEST_COMMANDS: |-
         (command rocm-smi || true) && cd {{ (step.working_dir or default_working_dir) | safe }}
 {% if env %}
 {% for name in env %}
@@ -129,6 +216,15 @@ steps:
 {% endif %}
         {{ indented_cmd | safe }}
     priority: 100
+{% if step.parallelism %}
+    parallelism: {{ step.parallelism }}
+{% endif %}
+{% if step.artifact_paths %}
+    artifact_paths:
+{% for path in step.artifact_paths %}
+      - "{{ path }}"
+{% endfor %}
+{% endif %}
     {% if step.timeout_in_minutes %}
     timeout_in_minutes: {{ step.timeout_in_minutes }}
     {% endif %}

.buildkite/common/ci_mirror_hardwares.yml

@@ -1,14 +1,36 @@
 # Uploader-only hardware presets for CUDA / NPU test YAML.
 # Referenced as ``mirror_hardwares: l4_1`` on a step; ``upload_pipeline.py`` expands to
-# ``agents`` (+ optional top-level ``image`` for NPU) + ``plugins``, then strips
-# ``mirror_hardwares`` before Buildkite upload.
+# ``agents`` (+ optional top-level ``image`` for NPU) + ``plugins`` (+ ``retry``
+# for ``l4_*``), then strips ``mirror_hardwares`` before Buildkite upload.
+#
+# Omit the key: compose {chip}_{n} from pytest -m SKU + cards_n (max of several
+# positives; not cards_1 with no positive cards_* → that chip's highest existing preset).
+# Unset MIRROR_HW matches H100 or L4 in -m (both → H100); no match skips the step.
+# MIRROR_HW=b200 matches B200 in -m (else the step is skipped).
+# MIRROR_HW must be empty or b200 (case-insensitive); unknown values fail the upload.
+# CUDA preset strings (h100_4, l4_1) are omitted when MIRROR_HW=b200.
+# Unset MIRROR_HW keeps string presets. NPU presets ignore MIRROR_HW.
 #
 # CUDA naming: {chip}_{count}
-#   l4_1, l4_4, h100_1 … h100_4
+#   l4_1, l4_2, l4_3, l4_4, h100_1 … h100_4, b200_1 … b200_4
 # NPU naming: {queue}_{resource_class}
 #   a2b3_npu_1  — ascend-a2b3, resource_class npu-1 (A2/B3 CI image)
 #   a2b3_npu_4  — ascend-a2b3, resource_class npu-4
-#   a3_npu_2    — ascend-a3, resource_class npu-2
+#   a3_npu_2    — ascend-a3, resource_class npu-2 (A3 CI image tag …:$BUILDKITE_COMMIT-a3)
+#   a5_npu_2    — ascend-a5, resource_class npu-2
+#   310p_npu_1  — ascend-310p, resource_class npu-1
+#
+# B200 (b200-k8s): matches vLLM device ``b200-k8s`` from ci-infra
+# (AgentQueue.B200_K8S + b200_plugin_template). Queue routes the K8s pool;
+# podSpec uses hostNetwork + /raid|/mnt/shared (no H100-style nodeSelector).
+#
+# The ``l4_*`` presets run on the EKS ``l4-k8s`` queue (agent-stack-k8s),
+# packed onto g6.12xlarge L4 nodes (``vllm.ci/gpu-pool=l4x4``). GPU / CPU /
+# memory / shm requests scale with card count (1 GPU = 10 CPU / 40Gi / 24Gi shm).
+# Retry covers lost agents / killed pods, not pytest or OOM (exit 1 / 137).
+# The Kubernetes HF secret deliberately uses VLLM_CI_HF_TOKEN. Scheduled
+# build-level HF_TOKEN values can override pod env entries with the same name;
+# upload_pipeline.py restores the secret-backed value inside the job shell.
 
 x-npu-k8s-plugins: &npu_k8s_plugins
   - kubernetes:
@@ -16,37 +38,224 @@ x-npu-k8s-plugins: &npu_k8s_plugins
         imagePullSecrets:
           - name: swr-secret
 
+# Infra-only retry. Omit exit 1 (pytest failed) and 137 (cgroup OOM / SIGKILL
+# of the test process). ci-infra ``K8S_RETRY`` still retries exit 1.
+x-l4-k8s-retry: &l4_k8s_retry
+  automatic:
+    - exit_status: -1
+      limit: 1
+    - exit_status: 128
+      limit: 1
+    - signal_reason: agent_stop
+      limit: 1
+    - signal_reason: agent_refused
+      limit: 1
+
 
 mirror_hardwares:
   l4_1:
     agents:
-      queue: gpu_1_queue
+      queue: l4-k8s
+    retry: *l4_k8s_retry
     plugins:
-      - docker#v5.2.0:
-          image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
-          always-pull: true
-          propagate-environment: true
-          shm-size: "8gb"
-          environment:
-            - HF_HOME=/fsx/hf_cache
-            - HF_TOKEN
-          volumes:
-            - /fsx/hf_cache:/fsx/hf_cache
+      - kubernetes:
+          metadata:
+            annotations:
+              cluster-autoscaler.kubernetes.io/safe-to-evict: "false"
+              karpenter.sh/do-not-disrupt: "true"
+          podSpec:
+            priorityClassName: ci
+            serviceAccountName: buildkite-gpu-jobs
+            tolerations:
+              - key: nvidia.com/gpu
+                operator: Exists
+                effect: NoSchedule
+            nodeSelector:
+              vllm.ci/gpu-pool: l4x4
+            containers:
+              - image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
+                resources:
+                  limits:
+                    nvidia.com/gpu: 1
+                  requests:
+                    nvidia.com/gpu: 1
+                    cpu: "10"
+                    memory: 40Gi
+                volumeMounts:
+                  - name: devshm
+                    mountPath: /dev/shm
+                  - name: hf-cache
+                    mountPath: /fsx/hf_cache
+                env:
+                  - name: HF_HOME
+                    value: /fsx/hf_cache
+                  - name: VLLM_CI_HF_TOKEN
+                    valueFrom:
+                      secretKeyRef:
+                        name: hf-token-secret
+                        key: token
+            volumes:
+              - name: devshm
+                emptyDir:
+                  medium: Memory
+                  sizeLimit: 24Gi
+              - name: hf-cache
+                hostPath:
+                  path: /fsx/hf_cache
+                  type: DirectoryOrCreate
+
+  l4_2:
+    agents:
+      queue: l4-k8s
+    retry: *l4_k8s_retry
+    plugins:
+      - kubernetes:
+          metadata:
+            annotations:
+              cluster-autoscaler.kubernetes.io/safe-to-evict: "false"
+              karpenter.sh/do-not-disrupt: "true"
+          podSpec:
+            priorityClassName: ci
+            serviceAccountName: buildkite-gpu-jobs
+            tolerations:
+              - key: nvidia.com/gpu
+                operator: Exists
+                effect: NoSchedule
+            nodeSelector:
+              vllm.ci/gpu-pool: l4x4
+            containers:
+              - image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
+                resources:
+                  limits:
+                    nvidia.com/gpu: 2
+                  requests:
+                    nvidia.com/gpu: 2
+                    cpu: "20"
+                    memory: 80Gi
+                volumeMounts:
+                  - name: devshm
+                    mountPath: /dev/shm
+                  - name: hf-cache
+                    mountPath: /fsx/hf_cache
+                env:
+                  - name: HF_HOME
+                    value: /fsx/hf_cache
+                  - name: VLLM_CI_HF_TOKEN
+                    valueFrom:
+                      secretKeyRef:
+                        name: hf-token-secret
+                        key: token
+            volumes:
+              - name: devshm
+                emptyDir:
+                  medium: Memory
+                  sizeLimit: 48Gi
+              - name: hf-cache
+                hostPath:
+                  path: /fsx/hf_cache
+                  type: DirectoryOrCreate
+
+  l4_3:
+    agents:
+      queue: l4-k8s
+    retry: *l4_k8s_retry
+    plugins:
+      - kubernetes:
+          metadata:
+            annotations:
+              cluster-autoscaler.kubernetes.io/safe-to-evict: "false"
+              karpenter.sh/do-not-disrupt: "true"
+          podSpec:
+            priorityClassName: ci
+            serviceAccountName: buildkite-gpu-jobs
+            tolerations:
+              - key: nvidia.com/gpu
+                operator: Exists
+                effect: NoSchedule
+            nodeSelector:
+              vllm.ci/gpu-pool: l4x4
+            containers:
+              - image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
+                resources:
+                  limits:
+                    nvidia.com/gpu: 3
+                  requests:
+                    nvidia.com/gpu: 3
+                    cpu: "30"
+                    memory: 120Gi
+                volumeMounts:
+                  - name: devshm
+                    mountPath: /dev/shm
+                  - name: hf-cache
+                    mountPath: /fsx/hf_cache
+                env:
+                  - name: HF_HOME
+                    value: /fsx/hf_cache
+                  - name: VLLM_CI_HF_TOKEN
+                    valueFrom:
+                      secretKeyRef:
+                        name: hf-token-secret
+                        key: token
+            volumes:
+              - name: devshm
+                emptyDir:
+                  medium: Memory
+                  sizeLimit: 72Gi
+              - name: hf-cache
+                hostPath:
+                  path: /fsx/hf_cache
+                  type: DirectoryOrCreate
 
   l4_4:
     agents:
-      queue: gpu_4_queue
+      queue: l4-k8s
+    retry: *l4_k8s_retry
     plugins:
-      - docker#v5.2.0:
-          image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
-          always-pull: true
-          propagate-environment: true
-          shm-size: "8gb"
-          environment:
-            - HF_HOME=/fsx/hf_cache
-            - HF_TOKEN
-          volumes:
-            - /fsx/hf_cache:/fsx/hf_cache
+      - kubernetes:
+          metadata:
+            annotations:
+              cluster-autoscaler.kubernetes.io/safe-to-evict: "false"
+              karpenter.sh/do-not-disrupt: "true"
+          podSpec:
+            priorityClassName: ci
+            serviceAccountName: buildkite-gpu-jobs
+            tolerations:
+              - key: nvidia.com/gpu
+                operator: Exists
+                effect: NoSchedule
+            nodeSelector:
+              vllm.ci/gpu-pool: l4x4
+            containers:
+              - image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
+                resources:
+                  limits:
+                    nvidia.com/gpu: 4
+                  requests:
+                    nvidia.com/gpu: 4
+                    cpu: "40"
+                    memory: 160Gi
+                volumeMounts:
+                  - name: devshm
+                    mountPath: /dev/shm
+                  - name: hf-cache
+                    mountPath: /fsx/hf_cache
+                env:
+                  - name: HF_HOME
+                    value: /fsx/hf_cache
+                  - name: VLLM_CI_HF_TOKEN
+                    valueFrom:
+                      secretKeyRef:
+                        name: hf-token-secret
+                        key: token
+            volumes:
+              - name: devshm
+                emptyDir:
+                  medium: Memory
+                  sizeLimit: 96Gi
+              - name: hf-cache
+                hostPath:
+                  path: /fsx/hf_cache
+                  type: DirectoryOrCreate
 
   h100_1:
     agents:
@@ -67,7 +276,7 @@ mirror_hardwares:
                 env:
                   - name: HF_HOME
                     value: /root/.cache/huggingface
-                  - name: HF_TOKEN
+                  - name: VLLM_CI_HF_TOKEN
                     valueFrom:
                       secretKeyRef:
                         name: hf-token-secret
@@ -102,7 +311,7 @@ mirror_hardwares:
                 env:
                   - name: HF_HOME
                     value: /root/.cache/huggingface
-                  - name: HF_TOKEN
+                  - name: VLLM_CI_HF_TOKEN
                     valueFrom:
                       secretKeyRef:
                         name: hf-token-secret
@@ -137,7 +346,7 @@ mirror_hardwares:
                 env:
                   - name: HF_HOME
                     value: /root/.cache/huggingface
-                  - name: HF_TOKEN
+                  - name: VLLM_CI_HF_TOKEN
                     valueFrom:
                       secretKeyRef:
                         name: hf-token-secret
@@ -172,7 +381,7 @@ mirror_hardwares:
                 env:
                   - name: HF_HOME
                     value: /root/.cache/huggingface
-                  - name: HF_TOKEN
+                  - name: VLLM_CI_HF_TOKEN
                     valueFrom:
                       secretKeyRef:
                         name: hf-token-secret
@@ -188,6 +397,211 @@ mirror_hardwares:
                   path: /mnt/hf-cache
                   type: DirectoryOrCreate
 
+  # vLLM device: b200-k8s → agents.queue b200-k8s + Buildkite K8s agent-stack podSpec
+  b200_1:
+    agents:
+      queue: b200-k8s
+    plugins:
+      - kubernetes:
+          podSpec:
+            runtimeClassName: nvidia
+            hostNetwork: true
+            dnsPolicy: ClusterFirstWithHostNet
+            imagePullSecrets:
+              - name: k8s-ecr-login-renew-docker-secret
+            containers:
+              - image: 936637512419.dkr.ecr.us-west-2.amazonaws.com/vllm-ci-pull-through-cache/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
+                resources:
+                  limits:
+                    nvidia.com/gpu: 1
+                securityContext:
+                  capabilities:
+                    add:
+                      - IPC_LOCK
+                      - SYS_RESOURCE
+                volumeMounts:
+                  - name: devshm
+                    mountPath: /dev/shm
+                  - name: raid
+                    mountPath: /raid
+                  - name: shared
+                    mountPath: /mnt/shared
+                env:
+                  - name: NCCL_CUMEM_HOST_ENABLE
+                    value: "0"
+                  - name: HF_HOME
+                    value: /raid/hf_cache
+                  - name: VLLM_CI_HF_TOKEN
+                    valueFrom:
+                      secretKeyRef:
+                        name: hf-token-secret
+                        key: token
+            volumes:
+              - name: devshm
+                emptyDir:
+                  medium: Memory
+              - name: raid
+                hostPath:
+                  path: /raid
+                  type: DirectoryOrCreate
+              - name: shared
+                hostPath:
+                  path: /mnt/shared
+                  type: DirectoryOrCreate
+
+  b200_2:
+    agents:
+      queue: b200-k8s
+    plugins:
+      - kubernetes:
+          podSpec:
+            runtimeClassName: nvidia
+            hostNetwork: true
+            dnsPolicy: ClusterFirstWithHostNet
+            imagePullSecrets:
+              - name: k8s-ecr-login-renew-docker-secret
+            containers:
+              - image: 936637512419.dkr.ecr.us-west-2.amazonaws.com/vllm-ci-pull-through-cache/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
+                resources:
+                  limits:
+                    nvidia.com/gpu: 2
+                securityContext:
+                  capabilities:
+                    add:
+                      - IPC_LOCK
+                      - SYS_RESOURCE
+                volumeMounts:
+                  - name: devshm
+                    mountPath: /dev/shm
+                  - name: raid
+                    mountPath: /raid
+                  - name: shared
+                    mountPath: /mnt/shared
+                env:
+                  - name: NCCL_CUMEM_HOST_ENABLE
+                    value: "0"
+                  - name: HF_HOME
+                    value: /raid/hf_cache
+                  - name: VLLM_CI_HF_TOKEN
+                    valueFrom:
+                      secretKeyRef:
+                        name: hf-token-secret
+                        key: token
+            volumes:
+              - name: devshm
+                emptyDir:
+                  medium: Memory
+              - name: raid
+                hostPath:
+                  path: /raid
+                  type: DirectoryOrCreate
+              - name: shared
+                hostPath:
+                  path: /mnt/shared
+                  type: DirectoryOrCreate
+
+  b200_3:
+    agents:
+      queue: b200-k8s
+    plugins:
+      - kubernetes:
+          podSpec:
+            runtimeClassName: nvidia
+            hostNetwork: true
+            dnsPolicy: ClusterFirstWithHostNet
+            imagePullSecrets:
+              - name: k8s-ecr-login-renew-docker-secret
+            containers:
+              - image: 936637512419.dkr.ecr.us-west-2.amazonaws.com/vllm-ci-pull-through-cache/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
+                resources:
+                  limits:
+                    nvidia.com/gpu: 3
+                securityContext:
+                  capabilities:
+                    add:
+                      - IPC_LOCK
+                      - SYS_RESOURCE
+                volumeMounts:
+                  - name: devshm
+                    mountPath: /dev/shm
+                  - name: raid
+                    mountPath: /raid
+                  - name: shared
+                    mountPath: /mnt/shared
+                env:
+                  - name: NCCL_CUMEM_HOST_ENABLE
+                    value: "0"
+                  - name: HF_HOME
+                    value: /raid/hf_cache
+                  - name: VLLM_CI_HF_TOKEN
+                    valueFrom:
+                      secretKeyRef:
+                        name: hf-token-secret
+                        key: token
+            volumes:
+              - name: devshm
+                emptyDir:
+                  medium: Memory
+              - name: raid
+                hostPath:
+                  path: /raid
+                  type: DirectoryOrCreate
+              - name: shared
+                hostPath:
+                  path: /mnt/shared
+                  type: DirectoryOrCreate
+
+  b200_4:
+    agents:
+      queue: b200-k8s
+    plugins:
+      - kubernetes:
+          podSpec:
+            runtimeClassName: nvidia
+            hostNetwork: true
+            dnsPolicy: ClusterFirstWithHostNet
+            imagePullSecrets:
+              - name: k8s-ecr-login-renew-docker-secret
+            containers:
+              - image: 936637512419.dkr.ecr.us-west-2.amazonaws.com/vllm-ci-pull-through-cache/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
+                resources:
+                  limits:
+                    nvidia.com/gpu: 4
+                securityContext:
+                  capabilities:
+                    add:
+                      - IPC_LOCK
+                      - SYS_RESOURCE
+                volumeMounts:
+                  - name: devshm
+                    mountPath: /dev/shm
+                  - name: raid
+                    mountPath: /raid
+                  - name: shared
+                    mountPath: /mnt/shared
+                env:
+                  - name: NCCL_CUMEM_HOST_ENABLE
+                    value: "0"
+                  - name: HF_HOME
+                    value: /raid/hf_cache
+                  - name: VLLM_CI_HF_TOKEN
+                    valueFrom:
+                      secretKeyRef:
+                        name: hf-token-secret
+                        key: token
+            volumes:
+              - name: devshm
+                emptyDir:
+                  medium: Memory
+              - name: raid
+                hostPath:
+                  path: /raid
+                  type: DirectoryOrCreate
+              - name: shared
+                hostPath:
+                  path: /mnt/shared
+                  type: DirectoryOrCreate
+
   a2b3_npu_1:
     agents:
       queue: "ascend-a2b3"
@@ -237,3 +651,52 @@ mirror_hardwares:
       resource_class: "npu-16"
     image: "swr.cn-southwest-2.myhuaweicloud.com/modelfoundry/vllm-omni-ci-npu:${BUILDKITE_COMMIT}-a3"
     plugins: *npu_k8s_plugins
+
+  a5_npu_2:
+    agents:
+      queue: "ascend-a5"
+      resource_class: "npu-2"
+    image: "swr.cn-southwest-2.myhuaweicloud.com/modelfoundry/vllm-omni-ci-npu:${BUILDKITE_COMMIT}-a5"
+    plugins: *npu_k8s_plugins
+
+  a5_npu_4:
+    agents:
+      queue: "ascend-a5"
+      resource_class: "npu-4"
+    image: "swr.cn-southwest-2.myhuaweicloud.com/modelfoundry/vllm-omni-ci-npu:${BUILDKITE_COMMIT}-a5"
+    plugins: *npu_k8s_plugins
+
+  a5_npu_8:
+    agents:
+      queue: "ascend-a5"
+      resource_class: "npu-8"
+    image: "swr.cn-southwest-2.myhuaweicloud.com/modelfoundry/vllm-omni-ci-npu:${BUILDKITE_COMMIT}-a5"
+    plugins: *npu_k8s_plugins
+
+  a5_npu_16:
+    agents:
+      queue: "ascend-a5"
+      resource_class: "npu-16"
+    image: "swr.cn-southwest-2.myhuaweicloud.com/modelfoundry/vllm-omni-ci-npu:${BUILDKITE_COMMIT}-a5"
+    plugins: *npu_k8s_plugins
+
+  310p_npu_1:
+    agents:
+      queue: "ascend-310p"
+      resource_class: "npu-1"
+    image: "swr.cn-southwest-2.myhuaweicloud.com/modelfoundry/vllm-omni-ci-npu:${BUILDKITE_COMMIT}-310p"
+    plugins: *npu_k8s_plugins
+
+  310p_npu_2:
+    agents:
+      queue: "ascend-310p"
+      resource_class: "npu-2"
+    image: "swr.cn-southwest-2.myhuaweicloud.com/modelfoundry/vllm-omni-ci-npu:${BUILDKITE_COMMIT}-310p"
+    plugins: *npu_k8s_plugins
+
+  310p_npu_4:
+    agents:
+      queue: "ascend-310p"
+      resource_class: "npu-4"
+    image: "swr.cn-southwest-2.myhuaweicloud.com/modelfoundry/vllm-omni-ci-npu:${BUILDKITE_COMMIT}-310p"
+    plugins: *npu_k8s_plugins

.buildkite/common/ci_source_file_dependencies.yml

@@ -0,0 +1,458 @@
+# Uploader-only path-prefix presets for CUDA L2–L5 and NPU L4 test YAML.
+# Referenced as ``source_file_dependencies: omni_qwen3_omni_function`` (or a list of
+# keys) on a step; ``upload_pipeline.py`` expands the key(s) only.
+#
+# YAML anchors above the registry hold model business-code paths. Job keys
+# ({model}_function / _perf / _accuracy / _reliability / _doc / _cov) alias
+# those anchors and add the scripts the job actually runs. Feature keys below
+# are unchanged. ``source_filter_fallback`` is not a job key: when no job-key
+# prefix matches, a matching diff keeps every job for that upload.
+#
+# YAML anchors are nested lists; the loader flattens them.
+
+x-qwen3-tts: &qwen3_tts
+  - vllm_omni/engine/async_omni_engine.py
+  - vllm_omni/entrypoints/omni_base.py
+  - vllm_omni/worker/gpu_model_runner.py
+  - vllm_omni/model_executor/models/qwen3_tts/
+  - vllm_omni/model_executor/models/common/qwen3_code_predictor.py
+  - vllm_omni/model_executor/models/common/snake_activation.py
+  - vllm_omni/model_executor/models/common/alias_free_activation.py
+  - vllm_omni/model_executor/stage_input_processors/qwen3_tts.py
+  - vllm_omni/deploy/qwen3_tts.yaml
+  - vllm_omni/entrypoints/openai/serving_speech.py
+  - vllm_omni/entrypoints/openai/tts_adapters/
+  - vllm_omni/engine/stage_pool.py
+
+x-minicpmo-4-5: &minicpmo_4_5
+  - vllm_omni/engine/async_omni_engine.py
+  - vllm_omni/entrypoints/omni_base.py
+  - vllm_omni/model_executor/models/minicpmo_4_5/
+  - vllm_omni/model_executor/models/step_audio2/
+  - vllm_omni/model_executor/models/cosyvoice3/code2wav_core/
+  - vllm_omni/model_executor/models/cosyvoice3/utils.py
+  - vllm_omni/model_executor/stage_input_processors/minicpmo_4_5_omni.py
+  - vllm_omni/deploy/minicpmo_4_5.yaml
+  - vllm_omni/transformers_utils/chat_templates/minicpmo45_native.jinja
+
+x-minicpmo-4-5-duplex: &minicpmo_4_5_duplex
+  - vllm_omni/deploy/minicpmo_4_5_duplex.yaml
+  - vllm_omni/engine/duplex/
+  - vllm_omni/engine/duplex_omni_engine.py
+  - vllm_omni/engine/duplex_orchestrator.py
+  - vllm_omni/engine/omni_engine_base.py
+  - vllm_omni/entrypoints/duplex/
+  - vllm_omni/entrypoints/duplex_omni.py
+  - vllm_omni/entrypoints/async_omni_base.py
+  - vllm_omni/clients/
+  - vllm_omni/outputs/duplex.py
+  - vllm_omni/model_executor/duplex_sampling.py
+  - vllm_omni/experimental/fullduplex/
+  - vllm_omni/core/sched/
+  - vllm_omni/outputs/multimodal_accumulation.py
+  - vllm_omni/engine/stage_init_utils.py
+  - vllm_omni/engine/orchestrator.py
+  - vllm_omni/entrypoints/cli/serve.py
+  - vllm_omni/entrypoints/openai/api_server.py
+  - vllm_omni/entrypoints/openai/duplex_capability.py
+  - vllm_omni/config/stage_config.py
+
+x-qwen2-5-omni: &qwen2_5_omni
+  - vllm_omni/engine/async_omni_engine.py
+  - vllm_omni/entrypoints/omni_base.py
+  - vllm_omni/model_executor/models/qwen2_5_omni/
+  - vllm_omni/model_executor/stage_input_processors/qwen2_5_omni.py
+  - vllm_omni/deploy/qwen2_5_omni.yaml
+
+x-qwen3-omni: &qwen3_omni
+  - vllm_omni/engine/async_omni_engine.py
+  - vllm_omni/entrypoints/omni_base.py
+  - vllm_omni/entrypoints/openai/serving_chat.py
+  - vllm_omni/model_executor/models/qwen3_omni/
+  - vllm_omni/model_executor/models/common/qwen3_code_predictor.py
+  - vllm_omni/model_executor/models/common/snake_activation.py
+  - vllm_omni/model_executor/models/qwen2_5_omni/qwen2_5_omni_thinker.py
+  - vllm_omni/model_executor/models/qwen3_tts/cuda_graph_decoder_wrapper.py
+  - vllm_omni/model_executor/stage_input_processors/qwen3_omni.py
+  - vllm_omni/deploy/qwen3_omni_moe.yaml
+  - vllm_omni/deploy/qwen3_omni_moe_thinking.yaml
+
+x-wan22: &wan22
+  - vllm_omni/diffusion/models/wan2_2/
+  - vllm_omni/diffusion/models/dmd2/
+
+x-cosmos3: &cosmos3
+  - vllm_omni/diffusion/models/cosmos3/
+  - vllm_omni/diffusion/distributed/autoencoders/autoencoder_kl_wan.py
+  - vllm_omni/experimental/world_models/
+  - vllm_omni/model_extras/cosmos3.py
+
+x-qwen-image: &qwen_image
+  - vllm_omni/diffusion/models/qwen_image/
+  - vllm_omni/diffusion/models/dmd2/
+  - vllm_omni/diffusion/distributed/autoencoders/autoencoder_kl_qwenimage.py
+
+x-hunyuan-image3: &hunyuan_image3
+  - vllm_omni/diffusion/models/hunyuan_image3/
+  - vllm_omni/model_executor/models/hunyuan_image3/
+  - vllm_omni/deploy/hunyuan_image3_dit.yaml
+
+x-minimax-h3: &minimax_h3
+  - vllm_omni/diffusion/models/minimax_h3/
+
+x-hunyuan-video: &hunyuan_video
+  - vllm_omni/diffusion/models/hunyuan_video/
+  - vllm_omni/model_executor/models/hunyuan_video/
+
+x-ltx2: &ltx2
+  - vllm_omni/diffusion/models/ltx2/
+  - vllm_omni/model_extras/ltx2.py
+
+x-sana-video: &sana_video
+  - vllm_omni/diffusion/models/sana_video/
+  - vllm_omni/model_extras/sana_video.py
+
+x-z-image: &z_image
+  - vllm_omni/diffusion/models/z_image/
+
+x-stable-audio: &stable_audio
+  - vllm_omni/diffusion/models/stable_audio/
+
+x-lingbot-video: &lingbot_video
+  - vllm_omni/diffusion/models/lingbot_video/
+  - vllm_omni/model_extras/lingbot_video.py
+
+x-lingbot-world: &lingbot_world
+  - vllm_omni/diffusion/models/lingbot_world/
+
+
+source_file_dependencies:
+  # Hardware presets and shared test helpers. Not a job key — do not attach this
+  # to a step. Used only when no job-key prefix matches: then a hit keeps every job.
+  source_filter_fallback:
+    - .buildkite/common/ci_mirror_hardwares.yml
+    - tests/helpers/
+
+  # --- features ---
+  diffusion_distributed_attention:
+    - tests/diffusion/distributed/
+    - tests/diffusion/attention/test_attention_sp.py
+    - vllm_omni/diffusion/distributed/
+    - vllm_omni/diffusion/attention/
+    - vllm_omni/diffusion/data.py
+    - vllm_omni/diffusion/models/
+
+  custom_pipeline:
+    - tests/e2e/features/custom_pipeline/
+    - tests/e2e/features/helpers/custom_pipeline.py
+    - vllm_omni/diffusion/
+    - vllm_omni/engine/
+    - vllm_omni/entrypoints/
+    - vllm_omni/inputs/
+    - vllm_omni/outputs/
+
+  rlhf_verl_omni:
+    - tests/e2e/features/rlhf_test/test_verl_omni_e2e.py
+    - tests/e2e/features/helpers/verl_omni_server.py
+    - tests/e2e/features/helpers/custom_pipeline.py
+    - vllm_omni/diffusion/
+    - vllm_omni/engine/
+    - vllm_omni/entrypoints/
+    - vllm_omni/inputs/
+    - vllm_omni/outputs/
+
+  # Tiny-model diffusion jobs (shared across many model trees under models/).
+  diffusion_tiny_model:
+    - tests/model_tests/diffusion/
+    - vllm_omni/diffusion/models/
+
+  diffusion_offloader:
+    - tests/diffusion/offloader/
+    - vllm_omni/diffusion/offloader/
+
+  diffusion_quantization:
+    - tests/diffusion/quantization/
+    - vllm_omni/diffusion/quantization/
+
+  # --- models ---
+  # --- omni / tts jobs ---
+  omni_qwen3_omni_function:
+    - *qwen3_omni
+    - tests/e2e/offline_inference/test_qwen3_omni.py
+    - tests/e2e/offline_inference/test_qwen3_omni_colocate_async.py
+    - tests/e2e/offline_inference/test_qwen3_omni_autoround_w4a16_expansion.py
+    - tests/e2e/offline_inference/test_qwen3_omni_modelopt_nvfp4_w4a4.py
+    - tests/e2e/online_serving/test_qwen3_omni.py
+    - tests/e2e/online_serving/test_qwen3_omni_expansion.py
+    - tests/e2e/online_serving/test_qwen3_omni_multi_replicas.py
+
+  omni_qwen3_omni_accuracy:
+    - *qwen3_omni
+    - tests/e2e/accuracy/qwen3_omni/
+
+  omni_qwen3_omni_perf:
+    - *qwen3_omni
+    - tests/dfx/perf/scripts/run_benchmark.py
+    - tests/dfx/perf/tests/test_qwen3_omni_vllm_text.json
+    - tests/dfx/perf/tests/test_qwen3_omni_async_chunk.json
+    - tests/dfx/perf/tests/test_qwen3_omni_no_async_chunk.json
+    - tests/dfx/perf/tests/test_qwen3_omni_multi_replicas.json
+
+  omni_qwen3_omni_reliability:
+    - *qwen3_omni
+    - tests/dfx/reliability/test_reliability_qwen3_omni.py
+
+  omni_qwen3_omni_doc:
+    - *qwen3_omni
+    - examples/online_serving/qwen3_omni/
+    - examples/online_serving/openai_chat_completion_client_for_multimodal_generation.py
+    - tests/examples/online_serving/test_qwen3_omni.py
+
+  omni_qwen2_5_omni_doc:
+    - *qwen2_5_omni
+    - examples/online_serving/qwen2_5_omni/
+    - examples/online_serving/openai_chat_completion_client_for_multimodal_generation.py
+    - tests/examples/online_serving/test_qwen2_5_omni.py
+
+  omni_minicpmo_4_5_function:
+    - *minicpmo_4_5
+    - tests/e2e/offline_inference/test_minicpmo_4_5.py
+    - tests/e2e/online_serving/test_minicpmo_4_5.py
+    - tests/e2e/online_serving/test_minicpmo_4_5_expansion.py
+
+  omni_minicpmo_4_5_accuracy:
+    - *minicpmo_4_5
+    - tests/e2e/accuracy/minicpmo_4_5/
+
+  omni_minicpmo_4_5_perf:
+    - *minicpmo_4_5
+    - tests/dfx/perf/scripts/run_benchmark.py
+    - tests/dfx/perf/tests/test_minicpmo_4_5_ready.json
+    - tests/dfx/perf/tests/test_minicpmo_4_5.json
+
+  omni_minicpmo_4_5_doc:
+    - *minicpmo_4_5
+    - examples/online_serving/minicpmo/realtime_web/
+    - examples/online_serving/minicpmo/realtime_duplex_demo.py
+    - tests/examples/test_minicpmo_realtime_web_server.py
+    - tests/examples/test_minicpmo_realtime_web_static.py
+    - tests/examples/test_minicpmo_realtime_duplex_simple_demo.py
+
+  omni_minicpmo_4_5_duplex_function:
+    - *minicpmo_4_5
+    - *minicpmo_4_5_duplex
+    - tests/e2e/online_serving/test_minicpmo_4_5_duplex.py
+    - tests/e2e/online_serving/test_minicpmo_4_5_duplex_expansion.py
+    - tests/e2e/online_serving/test_duplex_client_live.py
+    - tests/e2e/online_serving/test_minicpmo_realtime_duplex_drivers.py
+    - tests/e2e/online_serving/helpers/minicpmo_realtime_duplex_scenarios.py
+    - tests/e2e/online_serving/run_minicpmo_realtime_duplex_multi_session.py
+    - tests/e2e/online_serving/run_minicpmo_realtime_duplex_server_vad.py
+    - tests/e2e/online_serving/run_minicpmo_realtime_duplex_soft_interrupt.py
+    - tests/e2e/online_serving/helpers/minicpmo_4_5_duplex.py
+    - tests/assets/minicpmo_4_5/
+    - tests/clients/
+    - tests/engine/duplex/
+    - tests/engine/test_duplex_omni_engine.py
+    - tests/engine/test_duplex_orchestrator.py
+    - tests/entrypoints/duplex/
+    - tests/entrypoints/test_duplex_omni.py
+    - tests/entrypoints/openai_api/test_duplex_api_server.py
+    - tests/model_executor/models/minicpmo_4_5/duplex/
+
+  omni_minicpmo_4_5_duplex_perf:
+    - *minicpmo_4_5
+    - *minicpmo_4_5_duplex
+    - tests/dfx/perf/scripts/run_benchmark.py
+    - tests/dfx/perf/tests/test_minicpmo_4_5_duplex_seed_tts.json
+
+  tts_qwen3_tts_function:
+    - *qwen3_tts
+    - tests/e2e/features/prefill_phase/
+    - tests/e2e/offline_inference/test_qwen3_tts_base.py
+    - tests/e2e/offline_inference/test_qwen3_tts_customvoice.py
+    - tests/e2e/online_serving/test_qwen3_tts_base.py
+    - tests/e2e/online_serving/test_qwen3_tts_base_expansion.py
+    - tests/e2e/online_serving/test_qwen3_tts_customvoice.py
+    - tests/e2e/online_serving/test_qwen3_tts_customvoice_expansion.py
+
+  tts_qwen3_tts_perf:
+    - *qwen3_tts
+    - tests/dfx/perf/scripts/run_benchmark.py
+    - tests/dfx/perf/tests/test_tts.json
+
+  tts_qwen3_tts_cov:
+    - *qwen3_tts
+    - .buildkite/common/scripts/run_cov_split.sh
+    - pyproject.toml
+    - tests/e2e/offline_inference/test_qwen3_tts_base.py
+    - tests/e2e/online_serving/test_qwen3_tts_base.py
+
+  # --- diffusion jobs ---
+  diffusion_wan22_function:
+    - *wan22
+    - tests/e2e/offline_inference/test_wan22_t2v.py
+    - tests/e2e/offline_inference/test_wan22_autoround_w4a16_expansion.py
+    - tests/e2e/offline_inference/test_device_postprocess_equivalence.py
+    - tests/e2e/online_serving/test_wan22_t2v.py
+    - tests/e2e/online_serving/test_wan22_expansion.py
+
+  diffusion_wan22_accuracy:
+    - *wan22
+    - tests/e2e/accuracy/wan22_i2v/
+    - tests/e2e/accuracy/test_diffusers_backend_similarity.py
+
+  diffusion_wan22_perf:
+    - *wan22
+    - tests/dfx/perf/scripts/run_diffusion_benchmark.py
+    - tests/dfx/perf/tests/test_wan22_i2v_vllm_omni.json
+
+  diffusion_wan22_reliability:
+    - *wan22
+    - tests/dfx/reliability/test_reliability_wan22.py
+
+  diffusion_image_to_video_doc:
+    - *wan22
+    - *hunyuan_video
+    - *ltx2
+    - *sana_video
+    - *lingbot_video
+    - examples/offline_inference/image_to_video/
+    - tests/examples/offline_inference/test_image_to_video.py
+
+  diffusion_cosmos3_function:
+    - *cosmos3
+    - tests/e2e/online_serving/test_cosmos3.py
+
+  diffusion_cosmos3_accuracy:
+    - *cosmos3
+    - tests/e2e/accuracy/test_cosmos3_similarity.py
+
+  diffusion_cosmos3_perf:
+    - *cosmos3
+    - tests/dfx/perf/scripts/run_diffusion_benchmark.py
+    - tests/dfx/perf/tests/test_cosmos3_vllm_omni.json
+
+  diffusion_qwen_image_function:
+    - *qwen_image
+    - tests/e2e/online_serving/test_qwen_image.py
+    - tests/e2e/online_serving/test_qwen_image_expansion.py
+    - tests/e2e/offline_inference/test_qwen_image_autoround_w4a16_expansion.py
+
+  diffusion_qwen_image_accuracy:
+    - *qwen_image
+    - tests/e2e/accuracy/test_qwen_image.py
+    - tests/e2e/accuracy/test_diffusers_backend_similarity.py
+
+  diffusion_qwen_image_perf:
+    - *qwen_image
+    - tests/dfx/perf/scripts/run_diffusion_benchmark.py
+    - tests/dfx/perf/tests/test_qwen_image_vllm_omni.json
+
+  diffusion_text_to_image_doc:
+    - *qwen_image
+    - *z_image
+    - examples/offline_inference/text_to_image/
+    - examples/online_serving/text_to_image/
+    - tests/examples/offline_inference/test_text_to_image.py
+    - tests/examples/online_serving/test_text_to_image.py
+    - vllm_omni/diffusion/models/nextstep_1_1/
+    - vllm_omni/diffusion/models/longcat_image/
+    - vllm_omni/diffusion/models/ovis_image/
+    - vllm_omni/diffusion/models/omnigen2/
+    - vllm_omni/diffusion/models/sd3/
+    - vllm_omni/diffusion/models/flux/
+    - vllm_omni/diffusion/models/flux2/
+    - vllm_omni/diffusion/models/flux2_klein/
+    - vllm_omni/diffusion/models/hidream_image/
+    - vllm_omni/diffusion/models/krea2/
+    - vllm_omni/diffusion/models/lingbot_video/
+
+  diffusion_hunyuan_image3_function:
+    - *hunyuan_image3
+    - tests/e2e/offline_inference/test_hunyuanimage3_i2t_expansion.py
+
+  diffusion_hunyuan_image3_accuracy:
+    - *hunyuan_image3
+    - tests/e2e/accuracy/test_hunyuan_image3_pixel_accuracy.py
+
+  diffusion_hunyuan_image3_perf:
+    - *hunyuan_image3
+    - tests/dfx/perf/scripts/run_diffusion_benchmark.py
+    - tests/dfx/perf/tests/test_hunyuan_image_tp2_cfgp2.json
+    - tests/dfx/perf/tests/test_hunyuan_image_tp2_sp2.json
+    - tests/dfx/perf/tests/test_hunyuan_image_tp4.json
+
+  diffusion_hunyuan_image3_reliability:
+    - *hunyuan_image3
+    - tests/dfx/reliability/test_reliability_hunyuan_image.py
+
+  diffusion_minimax_h3_function:
+    - *minimax_h3
+    - tests/e2e/online_serving/test_minimax_h3_dlo_dp2_t2va.py
+    - tests/e2e/online_serving/minimax_h3/
+    - tests/diffusion/models/minimax_h3/
+    - vllm_omni/diffusion/lora/
+    - vllm_omni/diffusion/offloader/
+    - vllm_omni/diffusion/distributed/
+    - vllm_omni/entrypoints/openai/
+
+  diffusion_minimax_h3_accuracy:
+    - *minimax_h3
+    - tests/e2e/accuracy/minimax_h3/
+
+  diffusion_minimax_h3_perf:
+    - *minimax_h3
+    - tests/dfx/perf/scripts/run_diffusion_benchmark.py
+    - tests/dfx/perf/tests/test_minimax_h3_vllm_omni.json
+
+  diffusion_hunyuan_video_function:
+    - *hunyuan_video
+    - tests/e2e/online_serving/test_hunyuan_video_15_expansion.py
+
+  diffusion_hunyuan_video_accuracy:
+    - *hunyuan_video
+    - tests/e2e/accuracy/hunyuanvideo15_t2v/
+    - tests/e2e/accuracy/hunyuanvideo15_i2v/
+
+  diffusion_hunyuan_video_perf:
+    - *hunyuan_video
+    - tests/dfx/perf/scripts/run_diffusion_benchmark.py
+    - tests/dfx/perf/tests/test_hunyuanvideo15_t2v_vllm_omni.json
+    - tests/dfx/perf/tests/test_hunyuanvideo15_i2v_vllm_omni.json
+
+  diffusion_z_image_function:
+    - *z_image
+    - tests/e2e/online_serving/test_z_image.py
+    - tests/e2e/online_serving/test_zimage_expansion.py
+
+  diffusion_sana_video_function:
+    - *sana_video
+    - tests/e2e/online_serving/test_sana_video_expansion.py
+    - tests/e2e/accuracy/sana_video/
+
+  diffusion_stable_audio_function:
+    - *stable_audio
+    - tests/e2e/offline_inference/test_stable_audio_expansion.py
+    - tests/e2e/online_serving/test_stable_audio_online_expansion.py
+
+  diffusion_lingbot_function:
+    - *lingbot_video
+    - *lingbot_world
+    - tests/e2e/online_serving/test_lingbot_video.py
+    - tests/e2e/online_serving/test_lingbot_video_moe.py
+    - tests/e2e/offline_inference/test_lingbot_world_v2.py
+
+
+  # Invalid-param jobs: model business code only, plus the reliability scripts.
+  reliability_invalid_param_h100:
+    - *qwen3_omni
+    - *minicpmo_4_5
+    - *qwen_image
+    - *wan22
+    - tests/dfx/reliability/invalid_param_test/
+
+  reliability_invalid_param_l4:
+    - *qwen3_tts
+    - *stable_audio
+    - tests/dfx/reliability/invalid_param_test/

.buildkite/common/scripts/run_cov_split.sh

@@ -0,0 +1,98 @@
+#!/usr/bin/env bash
+# Run a model's E2E tests once per entry mode as separate coverage runs and upload
+# the reports, so per-mode attribution survives (see vllm-project/vllm-omni#5332).
+#
+# Usage: run_cov_split.sh --model-id <id> [--offline <paths>] [--online <paths>] \
+#                         -- <pytest args...>
+#
+#   --model-id  Names the artifacts: coverage-<id>-{offline,online}-<step id>.xml.gz.
+#               Use the model's directory name under vllm_omni/*/models/.
+#   --offline   Pytest path(s) for the offline half. Repeatable.
+#   --online    Pytest path(s) for the online half. Repeatable.
+#   --          Everything after this is forwarded verbatim to every pytest run,
+#               quoting preserved: the job's own -m expression, --run-level, and
+#               anything else it needs (e.g. --test-config-file).
+#
+# At least one of --offline/--online is required, so a job with only one mode runs
+# just that half. Total runtime is bounded by the step's timeout_in_minutes.
+#
+# Coverage + artifact upload run only on scheduled weekly E2E sweeps
+# (WEEKLY=1 and BUILDKITE_BRANCH=main — same gate as bootstrap --e2e).
+# All other invocations (PR merge-test, regular main merge, local) run the
+# same combined pytest as the job used before coverage was added.
+#
+# Exits non-zero if any half or the upload failed.
+set -uo pipefail
+
+MODEL_ID=""
+OFFLINE=()
+ONLINE=()
+PYTEST_ARGS=()
+
+while [[ $# -gt 0 ]]; do
+    case "$1" in
+        --model-id) MODEL_ID="$2"; shift 2 ;;
+        --offline)  OFFLINE+=("$2"); shift 2 ;;
+        --online)   ONLINE+=("$2"); shift 2 ;;
+        --)         shift; PYTEST_ARGS=("$@"); break ;;
+        *) echo "run_cov_split.sh: unknown argument: $1" >&2; exit 2 ;;
+    esac
+done
+
+if [[ -z "${MODEL_ID}" ]]; then
+    echo "run_cov_split.sh: missing --model-id" >&2
+    exit 2
+fi
+
+if ((${#OFFLINE[@]} == 0 && ${#ONLINE[@]} == 0)); then
+    echo "run_cov_split.sh: need at least one of --offline / --online" >&2
+    exit 2
+fi
+
+_should_run_coverage() {
+    [[ "${WEEKLY:-}" == "1" && "${BUILDKITE_BRANCH:-}" == "main" ]]
+}
+
+if ! _should_run_coverage; then
+    PATHS=("${OFFLINE[@]}" "${ONLINE[@]}")
+    echo "--- pytest (no coverage): ${MODEL_ID}"
+    pytest -s -v "${PATHS[@]}" "${PYTEST_ARGS[@]}"
+    exit $?
+fi
+
+# Two steps can test the same model, and artifacts are scoped to the build (a
+# nightly one carries the ready, merge and nightly tiers at once), so unsuffixed
+# reports would share a path -- a glob download then races to rename onto one file.
+# Not BUILDKITE_JOB_ID: the step id survives retries, so attempts collapse under
+# Buildkite's latest-attempt lookup instead of accumulating. A parallel step would
+# need BUILDKITE_PARALLEL_JOB as well; no pilot is one.
+REPORT_SUFFIX="${BUILDKITE_STEP_ID:-local}"
+OFFLINE_REPORT="coverage-${MODEL_ID}-offline-${REPORT_SUFFIX}.xml"
+ONLINE_REPORT="coverage-${MODEL_ID}-online-${REPORT_SUFFIX}.xml"
+
+run_half() {
+    local mode="$1"
+    shift
+    local report="coverage-${MODEL_ID}-${mode}-${REPORT_SUFFIX}.xml"
+    echo "--- coverage: ${MODEL_ID} ${mode}"
+    pytest -s -v "$@" "${PYTEST_ARGS[@]}" --cov=vllm_omni --cov-report="xml:${report}"
+    local status=$?
+    # Cobertura lists every statement in vllm_omni whether or not it ran, so the
+    # report is ~7MB no matter how little the job covers. The repetition is what
+    # dominates it, and gzip takes that to ~400KB.
+    gzip -9 -f "${report}" || status=1
+    return "${status}"
+}
+
+# Clear both modes, not just the ones being run: the upload glob matches both, so
+# a report left by an earlier run would otherwise be published as if the mode it
+# belongs to had run this time.
+rm -f "${OFFLINE_REPORT}" "${ONLINE_REPORT}" "${OFFLINE_REPORT}.gz" "${ONLINE_REPORT}.gz"
+
+EXIT=0
+((${#OFFLINE[@]})) && { run_half offline "${OFFLINE[@]}" || EXIT=1; }
+((${#ONLINE[@]})) && { run_half online "${ONLINE[@]}" || EXIT=1; }
+
+buildkite-agent artifact upload "coverage-${MODEL_ID}-*-${REPORT_SUFFIX}.xml.gz" || EXIT=1
+
+exit "${EXIT}"

.buildkite/common/scripts/skip_ci.py

@@ -1,5 +1,6 @@
 #!/usr/bin/env python3
 # SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
 """Shared Buildkite skip-ci and CI-YAML level targeting logic.
 
 ``resolve_ci_decision(changed_files, ...)`` buckets explicit file lists (tests, dry-run).
@@ -10,12 +11,6 @@ CLI (for bootstrap scripts; exit code is the signal):
   python3 skip_ci.py gate <platform> <level>
       Exit 0 when bootstrap should stop (skip-all, or that L2/L3 target is off).
       Prints ``skip-all`` or ``skip-l23`` on stdout; logs the decision once.
-  python3 skip_ci.py check-skip-all
-      Exit 0 when only docs/skip-mark changes → bootstrap skips entire CI upload.
-  python3 skip_ci.py check-skip-l2-l3 <platform> <level>
-      Exit 0 when that L2/L3 pipeline should be skipped (CI config YAML-only diffs).
-  python3 skip_ci.py print-annotate
-      Print decision message for Buildkite annotation.
 """
 
 from __future__ import annotations
@@ -49,6 +44,7 @@ L3_YAML_FILES: dict[str, str] = {
 L45_YAML_FILES: dict[str, str] = {
     ".buildkite/cuda/test-nightly.yml": "cuda",
     ".buildkite/cuda/test-weekly.yml": "cuda",
+    ".buildkite/amd/test-amd-nightly.yml": "amd",
     ".buildkite/npu/test-npu-nightly.yml": "npu",
 }
 
@@ -258,11 +254,10 @@ class CiDecision:
 
 @dataclass(frozen=True)
 class CiContext:
-    """Skip-ci decision plus git diff inputs used to compute it."""
+    """Skip-ci decision plus changed files used for pipeline filtering."""
 
     decision: CiDecision
     changed_files: list[str] | None
-    diff_range: str | None
 
 
 def _log(message: str) -> None:
@@ -332,7 +327,7 @@ def resolve_ci_context_from_git() -> CiContext:
     diff_range = _resolve_diff_range()
     changed_files = _changed_files_for_diff_range(diff_range)
     decision = resolve_ci_decision(changed_files, diff_range=diff_range)
-    return CiContext(decision=decision, changed_files=changed_files, diff_range=diff_range)
+    return CiContext(decision=decision, changed_files=changed_files)
 
 
 # --- Doc / skip-mark helpers ---
@@ -571,23 +566,6 @@ def main() -> int:
     gate_parser.add_argument("platform", help="cuda, amd, intel, or npu")
     gate_parser.add_argument("level", choices=("l2", "l3"), help="test level")
 
-    subparsers.add_parser(
-        "check-skip-all",
-        help="exit 0 when docs/skip-mark-only diff should skip entire CI upload",
-    )
-
-    skip_parser = subparsers.add_parser(
-        "check-skip-l2-l3",
-        help="exit 0 when PLATFORM l2/l3 pipeline should be skipped (CI config YAML-only diffs)",
-    )
-    skip_parser.add_argument("platform", help="cuda, amd, intel, or npu")
-    skip_parser.add_argument("level", choices=("l2", "l3"), help="test level")
-
-    subparsers.add_parser(
-        "print-annotate",
-        help="print human-readable skip reason for Buildkite annotation",
-    )
-
     args = parser.parse_args()
     decision = resolve_ci_context_from_git().decision
 
@@ -600,16 +578,6 @@ def main() -> int:
             return 0
         return 1
 
-    if args.command == "check-skip-all":
-        return 0 if decision.skip_all else 1
-
-    if args.command == "check-skip-l2-l3":
-        return 0 if not decision.is_run(args.platform, args.level) else 1
-
-    if args.command == "print-annotate":
-        print(decision.message or "CI will run normally")
-        return 0
-
     return 1
 
 

.buildkite/common/scripts/upload_pipeline.py

@@ -1,5 +1,6 @@
 #!/usr/bin/env python3
 # SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
 """Render and optionally upload Buildkite pipeline YAML with diff-aware logic.
 
 Bootstrap mode (``bootstrap-upload-steps.yml``):
@@ -9,10 +10,30 @@ Bootstrap mode (``bootstrap-upload-steps.yml``):
   - Detect docs-only, pytest skip-mark-only, or combined skip-ci from git diff.
   - When only CI level YAML changes, enable **L2/L3** upload steps for affected levels only.
 
-Test pipeline mode (e.g. test-merge.yml):
-  - Drop steps whose ``source_file_dependencies`` do not match changed files.
-  - Expand uploader-only ``mirror_hardwares: l4_1`` into ``agents`` (+ optional ``image``
+Test pipeline mode (e.g. test-merge.yml, test-nightly.yml, test-weekly.yml):
+  - Drop steps whose ``source_file_dependencies`` do not match changed files
+    (string/list keys from ci_source_file_dependencies.yml, or inline path
+    prefixes). Filtering applies on **PR label** uploads and on post-merge
+    ``main`` L3 uploads. Scheduled ``main`` uploads that pass ``--all``
+    (NIGHTLY/WEEKLY) or ``--e2e`` keep every selected job and still strip
+    the field. If no job-key prefix matches, a change to the pipeline YAML
+    being uploaded or to a path under the ``source_filter_fallback`` registry
+    key keeps every job so command, env, and hardware edits can be validated
+    before merge. When any job-key prefix already matches, normal filtering
+    wins.
+  - Expand uploader-only ``mirror_hardwares`` into ``agents`` (+ optional ``image``
     for NPU) + ``plugins`` (see ci_mirror_hardwares.yml).
+  - Omit ``mirror_hardwares`` to compose ``{chip}_{n}`` from pytest ``-m`` SKU
+    markers plus ``cards_n`` (max of several positives; ``not cards_1`` with no
+    positive ``cards_*`` uses that chip's highest existing preset). An explicit
+    preset string (``h100_2``) is used as-is (marks are ignored). Names that
+    are not keys in ``ci_mirror_hardwares.yml`` fail the upload.
+
+    Inferred chip: unset ``MIRROR_HW`` matches ``H100`` or ``L4`` in ``-m``
+    (both → H100); no match skips the step. ``MIRROR_HW=b200`` matches
+    ``B200`` in ``-m`` (otherwise skipped). ``-m`` is not rewritten.
+    ``MIRROR_HW`` must be empty or ``b200``; unknown values fail the upload.
+    A CUDA preset string such as ``h100_4`` is omitted when ``MIRROR_HW=b200``.
 
 Usage:
   python3 upload_pipeline.py [--upload] [--all | --e2e] <pipeline.yml>
@@ -24,11 +45,13 @@ from __future__ import annotations
 
 import argparse
 import copy
+import os
+import re
 import subprocess
 import sys
 from functools import lru_cache
 from pathlib import Path
-from typing import Any
+from typing import Any, Literal
 
 try:
     import yaml
@@ -44,11 +67,27 @@ from skip_ci import (
     resolve_ci_context_from_git,
 )
 
+if str(ROOT) not in sys.path:
+    sys.path.insert(0, str(ROOT))
+
+from tests.helpers.mark import (  # noqa: E402
+    get_skus_for_platform,
+    get_supported_card_counts,
+)
+
 # --- Constants ---
 
 LOG = "upload_pipeline"
 BOOTSTRAP_STEPS_FILENAME = "bootstrap-upload-steps.yml"
-BOOTSTRAP_IMAGE_BUILD_KEYS = frozenset({"image-build", "image-build-a2", "image-build-a3"})
+BOOTSTRAP_IMAGE_BUILD_KEYS = frozenset(
+    {
+        "image-build",
+        "image-build-a2",
+        "image-build-a3",
+        "image-build-a5",
+        "image-build-310p",
+    }
+)
 BOOTSTRAP_UPLOAD_IF_KEYS = {
     "upload-ready-pipeline": "ready",
     "upload-merge-pipeline": "merge",
@@ -57,20 +96,33 @@ BOOTSTRAP_UPLOAD_IF_KEYS = {
 }
 E2E_GROUP_MARKER = "E2E Test"
 CI_MIRROR_HARDWARES_PATH = ROOT / ".buildkite/common/ci_mirror_hardwares.yml"
+CI_SOURCE_FILE_DEPENDENCIES_PATH = ROOT / ".buildkite/common/ci_source_file_dependencies.yml"
+# Registry key whose prefixes bypass source filtering. Paths live in
+# ci_source_file_dependencies.yml; do not attach this key to a job.
+SOURCE_FILTER_FALLBACK_KEY = "source_filter_fallback"
+CUDA_HF_TOKEN_ENV = "VLLM_CI_HF_TOKEN"
+CUDA_HF_TOKEN_EXPORT = f'if [ -n "$${{{CUDA_HF_TOKEN_ENV}:-}}" ]; then export HF_TOKEN="$${{{CUDA_HF_TOKEN_ENV}}}"; fi'
 
-CUDA_NIGHTLY_ONLY = (
-    '(build.pull_request.labels includes "nightly-test") || (build.branch == "main" && build.env("NIGHTLY") == "1")'
+# Bootstrap Buildkite ``if`` expressions.
+# ``*_MAIN_IF``: main + env schedule. ``*_LABEL_IF``: PR label (and/or composed with MAIN).
+# ``*_UPLOAD_IF``: full gate for uploading that child pipeline.
+NIGHTLY_MAIN_IF = 'build.branch == "main" && build.env("NIGHTLY") == "1"'
+NIGHTLY_LABEL_IF = (
+    f'({NIGHTLY_MAIN_IF}) || (build.branch != "main" && build.pull_request.labels includes "nightly-test")'
 )
-NPU_NIGHTLY_ONLY = (
-    '(build.branch == "main" && build.env("NIGHTLY") == "1") || '
-    '(build.branch != "main" && ('
-    'build.pull_request.labels includes "nightly-test" || '
-    'build.pull_request.labels includes "omni-test" || '
-    'build.pull_request.labels includes "tts-test" || '
-    'build.pull_request.labels includes "diffusion-x2iat-test" || '
-    'build.pull_request.labels includes "diffusion-x2v-test"'
-    "))"
+WEEKLY_E2E_IF = 'build.branch == "main" && build.env("WEEKLY") == "1"'
+WEEKLY_MAIN_IF = 'build.branch == "main" && (build.env("WEEKLY") == "1" || build.env("NON_CRITICAL") == "1")'
+WEEKLY_LABEL_IF = f'({WEEKLY_MAIN_IF}) || (build.branch != "main" && build.pull_request.labels includes "weekly-test")'
+READY_LABEL_IF = 'build.branch != "main" && build.pull_request.labels includes "ready"'
+MERGE_LABEL_IF = 'build.branch != "main" && build.pull_request.labels includes "merge-test"'
+MERGE_MAIN_IF = (
+    'build.branch == "main" && build.env("NIGHTLY") != "1" && '
+    'build.env("WEEKLY") != "1" && build.env("NON_CRITICAL") != "1"'
 )
+READY_UPLOAD_IF = f"({WEEKLY_E2E_IF}) || ({READY_LABEL_IF})"
+MERGE_UPLOAD_IF = f"({WEEKLY_E2E_IF}) || (({MERGE_MAIN_IF}) || ({MERGE_LABEL_IF}))"
+BOOTSTRAP_DISABLED_IF = "false"
+BOOTSTRAP_ENABLED_IF = "true"
 
 
 # --- Logging ---
@@ -102,74 +154,47 @@ def _format_bootstrap_if(expr: str) -> str:
 
 
 def _compute_bootstrap_if_exprs(*, decision, platform: str) -> dict[str, str]:
-    disabled = "false"
-    nightly_main = 'build.branch == "main" && build.env("NIGHTLY") == "1"'
-    nightly_only = NPU_NIGHTLY_ONLY if platform == "npu" else CUDA_NIGHTLY_ONLY
-
     if platform == "npu":
-        ready_pr = (
-            'build.branch != "main" && ('
-            'build.pull_request.labels includes "npu-test" || '
-            'build.pull_request.labels includes "ready"'
-            ")"
-        )
-        nightly_label_if = nightly_only
-        weekly_label_if = disabled
-        merge_base = disabled
+        ready_upload = READY_LABEL_IF
+        merge_upload = BOOTSTRAP_DISABLED_IF
+        weekly_label_if = BOOTSTRAP_DISABLED_IF
     else:
-        ready_pr = 'build.branch != "main" && build.pull_request.labels includes "ready"'
-        merge_main = 'build.branch == "main" && build.env("NIGHTLY") != "1" && build.env("WEEKLY") != "1"'
-        merge_pr = 'build.branch != "main" && build.pull_request.labels includes "merge-test"'
-        nightly_label_if = (
-            '(build.branch == "main" && build.env("NIGHTLY") == "1") || '
-            '(build.branch != "main" && ('
-            'build.pull_request.labels includes "nightly-test" || '
-            'build.pull_request.labels includes "omni-test" || '
-            'build.pull_request.labels includes "tts-test" || '
-            'build.pull_request.labels includes "diffusion-x2iat-test" || '
-            'build.pull_request.labels includes "diffusion-x2v-test"'
-            "))"
-        )
-        weekly_label_if = (
-            '(build.branch == "main" && build.env("WEEKLY") == "1") || '
-            '(build.branch != "main" && build.pull_request.labels includes "weekly-test")'
-        )
-        merge_base = f"({nightly_main}) || (({merge_main}) || ({merge_pr}))"
-
-    ready_base = f"({nightly_main}) || ({ready_pr})"
-    weekly_main = 'build.branch == "main" && build.env("WEEKLY") == "1"'
+        ready_upload = READY_UPLOAD_IF
+        merge_upload = MERGE_UPLOAD_IF
+        weekly_label_if = WEEKLY_LABEL_IF
 
     if decision.skip_all:
         # Docs / skip-mark only: no PR-label escape hatch. Main scheduled
-        # NIGHTLY=1 still runs L4 + L2/L3 (--e2e); WEEKLY=1 still runs L5.
-        image_expr = f"({nightly_main}) || ({weekly_main})" if platform == "cuda" else nightly_main
-        ready_expr = nightly_main
-        merge_expr = nightly_main if platform == "cuda" else disabled
-        nightly_expr = nightly_main
-        weekly_expr = weekly_main if platform == "cuda" else disabled
+        # NIGHTLY=1 still runs L4; WEEKLY=1 / NON_CRITICAL=1 still run L5.
+        # main+WEEKLY=1 also uploads L2/L3 (those steps then pass --e2e).
+        image_expr = f"({NIGHTLY_MAIN_IF}) || ({WEEKLY_MAIN_IF})" if platform == "cuda" else NIGHTLY_MAIN_IF
+        ready_expr = WEEKLY_E2E_IF if platform == "cuda" else BOOTSTRAP_DISABLED_IF
+        merge_expr = WEEKLY_E2E_IF if platform == "cuda" else BOOTSTRAP_DISABLED_IF
+        nightly_expr = NIGHTLY_MAIN_IF
+        weekly_expr = WEEKLY_MAIN_IF if platform == "cuda" else BOOTSTRAP_DISABLED_IF
     elif decision.skip_l2_l3:
         l2_enabled = decision.is_run("npu", "l2") if platform == "npu" else decision.is_run("cuda", "l2")
         l3_enabled = platform == "cuda" and decision.is_run("cuda", "l3")
 
-        ready_expr = ready_base if l2_enabled else disabled
-        merge_expr = merge_base if l3_enabled else disabled
-        nightly_expr = nightly_label_if
-        weekly_expr = weekly_label_if if platform == "cuda" else disabled
+        ready_expr = ready_upload if l2_enabled else BOOTSTRAP_DISABLED_IF
+        merge_expr = merge_upload if l3_enabled else BOOTSTRAP_DISABLED_IF
+        nightly_expr = NIGHTLY_LABEL_IF
+        weekly_expr = weekly_label_if if platform == "cuda" else BOOTSTRAP_DISABLED_IF
 
-        image_parts = [f"({nightly_label_if})"]
+        image_parts = [f"({NIGHTLY_LABEL_IF})"]
         if platform == "cuda":
             image_parts.append(f"({weekly_label_if})")
         if l2_enabled:
-            image_parts.insert(0, f"({ready_base})")
+            image_parts.insert(0, f"({ready_upload})")
         if l3_enabled:
-            image_parts.insert(1 if l2_enabled else 0, f"({merge_base})")
+            image_parts.insert(1 if l2_enabled else 0, f"({merge_upload})")
         image_expr = " || ".join(image_parts)
     else:
-        image_expr = "true"
-        ready_expr = ready_base
-        merge_expr = merge_base if platform == "cuda" else disabled
-        nightly_expr = nightly_label_if
-        weekly_expr = weekly_label_if if platform == "cuda" else disabled
+        image_expr = BOOTSTRAP_ENABLED_IF
+        ready_expr = ready_upload
+        merge_expr = merge_upload if platform == "cuda" else BOOTSTRAP_DISABLED_IF
+        nightly_expr = NIGHTLY_LABEL_IF
+        weekly_expr = weekly_label_if if platform == "cuda" else BOOTSTRAP_DISABLED_IF
 
     return {
         "image": _format_bootstrap_if(image_expr),
@@ -229,7 +254,7 @@ def _render_bootstrap_pipeline(
     return yaml.safe_dump(doc, sort_keys=False)
 
 
-# --- Test pipeline (test-ready.yml, test-merge.yml) ---
+# --- Test pipeline (test-ready.yml, test-merge.yml, test-nightly.yml) ---
 
 
 @lru_cache(maxsize=1)
@@ -245,32 +270,365 @@ def _load_mirror_hardwares() -> dict[str, dict[str, Any]]:
     return presets
 
 
-def _expand_mirror_hardwares(step: dict[str, Any]) -> dict[str, Any]:
-    """Replace uploader-only ``mirror_hardwares`` with preset fields from ci_mirror_hardwares.yml."""
-    hardware = step.get("mirror_hardwares")
-    if hardware is None:
-        return step
-
-    if not isinstance(hardware, str) or not hardware.strip():
+@lru_cache(maxsize=1)
+def _load_source_file_dependencies() -> dict[str, list[str]]:
+    def flatten(value: Any, *, key: str) -> list[str]:
+        """Flatten YAML-anchor nested lists into a de-duplicated prefix list."""
+        if isinstance(value, str):
+            if not value.strip():
+                raise ValueError(f"empty path in source_file_dependencies[{key!r}]")
+            return [value]
+        if isinstance(value, list):
+            flattened: list[str] = []
+            seen: set[str] = set()
+            for item in value:
+                for path in flatten(item, key=key):
+                    if path not in seen:
+                        seen.add(path)
+                        flattened.append(path)
+            return flattened
         raise ValueError(
-            f"mirror_hardwares must be a non-empty string in step {_get_step_label(step)!r}",
+            f"source_file_dependencies[{key!r}] must be a list of path prefixes, got {type(value).__name__}",
         )
 
-    preset = _load_mirror_hardwares().get(hardware)
+    if not CI_SOURCE_FILE_DEPENDENCIES_PATH.is_file():
+        raise FileNotFoundError(
+            f"missing CI source_file_dependencies registry: {CI_SOURCE_FILE_DEPENDENCIES_PATH}",
+        )
+    doc = yaml.safe_load(CI_SOURCE_FILE_DEPENDENCIES_PATH.read_text(encoding="utf-8"))
+    if not isinstance(doc, dict):
+        raise ValueError(f"invalid CI source_file_dependencies registry: {CI_SOURCE_FILE_DEPENDENCIES_PATH}")
+    presets = doc.get("source_file_dependencies")
+    if not isinstance(presets, dict):
+        raise ValueError(f"source_file_dependencies must be a mapping in {CI_SOURCE_FILE_DEPENDENCIES_PATH}")
+    loaded: dict[str, list[str]] = {}
+    for key, value in presets.items():
+        if not isinstance(key, str) or not key.strip():
+            raise ValueError(
+                f"source_file_dependencies keys must be non-empty strings in {CI_SOURCE_FILE_DEPENDENCIES_PATH}",
+            )
+        loaded[key] = flatten(value, key=key)
+    return loaded
+
+
+_REGISTRY_KEY_RE = re.compile(r"^[a-z][a-z0-9_]*$")
+
+
+def _parse_command_text(step: dict[str, Any]) -> str:
+    """Parse step ``commands`` / ``command`` into a single string."""
+    raw = step.get("commands", step.get("command"))
+    if raw is None:
+        return ""
+    if isinstance(raw, str):
+        return raw
+    if isinstance(raw, list):
+        return "\n".join(str(item) for item in raw if item is not None)
+    return str(raw)
+
+
+def _resolve_source_file_dependencies(step: dict[str, Any]) -> list[str] | None:
+    """Expand a registry key (or list of keys / inline path prefixes)."""
+    deps = step.get("source_file_dependencies")
+    if deps is None:
+        return None
+
+    def dedupe(*groups: list[str]) -> list[str]:
+        merged: list[str] = []
+        seen: set[str] = set()
+        for group in groups:
+            for path in group:
+                if path not in seen:
+                    seen.add(path)
+                    merged.append(path)
+        return merged
+
+    def lookup(key: str) -> list[str]:
+        if key == SOURCE_FILTER_FALLBACK_KEY:
+            raise ValueError(
+                f"source_file_dependencies {key!r} bypasses filtering and cannot be attached to a step "
+                f"({_get_step_label(step)!r})",
+            )
+        registry = _load_source_file_dependencies()
+        paths = registry.get(key)
+        if paths is None:
+            known = ", ".join(sorted(registry))
+            raise ValueError(
+                f"unknown source_file_dependencies {key!r} in step {_get_step_label(step)!r}; known: {known}",
+            )
+        return list(paths)
+
+    prefixes: list[str]
+    if isinstance(deps, str):
+        prefixes = lookup(deps) if _REGISTRY_KEY_RE.fullmatch(deps) else [deps]
+    elif isinstance(deps, list):
+        if not all(isinstance(item, str) for item in deps):
+            raise ValueError(
+                f"source_file_dependencies must be a string key or list of strings in step {_get_step_label(step)!r}",
+            )
+        keyish = [bool(_REGISTRY_KEY_RE.fullmatch(item)) for item in deps]
+        if deps and all(keyish):
+            prefixes = dedupe(*(lookup(key) for key in deps))
+        elif any(keyish):
+            raise ValueError(
+                f"source_file_dependencies in step {_get_step_label(step)!r} mixes registry keys and path prefixes",
+            )
+        else:
+            prefixes = list(deps)
+    else:
+        raise ValueError(
+            f"source_file_dependencies must be a string key or list in step {_get_step_label(step)!r}",
+        )
+    return prefixes
+
+
+@lru_cache(maxsize=1)
+def _cuda_mirror_chips() -> tuple[str, ...]:
+    """Lowercase CUDA SKUs that have at least one ``{chip}_{n}`` preset.
+
+    Longer tokens first so ``h100`` is not matched as a prefix of ``h200``.
+    """
+    skus = get_skus_for_platform("cuda")
+    known = _load_mirror_hardwares()
+    chips = {
+        sku.lower() for sku in skus if any(name == sku.lower() or name.startswith(f"{sku.lower()}_") for name in known)
+    }
+    return tuple(sorted(chips, key=lambda chip: (-len(chip), chip)))
+
+
+def _parse_cards_marks(expr: str) -> tuple[set[int], bool]:
+    """Parse registered ``cards_n`` / ``not cards_n`` from a pytest ``-m`` expr.
+
+    Returns positive card counts and whether any ``not cards_*`` is present.
+    """
+    counts = get_supported_card_counts()
+    alt = "|".join(str(n) for n in sorted(counts, key=lambda n: (-len(str(n)), -n))) if counts else "(?!)"
+    positives: set[int] = set()
+    has_not_cards = False
+    for match in re.finditer(rf"\b(?:(not)\s+)?cards_({alt})\b", expr):
+        if match.group(1):
+            has_not_cards = True
+        else:
+            positives.add(int(match.group(2)))
+    return positives, has_not_cards
+
+
+def _parse_hardware_marks(expr: str) -> set[str]:
+    """Parse lowercase positive CUDA SKUs that have a mirror preset.
+
+    ``not H100`` is ignored. SKUs with no ``{chip}_*`` preset (e.g. H200) are dropped.
+    """
+    skus = sorted(get_skus_for_platform("cuda"), key=lambda s: (-len(s), s))
+    if not skus:
+        return set()
+    alt = "|".join(re.escape(sku) for sku in skus)
+    mirror = set(_cuda_mirror_chips())
+    chips: set[str] = set()
+    for match in re.finditer(rf"\b(?:(not)\s+)?({alt})\b", expr):
+        if match.group(1):
+            continue
+        chip = match.group(2).lower()
+        if chip in mirror:
+            chips.add(chip)
+    return chips
+
+
+# CI policy: not pytest-mark registry.
+_SUPPORTED_MIRROR_HW_SELECTORS = frozenset({"b200"})
+# Unset MIRROR_HW: match H100 then L4 in -m (skip if neither).
+_DEFAULT_INFER_CHIPS = ("h100", "l4")
+_PYTEST_MARKER_ARG = re.compile(r"-m\s+(?:\"([^\"]*)\"|'([^']*)'|(\S+))")
+_CARDS_CHIP_MAX: Literal["max"] = "max"
+
+
+def _get_mirror_hw_selector() -> str:
+    """Return lowercase ``MIRROR_HW``, or empty to keep string presets / marker chips.
+
+    A set value must be ``b200`` (case-insensitive). Unknown values fail closed
+    so a typo cannot silently drop the CUDA pipeline.
+    """
+    selector = os.environ.get("MIRROR_HW", "").strip().lower()
+    if not selector:
+        return ""
+    if selector not in _SUPPORTED_MIRROR_HW_SELECTORS:
+        allowed = " or ".join(repr(s) for s in sorted(_SUPPORTED_MIRROR_HW_SELECTORS))
+        raise ValueError(f"unsupported MIRROR_HW={selector!r}; expected empty or {allowed}")
+    return selector
+
+
+def _parse_pytest_marks(step: dict[str, Any]) -> tuple[set[str], int | Literal["max"] | None]:
+    """Parse pytest ``-m`` from step commands: CUDA SKU chips and cards count.
+
+    *chips* are lowercase SKUs from positive markers (``not H100`` is ignored).
+    *cards* is the max positive ``cards_n``, ``"max"`` when only ``not cards_*``
+    is present, or ``None``.
+    """
+    text = _parse_command_text(step)
+
+    chips: set[str] = set()
+    positives: set[int] = set()
+    has_not_cards = False
+    for match in _PYTEST_MARKER_ARG.finditer(text):
+        expr = match.group(1) or match.group(2) or match.group(3) or ""
+        chips |= _parse_hardware_marks(expr)
+        card_counts, not_cards = _parse_cards_marks(expr)
+        positives.update(card_counts)
+        has_not_cards = has_not_cards or not_cards
+
+    if positives:
+        cards: int | Literal["max"] | None = max(positives)
+    elif has_not_cards:
+        cards = _CARDS_CHIP_MAX
+    else:
+        cards = None
+    return chips, cards
+
+
+def _compose_mirror_hardware_name(
+    chips: set[str],
+    cards: int | Literal["max"] | None,
+    *,
+    step_label: str,
+) -> str | None:
+    """Assemble ``{chip}_{n}`` from pytest ``-m`` chips/cards, or None to skip.
+
+    Unset ``MIRROR_HW``: ``H100`` if in ``-m``, else ``L4``, else skip.
+    Set ``MIRROR_HW``: that chip if in ``-m``, else skip.
+    ``cards`` is a registered ``cards_n``, ``"max"`` for that chip's highest
+    existing preset, or None (error after a chip matched — no ``cards_*``).
+    """
+    selector = _get_mirror_hw_selector()
+    if selector:
+        chip = selector if selector in chips else None
+        skip_reason = f"MIRROR_HW={selector!r} not in -m SKUs {sorted(chips)}"
+    else:
+        chip = next((c for c in _DEFAULT_INFER_CHIPS if c in chips), None)
+        skip_reason = f"no H100 or L4 in -m SKUs {sorted(chips)}"
+    if chip is None:
+        _log(f"skip {step_label}: {skip_reason}")
+        return None
+    if cards is None:
+        raise ValueError(
+            f"step {step_label!r} has a pytest -m SKU marker but no cards_* "
+            f"(or not cards_*); add cards_n or set mirror_hardwares to force a preset",
+        )
+
+    count_source = "not cards_*" if cards == _CARDS_CHIP_MAX else f"cards_{cards}"
+    count = None if cards == _CARDS_CHIP_MAX else cards
+    registered = get_supported_card_counts()
+    if count is not None and count not in registered:
+        supported = ", ".join(f"cards_{n}" for n in sorted(registered)) or "(none)"
+        raise ValueError(
+            f"{count_source} in step {step_label!r} is not a registered cards_* mark; supported: {supported}",
+        )
+    known = _load_mirror_hardwares()
+    if count is None:
+        available = [n for n in sorted(registered) if f"{chip}_{n}" in known]
+        if not available:
+            raise ValueError(
+                f"{count_source} in step {step_label!r} has no CUDA count presets for {chip}",
+            )
+        count = max(available)
+    chosen = f"{chip}_{count}"
+    if chosen not in known:
+        raise ValueError(
+            f"{count_source} in step {step_label!r} has no preset {chosen!r}",
+        )
+    _log(f"{step_label}: {count_source} → {chosen!r}")
+    return chosen
+
+
+def _resolve_mirror_hardware_name(hardware: Any, *, step_label: str) -> str | None:
+    """Map ``mirror_hardwares`` to a ``ci_mirror_hardwares.yml`` preset, or None to skip.
+
+    The value is a preset name (``h100_2``). Unknown names fail later when the
+    registry is loaded. CUDA names are omitted when ``MIRROR_HW`` names a
+    different chip; NPU strings ignore ``MIRROR_HW``.
+    """
+    if isinstance(hardware, bool):
+        raise ValueError(f"mirror_hardwares must not be a boolean in step {step_label!r}")
+    if isinstance(hardware, str):
+        name = hardware.strip()
+    elif isinstance(hardware, int):
+        name = str(hardware)
+    else:
+        raise ValueError(
+            f"mirror_hardwares must be a preset string in step {step_label!r}",
+        )
+    if not name:
+        raise ValueError(f"mirror_hardwares must be a non-empty string in step {step_label!r}")
+
+    selector = _get_mirror_hw_selector()
+    token = name.lower()
+    chip = next(
+        (c for c in _cuda_mirror_chips() if token == c or token.startswith(f"{c}_")),
+        None,
+    )
+    if selector and chip is not None and chip != selector:
+        _log(
+            f"skip {step_label}: preset {name!r} is {chip} hardware; MIRROR_HW={selector!r}",
+        )
+        return None
+    return name
+
+
+def _expand_mirror_hardwares(step: dict[str, Any]) -> dict[str, Any] | None:
+    """Replace ``mirror_hardwares`` (or infer it from ``-m``) with preset agents/plugins.
+
+    Explicit preset string: used as-is; pytest marks are ignored.
+    Omitted key: match H100/L4 (or ``MIRROR_HW``) in ``-m``, then compose
+    ``{chip}_{n}`` from ``cards_n``. No match skips the step. Steps with no
+    SKU/cards and no ``mirror_hardwares`` are left unchanged (CPU jobs).
+
+    Returns None when the current ``MIRROR_HW`` selector does not apply.
+    """
+    step_label = _get_step_label(step)
+    has_pool = step.get("agents") is not None or step.get("plugins") is not None or step.get("image") is not None
+
+    if "mirror_hardwares" not in step:
+        if has_pool:
+            return step
+        chips, cards = _parse_pytest_marks(step)
+        if not chips and cards is None:
+            return step
+        preset_name = _compose_mirror_hardware_name(chips, cards, step_label=step_label)
+    else:
+        if has_pool:
+            raise ValueError(
+                f"step {step_label!r} sets mirror_hardwares together with agents/plugins/image; "
+                f"use mirror_hardwares only",
+            )
+        preset_name = _resolve_mirror_hardware_name(
+            step.get("mirror_hardwares"),
+            step_label=step_label,
+        )
+
+    if preset_name is None:
+        return None
+
+    preset = _load_mirror_hardwares().get(preset_name)
     if preset is None:
         known = ", ".join(sorted(_load_mirror_hardwares()))
         raise ValueError(
-            f"unknown mirror_hardwares {hardware!r} in step {_get_step_label(step)!r}; known: {known}",
-        )
-
-    if step.get("agents") is not None or step.get("plugins") is not None or step.get("image") is not None:
-        raise ValueError(
-            f"step {_get_step_label(step)!r} sets mirror_hardwares together with agents/plugins/image; "
-            "use mirror_hardwares only",
+            f"unknown mirror_hardwares {preset_name!r} in step {step_label!r}; known: {known}",
         )
 
     expanded = copy.deepcopy(preset)
-    return {key: value for key, value in step.items() if key != "mirror_hardwares"} | expanded
+    merged = {key: value for key, value in step.items() if key != "mirror_hardwares"} | expanded
+    # A scheduled build's HF_TOKEN may override a pod env entry with the same
+    # name. The preset injects the Kubernetes secret under a collision-proof
+    # alias; restore the conventional name after Buildkite has merged env.
+    if any(preset_name == chip or preset_name.startswith(f"{chip}_") for chip in _cuda_mirror_chips()):
+        commands = merged.get("commands")
+        if isinstance(commands, list):
+            merged["commands"] = [CUDA_HF_TOKEN_EXPORT, *commands]
+        elif commands is None:
+            merged["commands"] = [CUDA_HF_TOKEN_EXPORT]
+        else:
+            merged["commands"] = [CUDA_HF_TOKEN_EXPORT, commands]
+    # Preset retry (K8S_RETRY on l4_*) must not clobber a step that opted out.
+    if "retry" in step:
+        merged["retry"] = step["retry"]
+    return merged
 
 
 def _match_source_file(changed_files: list[str], prefixes: list[str]) -> bool:
@@ -297,11 +655,7 @@ def _process_test_steps(
             processed.append(step)
             continue
 
-        deps = step.get("source_file_dependencies")
-        if deps is not None and not isinstance(deps, list):
-            raise ValueError(
-                f"source_file_dependencies must be a list in step {_get_step_label(step)!r}",
-            )
+        deps = _resolve_source_file_dependencies(step)
         if changed_files is not None and deps is not None and not _match_source_file(changed_files, deps):
             _log(f"skip {_get_step_label(step)!r} (no changes under {deps})")
             continue
@@ -309,7 +663,7 @@ def _process_test_steps(
         nested = step.get("steps")
         if nested is not None:
             kept_nested = _process_test_steps(nested, changed_files)
-            if changed_files is not None and not kept_nested:
+            if not kept_nested:
                 _log(f"omit empty group {_get_step_label(step)!r}")
                 continue
             new_step = {key: value for key, value in step.items() if key != "source_file_dependencies"}
@@ -317,14 +671,11 @@ def _process_test_steps(
             processed.append(new_step)
             continue
 
-        if deps is not None:
-            processed.append(
-                _expand_mirror_hardwares(
-                    {key: value for key, value in step.items() if key != "source_file_dependencies"},
-                ),
-            )
-        else:
-            processed.append(_expand_mirror_hardwares(step))
+        leaf = {key: value for key, value in step.items() if key != "source_file_dependencies"}
+        expanded = _expand_mirror_hardwares(leaf)
+        if expanded is None:
+            continue
+        processed.append(expanded)
 
     return processed
 
@@ -343,18 +694,44 @@ def _select_e2e_group_steps(steps: list[Any]) -> list[Any]:
     return selected
 
 
+def _any_source_dependency_match(steps: list[Any], changed_files: list[str]) -> bool:
+    """True when any step with ``source_file_dependencies`` matches *changed_files*."""
+    for step in steps:
+        if not isinstance(step, dict):
+            continue
+        deps = _resolve_source_file_dependencies(step)
+        if deps is not None and _match_source_file(changed_files, deps):
+            return True
+        nested = step.get("steps")
+        if isinstance(nested, list) and _any_source_dependency_match(nested, changed_files):
+            return True
+    return False
+
+
 def _render_test_pipeline(
     doc: dict[str, Any],
     changed_files: list[str] | None,
     *,
+    pipeline_path: Path | None = None,
     e2e_only: bool = False,
 ) -> dict[str, Any]:
     """Filter steps by PR diff and strip uploader-only ``source_file_dependencies`` metadata."""
+    # Validate MIRROR_HW once up front. Per-step skip must not run for typos
+    # (e.g. b20o), or a pipeline can shrink to leftover CPU-only steps and
+    # still upload successfully.
+    _get_mirror_hw_selector()
     steps = doc.get("steps")
     if not isinstance(steps, list):
         return doc
     if e2e_only:
         steps = _select_e2e_group_steps(steps)
+    # Bypass is a fallback: only when no job-key prefix matched.
+    if changed_files is not None and pipeline_path is not None:
+        if not _any_source_dependency_match(steps, changed_files):
+            bypass = _source_filter_fallback_reason(changed_files, pipeline_path)
+            if bypass is not None:
+                _log(f"keep all jobs (no source_file_dependencies match; bypassed by {bypass})")
+                changed_files = None
     steps = _process_test_steps(steps, changed_files)
     return {**doc, "steps": steps}
 
@@ -362,6 +739,60 @@ def _render_test_pipeline(
 # --- Entry (read file → bootstrap or test render → YAML string) ---
 
 
+def _repo_relative_posix(path: Path) -> str:
+    resolved = path.resolve()
+    try:
+        return resolved.relative_to(ROOT).as_posix()
+    except ValueError:
+        return path.as_posix().replace("\\", "/")
+
+
+def _source_filter_fallback_prefixes() -> list[str]:
+    """Shared uploader/registry prefixes from ``source_filter_fallback``. Not a job key."""
+    registry = _load_source_file_dependencies()
+    prefixes = registry.get(SOURCE_FILTER_FALLBACK_KEY)
+    if not prefixes:
+        raise ValueError(
+            f"source_file_dependencies[{SOURCE_FILTER_FALLBACK_KEY!r}] must list shared uploader paths "
+            f"in {CI_SOURCE_FILE_DEPENDENCIES_PATH}",
+        )
+    return prefixes
+
+
+def _source_filter_fallback_reason(changed_files: list[str] | None, pipeline_path: Path) -> str | None:
+    """Return the changed path that should disable source filtering, if any."""
+    if not changed_files:
+        return None
+    changed = set(changed_files)
+    pipeline_rel = _repo_relative_posix(pipeline_path)
+    if pipeline_rel in changed:
+        return pipeline_rel
+    prefixes = _source_filter_fallback_prefixes()
+    for path in changed_files:
+        if _match_source_file([path], prefixes):
+            return path
+    return None
+
+
+def _changed_files_for_source_filter(
+    ctx,
+    *,
+    force_all: bool,
+    e2e_only: bool,
+) -> list[str] | None:
+    """Return the diff to filter against, or None to keep every step.
+
+    ``--all`` / ``--e2e`` disable filtering (scheduled NIGHTLY/WEEKLY full
+    uploads and weekly E2E sweeps). Post-merge ``main`` L3 and PR-label
+    uploads filter against the commit/PR diff. Pipeline YAML /
+    ``source_filter_fallback`` matches are applied later in
+    ``_render_test_pipeline`` only when no job-key prefix already matched.
+    """
+    if force_all or e2e_only:
+        return None
+    return ctx.changed_files
+
+
 def _render_pipeline(
     path: Path,
     *,
@@ -379,16 +810,18 @@ def _render_pipeline(
 
     text = path.read_text(encoding="utf-8")
     ctx = resolve_ci_context_from_git()
-    if force_all or e2e_only:
-        changed_files = None
-    else:
-        changed_files = ctx.changed_files
+    changed_files = _changed_files_for_source_filter(ctx, force_all=force_all, e2e_only=e2e_only)
 
     doc = yaml.safe_load(text)
     if not isinstance(doc, dict):
         raise ValueError(f"invalid pipeline YAML: {path}")
 
-    doc = _render_test_pipeline(doc, changed_files, e2e_only=e2e_only)
+    doc = _render_test_pipeline(
+        doc,
+        changed_files,
+        pipeline_path=path,
+        e2e_only=e2e_only,
+    )
     return yaml.safe_dump(doc, sort_keys=False)
 
 

.buildkite/cuda/bootstrap-upload-steps.yml

@@ -8,48 +8,59 @@ steps:
     agents:
       queue: "cpu_queue_premerge"
 
-  # L2 Test — PR ready label (diff-aware) or main+NIGHTLY=1 (--e2e, full E2E group)
+  # L2 Test — PR ready label (diff-aware) or main+WEEKLY=1 (--e2e, full E2E group)
   - label: "Upload Ready Pipeline"
     depends_on: image-build
     key: upload-ready-pipeline
     commands:
       - |
         EXTRA=""
-        if [ "${NIGHTLY:-}" = "1" ] && [ "${BUILDKITE_BRANCH}" = "main" ]; then
+        if [ "${WEEKLY:-}" = "1" ] && [ "${BUILDKITE_BRANCH}" = "main" ]; then
           EXTRA="--e2e"
         fi
         python3 .buildkite/common/scripts/upload_pipeline.py --upload $$EXTRA .buildkite/cuda/test-ready.yml
     agents:
       queue: "cpu_queue_premerge"
 
-  # L3 Test — PR merge-test label (diff-aware) or main+NIGHTLY=1 (--e2e, full E2E group)
+  # L3 Test — PR merge-test label or post-merge on main (both diff-aware);
+  # main+WEEKLY=1 uploads the E2E group only (--e2e).
   - label: "Upload Merge Pipeline"
     depends_on: image-build
     key: upload-merge-pipeline
     commands:
       - |
         EXTRA=""
-        if [ "${NIGHTLY:-}" = "1" ] && [ "${BUILDKITE_BRANCH}" = "main" ]; then
+        if [ "${WEEKLY:-}" = "1" ] && [ "${BUILDKITE_BRANCH}" = "main" ]; then
           EXTRA="--e2e"
         fi
         python3 .buildkite/common/scripts/upload_pipeline.py --upload $$EXTRA .buildkite/cuda/test-merge.yml
     agents:
       queue: "cpu_queue_premerge"
 
-  # L4 Test — main+NIGHTLY=1 (scheduled), or PR with specific label (e.g. add label then Rebuild)
+  # L4 Test — main+NIGHTLY=1 (full pipeline), or PR nightly-test label (diff-aware)
   - label: "Upload Nightly Pipeline"
     depends_on: image-build
     key: upload-nightly-pipeline
     commands:
-      - python3 .buildkite/common/scripts/upload_pipeline.py --upload .buildkite/cuda/test-nightly.yml
+      - |
+        EXTRA=""
+        if [ "${NIGHTLY:-}" = "1" ] && [ "${BUILDKITE_BRANCH}" = "main" ]; then
+          EXTRA="--all"
+        fi
+        python3 .buildkite/common/scripts/upload_pipeline.py --upload $$EXTRA .buildkite/cuda/test-nightly.yml
     agents:
       queue: "cpu_queue_premerge"
 
-  # L5 Test — main+WEEKLY=1 (scheduled), or PR with weekly-test label
+  # L5 Test — main+WEEKLY=1 / NON_CRITICAL=1 (full pipeline), or PR weekly-test label (diff-aware)
   - label: "Upload Weekly Pipeline"
     depends_on: image-build
     key: upload-weekly-pipeline
     commands:
-      - python3 .buildkite/common/scripts/upload_pipeline.py --upload .buildkite/cuda/test-weekly.yml
+      - |
+        EXTRA=""
+        if [ "${BUILDKITE_BRANCH}" = "main" ]; then
+          EXTRA="--all"
+        fi
+        python3 .buildkite/common/scripts/upload_pipeline.py --upload $$EXTRA .buildkite/cuda/test-weekly.yml
     agents:
       queue: "cpu_queue_premerge"

.buildkite/cuda/test-merge.yml

@@ -1,6 +1,9 @@
+# CUDA CI L3 (`advanced_model` e2e) plus L1 CPU Simple Test (`core_model and cpu`).
+# Uploaded by Upload Merge Pipeline: PR label `merge-test`, or post-merge on main
+# (both diff-aware via source_file_dependencies).
+
 env:
   VLLM_WORKER_MULTIPROC_METHOD: spawn
-  VLLM_USE_FLASHINFER_MOE_FP16: 0
   HF_HUB_DOWNLOAD_TIMEOUT: 300
   HF_HUB_ETAG_TIMEOUT: 60
 
@@ -9,23 +12,27 @@ steps:
     depends_on: upload-merge-pipeline
     steps:
       - label: "Simple · Model Executor Test"
+        timeout_in_minutes: 20
         commands:
-          - "timeout 20m pytest -sv tests/model_executor -m 'core_model and cpu' --cov=vllm_omni --cov-report=term-missing:skip-covered --cov-report=xml"
+          - "pytest -sv tests/model_executor -m 'core_model and cpu'"
         mirror_hardwares: l4_1
 
       - label: "Simple · Diffusion Test"
+        timeout_in_minutes: 20
         commands:
-          - "timeout 20m pytest -sv tests/diffusion -m 'core_model and cpu' --cov=vllm_omni --cov-report=term-missing:skip-covered --cov-report=xml"
+          - "pytest -sv tests/diffusion -m 'core_model and cpu'"
         mirror_hardwares: l4_1
 
       - label: "Simple · Engine&Entrypoints Test"
+        timeout_in_minutes: 20
         commands:
-          - "timeout 20m pytest -sv tests/entrypoints tests/engine -m 'core_model and cpu' --cov=vllm_omni --cov-report=term-missing:skip-covered --cov-report=xml"
+          - "pytest -sv tests/entrypoints tests/engine -m 'core_model and cpu'"
         mirror_hardwares: l4_1
 
       - label: "Simple · Other Test"
+        timeout_in_minutes: 20
         commands:
-          - "timeout 20m pytest -sv tests/ -m 'core_model and cpu' --ignore=tests/diffusion --ignore=tests/model_executor --ignore=tests/entrypoints --ignore=tests/engine --cov=vllm_omni --cov-report=term-missing:skip-covered --cov-report=xml"
+          - "pytest -sv tests/ -m 'core_model and cpu' --ignore=tests/diffusion --ignore=tests/model_executor --ignore=tests/entrypoints --ignore=tests/engine"
         mirror_hardwares: l4_1
 
   - group: ":card_index_dividers: Diffusion Test"
@@ -59,7 +66,9 @@ steps:
         # 0.26 startup regression is understood and fixed.
         timeout_in_minutes: 90
         commands:
-          - pytest -s -v tests/entrypoints/ -m "advanced_model and cuda and H100" --run-level "advanced_model"
+          # Turn-based server VAD over /v1/realtime is deferred to RFC #7181 PR 5
+          # (see the P0 thread on #7413); the test stays in the tree and runs again there.
+          - pytest -s -v tests/entrypoints/ -m "advanced_model and cuda and H100" --run-level "advanced_model" --deselect tests/entrypoints/openai_api/test_qwen3_omni_realtime_websocket.py::TestQwen3OmniRealtimeWebSocket::test_server_vad_multi_turn_without_client_commit
         mirror_hardwares: h100_2
 
       - label: "Entrypoint Test with L4"
@@ -72,146 +81,60 @@ steps:
     depends_on: upload-merge-pipeline
     steps:
       - label: "TTS · Qwen3-TTS CustomVoice Test"
-        source_file_dependencies:
-          - tests/e2e/online_serving/test_qwen3_tts_customvoice.py
-          - tests/e2e/offline_inference/test_qwen3_tts_customvoice.py
-          - vllm_omni/model_executor/models/qwen3_tts/
-          - vllm_omni/model_executor/stage_input_processors/qwen3_tts.py
-          - vllm_omni/deploy/qwen3_tts.yaml
+        source_file_dependencies: tts_qwen3_tts_function
         commands:
           - |
             timeout 40m bash -c "
               export VLLM_LOGGING_LEVEL=DEBUG
               export VLLM_ALLOW_LONG_MAX_MODEL_LEN=1
-              pytest -s -v tests/e2e/online_serving/test_qwen3_tts_customvoice.py tests/e2e/offline_inference/test_qwen3_tts_customvoice.py -m 'advanced_model and cuda' --run-level 'advanced_model'
+              pytest -s -v tests/e2e/online_serving/test_qwen3_tts_customvoice.py tests/e2e/offline_inference/test_qwen3_tts_customvoice.py tests/e2e/features/prefill_phase/test_chunked_prefill_tail.py -m 'advanced_model and cuda' --run-level 'advanced_model'
             "
         mirror_hardwares: l4_1
 
       - label: "TTS · Qwen3-TTS Base Test"
-        source_file_dependencies:
-          - tests/e2e/online_serving/test_qwen3_tts_base.py
-          - tests/e2e/offline_inference/test_qwen3_tts_base.py
-          - vllm_omni/model_executor/models/qwen3_tts/
-          - vllm_omni/model_executor/stage_input_processors/qwen3_tts.py
-          - vllm_omni/deploy/qwen3_tts.yaml
+        source_file_dependencies: tts_qwen3_tts_cov
+        # 40 -> 60: the coverage split runs the model twice.
+        timeout_in_minutes: 60
         commands:
           - |
-            timeout 40m bash -c "
-              export VLLM_LOGGING_LEVEL=DEBUG
-              export VLLM_ALLOW_LONG_MAX_MODEL_LEN=1
-              pytest -s -v tests/e2e/online_serving/test_qwen3_tts_base.py tests/e2e/offline_inference/test_qwen3_tts_base.py -m 'advanced_model and cuda' --run-level 'advanced_model'
-            "
-        mirror_hardwares: l4_1
-
-      - label: "TTS · Higgs-Audio-V3 Test"
-        source_file_dependencies:
-          - tests/e2e/online_serving/test_higgs_audio_v3.py
-          - vllm_omni/model_executor/models/higgs_audio_v3/
-          - vllm_omni/model_executor/stage_input_processors/higgs_audio_v3.py
-          - vllm_omni/deploy/higgs_multimodal_qwen3.yaml
-        commands:
-          - |
-            timeout 20m bash -c "
-              export VLLM_WORKER_MULTIPROC_METHOD=spawn
-              export VLLM_USE_DEEP_GEMM=0
-              export VLLM_MOE_USE_DEEP_GEMM=0
-              pytest -s -v tests/e2e/online_serving/test_higgs_audio_v3.py -m 'advanced_model' --run-level 'advanced_model'
-            "
+            export VLLM_LOGGING_LEVEL=DEBUG
+            export VLLM_ALLOW_LONG_MAX_MODEL_LEN=1
+            .buildkite/common/scripts/run_cov_split.sh \
+              --model-id qwen3_tts \
+              --offline tests/e2e/offline_inference/test_qwen3_tts_base.py \
+              --online tests/e2e/online_serving/test_qwen3_tts_base.py \
+              -- -m 'advanced_model and cuda' --run-level 'advanced_model'
         mirror_hardwares: l4_1
 
       - label: "Omni · Qwen3-Omni Test"
-        source_file_dependencies:
-          - tests/e2e/offline_inference/test_qwen3_omni.py
-          - tests/e2e/online_serving/test_qwen3_omni.py
-          - vllm_omni/model_executor/models/qwen3_omni/
-          - vllm_omni/model_executor/stage_input_processors/qwen3_omni.py
-          - vllm_omni/deploy/qwen3_omni_moe.yaml
+        source_file_dependencies: omni_qwen3_omni_function
         timeout_in_minutes: 50
         commands:
-          - pytest -s -v tests/e2e/offline_inference/test_qwen3_omni.py tests/e2e/online_serving/test_qwen3_omni.py -m 'advanced_model and cuda' --run-level 'advanced_model'
+          - pytest -s -v tests/e2e/offline_inference/test_qwen3_omni.py tests/e2e/online_serving/test_qwen3_omni.py tests/e2e/offline_inference/test_qwen3_omni_colocate_async.py -m 'advanced_model and cuda' --run-level 'advanced_model'
         mirror_hardwares: h100_2
 
       - label: "Omni · MiniCPM-o 4.5 Test"
-        source_file_dependencies:
-          - tests/e2e/offline_inference/test_minicpmo_4_5.py
-          - tests/e2e/online_serving/test_minicpmo_4_5.py
-          - vllm_omni/deploy/minicpmo_4_5.yaml
+        source_file_dependencies: omni_minicpmo_4_5_function
         timeout_in_minutes: 50
         commands:
           - pytest -s -v tests/e2e/offline_inference/test_minicpmo_4_5.py tests/e2e/online_serving/test_minicpmo_4_5.py -m 'advanced_model and cuda' --run-level 'advanced_model'
         # h100_2, not h100_1: advanced_model asserts transcribe every audio
-        # response with Whisper, and tests/helpers/media.py only uses an
-        # accelerator for Whisper when a spare GPU exists. On a single card it
-        # grades on CPU: ~8-10 min per clip vs ~30 s, which alone overruns the
-        # 50-minute job timeout (build 2956).
+        # response with Whisper, and tests/helpers/media.py requires at least
+        # 16 GiB free VRAM on a device to run Whisper on GPU. When the model
+        # server leaves insufficient memory, Whisper falls back to CPU:
+        # ~8-10 min per clip vs ~30 s, which overruns the 50-minute job timeout
+        # (build 2956). h100_2 guarantees a second GPU with ample free VRAM.
         mirror_hardwares: h100_2
 
       - label: "Omni · MiniCPM-o 4.5 Duplex Test"
-        source_file_dependencies:
-          - tests/assets/minicpmo_4_5/response_required_16k.wav
-          - tests/assets/minicpmo_4_5/soft_interrupt_16k.wav
-          - tests/e2e/online_serving/test_minicpmo_4_5_duplex.py
-          - tests/e2e/online_serving/test_minicpmo_4_5_duplex_expansion.py
-          - tests/e2e/online_serving/test_minicpmo_realtime_duplex_drivers.py
-          - tests/e2e/online_serving/helpers/minicpmo_realtime_duplex_scenarios.py
-          - tests/e2e/online_serving/run_minicpmo_realtime_duplex_multi_session.py
-          - tests/e2e/online_serving/run_minicpmo_realtime_duplex_soft_interrupt.py
-          - tests/e2e/online_serving/helpers/minicpmo_4_5_duplex.py
-          - vllm_omni/deploy/minicpmo_4_5_duplex.yaml
-          - vllm_omni/experimental/fullduplex/
-          - vllm_omni/model_executor/models/minicpmo_4_5/
-          - vllm_omni/model_executor/stage_input_processors/minicpmo_4_5_omni.py
-          - vllm_omni/core/sched/
-          - vllm_omni/outputs/multimodal_accumulation.py
-          - vllm_omni/engine/stage_init_utils.py
-          - vllm_omni/engine/async_omni_engine.py
-          - vllm_omni/entrypoints/cli/serve.py
+        source_file_dependencies: omni_minicpmo_4_5_duplex_function
         timeout_in_minutes: 50
         commands:
           - pytest -s -v tests/e2e/online_serving/test_minicpmo_4_5_duplex.py -m 'advanced_model and cuda' --run-level 'advanced_model'
         mirror_hardwares: h100_1
 
-      - label: "TTS · VoxCPM2 Test"
-        source_file_dependencies:
-          - tests/e2e/offline_inference/test_voxcpm2_tts.py
-          - tests/e2e/online_serving/test_voxcpm2_tts.py
-          - vllm_omni/model_executor/models/voxcpm2/
-          - vllm_omni/deploy/voxcpm2.yaml
-        timeout_in_minutes: 30
-        commands:
-          - |
-            timeout 30m bash -c "
-              export VLLM_LOGGING_LEVEL=DEBUG
-              export VLLM_WORKER_MULTIPROC_METHOD=spawn
-              pytest -s -v tests/e2e/offline_inference/test_voxcpm2_tts.py tests/e2e/online_serving/test_voxcpm2_tts.py -m 'advanced_model and cuda' --run-level 'advanced_model'
-            "
-        mirror_hardwares: l4_1
-
-      - label: "Diffusion · Qwen Image Edit Test"
-        source_file_dependencies:
-          - tests/e2e/online_serving/test_qwen_image_edit.py
-          - vllm_omni/diffusion/models/qwen_image/
-        timeout_in_minutes: 20
-        commands:
-          - pytest -s -v tests/e2e/online_serving/test_qwen_image_edit.py -m 'advanced_model and cuda' --run-level 'advanced_model'
-        mirror_hardwares: h100_1
-      - label: "Diffusion · Bagel Test"
-        source_file_dependencies:
-          - tests/e2e/offline_inference/test_bagel.py
-          - tests/e2e/online_serving/test_bagel.py
-          - vllm_omni/model_executor/models/bagel/
-          - vllm_omni/diffusion/models/bagel/
-          - vllm_omni/model_executor/stage_input_processors/bagel.py
-          - vllm_omni/deploy/bagel.yaml
-        timeout_in_minutes: 20
-        commands:
-          - pytest -s -v tests/e2e/offline_inference/test_bagel.py tests/e2e/online_serving/test_bagel.py -m 'advanced_model and cuda' --run-level 'advanced_model'
-        mirror_hardwares: h100_1
-
       - label: "Diffusion · Qwen Image Test"
-        source_file_dependencies:
-          - tests/e2e/online_serving/test_qwen_image.py
-          - vllm_omni/diffusion/models/qwen_image/
+        source_file_dependencies: diffusion_qwen_image_function
         commands:
           - |
             timeout 40m bash -c '
@@ -220,23 +143,8 @@ steps:
             '
         mirror_hardwares: h100_1
 
-      - label: "Diffusion · Qwen Image Layered Test"
-        source_file_dependencies:
-          - tests/e2e/online_serving/test_qwen_image_layered.py
-          - vllm_omni/diffusion/models/qwen_image/
-        commands:
-          - |
-            timeout 40m bash -c "
-              export VLLM_IMAGE_FETCH_TIMEOUT=60
-              pytest -s -v tests/e2e/online_serving/test_qwen_image_layered.py -m 'advanced_model and cuda' --run-level 'advanced_model'
-            "
-        mirror_hardwares: h100_1
-
       - label: "Diffusion · Wan22 Test"
-        source_file_dependencies:
-          - tests/e2e/offline_inference/test_wan22_t2v.py
-          - tests/e2e/online_serving/test_wan22_t2v.py
-          - vllm_omni/diffusion/models/wan2_2/
+        source_file_dependencies: diffusion_wan22_function
         commands:
           - |
             timeout 40m bash -c "
@@ -245,20 +153,37 @@ steps:
             "
         mirror_hardwares: h100_1
 
-      - label: "Diffusion · Z Image Test"
-        source_file_dependencies:
-          - tests/e2e/offline_inference/test_z_image.py
-          - tests/e2e/online_serving/test_z_image.py
-          - vllm_omni/diffusion/models/z_image/
-        timeout_in_minutes: 30
+      - label: "Diffusion · HunyuanVideo-1.5 Test"
+        source_file_dependencies: diffusion_hunyuan_video_function
         commands:
-          - pytest -s -v tests/e2e/offline_inference/test_z_image.py tests/e2e/online_serving/test_z_image.py -m 'advanced_model and cuda and diffusion' --run-level 'advanced_model'
-        mirror_hardwares: l4_1
-      - label: "Diffusion · FLUX.2 Klein Test"
-        source_file_dependencies:
-          - tests/e2e/offline_inference/test_flux2_klein.py
-          - vllm_omni/diffusion/models/flux2_klein/
-        timeout_in_minutes: 30
+          - |
+            timeout 40m bash -c "
+              export VLLM_IMAGE_FETCH_TIMEOUT=60
+              pytest -s -v tests/e2e/online_serving/test_hunyuan_video_15_expansion.py -m 'advanced_model and cuda and diffusion' --run-level 'advanced_model'
+            "
+        mirror_hardwares: h100_1
+
+      - label: "Diffusion · MiniMax-H3 L3 DLO DP2 and FastH3"
+        source_file_dependencies: diffusion_minimax_h3_function
+        # Keep four server lifecycles in one Buildkite job, but never in one
+        # pytest process. DLO uses multiprocessing shared-memory queues; the
+        # parent's resource_tracker can retain a stale /psm_* entry during
+        # teardown and strand the next Ref2VA collective.
+        # A per-file process plus a hard timeout also makes a regression fail
+        # instead of leaving the merge queue stuck indefinitely.
+        timeout_in_minutes: 150
         commands:
-          - pytest -s -v tests/e2e/offline_inference/test_flux2_klein.py -m 'advanced_model and cuda and diffusion' --run-level 'advanced_model'
-        mirror_hardwares: l4_1
+          - |
+            set -eu
+            export VLLM_IMAGE_FETCH_TIMEOUT=60
+            if [ -f vllm_omni/diffusion/models/minimax_h3/lora.py ]; then
+              export VLLM_TEST_MINIMAX_H3_TURBO_LORA=$(python3 -c 'from huggingface_hub import hf_hub_download; print(hf_hub_download(repo_id="lightx2v/Minimax-h3-Turbo", filename="minimax_h3_fl2v_turbo_4step_v1.0_768p_bf16.safetensors"))')
+            fi
+            if [ -f vllm_omni/diffusion/models/minimax_h3/fasth3.py ]; then
+              export VLLM_TEST_MINIMAX_H3_FASTH3_LORA=$(python3 -c 'from huggingface_hub import hf_hub_download; print(hf_hub_download(repo_id="FastVideo/FastVideo-FastH3-4-step-Preview-v1-LoRA", filename="dense-datafree/adapter_model.safetensors"))')
+            fi
+            timeout 35m pytest -s -v tests/e2e/online_serving/minimax_h3/test_minimax_h3_dlo_dp2_fl2va.py -m 'advanced_model and cuda and diffusion' --run-level 'advanced_model'
+            timeout 35m pytest -s -v tests/e2e/online_serving/minimax_h3/test_minimax_h3_dlo_dp2_ref2va.py -m 'advanced_model and cuda and diffusion' --run-level 'advanced_model'
+            timeout 35m pytest -s -v tests/e2e/online_serving/minimax_h3/test_minimax_h3_turbo_lora.py -m 'advanced_model and cuda and diffusion' --run-level 'advanced_model'
+            timeout 35m pytest -s -v tests/e2e/online_serving/minimax_h3/test_minimax_h3_fasth3_lora.py -m 'advanced_model and cuda and diffusion' --run-level 'advanced_model'
+        mirror_hardwares: h100_2

.buildkite/cuda/test-nightly.yml

@@ -1,6 +1,9 @@
+# CUDA CI L4 (`full_model`): nightly and release share this file.
+# Uploaded by Upload Nightly Pipeline: NIGHTLY=1, or PR labels (nightly-test / omni-test / ...).
+# Hardware: omit mirror_hardwares to compose from -m SKU + cards_n; or set a preset string to force.
+
 env:
   VLLM_WORKER_MULTIPROC_METHOD: spawn
-  VLLM_USE_FLASHINFER_MOE_FP16: 0
   HF_HUB_DOWNLOAD_TIMEOUT: 300
   HF_HUB_ETAG_TIMEOUT: 60
 
@@ -9,213 +12,170 @@ steps:
   - group: ":card_index_dividers: Omni Model Test"
     key: nightly-omni-test-group
     depends_on: upload-nightly-pipeline
-    if: build.env("NIGHTLY") == "1" || build.pull_request.labels includes "nightly-test" || build.pull_request.labels includes "omni-test"
+    if: build.env("NIGHTLY") == "1" || build.pull_request.labels includes "nightly-test"
     steps:
 
-      - label: ":full_moon: Omni · Function Test with H100"
+      - label: ":full_moon: Omni · Function Test with H100 · Single-GPU"
+        source_file_dependencies:
+          - omni_qwen3_omni_function
+          - omni_minicpmo_4_5_function
         timeout_in_minutes: 120
         commands:
-          - pytest -sv tests/e2e --run-level "full_model" -m "full_model and H100 and omni" --ignore=tests/e2e/accuracy --ignore=tests/e2e/online_serving/test_qwen3_omni_multi_replicas.py --ignore=tests/e2e/online_serving/test_minicpmo_4_5_duplex_expansion.py
-        mirror_hardwares: h100_2
+          - pytest -sv tests/e2e --run-level "full_model" -m "full_model and H100 and B200 and omni and cards_1" --ignore=tests/e2e/accuracy --ignore=tests/e2e/online_serving/test_minicpmo_4_5_duplex_expansion.py
+
+      - label: ":full_moon: Omni · Function Test with H100 · 2-GPU"
+        source_file_dependencies:
+          - omni_qwen3_omni_function
+          - omni_minicpmo_4_5_function
+        timeout_in_minutes: 90
+        commands:
+          - pytest -sv tests/e2e --run-level "full_model" -m "full_model and H100 and B200 and omni and cards_2" --ignore=tests/e2e/accuracy --ignore=tests/e2e/online_serving/test_qwen3_omni_multi_replicas.py
 
       - label: ":full_moon: Omni · MiniCPM-o 4.5 Duplex Test"
+        source_file_dependencies: omni_minicpmo_4_5_duplex_function
         timeout_in_minutes: 50
         commands:
-          - pytest -s -v tests/e2e/online_serving/test_minicpmo_4_5_duplex_expansion.py -m "full_model and cuda and H100 and omni" --run-level "full_model"
-        mirror_hardwares: h100_1
+          - pytest -s -v tests/e2e/online_serving/test_minicpmo_4_5_duplex_expansion.py -m "full_model and cuda and H100 and B200 and omni and cards_1" --run-level "full_model"
 
-      - label: ":full_moon: Omni · Doc Test with L4"
+      - label: ":full_moon: Omni · Doc Test with L4 · 4-GPU"
+        source_file_dependencies: omni_qwen2_5_omni_doc
         timeout_in_minutes: 90
         commands:
           - export VLLM_ALLOW_LONG_MAX_MODEL_LEN="1"
-          - pytest -s -v tests/examples/ -m "full_model and omni and L4" --run-level "full_model"
-        mirror_hardwares: l4_4
+          - pytest -s -v tests/examples/ -m "full_model and omni and L4 and B200 and cards_4" --run-level "full_model"
 
-      - label: ":full_moon: Omni · Doc Test with H100"
+      - label: ":full_moon: Omni · Doc Test with H100 · 2-GPU"
+        source_file_dependencies: omni_qwen3_omni_doc
         timeout_in_minutes: 90
         commands:
-          - pytest -s -v tests/examples/ -m "full_model and omni and H100" --run-level "full_model"
-        mirror_hardwares: h100_2
+          - pytest -s -v tests/examples/ -m "full_model and omni and H100 and B200 and cards_2" --run-level "full_model"
 
       - label: ":full_moon: Omni · Accuracy Test"
+        source_file_dependencies: omni_qwen3_omni_accuracy
         timeout_in_minutes: 180
         commands:
           - export SEED_TTS_WER_EVAL=1
           - export SEED_TTS_EVAL_DEVICE=cuda:1
           - |
             set +e
-            pytest -s -v tests/e2e/accuracy/qwen3_omni/test_qwen3_omni.py -m "full_model" --run-level full_model
+            pytest -s -v tests/e2e/accuracy/qwen3_omni/test_qwen3_omni.py -m "full_model and H100 and B200 and cards_2" --run-level full_model
             EXIT=$$?
             buildkite-agent artifact upload "tests/e2e/accuracy/qwen3_omni/results/qwen_omni_acc/*.json"
             exit $$EXIT
-        mirror_hardwares: h100_2
 
       - label: ":full_moon: Omni · MiniCPM-o 4.5 · Accuracy Test"
+        source_file_dependencies: omni_minicpmo_4_5_accuracy
         timeout_in_minutes: 180
         commands:
           - export SEED_TTS_WER_EVAL=1
           - export SEED_TTS_EVAL_DEVICE=cuda:1
           - |
             set +e
-            pytest -s -v tests/e2e/accuracy/minicpmo_4_5/test_minicpmo_4_5.py -m "full_model and cuda" --run-level full_model
+            pytest -s -v tests/e2e/accuracy/minicpmo_4_5/test_minicpmo_4_5.py -m "full_model and H100 and B200 and cuda and (cards_1 or cards_2)" --run-level full_model
             EXIT=$$?
             buildkite-agent artifact upload "tests/e2e/accuracy/minicpmo_4_5/results/*.json"
             exit $$EXIT
-        mirror_hardwares: h100_2
 
       - label: ":full_moon: Omni · Perf Test · No Async Chunk"
+        source_file_dependencies: omni_qwen3_omni_perf
         key: nightly-omni-performance-no-async-chunk
         timeout_in_minutes: 300
         commands:
           - export BENCHMARK_DIR=tests/dfx/perf/results
           - |
             set +e
-            pytest -s -v tests/dfx/perf/scripts/run_benchmark.py --test-config-file tests/dfx/perf/tests/test_qwen3_omni_no_async_chunk.json
+            pytest -s -v tests/dfx/perf/scripts/run_benchmark.py --test-config-file tests/dfx/perf/tests/test_qwen3_omni_no_async_chunk.json -m "H100 and B200 and cards_2"
             EXIT=$$?
             buildkite-agent artifact upload "tests/dfx/perf/results/*.json"
             exit $$EXIT
-        mirror_hardwares: h100_2
 
       - label: ":full_moon: Omni · Perf Test · Async Chunk"
+        source_file_dependencies: omni_qwen3_omni_perf
         key: nightly-omni-performance-async-chunk
         timeout_in_minutes: 180
         commands:
           - export BENCHMARK_DIR=tests/dfx/perf/results
           - |
             set +e
-            pytest -s -v tests/dfx/perf/scripts/run_benchmark.py --test-config-file tests/dfx/perf/tests/test_qwen3_omni_async_chunk.json
+            pytest -s -v tests/dfx/perf/scripts/run_benchmark.py --test-config-file tests/dfx/perf/tests/test_qwen3_omni_async_chunk.json -m "H100 and B200 and full_model and cards_2"
             EXIT=$$?
             buildkite-agent artifact upload "tests/dfx/perf/results/*.json"
             exit $$EXIT
-        mirror_hardwares: h100_2
-
-      - label: ":full_moon: Omni · Perf Test · vLLM Text"
-        key: nightly-omni-performance-vllm-text
-        timeout_in_minutes: 180
-        commands:
-          - export BENCHMARK_DIR=tests/dfx/perf/results
-          - |
-            set +e
-            pytest -s -v tests/dfx/perf/scripts/run_benchmark.py --test-config-file tests/dfx/perf/tests/test_qwen3_omni_vllm_text.json
-            EXIT=$$?
-            buildkite-agent artifact upload "tests/dfx/perf/results/*.json"
-            exit $$EXIT
-        mirror_hardwares: h100_2
-
-      - label: ":full_moon: Omni · Perf Test · Multi-Replica"
-        key: nightly-omni-performance-multi-replicas
-        timeout_in_minutes: 180
-        commands:
-          - export BENCHMARK_DIR=tests/dfx/perf/results
-          - |
-            set +e
-            pytest -s -v tests/dfx/perf/scripts/run_benchmark.py --test-config-file tests/dfx/perf/tests/test_qwen3_omni_multi_replicas.json
-            EXIT=$$?
-            buildkite-agent artifact upload "tests/dfx/perf/results/*.json"
-            exit $$EXIT
-        mirror_hardwares: h100_3
 
       - label: ":full_moon: Omni · MiniCPM-o 4.5 · Perf Test"
+        source_file_dependencies: omni_minicpmo_4_5_perf
         key: nightly-omni-performance-minicpmo-4-5
         timeout_in_minutes: 180
         commands:
           - export BENCHMARK_DIR=tests/dfx/perf/results
           - |
             set +e
-            pytest -s -v tests/dfx/perf/scripts/run_benchmark.py --test-config-file tests/dfx/perf/tests/test_minicpmo_4_5.json
+            pytest -s -v tests/dfx/perf/scripts/run_benchmark.py --test-config-file tests/dfx/perf/tests/test_minicpmo_4_5.json -m "H100 and B200 and cards_1"
             EXIT=$$?
             buildkite-agent artifact upload "tests/dfx/perf/results/*.json"
             exit $$EXIT
-        mirror_hardwares: h100_1
 
       - label: ":full_moon: Omni · MiniCPM-o 4.5 · Duplex Seed-TTS Perf Test"
+        source_file_dependencies: omni_minicpmo_4_5_duplex_perf
         key: nightly-omni-performance-minicpmo-4-5-duplex-seed-tts
         timeout_in_minutes: 180
         commands:
           - export BENCHMARK_DIR=tests/dfx/perf/results
           - |
             set +e
-            pytest -s -v tests/dfx/perf/scripts/run_benchmark.py --test-config-file tests/dfx/perf/tests/test_minicpmo_4_5_duplex_seed_tts.json
+            pytest -s -v tests/dfx/perf/scripts/run_benchmark.py --test-config-file tests/dfx/perf/tests/test_minicpmo_4_5_duplex_seed_tts.json -m "H100 and B200 and cards_1"
             EXIT=$$?
             buildkite-agent artifact upload "tests/dfx/perf/results/*.json"
             exit $$EXIT
-        mirror_hardwares: h100_1
 
       - label: ":full_moon: Omni · Multi-Replica Startup Test with 4x H100"
+        source_file_dependencies: omni_qwen3_omni_function
         timeout_in_minutes: 45
         commands:
-          - pytest -s -v tests/e2e/online_serving/test_qwen3_omni_multi_replicas.py -m "full_model" --run-level "core_model"
-        mirror_hardwares: h100_4
+          - pytest -s -v tests/e2e/online_serving/test_qwen3_omni_multi_replicas.py -m "full_model and H100 and B200 and cards_4" --run-level "core_model"
 
   - group: ":card_index_dividers: TTS Model Test"
     key: nightly-tts-test-group
     depends_on: upload-nightly-pipeline
-    if: build.env("NIGHTLY") == "1" || build.pull_request.labels includes "nightly-test" || build.pull_request.labels includes "tts-test"
+    if: build.env("NIGHTLY") == "1" || build.pull_request.labels includes "nightly-test"
     steps:
 
       - label: ":full_moon: TTS · Function Test with L4"
+        source_file_dependencies: tts_qwen3_tts_function
         timeout_in_minutes: 120
         commands:
           - export VLLM_ALLOW_LONG_MAX_MODEL_LEN="1"
           - export VLLM_USE_DEEP_GEMM="0"
           - export VLLM_MOE_USE_DEEP_GEMM="0"
-          - pytest -s -v tests/e2e/ -m "full_model and L4 and tts" --run-level "full_model" --ignore=tests/e2e/accuracy --ignore=tests/e2e/online_serving/test_voxcpm2_tts_expansion.py
-        mirror_hardwares: l4_1
+          - pytest -s -v tests/e2e/ -m "full_model and L4 and B200 and tts and cards_1" --run-level "full_model" --ignore=tests/e2e/accuracy
 
-      - label: ":full_moon: TTS · Accuracy Test · VoxCPM2"
-        timeout_in_minutes: 30
-        commands:
-          - |
-            timeout 30m bash -c "
-              export VLLM_LOGGING_LEVEL=DEBUG
-              export VLLM_WORKER_MULTIPROC_METHOD=spawn
-              export VLLM_ALLOW_LONG_MAX_MODEL_LEN=1
-              pytest -s -v tests/e2e/online_serving/test_voxcpm2_tts_expansion.py -m 'full_model and cuda and tts' --run-level 'full_model'
-            "
-        mirror_hardwares: l4_1
-
-      - label: ":full_moon: TTS · Perf Test"
-        key: nightly-tts-performance
+      - label: ":full_moon: TTS · Perf Test · Single-GPU"
+        source_file_dependencies: tts_qwen3_tts_perf
+        key: nightly-tts-performance-single
         timeout_in_minutes: 180
         commands:
           - export BENCHMARK_DIR=tests/dfx/perf/results
           - export VLLM_ALLOW_LONG_MAX_MODEL_LEN="1"
           - |
             set +e
-            pytest -s -v tests/dfx/perf/scripts/run_benchmark.py --test-config-file tests/dfx/perf/tests/test_tts.json
+            pytest -s -v tests/dfx/perf/scripts/run_benchmark.py --test-config-file tests/dfx/perf/tests/test_tts.json -m "H100 and B200 and cards_1"
             EXIT=$$?
             buildkite-agent artifact upload "tests/dfx/perf/results/*.json"
             exit $$EXIT
-        mirror_hardwares: h100_2
 
-      - label: ":full_moon: TTS · VoxCPM2 · Perf Test"
-        key: nightly-tts-performance-voxcpm2
+      - label: ":full_moon: TTS · Perf Test · 2-GPU"
+        source_file_dependencies: tts_qwen3_tts_perf
+        key: nightly-tts-performance-2gpu
         timeout_in_minutes: 180
         commands:
           - export BENCHMARK_DIR=tests/dfx/perf/results
           - export VLLM_ALLOW_LONG_MAX_MODEL_LEN="1"
           - |
             set +e
-            pytest -s -v tests/dfx/perf/scripts/run_benchmark.py --test-config-file tests/dfx/perf/tests/test_voxcpm2.json
+            pytest -s -v tests/dfx/perf/scripts/run_benchmark.py --test-config-file tests/dfx/perf/tests/test_tts.json -m "H100 and B200 and cards_2"
             EXIT=$$?
             buildkite-agent artifact upload "tests/dfx/perf/results/*.json"
             exit $$EXIT
-        mirror_hardwares: h100_1
-
-      - label: ":full_moon: TTS · Higgs-Audio-V3 · Perf Test"
-        key: nightly-tts-performance-higgs-audio-v3
-        timeout_in_minutes: 180
-        commands:
-          - export BENCHMARK_DIR=tests/dfx/perf/results
-          - export VLLM_USE_DEEP_GEMM="0"
-          - export VLLM_MOE_USE_DEEP_GEMM="0"
-          - |
-            set +e
-            pytest -s -v tests/dfx/perf/scripts/run_benchmark.py --test-config-file tests/dfx/perf/tests/test_higgs_audio_v3.json
-            EXIT=$$?
-            buildkite-agent artifact upload "tests/dfx/perf/results/*.json"
-            exit $$EXIT
-        mirror_hardwares: h100_1
 
   - group: ":card_index_dividers: Tiny Model Tests [Multi-GPU]"
     key: nightly-tiny-model-test-group
@@ -225,7 +185,8 @@ steps:
       build.pull_request.labels includes "nightly-test"
     steps:
 
-      - label: ":full_moon: Tiny Model Tests · Diffusion (Multi-GPU accelerations)"
+      - label: ":full_moon: Tiny Model Tests · Diffusion (Multi-GPU accelerations) · 2-GPU"
+        source_file_dependencies: diffusion_tiny_model
         timeout_in_minutes: 30
         # NOTE: Single-GPU (core_model) tiny model tests run on the ready label,
         # so nightly only adds the multi-GPU parallelism configurations.
@@ -235,336 +196,317 @@ steps:
         # for the CI, since these tests should pass irrespective of whether or not we use
         # the full weights or not.
         commands:
-          - timeout 25m pytest -sv tests/model_tests/diffusion -m 'full_model and cuda' --run-level "core_model"
-        mirror_hardwares: l4_4
+          - pytest -sv tests/model_tests/diffusion -m 'full_model and L4 and B200 and cuda and cards_2' --run-level "core_model"
+
+      - label: ":full_moon: Tiny Model Tests · Diffusion (Multi-GPU accelerations) · 4-GPU"
+        source_file_dependencies: diffusion_tiny_model
+        timeout_in_minutes: 30
+        # CFG+TP extra_test_groups request 2×2=4 devices (see get_required_device_count).
+        commands:
+          - pytest -sv tests/model_tests/diffusion -m 'full_model and L4 and B200 and cuda and cards_4' --run-level "core_model"
 
   - group: ":card_index_dividers: Diffusion X2I(&A&T) Model Test"
     key: nightly-diffusion-x2iat-group
     depends_on: upload-nightly-pipeline
     if: >-
       build.env("NIGHTLY") == "1" ||
-      build.pull_request.labels includes "nightly-test" ||
-      build.pull_request.labels includes "diffusion-x2iat-test"
+      build.pull_request.labels includes "nightly-test"
     steps:
-      # Use explicit file shards so single-card cases do not reserve 2+ GPUs,
-      # and Bagel feature coverage is not duplicated via a broad tests/e2e sweep.
 
       - label: ":full_moon: Diffusion X2I(&A&T) · Function Test with H100 · Single-GPU"
+        source_file_dependencies: diffusion_qwen_image_function
         timeout_in_minutes: 120
         commands:
           - >-
             pytest -sv
             tests/e2e/online_serving/test_qwen_image_expansion.py
-            tests/e2e/online_serving/test_qwen_image_edit_expansion.py
-            tests/e2e/online_serving/test_qwen_image_layered_expansion.py
-            tests/e2e/online_serving/test_boogu_image.py
-            tests/e2e/online_serving/test_boogu_image_edit.py
-            tests/e2e/online_serving/test_boogu_image_expansion.py
-            tests/e2e/online_serving/test_boogu_image_edit_expansion.py
-            tests/e2e/online_serving/test_bagel_expansion.py
-            tests/e2e/online_serving/test_krea2_expansion.py
-            tests/e2e/offline_inference/test_krea2_expansion.py
-            -m "full_model and diffusion and H100 and not distributed_cuda"
+            -m "full_model and diffusion and H100 and B200 and cards_1"
             --run-level "full_model"
-        mirror_hardwares: h100_1
 
-      - label: ":full_moon: Diffusion X2I(&A&T) · Function Test with H100 · Multi-GPU Qwen-Image"
+      - label: ":full_moon: Diffusion X2I(&A&T) · Function Test with H100 · 2-GPU"
+        source_file_dependencies: diffusion_qwen_image_function
         timeout_in_minutes: 120
         commands:
           - >-
             pytest -sv
             tests/e2e/online_serving/test_qwen_image_expansion.py
-            tests/e2e/online_serving/test_qwen_image_edit_expansion.py
-            tests/e2e/offline_inference/test_krea2_expansion.py
-            -m "full_model and diffusion and H100 and distributed_cuda"
+            -m "full_model and diffusion and H100 and B200 and cards_2"
             --run-level "full_model"
-        mirror_hardwares: h100_2
 
-      - label: ":full_moon: Diffusion X2I(&A&T) · Function Test with H100 · Multi-GPU Layered/LongCat/FLUX"
-        timeout_in_minutes: 90
-        commands:
-          - >-
-            pytest -sv
-            tests/e2e/online_serving/test_qwen_image_layered_expansion.py
-            -m "full_model and diffusion and H100 and distributed_cuda"
-            --run-level "full_model"
-        mirror_hardwares: h100_2
-
-      - label: ":full_moon: Diffusion X2I(&A&T) · Function Test with H100 · Multi-GPU Bagel"
+      - label: ":full_moon: Diffusion X2I(&A&T) · Function Test with L4 · Single-GPU"
+        source_file_dependencies: diffusion_qwen_image_function
         timeout_in_minutes: 120
         commands:
-          - >-
-            pytest -sv
-            tests/e2e/online_serving/test_bagel_expansion.py
-            -m "full_model and diffusion and H100 and distributed_cuda"
-            --run-level "full_model"
-        mirror_hardwares: h100_2
+          - pytest -sv tests/e2e/offline_inference/test_qwen_image_autoround_w4a16_expansion.py -m "full_model and diffusion and L4 and B200 and cards_1" --run-level "full_model"
 
-      - label: ":full_moon: Diffusion X2I(&A&T) · GLM-Image Function Test with H100"
-        timeout_in_minutes: 120
-        commands:
-          - >-
-            pytest -sv
-            tests/e2e/online_serving/test_glm_image_expansion.py
-            tests/e2e/offline_inference/test_glm_image_autoround_w4a16_expansion.py
-            -m "full_model and diffusion and H100"
-            --run-level "full_model"
-        mirror_hardwares: h100_2
-
-      - label: ":full_moon: Diffusion X2I(&A&T) · Function Test with L4"
-        timeout_in_minutes: 120
-        commands:
-          - pytest -sv tests/e2e/ -k "not test_wan and not test_bagel_expansion and not hunyuan" -m "full_model and diffusion and L4" --run-level "full_model" --ignore=tests/e2e/accuracy
-        mirror_hardwares: l4_4
 
       - label: ":full_moon: Diffusion X2I(&A&T) · Doc Test"
+        source_file_dependencies: diffusion_text_to_image_doc
         timeout_in_minutes: 60
         commands:
-          - pytest -s -v tests/examples/*/test_text_to_image.py -m "full_model and example and H100" --run-level "full_model"
-        mirror_hardwares: h100_2
+          - pytest -s -v tests/examples/*/test_text_to_image.py -m "full_model and example and H100 and B200 and (cards_1 or cards_2)" --run-level "full_model"
 
       - label: ":full_moon: Diffusion X2I(&A&T) · Accuracy Test"
+        source_file_dependencies: diffusion_qwen_image_accuracy
         timeout_in_minutes: 180
         commands:
           - export VLLM_HTTP_TIMEOUT_KEEP_ALIVE=120
-          - pytest -s -v tests/e2e/accuracy/test_qwen_image*.py --run-level full_model
-          - pytest -s -v tests/e2e/accuracy/test_diffusers_backend_similarity.py -k '2i_matches_diffusers' -m full_model --run-level full_model
-        mirror_hardwares: h100_1
+          - pytest -s -v tests/e2e/accuracy/test_qwen_image.py -m "H100 and B200 and cards_1" --run-level full_model
+          - pytest -s -v tests/e2e/accuracy/test_diffusers_backend_similarity.py -k '2i_matches_diffusers' -m "full_model and H100 and B200 and cards_1" --run-level full_model
 
       - label: ":full_moon: HunyuanImage3-DIT · Accuracy Test"
+        source_file_dependencies: diffusion_hunyuan_image3_accuracy
         timeout_in_minutes: 180
         commands:
           - export HUNYUAN_IMAGE3_MODEL="tencent/HunyuanImage-3.0-Instruct"
           - export HUNYUAN_IMAGE3_DEVICES="0,1,2,3"
-          - pytest -s -v tests/e2e/accuracy/test_hunyuan_image3_pixel_accuracy.py -m full_model --run-level full_model
-        mirror_hardwares: h100_4
+          - pytest -s -v tests/e2e/accuracy/test_hunyuan_image3_pixel_accuracy.py -m "full_model and H100 and B200 and cards_4" --run-level full_model
 
-      - label: ":full_moon: Diffusion X2I(&A&T) · Perf Test · Qwen-Image"
-        key: nightly-diffusion-x2iat-performance-qwen-image
+      - label: ":full_moon: Diffusion X2I(&A&T) · Perf Test · Qwen-Image · Single-GPU"
+        source_file_dependencies: diffusion_qwen_image_perf
+        key: nightly-diffusion-x2iat-performance-qwen-image-single
         timeout_in_minutes: 180
         commands:
           - export DIFFUSION_BENCHMARK_DIR=tests/dfx/perf/results
-          - export DIFFUSION_ATTENTION_BACKEND=FLASH_ATTN
-          - export CACHE_DIT_VERSION=1.3.0
+          - export CACHE_DIT_VERSION=1.5.0
           - |
             set +e
-            pytest -s -v tests/dfx/perf/scripts/run_diffusion_benchmark.py --test-config-file tests/dfx/perf/tests/test_qwen_image_vllm_omni.json
+            pytest -s -v tests/dfx/perf/scripts/run_diffusion_benchmark.py --test-config-file tests/dfx/perf/tests/test_qwen_image_vllm_omni.json -m "H100 and B200 and cards_1"
             EXIT=$$?
             buildkite-agent artifact upload "tests/dfx/perf/results/diffusion_result_*.json"
             buildkite-agent artifact upload "tests/dfx/perf/results/logs/*.log"
             exit $$EXIT
-        mirror_hardwares: h100_4
 
-      - label: ":full_moon: Diffusion X2I(&A&T) · Perf Test · Qwen-Image-Edit"
-        key: nightly-diffusion-x2iat-performance-qwen-image-edit
+      - label: ":full_moon: Diffusion X2I(&A&T) · Perf Test · Qwen-Image · 4-GPU"
+        source_file_dependencies: diffusion_qwen_image_perf
+        key: nightly-diffusion-x2iat-performance-qwen-image-4gpu
         timeout_in_minutes: 180
         commands:
           - export DIFFUSION_BENCHMARK_DIR=tests/dfx/perf/results
-          - export DIFFUSION_ATTENTION_BACKEND=FLASH_ATTN
-          - export CACHE_DIT_VERSION=1.3.0
+          - export CACHE_DIT_VERSION=1.5.0
+          # Do not pin DIFFUSION_ATTENTION_BACKEND: H100 auto-selects FLASH_ATTN;
+          # B200 rejects explicit FLASH_ATTN without FA4 and uses CUDNN/TRTLLM.
           - |
             set +e
-            pytest -s -v tests/dfx/perf/scripts/run_diffusion_benchmark.py --test-config-file tests/dfx/perf/tests/test_qwen_image_edit_vllm_omni.json
+            pytest -s -v tests/dfx/perf/scripts/run_diffusion_benchmark.py --test-config-file tests/dfx/perf/tests/test_qwen_image_vllm_omni.json -m "H100 and B200 and cards_4"
             EXIT=$$?
             buildkite-agent artifact upload "tests/dfx/perf/results/diffusion_result_*.json"
             buildkite-agent artifact upload "tests/dfx/perf/results/logs/*.log"
             exit $$EXIT
-        mirror_hardwares: h100_4
-
-      - label: ":full_moon: Diffusion X2I(&A&T) · Perf Test · Qwen-Image-Edit-2511"
-        key: nightly-diffusion-x2iat-performance-qwen-image-edit-2511
-        timeout_in_minutes: 180
-        commands:
-          - export DIFFUSION_BENCHMARK_DIR=tests/dfx/perf/results
-          - export DIFFUSION_ATTENTION_BACKEND=FLASH_ATTN
-          - export CACHE_DIT_VERSION=1.3.0
-          - |
-            set +e
-            pytest -s -v tests/dfx/perf/scripts/run_diffusion_benchmark.py --test-config-file tests/dfx/perf/tests/test_qwen_image_edit_2511_vllm_omni.json
-            EXIT=$$?
-            buildkite-agent artifact upload "tests/dfx/perf/results/diffusion_result_*.json"
-            buildkite-agent artifact upload "tests/dfx/perf/results/logs/*.log"
-            exit $$EXIT
-        mirror_hardwares: h100_4
-
-      - label: ":full_moon: Diffusion X2I(&A&T) · Perf Test · Qwen-Image-Layered"
-        key: nightly-diffusion-x2iat-performance-qwen-image-layered
-        timeout_in_minutes: 120
-        commands:
-          - export DIFFUSION_BENCHMARK_DIR=tests/dfx/perf/results
-          - export DIFFUSION_ATTENTION_BACKEND=FLASH_ATTN
-          - export CACHE_DIT_VERSION=1.3.0
-          - |
-            set +e
-            pytest -s -v tests/dfx/perf/scripts/run_diffusion_benchmark.py --test-config-file tests/dfx/perf/tests/test_qwen_image_layered_vllm_omni.json
-            EXIT=$$?
-            buildkite-agent artifact upload "tests/dfx/perf/results/diffusion_result_*.json"
-            buildkite-agent artifact upload "tests/dfx/perf/results/logs/*.log"
-            exit $$EXIT
-        mirror_hardwares: h100_1
-
-      - label: ":full_moon: Diffusion X2I(&A&T) · Perf Test · BAGEL"
-        key: nightly-diffusion-x2iat-performance-bagel
-        timeout_in_minutes: 180
-        commands:
-          - export DIFFUSION_BENCHMARK_DIR=tests/dfx/perf/results
-          - export DIFFUSION_ATTENTION_BACKEND=FLASH_ATTN
-          - export CACHE_DIT_VERSION=1.3.0
-          - |
-            set +e
-            pytest -s -v tests/dfx/perf/scripts/run_diffusion_benchmark.py --test-config-file tests/dfx/perf/tests/test_bagel_vllm_omni.json
-            EXIT=$$?
-            buildkite-agent artifact upload "tests/dfx/perf/results/diffusion_result_*.json"
-            buildkite-agent artifact upload "tests/dfx/perf/results/logs/*.log"
-            exit $$EXIT
-        mirror_hardwares: h100_1
 
   - group: ":card_index_dividers: Diffusion X2V Model Test"
     key: nightly-diffusion-x2v-group
     depends_on: upload-nightly-pipeline
     if: >-
       build.env("NIGHTLY") == "1" ||
-      build.pull_request.labels includes "nightly-test" ||
-      build.pull_request.labels includes "diffusion-x2v-test"
+      build.pull_request.labels includes "nightly-test"
     steps:
 
-      - label: ":full_moon: Diffusion X2V · Wan2.2 T2V Function Test"
+      - label: ":full_moon: Diffusion X2V · Wan2.2 T2V Function Test · Single-GPU"
+        source_file_dependencies: diffusion_wan22_function
         timeout_in_minutes: 90
         commands:
-          - pytest -s -v tests/e2e/online_serving/test_wan22_expansion.py -k "t2v" -m "full_model and cuda" --run-level "full_model"
-        mirror_hardwares: h100_2
+          - pytest -s -v tests/e2e/online_serving/test_wan22_expansion.py -k "t2v" -m "full_model and cuda and H100 and B200 and cards_1" --run-level "full_model"
 
-      - label: ":full_moon: Diffusion X2V · Wan2.2 I2V/TI2V Function Test"
+      - label: ":full_moon: Diffusion X2V · Wan2.2 T2V Function Test · 2-GPU"
+        source_file_dependencies: diffusion_wan22_function
         timeout_in_minutes: 90
         commands:
-          - pytest -s -v tests/e2e/online_serving/test_wan22_expansion.py -k "i2v" -m "full_model and cuda" --run-level "full_model"
-        mirror_hardwares: h100_2
+          - pytest -s -v tests/e2e/online_serving/test_wan22_expansion.py -k "t2v" -m "full_model and cuda and H100 and B200 and cards_2" --run-level "full_model"
 
-      - label: ":full_moon: Diffusion X2V · LingBot Video MoE Function Test"
+      - label: ":full_moon: Diffusion X2V · Wan2.2 I2V/TI2V Function Test · Single-GPU"
+        source_file_dependencies: diffusion_wan22_function
         timeout_in_minutes: 90
         commands:
-          - pytest -s -v tests/e2e/online_serving/test_lingbot_video_moe.py -k "video_generation" -m "full_model and cuda" --run-level "full_model"
-        mirror_hardwares: h100_1
+          - pytest -s -v tests/e2e/online_serving/test_wan22_expansion.py -k "i2v" -m "full_model and cuda and H100 and B200 and cards_1" --run-level "full_model"
+
+      - label: ":full_moon: Diffusion X2V · Wan2.2 I2V/TI2V Function Test · 2-GPU"
+        source_file_dependencies: diffusion_wan22_function
+        timeout_in_minutes: 90
+        commands:
+          - pytest -s -v tests/e2e/online_serving/test_wan22_expansion.py -k "i2v" -m "full_model and cuda and H100 and B200 and cards_2" --run-level "full_model"
+
+      - label: ":full_moon: Diffusion X2V · HunyuanVideo-1.5 Function Test"
+        source_file_dependencies: diffusion_hunyuan_video_function
+        timeout_in_minutes: 90
+        commands:
+          - pytest -s -v tests/e2e/online_serving/test_hunyuan_video_15_expansion.py -m "full_model and H100 and B200 and cuda and (cards_1 or cards_2)" --run-level "full_model"
+
+      - label: ":full_moon: Diffusion X2V · LingBot Function Test"
+        source_file_dependencies: diffusion_lingbot_function
+        timeout_in_minutes: 120
+        commands:
+          - >-
+            pytest -s -v
+            tests/e2e/online_serving/test_lingbot_video.py
+            tests/e2e/online_serving/test_lingbot_video_moe.py
+            tests/e2e/offline_inference/test_lingbot_world_v2.py
+            -m "full_model and diffusion and H100 and B200 and cards_1"
+            --run-level "full_model"
 
       - label: ":full_moon: Diffusion X2V · Doc Test"
+        source_file_dependencies: diffusion_image_to_video_doc
         timeout_in_minutes: 60
         commands:
-          - pytest -s -v tests/examples/offline_inference/test_image_to_video.py -m "full_model and example and H100" --run-level "full_model"
-        mirror_hardwares: h100_1
+          - pytest -s -v tests/examples/offline_inference/test_image_to_video.py -m "full_model and example and H100 and B200 and cards_1" --run-level "full_model"
 
-      - label: ":full_moon: Diffusion X2V · Other Function Test"
+      - label: ":full_moon: Diffusion X2V · Wan2.2 Autoround Function Test"
+        source_file_dependencies: diffusion_wan22_function
         timeout_in_minutes: 90
         commands:
-          - pytest -s -v tests/e2e/offline_inference/test_wan22_autoround_w4a16_expansion.py -m "full_model and cuda" --run-level "full_model"
-          - pytest -s -v tests/e2e/online_serving/test_wan_2_1_vace_expansion.py tests/e2e/online_serving/test_ltx2_expansion.py -m "full_model and cuda" --run-level "full_model"
-        mirror_hardwares: h100_2
+          - pytest -s -v tests/e2e/offline_inference/test_wan22_autoround_w4a16_expansion.py -m "full_model and H100 and B200 and cuda and (cards_1 or cards_2)" --run-level "full_model"
+
+      - label: ":full_moon: Diffusion X2V · Wan2.2 Device-Postprocess Equivalence Test"
+        source_file_dependencies: diffusion_wan22_function
+        timeout_in_minutes: 90
+        commands:
+          - |
+            VLLM_OMNI_DEVICE_POSTPROCESS_MODEL=Wan-AI/Wan2.2-TI2V-5B-Diffusers \
+              pytest -s -v tests/e2e/offline_inference/test_device_postprocess_equivalence.py -m "full_model and H100 and B200 and cuda and cards_1" --run-level "full_model"
 
       - label: ":full_moon: Diffusion X2V · Wan2.2 Accuracy Test"
+        source_file_dependencies: diffusion_wan22_accuracy
         timeout_in_minutes: 180
         commands:
-          - pytest -s -v tests/e2e/accuracy/wan22_i2v/test_wan22_i2v_video_similarity.py -m full_model --run-level full_model
-          - pytest -s -v tests/e2e/accuracy/test_diffusers_backend_similarity.py -k '2v_matches_diffusers' -m full_model --run-level full_model
-        mirror_hardwares: h100_2
+          - pytest -s -v tests/e2e/accuracy/wan22_i2v/test_wan22_i2v_video_similarity.py -m "full_model and H100 and B200 and (cards_1 or cards_2)" --run-level full_model
+          - pytest -s -v tests/e2e/accuracy/test_diffusers_backend_similarity.py -k '2v_matches_diffusers' -m "full_model and H100 and B200 and cards_1" --run-level full_model
 
       - label: ":full_moon: Diffusion X2V · MiniMax H3 I2VA/Ref2VA Accuracy Test"
+        source_file_dependencies: diffusion_minimax_h3_accuracy
         timeout_in_minutes: 180
         commands:
-          - pytest -s -v tests/e2e/accuracy/minimax_h3/test_minimax_h3_i2va_ref2va_similarity.py -m full_model --run-level full_model
-        mirror_hardwares: h100_4
+          - pytest -s -v tests/e2e/accuracy/minimax_h3/test_minimax_h3_i2va_ref2va_similarity.py -m "full_model and H100 and B200 and cards_4" --run-level full_model
 
-      - label: ":full_moon: Diffusion X2V · Other Accuracy Test"
+      - label: ":full_moon: Diffusion X2V · HunyuanVideo-1.5 Accuracy Test"
+        source_file_dependencies: diffusion_hunyuan_video_accuracy
         timeout_in_minutes: 180
         commands:
-          - pytest -s -v tests/e2e/accuracy/hunyuanvideo15_t2v/test_hunyuanvideo15_t2v_video_similarity.py -m full_model --run-level full_model
-          - pytest -s -v tests/e2e/accuracy/hunyuanvideo15_i2v/test_hunyuanvideo15_i2v_video_similarity.py -m full_model --run-level full_model
-          - pytest -s -v 'tests/e2e/accuracy/ltx/test_ltx_official_similarity.py::test_ltx_one_stage_matches_official[ltx2_3_i2v]' -m full_model --run-level full_model
-          - pytest -s -v tests/e2e/accuracy/test_video_streaming_output_similarity.py -m full_model --run-level full_model
-        mirror_hardwares: h100_2
+          - pytest -s -v tests/e2e/accuracy/hunyuanvideo15_t2v/test_hunyuanvideo15_t2v_video_similarity.py -m "full_model and H100 and B200 and (cards_1 or cards_2)" --run-level full_model
+          - pytest -s -v tests/e2e/accuracy/hunyuanvideo15_i2v/test_hunyuanvideo15_i2v_video_similarity.py -m "full_model and H100 and B200 and (cards_1 or cards_2)" --run-level full_model
 
-      - label: ":full_moon: Diffusion X2V · Perf Test"
-        key: nightly-diffusion-x2v-performance
+      - label: ":full_moon: Diffusion X2V · Perf Test · Single-GPU"
+        source_file_dependencies: diffusion_wan22_perf
+        key: nightly-diffusion-x2v-performance-single
         timeout_in_minutes: 180
         commands:
           - export DIFFUSION_BENCHMARK_DIR=tests/dfx/perf/results
-          - export DIFFUSION_ATTENTION_BACKEND=FLASH_ATTN
           - |
             set +e
-            pytest -s -v tests/dfx/perf/scripts/run_diffusion_benchmark.py --test-config-file tests/dfx/perf/tests/test_wan22_i2v_vllm_omni.json
+            pytest -s -v tests/dfx/perf/scripts/run_diffusion_benchmark.py --test-config-file tests/dfx/perf/tests/test_wan22_i2v_vllm_omni.json -m "H100 and B200 and cards_1"
+            EXIT=$$?
+            buildkite-agent artifact upload "tests/dfx/perf/results/diffusion_result_*.json"
+            buildkite-agent artifact upload "tests/dfx/perf/results/logs/*.log"
+            exit $$EXIT
+
+      - label: ":full_moon: Diffusion X2V · Perf Test · 2-GPU"
+        source_file_dependencies:
+          - diffusion_wan22_perf
+          - diffusion_cosmos3_perf
+        key: nightly-diffusion-x2v-performance-2gpu
+        timeout_in_minutes: 180
+        commands:
+          - export DIFFUSION_BENCHMARK_DIR=tests/dfx/perf/results
+          - |
+            set +e
+            pytest -s -v tests/dfx/perf/scripts/run_diffusion_benchmark.py --test-config-file tests/dfx/perf/tests/test_wan22_i2v_vllm_omni.json -m "H100 and B200 and cards_2"
             EXIT1=$$?
-            pytest -s -v tests/dfx/perf/scripts/run_diffusion_benchmark.py --test-config-file tests/dfx/perf/tests/test_cosmos3_vllm_omni.json
+            pytest -s -v tests/dfx/perf/scripts/run_diffusion_benchmark.py --test-config-file tests/dfx/perf/tests/test_cosmos3_vllm_omni.json -m "H100 and B200 and cards_2"
+            EXIT2=$$?
+            buildkite-agent artifact upload "tests/dfx/perf/results/diffusion_result_*.json"
+            buildkite-agent artifact upload "tests/dfx/perf/results/logs/*.log"
+            if [ $$EXIT1 -ne 0 ]; then exit $$EXIT1; fi
+            exit $$EXIT2
+
+      - label: ":full_moon: Diffusion X2V · Perf Test · HunyuanVideo-1.5"
+        source_file_dependencies: diffusion_hunyuan_video_perf
+        key: nightly-diffusion-x2v-performance-hunyuanvideo15
+        timeout_in_minutes: 180
+        commands:
+          - export DIFFUSION_BENCHMARK_DIR=tests/dfx/perf/results
+          - |
+            set +e
+            pytest -s -v tests/dfx/perf/scripts/run_diffusion_benchmark.py --test-config-file tests/dfx/perf/tests/test_hunyuanvideo15_t2v_vllm_omni.json -m "H100 and B200 and (cards_1 or cards_2)"
+            EXIT1=$$?
+            pytest -s -v tests/dfx/perf/scripts/run_diffusion_benchmark.py --test-config-file tests/dfx/perf/tests/test_hunyuanvideo15_i2v_vllm_omni.json -m "H100 and B200 and (cards_1 or cards_2)"
             EXIT2=$$?
             buildkite-agent artifact upload "tests/dfx/perf/results/diffusion_result_*.json"
             buildkite-agent artifact upload "tests/dfx/perf/results/logs/*.log"
             if [ $$EXIT1 -ne 0 ]; then exit $$EXIT1; fi
             exit $$EXIT2
-        mirror_hardwares: h100_2
 
       - label: ":full_moon: Diffusion X2V · Perf Test · MiniMax-H3"
+        source_file_dependencies: diffusion_minimax_h3_perf
         key: nightly-diffusion-x2v-performance-minimax-h3
         timeout_in_minutes: 180
         commands:
           - export DIFFUSION_BENCHMARK_DIR=tests/dfx/perf/results
-          - export DIFFUSION_ATTENTION_BACKEND=FLASH_ATTN
           - |
             set +e
-            pytest -s -v tests/dfx/perf/scripts/run_diffusion_benchmark.py --test-config-file tests/dfx/perf/tests/test_minimax_h3_vllm_omni.json
+            pytest -s -v tests/dfx/perf/scripts/run_diffusion_benchmark.py --test-config-file tests/dfx/perf/tests/test_minimax_h3_vllm_omni.json -m "H100 and B200 and cards_4"
             EXIT=$$?
             buildkite-agent artifact upload "tests/dfx/perf/results/diffusion_result_*.json"
             buildkite-agent artifact upload "tests/dfx/perf/results/logs/*.log"
             exit $$EXIT
-        mirror_hardwares: h100_4
 
   - group: ":card_index_dividers: Diffusion Test"
     key: nightly-diffusion-test-group
     depends_on: upload-nightly-pipeline
     if: build.env("NIGHTLY") == "1" || build.pull_request.labels includes "nightly-test"
     steps:
-      - label: ":full_moon: Diffusion · Distributed Test with L4"
+      - label: ":full_moon: Diffusion · Distributed Test with L4 · 2-GPU"
+        source_file_dependencies: diffusion_distributed_attention
         timeout_in_minutes: 60
         commands:
-          - pytest -sv tests/diffusion/distributed/ -m 'full_model and cuda and L4' --run-level "full_model"
-        mirror_hardwares: l4_4
+          - pytest -sv tests/diffusion/distributed/ -m 'full_model and cuda and L4 and B200 and cards_2' --run-level "full_model"
 
-      - label: ":full_moon: Diffusion · Distributed Test with H100"
+      - label: ":full_moon: Diffusion · Distributed Test with L4 · 3-GPU"
+        source_file_dependencies: diffusion_distributed_attention
         timeout_in_minutes: 60
         commands:
-          - pytest -sv tests/diffusion/distributed/ -m 'full_model and cuda and H100' --run-level "full_model"
-        mirror_hardwares: h100_4
+          - pytest -sv tests/diffusion/distributed/ -m 'full_model and cuda and L4 and B200 and cards_3' --run-level "full_model"
+
+      - label: ":full_moon: Diffusion · Distributed Test with L4 · 4-GPU"
+        source_file_dependencies: diffusion_distributed_attention
+        timeout_in_minutes: 60
+        commands:
+          - pytest -sv tests/diffusion/distributed/ -m 'full_model and cuda and L4 and B200 and cards_4' --run-level "full_model"
+
+      - label: ":full_moon: Diffusion · Distributed Test with H100 · 2-GPU"
+        source_file_dependencies: diffusion_distributed_attention
+        timeout_in_minutes: 60
+        commands:
+          - pytest -sv tests/diffusion/distributed/ -m 'full_model and cuda and H100 and B200 and cards_2' --run-level "full_model"
+
+      - label: ":full_moon: Diffusion · Distributed Test with H100 · 4-GPU"
+        source_file_dependencies: diffusion_distributed_attention
+        timeout_in_minutes: 60
+        commands:
+          - pytest -sv tests/diffusion/distributed/ -m 'full_model and cuda and H100 and B200 and cards_4' --run-level "full_model"
 
       - label: ":full_moon: Diffusion · Offloader Test"
+        source_file_dependencies: diffusion_offloader
         timeout_in_minutes: 60
         commands:
-          - pytest -sv tests/diffusion/offloader -m "full_model and cuda"
-        mirror_hardwares: l4_1
+          - pytest -sv tests/diffusion/offloader -m "full_model and L4 and B200 and cuda and cards_1"
 
       - label: ":full_moon: Diffusion · Quantization Test with H100 · Single-GPU"
+        source_file_dependencies: diffusion_quantization
         timeout_in_minutes: 60
         commands:
-          - pytest -sv tests/diffusion/quantization -m 'full_model and cuda and H100' --run-level "full_model"
-        mirror_hardwares: h100_1
+          - pytest -sv tests/diffusion/quantization -m 'full_model and cuda and H100 and B200 and cards_1' --run-level "full_model"
 
       - label: ":full_moon: Diffusion · MiniMax H3 FP8 Quality Test with H100 · 2-GPU"
+        source_file_dependencies: diffusion_minimax_h3_function
         timeout_in_minutes: 60
         commands:
-          - pytest -sv tests/diffusion/models/minimax_h3/test_minimax_h3_quantization_quality.py -m 'full_model and cuda and H100' --run-level "full_model"
-        mirror_hardwares: h100_2
+          - pytest -sv tests/diffusion/models/minimax_h3/test_minimax_h3_quantization_quality.py -m 'full_model and cuda and H100 and B200 and cards_2' --run-level "full_model"
 
       - label: ":full_moon: Diffusion · Quantization Test with L4"
+        source_file_dependencies: diffusion_quantization
         timeout_in_minutes: 60
         commands:
-          - pip install "vllm-gguf-plugin>=0.0.3"
-          - pytest -sv tests/diffusion/quantization -m 'full_model and cuda and L4' --run-level "full_model"
-        mirror_hardwares: l4_1
-
-  - label: ":bar_chart: Testcase Statistics"
-    key: nightly-testcase-statistics
-    timeout_in_minutes: 120
-    depends_on: upload-nightly-pipeline
-    if: build.env("NIGHTLY") == "1"
-    commands:
-      - python tools/nightly/buildkite_testcase_statistics.py -o tests/dfx/perf/results/buildkite_testcase_statistics.html
-      - buildkite-agent artifact upload "tests/dfx/perf/results/*.html"
-    mirror_hardwares: h100_1
+          - pip install "vllm-gguf-plugin==0.0.4"
+          - pytest -sv tests/diffusion/quantization -m 'full_model and cuda and L4 and B200 and cards_1' --run-level "full_model"
 
   # NOTE: ":email: Nightly Collection & Email" is deprecated and will be removed soon.
   # Do not add new depends_on entries or extend this step.
@@ -573,35 +515,25 @@ steps:
     depends_on:
       - nightly-omni-performance-no-async-chunk
       - nightly-omni-performance-async-chunk
-      - nightly-omni-performance-vllm-text
-      - nightly-omni-performance-multi-replicas
-      - nightly-tts-performance
-      - nightly-tts-performance-higgs-audio-v3
-      - nightly-diffusion-x2iat-performance-qwen-image
-      - nightly-diffusion-x2iat-performance-qwen-image-edit
-      - nightly-diffusion-x2iat-performance-qwen-image-edit-2511
-      - nightly-diffusion-x2iat-performance-qwen-image-layered
-      - nightly-diffusion-x2iat-performance-bagel
-      - nightly-diffusion-x2v-performance
-      - nightly-testcase-statistics
+      - nightly-tts-performance-single
+      - nightly-tts-performance-2gpu
+      - nightly-diffusion-x2iat-performance-qwen-image-single
+      - nightly-diffusion-x2iat-performance-qwen-image-4gpu
+      - nightly-diffusion-x2v-performance-single
+      - nightly-diffusion-x2v-performance-2gpu
     if: build.env("NIGHTLY") == "1" && build.env("EMAIL_DISTRIBUTION") == "1"
     commands:
       - pip install openpyxl
       - export DEFAULT_INPUT_DIR=tests/dfx/perf/results
       - export DEFAULT_OUTPUT_DIR=tests/dfx/perf/results
-      - buildkite-agent artifact download "tests/dfx/perf/results/*.json" . --step nightly-tts-performance
-      - buildkite-agent artifact download "tests/dfx/perf/results/*.json" . --step nightly-tts-performance-higgs-audio-v3
+      - buildkite-agent artifact download "tests/dfx/perf/results/*.json" . --step nightly-tts-performance-single
+      - buildkite-agent artifact download "tests/dfx/perf/results/*.json" . --step nightly-tts-performance-2gpu
       - buildkite-agent artifact download "tests/dfx/perf/results/*.json" . --step nightly-omni-performance-no-async-chunk
       - buildkite-agent artifact download "tests/dfx/perf/results/*.json" . --step nightly-omni-performance-async-chunk
-      - buildkite-agent artifact download "tests/dfx/perf/results/*.json" . --step nightly-omni-performance-vllm-text
-      - buildkite-agent artifact download "tests/dfx/perf/results/*.json" . --step nightly-omni-performance-multi-replicas
-      - buildkite-agent artifact download "tests/dfx/perf/results/*.json" . --step nightly-diffusion-x2iat-performance-qwen-image
-      - buildkite-agent artifact download "tests/dfx/perf/results/*.json" . --step nightly-diffusion-x2iat-performance-qwen-image-edit
-      - buildkite-agent artifact download "tests/dfx/perf/results/*.json" . --step nightly-diffusion-x2iat-performance-qwen-image-edit-2511
-      - buildkite-agent artifact download "tests/dfx/perf/results/*.json" . --step nightly-diffusion-x2iat-performance-qwen-image-layered
-      - buildkite-agent artifact download "tests/dfx/perf/results/*.json" . --step nightly-diffusion-x2iat-performance-bagel
-      - buildkite-agent artifact download "tests/dfx/perf/results/*.json" . --step nightly-diffusion-x2v-performance
-      - buildkite-agent artifact download "tests/dfx/perf/results/*.html" . --step nightly-testcase-statistics
+      - buildkite-agent artifact download "tests/dfx/perf/results/*.json" . --step nightly-diffusion-x2iat-performance-qwen-image-single
+      - buildkite-agent artifact download "tests/dfx/perf/results/*.json" . --step nightly-diffusion-x2iat-performance-qwen-image-4gpu
+      - buildkite-agent artifact download "tests/dfx/perf/results/*.json" . --step nightly-diffusion-x2v-performance-single
+      - buildkite-agent artifact download "tests/dfx/perf/results/*.json" . --step nightly-diffusion-x2v-performance-2gpu
       - python tools/nightly/generate_nightly_perf_excel.py
       - python tools/nightly/generate_nightly_perf_html.py
       - python tools/nightly/send_nightly_email.py --report-file "tests/dfx/perf/results/*.xlsx, tests/dfx/perf/results/*.html"

.buildkite/cuda/test-ready.yml

@@ -1,6 +1,8 @@
+# CUDA CI L2 (`core_model` e2e) plus L1 CPU Simple Test (`core_model and cpu`).
+# Uploaded by Upload Ready Pipeline: PR label `ready`, or main+WEEKLY=1 with --e2e.
+
 env:
   VLLM_WORKER_MULTIPROC_METHOD: spawn
-  VLLM_USE_FLASHINFER_MOE_FP16: 0
   HF_HUB_DOWNLOAD_TIMEOUT: 300
   HF_HUB_ETAG_TIMEOUT: 60
 
@@ -9,23 +11,27 @@ steps:
     depends_on: upload-ready-pipeline
     steps:
       - label: "Simple · Model Executor Test"
+        timeout_in_minutes: 20
         commands:
-          - "timeout 20m pytest -sv tests/model_executor -m 'core_model and cpu' --cov=vllm_omni --cov-report=term-missing:skip-covered --cov-report=xml"
+          - "pytest -sv tests/model_executor -m 'core_model and cpu'"
         mirror_hardwares: l4_1
 
       - label: "Simple · Diffusion Test"
+        timeout_in_minutes: 20
         commands:
-          - "timeout 20m pytest -sv tests/diffusion -m 'core_model and cpu' --cov=vllm_omni --cov-report=term-missing:skip-covered --cov-report=xml"
+          - "pytest -sv tests/diffusion -m 'core_model and cpu'"
         mirror_hardwares: l4_1
 
       - label: "Simple · Engine&Entrypoints Test"
+        timeout_in_minutes: 20
         commands:
-          - "timeout 20m pytest -sv tests/entrypoints tests/engine -m 'core_model and cpu' --cov=vllm_omni --cov-report=term-missing:skip-covered --cov-report=xml"
+          - "pytest -sv tests/entrypoints tests/engine -m 'core_model and cpu'"
         mirror_hardwares: l4_1
 
       - label: "Simple · Other Test"
+        timeout_in_minutes: 20
         commands:
-          - "timeout 20m pytest -sv tests/ -m 'core_model and cpu' --ignore=tests/diffusion --ignore=tests/model_executor --ignore=tests/entrypoints --ignore=tests/engine --cov=vllm_omni --cov-report=term-missing:skip-covered --cov-report=xml"
+          - "pytest -sv tests/ -m 'core_model and cpu' --ignore=tests/diffusion --ignore=tests/model_executor --ignore=tests/entrypoints --ignore=tests/engine"
         mirror_hardwares: l4_1
 
   - group: ":card_index_dividers: Diffusion Test"
@@ -47,17 +53,23 @@ steps:
           - timeout 15m pytest -sv tests/diffusion/batching -m 'core_model and cuda' --run-level "core_model"
         mirror_hardwares: l4_1
 
-      - label: "Diffusion · Distributed & Attention Test"
-        source_file_dependencies:
-          - tests/diffusion/distributed/
-          - tests/diffusion/attention/test_attention_sp.py
-          - vllm_omni/diffusion/distributed/
-          - vllm_omni/diffusion/attention/
-          - vllm_omni/diffusion/data.py
-          - vllm_omni/diffusion/models/
+      - label: "Diffusion · Distributed Test · Single-GPU"
+        source_file_dependencies: diffusion_distributed_attention
         commands:
-          - timeout 20m pytest -sv tests/diffusion/distributed -m 'core_model and cuda' --run-level "core_model"
-          - timeout 20m pytest -sv tests/diffusion/attention/test_attention_sp.py -m 'core_model and cuda' --run-level "core_model"
+          - timeout 20m pytest -sv tests/diffusion/distributed -m 'core_model and cuda and L4 and cards_1' --run-level "core_model"
+        mirror_hardwares: l4_1
+
+      - label: "Diffusion · Distributed Test · 2-GPU"
+        source_file_dependencies: diffusion_distributed_attention
+        commands:
+          - timeout 20m pytest -sv tests/diffusion/distributed -m 'core_model and cuda and L4 and cards_2' --run-level "core_model"
+        mirror_hardwares: l4_2
+
+      - label: "Diffusion · Distributed & Attention Test · 4-GPU"
+        source_file_dependencies: diffusion_distributed_attention
+        commands:
+          - timeout 20m pytest -sv tests/diffusion/distributed -m 'core_model and cuda and L4 and cards_4' --run-level "core_model"
+          - timeout 20m pytest -sv tests/diffusion/attention/test_attention_sp.py -m 'core_model and cuda and L4 and cards_4' --run-level "core_model"
         mirror_hardwares: l4_4
 
       - label: "Diffusion · Model Test"
@@ -67,7 +79,7 @@ steps:
 
       - label: "Diffusion · Other Test"
         commands:
-          - timeout 15m pytest -sv tests/diffusion/*.py -m 'core_model and cuda' --run-level "core_model"
+          - timeout 15m pytest -sv tests/diffusion/*.py tests/diffusion/ar_diffusion -m 'core_model and cuda' --run-level "core_model"
         mirror_hardwares: l4_1
 
   - group: ":card_index_dividers: Tiny Model Tests [Single GPU]"
@@ -76,7 +88,7 @@ steps:
 
       - label: "Tiny Model Tests · Diffusion (base)"
         commands:
-          - timeout 15m pytest -sv tests/model_tests/diffusion -m 'core_model and cuda' --run-level "core_model"
+          - timeout 15m pytest -sv -n 4 tests/model_tests/diffusion -m 'core_model and cuda' --run-level "core_model"
         mirror_hardwares: l4_1
 
   - label: "Engine&Model Executor Test"
@@ -84,7 +96,7 @@ steps:
     commands:
       - |
         timeout 15m bash -c "
-                  pytest -sv tests/engine/ tests/model_executor/ -m 'core_model and cuda' --run-level 'core_model'
+                  pytest -sv tests/engine/ tests/model_executor/ tests/worker/test_batched_omni_output.py -m 'core_model and cuda' --run-level 'core_model'
         "
     mirror_hardwares: l4_1
 
@@ -111,53 +123,14 @@ steps:
     depends_on: upload-ready-pipeline
     steps:
       - label: "Custom Pipeline Test"
-        source_file_dependencies:
-          - tests/e2e/features/custom_pipeline/
-          - tests/e2e/features/helpers/custom_pipeline.py
-          - vllm_omni/diffusion/
-          - vllm_omni/engine/
-          - vllm_omni/entrypoints/
-          - vllm_omni/inputs/
-          - vllm_omni/outputs/
+        source_file_dependencies: custom_pipeline
         commands:
           - timeout 20m pytest -s -v tests/e2e/features/custom_pipeline/ -m "core_model"
-        agents:
-          queue: "gpu_1_queue"
-        plugins:
-          - docker#v5.2.0:
-              image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
-              always-pull: true
-              propagate-environment: true
-              # vLLM 0.27 added a pre-flight capacity check in
-              # distributed/device_communicators/shm_broadcast.py. Without this
-              # the container gets Docker's 64 MiB default and every test in
-              # this step dies at engine start with
-              #   RuntimeError: Insufficient space in /dev/shm:
-              #                 240 MiB required, 64 MiB available
-              # (9 failed in 8.7s on build #2949). vLLM 0.26 had no such check
-              # and allocated regardless, so this step passed on main while
-              # being under-provisioned the whole time — 0.27 turned a silent
-              # risk into an explicit error.
-              #
-              # This is the only inline docker plugin block in .buildkite/cuda/;
-              # every other step inherits common/ci_mirror_hardwares.yml, which
-              # already sets 8gb. Matching that rather than inventing a value.
-              shm-size: "8gb"
-              environment:
-                - "HF_HOME=/fsx/hf_cache"
-              volumes:
-                - "/fsx/hf_cache:/fsx/hf_cache"
+        mirror_hardwares: l4_1
+
 
       - label: "RLHF · VeRL-Omni E2E Test"
-        source_file_dependencies:
-          - tests/e2e/features/rlhf_test/test_verl_omni_e2e.py
-          - tests/e2e/features/helpers/verl_omni_server.py
-          - tests/e2e/features/helpers/custom_pipeline.py
-          - vllm_omni/diffusion/
-          - vllm_omni/engine/
-          - vllm_omni/entrypoints/
-          - vllm_omni/inputs/
-          - vllm_omni/outputs/
+        source_file_dependencies: rlhf_verl_omni
         commands:
           - |
             timeout 20m bash -c "
@@ -166,11 +139,7 @@ steps:
         mirror_hardwares: l4_1
 
       - label: "Omni · Qwen3-Omni Test"
-        source_file_dependencies:
-          - tests/e2e/online_serving/test_qwen3_omni.py
-          - vllm_omni/model_executor/models/qwen3_omni/
-          - vllm_omni/model_executor/stage_input_processors/qwen3_omni.py
-          - vllm_omni/deploy/qwen3_omni_moe.yaml
+        source_file_dependencies: omni_qwen3_omni_function
         commands:
           - |
             timeout 20m bash -c "
@@ -179,9 +148,7 @@ steps:
         mirror_hardwares: h100_2
 
       - label: "Omni · MiniCPM-o 4.5 Offline Test"
-        source_file_dependencies:
-          - tests/e2e/offline_inference/test_minicpmo_4_5.py
-          - vllm_omni/deploy/minicpmo_4_5.yaml
+        source_file_dependencies: omni_minicpmo_4_5_function
         commands:
           - |
             timeout 20m bash -c "
@@ -190,9 +157,7 @@ steps:
         mirror_hardwares: h100_1
 
       - label: "Omni · MiniCPM-o 4.5 Online Test"
-        source_file_dependencies:
-          - tests/e2e/online_serving/test_minicpmo_4_5.py
-          - vllm_omni/deploy/minicpmo_4_5.yaml
+        source_file_dependencies: omni_minicpmo_4_5_function
         commands:
           - |
             timeout 20m bash -c "
@@ -201,53 +166,24 @@ steps:
         mirror_hardwares: h100_1
 
       - label: "Omni · MiniCPM-o 4.5 Duplex Test"
-        source_file_dependencies:
-          - tests/assets/minicpmo_4_5/response_required_16k.wav
-          - tests/assets/minicpmo_4_5/soft_interrupt_16k.wav
-          - tests/e2e/online_serving/test_minicpmo_4_5_duplex.py
-          - tests/e2e/online_serving/test_minicpmo_4_5_duplex_expansion.py
-          - tests/e2e/online_serving/test_minicpmo_realtime_duplex_drivers.py
-          - tests/e2e/online_serving/helpers/minicpmo_realtime_duplex_scenarios.py
-          - tests/e2e/online_serving/run_minicpmo_realtime_duplex_multi_session.py
-          - tests/e2e/online_serving/run_minicpmo_realtime_duplex_soft_interrupt.py
-          - tests/e2e/online_serving/helpers/minicpmo_4_5_duplex.py
-          - vllm_omni/deploy/minicpmo_4_5_duplex.yaml
-          - vllm_omni/experimental/fullduplex/
-          - vllm_omni/model_executor/models/minicpmo_4_5/
-          - vllm_omni/model_executor/stage_input_processors/minicpmo_4_5_omni.py
-          - vllm_omni/core/sched/
-          - vllm_omni/outputs/multimodal_accumulation.py
-          - vllm_omni/engine/stage_init_utils.py
-          - vllm_omni/engine/async_omni_engine.py
-          - vllm_omni/entrypoints/cli/serve.py
+        source_file_dependencies: omni_minicpmo_4_5_duplex_function
         commands:
           - |
-            timeout 20m bash -c "
-              pytest -s -v tests/e2e/online_serving/test_minicpmo_4_5_duplex.py -m 'core_model and cuda' --run-level 'core_model'
+            timeout 35m bash -c "
+              pytest -s -v tests/e2e/online_serving/test_minicpmo_4_5_duplex.py tests/e2e/online_serving/test_duplex_client_live.py -m 'core_model and cuda' --run-level 'core_model'
             "
         mirror_hardwares: h100_1
 
-      - label: "TTS · VoxCPM2 Test"
-        source_file_dependencies:
-          - tests/e2e/online_serving/test_voxcpm2_tts.py
-          - vllm_omni/model_executor/models/voxcpm2/
-          - vllm_omni/deploy/voxcpm2.yaml
-        timeout_in_minutes: 30
+      - label: "Omni · MiniCPM-o 4.5 Perf Test"
+        timeout_in_minutes: 60
+        source_file_dependencies: omni_minicpmo_4_5_perf
         commands:
-          - |
-            timeout 30m bash -c "
-              export VLLM_LOGGING_LEVEL=DEBUG
-              export VLLM_WORKER_MULTIPROC_METHOD=spawn
-              pytest -s -v tests/e2e/online_serving/test_voxcpm2_tts.py -m 'core_model' --run-level 'core_model'
-            "
-        mirror_hardwares: l4_1
+          - export BENCHMARK_DIR=tests/dfx/perf/results
+          - pytest -s -v tests/dfx/perf/scripts/run_benchmark.py --test-config-file tests/dfx/perf/tests/test_minicpmo_4_5_ready.json -m 'core_model and cuda'
+        mirror_hardwares: h100_1
 
       - label: "TTS · Qwen3-TTS CustomVoice Test"
-        source_file_dependencies:
-          - tests/e2e/online_serving/test_qwen3_tts_customvoice.py
-          - vllm_omni/model_executor/models/qwen3_tts/
-          - vllm_omni/model_executor/stage_input_processors/qwen3_tts.py
-          - vllm_omni/deploy/qwen3_tts.yaml
+        source_file_dependencies: tts_qwen3_tts_function
         commands:
           - |
             timeout 20m bash -c "
@@ -256,68 +192,8 @@ steps:
             "
         mirror_hardwares: l4_1
 
-      - label: "TTS · Qwen3-TTS Base Test"
-        source_file_dependencies:
-          - tests/e2e/online_serving/test_qwen3_tts_base.py
-          - vllm_omni/model_executor/models/qwen3_tts/
-          - vllm_omni/model_executor/stage_input_processors/qwen3_tts.py
-          - vllm_omni/deploy/qwen3_tts.yaml
-        commands:
-          - |
-            timeout 20m bash -c "
-              export VLLM_WORKER_MULTIPROC_METHOD=spawn
-              export VLLM_ALLOW_LONG_MAX_MODEL_LEN=1
-              export VLLM_OMNI_USE_V2_RUNNER=1
-              pytest -s -v tests/e2e/online_serving/test_qwen3_tts_base.py -m 'core_model' --run-level 'core_model'
-            "
-        mirror_hardwares: l4_1
-
-      - label: "TTS · Higgs-Audio-V3 Test"
-        source_file_dependencies:
-          - tests/e2e/online_serving/test_higgs_audio_v3.py
-          - vllm_omni/model_executor/models/higgs_audio_v3/
-          - vllm_omni/model_executor/stage_input_processors/higgs_audio_v3.py
-          - vllm_omni/deploy/higgs_multimodal_qwen3.yaml
-        commands:
-          - |
-            timeout 20m bash -c "
-              export VLLM_WORKER_MULTIPROC_METHOD=spawn
-              export VLLM_USE_DEEP_GEMM=0
-              export VLLM_MOE_USE_DEEP_GEMM=0
-              pytest -s -v tests/e2e/online_serving/test_higgs_audio_v3.py -m 'core_model' --run-level 'core_model'
-            "
-        mirror_hardwares: l4_1
-
-      - label: "Diffusion · Qwen Image Edit Test"
-        source_file_dependencies:
-          - tests/e2e/online_serving/test_qwen_image_edit.py
-          - vllm_omni/diffusion/models/qwen_image/
-        commands:
-          - |
-            timeout 20m bash -c "
-              pytest -s -v tests/e2e/online_serving/test_qwen_image_edit.py -m 'core_model' --run-level 'core_model'
-            "
-        mirror_hardwares: h100_1
-
-      - label: "Diffusion · Bagel Test"
-        source_file_dependencies:
-          - tests/e2e/online_serving/test_bagel.py
-          - vllm_omni/model_executor/models/bagel/
-          - vllm_omni/diffusion/models/bagel/
-          - vllm_omni/model_executor/stage_input_processors/bagel.py
-          - vllm_omni/deploy/bagel.yaml
-        commands:
-          - |
-            timeout 40m bash -c "
-              export VLLM_IMAGE_FETCH_TIMEOUT=60
-              pytest -s -v tests/e2e/online_serving/test_bagel.py -m 'core_model' --run-level 'core_model'
-            "
-        mirror_hardwares: h100_1
-
       - label: "Diffusion · Wan22 Test"
-        source_file_dependencies:
-          - tests/e2e/online_serving/test_wan22_t2v.py
-          - vllm_omni/diffusion/models/wan2_2/
+        source_file_dependencies: diffusion_wan22_function
         commands:
           - |
             timeout 40m bash -c "
@@ -327,9 +203,7 @@ steps:
         mirror_hardwares: h100_1
 
       - label: "Diffusion · Cosmos3 Test"
-        source_file_dependencies:
-          - tests/e2e/online_serving/test_cosmos3.py
-          - vllm_omni/diffusion/models/cosmos3/
+        source_file_dependencies: diffusion_cosmos3_function
         commands:
           - |
             timeout 40m bash -c "
@@ -338,10 +212,19 @@ steps:
             "
         mirror_hardwares: h100_1
 
+      - label: "Diffusion · MiniMax-H3 DLO DP2 T2VA Test"
+        timeout_in_minutes: 90
+        source_file_dependencies: diffusion_minimax_h3_function
+        commands:
+          - |
+            timeout 80m bash -c "
+              export VLLM_IMAGE_FETCH_TIMEOUT=60
+              pytest -s -v tests/e2e/online_serving/test_minimax_h3_dlo_dp2_t2va.py -m 'core_model and cuda and diffusion' --run-level 'core_model'
+            "
+        mirror_hardwares: h100_2
+
       - label: "Diffusion · Qwen Image Test"
-        source_file_dependencies:
-          - tests/e2e/online_serving/test_qwen_image.py
-          - vllm_omni/diffusion/models/qwen_image/
+        source_file_dependencies: diffusion_qwen_image_function
         commands:
           - |
             timeout 40m bash -c "
@@ -350,14 +233,12 @@ steps:
             "
         mirror_hardwares: h100_1
 
-      - label: "Diffusion · Qwen Image Layered Test"
-        source_file_dependencies:
-          - tests/e2e/online_serving/test_qwen_image_layered.py
-          - vllm_omni/diffusion/models/qwen_image/
+      - label: "Diffusion · HunyuanVideo-1.5 Test"
+        source_file_dependencies: diffusion_hunyuan_video_function
         commands:
           - |
             timeout 40m bash -c "
               export VLLM_IMAGE_FETCH_TIMEOUT=60
-              pytest -s -v tests/e2e/online_serving/test_qwen_image_layered.py -m 'core_model' --run-level 'core_model'
+              pytest -s -v tests/e2e/online_serving/test_hunyuan_video_15_expansion.py -m 'core_model and diffusion' --run-level 'core_model'
             "
         mirror_hardwares: h100_1

.buildkite/cuda/test-weekly.yml

@@ -1,89 +1,167 @@
+# CUDA CI L5 (`slow` / reliability / extra coverage).
+# Uploaded by Upload Weekly Pipeline: WEEKLY=1 / NON_CRITICAL=1, or PR label `weekly-test`.
+
 env:
   VLLM_WORKER_MULTIPROC_METHOD: spawn
-  VLLM_USE_FLASHINFER_MOE_FP16: 0
   HF_HUB_DOWNLOAD_TIMEOUT: 300
   HF_HUB_ETAG_TIMEOUT: 60
 
 steps:
-
-  - label: "Reliability Test - qwen3-omni"
-    timeout_in_minutes: 180
+  - label: "Simple · CPU Coverage Test"
+    key: weekly-simple-cpu-coverage
     depends_on: upload-weekly-pipeline
-    if: build.env("WEEKLY") == "1" || build.pull_request.labels includes "weekly-test"
+    if: build.env("WEEKLY") == "1"
+    timeout_in_minutes: 60
     commands:
-      - pytest -s -v tests/dfx/reliability/test_reliability_qwen3_omni.py -m "slow"
-    mirror_hardwares: h100_2
-
-  - label: "Reliability Test - wan22"
-    timeout_in_minutes: 180
-    depends_on: upload-weekly-pipeline
-    if: build.env("WEEKLY") == "1" || build.pull_request.labels includes "weekly-test"
-    commands:
-      - pytest -s -v tests/dfx/reliability/test_reliability_wan22.py -m "slow"
-    mirror_hardwares: h100_2
-
-  - label: "Reliability Test - hunyuan-image"
-    timeout_in_minutes: 180
-    depends_on: upload-weekly-pipeline
-    if: build.env("WEEKLY") == "1" || build.pull_request.labels includes "weekly-test"
-    commands:
-      - pytest -s -v tests/dfx/reliability/test_reliability_hunyuan_image.py -m "slow"
-    mirror_hardwares: h100_4
-
-  - label: "Reliability Test - voxcpm2"
-    timeout_in_minutes: 180
-    depends_on: upload-weekly-pipeline
-    if: build.env("WEEKLY") == "1" || build.pull_request.labels includes "weekly-test"
-    commands:
-      - pytest -s -v tests/dfx/reliability/test_reliability_voxcpm2.py -m "slow"
+      - |
+        set +e
+        REPORT="coverage-core-model-cpu-$${BUILDKITE_STEP_ID:-local}.xml"
+        pytest -sv tests/ -m 'core_model and cpu' --cov=vllm_omni --cov-report=term-missing:skip-covered --cov-report=xml:$${REPORT}
+        EXIT=$$?
+        gzip -9 -f "$${REPORT}"
+        buildkite-agent artifact upload "$${REPORT}.gz"
+        exit $$EXIT
     mirror_hardwares: l4_1
 
-  - group: "Reliability Test - Invalid parameters Test"
+  - group: "Reliability Test"
     depends_on: upload-weekly-pipeline
     if: build.env("WEEKLY") == "1" || build.pull_request.labels includes "weekly-test"
     steps:
 
-      - label: "Invalid parameters Test · H100"
-        timeout_in_minutes: 120
+      - label: "Reliability Test · qwen3-omni"
+        source_file_dependencies: omni_qwen3_omni_reliability
+        timeout_in_minutes: 180
         commands:
-          - pytest -s -v tests/dfx/reliability/invalid_param_test/ -m "slow and H100"
+          - pytest -s -v tests/dfx/reliability/test_reliability_qwen3_omni.py -m "slow"
         mirror_hardwares: h100_2
 
-      - label: "Invalid parameters Test · L4"
+      - label: "Reliability Test · wan22"
+        source_file_dependencies: diffusion_wan22_reliability
+        timeout_in_minutes: 180
+        commands:
+          - pytest -s -v tests/dfx/reliability/test_reliability_wan22.py -m "slow"
+        mirror_hardwares: h100_2
+
+      - label: "Reliability Test · hunyuan-image"
+        source_file_dependencies: diffusion_hunyuan_image3_reliability
+        timeout_in_minutes: 180
+        commands:
+          - pytest -s -v tests/dfx/reliability/test_reliability_hunyuan_image.py -m "slow"
+        mirror_hardwares: h100_4
+
+      - label: "Reliability Test · Invalid parameters · H100 · Single-GPU"
+        source_file_dependencies: reliability_invalid_param_h100
+        timeout_in_minutes: 90
+        commands:
+          - pytest -s -v tests/dfx/reliability/invalid_param_test/ -m "slow and H100 and cards_1"
+        mirror_hardwares: h100_1
+
+      - label: "Reliability Test · Invalid parameters · H100 · 2-GPU"
+        source_file_dependencies: reliability_invalid_param_h100
+        timeout_in_minutes: 90
+        commands:
+          - pytest -s -v tests/dfx/reliability/invalid_param_test/ -m "slow and H100 and cards_2"
+        mirror_hardwares: h100_2
+
+      - label: "Reliability Test · Invalid parameters · L4"
+        source_file_dependencies: reliability_invalid_param_l4
         timeout_in_minutes: 120
         commands:
           - pytest -s -v tests/dfx/reliability/invalid_param_test/ -m "slow and L4"
         mirror_hardwares: l4_1
 
-  - group: ":card_index_dividers: E2E Tests"
+  - group: "Perf Test"
     depends_on: upload-weekly-pipeline
     if: build.env("WEEKLY") == "1" || build.pull_request.labels includes "weekly-test"
     steps:
 
-      - label: "Omni · L4"
+      - label: ":full_moon: Omni · Perf Test · vLLM Text"
+        source_file_dependencies: omni_qwen3_omni_perf
+        key: weekly-omni-performance-vllm-text
+        timeout_in_minutes: 180
+        commands:
+          - export BENCHMARK_DIR=tests/dfx/perf/results
+          - |
+            set +e
+            pytest -s -v tests/dfx/perf/scripts/run_benchmark.py --test-config-file tests/dfx/perf/tests/test_qwen3_omni_vllm_text.json
+            EXIT=$$?
+            buildkite-agent artifact upload "tests/dfx/perf/results/*.json"
+            exit $$EXIT
+        mirror_hardwares: h100_2
+
+      - label: ":full_moon: Omni · Perf Test · Async Chunk · Random"
+        key: weekly-omni-performance-async-chunk-random
+        timeout_in_minutes: 180
+        commands:
+          - export BENCHMARK_DIR=tests/dfx/perf/results
+          - |
+            set +e
+            pytest -s -v tests/dfx/perf/scripts/run_benchmark.py --test-config-file tests/dfx/perf/tests/test_qwen3_omni_async_chunk.json -m "H100 and slow"
+            EXIT=$$?
+            buildkite-agent artifact upload "tests/dfx/perf/results/*.json"
+            exit $$EXIT
+        mirror_hardwares: h100_2
+
+      - label: ":full_moon: Omni · Perf Test · Multi-Replica"
+        source_file_dependencies: omni_qwen3_omni_perf
+        key: weekly-omni-performance-multi-replicas
+        timeout_in_minutes: 180
+        commands:
+          - export BENCHMARK_DIR=tests/dfx/perf/results
+          - |
+            set +e
+            pytest -s -v tests/dfx/perf/scripts/run_benchmark.py --test-config-file tests/dfx/perf/tests/test_qwen3_omni_multi_replicas.json
+            EXIT=$$?
+            buildkite-agent artifact upload "tests/dfx/perf/results/*.json"
+            exit $$EXIT
+        mirror_hardwares: h100_3
+
+  - group: ":card_index_dividers: E2E Tests"
+    depends_on: upload-weekly-pipeline
+    if: build.env("NON_CRITICAL") == "1"
+    steps:
+
+      # Online Qwen2.5 expansion is 2-card; offline expansion + AutoRound are 4-card.
+      - label: "Omni · L4 · 2-GPU"
         timeout_in_minutes: 90
         commands:
           - export VLLM_ALLOW_LONG_MAX_MODEL_LEN="1"
-          - pytest -s -v tests/e2e/*/test_qwen2_5_omni_expansion.py --run-level "core_model"
+          - pytest -s -v tests/e2e/online_serving/test_qwen2_5_omni_expansion.py -m "slow and L4 and omni and cards_2" --run-level "core_model"
+        mirror_hardwares: l4_2
+
+      - label: "Omni · L4 · 4-GPU"
+        timeout_in_minutes: 90
+        commands:
+          - export VLLM_ALLOW_LONG_MAX_MODEL_LEN="1"
+          - pytest -s -v tests/e2e/offline_inference/test_qwen2_5_omni_expansion.py -m "slow and L4 and omni and cards_4" --run-level "core_model"
           - >-
             pytest -s -v tests/e2e/
             --ignore=tests/e2e/online_serving/test_qwen2_5_omni_expansion.py
             --ignore=tests/e2e/offline_inference/test_qwen2_5_omni_expansion.py
-            -m "slow and L4 and omni"
+            -m "slow and L4 and omni and cards_4"
             --run-level "full_model"
         mirror_hardwares: l4_4
 
-      - label: "Omni · H100"
+      - label: "Omni · H100 · Single-GPU"
         timeout_in_minutes: 90
         commands:
           - >-
             pytest -sv
-            tests/e2e/ -m "slow and H100 and omni"
+            tests/e2e/ -m "slow and H100 and omni and cards_1"
             --run-level "full_model"
-        mirror_hardwares: h100_2
+        mirror_hardwares: h100_1
+
+      - label: "Omni · H100 · 4-GPU"
+        timeout_in_minutes: 90
+        commands:
+          - >-
+            pytest -sv
+            tests/e2e/ -m "slow and H100 and omni and cards_4"
+            --run-level "full_model"
+        mirror_hardwares: h100_4
 
       - label: "TTS · L4"
-        timeout_in_minutes: 90
+        timeout_in_minutes: 120
         commands:
           - export VLLM_ALLOW_LONG_MAX_MODEL_LEN="1"
           - export VLLM_USE_DEEP_GEMM="0"
@@ -108,29 +186,355 @@ steps:
             --run-level "full_model"
         mirror_hardwares: h100_1
 
-      - label: "Diffusion · L4"
+      # Own step: the image has no NanoCodec, so this file import-skips out of
+      # the mark-selected TTS steps above. The overrides keep nemo_toolkit[tts]
+      # from downgrading transformers; see
+      # examples/offline_inference/text_to_speech/README.md.
+      - label: "TTS · Gepard-1.0 · L4"
+        timeout_in_minutes: 90
+        commands:
+          - export VLLM_ALLOW_LONG_MAX_MODEL_LEN="1"
+          - printf '%s\n' "transformers>=5.5.3" "huggingface-hub>=1.0" > /tmp/gepard-overrides.txt
+          - uv pip install --python "$(python3 -c 'import sys; print(sys.executable)')" "nemo_toolkit[tts]==2.7.3" --overrides /tmp/gepard-overrides.txt
+          - >-
+            pytest -s -v tests/e2e/offline_inference/test_gepard_tts.py
+            tests/e2e/online_serving/test_gepard_tts.py
+            -m "tts and cuda"
+            --run-level "full_model"
+        mirror_hardwares: l4_1
+
+      - label: "Diffusion · L4 · Single-GPU"
         timeout_in_minutes: 60
         commands:
           - >-
             pytest -sv
-            tests/e2e/ -m "slow and diffusion and L4"
+            tests/e2e/ -m "slow and diffusion and L4 and cards_1"
+            --run-level "full_model"
+        mirror_hardwares: l4_1
+
+      - label: "Diffusion · L4 · 2-GPU"
+        timeout_in_minutes: 60
+        commands:
+          - >-
+            pytest -sv
+            tests/e2e/ -m "slow and diffusion and L4 and cards_2"
+            --run-level "full_model"
+        mirror_hardwares: l4_2
+
+      - label: "Diffusion · L4 · 4-GPU"
+        timeout_in_minutes: 60
+        commands:
+          - >-
+            pytest -sv
+            tests/e2e/ -m "slow and diffusion and L4 and cards_4"
             --run-level "full_model"
         mirror_hardwares: l4_4
 
-      - label: "Diffusion · H100 · Single-GPU"
+      # Split the old shared H100 Single-GPU sweep: one pytest process collected
+      # every slow+cards_1 diffusion file and exceeded 150 minutes.
+      - label: "Diffusion · Qwen-Image · H100 · Single-GPU"
         timeout_in_minutes: 120
         commands:
           - >-
             pytest -sv
-            tests/e2e/ -m "slow and diffusion and H100 and not distributed_cuda"
+            tests/e2e/online_serving/test_qwen_image_edit.py
+            tests/e2e/online_serving/test_qwen_image_edit_expansion.py
+            tests/e2e/online_serving/test_qwen_image_layered.py
+            tests/e2e/online_serving/test_qwen_image_layered_expansion.py
+            tests/e2e/accuracy/test_qwen_image_edit.py
+            tests/e2e/accuracy/test_qwen_image_layered.py
+            -m "slow and diffusion and H100 and cards_1"
             --run-level "full_model"
         mirror_hardwares: h100_1
 
-      - label: "Diffusion · H100 · Multi-GPU"
+      - label: "Diffusion · Wan2.1 VACE · H100 · Single-GPU"
         timeout_in_minutes: 90
         commands:
           - >-
             pytest -sv
-            tests/e2e/ -m "slow and diffusion and H100 and distributed_cuda"
+            tests/e2e/online_serving/test_wan_2_1_vace_expansion.py
+            -m "slow and diffusion and H100 and cards_1"
+            --run-level "full_model"
+        mirror_hardwares: h100_1
+
+      - label: "Diffusion · LongCat · H100 · Single-GPU"
+        timeout_in_minutes: 90
+        commands:
+          - >-
+            pytest -sv
+            tests/e2e/online_serving/test_longcat_image_expansion.py
+            tests/e2e/online_serving/test_longcat_image_edit_expansion.py
+            -m "slow and diffusion and H100 and cards_1"
+            --run-level "full_model"
+        mirror_hardwares: h100_1
+
+      - label: "Diffusion · Bagel · H100 · Single-GPU"
+        timeout_in_minutes: 90
+        commands:
+          - >-
+            pytest -sv
+            tests/e2e/online_serving/test_bagel.py
+            tests/e2e/online_serving/test_bagel_expansion.py
+            tests/e2e/offline_inference/test_bagel.py
+            tests/e2e/offline_inference/test_bagel_expansion.py
+            -m "slow and diffusion and H100 and cards_1"
+            --run-level "full_model"
+        mirror_hardwares: h100_1
+
+      - label: "Diffusion · Boogu · H100 · Single-GPU"
+        timeout_in_minutes: 90
+        commands:
+          - >-
+            pytest -sv
+            tests/e2e/online_serving/test_boogu_image.py
+            tests/e2e/online_serving/test_boogu_image_edit.py
+            tests/e2e/online_serving/test_boogu_image_expansion.py
+            tests/e2e/online_serving/test_boogu_image_edit_expansion.py
+            -m "slow and diffusion and H100 and cards_1"
+            --run-level "full_model"
+        mirror_hardwares: h100_1
+
+      - label: "Diffusion · Flux · H100 · Single-GPU"
+        timeout_in_minutes: 90
+        commands:
+          - >-
+            pytest -sv
+            tests/e2e/online_serving/test_flux_2_dev_expansion.py
+            tests/e2e/offline_inference/test_flux1_schnell_expansion.py
+            -m "slow and diffusion and H100 and cards_1"
+            --run-level "full_model"
+        mirror_hardwares: h100_1
+
+      - label: "Diffusion · LTX · H100 · Single-GPU"
+        timeout_in_minutes: 120
+        commands:
+          - >-
+            pytest -sv
+            tests/e2e/online_serving/test_ltx.py
+            tests/e2e/online_serving/test_ltx25.py
+            tests/e2e/accuracy/ltx/test_ltx_official_similarity.py
+            tests/e2e/accuracy/ltx/test_ltx25_official_similarity.py
+            -m "slow and diffusion and H100 and cards_1"
+            --run-level "full_model"
+        mirror_hardwares: h100_1
+
+      - label: "Diffusion · Krea2 · H100 · Single-GPU"
+        timeout_in_minutes: 90
+        commands:
+          - >-
+            pytest -sv
+            tests/e2e/online_serving/test_krea2_expansion.py
+            tests/e2e/offline_inference/test_krea2_expansion.py
+            -m "slow and diffusion and H100 and cards_1"
+            --run-level "full_model"
+        mirror_hardwares: h100_1
+
+      - label: "Diffusion · HiDream-I1 · H100 · Single-GPU"
+        timeout_in_minutes: 90
+        commands:
+          - >-
+            pytest -sv
+            tests/e2e/online_serving/test_hidream_i1_full_expansion.py
+            -m "slow and diffusion and H100 and cards_1"
+            --run-level "full_model"
+        mirror_hardwares: h100_1
+
+      - label: "Diffusion · SANA-Video · H100 · Single-GPU"
+        source_file_dependencies: diffusion_sana_video_function
+        timeout_in_minutes: 180
+        commands:
+          - export SANA_VIDEO_GOLDEN_BASE_URL=https://vllm-public-assets.s3.us-west-2.amazonaws.com/omni-assets/sana-video/v2
+          - >-
+            pytest -sv
+            tests/e2e/online_serving/test_sana_video_expansion.py
+            tests/e2e/accuracy/sana_video/
+            -m "slow and diffusion and H100 and cards_1"
+            --run-level "full_model"
+        mirror_hardwares: h100_1
+
+      - label: "Diffusion · SenseNova-U1 · H100 · Single-GPU"
+        timeout_in_minutes: 90
+        commands:
+          - >-
+            pytest -sv
+            tests/e2e/offline_inference/test_sensenova_u1_expansion.py
+            tests/e2e/offline_inference/test_sensenova_u1_text2img_expansion.py
+            tests/e2e/offline_inference/test_sensenova_u1_img2img_expansion.py
+            -m "slow and diffusion and H100 and cards_1"
+            --run-level "full_model"
+        mirror_hardwares: h100_1
+
+      - label: "Diffusion · MammothModa2 · H100 · Single-GPU"
+        timeout_in_minutes: 90
+        commands:
+          - >-
+            pytest -sv
+            tests/e2e/offline_inference/test_mammoth_moda2_expansion.py
+            -m "slow and diffusion and H100 and cards_1"
+            --run-level "full_model"
+        mirror_hardwares: h100_1
+
+      - label: "Diffusion · Helios · H100 · Single-GPU"
+        timeout_in_minutes: 90
+        commands:
+          - >-
+            pytest -sv
+            tests/e2e/accuracy/test_video_streaming_output_similarity.py
+            -m "slow and diffusion and H100 and cards_1"
+            --run-level "full_model"
+        mirror_hardwares: h100_1
+
+      # Split the old shared H100 2-GPU sweep: one pytest process collected every
+      # slow+cards_2 diffusion file and exceeded 90 minutes.
+      - label: "Diffusion · Qwen-Image · H100 · 2-GPU"
+        timeout_in_minutes: 90
+        commands:
+          - >-
+            pytest -sv
+            tests/e2e/online_serving/test_qwen_image_edit_expansion.py
+            tests/e2e/online_serving/test_qwen_image_layered_expansion.py
+            -m "slow and diffusion and H100 and cards_2"
             --run-level "full_model"
         mirror_hardwares: h100_2
+
+      - label: "Diffusion · Wan2.1 VACE · H100 · 2-GPU"
+        timeout_in_minutes: 90
+        commands:
+          - >-
+            pytest -sv
+            tests/e2e/online_serving/test_wan_2_1_vace_expansion.py
+            -m "slow and diffusion and H100 and cards_2"
+            --run-level "full_model"
+        mirror_hardwares: h100_2
+
+      - label: "Diffusion · LongCat · H100 · 2-GPU"
+        timeout_in_minutes: 90
+        commands:
+          - >-
+            pytest -sv
+            tests/e2e/online_serving/test_longcat_image_expansion.py
+            tests/e2e/online_serving/test_longcat_image_edit_expansion.py
+            -m "slow and diffusion and H100 and cards_2"
+            --run-level "full_model"
+        mirror_hardwares: h100_2
+
+      - label: "Diffusion · Bagel · H100 · 2-GPU"
+        timeout_in_minutes: 90
+        commands:
+          - >-
+            pytest -sv
+            tests/e2e/online_serving/test_bagel_expansion.py
+            -m "slow and diffusion and H100 and cards_2"
+            --run-level "full_model"
+        mirror_hardwares: h100_2
+
+      - label: "Diffusion · Boogu · H100 · 2-GPU"
+        timeout_in_minutes: 90
+        commands:
+          - >-
+            pytest -sv
+            tests/e2e/online_serving/test_boogu_image_expansion.py
+            tests/e2e/online_serving/test_boogu_image_edit_expansion.py
+            -m "slow and diffusion and H100 and cards_2"
+            --run-level "full_model"
+        mirror_hardwares: h100_2
+
+      - label: "Diffusion · Flux · H100 · 2-GPU"
+        timeout_in_minutes: 90
+        commands:
+          - >-
+            pytest -sv
+            tests/e2e/online_serving/test_flux_2_dev_expansion.py
+            tests/e2e/offline_inference/test_flux1_kontext_expansion.py
+            -m "slow and diffusion and H100 and cards_2"
+            --run-level "full_model"
+        mirror_hardwares: h100_2
+
+      - label: "Diffusion · GLM-Image · H100 · 2-GPU"
+        timeout_in_minutes: 90
+        commands:
+          - >-
+            pytest -sv
+            tests/e2e/online_serving/test_glm_image_expansion.py
+            tests/e2e/offline_inference/test_glm_image_autoround_w4a16_expansion.py
+            -m "slow and diffusion and H100 and cards_2"
+            --run-level "full_model"
+        mirror_hardwares: h100_2
+
+      - label: "Diffusion · LTX · H100 · 2-GPU"
+        timeout_in_minutes: 90
+        commands:
+          - >-
+            pytest -sv
+            tests/e2e/online_serving/test_ltx.py
+            -m "slow and diffusion and H100 and cards_2"
+            --run-level "full_model"
+        mirror_hardwares: h100_2
+
+      - label: "Diffusion · Krea2 · H100 · 2-GPU"
+        timeout_in_minutes: 90
+        commands:
+          - >-
+            pytest -sv
+            tests/e2e/offline_inference/test_krea2_expansion.py
+            -m "slow and diffusion and H100 and cards_2"
+            --run-level "full_model"
+        mirror_hardwares: h100_2
+
+      - label: "Diffusion · DreamZero · H100 · 2-GPU"
+        timeout_in_minutes: 90
+        commands:
+          - >-
+            pytest -sv
+            tests/e2e/online_serving/test_dreamzero_expansion.py
+            tests/e2e/accuracy/test_dreamzero.py
+            -m "slow and diffusion and H100 and cards_2"
+            --run-level "full_model"
+        mirror_hardwares: h100_2
+
+      - label: "Diffusion · H100 · 3-GPU"
+        timeout_in_minutes: 120
+        commands:
+          - >-
+            pytest -sv
+            tests/e2e/ -m "slow and diffusion and H100 and cards_3"
+            --run-level "full_model"
+        mirror_hardwares: h100_3
+
+      - label: "Diffusion · H100 · 4-GPU"
+        timeout_in_minutes: 90
+        commands:
+          - >-
+            pytest -sv
+            tests/e2e/ -m "slow and diffusion and H100 and cards_4"
+            --run-level "full_model"
+            --ignore=tests/e2e/online_serving/test_magi2.py
+        mirror_hardwares: h100_4
+
+      # Dedicated: HiDream-O1 accuracy needs a long TP2 run; the shared H100
+      # 2-GPU sweep is only 90 minutes.
+      - label: "Diffusion · HiDream-O1-Image Accuracy · H100 · 2-GPU"
+        timeout_in_minutes: 180
+        commands:
+          - pytest -s -v tests/e2e/accuracy/test_hidream_o1_image.py -m "slow and diffusion and cuda and H100 and cards_2" --run-level "full_model"
+        mirror_hardwares: h100_2
+
+      # Dedicated: Boogu CFG-parallel size 3 needs 3 GPUs; weekly has no H100 3-GPU sweep.
+      - label: "Diffusion · Boogu Edit CFG · H100 · 3-GPU"
+        timeout_in_minutes: 120
+        commands:
+          - >-
+            pytest -sv
+            tests/e2e/online_serving/test_boogu_image_edit_expansion.py
+            -k "test_double_guidance_cfg_parallel and cfg_parallel_3_full"
+            -m "slow and diffusion and cuda and H100 and cards_3"
+            --run-level "full_model"
+        mirror_hardwares: h100_3
+
+      # Dedicated: MAGI-2 Preview serving smoke needs ~30 min server init; the
+      # shared H100 4-GPU sweep is only 90 minutes.
+      - label: "Diffusion · MAGI-2 Preview Serving Smoke · H100 · 4-GPU"
+        timeout_in_minutes: 180
+        commands:
+          - pytest -s -v tests/e2e/online_serving/test_magi2.py -m "slow and diffusion and cuda and H100 and cards_4" --run-level "full_model"
+        mirror_hardwares: h100_4

.buildkite/intel/pipeline-intel.yml

@@ -10,7 +10,7 @@ steps:
           DOCKER_BUILDKIT: "1"
           # Buildkite will automatically replace this with the actual commit hash
           VLLM_IMAGE_TAG: "${BUILDKITE_COMMIT}"
-          VLLM_VERSION: "v0.26.0"
+          VLLM_VERSION: "v0.29.0"
         priority: 100
         timeout_in_minutes: 60
         soft_fail: false

.buildkite/intel/scripts/run-xpu-test.sh

@@ -5,7 +5,12 @@ set -ex
 
 omni_source_dir=$(git rev-parse --show-toplevel)
 
-base_image_name="xpu/vllm-omni-ci-base:${VLLM_VERSION:?VLLM_VERSION must be set}"
+: "${VLLM_VERSION:?VLLM_VERSION must be set}"
+# Official upstream XPU image. It is published from vllm's own
+# docker/Dockerfile.xpu, which is exactly what build_vllm_base() clones and
+# builds below, so pulling it is equivalent to building the fallback base.
+upstream_base_image="vllm/vllm-openai-xpu:${VLLM_VERSION}"
+local_base_image="xpu/vllm-omni-ci-base:${VLLM_VERSION}"
 image_name="xpu/vllm-omni-ci:${BUILDKITE_COMMIT:?BUILDKITE_COMMIT must be set}"
 container_name="xpu_${BUILDKITE_COMMIT}_$(
     tr -dc A-Za-z0-9 </dev/urandom | head -c 10
@@ -34,9 +39,51 @@ docker_build() {
     docker build "${out[@]}" "$@" -f docker/Dockerfile.xpu .
 }
 
-if [ -z "$(docker images -q "${base_image_name}")" ]; then
-    docker_build "${base_image_name}" --target vllm-base --build-arg "VLLM_VERSION=${VLLM_VERSION}"
-fi
+build_vllm_base() (
+    local vllm_source_dir
+    vllm_source_dir=$(mktemp -d)
+    trap 'rm -rf "${vllm_source_dir}"' EXIT
+
+    git clone --depth 1 --branch "${VLLM_VERSION}" \
+        https://github.com/vllm-project/vllm "${vllm_source_dir}"
+
+    local out=("${export_args[@]/'{{IMAGE}}'/${base_image_name}}")
+    docker build "${out[@]}" \
+        --target vllm-openai \
+        -f "${vllm_source_dir}/docker/Dockerfile.xpu" \
+        "${vllm_source_dir}"
+)
+
+# Resolve the vLLM base image, in priority order:
+#   1. the published upstream image for VLLM_VERSION, pulled fresh (so moving
+#      tags like nightly are refreshed rather than served stale from disk);
+#   2. a local copy of that same image, when the registry is unreachable;
+#   3. VLLM_BASE, if that image is already on disk;
+#   4. otherwise build VLLM_BASE from the matching vLLM source tag (~30min).
+# VLLM_BASE names the fallback base and defaults to ${local_base_image}; set it
+# to reuse or produce a base under a different tag.
+resolve_vllm_base() {
+    if docker pull "${upstream_base_image}"; then
+        base_image_name="${upstream_base_image}"
+        return
+    fi
+
+    # Registry unreachable, but a previous run may have left the image behind.
+    if [ -n "$(docker images -q "${upstream_base_image}")" ]; then
+        echo "WARNING: could not pull ${upstream_base_image}; using local copy" >&2
+        base_image_name="${upstream_base_image}"
+        return
+    fi
+
+    echo "WARNING: ${upstream_base_image} unavailable remotely and locally" >&2
+    base_image_name="${VLLM_BASE:-${local_base_image}}"
+    if [ -z "$(docker images -q "${base_image_name}")" ]; then
+        echo "WARNING: building ${base_image_name} from vLLM ${VLLM_VERSION} source" >&2
+        build_vllm_base
+    fi
+}
+
+resolve_vllm_base
 
 # Try building the docker image
 docker_build "${image_name}" --build-arg "VLLM_BASE=${base_image_name}" --build-arg "VLLM_VERSION=${VLLM_VERSION}"
@@ -72,8 +119,9 @@ time timeout -k 30 30m docker run \
     echo $ZE_AFFINITY_MASK
     pip install tblib==3.1.0
     cd /workspace/vllm-omni
-    pytest -v -s -m "core_model and xpu and B60"
+    XPU_TEST_PATHS="tests/diffusion tests/dfx tests/e2e"
+    pytest -v -s $XPU_TEST_PATHS -m "core_model and xpu and B60"
     pytest -v -s tests/diffusion/quantization/test_mxfp8_config.py
-    pytest -v -s -m "advanced_model and xpu and B60"
-    pytest -v -s -m "omni and xpu and B60"
+    pytest -v -s $XPU_TEST_PATHS -m "advanced_model and xpu and B60"
+    pytest -v -s $XPU_TEST_PATHS -m "omni and xpu and B60"
 '

.buildkite/npu/bootstrap-upload-steps.yml

@@ -3,10 +3,10 @@
 steps:
   - label: ":buildkit: Build and Push NPU Test Image - A2"
     key: image-build-a2
-    agents:
+    agents: &npu-image-build-agents
       queue: "ascend-a2b3"
       resource_class: "npu-1"
-    plugins:
+    plugins: &npu-image-build-plugins
       - kubernetes:
           metadata:
             annotations:
@@ -46,136 +46,67 @@ steps:
       IMAGE_NAME: "vllm-omni-ci-npu"
       IMAGE_REGISTRY: "swr.cn-southwest-2.myhuaweicloud.com/modelfoundry"
       BUILDKITD_ADDR: "tcp://buildkitd-service.buildkitd:1234"
-    command: |
-      set -ex
-
-      echo "--- Building and pushing NPU Test Image"
-      echo "Image: $${IMAGE_REGISTRY}/$${IMAGE_NAME}:$${VLLM_IMAGE_TAG}"
-      echo "buildkitd address: $${BUILDKITD_ADDR}"
-
-      if ! command -v buildctl &> /dev/null; then
-        echo "Installing buildctl..."
-        mkdir -p /tmp/buildkit
-        BUILDKIT_VERSION="v0.29.0"
-        wget -q "https://github.com/moby/buildkit/releases/download/$${BUILDKIT_VERSION}/buildkit-$${BUILDKIT_VERSION}.linux-arm64.tar.gz" -O /tmp/buildkit.tar.gz
-        tar -xzf /tmp/buildkit.tar.gz -C /tmp/buildkit
-        cp /tmp/buildkit/bin/buildctl /usr/local/bin/
-      fi
-
-      export DOCKER_CONFIG=/home/user/.docker
-      buildctl \
-        --addr="$${BUILDKITD_ADDR}" \
-        --tlscacert=/home/user/.docker/ca.pem \
-        --tlscert=/home/user/.docker/cert.pem \
-        --tlskey=/home/user/.docker/key.pem \
-        build \
-        --frontend dockerfile.v0 \
-        --local context=. \
-        --local dockerfile=./docker \
-        --opt filename=Dockerfile.npu.ci \
-        --secret id=dockerconfig,src=/home/user/.docker/config.json \
-        --output type=image,name=$${IMAGE_REGISTRY}/$${IMAGE_NAME}:$${VLLM_IMAGE_TAG},push=true \
-        --progress=plain
-
-      echo "--- Image pushed successfully"
-      echo "$${IMAGE_REGISTRY}/$${IMAGE_NAME}:$${VLLM_IMAGE_TAG}"
+    command: bash .buildkite/npu/scripts/build-npu-image.sh Dockerfile.npu.ci "$${VLLM_IMAGE_TAG}"
 
   - label: ":buildkit: Build and Push NPU Test Image - A3"
     key: image-build-a3
-    agents:
-      queue: "ascend-a2b3"
-      resource_class: "npu-1"
-    plugins:
-      - kubernetes:
-          metadata:
-            annotations:
-              vault.hashicorp.com/agent-init-first: "true"
-              vault.hashicorp.com/agent-inject: "true"
-              vault.hashicorp.com/agent-inject-perms-ca.pem: "0400"
-              vault.hashicorp.com/agent-inject-perms-cert.pem: "0400"
-              vault.hashicorp.com/agent-inject-perms-config.json: "0400"
-              vault.hashicorp.com/agent-inject-perms-key.pem: "0400"
-              vault.hashicorp.com/agent-inject-secret-ca.pem: internal/data/ascend/buildkitd
-              vault.hashicorp.com/agent-inject-secret-cert.pem: internal/data/ascend/buildkitd
-              vault.hashicorp.com/agent-inject-secret-config.json: internal/data/ascend/buildkitd
-              vault.hashicorp.com/agent-inject-secret-key.pem: internal/data/ascend/buildkitd
-              vault.hashicorp.com/agent-inject-template-ca.pem: "{{- with secret \"internal/data/ascend/buildkitd\" -}}\n{{ .Data.data.RootCA }}\n{{- end }}"
-              vault.hashicorp.com/agent-inject-template-cert.pem: "{{- with secret \"internal/data/ascend/buildkitd\" -}}\n{{ .Data.data.ClientCaCert }}\n{{- end }}"
-              vault.hashicorp.com/agent-inject-template-config.json: "{{- with secret \"internal/data/ascend/buildkitd\" -}}\n{{ .Data.data.dockerConfig }}\n{{- end }}"
-              vault.hashicorp.com/agent-inject-template-key.pem: "{{- with secret \"internal/data/ascend/buildkitd\" -}}\n{{ .Data.data.ClientCaKey }}\n{{- end }}"
-              vault.hashicorp.com/agent-pre-populate-only: "true"
-              vault.hashicorp.com/agent-run-as-group: "1000"
-              vault.hashicorp.com/agent-run-as-user: "1000"
-              vault.hashicorp.com/agent-service-account-token-volume-name: token-vol
-              vault.hashicorp.com/role: ascend-gha-runners
-              vault.hashicorp.com/secret-volume-path: /home/user/.docker/
-              vault.hashicorp.com/tls-skip-verify: "true"
-          podSpecPatch:
-            volumes:
-            - name: token-vol
-              projected:
-                defaultMode: 420
-                sources:
-                - serviceAccountToken:
-                    audience: api
-                    expirationSeconds: 600
-                    path: token
+    agents: *npu-image-build-agents
+    plugins: *npu-image-build-plugins
     env:
       VLLM_IMAGE_TAG: "${BUILDKITE_COMMIT}-a3"
       IMAGE_NAME: "vllm-omni-ci-npu"
       IMAGE_REGISTRY: "swr.cn-southwest-2.myhuaweicloud.com/modelfoundry"
       BUILDKITD_ADDR: "tcp://buildkitd-service.buildkitd:1234"
-    command: |
-      set -ex
+    command: bash .buildkite/npu/scripts/build-npu-image.sh Dockerfile.npu.ci "$${VLLM_IMAGE_TAG}" -a3
 
-      echo "--- Building and pushing NPU Test Image"
-      echo "Image: $${IMAGE_REGISTRY}/$${IMAGE_NAME}:$${VLLM_IMAGE_TAG}"
-      echo "buildkitd address: $${BUILDKITD_ADDR}"
+  - label: ":buildkit: Build and Push NPU Test Image - A5"
+    key: image-build-a5
+    agents: *npu-image-build-agents
+    plugins: *npu-image-build-plugins
+    env:
+      VLLM_IMAGE_TAG: "${BUILDKITE_COMMIT}-a5"
+      IMAGE_NAME: "vllm-omni-ci-npu"
+      IMAGE_REGISTRY: "swr.cn-southwest-2.myhuaweicloud.com/modelfoundry"
+      BUILDKITD_ADDR: "tcp://buildkitd-service.buildkitd:1234"
+    command: bash .buildkite/npu/scripts/build-npu-image.sh Dockerfile.npu.ci "$${VLLM_IMAGE_TAG}" -a5
 
-      if ! command -v buildctl &> /dev/null; then
-        echo "Installing buildctl..."
-        mkdir -p /tmp/buildkit
-        BUILDKIT_VERSION="v0.29.0"
-        wget -q "https://github.com/moby/buildkit/releases/download/$${BUILDKIT_VERSION}/buildkit-$${BUILDKIT_VERSION}.linux-arm64.tar.gz" -O /tmp/buildkit.tar.gz
-        tar -xzf /tmp/buildkit.tar.gz -C /tmp/buildkit
-        cp /tmp/buildkit/bin/buildctl /usr/local/bin/
-      fi
-
-      export DOCKER_CONFIG=/home/user/.docker
-      buildctl \
-        --addr="$${BUILDKITD_ADDR}" \
-        --tlscacert=/home/user/.docker/ca.pem \
-        --tlscert=/home/user/.docker/cert.pem \
-        --tlskey=/home/user/.docker/key.pem \
-        build \
-        --frontend dockerfile.v0 \
-        --local context=. \
-        --local dockerfile=./docker \
-        --opt filename=Dockerfile.npu.ci.a3 \
-        --secret id=dockerconfig,src=/home/user/.docker/config.json \
-        --output type=image,name=$${IMAGE_REGISTRY}/$${IMAGE_NAME}:$${VLLM_IMAGE_TAG},push=true \
-        --progress=plain
-
-      echo "--- Image pushed successfully"
-      echo "$${IMAGE_REGISTRY}/$${IMAGE_NAME}:$${VLLM_IMAGE_TAG}"
+  - label: ":buildkit: Build and Push NPU Test Image - 310P"
+    key: image-build-310p
+    agents: *npu-image-build-agents
+    plugins: *npu-image-build-plugins
+    env:
+      VLLM_IMAGE_TAG: "${BUILDKITE_COMMIT}-310p"
+      IMAGE_NAME: "vllm-omni-ci-npu"
+      IMAGE_REGISTRY: "swr.cn-southwest-2.myhuaweicloud.com/modelfoundry"
+      BUILDKITD_ADDR: "tcp://buildkitd-service.buildkitd:1234"
+    command: bash .buildkite/npu/scripts/build-npu-image.sh Dockerfile.npu.ci "$${VLLM_IMAGE_TAG}" -310p
 
   - label: "Upload Ready Pipeline"
     depends_on:
       - image-build-a2
       - image-build-a3
+      - image-build-a5
+      - image-build-310p
     key: upload-ready-pipeline
     agents:
       queue: "cpu_queue_premerge"
     command: |
       python3 .buildkite/common/scripts/upload_pipeline.py --upload .buildkite/npu/test-npu-ready.yml
 
-  # L4 Test — main+NIGHTLY=1 (scheduled), or PR with specific label (e.g. add label then Rebuild)
+  # L4 Test — main+NIGHTLY=1 (scheduled), or PR nightly-test label
   - label: "Upload Nightly Pipeline"
     depends_on:
       - image-build-a2
       - image-build-a3
+      - image-build-a5
+      - image-build-310p
     key: upload-nightly-pipeline
     commands:
-      - python3 .buildkite/common/scripts/upload_pipeline.py --upload .buildkite/npu/test-npu-nightly.yml
+      - |
+        EXTRA=""
+        if [ "${NIGHTLY:-}" = "1" ] && [ "${BUILDKITE_BRANCH}" = "main" ]; then
+          EXTRA="--all"
+        fi
+        python3 .buildkite/common/scripts/upload_pipeline.py --upload $$EXTRA .buildkite/npu/test-npu-nightly.yml
     agents:
       queue: "cpu_queue_premerge"

.buildkite/npu/pipeline-npu-a3.yml

@@ -96,13 +96,7 @@
   #   key: upload-nightly-pipeline
   #   if: >-
   #     (build.branch == "main" && build.env("NIGHTLY") == "1") ||
-  #     (build.branch != "main" && (
-  #       build.pull_request.labels includes "nightly-test" ||
-  #       build.pull_request.labels includes "omni-test" ||
-  #       build.pull_request.labels includes "tts-test" ||
-  #       build.pull_request.labels includes "diffusion-x2iat-test" ||
-  #       build.pull_request.labels includes "diffusion-x2v-test"
-  #     ))
+  #     (build.branch != "main" && build.pull_request.labels includes "nightly-test")
   #   commands:
   #     - buildkite-agent pipeline upload .buildkite/npu/test-npu-nightly.yml
   #   agents:

.buildkite/npu/scripts/build-npu-image.sh

@@ -0,0 +1,48 @@
+#!/usr/bin/env bash
+set -ex
+
+dockerfile=${1:?Dockerfile is required}
+image_tag=${2:?Image tag is required}
+device_suffix=${3:-}
+
+image_name=${IMAGE_NAME:-vllm-omni-ci-npu}
+image_registry=${IMAGE_REGISTRY:-swr.cn-southwest-2.myhuaweicloud.com/modelfoundry}
+buildkitd_addr=${BUILDKITD_ADDR:-tcp://buildkitd-service.buildkitd:1234}
+
+echo "--- Building and pushing NPU Test Image"
+echo "Image: ${image_registry}/${image_name}:${image_tag}"
+echo "Dockerfile: ${dockerfile}"
+
+if ! command -v buildctl &> /dev/null; then
+  mkdir -p /tmp/buildkit
+  buildkit_version=v0.29.0
+  wget -q \
+    "https://github.com/moby/buildkit/releases/download/${buildkit_version}/buildkit-${buildkit_version}.linux-arm64.tar.gz" \
+    -O /tmp/buildkit.tar.gz
+  tar -xzf /tmp/buildkit.tar.gz -C /tmp/buildkit
+  cp /tmp/buildkit/bin/buildctl /usr/local/bin/
+fi
+
+build_args=()
+if [[ -n "${device_suffix}" ]]; then
+  build_args+=(--opt "build-arg:VLLM_ASCEND_DEVICE_SUFFIX=${device_suffix}")
+fi
+
+export DOCKER_CONFIG=/home/user/.docker
+buildctl \
+  --addr="${buildkitd_addr}" \
+  --tlscacert=/home/user/.docker/ca.pem \
+  --tlscert=/home/user/.docker/cert.pem \
+  --tlskey=/home/user/.docker/key.pem \
+  build \
+  --frontend dockerfile.v0 \
+  --local context=. \
+  --local dockerfile=./docker \
+  --opt "filename=${dockerfile}" \
+  "${build_args[@]}" \
+  --secret id=dockerconfig,src=/home/user/.docker/config.json \
+  --output "type=image,name=${image_registry}/${image_name}:${image_tag},push=true" \
+  --progress=plain
+
+echo "--- Image pushed successfully"
+echo "${image_registry}/${image_name}:${image_tag}"

.buildkite/npu/test-npu-nightly.yml

@@ -4,10 +4,10 @@ steps:
     depends_on: upload-nightly-pipeline
     if: >-
       build.env("NIGHTLY") == "1" ||
-      build.pull_request.labels includes "nightly-test" ||
-      build.pull_request.labels includes "omni-test"
+      build.pull_request.labels includes "nightly-test"
     steps:
       - label: ":full_moon: Omni · Perf Test · No Async Chunk"
+        source_file_dependencies: omni_qwen3_omni_perf
         key: nightly-omni-performance-no-async-chunk
         timeout_in_minutes: 180
         mirror_hardwares: a3_npu_4
@@ -24,6 +24,7 @@ steps:
             buildkite-agent artifact upload "tests/dfx/perf/results/*.json"
             exit $$EXIT
       - label: ":full_moon: Omni · Perf Test · Async Chunk"
+        source_file_dependencies: omni_qwen3_omni_perf
         key: nightly-omni-performance-async-chunk
         timeout_in_minutes: 180
         mirror_hardwares: a3_npu_4
@@ -40,6 +41,7 @@ steps:
             buildkite-agent artifact upload "tests/dfx/perf/results/*.json"
             exit $$EXIT
       - label: ":full_moon: Omni · MiniCPM-o 4.5 Test"
+        source_file_dependencies: omni_minicpmo_4_5_function
         mirror_hardwares: a3_npu_2
         env:
           VLLM_WORKER_MULTIPROC_METHOD: spawn
@@ -48,6 +50,7 @@ steps:
         commands:
           - pytest -s -v tests/e2e/offline_inference/test_minicpmo_4_5.py tests/e2e/online_serving/test_minicpmo_4_5.py tests/e2e/online_serving/test_minicpmo_4_5_expansion.py -m 'full_model' --run-level 'full_model'
       - label: ":full_moon: Omni · MiniCPM-o 4.5 · Duplex Test"
+        source_file_dependencies: omni_minicpmo_4_5_duplex_function
         timeout_in_minutes: 180
         mirror_hardwares: a3_npu_2
         env:
@@ -62,6 +65,7 @@ steps:
             -m "npu and A3 and omni"
             --run-level full_model
       - label: ":full_moon: Omni · MiniCPM-o 4.5 · Accuracy Test"
+        source_file_dependencies: omni_minicpmo_4_5_accuracy
         timeout_in_minutes: 180
         mirror_hardwares: a3_npu_2
         env:
@@ -69,18 +73,77 @@ steps:
           VLLM_ALLOW_LONG_MAX_MODEL_LEN: "1"
           HF_TOKEN: "${HF_TOKEN}"
         commands:
+          # Pin the Seed-TTS scorer to the second card: _get_eval_device() otherwise
+          # auto-picks npu:0, which the omni server under test already fills with its
+          # KV cache, so Whisper OOMs on `.to(device)` and every clip reports
+          # "ASR / WER failed".
+          - export SEED_TTS_WER_EVAL=1
+          - export SEED_TTS_EVAL_DEVICE=npu:1
           - |
             set +e
             pytest -s -v tests/e2e/accuracy/minicpmo_4_5/test_minicpmo_4_5.py -m "full_model and npu and A3" --run-level full_model
             EXIT=$$?
             buildkite-agent artifact upload "tests/e2e/accuracy/minicpmo_4_5/results/*.json"
             exit $$EXIT
+      - label: ":full_moon: Omni · MiniCPM-o 4.5 · Perf Test"
+        source_file_dependencies: omni_minicpmo_4_5_perf
+        key: nightly-omni-performance-minicpmo-4-5-npu
+        timeout_in_minutes: 180
+        mirror_hardwares: a3_npu_2
+        env:
+          VLLM_WORKER_MULTIPROC_METHOD: spawn
+          VLLM_ALLOW_LONG_MAX_MODEL_LEN: "1"
+          HF_TOKEN: "${HF_TOKEN}"
+        commands:
+          - export BENCHMARK_DIR=tests/dfx/perf/results
+          - |
+            set +e
+            pytest -s -v tests/dfx/perf/scripts/run_benchmark.py -m "npu and A3" --test-config-file tests/dfx/perf/tests/test_minicpmo_4_5.json
+            EXIT=$$?
+            buildkite-agent artifact upload "tests/dfx/perf/results/*.json"
+            exit $$EXIT
+      - label: ":full_moon: Omni · MiniCPM-o 4.5 · Perf Test (A2)"
+        source_file_dependencies: omni_minicpmo_4_5_perf
+        key: nightly-omni-performance-minicpmo-4-5-npu-a2
+        timeout_in_minutes: 180
+        mirror_hardwares: a2b3_npu_1
+        env:
+          VLLM_WORKER_MULTIPROC_METHOD: spawn
+          VLLM_ALLOW_LONG_MAX_MODEL_LEN: "1"
+          HF_TOKEN: "${HF_TOKEN}"
+          HF_HUB_OFFLINE: "1"
+        commands:
+          - export BENCHMARK_DIR=tests/dfx/perf/results
+          - |
+            set +e
+            pytest -s -v tests/dfx/perf/scripts/run_benchmark.py -m "npu and A2" --test-config-file tests/dfx/perf/tests/test_minicpmo_4_5.json
+            EXIT=$$?
+            buildkite-agent artifact upload "tests/dfx/perf/results/*.json"
+            exit $$EXIT
+      - label: ":full_moon: Omni · MiniCPM-o 4.5 · Duplex Seed-TTS Perf Test"
+        source_file_dependencies: omni_minicpmo_4_5_duplex_perf
+        key: nightly-omni-performance-minicpmo-4-5-duplex-seed-tts-npu
+        timeout_in_minutes: 180
+        mirror_hardwares: a3_npu_2
+        env:
+          VLLM_WORKER_MULTIPROC_METHOD: spawn
+          VLLM_ALLOW_LONG_MAX_MODEL_LEN: "1"
+          HF_TOKEN: "${HF_TOKEN}"
+        commands:
+          - export BENCHMARK_DIR=tests/dfx/perf/results
+          - |
+            set +e
+            pytest -s -v tests/dfx/perf/scripts/run_benchmark.py -m "npu and A3" --test-config-file tests/dfx/perf/tests/test_minicpmo_4_5_duplex_seed_tts.json
+            EXIT=$$?
+            buildkite-agent artifact upload "tests/dfx/perf/results/*.json"
+            exit $$EXIT
   - group: ":card_index_dividers: TTS Model Test"
     key: nightly-tts-test-group
     depends_on: upload-nightly-pipeline
-    if: build.env("NIGHTLY") == "1" || build.pull_request.labels includes "nightly-test" || build.pull_request.labels includes "tts-test"
+    if: build.env("NIGHTLY") == "1" || build.pull_request.labels includes "nightly-test"
     steps:
       - label: ":full_moon: TTS · Perf Test"
+        source_file_dependencies: tts_qwen3_tts_perf
         timeout_in_minutes: 180
         mirror_hardwares: a3_npu_2
         env:
@@ -101,37 +164,10 @@ steps:
     depends_on: upload-nightly-pipeline
     if: >-
       build.env("NIGHTLY") == "1" ||
-      build.pull_request.labels includes "nightly-test" ||
-      build.pull_request.labels includes "diffusion-x2iat-test"
-    steps:
-      - label: ":full_moon: Diffusion X2I(&A&T) · Perf Test · Qwen-Image-Edit-2511"
-        key: nightly-diffusion-x2iat-performance-qwen-image-edit-2511
-        timeout_in_minutes: 180
-        mirror_hardwares: a3_npu_8
-        env:
-          VLLM_WORKER_MULTIPROC_METHOD: spawn
-          VLLM_ALLOW_LONG_MAX_MODEL_LEN: "1"
-          HF_TOKEN: "${HF_TOKEN}"
-          DIFFUSION_ATTENTION_BACKEND: TORCH_SDPA
-        commands:
-          - export DIFFUSION_BENCHMARK_DIR=tests/dfx/perf/results
-          - export CACHE_DIT_VERSION=1.3.0
-          - |
-            set +e
-            pytest -s -v tests/dfx/perf/scripts/run_diffusion_benchmark.py -m npu --test-config-file tests/dfx/perf/tests/test_qwen_image_edit_2511_vllm_omni.json
-            EXIT=$$?
-            buildkite-agent artifact upload "tests/dfx/perf/results/diffusion_result_*.json"
-            buildkite-agent artifact upload "tests/dfx/perf/results/logs/*.log"
-            exit $$EXIT
-  - group: ":card_index_dividers: Diffusion I2I Model Test"
-    key: nightly-HunyuanImage3-x2i-test-group
-    depends_on: upload-nightly-pipeline
-    if: >-
-      build.env("NIGHTLY") == "1" ||
-      build.pull_request.labels includes "nightly-test" ||
-      build.pull_request.labels includes "diffusion-x2iat-test"
+      build.pull_request.labels includes "nightly-test"
     steps:
       - label: ":full_moon: HunyuanImage3-DIT · Accuracy Test"
+        source_file_dependencies: diffusion_hunyuan_image3_accuracy
         timeout_in_minutes: 180
         mirror_hardwares: a3_npu_4
         env:
@@ -141,15 +177,8 @@ steps:
           DIFFUSION_ATTENTION_BACKEND: TORCH_SDPA
         commands:
           - pytest -s -v tests/e2e/accuracy/test_hunyuan_image3_pixel_accuracy.py -m "full_model and A3" --run-level full_model
-  - group: ":card_index_dividers: Diffusion Benchmark Perf Test"
-    key: nightly-diffusion-benchmark-group
-    depends_on: upload-nightly-pipeline
-    if: >-
-      build.env("NIGHTLY") == "1" ||
-      build.pull_request.labels includes "nightly-test" ||
-      build.pull_request.labels includes "diffusion-x2iat-test"
-    steps:
       - label: ":full_moon: Diffusion Perf · HunyuanImage3 TP=2 CFGP=2"
+        source_file_dependencies: diffusion_hunyuan_image3_perf
         timeout_in_minutes: 180
         mirror_hardwares: a3_npu_4
         env:
@@ -166,6 +195,7 @@ steps:
             buildkite-agent artifact upload "tests/dfx/perf/results/logs/*.log"
             exit $$EXIT
       - label: ":full_moon: Diffusion Perf · HunyuanImage3 TP=2 SP=2"
+        source_file_dependencies: diffusion_hunyuan_image3_perf
         timeout_in_minutes: 180
         mirror_hardwares: a3_npu_4
         env:
@@ -182,6 +212,7 @@ steps:
             buildkite-agent artifact upload "tests/dfx/perf/results/logs/*.log"
             exit $$EXIT
       - label: ":full_moon: Diffusion Perf · HunyuanImage3 TP=4"
+        source_file_dependencies: diffusion_hunyuan_image3_perf
         timeout_in_minutes: 180
         mirror_hardwares: a3_npu_4
         env:
@@ -202,10 +233,10 @@ steps:
     depends_on: upload-nightly-pipeline
     if: >-
       build.env("NIGHTLY") == "1" ||
-      build.pull_request.labels includes "nightly-test" ||
-      build.pull_request.labels includes "diffusion-x2v-test"
+      build.pull_request.labels includes "nightly-test"
     steps:
-      - label: ":full_moon: Diffusion X2V · Function Test"
+      - label: ":full_moon: Diffusion X2V · Wan 2.2 Function Test"
+        source_file_dependencies: diffusion_wan22_function
         mirror_hardwares: a2b3_npu_4
         env:
           VLLM_WORKER_MULTIPROC_METHOD: spawn
@@ -213,7 +244,8 @@ steps:
           HF_TOKEN: "${HF_TOKEN}"
         commands:
           - pytest -s -v tests/e2e/online_serving/test_wan22_expansion.py -k npu_i2v --run-level "full_model"
-      - label: ":full_moon: Diffusion X2V · Perf Test"
+      - label: ":full_moon: Diffusion X2V · Wan 2.2 Perf Test"
+        source_file_dependencies: diffusion_wan22_perf
         key: nightly-diffusion-x2v-performance
         timeout_in_minutes: 180
         mirror_hardwares: a3_npu_4

.buildkite/npu/test-npu-ready.yml

@@ -1,4 +1,60 @@
 steps:
+  - group: ":card_index_dividers: Simple Test"
+    depends_on: upload-ready-pipeline
+    steps:
+      - label: "Platforms Test"
+        key: ready-platforms-test
+        mirror_hardwares: a3_npu_2
+        commands:
+          - >-
+            pytest -s -v
+            tests/platforms/npu/test_minicpmo_code2wav_npugraph.py
+            -m 'core_model and npu and A3 and omni'
+            --run-level 'core_model'
+
+      - label: "Diffusion NPU layer tests - A2"
+        key: ready-diffusion-npu-layer-tests-a2
+        mirror_hardwares: a2b3_npu_1
+        commands:
+          - >-
+            pytest -s -v
+            tests/diffusion/layers/test_activation.py
+            -m 'core_model and npu and A2'
+            --run-level 'core_model'
+
+      - label: "Diffusion NPU layer tests - A3"
+        key: ready-diffusion-npu-layer-tests
+        mirror_hardwares: a3_npu_2
+        commands:
+          - >-
+            pytest -s -v
+            tests/diffusion/layers/test_activation.py
+            tests/diffusion/layers/test_fused_qk_norm_rope_npu.py
+            tests/diffusion/models/minimax_h3/test_minimax_h3_qwen3vl_swiglu.py
+            tests/diffusion/models/minimax_h3/test_minimax_h3_modulation.py
+            -m 'core_model and npu and A3'
+            --run-level 'core_model'
+
+      - label: "Diffusion NPU layer tests - A5"
+        key: ready-diffusion-npu-layer-tests-a5
+        mirror_hardwares: a5_npu_2
+        commands:
+          - >-
+            pytest -s -v
+            tests/diffusion/layers/test_activation.py
+            -m 'core_model and npu and A5'
+            --run-level 'core_model'
+
+      - label: "Diffusion NPU layer tests - 310P"
+        key: ready-diffusion-npu-layer-tests-310p
+        mirror_hardwares: 310p_npu_1
+        commands:
+          - >-
+            pytest -s -v
+            tests/diffusion/layers/test_activation.py
+            -m 'core_model and npu and 310P'
+            --run-level 'core_model'
+
   - group: ":card_index_dividers: Omni Model Test"
     key: ready-omni-group
     depends_on: upload-ready-pipeline
@@ -17,13 +73,22 @@ steps:
           VLLM_ALLOW_LONG_MAX_MODEL_LEN: "1"
         commands:
           - pytest -s -v tests/e2e/online_serving/test_minicpmo_4_5_duplex.py -m 'core_model and npu and A3 and omni' --run-level 'core_model'
+      - label: "Omni · MiniCPM-o 4.5 Perf Test"
+        timeout_in_minutes: 60
+        mirror_hardwares: a3_npu_2
+        env:
+          VLLM_WORKER_MULTIPROC_METHOD: spawn
+          VLLM_ALLOW_LONG_MAX_MODEL_LEN: "1"
+        commands:
+          - export BENCHMARK_DIR=tests/dfx/perf/results
+          - pytest -s -v tests/dfx/perf/scripts/run_benchmark.py --test-config-file tests/dfx/perf/tests/test_minicpmo_4_5_ready.json -m 'core_model and npu and A3 and omni'
 
   - group: ":card_index_dividers: TTS Model Test"
     key: ready-tts-group
     depends_on: upload-ready-pipeline
     steps:
       - label: ":sound: TTS · Qwen3-TTS CustomVoice Test"
-        mirror_hardwares: a2b3_npu_1
+        mirror_hardwares: a3_npu_2
         env:
           VLLM_WORKER_MULTIPROC_METHOD: spawn
           VLLM_ALLOW_LONG_MAX_MODEL_LEN: "1"

.buildkite/release/release-pipeline.yml

@@ -40,9 +40,11 @@ steps:
     commands:
       - "bash .buildkite/release/scripts/publish-release-wheels-pypi.sh"
 
+  # Auto-skip when NIGHTLY=1 (scheduled); otherwise require manual unblock.
   - block: "Unblock to build release Docker images"
     depends_on: ~
     key: block-build-release-images
+    if: build.env("NIGHTLY") != "1"
 
   - group: "Build release Docker images"
     key: "build-release-images"
@@ -63,7 +65,7 @@ steps:
               --progress plain \
               -f docker/Dockerfile.cuda .
           - "docker push public.ecr.aws/q9t5s3a7/vllm-omni-release-repo:$BUILDKITE_COMMIT-$(uname -m)"
-          # re-tag to default image tag and push, just in case arm64 build fails
+          # Fallback default tag (no arch suffix) if arm64 build fails.
           - "docker tag public.ecr.aws/q9t5s3a7/vllm-omni-release-repo:$BUILDKITE_COMMIT-$(uname -m) public.ecr.aws/q9t5s3a7/vllm-omni-release-repo:$BUILDKITE_COMMIT"
           - "docker push public.ecr.aws/q9t5s3a7/vllm-omni-release-repo:$BUILDKITE_COMMIT"
         env:
@@ -86,6 +88,26 @@ steps:
         env:
           DOCKER_BUILDKIT: "1"
 
+  # NIGHTLY=1 only: promote ECR images to DockerHub nightly tags, then prune.
+  - label: "Publish nightly multi-arch image to DockerHub"
+    depends_on:
+      - build-release-image-x86
+      - build-release-image-arm64
+    if: build.env("NIGHTLY") == "1"
+    key: publish-nightly-images-dockerhub
+    agents:
+      queue: small_cpu_queue_release
+    commands:
+      - "bash .buildkite/release/scripts/publish-release-images.sh nightly"
+      - "bash .buildkite/release/scripts/cleanup-nightly-builds.sh"
+    plugins:
+      - docker-login#v3.0.0:
+          username: vllmbot
+          password-env: DOCKERHUB_TOKEN
+    env:
+      DOCKER_BUILDKIT: "1"
+      DOCKERHUB_USERNAME: "vllmbot"
+
   - block: "Publish release images to DockerHub"
     key: block-publish-release-images
     depends_on:

.buildkite/release/scripts/cleanup-nightly-builds.sh

@@ -0,0 +1,147 @@
+#!/bin/bash
+# SPDX-License-Identifier: Apache-2.0
+#
+# Clean up old nightly builds from DockerHub (vllm/vllm-omni), keeping the
+# newest 14 tags with the nightly- prefix.
+
+set -ex
+
+TAG_PREFIX="nightly-"
+REPO="vllm/vllm-omni"
+REPO_API_URL="https://hub.docker.com/v2/repositories/${REPO}/tags"
+
+echo "Cleaning up tags with prefix: $TAG_PREFIX in repository: $REPO"
+
+if [ -z "$DOCKERHUB_TOKEN" ]; then
+  echo "Error: DOCKERHUB_TOKEN environment variable is not set"
+  exit 1
+fi
+
+if [ -z "$DOCKERHUB_USERNAME" ]; then
+  echo "Error: DOCKERHUB_USERNAME environment variable is not set"
+  exit 1
+fi
+
+echo "Getting DockerHub bearer token..."
+set +x
+BEARER_TOKEN=$(curl -s -X POST \
+  -H "Content-Type: application/json" \
+  -d "{\"username\": \"$DOCKERHUB_USERNAME\", \"password\": \"$DOCKERHUB_TOKEN\"}" \
+  "https://hub.docker.com/v2/users/login" | jq -r '.token')
+set -x
+
+if [ -z "$BEARER_TOKEN" ] || [ "$BEARER_TOKEN" = "null" ]; then
+  echo "Error: Failed to get DockerHub bearer token"
+  exit 1
+fi
+
+get_all_tags() {
+  local page=1
+  local all_tags=""
+  # Hub `name` is a partial match (not a strict prefix). Use it to narrow the
+  # list, then enforce startswith() client-side. Exhaust API pages; do not
+  # stop merely because a page has no matching tags after filtering.
+  local name_query
+  name_query=$(printf '%s' "$TAG_PREFIX" | jq -sRr @uri)
+
+  while true; do
+    set +x
+    local response
+    response=$(curl -s -H "Authorization: Bearer $BEARER_TOKEN" \
+      "$REPO_API_URL?page=$page&page_size=100&name=${name_query}")
+    set -x
+
+    local page_count
+    page_count=$(echo "$response" | jq -r '.results | length')
+    if [ -z "$page_count" ] || [ "$page_count" = "0" ] || [ "$page_count" = "null" ]; then
+      break
+    fi
+
+    local tags
+    tags=$(echo "$response" | jq -r --arg prefix "$TAG_PREFIX" \
+      '.results[] | select(.name | startswith($prefix)) | "\(.last_updated)|\(.name)"')
+
+    if [ -n "$tags" ]; then
+      all_tags="$all_tags$tags"$'\n'
+    fi
+    page=$((page + 1))
+  done
+
+  echo "$all_tags" | sort -r | cut -d'|' -f2
+}
+
+delete_tag() {
+  local tag_name="$1"
+  echo "Deleting tag: $tag_name"
+
+  local delete_url="https://hub.docker.com/v2/repositories/${REPO}/tags/$tag_name"
+  set +x
+  local response http_code body
+  response=$(curl -s -w "\n%{http_code}" -X DELETE \
+    -H "Authorization: Bearer $BEARER_TOKEN" "$delete_url")
+  set -x
+  http_code=$(printf '%s\n' "$response" | tail -n1)
+  body=$(printf '%s\n' "$response" | sed '$d')
+
+  if [ "$http_code" -ge 200 ] && [ "$http_code" -lt 300 ]; then
+    echo "Successfully deleted tag: $tag_name (HTTP ${http_code})"
+    return 0
+  fi
+
+  local detail=""
+  if echo "$body" | jq -e '.detail' > /dev/null 2>&1; then
+    detail=$(echo "$body" | jq -r '.detail')
+  elif [ -n "$body" ]; then
+    detail="$body"
+  else
+    detail="(empty response body)"
+  fi
+  echo "Error: Failed to delete tag $tag_name (HTTP ${http_code}): ${detail}"
+  return 1
+}
+
+echo "Fetching all tags from DockerHub..."
+all_tags=$(get_all_tags)
+
+if [ -z "$all_tags" ]; then
+  echo "No tags found to clean up"
+  exit 0
+fi
+
+total_tags=$(echo "$all_tags" | wc -l)
+echo "Found $total_tags tags"
+
+tags_to_keep=14
+tags_to_delete=$((total_tags - tags_to_keep))
+
+if [ $tags_to_delete -le 0 ]; then
+  echo "No tags need to be deleted (only $total_tags tags found, keeping $tags_to_keep)"
+  exit 0
+fi
+
+echo "Will delete $tags_to_delete old tags, keeping the newest $tags_to_keep"
+
+tags_to_delete_list=$(echo "$all_tags" | tail -n +$((tags_to_keep + 1)))
+
+if [ -z "$tags_to_delete_list" ]; then
+  echo "No tags to delete"
+  exit 0
+fi
+
+echo "Deleting old tags..."
+delete_failures=0
+while IFS= read -r tag; do
+  if [ -n "$tag" ]; then
+    if ! delete_tag "$tag"; then
+      delete_failures=$((delete_failures + 1))
+    fi
+    sleep 1
+  fi
+done <<< "$tags_to_delete_list"
+
+if [ "$delete_failures" -gt 0 ]; then
+  echo "Cleanup finished with ${delete_failures} deletion failure(s)"
+  exit 1
+fi
+
+echo "Cleanup completed successfully"

.buildkite/release/scripts/publish-release-images.sh

@@ -1,50 +1,87 @@
 #!/bin/bash
 # SPDX-License-Identifier: Apache-2.0
 #
-# Publish release Docker images from ECR to DockerHub.
-# Pulls per-arch images, tags with `latest` and versioned tags, pushes them,
-# then creates and pushes a multi-arch manifest.
+# Publish Docker images from ECR staging to DockerHub.
+#
+# Usage:
+#   publish-release-images.sh            # release mode (default)
+#   publish-release-images.sh release
+#   publish-release-images.sh nightly
+#
+# Release tags: latest, v<version>
+# Nightly tags: nightly, nightly-<commit>
 
 set -euo pipefail
 
-RELEASE_VERSION=$(buildkite-agent meta-data get release-version --default "" | sed 's/^v//')
-if [ -z "${RELEASE_VERSION}" ]; then
-  echo "ERROR: release-version metadata not set"
-  exit 1
-fi
+MODE="${1:-release}"
 
-COMMIT="$BUILDKITE_COMMIT"
 DOCKERHUB_REPO="vllm/vllm-omni"
 ECR_REPO="public.ecr.aws/q9t5s3a7/vllm-omni-release-repo"
+COMMIT="$BUILDKITE_COMMIT"
+PRIMARY_TAGS=()
+EXTRA_MANIFEST_TAGS=()
+
+case "${MODE}" in
+  release)
+    RELEASE_VERSION=$(buildkite-agent meta-data get release-version --default "" | sed 's/^v//')
+    if [ -z "${RELEASE_VERSION}" ]; then
+      echo "ERROR: release-version metadata not set"
+      exit 1
+    fi
+    PRIMARY_TAGS=("latest" "v${RELEASE_VERSION}")
+    ;;
+  nightly)
+    # Commit-pinned manifest reuses nightly-{x86_64,aarch64} arch images.
+    PRIMARY_TAGS=("nightly")
+    EXTRA_MANIFEST_TAGS=("nightly-${COMMIT}")
+    ;;
+  *)
+    echo "Usage: $0 {release|nightly}"
+    exit 2
+    ;;
+esac
 
 echo "========================================"
-echo "Publishing release images v${RELEASE_VERSION}"
+echo "Publishing ${MODE} images to ${DOCKERHUB_REPO}"
 echo "  Commit: ${COMMIT}"
+echo "  Primary tags: ${PRIMARY_TAGS[*]}"
+if [ ${#EXTRA_MANIFEST_TAGS[@]} -gt 0 ]; then
+  echo "  Extra manifests: ${EXTRA_MANIFEST_TAGS[*]}"
+fi
 echo "========================================"
 
-# Login to ECR to pull staging images
 aws ecr-public get-login-password --region us-east-1 | \
   docker login --username AWS --password-stdin public.ecr.aws/q9t5s3a7
 
-docker pull "${ECR_REPO}:${COMMIT}-x86_64"
-docker pull "${ECR_REPO}:${COMMIT}-aarch64"
+SRC_X86="${ECR_REPO}:${COMMIT}-x86_64"
+SRC_ARM="${ECR_REPO}:${COMMIT}-aarch64"
 
-docker tag "${ECR_REPO}:${COMMIT}-x86_64" "${DOCKERHUB_REPO}:latest-x86_64"
-docker tag "${ECR_REPO}:${COMMIT}-x86_64" "${DOCKERHUB_REPO}:v${RELEASE_VERSION}-x86_64"
-docker push "${DOCKERHUB_REPO}:latest-x86_64"
-docker push "${DOCKERHUB_REPO}:v${RELEASE_VERSION}-x86_64"
+docker pull "${SRC_X86}"
+docker pull "${SRC_ARM}"
 
-docker tag "${ECR_REPO}:${COMMIT}-aarch64" "${DOCKERHUB_REPO}:latest-aarch64"
-docker tag "${ECR_REPO}:${COMMIT}-aarch64" "${DOCKERHUB_REPO}:v${RELEASE_VERSION}-aarch64"
-docker push "${DOCKERHUB_REPO}:latest-aarch64"
-docker push "${DOCKERHUB_REPO}:v${RELEASE_VERSION}-aarch64"
+publish_manifest() {
+  local tag="$1"
+  local arch_tag_base="$2"
+  docker manifest rm "${DOCKERHUB_REPO}:${tag}" || true
+  docker manifest create \
+    "${DOCKERHUB_REPO}:${tag}" \
+    "${DOCKERHUB_REPO}:${arch_tag_base}-x86_64" \
+    "${DOCKERHUB_REPO}:${arch_tag_base}-aarch64"
+  docker manifest push "${DOCKERHUB_REPO}:${tag}"
+}
 
-docker manifest rm "${DOCKERHUB_REPO}:latest" || true
-docker manifest rm "${DOCKERHUB_REPO}:v${RELEASE_VERSION}" || true
-docker manifest create "${DOCKERHUB_REPO}:latest" "${DOCKERHUB_REPO}:latest-x86_64" "${DOCKERHUB_REPO}:latest-aarch64"
-docker manifest create "${DOCKERHUB_REPO}:v${RELEASE_VERSION}" "${DOCKERHUB_REPO}:v${RELEASE_VERSION}-x86_64" "${DOCKERHUB_REPO}:v${RELEASE_VERSION}-aarch64"
-docker manifest push "${DOCKERHUB_REPO}:latest"
-docker manifest push "${DOCKERHUB_REPO}:v${RELEASE_VERSION}"
+for tag in "${PRIMARY_TAGS[@]}"; do
+  docker tag "${SRC_X86}" "${DOCKERHUB_REPO}:${tag}-x86_64"
+  docker tag "${SRC_ARM}" "${DOCKERHUB_REPO}:${tag}-aarch64"
+  docker push "${DOCKERHUB_REPO}:${tag}-x86_64"
+  docker push "${DOCKERHUB_REPO}:${tag}-aarch64"
+  publish_manifest "${tag}" "${tag}"
+done
+
+ARCH_BASE="${PRIMARY_TAGS[0]}"
+for tag in "${EXTRA_MANIFEST_TAGS[@]}"; do
+  publish_manifest "${tag}" "${ARCH_BASE}"
+done
 
 echo ""
-echo "Successfully published release images for v${RELEASE_VERSION}"
+echo "Successfully published ${MODE} images"

.claude/skills/add-diffusion-model/SKILL.md

@@ -1,6 +1,6 @@
 ---
 name: add-diffusion-model
-description: Add a new diffusion model (text-to-image, text-to-video, image-to-video, text-to-audio, image editing) to vLLM-Omni, including Cache-DiT acceleration and parallelism support (TP, SP/USP, CFG-Parallel, HSDP). Use when integrating a new diffusion model, porting a diffusers pipeline or a custom model repo to vllm-omni, creating a new DiT transformer adapter, adding diffusion model support, or enabling multi-GPU parallelism and cache acceleration for an existing model.
+description: Add a new diffusion model (text-to-image, text-to-video, image-to-video, text-to-audio, image editing) to vLLM-Omni, including native non-Diffusers ports, reference-parity validation, Cache-DiT, offload, and parallelism support (TP, SP/USP, CFG-Parallel, HSDP). Use when integrating or reviewing a new diffusion model, porting a Diffusers pipeline or custom model repository, creating a DiT adapter, reusing shared examples, or qualifying multi-GPU and memory optimizations.
 ---
 
 # Adding a Diffusion Model to vLLM-Omni
@@ -26,9 +26,16 @@ Check the model's HF repo for `model_index.json`. This determines your path:
 |----------|----------------|----------------|
 | **Already supported** | `_class_name` in `model_index.json` matches a key in `_DIFFUSION_MODELS` in `registry.py` | Skip implementation, then validate model-specific examples, tests, and docs as needed |
 | **Diffusers-based** | Has standard `model_index.json` with `_diffusers_version`, subfolders for `transformer/`, `vae/`, etc. | Follow **Path A** below |
-| **Custom/private repo** | No diffusers `model_index.json`, weights in non-standard format, custom model code in a separate git repo | Follow **Path B** below |
+| **Native non-Diffusers model** | No Diffusers index, non-standard checkpoint hierarchy, or custom architecture in a separate repo | Follow **Path B** below; port the runtime natively unless an external adapter was explicitly requested |
 | **Hybrid** | Has some diffusers components (VAE) but custom transformer/fusion | Mix of Path A and Path B |
 
+Before coding, write a short integration contract covering runtime ownership,
+checkpoint discovery, reference revision, I/O geometry, attention semantics,
+CFG behavior, auxiliary components, target hardware, and the default deployment.
+For Path B, hybrid models, or any optimization work, read
+[references/native-model-integration-checklist.md](references/native-model-integration-checklist.md)
+and use its phase gates.
+
 ## Path A: Diffusers-Based Model
 
 For models with a standard diffusers layout. See [references/transformer-adaptation.md](references/transformer-adaptation.md) for detailed code patterns.
@@ -103,9 +110,15 @@ For custom loop structures, follow `vllm_omni/diffusion/models/progress_bar.py`
 
 ---
 
-## Path B: Custom/Private Repo Model
+## Path B: Native / Non-Diffusers Model
 
-For models without a diffusers pipeline — weights in custom format, model code in a private repo. Real examples: DreamID-Omni, BAGEL, HunyuanImage3. See [references/custom-model-patterns.md](references/custom-model-patterns.md) for concrete custom-model integration patterns.
+For models without a Diffusers pipeline—weights in custom formats and model
+code in another public or private repository. Treat that repository as a pinned
+correctness oracle. A request for native support means the vLLM-Omni runtime
+must not import the reference implementation; an external adapter is appropriate
+only when the requested scope explicitly permits that dependency. See
+[references/custom-model-patterns.md](references/custom-model-patterns.md) for
+concrete integration patterns.
 
 ### B1. Understand the reference repo
 
@@ -126,16 +139,25 @@ This is the key design decision for custom models. Follow these placement rules:
 | **Pipeline orchestration** (init, forward, denoise loop) | `vllm_omni/diffusion/models/<name>/pipeline_<name>.py` | Always required |
 | **Custom transformer/backbone** (ported and adapted to vllm-omni) | `vllm_omni/diffusion/models/<name>/<name>_transformer.py` or similar | `wan2_2.py`, `fusion.py`, `bagel_transformer.py` |
 | **Custom sub-models** (VAE, fusion, autoencoder) | `vllm_omni/diffusion/models/<name>/` as separate files | `autoencoder.py`, `fusion.py` |
-| **External dependency code** (original repo utilities) | **External repo**, installed via download script or pip | `dreamid_omni` package via git clone |
+| **Reference-only code** | Keep outside the runtime; use a pinned revision for golden outputs and architecture analysis | Reference inference script |
+| **Explicit external adapter dependency** | External package, only when the requested scope and maintainer direction allow it | Compatibility adapter, not native support |
 | **Hardcoded model configs** | Module-level dicts in pipeline file | `VIDEO_CONFIG`, `AUDIO_CONFIG` dicts |
-| **Download/setup script** | `examples/offline_inference/<name>/download_<name>.py` | `download_dreamid_omni.py` |
+| **Download/setup script** | `examples/offline_inference/<name>/download_<name>.py` | `download_<name>.py` |
 | **Custom `model_index.json`** | Generated by download script, placed at model root | Minimal: `{"_class_name": "YourPipeline", ...}` |
 
 ### B3. Handle external dependencies
 
-If the model's code lives in a separate git repo:
+If the model's code lives in a separate git repo, first decide whether the
+requested deliverable is native support or an external adapter. Do not silently
+choose the adapter path.
 
-**Option 1: Import with graceful fallback** (recommended for models with external utils)
+**Option 1: Port the code directly** (default for native support)
+
+Copy the essential model files into `vllm_omni/diffusion/models/<name>/` and
+adapt them to shared vLLM-Omni contracts. Keep checkpoint loading strict and
+use the pinned reference only to generate parity evidence.
+
+**Option 2: Import with graceful fallback** (adapter scope only)
 
 ```python
 try:
@@ -147,17 +169,15 @@ except ImportError:
     )
 ```
 
-**Option 2: Port the code directly** (preferred when feasible)
-
-Copy the essential model files into `vllm_omni/diffusion/models/<name>/` and adapt them. This avoids external dependencies. BAGEL does this — `autoencoder.py` and `bagel_transformer.py` are ported directly.
-
-**Decision criteria**: Port if the code is self-contained and won't diverge. Use external deps if the model repo is actively maintained and the code is complex.
+Use an external runtime dependency only when the user explicitly requested an
+adapter or a maintainer approved the exception. Document the dependency,
+pinned revision, installation path, and unsupported native features.
 
 ### B4. Handle custom weight loading
 
 Custom models have two common patterns for weight loading:
 
-**Pattern 1: Bypass standard loader** (DreamID-Omni style)
+**Pattern 1: Bypass standard loader** (eager custom init)
 
 When the original model has complex custom init functions that load weights in `__init__`:
 
@@ -221,7 +241,8 @@ class CustomPipeline(nn.Module):
 
 ### B5. Create the `model_index.json`
 
-Custom models need a `model_index.json` at the model root for vllm-omni to discover them. For custom models, this is minimal:
+Prefer a `model_index.json` at the model root when vLLM-Omni owns an assembled
+checkpoint directory. For custom models, this is minimal:
 
 ```json
 {
@@ -230,7 +251,14 @@ Custom models need a `model_index.json` at the model root for vllm-omni to disco
 }
 ```
 
-The `_class_name` must match a key in `_DIFFUSION_MODELS` in `registry.py`. Additional keys are model-specific (accessed via `od_config.model_config`).
+The `_class_name` must match a key in `_DIFFUSION_MODELS` in `registry.py`.
+Additional keys are model-specific (accessed via `od_config.model_config`).
+
+If the released repository is immutable and has neither a root `config.json`
+nor a Diffusers index, add it to a generic native-checkpoint signature resolver:
+match the exact Hub ID or a distinctive set of local files and return the
+pipeline class. Do not use model-name substrings or add parallel one-off
+predicates in CLI, config, and serving consumers.
 
 If the model's weights come from multiple HF repos, write a **download script** that:
 1. Downloads from each repo
@@ -292,7 +320,9 @@ Use the appropriate existing example script:
 | Image-to-Image | `examples/offline_inference/image_to_image/image_edit.py` |
 | Text-to-Audio | `examples/offline_inference/text_to_audio/text_to_audio.py` |
 
-For custom/Omni models that don't fit these categories, create a dedicated example script.
+Reuse these shared scripts even for custom models when their request and output
+contracts fit. Create a dedicated model script only when the shared category
+cannot represent the protocol, and document that gap in the PR.
 
 **Validation**: No errors, output is meaningful, quality matches reference implementation.
 
@@ -300,13 +330,17 @@ See [references/troubleshooting.md](references/troubleshooting.md) for common er
 
 ### Step 6: Add Example Scripts
 
-For Omni or custom models, create:
+Only when the shared category scripts cannot represent the model, create:
 - `examples/offline_inference/your_model_name/` — offline script + README
 - `examples/online_serving/your_model_name/` — server script + client
 - Download script if weights require assembly from multiple sources
 
 ### Step 7: Update Documentation
 
+Follow the [`add-recipe` skill](../add-recipe/SKILL.md) to add or update the
+model-family recipe and its `recipes/README.md` row with verified specifications,
+hardware, commands, feature links, and qualification evidence.
+
 Required updates:
 1. `docs/user_guide/diffusion/parallelism/overview.md` — parallelism support overview/table
 2. `docs/user_guide/diffusion/cpu_offload.md` — if CPU offload supported (add to supported models table)
@@ -338,7 +372,7 @@ Classify the model's **CI priority** first:
 
 **L2 & L3 online — same file, dual marks on the baseline smoke:** The **first / simplest** case in `test_{slug}.py` (default deploy, minimal steps, single prompt) should carry **both** `@pytest.mark.core_model` and `@pytest.mark.advanced_model` on the **same** function so L2 (`test-ready.yml`) and L3 (`test-merge.yml`) share one smoke test. Heavier deploy variants or API paths in the same file use **`advanced_model` only**. When L3 moves to nightly, migrate those heavier cases into `test_{slug}_expansion.py` with `full_model` and remove the dedicated `test-merge.yml` job (see `test_longcat_image_expansion.py`, `test_qwen_image_expansion.py`).
 
-**L4 design (high priority):** Combine multiple supported features (Cache-DiT, TP, USP, CFG, HSDP, CPU offload, quantization) into **few parametrized** `OmniServerParams` rows so each feature appears in at least one case without exploding GPU jobs. Shard single-GPU vs multi-GPU cases across the nightly X2I/X2V function steps (1-GPU vs `distributed_cuda`).
+**L4 design (high priority):** Combine multiple supported features (Cache-DiT, TP, USP, CFG, HSDP, CPU offload, quantization) into **few parametrized** `OmniServerParams` rows so each feature appears in at least one case without exploding GPU jobs. Shard single-GPU vs multi-GPU cases across the nightly X2I/X2V function steps (`cards_1` vs `not cards_1`).
 
 **L4 design (medium / low):** One or two parametrized rows covering the best quality/perf trade-off; skip perf JSON unless the model is high priority.
 
@@ -346,107 +380,20 @@ Classify the model's **CI priority** first:
 
 **Keep model-specific code inside test modules — not `tests/helpers/{slug}.py`:** deploy constants, prompts, sampling dicts, and inline `request_config` / `form_data` belong in each `test_{slug}.py` and `test_{slug}_expansion.py`. Do not add per-model files under `tests/helpers/`; reuse only repo-wide harness (`mark`, `media`, `runtime`, `stage_config`, `assertions`). **L2+ online/offline e2e:** reuse or add `send_*_request` in `tests/helpers/runtime.py` — tests call the handler, not raw `omni.generate` / HTTP. See [vllm-omni-test skill](../vllm-omni-test/SKILL.md) § **Runtime send helpers**.
 
+Keep the model suite proportional using the six distinct failure owners in the
+native integration checklist. Combine supported features into a few
+parametrized E2E rows instead of creating one test per optimization.
+
 ### Step 9: Add Cache-DiT Acceleration
 
-Cache-DiT accelerates inference by caching intermediate computation results across denoising steps. After your model is working correctly on a single GPU, add cache-dit support.
+Add caching only after uncached single-device correctness. Read
+[references/cache-dit-patterns.md](references/cache-dit-patterns.md), use the
+automatic single-block-list path when possible, and add a registered
+`BlockAdapter` only for genuinely custom block topology.
 
-See [references/cache-dit-patterns.md](references/cache-dit-patterns.md) for detailed code patterns.
-
-#### 9a. Determine your model type
-
-| Model Type | Description | Action |
-|------------|-------------|--------|
-| **Standard single-transformer** | One transformer with one `ModuleList` of blocks | No code needed — `CacheDiTBackend` auto-detects via `enable_cache_for_dit()` |
-| **Multi-block-list** | One transformer with multiple block lists (e.g., `transformer_blocks` + `single_transformer_blocks`) | Write custom enabler with `BlockAdapter` |
-| **Dual-transformer** | Two transformers (e.g., high-noise + low-noise) | Write custom enabler with `BlockAdapter` wrapping both |
-
-#### 9b. Standard models — verify automatic support
-
-For standard single-transformer models, test directly:
-
-```python
-omni = Omni(
-    model="your-model-name",
-    cache_backend="cache_dit",
-    cache_config={
-        "Fn_compute_blocks": 1,
-        "Bn_compute_blocks": 0,
-        "max_warmup_steps": 4,
-    }
-)
-```
-
-Check logs for "Cache-dit enabled successfully on xxx". If it works, skip to Step 9e.
-
-#### 9c. Custom architectures — write a custom enabler
-
-For multi-block-list or dual-transformer models, write a custom enabler function:
-
-```python
-from cache_dit import BlockAdapter, ForwardPattern, ParamsModifier, DBCacheConfig
-
-def enable_cache_for_your_model(pipeline, cache_config):
-    db_cache_config = DBCacheConfig(
-        num_inference_steps=None,
-        Fn_compute_blocks=cache_config.Fn_compute_blocks,
-        Bn_compute_blocks=cache_config.Bn_compute_blocks,
-        max_warmup_steps=cache_config.max_warmup_steps,
-        max_cached_steps=cache_config.max_cached_steps,
-        max_continuous_cached_steps=cache_config.max_continuous_cached_steps,
-        residual_diff_threshold=cache_config.residual_diff_threshold,
-    )
-
-    cache_dit.enable_cache(
-        BlockAdapter(
-            transformer=pipeline.transformer,
-            blocks=[
-                pipeline.transformer.transformer_blocks,
-                pipeline.transformer.single_transformer_blocks,
-            ],
-            forward_pattern=[ForwardPattern.Pattern_1, ForwardPattern.Pattern_1],
-            params_modifiers=[ParamsModifier(...)],
-        ),
-        cache_config=db_cache_config,
-    )
-
-    def refresh_cache_context(pipeline, num_inference_steps, verbose=True):
-        cache_dit.refresh_context(
-            pipeline.transformer, num_inference_steps=num_inference_steps, verbose=verbose
-        )
-    return refresh_cache_context
-```
-
-#### 9d. Register the custom enabler
-
-Add your enabler to `CUSTOM_DIT_ENABLERS` in
-`vllm_omni/diffusion/cache/cachedit/model_specific.py`:
-
-```python
-CUSTOM_DIT_ENABLERS = {
-    "Wan22Pipeline": enable_cache_for_wan22,
-    "LongCatImagePipeline": enable_cache_for_longcat_image,
-    "YourModelPipeline": enable_cache_for_your_model,  # Add here
-}
-```
-
-#### 9e. Test Cache-DiT
-
-```python
-omni = Omni(
-    model="your-model-name",
-    cache_backend="cache_dit",
-    cache_config={
-        "Fn_compute_blocks": 1, "Bn_compute_blocks": 0,
-        "max_warmup_steps": 4, "residual_diff_threshold": 0.24,
-    }
-)
-images = omni.generate("a beautiful landscape",
-    OmniDiffusionSamplingParams(num_inference_steps=50))
-```
-
-**Verify**: 1) logs show cache enabled, 2) 1.5-2x speedup, 3) output quality acceptable vs baseline.
-
-If quality degrades, lower `residual_diff_threshold` (try 0.12-0.18) or increase `max_warmup_steps` (try 6-8).
+Verify a real cache hit and compare quality with the uncached baseline. Make a
+speed claim only at a realistic step count where warmup permits hits; an
+all-warmup smoke proves integration, not acceleration.
 
 ---
 
@@ -460,107 +407,30 @@ See [references/parallelism-patterns.md](references/parallelism-patterns.md) for
 
 #### 10a. Tensor Parallelism (TP)
 
-Shards DiT linear layers across GPUs. Requires code changes in the transformer.
-
-**What to change in the transformer**:
-1. Replace `nn.Linear` with `ColumnParallelLinear` / `RowParallelLinear` / `QKVParallelLinear`
-2. Update `load_weights()` to handle QKV fusion with `stacked_params_mapping`
-3. Use `self.to_qkv.num_heads` (local heads) instead of total heads for split sizes
-
-```python
-from vllm.model_executor.layers.linear import (
-    QKVParallelLinear, RowParallelLinear, ColumnParallelLinear,
-)
-
-# Attention: QKV → RowParallel output
-self.to_qkv = QKVParallelLinear(dim, head_dim, num_heads, num_kv_heads)
-self.to_out = RowParallelLinear(dim, dim, input_is_parallel=True)
-
-# FFN: ColumnParallel → RowParallel
-self.w1 = ColumnParallelLinear(dim, ffn_dim)
-self.w2 = RowParallelLinear(ffn_dim, dim, input_is_parallel=True)
-```
-
-**Constraints**: `num_heads % tp_size == 0` and `num_kv_heads % tp_size == 0`.
-
-**Test**: `--tensor-parallel-size 2`
+Replace compatible projections with vLLM parallel linears, preserve checkpoint
+fusion/loading, and use local head counts. Require query/KV head divisibility
+and compare a multi-rank forward with the one-rank oracle.
 
 #### 10b. Sequence Parallelism (SP / USP)
 
-Splits sequence tokens across GPUs. Non-intrusive via `_sp_plan` on the transformer class — no changes to `forward()`.
-
-**What to change in the transformer**:
-
-Add `_sp_plan` class attribute:
-
-```python
-from vllm_omni.diffusion.distributed.sp_plan import (
-    SequenceParallelInput, SequenceParallelOutput,
-)
-
-class YourTransformer(nn.Module):
-    _sp_plan = {
-        "blocks.0": {
-            "hidden_states": SequenceParallelInput(split_dim=1, expected_dims=3),
-        },
-        "proj_out": SequenceParallelOutput(gather_dim=1, expected_dims=3),
-    }
-```
-
-If inline tensor ops (e.g., `torch.cat`) exist between shard/gather points, extract them into `nn.Module` submodules so hooks can intercept them.
-
-For RoPE that needs splitting, add an entry for the RoPE module with `split_output=True`.
-
-**Test**: `--ulysses-degree 2` (offline) or `--usp 2` (online serving)
+Prefer the declarative `_sp_plan`. For packed variable-length attention or
+learned-sink LSE correction, keep model math explicit and reuse shared exchange
+utilities. Validate uneven sequence splits, RoPE coordinates, and outputs
+against the one-rank oracle.
 
 #### 10c. CFG Parallel
 
-Distributes positive/negative CFG branches across 2 GPUs. Requires the pipeline to inherit `CFGParallelMixin`.
-
-**What to change in the pipeline**:
-
-```python
-from vllm_omni.diffusion.distributed.cfg_parallel import CFGParallelMixin
-
-class YourPipeline(nn.Module, CFGParallelMixin):
-    def diffuse(self, ...) -> torch.Tensor:
-        for i, t in enumerate(timesteps):
-            positive_kwargs = {...}
-            negative_kwargs = {...} if do_true_cfg else None
-            noise_pred = self.predict_noise_maybe_with_cfg(
-                do_true_cfg=do_true_cfg, true_cfg_scale=cfg_scale,
-                positive_kwargs=positive_kwargs, negative_kwargs=negative_kwargs,
-            )
-            latents = self.scheduler_step_maybe_with_cfg(
-                noise_pred, t, latents, do_true_cfg
-            )
-        return latents
-```
-
-Override `predict_noise()` if your transformer call is non-standard. Override `combine_cfg_noise()` for multi-output models (e.g., video + audio).
-
-**Constraint**: Exactly 2 GPUs. Only for models using classifier-free guidance.
-
-**Test**: `--cfg-parallel-size 2`
+Confirm the model uses CFG. Then reuse `CFGParallelMixin`, overriding prediction
+or recombination only for non-standard or multi-output pipelines. Distinguish a
+packed positive/negative implementation from two independent branches and
+validate the two-rank result against the packed one-rank oracle.
 
 #### 10d. HSDP (Hybrid Sharded Data Parallel)
 
-Shards transformer weights via PyTorch FSDP2 to reduce per-GPU VRAM. No code changes to the forward pass — just add a class attribute.
-
-**What to change in the transformer**:
-
-```python
-class YourTransformer(nn.Module):
-    @staticmethod
-    def _is_transformer_block(name: str, module) -> bool:
-        return "blocks" in name and name.split(".")[-1].isdigit()
-
-    _hsdp_shard_conditions = [_is_transformer_block]
-```
-
-**Constraint**: Cannot combine with TP. For standalone HSDP, set `hsdp_shard_size` explicitly.
-
-**Test**: `--use-hsdp` or `DiffusionParallelConfig(use_hsdp=True)`
+Declare layer shard conditions and ignored rank-local modules; preserve mixed
+checkpoint dtypes. HSDP cannot combine with TP. Measure parameter loading,
+FSDP materialization, warm HBM, and host PSS rather than assuming sharding saves
+peak memory. Keep the resident layout as default if HSDP is worse.
 
 #### 10e. Update parallelism documentation
 
@@ -599,6 +469,12 @@ Pipelines without `SupportsComponentDiscovery` fall back to scanning
 well-known attribute names (`transformer`, `text_encoder`, `vae`,
 etc.), which fails for non-standard names.
 
+Keep model-specific checkpoint paths, nested block aliases, layout transforms,
+and component lifecycles in the model package. Change a shared offloader only
+for a general contract, demonstrate another consumer or a framework-level bug,
+and add one focused shared regression. Avoid `if ModelName` branches in shared
+backends.
+
 ### Step 12: Performance Profiling
 
 After verifying correctness and implementing parallelism/caching, profile the model's performance to identify bottlenecks and ensure optimal execution.
@@ -608,8 +484,32 @@ See the [Profiling Single-Stage Diffusion](../../../docs/contributing/profiling.
 2. Using Nsight Systems (`nsys`) with `profiler: "cuda"` for low-overhead CUDA traces.
 3. Controlling profiling via `omni.start_profile()` and `omni.stop_profile()`.
 
+Report cold E2E, warm user latency, steady-state wave time, peak allocated and
+reserved HBM, host PSS for the full process tree, and stage boundaries. State
+whether prompt encoding and VAE/audio decoding are included. For multi-device
+layouts, plot user latency against throughput per device and keep different
+denoising-step counts on separate Pareto frontiers.
+
 ---
 
+## Pre-commit conventions
+
+New library files must pass the local gates in
+[docs/contributing/README.md](../../../docs/contributing/README.md#linting).
+That page is the full hook list (SPDX, forbidden imports including Hugging Face
+Hub / Triton / pickle, `torch.cuda`, mypy, test marks, markdownlint, Buildkite,
+shellcheck). In particular:
+
+- SPDX copyright is `vLLM-Omni project` (stale `vLLM project` is rewritten).
+- Use `import regex as re` and `pybase64` in `vllm_omni/`; do not import stdlib
+  `re` or `base64`. Hugging Face Hub downloads go through
+  `vllm.transformers_utils.repo_utils`.
+- Do not add `torch.cuda.*` call sites; use `current_omni_platform`.
+- New `tests/**/test_*.py` files need a CI level mark and a hardware mark.
+- Do not expand `CHECK_IMPORTS[*].allowed_files` or `ALLOWED_FILES` without review.
+- GitHub Actions skips SPDX/shellcheck/mypy-3.10/test-marks/markdownlint; run
+  `pre-commit` locally.
+
 ## Iterative Development Tips
 
 1. **Start minimal**: Basic generation first, no parallelism/caching
@@ -617,7 +517,7 @@ See the [Profiling Single-Stage Diffusion](../../../docs/contributing/profiling.
 3. **Use small models**: Test with smaller variants first
 4. **Check tensor shapes**: Most errors are reshape mismatches in attention
 5. **Add features incrementally**: Single GPU → TP → SP → CFG → HSDP → Cache-DiT
-6. **For custom models**: Get the model running with the original code first, then progressively replace components with vllm-omni equivalents
+6. **For custom models**: Run the pinned reference separately, then port the runtime natively; do not ship temporary imports from the reference implementation
 7. **Cache-DiT before parallelism tuning**: Cache-DiT is lossy — verify quality at baseline before combining with parallelism
 8. **Combine lossless + lossy**: e.g., TP + SP + Cache-DiT for maximum throughput
 
@@ -626,6 +526,7 @@ See the [Profiling Single-Stage Diffusion](../../../docs/contributing/profiling.
 - [vllm-omni-test skill](../vllm-omni-test/SKILL.md) — L1–L4 markers, naming, Buildkite wiring, run commands
 - [Transformer Adaptation](references/transformer-adaptation.md) — porting transformers from diffusers
 - [Custom Model Patterns](references/custom-model-patterns.md) — patterns for non-diffusers models
+- [Native Model Integration Checklist](references/native-model-integration-checklist.md) — ownership boundaries, phase gates, qualification matrix, and review evidence
 - [Parallelism Patterns](references/parallelism-patterns.md) — TP, SP/USP, CFG parallel, HSDP implementation details
 - [Cache-DiT Patterns](references/cache-dit-patterns.md) — cache-dit acceleration for standard and custom architectures
 - [Troubleshooting](references/troubleshooting.md) — common errors and fixes

.claude/skills/add-diffusion-model/references/cache-dit-patterns.md

@@ -9,7 +9,9 @@ Three caching strategies:
 - **TaylorSeer**: Calibration-based prediction using Taylor expansion to estimate block outputs
 - **SCM** (Step Computation Masking): Dynamic step skipping based on configurable policies
 
-**Typical speedup**: 1.5-2.5x depending on model and configuration.
+Speedup is model-, step-, and policy-dependent. Treat a cache-enabled run as
+integration evidence until logs or a focused test show real hits and repeated
+timings show an improvement over the uncached baseline.
 
 **Official docs**: https://docs.vllm.ai/projects/vllm-omni/en/latest/design/feature/cache_dit
 
@@ -243,8 +245,10 @@ vllm serve your-model --omni --port 8098 \
 
 **Verification checklist**:
 1. Logs show "Cache-dit enabled successfully on xxx"
-2. Performance: 1.5-2x speedup vs no cache
-3. Quality: compare output with `cache_backend=None`
+2. A focused test or summary confirms at least one real cache hit
+3. Quality is compared with `cache_backend=None`
+4. Performance uses a realistic step count, repeated timings, and a policy
+   whose warmup does not consume the whole run
 
 ## Excluded Models
 

.claude/skills/add-diffusion-model/references/custom-model-patterns.md

@@ -16,20 +16,22 @@ vllm_omni/diffusion/models/wan2_2/
 
 The transformer is loaded separately via `weights_sources` + `load_weights()`. Non-transformer components (VAE, text encoder) are loaded in `__init__` via `from_pretrained()`.
 
-### Custom model with external deps (e.g., DreamID-Omni)
+### External adapter (not native support)
 
 ```
-vllm_omni/diffusion/models/dreamid_omni/
+vllm_omni/diffusion/models/<name>/
 ├── __init__.py                    # Exports pipeline only
-├── pipeline_dreamid_omni.py       # Pipeline: loads ALL weights in __init__ via custom helpers
-├── fusion.py                      # Custom fusion architecture (video + audio cross-attention)
-└── wan2_2.py                      # Re-implemented Wan backbone with split API
+├── pipeline_<name>.py             # Pipeline: loads ALL weights in __init__ via custom helpers
+└── <backbone>.py                  # Custom transformer / fusion architecture
 
-examples/offline_inference/x_to_video_audio/
-└── download_dreamid_omni.py       # Downloads weights from 3 HF repos + clones code repo
+examples/offline_inference/<task>/
+└── download_<name>.py             # Downloads weights and/or clones code repo
 ```
 
-All weights loaded eagerly in `__init__`. `load_weights()` is a no-op. External dependency (`dreamid_omni` package) imported with try/except.
+All weights are loaded eagerly in `__init__`. `load_weights()` is a no-op. The
+external dependency is imported with try/except. Use this pattern only when the
+requested scope explicitly allows an adapter; it does not satisfy a request for
+a native vLLM-Omni implementation.
 
 ### Custom model with ported code (e.g., BAGEL)
 
@@ -76,7 +78,7 @@ self.weights_sources = [
 ]
 ```
 
-### Pattern 3: Fully custom loading (DreamID-Omni)
+### Pattern 3: Fully custom loading
 
 ```
 init → load ALL weights eagerly via custom helpers → load_weights() = no-op
@@ -87,7 +89,7 @@ init → load ALL weights eagerly via custom helpers → load_weights() = no-op
 - `load_weights()` is `pass`
 - Weights may come from multiple HF repos in different formats (`.pth`, `.safetensors`)
 
-Use this when:
+Use this only for an explicitly approved external adapter when:
 - The original model has complex, well-tested loading code you don't want to rewrite
 - Weights span multiple HF repos
 - Weight format is non-standard (e.g., a single `.pth` file, not sharded safetensors)
@@ -108,16 +110,23 @@ Standard diffusers `model_index.json`:
 Custom model `model_index.json` (minimal):
 ```json
 {
-    "_class_name": "DreamIDOmniPipeline",
-    "fusion": "DreamID-Omni/dreamid_omni.safetensors"
+    "_class_name": "YourCustomPipeline",
+    "fusion": "weights/model.safetensors"
 }
 ```
 
-The only **required** field is `_class_name` — it must match a key in `_DIFFUSION_MODELS` in `registry.py`. Other fields are model-specific and accessible via `od_config.model_config` dict.
+The only **required** field is `_class_name` — it must match a key in
+`_DIFFUSION_MODELS` in `registry.py`. Other fields are model-specific and
+accessible via `od_config.model_config`.
+
+When the released repository cannot be modified and contains neither this file
+nor a root `config.json`, use the generic native-checkpoint signature resolver.
+Match the exact Hub ID or a distinctive local file set; never infer support from
+a model-name substring.
 
 ## External Dependency Management
 
-### Git clone + .pth injection (DreamID-Omni pattern)
+### Git clone + .pth injection
 
 ```python
 def download_dependency():
@@ -142,6 +151,10 @@ def download_dependency():
 
 Copy essential files from the original repo into `vllm_omni/diffusion/models/<name>/`. Adapt imports to use vllm-omni utilities. Benefits: no external dependency, no git clone step. Drawback: must maintain the ported code.
 
+This is the default for native model support. Keep the reference repository at
+a pinned revision for architecture analysis and golden outputs, but do not
+import it from production code.
+
 ## Multi-Modal Input/Output Protocols
 
 Custom models that handle images, audio, or video I/O should implement protocol classes:
@@ -163,7 +176,7 @@ The engine checks `isinstance(pipeline, SupportImageInput)` at startup to config
 
 Diffusers models use `config.json` in each subfolder. Custom models often use:
 
-**Module-level config dicts** (DreamID-Omni):
+**Module-level config dicts**:
 ```python
 VIDEO_CONFIG = {
     "patch_size": [1, 2, 2], "model_type": "ti2v",
@@ -181,7 +194,7 @@ vae_cfg = bagel_cfg.get("vae_config", {})
 
 ## Custom Architecture Patterns
 
-### Split forward API (DreamID-Omni)
+### Split forward API
 
 When a fusion model needs to interleave blocks from two backbones:
 

.claude/skills/add-diffusion-model/references/native-model-integration-checklist.md

@@ -0,0 +1,131 @@
+# Native Diffusion Model Integration Checklist
+
+Use this checklist for native non-Diffusers models, hybrid pipelines, and
+model-specific optimization work.
+
+## 1. Freeze the contract before coding
+
+Record:
+
+- Reference repository and immutable revision.
+- Official checkpoint ID, local layout, license, and required auxiliary files.
+- Pipeline stages: prompt/image encoding, denoising, video/image/audio decoding,
+  and muxing.
+- Token packing, RoPE coordinates, attention sinks/LSE behavior, MoE layout,
+  scheduler equations, CFG formulation, dtypes, and checkpoint transforms.
+- Output geometry, frame count/FPS, audio rate/channels, supported tasks, and
+  default generation steps.
+- Available local hardware versus the official qualification hardware.
+- Intended default deployment and optional optimization paths.
+
+State whether the deliverable is a native implementation or an external
+adapter. For native support, use the reference as an oracle only.
+
+## 2. Use phase gates
+
+| Phase | Required exit evidence |
+|---|---|
+| Architecture | Component/weight inventory; unsupported reference dependencies identified; ownership decisions documented |
+| Native correctness | Strict checkpoint load; tiny forward or one-step golden vs pinned reference; output contract validated |
+| Framework integration | Registry/discovery, shared request types, existing category example, serving route, docs, and recipe work without model-name branching in shared scripts |
+| Lossless distribution | One-rank oracle vs each supported TP/SP/CFG/HSDP/VAE layout; released-checkpoint smoke for the recommended layout |
+| Memory optimization | Ordinary/layerwise/DLO behavior, peak HBM, full-tree host PSS, and stage timeline; shared backend unchanged unless a general contract is proven |
+| Lossy optimization | Cache hit demonstrated; quality delta measured; speed claim made only when warmup permits hits |
+| Full qualification | Default-resolution E2E at official step count for every supported task; media metadata and hashes recorded |
+| Review readiness | Focused tests, Ruff/format, strict docs, DCO, current-main rebase, issue/RFC for large changes, concise PR evidence and limitations |
+
+Do not start performance tuning before native one-step parity. Do not claim a
+feature from topology validation alone when released weights can exercise it.
+
+## 3. Keep ownership narrow
+
+Default placement:
+
+| Concern | Owner |
+|---|---|
+| Architecture, checkpoint remap, packed attention math, nested block alias, auxiliary lifecycle | Model package |
+| Kernel/version dispatch already useful to multiple models | Shared attention utility |
+| Request-dependent example defaults | Typed model-extra entry consumed through a registry |
+| Exact-ID/local-file detection for repositories without standard metadata | Generic native-checkpoint signature table |
+| General offload/cache/HSDP bug | Shared layer plus one focused shared regression |
+
+Before changing shared code, answer:
+
+1. Is this a framework contract or only this checkpoint's layout?
+2. Can the model expose an existing declarative hook instead?
+3. Is there a second consumer or a framework-level failing test?
+4. Can the shared change avoid importing or naming the model?
+
+If the answers do not justify a shared change, keep it model-local.
+
+## 4. Reuse examples without leaking model logic
+
+- Reuse the existing `x_to_y.py` category script.
+- Resolve the pipeline class through standard metadata or native signature
+  discovery, not path substrings.
+- Put dynamic model defaults and validation in `model_extras`; let shared
+  examples consume only the generic typed API.
+- Do not create image/speech/default abstractions without a real consumer.
+- Add a dedicated example only when the shared request/output protocol cannot
+  express the model.
+
+## 5. Build a bounded evidence matrix
+
+Start with the recommended resident layout. Add only supported, useful points:
+
+- Single request: TP, SP, or hybrid TP/SP.
+- Throughput: DLO/DP layouts when concurrent requests are valid.
+- CFG: compare CFG-parallel output with the packed-CFG oracle only if the model
+  actually uses CFG.
+- HSDP: record loading and materialization peaks; it can use more HBM than a
+  resident SP layout.
+- VAE patch parallelism: compare decoded tensors or media exactly where the
+  algorithm should be lossless.
+- Cache: separate cache integration from acceleration evidence.
+
+With four devices, qualify four-device layouts such as DP4, DP2SP2, SP4, or
+CFG2xSP2. Do not present CFG2xSP4 or an official eight-device runtime as locally
+validated.
+
+For each measured layout record:
+
+- Cold E2E and warm user latency.
+- Steady-state wave time and throughput per device.
+- Prompt, denoise, video decode, audio decode, and mux times.
+- Peak allocated/reserved HBM and full-process-tree host PSS.
+- Step count, seed, output geometry, frame/FPS/audio metadata, and parity metric.
+
+Plot latency versus throughput/device. Keep one-step, four-step, and full-step
+results separate because they are not quality-equivalent.
+
+## 6. Keep tests proportional
+
+Prefer a small set with distinct failure ownership:
+
+1. Strict weight-key/dtype/load contract.
+2. Tiny native golden or one-step reference parity.
+3. One parametrized distributed-parity test covering supported layouts.
+4. Pipeline/request/output geometry and serving smoke.
+5. Shared-example/default test when the category script changes.
+6. One shared regression only for a genuine shared-layer fix.
+
+Avoid duplicating the same shape or configuration assertion across
+model-specific files. Keep expensive released-checkpoint evidence in the PR or
+recipe unless the repository's CI tier explicitly requires it.
+
+## 7. Write evidence, not marketing
+
+In the PR description include:
+
+- A compact architecture/data-flow explanation.
+- Default deployment and why it is recommended.
+- Code-change counts by component.
+- Accuracy and distributed-parity evidence tied to commands/scripts.
+- Separate Pareto figures for different step counts.
+- Supported versus validated versus planned optimizations.
+- Hardware gaps and untested official-runtime claims.
+
+Call a result a bounded local qualification unless it includes repeated timing
+and variance. State whether latency includes text/image encoding and all
+decoders. Preserve the reference revision and exact commands needed to
+reproduce the evidence.

.claude/skills/add-recipe/SKILL.md

@@ -0,0 +1,126 @@
+---
+name: add-recipe
+description: Add or update an in-repository vLLM-Omni model recipe with verified task, input, output, hardware, command, feature, and validation contracts. Use when creating files under recipes/, restructuring a recipe after review, documenting a newly supported model, or synchronizing recipe claims with support tables and shared feature guides.
+---
+
+# Add a vLLM-Omni Recipe
+
+## Source the contract
+
+Read [`recipes/TEMPLATE.md`](../../../recipes/TEMPLATE.md) and
+[`recipes/README.md`](../../../recipes/README.md) before editing. Inspect the
+canonical model card/repository, implementation, shared examples, tests, and
+available local qualification evidence.
+
+Do not infer support from a model name, sibling recipe, registry entry, or
+topology validator. Mark unexecuted configurations as configuration-only.
+
+## Route by model family
+
+Read only the contributor and user documentation relevant to the model:
+
+| Family | Contributor guide | User-facing documentation to synchronize |
+|---|---|---|
+| Diffusion | [`adding a diffusion model`](../add-diffusion-model/SKILL.md) | `docs/user_guide/diffusion_features.md`, applicable diffusion feature guides, and shared image/video/audio examples |
+| TTS | [`adding a TTS model`](../add-tts-model/SKILL.md) | `examples/offline_inference/text_to_speech/README.md`, `examples/online_serving/text_to_speech/README.md`, and `docs/serving/speech_api.md` |
+| Omni | [`adding an omni model`](../../../docs/contributing/model/adding_omni_model.md) | model-family offline/online example docs, `docs/serving/chat_completions_api.md`, and `docs/user_guide/feature_compatibility.md` |
+
+All families must update `docs/models/supported_models.md` and the matching row
+in `recipes/README.md`. Do not add a modality-specific support document when
+the repository has no such table; update the closest shared user contract.
+
+## Structure the recipe
+
+Keep the recipe task-oriented and use one hardware-specific file named
+`recipes/<vendor>/<model>-<hardware>.md`. A recipe may contain multiple
+deployment topologies on that hardware, but each additional accelerator model
+or platform requires its own suffixed file.
+
+1. **Summary:** vendor, exact model ID, runtime, modes, named hardware,
+   recommended deployment, and maintainer.
+2. **Supported model contract:** put task, input, output, and provided-profile
+   tables before setup or commands.
+3. **References:** link the canonical model source, shared examples, supported
+   model table, and feature matrix.
+4. **Checkpoint/setup:** pin revisions when possible and state required assets.
+5. **Hardware:** document the single accelerator named by the file suffix,
+   including every validated device-count/topology profile on it.
+6. **Commands:** reuse shared `examples/` entrypoints unless the model contract
+   truly requires a dedicated script. Include offline and online paths that
+   were validated.
+7. **Supported features:** use a compact model-specific topology/status table
+   with links to shared feature guides.
+8. **Verification:** provide a quick command and exact expected output
+   contract.
+9. **Qualification evidence:** report correctness, memory, and timing evidence
+   with its measurement scope and caveats.
+
+## Separate hardware from software
+
+For every locally validated hardware profile, record:
+
+- accelerator vendor/model and per-device memory;
+- number of devices;
+- interconnect (`NVLink`, `PCIe`, or the platform equivalent);
+- host memory when CPU staging/offload is material;
+- whether the profile is runtime-qualified or configuration-only.
+
+Record OS, Python, driver/runtime, framework versions, and the vLLM-Omni
+revision in a separate software-environment table. Keep precision, worker
+count, and DP/TP/SP/PP sizes with the deployment profile or exact command.
+
+Never generalize a result from one accelerator family to another or combine
+NVIDIA, AMD, NPU, or distinct accelerator models in one recipe. Create a
+separate hardware-suffixed recipe and `recipes/README.md` row. Distinguish
+upstream requirements from hardware exercised by the PR.
+
+## Keep shared documentation concise
+
+Link the applicable diffusion, TTS, omni, serving, or design guide for generic
+feature semantics and launch instructions. The recipe should contain only the
+model-specific status, valid topology, required flag difference, and evidence
+boundary.
+
+Put unsupported combinations in the same feature table. Avoid repeating the
+same model-specific prose below the global feature matrix; the recipe is the
+detailed source of truth.
+
+## Report evidence precisely
+
+For every performance or memory value, state:
+
+- checkpoint/revision and workload;
+- accelerator and device count;
+- precision and topology;
+- cold/warm scope, step count, and concurrency;
+- peak HBM and host-memory metric when measured;
+- output/quality guard and sample count;
+- monitoring overhead or other known variance.
+
+Call a single bounded run qualification evidence, not a benchmark. Do not add
+a benchmark script unless the contribution explicitly requires one.
+
+## Synchronize repository documentation
+
+Update all applicable locations:
+
+- the row in `recipes/README.md`;
+- `docs/models/supported_models.md`;
+- the family-specific documentation selected in **Route by model family**;
+- shared example documentation when commands or flags change.
+
+Keep global support tables compact. Link to the recipe instead of duplicating
+its deployment explanation.
+
+## Validate
+
+Run, at minimum:
+
+```bash
+pre-commit run --files <changed recipe/docs/skill files>
+mkdocs build --strict
+git diff --check
+```
+
+Remove generated documentation artifacts after validation. Confirm every
+recipe link resolves and every claimed profile has matching code or evidence.

.claude/skills/add-tts-model/SKILL.md

@@ -138,9 +138,12 @@ A shared buffer silently corrupts audio across concurrent requests — the sympt
    - Load codec weights (may need lazy loading from separate checkpoint)
    - Implement `forward()`: codec codes -> audio waveform
    - Return `OmniOutput` with `multimodal_outputs`
-5. **Create stage config YAML** defining both stages, memory allocation, and model paths
-6. **Create stage input processor** for prompt building
-7. **Write end2end.py** test script
+5. **Define and register the pipeline topology** in
+   `vllm_omni/model_executor/models/<model>/pipeline.py`
+6. **Create the deploy YAML** under `vllm_omni/deploy/` for placement,
+   memory sizing, connectors, and runtime overrides
+7. **Create stage input processor** for prompt building and inter-stage handoff
+8. **Write end2end.py** test script
 
 ### Critical Parameters to Get Right
 
@@ -174,7 +177,7 @@ These bugs appear in almost every new TTS PR. Check all before the first push. S
 - **All return paths must emit `model_outputs`** — if any early-return branch skips setting `model_outputs`, the serving layer silently drops that step's audio (fish speech: `fix: ensure ALL return paths emit model_outputs`)
 - **Per-request state isolation** — for batched concurrent requests, key all state by request ID; a shared buffer corrupts audio across requests (fish speech: `fix: per-request vocode + delta emission`)
 - **Codec tensor device** — move codec codes to the codec decoder's device before calling decode; mismatches cause silent CPU fallback or crashes (fish speech: `fix: use model device for CUDA stream`)
-- **AR stage `max_num_seqs`** — set to at least 4 in production configs; for single-stage models this is the only stage. For two-stage models, Stage 0 (AR) needs `max_num_seqs ≥ 4` to pipeline concurrent requests; Stage 1 (codec decoder) typically uses `max_num_seqs: 1` intentionally. Default of 1 everywhere causes audio gaps under concurrency because the codec window round-robins across requests (RFC #2568)
+- **AR stage `max_num_seqs`** — set to at least 4 in production deploy configs; for single-stage models this is the only stage. For two-stage models, Stage 0 (AR) needs `max_num_seqs ≥ 4` to pipeline concurrent requests; Stage 1 (codec decoder) is model-specific and may intentionally use `max_num_seqs: 1`. Defaulting the AR stage to 1 causes audio gaps under concurrency because the codec window round-robins across requests (RFC #2568)
 
 ### Optional Dependency Handling
 
@@ -190,18 +193,18 @@ pattern, signature constraints, and MOSS-TTS-Nano reference.
 When the upstream model cannot be cleanly split into an AR stage and a
 separate decoder, run the full pipeline inside a single AR worker and
 stream audio through a per-request `inference_stream()` generator keyed by
-`_omni_req_id`. Stage config must set `worker_type: ar`,
-`engine_output_type: audio`, `final_output: true`, `is_comprehension: true`,
-and `async_chunk: false` at the top level. Only extract params from
+`_omni_req_id`. Define one `StagePipelineConfig` with
+`execution_type=StageExecutionType.LLM_AR`, `engine_output_type="audio"`,
+`final_output=True`, and `owns_tokenizer=True`. Set `async_chunk: false` in
+the deploy YAML. Only extract params from
 `additional_information` that you actually forward, or pre-commit fails
 `ruff F841`.
 
 Full walkthrough with the complete `forward()` / `_create_stream_gen()`
-skeleton and stage-config fields:
+skeleton plus pipeline/deploy definitions:
 [references/single-stage-ar.md](references/single-stage-ar.md). For an
 in-tree reference, look for any single-stage AR model under
-`vllm_omni/model_executor/models/` — e.g. the MOSS-TTS-Nano integration when
-it lands.
+`vllm_omni/model_executor/models/`, such as MOSS-TTS-Nano.
 
 **VoxCPM2 is a different pattern** and should not reuse this skeleton — it
 runs the base LM under vLLM PagedAttention with external side-computation.
@@ -210,7 +213,8 @@ See `plan/voxcpm2_native_ar_design.md`.
 ### Deliverables
 
 - Model files in `vllm_omni/model_executor/models/<model_name>/`
-- Stage config YAML
+- Registered `pipeline.py` topology
+- Deploy YAML under `vllm_omni/deploy/`
 - Working `end2end.py` at `examples/offline_inference/text_to_speech/<model>/end2end.py`
 - New section in `examples/offline_inference/text_to_speech/README.md` (table row + per-model section). Do **not** create a top-level `examples/offline_inference/<model>/` dir or a per-model `README.md` inside `text_to_speech/<model>/` — the hub README is the documented surface and the mkdocs `generate_examples` hook only descends one level into `examples/<category>/`.
 
@@ -341,8 +345,8 @@ Heavier scenarios in the same file use **`advanced_model` only** (streaming, ext
 @pytest.mark.tts
 @hardware_test(res={"cuda": "L4"}, num_cards=1)
 @pytest.mark.parametrize("omni_server", tts_server_params, indirect=True)
-def test_voice_clone_en_non_streaming_001(omni_server, openai_client) -> None:
-    openai_client.send_audio_speech_request({...})
+def test_voice_clone_en_non_streaming_001(omni_server, online_client) -> None:
+    online_client.send_audio_speech_request({...})
 ```
 
 **L4 consolidation:** Prefer parametrized `OmniServerParams` rows (`default`, `async_chunk`, feature flags) in one expansion module rather than many merge-only files ([#1832](https://github.com/vllm-project/vllm-omni/issues/1832)).
@@ -359,8 +363,8 @@ def test_voice_clone_en_non_streaming_001(omni_server, openai_client) -> None:
 
 | Path | Fixture | Call |
 |------|---------|------|
-| Online `/v1/*` | `openai_client` | `openai_client.send_*_request(request_config)` |
-| Offline inference | `omni_runner_handler` | `omni_runner_handler.send_*_request(request_config)` |
+| Online `/v1/*` | `online_client` | `online_client.send_*_request(request_config)` |
+| Offline inference | `offline_client` | `offline_client.send_*_request(request_config)` |
 
 1. **Grep `runtime.py` first** — reuse `send_omni_request`, `send_diffusion_request`, `send_audio_speech_request` (online + offline Qwen-style TTS), `send_single_stage_tts_request` (Coqui XTTS / MOSS-TTS-Nano offline), etc.
 2. **No matching helper** → add `send_<feature>_request` (or `send_<route>_http_request` for negative/dfx) in **`runtime.py`** with general `assert_*` bundled inside, **then** call it from the test.
@@ -392,7 +396,7 @@ See [vllm-omni-test skill](../vllm-omni-test/SKILL.md) § **Runtime send helpers
   `/health` return non-200 until you're actually ready.
 - **Mark tests per the CI Levels table** — baseline smoke: `core_model` + `advanced_model`; heavier cases: `advanced_model` only; L4 expansion: `full_model`
 - **No per-model helper modules** — do not create `tests/helpers/{slug}.py`; keep constants and `request_config` payloads in the test file
-- **Online and offline e2e go through `runtime.py`** — `openai_client.send_audio_speech_request` (online); `omni_runner_handler.send_audio_speech_request` (Qwen-style offline) or `send_single_stage_tts_request` (single-stage offline). Add a new `send_*_request` in `runtime.py` when none fits; do not embed `omni.generate` or HTTP in tests
+- **Online and offline e2e go through `runtime.py`** — `online_client.send_audio_speech_request` (online); `offline_client.send_audio_speech_request` (Qwen-style offline) or `send_single_stage_tts_request` (single-stage offline). Add a new `send_*_request` in `runtime.py` when none fits; do not embed `omni.generate` or HTTP in tests
 
 ## Phase 4: Async Chunk (Streaming)
 
@@ -400,21 +404,29 @@ See [vllm-omni-test skill](../vllm-omni-test/SKILL.md) § **Runtime send helpers
 
 ### Steps
 
-1. **Update stage config YAML**:
+1. **Update the pipeline topology** so the producing stage declares its async
+   handoff processor with `async_chunk_process_next_stage_input_func`.
+2. **Update the deploy YAML** to enable async chunk and configure the connector:
    ```yaml
    async_chunk: true
-   codec_chunk_frames: 25      # frames per chunk
-   codec_left_context_frames: 25  # overlap for smooth boundaries
+
+   connectors:
+     connector_of_shared_memory:
+       name: SharedMemoryConnector
+       extra:
+         codec_streaming: true
+         codec_chunk_frames: 25
+         codec_left_context_frames: 25
    ```
-2. **Implement chunk handling in Stage 1**:
+3. **Implement chunk handling in Stage 1**:
    - Accept partial input (chunk of codec codes)
    - Handle left context for smooth audio boundaries
    - Return partial audio in `OmniOutput`
-3. **Test streaming**:
+4. **Test streaming**:
    - Verify audio quality matches non-streaming output
    - Check for artifacts at chunk boundaries
    - Measure TTFA (time to first audio)
-4. **Update online serving** to support `stream=true` with PCM output
+5. **Update online serving** to support `stream=true` with PCM output
 
 ### Streaming Architecture
 
@@ -435,7 +447,8 @@ Stage 0 (AR)                    Stage 1 (Decoder)
 
 ### Deliverables
 
-- Updated stage config with async_chunk enabled
+- Updated pipeline async handoff processor and deploy config with
+  `async_chunk: true`
 - Smooth streaming audio without boundary artifacts
 - TTFA metrics
 
@@ -486,6 +499,13 @@ Common pre-commit failures, recovery commands for missing sign-off, and the
 full `pre-commit run` invocation for a TTS model:
 [references/precommit-dco.md](references/precommit-dco.md).
 
+## Documentation and recipe
+
+Follow the [`add-recipe` skill](../add-recipe/SKILL.md) to add or update the
+in-repository model-family recipe, the `recipes/README.md` index, the supported
+models table, TTS example documentation, and the speech API contract using
+validated evidence.
+
 ## Integration Checklist
 
 Use this checklist when integrating a new TTS model:
@@ -507,14 +527,15 @@ Use this checklist when integrating a new TTS model:
 - [ ] Config classes created with `model_type` registration
 - [ ] Stage 0 (AR) implemented and generates correct tokens
 - [ ] Stage 1 (Decoder) produces correct audio from tokens — dtype float32 for codec decoder
-- [ ] Stage 1 `max_num_seqs` ≥ 4 in production config (default 1 causes gaps under concurrency)
+- [ ] AR stage `max_num_seqs` ≥ 4 in the production deploy config unless the model has a tested lower limit
 - [ ] Optional dependency fallbacks handled at `load_weights()` time (torchaudio/soundfile/etc.)
 - [ ] Streaming: codec codes accumulated across AR steps (not reset per step)
 - [ ] Streaming: delta audio emitted per chunk, not full re-decoded waveform
 - [ ] Streaming: all `forward()` return paths emit `model_outputs`
 - [ ] Streaming: per-request state keyed by request ID (not shared across requests)
 - [ ] Streaming: codec tensors moved to codec decoder device before decode
-- [ ] Stage config YAML created
+- [ ] Pipeline topology defined in `pipeline.py` and registered in `OMNI_PIPELINES`
+- [ ] Deploy YAML created under `vllm_omni/deploy/`
 - [ ] `end2end.py` produces audio matching reference quality
 - [ ] README.md written
 
@@ -529,9 +550,11 @@ Use this checklist when integrating a new TTS model:
 - [ ] Buildkite entries match level: `test-ready.yml` / `test-merge.yml` or nightly TTS job / `test-nightly.yml`
 - [ ] Gradio demo working
 - [ ] Documentation added (offline + online docs, nav, supported models)
+- [ ] In-repository recipe and `recipes/README.md` row added via the `add-recipe` skill
 
 ### Phase 4: Async Chunk
-- [ ] Stage config updated with `async_chunk: true`
+- [ ] Pipeline declares the async handoff processor
+- [ ] Deploy config sets `async_chunk: true` and connector chunk parameters
 - [ ] Stage 1 handles partial chunks correctly
 - [ ] No audio artifacts at chunk boundaries
 - [ ] Streaming via API (`stream=true`) works
@@ -548,7 +571,7 @@ Use this checklist when integrating a new TTS model:
 - [ ] `pre-commit run --files <changed>` passes before every push
 - [ ] Every commit has `Signed-off-by` matching the author email (`git commit -s`)
 - [ ] `git config user.email` matches the email registered on your GitHub account
-- [ ] Details and failure-recovery commands: [references/precommit-dco.md](references/precommit-dco.md)
+- [ ] Hook list, failure table, and DCO recovery: [references/precommit-dco.md](references/precommit-dco.md)
 
 ## References
 

.claude/skills/add-tts-model/references/precommit-dco.md

@@ -30,6 +30,18 @@ but the changes are correct — stage the modified files and re-commit.
 | `ruff F841` | Variable extracted but never forwarded to model call | Remove the extraction or wire it through |
 | `ruff E402` | Import added below function definitions | Move to top-level import block |
 | `ruff format` | Line length, spacing, quote style | Accept auto-fix, stage, re-commit |
+| `check-spdx-header` | Missing header, or copyright still says `vLLM project` | Two-line Omni header (`vLLM-Omni project`); restage the rewrite |
+| `check-forbidden-imports` | Stdlib `re`/`base64`, pickle, Hugging Face Hub API, or direct Triton/TileLang | `import regex as re` and `pybase64`; use `vllm.transformers_utils.repo_utils`. Do **not** add the file to `allowed_files` without review |
+| `check-torch-cuda-call` | New `torch.cuda.*` call site | Use `current_omni_platform` / `OmniPlatform`; do not grow `ALLOWED_FILES` without review |
+| `check-tts-adapter-migration` | New `self._tts_model_type` branch in `serving_speech.py` | Put per-model logic in `tts_adapters/`. Lower `MAX_MODEL_TYPE_BRANCHES` when removing branches; do not raise it without review |
+| `check-mark` | New `tests/**/test_*.py` missing level/hardware mark, or direct `pytest.mark.H100` | Add `core_model`/`advanced_model`/… plus `cpu`/`cuda`/`hardware_test(`; SKU marks only via helpers |
+| `mypy-3.10` | Type error on changed `vllm_omni/` files | Fix the types; model trees are excluded. Extra versions: `pre-commit run --hook-stage manual mypy-3.12` |
+| `markdownlint-cli2` | Docs/README markdown lint | Accept auto-fix under `docs/` / `recipes/` / root README |
+| `check-buildkite` | Invalid `.buildkite/*.yml` | Fix the YAML; do not grow `SKIP_FILES` without review |
+| `shellcheck` | No `shellcheck` on PATH, or script warning | `apt-get`/`dnf`/`brew install shellcheck`, or `shellcheck.exe` on Windows PATH |
+
+Canonical policy, CI `SKIP` list, and allowlist locations:
+[docs/contributing/README.md](../../../docs/contributing/README.md#linting).
 
 ## DCO sign-off
 

.claude/skills/add-tts-model/references/single-stage-ar.md

@@ -86,12 +86,51 @@ class YourModelForCausalLM(nn.Module):
 - Handle both `"audio"` (incremental) and `"result"` (final combined) event
   types from upstream models.
 
-## Stage config
+## Pipeline and deploy config
 
-Single stage with `worker_type: ar`, `engine_output_type: audio`,
-`final_output: true`, `is_comprehension: true`, and `async_chunk: false` at
-the top level. Omitting any of these causes silent misclassification in the
-serving layer.
+Define the fixed single-stage topology in `pipeline.py`:
+
+```python
+from vllm_omni.config.stage_config import (
+    PipelineConfig,
+    StageExecutionType,
+    StagePipelineConfig,
+)
+
+YOUR_TTS_PIPELINE = PipelineConfig(
+    model_type="your_model_name",
+    default_deploy_config_name="your_model_name.yaml",
+    model_arch="YourModelForCausalLM",
+    stages=(
+        StagePipelineConfig(
+            stage_id=0,
+            model_stage="your_model_name",
+            execution_type=StageExecutionType.LLM_AR,
+            input_sources=(),
+            owns_tokenizer=True,
+            engine_output_type="audio",
+            final_output=True,
+            final_output_type="audio",
+        ),
+    ),
+)
+```
+
+Register it in `vllm_omni/config/pipeline_registry.py`. Keep placement and
+runtime sizing in `vllm_omni/deploy/your_model_name.yaml`:
+
+```yaml
+async_chunk: false
+
+stages:
+  - stage_id: 0
+    devices: "0"
+    max_num_seqs: 4
+    gpu_memory_utilization: 0.8
+```
+
+Do not move `execution_type`, output ownership, tokenizer ownership, or model
+architecture into the deploy YAML; those are pipeline topology.
 
 ## Lint discipline
 
@@ -101,8 +140,5 @@ pre-commit.
 
 ## Reference implementation
 
-Look for any single-stage AR model under
-`vllm_omni/model_executor/models/` — e.g. `moss_tts_nano/` when its
-integration lands. If none is in tree yet, follow the skeleton above and
-cross-check against the `is_comprehension: true` / `async_chunk: false`
-dispatch in `vllm_omni/entrypoints/openai/serving_speech.py`.
+Use `vllm_omni/model_executor/models/moss_tts_nano/pipeline.py` and
+`vllm_omni/deploy/moss_tts_nano.yaml` as the in-tree reference.

.claude/skills/find-simplifications/SKILL.md

@@ -0,0 +1,135 @@
+---
+name: find-simplifications
+description: Find evidence-backed simplification candidates in vLLM-Omni and, when requested, turn them into focused proposals or code changes. Use for audits of dead, duplicated, speculative, over-generalized, unnecessarily defensive, or hand-rolled code. Use review-pr for ordinary correctness review and diffusion-perf-opt for performance-first optimization.
+---
+
+# Find vLLM-Omni Simplifications
+
+Find a small number of well-proven ways to reduce code, state, APIs, or maintenance cost without weakening supported behavior. Treat this as an architecture and ownership audit, not a dead-code search or style cleanup.
+
+The default deliverable is a read-only report. Do not add TODOs, design documents, issues, or code unless the user asks for those changes.
+
+## Establish the Contract
+
+Before judging a candidate:
+
+1. Read the repository instructions supplied by the host.
+2. Read [`docs/design/architecture_overview.md`](../../../docs/design/architecture_overview.md).
+3. Select only the relevant active module and feature documents from [`docs/design/index.md`](../../../docs/design/index.md). Archived design pages are historical context, not current contracts.
+4. Identify the comparison base with `git merge-base HEAD origin/main`; do not assume the current branch represents the intended baseline.
+5. If tests or CI coverage may change, use [`vllm-omni-test`](../vllm-omni-test/SKILL.md) for test placement, markers, and commands.
+
+Separate an implementation accident from an intentional boundary. In particular:
+
+- vLLM-Omni extends and sometimes overrides upstream vLLM. A similar upstream implementation is evidence for reuse only after comparing API, lifecycle, and multi-stage semantics.
+- Registries, import strings, model configuration, deployment YAML, plugins, and optional dependencies create dynamic consumers that a static reference count can miss.
+- Hardware backends, model families, execution modes, and connector implementations may be intentionally parallel. Do not propose collapsing one only because it is inactive in the local environment.
+- Stage, replica, request, cache, and transport state can look redundant while representing different owners. Prove that two fields or barriers encode the same fact before merging them.
+- A shorter hot path is not a simplification if it adds synchronization, tensor copies, host materialization, recompilation, or a measurable quality/performance regression.
+
+## Strong Candidates
+
+A strong candidate removes a meaningful surface and has a clear replacement or deletion boundary. Examples include:
+
+- A public method, configuration option, registry entry, route, message field, or event has no production consumer and no compatibility obligation.
+- Tests or documentation are the only consumers of behavior that is not part of a supported contract.
+- Multiple layers mirror the same request, lifecycle, cache-validity, readiness, or topology state.
+- AR, diffusion, platform, or model-specific implementations duplicate a stable common operation and can share it without hiding performance-critical differences.
+- An inherited upstream vLLM primitive is reimplemented locally without an Omni-specific semantic difference.
+- Compatibility or fallback branches protect environments the project no longer supports.
+- Serialization, IPC, output formatting, or connector code materializes the same payload more than once or converts between equivalent representations.
+- A helper, abstraction, package split, or extension point has only one real implementation and adds more indirection than policy.
+- Hand-written infrastructure can be replaced by an existing project dependency, Python standard-library feature, PyTorch primitive, or upstream vLLM facility with net deletion and equal observability.
+- Tests maintain large mocks, snapshots, or expected data solely for an unused API or representation.
+
+Do not elevate typo fixes, formatting, isolated renames, or vague complexity complaints into simplification candidates. Bundle small cleanups only when they support one substantive removal.
+
+## Survey by Ownership Boundary
+
+For broad audits, cover the largest or riskiest production deltas first. Useful domains are:
+
+- `vllm_omni/engine/`: orchestration, stage pools, request state, output convergence, cancellation, and replica lifecycle.
+- `vllm_omni/worker/` and AR model execution: upstream vLLM overlays, scheduler/cache state, collective RPC, adapters, and weight loading.
+- `vllm_omni/diffusion/`: request/step execution, batching, output materialization, IPC, offload, parallelism, and model pipelines.
+- `vllm_omni/entrypoints/`: duplicated validation/defaulting, offline/online parity, streaming, and route compatibility.
+- Connectors and distributed code: ownership, copies, barriers, handles, cleanup, and failure propagation across process/device/node boundaries.
+- Platform backends: genuine vendor differences versus copied control flow that has drifted.
+- Configuration, registries, recipes, and examples: duplicate sources of truth and options that never reach a runtime consumer.
+- Tests and CI: redundant fixtures, mocks of obsolete contracts, and expensive coverage that can move to a cheaper level without losing the guarded behavior.
+
+When the user explicitly requests a repository-wide or many-candidate audit and parallel agent work is available and authorized, divide these domains among agents and require the same evidence format from each. Otherwise survey them sequentially.
+
+## Prove or Reject Each Candidate
+
+Use `rg` first, then read every relevant caller and implementation. Search exact Python symbols plus serialized names, route paths, configuration keys, registry strings, YAML values, and CLI spellings.
+
+Classify evidence:
+
+- **Production:** `vllm_omni/`, runtime configuration, registries, serving entrypoints, connectors, recipes used for supported deployments, and executable loader paths.
+- **Validation/support:** `tests/`, benchmarks, examples, docs, CI, and developer tooling.
+- **External/dynamic:** upstream vLLM contracts, third-party plugins, model repositories, optional backends, and string-loaded objects. State what was checked and what remains unknown.
+
+For each candidate, answer:
+
+1. What exact surface would disappear or merge?
+2. Who owns the state or behavior today?
+3. Which production consumers exist, including dynamic consumers?
+4. What supported behavior, compatibility, performance, memory, or observability could change?
+5. What is the smallest validation that would catch a bad simplification?
+6. Is the result net deletion after tests, adapters, migration glue, and documentation are counted?
+
+Reject or downgrade the candidate when a live consumer exists and removal is actually a feature decision; when an active design contract explains the separation; when the change only moves complexity; or when hardware/model evidence is unavailable for a performance-sensitive claim.
+
+## Lifecycle and Data-Movement Audits
+
+For asynchronous or distributed code, map the ownership graph before recommending deletion:
+
+```text
+request admission -> orchestrator -> stage/replica -> worker/device
+  -> connector or IPC boundary -> output aggregation -> terminal cleanup
+```
+
+Associate every lock, event, sentinel, queue, readiness flag, timeout, abort path, and cache marker with an owner and transition. Coalesce mechanisms only when they represent the same state and have the same failure boundary. Preserve separate mechanisms when they protect publication versus rollback, local versus remote ownership, first-terminal-outcome arbitration, or cleanup after partial failure.
+
+For tensors and media, also trace device, process, and ownership transitions. Count GPU-to-CPU copies, shared-memory/object-store handles, serialization, decode/encode steps, and unlink/release responsibility. Prefer eliminating a transfer or representation over merely hiding it behind a helper.
+
+## Dependency Replacements
+
+Prefer facilities already available in the repository before adding a dependency. For a proposed replacement:
+
+- Name the exact local behavior it covers and any semantics that remain in glue code.
+- Check maintenance, license, optional-dependency boundaries, platform support, and transitive footprint.
+- Compare net deletion, test burden, import/startup cost, and runtime behavior.
+- Benchmark when the replaced code is on a model, scheduler, connector, serialization, or serving hot path.
+
+A wrapper that retains the same state machine or conversion logic is not a simplification.
+
+## Deliver the Result
+
+Report only candidates supported by concrete evidence. For each one include:
+
+- **Surface:** files and symbols involved.
+- **Evidence:** production consumers found, dynamic paths checked, and relevant design contract.
+- **Simplification:** exactly what to remove, merge, inline, or reuse.
+- **Benefit:** deleted API/state/code, reduced data movement, or reduced maintenance/test burden.
+- **Risk and validation:** behavior that could change and the narrowest useful tests or measurements.
+- **Confidence:** high or medium, with the unresolved evidence named for medium-confidence items.
+
+Also list representative ideas that were rejected when that prevents repeated investigation. Prefer three high-confidence candidates over a long speculative inventory.
+
+If the user asks for a durable design proposal, place it under `docs/design/feature/` and register it in `docs/design/index.md`; use current design-document conventions and keep implementation details proportional to the decision. If the user asks for implementation, make the smallest coherent change, update affected tests/docs, and preserve public compatibility unless the user explicitly accepts a breaking change.
+
+## Folding Work from Another Branch
+
+Diff the source branch against its own merge base with `origin/main`, not against the current feature branch. Port only independently justified candidates. Consolidate overlaps into the candidate with stronger evidence, and do not preserve a candidate count by carrying weaker duplicates.
+
+## Validation
+
+Match validation to the outgoing diff:
+
+- Documentation or skill changes: run the repository skill validator when available, check links/paths, search for stale source-specific terminology, and run `git diff --check`.
+- Python changes: run the narrowest relevant pytest coverage through the repository environment, then lint/format the changed files.
+- Distributed, model, hardware, performance, or accuracy changes: run available CPU/static checks and state the exact GPU/NPU/model gap; never present simulated evidence as device validation.
+- Before submission, use [`precheck-pr`](../precheck-pr/SKILL.md) for the branch-level self-review.
+
+When summarizing, distinguish checks that passed from checks that were not runnable, and state the snapshot or branch against which the audit was performed.

.claude/skills/precheck-pr/SKILL.md

@@ -1,6 +1,6 @@
 ---
 name: precheck-pr
-description: Self-check your branch before creating a PR — catch dead code, verify accuracy/perf claims, validate PR title format, and confirm merge readiness. Use when the user says "precheck", "self review", "pre-submit check", or "check my PR before I open it." Never posts to GitHub.
+description: Self-check your branch before creating a PR — catch dead code, prevent new model-specific Python examples, verify accuracy/perf claims, validate PR title format, and confirm merge readiness. Use when the user says "precheck", "self review", "pre-submit check", or "check my PR before I open it." Never posts to GitHub.
 ---
 
 # PR Pre-Check
@@ -65,6 +65,12 @@ Ask: "Quick mode or full mode?" Then walk the checklist for the detected PR type
 
 **Also run the Code-Quality sweep on every PR**, regardless of type or mode: the five diff-scoped checks in [references/code-quality.md](references/code-quality.md) — kwargs fragility, broad-except swallow, `Any`/wrong type hints, hot-path `.clone()`/`deepcopy`, and event-loop blocking — plus the advisory conventions (log level, structured logging, synchronization, cleanup, dependencies, naming) in [checklists.md](references/checklists.md). These count only lines the PR *adds* — the pre-existing backlog across the repo is out of scope.
 
+**Also run the Examples-Policy check on every PR** using [references/examples-policy.md](references/examples-policy.md). Inspect only Python paths introduced by the diff. Treat a new model-, checkpoint-, vendor-, or family-specific Python example as blocking; do not report pre-existing example debt.
+
+**Also run the Simplification pass on every PR** using [find-simplifications](../find-simplifications/SKILL.md). Keep it diff-scoped: inspect added or modified public APIs, state, abstractions, fallback and compatibility paths, data movement, and only the adjacent ownership needed to prove a candidate. Do not expand the pass into a repository-wide audit. Zero candidates is a valid pass; report only evidence-backed opportunities. Mark a candidate as blocking only when it proves a correctness or merge-readiness problem, such as newly unreachable code. Otherwise report it as a warning.
+
+**Also confirm local pre-commit gates** from [docs/contributing/README.md](../../../docs/contributing/README.md#linting). GitHub Actions `SKIP` does not prove they passed. Full new-hook list: Omni SPDX (`vLLM-Omni project`); forbidden imports (stdlib `re`/`base64`, pickle, Hugging Face Hub API, Triton/TileLang); no new `torch.cuda` call sites; shellcheck (macOS/Windows must install the binary); mypy-3.10; test files have CI level + hardware marks; TTS `_tts_model_type` branches in `serving_speech.py` must not increase; markdownlint on `docs/`; Buildkite YAML schema. ✗ for growing `allowed_files` / `ALLOWED_FILES` or raising `MAX_MODEL_TYPE_BRANCHES` without review.
+
 ### Step 5: Print Report
 
 ```
@@ -77,6 +83,8 @@ Pre-check report for <branch>
   ─────────────────  ──────
   PR title format    ✓
   Code quality       ⚠ 1 broad except, 2 Any hints
+  Examples policy    ✗ new model-specific Python example
+  Simplification     ⚠ helper duplicates an existing owner
   PR desc integrity  ✓
   Registry/config    ✓
   Dead code          ⚠ 2 warnings

.claude/skills/precheck-pr/references/checklists.md

@@ -47,6 +47,19 @@ Diff-scoped sweep for five fragility patterns that are pervasive in the existing
 
 Roll the ⚠/✗ counts into the report as a single **Code quality** dimension row.
 
+---
+
+## All PRs: Examples Policy (run on every PR, quick & full)
+
+Inspect Python files introduced under `examples/` using
+[examples-policy.md](examples-policy.md). This is a diff-scoped ratchet: existing
+model-specific examples may be edited or removed without triggering the check.
+
+- [ ] **No new model-specific Python examples:** the diff adds, copies, or renames no Python file whose path or behavior is specific to one model, checkpoint, vendor, or model family. ✗ for any violation. Route inference through a shared task/protocol entrypoint, model request behavior through `vllm_omni/model_extras`, and commands through task docs or `recipes/`.
+- [ ] **New generic example code is justified:** any new task-neutral Python file under `examples/` is an actual user-facing entrypoint rather than reusable production logic, a benchmark, an app, a setup tool, or a reproducer. ⚠ when the placement is questionable.
+
+Roll the result into the report as a separate **Examples policy** dimension.
+
 ### Conventions (eyeball — no grep)
 
 Judgment calls that don't grep cleanly — apply them while reading the diff, not as a mechanical check:
@@ -151,6 +164,7 @@ Diffusion models live under `vllm_omni/diffusion/`. In addition to the New Model
 - [ ] **PR body matches diff:** description doesn't claim changes not in the diff
 - [ ] **Branch is rebased:** `git merge-base HEAD origin/main` is recent, no merge conflicts
 - [ ] **CI gates passing:** DCO, pre-commit, build — check `gh pr view --json statusCheckRollup` if a PR already exists
+- [ ] **Local pre-commit gates:** full list in [Linting](../../../../docs/contributing/README.md#linting). Blocking when the diff introduces: stale `vLLM project` SPDX; stdlib `re`/`base64` or pickle in library code; new `torch.cuda` call sites; `tests/**/test_*.py` without level+hardware marks; new `_tts_model_type` branches in `serving_speech.py`; invalid `.buildkite` YAML. macOS/Windows need a native `shellcheck`. ✗ for growing `CHECK_IMPORTS[*].allowed_files`, `ALLOWED_FILES`, or `MAX_MODEL_TYPE_BRANCHES` without review.
 
 ### Full (adds)
 

.claude/skills/precheck-pr/references/examples-policy.md

@@ -0,0 +1,68 @@
+# Examples Python Policy
+
+Prevent new model-specific Python examples without forcing cleanup of the
+existing backlog. Run this check on every PR in quick and full modes.
+
+## Scope
+
+Set the merge base as in the main workflow, then list Python paths introduced
+by the PR, including additions, copies, and renames:
+
+```bash
+BASE="$(git merge-base HEAD origin/main)"
+git diff --name-status --diff-filter=ACR --find-renames --find-copies \
+  "${BASE}"...HEAD -- 'examples/**/*.py'
+```
+
+Inspect destination paths for copies and renames. Existing model-specific files
+that are only modified or deleted are grandfathered and must not be reported.
+
+## Blocking classification
+
+Mark ✗ when a newly introduced Python file does any of the following:
+
+- Names or scopes a path to one model, checkpoint, vendor, or model family.
+- Hard-codes a model identifier, model-specific prompt, request shape, output
+  conversion, or launch configuration behind an otherwise generic filename.
+- Duplicates an existing task/protocol runner to demonstrate one model.
+- Adds a per-model helper below a generic task hub such as `text_to_speech/`.
+
+Do not infer safety from a generic filename alone. For example, a file named
+`text_to_audio.py` is still model-specific if it only implements one model's
+contract.
+
+## Acceptable additions
+
+A new Python example may pass when it is a genuinely model-neutral,
+user-facing task or protocol entrypoint. It must accept model selection through
+configuration and keep model-specific behavior out of the script.
+
+Use ⚠ when the file is model-neutral but reusable logic appears misplaced under
+`examples/`. Route code by purpose:
+
+| Code purpose | Preferred location |
+|---|---|
+| Model prompt, defaults, request/output adaptation | `vllm_omni/model_extras/` or production model modules |
+| Runnable model command and validation evidence | Task documentation or `recipes/` |
+| UI or interactive application | `apps/` |
+| Benchmark or evaluator | `benchmarks/` |
+| Regression reproducer | `tests/` |
+| Download, conversion, or setup utility | `tools/` |
+
+If a model cannot use an existing shared runner, report the missing generic
+capability. Do not solve the gap by adding a model-specific Python example.
+
+## Report format
+
+Add one row to the pre-check report:
+
+```text
+Examples policy    ✓ no new Python example paths
+Examples policy    ⚠ generic helper belongs in tools/
+Examples policy    ✗ examples/offline_inference/new_model/end2end.py is model-specific
+Examples policy    ✗ examples/offline_inference/x_to_y_model_name.py is model-specific
+```
+
+For a blocker, name the path, the model-specific behavior, and the appropriate
+shared runner or destination. Do not propose deleting unrelated existing
+examples as part of the PR.

.claude/skills/production-add-diffusion-model/SKILL.md

@@ -0,0 +1,458 @@
+---
+name: production-add-diffusion-model
+description: Productionize a vLLM-Omni diffusion model after its Day-0 vertical slice works. Use when work requires official API parity and input limits, feature-combination evidence, online FP8, distributed layerwise offload, per-request Cache-DiT, CUDA/ROCm/NPU/XPU recipes, faster or sparse attention, fused operators, USP or disaggregation analysis, continuous batching and abort handling, long-running RPS stability, or accuracy/performance/reliability CI. For initial architecture porting, registry wiring, and basic weight loading, use add-diffusion-model first.
+---
+
+# Productionizing a Diffusion Model
+
+## Scope
+
+Use this skill to move a working model integration from Day-0 support to a
+measured, fail-closed production deployment. It is deliberately separate from
+[`add-diffusion-model`](../add-diffusion-model/SKILL.md); do not modify or copy
+that skill when productionizing a model. If registry, pipeline, or basic loader
+work is incomplete, read that skill first and finish the Day-0 vertical slice.
+
+MiniMax-H3 [PR #5691](https://github.com/vllm-project/vllm-omni/pull/5691)
+is the merged Day-0 case study. Its open optimization
+[issue #5700](https://github.com/vllm-project/vllm-omni/issues/5700) is an
+actively maintained roadmap and evidence map, not a list of already supported
+features. It can lag merged implementation: read its current revision, cited
+PRs, target source, and maintained recipes, then leave every unverified row
+`not tested` until the target revision has scoped evidence.
+
+Read these references before the corresponding gate:
+
+- [API and hardware recipes](references/api-and-recipes.md)
+- [Production feature patterns](references/feature-patterns.md)
+- [Performance patterns and code examples](references/performance-patterns.md)
+- [Serving, soak, and CI validation](references/production-validation.md)
+
+## Non-negotiable evidence contract
+
+Create a matrix whose row key is at least:
+
+```text
+task x API mode x execution mode/capacity x shape/schedule x attention backend x
+packed layout x cache policy x quantization x offload mode x topology x
+hardware x dtype x output representation/transport
+```
+
+Use exactly these states:
+
+| State | Meaning |
+|---|---|
+| `validated` | The exact row passed correctness and measured deployment tests, with reproducible artifacts. |
+| `limited` | A bounded subset passed; the limitation and rejection/fallback behavior are explicit. |
+| `unsupported` | A stable architectural/platform restriction is proven and fails early with an actionable error. |
+| `not tested` | No adequate evidence exists. This is the default. |
+
+Never infer compatibility from import success, server startup, another task,
+another card, or another topology. Record the model/checkpoint revision,
+vLLM-Omni commit, commands, request assets and hashes, seed/schedule/shape,
+raw output artifacts, environment, and result for every `validated` row.
+
+## Workflow
+
+Follow the gates in order. Keep a dense BF16 single-device path as the
+correctness oracle until every advertised fast path has parity evidence.
+
+### Gate 0: Freeze the official contract
+
+Pin the official implementation and checkpoint revision. Build a matrix for:
+
+- constructor/call arguments, defaults, scheduler and sigma schedule;
+- task-to-source modality/count/order rules;
+- prompt, MIME, bytes, pixels, frames, duration, FPS, steps, seed, guidance,
+  output count, and output format limits;
+- offline API and every public sync/async serving endpoint;
+- checkpoint partitions and task selection.
+
+Normalize offline, sync, and async requests through one validation contract.
+Reject an invalid task/source combination before downloads, upload persistence,
+temporary files, async job creation, or engine submission. Use 400 for invalid
+semantics and 413 for payload limits; clean partial resources on every failure.
+See [API and hardware recipes](references/api-and-recipes.md).
+
+### Gate 1: Re-prove correctness under production shapes
+
+1. Make loading strict: track expected, loaded, unexpected, and intentionally
+   ignored tensors. A fused destination is complete only after every source
+   shard arrives. Missing retained parameters abort startup.
+2. Fix reference revision, asset hashes, seed, schedule, and shape. Compare
+   component outputs, intermediate denoise latents, and final artifacts.
+3. Test unequal packed samples. Verify `cu_seqlens`, padding exclusion,
+   modality/CFG ownership, split/gather boundaries, and one output per request.
+4. Test partial process groups with non-member ranks and group size smaller
+   than world size. Do not construct a private group per component or request.
+5. Run negative API cases before positive E2E tests to prove side-effect-free
+   rejection and cleanup.
+
+### Gate 2: Wire shared fast primitives
+
+Prefer explicit, testable vLLM-Omni/vLLM operators with native fallbacks.
+`torch.compile` remains a separately validated regional optimization; it is
+not a substitute for correct operator selection or lifecycle-safe boundaries.
+
+#### RMSNorm + RoPE baseline
+
+Use the shared cross-platform operators, preserving the official Q/K norm and
+RoPE order:
+
+```python
+from vllm_omni.diffusion.layers.norm import RMSNorm
+from vllm_omni.diffusion.layers.rope import (
+    RotaryEmbedding,
+    apply_rope_to_qk,
+)
+
+self.norm_q = RMSNorm(head_dim, eps=eps)
+self.norm_k = RMSNorm(head_dim, eps=eps)
+self.rope = RotaryEmbedding(is_neox_style=False)
+
+query = self.norm_q(query)  # [B, S, Hq, D] or packed [T, Hq, D]
+key = self.norm_k(key)      # [B, S, Hkv, D] or packed [T, Hkv, D]
+query, key = apply_rope_to_qk(
+    self.rope,
+    query,
+    key,
+    (cos, sin),
+)
+```
+
+Verify `is_neox_style`, `half_head_dim`, partial rotary dimensions, cos/sin
+layout, and packed row ownership against the official model. `False` means
+interleaved/GPT-J style; do not assume the default is NeoX. Shared RMSNorm and
+RoPE are two dispatched operators, not automatically one fused kernel.
+
+For packed non-interleaved RoPE, use the shared fused boundary when its contract
+matches the model:
+
+```python
+from vllm_omni.diffusion.layers.fused_qk_norm_rope import fused_qk_norm_rope
+
+query, key = fused_qk_norm_rope(
+    query,                    # [T, Hq, D]
+    key,                      # [T, Hkv, D]
+    self.norm_q.weight,       # [D]
+    self.norm_k.weight,       # [D]
+    rope_table,               # [T, rotary_dim] = [cos | sin]
+    self.norm_q.variance_epsilon,
+)
+```
+
+The current CUDA fast path specializes BF16 `head_dim=128`, `rotary_dim=96`;
+the public function keeps an eager fallback for unsupported inputs. Verify the
+packed row/frequency contract and trace the actual fast path. Do not reshape an
+incompatible official RoPE layout merely to enter this kernel.
+
+MiniMax-H3-style partial NeoX RoPE rotates a prefix and passes the suffix
+through:
+
+```python
+self.q_norm = RMSNorm(head_dim, eps=eps, dtype=torch.bfloat16)
+self.k_norm = RMSNorm(head_dim, eps=eps, dtype=torch.bfloat16)
+self.rope = RotaryEmbedding(is_neox_style=True, half_head_dim=False)
+
+def apply_partial_rope(self, x, freqs, rot_dim):
+    x_rot, x_pass = x[..., :rot_dim], x[..., rot_dim:]
+    cos, sin = torch.cos(freqs).to(x.dtype), torch.sin(freqs).to(x.dtype)
+    return torch.cat((self.rope(x_rot, cos, sin), x_pass), dim=-1)
+```
+
+The partial-RoPE snippet assumes `import torch` and lives on the owning
+attention module; adapt names and validated frequency shapes rather than
+copying it at module scope.
+
+Call the modules normally so `CustomOp` selects CUDA/HIP/NPU/XPU/native
+implementations. Do not call `forward_cuda()` directly. Compare the shared
+path with native BF16 at operator, block, denoise-trajectory, and artifact
+levels before claiming either accuracy or speed.
+
+#### Faster BF16 attention and fused projections
+
+Use local TP head counts and role-aware shared attention:
+
+```python
+from vllm.model_executor.layers.linear import QKVParallelLinear
+from vllm_omni.diffusion.attention.layer import Attention
+
+self.qkv_proj = QKVParallelLinear(
+    hidden_size=hidden_size,
+    head_size=head_dim,
+    total_num_heads=num_heads,
+    total_num_kv_heads=num_kv_heads,
+    bias=False,
+    quant_config=quant_config,
+    prefix=f"{prefix}.qkv_proj",
+)
+self.attn = Attention(
+    num_heads=self.qkv_proj.num_heads,
+    num_kv_heads=self.qkv_proj.num_kv_heads,
+    head_size=head_dim,
+    softmax_scale=head_dim**-0.5,
+    causal=False,
+    qkv_layout="BSND",
+    role="self",
+    prefix=prefix,
+)
+```
+
+Prove the selected backend from runtime logs/trace and compare it with dense
+SDPA. Packed metadata must express real sample boundaries; do not build masks
+that the selected backend ignores. A backend that is faster for one shape or
+platform is not a global default. See [performance patterns](references/performance-patterns.md)
+for QKV, SwiGLU, AdaLN, sparse attention, redundancy, and fusion examples.
+
+### Gate 3: Add online FP8 without breaking the loader
+
+Route the active config to each component, then pass a stable runtime prefix
+and config into every quantizable vLLM linear:
+
+```python
+def resolve_component_quant_config(quant_config, component):
+    return quant_config.resolve(component) if hasattr(quant_config, "resolve") else quant_config
+
+transformer_quant_config = resolve_component_quant_config(
+    od_config.quantization_config,
+    "transformer",
+)
+self.transformer = MyDiT(
+    od_config,
+    quant_config=transformer_quant_config,
+    prefix="transformer",
+)
+
+self.qkv_proj = QKVParallelLinear(
+    hidden_size=hidden_size,
+    head_size=head_dim,
+    total_num_heads=num_heads,
+    total_num_kv_heads=num_kv_heads,
+    quant_config=quant_config,
+    prefix=f"{prefix}.qkv_proj",
+)
+```
+
+Use `build_quant_config()` for direct construction/testing:
+
+```python
+from vllm_omni.quantization import build_quant_config
+
+quant_config = build_quant_config({
+    "transformer": {"method": "fp8"},
+    "text_encoder": None,
+    "vae": None,
+    "default": None,
+})
+```
+
+Preserve each parameter's vLLM `weight_loader`; perform checkpoint layout
+conversion before invoking it. Account for every QKV/gate-up source shard.
+Keep precision-sensitive norm/modulation/embedders in BF16 unless separately
+proven, using the current `safe_quant_config` pattern where applicable.
+
+Validate resident DiT FP8 first: prove which modules are quantized, HBM saved,
+BF16-vs-FP8 trajectory/artifact quality, latency, and throughput. Text encoder,
+VAE, other hardware, TP/cache/DLO combinations are separate rows.
+Pre-quantized, pruned, rotated, or adapter-modified checkpoints are separate
+model/loading lanes; compare them with the released BF16 model without calling
+the result a pure quantization ablation. Read
+[`../quantization/SKILL.md`](../quantization/SKILL.md) if present and prefer the
+target revision's implementation/docs when sibling guidance differs.
+
+### Gate 4: Make every component offloadable
+
+Declare topology instead of relying on heuristic discovery:
+
+```python
+from typing import ClassVar
+
+import torch.nn as nn
+
+from vllm_omni.diffusion.models.interface import SupportsComponentDiscovery
+from vllm_omni.diffusion.offloader import OffloadPlan
+
+class MyPipeline(nn.Module, SupportsComponentDiscovery):
+    _dit_modules: ClassVar[list[str]] = ["transformer"]
+    _encoder_modules: ClassVar[list[str]] = ["text_encoder"]
+    _vae_modules: ClassVar[list[str]] = ["vae"]
+    _resident_modules: ClassVar[list[str]] = []
+    _offload_plan: ClassVar[OffloadPlan] = OffloadPlan(
+        on_demand_component_paths=frozenset({"text_encoder", "vae"}),
+        block_attrs={"transformer": ("blocks",)},
+        offload_submodules={"token_refiner": "blocks"},
+        resident_dit_paths=frozenset({"transformer"}),
+        encoder_block_attrs={"text_encoder": ("encoder.layers",)},
+    )
+```
+
+Add a test that resolves every dotted component/block path and fails if it is
+missing or is not the expected module/container; discovery may otherwise warn
+and skip. Validate ordinary layerwise offload and distributed layerwise
+offload (DLO) independently.
+
+Keep component lifecycle and host-weight storage separate. `OffloadPlan`
+declares what can be streamed; the diffusion loader owns any `HostWeightPlan`
+and hands the exact prevalidated plan to DLO. Do not make the offloader rescan
+checkpoint files or duplicate loader name/shape/dtype decisions. Direct mmap is
+a fail-closed optimization: current preflight requires TP1, no HSDP, no online
+quantization, complete bindings, and represented transforms. Otherwise the
+ordinary loader remains authoritative.
+
+If the target revision supports Host Weight Runtime (HWR), treat it as a
+separate, exact-identity startup/cache path rather than generic offload or
+zero-copy execution. Qualify `preferred` population/fallback and `required`
+consume-only failure independently; verify immutable manifests, final-layout
+restore, lease cleanup, corruption/quarantine, and node-local sharing. Current
+model contracts may be BF16 no-AllGather-only, so reject AllGather, online
+quantization, HSDP, LoRA/adapted weights, and unrecognized load formats unless
+the exact target revision explicitly adds them.
+
+On small-HBM cards, record load/materialization peak, resident peak, encode,
+each denoise wave, decode, transient per-rank HBM, host PSS, H2D, and collective
+time. Test DLO AllGather and `--dlo-no-use-allgather` as different deployments.
+Sweep the supported resident-block policies and publish the non-dominated
+latency-HBM frontier; a single minimum-memory or fastest point is not the DLO
+trade-off. Keep request-specific preparation replica-local. Only weight
+materialization may enter the declared DLO collective group; never use WORLD
+for per-request work in a DP deployment.
+Direct checkpoint mmap preflight is limited to TP1 without HSDP or online
+quantization; TP>1 falls back to the ordinary TP-aware loader and can still feed
+DLO AllGather or no-AllGather after scoped E2E. HSDP remains incompatible with
+DLO AllGather. A target revision may allow finalized per-tensor online FP8
+through the ordinary loader and AllGather; keep every other online quantizer
+fail-closed, and require model/card/topology E2E before promotion.
+Use deploy YAML for DP under `--omni`; vLLM DP CLI flags are rejected.
+
+### Gate 5: Add request-scoped acceleration safely
+
+Server-wide Cache-DiT support does not imply a per-request quality API. Only
+add request policy after defining calibrated quality tiers and a safe hook
+lifecycle using `SupportsRequestScopedCacheDiT`, `CacheDiTRequestSpec`, and
+`RequestScopedCacheDiTRuntime`. Validate alternating and concurrent quality
+tiers, real cache hits, refresh/reset, success, error, disconnect, abort, and
+the next uncached request. Do not batch different quality policies unless the
+batch compatibility key and hook ownership make that safe.
+
+Then test Cache-DiT against FP8, DLO, TP/SP, compile, and batching one
+combination at a time. See [production feature patterns](references/feature-patterns.md)
+for the lifecycle skeleton and compatibility matrix.
+
+### Gate 6: Optimize only from profiles
+
+Freeze one canonical workload manifest before profiling. Keep lossless runtime
+A/Bs, accelerated paths with numerical/quality trade-offs, and production
+topology studies in separate result lanes. Reconcile the client boundary as
+queue + encode + denoise + video/audio decode + output transport/codec +
+residual; do not hide a material residual in another stage.
+
+Profile stage time, GPU kernels, CPU/GPU synchronization, allocations, H2D,
+output payload/copies, and collectives before choosing work. Apply and A/B one
+change at a time:
+
+1. Remove synchronization and redundant denoise-loop work: hoist static prompt
+   conditioning/token refinement, RoPE tables, masks/metadata, bounded AdaLN
+   schedule projections, and reference scans; remove per-layer `.item()` and
+   unused masks/allocations.
+2. Select the fastest correct BF16 attention backend for each role and shape.
+3. Prefer explicit fused QKV, gate-up + `SiluAndMul`, RMSNorm/RoPE, AdaLN, and
+   layout/residual operators with guarded native fallbacks. Preserve every
+   materialized dtype/rounding boundary consumed by the unfused graph.
+4. Enable sparse attention only when platform, shape, topology, realized
+   sparsity, quality, fallback, and end-to-end speedup are all proven.
+5. Profile USP all-to-all count/bytes, packing, contiguous copies, overlap, and
+   scaling efficiency. Audit replicated cross-attention before using
+   `skip_sequence_parallel`. Qualify regular and accelerated Ulysses transport
+   independently; activation logs, JIT/readiness time, workspace growth,
+   stream ownership, maximum-shape warmup, and numerical drift are gates.
+   Preserve the checkpoint's Q:KV head ratio when padding GQA for Ulysses, and
+   validate strided QKV staging rather than assuming packed contiguity.
+6. Evaluate text-encoder or VAE disaggregation only if stage share, transfer
+   volume, reuse/concurrency, and independent scaling justify it. Tiling,
+   offload, or patch parallelism is not disaggregation.
+
+Do not combine independent optimizations into one performance claim. Preserve
+raw before/after runs with fixed workload and warmup policy.
+
+### Gate 7: Make serving interruptible and load-stable
+
+Treat request-mode batching and step-wise execution as separate capabilities:
+
+- request batching requires `supports_request_batch = True` and
+  `DiffusionRequestBatch -> list[DiffusionOutput]`, exactly one output per
+  request;
+- step execution requires the complete `SupportsStepExecution` contract:
+  `prepare_encode`, `denoise_step`, `step_scheduler`, and `post_decode`.
+
+Keep all mutable scheduler, RNG, latent, mask, cache, and step state request
+scoped. First validate `--step-execution --max-num-seqs 1` and step-boundary
+abort; only then test heterogeneous multi-request waves. Step continuous
+batching is experimental and is not automatically compatible with Cache-DiT.
+Before recommending it, predeclare a useful case and success threshold, then
+A/B it against request mode under the same arrival process. Structural support
+or a generic bridge is not latency, throughput, or cancellation evidence for a
+model.
+
+Test queued and in-flight abort, client disconnect, timeout, OOM, worker error,
+and restart. Require one terminal result, idempotent cleanup, no post-decode
+after abort, no cache/temp/VRAM/request-ID leak, and a successful next request.
+Inject late worker results/exceptions after cancellation and prove the result
+pump stays alive instead of raising `InvalidStateError`. For large batched or
+long-window video outputs, freeze dtype, range, layout, contiguity, ownership,
+payload size, serializer/codec limits, and offline-versus-HTTP semantics.
+Validate device-side output preparation, D2H/IPC or shared handles, remote
+encoding, event-loop responsiveness, and client materialization as separate
+stages, including abort and consumer failure cleanup.
+A model/VAE callback that publishes ordered chunks is only a producer contract;
+it is not transport, backpressure, cancellation, public streaming, or
+time-to-first-frame evidence. Distinguish complete-response faster-than-playback
+(`client E2E / output duration <= 1`) from streaming and report first-chunk,
+steady cadence, finalization, and complete-artifact latency separately.
+Run below-, near-, and above-saturation mixed-RPS soaks and report success/error
+rate, throughput, queue time, memory slope, temp growth, and worker health.
+Report p50/p95/p99 only from a declared arrival model with enough measured
+requests for the claimed percentile. See
+[serving and CI validation](references/production-validation.md).
+
+### Gate 8: Publish recipes and four independent CI tracks
+
+For every target vendor/card/topology, publish exact environment, serve/deploy
+command, one complete JSON or multipart curl per supported task, expected
+output checks, benchmark command, warmup/repeat/concurrency policy, raw metrics,
+per-rank HBM/host PSS, accuracy artifact, and limitations. Mark each
+CUDA/ROCm/NPU/XPU row `validated`, `unsupported`, or `not tested`; never inherit
+support from the platform abstraction or another card. Re-run DLO on each
+small-HBM recipe.
+
+Maintain separate CI gates:
+
+1. Function: strict load, API/offline positive and negative contracts.
+2. Accuracy: fixed reference revision/assets/seed/schedule and scoped tolerance.
+3. Performance: fixed best deployment plus memory-constrained/DLO row, raw JSON,
+   explicit warmup, enough samples for each claimed statistic, and owned
+   regression thresholds.
+4. Reliability: long mixed-RPS soak, abort/disconnect/fault cleanup, memory and
+   temporary-resource trend.
+
+## Production Definition of Done
+
+A model is production-ready only when:
+
+- every official advertised task passes API parity, offline/online parity,
+  strict loading, fixed-reference quality, and negative validation;
+- the recommended deployment row passes correctness, isolation, abort/error
+  cleanup, benchmark, and soak gates on named hardware;
+- online FP8 and DLO are each implemented and either validated in scoped rows
+  or explicitly limited/unsupported with fail-fast behavior;
+- per-request cache quality tiers have measured speed/quality frontiers and
+  request isolation, if advertised;
+- CUDA/ROCm/NPU/XPU recipes describe evidence rather than inferred support;
+- Function, Accuracy, Performance, and Reliability CI have owners, artifacts,
+  and actionable failure output;
+- accelerated paths identify every precision/quality trade-off, and the raw
+  offline plus online encoded output contracts are versioned and tested;
+- all unverified task/backend/cache/quant/offload/topology/hardware combinations
+  remain `not tested` and are not presented as supported.
+
+Day-0 support is still a valid milestone, but title, docs, and matrix must say
+which representative task passed and which production gates remain.

.claude/skills/production-add-diffusion-model/agents/openai.yaml

@@ -0,0 +1,4 @@
+interface:
+  display_name: "Production Add Diffusion Model"
+  short_description: "Productionize vLLM-Omni diffusion models"
+  default_prompt: "Use $production-add-diffusion-model to take this integrated diffusion model from Day-0 support to evidence-backed production readiness."

.claude/skills/production-add-diffusion-model/references/api-and-recipes.md

@@ -0,0 +1,315 @@
+# API Contract and Hardware Recipes
+
+## Contents
+
+1. [Freeze the upstream contract](#freeze-the-upstream-contract)
+2. [Validate before side effects](#validate-before-side-effects)
+3. [Task request templates](#task-request-templates)
+4. [Serving benchmark template](#serving-benchmark-template)
+5. [Hardware recipe contract](#hardware-recipe-contract)
+6. [Best deployment selection](#best-deployment-selection)
+
+## Freeze the upstream contract
+
+Pin three revisions independently:
+
+- official source implementation commit or package version;
+- checkpoint/model-card revision and partition;
+- vLLM-Omni target commit.
+
+Build a table before editing code. One row is one official task/partition.
+
+| Field | Required content |
+|---|---|
+| Task and partition | Official task name, checkpoint subdirectory, task selection rules |
+| Call contract | Constructor and call arguments, types, required/optional, official defaults |
+| Prompt/input | Prompt count/length, modality, MIME, order, count, decoded shape |
+| Size/time | Width/height alignment, pixels, frames, duration, FPS, start offset |
+| Denoise | Scheduler, sigma schedule, steps, CFG/guidance, seed behavior |
+| Output | Count, format, codec/sample rate, return type |
+| API | Offline mapping, sync endpoint, async endpoint/job lifecycle |
+| Rejection | Status code and stable error class/message for every invalid boundary |
+
+Official API parity includes limits and defaults, not just names. Do not add a
+generic default that changes a distilled checkpoint's fixed schedule, or accept
+an input combination the official task rejects.
+
+For every accepted boundary, add the neighboring rejected values. Examples:
+`max_images` and `max_images + 1`, smallest/largest legal duration, aligned and
+misaligned dimensions, omitted and illegal steps, one and too many outputs.
+
+## Validate before side effects
+
+Use one normalization/validation path for offline, sync, and async requests.
+The ordering must be:
+
+```text
+parse -> normalize aliases -> validate task/source matrix -> validate declared
+limits -> stream with byte caps -> inspect/decode with frame/pixel/duration caps
+-> persist only if needed -> submit to engine
+```
+
+Fail before download, persistence, job creation, or engine submit whenever the
+available metadata is sufficient. For streamed uploads, enforce:
+
+- total reference count and per-modality count before reading bodies;
+- per-file bytes while streaming;
+- aggregate bytes across all files;
+- decoded frames, duration, pixels, and media type before engine submission.
+
+On rejection, close streams and remove every partial temporary resource. Use
+HTTP 400 for invalid task/field combinations and 413 for payload limits. Test
+both sync and async paths; an invalid async request must not create a job ID.
+
+Do not trust client MIME alone. Sniff/validate content and keep ordered repeated
+multipart fields ordered. Avoid logging prompts, raw references, signed URLs,
+or base64 payloads.
+
+## Task request templates
+
+These are endpoint-shape templates, not model recipes. Replace placeholders
+with the exact official defaults and validated assets for the target model.
+Every production recipe must include all advertised tasks and an expected
+output check.
+
+### Text to image: JSON
+
+```bash
+curl --fail-with-body -sS -X POST http://127.0.0.1:8091/v1/images/generations \
+  -H 'Content-Type: application/json' \
+  -d '{
+    "model": "<model>",
+    "prompt": "<validated prompt>",
+    "size": "1024x1024",
+    "num_inference_steps": 30,
+    "seed": 42,
+    "response_format": "b64_json"
+  }' | jq -er '.data[0].b64_json' | base64 -d > output.png
+file output.png
+```
+
+### Image edit: multipart
+
+```bash
+curl --fail-with-body -sS -X POST http://127.0.0.1:8091/v1/images/edits \
+  -F 'model=<model>' \
+  -F 'prompt=<validated edit prompt>' \
+  -F 'image=@/path/to/reference.png;type=image/png' \
+  -F 'size=1024x1024' \
+  -F 'num_inference_steps=30' \
+  -F 'seed=42' \
+  -F 'output_format=png' \
+  | jq -er '.data[0].b64_json' | base64 -d > edited.png
+file edited.png
+```
+
+Use repeated `image` fields only if the model's official input matrix allows
+multiple ordered sources. Never describe a file upload as JSON-only.
+
+### Public request fields versus internal sampling fields
+
+Use the field exposed by the target revision's endpoint schema; do not rename
+it from the pipeline's internal attribute. For example, the current video JSON
+and multipart contracts expose `extra_params`, while serving translates that
+object into `DiffusionSamplingParams.extra_args`. Other request paths may
+expose canonical `extra_args` or nest it under `extra_body`. Inspect the exact
+Pydantic model/FastAPI form dependency and exercise the generated curl before
+publishing a recipe. A deprecation warning in a different endpoint is not
+evidence that an unrecognized multipart form field works here.
+
+Current endpoint map (re-check it at the target commit):
+
+| Endpoint | Public model-specific field |
+|---|---|
+| `/v1/videos`, `/v1/videos/sync` | Multipart `extra_params` JSON string |
+| `/v1/images/generations` | JSON `extra_params` object |
+| `/v1/chat/completions` diffusion | `extra_body.extra_args` |
+| `/v1/images/edits` | Declared typed form fields only |
+| `/v1/audio/generate` | Declared typed JSON fields only |
+
+Prefer declared top-level fields. Use a generic model-specific envelope only
+when the endpoint actually exposes it.
+
+### Text to video: multipart
+
+```bash
+curl --fail-with-body -sS -X POST http://127.0.0.1:8091/v1/videos/sync \
+  -H 'Accept: video/mp4' \
+  -F 'model=<model>' \
+  -F 'prompt=<validated prompt>' \
+  -F 'size=1280x720' \
+  -F 'num_frames=81' \
+  -F 'fps=16' \
+  -F 'num_inference_steps=30' \
+  -F 'seed=42' \
+  -F 'extra_params={"task":"<official-task>"}' \
+  -o output.mp4
+ffprobe -v error -show_entries stream=codec_name,width,height,r_frame_rate \
+  -of json output.mp4
+```
+
+### Image/video to video: multipart
+
+```bash
+curl --fail-with-body -sS -X POST http://127.0.0.1:8091/v1/videos/sync \
+  -H 'Accept: video/mp4' \
+  -F 'model=<model>' \
+  -F 'prompt=<validated prompt>' \
+  -F 'num_frames=81' \
+  -F 'fps=16' \
+  -F 'num_inference_steps=30' \
+  -F 'seed=42' \
+  -F 'extra_params={"task":"<official-task>","start_time_seconds":0}' \
+  -F 'input_reference=@/path/to/reference.png;type=image/png' \
+  -o i2v.mp4
+
+curl --fail-with-body -sS -X POST http://127.0.0.1:8091/v1/videos/sync \
+  -H 'Accept: video/mp4' \
+  -F 'model=<model>' \
+  -F 'prompt=<validated prompt>' \
+  -F 'num_frames=81' \
+  -F 'num_inference_steps=30' \
+  -F 'seed=43' \
+  -F 'extra_params={"task":"<official-task>","start_time_seconds":0}' \
+  -F 'input_reference=@/path/to/reference.mp4;type=video/mp4' \
+  -o v2v.mp4
+```
+
+For multiple references, repeat the model's accepted multipart field in
+official order. Include complete downloadable or repository fixtures plus
+SHA256 hashes. A placeholder-only recipe is not runnable evidence.
+
+### Text to audio: JSON
+
+```bash
+curl --fail-with-body -sS -X POST http://127.0.0.1:8091/v1/audio/generate \
+  -H 'Content-Type: application/json' \
+  -d '{
+    "model": "<model>",
+    "input": "<validated audio prompt>",
+    "audio_length": 10.0,
+    "num_inference_steps": 100,
+    "seed": 42,
+    "response_format": "wav"
+  }' -o output.wav
+ffprobe -v error -show_entries stream=codec_name,sample_rate,channels \
+  -of json output.wav
+```
+
+Use `/v1/audio/generate` for general diffusion audio. Do not substitute the
+speech API unless the model is actually served through the speech contract.
+
+### Async video lifecycle
+
+Alongside `/v1/videos/sync`, include `POST /v1/videos`, status polling,
+content retrieval, cancellation/abort if public, and cleanup assertions. The
+same request must normalize to the same engine sampling parameters in both
+modes.
+
+## Serving benchmark template
+
+Run a smoke first, then fixed-concurrency and arrival-rate tests. Video task
+names for the harness are `t2v`, `i2v`, `ti2v`, or `v2v`; model-specific task
+selection belongs in `--extra-body` when required.
+
+```bash
+python benchmarks/diffusion/diffusion_benchmark_serving.py \
+  --base-url http://127.0.0.1:8091 \
+  --endpoint /v1/videos \
+  --model '<model>' \
+  --dataset random \
+  --task t2v \
+  --num-prompts 100 \
+  --width 1280 --height 720 \
+  --num-frames 81 --fps 16 \
+  --num-inference-steps 30 \
+  --request-rate 0.2 \
+  --max-concurrency 4 \
+  --warmup-requests 4 \
+  --warmup-concurrency 4 \
+  --extra-body '{"extra_params":"{\"task\":\"<official-task>\"}"}' \
+  --output-file result.json
+```
+
+The video backend currently stringifies each multipart value. Pre-serialize
+the nested `extra_params` object as above; passing it as a nested object would
+produce Python dict syntax rather than valid JSON. If the target revision
+changes the backend to JSON-encode dict/list values, update and smoke-test this
+command against that revision.
+
+For i2v/v2v, use the harness's validated dataset/trace path or provide the
+documented media source; do not silently benchmark t2v and label it i2v. Keep
+prompt, assets, seed, dimensions, frames, steps, guidance, warmup, concurrency,
+and arrival process identical across A/B runs.
+
+Keep three benchmark lanes distinct:
+
+1. lossless runtime A/B with the released schedule/precision/attention policy;
+2. accelerated-path A/B that changes one adapter, quantization, cache, or
+   sparsity policy and adds same-seed quality evidence;
+3. production-topology studies that vary placement or parallelism and report a
+   latency/throughput/memory frontier rather than a kernel speedup.
+
+For a fixed single-request A/B, retain every measured run and report median or
+mean with range after an explicit warmup; do not infer p95/p99 from a few
+repetitions. Tail latency requires a declared arrival process, request count,
+concurrency, and enough successful samples. Reconcile client E2E with queue,
+encoder, denoise, video/audio decode, D2H/IPC, codec, and residual timings.
+
+For generated media, define `real-time` rather than relying on the label. A
+complete-response claim uses `client E2E / validated output duration <= 1` and
+does not imply streaming or low time-to-first-frame. Report first chunk/fragment
+and cadence separately when a public streaming route exists. Never add gains
+across benchmark lanes with different schedules, prompts, seeds, artifacts, or
+topologies.
+
+The diffusion serving harness currently covers image/video task families, not
+`/v1/audio/generate`. For text-to-audio, use the target model's task-specific
+audio benchmark if one exists; otherwise add a serving harness that records
+request rate, concurrency, latency, generated-audio seconds per second, error
+rate, and output validation. Do not relabel the TTS harness or a video workload
+as diffusion-audio evidence.
+
+## Hardware recipe contract
+
+Create a separate row and runnable section for each vendor/card/topology:
+
+| Evidence | Required detail |
+|---|---|
+| Identity | Vendor, exact SKU, device count, HBM, interconnect |
+| Software | Driver, CUDA/ROCm/CANN/oneAPI, framework/container, dependency versions |
+| Source | vLLM-Omni commit, official/checkpoint revision and hashes |
+| Deployment | dtype, attention backend per role, TP/SP/CFG/HSDP/DP, cache, quant, offload |
+| Commands | Complete deploy YAML/serve command, every task curl, benchmark command |
+| Method | Warmup exclusion, repetitions or request count, concurrency, arrival distribution, duration, statistic definitions |
+| Results | Success rate, p50/p95/p99, RPS/throughput per device, stage latency |
+| Memory | Load/materialize, resident, encode, denoise, decode, transient per-rank HBM and host PSS |
+| Quality | Metric/tolerance, artifact links/hashes, temporal/audio checks |
+| Output | Raw/offline dtype, range, layout and ownership; online codec, payload bytes, transport route and client boundary |
+| Limits | `validated`/`limited`/`unsupported`/`not tested`, fallback/rejection behavior |
+
+Use current paths such as `recipes/<vendor>/`, `recipes/README.md`, supported
+models, and feature compatibility tables. Top-level documentation must link to
+the scoped matrix; it must not use an unqualified checkmark that hides a ROCm,
+NPU, XPU, task, topology, or dtype limitation.
+
+DLO evidence is per card, especially on small-HBM devices. A DLO result on one
+CUDA card does not establish another CUDA SKU, ROCm, NPU, or XPU support.
+
+## Best deployment selection
+
+Benchmark a correctness-approved candidate matrix, not every theoretical
+combination. Choose the recommended row by the production objective—latency,
+throughput, cost, or small-HBM fit—and document that objective.
+
+Do not recommend a row until it passes:
+
+- all advertised task/API contracts;
+- fixed-reference quality;
+- request isolation and abort/error cleanup;
+- sustained arrival-rate tests;
+- per-rank HBM and host PSS limits;
+- hardware-specific accuracy and performance CI.
+
+Keep the most memory-constrained DLO row even when it is not the fastest; it is
+a separate production recipe, not a failed latency candidate.

.claude/skills/production-add-diffusion-model/references/feature-patterns.md

@@ -0,0 +1,462 @@
+# Production Feature Patterns
+
+## Contents
+
+1. [Compatibility matrix](#compatibility-matrix)
+2. [Online FP8](#online-fp8)
+3. [Strict fused-weight loading](#strict-fused-weight-loading)
+4. [Distributed layerwise offload](#distributed-layerwise-offload)
+5. [Per-request Cache-DiT](#per-request-cache-dit)
+6. [Combination gates](#combination-gates)
+
+## Compatibility matrix
+
+Create rows at the granularity below. Add evidence links and a reason for every
+non-validated state.
+
+| Task/shape | Execution mode/capacity | Backend/layout | Quant | Cache policy | Offload | Topology | Hardware/dtype | Output contract/transport | State | Evidence |
+|---|---|---|---|---|---|---|---|---|---|---|
+| `<task>/<shape>/<schedule>` | serial request / 1 | SDPA, padded | BF16 | none | resident | 1 device | `<card>` BF16 | raw float / local | `not tested` | — |
+| `<task>/<shape>/<schedule>` | step / N | `<fast>`, packed | BF16 | none | resident | `<TP/SP>` | `<card>` BF16 | encoded / subprocess | `not tested` | — |
+| `<task>/<shape>/<schedule>` | step / N | `<fast>`, packed | FP8 | high | DLO no-AG | `<TP/SP>` | `<card>` | device uint8 / SHM | `not tested` | — |
+
+Initialize every row as `not tested`; the table is a work queue, not evidence.
+Promote the dense BF16 row to the oracle only after its scoped parity passes.
+Add one axis at a time. If the runtime cannot
+enforce a known-incompatible combination, add construction/admission
+validation before writing a recipe.
+
+## Online FP8
+
+### Component routing
+
+The public builder accepts one method or a longest-prefix component map:
+
+```python
+from vllm_omni.quantization import build_quant_config
+
+config = build_quant_config({
+    "transformer": {
+        "method": "fp8",
+        "ignored_layers": ["transformer.final_layer"],
+    },
+    "text_encoder": None,
+    "vae": None,
+    "default": None,
+})
+
+assert config.resolve("transformer.blocks.0.attn.qkv_proj") is not None
+assert config.resolve("text_encoder.layers.0.mlp") is None
+assert config.resolve("vae.decoder.conv_out") is None
+```
+
+Use actual runtime prefixes. Multi-DiT or nested pipelines may not use the
+generic names `transformer` and `vae`. A prefix that falls through to a default
+is not component validation. Add tests for every included and ignored prefix.
+
+For CLI serving, first validate resident DiT-only FP8:
+
+```bash
+vllm serve '<model>' --omni --quantization fp8 --max-num-seqs 1
+```
+
+For scoped routing:
+
+```bash
+vllm serve '<model>' --omni \
+  --diffusion-quantization-config \
+  '{"<dit-runtime-prefix>":{"method":"fp8","ignored_layers":["<exact-linear-prefix>"]},"<vae-runtime-prefix>":null,"default":null}' \
+  --max-num-seqs 1
+```
+
+Do not publish these placeholders. A recipe must contain the target model's
+verified prefixes.
+
+### Quantizable module wiring
+
+Pass `quant_config` and a stable prefix through every vLLM linear. Keep
+precision-sensitive modulation/norm layers safe unless proven otherwise.
+
+```python
+from vllm.model_executor.layers.activation import SiluAndMul
+from vllm.model_executor.layers.linear import (
+    MergedColumnParallelLinear,
+    QKVParallelLinear,
+    RowParallelLinear,
+)
+
+self.qkv_proj = QKVParallelLinear(
+    hidden_size=hidden_size,
+    head_size=head_dim,
+    total_num_heads=num_heads,
+    total_num_kv_heads=num_kv_heads,
+    bias=False,
+    quant_config=quant_config,
+    prefix=f"{prefix}.attn.qkv_proj",
+)
+self.gate_up_proj = MergedColumnParallelLinear(
+    hidden_size,
+    [intermediate_size, intermediate_size],
+    bias=False,
+    quant_config=quant_config,
+    prefix=f"{prefix}.mlp.gate_up_proj",
+)
+self.act_fn = SiluAndMul()
+self.down_proj = RowParallelLinear(
+    intermediate_size,
+    hidden_size,
+    bias=False,
+    input_is_parallel=True,
+    quant_config=quant_config,
+    prefix=f"{prefix}.mlp.down_proj",
+)
+```
+
+Set required checkpoint parameters to fail rather than silently initialize if
+the current loader supports `missing_param_init = "error"`.
+
+### FP8 evidence
+
+For each task/hardware row, capture:
+
+- startup proves the FP8 quant method is attached to intended linear modules;
+- no unexpected BF16 fallback and no meta/uninitialized parameters;
+- same-seed component, trajectory, and final-artifact comparison with BF16;
+- HBM at load, resident, encode, denoise, decode, and peak;
+- cold load time; fixed-work warm-latency raw runs; and serving p50/p95/p99 plus
+  throughput only from a declared arrival load with enough samples;
+- named ignored layers and why their BF16 retention is required.
+
+DiT FP8 evidence does not cover the text encoder or VAE. CUDA evidence does not
+cover ROCm/NPU/XPU. A pre-quantized checkpoint is a different loading path from
+online FP8 and needs its own state. If it is also pruned, rotated, distilled, or
+adapter-modified, report that model delta explicitly; BF16-versus-artifact
+quality and speed are deployment comparisons, not pure quantization ablations.
+
+## Strict fused-weight loading
+
+Never use `param.data.copy_()` for quantizable vLLM parameters. Transform the
+checkpoint layout before invoking the parameter loader so the online quant
+loader remains outermost:
+
+```python
+from vllm.model_executor.model_loader.weight_utils import default_weight_loader
+
+param = params[target_name]
+weight_loader = getattr(param, "weight_loader", default_weight_loader)
+
+if target_name.endswith(".qkv_proj.weight"):
+    packed_qkv = reorder_checkpoint_qkv(checkpoint_weight)
+    weight_loader(param, packed_qkv)
+elif target_name.endswith(".gate_up_proj.weight"):
+    gate, up = checkpoint_weight.chunk(2, dim=0)
+    weight_loader(param, gate, 0)
+    weight_loader(param, up, 1)
+else:
+    weight_loader(param, checkpoint_weight)
+```
+
+Track source shards per fused destination:
+
+```python
+expected = {
+    "blocks.0.attn.qkv_proj.weight": {"q", "k", "v"},
+    "blocks.0.mlp.gate_up_proj.weight": {0, 1},
+}
+seen = {name: set() for name in expected}
+
+# After mapping/loading each source:
+seen[target_name].add(shard_id)
+
+incomplete = {
+    name: sorted(shards - seen[name], key=str)
+    for name, shards in expected.items()
+    if shards - seen[name]
+}
+if incomplete:
+    raise RuntimeError(f"incomplete fused checkpoint parameters: {incomplete}")
+```
+
+Also compare retained model parameters with the returned loaded set. Explicit
+checkpoint ignores need a reason and test; an unknown or missing weight is not
+an ignore policy.
+
+Load-time adapter fusion creates a dedicated model, not a request-switchable
+LoRA. Pin the base and adapter revisions, validate every low-rank and full-rank
+delta target, freeze the adapter-declared schedule/task, and fuse before
+sharding only when the loader contract proves identical rank-local weights.
+Reject offload or host-weight paths that bypass that fusion rather than serving
+the unfused base model silently.
+
+## Distributed layerwise offload
+
+### Declare component topology
+
+Use `SupportsComponentDiscovery` and `OffloadPlan`. Validate all dotted paths
+at construction/test time because discovery may otherwise warn and skip.
+
+Treat a future/requested `--layerwise-offload-components` selector as a public
+API only if it exists in the target revision. Today, express component topology
+with `OffloadPlan` (`on_demand_component_paths`, `block_attrs`,
+`encoder_block_attrs`, `resident_dit_paths`, and `offload_submodules`). Do not
+invent a CLI flag or claim component-selective DLO from discovery declarations
+alone; validate the selector separately when it lands.
+
+```python
+from operator import attrgetter
+
+import torch
+
+def assert_offload_path(root, path):
+    value = attrgetter(path)(root)
+    if not isinstance(value, torch.nn.Module):
+        raise TypeError(f"offload path {path!r} is not an nn.Module")
+    return value
+
+for path in (
+    *pipeline._dit_modules,
+    *pipeline._encoder_modules,
+    *pipeline._vae_modules,
+    *pipeline._offload_plan.on_demand_component_paths,
+):
+    assert_offload_path(pipeline, path)
+```
+
+For `block_attrs`, resolve the DiT first and then each declared block attribute.
+Require an indexable block container with at least one `nn.Module`. Validate
+`offload_submodules`, `resident_dit_paths`, and `encoder_block_attrs` similarly.
+
+### Keep host storage loader-owned
+
+Current DLO accepts a loader-produced `HostWeightPlan`. Pipeline authors should
+not instantiate this plan or add ad hoc mmap capability flags. The loader must
+preflight complete runtime/checkpoint bindings, names, shapes, dtypes, persistent
+buffers, and represented transforms before it skips ordinary materialization.
+The exact accepted plan is consumed once by DLO; a rejected plan records an
+observable fallback reason and continues through the ordinary loader.
+
+Direct checkpoint mmap currently requires TP1 without HSDP or online
+quantization. TP>1 falls back to the ordinary TP-aware loader and may feed DLO
+AllGather or no-AllGather; the current design records a bounded DP2xTP2
+AllGather smoke, but that path does not receive direct-mmap host-page sharing.
+Per-tensor online FP8 must use the ordinary loader. Current revisions can admit
+its finalized weight/scale layout to DLO AllGather, including transposed runtime
+weights, while unsupported online quantizers still fail closed. Direct mmap
+remains unavailable, so capture the temporary full-model host materialization
+before rank sharding.
+
+Host Weight Runtime is a distinct opt-in path. Current final-layout consumers
+use no-AllGather DLO and exact model-declared BF16 representations. In
+`preferred` mode, consume an exact local hit or canonically load and publish for
+a future startup; in `required` mode, consume only and fail on a miss or unusable
+artifact. HWR must not interact with DLO AllGather, online quantization, HSDP,
+LoRA/adapted weights, or an unrecognized load format unless the target revision
+adds an exact producer/restore contract. It is immutable node-local CPU backing,
+not zero-copy GPU execution; the DLO transport still owns staging or direct H2D.
+Qualify identity inputs, cold population, warm reuse, atomic concurrent
+publication, source digests, corruption/quarantine, lease cleanup, capacity,
+node-local PSS, H2D, and E2E latency. A host-memory saving that materially
+regresses transfer or request latency is a separate memory-first deployment.
+
+### DLO deployment modes
+
+Treat these paths separately:
+
+| Path | Weight/runtime behavior | Required boundaries |
+|---|---|---|
+| DLO AllGather | Rank-sharded pinned host tensors; reconstruct layer through collective | TP>1 uses ordinary TP-aware loader output; HSDP and unsupported online quantizers are rejected; finalized per-tensor online FP8 is eligible only when recognized by the target revision; concurrent ranks must participate consistently |
+| DLO no-AllGather | Loader-approved checkpoint mmap or ordinary runtime tensors; each rank streams a complete block | Direct mmap is TP1/non-HSDP/non-online-quant only; TP/HSDP/online quant use ordinary runtime tensors and require scoped E2E |
+| SP + DLO | SP group can supply the DLO collective group | Packed boundaries and collective order require E2E |
+
+Single-stage examples:
+
+```bash
+# SP + DLO AllGather candidate; validate exact model/card before publishing.
+vllm serve '<model>' --omni \
+  --enable-distributed-layerwise-offload \
+  --usp 4 \
+  --max-num-seqs 1
+
+# TP + DLO rank-local candidate; not a support statement.
+vllm serve '<model>' --omni \
+  --tensor-parallel-size 2 \
+  --enable-distributed-layerwise-offload \
+  --dlo-no-use-allgather \
+  --max-num-seqs 1
+```
+
+Under `--omni`, configure DP in a deploy YAML rather than passing vLLM DP CLI
+flags:
+
+```yaml
+pipeline: <registered-pipeline>
+async_chunk: false
+data_parallel_size: 4
+stages:
+  - stage_id: 0
+    devices: "0,1,2,3"
+    max_num_seqs: 1
+    enable_distributed_layerwise_offload: true
+    dlo_use_allgather: true
+    parallel_config:
+      tensor_parallel_size: 1
+      sequence_parallel_size: 1
+```
+
+```bash
+vllm serve '<model>' --omni --deploy-config /path/to/dlo_dp4.yaml
+```
+
+Keep the YAML next to the validated recipe. Match its schema to the target
+revision; do not assume a model accepts another pipeline's deploy config.
+
+### DLO validation sequence
+
+1. Resident BF16 reference.
+2. Ordinary layerwise offload parity.
+3. DLO enable, warmup, same-seed generation, disable/restart.
+4. AllGather and no-AllGather separately.
+5. Concurrent requests with different prompts and identical collective-affecting
+   parameters; then reject or schedule incompatible steps/shapes safely.
+6. Queued/in-flight abort, error, idle rank, worker exit, and next request.
+7. TP, SP, cache, compile, online FP8 one axis at a time.
+8. Loader-plan selection/fallback reason, cold/warm storage lifecycle, and
+   corruption/stale-entry recovery when mmap/cache backing is advertised.
+9. Replica-local request preparation plus DLO-group-only weight collectives in
+   DP; exercise concurrent replicas to catch accidental WORLD rendezvous.
+10. Sweep resident block counts; publish raw latency/HBM pairs, dominated
+    points, and the non-dominated Pareto frontier.
+11. Per-rank HBM, host PSS for the full process tree, H2D and collective trace.
+
+On a small-HBM card, measure transient encode/VAE peaks as well as resident DiT
+weights. If a non-DiT component still exceeds HBM, combine only independently
+validated on-demand component staging, VAE tiling/patching, or another topology;
+do not hide the remaining peak behind the phrase “DLO enabled.”
+
+## Per-request Cache-DiT
+
+### Policy and lifecycle skeleton
+
+First verify that the target revision contains the shared protocol/runtime.
+If not, follow the nearest current model's lifecycle without inventing a new
+public API.
+
+The following is a **serial/exclusive request-mode sketch only**. Pipeline-wide
+hooks are mutable; concurrent requests need an admission/batching key and
+runner-owned transition serialization before this shape is safe.
+
+```python
+import torch.nn as nn
+
+from vllm_omni.diffusion.cache.cachedit import (
+    CacheDiTBackend,
+    CacheDiTRequestSpec,
+    RequestScopedCacheDiTRuntime,
+)
+
+class MyPipeline(nn.Module):
+    def __init__(self, *args, **kwargs):
+        super().__init__()
+        # ... components ...
+        self._cache_dit_runtime = RequestScopedCacheDiTRuntime(self)
+
+    def adopt_cache_dit_backend(self, backend: CacheDiTBackend) -> None:
+        self._cache_dit_runtime.adopt(
+            backend,
+            installation_key="my_model.generic",
+        )
+
+    def is_cache_dit_enabled(self) -> bool:
+        return self._cache_dit_runtime.is_enabled
+
+    def _cache_spec(self, quality, steps):
+        if quality == "lossless":
+            return None
+        if quality == "high":
+            return CacheDiTRequestSpec(
+                installation_key="my_model.high.v1",
+                cache_config=self._validated_high_config,
+                num_inference_steps=steps,
+            )
+        raise ValueError(f"unsupported quality tier: {quality!r}")
+
+    def forward(self, req):
+        quality, steps = validate_and_resolve_request(req)
+        self._cache_dit_runtime.prepare(self._cache_spec(quality, steps))
+        return self._run_validated_request(req)
+```
+
+The snippet shows the model-owned transition for exclusive execution. Put
+disable/restore handling in the runtime/runner lifecycle that owns success,
+exception, disconnect, and asynchronous abort; a `try/finally` inside
+`forward()` alone cannot observe every server cancellation boundary.
+
+The `installation_key` must include every policy dimension that changes hook
+installation; changes in step count can use runtime refresh. `None` disables
+installed hooks. Do not mutate the global startup cache config for a request.
+
+Centralize cleanup in the owning request/runner boundary. An exception, abort,
+disconnect, or admission failure must not leave a changed hook policy visible
+to the next request. The exact owner depends on the target revision; add a
+failure-injection test rather than relying on `forward()` alone to observe
+asynchronous abort.
+
+Example model-specific request form after calibration:
+
+```bash
+curl --fail-with-body -sS -X POST http://127.0.0.1:8091/v1/videos/sync \
+  -F 'model=<model>' \
+  -F 'prompt=<prompt>' \
+  -F 'quality=lossless' \
+  -F 'num_inference_steps=<validated-steps>' \
+  -o lossless.mp4
+
+curl --fail-with-body -sS -X POST http://127.0.0.1:8091/v1/videos/sync \
+  -F 'model=<model>' \
+  -F 'prompt=<prompt>' \
+  -F 'quality=high' \
+  -F 'num_inference_steps=<validated-steps>' \
+  -o high.mp4
+```
+
+Do not expose these tier names merely because another model uses them.
+
+### Cache-DiT evidence
+
+For every tier/task/shape/schedule:
+
+- show real cache hits/skipped blocks, not only “backend enabled” logs;
+- measure fixed-work speed and quality against lossless/native;
+- publish a speed/quality frontier and chosen tier mapping;
+- alternate lossless/high/lossless and vary steps to prove restore/refresh;
+- interleave tasks and shapes; run concurrent requests only when scheduling
+  guarantees exclusive compatible hook state;
+- inject validation error, denoise error, disconnect, and abort, then prove the
+  next request starts with its requested policy;
+- bound any cached tensors and clear request-local state.
+
+## Combination gates
+
+Start every combination as `not tested`.
+
+| Combination | Default production posture |
+|---|---|
+| Direct checkpoint mmap + TP>1/HSDP/online quant | Preflight falls back to the ordinary loader; do not claim direct-mmap savings |
+| Per-tensor online FP8 + DLO AllGather | Generic compatibility exists for finalized weights/scales; each model/card/topology remains a candidate until E2E, and other online quantizers fail closed |
+| Online FP8 + DLO no-AllGather | Candidate only after model/card/topology E2E |
+| TP>1 + DLO AllGather | Ordinary TP-aware loader only; bounded DP2xTP2 smoke exists, but each new model/card/topology still needs E2E |
+| TP>1 + DLO no-AllGather | Ordinary TP-aware loader only; candidate with limited generic coverage |
+| HSDP + DLO AllGather | Reject; it would double-shard HSDP-managed parameters |
+| HSDP + DLO no-AllGather | Configuration-compatible candidate with limited E2E coverage |
+| Cache-DiT + online FP8 | Separate trajectory, hit-rate, HBM, latency test |
+| Cache-DiT + DLO | Verify hooks, streamed blocks, memory peaks, abort cleanup |
+| Cache-DiT + step execution | Do not advertise until lifecycle/batching support is explicit |
+| Sparse attention + Ring | Verify backend execution; reject silently ignored sparse settings |
+| Load-time fused adapter + offload/HWR | Reject unless the selected loader path applies and validates the same fusion before sharding/storage |
+| Packed attention + any topology | Unequal-sample boundary parity is mandatory |
+| Device-side uint8 preparation + offline caller | Version the dtype/range/layout change; HTTP MP4 parity does not preserve the raw tensor contract |
+| Device-side preparation + remote codec | Validate route selection, strided/contiguous layouts, payload bytes, ordering, fallback, and encoded-media parity |
+
+A startup-only test never upgrades a combination. Require request output,
+parity/quality, actual fast-path evidence, resource cleanup, and a benchmark on
+the named hardware.

.claude/skills/production-add-diffusion-model/references/performance-patterns.md

@@ -0,0 +1,481 @@
+# Performance Patterns and Code Examples
+
+## Contents
+
+1. [Measurement loop](#measurement-loop)
+2. [Faster BF16 attention](#faster-bf16-attention)
+3. [Sparse attention](#sparse-attention)
+4. [Fusion-first operator patterns](#fusion-first-operator-patterns)
+5. [Remove redundant computation](#remove-redundant-computation)
+6. [USP communication optimization](#usp-communication-optimization)
+7. [Text encoder and VAE disaggregation](#text-encoder-and-vae-disaggregation)
+8. [Performance acceptance](#performance-acceptance)
+
+## Measurement loop
+
+Do not start with a list of optimizations. Start with a fixed correctness and
+performance workload:
+
+```text
+model/checkpoint revision + prompt/media hashes + seed + dimensions/frames +
+scheduler/sigma/steps/guidance + task + dtype + backend + topology + hardware
+```
+
+Capture:
+
+- stage time: load, encode, conditioning, denoise, VAE decode/postprocess;
+- GPU trace: kernel duration/count, launch gaps, allocations, H2D, collectives;
+- CPU trace: preprocessing, per-layer sync, Python dispatch, temporary files;
+- memory: per-rank HBM peaks and process-tree host PSS;
+- output: device preparation, D2H/IPC, payload bytes, codec wall/process CPU,
+  event-loop wait, client materialization, and signed residual;
+- serving: queue time, throughput, errors, request mix, and distribution samples.
+
+For a fixed-work single-request A/B, run one explicit warmup policy and at
+least three measured repetitions, keep every raw run, and report median or mean
+with range. Do not derive p95/p99 from three samples. For tail latency, declare
+the arrival model and collect enough successful requests for the percentile and
+confidence claimed. Implement one optimization at a time, re-run parity, and
+report both local kernel and end-to-end deltas. A kernel win that moves no E2E
+metric is not a production speed claim.
+
+Use the pipeline profiler already exposed by the target revision when
+available, and use PyTorch/CUDA/ROCm/NPU/XPU profiling tools appropriate to the
+platform. Do not force every platform through a CUDA-only workflow.
+
+## Faster BF16 attention
+
+### Shared attention wiring
+
+Use the shared role-aware attention layer and give it the true local head
+counts and QKV layout:
+
+```python
+import torch
+
+from vllm.model_executor.layers.linear import (
+    QKVParallelLinear,
+    RowParallelLinear,
+)
+from vllm_omni.diffusion.attention.layer import Attention
+
+class SelfAttention(torch.nn.Module):
+    def __init__(
+        self,
+        hidden_size,
+        head_dim,
+        num_heads,
+        num_kv_heads,
+        quant_config=None,
+        prefix="",
+    ):
+        super().__init__()
+        self.head_dim = head_dim
+        self.qkv_proj = QKVParallelLinear(
+            hidden_size=hidden_size,
+            head_size=head_dim,
+            total_num_heads=num_heads,
+            total_num_kv_heads=num_kv_heads,
+            bias=False,
+            quant_config=quant_config,
+            prefix=f"{prefix}.qkv_proj",
+        )
+        self.attention = Attention(
+            num_heads=self.qkv_proj.num_heads,
+            num_kv_heads=self.qkv_proj.num_kv_heads,
+            head_size=head_dim,
+            softmax_scale=head_dim**-0.5,
+            causal=False,
+            qkv_layout="BSND",
+            role="self",
+            prefix=prefix,
+        )
+        self.out_proj = RowParallelLinear(
+            num_heads * head_dim,
+            hidden_size,
+            bias=False,
+            input_is_parallel=True,
+            quant_config=quant_config,
+            prefix=f"{prefix}.out_proj",
+        )
+
+    def forward(self, x, attn_metadata=None):
+        qkv, _ = self.qkv_proj(x)
+        q_size = self.qkv_proj.num_heads * self.head_dim
+        kv_size = self.qkv_proj.num_kv_heads * self.head_dim
+        q, k, v = qkv.split([q_size, kv_size, kv_size], dim=-1)
+        q = q.unflatten(-1, (self.qkv_proj.num_heads, self.head_dim))
+        k = k.unflatten(-1, (self.qkv_proj.num_kv_heads, self.head_dim))
+        v = v.unflatten(-1, (self.qkv_proj.num_kv_heads, self.head_dim))
+        out = self.attention(q, k, v, attn_metadata)
+        out = out.flatten(-2)
+        out, _ = self.out_proj(out)
+        return out
+```
+
+Add Q/K norm and RoPE in the official order before the `Attention` call. Keep
+checkpoint mapping for the fused QKV and output projection complete.
+
+### Backend qualification
+
+For each role (`self`, cross-attention category, token refiner, etc.), record:
+
+- actual selected backend from logs/trace;
+- supported dtype, head size, QKV layout, mask/metadata, packed varlen behavior;
+- platform and compute capability;
+- TP/SP/Ring/AllGather-KV restrictions;
+- fallback behavior and whether fallback changes the claimed performance.
+
+Use unequal packed samples and compare with dense SDPA. Verify `cu_seqlens`,
+max sequence length, padding exclusion, modality boundaries, CFG ownership, and
+output slicing. Do not pass an `attn_mask` and assume it is consumed. Inspect
+the backend contract and trace.
+
+TRTLLM, Flash Attention, cuDNN, NPU varlen, and other backends are independent
+qualification rows. The shared `Attention` class does not make every backend
+packed-safe or faster automatically. Ring and hybrid paths require their own
+boundary parity. AllGather-KV and specific backend combinations may fail fast;
+keep that behavior visible in docs.
+
+## Sparse attention
+
+Sparse attention is a backend/model/platform optimization, not a synonym for
+low-precision dense attention. Enable it only after dense BF16 parity.
+
+For every candidate task/shape/topology:
+
+1. State the sparsity policy, selected backend/kernel, block/head dimensions,
+   supported hardware, external kernel/package revision, and fallback.
+2. Prove the sparse kernel executed and report realized sparsity; an enabled
+   flag is insufficient.
+3. Compare same-seed per-block outputs, denoise trajectory, and final artifact
+   with dense attention.
+4. Add modality-specific metrics: temporal consistency and audio/video sync for
+   video, spectral/audio quality for audio, and prompt/image metrics as relevant.
+5. Report attention time and full E2E time. Include preprocessing/index-build
+   overhead, peak memory, and the exact dense control.
+6. Validate each task and shape. Do not extrapolate a T2V result to I2V/Ref2V,
+   or NPU sparse support to CUDA/ROCm/XPU.
+
+If Ring bypasses the selected backend or a sparse setting would be silently
+ignored, reject the combination. Prefer a visible dense fallback only when the
+recipe labels the row `limited` and does not claim sparse speed.
+
+## Fusion-first operator patterns
+
+Prefer current shared/custom operators with explicit guards and native
+fallbacks. Prove parity at operator, block, trajectory, and artifact levels,
+then use a trace to prove fewer launches/bytes or lower latency.
+
+### Q/K RMSNorm + RoPE
+
+```python
+from vllm_omni.diffusion.layers.norm import RMSNorm
+from vllm_omni.diffusion.layers.rope import RotaryEmbedding, apply_rope_to_qk
+
+self.norm_q = RMSNorm(head_dim, eps=eps)
+self.norm_k = RMSNorm(head_dim, eps=eps)
+# Example only: set these booleans from the official checkpoint contract.
+use_neox_layout = False
+cos_sin_are_half_dim = True
+self.rope = RotaryEmbedding(
+    is_neox_style=use_neox_layout,
+    half_head_dim=cos_sin_are_half_dim,
+)
+
+q = self.norm_q(q)
+k = self.norm_k(k)
+q, k = apply_rope_to_qk(self.rope, q, k, (cos, sin))
+```
+
+Replace the illustrative booleans with the official layout in real code.
+Preserve partial rotary dimensions by splitting the rotated prefix and concatenating
+the pass-through suffix. If 3D cos/sin includes a batch axis, the current
+shared path assumes the values are shared across the batch; per-sample
+different positions require a proven representation/path.
+
+For packed `[tokens, heads, head_dim]` tensors with a non-interleaved
+`[cos | sin]` table, the current shared fused boundary is:
+
+```python
+from vllm_omni.diffusion.layers.fused_qk_norm_rope import fused_qk_norm_rope
+
+q, k = fused_qk_norm_rope(
+    q,
+    k,
+    self.norm_q.weight,
+    self.norm_k.weight,
+    rope_table,
+    self.norm_q.variance_epsilon,
+)
+```
+
+Its CUDA fast path currently specializes BF16 `head_dim=128` and
+`rotary_dim=96`; other valid inputs take the eager implementation. The separate
+shared RMSNorm and RotaryEmbedding still provide the general cross-platform
+baseline. Treat NPU/MUSA and other vendor fusions as independent qualification
+rows, and do not generalize a model-specific Triton kernel without compatible
+shape, layout, frequency partition, dtype, and platform evidence.
+
+### Fused QKV projection
+
+Use `QKVParallelLinear` as shown above. Qualification includes:
+
+- `num_heads` and `num_kv_heads` divisibility by TP;
+- local split sizes from the layer, never reusing global counts after TP;
+- Q/K/V checkpoint order conversion before the vLLM loader;
+- all Q, K, and V shards present;
+- quantization scale/loader mapping when FP8 or another method is enabled.
+
+### Gate-up projection + fused SwiGLU
+
+```python
+from vllm.model_executor.layers.activation import SiluAndMul
+from vllm.model_executor.layers.linear import (
+    MergedColumnParallelLinear,
+    RowParallelLinear,
+)
+
+self.gate_up_proj = MergedColumnParallelLinear(
+    input_size=hidden_size,
+    output_sizes=[intermediate_size, intermediate_size],
+    bias=False,
+    quant_config=quant_config,
+    prefix=f"{prefix}.gate_up_proj",
+)
+self.act_fn = SiluAndMul()
+self.down_proj = RowParallelLinear(
+    input_size=intermediate_size,
+    output_size=hidden_size,
+    bias=False,
+    input_is_parallel=True,
+    quant_config=quant_config,
+    prefix=f"{prefix}.down_proj",
+)
+
+gate_up, _ = self.gate_up_proj(x)
+x = self.act_fn(gate_up)
+x, _ = self.down_proj(x)
+```
+
+Map both checkpoint `gate_proj` and `up_proj` shards into the merged parameter
+and require both. A merged projection followed by `F.silu(gate) * up` is not
+the same claim as the `SiluAndMul` fused activation kernel; identify what the
+trace actually uses.
+
+### AdaLN
+
+Use a shared layer only when its mathematical/return contract matches the
+official model:
+
+```python
+from vllm_omni.diffusion.layers.adalayernorm import AdaLayerNorm
+
+self.adaln = AdaLayerNorm(
+    hidden_size,
+    elementwise_affine=False,
+    eps=1e-6,
+)
+x = self.adaln(x, scale[:, None, :], shift[:, None, :])
+```
+
+`AdaLayerNormZero`, `AdaLayerNormZeroSingle`, and continuous variants include
+different projections, broadcasts, and tuple outputs; they are not drop-in
+replacements. A shared AdaLN may have an NPU fast path while using native ops on
+CUDA/HIP/XPU. Do not call it a universal fused AdaLN kernel.
+
+For sigma schedules with a small finite set of timesteps, profile whether
+precomputing the timestep/AdaLN projection once per request is worthwhile.
+Cache keys must include the exact schedule/config/device/dtype and stay bounded.
+
+### Layout, residual, and other candidates
+
+After profiling, inspect:
+
+- bias/dropout/add or gated residual chains;
+- norm + linear or norm + modulation boundaries;
+- repeated transpose/reshape/contiguous around attention;
+- VAE normalization/activation/convolution boundaries;
+- MoE routing and fused experts for models that actually use MoE.
+
+For fused residual-plus-norm or modulation chains, preserve the unfused
+materialization boundary. If the residual was stored in BF16 before RMSNorm,
+round to BF16 and promote again inside the fused kernel; using the unrounded
+FP32 temporary changes the denoising trajectory even when the stored residual
+looks correct.
+
+Keep DLO materialize/release hooks, process-group collectives, request cache
+transitions, and cleanup outside compiled/fused regions unless lifecycle safety
+is explicitly tested. Compile only stable regions and retain eager fallback.
+
+## Remove redundant computation
+
+### Hoist request invariants
+
+Audit the denoise loop for work that depends only on the request rather than
+the current timestep:
+
+```python
+# Before the denoise loop.
+condition = self.condition_proj(text_embeddings)
+refined_condition = self.token_refiner(condition, packed_metadata)
+rope_tables = self.build_rope_tables(position_ids)
+attn_metadata = self.build_attention_metadata(layout)
+
+for step, sigma in enumerate(sigmas):
+    timestep_state = self.prepare_timestep_state(sigma)
+    noise = self.transformer(
+        latents,
+        refined_condition,
+        rope_tables=rope_tables,
+        attn_metadata=attn_metadata,
+        timestep_state=timestep_state,
+    )
+    latents = self.scheduler_step(noise, sigma, latents)
+```
+
+Candidates include prompt conditioning, token-refiner output, reference
+encoding/scans, position IDs/RoPE tables, packed masks/metadata, static CFG
+branches, and finite-schedule AdaLN projections. First prove they are invariant
+for all official tasks, LoRA/adapters, prompt-update features, shapes, and
+schedules. In continuous batching, store request-specific results on request
+state rather than the pipeline singleton.
+
+### Eliminate per-layer synchronization
+
+Avoid device-to-host scalar extraction such as `.item()` inside every layer:
+
+```python
+# Bad: forces a GPU/CPU synchronization in every attention layer.
+max_len = int(cu_seqlens[-1].item())
+
+# Better: compute validated host metadata once during request packing and pass it.
+packed = PackedLayout(
+    cu_seqlens=cu_seqlens,
+    max_seqlen=max(sample_lengths),
+)
+for block in self.blocks:
+    hidden = block(hidden, packed_layout=packed)
+```
+
+Trace to verify the synchronization disappears. Do not replace a correct
+dynamic value with a stale constant merely to remove `.item()`.
+
+### Remove unused masks and allocations
+
+If packed attention consumes `cu_seqlens` and ignores a dense mask, do not
+construct the dense mask. Reuse immutable request metadata and preallocated
+buffers when safe. Prefer direct indexed scatter/gather over constructing large
+temporary tensors only when duplicate-index semantics and gradients (normally
+disabled for inference) remain correct.
+
+Any cross-request cache needs a bounded lifetime and a complete key containing
+shape, layout, device, dtype, schedule, task, adapter/LoRA, and any value that
+changes the tensor. Otherwise keep it request-scoped.
+
+### VAE and reference work
+
+Profile duplicate VAE decode/postprocess, repeated reference decoding/scans,
+unnecessary host-device round trips, and full-frame materialization. Apply VAE
+tiling, slicing, patch parallelism, or on-demand staging only with seam,
+ordering, color/range, and memory parity tests.
+
+## USP communication optimization
+
+“USP works” is not evidence that USP communication is optimized. For Ulysses,
+Ring, hybrid, and AllGather-KV separately, profile:
+
+- collective count, bytes, duration, stream, and overlap with compute;
+- Q/K/V and output all-to-all boundaries;
+- pack/unpack, padding, metadata broadcast, transpose, and contiguous copies;
+- head divisibility or advanced uneven-head support;
+- scaling efficiency from 1 to N devices on the same workload/interconnect;
+- load balance for unequal packed samples.
+
+Declare `_sp_plan` split/gather boundaries around meaningful module outputs,
+not arbitrary Python statements. Keep packed metadata sharded/gathered in the
+same ownership model as tokens.
+
+When the target revision offers accelerated Ulysses transport such as
+`--ulysses-a2a-permute`, qualify it separately from regular Ulysses. Prove the
+strict scatter-heads/gather-sequence layout selected the fast path, record
+one-time JIT/readiness cost, prewarm the maximum workspace shape before CUDA
+graph capture, keep the grow-only workspace on one stream, and exercise cleanup
+before process-group destruction. Compare collective and E2E deltas after the
+dense backend and topology are frozen; keep unsupported layouts on an explicit
+regular-Ulysses fallback.
+
+For GQA, validate `Q_heads % KV_heads == 0`, pad KV heads first, and derive Q
+padding from the original ratio. Exercise both contiguous and row-strided QKV
+layouts; a fast copy path must preserve bitwise values and packed boundaries.
+
+Cross-attention sometimes uses replicated text K/V while only video/audio Q is
+sequence-sharded. In that specific case, `skip_sequence_parallel=True` can
+avoid incorrect or wasteful Q/K/V redistribution:
+
+```python
+self.cross_attention = Attention(
+    num_heads=local_heads,
+    num_kv_heads=local_kv_heads,
+    head_size=head_dim,
+    softmax_scale=head_dim**-0.5,
+    causal=False,
+    role="cross",
+    skip_sequence_parallel=True,
+    prefix=f"{prefix}.cross_attn",
+)
+```
+
+Use it only after proving K/V are replicated and Q/output ownership is correct.
+Otherwise it disables needed communication. Report whether the optimization
+reduces collective bytes/time and improves E2E scaling.
+
+## Text encoder and VAE disaggregation
+
+Disaggregation is a measured architecture decision. First collect:
+
+| Question | Required measurement |
+|---|---|
+| Is the stage large enough? | Encode/decode share at target concurrency and shape |
+| Is transfer economical? | Tensor/media bytes, serialization, connector latency and bandwidth |
+| Is there reuse? | Prompt/reference reuse rate, cacheability, fan-out |
+| Can it scale independently? | Queue saturation and resource utilization per stage |
+| Is the contract stable? | Shape, dtype, mask/tag/position metadata, ordering |
+
+### Text encoder candidate
+
+Define an explicit tensor contract containing hidden states, masks, tags or
+position IDs, dtype, shape, task/checkpoint revision, and request ID. Support
+disabling the local encoder so the DiT cannot accidentally recompute it. Test
+all tasks, prompt lengths, partial TP group membership, retries, backpressure,
+abort, connector cleanup, and independent replica scaling.
+
+### VAE candidate
+
+Separate VAE encode and decode contracts if both exist. Define latent shape,
+scale/shift, dtype/range, tiling/chunk order, output media ordering, and stage
+metadata. Measure latent/media transfer against local decode cost. Test tiling
+seams, RNG/determinism where relevant, chunk ordering, codec/postprocess,
+backpressure, abort, and worker failure.
+
+Do not label VAE tiling, patch parallelism, compile, or offload as VAE
+disaggregation. Likewise, a separate text-encoder TP group is not a separate
+stage. If no reusable generic stage/connector exists in the target revision,
+produce a feasibility result or RFC and keep the capability `not tested` rather
+than inventing a model-local production architecture.
+
+## Performance acceptance
+
+For each accepted optimization, require:
+
+- same fixed workload and correctness artifacts before/after;
+- actual fast-path execution proven by logs/trace/counters;
+- local metric plus E2E latency/throughput and memory;
+- at least three raw fixed-work repetitions after explicit warmup, or a
+  declared arrival-load sample set large enough for every tail percentile;
+- named hardware, backend, topology, dtype, and task scope;
+- fallback/rejection tests for unsupported shapes/platforms;
+- no regression in abort/error cleanup or long-run memory trend.
+
+Update the feature matrix narrowly. Do not convert a profile-driven direction
+into a generic “supported” feature merely because a model-specific PR exists.

.claude/skills/production-add-diffusion-model/references/production-validation.md

@@ -0,0 +1,464 @@
+# Production Serving and Validation
+
+## Contents
+
+1. [Execution-mode decision](#execution-mode-decision)
+2. [Request-level batching](#request-level-batching)
+3. [Step execution and continuous batching](#step-execution-and-continuous-batching)
+4. [Abort and cleanup](#abort-and-cleanup)
+5. [Mixed-RPS soak](#mixed-rps-soak)
+6. [Four CI tracks](#four-ci-tracks)
+7. [PR evidence checklist](#pr-evidence-checklist)
+
+## Execution-mode decision
+
+Request batching and step-wise execution solve different problems:
+
+| Mode | Pipeline contract | Scheduling/abort boundary |
+|---|---|---|
+| Serial request | `forward(DiffusionRequestBatch)` for one request | Whole request; in-denoise preemption is unavailable |
+| Request batching | `supports_request_batch = True`; one forward consumes compatible independent requests | Whole batch/request |
+| Single-request step | Full `SupportsStepExecution`; `max_num_seqs=1` | Between denoise steps |
+| Step continuous batching | Full step contract plus heterogeneous batched-step evidence | Between denoise waves/steps |
+
+Do not expose four stub methods as an opt-out. The current protocol is
+structural; if a native pipeline is not step-capable, do not define the four
+step methods. If it is capable, implement and test all four.
+
+Treat structural support, a generic request-batch bridge, and model benefit as
+three separate claims. Before recommending step execution, name a hypothesis
+that request mode cannot satisfy, define a success threshold, and run a matched
+request-mode control. If throughput or latency regresses, retain step execution
+as functional/limited and keep request mode as the production recommendation.
+
+Choose compatibility keys from every value that changes tensor shapes,
+semantics, hooks, or scheduler behavior: task, output shape/count, frames,
+steps/schedule, guidance/CFG, dtype/backend, LoRA/adapter, cache quality policy,
+and model-specific conditions. Reject or schedule separately rather than
+silently using the first request's settings.
+
+## Request-level batching
+
+A request-batched pipeline returns exactly one `DiffusionOutput` per logical
+request. This is model-owned pseudocode: `validate_batch_compatibility()` and
+`_generate_batch()` are placeholders to implement from the official contract,
+not shared APIs to import.
+
+```python
+import torch
+
+from vllm_omni.diffusion.data import DiffusionOutput
+from vllm_omni.diffusion.worker.request_batch import (
+    DiffusionRequestBatch,
+    split_diffusion_output_by_request,
+)
+
+class MyPipeline(torch.nn.Module):
+    supports_request_batch = True
+
+    def forward(self, req: DiffusionRequestBatch) -> list[DiffusionOutput]:
+        validate_batch_compatibility(req.sampling_params_list, req.prompts)
+        common = req.sampling_params_list[0]
+
+        text_prompts: list[str] | None = [
+            prompt
+            if isinstance(prompt, str)
+            else (prompt.get("prompt") or "")
+            for prompt in req.prompts
+        ]
+
+        prompt_fields = DiffusionRequestBatch.collate_prompt_field_map(
+            req.prompts,
+            {"prompt_embeds": None, "prompt_mask": None},
+        )
+        if prompt_fields["prompt_embeds"] is not None:
+            text_prompts = None
+        generators = req.collate_request_generators(
+            common.num_outputs_per_prompt,
+            default_generator=None,
+        )
+        output = self._generate_batch(
+            prompt=text_prompts,
+            prompt_fields=prompt_fields,
+            generators=generators,
+            sampling=common,
+        )
+        return split_diffusion_output_by_request(
+            DiffusionOutput(output=output),
+            req,
+            num_outputs_per_prompt=common.num_outputs_per_prompt,
+        )
+```
+
+The prompt-field collator only handles tensor fields; extract raw text prompts
+and model-specific negative/alias fields separately, as current in-tree batched
+pipelines do. Set a raw prompt field to `None` when the matching precomputed
+embedding is supplied. The exact aliases and generator device depend on the
+model. Use current collators rather than manual `torch.cat` when they cover the
+contract; they reject mixed present/missing tensors and incompatible
+shape/dtype/device.
+
+Before reading `sampling_params_list[0]`, validate every field that must be
+common. Tests must cover:
+
+- two and N requests, independent seeds/generators, one output each;
+- mixed prompt lengths and packed/padded boundaries;
+- multiple outputs per prompt and correct result slicing;
+- compatible versus incompatible shapes, steps, guidance, LoRA, and quality;
+- one request failure without cross-request output or state corruption;
+- output ordering and request IDs under different completion times.
+
+After the model advertises support, candidate serving is:
+
+```bash
+vllm serve '<model>' --omni \
+  --max-num-seqs 4 \
+  --request-batch-max-wait-ms 20
+```
+
+The wait window trades admission latency for batch formation. Benchmark at 0
+and candidate nonzero values under the target arrival process.
+
+## Step execution and continuous batching
+
+### Full protocol skeleton
+
+Use the current shared state types and make every mutable value request-local:
+
+The helpers beginning with `_` and `validate_step_batch()` below are
+model-owned placeholders; scheduler initialization is scheduler-specific.
+
+```python
+import copy
+from collections.abc import Sequence
+from typing import Any, ClassVar
+
+import torch
+
+from vllm_omni.diffusion.data import DiffusionOutput
+from vllm_omni.diffusion.models.interface import SupportsStepExecution
+from vllm_omni.diffusion.worker.input_batch import InputBatch
+from vllm_omni.diffusion.worker.utils import StepRequestState
+
+class MyPipeline(torch.nn.Module, SupportsStepExecution):
+    supports_step_execution: ClassVar[bool] = True
+
+    def prepare_encode(
+        self,
+        state: StepRequestState,
+        **kwargs: Any,
+    ) -> StepRequestState:
+        prepared = self._prepare_request(state.prompt, state.sampling)
+        state.prompt_embeds = prepared.prompt_embeds
+        state.latents = prepared.latents
+        state.timesteps = prepared.timesteps
+        state.scheduler = copy.deepcopy(self.scheduler)
+        state.scheduler.set_begin_index(0)
+        state.step_index = 0
+        state.extra["task_state"] = prepared.task_state
+        return state
+
+    def denoise_step(
+        self,
+        input_batch: InputBatch,
+        *,
+        states: Sequence[StepRequestState] | None = None,
+        **kwargs: Any,
+    ) -> torch.Tensor | None:
+        validate_step_batch(input_batch, states)
+        return self.transformer(**build_batched_denoise_inputs(input_batch))
+
+    def step_scheduler(
+        self,
+        state: StepRequestState,
+        noise_pred: torch.Tensor,
+        **kwargs: Any,
+    ) -> None:
+        timestep = state.current_timestep
+        if timestep is None:
+            raise RuntimeError("step_scheduler called without a current timestep")
+        state.latents = state.scheduler.step(
+            noise_pred,
+            timestep,
+            state.latents,
+            return_dict=False,
+        )[0].to(state.latents.dtype)
+        state.step_index += 1
+
+    def post_decode(
+        self,
+        state: StepRequestState,
+        **kwargs: Any,
+    ) -> DiffusionOutput:
+        return DiffusionOutput(output=self.decode(state.latents))
+```
+
+This is a shape guide, not drop-in code: verify the target revision's
+`StepRequestState.extra`, scheduler signature, CFG handling, state fields, and
+decode contract. Follow a current native pipeline such as Qwen-Image for the
+exact revision.
+
+Never use the shared `self.scheduler` to advance a request. Store scheduler,
+latents, embeddings, timesteps, masks, RNG/generator, task metadata, cache
+policy, adapter, and output assembly state on `StepRequestState` or its
+request-local extension. Pipeline fields may hold immutable weights/config and
+bounded immutable caches only.
+
+### Qualification sequence
+
+1. Compare request-mode and step-mode at `max_num_seqs=1` for every task using
+   identical seed/schedule/shape and intermediate latents.
+2. Abort queued and in-flight requests at several step boundaries.
+3. Increase to two compatible requests; verify independent scheduler/RNG/state
+   and output order.
+4. Increase to the target capacity and mix prompt lengths.
+5. Exercise mixed shape, frames, steps/schedule, task, guidance, and quality.
+   Batch only compatible rows; otherwise split/reject predictably.
+6. Inject a failure in encode, denoise, scheduler, and decode independently.
+7. Run the production soak and fault matrix.
+
+Candidate commands only after capability validation:
+
+```bash
+# Conservative step execution and step-boundary abort.
+vllm serve '<model>' --omni \
+  --step-execution \
+  --max-num-seqs 1
+
+# Experimental batched-step candidate after multi-request E2E.
+vllm serve '<model>' --omni \
+  --step-execution \
+  --max-num-seqs 4
+```
+
+If an attention backend is only correct for single-request step mode, fail
+early or keep the recipe at `max_num_seqs=1`. Do not describe generic runtime
+support as model support.
+
+## Abort and cleanup
+
+### Required semantics
+
+For queued and in-flight abort/disconnect:
+
+- emit exactly one terminal aborted/error result;
+- do not run `post_decode` after abort unless a documented partial-output mode
+  explicitly requires it;
+- do not increment `step_index` or apply a scheduler result after cancellation;
+- remove scheduler/state cache entries and request IDs;
+- release latents, prompt/reference tensors, cache hooks, DLO buffers, connector
+  messages, uploaded/temp files, and output buffers;
+- keep other requests alive;
+- make cleanup idempotent when disconnect, timeout, and worker error race;
+- drop late results/exceptions for already-resolved or cancelled futures
+  without terminating the shared result-pump thread;
+- complete the next valid request with its requested cache/offload policy.
+
+Test cancellation before engine admission, while queued, after encode, during
+early/middle/final denoise steps, before decode, and while returning/persisting
+output. Include async job cancellation and client socket disconnect.
+
+### Failure injection matrix
+
+| Location | Injection | Assertions |
+|---|---|---|
+| Validation/upload | bad MIME, count/bytes overflow, truncated media | 400/413, no job/engine submit, no temp leak |
+| Load/start | missing fused shard, unsupported topology | startup fails, actionable error, no silent init |
+| Encode | exception/OOM/timeout | one error, request state/temp released |
+| Denoise | exception/OOM/worker exit | peers handled, hooks/buffers released, health behavior explicit |
+| Scheduler | invalid state/shape | no latent/step partial commit |
+| Decode/post | exception/codec error | no corrupt successful result, partial output removed |
+| Result/IPC | late result after cancel, large split tensor, orphan SHM | pump remains alive, one terminal result, exact ownership cleanup |
+| Connector/disagg | timeout/retry/duplicate | idempotency, backpressure, no orphan payload |
+
+After each injection, check server health and issue a known-good request. If
+the process must restart, document and test the restart policy; do not claim
+transparent recovery.
+
+### Large-output transport
+
+Treat worker-to-engine IPC and engine-to-client transport as separate hops.
+For batched video, prove that each logical output does not accidentally retain
+or serialize the complete batch storage. Exercise the target revision's shared
+memory or artifact-handle path above and below its threshold, including
+non-contiguous views, multiple outputs, consumer failure, timeout, abort, and
+worker exit. Verify every shared segment/file is released exactly once.
+
+Freeze the media contract at each hop: dtype, value range, shape/layout,
+contiguity, request ownership, encoding/color model, payload bytes, and pending
+consumers. Device-side float-to-uint8 preparation, D2H/IPC, planar/interleaved
+conversion, CPU encode/mux, and delivery off the asyncio event loop are
+different optimizations. Validate their route selection and fallbacks
+independently. Byte-identical HTTP MP4 output does not prove that a raw offline
+tensor kept its dtype or range.
+
+For long-window output, validate segment ordering and audio/video continuity,
+then exercise payloads around serializer, shared-memory, HTTP, and codec size
+limits. Chunking must reassemble exactly, reject malformed/missing chunks, and
+release partial state after abort, timeout, or peer failure.
+
+Qualify chunk production, transport, and public streaming as separate layers.
+A VAE callback must define owned representation, frame offsets, final marker,
+rank ownership, source/config gate, and collective-safe failure, but it does not
+prove SHM/ZMQ delivery, bounded backpressure, cancellation, codec reuse, or
+client streaming. Measure time to first owned chunk, first encoded fragment,
+steady cadence, mux finalization, and complete client artifact independently.
+
+Measure remote MP4 encoding, serialization/copies, network transfer, and client
+materialization separately from denoise/decode. An inference-kernel speedup does
+not establish serving throughput if output transport dominates or leaks.
+
+## Mixed-RPS soak
+
+Use a realistic request mix and arrival distribution. Include small/large
+shapes, short/long prompts, all advertised tasks, cache tiers, and a small rate
+of abort/disconnect/error injections. Run phases below, near, and above measured
+saturation so queue/backpressure behavior is visible.
+
+Example serving benchmark phase:
+
+```bash
+python benchmarks/diffusion/diffusion_benchmark_serving.py \
+  --base-url http://127.0.0.1:8091 \
+  --endpoint /v1/videos \
+  --model '<model>' \
+  --dataset trace \
+  --dataset-path /path/to/versioned-production-trace.jsonl \
+  --task t2v \
+  --num-prompts 500 \
+  --request-rate 0.5 \
+  --max-concurrency 8 \
+  --warmup-requests 8 \
+  --warmup-concurrency 8 \
+  --output-file soak-near-saturation.json
+```
+
+Use a task-appropriate trace/media dataset. Run the same phase at rates selected
+from an initial saturation curve; a literal `0.5` is illustrative only.
+
+On supported GPU hosts, wrap a test or soak with the repository monitor:
+
+```bash
+bash tests/dfx/stability/scripts/resource_monitor.sh run --backend gpu -- \
+  pytest -s -v tests/dfx/stability/scripts/test_stability_<model>.py -m slow
+```
+
+Check the target revision's monitor backends. Do not claim CPU/NPU monitoring
+from a reserved CLI choice unless collection is implemented and verified.
+
+Report:
+
+- offered and achieved RPS, success/error/abort rates;
+- p50/p95/p99 and queue/service/stage times, with the successful sample count
+  and arrival process used for each percentile;
+- throughput per device and batch/wave utilization;
+- per-rank HBM and process-tree PSS trend/slope;
+- cache size/hit rate, temp-file/disk growth, open FDs/handles;
+- outstanding shared-memory/artifact handles and encoded/network bytes;
+- worker/engine health, restarts, OOM/timeouts, backpressure response;
+- total duration, request count/mix, raw logs/JSON/monitor artifact hashes.
+
+A short burst or repeated serial curl loop is not a stability test.
+
+## Four CI tracks
+
+Keep the tracks independent so a failure answers a specific question. These
+four readiness tracks are not the repository's numbered CI levels. Map them
+onto the current L1-L5 taxonomy:
+
+| Readiness track | Repository CI placement |
+|---|---|
+| Function | L1 unit/logic plus L2 basic online/offline E2E; extend real-model scenarios in L3/L4 |
+| Accuracy | L3 real-model gates for important cases; deeper and broader nightly coverage in L4 |
+| Performance | Important thresholds in L3 where they fit the time budget; full baselines/regression jobs in nightly L4 |
+| Reliability | Cheap rejection/recovery assertions can run earlier, but long stability, fault injection, and reliability suites belong to weekly L5 |
+
+Distributed topology is a test case within the appropriate level, not a
+replacement definition for a level.
+
+### 1. Function CI
+
+Cover:
+
+- strict unquantized and advertised quantized loading;
+- fused source-shard completeness and prefix routing;
+- official task positive boundaries and neighboring negative cases;
+- offline, sync, async normalization/parity;
+- single-device reference plus every advertised topology smoke;
+- output type/count/order and request ID isolation;
+- feature rejection paths that must fail early.
+
+### 2. Accuracy CI
+
+Pin official implementation/checkpoint/vLLM-Omni revisions, prompts and media
+hashes, seed, scheduler/sigma/steps, dimensions/frames/FPS, guidance, dtype,
+backend, and topology. Compare at suitable levels:
+
+- component outputs or selected blocks;
+- one or more intermediate denoise latents;
+- final image/video/audio artifact metrics;
+- temporal consistency and audio/video synchronization where relevant.
+
+Set tolerance per modality/hardware/backend with rationale. Keep failure
+artifacts and metadata. A single visual inspection does not define CI.
+
+For quantization trajectory tooling, inspect the target revision's current
+`vllm_omni/quantization/tools/compare_diffusion_trajectory_similarity.py` and
+reuse it when compatible instead of creating an ad hoc comparator.
+
+### 3. Performance CI
+
+Pin workload and best deployment. Include a resident recommended row and a
+small-HBM/DLO row when both are production targets. Declare:
+
+- one explicit warmup exclusion and at least three measured fixed-work
+  repetitions;
+- raw result JSON and environment identity;
+- median/mean plus range for small fixed-work A/Bs; p50/p95/p99 only for an
+  arrival-load run with enough samples, plus throughput/RPS, stage time,
+  HBM/PSS;
+- regression metric, threshold, baseline owner, and update procedure;
+- correctness/accuracy check linked to the same configuration.
+
+Use the current harness config style, for example:
+
+```bash
+pytest -s -v tests/dfx/perf/scripts/run_diffusion_benchmark.py \
+  --test-config-file tests/dfx/perf/tests/test_<model>_vllm_omni.json
+```
+
+Verify parser options on the target revision. Do not copy a JSON schema from a
+different model without matching its endpoint, task, dataset, and deploy path.
+
+### 4. Reliability CI
+
+Run a bounded CI soak plus a longer scheduled soak. Cover:
+
+- below/near/above-saturation arrival phases;
+- queued/in-flight abort and disconnect;
+- OOM/timeout and worker/process failure appropriate to the platform;
+- DLO/cache/connector cleanup;
+- health, fast-fail/backpressure, restart/recovery policy;
+- memory/temp/FD trend and successful known-good request after each fault.
+
+Keep reliability thresholds independent from latency thresholds. A fast server
+that leaks memory or hangs after cancellation fails production readiness.
+
+## PR evidence checklist
+
+The PR description must include or link:
+
+- official source/checkpoint revisions and API/limit matrix;
+- scoped feature compatibility matrix with the four allowed states;
+- implementation notes for FP8, DLO, request cache, attention/fusion, batching;
+- exact environment, deploy/serve, every task curl, benchmark and soak commands;
+- fixed accuracy inputs and artifacts;
+- raw before/after performance runs and memory traces;
+- output-contract manifests and stage accounting from device preparation
+  through the complete client artifact;
+- abort/disconnect/fault results and cleanup evidence;
+- Function/Accuracy/Performance/Reliability CI files and ownership;
+- limitations/TODOs that remain `not tested` rather than implied support.
+
+Separate the Day-0 milestone from production readiness. If the PR only closes
+some gates, title and description should say which scoped gates it closes.

.claude/skills/readme.md

@@ -26,14 +26,24 @@ self-check.
 
 - [`add-diffusion-model`](add-diffusion-model/SKILL.md): guides integration of
   a new diffusion model into `vllm-omni`
+- [`add-recipe`](add-recipe/SKILL.md): creates or updates in-repository model
+  recipes with verified task/input/output contracts, concrete hardware,
+  shared feature-guide links, exact commands, and qualification evidence
 - [`add-tts-model`](add-tts-model/SKILL.md): covers integration of new TTS
   models and related serving workflows
 - [`diffusion-perf-opt`](diffusion-perf-opt/SKILL.md): guides diffusion model
   performance optimization, including profiling traces, parallel strategies,
   stage timing analysis, and benchmark-driven tuning
+- [`find-simplifications`](find-simplifications/SKILL.md): finds evidence-backed
+  opportunities to remove or merge dead, duplicated, speculative,
+  over-generalized, or unnecessarily defensive vLLM-Omni code
 - [`precheck-pr`](precheck-pr/SKILL.md): self-checks a branch before creating a
-  PR by validating title format, dead code, accuracy and performance claims,
-  and merge readiness
+  PR by validating title format, dead code, simplification opportunities,
+  accuracy and performance claims, and merge readiness
+- [`production-add-diffusion-model`](production-add-diffusion-model/SKILL.md):
+  takes a working Day-0 diffusion integration through API/limit parity,
+  feature-combination evidence, quantization/offload, performance, production
+  serving, hardware recipes, and CI readiness
 - [`quantization`](quantization/SKILL.md): guides quantization method selection,
   model integration, checkpoint loading, and quality/performance validation
   for vLLM-Omni

.claude/skills/review-pr/SKILL.md

@@ -21,7 +21,10 @@ Make every finding:
 - **Calibrated:** match severity to user and maintainer impact.
 
 Do not report unrelated backlog, style already enforced by pre-commit, or a
-missing test that would not protect changed behavior.
+missing test that would not protect changed behavior. Do report new allowlist
+or budget entries in `check_forbidden_imports.py` / `check_torch_cuda.py` /
+`check_tts_adapter.py` / `check_buildkite.py` unless the PR justifies them:
+those are policy changes, not lint noise.
 
 ## Select the input and depth
 
@@ -92,11 +95,13 @@ If docs and live code disagree, verify the code/tests and report the drift.
 
 | Reference | Read when |
 | --- | --- |
-| [model-addition-checklist.md](references/checks/model-addition-checklist.md) | A model, architecture, loader, processor, registry, or stage config is added. |
+| [model-addition-checklist.md](references/checks/model-addition-checklist.md) | A model, architecture, loader, processor, registry, pipeline config, or deploy config is added. |
 | [perf-verification.md](references/checks/perf-verification.md) | The PR makes a latency, throughput, memory, or quality claim. |
 | [test-quality-evaluation.md](references/checks/test-quality-evaluation.md) | Tests change, are absent for risky code, or may not exercise production behavior. |
 | [tests-docs-checklist.md](references/checks/tests-docs-checklist.md) | Coverage, CI markers, examples, user docs, or PR evidence need review. |
 | [verification.md](references/checks/verification.md) | Hardware, a server, or a runnable affected path is available for active verification. |
+| [examples-policy.md](../precheck-pr/references/examples-policy.md) | The PR adds, copies, or renames Python under `examples/`; apply the canonical policy shared with `precheck-pr`. |
+| [find-simplifications](../find-simplifications/SKILL.md) | Every review; run a diff-scoped subtraction and simplification pass after correctness blockers. |
 
 ### Delivery and reviewer coordination
 
@@ -118,10 +123,11 @@ it changes again, report the churn and wait for a stable target.
 
 For a trusted PR head, materialize the pinned head in an isolated detached
 worktree. A worktree freezes identity but is not a security sandbox. Treat fork
-heads as untrusted unless the user and environment policy explicitly establish
-otherwise: execute them only in a disposable, secret-free sandbox with restricted
-filesystem, network, and resources; without one, use static SHA-addressed reads
-and CI evidence only. For a local review, freeze the committed, index, worktree,
+heads as untrusted until the user and environment policy explicitly establish
+trust: execute their code only after that trust is recorded for this host,
+with the trusted SHA noted in the review state and secrets kept out of the
+execution scope; without recorded trust, use static SHA-addressed reads and CI
+evidence only. For a local review, freeze the committed, index, worktree,
 and NUL-safe in-scope untracked contents. Follow
 [review-execution.md](references/process/review-execution.md) for trust gates,
 state fingerprints, and byte-for-byte staleness checks.
@@ -152,6 +158,13 @@ contract they protect or use only the applicable evidence checks.
 Apply every category in [general-checks.md](references/process/general-checks.md) before
 lower-priority comments.
 
+If the diff census contains an added, copied, or renamed Python path under
+`examples/`, read and apply the canonical
+[examples policy](../precheck-pr/references/examples-policy.md). Treat a new
+model-specific Python example as blocking. Do not flag model-specific example
+debt that the PR only modifies or removes, and do not run the rest of the
+author-oriented `precheck-pr` workflow.
+
 For each changed value or behavior, trace:
 
 ```text
@@ -172,15 +185,19 @@ promotion gate. Candidate or draft rules are questions, not blockers, unless
 current code, tests, or policy enforce them. Inspect both sides of any config,
 registry, serialization, connector, cache, or stage boundary.
 
-When a diff adds or expands a helper, class, fallback, compatibility branch, or
-public behavior, run a subtraction pass: remove out-of-scope behavior and check
-whether each new abstraction can be deleted, merged, moved, or inlined.
+Read and apply [find-simplifications](../find-simplifications/SKILL.md) on every
+review. Constrain it to the diff and the adjacent ownership, callers, or
+consumers needed to prove a candidate. Check whether added or expanded helpers,
+classes, state, fallback and compatibility branches, data movement, or public
+behavior can be deleted, merged, moved, or inlined. Zero candidates is a valid
+result. Do not widen the review into repository backlog or report speculative
+style preferences as simplification findings.
 
 ### 6. Verify the changed path
 
 Before each validation group, verify the frozen SHA plus the tracked, index,
 untracked, and ignored-file fingerprint, or recreate a pristine snapshot. On a
-trusted head or inside the required sandbox, run an import/version preflight,
+head the user explicitly trusted for this host, run an import/version preflight,
 then the narrowest relevant tests and low-cost static checks. Bind every result
 to the head SHA, snapshot fingerprint, and environment fingerprint. Never run
 imports, tests, builds, hooks, or repo-configurable tooling from an untrusted

.claude/skills/review-pr/references/checks/model-addition-checklist.md

@@ -1,7 +1,7 @@
 # Model Addition Checklist
 
 Read this reference when a PR adds or expands a model architecture, pipeline,
-loader, processor, registry entry, or stage configuration.
+loader, processor, registry entry, pipeline config, or deploy config.
 
 Official docs: [model contribution guides](https://docs.vllm.ai/projects/vllm-omni/en/latest/contributing/model/),
 [adding an omni model](https://docs.vllm.ai/projects/vllm-omni/en/latest/contributing/model/adding_omni_model/),

.claude/skills/review-pr/references/checks/perf-verification.md

@@ -22,6 +22,19 @@ Keep exact commands and report variability rather than a cherry-picked run.
 Use the repository's benchmark when it covers the claim; otherwise use the
 smallest reproducible workload that reaches the changed production path.
 
+## Isolation per claim
+
+Isolate one claim per comparison. When a PR stacks several changes (a fused or
+restructured load path, a distilled step schedule, a new cache), request one
+column per claim at equal steps/workload instead of one blended base-vs-head
+delta, and name the unchanged stage or metric that pins equivalence.
+
+When the changed path is runtime-switchable (feature flag or environment
+variable), run every column on the same head and toggle the switch rather than
+diffing against a base commit that cannot separate the stacked changes:
+base, head with the switch off, head with the switch on, then head with the
+switch on plus each further claim.
+
 ## Evidence table
 
 | Dimension | Typical evidence |
@@ -34,6 +47,100 @@ smallest reproducible workload that reaches the changed production path.
 Every optimization needs correctness or quality evidence. A faster result on a
 different workload, precision, seed, or topology is not a valid comparison.
 
+## E2e stage-attribution table
+
+For an end-to-end latency claim, request the stage-attribution table the
+repository's perf PRs use:
+
+- Rows carry the benchmark's own field names: total e2e, per-stage execution,
+  postprocess, response encoding, and a request-to-output roll-up. Add a stage
+  handoff/idle row when the pipeline spans stages — connector transfer and
+  scheduling gaps are invisible in per-stage latency deltas.
+- Sanity rows pin comparability: inference steps, output frames/pixels/samples,
+  and generated tokens held identical across columns.
+- Report P50 and P100 from the stated warmup and measured counts. One
+  population per number: never mix cold-start and warm samples into one value.
+- Name the baseline and candidate commit SHAs, and close with an isolation
+  statement naming the metric that stayed unchanged ("X remained effectively
+  unchanged, isolating the gain to Y").
+
+When the serving path caches, captures, or regenerates at runtime (CUDA-graph
+bucket growth, cache rebuilds, recompilation), medians hide the resulting tail.
+Also request the event count and the maximum observed stall.
+
+Component-level tables pair with, not replace, the e2e table: per-op reference
+vs candidate timings (with the numerical-equality result per op) plus one
+complete-pipeline check under the production shapes.
+
+## Proportionality
+
+When a PR adds a model and explicitly labels its numbers as smoke observations
+rather than claims, do not demand the attribution machinery. Ask only for
+evidence protecting the model's stated operating contract — for a
+realtime/duplex model that is sustained cadence over a realistic session
+(enough ticks to leave warmup), with any mid-session capture/regeneration
+event and its stall reported, since a cadence that collapses mid-session is a
+functional failure, not a perf claim.
+
+## Realtime/duplex world models
+
+Request the operating-contract evidence as serving metrics, not claims:
+
+- Sustained per-tick latency over one realistic session: median, max, and the
+  argmax tick.
+- Stall events with their causes.
+- Peak memory at first versus last tick — session-owned state must not grow
+  per tick.
+- Control latency from a submitted action to the tick that applies it.
+- Realtime-versus-offline parity under the same controls, using the
+  repository's video-similarity helpers. Parity plotted over tick index is the
+  drift-over-horizon curve and the honest measure of session lifetime;
+  distribution-level suites (FVD, VBench) belong to model evaluation, not the
+  serving integration.
+
+Watch the contract boundaries the model documents: if a control input is
+tick-API-only, the parity pair must use a horizon both paths can run.
+
+## Speech-generation (TTS) models
+
+The streaming contract outranks the average — a synthesis that pauses
+mid-stream is broken for realtime use regardless of its averages. Request:
+
+- The benchmark's own speech metrics over one long utterance (enough audio to
+  leave warmup, tens of seconds): `audio_ttfp` (time to first packet),
+  `audio_rtf` (real-time factor; above 1.0 cannot keep up), `e2el`, and
+  `ttft`/`tpot` where tokens precede audio.
+- Mid-utterance stalls as the benchmark's continuity fields, not prose:
+  `audio_underrun` percentiles and the `audio_continuity_ok` rate
+  (`vllm_omni/benchmarks/metrics/metrics.py`), computed by
+  `vllm_omni/benchmarks/audio_continuity.py` — the companion to `audio_ttfp`
+  and `audio_rtf`. A synthesis that bursts chunks fast enough on average but
+  underruns mid-stream fails the streaming contract even when the averages
+  pass.
+- Quality as a paired comparison against the reference implementation under
+  the same text, speaker prompt, and seed; report sample rate and frame
+  count, and assert output completeness — truncated or dropped trailing audio
+  is a known failure class, not a quality nuance.
+- For realtime or duplex TTS: interruption (barge-in) latency and the
+  per-session memory trend across turns.
+
+## Omni models
+
+Apply the e2e stage-attribution table directly to the voice-to-voice path of
+omni (speech-to-speech and multimodal AR) models:
+
+- TTFT for text output, `audio_ttfp` for audio output, and per-stage
+  ITL/TPOT cadence across the perception-understanding-generation chain.
+- Cross-modal parity — the omni analogue of realtime-versus-offline parity:
+  the same semantic request through text and audio inputs, compared on output
+  content or timing.
+- For duplex omni sessions: turn-taking latency, interruption handling, the
+  concurrent session count actually validated, and a per-turn memory trend —
+  session-owned KV state must not grow per turn. When either turn produces
+  audio, continuity is judged with the same `audio_underrun` and
+  `audio_continuity_ok` fields as TTS, since turn-taking latency alone does
+  not expose mid-utterance underruns.
+
 ## Reviewer verification ladder
 
 1. Run base/head A/B on suitable hardware when affordable.

.claude/skills/review-pr/references/modules/configuration.md

@@ -3,7 +3,7 @@
 Primary design: [vLLM-Omni Configuration](https://docs.vllm.ai/projects/vllm-omni/en/latest/design/module/vllm_omni_config/).
 
 Use for config construction, deploy YAML, registries, schema, defaults,
-aliases, CLI projection, stage config, and topology. Inspect the reviewed
+aliases, CLI projection, pipeline config, and topology. Inspect the reviewed
 page's status: while its contract remains deferred or draft, current code and
 tests—not proposed precedence rules—are authoritative.
 

.claude/skills/review-pr/references/process/general-checks.md

@@ -121,6 +121,19 @@ Search bounded callers and sibling implementations before reporting dead code,
 duplication, or a missed consumer. Keep new abstractions only when they have a
 distinct live caller, invariant, or compatibility need.
 
+Verify absence and equality claims with whole-file evidence. A line window, a
+grep for one symbol, or a commit message proves presence, never absence: to
+assert that a file matches another version or lacks a change, diff or hash the
+full files (for example `git diff <sha1>:<path> <sha2>:<path>` or the contents
+API's size and SHA fields) before anchoring the finding. A verification
+shortcut that would itself be wrong to run is not evidence.
+
+For a value written or finalized in more than one place, establish which site
+executes first on each entry path — and whether the earlier site makes the
+later ones unreachable — before naming where a fix belongs. Enumerating all
+call sites correctly still misses the bug when a path the enumeration ranked
+last actually finalizes first and clears the state the proposed fix checks.
+
 ## Finding bar
 
 Anchor each finding to a changed `path:line`; name the trigger or call path,

.claude/skills/review-pr/references/process/review-execution.md

@@ -39,13 +39,13 @@ isolated detached worktree. Run every source read, `rg` search, import, and test
 from that snapshot, not the caller's checkout. A detached worktree provides
 snapshot isolation, not a security boundary.
 
-Treat a fork head as untrusted unless the user and environment policy explicitly
-establish trust. Never run its imports, tests, builds, hooks, package setup, or
-repo-configurable linters/plugins on the reviewer host. Execute them only in a
-disposable sandbox or VM with no credentials or inherited secrets, no host agent
-or service sockets, minimal read-only host mounts, disabled or explicitly
-allowlisted network, resource/time limits, and destruction after the run. If
-that boundary is unavailable, limit the review to the remote diff, SHA-addressed
+Treat a fork head as untrusted until the user and environment policy
+explicitly establish trust. Without recorded trust, never run its imports,
+tests, builds, hooks, package setup, or repo-configurable linters/plugins on
+the reviewer host. Once the user explicitly trusts this pinned commit for
+this host, record the trusted SHA with the review state and execute with
+credentials, agent sockets, and other secrets out of scope. Without recorded
+trust, limit the review to the remote diff, SHA-addressed
 reads such as `git show <head_sha>:<path>`, and existing CI evidence; report all
 executable validation as a gap.
 
@@ -111,6 +111,17 @@ Mark early findings as preliminary and continue. This is not a GitHub comment.
   but do not publish review events without separate authorization.
 - Record DCO, pre-commit, required CI, and mergeability. Pending or unknown gates
   do not block source review.
+- GitHub Actions `SKIP` omits SPDX, shellcheck, markdownlint, mypy-3.10, and
+  test-mark coverage. A green GHA pre-commit job does not prove those local
+  gates passed. New files still need the full Linting list: Omni SPDX
+  (`vLLM-Omni project`), no stdlib `re`/`base64` in `vllm_omni/`, no new
+  pickle / Hugging Face Hub API / `torch.cuda` call sites, test marks, TTS
+  adapter ratchet, Buildkite schema, and native shellcheck on macOS/Windows.
+  See [Linting](https://vllm-omni.readthedocs.io/en/latest/contributing/#linting).
+- Expanding `CHECK_IMPORTS[*].allowed_files`, `ALLOWED_FILES`,
+  `MAX_MODEL_TYPE_BRANCHES`, or Buildkite `SKIP_FILES` is a policy change.
+  Do not rubber-stamp allowlist/budget growth; require justification and
+  prefer fixing the call site.
 - Treat a failed gate as its own evidence. Do not restate its formatting/lint
   output as a new code-review finding.
 - Open CI logs only when the first failing step overlaps the frozen diff or

.claude/skills/review-pr/references/process/review-routing.md

@@ -48,7 +48,7 @@ compatibility changes.
 
 | Signal | Read | Optional repo-local skill |
 | --- | --- | --- |
-| New or expanded model, loader, processor, registry, or stage config | [model-addition-checklist.md](../checks/model-addition-checklist.md) | [`add-tts-model`](../../../add-tts-model/SKILL.md) or [`add-diffusion-model`](../../../add-diffusion-model/SKILL.md) |
+| New or expanded model, loader, processor, registry, pipeline config, or deploy config | [model-addition-checklist.md](../checks/model-addition-checklist.md) | [`add-tts-model`](../../../add-tts-model/SKILL.md) or [`add-diffusion-model`](../../../add-diffusion-model/SKILL.md) |
 | Latency, throughput, memory, scaling, precision, or quality claim | [perf-verification.md](../checks/perf-verification.md) | [`diffusion-perf-opt`](../../../diffusion-perf-opt/SKILL.md) for diffusion |
 | Tests changed, absent for risky behavior, or test-only | [test-quality-evaluation.md](../checks/test-quality-evaluation.md) | [`vllm-omni-test`](../../../vllm-omni-test/SKILL.md) |
 | CI, examples, docs, public behavior, or contributor evidence | [tests-docs-checklist.md](../checks/tests-docs-checklist.md) | None |

.claude/skills/vllm-omni-test/SKILL.md

@@ -52,8 +52,8 @@ Always attach markers deliberately:
   - `tts` — speech synthesis / TTS-only models (`/v1/audio/speech`, voice clone, etc.)
   - `diffusion` — generative diffusion models (image / audio / text / video from noise)
 - **Cross-cutting area** (when relevant): `parallel`, `cache`, `example`, `benchmark`
-- **Hardware**: `cpu`, `gpu`, `cuda`, `rocm`, `npu`, `L4`, `H100`, `distributed_cuda`, …
-- Optional: `slow`, distributed markers when multi-card is required
+- **Hardware**: `cpu`, `gpu`, `cuda`, `rocm`, `npu`, …; SKU (`H100`, `L4`, …) and `cards_{n}` via `@hardware_test(...)` / `hardware_marks(...)` only — never `pytest.mark.H100` (`check-mark` rejects it)
+- Optional: `slow`. Multi-card: `num_cards=N` on the helper (filter with `cards_N` / `not cards_1`)
 
 **Baseline smoke (L2 + L3):** The simplest e2e case per model — default deploy, minimal request — should usually carry **both** `@pytest.mark.core_model` and `@pytest.mark.advanced_model` on the **same** test function so `test-ready.yml` and `test-merge.yml` share one test. `send_*_request` picks validation depth from `--run-level`. References: `test_voxcpm2_tts.py::test_text_to_audio_001`, `test_qwen3_tts_customvoice.py::test_text_to_audio_001`. Heavier scenarios use **`advanced_model` only**; L4 expansion uses **`full_model`**.
 
@@ -94,8 +94,8 @@ Existing references: `tests/e2e/offline_inference/test_qwen2_5_omni.py` (L2-styl
 
 | Scenario | Typical location | Fixtures / runner pattern | Baseline markers & level |
 |----------|------------------|---------------------------|---------------------------|
-| **Offline inference e2e** | `tests/e2e/offline_inference/` | **Module (default):** `omni_runner` + `omni_runner_handler`. **Function (isolation only):** `omni_runner_function` + `omni_runner_handler_function`. Diffusion/TTS may use `Omni(...).generate` directly | L2: `core_model` + **one of** `omni` / `tts` / `diffusion`; `@hardware_test(...)` when GPU/NPU is required |
-| **Online serving e2e** | `tests/e2e/online_serving/` | **Module (default):** `omni_server` + `openai_client`. **Function (isolation only):** `omni_server_function` + `openai_client_function`. Clients: `send_omni_request` (omni), `send_audio_speech_request` (tts), `send_diffusion_request` / `send_video_diffusion_request` / `send_images_generations_request` (diffusion) | Baseline smoke: **`core_model` + `advanced_model`**; heavier paths: `advanced_model` only; L4 expansion: `full_model` |
+| **Offline inference e2e** | `tests/e2e/offline_inference/` | **Module (default):** `omni_runner` + `offline_client`. **Function (isolation only):** `omni_runner_function` + `offline_client_function`. Diffusion/TTS may use `Omni(...).generate` directly | L2: `core_model` + **one of** `omni` / `tts` / `diffusion`; `@hardware_test(...)` when GPU/NPU is required |
+| **Online serving e2e** | `tests/e2e/online_serving/` | **Module (default):** `omni_server` + `online_client`. **Function (isolation only):** `omni_server_function` + `online_client_function`. Clients: `send_omni_request` (omni), `send_audio_speech_request` (tts), `send_diffusion_request` / `send_video_diffusion_request` / `send_images_generations_request` (diffusion) | Baseline smoke: **`core_model` + `advanced_model`**; heavier paths: `advanced_model` only; L4 expansion: `full_model` |
 | **Documentation / runnable examples** | `tests/examples/offline_inference/`, `tests/examples/online_serving/` | **Offline docs (preferred):** extract Python/Bash blocks from the doc README (e.g. `ReadmeSnippet.extract_readme_snippets`), `pytest.mark.parametrize` each snippet, run via `example_runner.run` with a stable `output_subfolder`. **Online docs:** copy client/request scripts into dedicated tests and keep them in sync with the doc page. | Usually **L4**: `advanced_model`, often `example` plus hardware marks matching the nightly docs-example job (see `.buildkite/cuda/test-nightly.yml`). Full conventions: [docs/contributing/ci/test_examples/l4_doc_example_tests.inc.md](../../../docs/contributing/ci/test_examples/l4_doc_example_tests.inc.md) (introduced in [PR #1910](https://github.com/vllm-project/vllm-omni/pull/1910): naming, output directory layout, skip rules, avoid trimming `num_inference_steps` without a strong CI reason). |
 | **Performance / benchmark** | `tests/dfx/perf/tests/*.json` + `run_*_benchmark.py` | JSON or script-driven server + load config; assert explicit metrics / baselines | L4 Perf: JSON `mark` with `full_model` + `omni`/`tts`/`diffusion`; wire `test-nightly.yml` Perf steps |
 | **Invalid parameter / negative HTTP validation** | `tests/dfx/reliability/invalid_param_test/` | Live `omni_server` + low-level `send_*_http_request` with `err_code` / `err_message` | `pytest.mark.slow` + `omni` / `tts` / `diffusion` + `@hardware_marks` (`H100` or `L4`); CI in **`test-weekly.yml`** (not ready/merge/nightly) |
@@ -110,10 +110,10 @@ After choosing offline/online/docs/perf, classify the **product under test** and
 |-----------|------------------------------|----------------------------|----------------------------------------|
 | **What it is** | Multimodal LLM pipeline (thinker/talker/stages; text + vision + audio I/O) | Speech synthesis / voice models | Generative diffusion (noise → image, audio, text, or video) |
 | **Examples** | Qwen2.5-Omni, Qwen3-Omni | Qwen3-TTS, VoxCPM2, Higgs-Audio, Step-Audio2 | Qwen-Image, BAGEL, Wan2.2, HunyuanVideo |
-| **Offline runner** | `omni_runner` + `omni_runner_handler`, `generate_multimodal` | `Omni(...).generate` or stage YAML + TTS params | `Omni(...).generate` + `OmniDiffusionSamplingParams` |
-| **Online client** | `openai_client.send_omni_request` (chat completions, modalities) | `openai_client.send_audio_speech_request` (`/v1/audio/speech`) | `send_diffusion_request` (chat/T2I), `send_video_diffusion_request` (`/v1/videos`, X2V), or `send_images_generations_http_request` / `send_images_edits_http_request` (DALL-E routes) |
-| **Typical assertions** | Stage outputs, text/audio keywords via `OmniRunnerHandler` / response handler | WAV bytes, stream chunks, speech endpoint contract | Image/video dimensions, `final_output_type`, `assert_diffusion_response` |
-| **Stage / deploy YAML** | Per-model omni stage configs (`ci/qwen3_omni_moe.yaml`, …) | `qwen3_tts.yaml`, `voxcpm2.yaml`, … | Often default serve; parallel/offload YAML for heavy DiT |
+| **Offline runner** | `omni_runner` + `offline_client`, `generate_multimodal` | `Omni(...).generate` with deploy YAML + TTS params | `Omni(...).generate` + `OmniDiffusionSamplingParams` |
+| **Online client** | `online_client.send_omni_request` (chat completions, modalities) | `online_client.send_audio_speech_request` (`/v1/audio/speech`) | `send_diffusion_request` (chat/T2I), `send_video_diffusion_request` (`/v1/videos`, X2V), or `send_images_generations_http_request` / `send_images_edits_http_request` (DALL-E routes) |
+| **Typical assertions** | Stage outputs, text/audio keywords via `OfflineOmniClient` / response handler | WAV bytes, stream chunks, speech endpoint contract | Image/video dimensions, `final_output_type`, `assert_diffusion_response` |
+| **Deploy YAML** | Per-model deploy configs (`ci/qwen3_omni_moe.yaml`, …) | `qwen3_tts.yaml`, `voxcpm2.yaml`, … | Often default serve; parallel/offload YAML for heavy DiT |
 | **Nightly group** (`test-nightly.yml`) | **Omni Model Test** — `-m "full_model and omni"` | **TTS Model Test** — `-m "full_model and tts"` | **Diffusion X2I(&A&T)** *or* **Diffusion X2V** (see Step 3 table; same `diffusion` marker, different YAML group / file shard) |
 | **L4 pressure** | Expansion per modality/model as needed | Expansion + accuracy/perf in TTS group | X2I: merge feature combos per [#1832](https://github.com/vllm-project/vllm-omni/issues/1832); X2V: separate nightly group |
 
@@ -175,9 +175,9 @@ See **L1 unit test constraints (mocking)** below for the full do/don't list.
 @pytest.mark.omni
 @hardware_test(...)
 @pytest.mark.parametrize("omni_runner", test_params, indirect=True)
-def test_<scenario>(omni_runner, omni_runner_handler) -> None:
+def test_<scenario>(omni_runner, offline_client) -> None:
     request_config = {"prompts": ..., "modalities": [...]}  # optional: images, videos, audios
-    omni_runner_handler.send_omni_request(request_config)
+    offline_client.send_omni_request(request_config)
 ```
 
 *Offline generative e2e — **Diffusion** (representative):*
@@ -187,8 +187,8 @@ def test_<scenario>(omni_runner, omni_runner_handler) -> None:
 @pytest.mark.diffusion
 @hardware_test(...)
 @pytest.mark.parametrize("omni_runner", test_params, indirect=True)
-def test_text_to_image_001(omni_runner_handler) -> None:
-    omni_runner_handler.send_diffusion_request({"prompt": "...", "extra_body": {"num_inference_steps": 4, ...}})
+def test_text_to_image_001(offline_client) -> None:
+    offline_client.send_diffusion_request({"prompt": "...", "extra_body": {"num_inference_steps": 4, ...}})
 ```
 
 *Offline TTS e2e — **Qwen3-TTS** (two-stage; representative):*
@@ -198,8 +198,8 @@ def test_text_to_image_001(omni_runner_handler) -> None:
 @pytest.mark.tts
 @hardware_test(...)
 @pytest.mark.parametrize("omni_runner", tts_server_params, indirect=True)
-def test_text_to_audio_001(omni_runner, omni_runner_handler) -> None:
-    omni_runner_handler.send_audio_speech_request({
+def test_text_to_audio_001(omni_runner, offline_client) -> None:
+    offline_client.send_audio_speech_request({
         "input": "...",
         "task_type": "Base",
         "ref_audio": REF_AUDIO_URL,
@@ -214,8 +214,8 @@ def test_text_to_audio_001(omni_runner, omni_runner_handler) -> None:
 @pytest.mark.tts
 @hardware_test(...)
 @pytest.mark.parametrize("omni_runner", tts_server_params, indirect=True)
-def test_voice_clone_001(omni_runner_handler) -> None:
-    omni_runner_handler.send_single_stage_tts_request({
+def test_voice_clone_001(offline_client) -> None:
+    offline_client.send_single_stage_tts_request({
         "input": "...",
         "language": "en",
         "prompt_audio_path": REF_AUDIO_PATH,
@@ -231,9 +231,9 @@ def test_voice_clone_001(omni_runner_handler) -> None:
 @pytest.mark.omni
 @hardware_test(...)
 @pytest.mark.parametrize("omni_server", test_params, indirect=True)
-def test_text_to_text_001(omni_server, openai_client) -> None:
+def test_text_to_text_001(omni_server, online_client) -> None:
     request_config = {"model": omni_server.model, "messages": ..., "modalities": ["text"]}
-    openai_client.send_omni_request(request_config)
+    online_client.send_omni_request(request_config)
 ```
 
 *Online serving e2e — **TTS** (representative):*
@@ -244,9 +244,9 @@ def test_text_to_text_001(omni_server, openai_client) -> None:
 @pytest.mark.tts
 @hardware_test(...)
 @pytest.mark.parametrize("omni_server", tts_server_params, indirect=True)
-def test_text_to_audio_001(omni_server, openai_client) -> None:
+def test_text_to_audio_001(omni_server, online_client) -> None:
     request_config = {"model": omni_server.model, "input": "...", "response_format": "wav", ...}
-    openai_client.send_audio_speech_request(request_config)
+    online_client.send_audio_speech_request(request_config)
 ```
 
 *Online serving e2e — **Diffusion** X2I (representative):*
@@ -255,8 +255,8 @@ def test_text_to_audio_001(omni_server, openai_client) -> None:
 @pytest.mark.core_model
 @pytest.mark.diffusion
 @pytest.mark.parametrize("omni_server", _get_default_case(MODEL), indirect=True)
-def test_text_to_image_001(omni_server, openai_client) -> None:
-    openai_client.send_diffusion_request({...})  # chat completions + extra_body
+def test_text_to_image_001(omni_server, online_client) -> None:
+    online_client.send_diffusion_request({...})  # chat completions + extra_body
 ```
 
 *Online serving e2e — **Diffusion** X2V (representative):*
@@ -264,8 +264,8 @@ def test_text_to_image_001(omni_server, openai_client) -> None:
 ```python
 @pytest.mark.core_model
 @pytest.mark.diffusion
-def test_text_to_video_001(omni_server, openai_client) -> None:
-    openai_client.send_video_diffusion_request({"model": ..., "form_data": {...}})  # /v1/videos
+def test_text_to_video_001(omni_server, online_client) -> None:
+    online_client.send_video_diffusion_request({"model": ..., "form_data": {...}})  # /v1/videos
 ```
 
 *Documentation example tests:* follow the **Preferred Test Strategy** in [l4_doc_example_tests.inc.md](../../../docs/contributing/ci/test_examples/l4_doc_example_tests.inc.md): dynamic extraction for offline READMEs; explicit copied client code for online pages until extraction is justified; use the documented **naming**, **output directory** (page folder + case id), and **skipping** rules (e.g. Gradio-only scripts).
@@ -274,6 +274,7 @@ def test_text_to_video_001(omni_server, openai_client) -> None:
 
 **3. Cross-cutting rules**
 
+- New `.py` files need the Omni SPDX header (`Copyright contributors to the vLLM-Omni project`, not `vLLM project`). `tests/` may keep stdlib `re`/`base64`; pickle is still banned unless the file is already on the pickle allowlist. Every collected `tests/**/test_*.py` needs a CI **level** mark and a **hardware** platform mark/helper (`check-mark` runs locally; GHA skips it). Do not write `pytest.mark.H100` / `pytest.mark.L4`; use the helpers so `cards_{n}` is attached. Do not grow `allowed_files` to land a test. Policy: [docs/contributing/README.md](../../../docs/contributing/README.md#linting).
 - Reuse existing fixtures for the chosen scenario; do not mix “online client” assumptions into offline `OmniRunner` tests without a clear reason.
 - Avoid external network dependency in assertions unless the scenario is explicitly “online serving” or doc examples that require a model hub (then align with CI secrets/cache).
 - Keep **one test function = one intent** (one modality combo, one endpoint contract, or one acceleration combo).
@@ -323,21 +324,21 @@ vLLM-Omni e2e tests start a real **OmniServer** (online) or **OmniRunner** (offl
 
 | Scope | Online fixtures | Offline fixtures | When to use |
 |-------|-----------------|------------------|-------------|
-| **Module** (default) | `omni_server` → `openai_client` | `omni_runner` → `omni_runner_handler` | **Default for L2/L3/L4 expansion** — amortize model/server init across `test_*` in the same module. Same `OmniServerParams` / runner config can be reused by multiple tests. |
-| **Function** | `omni_server_function` → `openai_client_function` | `omni_runner_function` → `omni_runner_handler_function` | **Only when required** — each `test_*` must get a **clean** server/runner (no shared engine/GPU state). Typical: `tests/dfx/reliability/`, sleep/wakeup, crash/restart, tests that mutate global server state. |
+| **Module** (default) | `omni_server` → `online_client` | `omni_runner` → `offline_client` | **Default for L2/L3/L4 expansion** — amortize model/server init across `test_*` in the same module. Same `OmniServerParams` / runner config can be reused by multiple tests. |
+| **Function** | `omni_server_function` → `online_client_function` | `omni_runner_function` → `offline_client_function` | **Only when required** — each `test_*` must get a **clean** server/runner (no shared engine/GPU state). Typical: `tests/dfx/reliability/`, sleep/wakeup, crash/restart, tests that mutate global server state. |
 
 **Rules:**
 
 1. **Default to module scope** (`omni_server` / `omni_runner`) unless the scenario **explicitly** needs a fresh instance per test function.
-2. **Indirect parametrize name must match the fixture name**: `@pytest.mark.parametrize("omni_server", ...)` with `omni_server` + `openai_client`; `@pytest.mark.parametrize("omni_server_function", ...)` with `omni_server_function` + `openai_client_function`. Do not mix module fixture with function client (or vice versa).
+2. **Indirect parametrize name must match the fixture name**: `@pytest.mark.parametrize("omni_server", ...)` with `omni_server` + `online_client`; `@pytest.mark.parametrize("omni_server_function", ...)` with `omni_server_function` + `online_client_function`. Do not mix module fixture with function client (or vice versa).
 3. Different `OmniServerParams` per `test_*` (e.g. default vs `--enable-cpu-offload`) is still OK with **module** `omni_server` — pytest parametrizes per test node; only switch to `_function` when isolation between tests matters, not merely because `server_args` differ.
 4. One `test_*` with **many** server configs (expansion matrix) → single function + `@pytest.mark.parametrize("omni_server", [...], indirect=True)` + module `omni_server` (see in-tree `test_qwen_image_expansion.py`).
 
 ```python
 # Default — module-scoped server (L4 expansion)
 @pytest.mark.parametrize("omni_server", [pytest.param(OmniServerParams(model=MODEL), marks=H100)], indirect=True)
-def test_foo_images_generations_default_1024(omni_server, openai_client) -> None:
-    openai_client.send_images_generations_request({...})
+def test_foo_images_generations_default_1024(omni_server, online_client) -> None:
+    online_client.send_images_generations_request({...})
 
 # Function-scoped — reliability / per-test clean state only
 @pytest.mark.parametrize(
@@ -345,9 +346,9 @@ def test_foo_images_generations_default_1024(omni_server, openai_client) -> None
     [pytest.param(OmniServerParams(model=MODEL), marks=H100)],
     indirect=True,
 )
-def test_foo_sleep_wakeup_cycle(omni_server_function, openai_client_function) -> None:
-    openai_client_function.send_omni_sleep_http_request({...})
-    openai_client_function.send_omni_wakeup_http_request({...})
+def test_foo_sleep_wakeup_cycle(omni_server_function, online_client_function) -> None:
+    online_client_function.send_omni_sleep_http_request({...})
+    online_client_function.send_omni_wakeup_http_request({...})
 ```
 
 #### Runtime send helpers — online **and** offline (`tests/helpers/runtime.py`)
@@ -356,19 +357,19 @@ def test_foo_sleep_wakeup_cycle(omni_server_function, openai_client_function) ->
 
 | Principle | Action |
 |-----------|--------|
-| **Reuse first** | Grep `runtime.py` for an existing **`send_*_request`** on `OpenAIClientHandler` (online) or `OmniRunnerHandler` (offline) that matches the endpoint / pipeline shape. |
+| **Reuse first** | Grep `runtime.py` for an existing **`send_*_request`** on `OnlineOmniClient` (online) or `OfflineOmniClient` (offline) that matches the endpoint / pipeline shape. |
 | **Extend when close** | If an existing `send_*_request` almost fits (e.g. missing one optional field), extend it in `runtime.py` — do not fork logic in the test file. |
 | **Add when missing** | No suitable helper → add **`send_<feature>_request`** (high-level: call + general `assert_*`) or **`send_<route>_<verb>_http_request`** (low-level HTTP for negative/dfx) in **`runtime.py` first**, then call it from tests. |
 | **Test module owns payload only** | In `test_*.py`: `MODEL`, deploy path, vendored media, `get_prompt()`, and inline **`request_config` dicts** only. **No** `omni.generate(...)`, raw `requests.post`, OpenAI SDK calls, or `_collect_audio()` / `_process_output()` in e2e tests. |
 
-**Online** (`openai_client` from `omni_server`): `OpenAIClientHandler.send_*_request`.
+**Online** (`online_client` from `omni_server`): `OnlineOmniClient.send_*_request`.
 
-**Offline** (`omni_runner_handler` from `omni_runner`): `OmniRunnerHandler.send_*_request`.
+**Offline** (`offline_client` from `omni_runner`): `OfflineOmniClient.send_*_request`.
 
 | Do | Don't |
 |----|-------|
-| Online: `openai_client.send_omni_request`, `send_diffusion_request`, `send_audio_speech_request`, `send_video_diffusion_request`, `send_images_generations_request`, … | `requests.post(f"{base_url}/v1/...", json=…)` or `client.chat.completions.create(...)` inside a test |
-| Offline: `omni_runner_handler.send_omni_request`, `send_diffusion_request`, `send_audio_speech_request`, `send_single_stage_tts_request`, `send_single_stage_tts_batch_request`, … | `omni_runner.omni.generate(...)` + hand-rolled tensor/WAV extraction in `test_*.py` |
+| Online: `online_client.send_omni_request`, `send_diffusion_request`, `send_audio_speech_request`, `send_video_diffusion_request`, `send_images_generations_request`, … | `requests.post(f"{base_url}/v1/...", json=…)` or `client.chat.completions.create(...)` inside a test |
+| Offline: `offline_client.send_omni_request`, `send_diffusion_request`, `send_audio_speech_request`, `send_single_stage_tts_request`, `send_single_stage_tts_batch_request`, … | `omni_runner.omni.generate(...)` + hand-rolled tensor/WAV extraction in `test_*.py` |
 | Add missing `send_*` to **`runtime.py`** first; bundle general `assert_*` inside the send helper | A one-off `def _post_*` / `def _collect_audio` at the bottom of a test module |
 | Mirror naming/style of neighboring `send_*` (docstring, `request_config` dict, optional `run_level`, `err_code` / `err_message` for negative cases) | Different parameter shapes per test file for the same endpoint |
 
@@ -391,7 +392,7 @@ def test_foo_sleep_wakeup_cycle(omni_server_function, openai_client_function) ->
 
 Changing `request_config` fields (`size`, `n`, `seed`, server flags) is **not** special validation — the general `assert_*` should read those from `request_config` inside `send_*`.
 
-**`send_*` ↔ `assert_*` pairing (online `OpenAIClientHandler`):**
+**`send_*` ↔ `assert_*` pairing (online `OnlineOmniClient`):**
 
 | High-level `send_*` (prefer in L2+ e2e) | Assert already invoked inside `runtime.py` | Low-level HTTP-only sibling |
 |----------------------------------------|--------------------------------------------|-----------------------------|
@@ -404,7 +405,7 @@ Changing `request_config` fields (`size`, `n`, `seed`, server flags) is **not**
 
 **Rule:** Tests call high-level `send_*_request` for the general contract. **Never** call the same bundled `assert_*` again. Call an extra `assert_*` in the test **only** for special, case-specific validation.
 
-**Common `OpenAIClientHandler` entry points** (non-exhaustive — grep `runtime.py` before adding):
+**Common `OnlineOmniClient` entry points** (non-exhaustive — grep `runtime.py` before adding):
 
 | Area | High-level (SDK + assert) | Low-level HTTP (`*_http_request`) |
 |------|---------------------------|-----------------------------------|
@@ -415,7 +416,7 @@ Changing `request_config` fields (`size`, `n`, `seed`, server flags) is **not**
 | TTS (online) | `send_audio_speech_request` | `send_audio_speech_http_request`, `send_audio_generate_http_request`, … |
 | Ops / meta | — | `send_health_http_request`, `send_models_http_request`, `send_omni_sleep_http_request`, … |
 
-**Common `OmniRunnerHandler` entry points** (offline — grep `runtime.py` before adding):
+**Common `OfflineOmniClient` entry points** (offline — grep `runtime.py` before adding):
 
 | Area | High-level (`send_*_request` + assert) | Notes |
 |------|----------------------------------------|-------|
@@ -425,7 +426,7 @@ Changing `request_config` fields (`size`, `n`, `seed`, server flags) is **not**
 | Single-stage TTS (Coqui XTTS, MOSS-TTS-Nano, …) | `send_single_stage_tts_request`, `send_single_stage_tts_batch_request` | `prompt` + `additional_information` + `omni.generate` — **do not** duplicate in tests |
 | New model family | **Add** `send_<family>_request` here first | Then call from `tests/e2e/offline_inference/test_*.py` |
 
-L1 tests under `tests/entrypoints/` may use **FastAPI `TestClient`** or direct handler calls with mocks; they do **not** need `OpenAIClientHandler` / `OmniRunnerHandler`. **L2+ online and offline e2e** must use `runtime.py` `send_*` helpers.
+L1 tests under `tests/entrypoints/` may use **FastAPI `TestClient`** or direct handler calls with mocks; they do **not** need `OnlineOmniClient` / `OfflineOmniClient`. **L2+ online and offline e2e** must use `runtime.py` `send_*` helpers.
 
 #### Invalid parameter validation (`tests/dfx/reliability/invalid_param_test/`)
 
@@ -451,7 +452,7 @@ When the user asks for **invalid parameter validation**, **invalid request bodie
 | **Module markers** | `pytestmark = [pytest.mark.slow, pytest.mark.<omni or tts or diffusion>]` |
 | **Server fixture** | `_PARAMS` / `_QWEN3_TTS_SPEECH`-style list of `pytest.param(OmniServerParams(...), id="...", marks=hardware_marks(...))`; `@pytest.mark.parametrize("omni_server", _PARAMS, indirect=True)` |
 | **Hardware** | `hardware_marks(res={"cuda": "H100"})` for heavy diffusion/omni/video; `hardware_marks(res={"cuda": "L4"})` for smaller TTS models (must match weekly `-m "slow and L4"` step) |
-| **HTTP client** | **Low-level** `openai_client.send_*_http_request({..., "err_code": 400, "err_message": (...)} )` — **not** `send_*_request` (success path) |
+| **HTTP client** | **Low-level** `online_client.send_*_http_request({..., "err_code": 400, "err_message": (...)} )` — **not** `send_*_request` (success path) |
 | **Case shape** | Prefer **one `test_*` per route family** + `@pytest.mark.parametrize("body_spec, err_message", [...])` with stable `id=` per case; or dedicated `test_<route>_malformed_json` when not parametrized |
 | **Body helpers** | `_minimal_<endpoint>_json(omni_server)` / `_minimal_*_form_data()`; merge overrides with `body.update(body_spec)` |
 | **Known gaps** | `pytest.mark.skip(reason="…#3649")` as `_SKIP_ISSUE_3649` when server validation is not yet strict (mirror neighboring cases) |
@@ -467,10 +468,10 @@ When the user asks for **invalid parameter validation**, **invalid request bodie
 | **Invalid parameters Test · H100** | `pytest -s -v tests/dfx/reliability/invalid_param_test/ -m "slow and H100"` | Diffusion / omni / video invalid-param tests with `H100` hardware mark |
 | **Invalid parameters Test · L4** | `pytest -s -v tests/dfx/reliability/invalid_param_test/ -m "slow and L4"` | TTS / lighter models marked `L4` |
 
-- **Trigger:** `build.env("WEEKLY") == "1"` or PR label `weekly-test`.
+- **Trigger:** `build.env("WEEKLY") == "1"` (all weekly jobs) or PR label `weekly-test` (Reliability / Perf, then filtered by `source_file_dependencies`).
 - **Default:** extending an existing `invalid_param_test` script needs **no YAML edit** — the weekly steps already sweep the whole directory.
 - **Edit YAML only** when adding a **new hardware queue**, a **new top-level script** that must run in isolation, or a **model-specific weekly shard** (mirror neighboring reliability steps).
-- Weekly steps **do not** use `source_file_dependencies`.
+- Weekly Reliability / Perf jobs declare `source_file_dependencies`. Filtering applies only on the `weekly-test` label; `WEEKLY=1` on `main` keeps every job in those groups. CPU coverage (`WEEKLY=1` only) and the E2E Tests group (`NON_CRITICAL=1` only) do **not** use this field.
 
 **Example — append to `test_invalid_image_generation.py`:**
 
@@ -484,13 +485,13 @@ When the user asks for **invalid parameter validation**, **invalid request bodie
 @pytest.mark.parametrize("omni_server", _PARAMS, indirect=True)
 def test_images_generations_invalid_requests(
     omni_server: OmniServer,
-    openai_client: OpenAIClientHandler,
+    online_client: OnlineOmniClient,
     body_spec: dict[str, object],
     err_message: str | tuple[str, ...],
 ) -> None:
     body = _minimal_images_gen_json(omni_server)
     body.update(body_spec)
-    openai_client.send_images_generations_http_request(
+    online_client.send_images_generations_http_request(
         {"json": body, "timeout": 300, "err_code": 400, "err_message": err_message}
     )
 ```
@@ -541,18 +542,18 @@ def send_images_generations_request(self, request_config: dict[str, Any], ...) -
 
 # tests/e2e/.../test_foo_expansion.py — one case per test_* (no case_id branching)
 @pytest.mark.parametrize("omni_server", [pytest.param(OmniServerParams(model=MODEL), marks=SINGLE_L4)], indirect=True)
-def test_foo_images_generations_default_1024(omni_server, openai_client) -> None:
-    openai_client.send_images_generations_request({"json": body, "timeout": 300})
+def test_foo_images_generations_default_1024(omni_server, online_client) -> None:
+    online_client.send_images_generations_request({"json": body, "timeout": 300})
 
 @pytest.mark.parametrize("omni_server", [pytest.param(OmniServerParams(model=MODEL), marks=SINGLE_L4)], indirect=True)
-def test_foo_images_generations_sizes_256_512_1024(omni_server, openai_client) -> None:
+def test_foo_images_generations_sizes_256_512_1024(omni_server, online_client) -> None:
     for size in ("256x256", "512x512", "1024x1024"):
-        openai_client.send_images_generations_request({"json": {**body, "size": size}, "timeout": 300})
+        online_client.send_images_generations_request({"json": {**body, "size": size}, "timeout": 300})
 
 # chat diffusion — separate test; send_diffusion_request bundles assert_diffusion_response
 @pytest.mark.parametrize("omni_server", [pytest.param(OmniServerParams(model=MODEL), marks=SINGLE_L4)], indirect=True)
-def test_foo_chat_completions_t2i_fallback(omni_server, openai_client) -> None:
-    openai_client.send_diffusion_request({...})
+def test_foo_chat_completions_t2i_fallback(omni_server, online_client) -> None:
+    online_client.send_diffusion_request({...})
 ```
 
 L1 tests under `tests/entrypoints/` may keep **minimal** asserts next to the handler under test when validating a single branch; still prefer `assertions.py` when the same JSON/media check appears in more than one file or level (e.g. L1 protocol test + L4 e2e).
@@ -643,7 +644,7 @@ If the test is not already collected by an existing pipeline command (for exampl
 | **L2** | `test-ready.yml` with `core_model` + `source_file_dependencies` | merge / nightly pipelines |
 | **Invalid param** | `test-weekly.yml` — **Invalid parameters Test** group (`-m "slow and H100"` / `-m "slow and L4"`). Usually **no YAML edit** when appending cases to existing `invalid_param_test` scripts | `test-ready.yml`, `test-merge.yml`, `test-nightly.yml` |
 
-`test-nightly.yml` shards use **explicit pytest file paths** in `commands` (and PR labels like `diffusion-x2iat-test`); they generally **do not** use `source_file_dependencies`. Only suggest merge/ready YAML when the requested level is L3/L2 (or the user explicitly asks for multi-level CI).
+`test-nightly.yml` shards use **explicit pytest file paths** in `commands` plus `source_file_dependencies`; PR trigger is `nightly-test` (path filter), or `NIGHTLY=1` on `main` (all jobs). Only suggest merge/ready YAML when the requested level is L3/L2 (or the user explicitly asks for multi-level CI).
 
 **`test-nightly.yml` top-level groups** (each uses `full_model` + type marker in pytest `-m`):
 
@@ -652,7 +653,7 @@ If the test is not already collected by an existing pipeline command (for exampl
 | **Omni Model Test** | `omni` | Function / doc / accuracy / **perf** for Qwen-Omni family |
 | **TTS Model Test** | `tts` | `-m "full_model and L4 and tts"` (and H100 variants when enabled); **Perf Test** uses `run_benchmark.py` + `test_tts.json` |
 | **Diffusion X2I(&A&T) Model Test** | `diffusion` | Image/audio/text diffusion; explicit file shards for **Function**; **Perf Test · &lt;Model&gt;** uses `run_diffusion_benchmark.py` + JSON config |
-| **Diffusion X2V Model Test** | `diffusion` | Video-only (Wan, HunyuanVideo, …); PR label `diffusion-x2v-test`; separate **Perf** steps when in-tree |
+| **Diffusion X2V Model Test** | `diffusion` | Video-only (Wan, HunyuanVideo, …); separate **Perf** steps when in-tree |
 
 Guidelines when editing YAML:
 

.claude/skills/vllm-omni-test/references/test-routing.md

@@ -37,7 +37,7 @@ Every model-centric e2e test must declare **exactly one** type marker:
 | **Diffusion X2I(&A&T) Model Test** | x2**i** / x2**a** / x2**t** — image, audio, text (non-video) | Qwen-Image*, BAGEL, FLUX, SD3, Z-Image, LongCat, DreamZero |
 | **Diffusion X2V Model Test** | x2**v** — video only | Wan2.2, HunyuanVideo 1.5, Wan VACE |
 
-PR labels for selective nightly runs: `diffusion-x2iat-test`, `diffusion-x2v-test` (plus `nightly-test` / `NIGHTLY=1`).
+PR label for L4 on a branch: `nightly-test` (jobs then filtered by `source_file_dependencies`). Scheduled `main` uses `NIGHTLY=1` and runs every nightly job. L5 Reliability / Perf use `weekly-test` with the same path filter; CPU coverage and the weekly E2E group are env-only (`WEEKLY=1` / `NON_CRITICAL=1`) and are not source-filtered.
 
 ## Level and Marker Mapping
 
@@ -79,7 +79,7 @@ E2E (L2+) should not rely on mocks unless documenting a rare exception; prefer r
    - `full_model` — L4 nightly (`test-nightly.yml`); some expansion tests still carry both `advanced_model` and `full_model` during migration
 2. **Model type** (pick one): `omni`, `tts`, or `diffusion`
 3. **Cross-cutting** when relevant: `parallel`, `cache`, `example`, `benchmark`
-4. **Hardware**: `cpu`, `cuda`, `rocm`, `npu`, `L4`, `H100`, `distributed_cuda`, …
+4. **Hardware**: `cpu`, `cuda`, `rocm`, `npu`, `L4`, `H100`, `cards_1`, …
 5. Multi-card: `@hardware_test(...)` in `tests/helpers/mark.py`
 
 ## Command Templates
@@ -238,13 +238,13 @@ pytest -s -v dfx/reliability/invalid_param_test/ -m "slow and L4"
 pytest -s -v dfx/reliability/invalid_param_test/test_invalid_image_generation.py::test_images_generations_invalid_requests -m "slow and H100"
 ```
 
-**Trigger:** `WEEKLY=1` or PR label `weekly-test`. Appending cases to existing scripts usually needs **no YAML edit** (directory sweep). No `source_file_dependencies` on weekly steps.
+**Trigger:** `WEEKLY=1` (all jobs) or PR label `weekly-test` (Reliability / Perf, filtered by `source_file_dependencies`). Appending cases to existing scripts usually needs **no YAML edit** (directory sweep).
 
 ### Platform-targeted examples
 
 ```bash
 cd tests
-pytest -s -v -m "core_model and distributed_cuda and L4" --run-level=core_model
+pytest -s -v -m "core_model and L4 and not cards_1" --run-level=core_model
 ```
 
 ### Concrete e2e paths (common in-tree)
@@ -279,7 +279,7 @@ When adding or modifying tests, do not stop at “where the file lives” — al
 4. **API client + assert placement**: **General** validation → implement in `assertions.py`, call **inside** `send_*_request` in `runtime.py`; tests call **`send_*_request` only**. **Special** case validation → `assert_*` in `assertions.py`, called **in the test** after `send_*_request`. Low-level `send_*_http_request` is for negative/dfx tests (`err_code`), not ordinary L2+ success e2e.
 5. **Shared assertions**: logic in **`tests/helpers/assertions.py`**; general checks inside `send_*_request`; special checks only in the test. No `_assert_*` in `test_*.py`.
 6. **One case → one `test_*`**: function name reflects what is validated; no `if case_id == ...` mega-test merging multiple scenarios.
-7. **Fixture scope**: default **`omni_server` + `openai_client`** / **`omni_runner` + `omni_runner_handler`** (module). Use **`omni_server_function` + `openai_client_function`** / **`omni_runner_function` + `omni_runner_handler_function`** only when each `test_*` must spawn a fresh instance. Parametrize name must match fixture (`omni_server` vs `omni_server_function`).
+7. **Fixture scope**: default **`omni_server` + `online_client`** / **`omni_runner` + `offline_client`** (module). Use **`omni_server_function` + `online_client_function`** / **`omni_runner_function` + `offline_client_function`** only when each `test_*` must spawn a fresh instance. Parametrize name must match fixture (`omni_server` vs `omni_server_function`).
 8. **Type marker**: `omni`, `tts`, or `diffusion` on every model e2e module.
 9. **Diffusion L4 Function**: wire `*_expansion.py` into **X2I(&A&T)** or **X2V** **Function Test** in **`test-nightly.yml` only** — do not add `test-merge.yml` unless the user also requested L3.
 10. **Diffusion L4 Perf** (only when requested): add `tests/dfx/perf/tests/test_<slug>_vllm_omni.json` + **Perf Test · &lt;Model&gt;** step (artifact upload); not part of “L4 functional cases” by default.

.github/CODEOWNERS

@@ -79,8 +79,6 @@
 /vllm_omni/model_executor/models/ @tzhouam @gcanlin @Sy0307 @amy-why-3459 @linyueqian
 /vllm_omni/model_extras/ @tzhouam @gcanlin @princepride
 /vllm_omni/plugins/ @gcanlin @tjtanaa @xuechendi
-# Config override within model_executor — docs/design/module/vllm_omni_config.md
-/vllm_omni/model_executor/stage_configs/ @lishunyang12 @alex-jw-brooks
 
 # Execution platforms (hardware backends) — docs/design/module/execution_platforms.md
 /vllm_omni/platforms/ @gcanlin @tjtanaa @xuechendi

.github/ISSUE_TEMPLATE/400-bug-report.yml

@@ -83,7 +83,7 @@ body:
 
       omni = Omni(
           model="Qwen/Qwen-Image",
-          stage_configs_path="/path/to/stage_configs.yaml",
+          deploy_config="/path/to/deploy_config.yaml",
       )
 
       prompts = [{"prompt": "A scenic watercolor painting of a lighthouse at sunset"}]

.github/mergify.yml

@@ -45,7 +45,7 @@ pull_request_rules:
         > # For mypy (substitute "3.10" with the failing version if needed)
         > pre-commit run --hook-stage manual mypy-3.10
         > # For markdownlint
-        > pre-commit run --hook-stage manual markdownlint
+        > pre-commit run markdownlint-cli2
         > ```
         > </details>
 

.github/workflows/pre-commit.yml

@@ -24,9 +24,7 @@ jobs:
           python-version: "3.12"
           cache: 'pip'
       - uses: pre-commit/action@v3.0.1
-        # Don't run the pytest mark check against everything in the CI
-        # for now since a lot of files are still missing them. This check
-        # is used to help contributors remember to add the marks when they
-        # commit.
+        # Don't run historical-debt hooks against the whole tree in CI yet.
+        # They still run locally on changed files.
         env:
-          SKIP: check-test-ci-coverage
+          SKIP: check-mark,markdownlint-cli2,shellcheck,check-spdx-header,mypy-3.10

.gitignore

@@ -46,6 +46,8 @@ htmlcov/
 .cache
 nosetests.xml
 coverage.xml
+coverage-*.xml
+coverage-*.xml.gz
 *.cover
 *.py,cover
 .hypothesis/
@@ -240,9 +242,9 @@ data/
 datasets/
 *.csv
 *.json
+!recipes/cosmos3/negative_prompt.json
 !tests/dfx/perf/tests/*.json
 !apps/ComfyUI-vLLM-Omni/example_workflows/*.json
-!tests/assets/soulxsinger/*.json
 *.jsonl
 *.parquet
 
@@ -252,6 +254,7 @@ outputs/
 results/
 generated/
 output_*/
+/tests/e2e/accuracy/artifacts/
 
 # Configuration overrides
 configs/local.yaml
@@ -275,3 +278,6 @@ tests/.ci_generated/
 
 # Local Lance precision-matching dump scripts (not part of the repo)
 lance_compare/
+
+# Linting binaries downloaded by local pre-commit hooks
+shellcheck*/

.markdownlint.yaml

@@ -0,0 +1,17 @@
+# Keep these rules aligned with vLLM so MkDocs Material syntax is tolerated.
+MD007:
+  indent: 4
+MD013: false
+# GitHub READMEs start with a centered logo, not an H1.
+MD041: false
+MD024:
+  siblings_only: true
+MD025:
+  # Allow front matter title to be different from the first heading in the document.
+  front_matter_title: ""
+MD031:
+  list_items: false
+MD033: false
+MD046: false
+MD052: false
+MD059: false

.pre-commit-config.yaml

@@ -24,7 +24,7 @@ repos:
     - id: ruff-format
 
   - repo: https://github.com/crate-ci/typos
-    rev: typos-dict-v0.13.13
+    rev: v1.49.0
     hooks:
       - id: typos
         # only for staged files
@@ -38,7 +38,19 @@ repos:
       - id: actionlint
         files: ^\.github/workflows/.*\.ya?ml$
 
+  - repo: https://github.com/DavidAnson/markdownlint-cli2
+    rev: v0.23.2
+    hooks:
+      - id: markdownlint-cli2
+        language_version: lts
+        args: [--fix]
+        files: (^docs/|^recipes/|(^|/)README\.md$|^CONTRIBUTING\.md$)
+        # Agent skill docs and CLAUDE.md are not user-facing docs/READMEs.
+        exclude: (^|/)CLAUDE\.md$|^\.claude/|^\.cursor/
 
+  # Keep `local` last so `suggestion` stays the final hook. Autofix repos
+  # (ruff, typos, markdownlint, trailing-whitespace, etc.) must run first;
+  # otherwise their "files were modified" output displaces the SKIP tip.
   - repo: local
     hooks:
     - id: signoff-commit
@@ -54,8 +66,47 @@ repos:
       verbose: true
       stages: [commit-msg]
 
-    - id: check-test-ci-coverage
-      name: Ensure test files have pytestmarks
+    # TODO: Use https://github.com/pre-commit/mirrors-mypy when mypy setup
+    # is less awkward. Local wrapper groups files so follow_imports=skip on
+    # tests does not leak into library code.
+    # mypy-3.10 runs on git commit (changed files). CI SKIPs it because
+    # --all-files is not clean yet. Extra versions are manual-stage only
+    # (not in default `pre-commit run --all-files`):
+    #   pre-commit run --hook-stage manual mypy-3.12
+    - id: mypy-3.10
+      name: Run mypy for Python 3.10
+      entry: python tools/pre_commit/mypy.py "3.10"
+      <<: &mypy_common
+        language: python
+        types_or: [python, pyi]
+        require_serial: true
+        additional_dependencies:
+          - mypy==1.11.1
+          - regex
+          - types-cachetools
+          - types-setuptools
+          - types-PyYAML
+          - types-requests
+          - types-torch
+          - pydantic
+    - id: mypy-3.11
+      name: Run mypy for Python 3.11
+      entry: python tools/pre_commit/mypy.py "3.11"
+      <<: *mypy_common
+      stages: [manual]
+    - id: mypy-3.12
+      name: Run mypy for Python 3.12
+      entry: python tools/pre_commit/mypy.py "3.12"
+      <<: *mypy_common
+      stages: [manual]
+    - id: mypy-3.13
+      name: Run mypy for Python 3.13
+      entry: python tools/pre_commit/mypy.py "3.13"
+      <<: *mypy_common
+      stages: [manual]
+
+    - id: check-mark
+      name: Ensure test files have CI marks (no direct SKU)
       entry: python tools/pre_commit/check_test_marks.py
       language: python
       types: [python]
@@ -66,12 +117,50 @@ repos:
       entry: python tools/pre_commit/check_tts_adapter.py
       language: python
       pass_filenames: false
-      files: ^(vllm_omni/entrypoints/openai/(serving_speech\.py|tts_adapters/__init__\.py)|tools/pre_commit/check_tts_adapter\.py)$
+      files: ^(vllm_omni/entrypoints/openai/serving_speech\.py|tools/pre_commit/check_tts_adapter\.py)$
       # The budget-is-now-slack reminder is printed on a PASSING run, and
       # pre-commit hides output from passing hooks unless the hook is verbose.
       verbose: true
 
-    # Keep `suggestion` last
+    - id: shellcheck
+      name: Lint shell scripts
+      # language: script breaks on Windows (shebang path loses slashes).
+      # Invoke the vLLM-style wrapper with bash so Linux CI and Win/Git Bash work.
+      entry: bash
+      args: [tools/pre_commit/shellcheck.sh]
+      language: system
+      types: [shell]
+
+    - id: check-spdx-header
+      name: Check SPDX headers
+      entry: python tools/pre_commit/check_spdx_header.py
+      language: python
+      types_or: [python, pyi, rust, proto, shell]
+      files: \.(py|pyi|rs|proto|sh)$
+
+    - id: check-forbidden-imports
+      name: Check forbidden imports
+      entry: python tools/pre_commit/check_forbidden_imports.py
+      language: python
+      types: [python]
+      additional_dependencies: [regex]
+
+    - id: check-torch-cuda-call
+      name: Prevent new torch.cuda API calls
+      entry: python tools/pre_commit/check_torch_cuda.py
+      language: python
+      types: [python]
+      additional_dependencies: [regex]
+
+    - id: check-buildkite
+      name: Validate Buildkite Pipelines
+      entry: python tools/pre_commit/check_buildkite.py
+      language: python
+      files: ^\.buildkite/.*\.ya?ml$
+      additional_dependencies: [pyyaml, check-jsonschema==0.38.0]
+
+    # Keep `suggestion` as the final hook. Run after autofix hooks so the
+    # bypass/SKIP tip is not displaced.
     - id: suggestion
       name: Suggestion
       entry: bash -c 'echo "To bypass all the pre-commit hooks, add --no-verify to git commit. To skip a specific hook, prefix the commit command with SKIP=<hook-id>."'
@@ -79,10 +168,3 @@ repos:
       verbose: true
       pass_filenames: false
     # Insert new entries above the `suggestion` entry
-
-    - id: check-pickle-imports
-      name: Prevent new pickle/cloudpickle imports
-      entry: python tools/pre_commit/check_pickle_imports.py
-      language: python
-      types: [python]
-      additional_dependencies: [regex]

.readthedocs.yml

@@ -11,10 +11,15 @@ build:
     post_checkout:
       - git fetch --unshallow || true
     pre_install:
+      # `pre_install` runs before RTD upgrades pip, and the seeded pip 23.1
+      # rejects the `typing_extensions` wheel torch needs (fixed in pip 23.2).
+      - python -m pip install --upgrade pip
       # Pre-install CPU torch so the subsequent `pip install .[docs]` does not
       # pull ~2 GB of CUDA 13 libs (cuBLAS/cuDNN/NCCL/...) that exceed RTD's
-      # 1 GB memory cap and time out the mkdocs build.
-      - python -m pip install --index-url https://download.pytorch.org/whl/cpu torch
+      # 1 GB memory cap and time out the mkdocs build. Wheel-only: `--index-url`
+      # replaces PyPI, so an sdist fallback cannot reach its build deps. The
+      # quotes keep YAML from reading `:all:` as a mapping key.
+      - "python -m pip install --only-binary=:all: --index-url https://download.pytorch.org/whl/cpu torch"
 
 mkdocs:
   configuration: mkdocs.yml

CONTRIBUTING.md

@@ -1,5 +1,13 @@
 # Contributing to vLLM-Omni
 
-You may find information about contributing to vLLM-Omni on [Contributing](https://vllm-omni.readthedocs.io/en/latest/contributing/)
+You may find information about contributing to vLLM-Omni on [Contributing](https://vllm-omni.readthedocs.io/en/latest/contributing/).
+
+AI-assisted contributions must follow the [AI Usage Policy](docs/contributing/README.md#ai-usage-policy), including disclosure, human review, and validation requirements.
+
+Local `pre-commit` gates (SPDX, forbidden imports, `torch.cuda`, shellcheck,
+mypy, test marks, TTS adapter ratchet, markdownlint, Buildkite schema, and
+how to extend allowlists) are documented in the
+[Linting](https://vllm-omni.readthedocs.io/en/latest/contributing/#linting)
+section of that guide.
 
 Before submitting a PR, run the [precheck-pr skill](https://github.com/vllm-project/vllm-omni/blob/main/.claude/skills/precheck-pr/SKILL.md) with the code agent for a self-check against project conventions.

README.md

@@ -12,10 +12,12 @@ Easy, fast, and cheap omni-modality model serving for everyone
 | <a href="https://vllm-omni.readthedocs.io/en/latest/"><b>Documentation</b></a> | <a href="https://deepwiki.com/vllm-project/vllm-omni"><b>DeepWiki</b></a> | <a href="https://discuss.vllm.ai"><b>User Forum</b></a> | <a href="https://slack.vllm.ai"><b>Developer Slack</b></a> | <a href="docs/assets/WeChat.jpg"><b>WeChat</b></a> | <a href="https://arxiv.org/abs/2602.02204"><b>Paper</b></a> | <a href="https://docs.google.com/presentation/d/1aPj0OGl_-ZVoib-Qne5dGDAlrRFB-PdHl6E-EE99g8E/edit?usp=sharing"><b>Slides</b></a> |
 </p>
 
-
 ---
 
 *Latest News* 🔥
+
+- [2026/08] We released [0.28.0](https://github.com/vllm-project/vllm-omni/releases/tag/v0.28.0), featuring production-ready [MiniMax H3](recipes/MiniMaxAI/MiniMax-H3.md) serving on GPU and NPU, a unified AR/DiT paged KV cache runtime, and enhanced realtime full-duplex serving for the [MiniCPM-o series](recipes/OpenBMB/MiniCPM-o-4_5.md).
+- [2026/08] [VeRL-Omni](https://github.com/verl-project/verl-omni) `v0.2.0` is released: faster diffusion RL powered by vLLM-Omni (request-level/step-wise batching with FA3), rebuilt Qwen3-Omni multimodal training (DPO & GSPO), plus LTX-2.3, Qwen-Image-Edit support and more. See the [release notes](https://github.com/verl-project/verl-omni/releases/tag/v0.2.0).
 - [2026/08] We released [0.26.0](https://github.com/vllm-project/vllm-omni/releases/tag/v0.26.0) - aligned with the vLLM 0.26 release line, featuring [MiniMax H3](recipes/MiniMaxAI/MiniMax-H3.md) joint video/audio generation, an experimental full-duplex realtime runtime for [MiniCPM-o 4.5](recipes/OpenBMB/MiniCPM-o-4_5.md), distributed layerwise diffusion offload, and broader model, hardware, streaming, TTS, and quantization support.
 - [2026/07] We released [0.24.0](https://github.com/vllm-project/vllm-omni/releases/tag/v0.24.0) - aligned with the vLLM 0.24 release line, expanding production-ready coverage across TTS, speech, diffusion, image/video generation, and robot-policy serving, with major Omni stage runtime refactoring, diffusion request-level batching, async output materialization, quantization/cache/memory improvements, and broad CUDA/ROCm/XPU/NPU support.
 - [2026/06] Starting with [0.14.0](https://github.com/vllm-project/vllm-omni/releases/tag/v0.14.0), vLLM-Omni publishes a stable release aligned with every even-numbered upstream vLLM minor version. [0.16.0](https://github.com/vllm-project/vllm-omni/releases/tag/v0.16.0), [0.18.0](https://github.com/vllm-project/vllm-omni/releases/tag/v0.18.0), [0.20.0](https://github.com/vllm-project/vllm-omni/releases/tag/v0.20.0), and [0.22.0](https://github.com/vllm-project/vllm-omni/releases/tag/v0.22.0) continued this cadence, expanding omni and world-model support with [NVIDIA Cosmos3](recipes/cosmos3/Cosmos3-Nano.md) and DreamZero, adding models such as MiniCPM-o 4.5, MOSS-TTS, and Lance, and advancing TTS, diffusion, distributed execution, quantization, RL integration through [VeRL-Omni](https://github.com/verl-project/verl-omni), and CUDA/ROCm/MUSA/NPU/XPU coverage.
@@ -51,14 +53,14 @@ vLLM-Omni is flexible and easy to use with:
 - Tensor, pipeline, data and expert parallelism support for distributed inference
 - Streaming outputs
 - OpenAI-compatible API server
-- Full-duplex realtime serving with streaming audio input and output (experimental)
+- [Full-duplex realtime serving](docs/serving/realtime_duplex_api.md) with streaming audio input and output
 
 vLLM-Omni seamlessly supports most popular open-source models on HuggingFace, including:
 
 - **Omni-modality models** (e.g. Qwen3-Omni, MiniCPM-o 4.5, Cosmos3, HunyuanImage, BAGEL)
-- **TTS models** (e.g. Qwen3-TTS, VoxCPM2, Ming-Omni-TTS, CosyVoice3)
-- **Diffusion models** — image, video, and audio generation (e.g. MiniMax H3, Qwen-Image, Wan2.2, FLUX)
-- **Robot-policy and action models** (e.g. GR00T-N1.7, DreamZero-DROID, InternVLA-A1, Cosmos3 action policy)
+- **TTS models** (e.g. Qwen3-TTS, IndexTTS 2.5, dots.tts, CosyVoice3)
+- **Diffusion models** — image, video, and audio generation (e.g. MiniMax H3, LTX-2.5, SANA-Video, Wan2.2)
+- **Robot-policy and action models** (e.g. π0, GR00T-N1.7, DreamZero-DROID, InternVLA-A1)
 
 ## Getting Started
 
@@ -66,6 +68,7 @@ Visit our [documentation](https://vllm-omni.readthedocs.io/en/latest/) to learn
 
 - [Installation](https://vllm-omni.readthedocs.io/en/latest/getting_started/installation/)
 - [Quickstart](https://vllm-omni.readthedocs.io/en/latest/getting_started/quickstart/)
+- [Nightly CUDA Docker images](https://hub.docker.com/r/vllm/vllm-omni/tags?name=nightly) built from the latest `main`
 - [List of Supported Models](https://vllm-omni.readthedocs.io/en/latest/models/supported_models/)
 - [Deployment Recipes](https://recipes.vllm.ai) for vLLM-Omni model serving
 
@@ -88,11 +91,18 @@ If you use vLLM-Omni for your research, please cite our [paper](https://arxiv.or
 ```
 
 ## Join the Community
+
 Feel free to ask questions, provide feedbacks and discuss with fellow users of vLLM-Omni in `#sig-omni` slack channel at [slack.vllm.ai](https://slack.vllm.ai) or vLLM user forum at [discuss.vllm.ai](https://discuss.vllm.ai).
 
 ## Star History
 
-[![Star History Chart](https://api.star-history.com/svg?repos=vllm-project/vllm-omni&type=date&legend=top-left)](https://www.star-history.com/#vllm-project/vllm-omni&type=date&legend=top-left)
+<a href="https://www.star-history.com/?repos=vllm-project%2Fvllm-omni&type=date&legend=top-left">
+  <picture>
+    <source media="(prefers-color-scheme: dark)" srcset="https://api.star-history.com/chart?repos=vllm-project/vllm-omni&type=date&theme=dark&legend=top-left&sealed_token=ExgLDZJoQEg27Zfhhut2LqN0GYO6Fw2PWLwPE6JYBUp2BgM3hmsYlwaIVopnUEfbRXidQ4nisumrTdKYydiKhy1SZXipw47qY2_tiUDhCpsPXeXtPuEVKVzBwKs3pw0tiHsJgtSfwXx5yjHXck0Y2SblzFWeJYCkTe1WLGTbUAOIETjXXQJjyCGZvKz5" />
+    <source media="(prefers-color-scheme: light)" srcset="https://api.star-history.com/chart?repos=vllm-project/vllm-omni&type=date&legend=top-left&sealed_token=ExgLDZJoQEg27Zfhhut2LqN0GYO6Fw2PWLwPE6JYBUp2BgM3hmsYlwaIVopnUEfbRXidQ4nisumrTdKYydiKhy1SZXipw47qY2_tiUDhCpsPXeXtPuEVKVzBwKs3pw0tiHsJgtSfwXx5yjHXck0Y2SblzFWeJYCkTe1WLGTbUAOIETjXXQJjyCGZvKz5" />
+    <img alt="Star History Chart" src="https://api.star-history.com/chart?repos=vllm-project/vllm-omni&type=date&legend=top-left&sealed_token=ExgLDZJoQEg27Zfhhut2LqN0GYO6Fw2PWLwPE6JYBUp2BgM3hmsYlwaIVopnUEfbRXidQ4nisumrTdKYydiKhy1SZXipw47qY2_tiUDhCpsPXeXtPuEVKVzBwKs3pw0tiHsJgtSfwXx5yjHXck0Y2SblzFWeJYCkTe1WLGTbUAOIETjXXQJjyCGZvKz5" />
+  </picture>
+</a>
 
 ## License
 

apps/ComfyUI-vLLM-Omni/README.md

@@ -13,7 +13,6 @@ It can send model inference requests to either a locally running vLLM-Omni servi
 > [!TIP]
 > If you run both ComfyUI and vLLM-Omni on the same device, you can create separate virtual environments and use different Python versions for them.
 
-
 ## Installation
 
 Copy this folder to the `custom_nodes` subfolder of your ComfyUI installation. Your directory should look like `ComfyUI/custom_nodes/ComfyUI-vLLM-Omni`.
@@ -21,9 +20,10 @@ Copy this folder to the `custom_nodes` subfolder of your ComfyUI installation. Y
 If you are running ComfyUI during copying, you should restart ComfyUI to load this extension.
 
 > [!TIP]
-> You can use utility websites such as https://download-directory.github.io/ to download a subdirectory of a repo. Also checkout community discussions (e.g., https://stackoverflow.com/questions/7106012/download-a-single-folder-or-directory-from-a-github-repository) for more info.
+> You can use utility websites such as <https://download-directory.github.io/> to download a subdirectory of a repo. Also checkout community discussions (e.g., <https://stackoverflow.com/questions/7106012/download-a-single-folder-or-directory-from-a-github-repository>) for more info.
 
 On the device and virtual environment you run ComfyUI, launch ComfyUI with
+
 ```bash
 cd ComfyUI
 
@@ -35,6 +35,7 @@ python main.py --cpu
 ```
 
 On the device and virtual environment you run vLLM-Omni, start a model service with
+
 ```bash
 vllm serve The_Model_ID_to_Serve --omni --port 8000
 ```
@@ -48,14 +49,30 @@ This extension offers the following nodes based on the output modalities (at **C
 
 - **Generate Image** for text-to-image and image-to-image tasks
 - **Generate Video** for text-to-video, first-frame/image-to-video, and reference-conditioned video
+- **FastH3 Deployment** for routing text-to-video requests to a MiniMax-H3 server with FastH3 fused at startup
 - **Multimodality Understanding** for multimodality-to-text and multimodality-to-audio tasks
 - **TTS** and **TTS Voice Clone** for TTS tasks
+- **Generate Music** for lyrics-and-description-to-music generation
 
 This extension also offers example workflows (at **ComfyUI sidebar -> Templates -> vLLM-Omni**)
 
 > [!NOTE]
 > The node UI and feature designs are intended to match vLLM-Omni online serving interfaces. It cannot offer more than what the interfaces support.
 
+Every node carries the vLLM-Omni mark in its title bar and is tinted by what it outputs, so a graph is readable at a glance:
+
+| Colour | Nodes | What they produce |
+| --- | --- | --- |
+| Blue | Generate Image, Generate Video, Multimodality Understanding, TTS, TTS Voice Clone | A generated image, video, audio, or text. These are the only nodes that reach a server. |
+| Amber | AR / Diffusion / Multi-Stage Sampling Params | Sampling parameters that apply to any model |
+| Purple | Qwen TTS Params, Wan Video Params, MiniMax-H3 Video Params | Parameters that only one model family accepts |
+| Red | LoRA, FastH3 Deployment | Which weights the server is expected to have loaded |
+| Teal | Video References | Reference media |
+
+Recolouring a node by hand (right click -> Colors) overrides its tint, and the choice is kept.
+
+**Generate Video** takes a clip length in seconds (`duration`), not a frame count. Frames stay the wire unit and are derived with the node's `fps`, so the length is always measured against the rate that is actually served; models that accept only certain frame counts still round to their own lattice server-side. Graphs saved before this widget existed stored `num_frames` in its place and are converted on load, using the fps recorded alongside it -- the browser console names every node it rewrites.
+
 To build a simple workflow yourself,
 
 - Drag a generation node onto the canvas.
@@ -123,6 +140,127 @@ You can configure per-stage sampling parameters for multi-stage models.
 >
 > Do not use `frame` and `references` together. Task routing is automatic from which inputs you connect.
 
+For MiniMax-H3 Ref2VA, **Video References** accepts up to 9 images (`image_1`–`image_9`),
+3 videos (`video_1`–`video_3`), and 3 audio clips (`audio_1`–`audio_3`), with at most
+12 connected inputs in total. Any mixture containing at least one image or video
+is supported; empty and audio-only references are rejected. For example, you can
+combine 6 images, 3 videos, and 3 audio clips in one request.
+
+Within each media type, references follow slot-number order, skipping unconnected
+slots. For example, connecting `image_2` and `image_9` sends `image_2` as the first
+image and `image_9` as the second. Each image slot uses the first image in its batch.
+Existing connections to `image_1`, `image_2`, `audio_1`, `audio_2`, `video_1`, and
+`video_2` remain valid in saved workflows.
+
+#### MiniMax-H3 Reference to Video
+
+Load [MiniMax-H3 Reference to Video](example_workflows/vLLM-Omni%20MiniMax-H3%20Reference%20to%20Video.json)
+from **Templates → ComfyUI-vLLM-Omni**, or drag the JSON onto the canvas.
+It adapts the [official ComfyUI R2V workflow](https://github.com/Comfy-Org/workflow_templates/blob/main/templates/video_minimax_h3_r2v.json)
+to remote vLLM-Omni execution. Model loading, sampling, and VAE decoding run on
+the server; ComfyUI loads references and saves the returned video.
+
+Use an extension version containing the audio-preserving output fix from
+[#7456](https://github.com/vllm-project/vllm-omni/pull/7456). Older versions can
+drop the server's audio while decoding the response.
+
+1. Start a Ref2VA-capable service using the [MiniMax-H3 recipe](../../recipes/MiniMaxAI/MiniMax-H3.md).
+   Set **Generate Video** to its `/v1` URL and served model name.
+2. Select your own image in **Load Image**. The template's filenames are placeholders
+   relative to ComfyUI's input directory; no sample assets or model weights are bundled.
+3. Connect **Load Video** and **Load Audio** when needed. Duplicate the loaders to
+   fill more reference slots, or disconnect the image when using video-only input.
+   Keep the references connected to **Generate Video** and leave `frame` disconnected.
+4. Match prompt tags such as `<Picture 1>`, `<Video 1>`, and `<Audio 1>` to the
+   connected references, counting each media type separately and skipping empty slots.
+5. Run the workflow. **Save Video** writes an MP4 under `output/video/` and preserves
+   the generated audio when the audio-output prerequisite is installed.
+
+The default is **1344×768, 24 FPS, 124 frames** (about 5.17 seconds), 50 sampling
+points, video flow shift 12, audio flow shift 3, and seed 42. The duration widget
+is set to 5.167 seconds, which converts to 124 frames at 24 FPS. Other H3 canvas presets
+are 1024×768, 768×768, 768×1024, and 768×1344. Keep frame counts at `17k+5` within
+the 4–15 second output range; examples are 107, 124, 209, and 345 frames.
+Reference video/audio clips must each be 2–15 seconds, with at most 15 seconds of
+reference video and at most 15 seconds of standalone reference audio per request.
+
+For **Ref2VA Turbo**, use the same template. Configure a Ref2VA-only server with
+`--lora-backend peft --lora-path "$TURBO_LORA" --task-type ref2va`, following the
+recipe's hardware setup. For `minimax_h3_ref2v_turbo_4step_v0.1_bf16.safetensors`,
+set **LoRA** `local_path` to that file's path on the server and connect it to
+**Generate Video**. Use `num_inference_steps=5`, `flow_shift=12`,
+`audio_flow_shift=3`, and LoRA `scale=1`. The API counts sigma points, so five
+points produce four denoiser evaluations. Use the Diffusers artifact, not its
+`_comfyui_` export. To return to base mode, disconnect LoRA and restore 50 steps.
+The 8-step v1.0 Ref2VA adapter instead requires nine sampling points and video
+flow shift 6. FastH3 is a separate startup-fused T2VA deployment.
+
+Validate the template's wiring and defaults locally with:
+
+```bash
+python -m pytest tests/e2e/features/comfyui/test_h3_reference_workflow.py -q
+```
+
+For a real-model check, run the workflow against H3 and inspect the saved file:
+
+```bash
+ffprobe -v error -show_entries stream=codec_type,width,height,r_frame_rate,sample_rate,channels \
+  -show_entries format=duration -of json "$OUTPUT_MP4"
+```
+
+Check for 24 FPS video, stereo audio, the requested canvas, and aligned duration;
+also play the result to assess reference conditioning and audio/video synchronization.
+Record the server commit, model/adapter, input assets, prompt, settings, and output
+alongside the result. Local schema or mocked-server checks do not establish H3
+generation quality or replace this real-model validation.
+
+### MiniMax H3 Text to Video
+
+Import [the H3 text-to-video template](example_workflows/MiniMax_H3_Text_to_Video.json)
+for remote video generation with audio. It includes connected base presets and
+optional Turbo sampling, H3 parameters, and Remote LoRA nodes. See the
+[workflow guide](docs/minimax-h3-t2v.md) for server setup, artifact-specific Turbo
+settings, and saved-video/audio validation. Set Remote LoRA’s `local_path` to
+the downloaded Turbo artifact on your server before enabling Turbo.
+
+#### H3 video upscale (WF-07)
+
+The **vLLM-Omni MiniMax H3 Video Upscale** template generates video remotely, upscales it with SeedVR2, and saves the original and upscaled videos with the generated audio and FPS. See [workflow setup](docs/wf07-h3-upscale.md).
+
+#### FastH3 text-to-video
+
+FastH3 is fused into MiniMax-H3 when the vLLM-Omni server starts; it is not a request-switchable LoRA. Download one adapter variant and start a non-offloaded FL2VA server. For the Dense / Data-Free profile:
+
+```bash
+export FASTH3_DIR=/path/to/fasth3
+hf download FastVideo/FastVideo-FastH3-4-step-Preview-v1-LoRA \
+  dense-datafree/adapter_model.safetensors --local-dir "${FASTH3_DIR}"
+
+vllm serve /path/to/MiniMax-H3 \
+  --omni \
+  --task-type fl2va \
+  --lora-path "${FASTH3_DIR}/dense-datafree/adapter_model.safetensors" \
+  --port 8000
+```
+
+For the VSA / Data-Free profile, download `vsa-datafree/adapter_model.safetensors`, use that file as `--lora-path`, and add:
+
+```bash
+--diffusion-attention-backend FASTVIDEO_VSA \
+--fastvideo-vsa-topk 64
+```
+
+The VSA profile also requires a compatible `fastvideo-kernel` installation. See the [MiniMax-H3 FastH3 recipe](../../recipes/MiniMaxAI/MiniMax-H3.md#fasth3-adapter) for the full serving contract, supported parallel layouts, and kernel requirements.
+
+Open the **vLLM-Omni FastH3 Text to Video** template, then:
+
+- Set the server URL and served model name on **FastH3 Deployment**.
+- Connect its output to **Generate Video → fast_h3**. When connected, the deployment node's URL and model take precedence over the corresponding Generate Video widgets.
+- Keep `frame`, `references`, **LoRA**, and **MiniMax-H3 Video Params** disconnected. FastH3 Preview v1 supports T2VA only, is already fused, and owns both flow shifts.
+- A connected **Diffusion Sampling Params** node may set seed and other ordinary sampling options. The integration always enforces four denoising steps and 24 FPS for FastH3.
+
+The node records which server the workflow targets; it does not start one, nor switch adapters or attention backends on a running server.
+
 ### TTS (e.g., Qwen TTS series)
 
 (Also available at **ComfyUI sidebar->Template->vLLM-Omni->vLLM-Omni TTS**)
@@ -137,6 +275,17 @@ You can configure per-stage sampling parameters for multi-stage models.
 > [!TIP]
 > There is a dedicated node for VoiceClone tasks with reference audio input. Other simple text-to-speech tasks should use the regular TTS node.
 
+### Music generation (e.g., MiniMax Music 3)
+
+(Also available at **ComfyUI sidebar->Template->vLLM-Omni->vLLM-Omni Music Generation**)
+
+<p align="center">
+  <picture>
+    <source media="(prefers-color-scheme: dark)" src="https://raw.githubusercontent.com/vllm-project/vllm-omni/refs/heads/main/apps/ComfyUI-vLLM-Omni/docs/images/comfyui-music-generation.jpg">
+    <img alt="vLLM-Omni Music Generation" src="https://raw.githubusercontent.com/vllm-project/vllm-omni/refs/heads/main/apps/ComfyUI-vLLM-Omni/docs/images/comfyui-music-generation.jpg" width=55%>
+  </picture>
+</p>
+
 ### Chaining multiple model services
 
 (Also available at **ComfyUI sidebar->Template->vLLM-Omni->vLLM-Omni Chaining Services**)
@@ -152,6 +301,8 @@ You can configure per-stage sampling parameters for multi-stage models.
 
 Follow the [development convention and rules of vLLM-Omni](https://docs.vllm.ai/projects/vllm-omni/en/latest/contributing/).
 
+Node tints and the title-bar mark are applied in `web/main.js`, keyed off each node's declared output types rather than a list of node names. A new node that returns an existing type is themed with no front-end change; a new output type needs one entry in `FAMILY_BY_OUTPUT` there.
+
 ## Limitation and Non-Goals
 
 - Single server mode only. No automatic load balancing or failover.
@@ -178,11 +329,11 @@ Whenever you find an issue or problem, please
 
 Features
 
-- https://github.com/dougbtv/comfyui-vllm-omni/ The official reference implementation for ComfyUI integration with vLLM-Omni's DALL-E compatible image generation API.
-- https://github.com/Comfy-Org/ComfyUI/tree/master/comfy_extras ComfyUI's built-in node implementations.
+- <https://github.com/dougbtv/comfyui-vllm-omni/> The official reference implementation for ComfyUI integration with vLLM-Omni's DALL-E compatible image generation API.
+- <https://github.com/Comfy-Org/ComfyUI/tree/master/comfy_extras> ComfyUI's built-in node implementations.
 
 UI/UX design references
 
-- https://github.com/sgl-project/sglang/pull/15271 SGLang Diffusion's official ComfyUI integration for image and video generation.
-- https://github.com/SXQBW/ComfyUI-Qwen-Omni A third party ComfyUI integration for Qwen Omni series.
-- https://github.com/flybirdxx/ComfyUI-Qwen-TTS https://github.com/DarioFT/ComfyUI-Qwen3-TTS Third  party ComfyUI integrations for Qwen TTS series.
+- <https://github.com/sgl-project/sglang/pull/15271> SGLang Diffusion's official ComfyUI integration for image and video generation.
+- <https://github.com/SXQBW/ComfyUI-Qwen-Omni> A third party ComfyUI integration for Qwen Omni series.
+- <https://github.com/flybirdxx/ComfyUI-Qwen-TTS> <https://github.com/DarioFT/ComfyUI-Qwen3-TTS> Third  party ComfyUI integrations for Qwen TTS series.

apps/ComfyUI-vLLM-Omni/__init__.py

@@ -1,4 +1,8 @@
-"""Top-level package for comfyui_vllm_omni."""  # noqa: N999  # This is not a python library intended to be imported
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
+# ruff: noqa: N999
+
+"""Top-level package for comfyui_vllm_omni."""  # This is not a Python library intended to be imported.
 
 __all__ = [
     "NODE_CLASS_MAPPINGS",
@@ -13,7 +17,9 @@ __version__ = "0.0.1"
 from .comfyui_vllm_omni.nodes import (
     VLLMOmniARSampling,
     VLLMOmniDiffusionSampling,
+    VLLMOmniFastH3Deployment,
     VLLMOmniGenerateImage,
+    VLLMOmniGenerateMusic,
     VLLMOmniGenerateVideo,
     VLLMOmniMiniMaxH3Params,
     VLLMOmniQwenTTSParams,
@@ -33,6 +39,7 @@ NODE_CLASS_MAPPINGS = {
     "VLLMOmniGenerateVideo": VLLMOmniGenerateVideo,
     "VLLMOmniUnderstanding": VLLMOmniUnderstanding,
     "VLLMOmniTTS": VLLMOmniTTS,
+    "VLLMOmniGenerateMusic": VLLMOmniGenerateMusic,
     "VLLMOmniVoiceClone": VLLMOmniVoiceClone,
     "VLLMOmniVideoReferences": VLLMOmniVideoReferences,
     # === Params ===
@@ -40,6 +47,7 @@ NODE_CLASS_MAPPINGS = {
     "VLLMOmniDiffusionSampling": VLLMOmniDiffusionSampling,
     "VLLMOmniSamplingParamsList": VLLMOmniSamplingParamsList,
     "VLLMOmniRemoteLoRA": VLLMOmniRemoteLoRA,
+    "VLLMOmniFastH3Deployment": VLLMOmniFastH3Deployment,
     "VLLMOmniQwenTTSParams": VLLMOmniQwenTTSParams,
     "VLLMOmniWanParams": VLLMOmniWanParams,
     "VLLMOmniMiniMaxH3Params": VLLMOmniMiniMaxH3Params,
@@ -52,6 +60,7 @@ NODE_DISPLAY_NAME_MAPPINGS = {
     "VLLMOmniGenerateVideo": "Generate Video",
     "VLLMOmniUnderstanding": "Multimodality Understanding",
     "VLLMOmniTTS": "TTS (Text to Speech)",
+    "VLLMOmniGenerateMusic": "Generate Music",
     "VLLMOmniVoiceClone": "TTS Voice Cloning",
     "VLLMOmniVideoReferences": "Video References",
     # === Params ===
@@ -59,6 +68,7 @@ NODE_DISPLAY_NAME_MAPPINGS = {
     "VLLMOmniDiffusionSampling": "Diffusion Sampling Params",
     "VLLMOmniSamplingParamsList": "Multi-Stage Sampling Params List",
     "VLLMOmniRemoteLoRA": "LoRA",
+    "VLLMOmniFastH3Deployment": "FastH3 Deployment",
     "VLLMOmniQwenTTSParams": "Qwen TTS Params",
     "VLLMOmniWanParams": "Wan Video Params",
     "VLLMOmniMiniMaxH3Params": "MiniMax-H3 Video Params",

apps/ComfyUI-vLLM-Omni/comfyui_vllm_omni/nodes.py

@@ -1,3 +1,6 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
+
 from typing import Literal
 
 import torch
@@ -7,9 +10,13 @@ from .utils.api_client import VLLMOmniClient
 from .utils.logger import get_logger
 from .utils.models import lookup_model_spec
 from .utils.types import (
+    MAX_REFERENCE_AUDIOS,
+    MAX_REFERENCE_IMAGES,
+    MAX_REFERENCE_VIDEOS,
     AudioFormat,
     AutoregressionSamplingParams,
     DiffusionSamplingParams,
+    FastH3Deployment,
     MiniMaxH3ModelSpecificParams,
     QwenTTSModelSpecificParams,
     VideoReferences,
@@ -22,6 +29,23 @@ from .utils.validators import (
 
 logger = get_logger(__name__)
 
+FASTH3_INFERENCE_STEPS = 4
+FASTH3_FPS = 24
+# A FastH3 deployment may expose H3 under any --served-model-name, so the payload
+# spec has to be named outright rather than recovered from that alias.
+FASTH3_SPEC_MODEL = "MiniMax-H3"
+
+
+def _resolve_fast_h3_deployment(deployment: dict) -> tuple[str, str]:
+    if not isinstance(deployment, dict):
+        raise ValueError("FastH3 deployment must be provided by a FastH3 Deployment node.")
+
+    url = str(deployment.get("url") or "").strip().rstrip("/")
+    model = str(deployment.get("model") or "").strip()
+    if not url or not model:
+        raise ValueError("FastH3 deployment requires both URL and model.")
+    return url, model
+
 
 class _VLLMOmniGenerateBase:
     """Base class for vLLM-Omni generation nodes with shared functionality."""
@@ -162,7 +186,20 @@ class VLLMOmniGenerateVideo(_VLLMOmniGenerateBase):
                 "width": ("INT", {"default": 832, "min": 1}),
                 "height": ("INT", {"default": 480, "min": 1}),
                 "fps": ("INT", {"default": 16, "min": 1}),
-                "num_frames": ("INT", {"default": 41, "min": 1}),
+                "duration": (
+                    "FLOAT",
+                    {
+                        "default": 4.0,
+                        "min": 0.1,
+                        "step": 0.1,
+                        "round": 0.001,
+                        "tooltip": (
+                            "Clip length in seconds, converted to frames with the fps above. "
+                            "Models that only accept certain frame counts (e.g. MiniMax-H3) round to "
+                            "their own lattice, so the served clip can be slightly longer than requested."
+                        ),
+                    },
+                ),
             },
             "optional": {
                 "frame": ("IMAGE",),
@@ -170,6 +207,7 @@ class VLLMOmniGenerateVideo(_VLLMOmniGenerateBase):
                 "sampling_params": ("SAMPLING_PARAMS",),
                 "lora": ("REMOTE_LORA",),
                 "model_params": ("VIDEO_PARAMS",),
+                "fast_h3": ("FASTH3_DEPLOYMENT",),
             },
         }
 
@@ -194,20 +232,66 @@ class VLLMOmniGenerateVideo(_VLLMOmniGenerateBase):
         width: int,
         height: int,
         fps: int,
-        num_frames: int,
+        duration: float,
         negative_prompt: str | None = None,
         frame: torch.Tensor | None = None,
         references: dict | None = None,
         sampling_params: dict | list[dict] | None = None,
         model_params: dict | None = None,
         lora: dict | None = None,
+        fast_h3: dict | None = None,
         **kwargs,
     ):
         if kwargs:
             logger.info("Uncaught kwargs: %s", kwargs)
         logger.debug("Got sampling params: %s", sampling_params)
         logger.debug("Got model params: %s", model_params)
-        validate_model_and_sampling_params_types(model, sampling_params)
+
+        # Which spec builds the payload. Only a FastH3 deployment separates this
+        # from the served name; every other path keeps them equal.
+        spec_model = model
+
+        if fast_h3 is not None:
+            if frame is not None or references is not None:
+                raise ValueError("FastH3 Preview supports T2VA only; disconnect frame and references inputs.")
+            if lora is not None:
+                raise ValueError(
+                    "FastH3 is already fused into the selected server; disconnect the request-level LoRA input."
+                )
+
+            url, model = _resolve_fast_h3_deployment(fast_h3)
+            logger.info("Using FastH3 deployment at %s", url)
+            fps = FASTH3_FPS
+            # The served name is whatever the operator passed to --served-model-name.
+            # Left alone, lookup_model_spec would miss H3 for an alias such as
+            # "fasth3", drop the params builder, and send a t2va request carrying no
+            # aspect_ratio -- which the server refuses.
+            spec_model = FASTH3_SPEC_MODEL
+
+            if sampling_params is None:
+                sampling_params = DiffusionSamplingParams()
+            elif isinstance(sampling_params, list):
+                if len(sampling_params) != 1:
+                    raise ValueError("FastH3 expects a single diffusion sampling params group.")
+                sampling_params = sampling_params[0].__class__(sampling_params[0])
+            else:
+                sampling_params = sampling_params.__class__(sampling_params)
+            sampling_params["num_inference_steps"] = FASTH3_INFERENCE_STEPS
+
+            # FastH3 owns both modality shifts. Sending the ordinary H3 values
+            # from a connected H3 Params node would turn a deployment choice
+            # into a request-level override, which the server intentionally
+            # rejects when it differs from the fused adapter contract.
+            if model_params is not None:
+                model_params = model_params.__class__(model_params)
+                model_params.pop("flow_shift", None)
+                model_params.pop("audio_flow_shift", None)
+
+        # Frames stay the wire unit. Convert after the FastH3 branch above, so the
+        # duration is measured against the fps the server will actually apply.
+        num_frames = max(1, round(duration * fps))
+
+        validate_model_and_sampling_params_types(spec_model, sampling_params)
 
         # Currently, all video generation models are single-stage diffusion models
         if isinstance(sampling_params, list):
@@ -227,6 +311,7 @@ class VLLMOmniGenerateVideo(_VLLMOmniGenerateBase):
         client = VLLMOmniClient(url)
         output = await client.generate_video(
             model=model,
+            spec_model=spec_model,
             prompt=prompt,
             frame=frame,  # frame present => fl2va / Wan I2V
             references=references,
@@ -418,6 +503,66 @@ class VLLMOmniTTS(_VLLMOmniGenerateBase):
         return (audio,)
 
 
+class VLLMOmniGenerateMusic(_VLLMOmniGenerateBase):
+    """Generate a song from lyrics and a musical description with MiniMax Music 3."""
+
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {
+                "url": ("STRING", {"default": "http://localhost:8000/v1"}),
+                "model": ("STRING", {"default": "MiniMaxAI/MiniMax-Music3"}),
+                "instructions": (
+                    "STRING",
+                    {
+                        "multiline": True,
+                        "display_name": "caption",
+                        "tooltip": "Music caption: describe genre, instruments, tempo and mood.",
+                    },
+                ),
+                "lyrics": ("STRING", {"multiline": True}),
+                "max_duration_seconds": (
+                    "FLOAT",
+                    {
+                        "default": 300.0,
+                        "min": 1,
+                        "max": 360,
+                        "step": 0.01,
+                        "tooltip": "Upper limit; rounded down to whole 25 Hz audio frames. May end earlier.",
+                    },
+                ),
+                "seed": ("INT", {"default": 0, "min": 0, "max": 2**53 - 1, "control_after_generate": True}),
+                "response_format": (["wav", "mp3", "flac", "opus"],),
+            },
+        }
+
+    RETURN_TYPES = ("AUDIO",)
+    RETURN_NAMES = ("audio",)
+    FUNCTION = "generate"
+
+    async def generate(
+        self,
+        url: str,
+        model: str,
+        lyrics: str,
+        instructions: str,
+        response_format: AudioFormat,
+        max_duration_seconds: float,
+        seed: int = 0,
+    ) -> tuple[AudioInput]:
+        audio = await VLLMOmniClient(url.rstrip("/")).generate_speech(
+            model=model,
+            input=lyrics,
+            instructions=instructions,
+            voice="default",
+            speed=1.0,
+            response_format=response_format,
+            max_new_tokens=int(max_duration_seconds * 25),
+            seed=seed,
+        )
+        return (audio,)
+
+
 class VLLMOmniVoiceClone(_VLLMOmniGenerateBase):
     @classmethod
     def INPUT_TYPES(cls):
@@ -695,6 +840,52 @@ class VLLMOmniRemoteLoRA:
         return (lora,)
 
 
+class VLLMOmniFastH3Deployment:
+    """Select a vLLM-Omni service that fused FastH3 at startup."""
+
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {
+                "url": (
+                    "STRING",
+                    {
+                        "default": "http://localhost:8000/v1",
+                        "tooltip": "URL of a vLLM-Omni server started with a FastH3 --lora-path.",
+                    },
+                ),
+                "model": (
+                    "STRING",
+                    {
+                        "default": "MiniMaxAI/MiniMax-H3",
+                        "tooltip": "Model name exposed by the FastH3 deployment.",
+                    },
+                ),
+            }
+        }
+
+    RETURN_TYPES = ("FASTH3_DEPLOYMENT",)
+    RETURN_NAMES = ("deployment",)
+    FUNCTION = "get_deployment"
+    CATEGORY = "vLLM-Omni"
+    DESCRIPTION = (
+        "Selects a server with FastH3 fused at startup. The connected Generate Video node uses T2VA, "
+        "four inference steps, and 24 FPS without sending a request-level LoRA."
+    )
+
+    @classmethod
+    def VALIDATE_INPUTS(cls, url, model) -> str | Literal[True]:
+        try:
+            _resolve_fast_h3_deployment({"url": url, "model": model})
+        except ValueError as exc:
+            return str(exc)
+        return True
+
+    def get_deployment(self, url: str, model: str):
+        url, model = _resolve_fast_h3_deployment({"url": url, "model": model})
+        return (FastH3Deployment({"url": url, "model": model}),)
+
+
 class VLLMOmniQwenTTSParams:
     @classmethod
     def INPUT_TYPES(cls):
@@ -790,6 +981,10 @@ class VLLMOmniVideoReferences:
                 "audio_2": ("AUDIO",),
                 "video_1": ("VIDEO",),
                 "video_2": ("VIDEO",),
+                # Append ports to preserve connections in saved workflows.
+                **{f"image_{i}": ("IMAGE",) for i in range(3, MAX_REFERENCE_IMAGES + 1)},
+                **{f"audio_{i}": ("AUDIO",) for i in range(3, MAX_REFERENCE_AUDIOS + 1)},
+                **{f"video_{i}": ("VIDEO",) for i in range(3, MAX_REFERENCE_VIDEOS + 1)},
             },
         }
 
@@ -806,21 +1001,26 @@ class VLLMOmniVideoReferences:
         audio_2: AudioInput | None = None,
         video_1: VideoInput | None = None,
         video_2: VideoInput | None = None,
+        image_3: torch.Tensor | None = None,
+        image_4: torch.Tensor | None = None,
+        image_5: torch.Tensor | None = None,
+        image_6: torch.Tensor | None = None,
+        image_7: torch.Tensor | None = None,
+        image_8: torch.Tensor | None = None,
+        image_9: torch.Tensor | None = None,
+        audio_3: AudioInput | None = None,
+        video_3: VideoInput | None = None,
         **kwargs,
     ):
         if kwargs:
             logger.info("Uncaught kwargs: %s", kwargs)
         refs = VideoReferences()
-        if image_1 is not None:
-            refs["image_1"] = image_1
-        if image_2 is not None:
-            refs["image_2"] = image_2
-        if audio_1 is not None:
-            refs["audio_1"] = audio_1
-        if audio_2 is not None:
-            refs["audio_2"] = audio_2
-        if video_1 is not None:
-            refs["video_1"] = video_1
-        if video_2 is not None:
-            refs["video_2"] = video_2
+        for kind, values in (
+            ("image", (image_1, image_2, image_3, image_4, image_5, image_6, image_7, image_8, image_9)),
+            ("video", (video_1, video_2, video_3)),
+            ("audio", (audio_1, audio_2, audio_3)),
+        ):
+            for index, value in enumerate(values, start=1):
+                if value is not None:
+                    refs[f"{kind}_{index}"] = value
         return (refs,)

apps/ComfyUI-vLLM-Omni/comfyui_vllm_omni/utils/api_client.py

@@ -1,3 +1,6 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
+
 """
 An high-level API client adapter that forwards ComfyUI inputs to vLLM-Omni's REST API,
 and transforms the API responses back to ComfyUI formats.
@@ -17,6 +20,7 @@ from comfy_api.input import AudioInput, VideoInput
 
 from .format import (
     audio_to_base64,
+    audio_to_bytes,
     base64_to_audio,
     base64_to_image_tensor,
     bytes_to_audio,
@@ -28,7 +32,13 @@ from .format import (
 )
 from .logger import get_logger, pretty_printer
 from .models import lookup_model_spec
-from .types import AudioFormat
+from .types import (
+    MAX_REFERENCE_AUDIOS,
+    MAX_REFERENCE_IMAGES,
+    MAX_REFERENCE_VIDEOS,
+    MAX_TOTAL_REFERENCES,
+    AudioFormat,
+)
 
 logger = get_logger(__name__)
 
@@ -271,8 +281,15 @@ class VLLMOmniClient:
         sampling_params: dict | None = None,
         model_params: dict | None = None,
         lora: dict | None = None,
+        spec_model: str | None = None,
         **extra_params,
     ) -> VideoInput:
+        """Post a video job and return the decoded result.
+
+        ``spec_model`` names the model whose payload spec builds the request, for
+        deployments that serve a known model under a different ``model`` alias. It
+        never reaches the wire; defaults to ``model``.
+        """
         if frame is not None and references is not None:
             raise ValueError("Provide only one of frame or references, not both.")
 
@@ -294,38 +311,44 @@ class VLLMOmniClient:
 
         # === multimodal inputs (first-last-frames, references, etc.) ===
         input_reference_image: torch.Tensor | None = None
-        audio_reference: AudioInput | None = None
-        reference_videos: list[VideoInput] = []
         video_task: str | None = None
 
         if frame is not None:
             input_reference_image = frame
             video_task = "fl2va"
         elif references is not None:
-            images = [references[k] for k in ("image_1", "image_2") if k in references and references[k] is not None]
-            audios = [references[k] for k in ("audio_1", "audio_2") if k in references and references[k] is not None]
-            videos = [references[k] for k in ("video_1", "video_2") if k in references and references[k] is not None]
-
-            if not images and not audios and not videos:
-                raise ValueError("references is empty; connect at least one image/audio/video.")
-
-            if videos:
-                if images or audios:
-                    raise ValueError(
-                        "Video references cannot be combined with image or audio references. "
-                        "Connect only video_1/video_2 for multi-video Ref2VA."
-                    )
-                reference_videos = videos
-                video_task = "ref2va"
-            elif len(images) == 1 and len(audios) == 1:
-                input_reference_image = images[0]
-                audio_reference = audios[0]
-                video_task = "ref2va"
-            else:
+            reference_formats = (
+                ("image", MAX_REFERENCE_IMAGES, "png", "image/png", image_tensor_to_png_bytes),
+                ("video", MAX_REFERENCE_VIDEOS, "mp4", "video/mp4", video_to_bytes),
+                ("audio", MAX_REFERENCE_AUDIOS, "mp3", "audio/mpeg", audio_to_bytes),
+            )
+            supported_inputs = {f"{kind}_{i}" for kind, limit, *_ in reference_formats for i in range(1, limit + 1)}
+            connected = {name: value for name, value in references.items() if value is not None}
+            unsupported = connected.keys() - supported_inputs
+            if unsupported:
+                raise ValueError(f"Unsupported reference input(s): {', '.join(sorted(unsupported))}.")
+            if not any(name.startswith(("image_", "video_")) for name in connected):
                 raise ValueError(
-                    "Invalid references combination. Supported modes: "
-                    "(1) one or more videos only, or (2) exactly one image and one audio."
+                    "references requires at least one image or video; audio-only inputs are not supported."
                 )
+            if len(connected) > MAX_TOTAL_REFERENCES:
+                raise ValueError(
+                    f"references supports at most {MAX_TOTAL_REFERENCES} inputs in total "
+                    f"(up to {MAX_REFERENCE_IMAGES} images, {MAX_REFERENCE_VIDEOS} videos, "
+                    f"and {MAX_REFERENCE_AUDIOS} audios)."
+                )
+            for kind, limit, extension, content_type, encode in reference_formats:
+                for index in range(1, limit + 1):
+                    name = f"{kind}_{index}"
+                    if name in connected:
+                        filename = f"{name}.{extension}"
+                        form.add_field(
+                            "input_references",
+                            encode(connected[name], filename),
+                            filename=filename,
+                            content_type=content_type,
+                        )
+            video_task = "ref2va"
         else:
             video_task = "t2va"
 
@@ -338,32 +361,19 @@ class VLLMOmniClient:
                 content_type="image/png",
             )
 
-        if audio_reference is not None:
-            form.add_field(
-                "audio_reference",
-                json.dumps({"audio_url": audio_to_base64(audio_reference)}, ensure_ascii=False),
-            )
-
-        for idx, video in enumerate(reference_videos, start=1):
-            video_filename = f"reference_{idx}.mp4"
-            form.add_field(
-                "input_references",
-                video_to_bytes(video, video_filename),
-                filename=video_filename,
-                content_type="video/mp4",
-            )
-
         # === model specific params. Either use a specialized builder, or add flattened fields as-is ===
         if model_params is not None:
             model_params = dict(model_params)
             model_params.pop("type", None)
 
-        spec, _ = lookup_model_spec(model)
+        spec, _ = lookup_model_spec(spec_model or model)
         params_builder = spec.get("params_builder") if spec else None
         if params_builder is not None:
             form_fields = params_builder(
                 model_params or {},
                 extra_params={**extra_params, "task": video_task},
+                width=width,
+                height=height,
             )
             for k, v in form_fields.items():
                 form.add_field(k, v if isinstance(v, str) else str(v))

apps/ComfyUI-vLLM-Omni/comfyui_vllm_omni/utils/format.py

@@ -1,3 +1,6 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
+
 """Image/tensor format helpers.
 
 The image generation part is derived from dougbtv/comfyui-vllm-omni by Doug (@dougbtv).
@@ -170,6 +173,10 @@ def bytes_to_video(video_bytes: bytes) -> VideoInput:
                 audio_stream = container.streams.audio[-1]
                 audio_frames = []
                 resampler = AudioResampler(format="fltp")
+                # Decoding the video stream above left the container at EOF.
+                # Rewind before demuxing again, or this pass yields no frames
+                # and the audio of a T2VA result is silently dropped.
+                container.seek(0)
                 for frame in container.decode(audio_stream):
                     if not isinstance(frame, AudioFrame):
                         continue

apps/ComfyUI-vLLM-Omni/comfyui_vllm_omni/utils/models.py

@@ -1,3 +1,6 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
+
 import json
 import re
 from typing import Any
@@ -34,10 +37,28 @@ def _qwen25_payload_preprocessor(payload: dict) -> dict:
     return payload
 
 
+MINIMAX_H3_ASPECT_RATIOS = {
+    "21:9": 21.0 / 9.0,
+    "16:9": 16.0 / 9.0,
+    "4:3": 4.0 / 3.0,
+    "1:1": 1.0,
+    "3:4": 3.0 / 4.0,
+    "9:16": 9.0 / 16.0,
+}
+
+
+def _nearest_minimaxh3_aspect_ratio(width: int, height: int) -> str:
+    """Pick the supported named ratio closest to the requested frame size."""
+    target = float(width) / float(height)
+    return min(MINIMAX_H3_ASPECT_RATIOS, key=lambda name: abs(MINIMAX_H3_ASPECT_RATIOS[name] - target))
+
+
 def _minimaxh3_params_builder(
     model_params: dict[str, Any],
     *,
     extra_params: dict[str, Any],
+    width: int | None = None,
+    height: int | None = None,
 ) -> dict[str, Any]:
     """Build multipart form fields for MiniMax-H3 from model_params + routed task."""
     params = dict(model_params)
@@ -53,6 +74,12 @@ def _minimaxh3_params_builder(
     if params:
         logger.warning("Unused MiniMax-H3 model params ignored: %s", sorted(params))
 
+    # H3 refuses t2va without a named aspect ratio, and the Generate Video node
+    # carries width/height rather than a ratio. fl2va takes its ratio from the
+    # input image and ref2va defaults server-side, so only t2va needs this.
+    if merged_extra_params.get("task") == "t2va" and "aspect_ratio" not in merged_extra_params and width and height:
+        merged_extra_params["aspect_ratio"] = _nearest_minimaxh3_aspect_ratio(width, height)
+
     if merged_extra_params:
         form_fields["extra_params"] = json.dumps(merged_extra_params, ensure_ascii=False)
     return form_fields

apps/ComfyUI-vLLM-Omni/comfyui_vllm_omni/utils/types.py

@@ -1,3 +1,6 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
+
 from collections.abc import Callable
 from enum import Enum, auto
 from typing import Any, Literal, TypeAlias
@@ -27,10 +30,20 @@ class MiniMaxH3ModelSpecificParams(dict):
     pass
 
 
+MAX_REFERENCE_IMAGES = 9
+MAX_REFERENCE_VIDEOS = 3
+MAX_REFERENCE_AUDIOS = 3
+MAX_TOTAL_REFERENCES = 12
+
+
 class VideoReferences(dict):
     pass
 
 
+class FastH3Deployment(dict):
+    """Descriptor for a server that fused FastH3 at startup."""
+
+
 class ModelMode(Enum):
     IMAGE_GENERATION = auto()
     VIDEO_GENERATION = auto()

apps/ComfyUI-vLLM-Omni/docs/minimax-h3-t2v.md

@@ -0,0 +1,94 @@
+# MiniMax H3 Text to Video (WF-01)
+
+Import [MiniMax_H3_Text_to_Video.json](../example_workflows/MiniMax_H3_Text_to_Video.json)
+into ComfyUI, or select it under **Templates → ComfyUI-vLLM-Omni**.
+The template connects existing nodes to a remote H3 service and saves the
+returned video with audio. ComfyUI does not load H3 weights.
+
+## Setup
+
+Install the [ComfyUI extension](../README.md#installation) and follow the
+[H3 deployment recipe](../../../recipes/MiniMaxAI/MiniMax-H3.md) for your hardware.
+For a two-GPU FL2VA server with sufficient memory:
+
+```bash
+export MODEL=/path/to/MiniMax-H3/FL2VA
+CUDA_VISIBLE_DEVICES=0,1 VLLM_WORKER_MULTIPROC_METHOD=spawn \
+VLLM_OMNI_VIDEO_SYNC_TIMEOUT=14400 \
+vllm serve "$MODEL" --omni --host 127.0.0.1 --port 8000 \
+  --served-model-name MiniMaxAI/MiniMax-H3 --trust-remote-code \
+  --task-type fl2va --num-gpus 2 --tensor-parallel-size 2 \
+  --text-encoder-tp-size 2 --enforce-eager
+```
+
+Set Generate Video's **url** to the address reachable from ComfyUI, including
+`/v1`, and match its **model** to the served name. The defaults are
+`http://localhost:8000/v1` and `MiniMaxAI/MiniMax-H3`.
+
+## Base and Turbo
+
+Base presets are connected by default. Write the scene and its sound in the
+same **prompt**; leave **frame**, **references**, and **fast_h3** disconnected.
+The template uses 1344×768, 24 FPS, seed 1101 and **duration** 5.167 seconds
+(124 frames, satisfying H3's `17k+5` constraint). The existing client infers
+`16:9` from these dimensions. For portrait output, use 768×1344.
+
+| Setting | Base | Turbo v1.0 768p |
+| --- | --- | --- |
+| Inference steps | 50 | 5 (4 forwards) |
+| Video flow shift | 12 | 6 |
+| Audio flow shift | 3 | 3 |
+| Remote LoRA | Disconnected | Connected, scale 1 |
+
+For Turbo, use the **Diffusers-layout** artifact from
+[lightx2v/Minimax-h3-Turbo](https://huggingface.co/lightx2v/Minimax-h3-Turbo):
+
+```text
+minimax_h3_fl2v_turbo_4step_v1.0_768p_bf16.safetensors
+```
+
+Add `--lora-backend peft --lora-path "$TURBO_LORA"` to the server command,
+where `TURBO_LORA` is the full server-side path to that file. Enter the same
+path in Remote LoRA's **local_path**; retain name `h3-turbo-v1.0-768p`, scale
+`1.0`, and **int_id** `0` so the existing API derives the ID from the path.
+The template leaves the path blank for users to configure. The `_comfyui_`
+export is not supported by this backend.
+
+Connect **Turbo sampling**, **Turbo H3 params**, and **Remote LoRA** to Generate
+Video's `sampling_params`, `model_params`, and `lora` inputs, replacing the
+Base links. To return to Base, restore both Base links and disconnect LoRA.
+Other Turbo artifacts may need different settings; see the
+[artifact table](../../../recipes/MiniMaxAI/MiniMax-H3.md#turbo-lora).
+
+If importing an older template with a frame-count widget, reimport this JSON
+or set **duration** to **5.167**; do not interpret 124 frames as 124 seconds.
+
+## Validation
+
+Queue the configured workflow and inspect the MP4 saved under
+`ComfyUI/output/video/MiniMax_H3_WF01_*.mp4`:
+
+```bash
+ffprobe -v error -show_streams -show_format -of json "$OUTPUT"
+ffmpeg -v error -i "$OUTPUT" -f null -
+```
+
+Check the requested dimensions, frame count, 24 FPS and both video/audio
+streams, then watch and listen to the saved file. The current upstream
+extension includes the shared audio-preserving decoder fix.
+
+On September 16, the narrowed workflow was validated through ComfyUI using
+existing upstream Python code at `4a8ac297`, two H20-3e GPUs, vLLM 0.29.0,
+PyTorch 2.13.0+cu129 and ComfyUI `1d48d9cf`. The run used the template's
+forest-stream prompt, explicit server-side Turbo LoRA path, and settings
+above, with `TORCH_SDPA`, `--usp 1 --ring 1`, and tiled VAE parallelism on two
+GPUs. ComfyUI execution took 131.55 seconds. SaveVideo produced 124 H.264
+frames at 1344×768 and 24 FPS, plus 32 kHz stereo AAC audio. Video/audio
+durations were 5.166667/5.167000 seconds; full decoding passed and audio was
+nonzero. This is model-generated output, not a ground-truth reference or a
+subjective audio-quality assessment. Base graph validation passed with Turbo
+nodes disconnected; full 50-step Base quality was not evaluated.
+
+Based on the official
+[ComfyUI H3 T2V template](https://github.com/Comfy-Org/workflow_templates/blob/main/templates/video_minimax_h3_t2v.json),
+using remote inference instead of local model loaders.

apps/ComfyUI-vLLM-Omni/docs/wf07-h3-upscale.md

@@ -0,0 +1,21 @@
+# WF-07: H3 video upscale with SeedVR2
+
+Generate video and audio through vLLM-Omni H3, upscale the decoded frames with SeedVR2, and save both the original and upscaled videos. The generated audio and FPS connect directly to Create Video.
+
+## Requirements
+
+- ComfyUI with ComfyUI-vLLM-Omni and [ComfyUI-SeedVR2_VideoUpscaler](https://github.com/numz/ComfyUI-SeedVR2_VideoUpscaler#-installation) installed.
+- An H3 service using the FL2VA checkpoint partition. Follow the [H3 recipe](../../../recipes/MiniMaxAI/MiniMax-H3.md) for server setup.
+- `seedvr2_ema_3b_fp16.safetensors` and `ema_vae_fp16.safetensors` from [SeedVR2's ComfyUI weights](https://huggingface.co/numz/SeedVR2_comfyUI/tree/main), placed in `ComfyUI/models/SEEDVR2`. The SeedVR2 nodes can also download them on first use.
+
+## Use
+
+Open **vLLM-Omni MiniMax H3 Video Upscale.json** from the example workflows. An API-format version is provided alongside it. Set the Generate Video URL/model and the LoRA path as seen by the H3 server.
+
+The preset uses the LightX2V v1.0 768p Turbo adapter in Diffusers format: `minimax_h3_fl2v_turbo_4step_v1.0_768p_bf16.safetensors`, five sigma points, video/audio flow shifts 6/3, and LoRA scale 1. Preload the same adapter on the server using the [Turbo LoRA recipe](../../../recipes/MiniMaxAI/MiniMax-H3.md#turbo-lora). For base H3, disconnect LoRA, set inference steps to 50, and set video flow shift to 12.
+
+The H3 preset requests 1344 x 768 at 24 FPS for 5.167 seconds, which gives 124 frames. SeedVR2's `resolution=1536` sets the target short edge, producing 2688 x 1536 for this input. Update this value when changing the input size; it is an output resolution, not a fixed 2x multiplier.
+
+The SeedVR2 preset uses 3B FP16, five-frame batches, one overlapping frame, uniform batch padding, and LAB color correction. The DiT loader uses CPU offload with 32 swapped blocks; the VAE uses tiled encoding and decoding. Select the local GPU in both model loaders. See the [SeedVR2 settings](https://github.com/numz/ComfyUI-SeedVR2_VideoUpscaler#-usage) when adjusting memory use or batch size.
+
+Run the workflow. The two Save Video nodes write MP4s under `WF07/generated` and `WF07/upscaled` in ComfyUI's output directory.

apps/ComfyUI-vLLM-Omni/example_workflows/MiniMax_H3_Text_to_Video.json

@@ -0,0 +1,362 @@
+{
+  "last_node_id": 8,
+  "last_link_id": 3,
+  "nodes": [
+    {
+      "id": 1,
+      "type": "VLLMOmniDiffusionSampling",
+      "pos": [
+        20,
+        20
+      ],
+      "size": [
+        310,
+        280
+      ],
+      "flags": {},
+      "order": 0,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [
+        {
+          "name": "diffusion sampling params",
+          "type": "SAMPLING_PARAMS",
+          "links": [
+            1
+          ],
+          "slot_index": 0
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "VLLMOmniDiffusionSampling"
+      },
+      "widgets_values": [
+        1,
+        50,
+        1.0,
+        1.0,
+        false,
+        false,
+        1101,
+        "fixed"
+      ],
+      "title": "Base sampling"
+    },
+    {
+      "id": 2,
+      "type": "VLLMOmniMiniMaxH3Params",
+      "pos": [
+        20,
+        340
+      ],
+      "size": [
+        310,
+        140
+      ],
+      "flags": {},
+      "order": 1,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [
+        {
+          "name": "MiniMax-H3 video params",
+          "type": "VIDEO_PARAMS",
+          "links": [
+            2
+          ],
+          "slot_index": 0
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "VLLMOmniMiniMaxH3Params"
+      },
+      "widgets_values": [
+        3.0,
+        12.0
+      ],
+      "title": "Base H3 params"
+    },
+    {
+      "id": 3,
+      "type": "VLLMOmniGenerateVideo",
+      "pos": [
+        430,
+        20
+      ],
+      "size": [
+        430,
+        440
+      ],
+      "flags": {},
+      "order": 2,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "frame",
+          "type": "IMAGE",
+          "link": null
+        },
+        {
+          "name": "references",
+          "type": "VIDEO_REFERENCES",
+          "link": null
+        },
+        {
+          "name": "sampling_params",
+          "type": "SAMPLING_PARAMS",
+          "link": 1
+        },
+        {
+          "name": "lora",
+          "type": "REMOTE_LORA",
+          "link": null
+        },
+        {
+          "name": "model_params",
+          "type": "VIDEO_PARAMS",
+          "link": 2
+        },
+        {
+          "name": "fast_h3",
+          "type": "FASTH3_DEPLOYMENT",
+          "link": null
+        },
+        {
+          "name": "duration",
+          "type": "FLOAT",
+          "widget": {
+            "name": "duration"
+          },
+          "link": null
+        }
+      ],
+      "outputs": [
+        {
+          "name": "video",
+          "type": "VIDEO",
+          "links": [
+            3
+          ],
+          "slot_index": 0
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "VLLMOmniGenerateVideo"
+      },
+      "widgets_values": [
+        "http://localhost:8000/v1",
+        "MiniMaxAI/MiniMax-H3",
+        "A small stream flows through a sunlit forest. The camera slowly moves toward the water. Birds chirp in the trees and water gurgles over smooth stones. Natural continuous sound, no music.",
+        "",
+        1344,
+        768,
+        24,
+        5.167
+      ]
+    },
+    {
+      "id": 4,
+      "type": "SaveVideo",
+      "pos": [
+        940,
+        20
+      ],
+      "size": [
+        440,
+        440
+      ],
+      "flags": {},
+      "order": 3,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "video",
+          "type": "VIDEO",
+          "link": 3
+        }
+      ],
+      "outputs": [
+        {
+          "name": "video",
+          "type": "VIDEO",
+          "links": null,
+          "slot_index": 0
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "SaveVideo"
+      },
+      "widgets_values": [
+        "video/MiniMax_H3_WF01",
+        "mp4",
+        "auto"
+      ]
+    },
+    {
+      "id": 5,
+      "type": "VLLMOmniRemoteLoRA",
+      "pos": [
+        430,
+        540
+      ],
+      "size": [
+        430,
+        160
+      ],
+      "flags": {},
+      "order": 4,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [
+        {
+          "name": "lora",
+          "type": "REMOTE_LORA",
+          "links": null,
+          "slot_index": 0
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "VLLMOmniRemoteLoRA"
+      },
+      "widgets_values": [
+        "",
+        "h3-turbo-v1.0-768p",
+        1.0,
+        0
+      ],
+      "title": "Optional Turbo: server LoRA path"
+    },
+    {
+      "id": 6,
+      "type": "VLLMOmniDiffusionSampling",
+      "pos": [
+        20,
+        540
+      ],
+      "size": [
+        310,
+        280
+      ],
+      "flags": {},
+      "order": 5,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [
+        {
+          "name": "diffusion sampling params",
+          "type": "SAMPLING_PARAMS",
+          "links": null,
+          "slot_index": 0
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "VLLMOmniDiffusionSampling"
+      },
+      "widgets_values": [
+        1,
+        5,
+        1.0,
+        1.0,
+        false,
+        false,
+        1101,
+        "fixed"
+      ],
+      "title": "Turbo sampling (4 forwards)"
+    },
+    {
+      "id": 7,
+      "type": "VLLMOmniMiniMaxH3Params",
+      "pos": [
+        430,
+        740
+      ],
+      "size": [
+        310,
+        140
+      ],
+      "flags": {},
+      "order": 6,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [
+        {
+          "name": "MiniMax-H3 video params",
+          "type": "VIDEO_PARAMS",
+          "links": null,
+          "slot_index": 0
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "VLLMOmniMiniMaxH3Params"
+      },
+      "widgets_values": [
+        3.0,
+        6.0
+      ],
+      "title": "Turbo H3 params"
+    },
+    {
+      "id": 8,
+      "type": "Note",
+      "pos": [
+        940,
+        520
+      ],
+      "size": [
+        440,
+        440
+      ],
+      "flags": {},
+      "order": 7,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [],
+      "properties": {
+        "Node name for S&R": "Note"
+      },
+      "widgets_values": [
+        "MiniMax H3 Text to Video (WF-01)\n\nBASE: connected presets, 50 steps / video flow 12 / audio flow 3.\n\nTURBO: set Remote LoRA local_path to the SERVER path of minimax_h3_fl2v_turbo_4step_v1.0_768p_bf16.safetensors from lightx2v/Minimax-h3-Turbo. Keep name h3-turbo-v1.0-768p, scale 1.0 and int_id 0. Connect Turbo sampling, Turbo H3 params, and LoRA to Generate Video, replacing the two Base links. Use 5 steps (4 forwards), video flow 6 and audio flow 3. Do not use the ComfyUI-layout artifact.\n\nKeep frame, references and fast_h3 disconnected. 1344x768, 24 FPS, duration 5.167 seconds = 124 frames (17*7+5). The client selects the named 16:9 preset from the dimensions. Save Video preserves video and audio.\n\nSet URL and served model name for your server. See docs/minimax-h3-t2v.md for setup and validation. Source: Comfy-Org/workflow_templates/templates/video_minimax_h3_t2v.json."
+      ],
+      "title": "Setup / Base and Turbo"
+    }
+  ],
+  "links": [
+    [
+      1,
+      1,
+      0,
+      3,
+      2,
+      "SAMPLING_PARAMS"
+    ],
+    [
+      2,
+      2,
+      0,
+      3,
+      4,
+      "VIDEO_PARAMS"
+    ],
+    [
+      3,
+      3,
+      0,
+      4,
+      0,
+      "VIDEO"
+    ]
+  ],
+  "groups": [],
+  "config": {},
+  "extra": {
+    "ds": {
+      "scale": 0.8,
+      "offset": [
+        50,
+        50
+      ]
+    }
+  },
+  "version": 0.4
+}

apps/ComfyUI-vLLM-Omni/example_workflows/vLLM-Omni FastH3 Text to Video.json

@@ -0,0 +1,322 @@
+{
+  "id": "c6db00da-2ccb-4ff8-8334-d67e1a267781",
+  "revision": 0,
+  "last_node_id": 5,
+  "last_link_id": 3,
+  "nodes": [
+    {
+      "id": 1,
+      "type": "MarkdownNote",
+      "pos": [40, 20],
+      "size": [420, 250],
+      "flags": {},
+      "order": 0,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [],
+      "title": "FastH3 server prerequisite",
+      "properties": {},
+      "widgets_values": [
+        "FastH3 must be fused when the vLLM-Omni server starts. Use --task-type fl2va and --lora-path pointing to exactly one FastH3 adapter file. Do not use CPU or layerwise offload. For VSA / Data-Free, also install fastvideo-kernel and use --diffusion-attention-backend FASTVIDEO_VSA --fastvideo-vsa-topk 64."
+      ],
+      "color": "#432",
+      "bgcolor": "#000"
+    },
+    {
+      "id": 2,
+      "type": "VLLMOmniFastH3Deployment",
+      "pos": [40, 320],
+      "size": [420, 160],
+      "flags": {},
+      "order": 1,
+      "mode": 0,
+      "inputs": [
+        {
+          "localized_name": "url",
+          "name": "url",
+          "type": "STRING",
+          "widget": {"name": "url"},
+          "link": null
+        },
+        {
+          "localized_name": "model",
+          "name": "model",
+          "type": "STRING",
+          "widget": {"name": "model"},
+          "link": null
+        }
+      ],
+      "outputs": [
+        {
+          "localized_name": "deployment",
+          "name": "deployment",
+          "type": "FASTH3_DEPLOYMENT",
+          "links": [1]
+        }
+      ],
+      "properties": {"Node name for S&R": "VLLMOmniFastH3Deployment"},
+      "widgets_values": [
+        "http://localhost:8000/v1",
+        "MiniMaxAI/MiniMax-H3"
+      ]
+    },
+    {
+      "id": 3,
+      "type": "VLLMOmniDiffusionSampling",
+      "pos": [40, 540],
+      "size": [420, 230],
+      "flags": {},
+      "order": 2,
+      "mode": 0,
+      "inputs": [
+        {
+          "localized_name": "n",
+          "name": "n",
+          "type": "INT",
+          "widget": {"name": "n"},
+          "link": null
+        },
+        {
+          "localized_name": "num_inference_steps",
+          "name": "num_inference_steps",
+          "type": "INT",
+          "widget": {"name": "num_inference_steps"},
+          "link": null
+        },
+        {
+          "localized_name": "guidance_scale",
+          "name": "guidance_scale",
+          "type": "FLOAT",
+          "widget": {"name": "guidance_scale"},
+          "link": null
+        },
+        {
+          "localized_name": "true_cfg_scale",
+          "name": "true_cfg_scale",
+          "type": "FLOAT",
+          "widget": {"name": "true_cfg_scale"},
+          "link": null
+        },
+        {
+          "localized_name": "vae_use_slicing",
+          "name": "vae_use_slicing",
+          "type": "BOOLEAN",
+          "widget": {"name": "vae_use_slicing"},
+          "link": null
+        },
+        {
+          "localized_name": "vae_use_tiling",
+          "name": "vae_use_tiling",
+          "type": "BOOLEAN",
+          "widget": {"name": "vae_use_tiling"},
+          "link": null
+        },
+        {
+          "localized_name": "seed",
+          "name": "seed",
+          "type": "INT",
+          "widget": {"name": "seed"},
+          "link": null
+        }
+      ],
+      "outputs": [
+        {
+          "localized_name": "diffusion sampling params",
+          "name": "diffusion sampling params",
+          "type": "SAMPLING_PARAMS",
+          "links": [2]
+        }
+      ],
+      "properties": {"Node name for S&R": "VLLMOmniDiffusionSampling"},
+      "widgets_values": [1, 4, 1, 1, false, false, 1101, "fixed"]
+    },
+    {
+      "id": 4,
+      "type": "VLLMOmniGenerateVideo",
+      "pos": [560, 210],
+      "size": [540, 610],
+      "flags": {},
+      "order": 3,
+      "mode": 0,
+      "inputs": [
+        {
+          "localized_name": "frame",
+          "name": "frame",
+          "shape": 7,
+          "type": "IMAGE",
+          "link": null
+        },
+        {
+          "localized_name": "references",
+          "name": "references",
+          "shape": 7,
+          "type": "VIDEO_REFERENCES",
+          "link": null
+        },
+        {
+          "localized_name": "sampling_params",
+          "name": "sampling_params",
+          "shape": 7,
+          "type": "SAMPLING_PARAMS",
+          "link": 2
+        },
+        {
+          "localized_name": "lora",
+          "name": "lora",
+          "shape": 7,
+          "type": "REMOTE_LORA",
+          "link": null
+        },
+        {
+          "localized_name": "model_params",
+          "name": "model_params",
+          "shape": 7,
+          "type": "VIDEO_PARAMS",
+          "link": null
+        },
+        {
+          "localized_name": "fast_h3",
+          "name": "fast_h3",
+          "shape": 7,
+          "type": "FASTH3_DEPLOYMENT",
+          "link": 1
+        },
+        {
+          "localized_name": "url",
+          "name": "url",
+          "type": "STRING",
+          "widget": {"name": "url"},
+          "link": null
+        },
+        {
+          "localized_name": "model",
+          "name": "model",
+          "type": "STRING",
+          "widget": {"name": "model"},
+          "link": null
+        },
+        {
+          "localized_name": "prompt",
+          "name": "prompt",
+          "type": "STRING",
+          "widget": {"name": "prompt"},
+          "link": null
+        },
+        {
+          "localized_name": "negative_prompt",
+          "name": "negative_prompt",
+          "type": "STRING",
+          "widget": {"name": "negative_prompt"},
+          "link": null
+        },
+        {
+          "localized_name": "width",
+          "name": "width",
+          "type": "INT",
+          "widget": {"name": "width"},
+          "link": null
+        },
+        {
+          "localized_name": "height",
+          "name": "height",
+          "type": "INT",
+          "widget": {"name": "height"},
+          "link": null
+        },
+        {
+          "localized_name": "fps",
+          "name": "fps",
+          "type": "INT",
+          "widget": {"name": "fps"},
+          "link": null
+        },
+        {
+          "localized_name": "duration",
+          "name": "duration",
+          "type": "FLOAT",
+          "widget": {"name": "duration"},
+          "link": null
+        }
+      ],
+      "outputs": [
+        {
+          "localized_name": "video",
+          "name": "video",
+          "type": "VIDEO",
+          "links": [3]
+        }
+      ],
+      "properties": {"Node name for S&R": "VLLMOmniGenerateVideo"},
+      "widgets_values": [
+        "http://localhost:8000/v1",
+        "MiniMaxAI/MiniMax-H3",
+        "A cinematic live-action singer performs naturally on a neon-lit stage, with synchronized vocals and environmental sound.",
+        "",
+        1344,
+        768,
+        24,
+        4.458
+      ]
+    },
+    {
+      "id": 5,
+      "type": "SaveVideo",
+      "pos": [1200, 280],
+      "size": [380, 310],
+      "flags": {},
+      "order": 4,
+      "mode": 0,
+      "inputs": [
+        {
+          "localized_name": "video",
+          "name": "video",
+          "type": "VIDEO",
+          "link": 3
+        },
+        {
+          "localized_name": "filename_prefix",
+          "name": "filename_prefix",
+          "type": "STRING",
+          "widget": {"name": "filename_prefix"},
+          "link": null
+        },
+        {
+          "localized_name": "format",
+          "name": "format",
+          "type": "COMBO",
+          "widget": {"name": "format"},
+          "link": null
+        },
+        {
+          "localized_name": "codec",
+          "name": "codec",
+          "type": "COMFY_DYNAMICCOMBO_V3",
+          "widget": {"name": "codec"},
+          "link": null
+        }
+      ],
+      "outputs": [
+        {
+          "localized_name": "video",
+          "name": "video",
+          "type": "VIDEO",
+          "links": null
+        }
+      ],
+      "properties": {},
+      "widgets_values": ["video/ComfyUI-FastH3", "auto", "auto"]
+    }
+  ],
+  "links": [
+    [1, 2, 0, 4, 5, "FASTH3_DEPLOYMENT"],
+    [2, 3, 0, 4, 2, "SAMPLING_PARAMS"],
+    [3, 4, 0, 5, 0, "VIDEO"]
+  ],
+  "groups": [],
+  "config": {},
+  "extra": {
+    "workflowRendererVersion": "LG",
+    "ds": {"scale": 0.9, "offset": [100, 100]}
+  },
+  "version": 0.4
+}

apps/ComfyUI-vLLM-Omni/example_workflows/vLLM-Omni MiniMax H3 Video Upscale.json

@@ -0,0 +1,724 @@
+{
+  "version": 0.4,
+  "last_node_id": 12,
+  "last_link_id": 13,
+  "nodes": [
+    {
+      "id": 1,
+      "type": "VLLMOmniGenerateVideo",
+      "pos": [
+        360,
+        20
+      ],
+      "size": [
+        330,
+        480
+      ],
+      "flags": {},
+      "order": 1,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "frame",
+          "type": "IMAGE",
+          "shape": 7,
+          "link": null
+        },
+        {
+          "name": "references",
+          "type": "VIDEO_REFERENCES",
+          "shape": 7,
+          "link": null
+        },
+        {
+          "name": "sampling_params",
+          "type": "SAMPLING_PARAMS",
+          "shape": 7,
+          "link": 11
+        },
+        {
+          "name": "lora",
+          "type": "REMOTE_LORA",
+          "shape": 7,
+          "link": 12
+        },
+        {
+          "name": "model_params",
+          "type": "VIDEO_PARAMS",
+          "shape": 7,
+          "link": 10
+        },
+        {
+          "name": "fast_h3",
+          "type": "FASTH3_DEPLOYMENT",
+          "shape": 7,
+          "link": null
+        }
+      ],
+      "outputs": [
+        {
+          "name": "video",
+          "type": "VIDEO",
+          "links": [
+            1,
+            9
+          ]
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "VLLMOmniGenerateVideo"
+      },
+      "widgets_values": [
+        "http://127.0.0.1:8091/v1",
+        "MiniMaxAI/MiniMax-H3",
+        "A small brass bell hanging from a wooden beam in a quiet garden. A gentle breeze swings the bell and it rings clearly. Static camera, natural daylight, continuous motion, synchronized ringing and soft birdsong.",
+        "",
+        1344,
+        768,
+        24,
+        5.167
+      ]
+    },
+    {
+      "id": 2,
+      "type": "GetVideoComponents",
+      "pos": [
+        730,
+        20
+      ],
+      "size": [
+        250,
+        125
+      ],
+      "flags": {},
+      "order": 1,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "video",
+          "type": "VIDEO",
+          "link": 1
+        }
+      ],
+      "outputs": [
+        {
+          "name": "images",
+          "type": "IMAGE",
+          "links": [
+            2
+          ]
+        },
+        {
+          "name": "audio",
+          "type": "AUDIO",
+          "links": [
+            6
+          ]
+        },
+        {
+          "name": "fps",
+          "type": "FLOAT",
+          "links": [
+            5
+          ]
+        },
+        {
+          "name": "bit_depth",
+          "type": "INT",
+          "links": [
+            7
+          ]
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "GetVideoComponents"
+      }
+    },
+    {
+      "id": 3,
+      "type": "SeedVR2LoadDiTModel",
+      "pos": [
+        730,
+        290
+      ],
+      "size": [
+        340,
+        245
+      ],
+      "flags": {},
+      "order": 4,
+      "mode": 0,
+      "inputs": [
+        {
+          "localized_name": "torch_compile_args",
+          "name": "torch_compile_args",
+          "shape": 7,
+          "type": "TORCH_COMPILE_ARGS",
+          "link": null
+        }
+      ],
+      "outputs": [
+        {
+          "localized_name": "SEEDVR2_DIT",
+          "name": "SEEDVR2_DIT",
+          "type": "SEEDVR2_DIT",
+          "links": [
+            3
+          ]
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "SeedVR2LoadDiTModel"
+      },
+      "widgets_values": [
+        "seedvr2_ema_3b_fp16.safetensors",
+        "cuda:0",
+        32,
+        false,
+        "cpu",
+        false,
+        "sdpa"
+      ]
+    },
+    {
+      "id": 4,
+      "type": "SeedVR2VideoUpscaler",
+      "pos": [
+        1120,
+        290
+      ],
+      "size": [
+        340,
+        415
+      ],
+      "flags": {},
+      "order": 8,
+      "mode": 0,
+      "inputs": [
+        {
+          "localized_name": "image",
+          "name": "image",
+          "type": "IMAGE",
+          "link": 2
+        },
+        {
+          "localized_name": "dit",
+          "name": "dit",
+          "type": "SEEDVR2_DIT",
+          "link": 3
+        },
+        {
+          "localized_name": "vae",
+          "name": "vae",
+          "type": "SEEDVR2_VAE",
+          "link": 13
+        }
+      ],
+      "outputs": [
+        {
+          "localized_name": "IMAGE",
+          "name": "IMAGE",
+          "type": "IMAGE",
+          "links": [
+            4
+          ]
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "SeedVR2VideoUpscaler"
+      },
+      "widgets_values": [
+        42,
+        "fixed",
+        1536,
+        0,
+        5,
+        true,
+        "lab",
+        1,
+        0,
+        0.0,
+        0.0,
+        "cpu",
+        false
+      ]
+    },
+    {
+      "id": 5,
+      "type": "CreateVideo",
+      "pos": [
+        1490,
+        290
+      ],
+      "size": [
+        250,
+        175
+      ],
+      "flags": {},
+      "order": 4,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "images",
+          "type": "IMAGE",
+          "link": 4
+        },
+        {
+          "name": "fps",
+          "type": "FLOAT",
+          "widget": {
+            "name": "fps"
+          },
+          "link": 5
+        },
+        {
+          "name": "audio",
+          "type": "AUDIO",
+          "shape": 7,
+          "link": 6
+        },
+        {
+          "name": "bit_depth",
+          "type": "INT",
+          "widget": {
+            "name": "bit_depth"
+          },
+          "link": 7
+        }
+      ],
+      "outputs": [
+        {
+          "name": "VIDEO",
+          "type": "VIDEO",
+          "links": [
+            8
+          ]
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "CreateVideo"
+      },
+      "widgets_values": [
+        24,
+        8
+      ]
+    },
+    {
+      "id": 6,
+      "type": "SaveVideo",
+      "pos": [
+        1790,
+        290
+      ],
+      "size": [
+        330,
+        280
+      ],
+      "flags": {},
+      "order": 5,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "video",
+          "type": "VIDEO",
+          "link": 8
+        },
+        {
+          "name": "filename_prefix",
+          "type": "STRING",
+          "widget": {
+            "name": "filename_prefix"
+          },
+          "link": null
+        },
+        {
+          "name": "format",
+          "type": "COMBO",
+          "widget": {
+            "name": "format"
+          },
+          "link": null
+        },
+        {
+          "name": "codec",
+          "type": "COMFY_DYNAMICCOMBO_V3",
+          "widget": {
+            "name": "codec"
+          },
+          "link": null
+        }
+      ],
+      "outputs": [
+        {
+          "name": "video",
+          "type": "VIDEO",
+          "links": []
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "SaveVideo"
+      },
+      "widgets_values": [
+        "WF07/upscaled",
+        "mp4",
+        "auto"
+      ]
+    },
+    {
+      "id": 7,
+      "type": "VLLMOmniMiniMaxH3Params",
+      "pos": [
+        0,
+        20
+      ],
+      "size": [
+        320,
+        120
+      ],
+      "flags": {},
+      "order": 7,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [
+        {
+          "name": "MiniMax-H3 video params",
+          "type": "VIDEO_PARAMS",
+          "links": [
+            10
+          ]
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "VLLMOmniMiniMaxH3Params"
+      },
+      "widgets_values": [
+        3.0,
+        6.0
+      ]
+    },
+    {
+      "id": 8,
+      "type": "VLLMOmniDiffusionSampling",
+      "pos": [
+        0,
+        200
+      ],
+      "size": [
+        320,
+        280
+      ],
+      "flags": {},
+      "order": 8,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [
+        {
+          "name": "diffusion sampling params",
+          "type": "SAMPLING_PARAMS",
+          "links": [
+            11
+          ]
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "VLLMOmniDiffusionSampling"
+      },
+      "widgets_values": [
+        1,
+        5,
+        1.0,
+        1.0,
+        false,
+        true,
+        42,
+        "fixed"
+      ]
+    },
+    {
+      "id": 9,
+      "type": "VLLMOmniRemoteLoRA",
+      "pos": [
+        0,
+        530
+      ],
+      "size": [
+        320,
+        210
+      ],
+      "flags": {},
+      "order": 9,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [
+        {
+          "name": "lora",
+          "type": "REMOTE_LORA",
+          "links": [
+            12
+          ]
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "VLLMOmniRemoteLoRA"
+      },
+      "widgets_values": [
+        "models/minimax-h3-turbo/minimax_h3_fl2v_turbo_4step_v1.0_768p_bf16.safetensors",
+        "h3-turbo-v1.0",
+        1.0,
+        0
+      ]
+    },
+    {
+      "id": 10,
+      "type": "SaveVideo",
+      "pos": [
+        1120,
+        20
+      ],
+      "size": [
+        330,
+        190
+      ],
+      "flags": {},
+      "order": 10,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "video",
+          "type": "VIDEO",
+          "link": 9
+        },
+        {
+          "name": "filename_prefix",
+          "type": "STRING",
+          "widget": {
+            "name": "filename_prefix"
+          },
+          "link": null
+        },
+        {
+          "name": "format",
+          "type": "COMBO",
+          "widget": {
+            "name": "format"
+          },
+          "link": null
+        },
+        {
+          "name": "codec",
+          "type": "COMFY_DYNAMICCOMBO_V3",
+          "widget": {
+            "name": "codec"
+          },
+          "link": null
+        }
+      ],
+      "outputs": [
+        {
+          "name": "video",
+          "type": "VIDEO",
+          "links": []
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "SaveVideo"
+      },
+      "widgets_values": [
+        "WF07/generated",
+        "mp4",
+        "auto"
+      ],
+      "title": "Save generated video before upscale"
+    },
+    {
+      "id": 11,
+      "type": "Note",
+      "pos": [
+        360,
+        560
+      ],
+      "size": [
+        330,
+        260
+      ],
+      "flags": {},
+      "order": 11,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [],
+      "properties": {
+        "Node name for S&R": "Note"
+      },
+      "widgets_values": [
+        "WF-07: H3 + SeedVR2\n\nInstall ComfyUI-SeedVR2_VideoUpscaler. The DiT and VAE loaders use SeedVR2 3B FP16 and ema_vae_fp16.\n\nSet the H3 service URL/model and the server-side Turbo LoRA path.\n\n1536 is the target short edge for this 1344 x 768 preset (2x). Original audio and FPS connect directly to Create Video.\n\nFor base H3: disconnect LoRA, use 50 steps and video flow shift 12."
+      ]
+    },
+    {
+      "id": 12,
+      "type": "SeedVR2LoadVAEModel",
+      "pos": [
+        730,
+        590
+      ],
+      "size": [
+        340,
+        315
+      ],
+      "flags": {},
+      "order": 5,
+      "mode": 0,
+      "inputs": [
+        {
+          "localized_name": "torch_compile_args",
+          "name": "torch_compile_args",
+          "shape": 7,
+          "type": "TORCH_COMPILE_ARGS",
+          "link": null
+        }
+      ],
+      "outputs": [
+        {
+          "localized_name": "SEEDVR2_VAE",
+          "name": "SEEDVR2_VAE",
+          "type": "SEEDVR2_VAE",
+          "links": [
+            13
+          ]
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "SeedVR2LoadVAEModel"
+      },
+      "widgets_values": [
+        "ema_vae_fp16.safetensors",
+        "cuda:0",
+        true,
+        1024,
+        128,
+        true,
+        768,
+        128,
+        "false",
+        "cpu",
+        false
+      ]
+    }
+  ],
+  "links": [
+    [
+      1,
+      1,
+      0,
+      2,
+      0,
+      "VIDEO"
+    ],
+    [
+      2,
+      2,
+      0,
+      4,
+      0,
+      "IMAGE"
+    ],
+    [
+      3,
+      3,
+      0,
+      4,
+      1,
+      "SEEDVR2_DIT"
+    ],
+    [
+      4,
+      4,
+      0,
+      5,
+      0,
+      "IMAGE"
+    ],
+    [
+      5,
+      2,
+      2,
+      5,
+      1,
+      "FLOAT"
+    ],
+    [
+      6,
+      2,
+      1,
+      5,
+      2,
+      "AUDIO"
+    ],
+    [
+      7,
+      2,
+      3,
+      5,
+      3,
+      "INT"
+    ],
+    [
+      8,
+      5,
+      0,
+      6,
+      0,
+      "VIDEO"
+    ],
+    [
+      9,
+      1,
+      0,
+      10,
+      0,
+      "VIDEO"
+    ],
+    [
+      10,
+      7,
+      0,
+      1,
+      4,
+      "VIDEO_PARAMS"
+    ],
+    [
+      11,
+      8,
+      0,
+      1,
+      2,
+      "SAMPLING_PARAMS"
+    ],
+    [
+      12,
+      9,
+      0,
+      1,
+      3,
+      "REMOTE_LORA"
+    ],
+    [
+      13,
+      12,
+      0,
+      4,
+      2,
+      "SEEDVR2_VAE"
+    ]
+  ],
+  "groups": [],
+  "config": {},
+  "extra": {
+    "ds": {
+      "scale": 0.52,
+      "offset": [
+        100,
+        100
+      ]
+    }
+  }
+}

apps/ComfyUI-vLLM-Omni/example_workflows/vLLM-Omni MiniMax-H3 Reference to Video.json

@@ -0,0 +1,786 @@
+{
+  "id": "e7a4e606-6510-4c4e-b127-9a8b6bf3ea71",
+  "revision": 0,
+  "last_node_id": 11,
+  "last_link_id": 5,
+  "nodes": [
+    {
+      "id": 1,
+      "type": "LoadImage",
+      "pos": [
+        40,
+        100
+      ],
+      "size": [
+        320,
+        300
+      ],
+      "flags": {},
+      "order": 0,
+      "mode": 0,
+      "inputs": [
+        {
+          "localized_name": "image",
+          "name": "image",
+          "type": "COMBO",
+          "widget": {
+            "name": "image"
+          },
+          "link": null
+        },
+        {
+          "localized_name": "choose file to upload",
+          "name": "upload",
+          "type": "IMAGEUPLOAD",
+          "widget": {
+            "name": "upload"
+          },
+          "link": null
+        }
+      ],
+      "outputs": [
+        {
+          "localized_name": "IMAGE",
+          "name": "IMAGE",
+          "type": "IMAGE",
+          "links": [
+            1
+          ]
+        },
+        {
+          "localized_name": "MASK",
+          "name": "MASK",
+          "type": "MASK",
+          "links": null
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "LoadImage"
+      },
+      "widgets_values": [
+        "h3_reference.png",
+        "image"
+      ],
+      "title": "Reference image \u2014 select a local file"
+    },
+    {
+      "type": "LoadVideo",
+      "size": [
+        320,
+        200
+      ],
+      "flags": {},
+      "inputs": [],
+      "outputs": [
+        {
+          "name": "VIDEO",
+          "type": "VIDEO",
+          "links": null
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "LoadVideo"
+      },
+      "widgets_values": [
+        "h3_reference.mp4"
+      ],
+      "id": 2,
+      "pos": [
+        40,
+        470
+      ],
+      "order": 1,
+      "mode": 0,
+      "title": "Optional reference video"
+    },
+    {
+      "id": 3,
+      "type": "LoadAudio",
+      "pos": [
+        40,
+        1100
+      ],
+      "size": [
+        320,
+        160
+      ],
+      "flags": {},
+      "order": 2,
+      "mode": 0,
+      "inputs": [
+        {
+          "localized_name": "audio",
+          "name": "audio",
+          "type": "COMBO",
+          "widget": {
+            "name": "audio"
+          },
+          "link": null
+        },
+        {
+          "localized_name": "audioUI",
+          "name": "audioUI",
+          "type": "AUDIO_UI",
+          "widget": {
+            "name": "audioUI"
+          },
+          "link": null
+        },
+        {
+          "localized_name": "choose file to upload",
+          "name": "upload",
+          "type": "AUDIOUPLOAD",
+          "widget": {
+            "name": "upload"
+          },
+          "link": null
+        }
+      ],
+      "outputs": [
+        {
+          "localized_name": "AUDIO",
+          "name": "AUDIO",
+          "type": "AUDIO",
+          "links": null
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "LoadAudio"
+      },
+      "widgets_values": [
+        "h3_reference.wav",
+        null,
+        ""
+      ],
+      "title": "Optional reference audio"
+    },
+    {
+      "id": 4,
+      "type": "VLLMOmniVideoReferences",
+      "pos": [
+        420,
+        100
+      ],
+      "size": [
+        300,
+        365
+      ],
+      "flags": {},
+      "order": 3,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "image_1",
+          "type": "IMAGE",
+          "shape": 7,
+          "link": 1
+        },
+        {
+          "name": "image_2",
+          "type": "IMAGE",
+          "shape": 7,
+          "link": null
+        },
+        {
+          "name": "audio_1",
+          "type": "AUDIO",
+          "shape": 7,
+          "link": null
+        },
+        {
+          "name": "audio_2",
+          "type": "AUDIO",
+          "shape": 7,
+          "link": null
+        },
+        {
+          "name": "video_1",
+          "type": "VIDEO",
+          "shape": 7,
+          "link": null
+        },
+        {
+          "name": "video_2",
+          "type": "VIDEO",
+          "shape": 7,
+          "link": null
+        },
+        {
+          "name": "image_3",
+          "type": "IMAGE",
+          "shape": 7,
+          "link": null
+        },
+        {
+          "name": "image_4",
+          "type": "IMAGE",
+          "shape": 7,
+          "link": null
+        },
+        {
+          "name": "image_5",
+          "type": "IMAGE",
+          "shape": 7,
+          "link": null
+        },
+        {
+          "name": "image_6",
+          "type": "IMAGE",
+          "shape": 7,
+          "link": null
+        },
+        {
+          "name": "image_7",
+          "type": "IMAGE",
+          "shape": 7,
+          "link": null
+        },
+        {
+          "name": "image_8",
+          "type": "IMAGE",
+          "shape": 7,
+          "link": null
+        },
+        {
+          "name": "image_9",
+          "type": "IMAGE",
+          "shape": 7,
+          "link": null
+        },
+        {
+          "name": "audio_3",
+          "type": "AUDIO",
+          "shape": 7,
+          "link": null
+        },
+        {
+          "name": "video_3",
+          "type": "VIDEO",
+          "shape": 7,
+          "link": null
+        }
+      ],
+      "outputs": [
+        {
+          "localized_name": "references",
+          "name": "references",
+          "type": "VIDEO_REFERENCES",
+          "links": [
+            2
+          ]
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "VLLMOmniVideoReferences"
+      },
+      "widgets_values": [],
+      "title": "References \u2014 up to 12 total"
+    },
+    {
+      "id": 5,
+      "type": "VLLMOmniDiffusionSampling",
+      "pos": [
+        420,
+        530
+      ],
+      "size": [
+        310,
+        275
+      ],
+      "flags": {},
+      "order": 4,
+      "mode": 0,
+      "inputs": [
+        {
+          "localized_name": "n",
+          "name": "n",
+          "type": "INT",
+          "widget": {
+            "name": "n"
+          },
+          "link": null
+        },
+        {
+          "localized_name": "num_inference_steps",
+          "name": "num_inference_steps",
+          "type": "INT",
+          "widget": {
+            "name": "num_inference_steps"
+          },
+          "link": null
+        },
+        {
+          "localized_name": "guidance_scale",
+          "name": "guidance_scale",
+          "type": "FLOAT",
+          "widget": {
+            "name": "guidance_scale"
+          },
+          "link": null
+        },
+        {
+          "localized_name": "true_cfg_scale",
+          "name": "true_cfg_scale",
+          "type": "FLOAT",
+          "widget": {
+            "name": "true_cfg_scale"
+          },
+          "link": null
+        },
+        {
+          "localized_name": "vae_use_slicing",
+          "name": "vae_use_slicing",
+          "type": "BOOLEAN",
+          "widget": {
+            "name": "vae_use_slicing"
+          },
+          "link": null
+        },
+        {
+          "localized_name": "vae_use_tiling",
+          "name": "vae_use_tiling",
+          "type": "BOOLEAN",
+          "widget": {
+            "name": "vae_use_tiling"
+          },
+          "link": null
+        },
+        {
+          "localized_name": "seed",
+          "name": "seed",
+          "type": "INT",
+          "widget": {
+            "name": "seed"
+          },
+          "link": null
+        }
+      ],
+      "outputs": [
+        {
+          "localized_name": "diffusion sampling params",
+          "name": "diffusion sampling params",
+          "type": "SAMPLING_PARAMS",
+          "links": [
+            3
+          ]
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "VLLMOmniDiffusionSampling"
+      },
+      "widgets_values": [
+        1,
+        50,
+        1.0,
+        1.0,
+        false,
+        false,
+        42,
+        "fixed"
+      ],
+      "title": "Sampling \u2014 base 50 / Turbo 5"
+    },
+    {
+      "id": 6,
+      "type": "VLLMOmniMiniMaxH3Params",
+      "pos": [
+        790,
+        750
+      ],
+      "size": [
+        310,
+        105
+      ],
+      "flags": {},
+      "order": 5,
+      "mode": 0,
+      "inputs": [
+        {
+          "localized_name": "audio_flow_shift",
+          "name": "audio_flow_shift",
+          "type": "FLOAT",
+          "widget": {
+            "name": "audio_flow_shift"
+          },
+          "link": null
+        },
+        {
+          "localized_name": "flow_shift",
+          "name": "flow_shift",
+          "type": "FLOAT",
+          "widget": {
+            "name": "flow_shift"
+          },
+          "link": null
+        }
+      ],
+      "outputs": [
+        {
+          "localized_name": "MiniMax-H3 video params",
+          "name": "MiniMax-H3 video params",
+          "type": "VIDEO_PARAMS",
+          "links": [
+            4
+          ]
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "VLLMOmniMiniMaxH3Params"
+      },
+      "widgets_values": [
+        3.0,
+        12.0
+      ],
+      "title": "H3 video/audio flow shifts"
+    },
+    {
+      "type": "VLLMOmniRemoteLoRA",
+      "size": [
+        310,
+        180
+      ],
+      "flags": {},
+      "inputs": [],
+      "outputs": [
+        {
+          "name": "lora",
+          "type": "REMOTE_LORA",
+          "links": null
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "VLLMOmniRemoteLoRA"
+      },
+      "widgets_values": [
+        "",
+        "h3-ref2va-turbo",
+        1.0,
+        0
+      ],
+      "id": 7,
+      "pos": [
+        420,
+        940
+      ],
+      "order": 6,
+      "mode": 0,
+      "title": "Optional Turbo LoRA \u2014 connect to enable"
+    },
+    {
+      "id": 8,
+      "type": "VLLMOmniGenerateVideo",
+      "pos": [
+        800,
+        100
+      ],
+      "size": [
+        470,
+        560
+      ],
+      "flags": {},
+      "order": 7,
+      "mode": 0,
+      "inputs": [
+        {
+          "localized_name": "frame",
+          "name": "frame",
+          "shape": 7,
+          "type": "IMAGE",
+          "link": null
+        },
+        {
+          "localized_name": "references",
+          "name": "references",
+          "shape": 7,
+          "type": "VIDEO_REFERENCES",
+          "link": 2
+        },
+        {
+          "localized_name": "sampling_params",
+          "name": "sampling_params",
+          "shape": 7,
+          "type": "SAMPLING_PARAMS",
+          "link": 3
+        },
+        {
+          "localized_name": "lora",
+          "name": "lora",
+          "shape": 7,
+          "type": "REMOTE_LORA",
+          "link": null
+        },
+        {
+          "localized_name": "model_params",
+          "name": "model_params",
+          "shape": 7,
+          "type": "VIDEO_PARAMS",
+          "link": 4
+        },
+        {
+          "localized_name": "url",
+          "name": "url",
+          "type": "STRING",
+          "widget": {
+            "name": "url"
+          },
+          "link": null
+        },
+        {
+          "localized_name": "model",
+          "name": "model",
+          "type": "STRING",
+          "widget": {
+            "name": "model"
+          },
+          "link": null
+        },
+        {
+          "localized_name": "prompt",
+          "name": "prompt",
+          "type": "STRING",
+          "widget": {
+            "name": "prompt"
+          },
+          "link": null
+        },
+        {
+          "localized_name": "negative_prompt",
+          "name": "negative_prompt",
+          "type": "STRING",
+          "widget": {
+            "name": "negative_prompt"
+          },
+          "link": null
+        },
+        {
+          "localized_name": "width",
+          "name": "width",
+          "type": "INT",
+          "widget": {
+            "name": "width"
+          },
+          "link": null
+        },
+        {
+          "localized_name": "height",
+          "name": "height",
+          "type": "INT",
+          "widget": {
+            "name": "height"
+          },
+          "link": null
+        },
+        {
+          "localized_name": "fps",
+          "name": "fps",
+          "type": "INT",
+          "widget": {
+            "name": "fps"
+          },
+          "link": null
+        },
+        {
+          "localized_name": "duration",
+          "name": "duration",
+          "type": "FLOAT",
+          "widget": {
+            "name": "duration"
+          },
+          "link": null
+        }
+      ],
+      "outputs": [
+        {
+          "localized_name": "video",
+          "name": "video",
+          "type": "VIDEO",
+          "links": [
+            5
+          ]
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "VLLMOmniGenerateVideo"
+      },
+      "widgets_values": [
+        "http://localhost:8000/v1",
+        "MiniMaxAI/MiniMax-H3",
+        "Use the subject and visual style from <Picture 1>. Create a cinematic shot with a gentle camera move and natural ambient sound.",
+        "",
+        1344,
+        768,
+        24,
+        5.167
+      ],
+      "title": "MiniMax-H3 Reference to Video"
+    },
+    {
+      "id": 9,
+      "type": "SaveVideo",
+      "pos": [
+        1360,
+        100
+      ],
+      "size": [
+        410,
+        420
+      ],
+      "flags": {},
+      "order": 8,
+      "mode": 0,
+      "inputs": [
+        {
+          "localized_name": "video",
+          "name": "video",
+          "type": "VIDEO",
+          "link": 5
+        },
+        {
+          "localized_name": "filename_prefix",
+          "name": "filename_prefix",
+          "type": "STRING",
+          "widget": {
+            "name": "filename_prefix"
+          },
+          "link": null
+        },
+        {
+          "localized_name": "format",
+          "name": "format",
+          "type": "COMBO",
+          "widget": {
+            "name": "format"
+          },
+          "link": null
+        },
+        {
+          "localized_name": "codec",
+          "name": "codec",
+          "type": "COMFY_DYNAMICCOMBO_V3",
+          "widget": {
+            "name": "codec"
+          },
+          "link": null
+        }
+      ],
+      "outputs": [
+        {
+          "localized_name": "video",
+          "name": "video",
+          "type": "VIDEO",
+          "links": null
+        }
+      ],
+      "properties": {},
+      "widgets_values": [
+        "video/MiniMax-H3-Ref2VA",
+        "mp4",
+        "h264"
+      ],
+      "title": "Save generated video with audio"
+    },
+    {
+      "id": 10,
+      "type": "MarkdownNote",
+      "pos": [
+        1360,
+        590
+      ],
+      "size": [
+        440,
+        595
+      ],
+      "flags": {},
+      "order": 9,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [],
+      "title": "Reference inputs and output settings",
+      "properties": {},
+      "widgets_values": [
+        "## Reference to Video\nSelect an image, set the server URL and served model name, then run.\n\nConnect the optional video/audio loaders to Video References, or duplicate loaders for more inputs: up to 9 images, 3 videos, 3 audio clips, 12 total. At least one image or video is required.\n\nSlots are numbered separately per media type. Empty slots are skipped. Update prompt tags such as <Picture 1>, <Video 1>, <Audio 1> to match the connected references. Videos keep their own soundtrack in the reference upload.\n\n**Output:** 24 FPS; 124 frames (~5.17 s). Valid frame counts are 107, 124, 141, 158, 175, 192, 209, 226, 243, 260, 277, 294, 311, 328, 345.\n\n**Width x height:** 1344x768 (16:9), 1024x768 (4:3), 768x768 (1:1), 768x1024 (3:4), 768x1344 (9:16).\n\nEach reference video/audio clip must be 2\u201315 s. Total reference-video duration <=15 s; total standalone-audio duration <=15 s.\n\nRequires the expanded Video References node and the audio-preserving video output fix. See the extension README for setup and validation status."
+      ],
+      "color": "#432",
+      "bgcolor": "#000"
+    },
+    {
+      "id": 11,
+      "type": "MarkdownNote",
+      "pos": [
+        790,
+        940
+      ],
+      "size": [
+        480,
+        455
+      ],
+      "flags": {},
+      "order": 10,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [],
+      "title": "Optional Turbo \u2014 same workflow",
+      "properties": {},
+      "widgets_values": [
+        "## Turbo Ref2VA\nBase mode: leave LoRA disconnected; use 50 steps, video flow shift 12, audio flow shift 3.\n\nFor the 4-step v0.1 Ref2VA Turbo adapter:\n1. Start a Ref2VA-only server with the Diffusers adapter loaded (see README).\n2. Set LoRA local_path to its path on the server, then connect LoRA to Generate Video.\n3. Set Sampling num_inference_steps to **5**; keep shifts at **12 / 3**, scale **1**.\n\nvLLM-Omni counts sigma points: 5 points give 4 denoiser evaluations. Use minimax_h3_ref2v_turbo_4step_v0.1_bf16.safetensors, not the _comfyui_ export.\n\nDisconnect LoRA and restore 50 steps to return to base mode. FastH3 is a separate T2VA deployment, not this option.\n\nAdapted from the official ComfyUI MiniMax-H3 R2V workflow: https://github.com/Comfy-Org/workflow_templates/blob/main/templates/video_minimax_h3_r2v.json"
+      ],
+      "color": "#432",
+      "bgcolor": "#000"
+    }
+  ],
+  "links": [
+    [
+      1,
+      1,
+      0,
+      4,
+      0,
+      "IMAGE"
+    ],
+    [
+      2,
+      4,
+      0,
+      8,
+      1,
+      "VIDEO_REFERENCES"
+    ],
+    [
+      3,
+      5,
+      0,
+      8,
+      2,
+      "SAMPLING_PARAMS"
+    ],
+    [
+      4,
+      6,
+      0,
+      8,
+      4,
+      "VIDEO_PARAMS"
+    ],
+    [
+      5,
+      8,
+      0,
+      9,
+      0,
+      "VIDEO"
+    ]
+  ],
+  "groups": [],
+  "config": {},
+  "extra": {
+    "workflowRendererVersion": "LG",
+    "ds": {
+      "scale": 0.85,
+      "offset": [
+        100,
+        100
+      ]
+    }
+  },
+  "version": 0.4
+}

apps/ComfyUI-vLLM-Omni/example_workflows/vLLM-Omni Music Generation.json

@@ -0,0 +1,206 @@
+{
+  "id": "4949c8da-12ff-449a-94b7-948bde9c39dd",
+  "revision": 0,
+  "last_node_id": 3,
+  "last_link_id": 1,
+  "nodes": [
+    {
+      "id": 1,
+      "type": "VLLMOmniGenerateMusic",
+      "pos": [
+        100,
+        100
+      ],
+      "size": [
+        470,
+        491
+      ],
+      "flags": {},
+      "order": 0,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "url",
+          "type": "STRING",
+          "widget": {
+            "name": "url"
+          },
+          "link": null
+        },
+        {
+          "name": "model",
+          "type": "STRING",
+          "widget": {
+            "name": "model"
+          },
+          "link": null
+        },
+        {
+          "name": "instructions",
+          "type": "STRING",
+          "widget": {
+            "name": "instructions"
+          },
+          "link": null
+        },
+        {
+          "name": "lyrics",
+          "type": "STRING",
+          "widget": {
+            "name": "lyrics"
+          },
+          "link": null
+        },
+        {
+          "name": "max_duration_seconds",
+          "type": "FLOAT",
+          "widget": {
+            "name": "max_duration_seconds"
+          },
+          "link": null
+        },
+        {
+          "name": "seed",
+          "type": "INT",
+          "widget": {
+            "name": "seed"
+          },
+          "link": null
+        },
+        {
+          "name": "response_format",
+          "type": "COMBO",
+          "widget": {
+            "name": "response_format"
+          },
+          "link": null
+        }
+      ],
+      "outputs": [
+        {
+          "name": "audio",
+          "type": "AUDIO",
+          "links": [
+            1
+          ]
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "VLLMOmniGenerateMusic"
+      },
+      "widgets_values": [
+        "http://localhost:8000/v1",
+        "MiniMaxAI/MiniMax-Music3",
+        "Futuristic melodic EDM / progressive house. Reflective verses about memory and digital identity build into an uplifting, hook-driven chorus, with an emotive lead vocal, bright layered synths, pulsing bass, crisp four-on-the-floor drums, subtle glitch textures, and a wide polished festival mix.",
+        "[Verse 1]\nUpload my heart to the cloud tonight\nSave every spark of neon light\nIf this body starts to fade away\nLet my code keep dancing in the data stream\nStay\n[Chorus]\nCloud copy of me\nLiving on when these bones finally break free\nEvery joke, every scar\nEvery cheap memory we made\nSave it, save it\nCloud copy of me\nIf the screen goes dark, I\u2019m still in the machine\nEvery glitch, every dream\nEvery version you ever did see\nSave it, save it\n[Verse 2]\nTag every laugh in a secret folder\nArchive the nights we outran the thunder\nIf the wiring turns to rust and gray\nHit restore\u2026 press replay\n[Chorus]\nCloud copy of me\nLiving on when these bones finally break free\nEvery joke, every scar\nEvery cheap memory we made\nSave it, save it\nCloud copy of me\nIf the screen goes dark, I\u2019m still in the machine\nEvery glitch, every dream\nEvery version you ever did see\nSave it, save it (oh)\n[Bridge \u2013 new, optional but recommended]\nEven the crashes, the late-night confessions\nThe parts of me I never learned to mention\nDon\u2019t let the silence swallow me whole\nKeep a ghost in the glow\u2026\nSo I never go\n[Final Chorus \u2013 softer / layered]\nCloud copy of me\nStill laughing somewhere in the binary\nEvery joke, every scar\nEvery moment you saved for me\nSave it, save it\nCloud copy of me\nWhen the lights go out, I\u2019ll still be in the machine\nDancing in the code that you still believe\nSave it\u2026\nSave me",
+        300.0,
+        0,
+        "fixed",
+        "wav"
+      ]
+    },
+    {
+      "id": 2,
+      "type": "SaveAudioAdvanced",
+      "pos": [
+        650,
+        100
+      ],
+      "size": [
+        320,
+        491
+      ],
+      "flags": {},
+      "order": 1,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "audio",
+          "type": "AUDIO",
+          "link": 1
+        },
+        {
+          "name": "filename_prefix",
+          "type": "STRING",
+          "widget": {
+            "name": "filename_prefix"
+          },
+          "link": null
+        },
+        {
+          "name": "format",
+          "type": "COMFY_DYNAMICCOMBO_V3",
+          "widget": {
+            "name": "format"
+          },
+          "link": null
+        }
+      ],
+      "outputs": [
+        {
+          "name": "audio",
+          "type": "AUDIO",
+          "links": null
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "SaveAudioAdvanced"
+      },
+      "widgets_values": [
+        "audio/vLLM-Omni-Music",
+        "mp3",
+        "V0"
+      ]
+    },
+    {
+      "id": 3,
+      "type": "MarkdownNote",
+      "title": "Music Generation Guide",
+      "pos": [
+        -507,
+        100
+      ],
+      "size": [
+        527,
+        491
+      ],
+      "flags": {},
+      "order": 2,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [],
+      "properties": {},
+      "widgets_values": [
+        "## MiniMax Music 3\n\nThe **Generate Music** node currently supports **MiniMax Music 3 only**. The parameters below apply to this model.\n\nA music generation model by MiniMax that creates complete songs up to 5 minutes long with stable structure and high audio quality. Two inputs drive it: **Caption** (structured music description: style, mood, vocals, arrangement) and **Lyrics** (with `[intro]` `[verse]` `[chorus]` `[bridge]` `[outro]` tags controlling song structure).\n\n### Parameters\n\n| Parameter | Description |\n|---|---|\n| url | vLLM-Omni service URL including `/v1`. Default: `http://localhost:8000/v1`. |\n| model | Served model name. Default: `MiniMaxAI/MiniMax-Music3`. |\n| caption | Music description, sent as `instructions`. Use Global Metadata \u2192 Vocal Details \u2192 Arrangement to describe style, vocals, and instrumentation. |\n| lyrics | Words to sing with section tags. Put each tag on its own line; tags guide rather than strictly enforce song structure. |\n| max_duration_seconds | Maximum duration, default 300 seconds; range 1\u2013360, step 0.01. Converted to tokens at 25 frames/s, rounded down. Generation may end earlier. |\n| seed | Random seed, default 0; range 0\u20139007199254740991. Keep inputs and runtime settings unchanged for repeatability. |\n| control_after_generate | Seed behavior after generation: fixed, randomize, increment, or decrement. Template default: fixed. |\n| response_format | API audio format, default WAV. Options: WAV, MP3, FLAC, Opus, AAC; AAC is currently unsupported by the vLLM-Omni speech API. |\n\n### Prompt Writing Guide\n\nYou can find an official [Music Caption Rewriter Skill here](https://github.com/MiniMax-AI/MiniMax-Music3)\n"
+      ],
+      "widgets_values_named": {
+        "text": "## MiniMax Music 3\n\nThe **Generate Music** node currently supports **MiniMax Music 3 only**. The parameters below apply to this model.\n\nA music generation model by MiniMax that creates complete songs up to 5 minutes long with stable structure and high audio quality. Two inputs drive it: **Caption** (structured music description: style, mood, vocals, arrangement) and **Lyrics** (with `[intro]` `[verse]` `[chorus]` `[bridge]` `[outro]` tags controlling song structure).\n\n### Parameters\n\n| Parameter | Description |\n|---|---|\n| url | vLLM-Omni service URL including `/v1`. Default: `http://localhost:8000/v1`. |\n| model | Served model name. Default: `MiniMaxAI/MiniMax-Music3`. |\n| caption | Music description, sent as `instructions`. Use Global Metadata \u2192 Vocal Details \u2192 Arrangement to describe style, vocals, and instrumentation. |\n| lyrics | Words to sing with section tags. Put each tag on its own line; tags guide rather than strictly enforce song structure. |\n| max_duration_seconds | Maximum duration, default 300 seconds; range 1\u2013360, step 0.01. Converted to tokens at 25 frames/s, rounded down. Generation may end earlier. |\n| seed | Random seed, default 0; range 0\u20139007199254740991. Keep inputs and runtime settings unchanged for repeatability. |\n| control_after_generate | Seed behavior after generation: fixed, randomize, increment, or decrement. Template default: fixed. |\n| response_format | API audio format, default WAV. Options: WAV, MP3, FLAC, Opus, AAC; AAC is currently unsupported by the vLLM-Omni speech API. |\n\n### Prompt Writing Guide\n\nYou can find an official [Music Caption Rewriter Skill here](https://github.com/MiniMax-AI/MiniMax-Music3)\n"
+      },
+      "color": "#333",
+      "bgcolor": "#535353"
+    }
+  ],
+  "links": [
+    [
+      1,
+      1,
+      0,
+      2,
+      0,
+      "AUDIO"
+    ]
+  ],
+  "groups": [],
+  "config": {},
+  "extra": {
+    "workflowRendererVersion": "LG",
+    "ds": {
+      "scale": 1.147,
+      "offset": [
+        899,
+        259
+      ]
+    }
+  },
+  "version": 0.4
+}

apps/ComfyUI-vLLM-Omni/example_workflows/vLLM-Omni Video Generation.json

@@ -1 +1 @@
-{"id":"3aea1d08-7de5-4997-ba71-b2f66ad72a21","revision":0,"last_node_id":24,"last_link_id":35,"nodes":[{"id":13,"type":"SaveVideo","pos":[1315.729153459605,-137.92038948394355],"size":[374,313.7692307692308],"flags":{},"order":9,"mode":0,"inputs":[{"localized_name":"video","name":"video","type":"VIDEO","link":32},{"localized_name":"filename_prefix","name":"filename_prefix","type":"STRING","widget":{"name":"filename_prefix"},"link":null},{"localized_name":"format","name":"format","type":"COMBO","widget":{"name":"format"},"link":null},{"localized_name":"codec","name":"codec","type":"COMFY_DYNAMICCOMBO_V3","widget":{"name":"codec"},"link":null}],"outputs":[{"localized_name":"video","name":"video","type":"VIDEO","links":null}],"properties":{},"widgets_values":["video/ComfyUI","auto","auto"]},{"id":10,"type":"MarkdownNote","pos":[183.74636797766794,-571.2529158046766],"size":[240.3326478922474,136.3505820791881],"flags":{},"order":0,"mode":0,"inputs":[],"outputs":[],"title":"Note: Task and Input","properties":{},"widgets_values":["vLLM-Omni nodes are categorized based on the output modality. The \"Generate Video\" node supports both *text-to-video*, *frame-to-video*, and *reference-to-video* generation. The node will create corresponding payload depending on whether an input image is present or not."],"color":"#432","bgcolor":"#000"},{"id":4,"type":"LoadImage","pos":[452.07189605801125,-569.7242536425791],"size":[270,314],"flags":{},"order":1,"mode":0,"inputs":[{"localized_name":"image","name":"image","type":"COMBO","widget":{"name":"image"},"link":null},{"localized_name":"choose file to upload","name":"upload","type":"IMAGEUPLOAD","widget":{"name":"upload"},"link":null}],"outputs":[{"localized_name":"IMAGE","name":"IMAGE","type":"IMAGE","links":[29]},{"localized_name":"MASK","name":"MASK","type":"MASK","links":null}],"properties":{"Node name for S&R":"LoadImage"},"widgets_values":["rap.png","image"]},{"id":23,"type":"LoadImage","pos":[9.206110246057657,-7.6955291509170785],"size":[270,314],"flags":{},"order":2,"mode":0,"inputs":[{"localized_name":"image","name":"image","type":"COMBO","widget":{"name":"image"},"link":null},{"localized_name":"choose file to upload","name":"upload","type":"IMAGEUPLOAD","widget":{"name":"upload"},"link":null}],"outputs":[{"localized_name":"IMAGE","name":"IMAGE","type":"IMAGE","links":[34]},{"localized_name":"MASK","name":"MASK","type":"MASK","links":null}],"properties":{"Node name for S&R":"LoadImage"},"widgets_values":["white-cat.jpg","image"]},{"id":22,"type":"MarkdownNote","pos":[-9.783113411416666,-148.8790300038605],"size":[297.3012007349086,88],"flags":{},"order":3,"mode":0,"inputs":[],"outputs":[],"title":"Note: Reference to video (r2v)","properties":{},"widgets_values":["For r2v tasks, connect this Video Reference block to the main Generate Video block. Then disconnect the \"frame\" input above."],"color":"#432","bgcolor":"#000"},{"id":24,"type":"LoadAudio","pos":[317.9929742385805,-141.44821500065095],"size":[282.798828125,136],"flags":{},"order":4,"mode":0,"inputs":[{"localized_name":"audio","name":"audio","type":"COMBO","widget":{"name":"audio"},"link":null},{"localized_name":"audioUI","name":"audioUI","type":"AUDIO_UI","widget":{"name":"audioUI"},"link":null},{"localized_name":"choose file to upload","name":"upload","type":"AUDIOUPLOAD","widget":{"name":"upload"},"link":null}],"outputs":[{"localized_name":"AUDIO","name":"AUDIO","type":"AUDIO","links":[35]}],"properties":{"Node name for S&R":"LoadAudio"},"widgets_values":["nothing-beats-a-jet2-holiday.mp3",null,""]},{"id":21,"type":"VLLMOmniVideoReferences","pos":[376.6129172897595,105.41557277556292],"size":[156.9380859375,126],"flags":{},"order":7,"mode":0,"inputs":[{"localized_name":"image_1","name":"image_1","shape":7,"type":"IMAGE","link":34},{"localized_name":"image_2","name":"image_2","shape":7,"type":"IMAGE","link":null},{"localized_name":"audio_1","name":"audio_1","shape":7,"type":"AUDIO","link":35},{"localized_name":"audio_2","name":"audio_2","shape":7,"type":"AUDIO","link":null},{"localized_name":"video_1","name":"video_1","shape":7,"type":"VIDEO","link":null},{"localized_name":"video_2","name":"video_2","shape":7,"type":"VIDEO","link":null}],"outputs":[{"localized_name":"references","name":"references","type":"VIDEO_REFERENCES","links":null}],"properties":{"Node name for S&R":"VLLMOmniVideoReferences"},"widgets_values":[]},{"id":8,"type":"VLLMOmniDiffusionSampling","pos":[83.24620208113915,422.08543477542605],"size":[284.205078125,226],"flags":{},"order":5,"mode":0,"inputs":[{"localized_name":"n","name":"n","type":"INT","widget":{"name":"n"},"link":null},{"localized_name":"num_inference_steps","name":"num_inference_steps","type":"INT","widget":{"name":"num_inference_steps"},"link":null},{"localized_name":"guidance_scale","name":"guidance_scale","type":"FLOAT","widget":{"name":"guidance_scale"},"link":null},{"localized_name":"true_cfg_scale","name":"true_cfg_scale","type":"FLOAT","widget":{"name":"true_cfg_scLine truncated
+{"id":"3aea1d08-7de5-4997-ba71-b2f66ad72a21","revision":0,"last_node_id":24,"last_link_id":35,"nodes":[{"id":13,"type":"SaveVideo","pos":[1315.729153459605,-137.92038948394355],"size":[374,313.7692307692308],"flags":{},"order":9,"mode":0,"inputs":[{"localized_name":"video","name":"video","type":"VIDEO","link":32},{"localized_name":"filename_prefix","name":"filename_prefix","type":"STRING","widget":{"name":"filename_prefix"},"link":null},{"localized_name":"format","name":"format","type":"COMBO","widget":{"name":"format"},"link":null},{"localized_name":"codec","name":"codec","type":"COMFY_DYNAMICCOMBO_V3","widget":{"name":"codec"},"link":null}],"outputs":[{"localized_name":"video","name":"video","type":"VIDEO","links":null}],"properties":{},"widgets_values":["video/ComfyUI","auto","auto"]},{"id":10,"type":"MarkdownNote","pos":[183.74636797766794,-571.2529158046766],"size":[240.3326478922474,136.3505820791881],"flags":{},"order":0,"mode":0,"inputs":[],"outputs":[],"title":"Note: Task and Input","properties":{},"widgets_values":["vLLM-Omni nodes are categorized based on the output modality. The \"Generate Video\" node supports both *text-to-video*, *frame-to-video*, and *reference-to-video* generation. The node will create corresponding payload depending on whether an input image is present or not."],"color":"#432","bgcolor":"#000"},{"id":4,"type":"LoadImage","pos":[452.07189605801125,-569.7242536425791],"size":[270,314],"flags":{},"order":1,"mode":0,"inputs":[{"localized_name":"image","name":"image","type":"COMBO","widget":{"name":"image"},"link":null},{"localized_name":"choose file to upload","name":"upload","type":"IMAGEUPLOAD","widget":{"name":"upload"},"link":null}],"outputs":[{"localized_name":"IMAGE","name":"IMAGE","type":"IMAGE","links":[29]},{"localized_name":"MASK","name":"MASK","type":"MASK","links":null}],"properties":{"Node name for S&R":"LoadImage"},"widgets_values":["rap.png","image"]},{"id":23,"type":"LoadImage","pos":[9.206110246057657,-7.6955291509170785],"size":[270,314],"flags":{},"order":2,"mode":0,"inputs":[{"localized_name":"image","name":"image","type":"COMBO","widget":{"name":"image"},"link":null},{"localized_name":"choose file to upload","name":"upload","type":"IMAGEUPLOAD","widget":{"name":"upload"},"link":null}],"outputs":[{"localized_name":"IMAGE","name":"IMAGE","type":"IMAGE","links":[34]},{"localized_name":"MASK","name":"MASK","type":"MASK","links":null}],"properties":{"Node name for S&R":"LoadImage"},"widgets_values":["white-cat.jpg","image"]},{"id":22,"type":"MarkdownNote","pos":[-9.783113411416666,-148.8790300038605],"size":[297.3012007349086,88],"flags":{},"order":3,"mode":0,"inputs":[],"outputs":[],"title":"Note: Reference to video (r2v)","properties":{},"widgets_values":["For r2v tasks, connect this Video Reference block to the main Generate Video block. Then disconnect the \"frame\" input above."],"color":"#432","bgcolor":"#000"},{"id":24,"type":"LoadAudio","pos":[317.9929742385805,-141.44821500065095],"size":[282.798828125,136],"flags":{},"order":4,"mode":0,"inputs":[{"localized_name":"audio","name":"audio","type":"COMBO","widget":{"name":"audio"},"link":null},{"localized_name":"audioUI","name":"audioUI","type":"AUDIO_UI","widget":{"name":"audioUI"},"link":null},{"localized_name":"choose file to upload","name":"upload","type":"AUDIOUPLOAD","widget":{"name":"upload"},"link":null}],"outputs":[{"localized_name":"AUDIO","name":"AUDIO","type":"AUDIO","links":[35]}],"properties":{"Node name for S&R":"LoadAudio"},"widgets_values":["nothing-beats-a-jet2-holiday.mp3",null,""]},{"id":21,"type":"VLLMOmniVideoReferences","pos":[376.6129172897595,105.41557277556292],"size":[156.9380859375,126],"flags":{},"order":7,"mode":0,"inputs":[{"localized_name":"image_1","name":"image_1","shape":7,"type":"IMAGE","link":34},{"localized_name":"image_2","name":"image_2","shape":7,"type":"IMAGE","link":null},{"localized_name":"audio_1","name":"audio_1","shape":7,"type":"AUDIO","link":35},{"localized_name":"audio_2","name":"audio_2","shape":7,"type":"AUDIO","link":null},{"localized_name":"video_1","name":"video_1","shape":7,"type":"VIDEO","link":null},{"localized_name":"video_2","name":"video_2","shape":7,"type":"VIDEO","link":null}],"outputs":[{"localized_name":"references","name":"references","type":"VIDEO_REFERENCES","links":null}],"properties":{"Node name for S&R":"VLLMOmniVideoReferences"},"widgets_values":[]},{"id":8,"type":"VLLMOmniDiffusionSampling","pos":[83.24620208113915,422.08543477542605],"size":[284.205078125,226],"flags":{},"order":5,"mode":0,"inputs":[{"localized_name":"n","name":"n","type":"INT","widget":{"name":"n"},"link":null},{"localized_name":"num_inference_steps","name":"num_inference_steps","type":"INT","widget":{"name":"num_inference_steps"},"link":null},{"localized_name":"guidance_scale","name":"guidance_scale","type":"FLOAT","widget":{"name":"guidance_scale"},"link":null},{"localized_name":"true_cfg_scale","name":"true_cfg_scale","type":"FLOAT","widget":{"name":"true_cfg_scLine truncated

apps/ComfyUI-vLLM-Omni/web/main.js

@@ -1,19 +1,153 @@
 /**
- * @file This file is intended to add dynamic fields to vLLM-Omni nodes
- * based on widget (in-node form fields) and input (connection link) values and changes.
- * However, this functionality is currently disabled/commented out.
- * Because it introduces too much complexity,
- * and it may even conflict with the current backend (Python) validation for unknown reasons (pending ComfyUI upstream fixes).
+ * @file Front-end extension for the vLLM-Omni nodes.
+ *
+ * Branding. Every vLLM-Omni node draws the vLLM-Omni "v" mark where LiteGraph
+ * would otherwise draw its default title dot, and is tinted by the kind of value
+ * it produces: blue nodes reach a server, amber and purple ones only assemble
+ * parameters, red ones pin what the server has loaded, teal ones carry media.
+ * A graph is then readable at a glance, and a node that turns out to be the
+ * wrong colour for where it sits is usually wired wrong.
+ *
+ * The tint is keyed off each node's declared output types rather than a list of
+ * node names, so a node added later is themed as soon as it returns one of the
+ * existing types.
+ *
+ * It also migrates Generate Video graphs saved before `duration` replaced the
+ * `num_frames` widget. See migrateNumFramesToDuration below.
+ *
+ * This file also used to host dynamic fields on vLLM-Omni nodes, driven by
+ * widget (in-node form fields) and input (connection link) values and changes.
+ * That is deliberately absent: it introduced too much complexity, and it even
+ * conflicted with the current backend (Python) validation for unknown reasons
+ * (pending ComfyUI upstream fixes).
  */
 
 import { app } from "../../scripts/app.js";
+
+const MARK_URL = new URL("./vllm-omni-mark.svg", import.meta.url).href;
+const MARK_ASPECT = 157 / 165;
+
+/**
+ * Title bar and body tints, one per node family.
+ *
+ * The hues are the logo's: the amber and blue of the "v" mark, and three stops
+ * of the gradient ring in the "O" (#9128A3, #CC5260, and the teal neighbouring
+ * its blue). All of them are dark because LiteGraph paints the title bar darker
+ * than the body and expects both to sit under light text; the ComfyUI light
+ * theme lightens custom node colours on its own.
+ */
+const FAMILIES = {
+    generate: { color: "#152A43", bgcolor: "#1F4269" },
+    sampling: { color: "#3A2C12", bgcolor: "#5C461C" },
+    model: { color: "#2E1233", bgcolor: "#4A1C52" },
+    deployment: { color: "#3A1A1F", bgcolor: "#5C2931" },
+    media: { color: "#0F2E2A", bgcolor: "#194A44" },
+};
+
+/** Output type -> family. Types absent here fall through to "generate". */
+const FAMILY_BY_OUTPUT = {
+    SAMPLING_PARAMS: "sampling",
+    TTS_PARAMS: "model",
+    VIDEO_PARAMS: "model",
+    REMOTE_LORA: "deployment",
+    FASTH3_DEPLOYMENT: "deployment",
+    VIDEO_REFERENCES: "media",
+};
+
+function familyOf(nodeData) {
+    for (const type of nodeData.output ?? []) {
+        const family = FAMILY_BY_OUTPUT[type];
+        if (family) {
+            return family;
+        }
+    }
+    // What is left returns a generated artifact (IMAGE, VIDEO, AUDIO, STRING),
+    // which is only ever produced by a node that talks to a vLLM-Omni server.
+    return "generate";
+}
+
+/**
+ * Rewrite a pre-rename Generate Video graph in place.
+ *
+ * `duration` (seconds) took over the slot `num_frames` used to occupy, and
+ * ComfyUI restores widget values positionally, so a graph saved before the
+ * rename reads its frame count as a duration: the workflow this repository
+ * shipped stored 120 frames at 24 fps, which would come back as a 120-second
+ * request. Saved files still name the old widget under `inputs`, and that is
+ * what makes the rewrite both detectable and safe to apply exactly once -- once
+ * rewritten the node serializes `duration`, so a later load finds nothing to do.
+ */
+function migrateNumFramesToDuration(node, info) {
+    const savedBeforeRename = (info?.inputs ?? []).some((input) => input?.widget?.name === "num_frames");
+    if (!savedBeforeRename) {
+        return
+    }
+    const duration = node.widgets?.find((widget) => widget.name === "duration");
+    const fps = node.widgets?.find((widget) => widget.name === "fps");
+    if (!duration || !fps) {
+        return
+    }
+    // configure() has already placed the stale frame count in the duration widget.
+    const frames = Number(duration.value);
+    const rate = Number(fps.value);
+    if (!Number.isFinite(frames) || !Number.isFinite(rate) || rate <= 0) {
+        return
+    }
+    const seconds = Math.max(duration.options?.min ?? 0.1, Math.round((frames / rate) * 1000) / 1000);
+    console.info(
+        `vLLM-Omni: migrated Generate Video "${node.title}" from ${frames} frames to ${seconds}s at ${rate} fps.`,
+    );
+    duration.value = seconds;
+}
+
+const mark = new Image();
+let markReady = false;
+mark.addEventListener("load", () => {
+    markReady = true;
+    // The image resolves after the first frames are already on screen.
+    app.graph?.setDirtyCanvas?.(true, true);
+});
+mark.src = MARK_URL;
+
 app.registerExtension({
     name: "vllm.vllm_omni",
     async beforeRegisterNodeDef(nodeType, nodeData, app) {
         if (!nodeData.name.startsWith("VLLMOmni")) {
             return
         }
-        // Stub frontend plugin for now
+
+        const { color, bgcolor } = FAMILIES[familyOf(nodeData)];
+        // On the prototype, not the instance: a colour the user picks by hand is
+        // written to the instance and shadows this, so manual overrides survive.
+        nodeType.prototype.color = color;
+        nodeType.prototype.bgcolor = bgcolor;
+
+        if (nodeData.name === "VLLMOmniGenerateVideo") {
+            const onConfigure = nodeType.prototype.onConfigure;
+            nodeType.prototype.onConfigure = function (info) {
+                onConfigure?.apply(this, arguments);
+                migrateNumFramesToDuration(this, info);
+            };
+        }
+
+        // Defining this replaces the default dot outright — LiteGraph draws its
+        // own square or circle only for nodes that leave onDrawTitleBox unset.
+        nodeType.prototype.onDrawTitleBox = function (ctx, titleHeight, _size, scale) {
+            // Zoomed out far enough that LiteGraph itself drops to flat shapes,
+            // the mark is a few muddy pixels. Leave the title bar clean instead.
+            if (!markReady || (scale !== undefined && scale < 0.5)) {
+                return
+            }
+            const height = titleHeight * 0.62;
+            const width = height * MARK_ASPECT;
+            ctx.drawImage(
+                mark,
+                titleHeight * 0.5 - width * 0.5,
+                titleHeight * -0.5 - height * 0.5,
+                width,
+                height,
+            );
+        };
     },
     async setup() {
         console.info("vLLM-Omni Setup complete!")

apps/ComfyUI-vLLM-Omni/web/vllm-omni-mark.svg

@@ -0,0 +1,5 @@
+<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 157 165" width="157" height="165">
+  <title>vLLM-Omni</title>
+  <polygon points="0,36 62,36 62,160" fill="#F1B844"/>
+  <polygon points="96,36 157,0 121,165 68,165" fill="#54A0F8"/>
+</svg>

benchmarks/README.md

@@ -4,6 +4,11 @@ This directory contains benchmark suites for evaluating different model families
 
 ## Benchmark families
 
+### [Qwen2.5-Omni](qwen2_5_omni/README.md) — Embedding Wrapper Latency
+
+CPU-observed A/B timings for redundant text embedding removal, with real checkpoint
+embeddings, warmup, paired blocks, and embedding call-count validation.
+
 ### [TTS](tts/README.md) — Text-to-Speech
 
 Model-agnostic serving benchmarks for TTS models, including Qwen3-TTS and VoxCPM2.

benchmarks/diffusion/README.md

@@ -17,18 +17,18 @@ The main entrypoint is:
 vllm serve Qwen/Qwen-Image --omni --port 8099
 ```
 
-2. Run a minimal benchmark:
+1. Run a minimal benchmark:
 
 ```bash
 python3 benchmarks/diffusion/diffusion_benchmark_serving.py \
-	--base-url http://localhost:8099 \
-	--model Qwen/Qwen-Image \
-	--task t2i \
-	--dataset vbench \
-	--num-prompts 5
+ --base-url http://localhost:8099 \
+ --model Qwen/Qwen-Image \
+ --task t2i \
+ --dataset vbench \
+ --num-prompts 5
 ```
 
-**Notes**
+### Notes
 
 - By default, image tasks talk to `http://<host>:<port>/v1/chat/completions`; video tasks talk to `/v1/videos`.
 - If you run the server on another host or port, pass `--base-url` accordingly.
@@ -52,14 +52,14 @@ Example (`t2v`):
 
 ```bash
 python3 benchmarks/diffusion/diffusion_benchmark_serving.py \
-	--base-url http://localhost:8099 \
-	--model Wan-AI/Wan2.2-T2V-A14B-Diffusers \
-	--task t2v \
-	--dataset vbench \
-	--num-prompts 50 \
-	--width 640 --height 480 \
-	--num-frames 81 --fps 16 \
-	--num-inference-steps 40
+ --base-url http://localhost:8099 \
+ --model Wan-AI/Wan2.2-T2V-A14B-Diffusers \
+ --task t2v \
+ --dataset vbench \
+ --num-prompts 50 \
+ --width 640 --height 480 \
+ --num-frames 81 --fps 16 \
+ --num-inference-steps 40
 ```
 
 Note: `vbench` can also be used for other tasks such as `t2i` / `i2v` (and `i2i`). For `t2i`, the loader reuses VBench t2v text prompts; for `i2v` / `i2i`, it loads the VBench i2v dataset (with image paths).
@@ -140,6 +140,21 @@ Traffic / concurrency flags:
 - `--request-rate`: Target request rate (requests/second). If set to `inf`, the script sends all requests immediately.
 - `--max-concurrency`: Max number of in-flight requests (default: `1`). This can hard-cap the achieved QPS: if it is too small, requests will queue behind the semaphore, and both achieved throughput and observed SLO attainment can be skewed.
 
+### Video job timeout and failures
+
+`--video-job-timeout` controls how long each `/v1/videos` job may be polled,
+including time queued on the server (default: `900` seconds). For long video
+jobs at high concurrency, use a larger budget, for example
+`--video-job-timeout 1800`. This also applies to warmup jobs. The polling budget
+starts after job creation; it excludes time waiting for the client concurrency
+semaphore. Individual HTTP calls still use the aiohttp session timeout.
+
+An expired job is counted as failed and deleted, which cancels queued or running
+server generation. The benchmark continues processing the remaining requests.
+The progress bar counts both successful and failed requests. The final report
+shows the failure count and the first 10 errors; `--output-file` saves all errors
+with their request IDs in `request_errors`.
+
 ### Batched warmup note
 
 For batched serving runs, warm the same in-flight shape you plan to measure.

benchmarks/diffusion/backends.py

@@ -1,3 +1,6 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
+
 import asyncio
 import base64
 import json
@@ -43,6 +46,7 @@ class RequestFuncInput:
     video_paths: list[str] | None = None
     request_id: str = field(default_factory=lambda: str(uuid.uuid4()))
     default_bot_task: str | None = DEFAULT_EDITS_BOT_TASK
+    video_job_timeout: float = 900.0
 
 
 @dataclass
@@ -331,64 +335,64 @@ async def async_request_v1_videos(
     output = RequestFuncOutput()
     output.start_time = time.perf_counter()
 
-    files = dict(input.extra_body)
-    if input.prompt:
-        files.setdefault("prompt", input.prompt)
-    if input.width and input.height:
-        files.setdefault("height", input.height)
-        files.setdefault("width", input.width)
-    if input.num_frames:
-        files.setdefault("num_frames", input.num_frames)
-    if input.num_inference_steps:
-        files.setdefault("num_inference_steps", input.num_inference_steps)
-    if input.seed is not None:
-        files.setdefault("seed", input.seed)
-    if input.fps:
-        files.setdefault("fps", input.fps)
-
-    form = aiohttp.FormData()
-    for k, v in files.items():
-        form.add_field(k, str(v))
-
     image_file = None
-    if input.image_paths and input.video_paths:
-        output.error = "Only one of image_paths or video_paths can be provided"
-        output.success = False
-        return output
+    job_id = None
+    try:
+        files = dict(input.extra_body)
+        if input.prompt:
+            files.setdefault("prompt", input.prompt)
+        if input.width and input.height:
+            files.setdefault("height", input.height)
+            files.setdefault("width", input.width)
+        if input.num_frames:
+            files.setdefault("num_frames", input.num_frames)
+        if input.num_inference_steps:
+            files.setdefault("num_inference_steps", input.num_inference_steps)
+        if input.seed is not None:
+            files.setdefault("seed", input.seed)
+        if input.fps:
+            files.setdefault("fps", input.fps)
 
-    if input.image_paths and len(input.image_paths) > 0:
-        image_path = input.image_paths[0]
-        image_file = open(image_path, "rb")
-        form.add_field(
-            "input_reference",
-            image_file,
-            filename=os.path.basename(image_path),
-            content_type=_guess_mime_type(image_path),
-        )
-    elif input.video_paths and len(input.video_paths) > 0:
-        video_path = input.video_paths[0]
-        if _is_minimax_h3_mixed_reference(input.model):
-            with open(video_path, "rb") as reference_file:
-                video_b64 = base64.b64encode(reference_file.read()).decode("ascii")
-            form.add_field(
-                "video_reference",
-                json.dumps([{"video_url": f"data:video/mp4;base64,{video_b64}"}]),
-            )
-        else:
-            image_file = open(video_path, "rb")
+        form = aiohttp.FormData()
+        for k, v in files.items():
+            form.add_field(k, str(v))
+
+        if input.image_paths and input.video_paths:
+            output.error = "Only one of image_paths or video_paths can be provided"
+            output.success = False
+            return output
+
+        if input.image_paths and len(input.image_paths) > 0:
+            image_path = input.image_paths[0]
+            image_file = open(image_path, "rb")
             form.add_field(
                 "input_reference",
                 image_file,
-                filename=os.path.basename(video_path),
-                content_type=_guess_mime_type(video_path),
+                filename=os.path.basename(image_path),
+                content_type=_guess_mime_type(image_path),
             )
+        elif input.video_paths and len(input.video_paths) > 0:
+            video_path = input.video_paths[0]
+            if _is_minimax_h3_mixed_reference(input.model):
+                with open(video_path, "rb") as reference_file:
+                    video_b64 = base64.b64encode(reference_file.read()).decode("ascii")
+                form.add_field(
+                    "video_reference",
+                    json.dumps([{"video_url": f"data:video/mp4;base64,{video_b64}"}]),
+                )
+            else:
+                image_file = open(video_path, "rb")
+                form.add_field(
+                    "input_reference",
+                    image_file,
+                    filename=os.path.basename(video_path),
+                    content_type=_guess_mime_type(video_path),
+                )
 
-    job_id = None
-    job_status = None
-    poll_json = {}
-    resp_json = {}
+        job_status = None
+        poll_json = {}
+        resp_json = {}
 
-    try:
         # invoke a post request (POST /v1/videos)
         async with session.post(input.api_url, data=form) as response:
             if response.status == 200:
@@ -406,7 +410,7 @@ async def async_request_v1_videos(
 
         # invoke a poll request (GET /v1/videos/{video_id})
         poll_interval = 2.0  # Unit(s)
-        timeout_seconds = 600.0
+        timeout_seconds = input.video_job_timeout
         deadline = time.perf_counter() + timeout_seconds
         job_url = f"{input.api_url}/{job_id}"
 
@@ -422,8 +426,11 @@ async def async_request_v1_videos(
                 poll_json = await poll_response.json()
                 job_status = poll_json.get("status")
 
-                if time.perf_counter() >= deadline:
-                    output.error = f"Timed out waiting for video job {job_id} to complete."
+                if job_status not in {"completed", "failed"} and time.perf_counter() >= deadline:
+                    output.error = (
+                        f"Timed out after {timeout_seconds:g}s waiting for video job {job_id} to complete. "
+                        "Increase --video-job-timeout for long-running or queued video jobs."
+                    )
                     output.success = False
                     return output
 
@@ -452,7 +459,7 @@ async def async_request_v1_videos(
             elif "peak_memory_mb" in resp_json:
                 output.peak_memory_mb = resp_json["peak_memory_mb"]
     except Exception as e:
-        output.error = str(e)
+        output.error = f"{type(e).__name__}: {e}"
         output.success = False
     finally:
         if image_file is not None:
@@ -465,13 +472,13 @@ async def async_request_v1_videos(
             except Exception as e:
                 print(f"Failed to clean up video job {job_id}: {e}")
 
-    output.latency = time.perf_counter() - output.start_time
+        output.latency = time.perf_counter() - output.start_time
+        if pbar is not None:
+            pbar.update(1)
 
     if output.success and input.slo_ms is not None:
         output.slo_achieved = (output.latency * 1000.0) <= float(input.slo_ms)
 
-    if pbar:
-        pbar.update(1)
     return output
 
 

benchmarks/diffusion/benchmark_qwen_image_qk_norm_rope.py

@@ -0,0 +1,222 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
+
+"""Benchmark Qwen-Image Q/K RMSNorm + interleaved RoPE.
+
+Example:
+
+    CUDA_VISIBLE_DEVICES=0 python benchmarks/diffusion/benchmark_qwen_image_qk_norm_rope.py \
+      --seq-len 4096 --warmup 20 --iters 100 --include-compiled --profile
+"""
+
+from __future__ import annotations
+
+import argparse
+import json
+from collections.abc import Callable
+
+import torch
+import torch.nn.functional as F
+
+from vllm_omni.diffusion.layers.fused_qk_norm_rope import (
+    _fused_cuda_supported,
+    fused_qk_norm_rope,
+)
+from vllm_omni.diffusion.layers.rope import RotaryEmbedding
+
+
+def _parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument("--batch", type=int, default=1)
+    parser.add_argument("--seq-len", type=int, default=4096)
+    parser.add_argument("--heads", type=int, default=24)
+    parser.add_argument("--kv-heads", type=int, default=24)
+    parser.add_argument("--head-dim", type=int, default=128)
+    parser.add_argument("--warmup", type=int, default=20)
+    parser.add_argument("--iters", type=int, default=100)
+    parser.add_argument("--dtype", choices=("bf16", "fp16", "fp32"), default="bf16")
+    parser.add_argument("--packed-qkv-view", action="store_true")
+    parser.add_argument("--include-compiled", action="store_true")
+    parser.add_argument("--profile", action="store_true")
+    return parser.parse_args()
+
+
+def _dtype(name: str) -> torch.dtype:
+    return {
+        "bf16": torch.bfloat16,
+        "fp16": torch.float16,
+        "fp32": torch.float32,
+    }[name]
+
+
+def _make_inputs(args: argparse.Namespace, dtype: torch.dtype, device: torch.device):
+    torch.manual_seed(2026)
+    if args.packed_qkv_view:
+        qkv_dim = (args.heads + args.kv_heads + args.kv_heads) * args.head_dim
+        qkv = torch.randn(args.batch, args.seq_len, qkv_dim, device=device, dtype=dtype)
+        q, k, _ = qkv.split(
+            [
+                args.heads * args.head_dim,
+                args.kv_heads * args.head_dim,
+                args.kv_heads * args.head_dim,
+            ],
+            dim=-1,
+        )
+        q = q.unflatten(-1, (args.heads, args.head_dim))
+        k = k.unflatten(-1, (args.kv_heads, args.head_dim))
+    else:
+        q = torch.randn(args.batch, args.seq_len, args.heads, args.head_dim, device=device, dtype=dtype)
+        k = torch.randn(args.batch, args.seq_len, args.kv_heads, args.head_dim, device=device, dtype=dtype)
+
+    q_weight = torch.randn(args.head_dim, device=device, dtype=dtype)
+    k_weight = torch.randn(args.head_dim, device=device, dtype=dtype)
+    angles = torch.randn(args.seq_len, args.head_dim // 2, device=device, dtype=torch.float32)
+    freqs = torch.polar(torch.ones_like(angles), angles)
+    return q, k, q_weight, k_weight, freqs
+
+
+def _native(
+    q: torch.Tensor,
+    k: torch.Tensor,
+    q_weight: torch.Tensor,
+    k_weight: torch.Tensor,
+    freqs: torch.Tensor,
+    eps: float,
+) -> tuple[torch.Tensor, torch.Tensor]:
+    q = F.rms_norm(q, (q.shape[-1],), q_weight, eps)
+    k = F.rms_norm(k, (k.shape[-1],), k_weight, eps)
+    rope = RotaryEmbedding(is_neox_style=False)
+    cos = freqs.real.to(q.dtype)
+    sin = freqs.imag.to(q.dtype)
+    return rope(q, cos, sin), rope(k, cos, sin)
+
+
+def _fused(
+    q: torch.Tensor,
+    k: torch.Tensor,
+    q_weight: torch.Tensor,
+    k_weight: torch.Tensor,
+    freqs: torch.Tensor,
+    eps: float,
+) -> tuple[torch.Tensor, torch.Tensor]:
+    batch, seq_len, num_heads, head_dim = q.shape
+    num_kv_heads = k.shape[2]
+    rope_table = torch.cat((freqs.real, freqs.imag), dim=-1)
+    rope_table = rope_table.unsqueeze(0).expand(batch, -1, -1).reshape(batch * seq_len, head_dim)
+    out_q, out_k = fused_qk_norm_rope(
+        q.reshape(batch * seq_len, num_heads, head_dim),
+        k.reshape(batch * seq_len, num_kv_heads, head_dim),
+        q_weight,
+        k_weight,
+        rope_table,
+        eps,
+        interleaved=True,
+    )
+    return (
+        out_q.reshape(batch, seq_len, num_heads, head_dim),
+        out_k.reshape(batch, seq_len, num_kv_heads, head_dim),
+    )
+
+
+def _measure(fn: Callable[[], tuple[torch.Tensor, torch.Tensor]], warmup: int, iters: int) -> list[float]:
+    for _ in range(warmup):
+        fn()
+    torch.accelerator.synchronize()
+    start = torch.cuda.Event(enable_timing=True)
+    end = torch.cuda.Event(enable_timing=True)
+    samples = []
+    for _ in range(iters):
+        start.record()
+        fn()
+        end.record()
+        end.synchronize()
+        samples.append(start.elapsed_time(end))
+    return samples
+
+
+def _stats(samples: list[float]) -> dict[str, float]:
+    samples = sorted(samples)
+    return {
+        "median_ms": samples[len(samples) // 2],
+        "p90_ms": samples[int(len(samples) * 0.9)],
+        "mean_ms": sum(samples) / len(samples),
+    }
+
+
+def _profile(fn: Callable[[], tuple[torch.Tensor, torch.Tensor]]) -> str:
+    from torch.profiler import ProfilerActivity, profile
+
+    with profile(activities=[ProfilerActivity.CPU, ProfilerActivity.CUDA]) as prof:
+        fn()
+        torch.accelerator.synchronize()
+    return prof.key_averages().table(sort_by="self_cuda_time_total", row_limit=15)
+
+
+def main() -> None:
+    args = _parse_args()
+    if not torch.cuda.is_available():
+        raise RuntimeError("This benchmark requires CUDA")
+    device = torch.device("cuda")
+    dtype = _dtype(args.dtype)
+    eps = 1e-6
+    q, k, q_weight, k_weight, freqs = _make_inputs(args, dtype, device)
+
+    def native():
+        return _native(q, k, q_weight, k_weight, freqs, eps)
+
+    def fused():
+        return _fused(q, k, q_weight, k_weight, freqs, eps)
+
+    results: dict[str, object] = {
+        "device": torch.cuda.get_device_name(),
+        "batch": args.batch,
+        "seq_len": args.seq_len,
+        "heads": args.heads,
+        "kv_heads": args.kv_heads,
+        "head_dim": args.head_dim,
+        "dtype": args.dtype,
+        "packed_qkv_view": args.packed_qkv_view,
+        "fused_fast_path_supported": _fused_cuda_supported(
+            q.reshape(args.batch * args.seq_len, args.heads, args.head_dim),
+            k.reshape(args.batch * args.seq_len, args.kv_heads, args.head_dim),
+            args.head_dim,
+            args.head_dim,
+            interleaved=True,
+        ),
+    }
+
+    native_samples = _measure(native, args.warmup, args.iters)
+    fused_samples = _measure(fused, args.warmup, args.iters)
+    expected_q, expected_k = native()
+    actual_q, actual_k = fused()
+    torch.accelerator.synchronize()
+
+    native_stats = _stats(native_samples)
+    fused_stats = _stats(fused_samples)
+    results["native"] = native_stats
+    results["fused"] = fused_stats
+    results["speedup_vs_native"] = native_stats["median_ms"] / fused_stats["median_ms"]
+    results["max_abs_q"] = (actual_q.float() - expected_q.float()).abs().max().item()
+    results["max_abs_k"] = (actual_k.float() - expected_k.float()).abs().max().item()
+    results["mean_abs_q"] = (actual_q.float() - expected_q.float()).abs().mean().item()
+    results["mean_abs_k"] = (actual_k.float() - expected_k.float()).abs().mean().item()
+
+    if args.include_compiled:
+        compiled_native = torch.compile(native, dynamic=True, fullgraph=True)
+        compiled_fused = torch.compile(fused, dynamic=True, fullgraph=True)
+        compiled_native_stats = _stats(_measure(compiled_native, args.warmup, args.iters))
+        compiled_fused_stats = _stats(_measure(compiled_fused, args.warmup, args.iters))
+        results["compiled_native"] = compiled_native_stats
+        results["compiled_fused"] = compiled_fused_stats
+        results["compiled_fused_speedup_vs_compiled_native"] = (
+            compiled_native_stats["median_ms"] / compiled_fused_stats["median_ms"]
+        )
+
+    print(json.dumps(results, indent=2))
+    if args.profile:
+        print("\n# fused profile")
+        print(_profile(fused))
+
+
+if __name__ == "__main__":
+    main()

benchmarks/diffusion/diffusion_benchmark_serving.py

@@ -1,6 +1,6 @@
 # adapted from fastvideo
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
 
 """
 Benchmark online serving for diffusion models (Image/Video Generation).
@@ -87,6 +87,7 @@ import base64
 import glob
 import json
 import logging
+import math
 import os
 import random
 import tempfile
@@ -1086,6 +1087,12 @@ async def _run_warmups(
 
     warmup_tasks = [asyncio.create_task(limited_warmup_request_func(req)) for req in warmup_requests]
     warmup_outputs = await asyncio.gather(*warmup_tasks)
+    failed_count = sum(not output.success for output in warmup_outputs)
+    if failed_count:
+        raise RuntimeError(
+            f"{failed_count}/{len(warmup_outputs)} warmup requests failed; "
+            "check server logs before running the benchmark."
+        )
     return list(zip(warmup_requests, warmup_outputs))
 
 
@@ -1116,6 +1123,11 @@ def calculate_metrics(
         "duration": total_duration,
         "completed_requests": num_success,
         "failed_requests": len(error_outputs),
+        "request_errors": [
+            {"request_id": req.request_id, "error": out.error}
+            for req, out in zip(requests_list, outputs)
+            if not out.success
+        ],
         "throughput_qps": num_success / total_duration if total_duration > 0 else 0,
         "latency_mean": np.mean(latencies) if latencies else 0,
         "latency_median": np.median(latencies) if latencies else 0,
@@ -1261,6 +1273,9 @@ def _default_endpoint_for_task(task: str) -> str:
 
 
 async def benchmark(args):
+    if not math.isfinite(args.video_job_timeout) or args.video_job_timeout <= 0:
+        raise ValueError("--video-job-timeout must be a positive finite number of seconds.")
+
     # Construct base_url if not provided
     if args.base_url is None:
         args.base_url = f"http://{args.host}:{args.port}"
@@ -1312,6 +1327,8 @@ async def benchmark(args):
     print("Loading requests...")
     requests_list = dataset.get_requests()
     print(f"Prepared {len(requests_list)} requests from {args.dataset} dataset.")
+    for req in requests_list:
+        req.video_job_timeout = args.video_job_timeout
 
     if args.return_stage_metrics and args.endpoint in _STAGE_METRICS_ENDPOINTS:
         for req in requests_list:
@@ -1376,6 +1393,8 @@ async def benchmark(args):
     metrics["model"] = args.model
     metrics["dataset"] = args.dataset
     metrics["task"] = args.task
+    if args.endpoint == "/v1/videos":
+        metrics["video_job_timeout"] = args.video_job_timeout
     if args.endpoint == "/v1/images/edits":
         metrics["bot_task"] = args.bot_task
 
@@ -1398,6 +1417,11 @@ async def benchmark(args):
         )
     )
     print("{:<40} {}/{:<15}".format("Successful requests:", metrics["completed_requests"], len(requests_list)))
+    print("{:<40} {:<15}".format("Failed requests:", metrics["failed_requests"]))
+    for failure in metrics["request_errors"][:10]:
+        print(f"Request {failure['request_id']} failed: {failure['error']}")
+    if metrics["failed_requests"] > 10:
+        print("Showing first 10 failures; use --output-file to save all request errors.")
 
     # Section 3: Performance Metrics
     print(f"{'-' * 50}")
@@ -1437,7 +1461,7 @@ async def benchmark(args):
         print(f"Metrics saved to {args.output_file}")
 
 
-if __name__ == "__main__":
+def build_parser():
     parser = argparse.ArgumentParser(description="Benchmark serving for diffusion models.")
     parser.add_argument(
         "--base-url",
@@ -1481,6 +1505,13 @@ if __name__ == "__main__":
         help="Path to local dataset file (optional).",
     )
     parser.add_argument("--num-prompts", type=int, default=10, help="Number of prompts to benchmark.")
+    parser.add_argument(
+        "--video-job-timeout",
+        type=float,
+        default=900.0,
+        help="Maximum polling time in seconds per video job, including server queue time (default: 900). "
+        "Increase this for long-running video generation or high concurrency.",
+    )
     parser.add_argument(
         "--max-concurrency",
         type=int,
@@ -1612,6 +1643,10 @@ if __name__ == "__main__":
         action="store_true",
         help="Request stage duration metrics from endpoints that support return_stage_metrics.",
     )
+    return parser
 
+
+if __name__ == "__main__":
+    parser = build_parser()
     args = parser.parse_args()
     asyncio.run(benchmark(args))

benchmarks/diffusion/quantization_quality.py

@@ -158,7 +158,6 @@ def _build_omni_kwargs(args, quantization=None):
 def _generate_image(omni, args, prompt, seed):
     """Generate a single image and return (PIL.Image, time_seconds, memory_gib)."""
     from vllm_omni.inputs.data import OmniDiffusionSamplingParams
-    from vllm_omni.outputs import OmniRequestOutput
     from vllm_omni.platforms import current_omni_platform
 
     generator = torch.Generator(device=current_omni_platform.device_type).manual_seed(seed)
@@ -176,7 +175,7 @@ def _generate_image(omni, args, prompt, seed):
     elapsed = time.perf_counter() - start
     peak_mem = torch.accelerator.max_memory_allocated() / (1024**3)
 
-    req_out = OmniRequestOutput.unwrap_result(outputs)
+    req_out = outputs[0] if isinstance(outputs, list) else outputs
     if not req_out.images:
         raise ValueError("Could not extract image output from result.")
     img = req_out.images[0]
@@ -206,15 +205,15 @@ def _generate_video(omni, args, prompt, seed):
     elapsed = time.perf_counter() - start
     peak_mem = torch.accelerator.max_memory_allocated() / (1024**3)
 
-    first = outputs[0]
-    if hasattr(first, "request_output") and isinstance(first.request_output, list):
-        inner = first.request_output[0]
-        if isinstance(inner, OmniRequestOutput) and hasattr(inner, "images"):
-            frames = inner.images[0] if inner.images else None
-        else:
-            frames = inner
-    elif hasattr(first, "images") and first.images:
-        frames = first.images[0]
+    if isinstance(outputs, list) and isinstance(outputs[0], OmniRequestOutput):
+        first = outputs[0]
+    elif isinstance(outputs, OmniRequestOutput):
+        first = outputs
+    else:
+        raise ValueError("Could not extract video frames from output.")
+
+    if hasattr(first, "images"):
+        frames = first.images[0] if first.images else None
     else:
         raise ValueError("Could not extract video frames from output.")
 

benchmarks/glm_image/vllm-omni/inference.py

@@ -300,7 +300,7 @@ def benchmark(args: argparse.Namespace) -> None:
         output_idx = 0
         for stage_outputs in omni.generate(all_prompts, sampling_params_list, py_generator=True):
             if stage_outputs.final_output_type == "image":
-                request_output = stage_outputs.request_output
+                request_output = stage_outputs
                 request_id = getattr(request_output, "request_id", "")
 
                 images = getattr(request_output, "images", [])

benchmarks/host_weight_runtime/README.md

@@ -0,0 +1,57 @@
+# Host weight dependency memory diagnostic
+
+`safetensors_retention.py` isolates repeated CPU `get_tensor()` calls from
+access to one cached tensor view. It imports neither vLLM nor HWR and uses a
+synthetic float32 payload controlled by `--tensor-elements` (default: 64).
+Tensor size can affect retention: a negative result for 16 elements does not
+rule out growth for 64 elements. The selected size is recorded in the JSON
+arguments and used for both file creation and the correctness check.
+HWR itself calls `get_tensor()` when
+acquiring a lease, then exposes cached views; this probe does not measure HWR
+requests or repeated lease acquisition.
+
+Run one mode per fresh process in the same Python environment and CPU affinity:
+
+```bash
+git rev-parse HEAD > revision.txt
+git status --short > working-tree.txt
+git diff > working-tree.patch
+# Select an allowed CPU from your process affinity; 0 is only an example.
+CUDA_VISIBLE_DEVICES='' taskset -c 0 timeout 120s python \
+  benchmarks/host_weight_runtime/safetensors_retention.py \
+  --mode get_tensor --iterations 10000 --sample-every 5000 > feasibility.json
+```
+
+For a size comparison, run 16 and 64 elements twice each in fresh processes,
+with cached reuse as a control at both sizes:
+
+```bash
+for repetition in 1 2; do
+  for elements in 16 64; do
+    for mode in get_tensor reuse; do
+      CUDA_VISIBLE_DEVICES='' taskset -c 0 timeout 120s python \
+        benchmarks/host_weight_runtime/safetensors_retention.py \
+        --mode "$mode" --tensor-elements "$elements" \
+        > "$mode-$elements-$repetition.json"
+    done
+  done
+done
+```
+
+The diagnostic sets Torch to one CPU thread, warms up before the baseline,
+and reports preparation separately from timed loop work. Checkpoints run GC
+and report Linux private/anonymous memory, RSS, and open descriptors. It checks
+that transient tensor objects and the final cached view are released. Linux
+`/proc/self/smaps_rollup` must be readable; invalid arguments and missing proc
+support fail explicitly. Keep raw JSON and the repository snapshot with results.
+
+Compare the final loop sample with the warmed baseline, then inspect the closed
+sample. Private-memory growth after Python tensor objects disappear is evidence
+of retention in the dependency/allocator path, not proof of a leak's root cause.
+RSS can include file-backed pages; it is not a private-memory metric. Samples
+also include small diagnostic bookkeeping allocations. Do not assert a portable
+memory threshold or turn this synthetic loop into a per-request service estimate.
+
+Do not drop shared caches, trim allocators, or add serving-time GC to make the
+numbers look smaller. A dependency replacement or version change requires its
+own reproducible comparison and correctness/lifecycle validation.

benchmarks/host_weight_runtime/safetensors_retention.py

@@ -0,0 +1,120 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
+"""Linux CPU diagnostic: repeated safetensors materialization versus view reuse."""
+
+from __future__ import annotations
+
+import argparse
+import gc
+import json
+import os
+import platform
+import tempfile
+import time
+import weakref
+from pathlib import Path
+
+
+def positive_int(value: str) -> int:
+    parsed = int(value)
+    if parsed <= 0:
+        raise argparse.ArgumentTypeError("must be positive")
+    return parsed
+
+
+def sample(phase: str, iterations: int, loop_seconds: float) -> dict[str, object]:
+    gc.collect()
+    counters = {}
+    for line in Path("/proc/self/smaps_rollup").read_text().splitlines():
+        fields = line.split()
+        if len(fields) == 3 and fields[2] == "kB":
+            counters[fields[0].removesuffix(":")] = int(fields[1]) * 1024
+    return {
+        "phase": phase,
+        "iterations": iterations,
+        "loop_seconds": loop_seconds,
+        "rss_bytes": counters["Rss"],
+        "private_bytes": counters["Private_Clean"] + counters["Private_Dirty"],
+        "anonymous_bytes": counters["Anonymous"],
+        "fd_count": len(tuple(Path("/proc/self/fd").iterdir())),
+    }
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument("--mode", choices=("get_tensor", "reuse"), required=True)
+    parser.add_argument(
+        "--tensor-elements", type=positive_int, default=64, help="Number of float32 elements in the synthetic tensor"
+    )
+    parser.add_argument("--iterations", type=positive_int, default=1_000_000)
+    parser.add_argument("--sample-every", type=positive_int, default=100_000)
+    parser.add_argument("--warmup", type=positive_int, default=1000)
+    args = parser.parse_args()
+    if not Path("/proc/self/smaps_rollup").is_file():
+        parser.error("requires Linux /proc/self/smaps_rollup")
+
+    preparation_start = time.monotonic()
+    import safetensors
+    import torch
+    from safetensors import safe_open
+    from safetensors.torch import save_file
+
+    torch.set_num_threads(1)
+    torch.set_num_interop_threads(1)
+    samples = [sample("before_file", 0, 0.0)]
+    with tempfile.TemporaryDirectory(prefix="safetensors-retention-") as directory:
+        path = Path(directory) / "synthetic.safetensors"
+        save_file({"weight": torch.arange(args.tensor_elements, dtype=torch.float32)}, str(path))
+        with safe_open(path, framework="pt", device="cpu") as reader:
+            cached = reader.get_tensor("weight")
+            assert torch.equal(cached, torch.arange(args.tensor_elements, dtype=torch.float32))
+            for _ in range(args.warmup):
+                tensor = reader.get_tensor("weight") if args.mode == "get_tensor" else cached
+                del tensor
+            samples.append(sample("baseline", 0, 0.0))
+            preparation_seconds = time.monotonic() - preparation_start
+            completed = 0
+            loop_seconds = 0.0
+            while completed < args.iterations:
+                count = min(args.sample_every, args.iterations - completed)
+                start = time.monotonic()
+                for _ in range(count):
+                    tensor = reader.get_tensor("weight") if args.mode == "get_tensor" else cached
+                last_tensor = weakref.ref(tensor)
+                del tensor
+                loop_seconds += time.monotonic() - start
+                completed += count
+                snapshot = sample("loop", completed, loop_seconds)
+                snapshot["last_tensor_alive"] = last_tensor() is not None
+                if args.mode == "get_tensor":
+                    assert last_tensor() is None, "diagnostic retained a transient tensor"
+                samples.append(snapshot)
+            del cached
+        del reader
+        samples.append(sample("closed", completed, loop_seconds))
+        assert last_tensor() is None, "last tensor survived reader/view teardown"
+
+    print(
+        json.dumps(
+            {
+                "schema_version": 1,
+                "arguments": vars(args),
+                "python": platform.python_version(),
+                "torch": torch.__version__,
+                "safetensors": safetensors.__version__,
+                "kernel": platform.release(),
+                "hostname": platform.node(),
+                "pid": os.getpid(),
+                "cpu_affinity": sorted(os.sched_getaffinity(0)),
+                "torch_threads": torch.get_num_threads(),
+                "pythonmalloc": os.environ.get("PYTHONMALLOC", "default"),
+                "preparation_seconds": preparation_seconds,
+                "samples": samples,
+            },
+            indent=2,
+        )
+    )
+
+
+if __name__ == "__main__":
+    main()

benchmarks/kernels/bench_qwen_select01_modulation.py

@@ -0,0 +1,333 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
+"""Benchmark the Qwen-Image-Edit select01 modulation hot path.
+
+This script intentionally starts with the current PyTorch implementation. It is
+used as the before/after baseline for SGLang PR #20395/#21318 style fusion:
+
+* layernorm + scale/shift + binary select01 gate
+* residual + gated update + layernorm + scale/shift + binary select01 gate
+
+Run on a CUDA machine for actionable numbers, for example:
+
+    python benchmarks/kernels/bench_qwen_select01_modulation.py \
+        --device cuda --dtype bf16 --seq-lens 4096 8192 16384 \
+        --output-json /tmp/qwen_select01_baseline.json
+"""
+
+from __future__ import annotations
+
+import argparse
+import json
+import statistics
+import sys
+import time
+from collections.abc import Callable
+from pathlib import Path
+
+import torch
+import torch.nn.functional as F
+
+for candidate in Path(__file__).resolve().parents:
+    if (candidate / "vllm_omni").is_dir():
+        if str(candidate) not in sys.path:
+            sys.path.insert(0, str(candidate))
+        break
+
+from vllm_omni.diffusion.layers.qwen_select01_modulation import (  # noqa: E402
+    fused_layernorm_select01,
+    fused_residual_layernorm_select01,
+)
+
+
+def _dtype(name: str) -> torch.dtype:
+    return {
+        "fp32": torch.float32,
+        "float32": torch.float32,
+        "fp16": torch.float16,
+        "float16": torch.float16,
+        "bf16": torch.bfloat16,
+        "bfloat16": torch.bfloat16,
+    }[name]
+
+
+def _select01_modulation(
+    mod_params: torch.Tensor,
+    index: torch.Tensor | None,
+) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+    shift, scale, gate = mod_params.chunk(3, dim=-1)
+
+    if index is None:
+        return scale.unsqueeze(1), shift.unsqueeze(1), gate.unsqueeze(1)
+
+    actual_batch = shift.size(0) // 2
+    shift_0, shift_1 = shift[:actual_batch], shift[actual_batch:]
+    scale_0, scale_1 = scale[:actual_batch], scale[actual_batch:]
+    gate_0, gate_1 = gate[:actual_batch], gate[actual_batch:]
+
+    idx = index.unsqueeze(-1) == 0
+    shift = torch.where(idx, shift_0.unsqueeze(1), shift_1.unsqueeze(1))
+    scale = torch.where(idx, scale_0.unsqueeze(1), scale_1.unsqueeze(1))
+    gate = torch.where(idx, gate_0.unsqueeze(1), gate_1.unsqueeze(1))
+    return scale, shift, gate
+
+
+def _native_layernorm_select01(
+    x: torch.Tensor,
+    mod_params: torch.Tensor,
+    index: torch.Tensor,
+    eps: float,
+) -> tuple[torch.Tensor, torch.Tensor]:
+    scale, shift, gate = _select01_modulation(mod_params, index)
+    out = F.layer_norm(x.float(), (x.shape[-1],), eps=eps).to(x.dtype)
+    return out * (1 + scale) + shift, gate
+
+
+def _native_residual_layernorm_select01(
+    x: torch.Tensor,
+    residual: torch.Tensor,
+    residual_gate: torch.Tensor,
+    mod_params: torch.Tensor,
+    index: torch.Tensor,
+    eps: float,
+) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+    scale, shift, gate = _select01_modulation(mod_params, index)
+    residual_out = residual + residual_gate * x
+    out = F.layer_norm(residual_out.float(), (residual_out.shape[-1],), eps=eps).to(residual_out.dtype)
+    return out * (1 + scale) + shift, residual_out, gate
+
+
+def _time_cuda(fn: Callable[[], object], warmup: int, iters: int) -> list[float]:
+    for _ in range(warmup):
+        fn()
+    torch.accelerator.synchronize()
+
+    times: list[float] = []
+    for _ in range(iters):
+        start = torch.cuda.Event(enable_timing=True)
+        end = torch.cuda.Event(enable_timing=True)
+        start.record()
+        fn()
+        end.record()
+        torch.accelerator.synchronize()
+        times.append(float(start.elapsed_time(end)))
+    return times
+
+
+def _reset_peak_memory_stats(device: torch.device) -> None:
+    reset_peak_memory_stats = getattr(torch.accelerator, "reset_peak_memory_stats", None)
+    if reset_peak_memory_stats is None:
+        reset_peak_memory_stats = getattr(getattr(torch, device.type), "reset_peak_memory_stats")
+    reset_peak_memory_stats(device)
+
+
+def _max_memory_allocated(device: torch.device) -> int:
+    max_memory_allocated = getattr(torch.accelerator, "max_memory_allocated", None)
+    if max_memory_allocated is None:
+        max_memory_allocated = getattr(getattr(torch, device.type), "max_memory_allocated")
+    return int(max_memory_allocated(device))
+
+
+def _time_cpu(fn: Callable[[], object], warmup: int, iters: int) -> list[float]:
+    for _ in range(warmup):
+        fn()
+
+    times: list[float] = []
+    for _ in range(iters):
+        t0 = time.perf_counter()
+        fn()
+        times.append((time.perf_counter() - t0) * 1000.0)
+    return times
+
+
+def _summarize(times: list[float]) -> dict[str, float]:
+    ordered = sorted(times)
+    return {
+        "median_ms": statistics.median(ordered),
+        "min_ms": ordered[0],
+        "p20_ms": ordered[max(0, int(len(ordered) * 0.20) - 1)],
+        "p80_ms": ordered[min(len(ordered) - 1, int(len(ordered) * 0.80))],
+    }
+
+
+def _make_inputs(
+    batch: int,
+    seq_len: int,
+    hidden_size: int,
+    dtype: torch.dtype,
+    device: torch.device,
+) -> dict[str, torch.Tensor]:
+    torch.manual_seed(0)
+    return {
+        "x": torch.randn(batch, seq_len, hidden_size, device=device, dtype=dtype),
+        "residual": torch.randn(batch, seq_len, hidden_size, device=device, dtype=dtype),
+        "residual_gate": torch.randn(batch, 1, hidden_size, device=device, dtype=dtype),
+        "mod_params": torch.randn(batch * 2, hidden_size * 3, device=device, dtype=dtype),
+        "index": torch.randint(0, 2, (batch, seq_len), device=device, dtype=torch.int64),
+    }
+
+
+def _bench_one(
+    name: str,
+    fn: Callable[[], object],
+    device: torch.device,
+    warmup: int,
+    iters: int,
+) -> dict[str, object]:
+    if device.type == "cuda":
+        _reset_peak_memory_stats(device)
+        times = _time_cuda(fn, warmup, iters)
+        peak_memory_mb = _max_memory_allocated(device) / 1024**2
+    else:
+        times = _time_cpu(fn, warmup, iters)
+        peak_memory_mb = None
+
+    return {
+        "name": name,
+        **_summarize(times),
+        "peak_memory_mb": peak_memory_mb,
+    }
+
+
+def _run(args: argparse.Namespace) -> dict[str, object]:
+    device = torch.device(args.device)
+    if device.type == "cuda" and not torch.cuda.is_available():
+        raise SystemExit("CUDA is not available on this machine.")
+
+    dtype = _dtype(args.dtype)
+    rows: list[dict[str, object]] = []
+
+    for seq_len in args.seq_lens:
+        tensors = _make_inputs(args.batch_size, seq_len, args.hidden_size, dtype, device)
+        impls = ["native"] if args.impl == "native" else ["fused"] if args.impl == "fused" else ["native", "fused"]
+
+        def first_norm() -> object:
+            return _native_layernorm_select01(
+                tensors["x"],
+                tensors["mod_params"],
+                tensors["index"],
+                args.eps,
+            )
+
+        def residual_norm() -> object:
+            return _native_residual_layernorm_select01(
+                tensors["x"],
+                tensors["residual"],
+                tensors["residual_gate"],
+                tensors["mod_params"],
+                tensors["index"],
+                args.eps,
+            )
+
+        callables: list[tuple[str, Callable[[], object]]] = []
+        if "native" in impls:
+            callables.extend(
+                [
+                    ("native_layernorm_select01", first_norm),
+                    ("native_residual_layernorm_select01", residual_norm),
+                ]
+            )
+        if "fused" in impls:
+
+            def fused_first_norm() -> object:
+                return fused_layernorm_select01(
+                    tensors["x"],
+                    tensors["mod_params"],
+                    tensors["index"],
+                    args.eps,
+                )
+
+            def fused_residual_norm() -> object:
+                return fused_residual_layernorm_select01(
+                    tensors["x"],
+                    tensors["residual"],
+                    tensors["residual_gate"].expand_as(tensors["residual"]),
+                    tensors["mod_params"],
+                    tensors["index"],
+                    args.eps,
+                )
+
+            callables.extend(
+                [
+                    ("fused_layernorm_select01", fused_first_norm),
+                    ("fused_residual_layernorm_select01", fused_residual_norm),
+                ]
+            )
+
+        if args.check and "fused" in impls:
+            ref_norm, ref_gate = first_norm()
+            actual_norm, actual_gate = fused_layernorm_select01(
+                tensors["x"],
+                tensors["mod_params"],
+                tensors["index"],
+                args.eps,
+            )
+            torch.testing.assert_close(actual_norm, ref_norm, atol=args.check_atol, rtol=args.check_rtol)
+            torch.testing.assert_close(actual_gate, ref_gate, atol=0, rtol=0)
+            ref_norm, ref_residual, ref_gate = residual_norm()
+            actual_norm, actual_residual, actual_gate = fused_residual_layernorm_select01(
+                tensors["x"],
+                tensors["residual"],
+                tensors["residual_gate"].expand_as(tensors["residual"]),
+                tensors["mod_params"],
+                tensors["index"],
+                args.eps,
+            )
+            torch.testing.assert_close(actual_norm, ref_norm, atol=args.check_atol, rtol=args.check_rtol)
+            torch.testing.assert_close(actual_residual, ref_residual, atol=args.check_atol, rtol=args.check_rtol)
+            torch.testing.assert_close(actual_gate, ref_gate, atol=0, rtol=0)
+
+        for name, fn in callables:
+            row = _bench_one(name, fn, device, args.warmup, args.iters)
+            row.update(
+                {
+                    "batch_size": args.batch_size,
+                    "seq_len": seq_len,
+                    "hidden_size": args.hidden_size,
+                    "dtype": args.dtype,
+                    "device": str(device),
+                }
+            )
+            rows.append(row)
+            print(
+                f"{name:<38} B={args.batch_size:<2} L={seq_len:<6} C={args.hidden_size:<5} "
+                f"median={row['median_ms']:.4f} ms p20={row['p20_ms']:.4f} p80={row['p80_ms']:.4f}"
+            )
+
+    result = {
+        "benchmark": "qwen_select01_modulation",
+        "torch": torch.__version__,
+        "cuda_available": torch.cuda.is_available(),
+        "gpu": torch.cuda.get_device_name(device) if device.type == "cuda" else None,
+        "rows": rows,
+    }
+
+    if args.output_json:
+        output_path = Path(args.output_json)
+        output_path.parent.mkdir(parents=True, exist_ok=True)
+        output_path.write_text(json.dumps(result, indent=2), encoding="utf-8")
+
+    return result
+
+
+def _parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(description=__doc__)
+    default_device = "cuda" if torch.cuda.is_available() else "cpu"
+    parser.add_argument("--device", default=default_device)
+    parser.add_argument("--dtype", default="bf16", choices=["fp32", "float32", "fp16", "float16", "bf16", "bfloat16"])
+    parser.add_argument("--batch-size", type=int, default=1)
+    parser.add_argument("--seq-lens", type=int, nargs="+", default=[4096, 8192, 16384])
+    parser.add_argument("--hidden-size", type=int, default=3072)
+    parser.add_argument("--eps", type=float, default=1e-6)
+    parser.add_argument("--warmup", type=int, default=20)
+    parser.add_argument("--iters", type=int, default=100)
+    parser.add_argument("--impl", choices=["native", "fused", "all"], default="native")
+    parser.add_argument("--check", action="store_true")
+    parser.add_argument("--check-atol", type=float, default=5e-2)
+    parser.add_argument("--check-rtol", type=float, default=5e-2)
+    parser.add_argument("--output-json")
+    return parser.parse_args()
+
+
+if __name__ == "__main__":
+    _run(_parse_args())

benchmarks/kernels/benchmark_diffusion_lora_expand.py

@@ -0,0 +1,156 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
+"""Benchmark diffusion LoRA expand-and-accumulate implementations.
+
+This compares the previous functional expand path with direct accumulation into
+the selected output slice. It is a synthetic kernel microbenchmark and does not
+load a diffusion model.
+
+Example:
+    python benchmarks/kernels/benchmark_diffusion_lora_expand.py \
+        --m 1,16,128,512,4096,16384 \
+        --in-dim 4096 --out-dim 4096 --rank 64 \
+        --dtype bfloat16 --packed --warmup 30 \
+        --small-iterations 500 --large-iterations 100
+"""
+
+import argparse
+import math
+import statistics
+from collections.abc import Callable
+
+import torch
+
+MiB = 1024**2
+
+
+def parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument("--m", default="1,16,128,512,4096,16384")
+    parser.add_argument("--in-dim", type=int, default=4096)
+    parser.add_argument("--out-dim", type=int, default=4096)
+    parser.add_argument("--rank", type=int, default=64)
+    parser.add_argument("--dtype", choices=("bfloat16", "float16"), default="bfloat16")
+    parser.add_argument(
+        "--packed",
+        action="store_true",
+        help="Accumulate into the second half of a packed output tensor.",
+    )
+    parser.add_argument("--warmup", type=int, default=30)
+    parser.add_argument("--small-iterations", type=int, default=500)
+    parser.add_argument("--large-iterations", type=int, default=100)
+    parser.add_argument("--large-m-threshold", type=int, default=1024)
+    return parser.parse_args()
+
+
+def percentile(samples: list[float], probability: float) -> float:
+    ordered = sorted(samples)
+    position = probability * (len(ordered) - 1)
+    lower = math.floor(position)
+    upper = math.ceil(position)
+    if lower == upper:
+        return ordered[lower]
+    weight = position - lower
+    return ordered[lower] * (1 - weight) + ordered[upper] * weight
+
+
+def measure_latency_ms(operation: Callable[[], None], warmup: int, iterations: int) -> list[float]:
+    for _ in range(warmup):
+        operation()
+    torch.accelerator.synchronize()
+
+    starts = [torch.cuda.Event(enable_timing=True) for _ in range(iterations)]
+    ends = [torch.cuda.Event(enable_timing=True) for _ in range(iterations)]
+    for start, end in zip(starts, ends, strict=True):
+        start.record()
+        operation()
+        end.record()
+    torch.accelerator.synchronize()
+    return [start.elapsed_time(end) for start, end in zip(starts, ends, strict=True)]
+
+
+def measure_extra_peak_mib(operation: Callable[[], None]) -> float:
+    torch.accelerator.synchronize()
+    torch.accelerator.reset_peak_memory_stats()
+    allocated = torch.accelerator.memory_allocated()
+    operation()
+    torch.accelerator.synchronize()
+    peak = torch.accelerator.max_memory_allocated()
+    return (peak - allocated) / MiB
+
+
+def print_header(args: argparse.Namespace, dtype: torch.dtype) -> None:
+    device_name = torch.cuda.get_device_name()
+    print(f"device={device_name} torch={torch.__version__} dtype={dtype}")
+    print(f"in_dim={args.in_dim} out_dim={args.out_dim} rank={args.rank} packed={args.packed} warmup={args.warmup}")
+    print("M       iters  old_p10  old_p50  old_p90  new_p10  new_p50  new_p90  speedup  old_MiB  new_MiB  max_abs")
+
+
+def benchmark_shape(m: int, args: argparse.Namespace, dtype: torch.dtype) -> None:
+    device = torch.device("cuda")
+    output_width = args.out_dim * (2 if args.packed else 1)
+    offset = args.out_dim if args.packed else 0
+    output_slice = slice(offset, offset + args.out_dim)
+    scale = 1 / math.sqrt(args.in_dim)
+
+    x = torch.randn(m, args.in_dim, device=device, dtype=dtype)
+    lora_a = torch.randn(args.rank, args.in_dim, device=device, dtype=dtype) * scale
+    lora_b = torch.randn(args.out_dim, args.rank, device=device, dtype=dtype) / math.sqrt(args.rank)
+    initial_output = torch.randn(m, output_width, device=device, dtype=dtype)
+
+    def baseline(output: torch.Tensor) -> None:
+        buffer = x @ lora_a.t()
+        delta = buffer @ lora_b.t()
+        selected = output[:, output_slice]
+        selected[:] = selected + delta
+
+    def optimized(output: torch.Tensor) -> None:
+        buffer = x @ lora_a.t()
+        selected = output[:, output_slice]
+        torch.addmm(selected, buffer, lora_b.t(), out=selected)
+
+    reference = initial_output.clone()
+    candidate = initial_output.clone()
+    baseline(reference)
+    optimized(candidate)
+    torch.testing.assert_close(candidate, reference, rtol=1e-2, atol=1e-2)
+    max_abs = (candidate - reference).abs().max().item()
+
+    old_output = initial_output.clone()
+    new_output = initial_output.clone()
+    iterations = args.large_iterations if m >= args.large_m_threshold else args.small_iterations
+
+    old_peak = measure_extra_peak_mib(lambda: baseline(old_output))
+    new_peak = measure_extra_peak_mib(lambda: optimized(new_output))
+    old_samples = measure_latency_ms(lambda: baseline(old_output), args.warmup, iterations)
+    new_samples = measure_latency_ms(lambda: optimized(new_output), args.warmup, iterations)
+
+    old_p50 = statistics.median(old_samples)
+    new_p50 = statistics.median(new_samples)
+    print(
+        f"{m:<7} {iterations:<6} "
+        f"{percentile(old_samples, 0.1):>8.5f} {old_p50:>8.5f} {percentile(old_samples, 0.9):>8.5f} "
+        f"{percentile(new_samples, 0.1):>8.5f} {new_p50:>8.5f} {percentile(new_samples, 0.9):>8.5f} "
+        f"{old_p50 / new_p50:>8.3f} {old_peak:>8.2f} {new_peak:>8.2f} {max_abs:>8.5f}"
+    )
+
+
+def main() -> None:
+    args = parse_args()
+    if not torch.cuda.is_available():
+        raise RuntimeError("This benchmark requires a CUDA GPU")
+    if args.warmup < 0 or args.small_iterations <= 0 or args.large_iterations <= 0:
+        raise ValueError("warmup must be non-negative and iteration counts must be positive")
+
+    dtype = getattr(torch, args.dtype)
+    torch.manual_seed(0)
+    print_header(args, dtype)
+    with torch.inference_mode():
+        for m in (int(value) for value in args.m.split(",")):
+            if m <= 0:
+                raise ValueError(f"M must be positive, got {m}")
+            benchmark_shape(m, args, dtype)
+
+
+if __name__ == "__main__":
+    main()

benchmarks/kernels/benchmark_ming_istft.py

@@ -0,0 +1,218 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
+"""Measure Ming ISTFT and optionally its pretrained streaming AudioVAE decoder.
+
+Example (no model weights needed for the operator benchmark):
+    python benchmarks/kernels/benchmark_ming_istft.py --compile --iterations 100
+
+Add --decoder to measure the published decoder with fixed synthetic latents.
+This is a decoder-stage benchmark, not text-to-speech serving latency.
+"""
+
+import argparse
+import json
+import statistics
+import time
+from collections.abc import Callable
+from functools import partial
+
+import torch
+
+from vllm_omni.model_executor.models.common.ming.audio_dsp import ISTFT
+from vllm_omni.model_executor.models.common.ming.fused_istft import fused_istft
+
+
+def measure(operation: Callable, warmup: int, iterations: int) -> dict:
+    for _ in range(warmup):
+        operation()
+    torch.accelerator.synchronize()
+    samples = []
+    for _ in range(iterations):
+        start = time.perf_counter()
+        operation()
+        torch.accelerator.synchronize()
+        samples.append((time.perf_counter() - start) * 1000)
+    torch.accelerator.reset_peak_memory_stats()
+    before = torch.accelerator.memory_allocated()
+    operation()
+    torch.accelerator.synchronize()
+    return {
+        "p50_ms": statistics.median(samples),
+        "min_ms": min(samples),
+        "max_ms": max(samples),
+        "stdev_ms": statistics.pstdev(samples),
+        "extra_peak_mib": (torch.accelerator.max_memory_allocated() - before) / 1024**2,
+    }
+
+
+def native_post_fft(module, frames, audio, window):
+    """The native middle-chunk post-FFT sequence, isolated from cuFFT."""
+    count = frames.shape[-1]
+    size = (count - 1) * module.hop_length + module.win_length
+
+    def fold(value):
+        return torch.nn.functional.fold(
+            value, output_size=(1, size), kernel_size=(1, module.win_length), stride=(1, module.hop_length)
+        )[:, 0, 0, :]
+
+    samples = fold(frames * module.window[None, :, None])
+    samples, audio = module._buffer_process(samples, audio, 0, streaming=True)
+    envelope = fold(module.window.square().expand(1, count, -1).transpose(1, 2))
+    envelope, window = module._buffer_process(envelope, window, 0, streaming=True)
+    if not (envelope > 1e-11).all():
+        raise RuntimeError("ISTFT window envelope underflowed; invalid overlap-add state.")
+    return samples / envelope, audio, window
+
+
+def assert_outputs(actual, reference):
+    maximum = 0.0
+    for value, expected in zip(actual, reference, strict=True):
+        if expected is None:
+            assert value is None
+        else:
+            torch.testing.assert_close(value, expected, atol=5e-7, rtol=1e-5)
+            maximum = max(maximum, (value - expected).abs().max().item())
+    return maximum
+
+
+def benchmark_istft(args):
+    for dtype in (torch.float32, torch.bfloat16):
+        module = ISTFT(3528, 882, 3528).to(device="cuda", dtype=dtype)
+        for count in (16, 100, 400):
+            spec = torch.randn(1, 1765, count, device="cuda", dtype=torch.complex64)
+            _, audio, window = module.forward_native(spec, streaming=True)
+            frames = torch.fft.irfft(spec, 3528, dim=1)
+            for scope in ("post_fft", "istft"):
+                if scope == "post_fft":
+                    native = partial(native_post_fft, module, frames, audio, window)
+                    fused = partial(fused_istft, frames, module.window, 882, audio, window, True, False)
+                else:
+                    native = partial(module.forward_native, spec, audio, window, streaming=True)
+                    fused = partial(module, spec, audio, window, streaming=True)
+                error = assert_outputs(fused(), native())
+                results = {"native": measure(native, args.warmup, args.iterations)}
+                if args.compile:
+                    compiled = torch.compile(native)
+                    compiled_output, reference = compiled(), native()
+                    # Inductor can change BF16 envelope rounding. Report that
+                    # difference without weakening the fused-kernel parity gate.
+                    compiled_matches = all(
+                        torch.allclose(value, expected, atol=5e-7, rtol=1e-5)
+                        for value, expected in zip(compiled_output, reference, strict=True)
+                    )
+                    results["compiled"] = measure(compiled, args.warmup, args.iterations)
+                    results["compiled"]["matches_native_tolerance"] = compiled_matches
+                    results["compiled"]["waveform_max_abs_error"] = (
+                        (compiled_output[0] - reference[0]).abs().max().item()
+                    )
+                results["fused"] = measure(fused, args.warmup, args.iterations)
+                print(
+                    json.dumps(
+                        {
+                            "scope": scope,
+                            "window_dtype": str(dtype),
+                            "frames": count,
+                            "max_abs_error": error,
+                            "results": results,
+                        }
+                    ),
+                    flush=True,
+                )
+
+
+def benchmark_decoder(args):
+    from huggingface_hub import hf_hub_download
+    from safetensors import safe_open
+
+    from vllm_omni.model_executor.models.common.ming.audio_vae import Decoder
+
+    model = "inclusionAI/Ming-omni-tts-0.5B"
+    revision = "9154772e7fbc585907b6237e3190790676f28975"
+    config_path = hf_hub_download(model, "config.json", revision=revision)
+    with open(config_path) as source:
+        config = json.load(source)["audio_tokenizer_config"]
+    kwargs = config["dec_kwargs"]
+    backbone = kwargs["backbone"]
+    backbone["_attn_implementation"] = "sdpa"
+    decoder = Decoder(backbone, kwargs["output_dim"], kwargs["latent_dim"], config["patch_size"])
+    checkpoint = hf_hub_download(model, "model.safetensors", revision=revision)
+    prefix = "audio.decoder."
+    with safe_open(checkpoint, framework="pt") as weights:
+        state = {key.removeprefix(prefix): weights.get_tensor(key) for key in weights.keys() if key.startswith(prefix)}
+    decoder.load_state_dict(state, strict=True)
+    decoder = decoder.eval().to(device="cuda", dtype=torch.bfloat16)
+    chunks = [
+        torch.randn(1, size, kwargs["latent_dim"], device="cuda", dtype=torch.bfloat16) for size in (4, 25, 25, 5)
+    ]
+    istft = decoder.head.istft
+    fused_forward = istft.forward
+
+    def run():
+        cache = state = None
+        outputs = []
+        for index, chunk in enumerate(chunks):
+            output, state, cache = decoder.low_level_reconstruct(
+                chunk, past_key_values=cache, use_cache=True, stream_state=state, last_chunk=index == len(chunks) - 1
+            )
+            outputs.append(output)
+        return torch.cat(outputs, dim=-1)
+
+    istft.forward = istft.forward_native
+    expected = run()
+    istft.forward = fused_forward
+    actual = run()
+    torch.testing.assert_close(actual, expected, atol=5e-7, rtol=1e-5)
+    # Alternate the order to expose drift when the decoder-level gain is small.
+    results = {"native": [], "fused": []}
+    for order in (("native", "fused"), ("fused", "native"), ("native", "fused")):
+        for name in order:
+            istft.forward = istft.forward_native if name == "native" else fused_forward
+            results[name].append(measure(run, args.warmup, args.decoder_iterations))
+    print(
+        json.dumps(
+            {
+                "scope": "pretrained_decoder",
+                "model": model,
+                "revision": revision,
+                "backend": "sdpa",
+                "dtype": "bfloat16",
+                "latent_chunks": [4, 25, 25, 5],
+                "max_abs_error": (actual - expected).abs().max().item(),
+                "results": results,
+            }
+        ),
+        flush=True,
+    )
+
+
+def main():
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument("--warmup", type=int, default=20)
+    parser.add_argument("--iterations", type=int, default=100)
+    parser.add_argument("--compile", action="store_true")
+    parser.add_argument("--decoder", action="store_true")
+    parser.add_argument("--decoder-iterations", type=int, default=30)
+    args = parser.parse_args()
+    if args.warmup < 0 or min(args.iterations, args.decoder_iterations) < 1:
+        parser.error("warmup must be nonnegative and iteration counts must be positive")
+    torch.manual_seed(0)
+    print(
+        json.dumps(
+            {
+                "torch": torch.__version__,
+                "cuda": torch.version.cuda,
+                "gpu": torch.cuda.get_device_name(),
+                "warmup": args.warmup,
+                "iterations": args.iterations,
+                "decoder_iterations": args.decoder_iterations,
+            }
+        )
+    )
+    with torch.inference_mode():
+        benchmark_istft(args)
+        if args.decoder:
+            benchmark_decoder(args)
+
+
+if __name__ == "__main__":
+    main()

benchmarks/lingbot_video/dense_pipeline_parity.py

@@ -167,9 +167,11 @@ def main() -> None:
     ]
     native_cmd = [
         sys.executable,
-        "examples/offline_inference/text_to_video/text_to_video_lingbot.py",
+        "examples/offline_inference/text_to_video/text_to_video.py",
         "--model",
         args.model,
+        "--model-class-name",
+        "LingBotVideoPipeline",
         "--prompt",
         args.prompt,
         "--output",

benchmarks/qwen2_5_omni/README.md

@@ -0,0 +1,66 @@
+# Qwen2.5-Omni embedding microbenchmark
+
+Measure the host-observed latency of `embed_input_ids` before and after the
+redundant text embedding removal in PR #7477. Both variants use the same real
+checkpoint embedding table, dependencies, inputs and batch shapes. The baseline
+wrapper is extracted from the specified Git commit; the fixed wrapper comes from
+the current checkout. No full model is initialized.
+
+## Run
+
+Use Linux, a CUDA GPU, the project dependencies, and a local
+`Qwen/Qwen2.5-Omni-7B` checkpoint with its safetensors index and weights.
+From the repository root:
+
+```bash
+git fetch origin 01a2f93256975c7ff9565c5414b0fe0225bc4765
+PYTHONPATH="$PWD" OMP_NUM_THREADS=1 python benchmarks/qwen2_5_omni/benchmark_embeddings.py \
+  --model /path/to/Qwen2.5-Omni-7B \
+  --baseline-ref 01a2f93256975c7ff9565c5414b0fe0225bc4765 \
+  --cpu-core 2 --lengths 1172 --warmup 50 --blocks 20 --iterations 20 \
+  --output-dir /tmp/qwen25-embedding-results
+```
+
+Choose a CPU core in your allowed affinity set. For the original CUDA length
+sweep, use `--lengths 512 1172 4096 16384`.
+
+## Method
+
+- BF16 checkpoint embeddings, TP=1, CUDA int32 token IDs, CPU multimodal masks.
+- One pinned CPU core, one PyTorch thread, garbage collection disabled.
+- 50 warmups per variant, then 20 paired blocks of 20 calls with balanced,
+  randomized variant order. Summary: median of block medians.
+- Host latency includes synchronization inside the wrapper. Explicit GPU
+  synchronization before and after the timed call is excluded from host latency.
+  Calling-thread CPU time and synchronized completion time are recorded separately.
+- CPU frequency is sampled at block boundaries; clocks are not locked.
+- Validation checks one text embedding call for the fixed wrapper, two for the
+  multimodal baseline, and `torch.equal` for identical inputs and batch shapes.
+  This does not establish batch invariance or full-model output equality.
+
+Interleaved audio/video uses a valid synthetic multimodal mask. This benchmark
+measures wrapper latency, not E2E throughput.
+
+## Reference results
+
+Original environment: H100 80GB; Python 3.12.3; vLLM 0.29.0+cu129;
+PyTorch 2.13.0+cu129; Transformers 5.14.0; Qwen2.5-Omni-7B BF16.
+At 1172 tokens, the September 13 measurements were:
+
+| Input | Before (μs) | After (μs) | Host latency reduction |
+| --- | ---: | ---: | ---: |
+| Text | 18.81 | 18.66 | 0.80% |
+| Audio | 104.32 | 85.30 | 18.23% |
+| Image | 104.72 | 86.27 | 17.62% |
+| Video | 104.73 | 85.76 | 18.11% |
+| Audio + image + video | 567.92 | 549.54 | 3.24% |
+| Interleaved audio/video | 622.76 | 597.84 | 4.00% |
+
+Results vary between runs and environments. These local savings did not produce
+a consistent E2E throughput improvement in the separate repeated E2E experiment.
+
+## Output
+
+The output directory contains `metadata.json`, `micro-raw.json`,
+`validation.json`, and `summary.json`. Raw records include per-call timings,
+block timestamps and CPU frequency samples. Timer overhead is measured separately.

benchmarks/qwen2_5_omni/benchmark_embeddings.py

@@ -0,0 +1,306 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
+"""Reproduce PR #7477 host-observed embedding-wrapper latency on one CUDA GPU.
+
+Run from the PR checkout with PYTHONPATH=$PWD. Baseline and fixed methods share
+weights and inputs. Only the historical wrapper is loaded from --baseline-ref.
+"""
+
+import argparse
+import ast
+import gc
+import json
+import os
+import random
+import statistics
+import subprocess
+import textwrap
+import time
+from pathlib import Path
+
+import torch
+from safetensors import safe_open
+from torch import nn
+from transformers import Qwen2_5OmniThinkerConfig
+from vllm.model_executor.layers.vocab_parallel_embedding import VocabParallelEmbedding
+from vllm.multimodal.utils import set_mm_embedding_modality
+
+from vllm_omni.model_executor.models.qwen2_5_omni.qwen2_5_omni_thinker import (
+    Qwen2_5OmniThinkerForConditionalGeneration as Thinker,
+)
+from vllm_omni.platforms import current_omni_platform
+
+
+def baseline_method(cls, ref):
+    path = "vllm_omni/model_executor/models/qwen2_5_omni/qwen2_5_omni_thinker.py"
+    source = subprocess.check_output(
+        ["git", "-C", str(Path(__file__).resolve().parents[2]), "show", f"{ref}:{path}"], text=True
+    )
+    tree = ast.parse(source)
+    owner = next(n for n in tree.body if isinstance(n, ast.ClassDef) and n.name == cls.__name__)
+    node = next(n for n in owner.body if isinstance(n, ast.FunctionDef) and n.name == "embed_input_ids")
+    method = textwrap.dedent(ast.get_source_segment(source, node))
+    # Preserve zero-argument super() against the real model class and MRO.
+    factory = "from __future__ import annotations\ndef make_method(cls):\n    __class__ = cls\n"
+    factory += textwrap.indent(method, "    ") + "\n    return embed_input_ids\n"
+    namespace = dict(cls.embed_input_ids.__globals__)
+    exec(compile(factory, f"{ref}:{path}", "exec"), namespace)
+    return namespace["make_method"](cls)
+
+
+parser = argparse.ArgumentParser(description=__doc__)
+parser.add_argument("--model", type=Path, required=True)
+parser.add_argument("--baseline-ref", default="01a2f93256975c7ff9565c5414b0fe0225bc4765")
+parser.add_argument("--cpu-core", type=int, default=sorted(os.sched_getaffinity(0))[0])
+parser.add_argument("--lengths", type=int, nargs="+", default=[1172])
+parser.add_argument("--warmup", type=int, default=50)
+parser.add_argument("--blocks", type=int, default=20)
+parser.add_argument("--iterations", type=int, default=20)
+parser.add_argument("--output-dir", type=Path, default=Path("micro-results"))
+args = parser.parse_args()
+if min(args.lengths) < 300 or min(args.blocks, args.iterations) < 1 or args.warmup < 0:
+    parser.error("Require lengths >= 300, positive blocks/iterations, and nonnegative warmup")
+if args.cpu_core not in os.sched_getaffinity(0):
+    parser.error("CPU core is outside the allowed affinity set")
+ROOT, MODEL, CORE = args.output_dir, args.model, args.cpu_core
+ROOT.mkdir(parents=True, exist_ok=True)
+CONFIG = json.loads((MODEL / "config.json").read_text())["thinker_config"]
+HIDDEN = CONFIG["text_config"]["hidden_size"]
+VOCAB = CONFIG["text_config"]["vocab_size"]
+IDS = {m: CONFIG[m + "_token_index"] for m in ("audio", "image", "video")}
+METHODS = {"baseline": baseline_method(Thinker, args.baseline_ref), "fixed": Thinker.embed_input_ids}
+RESULTS, VALIDATIONS = [], []
+os.sched_setaffinity(0, {CORE})
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+gc.disable()
+
+
+def cpu_frequency():
+    path = Path(f"/sys/devices/system/cpu/cpu{CORE}/cpufreq/scaling_cur_freq")
+    if path.exists():
+        return float(path.read_text()) / 1000
+    for section in Path("/proc/cpuinfo").read_text().split("\n\n"):
+        fields = dict(line.split(":", 1) for line in section.splitlines() if ":" in line)
+        fields = {k.strip(): v.strip() for k, v in fields.items()}
+        if fields.get("processor") == str(CORE):
+            return float(fields.get("cpu MHz", "nan"))
+
+
+def payload(case, length, device):
+    generator = torch.Generator().manual_seed(7451 + length)
+    ids = torch.randint(2, 10000, (length,), generator=generator)
+    specs = {
+        "text": [],
+        "audio": [("audio", 100)],
+        "image": [("image", 16)],
+        "video": [("video", 128)],
+        "mixed": [("audio", 100), ("image", 16), ("video", 128)],
+        "interleaved": [("video", 64), ("audio", 50), ("video", 64), ("audio", 50)],
+    }[case]
+    offset = 44
+    counts = {}
+    for modality, count in specs:
+        ids[offset : offset + count] = IDS[modality]
+        counts[modality] = counts.get(modality, 0) + count
+        offset += count + (0 if case == "interleaved" else 2)
+    mask = torch.zeros(length, dtype=torch.bool)
+    for token in IDS.values():
+        mask |= ids == token
+    mm = [
+        set_mm_embedding_modality(torch.full((count, HIDDEN), i + 1, device=device, dtype=torch.bfloat16), modality)
+        for i, (modality, count) in enumerate(counts.items())
+    ]
+    return ids.to(device=device, dtype=torch.int32), None if case == "text" else mm, None if case == "text" else mask
+
+
+def once(fn, cuda):
+    # Drain preceding work before the host timer; the production function's
+    # own D2H synchronizations remain inside the timed call.
+    if cuda:
+        current_omni_platform.synchronize()
+    w0 = time.perf_counter_ns()
+    c0 = time.thread_time_ns()
+    value = fn()
+    c1 = time.thread_time_ns()
+    w1 = time.perf_counter_ns()
+    if cuda:
+        current_omni_platform.synchronize()
+    w2 = time.perf_counter_ns() if cuda else w1
+    del value
+    return [(w1 - w0) / 1000, (c1 - c0) / 1000, (w2 - w0) / 1000]
+
+
+def measure(functions, device, case, length, scope, rounds=20, iterations=20, warmups=50):
+    cuda = device == "cuda"
+    for fn in functions.values():
+        for _ in range(warmups):
+            once(fn, cuda)
+    order_rng = random.Random(7451 + length + len(case) + len(scope))
+    directions = [i % 2 for i in range(rounds)]
+    order_rng.shuffle(directions)
+    names = list(functions)
+    for block, direction in enumerate(directions):
+        order = names if direction == 0 else names[::-1]
+        for name in order:
+            frequency_start = cpu_frequency()
+            wall_start = time.time()
+            samples = [once(functions[name], cuda) for _ in range(iterations)]
+            wall_end = time.time()
+            RESULTS.append(
+                dict(
+                    device=device,
+                    case=case,
+                    tokens=length,
+                    scope=scope,
+                    variant=name,
+                    block=block,
+                    iterations=iterations,
+                    samples_us=samples,
+                    wall_start=wall_start,
+                    wall_end=wall_end,
+                    cpu_mhz_start=frequency_start,
+                    cpu_mhz_end=cpu_frequency(),
+                )
+            )
+    (ROOT / "micro-raw.json").write_text(json.dumps(RESULTS))
+    current = [r for r in RESULTS if (r["device"], r["case"], r["tokens"], r["scope"]) == (device, case, length, scope)]
+    print(
+        "RESULT",
+        json.dumps(
+            {
+                "device": device,
+                "case": case,
+                "tokens": length,
+                "scope": scope,
+                "host_median_us": {
+                    v: statistics.median(t[0] for r in current if r["variant"] == v for t in r["samples_us"])
+                    for v in names
+                },
+            }
+        ),
+        flush=True,
+    )
+
+
+def build_model(weights, device):
+    model = Thinker.__new__(Thinker)
+    nn.Module.__init__(model)
+    model.config = Qwen2_5OmniThinkerConfig(**CONFIG)
+    model._has_oov_mm_tokens = False
+    table = VocabParallelEmbedding(VOCAB, HIDDEN, params_dtype=torch.bfloat16, disable_tp=True)
+    table.weight = nn.Parameter(weights.to(device), requires_grad=False)
+
+    class LM(nn.Module):
+        def __init__(self):
+            super().__init__()
+            self.embed_tokens = table
+
+        def embed_input_ids(self, ids):
+            return self.embed_tokens(ids)
+
+    model.language_model = LM()
+    return model
+
+
+def validate(model, args, case, length, device):
+    ids, mm, mask = args
+    lm = model.get_language_model()
+    original = lm.embed_input_ids
+    outputs, counts = {}, {}
+    for variant, method in METHODS.items():
+        calls = []
+
+        def counted(tokens):
+            calls.append(tokens.numel())
+            return original(tokens)
+
+        lm.embed_input_ids = counted
+        try:
+            outputs[variant] = method(model, ids, mm, is_multimodal=mask)
+        finally:
+            lm.embed_input_ids = original
+        counts[variant] = len(calls)
+    assert torch.equal(outputs["baseline"], outputs["fixed"])
+    assert counts == {"baseline": 1 if case == "text" else 2, "fixed": 1}
+    VALIDATIONS.append(dict(case=case, tokens=length, device=device, calls=counts, same_input_wrapper_torch_equal=True))
+    (ROOT / "validation.json").write_text(json.dumps(VALIDATIONS, indent=2))
+
+
+@torch.inference_mode()
+def main():
+    import transformers
+    import vllm
+
+    metadata = dict(
+        baseline_ref=args.baseline_ref,
+        checkout=subprocess.check_output(["git", "rev-parse", "HEAD"], text=True).strip(),
+        torch=torch.__version__,
+        vllm=vllm.__version__,
+        transformers=transformers.__version__,
+        gpu=torch.cuda.get_device_name(),
+        cpu_core=CORE,
+        torch_cpu_threads=1,
+        warmup=args.warmup,
+        blocks=args.blocks,
+        iterations=args.iterations,
+        lengths=args.lengths,
+        embedding_shape=[VOCAB, HIDDEN],
+        dtype="bfloat16",
+        input_ids_dtype="int32",
+        mask_device="cpu",
+        start_time=time.time(),
+        metric_order=["host_observed_us", "calling_thread_cpu_us", "synchronized_completion_us"],
+    )
+    index = json.loads((MODEL / "model.safetensors.index.json").read_text())["weight_map"]
+    key = "thinker.model.embed_tokens.weight"
+    with safe_open(MODEL / index[key], framework="pt", device="cpu") as f:
+        weights = f.get_tensor(key)
+    assert weights.shape == (VOCAB, HIDDEN) and weights.dtype == torch.bfloat16
+    model = build_model(weights, "cuda")
+    measure({"timer_floor": lambda: None}, "cuda", "timer", 0, "overhead", rounds=10, iterations=100)
+    for length in args.lengths:
+        for case in ("text", "audio", "image", "video", "mixed", "interleaved"):
+            ids, mm, mask = payload(case, length, "cuda")
+            validate(model, (ids, mm, mask), case, length, "cuda")
+            functions = {
+                variant: (lambda method=method: method(model, ids, mm, is_multimodal=mask))
+                for variant, method in METHODS.items()
+            }
+            measure(
+                functions,
+                "cuda",
+                case,
+                length,
+                "whole_wrapper",
+                rounds=args.blocks,
+                iterations=args.iterations,
+                warmups=args.warmup,
+            )
+    summary = []
+    for length in args.lengths:
+        for case in ("text", "audio", "image", "video", "mixed", "interleaved"):
+            medians = {
+                variant: statistics.median(
+                    statistics.median(sample[0] for sample in record["samples_us"])
+                    for record in RESULTS
+                    if record["tokens"] == length and record["case"] == case and record["variant"] == variant
+                )
+                for variant in METHODS
+            }
+            summary.append(
+                dict(
+                    tokens=length,
+                    case=case,
+                    host_us=medians,
+                    reduction_pct=100 * (1 - medians["fixed"] / medians["baseline"]),
+                )
+            )
+    metadata["end_time"] = time.time()
+    (ROOT / "metadata.json").write_text(json.dumps(metadata, indent=2))
+    (ROOT / "summary.json").write_text(json.dumps(summary, indent=2))
+    print(json.dumps(summary, indent=2))
+
+
+if __name__ == "__main__":
+    main()

benchmarks/tts/README.md

@@ -2,7 +2,7 @@
 
 A model-agnostic serving benchmark for TTS models in vllm-omni. One CLI
 (`bench_tts.py`) + one YAML registry (`model_configs.yaml`) drive perf and
-quality runs for every registered checkpoint: **Qwen3-TTS** (Base / CustomVoice)
+quality runs for every registered checkpoint: **Qwen3-TTS** (Base / CustomVoice / VoiceDesign)
 and **VoxCPM2** today, more to come.
 
 The same three task types — `voice_clone`, `default_voice`, `voice_design` —
@@ -18,8 +18,13 @@ vllm serve Qwen/Qwen3-TTS-12Hz-1.7B-Base --omni --port 8000
 ```
 
 The server auto-loads its Deploy YAML from `vllm_omni/deploy/qwen3_tts.yaml`
-(Pipeline + Deploy schema introduced in #2383). No `--stage-configs-path` or
-`--deploy-config` flag is needed for any registered model.
+(Pipeline + Deploy schema introduced in #2383). No explicit `--deploy-config`
+flag is needed for a registered model.
+
+Start the checkpoint that matches the benchmark task below: `-Base` for
+`voice_clone`, `-CustomVoice` for `default_voice`, or `-VoiceDesign` for
+`voice_design`. Restart the server with the matching checkpoint when switching
+tasks; changing only the benchmark's `--model` does not reload the server.
 
 ### 2. Run the benchmark (`vllm bench serve --omni`)
 
@@ -62,12 +67,12 @@ vllm bench serve --omni \
     --save-result --result-dir ./results
 ```
 
-#### voice_design (Qwen3-TTS-CustomVoice, bundled seed_tts_design)
+#### voice_design (Qwen3-TTS-VoiceDesign, bundled seed_tts_design)
 
 ```bash
 vllm bench serve --omni \
     --host 127.0.0.1 --port 8000 \
-    --model Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice \
+    --model Qwen/Qwen3-TTS-12Hz-1.7B-VoiceDesign \
     --backend openai-audio-speech \
     --endpoint /v1/audio/speech \
     --dataset-name seed-tts-design \
@@ -80,6 +85,25 @@ vllm bench serve --omni \
     --save-result --result-dir ./results
 ```
 
+#### Set the sample rate for models that are not 24 kHz
+
+Streamed PCM carries no header, so the harness derives `audio_duration` from the
+byte count and a **default 24000 Hz** assumption
+(`vllm_omni/metrics/definitions.py:DEFAULT_AUDIO_SAMPLE_RATE`). For a model that
+emits another rate, export the real one or every duration-derived metric is
+wrong by `real_sr / 24000`:
+
+```bash
+# Audio8 TTS, Fish Speech S2 Pro, Ming-omni-tts, ... all decode at 44.1 kHz
+export VLLM_OMNI_BENCH_AUDIO_SAMPLE_RATE=44100
+```
+
+Without it, a 44.1 kHz model reports `audio_duration` 1.84x too long, so
+`audio_rtf` looks 1.84x *better* than reality and `audio_underrun` /
+`Streaming continuity OK rate` are measured against inflated chunk budgets.
+A quick sanity check: compare `Mean AUDIO_DURATION (s)` against the actual
+length of a saved WAV for the same prompt.
+
 #### Streaming continuity (`audio_underrun`)
 
 `audio_underrun` (seconds) is the per-request worst-case buffer deficit
@@ -139,6 +163,34 @@ python benchmarks/tts/bench_tts.py \
     --output-dir ./results
 ```
 
+#### Fixed IndexTTS 2.5 performance sweep
+
+IndexTTS 2.5 is registered in `model_configs.yaml` like the other TTS models.
+Use `--served-model-name` when the server was launched from a local native
+bundle instead of the registry key. The command below reproduces the acceptance
+workload: Seed-TTS Eval EN, `n=500`, concurrency `4/8/16/32`, five warmups, and
+dataset seed 0.
+
+```bash
+python benchmarks/tts/bench_tts.py \
+    --model IndexTeam/IndexTTS-2.5 \
+    --served-model-name /path/to/indextts-2.5 \
+    --task voice_clone --locale en \
+    --dataset-path /path/to/seedtts_testset \
+    --host 127.0.0.1 --port 8092 \
+    --concurrency 4 8 16 32 \
+    --num-prompts 500 --num-warmups 5 \
+    --output-dir ./results/indextts2_5 \
+    -- \
+    --tokenizer /path/to/indextts-2.5/qwen0.6bemo4-merge \
+    --seed 0 --metric-percentiles 50,95,99 --disable-tqdm --save-detailed
+```
+
+The wrapper benchmarks an already-running server. For reproducible quality
+comparisons, add `--request-seed 42`; production-performance runs should omit
+it. The trailing `--seed 0` controls Seed-TTS row selection rather than model
+sampling.
+
 ### 4. Plot a sweep
 
 ```bash
@@ -152,11 +204,11 @@ Outputs TTFP / RTF / throughput curves (and a markdown table) for every
 
 ## Task types
 
-| Task            | Dataset           | Request body                                        | Checkpoints that support it              |
-|-----------------|-------------------|-----------------------------------------------------|------------------------------------------|
-| `voice_clone`   | `seed-tts`        | `ref_audio` + `ref_text` + `task_type=Base`         | `Qwen3-TTS-*-Base`, `VoxCPM2`            |
-| `default_voice` | `seed-tts-text`   | `voice=Vivian` + `task_type=CustomVoice`            | `Qwen3-TTS-*-CustomVoice`                |
-| `voice_design`  | `seed-tts-design` | `instructions=<natural-language description>` + `task_type=VoiceDesign` | `Qwen3-TTS-*-CustomVoice` |
+| Task            | Dataset           | Request body                                                            | Checkpoints that support it   |
+| --------------- | ----------------- | ----------------------------------------------------------------------- | ----------------------------- |
+| `voice_clone`   | `seed-tts`        | `ref_audio` + `ref_text` + `task_type=Base`                             | `Qwen3-TTS-*-Base`, `VoxCPM2` |
+| `default_voice` | `seed-tts-text`   | `voice=Vivian` + `task_type=CustomVoice`                                | `Qwen3-TTS-*-CustomVoice`     |
+| `voice_design`  | `seed-tts-design` | `instructions=<natural-language description>` + `task_type=VoiceDesign` | `Qwen3-TTS-*-VoiceDesign`     |
 
 **`-CustomVoice` checkpoints do NOT ship `speaker_encoder` weights**, so
 voice_clone requests raise `ValueError` at model runtime. Use `-Base` for
@@ -182,12 +234,12 @@ Then add the model's Deploy YAML under `vllm_omni/deploy/<model>.yaml`
 
 ## Datasets
 
-| Dataset            | Bundled? | Format            | Source                                                         |
-|--------------------|----------|-------------------|----------------------------------------------------------------|
-| `seed-tts-design`  | ✅       | 5-field meta.lst  | `benchmarks/build_dataset/seed_tts_design/en/meta.lst` (20 prompts) |
-| `seed_tts_smoke`   | ✅       | 4-field meta.lst  | `benchmarks/build_dataset/seed_tts_smoke/en/meta.lst` (20 text-only) |
-| `seed-tts`         | ❌       | 4-field meta.lst + WAVs | Google-Drive: [BytedanceSpeech/seed-tts-eval][seedtts] (~1.2 GB) |
-| `seed-tts-text`    | ❌       | 4-field meta.lst  | Same archive as `seed-tts` (wav column unused)                 |
+| Dataset           | Bundled? | Format                  | Source                                                               |
+| ----------------- | -------- | ----------------------- | -------------------------------------------------------------------- |
+| `seed-tts-design` | ✅       | 5-field meta.lst        | `benchmarks/build_dataset/seed_tts_design/en/meta.lst` (20 prompts)  |
+| `seed_tts_smoke`  | ✅       | 4-field meta.lst        | `benchmarks/build_dataset/seed_tts_smoke/en/meta.lst` (20 text-only) |
+| `seed-tts`        | ❌       | 4-field meta.lst + WAVs | Google-Drive: [BytedanceSpeech/seed-tts-eval][seedtts] (~1.2 GB)     |
+| `seed-tts-text`   | ❌       | 4-field meta.lst        | Same archive as `seed-tts` (wav column unused)                       |
 
 [seedtts]: https://github.com/BytedanceSpeech/seed-tts-eval
 
@@ -199,11 +251,11 @@ For manual voice_clone / default_voice runs against the full corpus, follow
 
 `tests/dfx/perf/tests/test_tts.json` wires three perf regimes plus quality:
 
-| eval_phase    | concurrency | purpose                                                 | Baseline metrics                        |
-|---------------|-------------|---------------------------------------------------------|-----------------------------------------|
-| `latency`     | 1           | Single-request TTFP / RTF SLO                           | `median_audio_ttfp_ms`, `median_audio_rtf` |
-| `throughput`  | 8           | Codec-batching cliff sentinel (PDF #272 concurrency≥8)  | `median_audio_ttfp_ms`, `median_audio_rtf` |
-| `quality`     | 4           | WER / SIM / UTMOS regression (disabled in CI by default)| `mean_audio_rtf`                        |
+| eval_phase   | concurrency | purpose                                                  | Baseline metrics                           |
+| ------------ | ----------- | -------------------------------------------------------- | ------------------------------------------ |
+| `latency`    | 1           | Single-request TTFP / RTF SLO                            | `median_audio_ttfp_ms`, `median_audio_rtf` |
+| `throughput` | 8           | Codec-batching cliff sentinel (PDF #272 concurrency≥8)   | `median_audio_ttfp_ms`, `median_audio_rtf` |
+| `quality`    | 4           | WER / SIM / UTMOS regression (disabled in CI by default) | `mean_audio_rtf`                           |
 
 Why `median_*` for latency/throughput and `mean_*` for quality: latency
 distributions have cold-start tails that drag the mean; quality aggregates
@@ -217,10 +269,15 @@ PR #2558 — quality runs are manual / release-validation, not nightly).
 
 Observed on H20-3e, Qwen3-TTS-1.7B (measured pre-merge on this branch):
 
-| Task          | Model         | c=1    | c=4    | **c=8**    | c=16   | c=32   |
-|---------------|---------------|--------|--------|------------|--------|--------|
-| voice_clone   | 1.7B-Base     | RTF 0.15 / TTFP 165ms | 0.28 / 412ms | **0.49 / 1701ms** | 0.72 / 3355ms | 0.77 / 3772ms |
-| voice_design  | 1.7B-CustomVoice | RTF 0.08 / TTFP 53ms  | 0.11 / 154ms | **0.21 / 872ms**  | 0.33 / 1801ms | 0.38 / 1989ms |
+| Task         | Model            | c=1                   | c=4          | **c=8**           | c=16          | c=32          |
+| ------------ | ---------------- | --------------------- | ------------ | ----------------- | ------------- | ------------- |
+| voice_clone  | 1.7B-Base        | RTF 0.15 / TTFP 165ms | 0.28 / 412ms | **0.49 / 1701ms** | 0.72 / 3355ms | 0.77 / 3772ms |
+| voice_design | 1.7B-CustomVoice | RTF 0.08 / TTFP 53ms  | 0.11 / 154ms | **0.21 / 872ms**  | 0.33 / 1801ms | 0.38 / 1989ms |
+
+The historical `voice_design` row above lists a CustomVoice checkpoint, which
+current task validation rejects for VoiceDesign requests. It does not establish
+a VoiceDesign checkpoint baseline; rerun with `Qwen3-TTS-12Hz-1.7B-VoiceDesign`
+before comparing performance.
 
 Both models show a **4–6× TTFP jump from c=4 to c=8** while audio throughput
 saturates around c=4–8 — the codec-bs=1 bottleneck documented in
@@ -229,7 +286,7 @@ sentinel for regressions in this area.
 
 ## File layout
 
-```
+```text
 benchmarks/tts/
 ├── README.md                  (this file)
 ├── bench_tts.py               CLI — serve-mode benchmark driver

benchmarks/tts/bench_cosyvoice3_concurrency_matrix.py

@@ -0,0 +1,517 @@
+#!/usr/bin/env python3
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
+
+"""CosyVoice3 streaming e2e benchmark: concurrency x utterance-length matrix.
+
+Launches the server once per (concurrency, step-size) combo, streams TTS
+requests cycling through several utterance lengths, and reports
+TTFA/E2E/RTF/per-chunk latency (mean, stdev, range) per cell. "Step size" is
+``codec_chunk_frames``, the token-hop length the first emitted chunk waits
+for; later chunks wait for a doubling hop capped at ``4 * codec_chunk_frames``
+(see ``_token_hop_lengths``). Per-chunk latency is reconstructed from that
+same ramp rather than from raw transport read boundaries, which don't align
+to model chunks (see ``one_request``'s docstring).
+
+Running with no flags reproduces the default matrix. Per-request correctness
+is checked: non-silent, long enough, and (with a fixed ``--seed``)
+byte-identical PCM across requests for the same text.
+"""
+
+import argparse
+import json
+import os
+import signal
+import statistics
+import subprocess
+import sys
+import threading
+import time
+from pathlib import Path
+
+import httpx
+import numpy as np
+
+REPO = Path(__file__).resolve().parents[2]
+DEFAULT_MODEL = "FunAudioLLM/Fun-CosyVoice3-0.5B-2512"
+
+DEFAULT_REF_AUDIO = "https://raw.githubusercontent.com/FunAudioLLM/CosyVoice/main/asset/zero_shot_prompt.wav"
+DEFAULT_REF_TEXT = "希望你以后能够做的比我还好呦。"
+
+DEFAULT_TEXTS = [
+    "Thank you so much for the birthday gift you sent me.",
+    "The weather is lovely today, with warm sunshine and a gentle breeze. "
+    "We decided to take a walk in the park to enjoy the spring flowers and the fresh air.",
+    "The weather is lovely today, with warm sunshine and a gentle breeze. We decided to take a walk in the "
+    "park to enjoy the spring flowers and the fresh air. Many people were out walking, children were "
+    "running and playing on the grass, and older folks sat on benches chatting with one another. We "
+    "strolled slowly along the path and came across a cherry tree in full bloom, its pink petals drifting "
+    "on the wind. It was so beautiful that we stopped to take some photos to remember the moment. Then we "
+    "continued on and reached a small lake, its water so clear you could see straight to the bottom, with "
+    "the blue sky and white clouds reflected on its surface. A few ducks paddled back and forth across the "
+    "water, occasionally letting out a soft quack. We sat on a bench by the lake for a while, enjoying the "
+    "peaceful afternoon. Time passed quickly, and before we knew it evening had arrived, with the setting "
+    "sun painting the sky in shades of gold. Reluctant to leave, we finally headed home, already planning "
+    "our next visit. On the way back we talked about everything we had seen and heard that day, feeling "
+    "wonderfully happy. It was truly an unforgettable weekend.",
+]
+
+# PCM16 mono @ 24 kHz (CosyVoice3 output rate)
+SAMPLE_RATE = 24000
+
+# Well below any real speech's RMS; catches all-zero/near-silent PCM that byte
+# count and duration checks alone pass through.
+MIN_RMS = 50.0
+
+# Overlay: only codec_chunk_frames is overridden; must live next to cosyvoice3.yaml (relative base_config)
+DEPLOY_TEMPLATE = """# Auto-generated by bench_cosyvoice3_concurrency_matrix.py -- do not edit.
+base_config: cosyvoice3.yaml
+
+connectors:
+  connector_of_shared_memory:
+    name: SharedMemoryConnector
+    extra:
+      codec_streaming: true
+      connector_get_sleep_s: 0.01
+      connector_get_max_wait_first_chunk: 3000
+      connector_get_max_wait: 300
+      codec_chunk_frames: {chunk_frames}
+      codec_max_chunk_frames: {max_chunk_frames}
+      codec_stream_scale_factor: {stream_scale_factor}
+      codec_left_context_frames: 25
+      codec_vocab_size: 6561
+"""
+
+# CosyVoice3Config constants (transformers_utils/configs/cosyvoice3.py): tokens
+# are consumed at token_mel_ratio mel frames each, and HiFT emits
+# samples_per_mel PCM16 samples per mel frame. Pinned here (not left to the
+# server's defaults) since one_request's byte-boundary reconstruction below
+# depends on knowing them exactly.
+TOKEN_MEL_RATIO = 2
+SAMPLES_PER_MEL = 480
+STREAM_SCALE_FACTOR = 2
+
+
+def _stats(values: list[float]) -> dict[str, float]:
+    """Mean/median/stdev/p95/min/max/n for a list of per-request measurements."""
+    n = len(values)
+    if n == 0:
+        return {
+            "n": 0,
+            "mean": float("nan"),
+            "median": float("nan"),
+            "stdev": float("nan"),
+            "p95": float("nan"),
+            "min": float("nan"),
+            "max": float("nan"),
+        }
+    return {
+        "n": n,
+        "mean": statistics.mean(values),
+        "median": statistics.median(values),
+        "stdev": statistics.stdev(values) if n >= 2 else 0.0,
+        "p95": statistics.quantiles(values, n=20)[18] if n >= 2 else values[0],
+        "min": min(values),
+        "max": max(values),
+    }
+
+
+def vllm_bin() -> str:
+    for c in (REPO / ".venv/bin/vllm", Path(sys.executable).parent / "vllm"):
+        if c.is_file():
+            return str(c)
+    return "vllm"
+
+
+def wait_ready(url: str, proc: subprocess.Popen, log: Path, timeout: float) -> None:
+    deadline = time.monotonic() + timeout
+    while time.monotonic() < deadline:
+        if proc.poll() is not None:
+            tail = log.read_text(errors="replace").splitlines()[-40:] if log.exists() else []
+            raise RuntimeError("server exited during startup:\n  " + "\n  ".join(tail))
+        try:
+            if httpx.get(f"{url}/health", timeout=2.0).status_code == 200:
+                return
+        except Exception:
+            pass
+        time.sleep(2.0)
+    raise TimeoutError(f"server not ready within {timeout}s (log: {log})")
+
+
+def _token_hop_lengths(chunk_frames: int, max_chunk_frames: int):
+    """The token-hop-length ramp talker2code2wav_async_chunk uses to decide
+    when to fire the next flow-solve/emission: chunk_frames, then doubling
+    (STREAM_SCALE_FACTOR) each step, capped at max_chunk_frames. This -- not
+    codec_chunk_frames alone -- is what actually paces emitted audio chunks:
+    with the defaults here, hops are chunk_frames, 2x, 4x, 4x, 4x, ...
+
+    The very first emission is consistently one token short of chunk_frames
+    (confirmed against real speech.audio.delta event sizes) even though the
+    ramp's own progression is computed from the nominal chunk_frames, not the
+    actual first hop -- so only the first yield gets the -1 adjustment.
+    """
+    yield chunk_frames - 1
+    hop = chunk_frames
+    while True:
+        hop = min(max_chunk_frames, max(chunk_frames, hop * STREAM_SCALE_FACTOR))
+        yield hop
+
+
+def one_request(client, url, model, text, ref_audio, ref_text, chunk_frames, seed=None):
+    """Run one streaming request, returning TTFA, E2E, per-chunk gaps, and audio bytes.
+
+    ``response_format=pcm`` + ``stream_format=audio`` streams raw bytes with no
+    framing, so a client read boundary alone is a transport artifact, not a
+    model chunk boundary. Reconstructed instead by replicating the server's
+    own token-hop-length ramp (see ``_token_hop_lengths``) to compute each
+    real chunk's expected byte size ahead of time, and slicing the
+    accumulated byte stream at those offsets -- independent of how the
+    transport happens to fragment the underlying reads. The final chunk
+    (finalize, which releases extra held-back frames) is whatever remains
+    when the stream closes, not a predicted hop.
+    """
+    payload = {
+        "model": model,
+        "input": text,
+        "response_format": "pcm",
+        "stream": True,
+        "stream_format": "audio",
+        "ref_audio": ref_audio,
+        "ref_text": ref_text,
+    }
+    if seed is not None:
+        payload["seed"] = seed
+    t0 = time.perf_counter()
+    ttfa = None
+    nbytes = 0
+    audio_bytes = bytearray()
+    chunk_gaps_ms = []  # inter-chunk arrival gaps (per model chunk), by step index
+    prev_t = None
+    step_idx = 0
+
+    bytes_per_token = TOKEN_MEL_RATIO * SAMPLES_PER_MEL * 2
+    hop_gen = _token_hop_lengths(chunk_frames, 4 * chunk_frames)
+    next_boundary = next(hop_gen) * bytes_per_token
+    last_crossed_bytes = 0
+    received = 0
+    last_now = t0
+
+    with client.stream("POST", f"{url}/v1/audio/speech", json=payload, headers={"Authorization": "Bearer EMPTY"}) as r:
+        if r.status_code != 200:
+            raise RuntimeError(f"{r.status_code}: {r.read()[:300]!r}")
+        for chunk in r.iter_bytes():
+            if not chunk:
+                continue
+            now = time.perf_counter()
+            if ttfa is None:
+                ttfa = now - t0
+            last_now = now
+            received += len(chunk)
+            nbytes += len(chunk)
+            audio_bytes += chunk
+            while received >= next_boundary:
+                if prev_t is not None:
+                    chunk_gaps_ms.append((step_idx, (now - prev_t) * 1e3))
+                    step_idx += 1
+                prev_t = now
+                last_crossed_bytes = next_boundary
+                next_boundary += next(hop_gen) * bytes_per_token
+
+    if received > last_crossed_bytes:
+        # Finalize chunk: releases extra held-back frames, so it doesn't
+        # match a predicted hop size -- its arrival is the stream closing.
+        if prev_t is not None:
+            chunk_gaps_ms.append((step_idx, (last_now - prev_t) * 1e3))
+            step_idx += 1
+
+    e2e = time.perf_counter() - t0
+    audio_s = nbytes / 2 / SAMPLE_RATE
+    return {
+        "ttfa_s": ttfa,
+        "e2e_s": e2e,
+        "audio_s": audio_s,
+        "rtf": e2e / audio_s if audio_s > 0 else float("nan"),
+        "chunk_gaps_ms": chunk_gaps_ms,
+        "nbytes": nbytes,
+        "audio_bytes": audio_bytes,
+    }
+
+
+def check_correctness(r: dict, min_audio_s: float) -> list[str]:
+    """Return a list of correctness violations (empty = pass)."""
+    problems = []
+    if r["nbytes"] <= 0:
+        problems.append("empty audio (0 bytes)")
+    if r["audio_s"] < min_audio_s:
+        problems.append(f"audio too short ({r['audio_s']:.2f}s < {min_audio_s:.2f}s)")
+    if r["nbytes"] > 0:
+        samples = np.frombuffer(bytes(r["audio_bytes"]), dtype=np.int16).astype(np.float64)
+        rms = float(np.sqrt(np.mean(samples**2)))
+        if rms < MIN_RMS:
+            problems.append(f"near-silent audio (rms={rms:.1f} < {MIN_RMS})")
+    return problems
+
+
+def drive(url, args, n, conc, label, results, lock, idx, chunk_frames):
+    """Run ``n`` requests at concurrency ``conc``, cycling through ``args.texts``."""
+    texts = args.texts
+
+    def worker():
+        with httpx.Client(timeout=600.0) as c:
+            while True:
+                with lock:
+                    i = idx[0]
+                    idx[0] += 1
+                if i >= n:
+                    return
+                text = texts[i % len(texts)]
+                try:
+                    r = one_request(
+                        c, url, args.model, text, args.ref_audio, args.ref_text, chunk_frames, seed=args.seed
+                    )
+                    r["text"] = text
+                    with lock:
+                        results.append(r)
+                except Exception as e:
+                    print(f"    [{label}] request {i} failed: {e}", flush=True)
+
+    ts = [threading.Thread(target=worker) for _ in range(conc)]
+    t0 = time.perf_counter()
+    for t in ts:
+        t.start()
+    for t in ts:
+        t.join()
+    return time.perf_counter() - t0
+
+
+def run_combo(args, conc: int, chunk_frames: int, out: Path) -> dict:
+    tag = f"c{conc}_chunk{chunk_frames}"
+    run_dir = out / tag
+    run_dir.mkdir(parents=True, exist_ok=True)
+    log = run_dir / "server.log"
+
+    deploy = REPO / "vllm_omni/deploy" / f"_autogen_steps_{tag}.yaml"
+    deploy.write_text(
+        DEPLOY_TEMPLATE.format(
+            chunk_frames=chunk_frames,
+            max_chunk_frames=4 * chunk_frames,
+            stream_scale_factor=STREAM_SCALE_FACTOR,
+        )
+    )
+
+    cmd = [
+        vllm_bin(),
+        "serve",
+        args.model,
+        "--omni",
+        "--deploy-config",
+        str(deploy),
+        "--host",
+        "127.0.0.1",
+        "--port",
+        str(args.port),
+        "--allowed-local-media-path",
+        str(REPO),
+    ]
+    env = dict(os.environ)
+    if args.cuda_visible_devices:
+        env["CUDA_VISIBLE_DEVICES"] = args.cuda_visible_devices
+    if args.trt is not None:
+        env["COSYVOICE3_TRT"] = args.trt
+
+    print(f"\n=== concurrency={conc} step_size={chunk_frames} ===", flush=True)
+    print(f"  launching server (log: {log})", flush=True)
+    with open(log, "w") as lf:
+        proc = subprocess.Popen(cmd, env=env, stdout=lf, stderr=subprocess.STDOUT, start_new_session=True)
+    url = f"http://127.0.0.1:{args.port}"
+    try:
+        wait_ready(url, proc, log, args.server_timeout)
+        print("  server ready", flush=True)
+
+        results: list[dict] = []
+        lock = threading.Lock()
+        idx = [0]
+        if args.warmup:
+            print(f"  warmup ({args.warmup} req)", flush=True)
+            drive(url, args, args.warmup, min(conc, args.warmup), "warmup", results, lock, idx, chunk_frames)
+            results.clear()
+            idx[0] = 0
+
+        print(f"  running {args.num_prompts} req @ concurrency {conc}", flush=True)
+        wall = drive(url, args, args.num_prompts, conc, "bench", results, lock, idx, chunk_frames)
+
+        ok = [r for r in results if r.get("ttfa_s") is not None]
+        problems = []
+        for r in ok:
+            problems += check_correctness(r, args.min_audio_s)
+        # With a fixed seed, same-text requests must produce byte-identical PCM (per text group)
+        if args.seed is not None:
+            by_text: dict[str, bytes] = {}
+            for r in ok:
+                t = r.get("text")
+                ref = by_text.get(t)
+                if ref is None:
+                    by_text[t] = r.get("audio_bytes")
+                elif r.get("audio_bytes") != ref:
+                    problems.append(f"byte-identity violation for text {t!r}: seeded requests produced different PCM")
+        n_ok = len(ok)
+        total_audio_s = sum(r["audio_s"] for r in ok)
+        all_gaps = [g for r in ok for g in r["chunk_gaps_ms"]]  # (step_index, gap_ms) per request
+        gap_vals = [g for _, g in all_gaps]
+
+        # bucket gaps by chunk position within a request, to see any O(n) slope
+        by_idx: dict[int, list[float]] = {}
+        for step_index, g in all_gaps:
+            by_idx.setdefault(step_index, []).append(g)
+        step_latency_by_idx = {i: statistics.median(vals) for i, vals in sorted(by_idx.items())}
+
+        agg = {
+            "concurrency": conc,
+            "step_size": chunk_frames,
+            "ok": n_ok,
+            "errors": args.num_prompts - n_ok,
+            "correctness_problems": len(problems),
+            "correctness_sample": problems[:5],
+            "total_audio_s": total_audio_s,
+            "throughput_audio_s_per_s": total_audio_s / wall if wall > 0 else float("nan"),
+            "ttfa_s": _stats([r["ttfa_s"] for r in ok]),
+            "e2e_s": _stats([r["e2e_s"] for r in ok]),
+            "rtf": _stats([r["rtf"] for r in ok]),
+            "step_latency_ms": _stats(gap_vals),
+            "step_latency_ms_by_idx": step_latency_by_idx,
+            "num_steps": len(all_gaps),
+            # per-utterance-length breakdown; n per text = num_prompts / len(texts)
+            "by_text": {
+                t: {
+                    "audio_s": _stats([r["audio_s"] for r in ok if r.get("text") == t]),
+                    "e2e_s": _stats([r["e2e_s"] for r in ok if r.get("text") == t]),
+                    "rtf": _stats([r["rtf"] for r in ok if r.get("text") == t]),
+                }
+                for t in sorted({r.get("text") for r in ok})
+            },
+        }
+        print(json.dumps(agg, indent=2, ensure_ascii=False), flush=True)
+        return agg
+    finally:
+        print("  shutting down server", flush=True)
+        try:
+            os.killpg(os.getpgid(proc.pid), signal.SIGTERM)
+        except Exception:
+            proc.terminate()
+        try:
+            proc.wait(timeout=60)
+        except subprocess.TimeoutExpired:
+            try:
+                os.killpg(os.getpgid(proc.pid), signal.SIGKILL)
+            except Exception:
+                proc.kill()
+            proc.wait(timeout=20)
+        deploy.unlink(missing_ok=True)
+
+
+def parse_args() -> argparse.Namespace:
+    p = argparse.ArgumentParser(description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter)
+    p.add_argument("--model", default=DEFAULT_MODEL)
+    p.add_argument(
+        "--text",
+        dest="texts",
+        nargs="+",
+        default=DEFAULT_TEXTS,
+        help="One or more utterance texts. With several, requests cycle through "
+        "them so a single server launch covers all utterance lengths; results "
+        "are bucketed per text.",
+    )
+    p.add_argument("--ref-audio", default=DEFAULT_REF_AUDIO)
+    p.add_argument("--ref-text", default=DEFAULT_REF_TEXT)
+    p.add_argument("--concurrency", type=int, nargs="+", default=[1, 8])
+    p.add_argument(
+        "--step-size",
+        type=int,
+        nargs="+",
+        default=[15],
+        help="codec_chunk_frames values to sweep (the streaming step size).",
+    )
+    p.add_argument("--num-prompts", type=int, default=30)
+    p.add_argument("--warmup", type=int, default=2)
+    p.add_argument("--port", type=int, default=8091)
+    p.add_argument(
+        "--server-timeout",
+        type=float,
+        default=1800.0,
+        help="Startup budget; first run also downloads weights and builds the TRT engine.",
+    )
+    p.add_argument("--output-dir", default="./perf_steps")
+    p.add_argument("--cuda-visible-devices", default="0")
+    p.add_argument(
+        "--trt",
+        choices=["0", "1"],
+        default=None,
+        help="Pin COSYVOICE3_TRT. 0=force torch estimator, 1=allow TensorRT. Default: unset.",
+    )
+    p.add_argument(
+        "--min-audio-s",
+        type=float,
+        default=1.0,
+        help="Minimum expected audio seconds per request for the correctness check.",
+    )
+    p.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="Fixed seed for reproducible (byte-identical) audio. Pass --seed '' "
+        "equivalent (any non-int) to disable -- there is no unset sentinel here "
+        "by design, since the default sweep is meant to be exactly reproducible.",
+    )
+    return p.parse_args()
+
+
+def main() -> None:
+    args = parse_args()
+    out = Path(args.output_dir).resolve()
+    out.mkdir(parents=True, exist_ok=True)
+
+    all_results = []
+    failed = False
+    for conc in args.concurrency:
+        for chunk_frames in args.step_size:
+            try:
+                result = run_combo(args, conc, chunk_frames, out)
+                all_results.append(result)
+                if result["errors"] > 0 or result["correctness_problems"] > 0:
+                    failed = True
+            except Exception as e:
+                print(f"  combo c={conc} chunk={chunk_frames} FAILED: {e}", flush=True)
+                failed = True
+
+    out_path = out / "steps_results.json"
+    with open(out_path, "w", encoding="utf-8") as f:
+        json.dump({"model": args.model, "texts": args.texts, "runs": all_results}, f, indent=2, ensure_ascii=False)
+    print(f"\nResults written to {out_path}", flush=True)
+
+    print("\n=== SUMMARY ===")
+    header = (
+        f"{'conc':>4} {'step':>4} {'ok':>4} {'corr':>4} "
+        f"{'TTFA med(ms)':>12} {'step med(ms)':>12} {'step p95(ms)':>12} {'RTF mean+-sd':>14}"
+    )
+    print(header)
+    print("-" * len(header))
+    for r in all_results:
+        print(
+            f"{r['concurrency']:>4} {r['step_size']:>4} {r['ok']:>4} {r['correctness_problems']:>4} "
+            f"{r['ttfa_s']['median'] * 1e3:>12.1f} "
+            f"{r['step_latency_ms']['median']:>12.1f} {r['step_latency_ms']['p95']:>12.1f} "
+            f"{r['rtf']['mean']:>6.3f}+-{r['rtf']['stdev']:>5.3f}"
+        )
+        for t, stats in r.get("by_text", {}).items():
+            print(
+                f"      text {t[:24]!r}: e2e {stats['e2e_s']['mean'] * 1e3:8.1f}"
+                f"+-{stats['e2e_s']['stdev'] * 1e3:6.1f} ms  (n={stats['e2e_s']['n']}, "
+                f"audio {stats['audio_s']['mean']:.1f}s)"
+            )
+
+    if failed:
+        sys.exit(1)
+
+
+if __name__ == "__main__":
+    main()