[CI/Build] Match perf warmups to benchmark concurrency

Use enough warmup requests to cover a full concurrent batch while preserving the existing two-request minimum.

Signed-off-by: Sy03 <24841857+Sy0307@users.noreply.github.com>

.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/test-amd-merge.yml

@@ -7,54 +7,64 @@ 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 --cov=vllm_omni --cov-report=term-missing:skip-covered --cov-report=xml:coverage-model-executor-$$BUILDKITE_PARALLEL_JOB.xml"
 
-    - label: "Simple · Diffusion & Model Executor Test (2/2)"
-      agent_pool: mi325_1
+    - label: "Simple · Diffusion 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
-        # 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 --cov=vllm_omni --cov-report=term-missing:skip-covered --cov-report=xml:coverage-diffusion-$$BUILDKITE_PARALLEL_JOB.xml"
+
+    - 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' --cov=vllm_omni --cov-report=term-missing:skip-covered --cov-report=xml"
 
     - 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 --cov=vllm_omni --cov-report=term-missing:skip-covered --cov-report=xml"
 
 - 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
@@ -62,7 +72,7 @@ steps:
         - export VLLM_LOGGING_LEVEL=DEBUG
         - export VLLM_WORKER_MULTIPROC_METHOD=spawn
         - |
-          timeout 20m bash -c '
+          bash -c '
             set +e
             pytest -s -v tests/diffusion/offloader/test_diffusion_cpu_offload.py
             EXIT1=\$?
@@ -72,7 +82,7 @@ steps:
           '
 
     - label: "Diffusion Sequence Parallelism Test (Need 4 GPUs)"
-      agent_pool: mi325_4
+      agent_pool: mi300_4
       depends_on: amd-build
       mirror_hardwares: [amdproduction]
       grade: Blocking
@@ -84,7 +94,7 @@ steps:
 
     # 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 +102,7 @@ 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 +110,7 @@ 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 +118,7 @@ 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 +131,7 @@ 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 +142,7 @@ 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 +157,7 @@ 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:
@@ -155,7 +165,7 @@ 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
@@ -163,22 +173,23 @@ steps:
         - 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'
 
-    - label: "Qwen3-TTS CustomVoice E2E Test"
-      agent_pool: mi325_1
+    - label: "Qwen3-TTS CustomVoice E2E Test · Shard %N/%t"
+      agent_pool: mi300_1
+      parallelism: 2
       depends_on: amd-build
       mirror_hardwares: [amdproduction]
       grade: Blocking
       commands:
         - |
-          timeout 50m bash -c '
+          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"
+            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" --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --shard-id=$$BUILDKITE_PARALLEL_JOB
           '
 
     - label: "Qwen3-TTS Base E2E Test"
-      agent_pool: mi325_1
+      agent_pool: mi300_1
       depends_on: amd-build
       mirror_hardwares: [amdproduction]
       grade: Blocking
@@ -196,7 +207,7 @@ steps:
 # 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 +219,7 @@ 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 +231,7 @@ 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-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 --cov=vllm_omni --cov-report=term-missing:skip-covered --cov-report=xml:coverage-model-executor-$$BUILDKITE_PARALLEL_JOB.xml"
+
+    - 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 --cov=vllm_omni --cov-report=term-missing:skip-covered --cov-report=xml"
+
+    - 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' --cov=vllm_omni --cov-report=term-missing:skip-covered --cov-report=xml"
 
     - 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 --cov=vllm_omni --cov-report=term-missing:skip-covered --cov-report=xml"
 
 - 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
@@ -96,7 +109,7 @@ steps:
         - timeout 20m pytest -s -v tests/diffusion/distributed/test_sequence_parallel.py -m core_model
 
     - label: "Diffusion GPU Worker Test"
-      agent_pool: mi325_2
+      agent_pool: mi300_2
       depends_on: amd-build
       mirror_hardwares: [amdproduction]
       grade: Blocking
@@ -104,7 +117,7 @@ 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
@@ -112,7 +125,7 @@ steps:
         - timeout 15m pytest -sv tests/diffusion/models/ -m 'core_model and cuda' --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 +139,7 @@ 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
@@ -137,7 +150,7 @@ 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
@@ -149,7 +162,7 @@ steps:
       '
 
 # - 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 +175,7 @@ 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
@@ -173,7 +186,7 @@ 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
@@ -186,7 +199,7 @@ steps:
 # 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 +211,7 @@ 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 +223,7 @@ 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

@@ -8,7 +8,12 @@
 # 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 CI image tag …:$BUILDKITE_COMMIT-a3)
+#   a3_npu_2    — ascend-a3, resource_class npu-2
+#
+# The ``l4_*`` presets run under the ``docker`` plugin, which does not put
+# ``buildkite-agent`` on PATH inside the container the way the ``kubernetes``
+# presets do, so they set ``mount-buildkite-agent: true``; without it any step
+# calling ``buildkite-agent`` (artifact upload, annotate, meta-data) fails there.
 
 x-npu-k8s-plugins: &npu_k8s_plugins
   - kubernetes:
@@ -26,6 +31,7 @@ mirror_hardwares:
           image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
           always-pull: true
           propagate-environment: true
+          mount-buildkite-agent: true
           shm-size: "8gb"
           environment:
             - HF_HOME=/fsx/hf_cache
@@ -41,6 +47,7 @@ mirror_hardwares:
           image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
           always-pull: true
           propagate-environment: true
+          mount-buildkite-agent: true
           shm-size: "8gb"
           environment:
             - HF_HOME=/fsx/hf_cache

.buildkite/common/scripts/run_cov_split.sh

@@ -0,0 +1,82 @@
+#!/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.
+#
+# 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 && ${#ONLINE[@]} == 0)); then
+    echo "run_cov_split.sh: need at least one of --offline / --online" >&2
+    exit 2
+fi
+
+# Two steps can test the same model, and artifacts are scoped to the build (a
+# nightly one carries the ready, merge and nightly tiers at once), so unsuffixed
+# reports would share a path -- a glob download then races to rename onto one file.
+# Not BUILDKITE_JOB_ID: the step id survives retries, so attempts collapse under
+# Buildkite's latest-attempt lookup instead of accumulating. A parallel step would
+# need BUILDKITE_PARALLEL_JOB as well; no pilot is one.
+REPORT_SUFFIX="${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/cuda/test-merge.yml

@@ -76,6 +76,9 @@ steps:
           - 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/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
         commands:
@@ -92,31 +95,24 @@ steps:
           - 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/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
+          - .buildkite/common/scripts/run_cov_split.sh
+          - pyproject.toml
+        # 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"
@@ -124,6 +120,10 @@ steps:
           - 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/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
         timeout_in_minutes: 50
@@ -135,10 +135,20 @@ steps:
         source_file_dependencies:
           - tests/e2e/offline_inference/test_minicpmo_4_5.py
           - tests/e2e/online_serving/test_minicpmo_4_5.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
         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).
         mirror_hardwares: h100_2
 
       - label: "Omni · MiniCPM-o 4.5 Duplex Test"
@@ -148,65 +158,38 @@ steps:
           - 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/minicpmo_realtime_duplex_scenarios.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/helpers/minicpmo_4_5_duplex.py
+          - vllm_omni/deploy/minicpmo_4_5.yaml
+          - 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/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/core/sched/
           - vllm_omni/outputs/multimodal_accumulation.py
           - vllm_omni/engine/stage_init_utils.py
           - vllm_omni/engine/async_omni_engine.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
         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_2
-
-      - 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/
+          - vllm_omni/diffusion/models/dmd2/
+          - vllm_omni/diffusion/distributed/autoencoders/autoencoder_kl_qwenimage.py
         commands:
           - |
             timeout 40m bash -c '
@@ -215,23 +198,12 @@ 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/
-        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/
+          - vllm_omni/diffusion/models/dmd2/
         commands:
           - |
             timeout 40m bash -c "
@@ -239,21 +211,3 @@ steps:
               pytest -s -v tests/e2e/offline_inference/test_wan22_t2v.py tests/e2e/online_serving/test_wan22_t2v.py -m 'advanced_model and cuda and diffusion' --run-level 'advanced_model'
             "
         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
-        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
-        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

.buildkite/cuda/test-nightly.yml

@@ -22,7 +22,7 @@ steps:
         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_2
+        mirror_hardwares: h100_1
 
       - label: ":full_moon: Omni · Doc Test with L4"
         timeout_in_minutes: 90
@@ -50,6 +50,19 @@ steps:
             exit $$EXIT
         mirror_hardwares: h100_2
 
+      - label: ":full_moon: Omni · MiniCPM-o 4.5 · Accuracy Test"
+        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
+            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"
         key: nightly-omni-performance-no-async-chunk
         timeout_in_minutes: 300
@@ -102,6 +115,32 @@ steps:
             exit $$EXIT
         mirror_hardwares: h100_3
 
+      - label: ":full_moon: Omni · MiniCPM-o 4.5 · Perf Test"
+        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
+            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"
+        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
+            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"
         timeout_in_minutes: 45
         commands:
@@ -120,19 +159,7 @@ steps:
           - 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
-
-      - 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'
-            "
+          - pytest -s -v tests/e2e/ -m "full_model and L4 and tts" --run-level "full_model" --ignore=tests/e2e/accuracy
         mirror_hardwares: l4_1
 
       - label: ":full_moon: TTS · Perf Test"
@@ -149,35 +176,6 @@ steps:
             exit $$EXIT
         mirror_hardwares: h100_2
 
-      - label: ":full_moon: TTS · VoxCPM2 · Perf Test"
-        key: nightly-tts-performance-voxcpm2
-        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
-            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
     depends_on: upload-nightly-pipeline
@@ -207,8 +205,7 @@ steps:
       build.pull_request.labels includes "nightly-test" ||
       build.pull_request.labels includes "diffusion-x2iat-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.
+      # Use explicit file shards so single-card cases do not reserve 2+ GPUs.
 
       - label: ":full_moon: Diffusion X2I(&A&T) · Function Test with H100 · Single-GPU"
         timeout_in_minutes: 120
@@ -216,15 +213,6 @@ steps:
           - >-
             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"
             --run-level "full_model"
         mirror_hardwares: h100_1
@@ -235,47 +223,14 @@ steps:
           - >-
             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"
             --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"
-        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
-
-      - 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
+          - pytest -sv tests/e2e/ -k "not test_wan 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"
@@ -288,7 +243,7 @@ steps:
         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_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
 
@@ -306,7 +261,7 @@ steps:
         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
@@ -316,70 +271,6 @@ steps:
             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
-        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_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-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
@@ -401,12 +292,6 @@ steps:
           - 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
 
-      - label: ":full_moon: Diffusion X2V · LingBot Video MoE Function Test"
-        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
-
       - label: ":full_moon: Diffusion X2V · Doc Test"
         timeout_in_minutes: 60
         commands:
@@ -417,7 +302,6 @@ steps:
         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
 
       - label: ":full_moon: Diffusion X2V · Wan2.2 Accuracy Test"
@@ -427,10 +311,10 @@ steps:
           - 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
 
-      - label: ":full_moon: Diffusion X2V · MiniMax H3 T2VA Accuracy Test"
+      - label: ":full_moon: Diffusion X2V · MiniMax H3 I2VA/Ref2VA Accuracy Test"
         timeout_in_minutes: 180
         commands:
-          - pytest -s -v tests/e2e/accuracy/minimax_h3/test_minimax_h3_t2va_similarity.py -m full_model --run-level full_model
+          - 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
 
       - label: ":full_moon: Diffusion X2V · Other Accuracy Test"
@@ -438,8 +322,6 @@ steps:
         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
 
       - label: ":full_moon: Diffusion X2V · Perf Test"
@@ -517,16 +399,6 @@ steps:
           - 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
-
   # NOTE: ":email: Nightly Collection & Email" is deprecated and will be removed soon.
   # Do not add new depends_on entries or extend this step.
   - label: ":email: Nightly Collection & Email"
@@ -537,32 +409,20 @@ steps:
       - 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
     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-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
       - 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

@@ -128,6 +128,21 @@ steps:
               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:
@@ -154,6 +169,10 @@ steps:
         source_file_dependencies:
           - tests/e2e/online_serving/test_qwen3_omni.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
         commands:
@@ -166,24 +185,34 @@ steps:
       - label: "Omni · MiniCPM-o 4.5 Offline Test"
         source_file_dependencies:
           - tests/e2e/offline_inference/test_minicpmo_4_5.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
         commands:
           - |
             timeout 20m bash -c "
               pytest -s -v tests/e2e/offline_inference/test_minicpmo_4_5.py -m 'core_model' --run-level 'core_model'
             "
-        mirror_hardwares: h100_2
+        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/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
         commands:
           - |
             timeout 20m bash -c "
               pytest -s -v tests/e2e/online_serving/test_minicpmo_4_5.py -m 'core_model' --run-level 'core_model'
             "
-        mirror_hardwares: h100_2
+        mirror_hardwares: h100_1
 
       - label: "Omni · MiniCPM-o 4.5 Duplex Test"
         source_file_dependencies:
@@ -192,45 +221,53 @@ steps:
           - 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/minicpmo_realtime_duplex_scenarios.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/helpers/minicpmo_4_5_duplex.py
+          - vllm_omni/deploy/minicpmo_4_5.yaml
+          - 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/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/core/sched/
           - vllm_omni/outputs/multimodal_accumulation.py
           - vllm_omni/engine/stage_init_utils.py
           - vllm_omni/engine/async_omni_engine.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
         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'
             "
-        mirror_hardwares: h100_2
+        mirror_hardwares: h100_1
 
-      - label: "TTS · VoxCPM2 Test"
+      - label: "Omni · MiniCPM-o 4.5 Perf Test"
+        timeout_in_minutes: 60
         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
+          - tests/dfx/perf/tests/test_minicpmo_4_5_ready.json
+          - vllm_omni/deploy/minicpmo_4_5.yaml
+          - vllm_omni/model_executor/models/minicpmo_4_5/
+          - vllm_omni/model_executor/stage_input_processors/minicpmo_4_5_omni.py
         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/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
         commands:
@@ -245,6 +282,9 @@ steps:
         source_file_dependencies:
           - tests/e2e/online_serving/test_qwen3_tts_base.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
         commands:
@@ -257,52 +297,11 @@ steps:
             "
         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/
+          - vllm_omni/diffusion/models/dmd2/
         commands:
           - |
             timeout 40m bash -c "
@@ -315,6 +314,9 @@ steps:
         source_file_dependencies:
           - tests/e2e/online_serving/test_cosmos3.py
           - 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
         commands:
           - |
             timeout 40m bash -c "
@@ -327,6 +329,8 @@ steps:
         source_file_dependencies:
           - tests/e2e/online_serving/test_qwen_image.py
           - vllm_omni/diffusion/models/qwen_image/
+          - vllm_omni/diffusion/models/dmd2/
+          - vllm_omni/diffusion/distributed/autoencoders/autoencoder_kl_qwenimage.py
         commands:
           - |
             timeout 40m bash -c "
@@ -334,15 +338,3 @@ steps:
               pytest -s -v tests/e2e/online_serving/test_qwen_image.py -m 'core_model' --run-level 'core_model'
             "
         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 'core_model' --run-level 'core_model'
-            "
-        mirror_hardwares: h100_1

.buildkite/cuda/test-weekly.yml

@@ -30,14 +30,6 @@ steps:
       - 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"
-    mirror_hardwares: l4_1
-
   - group: "Reliability Test - Invalid parameters Test"
     depends_on: upload-weekly-pipeline
     if: build.env("WEEKLY") == "1" || build.pull_request.labels includes "weekly-test"
@@ -108,6 +100,22 @@ steps:
             --run-level "full_model"
         mirror_hardwares: h100_1
 
+      # 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: 60
+        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
+            -m "tts and cuda"
+            --run-level "full_model"
+        mirror_hardwares: l4_1
+
       - label: "Diffusion · L4"
         timeout_in_minutes: 60
         commands:
@@ -133,4 +141,4 @@ steps:
             pytest -sv
             tests/e2e/ -m "slow and diffusion and H100 and distributed_cuda"
             --run-level "full_model"
-        mirror_hardwares: h100_2
+        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: "v0.27.0"
         priority: 100
         timeout_in_minutes: 60
         soft_fail: false

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

@@ -34,8 +34,23 @@ docker_build() {
     docker build "${out[@]}" "$@" -f docker/Dockerfile.xpu .
 }
 
+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" \
+        "${vllm_source_dir}"
+)
+
 if [ -z "$(docker images -q "${base_image_name}")" ]; then
-    docker_build "${base_image_name}" --target vllm-base --build-arg "VLLM_VERSION=${VLLM_VERSION}"
+    build_vllm_base
 fi
 
 # Try building the docker image
@@ -72,8 +87,9 @@ 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/test-npu-nightly.yml

@@ -39,7 +39,7 @@ steps:
             EXIT=$$?
             buildkite-agent artifact upload "tests/dfx/perf/results/*.json"
             exit $$EXIT
-      - label: ":robot: Omni · MiniCPM-o 4.5 Test"
+      - label: ":full_moon: Omni · MiniCPM-o 4.5 Test"
         mirror_hardwares: a3_npu_2
         env:
           VLLM_WORKER_MULTIPROC_METHOD: spawn
@@ -47,6 +47,40 @@ steps:
           HF_TOKEN: "${HF_TOKEN}"
         commands:
           - pytest -s -v tests/e2e/offline_inference/test_minicpmo_4_5.py tests/e2e/online_serving/test_minicpmo_4_5.py tests/e2e/online_serving/test_minicpmo_4_5_expansion.py -m 'full_model' --run-level 'full_model'
+      - label: ":full_moon: Omni · MiniCPM-o 4.5 · Duplex Test"
+        timeout_in_minutes: 180
+        mirror_hardwares: a3_npu_2
+        env:
+          VLLM_WORKER_MULTIPROC_METHOD: spawn
+          VLLM_ALLOW_LONG_MAX_MODEL_LEN: "1"
+          HF_TOKEN: "${HF_TOKEN}"
+        commands:
+          - >-
+            pytest -s -v
+            tests/e2e/online_serving/test_minicpmo_4_5_duplex.py
+            tests/e2e/online_serving/test_minicpmo_4_5_duplex_expansion.py
+            -m "npu and A3 and omni"
+            --run-level full_model
+      - label: ":full_moon: Omni · MiniCPM-o 4.5 · Accuracy Test"
+        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:
+          # 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
   - group: ":card_index_dividers: TTS Model Test"
     key: nightly-tts-test-group
     depends_on: upload-nightly-pipeline
@@ -68,33 +102,6 @@ steps:
             EXIT=$$?
             buildkite-agent artifact upload "tests/dfx/perf/results/*.json"
             exit $$EXIT
-  - 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"
-    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

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

@@ -1,24 +1,42 @@
 steps:
+  - label: "Platforms Test"
+    key: ready-platforms-test
+    depends_on: upload-ready-pipeline
+    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'
+
   - group: ":card_index_dividers: Omni Model Test"
     key: ready-omni-group
     depends_on: upload-ready-pipeline
     steps:
-      - label: ":robot: Omni · MiniCPM-o 4.5 Test"
-        agents:
-          queue: "ascend-a3"
-          resource_class: "npu-2"
-        image: "swr.cn-southwest-2.myhuaweicloud.com/modelfoundry/vllm-omni-ci-npu:${BUILDKITE_COMMIT}-a3"
-        plugins:
-          - kubernetes:
-              podSpecPatch:
-                imagePullSecrets:
-                  - name: swr-secret
+      - label: "Omni · MiniCPM-o 4.5 Test"
+        mirror_hardwares: a3_npu_2
         env:
           VLLM_WORKER_MULTIPROC_METHOD: spawn
           VLLM_ALLOW_LONG_MAX_MODEL_LEN: "1"
-          HF_TOKEN: "${HF_TOKEN}"
         commands:
-          - pytest -s -v tests/e2e/offline_inference/test_minicpmo_4_5.py tests/e2e/online_serving/test_minicpmo_4_5.py -m 'core_model' --run-level 'core_model'
+          - pytest -s -v tests/e2e/offline_inference/test_minicpmo_4_5.py tests/e2e/online_serving/test_minicpmo_4_5.py -m 'core_model and npu and A3 and omni' --run-level 'core_model'
+      - label: "Omni · MiniCPM-o 4.5 Duplex Test"
+        mirror_hardwares: a3_npu_2
+        env:
+          VLLM_WORKER_MULTIPROC_METHOD: spawn
+          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

.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/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 +65,8 @@ 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.
+
 ### Step 5: Print Report
 
 ```
@@ -77,6 +79,7 @@ Pre-check report for <branch>
   ─────────────────  ──────
   PR title format    ✓
   Code quality       ⚠ 1 broad except, 2 Any hints
+  Examples policy    ✗ new model-specific Python example
   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:

.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/readme.md

@@ -1,9 +1,9 @@
-# Claude Skills for vLLM-Omni
+# Repository Skills for vLLM-Omni
 
-This directory contains Claude Code skills maintained for the `vllm-omni`
-repository. These skills capture repeatable workflows for common contributor
-tasks such as model integration, pull request review, and release note
-generation.
+This directory contains repository-scale skills maintained for `vllm-omni`.
+They capture repeatable workflows for common contributor and maintainer tasks
+such as model integration, CI-aligned testing, performance optimization, and
+pull request review.
 
 ## Directory Structure
 
@@ -14,31 +14,41 @@ include:
 - `references/`: focused reference material used by the skill
 - `scripts/`: small helper scripts used by the skill
 
+## Using the Skills
+
+Coding agents that discover repository skills can invoke the relevant skill by
+name. For other agents, point them to the skill's `SKILL.md` and ask them to
+read it before changing code. Combine a domain skill with `vllm-omni-test` when
+tests or CI coverage change, then use `precheck-pr` for the contributor's final
+self-check.
+
 ## Available Skills
 
-- `add-diffusion-model`: guides integration of a new diffusion model into
-  `vllm-omni`
-- `diffusion-perf-opt`: guides diffusion model performance optimization,
-  including profiling traces, parallel strategies, stage timing analysis, and
-  benchmark-driven tuning
-- `quantization`: guides quantization method selection, model integration,
-  checkpoint loading, and quality/performance validation for vLLM-Omni
-- `add-omni-model`: covers addition of new omni-modality model support
-- `add-tts-model`: covers integration of new TTS models and related serving
-  workflows
-- `generate-release-note`: helps prepare release notes for repository changes
-- `precheck-pr`: self-check a branch before creating a PR — validates PR title
-  format, catches dead code, verifies accuracy/perf claims, and confirms merge
-  readiness
-- `vllm-omni-test`: guides generation and execution of CI-aligned tests (L1–L4),
-  pytest marker selection (`core_model` / `advanced_model` / `full_model`,
-  `omni` / `tts` / `diffusion`), Buildkite wiring (`test-ready.yml`,
-  `test-merge.yml`, `test-nightly.yml`, `test-weekly.yml`), and copy-paste
-  local plus CI-like `pytest` commands; see `references/test-routing.md` for
-  level-to-command mapping
-- `review-pr`: provides a structured workflow for reviewing pull requests
-- `vllm-omni-npu-model-runner-upgrade`: upgrades NPU model runners to align with the
-  latest vllm-ascend NPUModelRunner
+- [`add-diffusion-model`](add-diffusion-model/SKILL.md): guides integration of
+  a new diffusion model into `vllm-omni`
+- [`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
+- [`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
+- [`quantization`](quantization/SKILL.md): guides quantization method selection,
+  model integration, checkpoint loading, and quality/performance validation
+  for vLLM-Omni
+- [`review-pr`](review-pr/SKILL.md): provides a frozen-snapshot, contract-aware
+  workflow for maintainers and reviewers; contributors should use
+  `precheck-pr` for self-review
+- [`vllm-omni-npu-model-runner-upgrade`](vllm-omni-npu-upgrade/SKILL.md):
+  upgrades NPU model runners to align with the latest vllm-ascend
+  `NPUModelRunner`
+- [`vllm-omni-test`](vllm-omni-test/SKILL.md): guides generation and execution
+  of CI-aligned tests (L1–L4), pytest marker selection (`core_model` /
+  `advanced_model` / `full_model`, `omni` / `tts` / `diffusion`), Buildkite
+  wiring (`test-ready.yml`, `test-merge.yml`, `test-nightly.yml`,
+  `test-weekly.yml`), and copy-paste local plus CI-like `pytest` commands; see
+  `references/test-routing.md` for level-to-command mapping
 
 ## Maintenance Guidelines
 

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

@@ -97,6 +97,7 @@ If docs and live code disagree, verify the code/tests and report the drift.
 | [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`. |
 
 ### Delivery and reviewer coordination
 
@@ -152,6 +153,13 @@ 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

.gitignore

@@ -46,6 +46,8 @@ htmlcov/
 .cache
 nosetests.xml
 coverage.xml
+coverage-*.xml
+coverage-*.xml.gz
 *.cover
 *.py,cover
 .hypothesis/
@@ -252,6 +254,7 @@ outputs/
 results/
 generated/
 output_*/
+/tests/e2e/accuracy/artifacts/
 
 # Configuration overrides
 configs/local.yaml

.pre-commit-config.yaml

@@ -66,7 +66,7 @@ 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

.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

README.md

@@ -16,6 +16,7 @@ Easy, fast, and cheap omni-modality model serving for everyone
 ---
 
 *Latest News* 🔥
+- [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.

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/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/tts/README.md

@@ -139,6 +139,34 @@ 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

benchmarks/tts/bench_tts.py

@@ -84,12 +84,17 @@ def build_bench_args(
     result_filename: str | None,
     extra_cli_args: list[str],
     output_len: int | None = None,
+    served_model_name: str | None = None,
+    num_warmups: int = 2,
+    request_seed: int | None = None,
 ) -> list[str]:
     """Build the ``vllm bench serve --omni`` command for one (task, concurrency) run."""
     dataset_name = _TASK_TO_DATASET[task]
     backend: str = model_cfg["backend"]
     endpoint: str = model_cfg["endpoint"]
-    task_extra_body: dict[str, Any] = (model_cfg.get("task_extra_body") or {}).get(task) or {}
+    task_extra_body: dict[str, Any] = dict((model_cfg.get("task_extra_body") or {}).get(task) or {})
+    if request_seed is not None:
+        task_extra_body["seed"] = request_seed
 
     # Resolve dataset path
     if dataset_path:
@@ -109,7 +114,7 @@ def build_bench_args(
         "--port",
         str(port),
         "--model",
-        model,
+        served_model_name or model,
         "--backend",
         backend,
         "--endpoint",
@@ -119,7 +124,7 @@ def build_bench_args(
         "--num-prompts",
         str(num_prompts),
         "--num-warmups",
-        "2",
+        str(num_warmups),
         "--percentile-metrics",
         "ttft,e2el,audio_rtf,audio_ttfp,audio_duration,audio_underrun",
     ]
@@ -152,7 +157,8 @@ def build_bench_args(
         if result_filename:
             cmd += ["--result-filename", result_filename]
 
-    cmd += extra_cli_args
+    passthrough = extra_cli_args[1:] if extra_cli_args[:1] == ["--"] else extra_cli_args
+    cmd += passthrough
     return cmd
 
 
@@ -225,6 +231,12 @@ def main() -> None:
     parser.add_argument(
         "--model", required=True, help="HuggingFace model ID (e.g. Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice)"
     )
+    parser.add_argument(
+        "--served-model-name",
+        default=None,
+        help="Model name/path accepted by the running server. Use this when --model selects a registry entry "
+        "but the server was launched from a local bundle path.",
+    )
     parser.add_argument("--task", default="all", help="Task type: voice_clone | default_voice | voice_design | all")
     parser.add_argument("--locale", default="en", choices=["en", "zh"])
     parser.add_argument("--concurrency", type=int, nargs="+", default=[1, 4], metavar="N")
@@ -236,6 +248,13 @@ def main() -> None:
         metavar="N",
         help="Number of prompts per run. If one value, applied to all concurrency levels.",
     )
+    parser.add_argument("--num-warmups", type=int, default=2, help="Warmup requests before each measured run")
+    parser.add_argument(
+        "--request-seed",
+        type=int,
+        default=None,
+        help="Optional model sampling seed added to every request body for reproducible quality runs",
+    )
     parser.add_argument(
         "--dataset-path", default=None, help="Root of seed-tts-eval dataset (required for voice_clone/default_voice)"
     )
@@ -256,6 +275,9 @@ def main() -> None:
     parser.add_argument("extra", nargs=argparse.REMAINDER, help="Extra args passed directly to vllm bench serve")
     args = parser.parse_args()
 
+    if args.num_warmups < 0:
+        parser.error("--num-warmups cannot be negative")
+
     model_configs = load_model_configs(Path(args.model_configs))
     if args.model not in model_configs:
         known = "\n  ".join(model_configs.keys())
@@ -308,6 +330,9 @@ def main() -> None:
                 result_filename=result_filename,
                 extra_cli_args=args.extra or [],
                 output_len=args.output_len,
+                served_model_name=args.served_model_name,
+                num_warmups=args.num_warmups,
+                request_seed=args.request_seed,
             )
             result = run_one_benchmark(cmd)
             if result is not None:

benchmarks/tts/model_configs.yaml

@@ -7,6 +7,17 @@
 # is tracked here.
 
 models:
+  IndexTeam/IndexTTS-2.5:
+    supported_tasks: [voice_clone]
+    backend: openai-audio-speech
+    endpoint: /v1/audio/speech
+    trust_remote_code: true
+    task_extra_body:
+      voice_clone:
+        extra_params:
+          lang: en
+          text_normalization: true
+
   # -CustomVoice checkpoints lack speaker_encoder weights, so voice_clone is
   # NOT supported (an attempt raises ValueError from _extract_speaker_embedding
   # at model runtime).  Use -Base for voice_clone.

docker/Dockerfile.ci

@@ -1,7 +1,24 @@
-# Target vllm version: v0.26.0 (releases/v0.26.0 branch).
+# Target vllm version: v0.27.0 (releases/v0.27.0 branch).
 # Keep VLLM_BASE_TAG aligned with the target vLLM release on every rebase.
+#
+# Release-boundary form: there is no precompiled-wheel pin here, because the
+# base image already ships exactly the target vLLM. Everything the runtime
+# needs — torch, flashinfer, nvJitLink, numpy — comes from that image as one
+# self-consistent set.
+#
+# This block reappears only when the rebase targets a vLLM commit that is NOT a
+# released tag. During the 0.27 rebase, targeting v0.27.0rc1 on top of the
+# v0.26.0 base image required reinstalling vLLM from wheels.vllm.ai and then
+# hand-repairing the fallout: pinning torch ahead of flashinfer (installing it
+# after bound flashinfer's TVM FFI runtime to the old torch and aborted engine
+# core init), registering libnvJitLink.so.13 with ldconfig, and re-pinning numpy
+# under numba's ceiling. All of it was compensation for a base image and a vLLM
+# that disagreed. At a real release boundary they agree, so none of it is
+# needed. If you pin an untagged commit again, restore the reinstall block —
+# agent/tasks/20_update_wheel_url.sh fails the run rather than letting the
+# Dockerfile silently ship the wrong vLLM.
 ARG VLLM_BASE_IMAGE=vllm/vllm-openai
-ARG VLLM_BASE_TAG=v0.26.0
+ARG VLLM_BASE_TAG=v0.27.0
 
 # We use registry-based layer caching for the dependencies,
 # where the tag is determined based on the hash of the contents

docker/Dockerfile.npu

@@ -1,5 +1,5 @@
 ARG VLLM_ASCEND_IMAGE=quay.io/atlas-ci/vllm-ascend
-ARG VLLM_ASCEND_TAG=v0.26.0
+ARG VLLM_ASCEND_TAG=v0.27.0
 FROM ${VLLM_ASCEND_IMAGE}:${VLLM_ASCEND_TAG}
 
 # WORKDIR /vllm-workspace/vllm

docker/Dockerfile.npu.a3

@@ -1,5 +1,5 @@
 ARG VLLM_ASCEND_IMAGE=quay.io/atlas-ci/vllm-ascend
-ARG VLLM_ASCEND_TAG=v0.26.0-a3
+ARG VLLM_ASCEND_TAG=v0.27.0-a3
 FROM ${VLLM_ASCEND_IMAGE}:${VLLM_ASCEND_TAG}
 
 # WORKDIR /vllm-workspace/vllm

docker/Dockerfile.npu.ci

@@ -1,5 +1,5 @@
 ARG VLLM_ASCEND_IMAGE=quay.io/atlas-ci/vllm-ascend
-ARG VLLM_ASCEND_TAG=v0.26.0
+ARG VLLM_ASCEND_TAG=v0.27.0
 FROM ${VLLM_ASCEND_IMAGE}:${VLLM_ASCEND_TAG}
 
 ARG APP_DIR=/vllm-workspace/vllm-omni
@@ -45,6 +45,7 @@ RUN source /usr/local/Ascend/ascend-toolkit/set_env.sh && \
       "hyperpyyaml>=1.2.0" \
       "onnx>=1.16.0" \
       "librosa>=0.10.0" \
+      "datasets>=3.0.0" \
       --index-url http://cache-service.nginx-pypi-cache.svc.cluster.local/pypi/simple \
       --trusted-host cache-service.nginx-pypi-cache.svc.cluster.local \
       --default-timeout=60 && \

docker/Dockerfile.npu.ci.a3

@@ -1,5 +1,5 @@
 ARG VLLM_ASCEND_IMAGE=quay.io/atlas-ci/vllm-ascend
-ARG VLLM_ASCEND_TAG=v0.26.0-a3
+ARG VLLM_ASCEND_TAG=v0.27.0-a3
 FROM ${VLLM_ASCEND_IMAGE}:${VLLM_ASCEND_TAG}
 
 ARG APP_DIR=/vllm-workspace/vllm-omni
@@ -46,6 +46,7 @@ RUN source /usr/local/Ascend/ascend-toolkit/set_env.sh && \
       "onnx>=1.16.0" \
       "librosa>=0.10.0" \
       "torchmetrics>=1.9.0" \
+      "datasets>=3.0.0" \
       --index-url http://cache-service.nginx-pypi-cache.svc.cluster.local/pypi/simple \
       --trusted-host cache-service.nginx-pypi-cache.svc.cluster.local \
       --default-timeout=60 && \

docker/Dockerfile.rocm

@@ -1,4 +1,4 @@
-ARG BASE_IMAGE=vllm/vllm-openai-rocm:v0.26.0
+ARG BASE_IMAGE=vllm/vllm-openai-rocm:v0.27.0
 FROM ${BASE_IMAGE} AS base
 
 # Declare a variable to know if we want to use the nightly build or the stable build.
@@ -10,7 +10,7 @@ FROM ${BASE_IMAGE} AS base
 #    we should swap over to use stable release ASAP.
 #    We should avoid relying on custom commits.
 ARG USE_NIGHTLY_BUILD=0
-ARG VLLM_VERSION_OR_COMMIT_HASH=v0.26.0
+ARG VLLM_VERSION_OR_COMMIT_HASH=v0.27.0
 ARG ARG_PYTORCH_ROCM_ARCH
 ENV PYTORCH_ROCM_ARCH=${ARG_PYTORCH_ROCM_ARCH:-${PYTORCH_ROCM_ARCH}}
 
@@ -79,9 +79,16 @@ RUN printf '%s\n' \
     '    os._exit(_exit_code)' \
     > ${COMMON_WORKDIR}/vllm-omni/conftest.py
 
+# Some optional dependencies pull in the CPU ONNX Runtime distribution. The
+# CPU and ROCm wheels own the same import path and cannot coexist, so remove
+# the CPU wheel and reinstall the ROCm wheel after dependency resolution.
 RUN cd ${COMMON_WORKDIR}/vllm-omni && \
     uv pip install --python "$(python3 -c 'import sys; print(sys.executable)')" \
         --no-cache-dir ".[dev]" --no-build-isolation && \
+    python3 -m pip uninstall -y onnxruntime && \
+    python3 -m pip install --no-cache-dir --force-reinstall --no-deps \
+        "onnxruntime-rocm>=1.22.2" && \
+    python3 -c "import importlib.metadata as m; import onnxruntime as ort; assert 'ROCMExecutionProvider' in ort.get_available_providers(), ort.get_available_providers(); assert m.version('onnxruntime-rocm'); assert not any(d.metadata['Name'] == 'onnxruntime' for d in m.distributions())" && \
     python3 -c "from torchcodec.decoders import VideoDecoder"
 
 RUN ln -sf /usr/bin/python3 /usr/bin/python

docker/Dockerfile.xpu

@@ -88,7 +88,7 @@ ENV UV_EXTRA_INDEX_URL=${PIP_EXTRA_INDEX_URL}
 ENV UV_INDEX_STRATEGY="unsafe-best-match"
 ENV UV_LINK_MODE="copy"
 
-ARG VLLM_VERSION=v0.26.0
+ARG VLLM_VERSION=v0.27.0
 RUN git clone -b ${VLLM_VERSION} https://github.com/vllm-project/vllm
 WORKDIR /workspace/vllm
 
@@ -108,7 +108,7 @@ ENV VLLM_WORKER_MULTIPROC_METHOD=spawn
 
 RUN --mount=type=cache,target=/root/.cache/uv \
     uv pip uninstall triton triton-xpu && \
-    uv pip install triton-xpu==3.7.1 && \
+    uv pip install triton-xpu==3.7.2 && \
     uv pip install --no-build-isolation --no-deps .
 
 CMD ["/bin/bash"]
@@ -144,12 +144,13 @@ WORKDIR /workspace/vllm-omni
 COPY . .
 
 ENV VLLM_OMNI_TARGET_DEVICE=xpu
+# Keep the upstream oneCCL packages paired with the XPU torch runtime.
 RUN --mount=type=cache,target=/root/.cache/uv \
     uv pip install --no-build-isolation \
     --extra-index-url https://mirrors.ustc.edu.cn/pypi/simple/ \
     ".[dev]" && \
-    uv pip uninstall triton oneccl oneccl-devel && \
-    uv pip install triton-xpu==3.7.1 --reinstall
+    uv pip uninstall triton && \
+    uv pip install triton-xpu==3.7.2 --reinstall
 
 FROM vllm-omni AS vllm-omni-openai
 

docs/.nav.yml

@@ -40,30 +40,13 @@ nav:
   - Models:
     - models/supported_models.md
   - Features:
-    - Sleep Mode: features/sleep_mode.md
-    - Diffusion Features:
-      - Overview: user_guide/diffusion_features.md
-      - Feature Compatibility: user_guide/feature_compatibility.md
-      - Regional Compilation: user_guide/diffusion/regional_compilation.md
-      - Cache Acceleration:
-        - TeaCache: user_guide/diffusion/cache_acceleration/teacache.md
-        - Cache-DiT: user_guide/diffusion/cache_acceleration/cache_dit.md
-      - Attention Backends: user_guide/diffusion/attention_backends.md
-      - Frame Interpolation: user_guide/diffusion/frame_interpolation.md
-      - Parallelism:
-        - Overview: user_guide/diffusion/parallelism/overview.md
-        - CFG Parallel: user_guide/diffusion/parallelism/cfg_parallel.md
-        - Expert Parallel: user_guide/diffusion/parallelism/expert_parallel.md
-        - Hybrid Sharded Data Parallel: user_guide/diffusion/parallelism/hsdp.md
-        - Sequence Parallel: user_guide/diffusion/parallelism/sequence_parallel.md
-        - Tensor Parallel: user_guide/diffusion/parallelism/tensor_parallel.md
-        - VAE Parallelism: user_guide/diffusion/parallelism/vae_parallelism.md
-      - CPU Offloading: user_guide/diffusion/cpu_offload.md
-      - LoRA: user_guide/diffusion/lora.md
-      - Custom Pipeline: features/custom_pipeline.md
-      - Execution Modes: user_guide/diffusion/execution_modes.md
+    - Overview: features/README.md
+    - Runtime and Stage Execution:
+      - Execution Modes and Streaming: user_guide/diffusion/execution_modes.md
+      - Prefill-Decode Disaggregation (experimental): features/pd_disaggregation.md
+      - Sleep Mode: features/sleep_mode.md
     - Quantization:
-      - Quantization: user_guide/quantization/overview.md
+      - Overview: user_guide/quantization/overview.md
       - Online Quantization: user_guide/quantization/online.md
       - Quantized KV Cache: user_guide/quantization/quantized_kvcache.md
       - FP8 W8A8: user_guide/quantization/fp8.md
@@ -72,9 +55,43 @@ nav:
       - MXFP8 W8A8: user_guide/quantization/mxfp8.md
       - MXFP4 W4A4: user_guide/quantization/mxfp4.md
       - ModelOpt: user_guide/quantization/modelopt.md
-      - GGUF: user_guide/quantization/gguf.md
       - AutoRound: user_guide/quantization/autoround.md
       - msModelSlim: user_guide/quantization/msmodelslim.md
+    - Diffusion Acceleration:
+      - Overview and Compatibility:
+        - Overview: user_guide/diffusion_features.md
+        - Feature Compatibility: user_guide/feature_compatibility.md
+      - CPU Offloading:
+        - Overview: user_guide/diffusion/cpu_offload.md
+        - Model-Level: user_guide/diffusion/offloader/cpu_offload.md
+        - Layerwise: user_guide/diffusion/offloader/layerwise_offload.md
+        - Distributed Layerwise: user_guide/diffusion/offloader/distributed_layerwise_offload.md
+      - Cache Acceleration:
+        - TeaCache: user_guide/diffusion/cache_acceleration/teacache.md
+        - Cache-DiT: user_guide/diffusion/cache_acceleration/cache_dit.md
+      - Parallelism:
+        - Overview: user_guide/diffusion/parallelism/overview.md
+        - CFG Parallel: user_guide/diffusion/parallelism/cfg_parallel.md
+        - Expert Parallel: user_guide/diffusion/parallelism/expert_parallel.md
+        - Hybrid Sharded Data Parallel: user_guide/diffusion/parallelism/hsdp.md
+        - Sequence Parallel: user_guide/diffusion/parallelism/sequence_parallel.md
+        - Tensor Parallel: user_guide/diffusion/parallelism/tensor_parallel.md
+        - VAE Parallelism: user_guide/diffusion/parallelism/vae_parallelism.md
+      - Attention Backends:
+        - Overview: user_guide/diffusion/attention_backends.md
+        - Dense Backends: user_guide/diffusion/attention_backends/dense_backends.md
+        - TRTLLM: user_guide/diffusion/attention_backends/trtllm.md
+        - SageAttention: user_guide/diffusion/attention_backends/sage.md
+        - Hugging Face Hub: user_guide/diffusion/attention_backends/huggingface_hub.md
+        - RainFusion: user_guide/diffusion/attention_backends/rainfusion.md
+      - Regional Compilation: user_guide/diffusion/regional_compilation.md
+      - Frame Interpolation: user_guide/diffusion/frame_interpolation.md
+      - Startup and Loading: user_guide/diffusion/startup_and_loading.md
+      - LoRA: user_guide/diffusion/lora.md
+    - Experimental:
+      - Session State Manager: features/session_state_manager.md
+  - Integrations:
+    - Custom Diffusion Pipeline: features/custom_pipeline.md
     - ComfyUI: features/comfyui.md
   - Fault Tolerance:
     - Failure Modes: user_guide/fault_injection_reliability_matrix.md
@@ -100,6 +117,7 @@ nav:
         - Async Diffusion Output: design/feature/async_diffusion_output.md
         - Async Omni Output Materialization: design/feature/omni_async_output_materialization.md
         - Automatic Prefix Caching: design/feature/prefix_caching.md
+        - Realtime AR-Diffusion: design/feature/realtime_ar_diffusion.md
       - Communication:
         - OmniConnector Implementations:
           - Mooncake Store Connector: design/feature/omni_connectors/mooncake_store_connector.md
@@ -108,6 +126,7 @@ nav:
           - Shared Memory Connector: design/feature/omni_connectors/shared_memory_connector.md
           - Yuanrong Store Connector: design/feature/omni_connectors/yuanrong_connector.md
           - Yuanrong Transfer Engine Connector: design/feature/omni_connectors/yuanrong_transfer_engine_connector.md
+      - Quantization: design/feature/quantization.md
       - Diffusion Acceleration:
         - Parallelism:
           - CFG-Parallel: design/feature/cfg_parallel.md
@@ -117,13 +136,17 @@ nav:
           - Sequence Parallel: design/feature/sequence_parallel.md
           - Tensor Parallel: design/feature/tensor_parallel.md
           - VAE Patch Parallelism: design/feature/vae_parallel.md
-        - Attention Backends:
+        - Attention Optimization:
+          - Backend Selection: design/feature/attention_backend_selection.md
           - Skip-Softmax: design/feature/skip_softmax.md
-        - Quantization: design/feature/quantization.md
+        - CPU Offloading:
+          - Overview: design/feature/offloader/README.md
+          - Model-Level: design/feature/offloader/cpu_offload.md
+          - Layerwise: design/feature/offloader/layerwise_offload.md
+          - Distributed Layerwise: design/feature/offloader/distributed_layerwise_offload.md
         - Cache-DiT: design/feature/cache_dit.md
         - TeaCache: design/feature/teacache.md
         - Diffusion Continuous Batching: design/feature/diffusion_continuous_batching.md
-        - Distributed Layerwise Offload: design/feature/distributed_layerwise_offload.md
     - Infrastructure and Performance:
       - Prometheus Metrics: design/metrics.md
       - Speech Generation Performance Optimizations: design/qwen3_omni_tts_performance_optimization.md
@@ -157,7 +180,8 @@ nav:
 - CLI Reference: cli
 - Community:
   - Governance: community/governance.md
-  - community/*
+  - Contact Us: community/contact_us.md
+  - Community Meetings: community/meetings.md
   - Slack: https://slack.vllm.ai
   - Blog: https://blog.vllm.ai
   - Forum: https://discuss.vllm.ai

docs/community/contact_us.md

@@ -1,5 +1,9 @@
 # Contact Us
 
-- For technical questions and feature requests, please use GitHub [Issues](https://github.com/vllm-project/vllm-omni/issues)
-- For coordinating contributions and development and discussing with other users and developers, please join `sig-omni` channel in our [Slack](https://slack.vllm.ai/) or use the [vLLM Forum](https://discuss.vllm.ai/)
-- For security disclosures, please use GitHub's [Security Advisories](https://github.com/vllm-project/vllm-omni/security/advisories) feature
+- For bugs and feature requests, search or open a GitHub [issue](https://github.com/vllm-project/vllm-omni/issues).
+- For technical and usage questions, join the `sig-omni` channel in [vLLM Slack](https://slack.vllm.ai/) or use the [vLLM Forum](https://discuss.vllm.ai/).
+- For coordinating contributions and development, use `sig-omni`, the forum, or join a [community meeting](meetings.md).
+
+## Security Disclosures
+
+Do not share vulnerability details in public GitHub issues, pull requests, discussions, or Slack channels. To arrange a private report, send a direct message to one of the [lead maintainers](governance.md#lead-maintainers) in [vLLM Slack](https://slack.vllm.ai/). If you cannot send a direct message, you may post a detail-free request in `sig-omni` asking a lead maintainer to contact you privately; do not include any vulnerability details in that request.

docs/community/governance.md

@@ -52,31 +52,11 @@ Committers have write access and merge rights. They typically have deep expertis
 - [@ZJY0516](https://github.com/ZJY0516): Diffusion attention backend, kernel fusion, and CustomOp
 - [@yuanheng-zhao](https://github.com/yuanheng-zhao): Diffusion cache, offload, and Omni Support
 
-## Diffusion Workload Division
+## Reviewer Routing
 
-This section breaks down diffusion-related responsibilities. If you have PRs touching these areas, please ping the listed owners for review.
+For path-based reviewer routing, consult [`.github/CODEOWNERS`](https://github.com/vllm-project/vllm-omni/blob/main/.github/CODEOWNERS). Detailed subsystem boundaries and safe-change guidance live in the [module design documents](../design/index.md#module-design-documents).
 
-### Runtime-related
-
-- Diffusion models: @RuixiangMa, @wtomin, @xuechendi
-
-### Optimization
-
-- Parallel: @RuixiangMa, @wtomin
-- Attention backend: @ZJY0516
-- Cache: @yuanheng-zhao, @RuixiangMa, @SamitHuang
-- Offload: @yuanheng-zhao
-- Intel GPU support: @xuechendi
-- Quantization: @lishunyang12, @david6666666
-- Kernel fusion / communication-computation: @ZJY0516
-
-### Docs & Test
-
-- Diffusion docs: @RuixiangMa, @wtomin
-
-## Meetings
-
-Committers hold **bi-weekly meetings** to discuss future directions and collaborations of the project.
+`CODEOWNERS` is operational routing metadata, not a governance or approval policy. Being listed does not grant committer status or merge rights, and it does not mean every listed reviewer must approve a change. The maintainer roster above is authoritative for project roles and merge rights.
 
 ## Committer Nomination Process
 

docs/community/meetings.md

@@ -0,0 +1,8 @@
+# Community Meetings
+
+vLLM-Omni hosts developer-facing meetings for Chinese- and English-language audiences. Schedules and access details can change, so use these maintained documents as the source of truth:
+
+- [Chinese-language meeting document](https://docs.google.com/document/d/1pdUBiS_7mdOUNDtdwy-9OUf7jsMWN324BIbps5olDME/edit?tab=t.0)
+- [English-language meeting document](https://docs.google.com/document/d/19AlD08T8uJ7wqfs8VtYCi0iGMTwlTYVVlAxFZ3S7OIg/edit?tab=t.0)
+
+Each document contains the current schedule, access details, agendas, and past meeting notes. For broader vLLM community events, see the [vLLM events calendar](https://vllm.ai/events/).

docs/community/meetups.md

@@ -1 +0,0 @@
-# Meetups

docs/community/volunteers.md

@@ -1,12 +0,0 @@
-# Volunteers for Bugfix and CI
-
-We encourage you to check current docs and [issues](https://github.com/vllm-project/vllm-omni/issues) to find possible solutions for your questions. If non of these can solve it, please propose an issue to describe your questions about bug or CI problems for developing.
-
-If you have urgent need for locating and solving bugfix or CI problems, please find community volunteers below.
-
-| Dec 4-Dec 12 | Dec 15-Dec 19 | Dec 22-Dec 26 | Dec 29- Jan 2, 2026| Jan 5-Jan 9 | Jan 12-Jan 16 |
-|----------|----------|----------|----------|----------|----------|
-| <a href="https://github.com/congw729">Conw729</a> | <a href="https://github.com/yinpeiqi">yinpeiqi</a> | <a href="https://github.com/tzhouam">tzhouam</a> | <a href="https://github.com/SamitHuang">SamitHuang</a> | <a href="https://github.com/gcanlin">gcanlin</a> | <a href="https://github.com/natureofnature">natureofnature</a> |
-| <a href="https://github.com/david6666666">david6666666</a> | <a href="https://github.com/R2-Y">R2-Y</a> | <a href="https://github.com/hsliuustc0106">hsliuustc0106</a> | <a href="https://github.com/Gaohan123">Gaohan123</a> | <a href="https://github.com/ZJY0516">ZJY0516</a> | <a href="https://github.com/qibaoyuan">qibaoyuan</a> |
-
-We kindly welcome more contributors to fix bugs and contribute new features!

docs/configuration/README.md

@@ -14,10 +14,6 @@ For introduction, please check [Introduction for stage config](./stage_configs.m
 
 - **[GPU Memory Calculation and Configuration](./gpu_memory_utilization.md)** - Guide on how to calculate memory requirements and set up `gpu_memory_utilization` for optimal performance
 
-## Multi-Stage Recipes
-
-- **[Prefill-Decode Disaggregation](./pd_disaggregation.md)** - How to derive a PD-aware Qwen3-Omni stage config from the default config without introducing another bundled YAML
-
 ## Optimization Features
 
 - **[Diffusion Features Overview](../user_guide/diffusion_features.md)** - Complete overview of all diffusion model features and supported models

docs/configuration/pd_disaggregation.md

@@ -1,171 +0,0 @@
-# Prefill-Decode (PD) Disaggregation
-
-PD disaggregation splits the Qwen3-Omni thinker into separate prefill and decode
-stages so prompt processing and token generation can run on different workers.
-
-This is documented as a stage-config recipe instead of a bundled YAML because the
-deployment-specific values usually change per environment:
-
-- GPU placement
-- `tensor_parallel_size`
-- connector backend and connector ports
-- connector IPs or bootstrap addresses
-
-Start from the [default Qwen3-Omni stage config](gh-file:vllm_omni/deploy/qwen3_omni_moe.yaml)
-and copy it to your own file, for example `qwen3_omni_pd.yaml`. Then apply the
-changes below.
-
-## Requirements
-
-- 3+ GPUs for a basic layout: prefill, decode, and talker+code2wav
-- A KV connector supported by vLLM, such as `MooncakeConnector`
-- Matching `tensor_parallel_size` on the prefill and decode thinker stages
-
-## 1. Split the thinker into prefill and decode stages
-
-Replace the original thinker stage with two stages:
-
-```yaml
-stage_args:
-  - stage_id: 0
-    stage_type: llm
-    is_prefill_only: true
-    runtime:
-      devices: "0"
-    engine_args:
-      max_num_seqs: 16
-      model_stage: thinker
-      model_arch: Qwen3OmniMoeForConditionalGeneration
-      worker_type: ar
-      scheduler_cls: vllm_omni.core.sched.omni_ar_scheduler.OmniARScheduler
-      gpu_memory_utilization: 0.9
-      enforce_eager: true
-      trust_remote_code: true
-      engine_output_type: latent
-      distributed_executor_backend: "mp"
-      enable_prefix_caching: false
-      max_num_batched_tokens: 32768
-      hf_config_name: thinker_config
-      tensor_parallel_size: 1
-      kv_transfer_config:
-        kv_connector: "MooncakeConnector"
-        kv_role: "kv_producer"
-        kv_rank: 0
-        kv_parallel_size: 2
-        kv_connector_extra_config:
-          mooncake_bootstrap_port: 25201
-    final_output: false
-    is_comprehension: true
-    default_sampling_params:
-      temperature: 0.4
-      top_p: 0.9
-      top_k: 1
-      max_tokens: 2048
-      seed: 42
-      detokenize: True
-      repetition_penalty: 1.05
-
-  - stage_id: 1
-    stage_type: llm
-    is_decode_only: true
-    runtime:
-      devices: "1"
-    engine_args:
-      max_num_seqs: 64
-      model_stage: thinker
-      model_arch: Qwen3OmniMoeForConditionalGeneration
-      worker_type: ar
-      scheduler_cls: vllm_omni.core.sched.omni_ar_scheduler.OmniARScheduler
-      gpu_memory_utilization: 0.9
-      enforce_eager: true
-      trust_remote_code: true
-      engine_output_type: latent
-      distributed_executor_backend: "mp"
-      enable_prefix_caching: false
-      max_num_batched_tokens: 32768
-      hf_config_name: thinker_config
-      tensor_parallel_size: 1
-      kv_transfer_config:
-        kv_connector: "MooncakeConnector"
-        kv_role: "kv_consumer"
-        kv_rank: 1
-        kv_parallel_size: 2
-        kv_connector_extra_config:
-          mooncake_bootstrap_port: 25202
-    engine_input_source: [0]
-    final_output: true
-    final_output_type: text
-    is_comprehension: true
-    default_sampling_params:
-      temperature: 0.4
-      top_p: 0.9
-      top_k: 1
-      max_tokens: 2048
-      seed: 42
-      detokenize: True
-      repetition_penalty: 1.05
-```
-
-Notes:
-
-- `is_prefill_only: true` marks the thinker stage that only saves KV.
-- `is_decode_only: true` marks the thinker stage that resumes from remote KV.
-- `kv_transfer_config` is required on both stages.
-- The orchestrator forces the prefill stage to run with `max_tokens=1`, so the
-  prefill side only processes the prompt and exports KV.
-
-## 2. Shift the downstream stages by one index
-
-After inserting the extra thinker stage, renumber the remaining stages:
-
-```yaml
-  - stage_id: 2
-    runtime:
-      devices: "2"
-    engine_input_source: [1]
-    sync_process_input_func: vllm_omni.model_executor.stage_input_processors.qwen3_omni.thinker2talker_token_only
-
-  - stage_id: 3
-    runtime:
-      devices: "2"
-    engine_args:
-      max_num_seqs: 1
-    engine_input_source: [2]
-    sync_process_input_func: vllm_omni.model_executor.stage_input_processors.qwen3_omni.talker2code2wav_token_only
-```
-
-Compared with the default Qwen3-Omni config:
-
-- the talker becomes stage `2` instead of stage `1`
-- the code2wav stage becomes stage `3` instead of stage `2`
-- the talker now reads from decode stage `1`
-
-## 3. Add runtime edges for the four-stage pipeline
-
-```yaml
-runtime:
-  enabled: true
-  edges:
-    - from: 0
-      to: 1
-    - from: 1
-      to: 2
-    - from: 2
-      to: 3
-```
-
-## 4. Launch with your custom config
-
-```bash
-vllm serve Qwen/Qwen3-Omni-30B-A3B-Instruct --omni --port 8091 \
-    --deploy-config /path/to/qwen3_omni_pd.yaml
-```
-
-## Operational Notes
-
-- `MooncakeConnector` does not support heterogeneous TP sizes across the PD
-  pair. Keep prefill and decode at the same `tensor_parallel_size`.
-- If the thinker requires TP=2, both thinker stages must use TP=2 and be given
-  separate GPU sets, for example `"0,1"` for prefill and `"2,3"` for decode.
-- Choose connector ports and addresses that match your deployment. The values
-  shown above are examples only.

docs/configuration/stage_configs.md

@@ -38,11 +38,12 @@ Each entry under `stages:` accepts any `StageDeployConfig` field directly (no ne
 |-------|------|----------|---------|-------------|
 | `stage_id` | int | required | — | Stage identity; matched against `PipelineConfig.stages[*].stage_id`. |
 | `max_num_seqs` | int | optional | `64` | Max concurrent sequences per stage. |
-| `gpu_memory_utilization` | float | optional | `0.9` | Per-stage memory budget. |
+| `gpu_memory_utilization` | float | optional | `0.9` | Per-stage total memory target; used for automatic KV-cache sizing. |
+| `kv_cache_memory_bytes` | int \| null | optional | `null` | Explicit per-rank KV-cache budget in bytes; overrides automatic sizing. |
 | `tensor_parallel_size` | int | optional | `1` | TP degree for this stage. |
 | `enforce_eager` | bool | optional | `false` | Disable CUDA graphs. |
-| `max_num_batched_tokens` | int | optional | `32768` | Prefill budget. |
-| `max_model_len` | int \| null | optional | `null` | Per-stage context length (auto-sets `VLLM_ALLOW_LONG_MAX_MODEL_LEN=1` when larger than HF default). |
+| `max_num_batched_tokens` | int | optional | `32768` | Per-stage prefill/token budget; also contributes to the native maximum in-flight token limit. |
+| `max_model_len` | int \| null | optional | `null` | Per-sequence context or KV length; `-1` enables native cache-capacity auto-fitting, while values above the HF default auto-set `VLLM_ALLOW_LONG_MAX_MODEL_LEN=1`. |
 | `async_scheduling` | bool \| null | optional | `null` | Per-stage async scheduling toggle. |
 | `devices` | str | optional | `"0"` | `CUDA_VISIBLE_DEVICES`-style device list. |
 | `output_connectors` | dict \| null | optional | `null` | Keyed by `to_stage_<n>`; values are names registered under top-level `connectors:`. |

docs/contributing/README.md

@@ -3,7 +3,7 @@
 Thank you for your interest in contributing to vLLM-Omni! This document provides guidelines and instructions for contributing.
 
 !!! note
-    We host weekly developer-facing online meetings to discuss milestones and updates **every Tuesday at 19:30 PDT**. Meeting link as well as the past meeting notes can be found [here](https://tinyurl.com/vllm-omni-meeting).
+    vLLM-Omni hosts developer-facing meetings for Chinese- and English-language audiences. See [Community Meetings](../community/meetings.md) for current schedules, access details, agendas, and past notes.
 
 ## Getting Started
 
@@ -76,12 +76,53 @@ Please refer to the [test instructions](./ci/test_execution_guide.md) for detail
 !!! warning
     Currently, not all unit tests pass when run on CPU platforms. If you don't have access to a GPU platform to run unit tests locally, rely on the continuous integration system to run the tests for now.
 
+### Using repository skills with coding agents
+
+vLLM-Omni maintains [repository-scale skills](https://github.com/vllm-project/vllm-omni/tree/main/.claude/skills) for common coding, testing, performance, and review workflows. These skills capture repository-specific structure, validation requirements, and evidence standards that a general coding agent may not know.
+
+If your coding agent discovers repository skills automatically, ask it to use the relevant skill by name. Otherwise, point it to the linked `SKILL.md` and ask it to read the file before changing code. Skills can be combined: use a domain skill for the implementation, `vllm-omni-test` for test and CI decisions, and `precheck-pr` before opening the pull request.
+
+| Task | Repository skill |
+| --- | --- |
+| Add or extend a diffusion model | [`add-diffusion-model`](https://github.com/vllm-project/vllm-omni/blob/main/.claude/skills/add-diffusion-model/SKILL.md) |
+| Add or extend a text-to-speech model | [`add-tts-model`](https://github.com/vllm-project/vllm-omni/blob/main/.claude/skills/add-tts-model/SKILL.md) |
+| Add, select, or debug quantization | [`quantization`](https://github.com/vllm-project/vllm-omni/blob/main/.claude/skills/quantization/SKILL.md) |
+| Diagnose or optimize diffusion performance | [`diffusion-perf-opt`](https://github.com/vllm-project/vllm-omni/blob/main/.claude/skills/diffusion-perf-opt/SKILL.md) |
+| Upgrade an NPU model runner | [`vllm-omni-npu-model-runner-upgrade`](https://github.com/vllm-project/vllm-omni/blob/main/.claude/skills/vllm-omni-npu-upgrade/SKILL.md) |
+| Add a regression test, choose L1-L4 coverage, or wire Buildkite | [`vllm-omni-test`](https://github.com/vllm-project/vllm-omni/blob/main/.claude/skills/vllm-omni-test/SKILL.md) |
+| Self-check a branch before opening a pull request | [`precheck-pr`](https://github.com/vllm-project/vllm-omni/blob/main/.claude/skills/precheck-pr/SKILL.md) |
+| Perform a maintainer or reviewer review of an existing pull request or local branch | [`review-pr`](https://github.com/vllm-project/vllm-omni/blob/main/.claude/skills/review-pr/SKILL.md) |
+
+Use the following workflow when collaborating with a coding agent:
+
+1. **Ground the task.** Provide the issue, RFC, or design document that defines the expected behavior. Ask the agent to state its assumptions and identify the affected module or feature contracts before editing code.
+2. **Select the narrowest applicable skills.** Start with the domain skill that owns the change. Add `vllm-omni-test` whenever production behavior or tests change. For a bug fix, ask for the smallest reproducer and a regression test tied to the original symptom.
+3. **Require executable evidence.** Ask the agent to run the narrowest relevant checks and report the exact commands and results. It must identify tests it could not run because of unavailable hardware, model weights, credentials, or dependencies; an unrun check is not a pass.
+4. **Inspect the resulting diff.** Confirm that the implementation stays within the issue or RFC, does not overwrite unrelated work, and includes the required tests, documentation, and benchmark or accuracy evidence.
+5. **Run the pre-submit workflow.** Use `precheck-pr`, then run the applicable local tests and pre-commit hooks before opening the pull request.
+
+For example:
+
+```text
+Use add-diffusion-model and vllm-omni-test to implement the linked issue.
+Read the relevant design documents first, keep the change within the issue,
+and report every test command, result, and hardware validation gap.
+```
+
+```text
+Use vllm-omni-test to turn this bug reproducer into the smallest deterministic
+regression test. Prefer L1 CPU coverage; if the failure requires real weights
+or serving, explain the required L2/L3 environment and provide the exact command.
+```
+
+Repository skills guide the agent, but they do not replace contributor judgment, the accepted issue or RFC, required test evidence, or maintainer review. Review all generated changes before submitting them.
+
 ## Issues
 
 If you encounter a bug or have a feature request, please search existing issues first to see if it has already been reported. If not, please file a new issue, providing as much relevant information as possible.
 
 !!! important
-    If you discover a security vulnerability, please report it by creating a GitHub issue with the `security` label.
+    Do not report suspected security vulnerabilities through a public issue, pull request, discussion, or Slack channel. Follow the [security disclosure instructions](../community/contact_us.md#security-disclosures) to arrange a private report.
 
 ## Pull Requests & Code Reviews
 

docs/contributing/ci/ci_settings.md

@@ -327,6 +327,99 @@ On PR builds, `upload_pipeline.py` logs `skip '…' (no changes under …)` for
 - Use **groups** for dashboard readability; keep **E2E Test** as the group name CUDA diff filtering expects for `--e2e` nightly runs on main.
 - Prefix labels by model domain: **Omni ·**, **TTS ·**, **Diffusion ·**, **Simple ·**, etc., matching existing steps.
 
+### Per-model coverage
+
+Pilot (v1): one Merge-tier (L3) job in `.buildkite/cuda/test-merge.yml` uploads
+per-model, per-entry-mode coverage as Buildkite artifacts — "TTS · Qwen3-TTS Base
+Test". It runs on a single GPU so the pilot is cheap to reproduce. (A second
+diffusion pilot, "Diffusion · Bagel Test", was retired from merge; reintroduce
+another diffusion coverage job if a second pilot is needed again.) This is a
+pilot, not a rollout: no dashboard/visualization lives in this repo, and nothing
+is gated on coverage.
+
+**Naming convention**: `coverage-<model_id>-<mode>-<step_id>.xml.gz`, where
+`<model_id>` is the model's directory name under
+`vllm_omni/diffusion/models/<model_id>/` or
+`vllm_omni/model_executor/models/<model_id>/` (e.g. `z_image`, `qwen3_omni`),
+`<mode>` is `online` or `offline`, and `<step_id>` is `BUILDKITE_STEP_ID`
+(`local` outside Buildkite). `gunzip` before feeding a report to a Cobertura
+consumer.
+
+Reports are gzipped because `--cov=vllm_omni` sets coverage's `source`, which
+makes it walk the package and emit one `<line>` element per statement for every
+file — including the ones a single model's tests never import. That inventory,
+not the coverage, is the size: a Bagel report is 7.1 MB, of which 96.5% is
+`<line>` elements and 77% belongs to the 840 of 1169 files that run never
+imported. A run covering zero lines produces the same 7 MB. gzip takes it to
+roughly 410 KB and changes nothing about the data.
+
+`<step_id>` is what keeps two steps that cover the same model apart. Artifacts are
+scoped to the build, and a nightly build carries the ready, merge and nightly tiers
+at once, so several same-model steps land in one namespace — `qwen3_tts` already has
+both a Base and a CustomVoice job, and `minicpmo_4_5` a base and a duplex one. Two
+uploads on one path do not fail loudly: an exact-name download can report the path
+as ambiguous, and a glob download fetches every match concurrently and renames each
+onto the same destination, so whichever finishes last wins. The step id remains the
+logical step identifier across retries, while Buildkite's default artifact lookup
+selects only the latest attempt — `BUILDKITE_JOB_ID` would instead give every
+attempt its own filename, leaving a failed attempt's partial report to be picked up
+alongside the real one. A step using `parallelism`/matrix would need
+`BUILDKITE_PARALLEL_JOB` added, since its jobs share a step id; neither pilot is
+one.
+
+**Opting in a new model**: replace the job's combined `pytest` command with
+[`run_cov_split.sh`](https://github.com/vllm-project/vllm-omni/blob/main/.buildkite/common/scripts/run_cov_split.sh),
+which runs each mode, names the reports, uploads them, and fails the step if
+either half or the upload failed:
+
+```yaml
+commands:
+  - |
+    .buildkite/common/scripts/run_cov_split.sh \
+      --model-id <model_id> \
+      --offline <offline test path> \
+      --online <online test path> \
+      -- <the job's existing pytest flags, e.g. -m '...' --run-level '...'>
+```
+
+Everything after `--` is forwarded verbatim to both runs with quoting preserved,
+so an expression like `-m 'advanced_model and cuda'` survives intact and a job can
+pass anything else it needs. `--offline`/`--online` are repeatable and only one is
+required, so a job with tests in a single mode runs just that half. Runtime is
+bounded solely by the step's `timeout_in_minutes` — raise it, since the split
+loads the model once per mode.
+
+Before splitting, confirm each half actually collects at least one test under the job's forwarded filters — pytest exits with code 5 ("no tests collected") if a half is empty, which the script reports as a failure, red-ing a job that used to pass as one combined invocation.
+
+Also check the tests' `num_cards` against the job's `mirror_hardwares` preset.
+`cuda_marks` turns `num_cards > 1` into `skipif(device_count() < num_cards)`, so a
+test asking for more GPUs than the preset provides is silently skipped and its
+coverage artifact comes back empty. Splitting such a job produces a file that
+looks fine and measures nothing — "Diffusion · Z Image Test" is the current
+example (online tests want 4 cards, the job runs on `l4_1`), which is why it is
+not a coverage pilot.
+
+Online-mode coverage depends on the `[tool.coverage.run]` `patch`/`sigterm`
+settings in `pyproject.toml`: online tests launch the server with
+`subprocess.Popen` and stop it with SIGTERM, and without those settings the XML
+reflects only the pytest parent process, not the server's code paths.
+
+The upload depends on `buildkite-agent` being callable inside the container. The
+`kubernetes` presets (`h100_*`, `*_npu_*`) provide it; the `docker` ones (`l4_*`)
+only do because they set `mount-buildkite-agent: true`. A new docker preset that
+runs a coverage job needs the same.
+
+List both `run_cov_split.sh` and `pyproject.toml` in every opted-in job's
+`source_file_dependencies` — both change what the job measures, so without them a
+change there is filtered out of normal PR builds and only surfaces in a later
+nightly. `tests/buildkite/test_upload_pipeline.py` covers the filter behavior with
+a synthetic job (it does not pin real merge labels).
+Editing only the surrounding CI YAML still does not schedule them, so a PR that
+touches just the wiring needs a full E2E run (or the commands run on a GPU host)
+to produce artifacts. When checking a new model's
+artifacts, compare `lines-covered` between the online and offline XML rather than
+just confirming both files exist.
+
 ### Validation checklist
 
 | Check | Command / location |

docs/contributing/ci/test_execution_guide.md

@@ -184,9 +184,6 @@ Failed jobs: 1/2
     # Reliability: HunyuanImage DiT (H100 × 4)
     pytest -s -v dfx/reliability/test_reliability_hunyuan_image.py -m slow
 
-    # Reliability: VoxCPM2 (L4)
-    pytest -s -v dfx/reliability/test_reliability_voxcpm2.py -m slow
-
     ```
 
     The latest L5 CI jobs (reliability + invalid-parameter weekly steps) are in [test-weekly.yml](https://github.com/vllm-project/vllm-omni/blob/main/.buildkite/cuda/test-weekly.yml).

docs/contributing/ci/test_writing_guide.md

@@ -562,7 +562,6 @@ L5 level testing focuses on the performance of model services under ***long-runn
     -   `test_reliability_qwen3_omni.py` — Qwen3-Omni chat / multimodal reliability (GPU OOM, process kill, recovery, error contract under `--async-chunk` vs default).
     -   `test_reliability_wan22.py` — Wan2.2 I2V video API reliability (`/v1/videos` under OOM and process kill, recovery).
     -   `test_reliability_hunyuan_image.py` — HunyuanImage-3.0-Instruct DiT-only reliability (`/v1/images/generations` under OOM and process kill; deploy `hunyuan_image3_dit.yaml`, H100 × 4).
-    -   `test_reliability_voxcpm2.py` — VoxCPM2 TTS reliability (`/v1/audio/speech` under OOM and process kill; deploy `voxcpm2.yaml`, L4 × 1).
     -   `helpers.py` — Shared primitives used by current suites: raw HTTP probes for `/v1/chat/completions` and `/health`, OpenAI-style error parsing, GPU OOM sidecar (`inject_gpu_oom` / `stop_gpu_oom_hogs`), and `pgrep`-based process-kill injector construction (`make_process_kill_fault_injector`).
     -   `conftest.py` — `fault_injector` and `omni_server_after_fault` / `omni_server_after_fault_function` fixtures to run a callable **after** the server is ready.
     -   `README.md` — Short local run commands for this directory.
@@ -662,13 +661,12 @@ Reliability tests are **short fault-injection** integration runs (L5 **(b)** in
 | `test_reliability_qwen3_omni.py` | Qwen3-Omni: OOM vs **text vs text+audio** error contract, large multimodal chat under OOM, concurrent pressure, **SIGKILL** on `VLLM::Worker`, `/health` → 503 + fast-fail + concurrent chat; optional OOM recovery scenario (may be skipped while tracked in issues). |
 | `test_reliability_wan22.py` | Wan2.2 I2V: large `/v1/videos` under OOM, **SIGKILL** on `multiprocessing.spawn` chain, health / fast-fail / concurrent video requests; optional recovery test (may be skipped). |
 | `test_reliability_hunyuan_image.py` | HunyuanImage DiT-only: large `/v1/images/generations` under OOM, **SIGKILL** on `vLLM-Omni::` workers and serve/tree targets, health / fast-fail / concurrent image requests; some OOM/recovery cases may be skipped while tracked in issues. |
-| `test_reliability_voxcpm2.py` | VoxCPM2: `/v1/audio/speech` under OOM (error contract), **SIGKILL** on `VLLM::` workers and serve/tree targets, health / fast-fail / concurrent speech requests; some OOM cases may be skipped while tracked in issues. |
 | `README.md` | Minimal run / collect examples. |
 
 ##### Parametrization and markers
 
 - Each test module defines a **`RELIABILITY_SCENARIOS`** list (`test_name`, `server_params`: model, `stage_config_name` or diffusion `server_args`, etc.). **`create_reliability_omni_server_params()`** in `tests/dfx/conftest.py` resolves stage paths (including XPU substitutions where applicable) and builds **`OmniServerParams`** lists consumed by **`@pytest.mark.parametrize(..., indirect=True)`** on `omni_server` or `omni_server_function`.
-- Cases are tagged **`@pytest.mark.slow`** for weekly / selective CI. GPU-heavy suites use **`@hardware_test(res={"cuda": "H100"}, num_cards=...)`** or **`@hardware_test(res={"cuda": "L4"}, num_cards=1)`** (Qwen3-Omni **2**× H100; Wan2.2 **1**× H100; HunyuanImage DiT **4**× H100; VoxCPM2 **1**× L4).
+- Cases are tagged **`@pytest.mark.slow`** for weekly / selective CI. GPU-heavy suites use **`@hardware_test(res={"cuda": "H100"}, num_cards=...)`** (Qwen3-Omni **2**× H100; Wan2.2 **1**× H100; HunyuanImage DiT **4**× H100).
 - **Process-kill** tests use **`@pytest.mark.skipif(os.name == "nt", ...)`** because injection uses POSIX **`pgrep` / `kill`**.
 
 ##### CI trigger
@@ -680,13 +678,11 @@ Weekly Buildkite (`.buildkite/cuda/test-weekly.yml`) runs one step per model sui
 | Reliability Test - qwen3-omni | `test_reliability_qwen3_omni.py` | H100 × 2 (`mithril-h100-pool`) |
 | Reliability Test - wan22 | `test_reliability_wan22.py` | H100 × 2 (`mithril-h100-pool`) |
 | Reliability Test - hunyuan-image | `test_reliability_hunyuan_image.py` | H100 × 4 (`mithril-h100-pool`) |
-| Reliability Test - voxcpm2 | `test_reliability_voxcpm2.py` | L4 × 1 (`gpu_1_queue`) |
 
 ```bash
 pytest -s -v tests/dfx/reliability/test_reliability_qwen3_omni.py -m "slow"
 pytest -s -v tests/dfx/reliability/test_reliability_wan22.py -m "slow"
 pytest -s -v tests/dfx/reliability/test_reliability_hunyuan_image.py -m "slow"
-pytest -s -v tests/dfx/reliability/test_reliability_voxcpm2.py -m "slow"
 ```
 
 ##### Local commands
@@ -696,7 +692,6 @@ pytest --collect-only tests/dfx/reliability
 pytest -s -v tests/dfx/reliability/test_reliability_qwen3_omni.py -m slow
 pytest -s -v tests/dfx/reliability/test_reliability_wan22.py -m slow
 pytest -s -v tests/dfx/reliability/test_reliability_hunyuan_image.py -m slow
-pytest -s -v tests/dfx/reliability/test_reliability_voxcpm2.py -m slow
 ```
 
 ##### Adding a new model suite
@@ -709,4 +704,4 @@ pytest -s -v tests/dfx/reliability/test_reliability_voxcpm2.py -m slow
 </details>
 
 -   -   ***Stability***: `pytest -s -v tests/dfx/stability/scripts/test_stability_qwen3_omni.py` or `pytest -s -v tests/dfx/stability/scripts/test_stability_wan22.py` (or add `test_stability_<model>.py` alongside a matching JSON config)
-    -   ***Reliability***: `pytest -s -v tests/dfx/reliability/test_reliability_<model>.py -m slow` (current suites: `qwen3_omni`, `wan22`, `hunyuan_image`, `voxcpm2`; add `test_reliability_<suite>.py` and a matching step in `.buildkite/cuda/test-weekly.yml` for new models)
+    -   ***Reliability***: `pytest -s -v tests/dfx/reliability/test_reliability_<model>.py -m slow` (current suites: `qwen3_omni`, `wan22`, `hunyuan_image`; add `test_reliability_<suite>.py` and a matching step in `.buildkite/cuda/test-weekly.yml` for new models)

docs/design/ar_diffusion_pipeline_capability.md

@@ -4,6 +4,9 @@
 > `experimental/ar_diffusion` contract. It follows the ownership and lifecycle
 > direction in [#5137](https://github.com/vllm-project/vllm-omni/issues/5137),
 > but it is not yet part of the stable module-design hierarchy.
+>
+> The end-to-end request, identity, failure, and memory contracts are documented
+> in [Realtime AR-Diffusion sessions](feature/realtime_ar_diffusion.md).
 
 `ARDiffusionEngine` is an opt-in diffusion backend for pipelines that keep
 autoregressive attention KV across requests. The engine only selects
@@ -51,14 +54,13 @@ time, and its inherited batch/step entry points also fail explicitly. Supporting
 those paths requires a batch-aware state-binding contract rather than silently
 bypassing `bind_ar_diffusion_state()`.
 
-The current implementation retains one session per model instance. A request
-for a different `session_id` closes the resident session through the same LRU
-lifecycle used by explicit close and failure cleanup before allocating the new
-one. Self-attention and cross-attention memory are therefore sized for one
-resident session even if the pipeline declares a larger `session_capacity`.
-This prioritizes rollout throughput and prevents block-pool exhaustion before a
-count-based eviction can run. Multi-session residency can be added later with a
-memory-backed capacity calculation or allocation-pressure eviction.
+The implementation may retain multiple sessions per model instance. It selects
+the largest feasible capacity up to the pipeline's `session_capacity`, based on
+self-attention pages, cross-attention allocations, scratch, and declared
+model-owned state. One session is admitted whenever it fits actual free device
+memory; `gpu_memory_fraction` controls additional residency. Requests beyond
+that capacity evict the least-recently-used session through the same lifecycle
+hook used by explicit close and failure cleanup.
 
 Scratch storage is also capability-driven. `frames_per_block` reserves space for
 an uncommitted current video block, while `max_scratch_tokens_per_branch`

docs/design/architecture_overview.md

@@ -406,7 +406,7 @@ the source for configuration, compatibility, and implementation details.
   stage contract.
 * **Quantization and memory efficiency:** [Quantization](feature/quantization.md)
   resolves per-pipeline or per-component quantization configurations, while
-  [Distributed Layerwise Offload](feature/distributed_layerwise_offload.md)
+  [Distributed Layerwise Offload](feature/offloader/distributed_layerwise_offload.md)
   streams diffusion blocks from host memory within the existing parallel
   topology.
 

docs/design/feature/async_diffusion_output.md

@@ -117,7 +117,22 @@ After (async D2H):
 
 ### Batch Split
 
-When `execute_model_batch` returns a `COMPUTE_DONE` with a single `async_output_id` for the entire batch, the executor splits it into per-request `async_output_id`s (format `{batch_id}/{request_id}`) via `_batch_split_map`. When `OUTPUT_READY` arrives, the pump extracts per-request results from the batch output and resolves each request's output future independently.
+When `execute_model_batch` returns a `COMPUTE_DONE` with a single `async_output_id` for the entire batch, the executor splits it into per-request `async_output_id`s (format `{batch_id}/{request_id}`) via `_batch_split_map`. When `OUTPUT_READY` arrives, the pump extracts per-request results from the batch output and resolves each request's output future independently. If `OUTPUT_READY` arrives before `execute_batch` registers the split map, `execute_batch` directly adopts the early output.
+
+### Reliability, Lifecycle & Timeout Behavior
+
+1. **Drain Before Memory Release (`drain_async_outputs` / `_ASYNC_OUTPUT_DRAIN_TIMEOUT_S`)**:
+   When workers transition into sleep states (`handle_sleep_task`) or invoke memory-releasing methods, `WorkerProc.drain_async_outputs()` blocks up to `_ASYNC_OUTPUT_DRAIN_TIMEOUT_S` (10.0s) until all in-flight background D2H/SHM packing tasks finish. This acts as a synchronization barrier preventing worker sleep cycles from releasing device tensors while the side CUDA stream is still actively reading them. Draining typically completes within 10–50 ms under normal operation.
+
+2. **Timeout Diagnostics (`_ASYNC_OUTPUT_TIMEOUT`)**:
+   The engine waits for `wait_output_ready()` with a timeout of `_ASYNC_OUTPUT_TIMEOUT` (30.0s). If a timeout occurs, the executor logs detailed diagnostic bookkeeping (pending futures, cached outputs, and batch split maps) to identify whether the stall occurred in the worker, pump, or waiter.
+
+3. **Queue Serialization & Atomic Future Resolution**:
+   - `result_mq` writes are serialized with `_result_mq_lock` to protect the single-writer `MessageQueue` against concurrent writes from the worker main loop (`COMPUTE_DONE`) and background thread (`OUTPUT_READY`).
+   - `ResultPumpThread` unpacks shared-memory payloads first, and then atomically resolves waiting futures or populates `_completed_outputs` under `_futures_lock`, avoiding race conditions with concurrent `wait_output_ready()` invocations.
+
+4. **Storage-Aware IPC Packing**:
+   `pack_diffusion_output_shm()` evaluates `max(view_bytes, storage_bytes)` against `_SHM_TENSOR_THRESHOLD` (1 MB) to ensure that per-request slices sharing a large batch storage are moved to POSIX shared memory rather than serialized through pickle over the MessageQueue inline buffer.
 
 ## Model Coverage
 
@@ -131,6 +146,7 @@ All models running in request-mode (`step_execution=False`, the default) automat
 |---|---|---|
 | **HunyuanImage-3.0** | Image | `False` |
 | **Qwen-Image** | Image | `True` |
+| **LTX-2.3** | Video / Audio | `False` |
 
 Other models with `step_execution=False` are also supported but not yet verified.
 

docs/design/feature/attention_backend_selection.md

@@ -0,0 +1,89 @@
+# Diffusion Attention Backend Selection
+
+This document defines the selection and extension contract for diffusion
+attention backends. User-facing backend choices, installation, and tuning are
+in the [attention backend guides](../../user_guide/diffusion/attention_backends.md).
+
+## Scope
+
+The contract applies to DiT and other diffusion attention layers. It is
+separate from vLLM's autoregressive attention selector.
+
+The selection path has four responsibilities:
+
+1. normalize user configuration into `AttentionConfig` and `AttentionSpec`;
+2. resolve a spec for an attention role;
+3. ask the active platform to validate an explicit backend or choose a
+   hardware default; and
+4. load the selected `AttentionBackend` class from the registry path.
+
+## Resolution contract
+
+`get_attn_backend_for_role()` resolves configuration in this order:
+
+1. exact `per_role[role]`;
+2. category `per_role[role_category]`;
+3. `default`;
+4. platform default.
+
+An explicit resolution returns both the backend class and its
+`AttentionSpec`. A platform-default resolution returns the class and `None`.
+Layers must therefore treat the spec as optional and must not infer that a
+missing spec means SDPA.
+
+Class resolution is cached by backend name, head size, and whether TRTLLM may
+be selected automatically. Logging is cached separately by role and source so
+the startup record identifies why each role received its backend.
+
+## Registry and platform boundary
+
+`DiffusionAttentionBackendEnum` maps stable configuration names to default
+qualified class paths. `register_diffusion_backend()` may replace a path at
+runtime without changing the public enum value.
+
+The active `OmniPlatform` owns compatibility policy through
+`get_diffusion_attn_backend_cls()`. It must:
+
+- validate explicit selections and fail with an actionable error when the
+  requested kernel cannot run;
+- choose only an available, compatible default when no backend is explicit;
+- account for head size and other platform-visible constraints; and
+- return a qualified class path implementing `AttentionBackend`.
+
+The selector must not duplicate device capability or package-availability
+policy that belongs to the platform.
+
+## Typed backend options
+
+Backend-specific settings remain typed fields on `AttentionSpec` rather than
+unstructured keyword dictionaries:
+
+- `quant` is consumed by FlashInfer and TRTLLM with backend-specific value
+  validation;
+- `skip_softmax` is consumed by TRTLLM; and
+- `block_sparse` is consumed by block-sparse backends such as RainFusion.
+
+A backend reads only the fields it owns and rejects incompatible values. New
+options should be added to a shared typed spec only when more than one backend
+shares their semantics; otherwise add a dedicated typed block.
+
+## Model integration contract
+
+Each diffusion attention call declares a stable role and, when appropriate, a
+broader category. Model-specific roles permit precise overrides; categories
+preserve common `self` or `cross` policies. A model also decides whether its
+path is compatible with automatic TRTLLM selection, because only the model
+knows whether masking and packing satisfy the kernel contract.
+
+## Adding or changing a backend
+
+An implementation change is complete only when it:
+
+1. implements the `AttentionBackend` interface and exposes a stable enum name;
+2. adds platform validation and default routing where appropriate;
+3. defines and validates any typed configuration;
+4. tests explicit selection, default selection, and incompatible paths;
+5. documents installation, fallback behavior, and quality implications in the
+   corresponding user guide; and
+6. updates the overview matrix without moving algorithm details back into the
+   landing page.

docs/design/feature/distributed_layerwise_offload.md

@@ -1,156 +0,0 @@
-# Distributed Layerwise Offload
-
-This document describes distributed layerwise offload (DLO) for diffusion
-models. DLO keeps only a small number of DiT blocks on the accelerator and
-streams the remaining blocks from host memory. The distributed backend can
-either shard those host-side weights across an existing parallel group or keep
-the standard loader's rank-local weights and avoid an additional collective.
-
-For user-facing commands, see the
-[distributed layerwise offloading guide](../../user_guide/diffusion/cpu_offload.md)
-and the [Cosmos3 recipe](https://github.com/vllm-project/vllm-omni/blob/main/recipes/cosmos3/Cosmos3-DistOffload.md).
-
-## Status
-
-DLO is implemented for multi-device diffusion execution. The default
-AllGather path is the primary path for DP and SP deployments. The
-`--dlo-no-use-allgather` path is a rank-local compatibility mode: it is useful
-for standard-loader sharding, workstation bring-up, and systems where an
-additional DLO collective is undesirable, but it does not reduce host weight
-storage across ranks.
-
-The compatibility matrix below describes the current implementation. The
-unit-level guards are covered, but not every parallelism combination has a
-full model-and-hardware end-to-end test.
-
-## Design
-
-### DLO consumes the existing parallel topology
-
-DLO does not create a new DP, TP, or SP topology. It reads the configured
-`DiffusionParallelConfig` and attaches offload hooks to the DiT blocks after the
-standard distributed groups have been initialized.
-
-The DLO weight-sharding group is selected as follows:
-
-1. Use the existing DP group when `data_parallel_size > 1`.
-2. When DP is one and SP is greater than one, use the SP group.
-3. Otherwise, run rank-locally without a DLO process group.
-
-TP is deliberately not used as DLO's AllGather group. HSDP has its own
-parameter-sharding lifecycle and is not allowed to be sharded a second time by
-DLO's AllGather path.
-
-### AllGather path
-
-With the default `dlo_use_allgather=True`, each rank stores approximately
-`1 / group_size` of each streamable block in pinned host memory. The next
-block's shard is copied to a device buffer and reconstructed with
-`all_gather_into_tensor` on a communication stream while the current block is
-executing.
-
-```text
-Compute:    [Block N]             [Block N+1]          [Block N+2]
-H2D:                      [shard N+1]           [shard N+2]
-AllGather:                [full N+1]             [full N+2]
-Buffers:    [current slot]       [prefetch slot]       [current slot]
-```
-
-![DLO double-buffer prefetch pipeline](../figures/dlo/dlo_pipeline.gif)
-
-The backend uses two shared device buffers, so accelerator weight residency is
-bounded by the largest streamed blocks rather than the complete model.
-
-When DP is greater than one, the engine can process one request per DP rank in
-the same denoising wave. Because AllGather is a collective, all participating
-requests must take the same execution path at every denoising step.
-
-### Rank-local path without DLO AllGather
-
-With `--dlo-no-use-allgather`, DLO forces its internal offload shard size to
-one. The regular model loader remains responsible for preparing each rank's
-weights, including TP-local tensors or HSDP-managed parameters. DLO then
-streams those rank-local tensors block by block using H2D copies only.
-
-This mode means:
-
-- DP still provides independent replicas, but DLO does not shard weights
-  across DP ranks.
-- SP still performs its normal activation/attention collectives, but DLO does
-  not shard weights across SP ranks.
-- TP/HSDP/SP collectives, if configured, are not disabled by this flag; only
-  DLO's additional weight AllGather is disabled.
-- Pure DP deployments keep a full host-side model copy per rank, subject to
-  shared OS page-cache behavior.
-- The scheduler does not require a synchronized DP request wave for DLO.
-
-This path is intentionally not implemented by reusing the AllGather mmap
-loader. It relies on the standard loader so model-specific TP/HSDP transforms
-remain intact.
-
-## Parallelism compatibility
-
-| Parallelism | DLO + AllGather | DLO without AllGather |
-|---|---|---|
-| **DP** | Supported primary path. DLO shards host weights across the DP group and can run DP multi-concurrency. | Supported rank-local path. DP replicas remain independent; no DLO weight sharding or cross-DP DLO collective. |
-| **SP** | Supported in the implementation. With DP=1, DLO uses the SP group for host-weight sharding; SP still shards sequence/activation work. | SP remains active, but DLO keeps standard-loader rank-local weights and adds no SP weight collective. |
-| **TP > 1** | Unsupported in the DLO mmap path. TP-aware loader callbacks are bypassed, so the backend rejects this configuration when it enters that path. | Standard loading is retained and TP-local tensors can be streamed. This is the intended compatibility path, but it still needs broader model and hardware validation. |
-| **HSDP** | Rejected. HSDP has already sharded parameters, so DLO AllGather would double-shard them. | Accepted by configuration. HSDP owns parameter sharding and its own gathers; DLO only stages rank-local parameters. End-to-end coverage is limited. |
-
-### Combined dimensions
-
-- **DP + SP:** DLO uses the DP group for weight sharding when DP is greater
-  than one; SP continues to use its own sequence-parallel group. If DP is one,
-  the SP group becomes DLO's sharding group in AllGather mode.
-- **DP + TP/SP without AllGather:** standard model loading defines the
-  rank-local tensor layout. DLO adds no cross-DP, cross-TP, or cross-SP weight
-  collective.
-- **HSDP + SP:** the general parallel configuration permits HSDP over SP, but
-  DLO must use `--dlo-no-use-allgather`. HSDP remains responsible for weight
-  materialization and synchronization.
-- **HSDP + DP or TP:** rejected independently by the diffusion parallel
-  configuration.
-
-## Request and loading constraints
-
-AllGather DP multi-concurrency requires:
-
-- explicit `num_inference_steps`;
-- the same `num_inference_steps` for all requests in a wave; and
-- identical request arguments that affect the collective execution path.
-
-The no-AllGather path does not impose these DLO-specific synchronized-wave
-requirements.
-
-The mmap loader is used only by the supported DLO+AllGather path when the
-model has a compatible checkpoint layout. Online quantization is incompatible
-with that sharded mmap path; use `--dlo-no-use-allgather` or disable online
-quantization. Models that do not meet the mmap requirements use the regular
-loader path.
-
-## Validation coverage
-
-Current source-level validation includes:
-
-- HSDP + DLO + AllGather rejection;
-- HSDP + DLO without AllGather acceptance at configuration level;
-- TP rejection in the DLO+AllGather mmap path;
-- resident-layer requests requiring no-AllGather;
-- DP request-wave validation for denoising-step compatibility;
-- sharding, double-buffer, AllGather-size, and heterogeneous-block regression
-  tests.
-
-The highest-value missing coverage is end-to-end numerical comparison against
-ordinary layerwise offload for DP+SP, TP+no-AllGather, and HSDP+SP+no-AllGather
-on the target CUDA/NCCL or CANN/HCCL hardware.
-
-## Recommendations
-
-- Use **DP + DLO AllGather** for the supported throughput and host-memory
-  scaling path.
-- Use **SP + DLO AllGather** for long-sequence workloads when DP concurrency is
-  not the goal.
-- Use **no-AllGather** to bring up TP or workstation PCIe configurations, with
-  the expectation of higher host-memory use and lower validation confidence.
-- Prefer **HSDP alone** for production HSDP deployments until the combined
-  HSDP + DLO no-AllGather path has broader end-to-end coverage.

docs/design/feature/offloader/README.md

@@ -0,0 +1,80 @@
+# CPU Offloading
+
+This document defines the shared architecture for diffusion CPU offloading.
+The three strategies have separate user and design pages:
+
+| Strategy | User guide | Design contract |
+| --- | --- | --- |
+| Model-level | [Guide](../../../user_guide/diffusion/offloader/cpu_offload.md) | [Design](cpu_offload.md) |
+| Layerwise | [Guide](../../../user_guide/diffusion/offloader/layerwise_offload.md) | [Design](layerwise_offload.md) |
+| Distributed layerwise | [Guide](../../../user_guide/diffusion/offloader/distributed_layerwise_offload.md) | [Design](distributed_layerwise_offload.md) |
+
+## Strategy selection
+
+`OffloadConfig.from_od_config()` converts diffusion configuration into one
+`OffloadStrategy`. If callers set multiple flags, selection is deterministic:
+
+1. `DISTRIBUTED_LAYERWISE`;
+2. `LAYERWISE`;
+3. `MODEL_LEVEL`; and
+4. `NONE`.
+
+The factory logs when a higher-priority strategy disables another. It derives
+parallel and HSDP state from `DiffusionParallelConfig`; callers do not provide
+a separate offload group size.
+
+`get_offload_backend()` then validates platform offload support, resolves the
+device, and creates exactly one backend. Returning `None` means offloading is
+disabled or unsupported and must not leave partially installed hooks.
+
+## Shared lifecycle
+
+Every backend implements `OffloadBackend`:
+
+- `enable(pipeline)` discovers modules, establishes initial residency, and
+  installs hooks;
+- `disable()` removes owned hooks and resources; and
+- `is_enabled()` reports lifecycle state.
+
+`disable()` does not promise to restore the pipeline's original device
+placement. The caller owns any subsequent rematerialization.
+
+Hooks are registered through `HookRegistry` and `ModelHook`; offload backends
+must use distinct hook names and remove only hooks they own.
+
+## Discovery and topology
+
+Pipeline component discovery prefers `SupportsComponentDiscovery` declarations:
+
+- `_dit_modules`;
+- `_encoder_modules`;
+- `_vae_modules`; and
+- `_resident_modules`.
+
+Dotted paths are supported. Legacy pipelines may use the fallback scan of
+well-known attribute names, but new integrations should declare components
+explicitly.
+
+Ordinary layerwise offload obtains topology from declared
+`_layerwise_offload_blocks_attrs` (with a deprecated singular-name
+compatibility path). Distributed layerwise offload may instead use a pipeline
+`OffloadPlan`, then falls back to declared block attributes and its documented
+compatibility discovery. Discovery metadata describes structure only; the
+backend remains responsible for transfer, synchronization, and storage
+ownership.
+
+## Cross-strategy invariants
+
+- At most one strategy owns offload hooks for a pipeline.
+- A parameter has one authoritative host representation while offloaded.
+- Device storage is not freed until dependent compute or transfer work is
+  complete.
+- Non-persistent and model-specific buffers remain correct after movement or
+  checkpoint rematerialization.
+- Unsupported parallel or loading combinations fail before hooks mutate the
+  model.
+- Platform streams, events, synchronization, and cache management go through
+  the vLLM-Omni platform abstraction.
+
+The diffusion [Offloader module design](../../module/diffusion/offloader.md)
+describes how these feature contracts fit into the larger diffusion runtime.

docs/design/feature/offloader/cpu_offload.md

@@ -0,0 +1,56 @@
+# Model-Level Offload
+
+Model-level offload enforces mutual exclusion between pipeline component
+groups. For commands and model support, see the
+[model-level user guide](../../../user_guide/diffusion/offloader/cpu_offload.md).
+
+## Generic hook path
+
+`ModelLevelOffloadBackend.enable()` discovers DiTs, encoders, and VAEs. It
+makes encoders and VAEs device resident for initialization, then installs
+`SequentialOffloadHook` instances:
+
+- each DiT hook offloads all encoders and other DiTs before loading itself;
+- each encoder hook offloads every DiT before loading itself; and
+- VAEs are not part of the mutual-exclusion hook set.
+
+The hook moves parameters and buffers without recursively calling
+`module.to()`. This avoids recursion through modules that retain references to
+other transformer blocks. Host tensors are pinned when configured; DTensors
+are not pinned through the ordinary tensor path.
+
+Before forward returns control to the model, the hook synchronizes the active
+platform so compute cannot race the transfer. XPU uses blocking device-to-host
+copies because its allocator cannot safely evict storage while an asynchronous
+copy is pending.
+
+## Split-model delegation
+
+If a pipeline provides callable `enable_omni_model_cpu_offload`, the backend
+delegates instead of installing generic encoder/DiT hooks. This supports models
+whose mutually exclusive components are nested inside one transformer, such
+as a reasoner and generator.
+
+The delegated pipeline owns its internal contexts, but must preserve the same
+contract:
+
+- only the executing component is device resident;
+- enable and disable are idempotent;
+- transfers use the supplied device, pinning, and HSDP settings; and
+- disabling removes all model-local offload hooks or contexts.
+
+The delegation is currently duck typed. If another independent implementation
+is added, formalize the enable/disable pair as a protocol rather than expanding
+attribute-name heuristics.
+
+## Extension invariants
+
+- New ordinary pipelines declare both DiT and encoder groups through
+  `SupportsComponentDiscovery`.
+- Models with multiple DiTs offload the inactive DiTs as well as encoders.
+- Empty or parameterless modules do not trigger transfers.
+- Hook removal does not move modules; callers must not assume original
+  residency is restored.
+
+Shared selection and lifecycle behavior is defined in the
+[CPU Offloading design](README.md).

docs/design/feature/offloader/distributed_layerwise_offload.md

@@ -0,0 +1,308 @@
+# Distributed Layerwise Offload
+
+This document describes distributed layerwise offload (DLO) for diffusion
+models. DLO keeps only a small number of DiT blocks on the accelerator and
+streams the remaining blocks from host memory. The distributed backend can
+either shard those host-side weights across an existing parallel group or keep
+complete rank-local block sources and avoid an additional collective.
+
+For user-facing commands, see the
+[distributed layerwise offloading guide](../../../user_guide/diffusion/offloader/distributed_layerwise_offload.md)
+and the [Cosmos3 recipe](https://github.com/vllm-project/vllm-omni/blob/main/recipes/cosmos3/Cosmos3-DistOffload.md).
+
+## Feature compatibility
+
+Host-storage optimization and runtime compatibility are separate decisions.
+When direct checkpoint mmap is unavailable, DLO can still use tensors produced
+by the ordinary loader. "Compatibility path" below means that fallback is
+implemented but has less end-to-end coverage than the primary path.
+
+Legend: ✅ supported, ⚠️ compatibility path or limited validation, ❌ unsupported.
+
+| Feature | DLO + AllGather | DLO without AllGather |
+|---|---|---|
+| **DP** | ✅ Primary path; host weights are sharded across the DP group. | ✅ Each DP rank streams complete rank-local blocks. |
+| **SP** | ✅ When DP=1, DLO uses the SP group for weight sharding. | ✅ SP remains active without a DLO weight collective. |
+| **TP > 1** | ⚠️ Ordinary TP-aware loader only; no direct checkpoint mmap. | ⚠️ Ordinary TP-aware loader only; no direct checkpoint mmap. |
+| **HSDP** | ❌ Rejected to avoid double-sharding parameters. | ⚠️ Limited end-to-end coverage. |
+| **Per-tensor online FP8 linears** | ✅ Ordinary loader finalizes weights and scales before DLO sharding. | ✅ Ordinary loader retains complete rank-local tensors. |
+| **Other online quantization methods** | ❌ Rejected until runtime packing and scale layouts are validated. | ⚠️ Allowed through the ordinary loader; validation is method-specific. |
+| **Model-level or standard layerwise CPU offload** | ❌ Disabled because DLO takes priority. | ❌ Disabled because DLO takes priority. |
+| **Resident leading layers** | ❌ Rejected. | ✅ Requires eligible resident paths in the model's `OffloadPlan`. |
+
+See [Parallelism compatibility](#parallelism-compatibility) and
+[Request and loading constraints](#request-and-loading-constraints) for the
+detailed contracts and validation boundaries.
+
+## Status
+
+DLO is implemented for multi-device diffusion execution. The default
+AllGather path is the primary path for DP and SP deployments. The
+`--dlo-no-use-allgather` path streams complete blocks independently and adds no
+DLO weight collective.
+
+Host storage is selected separately from the transfer protocol. The loader can
+produce a direct-checkpoint mmap plan for a proven-compatible runtime layout;
+otherwise it uses the ordinary loader. Consequently, no-AllGather replicas on
+the same node can share immutable checkpoint pages when direct mmap is
+selected, while the ordinary-loader fallback still keeps a private runtime
+copy per process.
+
+The Phase A shared-mmap support boundary is TP1. TP greater than one is an
+ordinary-loader compatibility path: DLO can consume the resulting TP-local
+tensors, but those configurations do not use checkpoint mmap and must not be
+used to claim shared-mmap host-memory savings.
+
+The compatibility matrix below describes the current implementation. The
+unit-level guards are covered, but not every parallelism combination has a
+full model-and-hardware end-to-end test.
+
+## Design
+
+### DLO consumes the existing parallel topology
+
+DLO does not create a new DP, TP, or SP topology. It reads the configured
+`DiffusionParallelConfig` and attaches offload hooks to the DiT blocks after the
+standard distributed groups have been initialized.
+
+The DLO weight-sharding group is selected as follows:
+
+1. Use the existing DP group when `data_parallel_size > 1`.
+2. When DP is one and SP is greater than one, use the SP group.
+3. Otherwise, run rank-locally without a DLO process group.
+
+TP is deliberately not used as DLO's AllGather group. HSDP has its own
+parameter-sharding lifecycle and is not allowed to be sharded a second time by
+DLO's AllGather path.
+
+### The loader owns host-weight planning
+
+Before it decides whether ordinary weight materialization can be skipped, the
+diffusion loader builds one `HostWeightPlan`. A direct-checkpoint mmap plan is
+accepted only when preflight proves all of the following:
+
+- every required DiT parameter and persistent buffer has exactly one source;
+- runtime names, checkpoint keys, shapes, and dtypes match;
+- the runtime topology is TP1 without HSDP or online quantization; and
+- every custom loader operation is represented by a loader-owned checkpoint
+  adapter.
+
+The exact plan object is handed to DLO. The backend does not rescan checkpoint
+files, repeat the capability decision, or reconstruct names from its block
+topology. If preflight fails, the loader materializes weights normally and DLO
+consumes those runtime tensors.
+
+The plan owns only dedicated DiT component sources. If a pipeline also exposes
+ordinary sources for a text encoder or another non-DiT component, the loader
+still consumes those sources and includes their loaded names in its strict
+coverage check. Only the source prefixes covered by the plan skip ordinary
+materialization. A source that mixes DiT and non-DiT weights fails closed to
+the complete ordinary-loader path because it cannot be skipped safely as a
+unit.
+
+This boundary keeps checkpoint semantics out of DLO and avoids model-pipeline
+flags such as `_supports_mmap_loading` or parameter attributes for mmap-only
+transforms. Model-specific direct-layout knowledge, when required, lives in a
+checkpoint adapter beside the ordinary loader.
+
+### AllGather path
+
+With the default `dlo_use_allgather=True`, each rank stores approximately
+`1 / group_size` of each streamable block in pinned host memory. The next
+block's shard is copied to a device buffer and reconstructed with
+`all_gather_into_tensor` on a communication stream while the current block is
+executing.
+
+```text
+Compute:    [Block N]             [Block N+1]          [Block N+2]
+H2D:                      [shard N+1]           [shard N+2]
+AllGather:                [full N+1]             [full N+2]
+Buffers:    [current slot]       [prefetch slot]       [current slot]
+```
+
+![DLO double-buffer prefetch pipeline](../../figures/dlo/dlo_pipeline.gif)
+
+The backend uses two shared device buffers, so accelerator weight residency is
+bounded by the largest streamed blocks rather than the complete model.
+
+When direct checkpoint mmap is selected, the checkpoint mappings are only the
+source used to prepare each rank's persistent shard. They can be closed after
+shard preparation. Across the AllGather group, those private shards total
+approximately one runtime model copy.
+
+An effective DLO group size of one performs no collective, even when
+`dlo_use_allgather=True`; it follows the rank-local transfer path described
+below.
+
+When DP is greater than one, the engine can process one request per DP rank in
+the same denoising wave. Because AllGather is a collective, all participating
+requests must take the same execution path at every denoising step.
+
+### Rank-local path without DLO AllGather
+
+With `--dlo-no-use-allgather`, DLO forces its internal offload shard size to
+one and streams complete blocks using H2D copies only. The host backing may be
+either a loader-approved checkpoint mapping or ordinary runtime tensors.
+
+For direct mmap, each process retains immutable safetensors views and uses two
+bounded pinned host staging slots. Processes on the same node that map the same
+files share physical checkpoint pages through the OS page cache. This removes
+the persistent private full-model copy per pure-DP process, but each process
+still packs and transfers every complete block. Sharing is node-local; each
+node has its own page cache.
+
+When direct mmap preflight fails, the regular model loader remains responsible
+for preparing each rank's weights, including TP-local tensors or HSDP-managed
+parameters. In that fallback, each pure-DP process keeps a private full runtime
+copy.
+
+This mode means:
+
+- DP still provides independent replicas, but DLO does not shard weights
+  across DP ranks.
+- SP still performs its normal activation/attention collectives, but DLO does
+  not shard weights across SP ranks.
+- TP/HSDP/SP collectives, if configured, are not disabled by this flag; only
+  DLO's additional weight AllGather is disabled.
+- Pure DP deployments share one checkpoint-backed copy per node when direct
+  mmap is selected; the ordinary-loader fallback keeps one private runtime
+  copy per rank.
+- The scheduler does not require a synchronized DP request wave for DLO.
+
+## Parallelism compatibility
+
+| Parallelism | DLO + AllGather | DLO without AllGather |
+|---|---|---|
+| **DP** | Supported primary path. DLO shards host weights across the DP group and can run DP multi-concurrency. | Supported rank-local path. Compatible TP1 replicas can share checkpoint pages on each node; fallback runtime tensors remain private. |
+| **SP** | Supported in the implementation. With DP=1, DLO uses the SP group for host-weight sharding; SP still shards sequence/activation work. | SP remains active, but DLO keeps standard-loader rank-local weights and adds no SP weight collective. |
+| **TP > 1** | Outside the Phase A shared-mmap support scope. The loader falls back before mutation, preserves TP-local layouts, and DLO may apply DP/SP host sharding to those ordinary runtime tensors. | Outside the Phase A shared-mmap support scope. The ordinary TP-aware loader produces rank-local tensors, which DLO streams without an additional weight collective; DP replicas retain private runtime storage. |
+| **HSDP** | Rejected. HSDP has already sharded parameters, so DLO AllGather would double-shard them. | Accepted by configuration. HSDP owns parameter sharding and its own gathers; DLO only stages rank-local parameters. End-to-end coverage is limited. |
+
+### Combined dimensions
+
+- **DP + SP:** DLO uses the DP group for weight sharding when DP is greater
+  than one; SP continues to use its own sequence-parallel group. If DP is one,
+  the SP group becomes DLO's sharding group in AllGather mode.
+- **DP + TP/SP without AllGather:** standard model loading defines the
+  rank-local tensor layout. DLO adds no cross-DP, cross-TP, or cross-SP weight
+  collective.
+- **HSDP + SP:** the general parallel configuration permits HSDP over SP, but
+  DLO must use `--dlo-no-use-allgather`. HSDP remains responsible for weight
+  materialization and synchronization.
+- **HSDP + DP or TP:** rejected independently by the diffusion parallel
+  configuration.
+
+## Request and loading constraints
+
+AllGather DP multi-concurrency requires:
+
+- explicit `num_inference_steps`;
+- the same `num_inference_steps` for all requests in a wave; and
+- identical request arguments that affect the collective execution path.
+
+The no-AllGather path does not impose these DLO-specific synchronized-wave
+requirements.
+
+Direct checkpoint mmap can back either transfer path. It is currently limited
+to proven TP1, non-HSDP, non-online-quantized layouts. Other layouts use the
+ordinary loader. Per-tensor online FP8 linears can use DLO AllGather after the
+ordinary loader finalizes their runtime weights and scales; DLO then shards and
+reconstructs those tensors with their recorded layouts. Other online methods
+must use `--dlo-no-use-allgather` or disable online quantization until their
+runtime layouts are validated.
+
+A normalized runtime mmap cache, built through the ordinary loader, is the
+proposed general mechanism for sharing transformed TP or quantized layouts.
+That cache and its publication/lifecycle protocol are intentionally outside
+this phase; see [RFC #6195](https://github.com/vllm-project/vllm-omni/issues/6195).
+
+## Validation coverage
+
+Current source-level validation includes:
+
+- HSDP + DLO + AllGather rejection;
+- HSDP + DLO without AllGather acceptance at configuration level;
+- loader preflight fallback for TP, HSDP, online quantization, unknown custom
+  loaders, missing keys, and shape/dtype mismatches;
+- ordinary-loader fallback for per-tensor online FP8 linears followed by DLO
+  sharding of finalized weights and scales;
+- exact loader-to-backend plan transfer and ordinary-loader fallback;
+- rank-local mmap source retention, bounded two-slot staging, and adapter
+  transforms without parameter-side flags;
+- resident-layer requests requiring no-AllGather;
+- DP request-wave validation for denoising-step compatibility;
+- sharding, double-buffer, AllGather-size, and heterogeneous-block regression
+  tests.
+
+### B300 parallel-topology smoke matrix
+
+A four-GPU B300 smoke test covered MiniMax-H3 FL2VA with the same prompt, seed,
+CUDNN attention backend, 256x256 output, two denoising steps, and
+`dlo_resident_layers=0`. The TP2 rows used DiT DP2xTP2 with the text encoder and
+VAEs at TP1. They validate the ordinary-loader fallback only, not direct mmap
+or shared-mmap host-memory savings.
+
+| Configuration | Result | Warm E2E | Peak device memory | Host PSS |
+|---|---:|---:|---:|---:|
+| DP4xTP1 AllGather | Passed, 4 concurrent requests | 2.87 s / 4 requests | 13.84 GiB | 211.99 GiB |
+| DP4xTP1 no-AllGather | Passed, 1 request | 15.02 s | 13.23 GiB | 187.77 GiB |
+| DP2xTP2 AllGather | Passed, 2 concurrent requests | 4.16 s / 2 requests | 12.50 GiB | 211.97 GiB |
+| DP2xTP2 no-AllGather | Passed, 1 request | 3.51 s | 11.88 GiB | 314.01 GiB |
+
+Within each topology, the AllGather and no-AllGather video and audio outputs
+were byte-identical. All four runs completed without an `ERROR` or traceback
+and released their device allocations. For DP4xTP1, no-AllGather direct mmap
+reduced total PSS by 24.22 GiB (11.4%) and `Private_Dirty` from 211.33 to
+125.32 GiB (40.7%) relative to AllGather. For DP2xTP2, preflight selected the
+ordinary loader as designed; no-AllGather PSS was 314.01 GiB, about 48% above
+AllGather, because DP replicas did not share checkpoint-backed runtime
+weights. This is a functional and memory smoke test, not a production-quality
+performance or output-quality benchmark.
+
+### Host-memory measurement
+
+A two-worker MiniMax-H3 FL2VA measurement on one L20X node compared the
+ordinary-loader fallback with direct mmap. Both runs used
+DP=2, TP=1, no DLO AllGather, BF16 weights, two denoising steps, and a
+256x256 four-second request. The ordinary-loader workers were sampled after
+initialization. The mmap workers were sampled after one completed request, so
+the checkpoint working set had been faulted into the page cache; this is the
+more conservative point for mmap.
+
+The values below come from `/proc/<worker>/smaps_rollup` and include the whole
+worker, not only the DiT. The stable rank-to-rank difference comes from other
+pipeline components, so each worker should be compared with the same worker in
+the other storage mode.
+
+| Worker | Ordinary RSS | mmap RSS | Ordinary PSS | mmap PSS | PSS reduction |
+|---|---:|---:|---:|---:|---:|
+| DP worker 0 | 168.27 GiB | 132.76 GiB | 167.84 GiB | 101.43 GiB | 66.40 GiB |
+| DP worker 1 | 116.19 GiB | 79.97 GiB | 115.73 GiB | 48.64 GiB | 67.09 GiB |
+| **Two-worker total** | — | — | **283.56 GiB** | **150.08 GiB** | **133.48 GiB (47.1%)** |
+
+The direct-mmap workers each reported 62.45 GiB `Shared_Clean` but only
+31.20 GiB `Pss_File`, which is the proportional charge expected when the same
+resident checkpoint pages are mapped by two workers. `Private_Dirty` also fell
+from 167.53 to 70.24 GiB for worker 0 and from 115.40 to 17.44 GiB for worker
+1, a reduction of about 97–98 GiB per worker. RSS understates this benefit
+because it counts a shared physical page in every process that maps it; summed
+PSS is the appropriate node-memory comparison.
+
+The highest-value missing coverage is broader end-to-end numerical and
+lifecycle comparison against ordinary layerwise offload for DP+SP,
+HSDP+SP+no-AllGather, and TP greater than one across additional models and
+target CUDA/NCCL or CANN/HCCL hardware. That broader TP coverage does not
+change the Phase A direct-mmap TP1 support boundary.
+
+## Recommendations
+
+- Use **DP + DLO AllGather** for the supported throughput and host-memory
+  scaling path.
+- Use **SP + DLO AllGather** for long-sequence workloads when DP concurrency is
+  not the goal.
+- Use **no-AllGather** when independent replica execution is required. TP1
+  direct-mmap deployments can share checkpoint pages per node; other layouts
+  retain the ordinary loader's private host memory behavior and are outside
+  the Phase A shared-mmap support scope.
+- Prefer **HSDP alone** for production HSDP deployments until the combined
+  HSDP + DLO no-AllGather path has broader end-to-end coverage.

docs/design/feature/offloader/layerwise_offload.md

@@ -0,0 +1,57 @@
+# Layerwise Offload
+
+Layerwise offload streams transformer blocks through one accelerator. For
+commands and model support, see the
+[layerwise user guide](../../../user_guide/diffusion/offloader/layerwise_offload.md).
+
+## Hook ring
+
+`LayerWiseOffloadBackend` discovers every streamable block and registers one
+`LayerwiseOffloadHook` per block. Hooks form a ring: each hook owns the pinned
+host representation of its registered block and a reference to the next block
+it must prefetch; the final hook prefetches the first block for the next
+denoising iteration.
+
+During initialization, parameters and buffers are grouped by dtype and copied
+into contiguous pinned CPU tensors. Original tensor storage is replaced with
+an empty or meta placeholder while identity and module structure are retained.
+Metadata records names, offsets, shapes, and dtypes for rematerialization.
+
+## Forward lifecycle
+
+For each block:
+
+1. the previous hook's completion event ensures this block is materialized;
+2. the copy stream asynchronously allocates and fills the next block's device
+   tensors;
+3. parameter and buffer storage is rebound to views of those tensors;
+4. the current block computes on the normal stream; and
+5. its post-forward hook replaces device storage with placeholders after all
+   dependent work is safe.
+
+The copy stream waits on the compute stream before reusing storage, and the
+compute stream waits on the per-block prefetch event before forward. These
+dependencies, not incidental global synchronization, define correctness.
+
+## Topology contract
+
+A DiT declares ordered block-container attributes using
+`_layerwise_offload_blocks_attrs` or an `OffloadPlan`. Each resolved item must
+be an executable `nn.Module`; ordering must match forward execution. Multiple
+containers are concatenated in execution order.
+
+Non-block DiT modules, encoders, VAEs, and declared resident modules are not
+streamed by this backend.
+
+## Invariants and limitations
+
+- Each offloaded tensor has one pinned host backing store.
+- Parameter identity is preserved when storage is rebound.
+- Blocks with heterogeneous shapes and dtypes use their recorded metadata;
+  implementations must not assume identical blocks.
+- The ring is initialized before the first forward so block zero is available.
+- This backend is single-device; distributed weight reconstruction belongs to
+  [Distributed Layerwise Offload](distributed_layerwise_offload.md).
+
+Shared selection and lifecycle behavior is defined in the
+[CPU Offloading design](README.md).

docs/design/feature/quantization.md

@@ -76,6 +76,12 @@ equivalent. `auto-round` and `auto_round` use the Omni INC/AutoRound adapter.
 ModelOpt checkpoint metadata is detected from its method, producer metadata, or
 `quant_algo`, then mapped to the corresponding vLLM ModelOpt configuration.
 
+Unless the input is a per-component dictionary, the resolved configuration is
+passed unchanged to every quantization-aware component constructed by the
+pipeline. It affects only layers supported by that component's quantization
+method; it does not rewrite arbitrary `torch.nn` modules. A per-component
+dictionary narrows this global scope through runtime layer-prefix routing.
+
 The factory uses Omni-specific overrides for methods whose diffusion behavior
 or platform dispatch is not provided by the vLLM registry. The current override
 set includes INT8, BitsAndBytes, MXFP8, MXFP4, MXFP4 dual-scale, and INC/
@@ -154,9 +160,11 @@ quantizable vLLM layers. Each method configuration is responsible for:
 Model code must keep precision-sensitive or unsupported modules unquantized.
 The shared `safe_quant_config()` helper is used by model integrations for
 normalization and modulation layers where a generic FP8 configuration would be
-unsafe. Text encoders, audio/vision encoders, VAEs, schedulers, and tokenizers
-are not implicitly quantized just because the main transformer is quantized;
-they require explicit component routing and method support.
+unsafe. Encoders exposed through the quantization factory receive a global
+configuration just like the main transformer, but only their supported vLLM
+quantizable layers are affected. Ordinary `torch.nn` encoder layers, VAEs,
+schedulers, and tokenizers are not rewritten automatically. A component map can
+select a narrower scope, but it cannot add quantization support to a layer.
 
 Weight quantization and quantized KV cache are separate paths. The former is
 resolved through `quantization_config`; runtime attention/KV-cache quantization
@@ -182,12 +190,15 @@ configured parallelism strategy:
 - `WeightsMapper` transformations must be applied before matching ignored-layer
   lists or component prefixes.
 
-Distributed layerwise offload has an additional boundary: its sharded mmap
-AllGather path is not compatible with online quantization because the path
-expects already materialized checkpoint weights. Use the rank-local
-`--dlo-no-use-allgather` path or a compatible pre-quantized checkpoint when
-combining these features. See the [DLO design document](distributed_layerwise_offload.md)
-for the full compatibility matrix.
+Distributed layerwise offload has an additional boundary: direct checkpoint
+mmap cannot prepare online-quantized tensors because it bypasses the ordinary
+loader that creates their runtime weights and scales. Per-tensor online FP8
+linears can still use DLO AllGather through the ordinary loader: each rank first
+finalizes the FP8 tensors, then DLO shards those runtime tensors for
+reconstruction during prefetch. Other online methods remain limited to
+`--dlo-no-use-allgather` until their packing and scale layouts are validated.
+See the [DLO design document](offloader/distributed_layerwise_offload.md) for
+the full compatibility matrix.
 
 ## Adding a quantization backend
 

docs/design/feature/realtime_ar_diffusion.md

@@ -0,0 +1,213 @@
+# Realtime AR-Diffusion sessions
+
+> **Status:** experimental. This document describes the internal execution
+> contract implemented under `vllm_omni.experimental.ar_diffusion`. It does not
+> define a public HTTP or WebSocket schema.
+
+## Scope
+
+Realtime AR-Diffusion serves world models that generate one autoregressive
+latent block per request while retaining model state across requests. The
+generic runtime owns identity, event ordering, session lifecycle, routing, and
+KV allocation. A model adapter owns model-specific controls and conditioning.
+
+LingBot World v2 is the first single-KV-branch integration. DreamZero uses the
+same runner capability with two logical CFG branches.
+
+## Identities
+
+Four identifiers have distinct lifetimes:
+
+| Identifier | Lifetime | Owner | Contract |
+| --- | --- | --- | --- |
+| `session_id` | One persistent world | Session manager | Selects runner KV and model-owned state. |
+| `event_id` | One control/prompt update | Transport adapter | Unique and monotonically increasing within a session, including across reset. |
+| `chunk_index` | One committed AR block position | Session | Contiguous from zero; reset starts again at zero. |
+| `request_id` | One chunk execution | Session | Correlates the tick with its returned metadata. |
+
+`AsyncOmni` keeps its normal UUID-suffixed engine routing ID. The tick
+`request_id` remains inside the immutable AR-Diffusion snapshot and the
+standard output metadata, while caller-visible `OmniRequestOutput.request_id`
+continues to use AsyncOmni's existing external-ID contract. Protocol identity
+therefore does not require a private entrypoint override.
+
+## Request lifecycle
+
+1. `ARDiffusionSession.accept_event()` validates and queues an immutable
+   `ARDiffusionSessionEvent`. Prompt and opaque control tracks may be updated in
+   the same event.
+2. `ARDiffusionSession.next_chunk()` serializes chunk execution and calls
+   `_build_tick_locked()`. The resulting `ARDiffusionTickRequest` is a snapshot
+   of the current prompt, reduced controls, ordered event IDs, and chunk index.
+3. `ARDiffusionOmniTickConsumer.execute_tick()` copies the stage sampling
+   parameters, stores the typed tick in
+   `sampling_params.extra_args["ar_diffusion_tick"]`, constructs the normal Omni
+   prompt through its `prompt_provider`, and invokes `AsyncOmni.generate()`.
+4. `AsyncOmni` allocates its normal unique engine routing ID and submits one
+   `OmniDiffusionRequest` through the orchestrator.
+5. `ARDiffusionModelRunner._request_session()` parses the typed tick.
+   `_get_or_create_session()` obtains runner-owned paged KV, and
+   `bind_ar_diffusion_state()` exposes it to the pipeline only for the duration
+   of `execute_model()`.
+6. The model pipeline validates the tick and generates exactly one latent
+   block. LingBot interprets camera controls in
+   `LingBotCameraControlReducer` and
+   `LingBotWorldCausalDMDPipeline.forward()`; the generic runtime does not parse
+   camera or action schemas.
+7. The pipeline returns one `DiffusionOutput` with the generated payload and an
+   `ARDiffusionChunkMetadata` mapping under `metadata["ar_diffusion"]`.
+   `AsyncOmni` exposes it as
+   `OmniRequestOutput.multimodal_output["metadata"]["ar_diffusion"]`.
+8. The consumer parses the standard envelope. The session commits its reducer
+   state, prompt, controls, event IDs, and next chunk index only when returned
+   metadata exactly equals the submitted tick snapshot.
+
+The metadata envelope is:
+
+```json
+{
+  "metadata": {
+    "ar_diffusion": {
+      "session_id": "world-7",
+      "request_id": "ar-world-7-3-...",
+      "chunk_index": 3,
+      "applied_event_ids": [10, 11]
+    }
+  }
+}
+```
+
+## Transaction and failure semantics
+
+Accepted events are not acknowledged as applied until the model output and
+metadata are validated. Reducer `prepare()` is speculative; reducer `commit()`
+and the session snapshot form one logical commit. A model, metadata, or reducer
+failure leaves the chunk index and queued events uncommitted and transitions the
+session to `FAILED`.
+
+After a tick failure, worker state is closed because a forward may have
+partially modified or already released KV. The same chunk is not retried in
+place. A caller must explicitly reset the session, starting again at chunk
+zero, or close it.
+
+Explicit close follows this state machine:
+
+```text
+ACTIVE or FAILED
+       |
+       v
+    CLOSING ---- worker cleanup succeeds ----> CLOSED
+       |
+       +------- worker cleanup fails --------> CLEANUP_FAILED
+                                                   |
+                                                   +-- close retry --> CLOSING
+```
+
+`CLEANUP_FAILED` is a local tombstone. The session manager retains it, rejects
+creation of the same `session_id`, and rejects reset. Only an idempotent close
+retry that succeeds removes the session. Transport disconnect uses the same
+path.
+
+## KV ownership and capacity
+
+`ARDiffusionKVCacheSpec` is the pipeline-owned immutable geometry contract. It
+declares:
+
+- TP-local self-attention geometry, AR block size, sink, and recent window;
+- logical KV branches and their worker-local slots;
+- fixed-length cross-attention caches;
+- scratch tokens required by an uncommitted forward;
+- persistent model-owned CUDA bytes per resident session; and
+- the requested resident-session capacity.
+
+The runner computes:
+
+```text
+required(capacity) =
+    scratch
+  + managed_self_attention_pages(capacity)
+  + capacity * cross_attention_bytes_per_session
+  + capacity * model_owned_state_bytes_per_session
+```
+
+`gpu_memory_fraction` is a soft expansion budget:
+
+```text
+configured_budget = available_device_bytes * gpu_memory_fraction
+effective_budget = max(configured_budget, required(1))
+```
+
+If `required(1)` exceeds actual available device memory, initialization fails.
+Otherwise the runner always admits one viable session and uses the configured
+fraction to select additional resident sessions up to the model-declared cap.
+All model-owned reservations are deducted before the paged self-attention pool
+is allocated.
+
+LingBot reports its persistent full-horizon image-condition tensor. DreamZero
+reports the measured 603 MiB per-session Wan VAE causal-convolution state.
+
+## Concurrency and routing limits
+
+The current runner executes one request at a time:
+
+- `max_num_seqs` must be one;
+- request batching and diffusion step execution are rejected;
+- one AR block is generated per engine request; and
+- an AR-Diffusion stage must have exactly one replica.
+
+The single-replica restriction is intentional. Multi-replica support requires
+session-affine routing so every tick, reset, and close reaches the worker that
+owns the session. Random or round-robin routing is not correct.
+
+Multiple resident sessions may share one runner. Capacity selection accounts
+for all persistent pools, and the runner performs LRU eviction through the same
+model lifecycle hook used by explicit close.
+
+## Model boundary
+
+The generic runtime transports controls as opaque
+`ARDiffusionControlInput(track, schema, data)` values. Model adapters validate
+and reduce those values. This keeps target/velocity, key state, pose,
+intrinsics, and other model-specific conventions out of the session protocol.
+
+For LingBot World v2:
+
+- `lingbot_world/actions.py` parses key-state and trajectory controls;
+- `lingbot_world/camera.py` loads or constructs camera trajectories and Plücker embeddings;
+- `lingbot_world/transformer.py` implements checkpoint-compatible causal attention; and
+- `lingbot_world/pipeline.py` constructs conditioning, owns small non-KV session state, and
+  produces one block plus the standard metadata envelope.
+
+## Non-goals
+
+This contract does not currently provide:
+
+- a public HTTP or WebSocket request schema (the structured-interaction
+  frontend is tracked separately in
+  [#5527](https://github.com/vllm-project/vllm-omni/pull/5527));
+- camera/action semantics shared by every world model;
+- session migration or replication across workers;
+- retry of an ambiguous partially executed chunk;
+- cross-stage KV transfer; or
+- stateful streaming VAE decode.
+
+Those features can be layered on top without adding LingBot-specific identity
+or lifecycle fields to the generic runtime.
+
+## Required regression coverage
+
+CPU contract tests cover:
+
+- separation of tick correlation IDs from internal engine routing IDs;
+- event ordering, metadata equality, and reducer fail-closed commit;
+- close/disconnect failure tombstones and cleanup retry;
+- capacity one under the default fraction for shipped LingBot geometry;
+- capacity expansion under a tuned fraction;
+- DreamZero model-owned state accounting;
+- runner session reuse, reset, close, LRU eviction, and forward cleanup; and
+- LingBot model registration, imports, camera/action reduction, and one-block
+  output metadata.
+
+GPU validation must additionally exercise real model loading, action input,
+prompt switching, rolling-window replacement, reset, close, finite output
+latents, and exact metadata/request identity.

docs/design/feature/skip_softmax.md

@@ -2,8 +2,10 @@
 
 Skip-Softmax is the sparse-attention mode of the `TRTLLM_ATTN` backend. Usage — the config keys
 and how to pick an operating point — is in
-[Diffusion Attention Backends](../../user_guide/diffusion/attention_backends.md). This page explains
-the algorithm.
+[TRTLLM Attention](../../user_guide/diffusion/attention_backends/trtllm.md#skip-softmax).
+The shared selector contract is documented in
+[Diffusion Attention Backend Selection](attention_backend_selection.md). This
+page explains the algorithm.
 
 ## Motivation
 

docs/design/index.md

@@ -11,6 +11,10 @@ in the active navigation.
 
 ## Feature Design Documents
 
+For user-facing configuration and current compatibility, see the
+[Features overview](../features/README.md). A design document defines an
+implementation contract; it is not, by itself, a general support claim.
+
 ### Runtime and stage execution
 
 - [Disaggregated Inference](feature/disaggregated_inference.md)
@@ -18,6 +22,7 @@ in the active navigation.
 - [Async Diffusion Output](feature/async_diffusion_output.md)
 - [Async Omni Output Materialization](feature/omni_async_output_materialization.md)
 - [Automatic Prefix Caching in Omni Models](feature/prefix_caching.md)
+- [Realtime AR-Diffusion Sessions](feature/realtime_ar_diffusion.md)
 
 ### Communication
 
@@ -30,6 +35,10 @@ in the active navigation.
 - [Yuanrong Store Connector](feature/omni_connectors/yuanrong_connector.md)
 - [Yuanrong Transfer Engine Connector](feature/omni_connectors/yuanrong_transfer_engine_connector.md)
 
+### Quantization
+
+- [Quantization](feature/quantization.md)
+
 ### Diffusion acceleration
 
 #### Parallelism
@@ -42,18 +51,25 @@ in the active navigation.
 - [Tensor Parallel](feature/tensor_parallel.md)
 - [VAE Patch Parallelism](feature/vae_parallel.md)
 
-#### Attention Backends
+#### Attention optimization
 
+The [Diffusion Attention Backends](../user_guide/diffusion/attention_backends.md)
+guides list selectable backends, platform defaults, installation, and tuning.
+The design contracts separate selection mechanics from backend algorithms:
+
+- [Attention Backend Selection](feature/attention_backend_selection.md)
 - [Skip-Softmax](feature/skip_softmax.md)
 
-#### Quantization
+#### CPU offloading
 
-- [Quantization](feature/quantization.md)
+- [Overview and Shared Contracts](feature/offloader/README.md)
+- [Model-Level Offload](feature/offloader/cpu_offload.md)
+- [Layerwise Offload](feature/offloader/layerwise_offload.md)
+- [Distributed Layerwise Offload](feature/offloader/distributed_layerwise_offload.md)
 
 - [Cache-DiT](feature/cache_dit.md)
 - [TeaCache](feature/teacache.md)
 - [Diffusion Continuous Batching](feature/diffusion_continuous_batching.md)
-- [Distributed Layerwise Offload](feature/distributed_layerwise_offload.md)
 
 ## Infrastructure and Performance
 

docs/design/metrics.md

@@ -1,6 +1,6 @@
 # Prometheus Metrics Design
 
-This document describes how vLLM-Omni exposes Prometheus metrics for multi-stage pipelines, the constraints that shaped the design, and how the pipeline-level metrics coexist with upstream vLLM per-engine metrics. This iteration covers the text and audio surface; diffusion / image / video families land in a follow-up.
+This document describes how vLLM-Omni exposes Prometheus metrics for multi-stage pipelines, the constraints that shaped the design, and how the pipeline-level metrics coexist with upstream vLLM per-engine metrics.
 
 ## Objectives
 

docs/design/module/vllm_omni_config.md

@@ -18,8 +18,8 @@ validation_paths:
   - tests/config/**
 upstream_refs:
   - vllm.config
-last_reviewed: 2026-08-07
-last_verified_commit: 3d7fc3b9ba3cac88d579d4dc35b78b0b641675fc
+last_reviewed: 2026-08-11
+last_verified_commit: e356708da4cd39992405686dacfc32b18fcccc7f
 ---
 
 # vLLM-Omni configuration
@@ -39,11 +39,66 @@ The PR review proposal to capture environment-derived values in configuration
 objects at initialization is an open design question, not a current
 invariant.
 
+## Current structured-stage reuse behavior
+
+The current RFC #4021 implementation uses upstream vLLM config classes as the
+schema source for AR and generation stages:
+
+- `OmniStageLoadConfig` inherits `vllm.config.LoadConfig`.
+- `OmniStageCacheConfig` inherits `vllm.config.CacheConfig`.
+- `OmniStageSchedulerConfig` inherits `vllm.config.SchedulerConfig`.
+- `OmniStageParallelConfig` inherits `vllm.config.ParallelConfig`.
+
+This inheritance preserves the existing load, cache, scheduler, and parallel
+concern boundaries while reducing duplicate downstream declarations. It does
+not, by itself, make every inherited field an effective vLLM-Omni runtime
+option. Construction and runtime consumption are separate surfaces:
+
+| Surface | Current behavior |
+| --- | --- |
+| Structured schema | Upstream dataclass fields are inherited. Their defaults, default factories, and applicable Pydantic validation can participate when the structured object is constructed. |
+| Omni input ownership | Pipeline/deploy/CLI construction continues to accept only fields with an existing structured owner. Direct construction of an inherited sub-config can expose a wider upstream schema. |
+| Engine projection | For AR and generation stages, reusable fields are discovered by intersecting each upstream config dataclass with upstream `EngineArgs`. Only constructor-explicit values are emitted; inherited defaults remain deferred to terminal materialization. |
+| Terminal materialization | The engine-owning process constructs the final upstream `VllmConfig` and performs model-, platform-, rank-, port-, and backend-dependent initialization. |
+
+The generated projection maps include known upstream naming differences such
+as `cache_dtype` to `kv_cache_dtype`, `policy` to `scheduling_policy`, and
+`data_parallel_master_ip` to `data_parallel_address`. A field added to both an
+upstream concern config and `EngineArgs` therefore enters the AR/generation
+projection without requiring a second downstream allowlist. Explicit inherited
+fields that have no projection raise an error rather than being silently
+dropped.
+
+Ownership exclusions remain deliberate. Stage topology owns `scheduler_cls`,
+cache owns `disable_hybrid_kv_cache_manager`, and Omni runtime owns
+`distributed_executor_backend` and `worker_cls`; vLLM's private API-process
+fields are terminal internals. These exclusions prevent one input from being
+constructed or projected through two config concerns. Effective-engine-argument
+tests cover both the dynamic mapping and this exclusion boundary.
+
+`CompilationConfig` and `ProfilerConfig` are direct structured-stage fields:
+`stage.compilation_config` and `stage.profiler_config`. They accept mapping/YAML
+inputs and complete upstream value objects. Mapping inputs are validated and
+materialized as their concrete upstream types when the structured config is
+constructed, so downstream consumers see one resolved representation.
+Prebuilt upstream objects retain their type across the structured projection
+boundary.
+
+Quantization remains on its existing Omni-owned transport contract. This
+reuse step does not adopt upstream `QuantizationConfigArgs`, because the
+current engine-specific quantization split needs to be resolved separately.
+
+Diffusion keeps its existing engine-field ownership and projection surface.
+Shared Python inheritance must not be treated as evidence that an LLM-only
+parallel input is effective for a diffusion stage.
+
 ## Safe-change guide while deferred
 
 Changes should trace every affected producer to its runtime consumer and test
 the structured, legacy, CLI, and deployment paths that are actually supported.
-Do not infer a stable contract from this placeholder page.
+For inherited vLLM fields, tests should separately cover constructor behavior
+and effective `EngineArgs` projection. Do not infer a stable contract from
+this placeholder page.
 
 ## Promotion gate
 

docs/features/README.md

@@ -0,0 +1,73 @@
+# Features
+
+Use this section to choose and configure vLLM-Omni runtime and optimization
+features. The navigation follows the active
+[Feature Design](../design/index.md#feature-design-documents) taxonomy whenever
+an implementation contract has a user-facing workflow.
+
+!!! note "User guides, design documents, and recipes"
+
+    User guides explain how to enable a feature and where it is supported.
+    Design documents define internal contracts and do not, by themselves,
+    imply general or production support. For model-specific launch flags and
+    hardware requirements, use the linked examples and validated recipes from
+    [Supported Models](../models/supported_models.md).
+
+## Runtime and Stage Execution
+
+| Goal | User guide | Related design contract |
+| --- | --- | --- |
+| Choose serial, batched, step-wise, or streaming diffusion execution | [Execution Modes and Streaming](../user_guide/diffusion/execution_modes.md) | [Diffusion Continuous Batching](../design/feature/diffusion_continuous_batching.md), [Async Diffusion Output](../design/feature/async_diffusion_output.md) |
+| Reclaim stage memory without restarting the server | [Sleep Mode](sleep_mode.md) | Runtime lifecycle behavior is documented in the user guide |
+
+The [Prefill-Decode Disaggregation design reference](pd_disaggregation.md)
+describes the experimental Qwen3-Omni runtime topology and its current
+integration boundaries. It is not yet a supported launch recipe.
+
+Some runtime designs are deliberately not promoted as standalone User Guide
+features yet:
+
+- Generic, OmniConnector-based
+  [Disaggregated Inference](../design/feature/disaggregated_inference.md) is not
+  yet a standalone User Guide feature. Its
+  [OmniConnector implementations](../design/index.md#communication) describe
+  transport contracts, while practical connector configuration remains under
+  [Pipeline and deploy configurations](../configuration/stage_configs.md).
+- [Async Chunk](../design/feature/async_chunk.md) and
+  [Async Omni Output Materialization](../design/feature/omni_async_output_materialization.md)
+  are model- and pipeline-dependent. Use the selected model's deploy
+  configuration and recipe for the supported settings.
+- [Automatic Prefix Caching](../design/feature/prefix_caching.md) remains a
+  design-level contract until its user-facing configuration and compatibility
+  surface is consolidated.
+
+## Quantization
+
+Quantization is a cross-model feature rather than a diffusion-only
+optimization. The unified [`quantization_config` guide](../user_guide/quantization/overview.md)
+covers diffusion-only models, multi-stage omni/TTS models, and multi-stage
+diffusion models. Its [design contract](../design/feature/quantization.md)
+defines the shared configuration and backend extension points.
+
+## Diffusion Acceleration
+
+| Goal | User guide | Related design contract |
+| --- | --- | --- |
+| Compare acceleration methods and supported combinations | [Overview](../user_guide/diffusion_features.md), [Feature Compatibility](../user_guide/feature_compatibility.md) | [Diffusion acceleration designs](../design/index.md#diffusion-acceleration) |
+| Move weights between host and device memory | [CPU Offloading](../user_guide/diffusion/cpu_offload.md) | [CPU Offloading](../design/feature/offloader/README.md) |
+| Reuse denoising computation | [Cache Acceleration](../user_guide/diffusion/cache_acceleration/cache_dit.md) | [Cache-DiT](../design/feature/cache_dit.md), [TeaCache](../design/feature/teacache.md) |
+| Distribute diffusion work across devices | [Parallelism](../user_guide/diffusion/parallelism/overview.md) | [Parallelism designs](../design/index.md#parallelism) |
+| Select dense, sparse, or quantized attention paths | [Attention Backends](../user_guide/diffusion/attention_backends.md) | [Attention Backend Selection](../design/feature/attention_backend_selection.md) |
+| Compile repeated diffusion regions | [Regional Compilation](../user_guide/diffusion/regional_compilation.md) | User-facing optimization guide |
+| Add generated video frames | [Frame Interpolation](../user_guide/diffusion/frame_interpolation.md) | User-facing extension guide |
+| Reduce diffusion model startup time | [Startup and Loading](../user_guide/diffusion/startup_and_loading.md) | User-facing loading guide |
+| Apply diffusion adapters | [LoRA](../user_guide/diffusion/lora.md) | User-facing extension guide |
+
+The [Pipeline Parallelism guide](../user_guide/diffusion/parallelism/pipeline_parallel.md)
+remains available by direct link, but it is not promoted in the primary User
+Guide navigation until its user-facing support and placement are settled.
+
+## Experimental
+
+[Session State Manager](session_state_manager.md) is opt-in and experimental.
+Its APIs and compatibility may change without notice.

docs/features/custom_pipeline.md

@@ -101,7 +101,7 @@ outputs = omni.generate(
 )
 
 # Access custom trajectory data
-output = outputs[0].request_output
+output = outputs[0]
 print(f"Trajectory timesteps shape: {output.metrics['trajectory_timesteps'].shape}")
 print(f"Trajectory latents shape: {output.latents.shape}")
 ```

docs/features/pd_disaggregation.md

@@ -0,0 +1,140 @@
+# Prefill-Decode (PD) Disaggregation (experimental)
+
+!!! warning "Experimental design reference"
+    PD disaggregation is only partially integrated on the current `main` branch.
+    The runtime contains PD detection and routing scaffolding, but there is no
+    supported deploy overlay or end-to-end validated launch recipe yet. Treat this
+    page as a description of the current design, not as production guidance.
+
+Prefill-decode disaggregation splits the Qwen3-Omni Thinker into two logical
+stages: a prefill worker that processes the prompt and produces KV cache, and a
+decode worker that imports that cache and generates tokens. The Talker and
+Code2Wav stages remain downstream; PD does not disaggregate the entire
+Qwen3-Omni pipeline.
+
+## Current design
+
+### Runtime topology
+
+```mermaid
+flowchart LR
+    client["Client request"] --> orchestrator["Omni Orchestrator"]
+    orchestrator -->|"original prompt<br/>max_tokens = 1"| prefill["Thinker prefill<br/>KV producer"]
+    prefill -->|"KV blocks"| connector[("vLLM KV connector")]
+    prefill -->|"KV routing metadata"| orchestrator
+    orchestrator -->|"original prompt<br/>remote KV parameters"| decode["Thinker decode<br/>KV consumer"]
+    connector -->|"remote KV load"| decode
+    decode -->|"generated text"| client
+```
+
+The orchestrator owns the request lifecycle, while the vLLM KV connector owns
+the bulk KV transfer. The original prompt is submitted to both Thinker stages:
+the decode worker still needs the request tokens and multimodal features to
+construct its request, while the imported KV cache prevents it from repeating
+the completed prefill computation.
+
+For the full Qwen3-Omni pipeline, splitting the Thinker adds one logical stage:
+
+| Stage | Role | Output or handoff |
+| --- | --- | --- |
+| 0 | Thinker prefill | Saves prompt KV and prompt-side multimodal output |
+| 1 | Thinker decode | Loads remote KV, generates text, and produces decode-side conditioning |
+| 2 | Talker | Converts combined Thinker conditioning into codec codes |
+| 3 | Code2Wav | Converts codec codes into audio |
+
+Text remains a final output of the decode stage, and audio remains a final
+output of Code2Wav.
+
+### Request lifecycle
+
+1. At startup, PD detection looks for one stage marked `is_prefill_only` and
+   one downstream stage marked `is_decode_only`. The decode stage must list the
+   prefill stage as an input source.
+2. The orchestrator clones the Thinker sampling parameters for prefill, forces
+   `max_tokens=1`, clears stop conditions, and sets producer-side
+   `kv_transfer_params`. This makes prefill finish by length after exporting KV.
+3. When prefill finishes, the orchestrator captures its KV routing metadata and
+   Qwen multimodal output under the original request ID.
+4. The orchestrator submits the original prompt to decode with consumer-side
+   `kv_transfer_params`. Those parameters identify the transfer, prefill engine,
+   bootstrap endpoint, and, when required by the connector, remote request.
+5. The decode engine imports the KV cache and continues generation. The normal
+   Omni stage path then carries Thinker conditioning through Talker and
+   Code2Wav.
+
+### Control and data paths
+
+PD uses two different connector layers. They solve different handoffs and are
+not interchangeable.
+
+| Path | Payload | Owner |
+| --- | --- | --- |
+| PD control path | Request ID, original prompt, `transfer_id`, bootstrap address, engine ID, and connector-specific request metadata | Omni Orchestrator |
+| PD data path | Thinker attention KV blocks | vLLM KV connector |
+| Downstream model path | Thinker hidden states and embeddings, Talker codec codes, and audio tensors | vLLM-Omni stage connectors and processors |
+
+### Detection and validation
+
+The current runtime accepts one logical prefill/decode pair per pipeline. Each
+stage may use tensor parallelism, but both sides of the pair must satisfy these
+invariants:
+
+- both stages define `kv_transfer_config`
+- prefill uses `kv_producer` or `kv_both`
+- decode uses `kv_consumer` or `kv_both`
+- both stages use the same KV connector
+- `kv_buffer_device` and `kv_buffer_size` agree when set on both stages
+- both stages use the same `tensor_parallel_size`
+
+These checks happen before request processing so an inconsistent PD topology
+fails during engine initialization.
+
+## What is implemented today
+
+| Area | Current state |
+| --- | --- |
+| Pair detection and topology validation | Implemented in the shared PD helper |
+| Prefill sampling preparation | Implemented for sync and async entrypoints |
+| Prefill-to-decode routing | Implemented in the current Orchestrator |
+| Deploy-based configuration | Not yet migrated; the supported `stages:` deploy schema cannot currently expand Qwen3-Omni into the four-stage PD topology |
+| Qwen3 Thinker-to-Talker handoff | Bridge state and embedding-merge helpers exist, but the merge is not wired into the live stage processor |
+| Connector compatibility | Incomplete; the Orchestrator currently requires `remote_request_id`, while the pinned Mooncake connector completes prefill without returning that metadata |
+| Validation coverage | PD entrypoint tests and the PD deploy overlay in online end-to-end tests are temporarily disabled during migration |
+| Performance | No non-regression or throughput/latency result is established for the current path |
+
+Because these pieces are still converging, the previous legacy `stage_args` YAML
+and `vllm serve` command have been removed. The serve CLI now accepts the
+deploy-based `stages:` schema, so retaining the legacy recipe would imply a
+launch path that current `main` rejects.
+
+## Configuration contract under development
+
+The intended deploy integration starts from the
+[default Qwen3-Omni deploy config](gh-file:vllm_omni/deploy/qwen3_omni_moe.yaml)
+and expands its frozen three-stage pipeline into the four stages shown above.
+A complete integration must:
+
+- split the Thinker definition without duplicating the Talker or Code2Wav model
+  contracts
+- remap downstream stage IDs, input sources, and stage connectors
+- inject producer and consumer `kv_transfer_config` values into the two Thinker
+  stages
+- allocate separate device sets for prefill and decode
+- preserve both final outputs: decode text and Code2Wav audio
+- provide connector-specific routing metadata without hard-coding one connector
+  contract
+
+The required number of GPUs depends on tensor-parallel sizes and whether Talker
+and Code2Wav share a device. A basic TP=1 layout uses one device for prefill, one
+for decode, and one for the two downstream audio stages.
+
+## Implementation map
+
+- PD detection, validation, and sampling-parameter preparation:
+  [`PDDisaggregationMixin`](gh-file:vllm_omni/entrypoints/pd_utils.py)
+- startup detection and connector metadata:
+  [`AsyncOmniEngine`](gh-file:vllm_omni/engine/async_omni_engine.py)
+- request-scoped prefill-to-decode routing:
+  [`Orchestrator`](gh-file:vllm_omni/engine/orchestrator.py)
+- Qwen3 downstream conditioning helpers:
+  [Qwen3-Omni stage input processor](gh-file:vllm_omni/model_executor/stage_input_processors/qwen3_omni.py)

docs/getting_started/quickstart.md

@@ -47,7 +47,7 @@ if __name__ == "__main__":
     omni = Omni(model="Tongyi-MAI/Z-Image-Turbo")
     prompt = "a cup of coffee on the table"
     outputs = omni.generate(prompt)
-    images = outputs[0].request_output.images
+    images = outputs[0].images
     images[0].save("coffee.png")
 ```
 
@@ -75,8 +75,7 @@ if __name__ == "__main__":
     ]
     omni_outputs = omni.generate(prompts)
     for i_prompt, prompt_output in enumerate(omni_outputs):
-        this_request_output = prompt_output.request_output
-        this_images = this_request_output.images
+        this_images = prompt_output.images
         for i_image, image in enumerate(this_images):
             image.save(f"p{i_prompt}-img{i_image}.jpg")
             print("saved to", f"p{i_prompt}-img{i_image}.jpg")

docs/models/supported_models.md

@@ -37,26 +37,37 @@ th {
 | `Krea2Pipeline` | Krea 2 (Raw + Turbo) | `krea/Krea-2-Raw`, `krea/Krea-2-Turbo` | ✅︎ | | | | — |
 | `WanPipeline` | Wan2.1-T2V, Wan2.2-T2V, Wan2.2-TI2V | `Wan-AI/Wan2.1-T2V-1.3B-Diffusers`, `Wan-AI/Wan2.1-T2V-14B-Diffusers`, `Wan-AI/Wan2.2-T2V-A14B-Diffusers`, `Wan-AI/Wan2.2-TI2V-5B-Diffusers` | ✅︎ | ✅︎ | ✅︎ | ✅︎ | — |
 | `WanImageToVideoPipeline` | Wan2.2-I2V | `Wan-AI/Wan2.2-I2V-A14B-Diffusers` | ✅︎ | ✅︎ | ✅︎ | ✅︎ | — |
+| `LingBotWorldCausalDMDPipeline` | LingBot-World 2.0 v1 (experimental) | `robbyant/lingbot-world-v2-14b-causal-fast-diffusers` | Experimental | | | | — |
 | `Cosmos3OmniDiffusersPipeline` | Cosmos3 T2I, T2V, I2V, V2V, T2V with sound, action policy | `nvidia/Cosmos3-Nano`, `nvidia/Cosmos3-Super` | ✅︎ | ✅︎ | ✅︎ | ✅︎ | — |
+| `SanaWmPipeline` | SANA-WM | `BBBBruce/SANA-WM_bidirectional-stage1-diffusers` | ✅︎ | | | | [Repository](https://github.com/vllm-project/vllm-omni/blob/main/recipes/NVIDIA/SANA-WM.md) |
 | `Wan22S2VPipeline` | Wan2.2-S2V | `Wan-AI/Wan2.2-S2V-14B` | ✅︎ | ✅︎ | ✅︎ | ✅︎ | — |
 | `Wan22VACEPipeline` | Wan2.1-VACE | `Wan-AI/Wan2.1-VACE-1.3B-diffusers`, `Wan-AI/Wan2.1-VACE-14B-diffusers` | ✅︎ | ✅︎ | ✅︎ | ✅︎ | — |
 | `Wan22VACEPipeline` | Wan2.2-VACE | `Pyros13/Wan2.2-VACE-Fun-A14B-Diffusers` | ✅︎ |   |   |   | — |
 | `LTX2Pipeline` | LTX-2 / LTX-2.3 one-stage T2V and I2V | `Lightricks/LTX-2`, `diffusers/LTX-2.3-Diffusers` | ✅︎ | ✅︎ | | | — |
-| `LTX2DistilledPipeline` | LTX-2 distilled two-stage T2V and I2V | `rootonchair/LTX-2-19b-distilled` | ✅︎ | ✅︎ | | | — |
+| `LTX2TwoStagePipeline` | LTX-2 / LTX-2.3 ordinary two-stage T2V and I2V | `Lightricks/LTX-2`, `diffusers/LTX-2.3-Diffusers` + matching Lightricks LoRA and upsampler | ✅︎ | ✅︎ | | | — |
+| `LTX2DistilledOneStagePipeline` | LTX-2 / LTX-2.3 merged-distilled one-stage T2V and I2V | `rootonchair/LTX-2-19b-distilled`, `diffusers/LTX-2.3-Distilled-Diffusers` | ✅︎ | ✅︎ | | | — |
+| `LTX2DistilledTwoStagePipeline` | LTX-2 / LTX-2.3 merged-distilled two-stage T2V and I2V | `rootonchair/LTX-2-19b-distilled`, `diffusers/LTX-2.3-Distilled-Diffusers` + matching Lightricks upsampler | ✅︎ | ✅︎ | | | — |
+| `LTX2Pipeline` | LTX-2.5 Full/SFT one-stage T2V and I2V | `Lightricks/LTX-2.5-Diffusers` | ✅︎ | | | | [Repository](https://github.com/vllm-project/vllm-omni/blob/main/recipes/LTX/LTX-2.5.md) |
+| `LTX2TwoStagePipeline` | LTX-2.5 Full/SFT two-stage T2V and I2V | `Lightricks/LTX-2.5-Diffusers` + `Lightricks/LTX-2.5` LoRA | ✅︎ | | | | [Repository](https://github.com/vllm-project/vllm-omni/blob/main/recipes/LTX/LTX-2.5.md) |
+| `LTX2DistilledOneStagePipeline` | LTX-2.5 merged-distilled one-stage T2V and I2V | `Lightricks/LTX-2.5-Diffusers` | ✅︎ | | | | [Repository](https://github.com/vllm-project/vllm-omni/blob/main/recipes/LTX/LTX-2.5.md) |
+| `LTX2DistilledTwoStagePipeline` | LTX-2.5 merged-distilled two-stage T2V and I2V | `Lightricks/LTX-2.5-Diffusers` | ✅︎ | | | | [Repository](https://github.com/vllm-project/vllm-omni/blob/main/recipes/LTX/LTX-2.5.md) |
 | `LingBotVideoPipeline` | LingBot-Video dense and MoE T2I, T2V, TI2V | `robbyant/lingbot-video-dense-1.3b`, `robbyant/lingbot-video-moe-30b-a3b` | ✅︎ | | | | — |
 | `MiniMaxH3Pipeline` | MiniMax H3 T2VA, first/last-keyframe FL2VA, and mixed-reference Ref2VA | `MiniMaxAI/MiniMax-H3` | ✅︎ | ✅︎ | | | [Published](https://recipes.vllm.ai/MiniMaxAI/MiniMax-H3) |
 | `DreamZeroPipeline` | DreamZero-DROID | `GEAR-Dreams/DreamZero-DROID` | ✅︎ | ✅︎ | ✅︎ | ✅︎ | — |
+| `Pi0Pipeline` | π0 (Pi-Zero) VLA | `lerobot/pi0_base` | ✅︎ | | | | [Repository](https://github.com/vllm-project/vllm-omni/blob/main/recipes/lerobot/Pi0.md) |
 | `HeliosPipeline`, `HeliosPyramidPipeline` | Helios | `BestWishYsh/Helios-Base`, `BestWishYsh/Helios-Mid`, `BestWishYsh/Helios-Distilled` | ✅︎ | ✅︎ | ✅︎ | | — |
 | `MagiHumanPipeline` | MagiHuman | `SII-GAIR/daVinci-MagiHuman-Base-1080p` | ✅︎ | ✅︎ | | | — |
 | `OvisImagePipeline` | Ovis-Image | `OvisAI/Ovis-Image` | ✅︎ | ✅︎ | | ✅︎ | — |
 | `LongcatImagePipeline` | LongCat-Image | `meituan-longcat/LongCat-Image` | ✅︎ | ✅︎ | ✅︎ | ✅︎ | — |
 | `LongCatImageEditPipeline` | LongCat-Image-Edit | `meituan-longcat/LongCat-Image-Edit` | ✅︎ | ✅︎ | ✅︎ | ✅︎ | — |
+| `LongCatVideoAvatarPipeline` | LongCat-Video-Avatar-1.5 A2V/AI2V (native single-speaker and multi-speaker AI2V/AVC) | `meituan-longcat/LongCat-Video-Avatar-1.5` | ✅︎ | | | | [Repository](https://github.com/vllm-project/vllm-omni/blob/main/recipes/meituan-longcat/LongCat-Video-Avatar-1.5.md) |
 | `BooguImagePipeline` | Boogu-Image | `Boogu/Boogu-Image-0.1-Base`, `Boogu/Boogu-Image-0.1-Edit` | ✅︎ | | | | — |
 | `StableDiffusionXLPipeline` | Stable-Diffusion-XL | `stabilityai/stable-diffusion-xl-base-1.0` | ✅︎ | ✅︎ | ✅︎ | ✅︎ | — |
 | `StableDiffusion3Pipeline` | Stable-Diffusion-3 | `stabilityai/stable-diffusion-3.5-medium` | ✅︎ | ✅︎ | | ✅︎ | — |
 | `CosyVoice3Model` | CosyVoice3 | `FunAudioLLM/Fun-CosyVoice3-0.5B-2512` | ✅︎ | ✅︎ | | ✅︎ | — |
 | `OmniVoiceModel` | OmniVoice | `k2-fsa/OmniVoice` | ✅︎ | | | | — |
 | `VoxCPM2TalkerForConditionalGeneration` | VoxCPM2 | `openbmb/VoxCPM2` | ✅︎ | | | | — |
+| `DotsTTSForConditionalGeneration` | dots.tts | `rednote-hilab/dots.tts-soar` | ✅︎ | | | | — |
 | `MammothModa2ForConditionalGeneration` | MammothModa2-Preview | `bytedance-research/MammothModa2-Preview` | ✅︎ | ✅︎ | | | — |
 | `MammothModa2ForConditionalGeneration` | MammothModa2-Dev (AR-only image understanding) | `bytedance-research/MammothModa2-Dev` | ✅︎ | | | | — |
 | `Flux2KleinPipeline` | FLUX.2-klein | `black-forest-labs/FLUX.2-klein-4B`, `black-forest-labs/FLUX.2-klein-9B` | ✅︎ | ✅︎ | ✅︎ | ✅︎ | — |
@@ -68,6 +79,7 @@ th {
 | `SoulXSingerPipeline` | SoulX-Singer (SVS) | `Soul-AILab/SoulX-Singer` | ✅︎ | | | | — |
 | `SoulXSingerSVCPipeline` | SoulX-Singer-SVC | `Soul-AILab/SoulX-Singer` (`model-svc.pt`) | ✅︎ | | | | — |
 | `AudioXPipeline` | AudioX | `zhangj1an/AudioX` | ✅︎ | ✅︎ | | | — |
+| `MiniMaxMusic3ForConditionalGeneration` | MiniMax Music 3 (text-to-music) | `MiniMaxAI/MiniMax-Music3` | ✅︎ | | | | — |
 | `Qwen3TTSForConditionalGeneration` | Qwen3-TTS-12Hz-1.7B-CustomVoice | `Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice` | ✅︎ | | | | [Published](https://recipes.vllm.ai/Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice) |
 | `Qwen3TTSForConditionalGeneration` | Qwen3-TTS-12Hz-1.7B-VoiceDesign | `Qwen/Qwen3-TTS-12Hz-1.7B-VoiceDesign` | ✅︎ | ✅︎ | ✅︎ | ✅︎ | — |
 | `Qwen3TTSForConditionalGeneration` | Qwen3-TTS-12Hz-1.7B-Base | `Qwen/Qwen3-TTS-12Hz-1.7B-Base` | ✅︎ | ✅︎ | ✅︎ | ✅︎ | — |
@@ -79,6 +91,7 @@ th {
 | `HiggsAudioV2ForConditionalGeneration` | Higgs-Audio v2 | `bosonai/higgs-audio-v2-generation-3B-base` | ✅︎ | | | | — |
 | `HiggsMultimodalQwen3ForConditionalGeneration` | Higgs-Audio v3 (TTS) | `bosonai/higgs-audio-v3-tts-4b` | ✅︎ | | | | [Published](https://recipes.vllm.ai/bosonai/higgs-audio-v3-tts-4b) |
 | `IndexTTS2TalkerForConditionalGeneration` | IndexTTS-2 | `IndexTeam/IndexTTS-2` | ✅︎ | | | | — |
+| `IndexTTS25TalkerForConditionalGeneration` | IndexTTS-2.5 | native `checkpoints/` bundle | ✅︎ | | | | [Repository](https://github.com/vllm-project/vllm-omni/blob/main/recipes/IndexTeam/IndexTTS-2_5.md) |
 | `NextStep11Pipeline` | NextStep-1.1 | `stepfun-ai/NextStep-1.1` | ✅︎ | ✅︎ | | ✅︎ | — |
 | `MiMoAudioModel` | MiMo-Audio-7B-Instruct | `XiaomiMiMo/MiMo-Audio-7B-Instruct` | ✅︎ | ✅︎ | | | — |
 | `MiMoV2ASRForCausalLM` | MiMo-V2.5-ASR | `XiaomiMiMo/MiMo-V2.5-ASR` | ✅︎ | ✅︎ | | | — |
@@ -94,6 +107,7 @@ th {
 |`DyninOmniForConditionalGeneration` | Dynin-Omni | `snu-aidas/Dynin-Omni` | ✅︎ | | | | — |
 | `MiniCPMO45OmniForConditionalGeneration` | MiniCPM-o 4.5 | `openbmb/MiniCPM-o-4_5` | ✅︎ | | ✅︎ | | — |
 | `ErnieImagePipeline` | ERNIE-Image | `baidu/ERNIE-Image`, `baidu/ERNIE-Image-Turbo` | ✅︎ | ✅︎ | ✅︎ | ✅︎ | — |
+| `GepardTalkerForConditionalGeneration` | Gepard-1.0 | `nineninesix/gepard-1.0` | ✅︎ | | | | — |
 |`HiDreamImagePipeline` | HiDream-I1-Full | `HiDream-ai/HiDream-I1-Full` | ✅︎ | ✅︎ | | | — |
 
 The `Recipe` column links the corresponding deployment recipe. `Published` is
@@ -107,6 +121,10 @@ support metadata until their recipe is audited.
 
 ✅︎ indicates the model is supported on that backend. Empty cells mean not listed as supported on that backend.
 
+See the [LingBot-World 2.0 recipe](https://github.com/vllm-project/vllm-omni/blob/main/recipes/Robbyant/LingBot-World-2.0.md)
+for offline and experimental realtime usage, supported controls, validation,
+and current limitations.
+
 <sup>H3</sup> MiniMax H3 AMD GPU support is validated on gfx942 (MI300X) and
 gfx950 (MI350) in BF16, using the AITER `FLASH_ATTN` diffusion attention backend.
 Other AMD SKUs (e.g. MI325X) are not listed until their

docs/serving/speech_api.md

@@ -141,7 +141,20 @@ The response shape depends on the streaming parameters:
 
 **Non-streaming (default).** With `stream=false` and no `stream_format`, returns the
 complete clip as binary audio data with an appropriate `Content-Type` header (e.g.
-`audio/wav`). The raw-bytes body has no JSON carrier, so no `usage` is reported.
+`audio/wav`). Because the raw-bytes body has no JSON carrier, successful
+non-streaming responses from non-diffusion speech servers report usage through
+response headers:
+
+| Header | Description |
+| --- | --- |
+| `x-vllm-omni-input-tokens` | Total input tokens (`text_tokens` + `audio_tokens`). |
+| `x-vllm-omni-output-tokens` | Generated codec/audio tokens. |
+| `x-vllm-omni-total-tokens` | `input_tokens` + `output_tokens`. |
+| `x-vllm-omni-input-text-tokens` | Tokens from the synthesized text (`input` plus `instructions`). |
+| `x-vllm-omni-input-audio-tokens` | Reference-audio codec frames, non-zero only for in-context voice cloning. |
+
+Diffusion-mode speech servers route through a separate response path and do not
+emit these headers.
 
 **Raw audio stream** (`stream_format="audio"`). Streams raw audio bytes (PCM or
 WAV) as they are decoded.

docs/serving/videos_api.md

@@ -167,6 +167,14 @@ transfer options such as `control_path`, `control_guidance`,
 `num_conditional_frames`, `show_control_condition`, and `show_input`; see the
 Cosmos3 recipe for complete examples.
 
+HTTP redirects for `image_reference.image_url` follow vLLM's
+`VLLM_MEDIA_URL_ALLOW_REDIRECTS` setting. Before starting the server, set it to
+`1` (the default) to allow redirects or `0` to reject them. A redirect target
+can differ from the original host, and vLLM-Omni does not yet fully implement
+upstream vLLM's media URL allowlist protection. Deployments should therefore
+treat remote media URLs as untrusted and choose this setting as part of their
+URL access policy.
+
 ### Speech-to-Video
 
 For models that support audio-driven generation (e.g., Wan2.2-S2V), pass both

docs/usage/faq.md

@@ -10,12 +10,12 @@ A: Refer to [GPU memory calculation and configuration](../configuration/gpu_memo
 
 > Q: I encounter some bugs or CI problems, which is urgent. How can I solve it?
 
-A: At first, you can check current [issues](https://github.com/vllm-project/vllm-omni/issues) to find possible solutions. If none of these satisfy your demand and it is urgent, please find these [volunteers](https://docs.vllm.ai/projects/vllm-omni/en/latest/community/volunteers/) for help.
+A: Search the current [issues](https://github.com/vllm-project/vllm-omni/issues) first. If no existing issue matches a bug or CI failure, open a [bug report](https://github.com/vllm-project/vllm-omni/issues/new?template=400-bug-report.yml) with a minimal reproducer, environment details, and relevant logs. For technical or usage questions, use the `sig-omni` channel in [vLLM Slack](https://slack.vllm.ai/) or the [vLLM Forum](https://discuss.vllm.ai/).
 
 > Q: Does vLLM-Omni support AWQ or any other quantization?
 
-A: We plan to introduce GGUF FP8 prequantized models and online FP8 quantization in version 0.16.0. Support for other quantization types will follow in future releases. For details, please see our [Q1 quantization roadmap](https://github.com/vllm-project/vllm-omni/issues/1057).
+A: AWQ is available through the inherited vLLM quantization registry, and vLLM-Omni supports several additional methods. Actual support and validation vary by model, component, hardware, and checkpoint. See the [quantization overview](../user_guide/quantization/overview.md) for validated combinations and method-specific guides.
 
 > Q: Does vLLM-Omni support multimodal streaming input and output?
 
-A: Not yet. We already put it on the [Roadmap](https://github.com/vllm-project/vllm-omni/issues/165). Please stay tuned!
+A: Yes, for specific models and serving paths. Current capabilities include [streaming video input](../serving/video_stream_api.md), [streaming diffusion output](../user_guide/diffusion/execution_modes.md#streaming-output), and experimental full-duplex real-time audio described in the [project overview](../README.md). Check the model and API documentation because input and output support is model-specific.

docs/usage/metrics.md

@@ -1,23 +1,23 @@
 # Production Metrics
 
-vLLM-Omni exposes Prometheus metrics via the `/metrics` endpoint on the OpenAI-compatible API server. This page covers the text and audio surface; diffusion / image / video metrics are tracked in a follow-up PR.
+vLLM-Omni exposes Prometheus metrics via the `/metrics` endpoint on the OpenAI-compatible API server.
 
 ```bash
 vllm-omni serve Qwen/Qwen3-Omni-30B-A3B-Instruct --port 8000 --log-stats
 curl http://localhost:8000/metrics
 ```
 
-**`--log-stats` is required to populate metric data.** Without the flag, the endpoint still returns `200 OK` but the upstream `vllm:*` wrap is not registered at all, and the 15 `vllm:omni_*` families are registered as placeholders with no data written to them. This keeps the runtime cost essentially zero for deployments that don't need monitoring. With the flag, all ~80 families populate.
+**`--log-stats` is required to populate metric data.** Without the flag, the endpoint still returns `200 OK` but the upstream `vllm:*` wrap is not registered at all, and the 15 `vllm_omni:*` families are registered as placeholders with no data written to them. This keeps the runtime cost essentially zero for deployments that don't need monitoring. With the flag, all ~80 families populate.
 
 ## Metric Namespaces
 
 | Prefix | Source | Present when |
 |--------|--------|--------------|
-| `vllm:omni_` | vLLM-Omni orchestrator / audio modality / cross-stage transfer | Pipeline-dependent |
+| `vllm_omni:` | vLLM-Omni orchestrator / audio modality / cross-stage transfer | Pipeline-dependent |
 | `vllm:` | Upstream vLLM engine, wrapped by `OmniPrometheusStatLogger` to expose `{stage, replica}` | Pipeline includes an LLM (AR) stage |
 | `http_` / `process_` | Uvicorn / Python runtime | Always |
 
-## Pipeline-Level Metrics (`vllm:omni_`)
+## Pipeline-Level Metrics (`vllm_omni:`)
 
 Defined in `vllm_omni/metrics/prometheus.py`. Track request lifecycle across the full multi-stage pipeline.
 
@@ -25,42 +25,66 @@ Defined in `vllm_omni/metrics/prometheus.py`. Track request lifecycle across the
 
 | Metric | Type | Labels | Description |
 |--------|------|--------|-------------|
-| `vllm:omni_num_requests_running` | Gauge | `model_name` | Pipeline-global in-flight requests (dispatched to engine, not yet finalized) |
-| `vllm:omni_num_requests_waiting` | Gauge | `model_name` | Requests waiting in the Orchestrator queue |
-| `vllm:omni_requests_success_total` | Counter | `model_name`, `finished_reason` | Total requests by completion reason. `finished_reason` ∈ {`stop`, `length`, `abort`, ...} mirroring upstream `vllm:request_success_total`; aborts cover client disconnect / cancellation paths in addition to upstream `FinishReason.ABORT` |
+| `vllm_omni:num_requests_running` | Gauge | `model_name` | Pipeline-global in-flight requests (dispatched to engine, not yet finalized) |
+| `vllm_omni:num_requests_waiting` | Gauge | `model_name` | Requests waiting in the Orchestrator queue |
+| `vllm_omni:requests_success_total` | Counter | `model_name`, `finished_reason` | Total requests by completion reason. `finished_reason` ∈ {`stop`, `length`, `abort`, ...} mirroring upstream `vllm:request_success_total`; aborts cover client disconnect / cancellation paths in addition to upstream `FinishReason.ABORT` |
 
 ### Latency
 
 | Metric | Type | Labels | Description |
 |--------|------|--------|-------------|
-| `vllm:omni_e2e_request_latency_s` | Histogram | `model_name` | Pipeline-global end-to-end request latency in seconds |
+| `vllm_omni:e2e_request_latency_s` | Histogram | `model_name` | Pipeline-global end-to-end request latency in seconds |
 
-## Audio Modality Metrics (`vllm:omni_`)
+## Audio Modality Metrics (`vllm_omni:`)
 
 Emitted at request finalize, except for `audio_ttfp_s` (streaming-hook at the first audio packet) and `audio_underrun_s` / `audio_continuity_ok_total` (streaming finalize, after the chunk stream is exhausted). All carry `{model_name, stage, replica}` plus the listed extra label.
 
 | Metric | Type | Extra label | Description |
 |--------|------|-------------|-------------|
-| `vllm:omni_audio_ttfp_s` | Histogram | — | Time from request arrival to first audio packet/frame |
-| `vllm:omni_audio_duration_s` | Histogram | — | Audio content duration (`audio_frames / sample_rate`) |
-| `vllm:omni_audio_rtf` | Histogram | — | Real-time factor (`stage_gen_time_s / audio_duration_s`); streaming TTS SLO red line `< 1`; uses `RTF_BUCKETS` |
-| `vllm:omni_audio_frames_total` | Counter | — | Cumulative audio frame count; throughput via `rate()` |
-| `vllm:omni_audio_underrun_s` | Histogram | — | Per-request worst-case player deficit; `> 0` indicates listener heard silent gaps |
-| `vllm:omni_audio_continuity_ok_total` | Counter | `threshold_ms` | Incremented when the request's worst underrun stayed below `threshold_ms` |
-| `vllm:omni_audio_skipped_requests_total` | Counter | `reason` | Silent-loss counter — code2wav rejected malformed codec input and returned `200 OK` with empty audio |
+| `vllm_omni:audio_ttfp_s` | Histogram | — | Time from request arrival to first audio packet/frame |
+| `vllm_omni:audio_duration_s` | Histogram | — | Audio content duration (`audio_frames / sample_rate`) |
+| `vllm_omni:audio_rtf` | Histogram | — | Real-time factor (`stage_gen_time_s / audio_duration_s`); streaming TTS SLO red line `< 1`; uses `RTF_BUCKETS` |
+| `vllm_omni:audio_frames_total` | Counter | — | Cumulative audio frame count; throughput via `rate()` |
+| `vllm_omni:audio_underrun_s` | Histogram | — | Per-request worst-case player deficit; `> 0` indicates listener heard silent gaps |
+| `vllm_omni:audio_continuity_ok_total` | Counter | `threshold_ms` | Incremented when the request's worst underrun stayed below `threshold_ms` |
+| `vllm_omni:audio_skipped_requests_total` | Counter | `reason` | Silent-loss counter — code2wav rejected malformed codec input and returned `200 OK` with empty audio |
 
 The continuity math comes from `vllm_omni/benchmarks/audio_continuity.py::compute_continuity_stats` so the server-side observation aligns with the bench-side definition.
 
-## Cross-Stage Transfer Metrics (`vllm:omni_`)
+## Diffusion Metrics (`vllm_omni:`)
+
+Per-request timing breakdowns for diffusion (image/video) stages. Emitted at request finalize when `stage_metrics.diffusion_metrics` is present. All carry `{model_name, stage, replica}`.
+
+| Metric | Type | Description |
+|--------|------|-------------|
+| `vllm_omni:diffusion_exec_s` | Histogram | DiT forward pass execution time per request in seconds |
+| `vllm_omni:diffusion_exec_per_step_s` | Histogram | DiT forward pass execution time per denoising step in seconds (`exec_s / num_inference_steps`) |
+| `vllm_omni:diffusion_preprocess_s` | Histogram | Diffusion input preprocessing time per request in seconds |
+| `vllm_omni:diffusion_postprocess_s` | Histogram | Diffusion output postprocessing (VAE decode) time per request in seconds |
+
+`diffusion_exec_s` uses `SECONDS_BUCKETS` (50 ms → 300 s); the other three use `SECONDS_FAST_BUCKETS` (1 ms → 60 s). `diffusion_exec_per_step_s` is skipped when `num_inference_steps` is not set.
+
+### Example PromQL
+
+```promql
+# P99 DiT exec time by stage
+histogram_quantile(0.99, rate(vllm_omni:diffusion_exec_s_bucket[5m]))
+
+# Ratio of preprocess to total exec wall time
+rate(vllm_omni:diffusion_preprocess_s_sum[5m])
+  / rate(vllm_omni:diffusion_exec_s_sum[5m])
+```
+
+## Cross-Stage Transfer Metrics (`vllm_omni:`)
 
 Per-physical-transfer histograms tracking the data hop between adjacent stages. Labels `{model_name, from_stage, from_replica, to_stage, to_replica}` let dashboards attribute latency to specific replica edges. `from_replica` / `to_replica` are resolved from the orchestrator's sticky-routing binding (`stage_pool.get_bound_replica_id(request_id)`), so no extra plumbing through `TransferEdgeStats` is needed.
 
 | Metric | Type | Description |
 |--------|------|-------------|
-| `vllm:omni_transfer_size_bytes` | Histogram | Per-transfer payload size in bytes |
-| `vllm:omni_transfer_tx_s` | Histogram | Sender-side time (serialize + submit to connector) |
-| `vllm:omni_transfer_rx_s` | Histogram | Receiver-side time (recv + deserialize) |
-| `vllm:omni_transfer_in_flight_s` | Histogram | Network in-flight time (TX done → RX recv start) |
+| `vllm_omni:transfer_size_bytes` | Histogram | Per-transfer payload size in bytes |
+| `vllm_omni:transfer_tx_s` | Histogram | Sender-side time (serialize + submit to connector) |
+| `vllm_omni:transfer_rx_s` | Histogram | Receiver-side time (recv + deserialize) |
+| `vllm_omni:transfer_in_flight_s` | Histogram | Network in-flight time (TX done → RX recv start) |
 
 ## vLLM Engine Metrics (`vllm:`)
 
@@ -83,9 +107,9 @@ For the full list of upstream metrics, see [the vLLM docs](https://github.com/vl
 
 | Metric group | Multi-stage LLM (Qwen3-Omni) |
 |---|---|
-| `vllm:omni_` request tracking + latency | With `--log-stats` |
-| `vllm:omni_` audio modality | With `--log-stats`, if pipeline has a talker stage |
-| `vllm:omni_` transfer | With `--log-stats`, if pipeline has ≥ 2 stages |
+| `vllm_omni:` request tracking + latency | With `--log-stats` |
+| `vllm_omni:` audio modality | With `--log-stats`, if pipeline has a talker stage |
+| `vllm_omni:` transfer | With `--log-stats`, if pipeline has ≥ 2 stages |
 | `vllm:` engine metrics (per `(stage, replica)`) | With `--log-stats` |
 | `vllm:` MFU metrics | With `--log-stats --enable-mfu-metrics` |
 
@@ -94,5 +118,5 @@ For the full list of upstream metrics, see [the vLLM docs](https://github.com/vl
 - All time-bearing metrics use the `_s` suffix (values in seconds). Buckets are `SECONDS_BUCKETS` for e2e / generation-style values and `SECONDS_FAST_BUCKETS` (1 ms → 60 s) for the fine-grained transfer and audio-underrun values.
 - Counters use the `_total` suffix (auto-appended by `prometheus_client`).
 - Sizes use the `_bytes` suffix.
-- All omni-specific families are prefixed `vllm:omni_`. The upstream `unregister_vllm_metrics()` function is monkey-patched (see `vllm_omni/patch.py`) to a scoped version that still strips upstream `vllm:*` collectors so multi-engine init within one process does not crash on duplicate registration, but preserves anything prefixed `vllm:omni_` / `vllm_omni`.
-- Text and audio first-output use distinct families (`vllm:time_to_first_token_seconds` reused from upstream for text; `vllm:omni_audio_ttfp_s` for audio) rather than a single metric with a `modality` label.
+- All omni-specific families are prefixed `vllm_omni:`. The upstream `unregister_vllm_metrics()` function is monkey-patched (see `vllm_omni/patch.py`) to a scoped version that still strips upstream `vllm:*` collectors so multi-engine init within one process does not crash on duplicate registration, but preserves anything prefixed `vllm_omni:`.
+- Text and audio first-output use distinct families (`vllm:time_to_first_token_seconds` reused from upstream for text; `vllm_omni:audio_ttfp_s` for audio) rather than a single metric with a `modality` label.

docs/user_guide/diffusion/attention_backends.md

@@ -1,494 +1,192 @@
 # Diffusion Attention Backends
 
-This document describes the diffusion attention backends available in vLLM-Omni, how to select them globally and per-role, the per-platform defaults, and how to use SageAttention.
+Use this page to select and configure a diffusion attention backend. Backend-
+specific installation, tuning, and compatibility details live in separate
+guides so that the selection contract stays easy to scan.
 
-## Overview
+Diffusion attention backend selection applies to DiT attention in image and
+video generation models. It does **not** change autoregressive LLM attention,
+which uses vLLM's attention selector.
 
-Diffusion attention backend selection is resolved in `vllm_omni.diffusion.attention.selector`. It looks up the backend from a structured `AttentionConfig` carried on `OmniDiffusionConfig` and falls back to the platform default when nothing is configured.
+For the internal selector, registry, and platform contract, see
+[Attention Backend Selection](../../design/feature/attention_backend_selection.md).
 
-This backend is used by diffusion attention layers such as the DiT attention in video and image generation models. It does **not** affect autoregressive (LLM) attention paths — those go through vLLM's own attention backend selector.
+## Choose a guide
 
-The full set of backends and their platform defaults is in the **Backend Options** and **Platform Defaults** sections below. If no attention backend is configured, vLLM-Omni asks the current platform to choose the default.
+| Need | Guide |
+| --- | --- |
+| Select a conservative or platform-native dense kernel | [Dense Backends](attention_backends/dense_backends.md) |
+| Use FlashInfer trtllm-gen FMHA, Skip-Softmax, or TRTLLM SAGE quantization | [TRTLLM Attention](attention_backends/trtllm.md) |
+| Install and use SageAttention 2.2 or SageAttention3 | [SageAttention](attention_backends/sage.md) |
+| Match training or rollout kernels loaded from Hugging Face | [Hugging Face Hub Backends](attention_backends/huggingface_hub.md) |
+| Use block-sparse video attention on Ascend NPU | [RainFusion](attention_backends/rainfusion.md) |
 
-## Backend Options
-
-| Value | Notes |
-|---|---|
-| `TRTLLM_ATTN` | FlashInfer's trtllm-gen FMHA (TensorRT-LLM's generated kernels, vendored by FlashInfer). Dense BF16, GQA native, `head_dim=128`. Datacenter Blackwell only (sm_100 / sm_103). Packed paths can provide `cu_seqlens` directly. Supports optional **Skip-Softmax** sparse attention — see below. Requires `flashinfer`. |
-| `FLASH_ATTN` | Wraps FlashAttention 4 on Blackwell when `flash-attn-4` is installed, then falls back to FlashAttention 3/2. Default on Hopper / Ada / Ampere when a compatible FlashAttention package is installed. |
-| `CUDNN_ATTN` | Pins `sdpa_kernel([CUDNN_ATTENTION])`. Default on Blackwell (sm_10x / sm_12x) with cuDNN ≥ 9.5. Wins on mask-heavy DiTs (HunyuanVideo-1.5: 2× e2e vs SDPA). |
-| `FLASHINFER_ATTN` | Uses FlashInfer's batch-prefill wrapper and supports mixed Q/K and V input dtypes through backend-specific configuration. Used as Blackwell fallback when cuDNN is unavailable. Requires `flashinfer` >= 0.6.16rc1 for mixed-dtype configurations. |
-| `TORCH_SDPA` | PyTorch `scaled_dot_product_attention` with the default backend dispatcher. Most conservative; always available. |
-| `SAGE_ATTN` | SageAttention 2.2 — INT8-quantized attention with FP16 accumulation. Lossy but typically visually indistinguishable on diffusion outputs. Requires `sageattention`. |
-| `SAGE_ATTN_3` | Requires `sageattn3` from `SageAttention/sageattention3_blackwell`. CUDA only, intended for Blackwell GPUs, with GQA/MQA requests falling back to PyTorch SDPA. |
-| `FLASH_ATTN_HUB` | FlashAttention 2 from HuggingFace `kernels` library. Useful for train/rollout alignment. |
-| `FLASH_ATTN_3_HUB` | FlashAttention 3 from HuggingFace `kernels` library. CUDA Hopper (sm_90+) only; falls back to `FLASH_ATTN_HUB` on older GPUs. |
-| `RAINFUSION_ATTN` | MindIE-SD **RainFusion** block-sparse video attention — see [below](#rainfusion_attn-backend-and-block-sparse-video-attention). Ascend NPU only; requires `mindiesd`. Delegates to `FLASH_ATTN` for anything that is not a packed video sequence. |
+## Backend options
 
+| Value | Family | Primary use | Detail |
+| --- | --- | --- | --- |
+| `TORCH_SDPA` | Dense | Conservative reference; always available | [Dense Backends](attention_backends/dense_backends.md#torch_sdpa) |
+| `FLASH_ATTN` | Dense | FlashAttention 4/3/2 depending on the installed package and GPU | [Dense Backends](attention_backends/dense_backends.md#flash_attn) |
+| `CUDNN_ATTN` | Dense | Mask-heavy DiTs on Blackwell with cuDNN 9.5 or newer | [Dense Backends](attention_backends/dense_backends.md#cudnn_attn) |
+| `FLASHINFER_ATTN` | Dense or quantized | FlashInfer batch prefill; optional mixed Q/K and V dtypes | [Dense Backends](attention_backends/dense_backends.md#flashinfer_attn) |
+| `TRTLLM_ATTN` | Dense, sparse, or quantized | Datacenter Blackwell with `head_dim=128` and compatible packed paths | [TRTLLM Attention](attention_backends/trtllm.md) |
+| `SAGE_ATTN` | Quantized | SageAttention 2.2 INT8 attention | [SageAttention](attention_backends/sage.md#sage_attn) |
+| `SAGE_ATTN_3` | Quantized | SageAttention3 on Blackwell | [SageAttention](attention_backends/sage.md#sage_attn_3) |
+| `FLASH_ATTN_HUB` | Hub kernel | FlashAttention 2 from Hugging Face `kernels` | [Hugging Face Hub Backends](attention_backends/huggingface_hub.md) |
+| `FLASH_ATTN_3_HUB` | Hub kernel | FlashAttention 3 from Hugging Face `kernels` on Hopper or newer | [Hugging Face Hub Backends](attention_backends/huggingface_hub.md) |
+| `RAINFUSION_ATTN` | Block sparse | MindIE-SD RainFusion video attention on Ascend NPU | [RainFusion](attention_backends/rainfusion.md) |
 
 ## Configuration
 
-Diffusion attention backends can be configured three ways, in priority order:
+Backend selection is resolved in this order:
 
-1. **`--diffusion-attention-config`** — structured per-role config (highest priority).
-2. **`--diffusion-attention-backend` / `DIFFUSION_ATTENTION_BACKEND` env var** — global shorthand that sets the default backend.
-3. **Platform default** — used when nothing is configured.
+1. `--diffusion-attention-config` per-role configuration.
+2. `--diffusion-attention-backend` or `DIFFUSION_ATTENTION_BACKEND` as a
+   global default.
+3. The current platform's default.
 
-`--diffusion-attention-backend` is shorthand for `--diffusion-attention-config.default.backend`. It may be combined with `--diffusion-attention-config.per_role.*` overrides, but is mutually exclusive with `--diffusion-attention-config.default.backend`.
+`--diffusion-attention-backend` is shorthand for
+`--diffusion-attention-config.default.backend`. Do not pass it together with an
+explicit `default.backend` in the structured configuration.
 
 ### Global default
 
-Set the default backend for every diffusion attention layer:
-
 ```bash
-# CLI flag
-vllm-omni serve <model> --diffusion-attention-backend SAGE_ATTN
+vllm-omni serve <model> --diffusion-attention-backend FLASH_ATTN
 
-# Environment variable (also recognized for backwards compatibility)
-export DIFFUSION_ATTENTION_BACKEND=SAGE_ATTN
+# Backwards-compatible environment variable
+export DIFFUSION_ATTENTION_BACKEND=FLASH_ATTN
 ```
 
 ### Per-role configuration
 
-Roles are free-form strings declared by each diffusion model. The two common categories are `"self"` and `"cross"`; model-specific roles (e.g. `"ltx2.audio_to_video"`) may also be declared. A role string is matched in this order:
+Roles are declared by each diffusion model. Common categories are `self` and
+`cross`; a model may also use a more specific role such as
+`ltx2.audio_to_video`. Resolution order is:
 
-1. Exact `per_role[role]` match
-2. `per_role[role_category]` fallback (e.g. `"ltx2.audio_to_video"` → `"cross"`)
-3. `default`
-4. Platform default
-
-Use vLLM-style dotted flags or one JSON blob:
+1. Exact `per_role[role]` match.
+2. Category `per_role[role_category]` match.
+3. `default`.
+4. Platform default.
 
 ```bash
 # Dotted flags
 vllm-omni serve <model> \
-    --diffusion-attention-config.default.backend FLASH_ATTN \
-    --diffusion-attention-config.per_role.cross.backend TORCH_SDPA
+  --diffusion-attention-config.default.backend FLASH_ATTN \
+  --diffusion-attention-config.per_role.cross.backend TORCH_SDPA
 
-# JSON
+# Equivalent JSON
 vllm-omni serve <model> \
-    --diffusion-attention-config '{"default":{"backend":"FLASH_ATTN"},"per_role":{"cross":{"backend":"TORCH_SDPA"}}}'
+  --diffusion-attention-config \
+  '{"default":{"backend":"FLASH_ATTN"},"per_role":{"cross":{"backend":"TORCH_SDPA"}}}'
 ```
 
-A backend that needs configuration exposes it as a typed field on the spec:
-`TRTLLM_ATTN`'s Skip-Softmax (see [below](#trtllm_attn-backend-and-skip-softmax)) and
-`RAINFUSION_ATTN`'s RainFusion block (see [below](#rainfusion_attn-backend-and-block-sparse-video-attention)):
-
-```bash
---diffusion-attention-config.default.backend TRTLLM_ATTN \
---diffusion-attention-config.default.skip_softmax.target_sparsity 0.5
-```
-
-### Programmatic API
-
-When constructing `OmniDiffusionConfig` directly:
+### Python API
 
 ```python
-from vllm_omni.diffusion.data import AttentionConfig, AttentionSpec, AttnQuantSpec, OmniDiffusionConfig
+from vllm_omni.diffusion.data import (
+    AttentionConfig,
+    AttentionSpec,
+    OmniDiffusionConfig,
+)
 
 config = OmniDiffusionConfig(
     diffusion_attention_config=AttentionConfig(
         default=AttentionSpec(backend="FLASH_ATTN"),
-        per_role={
-            "cross": AttentionSpec(backend="TORCH_SDPA"),
-        },
+        per_role={"cross": AttentionSpec(backend="TORCH_SDPA")},
     ),
     ...,
 )
 ```
 
-A plain dict is also accepted and normalized to `AttentionConfig`.
+Backend-specific typed blocks are documented with their consumers:
 
-#### FlashInfer QK16/V8 quantized attention
+- `quant`: [FlashInfer](attention_backends/dense_backends.md#flashinfer-quantized-attention)
+  and [TRTLLM SAGE](attention_backends/trtllm.md#sage-quantization).
+- `skip_softmax`: [TRTLLM Skip-Softmax](attention_backends/trtllm.md#skip-softmax).
+- `block_sparse`: [RainFusion](attention_backends/rainfusion.md#configuration).
 
-`FLASHINFER_ATTN` accepts quantization options through `AttentionSpec.quant`. For QK16/V8:
-
-```python
-config = OmniDiffusionConfig(
-    diffusion_attention_config=AttentionConfig(
-        default=AttentionSpec(
-            backend="FLASHINFER_ATTN",
-            quant=AttnQuantSpec(
-                dtype_qk="bfloat16",
-                dtype_vo="fp8_e4m3",
-            ),
-        ),
-    ),
-    ...,
-)
-```
-
-`dtype_qk` configures the Q and K input dtypes. `dtype_vo` configures the V input dtype.
-
-## Platform Defaults
+## Platform defaults
 
 ### Blackwell (sm_100 / sm_103 / sm_120 / sm_121)
 
-Auto-route preference, in order:
+The CUDA auto-route preference is:
 
-1. `TRTLLM_ATTN` — on **datacenter** Blackwell (sm_100 / sm_103) when `flashinfer` is installed, the model's `head_dim` is 128, **and the model declares a compatible packed/mask-free path**
-2. `CUDNN_ATTN` — when cuDNN ≥ 9.5 is available (ships in PyTorch 2.5+ wheels)
-3. `FLASHINFER_ATTN` — when `flashinfer` is installed but cuDNN < 9.5
-4. `FLASH_ATTN` — when `flash-attn` is installed with the Blackwell CUTE kernel
-5. `TORCH_SDPA` — last resort
+1. `TRTLLM_ATTN` on datacenter Blackwell (sm_100/sm_103) when FlashInfer is
+   available, `head_dim=128`, and the model declares a compatible packed or
+   mask-free path.
+2. `CUDNN_ATTN` when cuDNN 9.5 or newer is available.
+3. `FLASHINFER_ATTN` when FlashInfer is available but cuDNN is too old.
+4. `FLASH_ATTN` when a compatible package is installed.
+5. `TORCH_SDPA`.
 
-`TRTLLM_ATTN` is skipped on workstation Blackwell (sm_120 / sm_121) and for any `head_dim != 128`, so those GPUs keep the `CUDNN_ATTN` route described below.
+`TRTLLM_ATTN` is not auto-selected on workstation Blackwell
+(sm_120/sm_121), for other head dimensions, or for paths that require an
+unsupported mask.
 
-`TRTLLM_ATTN` outranks `CUDNN_ATTN` on datacenter Blackwell for compatible packed/mask-free pipelines. Workstation Blackwell (sm_120 / sm_121) and pipelines that require attention masks retain their normal fallback.
+### Hopper, Ada, and Ampere
 
-The startup log line `Defaulting to diffusion attention backend CUDNN_ATTN (Blackwell sm_120, cuDNN 91002)` confirms the route.
+The CUDA auto-route uses `FLASH_ATTN` when available and otherwise falls back
+to `TORCH_SDPA`. `CUDNN_ATTN` and `FLASHINFER_ATTN` remain explicit options.
 
-**Why CUDNN_ATTN by default**: on mask-heavy diffusion models (HunyuanVideo-1.5, Qwen-Image), cuDNN's pinned FMHA kernel sidesteps a PyTorch SDPA dispatch quirk where the unpinned dispatcher picks `EFFICIENT_ATTENTION` (~25 ms) for masked calls instead of cuDNN (~11 ms). The pin gives 2× e2e on HV-1.5 with no regression on lighter models.
+Other platforms validate an explicit backend and choose their own default
+through the platform implementation. Check the startup log to confirm the
+resolved backend.
 
-### Hopper (sm_90) / Ada (sm_89) / Ampere (sm_80–sm_86)
+## Choosing a backend manually
 
-Auto-route preference:
+Override the platform default when you need:
 
-1. `FLASH_ATTN` — when `flash-attn` is installed
-2. `TORCH_SDPA` — fallback
+- a correctness reference (`TORCH_SDPA`);
+- a backend-specific workaround;
+- training/rollout kernel alignment (Hub backends); or
+- an explicitly validated sparse or quantized speedup.
 
-`CUDNN_ATTN` and `FLASHINFER_ATTN` are still selectable via env var on these GPUs but are not in the auto-route — FlashAttention 2 is the well-tuned path on pre-Blackwell hardware.
+The startup log prints the resolved backend and whether it came from explicit
+configuration or platform defaulting. If no resolution message appears, check
+earlier logs for diffusion-stage initialization failures.
 
-## TRTLLM_ATTN Backend and Skip-Softmax
+## Reference benchmark
 
-`TRTLLM_ATTN` runs FlashInfer's trtllm-gen FMHA. Dense it is a BF16 backend like the others; on top of
-that it can enable **Skip-Softmax**, a sparse-attention mode that trades a little fidelity for speed
-(algorithm: [design doc](../../design/feature/skip_softmax.md)).
+The following BF16 results were measured on an sm_120 RTX Pro 6000 Blackwell
+with the same prompt and seed across runs. Treat them as reference results, not
+portable guarantees.
 
-Enable it through the typed `skip_softmax` block on the attention spec:
-
-| Key | Valid values | Meaning |
-|---|---|---|
-| `target_sparsity` | finite, `[0, 1]` | Operating point on a calibrated curve. Needs a calibration for the model. |
-| `threshold` | finite, `≥ 0` | Direct threshold, no calibration needed. Mutually exclusive with `target_sparsity`. |
-| `disabled_until_timestep` | finite, `[0, 1]` | Holds early, high-noise steps dense; skip turns on once the normalized timestep `t ≤ D`. |
-
-```bash
-vllm-omni serve Wan-AI/Wan2.2-T2V-A14B-Diffusers \
-  --diffusion-attention-config '{"default": {"backend": "TRTLLM_ATTN",
-      "skip_softmax": {"target_sparsity": 0.65, "disabled_until_timestep": 0.86}}}'
-```
-
-Programmatically the same block is a typed `SkipSoftmaxSpec` (values validated at construction):
-
-```python
-from vllm_omni.diffusion.data import AttentionConfig, AttentionSpec, SkipSoftmaxSpec
-
-AttentionConfig(
-    default=AttentionSpec(
-        backend="TRTLLM_ATTN",
-        skip_softmax=SkipSoftmaxSpec(target_sparsity=0.65, disabled_until_timestep=0.86),
-    ),
-)
-```
-
-**Start at `target_sparsity=0.65, disabled_until_timestep=0.86`** and raise `target_sparsity` only
-as far as your quality bar allows — the output diverges from dense as it climbs. The gain grows with
-sequence length, since Skip-Softmax only accelerates attention. Dense BF16 is `TRTLLM_ATTN` with no
-`skip_softmax` block.
-
-Requires datacenter Blackwell with `head_dim == 128`; elsewhere, or without FlashInfer, selecting
-`TRTLLM_ATTN` raises rather than silently degrading.
-
-## TRTLLM_ATTN SAGE Quantization
-
-On top of dense/Skip-Softmax, `TRTLLM_ATTN` can run **SAGE** (SageAttention) — the trtllm-gen
-FMHA kernel with per-block quantized Q/K and per-channel fp8 V, so the two attention matmuls run
-in low precision. Enable it through the typed `quant` block on the attention spec:
-
-| Key | Valid values | Meaning |
-|---|---|---|
-| `dtype_qk` | `int8`, `fp8_e4m3` | Quantization dtype for Q and K. Absent ⇒ dense (SAGE off). |
-| `q_block_size` | `1`, `4`, `16` | Per-token block size for Q scales (default `1`). |
-| `k_block_size` | `1`, `4`, `16` | Per-token block size for K scales (default `16`). |
-
-V is always quantized per-channel to fp8_e4m3 and K-smoothing is applied inside the routine, so
-neither is a user knob. Only block sizes `1`, `4`, `16` have compiled kernels.
-
-The `quant` block is a shared `AttnQuantSpec` consumed by more than one backend; each reads the
-fields that apply and rejects the rest. `TRTLLM_ATTN` (SAGE) reads `dtype_qk` (`int8` / `fp8_e4m3`),
-`q_block_size`, and `k_block_size`; `FLASHINFER_ATTN` reads `dtype_qk` (`float16` / `bfloat16`),
-`dtype_vo`, and `flashinfer_backend` (see its section). A `dtype_qk` from the wrong set raises.
-
-```bash
-vllm-omni serve Wan-AI/Wan2.2-T2V-A14B-Diffusers \
-  --diffusion-attention-config '{"default": {"backend": "TRTLLM_ATTN",
-      "quant": {"dtype_qk": "fp8_e4m3", "q_block_size": 1, "k_block_size": 16}}}'
-```
-
-```python
-from vllm_omni.diffusion.data import AttentionConfig, AttentionSpec, AttnQuantSpec
-
-AttentionConfig(
-    default=AttentionSpec(
-        backend="TRTLLM_ATTN",
-        quant=AttnQuantSpec(dtype_qk="fp8_e4m3", q_block_size=1, k_block_size=16),
-    ),
-)
-```
-
-**Requirements.** Needs FlashInfer ≥ 0.6.16rc1. Kernel
-availability is arch-dependent: `fp8_e4m3` QK has kernels on both **SM100** (B200) and **SM103**
-(B300); `int8` QK kernels are compiled for **SM100 only**.
-
-## RAINFUSION_ATTN Backend and Block-Sparse Video Attention
-
-`RAINFUSION_ATTN` runs MindIE-SD's RainFusion (`rf_v2`) kernel on Ascend NPU. It pools each
-128-token block of keys, ranks them per query block, and computes attention against only the
-top-scoring ones. Video tokens are rearranged into `(t, h, w)` order first, so the kept blocks are
-spatiotemporal neighbours rather than arbitrary rows of the packed sequence.
-
-Only the video segment is sparsified. The prefix rows (text, visual conditions, audio) and the
-first-frame blocks are always kept dense, which is why the realized sparsity is lower than the
-nominal `sparsity` you configure. Anything the kernel cannot handle — a warmup denoise step, an
-exempt layer, a sequence with no published video geometry, a video segment under 32 blocks —
-delegates to `FLASH_ATTN`, so a model can select this backend unconditionally.
-
-Enable it through the typed `block_sparse` block on the attention spec, shared by every
-block-sparse backend:
-
-| Key | Valid values | Meaning |
-|---|---|---|
-| `sparsity` | finite, `[0, 1]` | Nominal fraction of key blocks dropped per query block. `0` disables sparsity. Defaults to `0.8`. |
-| `start_step` | `≥ 0` | Keep the first N denoise steps dense. Layout is decided early, so these steps dominate structural fidelity. |
-| `skip_layers` | index selector, e.g. `"0-3,38"` | DiT blocks that always stay dense. |
-
-```bash
-vllm-omni serve MiniMaxAI/MiniMax-H3 \
-  --diffusion-attention-config '{"default": {"backend": "RAINFUSION_ATTN",
-      "block_sparse": {"sparsity": 0.8, "start_step": 0}}}'
-```
-
-Programmatically the same block is a typed `BlockSparseSpec` (values validated at construction):
-
-```python
-from vllm_omni.diffusion.data import AttentionConfig, AttentionSpec, BlockSparseSpec
-
-AttentionConfig(
-    default=AttentionSpec(
-        backend="RAINFUSION_ATTN",
-        block_sparse=BlockSparseSpec(sparsity=0.8, start_step=0, skip_layers="0-1"),
-    ),
-)
-```
-
-**Tune in the order `start_step` → `sparsity` → `skip_layers`.** Raise `start_step` first: it is the
-cheapest way to recover structure, because the early high-noise steps decide global layout while the
-later steps only refine texture. Then walk `sparsity` up to your quality bar. Reach for
-`skip_layers` last, once an A/B against dense at the same seed points at specific blocks.
-
-Requires Ascend NPU with `mindiesd`; selecting it on any other platform raises. It is also
-incompatible with ring sequence parallelism, since `rf_v2` needs the whole key sequence to rank
-blocks — use Ulysses SP (`ring_degree=1`).
-
-### Which geometries run sparse
-
-MindIE-SD rf_v2 handles arbitrary video grids by spatially rearranging video first and, when
-necessary, promoting a real-video suffix to its always-kept prefix segment. It never pads: rf_v2
-does not consume an attention mask for padded keys. The remaining video rows stay sparse, while the
-promoted rows remain visible to every query and key.
-
-For example, MiniMax-H3 at 1344x768 has grid `(62, 24, 42)`: its 62496 video rows leave a 32-row
-128-block residue, and its two non-8-wide latent columns form a 2928-row rearranged suffix.
-MindIE-SD promotes 2976 rows, leaving 59520 sparse rows (`465 × 128`). This keeps the block mask
-and kernel tiling aligned, at the cost of retaining more blocks.
-
-8x8-aligned spatial grids remain preferable for performance. A latent `h` or `w` not divisible by 8
-creates an always-kept residual suffix and lowers realized sparsity. "Multiple of 8" on the latent
-grid means **width and height that are multiples of 256**. This path requires a MindIE-SD release
-that implements protected video tails for rf_v2.
-
-## End-to-End Benchmark (BF16, sm_120 RTX Pro 6000 Blackwell)
-
-Same prompt and seed across runs. `Total generation time` from `text_to_video.py` / `text_to_image.py`.
-
-| Model | Shape | TORCH_SDPA | CUDNN_ATTN | FLASHINFER_ATTN |
-|---|---|---|---|---|
+| Model | Shape | `TORCH_SDPA` | `CUDNN_ATTN` | `FLASHINFER_ATTN` |
+| --- | --- | ---: | ---: | ---: |
 | HunyuanVideo-1.5 (T2V) | 480p / 33f / 50 steps | 147.05 s | **73.02 s** | 127.84 s |
 | Wan 2.2 14B (T2V) | 480p / 33f / 40 steps | 117.75 s | 117.17 s | **115.07 s** |
 | Qwen-Image (T2I) | 1024² / 50 steps | 17.41 s | **15.14 s** | 16.02 s |
 | FLUX.2-dev (T2I) | 1024² / 50 steps, TP=2 | 53.62 s | **53.30 s** | 54.94 s |
 
-Pattern: mask-heavy DiTs (HV-1.5, Qwen-Image) favor `CUDNN_ATTN`; lighter-mask DiTs and TP-saturated configs (Wan 2.2, FLUX.2 TP=2) tie within noise.
+Mask-heavy DiTs favored `CUDNN_ATTN`; lighter-mask or TP-saturated workloads
+were close enough that users should benchmark their exact model and shape.
 
-## Known Limitations
+## Compatibility anchors
 
-### LTX-2.0: `CUDNN_ATTN` crashes under torch.compile
+The following headings preserve links to sections that moved into dedicated
+guides.
 
-LTX-2's audio attention has a symbolic head_dim under torch.compile tracing. cuDNN's SDPA backend selector rejects symbolic dims and Dynamo aborts compilation. Tracked in [#3121](https://github.com/vllm-project/vllm-omni/issues/3121).
+## TRTLLM_ATTN Backend and Skip-Softmax
 
-**Workaround**: explicitly select `FLASHINFER_ATTN` or `TORCH_SDPA` for LTX-2.0:
+See [TRTLLM Attention](attention_backends/trtllm.md#skip-softmax).
 
-```bash
-DIFFUSION_ATTENTION_BACKEND=FLASHINFER_ATTN python examples/offline_inference/text_to_video/text_to_video.py \
-    --model Lightricks/LTX-2 ...
-```
+## TRTLLM_ATTN SAGE Quantization
 
-### FlashAttention-4 on Blackwell
+See [TRTLLM Attention](attention_backends/trtllm.md#sage-quantization).
 
-Install the optional CUDA 13 extra to use the CuTe-based FA4 path:
+## RAINFUSION_ATTN Backend and Block-Sparse Video Attention
 
-```bash
-pip install 'vllm-omni[fa4]'
-```
-
-`FLASH_ATTN` prefers `flash_attn.cute` on Blackwell and falls back to FA3/FA2
-when it is unavailable. Version `4.0.0b18` is required; earlier beta
-wheels had known JIT failures on Blackwell.
-
-## Choosing a Backend Manually
-
-### When to override the default
-
-- **Quality validation**: compare a new backend against `TORCH_SDPA` as the reference, since SDPA's default dispatcher is the most extensively tested.
-- **Lossy speedup hunting**: try `SAGE_ATTN` (INT8 quantized) on diffusion outputs — typically indistinguishable visually but always validate.
-- **Workaround for known issues**: see Known Limitations above.
-
-### Verifying which backend is in use
-
-The startup log prints one of:
-
-```
-Using diffusion attention backend 'CUDNN_ATTN'           # explicit override
-Defaulting to diffusion attention backend CUDNN_ATTN ... # auto-route
-Defaulting to diffusion attention backend SDPA           # nothing else available
-```
-
-If you don't see one of these, the model didn't reach diffusion stage init — check earlier logs for failures.
+See [RainFusion](attention_backends/rainfusion.md).
 
 ## SageAttention Installation
 
-vLLM-Omni expects SageAttention to be installed into the same Python environment as vLLM-Omni.
-
-Build from source:
-
-```bash
-git clone https://github.com/thu-ml/SageAttention.git
-cd SageAttention
-
-export EXT_PARALLEL=4 NVCC_APPEND_FLAGS="--threads 8" MAX_JOBS=32
-pip install . --no-build-isolation
-```
-
-Quick check:
-
-```bash
-python -c "import sageattention; print(sageattention.__file__)"
-```
+See [SageAttention](attention_backends/sage.md#installation).
 
 ## SageAttention3 Installation
 
-vLLM-Omni expects SageAttention3 to be installed into the same Python environment as vLLM-Omni.
-
-Build from source:
-
-```bash
-git clone https://github.com/thu-ml/SageAttention.git
-cd SageAttention/sageattention3_blackwell
-python setup.py install
-```
-
-Quick check:
-
-```bash
-python -c "import sageattn3; print(sageattn3.__file__)"
-```
-
-Notes:
-
-- `SAGE_ATTN_3` is only selected on CUDA when `sageattn3` is importable and the GPU is Blackwell-class.
-- SageAttention3's Blackwell kernel assumes `Hq == Hkv`. In vLLM-Omni, GQA/MQA diffusion requests fall back to PyTorch SDPA for correctness.
+See [SageAttention](attention_backends/sage.md#sageattention3-installation).
 
 ## HuggingFace Kernels Hub Backends
 
-To achieve perfect numerical consistency between **training** (typically using HuggingFace Diffusers with Hub kernels) and **serving/rollout** (using vLLM-Omni), you can use the Hub-based attention backends. This eliminates numerical drift / sampling divergence caused by executing different local kernel versions during rollout.
-
-The following backend options are supported:
-- `FLASH_ATTN_HUB` (HuggingFace `kernels-community/flash-attn2`)
-- `FLASH_ATTN_3_HUB` (HuggingFace `kernels-community/flash-attn3`, Hopper sm_90+ only)
-
-### Installation
-
-To use these backends, you must install the `kernels` library:
-
-```bash
-pip install kernels==0.14.1
-```
-
-If the `kernels` library is not available in the environment, vLLM-Omni will log a warning and fall back gracefully to the corresponding local backend implementations (`FLASH_ATTN`). On CUDA GPUs below Hopper (compute capability < 9.0), `FLASH_ATTN_3_HUB` falls back to `FLASH_ATTN_HUB`.
-
-### Usage
-
-Select a Hub backend using the global CLI flag or environment variables:
-
-```bash
-# Environment variable
-export DIFFUSION_ATTENTION_BACKEND=FLASH_ATTN_HUB
-
-# CLI flag
-vllm-omni serve <model> --diffusion-attention-backend FLASH_ATTN_HUB
-```
-
-## Usage Examples
-
-### Default (auto-route)
-
-```bash
-python examples/offline_inference/text_to_video/text_to_video.py \
-    --model hunyuanvideo-community/HunyuanVideo-1.5-Diffusers-480p_t2v \
-    --prompt "A dog running across a field of golden wheat." \
-    --height 480 --width 832 --num-frames 33 \
-    --num-inference-steps 50 --seed 42 --guidance-scale 6.0 \
-    --output hv15.mp4
-```
-
-On Blackwell this picks `CUDNN_ATTN` automatically. Check the log for the `Defaulting to ...` line.
-
-### Explicit backend selection
-
-```bash
-DIFFUSION_ATTENTION_BACKEND=FLASHINFER_ATTN python examples/offline_inference/text_to_video/text_to_video.py \
-    --model Lightricks/LTX-2 \
-    --prompt "A dog running across a field of golden wheat." \
-    --height 480 --width 832 --num-frames 33 \
-    --num-inference-steps 40 --seed 42 --guidance-scale 4.0 \
-    --output ltx2.mp4
-```
-
-### SageAttention (lossy)
-
-```bash
-DIFFUSION_ATTENTION_BACKEND=SAGE_ATTN python examples/offline_inference/text_to_video/text_to_video.py \
-    --model hunyuanvideo-community/HunyuanVideo-1.5-Diffusers-480p_t2v \
-    --prompt "A dog running across a field of golden wheat." \
-    --height 480 --width 832 --num-frames 33 \
-    --num-inference-steps 30 --seed 42 --guidance-scale 6.0 \
-    --tensor-parallel-size 2 \
-    --output hv15_sage.mp4
-```
-
-Example: Wan2.2 TI2V 5B
-
-```bash
-DIFFUSION_ATTENTION_BACKEND=SAGE_ATTN python examples/offline_inference/text_to_video/text_to_video.py \
-    --model Wan-AI/Wan2.2-TI2V-5B-Diffusers \
-    --prompt "A dog running across a field of golden wheat." \
-    --height 704 --width 1280 --num-frames 49 \
-    --num-inference-steps 30 --seed 42 --guidance-scale 5.0 \
-    --tensor-parallel-size 2 \
-    --output outputs/wan22_sage.mp4
-```
-
-### Enable SageAttention3
-
-Example:
-
-```bash
-DIFFUSION_ATTENTION_BACKEND=SAGE_ATTN_3 python examples/offline_inference/text_to_video/text_to_video.py \
-    --model hunyuanvideo-community/HunyuanVideo-1.5-Diffusers-480p_t2v \
-    --prompt "A dog running across a field of golden wheat." \
-    --height 480 --width 832 --num-frames 33 \
-    --num-inference-steps 30 --seed 42 --guidance-scale 6.0 \
-    --tensor-parallel-size 2 \
-    --output outputs/hv15_sage3.mp4
-```
+See [Hugging Face Hub Backends](attention_backends/huggingface_hub.md).

docs/user_guide/diffusion/attention_backends/dense_backends.md

@@ -0,0 +1,110 @@
+# Dense Attention Backends
+
+Use these backends when you want dense attention without backend-specific
+sparsity. Start with `TORCH_SDPA` for correctness comparisons, then benchmark
+the fastest compatible kernel for your model, shape, and hardware.
+
+For selection precedence and per-role configuration, see the
+[attention backend overview](../attention_backends.md).
+
+## `TORCH_SDPA`
+
+`TORCH_SDPA` calls PyTorch `scaled_dot_product_attention` and lets PyTorch's
+dispatcher choose the implementation. It is always available and is the most
+conservative reference when validating another backend.
+
+```bash
+vllm-omni serve <model> --diffusion-attention-backend TORCH_SDPA
+```
+
+## `FLASH_ATTN`
+
+`FLASH_ATTN` uses the installed FlashAttention implementation. On Blackwell it
+prefers the CuTe-based FlashAttention 4 path and falls back to
+FlashAttention 3 or 2 when that path is unavailable. On Hopper, Ada, and
+Ampere it is the preferred automatic route when a compatible package is
+installed.
+
+```bash
+vllm-omni serve <model> --diffusion-attention-backend FLASH_ATTN
+```
+
+### FlashAttention 4 on Blackwell
+
+Install the optional CUDA 13 extra:
+
+```bash
+pip install 'vllm-omni[fa4]'
+```
+
+Version `4.0.0b18` is required; earlier beta wheels had known JIT failures on
+Blackwell. If the CuTe path is unavailable, the backend falls back to the
+compatible FlashAttention 3 or 2 path.
+
+## `CUDNN_ATTN`
+
+`CUDNN_ATTN` pins PyTorch SDPA to `CUDNN_ATTENTION`. It is particularly useful
+for mask-heavy DiTs and is automatically preferred on Blackwell when cuDNN
+9.5 or newer is available and the higher-priority TRTLLM route is not
+compatible.
+
+```bash
+vllm-omni serve <model> --diffusion-attention-backend CUDNN_ATTN
+```
+
+### LTX-2.0 limitation
+
+LTX-2 audio attention has a symbolic head dimension during `torch.compile`
+tracing. The cuDNN SDPA selector rejects that symbolic dimension and Dynamo
+aborts compilation. This is tracked in
+[issue #3121](https://github.com/vllm-project/vllm-omni/issues/3121).
+
+Use `FLASHINFER_ATTN` or `TORCH_SDPA` as a workaround:
+
+```bash
+DIFFUSION_ATTENTION_BACKEND=FLASHINFER_ATTN \
+  python examples/offline_inference/text_to_video/text_to_video.py \
+  --model Lightricks/LTX-2 ...
+```
+
+## `FLASHINFER_ATTN`
+
+`FLASHINFER_ATTN` uses FlashInfer's batch-prefill wrapper. It is an explicit
+option on CUDA platforms and an automatic Blackwell fallback when FlashInfer
+is installed but cuDNN is too old for `CUDNN_ATTN`.
+
+```bash
+vllm-omni serve <model> --diffusion-attention-backend FLASHINFER_ATTN
+```
+
+### FlashInfer quantized attention
+
+The backend accepts an `AttentionSpec.quant` block. For QK16/V8, keep Q and K
+in FP16 or BF16 and use FP8 E4M3 for V:
+
+```python
+from vllm_omni.diffusion.data import (
+    AttentionConfig,
+    AttentionSpec,
+    AttnQuantSpec,
+    OmniDiffusionConfig,
+)
+
+config = OmniDiffusionConfig(
+    diffusion_attention_config=AttentionConfig(
+        default=AttentionSpec(
+            backend="FLASHINFER_ATTN",
+            quant=AttnQuantSpec(
+                dtype_qk="bfloat16",
+                dtype_vo="fp8_e4m3",
+            ),
+        ),
+    ),
+    ...,
+)
+```
+
+`dtype_qk` controls Q and K; `dtype_vo` controls V. Mixed-dtype
+configurations require FlashInfer 0.6.16rc1 or newer. The shared quantization
+schema is also consumed by TRTLLM, but each backend validates its own allowed
+fields and values; see [TRTLLM SAGE quantization](trtllm.md#sage-quantization).

docs/user_guide/diffusion/attention_backends/huggingface_hub.md

@@ -0,0 +1,32 @@
+# Hugging Face Hub Attention Backends
+
+Hub backends load the same published kernels commonly used with Hugging Face
+Diffusers. They are useful when training and serving must use matching kernel
+implementations to reduce numerical drift and sampling divergence.
+
+| Backend | Kernel | Platform |
+| --- | --- | --- |
+| `FLASH_ATTN_HUB` | `kernels-community/flash-attn2` | Compatible CUDA GPUs |
+| `FLASH_ATTN_3_HUB` | `kernels-community/flash-attn3` | Hopper sm_90 or newer |
+
+## Installation
+
+```bash
+pip install kernels==0.14.1
+```
+
+If `kernels` is unavailable, vLLM-Omni warns and falls back to the
+corresponding local FlashAttention implementation. On CUDA devices older than
+Hopper, `FLASH_ATTN_3_HUB` falls back to `FLASH_ATTN_HUB`.
+
+## Usage
+
+```bash
+export DIFFUSION_ATTENTION_BACKEND=FLASH_ATTN_HUB
+
+# Equivalent CLI selection
+vllm-omni serve <model> --diffusion-attention-backend FLASH_ATTN_HUB
+```
+
+For per-role kernel alignment, configure the Hub backend on only the relevant
+role; see the [attention backend overview](../attention_backends.md#per-role-configuration).

docs/user_guide/diffusion/attention_backends/rainfusion.md

@@ -0,0 +1,72 @@
+# RainFusion Attention
+
+`RAINFUSION_ATTN` runs MindIE-SD RainFusion (`rf_v2`) block-sparse video
+attention on Ascend NPU. It pools 128-token key blocks, ranks them per query
+block, and attends to the highest-scoring blocks after arranging video tokens
+in `(t, h, w)` order.
+
+Only the video segment is sparse. Prefix rows such as text, visual conditions,
+and audio, plus first-frame blocks, remain dense. Unsupported calls—including
+warmup steps, skipped layers, missing video geometry, and video segments below
+32 blocks—delegate to `FLASH_ATTN`, so compatible models can select RainFusion
+globally.
+
+## Configuration
+
+| Key | Valid values | Meaning |
+| --- | --- | --- |
+| `sparsity` | finite, `[0, 1]` | Nominal dropped-key-block fraction; default `0.8`; `0` disables sparsity |
+| `start_step` | integer, `>= 0` | Number of early denoise steps kept dense |
+| `skip_layers` | selector such as `"0-3,38"` | DiT blocks kept dense |
+
+```bash
+vllm-omni serve MiniMaxAI/MiniMax-H3 \
+  --diffusion-attention-config '{"default":{"backend":"RAINFUSION_ATTN",\
+    "block_sparse":{"sparsity":0.8,"start_step":0}}}'
+```
+
+```python
+from vllm_omni.diffusion.data import AttentionConfig, AttentionSpec, BlockSparseSpec
+
+config = AttentionConfig(
+    default=AttentionSpec(
+        backend="RAINFUSION_ATTN",
+        block_sparse=BlockSparseSpec(
+            sparsity=0.8,
+            start_step=0,
+            skip_layers="0-1",
+        ),
+    ),
+)
+```
+
+Tune `start_step`, then `sparsity`, then `skip_layers`. Increasing the dense
+early-step window is usually the cheapest way to recover global structure;
+use layer exclusions only after a same-seed dense comparison identifies
+sensitive blocks.
+
+## Requirements and compatibility
+
+RainFusion requires Ascend NPU and `mindiesd`. Selecting it on another
+platform raises. It is incompatible with ring sequence parallelism because
+the kernel needs the complete key sequence for ranking; use Ulysses sequence
+parallelism with `ring_degree=1`.
+
+## Geometry handling
+
+RainFusion handles arbitrary video grids by rearranging video spatially and,
+when necessary, promoting a real-video suffix to the always-kept prefix. It
+does not pad because `rf_v2` does not consume a padding attention mask.
+
+For MiniMax-H3 at 1344x768, grid `(62, 24, 42)` produces 62,496 video rows.
+The implementation promotes 2,976 rows and leaves 59,520 sparse rows
+(`465 x 128`) so the mask and kernel tiling remain aligned.
+
+Spatial grids aligned to 8x8 are still preferable. A latent height or width
+not divisible by 8 creates an always-kept suffix and reduces realized
+sparsity. At the image level, that means width and height divisible by 256.
+Protected video tails require a compatible MindIE-SD release.
+
+For common configuration and selector behavior, see the
+[attention backend overview](../attention_backends.md) and the
+[backend selection design](../../../design/feature/attention_backend_selection.md).

docs/user_guide/diffusion/attention_backends/sage.md

@@ -0,0 +1,63 @@
+# SageAttention
+
+SageAttention backends provide lossy low-precision attention for diffusion
+models. Validate output quality against `TORCH_SDPA` at the same seed before
+using either backend in production.
+
+## `SAGE_ATTN`
+
+`SAGE_ATTN` uses SageAttention 2.2 with INT8-quantized attention and FP16
+accumulation.
+
+### Installation
+
+Install SageAttention into the same environment as vLLM-Omni:
+
+```bash
+git clone https://github.com/thu-ml/SageAttention.git
+cd SageAttention
+export EXT_PARALLEL=4 NVCC_APPEND_FLAGS="--threads 8" MAX_JOBS=32
+pip install . --no-build-isolation
+```
+
+Verify the installation:
+
+```bash
+python -c "import sageattention; print(sageattention.__file__)"
+```
+
+Select it globally:
+
+```bash
+vllm-omni serve <model> --diffusion-attention-backend SAGE_ATTN
+```
+
+## `SAGE_ATTN_3`
+
+`SAGE_ATTN_3` uses the SageAttention3 Blackwell implementation.
+
+### SageAttention3 installation
+
+```bash
+git clone https://github.com/thu-ml/SageAttention.git
+cd SageAttention/sageattention3_blackwell
+python setup.py install
+```
+
+Verify the installation:
+
+```bash
+python -c "import sageattn3; print(sageattn3.__file__)"
+```
+
+```bash
+vllm-omni serve <model> --diffusion-attention-backend SAGE_ATTN_3
+```
+
+`SAGE_ATTN_3` requires CUDA, an importable `sageattn3`, and a Blackwell-class
+GPU. Its kernel assumes the query-head count equals the key/value-head count.
+GQA and MQA diffusion calls therefore fall back to PyTorch SDPA for
+correctness.
+
+For common configuration and platform routing, see the
+[attention backend overview](../attention_backends.md).

docs/user_guide/diffusion/attention_backends/trtllm.md

@@ -0,0 +1,103 @@
+# TRTLLM Attention
+
+`TRTLLM_ATTN` runs FlashInfer's trtllm-gen FMHA kernels. It supports dense
+BF16 attention plus optional Skip-Softmax sparsity or SAGE quantization.
+
+For common selection and per-role configuration, see the
+[attention backend overview](../attention_backends.md).
+
+## Requirements
+
+Selecting `TRTLLM_ATTN` requires:
+
+- datacenter Blackwell, sm_100 or sm_103;
+- `head_dim=128`;
+- FlashInfer; and
+- a model-declared packed or mask-free compatible path.
+
+It is not supported on workstation Blackwell (sm_120/sm_121). An incompatible
+explicit selection raises instead of silently changing the backend.
+
+## Dense mode
+
+Select `TRTLLM_ATTN` without `skip_softmax` or `quant` to run dense BF16:
+
+```bash
+vllm-omni serve <model> --diffusion-attention-backend TRTLLM_ATTN
+```
+
+## Skip-Softmax
+
+Skip-Softmax trades some fidelity for speed by skipping low-contribution
+attention work. See the [feature design](../../../design/feature/skip_softmax.md)
+for the algorithm and implementation contract.
+
+| Key | Valid values | Meaning |
+| --- | --- | --- |
+| `target_sparsity` | finite, `[0, 1]` | Point on a model-calibrated curve |
+| `threshold` | finite, `>= 0` | Direct threshold; mutually exclusive with `target_sparsity` |
+| `disabled_until_timestep` | finite, `[0, 1]` | Keep early high-noise steps dense until normalized `t <= D` |
+
+```bash
+vllm-omni serve Wan-AI/Wan2.2-T2V-A14B-Diffusers \
+  --diffusion-attention-config '{"default":{"backend":"TRTLLM_ATTN",\
+    "skip_softmax":{"target_sparsity":0.65,"disabled_until_timestep":0.86}}}'
+```
+
+```python
+from vllm_omni.diffusion.data import AttentionConfig, AttentionSpec, SkipSoftmaxSpec
+
+config = AttentionConfig(
+    default=AttentionSpec(
+        backend="TRTLLM_ATTN",
+        skip_softmax=SkipSoftmaxSpec(
+            target_sparsity=0.65,
+            disabled_until_timestep=0.86,
+        ),
+    ),
+)
+```
+
+Start with `target_sparsity=0.65` and `disabled_until_timestep=0.86`. Increase
+sparsity only after comparing quality against dense output at the same seed.
+The benefit grows with sequence length because only attention is accelerated.
+
+## SAGE quantization
+
+TRTLLM SAGE quantizes Q/K per block and V per channel so both attention
+matrix multiplications use low precision.
+
+| Key | Valid values | Meaning |
+| --- | --- | --- |
+| `dtype_qk` | `int8`, `fp8_e4m3` | Q/K quantization dtype; absent means dense |
+| `q_block_size` | `1`, `4`, `16` | Q scale block size; default `1` |
+| `k_block_size` | `1`, `4`, `16` | K scale block size; default `16` |
+
+V is always quantized per channel to FP8 E4M3 and K-smoothing is internal.
+Only block sizes with compiled kernels are accepted.
+
+```bash
+vllm-omni serve Wan-AI/Wan2.2-T2V-A14B-Diffusers \
+  --diffusion-attention-config '{"default":{"backend":"TRTLLM_ATTN",\
+    "quant":{"dtype_qk":"fp8_e4m3","q_block_size":1,"k_block_size":16}}}'
+```
+
+```python
+from vllm_omni.diffusion.data import AttentionConfig, AttentionSpec, AttnQuantSpec
+
+config = AttentionConfig(
+    default=AttentionSpec(
+        backend="TRTLLM_ATTN",
+        quant=AttnQuantSpec(
+            dtype_qk="fp8_e4m3",
+            q_block_size=1,
+            k_block_size=16,
+        ),
+    ),
+)
+```
+
+This path requires FlashInfer 0.6.16rc1 or newer. FP8 Q/K kernels are
+available on sm_100 and sm_103; INT8 Q/K kernels are compiled for sm_100 only.
+The shared `AttnQuantSpec` is backend typed: values intended for
+`FLASHINFER_ATTN` are rejected here and vice versa.

docs/user_guide/diffusion/cpu_offload.md

@@ -1,389 +1,75 @@
 # CPU Offloading for Diffusion Models
 
-## Overview
+vLLM-Omni provides three CPU-offload strategies for diffusion models. Choose
+the coarsest strategy that meets your memory target: finer-grained streaming
+saves more device memory but adds more transfer and setup overhead.
 
-vLLM-Omni provides two offloading strategies to reduce GPU memory usage for diffusion models:
+For the shared factory, discovery, and lifecycle contract, see the
+[CPU Offloading design](../../design/feature/offloader/README.md).
 
-1. **Model-level (Sequential) Offloading**: Mutual exclusion between DiT model and encoder - only one is on GPU at a time.
-2. **Layerwise (Blockwise) Offloading**: Keeps only one transformer block on GPU at a time with compute-memory overlap.
+## Choose a strategy
 
-Both strategies use pinned memory for faster CPU-GPU transfers. The strategies are **mutually exclusive** for now - if both are enabled, layerwise takes priority.
+| Strategy | Device residency | Parallel scope | Guide |
+| --- | --- | --- | --- |
+| Model-level (sequential) | One pipeline component group at a time | Single device | [Model-Level Offloading](offloader/cpu_offload.md) |
+| Layerwise (blockwise) | One transformer block, with next-block prefetch | Single device | [Layerwise Offloading](offloader/layerwise_offload.md) |
+| Distributed layerwise | Fixed two-block device buffer; optional host-weight sharding and AllGather | Multiple GPU/NPU ranks | [Distributed Layerwise Offloading](offloader/distributed_layerwise_offload.md) |
 
+All strategies use pinned host memory for faster transfers where applicable.
+Configuration priority is:
+
+1. Distributed layerwise offloading.
+2. Layerwise offloading.
+3. Model-level offloading.
+
+Treat the flags as mutually exclusive. If more than one is enabled, the
+higher-priority strategy is selected.
+
+## Quick selection
+
+- Use [model-level offloading](offloader/cpu_offload.md) when swapping whole
+  encoders and DiTs is enough to fit and phase-boundary transfers are
+  acceptable.
+- Use [layerwise offloading](offloader/layerwise_offload.md) for compute-heavy video
+  DiTs where block transfers can overlap computation.
+- Use [distributed layerwise offloading](offloader/distributed_layerwise_offload.md)
+  when a multi-rank deployment also needs bounded device residency and,
+  optionally, sharded host weights.
+
+## Supported models
+
+| Architecture | Example models | DiT class | Model-level | Layerwise | Distributed layerwise | Layerwise block attributes |
+| --- | --- | --- | :---: | :---: | :---: | --- |
+| Flux2Pipeline | `black-forest-labs/FLUX.2-dev` | `Flux2Transformer2DModel` | yes | yes | — | `transformer_blocks`, `single_transformer_blocks` |
+| LongCatImagePipeline | `meituan-longcat/LongCat-Image` | `LongCatImageTransformer2DModel` | — | yes | — | `transformer_blocks`, `single_transformer_blocks` |
+| NextStep11Pipeline | `stepfun-ai/NextStep-1.1` | `NextStepModel` | — | yes | — | `layers` |
+| OvisImagePipeline | `AIDC-AI/Ovis-Image-7B` | `OvisImageTransformer2DModel` | — | yes | — | `transformer` |
+| QwenImagePipeline | `Qwen/Qwen-Image` | `QwenImageTransformer2DModel` | yes | yes | — | `transformer_blocks` |
+| StableDiffusionXLPipeline | `stabilityai/stable-diffusion-xl-base-1.0` | `SDXLUNet2DConditionModel` | yes | yes | — | `down_blocks`, `up_blocks` |
+| StableDiffusion3Pipeline | `stabilityai/stable-diffusion-3.5-medium` | `SD3Transformer2DModel` | — | yes | — | `transformer_blocks` |
+| Wan22I2VPipeline | `Wan-AI/Wan2.2-I2V-A14B-Diffusers` | `WanTransformer3DModel` | yes | yes | — | `blocks` |
+| Wan22Pipeline | `Wan-AI/Wan2.2-T2V-A14B-Diffusers` | `WanTransformer3DModel` | yes | yes | — | `blocks` |
+| SoulXSingerPipeline / SoulXSingerSVCPipeline | `Soul-AILab/SoulX-Singer` | `DiffLlama` | yes | yes | — | `layers` |
+| BagelPipeline | `ByteDance-Seed/BAGEL-7B-MoT` | `Qwen2MoTModel` | — | yes | — | `layers`, customized modules |
+| Cosmos3OmniDiffusersPipeline | `nvidia/Cosmos3-Nano`, `nvidia/Cosmos3-Super` | `Cosmos3VFMTransformer`, `Cosmos3LanguageModel` | yes | yes | yes | `layers`, `gen_layers` |
+
+Model-level support requires discoverable DiT and encoder components.
+Layerwise support requires transformer block topology. Distributed support
+reuses that topology but still requires validation for the model, checkpoint,
+and parallel configuration.
+
+## Compatibility anchors
+
+These headings preserve links to sections that moved into dedicated guides.
 
 ## Model-level (Sequential) Offloading
 
-### How It Works
-
-Model-level offloading implements mutual exclusion between DiT transformer and encoder modules using pre forward hooks:
-
-- **When encoders run**: DiT transformer is offloaded to CPU
-- **When DiT runs**: Encoders are offloaded to CPU, if more than one dit models, only one loaded on GPU, others get offloaded to CPU.
-- **VAE**: Stays resident on GPU
-
-Before each module's forward pass, the hook automatically moves it to GPU while offloading the other module group to CPU. Transfers use pinned memory for speed.
-
-### Usage
-
-**Python API:**
-```python
-from vllm_omni import Omni
-
-m = Omni(model="Wan-AI/Wan2.2-T2V-A14B-Diffusers", enable_cpu_offload=True)
-```
-
-**CLI:**
-```bash
-vllm serve Wan-AI/Wan2.2-T2V-A14B-Diffusers --omni --enable-cpu-offload
-```
-
-### To Support a Model
-
-Implement the `SupportsComponentDiscovery` protocol to declare which
-submodules serve as pipeline components (used by offloading, HSDP
-sharding, and other framework features):
-
-```python
-from typing import ClassVar
-from vllm_omni.diffusion.models.interface import SupportsComponentDiscovery
-
-class MyPipeline(nn.Module, SupportsComponentDiscovery):
-    _dit_modules: ClassVar[list[str]] = ["transformer"]
-    _encoder_modules: ClassVar[list[str]] = ["text_encoder", "vision_model"]
-    _vae_modules: ClassVar[list[str]] = ["vae"]
-    _resident_modules: ClassVar[list[str]] = []  # optional
-
-    def __init__(self):
-        super().__init__()
-        self.transformer = ...     # DiT — stays on GPU during denoising
-        self.text_encoder = ...    # Encoder — offloaded to CPU during denoising
-        self.vision_model = ...    # Encoder — offloaded to CPU during denoising
-        self.vae = ...             # VAE — always on GPU
-```
-
-- `_dit_modules`: attribute names of denoising submodules (kept on GPU
-  during the diffusion loop).
-- `_encoder_modules`: attribute names of encoder/vision submodules
-  (offloaded to CPU during the diffusion loop).
-- `_vae_modules`: attribute names of VAE(s) (always kept on GPU, not
-  part of the mutual exclusion hooks).
-- `_resident_modules`: attribute names of small submodules that must
-  stay on GPU during layerwise offloading (e.g. embedders, connectors).
-  Optional — defaults to `[]`.
-
-All attribute names support dotted paths for nested submodules
-(e.g. `"pipe.transformer"`, `"bagel.time_embedder"`).
-
-Both DiT and encoder lists are needed because the offload hooks use
-mutual exclusion: when one group runs, the other moves to CPU.
-
-### Limitations
-- Cold start latency increases
-- Adds overhead from CPU-GPU transfers between encoder and denoising phases
-- Support single GPU only for now
-
-
-### Component offloading for split models (e.g. Cosmos3)
-
-Some models split their transformer into mutually-exclusive *components* that run
-in different phases of a single forward pass rather than as separate pipeline
-components -- e.g. Cosmos3's understanding (reasoner) component runs once per
-generation while the generation (generator) component runs every denoising step.
-Such models have no separate text encoder to swap against, so the transformer
-owns a small model-local offload path and wraps each phase with
-`with self._offload_context(name):`
-
-```python
-class Cosmos3VFMTransformer(nn.Module):
-    def forward(self, ...):
-        with self._offload_context("reasoner"):
-            ...  # understanding pass, runs once
-        with self._offload_context("generator"):
-            ...  # denoising pass, runs every step
-```
-
-Model-level offloading then keeps exactly one component GPU-resident at a time
-(the other on CPU), reusing the same `SequentialOffloadHook` `.to()` movers. The
-pipeline opts in by exposing `enable_omni_model_cpu_offload` (which drives the
-transformer's `enable_model_cpu_offload` and pins the VAE). Layerwise offloading
-works for these models too -- each component declares its own block container via
-`_layerwise_offload_blocks_attrs`.
-
+See [Model-Level Offloading](offloader/cpu_offload.md).
 
 ## Layerwise (Blockwise) Offloading
 
-### How It Works
-
-Layerwise offloading keeps only one transformer block on GPU at a time.
-
-As each block completes, the next block is prefetched to GPU while the current block is freed. The pre and forward hooks utilized by layerwise offloading apply a separate CUDA stream (`copy_stream`) to overlap weight transfer with computation, and retain flattened tensors in pinned CPU memory for block parameters re-materialization. Encoders, VAE, and non-block DiT modules (embeddings, norms) always stay on GPU.
-
-**Execution Flow:**
-
-| Block | Pre-forward Hook | Forward | Post-forward Hook |
-|-------|------------------|---------|-------------------|
-| block-0 | Prefetch block-1 (async) | Compute block-0 | Free block-0 |
-| block-1 | Prefetch block-2 (async) | Compute block-1 | Free block-1 |
-| ... | ... | ... | ... |
-| block-(n-1) | **Prefetch block-0** (async) | Compute block-(n-1) | Free block-(n-1) |
-
-Each transformer block has a `LayerwiseOffloadHook` that prefetches the next block before forward and frees the current block after forward.
-
-Layerwise offloading is primarily recommended for large **video generation models** where the compute cost per block is high enough to effectively overlap with memory prefetch operations. For example, Wan2.2 T2V and I2V pipelines.
-
-### Usage
-
-**Python API:**
-```python
-from vllm_omni import Omni
-
-# Text-to-video
-m = Omni(model="Wan-AI/Wan2.2-T2V-A14B-Diffusers", enable_layerwise_offload=True)
-
-# Or image-to-video
-m = Omni(model="Wan-AI/Wan2.2-I2V-A14B-Diffusers", enable_layerwise_offload=True)
-```
-
-**CLI:**
-```bash
-# Text-to-video
-vllm serve Wan-AI/Wan2.2-T2V-A14B-Diffusers --omni --enable-layerwise-offload
-
-# Or image-to-video
-vllm serve Wan-AI/Wan2.2-I2V-A14B-Diffusers --omni --enable-layerwise-offload
-```
-
-### To Support a Model
-
-Models must define the blocks attribute name for layerwise offloading:
-
-```python
-class WanTransformer3DModel(nn.Module):
-    _layerwise_offload_blocks_attrs = ["blocks"]  # Attribute names containing transformer blocks
-
-    def __init__(self):
-        self.blocks = nn.ModuleList([...])  # Transformer blocks
-```
-
-For models with multiple block types:
-
-```python
-class Flux2Transformer2DModel(nn.Module):
-    _layerwise_offload_blocks_attrs = ["transformer_blocks", "single_transformer_blocks"]
-```
-
-### Limitations
-- Cold start latency increases because offloaded components must be moved to CPU
-  during setup; layerwise offload may add extra weight consolidation and pinning
-  work.
-- Performance depends on compute cost and H2D bandwidth as well
-- Support single GPU only for now
-
+See [Layerwise Offloading](offloader/layerwise_offload.md).
 
 ## Distributed Layerwise Offloading
 
-### How It Works
-
-Distributed layerwise offloading extends single-GPU layerwise offloading to
-multi-device deployments. In the default mode, each rank stores only
-**1/dp_size** of the model weights in host memory; full layer weights are
-reconstructed at runtime via **AllGather** on a dedicated communication stream,
-overlapped with computation via a fixed double-buffer scheme. The
-`--dlo-no-use-allgather` mode instead streams standard-loader rank-local
-weights independently and does not shard weights across ranks.
-
-For the implementation design and compatibility matrix, see
-[Distributed Layerwise Offload](../../design/feature/distributed_layerwise_offload.md).
-
-**Key features:**
-- **Weight sharding + AllGather**: each rank stores 1/dp_size of weights,
-  reconstructed per-layer via `all_gather_into_tensor`
-- **Fixed double-buffer**: exactly 2 transformer blocks on each device at any
-  time, regardless of model size
-- **DP multi-concurrency**: N concurrent requests processed in parallel
-  (AllGather only gathers weight shards, which are request-independent)
-- **mmap weight loading**: weights loaded as mmap views pointing to shared OS
-  page cache, eliminating O(dp_size × model_size) RSS during model creation
-- **Platform-agnostic**: works on NVIDIA GPU (CUDA/NCCL) and Ascend NPU
-  (CANN/HCCL) via vLLM-Omni's platform abstraction
-
-**Execution Flow (per device):**
-
-```
-Compute Stream:  [Layer N]          [Layer N+1]          [Layer N+2]
-H2D Stream:      [H2D Shard N+1]   [H2D Shard N+2]
-AllGather:       [AG N+1]          [AG N+2]
-Slot Usage:      Slot0: Layer N    Slot1: Layer N+1
-```
-
-### Usage
-
-**CLI:**
-```bash
-# 4× GPU/NPU with AllGather (recommended)
-vllm serve /path/to/model --omni \
-  --enable-distributed-layerwise-offload \
-  --data-parallel-size 4
-
-# Without DLO AllGather (each rank streams standard-loader rank-local weights)
-vllm serve /path/to/model --omni \
-  --enable-distributed-layerwise-offload \
-  --data-parallel-size 4 \
-  --dlo-no-use-allgather
-
-# With SP instead of DP (long sequences)
-vllm serve /path/to/model --omni \
-  --enable-distributed-layerwise-offload \
-  --usp 4
-```
-
-**Python API:**
-```python
-from vllm_omni import Omni
-
-m = Omni(
-    model="/path/to/model",
-    enable_distributed_layerwise_offload=True,
-    dlo_use_allgather=True,  # default
-)
-```
-
-### CLI Flags
-
-| Flag | Description | Default |
-|------|-------------|---------|
-| `--enable-distributed-layerwise-offload` | Enable DLO | `false` |
-| `--data-parallel-size N` | Number of DP ranks; DLO weight-sharding group when AllGather is enabled | `1` |
-| `--dlo-use-allgather` | Shard + AllGather (saves CPU, requires concurrent requests) | `true` |
-| `--dlo-no-use-allgather` | Use standard-loader rank-local weights; no DLO sharding or AllGather | `false` |
-
-### How model weights are loaded (mmap path)
-
-When DLO + AllGather is active, the offloader:
-
-1. **Saves non-persistent buffers** (e.g. RoPE `inv_freq`, timestep `freqs`)
-   from the normally-created transformer
-2. **Converts to meta device** via `to_empty(device="meta")`, releasing random
-   initialization weights
-3. **Loads checkpoint weights as mmap views** via `safe_open().get_tensor()`,
-   which return views into the OS page cache (shared across all ranks, 0 RSS)
-4. **Calls `post_load_weights()`** to apply model-specific dtype conversions
-   (e.g. Cosmos3's `time_embedder` → FP32)
-5. **Restores non-persistent buffers** from saved copies
-
-This approach requires **zero model-specific code changes** — no pipeline or
-transformer modifications are needed.
-
-### OffloadPlan (declarative topology metadata)
-
-Models can optionally declare an `OffloadPlan` class variable to provide
-topology metadata (block attribute names, submodules to offload) without
-any offload-specific logic:
-
-```python
-from vllm_omni.diffusion.offloader import OffloadPlan
-
-class MyPipeline(nn.Module):
-    _dit_modules = ["transformer"]
-    _offload_plan = OffloadPlan(
-        block_attrs={"transformer": ("blocks",)},
-        offload_submodules={"context_encoder": "layers"},
-    )
-```
-
-When not declared, the offloader falls back to `_layerwise_offload_blocks_attrs`
-and heuristic attribute search.  This is backward-compatible — existing models
-work without any changes.
-
-### DP Multi-concurrency
-
-When `--data-parallel-size > 1` and AllGather is enabled, the scheduler batches
-up to `dp_size` requests per denoise step.  Each DP rank processes a different
-request while AllGather synchronizes only weight shards (request-independent).
-
-**Requirements for concurrent requests:**
-- `num_inference_steps` must be specified explicitly (None is not allowed)
-- All concurrent requests must have the same `num_inference_steps` value
-- These constraints exist because AllGather is a collective that requires
-  every rank to participate at each denoise step
-
-### Limitations
-- Online quantization (FP8) is incompatible with mmap loading — falls back to
-  regular loading only when the model is outside the DLO+AllGather mmap path;
-  online quantization with DLO+AllGather is rejected. Use
-  `--dlo-no-use-allgather` or disable online quantization.
-- Tensor Parallel > 1 is rejected in the DLO+AllGather mmap path because TP-aware
-  weight-loader callbacks are bypassed. The no-AllGather path retains the
-  standard TP loader and is experimental.
-- HSDP + AllGather is rejected because it would double-shard parameters. HSDP
-  with no-AllGather is accepted at configuration level, but has limited
-  end-to-end validation.
-- `num_inference_steps=None` is not allowed in DP multi-concurrency mode
-
-**Module Discovery**
-
-The offloader discovers pipeline components in two ways:
-
-1. **Protocol-based** (preferred): If the pipeline implements
-    `SupportsComponentDiscovery`, its `_dit_modules`, `_encoder_modules`,
-    `_vae_modules`, and `_resident_modules` class variables are used
-    directly.  All attribute names support dotted paths (e.g.
-    `"pipe.transformer"`, `"bagel.time_embedder"`) for nested submodules.
-
-2. **Fallback attribute scan**: Otherwise, the offloader scans for
-    well-known attribute names:
-    - **DiT modules**: `transformer`, `transformer_2`, `dit`, `sr_dit`, `language_model`, `transformer_blocks`, `model`
-    - **Encoders**: `text_encoder`, `text_encoder_2`, `text_encoder_3`, `image_encoder`
-    - **VAE**: `vae`, `audio_vae`
-
-**Hook System**
-
-Both strategies use vLLM-Omni's hook registry system (`HookRegistry` and `ModelHook`) to register pre/post forward callbacks on modules, enabling automatic swapping without modifying model code.
-
-**Backend Architecture**
-
-```
-OffloadBackend (base class)
-├── ModelLevelOffloadBackend → uses SequentialOffloadHook (.to() swap)
-│                              (delegates to a pipeline's enable_omni_model_cpu_offload
-│                               for split models like Cosmos3)
-├── LayerWiseOffloadBackend → uses LayerwiseOffloadHook
-│                          (single-GPU, full weights on host)
-└── DistributedLayerwiseOffloadBackend → uses DistributedLayerwiseOffloadHook
-                                         (multi-GPU, sharded AllGather or rank-local mode)
-```
-
-Factory function `get_offload_backend()` selects the appropriate backend based on
-configuration.
-
-For split models, `ModelLevelOffloadBackend.enable()` detects a pipeline's
-`enable_omni_model_cpu_offload` hook and delegates to it; Cosmos3 then swaps its
-reasoner/generator components inside the model forward pass.
-
-
-## Supported Models
-
-| Architecture | Example Models | DiT Class | Model-Level Offload | Layerwise Offload | Distributed Layerwise Offload | Blocks Attrs (Layerwise specific) |
-|--------------|----------------|-----------|---------------------|-------------------|-------------------------------|-----------------------------------|
-| Flux2Pipeline | `black-forest-labs/FLUX.2-dev` | `Flux2Transformer2DModel` | ✓ | ✓ | - | `"transformer_blocks"`, `"single_transformer_blocks"` |
-| LongCatImagePipeline | `meituan-longcat/LongCat-Image` | `LongCatImageTransformer2DModel` | - | ✓ | - | `"transformer_blocks"`, `"single_transformer_blocks"` |
-| NextStep11Pipeline | `stepfun-ai/NextStep-1.1` | `NextStepModel` | - | ✓ | - | `"layers"` |
-| OvisImagePipeline | `AIDC-AI/Ovis-Image-7B` | `OvisImageTransformer2DModel` | - | ✓ | - | `"transformer"` |
-| QwenImagePipeline | `Qwen/Qwen-Image` | `QwenImageTransformer2DModel` | ✓ | ✓ | - | `"transformer_blocks"` |
-| StableDiffusionXLPipeline | `stabilityai/stable-diffusion-xl-base-1.0` | `SDXLUNet2DConditionModel` | ✓ | ✓ | - | `"down_blocks"`, `"up_blocks"` |
-| StableDiffusion3Pipeline | `stabilityai/stable-diffusion-3.5-medium` | `SD3Transformer2DModel` | - | ✓ | - | `"transformer_blocks"` |
-| Wan22I2VPipeline | `Wan-AI/Wan2.2-I2V-A14B-Diffusers` | `WanTransformer3DModel` | ✓ | ✓ | - | `"blocks"` |
-| Wan22Pipeline | `Wan-AI/Wan2.2-T2V-A14B-Diffusers` | `WanTransformer3DModel` | ✓ | ✓ | - | `"blocks"` |
-| SoulXSingerPipeline / SoulXSingerSVCPipeline | `Soul-AILab/SoulX-Singer` | `DiffLlama` (`cfm_decoder.model.diff_estimator`) | ✓ | ✓ | - | `"layers"` |
-| BagelPipeline | `ByteDance-Seed/BAGEL-7B-MoT` | `Qwen2MoTModel` | - | ✓ | - | `"layers"`, `"customized modules"` |
-| Cosmos3OmniDiffusersPipeline | `nvidia/Cosmos3-Nano`, `nvidia/Cosmos3-Super` | `Cosmos3VFMTransformer`, `Cosmos3LanguageModel` | ✓ | ✓ | ✓ | `"layers"`, `"gen_layers"` |
-
-**Notes:**
-- Model-Level Offloading is expected to be supported by all common diffusion models (DiT and encoders) naturally
-- Layerwise Offloading requires DiT class to define `_layerwise_offload_blocks_attrs` pointing to transformer blocks
-- Distributed Layerwise Offloading reuses the layerwise block topology, but the
-  model and parallelism combination still needs validation. AllGather mode
-  additionally requires a compatible checkpoint/mmap layout; no-AllGather mode
-  retains the standard loader's rank-local tensors. See the
-  [DLO design](../../design/feature/distributed_layerwise_offload.md) and
-  [Cosmos3 DistOffload recipe](https://github.com/vllm-project/vllm-omni/blob/main/recipes/cosmos3/Cosmos3-DistOffload.md)
-  for usage examples.
+See [Distributed Layerwise Offloading](offloader/distributed_layerwise_offload.md).

docs/user_guide/diffusion/lora.md

@@ -1,14 +1,18 @@
 # LoRA (Low-Rank Adaptation) Guide
 
-LoRA (Low-Rank Adaptation) enables fine-tuning diffusion models by adding trainable low-rank matrices to existing model weights. vLLM-Omni currently supports PEFT-style LoRA adapters, allowing you to customize model behavior without modifying the base model weights.
+LoRA (Low-Rank Adaptation) enables fine-tuning diffusion models by adding trainable low-rank matrices to existing model weights. vLLM-Omni supports two LoRA backends: **PEFT** for PEFT-style adapters, and **Distill** usually for few-steps inference. **PEFT** backend allows you to customize model behavior without modifying the base model weights. **Distill** backend fuses LoRA weights into the base model at initialization.
 
 ## Overview
 
-LoRA adapters are lightweight, model-specific fine-tuning weights that can be dynamically loaded and applied to diffusion models. vLLM-Omni uses a unified LoRA handling mechanism similar to vLLM with LRU cache management.
+LoRA adapters are lightweight, model-specific fine-tuning weights that can be applied to diffusion models in two ways:
+
+- **PEFT backend** (`--lora-backend peft`, default): Loads a PEFT-format adapter folder via `DiffusionLoRAManager`. Adapters are cached (LRU) and activated per request via `LoRARequest`. It uses a unified LoRA handling mechanisms similar to vLLM with LRU cache management.
+- **Distill backend** (`--lora-backend distill`): Calls `pipeline.load_lora_weights` once at initialization to fuse one or more concrete checkpoint files directly into the base weights. Typically used for distilled few-step LoRAs (e.g. Lightning, LightX2V).
 
 ## LoRA Adapter Format
 
-LoRA adapters must be in **PEFT (Parameter-Efficient Fine-Tuning)** format. A typical LoRA adapter directory structure:
+### PEFT (Parameter-Efficient Fine-Tuning) format (default)
+A typical PEFT-format LoRA adapter directory structure:
 
 ```
 lora_adapter/
@@ -25,19 +29,20 @@ The `adapter_config.json` file contains metadata about the LoRA adapter, includi
 
 ### Offline Inference
 
-#### Pre-loaded LoRA
+#### PEFT backend: pre-loaded LoRA
 
-Load a LoRA adapter at initialization. This adapter is pre-loaded into the cache and can be activated by requests:
+Load a PEFT-format LoRA adapter at initialization. The adapter is pre-loaded into the cache and can be activated per request:
 
 ```python
 from vllm_omni import Omni
 from vllm_omni.lora.request import LoRARequest
 
-lora_path="/path/to/lora_adapter"
+lora_path = "/path/to/lora_adapter"
 
 omni = Omni(
     model="stabilityai/stable-diffusion-3.5-medium",
-    lora_path=lora_path
+    lora_path=lora_path,
+    lora_backend="peft",  # default, can be omitted
 )
 
 lora_request = LoRARequest(
@@ -53,6 +58,98 @@ outputs = omni.generate(
 )
 ```
 
+#### Distill backend: fuse distilled LoRA at init
+
+For distilled few-step LoRAs, pass `lora_backend="distill"` together with one or more concrete `.safetensors` files. The weights are fused into the base model once at init; subsequent `generate()` calls do not need a `LoRARequest`.
+
+##### Supported pipelines
+
+For Qwen-Image and Wan pipelines, the distill backend calls `pipeline.load_lora_weights(...)` during worker initialization.
+
+| Pipeline | Supported distilled LoRA repo | Notes |
+|----------|--------------------------------|-------|
+| `QwenImagePipeline` | `lightx2v/Qwen-Image-2512-Lightning` | Used with Qwen-Image-2512 Lightning-style few-step inference. |
+| `Wan22Pipeline` | `lightx2v/Wan2.1-Distill-Loras`, `lightx2v/Wan2.2-Distill-Loras` | Wan2.1 uses one LoRA file. For dual-transformer Wan2.2 MoE, pass high-noise then low-noise LoRA files. |
+| `Wan22I2VPipeline` | `lightx2v/Wan2.2-Distill-Loras` | For dual-transformer Wan2.2 MoE, pass high-noise then low-noise LoRA files. |
+
+Other diffusion pipelines are not currently listed as supporting distilled LoRA. Use the PEFT backend for request-time adapters, or bake converted weights into a local Diffusers directory before serving.
+
+Single-file example (Qwen-Image-Lightning):
+
+```python
+from vllm_omni import Omni
+
+omni = Omni(
+    model="Qwen/Qwen-Image-2512",
+    lora_path="/path/to/Qwen-Image-2512-Lightning.safetensors",
+    lora_backend="distill",
+)
+
+outputs = omni.generate(prompt="A piece of cheesecake")
+```
+
+Multi-file example (Wan2.2 MoE, high + low noise):
+
+```python
+from vllm_omni import Omni
+
+omni = Omni(
+    model="Wan-AI/Wan2.2-T2V-A14B-Diffusers",
+    lora_path=[
+        "/path/to/wan2.2_high_noise_lora.safetensors",   # -> transformer
+        "/path/to/wan2.2_low_noise_lora.safetensors",    # -> transformer_2
+    ],
+    lora_backend="distill",
+)
+```
+
+The CLI examples under `examples/offline_inference/` accept the same flags, e.g.:
+
+```bash
+python examples/offline_inference/text_to_video/text_to_video.py \
+  --model Wan-AI/Wan2.2-T2V-A14B-Diffusers \
+  --lora-backend distill \
+  --lora-path /path/to/high.safetensors /path/to/low.safetensors \
+  --prompt "A cat playing with yarn"
+```
+
+##### Online serving
+
+Distilled LoRAs can also be fused when an online diffusion server starts. The
+adapter remains active for the lifetime of that server, so requests do not pass
+a per-request `lora` field.
+
+```bash
+vllm serve Qwen/Qwen-Image-2512 \
+  --omni \
+  --port 8091 \
+  --lora-backend distill \
+  --lora-path /path/to/Qwen-Image-2512-Lightning-4steps.safetensors
+```
+
+Then send a normal generation request with the sampling settings expected by
+the distilled checkpoint:
+
+```bash
+curl -s http://localhost:8091/v1/chat/completions \
+  -H "Content-Type: application/json" \
+  -d '{
+    "messages": [
+      {"role": "user", "content": "A Chinese female college student, around 20 years old, with a very short haircut that conveys a gentle, artistic vibe. Her hair naturally falls to partially cover her cheeks, projecting a tomboyish yet charming demeanor. She has cool-toned fair skin and delicate features, with a slightly shy yet subtly confident expression—her mouth crooked in a playful, youthful smirk. She wears an off-shoulder top, revealing one shoulder, with a well-proportioned figure. The image is framed as a close-up selfie: she dominates the foreground, while the background clearly shows her dormitory—a neatly made bed with white linens on the top bunk, a tidy study desk with organized stationery, and wooden cabinets and drawers. The photo is captured on a smartphone under soft, even ambient lighting, with natural tones, high clarity, and a bright, lively atmosphere full of youthful, everyday energy."}
+    ],
+    "extra_body": {
+      "height": 1024,
+      "width": 1024,
+      "num_inference_steps": 4,
+      "true_cfg_scale": 1.0,
+      "seed": 42
+    }
+  }' | jq -r '.choices[0].message.content[0].image_url.url' | cut -d',' -f2 | base64 -d > qwen_image_distill_lora_output.png
+```
+
+For Wan2.2 MoE serving, pass the high-noise checkpoint first and the low-noise
+checkpoint second to `--lora-path`.
+
 !!! note "Server-side Path Requirement"
     The LoRA adapter path (`local_path`) must be readable on the **server** machine. If your client and server are on different machines, ensure the LoRA adapter is accessible via a shared mount or copied to the server.
 
@@ -141,6 +238,7 @@ Notes:
 
 - This route avoids runtime LoRA loading changes in vLLM-Omni when you choose to bake converted weights into a local Diffusers directory.
 - Output quality and speed depend on the replacement checkpoints and sampling params you choose.
+- If you only need to fuse distilled LoRAs into a Wan2.2 checkpoint at load time (without the full LightX2V convert + assemble pipeline), you can instead pass them directly via `--lora-backend distill --lora-path <high>.safetensors <low>.safetensors`. See the [Distill backend](#distill-backend-fuse-distilled-lora-at-init) section above.
 
 
 ## See Also

docs/user_guide/diffusion/offloader/cpu_offload.md

@@ -0,0 +1,81 @@
+# Model-Level Offloading
+
+Model-level, or sequential, offloading keeps only the pipeline component group
+currently executing on the accelerator. It is the simplest offload strategy
+and is selected with `--enable-cpu-offload`.
+
+## How it works
+
+Pre-forward hooks enforce mutual exclusion between DiT and encoder modules:
+
+- before an encoder runs, the DiT moves to CPU;
+- before a DiT runs, encoders and other DiTs move to CPU; and
+- VAE modules remain on the accelerator.
+
+Pinned host memory reduces transfer overhead. Transfers occur at phase
+boundaries, so cold-start and encoder-to-denoiser transitions become slower.
+
+## Usage
+
+```python
+from vllm_omni import Omni
+
+omni = Omni(
+    model="Wan-AI/Wan2.2-T2V-A14B-Diffusers",
+    enable_cpu_offload=True,
+)
+```
+
+```bash
+vllm serve Wan-AI/Wan2.2-T2V-A14B-Diffusers \
+  --omni --enable-cpu-offload
+```
+
+## Model integration
+
+Pipelines should implement `SupportsComponentDiscovery`:
+
+```python
+from typing import ClassVar
+
+from vllm_omni.diffusion.models.interface import SupportsComponentDiscovery
+
+
+class MyPipeline(nn.Module, SupportsComponentDiscovery):
+    _dit_modules: ClassVar[list[str]] = ["transformer"]
+    _encoder_modules: ClassVar[list[str]] = ["text_encoder", "vision_model"]
+    _vae_modules: ClassVar[list[str]] = ["vae"]
+    _resident_modules: ClassVar[list[str]] = []
+```
+
+All entries may be dotted paths. DiT and encoder lists are both required for
+mutual exclusion. VAE modules are pinned but not swapped; resident modules are
+small modules that must stay on the accelerator for layerwise paths.
+
+## Split-model components
+
+Some models, such as Cosmos3, split one transformer into mutually exclusive
+components that run in different phases. The pipeline exposes
+`enable_omni_model_cpu_offload`, and the backend delegates to the model-local
+contexts:
+
+```python
+class Cosmos3VFMTransformer(nn.Module):
+    def forward(self, ...):
+        with self._offload_context("reasoner"):
+            ...
+        with self._offload_context("generator"):
+            ...
+```
+
+This preserves the same invariant—exactly one component is device resident—
+while reusing sequential `.to()` movers.
+
+## Limitations
+
+- Single device only.
+- Higher cold-start latency.
+- Transfers between encoder and denoising phases add latency.
+
+See the [model-level design](../../../design/feature/offloader/cpu_offload.md)
+for lifecycle and extension invariants.

docs/user_guide/diffusion/offloader/distributed_layerwise_offload.md

@@ -0,0 +1,156 @@
+# Distributed Layerwise Offloading
+
+Distributed layerwise offloading (DLO) extends block streaming to multi-device
+deployments. With AllGather enabled, each rank stores roughly `1 / dp_size` of
+the host weights and reconstructs each layer at runtime. Without AllGather,
+each rank streams a complete block independently. Compatible TP1 deployments
+can share checkpoint-backed host pages among processes on the same node;
+otherwise DLO streams the ordinary loader's rank-local tensors.
+
+See the [DLO feature design](../../../design/feature/offloader/distributed_layerwise_offload.md)
+for the implementation contract and compatibility matrix.
+
+## Execution model
+
+DLO overlaps three operations with a fixed two-block device buffer:
+
+```text
+Compute stream:  [Layer N]          [Layer N+1]        [Layer N+2]
+H2D stream:      [H2D shard N+1]    [H2D shard N+2]
+AllGather:       [AG N+1]           [AG N+2]
+Slots:           slot 0: Layer N    slot 1: Layer N+1
+```
+
+AllGather communicates only request-independent weight shards, so data-
+parallel ranks may process different requests concurrently.
+
+## Usage
+
+```bash
+# Four ranks with sharded host weights and AllGather
+vllm serve /path/to/model --omni \
+  --enable-distributed-layerwise-offload \
+  --data-parallel-size 4
+
+# Standard-loader rank-local weights, without DLO AllGather
+vllm serve /path/to/model --omni \
+  --enable-distributed-layerwise-offload \
+  --data-parallel-size 4 \
+  --dlo-no-use-allgather
+
+# Sequence parallel deployment
+vllm serve /path/to/model --omni \
+  --enable-distributed-layerwise-offload \
+  --usp 4
+```
+
+```python
+from vllm_omni import Omni
+
+omni = Omni(
+    model="/path/to/model",
+    enable_distributed_layerwise_offload=True,
+    dlo_use_allgather=True,
+)
+```
+
+## Flags
+
+| Flag | Meaning | Default |
+| --- | --- | --- |
+| `--enable-distributed-layerwise-offload` | Enable DLO | `false` |
+| `--data-parallel-size N` | DP ranks and AllGather weight-sharding group | `1` |
+| `--dlo-use-allgather` | Shard host weights and reconstruct with AllGather | `true` |
+| `--dlo-no-use-allgather` | Stream complete rank-local blocks without a DLO weight collective | `false` |
+| `--dlo-resident-layers N` | Keep N leading main-DiT blocks on device; requires no-AllGather and model-declared resident paths | `0` |
+
+## Host-weight loading
+
+The diffusion loader chooses host storage before DLO is enabled. It first
+attempts to build a complete, validated direct-checkpoint mmap plan. If names,
+coverage, shape, dtype, topology, or loader-callback compatibility cannot be
+proven, it runs the ordinary model loader instead. DLO consumes that result and
+does not make a second checkpoint-compatibility decision.
+
+The shared-mmap optimization in this phase is supported only with TP1. TP
+greater than one falls back before model mutation to ordinary TP-aware loading.
+DLO may still consume those TP-local tensors, but this is a compatibility path:
+it does not share checkpoint-backed runtime weights across DP replicas and
+provides no shared-mmap host-memory guarantee.
+
+The mmap plan skips only dedicated DiT weight sources. Other component sources,
+such as a text encoder loaded through the shared diffusion loader, continue to
+use their ordinary component loader. A checkpoint source that mixes DiT and
+non-DiT weights falls back completely rather than leaving an unplanned
+component uninitialized.
+
+With direct checkpoint mmap, the loader:
+
+1. saves non-persistent buffers such as RoPE frequencies;
+2. moves the normally created transformer to the meta device;
+3. loads checkpoint tensors as mmap views backed by the shared OS page cache;
+4. applies any loader-owned bounded layout adapters while packing blocks;
+5. restores saved non-persistent buffers; and
+6. preserves `post_load_weights()` and `validate_loaded_weights()` lifecycle
+   hooks.
+
+For AllGather with a group larger than one, each process copies only its
+persistent shard and then releases the source mapping. For no-AllGather, each
+process keeps the mapping open and packs complete blocks through two bounded
+pinned staging slots. Processes mapping the same files on one node share the
+immutable pages; no-AllGather still performs a complete-block H2D copy in each
+process.
+
+When the effective DLO group size is one, `dlo_use_allgather=True` does not
+perform a collective and uses the same rank-local transfer behavior.
+
+## Declarative topology
+
+Models may declare an `OffloadPlan` instead of embedding offload logic:
+
+```python
+from vllm_omni.diffusion.offloader import OffloadPlan
+
+
+class MyPipeline(nn.Module):
+    _dit_modules = ["transformer"]
+    _offload_plan = OffloadPlan(
+        block_attrs={"transformer": ("blocks",)},
+        offload_submodules={"context_encoder": "layers"},
+    )
+```
+
+When no plan exists, discovery falls back to
+`_layerwise_offload_blocks_attrs` and then heuristic attribute lookup.
+
+## Data-parallel concurrency
+
+With `data_parallel_size > 1` and AllGather enabled, the scheduler can process
+up to `dp_size` requests per denoising step. Every concurrent request must set
+the same explicit `num_inference_steps`; `None` is rejected because every rank
+must enter each collective.
+
+## Limitations
+
+- Direct checkpoint mmap currently requires TP1. TP greater than one is
+  outside the Phase A shared-mmap support scope and falls back before model
+  mutation to the ordinary TP-aware loader. DLO can stream that runtime layout,
+  but it provides no shared-mmap host-memory benefit or guarantee.
+- HSDP plus AllGather is rejected to avoid double sharding. HSDP without
+  AllGather has limited end-to-end validation.
+- Per-tensor online FP8 linears use the ordinary loader and can run with either
+  DLO transfer path. With AllGather, every rank temporarily materializes the
+  complete FP8 model in host memory before DLO retains only its shard. Other
+  online quantization methods require no-AllGather until their runtime layouts
+  are validated.
+- Resident leading layers require `--dlo-no-use-allgather` and a model
+  `OffloadPlan` that declares eligible `resident_dit_paths`.
+- DP concurrency requires an explicit, identical inference-step count.
+
+Sharing transformed TP or quantized runtime layouts through a normalized mmap
+cache is a follow-up design in
+[RFC #6195](https://github.com/vllm-project/vllm-omni/issues/6195), not part of
+the direct-checkpoint path.
+
+See the [Cosmos3 DistOffload recipe](https://github.com/vllm-project/vllm-omni/blob/main/recipes/cosmos3/Cosmos3-DistOffload.md)
+for an end-to-end example.

docs/user_guide/diffusion/offloader/layerwise_offload.md

@@ -0,0 +1,65 @@
+# Layerwise Offloading
+
+Layerwise, or blockwise, offloading keeps one transformer block on the
+accelerator and prefetches the next block while the current block computes.
+It is best suited to compute-heavy video DiTs whose block execution can hide
+host-to-device transfers.
+
+## Execution flow
+
+Each block has a pre-forward and post-forward hook. Parameters are consolidated
+in pinned host tensors and rematerialized for execution on a dedicated copy
+stream.
+
+| Block | Pre-forward hook | Forward | Post-forward hook |
+| --- | --- | --- | --- |
+| block 0 | Prefetch block 1 | Compute block 0 | Free block 0 |
+| block 1 | Prefetch block 2 | Compute block 1 | Free block 1 |
+| ... | ... | ... | ... |
+| last block | Prefetch block 0 | Compute last block | Free last block |
+
+Encoders, VAE modules, and non-block DiT modules such as embeddings and norms
+remain device resident.
+
+## Usage
+
+```python
+from vllm_omni import Omni
+
+omni = Omni(
+    model="Wan-AI/Wan2.2-T2V-A14B-Diffusers",
+    enable_layerwise_offload=True,
+)
+```
+
+```bash
+vllm serve Wan-AI/Wan2.2-T2V-A14B-Diffusers \
+  --omni --enable-layerwise-offload
+```
+
+## Model integration
+
+Transformer classes declare containers of executable blocks:
+
+```python
+class WanTransformer3DModel(nn.Module):
+    _layerwise_offload_blocks_attrs = ["blocks"]
+
+
+class Flux2Transformer2DModel(nn.Module):
+    _layerwise_offload_blocks_attrs = [
+        "transformer_blocks",
+        "single_transformer_blocks",
+    ]
+```
+
+See the [layerwise design](../../../design/feature/offloader/layerwise_offload.md) for
+the discovery and hook invariants. `OffloadPlan` is a separate declarative
+topology path for distributed layerwise offload.
+
+## Limitations
+
+- Single device only.
+- Setup consolidates and pins block parameters, increasing cold-start time.
+- Performance depends on block compute time and host-to-device bandwidth;
+  lightweight blocks may not hide transfers.

docs/user_guide/diffusion/parallelism/hsdp.md

@@ -64,6 +64,41 @@ omni = Omni(
 )
 ```
 
+### Combined with Online FP8 Quantization
+
+HSDP can be combined with online FP8 quantization for diffusion models that
+support both features. HSDP shards the quantized weights across GPUs to reduce
+the per-GPU weight memory footprint.
+
+```python
+from vllm_omni import Omni
+from vllm_omni.diffusion.data import DiffusionParallelConfig
+
+omni = Omni(
+    model="<your-model>",
+    quantization="fp8",
+    parallel_config=DiffusionParallelConfig(
+        use_hsdp=True,
+        hsdp_shard_size=2,
+    ),
+)
+```
+
+For online serving:
+
+```bash
+vllm serve <your-model> \
+    --omni \
+    --quantization fp8 \
+    --use-hsdp \
+    --hsdp-shard-size 2
+```
+
+The model must support both HSDP and online FP8 quantization. HSDP remains
+incompatible with Tensor Parallelism. See the
+[FP8 Quantization Guide](../../quantization/fp8.md) for supported models and
+hardware.
+
 ---
 
 ## Example Script
@@ -117,7 +152,8 @@ In `DiffusionParallelConfig`:
 **Constraints:**
 
 - `hsdp_replicate_size × hsdp_shard_size == world_size`
-- HSDP cannot be used with Tensor Parallelism (`tensor_parallel_size` must be 1)
+- HSDP cannot be used with Tensor Parallelism, ordinary Data Parallelism,
+  Pipeline Parallelism, or Expert Parallelism.
 
 ---
 
@@ -147,3 +183,4 @@ For detailed instructions on adding HSDP support to new models, see the [HSDP Co
 1. ✅ **Enable HSDP** - Set `use_hsdp=True` and `hsdp_shard_size` to reduce per-GPU memory for large models
 2. ✅ **Combine with SP** - Use together with `ulysses_degree` for video models requiring both memory reduction and sequence parallelism
 3. ⚠️ **Incompatible with TP** - `tensor_parallel_size` must be 1 when HSDP is enabled
+4. ✅ **Combine with online FP8** - Use `quantization="fp8"` or `--quantization fp8` to reduce the sharded weight memory footprint

docs/user_guide/diffusion/parallelism/sequence_parallel.md

@@ -164,10 +164,12 @@ In `DiffusionParallelConfig`:
 | `ulysses_degree` | int | 1 | Number of GPUs for Ulysses-SP. Uses all-to-all communication. |
 | `ring_degree` | int | 1 | Number of GPUs for Ring-Attention. Uses P2P ring communication. |
 | `ulysses_mode` | str | `"default"` | Ulysses attention mode. Set to `"advanced_uaa"` to handle arbitrary sequence lengths and head counts without padding. |
+| `mask_sp_padding` | bool | `False` | When the sequence length is not divisible by the SP size, tokens are auto-padded with zeros. Set to `True` to mask those padding tokens (strict, but uses the slower varlen attention path); the default `False` leaves them unmasked, keeping the fast path with negligible numerical impact. |
 
 **Notes:**
 - Total sequence parallel size equals to `ulysses_degree × ring_degree`
 - Degrees must evenly divide the sequence length for optimal performance (or use `ulysses_mode="advanced_uaa"` for Ulysses-SP)
+- `mask_sp_padding` is an experimental feature, currently only supported by `Wan2.2`, `Wan2.2 Vace`, `Qwen-Image`, `Flux 2`, and `HunyuanVideo 1.5`
 
 
 ## Best Practices

docs/user_guide/diffusion/startup_and_loading.md

@@ -0,0 +1,60 @@
+# Diffusion Startup and Loading
+
+Large diffusion models can take several minutes to load at startup. vLLM-Omni
+loads safetensors shards in parallel to reduce this initialization time.
+
+Multi-thread weight loading is enabled by default with four threads. No
+configuration is needed for the default behavior.
+
+## Configuration
+
+| Parameter | CLI flag | Default | Description |
+| --- | --- | --- | --- |
+| `enable_multithread_weight_load` | `--disable-multithread-weight-load` | `True` | Pass the flag to disable multi-thread loading |
+| `num_weight_load_threads` | `--num-weight-load-threads` | `4` | Number of parallel weight-loading threads |
+
+!!! tip
+
+    The default balances startup speed and disk I/O contention. Fast NVMe
+    storage may benefit from more threads, while network storage or hard disks
+    may not.
+
+## Online Serving
+
+```bash
+# Default: multi-thread loading with four threads
+vllm serve Qwen/Qwen-Image --omni --port 8091
+
+# Increase the thread count
+vllm serve Wan-AI/Wan2.2-I2V-A14B-Diffusers --omni \
+  --num-weight-load-threads 8
+
+# Disable multi-thread loading
+vllm serve Qwen/Qwen-Image --omni --disable-multithread-weight-load
+```
+
+## Offline Inference
+
+```python
+from vllm_omni import Omni
+
+# Default: multi-thread loading with four threads
+omni = Omni(model="Qwen/Qwen-Image")
+
+# Increase the thread count
+omni = Omni(
+    model="Wan-AI/Wan2.2-I2V-A14B-Diffusers",
+    num_weight_load_threads=8,
+)
+```
+
+## Reference Benchmarks
+
+The following measurements were collected on NVIDIA H800 hardware. Treat them
+as reference results rather than a guarantee for other storage or hardware
+configurations.
+
+| Model | Sequential loading | Multi-thread loading | Speedup |
+| --- | ---: | ---: | ---: |
+| **Qwen/Qwen-Image** (53.7 GiB) | 168 s | 27 s | **6.2x** |
+| **Wan-AI/Wan2.2-I2V-A14B-Diffusers** (64.5 GiB) | 283 s | 56 s | **5.1x** |

docs/user_guide/diffusion_features.md

@@ -49,7 +49,7 @@ Parallelism methods distribute computation across GPUs without quality loss (mat
 
 | Method | Description | Best For |
 |--------|-------------|----------|
-| **[Multi-Thread Weight Loading](#multi-thread-weight-loading)** | Loads safetensors shards in parallel using a thread pool | All diffusion models; reduces startup from minutes to seconds |
+| **[Multi-Thread Weight Loading](diffusion/startup_and_loading.md)** | Loads safetensors shards in parallel using a thread pool | All diffusion models; reduces startup from minutes to seconds |
 
 **Note:** Some acceleration methods can be combined together for optimized performance. See [Feature Compatibility Table](#feature-compatibility) and [Feature Compatibility Tutorial](feature_compatibility.md) for detailed configuration examples.
 
@@ -100,11 +100,16 @@ The following tables show which models support each feature:
 
 - **🔀SP (Ulysses & Ring)**: Includes both Ulysses-SP and Ring-Attention methods
 - ✅ = Fully supported
+- ✅* = Supported with the constraint listed below the table
 - ❌ = Not supported
+- ❓ = Not verified; not recommended
 
 > Notes:
 
-> 1. CPU Offload has two methods: Module-wise (default for models with DiT + text encoder) and Layerwise. The tables below show **Layerwise support** only. Split models like Cosmos3 (no separate text encoder) swap their reasoner/generator components for module-wise offload; see the [CPU Offload Guide](diffusion/cpu_offload.md).
+> 1. CPU Offload has three strategies: model-level, layerwise, and distributed
+>    layerwise. The tables below show **layerwise support** only. Split models
+>    like Cosmos3 swap their reasoner/generator components for model-level
+>    offload; see the [CPU Offload Guide](diffusion/cpu_offload.md).
 > 2. The **💾Quantization** column is collapsed for readability. See [Quantization](quantization/overview.md) for per-method and per-model support details.
 
 ### ImageGen
@@ -119,7 +124,7 @@ The following tables show which models support each feature:
 | **FLUX.2-dev**           |     ✅     |     ✅      |           ✅           |       ✅        |         ✅         |          ❌          |   ✅    |             ✅             |          ✅           |       ❌        |        ❌         |
 | **GLM-Image**            |     ❌     |     ❌      |           ❌           |       ✅        |         ✅         |          ❌          |   ✅    |             ❌             |          ❌           |       ❌        |        ❌         |
 | **Hidream-I1-Full**        |     ❌     |     ❌      |           ❌           |       ❌        |         ✅         |          ❌          |   ❌    |             ❌             |          ❌           |       ❌        |        ❌         |
-| **HunyuanImage3**        |     ❌     |     ✅      |           ❌           |       ❌        |         ✅         |          ❌          |   ❌    |             ❌             |          ❌           |       ✅        |        ❌         |
+| **HunyuanImage3**        |     ❌     |     ✅      |           ❌           |       ❌        |         ✅         |          ❌          |   ❌    |             ❌             |          ❌           |       ✅        |        ✅*        |
 | **Krea 2**               |     ❌     |     ✅      |           ❌           |       ❌        |         ❌         |          ❌          |   ✅    |             ✅             |      ✅ (decode)      |       ❌        |        ❌         |
 | **LongCat-Image**        |     ✅     |     ✅      |           ✅           |       ✅        |         ✅         |          ❌          |   ❌    |             ✅             |          ❌           |       ❌        |        ❌         |
 | **LongCat-Image-Edit**   |     ✅     |     ✅      |           ✅           |       ✅        |         ✅         |          ❌          |   ❌    |             ✅             |          ❌           |       ❌        |        ❌         |
@@ -145,6 +150,7 @@ The following tables show which models support each feature:
 > 2. `Tongyi-MAI/Z-Image-Turbo` and `SII-GAIR/daVinci-MagiHuman-Base-1080p` are distilled models with minimal NFEs; CFG-Parallel is not necessary.
 > 3. Cosmos3 T2I uses `Cosmos3OmniDiffusersPipeline` with `modalities=["image"]`. Model-level CPU offload swaps the nested UND reasoner and GEN generator pathways; layerwise offload remains available for blockwise GEN/UND offload.
 > 4. Krea 2 currently supports single-GPU inference plus LoRA, Cache-DiT, HSDP, CPU/layerwise offload, and VAE-patch-parallel (decode). TP/SP/CFG-Parallel are not yet wired. The few-step distilled (Turbo) checkpoint uses `is_distilled=true` (fixed timestep shift `mu=1.15`); generate at 2048x2048 by default with `num_inference_steps≈8` and `guidance_scale=0`. The Raw checkpoint uses 1024x1024, `num_inference_steps=28`, and `guidance_scale=4.5`.
+> 5. HunyuanImage3 supports step execution. Multi-request step execution requires `TORCH_SDPA`; see [Diffusion Execution Modes](diffusion/execution_modes.md#step-execution).
 
 ### VideoGen
 
@@ -155,16 +161,29 @@ The following tables show which models support each feature:
 | **Wan2.1-VACE**              |     ❌     |     ✅      |           ✅           |       ✅        |         ✅         |         ❌         |   ✅    |             ✅             |      ✅ (decode)      |       ❌        |        ❌         |
 | **LTX-2**                    |     ❌     |     ✅      |           ✅           |       ✅        |         ✅         |         ❌         |   ✅    |             ✅             |      ✅ (decode)      |       ❌        |        ❌         |
 | **LTX-2.3**                  |     ❌     |     ✅      |           ✅           |       ✅        |         ✅         |         ❌         |   ✅    |             ✅             |      ✅ (decode)      |       ❌        |        ❌         |
-| **Helios**                   |     ❌     |     ✅      |           ✅           |       ✅        |         ✅         |         ❌         |   ✅    |             ✅             |          ❌           |       ❌        |        ❌         |
+| **LTX-2.5**                  |     ❌     | ❓ (one-stage) | ❓ (Ulysses only) | ✅ (Full only) |         ❓         |         ❌         |   ✅    |             ✅             |      ✅ (decode)      |    ❓ (FP8)     |        ❌         |
+| **Helios**                   |     ❌     |     ✅      |           ✅           |       ✅        |         ✅         |         ❌         |   ✅    |             ✅             |          ❌           |       ❌        |        ✅*        |
 | **HunyuanVideo-1.5 T2V I2V** |     ❌     |     ✅      |           ✅           |       ✅        |         ✅         |         ❌         |   ✅    |             ✅             |  ✅ (encode/decode)   |       ✅        |        ❌         |
 | **DreamID-Omni**             |     ❌     |     ❌      |           ❌           |       ✅        |         ❌         |         ❌         |   ✅    |             ✅             |          ❌           |       ❌        |        ❌         |
 | **Cosmos3**                  |     ❌     |     ✅      |           ✅           |       ✅        |         ✅         |         ❌         |   ✅    |             ✅             |  ✅ (encode/decode)   |       ✅        |        ❌         |
+| **LongCat-Video-Avatar-1.5** |     ❌     |     ❌      |           ❌           |       ❌        |         ❌         |         ❌         |   ❌    |             ❌             |          ❌           |       ❌        |        ❌         |
 | **MiniMax-H3**               | ✅ (FL2VA) |     ✅      |           ✅           |       ❌        |       ✅ (DiT/TE)  |         ❌         |   ✅    |             ✅             |       ✅ (tile)       |      ✅ (DiT)      |        ❌         |
+| **SANA-WM**                  |     ❌     |     ❌      |          ❌<sup>5</sup> |       ✅        |         ✅         |         ❌         |   ❌    |             ❌             |          ❌           |       ❌        |        ❌         |
+
+> Notes:
+> 5. SANA-WM cannot support sequence parallelism: its bidirectional gated delta
+>    recurrence carries state across frames, so a rank cannot denoise a slice of
+>    the token sequence in isolation. Doing so would need a distributed scan or
+>    an all-gather before every GDN block. The remaining ❌ columns are simply
+>    unvalidated on this model, not known-broken.
+
+> **Step execution note:** Helios supports single-request step execution only;
+> use `max_num_seqs=1`.
 
 **Frame Interpolation Support**
 
 - **Supported**: Wan2.2 text-to-video, image-to-video, and TI2V pipelines
-- **Not supported**: Wan2.1-VACE, LTX-2, LTX-2.3, Helios, HunyuanVideo-1.5, DreamID-Omni
+- **Not supported**: Wan2.1-VACE, LTX-2, LTX-2.3, LTX-2.5, Helios, HunyuanVideo-1.5, DreamID-Omni, SANA-WM
 
 ### AudioGen
 
@@ -189,7 +208,7 @@ The following tables show which models support each feature:
 | **🔀Ring-Attn** | ✅ | ✅ | ✅ | | | | | | | | | | | |
 | **🔀CFG-Parallel** | ✅ | ✅ | ✅ | ✅ | | | | | | | | | | |
 | **🔀Tensor Parallel** | ✅ | ✅ | ✅ | ✅ | ✅ | | | | | | | | | |
-| **🔀HSDP** | ❓ | ❓ | ❓ | ❓ | ❓ | ❌ | | | | | | | | |
+| **🔀HSDP** | ❓ | ❓ | ✅ | ❓ | ✅ | ❌ | | | | | | | | |
 | **🔀Expert Parallel** | ❓ | ❓ | ❓ | ❓ | ❓ | ❓ | ❓ | | | | | | | |
 | **💾CPU Offloading (Layerwise)** | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | | | | | | |
 | **💾CPU Offloading (Module-wise)** | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❓ | ❓ | ❌ | | | | | |
@@ -200,97 +219,43 @@ The following tables show which models support each feature:
 
 !!! info
 
-    1. Tensor Parallel and HSDP are not compatible.
+    1. HSDP can be combined with Ulysses-SP or CFG-Parallel. Tensor Parallel
+       and HSDP are not compatible; other HSDP combinations in the table
+       remain unverified.
     2. TeaCache and Cache-DiT are not compatible.
     3. CPU Offloading (Layerwise) and CPU Offloading (Module-wise) are not compatible.
-    4. CPU Offloading (Layerwise) supports single-card for now.
-    5. Using FP8-Quant as an example of qunatization methods.
+    4. The CPU Offloading (Layerwise) row describes local layerwise offload.
+       Multi-device Distributed Layerwise Offload has a separate topology and
+       compatibility matrix in the [Distributed Layerwise Offloading guide](diffusion/offloader/distributed_layerwise_offload.md).
+    5. The compatibility matrix uses FP8 as the representative quantization method.
     6. Step Execution is not compatible with any diffusion cache backend. LoRA is supported, but each scheduled batch must use a single adapter (requests with different `lora_request` or `lora_scale` are kept in separate batches).
 
 
 ## Multi-Thread Weight Loading
 
-Large diffusion models can take several minutes to load weights at startup (e.g., ~3 min for Qwen-Image, ~5 min for Wan2.2 I2V 14B). Multi-thread weight loading speeds up this process by loading safetensors shards in parallel using a thread pool instead of sequentially.
-
-This optimization is **enabled by default** with 4 threads. No configuration is needed for the default behavior.
-
-### Configuration
-
-| Parameter | CLI Flag | Default | Description |
-|-----------|----------|---------|-------------|
-| `enable_multithread_weight_load` | `--disable-multithread-weight-load` | `True` (enabled) | Pass the flag to disable multi-thread loading |
-| `num_weight_load_threads` | `--num-weight-load-threads` | `4` | Number of threads for parallel weight loading |
-
-!!! tip
-    The default of 4 threads balances speed and disk I/O contention. On fast NVMe storage you may benefit from more threads (e.g., 8). On HDD or network storage, the default of 4 avoids saturating I/O bandwidth.
-
-### Online Serving
-
-```bash
-# Default (multi-thread enabled, 4 threads)
-vllm serve Qwen/Qwen-Image --omni --port 8091
-
-# Custom thread count
-vllm serve Wan-AI/Wan2.2-I2V-A14B-Diffusers --omni --num-weight-load-threads 8
-
-# Disable multi-thread loading
-vllm serve Qwen/Qwen-Image --omni --disable-multithread-weight-load
-```
-
-### Offline Inference
-
-```python
-from vllm_omni import Omni
-
-# Default (multi-thread enabled, 4 threads)
-omni = Omni(model="Qwen/Qwen-Image")
-
-# Custom thread count
-omni = Omni(
-    model="Wan-AI/Wan2.2-I2V-A14B-Diffusers",
-    num_weight_load_threads=8,
-)
-```
-
-### Benchmarks
-
-Measured on NVIDIA H800:
-
-| Model | Before | After | Speedup |
-|-------|--------|-------|---------|
-| **Qwen/Qwen-Image** (53.7 GiB) | 168s | 27s | **6.2x** |
-| **Wan-AI/Wan2.2-I2V-A14B-Diffusers** (64.5 GiB) | 283s | 56s | **5.1x** |
+The loading guide now lives at [Diffusion Startup and
+Loading](diffusion/startup_and_loading.md). This heading remains so existing
+links to this section continue to work.
 
 ## Learn More
 
-**Cache Acceleration:**
+The Diffusion Acceleration navigation groups the remaining guides as follows:
 
-- **[TeaCache Configuration Guide](diffusion/cache_acceleration/teacache.md)** - Parameter tuning, performance tips, troubleshooting
-- **[Cache-DiT Advanced Guide](diffusion/cache_acceleration/cache_dit.md)** - DBCache, TaylorSeer, SCM techniques and optimization
+| Area | Guide |
+| --- | --- |
+| Compatibility | [Feature Compatibility](feature_compatibility.md) |
+| CPU offloading | [CPU Offloading](diffusion/cpu_offload.md) |
+| Cache acceleration | [TeaCache](diffusion/cache_acceleration/teacache.md), [Cache-DiT](diffusion/cache_acceleration/cache_dit.md) |
+| Parallelism | [Parallelism Overview](diffusion/parallelism/overview.md) |
+| Attention | [Attention Backends](diffusion/attention_backends.md) |
+| Compilation | [Regional Compilation](diffusion/regional_compilation.md) |
+| Video extension | [Frame Interpolation](diffusion/frame_interpolation.md) |
+| Startup | [Startup and Loading](diffusion/startup_and_loading.md) |
+| Adapters | [LoRA](diffusion/lora.md) |
 
-**Parallelism Methods:**
+Related cross-model and runtime features are documented separately:
 
-- **[Parallelism Overview](diffusion/parallelism/overview.md)** - Tensor Parallelism, Sequence Parallelism, CFG Parallelism, Pipeline Parallelism, HSDP, and Expert Parallelism
-
-**Memory Optimization:**
-
-- **[CPU Offload Guide](diffusion/cpu_offload.md)** - Offload model components to CPU, reduce GPU memory usage
-- **[VAE Parallelism Guide](diffusion/parallelism/vae_parallelism.md)** - Distribute VAE decode work across GPUs for high-resolution images and videos
-- **[Quantization](quantization/overview.md)** - Quantization methods for diffusion, multi-stage omni/TTS, and multi-stage diffusion models
-
-**Extensions:**
-
-- **[LoRA Inference Guide](diffusion/lora.md)** - Low-Rank Adaptation for style customization and fine-tuning
-- **[Frame Interpolation Guide](diffusion/frame_interpolation.md)** - Worker-side post-generation video frame interpolation for smoother motion
-
-**Execution Modes:**
-
-- **[Diffusion Execution Modes](diffusion/execution_modes.md)** - Configure request batching, step execution, continuous batching, and streaming output
-
-**Startup Optimization:**
-
-- **[Multi-Thread Weight Loading](#multi-thread-weight-loading)** - Speed up model startup by loading safetensors shards in parallel
-
-**Advanced Topics:**
-
-- **[Feature Compatibility](feature_compatibility.md)** - How to combine multiple features for maximum performance
+- [Quantization](quantization/overview.md) covers diffusion-only models,
+  multi-stage omni/TTS models, and multi-stage diffusion models.
+- [Execution Modes and Streaming](diffusion/execution_modes.md) covers the
+  diffusion runtime, including batching, step execution, and streaming output.

docs/user_guide/examples/offline_inference/glm_image.md

@@ -43,7 +43,7 @@ if __name__ == "__main__":
             ),
         ],
     )
-    outputs[0].request_output.images[0].save("output.png")
+    outputs[0].images[0].save("output.png")
 ```
 
 ## Image-to-Image (Image Editing)
@@ -72,7 +72,7 @@ if __name__ == "__main__":
             ),
         ],
     )
-    outputs[0].request_output.images[0].save("output.png")
+    outputs[0].images[0].save("output.png")
 ```
 
 ## Generation Parameters

docs/user_guide/examples/offline_inference/text_to_image.md

@@ -55,7 +55,7 @@ if __name__ == "__main__":
     omni = Omni(model="Qwen/Qwen-Image")
     prompt = "a cup of coffee on the table"
     outputs = omni.generate(prompt)
-    images = outputs[0].request_output.images
+    images = outputs[0].images
     images[0].save("coffee.png")
 ```
 
@@ -88,8 +88,9 @@ python text_to_image.py \
 | `--vae-use-tiling` | flag | off | Enable VAE tiling for memory optimization |
 | `--cfg-parallel-size` | int | `1` | Set to `2` to enable CFG Parallel |
 | `--enable-cpu-offload` | flag | off | Enable CPU offloading for diffusion models |
-| `--lora-path` | str | — | Path to PEFT LoRA adapter folder |
+| `--lora-path` | str | — | Path to PEFT LoRA adapter folder or checkpoint file |
 | `--lora-scale` | float | `1.0` | Scale factor for LoRA weights |
+| `--lora-backend` | str |`"peft"`| LoRA backend for loading LoRA adapters. Default: peft. Available options: peft, distill |
 
 **NextStep-1.1 specific arguments:**
 
@@ -178,7 +179,7 @@ if __name__ == "__main__":
     ]
     outputs = omni.generate(prompts)
     for i, output in enumerate(outputs):
-        output.request_output.images[0].save(f"{i}.jpg")
+        output.images[0].save(f"{i}.jpg")
 ```
 
 !!! info
@@ -206,7 +207,7 @@ if __name__ == "__main__":
         }
     ])
     for i, output in enumerate(outputs):
-        output.request_output.images[0].save(f"{i}.jpg")
+        output.images[0].save(f"{i}.jpg")
 ```
 
 You can also pass a negative prompt via the CLI argument `--negative-prompt`:

docs/user_guide/examples/offline_inference/text_to_video.md

@@ -12,6 +12,7 @@ For backend selection and SageAttention usage, see the [Diffusion Attention Back
 | Model | Default Resolution | Default Frames | Default Steps | Guidance | VRAM (BF16) |
 |---|---|---|---|---|---|
 | `Wan-AI/Wan2.2-T2V-A14B-Diffusers` | 720x1280 | 81 | 40 | 4.0 | ~60 GiB |
+| `robbyant/lingbot-video-dense-1.3b` / `robbyant/lingbot-video-moe-30b-a3b` | 192x320 | 9 | 2 | 3.0 | ~68 GiB (MoE smoke) |
 | `Lightricks/LTX-2` | 512x768 | 121 | 40 | video 3.0 / audio 7.0 | Model-dependent |
 | `hunyuanvideo-community/HunyuanVideo-1.5-Diffusers-480p_t2v` | 480x832 | 121 | 50 | 6.0 | 1×A100 80GB |
 | `hunyuanvideo-community/HunyuanVideo-1.5-Diffusers-720p_t2v` | 720x1280 | 121 | 50 | 6.0 | FP8 + VAE tiling required |
@@ -35,6 +36,25 @@ python text_to_video.py \
   --output t2v_out.mp4
 ```
 
+### LingBot-Video
+
+The shared runner recognizes both official LingBot checkpoint IDs and selects
+the native `LingBotVideoPipeline` automatically:
+
+```bash
+python text_to_video.py \
+  --model robbyant/lingbot-video-dense-1.3b \
+  --prompt "a robotic arm picks up a red block" \
+  --height 192 --width 320 --num-frames 9 --num-inference-steps 2 \
+  --guidance-scale 3.0 --flow-shift 3.0 --fps 24 \
+  --output lingbot_t2v.mp4
+```
+
+Use `robbyant/lingbot-video-moe-30b-a3b` for the MoE checkpoint. For local
+paths without `lingbot` in the directory name, add
+`--model-class-name LingBotVideoPipeline`. Model-only options use
+`--extra-body`, for example `'{"batch_cfg": true, "output_type": "np"}'`.
+
 ### LTX-2
 
 ```bash
@@ -125,6 +145,9 @@ python text_to_video.py \
 - `--audio-sample-rate`: fallback audio sample rate when the pipeline returns audio.
 - `--quantization`: quantization method (`fp8` for FP8, `gguf` for GGUF).
 - `--flow-shift`: scheduler flow_shift parameter.
+- `--lora-path`: path to PEFT LoRA adapter folder or checkpoint file.
+- `--lora-scale`: scale factor for LoRA weights.
+- `--lora-backend`: backend for loading LoRA adapters. Default: peft. Available options: peft, distill.
 
 ### Wan2.2-specific
 

docs/user_guide/fault_injection_reliability_matrix.md

@@ -65,7 +65,7 @@ More failure modes (for example, network jitter and network interruption) will b
 
 ## Source of Conclusions
 
-The behaviors and conclusions above are summarized from current fault injection validation results on **`Qwen3-Omni`**, **`Wan2.2`**, **`HunyuanImage-3.0-Instruct`** (DiT-only, `/v1/images/generations`), and **`VoxCPM2`** (`/v1/audio/speech`). Automated coverage lives under `tests/dfx/reliability/` (`test_reliability_qwen3_omni.py`, `test_reliability_wan22.py`, `test_reliability_hunyuan_image.py`, `test_reliability_voxcpm2.py`) and runs weekly via `.buildkite/cuda/test-weekly.yml`.
+The behaviors and conclusions above are summarized from current fault injection validation results on **`Qwen3-Omni`**, **`Wan2.2`**, and **`HunyuanImage-3.0-Instruct`** (DiT-only, `/v1/images/generations`). Automated coverage lives under `tests/dfx/reliability/` (`test_reliability_qwen3_omni.py`, `test_reliability_wan22.py`, `test_reliability_hunyuan_image.py`) and runs weekly via `.buildkite/cuda/test-weekly.yml`.
 
 ### Example: `SIGTERM` fault injection log (Qwen3-Omni)
 

docs/user_guide/feature_compatibility.md

@@ -165,7 +165,7 @@ vllm serve Qwen/Qwen-Image --omni --port 8091 \
 
 - **Tensor Parallelism — Text Encoder Not Sharded**: TP currently only shards the DiT blocks. Each TP rank retains a **full copy of the text encoder weights**, leading to significant GPU memory overhead proportional to TP degree. Tracked in [Issue #771](https://github.com/vllm-project/vllm-omni/issues/771).
 
-- **CPU Offloading — Two Modes Are Mutually Exclusive**: Model-level offload (`enable_cpu_offload`) and layerwise offload (`enable_layerwise_offload`) cannot be used simultaneously. If both are set, layerwise takes priority and model-level is silently ignored.
+- **CPU Offloading — Strategies Are Mutually Exclusive**: Distributed layerwise, layerwise, and model-level offload should not be enabled together. Selection priority is distributed layerwise, then layerwise, then model-level.
 
 - **CPU Offloading — VAE stays on GPU**: Both offloading strategies keep the VAE on GPU at all times. For high-resolution generation, VAE decode can still cause OOM. Mitigate by combining with `vae_use_tiling=True` or VAE Patch Parallelism.
 

docs/user_guide/quantization/fp8.md

@@ -80,12 +80,14 @@ warmup_quack_fp8([(14040, 2048, 6144), (14040, 2048, 2048)])
 
 ## Model Type Support
 
-### Diffusion Model (Qwen-Image, Wan2.2)
+### Diffusion Models
 
 | Model | HF models | Online | Pre-calibrated | Recommendation | `ignored_layers` | Text-Encoder quantization |
 |-------|-----------|:-------:|:------:|----------------|------------------|------------------|
 | Qwen-Image | `Qwen/Qwen-Image`, `Qwen/Qwen-Image-2512` | Yes | Yes | Skip sensitive image-stream MLPs when quality regresses | `img_mlp` | |
 | Wan2.2 | Wan2.2 diffusion pipelines | Not validated | Not validated | Validate against BF16 before documenting as supported | TBD | |
+| LTX-2 | `Lightricks/LTX-2`, `rootonchair/LTX-2-19b-distilled` | Yes | Not validated | Transformer only; use dynamic phase LoRA for ordinary two-stage | None | |
+| LTX-2.3 | `diffusers/LTX-2.3-Diffusers`, `diffusers/LTX-2.3-Distilled-Diffusers` | Yes | Not validated | Transformer only; use dynamic phase LoRA for ordinary two-stage | None | |
 | Z-Image | `Tongyi-MAI/Z-Image-Turbo` | Yes | Yes | All layers | None | ✅︎ |
 | FLUX.1 | `black-forest-labs/FLUX.1-dev`, `black-forest-labs/FLUX.1-schnell` | Yes | Yes | All layers | None | |
 | FLUX.2-dev | `black-forest-labs/FLUX.2-dev` | Yes | Not validated | All layers | None | ✅︎ |
@@ -93,7 +95,7 @@ warmup_quack_fp8([(14040, 2048, 6144), (14040, 2048, 2048)])
 | HunyuanImage-3.0 | `tencent/HunyuanImage-3.0`, `tencent/HunyuanImage-3.0-Instruct` | Yes | Yes | All layers; use the Hunyuan stage config for multi-stage runs | None | |
 | HunyuanVideo-1.5 | `hunyuanvideo-community/HunyuanVideo-1.5-Diffusers-480p_t2v`, `720p_t2v`, `480p_i2v` | Yes | Yes | All layers | None | |
 | Cosmos3 | `nvidia/Cosmos3-Nano`, `nvidia/Cosmos3-Super` | Yes | Not validated | All layers | None | |
-| MiniMax H3 | `MiniMaxAI/MiniMax-H3` | Yes | Not validated | DiT linears except patch, timestep, and final projections; not compatible with layerwise offload | None | |
+| MiniMax-H3 | `MiniMaxAI/MiniMax-H3` (`FL2VA` / `Ref2VA`) | Yes | Not validated | `quantization="fp8"` quantizes eligible DiT and text-encoder linears; mixed-precision input/output heads stay FP32 | None | ✅︎ |
 
 ### Multi-Stage Omni/TTS Model (Qwen3-Omni, Qwen3-TTS)
 
@@ -136,6 +138,39 @@ outputs = omni.generate(
 )
 ```
 
+### Global and per-component scope
+
+A plain, global quantization configuration is passed unchanged to every
+quantization-aware component constructed by a pipeline. This includes an
+eligible encoder when that encoder is implemented with vLLM quantizable
+layers; it does not rewrite arbitrary `torch.nn` modules. A structured
+component map is the only way to narrow that scope. Pipeline integrations that
+do not yet expose an encoder through the quantization factory remain DiT-only.
+
+For a pipeline that exposes both a transformer and a quantization-aware text
+encoder, the scope is:
+
+| Configuration | Transformer | Text encoder | Components without supported quantizable layers |
+|---------------|-------------|--------------|-------------------------------------------------|
+| `quantization="fp8"` | FP8 | FP8 | checkpoint precision |
+| `{"transformer": {"method": "fp8"}}` | FP8 | checkpoint precision | checkpoint precision |
+| `{"text_encoder": {"method": "fp8"}}` | checkpoint precision | FP8 | checkpoint precision |
+
+Use `quantization_config` for component-selective Python configuration. For
+example, quantize only the text decoder:
+
+```python
+omni = Omni(
+    model="<your-model>",
+    quantization_config={"text_encoder": {"method": "fp8"}},
+)
+```
+
+Component keys are runtime prefixes exposed by the pipeline integration; common
+keys include `transformer` and `text_encoder`. Entries may be combined, and
+`ignored_layers` can keep named eligible layers in checkpoint precision. Check
+the model recipe for supported components and their runtime prefixes.
+
 CLI:
 
 ```bash
@@ -144,6 +179,24 @@ python text_to_image.py --model <your-model> --quantization fp8 --ignored-layers
 vllm serve <your-model> --omni --quantization fp8
 ```
 
+### Distributed layerwise offload
+
+Per-tensor online FP8 linears can be combined with DLO's default AllGather path:
+
+```bash
+vllm serve <your-model> \
+  --omni \
+  --quantization fp8 \
+  --enable-distributed-layerwise-offload
+```
+
+Direct checkpoint mmap is not used for this combination. Each rank runs the
+ordinary loader to create the final FP8 weights and scales before DLO shards
+them, so startup temporarily materializes the complete FP8 model in host memory
+on every rank. Runtime DLO host residency is sharded normally. Use
+`--dlo-no-use-allgather` when the synchronized AllGather request-wave contract
+does not fit the workload.
+
 ## Parameters
 
 | Parameter | Type | Default | Description |

docs/user_guide/quantization/overview.md

@@ -12,20 +12,20 @@ For the internal architecture and backend extension points, see the
 
 | Mode | Guide | Description | Methods |
 |------|-------|-------------|---------|
-| Online quantization | [Online Quantization](online.md) | vLLM-Omni computes quantized weights and scales while loading the model. | FP8 W8A8, Int8 W8A8, MXFP8 W8A8, MXFP4 W4A4 |
+| Online quantization | [Online Quantization](online.md) | vLLM-Omni computes quantized weights and scales while loading the model. | FP8 W8A8, Int8 W8A8, BitsAndBytes W4, MXFP8 W8A8, MXFP4 W4A4 |
 | Runtime attention quantization | [Quantized KV Cache](quantized_kvcache.md) | vLLM-Omni dynamically quantizes eligible diffusion Flash Attention tensors during inference. | FP8 FA |
-| Pre-quantized checkpoints | Method-specific guides | The checkpoint or an offline quantizer provides quantized weights and scales before serving. | ModelOpt, GGUF, AutoRound, msModelSlim, serialized Int8, offline MXFP8, offline MXFP4 DualScale |
+| Pre-quantized checkpoints | Method-specific guides | The checkpoint or an offline quantizer provides quantized weights and scales before serving. | ModelOpt, AutoRound, msModelSlim, serialized Int8, offline MXFP8, offline MXFP4 DualScale |
 
 ## Hardware Support
 
-| Device | FP8 W8A8 | Int8 W8A8 | ModelOpt | MXFP8 W8A8 | MXFP4 W4A4 | GGUF | AutoRound | msModelSlim |
-|--------|----------|-----------|----------|------------|------------|------|-----------|-------------|
-| NVIDIA Blackwell GPU (SM 100+) | ✅ | ✅ | ✅ | ⭕ | ⭕ | ✅ | ✅ | ❌ |
-| NVIDIA Ada/Hopper GPU (SM 89+) | ✅ | ✅ | ✅ | ⭕ | ⭕ | ✅ | ✅ | ❌ |
-| NVIDIA Ampere GPU (SM 80+) | ✅ | ✅ | ⭕ | ⭕ | ⭕ | ✅ | ✅ | ❌ |
-| AMD ROCm | ⭕ | ⭕ | ⭕ | ⭕ | ✅ | ⭕ | ⭕ | ❌ |
-| Intel XPU | ⭕ | ⭕ | ⭕ | ⭕ | ⭕ | ⭕ | ✅ | ❌ |
-| Ascend NPU | ❌ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ✅ |
+| Device | FP8 W8A8 | Int8 W8A8 | BitsAndBytes W4 | ModelOpt | MXFP8 W8A8 | MXFP4 W4A4 | AutoRound | msModelSlim |
+|--------|----------|-----------|------------------|----------|------------|------------|-----------|-------------|
+| NVIDIA Blackwell GPU (SM 100+) | ✅ | ✅ | ✅ | ✅ | ⭕ | ⭕ | ✅ | ❌ |
+| NVIDIA Ada/Hopper GPU (SM 89+) | ✅ | ✅ | ✅ | ✅ | ⭕ | ⭕ | ✅ | ❌ |
+| NVIDIA Ampere GPU (SM 80+) | ✅ | ✅ | ✅ | ⭕ | ⭕ | ⭕ | ✅ | ❌ |
+| AMD ROCm | ⭕ | ⭕ | ❌ | ⭕ | ⭕ | ✅ | ⭕ | ❌ |
+| Intel XPU | ⭕ | ⭕ | ❌ | ⭕ | ⭕ | ⭕ | ✅ | ❌ |
+| Ascend NPU | ❌ | ✅ | ❌ | ❌ | ✅ | ✅ | ❌ | ✅ |
 
 Legend: `✅` supported, `❌` unsupported, `⭕` not verified in this
 guide. FP8 on Ampere may use a weight-only path where available.
@@ -43,10 +43,10 @@ otherwise.
 |--------|-------|------|----------------|--------|
 | FP8 W8A8 | [FP8](fp8.md) | Online W8A8 or checkpoint FP8 | Qwen-Image; Wan2.2 is not validated | Validated for Qwen-Image family and other DiT models |
 | Int8 W8A8 | [Int8](int8.md) | Online or serialized W8A8 | Qwen-Image; Wan2.2 is not validated | Validated for Qwen-Image and Z-Image |
+| BitsAndBytes W4 | [BitsAndBytes](bitsandbytes.md) | Online W4 weight-only | Z-Image-Turbo; Qwen-Image/Wan2.2 not validated | Validated for Z-Image-Turbo |
 | ModelOpt | [ModelOpt](modelopt.md) | Pre-quantized FP8 checkpoints | Qwen-Image, Z-Image, FLUX.2, HunyuanImage-3.0 | Validated for ModelOpt FP8 diffusion checkpoints |
 | MXFP8 W8A8 | [MXFP8](mxfp8.md) | Online W8A8 or offline pre-quantized | Wan2.2-T2V-A14B, I2V-A14B, TI2V-5B | Ascend NPU only; validated for Wan2.2 |
 | MXFP4 W4A4 | [MXFP4](mxfp4.md) | `mxfp4`: online single-scale only; `mxfp4_dualscale`: online or offline dual-scale (offline recommended) | Wan2.2-T2V-A14B, I2V-A14B | Ascend NPU only; validated for Wan2.2 A14B cascade models; TI2V-5B not supported; offline `mxfp4_dualscale` uses calibrated `mul_scale` for best accuracy |
-| GGUF | [GGUF](gguf.md) | Pre-quantized transformer weights | Qwen-Image | Validated where a model-specific GGUF adapter exists |
 | AutoRound | [AutoRound](autoround.md) | Pre-quantized W4A16 checkpoints | FLUX.1-dev; Qwen-Image/Wan2.2 not validated | Checkpoint-driven |
 | msModelSlim | [msModelSlim](msmodelslim.md) | Pre-quantized Ascend checkpoints | Wan2.2 recipe; HunyuanImage-3.0 inference target | Ascend/NPU path |
 
@@ -61,9 +61,9 @@ in BF16 unless the model guide explicitly adds support.
 |--------|-------|-------|----------------|--------|
 | ModelOpt | [ModelOpt](modelopt.md) | Thinker or language-model checkpoint config | Qwen3-Omni thinker | ModelOpt checkpoint path |
 | Int8 | [Int8](int8.md) | Not currently validated for omni/TTS stages | Qwen3-Omni, Qwen3-TTS | Not validated |
+| BitsAndBytes | [BitsAndBytes](bitsandbytes.md) | Not currently validated for omni/TTS stages | Qwen3-Omni, Qwen3-TTS | Not validated |
 | MXFP8 | [MXFP8](mxfp8.md) | Not currently validated for omni/TTS stages | Qwen3-Omni, Qwen3-TTS | Not validated |
 | MXFP4 | [MXFP4](mxfp4.md) | Not currently validated for omni/TTS stages | Qwen3-Omni, Qwen3-TTS | Not validated |
-| GGUF | [GGUF](gguf.md) | Not currently validated for omni/TTS stages | Qwen3-Omni, Qwen3-TTS | Not validated |
 | AutoRound | [AutoRound](autoround.md) | Thinker or language-model checkpoint config | Qwen2.5-Omni, Qwen3-Omni | Supported through AutoRound checkpoints |
 | msModelSlim | [msModelSlim](msmodelslim.md) | Not currently validated for omni/TTS stages | Qwen3-Omni, Qwen3-TTS | Not validated |
 
@@ -76,10 +76,10 @@ attached to the intended stage rather than applied globally.
 |--------|-------|-------|----------------|--------|
 | FP8 | [FP8](fp8.md) | Stage-specific DiT or transformer module | BAGEL, GLM-Image | Requires model-specific validation |
 | Int8 | [Int8](int8.md) | Stage-specific DiT or transformer module | BAGEL, GLM-Image | Requires model-specific validation |
+| BitsAndBytes | [BitsAndBytes](bitsandbytes.md) | Stage-specific transformer or DiT module | BAGEL, GLM-Image | Not validated |
 | ModelOpt | [ModelOpt](modelopt.md) | Checkpoint-defined diffusion stage | BAGEL, GLM-Image | Requires model-specific validation |
 | MXFP8 | [MXFP8](mxfp8.md) | Stage-specific DiT or transformer module | BAGEL, GLM-Image | Not validated |
 | MXFP4 | [MXFP4](mxfp4.md) | Stage-specific DiT or transformer module | BAGEL, GLM-Image | Not validated |
-| GGUF | [GGUF](gguf.md) | Stage-specific transformer weights | BAGEL, GLM-Image | No validated adapter listed |
 | AutoRound | [AutoRound](autoround.md) | Checkpoint-defined stage | BAGEL, GLM-Image | No validated checkpoint listed |
 | msModelSlim | [msModelSlim](msmodelslim.md) | Ascend-generated stage weights | GLM-Image | Requires model-specific adaptation |
 
@@ -106,7 +106,7 @@ config = build_quant_config({
 
 | Component | Default quantized? | Notes |
 |-----------|--------------------|-------|
-| Diffusion transformer | Yes | Primary target for FP8, Int8, ModelOpt, MXFP8, MXFP4, GGUF, AutoRound, and msModelSlim |
+| Diffusion transformer | Yes | Primary target for FP8, Int8, BitsAndBytes, ModelOpt, MXFP8, MXFP4, AutoRound, and msModelSlim |
 | Text encoder | No | Keep BF16 unless a method-specific guide documents support |
 | VAE | No | Keep BF16; storage-only paths are method-specific |
 | Scheduler/tokenizer | No | Loaded from the base model repository |
@@ -139,8 +139,8 @@ from vllm_omni.quantization import build_quant_config
 
 build_quant_config("fp8")
 build_quant_config({"method": "fp8", "activation_scheme": "static"})
+build_quant_config("bitsandbytes")
 build_quant_config("auto-round", bits=4, group_size=128)
-build_quant_config({"method": "gguf", "gguf_model": "/path/to/model.gguf"})
 build_quant_config({"transformer": {"method": "fp8"}, "vae": None})
 build_quant_config(None)
 ```

examples/offline_inference/aura_omni/end2end.py

@@ -24,7 +24,6 @@ from vllm.assets.video import VideoAsset, video_to_ndarrays
 from vllm.multimodal.media.audio import load_audio
 from vllm.utils.argparse_utils import FlexibleArgumentParser
 
-from vllm_omni.engine.arg_utils import nullify_stage_engine_defaults
 from vllm_omni.entrypoints.omni import Omni
 from vllm_omni.model_executor.stage_input_processors.aura_omni import (
     DEFAULT_QWEN3_TTS_REF_TEXT,
@@ -131,7 +130,7 @@ def build_sampling_params() -> list[SamplingParams]:
 
 
 def save_stage_output(stage_outputs, output_dir: str) -> None:
-    output = stage_outputs.request_output
+    output = stage_outputs
     request_id = output.request_id
     if stage_outputs.final_output_type == "text":
         text_output = output.outputs[0].text
@@ -228,7 +227,6 @@ def parse_args():
     parser.add_argument("--init-timeout", type=int, default=2000, help="Overall initialization timeout.")
     parser.add_argument("--stage-init-timeout", type=int, default=2000, help="Per-stage initialization timeout.")
     parser.add_argument("--log-stats", action="store_true", default=False, help="Enable stats logging.")
-    nullify_stage_engine_defaults(parser)
     return parser.parse_args()
 
 

examples/offline_inference/covo_audio/end2end.py

@@ -59,7 +59,7 @@ def main(args):
 
     omni = Omni(
         model=args.model_name,
-        stage_configs_path=args.stage_configs_path,
+        deploy_config=args.deploy_config,
         log_stats=args.log_stats,
         stage_init_timeout=args.stage_init_timeout,
         batch_timeout=args.batch_timeout,
@@ -104,7 +104,7 @@ def main(args):
     os.makedirs(output_dir, exist_ok=True)
 
     for stage_outputs in omni_outputs:
-        output = stage_outputs.request_output
+        output = stage_outputs
         if stage_outputs.final_output_type == "text":
             request_id = output.request_id
             text_output = output.outputs[0].text
@@ -167,10 +167,10 @@ def parse_args():
         help="Sampling rate for audio loading (default: 16000).",
     )
     parser.add_argument(
-        "--stage-configs-path",
+        "--deploy-config",
         type=str,
         default=None,
-        help="Path to stage configs YAML file.",
+        help="Path to a deploy config YAML file.",
     )
     parser.add_argument(
         "--log-stats",