Merge branch 'main' into codex-split-attention-offload-guides

Signed-off-by: Joshna Medisetty <joshna.medisetty@intel.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: Chendi.Xue <chendi.xue@intel.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.
+# Run vLLM-Omni ROCm tests directly in the MI300 Kubernetes pod. The pod's
-set -o pipefail
+# container image is selected by test-template-amd-omni.j2; MI300 has no DinD.
+set -euo pipefail
 
-# Export Python path
+: "${PYTHONFAULTHANDLER:=1}"
-export PYTHONPATH=".."
+: "${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
+expected_gpus="${VLLM_CI_EXPECTED_GPU_COUNT:-1}"
-cleanup_docker() {
+python3 - <<'PY'
-  # Get Docker's root directory
+import os
-  docker_root=$(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
-}
 
-# Call the cleanup docker function
+import torch
-cleanup_docker
 
-echo "--- Pulling container"
+expected = int(os.environ.get("VLLM_CI_EXPECTED_GPU_COUNT", "1"))
-## Temporary change to use AMD Docker Hub to store the vllm-omni image
+assert torch.version.hip, "PyTorch is not a ROCm build"
-# to bypass the rate limit issue with ECR Public Gallery.
+assert torch.cuda.is_available(), "ROCm GPU is not available to PyTorch"
-# Images are now stored in a separate repository for vllm-omni, instead of vllm-ci.
+actual = torch.cuda.device_count()
-# TODO: @tjtanaa point back to ECR Public Gallery
+assert actual == expected, f"Expected {expected} ROCm GPU(s), found {actual}"
-# once the amd agents are configured to use ECR Public Gallery.
+PY
-# 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)"
 
-# TODO: @tjtanaa uncomment this once the amd agents are configured to use ECR Public Gallery.
+echo "Commands:${commands}"
-# # Install AWS CLI to authenticate to ECR Public Gallery to get higher rate limit for pulling images
+if /bin/bash -o pipefail -c '
-# sudo apt-get update && sudo apt-get install -y awscli
+set -E
-##  Use safe docker login helper to prevent race conditions
+test_status=0
-# source "$(dirname "${BASH_SOURCE[0]}")/../../common/scripts/docker_login_ecr_public.sh"
+trap '\''
-# safe_docker_login_ecr_public
+    command_status=$?
-
+    # Keep running subsequent commands, but retain a failure for the final exit.
-## Pull the container from AMD Docker Hub
+    # Prefer a test failure over pytest "no tests collected" when both occur.
-
+    if (( test_status == 0 || (test_status == 5 && command_status != 5) )); then
-docker pull "${image_name}"
+        test_status=${command_status}
-
-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}"
     fi
-  done
+'\'' ERR
+eval "$1"
+exit "${test_status}"
+' _ "${commands}"; then
+    exit 0
 else
-  echo "Render devices: $BUILDKITE_AGENT_META_DATA_RENDER_DEVICES"
+    exit_code=$?
-  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}"
 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)"
+    - label: "Simple · Model Executor Test · Shard %N/%t"
-      agent_pool: mi325_1
+      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.
+        - "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"
-        - "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"
 
-    - label: "Simple · Diffusion & Model Executor Test (2/2)"
+    - label: "Simple · Diffusion Test · Shard %N/%t"
-      agent_pool: mi325_1
+      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.
+        # ignore test_teacache_extractors.py because it use rocm gemm kernel from vLLM
-        - "timeout 60m pytest -sv tests/model_executor -m 'core_model and cpu' --cov=vllm_omni --cov-report=term-missing:skip-covered --cov-report=xml"
+        # 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"
+    - label: "Qwen3-TTS CustomVoice E2E Test · Shard %N/%t"
-      agent_pool: mi325_1
+      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"
+    - label: "Simple · Model Executor Test · Shard %N/%t"
-      agent_pool: mi325_1
+      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
-        - |
+        - pytest -s -v tests/diffusion/offloader -m "core_model and cuda" --run-level "core_model"
-          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))
-          '
 
     - 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
+   Last synced: 2026-08-04
-   Modifications: Removed unused CUDA/NVIDIA logic, keeping only AMD tests
+   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

@@ -9,6 +9,11 @@
 #   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
+#
+# 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

@@ -94,13 +94,19 @@ steps:
           - vllm_omni/model_executor/models/qwen3_tts/
           - 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_LOGGING_LEVEL=DEBUG
+            export VLLM_ALLOW_LONG_MAX_MODEL_LEN=1
-              export VLLM_ALLOW_LONG_MAX_MODEL_LEN=1
+            .buildkite/common/scripts/run_cov_split.sh \
-              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'
+              --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: "TTS · Higgs-Audio-V3 Test"
@@ -203,9 +209,17 @@ steps:
           - vllm_omni/diffusion/models/bagel/
           - vllm_omni/model_executor/stage_input_processors/bagel.py
           - vllm_omni/deploy/bagel.yaml
-        timeout_in_minutes: 20
+          - .buildkite/common/scripts/run_cov_split.sh
+          - pyproject.toml
+        # 20 -> 40: the coverage split runs the model twice.
+        timeout_in_minutes: 40
         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'
+          - |
+            .buildkite/common/scripts/run_cov_split.sh \
+              --model-id bagel \
+              --offline tests/e2e/offline_inference/test_bagel.py \
+              --online 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"

.buildkite/cuda/test-nightly.yml

@@ -456,7 +456,7 @@ 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"
+          - pytest -s -v tests/e2e/online_serving/test_wan_2_1_vace_expansion.py tests/e2e/online_serving/test_ltx.py -m "full_model and cuda" --run-level "full_model"
         mirror_hardwares: h100_2
 
       - label: ":full_moon: Diffusion X2V · Wan2.2 Accuracy Test"
@@ -477,10 +477,16 @@ 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 · LTX Accuracy Test"
+        timeout_in_minutes: 180
+        commands:
+          - pytest -s -v 'tests/e2e/accuracy/ltx/test_ltx_official_similarity.py::test_ltx_matches_official[ltx2_3_i2v]' -m full_model --run-level full_model
+          - pytest -s -v 'tests/e2e/accuracy/ltx/test_ltx_official_similarity.py::test_ltx_matches_official[ltx23_two_stage_layer_fused_i2v]' -m full_model --run-level full_model
+        mirror_hardwares: h100_1
+
       - label: ":full_moon: Diffusion X2V · Perf Test"
         key: nightly-diffusion-x2v-performance
         timeout_in_minutes: 180

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

@@ -87,8 +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 $XPU_TEST_PATHS -m "advanced_model and xpu and B60"
-    pytest -v -s -m "omni and xpu and B60"
+    pytest -v -s $XPU_TEST_PATHS -m "omni and xpu and B60"
 '

.gitignore

@@ -46,6 +46,8 @@ htmlcov/
 .cache
 nosetests.xml
 coverage.xml
+coverage-*.xml
+coverage-*.xml.gz
 *.cover
 *.py,cover
 .hypothesis/

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.

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

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

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

docs/.nav.yml

@@ -60,7 +60,11 @@ nav:
       - Overview and Compatibility:
         - Overview: user_guide/diffusion_features.md
         - Feature Compatibility: user_guide/feature_compatibility.md
-      - CPU Offloading: user_guide/diffusion/cpu_offload.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
@@ -72,7 +76,13 @@ nav:
         - 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: user_guide/diffusion/attention_backends.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
@@ -125,11 +135,16 @@ nav:
           - Tensor Parallel: design/feature/tensor_parallel.md
           - VAE Patch Parallelism: design/feature/vae_parallel.md
         - Attention Optimization:
+          - Backend Selection: design/feature/attention_backend_selection.md
           - Skip-Softmax: design/feature/skip_softmax.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
@@ -163,7 +178,8 @@ 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 bugs and feature requests, search or open a GitHub [issue](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 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 security disclosures, please use GitHub's [Security Advisories](https://github.com/vllm-project/vllm-omni/security/advisories) feature
+- 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
+`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.
-
-- 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.
 
 ## 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/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
 
@@ -122,7 +122,7 @@ Repository skills guide the agent, but they do not replace contributor judgment,
 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 +327,96 @@ 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): two Merge-tier (L3) jobs in `.buildkite/cuda/test-merge.yml` upload
+per-model, per-entry-mode coverage as Buildkite artifacts — "Diffusion · Bagel
+Test" and "TTS · Qwen3-TTS Base Test". Both run on a single GPU so the pilot is
+cheap to reproduce. 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` asserts this for the pilots.
+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/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/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/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

@@ -7,7 +7,7 @@ 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)
+[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).
 
 ## Status
@@ -56,7 +56,7 @@ AllGather:                [full N+1]             [full N+2]
 Buffers:    [current slot]       [prefetch slot]       [current slot]
 ```
 
-![DLO double-buffer prefetch pipeline](../figures/dlo/dlo_pipeline.gif)
+![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.

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

@@ -186,7 +186,7 @@ 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)
+combining these features. See the [DLO design document](offloader/distributed_layerwise_offload.md)
 for the full compatibility matrix.
 
 ## Adding a quantization backend

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
+[TRTLLM Attention](../../user_guide/diffusion/attention_backends/trtllm.md#skip-softmax).
-the algorithm.
+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

@@ -53,15 +53,22 @@ implementation contract; it is not, by itself, a general support claim.
 #### Attention optimization
 
 The [Diffusion Attention Backends](../user_guide/diffusion/attention_backends.md)
-guide lists all selectable backends and their platform defaults. Skip-Softmax
+guides list selectable backends, platform defaults, installation, and tuning.
-is the backend-related optimization with a standalone feature-design contract:
+The design contracts separate selection mechanics from backend algorithms:
 
+- [Attention Backend Selection](feature/attention_backend_selection.md)
 - [Skip-Softmax](feature/skip_softmax.md)
 
+#### CPU offloading
+
+- [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/features/README.md

@@ -48,10 +48,10 @@ defines the shared configuration and backend extension points.
 | 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) | [Distributed Layerwise Offload](../design/feature/distributed_layerwise_offload.md) |
+| 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) | [Skip-Softmax](../design/feature/skip_softmax.md) |
+| 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 |

docs/models/supported_models.md

@@ -42,7 +42,8 @@ th {
 | `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 | ✅︎ | ✅︎ | | | — |
+| `LTX2DistilledPipeline` | LTX-2 / LTX-2.3 full-distilled two-stage T2V and I2V | `rootonchair/LTX-2-19b-distilled`, `diffusers/LTX-2.3-Distilled-Diffusers` + `Lightricks/LTX-2.3` upsampler | ✅︎ | ✅︎ | | | — |
 | `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` | ✅︎ | ✅︎ | ✅︎ | ✅︎ | — |
@@ -58,6 +59,7 @@ th {
 | `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` | ✅︎ | ✅︎ | ✅︎ | ✅︎ | — |

docs/usage/faq.md

@@ -10,15 +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: AWQ is not currently listed as a validated vLLM-Omni path. Quantization
+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.
-support depends on the model, pipeline stage, checkpoint format, and hardware
-backend. See the current [quantization support matrix](../user_guide/quantization/overview.md)
-for online methods, pre-quantized checkpoints, and model-level validation.
 
 > 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/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
+## 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. |
 
+| 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).
+1. `--diffusion-attention-config` per-role configuration.
-2. **`--diffusion-attention-backend` / `DIFFUSION_ATTENTION_BACKEND` env var** — global shorthand that sets the default backend.
+2. `--diffusion-attention-backend` or `DIFFUSION_ATTENTION_BACKEND` as a
-3. **Platform default** — used when nothing is configured.
+   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 FLASH_ATTN
-vllm-omni serve <model> --diffusion-attention-backend SAGE_ATTN
 
-# Environment variable (also recognized for backwards compatibility)
+# Backwards-compatible environment variable
-export DIFFUSION_ATTENTION_BACKEND=SAGE_ATTN
+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
+1. Exact `per_role[role]` match.
-2. `per_role[role_category]` fallback (e.g. `"ltx2.audio_to_video"` → `"cross"`)
+2. Category `per_role[role_category]` match.
-3. `default`
+3. `default`.
-4. Platform default
+4. Platform default.
-
-Use vLLM-style dotted flags or one JSON blob:
 
 ```bash
 # Dotted flags
 vllm-omni serve <model> \
-    --diffusion-attention-config.default.backend FLASH_ATTN \
+  --diffusion-attention-config.default.backend FLASH_ATTN \
-    --diffusion-attention-config.per_role.cross.backend TORCH_SDPA
+  --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:
+### Python API
-`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
-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={
+        per_role={"cross": AttentionSpec(backend="TORCH_SDPA")},
-            "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:
+## Platform defaults
-
-```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
 
 ### 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**
+1. `TRTLLM_ATTN` on datacenter Blackwell (sm_100/sm_103) when FlashInfer is
-2. `CUDNN_ATTN` — when cuDNN ≥ 9.5 is available (ships in PyTorch 2.5+ wheels)
+   available, `head_dim=128`, and the model declares a compatible packed or
-3. `FLASHINFER_ATTN` — when `flashinfer` is installed but cuDNN < 9.5
+   mask-free path.
-4. `FLASH_ATTN` — when `flash-attn` is installed with the Blackwell CUTE kernel
+2. `CUDNN_ATTN` when cuDNN 9.5 or newer is available.
-5. `TORCH_SDPA` — last resort
+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
+- a correctness reference (`TORCH_SDPA`);
-2. `TORCH_SDPA` — fallback
+- 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
+The following BF16 results were measured on an sm_120 RTX Pro 6000 Blackwell
-that it can enable **Skip-Softmax**, a sparse-attention mode that trades a little fidelity for speed
+with the same prompt and seed across runs. Treat them as reference results, not
-(algorithm: [design doc](../../design/feature/skip_softmax.md)).
+portable guarantees.
 
-Enable it through the typed `skip_softmax` block on the attention spec:
+| Model | Shape | `TORCH_SDPA` | `CUDNN_ATTN` | `FLASHINFER_ATTN` |
-
+| --- | --- | ---: | ---: | ---: |
-| 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 |
-|---|---|---|---|---|
 | 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
+## TRTLLM_ATTN SAGE Quantization
-DIFFUSION_ATTENTION_BACKEND=FLASHINFER_ATTN python examples/offline_inference/text_to_video/text_to_video.py \
-    --model Lightricks/LTX-2 ...
-```
 
-### 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
+See [RainFusion](attention_backends/rainfusion.md).
-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.
 
 ## SageAttention Installation
 
-vLLM-Omni expects SageAttention to be installed into the same Python environment as vLLM-Omni.
+See [SageAttention](attention_backends/sage.md#installation).
-
-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__)"
-```
 
 ## SageAttention3 Installation
 
-vLLM-Omni expects SageAttention3 to be installed into the same Python environment as vLLM-Omni.
+See [SageAttention](attention_backends/sage.md#sageattention3-installation).
-
-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.
 
 ## 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.
+See [Hugging Face Hub Backends](attention_backends/huggingface_hub.md).
-
-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
-```

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.
+## Choose a strategy
-2. **Layerwise (Blockwise) Offloading**: Keeps only one transformer block on GPU at a time with compute-memory overlap.
 
-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
+See [Model-Level Offloading](offloader/cpu_offload.md).
-
-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`.
-
 
 ## Layerwise (Blockwise) Offloading
 
-### How It Works
+See [Layerwise Offloading](offloader/layerwise_offload.md).
-
-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
-
 
 ## Distributed Layerwise Offloading
 
-### How It Works
+See [Distributed Layerwise Offloading](offloader/distributed_layerwise_offload.md).
-
-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.

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 +1,118 @@
+# 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 its standard-loader rank-local weights independently.
+
+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` | Keep standard-loader rank-local weights | `false` |
+| `--dlo-resident-layers N` | Keep N leading main-DiT blocks on device; requires no-AllGather and model-declared resident paths | `0` |
+
+## mmap weight loading
+
+The DLO plus AllGather path:
+
+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. calls model-specific `post_load_weights()` conversions; and
+5. restores the saved non-persistent buffers.
+
+This avoids one full checkpoint RSS copy per rank and does not require
+model-specific loading code.
+
+## 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
+
+- Online FP8 quantization is rejected with the DLO plus AllGather mmap path.
+  Use `--dlo-no-use-allgather` or disable online quantization.
+- Tensor parallel size greater than one is rejected in the mmap path because
+  it bypasses TP-aware loader callbacks. The no-AllGather path retains the
+  standard loader and remains experimental with TP.
+- HSDP plus AllGather is rejected to avoid double sharding. HSDP without
+  AllGather has limited end-to-end validation.
+- 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.
+
+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
@@ -147,3 +182,4 @@ 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_features.md

@@ -92,7 +92,7 @@ current support.
 |--------|--------------|-------------|----------|
 | **[FP8](quantization/fp8.md)** | `quantization="fp8"` | FP8 W8A8 on validated transformer stages | Memory reduction, inference speedup |
 | **[INT8](quantization/int8.md)** | `quantization="int8"` | INT8 W8A8 on validated transformer stages | Memory reduction, broad GPU compatibility |
-| **[BitsAndBytes W4](quantization/bitsandbytes.md)** | `quantization="bitsandbytes"` | Online NF4/FP4 weight-only quantization on validated CUDA transformer stages | Reducing transformer weight memory without preparing a separate checkpoint |
+| **[GGUF](quantization/gguf.md)** | `quantization="gguf"` | Native GGUF transformer-only weights (Q4, Q8, etc.) | Memory reduction on consumer GPUs |
 
 ## Supported Models
 
@@ -106,7 +106,10 @@ The following tables show which models support each feature:
 
 > 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
@@ -214,7 +217,7 @@ The following tables show which models support each feature:
     3. CPU Offloading (Layerwise) and CPU Offloading (Module-wise) are not compatible.
     4. The CPU Offloading (Layerwise) row describes local layerwise offload.
        Multi-device Distributed Layerwise Offload has a separate topology and
-       compatibility matrix in the [CPU Offloading guide](diffusion/cpu_offload.md#distributed-layerwise-offloading).
+       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).
 

docs/user_guide/examples/offline_inference/text_to_video.md

@@ -143,8 +143,7 @@ python text_to_video.py \
 - `--enable-layerwise-offload`: enable layerwise offloading on DiT modules.
 - `--frame-rate`: generation FPS for pipelines that require it (e.g., LTX2).
 - `--audio-sample-rate`: fallback audio sample rate when the pipeline returns audio.
-- `--quantization`: online quantization method, for example `fp8`; supported
+- `--quantization`: quantization method (`fp8` for FP8, `gguf` for GGUF).
-  pre-quantized checkpoints are normally auto-detected from checkpoint metadata.
 - `--flow-shift`: scheduler flow_shift parameter.
 - `--lora-path`: path to PEFT LoRA adapter folder or checkpoint file.
 - `--lora-scale`: scale factor for LoRA weights.

docs/user_guide/examples/online_serving/diffusers_pipeline_adapter.md

@@ -90,8 +90,8 @@ vllm serve "Qwen/Qwen-Image" \
     --quantization-config '{"method": "fp8"}'
 ```
 
-Other vLLM-Omni quantization methods, such as `modelopt`, `mxfp4`, `mxfp8`,
+Other vLLM-Omni quantization methods, such as `gguf`, `modelopt`, `mxfp4`,
-serialized checkpoints, static FP8 configs, and layer-name skip lists
+`mxfp8`, serialized checkpoints, static FP8 configs, and layer-name skip lists
 such as `ignored_layers`, are intentionally not translated by this compatibility
 shortcut. Use Diffusers-native configuration through `diffusers_load_kwargs` or
 a native vLLM-Omni pipeline for those cases.

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/autoround.md

@@ -10,10 +10,6 @@ checkpoint's `config.json` and auto-detects
 AutoRound is static quantization: no `--quantization` flag is needed at
 inference time when the checkpoint already contains the quantization config.
 
-The validation terms on this page follow the
-[support levels](overview.md#support-levels) defined in the overview. A listed
-checkpoint is not enough by itself to establish end-to-end model support.
-
 ## Hardware Support
 
 | Device | Support |
@@ -25,57 +21,46 @@ checkpoint is not enough by itself to establish end-to-end model support.
 | Intel XPU | ✅ |
 | Ascend NPU | ❌ |
 
-Legend: `✅` backend available, `❌` unsupported, `⭕` not verified in this
+Legend: `✅` supported, `❌` unsupported, `⭕` not verified in this
-guide. This table is method-level backend availability. AutoRound is an Intel
+guide. AutoRound is Intel-supported.
-project; vLLM selects a compatible CUDA or XPU compute backend at load time.
 
 ## Model Type Support
 
 ### Diffusion Model (Qwen-Image, Wan2.2)
 
-| Model | Checkpoint | Scope | Scheme | Validation |
+| Model | Checkpoint | Scope | Scheme | Backend |
-|-------|------------|-------|--------|------------|
+|-------|------------|-------|--------|---------|
-| FLUX.1-dev | `vllm-project-org/FLUX.1-dev-AutoRound-w4a16` | Diffusion transformer | W4A16 | Validated |
+| FLUX.1-dev | `vllm-project-org/FLUX.1-dev-AutoRound-w4a16` | Diffusion transformer | W4A16 | GPTQ-Marlin or Intel-supported AutoRound backend |
-| Qwen-Image | `INC4AI/Qwen-Image-AutoRound-W4A16` | Diffusion transformer | W4A16 | Validated with end-to-end generation and quality comparison |
+| Qwen-Image | Not listed | Diffusion transformer | W4A16 | Not validated |
-| Wan2.2-I2V | `Intel/Wan2.2-I2V-A14B-Diffusers-int4-AutoRound` | Both diffusion transformers | W4A16 | CI-backed |
+| Wan2.2-I2V | `Intel/Wan2.2-I2V-A14B-Diffusers-int4-AutoRound` | Diffusion transformer | W4A16 | GPTQ-Marlin or Intel-supported AutoRound backend |
-| Wan2.2-T2V | `Intel/Wan2.2-T2V-A14B-Diffusers-int4-AutoRound` | Both diffusion transformers | W4A16 | CI-backed |
+| Wan2.2-T2V | `Intel/Wan2.2-T2V-A14B-Diffusers-int4-AutoRound` | Diffusion transformer | W4A16 | GPTQ-Marlin or Intel-supported AutoRound backend |
-| Wan2.2-TI2V | `Intel/Wan2.2-TI2V-5B-Diffusers-int4-AutoRound` | Diffusion transformer | W4A16 | Validated; not in the scheduled Wan AutoRound job |
+| Wan2.2-TI2V | `Intel/Wan2.2-TI2V-5B-Diffusers-int4-AutoRound` | Diffusion transformer | W4A16 | GPTQ-Marlin or Intel-supported AutoRound backend |
-
-CUDA execution uses GPTQ-Marlin or another compatible vLLM AutoRound backend;
-Intel XPU uses the Intel-supported backend. The text encoder and VAE stay in
-BF16 for all entries above.
 
 ### World Model (Cosmos3)
 
-| Model | Checkpoint | Scope | Scheme | Validation |
+| Model | Checkpoint | Scope | Scheme | Backend |
-|-------|------------|-------|--------|------------|
+|-------|------------|-------|--------|---------|
-| Cosmos3-Nano | `Intel/Cosmos3-Nano-int4-AutoRound` | World-model transformer | W4A16 | Integrated; no model-specific end-to-end test is maintained in-tree |
+| Cosmos3-Nano | `Intel/Cosmos3-Nano-int4-AutoRound` | World-model transformer | W4A16 | GPTQ-Marlin or Intel-supported AutoRound backend |
-| Cosmos3-Super | `Intel/Cosmos3-Super-int4-AutoRound` | World-model transformer | W4A16 | Validated with a manual serving run |
+| Cosmos3-Super | `Intel/Cosmos3-Super-int4-AutoRound` | World-model transformer | W4A16 | GPTQ-Marlin or Intel-supported AutoRound backend |
-
-The Cosmos VAE and guardrail components stay in BF16.
 
 ### Multi-Stage Omni/TTS Model (Qwen3-Omni, Qwen3-TTS)
 
-| Model | Checkpoint | Scope | Scheme | Validation |
+| Model | Checkpoint | Scope | Scheme | Backend |
-|-------|------------|-------|--------|------------|
+|-------|------------|-------|--------|---------|
-| Qwen2.5-Omni-7B | `Intel/Qwen2.5-Omni-7B-int4-AutoRound` | Thinker language-model stage | W4A16 | Validated outside scheduled CI |
+| Qwen2.5-Omni-7B | `Intel/Qwen2.5-Omni-7B-int4-AutoRound` | Language-model stage | W4A16 | AutoRound |
-| Qwen3-Omni-30B-A3B-Instruct | `Intel/Qwen3-Omni-30B-A3B-Instruct-int4-AutoRound` | Thinker language-model stage | W4A16 | Validated outside scheduled CI |
+| Qwen3-Omni-30B-A3B-Instruct | `Intel/Qwen3-Omni-30B-A3B-Instruct-int4-AutoRound` | Thinker language-model stage | W4A16 | AutoRound |
 | Qwen3-TTS | Not listed | TTS language-model stage | W4A16 | Not validated |
 
 AutoRound support is checkpoint-driven. A model is supported when its
 checkpoint uses a compatible INC/AutoRound config and the target stage maps to
 vLLM-Omni's runtime module names.
 
-For Qwen-Omni, only the thinker language model uses the AutoRound checkpoint.
-Audio and vision encoders, the talker, and waveform-decoder stages stay in
-BF16.
-
 ### Multi-Stage Diffusion Model (BAGEL, GLM-Image)
 
-| Model | Checkpoint | Scope | Validation |
+| Model | Scope | Status | Notes |
-|-------|------------|-------|------------|
+|-------|-------|--------|-------|
-| GLM-Image | `Intel/GLM-Image-int4-AutoRound` | Diffusion transformer | CI-backed |
+| GLM-Image | Diffusion transformer  | ✅ | `Intel/GLM-Image-int4-AutoRound` |
-| BAGEL | Not listed | Checkpoint-defined diffusion or transformer stage | Not validated |
+| BAGEL | Checkpoint-defined diffusion or transformer stage | Not validated | Requires a compatible AutoRound checkpoint |
 
 ## Configuration
 
@@ -134,14 +119,6 @@ The checkpoint should contain a config like:
 At load time, vLLM-Omni builds an `OmniINCConfig`, remaps checkpoint block names
 to runtime module names, and selects the matching vLLM compute backend.
 
-Checkpoint auto-detection and successful layer construction establish the
-**Integrated** level. Promote a model to **Validated** only after the named
-checkpoint completes an end-to-end generation or multimodal request and its
-output passes a model-appropriate sanity or quality check. Compare against the
-BF16 baseline for quality-sensitive image and video paths. Promote it to
-**CI-backed** only when that named checkpoint is selected by scheduled hardware
-CI.
-
 Example checkpoint creation:
 
 ```bash

docs/user_guide/quantization/bitsandbytes.md

@@ -9,10 +9,6 @@ stay in BF16/FP16. No pre-quantized checkpoint is required.
 This is an online (dynamic) path: load the normal HuggingFace checkpoint and
 quantize during model loading via the `bitsandbytes` CUDA kernels.
 
-The validation terms on this page follow the
-[support levels](overview.md#support-levels). BitsAndBytes kernel support does
-not imply that every diffusion transformer has been quality-validated.
-
 ## Hardware Support
 
 | Device | Support |
@@ -25,7 +21,7 @@ not imply that every diffusion transformer has been quality-validated.
 | Intel XPU | ❌ |
 | Ascend NPU | ❌ |
 
-Legend: `✅` backend available, `❌` unsupported. Non-CUDA platforms raise in
+Legend: `✅` supported, `❌` unsupported. Non-CUDA platforms raise in
 `get_quant_method()`; CUDA requires compute capability 7.5+ (`SM 75+`).
 
 Requires the optional `bitsandbytes` package (`pip install bitsandbytes`).
@@ -34,11 +30,11 @@ Requires the optional `bitsandbytes` package (`pip install bitsandbytes`).
 
 ### Diffusion Model (Qwen-Image, Wan2.2)
 
-| Model | HF models | Scope | Level | Recommendation |
+| Model | HF models | CUDA | Mode | Recommendation |
-|-------|-----------|-------|-------|----------------|
+|-------|-----------|:----:|------|----------------|
-| Z-Image | `Tongyi-MAI/Z-Image-Turbo` | Diffusion transformer | Validated | All heavy linear layers; sensitive embedders stay BF16 |
+| Z-Image | `Tongyi-MAI/Z-Image-Turbo` | Yes | Online W4 weight-only | All heavy linear layers; sensitive embedders stay BF16 |
-| Qwen-Image | `Qwen/Qwen-Image`, `Qwen/Qwen-Image-2512` | Diffusion transformer | Not validated | Compare with BF16 before enabling |
+| Qwen-Image | `Qwen/Qwen-Image`, `Qwen/Qwen-Image-2512` | Not validated | Online W4 weight-only | Compare vs BF16 before enabling |
-| Wan2.2 | Wan2.2 diffusion pipelines | One or both diffusion transformers | Not validated | Validate quality and cascade behavior before enabling |
+| Wan2.2 | Wan2.2 diffusion pipelines | Not validated | Online W4 weight-only | Validate before enabling in docs |
 
 Other diffusion models may work if their transformer uses supported linear
 layers, but they are not validated in this guide.
@@ -108,10 +104,6 @@ Multi-GPU tensor parallelism (`tensor_parallel_size` > 1) is not validated for
 BitsAndBytes in diffusion models; each rank quantizes its own weight shard
 independently.
 
-Z-Image-Turbo is validated outside scheduled full-model BitsAndBytes CI. The
-real CUDA kernel smoke tests cover quantization and matrix multiplication, but
-do not establish model-level quality for unlisted architectures.
-
 If quality regresses, use `ignored_layers` to keep sensitive projections in BF16
 (for example `to_out` or `w2`).
 

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 | ✅︎ |

docs/user_guide/quantization/modelopt.md

@@ -13,11 +13,6 @@ components, such as the tokenizer, scheduler, text encoder, vision/audio
 encoder, and VAE, on the base checkpoint unless a model-specific guide says
 otherwise.
 
-The validation terms on this page follow the
-[support levels](overview.md#support-levels) defined in the overview. ModelOpt
-format detection and checkpoint adaptation are shared runtime integrations;
-each model and stage still requires its own end-to-end validation.
-
 !!! note
     ModelOpt checkpoints are pre-quantized checkpoints. Do not pass
     `--quantization fp8` for these checkpoints. The checkpoint
@@ -63,11 +58,11 @@ stay unquantized unless the model-specific recipe validates otherwise.
 | Intel XPU | ⭕ |
 | Ascend NPU | ❌ |
 
-Legend: `✅` backend available, `❌` unsupported, `⭕` not verified in this
+Legend: `✅` supported, `❌` unsupported, `⭕` not verified in this guide.
-guide. The optional CUTLASS FP8 runtime override requires CUDA SM89+. ModelOpt
+The optional CUTLASS FP8 runtime override requires CUDA SM89+. ModelOpt NVFP4
-NVFP4 and mixed FP8/NVFP4 diffusion checkpoints are currently validated on
+and mixed FP8/NVFP4 diffusion checkpoints are currently validated on Blackwell
-Blackwell CUDA systems in the recipes below; other CUDA generations require
+CUDA systems in the recipes below; other CUDA generations require separate
-separate backend and quality validation.
+backend and quality validation.
 
 ## Model Type Support
 
@@ -84,9 +79,6 @@ separate backend and quality validation.
 | HunyuanImage-3.0 | `feizhai123/hunyuan-image3-modelopt-mixed-experts-nvfp4-dense-fp8` | MoE diffusion transformer | Validated for ModelOpt mixed FP8/NVFP4 checkpoints |
 | Wan2.2 | Not available | Diffusion transformer | Not validated |
 
-The validated diffusion entries above have named checkpoint and recipe
-evidence. They are not all exercised by scheduled full-model CI.
-
 For full serving commands and benchmark context, see
 [`recipes/Qwen/Qwen-Image.md`](https://github.com/vllm-project/vllm-omni/blob/main/recipes/Qwen/Qwen-Image.md)
 and
@@ -96,8 +88,8 @@ and
 
 | Model | Scope | Status |
 |-------|-------|--------|
-| Qwen3-Omni | Thinker language-model stage (FP8) | Validated outside scheduled CI |
+| Qwen3-Omni | Thinker language-model stage | ModelOpt FP8 checkpoint path |
-| Qwen3-Omni | Thinker language-model stage (W4A4 NVFP4) | CI-backed for the H100 Marlin fallback; native Blackwell path is recipe-validated |
+| Qwen3-Omni | Thinker language-model stage (W4A4 NVFP4) | Validated; see [Qwen3-Omni NVFP4 W4A4](#qwen3-omni-nvfp4-w4a4-thinker) below |
 | Qwen3-TTS | TTS language-model stage | Not validated |
 
 Audio encoder, vision encoder, talker, and code2wav stages stay in BF16 unless
@@ -110,9 +102,9 @@ Qwen3-Omni-30B-A3B-Instruct thinker language model. The thinker text body
 (attention + MoE experts) is quantized to NVFP4 with FP8 per-tensor input
 scales; the audio encoder, vision encoder, talker, and code2wav stay in BF16.
 
-| Variant | HF checkpoint | Hardware and validation |
+| Variant | HF checkpoint | Hardware |
-|---------|---------------|-------------------------|
+|---------|---------------|----------|
-| W4A4 NVFP4 (full thinker) | `YihongJin/Qwen3-Omni-30B-A3B-Instruct-NVFP4-W4A4-full-thinker-awqclip` | sm_100+ native FlashInfer FP4 GEMM; H100 Marlin fallback is CI-backed |
+| W4A4 NVFP4 (full thinker) | `YihongJin/Qwen3-Omni-30B-A3B-Instruct-NVFP4-W4A4-full-thinker-awqclip` | sm_100+ (Blackwell, FlashInfer FP4 GEMM) |
 
 Calibration uses ModelOpt `mtq.NVFP4_DEFAULT_CFG` with the `awq_clip` algorithm
 on 1024 ultrachat samples chat-templated through the Qwen3-Omni tokenizer.
@@ -240,6 +232,3 @@ omni = Omni(
 5. Keep `--quantization fp8` for online FP8 from BF16 checkpoints; use this
    ModelOpt path only when the checkpoint already contains ModelOpt quantized
    weights and scales.
-6. The Qwen3-Omni full-model CI smoke exercises the H100 Marlin fallback. A
-   scheduled Blackwell test is still required before treating the native
-   FlashInfer FP4 kernel path as CI-backed.

docs/user_guide/quantization/msmodelslim.md

@@ -3,17 +3,12 @@
 ## Overview
 
 [msModelSlim](https://github.com/Ascend/msmodelslim) is an Ascend compression
-toolkit for producing pre-quantized model checkpoints. vLLM-Omni consumes its
+toolkit for producing pre-quantized model checkpoints. In vLLM-Omni, these
-output either through a compatible generic Ascend checkpoint path or through
+checkpoints run through the Ascend/NPU path with `--quantization ascend`.
-the native MXFP8 and MXFP4 merge workflows documented in this guide.
 
 msModelSlim is static quantization: quantized weights are generated offline
 before vLLM-Omni inference starts.
 
-The validation terms on this page follow the
-[support levels](overview.md#support-levels). A producer recipe or hardware
-backend alone is an integration point, not model-level validation.
-
 ## Hardware Support
 
 | Device | Support |
@@ -25,28 +20,23 @@ backend alone is an integration point, not model-level validation.
 | Intel XPU | ❌ |
 | Ascend NPU | ✅ |
 
-Legend: `✅` backend available, `❌` unsupported, `⭕` not verified in this
+Legend: `✅` supported, `❌` unsupported, `⭕` not verified in this
 guide.
 
 ## Model Type Support
 
 ### Diffusion Model (Qwen-Image, Wan2.2)
 
-| Model | Base model | Scope | Level | Notes |
+| Model | Base model | Scope | Hardware | Notes |
-|-------|------------|-------|-------|-------|
+|-------|------------|-------|----------|-------|
-| Wan2.2 | Wan2.2 diffusion weights | DiT or diffusion stage | Integrated | Upstream msModelSlim provides a producer recipe; use the [MXFP8](mxfp8.md) or [MXFP4](mxfp4.md) workflow for vLLM-Omni's native merged formats |
+| Wan2.2 | Wan2.2 diffusion weights | DiT or diffusion stage | Ascend NPU | Upstream msModelSlim provides a Wan2.2 quantization recipe; vLLM-Omni inference validation is not listed |
-| Qwen-Image | `Qwen/Qwen-Image`, `Qwen/Qwen-Image-2512` | DiT or diffusion stage | Not validated | No supported checkpoint workflow is documented |
+| Qwen-Image | `Qwen/Qwen-Image`, `Qwen/Qwen-Image-2512` | DiT or diffusion stage | Ascend NPU | Not validated in this guide |
-| HunyuanImage-3.0 | `tencent/HunyuanImage-3.0`, `tencent/HunyuanImage-3.0-Instruct` | DiT or diffusion stage | Integrated | Generate quantized weights with the HunyuanImage-3.0 msModelSlim adaptation and validate against BF16 |
+| HunyuanImage-3.0 | `tencent/HunyuanImage-3.0`, `tencent/HunyuanImage-3.0-Instruct` | DiT or diffusion stage | Ascend A2/A3 NPU | Generate quantized weights with the HunyuanImage-3.0 msModelSlim adaptation |
 
 Public Hugging Face quantized weights are not available yet. Use the
 [HunyuanImage-3.0 msModelSlim adaptation](https://gitcode.com/betta18/msmodelslim/tree/hyimage3_mxfp8)
 to generate the checkpoint manually.
 
-msModelSlim is also the producer used by the native offline MXFP8 and MXFP4
-DualScale workflows. Those workflows preprocess and annotate each transformer
-checkpoint and are auto-detected at inference; follow their method pages rather
-than passing `--quantization ascend`.
-
 ### Multi-Stage Omni/TTS Model (Qwen3-Omni, Qwen3-TTS)
 
 | Model | Scope | Status | Notes |
@@ -63,32 +53,25 @@ than passing `--quantization ascend`.
 
 ## Configuration
 
-For a compatible generic Ascend checkpoint, select the Ascend quantization
-method explicitly:
-
 Offline inference:
 
 ```bash
-python text_to_image.py --model <compatible-ascend-checkpoint> --quantization ascend
+python text_to_image.py --model <quantized-model-path> --quantization ascend
 ```
 
 Online serving:
 
 ```bash
-vllm serve <compatible-ascend-checkpoint> --omni --quantization ascend
+vllm serve <quantized-model-path> --omni --quantization ascend
 ```
 
 ## Parameters
 
 | Parameter | Type | Default | Description |
 |-----------|------|---------|-------------|
-| `--quantization` | str | - | Use `ascend` only for a compatible generic Ascend-format checkpoint; omit it for native merged MXFP8/MXFP4 checkpoints |
+| `--quantization` | str | - | Use `ascend` for msModelSlim-produced checkpoints |
 | `model` | str | - | Path to the quantized checkpoint generated by Ascend tooling |
 
-Do not pass `--quantization ascend` for a Wan checkpoint processed by the native
-MXFP8 or MXFP4 merge tools. Their transformer metadata selects `mxfp8` or
-`mxfp4_dualscale` automatically.
-
 Example msModelSlim command for a Wan2.2 W8A8 checkpoint:
 
 ```bash
@@ -110,6 +93,3 @@ For HunyuanImage-3.0, use the Hunyuan-specific adaptation linked above.
    tooling; it is not a load-time CUDA quantizer.
 3. Keep the quantized checkpoint aligned with the same model architecture and
    stage config used for BF16 inference.
-4. No named full-model msModelSlim checkpoint is currently scheduled in NPU
-   CI. Treat new producer outputs as **Integrated** until an end-to-end BF16
-   comparison establishes **Validated** support.

docs/user_guide/quantization/mxfp4.md

@@ -13,10 +13,6 @@ vLLM-Omni provides two quantization methods with different scale structures:
 | `mxfp4` | Single-scale (per-32 fine only) | Online only | Quick accuracy baseline; no checkpoint prep needed |
 | `mxfp4_dualscale` | Dual-scale (fine per-32 + coarse per-512 + per-channel `mul_scale`) | Online + Offline | Production; better accuracy; offline recommended |
 
-The validation terms on this page follow the
-[support levels](overview.md#support-levels). Current model guidance covers
-Wan diffusion transformers only; the text encoder and VAE stay in BF16.
-
 !!! tip "Recommended: `mxfp4_dualscale` offline"
     For production deployments, use the `mxfp4_dualscale` offline mode with a
     pre-quantized checkpoint produced by msModelSlim. Offline checkpoints load
@@ -38,45 +34,26 @@ Wan diffusion transformers only; the text encoder and VAE stay in BF16.
 
 ## Hardware Support
 
-| Device | Online `mxfp4` | Online `mxfp4_dualscale` | Offline `mxfp4_dualscale` |
+| Device | Support |
-|--------|-----------------|---------------------------|----------------------------|
+|--------|---------|
-| NVIDIA Blackwell GPU (SM 100+) | ❌ | ❌ | ❌ |
+| NVIDIA Blackwell GPU (SM 100+) | ⭕ |
-| NVIDIA Ada/Hopper GPU (SM 89+) | ❌ | ❌ | ❌ |
+| NVIDIA Ada/Hopper GPU (SM 89+) | ⭕ |
-| NVIDIA Ampere GPU (SM 80+) | ❌ | ❌ | ❌ |
+| NVIDIA Ampere GPU (SM 80+) | ⭕ |
-| AMD ROCm (gfx950 / MI355X) | ✅ | ❌ | ❌ |
+| AMD ROCm (gfx950 / MI355X) | ✅ |
-| Intel XPU | ❌ | ❌ | ❌ |
+| Intel XPU | ⭕ |
-| Ascend NPU (Atlas 950 A5) | ✅ | ✅ | ✅ |
+| Ascend NPU (Atlas 950 A5) | ✅ |
 
-Legend: `✅` backend available, `❌` unsupported, `⭕` not verified in this
+Legend: `✅` supported, `❌` unsupported, `⭕` not verified in this guide.
-guide. Backend availability does not imply full-model checkpoint validation.
-
-!!! note "Why NVIDIA is marked unsupported"
-    This table describes the methods currently wired into vLLM-Omni, not every
-    low-precision format NVIDIA hardware could execute. `DiffusionMXFP4Config`
-    has runtime implementations only for Ascend and ROCm gfx950; on CUDA it
-    raises `NotImplementedError` because no CUDA MXFP4 linear method or
-    checkpoint loader is registered.
-
-    On NVIDIA, use [ModelOpt](modelopt.md) NVFP4 for the documented 4-bit
-    checkpoint path. NVFP4 and OCP MXFP4 have different checkpoint metadata,
-    scale layouts, and kernel contracts, so they are not interchangeable. A
-    native CUDA MXFP4 path would require a CUDA linear implementation, loader
-    contract, and model-level output validation.
-
-The ROCm path dispatches online single-scale `mxfp4` to AITER on gfx950. It has
-kernel and dispatch coverage, but this guide does not yet claim end-to-end
-quality validation for a named Wan checkpoint on ROCm. DualScale and
-pre-quantized MXFP4 checkpoints are not supported on ROCm.
 
 ## Model Type Support
 
 ### Diffusion Model (Wan2.2)
 
-| Model | Online (Ascend) | Offline (Ascend) | Validation and notes |
+| Model | Online | Offline | Notes |
-|-------|-----------------|------------------|----------------------|
+|-------|--------|---------|-------|
-| Wan2.2-T2V-A14B | `mxfp4` / `mxfp4_dualscale` | `mxfp4_dualscale` | Validated outside scheduled full-model offline NPU CI; both cascade transformers are quantized |
+| Wan2.2-T2V-A14B | `mxfp4` / `mxfp4_dualscale` | `mxfp4_dualscale` | MoE cascade (`transformer` + `transformer_2`); both transformers quantized with the same config |
-| Wan2.2-I2V-A14B | `mxfp4` / `mxfp4_dualscale` | `mxfp4_dualscale` | Validated outside scheduled full-model offline NPU CI; same cascade scheme as T2V-A14B |
+| Wan2.2-I2V-A14B | `mxfp4` / `mxfp4_dualscale` | `mxfp4_dualscale` | MoE cascade; same scheme as T2V-A14B |
-| Wan2.2-TI2V-5B | ❌ | ❌ | Unsupported because W4A4 causes unacceptable accuracy loss |
+| Wan2.2-TI2V-5B | ❌ | ❌ | Parameter count too small; W4A4 causes unacceptable accuracy loss |
 
 The choice between `mxfp4` and `mxfp4_dualscale` in **online mode** is about
 quantization quality, not model compatibility — both work on cascade (A14B) and
@@ -106,13 +83,12 @@ by the merge script); there is no offline `mxfp4` single-scale format.
 
 ### `mxfp4` — Single-Scale Online Mode
 
-Online mode requires no pre-processing. On Ascend, vLLM-Omni quantizes BF16
+Online mode requires no pre-processing. vLLM-Omni quantizes BF16 weights to
-weights at load time using `npu_dynamic_mx_quant`. On ROCm gfx950, the same
+MXFP4 at load time using `npu_dynamic_mx_quant`. A single block scale
-`mxfp4` config dispatches to the AITER online kernel. A single block scale
 (`float8_e8m0fnu`, one per 32 K elements) is computed on the fly; no
-calibration `mul_scale` is available. The config applies equally to
+calibration `mul_scale` is available. Applies equally to single-transformer
-single-transformer and cascade (A14B) models, but named Wan model validation is
+and cascade (A14B) models — both transformers in a cascade receive the same
-currently limited to the documented Ascend path.
+quantization config automatically.
 
 ```python
 from vllm_omni import Omni
@@ -135,10 +111,10 @@ vllm serve <your-model> --omni --quantization mxfp4
 
 ### `mxfp4_dualscale` — DualScale Online Mode
 
-Ascend-only online DualScale mode computes both fine and coarse scales on the
+Online DualScale mode computes both fine and coarse scales on the fly from BF16
-fly from BF16 weights using `npu_dynamic_dual_level_mx_quant`. It applies
+weights using `npu_dynamic_dual_level_mx_quant`. Applies equally to
-equally to single-transformer and cascade (A14B) models. Compared to `mxfp4`
+single-transformer and cascade (A14B) models. Compared to `mxfp4` online,
-online, DualScale provides better quantization accuracy at higher compute cost.
+DualScale provides better quantization accuracy at higher compute cost.
 
 The default configuration keeps the leading 5 transformer blocks in BF16
 (`num_bf16_fallback_layers=5`). Accuracy evaluation on Wan2.2-A14B shows this
@@ -365,11 +341,6 @@ unless they appear in `ignored_layers`.
    quantization. Use `ignored_layers` and `num_bf16_fallback_layers` to trade
    off compression vs. accuracy for precision-sensitive layers.
 
-8. The native methods, merge mappings, and per-transformer cascade metadata
-   have unit coverage. The current Buildkite configuration does not schedule a
-   named full-model offline MXFP4 checkpoint on NPU, so this path is
-   **Validated**, not **CI-backed**.
-
 ## Adapting MXFP4 for a New Model
 
 This section is aimed at developers who want to add MXFP4 support to a model

docs/user_guide/quantization/mxfp8.md

@@ -15,45 +15,30 @@ This method supports three modes:
 | **Offline (Native)** | msModelSlim-exported MXFP8 weights converted to diffusers format via `merge_mxfp8_checkpoint.py` — weights and scales are loaded directly from the preprocessed checkpoint |
 | **Offline (AutoRound)** | AutoRound MXFP8 checkpoints with `data_type="mx_fp"` — auto-detected from `config.json` |
 
-The validation terms on this page follow the
-[support levels](overview.md#support-levels). Current model guidance covers
-Wan diffusion transformers only; the text encoder and VAE stay in BF16.
-
 ## Hardware Support
 
 | Device | Online | Offline (Native) | Offline (AutoRound) |
 |--------|--------|------------------|---------------------|
-| NVIDIA Blackwell GPU (SM 100+) | ❌ | ❌ | ⭕ |
+| NVIDIA Blackwell GPU (SM 100+) | ⭕ | ⭕ | ⭕ |
-| NVIDIA Ada/Hopper GPU (SM 89+) | ❌ | ❌ | ⭕ |
+| NVIDIA Ada/Hopper GPU (SM 89+) | ⭕ | ⭕ | ⭕ |
-| NVIDIA Ampere GPU (SM 80+) | ❌ | ❌ | ⭕ |
+| NVIDIA Ampere GPU (SM 80+) | ⭕ | ⭕ | ⭕ |
-| AMD ROCm | ❌ | ❌ | ⭕ |
+| AMD ROCm | ⭕ | ⭕ | ⭕ |
 | Intel XPU | ✅ | ❌ | ✅ |
 | Ascend NPU (Atlas 950 A5) | ✅ | ✅ | ⭕ |
 
-Legend: `✅` backend available, `❌` unsupported, `⭕` not verified in this
+Legend: `✅` supported, `❌` unsupported, `⭕` not verified in this guide.
-guide. Backend availability does not imply full-model checkpoint validation.
 
-The native `DiffusionMXFP8Config` currently dispatches only to Intel XPU and
+**Note**: Intel XPU only supports AutoRound MXFP8 for offline mode. Use AutoRound quantized checkpoints or online mode for XPU.
-Ascend NPU implementations. CUDA reaches `NotImplementedError`, so the NVIDIA
-rows are marked unsupported even though this is an implementation gap rather
-than a general claim about NVIDIA hardware capabilities. On NVIDIA, use
-[FP8](fp8.md) for load-time quantization or [ModelOpt](modelopt.md) FP8 for
-pre-quantized checkpoints. Adding native CUDA MXFP8 requires a registered CUDA
-linear method, a compatible checkpoint contract, and model-level validation.
-
-Native `mxfp8` is available online on Intel XPU and online or offline on
-Ascend. Intel XPU only supports AutoRound MXFP8 for offline mode. The AutoRound
-column is a separate checkpoint format and compute path.
 
 ## Model Type Support
 
 ### Diffusion Model (Wan2.2)
 
-| Model | Mode | Scope | Validation |
+| Model | Mode | Notes |
-|-------|------|-------|------------|
+|-------|------|-------|
-| Wan2.2-T2V-A14B | Online + Offline | MoE cascade: `transformer` + `transformer_2` | Validated outside scheduled full-model offline NPU CI |
+| Wan2.2-T2V-A14B | Online + Offline | MoE cascade; quantizes two transformers (`transformer` + `transformer_2`) |
-| Wan2.2-I2V-A14B | Online + Offline | MoE cascade: `transformer` + `transformer_2` | Validated outside scheduled full-model offline NPU CI |
+| Wan2.2-I2V-A14B | Online + Offline | MoE cascade; quantizes two transformers (`transformer` + `transformer_2`) |
-| Wan2.2-TI2V-5B | Online + Offline | Single diffusion transformer | Validated outside scheduled full-model offline NPU CI |
+| Wan2.2-TI2V-5B | Online + Offline | Single transformer |
 
 ### Multi-Stage Omni/TTS Model (Qwen3-Omni, Qwen3-TTS)
 
@@ -233,7 +218,3 @@ omni = Omni(model="<autoround-mxfp8-model>")
    `--output-dir`), regenerate it with the current script. The old script
    wrote a separate `quantization_config.json` that is not read by vLLM-Omni;
    the current script injects the config directly into `transformer/config.json`.
-4. The native loader, merge mappings, and Wan cascade config propagation have
-   unit coverage. The current Buildkite configuration does not schedule a
-   named full-model offline MXFP8 checkpoint on NPU, so this path is
-   **Validated**, not **CI-backed**.

docs/user_guide/quantization/online.md

@@ -6,7 +6,7 @@ Online quantization means vLLM-Omni computes quantized weights and scales while
 loading the model. Use it when you want memory savings without preparing a
 separate quantized checkpoint.
 
-This mode is different from pre-quantized checkpoint formats such as ModelOpt,
+This mode is different from pre-quantized checkpoint formats such as GGUF,
 AutoRound, msModelSlim, or serialized Int8 checkpoints. Those formats are
 prepared before serving and are documented in their method-specific guides.
 For MXFP8 and MXFP4, use this page for load-time quantization from BF16
@@ -15,22 +15,18 @@ by msModelSlim and the merge tools.
 
 ## Hardware Support
 
-| Device | FP8 W8A8 | Int8 W8A8 | BitsAndBytes W4 | MXFP8 W8A8 | MXFP4 W4A4 |
+| Device | FP8 W8A8 | Int8 W8A8 | MXFP8 W8A8 | MXFP4 W4A4 |
-|--------|----------|-----------|------------------|------------|------------|
+|--------|----------|-----------|------------|------------|
-| NVIDIA Blackwell GPU (SM 100+) | ✅ | ✅ | ✅ | ❌ | ❌ |
+| NVIDIA Blackwell GPU (SM 100+) | ✅ | ✅ | ⭕ | ⭕ |
-| NVIDIA Ada/Hopper GPU (SM 89+) | ✅ | ✅ | ✅ | ❌ | ❌ |
+| NVIDIA Ada/Hopper GPU (SM 89+) | ✅ | ✅ | ⭕ | ⭕ |
-| NVIDIA Ampere GPU (SM 80+) | ✅ | ✅ | ✅ | ❌ | ❌ |
+| NVIDIA Ampere GPU (SM 80+) | ✅ | ✅ | ⭕ | ⭕ |
-| AMD ROCm (gfx950) | ⭕ | ⭕ | ❌ | ❌ | ✅ |
+| AMD ROCm | ⭕ | ⭕ | ⭕ | ⭕ |
-| Intel XPU | ⭕ | ⭕ | ❌ | ✅ | ❌ |
+| Intel XPU | ⭕ | ⭕ | ✅ | ⭕ |
-| Ascend NPU | ❌ | ✅ | ❌ | ✅ | ✅ |
+| Ascend NPU | ❌ | ✅ | ✅ | ✅ |
 
-Legend: `✅` backend available, `❌` unsupported, `⭕` not verified in this
+Legend: `✅` supported, `❌` unsupported, `⭕` not verified in this
-guide. These entries describe method-level backend availability, not validation
+guide. FP8 on Ampere may use a weight-only path where available. MXFP8 and
-for every model. FP8 on Ampere may use a weight-only path where available.
+MXFP4 are documented for the Ascend NPU path.
-On ROCm gfx950, MXFP4 means online single-scale `mxfp4`; online and offline
-`mxfp4_dualscale` are Ascend-only. The native MXFP8/MXFP4 configs do not
-currently register CUDA implementations; use FP8 or ModelOpt FP8/NVFP4 on
-NVIDIA instead.
 
 ## Model Type Support
 
@@ -40,9 +36,8 @@ NVIDIA instead.
 |--------|-------|----------------|--------|
 | FP8 W8A8 | [FP8](fp8.md) | Qwen-Image; Wan2.2 is not validated | Validated for Qwen-Image family and other DiT models |
 | Int8 W8A8 | [Int8](int8.md) | Qwen-Image; Wan2.2 is not validated | Validated for Qwen-Image and Z-Image |
-| BitsAndBytes W4 | [BitsAndBytes](bitsandbytes.md) | Z-Image-Turbo; Qwen-Image and Wan2.2 are not validated | Validated for Z-Image-Turbo |
 | MXFP8 W8A8 | [MXFP8](mxfp8.md) | Wan2.2-T2V-A14B, Wan2.2-I2V-A14B, Wan2.2-TI2V-5B | Validated on Ascend NPU and Intel XPU |
-| MXFP4 W4A4 | [MXFP4](mxfp4.md) | Wan2.2-T2V-A14B, Wan2.2-I2V-A14B | Validated on Ascend NPU; the ROCm gfx950 backend is integrated but not model-validated; TI2V-5B is unsupported |
+| MXFP4 W4A4 | [MXFP4](mxfp4.md) | Wan2.2-T2V-A14B, Wan2.2-I2V-A14B | Ascend NPU only; TI2V-5B is not supported |
 
 ### Multi-Stage Omni/TTS Model (Qwen3-Omni, Qwen3-TTS)
 
@@ -64,7 +59,6 @@ from vllm_omni import Omni
 
 omni_fp8 = Omni(model="<your-model>", quantization="fp8")
 omni_int8 = Omni(model="<your-model>", quantization="int8")
-omni_bnb = Omni(model="Tongyi-MAI/Z-Image-Turbo", quantization="bitsandbytes")
 omni_mxfp8 = Omni(model="<your-model>", quantization="mxfp8")
 omni_mxfp4 = Omni(model="<your-model>", quantization="mxfp4")
 omni_mxfp4_dualscale = Omni(model="<your-model>", quantization="mxfp4_dualscale")
@@ -75,7 +69,6 @@ CLI:
 ```bash
 vllm serve <your-model> --omni --quantization fp8
 vllm serve <your-model> --omni --quantization int8
-vllm serve Tongyi-MAI/Z-Image-Turbo --omni --quantization bitsandbytes
 vllm serve <your-model> --omni --quantization mxfp8
 vllm serve <your-model> --omni --quantization mxfp4
 vllm serve <your-model> --omni --quantization mxfp4_dualscale
@@ -92,19 +85,14 @@ config = build_quant_config({
 })
 ```
 
-Component routing is an extension mechanism, not automatic support for every
-pipeline component. See [Quantization Scope](overview.md#quantization-scope)
-before applying it to an encoder, VAE, or decoder.
-
 ## Parameters
 
 | Parameter | Methods | Description |
 |-----------|---------|-------------|
-| `method` | FP8, Int8, BitsAndBytes, MXFP8, MXFP4 | Quantization method: `"fp8"`, `"int8"`, `"bitsandbytes"`, `"mxfp8"`, `"mxfp4"`, or `"mxfp4_dualscale"` |
+| `method` | FP8, Int8, MXFP8, MXFP4 | Quantization method: `"fp8"`, `"int8"`, `"mxfp8"`, `"mxfp4"`, or `"mxfp4_dualscale"` |
-| `ignored_layers` | FP8, Int8, BitsAndBytes, MXFP8, MXFP4 | Layer name patterns to keep in BF16/FP16 |
+| `ignored_layers` | FP8, Int8, MXFP8, MXFP4 | Layer name patterns to keep in BF16/FP16 |
 | `activation_scheme` | FP8, Int8 | The runtime value `"dynamic"` selects online activation scaling |
 | `weight_block_size` | FP8 | Optional block-wise FP8 weight quantization size |
-| `quant_type` | BitsAndBytes | Weight format: `"nf4"` (default) or `"fp4"` |
 | `num_bf16_fallback_layers` | MXFP4 DualScale | Leading transformer blocks to keep in BF16 for online `mxfp4_dualscale`; defaults to `5` |
 
 ## Validation and Notes

docs/user_guide/quantization/overview.md

@@ -16,123 +16,72 @@ For the internal architecture and backend extension points, see the
 | 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, AutoRound, msModelSlim, serialized Int8, offline MXFP8, offline MXFP4 DualScale |
 
-Online quantization starts from a normal BF16/FP16 checkpoint and repeats the
+## Hardware Support
-conversion on each fresh model load. It is the simplest path for experiments.
-Pre-quantized checkpoints store the packed weights and scales produced by a
-separate tool; they are easier to reuse across deployments and can carry
-calibrated or mixed-precision policies that load-time conversion cannot
-reproduce.
 
-## Support Levels
+| 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 | ❌ | ✅ | ❌ | ❌ | ✅ | ✅ | ❌ | ✅ |
 
-Quantization support is specific to a model, component, checkpoint format, and
+Legend: `✅` supported, `❌` unsupported, `⭕` not verified in this
-hardware backend. The tables in this guide use the following levels:
+guide. FP8 on Ampere may use a weight-only path where available.
 
-| Level | Meaning |
+## Model Type Support
-|-------|---------|
-| **CI-backed** | A named model and quantized checkpoint are exercised by scheduled full-model hardware CI. |
-| **Validated** | A named model and checkpoint have end-to-end or quality-validation evidence, but the check is not necessarily scheduled in CI. |
-| **Integrated** | The config, loader, quantized layers, or model wiring exists, but this guide has no completed end-to-end validation for the named model and checkpoint. |
-| **Not validated** | There is no current support recommendation for this model/component combination. |
-| **Unsupported** | The combination is intentionally rejected or known not to work. |
 
-A method being available on a device does not make every model on that device
+### Diffusion Model (Qwen-Image, Wan2.2)
-supported. Likewise, successfully detecting a checkpoint's
-`quantization_config` establishes loader compatibility, not output quality.
 
-## Hardware × Quantization Method
+These models run a diffusion transformer as the primary inference module. The
+default quantization target is the transformer; tokenizer, scheduler, text
+encoder, and VAE stay on the base checkpoint unless a method guide says
+otherwise.
 
-This matrix uses concrete inference paths as columns. Producer toolkits are not
+| Method | Guide | Mode | Example models | Status |
-separate methods: NVIDIA ModelOpt produces the ModelOpt FP8/NVFP4 formats,
+|--------|-------|------|----------------|--------|
-Intel AutoRound produces AutoRound checkpoints, and Ascend msModelSlim produces
+| 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 |
-checkpoints consumed by the native MXFP or compatible Ascend paths.
+| 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 |
+| 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 |
 
-| Hardware | [Online FP8](fp8.md) | [Int8](int8.md) | [BitsAndBytes W4](bitsandbytes.md) | [ModelOpt FP8](modelopt.md) | [ModelOpt NVFP4](modelopt.md#supported-modelopt-checkpoint-formats) | [AutoRound W4A16](autoround.md) | [OCP MXFP8](mxfp8.md) | [OCP MXFP4](mxfp4.md) |
+### Multi-Stage Omni/TTS Model (Qwen3-Omni, Qwen3-TTS)
-|----------|-----------------------|------------------|------------------------------------------|--------------------------------|-------------------------------------------------------------------------|--------------------------------------|--------------------------|--------------------------|
-| NVIDIA Blackwell (SM 100+) | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
-| NVIDIA Hopper (SM 90) | ✅ | ✅ | ✅ | ✅ | ✅ (Marlin fallback) | ✅ | ❌ | ❌ |
-| NVIDIA Ada (SM 89) | ✅ | ✅ | ✅ | ✅ | ⭕ | ✅ | ❌ | ❌ |
-| NVIDIA Ampere (SM 80+) | ✅ | ✅ | ✅ | ⭕ | ⭕ | ✅ | ❌ | ❌ |
-| AMD ROCm (gfx950 for MXFP4) | ⭕ | ⭕ | ❌ | ⭕ | ⭕ | ⭕ | ❌ | ✅ (online only) |
-| Intel XPU | ⭕ | ⭕ | ❌ | ⭕ | ⭕ | ✅ | ✅ | ❌ |
-| Ascend NPU | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | ✅ | ✅ |
 
-Legend: `✅` backend available, `❌` rejected or no implementation, `⭕` not
+These models combine an AR language model with audio, vision, talker, or TTS
-verified in this guide. This matrix does not establish model-level support; use
+stages. Quantization is scoped to the AR language-model stage when the
-the model-and-stage tables below for that.
+checkpoint contains a supported `quantization_config`; the non-AR stages stay
+in BF16 unless the model guide explicitly adds support.
 
-The Int8 column covers online and serialized Int8 paths. OCP MXFP8 is online on
+| Method | Guide | Scope | Example models | Status |
-Intel XPU and online/offline on Ascend; Intel's separate AutoRound MXFP8
+|--------|-------|-------|----------------|--------|
-checkpoint path is described in the [MXFP8 guide](mxfp8.md). OCP MXFP4 on AMD
+| ModelOpt | [ModelOpt](modelopt.md) | Thinker or language-model checkpoint config | Qwen3-Omni thinker | ModelOpt checkpoint path |
-means online single-scale `mxfp4` on gfx950, while MXFP4 DualScale is
+| Int8 | [Int8](int8.md) | Not currently validated for omni/TTS stages | Qwen3-Omni, Qwen3-TTS | Not validated |
-Ascend-only. NVIDIA's NVFP4 and OCP MXFP4 have different checkpoint metadata,
+| BitsAndBytes | [BitsAndBytes](bitsandbytes.md) | Not currently validated for omni/TTS stages | Qwen3-Omni, Qwen3-TTS | Not validated |
-scale layouts, and kernel contracts. Mixed ModelOpt FP8/NVFP4 checkpoints also
+| MXFP8 | [MXFP8](mxfp8.md) | Not currently validated for omni/TTS stages | Qwen3-Omni, Qwen3-TTS | Not validated |
-require a model-specific mixed-precision policy and validation. FP8 on Ampere
+| MXFP4 | [MXFP4](mxfp4.md) | Not currently validated for omni/TTS stages | Qwen3-Omni, Qwen3-TTS | Not validated |
-may use a weight-only path where available.
+| 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 |
 
-## Current Model and Stage Support
+### Multi-Stage Diffusion Model (BAGEL, GLM-Image)
 
-The current offline checkpoint guidance is transformer-first. A listed
+These models split generation across multiple stages. Quantization must be
-checkpoint quantizes the stage named in the **Scope** column; it does not imply
+attached to the intended stage rather than applied globally.
-that the model's encoders, VAE, decoder, or other stages are also quantized.
 
-### Vendor and Runtime Ecosystems
+| Method | Guide | Scope | Example models | Status |
-
+|--------|-------|-------|----------------|--------|
-The producer and runtime backend are related, but they are not the same support
+| FP8 | [FP8](fp8.md) | Stage-specific DiT or transformer module | BAGEL, GLM-Image | Requires model-specific validation |
-claim. A producer can emit several checkpoint formats, and each format still
+| Int8 | [Int8](int8.md) | Stage-specific DiT or transformer module | BAGEL, GLM-Image | Requires model-specific validation |
-needs a compatible vLLM-Omni loader, model-stage wiring, and validation. A
+| BitsAndBytes | [BitsAndBytes](bitsandbytes.md) | Stage-specific transformer or DiT module | BAGEL, GLM-Image | Not validated |
-runtime may also provide online quantization without an offline producer path.
+| 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 |
-| Ecosystem | Current paths | Current vLLM-Omni scope | Model evidence |
+| MXFP4 | [MXFP4](mxfp4.md) | Stage-specific DiT or transformer module | BAGEL, GLM-Image | Not validated |
-|-----------|---------------|-------------------------|----------------|
+| AutoRound | [AutoRound](autoround.md) | Checkpoint-defined stage | BAGEL, GLM-Image | No validated checkpoint listed |
-| NVIDIA | ModelOpt FP8, NVFP4, and mixed FP8/NVFP4 | Diffusion transformer or Qwen-Omni thinker | Qwen-Image and other named diffusion checkpoints are validated; Qwen3-Omni NVFP4 has H100 CI for the Marlin fallback |
+| msModelSlim | [msModelSlim](msmodelslim.md) | Ascend-generated stage weights | GLM-Image | Requires model-specific adaptation |
-| Intel | AutoRound W4A16 and AutoRound MXFP8 | Diffusion/world-model transformer or Qwen-Omni thinker | Wan A14B and GLM-Image W4A16 are CI-backed; Qwen-Image, Qwen-Omni, and Cosmos3-Super have validation outside scheduled CI |
-| AMD ROCm | Online OCP MXFP4 through AITER on gfx950; no pre-quantized MXFP4 or AMD Quark workflow is integrated | Diffusion transformer linear layers | Integrated backend with dispatch and weight-allocation unit coverage; no named-model quality validation or scheduled gfx950 full-model CI |
-| Ascend | msModelSlim output through native MXFP8/MXFP4 merge workflows or a compatible generic Ascend format | Wan diffusion transformers | Native Wan MXFP8/MXFP4 paths are validated outside scheduled full-model NPU CI; other model families remain integrated or unvalidated |
-
-### Pre-Quantized Model Matrix
-
-| Model family | Method | Scope | Level | Notes |
-|--------------|--------|-------|-------|-------|
-| Qwen-Image | [ModelOpt](modelopt.md) FP8 or mixed FP8/NVFP4 | Diffusion transformer | Validated | Named Qwen-Image 2512 checkpoints and recipes are available. Text encoder and VAE stay BF16. |
-| Qwen-Image | [AutoRound](autoround.md) W4A16 | Diffusion transformer | Validated | Uses `INC4AI/Qwen-Image-AutoRound-W4A16`; text encoder and VAE stay BF16. |
-| Wan2.2 I2V/T2V A14B | [AutoRound](autoround.md) W4A16 | Both diffusion transformers | CI-backed | The high-noise and low-noise DiTs are quantized; text encoder and VAE stay BF16. |
-| Wan2.2 TI2V 5B | [AutoRound](autoround.md) W4A16 | Diffusion transformer | Validated | Named checkpoint available; not covered by the scheduled Wan AutoRound job. |
-| Wan2.2 | [MXFP8](mxfp8.md) | One or both diffusion transformers | Validated | Native Ascend and AutoRound checkpoint formats are integrated; no scheduled full-model offline NPU job. |
-| Wan2.2 I2V/T2V A14B | [MXFP4 DualScale](mxfp4.md) | Both diffusion transformers | Validated | Offline calibrated checkpoints are recommended; no scheduled full-model offline NPU job. |
-| Qwen2.5-Omni | [AutoRound](autoround.md) W4A16 | Thinker language model | Validated | Encoders and output/audio stages stay BF16. |
-| Qwen3-Omni | [ModelOpt](modelopt.md) FP8 | Thinker language model | Validated | Tested outside scheduled CI. |
-| Qwen3-Omni | [ModelOpt](modelopt.md) NVFP4 | Thinker language model | CI-backed | H100 CI exercises the Marlin fallback; the native Blackwell FP4 path has separate recipe validation. |
-| Qwen3-Omni | [AutoRound](autoround.md) W4A16 | Thinker language model | Validated | Audio encoder, vision encoder, talker, and Code2Wav stay BF16. |
-| GLM-Image | [AutoRound](autoround.md) W4A16 | Diffusion transformer stage | CI-backed | Text-to-image and image-to-image full-model smoke coverage. |
-| Cosmos3-Super | [AutoRound](autoround.md) W4A16 | World-model transformer | Validated | Manually validated serving path; VAE and other components stay BF16. |
-| Cosmos3-Nano | [AutoRound](autoround.md) W4A16 | World-model transformer | Integrated | Named checkpoint is listed, but no model-specific end-to-end test is maintained in-tree. |
-| MiniMax H3 | None | None | Not validated | No offline vendor checkpoint path is documented; online FP8 targets the DiT only. |
-
-The method pages contain the complete checkpoint lists, including FLUX,
-Z-Image, and HunyuanImage-3.0 variants.
-
-### Online Quantization
-
-| Model family | Methods | Scope | Level |
-|--------------|---------|-------|-------|
-| Qwen-Image | FP8, Int8 | Diffusion transformer | Validated |
-| Z-Image | Int8, BitsAndBytes W4 | Diffusion transformer | Validated |
-| Wan2.2 | MXFP8 | One or both diffusion transformers | Validated on the documented XPU/NPU paths |
-| Wan2.2 A14B | MXFP4 | Both diffusion transformers | Validated on Ascend; the ROCm gfx950 online backend is integrated but not model-validated |
-| MiniMax H3 | FP8 | DiT and optional reference DiT | Validated |
-| Qwen3-Omni | Generic online methods, such as FP8 | Thinker language model only | Integrated; not model-validated as an online recommendation |
-
-See [Online Quantization](online.md) for configuration and hardware details.
-
-### Unlisted and Long-Tail Models
-
-For models not listed above, generic quantized linear layers or compatible
-checkpoint metadata are only an integration starting point. The default support
-target is the model's primary transformer, DiT, or AR language-model stage.
-Text/audio/vision encoders, VAEs, vocoders, decoders, and output stages remain
-BF16 until a model-specific integration supplies quant-aware layers, checkpoint
-mapping, end-to-end output validation, and backend coverage.
 
 !!! note
     "Online quantization" means vLLM-Omni computes the quantization data while
@@ -143,8 +92,8 @@ mapping, end-to-end output validation, and backend coverage.
 
 ### Diffusion Model (Qwen-Image, Wan2.2)
 
-The default target is the diffusion transformer. `build_quant_config()` can
+The default target is the diffusion transformer. Component routing is available
-construct a component router:
+through `build_quant_config()`:
 
 ```python
 from vllm_omni.quantization import build_quant_config
@@ -155,19 +104,11 @@ config = build_quant_config({
 })
 ```
 
-!!! warning "Routing syntax is not a support claim"
-    A component dictionary only selects a quantization config by module prefix.
-    It does not make an encoder or VAE quantizable. The model must construct
-    that component from quantization-aware layers, and an offline checkpoint
-    must contain matching packed weights and scales. Current user guidance
-    keeps diffusion text encoders and VAEs in BF16.
-
 | Component | Default quantized? | Notes |
 |-----------|--------------------|-------|
-| Diffusion transformer | Yes, for listed model/method pairs | Primary target for FP8, Int8, BitsAndBytes, ModelOpt, MXFP8, MXFP4, AutoRound, and msModelSlim |
+| Diffusion transformer | Yes | Primary target for FP8, Int8, BitsAndBytes, ModelOpt, MXFP8, MXFP4, AutoRound, and msModelSlim |
-| Second diffusion transformer | Yes, for listed Wan2.2 A14B methods | `transformer` and `transformer_2` read their own offline checkpoint metadata; arbitrary mixed methods are not currently validated |
 | Text encoder | No | Keep BF16 unless a method-specific guide documents support |
-| VAE | No | Keep BF16 |
+| VAE | No | Keep BF16; storage-only paths are method-specific |
 | Scheduler/tokenizer | No | Loaded from the base model repository |
 
 ### Multi-Stage Omni/TTS Model (Qwen3-Omni, Qwen3-TTS)
@@ -180,18 +121,6 @@ config = build_quant_config({
 | Talker or TTS stage | No | BF16 unless model-specific support is documented |
 | Code2Wav | No | BF16 |
 
-An explicit `ComponentQuantizationConfig` can be used by model developers to
-route different configs to `audio_tower`, `visual`, and `language_model`, but
-the supported Qwen-Omni checkpoints currently quantize only the thinker
-language model.
-
-### World Model (Cosmos3)
-
-| Component | Default quantized? | Notes |
-|-----------|--------------------|-------|
-| World-model transformer | AutoRound checkpoints only | Cosmos3-Super is validated; Cosmos3-Nano is integrated but not end-to-end tested in-tree |
-| VAE and guardrail components | No | BF16 |
-
 ### Multi-Stage Diffusion Model (BAGEL, GLM-Image)
 
 | Component | Default quantized? | Notes |
@@ -212,7 +141,6 @@ 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)
-# Component routing syntax; model-specific support is still required.
 build_quant_config({"transformer": {"method": "fp8"}, "vae": None})
 build_quant_config(None)
 ```

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,
@@ -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",

examples/offline_inference/dynin_omni/end2end.py

@@ -876,7 +876,7 @@ def parse_args(repo_root: Path) -> argparse.Namespace:
     parser.add_argument("--task", type=str, required=True, choices=TASK_CHOICES)
     parser.add_argument("--model", type=str, required=True, help="HF repo id or local model directory.")
     parser.add_argument(
-        "--stage-config-path",
+        "--deploy-config",
         type=str,
         default=str(repo_root / "vllm_omni/deploy/dynin_omni.yaml"),
         help="Path to deploy config yaml.",
@@ -1394,8 +1394,8 @@ def main() -> None:
 
     from vllm_omni.entrypoints.omni import Omni
 
-    stage_config_path = str(Path(args.stage_config_path).expanduser())
+    deploy_config_path = str(Path(args.deploy_config).expanduser())
-    omni = Omni(model=model_source, stage_configs_path=stage_config_path, dtype=args.dtype)
+    omni = Omni(model=model_source, deploy_config=deploy_config_path, dtype=args.dtype)
     sampling_params_list = [
         SamplingParams(max_tokens=int(args.max_tokens_per_stage), temperature=0.0, top_p=1.0, detokenize=False)
         for _ in range(omni.num_stages)

examples/offline_inference/hunyuan_image3/end2end.py

@@ -73,7 +73,6 @@ def parse_args():
     )
     parser.add_argument("--sys-type", type=str, default=None, help="Override system prompt type.")
     parser.add_argument("--deploy-config", type=str, default=None, help="Custom deploy YAML path.")
-    parser.add_argument("--stage-configs-path", type=str, default=None, help="Custom legacy stage config YAML path.")
     parser.add_argument(
         "--stream",
         action="store_true",
@@ -146,13 +145,7 @@ def main():
     else:
         bot_task = args.bot_task
 
-    if args.deploy_config is not None and args.stage_configs_path is not None:
+    deploy_config = args.deploy_config or _MODALITY_DEFAULT_DEPLOY_CONFIG[args.modality]
-        raise ValueError("--deploy-config and --stage-configs-path are mutually exclusive.")
-
-    deploy_config = args.deploy_config
-    stage_configs_path = args.stage_configs_path
-    if deploy_config is None and stage_configs_path is None:
-        deploy_config = _MODALITY_DEFAULT_DEPLOY_CONFIG[args.modality]
 
     omni_kwargs = {
         "model": args.model,
@@ -168,10 +161,7 @@ def main():
 
     if additional_config is not None:
         omni_kwargs["additional_config"] = additional_config
-    if deploy_config is not None:
+    omni_kwargs["deploy_config"] = deploy_config
-        omni_kwargs["deploy_config"] = deploy_config
-    else:
-        omni_kwargs["stage_configs_path"] = stage_configs_path
 
     omni = Omni(**omni_kwargs)
 
@@ -272,10 +262,7 @@ def main():
     print(f"  Modality: {args.modality}")
     print(f"  Prompt task: {task}")
     print(f"  Bot task: {bot_task}")
-    if deploy_config is not None:
+    print(f"  Deploy config: {deploy_config}")
-        print(f"  Deploy config: {deploy_config}")
-    else:
-        print(f"  Stage config: {stage_configs_path}")
     print(f"  Num stages: {omni.num_stages}")
     if args.modality in ("text2img", "img2img"):
         print(f"  Inference steps: {args.steps}")

examples/offline_inference/minicpmo/end2end.py

@@ -452,12 +452,6 @@ def parse_args():
         default=None,
         help="Path to a .txt file with one prompt per line (preferred).",
     )
-    parser.add_argument(
-        "--stage-configs-path",
-        type=str,
-        default=None,
-        help="Path to a stage configs file (deprecated; prefer --deploy-config).",
-    )
     parser.add_argument(
         "--deploy-config",
         type=str,

examples/offline_inference/qwen3_omni/end2end.py

@@ -484,10 +484,10 @@ def parse_args():
         help="Path to a .txt file with one prompt per line (preferred).",
     )
     parser.add_argument(
-        "--stage-configs-path",
+        "--deploy-config",
         type=str,
         default=None,
-        help="Path to a stage configs file.",
+        help="Path to a deploy config YAML.",
     )
     parser.add_argument(
         "--video-path",

examples/offline_inference/step_audio2/end2end.py

@@ -232,15 +232,8 @@ def main(args):
         raise ValueError(f"Unknown query type: {args.query_type}")
 
     # Initialize vLLM-Omni with Step-Audio2
-    # Resolve deploy config path. Keep explicit stage_configs_path for legacy
+    # Resolve an explicit deploy config or use the bundled task default.
-    # custom configs, but use bundled deploy configs by default.
+    if args.deploy_config:
-    deploy_config_path = None
-    stage_config_path = None
-    if args.stage_configs_path and args.deploy_config:
-        raise ValueError("--stage-configs-path and --deploy-config are mutually exclusive")
-    if args.stage_configs_path:
-        stage_config_path = args.stage_configs_path
-    elif args.deploy_config:
         deploy_config_path = args.deploy_config
     else:
         configs_dir = Path(__file__).parent.parent.parent.parent / "vllm_omni/deploy"
@@ -262,10 +255,7 @@ def main(args):
         "ray_address": args.ray_address,
         "trust_remote_code": True,
     }
-    if stage_config_path is not None:
+    omni_kwargs["deploy_config"] = deploy_config_path
-        omni_kwargs["stage_configs_path"] = stage_config_path
-    else:
-        omni_kwargs["deploy_config"] = deploy_config_path
 
     omni_llm = Omni(**omni_kwargs)
 
@@ -418,12 +408,6 @@ def parse_args():
             "  - HuggingFace ID: stepfun-ai/Step-Audio-2-mini"
         ),
     )
-    parser.add_argument(
-        "--stage-configs-path",
-        type=str,
-        default=None,
-        help="Path to a legacy stage config YAML file (default: use bundled deploy config)",
-    )
     parser.add_argument(
         "--deploy-config",
         type=str,

examples/offline_inference/text_to_image/text_to_image.py

@@ -92,12 +92,6 @@ def parse_args() -> argparse.Namespace:
         "meituan-longcat/LongCat-Image, OvisAI/Ovis-Image, "
         "stabilityai/stable-diffusion-3.5-medium, Tongyi-MAI/Z-Image-Turbo and etc.",
     )
-    parser.add_argument(
-        "--stage-configs-path",
-        type=str,
-        default=None,
-        help="[Deprecated] Path to a legacy stage_args-format YAML. Prefer --deploy-config.",
-    )
     parser.add_argument(
         "--deploy-config",
         type=str,
@@ -475,8 +469,6 @@ def main():
         omni_kwargs["tensor_parallel_size"] = args.tensor_parallel_size
     if args.enforce_eager is not None:
         omni_kwargs["enforce_eager"] = args.enforce_eager
-    if args.stage_configs_path:
-        omni_kwargs["stage_configs_path"] = args.stage_configs_path
     if args.deploy_config:
         omni_kwargs["deploy_config"] = args.deploy_config
     if use_nextstep:
@@ -515,8 +507,8 @@ def main():
     print(f"  Image size: {args.width}x{args.height}")
     if args.lora_path:
         print(f"  LoRA: scale={args.lora_scale}")
-    if args.stage_configs_path:
+    if args.deploy_config:
-        print(f"  stage-configs-path: {args.stage_configs_path}")
+        print(f"  deploy-config: {args.deploy_config}")
     print(f"{'=' * 60}\n")
 
     # Build LoRA request when --lora-path is set

examples/offline_inference/text_to_speech/README.md

@@ -23,6 +23,7 @@ list of supported architectures across all modalities, see
 | OmniVoice | `k2-fsa/OmniVoice` | 2 (gen + dec) | ✓ | — | voice design, language hint | 24 kHz |
 | Qwen3-TTS | `Qwen/Qwen3-TTS-12Hz-1.7B-{CustomVoice,VoiceDesign,Base}` | 2 (talker + code2wav) | ✓ (Base) | ✓ | 3 task variants | 24 kHz |
 | VoxCPM2 | `openbmb/VoxCPM2` | single (native AR) | ✓ | ✓ (online) | continuation | 48 kHz |
+| dots.tts | `rednote-hilab/dots.tts-soar` | single (native AR) | — (not wired yet) | — | — | 48 kHz |
 | IndexTTS-2 | `IndexTeam/IndexTTS-2` | 2 (AR talker + S2Mel DiT + BigVGAN) | ✓ (required) | — | emotion control (`--emo-audio`, `--emo-text`, `--emo-vector`) | 22.05 kHz |
 | IndexTTS-2.5 | native `checkpoints/` bundle | 2 (AR talker + EnhancedCodec + S2Mel DiT + BigVGAN) | ✓ (required) | — | multilingual (`--lang`) + emotion control | 22.05 kHz |
 | Voxtral TTS | `mistralai/Voxtral-4B-TTS-2603` | varies | ✓ | ✓ | voice presets | 24 kHz |
@@ -430,6 +431,31 @@ Streaming is exposed through the online OpenAI Speech API (`stream=true`). See [
 
 ---
 
+## dots.tts
+
+Single-stage native AR TTS at 48 kHz (rednote-hilab). Pipeline: `Qwen2.5-1.5B base LM → DiT (10-step Euler flow matching) → patch_encoder AR loopback → AudioVAE (streaming decode)`. Same "vLLM-native base LM + side-path computation" pattern as VoxCPM2, with a plain Qwen2 backbone instead of MiniCPM4 and no FSQ / residual LM stage.
+
+### Quick start
+```bash
+python examples/offline_inference/text_to_speech/dots_tts/end2end.py \
+    --model rednote-hilab/dots.tts-soar \
+    --text "Hello, this is a test of dots TTS running on vLLM Omni."
+```
+
+### Voice cloning
+Not wired in this release — generation is zero-shot only. The CAM++ x-vector speaker encoder weights load, but `end2end.py` has no `--ref-audio`/`--ref-text` flags yet.
+
+### Streaming
+The AudioVAE decoder has an internal streaming path (`init_stream_state` / `stream_step` / `stream_flush`) used to avoid boundary artifacts between 160 ms patches, but it is not yet exposed through an online serving endpoint or example.
+
+### Notes
+- Output: 48 kHz mono WAV.
+- Deploy config: `vllm_omni/deploy/dots_tts.yaml` (auto-loaded by HF `model_type`).
+- Checkpoints: `rednote-hilab/dots.tts-soar` is the validated default. `dots.tts-base` shares the same architecture but is unvalidated in this repo. `dots.tts-mf` (MeanFlow, 2-4 step) is not supported yet.
+- Known limitation: no CUDA graph capture and no batched side-path computation yet, so concurrent requests do not currently scale (each request's DiT Euler steps run serially). See `recipes/rednote-hilab/dots.tts.md` for details and the roadmap.
+
+---
+
 ## IndexTTS-2 and IndexTTS-2.5
 
 Both versions use a 2-stage TTS pipeline and produce 22.05 kHz mono audio. Stage 0 is a GPT AR talker. IndexTTS-2 Stage 1 uses RepCodec semantic embeddings, GPT latent, S2Mel CFM/DiT, and BigVGAN. IndexTTS-2.5 replaces RepCodec with EnhancedCodec and uses the official code-only path (`use_gpt_latent=false`). Every request requires reference audio for zero-shot voice cloning.

examples/offline_inference/text_to_speech/dots_tts/end2end.py

@@ -0,0 +1,125 @@
+"""Offline dots.tts inference example (single-stage native AR pipeline).
+
+Uses the single-stage native AR config (vllm_omni/deploy/dots_tts.yaml).
+Reference checkpoint: rednote-hilab/dots.tts-soar.
+"""
+
+from __future__ import annotations
+
+import os
+import time
+from pathlib import Path
+
+import soundfile as sf
+import torch
+
+from vllm_omni import Omni
+from vllm_omni.utils.tracking_parser import TrackingArgumentParser
+
+REPO_ROOT = Path(__file__).resolve().parents[4]
+SAMPLE_RATE = 48_000
+
+
+def parse_args():
+    parser = TrackingArgumentParser(description="Offline dots.tts native AR inference")
+    parser.add_argument(
+        "--model",
+        type=str,
+        default="rednote-hilab/dots.tts-soar",
+        help="dots.tts model path or HuggingFace repo ID.",
+    )
+    parser.add_argument(
+        "--text",
+        type=str,
+        default="Hello, this is a test of dots TTS running on vLLM Omni.",
+        help="Text to synthesize.",
+    )
+    parser.add_argument(
+        "--output-dir",
+        type=str,
+        default="output_audio",
+        help="Directory for output WAV files.",
+    )
+    parser.add_argument(
+        "--deploy-config",
+        type=str,
+        default=None,
+        help="Override the deploy config path. If unset, auto-loads "
+        "vllm_omni/deploy/dots_tts.yaml based on the HF model_type.",
+    )
+    return parser.parse_args()
+
+
+def extract_audio(multimodal_output: dict) -> torch.Tensor:
+    """Extract the final complete audio tensor from multimodal output.
+
+    The output processor concatenates per-step delta tensors under
+    ``model_outputs``.  Falls back to ``audio`` for backwards compat.
+    """
+    audio = multimodal_output.get("model_outputs")
+    if audio is None:
+        audio = multimodal_output.get("audio")
+    if audio is None:
+        raise ValueError(f"No audio key in multimodal_output: {list(multimodal_output.keys())}")
+
+    if isinstance(audio, list):
+        valid = [torch.as_tensor(a).float().cpu().reshape(-1) for a in audio if a is not None]
+        if not valid:
+            raise ValueError("Audio list is empty or all elements are None.")
+        return torch.cat(valid, dim=0) if len(valid) > 1 else valid[0]
+
+    return torch.as_tensor(audio).float().cpu().reshape(-1)
+
+
+def main():
+    args = parse_args()
+
+    output_dir = Path(args.output_dir)
+    output_dir.mkdir(parents=True, exist_ok=True)
+
+    engine = Omni(
+        model=args.model,
+        deploy_config=args.deploy_config,
+    )
+
+    from transformers import AutoTokenizer
+
+    from vllm_omni.model_executor.models.dots_tts.dots_tts_prompt import (
+        build_dots_tts_prompt,
+    )
+
+    tokenizer = AutoTokenizer.from_pretrained(args.model, trust_remote_code=True)
+
+    prompt = build_dots_tts_prompt(
+        tokenizer=tokenizer,
+        text=args.text,
+    )
+
+    print(f"Model       : {args.model}")
+    print(f"Text        : {args.text}")
+    print(f"Output dir  : {output_dir}")
+    print(f"Prompt len  : {len(prompt['prompt_token_ids'])} tokens")
+
+    t_start = time.perf_counter()
+    outputs = engine.generate([prompt])
+    elapsed = time.perf_counter() - t_start
+
+    request_output = outputs[0]
+    mm = request_output.outputs[0].multimodal_output
+    audio = extract_audio(mm)
+
+    duration = audio.numel() / SAMPLE_RATE
+    rtf = elapsed / duration if duration > 0 else float("inf")
+
+    output_path = output_dir / "output.wav"
+    sf.write(str(output_path), audio.numpy(), SAMPLE_RATE, format="WAV")
+
+    print(f"Saved       : {output_path}")
+    print(f"Duration    : {duration:.2f}s")
+    print(f"Inference   : {elapsed:.2f}s")
+    print(f"RTF         : {rtf:.3f}")
+
+
+if __name__ == "__main__":
+    os.environ["VLLM_WORKER_MULTIPROC_METHOD"] = "spawn"
+    main()

examples/offline_inference/text_to_speech/fish_speech/end2end.py

@@ -38,7 +38,7 @@ from vllm_omni.utils.tracking_parser import TrackingArgumentParser
 logger = logging.getLogger(__name__)
 
 DEFAULT_MODEL = "fishaudio/s2-pro"
-DEFAULT_STAGE_CONFIG = os.path.join(
+DEFAULT_DEPLOY_CONFIG = os.path.join(
     os.path.dirname(__file__),
     "..",
     "..",
@@ -152,10 +152,7 @@ def _build_omni_kwargs(args, model_name: str) -> dict:
         log_stats=args.log_stats,
         stage_init_timeout=args.stage_init_timeout,
     )
-    if args.deploy_config:
+    omni_kwargs["deploy_config"] = args.deploy_config or DEFAULT_DEPLOY_CONFIG
-        omni_kwargs["deploy_config"] = args.deploy_config
-    else:
-        omni_kwargs["stage_configs_path"] = args.stage_configs_path or DEFAULT_STAGE_CONFIG
     return omni_kwargs
 
 
@@ -278,12 +275,6 @@ def parse_args():
         default=None,
         help="Deploy config YAML path (under vllm_omni/deploy/).",
     )
-    parser.add_argument(
-        "--stage-configs-path",
-        type=str,
-        default=None,
-        help="Path to stage configs YAML (legacy, prefer --deploy-config).",
-    )
     parser.add_argument(
         "--output-dir",
         default="output_audio",

examples/offline_inference/text_to_speech/glm_tts/end2end.py

@@ -145,7 +145,7 @@ def main(args):
 
     omni = Omni(
         model=args.model,
-        stage_configs_path=deploy_config_path,
+        deploy_config=deploy_config_path,
         log_stats=args.log_stats,
         stage_init_timeout=args.stage_init_timeout,
     )

examples/offline_inference/text_to_speech/qwen3_tts/end2end.py

@@ -490,10 +490,10 @@ def parse_args():
         help="Path to a .txt file with one prompt per line (preferred).",
     )
     parser.add_argument(
-        "--stage-configs-path",
+        "--deploy-config",
         type=str,
         default=None,
-        help="Path to a stage configs file.",
+        help="Path to a deploy config YAML.",
     )
     parser.add_argument(
         "--audio-path",

mkdocs.yml

@@ -73,12 +73,18 @@ plugins:
   - awesome-nav
   - redirects:
       redirect_maps:
+        community/meetups.md: community/meetings.md
+        community/volunteers.md: community/contact_us.md
         user_guide/diffusion/request_batching.md: user_guide/diffusion/execution_modes.md
         user_guide/diffusion/step_execution.md: user_guide/diffusion/execution_modes.md
         user_guide/diffusion/cpu_offload_diffusion.md: user_guide/diffusion/cpu_offload.md
         user_guide/quantization/index.md: user_guide/quantization/overview.md
         design/feature/diffusion_request_level_batching.md: design/feature/diffusion_continuous_batching.md
         design/feature/diffusion_step_execution.md: design/feature/diffusion_continuous_batching.md
+        design/feature/cpu_offloading.md: design/feature/offloader/README.md
+        design/feature/model_level_offload.md: design/feature/offloader/cpu_offload.md
+        design/feature/layerwise_offload.md: design/feature/offloader/layerwise_offload.md
+        design/feature/distributed_layerwise_offload.md: design/feature/offloader/distributed_layerwise_offload.md
   - glightbox
   - git-revision-date-localized:
       # exclude files

pyproject.toml

@@ -40,6 +40,7 @@ dev = [
     "pytest-asyncio==1.4.0",
     "pytest-cov==7.1.0",
     "pytest-mock==3.15.1",
+    "pytest-shard==0.1.2",
     "websockets==16.1",
     "pytest-subtests==0.15.0",
     "datasets==3.6.0",
@@ -301,6 +302,21 @@ filterwarnings = [
     "ignore:.*does not have '__bases__' attribute.*:UserWarning",
 ]
 
+[tool.coverage.run]
+core = "sysmon"
+source = ["vllm_omni"]
+# Online-serving tests launch the server via subprocess.Popen and stop it with
+# SIGTERM. pytest-cov no longer instruments subprocesses implicitly, and a
+# SIGTERM-ed process discards its data, so without patch/sigterm a coverage run
+# over an online test measures only the pytest parent.
+patch = ["subprocess"]
+concurrency = ["multiprocessing", "thread"]
+sigterm = true
+
+[tool.coverage.report]
+show_missing = true
+skip_covered = true
+
 [tool.typos.default]
 extend-ignore-identifiers-re = [
     ".*_thw",

recipes/LTX/LTX-2.md

@@ -4,11 +4,20 @@
 
 ## Pipelines
 
-| Model | Checkpoint | Task | `--model-class-name` | Batching |
+| `--model-class-name` | Task | Required checkpoint repositories |
-|---|---|---|---|---|
+|---|---|---|
-| LTX-2 | `Lightricks/LTX-2` | One-stage T2V/I2V | `LTX2Pipeline` | Yes |
+| `LTX2Pipeline` | LTX-2 one-stage T2V/I2V | `Lightricks/LTX-2` |
-| LTX-2 distilled | `rootonchair/LTX-2-19b-distilled` | Two-stage T2V/I2V | `LTX2DistilledPipeline` | No |
+| `LTX2TwoStagePipeline` | LTX-2 ordinary two-stage T2V/I2V | `Lightricks/LTX-2` |
-| LTX-2.3 | `diffusers/LTX-2.3-Diffusers` | One-stage T2V/I2V | `LTX2Pipeline` | Yes |
+| `LTX2DistilledPipeline` | LTX-2 full-distilled two-stage T2V/I2V | `rootonchair/LTX-2-19b-distilled` |
+| `LTX2Pipeline` | LTX-2.3 one-stage T2V/I2V | `diffusers/LTX-2.3-Diffusers` |
+| `LTX2TwoStagePipeline` | LTX-2.3 ordinary two-stage T2V/I2V | `diffusers/LTX-2.3-Diffusers`<br>`Lightricks/LTX-2.3` |
+| `LTX2DistilledPipeline` | LTX-2.3 full-distilled two-stage T2V/I2V | `diffusers/LTX-2.3-Distilled-Diffusers`<br>`Lightricks/LTX-2.3` |
+
+Repositories in the table are download units. A full pipeline repository
+contains the Transformer, text encoder, connectors, VAEs, vocoder, scheduler,
+and tokenizer; an additional repository supplies LoRA or upsampler sidecars.
+The runtime first searches the model root for sidecars under their official
+filenames, then downloads them from the matching Lightricks Hub repository.
 
 `LTX2Pipeline` is the unified one-stage entry. Checkpoint metadata selects the
 LTX-2 or LTX-2.3 profile; omitting an image selects T2V, while one initial
@@ -16,10 +25,10 @@ image selects I2V. Both one-stage repositories declare this class, so
 `--model-class-name` is optional. LTX-2.3 requires the Diffusers checkpoint;
 the raw `Lightricks/LTX-2.3` safetensors repository is not directly loadable.
 
-`LTX2DistilledPipeline` is the unified distilled entry: omit `image` for T2V
+`LTX2TwoStagePipeline` samples the regular model at half resolution, upsamples,
-or provide one initial image for I2V. As in the official LTX pipelines,
+then refines with the distilled LoRA. `LTX2DistilledPipeline` uses a fully
-distilled mode is selected explicitly rather than inferred from the checkpoint
+merged distilled Transformer in both stages. Both entries support T2V and I2V;
-name or metadata; always pass `--model-class-name LTX2DistilledPipeline`.
+select their class explicitly.
 
 ## API Migration
 
@@ -68,24 +77,28 @@ The model recipe also supplies the default negative prompt. Top-level
 API defaults `num_frames` to `1`, so set it explicitly; the offline LTX scripts
 default to `121`.
 
-## Distilled Two-Stage Defaults
+## Two-Stage Defaults
 
-| Parameter | Value |
+| Parameter | Ordinary | Full-distilled |
-|---|---:|
+|---|---:|---:|
-| Final width × height | 1536 × 1024 |
+| Final width × height | 1536 × 1024 | 1536 × 1024 |
-| Stage 1 width × height | 768 × 512 |
+| Stage 1 width × height | 768 × 512 | 768 × 512 |
-| Frames / frame rate | 121 / 24 |
+| Frames / frame rate | 121 / 24 | 121 / 24 |
-| Stage 1 / Stage 2 denoise steps | 8 / 3 |
+| Stage 1 / Stage 2 steps | 40 (LTX-2) or 30 (LTX-2.3) / 3 | 8 / 3 |
-| Guidance | Fixed positive-only |
+| Guidance | Stage 1 guided; Stage 2 positive-only | Fixed positive-only |
 
-The dimensions passed to the API are the final output dimensions and must be
+API dimensions are final dimensions and must be divisible by 64. All LTX
-divisible by 64. Both one- and two-stage requests require `num_frames = 8k+1`.
+requests require `num_frames = 8k+1`. Ordinary Stage 1 uses the LTX-2 or
-The distilled recipe fixes both sigma schedules; `num_inference_steps` must be
+LTX-2.3 one-stage defaults shown above. Distilled schedules are fixed, so
-`8` when supplied.
+`num_inference_steps`, when supplied, must be `8`. Both entries reject custom
+sigmas and input latents.
+
+Ordinary two-stage uses layer-fused LoRA for an unquantized BF16 Transformer
+and automatically switches to dynamic LoRA when quantization is enabled.
 
 ## Serving
 
-Start either checkpoint:
+Start either one-stage checkpoint:
 
 ```bash
 vllm serve Lightricks/LTX-2 --omni --stage-init-timeout 600
@@ -95,11 +108,19 @@ vllm serve Lightricks/LTX-2 --omni --stage-init-timeout 600
 vllm serve diffusers/LTX-2.3-Diffusers --omni --stage-init-timeout 600
 ```
 
-Start the distilled checkpoint with its two-stage entry:
+Start a two-stage checkpoint with its explicit class:
 
 ```bash
 vllm serve rootonchair/LTX-2-19b-distilled --omni \
   --model-class-name LTX2DistilledPipeline --stage-init-timeout 600
+# LTX-2.3 full-distilled; the v1.1 x2 upsampler is resolved when absent
+vllm serve diffusers/LTX-2.3-Distilled-Diffusers --omni \
+  --model-class-name LTX2DistilledPipeline --stage-init-timeout 600
+# Ordinary LTX-2.3 two-stage
+vllm serve diffusers/LTX-2.3-Diffusers --omni \
+  --model-class-name LTX2TwoStagePipeline \
+  --enable-layerwise-offload \
+  --stage-init-timeout 600
 ```
 
 The same server handles T2V and I2V. A T2V request using the selected recipe's
@@ -135,8 +156,9 @@ together with `input_reference`.
 
 ## Guidance
 
-LTX supports independent video/audio CFG, spatio-temporal guidance (STG),
+One-stage and ordinary Stage 1 support independent video/audio CFG,
-cross-modality guidance, and rescaling:
+spatio-temporal guidance (STG), cross-modality guidance, and rescaling.
+Distilled stages and ordinary Stage 2 are fixed positive-only.
 
 | Parameter | Default | Effect | Alias |
 |---|---:|---|---|
@@ -172,9 +194,9 @@ complete guidance plan in parallel. Despite the legacy `cfg` name, it covers
 text CFG, STG, and cross-modality guidance passes; guidance rescaling is
 applied after all pass predictions are gathered.
 
-The default one-stage recipe has four Transformer passes per denoise step:
+The default one-stage recipe and ordinary Stage 1 have four Transformer passes
-`cond`, `uncond`, `ptb` (STG), and `mod` (cross-modality). The useful balanced
+per denoise step: `cond`, `uncond`, `ptb` (STG), and `mod`
-configurations are therefore:
+(cross-modality). The useful balanced configurations are therefore:
 
 | `--cfg-parallel-size` | Passes per rank | Guidance-slot utilization | Notes |
 |---:|---:|---:|---|
@@ -195,9 +217,8 @@ vllm serve diffusers/LTX-2.3-Diffusers --omni \
 ```
 
 The total device count is the product of `cfg_parallel_size` and the other
-configured parallel dimensions. Distilled pipelines use positive-only
+configured parallel dimensions. Positive-only distilled phases do not benefit
-guidance, so `cfg_parallel_size > 1` does not provide guidance speedup for
+from `cfg_parallel_size > 1`.
-them.
 
 ### Python API
 
@@ -242,22 +263,22 @@ noted below.
 | `image` | image or batch, `None` | Direct value wins over request images; no image selects T2V. I2V accepts one image per prompt, and a batch cannot mix T2V/I2V. |
 | `prompt` | string or list, `None` | Positive-text fallback; request prompts win. Mutually exclusive with `prompt_embeds`. |
 | `negative_prompt` | string or list, `None` | Fallback after request values and before the recipe default; mutually exclusive with negative embeddings. |
-| `height` | `int`, `None` | Request → direct value → recipe default; divisible by 32 one-stage or 64 distilled. |
+| `height` | `int`, `None` | Request → direct value → recipe default; divisible by 32 one-stage or 64 two-stage. |
 | `width` | `int`, `None` | Same precedence and alignment as `height`. |
 | `num_frames` | `int`, `None` | Request → direct value → recipe default; must be `8k+1` and also determines audio duration with `frame_rate`. |
 | `frame_rate` | `float`, `None` | Request `frame_rate` → request `fps` → direct value → recipe default. |
-| `num_inference_steps` | `int`, `None` | Request → direct value → recipe default; minimum 2 one-stage, fixed at 8 for distilled Stage 1. Custom one-stage `sigmas` determine actual steps. |
+| `num_inference_steps` | `int`, `None` | Request → direct value → recipe default; minimum 2 one-stage and ordinary Stage 1, fixed at 8 for distilled Stage 1. Custom one-stage `sigmas` determine actual steps. |
 | `sigmas` | list of float, `None` | One-stage only. Request values win; every request in a fused batch must use the same schedule. |
 | `timesteps` | list of int, `None` | Compatibility slot; LTX accepts only `None`. Use `sigmas`. |
-| `guidance_scale` | `float`, `None` | One-stage common video/audio CFG fallback; an explicit request value wins. Distilled two-stage uses fixed positive-only guidance. |
+| `guidance_scale` | `float`, `None` | One-stage and ordinary Stage 1 common video/audio CFG fallback; an explicit request value wins. Distilled phases use fixed positive-only guidance. |
 | `guidance_rescale` | `float`, `None` | Accepts only `None` or `0.0`; use the modality rescale fields. |
 | `noise_scale` | `float`, `0.0` | Compatibility slot; LTX accepts only `0.0`. |
 | `num_videos_per_prompt` | `int`, `1` | Output-count fallback; positive request `num_outputs_per_prompt` wins. |
 | `generator` | generator or list, `None` | Explicit RNG; otherwise request generators/seeds are collated. Lists must match the effective output batch. |
-| `latents` | tensor, `None` | Request tensors win. One-stage accepts packed `[B, S, C]` and validated unpacked video layouts. |
+| `latents` | tensor, `None` | One-stage only. Request tensors win; packed `[B, S, C]` and validated unpacked video layouts are accepted. |
-| `audio_latents` | tensor, `None` | One-stage initial audio latents; request tensors win and are collated. |
+| `audio_latents` | tensor, `None` | One-stage only. Request tensors win and are collated. |
 | `prompt_embeds` | tensor, `None` | Precomputed positive conditioning; requires `prompt_attention_mask` and cannot accompany `prompt`. |
-| `negative_prompt_embeds` | tensor, `None` | One-stage precomputed negative conditioning; requires its mask and cannot accompany `negative_prompt`. |
+| `negative_prompt_embeds` | tensor, `None` | One-stage and ordinary Stage 1 conditioning; requires its mask and cannot accompany `negative_prompt`. |
 | `prompt_attention_mask` | tensor, `None` | Mask for positive embeddings; request `prompt_attention_mask` or `attention_mask` wins. |
 | `negative_prompt_attention_mask` | tensor, `None` | Mask for negative embeddings; request negative mask fields win. |
 | `decode_timestep` | float or list, `0.0` | Video-VAE decode timestep; request value wins. Lists may match 1, prompt batch, or output batch. |
@@ -273,13 +294,13 @@ request prompt payload; LTX guidance fields live in sampling `extra_args`.
 
 ### Recipe-Specific Request Capabilities
 
-| Override | One-stage | Distilled two-stage |
+| Override | One-stage | Ordinary two-stage | Distilled two-stage |
-|---|---|---|
+|---|---|---|---|
-| Guidance | Supported | Fixed positive-only |
+| Guidance | Supported | Stage 1 only; Stage 2 is positive-only | Fixed positive-only |
-| Negative prompt/embeddings | Supported | Rejected |
+| Negative prompt/embeddings | Supported | Supported by Stage 1 | Rejected |
-| `num_inference_steps` | Supported | Fixed at 8 for Stage 1; Stage 2 uses 3 |
+| `num_inference_steps` | Supported | Controls Stage 1; Stage 2 uses 3 | Fixed at 8 for Stage 1; Stage 2 uses 3 |
-| Custom `sigmas` | Supported | Rejected; both phases use fixed schedules |
+| Custom `sigmas` | Supported | Rejected | Rejected; both phases use fixed schedules |
-| Video/audio latents | Supported | Rejected |
+| Video/audio latents | Supported | Rejected | Rejected |
 
 These capability checks apply equally to direct `forward` keywords and
 values in `OmniDiffusionSamplingParams`; unsupported values fail instead of
@@ -309,6 +330,10 @@ bundled offline CLI do not currently expose `sigmas`.
   `--max-num-seqs 1`.
 - `--cfg-parallel-size` shards the complete LTX guidance plan, including STG,
   modality guidance, and rescale-compatible prediction gathering.
+- Sequence parallelism may pad audio latents. Pure Ulysses masks the padding;
+  Ring cannot, so audio length must be SP-divisible. Use `ring_degree=1` or a
+  divisible request shape.
+- Cache-DiT is one-stage only; multi-stage configurations are rejected.
 
 ## Operational Notes
 
@@ -323,14 +348,15 @@ bundled offline CLI do not currently expose `sigmas`.
   request parameter.
 - For benchmarks, use `tests/dfx/perf/tests/test_ltx2_vllm_omni.json` with
   `tests/dfx/perf/scripts/run_diffusion_benchmark.py`.
-- Current two-stage support is limited to the distilled checkpoint; official
+- Ordinary and full-distilled two-stage T2V/I2V are supported; HQ execution
-  LTX two-stage/HQ execution remains out of scope.
+  remains out of scope.
 
 ## References
 
 - <https://huggingface.co/Lightricks/LTX-2>
 - <https://huggingface.co/Lightricks/LTX-2.3>
 - <https://huggingface.co/diffusers/LTX-2.3-Diffusers>
+- <https://huggingface.co/diffusers/LTX-2.3-Distilled-Diffusers>
 - [Online video generation](../../docs/user_guide/examples/online_serving/text_to_video.md)
 - [Diffusion execution modes](../../docs/user_guide/diffusion/execution_modes.md)
 - [T2V offline example](../../examples/offline_inference/text_to_video/text_to_video.md)

recipes/rednote-hilab/dots.tts.md

@@ -0,0 +1,152 @@
+# dots.tts
+
+> Offline continuous-AR TTS at 48 kHz (rednote-hilab)
+
+## Summary
+
+- Vendor: rednote-hilab
+- Model: `rednote-hilab/dots.tts-soar`
+- Task: Text-to-speech, zero-shot synthesis only
+- Mode: Offline end-to-end example only (not yet wired for online serving —
+  see [Known limitations](#known-limitations))
+- Maintainer: Community
+
+## When to use this recipe
+
+Use this recipe as a known-good starting point for running
+`rednote-hilab/dots.tts-soar` offline on vLLM-Omni on consumer-class GPUs.
+dots.tts is a ~1.7B-parameter continuous-AR TTS model (Qwen2.5-1.5B base LM
++ 344M DiT flow-matching head + 180M AudioVAE) that emits 48 kHz mono audio.
+It follows the same "vLLM-native base LM + side-path computation" pattern as
+VoxCPM2 — single-stage pipeline
+`Qwen2.5-1.5B base LM → DiT (10-step Euler flow matching) → patch_encoder AR
+loopback → AudioVAE (streaming decode)` — with a plain Qwen2 backbone
+instead of MiniCPM4, and no FSQ / residual-LM stage.
+
+This is an early integration (see [Known limitations](#known-limitations)
+below before depending on it for anything beyond offline experimentation).
+
+## References
+
+- Offline end-to-end script:
+  [`examples/offline_inference/text_to_speech/dots_tts/end2end.py`](../../examples/offline_inference/text_to_speech/dots_tts/end2end.py)
+- Example guide:
+  [`examples/offline_inference/text_to_speech/README.md`](../../examples/offline_inference/text_to_speech/README.md#dotstts)
+- Default deploy config:
+  [`vllm_omni/deploy/dots_tts.yaml`](../../vllm_omni/deploy/dots_tts.yaml)
+- Talker / pipeline source:
+  [`vllm_omni/model_executor/models/dots_tts/`](../../vllm_omni/model_executor/models/dots_tts/)
+- Upstream: [rednote-hilab/dots.tts](https://github.com/rednote-hilab/dots.tts)
+
+## Hardware Support
+
+This recipe documents one tested 16 GB consumer-GPU configuration. Other
+vendor sections (ROCm, NPU) and larger-VRAM configurations are welcome as
+community validation lands.
+
+## GPU
+
+### 1 x RTX 5080 16GB (Single GPU, Minimum Recommended)
+
+dots.tts (~1.7B params across the base LM + DiT + AudioVAE + CAM++
+speaker encoder, bfloat16) fits comfortably on a single 16 GB GPU. The
+bundled default config at
+[`vllm_omni/deploy/dots_tts.yaml`](../../vllm_omni/deploy/dots_tts.yaml)
+(`gpu_memory_utilization: 0.8`, `max_num_seqs: 4`, `enforce_eager: true`,
+`enable_prefix_caching: false`) loads cleanly with ~5.1 GiB for model
+weights and ~0.3 GiB peak activation; the remainder of the configured
+budget is available for KV cache. Total resident footprint at idle is
+roughly **7-8 GiB / 16 GB** — the only tight spot in the full CUDA-Graph
+roadmap would be step 8's graph capture (not implemented yet; this
+release runs `enforce_eager: true`, so it doesn't apply today).
+
+#### Environment
+
+- OS: Linux (WSL2)
+- Python: 3.12
+- Driver / runtime: NVIDIA driver 595.95
+- torch: 2.11.0+cu130
+- vLLM: 0.26.0
+- vLLM-Omni: 0.22.1.dev (current `main`)
+
+#### Command
+
+```bash
+python examples/offline_inference/text_to_speech/dots_tts/end2end.py \
+    --model rednote-hilab/dots.tts-soar \
+    --text "Hello, this is a test of dots TTS running on vLLM Omni."
+```
+
+The deploy config at
+[`vllm_omni/deploy/dots_tts.yaml`](../../vllm_omni/deploy/dots_tts.yaml)
+is loaded automatically by the model registry (HF `model_type=dots_tts`).
+Pass `--deploy-config <path>` to override.
+
+#### Verification
+
+**T1 — offline zero-shot synthesis**:
+
+```bash
+python examples/offline_inference/text_to_speech/dots_tts/end2end.py \
+    --model rednote-hilab/dots.tts-soar \
+    --text "Hello, this is a test of dots TTS running on vLLM Omni."
+```
+
+Observed: `output_audio/output.wav`, 3.52 s @ 48 kHz mono. Single
+one-shot process (init → one `generate()` → exit), so the reported
+numbers include engine init and first-request warmup, not just steady-state
+per-step throughput — same caveat as VoxCPM2's recipe. `Inference: 5.52s`,
+`RTF: 1.569`.
+
+Weight-loading breakdown from the same run (all tensors matched, no
+missing/extra keys): 951/951 AudioVAE, 244/244 DiT, 270/270
+patch_encoder, 198 Qwen2, 938/938 CAM++ speaker encoder.
+
+Whisper transcription of the output matched the input text with no
+dropped leading word (confirms the streaming-vocoder patch-boundary fix
+described in [Known limitations](#known-limitations)).
+
+#### Notes
+
+- Output: 48 kHz mono WAV.
+- Checkpoints: `rednote-hilab/dots.tts-soar` is the validated default
+  used throughout this recipe. `rednote-hilab/dots.tts-base` shares the
+  same architecture but is unvalidated in this repo. `rednote-hilab/dots.tts-mf`
+  (MeanFlow, 2-4 step) is not supported — see below.
+- `enforce_eager: true` and `enable_prefix_caching: false` in the deploy
+  config are load-bearing, not just conservative defaults: with prefix
+  caching enabled, vLLM-Omni's prefix-cache multimodal-output merge path
+  does not preserve the `sparse_audio` marker this model relies on to
+  route audio output correctly, and generation silently truncates to a
+  single ~160 ms patch. Do not override `enable_prefix_caching` for this
+  model until that framework-level gap is fixed.
+
+## Known limitations
+
+- **No online serving yet.** dots.tts is not registered in
+  `vllm_omni/entrypoints/openai/tts_adapters/` (the framework other TTS
+  models in this repo use for `/v1/audio/speech`), so `vllm serve
+  rednote-hilab/dots.tts-soar --omni` will start but the OpenAI-compatible
+  speech endpoint does not know how to build request params for it yet.
+  Offline `Omni()` (as used in `end2end.py`) is the only supported path
+  today.
+- **Voice cloning is not wired.** The CAM++ x-vector speaker encoder
+  weights load and are exercised by `load_weights()`, but nothing in the
+  prompt builder consumes a reference audio yet — generation is zero-shot
+  only. `end2end.py` has no `--ref-audio`/`--ref-text` flags.
+- **`dots.tts-mf` (MeanFlow, 2-4 step) checkpoint is not supported.** Only
+  the fixed 10-step Euler DiT sampler used by `dots.tts-soar` /
+  `dots.tts-base` is implemented.
+- **No CUDA graph capture.** The talker runs fully eager. voxcpm2's three
+  captured graphs (base LM decode, CFM solver, VAE decode) have no
+  dots.tts equivalent yet.
+- **Concurrent requests do not scale.** Each request's 10-step DiT Euler
+  integration runs serially in the side path (no cross-request batching,
+  unlike voxcpm2's `enable_batched_cfm`). A community review of this
+  integration measured no throughput gain at `c=4` concurrent requests
+  versus `c=1`.
+- **`SamplingParams.seed` does not control audio-generation randomness.**
+  The DiT's flow-matching noise is deterministically derived per-request
+  from a fixed internal seed (reproducible run-to-run, matching
+  voxcpm2's `deterministic_cfm_seed` convention) rather than from the
+  caller-supplied `seed` field — the same limitation voxcpm2 has today.

tests/buildkite/test_upload_pipeline.py

@@ -5,6 +5,7 @@ import sys
 from pathlib import Path
 
 import pytest
+import yaml
 
 sys.path.insert(0, str(Path(__file__).resolve().parents[2] / ".buildkite" / "common" / "scripts"))
 
@@ -139,3 +140,48 @@ def test_mirror_hardwares_a2b3_npu_4_expands_agents_image_and_plugins() -> None:
     assert step["plugins"][0]["kubernetes"]["podSpecPatch"]["imagePullSecrets"] == [
         {"name": "swr-secret"},
     ]
+
+
+COVERAGE_PILOT_LABELS = ("Diffusion · Bagel Test", "TTS · Qwen3-TTS Base Test")
+MERGE_PIPELINE = Path(".buildkite/cuda/test-merge.yml")
+
+
+def _surviving_labels(changed_files: list[str]) -> set[str]:
+    doc = yaml.safe_load(MERGE_PIPELINE.read_text(encoding="utf-8"))
+    rendered = _render_test_pipeline(doc, changed_files=changed_files)
+    labels: set[str] = set()
+
+    def walk(steps: list | None) -> None:
+        for step in steps or []:
+            if not isinstance(step, dict):
+                continue
+            if "label" in step:
+                labels.add(step["label"])
+            walk(step.get("steps"))
+
+    walk(rendered.get("steps"))
+    return labels
+
+
+@pytest.mark.parametrize(
+    "changed_file",
+    [
+        ".buildkite/common/scripts/run_cov_split.sh",
+        "pyproject.toml",
+    ],
+)
+def test_coverage_pilots_selected_by_their_shared_inputs(changed_file: str) -> None:
+    """The split helper and the coverage config both change what these jobs measure.
+
+    Without them in source_file_dependencies the jobs are filtered out of normal
+    PR builds, so a regression would only surface in a later nightly.
+    """
+    labels = _surviving_labels([changed_file])
+    for pilot in COVERAGE_PILOT_LABELS:
+        assert pilot in labels, f"{pilot} dropped when {changed_file} changed"
+
+
+def test_coverage_pilots_not_selected_by_unrelated_change() -> None:
+    labels = _surviving_labels(["vllm_omni/entrypoints/openai/serving_chat.py"])
+    for pilot in COVERAGE_PILOT_LABELS:
+        assert pilot not in labels

tests/config/test_config_factory.py

@@ -6,7 +6,7 @@ Unit tests for StageConfigFactory and related classes.
 
 import importlib
 import warnings
-from dataclasses import dataclass, fields
+from dataclasses import dataclass
 from pathlib import Path
 from unittest.mock import patch
 
@@ -34,9 +34,8 @@ from vllm_omni.config.stage_config import (
     load_deploy_config,
     merge_pipeline_deploy,
     pipeline_cfg_resolver,
-    strip_parent_engine_args,
 )
-from vllm_omni.engine.arg_utils import SHARED_FIELDS, EngineArgs, internal_blacklist_keys
+from vllm_omni.engine.arg_utils import SHARED_FIELDS, internal_blacklist_keys
 
 pytestmark = [pytest.mark.core_model, pytest.mark.cpu]
 
@@ -475,7 +474,7 @@ class TestStageConfigFactory:
         cli_overrides = {
             "gpu_memory_utilization": 0.9,
             "model": "some_model",  # Internal
-            "stage_configs_path": "/path",  # Internal
+            "deploy_config": "/path",  # Internal
             "batch_timeout": 10,  # Internal
         }
 
@@ -483,7 +482,7 @@ class TestStageConfigFactory:
 
         assert overrides["gpu_memory_utilization"] == 0.9
         assert "model" not in overrides
-        assert "stage_configs_path" not in overrides
+        assert "deploy_config" not in overrides
         assert "batch_timeout" not in overrides
 
     def test_per_stage_override_excludes_internal_keys(self):
@@ -525,49 +524,6 @@ class TestStageResolutionHelpers:
         assert "model" not in overrides
         assert "parallel_config" not in overrides
 
-    def test_strip_parent_engine_args_reports_only_surprising_parent_overrides(self):
-        parent_fields = {f.name: f for f in fields(EngineArgs)}
-        filtered, overridden = strip_parent_engine_args(
-            {
-                "model": "some/model",
-                "stage_configs_path": "/tmp/stages.yaml",
-                "tensor_parallel_size": 4,
-                "worker_extension_cls": "some.Extension",
-                "custom_pipeline_args": {"pipeline_class": "demo.Pipeline"},
-            },
-            parent_fields=parent_fields,
-            keep_keys={"worker_extension_cls"},
-            strip_keys={"stage_configs_path"},
-            no_warn_keys={"model"},
-        )
-
-        assert filtered == {
-            "worker_extension_cls": "some.Extension",
-            "custom_pipeline_args": {"pipeline_class": "demo.Pipeline"},
-        }
-        assert overridden == ["tensor_parallel_size"]
-
-    def test_strip_parent_engine_args_keeps_allowed_media_access_controls(self):
-        parent_fields = {f.name: f for f in fields(EngineArgs)}
-        filtered, overridden = strip_parent_engine_args(
-            {
-                "model": "some/model",
-                "stage_configs_path": "/tmp/stages.yaml",
-                "allowed_local_media_path": "/data/qwentts",
-                "allowed_media_domains": ["example.com"],
-            },
-            parent_fields=parent_fields,
-            keep_keys={"allowed_local_media_path", "allowed_media_domains"},
-            strip_keys={"stage_configs_path"},
-            no_warn_keys={"model"},
-        )
-
-        assert filtered == {
-            "allowed_local_media_path": "/data/qwentts",
-            "allowed_media_domains": ["example.com"],
-        }
-        assert overridden == []
-
 
 class TestPipelineDiscovery:
     """Tests for the central pipeline registry (``OMNI_PIPELINES``)."""

tests/config/test_omni_config.py

@@ -471,6 +471,7 @@ def test_sub_config_fields_match_structured_scopes():
         "trust_remote_code",
         "dtype",
         "attention_backend",
+        "attention_config",
         "moe_backend",
         "hf_overrides",
         "limit_mm_per_prompt",

tests/diffusion/ar_diffusion/test_paged_attention.py

@@ -60,9 +60,18 @@ def _dense_attention(query: torch.Tensor, key: torch.Tensor, value: torch.Tensor
     return torch.einsum("bhqk,bkhd->bqhd", probs, value)
 
 
-def _cuda_flash_attn_usable() -> bool:
+def _gpu_flash_attn_usable() -> bool:
     if not torch.cuda.is_available():
         return False
+    if torch.version.hip is not None:
+        for module in ("aiter", "flash_attn"):
+            try:
+                imported = __import__(module, fromlist=["flash_attn_varlen_func"])
+                if getattr(imported, "flash_attn_varlen_func", None) is not None:
+                    return True
+            except ImportError:
+                pass
+        return False
     try:
         spec = find_spec("vllm.vllm_flash_attn")
         if spec is None or spec.origin is None:
@@ -227,12 +236,11 @@ def test_paged_attention_matches_dense_reference_cpu(history_chunks, action_len,
     assert st.adapter(POS).completed_chunks == before + (1 if commit_current else 0)
 
 
-@pytest.mark.skipif(not _cuda_flash_attn_usable(), reason="usable CUDA FlashAttention is required")
+@pytest.mark.skipif(not _gpu_flash_attn_usable(), reason="usable GPU FlashAttention is required")
 @pytest.mark.parametrize("history_chunks", [1, 3])
 @pytest.mark.parametrize("action_len", [0, 3])
 @pytest.mark.parametrize("commit_current", [False, True])
 def test_paged_attention_matches_dense_reference_gpu(history_chunks, action_len, commit_current):
-    pytest.importorskip("vllm.vllm_flash_attn")
     torch.manual_seed(0)
     device = torch.device("cuda")
     dtype = torch.float16

tests/diffusion/attention/test_flash_attn.py

@@ -265,10 +265,15 @@ def test_fa_vs_sdpa():
     print(f"Max absolute difference (valid region): {max_diff:.6f}")
     print(f"Mean absolute difference (valid region): {mean_diff:.6f}")
 
-    # Assert that outputs are close
+    # AITER's ROCm BF16 kernel is deterministic, but differs from ROCm SDPA by
-    # Using higher tolerance for bfloat16 and different implementations
+    # about 1.3% relative error. An absolute-only bound misclassifies that
-    assert max_diff < 0.01, f"Max difference {max_diff} exceeds threshold 0.01"
+    # expected rounding difference for larger output values (for example,
-    assert mean_diff < 0.001, f"Mean difference {mean_diff} exceeds threshold 0.001"
+    # -2.65625 vs -2.6875). Keep the tighter historical bounds elsewhere.
+    if current_omni_platform.is_rocm():
+        torch.testing.assert_close(output_fa_valid, output_sdpa_valid, rtol=1e-2, atol=1.5e-2)
+    else:
+        assert max_diff < 0.01, f"Max difference {max_diff} exceeds threshold 0.01"
+        assert mean_diff < 0.001, f"Mean difference {mean_diff} exceeds threshold 0.001"
 
     print("✓ Case 2 PASSED: FA and SDPA outputs are very close!")
 

tests/diffusion/distributed/test_sequence_parallel.py

@@ -416,8 +416,8 @@ def test_sp_correctness_advanced(model_name: str):
 
 
 @pytest.mark.skipif(
-    not (current_omni_platform.is_cuda() or current_omni_platform.is_xpu()),
+    not (current_omni_platform.is_cuda() or current_omni_platform.is_rocm() or current_omni_platform.is_xpu()),
-    reason="Only tested on CUDA and XPU",
+    reason="Only tested on CUDA, ROCM and XPU",
 )
 @pytest.mark.diffusion
 @pytest.mark.parallel
@@ -447,8 +447,8 @@ def test_local_sp_padding_mask(monkeypatch: pytest.MonkeyPatch) -> None:
 
 
 @pytest.mark.skipif(
-    not (current_omni_platform.is_cuda() or current_omni_platform.is_xpu()),
+    not (current_omni_platform.is_cuda() or current_omni_platform.is_rocm() or current_omni_platform.is_xpu()),
-    reason="Only tested on CUDA and XPU",
+    reason="Only tested on CUDA, ROCM and XPU",
 )
 @pytest.mark.diffusion
 @pytest.mark.parallel
@@ -468,8 +468,8 @@ def test_local_sp_padding_mask_no_padding(monkeypatch: pytest.MonkeyPatch) -> No
 
 
 @pytest.mark.skipif(
-    not (current_omni_platform.is_cuda() or current_omni_platform.is_xpu()),
+    not (current_omni_platform.is_cuda() or current_omni_platform.is_rocm() or current_omni_platform.is_xpu()),
-    reason="Only tested on CUDA and XPU",
+    reason="Only tested on CUDA, ROCM and XPU",
 )
 @pytest.mark.diffusion
 @pytest.mark.parallel

tests/diffusion/layers/test_fused_qk_norm_rope.py

@@ -0,0 +1,82 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+import pytest
+import torch
+import torch.nn.functional as F
+from vllm.triton_utils import HAS_TRITON
+
+pytestmark = [pytest.mark.core_model, pytest.mark.cuda, pytest.mark.diffusion]
+
+_HEAD_DIM = 128
+_ROTARY_DIM = 96
+_EPS = 1e-5
+
+
+def _reference(q, k, q_weight, k_weight, rope_table):
+    q = F.rms_norm(q, (_HEAD_DIM,), q_weight, _EPS)
+    k = F.rms_norm(k, (_HEAD_DIM,), k_weight, _EPS)
+    half = _ROTARY_DIM // 2
+    cos = rope_table[..., :half].unsqueeze(1)
+    sin = rope_table[..., half:].unsqueeze(1)
+
+    def apply(x):
+        first = x[..., :half]
+        second = x[..., half:_ROTARY_DIM]
+        return torch.cat(
+            (
+                first * cos - second * sin,
+                second * cos + first * sin,
+                x[..., _ROTARY_DIM:],
+            ),
+            dim=-1,
+        )
+
+    return apply(q), apply(k)
+
+
+@pytest.mark.skipif(not torch.cuda.is_available(), reason="CUDA required")
+@pytest.mark.skipif(not HAS_TRITON, reason="Triton required")
+@pytest.mark.parametrize("seq_len", [1, 257, 1024])
+def test_fused_qk_norm_rope_matches_bf16_reference(seq_len):
+    from vllm_omni.diffusion.layers.fused_qk_norm_rope import (
+        fused_qk_norm_rope,
+    )
+
+    torch.manual_seed(17)
+    heads = 14
+    qkv = torch.randn(
+        seq_len,
+        heads * _HEAD_DIM * 3,
+        device="cuda",
+        dtype=torch.bfloat16,
+    )
+    q = qkv[:, : heads * _HEAD_DIM].view(seq_len, heads, _HEAD_DIM)
+    k = qkv[:, heads * _HEAD_DIM : 2 * heads * _HEAD_DIM].view(
+        seq_len,
+        heads,
+        _HEAD_DIM,
+    )
+    q_weight = torch.randn(_HEAD_DIM, device="cuda", dtype=torch.bfloat16)
+    k_weight = torch.randn(_HEAD_DIM, device="cuda", dtype=torch.bfloat16)
+    freqs = torch.randn(seq_len, _ROTARY_DIM // 2, device="cuda")
+    rope_table = torch.cat((torch.cos(freqs), torch.sin(freqs)), dim=-1).to(torch.bfloat16)
+
+    expected_q, expected_k = _reference(
+        q,
+        k,
+        q_weight,
+        k_weight,
+        rope_table,
+    )
+    actual_q, actual_k = fused_qk_norm_rope(
+        q,
+        k,
+        q_weight,
+        k_weight,
+        rope_table,
+        _EPS,
+    )
+
+    torch.testing.assert_close(actual_q, expected_q, atol=0.0625, rtol=0.02)
+    torch.testing.assert_close(actual_k, expected_k, atol=0.0625, rtol=0.02)

tests/diffusion/lora/test_images_generations_lora.py

@@ -24,11 +24,20 @@ from safetensors.torch import save_file
 
 from tests.helpers.mark import hardware_test
 from tests.helpers.runtime import OmniServer, OmniServerParams
+from vllm_omni.platforms import current_omni_platform
 
 os.environ["VLLM_WORKER_MULTIPROC_METHOD"] = "spawn"
 
 MODEL = "Tongyi-MAI/Z-Image-Turbo"
 
+# ROCm's default FLASH_ATTN diffusion backend resolves to AITER. On a cold CI
+# worker, JIT-building its MHA kernel can take longer than the server readiness
+# timeout. This test validates LoRA switching rather than attention-backend or
+# compiled-execution performance, so use the eager SDPA path on ROCm only.
+ROCM_SERVER_ARGS = (
+    ["--enforce-eager", "--diffusion-attention-backend", "TORCH_SDPA"] if current_omni_platform.is_rocm() else []
+)
+
 
 PROMPT = "a photo of a cat sitting on a laptop keyboard"
 SIZE = "256x256"
@@ -42,7 +51,14 @@ server_params = [
             server_args=[
                 "--num-gpus",
                 "1",
+                *ROCM_SERVER_ARGS,
             ],
+            # A cold-cache Z-Image load can take slightly more than the
+            # orchestrator's 900s default on CI (weights alone have been
+            # observed at ~690s and full orchestrator readiness at ~906s).
+            # Match the large-model offline runner's startup allowance without
+            # weakening timeout detection for every online-serving test.
+            init_timeout=1800,
         ),
         id="zimage_turbo_lora",
     ),

tests/diffusion/models/ltx2/test_ltx2_latents.py

@@ -23,7 +23,7 @@ def _make_pipeline(pipeline_cls, sequence_parallel_size: int = 1):
     return pipeline
 
 
-def test_prepare_video_latents_samples_directly_in_packed_token_space():
+def test_prepare_video_latents_matches_official_values_and_token_major_layout():
     pipeline = _make_pipeline(LTX2Pipeline)
     pipeline.vae_spatial_compression_ratio = 8
     pipeline.vae_temporal_compression_ratio = 8
@@ -44,6 +44,7 @@ def test_prepare_video_latents_samples_directly_in_packed_token_space():
     )
 
     torch.testing.assert_close(actual, expected, rtol=0, atol=0)
+    assert actual.stride()[1:] == (1, actual.shape[1])
 
 
 def test_prepare_audio_latents_samples_directly_in_packed_token_space():
@@ -81,6 +82,42 @@ def test_prepare_audio_latents_pads_generated_dummy_length_for_sp():
     assert original_num_frames == 1
     assert padded_num_frames == 2
     assert latents.shape == (1, 2, 128)
+    torch.testing.assert_close(latents[:, 1:], torch.zeros_like(latents[:, 1:]))
+
+
+def test_prepare_audio_latents_request_rng_is_invariant_to_sp_padding():
+    pipeline_sp1 = _make_pipeline(LTX2Pipeline, sequence_parallel_size=1)
+    pipeline_sp4 = _make_pipeline(LTX2Pipeline, sequence_parallel_size=4)
+    generator_sp1 = torch.Generator().manual_seed(42)
+    generator_sp4 = torch.Generator().manual_seed(42)
+
+    latents_sp1, _, _ = pipeline_sp1.prepare_audio_latents(
+        batch_size=1,
+        num_channels_latents=2,
+        num_mel_bins=8,
+        audio_latent_length=3,
+        dtype=torch.float32,
+        device=torch.device("cpu"),
+        generator=generator_sp1,
+    )
+    latents_sp4, _, _ = pipeline_sp4.prepare_audio_latents(
+        batch_size=1,
+        num_channels_latents=2,
+        num_mel_bins=8,
+        audio_latent_length=3,
+        dtype=torch.float32,
+        device=torch.device("cpu"),
+        generator=generator_sp4,
+    )
+
+    torch.testing.assert_close(latents_sp4[:, :3], latents_sp1, rtol=0, atol=0)
+    torch.testing.assert_close(latents_sp4[:, 3:], torch.zeros_like(latents_sp4[:, 3:]))
+    torch.testing.assert_close(
+        torch.randn(8, generator=generator_sp4),
+        torch.randn(8, generator=generator_sp1),
+        rtol=0,
+        atol=0,
+    )
 
 
 def test_prepare_audio_latents_pads_packed_sequence_dim_for_provided_latents():
@@ -137,7 +174,9 @@ def test_prepare_audio_latents_accepts_already_padded_4d_latents_for_sp():
     assert original_num_frames == 10
     assert padded_num_frames == 12
     assert padded.shape == (1, 12, 8)
-    torch.testing.assert_close(padded, latent_ops.pack_audio_latents(latents))
+    packed = latent_ops.pack_audio_latents(latents)
+    torch.testing.assert_close(padded[:, :10], packed[:, :10])
+    torch.testing.assert_close(padded[:, 10:], torch.zeros_like(padded[:, 10:]))
 
 
 def test_resolve_audio_latent_length_preserves_legacy_4d_shape_inference():

tests/diffusion/models/ltx2/test_ltx2_parallel.py

@@ -12,6 +12,19 @@ import torch
 pytestmark = [pytest.mark.core_model, pytest.mark.cpu]
 
 
+def test_velocity_from_x0_uses_official_host_scalar_sigma(mocker):
+    from vllm_omni.diffusion.models.ltx2.ltx2_guidance import velocity_from_x0
+
+    sigma = mocker.MagicMock()
+    sigma.to.return_value.item.return_value = 0.5
+
+    actual = velocity_from_x0(torch.tensor([1.0]), torch.tensor([0.0]), sigma)
+
+    sigma.to.assert_called_once_with(torch.float32)
+    sigma.to.return_value.item.assert_called_once_with()
+    torch.testing.assert_close(actual, torch.tensor([2.0]))
+
+
 class TestCFGParallelHelpers:
     """Test LTX-2.3 CFG helper math without loading model weights."""
 
@@ -199,6 +212,7 @@ class TestCFGParallelHelpers:
             ),
             attention_kwargs=None,
             audio_scheduler=SimpleNamespace(sigmas=torch.stack([audio_sigma])),
+            original_audio_num_frames=state.audio.shape[1],
         )
 
         actual_video, actual_audio = executor.predict_parallel_guidance(

tests/diffusion/models/ltx2/test_ltx2_phase_adapter.py

@@ -0,0 +1,483 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+"""Unit tests for the fixed LTX refinement-phase adapter."""
+
+from types import SimpleNamespace
+
+import pytest
+import torch
+from safetensors.torch import save_file
+
+from vllm_omni.diffusion.models.ltx2 import ltx2_phase_adapter
+from vllm_omni.diffusion.models.ltx2.ltx2_adapter_parser import (
+    AdapterTarget,
+    parse_ltx_adapter,
+)
+from vllm_omni.diffusion.models.ltx2.ltx2_components import LTX23_TWO_STAGE_COMPONENT_PROFILE
+from vllm_omni.diffusion.models.ltx2.ltx2_phase_adapter import LTXPhaseAdapterRuntime
+from vllm_omni.diffusion.models.ltx2.ltx2_recipes import (
+    LTX2_TWO_STAGE_RECIPE,
+    LTX_DISTILLED_ADAPTER_SLOT,
+)
+
+pytestmark = [pytest.mark.core_model, pytest.mark.cpu]
+
+
+def _target(
+    lora_a,
+    lora_b,
+    *,
+    source_module="proj",
+    module="proj",
+    target_slice=None,
+):
+    return AdapterTarget(
+        source_module=source_module,
+        module=module,
+        slice=target_slice,
+        a_key="a",
+        b_key="b",
+        a_shape=tuple(lora_a.shape),
+        b_shape=tuple(lora_b.shape),
+    )
+
+
+def _wrap(
+    layer,
+    lora_a,
+    lora_b,
+    *,
+    source_module="proj",
+    module="proj",
+    target_slice=None,
+    layer_fused=False,
+):
+    target = _target(
+        lora_a,
+        lora_b,
+        source_module=source_module,
+        module=module,
+        target_slice=target_slice,
+    )
+    wrapper = ltx2_phase_adapter._PhaseAdapterLinear(
+        layer,
+        [target],
+        adapter_dtype=lora_a.dtype,
+        layer_fused=layer_fused,
+    )
+    wrapper.load_target(target, lora_a, lora_b)
+    wrapper.set_enabled(True)
+    return wrapper
+
+
+class _LoRAAttention(torch.nn.Module):
+    def __init__(self) -> None:
+        super().__init__()
+        self.to_qkv = torch.nn.Identity()
+
+
+class _LoRATransformer(torch.nn.Module):
+    stacked_params_mapping = ((".attn1.to_qkv", ".attn1.to_q", "q"),)
+
+    def __init__(self) -> None:
+        super().__init__()
+        self.proj_in = torch.nn.Linear(4, 6, bias=False)
+        block = torch.nn.Module()
+        block.attn1 = _LoRAAttention()
+        self.transformer_blocks = torch.nn.ModuleList([block])
+
+
+def test_ltx_adapter_parser_maps_official_and_packed_module_names(tmp_path):
+    adapter_path = tmp_path / "distilled.safetensors"
+    save_file(
+        {
+            "diffusion_model.text_embedding_projection.aggregate_embed.lora_A.weight": torch.randn(2, 4),
+            "diffusion_model.patchify_proj.lora_A.weight": torch.randn(2, 4),
+            "diffusion_model.patchify_proj.lora_B.weight": torch.randn(6, 2),
+            "diffusion_model.transformer_blocks.0.attn1.to_q.lora_A.weight": torch.randn(2, 4),
+            "diffusion_model.transformer_blocks.0.attn1.to_q.lora_B.weight": torch.randn(4, 2),
+        },
+        adapter_path,
+    )
+
+    manifest = parse_ltx_adapter(_LoRATransformer(), str(adapter_path))
+    assert {(target.source_module, target.module, target.slice) for target in manifest.targets} == {
+        ("proj_in", "proj_in", None),
+        ("transformer_blocks.0.attn1.to_q", "transformer_blocks.0.attn1.to_qkv", "q"),
+    }
+
+
+def test_ltx_adapter_parser_rejects_unpaired_b_tensor(tmp_path):
+    adapter_path = tmp_path / "distilled.safetensors"
+    save_file(
+        {"diffusion_model.patchify_proj.lora_B.weight": torch.randn(6, 2)},
+        adapter_path,
+    )
+
+    with pytest.raises(ValueError, match="Missing paired LoRA A tensors"):
+        parse_ltx_adapter(_LoRATransformer(), str(adapter_path))
+
+
+def _pipeline(*, dtype, quantization_config):
+    return SimpleNamespace(
+        pipeline_recipe=LTX2_TWO_STAGE_RECIPE,
+        component_profile=LTX23_TWO_STAGE_COMPONENT_PROFILE,
+        transformer=object(),
+        od_config=SimpleNamespace(
+            model="unused",
+            lora_path=None,
+            dtype=dtype,
+            quantization_config=quantization_config,
+        ),
+    )
+
+
+@pytest.mark.parametrize(
+    ("dtype", "quantization_config", "layer_fused"),
+    [
+        (torch.float32, object(), False),
+        (torch.bfloat16, None, True),
+    ],
+)
+def test_ltx_phase_adapter_selects_execution_from_quantization(
+    monkeypatch,
+    dtype,
+    quantization_config,
+    layer_fused,
+):
+    pipeline = _pipeline(dtype=dtype, quantization_config=quantization_config)
+    manifest = object()
+    captured = {}
+
+    def resolve_artifact(*_args, **kwargs):
+        assert not kwargs
+        return "adapter"
+
+    def parse(transformer, path):
+        assert transformer is pipeline.transformer
+        assert path == "adapter"
+        return manifest
+
+    class Runtime:
+        def __init__(self, transformer, parsed_manifest, **kwargs):
+            captured.update(transformer=transformer, manifest=parsed_manifest, **kwargs)
+
+    monkeypatch.setattr(ltx2_phase_adapter, "resolve_ltx_artifact", resolve_artifact)
+    monkeypatch.setattr(ltx2_phase_adapter, "parse_ltx_adapter", parse)
+    monkeypatch.setattr(ltx2_phase_adapter, "LTXPhaseAdapterRuntime", Runtime)
+
+    runtime = ltx2_phase_adapter.build_ltx_phase_adapter(pipeline)
+
+    assert isinstance(runtime, Runtime)
+    assert captured == {
+        "transformer": pipeline.transformer,
+        "manifest": manifest,
+        "dtype": dtype,
+        "layer_fused": layer_fused,
+    }
+
+
+def test_ltx_phase_adapter_rejects_non_bfloat16_layer_fused(monkeypatch):
+    pipeline = _pipeline(dtype=torch.float16, quantization_config=None)
+    monkeypatch.setattr(ltx2_phase_adapter, "resolve_ltx_artifact", lambda *_args, **_kwargs: "adapter")
+
+    with pytest.raises(ValueError, match="bfloat16"):
+        ltx2_phase_adapter.build_ltx_phase_adapter(pipeline)
+
+
+def test_ltx_phase_adapter_keeps_one_transformer_and_switches_a_fixed_slot(tmp_path):
+    class TinyTransformer(torch.nn.Module):
+        stacked_params_mapping = ()
+
+        def __init__(self) -> None:
+            super().__init__()
+            self.proj = torch.nn.Linear(2, 2, bias=False)
+
+    adapter_path = tmp_path / "distilled.safetensors"
+    save_file(
+        {
+            "diffusion_model.proj.lora_A.weight": torch.tensor([[1.0, 2.0]]),
+            "diffusion_model.proj.lora_B.weight": torch.tensor([[3.0], [4.0]]),
+        },
+        adapter_path,
+    )
+    transformer = TinyTransformer()
+    runtime = LTXPhaseAdapterRuntime(
+        transformer,
+        parse_ltx_adapter(transformer, str(adapter_path)),
+        dtype=torch.float32,
+    )
+
+    assert runtime.base_parameter_name("proj.weight") == "proj.base_layer.weight"
+    assert {name for name, _ in transformer.named_parameters()} == {"proj.base_layer.weight"}
+    with torch.no_grad():
+        transformer.proj.base_layer.weight.copy_(torch.eye(2))
+    runtime.finalize()
+
+    x = torch.tensor([[1.0, 1.0]])
+    runtime.activate(None)
+    torch.testing.assert_close(transformer.proj(x), torch.tensor([[1.0, 1.0]]))
+    runtime.activate(LTX_DISTILLED_ADAPTER_SLOT)
+    torch.testing.assert_close(transformer.proj(x), torch.tensor([[10.0, 13.0]]))
+
+
+def test_ltx_layer_fused_adapter_matches_official_bf16_weight_fusion():
+    layer = torch.nn.Linear(4, 3, bias=True, dtype=torch.bfloat16)
+    base_weight = torch.tensor(
+        [
+            [0.5, -0.25, 0.125, 0.75],
+            [-0.375, 0.625, 0.25, -0.5],
+            [0.875, -0.125, -0.75, 0.375],
+        ],
+        dtype=torch.bfloat16,
+    )
+    lora_a = torch.tensor(
+        [
+            [0.1015625, -0.203125, 0.3046875, -0.40625],
+            [0.5078125, -0.609375, 0.7109375, -0.8125],
+        ],
+        dtype=torch.bfloat16,
+    )
+    lora_b = torch.tensor(
+        [[0.9140625, -0.1171875], [-0.21875, 0.3203125], [0.421875, -0.5234375]],
+        dtype=torch.bfloat16,
+    )
+    with torch.no_grad():
+        layer.weight.copy_(base_weight)
+        layer.bias.copy_(torch.tensor([0.125, -0.25, 0.5], dtype=torch.bfloat16))
+    wrapper = _wrap(layer, lora_a, lora_b, layer_fused=True)
+
+    expected_weight = torch.matmul(lora_b, lora_a)
+    expected_weight.add_(base_weight)
+    x = torch.tensor([[0.6015625, -0.703125, 0.8046875, -0.90625]], dtype=torch.bfloat16)
+    expected = torch.nn.functional.linear(x, expected_weight, layer.bias)
+
+    assert torch.equal(wrapper(x), expected)
+    assert torch.equal(layer.weight, base_weight)
+
+
+def test_ltx_layer_fused_row_parallel_uses_one_collective(monkeypatch):
+    from vllm.model_executor.layers.linear import UnquantizedLinearMethod
+
+    class FakeRow(torch.nn.Module):
+        input_size = 4
+        output_size = 2
+        input_size_per_partition = 2
+        input_is_parallel = True
+        tp_rank = 0
+        tp_size = 2
+        reduce_results = True
+        skip_bias_add = False
+        return_bias = False
+
+        def __init__(self):
+            super().__init__()
+            self.weight = torch.nn.Parameter(torch.eye(2, dtype=torch.bfloat16))
+            self.bias = torch.nn.Parameter(torch.tensor([0.25, -0.5], dtype=torch.bfloat16))
+            self.quant_method = UnquantizedLinearMethod()
+
+    monkeypatch.setattr(ltx2_phase_adapter, "RowParallelLinear", FakeRow)
+    reductions = []
+
+    def all_reduce(value):
+        reductions.append(value.clone())
+        return value + 1
+
+    monkeypatch.setattr(ltx2_phase_adapter, "tensor_model_parallel_all_reduce", all_reduce)
+    layer = FakeRow()
+    lora_a = torch.tensor([[0.5, -0.25, 0.75, -0.125]], dtype=torch.bfloat16)
+    lora_b = torch.tensor([[0.25], [-0.5]], dtype=torch.bfloat16)
+    wrapper = _wrap(layer, lora_a, lora_b, layer_fused=True)
+
+    expected_weight = torch.matmul(lora_b, lora_a[:, :2])
+    expected_weight.add_(layer.weight)
+    x = torch.tensor([[1.0, 2.0]], dtype=torch.bfloat16)
+    expected_local = torch.nn.functional.linear(x, expected_weight, layer.bias)
+
+    actual = wrapper(x)
+    assert len(reductions) == 1
+    assert torch.equal(reductions[0], expected_local)
+    assert torch.equal(actual, expected_local + 1)
+
+
+def test_ltx_layer_fused_column_parallel_uses_local_weight_shard(monkeypatch):
+    from vllm.model_executor.layers.linear import UnquantizedLinearMethod
+
+    class FakeColumn(torch.nn.Module):
+        input_size = 3
+        output_size = 4
+        output_size_per_partition = 2
+        tp_rank = 1
+        tp_size = 2
+        gather_output = False
+        skip_bias_add = False
+        return_bias = False
+
+        def __init__(self):
+            super().__init__()
+            self.weight = torch.nn.Parameter(
+                torch.tensor([[0.5, -0.25, 0.125], [-0.375, 0.625, 0.25]], dtype=torch.bfloat16)
+            )
+            self.bias = torch.nn.Parameter(torch.tensor([0.25, -0.5], dtype=torch.bfloat16))
+            self.quant_method = UnquantizedLinearMethod()
+
+    monkeypatch.setattr(ltx2_phase_adapter, "ColumnParallelLinear", FakeColumn)
+    layer = FakeColumn()
+    base_weight = layer.weight.detach().clone()
+    lora_a = torch.tensor([[0.5, -0.25, 0.75]], dtype=torch.bfloat16)
+    lora_b = torch.tensor([[0.25], [-0.5], [0.75], [-1.0]], dtype=torch.bfloat16)
+    wrapper = _wrap(layer, lora_a, lora_b, layer_fused=True)
+
+    expected_weight = torch.matmul(lora_b[2:], lora_a)
+    expected_weight.add_(base_weight)
+    x = torch.tensor([[1.0, -0.5, 0.25]], dtype=torch.bfloat16)
+    expected = torch.nn.functional.linear(x, expected_weight, layer.bias)
+
+    assert torch.equal(wrapper(x), expected)
+    assert torch.equal(layer.weight, base_weight)
+
+
+def test_ltx_layer_fused_packed_qkv_updates_only_the_target_slice(monkeypatch):
+    from vllm.model_executor.layers.linear import UnquantizedLinearMethod
+
+    class FakeColumn(torch.nn.Module):
+        pass
+
+    class FakeQKV(FakeColumn):
+        input_size = 4
+        head_size = 2
+        v_head_size = 2
+        total_num_heads = 2
+        total_num_kv_heads = 2
+        num_heads = 1
+        num_kv_heads = 1
+        num_kv_head_replicas = 1
+        tp_rank = 1
+        tp_size = 2
+        gather_output = False
+        skip_bias_add = True
+        return_bias = True
+
+        def __init__(self):
+            super().__init__()
+            self.weight = torch.nn.Parameter(
+                torch.arange(24, dtype=torch.float32).reshape(6, 4).to(torch.bfloat16) / 16
+            )
+            self.bias = torch.nn.Parameter(torch.arange(6, dtype=torch.float32).to(torch.bfloat16) / 8)
+            self.quant_method = UnquantizedLinearMethod()
+
+    monkeypatch.setattr(ltx2_phase_adapter, "ColumnParallelLinear", FakeColumn)
+    monkeypatch.setattr(ltx2_phase_adapter, "QKVParallelLinear", FakeQKV)
+    layer = FakeQKV()
+    base_weight = layer.weight.detach().clone()
+    lora_a = torch.tensor([[0.5, -0.25, 0.75, -0.125]], dtype=torch.bfloat16)
+    lora_b = torch.tensor([[0.25], [-0.5], [0.75], [-1.0]], dtype=torch.bfloat16)
+    wrapper = _wrap(
+        layer,
+        lora_a,
+        lora_b,
+        source_module="to_k",
+        module="to_qkv",
+        target_slice="k",
+        layer_fused=True,
+    )
+
+    expected_weight = base_weight.clone()
+    expected_k = torch.matmul(lora_b[2:], lora_a)
+    expected_k.add_(base_weight[2:4])
+    expected_weight[2:4].copy_(expected_k)
+    x = torch.tensor([[1.0, -0.5, 0.25, -0.125]], dtype=torch.bfloat16)
+    expected_output = torch.nn.functional.linear(x, expected_weight)
+
+    actual_output, returned_bias = wrapper(x)
+    assert torch.equal(actual_output, expected_output)
+    assert returned_bias is layer.bias
+    assert torch.equal(layer.weight, base_weight)
+
+
+@pytest.mark.parametrize(
+    ("layer_factory", "target_slice", "expected_a", "expected_b", "expected_output"),
+    [
+        (
+            lambda: type(
+                "FakeColumn",
+                (torch.nn.Module,),
+                {
+                    "input_size": 4,
+                    "output_size": 4,
+                    "output_size_per_partition": 2,
+                    "tp_rank": 1,
+                    "gather_output": False,
+                    "forward": lambda self, x: x.new_zeros((*x.shape[:-1], 2)),
+                },
+            )(),
+            None,
+            torch.tensor([[1.0, 10.0, 100.0, 1000.0]]),
+            torch.tensor([[1.0], [2.0], [3.0], [4.0]]),
+            torch.tensor([[12963.0, 17284.0]]),
+        ),
+        (
+            lambda: type(
+                "FakeRow",
+                (torch.nn.Module,),
+                {
+                    "input_size": 4,
+                    "output_size": 2,
+                    "input_size_per_partition": 2,
+                    "tp_rank": 1,
+                    "reduce_results": False,
+                    "forward": lambda self, x: x.new_zeros((*x.shape[:-1], 2)),
+                },
+            )(),
+            None,
+            torch.tensor([[1.0, 10.0, 100.0, 1000.0]]),
+            torch.tensor([[2.0], [3.0]]),
+            torch.tensor([[8600.0, 12900.0]]),
+        ),
+        (
+            lambda: type(
+                "FakeQKV",
+                (torch.nn.Module,),
+                {
+                    "input_size": 4,
+                    "head_size": 2,
+                    "v_head_size": 2,
+                    "total_num_heads": 2,
+                    "total_num_kv_heads": 2,
+                    "num_heads": 1,
+                    "num_kv_heads": 1,
+                    "num_kv_head_replicas": 1,
+                    "tp_rank": 1,
+                    "forward": lambda self, x: x.new_zeros((*x.shape[:-1], 6)),
+                },
+            )(),
+            "k",
+            torch.tensor([[1.0, 10.0, 100.0, 1000.0]]),
+            torch.tensor([[1.0], [2.0], [3.0], [4.0]]),
+            torch.tensor([[0.0, 0.0, 12963.0, 17284.0, 0.0, 0.0]]),
+        ),
+    ],
+)
+def test_ltx_phase_adapter_uses_local_tp_lora_slices(
+    monkeypatch,
+    layer_factory,
+    target_slice,
+    expected_a,
+    expected_b,
+    expected_output,
+):
+    layer = layer_factory()
+    layer.register_buffer("device_anchor", torch.empty(0))
+    layer_type = type(layer)
+    if target_slice == "k":
+        monkeypatch.setattr(ltx2_phase_adapter, "QKVParallelLinear", layer_type)
+    elif "Column" in layer_type.__name__:
+        monkeypatch.setattr(ltx2_phase_adapter, "ColumnParallelLinear", layer_type)
+    else:
+        monkeypatch.setattr(ltx2_phase_adapter, "RowParallelLinear", layer_type)
+
+    wrapper = _wrap(layer, expected_a, expected_b, target_slice=target_slice)
+    torch.testing.assert_close(wrapper(torch.tensor([[1.0, 2.0, 3.0, 4.0]])), expected_output)
+    expected_local_a_width = getattr(layer, "input_size_per_partition", layer.input_size)
+    assert wrapper.pieces[0].lora_a.shape[-1] == expected_local_a_width

tests/diffusion/models/ltx2/test_ltx2_pipeline.py

@@ -13,13 +13,21 @@ import pytest
 import torch
 
 from vllm_omni.diffusion.data import DiffusionOutput
-from vllm_omni.diffusion.models.ltx2 import ltx2_components
+from vllm_omni.diffusion.models.ltx2 import (
+    ltx2_components,
+    ltx2_transformer,
+)
 from vllm_omni.diffusion.models.ltx2.ltx2_components import (
     LTX2_COMPONENT_PROFILE,
     LTX2_DISTILLED_COMPONENT_PROFILE,
+    LTX2_TWO_STAGE_COMPONENT_PROFILE,
     LTX23_COMPONENT_PROFILE,
+    LTX23_DISTILLED_COMPONENT_PROFILE,
+    LTX23_TWO_STAGE_COMPONENT_PROFILE,
     _install_connector_attention,
     detect_ltx_model_version,
+    resolve_ltx_artifact,
+    resolve_ltx_component_profile,
 )
 from vllm_omni.diffusion.models.ltx2.ltx2_conditioning import LTXI2VConditioningMixin
 from vllm_omni.diffusion.models.ltx2.ltx2_denoise import (
@@ -33,6 +41,7 @@ from vllm_omni.diffusion.models.ltx2.ltx2_guidance import (
     LTXGuidancePlan,
     LTXGuidanceSpec,
     LTXModalityGuidance,
+    _repeat_batch,
     build_perturbation_kwargs,
     combine_guided_x0,
 )
@@ -40,18 +49,29 @@ from vllm_omni.diffusion.models.ltx2.ltx2_latents import LTXAVState
 from vllm_omni.diffusion.models.ltx2.ltx2_recipes import (
     LTX2_DISTILLED_TWO_STAGE_RECIPE,
     LTX2_ONE_STAGE_RECIPE,
+    LTX2_TWO_STAGE_RECIPE,
+    LTX23_DISTILLED_TWO_STAGE_RECIPE,
     LTX23_ONE_STAGE_RECIPE,
+    LTX23_TWO_STAGE_RECIPE,
     LTX_DEFAULT_NEGATIVE_PROMPT,
     LTX_POSITIVE_ONLY_RECIPE,
+    resolve_ltx_pipeline_recipe,
+)
+from vllm_omni.diffusion.models.ltx2.ltx2_request import (
+    LTXRequestInputs,
+    validate_ltx_checkpoint,
+    validate_pipeline_request,
 )
-from vllm_omni.diffusion.models.ltx2.ltx2_request import LTXRequestInputs, validate_pipeline_request
 from vllm_omni.diffusion.models.ltx2.ltx2_runtime import LTXRuntime
 from vllm_omni.diffusion.models.ltx2.pipeline_ltx2 import (
     LTX2I2VDMD2Pipeline,
     LTX2Pipeline,
     LTX2T2VDMD2Pipeline,
 )
-from vllm_omni.diffusion.models.ltx2.pipeline_ltx2_two_stage import LTX2DistilledPipeline
+from vllm_omni.diffusion.models.ltx2.pipeline_ltx2_two_stage import (
+    LTX2DistilledPipeline,
+    LTX2TwoStagePipeline,
+)
 
 pytestmark = [pytest.mark.core_model, pytest.mark.cpu]
 
@@ -104,10 +124,22 @@ def test_ltx_public_entries_share_runtime_and_keep_recipe_boundaries():
     )
 
     assert issubclass(LTX2Pipeline, LTXRuntime)
+    assert issubclass(LTX2TwoStagePipeline, LTXRuntime)
     assert issubclass(LTX2DistilledPipeline, LTXRuntime)
     assert LTX2Pipeline.component_profile is LTX2_COMPONENT_PROFILE
     assert LTX2Pipeline.pipeline_recipe is LTX2_ONE_STAGE_RECIPE
+    assert LTX2TwoStagePipeline.component_profile is LTX2_TWO_STAGE_COMPONENT_PROFILE
+    assert LTX2TwoStagePipeline.pipeline_recipe is LTX2_TWO_STAGE_RECIPE
     assert LTX2DistilledPipeline.component_profile is LTX2_DISTILLED_COMPONENT_PROFILE
+    for pipeline_cls, profile in (
+        (LTX2Pipeline, LTX2_COMPONENT_PROFILE),
+        (LTX2TwoStagePipeline, LTX2_TWO_STAGE_COMPONENT_PROFILE),
+        (LTX2DistilledPipeline, LTX2_DISTILLED_COMPONENT_PROFILE),
+    ):
+        assert pipeline_cls._dit_modules == list(profile.dit_modules)
+        assert pipeline_cls._encoder_modules == list(profile.encoder_modules)
+        assert pipeline_cls._vae_modules == list(profile.vae_modules)
+        assert pipeline_cls._resident_modules == list(profile.resident_modules)
     assert LTX2_COMPONENT_PROFILE.video_vae_cls is DistributedAutoencoderKLLTX2Video
     assert LTX23_COMPONENT_PROFILE.video_vae_cls is DistributedAutoencoderKLLTX2Video
 
@@ -123,6 +155,72 @@ def test_ltx_checkpoint_version_detection_uses_metadata(tmp_path):
     assert detect_ltx_model_version(str(tmp_path)) == "2"
 
 
+def test_ltx_artifact_falls_back_to_unpinned_hub_download(tmp_path, monkeypatch):
+    filename = "ltx-sidecar.safetensors"
+    calls = []
+
+    def fake_download(**kwargs):
+        calls.append(kwargs)
+        return "/cache/ltx-sidecar.safetensors"
+
+    monkeypatch.setattr(ltx2_components, "hf_hub_download", fake_download)
+
+    assert (
+        resolve_ltx_artifact(str(tmp_path / "model"), "Lightricks/test", filename) == "/cache/ltx-sidecar.safetensors"
+    )
+    assert calls == [{"repo_id": "Lightricks/test", "filename": filename}]
+
+
+def test_ltx_artifact_prefers_model_root(tmp_path, monkeypatch):
+    filename = "ltx-sidecar.safetensors"
+    expected = tmp_path / filename
+    expected.write_bytes(b"sidecar")
+    monkeypatch.setattr(ltx2_components, "hf_hub_download", lambda **_kwargs: pytest.fail("unexpected Hub lookup"))
+
+    assert resolve_ltx_artifact(str(tmp_path), "Lightricks/test", filename) == str(expected)
+
+
+@pytest.mark.parametrize(
+    ("expected_kind", "scheduler_config"),
+    [
+        ("regular", {"use_dynamic_shifting": True, "shift_terminal": 0.1}),
+        ("distilled", {"use_dynamic_shifting": False, "shift_terminal": None}),
+    ],
+)
+def test_ltx_checkpoint_validation_accepts_matching_scheduler_metadata(expected_kind, scheduler_config):
+    validate_ltx_checkpoint(
+        scheduler_config,
+        expected_kind=expected_kind,
+        pipeline_name="TestLTXPipeline",
+    )
+
+
+@pytest.mark.parametrize(
+    ("expected_kind", "scheduler_config", "error"),
+    [
+        ("regular", {"use_dynamic_shifting": False, "shift_terminal": None}, "regular non-distilled"),
+        ("distilled", {"use_dynamic_shifting": True, "shift_terminal": 0.1}, "merged distilled"),
+    ],
+)
+def test_ltx_checkpoint_validation_rejects_mismatched_scheduler_metadata(expected_kind, scheduler_config, error):
+    with pytest.raises(ValueError, match=error):
+        validate_ltx_checkpoint(
+            scheduler_config,
+            expected_kind=expected_kind,
+            pipeline_name="TestLTXPipeline",
+        )
+
+
+def test_ltx_rejects_advanced_uaa_before_component_initialization():
+    od_config = SimpleNamespace(
+        model="unused",
+        parallel_config=SimpleNamespace(ulysses_mode="advanced_uaa"),
+    )
+
+    with pytest.raises(ValueError, match="does not support ulysses_mode='advanced_uaa'"):
+        LTX2Pipeline(od_config=od_config)
+
+
 def test_ltx23_checkpoint_selects_version_specific_one_stage_profile(tmp_path, monkeypatch):
     from vllm_omni.diffusion.models.ltx2 import ltx2_runtime
 
@@ -144,16 +242,104 @@ def test_ltx23_checkpoint_selects_version_specific_one_stage_profile(tmp_path, m
     assert pipe.reports_stage_durations
 
 
-def test_ltx_one_stage_entry_exposes_batch_image_and_distributed_decode_contracts():
+def test_ltx_multistage_rejects_cache_dit_before_component_initialization(tmp_path, monkeypatch):
+    from vllm_omni.diffusion.models.ltx2 import ltx2_runtime
+
+    (tmp_path / "model_index.json").write_text(json.dumps({"vocoder": ["ltx2", "LTX2Vocoder"]}))
+    initialized = False
+
+    def stub_components(*_args, **_kwargs):
+        nonlocal initialized
+        initialized = True
+
+    monkeypatch.setattr(ltx2_runtime, "initialize_pipeline_components", stub_components)
+
+    with pytest.raises(ValueError, match="only one-stage LTX recipes"):
+        LTX2TwoStagePipeline(
+            od_config=SimpleNamespace(
+                model=str(tmp_path),
+                cache_backend="cache_dit",
+            )
+        )
+
+    assert not initialized
+
+
+def test_ltx23_checkpoint_selects_version_specific_two_stage_profiles(tmp_path, monkeypatch):
+    from vllm_omni.diffusion.models.ltx2 import ltx2_runtime
+
+    (tmp_path / "model_index.json").write_text(json.dumps({"vocoder": ["ltx2", "LTX2VocoderWithBWE"]}))
+
+    def stub_components(pipe, od_config):
+        pipe.od_config = od_config
+        pipe.vae_spatial_compression_ratio = 32
+
+    monkeypatch.setattr(ltx2_runtime, "initialize_pipeline_components", stub_components)
+    monkeypatch.setattr(LTXRuntime, "setup_diffusion_pipeline_profiler", lambda *_args, **_kwargs: None)
+    monkeypatch.setattr(ltx2_runtime, "build_ltx_phase_adapter", lambda *_args: SimpleNamespace())
+    pipe = LTX2TwoStagePipeline(
+        od_config=SimpleNamespace(
+            model=str(tmp_path),
+            enable_diffusion_pipeline_profiler=False,
+            lora_path=None,
+            model_config={},
+        )
+    )
+
+    assert pipe.model_version == "2.3"
+    assert pipe.component_profile is LTX23_TWO_STAGE_COMPONENT_PROFILE
+    assert pipe.pipeline_recipe is LTX23_TWO_STAGE_RECIPE
+
+
+def test_ltx23_checkpoint_selects_full_distilled_profile_and_recipe(tmp_path, monkeypatch):
+    from vllm_omni.diffusion.models.ltx2 import ltx2_runtime
+
+    (tmp_path / "model_index.json").write_text(json.dumps({"vocoder": ["ltx2", "LTX2VocoderWithBWE"]}))
+
+    def stub_components(pipe, od_config):
+        pipe.od_config = od_config
+        pipe.vae_spatial_compression_ratio = 32
+
+    monkeypatch.setattr(ltx2_runtime, "initialize_pipeline_components", stub_components)
+    monkeypatch.setattr(LTXRuntime, "setup_diffusion_pipeline_profiler", lambda *_args, **_kwargs: None)
+
+    pipe = LTX2DistilledPipeline(
+        od_config=SimpleNamespace(
+            model=str(tmp_path),
+            enable_diffusion_pipeline_profiler=False,
+        )
+    )
+
+    assert pipe.model_version == "2.3"
+    assert pipe.component_profile is LTX23_DISTILLED_COMPONENT_PROFILE
+    assert pipe.pipeline_recipe is LTX23_DISTILLED_TWO_STAGE_RECIPE
+    assert pipe._phase_adapter is None
+    assert pipe.preserve_sp_padded_audio_duration
+    assert pipe.reports_stage_durations
+
+
+def test_ltx_entries_expose_batch_image_and_distributed_decode_contracts():
     assert LTX2Pipeline.supports_request_batch
     assert LTX2Pipeline.support_image_input
     assert LTX2Pipeline.unified_text_image_entry
     assert LTX2Pipeline.distributed_video_decode
+    assert not LTX2TwoStagePipeline.supports_request_batch
+    assert LTX2TwoStagePipeline.support_image_input
+    assert LTX2TwoStagePipeline.unified_text_image_entry
     assert not LTX2DistilledPipeline.supports_request_batch
     assert LTX2DistilledPipeline.support_image_input
     assert LTX2DistilledPipeline.unified_text_image_entry
 
 
+def test_ltx_two_stage_resolves_request_embedded_images():
+    pipe = object.__new__(LTX2TwoStagePipeline)
+    request_inputs = SimpleNamespace(latents=None)
+    image = torch.zeros(3, 8, 8)
+    req = SimpleNamespace(prompts=[{"prompt": "image conditioned", "multi_modal_data": {"image": image}}])
+
+    assert pipe._resolve_request_image(req, None, request_inputs) is image
+
+
 def test_ltx_dmd2_entries_retain_their_task_boundaries():
     assert not LTX2T2VDMD2Pipeline.support_image_input
     assert LTX2I2VDMD2Pipeline.support_image_input
@@ -330,6 +516,15 @@ def test_ltx_guidance_uses_official_close_to_disabled_semantics():
     assert LTXGuidancePlan.build(LTXGuidanceSpec(video=guidance)).names == ("cond",)
 
 
+def test_ltx_positive_only_guidance_preserves_token_major_layout():
+    tensor = torch.arange(24).view(1, 3, 8).transpose(1, 2).contiguous().transpose(1, 2)
+
+    repeated = _repeat_batch(tensor, 1)
+
+    assert repeated is tensor
+    assert repeated.stride() == tensor.stride()
+
+
 def test_ltx_guidance_rescale_is_invariant_to_request_batching():
     cond = torch.tensor(
         [
@@ -380,6 +575,32 @@ def test_ltx_guidance_rescale_handles_zero_variance_prediction():
     torch.testing.assert_close(actual, torch.zeros_like(cond))
 
 
+def test_ltx_audio_guidance_rescale_ignores_sp_padding_tokens():
+    logical_cond = torch.tensor([[[1.0, 2.0], [3.0, 5.0]]])
+    logical_uncond = torch.tensor([[[0.5, -1.0], [2.0, 1.0]]])
+    guidance = LTXModalityGuidance(cfg_scale=3.0, rescale_scale=1.0)
+    expected = combine_guided_x0(
+        cond=logical_cond,
+        uncond_text=logical_uncond,
+        uncond_perturbed=0.0,
+        uncond_modality=0.0,
+        guidance=guidance,
+    )
+    padded_cond = torch.cat([logical_cond, torch.tensor([[[1000.0, -1000.0]]])], dim=1)
+    padded_uncond = torch.cat([logical_uncond, torch.tensor([[[-500.0, 500.0]]])], dim=1)
+
+    actual = combine_guided_x0(
+        cond=padded_cond,
+        uncond_text=padded_uncond,
+        uncond_perturbed=0.0,
+        uncond_modality=0.0,
+        guidance=guidance,
+        rescale_token_count=2,
+    )
+
+    torch.testing.assert_close(actual[:, :2], expected)
+
+
 def test_ltx_guidance_builds_per_sample_stg_and_modality_masks():
     plan = LTXGuidancePlan.build(LTX23_ONE_STAGE_RECIPE.request_guidance)
     perturbations = build_perturbation_kwargs(plan, batch_size=2, reference=torch.ones(8, 1, 1))
@@ -547,11 +768,19 @@ def test_ltx_connector_attention_dispatches_through_omni_kernel(has_learned_regi
 
 def test_ltx2_distilled_pipeline_shares_recipe_runtime():
     assert issubclass(LTX2DistilledPipeline, LTXI2VConditioningMixin)
-    assert issubclass(LTX2DistilledPipeline, LTX2Pipeline)
+    assert not issubclass(LTX2DistilledPipeline, LTX2Pipeline)
     assert LTX2DistilledPipeline._run_recipe is LTXRuntime._run_recipe
     assert "_forward_request" not in LTX2DistilledPipeline.__dict__
 
 
+def test_ltx_two_stage_entry_is_a_thin_pipeline_variant():
+    assert "_run_recipe" not in LTX2TwoStagePipeline.__dict__
+    assert "_forward_request" not in LTX2TwoStagePipeline.__dict__
+    assert "load_weights" not in LTX2TwoStagePipeline.__dict__
+    assert issubclass(LTX2TwoStagePipeline, LTXI2VConditioningMixin)
+    assert not issubclass(LTX2TwoStagePipeline, LTX2Pipeline)
+
+
 def test_ltx_variants_share_denoise_loop_and_i2v_conditioning():
     assert "_denoise_loop" not in LTX2Pipeline.__dict__
     assert issubclass(LTX2Pipeline, LTXI2VConditioningMixin)
@@ -594,7 +823,8 @@ def test_denoise_executor_owns_progress_and_interrupt():
     torch.testing.assert_close(state.audio, torch.tensor(11.0))
 
 
-def test_ltx2_distilled_two_stage_executes_declarative_phase_plan():
+@pytest.mark.parametrize("pipeline_cls", [LTX2TwoStagePipeline, LTX2DistilledPipeline])
+def test_ltx_two_stage_executes_declarative_i2v_phase_plan(pipeline_cls):
     request_inputs = LTXRequestInputs(
         prompt="prompt",
         negative_prompt="",
@@ -671,12 +901,15 @@ def test_ltx2_distilled_two_stage_executes_declarative_phase_plan():
             return latents.repeat_interleave(2, dim=-2).repeat_interleave(2, dim=-1)
 
     prompt_context_sentinel = prompt_context
-    pipeline = object.__new__(LTX2DistilledPipeline)
+    pipeline = object.__new__(pipeline_cls)
     torch.nn.Module.__init__(pipeline)
     pipeline.device = torch.device("cpu")
     pipeline.vae_spatial_compression_ratio = 32
     pipeline.vae_temporal_compression_ratio = 8
     pipeline.latent_upsampler = FakeUpsampler()
+    activated_slots = []
+    if pipeline_cls is LTX2TwoStagePipeline:
+        pipeline._phase_adapter = SimpleNamespace(activate=activated_slots.append)
     object.__setattr__(pipeline, "_resolve_request_inputs", resolve_request_inputs)
     object.__setattr__(pipeline, "run_phase", run_phase)
     object.__setattr__(pipeline, "decode_phase", decode_phase)
@@ -686,14 +919,18 @@ def test_ltx2_distilled_two_stage_executes_declarative_phase_plan():
 
     assert len(phase_calls) == 2
     assert phase_calls[1][2] is prompt_context_sentinel
+    if pipeline_cls is LTX2TwoStagePipeline:
+        assert activated_slots == [None, "ltx_distilled"]
     torch.testing.assert_close(output.output[0], torch.full((1, 128, 1, 2, 2), 3.0))
-    torch.testing.assert_close(output.output[1], torch.full((1, 8, 1, 2), 4.0))
+    expected_audio = 2.0 if pipeline_cls is LTX2TwoStagePipeline else 4.0
+    torch.testing.assert_close(output.output[1], torch.full((1, 8, 1, 2), expected_audio))
 
 
-def test_ltx2_distilled_stage2_reapplies_final_resolution_i2v_conditioning():
+@pytest.mark.parametrize("pipeline_cls", [LTX2TwoStagePipeline, LTX2DistilledPipeline])
+def test_ltx_two_stage_stage2_reapplies_final_resolution_i2v_conditioning(pipeline_cls):
     from vllm_omni.diffusion.models.ltx2 import ltx2_latents
 
-    pipeline = object.__new__(LTX2DistilledPipeline)
+    pipeline = object.__new__(pipeline_cls)
     torch.nn.Module.__init__(pipeline)
     pipeline.vae_spatial_compression_ratio = 32
     pipeline.vae_temporal_compression_ratio = 8
@@ -750,6 +987,18 @@ def test_ltx2_distilled_two_stage_recipe_is_fixed_positive_only():
     assert (LTX2_DISTILLED_TWO_STAGE_RECIPE.height, LTX2_DISTILLED_TWO_STAGE_RECIPE.width) == (1024, 1536)
 
 
+def test_ltx23_full_distilled_route_uses_merged_weights_without_lora():
+    assert resolve_ltx_component_profile("distilled_two_stage", "2.3") is LTX23_DISTILLED_COMPONENT_PROFILE
+    assert resolve_ltx_pipeline_recipe("distilled_two_stage", "2.3") is LTX23_DISTILLED_TWO_STAGE_RECIPE
+    assert LTX23_DISTILLED_TWO_STAGE_RECIPE is LTX2_DISTILLED_TWO_STAGE_RECIPE
+    assert all(phase.adapter_slot is None for phase in LTX23_DISTILLED_TWO_STAGE_RECIPE.phases)
+    assert LTX23_DISTILLED_COMPONENT_PROFILE.distilled_lora_filename is None
+    assert LTX23_DISTILLED_COMPONENT_PROFILE.vocoder_cls is LTX23_COMPONENT_PROFILE.vocoder_cls
+    assert LTX23_DISTILLED_COMPONENT_PROFILE.latent_upsampler_filename == (
+        "ltx-2.3-spatial-upscaler-x2-1.1.safetensors"
+    )
+
+
 @pytest.mark.parametrize(
     ("request_inputs", "error"),
     [
@@ -958,6 +1207,98 @@ def test_ltx_custom_sigmas_bypass_scheduler_shifting():
     torch.testing.assert_close(timesteps, expected[:-1] * scheduler.config.num_train_timesteps)
 
 
+@pytest.mark.parametrize(
+    ("recipe", "steps", "stg_block"),
+    [
+        (LTX2_TWO_STAGE_RECIPE, 40, 29),
+        (LTX23_TWO_STAGE_RECIPE, 30, 28),
+    ],
+)
+def test_ltx_official_two_stage_recipes_only_vary_by_model_defaults(recipe, steps, stg_block):
+    stage1, stage2 = recipe.phases
+    assert (recipe.width, recipe.height) == (1536, 1024)
+    assert recipe.num_inference_steps == steps
+    assert stage1.spatial_downscale == 2
+    assert stage1.adapter_slot is None
+    assert stage1.guidance.video.stg_blocks == (stg_block,)
+    assert stage1.guidance.audio.stg_blocks == (stg_block,)
+    assert stage2.input_transform == "spatial_upsample"
+    assert stage2.adapter_slot == "ltx_distilled"
+    assert stage2.guidance == LTXGuidanceSpec.positive_only()
+    assert stage2.num_inference_steps == 3
+    assert stage2.sigmas == (0.909375, 0.725, 0.421875, 0.0)
+    # Official Stage 2 refines video only; its audio result is discarded.
+    assert (recipe.video_output_phase, recipe.audio_output_phase) == (1, 0)
+
+
+def test_ltx_two_stage_entries_select_the_official_adapter_slots():
+    phase_switches = []
+    ordinary_pipeline = object.__new__(LTX2TwoStagePipeline)
+    ordinary_pipeline._phase_adapter = SimpleNamespace(activate=phase_switches.append)
+    for phase in LTX2_TWO_STAGE_RECIPE.phases:
+        ordinary_pipeline._enter_phase(phase)
+    assert phase_switches == [None, "ltx_distilled"]
+
+    distilled_pipeline = object.__new__(LTX2DistilledPipeline)
+    for phase in LTX2_DISTILLED_TWO_STAGE_RECIPE.phases:
+        distilled_pipeline._enter_phase(phase)
+
+
+def test_ltx23_gate_projection_declares_its_global_column_parallel_layout(monkeypatch):
+    class FakeColumn(torch.nn.Module):
+        def __init__(
+            self,
+            input_size,
+            output_size,
+            *,
+            bias,
+            gather_output,
+            return_bias,
+            quant_config,
+            prefix,
+        ):
+            super().__init__()
+            self.input_size = input_size
+            self.output_size = output_size
+            self.bias_enabled = bias
+            self.gather_output = gather_output
+            self.return_bias = return_bias
+            self.quant_config = quant_config
+            self.prefix = prefix
+
+    class FakeRow(torch.nn.Module):
+        def __init__(self, *args, **kwargs):
+            super().__init__()
+
+    class FakeAttention(torch.nn.Module):
+        def __init__(self, **kwargs):
+            super().__init__()
+
+    monkeypatch.setattr(ltx2_transformer, "ColumnParallelLinear", FakeColumn)
+    monkeypatch.setattr(ltx2_transformer, "RowParallelLinear", FakeRow)
+    monkeypatch.setattr(ltx2_transformer, "_LTX2ParallelAttention", FakeAttention)
+    monkeypatch.setattr(ltx2_transformer, "get_tensor_model_parallel_world_size", lambda: 2)
+
+    attention = ltx2_transformer.LTX2Attention(
+        query_dim=16,
+        heads=4,
+        dim_head=4,
+        qk_norm="rms_norm",
+        apply_gated_attention=True,
+        pack_qkv=False,
+        prefix="transformer_blocks.0.attn1",
+    )
+
+    gate = attention.to_gate_logits
+    assert isinstance(gate, FakeColumn)
+    assert gate.input_size == 16
+    assert gate.output_size == 4
+    assert gate.gather_output is False
+    assert gate.return_bias is False
+    assert gate.quant_config is None
+    assert gate.prefix == "transformer_blocks.0.attn1.to_gate_logits"
+
+
 class TestLTXRequestParsing:
     def test_request_resolves_official_and_explicit_negative_prompts_per_sample(self):
         from vllm_omni.diffusion.request import OmniDiffusionRequest
@@ -1432,6 +1773,7 @@ class TestPipelineComponents:
 
     def test_distilled_profile_keeps_upsampler_in_managed_resident_components(self):
         assert LTX2_DISTILLED_COMPONENT_PROFILE.resident_modules == ("vocoder", "latent_upsampler")
+        assert LTX23_DISTILLED_COMPONENT_PROFILE.resident_modules == ("vocoder", "latent_upsampler")
 
 
 class TestLTX23DecodeConditioning:
@@ -1477,6 +1819,11 @@ class TestRegistryIntegration:
         from vllm_omni.diffusion.registry import _DIFFUSION_MODELS
 
         assert _DIFFUSION_MODELS["LTX2Pipeline"] == ("ltx2", "pipeline_ltx2", "LTX2Pipeline")
+        assert _DIFFUSION_MODELS["LTX2TwoStagePipeline"] == (
+            "ltx2",
+            "pipeline_ltx2_two_stage",
+            "LTX2TwoStagePipeline",
+        )
         assert _DIFFUSION_MODELS["LTX2DistilledPipeline"] == (
             "ltx2",
             "pipeline_ltx2_two_stage",
@@ -1488,6 +1835,9 @@ class TestRegistryIntegration:
             "LTX23ImageToVideoPipeline",
             "LTX2TwoStagesPipeline",
             "LTX2ImageToVideoTwoStagesPipeline",
+            "LTX2ImageToVideoTwoStagePipeline",
+            "LTX23TwoStagePipeline",
+            "LTX23ImageToVideoTwoStagePipeline",
         }
         assert removed_entries.isdisjoint(_DIFFUSION_MODELS)
 
@@ -1497,6 +1847,7 @@ class TestRegistryIntegration:
 
         expected = [
             "LTX2Pipeline",
+            "LTX2TwoStagePipeline",
             "LTX2DistilledPipeline",
             "LTX2T2VDMD2Pipeline",
             "LTX2I2VDMD2Pipeline",
@@ -1509,6 +1860,7 @@ class TestRegistryIntegration:
         "model_class_name",
         [
             "LTX2Pipeline",
+            "LTX2TwoStagePipeline",
             "LTX2DistilledPipeline",
             "LTX2T2VDMD2Pipeline",
             "LTX2I2VDMD2Pipeline",

tests/diffusion/models/ltx2/test_ltx2_transformer.py

@@ -1,13 +1,20 @@
 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
+from types import SimpleNamespace
+
 import pytest
+import torch
 from torch import nn
 
+from vllm_omni.diffusion.attention.backends.abstract import AttentionMetadata
 from vllm_omni.diffusion.cache.cachedit import CacheDiTAdapterConfig
-from vllm_omni.diffusion.models.ltx2.ltx2_transformer import LTX2VideoTransformer3DModel, _make_rms_norm
+from vllm_omni.diffusion.models.ltx2.ltx2_transformer import (
-
+    LTX2AudioVideoAttnProcessor,
-pytestmark = [pytest.mark.core_model, pytest.mark.cpu]
+    LTX2VideoTransformer3DModel,
+    _LTX2ParallelAttention,
+    _make_rms_norm,
+)
 
 pytestmark = [pytest.mark.core_model, pytest.mark.diffusion, pytest.mark.cpu]
 
@@ -66,3 +73,83 @@ def test_ltx_sp_plan_shards_video_and_audio_timesteps_together(rope_type):
     assert video_timestep.expected_dims == audio_timestep.expected_dims == 2
     assert not video_timestep.split_output
     assert not audio_timestep.split_output
+
+
+def test_ltx_sp_keeps_cross_attention_key_padding_mask(monkeypatch):
+    monkeypatch.setattr(LTX2AudioVideoAttnProcessor, "_is_sp_enabled", staticmethod(lambda: True))
+    processor = LTX2AudioVideoAttnProcessor()
+    hidden_states = torch.zeros(1, 3, 4)
+    encoder_hidden_states = torch.zeros(1, 4, 4)
+    key_padding_mask = torch.tensor([[True, True, True, False]])
+
+    actual = processor._prepare_attention_mask(
+        SimpleNamespace(),
+        hidden_states,
+        encoder_hidden_states,
+        key_padding_mask,
+        batch_size=1,
+        sequence_length=4,
+    )
+
+    torch.testing.assert_close(actual, key_padding_mask[:, None, None, :])
+
+
+def _fake_ltx_parallel_attention():
+    calls = []
+
+    class Backend:
+        @staticmethod
+        def get_name():
+            return "FLASH_ATTN"
+
+        @staticmethod
+        def supports_attention_mask():
+            return True
+
+    def run(name):
+        def forward(query, _key, _value, _metadata):
+            calls.append(name)
+            return torch.zeros_like(query)
+
+        return forward
+
+    layer = object.__new__(_LTX2ParallelAttention)
+    nn.Module.__init__(layer)
+    layer.attn_backend = Backend
+    layer.attention = SimpleNamespace(forward=run("native"))
+    layer.sdpa_fallback = SimpleNamespace(forward=run("sdpa"))
+    layer.backend_pref = "FLASH_ATTN"
+    return layer, calls
+
+
+@pytest.mark.parametrize(("query_length", "key_length"), [(4, 4), (3, 4)])
+def test_ltx_masked_cross_attention_uses_sdpa(query_length, key_length):
+    layer, calls = _fake_ltx_parallel_attention()
+    query = torch.zeros(1, query_length, 2, 4, dtype=torch.bfloat16)
+    key = value = torch.zeros(1, key_length, 2, 4, dtype=torch.bfloat16)
+    key_padding_mask = torch.tensor([[True] * (key_length - 1) + [False]])
+    metadata = AttentionMetadata(attn_mask=key_padding_mask[:, None, None, :])
+
+    layer._run_local_attention(query, key, value, metadata)
+
+    assert calls == ["sdpa"]
+
+
+def test_ltx_masked_self_attention_uses_sdpa():
+    layer, calls = _fake_ltx_parallel_attention()
+    query = key = value = torch.zeros(1, 4, 2, 4, dtype=torch.bfloat16)
+    key_padding_mask = torch.tensor([[True, True, True, False]])
+    metadata = AttentionMetadata(attn_mask=key_padding_mask[:, None, None, :])
+
+    layer._run_local_attention(query, key, value, metadata)
+
+    assert calls == ["sdpa"]
+
+
+def test_ltx_unmasked_cross_attention_keeps_native_backend():
+    layer, calls = _fake_ltx_parallel_attention()
+    query = key = value = torch.zeros(1, 4, 2, 4, dtype=torch.bfloat16)
+
+    layer._run_local_attention(query, key, value, None)
+
+    assert calls == ["native"]

tests/diffusion/models/minimax_h3/test_minimax_h3_contract.py

@@ -2,6 +2,7 @@
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 import json
+import sys
 from multiprocessing.reduction import ForkingPickler
 from types import SimpleNamespace
 from unittest.mock import Mock
@@ -1428,34 +1429,219 @@ def test_r7_r8_ref2va_video_segment_matrix(monkeypatch, tmp_path, case, start_ti
     assert transcode_calls[0][1]["target_frame_count"] == 209
 
 
-def test_ref2va_qwen_sampling_decodes_prepared_video_once(monkeypatch):
+def test_ref2va_qwen_sampling_uses_one_selective_decode(monkeypatch):
     from vllm_omni.diffusion.models.minimax_h3 import reference_video as reference_video_module
 
-    decoded = np.arange(25 * 2 * 2 * 3, dtype=np.uint8).reshape(25, 2, 2, 3)
+    decoded = np.arange(3 * 2 * 2 * 3, dtype=np.uint8).reshape(3, 2, 2, 3)
-    load_calls = []
+    decode_calls = []
     monkeypatch.setattr(
         reference_video_module,
         "_probe_video",
-        lambda _path: {"frame_count": len(decoded)},
+        lambda _path: {
+            "frame_count": 25,
+            "width": 2,
+            "height": 2,
+        },
     )
     monkeypatch.setattr(
         reference_video_module,
-        "load_video_frames",
+        "_decode_video_frames_ffmpeg",
-        lambda path: load_calls.append(path) or decoded,
+        lambda path, **kwargs: decode_calls.append((path, kwargs)) or decoded,
     )
 
     sampled = reference_video_module.sample_reference_video_frames("prepared.mp4")
 
-    assert load_calls == ["prepared.mp4"]
+    assert decode_calls == [
-    for frame, expected in zip(
+        (
-        sampled["frames"],
+            "prepared.mp4",
-        (decoded[0], decoded[12], decoded[24]),
+            {
-        strict=True,
+                "frame_count": 25,
-    ):
+                "indices": [0, 12, 24],
+                "width": 2,
+                "height": 2,
+            },
+        )
+    ]
+    for frame, expected in zip(sampled["frames"], decoded, strict=True):
         np.testing.assert_array_equal(frame, expected)
     assert sampled["block_timestamps"] == [0.25, 1.0]
 
 
+@pytest.mark.parametrize(
+    ("indices", "frame_count"),
+    [([0, 12], 25), (None, 2)],
+)
+def test_ref2va_ffmpeg_decode_reads_exact_rgb_frames(monkeypatch, indices, frame_count):
+    from vllm_omni.diffusion.models.minimax_h3 import reference_video as reference_video_module
+
+    expected = np.arange(2 * 2 * 3 * 3, dtype=np.uint8).reshape(2, 2, 3, 3)
+    observed = []
+
+    def fake_run(command, **kwargs):
+        observed.append((command, kwargs))
+        return SimpleNamespace(stdout=expected.tobytes())
+
+    monkeypatch.setattr(reference_video_module.subprocess, "run", fake_run)
+
+    actual = reference_video_module._decode_video_frames_ffmpeg(
+        "prepared.mp4",
+        frame_count=frame_count,
+        indices=indices,
+        width=3,
+        height=2,
+    )
+
+    np.testing.assert_array_equal(actual, expected)
+    command, kwargs = observed[0]
+    assert command[0] == "ffmpeg"
+    assert command[command.index("-frames:v") + 1] == "2"
+    if indices is None:
+        assert "-vf" not in command
+    else:
+        assert command[command.index("-vf") + 1] == "select=eq(n\\,0)+eq(n\\,12)"
+    assert command[-1] == "pipe:1"
+    assert kwargs == {"check": True, "capture_output": True}
+
+
+def test_ref2va_load_video_frames_uses_ffmpeg_without_decord(monkeypatch):
+    from vllm_omni.diffusion.models.minimax_h3 import reference_video as reference_video_module
+
+    expected = np.zeros((2, 2, 3, 3), dtype=np.uint8)
+    decode_calls = []
+    monkeypatch.setitem(sys.modules, "decord", None)
+    monkeypatch.setattr(
+        reference_video_module,
+        "_probe_video",
+        lambda _path: {
+            "frame_count": 2,
+            "width": 3,
+            "height": 2,
+        },
+    )
+    monkeypatch.setattr(
+        reference_video_module,
+        "_decode_video_frames_ffmpeg",
+        lambda path, **kwargs: decode_calls.append((path, kwargs)) or expected,
+    )
+
+    actual = reference_video_module.load_video_frames("prepared.mp4")
+
+    assert actual is expected
+    assert decode_calls == [
+        (
+            "prepared.mp4",
+            {
+                "frame_count": 2,
+                "width": 3,
+                "height": 2,
+            },
+        )
+    ]
+
+
+@pytest.mark.parametrize(
+    ("kwargs", "message"),
+    [
+        ({"frame_count": 0, "width": 3, "height": 2}, "video has no frames"),
+        ({"frame_count": 1, "width": 0, "height": 2}, "invalid dimensions"),
+        (
+            {"frame_count": 1, "indices": [], "width": 3, "height": 2},
+            "invalid frame indices",
+        ),
+    ],
+)
+def test_ref2va_ffmpeg_decode_rejects_invalid_metadata(kwargs, message):
+    from vllm_omni.diffusion.models.minimax_h3 import reference_video as reference_video_module
+    from vllm_omni.errors import OmniClientError
+
+    with pytest.raises(OmniClientError, match=message):
+        reference_video_module._decode_video_frames_ffmpeg("prepared.mp4", **kwargs)
+
+
+def test_ref2va_probe_uses_container_frame_count_without_scan(monkeypatch):
+    from vllm_omni.diffusion.models.minimax_h3 import reference_video as reference_video_module
+
+    calls = []
+    probe = {
+        "streams": [
+            {
+                "codec_type": "video",
+                "width": 4,
+                "height": 2,
+                "r_frame_rate": "24/1",
+                "nb_frames": "3",
+                "duration": "0.125",
+                "sample_aspect_ratio": "1:1",
+            }
+        ],
+        "format": {"duration": "0.125", "size": "123"},
+    }
+    monkeypatch.setattr(
+        reference_video_module.subprocess,
+        "run",
+        lambda command, **kwargs: calls.append((command, kwargs)) or SimpleNamespace(stdout=json.dumps(probe)),
+    )
+
+    metadata = reference_video_module._probe_video("prepared.mp4")
+
+    assert metadata["frame_count"] == 3
+    assert len(calls) == 1
+    assert "-count_frames" not in calls[0][0]
+
+
+def test_ref2va_probe_counts_frames_only_when_metadata_is_missing(monkeypatch):
+    from vllm_omni.diffusion.models.minimax_h3 import reference_video as reference_video_module
+
+    calls = []
+    stream = {
+        "codec_type": "video",
+        "width": 4,
+        "height": 2,
+        "r_frame_rate": "24/1",
+        "duration": "0.125",
+        "sample_aspect_ratio": "1:1",
+    }
+    first_probe = {
+        "streams": [stream],
+        "format": {"duration": "0.125", "size": "123"},
+    }
+    second_probe = {
+        "streams": [{**stream, "nb_read_frames": "3"}],
+        "format": {"duration": "0.125", "size": "123"},
+    }
+    monkeypatch.setattr(
+        reference_video_module.subprocess,
+        "run",
+        lambda command, **kwargs: calls.append((command, kwargs))
+        or SimpleNamespace(stdout=json.dumps(first_probe if len(calls) == 1 else second_probe)),
+    )
+
+    metadata = reference_video_module._probe_video("prepared.mp4")
+
+    assert metadata["frame_count"] == 3
+    assert len(calls) == 2
+    assert "-count_frames" in calls[1][0]
+
+
+def test_ref2va_ffmpeg_decode_rejects_incomplete_output(monkeypatch):
+    from vllm_omni.diffusion.models.minimax_h3 import reference_video as reference_video_module
+    from vllm_omni.errors import OmniClientError
+
+    monkeypatch.setattr(
+        reference_video_module.subprocess,
+        "run",
+        lambda *_args, **_kwargs: SimpleNamespace(stdout=b"short"),
+    )
+
+    with pytest.raises(OmniClientError, match="decoded 5 bytes.*expected 18"):
+        reference_video_module._decode_video_frames_ffmpeg(
+            "prepared.mp4",
+            frame_count=1,
+            width=3,
+            height=2,
+        )
+
+
 def test_ref2va_two_video_recipe_tolerates_container_rounding(monkeypatch, tmp_path):
     from vllm_omni.diffusion.models.minimax_h3 import reference_video as reference_video_module
 

tests/diffusion/models/minimax_h3/test_minimax_h3_qwen3vl_rope.py

@@ -0,0 +1,104 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+import pytest
+import torch
+
+from vllm_omni.platforms import current_omni_platform
+
+pytestmark = [pytest.mark.core_model, pytest.mark.diffusion]
+
+
+@pytest.mark.skipif(not current_omni_platform.is_npu(), reason="requires Ascend NPU")
+@pytest.mark.parametrize("num_q_heads,num_kv_heads,seq_len", [(8, 2, 16), (64, 8, 1)])
+def test_qwen3_vl_npu_rope_matches_bnsd_reference(
+    num_q_heads: int,
+    num_kv_heads: int,
+    seq_len: int,
+) -> None:
+    from vllm_omni.platforms.npu.models.minimax_h3 import _apply_rotary_pos_emb_npu
+
+    device = torch.device("npu")
+    dtype = torch.bfloat16
+    q = torch.randn(1, num_q_heads, seq_len, 128, device=device, dtype=dtype)
+    k = torch.randn(1, num_kv_heads, seq_len, 128, device=device, dtype=dtype)
+    cos = torch.randn(1, seq_len, 128, device=device, dtype=dtype)
+    sin = torch.randn(1, seq_len, 128, device=device, dtype=dtype)
+
+    q_out, k_out = _apply_rotary_pos_emb_npu(q, k, cos, sin)
+    cos_bnsd = cos.unsqueeze(1)
+    sin_bnsd = sin.unsqueeze(1)
+
+    def rotate_half(x: torch.Tensor) -> torch.Tensor:
+        x1, x2 = x.chunk(2, dim=-1)
+        return torch.cat((-x2, x1), dim=-1)
+
+    torch.testing.assert_close(
+        q_out,
+        q * cos_bnsd + rotate_half(q) * sin_bnsd,
+        atol=2e-2,
+        rtol=2e-2,
+    )
+    torch.testing.assert_close(
+        k_out,
+        k * cos_bnsd + rotate_half(k) * sin_bnsd,
+        atol=2e-2,
+        rtol=2e-2,
+    )
+
+
+@pytest.mark.skipif(not current_omni_platform.is_npu(), reason="requires Ascend NPU")
+def test_qwen3_vl_npu_rope_preserves_batched_mrope_positions() -> None:
+    """Keep per-sample Qwen3-VL M-RoPE frequencies on the BNSD NPU path."""
+    from types import SimpleNamespace
+
+    from vllm_omni.diffusion.models.minimax_h3.encoder import (
+        MiniMaxH3Qwen3VLTextRotaryEmbedding,
+    )
+    from vllm_omni.platforms.npu.models.minimax_h3 import _apply_rotary_pos_emb_npu
+
+    device = torch.device("npu")
+    dtype = torch.bfloat16
+    batch_size, seq_len, head_dim = 2, 16, 128
+    config = SimpleNamespace(
+        hidden_size=head_dim,
+        num_attention_heads=1,
+        rope_parameters={
+            "rope_theta": 1_000_000.0,
+            "mrope_section": [24, 20, 20],
+        },
+    )
+    rope = MiniMaxH3Qwen3VLTextRotaryEmbedding(config).to(device)
+    positions = torch.arange(seq_len, device=device)
+    position_ids = torch.stack(
+        (
+            torch.stack((positions, positions + 7)),
+            torch.stack((positions * 2, positions * 2 + 3)),
+            torch.stack((positions * 3, positions * 3 + 5)),
+        )
+    )
+    q = torch.randn(batch_size, 8, seq_len, head_dim, device=device, dtype=dtype)
+    k = torch.randn(batch_size, 2, seq_len, head_dim, device=device, dtype=dtype)
+    cos, sin = rope(q, position_ids)
+
+    q_out, k_out = _apply_rotary_pos_emb_npu(q, k, cos, sin)
+    cos_bnsd = cos.unsqueeze(1)
+    sin_bnsd = sin.unsqueeze(1)
+
+    def rotate_half(x: torch.Tensor) -> torch.Tensor:
+        x1, x2 = x.chunk(2, dim=-1)
+        return torch.cat((-x2, x1), dim=-1)
+
+    assert not torch.equal(cos[0], cos[1])
+    torch.testing.assert_close(
+        q_out,
+        q * cos_bnsd + rotate_half(q) * sin_bnsd,
+        atol=2e-2,
+        rtol=2e-2,
+    )
+    torch.testing.assert_close(
+        k_out,
+        k * cos_bnsd + rotate_half(k) * sin_bnsd,
+        atol=2e-2,
+        rtol=2e-2,
+    )

tests/diffusion/models/minimax_h3/test_minimax_h3_qwen3vl_swiglu.py

@@ -0,0 +1,46 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+import pytest
+import torch
+import torch.nn.functional as F
+
+from vllm_omni.platforms import current_omni_platform
+
+pytestmark = [pytest.mark.core_model, pytest.mark.diffusion]
+
+
+@pytest.mark.skipif(not current_omni_platform.is_npu(), reason="requires Ascend NPU")
+@pytest.mark.parametrize("shape", [(1, 16, 256), (2, 1, 256)])
+def test_qwen3_vl_npu_swiglu_matches_split_reference(
+    shape: tuple[int, int, int],
+) -> None:
+    from vllm_omni.platforms.npu.models.minimax_h3 import npu_swiglu_from_packed
+
+    packed = torch.randn(*shape, device="npu", dtype=torch.bfloat16)
+    gate, up = packed.chunk(2, dim=-1)
+
+    torch.testing.assert_close(
+        npu_swiglu_from_packed(packed),
+        F.silu(gate) * up,
+        atol=2e-2,
+        rtol=2e-2,
+    )
+
+
+@pytest.mark.skipif(not current_omni_platform.is_npu(), reason="requires Ascend NPU")
+def test_qwen3_vl_npu_swiglu_matches_two_projection_reference() -> None:
+    from vllm_omni.platforms.npu.models.minimax_h3 import npu_swiglu_from_packed
+
+    x = torch.randn(1, 16, 128, device="npu", dtype=torch.bfloat16)
+    weight = torch.randn(256, 128, device="npu", dtype=torch.bfloat16)
+    packed = F.linear(x, weight)
+    gate = F.linear(x, weight[:128])
+    up = F.linear(x, weight[128:])
+
+    torch.testing.assert_close(
+        npu_swiglu_from_packed(packed),
+        F.silu(gate) * up,
+        atol=2e-2,
+        rtol=2e-2,
+    )

tests/diffusion/quantization/test_hsdp_fp8.py

@@ -36,15 +36,23 @@ class _ToyKernel:
 # --- helper to build a toy module simulating online-FP8 post-load state ---
 
 
-def _make_fp8_toy_module(out_features: int = 16, in_features: int = 32):
+def _make_fp8_toy_module(
+    out_features: int = 16,
+    in_features: int = 32,
+    *,
+    quant_method_cls=None,
+):
     """Return a module whose weight is ``qweight.t()`` (non-contiguous) and whose
     ``quant_method`` is an ``Fp8LinearMethod`` with a ``_ToyKernel``.
     """
     from vllm.model_executor.layers.quantization.fp8 import Fp8LinearMethod
 
+    if quant_method_cls is None:
+        quant_method_cls = Fp8LinearMethod
+
     # Create the quant-method instance without calling __init__ (avoids
     # heavyweight upstream config dependencies).  isinstance() still works.
-    qm = object.__new__(Fp8LinearMethod)
+    qm = object.__new__(quant_method_cls)
     qm.fp8_linear = _ToyKernel()
 
     # randn does not support float8 on CPU; zeros does and the test only
@@ -104,6 +112,28 @@ def test_rewrites_non_contiguous_weight():
     assert module.weight.shape == (16, 32)
 
 
+def test_rewrites_per_tensor_online_fp8_weight():
+    """Regression test for vLLM's new online FP8 frontend (#45463)."""
+    from vllm.model_executor.layers.quantization.online.fp8 import (
+        Fp8PerTensorOnlineLinearMethod,
+    )
+
+    module = _make_fp8_toy_module(
+        16,
+        32,
+        quant_method_cls=Fp8PerTensorOnlineLinearMethod,
+    )
+
+    n_rewritten_layers = prepare_fp8_layers_for_fsdp(module)
+
+    assert n_rewritten_layers == 1
+    assert module.weight.is_contiguous()
+    assert module.weight.shape == (16, 32)
+    w, *_ = module.quant_method.fp8_linear._get_layer_params(module)
+    assert w.shape == (32, 16)
+    assert w.stride() == (1, 32)
+
+
 def test_patched_get_layer_params_returns_transposed_view():
     module = _make_fp8_toy_module(16, 32)
     prepare_fp8_layers_for_fsdp(module)

tests/diffusion/quantization/test_quantization_fp8.py

@@ -199,7 +199,7 @@ def _generate_bagel_image(
     config_path = get_deploy_config_path("ci/bagel.yaml")
     omni_kwargs: dict[str, Any] = {
         "model": "ByteDance-Seed/BAGEL-7B-MoT",
-        "stage_configs_path": config_path,
+        "deploy_config": config_path,
         "stage_init_timeout": 300,
     }
     if diffusion_quantization_config:

tests/distributed/omni_connectors/test_bagel_mooncake_connector.py

@@ -314,7 +314,7 @@ def test_bagel_text2img_mooncake_connector(run_level):
         temp_config_file = _resolve_deploy_config(temp_config_file, run_level)
         with OmniRunner(
             "ByteDance-Seed/BAGEL-7B-MoT",
-            stage_configs_path=temp_config_file,
+            deploy_config=temp_config_file,
             stage_init_timeout=300,
         ) as runner:
             generated_image = _generate_bagel_image(runner.omni)

tests/distributed/omni_connectors/test_bagel_shared_memory_connector.py

@@ -251,7 +251,7 @@ def test_bagel_img2img_shared_memory_connector(run_level):
     config_path = _resolve_deploy_config(BAGEL_CI_DEPLOY, run_level)
     with OmniRunner(
         "ByteDance-Seed/BAGEL-7B-MoT",
-        stage_configs_path=config_path,
+        deploy_config=config_path,
     ) as runner:
         generated_image = _generate_bagel_img2img(runner.omni, input_image)
         if run_level == "advanced_model":
@@ -266,7 +266,7 @@ def test_bagel_text2img_shared_memory_connector(run_level):
     config_path = _resolve_deploy_config(BAGEL_CI_DEPLOY, run_level)
     with OmniRunner(
         "ByteDance-Seed/BAGEL-7B-MoT",
-        stage_configs_path=config_path,
+        deploy_config=config_path,
     ) as runner:
         generated_image = _generate_bagel_text2img(runner.omni)
         if run_level == "advanced_model":

tests/e2e/accuracy/ltx/run_ltx_reference.py

@@ -20,6 +20,11 @@ from PIL import Image
 def _parse_args() -> argparse.Namespace:
     parser = argparse.ArgumentParser()
     parser.add_argument("--backend", choices=("official", "omni"), required=True)
+    parser.add_argument(
+        "--pipeline-kind",
+        choices=("one_stage", "distilled", "two_stage"),
+        default="one_stage",
+    )
     parser.add_argument("--request", type=Path, required=True)
     parser.add_argument("--output-dir", type=Path, required=True)
     parser.add_argument("--model")
@@ -27,6 +32,8 @@ def _parse_args() -> argparse.Namespace:
     parser.add_argument("--official-root", type=Path)
     parser.add_argument("--checkpoint", type=Path)
     parser.add_argument("--gemma-root", type=Path)
+    parser.add_argument("--spatial-upsampler", type=Path)
+    parser.add_argument("--distilled-lora", type=Path)
     parser.add_argument("--enable-layerwise-offload", action="store_true")
     return parser.parse_args()
 
@@ -91,9 +98,13 @@ def _configure_official_sdpa(pipeline: Any) -> None:
         attention=attention,
         masked_attention=attention,
     )
-    owners_and_attributes = (
+    owners_and_attributes = [
-        (pipeline.stage, "_transformer_builder"),
         (pipeline.prompt_encoder, "_embeddings_processor_builder"),
+    ]
+    owners_and_attributes.extend(
+        (stage, "_transformer_builder")
+        for name in ("stage", "stage_1", "stage_2")
+        if (stage := getattr(pipeline, name, None)) is not None
     )
     for owner, attribute in owners_and_attributes:
         builder = getattr(owner, attribute)
@@ -104,6 +115,12 @@ def _configure_official_sdpa(pipeline: Any) -> None:
         )
 
 
+def _require_path(path: Path | None, description: str) -> str:
+    if path is None or not path.is_file():
+        raise ValueError(f"Official {description} is required and must exist: {path}")
+    return str(path)
+
+
 @torch.inference_mode()
 def _run_official(args: argparse.Namespace, request: dict[str, Any]) -> None:
     if args.official_root is None or args.checkpoint is None or args.gemma_root is None:
@@ -113,16 +130,54 @@ def _run_official(args: argparse.Namespace, request: dict[str, Any]) -> None:
     torch.backends.cuda.enable_cudnn_sdp(False)
 
     from ltx_core.components.guiders import MultiModalGuiderParams
+    from ltx_core.loader import LTXV_LORA_COMFY_RENAMING_MAP, LoraPathStrengthAndSDOps
+    from ltx_pipelines.distilled import DistilledPipeline
     from ltx_pipelines.ti2vid_one_stage import TI2VidOneStagePipeline
+    from ltx_pipelines.ti2vid_two_stages import TI2VidTwoStagesPipeline
     from ltx_pipelines.utils.args import ImageConditioningInput
     from ltx_pipelines.utils.types import OffloadMode
 
-    pipeline = TI2VidOneStagePipeline(
+    if request.get("pipeline_kind", "one_stage") != args.pipeline_kind:
-        checkpoint_path=str(args.checkpoint),
+        raise ValueError(
-        gemma_root=str(args.gemma_root),
+            f"Request pipeline kind {request.get('pipeline_kind')!r} does not match {args.pipeline_kind!r}"
-        loras=(),
+        )
-        offload_mode=OffloadMode.CPU if args.enable_layerwise_offload else OffloadMode.NONE,
+    checkpoint = _require_path(args.checkpoint, "checkpoint")
-    )
+    gemma_root = str(args.gemma_root)
+    if not Path(gemma_root).is_dir():
+        raise ValueError(f"Official Gemma root is required and must be a directory: {gemma_root}")
+    offload_mode = OffloadMode.CPU if args.enable_layerwise_offload else OffloadMode.NONE
+    pipeline: Any
+    if args.pipeline_kind == "one_stage":
+        pipeline = TI2VidOneStagePipeline(
+            checkpoint_path=checkpoint,
+            gemma_root=gemma_root,
+            loras=(),
+            offload_mode=offload_mode,
+        )
+    elif args.pipeline_kind == "distilled":
+        pipeline = DistilledPipeline(
+            distilled_checkpoint_path=checkpoint,
+            spatial_upsampler_path=_require_path(args.spatial_upsampler, "spatial upsampler"),
+            gemma_root=gemma_root,
+            loras=(),
+            offload_mode=offload_mode,
+        )
+    else:
+        distilled_lora = [
+            LoraPathStrengthAndSDOps(
+                path=_require_path(args.distilled_lora, "distilled LoRA"),
+                strength=1.0,
+                sd_ops=LTXV_LORA_COMFY_RENAMING_MAP,
+            )
+        ]
+        pipeline = TI2VidTwoStagesPipeline(
+            checkpoint_path=checkpoint,
+            distilled_lora=distilled_lora,
+            spatial_upsampler_path=_require_path(args.spatial_upsampler, "spatial upsampler"),
+            gemma_root=gemma_root,
+            loras=(),
+            offload_mode=offload_mode,
+        )
     _configure_official_sdpa(pipeline)
     image_path = request.get("image")
     images = (
@@ -140,34 +195,45 @@ def _run_official(args: argparse.Namespace, request: dict[str, Any]) -> None:
             )
         ]
     )
-    video, audio = pipeline(
+    if args.pipeline_kind == "distilled":
-        prompt=request["prompt"],
+        video, audio = pipeline(
-        negative_prompt=request["negative_prompt"],
+            prompt=request["prompt"],
-        seed=request["seed"],
+            seed=request["seed"],
-        height=request["height"],
+            height=request["height"],
-        width=request["width"],
+            width=request["width"],
-        num_frames=request["num_frames"],
+            num_frames=request["num_frames"],
-        frame_rate=request["fps"],
+            frame_rate=request["fps"],
-        num_inference_steps=request["num_inference_steps"],
+            images=images,
-        video_guider_params=MultiModalGuiderParams(
+        )
-            cfg_scale=request["video_cfg_scale"],
+    else:
-            stg_scale=request["video_stg_scale"],
+        video, audio = pipeline(
-            rescale_scale=request["video_rescale_scale"],
+            prompt=request["prompt"],
-            modality_scale=request["video_modality_scale"],
+            negative_prompt=request["negative_prompt"],
-            skip_step=0,
+            seed=request["seed"],
-            stg_blocks=request["video_stg_blocks"],
+            height=request["height"],
-        ),
+            width=request["width"],
-        audio_guider_params=MultiModalGuiderParams(
+            num_frames=request["num_frames"],
-            cfg_scale=request["audio_cfg_scale"],
+            frame_rate=request["fps"],
-            stg_scale=request["audio_stg_scale"],
+            num_inference_steps=request["num_inference_steps"],
-            rescale_scale=request["audio_rescale_scale"],
+            video_guider_params=MultiModalGuiderParams(
-            modality_scale=request["audio_modality_scale"],
+                cfg_scale=request["video_cfg_scale"],
-            skip_step=0,
+                stg_scale=request["video_stg_scale"],
-            stg_blocks=request["audio_stg_blocks"],
+                rescale_scale=request["video_rescale_scale"],
-        ),
+                modality_scale=request["video_modality_scale"],
-        images=images,
+                skip_step=0,
-        max_batch_size=4,
+                stg_blocks=request["video_stg_blocks"],
-    )
+            ),
+            audio_guider_params=MultiModalGuiderParams(
+                cfg_scale=request["audio_cfg_scale"],
+                stg_scale=request["audio_stg_scale"],
+                rescale_scale=request["audio_rescale_scale"],
+                modality_scale=request["audio_modality_scale"],
+                skip_step=0,
+                stg_blocks=request["audio_stg_blocks"],
+            ),
+            images=images,
+            max_batch_size=4,
+        )
     video_tensor = torch.cat([chunk.detach().cpu() for chunk in video], dim=0)
     _save_outputs(
         args.output_dir,
@@ -179,6 +245,7 @@ def _run_official(args: argparse.Namespace, request: dict[str, Any]) -> None:
             "attention_backend": "torch_sdpa",
             "official_revision": os.environ.get("VLLM_TEST_LTX_OFFICIAL_REVISION"),
             "checkpoint": str(args.checkpoint),
+            "pipeline_kind": args.pipeline_kind,
         },
     )
 
@@ -193,15 +260,9 @@ def _unwrap_omni_output(output: Any) -> tuple[Any, Any, int]:
         multimodal_output = frames.multimodal_output or {}
         audio = multimodal_output.get("audio")
         audio_sample_rate = multimodal_output.get("audio_sample_rate")
-        if frames.is_pipeline_output and isinstance(frames.request_output, OmniRequestOutput):
+        if not frames.images:
-            frames = frames.request_output
+            raise ValueError("No video frames found in OmniRequestOutput")
-            multimodal_output = frames.multimodal_output or {}
+        frames = frames.images
-            audio = multimodal_output.get("audio", audio)
-            audio_sample_rate = multimodal_output.get("audio_sample_rate", audio_sample_rate)
-        if isinstance(frames, OmniRequestOutput):
-            if not frames.images:
-                raise ValueError("No video frames found in OmniRequestOutput")
-            frames = frames.images
 
     if isinstance(frames, list) and len(frames) == 1:
         frames = frames[0]
@@ -253,6 +314,10 @@ def _canonical_video(video: Any) -> torch.Tensor:
 def _run_omni(args: argparse.Namespace, request: dict[str, Any]) -> None:
     if args.model is None or args.model_class_name is None:
         raise ValueError("Omni backend requires --model and --model-class-name")
+    if request.get("pipeline_kind", "one_stage") != args.pipeline_kind:
+        raise ValueError(
+            f"Request pipeline kind {request.get('pipeline_kind')!r} does not match {args.pipeline_kind!r}"
+        )
 
     attention_config = {"default": {"backend": "TORCH_SDPA"}}
 
@@ -273,7 +338,8 @@ def _run_omni(args: argparse.Namespace, request: dict[str, Any]) -> None:
         parallel_config=DiffusionParallelConfig(),
     )
     try:
-        model_class_name = get_model_class_name(omni)
+        detected_model_class_name = get_model_class_name(omni)
+        model_class_name = args.model_class_name
         sampling_params = OmniDiffusionSamplingParams(
             height=request["height"],
             width=request["width"],
@@ -289,10 +355,9 @@ def _run_omni(args: argparse.Namespace, request: dict[str, Any]) -> None:
             key: value for key, value in request.items() if key.startswith("video_") or key.startswith("audio_")
         }
         apply_declared_extra_args(sampling_params, get_extra_body_params(model_class_name), guidance)
-        prompt: dict[str, Any] = {
+        prompt: dict[str, Any] = {"prompt": request["prompt"]}
-            "prompt": request["prompt"],
+        if request.get("negative_prompt"):
-            "negative_prompt": request["negative_prompt"],
+            prompt["negative_prompt"] = request["negative_prompt"]
-        }
         image_path = request.get("image")
         if image_path is not None:
             with Image.open(str(image_path)) as source_image:
@@ -312,6 +377,8 @@ def _run_omni(args: argparse.Namespace, request: dict[str, Any]) -> None:
                 "attention_backend": "torch_sdpa",
                 "model": args.model,
                 "model_class_name": model_class_name,
+                "model_config_class_name": detected_model_class_name,
+                "pipeline_kind": args.pipeline_kind,
             },
         )
     finally:

tests/e2e/accuracy/ltx/test_ltx_official_similarity.py

@@ -3,10 +3,9 @@
 
 """E2E accuracy guard against a pinned Lightricks LTX pipeline revision.
 
-The comparison runs both runtimes through PyTorch SDPA and uses
+The original reduced one-stage guards remain unchanged. Four complementary
-``max_batch_size=4`` in the official reference to match Omni's fused guidance
+default-shape cases cover both model versions, both two-stage weight families,
-batch. Video and audio guidance use the official non-HQ one-stage defaults;
+and T2V/I2V without expanding every combination into a duplicate golden.
-only the generation shape and step count are reduced for CI runtime.
 """
 
 from __future__ import annotations
@@ -17,8 +16,9 @@ import shutil
 import subprocess
 import sys
 import time
-from dataclasses import dataclass
+from dataclasses import asdict, dataclass, replace
 from pathlib import Path
+from typing import Literal
 
 import numpy as np
 import pytest
@@ -53,74 +53,260 @@ NEGATIVE_PROMPT = (
 
 # Both runtimes use PyTorch SDPA with the current Torch dispatch defaults.
 ATTENTION_BACKEND = "torch_sdpa"
-VIDEO_SSIM_MEAN_THRESHOLD = 0.95
+
-VIDEO_SSIM_MIN_THRESHOLD = 0.90
+
-VIDEO_PSNR_MEAN_THRESHOLD = 30.0
+@dataclass(frozen=True)
-AUDIO_RELATIVE_L2_THRESHOLD = 0.2
+class LTXArtifact:
-AUDIO_COSINE_THRESHOLD = 0.95
+    repo_id: str
+    filename: str
+    revision: str
+    env: str
+    repo_type: Literal["model", "dataset"] = "model"
+
+
+@dataclass(frozen=True)
+class LTXAccuracyThresholds:
+    video_ssim_mean: float
+    video_ssim_min: float
+    video_psnr_mean_db: float
+    # Zero disables gating for waveform-sensitive audio that is not expected
+    # to align with the official output.
+    audio_relative_l2: float
+    audio_cosine_similarity: float
+
+
+STRICT_THRESHOLDS = LTXAccuracyThresholds(
+    video_ssim_mean=0.95,
+    video_ssim_min=0.90,
+    video_psnr_mean_db=30.0,
+    audio_relative_l2=0.20,
+    audio_cosine_similarity=0.95,
+)
 
 
 @dataclass(frozen=True)
 class LTXAccuracyCase:
     name: str
+    pipeline_kind: Literal["one_stage", "distilled", "two_stage"]
     model_id: str
     model_revision: str
     model_env: str
     model_class_name: str
-    checkpoint_repo: str
+    checkpoint: LTXArtifact
-    checkpoint_filename: str
+    width: int
-    checkpoint_revision: str
+    height: int
-    checkpoint_env: str
+    num_frames: int
-    stg_block: int
+    num_inference_steps: int
+    seed: int
+    stg_block: int | None
+    thresholds: LTXAccuracyThresholds
     prompt: str = PROMPT
-    image_repo: str | None = None
+    gemma_model_id: str | None = None
-    image_filename: str | None = None
+    gemma_model_revision: str | None = None
-    image_revision: str | None = None
+    gemma_model_env: str | None = None
+    image: LTXArtifact | None = None
+    spatial_upsampler: LTXArtifact | None = None
+    distilled_lora: LTXArtifact | None = None
+    # Official uses block streaming; Omni uses layerwise offload. Both preserve
+    # bf16 arithmetic while keeping full-resolution two-stage cases on one H100.
+    enable_layerwise_offload: bool = False
 
 
-CASES = (
+LTX2_REVISION = "47da56e2ad66ce4125a9922b4a8826bf407f9d0a"
+LTX23_REVISION = "4229404625088d21c4f112eb640fb04a0900ee25"
+LTX2_CHECKPOINT = LTXArtifact(
+    repo_id="Lightricks/LTX-2",
+    filename="ltx-2-19b-dev.safetensors",
+    revision=LTX2_REVISION,
+    env="VLLM_TEST_LTX2_OFFICIAL_CHECKPOINT",
+)
+LTX2_DISTILLED_CHECKPOINT = replace(
+    LTX2_CHECKPOINT,
+    filename="ltx-2-19b-distilled.safetensors",
+    env="VLLM_TEST_LTX2_DISTILLED_OFFICIAL_CHECKPOINT",
+)
+LTX2_UPSAMPLER = replace(
+    LTX2_CHECKPOINT,
+    filename="ltx-2-spatial-upscaler-x2-1.0.safetensors",
+    env="VLLM_TEST_LTX2_UPSAMPLER",
+)
+LTX2_DISTILLED_LORA = replace(
+    LTX2_CHECKPOINT,
+    filename="ltx-2-19b-distilled-lora-384.safetensors",
+    env="VLLM_TEST_LTX2_DISTILLED_LORA",
+)
+LTX23_CHECKPOINT = LTXArtifact(
+    repo_id="Lightricks/LTX-2.3",
+    filename="ltx-2.3-22b-dev.safetensors",
+    revision=LTX23_REVISION,
+    env="VLLM_TEST_LTX23_OFFICIAL_CHECKPOINT",
+)
+LTX23_DISTILLED_CHECKPOINT = replace(
+    LTX23_CHECKPOINT,
+    # The pinned Diffusers checkpoint contains the original merged distilled
+    # Transformer. Keep the reference Transformer on the same version; the
+    # spatial upsampler is versioned independently and remains on 1.1 below.
+    filename="ltx-2.3-22b-distilled.safetensors",
+    env="VLLM_TEST_LTX23_DISTILLED_OFFICIAL_CHECKPOINT",
+)
+LTX23_UPSAMPLER = replace(
+    LTX23_CHECKPOINT,
+    filename="ltx-2.3-spatial-upscaler-x2-1.1.safetensors",
+    env="VLLM_TEST_LTX23_UPSAMPLER",
+)
+LTX23_DISTILLED_LORA = replace(
+    LTX23_CHECKPOINT,
+    filename="ltx-2.3-22b-distilled-lora-384-1.1.safetensors",
+    env="VLLM_TEST_LTX23_DISTILLED_LORA",
+)
+I2V_IMAGE = LTXArtifact(
+    repo_id="huggingface/documentation-images",
+    filename="diffusers/svd/rocket.png",
+    revision="645d8364f0c7a101180b364811b5a11a362e4010",
+    env="VLLM_TEST_LTX_I2V_IMAGE",
+    repo_type="dataset",
+)
+
+
+LEGACY_CASES = (
     LTXAccuracyCase(
         name="ltx2",
+        pipeline_kind="one_stage",
         model_id="Lightricks/LTX-2",
-        model_revision="47da56e2ad66ce4125a9922b4a8826bf407f9d0a",
+        model_revision=LTX2_REVISION,
         model_env="VLLM_TEST_LTX2_MODEL",
         model_class_name="LTX2Pipeline",
-        checkpoint_repo="Lightricks/LTX-2",
+        checkpoint=LTX2_CHECKPOINT,
-        checkpoint_filename="ltx-2-19b-dev.safetensors",
+        width=512,
-        checkpoint_revision="47da56e2ad66ce4125a9922b4a8826bf407f9d0a",
+        height=384,
-        checkpoint_env="VLLM_TEST_LTX2_OFFICIAL_CHECKPOINT",
+        num_frames=25,
+        num_inference_steps=20,
+        seed=42,
         stg_block=29,
+        thresholds=STRICT_THRESHOLDS,
     ),
     LTXAccuracyCase(
         name="ltx2_3",
+        pipeline_kind="one_stage",
         model_id="diffusers/LTX-2.3-Diffusers",
         model_revision="8eee8edcf067e838b843f926ec4d4cc9b2be1aaf",
         model_env="VLLM_TEST_LTX23_MODEL",
         model_class_name="LTX2Pipeline",
-        checkpoint_repo="Lightricks/LTX-2.3",
+        checkpoint=LTX23_CHECKPOINT,
-        checkpoint_filename="ltx-2.3-22b-dev.safetensors",
+        width=512,
-        checkpoint_revision="4229404625088d21c4f112eb640fb04a0900ee25",
+        height=384,
-        checkpoint_env="VLLM_TEST_LTX23_OFFICIAL_CHECKPOINT",
+        num_frames=25,
+        num_inference_steps=20,
+        seed=42,
         stg_block=28,
+        thresholds=STRICT_THRESHOLDS,
     ),
     LTXAccuracyCase(
         name="ltx2_3_i2v",
+        pipeline_kind="one_stage",
         model_id="diffusers/LTX-2.3-Diffusers",
         model_revision="8eee8edcf067e838b843f926ec4d4cc9b2be1aaf",
         model_env="VLLM_TEST_LTX23_MODEL",
         model_class_name="LTX2Pipeline",
-        checkpoint_repo="Lightricks/LTX-2.3",
+        checkpoint=LTX23_CHECKPOINT,
-        checkpoint_filename="ltx-2.3-22b-dev.safetensors",
+        width=512,
-        checkpoint_revision="4229404625088d21c4f112eb640fb04a0900ee25",
+        height=384,
-        checkpoint_env="VLLM_TEST_LTX23_OFFICIAL_CHECKPOINT",
+        num_frames=25,
+        num_inference_steps=20,
+        seed=42,
         stg_block=28,
-        image_repo="huggingface/documentation-images",
+        thresholds=STRICT_THRESHOLDS,
-        image_filename="diffusers/svd/rocket.png",
+        image=I2V_IMAGE,
-        image_revision="645d8364f0c7a101180b364811b5a11a362e4010",
     ),
 )
 
 
+TWO_STAGE_CASES = (
+    LTXAccuracyCase(
+        name="ltx2_distilled_t2v",
+        pipeline_kind="distilled",
+        model_id="rootonchair/LTX-2-19b-distilled",
+        model_revision="388e2846f54aae51687498ffb6b27c7c2c9ce9e5",
+        model_env="VLLM_TEST_LTX2_DISTILLED_MODEL",
+        model_class_name="LTX2DistilledPipeline",
+        checkpoint=LTX2_DISTILLED_CHECKPOINT,
+        spatial_upsampler=LTX2_UPSAMPLER,
+        width=1536,
+        height=1024,
+        num_frames=121,
+        num_inference_steps=8,
+        seed=10,
+        stg_block=None,
+        gemma_model_id="Lightricks/LTX-2",
+        gemma_model_revision=LTX2_REVISION,
+        gemma_model_env="VLLM_TEST_LTX2_MODEL",
+        thresholds=STRICT_THRESHOLDS,
+    ),
+    LTXAccuracyCase(
+        name="ltx23_distilled_i2v",
+        pipeline_kind="distilled",
+        model_id="diffusers/LTX-2.3-Distilled-Diffusers",
+        model_revision="432e0d3c2d1769aaa4d295f9243f7062bf6b47ee",
+        model_env="VLLM_TEST_LTX23_DISTILLED_MODEL",
+        model_class_name="LTX2DistilledPipeline",
+        checkpoint=LTX23_DISTILLED_CHECKPOINT,
+        spatial_upsampler=LTX23_UPSAMPLER,
+        width=1536,
+        height=1024,
+        num_frames=121,
+        num_inference_steps=8,
+        seed=10,
+        stg_block=None,
+        gemma_model_id="diffusers/LTX-2.3-Diffusers",
+        gemma_model_revision="8eee8edcf067e838b843f926ec4d4cc9b2be1aaf",
+        gemma_model_env="VLLM_TEST_LTX23_MODEL",
+        thresholds=STRICT_THRESHOLDS,
+        image=I2V_IMAGE,
+    ),
+    LTXAccuracyCase(
+        name="ltx2_two_stage_layer_fused_t2v",
+        pipeline_kind="two_stage",
+        model_id="Lightricks/LTX-2",
+        model_revision=LTX2_REVISION,
+        model_env="VLLM_TEST_LTX2_MODEL",
+        model_class_name="LTX2TwoStagePipeline",
+        checkpoint=LTX2_CHECKPOINT,
+        spatial_upsampler=LTX2_UPSAMPLER,
+        distilled_lora=LTX2_DISTILLED_LORA,
+        width=1536,
+        height=1024,
+        num_frames=121,
+        num_inference_steps=40,
+        seed=10,
+        stg_block=29,
+        enable_layerwise_offload=True,
+        thresholds=STRICT_THRESHOLDS,
+    ),
+    LTXAccuracyCase(
+        name="ltx23_two_stage_layer_fused_i2v",
+        pipeline_kind="two_stage",
+        model_id="diffusers/LTX-2.3-Diffusers",
+        model_revision="8eee8edcf067e838b843f926ec4d4cc9b2be1aaf",
+        model_env="VLLM_TEST_LTX23_MODEL",
+        model_class_name="LTX2TwoStagePipeline",
+        checkpoint=LTX23_CHECKPOINT,
+        spatial_upsampler=LTX23_UPSAMPLER,
+        distilled_lora=LTX23_DISTILLED_LORA,
+        width=1536,
+        height=1024,
+        num_frames=121,
+        num_inference_steps=30,
+        seed=10,
+        stg_block=28,
+        enable_layerwise_offload=True,
+        thresholds=STRICT_THRESHOLDS,
+        image=I2V_IMAGE,
+    ),
+)
+
+CASES = (*LEGACY_CASES, *TWO_STAGE_CASES)
+
+
 def _run(command: list[str], *, env: dict[str, str], timeout: int = 1800) -> None:
     start = time.perf_counter()
     subprocess.run(command, env=env, timeout=timeout, check=True)
@@ -209,6 +395,7 @@ def _resolve_model(case: LTXAccuracyCase) -> Path:
                 "model_index.json",
                 "audio_vae/*",
                 "connectors/*",
+                "latent_upsampler/*",
                 "processor/*",
                 "scheduler/*",
                 "text_encoder/config.json",
@@ -223,74 +410,82 @@ def _resolve_model(case: LTXAccuracyCase) -> Path:
     )
 
 
-def _resolve_gemma_root(model: Path) -> Path:
+def _resolve_gemma_root(case: LTXAccuracyCase, model: Path) -> Path:
     configured_root = os.environ.get("VLLM_TEST_LTX_GEMMA_ROOT")
     if configured_root:
         root = Path(configured_root)
         assert root.is_dir(), f"Gemma root not found: {root}"
         return root
+    if case.gemma_model_id is not None:
+        configured_model = os.environ.get(case.gemma_model_env or "")
+        if configured_model and Path(configured_model).is_dir():
+            return Path(configured_model)
+        return Path(
+            snapshot_download(
+                repo_id=case.gemma_model_id,
+                revision=case.gemma_model_revision,
+                allow_patterns=[
+                    "processor/*",
+                    "text_encoder/*",
+                    "tokenizer/*",
+                ],
+            )
+        )
     return model
 
 
-def _resolve_checkpoint(case: LTXAccuracyCase, model: Path) -> Path:
+def _resolve_artifact(artifact: LTXArtifact, model: Path | None = None) -> Path:
-    configured_checkpoint = os.environ.get(case.checkpoint_env)
+    configured_path = os.environ.get(artifact.env)
-    if configured_checkpoint:
+    if configured_path:
-        checkpoint = Path(configured_checkpoint)
+        path = Path(configured_path)
-        assert checkpoint.is_file(), f"Official checkpoint not found: {checkpoint}"
+        assert path.is_file(), f"Configured LTX artifact not found: {path}"
-        return checkpoint
+        return path
-    model_checkpoint = model / case.checkpoint_filename
+    if model is not None:
-    if model_checkpoint.is_file():
+        model_path = model / artifact.filename
-        return model_checkpoint
+        if model_path.is_file():
+            return model_path
     return Path(
         hf_hub_download(
-            repo_id=case.checkpoint_repo,
+            repo_id=artifact.repo_id,
-            filename=case.checkpoint_filename,
+            repo_type=None if artifact.repo_type == "model" else artifact.repo_type,
-            revision=case.checkpoint_revision,
+            filename=artifact.filename,
+            revision=artifact.revision,
         )
     )
 
 
 def _resolve_image(case: LTXAccuracyCase) -> Path | None:
-    if case.image_filename is None:
+    if case.image is None:
         return None
-    if case.image_repo is None or case.image_revision is None:
+    return _resolve_artifact(case.image)
-        raise ValueError(f"Incomplete image source for LTX accuracy case {case.name!r}.")
-    configured_image = os.environ.get("VLLM_TEST_LTX_I2V_IMAGE")
-    if configured_image:
-        image = Path(configured_image)
-        assert image.is_file(), f"LTX I2V conditioning image not found: {image}"
-        return image
-    return Path(
-        hf_hub_download(
-            repo_id=case.image_repo,
-            repo_type="dataset",
-            filename=case.image_filename,
-            revision=case.image_revision,
-        )
-    )
 
 
 def _request(case: LTXAccuracyCase, image: Path | None) -> dict[str, object]:
     request: dict[str, object] = {
+        "pipeline_kind": case.pipeline_kind,
         "prompt": case.prompt,
-        "negative_prompt": NEGATIVE_PROMPT,
+        "negative_prompt": "" if case.pipeline_kind == "distilled" else NEGATIVE_PROMPT,
-        "width": 512,
+        "width": case.width,
-        "height": 384,
+        "height": case.height,
-        "num_frames": 25,
+        "num_frames": case.num_frames,
         "fps": 24,
-        "num_inference_steps": 20,
+        "num_inference_steps": case.num_inference_steps,
-        "seed": 42,
+        "seed": case.seed,
-        "video_cfg_scale": 3.0,
-        "audio_cfg_scale": 7.0,
-        "video_stg_scale": 1.0,
-        "audio_stg_scale": 1.0,
-        "video_modality_scale": 3.0,
-        "audio_modality_scale": 3.0,
-        "video_rescale_scale": 0.7,
-        "audio_rescale_scale": 0.7,
-        "video_stg_blocks": [case.stg_block],
-        "audio_stg_blocks": [case.stg_block],
     }
+    if case.pipeline_kind != "distilled":
+        assert case.stg_block is not None
+        request.update(
+            video_cfg_scale=3.0,
+            audio_cfg_scale=7.0,
+            video_stg_scale=1.0,
+            audio_stg_scale=1.0,
+            video_modality_scale=3.0,
+            audio_modality_scale=3.0,
+            video_rescale_scale=0.7,
+            audio_rescale_scale=0.7,
+            video_stg_blocks=[case.stg_block],
+            audio_stg_blocks=[case.stg_block],
+        )
     if image is not None:
         request["image"] = str(image.resolve())
     return request
@@ -299,20 +494,30 @@ def _request(case: LTXAccuracyCase, image: Path | None) -> dict[str, object]:
 def _video_metrics(reference: np.ndarray, prediction: np.ndarray) -> dict[str, float]:
     assert reference.shape == prediction.shape
     assert reference.ndim == 4 and reference.shape[-1] == 3
+    device = torch.device("cpu")
+    ssim = StructuralSimilarityIndexMeasure(data_range=1.0).to(device)
+    psnr = PeakSignalNoiseRatio(data_range=1.0).to(device)
     ssim_scores: list[float] = []
     psnr_scores: list[float] = []
-    for reference_frame, prediction_frame in zip(reference, prediction, strict=True):
+    max_abs = 0.0
-        reference_tensor = torch.from_numpy(reference_frame).permute(2, 0, 1).unsqueeze(0)
+    absolute_error_sum = 0.0
-        prediction_tensor = torch.from_numpy(prediction_frame).permute(2, 0, 1).unsqueeze(0)
+    with torch.inference_mode():
-        ssim_scores.append(float(StructuralSimilarityIndexMeasure(data_range=1.0)(prediction_tensor, reference_tensor)))
+        for reference_frame, prediction_frame in zip(reference, prediction, strict=True):
-        psnr_scores.append(float(PeakSignalNoiseRatio(data_range=1.0)(prediction_tensor, reference_tensor)))
+            reference_tensor = torch.from_numpy(reference_frame).permute(2, 0, 1).unsqueeze(0).to(device)
-    difference = np.abs(reference.astype(np.float64) - prediction.astype(np.float64))
+            prediction_tensor = torch.from_numpy(prediction_frame).permute(2, 0, 1).unsqueeze(0).to(device)
+            ssim_scores.append(float(ssim(prediction_tensor, reference_tensor)))
+            psnr_scores.append(float(psnr(prediction_tensor, reference_tensor)))
+            ssim.reset()
+            psnr.reset()
+            difference = np.abs(reference_frame.astype(np.float64) - prediction_frame.astype(np.float64))
+            max_abs = max(max_abs, float(difference.max()))
+            absolute_error_sum += float(difference.sum())
     return {
         "ssim_mean": float(np.mean(ssim_scores)),
         "ssim_min": float(np.min(ssim_scores)),
         "psnr_mean_db": float(np.mean(psnr_scores)),
-        "max_abs": float(difference.max()),
+        "max_abs": max_abs,
-        "mean_abs": float(difference.mean()),
+        "mean_abs": absolute_error_sum / reference.size,
     }
 
 
@@ -343,13 +548,19 @@ def _audio_metrics(reference: np.ndarray, prediction: np.ndarray) -> dict[str, f
 @pytest.mark.benchmark
 @pytest.mark.diffusion
 @hardware_test(res={"cuda": "H100"}, num_cards=1)
-def test_ltx_one_stage_matches_official(case: LTXAccuracyCase, accuracy_artifact_root: Path) -> None:
+def test_ltx_matches_official(case: LTXAccuracyCase, accuracy_artifact_root: Path) -> None:
-    """Compare official and Omni raw AV outputs from the same E2E request."""
+    """Compare official and Omni raw AV outputs from the same single-H100 request."""
+    configured_artifact_root = os.environ.get("VLLM_TEST_LTX_ARTIFACT_ROOT")
+    if configured_artifact_root:
+        accuracy_artifact_root = Path(configured_artifact_root)
+        accuracy_artifact_root.mkdir(parents=True, exist_ok=True)
     output_root = reset_artifact_dir(accuracy_artifact_root / "ltx_official" / case.name)
     official_root, official_revision = _official_source(accuracy_artifact_root / "ltx_official")
     model = _resolve_model(case)
-    gemma_root = _resolve_gemma_root(model)
+    gemma_root = _resolve_gemma_root(case, model)
-    checkpoint = _resolve_checkpoint(case, model)
+    checkpoint = _resolve_artifact(case.checkpoint, model)
+    spatial_upsampler = _resolve_artifact(case.spatial_upsampler, model) if case.spatial_upsampler is not None else None
+    distilled_lora = _resolve_artifact(case.distilled_lora, model) if case.distilled_lora is not None else None
     image = _resolve_image(case)
     request_path = output_root / "request.json"
     request_path.write_text(json.dumps(_request(case, image), indent=2) + "\n")
@@ -359,8 +570,13 @@ def test_ltx_one_stage_matches_official(case: LTXAccuracyCase, accuracy_artifact
         str(runner),
         "--request",
         str(request_path),
+        "--pipeline-kind",
+        case.pipeline_kind,
     ]
-    if os.environ.get("VLLM_TEST_LTX_ENABLE_LAYERWISE_OFFLOAD", "").lower() in {"1", "true", "yes", "on"}:
+    enable_layerwise_offload = case.enable_layerwise_offload or os.environ.get(
+        "VLLM_TEST_LTX_ENABLE_LAYERWISE_OFFLOAD", ""
+    ).lower() in {"1", "true", "yes", "on"}
+    if enable_layerwise_offload:
         runner_args.append("--enable-layerwise-offload")
     env = os.environ.copy()
     env["VLLM_TEST_LTX_OFFICIAL_REVISION"] = official_revision
@@ -372,10 +588,9 @@ def test_ltx_one_stage_matches_official(case: LTXAccuracyCase, accuracy_artifact
     )
 
     official_output = output_root / "official"
-    _run(
+    official_args = _official_runner_prefix() + runner_args
-        _official_runner_prefix()
+    official_args.extend(
-        + runner_args
+        [
-        + [
             "--backend",
             "official",
             "--output-dir",
@@ -386,12 +601,16 @@ def test_ltx_one_stage_matches_official(case: LTXAccuracyCase, accuracy_artifact
             str(checkpoint),
             "--gemma-root",
             str(gemma_root),
-        ],
+        ]
-        env=env,
     )
+    if spatial_upsampler is not None:
+        official_args.extend(["--spatial-upsampler", str(spatial_upsampler)])
+    if distilled_lora is not None:
+        official_args.extend(["--distilled-lora", str(distilled_lora)])
+    _run(official_args, env=env)
 
     omni_output = output_root / "omni"
-    _run(
+    omni_args = (
         [sys.executable]
         + runner_args
         + [
@@ -403,9 +622,9 @@ def test_ltx_one_stage_matches_official(case: LTXAccuracyCase, accuracy_artifact
             str(model),
             "--model-class-name",
             case.model_class_name,
-        ],
+        ]
-        env=env,
     )
+    _run(omni_args, env=env)
 
     official_metadata = json.loads((official_output / "metadata.json").read_text())
     omni_metadata = json.loads((omni_output / "metadata.json").read_text())
@@ -423,18 +642,27 @@ def test_ltx_one_stage_matches_official(case: LTXAccuracyCase, accuracy_artifact
     result = {
         "case": case.name,
         "task": "i2v" if image is not None else "t2v",
+        "pipeline_kind": case.pipeline_kind,
+        "model_class_name": case.model_class_name,
         "attention_backend": ATTENTION_BACKEND,
         "official_revision": official_revision,
         "model_revision": case.model_revision,
-        "checkpoint_revision": case.checkpoint_revision,
+        "checkpoint_revision": case.checkpoint.revision,
+        "spatial_upsampler_revision": (case.spatial_upsampler.revision if case.spatial_upsampler is not None else None),
+        "distilled_lora_revision": case.distilled_lora.revision if case.distilled_lora is not None else None,
+        "enable_layerwise_offload": enable_layerwise_offload,
+        "thresholds": asdict(case.thresholds),
+        "request": _request(case, image),
         "video": video_metrics,
         "audio": audio_metrics,
     }
     (output_root / "metrics.json").write_text(json.dumps(result, indent=2) + "\n")
     print(json.dumps(result, indent=2))
 
-    assert video_metrics["ssim_mean"] >= VIDEO_SSIM_MEAN_THRESHOLD
+    assert video_metrics["ssim_mean"] >= case.thresholds.video_ssim_mean
-    assert video_metrics["ssim_min"] >= VIDEO_SSIM_MIN_THRESHOLD
+    assert video_metrics["ssim_min"] >= case.thresholds.video_ssim_min
-    assert video_metrics["psnr_mean_db"] >= VIDEO_PSNR_MEAN_THRESHOLD
+    assert video_metrics["psnr_mean_db"] >= case.thresholds.video_psnr_mean_db
-    assert audio_metrics["relative_l2"] <= AUDIO_RELATIVE_L2_THRESHOLD
+    if case.thresholds.audio_relative_l2 > 0:
-    assert audio_metrics["cosine_similarity"] >= AUDIO_COSINE_THRESHOLD
+        assert audio_metrics["relative_l2"] <= case.thresholds.audio_relative_l2
+    if case.thresholds.audio_cosine_similarity > 0:
+        assert audio_metrics["cosine_similarity"] >= case.thresholds.audio_cosine_similarity

tests/e2e/offline_inference/test_audiox_model_expansion.py

@@ -17,7 +17,7 @@ from vllm_omni.platforms import current_omni_platform
 # Override in CI if needed: AUDIOX_TEST_MODEL=<model-or-local-path>
 models = [os.environ.get("AUDIOX_TEST_MODEL", "zhangj1an/audiox_random")]
 
-# (model, stage_configs_path, extra_omni_kwargs) for ``omni_runner`` indirect parametrize
+# (model, deploy_config_path, extra_omni_kwargs) for ``omni_runner`` indirect parametrize
 _OMNI_RUNNER_PARAMS = [(m, None, {"model_class_name": "AudioXPipeline"}) for m in models]
 
 pytestmark = [

tests/e2e/offline_inference/test_bagel_expansion.py

@@ -43,7 +43,7 @@ MODEL = "ByteDance-Seed/BAGEL-7B-MoT"
 BAGEL_STAGE_CONFIG = get_deploy_config_path("ci/bagel.yaml")
 DEFAULT_PROMPT = "<|im_start|>A cute cat<|im_end|>"
 
-# (model, stage_configs_path) for ``@pytest.mark.parametrize("omni_runner", ..., indirect=True)``
+# (model, deploy_config_path) for ``@pytest.mark.parametrize("omni_runner", ..., indirect=True)``
 _OMNI_RUNNER_PARAM = (MODEL, BAGEL_STAGE_CONFIG)