[Diffusion] Register runtime cache for direct DLO H2D

Avoid recurrent mmap-to-pinned packing in no-AllGather DLO by registering transform-complete runtime-cache mappings under an explicit per-worker budget. Copy mapped tensors directly into the rotating device buffers while preserving bounded-staging fallback and ordered shutdown cleanup. Signed-off-by: Hongsheng Liu <liuhongsheng4@huawei.com>
[Diffusion] Document runtime cache hardware tradeoffs
2026-08-17 11:50:32 +08:00 · 2026-08-17 00:43:06 +08:00 · 2026-08-16 23:26:25 +08:00 · 2026-08-16 23:15:48 +08:00 · 2026-08-16 11:22:28 +08:00 · 2026-08-16 11:02:38 +08:00
357 changed files with 38473 additions and 4216 deletions
@@ -1,153 +1,103 @@
 #!/bin/bash
-# vllm-omni customized version
-# Based on: vllm/.buildkite/scripts/hardware_ci/run-amd-test.sh
-# Last synced: 2025-12-15
-# Modifications: docker image name for vllm-omni

-# This script runs test inside the corresponding ROCm docker container.
-set -o pipefail
+# Run vLLM-Omni ROCm tests directly in the MI300 Kubernetes pod. The pod's
+# container image is selected by test-template-amd-omni.j2; MI300 has no DinD.
+set -euo pipefail

-# Export Python path
-export PYTHONPATH=".."
+: "${PYTHONFAULTHANDLER:=1}"
+: "${HF_HOME:=/home/buildkite-agent/huggingface}"
+: "${HF_HUB_DOWNLOAD_TIMEOUT:=300}"
+: "${HF_HUB_ETAG_TIMEOUT:=60}"
+: "${MIOPEN_DEBUG_CONV_DIRECT:=0}"
+: "${MIOPEN_DEBUG_CONV_GEMM:=0}"
+: "${VLLM_ROCM_USE_AITER:=0}"
+export PYTHONFAULTHANDLER HF_HOME HF_HUB_DOWNLOAD_TIMEOUT HF_HUB_ETAG_TIMEOUT
+export MIOPEN_DEBUG_CONV_DIRECT MIOPEN_DEBUG_CONV_GEMM VLLM_ROCM_USE_AITER
+export PYTORCH_ROCM_ARCH=""
+export PYTHONPATH="${PYTHONPATH:-..}"
+
+if [[ "${VLLM_CI_DOCKER_DISABLED:-0}" != "1" ]]; then
+    echo "Error: MI300 CI must run natively with Docker disabled." >&2
+    exit 1
+fi
+
+if [[ -n "${TEST_COMMANDS:-}" ]]; then
+    commands="${TEST_COMMANDS}"
+else
+    commands="$*"
+fi
+if [[ -z "${commands}" ]]; then
+    echo "Error: No test commands provided." >&2
+    exit 1
+fi
+
+echo "--- Native in-pod ROCm CI"
+
+job_id="${BUILDKITE_JOB_ID:-${BUILDKITE_PARALLEL_JOB:-local}}"
+job_id="${job_id//[^A-Za-z0-9_.-]/_}"
+native_cache_root="/tmp/vllm-omni-native-${job_id}"
+TMPDIR="${native_cache_root}/tmp"
+TORCHINDUCTOR_CACHE_DIR="${native_cache_root}/torchinductor"
+TRITON_CACHE_DIR="${native_cache_root}/triton"
+VLLM_CACHE_ROOT="${native_cache_root}/vllm"
+XDG_CACHE_HOME="${native_cache_root}/xdg"
+HF_DATASETS_CACHE="${native_cache_root}/huggingface/datasets"
+export TMPDIR TORCHINDUCTOR_CACHE_DIR TRITON_CACHE_DIR VLLM_CACHE_ROOT
+export XDG_CACHE_HOME HF_DATASETS_CACHE
+mkdir -p "${TMPDIR}" "${TORCHINDUCTOR_CACHE_DIR}" "${TRITON_CACHE_DIR}" \
+    "${VLLM_CACHE_ROOT}" "${XDG_CACHE_HOME}" "${HF_HOME}" \
+    "${HF_DATASETS_CACHE}"
+
+if [[ "${VLLM_CI_REQUIRE_PERSISTENT_HF_CACHE:-0}" == "1" ]]; then
+    if ! command -v findmnt >/dev/null 2>&1; then
+        echo "Error: findmnt is required to verify the Hugging Face cache mount." >&2
+        exit 1
+    fi
+    hf_mount=$(findmnt -n -T "${HF_HOME}" -o TARGET 2>/dev/null || true)
+    if [[ -z "${hf_mount}" || "${hf_mount}" == "/" ]]; then
+        echo "Error: MI300 CI requires a persistent volume at or above ${HF_HOME}." >&2
+        exit 1
+    fi
+fi

-# Print ROCm version
-echo "--- ROCm info"
 rocminfo

-# cleanup older docker images
-cleanup_docker() {
-  # Get Docker's root directory
-  docker_root=$(docker info -f '{{.DockerRootDir}}')
-  if [ -z "$docker_root" ]; then
-    echo "Failed to determine Docker root directory."
-    exit 1
-  fi
-  echo "Docker root directory: $docker_root"
-  # Check disk usage of the filesystem where Docker's root directory is located
-  disk_usage=$(df "$docker_root" | tail -1 | awk '{print $5}' | sed 's/%//')
-  # Define the threshold
-  threshold=70
-  if [ "$disk_usage" -gt "$threshold" ]; then
-    echo "Disk usage is above $threshold%. Cleaning up Docker images and volumes..."
-    # Remove dangling images (those that are not tagged and not used by any container)
-    docker image prune -f
-    # Remove unused volumes / force the system prune for old images as well.
-    docker volume prune -f && docker system prune --force --filter "until=72h" --all
-    echo "Docker images and volumes cleanup completed."
-  else
-    echo "Disk usage is below $threshold%. No cleanup needed."
-  fi
-}
+expected_gpus="${VLLM_CI_EXPECTED_GPU_COUNT:-1}"
+python3 - <<'PY'
+import os

-# Call the cleanup docker function
-cleanup_docker
+import torch

-echo "--- Pulling container"
-## Temporary change to use AMD Docker Hub to store the vllm-omni image
-# to bypass the rate limit issue with ECR Public Gallery.
-# Images are now stored in a separate repository for vllm-omni, instead of vllm-ci.
-# TODO: @tjtanaa point back to ECR Public Gallery
-# once the amd agents are configured to use ECR Public Gallery.
-# image_name="public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:${BUILDKITE_COMMIT}-rocm-omni"
-image_name="rocm/vllm-omni:${BUILDKITE_COMMIT}"
-container_name="rocm_${BUILDKITE_COMMIT}_$(tr -dc A-Za-z0-9 < /dev/urandom | head -c 10; echo)"
+expected = int(os.environ.get("VLLM_CI_EXPECTED_GPU_COUNT", "1"))
+assert torch.version.hip, "PyTorch is not a ROCm build"
+assert torch.cuda.is_available(), "ROCm GPU is not available to PyTorch"
+actual = torch.cuda.device_count()
+assert actual == expected, f"Expected {expected} ROCm GPU(s), found {actual}"
+PY

-# TODO: @tjtanaa uncomment this once the amd agents are configured to use ECR Public Gallery.
-# # Install AWS CLI to authenticate to ECR Public Gallery to get higher rate limit for pulling images
-# sudo apt-get update && sudo apt-get install -y awscli
-##  Use safe docker login helper to prevent race conditions
-# source "$(dirname "${BASH_SOURCE[0]}")/../../common/scripts/docker_login_ecr_public.sh"
-# safe_docker_login_ecr_public
-
-## Pull the container from AMD Docker Hub
-
-docker pull "${image_name}"
-
-remove_docker_container() {
-   docker ps -aq --filter "name=${container_name}" | xargs -r docker rm -f || true
-   docker image rm -f "${image_name}" || true
-}
-trap remove_docker_container EXIT
-
-echo "--- Running container"
-
-HF_CACHE="$(realpath ~)/huggingface"
-mkdir -p "${HF_CACHE}"
-HF_MOUNT="/root/.cache/huggingface"
-
-if [[ -n "${TEST_COMMAND:-}" ]]; then
-    commands="$TEST_COMMAND"
-else
-    commands="$@"
-fi
-echo "Commands:$commands"
-
-PARALLEL_JOB_COUNT=8
-MYPYTHONPATH=".."
-
-# Test that we're launching on the machine that has
-# proper access to GPUs
-render_gid=$(getent group render | cut -d: -f3)
-if [[ -z "$render_gid" ]]; then
-  echo "Error: 'render' group not found. This is required for GPU access." >&2
-  exit 1
-fi
-
-# check if the command contains shard flag, we will run all shards in parallel because the host have 8 GPUs.
-# TODO: @tjtanaa reenable to run VLLM_ROCM_USE_AITER=1 when AITER is shipped with prebuilt kernels.
-if [[ $commands == *"--shard-id="* ]]; then
-  # assign job count as the number of shards used
-  commands=$(echo "$commands" | sed -E "s/--num-shards[[:blank:]]*=[[:blank:]]*[0-9]*/--num-shards=${PARALLEL_JOB_COUNT} /g" | sed 's/ \\ / /g')
-  for GPU in $(seq 0 $(($PARALLEL_JOB_COUNT-1))); do
-    # assign shard-id for each shard
-    commands_gpu=$(echo "$commands" | sed -E "s/--shard-id[[:blank:]]*=[[:blank:]]*[0-9]*/--shard-id=${GPU} /g" | sed 's/ \\ / /g')
-    echo "Shard ${GPU} commands:$commands_gpu"
-    echo "Render devices: $BUILDKITE_AGENT_META_DATA_RENDER_DEVICES"
-    docker run \
-        --device /dev/kfd $BUILDKITE_AGENT_META_DATA_RENDER_DEVICES \
-        --network=host \
-        --shm-size=16gb \
-        --group-add "$render_gid" \
-        -e MIOPEN_DEBUG_CONV_DIRECT=0 \
-        -e MIOPEN_DEBUG_CONV_GEMM=0 \
-        -e VLLM_ROCM_USE_AITER=0 \
-        -e HIP_VISIBLE_DEVICES="${GPU}" \
-        -e HF_TOKEN \
-        -v "${HF_CACHE}:${HF_MOUNT}" \
-        -e "HF_HOME=${HF_MOUNT}" \
-        -e "PYTHONPATH=${MYPYTHONPATH}" \
-        --name "${container_name}_${GPU}" \
-        "${image_name}" \
-        /bin/bash -c "${commands_gpu}" \
-        |& while read -r line; do echo ">>Shard $GPU: $line"; done &
-    PIDS+=($!)
-  done
-  #wait for all processes to finish and collect exit codes
-  for pid in "${PIDS[@]}"; do
-    wait "${pid}"
-    STATUS+=($?)
-  done
-  for st in "${STATUS[@]}"; do
-    if [[ ${st} -ne 0 ]]; then
-      echo "One of the processes failed with $st"
-      exit "${st}"
+echo "Commands:${commands}"
+if /bin/bash -o pipefail -c '
+set -E
+test_status=0
+trap '\''
+    command_status=$?
+    # Keep running subsequent commands, but retain a failure for the final exit.
+    # Prefer a test failure over pytest "no tests collected" when both occur.
+    if (( test_status == 0 || (test_status == 5 && command_status != 5) )); then
+        test_status=${command_status}
    fi
-  done
+'\'' ERR
+eval "$1"
+exit "${test_status}"
+' _ "${commands}"; then
+    exit 0
 else
-  echo "Render devices: $BUILDKITE_AGENT_META_DATA_RENDER_DEVICES"
-  docker run \
-          --device /dev/kfd $BUILDKITE_AGENT_META_DATA_RENDER_DEVICES \
-          --network=host \
-          --shm-size=16gb \
-          --group-add "$render_gid" \
-          -e MIOPEN_DEBUG_CONV_DIRECT=0 \
-          -e MIOPEN_DEBUG_CONV_GEMM=0 \
-          -e VLLM_ROCM_USE_AITER=0 \
-          -e HF_TOKEN \
-          -v "${HF_CACHE}:${HF_MOUNT}" \
-          -e "HF_HOME=${HF_MOUNT}" \
-          -e "PYTHONPATH=${MYPYTHONPATH}" \
-          --name "${container_name}" \
-          "${image_name}" \
-          /bin/bash -c "${commands}"
+    exit_code=$?
 fi
+
+if [[ ${exit_code} -eq 5 && "${VLLM_CI_ALLOW_NO_TESTS:-0}" == "1" ]]; then
+    echo "Pytest collected no tests; VLLM_CI_ALLOW_NO_TESTS=1, treating exit code 5 as success."
+    exit 0
+fi
+
+exit "${exit_code}"
@@ -7,54 +7,64 @@ steps:

 - group: ":card_index_dividers: Simple Test"
  steps:
-    - label: "Simple · Diffusion & Model Executor Test (1/2)"
-      agent_pool: mi325_1
+    - label: "Simple · Model Executor Test · Shard %N/%t"
+      agent_pool: mi300_1
+      parallelism: 2
      depends_on: amd-build
-      timeout_in_minutes: 70
      mirror_hardwares: [amdproduction]
      grade: Blocking
      commands:
        - export VLLM_ROCM_USE_AITER=0
        # ignore test_teacache_extractors.py because it use rocm gemm kernel from vLLM
        # that is not supported on CPU
-        # CPU tests on ROCm gpu machines are very slow, so we give them more time.
-        - "timeout 60m pytest -sv tests/diffusion -m 'core_model and cpu' --ignore=tests/diffusion/cache/test_teacache_extractors.py --cov=vllm_omni --cov-report=term-missing:skip-covered --cov-report=xml"
+        - "pytest -sv tests/model_executor -m 'core_model and cpu' --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --shard-id=$$BUILDKITE_PARALLEL_JOB --cov=vllm_omni --cov-report=term-missing:skip-covered --cov-report=xml:coverage-model-executor-$$BUILDKITE_PARALLEL_JOB.xml"

-    - label: "Simple · Diffusion & Model Executor Test (2/2)"
-      agent_pool: mi325_1
+    - label: "Simple · Diffusion Test · Shard %N/%t"
+      agent_pool: mi300_1
+      parallelism: 2
      depends_on: amd-build
-      timeout_in_minutes: 70
      mirror_hardwares: [amdproduction]
      grade: Blocking
      commands:
        - export VLLM_ROCM_USE_AITER=0
-        # CPU tests on ROCm gpu machines are very slow, so we give them more time.
-        - "timeout 60m pytest -sv tests/model_executor -m 'core_model and cpu' --cov=vllm_omni --cov-report=term-missing:skip-covered --cov-report=xml"
+        # ignore test_teacache_extractors.py because it use rocm gemm kernel from vLLM
+        # that is not supported on CPU
+        - "pytest -sv tests/diffusion -m 'core_model and cpu' --ignore=tests/diffusion/cache/test_teacache_extractors.py --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --shard-id=$$BUILDKITE_PARALLEL_JOB --cov=vllm_omni --cov-report=term-missing:skip-covered --cov-report=xml:coverage-diffusion-$$BUILDKITE_PARALLEL_JOB.xml"
+
+    - label: "Simple · Engine&Entrypoints Test"
+      agent_pool: mi300_1
+      depends_on: amd-build
+      mirror_hardwares: [amdproduction]
+      grade: Blocking
+      commands:
+        - export VLLM_ROCM_USE_AITER=0
+        - "pytest -sv tests/entrypoints tests/engine -m 'core_model and cpu' --cov=vllm_omni --cov-report=term-missing:skip-covered --cov-report=xml"

    - label: "Simple · Other Test"
-      agent_pool: mi325_1
+      agent_pool: mi300_1
      depends_on: amd-build
-      timeout_in_minutes: 50
      mirror_hardwares: [amdproduction]
      grade: Blocking
      commands:
        - export VLLM_ROCM_USE_AITER=0
        # ignore test_teacache_extractors.py because it use rocm gemm kernel from vLLM
        # that is not supported on CPU
-        - "timeout 40m pytest -sv tests/diffusion tests/model_executor -m 'core_model and cpu' --ignore=tests/diffusion/cache/test_teacache_extractors.py --cov=vllm_omni --cov-report=term-missing:skip-covered --cov-report=xml"
+        - "pytest -sv tests/ -m 'core_model and cpu' --ignore=tests/diffusion --ignore=tests/model_executor --ignore=tests/entrypoints --ignore=tests/engine  --ignore=tests/diffusion/cache/test_teacache_extractors.py --cov=vllm_omni --cov-report=term-missing:skip-covered --cov-report=xml"

 - group: ":card_index_dividers: Diffusion Test"
  steps:
    - label: "Diffusion · LoRA Test"
-      agent_pool: mi325_1
+      agent_pool: mi300_1
      depends_on: amd-build
      mirror_hardwares: [amdproduction]
      grade: Blocking
+      artifact_paths:
+        - lora-server.log
      commands:
-        - timeout 20m pytest -s -v tests/diffusion/lora/test_images_generations_lora.py
+        - pytest -sv tests/diffusion/lora/ -m 'advanced_model and cuda' --run-level "advanced_model" 2>&1 | tee lora-server.log

    - label: "Diffusion Model CPU offloading Test"
-      agent_pool: mi325_1
+      agent_pool: mi300_1
      depends_on: amd-build
      mirror_hardwares: [amdproduction]
      grade: Blocking
@@ -62,7 +72,7 @@ steps:
        - export VLLM_LOGGING_LEVEL=DEBUG
        - export VLLM_WORKER_MULTIPROC_METHOD=spawn
        - |
-          timeout 20m bash -c '
+          bash -c '
            set +e
            pytest -s -v tests/diffusion/offloader/test_diffusion_cpu_offload.py
            EXIT1=\$?
@@ -72,7 +82,7 @@ steps:
          '

    - label: "Diffusion Sequence Parallelism Test (Need 4 GPUs)"
-      agent_pool: mi325_4
+      agent_pool: mi300_4
      depends_on: amd-build
      mirror_hardwares: [amdproduction]
      grade: Blocking
@@ -84,7 +94,7 @@ steps:

    # merge-only tests
    - label: "Diffusion Tensor Parallelism Test"
-      agent_pool: mi325_2
+      agent_pool: mi300_2
      depends_on: amd-build
      mirror_hardwares: [amdproduction]
      grade: Blocking
@@ -92,7 +102,7 @@ steps:
        - pytest -s -v tests/diffusion/distributed/test_tensor_parallel.py

    - label: "Diffusion Model Test"
-      agent_pool: mi325_1
+      agent_pool: mi300_1
      depends_on: amd-build
      mirror_hardwares: [amdproduction]
      grade: Blocking
@@ -100,7 +110,7 @@ steps:
        - timeout 30m pytest -s -v tests/e2e/offline_inference/test_z_image.py

    - label: "Diffusion GPU Worker Test"
-      agent_pool: mi325_2
+      agent_pool: mi300_2
      depends_on: amd-build
      mirror_hardwares: [amdproduction]
      grade: Blocking
@@ -108,7 +118,7 @@ steps:
        - timeout 20m pytest -s -v tests/diffusion/test_diffusion_worker.py

    - label: "Diffusion · Wan22 Test"
-      agent_pool: mi325_1
+      agent_pool: mi300_1
      depends_on: amd-build
      mirror_hardwares: [amdproduction]
      grade: Blocking
@@ -121,7 +131,7 @@ steps:

 ## ISSUE depends on `diffusers` package: https://github.com/huggingface/diffusers/issues/13274
 # - label: "Audio Generation Model Test"
-#   agent_pool: mi325_1
+#   agent_pool: mi300_1
 #   depends_on: amd-build
 #   mirror_hardwares: [amdproduction]
 #   grade: Blocking
@@ -132,7 +142,7 @@ steps:
 #     - timeout 20m pytest -s -v tests/e2e/offline_inference/test_stable_audio_expansion.py -m "advanced_model and diffusion and L4" --run-level advanced_model

 - label: "Engine Test"
-  agent_pool: mi325_1
+  agent_pool: mi300_1
  depends_on: amd-build
  mirror_hardwares: [amdproduction]
  grade: Blocking
@@ -147,7 +157,7 @@ steps:
 - group: ":card_index_dividers: E2E Test"
  steps:
    - label: "Omni Model Test Qwen2-5-Omni"
-      agent_pool: mi325_4
+      agent_pool: mi300_4
      depends_on: amd-build
      mirror_hardwares: [amdproduction]
      commands:
@@ -155,7 +165,7 @@ steps:
        - pytest -s -v tests/e2e/*/test_qwen2_5_omni_expansion.py --run-level "core_model"

    - label: "Omni Model Test Qwen3-Omni"
-      agent_pool: mi325_2
+      agent_pool: mi300_2
      depends_on: amd-build
      mirror_hardwares: [amdproduction]
      grade: Blocking
@@ -163,22 +173,23 @@ steps:
        - export VLLM_WORKER_MULTIPROC_METHOD=spawn
        - pytest -s -v tests/e2e/offline_inference/test_qwen3_omni.py tests/e2e/online_serving/test_qwen3_omni.py -m 'advanced_model and cuda' --run-level 'advanced_model'

-    - label: "Qwen3-TTS CustomVoice E2E Test"
-      agent_pool: mi325_1
+    - label: "Qwen3-TTS CustomVoice E2E Test · Shard %N/%t"
+      agent_pool: mi300_1
+      parallelism: 2
      depends_on: amd-build
      mirror_hardwares: [amdproduction]
      grade: Blocking
      commands:
        - |
-          timeout 50m bash -c '
+          bash -c '
            export VLLM_LOGGING_LEVEL=DEBUG
            export VLLM_WORKER_MULTIPROC_METHOD=spawn
            export VLLM_ALLOW_LONG_MAX_MODEL_LEN="1"
-            pytest -s -v tests/e2e/online_serving/test_qwen3_tts_customvoice.py tests/e2e/offline_inference/test_qwen3_tts_customvoice.py -m "advanced_model" --run-level "advanced_model"
+            pytest -s -v tests/e2e/online_serving/test_qwen3_tts_customvoice.py tests/e2e/offline_inference/test_qwen3_tts_customvoice.py -m "advanced_model" --run-level "advanced_model" --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --shard-id=$$BUILDKITE_PARALLEL_JOB
          '

    - label: "Qwen3-TTS Base E2E Test"
-      agent_pool: mi325_1
+      agent_pool: mi300_1
      depends_on: amd-build
      mirror_hardwares: [amdproduction]
      grade: Blocking
@@ -196,7 +207,7 @@ steps:
 # Need to debug before reneable numerical changes across large PRs
 # # split Bagel Model Test with H100 (Real Weights) into three tests
 # - label: "Bagel Text2Img Model Test (1/3)"
-#   agent_pool: mi325_1
+#   agent_pool: mi300_1
 #   depends_on: amd-build
 #   mirror_hardwares: [amdproduction]
 #   grade: Blocking
@@ -208,7 +219,7 @@ steps:
 #     - timeout 30m pytest -s -v tests/e2e/offline_inference/test_bagel.py -m "advanced_model" --run-level "advanced_model" -k "shared_memory" -k "rocm"

 # - label: "Bagel Img2Img Model Test (2/3)"
-#   agent_pool: mi325_1
+#   agent_pool: mi300_1
 #   depends_on: amd-build
 #   mirror_hardwares: [amdproduction]
 #   grade: Blocking
@@ -220,7 +231,7 @@ steps:
 #     - timeout 30m pytest -s -v tests/e2e/offline_inference/test_bagel.py -m "advanced_model" --run-level "advanced_model" -k "rocm"

 # - label: "Bagel Online Serving Test (3/3)"
-#   agent_pool: mi325_1
+#   agent_pool: mi300_1
 #   depends_on: amd-build
 #   mirror_hardwares: [amdproduction]
 #   grade: Blocking
@@ -7,8 +7,9 @@ steps:

 - group: ":card_index_dividers: Simple Test"
  steps:
-    - label: "Simple · Diffusion & Model Executor Test"
-      agent_pool: mi325_1
+    - label: "Simple · Model Executor Test · Shard %N/%t"
+      agent_pool: mi300_1
+      parallelism: 2
      depends_on: amd-build
      mirror_hardwares: [amdproduction]
      grade: Blocking
@@ -16,10 +17,30 @@ steps:
        - export VLLM_ROCM_USE_AITER=0
        # ignore test_teacache_extractors.py because it use rocm gemm kernel from vLLM
        # that is not supported on CPU
-        - "timeout 40m pytest -sv tests/diffusion tests/model_executor -m 'core_model and cpu' --ignore=tests/diffusion/cache/test_teacache_extractors.py --cov=vllm_omni --cov-report=term-missing:skip-covered --cov-report=xml"
+        - "pytest -sv tests/model_executor -m 'core_model and cpu' --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --shard-id=$$BUILDKITE_PARALLEL_JOB --cov=vllm_omni --cov-report=term-missing:skip-covered --cov-report=xml:coverage-model-executor-$$BUILDKITE_PARALLEL_JOB.xml"
+
+    - label: "Simple · Diffusion Test"
+      agent_pool: mi300_1
+      depends_on: amd-build
+      mirror_hardwares: [amdproduction]
+      grade: Blocking
+      commands:
+        - export VLLM_ROCM_USE_AITER=0
+        # ignore test_teacache_extractors.py because it use rocm gemm kernel from vLLM
+        # that is not supported on CPU
+        - "pytest -sv tests/diffusion -m 'core_model and cpu'  --ignore=tests/diffusion/cache/test_teacache_extractors.py --cov=vllm_omni --cov-report=term-missing:skip-covered --cov-report=xml"
+
+    - label: "Simple · Engine&Entrypoints Test"
+      agent_pool: mi300_1
+      depends_on: amd-build
+      mirror_hardwares: [amdproduction]
+      grade: Blocking
+      commands:
+        - export VLLM_ROCM_USE_AITER=0
+        - "pytest -sv tests/entrypoints tests/engine -m 'core_model and cpu' --cov=vllm_omni --cov-report=term-missing:skip-covered --cov-report=xml"

    - label: "Simple · Other Test"
-      agent_pool: mi325_1
+      agent_pool: mi300_1
      depends_on: amd-build
      mirror_hardwares: [amdproduction]
      grade: Blocking
@@ -27,19 +48,19 @@ steps:
        - export VLLM_ROCM_USE_AITER=0
        # ignore test_teacache_extractors.py because it use rocm gemm kernel from vLLM
        # that is not supported on CPU
-        - "timeout 40m pytest -sv tests/diffusion tests/model_executor -m 'core_model and cpu' --ignore=tests/diffusion/cache/test_teacache_extractors.py --cov=vllm_omni --cov-report=term-missing:skip-covered --cov-report=xml"
+        - "pytest -sv tests/ -m 'core_model and cpu' --ignore=tests/diffusion --ignore=tests/model_executor --ignore=tests/entrypoints --ignore=tests/engine  --ignore=tests/diffusion/cache/test_teacache_extractors.py --cov=vllm_omni --cov-report=term-missing:skip-covered --cov-report=xml"

 - label: "Custom Pipeline Test"
-  agent_pool: mi325_1
+  agent_pool: mi300_1
  depends_on: amd-build
  mirror_hardwares: [amdproduction]
  grade: Blocking
  commands:
-    - timeout 20m pytest -s -v tests/e2e/features/custom_pipeline/ -m "core_model"
+    - pytest -s -v tests/e2e/features/custom_pipeline/ -m "core_model"


 - label: "Voxtral TTS CUDA Unit Test"
-  agent_pool: mi325_1
+  agent_pool: mi300_1
  depends_on: amd-build
  mirror_hardwares: [amdproduction]
  grade: Blocking
@@ -49,33 +70,25 @@ steps:
 - group: ":card_index_dividers: Diffusion Test"
  steps:
    - label: "Diffusion · Batch Test"
-      agent_pool: mi325_1
+      agent_pool: mi300_1
      depends_on: amd-build
      mirror_hardwares: [amdproduction]
      grade: Blocking
      commands:
-        - timeout 20m pytest -sv tests/diffusion/batching -m 'core_model and cuda' --run-level "core_model"
+        - pytest -sv tests/diffusion/batching -m 'core_model and cuda' --run-level "core_model"

    - label: "Diffusion · Offloader Test"
-      agent_pool: mi325_1
+      agent_pool: mi300_1
      depends_on: amd-build
      mirror_hardwares: [amdproduction]
      grade: Blocking
      commands:
        - export VLLM_LOGGING_LEVEL=DEBUG
        - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-        - |
-          timeout 20m bash -c '
-            set +e
-            pytest -s -v tests/diffusion/offloader/test_diffusion_cpu_offload.py
-            EXIT1=\$?
-            pytest -s -v tests/diffusion/offloader/test_diffusion_layerwise_offload.py
-            EXIT2=\$?
-            exit \$((EXIT1 | EXIT2))
-          '
+        - pytest -s -v tests/diffusion/offloader -m "core_model and cuda" --run-level "core_model"

    - label: "Diffusion · Cache Test"
-      agent_pool: mi325_1
+      agent_pool: mi300_1
      depends_on: amd-build
      mirror_hardwares: [amdproduction]
      grade: Blocking
@@ -83,10 +96,10 @@ steps:
        - export GPU_ARCHS=gfx942
        - export VLLM_LOGGING_LEVEL=DEBUG
        - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-        - timeout 15m pytest -s -v -m "core_model and cache and diffusion and not distributed_cuda and L4"
+        - pytest -sv tests/diffusion/cache -m 'core_model and cuda' --run-level "core_model"

    - label: "Diffusion Sequence Parallelism Test"
-      agent_pool: mi325_2
+      agent_pool: mi300_2
      depends_on: amd-build
      mirror_hardwares: [amdproduction]
      grade: Blocking
@@ -96,7 +109,7 @@ steps:
        - timeout 20m pytest -s -v tests/diffusion/distributed/test_sequence_parallel.py -m core_model

    - label: "Diffusion GPU Worker Test"
-      agent_pool: mi325_2
+      agent_pool: mi300_2
      depends_on: amd-build
      mirror_hardwares: [amdproduction]
      grade: Blocking
@@ -104,7 +117,7 @@ steps:
        - timeout 20m pytest -s -v tests/diffusion/test_diffusion_worker.py

    - label: "Diffusion · Model Test"
-      agent_pool: mi325_1
+      agent_pool: mi300_1
      depends_on: amd-build
      mirror_hardwares: [amdproduction]
      grade: Blocking
@@ -112,7 +125,7 @@ steps:
        - timeout 15m pytest -sv tests/diffusion/models/ -m 'core_model and cuda' --run-level "core_model"

    - label: "Diffusion · Wan22 Test"
-      agent_pool: mi325_1
+      agent_pool: mi300_1
      depends_on: amd-build
      mirror_hardwares: [amdproduction]
      grade: Blocking
@@ -126,7 +139,7 @@ steps:

 ## ISSUE depends on `diffusers` package: https://github.com/huggingface/diffusers/issues/13274
 # - label: "Audio Generation Model Test"
-#   agent_pool: mi325_1
+#   agent_pool: mi300_1
 #   depends_on: amd-build
 #   mirror_hardwares: [amdproduction]
 #   grade: Blocking
@@ -137,7 +150,7 @@ steps:
 #     - timeout 20m pytest -s -v tests/e2e/offline_inference/test_stable_audio_expansion.py -m "advanced_model and diffusion and L4" --run-level advanced_model

 - label: "Engine Test"
-  agent_pool: mi325_1
+  agent_pool: mi300_1
  depends_on: amd-build
  mirror_hardwares: [amdproduction]
  grade: Blocking
@@ -149,7 +162,7 @@ steps:
      '

 # - label: "Omni Model Test Qwen3-Omni"
-#   agent_pool: mi325_2
+#   agent_pool: mi300_2
 #   depends_on: amd-build
 #   mirror_hardwares: [amdproduction]
 #   grade: Blocking
@@ -162,7 +175,7 @@ steps:
 - group: ":card_index_dividers: E2E Test"
  steps:
    - label: "Qwen3-TTS E2E Test"
-      agent_pool: mi325_1
+      agent_pool: mi300_1
      depends_on: amd-build
      mirror_hardwares: [amdproduction]
      grade: Blocking
@@ -173,7 +186,7 @@ steps:
        - timeout 30m pytest -s -v tests/e2e/online_serving/test_qwen3_tts_customvoice.py -m "core_model" --run-level "core_model"

    - label: "CosyVoice3-TTS E2E Test"
-      agent_pool: mi325_1
+      agent_pool: mi300_1
      depends_on: amd-build
      mirror_hardwares: [amdproduction]
      grade: Blocking
@@ -186,7 +199,7 @@ steps:
 # TODO: Bagel test on ROCm is very unstable. @tjtanaa
 # Need to debug before reneable numerical changes across large PRs
 # - label: "Bagel Text2Img Model Test"
-#   agent_pool: mi325_1
+#   agent_pool: mi300_1
 #   depends_on: amd-build
 #   mirror_hardwares: [amdproduction]
 #   grade: Blocking
@@ -198,7 +211,7 @@ steps:
 #     - timeout 30m pytest -s -v tests/e2e/offline_inference/test_bagel.py -m "core_model" --run-level "core_model" -k "rocm"

 # - label: "Bagel Img2Img Model Test"
-#   agent_pool: mi325_1
+#   agent_pool: mi300_1
 #   depends_on: amd-build
 #   mirror_hardwares: [amdproduction]
 #   grade: Blocking
@@ -210,7 +223,7 @@ steps:
 #     - timeout 30m pytest -s -v tests/e2e/offline_inference/test_bagel.py -m "core_model" --run-level "core_model" -k "rocm"

 # - label: "Bagel Online Serving Test"
-#   agent_pool: mi325_1
+#   agent_pool: mi300_1
 #   depends_on: amd-build
 #   mirror_hardwares: [amdproduction]
 #   grade: Blocking
@@ -1,11 +1,65 @@
 {# vllm-omni customized version
   Based on: https://github.com/vllm-project/ci-infra/blob/main/buildkite/test-template-amd.j2
-   Last synced: 2025-12-15
-   Modifications: Removed unused CUDA/NVIDIA logic, keeping only AMD tests
+   Last synced: 2026-08-04
+   Modifications: Use the shared MI300 queues and Omni command-passing contract
 #}
 {% set docker_image_amd = "rocm/vllm-omni:$BUILDKITE_COMMIT" %}
 {% set default_working_dir = "/app/vllm-omni" %}

+{# MI300 runs tests natively in Kubernetes; Docker-in-Docker is unavailable. #}
+{% macro amd_native_k8s_plugin(indent, step) -%}
+{{ indent }}plugins:
+{{ indent }}  - kubernetes:
+{{ indent }}      podSpecPatch:
+{{ indent }}        automountServiceAccountToken: false
+{{ indent }}        securityContext:
+{{ indent }}          seccompProfile:
+{{ indent }}            type: RuntimeDefault
+{{ indent }}        imagePullSecrets:
+{{ indent }}          - name: docker-config
+{{ indent }}        containers:
+{{ indent }}          - name: container-0
+{{ indent }}            image: {{ docker_image_amd }}
+{{ indent }}            imagePullPolicy: Always
+{{ indent }}            securityContext:
+{{ indent }}              allowPrivilegeEscalation: false
+{{ indent }}              capabilities:
+{{ indent }}                add:
+{{ indent }}                  - IPC_LOCK
+{{ indent }}            resources:
+{{ indent }}              limits:
+{{ indent }}                amd.com/gpu: "{{ step.agent_pool[-1:] }}"
+{{ indent }}              requests:
+{{ indent }}                amd.com/gpu: "{{ step.agent_pool[-1:] }}"
+{{ indent }}            volumeMounts:
+{{ indent }}              - name: devshm
+{{ indent }}                mountPath: /dev/shm
+{{ indent }}            env:
+{{ indent }}              - name: AMD_CI_RUNTIME
+{{ indent }}                value: native
+{{ indent }}              - name: NATIVE_CI
+{{ indent }}                value: "true"
+{{ indent }}              - name: VLLM_CI_DOCKER_DISABLED
+{{ indent }}                value: "1"
+{{ indent }}              - name: VLLM_CI_EXPECTED_GPU_COUNT
+{{ indent }}                value: "{{ step.agent_pool[-1:] }}"
+{{ indent }}              - name: HF_HOME
+{{ indent }}                value: /home/buildkite-agent/huggingface
+{{ indent }}              - name: MIOPEN_DEBUG_CONV_DIRECT
+{{ indent }}                value: "0"
+{{ indent }}              - name: MIOPEN_DEBUG_CONV_GEMM
+{{ indent }}                value: "0"
+{{ indent }}              - name: VLLM_ROCM_USE_AITER
+{{ indent }}                value: "0"
+{{ indent }}              - name: PYTORCH_ROCM_ARCH
+{{ indent }}                value: ""
+{{ indent }}        volumes:
+{{ indent }}          - name: devshm
+{{ indent }}            emptyDir:
+{{ indent }}              medium: Memory
+{{ indent }}              sizeLimit: 16Gi
+{%- endmacro %}
+
 {% if env %}
 env:
 {% for name in env %}
@@ -53,19 +107,31 @@ steps:
          {% if step.agent_pool %}
          queue: amd_{{ step.agent_pool }}
          {% else %}
-          queue: amd_mi325_1
+          queue: amd_mi300_1
          {% endif %}
+{{ amd_native_k8s_plugin("        ", step) }}
 {% set cmd_body = (step.command or (step.commands | join("\n"))) | trim %}
 {% set indented_cmd = cmd_body | replace("\n", "\n            ") %}
        command: bash .buildkite/amd/scripts/run-amd-test.sh
        env:
-          DOCKER_BUILDKIT: "1"
+          AMD_CI_RUNTIME: "native"
+          NATIVE_CI: "true"
+          VLLM_CI_DOCKER_DISABLED: "1"
+          VLLM_CI_EXPECTED_GPU_COUNT: "{{ step.agent_pool[-1:] }}"
+          VLLM_CI_REQUIRE_PERSISTENT_HF_CACHE: "1"
+          HF_HOME: "/home/buildkite-agent/huggingface"
+          HF_HUB_DOWNLOAD_TIMEOUT: "300"
+          HF_HUB_ETAG_TIMEOUT: "60"
+          MIOPEN_DEBUG_CONV_DIRECT: "0"
+          MIOPEN_DEBUG_CONV_GEMM: "0"
+          VLLM_ROCM_USE_AITER: "0"
+          PYTORCH_ROCM_ARCH: ""
 {% if step.env %}
 {% for name in step.env %}
          {{ name }}: "{{ step.env[name] }}"
 {% endfor %}
 {% endif %}
-          TEST_COMMAND: |-
+          TEST_COMMANDS: |-
            (command rocm-smi || true) && cd {{ (step.working_dir or default_working_dir) | safe }}
 {% if env %}
 {% for name in env %}
@@ -82,6 +148,15 @@ steps:
 {% endif %}
            {{ indented_cmd | safe }}
        priority: 100
+{% if step.parallelism %}
+        parallelism: {{ step.parallelism }}
+{% endif %}
+{% if step.artifact_paths %}
+        artifact_paths:
+{% for path in step.artifact_paths %}
+          - "{{ path }}"
+{% endfor %}
+{% endif %}
        {% if step.timeout_in_minutes %}
        timeout_in_minutes: {{ step.timeout_in_minutes }}
        {% endif %}
@@ -100,19 +175,31 @@ steps:
      {% if step.agent_pool %}
      queue: amd_{{ step.agent_pool }}
      {% else %}
-      queue: amd_mi325_1
+      queue: amd_mi300_1
      {% endif %}
+{{ amd_native_k8s_plugin("    ", step) }}
 {% set cmd_body = (step.command or (step.commands | join("\n"))) | trim %}
 {% set indented_cmd = cmd_body | replace("\n", "\n        ") %}
    command: bash .buildkite/amd/scripts/run-amd-test.sh
    env:
-      DOCKER_BUILDKIT: "1"
+      AMD_CI_RUNTIME: "native"
+      NATIVE_CI: "true"
+      VLLM_CI_DOCKER_DISABLED: "1"
+      VLLM_CI_EXPECTED_GPU_COUNT: "{{ step.agent_pool[-1:] }}"
+      VLLM_CI_REQUIRE_PERSISTENT_HF_CACHE: "1"
+      HF_HOME: "/home/buildkite-agent/huggingface"
+      HF_HUB_DOWNLOAD_TIMEOUT: "300"
+      HF_HUB_ETAG_TIMEOUT: "60"
+      MIOPEN_DEBUG_CONV_DIRECT: "0"
+      MIOPEN_DEBUG_CONV_GEMM: "0"
+      VLLM_ROCM_USE_AITER: "0"
+      PYTORCH_ROCM_ARCH: ""
 {% if step.env %}
 {% for name in step.env %}
      {{ name }}: "{{ step.env[name] }}"
 {% endfor %}
 {% endif %}
-      TEST_COMMAND: |-
+      TEST_COMMANDS: |-
        (command rocm-smi || true) && cd {{ (step.working_dir or default_working_dir) | safe }}
 {% if env %}
 {% for name in env %}
@@ -129,6 +216,15 @@ steps:
 {% endif %}
        {{ indented_cmd | safe }}
    priority: 100
+{% if step.parallelism %}
+    parallelism: {{ step.parallelism }}
+{% endif %}
+{% if step.artifact_paths %}
+    artifact_paths:
+{% for path in step.artifact_paths %}
+      - "{{ path }}"
+{% endfor %}
+{% endif %}
    {% if step.timeout_in_minutes %}
    timeout_in_minutes: {{ step.timeout_in_minutes }}
    {% endif %}
@@ -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
@@ -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}"
@@ -76,6 +76,9 @@ steps:
          - tests/e2e/online_serving/test_qwen3_tts_customvoice.py
          - tests/e2e/offline_inference/test_qwen3_tts_customvoice.py
          - vllm_omni/model_executor/models/qwen3_tts/
+          - vllm_omni/model_executor/models/common/qwen3_code_predictor.py
+          - vllm_omni/model_executor/models/common/snake_activation.py
+          - vllm_omni/model_executor/models/common/alias_free_activation.py
          - vllm_omni/model_executor/stage_input_processors/qwen3_tts.py
          - vllm_omni/deploy/qwen3_tts.yaml
        commands:
@@ -92,31 +95,24 @@ steps:
          - tests/e2e/online_serving/test_qwen3_tts_base.py
          - tests/e2e/offline_inference/test_qwen3_tts_base.py
          - vllm_omni/model_executor/models/qwen3_tts/
+          - vllm_omni/model_executor/models/common/qwen3_code_predictor.py
+          - vllm_omni/model_executor/models/common/snake_activation.py
+          - vllm_omni/model_executor/models/common/alias_free_activation.py
          - vllm_omni/model_executor/stage_input_processors/qwen3_tts.py
          - vllm_omni/deploy/qwen3_tts.yaml
+          - .buildkite/common/scripts/run_cov_split.sh
+          - pyproject.toml
+        # 40 -> 60: the coverage split runs the model twice.
+        timeout_in_minutes: 60
        commands:
          - |
-            timeout 40m bash -c "
-              export VLLM_LOGGING_LEVEL=DEBUG
-              export VLLM_ALLOW_LONG_MAX_MODEL_LEN=1
-              pytest -s -v tests/e2e/online_serving/test_qwen3_tts_base.py tests/e2e/offline_inference/test_qwen3_tts_base.py -m 'advanced_model and cuda' --run-level 'advanced_model'
-            "
-        mirror_hardwares: l4_1
-
-      - label: "TTS · Higgs-Audio-V3 Test"
-        source_file_dependencies:
-          - tests/e2e/online_serving/test_higgs_audio_v3.py
-          - vllm_omni/model_executor/models/higgs_audio_v3/
-          - vllm_omni/model_executor/stage_input_processors/higgs_audio_v3.py
-          - vllm_omni/deploy/higgs_multimodal_qwen3.yaml
-        commands:
-          - |
-            timeout 20m bash -c "
-              export VLLM_WORKER_MULTIPROC_METHOD=spawn
-              export VLLM_USE_DEEP_GEMM=0
-              export VLLM_MOE_USE_DEEP_GEMM=0
-              pytest -s -v tests/e2e/online_serving/test_higgs_audio_v3.py -m 'advanced_model' --run-level 'advanced_model'
-            "
+            export VLLM_LOGGING_LEVEL=DEBUG
+            export VLLM_ALLOW_LONG_MAX_MODEL_LEN=1
+            .buildkite/common/scripts/run_cov_split.sh \
+              --model-id qwen3_tts \
+              --offline tests/e2e/offline_inference/test_qwen3_tts_base.py \
+              --online tests/e2e/online_serving/test_qwen3_tts_base.py \
+              -- -m 'advanced_model and cuda' --run-level 'advanced_model'
        mirror_hardwares: l4_1

      - label: "Omni · Qwen3-Omni Test"
@@ -124,6 +120,10 @@ steps:
          - tests/e2e/offline_inference/test_qwen3_omni.py
          - tests/e2e/online_serving/test_qwen3_omni.py
          - vllm_omni/model_executor/models/qwen3_omni/
+          - vllm_omni/model_executor/models/common/qwen3_code_predictor.py
+          - vllm_omni/model_executor/models/common/snake_activation.py
+          - vllm_omni/model_executor/models/qwen2_5_omni/qwen2_5_omni_thinker.py
+          - vllm_omni/model_executor/models/qwen3_tts/cuda_graph_decoder_wrapper.py
          - vllm_omni/model_executor/stage_input_processors/qwen3_omni.py
          - vllm_omni/deploy/qwen3_omni_moe.yaml
        timeout_in_minutes: 50
@@ -135,6 +135,11 @@ steps:
        source_file_dependencies:
          - tests/e2e/offline_inference/test_minicpmo_4_5.py
          - tests/e2e/online_serving/test_minicpmo_4_5.py
+          - vllm_omni/model_executor/models/minicpmo_4_5/
+          - vllm_omni/model_executor/models/step_audio2/
+          - vllm_omni/model_executor/models/cosyvoice3/code2wav_core/
+          - vllm_omni/model_executor/models/cosyvoice3/utils.py
+          - vllm_omni/model_executor/stage_input_processors/minicpmo_4_5_omni.py
          - vllm_omni/deploy/minicpmo_4_5.yaml
        timeout_in_minutes: 50
        commands:
@@ -156,62 +161,35 @@ steps:
          - tests/e2e/online_serving/helpers/minicpmo_realtime_duplex_scenarios.py
          - tests/e2e/online_serving/run_minicpmo_realtime_duplex_multi_session.py
          - tests/e2e/online_serving/run_minicpmo_realtime_duplex_soft_interrupt.py
+          - vllm_omni/deploy/minicpmo_4_5.yaml
          - tests/e2e/online_serving/helpers/minicpmo_4_5_duplex.py
          - vllm_omni/deploy/minicpmo_4_5_duplex.yaml
          - vllm_omni/experimental/fullduplex/
          - vllm_omni/model_executor/models/minicpmo_4_5/
+          - vllm_omni/model_executor/models/step_audio2/
+          - vllm_omni/model_executor/models/cosyvoice3/code2wav_core/
+          - vllm_omni/model_executor/models/cosyvoice3/utils.py
          - vllm_omni/model_executor/stage_input_processors/minicpmo_4_5_omni.py
          - vllm_omni/core/sched/
          - vllm_omni/outputs/multimodal_accumulation.py
          - vllm_omni/engine/stage_init_utils.py
          - vllm_omni/engine/async_omni_engine.py
+          - vllm_omni/engine/orchestrator.py
          - vllm_omni/entrypoints/cli/serve.py
+          - vllm_omni/entrypoints/openai/api_server.py
+          - vllm_omni/entrypoints/openai/duplex_capability.py
+          - vllm_omni/config/stage_config.py
        timeout_in_minutes: 50
        commands:
          - pytest -s -v tests/e2e/online_serving/test_minicpmo_4_5_duplex.py -m 'advanced_model and cuda' --run-level 'advanced_model'
        mirror_hardwares: h100_1

-      - label: "TTS · VoxCPM2 Test"
-        source_file_dependencies:
-          - tests/e2e/offline_inference/test_voxcpm2_tts.py
-          - tests/e2e/online_serving/test_voxcpm2_tts.py
-          - vllm_omni/model_executor/models/voxcpm2/
-          - vllm_omni/deploy/voxcpm2.yaml
-        timeout_in_minutes: 30
-        commands:
-          - |
-            timeout 30m bash -c "
-              export VLLM_LOGGING_LEVEL=DEBUG
-              export VLLM_WORKER_MULTIPROC_METHOD=spawn
-              pytest -s -v tests/e2e/offline_inference/test_voxcpm2_tts.py tests/e2e/online_serving/test_voxcpm2_tts.py -m 'advanced_model and cuda' --run-level 'advanced_model'
-            "
-        mirror_hardwares: l4_1
-
-      - label: "Diffusion · Qwen Image Edit Test"
-        source_file_dependencies:
-          - tests/e2e/online_serving/test_qwen_image_edit.py
-          - vllm_omni/diffusion/models/qwen_image/
-        timeout_in_minutes: 20
-        commands:
-          - pytest -s -v tests/e2e/online_serving/test_qwen_image_edit.py -m 'advanced_model and cuda' --run-level 'advanced_model'
-        mirror_hardwares: h100_1
-      - label: "Diffusion · Bagel Test"
-        source_file_dependencies:
-          - tests/e2e/offline_inference/test_bagel.py
-          - tests/e2e/online_serving/test_bagel.py
-          - vllm_omni/model_executor/models/bagel/
-          - vllm_omni/diffusion/models/bagel/
-          - vllm_omni/model_executor/stage_input_processors/bagel.py
-          - vllm_omni/deploy/bagel.yaml
-        timeout_in_minutes: 20
-        commands:
-          - pytest -s -v tests/e2e/offline_inference/test_bagel.py tests/e2e/online_serving/test_bagel.py -m 'advanced_model and cuda' --run-level 'advanced_model'
-        mirror_hardwares: h100_1
-
      - label: "Diffusion · Qwen Image Test"
        source_file_dependencies:
          - tests/e2e/online_serving/test_qwen_image.py
          - vllm_omni/diffusion/models/qwen_image/
+          - vllm_omni/diffusion/models/dmd2/
+          - vllm_omni/diffusion/distributed/autoencoders/autoencoder_kl_qwenimage.py
        commands:
          - |
            timeout 40m bash -c '
@@ -220,23 +198,12 @@ steps:
            '
        mirror_hardwares: h100_1

-      - label: "Diffusion · Qwen Image Layered Test"
-        source_file_dependencies:
-          - tests/e2e/online_serving/test_qwen_image_layered.py
-          - vllm_omni/diffusion/models/qwen_image/
-        commands:
-          - |
-            timeout 40m bash -c "
-              export VLLM_IMAGE_FETCH_TIMEOUT=60
-              pytest -s -v tests/e2e/online_serving/test_qwen_image_layered.py -m 'advanced_model and cuda' --run-level 'advanced_model'
-            "
-        mirror_hardwares: h100_1
-
      - label: "Diffusion · Wan22 Test"
        source_file_dependencies:
          - tests/e2e/offline_inference/test_wan22_t2v.py
          - tests/e2e/online_serving/test_wan22_t2v.py
          - vllm_omni/diffusion/models/wan2_2/
+          - vllm_omni/diffusion/models/dmd2/
        commands:
          - |
            timeout 40m bash -c "
@@ -244,21 +211,3 @@ steps:
              pytest -s -v tests/e2e/offline_inference/test_wan22_t2v.py tests/e2e/online_serving/test_wan22_t2v.py -m 'advanced_model and cuda and diffusion' --run-level 'advanced_model'
            "
        mirror_hardwares: h100_1
-
-      - label: "Diffusion · Z Image Test"
-        source_file_dependencies:
-          - tests/e2e/offline_inference/test_z_image.py
-          - tests/e2e/online_serving/test_z_image.py
-          - vllm_omni/diffusion/models/z_image/
-        timeout_in_minutes: 30
-        commands:
-          - pytest -s -v tests/e2e/offline_inference/test_z_image.py tests/e2e/online_serving/test_z_image.py -m 'advanced_model and cuda and diffusion' --run-level 'advanced_model'
-        mirror_hardwares: l4_1
-      - label: "Diffusion · FLUX.2 Klein Test"
-        source_file_dependencies:
-          - tests/e2e/offline_inference/test_flux2_klein.py
-          - vllm_omni/diffusion/models/flux2_klein/
-        timeout_in_minutes: 30
-        commands:
-          - pytest -s -v tests/e2e/offline_inference/test_flux2_klein.py -m 'advanced_model and cuda and diffusion' --run-level 'advanced_model'
-        mirror_hardwares: l4_1
@@ -456,9 +456,15 @@ 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 · LTX-2.5 Online Test"
+        timeout_in_minutes: 60
+        commands:
+          - pytest -s -v tests/e2e/online_serving/test_ltx25.py -m "full_model and cuda" --run-level "full_model"
+        mirror_hardwares: h100_1
+
      - label: ":full_moon: Diffusion X2V · Wan2.2 Accuracy Test"
        timeout_in_minutes: 180
        commands:
@@ -477,10 +483,23 @@ 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 · LTX-2.5 Official Parity Test"
+        timeout_in_minutes: 180
+        commands:
+          - pytest -s -v tests/e2e/accuracy/ltx/test_ltx25_official_similarity.py::test_ltx25_distilled_two_stage_matches_official -m full_model --run-level full_model
+          - pytest -s -v tests/e2e/accuracy/ltx/test_ltx25_official_similarity.py::test_ltx25_full_matches_official -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
@@ -169,6 +169,10 @@ steps:
        source_file_dependencies:
          - tests/e2e/online_serving/test_qwen3_omni.py
          - vllm_omni/model_executor/models/qwen3_omni/
+          - vllm_omni/model_executor/models/common/qwen3_code_predictor.py
+          - vllm_omni/model_executor/models/common/snake_activation.py
+          - vllm_omni/model_executor/models/qwen2_5_omni/qwen2_5_omni_thinker.py
+          - vllm_omni/model_executor/models/qwen3_tts/cuda_graph_decoder_wrapper.py
          - vllm_omni/model_executor/stage_input_processors/qwen3_omni.py
          - vllm_omni/deploy/qwen3_omni_moe.yaml
        commands:
@@ -181,6 +185,11 @@ steps:
      - label: "Omni · MiniCPM-o 4.5 Offline Test"
        source_file_dependencies:
          - tests/e2e/offline_inference/test_minicpmo_4_5.py
+          - vllm_omni/model_executor/models/minicpmo_4_5/
+          - vllm_omni/model_executor/models/step_audio2/
+          - vllm_omni/model_executor/models/cosyvoice3/code2wav_core/
+          - vllm_omni/model_executor/models/cosyvoice3/utils.py
+          - vllm_omni/model_executor/stage_input_processors/minicpmo_4_5_omni.py
          - vllm_omni/deploy/minicpmo_4_5.yaml
        commands:
          - |
@@ -192,6 +201,11 @@ steps:
      - label: "Omni · MiniCPM-o 4.5 Online Test"
        source_file_dependencies:
          - tests/e2e/online_serving/test_minicpmo_4_5.py
+          - vllm_omni/model_executor/models/minicpmo_4_5/
+          - vllm_omni/model_executor/models/step_audio2/
+          - vllm_omni/model_executor/models/cosyvoice3/code2wav_core/
+          - vllm_omni/model_executor/models/cosyvoice3/utils.py
+          - vllm_omni/model_executor/stage_input_processors/minicpmo_4_5_omni.py
          - vllm_omni/deploy/minicpmo_4_5.yaml
        commands:
          - |
@@ -210,16 +224,24 @@ steps:
          - tests/e2e/online_serving/helpers/minicpmo_realtime_duplex_scenarios.py
          - tests/e2e/online_serving/run_minicpmo_realtime_duplex_multi_session.py
          - tests/e2e/online_serving/run_minicpmo_realtime_duplex_soft_interrupt.py
+          - vllm_omni/deploy/minicpmo_4_5.yaml
          - tests/e2e/online_serving/helpers/minicpmo_4_5_duplex.py
          - vllm_omni/deploy/minicpmo_4_5_duplex.yaml
          - vllm_omni/experimental/fullduplex/
          - vllm_omni/model_executor/models/minicpmo_4_5/
+          - vllm_omni/model_executor/models/step_audio2/
+          - vllm_omni/model_executor/models/cosyvoice3/code2wav_core/
+          - vllm_omni/model_executor/models/cosyvoice3/utils.py
          - vllm_omni/model_executor/stage_input_processors/minicpmo_4_5_omni.py
          - vllm_omni/core/sched/
          - vllm_omni/outputs/multimodal_accumulation.py
          - vllm_omni/engine/stage_init_utils.py
          - vllm_omni/engine/async_omni_engine.py
+          - vllm_omni/engine/orchestrator.py
          - vllm_omni/entrypoints/cli/serve.py
+          - vllm_omni/entrypoints/openai/api_server.py
+          - vllm_omni/entrypoints/openai/duplex_capability.py
+          - vllm_omni/config/stage_config.py
        commands:
          - |
            timeout 20m bash -c "
@@ -239,25 +261,13 @@ steps:
          - pytest -s -v tests/dfx/perf/scripts/run_benchmark.py --test-config-file tests/dfx/perf/tests/test_minicpmo_4_5_ready.json -m 'core_model and cuda'
        mirror_hardwares: h100_1

-      - label: "TTS · VoxCPM2 Test"
-        source_file_dependencies:
-          - tests/e2e/online_serving/test_voxcpm2_tts.py
-          - vllm_omni/model_executor/models/voxcpm2/
-          - vllm_omni/deploy/voxcpm2.yaml
-        timeout_in_minutes: 30
-        commands:
-          - |
-            timeout 30m bash -c "
-              export VLLM_LOGGING_LEVEL=DEBUG
-              export VLLM_WORKER_MULTIPROC_METHOD=spawn
-              pytest -s -v tests/e2e/online_serving/test_voxcpm2_tts.py -m 'core_model' --run-level 'core_model'
-            "
-        mirror_hardwares: l4_1
-
      - label: "TTS · Qwen3-TTS CustomVoice Test"
        source_file_dependencies:
          - tests/e2e/online_serving/test_qwen3_tts_customvoice.py
          - vllm_omni/model_executor/models/qwen3_tts/
+          - vllm_omni/model_executor/models/common/qwen3_code_predictor.py
+          - vllm_omni/model_executor/models/common/snake_activation.py
+          - vllm_omni/model_executor/models/common/alias_free_activation.py
          - vllm_omni/model_executor/stage_input_processors/qwen3_tts.py
          - vllm_omni/deploy/qwen3_tts.yaml
        commands:
@@ -272,6 +282,9 @@ steps:
        source_file_dependencies:
          - tests/e2e/online_serving/test_qwen3_tts_base.py
          - vllm_omni/model_executor/models/qwen3_tts/
+          - vllm_omni/model_executor/models/common/qwen3_code_predictor.py
+          - vllm_omni/model_executor/models/common/snake_activation.py
+          - vllm_omni/model_executor/models/common/alias_free_activation.py
          - vllm_omni/model_executor/stage_input_processors/qwen3_tts.py
          - vllm_omni/deploy/qwen3_tts.yaml
        commands:
@@ -284,52 +297,11 @@ steps:
            "
        mirror_hardwares: l4_1

-      - label: "TTS · Higgs-Audio-V3 Test"
-        source_file_dependencies:
-          - tests/e2e/online_serving/test_higgs_audio_v3.py
-          - vllm_omni/model_executor/models/higgs_audio_v3/
-          - vllm_omni/model_executor/stage_input_processors/higgs_audio_v3.py
-          - vllm_omni/deploy/higgs_multimodal_qwen3.yaml
-        commands:
-          - |
-            timeout 20m bash -c "
-              export VLLM_WORKER_MULTIPROC_METHOD=spawn
-              export VLLM_USE_DEEP_GEMM=0
-              export VLLM_MOE_USE_DEEP_GEMM=0
-              pytest -s -v tests/e2e/online_serving/test_higgs_audio_v3.py -m 'core_model' --run-level 'core_model'
-            "
-        mirror_hardwares: l4_1
-
-      - label: "Diffusion · Qwen Image Edit Test"
-        source_file_dependencies:
-          - tests/e2e/online_serving/test_qwen_image_edit.py
-          - vllm_omni/diffusion/models/qwen_image/
-        commands:
-          - |
-            timeout 20m bash -c "
-              pytest -s -v tests/e2e/online_serving/test_qwen_image_edit.py -m 'core_model' --run-level 'core_model'
-            "
-        mirror_hardwares: h100_1
-
-      - label: "Diffusion · Bagel Test"
-        source_file_dependencies:
-          - tests/e2e/online_serving/test_bagel.py
-          - vllm_omni/model_executor/models/bagel/
-          - vllm_omni/diffusion/models/bagel/
-          - vllm_omni/model_executor/stage_input_processors/bagel.py
-          - vllm_omni/deploy/bagel.yaml
-        commands:
-          - |
-            timeout 40m bash -c "
-              export VLLM_IMAGE_FETCH_TIMEOUT=60
-              pytest -s -v tests/e2e/online_serving/test_bagel.py -m 'core_model' --run-level 'core_model'
-            "
-        mirror_hardwares: h100_1
-
      - label: "Diffusion · Wan22 Test"
        source_file_dependencies:
          - tests/e2e/online_serving/test_wan22_t2v.py
          - vllm_omni/diffusion/models/wan2_2/
+          - vllm_omni/diffusion/models/dmd2/
        commands:
          - |
            timeout 40m bash -c "
@@ -342,6 +314,9 @@ steps:
        source_file_dependencies:
          - tests/e2e/online_serving/test_cosmos3.py
          - vllm_omni/diffusion/models/cosmos3/
+          - vllm_omni/diffusion/distributed/autoencoders/autoencoder_kl_wan.py
+          - vllm_omni/experimental/world_models/
+          - vllm_omni/model_extras/cosmos3.py
        commands:
          - |
            timeout 40m bash -c "
@@ -354,6 +329,8 @@ steps:
        source_file_dependencies:
          - tests/e2e/online_serving/test_qwen_image.py
          - vllm_omni/diffusion/models/qwen_image/
+          - vllm_omni/diffusion/models/dmd2/
+          - vllm_omni/diffusion/distributed/autoencoders/autoencoder_kl_qwenimage.py
        commands:
          - |
            timeout 40m bash -c "
@@ -361,15 +338,3 @@ steps:
              pytest -s -v tests/e2e/online_serving/test_qwen_image.py -m 'core_model' --run-level 'core_model'
            "
        mirror_hardwares: h100_1
-
-      - label: "Diffusion · Qwen Image Layered Test"
-        source_file_dependencies:
-          - tests/e2e/online_serving/test_qwen_image_layered.py
-          - vllm_omni/diffusion/models/qwen_image/
-        commands:
-          - |
-            timeout 40m bash -c "
-              export VLLM_IMAGE_FETCH_TIMEOUT=60
-              pytest -s -v tests/e2e/online_serving/test_qwen_image_layered.py -m 'core_model' --run-level 'core_model'
-            "
-        mirror_hardwares: h100_1
@@ -108,6 +108,22 @@ steps:
            --run-level "full_model"
        mirror_hardwares: h100_1

+      # Own step: the image has no NanoCodec, so this file import-skips out of
+      # the mark-selected TTS steps above. The overrides keep nemo_toolkit[tts]
+      # from downgrading transformers; see
+      # examples/offline_inference/text_to_speech/README.md.
+      - label: "TTS · Gepard-1.0 · L4"
+        timeout_in_minutes: 60
+        commands:
+          - export VLLM_ALLOW_LONG_MAX_MODEL_LEN="1"
+          - printf '%s\n' "transformers>=5.5.3" "huggingface-hub>=1.0" > /tmp/gepard-overrides.txt
+          - uv pip install --python "$(python3 -c 'import sys; print(sys.executable)')" "nemo_toolkit[tts]==2.7.3" --overrides /tmp/gepard-overrides.txt
+          - >-
+            pytest -s -v tests/e2e/offline_inference/test_gepard_tts.py
+            -m "tts and cuda"
+            --run-level "full_model"
+        mirror_hardwares: l4_1
+
      - label: "Diffusion · L4"
        timeout_in_minutes: 60
        commands:
@@ -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 -m "omni and xpu and B60"
+    pytest -v -s $XPU_TEST_PATHS -m "advanced_model and xpu and B60"
+    pytest -v -s $XPU_TEST_PATHS -m "omni and xpu and B60"
 '
@@ -46,6 +46,8 @@ htmlcov/
 .cache
 nosetests.xml
 coverage.xml
+coverage-*.xml
+coverage-*.xml.gz
 *.cover
 *.py,cover
 .hypothesis/
@@ -252,6 +254,7 @@ outputs/
 results/
 generated/
 output_*/
+/tests/e2e/accuracy/artifacts/

 # Configuration overrides
 configs/local.yaml
@@ -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.
@@ -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
@@ -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
@@ -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
@@ -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
@@ -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
@@ -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
@@ -106,6 +116,7 @@ nav:
        - Async Diffusion Output: design/feature/async_diffusion_output.md
        - Async Omni Output Materialization: design/feature/omni_async_output_materialization.md
        - Automatic Prefix Caching: design/feature/prefix_caching.md
+        - Realtime AR-Diffusion: design/feature/realtime_ar_diffusion.md
      - Communication:
        - OmniConnector Implementations:
          - Mooncake Store Connector: design/feature/omni_connectors/mooncake_store_connector.md
@@ -125,11 +136,16 @@ 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 +179,8 @@ 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
@@ -1,5 +1,9 @@
 # Contact Us

- For technical questions and feature requests, please use GitHub [Issues](https://github.com/vllm-project/vllm-omni/issues)
- For coordinating contributions and development and discussing with other users and developers, please join `sig-omni` channel in our [Slack](https://slack.vllm.ai/) or use the [vLLM Forum](https://discuss.vllm.ai/)
- For security disclosures, please use GitHub's [Security Advisories](https://github.com/vllm-project/vllm-omni/security/advisories) feature
+- For bugs and feature requests, search or open a GitHub [issue](https://github.com/vllm-project/vllm-omni/issues).
+- For technical and usage questions, join the `sig-omni` channel in [vLLM Slack](https://slack.vllm.ai/) or use the [vLLM Forum](https://discuss.vllm.ai/).
+- For coordinating contributions and development, use `sig-omni`, the forum, or join a [community meeting](meetings.md).
+
+## Security Disclosures
+
+Do not share vulnerability details in public GitHub issues, pull requests, discussions, or Slack channels. To arrange a private report, send a direct message to one of the [lead maintainers](governance.md#lead-maintainers) in [vLLM Slack](https://slack.vllm.ai/). If you cannot send a direct message, you may post a detail-free request in `sig-omni` asking a lead maintainer to contact you privately; do not include any vulnerability details in that request.
@@ -52,31 +52,11 @@ Committers have write access and merge rights. They typically have deep expertis
 - [@ZJY0516](https://github.com/ZJY0516): Diffusion attention backend, kernel fusion, and CustomOp
 - [@yuanheng-zhao](https://github.com/yuanheng-zhao): Diffusion cache, offload, and Omni Support

-## Diffusion Workload Division
+## Reviewer Routing

-This section breaks down diffusion-related responsibilities. If you have PRs touching these areas, please ping the listed owners for review.
+For path-based reviewer routing, consult [`.github/CODEOWNERS`](https://github.com/vllm-project/vllm-omni/blob/main/.github/CODEOWNERS). Detailed subsystem boundaries and safe-change guidance live in the [module design documents](../design/index.md#module-design-documents).

-### Runtime-related
-
- Diffusion models: @RuixiangMa, @wtomin, @xuechendi
-
-### Optimization
-
- Parallel: @RuixiangMa, @wtomin
- Attention backend: @ZJY0516
- Cache: @yuanheng-zhao, @RuixiangMa, @SamitHuang
- Offload: @yuanheng-zhao
- Intel GPU support: @xuechendi
- Quantization: @lishunyang12, @david6666666
- Kernel fusion / communication-computation: @ZJY0516
-
-### Docs & Test
-
- Diffusion docs: @RuixiangMa, @wtomin
-
-## Meetings
-
-Committers hold **bi-weekly meetings** to discuss future directions and collaborations of the project.
+`CODEOWNERS` is operational routing metadata, not a governance or approval policy. Being listed does not grant committer status or merge rights, and it does not mean every listed reviewer must approve a change. The maintainer roster above is authoritative for project roles and merge rights.

 ## Committer Nomination Process

@@ -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/).
@@ -1 +0,0 @@
-# Meetups
@@ -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!
@@ -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

@@ -327,6 +327,99 @@ On PR builds, `upload_pipeline.py` logs `skip '…' (no changes under …)` for
 - Use **groups** for dashboard readability; keep **E2E Test** as the group name CUDA diff filtering expects for `--e2e` nightly runs on main.
 - Prefix labels by model domain: **Omni ·**, **TTS ·**, **Diffusion ·**, **Simple ·**, etc., matching existing steps.

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

 | Check | Command / location |
@@ -4,6 +4,9 @@
 > `experimental/ar_diffusion` contract. It follows the ownership and lifecycle
 > direction in [#5137](https://github.com/vllm-project/vllm-omni/issues/5137),
 > but it is not yet part of the stable module-design hierarchy.
+>
+> The end-to-end request, identity, failure, and memory contracts are documented
+> in [Realtime AR-Diffusion sessions](feature/realtime_ar_diffusion.md).

 `ARDiffusionEngine` is an opt-in diffusion backend for pipelines that keep
 autoregressive attention KV across requests. The engine only selects
@@ -51,14 +54,13 @@ time, and its inherited batch/step entry points also fail explicitly. Supporting
 those paths requires a batch-aware state-binding contract rather than silently
 bypassing `bind_ar_diffusion_state()`.

-The current implementation retains one session per model instance. A request
-for a different `session_id` closes the resident session through the same LRU
-lifecycle used by explicit close and failure cleanup before allocating the new
-one. Self-attention and cross-attention memory are therefore sized for one
-resident session even if the pipeline declares a larger `session_capacity`.
-This prioritizes rollout throughput and prevents block-pool exhaustion before a
-count-based eviction can run. Multi-session residency can be added later with a
-memory-backed capacity calculation or allocation-pressure eviction.
+The implementation may retain multiple sessions per model instance. It selects
+the largest feasible capacity up to the pipeline's `session_capacity`, based on
+self-attention pages, cross-attention allocations, scratch, and declared
+model-owned state. One session is admitted whenever it fits actual free device
+memory; `gpu_memory_fraction` controls additional residency. Requests beyond
+that capacity evict the least-recently-used session through the same lifecycle
+hook used by explicit close and failure cleanup.

 Scratch storage is also capability-driven. `frames_per_block` reserves space for
 an uncommitted current video block, while `max_scratch_tokens_per_branch`
@@ -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.

@@ -117,7 +117,22 @@ After (async D2H):

 ### Batch Split

-When `execute_model_batch` returns a `COMPUTE_DONE` with a single `async_output_id` for the entire batch, the executor splits it into per-request `async_output_id`s (format `{batch_id}/{request_id}`) via `_batch_split_map`. When `OUTPUT_READY` arrives, the pump extracts per-request results from the batch output and resolves each request's output future independently.
+When `execute_model_batch` returns a `COMPUTE_DONE` with a single `async_output_id` for the entire batch, the executor splits it into per-request `async_output_id`s (format `{batch_id}/{request_id}`) via `_batch_split_map`. When `OUTPUT_READY` arrives, the pump extracts per-request results from the batch output and resolves each request's output future independently. If `OUTPUT_READY` arrives before `execute_batch` registers the split map, `execute_batch` directly adopts the early output.
+
+### Reliability, Lifecycle & Timeout Behavior
+
+1. **Drain Before Memory Release (`drain_async_outputs` / `_ASYNC_OUTPUT_DRAIN_TIMEOUT_S`)**:
+   When workers transition into sleep states (`handle_sleep_task`) or invoke memory-releasing methods, `WorkerProc.drain_async_outputs()` blocks up to `_ASYNC_OUTPUT_DRAIN_TIMEOUT_S` (10.0s) until all in-flight background D2H/SHM packing tasks finish. This acts as a synchronization barrier preventing worker sleep cycles from releasing device tensors while the side CUDA stream is still actively reading them. Draining typically completes within 10–50 ms under normal operation.
+
+2. **Timeout Diagnostics (`_ASYNC_OUTPUT_TIMEOUT`)**:
+   The engine waits for `wait_output_ready()` with a timeout of `_ASYNC_OUTPUT_TIMEOUT` (30.0s). If a timeout occurs, the executor logs detailed diagnostic bookkeeping (pending futures, cached outputs, and batch split maps) to identify whether the stall occurred in the worker, pump, or waiter.
+
+3. **Queue Serialization & Atomic Future Resolution**:
+   - `result_mq` writes are serialized with `_result_mq_lock` to protect the single-writer `MessageQueue` against concurrent writes from the worker main loop (`COMPUTE_DONE`) and background thread (`OUTPUT_READY`).
+   - `ResultPumpThread` unpacks shared-memory payloads first, and then atomically resolves waiting futures or populates `_completed_outputs` under `_futures_lock`, avoiding race conditions with concurrent `wait_output_ready()` invocations.
+
+4. **Storage-Aware IPC Packing**:
+   `pack_diffusion_output_shm()` evaluates `max(view_bytes, storage_bytes)` against `_SHM_TENSOR_THRESHOLD` (1 MB) to ensure that per-request slices sharing a large batch storage are moved to POSIX shared memory rather than serialized through pickle over the MessageQueue inline buffer.

 ## Model Coverage

@@ -131,6 +146,7 @@ All models running in request-mode (`step_execution=False`, the default) automat
 |---|---|---|
 | **HunyuanImage-3.0** | Image | `False` |
 | **Qwen-Image** | Image | `True` |
+| **LTX-2.3** | Video / Audio | `False` |

 Other models with `step_execution=False` are also supported but not yet verified.

@@ -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.
@@ -1,156 +0,0 @@
-# Distributed Layerwise Offload
-
-This document describes distributed layerwise offload (DLO) for diffusion
-models. DLO keeps only a small number of DiT blocks on the accelerator and
-streams the remaining blocks from host memory. The distributed backend can
-either shard those host-side weights across an existing parallel group or keep
-the standard loader's rank-local weights and avoid an additional collective.
-
-For user-facing commands, see the
-[distributed layerwise offloading guide](../../user_guide/diffusion/cpu_offload.md)
-and the [Cosmos3 recipe](https://github.com/vllm-project/vllm-omni/blob/main/recipes/cosmos3/Cosmos3-DistOffload.md).
-
-## Status
-
-DLO is implemented for multi-device diffusion execution. The default
-AllGather path is the primary path for DP and SP deployments. The
-`--dlo-no-use-allgather` path is a rank-local compatibility mode: it is useful
-for standard-loader sharding, workstation bring-up, and systems where an
-additional DLO collective is undesirable, but it does not reduce host weight
-storage across ranks.
-
-The compatibility matrix below describes the current implementation. The
-unit-level guards are covered, but not every parallelism combination has a
-full model-and-hardware end-to-end test.
-
-## Design
-
-### DLO consumes the existing parallel topology
-
-DLO does not create a new DP, TP, or SP topology. It reads the configured
-`DiffusionParallelConfig` and attaches offload hooks to the DiT blocks after the
-standard distributed groups have been initialized.
-
-The DLO weight-sharding group is selected as follows:
-
-1. Use the existing DP group when `data_parallel_size > 1`.
-2. When DP is one and SP is greater than one, use the SP group.
-3. Otherwise, run rank-locally without a DLO process group.
-
-TP is deliberately not used as DLO's AllGather group. HSDP has its own
-parameter-sharding lifecycle and is not allowed to be sharded a second time by
-DLO's AllGather path.
-
-### AllGather path
-
-With the default `dlo_use_allgather=True`, each rank stores approximately
-`1 / group_size` of each streamable block in pinned host memory. The next
-block's shard is copied to a device buffer and reconstructed with
-`all_gather_into_tensor` on a communication stream while the current block is
-executing.
-
-```text
-Compute:    [Block N]             [Block N+1]          [Block N+2]
-H2D:                      [shard N+1]           [shard N+2]
-AllGather:                [full N+1]             [full N+2]
-Buffers:    [current slot]       [prefetch slot]       [current slot]
-```
-
-![DLO double-buffer prefetch pipeline](../figures/dlo/dlo_pipeline.gif)
-
-The backend uses two shared device buffers, so accelerator weight residency is
-bounded by the largest streamed blocks rather than the complete model.
-
-When DP is greater than one, the engine can process one request per DP rank in
-the same denoising wave. Because AllGather is a collective, all participating
-requests must take the same execution path at every denoising step.
-
-### Rank-local path without DLO AllGather
-
-With `--dlo-no-use-allgather`, DLO forces its internal offload shard size to
-one. The regular model loader remains responsible for preparing each rank's
-weights, including TP-local tensors or HSDP-managed parameters. DLO then
-streams those rank-local tensors block by block using H2D copies only.
-
-This mode means:
-
- DP still provides independent replicas, but DLO does not shard weights
-  across DP ranks.
- SP still performs its normal activation/attention collectives, but DLO does
-  not shard weights across SP ranks.
- TP/HSDP/SP collectives, if configured, are not disabled by this flag; only
-  DLO's additional weight AllGather is disabled.
- Pure DP deployments keep a full host-side model copy per rank, subject to
-  shared OS page-cache behavior.
- The scheduler does not require a synchronized DP request wave for DLO.
-
-This path is intentionally not implemented by reusing the AllGather mmap
-loader. It relies on the standard loader so model-specific TP/HSDP transforms
-remain intact.
-
-## Parallelism compatibility
-
-| Parallelism | DLO + AllGather | DLO without AllGather |
-|---|---|---|
-| **DP** | Supported primary path. DLO shards host weights across the DP group and can run DP multi-concurrency. | Supported rank-local path. DP replicas remain independent; no DLO weight sharding or cross-DP DLO collective. |
-| **SP** | Supported in the implementation. With DP=1, DLO uses the SP group for host-weight sharding; SP still shards sequence/activation work. | SP remains active, but DLO keeps standard-loader rank-local weights and adds no SP weight collective. |
-| **TP > 1** | Unsupported in the DLO mmap path. TP-aware loader callbacks are bypassed, so the backend rejects this configuration when it enters that path. | Standard loading is retained and TP-local tensors can be streamed. This is the intended compatibility path, but it still needs broader model and hardware validation. |
-| **HSDP** | Rejected. HSDP has already sharded parameters, so DLO AllGather would double-shard them. | Accepted by configuration. HSDP owns parameter sharding and its own gathers; DLO only stages rank-local parameters. End-to-end coverage is limited. |
-
-### Combined dimensions
-
- **DP + SP:** DLO uses the DP group for weight sharding when DP is greater
-  than one; SP continues to use its own sequence-parallel group. If DP is one,
-  the SP group becomes DLO's sharding group in AllGather mode.
- **DP + TP/SP without AllGather:** standard model loading defines the
-  rank-local tensor layout. DLO adds no cross-DP, cross-TP, or cross-SP weight
-  collective.
- **HSDP + SP:** the general parallel configuration permits HSDP over SP, but
-  DLO must use `--dlo-no-use-allgather`. HSDP remains responsible for weight
-  materialization and synchronization.
- **HSDP + DP or TP:** rejected independently by the diffusion parallel
-  configuration.
-
-## Request and loading constraints
-
-AllGather DP multi-concurrency requires:
-
- explicit `num_inference_steps`;
- the same `num_inference_steps` for all requests in a wave; and
- identical request arguments that affect the collective execution path.
-
-The no-AllGather path does not impose these DLO-specific synchronized-wave
-requirements.
-
-The mmap loader is used only by the supported DLO+AllGather path when the
-model has a compatible checkpoint layout. Online quantization is incompatible
-with that sharded mmap path; use `--dlo-no-use-allgather` or disable online
-quantization. Models that do not meet the mmap requirements use the regular
-loader path.
-
-## Validation coverage
-
-Current source-level validation includes:
-
- HSDP + DLO + AllGather rejection;
- HSDP + DLO without AllGather acceptance at configuration level;
- TP rejection in the DLO+AllGather mmap path;
- resident-layer requests requiring no-AllGather;
- DP request-wave validation for denoising-step compatibility;
- sharding, double-buffer, AllGather-size, and heterogeneous-block regression
-  tests.
-
-The highest-value missing coverage is end-to-end numerical comparison against
-ordinary layerwise offload for DP+SP, TP+no-AllGather, and HSDP+SP+no-AllGather
-on the target CUDA/NCCL or CANN/HCCL hardware.
-
-## Recommendations
-
- Use **DP + DLO AllGather** for the supported throughput and host-memory
-  scaling path.
- Use **SP + DLO AllGather** for long-sequence workloads when DP concurrency is
-  not the goal.
- Use **no-AllGather** to bring up TP or workstation PCIe configurations, with
-  the expectation of higher host-memory use and lower validation confidence.
- Prefer **HSDP alone** for production HSDP deployments until the combined
-  HSDP + DLO no-AllGather path has broader end-to-end coverage.
@@ -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.
@@ -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).
@@ -0,0 +1,472 @@
+# Distributed Layerwise Offload
+
+This document describes distributed layerwise offload (DLO) for diffusion
+models. DLO keeps only a small number of DiT blocks on the accelerator and
+streams the remaining blocks from host memory. The distributed backend can
+either shard those host-side weights across an existing parallel group or keep
+complete rank-local block sources and avoid an additional collective.
+
+For user-facing commands, see the
+[distributed layerwise offloading guide](../../../user_guide/diffusion/offloader/distributed_layerwise_offload.md)
+and the [Cosmos3 recipe](https://github.com/vllm-project/vllm-omni/blob/main/recipes/cosmos3/Cosmos3-DistOffload.md).
+
+## Status
+
+DLO is implemented for multi-device diffusion execution. The default
+AllGather path is the primary path for DP and SP deployments. The
+`--dlo-no-use-allgather` path streams complete blocks independently and adds no
+DLO weight collective.
+
+Host storage is selected separately from the transfer protocol. The loader can
+produce a direct-checkpoint mmap plan for a proven-compatible runtime layout;
+otherwise it uses the ordinary loader. In no-AllGather mode, an opt-in Phase B
+runtime cache can normalize those final ordinary-loader DiT tensors into an
+immutable node-local mmap entry. Consequently, replicas can share either
+proven-compatible checkpoint pages or equivalent final runtime layouts. An
+opt-in CUDA registration budget can make the complete final mapping directly
+H2D-copyable, avoiding the recurrent host packing otherwise required by mmap.
+
+The Phase A shared-mmap support boundary is TP1. TP greater than one is an
+ordinary-loader compatibility path: DLO can consume the resulting TP-local
+tensors. Direct checkpoint mmap still does not cover TP, but the Phase B cache
+can share each final TP coordinate among equivalent processes.
+
+The compatibility matrix below describes the current implementation. The
+unit-level guards are covered, but not every parallelism combination has a
+full model-and-hardware end-to-end test.
+
+## Design
+
+### DLO consumes the existing parallel topology
+
+DLO does not create a new DP, TP, or SP topology. It reads the configured
+`DiffusionParallelConfig` and attaches offload hooks to the DiT blocks after the
+standard distributed groups have been initialized.
+
+The DLO weight-sharding group is selected as follows:
+
+1. Use the existing DP group when `data_parallel_size > 1`.
+2. When DP is one and SP is greater than one, use the SP group.
+3. Otherwise, run rank-locally without a DLO process group.
+
+TP is deliberately not used as DLO's AllGather group. HSDP has its own
+parameter-sharding lifecycle and is not allowed to be sharded a second time by
+DLO's AllGather path.
+
+### The loader owns host-weight planning
+
+Before it decides whether ordinary weight materialization can be skipped, the
+diffusion loader builds one `HostWeightPlan`. A direct-checkpoint mmap plan is
+accepted only when preflight proves all of the following:
+
+- every required DiT parameter and persistent buffer has exactly one source;
+- runtime names, checkpoint keys, shapes, and dtypes match;
+- the runtime topology is TP1 without HSDP or online quantization; and
+- every custom loader operation is represented by a loader-owned checkpoint
+  adapter.
+
+The exact plan object is handed to DLO. The backend does not rescan checkpoint
+files, repeat the capability decision, or reconstruct names from its block
+topology. If preflight fails, the loader materializes weights normally and DLO
+consumes those runtime tensors.
+
+The plan owns only dedicated DiT component sources. If a pipeline also exposes
+ordinary sources for a text encoder or another non-DiT component, the loader
+still consumes those sources and includes their loaded names in its strict
+coverage check. Only the source prefixes covered by the plan skip ordinary
+materialization. A source that mixes DiT and non-DiT weights fails closed to
+the complete ordinary-loader path because it cannot be skipped safely as a
+unit.
+
+This boundary keeps checkpoint semantics out of DLO and avoids model-pipeline
+flags such as `_supports_mmap_loading` or parameter attributes for mmap-only
+transforms. Model-specific direct-layout knowledge, when required, lives in a
+checkpoint adapter beside the ordinary loader.
+
+### Post-load runtime cache
+
+The Phase B cache preserves the same ownership boundary. DLO neither derives
+cache keys nor interprets model loaders. When enabled with no-AllGather, every
+rank first completes the ordinary loader, custom callbacks, post-load
+processing, strict validation, and calibration. The loader then:
+
+1. discovers final CPU DiT parameters and persistent buffers;
+2. rejects unsupported non-contiguous, aliased, quantized, distributed, or
+   ambiguous layouts before changing the model;
+3. hashes tensor names, metadata, and final bytes;
+4. builds or joins one content-addressed entry under a strict POSIX lock; and
+5. returns a `HostWeightPlan(backing_kind="runtime_cache")` only after file,
+   manifest, and remapped-tensor content validation succeeds.
+
+The identity excludes DP size/rank and SP rank. It includes TP world size/rank
+and conservative SP implementation/world-size guards. DP replicas and
+equivalent SP ranks can therefore share; different TP coordinates cannot.
+Final tensor content is authoritative, so an unmodeled loader/config change
+that changes bytes creates a different entry instead of risking silent
+sharing.
+
+Publication writes sharded safetensors and a JSON manifest into a same-
+filesystem temporary directory, flushes them, and atomically renames the
+directory. The manifest includes per-file SHA-256 digests and the final tensor
+content digest. Kernel `flock` ownership distinguishes a live writer from a
+stale lock file: process death releases the lock automatically. Stale temporary
+entries are removed only by the next lock owner. A bounded wait or any cache
+error retains the already-valid ordinary tensors.
+
+DLO maps this plan without moving the DiT to `meta` and without rerunning any
+post-load hook. It changes tensor storage while preserving `Parameter` objects,
+then runs the existing `gc.collect()`/`malloc_trim()` cleanup before installing
+streaming hooks. The source mappings remain open through DLO disable. With the
+default zero registration budget they feed the existing bounded two-slot host
+staging path. Every streamed block is first packed from pageable mmap storage
+into one of those pinned slots and is then copied to the device. This removes
+model-sized private pinned storage, but adds a recurrent host-to-host copy and
+can amplify TP synchronization waits when several ranks compete for host
+bandwidth.
+
+This version is deliberately opt-in and steady-state-oriented. It requires
+roughly one local-disk model copy per runtime identity, has no eviction, and
+still performs ordinary loading plus full content-hash passes in every rank.
+Skip-load-on-hit is a separate follow-up after loader side effects can be
+modeled safely.
+
+### Registered runtime-cache transfer
+
+`dlo_runtime_cache_pin_limit_gib > 0` opts a CUDA worker into complete
+runtime-cache registration. It also changes loader selection: no-AllGather
+uses the ordinary loader and final runtime cache even when a TP1 direct-
+checkpoint plan is available. A checkpoint binding can carry a deferred
+adapter such as grouped-QKV reordering. Registering that raw source would not
+remove the adapter's recurrent CPU allocation/copy, whereas the runtime cache
+contains the transform-complete final tensors.
+
+After mapping and validating every final tensor, DLO groups source storages by
+backing file, page-aligns their address ranges, and coalesces overlapping or
+adjacent ranges within that mapping. The complete byte count is checked
+against the per-worker budget before any CUDA call. It then registers every
+range with the read-only CUDA host-registration flag and verifies that PyTorch
+recognizes every mapped tensor as pinned. Failure rolls back any ranges already
+registered and preserves the existing two-slot staging path. Partial direct
+transfer is deliberately unsupported. A rollback failure aborts worker
+initialization so a still-registered range is retained until process/context
+teardown rather than being unsafely unmapped.
+
+On success, the hooks copy each immutable tensor view directly into its offset
+in the existing flattened rotating device buffer on the copy stream. Parameter
+repointing, two-block HBM residency, and H2D/compute overlap are unchanged. The
+implementation currently issues more H2D copy operations than block packing,
+but avoids the much larger recurrent CPU copy. A packed cache schema should be
+considered only if launch-count measurements justify its added format and
+publication complexity.
+
+Registration is per process and CUDA context; physical file-cache pages remain
+shared and are not copied into one private host allocation per worker. The
+operator must nevertheless budget page-locked memory and satisfy OS/CUDA
+registration limits for every worker. During shutdown, DLO first synchronizes
+outstanding device work, unregisters all ranges, and then closes the safetensors
+mappings. Worker teardown invokes this cleanup before destroying the
+distributed environment and CUDA context.
+
+### AllGather path
+
+With the default `dlo_use_allgather=True`, each rank stores approximately
+`1 / group_size` of each streamable block in pinned host memory. The next
+block's shard is copied to a device buffer and reconstructed with
+`all_gather_into_tensor` on a communication stream while the current block is
+executing.
+
+```text
+Compute:    [Block N]             [Block N+1]          [Block N+2]
+H2D:                      [shard N+1]           [shard N+2]
+AllGather:                [full N+1]             [full N+2]
+Buffers:    [current slot]       [prefetch slot]       [current slot]
+```
+
+![DLO double-buffer prefetch pipeline](../../figures/dlo/dlo_pipeline.gif)
+
+The backend uses two shared device buffers, so accelerator weight residency is
+bounded by the largest streamed blocks rather than the complete model.
+
+When direct checkpoint mmap is selected, the checkpoint mappings are only the
+source used to prepare each rank's persistent shard. They can be closed after
+shard preparation. Across the AllGather group, those private shards total
+approximately one runtime model copy.
+
+An effective DLO group size of one performs no collective, even when
+`dlo_use_allgather=True`; it follows the rank-local transfer path described
+below.
+
+When DP is greater than one, the engine can process one request per DP rank in
+the same denoising wave. Because AllGather is a collective, all participating
+requests must take the same execution path at every denoising step.
+
+Fast NVLink/NVSwitch changes collective cost, not this execution contract. For
+independently scheduled replicas—even on one NVSwitch baseboard—no-AllGather
+plus shared mmap storage avoids coupling their request schedules and failure
+domains.
+
+### Rank-local path without DLO AllGather
+
+With `--dlo-no-use-allgather`, DLO forces its internal offload shard size to
+one and streams complete blocks using H2D copies only. The host backing may be
+a loader-approved checkpoint mapping, a normalized runtime-cache mapping, or
+ordinary runtime tensors.
+
+For direct mmap, each process retains immutable safetensors views and uses two
+bounded pinned host staging slots. Processes on the same node that map the same
+files share physical checkpoint pages through the OS page cache. This removes
+the persistent private full-model copy per pure-DP process, but each process
+still packs and transfers every complete block. Sharing is node-local; each
+node has its own page cache.
+
+For a runtime-cache mapping, a positive registration budget replaces those
+host slots with direct copies from the shared final tensor views. A zero budget
+or failed registration retains the bounded staging behavior.
+
+When direct mmap preflight fails, the regular model loader remains responsible
+for preparing each rank's weights, including TP-local tensors or HSDP-managed
+parameters. Supported CPU layouts may then enter the opt-in runtime cache.
+Unsupported layouts keep one private runtime copy per process.
+
+This mode means:
+
+- DP still provides independent replicas, but DLO does not shard weights
+  across DP ranks.
+- SP still performs its normal activation/attention collectives, but DLO does
+  not shard weights across SP ranks.
+- TP/HSDP/SP collectives, if configured, are not disabled by this flag; only
+  DLO's additional weight AllGather is disabled.
+- Pure DP deployments share one checkpoint-backed copy per node when direct
+  mmap is selected. The runtime cache can share equivalent ordinary-loader
+  outputs; its fallback keeps one private runtime copy per rank.
+- The scheduler does not require a synchronized DP request wave for DLO.
+
+## Parallelism compatibility
+
+| Parallelism | DLO + AllGather | DLO without AllGather |
+|---|---|---|
+| **DP** | Supported primary path. DLO shards host weights across the DP group and can run DP multi-concurrency. | Supported rank-local path. Compatible TP1 replicas can share checkpoint pages; the optional runtime cache shares equivalent final layouts while excluding DP rank from its identity. CUDA workers may register the complete final mapping for direct H2D. |
+| **SP** | Supported in the implementation. With DP=1, DLO uses the SP group for host-weight sharding; SP still shards sequence/activation work. | SP remains active without a DLO weight collective. Runtime-cache v1 excludes SP rank and includes conservative SP implementation/world-size guards. Registered transfer does not remove ordinary SP activation/attention collectives. |
+| **TP > 1** | Outside the Phase A direct-mmap scope. The loader preserves TP-local layouts and DLO may apply DP/SP host sharding to ordinary runtime tensors. | The ordinary TP-aware loader produces rank-local tensors. Runtime-cache v1 creates one entry per TP coordinate, shareable by equivalent DP/SP processes. Registering those final mappings removes staging-driven TP skew but does not remove ordinary TP collectives. |
+| **HSDP** | Rejected. HSDP has already sharded parameters, so DLO AllGather would double-shard them. | Accepted by configuration. HSDP owns parameter sharding and its own gathers; DLO only stages rank-local parameters. End-to-end coverage is limited. |
+
+### Combined dimensions
+
+- **DP + SP:** DLO uses the DP group for weight sharding when DP is greater
+  than one; SP continues to use its own sequence-parallel group. If DP is one,
+  the SP group becomes DLO's sharding group in AllGather mode.
+- **DP + TP/SP without AllGather:** standard model loading defines the
+  rank-local tensor layout. DLO adds no cross-DP, cross-TP, or cross-SP weight
+  collective.
+- **HSDP + SP:** the general parallel configuration permits HSDP over SP, but
+  DLO must use `--dlo-no-use-allgather`. HSDP remains responsible for weight
+  materialization and synchronization.
+- **HSDP + DP or TP:** rejected independently by the diffusion parallel
+  configuration.
+
+## Request and loading constraints
+
+AllGather DP multi-concurrency requires:
+
+- explicit `num_inference_steps`;
+- the same `num_inference_steps` for all requests in a wave; and
+- identical request arguments that affect the collective execution path.
+
+The no-AllGather path does not impose these DLO-specific synchronized-wave
+requirements.
+
+Direct checkpoint mmap can back either transfer path. It is currently limited
+to proven TP1, non-HSDP, non-online-quantized layouts. Other layouts use the
+ordinary loader. Online quantization remains incompatible with DLO AllGather;
+use `--dlo-no-use-allgather` or disable online quantization. Runtime-cache v1
+is no-AllGather only and rejects all quantization configurations, HSDP/DTensor,
+expert parallelism, CFG parallelism, PP, non-contiguous tensors, and
+tied/shared storage.
+
+## Validation coverage
+
+Current source-level validation includes:
+
+- HSDP + DLO + AllGather rejection;
+- HSDP + DLO without AllGather acceptance at configuration level;
+- loader preflight fallback for TP, HSDP, online quantization, unknown custom
+  loaders, missing keys, and shape/dtype mismatches;
+- exact loader-to-backend plan transfer and ordinary-loader fallback;
+- rank-local mmap source retention, bounded two-slot staging, and adapter
+  transforms without parameter-side flags;
+- runtime-cache content identity, sharded atomic publication, concurrent
+  writer election, corruption rebuild, lock timeout, and fail-closed layout
+  rejection;
+- loader ordering through final post-load mutation and DLO remapping without
+  rerunning post-load hooks;
+- all-or-nothing page-aligned registration, budget rejection, partial-failure
+  rollback, direct-copy staging bypass, and explicit shutdown cleanup;
+- resident-layer requests requiring no-AllGather;
+- DP request-wave validation for denoising-step compatibility;
+- sharding, double-buffer, AllGather-size, and heterogeneous-block regression
+  tests.
+
+### B300 parallel-topology smoke matrix
+
+A four-GPU B300 smoke test covered MiniMax-H3 FL2VA with the same prompt, seed,
+CUDNN attention backend, 256x256 output, two denoising steps, and
+`dlo_resident_layers=0`. The TP2 rows used DiT DP2xTP2 with the text encoder and
+VAEs at TP1. They validate the ordinary-loader fallback only, not direct mmap
+or shared-mmap host-memory savings.
+
+| Configuration | Result | Warm E2E | Peak device memory | Host PSS |
+|---|---:|---:|---:|---:|
+| DP4xTP1 AllGather | Passed, 4 concurrent requests | 2.87 s / 4 requests | 13.84 GiB | 211.99 GiB |
+| DP4xTP1 no-AllGather | Passed, 1 request | 15.02 s | 13.23 GiB | 187.77 GiB |
+| DP2xTP2 AllGather | Passed, 2 concurrent requests | 4.16 s / 2 requests | 12.50 GiB | 211.97 GiB |
+| DP2xTP2 no-AllGather | Passed, 1 request | 3.51 s | 11.88 GiB | 314.01 GiB |
+
+Within each topology, the AllGather and no-AllGather video and audio outputs
+were byte-identical. All four runs completed without an `ERROR` or traceback
+and released their device allocations. For DP4xTP1, no-AllGather direct mmap
+reduced total PSS by 24.22 GiB (11.4%) and `Private_Dirty` from 211.33 to
+125.32 GiB (40.7%) relative to AllGather. For DP2xTP2, preflight selected the
+ordinary loader as designed; no-AllGather PSS was 314.01 GiB, about 48% above
+AllGather, because DP replicas did not share checkpoint-backed runtime
+weights. This is a functional and memory smoke test, not a production-quality
+performance or output-quality benchmark.
+
+### Host-memory measurement
+
+A two-worker MiniMax-H3 FL2VA measurement on one L20X node compared the
+ordinary-loader fallback with direct mmap. Both runs used
+DP=2, TP=1, no DLO AllGather, BF16 weights, two denoising steps, and a
+256x256 four-second request. The ordinary-loader workers were sampled after
+initialization. The mmap workers were sampled after one completed request, so
+the checkpoint working set had been faulted into the page cache; this is the
+more conservative point for mmap.
+
+The values below come from `/proc/<worker>/smaps_rollup` and include the whole
+worker, not only the DiT. The stable rank-to-rank difference comes from other
+pipeline components, so each worker should be compared with the same worker in
+the other storage mode.
+
+| Worker | Ordinary RSS | mmap RSS | Ordinary PSS | mmap PSS | PSS reduction |
+|---|---:|---:|---:|---:|---:|
+| DP worker 0 | 168.27 GiB | 132.76 GiB | 167.84 GiB | 101.43 GiB | 66.40 GiB |
+| DP worker 1 | 116.19 GiB | 79.97 GiB | 115.73 GiB | 48.64 GiB | 67.09 GiB |
+| **Two-worker total** | — | — | **283.56 GiB** | **150.08 GiB** | **133.48 GiB (47.1%)** |
+
+The direct-mmap workers each reported 62.45 GiB `Shared_Clean` but only
+31.20 GiB `Pss_File`, which is the proportional charge expected when the same
+resident checkpoint pages are mapped by two workers. `Private_Dirty` also fell
+from 167.53 to 70.24 GiB for worker 0 and from 115.40 to 17.44 GiB for worker
+1, a reduction of about 97–98 GiB per worker. RSS understates this benefit
+because it counts a shared physical page in every process that maps it; summed
+PSS is the appropriate node-memory comparison.
+
+### L20X runtime-cache smoke matrix
+
+Phase B was also exercised on four L20X GPUs in one NVLink domain with
+MiniMax-H3 FL2VA, CUDNN attention, eager execution, 256x256 four-second
+outputs, two denoising steps, `dlo_resident_layers=0`, text-encoder TP1, and
+VAE TP1. This is a functional, memory, and profiler smoke test rather than a
+production-quality performance benchmark. Host values below are node-wide
+maxima during request waves from descendant-process `smaps_rollup`; they do
+not represent the private startup peak. HBM is the engine-reported request
+peak because sampled `nvidia-smi` data can miss short-lived allocations.
+
+One internal DP2xTP2 engine compared the three host-storage/transfer choices.
+Each measured wave contained two requests. NVLink TX is the physical per-link
+counter delta summed across the four GPUs; RX had the same value and is not
+added again here.
+
+| Mode | Two-request wave | Throughput | Host PSS | `Private_Dirty` | Engine peak HBM | NVLink TX / wave |
+|---|---:|---:|---:|---:|---:|---:|
+| AllGather, ordinary loader | 3.41 s | 0.587 req/s | 210.70 GiB | 209.72 GiB | 12,498 MB | 73.37 GiB |
+| no-AllGather, private loader weights | 6.96 s | 0.287 req/s | 312.56 GiB | 311.56 GiB | 11,882 MB | 19.78 GiB |
+| no-AllGather, runtime cache (cold publish) | 18.94 s | 0.106 req/s | 178.71 GiB | 116.00 GiB | 11,882 MB | 19.78 GiB |
+| no-AllGather, runtime cache (warm reuse) | 24.23 s | 0.083 req/s | 178.29 GiB | 115.58 GiB | 11,882 MB | 19.78 GiB |
+
+The no-AllGather cache preserved the no-AllGather NVLink traffic and HBM
+footprint while reducing node PSS by about 134 GiB relative to private loader
+weights. Cold and warm labels describe cache publication or reuse during
+initialization; both request phases ran after initialization. A warm hit still
+performed ordinary loading and full content hashing, so it was not a startup
+fast path. All measured no-AllGather private/cache outputs were byte-identical
+for matching seeds.
+
+The deployment shape targeted by the cache was tested separately as two
+independent DP1xTP2 engines on the same node. Both engines mapped the same two
+TP-coordinate entries, including the same file device/inode tuples; the cache
+occupied 66,281,995,903 bytes. A start barrier kept all four measured requests
+ahead of profiling.
+
+| Storage | Mean request latency, engines A / B | Combined throughput | Host PSS | `Private_Dirty` | Engine peak HBM |
+|---|---:|---:|---:|---:|---:|
+| Private loader weights | 3.50 / 3.60 s | 0.463 req/s | 366.20 GiB | 365.22 GiB | 11,882 MB |
+| Shared runtime cache | 8.96 / 12.30 s | 0.142 req/s | 232.23 GiB | 169.52 GiB | 11,882 MB |
+
+The shared cache reduced request-time PSS by 133.97 GiB (36.6%) and
+`Private_Dirty` by 195.70 GiB (53.6%), but reduced combined throughput by
+69.3% in this host-contention smoke. Across one profiled request per engine,
+aggregate GPU compute was unchanged (1.359 versus 1.357 seconds) and H2D
+payload was exactly 264.24 GiB in both modes. Aggregate CPU `aten::copy_` time
+rose from 3.75 to 32.49 seconds, while communication-kernel time rose from
+3.56 to 5.80 seconds as TP ranks waited on skewed staging. These durations are
+summed across ranks and can overlap; they identify the mechanism, not wall
+time. Matching private/cache outputs were byte-identical.
+
+The result confirms that zero-budget runtime-cache staging is a host-capacity
+option, not a latency-neutral replacement for private pinned weights. Broader
+SP, model, platform, and production-quality validation remains tracked in
+[issue #6231](https://github.com/vllm-project/vllm-omni/issues/6231). The
+historical B300 rows above predate the runtime cache and must not be treated as
+Phase B memory validation.
+
+### L20X registered-runtime-cache smoke
+
+The read-only registered path was then tested with the same MiniMax-H3
+workload. These are functional and mechanism-validation measurements, not a
+production benchmark. The TP1 pair used one request per wave. Both modes moved
+exactly 121.86 GiB H2D in the profiled request and produced byte-identical video
+and audio for both measured seeds.
+
+| TP1 no-AllGather mode | Mean wave | Throughput | Request-time host PSS | `Private_Dirty` | Engine peak HBM | CPU `aten::copy_` | GPU compute |
+|---|---:|---:|---:|---:|---:|---:|---:|
+| Checkpoint mmap + bounded staging | 33.87 s | 0.0295 req/s | 133.65 GiB | 71.16 GiB | 13,226 MB | 21.91 s | 419.28 ms |
+| Final runtime cache + registered direct H2D | 4.88 s | 0.2055 req/s | 129.41 GiB | 66.93 GiB | 13,226 MB | 1.64 s | 419.05 ms |
+
+Registration covered 61.73 GiB in 13 page-aligned ranges. Mean wave latency
+fell by 85.6% and throughput increased 6.94x. GPU compute and H2D payload were
+unchanged, identifying recurrent host transform/packing as the removed work.
+The direct path issued 2,425 H2D events versus 1,945 for packed staging, but
+that launch increase was much smaller than the eliminated CPU cost.
+
+Two independent TP2 engines also registered the same two TP-coordinate cache
+entries across four workers. Compared with the zero-budget cache, mean wave
+latency improved from 9.21/12.55 seconds to 3.80/3.86 seconds for engines A/B;
+the private-loader reference was 3.61/3.72 seconds. Registered request-time
+host PSS was 224.26 GiB versus 232.23 GiB staged and 366.20 GiB private.
+Aggregate CPU `aten::copy_` time was 3.83 seconds registered, 32.49 seconds
+staged, and 3.75 seconds private; aggregate GPU compute and the 264.24 GiB H2D
+payload were effectively unchanged. Ordinary TP communication remained, but
+the extra wait caused by staging-driven rank skew largely disappeared.
+
+All workers explicitly unregistered before releasing their mappings, the
+system's acquired-minus-released pin counter returned to its pre-run value,
+and every GPU allocation was released. Cold runtime-cache publication remained
+expensive because it still performs ordinary loading and multiple full-content
+passes; that startup problem is outside the registered transfer scope.
+
+## Recommendations
+
+- Use **DP + DLO AllGather** for the supported throughput and host-memory
+  scaling path.
+- Use **SP + DLO AllGather** for long-sequence workloads when DP concurrency is
+  not the goal.
+- Use **no-AllGather** when independent replica execution is required. TP1
+  direct-mmap deployments can share checkpoint pages per node. Enable the
+  runtime cache only when independent replicas have repeated final layouts and
+  benchmark the target CPU/memory topology, especially for matching
+  coordinates across multiple TP engines. On CUDA, provide an explicit
+  registration budget large enough for the complete final layout when
+  recurrent staging cost is unacceptable; leave it at zero when page-locking
+  that layout is not operationally acceptable.
+- Prefer **HSDP alone** for production HSDP deployments until the combined
+  HSDP + DLO no-AllGather path has broader end-to-end coverage.
@@ -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).
@@ -76,6 +76,12 @@ equivalent. `auto-round` and `auto_round` use the Omni INC/AutoRound adapter.
 ModelOpt checkpoint metadata is detected from its method, producer metadata, or
 `quant_algo`, then mapped to the corresponding vLLM ModelOpt configuration.

+Unless the input is a per-component dictionary, the resolved configuration is
+passed unchanged to every quantization-aware component constructed by the
+pipeline. It affects only layers supported by that component's quantization
+method; it does not rewrite arbitrary `torch.nn` modules. A per-component
+dictionary narrows this global scope through runtime layer-prefix routing.
+
 The factory uses Omni-specific overrides for methods whose diffusion behavior
 or platform dispatch is not provided by the vLLM registry. The current override
 set includes INT8, BitsAndBytes, MXFP8, MXFP4, MXFP4 dual-scale, and INC/
@@ -154,9 +160,11 @@ quantizable vLLM layers. Each method configuration is responsible for:
 Model code must keep precision-sensitive or unsupported modules unquantized.
 The shared `safe_quant_config()` helper is used by model integrations for
 normalization and modulation layers where a generic FP8 configuration would be
-unsafe. Text encoders, audio/vision encoders, VAEs, schedulers, and tokenizers
-are not implicitly quantized just because the main transformer is quantized;
-they require explicit component routing and method support.
+unsafe. Encoders exposed through the quantization factory receive a global
+configuration just like the main transformer, but only their supported vLLM
+quantizable layers are affected. Ordinary `torch.nn` encoder layers, VAEs,
+schedulers, and tokenizers are not rewritten automatically. A component map can
+select a narrower scope, but it cannot add quantization support to a layer.

 Weight quantization and quantized KV cache are separate paths. The former is
 resolved through `quantization_config`; runtime attention/KV-cache quantization
@@ -186,7 +194,7 @@ 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
@@ -0,0 +1,213 @@
+# Realtime AR-Diffusion sessions
+
+> **Status:** experimental. This document describes the internal execution
+> contract implemented under `vllm_omni.experimental.ar_diffusion`. It does not
+> define a public HTTP or WebSocket schema.
+
+## Scope
+
+Realtime AR-Diffusion serves world models that generate one autoregressive
+latent block per request while retaining model state across requests. The
+generic runtime owns identity, event ordering, session lifecycle, routing, and
+KV allocation. A model adapter owns model-specific controls and conditioning.
+
+LingBot World v2 is the first single-KV-branch integration. DreamZero uses the
+same runner capability with two logical CFG branches.
+
+## Identities
+
+Four identifiers have distinct lifetimes:
+
+| Identifier | Lifetime | Owner | Contract |
+| --- | --- | --- | --- |
+| `session_id` | One persistent world | Session manager | Selects runner KV and model-owned state. |
+| `event_id` | One control/prompt update | Transport adapter | Unique and monotonically increasing within a session, including across reset. |
+| `chunk_index` | One committed AR block position | Session | Contiguous from zero; reset starts again at zero. |
+| `request_id` | One chunk execution | Session | Correlates the tick with its returned metadata. |
+
+`AsyncOmni` keeps its normal UUID-suffixed engine routing ID. The tick
+`request_id` remains inside the immutable AR-Diffusion snapshot and the
+standard output metadata, while caller-visible `OmniRequestOutput.request_id`
+continues to use AsyncOmni's existing external-ID contract. Protocol identity
+therefore does not require a private entrypoint override.
+
+## Request lifecycle
+
+1. `ARDiffusionSession.accept_event()` validates and queues an immutable
+   `ARDiffusionSessionEvent`. Prompt and opaque control tracks may be updated in
+   the same event.
+2. `ARDiffusionSession.next_chunk()` serializes chunk execution and calls
+   `_build_tick_locked()`. The resulting `ARDiffusionTickRequest` is a snapshot
+   of the current prompt, reduced controls, ordered event IDs, and chunk index.
+3. `ARDiffusionOmniTickConsumer.execute_tick()` copies the stage sampling
+   parameters, stores the typed tick in
+   `sampling_params.extra_args["ar_diffusion_tick"]`, constructs the normal Omni
+   prompt through its `prompt_provider`, and invokes `AsyncOmni.generate()`.
+4. `AsyncOmni` allocates its normal unique engine routing ID and submits one
+   `OmniDiffusionRequest` through the orchestrator.
+5. `ARDiffusionModelRunner._request_session()` parses the typed tick.
+   `_get_or_create_session()` obtains runner-owned paged KV, and
+   `bind_ar_diffusion_state()` exposes it to the pipeline only for the duration
+   of `execute_model()`.
+6. The model pipeline validates the tick and generates exactly one latent
+   block. LingBot interprets camera controls in
+   `LingBotCameraControlReducer` and
+   `LingBotWorldCausalDMDPipeline.forward()`; the generic runtime does not parse
+   camera or action schemas.
+7. The pipeline returns one `DiffusionOutput` with the generated payload and an
+   `ARDiffusionChunkMetadata` mapping under `metadata["ar_diffusion"]`.
+   `AsyncOmni` exposes it as
+   `OmniRequestOutput.multimodal_output["metadata"]["ar_diffusion"]`.
+8. The consumer parses the standard envelope. The session commits its reducer
+   state, prompt, controls, event IDs, and next chunk index only when returned
+   metadata exactly equals the submitted tick snapshot.
+
+The metadata envelope is:
+
+```json
+{
+  "metadata": {
+    "ar_diffusion": {
+      "session_id": "world-7",
+      "request_id": "ar-world-7-3-...",
+      "chunk_index": 3,
+      "applied_event_ids": [10, 11]
+    }
+  }
+}
+```
+
+## Transaction and failure semantics
+
+Accepted events are not acknowledged as applied until the model output and
+metadata are validated. Reducer `prepare()` is speculative; reducer `commit()`
+and the session snapshot form one logical commit. A model, metadata, or reducer
+failure leaves the chunk index and queued events uncommitted and transitions the
+session to `FAILED`.
+
+After a tick failure, worker state is closed because a forward may have
+partially modified or already released KV. The same chunk is not retried in
+place. A caller must explicitly reset the session, starting again at chunk
+zero, or close it.
+
+Explicit close follows this state machine:
+
+```text
+ACTIVE or FAILED
+       |
+       v
+    CLOSING ---- worker cleanup succeeds ----> CLOSED
+       |
+       +------- worker cleanup fails --------> CLEANUP_FAILED
+                                                   |
+                                                   +-- close retry --> CLOSING
+```
+
+`CLEANUP_FAILED` is a local tombstone. The session manager retains it, rejects
+creation of the same `session_id`, and rejects reset. Only an idempotent close
+retry that succeeds removes the session. Transport disconnect uses the same
+path.
+
+## KV ownership and capacity
+
+`ARDiffusionKVCacheSpec` is the pipeline-owned immutable geometry contract. It
+declares:
+
+- TP-local self-attention geometry, AR block size, sink, and recent window;
+- logical KV branches and their worker-local slots;
+- fixed-length cross-attention caches;
+- scratch tokens required by an uncommitted forward;
+- persistent model-owned CUDA bytes per resident session; and
+- the requested resident-session capacity.
+
+The runner computes:
+
+```text
+required(capacity) =
+    scratch
+  + managed_self_attention_pages(capacity)
+  + capacity * cross_attention_bytes_per_session
+  + capacity * model_owned_state_bytes_per_session
+```
+
+`gpu_memory_fraction` is a soft expansion budget:
+
+```text
+configured_budget = available_device_bytes * gpu_memory_fraction
+effective_budget = max(configured_budget, required(1))
+```
+
+If `required(1)` exceeds actual available device memory, initialization fails.
+Otherwise the runner always admits one viable session and uses the configured
+fraction to select additional resident sessions up to the model-declared cap.
+All model-owned reservations are deducted before the paged self-attention pool
+is allocated.
+
+LingBot reports its persistent full-horizon image-condition tensor. DreamZero
+reports the measured 603 MiB per-session Wan VAE causal-convolution state.
+
+## Concurrency and routing limits
+
+The current runner executes one request at a time:
+
+- `max_num_seqs` must be one;
+- request batching and diffusion step execution are rejected;
+- one AR block is generated per engine request; and
+- an AR-Diffusion stage must have exactly one replica.
+
+The single-replica restriction is intentional. Multi-replica support requires
+session-affine routing so every tick, reset, and close reaches the worker that
+owns the session. Random or round-robin routing is not correct.
+
+Multiple resident sessions may share one runner. Capacity selection accounts
+for all persistent pools, and the runner performs LRU eviction through the same
+model lifecycle hook used by explicit close.
+
+## Model boundary
+
+The generic runtime transports controls as opaque
+`ARDiffusionControlInput(track, schema, data)` values. Model adapters validate
+and reduce those values. This keeps target/velocity, key state, pose,
+intrinsics, and other model-specific conventions out of the session protocol.
+
+For LingBot World v2:
+
+- `lingbot_world/actions.py` parses key-state and trajectory controls;
+- `lingbot_world/camera.py` loads or constructs camera trajectories and Plücker embeddings;
+- `lingbot_world/transformer.py` implements checkpoint-compatible causal attention; and
+- `lingbot_world/pipeline.py` constructs conditioning, owns small non-KV session state, and
+  produces one block plus the standard metadata envelope.
+
+## Non-goals
+
+This contract does not currently provide:
+
+- a public HTTP or WebSocket request schema (the structured-interaction
+  frontend is tracked separately in
+  [#5527](https://github.com/vllm-project/vllm-omni/pull/5527));
+- camera/action semantics shared by every world model;
+- session migration or replication across workers;
+- retry of an ambiguous partially executed chunk;
+- cross-stage KV transfer; or
+- stateful streaming VAE decode.
+
+Those features can be layered on top without adding LingBot-specific identity
+or lifecycle fields to the generic runtime.
+
+## Required regression coverage
+
+CPU contract tests cover:
+
+- separation of tick correlation IDs from internal engine routing IDs;
+- event ordering, metadata equality, and reducer fail-closed commit;
+- close/disconnect failure tombstones and cleanup retry;
+- capacity one under the default fraction for shipped LingBot geometry;
+- capacity expansion under a tuned fraction;
+- DreamZero model-owned state accounting;
+- runner session reuse, reset, close, LRU eviction, and forward cleanup; and
+- LingBot model registration, imports, camera/action reduction, and one-block
+  output metadata.
+
+GPU validation must additionally exercise real model loading, action input,
+prompt switching, rolling-window replacement, reset, close, finite output
+latents, and exact metadata/request identity.
@@ -2,8 +2,10 @@

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

 ## Motivation

@@ -22,6 +22,7 @@ implementation contract; it is not, by itself, a general support claim.
 - [Async Diffusion Output](feature/async_diffusion_output.md)
 - [Async Omni Output Materialization](feature/omni_async_output_materialization.md)
 - [Automatic Prefix Caching in Omni Models](feature/prefix_caching.md)
+- [Realtime AR-Diffusion Sessions](feature/realtime_ar_diffusion.md)

 ### Communication

@@ -53,15 +54,22 @@ 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
-is the backend-related optimization with a standalone feature-design contract:
+guides list selectable backends, platform defaults, installation, and tuning.
+The design contracts separate selection mechanics from backend algorithms:

+- [Attention Backend Selection](feature/attention_backend_selection.md)
 - [Skip-Softmax](feature/skip_softmax.md)

+#### 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

@@ -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 |
@@ -101,7 +101,7 @@ outputs = omni.generate(
 )

 # Access custom trajectory data
-output = outputs[0].request_output
+output = outputs[0]
 print(f"Trajectory timesteps shape: {output.metrics['trajectory_timesteps'].shape}")
 print(f"Trajectory latents shape: {output.latents.shape}")
 ```
@@ -47,7 +47,7 @@ if __name__ == "__main__":
    omni = Omni(model="Tongyi-MAI/Z-Image-Turbo")
    prompt = "a cup of coffee on the table"
    outputs = omni.generate(prompt)
-    images = outputs[0].request_output.images
+    images = outputs[0].images
    images[0].save("coffee.png")
 ```

@@ -75,8 +75,7 @@ if __name__ == "__main__":
    ]
    omni_outputs = omni.generate(prompts)
    for i_prompt, prompt_output in enumerate(omni_outputs):
-        this_request_output = prompt_output.request_output
-        this_images = this_request_output.images
+        this_images = prompt_output.images
        for i_image, image in enumerate(this_images):
            image.save(f"p{i_prompt}-img{i_image}.jpg")
            print("saved to", f"p{i_prompt}-img{i_image}.jpg")
@@ -37,12 +37,19 @@ th {
 | `Krea2Pipeline` | Krea 2 (Raw + Turbo) | `krea/Krea-2-Raw`, `krea/Krea-2-Turbo` | ✅︎ | | | | — |
 | `WanPipeline` | Wan2.1-T2V, Wan2.2-T2V, Wan2.2-TI2V | `Wan-AI/Wan2.1-T2V-1.3B-Diffusers`, `Wan-AI/Wan2.1-T2V-14B-Diffusers`, `Wan-AI/Wan2.2-T2V-A14B-Diffusers`, `Wan-AI/Wan2.2-TI2V-5B-Diffusers` | ✅︎ | ✅︎ | ✅︎ | ✅︎ | — |
 | `WanImageToVideoPipeline` | Wan2.2-I2V | `Wan-AI/Wan2.2-I2V-A14B-Diffusers` | ✅︎ | ✅︎ | ✅︎ | ✅︎ | — |
+| `LingBotWorldCausalDMDPipeline` | LingBot-World 2.0 v1 (experimental) | `robbyant/lingbot-world-v2-14b-causal-fast-diffusers` | Experimental | | | | — |
 | `Cosmos3OmniDiffusersPipeline` | Cosmos3 T2I, T2V, I2V, V2V, T2V with sound, action policy | `nvidia/Cosmos3-Nano`, `nvidia/Cosmos3-Super` | ✅︎ | ✅︎ | ✅︎ | ✅︎ | — |
 | `Wan22S2VPipeline` | Wan2.2-S2V | `Wan-AI/Wan2.2-S2V-14B` | ✅︎ | ✅︎ | ✅︎ | ✅︎ | — |
 | `Wan22VACEPipeline` | Wan2.1-VACE | `Wan-AI/Wan2.1-VACE-1.3B-diffusers`, `Wan-AI/Wan2.1-VACE-14B-diffusers` | ✅︎ | ✅︎ | ✅︎ | ✅︎ | — |
 | `Wan22VACEPipeline` | Wan2.2-VACE | `Pyros13/Wan2.2-VACE-Fun-A14B-Diffusers` | ✅︎ |   |   |   | — |
 | `LTX2Pipeline` | LTX-2 / LTX-2.3 one-stage T2V and I2V | `Lightricks/LTX-2`, `diffusers/LTX-2.3-Diffusers` | ✅︎ | ✅︎ | | | — |
-| `LTX2DistilledPipeline` | LTX-2 distilled two-stage T2V and I2V | `rootonchair/LTX-2-19b-distilled` | ✅︎ | ✅︎ | | | — |
+| `LTX2TwoStagePipeline` | LTX-2 / LTX-2.3 ordinary two-stage T2V and I2V | `Lightricks/LTX-2`, `diffusers/LTX-2.3-Diffusers` + matching Lightricks LoRA and upsampler | ✅︎ | ✅︎ | | | — |
+| `LTX2DistilledOneStagePipeline` | LTX-2 / LTX-2.3 merged-distilled one-stage T2V and I2V | `rootonchair/LTX-2-19b-distilled`, `diffusers/LTX-2.3-Distilled-Diffusers` | ✅︎ | ✅︎ | | | — |
+| `LTX2DistilledTwoStagePipeline` | LTX-2 / LTX-2.3 merged-distilled two-stage T2V and I2V | `rootonchair/LTX-2-19b-distilled`, `diffusers/LTX-2.3-Distilled-Diffusers` + matching Lightricks upsampler | ✅︎ | ✅︎ | | | — |
+| `LTX2Pipeline` | LTX-2.5 Full/SFT one-stage T2V and I2V | `Lightricks/LTX-2.5-Diffusers` | ✅︎ | | | | [Repository](https://github.com/vllm-project/vllm-omni/blob/main/recipes/LTX/LTX-2.5.md) |
+| `LTX2TwoStagePipeline` | LTX-2.5 Full/SFT two-stage T2V and I2V | `Lightricks/LTX-2.5-Diffusers` + `Lightricks/LTX-2.5` LoRA | ✅︎ | | | | [Repository](https://github.com/vllm-project/vllm-omni/blob/main/recipes/LTX/LTX-2.5.md) |
+| `LTX2DistilledOneStagePipeline` | LTX-2.5 merged-distilled one-stage T2V and I2V | `Lightricks/LTX-2.5-Diffusers` | ✅︎ | | | | [Repository](https://github.com/vllm-project/vllm-omni/blob/main/recipes/LTX/LTX-2.5.md) |
+| `LTX2DistilledTwoStagePipeline` | LTX-2.5 merged-distilled two-stage T2V and I2V | `Lightricks/LTX-2.5-Diffusers` | ✅︎ | | | | [Repository](https://github.com/vllm-project/vllm-omni/blob/main/recipes/LTX/LTX-2.5.md) |
 | `LingBotVideoPipeline` | LingBot-Video dense and MoE T2I, T2V, TI2V | `robbyant/lingbot-video-dense-1.3b`, `robbyant/lingbot-video-moe-30b-a3b` | ✅︎ | | | | — |
 | `MiniMaxH3Pipeline` | MiniMax H3 T2VA, first/last-keyframe FL2VA, and mixed-reference Ref2VA | `MiniMaxAI/MiniMax-H3` | ✅︎ | ✅︎ | | | [Published](https://recipes.vllm.ai/MiniMaxAI/MiniMax-H3) |
 | `DreamZeroPipeline` | DreamZero-DROID | `GEAR-Dreams/DreamZero-DROID` | ✅︎ | ✅︎ | ✅︎ | ✅︎ | — |
@@ -58,6 +65,7 @@ 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` | ✅︎ | ✅︎ | ✅︎ | ✅︎ | — |
@@ -96,6 +104,7 @@ th {
 |`DyninOmniForConditionalGeneration` | Dynin-Omni | `snu-aidas/Dynin-Omni` | ✅︎ | | | | — |
 | `MiniCPMO45OmniForConditionalGeneration` | MiniCPM-o 4.5 | `openbmb/MiniCPM-o-4_5` | ✅︎ | | ✅︎ | | — |
 | `ErnieImagePipeline` | ERNIE-Image | `baidu/ERNIE-Image`, `baidu/ERNIE-Image-Turbo` | ✅︎ | ✅︎ | ✅︎ | ✅︎ | — |
+| `GepardTalkerForConditionalGeneration` | Gepard-1.0 | `nineninesix/gepard-1.0` | ✅︎ | | | | — |
 |`HiDreamImagePipeline` | HiDream-I1-Full | `HiDream-ai/HiDream-I1-Full` | ✅︎ | ✅︎ | | | — |

 The `Recipe` column links the corresponding deployment recipe. `Published` is
@@ -109,6 +118,10 @@ support metadata until their recipe is audited.

 ✅︎ indicates the model is supported on that backend. Empty cells mean not listed as supported on that backend.

+See the [LingBot-World 2.0 recipe](https://github.com/vllm-project/vllm-omni/blob/main/recipes/Robbyant/LingBot-World-2.0.md)
+for offline and experimental realtime usage, supported controls, validation,
+and current limitations.
+
 <sup>H3</sup> MiniMax H3 AMD GPU support is validated on gfx942 (MI300X) and
 gfx950 (MI350) in BF16, using the AITER `FLASH_ATTN` diffusion attention backend.
 Other AMD SKUs (e.g. MI325X) are not listed until their
@@ -10,12 +10,12 @@ A: Refer to [GPU memory calculation and configuration](../configuration/gpu_memo

 > Q: I encounter some bugs or CI problems, which is urgent. How can I solve it?

-A: At first, you can check current [issues](https://github.com/vllm-project/vllm-omni/issues) to find possible solutions. If none of these satisfy your demand and it is urgent, please find these [volunteers](https://docs.vllm.ai/projects/vllm-omni/en/latest/community/volunteers/) for help.
+A: Search the current [issues](https://github.com/vllm-project/vllm-omni/issues) first. If no existing issue matches a bug or CI failure, open a [bug report](https://github.com/vllm-project/vllm-omni/issues/new?template=400-bug-report.yml) with a minimal reproducer, environment details, and relevant logs. For technical or usage questions, use the `sig-omni` channel in [vLLM Slack](https://slack.vllm.ai/) or the [vLLM Forum](https://discuss.vllm.ai/).

 > Q: Does vLLM-Omni support AWQ or any other quantization?

-A: We plan to introduce GGUF FP8 prequantized models and online FP8 quantization in version 0.16.0. Support for other quantization types will follow in future releases. For details, please see our [Q1 quantization roadmap](https://github.com/vllm-project/vllm-omni/issues/1057).
+A: AWQ is available through the inherited vLLM quantization registry, and vLLM-Omni supports several additional methods. Actual support and validation vary by model, component, hardware, and checkpoint. See the [quantization overview](../user_guide/quantization/overview.md) for validated combinations and method-specific guides.

 > Q: Does vLLM-Omni support multimodal streaming input and output?

-A: Not yet. We already put it on the [Roadmap](https://github.com/vllm-project/vllm-omni/issues/165). Please stay tuned!
+A: Yes, for specific models and serving paths. Current capabilities include [streaming video input](../serving/video_stream_api.md), [streaming diffusion output](../user_guide/diffusion/execution_modes.md#streaming-output), and experimental full-duplex real-time audio described in the [project overview](../README.md). Check the model and API documentation because input and output support is model-specific.
@@ -1,494 +1,192 @@
 # Diffusion Attention Backends

-This document describes the diffusion attention backends available in vLLM-Omni, how to select them globally and per-role, the per-platform defaults, and how to use SageAttention.
+Use this page to select and configure a diffusion attention backend. Backend-
+specific installation, tuning, and compatibility details live in separate
+guides so that the selection contract stays easy to scan.

-## Overview
+Diffusion attention backend selection applies to DiT attention in image and
+video generation models. It does **not** change autoregressive LLM attention,
+which uses vLLM's attention selector.

-Diffusion attention backend selection is resolved in `vllm_omni.diffusion.attention.selector`. It looks up the backend from a structured `AttentionConfig` carried on `OmniDiffusionConfig` and falls back to the platform default when nothing is configured.
+For the internal selector, registry, and platform contract, see
+[Attention Backend Selection](../../design/feature/attention_backend_selection.md).

-This backend is used by diffusion attention layers such as the DiT attention in video and image generation models. It does **not** affect autoregressive (LLM) attention paths — those go through vLLM's own attention backend selector.
+## Choose a guide

-The full set of backends and their platform defaults is in the **Backend Options** and **Platform Defaults** sections below. If no attention backend is configured, vLLM-Omni asks the current platform to choose the default.
+| Need | Guide |
+| --- | --- |
+| Select a conservative or platform-native dense kernel | [Dense Backends](attention_backends/dense_backends.md) |
+| Use FlashInfer trtllm-gen FMHA, Skip-Softmax, or TRTLLM SAGE quantization | [TRTLLM Attention](attention_backends/trtllm.md) |
+| Install and use SageAttention 2.2 or SageAttention3 | [SageAttention](attention_backends/sage.md) |
+| Match training or rollout kernels loaded from Hugging Face | [Hugging Face Hub Backends](attention_backends/huggingface_hub.md) |
+| Use block-sparse video attention on Ascend NPU | [RainFusion](attention_backends/rainfusion.md) |

-## Backend Options
-
-| Value | Notes |
-|---|---|
-| `TRTLLM_ATTN` | FlashInfer's trtllm-gen FMHA (TensorRT-LLM's generated kernels, vendored by FlashInfer). Dense BF16, GQA native, `head_dim=128`. Datacenter Blackwell only (sm_100 / sm_103). Packed paths can provide `cu_seqlens` directly. Supports optional **Skip-Softmax** sparse attention — see below. Requires `flashinfer`. |
-| `FLASH_ATTN` | Wraps FlashAttention 4 on Blackwell when `flash-attn-4` is installed, then falls back to FlashAttention 3/2. Default on Hopper / Ada / Ampere when a compatible FlashAttention package is installed. |
-| `CUDNN_ATTN` | Pins `sdpa_kernel([CUDNN_ATTENTION])`. Default on Blackwell (sm_10x / sm_12x) with cuDNN ≥ 9.5. Wins on mask-heavy DiTs (HunyuanVideo-1.5: 2× e2e vs SDPA). |
-| `FLASHINFER_ATTN` | Uses FlashInfer's batch-prefill wrapper and supports mixed Q/K and V input dtypes through backend-specific configuration. Used as Blackwell fallback when cuDNN is unavailable. Requires `flashinfer` >= 0.6.16rc1 for mixed-dtype configurations. |
-| `TORCH_SDPA` | PyTorch `scaled_dot_product_attention` with the default backend dispatcher. Most conservative; always available. |
-| `SAGE_ATTN` | SageAttention 2.2 — INT8-quantized attention with FP16 accumulation. Lossy but typically visually indistinguishable on diffusion outputs. Requires `sageattention`. |
-| `SAGE_ATTN_3` | Requires `sageattn3` from `SageAttention/sageattention3_blackwell`. CUDA only, intended for Blackwell GPUs, with GQA/MQA requests falling back to PyTorch SDPA. |
-| `FLASH_ATTN_HUB` | FlashAttention 2 from HuggingFace `kernels` library. Useful for train/rollout alignment. |
-| `FLASH_ATTN_3_HUB` | FlashAttention 3 from HuggingFace `kernels` library. CUDA Hopper (sm_90+) only; falls back to `FLASH_ATTN_HUB` on older GPUs. |
-| `RAINFUSION_ATTN` | MindIE-SD **RainFusion** block-sparse video attention — see [below](#rainfusion_attn-backend-and-block-sparse-video-attention). Ascend NPU only; requires `mindiesd`. Delegates to `FLASH_ATTN` for anything that is not a packed video sequence. |
+## Backend options

+| Value | Family | Primary use | Detail |
+| --- | --- | --- | --- |
+| `TORCH_SDPA` | Dense | Conservative reference; always available | [Dense Backends](attention_backends/dense_backends.md#torch_sdpa) |
+| `FLASH_ATTN` | Dense | FlashAttention 4/3/2 depending on the installed package and GPU | [Dense Backends](attention_backends/dense_backends.md#flash_attn) |
+| `CUDNN_ATTN` | Dense | Mask-heavy DiTs on Blackwell with cuDNN 9.5 or newer | [Dense Backends](attention_backends/dense_backends.md#cudnn_attn) |
+| `FLASHINFER_ATTN` | Dense or quantized | FlashInfer batch prefill; optional mixed Q/K and V dtypes | [Dense Backends](attention_backends/dense_backends.md#flashinfer_attn) |
+| `TRTLLM_ATTN` | Dense, sparse, or quantized | Datacenter Blackwell with `head_dim=128` and compatible packed paths | [TRTLLM Attention](attention_backends/trtllm.md) |
+| `SAGE_ATTN` | Quantized | SageAttention 2.2 INT8 attention | [SageAttention](attention_backends/sage.md#sage_attn) |
+| `SAGE_ATTN_3` | Quantized | SageAttention3 on Blackwell | [SageAttention](attention_backends/sage.md#sage_attn_3) |
+| `FLASH_ATTN_HUB` | Hub kernel | FlashAttention 2 from Hugging Face `kernels` | [Hugging Face Hub Backends](attention_backends/huggingface_hub.md) |
+| `FLASH_ATTN_3_HUB` | Hub kernel | FlashAttention 3 from Hugging Face `kernels` on Hopper or newer | [Hugging Face Hub Backends](attention_backends/huggingface_hub.md) |
+| `RAINFUSION_ATTN` | Block sparse | MindIE-SD RainFusion video attention on Ascend NPU | [RainFusion](attention_backends/rainfusion.md) |

 ## Configuration

-Diffusion attention backends can be configured three ways, in priority order:
+Backend selection is resolved in this order:

-1. **`--diffusion-attention-config`** — structured per-role config (highest priority).
-2. **`--diffusion-attention-backend` / `DIFFUSION_ATTENTION_BACKEND` env var** — global shorthand that sets the default backend.
-3. **Platform default** — used when nothing is configured.
+1. `--diffusion-attention-config` per-role configuration.
+2. `--diffusion-attention-backend` or `DIFFUSION_ATTENTION_BACKEND` as a
+   global default.
+3. The current platform's default.

-`--diffusion-attention-backend` is shorthand for `--diffusion-attention-config.default.backend`. It may be combined with `--diffusion-attention-config.per_role.*` overrides, but is mutually exclusive with `--diffusion-attention-config.default.backend`.
+`--diffusion-attention-backend` is shorthand for
+`--diffusion-attention-config.default.backend`. Do not pass it together with an
+explicit `default.backend` in the structured configuration.

 ### Global default

-Set the default backend for every diffusion attention layer:
-
 ```bash
-# CLI flag
-vllm-omni serve <model> --diffusion-attention-backend SAGE_ATTN
+vllm-omni serve <model> --diffusion-attention-backend FLASH_ATTN

-# Environment variable (also recognized for backwards compatibility)
-export DIFFUSION_ATTENTION_BACKEND=SAGE_ATTN
+# Backwards-compatible environment variable
+export DIFFUSION_ATTENTION_BACKEND=FLASH_ATTN
 ```

 ### Per-role configuration

-Roles are free-form strings declared by each diffusion model. The two common categories are `"self"` and `"cross"`; model-specific roles (e.g. `"ltx2.audio_to_video"`) may also be declared. A role string is matched in this order:
+Roles are declared by each diffusion model. Common categories are `self` and
+`cross`; a model may also use a more specific role such as
+`ltx2.audio_to_video`. Resolution order is:

-1. Exact `per_role[role]` match
-2. `per_role[role_category]` fallback (e.g. `"ltx2.audio_to_video"` → `"cross"`)
-3. `default`
-4. Platform default
-
-Use vLLM-style dotted flags or one JSON blob:
+1. Exact `per_role[role]` match.
+2. Category `per_role[role_category]` match.
+3. `default`.
+4. Platform default.

 ```bash
 # Dotted flags
 vllm-omni serve <model> \
-    --diffusion-attention-config.default.backend FLASH_ATTN \
-    --diffusion-attention-config.per_role.cross.backend TORCH_SDPA
+  --diffusion-attention-config.default.backend FLASH_ATTN \
+  --diffusion-attention-config.per_role.cross.backend TORCH_SDPA

-# JSON
+# Equivalent JSON
 vllm-omni serve <model> \
-    --diffusion-attention-config '{"default":{"backend":"FLASH_ATTN"},"per_role":{"cross":{"backend":"TORCH_SDPA"}}}'
+  --diffusion-attention-config \
+  '{"default":{"backend":"FLASH_ATTN"},"per_role":{"cross":{"backend":"TORCH_SDPA"}}}'
 ```

-A backend that needs configuration exposes it as a typed field on the spec:
-`TRTLLM_ATTN`'s Skip-Softmax (see [below](#trtllm_attn-backend-and-skip-softmax)) and
-`RAINFUSION_ATTN`'s RainFusion block (see [below](#rainfusion_attn-backend-and-block-sparse-video-attention)):
-
-```bash
--diffusion-attention-config.default.backend TRTLLM_ATTN \
--diffusion-attention-config.default.skip_softmax.target_sparsity 0.5
-```
-
-### Programmatic API
-
-When constructing `OmniDiffusionConfig` directly:
+### Python API

 ```python
-from vllm_omni.diffusion.data import AttentionConfig, AttentionSpec, AttnQuantSpec, OmniDiffusionConfig
+from vllm_omni.diffusion.data import (
+    AttentionConfig,
+    AttentionSpec,
+    OmniDiffusionConfig,
+)

 config = OmniDiffusionConfig(
    diffusion_attention_config=AttentionConfig(
        default=AttentionSpec(backend="FLASH_ATTN"),
-        per_role={
-            "cross": AttentionSpec(backend="TORCH_SDPA"),
-        },
+        per_role={"cross": AttentionSpec(backend="TORCH_SDPA")},
    ),
    ...,
 )
 ```

-A plain dict is also accepted and normalized to `AttentionConfig`.
+Backend-specific typed blocks are documented with their consumers:

-#### FlashInfer QK16/V8 quantized attention
+- `quant`: [FlashInfer](attention_backends/dense_backends.md#flashinfer-quantized-attention)
+  and [TRTLLM SAGE](attention_backends/trtllm.md#sage-quantization).
+- `skip_softmax`: [TRTLLM Skip-Softmax](attention_backends/trtllm.md#skip-softmax).
+- `block_sparse`: [RainFusion](attention_backends/rainfusion.md#configuration).

-`FLASHINFER_ATTN` accepts quantization options through `AttentionSpec.quant`. For QK16/V8:
-
-```python
-config = OmniDiffusionConfig(
-    diffusion_attention_config=AttentionConfig(
-        default=AttentionSpec(
-            backend="FLASHINFER_ATTN",
-            quant=AttnQuantSpec(
-                dtype_qk="bfloat16",
-                dtype_vo="fp8_e4m3",
-            ),
-        ),
-    ),
-    ...,
-)
-```
-
-`dtype_qk` configures the Q and K input dtypes. `dtype_vo` configures the V input dtype.
-
-## Platform Defaults
+## Platform defaults

 ### Blackwell (sm_100 / sm_103 / sm_120 / sm_121)

-Auto-route preference, in order:
+The CUDA auto-route preference is:

-1. `TRTLLM_ATTN` — on **datacenter** Blackwell (sm_100 / sm_103) when `flashinfer` is installed, the model's `head_dim` is 128, **and the model declares a compatible packed/mask-free path**
-2. `CUDNN_ATTN` — when cuDNN ≥ 9.5 is available (ships in PyTorch 2.5+ wheels)
-3. `FLASHINFER_ATTN` — when `flashinfer` is installed but cuDNN < 9.5
-4. `FLASH_ATTN` — when `flash-attn` is installed with the Blackwell CUTE kernel
-5. `TORCH_SDPA` — last resort
+1. `TRTLLM_ATTN` on datacenter Blackwell (sm_100/sm_103) when FlashInfer is
+   available, `head_dim=128`, and the model declares a compatible packed or
+   mask-free path.
+2. `CUDNN_ATTN` when cuDNN 9.5 or newer is available.
+3. `FLASHINFER_ATTN` when FlashInfer is available but cuDNN is too old.
+4. `FLASH_ATTN` when a compatible package is installed.
+5. `TORCH_SDPA`.

-`TRTLLM_ATTN` is skipped on workstation Blackwell (sm_120 / sm_121) and for any `head_dim != 128`, so those GPUs keep the `CUDNN_ATTN` route described below.
+`TRTLLM_ATTN` is not auto-selected on workstation Blackwell
+(sm_120/sm_121), for other head dimensions, or for paths that require an
+unsupported mask.

-`TRTLLM_ATTN` outranks `CUDNN_ATTN` on datacenter Blackwell for compatible packed/mask-free pipelines. Workstation Blackwell (sm_120 / sm_121) and pipelines that require attention masks retain their normal fallback.
+### Hopper, Ada, and Ampere

-The startup log line `Defaulting to diffusion attention backend CUDNN_ATTN (Blackwell sm_120, cuDNN 91002)` confirms the route.
+The CUDA auto-route uses `FLASH_ATTN` when available and otherwise falls back
+to `TORCH_SDPA`. `CUDNN_ATTN` and `FLASHINFER_ATTN` remain explicit options.

-**Why CUDNN_ATTN by default**: on mask-heavy diffusion models (HunyuanVideo-1.5, Qwen-Image), cuDNN's pinned FMHA kernel sidesteps a PyTorch SDPA dispatch quirk where the unpinned dispatcher picks `EFFICIENT_ATTENTION` (~25 ms) for masked calls instead of cuDNN (~11 ms). The pin gives 2× e2e on HV-1.5 with no regression on lighter models.
+Other platforms validate an explicit backend and choose their own default
+through the platform implementation. Check the startup log to confirm the
+resolved backend.

-### Hopper (sm_90) / Ada (sm_89) / Ampere (sm_80–sm_86)
+## Choosing a backend manually

-Auto-route preference:
+Override the platform default when you need:

-1. `FLASH_ATTN` — when `flash-attn` is installed
-2. `TORCH_SDPA` — fallback
+- a correctness reference (`TORCH_SDPA`);
+- a backend-specific workaround;
+- training/rollout kernel alignment (Hub backends); or
+- an explicitly validated sparse or quantized speedup.

-`CUDNN_ATTN` and `FLASHINFER_ATTN` are still selectable via env var on these GPUs but are not in the auto-route — FlashAttention 2 is the well-tuned path on pre-Blackwell hardware.
+The startup log prints the resolved backend and whether it came from explicit
+configuration or platform defaulting. If no resolution message appears, check
+earlier logs for diffusion-stage initialization failures.

-## TRTLLM_ATTN Backend and Skip-Softmax
+## Reference benchmark

-`TRTLLM_ATTN` runs FlashInfer's trtllm-gen FMHA. Dense it is a BF16 backend like the others; on top of
-that it can enable **Skip-Softmax**, a sparse-attention mode that trades a little fidelity for speed
-(algorithm: [design doc](../../design/feature/skip_softmax.md)).
+The following BF16 results were measured on an sm_120 RTX Pro 6000 Blackwell
+with the same prompt and seed across runs. Treat them as reference results, not
+portable guarantees.

-Enable it through the typed `skip_softmax` block on the attention spec:
-
-| Key | Valid values | Meaning |
-|---|---|---|
-| `target_sparsity` | finite, `[0, 1]` | Operating point on a calibrated curve. Needs a calibration for the model. |
-| `threshold` | finite, `≥ 0` | Direct threshold, no calibration needed. Mutually exclusive with `target_sparsity`. |
-| `disabled_until_timestep` | finite, `[0, 1]` | Holds early, high-noise steps dense; skip turns on once the normalized timestep `t ≤ D`. |
-
-```bash
-vllm-omni serve Wan-AI/Wan2.2-T2V-A14B-Diffusers \
-  --diffusion-attention-config '{"default": {"backend": "TRTLLM_ATTN",
-      "skip_softmax": {"target_sparsity": 0.65, "disabled_until_timestep": 0.86}}}'
-```
-
-Programmatically the same block is a typed `SkipSoftmaxSpec` (values validated at construction):
-
-```python
-from vllm_omni.diffusion.data import AttentionConfig, AttentionSpec, SkipSoftmaxSpec
-
-AttentionConfig(
-    default=AttentionSpec(
-        backend="TRTLLM_ATTN",
-        skip_softmax=SkipSoftmaxSpec(target_sparsity=0.65, disabled_until_timestep=0.86),
-    ),
-)
-```
-
-**Start at `target_sparsity=0.65, disabled_until_timestep=0.86`** and raise `target_sparsity` only
-as far as your quality bar allows — the output diverges from dense as it climbs. The gain grows with
-sequence length, since Skip-Softmax only accelerates attention. Dense BF16 is `TRTLLM_ATTN` with no
-`skip_softmax` block.
-
-Requires datacenter Blackwell with `head_dim == 128`; elsewhere, or without FlashInfer, selecting
-`TRTLLM_ATTN` raises rather than silently degrading.
-
-## TRTLLM_ATTN SAGE Quantization
-
-On top of dense/Skip-Softmax, `TRTLLM_ATTN` can run **SAGE** (SageAttention) — the trtllm-gen
-FMHA kernel with per-block quantized Q/K and per-channel fp8 V, so the two attention matmuls run
-in low precision. Enable it through the typed `quant` block on the attention spec:
-
-| Key | Valid values | Meaning |
-|---|---|---|
-| `dtype_qk` | `int8`, `fp8_e4m3` | Quantization dtype for Q and K. Absent ⇒ dense (SAGE off). |
-| `q_block_size` | `1`, `4`, `16` | Per-token block size for Q scales (default `1`). |
-| `k_block_size` | `1`, `4`, `16` | Per-token block size for K scales (default `16`). |
-
-V is always quantized per-channel to fp8_e4m3 and K-smoothing is applied inside the routine, so
-neither is a user knob. Only block sizes `1`, `4`, `16` have compiled kernels.
-
-The `quant` block is a shared `AttnQuantSpec` consumed by more than one backend; each reads the
-fields that apply and rejects the rest. `TRTLLM_ATTN` (SAGE) reads `dtype_qk` (`int8` / `fp8_e4m3`),
-`q_block_size`, and `k_block_size`; `FLASHINFER_ATTN` reads `dtype_qk` (`float16` / `bfloat16`),
-`dtype_vo`, and `flashinfer_backend` (see its section). A `dtype_qk` from the wrong set raises.
-
-```bash
-vllm-omni serve Wan-AI/Wan2.2-T2V-A14B-Diffusers \
-  --diffusion-attention-config '{"default": {"backend": "TRTLLM_ATTN",
-      "quant": {"dtype_qk": "fp8_e4m3", "q_block_size": 1, "k_block_size": 16}}}'
-```
-
-```python
-from vllm_omni.diffusion.data import AttentionConfig, AttentionSpec, AttnQuantSpec
-
-AttentionConfig(
-    default=AttentionSpec(
-        backend="TRTLLM_ATTN",
-        quant=AttnQuantSpec(dtype_qk="fp8_e4m3", q_block_size=1, k_block_size=16),
-    ),
-)
-```
-
-**Requirements.** Needs FlashInfer ≥ 0.6.16rc1. Kernel
-availability is arch-dependent: `fp8_e4m3` QK has kernels on both **SM100** (B200) and **SM103**
-(B300); `int8` QK kernels are compiled for **SM100 only**.
-
-## RAINFUSION_ATTN Backend and Block-Sparse Video Attention
-
-`RAINFUSION_ATTN` runs MindIE-SD's RainFusion (`rf_v2`) kernel on Ascend NPU. It pools each
-128-token block of keys, ranks them per query block, and computes attention against only the
-top-scoring ones. Video tokens are rearranged into `(t, h, w)` order first, so the kept blocks are
-spatiotemporal neighbours rather than arbitrary rows of the packed sequence.
-
-Only the video segment is sparsified. The prefix rows (text, visual conditions, audio) and the
-first-frame blocks are always kept dense, which is why the realized sparsity is lower than the
-nominal `sparsity` you configure. Anything the kernel cannot handle — a warmup denoise step, an
-exempt layer, a sequence with no published video geometry, a video segment under 32 blocks —
-delegates to `FLASH_ATTN`, so a model can select this backend unconditionally.
-
-Enable it through the typed `block_sparse` block on the attention spec, shared by every
-block-sparse backend:
-
-| Key | Valid values | Meaning |
-|---|---|---|
-| `sparsity` | finite, `[0, 1]` | Nominal fraction of key blocks dropped per query block. `0` disables sparsity. Defaults to `0.8`. |
-| `start_step` | `≥ 0` | Keep the first N denoise steps dense. Layout is decided early, so these steps dominate structural fidelity. |
-| `skip_layers` | index selector, e.g. `"0-3,38"` | DiT blocks that always stay dense. |
-
-```bash
-vllm-omni serve MiniMaxAI/MiniMax-H3 \
-  --diffusion-attention-config '{"default": {"backend": "RAINFUSION_ATTN",
-      "block_sparse": {"sparsity": 0.8, "start_step": 0}}}'
-```
-
-Programmatically the same block is a typed `BlockSparseSpec` (values validated at construction):
-
-```python
-from vllm_omni.diffusion.data import AttentionConfig, AttentionSpec, BlockSparseSpec
-
-AttentionConfig(
-    default=AttentionSpec(
-        backend="RAINFUSION_ATTN",
-        block_sparse=BlockSparseSpec(sparsity=0.8, start_step=0, skip_layers="0-1"),
-    ),
-)
-```
-
-**Tune in the order `start_step` → `sparsity` → `skip_layers`.** Raise `start_step` first: it is the
-cheapest way to recover structure, because the early high-noise steps decide global layout while the
-later steps only refine texture. Then walk `sparsity` up to your quality bar. Reach for
-`skip_layers` last, once an A/B against dense at the same seed points at specific blocks.
-
-Requires Ascend NPU with `mindiesd`; selecting it on any other platform raises. It is also
-incompatible with ring sequence parallelism, since `rf_v2` needs the whole key sequence to rank
-blocks — use Ulysses SP (`ring_degree=1`).
-
-### Which geometries run sparse
-
-MindIE-SD rf_v2 handles arbitrary video grids by spatially rearranging video first and, when
-necessary, promoting a real-video suffix to its always-kept prefix segment. It never pads: rf_v2
-does not consume an attention mask for padded keys. The remaining video rows stay sparse, while the
-promoted rows remain visible to every query and key.
-
-For example, MiniMax-H3 at 1344x768 has grid `(62, 24, 42)`: its 62496 video rows leave a 32-row
-128-block residue, and its two non-8-wide latent columns form a 2928-row rearranged suffix.
-MindIE-SD promotes 2976 rows, leaving 59520 sparse rows (`465 × 128`). This keeps the block mask
-and kernel tiling aligned, at the cost of retaining more blocks.
-
-8x8-aligned spatial grids remain preferable for performance. A latent `h` or `w` not divisible by 8
-creates an always-kept residual suffix and lowers realized sparsity. "Multiple of 8" on the latent
-grid means **width and height that are multiples of 256**. This path requires a MindIE-SD release
-that implements protected video tails for rf_v2.
-
-## End-to-End Benchmark (BF16, sm_120 RTX Pro 6000 Blackwell)
-
-Same prompt and seed across runs. `Total generation time` from `text_to_video.py` / `text_to_image.py`.
-
-| Model | Shape | TORCH_SDPA | CUDNN_ATTN | FLASHINFER_ATTN |
-|---|---|---|---|---|
+| Model | Shape | `TORCH_SDPA` | `CUDNN_ATTN` | `FLASHINFER_ATTN` |
+| --- | --- | ---: | ---: | ---: |
 | HunyuanVideo-1.5 (T2V) | 480p / 33f / 50 steps | 147.05 s | **73.02 s** | 127.84 s |
 | Wan 2.2 14B (T2V) | 480p / 33f / 40 steps | 117.75 s | 117.17 s | **115.07 s** |
 | Qwen-Image (T2I) | 1024² / 50 steps | 17.41 s | **15.14 s** | 16.02 s |
 | FLUX.2-dev (T2I) | 1024² / 50 steps, TP=2 | 53.62 s | **53.30 s** | 54.94 s |

-Pattern: mask-heavy DiTs (HV-1.5, Qwen-Image) favor `CUDNN_ATTN`; lighter-mask DiTs and TP-saturated configs (Wan 2.2, FLUX.2 TP=2) tie within noise.
+Mask-heavy DiTs favored `CUDNN_ATTN`; lighter-mask or TP-saturated workloads
+were close enough that users should benchmark their exact model and shape.

-## Known Limitations
+## Compatibility anchors

-### LTX-2.0: `CUDNN_ATTN` crashes under torch.compile
+The following headings preserve links to sections that moved into dedicated
+guides.

-LTX-2's audio attention has a symbolic head_dim under torch.compile tracing. cuDNN's SDPA backend selector rejects symbolic dims and Dynamo aborts compilation. Tracked in [#3121](https://github.com/vllm-project/vllm-omni/issues/3121).
+## TRTLLM_ATTN Backend and Skip-Softmax

-**Workaround**: explicitly select `FLASHINFER_ATTN` or `TORCH_SDPA` for LTX-2.0:
+See [TRTLLM Attention](attention_backends/trtllm.md#skip-softmax).

-```bash
-DIFFUSION_ATTENTION_BACKEND=FLASHINFER_ATTN python examples/offline_inference/text_to_video/text_to_video.py \
-    --model Lightricks/LTX-2 ...
-```
+## TRTLLM_ATTN SAGE Quantization

-### FlashAttention-4 on Blackwell
+See [TRTLLM Attention](attention_backends/trtllm.md#sage-quantization).

-Install the optional CUDA 13 extra to use the CuTe-based FA4 path:
+## RAINFUSION_ATTN Backend and Block-Sparse Video Attention

-```bash
-pip install 'vllm-omni[fa4]'
-```
-
-`FLASH_ATTN` prefers `flash_attn.cute` on Blackwell and falls back to FA3/FA2
-when it is unavailable. Version `4.0.0b18` is required; earlier beta
-wheels had known JIT failures on Blackwell.
-
-## Choosing a Backend Manually
-
-### When to override the default
-
- **Quality validation**: compare a new backend against `TORCH_SDPA` as the reference, since SDPA's default dispatcher is the most extensively tested.
- **Lossy speedup hunting**: try `SAGE_ATTN` (INT8 quantized) on diffusion outputs — typically indistinguishable visually but always validate.
- **Workaround for known issues**: see Known Limitations above.
-
-### Verifying which backend is in use
-
-The startup log prints one of:
-
-```
-Using diffusion attention backend 'CUDNN_ATTN'           # explicit override
-Defaulting to diffusion attention backend CUDNN_ATTN ... # auto-route
-Defaulting to diffusion attention backend SDPA           # nothing else available
-```
-
-If you don't see one of these, the model didn't reach diffusion stage init — check earlier logs for failures.
+See [RainFusion](attention_backends/rainfusion.md).

 ## SageAttention Installation

-vLLM-Omni expects SageAttention to be installed into the same Python environment as vLLM-Omni.
-
-Build from source:
-
-```bash
-git clone https://github.com/thu-ml/SageAttention.git
-cd SageAttention
-
-export EXT_PARALLEL=4 NVCC_APPEND_FLAGS="--threads 8" MAX_JOBS=32
-pip install . --no-build-isolation
-```
-
-Quick check:
-
-```bash
-python -c "import sageattention; print(sageattention.__file__)"
-```
+See [SageAttention](attention_backends/sage.md#installation).

 ## SageAttention3 Installation

-vLLM-Omni expects SageAttention3 to be installed into the same Python environment as vLLM-Omni.
-
-Build from source:
-
-```bash
-git clone https://github.com/thu-ml/SageAttention.git
-cd SageAttention/sageattention3_blackwell
-python setup.py install
-```
-
-Quick check:
-
-```bash
-python -c "import sageattn3; print(sageattn3.__file__)"
-```
-
-Notes:
-
- `SAGE_ATTN_3` is only selected on CUDA when `sageattn3` is importable and the GPU is Blackwell-class.
- SageAttention3's Blackwell kernel assumes `Hq == Hkv`. In vLLM-Omni, GQA/MQA diffusion requests fall back to PyTorch SDPA for correctness.
+See [SageAttention](attention_backends/sage.md#sageattention3-installation).

 ## HuggingFace Kernels Hub Backends

-To achieve perfect numerical consistency between **training** (typically using HuggingFace Diffusers with Hub kernels) and **serving/rollout** (using vLLM-Omni), you can use the Hub-based attention backends. This eliminates numerical drift / sampling divergence caused by executing different local kernel versions during rollout.
-
-The following backend options are supported:
- `FLASH_ATTN_HUB` (HuggingFace `kernels-community/flash-attn2`)
- `FLASH_ATTN_3_HUB` (HuggingFace `kernels-community/flash-attn3`, Hopper sm_90+ only)
-
-### Installation
-
-To use these backends, you must install the `kernels` library:
-
-```bash
-pip install kernels==0.14.1
-```
-
-If the `kernels` library is not available in the environment, vLLM-Omni will log a warning and fall back gracefully to the corresponding local backend implementations (`FLASH_ATTN`). On CUDA GPUs below Hopper (compute capability < 9.0), `FLASH_ATTN_3_HUB` falls back to `FLASH_ATTN_HUB`.
-
-### Usage
-
-Select a Hub backend using the global CLI flag or environment variables:
-
-```bash
-# Environment variable
-export DIFFUSION_ATTENTION_BACKEND=FLASH_ATTN_HUB
-
-# CLI flag
-vllm-omni serve <model> --diffusion-attention-backend FLASH_ATTN_HUB
-```
-
-## Usage Examples
-
-### Default (auto-route)
-
-```bash
-python examples/offline_inference/text_to_video/text_to_video.py \
-    --model hunyuanvideo-community/HunyuanVideo-1.5-Diffusers-480p_t2v \
-    --prompt "A dog running across a field of golden wheat." \
-    --height 480 --width 832 --num-frames 33 \
-    --num-inference-steps 50 --seed 42 --guidance-scale 6.0 \
-    --output hv15.mp4
-```
-
-On Blackwell this picks `CUDNN_ATTN` automatically. Check the log for the `Defaulting to ...` line.
-
-### Explicit backend selection
-
-```bash
-DIFFUSION_ATTENTION_BACKEND=FLASHINFER_ATTN python examples/offline_inference/text_to_video/text_to_video.py \
-    --model Lightricks/LTX-2 \
-    --prompt "A dog running across a field of golden wheat." \
-    --height 480 --width 832 --num-frames 33 \
-    --num-inference-steps 40 --seed 42 --guidance-scale 4.0 \
-    --output ltx2.mp4
-```
-
-### SageAttention (lossy)
-
-```bash
-DIFFUSION_ATTENTION_BACKEND=SAGE_ATTN python examples/offline_inference/text_to_video/text_to_video.py \
-    --model hunyuanvideo-community/HunyuanVideo-1.5-Diffusers-480p_t2v \
-    --prompt "A dog running across a field of golden wheat." \
-    --height 480 --width 832 --num-frames 33 \
-    --num-inference-steps 30 --seed 42 --guidance-scale 6.0 \
-    --tensor-parallel-size 2 \
-    --output hv15_sage.mp4
-```
-
-Example: Wan2.2 TI2V 5B
-
-```bash
-DIFFUSION_ATTENTION_BACKEND=SAGE_ATTN python examples/offline_inference/text_to_video/text_to_video.py \
-    --model Wan-AI/Wan2.2-TI2V-5B-Diffusers \
-    --prompt "A dog running across a field of golden wheat." \
-    --height 704 --width 1280 --num-frames 49 \
-    --num-inference-steps 30 --seed 42 --guidance-scale 5.0 \
-    --tensor-parallel-size 2 \
-    --output outputs/wan22_sage.mp4
-```
-
-### Enable SageAttention3
-
-Example:
-
-```bash
-DIFFUSION_ATTENTION_BACKEND=SAGE_ATTN_3 python examples/offline_inference/text_to_video/text_to_video.py \
-    --model hunyuanvideo-community/HunyuanVideo-1.5-Diffusers-480p_t2v \
-    --prompt "A dog running across a field of golden wheat." \
-    --height 480 --width 832 --num-frames 33 \
-    --num-inference-steps 30 --seed 42 --guidance-scale 6.0 \
-    --tensor-parallel-size 2 \
-    --output outputs/hv15_sage3.mp4
-```
+See [Hugging Face Hub Backends](attention_backends/huggingface_hub.md).
@@ -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).
@@ -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).
@@ -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).
@@ -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).
@@ -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.
@@ -1,389 +1,75 @@
 # CPU Offloading for Diffusion Models

-## Overview
+vLLM-Omni provides three CPU-offload strategies for diffusion models. Choose
+the coarsest strategy that meets your memory target: finer-grained streaming
+saves more device memory but adds more transfer and setup overhead.

-vLLM-Omni provides two offloading strategies to reduce GPU memory usage for diffusion models:
+For the shared factory, discovery, and lifecycle contract, see the
+[CPU Offloading design](../../design/feature/offloader/README.md).

-1. **Model-level (Sequential) Offloading**: Mutual exclusion between DiT model and encoder - only one is on GPU at a time.
-2. **Layerwise (Blockwise) Offloading**: Keeps only one transformer block on GPU at a time with compute-memory overlap.
+## Choose a strategy

-Both strategies use pinned memory for faster CPU-GPU transfers. The strategies are **mutually exclusive** for now - if both are enabled, layerwise takes priority.
+| Strategy | Device residency | Parallel scope | Guide |
+| --- | --- | --- | --- |
+| Model-level (sequential) | One pipeline component group at a time | Single device | [Model-Level Offloading](offloader/cpu_offload.md) |
+| Layerwise (blockwise) | One transformer block, with next-block prefetch | Single device | [Layerwise Offloading](offloader/layerwise_offload.md) |
+| Distributed layerwise | Fixed two-block device buffer; optional host-weight sharding and AllGather | Multiple GPU/NPU ranks | [Distributed Layerwise Offloading](offloader/distributed_layerwise_offload.md) |

+All strategies use pinned host memory for faster transfers where applicable.
+Configuration priority is:
+
+1. Distributed layerwise offloading.
+2. Layerwise offloading.
+3. Model-level offloading.
+
+Treat the flags as mutually exclusive. If more than one is enabled, the
+higher-priority strategy is selected.
+
+## Quick selection
+
+- Use [model-level offloading](offloader/cpu_offload.md) when swapping whole
+  encoders and DiTs is enough to fit and phase-boundary transfers are
+  acceptable.
+- Use [layerwise offloading](offloader/layerwise_offload.md) for compute-heavy video
+  DiTs where block transfers can overlap computation.
+- Use [distributed layerwise offloading](offloader/distributed_layerwise_offload.md)
+  when a multi-rank deployment also needs bounded device residency and,
+  optionally, sharded host weights.
+
+## Supported models
+
+| Architecture | Example models | DiT class | Model-level | Layerwise | Distributed layerwise | Layerwise block attributes |
+| --- | --- | --- | :---: | :---: | :---: | --- |
+| Flux2Pipeline | `black-forest-labs/FLUX.2-dev` | `Flux2Transformer2DModel` | yes | yes | — | `transformer_blocks`, `single_transformer_blocks` |
+| LongCatImagePipeline | `meituan-longcat/LongCat-Image` | `LongCatImageTransformer2DModel` | — | yes | — | `transformer_blocks`, `single_transformer_blocks` |
+| NextStep11Pipeline | `stepfun-ai/NextStep-1.1` | `NextStepModel` | — | yes | — | `layers` |
+| OvisImagePipeline | `AIDC-AI/Ovis-Image-7B` | `OvisImageTransformer2DModel` | — | yes | — | `transformer` |
+| QwenImagePipeline | `Qwen/Qwen-Image` | `QwenImageTransformer2DModel` | yes | yes | — | `transformer_blocks` |
+| StableDiffusionXLPipeline | `stabilityai/stable-diffusion-xl-base-1.0` | `SDXLUNet2DConditionModel` | yes | yes | — | `down_blocks`, `up_blocks` |
+| StableDiffusion3Pipeline | `stabilityai/stable-diffusion-3.5-medium` | `SD3Transformer2DModel` | — | yes | — | `transformer_blocks` |
+| Wan22I2VPipeline | `Wan-AI/Wan2.2-I2V-A14B-Diffusers` | `WanTransformer3DModel` | yes | yes | — | `blocks` |
+| Wan22Pipeline | `Wan-AI/Wan2.2-T2V-A14B-Diffusers` | `WanTransformer3DModel` | yes | yes | — | `blocks` |
+| SoulXSingerPipeline / SoulXSingerSVCPipeline | `Soul-AILab/SoulX-Singer` | `DiffLlama` | yes | yes | — | `layers` |
+| BagelPipeline | `ByteDance-Seed/BAGEL-7B-MoT` | `Qwen2MoTModel` | — | yes | — | `layers`, customized modules |
+| Cosmos3OmniDiffusersPipeline | `nvidia/Cosmos3-Nano`, `nvidia/Cosmos3-Super` | `Cosmos3VFMTransformer`, `Cosmos3LanguageModel` | yes | yes | yes | `layers`, `gen_layers` |
+
+Model-level support requires discoverable DiT and encoder components.
+Layerwise support requires transformer block topology. Distributed support
+reuses that topology but still requires validation for the model, checkpoint,
+and parallel configuration.
+
+## Compatibility anchors
+
+These headings preserve links to sections that moved into dedicated guides.

 ## Model-level (Sequential) Offloading

-### How It Works
-
-Model-level offloading implements mutual exclusion between DiT transformer and encoder modules using pre forward hooks:
-
- **When encoders run**: DiT transformer is offloaded to CPU
- **When DiT runs**: Encoders are offloaded to CPU, if more than one dit models, only one loaded on GPU, others get offloaded to CPU.
- **VAE**: Stays resident on GPU
-
-Before each module's forward pass, the hook automatically moves it to GPU while offloading the other module group to CPU. Transfers use pinned memory for speed.
-
-### Usage
-
-**Python API:**
-```python
-from vllm_omni import Omni
-
-m = Omni(model="Wan-AI/Wan2.2-T2V-A14B-Diffusers", enable_cpu_offload=True)
-```
-
-**CLI:**
-```bash
-vllm serve Wan-AI/Wan2.2-T2V-A14B-Diffusers --omni --enable-cpu-offload
-```
-
-### To Support a Model
-
-Implement the `SupportsComponentDiscovery` protocol to declare which
-submodules serve as pipeline components (used by offloading, HSDP
-sharding, and other framework features):
-
-```python
-from typing import ClassVar
-from vllm_omni.diffusion.models.interface import SupportsComponentDiscovery
-
-class MyPipeline(nn.Module, SupportsComponentDiscovery):
-    _dit_modules: ClassVar[list[str]] = ["transformer"]
-    _encoder_modules: ClassVar[list[str]] = ["text_encoder", "vision_model"]
-    _vae_modules: ClassVar[list[str]] = ["vae"]
-    _resident_modules: ClassVar[list[str]] = []  # optional
-
-    def __init__(self):
-        super().__init__()
-        self.transformer = ...     # DiT — stays on GPU during denoising
-        self.text_encoder = ...    # Encoder — offloaded to CPU during denoising
-        self.vision_model = ...    # Encoder — offloaded to CPU during denoising
-        self.vae = ...             # VAE — always on GPU
-```
-
- `_dit_modules`: attribute names of denoising submodules (kept on GPU
-  during the diffusion loop).
- `_encoder_modules`: attribute names of encoder/vision submodules
-  (offloaded to CPU during the diffusion loop).
- `_vae_modules`: attribute names of VAE(s) (always kept on GPU, not
-  part of the mutual exclusion hooks).
- `_resident_modules`: attribute names of small submodules that must
-  stay on GPU during layerwise offloading (e.g. embedders, connectors).
-  Optional — defaults to `[]`.
-
-All attribute names support dotted paths for nested submodules
-(e.g. `"pipe.transformer"`, `"bagel.time_embedder"`).
-
-Both DiT and encoder lists are needed because the offload hooks use
-mutual exclusion: when one group runs, the other moves to CPU.
-
-### Limitations
- Cold start latency increases
- Adds overhead from CPU-GPU transfers between encoder and denoising phases
- Support single GPU only for now
-
-
-### Component offloading for split models (e.g. Cosmos3)
-
-Some models split their transformer into mutually-exclusive *components* that run
-in different phases of a single forward pass rather than as separate pipeline
-components -- e.g. Cosmos3's understanding (reasoner) component runs once per
-generation while the generation (generator) component runs every denoising step.
-Such models have no separate text encoder to swap against, so the transformer
-owns a small model-local offload path and wraps each phase with
-`with self._offload_context(name):`
-
-```python
-class Cosmos3VFMTransformer(nn.Module):
-    def forward(self, ...):
-        with self._offload_context("reasoner"):
-            ...  # understanding pass, runs once
-        with self._offload_context("generator"):
-            ...  # denoising pass, runs every step
-```
-
-Model-level offloading then keeps exactly one component GPU-resident at a time
-(the other on CPU), reusing the same `SequentialOffloadHook` `.to()` movers. The
-pipeline opts in by exposing `enable_omni_model_cpu_offload` (which drives the
-transformer's `enable_model_cpu_offload` and pins the VAE). Layerwise offloading
-works for these models too -- each component declares its own block container via
-`_layerwise_offload_blocks_attrs`.
-
+See [Model-Level Offloading](offloader/cpu_offload.md).

 ## Layerwise (Blockwise) Offloading

-### How It Works
-
-Layerwise offloading keeps only one transformer block on GPU at a time.
-
-As each block completes, the next block is prefetched to GPU while the current block is freed. The pre and forward hooks utilized by layerwise offloading apply a separate CUDA stream (`copy_stream`) to overlap weight transfer with computation, and retain flattened tensors in pinned CPU memory for block parameters re-materialization. Encoders, VAE, and non-block DiT modules (embeddings, norms) always stay on GPU.
-
-**Execution Flow:**
-
-| Block | Pre-forward Hook | Forward | Post-forward Hook |
-|-------|------------------|---------|-------------------|
-| block-0 | Prefetch block-1 (async) | Compute block-0 | Free block-0 |
-| block-1 | Prefetch block-2 (async) | Compute block-1 | Free block-1 |
-| ... | ... | ... | ... |
-| block-(n-1) | **Prefetch block-0** (async) | Compute block-(n-1) | Free block-(n-1) |
-
-Each transformer block has a `LayerwiseOffloadHook` that prefetches the next block before forward and frees the current block after forward.
-
-Layerwise offloading is primarily recommended for large **video generation models** where the compute cost per block is high enough to effectively overlap with memory prefetch operations. For example, Wan2.2 T2V and I2V pipelines.
-
-### Usage
-
-**Python API:**
-```python
-from vllm_omni import Omni
-
-# Text-to-video
-m = Omni(model="Wan-AI/Wan2.2-T2V-A14B-Diffusers", enable_layerwise_offload=True)
-
-# Or image-to-video
-m = Omni(model="Wan-AI/Wan2.2-I2V-A14B-Diffusers", enable_layerwise_offload=True)
-```
-
-**CLI:**
-```bash
-# Text-to-video
-vllm serve Wan-AI/Wan2.2-T2V-A14B-Diffusers --omni --enable-layerwise-offload
-
-# Or image-to-video
-vllm serve Wan-AI/Wan2.2-I2V-A14B-Diffusers --omni --enable-layerwise-offload
-```
-
-### To Support a Model
-
-Models must define the blocks attribute name for layerwise offloading:
-
-```python
-class WanTransformer3DModel(nn.Module):
-    _layerwise_offload_blocks_attrs = ["blocks"]  # Attribute names containing transformer blocks
-
-    def __init__(self):
-        self.blocks = nn.ModuleList([...])  # Transformer blocks
-```
-
-For models with multiple block types:
-
-```python
-class Flux2Transformer2DModel(nn.Module):
-    _layerwise_offload_blocks_attrs = ["transformer_blocks", "single_transformer_blocks"]
-```
-
-### Limitations
- Cold start latency increases because offloaded components must be moved to CPU
-  during setup; layerwise offload may add extra weight consolidation and pinning
-  work.
- Performance depends on compute cost and H2D bandwidth as well
- Support single GPU only for now
-
+See [Layerwise Offloading](offloader/layerwise_offload.md).

 ## Distributed Layerwise Offloading

-### How It Works
-
-Distributed layerwise offloading extends single-GPU layerwise offloading to
-multi-device deployments. In the default mode, each rank stores only
-**1/dp_size** of the model weights in host memory; full layer weights are
-reconstructed at runtime via **AllGather** on a dedicated communication stream,
-overlapped with computation via a fixed double-buffer scheme. The
-`--dlo-no-use-allgather` mode instead streams standard-loader rank-local
-weights independently and does not shard weights across ranks.
-
-For the implementation design and compatibility matrix, see
-[Distributed Layerwise Offload](../../design/feature/distributed_layerwise_offload.md).
-
-**Key features:**
- **Weight sharding + AllGather**: each rank stores 1/dp_size of weights,
-  reconstructed per-layer via `all_gather_into_tensor`
- **Fixed double-buffer**: exactly 2 transformer blocks on each device at any
-  time, regardless of model size
- **DP multi-concurrency**: N concurrent requests processed in parallel
-  (AllGather only gathers weight shards, which are request-independent)
- **mmap weight loading**: weights loaded as mmap views pointing to shared OS
-  page cache, eliminating O(dp_size × model_size) RSS during model creation
- **Platform-agnostic**: works on NVIDIA GPU (CUDA/NCCL) and Ascend NPU
-  (CANN/HCCL) via vLLM-Omni's platform abstraction
-
-**Execution Flow (per device):**
-
-```
-Compute Stream:  [Layer N]          [Layer N+1]          [Layer N+2]
-H2D Stream:      [H2D Shard N+1]   [H2D Shard N+2]
-AllGather:       [AG N+1]          [AG N+2]
-Slot Usage:      Slot0: Layer N    Slot1: Layer N+1
-```
-
-### Usage
-
-**CLI:**
-```bash
-# 4× GPU/NPU with AllGather (recommended)
-vllm serve /path/to/model --omni \
-  --enable-distributed-layerwise-offload \
-  --data-parallel-size 4
-
-# Without DLO AllGather (each rank streams standard-loader rank-local weights)
-vllm serve /path/to/model --omni \
-  --enable-distributed-layerwise-offload \
-  --data-parallel-size 4 \
-  --dlo-no-use-allgather
-
-# With SP instead of DP (long sequences)
-vllm serve /path/to/model --omni \
-  --enable-distributed-layerwise-offload \
-  --usp 4
-```
-
-**Python API:**
-```python
-from vllm_omni import Omni
-
-m = Omni(
-    model="/path/to/model",
-    enable_distributed_layerwise_offload=True,
-    dlo_use_allgather=True,  # default
-)
-```
-
-### CLI Flags
-
-| Flag | Description | Default |
-|------|-------------|---------|
-| `--enable-distributed-layerwise-offload` | Enable DLO | `false` |
-| `--data-parallel-size N` | Number of DP ranks; DLO weight-sharding group when AllGather is enabled | `1` |
-| `--dlo-use-allgather` | Shard + AllGather (saves CPU, requires concurrent requests) | `true` |
-| `--dlo-no-use-allgather` | Use standard-loader rank-local weights; no DLO sharding or AllGather | `false` |
-
-### How model weights are loaded (mmap path)
-
-When DLO + AllGather is active, the offloader:
-
-1. **Saves non-persistent buffers** (e.g. RoPE `inv_freq`, timestep `freqs`)
-   from the normally-created transformer
-2. **Converts to meta device** via `to_empty(device="meta")`, releasing random
-   initialization weights
-3. **Loads checkpoint weights as mmap views** via `safe_open().get_tensor()`,
-   which return views into the OS page cache (shared across all ranks, 0 RSS)
-4. **Calls `post_load_weights()`** to apply model-specific dtype conversions
-   (e.g. Cosmos3's `time_embedder` → FP32)
-5. **Restores non-persistent buffers** from saved copies
-
-This approach requires **zero model-specific code changes** — no pipeline or
-transformer modifications are needed.
-
-### OffloadPlan (declarative topology metadata)
-
-Models can optionally declare an `OffloadPlan` class variable to provide
-topology metadata (block attribute names, submodules to offload) without
-any offload-specific logic:
-
-```python
-from vllm_omni.diffusion.offloader import OffloadPlan
-
-class MyPipeline(nn.Module):
-    _dit_modules = ["transformer"]
-    _offload_plan = OffloadPlan(
-        block_attrs={"transformer": ("blocks",)},
-        offload_submodules={"context_encoder": "layers"},
-    )
-```
-
-When not declared, the offloader falls back to `_layerwise_offload_blocks_attrs`
-and heuristic attribute search.  This is backward-compatible — existing models
-work without any changes.
-
-### DP Multi-concurrency
-
-When `--data-parallel-size > 1` and AllGather is enabled, the scheduler batches
-up to `dp_size` requests per denoise step.  Each DP rank processes a different
-request while AllGather synchronizes only weight shards (request-independent).
-
-**Requirements for concurrent requests:**
- `num_inference_steps` must be specified explicitly (None is not allowed)
- All concurrent requests must have the same `num_inference_steps` value
- These constraints exist because AllGather is a collective that requires
-  every rank to participate at each denoise step
-
-### Limitations
- Online quantization (FP8) is incompatible with mmap loading — falls back to
-  regular loading only when the model is outside the DLO+AllGather mmap path;
-  online quantization with DLO+AllGather is rejected. Use
-  `--dlo-no-use-allgather` or disable online quantization.
- Tensor Parallel > 1 is rejected in the DLO+AllGather mmap path because TP-aware
-  weight-loader callbacks are bypassed. The no-AllGather path retains the
-  standard TP loader and is experimental.
- HSDP + AllGather is rejected because it would double-shard parameters. HSDP
-  with no-AllGather is accepted at configuration level, but has limited
-  end-to-end validation.
- `num_inference_steps=None` is not allowed in DP multi-concurrency mode
-
-**Module Discovery**
-
-The offloader discovers pipeline components in two ways:
-
-1. **Protocol-based** (preferred): If the pipeline implements
-    `SupportsComponentDiscovery`, its `_dit_modules`, `_encoder_modules`,
-    `_vae_modules`, and `_resident_modules` class variables are used
-    directly.  All attribute names support dotted paths (e.g.
-    `"pipe.transformer"`, `"bagel.time_embedder"`) for nested submodules.
-
-2. **Fallback attribute scan**: Otherwise, the offloader scans for
-    well-known attribute names:
-    - **DiT modules**: `transformer`, `transformer_2`, `dit`, `sr_dit`, `language_model`, `transformer_blocks`, `model`
-    - **Encoders**: `text_encoder`, `text_encoder_2`, `text_encoder_3`, `image_encoder`
-    - **VAE**: `vae`, `audio_vae`
-
-**Hook System**
-
-Both strategies use vLLM-Omni's hook registry system (`HookRegistry` and `ModelHook`) to register pre/post forward callbacks on modules, enabling automatic swapping without modifying model code.
-
-**Backend Architecture**
-
-```
-OffloadBackend (base class)
-├── ModelLevelOffloadBackend → uses SequentialOffloadHook (.to() swap)
-│                              (delegates to a pipeline's enable_omni_model_cpu_offload
-│                               for split models like Cosmos3)
-├── LayerWiseOffloadBackend → uses LayerwiseOffloadHook
-│                          (single-GPU, full weights on host)
-└── DistributedLayerwiseOffloadBackend → uses DistributedLayerwiseOffloadHook
-                                         (multi-GPU, sharded AllGather or rank-local mode)
-```
-
-Factory function `get_offload_backend()` selects the appropriate backend based on
-configuration.
-
-For split models, `ModelLevelOffloadBackend.enable()` detects a pipeline's
-`enable_omni_model_cpu_offload` hook and delegates to it; Cosmos3 then swaps its
-reasoner/generator components inside the model forward pass.
-
-
-## Supported Models
-
-| Architecture | Example Models | DiT Class | Model-Level Offload | Layerwise Offload | Distributed Layerwise Offload | Blocks Attrs (Layerwise specific) |
-|--------------|----------------|-----------|---------------------|-------------------|-------------------------------|-----------------------------------|
-| Flux2Pipeline | `black-forest-labs/FLUX.2-dev` | `Flux2Transformer2DModel` | ✓ | ✓ | - | `"transformer_blocks"`, `"single_transformer_blocks"` |
-| LongCatImagePipeline | `meituan-longcat/LongCat-Image` | `LongCatImageTransformer2DModel` | - | ✓ | - | `"transformer_blocks"`, `"single_transformer_blocks"` |
-| NextStep11Pipeline | `stepfun-ai/NextStep-1.1` | `NextStepModel` | - | ✓ | - | `"layers"` |
-| OvisImagePipeline | `AIDC-AI/Ovis-Image-7B` | `OvisImageTransformer2DModel` | - | ✓ | - | `"transformer"` |
-| QwenImagePipeline | `Qwen/Qwen-Image` | `QwenImageTransformer2DModel` | ✓ | ✓ | - | `"transformer_blocks"` |
-| StableDiffusionXLPipeline | `stabilityai/stable-diffusion-xl-base-1.0` | `SDXLUNet2DConditionModel` | ✓ | ✓ | - | `"down_blocks"`, `"up_blocks"` |
-| StableDiffusion3Pipeline | `stabilityai/stable-diffusion-3.5-medium` | `SD3Transformer2DModel` | - | ✓ | - | `"transformer_blocks"` |
-| Wan22I2VPipeline | `Wan-AI/Wan2.2-I2V-A14B-Diffusers` | `WanTransformer3DModel` | ✓ | ✓ | - | `"blocks"` |
-| Wan22Pipeline | `Wan-AI/Wan2.2-T2V-A14B-Diffusers` | `WanTransformer3DModel` | ✓ | ✓ | - | `"blocks"` |
-| SoulXSingerPipeline / SoulXSingerSVCPipeline | `Soul-AILab/SoulX-Singer` | `DiffLlama` (`cfm_decoder.model.diff_estimator`) | ✓ | ✓ | - | `"layers"` |
-| BagelPipeline | `ByteDance-Seed/BAGEL-7B-MoT` | `Qwen2MoTModel` | - | ✓ | - | `"layers"`, `"customized modules"` |
-| Cosmos3OmniDiffusersPipeline | `nvidia/Cosmos3-Nano`, `nvidia/Cosmos3-Super` | `Cosmos3VFMTransformer`, `Cosmos3LanguageModel` | ✓ | ✓ | ✓ | `"layers"`, `"gen_layers"` |
-
-**Notes:**
- Model-Level Offloading is expected to be supported by all common diffusion models (DiT and encoders) naturally
- Layerwise Offloading requires DiT class to define `_layerwise_offload_blocks_attrs` pointing to transformer blocks
- Distributed Layerwise Offloading reuses the layerwise block topology, but the
-  model and parallelism combination still needs validation. AllGather mode
-  additionally requires a compatible checkpoint/mmap layout; no-AllGather mode
-  retains the standard loader's rank-local tensors. See the
-  [DLO design](../../design/feature/distributed_layerwise_offload.md) and
-  [Cosmos3 DistOffload recipe](https://github.com/vllm-project/vllm-omni/blob/main/recipes/cosmos3/Cosmos3-DistOffload.md)
-  for usage examples.
+See [Distributed Layerwise Offloading](offloader/distributed_layerwise_offload.md).
@@ -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.
@@ -0,0 +1,262 @@
+# Distributed Layerwise Offloading
+
+Distributed layerwise offloading (DLO) extends block streaming to multi-device
+deployments. With AllGather enabled, each rank stores roughly `1 / dp_size` of
+the host weights and reconstructs each layer at runtime. Without AllGather,
+each rank streams a complete block independently. Compatible TP1 deployments
+can share checkpoint-backed host pages among processes on the same node;
+otherwise DLO streams the ordinary loader's rank-local tensors. The optional
+runtime-weight cache extends node-local sharing to final ordinary-loader
+layouts, including matching TP coordinates across independent engines. On
+CUDA, an explicit registration budget can make those shared mappings directly
+H2D-copyable and avoid recurrent mmap-to-pinned host packing.
+
+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 request-independent weight shards, but every member of
+the transfer group must request the same next block in the same collective
+order. A synchronized DP wave may contain different prompts only when those
+requests follow the same denoising and block-execution path.
+
+## 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
+
+# Independently scheduled replicas sharing final runtime weights on one node
+vllm serve /path/to/model --omni \
+  --enable-distributed-layerwise-offload \
+  --tensor-parallel-size 2 \
+  --dlo-no-use-allgather \
+  --dlo-enable-runtime-cache \
+  --dlo-runtime-cache-dir /var/cache/vllm-omni/dlo-runtime-weights \
+  --dlo-runtime-cache-pin-limit-gib 80
+
+# 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=False,
+    dlo_enable_runtime_cache=True,
+    dlo_runtime_cache_dir="/var/cache/vllm-omni/dlo-runtime-weights",
+    dlo_runtime_cache_pin_limit_gib=80,
+)
+```
+
+## Flags
+
+| Flag | Meaning | Default |
+| --- | --- | --- |
+| `--enable-distributed-layerwise-offload` | Enable DLO | `false` |
+| `--data-parallel-size N` | DP ranks and AllGather weight-sharding group | `1` |
+| `--dlo-use-allgather` | Shard host weights and reconstruct with AllGather | `true` |
+| `--dlo-no-use-allgather` | Stream complete rank-local blocks without a DLO weight collective | `false` |
+| `--dlo-enable-runtime-cache` | Normalize final ordinary-loader DiT weights into a node-local mmap cache; requires no-AllGather | `false` |
+| `--dlo-runtime-cache-dir PATH` | Shared local-disk cache root | `~/.cache/vllm-omni/dlo-runtime-weights` |
+| `--dlo-runtime-cache-lock-timeout SECONDS` | Maximum wait for another process publishing the same layout | `600` |
+| `--dlo-runtime-cache-pin-limit-gib GIB` | Per-worker limit for registering the complete read-only runtime-cache mapping for direct H2D; zero keeps bounded host staging | `0` |
+| `--dlo-resident-layers N` | Keep N leading main-DiT blocks on device; requires no-AllGather and model-declared resident paths | `0` |
+
+## Host-weight loading
+
+The diffusion loader chooses host storage before DLO is enabled. It first
+attempts to build a complete, validated direct-checkpoint mmap plan. If names,
+coverage, shape, dtype, topology, or loader-callback compatibility cannot be
+proven, it runs the ordinary model loader instead. DLO consumes that result and
+does not make a second checkpoint-compatibility decision.
+
+The shared-mmap optimization in this phase is supported only with TP1. TP
+greater than one falls back before model mutation to ordinary TP-aware loading.
+DLO may still consume those TP-local tensors, but this is a compatibility path:
+it does not share checkpoint-backed runtime weights across DP replicas and
+provides no shared-mmap host-memory guarantee.
+
+When `--dlo-enable-runtime-cache` is set with no-AllGather, ordinary loading
+gains a second storage opportunity. Every rank completes loader callbacks,
+post-load processing, validation, and calibration. It then hashes the final
+CPU DiT parameters and persistent buffers. One process per equivalent runtime
+layout publishes sharded safetensors plus a checksummed manifest; peers
+validate and mmap the same files. DLO replaces only tensor storage, preserves
+the final `Parameter` objects, and releases the private allocator-backed copies
+before installing its streaming hooks.
+
+A positive `--dlo-runtime-cache-pin-limit-gib` deliberately selects this final
+runtime layout even when a direct TP1 checkpoint plan is available. Raw
+checkpoint mappings may require deferred loader transforms such as grouped-QKV
+reordering; registering them would still leave recurrent host work. Publishing
+the fully transformed layout once makes every streamed source directly
+copyable.
+
+The runtime layout identity excludes DP rank and SP rank. It includes TP world
+size/rank and conservative SP implementation/world-size guards. Therefore two
+independent `TP=2` engines normally create two entries: both TP0 workers map
+one entry and both TP1 workers map the other. Cache sharing does not create a
+process group or an inference-time collective.
+
+With a zero registration budget, cached tensors remain pageable mmap views.
+DLO packs each streamed block into one of two bounded pinned host slots before
+the asynchronous H2D copy. This lowers steady-state PSS but adds recurrent
+host-to-host copies; competing TP ranks or engines can turn DDR contention and
+staging skew into collective wait time.
+
+With a positive budget on CUDA, DLO instead attempts to register the complete
+page-aligned cache mapping as read-only pinned host memory. Registration is
+all-or-nothing: the total mapped range must fit the per-worker budget, and any
+CUDA registration or verification failure rolls back partial registrations and
+uses the bounded staging path. If CUDA cannot roll back a partial registration,
+worker initialization aborts instead of unmapping memory that CUDA may still
+own. On success, each immutable tensor view is copied directly into the
+existing rotating device buffers on the H2D stream, so no private host staging
+slots are allocated.
+
+CUDA registration is process-local, but it does not clone file contents. The
+same physical page-cache pages remain shared among equivalent workers; each
+worker does, however, hold its own registration reference and must satisfy the
+host's page-locking limits. Shutdown synchronizes outstanding GPU work,
+unregisters every range, and only then closes the mmap handles.
+
+The mmap plan skips only dedicated DiT weight sources. Other component sources,
+such as a text encoder loaded through the shared diffusion loader, continue to
+use their ordinary component loader. A checkpoint source that mixes DiT and
+non-DiT weights falls back completely rather than leaving an unplanned
+component uninitialized.
+
+With direct checkpoint mmap, the loader:
+
+1. saves non-persistent buffers such as RoPE frequencies;
+2. moves the normally created transformer to the meta device;
+3. loads checkpoint tensors as mmap views backed by the shared OS page cache;
+4. applies any loader-owned bounded layout adapters while packing blocks;
+5. restores saved non-persistent buffers; and
+6. preserves `post_load_weights()` and `validate_loaded_weights()` lifecycle
+   hooks.
+
+For AllGather with a group larger than one, each process copies only its
+persistent shard and then releases the source mapping. For no-AllGather, each
+process keeps the mapping open and packs complete blocks through two bounded
+pinned staging slots. Processes mapping the same files on one node share the
+immutable pages; no-AllGather still performs a complete-block H2D copy in each
+process.
+
+When the effective DLO group size is one, `dlo_use_allgather=True` does not
+perform a collective and uses the same rank-local transfer behavior.
+
+## Runtime-cache lifecycle
+
+Runtime-cache publication uses a per-layout POSIX file lock, a temporary
+directory on the same filesystem, file and final-tensor SHA-256 validation,
+`fsync`, and an atomic directory rename. A crashed writer releases its kernel
+lock; the next writer removes only that layout's stale temporary directories
+while holding the lock. Lock timeout, unsupported layout, insufficient disk,
+or validation failure keeps the already-valid ordinary tensors.
+
+The cache root is deliberately independent of `VLLM_CACHE_ROOT`, because
+vLLM-Omni may isolate that variable between stage replicas. All consumers that
+should share must receive the same `--dlo-runtime-cache-dir` on a local,
+disk-backed filesystem. Do not place the cache on tmpfs when the goal is lower
+host-memory PSS, and do not use a cross-node directory in this version.
+
+There is no automatic eviction. Budget roughly one final DiT copy per distinct
+TP coordinate, dtype/layout, and model version, plus temporary space while an
+entry is published. Operators may remove entries or the complete cache root
+only after all workers using its mmap files have stopped. The feature is
+opt-in for this reason.
+
+For direct H2D, set the registration limit above the complete page-aligned
+layout size reported in the worker log. Partial registration is intentionally
+unsupported because mixing direct and staged tensors would complicate buffer
+lifetime and make performance topology-dependent. Registration is CUDA-only;
+an insufficient budget, unsupported runtime, OS/CUDA pinning limit, or
+registration error is reported and falls back to bounded staging.
+
+## Declarative topology
+
+Models may declare an `OffloadPlan` instead of embedding offload logic:
+
+```python
+from vllm_omni.diffusion.offloader import OffloadPlan
+
+
+class MyPipeline(nn.Module):
+    _dit_modules = ["transformer"]
+    _offload_plan = OffloadPlan(
+        block_attrs={"transformer": ("blocks",)},
+        offload_submodules={"context_encoder": "layers"},
+    )
+```
+
+When no plan exists, discovery falls back to
+`_layerwise_offload_blocks_attrs` and then heuristic attribute lookup.
+
+## Data-parallel concurrency
+
+With `data_parallel_size > 1` and AllGather enabled, the scheduler can process
+up to `dp_size` requests per denoising step. Every concurrent request must set
+the same explicit `num_inference_steps`; `None` is rejected because every rank
+must enter each collective.
+
+## Limitations
+
+- Direct checkpoint mmap currently requires TP1. TP greater than one falls
+  back to the ordinary TP-aware loader. With the optional runtime cache,
+  equivalent processes at the same TP coordinate can then share that final
+  rank-local layout in no-AllGather mode. A positive registration budget also
+  chooses the final runtime cache at TP1 so deferred checkpoint transforms are
+  not repeated during inference.
+- HSDP plus AllGather is rejected to avoid double sharding. HSDP without
+  AllGather has limited end-to-end validation.
+- Online quantization uses the ordinary loader with no-AllGather. It remains
+  incompatible with DLO AllGather and runtime-cache v1. Runtime-cache v1 also
+  rejects other quantized, non-contiguous, aliased/tied, device-only, HSDP,
+  expert-parallel, CFG-parallel, and PP layouts.
+- 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.
+
+The runtime cache improves steady-state host PSS, not startup peak: every rank
+loads private weights before remapping, and validation adds full content-hash
+passes. Direct H2D removes recurrent staging but does not change that startup
+contract. Skip-load-on-cache-hit and cache eviction remain follow-up work in
+[RFC #6195](https://github.com/vllm-project/vllm-omni/issues/6195).
+
+Zero-budget staging is also not latency-neutral. In the four-L20X MiniMax-H3
+smoke documented in the L20X runtime-cache matrix in the
+[feature design](../../../design/feature/offloader/distributed_layerwise_offload.md#l20x-runtime-cache-smoke-matrix),
+two independent TP2 engines reduced request-time host PSS by 36.6% with the
+cache, but recurrent mmap-to-pinned staging reduced combined throughput by
+69.3%. The registered path recovered near-private per-engine latency while
+retaining the shared-memory benefit; a matched TP1 run improved mean wave
+latency from 33.87 seconds to 4.88 seconds with byte-identical outputs. See the
+feature design for the full smoke-test caveats and metrics.
+
+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.
@@ -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.
@@ -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
@@ -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
@@ -158,6 +161,7 @@ The following tables show which models support each feature:
 | **Wan2.1-VACE**              |     ❌     |     ✅      |           ✅           |       ✅        |         ✅         |         ❌         |   ✅    |             ✅             |      ✅ (decode)      |       ❌        |        ❌         |
 | **LTX-2**                    |     ❌     |     ✅      |           ✅           |       ✅        |         ✅         |         ❌         |   ✅    |             ✅             |      ✅ (decode)      |       ❌        |        ❌         |
 | **LTX-2.3**                  |     ❌     |     ✅      |           ✅           |       ✅        |         ✅         |         ❌         |   ✅    |             ✅             |      ✅ (decode)      |       ❌        |        ❌         |
+| **LTX-2.5**                  |     ❌     | ❓ (one-stage) | ❓ (Ulysses only) | ✅ (Full only) |         ❓         |         ❌         |   ✅    |             ✅             |      ✅ (decode)      |    ❓ (FP8)     |        ❌         |
 | **Helios**                   |     ❌     |     ✅      |           ✅           |       ✅        |         ✅         |         ❌         |   ✅    |             ✅             |          ❌           |       ❌        |        ✅*        |
 | **HunyuanVideo-1.5 T2V I2V** |     ❌     |     ✅      |           ✅           |       ✅        |         ✅         |         ❌         |   ✅    |             ✅             |  ✅ (encode/decode)   |       ✅        |        ❌         |
 | **DreamID-Omni**             |     ❌     |     ❌      |           ❌           |       ✅        |         ❌         |         ❌         |   ✅    |             ✅             |          ❌           |       ❌        |        ❌         |
@@ -171,7 +175,7 @@ The following tables show which models support each feature:
 **Frame Interpolation Support**

 - **Supported**: Wan2.2 text-to-video, image-to-video, and TI2V pipelines
- **Not supported**: Wan2.1-VACE, LTX-2, LTX-2.3, Helios, HunyuanVideo-1.5, DreamID-Omni
+- **Not supported**: Wan2.1-VACE, LTX-2, LTX-2.3, LTX-2.5, Helios, HunyuanVideo-1.5, DreamID-Omni

 ### AudioGen

@@ -214,7 +218,7 @@ 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).

@@ -43,7 +43,7 @@ if __name__ == "__main__":
            ),
        ],
    )
-    outputs[0].request_output.images[0].save("output.png")
+    outputs[0].images[0].save("output.png")
 ```

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

 ## Generation Parameters
@@ -55,7 +55,7 @@ if __name__ == "__main__":
    omni = Omni(model="Qwen/Qwen-Image")
    prompt = "a cup of coffee on the table"
    outputs = omni.generate(prompt)
-    images = outputs[0].request_output.images
+    images = outputs[0].images
    images[0].save("coffee.png")
 ```

@@ -179,7 +179,7 @@ if __name__ == "__main__":
    ]
    outputs = omni.generate(prompts)
    for i, output in enumerate(outputs):
-        output.request_output.images[0].save(f"{i}.jpg")
+        output.images[0].save(f"{i}.jpg")
 ```

 !!! info
@@ -207,7 +207,7 @@ if __name__ == "__main__":
        }
    ])
    for i, output in enumerate(outputs):
-        output.request_output.images[0].save(f"{i}.jpg")
+        output.images[0].save(f"{i}.jpg")
 ```

 You can also pass a negative prompt via the CLI argument `--negative-prompt`:
@@ -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.

@@ -80,12 +80,14 @@ warmup_quack_fp8([(14040, 2048, 6144), (14040, 2048, 2048)])

 ## Model Type Support

-### Diffusion Model (Qwen-Image, Wan2.2)
+### Diffusion Models

 | Model | HF models | Online | Pre-calibrated | Recommendation | `ignored_layers` | Text-Encoder quantization |
 |-------|-----------|:-------:|:------:|----------------|------------------|------------------|
 | Qwen-Image | `Qwen/Qwen-Image`, `Qwen/Qwen-Image-2512` | Yes | Yes | Skip sensitive image-stream MLPs when quality regresses | `img_mlp` | |
 | Wan2.2 | Wan2.2 diffusion pipelines | Not validated | Not validated | Validate against BF16 before documenting as supported | TBD | |
+| LTX-2 | `Lightricks/LTX-2`, `rootonchair/LTX-2-19b-distilled` | Yes | Not validated | Transformer only; use dynamic phase LoRA for ordinary two-stage | None | |
+| LTX-2.3 | `diffusers/LTX-2.3-Diffusers`, `diffusers/LTX-2.3-Distilled-Diffusers` | Yes | Not validated | Transformer only; use dynamic phase LoRA for ordinary two-stage | None | |
 | Z-Image | `Tongyi-MAI/Z-Image-Turbo` | Yes | Yes | All layers | None | ✅︎ |
 | FLUX.1 | `black-forest-labs/FLUX.1-dev`, `black-forest-labs/FLUX.1-schnell` | Yes | Yes | All layers | None | |
 | FLUX.2-dev | `black-forest-labs/FLUX.2-dev` | Yes | Not validated | All layers | None | ✅︎ |
@@ -93,7 +95,7 @@ warmup_quack_fp8([(14040, 2048, 6144), (14040, 2048, 2048)])
 | HunyuanImage-3.0 | `tencent/HunyuanImage-3.0`, `tencent/HunyuanImage-3.0-Instruct` | Yes | Yes | All layers; use the Hunyuan stage config for multi-stage runs | None | |
 | HunyuanVideo-1.5 | `hunyuanvideo-community/HunyuanVideo-1.5-Diffusers-480p_t2v`, `720p_t2v`, `480p_i2v` | Yes | Yes | All layers | None | |
 | Cosmos3 | `nvidia/Cosmos3-Nano`, `nvidia/Cosmos3-Super` | Yes | Not validated | All layers | None | |
-| MiniMax H3 | `MiniMaxAI/MiniMax-H3` | Yes | Not validated | DiT linears except patch, timestep, and final projections; not compatible with layerwise offload | None | |
+| MiniMax-H3 | `MiniMaxAI/MiniMax-H3` (`FL2VA` / `Ref2VA`) | Yes | Not validated | `quantization="fp8"` quantizes eligible DiT and text-encoder linears; mixed-precision input/output heads stay FP32 | None | ✅︎ |

 ### Multi-Stage Omni/TTS Model (Qwen3-Omni, Qwen3-TTS)

@@ -136,6 +138,39 @@ outputs = omni.generate(
 )
 ```

+### Global and per-component scope
+
+A plain, global quantization configuration is passed unchanged to every
+quantization-aware component constructed by a pipeline. This includes an
+eligible encoder when that encoder is implemented with vLLM quantizable
+layers; it does not rewrite arbitrary `torch.nn` modules. A structured
+component map is the only way to narrow that scope. Pipeline integrations that
+do not yet expose an encoder through the quantization factory remain DiT-only.
+
+For a pipeline that exposes both a transformer and a quantization-aware text
+encoder, the scope is:
+
+| Configuration | Transformer | Text encoder | Components without supported quantizable layers |
+|---------------|-------------|--------------|-------------------------------------------------|
+| `quantization="fp8"` | FP8 | FP8 | checkpoint precision |
+| `{"transformer": {"method": "fp8"}}` | FP8 | checkpoint precision | checkpoint precision |
+| `{"text_encoder": {"method": "fp8"}}` | checkpoint precision | FP8 | checkpoint precision |
+
+Use `quantization_config` for component-selective Python configuration. For
+example, quantize only the text decoder:
+
+```python
+omni = Omni(
+    model="<your-model>",
+    quantization_config={"text_encoder": {"method": "fp8"}},
+)
+```
+
+Component keys are runtime prefixes exposed by the pipeline integration; common
+keys include `transformer` and `text_encoder`. Entries may be combined, and
+`ignored_layers` can keep named eligible layers in checkpoint precision. Check
+the model recipe for supported components and their runtime prefixes.
+
 CLI:

 ```bash
@@ -24,7 +24,6 @@ from vllm.assets.video import VideoAsset, video_to_ndarrays
 from vllm.multimodal.media.audio import load_audio
 from vllm.utils.argparse_utils import FlexibleArgumentParser

-from vllm_omni.engine.arg_utils import nullify_stage_engine_defaults
 from vllm_omni.entrypoints.omni import Omni
 from vllm_omni.model_executor.stage_input_processors.aura_omni import (
    DEFAULT_QWEN3_TTS_REF_TEXT,
@@ -228,7 +227,6 @@ def parse_args():
    parser.add_argument("--init-timeout", type=int, default=2000, help="Overall initialization timeout.")
    parser.add_argument("--stage-init-timeout", type=int, default=2000, help="Per-stage initialization timeout.")
    parser.add_argument("--log-stats", action="store_true", default=False, help="Enable stats logging.")
-    nullify_stage_engine_defaults(parser)
    return parser.parse_args()


@@ -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",
@@ -0,0 +1,288 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""Generate a LingBot-World v2 video from an image and camera trajectory.
+
+The official checkpoint is licensed separately under CC BY-NC-SA and is
+restricted to non-commercial use. This vLLM-Omni integration remains
+Apache-2.0 licensed.
+
+Example:
+    python examples/offline_inference/diffusion/lingbot_world_v2.py \
+        --prompt "The camera moves slowly forward through the scene." \
+        --image /path/to/first_frame.png \
+        --action-dir /path/to/actions/forward \
+        --output lingbot_world_v2.mp4
+
+``--action-dir`` must contain ``poses.npy`` and ``intrinsics.npy``. The
+example configures its parent as the trusted action root and sends the
+contained directory through ``extra_args["action_path"]``.
+"""
+
+from __future__ import annotations
+
+import argparse
+import math
+from collections.abc import Sequence
+from dataclasses import dataclass
+from pathlib import Path
+from typing import Any
+
+import numpy as np
+
+_MODEL = "robbyant/lingbot-world-v2-14b-causal-fast-diffusers"
+_NUM_INFERENCE_STEPS = 4
+_MAX_SEQUENCE_LENGTH = 512
+_MAX_PIXEL_AREA = 480 * 832
+_MAX_RAW_FRAMES = 117
+_MAX_SOURCE_CAMERA_FRAMES = 4096
+_TEMPORAL_COMPRESSION = 4
+_LATENT_FRAMES_PER_BLOCK = 3
+_SPATIAL_ALIGNMENT = 16
+
+
+@dataclass(frozen=True)
+class LingBotPaths:
+    image: Path
+    action_dir: Path
+    action_root: Path
+    action_relative: Path
+    camera_frames: int
+    output: Path
+
+
+def _validate_camera_arrays(action_dir: Path) -> int:
+    """Validate the bounded camera contract before starting the engine."""
+
+    try:
+        poses = np.load(action_dir / "poses.npy", mmap_mode="r", allow_pickle=False)
+        intrinsics = np.load(action_dir / "intrinsics.npy", mmap_mode="r", allow_pickle=False)
+    except (OSError, TypeError, ValueError):
+        raise ValueError("Camera files must be valid numeric poses.npy and intrinsics.npy arrays.") from None
+    if poses.ndim != 3 or poses.shape[1:] != (4, 4):
+        raise ValueError("poses.npy must have shape [frames, 4, 4].")
+    if intrinsics.ndim != 2 or intrinsics.shape[1:] != (4,):
+        raise ValueError("intrinsics.npy must have shape [frames, 4].")
+    if poses.shape[0] != intrinsics.shape[0]:
+        raise ValueError("poses.npy and intrinsics.npy must contain the same frame count.")
+    if not 1 <= poses.shape[0] <= _MAX_SOURCE_CAMERA_FRAMES:
+        raise ValueError("Camera arrays must contain between 1 and 4096 source frames.")
+    if poses.dtype.kind not in "fiu" or intrinsics.dtype.kind not in "fiu":
+        raise ValueError("Camera arrays must contain real numeric values.")
+    if not np.isfinite(poses).all() or not np.isfinite(intrinsics).all():
+        raise ValueError("Camera arrays must contain only finite values.")
+    return int(poses.shape[0])
+
+
+def parse_args(argv: Sequence[str] | None = None) -> argparse.Namespace:
+    parser = argparse.ArgumentParser(
+        description="Generate a LingBot-World v2 video from a first-frame image and camera trajectory."
+    )
+    parser.add_argument("--model", default=_MODEL, help="Official Hugging Face model ID or local checkpoint path.")
+    parser.add_argument("--prompt", required=True, help="Text description of the scene and desired motion.")
+    parser.add_argument("--image", required=True, help="Path to the first-frame image.")
+    parser.add_argument(
+        "--action-dir",
+        required=True,
+        help="Camera action directory containing poses.npy and intrinsics.npy.",
+    )
+    parser.add_argument("--height", type=int, default=480, help="Output height; must be divisible by 16.")
+    parser.add_argument("--width", type=int, default=832, help="Output width; must be divisible by 16.")
+    parser.add_argument(
+        "--num-frames",
+        type=int,
+        default=9,
+        help="Raw output frame count. Nine frames are one three-latent-frame causal block.",
+    )
+    parser.add_argument("--seed", type=int, default=42, help="Random seed.")
+    parser.add_argument(
+        "--tensor-parallel-size",
+        type=int,
+        default=1,
+        help="Number of GPUs used for tensor parallelism inside the DiT.",
+    )
+    parser.add_argument("--flow-shift", type=float, default=5.0, help="Positive FlowUniPC scheduler shift.")
+    parser.add_argument("--fps", type=int, default=16, help="Frames per second in the exported MP4.")
+    parser.add_argument("--output", default="lingbot_world_v2.mp4", help="Output MP4 path.")
+    parser.add_argument(
+        "--enforce-eager",
+        action="store_true",
+        help="Disable compilation and force eager execution.",
+    )
+    return parser.parse_args(argv)
+
+
+def resolve_cli_paths(args: argparse.Namespace) -> LingBotPaths:
+    """Canonicalize local inputs and construct a safe trusted-root mapping."""
+
+    image = Path(args.image).expanduser().resolve()
+    if not image.is_file():
+        raise ValueError("--image must point to an existing file.")
+
+    action_dir = Path(args.action_dir).expanduser().resolve()
+    if not action_dir.is_dir():
+        raise ValueError("--action-dir must point to an existing directory.")
+    if not all((action_dir / name).is_file() for name in ("poses.npy", "intrinsics.npy")):
+        raise ValueError("--action-dir must contain poses.npy and intrinsics.npy.")
+    camera_frames = _validate_camera_arrays(action_dir)
+
+    action_root = action_dir.parent.resolve()
+    try:
+        action_relative = action_dir.relative_to(action_root)
+    except ValueError:
+        raise ValueError("--action-dir must be contained by its canonical trusted action root.") from None
+    if not action_relative.parts:
+        raise ValueError("--action-dir must name a directory contained by a trusted action root.")
+
+    output = Path(args.output).expanduser().resolve()
+    return LingBotPaths(
+        image=image,
+        action_dir=action_dir,
+        action_root=action_root,
+        action_relative=action_relative,
+        camera_frames=camera_frames,
+        output=output,
+    )
+
+
+def _positive_finite(value: float, flag: str) -> float:
+    result = float(value)
+    if not math.isfinite(result) or result <= 0:
+        raise ValueError(f"{flag} must be a positive finite number.")
+    return result
+
+
+def build_omni_kwargs(
+    args: argparse.Namespace,
+    paths: LingBotPaths,
+    *,
+    parallel_config: Any,
+) -> dict[str, Any]:
+    """Build Omni configuration without overriding checkpoint class discovery."""
+
+    if args.tensor_parallel_size <= 0:
+        raise ValueError("--tensor-parallel-size must be a positive integer.")
+    flow_shift = _positive_finite(args.flow_shift, "--flow-shift")
+    model_path = Path(args.model).expanduser()
+    model = str(model_path.resolve()) if model_path.exists() else args.model
+    return {
+        "model": model,
+        "flow_shift": flow_shift,
+        "parallel_config": parallel_config,
+        "enforce_eager": args.enforce_eager,
+        "model_config": {"lingbot_action_root": str(paths.action_root)},
+    }
+
+
+def build_request(
+    args: argparse.Namespace,
+    paths: LingBotPaths,
+) -> tuple[dict[str, Any], dict[str, Any]]:
+    """Build the prompt and fixed v1 causal-DMD sampling contract."""
+
+    prompt_text = args.prompt.strip()
+    if not prompt_text:
+        raise ValueError("--prompt must contain non-whitespace text.")
+    if args.height <= 0 or args.width <= 0:
+        raise ValueError("--height and --width must be positive integers.")
+    if args.height % _SPATIAL_ALIGNMENT or args.width % _SPATIAL_ALIGNMENT:
+        raise ValueError("--height and --width must each be divisible by 16.")
+    if args.height * args.width > _MAX_PIXEL_AREA:
+        raise ValueError("--height * --width must not exceed 480 * 832 pixels.")
+    if args.num_frames <= 0:
+        raise ValueError("--num-frames must be a positive integer.")
+    if args.num_frames > _MAX_RAW_FRAMES:
+        raise ValueError("--num-frames must not exceed 117 raw frames.")
+    if args.num_frames > paths.camera_frames:
+        raise ValueError("Camera arrays must contain at least --num-frames frames.")
+    if (args.num_frames - 1) % _TEMPORAL_COMPRESSION:
+        raise ValueError("--num-frames must satisfy (num_frames - 1) divisible by 4.")
+    latent_frames = (args.num_frames - 1) // _TEMPORAL_COMPRESSION + 1
+    if latent_frames % _LATENT_FRAMES_PER_BLOCK:
+        raise ValueError("--num-frames must map to whole three-frame latent blocks.")
+    if args.fps <= 0:
+        raise ValueError("--fps must be a positive integer.")
+    flow_shift = _positive_finite(args.flow_shift, "--flow-shift")
+
+    prompt = {
+        "prompt": prompt_text,
+        "multi_modal_data": {"image": str(paths.image)},
+    }
+    sampling_kwargs = {
+        "height": args.height,
+        "width": args.width,
+        "num_frames": args.num_frames,
+        "num_inference_steps": _NUM_INFERENCE_STEPS,
+        "max_sequence_length": _MAX_SEQUENCE_LENGTH,
+        "seed": args.seed,
+        "fps": args.fps,
+        "extra_args": {
+            "action_path": paths.action_relative.as_posix(),
+            "flow_shift": flow_shift,
+        },
+    }
+    return prompt, sampling_kwargs
+
+
+def extract_video_array(outputs: Any) -> np.ndarray:
+    """Unwrap one standard Omni diffusion video into ``[frames, H, W, C]``."""
+
+    if not isinstance(outputs, list) or len(outputs) != 1:
+        raise ValueError("Expected exactly one Omni output for the single LingBot request.")
+    output = outputs[0]
+    request_output = getattr(output, "request_output", None)
+    if request_output is not None and getattr(request_output, "images", None):
+        output = request_output
+
+    images = getattr(output, "images", None)
+    if not images:
+        multimodal_output = getattr(output, "multimodal_output", None) or {}
+        images = multimodal_output.get("video")
+    if images is None:
+        raise ValueError("No video frames found in Omni output.")
+    if isinstance(images, list) and len(images) == 1:
+        images = images[0]
+
+    video: np.ndarray = np.asarray(images)
+    if video.ndim == 5:
+        if video.shape[0] != 1:
+            raise ValueError(f"Expected one generated video, got batch shape {video.shape}.")
+        video = video[0]
+    if video.ndim != 4 or video.shape[-1] not in (3, 4):
+        raise ValueError(f"Expected video shape [frames, height, width, channels], got {video.shape}.")
+    return video
+
+
+def main(argv: Sequence[str] | None = None) -> Path:
+    args = parse_args(argv)
+    paths = resolve_cli_paths(args)
+    prompt, sampling_kwargs = build_request(args, paths)
+    paths.output.parent.mkdir(parents=True, exist_ok=True)
+
+    # Keep vLLM imports below pure CLI/path validation so ``--help`` and the
+    # helper tests work on machines without the native vLLM runtime.
+    from diffusers.utils import export_to_video
+
+    from vllm_omni.diffusion.data import DiffusionParallelConfig
+    from vllm_omni.entrypoints.omni import Omni
+    from vllm_omni.inputs.data import OmniDiffusionSamplingParams
+
+    parallel_config = DiffusionParallelConfig(tensor_parallel_size=args.tensor_parallel_size)
+    omni_kwargs = build_omni_kwargs(args, paths, parallel_config=parallel_config)
+    omni = Omni(**omni_kwargs)
+    try:
+        outputs = omni.generate(
+            prompt,
+            OmniDiffusionSamplingParams(**sampling_kwargs),
+            use_tqdm=False,
+        )
+    finally:
+        omni.close()
+
+    frames = extract_video_array(outputs)
+    export_to_video(frames, str(paths.output), fps=args.fps)
+    print(f"Saved {frames.shape[0]} frames to {paths.output}")
+    return paths.output
+
+
+if __name__ == "__main__":
+    main()
@@ -0,0 +1,239 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""Run LingBot-World 2.0 as an in-process realtime AR-Diffusion session.
+
+This example exercises the internal session API. It is not a public HTTP or
+WebSocket protocol. Each JSONL event advances the world by one three-latent
+frame AR block and writes the latent plus its identity metadata. The current
+scoped implementation supports at most ten ticks per generation epoch because
+its image condition is bounded to 117 pixel frames; reset or create a session
+to start a new world.
+"""
+
+from __future__ import annotations
+
+import argparse
+import asyncio
+import json
+import math
+import time
+from collections.abc import Sequence
+from pathlib import Path
+from typing import Any
+
+_MODEL = "robbyant/lingbot-world-v2-14b-causal-fast-diffusers"
+_CAMERA_ACTION_SCHEMA = "lingbot.camera_actions.v1"
+_FRAMES_PER_BLOCK = 3
+# ((117 pixel frames - 1) / VAE temporal factor 4 + 1) / 3 latent frames.
+_MAX_REALTIME_TICKS = 10
+
+
+def _camera_event_data(frames: list[list[str]]) -> dict[str, Any]:
+    """Build the event-side script consumed by LingBotCameraControlReducer."""
+    return {"mode": "script", "frames": frames}
+
+
+def parse_args(argv: Sequence[str] | None = None) -> argparse.Namespace:
+    parser = argparse.ArgumentParser(description="Run an in-process realtime LingBot-World 2.0 session.")
+    parser.add_argument("--model", default=_MODEL, help="Hugging Face model ID or local checkpoint path.")
+    parser.add_argument("--image", required=True, help="Initial RGB image.")
+    parser.add_argument("--prompt", required=True, help="Initial scene prompt.")
+    parser.add_argument(
+        "--events",
+        required=True,
+        help="JSONL file with one prompt/action event per AR block; at most 10 events.",
+    )
+    parser.add_argument("--output-dir", required=True, help="Directory for chunk latents and metadata.")
+    parser.add_argument("--session-id", default="lingbot-world", help="Persistent world session identifier.")
+    parser.add_argument("--height", type=int, default=480)
+    parser.add_argument("--width", type=int, default=832)
+    parser.add_argument("--seed", type=int, default=42)
+    parser.add_argument("--gpu-memory-fraction", type=float, default=0.1)
+    parser.add_argument("--tensor-parallel-size", type=int, default=1)
+    parser.add_argument("--enforce-eager", action="store_true")
+    return parser.parse_args(argv)
+
+
+def _load_events(path: Path) -> list[dict[str, Any]]:
+    events: list[dict[str, Any]] = []
+    for line_number, raw_line in enumerate(path.read_text().splitlines(), start=1):
+        line = raw_line.strip()
+        if not line:
+            continue
+        try:
+            value = json.loads(line)
+        except json.JSONDecodeError as exc:
+            raise ValueError(f"events line {line_number} is not valid JSON: {exc.msg}.") from None
+        if not isinstance(value, dict):
+            raise ValueError(f"events line {line_number} must be a JSON object.")
+        event_id = value.get("event_id")
+        if isinstance(event_id, bool) or not isinstance(event_id, int) or event_id < 0:
+            raise ValueError(f"events line {line_number} requires a non-negative integer event_id.")
+        prompt = value.get("prompt")
+        if prompt is not None and (not isinstance(prompt, str) or not prompt.strip()):
+            raise ValueError(f"events line {line_number} prompt must be a non-empty string.")
+        frames = value.get("frames")
+        if frames is not None:
+            if not isinstance(frames, list) or len(frames) != _FRAMES_PER_BLOCK:
+                raise ValueError(f"events line {line_number} frames must contain exactly three lists.")
+            for frame in frames:
+                if not isinstance(frame, list) or any(not isinstance(action, str) for action in frame):
+                    raise ValueError(f"events line {line_number} frames must contain only action strings.")
+        if prompt is None and frames is None:
+            raise ValueError(f"events line {line_number} must update prompt and/or frames.")
+        events.append({"event_id": event_id, "prompt": prompt, "frames": frames})
+    if not events:
+        raise ValueError("events file must contain at least one event.")
+    if len(events) > _MAX_REALTIME_TICKS:
+        raise ValueError(
+            "events file must contain at most "
+            f"{_MAX_REALTIME_TICKS} events because the current LingBot realtime "
+            "image-condition horizon is 117 pixel frames."
+        )
+    event_ids = [event["event_id"] for event in events]
+    if event_ids != sorted(set(event_ids)):
+        raise ValueError("event_id values must be unique and strictly increasing.")
+    return events
+
+
+def _validate_args(args: argparse.Namespace) -> tuple[Path, Path, Path]:
+    image = Path(args.image).expanduser().resolve()
+    events = Path(args.events).expanduser().resolve()
+    output_dir = Path(args.output_dir).expanduser().resolve()
+    if not image.is_file():
+        raise ValueError("--image must point to an existing file.")
+    if not events.is_file():
+        raise ValueError("--events must point to an existing JSONL file.")
+    if not args.prompt.strip():
+        raise ValueError("--prompt must contain non-whitespace text.")
+    if args.height <= 0 or args.width <= 0 or args.height % 16 or args.width % 16:
+        raise ValueError("--height and --width must be positive multiples of 16.")
+    if args.tensor_parallel_size <= 0:
+        raise ValueError("--tensor-parallel-size must be positive.")
+    if not math.isfinite(args.gpu_memory_fraction) or not 0 < args.gpu_memory_fraction <= 1:
+        raise ValueError("--gpu-memory-fraction must be in (0, 1].")
+    return image, events, output_dir
+
+
+async def run(argv: Sequence[str] | None = None) -> Path:
+    args = parse_args(argv)
+    image, events_path, output_dir = _validate_args(args)
+    events = _load_events(events_path)
+    output_dir.mkdir(parents=True, exist_ok=True)
+
+    # Imports stay below pure input validation so --help and helper tests do not
+    # require a CUDA-enabled vLLM installation.
+    import torch
+
+    from vllm_omni.diffusion.data import DiffusionParallelConfig
+    from vllm_omni.diffusion.models.lingbot_world.actions import LingBotCameraControlReducer
+    from vllm_omni.entrypoints.async_omni import AsyncOmni
+    from vllm_omni.experimental.ar_diffusion.consumer import ARDiffusionOmniTickConsumer
+    from vllm_omni.experimental.ar_diffusion.session import (
+        ARDiffusionSessionEvent,
+        ARDiffusionSessionManager,
+        ARDiffusionWorkerLifecycle,
+    )
+    from vllm_omni.experimental.ar_diffusion.tick_protocol import ARDiffusionControlInput
+    from vllm_omni.inputs.data import OmniDiffusionSamplingParams
+
+    engine = AsyncOmni(
+        model=args.model,
+        engine_backend="vllm_omni.experimental.ar_diffusion.engine.ARDiffusionEngine",
+        enforce_eager=args.enforce_eager,
+        parallel_config=DiffusionParallelConfig(tensor_parallel_size=args.tensor_parallel_size),
+        max_num_seqs=1,
+        model_config={
+            "ar_diffusion_height": args.height,
+            "ar_diffusion_width": args.width,
+            "ar_diffusion_kv_config": {
+                "gpu_memory_fraction": args.gpu_memory_fraction,
+                "warmup_cudagraph": True,
+            },
+        },
+    )
+    sampling = OmniDiffusionSamplingParams(
+        height=args.height,
+        width=args.width,
+        num_frames=9,
+        num_inference_steps=4,
+        max_sequence_length=512,
+        seed=args.seed,
+        output_type="latent",
+        extra_args={"flow_shift": 5.0},
+    )
+    consumer = ARDiffusionOmniTickConsumer(
+        engine,
+        prompt_provider=lambda tick: {
+            "prompt": tick.prompt,
+            "multi_modal_data": {"image": str(image)},
+        },
+        sampling_params_list=[sampling],
+        diffusion_stage_id=0,
+    )
+    manager = ARDiffusionSessionManager(
+        tick_consumer=consumer,
+        lifecycle=ARDiffusionWorkerLifecycle(engine, stage_ids=[0], timeout=180.0),
+        max_pending_events=32,
+        control_reducer_factory=LingBotCameraControlReducer,
+    )
+    session = await manager.create_session(args.session_id)
+    measurements: list[dict[str, Any]] = []
+    try:
+        for chunk_index, event in enumerate(events):
+            controls = ()
+            if event["frames"] is not None:
+                controls = (
+                    ARDiffusionControlInput(
+                        track="camera",
+                        schema=_CAMERA_ACTION_SCHEMA,
+                        data=_camera_event_data(event["frames"]),
+                    ),
+                )
+            await session.accept_event(
+                ARDiffusionSessionEvent(
+                    event_id=event["event_id"],
+                    prompt=event["prompt"]
+                    if event["prompt"] is not None
+                    else (args.prompt if chunk_index == 0 else None),
+                    controls=controls,
+                )
+            )
+            started = time.perf_counter()
+            output = await session.next_chunk()
+            elapsed = time.perf_counter() - started
+            if len(output.images) != 1 or not isinstance(output.images[0], torch.Tensor):
+                raise RuntimeError("Expected one latent tensor from each realtime LingBot chunk.")
+            latent = output.images[0].detach().float().cpu()
+            metadata = consumer.chunk_metadata(output).to_dict()
+            torch.save(latent, output_dir / f"chunk_{chunk_index:03d}.pt")
+            (output_dir / f"chunk_{chunk_index:03d}.json").write_text(
+                json.dumps(metadata, indent=2, sort_keys=True) + "\n"
+            )
+            measurements.append(
+                {
+                    "chunk_index": chunk_index,
+                    "latency_seconds": elapsed,
+                    "shape": list(latent.shape),
+                    "finite": bool(torch.isfinite(latent).all()),
+                    "metadata": metadata,
+                }
+            )
+            print(json.dumps(measurements[-1], sort_keys=True), flush=True)
+    finally:
+        try:
+            await manager.close_session(args.session_id)
+        finally:
+            engine.shutdown()
+
+    summary_path = output_dir / "summary.json"
+    summary_path.write_text(json.dumps({"chunks": measurements}, indent=2, sort_keys=True) + "\n")
+    return summary_path
+
+
+def main(argv: Sequence[str] | None = None) -> Path:
+    return asyncio.run(run(argv))
+
+
+if __name__ == "__main__":
+    main()
@@ -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())
-    omni = Omni(model=model_source, stage_configs_path=stage_config_path, dtype=args.dtype)
+    deploy_config_path = str(Path(args.deploy_config).expanduser())
+    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)
@@ -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:
-        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]
+    deploy_config = args.deploy_config or _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
-    else:
-        omni_kwargs["stage_configs_path"] = stage_configs_path
+    omni_kwargs["deploy_config"] = deploy_config

    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}")
-    else:
-        print(f"  Stage config: {stage_configs_path}")
+    print(f"  Deploy config: {deploy_config}")
    print(f"  Num stages: {omni.num_stages}")
    if args.modality in ("text2img", "img2img"):
        print(f"  Inference steps: {args.steps}")
@@ -96,7 +96,7 @@ if __name__ == "__main__":
        ),
    )
    from diffusers.utils import export_to_video
-    frames = outputs[0].request_output.images
+    frames = outputs[0].images
    export_to_video(frames, "quick_test_i2v.mp4", fps=16)
 ```

@@ -68,6 +68,7 @@ from vllm_omni.model_extras import (
 from vllm_omni.model_extras import (
    get_extra_body_params,
    get_model_class_name,
+    should_preserve_reference_image_size,
 )
 from vllm_omni.outputs import OmniRequestOutput
 from vllm_omni.platforms import current_omni_platform
@@ -454,7 +455,13 @@ def main():

    media_inputs: dict[str, Any] = {}
    if image is not None:
-        media_inputs["image"] = image.resize((width, height), PIL.Image.Resampling.LANCZOS)
+        preserve_image_size = should_preserve_reference_image_size(
+            model_class_name,
+            model=args.model,
+        )
+        media_inputs["image"] = (
+            image if preserve_image_size else image.resize((width, height), PIL.Image.Resampling.LANCZOS)
+        )
    if last_image is not None:
        media_inputs["last_image"] = last_image.resize((width, height), PIL.Image.Resampling.LANCZOS)
    if mask_image is not None:
@@ -82,9 +82,6 @@ def parse_args():
    # ``max_num_seqs=1`` …) are passed as flat kwargs to ``Omni`` below
    # and ``create_default_diffusion`` materializes the stage config.

-    from vllm_omni.engine.arg_utils import nullify_stage_engine_defaults
-
-    nullify_stage_engine_defaults(p)
    return p.parse_args()


@@ -335,7 +335,7 @@ def main(args):
    print(f"use_tts (tts_bos): {use_tts}")

    for stage_outputs in omni_generator:
-        output = stage_outputs.request_output
+        output = stage_outputs
        if stage_outputs.final_output_type == "text":
            request_id = output.request_id
            text_output = output.outputs[0].text
@@ -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,
@@ -1,21 +1,32 @@
 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-"""Reproduce MiniMax-H3 request-mode and DLO/DP2 lifecycle behavior.
+"""Validate MiniMax-H3 request-mode and DLO lifecycle behavior.
+
+The DLO modes accept any positive data-parallel size.  With AllGather enabled,
+the script submits one complete DP wave and verifies recovery after an invalid
+wave.  It also checks that every mode shuts down all worker processes cleanly.

 Examples:

    VLLM_WORKER_MULTIPROC_METHOD=spawn \
    VLLM_OMNI_VIDEO_SYNC_TIMEOUT=14400 \
    VLLM_OMNI_DLO_DP_WAVE_TIMEOUT=600 \
+    CUDA_VISIBLE_DEVICES=0,1,2,3 \
+    python examples/offline_inference/minimax_h3/dlo_lifecycle.py \
+        --model /path/to/MiniMax-H3/FL2VA --mode dlo --dp-size 4
+
+    # Rank-local mmap storage, with no DLO collective or DP wave scheduling.
+    VLLM_WORKER_MULTIPROC_METHOD=spawn \
+    VLLM_OMNI_VIDEO_SYNC_TIMEOUT=14400 \
    CUDA_VISIBLE_DEVICES=0,1 \
-    python examples/offline_inference/minimax_h3/dlo_dp2_lifecycle.py \
-        --model /path/to/MiniMax-H3/FL2VA --mode dlo-dp2
+    python examples/offline_inference/minimax_h3/dlo_lifecycle.py \
+        --model /path/to/MiniMax-H3/FL2VA --mode dlo-no-allgather --dp-size 2

    VLLM_WORKER_MULTIPROC_METHOD=spawn \
    VLLM_OMNI_VIDEO_SYNC_TIMEOUT=14400 \
    CUDA_VISIBLE_DEVICES=0,1 \
-    python examples/offline_inference/minimax_h3/dlo_dp2_lifecycle.py \
-        --model /path/to/MiniMax-H3/FL2VA --mode request
+    python examples/offline_inference/minimax_h3/dlo_lifecycle.py \
+        --model /path/to/MiniMax-H3/FL2VA --mode request --tp-size 2
 """

 from __future__ import annotations
@@ -40,14 +51,33 @@ DEFAULT_PROMPTS = (
 )


+def positive_int(value: str) -> int:
+    parsed = int(value)
+    if parsed < 1:
+        raise argparse.ArgumentTypeError("must be a positive integer")
+    return parsed
+
+
 def parse_args() -> argparse.Namespace:
    parser = argparse.ArgumentParser(description=__doc__)
    parser.add_argument("--model", required=True, help="Path to MiniMax-H3/FL2VA")
    parser.add_argument(
        "--mode",
-        choices=("dlo-dp2", "request"),
+        choices=("dlo", "dlo-no-allgather", "request"),
        required=True,
-        help="DLO AllGather DP2 or ordinary non-DLO TP2 request mode",
+        help="DLO (with or without AllGather) or ordinary non-DLO request mode",
+    )
+    parser.add_argument(
+        "--dp-size",
+        type=positive_int,
+        default=2,
+        help="DLO data-parallel size (for example 1, 2, 4, or 8; default: 2)",
+    )
+    parser.add_argument(
+        "--tp-size",
+        type=positive_int,
+        default=2,
+        help="Tensor-parallel size used only by request mode (default: 2)",
    )
    parser.add_argument("--steps", type=int, default=2)
    parser.add_argument("--duration", type=float, default=5.0)
@@ -60,10 +90,11 @@ def parse_args() -> argparse.Namespace:


 def engine_kwargs(args: argparse.Namespace) -> dict[str, Any]:
+    is_dlo = args.mode in {"dlo", "dlo-no-allgather"}
    common: dict[str, Any] = {
        "model": args.model,
        "trust_remote_code": True,
-        "num_gpus": 2,
+        "num_gpus": args.dp_size if is_dlo else args.tp_size,
        "usp": 1,
        "ring": 1,
        "vae_parallel_mode": "tile",
@@ -74,22 +105,22 @@ def engine_kwargs(args: argparse.Namespace) -> dict[str, Any]:
        "stage_init_timeout": args.init_timeout,
        "init_timeout": args.init_timeout,
    }
-    if args.mode == "dlo-dp2":
+    if is_dlo:
        common.update(
            tensor_parallel_size=1,
-            data_parallel_size=2,
+            data_parallel_size=args.dp_size,
            text_encoder_tp_size=1,
            vae_patch_parallel_size=1,
            enable_distributed_layerwise_offload=True,
-            dlo_use_allgather=True,
+            dlo_use_allgather=args.mode == "dlo",
            dlo_resident_layers=0,
        )
    else:
        common.update(
-            tensor_parallel_size=2,
+            tensor_parallel_size=args.tp_size,
            data_parallel_size=1,
-            text_encoder_tp_size=2,
-            vae_patch_parallel_size=2,
+            text_encoder_tp_size=args.tp_size,
+            vae_patch_parallel_size=args.tp_size,
            enable_distributed_layerwise_offload=False,
        )
    return common
@@ -172,7 +203,8 @@ async def run(args: argparse.Namespace) -> dict[str, Any]:
        "engine_kwargs": engine_kwargs(args),
    }
    try:
-        if args.mode == "dlo-dp2":
+        uses_dp_wave = args.mode == "dlo" and args.dp_size > 1
+        if uses_dp_wave:
            started = time.perf_counter()
            asymmetric = await asyncio.gather(
                generate_one(
@@ -182,12 +214,15 @@ async def run(args: argparse.Namespace) -> dict[str, Any]:
                    prompt="",
                    seed=1001,
                ),
-                generate_one(
-                    engine,
-                    args,
-                    request_id="invalid-peer",
-                    prompt=DEFAULT_PROMPTS[0],
-                    seed=1002,
+                *(
+                    generate_one(
+                        engine,
+                        args,
+                        request_id=f"invalid-peer-{index}",
+                        prompt=DEFAULT_PROMPTS[index % len(DEFAULT_PROMPTS)],
+                        seed=1001 + index,
+                    )
+                    for index in range(1, args.dp_size)
                ),
                return_exceptions=True,
            )
@@ -195,18 +230,18 @@ async def run(args: argparse.Namespace) -> dict[str, Any]:
            summary["asymmetric_errors"] = [
                f"{type(result).__name__}: {result}" for result in asymmetric if isinstance(result, BaseException)
            ]
-            if len(summary["asymmetric_errors"]) != 2:
-                raise RuntimeError("Both requests in the asymmetric DP wave must fail before dispatch")
+            if len(summary["asymmetric_errors"]) != args.dp_size:
+                raise RuntimeError("Every request in the asymmetric DP wave must fail before dispatch")

        started = time.perf_counter()
-        request_count = 2 if args.mode == "dlo-dp2" else 1
+        request_count = args.dp_size if uses_dp_wave else 1
        outputs = await asyncio.gather(
            *(
                generate_one(
                    engine,
                    args,
                    request_id=f"recovery-{index}",
-                    prompt=DEFAULT_PROMPTS[index],
+                    prompt=DEFAULT_PROMPTS[index % len(DEFAULT_PROMPTS)],
                    seed=2000 + index,
                )
                for index in range(request_count)
@@ -67,7 +67,7 @@ def main() -> None:
    parser.add_argument("--output-dir", default="results/nemotron_voicechat")
    parser.add_argument("--system-prompt", default=DEFAULT_SYSTEM_PROMPT)
    parser.add_argument(
-        "--stage-configs-path",
+        "--deploy-config",
        default=None,
        help=(
            "Deploy yaml override. Pass vllm_omni/deploy/nemotron_labs_voicechat_streaming.yaml "
@@ -134,8 +134,8 @@ def main() -> None:
    ]

    omni_kwargs: dict = {}
-    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
    omni = Omni(model=str(ckpt), trust_remote_code=True, **omni_kwargs)
    outputs = omni.generate(inputs, sampling_params_list)

@@ -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",
@@ -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
-    # custom configs, but use bundled deploy configs by default.
-    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:
+    # Resolve an explicit deploy config or use the bundled task default.
+    if 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["stage_configs_path"] = stage_config_path
-    else:
-        omni_kwargs["deploy_config"] = deploy_config_path
+    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,
@@ -74,7 +74,7 @@ if __name__ == "__main__":
    omni = Omni(model="Qwen/Qwen-Image")
    prompt = "a cup of coffee on the table"
    outputs = omni.generate(prompt)
-    images = outputs[0].request_output.images
+    images = outputs[0].images
    images[0].save("coffee.png")
 ```

@@ -265,7 +265,7 @@ if __name__ == "__main__":
    ]
    outputs = omni.generate(prompts)
    for i, output in enumerate(outputs):
-        output.request_output.images[0].save(f"{i}.jpg")
+        output.images[0].save(f"{i}.jpg")
 ```

 !!! info
@@ -301,7 +301,7 @@ if __name__ == "__main__":
        }
    ])
    for i, output in enumerate(outputs):
-        output.request_output.images[0].save(f"{i}.jpg")
+        output.images[0].save(f"{i}.jpg")
 ```

 You can also pass a negative prompt via the CLI argument `--negative-prompt`:
@@ -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:
-        print(f"  stage-configs-path: {args.stage_configs_path}")
+    if args.deploy_config:
+        print(f"  deploy-config: {args.deploy_config}")
    print(f"{'=' * 60}\n")

    # Build LoRA request when --lora-path is set
@@ -16,6 +16,7 @@ list of supported architectures across all modalities, see
 |---|---|---|---|---|---|---|
 | CosyVoice3 | `FunAudioLLM/Fun-CosyVoice3-0.5B-2512` | 2 (talker + code2wav) | ✓ | ✓ | — | 24 kHz |
 | Fish Speech S2 Pro | `fishaudio/s2-pro` | dual-AR | ✓ | ✓ | — | 44.1 kHz |
+| Gepard-1.0 | `nineninesix/gepard-1.0` | single (native AR + NanoCodec) | — (zero-shot; cloning WIP) | — (serving WIP) | zero-shot | 22.05 kHz |
 | GLM-TTS | `zai-org/GLM-TTS` | 2 (AR + DiT) | ✓ (required) | ✓ | — | 24 kHz |
 | Ming-omni-tts | `inclusionAI/Ming-omni-tts-0.5B` | 2 (AR + audio VAE) | ✓ | ✓ | style / IP / dialect / TTA / podcast | 44.1 kHz |
 | Ming-flash-omni-TTS | `Jonathan1909/Ming-flash-omni-2.0` | single (talker only) | — (caption-controlled) | — | style / IP / basic captions | 44.1 kHz |
@@ -23,6 +24,7 @@ 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 |
@@ -394,6 +396,58 @@ python examples/offline_inference/text_to_speech/qwen3_tts/end2end.py \

 ---

+## Gepard-1.0
+
+Single-stage native AR TTS at 22.05 kHz. Pipeline: `Qwen3.5 backbone → 32 FSQ codebook heads (one frame/step) → NeMo NanoCodec`. Backbone runs under vLLM paged attention; the 32-head sampling + learned embedding feedback ride the native-AR runner hooks. `enforce_eager` (CUDA graph is a perf follow-up).
+
+### Prerequisites
+The NanoCodec decoder needs NeMo (NVIDIA Open Model License), installed separately:
+
+```bash
+uv pip install "nemo_toolkit[tts]==2.7.3" \
+    --overrides <(printf '%s\n' "transformers>=5.5.3" "huggingface-hub>=1.0")
+```
+
+The overrides are what make that command work. `nemo_toolkit[tts]` declares
+`transformers~=4.57.0`, so a plain install downgrades the version
+`requirements/common.txt` pins and takes `huggingface-hub` back to 0.x with it — after
+which `import vllm` fails. Both packages are named here rather than left to resolve on
+their own: transformers 5.x happens to require `huggingface-hub` 1.x, but relying on
+that would make the second repair an accident of the first. `--constraint` cannot be
+used instead, since it narrows a version range rather than overriding a dependency a
+package declares, and the two sets never intersect. NeMo 2.7.3 decodes correctly
+against transformers 5.x, so nothing is lost by holding it there.
+
+(The same clash is why this is documented here rather than declared in
+`pyproject.toml`: an extra resolves together with the base dependencies, so
+`vllm-omni[gepard]` would be unsatisfiable rather than opt-in. SoulX-Singer English
+SVS documents `nemo_toolkit[asr]` the same way.)
+
+On a host whose CUDA toolkit cannot build kernels for the GPU — no `nvcc`/`ninja`, or a
+consumer Blackwell (`sm_120`) card — also set `VLLM_USE_FLASHINFER_SAMPLER=0`. vLLM's
+sampler JIT-compiles through FlashInfer on first use, and where that build fails the engine
+dies inside `profile_run` long after every import has succeeded, which reads as a model
+failure rather than an environment one. The PyTorch sampler it falls back to costs nothing
+here.
+
+### Quick start (zero-shot, default voice)
+```bash
+python examples/offline_inference/text_to_speech/gepard/end2end.py \
+    --text "Hello, this is a Gepard demo."
+```
+
+### Voice cloning
+Not yet — PR1 is zero-shot only (the learned `null_prefix` default voice). Reference-audio cloning (`ref_compressor` speaker prefix) is a follow-up PR.
+
+### Notes
+- Output: 22.05 kHz mono WAV.
+- First run downloads two checkpoints, not one: the model itself, and the NanoCodec decoder it names in `codec_id` (`nvidia/nemo-nano-codec-22khz-1.89kbps-21.5fps` by default). Both are fetched automatically; expect the first startup to be correspondingly slower.
+- Offline only for now; the `/v1/audio/speech` adapter is a follow-up PR.
+- Deploy config: `vllm_omni/deploy/gepard.yaml` (the example's default; copy it and pass `--deploy-config` to change it).
+- Generation length and reproducibility are stage settings, not script flags. One output token is one audio frame, so `max_tokens` in the YAML is the frame budget; `seed` makes the in-model 32-head sampling reproducible. The script deliberately passes no `SamplingParams`: one supplied by a caller replaces the stage defaults wholesale rather than merging, which would drop the pipeline's stop token and run every request to `max_tokens`.
+- Text length: short texts are repeated internally to match the training layout (the checkpoint's `text_repetition` block); the upper bound is the stage's `max_model_len`, enforced by the engine. Empty text is rejected rather than voiced.
+- `VLLM_GEPARD_GREEDY=1` swaps the 32-head Gumbel-max sampling for argmax, for reproducible comparisons that do not depend on a seed.
+
 ## VoxCPM2

 Single-stage native AR TTS at 48 kHz. Pipeline: `feat_encoder → MiniCPM4 → FSQ → residual_lm → LocDiT → AudioVAE`.
@@ -430,6 +484,31 @@ 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.
@@ -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()
@@ -38,17 +38,6 @@ from vllm_omni.utils.tracking_parser import TrackingArgumentParser
 logger = logging.getLogger(__name__)

 DEFAULT_MODEL = "fishaudio/s2-pro"
-DEFAULT_STAGE_CONFIG = os.path.join(
-    os.path.dirname(__file__),
-    "..",
-    "..",
-    "..",
-    "..",
-    "vllm_omni",
-    "model_executor",
-    "stage_configs",
-    "fish_speech_s2_pro.yaml",
-)


 def build_prompt(
@@ -154,8 +143,6 @@ def _build_omni_kwargs(args, model_name: str) -> dict:
    )
    if args.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 +265,6 @@ 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",
@@ -0,0 +1,139 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""Offline inference example for Gepard-1.0 TTS via vLLM-Omni.
+
+Single-stage native-AR pipeline: a Qwen3.5 backbone samples one 32-code FSQ
+frame per step; the NeMo NanoCodec decodes the frames to a 22.05 kHz mono
+waveform. Zero-shot: the default learned voice, no reference audio.
+
+Generation length and seed are stage settings rather than flags here. One
+output token is one audio frame, so ``max_tokens`` in the deploy YAML is the
+frame budget; pass a copy of the YAML via ``--deploy-config`` to change it.
+
+Usage:
+  python end2end.py --text "Hello, this is Gepard speaking."
+"""
+
+from __future__ import annotations
+
+import os
+from pathlib import Path
+
+import soundfile as sf
+import torch
+
+import vllm_omni
+from vllm_omni.utils.tracking_parser import TrackingArgumentParser
+
+os.environ.setdefault("VLLM_WORKER_MULTIPROC_METHOD", "spawn")
+
+from vllm_omni import Omni  # noqa: E402
+
+MODEL = "nineninesix/gepard-1.0"
+SAMPLE_RATE = 22050
+# Absolute: the loader hands the path straight to OmegaConf, which resolves a
+# bare name against the current working directory, not the package.
+DEFAULT_DEPLOY_CONFIG = str(Path(vllm_omni.__file__).resolve().parent / "deploy" / "gepard.yaml")
+
+
+def build_request(text: str) -> dict:
+    """Build an Omni request payload for Gepard (zero-shot, no ref audio).
+
+    The model only consumes the [speaker slots, SOT, text, EOT, SOS] layout, so
+    a bare text prompt would have no speaker slots and no SOS.
+    """
+    from transformers import AutoTokenizer
+
+    from vllm_omni.model_executor.models.gepard.configuration_gepard import GepardConfig
+    from vllm_omni.model_executor.models.gepard.prompt import build_gepard_prompt_ids
+
+    # From the checkpoint's own sidecar, which is what the worker builds its
+    # config from too. A layout that disagrees with the trained one does not
+    # fail, it just degrades the speech.
+    cfg = GepardConfig.from_checkpoint(MODEL)
+    tokenizer = AutoTokenizer.from_pretrained(MODEL, trust_remote_code=True)
+    prompt_token_ids = build_gepard_prompt_ids(
+        tokenizer(text, add_special_tokens=False)["input_ids"],
+        config=cfg,
+    )
+    return {
+        "prompt_token_ids": prompt_token_ids,
+        "additional_information": {"text": [text]},
+    }
+
+
+def save_audio(waveform: torch.Tensor, path: str, sample_rate: int = SAMPLE_RATE) -> None:
+    sf.write(path, waveform.float().numpy(), sample_rate)
+    seconds = waveform.numel() / sample_rate
+    print(
+        f"  Saved {path} ({tuple(waveform.shape)}, {sample_rate} Hz, "
+        f"{seconds:.2f}s, peak {float(waveform.abs().max()):.3f})"
+    )
+
+
+def main(args) -> None:
+    omni = Omni(
+        model=MODEL,
+        deploy_config=args.deploy_config,
+        stage_init_timeout=args.stage_init_timeout,
+        init_timeout=args.init_timeout,
+    )
+
+    # No explicit SamplingParams: one supplied here would replace the stage
+    # defaults rather than merge over them, dropping the pipeline's stop token.
+    output_dir = Path(args.output_dir)
+    output_dir.mkdir(parents=True, exist_ok=True)
+
+    print(f"Synthesizing: {args.text!r}")
+    inputs = build_request(text=args.text)
+
+    for i, stage_outputs in enumerate(omni.generate(inputs)):
+        for j, out in enumerate(stage_outputs.outputs):
+            mm = out.multimodal_output
+            if mm is None:
+                print(f"  [req {i}] No audio output.")
+                continue
+            # The consolidation path renames "model_outputs" to "audio".
+            audio = mm.get("audio")
+            if audio is None:
+                audio = mm.get("model_outputs")
+            if isinstance(audio, list):
+                audio = audio[0] if len(audio) else None
+            if audio is None:
+                print(f"  [req {i}] No waveform in multimodal_output (keys: {list(mm)}).")
+                continue
+            sr_tensor = mm.get("sr")
+            sr = int(sr_tensor.item()) if hasattr(sr_tensor, "item") else SAMPLE_RATE
+            save_audio(audio.cpu(), str(output_dir / f"output_{i}_{j}.wav"), sr)
+
+    print("Done.")
+
+
+def parse_args():
+    parser = TrackingArgumentParser(description="Gepard-1.0 offline TTS inference (zero-shot)")
+    parser.add_argument("--text", default="Hello, this is Gepard speaking.", help="Text to synthesize.")
+    parser.add_argument(
+        "--output-dir",
+        default=os.path.join(
+            os.environ.get("XDG_CACHE_HOME", os.path.join(os.path.expanduser("~"), ".cache")),
+            "gepard_output",
+        ),
+        help="Directory for WAV outputs (default: ~/.cache/gepard_output).",
+    )
+    parser.add_argument(
+        "--deploy-config",
+        default=DEFAULT_DEPLOY_CONFIG,
+        help=(
+            "Path to the deploy YAML (default: the packaged vllm_omni/deploy/gepard.yaml). This must "
+            "stay set: without the YAML's `pipeline: gepard` pin the architectures fallback routes "
+            "Qwen3_5ForCausalLM to the diffusion registry. Copy it to change max_tokens or seed."
+        ),
+    )
+    # The upstream defaults are too tight for a cold start.
+    parser.add_argument("--stage-init-timeout", type=int, default=900)
+    parser.add_argument("--init-timeout", type=int, default=1800)
+    return parser.parse_args()
+
+
+if __name__ == "__main__":
+    main(parse_args())
@@ -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,
    )
@@ -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",
@@ -32,11 +32,10 @@ from end2end import _estimate_prompt_len
 from vllm.utils.argparse_utils import FlexibleArgumentParser

 from vllm_omni import Omni
-from vllm_omni.engine.arg_utils import nullify_stage_engine_defaults
 from vllm_omni.utils.forced_aligner import align, build_forced_aligner_config


-def _default_stage_config() -> str:
+def _default_deploy_config() -> str:
    repo_root = Path(__file__).resolve().parents[4]
    return str(repo_root / "vllm_omni" / "deploy" / "qwen3_tts.yaml")

@@ -109,7 +108,7 @@ def main(args: Any) -> None:
        "output_dir",
    ):
        omni_kwargs.pop(key, None)
-    omni_kwargs["stage_configs_path"] = args.stage_configs_path or _default_stage_config()
+    omni_kwargs["deploy_config"] = args.deploy_config or _default_deploy_config()
    omni_kwargs["log_stats"] = args.log_stats
    omni = Omni(**omni_kwargs)

@@ -156,7 +155,7 @@ def parse_args() -> Any:
        default=None,
        help="Optional YAML file for forced aligner settings (incl. gpu_memory_utilization)",
    )
-    parser.add_argument("--stage-configs-path", default=None, help="Qwen3-TTS deploy YAML")
+    parser.add_argument("--deploy-config", default=None, help="Qwen3-TTS deploy YAML")
    parser.add_argument("--text", default="Hello world.", help="Text to synthesize and align")
    parser.add_argument("--language", default="English", help="Qwen3-TTS language field")
    parser.add_argument("--speaker", default="Vivian", help="CustomVoice speaker name")
@@ -164,7 +163,6 @@ def parse_args() -> Any:
    parser.add_argument("--max-new-tokens", type=int, default=2048, help="TTS max_new_tokens")
    parser.add_argument("--output-dir", default="output_audio", help="Directory for WAV and JSON sidecar")
    parser.add_argument("--log-stats", action="store_true", default=False, help="Enable vLLM-Omni stats logging")
-    nullify_stage_engine_defaults(parser)
    return parser.parse_args()


@@ -570,6 +570,8 @@ def main():
        num_frames=args.num_frames,
        extra_args=extra_args,
    )
+    if args.frame_rate is not None:
+        sampling_kwargs["frame_rate"] = args.frame_rate
    if args.guidance_scale_high is not None:
        sampling_kwargs["guidance_scale_2"] = args.guidance_scale_high

@@ -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
@@ -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",
@@ -45,7 +45,7 @@ MusicGen-style delay pattern).
 **Online serving:**

 ```bash
-vllm-omni serve bosonai/higgs-audio-v3-tts-4b \
+vllm serve bosonai/higgs-audio-v3-tts-4b \
    --host 0.0.0.0 --port 8095 \
    --trust-remote-code --omni
 ```
@@ -0,0 +1,175 @@
+# LTX-2.5
+
+> Text-to-video and first-frame image-to-video generation with synchronized audio
+
+## Overview
+
+vLLM-Omni supports the gated
+[`Lightricks/LTX-2.5-Diffusers`](https://huggingface.co/Lightricks/LTX-2.5-Diffusers)
+checkpoint for T2V and first-frame I2V. Every output includes synchronized
+48 kHz stereo audio.
+
+The raw [`Lightricks/LTX-2.5`](https://huggingface.co/Lightricks/LTX-2.5)
+repository is not directly loadable with `--model`. It supplies the official
+upsampler and LoRA sidecars used by the Full/SFT two-stage pipeline, so accept
+both model licenses and authenticate before first use.
+
+## Choose a pipeline
+
+| Pipeline | Mode | Default size | Schedule |
+|---|---|---:|---:|
+| `LTX2Pipeline` | Full/SFT one-stage | 960x544 | 30 steps |
+| `LTX2TwoStagePipeline` | Full/SFT two-stage | 1920x1088 | 30 + 3 steps |
+| `LTX2DistilledOneStagePipeline` | Distilled one-stage | 960x544 | 8 steps |
+| `LTX2DistilledTwoStagePipeline` | Distilled two-stage | 1920x1088 | 8 + 3 steps |
+
+Both two-stage pipelines generate at half resolution, apply the official x2
+latent upsampler, and run a three-step refinement stage. Select the class with
+`--model-class-name`; no `--task-type` flag is required. Supplying one initial
+image selects I2V, while omitting it selects T2V.
+
+## Prerequisites
+
+```bash
+hf auth login
+export MODEL=Lightricks/LTX-2.5-Diffusers
+```
+
+Install matching vLLM and vLLM-Omni versions, and ensure `ffmpeg` and
+`ffprobe` are on `PATH`. I2V requires PyAV backed by an FFmpeg build with the
+`libx264` encoder.
+
+## Hardware
+
+| GPU | Status | Recommended scope |
+|---|---|---|
+| NVIDIA B300 | Verified | All four canonical pipelines |
+| NVIDIA B200 or H200 | Capacity-based recommendation; not yet verified | All four canonical pipelines |
+| NVIDIA GB200 or GB300 | Capacity-based recommendation; not yet verified | All four canonical pipelines |
+| NVIDIA H100 80 GB | FP8 recipe | Distilled one-stage at 960x544 |
+
+The 1920x1088 two-stage examples require about 114 GB of peak GPU memory.
+80 GB GPUs do not have enough safety margin for the canonical two-stage
+configuration; use a larger GPU or reduce the output configuration.
+
+### H100 80 GB FP8
+
+Use the distilled one-stage pipeline with FP8 and cuDNN attention:
+
+```bash
+vllm serve Lightricks/LTX-2.5-Diffusers \
+  --omni \
+  --model-class-name LTX2DistilledOneStagePipeline \
+  --quantization fp8 \
+  --diffusion-attention-backend CUDNN_ATTN \
+  --host 0.0.0.0 \
+  --port 8000
+```
+
+Use the T2V or I2V request below after the server is ready.
+
+## Offline inference
+
+Choose values from the pipeline table. For example, the distilled two-stage
+path uses:
+
+```bash
+export PIPELINE=LTX2DistilledTwoStagePipeline
+export WIDTH=1920
+export HEIGHT=1088
+export STEPS=8
+```
+
+### Text-to-video
+
+```bash
+python examples/offline_inference/text_to_video/text_to_video.py \
+  --model "${MODEL}" \
+  --model-class-name "${PIPELINE}" \
+  --prompt "A cinematic shot of a red fox walking through a snowy forest at dawn, the camera tracking alongside, snow crunching underfoot." \
+  --width "${WIDTH}" \
+  --height "${HEIGHT}" \
+  --num-frames 121 \
+  --num-inference-steps "${STEPS}" \
+  --frame-rate 24 \
+  --fps 24 \
+  --seed 42 \
+  --output ltx25-t2v.mp4
+```
+
+### First-frame image-to-video
+
+```bash
+python examples/offline_inference/image_to_video/image_to_video.py \
+  --model "${MODEL}" \
+  --model-class-name "${PIPELINE}" \
+  --image /absolute/path/to/first-frame.png \
+  --prompt "The red fox walks forward while the camera tracks alongside." \
+  --width "${WIDTH}" \
+  --height "${HEIGHT}" \
+  --num-frames 121 \
+  --num-inference-steps "${STEPS}" \
+  --frame-rate 24 \
+  --fps 24 \
+  --seed 42 \
+  --output ltx25-i2v.mp4
+```
+
+LTX-2.5 uses the official CRF-18 first-frame conditioning path by default.
+
+## Online serving
+
+Start one server for the selected pipeline:
+
+```bash
+vllm serve "${MODEL}" \
+  --omni \
+  --model-class-name "${PIPELINE}" \
+  --host 0.0.0.0 \
+  --port 8000 \
+  --stage-init-timeout 900
+```
+
+T2V request:
+
+```bash
+curl -sS --fail-with-body \
+  -X POST http://127.0.0.1:8000/v1/videos/sync \
+  -F 'prompt=A cinematic shot of a red fox walking through a snowy forest at dawn, the camera tracking alongside, snow crunching underfoot.' \
+  -o ltx25-online-t2v.mp4
+```
+
+For I2V, add exactly one first frame:
+
+```bash
+export FIRST_FRAME=/absolute/path/to/first-frame.png
+
+curl -sS --fail-with-body \
+  -X POST http://127.0.0.1:8000/v1/videos/sync \
+  -F "input_reference=@${FIRST_FRAME};type=image/png" \
+  -F 'prompt=The red fox walks forward while the camera tracks alongside.' \
+  -o ltx25-online-i2v.mp4
+```
+
+Restart the server after changing `PIPELINE` because each class selects a
+different weight layout and execution topology.
+
+## Constraints
+
+- Online `width`, `height`, `num_frames`, `fps`, and `num_inference_steps`
+  are optional and use the selected pipeline defaults when omitted. `seed` is
+  optional and random when omitted.
+- `num_frames` must be `8k+1` when overridden.
+- One-stage width and height must be divisible by 32; two-stage final
+  dimensions must be divisible by 64.
+- Full/SFT pipelines accept negative prompts. Distilled pipelines are
+  positive-only and reject negative prompts.
+- I2V accepts exactly one initial image per prompt.
+- The supported `--model` value is `Lightricks/LTX-2.5-Diffusers`.
+
+## References
+
+- [LTX-2.5-Diffusers model card](https://huggingface.co/Lightricks/LTX-2.5-Diffusers)
+- [Official LTX-2 repository](https://github.com/Lightricks/LTX-2)
+- [Generic T2V example](../../examples/offline_inference/text_to_video/text_to_video.py)
+- [Generic I2V example](../../examples/offline_inference/image_to_video/image_to_video.py)
@@ -4,11 +4,22 @@

 ## Pipelines

-| Model | Checkpoint | Task | `--model-class-name` | Batching |
-|---|---|---|---|---|
-| LTX-2 | `Lightricks/LTX-2` | One-stage T2V/I2V | `LTX2Pipeline` | Yes |
-| LTX-2 distilled | `rootonchair/LTX-2-19b-distilled` | Two-stage T2V/I2V | `LTX2DistilledPipeline` | No |
-| LTX-2.3 | `diffusers/LTX-2.3-Diffusers` | One-stage T2V/I2V | `LTX2Pipeline` | Yes |
+| `--model-class-name` | Task | Required checkpoint repositories |
+|---|---|---|
+| `LTX2Pipeline` | LTX-2 one-stage T2V/I2V | `Lightricks/LTX-2` |
+| `LTX2TwoStagePipeline` | LTX-2 ordinary two-stage T2V/I2V | `Lightricks/LTX-2` |
+| `LTX2DistilledOneStagePipeline` | LTX-2 merged-distilled one-stage T2V/I2V | `rootonchair/LTX-2-19b-distilled` |
+| `LTX2DistilledTwoStagePipeline` | LTX-2 merged-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` |
+| `LTX2DistilledOneStagePipeline` | LTX-2.3 merged-distilled one-stage T2V/I2V | `diffusers/LTX-2.3-Distilled-Diffusers` |
+| `LTX2DistilledTwoStagePipeline` | LTX-2.3 merged-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 +27,13 @@ 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
-or provide one initial image for I2V. As in the official LTX pipelines,
-distilled mode is selected explicitly rather than inferred from the checkpoint
-name or metadata; always pass `--model-class-name LTX2DistilledPipeline`.
+`LTX2TwoStagePipeline` samples the regular model at half resolution, upsamples,
+then refines with the distilled LoRA. `LTX2DistilledOneStagePipeline` uses a
+merged distilled Transformer without upsampling, while
+`LTX2DistilledTwoStagePipeline` uses it in both stages. All entries support
+T2V and I2V; select their class explicitly. The deprecated
+`LTX2DistilledPipeline` name remains an alias for
+`LTX2DistilledTwoStagePipeline`.

 ## API Migration

@@ -43,8 +57,8 @@ The consolidation also removes these registry names without aliases:
 | `LTX23Pipeline` | `LTX2Pipeline`; checkpoint metadata selects LTX-2.3 |
 | `LTX2ImageToVideoPipeline` | `LTX2Pipeline` with `image=` |
 | `LTX23ImageToVideoPipeline` | `LTX2Pipeline` with `image=`; checkpoint metadata selects LTX-2.3 |
-| `LTX2TwoStagesPipeline` | `LTX2DistilledPipeline` |
-| `LTX2ImageToVideoTwoStagesPipeline` | `LTX2DistilledPipeline` with `image=` |
+| `LTX2TwoStagesPipeline` | `LTX2DistilledTwoStagePipeline` |
+| `LTX2ImageToVideoTwoStagesPipeline` | `LTX2DistilledTwoStagePipeline` with `image=` |

 Passing any second positional argument now raises `TypeError`. These changes
 affect direct Python callers and explicit `--model-class-name` overrides;
@@ -68,24 +82,28 @@ 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 |
-|---|---:|
-| Final width × height | 1536 × 1024 |
-| Stage 1 width × height | 768 × 512 |
-| Frames / frame rate | 121 / 24 |
-| Stage 1 / Stage 2 denoise steps | 8 / 3 |
-| Guidance | Fixed positive-only |
+| Parameter | Ordinary | Full-distilled |
+|---|---:|---:|
+| Final width × height | 1536 × 1024 | 1536 × 1024 |
+| Stage 1 width × height | 768 × 512 | 768 × 512 |
+| Frames / frame rate | 121 / 24 | 121 / 24 |
+| Stage 1 / Stage 2 steps | 40 (LTX-2) or 30 (LTX-2.3) / 3 | 8 / 3 |
+| 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
-divisible by 64. Both one- and two-stage requests require `num_frames = 8k+1`.
-The distilled recipe fixes both sigma schedules; `num_inference_steps` must be
-`8` when supplied.
+API dimensions are final dimensions and must be divisible by 64. All LTX
+requests require `num_frames = 8k+1`. Ordinary Stage 1 uses the LTX-2 or
+LTX-2.3 one-stage defaults shown above. Distilled schedules are fixed, so
+`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 +113,19 @@ 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
+  --model-class-name LTX2DistilledTwoStagePipeline --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 LTX2DistilledTwoStagePipeline --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 +161,9 @@ together with `input_reference`.

 ## Guidance

-LTX supports independent video/audio CFG, spatio-temporal guidance (STG),
-cross-modality guidance, and rescaling:
+One-stage and ordinary Stage 1 support independent video/audio CFG,
+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 +199,9 @@ 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:
-`cond`, `uncond`, `ptb` (STG), and `mod` (cross-modality). The useful balanced
-configurations are therefore:
+The default one-stage recipe and ordinary Stage 1 have four Transformer passes
+per denoise step: `cond`, `uncond`, `ptb` (STG), and `mod`
+(cross-modality). The useful balanced configurations are therefore:

 | `--cfg-parallel-size` | Passes per rank | Guidance-slot utilization | Notes |
 |---:|---:|---:|---|
@@ -195,9 +222,8 @@ 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
-guidance, so `cfg_parallel_size > 1` does not provide guidance speedup for
-them.
+configured parallel dimensions. Positive-only distilled phases do not benefit
+from `cfg_parallel_size > 1`.

 ### Python API

@@ -242,22 +268,22 @@ 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. |
-| `audio_latents` | tensor, `None` | One-stage initial audio latents; request tensors win and are collated. |
+| `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 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 +299,13 @@ request prompt payload; LTX guidance fields live in sampling `extra_args`.

 ### Recipe-Specific Request Capabilities

-| Override | One-stage | Distilled two-stage |
-|---|---|---|
-| Guidance | Supported | Fixed positive-only |
-| Negative prompt/embeddings | Supported | Rejected |
-| `num_inference_steps` | Supported | Fixed at 8 for Stage 1; Stage 2 uses 3 |
-| Custom `sigmas` | Supported | Rejected; both phases use fixed schedules |
-| Video/audio latents | Supported | Rejected |
+| Override | One-stage | Ordinary two-stage | Distilled two-stage |
+|---|---|---|---|
+| Guidance | Supported | Stage 1 only; Stage 2 is positive-only | Fixed positive-only |
+| Negative prompt/embeddings | Supported | Supported by Stage 1 | Rejected |
+| `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 | Rejected; both phases use fixed schedules |
+| 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 +335,10 @@ 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 +353,15 @@ 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
-  LTX two-stage/HQ execution remains out of scope.
+- Ordinary and full-distilled two-stage T2V/I2V are supported; HQ execution
+  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)
@@ -0,0 +1,195 @@
+# MiniMax-H3 on a single NPU 950PR
+
+This recipe runs MiniMax-H3 with online INT8 quantization on one NPU 950PR
+NPU (128 GB HiBL 1.0 HBM). It covers the single-card T2VA configuration at
+1024x576. For the eight-card Atlas 800I A3 BF16 route at 768P, see
+[MiniMax-H3-NPU.md](MiniMax-H3-NPU.md).
+
+## Capacity requirements
+
+| Resource | Requirement |
+| --- | ---: |
+| NPU | 1x NPU 950PR |
+| NPU HBM | 128 GiB (131,072 MiB reported by `npu-smi`) |
+| Observed HBM high-water mark | 118,442 MiB (90.4% of capacity) |
+| Checkpoint storage | 135 GiB per partition, local disk strongly preferred |
+| Container shared memory | 8 GiB, or a writable local filesystem for spill |
+
+The observed high-water mark leaves about 12.3 GiB of headroom, so this card is
+sized for exactly one resident task partition at 1024x576. Do not expect a
+larger output shape or concurrency greater than one to fit.
+
+Because a single 950PR holds the whole partition in HBM, this route does not
+need layerwise offload, and the 200 GiB system-RAM floor carried by the 72 GiB
+GPU recipes does not apply. The validated run completed inside a container
+limited to 32 GiB of RAM. Passing `--enable-layerwise-offload` on this card is
+actively harmful: it stages weights in non-reclaimable host memory and the
+container's OOM killer terminates the server with exit code -9.
+
+
+## Environment
+
+- Host architecture: x86_64
+-  driver: 25.7.rc1.6 (hal 7.35.23)
+-  firmware: 9.0.0.105.229
+- CANN toolkit: 9.1.0 (`/usr/local//cann-9.1.0`)
+- npu-smi: 25.7.rc1.6
+- Python: 3.12.13
+- PyTorch: 2.10.0+cpu
+- torch_npu: 2.10.0.post2
+- vLLM: 0.26.0
+- vLLM-Omni: 0.26.1.dev103+g584d78c67.npu (commit `584d78c6`)
+
+Install vLLM-Omni from a checkout with MiniMax-H3 support:
+
+```bash
+uv venv
+source .venv/bin/activate
+uv pip install -e .
+```
+
+## Start a server
+
+```bash
+export MODEL=/path/to/MiniMax-H3/FL2VA
+export PORT=8000
+export VLLM_WORKER_MULTIPROC_METHOD=spawn
+export VLLM_OMNI_VIDEO_SYNC_TIMEOUT=14400
+
+vllm serve "${MODEL}" \
+  --omni \
+  --host 0.0.0.0 \
+  --port "${PORT}" \
+  --trust-remote-code \
+  --num-gpus 1 \
+  --tensor-parallel-size 1 \
+  --usp 1 \
+  --ring 1 \
+  --text-encoder-tp-size 1 \
+  --vae-patch-parallel-size 1 \
+  --vae-parallel-mode tile \
+  --vae-use-tiling \
+  --init-timeout 14400 \
+  --stage-init-timeout 14400 \
+  --quantization int8
+```
+
+Do not pass `--diffusion-attention-backend CUDNN_ATTN`. cuDNN attention has no
+NPU implementation and fails at the first denoise step. Leave the backend
+unset: it resolves to `TORCH_SDPA`, or to `FLASH_ATTN` when MindIE-SD is
+installed (see [MiniMax-H3-NPU.md § Environment](MiniMax-H3-NPU.md#environment)).
+
+H3 is CFG-distilled, so `--cfg-parallel-size` must remain 1.
+
+## Validated evidence
+
+Measured on one NPU 950PR with the server command above plus
+`--enable-diffusion-pipeline-profiler`, generating a 5 s 1024x576 T2VA clip from
+the request in [§ T2VA request example](#t2va-request-example). MiniMax-H3
+requested 60 denoise steps and executed 59 denoise updates, so per-step latency
+is `denoise / 59`. `ffprobe` confirms every returned MP4 is 1024x576, 24 fps,
+124 video frames, 5.175 s, with a 32 kHz audio track of 165,600 samples.
+
+Five requests were issued back to back with no idle time between them. The card
+is measurably faster on the first request than in steady state, so both regimes
+are reported.
+
+| Stage | First request, cold card | Thermal steady state |
+| --- | ---: | ---: |
+| End-to-end request | 462.13 s | 506.38 s |
+| Text encode | 0.55 s | 0.05 s |
+| Denoise (59 updates) | 429.29 s | 471.74 s |
+| Per denoise update | 7,276 ms | 7,996 ms |
+| VAE decode | 8.43 s | 8.14 s |
+| Worker-to-server handoff | 24 s | 27 s |
+| MP4 muxing | 0.57 s | 0.50 s |
+| Server-reported `denoise_step_latency_ms` | 7,702 ms | 8,440 ms |
+| Peak HBM (`npu-smi`) | 118,442 MiB | 118,442 MiB |
+| Average NPU power | 566.9 W | 556.9 W |
+| Peak NPU temperature | 104 C | 104 C |
+
+Sample counts differ per row. The cold column is a single request. In the steady
+column, text encode, denoise and VAE decode are means over the four subsequent
+requests — denoise spanned 470.37 to 473.21 s, a 0.60% spread — while
+end-to-end, handoff and muxing are means over the two of those four that
+survived the handoff timeout described above.
+
+**Report the steady-state column.** The cold card completes denoise 9.9% faster
+while drawing only 1.8% more average power, and both regimes reach the same
+104 C ceiling, so the first request is buying a short window of higher clocks
+before the die saturates. Any benchmark that issues a single request against an
+idle card will overstate throughput by about 10%.
+
+**Do not read `denoise_step_latency_ms` as a per-step time.** The server divides
+the entire stage wall time by the 60 *requested* steps, so it absorbs text
+encoding, VAE decode and the 837 MiB handoff. All three successful requests
+satisfy `denoise_step_latency_ms == e2e_stage_wall_time_ms / 60` exactly. The
+per-update figure in the table is `diffuse / 59` from the pipeline profiler.
+
+Denoise accounts for 93% of the steady-state request and the handoff accounts
+for 5.3%, so keeping the transfer in shared memory should bring end-to-end down
+to roughly 480 s.
+
+Peak HBM is sampled externally with `npu-smi` at a 3 s interval. It is a
+reserved high-water mark rather than live allocation: it reads 111,272 MiB after
+weight loading and before the first request, and does not fall back after
+requests complete. Re-measure for longer outputs, a different output shape, or
+concurrency greater than one.
+
+### Startup
+
+Weight loading took **31 min 50 s** on the validated host (first log line
+08:04:59, `Application startup complete` 08:36:49), with the checkpoint on a
+shared GlusterFS network mount. This is why the server command sets
+`--init-timeout 14400` and `--stage-init-timeout 14400`; the stock 600 s and
+1800 s timeouts both abort mid-load with `TimeoutError` and exit code 143.
+Reduce both timeouts when the partition is staged on local disk.
+
+## T2VA request example
+
+```bash
+export API_URL="http://127.0.0.1:${PORT}/v1/videos/sync"
+
+curl -sS --max-time 1800 -X POST "${API_URL}" \
+  -F 'prompt=At night, three cats march into a bedroom playing tiny brass instruments, then abruptly file out, with synchronized room ambience.' \
+  -F 'width=1024' \
+  -F 'height=576' \
+  -F 'aspect_ratio=16:9' \
+  -F 'fps=24' \
+  -F 'num_inference_steps=60' \
+  -F 'flow_shift=12' \
+  -F 'seed=1101' \
+  -F 'extra_params={"task":"t2va","duration":5,"audio_flow_shift":3.0}' \
+  -o t2va.mp4
+```
+
+## Known limitations
+
+- INT8 online quantization is validated for T2VA on this card. Use BF16 for
+  FL2VA and Ref2VA, or re-measure before relying on it.
+- Single-card serving loads one task partition at a time. For Ref2VA, stop the
+  server and restart it against the `Ref2VA` directory.
+- Sustained load drives the die to 104 C and costs about 10% of denoise
+  throughput relative to a cold card. Warm the server with at least one full
+  request before measuring, and treat steady state as the reportable number.
+- Requests can fail during result handoff on hosts where the IPC spill path is
+  slow. See [§ Result handoff and shared
+  memory](#result-handoff-and-shared-memory).
+- Loading the checkpoint from a network mount takes over 30 minutes and forces
+  very large init timeouts. Stage the partition on local disk when possible.
+- The configuration measured here is 1024x576 at 60 steps, which differs from
+  the 1344x768 at 50 steps used by the GPU recipes in this directory. Denoise
+  cost scales superlinearly with token count, so the numbers are not directly
+  comparable across recipes.
+- The image ships a `triton` package whose  backend is not built
+  (`No module named 'triton._C.libtriton.'`). vLLM logs this as an error
+  at startup and disables Triton; the diffusion path does not need it and the
+  run completes normally.
+
+## Additional resources
+
+- [MiniMax-H3.md](MiniMax-H3.md) — full GPU guide
+- [MiniMax-H3-NPU.md](MiniMax-H3-NPU.md) — eight-card Atlas 800I A3 BF16 guide
+- [Int8 quantization](../../docs/user_guide/quantization/int8.md)
+- [Supported models](../../docs/models/supported_models.md)
+- [Video API](../../docs/serving/videos_api.md)
@@ -357,12 +357,19 @@ No restart is needed: `task=fl2va` routes to `FL2VA/transformer`, while

 ### Online FP8 quantization

-MiniMax H3 supports load-time FP8 quantization of the DiT. The checkpoint
-remains BF16 on disk; vLLM-Omni quantizes eligible weights while loading and
-uses dynamic activation scaling during inference. By default, attention and
-MLP linears in the token refiner and main DiT blocks, the condition
-projection, and all AdaLN projections use FP8. Patch, timestep, and final
-projections remain FP32; the text encoder and VAEs are unchanged.
+MiniMax H3 supports online FP8 quantization of both the DiT and the Qwen3-VL
+text decoder. The checkpoint remains BF16 on disk; vLLM-Omni creates FP8
+weights at runtime and uses dynamic activation scaling during inference. By
+default, `--quantization fp8` quantizes eligible attention and MLP linears in
+the text decoder, token refiner, and main DiT blocks, as well as the DiT
+condition and AdaLN projections. The Qwen vision tower, embeddings, norms,
+RoPE, both VAEs, and the model's FP32 patch, timestep, and output projections
+keep checkpoint precision.
+
+To select a component, use `--diffusion-quantization-config` with
+`{"transformer":{"method":"fp8"}}` for DiT-only FP8 or
+`{"text_encoder":{"method":"fp8"}}` for text-decoder-only FP8. The two
+entries can be combined. The shorthand below enables both components.

 Add this option to an existing H3 server command:

@@ -370,6 +377,28 @@ Add this option to an existing H3 server command:
 --quantization fp8
 ```

+#### Single 96 GB GPU, no-offload capacity check
+
+Use the FL2VA-only partition for this capacity test. Loading the combined
+service would also load the Ref2VA DiT and would test a different memory
+budget. A no-offload capacity check should contain none of
+`--enable-cpu-offload`, `--enable-layerwise-offload`, or
+`--enable-distributed-layerwise-offload`. VAE tiling changes decode placement
+but does not offload model weights to the CPU.
+
+The run passes the capacity check when the server initializes, the request
+finishes without CUDA OOM or Xid errors, `peak_used_mib` remains below the
+card's reported `memory_total_mib`, and `ffprobe` reports H.264 video plus
+32 kHz stereo AAC audio. Report the measured headroom rather than assuming
+that every nominal 96 GB SKU exposes the same MiB total.
+
+As a capacity proxy only, the same five-second case on one B300 measured a
+92,946 MiB whole-device first-request peak and a 92,146 MiB worker peak. Its
+encode, diffuse, decode, and client wall times were 8.664 s, 134.504 s,
+6.016 s, and 151.583 s. These numbers suggest that a 96 GB card may fit, but
+they are not an RTX PRO 6000 validation: kernels, allocator behavior, and
+reported device capacity differ across GPUs.
+
 Use `ignored_layers` to keep any otherwise eligible linear in BF16. H3
 resolves the `transformer` component before constructing the DiT, so names do
 not start with `transformer.`. Entries are exact runtime linear prefixes; a
@@ -386,10 +415,10 @@ For example, keep the first main block's attention projections in BF16 with:
  '{"transformer":{"method":"fp8","ignored_layers":["blocks.0.attn.qkv_proj","blocks.0.attn.out_proj"]}}'
 ```

-The structured option replaces `--quantization fp8`. Online FP8 is currently
-incompatible with H3 layerwise offload because the offload path produces a
-weight stride rejected by the Cutlass FP8 kernel. Use resident FP8 with tensor
-parallelism and VAE tiling instead.
+The structured option replaces `--quantization fp8`. Online FP8 can be used
+with H3 layerwise offload and with distributed layerwise offload's full-weight
+per-rank path (`--dlo-no-use-allgather`). The sharded DLO AllGather path is not
+supported for runtime-created FP8 weights.

 ## AMD ROCm (gfx942 / gfx950)

@@ -750,6 +779,13 @@ audio spectral cosine 0.9589 (minimum 0.80), and audio RMS ratio 0.9342. The
 resident per-GPU peak was 68.52 GiB for BF16 and 53.51 GiB for FP8, a 22%
 reduction.

+For direct human inspection, see the external
+[BF16 versus global-FP8 comparison](https://lishunyang12.github.io/vllm-omni-rankings/scripts/minimax_h3_global_fp8_vs_bf16/),
+which includes matched five-second T2VA, I2VA, and Ref2VA videos. The
+[comparison sources](https://github.com/lishunyang12/vllm-omni-rankings/tree/main/scripts/minimax_h3_global_fp8_vs_bf16)
+also record per-task fidelity metrics and provenance without storing generated
+media in this repository.
+
 ## TeaCache acceleration

 TeaCache reuses DiT block residuals across denoising steps when consecutive
@@ -834,7 +870,8 @@ vllm serve "${MODEL_ROOT}/FL2VA" \
  H3 native `tile` mode only.
 - A U2 x Ring2 hybrid currently fails with an attention-mask length mismatch; use
  pure Ulysses.
- Online FP8 is not compatible with layerwise offload.
+- Runtime-created FP8 weights do not support the sharded DLO AllGather path;
+  use `--dlo-no-use-allgather` when combining online FP8 with DLO.
 - TeaCache and Cache-DiT cannot be enabled on the same server.
 - Image+audio Ref2VA accepts exactly one image and one audio reference.
 - Video Ref2VA accepts one or more video files, but not an additional standalone
@@ -51,7 +51,7 @@ NemotronH (~3B active parameters). Pick the pipeline by task:
 `model_type` auto-resolves `audex_tts.yaml`):

 ```bash
-vllm-omni serve nvidia/Nemotron-Labs-Audex-2B \
+vllm serve nvidia/Nemotron-Labs-Audex-2B \
    --host 0.0.0.0 --port 8097 \
    --trust-remote-code --omni
 # other modes: add --stage-configs-path vllm_omni/deploy/audex_{tta,thinker_only,s2s}.yaml
@@ -109,7 +109,7 @@ The 30B REQUIRES its explicit deploy yaml (the shared HF `model_type`
 auto-resolves to the 2B-tuned config otherwise):

 ```bash
-vllm-omni serve nvidia/Nemotron-Labs-Audex-30B-A3B \
+vllm serve nvidia/Nemotron-Labs-Audex-30B-A3B \
    --host 0.0.0.0 --port 8097 \
    --trust-remote-code --omni \
    --stage-configs-path vllm_omni/deploy/audex_tts_30b.yaml
@@ -55,10 +55,12 @@ recipes/
 | [`MiniMaxAI/MiniMax-H3-MUSA.md`](./MiniMaxAI/MiniMax-H3-MUSA.md) | MiniMax H3 T2VA, FL2VA, and Ref2VA on MUSA | 4x MTT S5000 (TP4 Ref2VA profile) |
 | [`krea/Krea-2.md`](./krea/Krea-2.md) | Text-to-image (Turbo + Raw), offline + online, with LoRA | 1x H100 80GB |
 | [`LTX/LTX-2.md`](./LTX/LTX-2.md) | LTX-2/LTX-2.3 text-to-video and image-to-video with synchronized audio | H200 141GB / 96GB-class GPU |
+| [`LTX/LTX-2.5.md`](./LTX/LTX-2.5.md) | LTX-2.5-Diffusers: Full/SFT one-stage and distilled two-stage T2V/I2V with synchronized audio | NVIDIA B300; cuDNN-qualified |
 | [`MammothModa2/MammothModa2.md`](./MammothModa2/MammothModa2.md) | Preview and Dev text-to-image (AR → DiT); Dev text/image understanding | Preview: 1x L40S 48GB / 1x ≥40GB GPU; Dev: 1x NVIDIA GPU with sufficient cache headroom |
 | [`meituan-longcat/LongCat-Video-Avatar-1.5.md`](./meituan-longcat/LongCat-Video-Avatar-1.5.md) | Audio-driven avatar video generation (AT2V / AI2V, single- and multi-speaker, AVC continuation) | 1x H100 80GB |
 | [`mistralai/Voxtral-TTS.md`](./mistralai/Voxtral-TTS.md) | Online serving for TTS | 1x RTX 4090 24GB |
 | [`Robbyant/LingBot-Video.md`](./Robbyant/LingBot-Video.md) | Native dense and MoE text-to-video serving | Dense: 1x L20X; MoE: 1x L20X (~67.7 GiB peak) |
+| [`Robbyant/LingBot-World-2.0.md`](./Robbyant/LingBot-World-2.0.md) | Offline and experimental realtime interactive world generation | 1x H200/B200 or 2x B200 for TP=2 |
 | [`cosmos3/Cosmos3-Nano.md`](./cosmos3/Cosmos3-Nano.md) | Text-to-image, text-to-video, image-to-video, video-to-video generation, text to video with sound, action policy | 1x H200 141GB / B300 |
 | [`cosmos3/Cosmos3-Edge.md`](./cosmos3/Cosmos3-Edge.md) | T2I / T2V / I2V + action (Physical-AI) generation on the Nemotron-based Edge transformer | 1x GPU (~8 GiB) |
 | [`cosmos3/Cosmos3-Super.md`](./cosmos3/Cosmos3-Super.md) | 64B T2I / T2V / I2V / V2V generation (+ optional audio) / Action policy | 8x H200/H100/A100 / 2x H200 / B300 |
@@ -0,0 +1,122 @@
+# LingBot-World 2.0
+
+> Offline and experimental realtime interactive world generation
+
+## Summary
+
+- Vendor: Robbyant
+- Model: `robbyant/lingbot-world-v2-14b-causal-fast-diffusers`
+- Task: image-conditioned interactive world generation
+- Modes: offline trajectory replay and in-process realtime AR-Diffusion ticks
+- Hardware validated: NVIDIA H200 and B200
+- Maintainer: Community
+
+The checkpoint is separately licensed under CC BY-NC-SA and restricted to
+non-commercial use. The vLLM-Omni integration code remains Apache-2.0.
+
+## Offline generation
+
+The offline path consumes one source image and an action directory containing
+`poses.npy` with shape `[frames, 4, 4]` and `intrinsics.npy` with shape
+`[frames, 4]`.
+
+```bash
+python examples/offline_inference/diffusion/lingbot_world_v2.py \
+  --prompt "The camera moves slowly forward through the scene." \
+  --image /path/to/first_frame.png \
+  --action-dir /path/to/actions/forward \
+  --num-frames 81 \
+  --output lingbot_world_v2.mp4
+```
+
+The current causal-fast checkpoint uses four DMD sampling steps. Raw frame
+counts must be `9 + 12k`, up to 117 frames. Camera files may contain a
+longer bounded source trajectory (the official example contains 269 frames);
+the runtime consumes only the prefix needed by the request.
+
+## Realtime in-process generation
+
+The realtime example keeps AR-Diffusion KV and model-owned state across
+requests. Each JSONL line describes the prompt and/or three latent-frame camera
+actions applied at the next chunk boundary:
+
+```json
+{"event_id":1,"prompt":"A road through a forest","frames":[["j"],[],[]]}
+{"event_id":2,"frames":[["w"],["w"],["w"]]}
+{"event_id":3,"prompt":"The road enters a snowy valley","frames":[[],[],[]]}
+```
+
+Run:
+
+```bash
+python examples/offline_inference/diffusion/lingbot_world_v2_realtime.py \
+  --image /path/to/first_frame.png \
+  --events /path/to/events.jsonl \
+  --output-dir /tmp/lingbot-realtime \
+  --gpu-memory-fraction 0.6
+```
+
+The example writes one latent tensor and one metadata JSON file per chunk. It
+exercises the same `ARDiffusionSessionManager -> ARDiffusionOmniTickConsumer
+-> AsyncOmni -> ARDiffusionEngine` path used by a future HTTP or WebSocket
+transport.
+
+This PR intentionally does not define a public realtime HTTP/WebSocket schema.
+An online serving client should be added together with the public transport
+API rather than exposing LingBot-specific event fields from the generic model
+runtime. The structured camera-interaction frontend and transport work is
+tracked separately in
+[vllm-project/vllm-omni#5527](https://github.com/vllm-project/vllm-omni/pull/5527).
+
+## Realtime identity and controls
+
+- `session_id` identifies the persistent world and its worker-owned state.
+- `event_id` identifies a prompt/control update and remains monotonic across reset.
+- `chunk_index` is contiguous from zero and restarts from zero after reset.
+- `request_id` correlates one chunk snapshot with its output metadata.
+- AsyncOmni uses a separate UUID-suffixed internal engine routing ID.
+
+The generic runtime transports controls as opaque snapshots. LingBot's adapter
+accepts:
+
+- `lingbot.camera_actions.v1` for per-latent-frame key states such as `w`, `a`,
+  `s`, `d`, `i`, `j`, `k`, and `l`;
+- `lingbot.camera_trajectory.v1` for explicit pose/intrinsics trajectories.
+
+## Validation
+
+Real-checkpoint validation uses 480x832 output, four DMD steps, and seed 42.
+The exercised matrix includes:
+
+- TP=1 and TP=2 execution;
+- two interleaved resident sessions;
+- action input and prompt switching;
+- seven contiguous chunks crossing the sink plus recent rolling window;
+- direct versus paged replay;
+- CUDA Graph execution;
+- reset, close, failure cleanup, and exact metadata matching; and
+- nine-frame VAE decode.
+
+Official `generate.py` accuracy and performance numbers, together with
+generation artifacts, are recorded in the PR validation section for the exact
+tested commit.
+
+## Current limitations
+
+- Only the 14B causal-fast checkpoint is supported.
+- The realtime control plane is internal; there is no public server transport yet.
+- AR-Diffusion stages currently require one replica because session-affine
+  routing across replicas is not implemented.
+- One AR block is generated per request and `max_num_seqs` must be one.
+- Stateful streaming VAE decode is not implemented; the realtime example emits
+  latent chunks.
+- SP/USP, pipeline/CFG parallelism, HSDP, VAE parallelism, quantization,
+  Cache-DiT, TeaCache, causal-pretrain, and the 1.3B checkpoint are not claimed.
+- No AMD GPU, Ascend NPU, or Intel GPU support is claimed.
+
+## References
+
+- Checkpoint: <https://huggingface.co/robbyant/lingbot-world-v2-14b-causal-fast-diffusers>
+- Official implementation: <https://github.com/robbyant/lingbot-world-v2>
+- Offline example: [`examples/offline_inference/diffusion/lingbot_world_v2.py`](../../examples/offline_inference/diffusion/lingbot_world_v2.py)
+- Realtime design: [`docs/design/feature/realtime_ar_diffusion.md`](../../docs/design/feature/realtime_ar_diffusion.md)
@@ -78,7 +78,7 @@ python examples/offline_inference/text_to_audio/text_to_audio.py \
 Start the online serving endpoint:

 ```bash
-vllm-omni serve /path/to/stable-audio-open-1.0 \
+vllm serve /path/to/stable-audio-open-1.0 \
  --host 0.0.0.0 \
  --port 8091 \
  --gpu-memory-utilization 0.9 \
@@ -39,6 +39,9 @@ platform section, document one or more tested hardware configurations.

 #### Command

+Serve commands use `vllm serve <model> --omni`. The `--omni` flag is what
+selects the Omni pipeline, so it is required and cannot be dropped.
+
 ```bash
 # Add the exact command(s) here
 ```
@@ -39,7 +39,7 @@ Other hardware is welcome as community validation lands.
 Launch the two-stage talker:

 ```bash
-vllm-omni serve inclusionAI/Ming-omni-tts-0.5B \
+vllm serve inclusionAI/Ming-omni-tts-0.5B \
    --deploy-config vllm_omni/deploy/ming_tts.yaml \
    --omni \
    --port 8091 \
@@ -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.
@@ -44,20 +44,20 @@ conditioned on reference audio and its transcript.
 Start the server from the repository root:

 ```bash
-vllm-omni serve zai-org/GLM-TTS --omni --trust-remote-code --port 8091
+vllm serve zai-org/GLM-TTS --omni --trust-remote-code --port 8091
 ```

 Async chunking is enabled by default in the bundled deployment config. For
 the sync (non-streaming) path:

 ```bash
-vllm-omni serve zai-org/GLM-TTS --omni --trust-remote-code --port 8091 --no-async-chunk
+vllm serve zai-org/GLM-TTS --omni --trust-remote-code --port 8091 --no-async-chunk
 ```

 Use a custom deploy config for advanced cases:

 ```bash
-vllm-omni serve zai-org/GLM-TTS --omni --trust-remote-code --port 8091 \
+vllm serve zai-org/GLM-TTS --omni --trust-remote-code --port 8091 \
  --deploy-config /path/to/your_glm_tts_overrides.yaml
 ```

@@ -1,15 +1,15 @@
 # Common dependencies for all platforms
-# transformers: kept exactly aligned with the upstream vLLM constraint.
-# vLLM pins ">= 4.56.0, != 5.0.*, != 5.1.*, != 5.2.*, != 5.3.*, != 5.4.*, != 5.5.0".
-# The previous extra "< 5.9.0" exclusion was removed when re-aligning with
-# upstream: transformers 5.9 changed _adjust_missing_and_unexpected_keys from
+# transformers: required unified multimodal model classes are unavailable
+# before 5.10.1; 5.10.0 is yanked, and 5.15 has a known construction
+# regression. transformers 5.9 also changed
+# _adjust_missing_and_unexpected_keys from
 # list+list to set|set, breaking models that declare
 # _keys_to_ignore_on_load_unexpected as a list (e.g. OpenMOSS-Team/MOSS-TTS-Nano)
 # with "TypeError: unsupported operand type(s) for |: 'list' and 'set'".
 # That breakage is now handled in code instead of by pinning:
 # MossTTSNanoForGeneration wraps weight loading in
 # transformers_keys_to_ignore_compat (vllm_omni/model_executor/models/utils.py).
-transformers >= 5.5.3
+transformers >= 5.10.1, < 5.15
 av>=14.0.0
 omegaconf>=2.3.0
 diffusers==0.38.0
Author	SHA1	Message	Date
Hongsheng Liu	f1abb13958	[Diffusion] Register runtime cache for direct DLO H2D Avoid recurrent mmap-to-pinned packing in no-AllGather DLO by registering transform-complete runtime-cache mappings under an explicit per-worker budget. Copy mapped tensors directly into the rotating device buffers while preserving bounded-staging fallback and ordered shutdown cleanup. Signed-off-by: Hongsheng Liu <liuhongsheng4@huawei.com>	2026-08-17 11:50:32 +08:00
Hongsheng Liu	b732ad2cea	[Diffusion] Document runtime cache hardware tradeoffs Signed-off-by: Hongsheng Liu <liuhongsheng4@huawei.com>	2026-08-17 00:43:06 +08:00
Hongsheng Liu	99347e1ba7	[Diffusion] Tighten runtime cache compatibility guards Signed-off-by: Hongsheng Liu <liuhongsheng4@huawei.com>	2026-08-16 23:26:25 +08:00
Hongsheng Liu	db86a6a078	[Diffusion] Add node-local runtime weight cache for DLO Signed-off-by: Hongsheng Liu <liuhongsheng4@huawei.com>	2026-08-16 23:15:48 +08:00
Yukim1andHongsheng Liu	baba7d1e2e	[Refactor] Remove legacy stage_args YAML loader (#6200 ) Signed-off-by: zwhzzz0821 <2831474076@qq.com> Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com>	2026-08-16 11:22:28 +08:00
Hongsheng Liu	9d2bb23ff6	[Diffusion] Add loader-owned host-weight plans for DLO (TP=1) (#6213 ) Signed-off-by: Hongsheng Liu <liuhongsheng4@huawei.com>	2026-08-16 11:02:38 +08:00
Jack Chen zhangbin.327 Hongsheng Liu	3875f8ded4	[Core] Add realtime AR-Diffusion tick sessions for LingBot World 2.0 (#5491 ) Signed-off-by: zhangbin.327 <zhangbin.327@bytedance.com> Signed-off-by: Jack47 <scsvip@gmail.com> Co-authored-by: zhangbin.327 <zhangbin.327@bytedance.com> Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com>	2026-08-16 09:11:20 +08:00
Yueqian Lin	9798bc3a8d	[Docs] Unify recipe serve commands on `vllm serve --omni` (#6221 ) Signed-off-by: Yueqian Lin <linyueqian@outlook.com>	2026-08-15 21:18:47 +00:00
Anurag	68bf8c2c42	[Bugfix] Prevent DiffusionResultPump crash on cancelled futures (#5793 ) (#5983 ) Signed-off-by: ANURAG KANADE <anuragkanade6@gmail.com>	2026-08-15 16:19:45 +00:00
yiminghub2024andHongsheng Liu	444485650b	Add MiniMax-H3 recipe for NPU 950PR (#6120 ) Signed-off-by: yiminghub2024 <482890@qq.com> Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com>	2026-08-16 00:04:38 +08:00
hurukawa	2a315e1a2d	[Perf] Support request-level batching for Wan2.2 pipelines (#5676 ) Signed-off-by: nagisa-kun <1434936049@qq.com> Signed-off-by: nagisa <1434936049@qq.com>	2026-08-15 23:58:36 +08:00
SYLARandHongsheng Liu	d1e230c95b	[Diffusion][Quantization] Enable MiniMax-H3 global FP8 with DLO (#5910 ) Signed-off-by: lishunyang12 <lishunyang12@163.com> Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com>	2026-08-15 23:01:26 +08:00
m0g3r	09b6a1604a	[Bugfix] Restore per-stage runtime env during launch (#6214 ) Signed-off-by: m0g3r <87276771+m0g3r@users.noreply.github.com>	2026-08-15 13:00:27 +00:00
Yueqian Lin	555336e887	[CI] Refresh Voxtral-4B-TTS perf baselines from a clean window (#6032 ) Signed-off-by: Yueqian Lin <linyueqian@outlook.com>	2026-08-15 11:51:56 +00:00
Samit	136cb27b05	[Bugfix] Fix lost async diffusion outputs in request-level batching (#6023 ) Signed-off-by: samithuang <285365963@qq.com>	2026-08-15 18:27:51 +08:00
SYLAR	ce6ab0ca46	[Model] Add LTX-2.5 support (#6070 ) Signed-off-by: lishunyang12 <lishunyang12@163.com>	2026-08-15 18:27:23 +08:00
NumberWan	f39b8684eb	[BugFix][Nightly CI] Adjust Qwen-Image accuracy thresholds for stable FA-deterministic mode (#5963 ) Signed-off-by: NumberWan <wantszkin2003@gmail.com>	2026-08-15 09:38:30 +00:00
MrlixiangWE	e57574ae29	[Bugfix] Carry ec_transfer_params and num_cache_creation_tokens on OmniRequestOutput (#6152 ) Signed-off-by: MrlixiangWE <mrdanaer@gmail.com>	2026-08-15 08:37:27 +00:00
mokeke mokashliu WeiQing Chen	7b76b6446d	[Model][Performance] Optimize MiniMax-H3 strict Ulysses boundaries (#6173 ) Signed-off-by: mokashliu <mokashliu@tencent.com> Co-authored-by: mokashliu <mokashliu@tencent.com> Co-authored-by: WeiQing Chen <40507679+david6666666@users.noreply.github.com>	2026-08-15 16:27:52 +08:00
hotTea jiangweifeng.dev Hongsheng Liu	a70d4d4b3c	[codex] Fix diffusion worker timeout for large broadcast payloads (#4845 ) Signed-off-by: jiangweifeng.dev <jiangweifeng.dev@bytedance.com> Co-authored-by: jiangweifeng.dev <jiangweifeng.dev@bytedance.com> Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com>	2026-08-15 16:03:04 +08:00
Hongsheng Liu	1260ceaae6	[Docs] Split diffusion attention and CPU offload guides (#6075 ) Signed-off-by: Hongsheng Liu <liuhongsheng4@huawei.com>	2026-08-15 14:41:09 +08:00
Ma1oneZhang	6f36fd3b78	[Bugfix] Preserve object storage URIs during model resolution (#5036 ) Signed-off-by: Ziyang Zhang <hafuhafu@qq.com>	2026-08-15 14:11:48 +08:00
mjZhaoElaine Claude Opus 5 Yueqian Lin	35cacf020c	[Model] Add Gepard-1.0 native-AR FSQ/NanoCodec TTS (offline inference) (#5666 ) Signed-off-by: Mengjie Zhao <zmj0129@gmail.com> Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com> Co-authored-by: Yueqian Lin <70319226+linyueqian@users.noreply.github.com>	2026-08-14 22:06:05 -07:00
wangyu	d671e6b0db	[CI] Expand E2E source_file_dependencies for shared model code (#5994 ) Signed-off-by: wangyu <410167048@qq.com> Signed-off-by: wangyu <53896905+yenuo26@users.noreply.github.com>	2026-08-15 12:01:20 +08:00
MrlixiangWE	b95c73742d	[Bugfix] Stop reading the removed OmniRequestOutput.request_output accessor (#6172 ) Signed-off-by: MrlixiangWE <mrdanaer@gmail.com>	2026-08-15 07:07:28 +08:00
Alex Brooks	fe54ecc388	[Frontend] Loosen Validation for TTS Models with No Uploaded Speakers (#5878 ) Signed-off-by: Alex Brooks <albrooks@redhat.com>	2026-08-14 17:05:11 -06:00
Joshna-Medisetty Cursor Chendi.Xue	4b6d6d5cf7	[Bugfix][XPU][Tests]Scope XPU pytest to explicit paths to fix collection crash (#6175 ) 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>	2026-08-14 15:23:41 -05:00
Alex Brooks	f7a5146093	[Feat/Bugfix] Fix API Server model tag <-> model sync & flag normalization (#3805 ) Signed-off-by: Alex Brooks <albrooks@redhat.com>	2026-08-14 19:13:00 +00:00
Wenjia Li	8ecd1f6d5c	[Perf][NPU] Fuse MiniMax H3 Qwen3-VL SwiGLU (#6167 ) Signed-off-by: Wenjia Li <wjialish@gmail.com>	2026-08-14 11:47:17 +00:00
Wenjia Li	ffd11c181e	[Perf][NPU] Fuse MiniMax H3 Qwen3-VL RoPE (#6061 ) Signed-off-by: Wenjia Li <wjialish@gmail.com>	2026-08-14 09:49:26 +00:00
R0CKSTAR	728640f4d9	[Bugfix][MUSA][Qwen Image] Avoid complex RoPE alias guards (#6110 ) Signed-off-by: Xiaodong Ye <xiaodong.ye@mthreads.com>	2026-08-14 15:51:23 +08:00
R0CKSTAR	022d7a7695	[Perf][Diffusion] Avoid redundant MiniMax-H3 reference video scans (#6064 ) Signed-off-by: Xiaodong Ye <xiaodong.ye@mthreads.com>	2026-08-14 15:49:26 +08:00
baonudesifeizhai	b300f66656	[Bugfix] Fix HSDP compatibility with the new online FP8 linear method (#5677 ) Signed-off-by: roG0d <baonudesifeizhai@gmail.com>	2026-08-14 15:45:55 +08:00
TJianandHongsheng Liu	bc73546c06	[ROCm] [CI] Migrate CI to mi300x for v0.27.x (#5886 ) Signed-off-by: tjtanaa <tunjian.tan@embeddedllm.com> Signed-off-by: TJian <tunjian.tan@embeddedllm.com> Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com>	2026-08-14 15:38:47 +08:00
Shenglei Fu Claude Sonnet 5 Alicia wangyu Hongsheng Liu	2b08419d8b	[CI] Per-model, per-entry-mode code coverage collection (#5593 ) Signed-off-by: Shenglei Fu <sfu@confluent.io> Signed-off-by: Shenglei Fu <117230642+ShengleiFu@users.noreply.github.com> Co-authored-by: Claude Sonnet 5 <noreply@anthropic.com> Co-authored-by: Alicia <115451386+congw729@users.noreply.github.com> Co-authored-by: wangyu <53896905+yenuo26@users.noreply.github.com> Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com>	2026-08-14 15:36:56 +08:00
zhengjiaandCursor	5d50bdc2ab	[Perf] Cap diffusion worker thread count before spawn (#6165 ) Signed-off-by: zjli2013 <leezhengjiang@126.com> Co-authored-by: Cursor <cursoragent@cursor.com>	2026-08-14 12:14:55 +08:00
Samit	8a5bf2fb3f	[doc]: Update latest news section for verl-omni release (#6187 ) Signed-off-by: Samit <285365963@qq.com>	2026-08-14 12:04:15 +08:00
Bo Li lishunyang12 Claude Opus 5	596c16a550	[Kernel] Fuse Q/K RMSNorm and RoPE (#5990 ) Signed-off-by: Bo Li <22713281+bobboli@users.noreply.github.com> Signed-off-by: lishunyang12 <lishunyang12@163.com> Signed-off-by: Mengjie Zhao <zmj0129@gmail.com> Co-authored-by: lishunyang12 <lishunyang12@163.com> Co-authored-by: Claude Opus 5 <noreply@anthropic.com>	2026-08-14 00:46:59 +00:00
Lucas(Ruijie)	90a08c0eda	[Model][TTS][WIP] Add dots.tts (rednote-hilab) — continuous-AR 48kHz TTS (#4765 ) Signed-off-by: Moore-Z <zhumoore9661996@gmail.com>	2026-08-13 17:58:33 +00:00
Sy03	dbc0dd6d85	[Bugfix][Higgs-Audio-V3] Disable XQA decode (#6068 ) Signed-off-by: Sy03 <1370724210@qq.com> Signed-off-by: Sysy <1370724210@qq.com>	2026-08-13 10:00:50 -07:00
Mu GuanLin	83d1f11449	[Feat] LTX Standard Two-Stage Pipeline (#5500 ) Signed-off-by: mglyn <1203789601@qq.com>	2026-08-14 00:13:14 +08:00
Yukim1andHongsheng Liu	4b66f62e10	[Refactor] Remove internal stage_configs_path plumbing (#5741 ) Signed-off-by: zwhzzz0821 <2831474076@qq.com> Signed-off-by: Hongsheng Liu <liuhongsheng4@huawei.com> Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com>	2026-08-13 21:29:03 +08:00
Weiming Liao	f41451f64c	[NPU] Upgrade to v0.27.0 (#6096 ) Signed-off-by: Weiming Liao <liaowm5@gmail.com>	2026-08-13 20:01:01 +08:00
Hongsheng Liu	bbe6ccc512	[Doc] Refresh community documentation (#6141 ) Signed-off-by: Hongsheng Liu <liuhongsheng4@huawei.com>	2026-08-13 17:04:06 +08:00
Canlin Guo	bbb709cfb5	[BugFix] Use MOSS-TTS-Local official sample params (#6156 ) Signed-off-by: Canlin Guo <canlinguosdu@gmail.com>	2026-08-13 08:56:26 +00:00