[Rebase] Rebase to vllm 0.23.0 (#4286)

Signed-off-by: tzhouam <tzhouam@connect.ust.hk>

.buildkite/pipeline-intel.yaml

@@ -10,7 +10,7 @@ steps:
           DOCKER_BUILDKIT: "1"
           # Buildkite will automatically replace this with the actual commit hash
           VLLM_IMAGE_TAG: "${BUILDKITE_COMMIT}"
-          VLLM_VERSION: "v0.22.0"
+          VLLM_VERSION: "v0.23.0"
         priority: 100
         timeout_in_minutes: 60
         soft_fail: false

.buildkite/pipeline.yml

@@ -1,14 +1,14 @@
 # Document 1: Buildkite loads only this block on first parse. The next step resolves docs-only skip-ci
 # from git diff, then uploads document 2. When docs-only skip applies, image-build still runs for the nightly
-# exception in upload_pipeline_with_skip_ci.sh (nightly-test label or main NIGHTLY=1); otherwise upload-nightly
+# exception in upload_pipeline.py (nightly-test label or main NIGHTLY=1); otherwise upload-nightly
 # can be skipped together with test-ready/test-merge. Weekly uploads also depend on image-build.
 
-# Document 2: appended after `---`; same file, read by upload_pipeline_with_skip_ci.sh (not evaluated as a second pipeline by Buildkite). Child uploads include test-ready.yml, test-merge.yml, test-nightly.yml, and test-weekly.yml.
+# Document 2: appended after `---`; same file, read by upload_pipeline.py (not evaluated as a second pipeline by Buildkite). Child uploads include test-ready.yml, test-merge.yml, test-nightly.yml, and test-weekly.yml.
 steps:
   - label: ":github: Resolve skip-ci & upload pipeline"
     key: upload-ci-pipeline
     commands:
-      - "bash .buildkite/scripts/upload_pipeline_with_skip_ci.sh"
+      - "python3 .buildkite/scripts/upload_pipeline.py --upload .buildkite/pipeline.yml"
     agents:
       queue: "cpu_queue_premerge"
 
@@ -31,7 +31,7 @@ steps:
     key: upload-ready-pipeline
     if: __UPLOAD_READY_IF__
     commands:
-      - buildkite-agent pipeline upload .buildkite/test-ready.yml
+      - python3 .buildkite/scripts/upload_pipeline.py --upload .buildkite/test-ready.yml
     agents:
       queue: "cpu_queue_premerge"
 
@@ -41,7 +41,7 @@ steps:
     key: upload-merge-pipeline
     if: __UPLOAD_MERGE_IF__
     commands:
-      - buildkite-agent pipeline upload .buildkite/test-merge.yml
+      - python3 .buildkite/scripts/upload_pipeline.py --upload .buildkite/test-merge.yml
     agents:
       queue: "cpu_queue_premerge"
 

.buildkite/rebase-pipeline.yaml

@@ -14,7 +14,7 @@ steps:
     depends_on: image-build
     key: upload-ready-pipeline
     commands:
-      - buildkite-agent pipeline upload .buildkite/test-ready.yml
+      - python3 .buildkite/scripts/upload_pipeline.py --upload .buildkite/test-ready.yml
     agents:
       queue: "cpu_queue_premerge"
 
@@ -23,7 +23,7 @@ steps:
     depends_on: image-build
     key: upload-merge-pipeline
     commands:
-      - buildkite-agent pipeline upload .buildkite/test-merge.yml
+      - python3 .buildkite/scripts/upload_pipeline.py --upload .buildkite/test-merge.yml
     agents:
       queue: "cpu_queue_premerge"
 

.buildkite/scripts/upload_pipeline.py

@@ -0,0 +1,308 @@
+#!/usr/bin/env python3
+# SPDX-License-Identifier: Apache-2.0
+"""Render and optionally upload Buildkite pipeline YAML with diff-aware logic.
+
+Bootstrap mode (pipeline.yml with __IMAGE_BUILD_IF__ placeholders):
+  - Detect docs-only PR/main changes and substitute skip-ci ``if`` expressions.
+
+Test pipeline mode (e.g. test-merge.yml):
+  - Drop steps or groups whose ``source_file_dependencies`` do not match changed files.
+  - ``source_file_dependencies`` may be set on a leaf step or on a group step.
+  - Always strip ``source_file_dependencies`` from uploaded YAML (Buildkite does not
+    recognize this uploader-only key).
+
+Usage:
+  python3 upload_pipeline.py [--upload] <pipeline.yml>
+
+  # Bootstrap (replaces upload_pipeline_with_skip_ci.sh):
+  python3 upload_pipeline.py --upload .buildkite/pipeline.yml
+
+  # Test pipeline (replaces upload_test_pipeline_with_diff_skip.py):
+  python3 upload_pipeline.py --upload .buildkite/test-merge.yml
+"""
+
+from __future__ import annotations
+
+import argparse
+import os
+import subprocess
+import sys
+from pathlib import Path
+from typing import Any
+
+try:
+    import yaml
+except ImportError:
+    subprocess.check_call(
+        [sys.executable, "-m", "pip", "install", "-q", "pyyaml"],
+    )
+    import yaml
+
+LOG = "upload_pipeline"
+ROOT = Path(__file__).resolve().parent.parent.parent
+DOC_SEP = "\n---\n"
+BOOTSTRAP_MARKER = "__IMAGE_BUILD_IF__"
+
+
+def _log(message: str) -> None:
+    print(f"{LOG}: {message}", file=sys.stderr)
+
+
+def _git(*args: str) -> subprocess.CompletedProcess[str]:
+    return subprocess.run(
+        ["git", *args],
+        check=False,
+        capture_output=True,
+        text=True,
+    )
+
+
+def resolve_changed_files() -> list[str] | None:
+    """Return changed file paths, or None when diff cannot be resolved."""
+    is_pr = os.environ.get("BUILDKITE_PULL_REQUEST", "false") != "false" and os.environ.get(
+        "BUILDKITE_PULL_REQUEST", ""
+    )
+    commit = os.environ.get("BUILDKITE_COMMIT", "")
+
+    if is_pr:
+        base_branch = os.environ.get("BUILDKITE_PULL_REQUEST_BASE_BRANCH", "main")
+        base_ref = f"origin/{base_branch}"
+        if _git("rev-parse", "--verify", base_ref).returncode != 0:
+            _log(f"origin/{base_branch} not found locally; trying fetch")
+            _git("fetch", "--depth=200", "origin", base_branch)
+        if _git("rev-parse", "--verify", base_ref).returncode != 0:
+            if _git("rev-parse", "--verify", base_branch).returncode == 0:
+                base_ref = base_branch
+            else:
+                _log(f"cannot resolve PR base {base_branch}; using safe defaults")
+                return None
+        diff_range = f"{base_ref}...{commit}"
+    elif os.environ.get("BUILDKITE_BRANCH", "") == "main":
+        if _git("rev-parse", "--verify", f"{commit}^").returncode != 0:
+            _log("main commit has no parent; using safe defaults")
+            return None
+        diff_range = f"{commit}^..{commit}"
+    else:
+        _log("not PR/main build; using safe defaults")
+        return None
+
+    result = _git("diff", "--name-only", diff_range)
+    if result.returncode != 0:
+        _log(f"git diff failed ({diff_range}); using safe defaults")
+        return None
+
+    files = [line for line in result.stdout.splitlines() if line.strip()]
+    _log(f"{len(files)} changed file(s)")
+    return files
+
+
+def is_docs_only_change(changed_files: list[str]) -> bool:
+    has_any = False
+    for file_path in changed_files:
+        if not file_path:
+            continue
+        has_any = True
+        if file_path.startswith("docs/"):
+            continue
+        if file_path.endswith(".md"):
+            continue
+        if file_path == "mkdocs.yml":
+            continue
+        return False
+    return has_any
+
+
+def resolve_skip_ci(changed_files: list[str] | None) -> bool:
+    if changed_files is None:
+        _log("skip-ci=0 (could not resolve changed files)")
+        return False
+    if is_docs_only_change(changed_files):
+        _log("docs-only change detected; skip-ci=1")
+        return True
+    _log("non-doc changes detected; skip-ci=0")
+    return False
+
+
+def render_bootstrap_pipeline(text: str, *, skip_ci: bool) -> str:
+    if DOC_SEP in text:
+        _, continuation = text.split(DOC_SEP, 1)
+    else:
+        continuation = text
+
+    nightly_only = (
+        '(build.pull_request.labels includes "nightly-test") || (build.branch == "main" && build.env("NIGHTLY") == "1")'
+    )
+    if skip_ci:
+        image_if = f"'{nightly_only}'"
+        ready_if = "'false'"
+        merge_if = "'false'"
+    else:
+        image_if = "'true'"
+        ready_if = '\'build.branch != "main" && build.pull_request.labels includes "ready"\''
+        merge_if = (
+            '\'(build.branch == "main" && build.env("NIGHTLY") != "1" '
+            '&& build.env("WEEKLY") != "1") || '
+            '(build.branch != "main" && build.pull_request.labels includes "merge-test")\''
+        )
+
+    return (
+        continuation.replace("__IMAGE_BUILD_IF__", image_if)
+        .replace("__UPLOAD_READY_IF__", ready_if)
+        .replace("__UPLOAD_MERGE_IF__", merge_if)
+    )
+
+
+def _matches_dependencies(changed_files: list[str], prefixes: list[str]) -> bool:
+    for path in changed_files:
+        for prefix in prefixes:
+            normalized = prefix.rstrip("/")
+            if path == normalized or path.startswith(f"{normalized}/"):
+                return True
+    return False
+
+
+def _strip_source_file_dependencies(step: dict[str, Any]) -> dict[str, Any]:
+    return {key: value for key, value in step.items() if key != "source_file_dependencies"}
+
+
+def _step_label(step: dict[str, Any]) -> str:
+    return str(step.get("group") or step.get("label") or "<step>")
+
+
+def _filter_steps(steps: list[Any], changed_files: list[str]) -> list[Any]:
+    filtered: list[Any] = []
+    for step in steps:
+        if not isinstance(step, dict):
+            filtered.append(step)
+            continue
+
+        deps = step.get("source_file_dependencies")
+        if deps is not None and not isinstance(deps, list):
+            raise ValueError(
+                f"source_file_dependencies must be a list in step {_step_label(step)!r}",
+            )
+        if deps is not None and not _matches_dependencies(changed_files, deps):
+            _log(
+                f"skip {_step_label(step)!r} (no changes under {deps})",
+            )
+            continue
+
+        nested = step.get("steps")
+        if nested is not None:
+            kept_nested = _filter_steps(nested, changed_files)
+            if not kept_nested:
+                _log(f"omit empty group {_step_label(step)!r}")
+                continue
+            new_step = _strip_source_file_dependencies(step)
+            new_step["steps"] = kept_nested
+            filtered.append(new_step)
+            continue
+
+        if deps is not None:
+            filtered.append(_strip_source_file_dependencies(step))
+        else:
+            filtered.append(step)
+
+    return filtered
+
+
+def _strip_uploader_metadata_from_steps(steps: list[Any]) -> list[Any]:
+    """Remove uploader-only keys while keeping all steps (no diff filtering)."""
+    stripped: list[Any] = []
+    for step in steps:
+        if not isinstance(step, dict):
+            stripped.append(step)
+            continue
+
+        deps = step.get("source_file_dependencies")
+        if deps is not None and not isinstance(deps, list):
+            raise ValueError(
+                f"source_file_dependencies must be a list in step {_step_label(step)!r}",
+            )
+
+        nested = step.get("steps")
+        new_step = _strip_source_file_dependencies(step)
+        if nested is not None:
+            new_step["steps"] = _strip_uploader_metadata_from_steps(nested)
+        stripped.append(new_step)
+
+    return stripped
+
+
+def render_test_pipeline(
+    doc: dict[str, Any],
+    changed_files: list[str] | None,
+) -> dict[str, Any]:
+    steps = doc.get("steps")
+    if not isinstance(steps, list):
+        return doc
+    if changed_files is not None:
+        steps = _filter_steps(steps, changed_files)
+    else:
+        steps = _strip_uploader_metadata_from_steps(steps)
+    return {**doc, "steps": steps}
+
+
+def resolve_pipeline_path(arg: str) -> Path:
+    path = Path(arg)
+    if path.is_absolute():
+        return path
+    return ROOT / path
+
+
+def render_pipeline(path: Path) -> str:
+    text = path.read_text(encoding="utf-8")
+    changed_files = resolve_changed_files()
+
+    if BOOTSTRAP_MARKER in text:
+        rendered = render_bootstrap_pipeline(text, skip_ci=resolve_skip_ci(changed_files))
+        return rendered
+
+    doc = yaml.safe_load(text)
+    if not isinstance(doc, dict):
+        raise ValueError(f"invalid pipeline YAML: {path}")
+
+    doc = render_test_pipeline(doc, changed_files)
+
+    return yaml.safe_dump(doc, sort_keys=False)
+
+
+def upload_to_buildkite(content: str) -> None:
+    subprocess.run(
+        ["buildkite-agent", "pipeline", "upload"],
+        input=content,
+        text=True,
+        check=True,
+    )
+
+
+def main() -> int:
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument(
+        "pipeline",
+        nargs="?",
+        default=".buildkite/pipeline.yml",
+        help="Pipeline YAML path (default: .buildkite/pipeline.yml)",
+    )
+    parser.add_argument(
+        "--upload",
+        action="store_true",
+        help="Pipe rendered YAML to buildkite-agent pipeline upload",
+    )
+    args = parser.parse_args()
+
+    path = resolve_pipeline_path(args.pipeline)
+    if not path.is_file():
+        _log(f"missing pipeline file: {path}")
+        return 1
+
+    rendered = render_pipeline(path)
+    if args.upload:
+        upload_to_buildkite(rendered)
+    else:
+        sys.stdout.write(rendered)
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

.buildkite/scripts/upload_pipeline_with_skip_ci.sh

@@ -1,146 +0,0 @@
-#!/usr/bin/env bash
-# Evaluate docs-only skip-ci and upload continuation steps from the given pipeline YAML
-# (YAML document after the first `---`). Buildkite `if` is evaluated at upload time.
-#
-# Usage: upload_pipeline_with_skip_ci.sh [pipeline_yaml]
-#   pipeline_yaml: path relative to repo root or absolute (default: .buildkite/pipeline.yml)
-set -euo pipefail
-
-ROOT="$(cd "$(dirname "${BASH_SOURCE[0]}")/../.." && pwd)"
-PIPELINE_ARG="${1:-.buildkite/pipeline.yml}"
-if [[ "${PIPELINE_ARG}" = /* ]]; then
-  PIPELINE_YML="${PIPELINE_ARG}"
-else
-  PIPELINE_YML="${ROOT}/${PIPELINE_ARG}"
-fi
-
-# Prints a single digit to stdout: 1 = skip image CI, 0 = run. Logs go to stderr.
-is_docs_only_change() {
-  local file_path
-  local has_any=0
-
-  while IFS= read -r file_path; do
-    [[ -z "${file_path}" ]] && continue
-    has_any=1
-
-    if [[ "${file_path}" == docs/* ]]; then
-      continue
-    fi
-    if [[ "${file_path}" == *.md ]]; then
-      continue
-    fi
-    if [[ "${file_path}" == "mkdocs.yml" ]]; then
-      continue
-    fi
-    return 1
-  done
-
-  [[ "${has_any}" -eq 1 ]]
-}
-
-resolve_skip_ci() {
-  local is_pr_build=0
-  local files
-  local base_branch base_ref
-
-  if [[ "${BUILDKITE_PULL_REQUEST:-false}" != "false" && -n "${BUILDKITE_PULL_REQUEST:-}" ]]; then
-    is_pr_build=1
-  fi
-
-  if [[ "${is_pr_build}" -eq 1 ]]; then
-    base_branch="${BUILDKITE_PULL_REQUEST_BASE_BRANCH:-main}"
-    if ! git rev-parse --verify "origin/${base_branch}" >/dev/null 2>&1; then
-      echo "resolve_skip_ci: origin/${base_branch} not found locally; trying fetch" >&2
-      git fetch --depth=200 origin "${base_branch}" >/dev/null 2>&1 || true
-    fi
-
-    base_ref=""
-    if git rev-parse --verify "origin/${base_branch}" >/dev/null 2>&1; then
-      base_ref="origin/${base_branch}"
-    elif git rev-parse --verify "${base_branch}" >/dev/null 2>&1; then
-      base_ref="${base_branch}"
-    else
-      echo "resolve_skip_ci: cannot resolve PR base ${base_branch}; skip-ci=0" >&2
-      echo -n 0
-      return 0
-    fi
-
-    if ! files="$(git diff --name-only "${base_ref}...${BUILDKITE_COMMIT}" 2>/dev/null)"; then
-      echo "resolve_skip_ci: failed to compute PR changed files; skip-ci=0" >&2
-      echo -n 0
-      return 0
-    fi
-  elif [[ "${BUILDKITE_BRANCH:-}" == "main" ]]; then
-    if ! git rev-parse --verify "${BUILDKITE_COMMIT}^" >/dev/null 2>&1; then
-      echo "resolve_skip_ci: commit has no parent on main; skip-ci=0" >&2
-      echo -n 0
-      return 0
-    fi
-    if ! files="$(git diff --name-only "${BUILDKITE_COMMIT}^..${BUILDKITE_COMMIT}" 2>/dev/null)"; then
-      echo "resolve_skip_ci: failed to compute main changed files; skip-ci=0" >&2
-      echo -n 0
-      return 0
-    fi
-  else
-    echo "resolve_skip_ci: not PR/main build; skip-ci=0" >&2
-    echo -n 0
-    return 0
-  fi
-
-  if is_docs_only_change <<< "${files}"; then
-    echo "resolve_skip_ci: docs-only change detected; skip-ci=1" >&2
-    echo -n 1
-    return 0
-  fi
-
-  echo "resolve_skip_ci: non-doc changes detected; skip-ci=0" >&2
-  echo -n 0
-}
-
-SKIP_CI="$(resolve_skip_ci)"
-
-if [[ ! -f "${PIPELINE_YML}" ]]; then
-  echo "upload_pipeline_with_skip_ci: missing ${PIPELINE_YML}" >&2
-  exit 1
-fi
-
-export ROOT SKIP_CI PIPELINE_YML
-python3 <<'PY' | buildkite-agent pipeline upload
-import os
-import pathlib
-
-path = pathlib.Path(os.environ["PIPELINE_YML"])
-text = path.read_text(encoding="utf-8")
-sep = "\n---\n"
-# Two supported layouts:
-#   - multi-doc: doc 1 = bootstrap (loaded by Buildkite), doc 2 (after `---`) = uploaded steps.
-#   - single-doc: whole file is the uploaded steps; caller (UI init step) is the bootstrap.
-if sep in text:
-    _, continuation = text.split(sep, 1)
-else:
-    continuation = text
-
-skip = os.environ.get("SKIP_CI") == "1"
-# When docs-only skip-ci: skip default CI image, but still build for L4 nightly (PR label nightly-test or
-# main NIGHTLY=1), otherwise upload-nightly (depends_on image-build) would be skipped too.
-nightly_only = (
-    '(build.pull_request.labels includes "nightly-test") '
-    '|| (build.branch == "main" && build.env("NIGHTLY") == "1")'
-)
-# Placeholder in pipeline.yml is `if: __IMAGE_BUILD_IF__` (valid YAML); replace value only.
-if skip:
-    rep = f"'{nightly_only}'"
-    ready_rep = "'false'"
-    merge_rep = "'false'"
-else:
-    rep = "'true'"
-    ready_rep = "'build.branch != \"main\" && build.pull_request.labels includes \"ready\"'"
-    merge_rep = "'(build.branch == \"main\" && build.env(\"NIGHTLY\") != \"1\" && build.env(\"WEEKLY\") != \"1\") || (build.branch != \"main\" && build.pull_request.labels includes \"merge-test\")'"
-rendered = (
-    continuation
-    .replace("__IMAGE_BUILD_IF__", rep)
-    .replace("__UPLOAD_READY_IF__", ready_rep)
-    .replace("__UPLOAD_MERGE_IF__", merge_rep)
-)
-print(rendered, end="")
-PY

.buildkite/test-amd-merge.yml

@@ -1,54 +1,111 @@
+env:
+  VLLM_WORKER_MULTIPROC_METHOD: spawn
+  HF_HUB_DOWNLOAD_TIMEOUT: 300
+  HF_HUB_ETAG_TIMEOUT: 60
+
 steps:
 
-- label: "Simple · Diffusion & Model Executor Test"
-  agent_pool: mi325_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"
+- group: ":card_index_dividers: Simple Test"
+  steps:
+    - label: "Simple · Diffusion & Model Executor Test"
+      agent_pool: mi325_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"
 
-- label: "Simple · Other Test"
-  agent_pool: mi325_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"
+    - label: "Simple · Other Test"
+      agent_pool: mi325_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"
 
-- label: "Diffusion Images API LoRA E2E"
-  agent_pool: mi325_1
-  depends_on: amd-build
-  mirror_hardwares: [amdproduction]
-  grade: Blocking
-  commands:
-    - timeout 20m pytest -s -v tests/diffusion/lora/test_images_generations_lora.py
+- group: ":card_index_dividers: Diffusion Test"
+  steps:
+    - label: "Diffusion · LoRA Test"
+      agent_pool: mi325_1
+      depends_on: amd-build
+      mirror_hardwares: [amdproduction]
+      grade: Blocking
+      commands:
+        - timeout 20m pytest -s -v tests/diffusion/lora/test_images_generations_lora.py
 
-- label: "Diffusion Model CPU offloading Test"
-  agent_pool: mi325_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))
-      '
+    - label: "Diffusion Model CPU offloading Test"
+      agent_pool: mi325_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))
+          '
+
+    - label: "Diffusion Sequence Parallelism Test (Need 4 GPUs)"
+      agent_pool: mi325_4
+      depends_on: amd-build
+      mirror_hardwares: [amdproduction]
+      grade: Blocking
+      commands:
+        - export VLLM_LOGGING_LEVEL=DEBUG
+        - export VLLM_WORKER_MULTIPROC_METHOD=spawn
+        - timeout 20m pytest -s -v tests/diffusion/distributed/test_sequence_parallel.py
+        - timeout 20m pytest -s -v tests/diffusion/distributed/test_ulysses_uaa_perf.py
+
+    # merge-only tests
+    - label: "Diffusion Tensor Parallelism Test"
+      agent_pool: mi325_2
+      depends_on: amd-build
+      mirror_hardwares: [amdproduction]
+      grade: Blocking
+      commands:
+        - pytest -s -v tests/diffusion/distributed/test_zimage_parallelism.py
+
+    - label: "Diffusion Model Test"
+      agent_pool: mi325_1
+      depends_on: amd-build
+      mirror_hardwares: [amdproduction]
+      grade: Blocking
+      commands:
+        - timeout 30m pytest -s -v tests/e2e/offline_inference/test_z_image.py
+
+    - label: "Diffusion GPU Worker Test"
+      agent_pool: mi325_2
+      depends_on: amd-build
+      mirror_hardwares: [amdproduction]
+      grade: Blocking
+      commands:
+        - timeout 20m pytest -s -v tests/diffusion/test_diffusion_worker.py
+
+    - label: "Diffusion · Wan22 Test"
+      agent_pool: mi325_1
+      depends_on: amd-build
+      mirror_hardwares: [amdproduction]
+      grade: Blocking
+      commands:
+        - |
+          timeout 40m bash -c "
+            export VLLM_IMAGE_FETCH_TIMEOUT=60
+            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'
+          "
 
 ## ISSUE depends on `diffusers` package: https://github.com/huggingface/diffusers/issues/13274
 # - label: "Audio Generation Model Test"
@@ -62,42 +119,6 @@ steps:
 #     - export VLLM_WORKER_MULTIPROC_METHOD=spawn
 #     - timeout 20m pytest -s -v tests/e2e/offline_inference/test_stable_audio_expansion.py -m "advanced_model and diffusion and L4" --run-level advanced_model
 
-- label: "Diffusion Sequence Parallelism Test (Need 4 GPUs)"
-  agent_pool: mi325_4
-  depends_on: amd-build
-  mirror_hardwares: [amdproduction]
-  grade: Blocking
-  commands:
-    - export VLLM_LOGGING_LEVEL=DEBUG
-    - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-    - timeout 20m pytest -s -v tests/diffusion/distributed/test_sequence_parallel.py
-    - timeout 20m pytest -s -v tests/diffusion/distributed/test_ulysses_uaa_perf.py
-
-# merge-only tests
-- label: "Diffusion Tensor Parallelism Test"
-  agent_pool: mi325_2
-  depends_on: amd-build
-  mirror_hardwares: [amdproduction]
-  grade: Blocking
-  commands:
-    - pytest -s -v tests/diffusion/distributed/test_zimage_parallelism.py
-
-- label: "Diffusion Model Test"
-  agent_pool: mi325_1
-  depends_on: amd-build
-  mirror_hardwares: [amdproduction]
-  grade: Blocking
-  commands:
-    - timeout 30m pytest -s -v tests/e2e/offline_inference/test_z_image.py
-
-- label: "Diffusion GPU Worker Test"
-  agent_pool: mi325_2
-  depends_on: amd-build
-  mirror_hardwares: [amdproduction]
-  grade: Blocking
-  commands:
-    - timeout 20m pytest -s -v tests/diffusion/test_diffusion_worker.py
-
 - label: "Engine Test"
   agent_pool: mi325_1
   depends_on: amd-build
@@ -111,50 +132,65 @@ steps:
 # This model seems to be important, so I am trying to run it here.
 # Since this is from test-nightly.yml,
 # I keep this non-blocking.
-- label: "Omni Model Test Qwen2-5-Omni"
-  agent_pool: mi325_4
-  depends_on: amd-build
-  mirror_hardwares: [amdproduction]
-  commands:
-    - export VLLM_ALLOW_LONG_MAX_MODEL_LEN="1"
-    - pytest -s -v tests/e2e/*/test_qwen2_5_omni_expansion.py --run-level "core_model"
+- group: ":card_index_dividers: E2E Test"
+  steps:
+    - label: "Omni Model Test Qwen2-5-Omni"
+      agent_pool: mi325_4
+      depends_on: amd-build
+      mirror_hardwares: [amdproduction]
+      commands:
+        - export VLLM_ALLOW_LONG_MAX_MODEL_LEN="1"
+        - 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
-  depends_on: amd-build
-  mirror_hardwares: [amdproduction]
-  grade: Blocking
-  commands:
-    - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-    - pytest -s -v tests/e2e/offline_inference/test_qwen3_omni.py tests/e2e/online_serving/test_qwen3_omni.py -m 'advanced_model and cuda' --run-level 'advanced_model'
+    - label: "Omni Model Test Qwen3-Omni"
+      agent_pool: mi325_2
+      depends_on: amd-build
+      mirror_hardwares: [amdproduction]
+      grade: Blocking
+      commands:
+        - export VLLM_WORKER_MULTIPROC_METHOD=spawn
+        - pytest -s -v tests/e2e/offline_inference/test_qwen3_omni.py tests/e2e/online_serving/test_qwen3_omni.py -m 'advanced_model and cuda' --run-level 'advanced_model'
 
-- label: "Qwen3-TTS CustomVoice E2E Test"
-  agent_pool: mi325_1
-  depends_on: amd-build
-  mirror_hardwares: [amdproduction]
-  grade: Blocking
-  commands:
-    - |
-      timeout 20m 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"
-      '
+    - label: "Qwen3-TTS CustomVoice E2E Test"
+      agent_pool: mi325_1
+      depends_on: amd-build
+      mirror_hardwares: [amdproduction]
+      grade: Blocking
+      commands:
+        - |
+          timeout 20m 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"
+          '
 
-- label: "Qwen3-TTS Base E2E Test"
-  agent_pool: mi325_1
-  depends_on: amd-build
-  mirror_hardwares: [amdproduction]
-  grade: Blocking
-  commands:
-    - |
-      timeout 30m bash -c '
-        export VLLM_LOGGING_LEVEL=DEBUG
-        export VLLM_WORKER_MULTIPROC_METHOD=spawn
-        export VLLM_ALLOW_LONG_MAX_MODEL_LEN="1"
-        pytest -s -v tests/e2e/online_serving/test_qwen3_tts_base.py tests/e2e/offline_inference/test_qwen3_tts_base.py -m "advanced_model" --run-level "advanced_model"
-      '
+    - label: "Qwen3-TTS Base E2E Test"
+      agent_pool: mi325_1
+      depends_on: amd-build
+      mirror_hardwares: [amdproduction]
+      grade: Blocking
+      commands:
+        - |
+          timeout 30m bash -c '
+            export VLLM_LOGGING_LEVEL=DEBUG
+            export VLLM_WORKER_MULTIPROC_METHOD=spawn
+            export VLLM_ALLOW_LONG_MAX_MODEL_LEN="1"
+            pytest -s -v tests/e2e/online_serving/test_qwen3_tts_base.py tests/e2e/offline_inference/test_qwen3_tts_base.py -m "advanced_model" --run-level "advanced_model"
+          '
+
+    - label: "Voxtral-TTS E2E Test"
+      agent_pool: mi325_1
+      depends_on: amd-build
+      mirror_hardwares: [amdproduction]
+      grade: Blocking
+      commands:
+        - |
+          timeout 20m bash -c '
+            export VLLM_LOGGING_LEVEL=DEBUG
+            export VLLM_WORKER_MULTIPROC_METHOD=spawn
+            pytest -s -v tests/e2e/online_serving/test_voxtral_tts.py tests/e2e/offline_inference/test_voxtral_tts.py -m "advanced_model" --run-level "advanced_model"
+          '
 
 # TODO: Bagel test on ROCm is very unstable. @tjtanaa
 # Need to debug before reneable numerical changes across large PRs
@@ -195,28 +231,3 @@ steps:
 #     - export VLLM_WORKER_MULTIPROC_METHOD=spawn
 #     - export VLLM_ROCM_USE_AITER_RMSNORM=0
 #     - timeout 40m pytest -s -v tests/e2e/online_serving/test_bagel.py -m "advanced_model" --run-level "advanced_model" -k "rocm"
-
-- label: "Diffusion · Wan22 Test"
-  agent_pool: mi325_1
-  depends_on: amd-build
-  mirror_hardwares: [amdproduction]
-  grade: Blocking
-  commands:
-    - |
-      timeout 40m bash -c "
-        export VLLM_IMAGE_FETCH_TIMEOUT=60
-        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'
-      "
-
-- label: "Voxtral-TTS E2E Test"
-  agent_pool: mi325_1
-  depends_on: amd-build
-  mirror_hardwares: [amdproduction]
-  grade: Blocking
-  commands:
-    - |
-      timeout 20m bash -c '
-        export VLLM_LOGGING_LEVEL=DEBUG
-        export VLLM_WORKER_MULTIPROC_METHOD=spawn
-        pytest -s -v tests/e2e/online_serving/test_voxtral_tts.py tests/e2e/offline_inference/test_voxtral_tts.py -m "advanced_model" --run-level "advanced_model"
-      '

.buildkite/test-amd-ready.yaml

@@ -1,42 +1,33 @@
+env:
+  VLLM_WORKER_MULTIPROC_METHOD: spawn
+  HF_HUB_DOWNLOAD_TIMEOUT: 300
+  HF_HUB_ETAG_TIMEOUT: 60
+
 steps:
 
-- label: "Simple · Diffusion & Model Executor Test"
-  agent_pool: mi325_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
-    - "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"
+- group: ":card_index_dividers: Simple Test"
+  steps:
+    - label: "Simple · Diffusion & Model Executor Test"
+      agent_pool: mi325_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
+        - "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"
 
-- label: "Simple · Other Test"
-  agent_pool: mi325_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
-    - "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"
-
-- label: "Voxtral TTS CUDA Unit Test"
-  agent_pool: mi325_1
-  depends_on: amd-build
-  mirror_hardwares: [amdproduction]
-  grade: Blocking
-  commands:
-    - timeout 10m pytest -s -v tests/model_executor/models/voxtral_tts/test_cuda_graph_acoustic_transformer.py
-
-- label: "Diffusion · Batch Test"
-  agent_pool: mi325_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"
+    - label: "Simple · Other Test"
+      agent_pool: mi325_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
+        - "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"
 
 - label: "Custom Pipeline Test"
   agent_pool: mi325_1
@@ -46,23 +37,92 @@ steps:
   commands:
     - timeout 20m pytest -s -v tests/e2e/offline_inference/custom_pipeline/ -m "core_model"
 
-- label: "Diffusion Model CPU offloading Test"
+
+- label: "Voxtral TTS CUDA Unit Test"
   agent_pool: mi325_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))
-      '
+    - timeout 10m pytest -s -v tests/model_executor/models/voxtral_tts/test_cuda_graph_acoustic_transformer.py
+
+- group: ":card_index_dividers: Diffusion Test"
+  steps:
+    - label: "Diffusion · Batch Test"
+      agent_pool: mi325_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"
+
+    - label: "Diffusion · Offloader Test"
+      agent_pool: mi325_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))
+          '
+
+    - label: "Diffusion · Cache Test"
+      agent_pool: mi325_1
+      depends_on: amd-build
+      mirror_hardwares: [amdproduction]
+      grade: Blocking
+      commands:
+        - 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"
+
+    - label: "Diffusion Sequence Parallelism Test"
+      agent_pool: mi325_2
+      depends_on: amd-build
+      mirror_hardwares: [amdproduction]
+      grade: Blocking
+      commands:
+        - export VLLM_LOGGING_LEVEL=DEBUG
+        - export VLLM_WORKER_MULTIPROC_METHOD=spawn
+        - timeout 20m pytest -s -v tests/diffusion/distributed/test_sequence_parallel.py -m core_model
+
+    - label: "Diffusion GPU Worker Test"
+      agent_pool: mi325_2
+      depends_on: amd-build
+      mirror_hardwares: [amdproduction]
+      grade: Blocking
+      commands:
+        - timeout 20m pytest -s -v tests/diffusion/test_diffusion_worker.py
+
+    - label: "Diffusion · Model Test"
+      agent_pool: mi325_1
+      depends_on: amd-build
+      mirror_hardwares: [amdproduction]
+      grade: Blocking
+      commands:
+        - timeout 15m pytest -sv tests/diffusion/models/ -m 'core_model and cuda' --run-level "core_model"
+
+    - label: "Diffusion · Wan22 Test"
+      agent_pool: mi325_1
+      depends_on: amd-build
+      mirror_hardwares: [amdproduction]
+      grade: Blocking
+      commands:
+        - |
+          timeout 40m bash -c "
+            export VLLM_TEST_CLEAN_GPU_MEMORY=1
+            export VLLM_IMAGE_FETCH_TIMEOUT=60
+            pytest -s -v tests/e2e/online_serving/test_wan22_t2v.py -m 'core_model' --run-level 'core_model'
+          "
 
 ## ISSUE depends on `diffusers` package: https://github.com/huggingface/diffusers/issues/13274
 # - label: "Audio Generation Model Test"
@@ -76,56 +136,6 @@ steps:
 #     - export VLLM_WORKER_MULTIPROC_METHOD=spawn
 #     - timeout 20m pytest -s -v tests/e2e/offline_inference/test_stable_audio_expansion.py -m "advanced_model and diffusion and L4" --run-level advanced_model
 
-- label: "Diffusion Cache Backend Test"
-  agent_pool: mi325_1
-  depends_on: amd-build
-  mirror_hardwares: [amdproduction]
-  grade: Blocking
-  commands:
-    - 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"
-
-- label: "Diffusion Sequence Parallelism Test"
-  agent_pool: mi325_2
-  depends_on: amd-build
-  mirror_hardwares: [amdproduction]
-  grade: Blocking
-  commands:
-    - export VLLM_LOGGING_LEVEL=DEBUG
-    - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-    - timeout 20m pytest -s -v tests/diffusion/distributed/test_sequence_parallel.py -m core_model
-
-- label: "Diffusion GPU Worker Test"
-  agent_pool: mi325_2
-  depends_on: amd-build
-  mirror_hardwares: [amdproduction]
-  grade: Blocking
-  commands:
-    - timeout 20m pytest -s -v tests/diffusion/test_diffusion_worker.py
-
-- label: "Diffusion · Model Test"
-  agent_pool: mi325_1
-  depends_on: amd-build
-  mirror_hardwares: [amdproduction]
-  grade: Blocking
-  commands:
-    - timeout 15m pytest -sv tests/diffusion/models/ -m 'core_model and cuda' --run-level "core_model"
-
-- label: "Diffusion · Wan22 Test"
-  agent_pool: mi325_1
-  depends_on: amd-build
-  mirror_hardwares: [amdproduction]
-  grade: Blocking
-  commands:
-    - |
-      timeout 40m bash -c "
-        export VLLM_TEST_CLEAN_GPU_MEMORY=1
-        export VLLM_IMAGE_FETCH_TIMEOUT=60
-        pytest -s -v tests/e2e/online_serving/test_wan22_t2v.py -m 'core_model' --run-level 'core_model'
-      "
-
 - label: "Engine Test"
   agent_pool: mi325_1
   depends_on: amd-build
@@ -149,43 +159,56 @@ steps:
 #     - timeout 10m pytest -s -v tests/e2e/offline_inference/test_qwen3_omni.py
 #     - timeout 20m pytest -s -v tests/e2e/online_serving/test_qwen3_omni.py -m "core_model" --run-level "core_model"
 
-- label: "MiMo-Audio E2E Test with H100"
-  agent_pool: mi325_1
-  depends_on: amd-build
-  mirror_hardwares: [amdproduction]
-  grade: Blocking
-  commands:
-    - |
-      timeout 30m bash -c '
-        export VLLM_LOGGING_LEVEL=DEBUG
-        export VLLM_WORKER_MULTIPROC_METHOD=spawn
-        pytest -s -v tests/e2e/online_serving/test_mimo_audio.py -m "core_model" --run-level "core_model"
-      '
+- group: ":card_index_dividers: E2E Test"
+  steps:
+    - label: "MiMo-Audio E2E Test with H100"
+      agent_pool: mi325_1
+      depends_on: amd-build
+      mirror_hardwares: [amdproduction]
+      grade: Blocking
+      commands:
+        - |
+          timeout 30m bash -c '
+            export VLLM_LOGGING_LEVEL=DEBUG
+            export VLLM_WORKER_MULTIPROC_METHOD=spawn
+            pytest -s -v tests/e2e/online_serving/test_mimo_audio.py -m "core_model" --run-level "core_model"
+          '
 
-- label: "Qwen3-TTS E2E Test"
-  agent_pool: mi325_1
-  depends_on: amd-build
-  mirror_hardwares: [amdproduction]
-  grade: Blocking
-  commands:
-    - export VLLM_LOGGING_LEVEL=DEBUG
-    - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-    - export VLLM_ALLOW_LONG_MAX_MODEL_LEN="1"
-    - timeout 30m pytest -s -v tests/e2e/online_serving/test_qwen3_tts_customvoice.py -m "core_model" --run-level "core_model"
+    - label: "Qwen3-TTS E2E Test"
+      agent_pool: mi325_1
+      depends_on: amd-build
+      mirror_hardwares: [amdproduction]
+      grade: Blocking
+      commands:
+        - export VLLM_LOGGING_LEVEL=DEBUG
+        - export VLLM_WORKER_MULTIPROC_METHOD=spawn
+        - export VLLM_ALLOW_LONG_MAX_MODEL_LEN="1"
+        - timeout 30m pytest -s -v tests/e2e/online_serving/test_qwen3_tts_customvoice.py -m "core_model" --run-level "core_model"
 
-- label: "Voxtral-TTS E2E Test"
-  agent_pool: mi325_1
-  depends_on: amd-build
-  mirror_hardwares: [amdproduction]
-  grade: Blocking
-  commands:
-    - |
-      timeout 20m bash -c '
-        export VLLM_LOGGING_LEVEL=DEBUG
-        export VLLM_WORKER_MULTIPROC_METHOD=spawn
-        pytest -s -v tests/e2e/online_serving/test_voxtral_tts.py -m "advanced_model" --run-level "advanced_model"
-        pytest -s -v tests/e2e/offline_inference/test_voxtral_tts.py -m "advanced_model" --run-level "advanced_model"
-      '
+    - label: "Voxtral-TTS E2E Test"
+      agent_pool: mi325_1
+      depends_on: amd-build
+      mirror_hardwares: [amdproduction]
+      grade: Blocking
+      commands:
+        - |
+          timeout 20m bash -c '
+            export VLLM_LOGGING_LEVEL=DEBUG
+            export VLLM_WORKER_MULTIPROC_METHOD=spawn
+            pytest -s -v tests/e2e/online_serving/test_voxtral_tts.py -m "advanced_model" --run-level "advanced_model"
+            pytest -s -v tests/e2e/offline_inference/test_voxtral_tts.py -m "advanced_model" --run-level "advanced_model"
+          '
+
+    - label: "CosyVoice3-TTS E2E Test"
+      agent_pool: mi325_1
+      depends_on: amd-build
+      mirror_hardwares: [amdproduction]
+      grade: Blocking
+      commands:
+        - |
+          timeout 20m bash -c '
+            pytest -s -v tests/e2e/online_serving/test_cosyvoice3_tts_expansion.py -m "full_model" --run-level "core_model"
+          '
 
 # TODO: Bagel test on ROCm is very unstable. @tjtanaa
 # Need to debug before reneable numerical changes across large PRs
@@ -225,14 +248,3 @@ steps:
 #     - export VLLM_WORKER_MULTIPROC_METHOD=spawn
 #     - export VLLM_ROCM_USE_AITER_RMSNORM=0
 #     - timeout 40m pytest -s -v tests/e2e/online_serving/test_bagel.py -m "core_model" --run-level "core_model" -k "rocm"
-
-- label: "CosyVoice3-TTS E2E Test"
-  agent_pool: mi325_1
-  depends_on: amd-build
-  mirror_hardwares: [amdproduction]
-  grade: Blocking
-  commands:
-    - |
-      timeout 20m bash -c '
-        pytest -s -v tests/e2e/online_serving/test_cosyvoice3_tts_expansion.py -m "full_model" --run-level "core_model"
-      '

.buildkite/test-merge.yml

@@ -149,7 +149,7 @@ steps:
         commands:
           - pytest -s -v tests/entrypoints/ -m "advanced_model and cuda and L4" --run-level "advanced_model"
         agents:
-          queue: "gpu_4_queue" # g6.4xlarge instance on AWS, has 1 L4 GPU
+          queue: "gpu_4_queue" # g6.12xlarge instance on AWS, has 4 L4 GPU
         plugins:
           - docker#v5.2.0:
               image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
@@ -164,7 +164,35 @@ steps:
   - group: ":card_index_dividers: E2E Test"
     depends_on: upload-merge-pipeline
     steps:
+      - label: "TTS · Step-Audio2 E2E Test"
+        timeout_in_minutes: 20
+        commands:
+          - |
+            timeout 20m bash -c "
+              export VLLM_LOGGING_LEVEL=DEBUG
+              export VLLM_WORKER_MULTIPROC_METHOD=spawn
+              pytest -s -v tests/e2e/offline_inference/test_step_audio2.py -m 'advanced_model' --run-level 'advanced_model'
+            "
+        agents:
+          queue: "gpu_4_queue"
+        plugins:
+          - docker#v5.2.0:
+              image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
+              always-pull: true
+              propagate-environment: true
+              shm-size: "8gb"
+              environment:
+                - "HF_HOME=/fsx/hf_cache"
+              volumes:
+                - "/fsx/hf_cache:/fsx/hf_cache"
+
       - label: "TTS · Qwen3-TTS CustomVoice Test"
+        source_file_dependencies:
+          - tests/e2e/online_serving/test_qwen3_tts_customvoice.py
+          - tests/e2e/offline_inference/test_qwen3_tts_customvoice.py
+          - vllm_omni/model_executor/models/qwen3_tts/
+          - vllm_omni/model_executor/stage_input_processors/qwen3_tts.py
+          - vllm_omni/deploy/qwen3_tts.yaml
         commands:
           - |
             timeout 20m bash -c "
@@ -173,7 +201,7 @@ steps:
               pytest -s -v tests/e2e/online_serving/test_qwen3_tts_customvoice.py tests/e2e/offline_inference/test_qwen3_tts_customvoice.py -m 'advanced_model and cuda' --run-level 'advanced_model'
             "
         agents:
-          queue: "gpu_4_queue"
+          queue: "gpu_1_queue" # g6.4xlarge instance on AWS, has 1 L4 GPU
         plugins:
           - docker#v5.2.0:
               image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
@@ -186,6 +214,12 @@ steps:
                 - "/fsx/hf_cache:/fsx/hf_cache"
 
       - label: "TTS · Qwen3-TTS Base Test"
+        source_file_dependencies:
+          - tests/e2e/online_serving/test_qwen3_tts_base.py
+          - tests/e2e/offline_inference/test_qwen3_tts_base.py
+          - vllm_omni/model_executor/models/qwen3_tts/
+          - vllm_omni/model_executor/stage_input_processors/qwen3_tts.py
+          - vllm_omni/deploy/qwen3_tts.yaml
         commands:
           - |
             timeout 20m bash -c "
@@ -194,7 +228,34 @@ steps:
               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'
             "
         agents:
-          queue: "gpu_4_queue"
+          queue: "gpu_1_queue" # g6.4xlarge instance on AWS, has 1 L4 GPU
+        plugins:
+          - docker#v5.2.0:
+              image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
+              always-pull: true
+              propagate-environment: true
+              shm-size: "8gb"
+              environment:
+                - "HF_HOME=/fsx/hf_cache"
+              volumes:
+                - "/fsx/hf_cache:/fsx/hf_cache"
+
+      - label: "TTS · Ming-Omni-TTS Test"
+        source_file_dependencies:
+          - tests/e2e/online_serving/test_ming_tts.py
+          - tests/e2e/offline_inference/test_ming_tts.py
+          - vllm_omni/model_executor/models/ming_tts/
+          - vllm_omni/model_executor/stage_input_processors/ming_tts.py
+          - vllm_omni/deploy/ming_tts.yaml
+        timeout_in_minutes: 20
+        commands:
+          - |
+            timeout 20m bash -c "
+              export VLLM_LOGGING_LEVEL=DEBUG
+              pytest -s -v tests/e2e/online_serving/test_ming_tts.py tests/e2e/offline_inference/test_ming_tts.py -m 'advanced_model and cuda' --run-level 'advanced_model'
+            "
+        agents:
+          queue: "gpu_1_queue"
         plugins:
           - docker#v5.2.0:
               image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
@@ -207,6 +268,12 @@ steps:
                 - "/fsx/hf_cache:/fsx/hf_cache"
 
       - label: "TTS · Voxtral-TTS Test"
+        source_file_dependencies:
+          - tests/e2e/offline_inference/test_voxtral_tts.py
+          - tests/e2e/online_serving/test_voxtral_tts.py
+          - vllm_omni/model_executor/models/voxtral_tts/
+          - vllm_omni/model_executor/stage_input_processors/voxtral_tts.py
+          - vllm_omni/deploy/voxtral_tts.yaml
         timeout_in_minutes: 20
         commands:
           - |
@@ -249,6 +316,12 @@ steps:
                       type: DirectoryOrCreate
 
       - label: "TTS · GLM-TTS Test"
+        source_file_dependencies:
+          - tests/e2e/online_serving/test_glm_tts.py
+          - tests/e2e/offline_inference/test_glm_tts.py
+          - vllm_omni/model_executor/models/glm_tts/
+          - vllm_omni/model_executor/stage_input_processors/glm_tts.py
+          - vllm_omni/deploy/glm_tts.yaml
         timeout_in_minutes: 20
         commands:
           - |
@@ -291,6 +364,12 @@ steps:
                       type: DirectoryOrCreate
 
       - label: "Omni · Qwen3-Omni Test"
+        source_file_dependencies:
+          - tests/e2e/offline_inference/test_qwen3_omni.py
+          - tests/e2e/online_serving/test_qwen3_omni.py
+          - vllm_omni/model_executor/models/qwen3_omni/
+          - vllm_omni/model_executor/stage_input_processors/qwen3_omni.py
+          - vllm_omni/deploy/qwen3_omni_moe.yaml
         timeout_in_minutes: 30
         commands:
           - pytest -s -v tests/e2e/offline_inference/test_qwen3_omni.py tests/e2e/online_serving/test_qwen3_omni.py -m 'advanced_model and cuda' --run-level 'advanced_model'
@@ -329,6 +408,11 @@ steps:
                       type: DirectoryOrCreate
 
       - label: "Omni · Mimo Audio Test"
+        source_file_dependencies:
+          - tests/e2e/online_serving/test_mimo_audio.py
+          - vllm_omni/model_executor/models/mimo_audio/
+          - vllm_omni/model_executor/stage_input_processors/mimo_audio.py
+          - vllm_omni/deploy/mimo_audio.yaml
         timeout_in_minutes: 30
         commands:
           - pytest -s -v tests/e2e/online_serving/test_mimo_audio.py -m 'advanced_model and cuda' --run-level 'advanced_model'
@@ -366,37 +450,19 @@ steps:
                       path: /mnt/hf-cache
                       type: DirectoryOrCreate
 
-      - label: "Omni · VoxCPM2 Test"
-        timeout_in_minutes: 20
-        commands:
-          - |
-            timeout 20m bash -c "
-              export VLLM_LOGGING_LEVEL=DEBUG
-              export VLLM_WORKER_MULTIPROC_METHOD=spawn
-              pytest -s -v tests/e2e/offline_inference/test_voxcpm2.py -m 'advanced_model and cuda' --run-level 'advanced_model'
-            "
-        agents:
-          queue: "gpu_1_queue"
-        plugins:
-          - docker#v5.2.0:
-              image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
-              always-pull: true
-              propagate-environment: true
-              shm-size: "8gb"
-              environment:
-                - "HF_HOME=/fsx/hf_cache"
-                - "HF_TOKEN"
-              volumes:
-                - "/fsx/hf_cache:/fsx/hf_cache"
-
-      - label: "TTS · VoxCPM2 Online Test"
+      - 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/online_serving/test_voxcpm2_tts.py -m 'advanced_model and cuda' --run-level 'advanced_model'
+              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'
             "
         agents:
           queue: "gpu_1_queue"
@@ -413,6 +479,12 @@ steps:
                 - "/fsx/hf_cache:/fsx/hf_cache"
 
       - label: "Omni · Higgs-Audio V2 Test"
+        source_file_dependencies:
+          - tests/e2e/online_serving/test_higgs_audio_v2.py
+          - tests/e2e/offline_inference/test_higgs_audio_v2.py
+          - vllm_omni/model_executor/models/higgs_audio_v2/
+          - vllm_omni/model_executor/stage_input_processors/higgs_audio_v2.py
+          - vllm_omni/deploy/higgs_audio_v2.yaml
         timeout_in_minutes: 30
         commands:
           - |
@@ -438,6 +510,9 @@ steps:
                 - "/fsx/hf_cache:/fsx/hf_cache"
 
       - 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'
@@ -476,6 +551,9 @@ steps:
                       type: DirectoryOrCreate
 
       - label: "Diffusion · Longcat Image Edit Test"
+        source_file_dependencies:
+          - tests/e2e/online_serving/test_longcat_image_edit.py
+          - vllm_omni/diffusion/models/longcat_image/
         timeout_in_minutes: 20
         commands:
           - pytest -s -v tests/e2e/online_serving/test_longcat_image_edit.py -m 'advanced_model and cuda' --run-level 'advanced_model'
@@ -514,6 +592,9 @@ steps:
                       type: DirectoryOrCreate
 
       - label: "Diffusion · Longcat Image Test"
+        source_file_dependencies:
+          - tests/e2e/online_serving/test_longcat_image.py
+          - vllm_omni/diffusion/models/longcat_image/
         timeout_in_minutes: 20
         commands:
           - pytest -s -v tests/e2e/online_serving/test_longcat_image.py -m 'advanced_model and cuda' --run-level 'advanced_model'
@@ -552,6 +633,13 @@ steps:
                       type: DirectoryOrCreate
 
       - 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'
@@ -591,6 +679,9 @@ steps:
 
 
       - label: "Diffusion · Qwen Image Test"
+        source_file_dependencies:
+          - tests/e2e/online_serving/test_qwen_image.py
+          - vllm_omni/diffusion/models/qwen_image/
         commands:
           - |
             timeout 40m bash -c '
@@ -632,6 +723,9 @@ steps:
                       type: DirectoryOrCreate
 
       - 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 "
@@ -673,6 +767,10 @@ steps:
                       type: DirectoryOrCreate
 
       - 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/
         commands:
           - |
             timeout 40m bash -c "
@@ -714,6 +812,9 @@ steps:
                       type: DirectoryOrCreate
 
       - label: "Diffusion · HunyuanVideo 1.5 Test"
+        source_file_dependencies:
+          - tests/e2e/online_serving/test_hunyuan_video_15.py
+          - vllm_omni/diffusion/models/hunyuan_video/
         timeout_in_minutes: 40
         commands:
           - |
@@ -755,6 +856,10 @@ steps:
                       type: DirectoryOrCreate
 
       - 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'
@@ -772,6 +877,10 @@ steps:
                 - "/fsx/hf_cache:/fsx/hf_cache"
 
       - label: "Diffusion · FLUX.2 Klein Test"
+        source_file_dependencies:
+          - tests/e2e/offline_inference/test_flux2_klein.py
+          - tests/e2e/online_serving/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 tests/e2e/online_serving/test_flux2_klein.py -m 'advanced_model and cuda and diffusion' --run-level 'advanced_model'
@@ -789,6 +898,9 @@ steps:
                 - "/fsx/hf_cache:/fsx/hf_cache"
 
       - label: "Diffusion · SD3 Test"
+        source_file_dependencies:
+          - tests/e2e/online_serving/test_sd3.py
+          - vllm_omni/diffusion/models/sd3/
         timeout_in_minutes: 45
         commands:
           - |

.buildkite/test-nightly.yml

@@ -568,6 +568,55 @@ steps:
                       path: /mnt/hf-cache
                       type: DirectoryOrCreate
 
+  - label: ":full_moon: TTS · Higgs-Audio-V3 · Perf Test"
+    key: nightly-tts-performance-higgs-audio-v3
+    depends_on: upload-nightly-pipeline
+    if: build.env("NIGHTLY") == "1" || build.pull_request.labels includes "nightly-test" || build.pull_request.labels includes "tts-test"
+    timeout_in_minutes: 180
+    commands:
+      - export BENCHMARK_DIR=tests/dfx/perf/results
+      - export VLLM_USE_DEEP_GEMM="0"
+      - export VLLM_MOE_USE_DEEP_GEMM="0"
+      - |
+        set +e
+        pytest -s -v tests/dfx/perf/scripts/run_benchmark.py --test-config-file tests/dfx/perf/tests/test_higgs_audio_v3.json
+        EXIT=$$?
+        buildkite-agent artifact upload "tests/dfx/perf/results/*.json"
+        exit $$EXIT
+    agents:
+      queue: "mithril-h100-pool"
+    plugins:
+      - kubernetes:
+          podSpec:
+            containers:
+              - image: 936637512419.dkr.ecr.us-west-2.amazonaws.com/vllm-ci-pull-through-cache/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
+                resources:
+                  limits:
+                    nvidia.com/gpu: 1
+                volumeMounts:
+                  - name: devshm
+                    mountPath: /dev/shm
+                  - name: hf-cache
+                    mountPath: /root/.cache/huggingface
+                env:
+                  - name: HF_HOME
+                    value: /root/.cache/huggingface
+                  - name: HF_TOKEN
+                    valueFrom:
+                      secretKeyRef:
+                        name: hf-token-secret
+                        key: token
+            nodeSelector:
+              node.kubernetes.io/instance-type: gpu-h100-sxm
+            volumes:
+              - name: devshm
+                emptyDir:
+                  medium: Memory
+              - name: hf-cache
+                hostPath:
+                  path: /mnt/hf-cache
+                  type: DirectoryOrCreate
+
   # Diffusion X2I suite: x2i / x2a / x2t and related non-video paths; x2v is only in "Diffusion X2V Model Test" below.
   - group: ":card_index_dividers: Diffusion X2I(&A&T) Model Test"
     key: nightly-diffusion-x2iat-group
@@ -591,6 +640,7 @@ steps:
             tests/e2e/online_serving/test_longcat_image_edit_expansion.py
             tests/e2e/online_serving/test_flux_2_dev_expansion.py
             tests/e2e/online_serving/test_bagel_expansion.py
+            tests/e2e/offline_inference/test_sensenova_u1_expansion.py
             -m "full_model and diffusion and H100 and not distributed_cuda"
             --run-level "full_model"
         agents:
@@ -1462,6 +1512,7 @@ steps:
       - nightly-omni-performance-vllm-text
       - nightly-omni-performance-multi-replicas
       - nightly-tts-performance
+      - nightly-tts-performance-higgs-audio-v3
       - nightly-diffusion-x2iat-performance-qwen-image
       - nightly-diffusion-x2iat-performance-qwen-image-edit
       - nightly-diffusion-x2iat-performance-qwen-image-edit-2511
@@ -1475,6 +1526,7 @@ steps:
       - export DEFAULT_INPUT_DIR=tests/dfx/perf/results
       - export DEFAULT_OUTPUT_DIR=tests/dfx/perf/results
       - buildkite-agent artifact download "tests/dfx/perf/results/*.json" . --step nightly-tts-performance
+      - buildkite-agent artifact download "tests/dfx/perf/results/*.json" . --step nightly-tts-performance-higgs-audio-v3
       - buildkite-agent artifact download "tests/dfx/perf/results/*.json" . --step nightly-omni-performance-no-async-chunk
       - buildkite-agent artifact download "tests/dfx/perf/results/*.json" . --step nightly-omni-performance-async-chunk
       - buildkite-agent artifact download "tests/dfx/perf/results/*.json" . --step nightly-omni-performance-vllm-text

.buildkite/test-ready.yml

@@ -215,10 +215,78 @@ steps:
                   path: /mnt/hf-cache
                   type: DirectoryOrCreate
 
+  - group: ":card_index_dividers: Entrypoints Test"
+    depends_on: upload-ready-pipeline
+    steps:
+      - label: "Entrypoints Test with H100"
+        commands:
+          - |
+            timeout 30m bash -c "
+              export VLLM_WORKER_MULTIPROC_METHOD=spawn
+              pytest -sv tests/entrypoints/ -m 'core_model and cuda and H100' --run-level 'core_model'
+            "
+        agents:
+          queue: "mithril-h100-pool"
+        plugins:
+          - kubernetes:
+              podSpec:
+                containers:
+                  - image: 936637512419.dkr.ecr.us-west-2.amazonaws.com/vllm-ci-pull-through-cache/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
+                    resources:
+                      limits:
+                        nvidia.com/gpu: 2
+                    volumeMounts:
+                      - name: devshm
+                        mountPath: /dev/shm
+                      - name: hf-cache
+                        mountPath: /root/.cache/huggingface
+                    env:
+                      - name: HF_HOME
+                        value: /root/.cache/huggingface
+                nodeSelector:
+                  node.kubernetes.io/instance-type: gpu-h100-sxm
+                volumes:
+                  - name: devshm
+                    emptyDir:
+                      medium: Memory
+                  - name: hf-cache
+                    hostPath:
+                      path: /mnt/hf-cache
+                      type: DirectoryOrCreate
+
+      - label: "Entrypoints Test with L4"
+        commands:
+          - |
+            timeout 30m bash -c "
+              export VLLM_WORKER_MULTIPROC_METHOD=spawn
+              pytest -sv tests/entrypoints/ -m 'core_model and cuda and L4' --run-level 'core_model'
+            "
+        agents:
+          queue: "gpu_1_queue"
+        plugins:
+          - docker#v5.2.0:
+              image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
+              always-pull: true
+              propagate-environment: true
+              shm-size: "8gb"
+              environment:
+                - "HF_HOME=/fsx/hf_cache"
+                - "HF_TOKEN"
+              volumes:
+                - "/fsx/hf_cache:/fsx/hf_cache"
+
   - group: ":card_index_dividers: E2E Test"
     depends_on: upload-ready-pipeline
     steps:
       - label: "RLHF · VeRL-Omni E2E Test"
+        source_file_dependencies:
+          - tests/e2e/offline_inference/rlhf_test/test_verl_omni_e2e.py
+          - tests/e2e/offline_inference/custom_pipeline/
+          - vllm_omni/diffusion/
+          - vllm_omni/engine/
+          - vllm_omni/entrypoints/
+          - vllm_omni/inputs/
+          - vllm_omni/outputs/
         commands:
           - |
             timeout 20m bash -c "
@@ -239,6 +307,11 @@ steps:
                 - "/fsx/hf_cache:/fsx/hf_cache"
 
       - label: "Omni · Qwen3-Omni Test"
+        source_file_dependencies:
+          - tests/e2e/online_serving/test_qwen3_omni.py
+          - vllm_omni/model_executor/models/qwen3_omni/
+          - vllm_omni/model_executor/stage_input_processors/qwen3_omni.py
+          - vllm_omni/deploy/qwen3_omni_moe.yaml
         commands:
           - |
             timeout 20m bash -c "
@@ -273,73 +346,18 @@ steps:
                       path: /mnt/hf-cache
                       type: DirectoryOrCreate
 
-      - label: "Omni · Qwen3-Omni Realtime WebSocket"
-        commands:
-          - |
-            timeout 30m bash -c "
-              export VLLM_WORKER_MULTIPROC_METHOD=spawn
-              pytest -s -v tests/entrypoints/openai_api/test_qwen3_omni_realtime_websocket.py::TestQwen3OmniRealtimeWebSocket::test_streaming_audio_input_pcm_output_async_chunk_smoke -m 'core_model' --run-level 'core_model'
-            "
-        agents:
-          queue: "mithril-h100-pool"
-        plugins:
-          - kubernetes:
-              podSpec:
-                containers:
-                  - image: 936637512419.dkr.ecr.us-west-2.amazonaws.com/vllm-ci-pull-through-cache/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
-                    resources:
-                      limits:
-                        nvidia.com/gpu: 2
-                    volumeMounts:
-                      - name: devshm
-                        mountPath: /dev/shm
-                      - name: hf-cache
-                        mountPath: /root/.cache/huggingface
-                    env:
-                      - name: HF_HOME
-                        value: /root/.cache/huggingface
-                nodeSelector:
-                  node.kubernetes.io/instance-type: gpu-h100-sxm
-                volumes:
-                  - name: devshm
-                    emptyDir:
-                      medium: Memory
-                  - name: hf-cache
-                    hostPath:
-                      path: /mnt/hf-cache
-                      type: DirectoryOrCreate
-
-      - label: "Omni · VoxCPM2 Test"
-        timeout_in_minutes: 20
-        commands:
-          - |
-            timeout 20m bash -c "
-              export VLLM_LOGGING_LEVEL=DEBUG
-              export VLLM_WORKER_MULTIPROC_METHOD=spawn
-              pytest -s -v tests/e2e/offline_inference/test_voxcpm2.py -m 'core_model' --run-level 'core_model'
-            "
-        agents:
-          queue: "gpu_1_queue"
-        plugins:
-          - docker#v5.2.0:
-              image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
-              always-pull: true
-              propagate-environment: true
-              shm-size: "8gb"
-              environment:
-                - "HF_HOME=/fsx/hf_cache"
-                - "HF_TOKEN"
-              volumes:
-                - "/fsx/hf_cache:/fsx/hf_cache"
-
-      - label: "TTS · VoxCPM2 Online Test"
+      - 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 and tts' --run-level 'core_model'
+              pytest -s -v tests/e2e/online_serving/test_voxcpm2_tts.py -m 'core_model' --run-level 'core_model'
             "
         agents:
           queue: "gpu_1_queue"
@@ -356,6 +374,11 @@ steps:
                 - "/fsx/hf_cache:/fsx/hf_cache"
 
       - label: "TTS · Qwen3-TTS CustomVoice Test"
+        source_file_dependencies:
+          - tests/e2e/online_serving/test_qwen3_tts_customvoice.py
+          - vllm_omni/model_executor/models/qwen3_tts/
+          - vllm_omni/model_executor/stage_input_processors/qwen3_tts.py
+          - vllm_omni/deploy/qwen3_tts.yaml
         commands:
           - |
             timeout 20m bash -c "
@@ -363,7 +386,7 @@ steps:
               pytest -s -v tests/e2e/online_serving/test_qwen3_tts_customvoice.py -m 'core_model' --run-level 'core_model'
             "
         agents:
-          queue: "gpu_4_queue"
+          queue: "gpu_1_queue"
         plugins:
           - docker#v5.2.0:
               image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
@@ -377,6 +400,11 @@ steps:
                 - "/fsx/hf_cache:/fsx/hf_cache"
 
       - label: "TTS · Qwen3-TTS Base Test"
+        source_file_dependencies:
+          - tests/e2e/online_serving/test_qwen3_tts_base.py
+          - vllm_omni/model_executor/models/qwen3_tts/
+          - vllm_omni/model_executor/stage_input_processors/qwen3_tts.py
+          - vllm_omni/deploy/qwen3_tts.yaml
         commands:
           - |
             timeout 20m bash -c "
@@ -386,7 +414,35 @@ steps:
               pytest -s -v tests/e2e/online_serving/test_qwen3_tts_base.py -m 'core_model' --run-level 'core_model'
             "
         agents:
-          queue: "gpu_4_queue"
+          queue: "gpu_1_queue"
+        plugins:
+          - docker#v5.2.0:
+              image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
+              always-pull: true
+              propagate-environment: true
+              shm-size: "8gb"
+              environment:
+                - "HF_HOME=/fsx/hf_cache"
+                - "HF_TOKEN"
+              volumes:
+                - "/fsx/hf_cache:/fsx/hf_cache"
+
+      - label: "TTS · Higgs-Audio-V3 Test"
+        source_file_dependencies:
+          - tests/e2e/online_serving/test_higgs_audio_v3.py
+          - vllm_omni/model_executor/models/higgs_audio_v3/
+          - vllm_omni/model_executor/stage_input_processors/higgs_audio_v3.py
+          - vllm_omni/deploy/higgs_multimodal_qwen3.yaml
+        commands:
+          - |
+            timeout 20m bash -c "
+              export VLLM_WORKER_MULTIPROC_METHOD=spawn
+              export VLLM_USE_DEEP_GEMM=0
+              export VLLM_MOE_USE_DEEP_GEMM=0
+              pytest -s -v tests/e2e/online_serving/test_higgs_audio_v3.py -m 'core_model' --run-level 'core_model'
+            "
+        agents:
+          queue: "gpu_1_queue"
         plugins:
           - docker#v5.2.0:
               image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
@@ -400,6 +456,9 @@ steps:
                 - "/fsx/hf_cache:/fsx/hf_cache"
 
       - 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 "
@@ -435,6 +494,12 @@ steps:
                       type: DirectoryOrCreate
 
       - 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 "
@@ -476,6 +541,9 @@ steps:
                       type: DirectoryOrCreate
 
       - label: "Diffusion · Wan22 Test"
+        source_file_dependencies:
+          - tests/e2e/online_serving/test_wan22_t2v.py
+          - vllm_omni/diffusion/models/wan2_2/
         commands:
           - |
             timeout 40m bash -c "
@@ -518,6 +586,9 @@ steps:
 
 
       - label: "Diffusion · Qwen Image Test"
+        source_file_dependencies:
+          - tests/e2e/online_serving/test_qwen_image.py
+          - vllm_omni/diffusion/models/qwen_image/
         commands:
           - |
             timeout 40m bash -c "
@@ -559,6 +630,9 @@ steps:
                       type: DirectoryOrCreate
 
       - 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 "

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

@@ -6,6 +6,13 @@
 {% set docker_image_amd = "rocm/vllm-omni:$BUILDKITE_COMMIT" %}
 {% set default_working_dir = "/app/vllm-omni" %}
 
+{% if env %}
+env:
+{% for name in env %}
+  {{ name }}: "{{ env[name] }}"
+{% endfor %}
+{% endif %}
+steps:
   - group: "AMD Tests"
     depends_on: ~
     steps:
@@ -31,8 +38,15 @@
         agents:
           queue: amd-cpu
 
-    {% for step in steps %}
-    {% if step.mirror_hardwares and mirror_hw in step.mirror_hardwares %}
+{% for step_group in steps %}
+{% if step_group.group %}
+  - group: "{{ step_group.group }}"
+{% if step_group.depends_on %}
+    depends_on: {{ step_group.depends_on }}
+{% endif %}
+    steps:
+{% for step in step_group.steps %}
+{% if step.mirror_hardwares and mirror_hw in step.mirror_hardwares %}
       - label: "{{ step.agent_pool }}: {{ step.label }}"
         depends_on: amd-build
         agents:
@@ -46,8 +60,23 @@
         command: bash .buildkite/scripts/hardware_ci/run-amd-test.sh
         env:
           DOCKER_BUILDKIT: "1"
+{% if step.env %}
+{% for name in step.env %}
+          {{ name }}: "{{ step.env[name] }}"
+{% endfor %}
+{% endif %}
           TEST_COMMAND: |-
             (command rocm-smi || true) && cd {{ (step.working_dir or default_working_dir) | safe }}
+{% if env %}
+{% for name in env %}
+            export {{ name }}="{{ env[name] }}"
+{% endfor %}
+{% endif %}
+{% if step.env %}
+{% for name in step.env %}
+            export {{ name }}="{{ step.env[name] }}"
+{% endfor %}
+{% endif %}
 {% if "mi250" in step.agent_pool %}
             python3 -m pip uninstall -y amd-aiter
 {% endif %}
@@ -61,5 +90,52 @@
         {% else %}
         soft_fail: true
         {% endif%}
+{% endif %}
+{% endfor %}
+{% elif step_group.mirror_hardwares and mirror_hw in step_group.mirror_hardwares %}
+{% set step = step_group %}
+  - label: "{{ step.agent_pool }}: {{ step.label }}"
+    depends_on: amd-build
+    agents:
+      {% if step.agent_pool %}
+      queue: amd_{{ step.agent_pool }}
+      {% else %}
+      queue: amd_mi325_1
+      {% endif %}
+{% set cmd_body = (step.command or (step.commands | join("\n"))) | trim %}
+{% set indented_cmd = cmd_body | replace("\n", "\n        ") %}
+    command: bash .buildkite/scripts/hardware_ci/run-amd-test.sh
+    env:
+      DOCKER_BUILDKIT: "1"
+{% if step.env %}
+{% for name in step.env %}
+      {{ name }}: "{{ step.env[name] }}"
+{% endfor %}
+{% endif %}
+      TEST_COMMAND: |-
+        (command rocm-smi || true) && cd {{ (step.working_dir or default_working_dir) | safe }}
+{% if env %}
+{% for name in env %}
+        export {{ name }}="{{ env[name] }}"
+{% endfor %}
+{% endif %}
+{% if step.env %}
+{% for name in step.env %}
+        export {{ name }}="{{ step.env[name] }}"
+{% endfor %}
+{% endif %}
+{% if "mi250" in step.agent_pool %}
+        python3 -m pip uninstall -y amd-aiter
+{% endif %}
+        {{ indented_cmd | safe }}
+    priority: 100
+    {% if step.timeout_in_minutes %}
+    timeout_in_minutes: {{ step.timeout_in_minutes }}
     {% endif %}
-    {% endfor %}
+    {% if step.grade and step.grade == "Blocking" %}
+    soft_fail: false
+    {% else %}
+    soft_fail: true
+    {% endif%}
+{% endif %}
+{% endfor %}

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

@@ -0,0 +1,97 @@
+---
+name: precheck-pr
+description: Self-check your branch before creating a PR — catch dead code, verify accuracy/perf claims, validate PR title format, and confirm merge readiness. Use when the user says "precheck", "self review", "pre-submit check", or "check my PR before I open it." Never posts to GitHub.
+---
+
+# PR Pre-Check
+
+Self-review your branch before creating a PR against `vllm-project/vllm-omni`. Two modes: **quick** catches showstoppers, **full** does a thorough maintainer-grade review. Never posts to GitHub; the report is for the contributor's terminal only.
+
+## Mode Selection
+
+| Mode | When | Time |
+|------|------|------|
+| **Quick** | About to push, final sanity check | ~3 min |
+| **Full** | Ready for review, want maintainer-level scan | ~10 min |
+
+Default to quick if unsure. Run full before marking a PR "ready for review."
+
+## Workflow
+
+### Step 1: Detect Base Branch
+
+```bash
+BASE_SHA=$(git merge-base HEAD origin/main 2>/dev/null \
+         || git merge-base HEAD main 2>/dev/null \
+         || echo origin/main)
+echo "diffing against ${BASE_SHA}"
+git diff --name-only ${BASE_SHA}...HEAD
+```
+
+### Step 2: Validate PR Title
+
+Check the most recent commit message (or branch name if no commit yet) against the project convention. Valid prefixes:
+
+| Prefix | Applies to |
+|--------|-----------|
+| `[Bugfix]` | Bug fixes |
+| `[CI/Build]` | Build or CI improvements |
+| `[Doc]` | Documentation changes |
+| `[Model]` | New/improved models (include model name) |
+| `[Frontend]` | Frontend changes (API server, OmniLLM class, etc.) |
+| `[Kernel]` | CUDA/kernel changes |
+| `[Core]` | Core logic changes (OmniProcessor, OmniARScheduler, etc.) |
+| `[Hardware][Vendor]` | Hardware-specific (e.g., `[Hardware][Ascend]`) |
+| `[Misc]` | Other changes (use sparingly) |
+
+✗ if: missing prefix, wrong case (`[bugfix]`), or WIP/Draft in title.
+⚠ if: `[Model]` prefix without the model identifier (e.g., `[Model] Add new model` — should be `[Model] Add <ModelName> ...`).
+
+### Step 3: Categorize the PR
+
+| Diff contains | PR type |
+|---------------|---------|
+| New files under `vllm_omni/model_executor/models/<name>/` | **New Model** |
+| Changes to `vllm_omni/diffusion/` | **Diffusion Model** |
+| `[Bugfix]` prefix or single-file fix | **Bug Fix** |
+| Perf/benchmark/throughput claims in commit msg or diff | **Performance** |
+| Everything else | **General** |
+
+If multiple rows apply (e.g., a diffusion model is also a new model), union the checklists.
+
+### Step 4: Run Checklist
+
+Ask: "Quick mode or full mode?" Then walk the checklist for the detected PR type from [references/checklists.md](references/checklists.md). Each item produces ✓, ✗, or ⚠.
+
+### Step 5: Print Report
+
+```
+Pre-check report for <branch>
+
+  Mode: quick | full
+  Type: <new-model | diffusion-model | bug-fix | perf | general>
+
+  Dimension          Result
+  ─────────────────  ──────
+  PR title format    ✓
+  PR desc integrity  ✓
+  Registry/config    ✓
+  Dead code          ⚠ 2 warnings
+  Accuracy           ✓
+  Benchmark          ✗ missing software versions
+
+  Verdict: 1 blocking | 2 warnings | recommend fixing ✗ before PR
+```
+
+**Severity:**
+
+| Mark | Meaning |
+|------|---------|
+| ✗ | Blocking — fix before opening PR |
+| ⚠ | Warning — consider fixing |
+| ✓ | Pass |
+| — | Skipped (not applicable) |
+
+### Stop Here
+
+Do not post comments, open PRs, or modify files. The report is for the contributor's terminal only.

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

@@ -0,0 +1,135 @@
+# Pre-Check Checklists
+
+Each item is a concrete check to run against the diff. ✗ = fix before PR. ⚠ = consider fixing.
+
+---
+
+## All PRs: Title Format (check first)
+
+PR title must follow the project convention documented in `docs/contributing/README.md`:
+
+| Prefix | Applies to |
+|--------|-----------|
+| `[Bugfix]` | Bug fixes |
+| `[CI/Build]` | Build or CI improvements |
+| `[Doc]` | Documentation changes |
+| `[Model]` | New/improved models (include model name) |
+| `[Frontend]` | Frontend changes (API server, OmniLLM class, etc.) |
+| `[Kernel]` | CUDA/kernel changes |
+| `[Core]` | Core logic changes (OmniProcessor, OmniARScheduler, etc.) |
+| `[Hardware][Vendor]` | Hardware-specific (e.g., `[Hardware][Ascend]`) |
+| `[Misc]` | Other changes (use sparingly) |
+
+**Examples:**
+```
+✓ [Model] Add MiniCPM-o 4.5 support
+✓ [Bugfix] Fix Ovis image text encoder dtype
+✓ [Core] Qwen3-Omni performance optimization
+✓ [Hardware][Ascend] Fix VoxCPM2 on Ascend
+✗ Fix Ovis image text encoder dtype              ← missing prefix
+✗ [bugfix] Fix Ovis image text encoder           ← wrong case
+✗ [WIP] Add model                                ← WIP in title
+```
+
+- [ ] **Title has valid prefix** — ✗ if missing prefix, wrong case, or WIP/Draft in title. For `[Model]` PRs, the title MUST also include the model identifier (e.g., `[Model] Add <ModelName> ...`). ⚠ if missing the model name.
+
+---
+
+## Bug Fix PRs
+
+### Quick
+
+- [ ] **PR body has 4 sections:** what broke (error message), repro steps, root cause, fix description
+- [ ] **Fix is minimal:** diff touches only files directly related to the bug. No unrelated refactors or style changes.
+- [ ] **Repro is runnable:** the PR body includes a copy-paste command that reproduces the bug
+- [ ] **Branch is rebased:** `git merge-base HEAD origin/main` is recent, no merge conflicts
+
+### Full (adds)
+
+- [ ] **Regression test exists:** `git diff --name-only` includes at least one `tests/` file
+- [ ] **Regression test reproduces the original error:** the test would fail on main, pass on this branch
+- [ ] **Fix matches root cause exactly:** no "fix the symptom + something else" — if the root cause is dtype, the fix is dtype, not dtype + reformatting
+- [ ] **No silent failure risk:** no bare `except: pass`, no `try: ... except: return None`, no empty fallback added
+- [ ] **Upstream pattern match:** the fix follows the same pattern as existing code for similar cases (grep for analogous `from_pretrained` calls, etc.)
+- [ ] **Environment documented:** torch/transformers/vllm versions listed if the bug is version-dependent
+
+---
+
+## Performance PRs
+
+### Quick
+
+- [ ] **Before/after numbers in PR body:** at minimum, one metric (RTF, RPS, latency, VRAM) with main vs PR values
+- [ ] **Hardware specified:** GPU model, count, VRAM. "L4" or "H20" is sufficient. "tested on GPU" is not.
+- [ ] **Concurrency scaling:** if the optimization claims better throughput, numbers at 2+ concurrency levels
+- [ ] **No unexplained regressions:** if any metric goes down at any concurrency level, explain why
+- [ ] **Branch is rebased:** `git merge-base HEAD origin/main` is recent, no merge conflicts
+
+### Full (adds)
+
+- [ ] **Software versions:** torch, CUDA, vllm, vllm-omni versions in PR body
+- [ ] **Warmup stated:** "1 warmup + 3 measured" or equivalent
+- [ ] **Stddev or range:** not just a single best run. "average of 2 rounds" is minimum.
+- [ ] **VRAM measured:** peak GPU memory before and after
+- [ ] **Benchmark script checked in:** under `tests/` or `examples/`, runnable by reviewer
+- [ ] **Exact command line in PR:** not pseudocode, not "run the benchmark"
+- [ ] **Pytest summary included:** `36 passed, 3 skipped, 17 warnings in 2897.66s` — proves the whole suite passes
+- [ ] **Config changes documented:** if `max_num_seqs`, `gpu_memory_utilization`, or deploy YAML changed, old and new values stated with reasoning
+
+---
+
+## New Model PRs
+
+### Quick
+
+- [ ] **PR body file list matches `git diff --name-only`:** every file path in the description exists in the diff
+- [ ] **Registry entries match claimed count:** count `_OMNI_MODELS` or `_DIFFUSION_MODELS` entries added vs claimed in body
+- [ ] **Pipeline config `custom_process_input_func` paths resolve:** each referenced module.function exists
+- [ ] **`__init__.py` exports match registry:** each registered class appears in `__init__.py` and `__all__`
+- [ ] **Shell script modes match README:** `run_curl.sh` supports every mode the README documents
+- [ ] **Branch is rebased:** `git merge-base HEAD origin/main` is recent, no merge conflicts
+
+### Full (adds)
+
+- [ ] **Dead code scan (5 patterns):**
+  1. Dead `forward()` — defined but never called through the pipeline path
+  2. Dead factory/builder functions — instantiated nowhere
+  3. Dead wrappers — thin wrappers with no callers
+  4. Dead branch guards — `if version > X` where X is always satisfied
+  5. Unused parameters — `__init__` args never accessed
+- [ ] **Copy-paste detection:** no string constant defined in 3+ files, no cross-module validation duplication, no near-identical shape coercion fns
+- [ ] **Import hygiene:** no re-export-only imports, no module-level side effects without docs, no `import os` in function body when already at top. Function-body imports are acceptable when the import is only reachable from a closure or factory return value (i.e., `import` would otherwise execute at module load even though the user may never trigger that path). Either lift to top-of-file or annotate with `# noqa: PLC0415`.
+- [ ] **Accuracy:** at least one test compares output against reference; audio = valid WAV at correct sample rate; determinism verified if seed claimed. ⚠ if missing and not explicitly deferred to a follow-up PR or RFC section; — if explicitly deferred (the deferral target SHOULD be linked in the PR body).
+- [ ] **Performance:** hardware, software versions, warmup, and metrics (RTF, VRAM, RPS, latency) stated. ⚠ if missing and not explicitly deferred; — if explicitly deferred.
+- [ ] **Benchmark settings:** model config (TP, PP, max_model_len, enforce_eager, quant), runtime config (batch, input/output spec), environment versions
+- [ ] **Benchmark script checked in + exact command line + pytest summary**
+
+---
+
+## Diffusion Model PRs
+
+Diffusion models live under `vllm_omni/diffusion/`. In addition to the New Model checklist above, verify:
+
+- [ ] **Transformer adapter follows contract:** inherits from `nn.Module`, implements `load_weights()`, uses `vllm_omni.diffusion.attention.layer.Attention`
+- [ ] **Pipeline has both offline and online paths:** `forward()` for batch generation, streaming path for online serving
+- [ ] **Cache-DiT integration:** at least one acceleration (TeaCache, FBCache, etc.) is wired up. ✗ for new **top-level** pipelines. For **subclass** pipelines that inherit `__call__` from a parent where Cache-DiT is already absent, downgrade to ⚠ and recommend filing a follow-up issue against the parent pipeline family.
+- [ ] **Parallelism config:** TP/SP/USP/CFG-Parallel options exposed in pipeline config
+- [ ] **Docs table updated:** model added to `docs/models/supported_models.md` with correct metadata
+- [ ] **E2E test exists:** at least one test exercises the full generate path
+
+---
+
+## General PRs
+
+### Quick
+
+- [ ] **PR body matches diff:** description doesn't claim changes not in the diff
+- [ ] **Branch is rebased:** `git merge-base HEAD origin/main` is recent, no merge conflicts
+- [ ] **CI gates passing:** DCO, pre-commit, build — check `gh pr view --json statusCheckRollup` if a PR already exists
+
+### Full (adds)
+
+- [ ] **No new dead code in changed files:** scan for unreachable paths, unused imports
+- [ ] **Test coverage:** if core code changed (`engine/`, `stages/`, `connectors/`), tests are added or existing tests cover the path
+- [ ] **Import hygiene:** no redundant imports added
+- [ ] **No unrelated changes:** diff scope matches PR description

.claude/skills/quantization/SKILL.md

@@ -0,0 +1,103 @@
+---
+name: quantization
+description: Work on vLLM-Omni quantization for diffusion, autoregressive, omni, or multi-stage models. Use when choosing or adding methods such as fp8, int8, gguf, mxfp8, mxfp4, mxfp4_dualscale, ModelOpt, AutoRound, INC, msModelSlim, awq, or gptq; debugging quantized loading; or validating memory, speed, and output quality.
+---
+
+# vLLM-Omni Quantization
+
+Use this skill for quantization work in `vllm-omni`. Start from the local
+code and docs, then pick the smallest validated path for the model, method,
+hardware, and quality target.
+
+## First Checks
+
+Before changing code or recommending a command, identify:
+
+- Model and task: AR text, omni/TTS, text-to-image, text-to-video, or multi-stage diffusion.
+- Hardware backend: CUDA generation, Ascend NPU, Intel XPU, or another platform.
+- Quantization mode: online from BF16/FP16 weights, pre-quantized checkpoint, runtime KV/FA quantization, or per-component routing.
+- Scope: transformer/DiT, thinker language model, text encoder, VAE, audio/vision encoder, talker, or a specific stage.
+- Validation target: memory reduction, latency/throughput, output quality, or all three.
+
+Do not assume a method is supported just because the CLI accepts the string.
+Check `docs/user_guide/quantization/` and the closest implementation first.
+
+## Quick Decision
+
+| Task | Start With |
+|------|------------|
+| Choose a method or command | `references/methods.md` and `references/modality-compat.md` |
+| Use `build_quant_config()` or per-component routing | `references/methods.md` |
+| Work on diffusion quantization | `references/diffusion.md` |
+| Add quantization to a new model | `references/adding-models.md` |
+| Convert or load ModelOpt FP8 checkpoints | `references/modelopt-fp8.md` |
+| Debug ModelOpt, GGUF, AutoRound, MXFP, or serialized Int8 loading | `references/diffusion.md` and method docs under `docs/user_guide/quantization/` |
+
+## Current Runtime Shape
+
+The unified entrypoint is:
+
+```python
+from vllm_omni.quantization import build_quant_config
+```
+
+It supports method strings, flat method dictionaries, per-component
+dictionaries, existing `QuantizationConfig` objects, and `None`. The factory
+delegates generic methods to upstream `vllm` and keeps vLLM-Omni overrides for
+diffusion or omni-specific routing:
+
+- `gguf`
+- `int8`
+- `mxfp8`
+- `mxfp4`
+- `mxfp4_dualscale`
+- `inc`, `auto-round`, `auto_round`
+- ModelOpt auto-detected configs: `modelopt`, `modelopt_fp4`, `modelopt_mixed`
+
+Use `ComponentQuantizationConfig` when only one stage or component should be
+quantized. Pre-quantized ModelOpt-style checkpoints should not spill into
+vision/audio encoders that have no corresponding scale tensors.
+
+## Ownership Boundary
+
+- Upstream `vllm` owns generic quantization configs, kernels, loader semantics,
+  hardware capability rules, and generic AR quantization methods.
+- `vllm-omni` owns unified config routing, diffusion-specific wrappers,
+  component scoping, GGUF or ModelOpt checkpoint adapters, model-specific
+  prefix mapping, docs, examples, and validation.
+
+If a new method needs missing generic kernels or loader behavior, fix upstream
+`vllm` first. In `vllm-omni`, add thin integration and model wiring.
+
+## Working Loop
+
+1. Read the local method doc under `docs/user_guide/quantization/`.
+2. Inspect the active factory and config class in `vllm_omni/quantization/`.
+3. Inspect the target model pipeline and transformer for stable prefixes and
+   whether every relevant vLLM linear layer receives `quant_config`.
+4. For pre-quantized checkpoints, inspect `transformer/config.json` and the
+   checkpoint tensor names before running full generation.
+5. Validate with the same prompt, seed, size, scheduler, steps, dtype, eager or
+   non-eager mode, and parallelism as the BF16 baseline.
+6. Report quality, latency, and memory separately. Do not call a method
+   supported until quality is checked.
+
+## Common Mistakes
+
+| Symptom | Likely Cause | Fix |
+|---------|--------------|-----|
+| `--quantization` has no visible effect | Wrong scope or unsupported model path | Check component routing and method docs |
+| Some layers stay BF16 unexpectedly | `quant_config` was not threaded into all vLLM linear layers | Audit transformer constructors and prefixes |
+| Quality collapses but loading succeeds | Too many sensitive layers were quantized | Add model-specific `ignored_layers` and compare to BF16 |
+| ModelOpt checkpoint loads but output is corrupted | Prefixes, packed-module mapping, scale routing, or BF16 fallback is wrong | Use `references/modelopt-fp8.md` |
+| GGUF shape or tensor mismatch | Missing architecture-specific adapter | Add explicit adapter mapping; avoid generic fallback |
+| Online and offline MXFP4 disagree | Wrong `mxfp4` vs `mxfp4_dualscale` mode | Check `docs/user_guide/quantization/mxfp4.md` |
+| Quantized path is slower than BF16 | Kernel path, compile mode, dtype, or shape mismatch | Compare eager-to-eager and non-eager-to-non-eager |
+
+## References
+
+- General methods and config routing: [references/methods.md](references/methods.md)
+- Modality and model scope: [references/modality-compat.md](references/modality-compat.md)
+- Diffusion implementation patterns: [references/diffusion.md](references/diffusion.md)
+- Adding model support: [references/adding-models.md](references/adding-models.md)
+- ModelOpt FP8 conversion and loading: [references/modelopt-fp8.md](references/modelopt-fp8.md)

.claude/skills/quantization/references/adding-models.md

@@ -0,0 +1,70 @@
+# Adding Quantization Support to a Model
+
+Use this checklist when extending quantization to a new architecture or stage.
+
+## Classify Ownership
+
+| Case | Owner |
+|------|-------|
+| New generic AR method, kernel, or loader semantics | upstream `vllm` first |
+| Existing upstream method needs diffusion wiring | `vllm-omni` model and config integration |
+| GGUF or ModelOpt tensor mapping for a diffusion architecture | `vllm-omni` adapter |
+| NPU MXFP or msModelSlim diffusion path | `vllm-omni` platform/method wrapper and docs |
+| Multi-stage omni checkpoint quantizes only the thinker LM | `vllm-omni` component scoping |
+
+Avoid building a private quantization stack in model code.
+
+## Model Wiring Checklist
+
+1. Find the closest supported model implementation.
+2. Confirm the model uses vLLM linear layers that accept `quant_config`.
+3. Add `quant_config` to the transformer constructor.
+4. Pass stable `prefix` values through every linear layer.
+5. Handle fused modules and packed mappings explicitly.
+6. Ensure `load_weights()` and any `WeightsMapper` remap `ignored_layers` or
+   `modules_to_not_convert`.
+7. Keep tokenizer, scheduler, VAE, text encoder, audio/vision encoders, and
+   talker BF16 unless method-specific docs validate them.
+8. Add or update a method doc under `docs/user_guide/quantization/`.
+
+## Checkpoint Adapter Checklist
+
+For pre-quantized checkpoints:
+
+- Inspect `transformer/config.json` first.
+- Check `quant_method`, `method`, `producer.name`, `quant_algo`, serialized
+  flags, `ignored_layers`, and any method-specific fields.
+- Verify tensor name mapping before generation.
+- Verify packed modules such as `to_qkv`, `add_kv_proj`, and `w13`.
+- Confirm scale tensor names and shapes.
+- Allow dequantization back to BF16 for intentionally skipped target layers.
+
+## Sensitive Layer Discovery
+
+Start conservative. Quantize the backbone first and keep output-side or routing
+layers BF16 until quality is proven.
+
+Frequently sensitive areas:
+
+- attention output projections such as `to_out`
+- FFN down projections such as `w2`
+- final projections such as `proj_out`, `final_layer`, and output heads
+- modulation, timestep, position, image, or text embedders
+- refiner blocks
+- MoE routers, experts selected by routing, and layernorm-adjacent modules
+- VAE, text/vision towers, audio encoders, and codec/talker stages
+
+Do not reuse one model family's ignore list blindly.
+
+## Tests and Evidence
+
+Minimum evidence for a PR:
+
+- loader/config test or smoke command
+- fixed-seed BF16 vs quantized output comparison
+- quality metrics and saved outputs for precision-changing work
+- memory and latency comparison under identical execution mode
+- failure behavior for unsupported models or missing checkpoint fields
+
+When the PR adds a method to a support table, include the exact model,
+hardware, command, quality threshold, and artifact path used for validation.

.claude/skills/quantization/references/diffusion.md

@@ -0,0 +1,113 @@
+# Diffusion Quantization Reference
+
+Use this file for DiT quantization, config routing, checkpoint adapters, and
+quality or loader debugging.
+
+## Supported Method Families
+
+The current local diffusion quantization families are:
+
+- CUDA/GPU online or checkpoint: `fp8`, `int8`, `gguf`, ModelOpt
+- Ascend NPU: `int8`, `mxfp8`, `mxfp4`, `mxfp4_dualscale`, msModelSlim paths
+- Intel XPU: AutoRound/INC and online MXFP8 paths where documented
+
+Always cross-check the method doc before claiming model support. The CLI
+accepting a method string is not enough.
+
+## Implementation Starting Points
+
+| Area | Files |
+|------|-------|
+| Unified factory | `vllm_omni/quantization/factory.py` |
+| Per-component routing | `vllm_omni/quantization/component_config.py` |
+| GGUF diffusion method | `vllm_omni/quantization/gguf_config.py` |
+| Int8 diffusion method | `vllm_omni/quantization/int8_config.py` |
+| MXFP8 diffusion method | `vllm_omni/quantization/mxfp8_config.py` |
+| MXFP4 and dual-scale method | `vllm_omni/quantization/mxfp4_config.py` |
+| AutoRound/INC omni extension | `vllm_omni/quantization/inc_config.py` |
+| ModelOpt checkpoint adapters | `vllm_omni/diffusion/model_loader/checkpoint_adapters/modelopt.py` |
+| FP8 scaled-mm runtime patches | `vllm_omni/patch.py` |
+| Similarity tool | `vllm_omni/quantization/tools/compare_diffusion_trajectory_similarity.py` |
+
+## Threading `quant_config`
+
+For a diffusion transformer:
+
+1. Accept `quant_config: QuantizationConfig | None` in the transformer.
+2. Pass `quant_config` and stable `prefix` values to every vLLM linear layer.
+3. Preserve packed-module mappings for fused QKV, FFN, or model-specific fused
+   projections.
+4. Call `apply_vllm_mapper()` on configs that need HF-to-vLLM name remapping.
+5. For component configs, keep prefixes specific enough to avoid quantizing VAE
+   or encoders by accident.
+
+Common skipped areas are `proj_out`, output projections, modulation layers,
+embedders, refiner blocks, MoE routers, and layernorm-adjacent paths.
+
+## Online vs Offline
+
+Online quantization:
+
+- starts from BF16/FP16 checkpoint weights
+- computes or stores quantized weights at load time
+- typically uses `--quantization <method>` or `quantization="<method>"`
+
+Offline or pre-quantized checkpoints:
+
+- include quantized weights and scales on disk
+- usually carry `quantization_config` in `transformer/config.json`
+- should auto-detect when the config is complete
+- may require model-specific adapter logic for tensor names and packed modules
+
+Do not mix online flags with incompatible offline checkpoint configs. If the
+active config and disk config differ, `resolve_quant_config_from_disk()` should
+raise or rebuild rather than silently loading corrupted weights.
+
+## GGUF Rules
+
+- Require `quantization_config.gguf_model`.
+- Use the normal base model for tokenizer, scheduler, text encoder, and VAE.
+- Add explicit architecture adapters for tensor-name mapping.
+- Use `named_parameters()` and `named_buffers()`, not `state_dict()`, to
+  discover loadable names.
+- Guard fused QKV/KV rewrites so tensors are not rewritten twice.
+- Flatten N-D activations around the GGUF matmul and restore shape.
+
+Unsupported architectures should fail clearly instead of falling back to a fake
+generic mapper.
+
+## MXFP Rules
+
+- `mxfp8` is W8A8 microscaling, currently centered on NPU and documented XPU
+  behavior.
+- `mxfp4` is online single-scale W4A4 on NPU.
+- `mxfp4_dualscale` is online/offline dual-scale W4A4 with BF16 fallback; the
+  offline path is the production path when calibrated `mul_scale` exists.
+- Offline dual-scale checkpoints must inject per-transformer `ignored_layers`
+  into config because cascade transformers can have different BF16 fallback
+  sets.
+
+Do not load an offline dual-scale checkpoint with the online single-scale path.
+
+## Common Failures
+
+| Symptom | Likely Cause | Fix |
+|---------|--------------|-----|
+| `Unknown quantization method` | Factory override or upstream registration is missing | Add the right override or upstream method |
+| Missing scale tensors | Wrong scope, wrong adapter, or BF16 layer not allowed to dequantize | Check config scope and adapter scale routing |
+| Output rank collapses from 3-D to 2-D | FP8 scaled-mm kernel ignored output shape | Check the patch in `vllm_omni/patch.py` |
+| Quality degrades silently | Sensitive layers quantized | Add `ignored_layers`; compare to BF16 |
+| Only one cascade transformer works | Disk config differs per transformer but active config is reused | Rebuild from each transformer's config |
+| Quantized path slower than BF16 | Kernel, dtype, compile, or benchmark mode mismatch | Compare eager-to-eager and non-eager-to-non-eager |
+
+## Verification
+
+For a new or changed diffusion quantization path, run:
+
+1. Config and loader smoke test.
+2. One fixed-seed generation against BF16.
+3. Numeric comparison with the similarity tool or equivalent metrics.
+4. Serving test if the path is user-facing online.
+5. Memory and latency check under the same execution mode as the baseline.
+
+Keep output artifacts when precision changes are part of the PR.

.claude/skills/quantization/references/methods.md

@@ -0,0 +1,82 @@
+# General Quantization Methods
+
+Use this file for method selection and unified `vllm-omni` config syntax. For
+diffusion-specific wiring, see `diffusion.md`. For ModelOpt checkpoint export
+and loading details, see `modelopt-fp8.md`.
+
+## Unified Entry Point
+
+`vllm_omni.quantization.build_quant_config()` accepts:
+
+- `None` or `"none"` to disable quantization
+- a method string such as `"fp8"`, `"int8"`, `"mxfp8"`, `"awq"`, or `"auto-round"`
+- a flat dict with `"method"` or `"quant_method"`
+- a per-component dict such as `{"transformer": {"method": "fp8"}, "vae": None}`
+- an already-built `QuantizationConfig`
+
+Per-component dicts are routed by `ComponentQuantizationConfig` using
+longest-prefix matching. Prefix stability matters: if a model remaps HF names
+to vLLM names, apply the mapper to `ignored_layers` or component prefixes.
+
+## Local Overrides
+
+The local factory delegates most methods to upstream `vllm`, but has vLLM-Omni
+overrides for methods that need diffusion, NPU, XPU, or component behavior:
+
+| Method | Local Config | Main Use |
+|--------|--------------|----------|
+| `gguf` | `DiffusionGGUFConfig` | Pre-quantized diffusion transformer weights |
+| `int8` | `DiffusionInt8Config` | Online or serialized W8A8 diffusion transformers |
+| `mxfp8` | `DiffusionMXFP8Config` | W8A8 MXFP8 diffusion on NPU and online XPU path |
+| `mxfp4` | `DiffusionMXFP4Config` | Online W4A4 MXFP4 diffusion on NPU |
+| `mxfp4_dualscale` | `DiffusionMXFP4DualScaleMixedConfig` | Online/offline W4A4 dual-scale plus BF16 fallback on NPU |
+| `inc`, `auto-round`, `auto_round` | `OmniINCConfig` | AutoRound/INC checkpoint loading, including MXFP8 checkpoint metadata |
+| ModelOpt config dicts | upstream ModelOpt configs through auto-detect | FP8, NVFP4, or mixed ModelOpt checkpoints |
+
+Check `SUPPORTED_QUANTIZATION_METHODS` in `vllm_omni/quantization/factory.py`
+before adding a new method name.
+
+## AR and Generic Quantization
+
+For AR-backed models, upstream `vllm` methods usually apply:
+
+- `awq` and `gptq`: 4-bit or 8-bit weight quantization with calibration,
+  commonly used for text and AR stages.
+- `fp8`: FP8 weight path or KV-cache FP8 depending on flags and hardware.
+- `inc` or `auto-round`: pre-quantized checkpoints with Intel Neural
+  Compressor / AutoRound metadata.
+
+Use these for the AR language model stage only when the checkpoint and model
+guide support it. Omni and TTS models often keep audio, vision, talker, and
+codec stages in BF16.
+
+## Diffusion and Offline Methods
+
+For diffusion transformers, prefer the method-specific docs:
+
+- `docs/user_guide/quantization/fp8.md`
+- `docs/user_guide/quantization/int8.md`
+- `docs/user_guide/quantization/gguf.md`
+- `docs/user_guide/quantization/modelopt.md`
+- `docs/user_guide/quantization/mxfp8.md`
+- `docs/user_guide/quantization/mxfp4.md`
+- `docs/user_guide/quantization/autoround.md`
+- `docs/user_guide/quantization/msmodelslim.md`
+
+Pre-quantized checkpoints should normally auto-detect from config. Runtime
+flags such as `--force-cutlass-fp8`, `--linear-backend cutlass`, and
+`--moe-backend cutlass` select kernels; they do not quantize a BF16 checkpoint.
+
+## Validation
+
+Use fixed-seed A/B comparisons and record:
+
+- model path and quantized checkpoint path
+- method config and `ignored_layers`
+- prompt, size, frame count, steps, scheduler, dtype, and seed
+- eager/non-eager mode, compile flags, and parallelism
+- output metrics such as PSNR, MAE, cosine similarity, SSIM, LPIPS, plus visual artifacts
+- latency and peak memory, reported separately from quality
+
+The helper `vllm_omni.quantization.tools.compare_diffusion_trajectory_similarity`
+is the default lightweight image/video comparison tool for diffusion paths.

.claude/skills/quantization/references/modality-compat.md

@@ -0,0 +1,47 @@
+# Quantization Compatibility by Modality
+
+Use this as a routing guide. The source of truth is the local method docs under
+`docs/user_guide/quantization/` and the current model implementation.
+
+## Summary
+
+- AR and omni language-model quantization mostly follows upstream `vllm`, with
+  vLLM-Omni scoping for multi-stage models.
+- Diffusion quantization is method-specific and model-specific. It usually
+  targets the diffusion transformer only.
+- Pre-quantized checkpoints should quantize only the component that has
+  quantized weights and scale tensors.
+
+## Model Type Matrix
+
+| Model Type | Typical Quantization Scope | Notes |
+|------------|----------------------------|-------|
+| AR text models | Language model | Upstream `vllm` methods usually apply |
+| Omni/TTS models | Thinker or AR language-model stage | Audio, vision, talker, and codec stages usually stay BF16 |
+| DiT image models | Diffusion transformer | Check FP8, Int8, GGUF, ModelOpt, AutoRound docs |
+| DiT video models | Diffusion transformer | HunyuanVideo FP8 and Wan2.2 MXFP paths are documented; validate other video paths |
+| Multi-stage diffusion | Selected stage only | Use per-component routing and per-stage validation |
+| Audio diffusion | Usually unsupported | Prefer dtype, offload, cache, or parallelism unless docs say otherwise |
+
+## Method Hints
+
+| Goal | Candidate |
+|------|-----------|
+| AR memory reduction | AWQ, GPTQ, AutoRound, ModelOpt where checkpoint supports it |
+| CUDA diffusion W8A8 | FP8 or Int8 |
+| Pre-quantized CUDA diffusion checkpoint | ModelOpt or GGUF |
+| Ascend NPU diffusion | Int8, MXFP8, MXFP4, MXFP4 DualScale, msModelSlim |
+| Intel XPU checkpoint | AutoRound/INC paths |
+| Conservative image quality | selective FP8/Int8 with `ignored_layers` |
+| Unsupported diffusion model | cache acceleration, CPU/layerwise offload, dtype, TP/SP/CFG parallelism |
+
+## Validation Rule
+
+Compatibility means more than successful loading. A method is compatible only
+after it has:
+
+- correct component scope
+- complete checkpoint/config metadata if offline
+- fixed-seed quality comparison against BF16
+- acceptable latency and memory behavior
+- docs or examples that state model, method, hardware, and caveats

.claude/skills/quantization/references/modelopt-fp8.md

@@ -0,0 +1,122 @@
+# ModelOpt FP8 Conversion and Loading
+
+Use this when exporting or loading ModelOpt FP8 checkpoints for diffusion
+transformers. This reference captures the practical conversion lessons from the
+2026-04-21 ModelOpt FP8 paste and aligns them with the current
+`vllm-omni` ModelOpt adapter.
+
+## Core Rule
+
+Treat ModelOpt FP8 as selective checkpoint quantization, not one-click
+whole-model FP8. Quantize the main transformer or DiT backbone, keep sensitive
+or peripheral modules BF16, serialize a complete ModelOpt checkpoint, and let
+vLLM-Omni auto-detect it from config.
+
+Do not pass `--quantization fp8` to force a ModelOpt checkpoint path. The
+checkpoint's `quantization_config` should select `modelopt`.
+
+## Conversion Strategy
+
+Prefer:
+
+- backbone attention Q/K/V and FFN expansion projections
+- conservative per-model ignore lists
+- MHA quantizers disabled until basic linear quantization is stable
+- full module-name prefixes in ignore/routing logic
+- BF16 fallback for output-side and routing-sensitive layers
+
+Avoid quantizing by default:
+
+- VAE
+- text, vision, or audio encoders
+- embedders and modulation layers
+- final heads and output projections
+- refiner modules
+- MoE routers and layernorm-sensitive blocks
+
+## Model-Specific Starting Points
+
+| Model Family | Quantize First | Keep BF16 First |
+|--------------|----------------|-----------------|
+| Z-Image | attention q/k/v, FFN w1/w3 | `to_out`, `w2`, `noise_refiner`, `context_refiner`, embedders, modulation, final/output |
+| Qwen-Image | transformer backbone linear layers | `proj_out`, `img_in`, `txt_in`, norms, modulation, position/time/embed layers |
+| FLUX2-dev / FLUX2-klein | backbone linear layers | `proj_out`, context/x embedders, `norm_out`, stream modulation, time embeddings |
+| HunyuanImage-3.0 | `model.model` backbone | VAE, vision tower, token/time/image embedders, final image head, LM head, routers, layernorm-sensitive blocks |
+
+These are starting points, not support claims. Validate with BF16 outputs.
+
+## Required Checkpoint Metadata
+
+`transformer/config.json` must include a complete `quantization_config`.
+At minimum it should identify ModelOpt and the algorithm:
+
+```json
+{
+  "quantization_config": {
+    "producer": {"name": "modelopt"},
+    "quant_method": "modelopt",
+    "quant_algo": "FP8"
+  }
+}
+```
+
+Current vLLM-Omni auto-detects:
+
+- `FP8` and `FP8_PER_CHANNEL_PER_TOKEN` -> `modelopt`
+- `NVFP4` -> `modelopt_fp4`
+- `MIXED_PRECISION` -> `modelopt_mixed`
+
+Packed modules and scales must map correctly:
+
+- `to_qkv` from `to_q`, `to_k`, `to_v`
+- `add_kv_proj` from `add_q_proj`, `add_k_proj`, `add_v_proj`
+- `w13` from `w1`, `w3`
+- `.input_scale`, `.weight_scale`, `.weight_scale_2`, `.weight_scale_inv`
+
+For selective FP8, target layers that remain BF16 must be allowed to
+dequantize from FP8 weights using the matching `weight_scale`, or the adapter
+must skip the quantized scale tensors for that BF16 target.
+
+## Runtime Signals
+
+Healthy loading should show ModelOpt auto-detection and a ModelOpt linear
+kernel path. Useful signals include:
+
+- auto-detected quantization from config
+- selected ModelOpt FP8 linear method
+- CUTLASS or FlashInfer FP8 scaled-mm backend when enabled and supported
+- non-eager path compiling the transformer when expected
+
+If online serving fails while offline inference works, inspect warmup and
+serving initialization before changing checkpoint conversion.
+
+## Validation Order
+
+1. Check config completeness.
+2. Run an export checker such as `examples/quantization/check_modelopt_fp8_export.py`.
+3. Confirm actual FP8 tensors and expected disk-size reduction.
+4. Run one offline generation with fixed prompt, seed, size, and steps.
+5. Run serving with `vllm serve --omni` and verify `/v1/images/generations`.
+6. Compare against BF16 using visual review plus metrics such as LPIPS, PSNR,
+   MAE, cosine similarity, and SSIM.
+7. Compare performance only under matched modes:
+   - BF16 eager vs FP8 eager
+   - BF16 non-eager vs FP8 non-eager
+
+FP8 is not automatically faster. Check whether execution reaches FP8 kernels,
+whether compile mode is comparable, and whether the workload shape fits the
+kernel path.
+
+## Triage
+
+| Symptom | First Check |
+|---------|-------------|
+| Startup failure | `config.json`, `model_index.json`, `quantization_config`, missing scales |
+| Loads but output is corrupted | Over-quantized sensitive layers or wrong full-name prefixes |
+| Offline works but online fails | Dummy warmup, serving init, compile path |
+| FP8 is not faster | Kernel selection, dtype, compile mode, workload shape |
+| BF16 fallback fails | Missing `weight_scale` for dequantization or wrong packed-module mapping |
+
+The success criterion is stable, correctly scoped checkpoint loading with
+quality preserved enough for the target use case, not maximizing the number of
+FP8 layers.

.claude/skills/readme.md

@@ -21,11 +21,18 @@ include:
 - `diffusion-perf-opt`: guides diffusion model performance optimization,
   including profiling traces, parallel strategies, stage timing analysis, and
   benchmark-driven tuning
+- `quantization`: guides quantization method selection, model integration,
+  checkpoint loading, and quality/performance validation for vLLM-Omni
 - `add-omni-model`: covers addition of new omni-modality model support
 - `add-tts-model`: covers integration of new TTS models and related serving
   workflows
 - `generate-release-note`: helps prepare release notes for repository changes
+- `precheck-pr`: self-check a branch before creating a PR — validates PR title
+  format, catches dead code, verifies accuracy/perf claims, and confirms merge
+  readiness
 - `review-pr`: provides a structured workflow for reviewing pull requests
+- `vllm-omni-npu-model-runner-upgrade`: upgrades NPU model runners to align with the
+  latest vllm-ascend NPUModelRunner
 
 ## Maintenance Guidelines
 

.github/PULL_REQUEST_TEMPLATE.md

@@ -1,21 +1,15 @@
 <!-- markdownlint-disable -->
-PLEASE FILL IN THE PR DESCRIPTION HERE ENSURING ALL CHECKLIST ITEMS (AT THE BOTTOM) HAVE BEEN CONSIDERED.
+PLEASE FILL IN THE PR DESCRIPTION HERE.
 
 ## Purpose
 
 ## Test Plan
 
+**vLLM Version:**
+
+**vLLM-Omni Commit:**
+
 ## Test Result
 
----
-<details>
-<summary> Essential Elements of an Effective PR Description Checklist </summary>
-
-- [ ] The purpose of the PR, such as "Fix some issue (link existing issues this PR will resolve)".
-- [ ] The test plan. Please provide the test scripts & test commands. Please state the reasons if your codes don't require additional test scripts. For test file guidelines, please check the [test style doc](https://docs.vllm.ai/projects/vllm-omni/en/latest/contributing/ci/tests_style/)
-- [ ] The test results. Please paste the results comparison before and after, or the e2e results.
-- [ ] (Optional) The necessary documentation update, such as updating `supported_models.md` and `examples` for a new model. **Please run `mkdocs serve` to sync the documentation editions to `./docs`.**
-- [ ] (Optional) Release notes update. If your change is user-facing, please update the release notes draft.
-</details>
-
-**BEFORE SUBMITTING, PLEASE READ <https://github.com/vllm-project/vllm-omni/blob/main/CONTRIBUTING.md>** (anything written below this line will be removed by GitHub Actions)
+**BEFORE SUBMITTING:** read [CONTRIBUTING.md](https://github.com/vllm-project/vllm-omni/blob/main/CONTRIBUTING.md) and run the [precheck-pr skill](https://github.com/vllm-project/vllm-omni/blob/main/.claude/skills/precheck-pr/SKILL.md) with the code agent for a self-check against project conventions.
+(anything written below this line will be removed by GitHub Actions)

CONTRIBUTING.md

@@ -1,3 +1,5 @@
 # Contributing to vLLM-Omni
 
 You may find information about contributing to vLLM-Omni on [Contributing](https://vllm-omni.readthedocs.io/en/latest/contributing/)
+
+Before submitting a PR, run the [precheck-pr skill](https://github.com/vllm-project/vllm-omni/blob/main/.claude/skills/precheck-pr/SKILL.md) with the code agent for a self-check against project conventions.

README.md

@@ -16,6 +16,7 @@ Easy, fast, and cheap omni-modality model serving for everyone
 ---
 
 *Latest News* 🔥
+- [2026/06] We released [0.22.0](https://github.com/vllm-project/vllm-omni/releases/tag/v0.22.0) - an **omnimodal world-model** release aligned with vLLM 0.22, featuring [Nvidia Cosmos3](recipes/cosmos3/Cosmos3-Nano.md)/DreamZero world model support, expanded quantization coverage across Blackwell/NPU/XPU, TTS production improvements, new models including MiniCPM-o 4.5, MOSS-TTS, and Lance, plus RL integration with [VeRL-Omni](https://github.com/verl-project/verl-omni).
 - [2026/05] We released [0.20.0](https://github.com/vllm-project/vllm-omni/releases/tag/v0.20.0) - refreshes the serving/runtime stack for large-scale omni workloads, and improves diffusion model performance, quantization, and hardware readiness across CUDA, ROCm, MUSA, NPU, and XPU backends.
 - [2026/03] We released [0.18.0](https://github.com/vllm-project/vllm-omni/releases/tag/v0.18.0) - strengthens the core runtime through a large entrypoint refactor and scheduler/runtime cleanups, expands unified quantization and diffusion execution, broadens multimodal model coverage, and improves production readiness across audio, omni, image, video, RL, and multi-platform deployments.
 - [2026/03] Check out our first public [project deepdive](https://youtu.be/sgwNfsNnR9I) at the vLLM Hong Kong Meetup!
@@ -56,8 +57,9 @@ vLLM-Omni is flexible and easy to use with:
 
 vLLM-Omni seamlessly supports most popular open-source models on HuggingFace, including:
 
-- Omni-modality models (e.g. Qwen-Omni)
-- Multi-modality generation models (e.g. Qwen-Image)
+- **Omni-modality models** (e.g. Qwen3-Omni, Cosmos, HunyuanImage, BAGEL)
+- **TTS models** (e.g. Qwen3-TTS, VoxCPM2, Ming-Omni-TTS, CosyVoice3)
+- **Diffusion models** — image, video, and audio generation (e.g. Qwen-Image, Wan2.2, FLUX)
 
 ## Getting Started
 

benchmarks/tts/model_configs.yaml

@@ -33,6 +33,15 @@ models:
       voice_clone:
         task_type: Base
 
+  bosonai/higgs-audio-v3-tts-4b:
+    supported_tasks: [default_voice, voice_clone]
+    backend: openai-audio-speech
+    endpoint: /v1/audio/speech
+    trust_remote_code: true
+    task_extra_body:
+      default_voice: {}
+      voice_clone: {}
+
   openbmb/VoxCPM2:
     supported_tasks: [voice_clone, default_voice]
     trust_remote_code: true

docker/Dockerfile.ci

@@ -2,7 +2,7 @@
 # The pinned VLLM_PRECOMPILED_WHEEL_COMMIT and VLLM_BASE_TAG below must
 # be updated on every rebase to match the new upstream vllm commit/tag.
 ARG VLLM_BASE_IMAGE=vllm/vllm-openai
-ARG VLLM_BASE_TAG=v0.22.0
+ARG VLLM_BASE_TAG=v0.23.0
 FROM ${VLLM_BASE_IMAGE}:${VLLM_BASE_TAG}
 ARG APP_DIR=/workspace/vllm-omni
 WORKDIR ${APP_DIR}
@@ -16,6 +16,7 @@ RUN apt-get update && \
 
 RUN uv pip install --system ".[dev]"
 
+
 RUN ln -sf /usr/bin/python3 /usr/bin/python
 
 ENTRYPOINT []

docker/Dockerfile.cuda

@@ -1,4 +1,4 @@
-ARG BASE_IMAGE=vllm/vllm-openai:v0.22.0
+ARG BASE_IMAGE=vllm/vllm-openai:v0.23.0
 FROM ${BASE_IMAGE}
 
 ARG COMMON_WORKDIR=/app

docker/Dockerfile.rocm

@@ -1,4 +1,4 @@
-ARG BASE_IMAGE=vllm/vllm-openai-rocm:v0.22.0
+ARG BASE_IMAGE=vllm/vllm-openai-rocm:v0.23.0
 FROM ${BASE_IMAGE} AS base
 
 # Declare a variable to know if we want to use the nightly build or the stable build.

docker/Dockerfile.xpu

@@ -88,7 +88,7 @@ ENV UV_EXTRA_INDEX_URL=${PIP_EXTRA_INDEX_URL}
 ENV UV_INDEX_STRATEGY="unsafe-best-match"
 ENV UV_LINK_MODE="copy"
 
-ARG VLLM_VERSION=v0.22.0
+ARG VLLM_VERSION=v0.23.0
 RUN git clone -b ${VLLM_VERSION} https://github.com/vllm-project/vllm
 WORKDIR /workspace/vllm
 
@@ -142,15 +142,12 @@ WORKDIR /workspace/vllm-omni
 COPY . .
 
 ENV VLLM_OMNI_TARGET_DEVICE=xpu
-RUN uv pip install --no-cache-dir ".[dev]" --no-build-isolation
-
-# FIX triton
-RUN --mount=type=cache,target=/root/.cache/uv \
-    uv pip uninstall triton triton-xpu && \
-    uv pip install triton-xpu==3.7.0
-
-# remove torch bundled oneccl to avoid conflicts
 RUN --mount=type=cache,target=/root/.cache/uv \
+    echo "triton @ https://example.com/blocked" > /tmp/vllm_omni_constraints.txt && \
+    PIP_CONSTRAINT=/tmp/vllm_omni_constraints.txt \
+    uv pip install --no-build-isolation \
+    --extra-index-url https://mirrors.ustc.edu.cn/pypi/simple/ \
+    ".[dev]" && \
     uv pip uninstall oneccl oneccl-devel
 
 FROM vllm-omni AS vllm-omni-openai

docs/.nav.yml

@@ -87,6 +87,8 @@ nav:
       - AutoRound: user_guide/quantization/autoround.md
       - msModelSlim: user_guide/quantization/msmodelslim.md
     - ComfyUI: features/comfyui.md
+  - Fault Tolerance:
+    - Failure Modes: user_guide/fault_injection_reliability_matrix.md
 - Developer Guide:
   - General:
     - contributing/README.md

docs/api/README.md

@@ -92,7 +92,6 @@ Core scheduling and caching components.
 Configuration classes.
 
 - [vllm_omni.config.model.OmniModelConfig][]
-- [vllm_omni.config.stage_config.ModelPipeline][]
 - [vllm_omni.config.stage_config.StageConfig][]
 - [vllm_omni.config.stage_config.StageConfigFactory][]
 - [vllm_omni.config.stage_config.StageType][]

docs/contributing/README.md

@@ -116,6 +116,16 @@ Only specific types of PRs will be reviewed. The PR title is prefixed appropriat
 !!! note
     If the PR spans more than one category, please include all relevant prefixes.
 
+### Pre-Check Before Submitting
+
+Before submitting a PR, run the [precheck-pr skill](https://github.com/vllm-project/vllm-omni/blob/main/.claude/skills/precheck-pr/SKILL.md) with the code agent for a self-review against project conventions:
+
+The skill offers two modes:
+- **Quick (~3 min):** catches showstoppers — PR title format, missing benchmark claims, rebase status
+- **Full (~10 min):** thorough maintainer-grade review — dead code scan, copy-paste detection, import hygiene
+
+The precheck covers five PR types: Bug Fix, Performance, New Model, Diffusion Model, and General. Each type has a tailored checklist that validates evidence quality (repro steps, A/B benchmarks, registry entries, etc.). See the [precheck-pr skill](https://github.com/vllm-project/vllm-omni/blob/main/.claude/skills/precheck-pr/SKILL.md) for the full checklist.
+
 ### Local Test
 Please run the L1 and L2 test cases locally first and attach the results before contacting us to add the "ready" label. Please refer to the [test instructions](./ci/test_guide.md) for running the test cases.
 

docs/contributing/ci/CI_5levels.md

@@ -268,6 +268,120 @@ Before entering specific testing levels, the project establishes two common spec
 1.  ***PR Checklist ([Tests Style](../ci/tests_style.md))***: This template defines the self-check items that must be completed before submitting a code review (Pull Request). It ensures that each code change meets basic requirements such as code style, dependency updates, and documentation synchronization before entering the automated testing pipeline, serving as the first manual line of defense for quality assurance.
 2.  ***CI Failure Explanation ([CI Failures](../ci/failures.md))***: This document archives and explains common failure patterns in the Continuous Integration (CI) pipeline, error log interpretation, and preliminary troubleshooting steps. It helps developers and testers quickly diagnose the causes of automated test failures, improving problem-solving efficiency.
 
+## Notes
+
+### Diff-aware Buildkite uploads (`source_file_dependencies`)
+
+L2 (`.buildkite/test-ready.yml`) and L3 (`.buildkite/test-merge.yml`) pipelines can **skip unrelated GPU jobs at upload time** based on the PR diff. This is implemented by `.buildkite/scripts/upload_pipeline.py`, which filters steps before calling `buildkite-agent pipeline upload`.
+
+#### What `source_file_dependencies` is
+
+- A **uploader-only** YAML key on a Buildkite step or group. **Buildkite does not understand it**; `upload_pipeline.py` always **removes** it from the YAML that is uploaded.
+- A list of path **prefixes** (directories or individual files). If **any** changed file in the diff equals a prefix or starts with `prefix/`, the step is kept; otherwise the step (or entire group) is **omitted** from the uploaded pipeline.
+
+#### When filtering runs
+
+| Build context | Changed files used for matching |
+| --- | --- |
+| Pull request | `git diff --name-only origin/<base>...<BUILDKITE_COMMIT>` |
+| `main` branch push | `git diff --name-only <commit>^..<commit>` |
+| Other (e.g. local dry-run, non-PR branch) | Diff cannot be resolved → **no filtering**; all steps are uploaded and `source_file_dependencies` is still stripped |
+
+Docs-only PRs are handled earlier in bootstrap (`pipeline.yml`) via skip-ci logic; `source_file_dependencies` applies only to the **uploaded** L2/L3 test pipelines.
+
+#### Where it is configured
+
+| Level | Pipeline file | Upload entry |
+| --- | --- | --- |
+| L2 | `.buildkite/test-ready.yml` | `upload_pipeline.py --upload .buildkite/test-ready.yml` (from `pipeline.yml`) |
+| L3 | `.buildkite/test-merge.yml` | `upload_pipeline.py --upload .buildkite/test-merge.yml` (from `pipeline.yml`) |
+
+Steps **without** `source_file_dependencies` are always uploaded (subject to the usual label conditions: `ready` for L2, `merge-test` for L3).
+
+#### Current skip policy (L2 / L3)
+
+To balance CI cost and coverage:
+
+- **Always run** (no `source_file_dependencies`): baseline groups outside E2E Test—e.g. Simple Test, Diffusion unit tests, Engine/Model Executor, Distributed, Custom Pipeline, Entrypoints (L2), LoRA / Entrypoints (L3).
+- **Diff-gated** (`source_file_dependencies` set): **every leaf job under the E2E Test group** in `test-ready.yml` and `test-merge.yml`, regardless of queue (`mithril-h100-pool`, `gpu_1_queue`, or `gpu_4_queue`). Each step lists the smallest set of prefixes that should trigger it—typically:
+  - pytest file(s) exercised by the job (online and/or offline);
+  - model code under `vllm_omni/model_executor/models/` or `vllm_omni/diffusion/models/`;
+  - related `vllm_omni/model_executor/stage_input_processors/` and `vllm_omni/deploy/*.yaml` when applicable.
+
+Adding a new E2E step: add `source_file_dependencies` on the leaf job with those prefixes. Prefer **per-step** deps rather than a broad group-level list unless every child shares the same paths.
+
+#### YAML examples
+
+H100 E2E (kubernetes / `mithril-h100-pool`):
+
+```yaml
+      - label: "Diffusion · Qwen Image Test"
+        source_file_dependencies:
+          - tests/e2e/online_serving/test_qwen_image.py
+          - vllm_omni/diffusion/models/qwen_image/
+        commands:
+          - pytest -s -v tests/e2e/online_serving/test_qwen_image.py -m 'core_model' ...
+        agents:
+          queue: "mithril-h100-pool"
+```
+
+Docker E2E (`gpu_1_queue` / `gpu_4_queue`)—same key, same upload-time filtering:
+
+```yaml
+      - label: "TTS · Qwen3-TTS CustomVoice Test"
+        source_file_dependencies:
+          - tests/e2e/online_serving/test_qwen3_tts_customvoice.py
+          - vllm_omni/model_executor/models/qwen3_tts/
+          - vllm_omni/model_executor/stage_input_processors/qwen3_tts.py
+          - vllm_omni/deploy/qwen3_tts.yaml
+        commands:
+          - pytest -s -v tests/e2e/online_serving/test_qwen3_tts_customvoice.py ...
+        agents:
+          queue: "gpu_4_queue"
+```
+
+A **group** may also define `source_file_dependencies`; nested steps inherit filtering as a unit—the whole group is dropped if no prefix matches.
+
+#### Local dry-run
+
+```bash
+# Render filtered YAML to stdout (no upload)
+python3 .buildkite/scripts/upload_pipeline.py .buildkite/test-ready.yml
+
+# Confirm uploader-only keys are stripped
+python3 .buildkite/scripts/upload_pipeline.py .buildkite/test-merge.yml | grep source_file_dependencies
+# (no output expected)
+```
+
+On a PR build, Buildkite logs from `upload_pipeline.py` include lines such as `skip '…' (no changes under …)` for omitted steps.
+
+#### Related
+
+- Implementation: `.buildkite/scripts/upload_pipeline.py`
+- L2/L3 diff skip does **not** replace label-based triggers (`ready`, `merge-test`); it only reduces which steps appear **after** the pipeline is already scheduled.
+
+### Test helper environment variables
+
+Some shared helpers under `tests/helpers/` honor optional environment variables for local debugging. These are **not** set in CI by default.
+
+| Variable | Accepted values | Description |
+| -------- | --------------- | ----------- |
+| `VLLM_OMNI_KEEP_REQUEST_MEDIA` | `1`, `true`, `yes` (case-insensitive) | When enabled, temporary WAV files created by `tests.helpers.media.convert_audio_bytes_to_text` are **not** deleted when the pytest process exits. By default, each call writes a unique file under the system temp directory via `tempfile.mkstemp` and registers `atexit` cleanup. Use this when debugging audio output validation (Whisper transcription, keyword checks, text–audio similarity). The saved path is logged as `audio data is saved: <path>`. |
+
+Example (Linux / macOS):
+
+```bash
+export VLLM_OMNI_KEEP_REQUEST_MEDIA=1
+pytest -s -v tests/e2e/online_serving/test_qwen3_omni.py -k test_mix_to_text_audio
+```
+
+Example (Windows PowerShell):
+
+```powershell
+$env:VLLM_OMNI_KEEP_REQUEST_MEDIA = "1"
+pytest -s -v tests/e2e/online_serving/test_qwen3_omni.py -k test_mix_to_text_audio
+```
+
 ## Chapter 1: L1 & L2 Level Testing - Unit Testing and Basic End-to-End Verification
 
 ### 1.1 Testing Purpose
@@ -302,6 +416,7 @@ A clear directory structure is key to managing test cases efficiently.
 ### 1.4 Execution Method and Example
 
 -   ***Trigger Timing***: **`PR with ready label`**. That is, when a developer adds a "ready for review" or similar label to a PR on platforms like GitHub, L1 and L2 tests are automatically triggered.
+-   ***Diff-aware step skipping***: On L2, **E2E Test** jobs may be omitted at pipeline upload when the PR diff does not touch their [`source_file_dependencies`](#diff-aware-buildkite-uploads-source_file_dependencies) prefixes (H100 and docker queues alike); non-E2E groups still always upload. See [Diff-aware Buildkite uploads](#diff-aware-buildkite-uploads-source_file_dependencies).
 -   ***Execution Environment***: L1 uses ***CPU*** environment; L2 requires ***GPU*** environment.
 -   ***Script Example***:
 
@@ -404,7 +519,8 @@ L3 level testing executes after code is merged into the main branch. Its core pu
 
 ### 2.4 Execution Method and Example
 
--   ***Trigger Timing***: **`PR Merged`**. Automatically triggered after code review is approved and merged into the main branch.
+-   ***Trigger Timing***: **`PR Merged`**. Automatically triggered after code review is approved and merged into the main branch (typically via `merge-test` label on the PR before merge).
+-   ***Diff-aware step skipping***: Same [`source_file_dependencies`](#diff-aware-buildkite-uploads-source_file_dependencies) mechanism as L2—**all E2E Test** leaf steps are diff-gated; other L3 groups always upload. See [Diff-aware Buildkite uploads](#diff-aware-buildkite-uploads-source_file_dependencies).
 -   ***Execution Environment***: ***GPU*** servers.
 -   ***Script Example***:
 
@@ -520,6 +636,8 @@ L3 level testing executes after code is merged into the main branch. Its core pu
 
     **Implicit Validation**: The `send_omni_request` and `send_diffusion_request` methods internally includes validation logic dynamically selected based on the --run-level parameter: core_model performs basic validation, while advanced_model and full_model perform deep validation.
 
+    **Audio output debugging**: Deep validation may transcribe returned audio via `convert_audio_bytes_to_text` (Whisper). If an audio keyword or text–audio similarity assertion fails, set `VLLM_OMNI_KEEP_REQUEST_MEDIA=1` before running pytest to keep the intermediate WAV files for inspection (see [Test helper environment variables](#test-helper-environment-variables)).
+
 -   ***Run Command (L3 merge)***: `pytest -s -v /tests/e2e/online_serving/test_{model_name}.py -m advanced_model --run-level=advanced_model`
 
 -   ***Run Command (L4 nightly expansion)***: `pytest -s -v /tests/e2e/online_serving/test_{model_name}_expansion.py -m full_model --run-level=full_model`

docs/contributing/ci/failures.md

@@ -69,6 +69,8 @@ If the failure is already tracked, leave a comment on your PR noting that the fa
 
 If the failure appears to be new, investigate whether your changes caused it.
 
+For resilience behavior under injected process signals/OOM pressure (validation scenarios, not root-cause equivalence), see [Failure Mode Injection Scenarios and Expected Behavior Matrix](../../user_guide/fault_injection_reliability_matrix.md).
+
 ### Reading Buildkite Logs
 
 1. Click **Details** next to the Buildkite check on your PR.

docs/design/feature/disaggregated_inference.md

@@ -8,6 +8,7 @@ Backend-specific setup lives in separate docs:
 - [SharedMemoryConnector](omni_connectors/shared_memory_connector.md)
 - [MooncakeStoreConnector](omni_connectors/mooncake_store_connector.md)
 - [MooncakeTransferEngineConnector](omni_connectors/mooncake_transfer_engine_connector.md)
+- [MoriTransferEngineConnector](omni_connectors/mori_transfer_engine_connector.md)
 - [YuanrongConnector](omni_connectors/yuanrong_connector.md)
 - [YuanrongTransferEngineConnector](omni_connectors/yuanrong_transfer_engine_connector.md)
 
@@ -23,6 +24,7 @@ Current connectors operate in D2H2D (device to host to device) mode.
 | Single node | SharedMemoryConnector | Auto-configured if no connector is specified. |
 | Multi node (Mooncake Store) | MooncakeStoreConnector | TCP-based, requires Mooncake Master + metadata server. |
 | Multi node (Mooncake RDMA) | MooncakeTransferEngineConnector | RDMA/TCP direct transfer with managed memory pool. Fastest. |
+| Multi node (Mori RDMA) | MoriTransferEngineConnector | RDMA direct transfer via Mori IOEngine. |
 | Multi node (Yuanrong) | YuanrongConnector | Requires Yuanrong Datasystem + etcd. |
 | Ascend NPU P2P (Yuanrong TE) | YuanrongTransferEngineConnector | Uses Yuanrong TransferEngine directly. Configure NPU device IPv4 and `memory_pool_device: "npu"`. |
 

docs/design/feature/omni_connectors/mori_transfer_engine_connector.md

@@ -0,0 +1,76 @@
+# MoriTransferEngineConnector
+
+## When to Use
+
+Currently supports intra-node deployment with Mori.
+
+As noted in #1742, inter-node support will be added back in a future
+refactor.
+
+## Mechanism
+
+Uses Mori's `IOEngine` / `MemoryDesc` API for zero-copy RDMA transfers.
+
+- Data Plane: RDMA (InfiniBand/RoCE) with managed memory pool.
+- Control Plane: ZMQ for pull-request handshake and async completion.
+
+## Installation
+
+See the [Mori repository](https://github.com/ROCm/mori) for installation instructions.
+
+## Configuration
+
+Mori is configured through the new deploy-config schema (see
+[`docs/configuration/stage_configs.md`](../../../configuration/stage_configs.md)).
+Define the connector at the top level of the deploy YAML and reference it
+by name from each stage's `input_connectors` / `output_connectors`:
+
+```yaml
+connectors:
+  mori_connector:
+    name: MoriTransferEngineConnector
+    extra:
+      host: "auto"
+      zmq_port: 50051
+      device_name: ""
+      memory_pool_size: 536870912
+      memory_pool_device: "cuda"
+
+stages:
+  - stage_id: 0
+    output_connectors:
+      to_stage_1: mori_connector
+
+  - stage_id: 1
+    input_connectors:
+      from_stage_0: mori_connector
+```
+
+A ready-to-run intra-node example for Qwen3-Omni-MoE on AMD MI300X lives
+at `vllm_omni/deploy/qwen3_omni_moe_mori_intranode.yaml` and can be loaded
+with:
+
+```bash
+vllm-omni serve Qwen/Qwen3-Omni-30B-A3B-Instruct --omni --log-stats \
+    --deploy-config vllm_omni/deploy/qwen3_omni_moe_mori_intranode.yaml
+```
+
+The yaml wires `MoriTransferEngineConnector` (with `backend_type: xgmi`) to
+the chunk_transfer_adapter path (`async_chunk: true`) so stage-to-stage
+hidden-state and codec-frame streams ship GPU-to-GPU over AMD Infinity
+Fabric instead of SHM.  Qwen2.5-Omni + Mori is not yet functional on
+the chunk path: it needs `thinker2talker_async_chunk` /
+`talker2code2wav_async_chunk` input processors that do not exist yet
+(the orchestrator-level path the upstream PR originally targeted was
+lost during the entrypoints → engine refactor and removed in #1742).
+
+Parameters:
+
+- host: local RDMA IP (`"auto"` for auto-detect).
+- zmq_port: ZMQ base port for control-plane communication.
+- device_name: RDMA device (e.g., `"mlx5_0"`), empty for auto-detect.
+- memory_pool_size: RDMA memory pool size in bytes.
+- memory_pool_device: `"cpu"` (pinned) or `"cuda"` (GPUDirect / XGMI RDMA).
+
+For more details, refer to the
+[Mori repository](https://github.com/ROCm/mori).

docs/design/module/dit_module.md

@@ -305,8 +305,8 @@ class GPUWorker:
         )
 
         # Load model
-        model_loader = DiffusersPipelineLoader(load_config)
-        self.pipeline = model_loader.load_model(od_config, load_device=f"cuda:{rank}")
+        model_loader = DiffusersPipelineLoader(load_config, self.od_config)
+        self.pipeline = model_loader.load_model(load_device=f"cuda:{rank}")
 
         # Setup cache backend
         from vllm_omni.diffusion.cache.selector import get_cache_backend

docs/getting_started/installation/gpu.md

@@ -34,7 +34,7 @@ vLLM-Omni is a Python library that supports the following GPU variants. The libr
 
 ### Pre-built wheels
 
-Note: Pre-built wheels are currently available for vLLM-Omni 0.11.0rc1, 0.12.0rc1, 0.14.0rc1, 0.14.0, 0.16.0, 0.18.0, 0.20.0, 0.21.0, and 0.22.0. If you need a newer unreleased revision, please [build from source](https://docs.vllm.ai/projects/vllm-omni/en/latest/getting_started/installation/gpu/#build-wheel-from-source).
+Note: Pre-built wheels are currently available for vLLM-Omni 0.11.0rc1, 0.12.0rc1, 0.14.0rc1, 0.14.0, 0.16.0, 0.18.0, 0.20.0, 0.21.0, 0.22.0, and 0.23.0. If you need a newer unreleased revision, please [build from source](https://docs.vllm.ai/projects/vllm-omni/en/latest/getting_started/installation/gpu/#build-wheel-from-source).
 
 === "NVIDIA CUDA"
 

docs/getting_started/installation/gpu/cuda.inc.md

@@ -20,7 +20,7 @@ Therefore, it is recommended to install vLLM and vLLM-Omni with a **fresh new**
 
 vLLM-Omni is built based on vLLM. Please install it with command below.
 ```bash
-uv pip install vllm==0.22.0 --torch-backend=auto
+uv pip install vllm==0.23.0 --torch-backend=auto
 ```
 
 #### Installation of vLLM-Omni
@@ -39,13 +39,13 @@ uv pip install 'vllm-omni[demo]'
 # --8<-- [start:build-wheel-from-source]
 
 #### Installation of vLLM
-If you do not need to modify source code of vLLM, you can directly install the stable 0.22.0 release version of the library
+If you do not need to modify source code of vLLM, you can directly install the stable 0.23.0 release version of the library
 
 ```bash
-uv pip install vllm==0.22.0 --torch-backend=auto
+uv pip install vllm==0.23.0 --torch-backend=auto
 ```
 
-The 0.22.0 release of vLLM ships CUDA 13.0-compatible binaries by default. If you need a different CUDA variant or want to reuse an existing PyTorch installation, build vLLM from source instead.
+The 0.23.0 release of vLLM ships CUDA 13.0-compatible binaries by default. If you need a different CUDA variant or want to reuse an existing PyTorch installation, build vLLM from source instead.
 
 #### Installation of vLLM-Omni
 Since vllm-omni is rapidly evolving, it's recommended to install it from source
@@ -66,12 +66,12 @@ If you want to check, modify or debug with source code of vLLM, install the libr
 ```bash
 git clone https://github.com/vllm-project/vllm.git
 cd vllm
-git checkout v0.22.0
+git checkout v0.23.0
 ```
 Set up environment variables to get pre-built wheels. If there are internet problems, just download the whl file manually. And set `VLLM_PRECOMPILED_WHEEL_LOCATION` as your local absolute path of whl file.
 ```bash
-#For CUDA 13.0 (the default for v0.22.0; the wheel filename has no `+cu130` suffix)
-export VLLM_PRECOMPILED_WHEEL_LOCATION=https://github.com/vllm-project/vllm/releases/download/v0.22.0/vllm-0.22.0-cp38-abi3-manylinux_2_35_x86_64.whl
+#For CUDA 13.0 (the default for v0.23.0; the wheel filename has no `+cu130` suffix)
+export VLLM_PRECOMPILED_WHEEL_LOCATION=https://github.com/vllm-project/vllm/releases/download/v0.23.0/vllm-0.23.0-cp38-abi3-manylinux_2_28_x86_64.whl
 ```
 Install vllm with command below (If you have no existing PyTorch).
 ```bash
@@ -102,7 +102,7 @@ docker run --runtime nvidia --gpus 2 \
     --env "HF_TOKEN=$HF_TOKEN" \
     -p 8091:8091 \
     --ipc=host \
-    vllm/vllm-omni:v0.22.0 \
+    vllm/vllm-omni:v0.23.0 \
     vllm serve Qwen/Qwen3-Omni-30B-A3B-Instruct --omni --port 8091
 ```
 

docs/getting_started/installation/gpu/rocm.inc.md

@@ -13,7 +13,7 @@ vLLM-Omni current recommends the steps in under setup through Docker Images.
 
 vLLM-Omni is built based on vLLM. Please install it with command below.
 ```bash
-uv pip install vllm==0.22.0+rocm722 --extra-index-url https://wheels.vllm.ai/rocm/0.22.0/rocm722
+uv pip install vllm==0.23.0+rocm722 --extra-index-url https://wheels.vllm.ai/rocm/0.23.0/rocm722
 ```
 
 #### Installation of vLLM-Omni
@@ -34,13 +34,13 @@ uv pip install onnxruntime-rocm
 # --8<-- [start:build-wheel-from-source]
 
 #### Installation of vLLM
-If you do not need to modify source code of vLLM, you can directly install the stable 0.22.0 release version of the library
+If you do not need to modify source code of vLLM, you can directly install the stable 0.23.0 release version of the library
 
 ```bash
-uv pip install vllm==0.22.0+rocm722 --extra-index-url https://wheels.vllm.ai/rocm/0.22.0/rocm722
+uv pip install vllm==0.23.0+rocm722 --extra-index-url https://wheels.vllm.ai/rocm/0.23.0/rocm722
 ```
 
-The pre-built 0.22.0 vLLM wheel targets ROCm 7.2.2. If you need a different ROCm stack or want to reuse an existing PyTorch installation, build vLLM from source instead.
+The pre-built 0.23.0 vLLM wheel targets ROCm 7.2.2. If you need a different ROCm stack or want to reuse an existing PyTorch installation, build vLLM from source instead.
 
 #### Installation of vLLM-Omni
 Since vllm-omni is rapidly evolving, it's recommended to install it from source
@@ -58,7 +58,7 @@ If you want to check, modify or debug with source code of vLLM, install the libr
 ```bash
 git clone https://github.com/vllm-project/vllm.git
 cd vllm
-git checkout v0.22.0
+git checkout v0.23.0
 python3 -m pip install -r requirements/rocm.txt
 python3 setup.py develop
 ```
@@ -130,7 +130,7 @@ docker run --rm \
   -v ~/.cache/huggingface:/root/.cache/huggingface \
   --env "HF_TOKEN=$HF_TOKEN" \
   -p 8091:8091 \
-  vllm/vllm-omni-rocm:v0.22.0 \
+  vllm/vllm-omni-rocm:v0.23.0 \
   --model Qwen/Qwen3-Omni-30B-A3B-Instruct --omni --port 8091
 ```
 
@@ -149,7 +149,7 @@ docker run --rm -it \
   -v ~/.cache/huggingface:/root/.cache/huggingface \
   --env "HF_TOKEN=$HF_TOKEN" \
   --entrypoint bash \
-  vllm/vllm-omni-rocm:v0.22.0
+  vllm/vllm-omni-rocm:v0.23.0
 ```
 
 # --8<-- [end:pre-built-images]

docs/getting_started/quickstart.md

@@ -19,10 +19,10 @@ uv venv --python 3.12 --seed
 source .venv/bin/activate
 
 # On CUDA
-uv pip install vllm==0.22.0 --torch-backend=auto
+uv pip install vllm==0.23.0 --torch-backend=auto
 
 # On ROCm
-uv pip install vllm==0.22.0+rocm721 --extra-index-url https://wheels.vllm.ai/rocm/0.22.0/rocm721
+uv pip install vllm==0.23.0+rocm722 --extra-index-url https://wheels.vllm.ai/rocm/0.23.0/rocm722
 
 git clone https://github.com/vllm-project/vllm-omni.git
 cd vllm-omni
@@ -35,7 +35,7 @@ For additional installation methods — please see the [installation guide](inst
 !!! note
     It is important to install the same major & minor version of vLLM and vLLM Omni, otherwise things may not work as expected. If the versions are misaligned, you will see a warning when you import vLLM Omni.
 
-    If you are seeing strange behavior with the `vllm` command not handling the `--omni` flag correctly, you most likely have a version mismatch with vLLM < `0.22.0` and vLLM Omni `0.22.0`, as vLLM Omni no longer hijacks the vLLM entrypoint. Updating vLLM should resolve this issue.
+    If you are seeing strange behavior with the `vllm` command not handling the `--omni` flag correctly, you most likely have a version mismatch with vLLM < `0.23.0` and vLLM Omni `0.23.0`, as vLLM Omni no longer hijacks the vLLM entrypoint. Updating vLLM should resolve this issue.
 
 ## Offline Inference
 

docs/models/supported_models.md

@@ -33,7 +33,7 @@ th {
 | `ZImagePipeline` | Z-Image | `Tongyi-MAI/Z-Image-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` | ✅︎ | ✅︎ | ✅︎ | ✅︎ |
-| `Cosmos3OmniDiffusersPipeline` | Cosmos3 T2I, T2V, I2V, T2V with sound, action policy | `nvidia/Cosmos3-Nano` | ✅︎ | | | |
+| `Cosmos3OmniDiffusersPipeline` | Cosmos3 T2I, T2V, I2V, V2V, T2V with sound, action policy | `nvidia/Cosmos3-Nano` | ✅︎ | ✅︎ | ✅︎ | ✅︎ |
 | `WanSpeechToVideoPipeline` | 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` | ✅︎ | ✅︎ | ✅︎ | ✅︎ |
 | `LTX2Pipeline` | LTX-2-T2V | `Lightricks/LTX-2` | ✅︎ | ✅︎ | | |
@@ -42,7 +42,7 @@ th {
 | `LTX2ImageToVideoTwoStagesPipeline` | LTX-2-I2V | `rootonchair/LTX-2-19b-distilled` | ✅︎ | ✅︎ | | |
 | `LTX23Pipeline` | LTX-2.3-T2V | `dg845/LTX-2.3-Diffusers` | ✅︎ | ✅︎ | | |
 | `LTX23ImageToVideoPipeline` | LTX-2.3-I2V | `dg845/LTX-2.3-Diffusers` | ✅︎ | ✅︎ | | |
-| `DreamZeroPipeline` | DreamZero-DROID | `GEAR-Dreams/DreamZero-DROID` | ✅︎ | | | |
+| `DreamZeroPipeline` | DreamZero-DROID | `GEAR-Dreams/DreamZero-DROID` | ✅︎ | ✅︎ | ✅︎ | ✅︎ |
 | `HeliosPipeline`, `HeliosPyramidPipeline` | Helios | `BestWishYsh/Helios-Base`, `BestWishYsh/Helios-Mid`, `BestWishYsh/Helios-Distilled` | ✅︎ | ✅︎ | ✅︎ | |
 | `MagiHumanPipeline` | MagiHuman | `SII-GAIR/daVinci-MagiHuman-Base-1080p` | ✅︎ | ✅︎ | | |
 | `OvisImagePipeline` | Ovis-Image | `OvisAI/Ovis-Image` | ✅︎ | ✅︎ | | ✅︎ |
@@ -50,6 +50,8 @@ th {
 | `LongCatImageEditPipeline` | LongCat-Image-Edit | `meituan-longcat/LongCat-Image-Edit` | ✅︎ | ✅︎ | ✅︎ | ✅︎ |
 | `StableDiffusion3Pipeline` | Stable-Diffusion-3 | `stabilityai/stable-diffusion-3.5-medium` | ✅︎ | ✅︎ | | ✅︎ |
 | `CosyVoice3Model` | CosyVoice3 | `FunAudioLLM/Fun-CosyVoice3-0.5B-2512` | ✅︎ | ✅︎ | | ✅︎ |
+| `OmniVoiceModel` | OmniVoice | `k2-fsa/OmniVoice` | ✅︎ | | | |
+| `VoxCPM2TalkerForConditionalGeneration` | VoxCPM2 | `openbmb/VoxCPM2` | ✅︎ | | | |
 | `MammothModa2ForConditionalGeneration` | MammothModa2-Preview | `bytedance-research/MammothModa2-Preview` | ✅︎ | ✅︎ | | |
 | `Flux2KleinPipeline` | FLUX.2-klein | `black-forest-labs/FLUX.2-klein-4B`, `black-forest-labs/FLUX.2-klein-9B` | ✅︎ | ✅︎ | ✅︎ | ✅︎ |
 | `FluxKontextPipeline` | FLUX.1-Kontext-dev | `black-forest-labs/FLUX.1-Kontext-dev` | ✅︎ | ✅︎ | | |
@@ -61,7 +63,12 @@ th {
 | `Qwen3TTSForConditionalGeneration` | Qwen3-TTS-12Hz-1.7B-CustomVoice | `Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice` | ✅︎ | ✅︎ | ✅︎ | ✅︎ |
 | `Qwen3TTSForConditionalGeneration` | Qwen3-TTS-12Hz-1.7B-VoiceDesign | `Qwen/Qwen3-TTS-12Hz-1.7B-VoiceDesign` | ✅︎ | ✅︎ | ✅︎ | ✅︎ |
 | `Qwen3TTSForConditionalGeneration` | Qwen3-TTS-12Hz-1.7B-Base | `Qwen/Qwen3-TTS-12Hz-0.6B-Base` | ✅︎ | ✅︎ | ✅︎ | ✅︎ |
+| `MingTTSForConditionalGeneration` | Ming-omni-tts dense 0.5B | `inclusionAI/Ming-omni-tts-0.5B` | ✅︎ | | | |
 | `GLMTTSForConditionalGeneration` | GLM-TTS | `zai-org/GLM-TTS` | ✅︎ | | | |
+| `MossTTSNanoForCausalLM` | MOSS-TTS-Nano | `OpenMOSS-Team/MOSS-TTS-Nano` | ✅︎ | | | |
+| `MossTTSDelayModel` | MOSS-TTS, MOSS-TTSD, MOSS-SoundEffect, MOSS-VoiceGenerator | `OpenMOSS-Team/MOSS-VoiceGenerator` | ✅︎ | | | |
+| `MossTTSRealtime` | MOSS-TTS-Realtime | `OpenMOSS-Team/MOSS-TTS-Realtime` | ✅︎ | | | |
+| `HiggsAudioV2ForConditionalGeneration` | Higgs-Audio v2 | `bosonai/higgs-audio-v2-generation-3B-base` | ✅︎ | | | |
 | `NextStep11Pipeline` | NextStep-1.1 | `stepfun-ai/NextStep-1.1` | ✅︎ | ✅︎ | | ✅︎ |
 | `MiMoAudioModel` | MiMo-Audio-7B-Instruct | `XiaomiMiMo/MiMo-Audio-7B-Instruct` | ✅︎ | ✅︎ | | |
 | `MiMoV2ASRForCausalLM` | MiMo-V2.5-ASR | `XiaomiMiMo/MiMo-V2.5-ASR` | ✅︎ | ✅︎ | | |
@@ -69,11 +76,13 @@ th {
 | `FishSpeechSlowARForConditionalGeneration` | Fish Speech S2 Pro | `fishaudio/s2-pro` | ✅︎ | ✅︎ | | |
 | `DreamIDOmniPipeline` | DreamID-Omni | `XuGuo699/DreamID-Omni` | ✅︎ | ✅︎ | | |
 | `SenseNovaU1Pipeline` | SenseNova-U1 (DiT-only) | `SenseNova/SenseNova-U1-8B-MoT` | ✅︎ | | | |
+| `LancePipeline` | Lance | `bytedance-research/Lance` | ✅︎ | | | |
 | `HunyuanVideo15Pipeline` | HunyuanVideo-1.5-T2V | `hunyuanvideo-community/HunyuanVideo-1.5-Diffusers-480p_t2v`, `hunyuanvideo-community/HunyuanVideo-1.5-Diffusers-720p_t2v` | ✅︎ | ✅︎ | | |
 | `HunyuanVideo15ImageToVideoPipeline` | HunyuanVideo-1.5-I2V | `hunyuanvideo-community/HunyuanVideo-1.5-Diffusers-480p_i2v`, `hunyuanvideo-community/HunyuanVideo-1.5-Diffusers-720p_i2v` | ✅︎ | ✅︎ | | |
 | `VoxtralTTSForConditionalGeneration` | Voxtral TTS | `mistralai/Voxtral-4B-TTS-2603` | ✅︎ | ✅︎ | | |
 | `CovoAudioForConditionalGeneration` | Covo-Audio-Chat | `tencent/Covo-Audio-Chat` | ✅︎ | | | |
 |`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` | ✅︎ | ✅︎ | ✅︎ | ✅︎ |
 |`HiDreamImagePipeline` | HiDream-I1-Full | `HiDream-ai/HiDream-I1-Full` | ✅︎ | ✅︎ | | |
 

docs/serving/speech_api.md

@@ -114,9 +114,10 @@ Content-Type: application/json
 | `instructions` | string | "" | Voice style/emotion instructions |
 | `max_new_tokens` | integer | 2048 | Maximum tokens to generate |
 | `initial_codec_chunk_frames` | integer | null | Per-request initial chunk size override for TTFA tuning. When null, IC is computed dynamically based on server load. |
+| `non_streaming_mode` | bool | null | Qwen3-TTS prompt construction mode override. Does not affect HTTP response streaming or async-chunk pipelining. When null, Qwen3-TTS uses model defaults: Base=false, CustomVoice/VoiceDesign=true. |
 | `stream` | bool | false | Stream raw PCM chunks as they are decoded (requires `response_format="pcm"`) |
 
-**Supported languages:** Auto, Chinese, English, Japanese, Korean, German, French, Russian, Portuguese, Spanish, Italian
+**Supported languages:** Only applicable to Qwen3-TTS. Derived from the model configuration (`talker_config.codec_language_id` in the checkpoint's `config.json`), plus `Auto`, which is always accepted. Official Qwen3-TTS checkpoints support: Auto, Chinese, English, Japanese, Korean, German, French, Russian, Portuguese, Spanish, Italian.
 
 #### Voice Clone Parameters (Base task)
 
@@ -332,7 +333,8 @@ curl -X POST http://localhost:8091/v1/audio/speech \
         "input": "Hello, this is a cloned voice",
         "task_type": "Base",
         "ref_audio": "https://example.com/reference.wav",
-        "ref_text": "Original transcript of the reference audio"
+        "ref_text": "Original transcript of the reference audio",
+        "non_streaming_mode": true
     }' --output cloned.wav
 ```
 
@@ -454,6 +456,7 @@ Content-Type: application/json
 | `ref_audio` | string | null | Default reference audio (Base task) |
 | `ref_text` | string | null | Default reference transcript (Base task) |
 | `max_new_tokens` | integer | null | Default max tokens |
+| `non_streaming_mode` | bool | null | Default Qwen3-TTS prompt construction mode override. Does not affect HTTP response streaming or async-chunk pipelining. When null, Qwen3-TTS uses model defaults: Base=false, CustomVoice/VoiceDesign=true. |
 
 Each item in the `items` array requires only `input` (the text). All other fields are optional and override the batch-level defaults when set:
 
@@ -469,6 +472,7 @@ Each item in the `items` array requires only `input` (the text). All other field
 | `ref_audio` | string | Override reference audio |
 | `ref_text` | string | Override reference transcript |
 | `max_new_tokens` | integer | Override max tokens |
+| `non_streaming_mode` | bool | Override Qwen3-TTS prompt construction mode. Does not affect HTTP response streaming or async-chunk pipelining. When null, inherits the batch-level value (then the model default). |
 
 ### Response Format
 

docs/serving/videos_api.md

@@ -67,8 +67,10 @@ curl -L "http://localhost:8091/v1/videos/${video_id}/content" -o output.mp4
 
 | Parameter | Type | Default | Description |
 |-----------|------|---------|-------------|
-| `input_reference` | file | null | Uploaded reference image for image-to-video requests |
-| `image_reference` | string | null | JSON-encoded reference image payload; do not combine with `input_reference` |
+| `input_reference` | file | null | Uploaded reference image or video for image-to-video/video-to-video requests |
+| `image_reference` | string | null | JSON-encoded reference image payload; do not combine with `input_reference` or `video_reference` |
+| `video_reference` | string | null | JSON-encoded reference video payload; do not combine with `input_reference` or `image_reference` |
+| `audio_reference` | string | null | JSON-encoded audio reference for speech-to-video: `{"audio_url": "..."}` — supports HTTP(s) URLs or base64 data URLs |
 | `width` | integer | model default | Output video width |
 | `height` | integer | model default | Output video height |
 | `num_frames` | integer | model default | Number of generated frames |
@@ -80,6 +82,8 @@ curl -L "http://localhost:8091/v1/videos/${video_id}/content" -o output.mp4
 | `flow_shift` | number | null | Scheduler flow-shift value |
 | `true_cfg_scale` | number | null | True CFG scale when supported by the model |
 | `seed` | integer | null | Random seed for reproducibility |
+| `generate_sound` | boolean | false | Request model-generated audio for video models that support sound generation |
+| `sound_duration` | number | null | Duration in seconds for generated audio; defaults to generated video duration |
 | `negative_prompt` | string | null | Text describing what to avoid in the generated video |
 | `enable_frame_interpolation` | boolean | null | Enable post-generation frame interpolation |
 | `frame_interpolation_exp` | integer | null | Interpolation exponent; `1=2x`, `2=4x`, and so on |
@@ -123,6 +127,60 @@ curl -s http://localhost:8091/v1/videos \
   -F "fps=16"
 ```
 
+### Video-to-Video
+
+For models that support video conditioning, upload the reference video with
+`input_reference`:
+
+```bash
+curl -s http://localhost:8091/v1/videos \
+  -F "prompt=continue this motion with consistent subjects and lighting" \
+  -F "input_reference=@input.mp4;type=video/mp4" \
+  -F "width=1280" \
+  -F "height=720" \
+  -F "num_frames=80" \
+  -F "fps=16"
+```
+
+You can also pass a JSON-safe video URL or `data:video/...;base64,...` payload
+through `video_reference`. Do not send `video_reference` together with
+`input_reference` or `image_reference`.
+
+```bash
+curl -s http://localhost:8091/v1/videos \
+  -F "prompt=continue this motion with consistent subjects and lighting" \
+  -F 'video_reference={"video_url":"https://example.com/input.mp4"}' \
+  -F "width=1280" \
+  -F "height=720" \
+  -F "num_frames=80" \
+  -F "fps=16"
+```
+
+JSON references currently support `image_url`/`video_url`; `file_id` references
+are not implemented yet. Models may expose additional V2V controls through
+`extra_params`. For example, Cosmos3 supports
+`condition_frame_indexes_vision` and `condition_video_keep` to select which
+decoded reference frames are used as clean conditioning.
+
+### Speech-to-Video
+
+For models that support audio-driven generation (e.g., Wan2.2-S2V), pass both
+an image reference and an audio reference. The `audio_reference` field accepts a
+JSON string with `audio_url` pointing to an HTTP(s) URL or base64 data URL.
+
+```bash
+curl -s http://localhost:8091/v1/videos \
+  -F "prompt=A person singing" \
+  -F 'image_reference={"image_url": "https://example.com/face.png"}' \
+  -F 'audio_reference={"audio_url": "https://example.com/speech.mp3"}' \
+  -F "width=832" \
+  -F "height=480" \
+  -F "num_inference_steps=40" \
+  -F "guidance_scale=4.5" \
+  -F "fps=16"
+```
+
+
 ### Synchronous Generation
 
 ```bash
@@ -146,7 +204,12 @@ export VLLM_OMNI_STORAGE_PATH=/var/tmp/vllm-omni-videos
 
 ## Model-Specific Examples
 
-For complete text-to-video and image-to-video walkthroughs, see:
+For complete text-to-video, image-to-video, and model-specific video-to-video
+walkthroughs, see:
 
 - [Text-to-Video](../user_guide/examples/online_serving/text_to_video.md)
 - [Image-to-Video](../user_guide/examples/online_serving/image_to_video.md)
+- [Speech-to-Video](../user_guide/examples/online_serving/speech_to_video.md)
+  for Wan2.2-S2V audio-driven lip-sync generation
+- [Cosmos3 recipes](https://github.com/vllm-project/vllm-omni/blob/main/recipes/cosmos3/Cosmos3-Nano.md)
+  for model-specific video-to-video examples and conditioning controls

docs/user_guide/diffusion_features.md

@@ -148,7 +148,7 @@ The following tables show which models support each feature:
 | **Wan2.2-S2V**               |     ❌     |     ✅      |           ❌           |       ❌        |         ✅         |         ❌         |   ❌    |             ✅             |  ✅ (encode/decode)   |       ❌        |        ❌         |
 | **Wan2.1-VACE**              |     ❌     |     ✅      |           ✅           |       ✅        |         ✅         |         ❌         |   ✅    |             ✅             |      ✅ (decode)      |       ❌        |        ❌         |
 | **LTX-2**                    |     ❌     |     ✅      |           ✅           |       ✅        |         ✅         |         ❌         |   ✅    |             ✅             |          ❌           |       ❌        |        ❌         |
-| **LTX-2.3**                  |     ❌     |     ✅      |           ✅           |       ✅        |         ✅         |         ❌         |   ❌    |             ❌             |          ❌           |       ❌        |        ❌         |
+| **LTX-2.3**                  |     ❌     |     ✅      |           ✅           |       ✅        |         ✅         |         ❌         |   ❌    |             ✅             |      ✅ (decode)      |       ❌        |        ❌         |
 | **Helios**                   |     ❌     |     ✅      |           ✅           |       ✅        |         ✅         |         ❌         |   ✅    |             ✅             |          ❌           |       ❌        |        ❌         |
 | **HunyuanVideo-1.5 T2V I2V** |     ❌     |     ✅      |           ✅           |       ✅        |         ✅         |         ❌         |   ✅    |             ✅             |  ✅ (encode/decode)   |       ✅        |        ❌         |
 | **DreamID-Omni**             |     ❌     |     ❌      |           ❌           |       ✅        |         ❌         |         ❌         |   ✅    |             ✅             |          ❌           |       ❌        |        ❌         |

docs/user_guide/examples/online_serving/speech_to_video.md

@@ -0,0 +1,128 @@
+# Speech-To-Video
+
+Source <https://github.com/vllm-project/vllm-omni/tree/main/examples/online_serving/speech_to_video>.
+
+This example demonstrates how to deploy the Wan2.2 speech-to-video (S2V) model for online video generation using vLLM-Omni.
+
+## Supported Models
+
+| Model | Model ID |
+|-------|----------|
+| Wan2.2 S2V (14B) | `Wan-AI/Wan2.2-S2V-14B` |
+
+## Start Server
+
+### Basic Start
+
+```bash
+VLLM_WORKER_MULTIPROC_METHOD=spawn \
+vllm serve Wan-AI/Wan2.2-S2V-14B --omni \
+  --model-class-name WanS2VPipeline \
+  --tensor-parallel-size 2 \
+  --flow-shift 3.0 \
+  --vae-use-slicing --vae-use-tiling \
+  --cache-backend cache_dit \
+  --port 8091
+```
+
+### Start with Script
+
+```bash
+bash run_server.sh
+```
+
+The script allows overriding:
+- `MODEL` (default: `Wan-AI/Wan2.2-S2V-14B`)
+- `PORT` (default: `8091`)
+- `FLOW_SHIFT` (default: `3.0`)
+- `TP` (default: `2`)
+- `CACHE_BACKEND` (default: `cache_dit`)
+
+### Start with 4 GPUs
+
+```bash
+TP=4 bash run_server.sh
+```
+
+## API Endpoints
+
+The server provides OpenAI-compatible video generation endpoints:
+
+- `POST /v1/videos`: create a video generation job (async)
+- `POST /v1/videos/sync`: generate a video and return raw bytes (sync, for benchmarks)
+- `GET /v1/videos/{video_id}`: retrieve the current job status and metadata
+- `GET /v1/videos`: list stored video jobs
+- `GET /v1/videos/{video_id}/content`: download the generated video file
+- `DELETE /v1/videos/{video_id}`: delete the job and any stored output
+
+## Sync API (Recommended for Testing)
+
+`POST /v1/videos/sync` blocks until generation completes and returns the raw
+video bytes (`video/mp4`) directly in the response body.
+
+```bash
+bash run_curl_speech_to_video.sh
+```
+
+## Async Job API
+
+`POST /v1/videos` creates an asynchronous job and returns immediately.
+
+```bash
+no_proxy=127.0.0.1 \
+create_response=$(curl -sS -X POST http://127.0.0.1:8091/v1/videos \
+  -H "Accept: application/json" \
+  -F "prompt=A person singing" \
+  -F 'image_reference={"image_url": "https://raw.githubusercontent.com/Wan-Video/Wan2.2/main/examples/Five%20Hundred%20Miles.png"}' \
+  -F 'audio_reference={"audio_url": "https://raw.githubusercontent.com/Wan-Video/Wan2.2/main/examples/Five%20Hundred%20Miles.MP3"}' \
+  -F "width=832" -F "height=480" \
+  -F "num_inference_steps=40" \
+  -F "guidance_scale=4.5" \
+  -F "fps=16")
+
+video_id=$(echo "$create_response" | jq -r '.id')
+
+# Poll until complete
+while true; do
+  status=$(curl -s "http://127.0.0.1:8091/v1/videos/${video_id}" | jq -r '.status')
+  if [ "$status" = "completed" ]; then
+    break
+  fi
+  if [ "$status" = "failed" ]; then
+    echo "Video generation failed"
+    exit 1
+  fi
+  sleep 2
+done
+
+# Download
+curl -L "http://127.0.0.1:8091/v1/videos/${video_id}/content" -o s2v_output.mp4
+```
+
+## Request Parameters
+
+| Parameter | Type | Default | Description |
+|-----------|------|---------|-------------|
+| `prompt` | str | - | Text description of the desired video |
+| `image_reference` | JSON str | - | Image reference: `{"image_url": "..."}` — supports HTTP(s) URLs or base64 data URLs |
+| `audio_reference` | JSON str | - | Audio reference: `{"audio_url": "..."}` — supports HTTP(s) URLs or base64 data URLs |
+| `width` | int | None | Video width in pixels |
+| `height` | int | None | Video height in pixels |
+| `fps` | int | None | Output video frame rate |
+| `num_inference_steps` | int | None | Number of denoising steps (40 recommended) |
+| `guidance_scale` | float | None | CFG guidance scale |
+| `seed` | int | None | Random seed for reproducibility |
+
+## Notes
+
+- S2V requires both a reference image (`image_reference`) and an audio reference
+  (`audio_reference`). The generated video will show a person matching the reference
+  image with lip movements synchronized to the audio.
+- `audio_reference` accepts a JSON string: `{"audio_url": "..."}` where the URL can be:
+  - An HTTP/HTTPS URL (e.g., `https://example.com/audio.mp3`)
+  - A base64 data URL (e.g., `data:audio/mp3;base64,...`)
+- `--model-class-name WanS2VPipeline` is required on the server to select the
+  S2V pipeline (distinct from the T2V/I2V pipelines).
+- `--cache-backend cache_dit` enables DiT caching for ~2x speedup on cached steps.
+- Audio is muxed into the output MP4 automatically.
+- Always pass `fps=16` to ensure correct video/audio alignment.

docs/user_guide/fault_injection_reliability_matrix.md

@@ -0,0 +1,117 @@
+# vLLM-Omni Failure Mode Injection Scenarios and Expected Behavior Matrix
+
+## Background and Goals
+
+The online serving pipeline of vLLM-Omni can be abstracted as: `Serve` (service entry process) -> `Engine` (model orchestration unit) -> `Worker` (execution process) -> `GPU` (compute resource).
+
+- `Serve` provides external APIs, accepts requests, and manages lifecycle.
+- `Engine` handles request scheduling and inference orchestration; the number and organization of engines can vary by model type.
+- `Worker` performs concrete model computation, usually bound to underlying GPU resources.
+- `GPU` is the resource layer that ultimately carries inference compute and memory usage.
+
+<div markdown style="text-align: center">
+
+```mermaid
+%%{init: {'flowchart': {'useMaxWidth': false}}}%%
+flowchart TD
+    R(request)--> A(serve)
+    A --> B(engine)
+    A --> C(engine)
+    A --> D(engine)
+    B --> E(worker)
+    C --> F(worker)
+    D --> I(worker)
+    E --> G(GPU)
+    F --> H(GPU)
+    I --> L(GPU)
+```
+
+</div>
+
+The diagram above shows an abstract relationship. Omni models and Diffusion models share the same hierarchy; differences mainly come from engine decomposition, worker count, and scheduling strategy.
+
+In real customer environments, the system will inevitably encounter accidental operations or environmental disturbances, such as unintended process termination, container/node jitter, or short-term resource contention. From the system perspective, many of these events map to two fault categories: "process receives abnormal signals" or "GPU memory is squeezed." For example, pressing `Ctrl+C` essentially sends a `SIGINT` signal to the `serve` process; running `kill -15 <pid>` maps to graceful termination via `SIGTERM`; running `kill -9 <pid>` maps to forced termination via `SIGKILL`; when `docker kill` is executed or a Pod is deleted, the container runtime usually sends `SIGKILL`.
+
+This document currently covers two core failure modes, and the system is expected to show predictable and observable behavior under each mode:
+
+- **Failure Mode 1: Process receives abnormal signals (`SIGINT` / `SIGTERM` / `SIGKILL`)**  
+  Focus on three aspects: whether processes exit as expected and complete cleanup, whether requests fail fast or connections are interrupted, and whether GPU resources are eventually fully released with no residue.
+- **Failure Mode 2: OOM (occupy all free GPU memory via an extra process)**  
+  Focus on two aspects: whether service health degrades to the expected state (for example, `503`), and whether requests fail within an acceptable time without hanging.
+
+More failure modes (for example, network jitter and network interruption) will be added in future iterations to improve end-to-end reliability validation coverage.
+
+> Note: independent fault injection on the `engine` component is not covered in the current version and will be added gradually in future versions.
+
+## Fault Injection Scenario Matrix
+
+| Scenario | Fault Type | System Behavior | Current Status |
+|------|----------------------------|----------|----------|
+| No load | Send `SIGKILL` to Worker | Worker process is killed immediately; main service detects child-process loss and turns unavailable; API enters stable 5xx |  |
+| No load | Send `SIGTERM` to Worker | Worker exits after receiving termination signal; main service is marked unavailable; API enters stable 5xx |  |
+| No load | Send `SIGKILL` to serve main process | serve main process exits instantly; request connections are interrupted; related child processes are cleaned up with no residue; GPU memory is released quickly | [#3725](https://github.com/vllm-project/vllm-omni/issues/3725) <br>[#43060](https://github.com/vllm-project/vllm/issues/43060) |
+| No load | Send `SIGTERM` to serve main process | serve enters graceful shutdown and stops serving; then exits and completes cleanup; GPU memory is released |  |
+| No load | Send `SIGINT` to serve main process (equivalent to `Ctrl+C`) | Triggers serve shutdown path; service stops responding and becomes unavailable; related child processes exit and resources are released |  |
+| No load | Send `SIGKILL` to all related processes | All related processes terminate immediately; service becomes unavailable at once; no residual processes remain; GPU memory is released quickly |  |
+| No load | Send `SIGTERM` to all related processes | All processes enter exit flow and complete shutdown; service becomes unavailable; resource release is completed |  |
+| Under load | Send `SIGKILL` to Worker | In-flight requests are hard interrupted (5xx/connection drop); main service becomes unavailable | [#3683](https://github.com/vllm-project/vllm-omni/issues/3683) |
+| Under load | Send `SIGTERM` to Worker | In-flight requests are canceled or fail fast; main service becomes unavailable |  |
+| Under load | Send `SIGKILL` to serve main process | serve is hard-killed and current connections are interrupted; in-flight requests fail; after cleanup there are no residual processes and GPU memory is released | [#3683](https://github.com/vllm-project/vllm-omni/issues/3683) |
+| Under load | Send `SIGTERM` to serve main process | serve stops accepting new requests and executes shutdown flow; in-flight requests fail; no residue remains and GPU memory is released | [#3683](https://github.com/vllm-project/vllm-omni/issues/3683) |
+| Under load | Send `SIGINT` to serve main process (equivalent to `Ctrl+C`) | `Ctrl+C`-style serve shutdown; in-flight requests fail (5xx/connection interruption); service unavailable; after exit there is no residue and GPU memory is released | [#3683](https://github.com/vllm-project/vllm-omni/issues/3683) |
+| Under load | Send `SIGKILL` to all related processes | All processes terminate instantly; all in-flight requests fail; service becomes unavailable immediately; GPU memory is released quickly |  |
+| Under load | Send `SIGTERM` to all related processes | All processes exit gracefully; in-flight requests fail; service unavailable; GPU memory is released | [#3683](https://github.com/vllm-project/vllm-omni/issues/3683) |
+| OOM | Occupy all free GPU memory via an extra process | After OOM injection process starts, GPU memory is continuously saturated; service enters unavailable/degraded state and health check drops to 503; different request types (chat/speech, etc.) fail fast within a fixed time and return 500 (no hanging) | [#4285](https://github.com/vllm-project/vllm-omni/issues/4285) |
+
+## Source of Conclusions
+
+The behaviors and conclusions above are summarized from current fault injection validation results on `Qwen3-Omni` and `Wan2.2`.
+
+### Example: `SIGTERM` fault injection log (Qwen3-Omni)
+
+The following full log is from a real run where a `SIGTERM` signal was sent to the serve root process.
+
+```text
+Launching OmniServer with: /workspace/.venv/bin/python3 -m vllm_omni.entrypoints.cli.main serve Qwen/Qwen3-Omni-30B-A3B-Instruct --host 127.0.0.1 --port 60675 --omni --async-chunk --stage-init-timeout 600 --init-timeout 900 --log-stats --stage-configs-path vllm-omni/vllm_omni/deploy/qwen3_omni_moe.yaml
+Server ready on 127.0.0.1:60675 (OmniServer startup took 206.023s)
+OmniServer started successfully
+[reliability][process-kill] current_server_proc pid=556855 name=python3 cmdline=/workspace/.venv/bin/python3 -m vllm_omni.entrypoints.cli.main serve Qwen/Qwen3-Omni-30B-A3B-Instruct --host 127.0.0.1 --port 60675 --omni --async-chunk --stage-init-timeout 600 --init-timeout 900 --log-stats --stage-configs-path vllm-omni/vllm_omni/deploy/qwen3_omni_moe.yaml
+[reliability][process-kill] current_server_proc pid=557118 name=python3 cmdline=/workspace/.venv/bin/python3 -c from multiprocessing.resource_tracker import main;main(57)
+[reliability][process-kill] current_server_proc pid=557119 name=VLLM::StageEngineCoreProc_noid_replica0_DP0 cmdline=VLLM::StageEngineCoreProc_noid_replica0_DP0
+[reliability][process-kill] current_server_proc pid=557122 name=VLLM::StageEngineCoreProc_noid_replica0_DP0 cmdline=VLLM::StageEngineCoreProc_noid_replica0_DP0
+[reliability][process-kill] current_server_proc pid=558201 name=VLLM::StageEngineCoreProc_noid_replica0_DP0 cmdline=VLLM::StageEngineCoreProc_noid_replica0_DP0
+
+[reliability][process-kill] root-kill pid=556855 name=python3 signal=SIGTERM cmdline=/workspace/.venv/bin/python3 -m vllm_omni.entrypoints.cli.main serve Qwen/Qwen3-Omni-30B-A3B-Instruct --host 127.0.0.1 --port 60675 --omni --async-chunk --stage-init-timeout 600 --init-timeout 900 --log-stats --stage-configs-path vllm-omni/vllm_omni/deploy/qwen3_omni_moe.yaml
+[0;36m(APIServer pid=556855)[0;0m INFO 05-19 12:24:51 [omni_base.py:463] [AsyncOmni] Shutting down
+[0;36m(APIServer pid=556855)[0;0m INFO 05-19 12:24:51 [async_omni_engine.py:2397] [AsyncOmniEngine] Shutting down Orchestrator
+[0;36m(APIServer pid=556855)[0;0m INFO 05-19 12:24:51 [orchestrator.py:351] [Orchestrator] Received shutdown signal
+[0;36m(APIServer pid=556855)[0;0m INFO 05-19 12:24:51 [orchestrator.py:1359] [Orchestrator] Shutting down all 3 client(s)
+
+[0;36m(StageEngineCoreProc pid=557119)[0;0m [32mINFO[0m [90m05-19 12:24:51[0m [90m[core.py:1242][0m Shutdown initiated (timeout=0)
+[0;36m(StageEngineCoreProc pid=557119)[0;0m [32mINFO[0m [90m05-19 12:24:51[0m [90m[core.py:1265][0m Shutdown complete
+[0;36m(APIServer pid=556855)[0;0m INFO:     Shutting down
+[0;36m(APIServer pid=556855)[0;0m INFO:     Waiting for application shutdown.
+[0;36m(APIServer pid=556855)[0;0m INFO:     Application shutdown complete.
+[0;36m(APIServer pid=556855)[0;0m INFO:     Finished server process [556855]
+[0;36m(APIServer pid=556855)[0;0m INFO 05-19 12:24:51 [omni_base.py:463] [AsyncOmni] Shutting down
+[rank0]:[W519 12:24:53.427894833 ProcessGroupNCCL.cpp:1575] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator())
+[0;36m(APIServer pid=556855)[0;0m INFO 05-19 12:24:56 [stage_pool.py:731] [StagePool] Stage 0 replica 0 shut down
+[0;36m(StageEngineCoreProc pid=557122)[0;0m [32mINFO[0m [90m05-19 12:24:56[0m [90m[core.py:1242][0m Shutdown initiated (timeout=0)
+[0;36m(StageEngineCoreProc pid=557122)[0;0m [32mINFO[0m [90m05-19 12:24:56[0m [90m[core.py:1265][0m Shutdown complete
+[rank0]:[W519 12:24:57.569425065 ProcessGroupNCCL.cpp:1575] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator())
+[0;36m(APIServer pid=556855)[0;0m INFO 05-19 12:25:00 [stage_pool.py:731] [StagePool] Stage 1 replica 0 shut down
+[0;36m(StageEngineCoreProc pid=558201)[0;0m [32mINFO[0m [90m05-19 12:25:00[0m [90m[core.py:1242][0m Shutdown initiated (timeout=0)
+[0;36m(StageEngineCoreProc pid=558201)[0;0m [32mINFO[0m [90m05-19 12:25:00[0m [90m[core.py:1265][0m Shutdown complete
+[0;36m(APIServer pid=556855)[0;0m WARNING 05-19 12:25:01 [async_omni_engine.py:2406] [AsyncOmniEngine] Orchestrator thread did not exit in time
+[rank0]:[W519 12:25:01.867242016 ProcessGroupNCCL.cpp:1575] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator())
+```
+
+### How this maps to the matrix
+
+- This log corresponds to the `Send SIGTERM to serve main process` scenario in the matrix.
+- The `root-kill ... signal=SIGTERM` line confirms the fault injection action is applied as expected.
+- The sequence from `Shutting down Orchestrator` to stage-level `Shutdown complete` and `Stage ... shut down` matches the expected graceful shutdown path and child-process cleanup behavior described in the matrix.
+
+## Related Documentation
+
+- For practical CI debugging procedures and triage guidance, see [CI Failures](../contributing/ci/failures.md).

docs/user_guide/quantization/fp8.md

@@ -28,6 +28,50 @@ in deep DiT blocks.
 Legend: `✅` supported, `❌` unsupported, `⭕` not verified in this
 guide. FP8 on Ampere may use a weight-only path where available.
 
+### Faster FP8 GEMM on Blackwell (quack)
+
+On Blackwell (SM 100+), vLLM runs FP8 linears through the FlashInfer kernel, which
+applies the bias as a separate kernel after the GEMM. On the small GEMMs in video
+DiTs this bias add is a significant overhead. Installing the optional `quack` kernel
+lets vLLM-Omni fuse `alpha * (A @ B) + bias` into a single CuteDSL GEMM, recovering
+that overhead (e.g. HunyuanVideo-1.5 FP8 goes from slower-than-BF16 to faster).
+
+```bash
+# CUDA 12.9
+pip install vllm-omni[quack]
+
+# CUDA 13.x
+pip install 'quack-kernels[cu13]' --extra-index-url https://download.pytorch.org/whl/cu130
+```
+
+It is enabled automatically once installed (no flag needed) and is **Blackwell-only**:
+on Hopper/Ada the CUTLASS FP8 kernel already fuses bias, so quack is not used there.
+Set `VLLM_OMNI_USE_QUACK_FP8=0` to force the FlashInfer path. If `quack-kernels` is
+not installed, FP8 still works — it just keeps the unfused FlashInfer path.
+
+#### Compile cache and warmup
+
+quack JIT-compiles its kernel once per distinct GEMM shape (tens of seconds, longer
+the first time across all autotuned configs). The compiled `.o` files are cached on
+disk and reused on later runs, so this is a one-time cost — **not** per request.
+
+vLLM-Omni points that cache at `~/.cache/vllm_omni/quack` (override with
+`QUACK_CACHE_DIR`) instead of quack's default under `/tmp`, so it survives restarts.
+In containers, set `QUACK_CACHE_DIR` to a mounted/persistent path — or bake it into
+the image — so the first cold start does not recompile. The engine's startup dummy
+run already exercises the kernels, so with a warm cache the first real request is fast.
+
+To pre-warm specific shapes (e.g. at image build time):
+
+```python
+from vllm_omni.quantization.quack_fp8 import warmup_quack_fp8
+# (M, K, N) per linear; M = number of tokens for your resolution/frame count
+warmup_quack_fp8([(14040, 2048, 6144), (14040, 2048, 2048)])
+```
+
+> The PyPI package is `quack-kernels` (imported as `quack`); plain `pip install
+> quack` is an unrelated statistics library. Requires CUDA 12.9+ and Python 3.12.
+
 ## Model Type Support
 
 ### Diffusion Model (Qwen-Image, Wan2.2)

examples/offline_inference/cosmos3/inputs/v2v.json

@@ -0,0 +1,14 @@
+{
+    "prompt": "A robotic arm, primarily white with black joints and cables, is shown in a clean, modern indoor setting with a white tabletop. The arm, equipped with a gripper holding a small, light green pitcher, is positioned above a clear glass containing a reddish-brown liquid and a spoon. The robotic arm is in the process of pouring a transparent liquid into the glass. To the left of the pitcher, there is an opened jar with a similar reddish-brown substance visible through its transparent body. In the background, a vase with white flowers and a brown couch are partially visible, adding to the contemporary ambiance. The lighting is bright, casting soft shadows on the table. The robotic arm's movements are smooth and controlled, demonstrating precision in its task. As the video progresses, the robotic arm completes the pour, leaving the glass half-filled with the reddish-brown liquid. The jar remains untouched throughout the sequence, and the spoon inside the glass remains stationary. The other robotic arm on the right side also stays stationary throughout the video. The final frame captures the robotic arm with the pitcher finishing the pour, with the glass now filled to a higher level, while the pitcher is slightly tilted but still held securely by the gripper.",
+    "negative_prompt": "",
+    "vision_path": "https://github.com/nvidia-cosmos/cosmos-dependencies/raw/refs/heads/assets/cosmos3/inputs/vision/robot_pouring.mp4",
+    "height": 720,
+    "width": 1280,
+    "num_frames": 189,
+    "num_inference_steps": 35,
+    "guidance_scale": 6.0,
+    "flow_shift": 10.0,
+    "fps": 24,
+    "condition_frame_indexes_vision": [0, 1],
+    "condition_video_keep": "first"
+}

examples/offline_inference/image_to_image/image_edit.py

@@ -143,6 +143,20 @@ def parse_args() -> argparse.Namespace:
         default=None,
         required=False,
     )
+    parser.add_argument(
+        "--width",
+        type=int,
+        default=None,
+        metavar="W",
+        help="Output image width in pixels. Default: None (pipeline's default).",
+    )
+    parser.add_argument(
+        "--height",
+        type=int,
+        default=None,
+        metavar="H",
+        help="Output image height in pixels. Default: None (pipeline's default).",
+    )
     parser.add_argument(
         "--seed",
         type=int,
@@ -236,8 +250,8 @@ def parse_args() -> argparse.Namespace:
     parser.add_argument(
         "--resolution",
         type=int,
-        default=640,
-        help="Bucket in (640, 1024) to determine the condition and output resolution",
+        default=None,
+        help="Bucket in (640, 1024) to determine the condition and output resolution. If width and height are not provided, this will be set to default 640.",
     )
 
     parser.add_argument(
@@ -384,6 +398,15 @@ def parse_args() -> argparse.Namespace:
 def main():
     args = parse_args()
 
+    if args.resolution and (args.width or args.height):
+        raise ValueError("--resolution and --width/--height cannot be specified together")
+    if args.width is not None and args.width <= 0:
+        raise ValueError("--width must be a positive integer")
+    if args.height is not None and args.height <= 0:
+        raise ValueError("--height must be a positive integer")
+    if not args.width and not args.height and not args.resolution:
+        args.resolution = 640
+
     # Validate input images exist and load them
     input_images = []
     for image_path in args.image:
@@ -464,6 +487,7 @@ def main():
     print(
         f"  Parallel configuration: ulysses_degree={args.ulysses_degree}, ring_degree={args.ring_degree}, cfg_parallel_size={args.cfg_parallel_size}, tensor_parallel_size={args.tensor_parallel_size}, enable_expert_parallel: {args.enable_expert_parallel}"
     )
+    print(f"  Output image size: {args.width or 'default'}x{args.height or 'default'}")
     print(f"{'=' * 60}\n")
 
     generation_start = time.perf_counter()
@@ -473,22 +497,26 @@ def main():
         omni.start_profile()
 
     # Generate edited image
+    sampling_params = OmniDiffusionSamplingParams(
+        width=args.width,
+        height=args.height,
+        generator=generator,
+        true_cfg_scale=args.cfg_scale,
+        guidance_scale=args.guidance_scale,
+        guidance_scale_2=args.guidance_scale_2,
+        num_inference_steps=args.num_inference_steps,
+        num_outputs_per_prompt=args.num_outputs_per_prompt,
+        layers=args.layers,
+    )
+    if args.resolution:
+        sampling_params.resolution = args.resolution
     outputs = omni.generate(
         {
             "prompt": args.prompt,
             "negative_prompt": args.negative_prompt,
             "multi_modal_data": {"image": input_image},
         },
-        OmniDiffusionSamplingParams(
-            generator=generator,
-            true_cfg_scale=args.cfg_scale,
-            guidance_scale=args.guidance_scale,
-            guidance_scale_2=args.guidance_scale_2,
-            num_inference_steps=args.num_inference_steps,
-            num_outputs_per_prompt=args.num_outputs_per_prompt,
-            layers=args.layers,
-            resolution=args.resolution,
-        ),
+        sampling_params,
     )
     generation_end = time.perf_counter()
     generation_time = generation_end - generation_start

examples/offline_inference/step_audio2/README.md

@@ -0,0 +1,375 @@
+# Step-Audio2 Offline Inference Examples
+
+This directory contains examples for running offline inference with Step-Audio2 using vLLM-Omni.
+
+## Model Overview
+
+Step-Audio2 is a two-stage audio model:
+
+- **Stage 0 (Thinker)**: Audio understanding → Text + Audio tokens
+  - Input: Audio (16kHz)
+  - Output: Text transcription + Audio tokens for synthesis
+
+- **Stage 1 (Token2Wav)**: Audio synthesis
+  - Input: Audio tokens + Speaker prompt wav
+  - Output: Synthesized audio waveform (24kHz)
+
+## Hardware Requirements
+
+| Mode | GPU Configuration | VRAM Required |
+|------|-------------------|---------------|
+| ASR (S2T) | 1x GPU | ~20-25GB |
+| TTS/S2ST (single GPU) | 1x GPU | ~40-50GB |
+| TTS/S2ST (multi GPU) | 2x GPU | GPU0: ~28GB, GPU1: ~22GB |
+
+**Tested on:**
+- 1x NVIDIA H100 80GB (single-card S2ST)
+- 2x NVIDIA A10 40GB (multi-card S2ST)
+
+**Notes:**
+- Single GPU mode requires high VRAM due to both stages sharing memory
+- Multi GPU mode separates Stage 0 (Thinker) and Stage 1 (Token2Wav) across GPUs
+- VRAM usage can be adjusted via `gpu_memory_utilization` in stage config
+
+## Performance Benchmark
+
+### vLLM-Omni vs Official Step-Audio2
+
+Single request latency comparison between vLLM-Omni and official Step-Audio2 implementation.
+
+| Task | Tokens | vllm-omni | Step-Audio2 | Speedup |
+|------|--------|-----------|-------------|---------|
+| S2ST | ~85 | 5.45s | 7.36s | **1.35x** |
+| S2ST | ~160 | 6.67s | 13.92s | **2.09x** |
+| S2ST | ~315 | 9.42s | 31.50s | **3.34x** |
+| TTS | ~1024 | 16.65s | ~87s | **~5.2x** |
+
+**Key observations:**
+- Speedup increases with sequence length due to vLLM's efficient KV cache management
+- TTS (pure generation) shows the largest speedup (~5x)
+- S2ST benefits from optimized multi-stage pipeline
+
+**Benchmark environment:**
+- GPU: NVIDIA H100 80GB (single card)
+- Model: Step-Audio2-mini
+- Warmup: 1 run, Measured: 3 runs (averaged)
+
+### Async Chunk Streaming Performance
+
+Comparison between sequential (non-async) and async chunk modes via `/v1/audio/speech` TTS endpoint.
+
+| Mode | Mean TTFP | Mean E2E | Mean RTF | Audio Throughput |
+|------|-----------|----------|----------|-----------------|
+| Sequential | 4316ms | 4316ms | 0.938 | 1.07x realtime |
+| **Async Chunk** | **1437ms** | 4362ms | 0.949 | 1.06x realtime |
+| **Improvement** | **-67% (3x faster)** | ~same | ~same | ~same |
+
+**Key observations:**
+- Async chunk reduces **time-to-first-audio (TTFP) by 67%** (4.3s → 1.4s)
+- E2E latency remains comparable — async chunk overlaps Thinker decode with Token2Wav synthesis
+- RTF < 1 in both modes (real-time capable)
+- Sequential mode: TTFP ≈ E2E (must wait for all audio tokens before synthesis starts)
+- Async chunk mode: Token2Wav starts after first 28 tokens (chunk_size=25 + lookahead=3)
+
+**Benchmark environment:**
+- GPU: 4x NVIDIA RTX 3090 24GB (TP=2 for Thinker, 1 GPU for Token2Wav)
+- Model: Step-Audio2-mini
+- Endpoint: `/v1/audio/speech` (10 prompts, concurrency=1)
+- Measured via `bench_tts_serve.py`
+
+## Installation
+
+Make sure you have installed vLLM-Omni and all required dependencies:
+
+```bash
+# Install vLLM-Omni
+pip install vllm-omni
+
+# Install Step-Audio2 (REQUIRED for Token2Wav stage)
+pip install step-audio2
+```
+
+## Model Setup
+
+### Option 1: Auto-download from HuggingFace (Recommended)
+
+The script will **automatically download** the model on first run:
+
+```bash
+# Just run without specifying --model, it will auto-download stepfun-ai/Step-Audio2-mini
+python end2end.py --query-type audio_to_text
+
+# Or explicitly specify the HuggingFace model
+python end2end.py --query-type audio_to_text --model stepfun-ai/Step-Audio2-mini
+```
+
+Models will be cached in `~/.cache/huggingface/hub/` for future use.
+
+**Available models**:
+- `stepfun-ai/Step-Audio2-mini` (smaller, faster)
+- `stepfun-ai/Step-Audio2-7B` (larger, better quality)
+
+### Option 2: Manual Download (for offline use)
+
+Download and use locally:
+
+```bash
+# Download from HuggingFace
+huggingface-cli download stepfun-ai/Step-Audio2-mini --local-dir ./models/Step-Audio2-mini
+
+# Then use the local path
+python end2end.py --query-type audio_to_text --model ./models/Step-Audio2-mini
+```
+
+Ensure the model directory contains:
+```
+Step-Audio2-mini/
+├── config.json
+├── model.safetensors (or pytorch_model.bin)
+├── tokenizer.json
+├── tokenizer_config.json
+└── token2wav/                           # Token2Wav models (REQUIRED)
+    ├── speech_tokenizer_v2_25hz.onnx   # Audio tokenizer
+    ├── campplus.onnx                    # Speaker encoder
+    ├── flow.yaml                        # Flow model config
+    ├── flow.pt                          # Flow model weights
+    └── hift.pt                          # HiFT vocoder weights
+```
+
+## Usage Examples
+
+### 1. Audio to Text (ASR - Speech Recognition)
+
+Transcribe audio to text:
+
+```bash
+# Quick start - Using default model and test audio
+python end2end.py --query-type audio_to_text
+
+# Using your own audio file (model will auto-download)
+python end2end.py --query-type audio_to_text \
+    --audio-path /path/to/input.wav
+
+# With specific model
+python end2end.py --query-type audio_to_text \
+    --audio-path /path/to/input.wav \
+    --model stepfun-ai/Step-Audio2-7B
+
+# With custom question
+python end2end.py --query-type audio_to_text \
+    --audio-path input.wav \
+    --question "What is the speaker saying?"
+```
+
+**Output**: Text transcription saved to `output_step_audio2/00000_text.txt`
+
+### 2. Text to Audio (TTS - Speech Synthesis)
+
+Convert text to speech:
+
+```bash
+# Basic TTS (model auto-downloads)
+python end2end.py --query-type text_to_audio \
+    --text "Hello, this is a test of Step Audio 2 synthesis."
+
+# With specific model
+python end2end.py --query-type text_to_audio \
+    --text "Hello, this is a test." \
+    --model stepfun-ai/Step-Audio2-7B
+```
+
+**Note**: Speaker voice is controlled by the `STEP_AUDIO2_DEFAULT_PROMPT_WAV` environment variable or the default prompt wav bundled with the model.
+
+**Output**:
+- Text: `output_step_audio2/00000_text.txt`
+- Audio: `output_step_audio2/00000_output.wav` (24kHz)
+
+### 3. Audio to Audio (Voice Conversion)
+
+Process input audio and generate output audio:
+
+```bash
+# Basic voice conversion (model auto-downloads)
+python end2end.py --query-type audio_to_audio \
+    --audio-path /path/to/source_audio.wav
+
+# With specific model
+python end2end.py --query-type audio_to_audio \
+    --audio-path source.wav \
+    --model stepfun-ai/Step-Audio2-7B
+```
+
+This mode:
+1. Understands the content in `--audio-path` (source)
+2. Generates audio output with the default voice
+
+**Note**: To use a custom speaker voice, set the `STEP_AUDIO2_DEFAULT_PROMPT_WAV` environment variable.
+
+### Advanced Options
+
+```bash
+# Use custom stage configuration
+python end2end.py --query-type audio_to_text \
+    --stage-configs-path /path/to/custom_config.yaml
+
+# Multiple prompts (for batch testing)
+python end2end.py --query-type audio_to_text \
+    --audio-path input.wav \
+    --num-prompts 5
+
+# Custom output directory
+python end2end.py --query-type text_to_audio \
+    --text "Test synthesis" \
+    --output-dir ./my_outputs
+
+# Enable detailed logging
+python end2end.py --query-type audio_to_text \
+    --audio-path input.wav \
+    --enable-stats
+
+# Adjust generation parameters
+python end2end.py --query-type audio_to_text \
+    --audio-path input.wav \
+    --max-tokens 2048
+
+# Use custom speaker voice via environment variable
+STEP_AUDIO2_DEFAULT_PROMPT_WAV=/path/to/speaker.wav python end2end.py \
+    --query-type text_to_audio \
+    --text "Hello world"
+
+# Use Ray backend for distributed processing
+python end2end.py --query-type text_to_audio \
+    --text "Hello world" \
+    --worker-backend ray \
+    --ray-address "auto"
+```
+
+## Configuration
+
+### Stage Configuration
+
+The default configuration (`step_audio_2.yaml`) uses:
+
+- **Stage 0 (Thinker)**: GPU 0, 80% memory
+- **Stage 1 (Token2Wav)**: GPU 1, 30% memory
+
+For **single GPU** setup, edit the config to use `devices: "0"` for both stages.
+
+### Sampling Parameters
+
+- **Thinker (Stage 0)**:
+  - Temperature: 0.7 (balanced creativity)
+  - Top-p: 0.9
+  - Max tokens: 1024 (configurable)
+
+- **Token2Wav (Stage 1)**:
+  - Temperature: 0.0 (deterministic)
+  - Operates in generation mode (not sampling)
+
+## Common Issues
+
+### 1. ImportError: No module named 's3tokenizer'
+
+**Solution**: Install Step-Audio2 package:
+```bash
+pip install step-audio2
+```
+
+### 2. FileNotFoundError: prompt_wav file not found
+
+**Solution**: Set the `STEP_AUDIO2_DEFAULT_PROMPT_WAV` environment variable to a valid audio file:
+```bash
+export STEP_AUDIO2_DEFAULT_PROMPT_WAV=/path/to/speaker.wav
+python end2end.py --query-type text_to_audio --text "Hello"
+```
+Or ensure the default prompt wav (`default_female.wav`) exists in your model directory.
+
+### 3. FileNotFoundError: token2wav models not found
+
+**Solution**: Ensure your model directory has the complete `token2wav/` subdirectory with all ONNX and PyTorch models.
+
+### 4. CUDA Out of Memory
+
+**Solutions**:
+- Use single GPU mode (set both stages to `devices: "0"`)
+- Reduce `gpu_memory_utilization` in config
+- Reduce `max_num_batched_tokens`
+- Process fewer prompts at once
+
+### 5. Model not found in registry
+
+**Solution**: Ensure you're using vLLM-Omni's entry point with `--omni` flag or install vllm-omni properly:
+```bash
+pip install vllm-omni
+```
+
+## Output Files
+
+The script generates files in the output directory (default: `output_step_audio2/`):
+
+```
+output_step_audio2/
+├── 00000_text.txt        # Text output from Thinker stage
+├── 00000_output.wav      # Audio output from Token2Wav stage (24kHz)
+├── 00001_text.txt        # (if multiple prompts)
+└── 00001_output.wav
+```
+
+## Performance Tips
+
+1. **First run is slow**: Stage initialization takes 20-60 seconds
+2. **Single GPU**: Set both stages to `devices: "0"` in config
+3. **Multiple prompts**: Use `--num-prompts N` for batch testing
+4. **Ray backend**: For multi-node or advanced scheduling
+5. **Logging**: Use `--enable-stats` to debug performance issues
+
+## Speaker Voice Configuration
+
+The Token2Wav stage requires a speaker prompt wav for voice conditioning. It is automatically resolved in this order:
+
+1. `STEP_AUDIO2_DEFAULT_PROMPT_WAV` environment variable (if set)
+2. `{model_dir}/assets/default_female.wav`
+3. `{model_dir}/default_female.wav`
+
+If none are found, set the environment variable explicitly:
+```bash
+export STEP_AUDIO2_DEFAULT_PROMPT_WAV=/path/to/speaker.wav
+```
+
+**Guidelines for custom speaker prompt:**
+
+- **Duration**: 3-10 seconds recommended
+- **Quality**: Clean audio, minimal background noise
+- **Format**: WAV, MP3, FLAC (will be resampled internally)
+- **Content**: Clear speech, representative of target voice
+
+## Example Workflow
+
+Complete example from audio to final output:
+
+```bash
+# 1. ASR: Transcribe audio
+python end2end.py --query-type audio_to_text \
+    --audio-path interview.wav \
+    --model ./models/Step-Audio2-7B \
+    --output-dir ./outputs
+
+# 2. Check the transcription
+cat ./outputs/00000_text.txt
+
+# 3. TTS: Synthesize new speech (with custom voice)
+STEP_AUDIO2_DEFAULT_PROMPT_WAV=./speaker_samples/female_voice.wav \
+python end2end.py --query-type text_to_audio \
+    --text "The quick brown fox jumps over the lazy dog" \
+    --model ./models/Step-Audio2-7B \
+    --output-dir ./outputs
+
+# 4. Listen to the result
+# Audio saved to: ./outputs/00000_output.wav
+```
+
+## References
+
+- [Step-Audio2 Paper](https://arxiv.org/abs/2507.16632)
+- [vLLM-Omni Documentation](https://vllm-omni.readthedocs.io)
+- [Model on HuggingFace](https://huggingface.co/stepfun-ai/Step-Audio-2-mini)

examples/offline_inference/step_audio2/end2end.py

@@ -0,0 +1,527 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""
+This example shows how to use vLLM-Omni for running offline inference
+with Step-Audio2 model for audio understanding and generation.
+
+Step-Audio2 is a two-stage model:
+  - Stage 0 (Thinker): Audio understanding → Text + Audio tokens
+  - Stage 1 (Token2Wav): Audio tokens → Waveform (24kHz)
+
+Usage examples:
+  # Audio to text (ASR)
+  python end2end.py --query-type audio_to_text --audio-path input.wav --model-path "$MODEL_PATH"
+
+  # Text to audio (TTS with default voice)
+  python end2end.py --query-type text_to_audio --model-path "$MODEL_PATH"
+
+  # Audio to audio (Voice conversion with default voice)
+  python end2end.py --query-type audio_to_audio --audio-path input.wav --model-path "$MODEL_PATH"
+"""
+
+import os
+import sys
+from pathlib import Path
+from typing import NamedTuple
+
+import librosa  # noqa: TID251
+import numpy as np
+import soundfile as sf
+import torch
+from vllm.assets.audio import AudioAsset
+from vllm.sampling_params import SamplingParams
+from vllm.utils.argparse_utils import FlexibleArgumentParser
+
+from vllm_omni.entrypoints.omni import Omni
+
+SEED = 42
+
+
+class QueryResult(NamedTuple):
+    inputs: dict
+    limit_mm_per_prompt: dict[str, int]
+
+
+def get_audio_to_text_query(
+    audio_path: str | None = None,
+    question: str | None = None,
+    sampling_rate: int = 16000,
+) -> QueryResult:
+    """
+    Audio Speech Recognition (ASR) - Convert audio to text.
+
+    Args:
+        audio_path: Path to audio file
+        question: Question about the audio (optional)
+        sampling_rate: Target sampling rate (16kHz for Step-Audio2)
+
+    Returns:
+        QueryResult with prompt and audio data
+    """
+    if question is None:
+        question = "请记录下你所听到的语音内容。"
+
+    prompt = (
+        f"<|im_start|>system\n{question}<|im_end|>\n<|im_start|>user\n<audio_patch><|im_end|>\n<|im_start|>assistant\n"
+    )
+
+    # Load audio
+    if audio_path:
+        if not os.path.exists(audio_path):
+            raise FileNotFoundError(f"Audio file not found: {audio_path}")
+        audio_signal, sr = librosa.load(audio_path, sr=sampling_rate)
+        audio_data = (audio_signal.astype(np.float32), sr)
+    else:
+        # Use default test audio
+        audio_data = AudioAsset("mary_had_lamb").audio_and_sample_rate
+        # Resample to 16kHz if needed
+        if audio_data[1] != sampling_rate:
+            audio_signal = librosa.resample(audio_data[0], orig_sr=audio_data[1], target_sr=sampling_rate)
+            audio_data = (audio_signal.astype(np.float32), sampling_rate)
+
+    return QueryResult(
+        inputs={
+            "prompt": prompt,
+            "multi_modal_data": {
+                "audio": audio_data,
+            },
+        },
+        limit_mm_per_prompt={"audio": 1},
+    )
+
+
+def get_text_to_audio_query(
+    text: str | None = None,
+    sampling_rate: int = 16000,
+) -> QueryResult:
+    """
+    Text-to-Speech (TTS) - Convert text to audio.
+
+    Args:
+        text: Text to synthesize
+        sampling_rate: Target sampling rate (16kHz for Step-Audio2)
+
+    Returns:
+        QueryResult with text prompt and speaker wav path
+
+    Note:
+        For audio generation, the prompt must end with "<tts_start>"
+        to signal the model to generate audio tokens instead of just text.
+    """
+    if text is None:
+        text = "Hello, this is a test of Step Audio 2 text to speech synthesis."
+
+    prompt = (
+        "<|im_start|>system\n"
+        "请逐字朗读用户提供的文本，不要改写或补充。<|im_end|>\n"
+        "<|im_start|>user\n"
+        f"{text}<|im_end|>\n"
+        "<|im_start|>assistant\n"
+        "<tts_start>"
+    )
+
+    inputs = {
+        "prompt": prompt,
+    }
+
+    return QueryResult(
+        inputs=inputs,
+        limit_mm_per_prompt={},
+    )
+
+
+def get_audio_to_audio_query(
+    audio_path: str | None = None,
+    question: str | None = None,
+    sampling_rate: int = 16000,
+) -> QueryResult:
+    """
+    Audio-to-Audio - Voice conversion/cloning.
+
+    Args:
+        audio_path: Path to source audio file
+        question: Question/instruction about the audio
+        sampling_rate: Target sampling rate (16kHz for Step-Audio2)
+
+    Returns:
+        QueryResult with audio input
+
+    Note:
+        This mode processes input audio and generates output audio.
+        Speaker voice is controlled by STEP_AUDIO2_DEFAULT_PROMPT_WAV env var.
+    """
+    if question is None:
+        question = "请仔细聆听这段语音，然后复述其内容。"
+
+    prompt = (
+        "<|im_start|>system\n"
+        f"{question}<|im_end|>\n"
+        "<|im_start|>user\n"
+        "<audio_patch><|im_end|>\n"
+        "<|im_start|>assistant\n"
+        "<tts_start>"
+    )
+
+    # Load source audio
+    if audio_path:
+        if not os.path.exists(audio_path):
+            raise FileNotFoundError(f"Audio file not found: {audio_path}")
+        audio_signal, sr = librosa.load(audio_path, sr=sampling_rate)
+        audio_data = (audio_signal.astype(np.float32), sr)
+
+    else:
+        audio_data = AudioAsset("mary_had_lamb").audio_and_sample_rate
+        if audio_data[1] != sampling_rate:
+            audio_signal = librosa.resample(audio_data[0], orig_sr=audio_data[1], target_sr=sampling_rate)
+            audio_data = (audio_signal.astype(np.float32), sampling_rate)
+
+    return QueryResult(
+        inputs={
+            "prompt": prompt,
+            "multi_modal_data": {
+                "audio": audio_data,
+            },
+        },
+        limit_mm_per_prompt={"audio": 1},
+    )
+
+
+query_map = {
+    "audio_to_text": get_audio_to_text_query,
+    "text_to_audio": get_text_to_audio_query,
+    "audio_to_audio": get_audio_to_audio_query,
+}
+
+
+def main(args):
+    # Use local model path or HuggingFace model name
+    if not args.model:
+        sys.exit(1)
+
+    model_name = args.model
+
+    # Check if it's a local path
+    if os.path.isdir(model_name):
+        # Verify essential files exist
+        config_path = os.path.join(model_name, "config.json")
+        if not os.path.exists(config_path):
+            sys.exit(1)
+
+    # Get query configuration
+    query_func = query_map[args.query_type]
+
+    # Build query based on type
+    if args.query_type == "audio_to_text":
+        query_result = query_func(
+            audio_path=args.audio_path,
+            question=args.question,
+            sampling_rate=args.sampling_rate,
+        )
+    elif args.query_type == "text_to_audio":
+        query_result = query_func(
+            text=args.text,
+            sampling_rate=args.sampling_rate,
+        )
+    elif args.query_type == "audio_to_audio":
+        query_result = query_func(
+            audio_path=args.audio_path,
+            question=args.question,
+            sampling_rate=args.sampling_rate,
+        )
+    else:
+        raise ValueError(f"Unknown query type: {args.query_type}")
+
+    # Initialize vLLM-Omni with Step-Audio2
+    # Resolve stage config path
+    if args.stage_configs_path:
+        stage_config_path = args.stage_configs_path
+    else:
+        configs_dir = Path(__file__).parent.parent.parent.parent / "vllm_omni/model_executor/stage_configs"
+        if args.query_type == "audio_to_text":
+            # ASR only needs Thinker (Stage 0), no Token2Wav
+            stage_config_path = str(configs_dir / "step_audio_2_asr.yaml")
+            if not os.path.exists(stage_config_path):
+                stage_config_path = str(configs_dir / "step_audio_2.yaml")
+        else:
+            # TTS/S2ST need both stages
+            stage_config_path = str(configs_dir / "step_audio_2.yaml")
+
+    omni_llm = Omni(
+        model=model_name,
+        log_stats=args.enable_stats,
+        log_file=("step_audio2_pipeline.log" if args.enable_stats else None),
+        init_sleep_seconds=args.init_sleep_seconds,
+        batch_timeout=args.batch_timeout,
+        init_timeout=args.init_timeout,
+        shm_threshold_bytes=args.shm_threshold_bytes,
+        stage_configs_path=stage_config_path,
+        worker_backend=args.worker_backend,
+        ray_address=args.ray_address,
+    )
+
+    # Configure sampling parameters for each stage
+    # Stage 0 (Thinker): Generate text and audio tokens
+    # Note: For audio generation tasks, we need to allow the model to generate audio tokens
+    # Audio tokens are in range [151696, 158257], so we should not stop too early
+    # Qwen EOS token is typically 151645 (<|im_end|>), but we want to allow audio generation
+    # So we don't set stop_token_ids for audio tasks, but we do set it for text-only tasks
+    stop_token_ids = None
+    if args.query_type == "audio_to_text":
+        # For ASR, stop at EOS token
+        stop_token_ids = [151645]  # Qwen EOS token <|im_end|>
+
+    # Adjust sampling parameters based on query type
+    # For audio generation tasks, we need higher repetition_penalty to avoid text loops
+    # and allow the model to transition to audio token generation
+    repetition_penalty = 1.1 if args.query_type in ["text_to_audio", "audio_to_audio"] else 1.05
+
+    thinker_sampling_params = SamplingParams(
+        temperature=0.7,
+        top_p=0.9,
+        top_k=-1,
+        max_tokens=args.max_tokens,
+        seed=SEED,
+        detokenize=True,
+        repetition_penalty=repetition_penalty,
+        stop_token_ids=stop_token_ids,
+    )
+
+    # Stage 1 (Token2Wav): Convert audio tokens to waveform
+    # Note: This stage doesn't actually sample, it's deterministic generation
+    token2wav_sampling_params = SamplingParams(
+        temperature=0.0,
+        top_p=1.0,
+        top_k=-1,
+        max_tokens=1,
+        seed=SEED,
+        detokenize=False,
+    )
+
+    if args.query_type == "audio_to_text":
+        # ASR: single stage, only Thinker
+        sampling_params_list = [thinker_sampling_params]
+    else:
+        # TTS/S2ST: two stages
+        sampling_params_list = [thinker_sampling_params, token2wav_sampling_params]
+
+    # Prepare prompts
+    if args.num_prompts > 1:
+        prompts = [query_result.inputs for _ in range(args.num_prompts)]
+    else:
+        prompts = [query_result.inputs]
+
+    # Run inference
+    omni_outputs = omni_llm.generate(prompts, sampling_params_list)
+
+    # Create output directory
+    output_dir = args.output_dir
+    os.makedirs(output_dir, exist_ok=True)
+
+    def _get_request_index(req_id: str, fallback: int) -> int:
+        """Parse request index from vLLM request_id.
+
+        vLLM may return IDs like "0_<uuid>". Use the leading integer if present,
+        otherwise fall back to the loop index.
+        """
+        try:
+            return int(req_id)
+        except (TypeError, ValueError):
+            if isinstance(req_id, str):
+                head = req_id.split("_", 1)[0]
+                if head.isdigit():
+                    return int(head)
+        return fallback
+
+    # Process outputs from each stage
+    for stage_idx, stage_outputs in enumerate(omni_outputs):
+        # request_output may be a single object or a list
+        raw_output = stage_outputs.request_output
+        outputs_list = raw_output if isinstance(raw_output, list) else [raw_output]
+
+        if stage_outputs.final_output_type == "text":
+            # Stage 0 (Thinker) text output
+            for i, output in enumerate(outputs_list):
+                request_id = _get_request_index(output.request_id, i)
+                text_output = output.outputs[0].text
+
+                # Save to file
+                out_txt = os.path.join(output_dir, f"{request_id:05d}_text.txt")
+                with open(out_txt, "w", encoding="utf-8") as f:
+                    f.write(f"Prompt:\n{prompts[request_id]['prompt']}\n\n")
+                    f.write(f"Output:\n{text_output}\n")
+                print(f"Text output saved to: {out_txt}")
+                print(f"Text: {text_output[:200]}")
+
+        elif stage_outputs.final_output_type == "audio":
+            # Stage 1 (Token2Wav) audio output
+            for i, output in enumerate(outputs_list):
+                request_id = _get_request_index(output.request_id, i)
+
+                # Get audio tensor (24kHz)
+                mm_out = {}
+                if hasattr(output, "multimodal_output") and output.multimodal_output:
+                    mm_out = output.multimodal_output
+                elif output.outputs and hasattr(output.outputs[0], "multimodal_output"):
+                    mm_out = output.outputs[0].multimodal_output or {}
+                audio_tensor = None
+                for key in ("audio", "wav", "waveform", "audio_pcm", "pcm", "model_outputs"):
+                    if key in mm_out and mm_out[key] is not None:
+                        audio_tensor = mm_out[key]
+                        break
+
+                if audio_tensor is not None:
+                    # Save audio file
+                    output_wav = os.path.join(output_dir, f"{request_id:05d}_output.wav")
+                    if isinstance(audio_tensor, list):
+                        chunks = []
+                        for chunk in audio_tensor:
+                            if chunk is None:
+                                continue
+                            if isinstance(chunk, torch.Tensor):
+                                chunk_np = chunk.detach().cpu().numpy()
+                            else:
+                                chunk_np = np.asarray(chunk)
+                            if chunk_np.size > 0:
+                                chunks.append(chunk_np)
+                        if not chunks:
+                            continue
+                        audio_numpy = np.concatenate(chunks, axis=-1)
+                    elif isinstance(audio_tensor, torch.Tensor):
+                        audio_numpy = audio_tensor.detach().cpu().numpy()
+                    else:
+                        audio_numpy = np.asarray(audio_tensor)
+                    sf.write(output_wav, audio_numpy, samplerate=24000)
+
+
+def parse_args():
+    parser = FlexibleArgumentParser(description="Demo on using vLLM-Omni for offline inference with Step-Audio2")
+
+    # Model and config
+    parser.add_argument(
+        "--model",
+        "-m",
+        type=str,
+        required=True,
+        help=(
+            "Model path (required). Can be:\n"
+            "  - Local path: /path/to/step-audio-2\n"
+            "  - HuggingFace ID: stepfun-ai/Step-Audio2-mini"
+        ),
+    )
+    parser.add_argument(
+        "--stage-configs-path",
+        type=str,
+        default=None,
+        help="Path to stage config YAML file (default: use built-in step_audio_2.yaml)",
+    )
+
+    # Query configuration
+    parser.add_argument(
+        "--query-type",
+        "-q",
+        type=str,
+        default="audio_to_text",
+        choices=query_map.keys(),
+        help="Query type: audio_to_text (ASR), text_to_audio (TTS), audio_to_audio (voice conversion)",
+    )
+    parser.add_argument(
+        "--audio-path",
+        "-a",
+        type=str,
+        default=None,
+        help="Path to input audio file (for audio_to_text and audio_to_audio)",
+    )
+    parser.add_argument(
+        "--text",
+        "-t",
+        type=str,
+        default=None,
+        help="Text to synthesize (for text_to_audio mode)",
+    )
+    parser.add_argument(
+        "--question",
+        type=str,
+        default=None,
+        help="Question/instruction about the audio",
+    )
+    parser.add_argument(
+        "--sampling-rate",
+        type=int,
+        default=16000,
+        help="Audio sampling rate for input (default: 16000 Hz)",
+    )
+
+    # Generation parameters
+    parser.add_argument(
+        "--max-tokens",
+        type=int,
+        default=1024,
+        help="Maximum tokens to generate in Thinker stage (default: 1024)",
+    )
+    parser.add_argument(
+        "--num-prompts",
+        type=int,
+        default=1,
+        help="Number of times to repeat the prompt (for testing)",
+    )
+
+    # System configuration
+    parser.add_argument(
+        "--enable-stats",
+        action="store_true",
+        default=False,
+        help="Enable detailed statistics logging",
+    )
+    parser.add_argument(
+        "--init-sleep-seconds",
+        type=int,
+        default=20,
+        help="Sleep seconds after starting each stage (default: 20)",
+    )
+    parser.add_argument(
+        "--batch-timeout",
+        type=int,
+        default=5,
+        help="Batch timeout in seconds (default: 5)",
+    )
+    parser.add_argument(
+        "--init-timeout",
+        type=int,
+        default=300,
+        help="Initialization timeout in seconds (default: 300)",
+    )
+    parser.add_argument(
+        "--shm-threshold-bytes",
+        type=int,
+        default=65536,
+        help="Shared memory threshold in bytes (default: 65536)",
+    )
+    parser.add_argument(
+        "--worker-backend",
+        type=str,
+        default="multi_process",
+        choices=["multi_process", "ray"],
+        help="Worker backend (default: multi_process)",
+    )
+    parser.add_argument(
+        "--ray-address",
+        type=str,
+        default=None,
+        help="Ray cluster address (for ray backend)",
+    )
+
+    # Output
+    parser.add_argument(
+        "--output-dir",
+        "-o",
+        type=str,
+        default="output_step_audio2",
+        help="Output directory for generated files (default: output_step_audio2)",
+    )
+
+    return parser.parse_args()
+
+
+if __name__ == "__main__":
+    args = parse_args()
+    main(args)

examples/offline_inference/text_to_speech/README.md

@@ -17,6 +17,7 @@ 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 |
 | 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 |
 | MOSS-TTS-Nano | `OpenMOSS-Team/MOSS-TTS-Nano` | single (AR + codec) | ✓ (required) | ✓ | voice_clone, continuation | 48 kHz |
 | OmniVoice | `k2-fsa/OmniVoice` | 2 (gen + dec) | ✓ | — | voice design, language hint | 24 kHz |
@@ -159,6 +160,46 @@ Streaming requires `async_chunk: true` in the stage config.
 
 ---
 
+## Ming-omni-tts
+
+Dense 0.5B two-stage TTS pipeline (`AR + flow` + audio VAE) at 44.1 kHz. The example covers style, IP voice, music-only generation, text-to-audio events, emotion, dialect, zero-shot cloning, podcast, speech+BGM, and speech+environment-sound cases.
+
+### Quick start
+```bash
+python examples/offline_inference/text_to_speech/ming_tts/end2end.py \
+    --case style \
+    --deploy-config vllm_omni/deploy/ming_tts.yaml \
+    --enforce-eager
+```
+
+### Voice cloning
+```bash
+python examples/offline_inference/text_to_speech/ming_tts/end2end.py \
+    --case zero_shot \
+    --ref-audio /path/to/reference.wav \
+    --ref-text "在此奉劝大家别乱打美白针。" \
+    --deploy-config vllm_omni/deploy/ming_tts.yaml \
+    --enforce-eager
+```
+
+### Streaming
+```bash
+python examples/offline_inference/text_to_speech/ming_tts/end2end.py \
+    --case basic \
+    --ref-audio /path/to/reference.wav \
+    --streaming \
+    --deploy-config vllm_omni/deploy/ming_tts.yaml \
+    --enforce-eager
+```
+
+### Notes
+- `style`, `ip`, `bgm`, and `tta` do not require reference audio.
+- Reference-audio cases use `--ref-audio`; `zero_shot` also requires `--ref-text`.
+- `podcast` uses multiple references via `--ref-audio-paths`.
+- Full case details live in [`ming_tts/README.md`](ming_tts/README.md).
+
+---
+
 ## Ming-flash-omni-TTS
 
 Standalone talker-only deployment of Ming-flash-omni-2.0 at 44.1 kHz. Voice is controlled through caption fields (`风格` / `IP` / `语速`/`基频`/`音量`) rather than reference audio.

examples/offline_inference/text_to_speech/higgs_audio_v3/end2end.py

@@ -0,0 +1,204 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""Offline higgs-audio v3 inference example.
+
+Runs Stage 0 (Qwen3 talker) + Stage 1 (HiggsAudio codec) end-to-end
+through the vLLM-Omni engine without going through the HTTP server, and
+saves a 24 kHz mono WAV per prompt.
+
+Example (plain TTS):
+
+    python examples/offline_inference/text_to_speech/higgs_audio_v3/end2end.py \\
+        --texts "Hello world." \\
+        --output-dir results/higgs_v3_wavs
+
+Example (voice clone):
+
+    python examples/offline_inference/text_to_speech/higgs_audio_v3/end2end.py \\
+        --texts "Hello world." \\
+        --ref-audio path/to/reference.wav \\
+        --ref-text "Transcript of the reference clip." \\
+        --output-dir results/higgs_v3_wavs
+"""
+
+from __future__ import annotations
+
+import os
+
+os.environ.setdefault("VLLM_USE_DEEP_GEMM", "0")
+os.environ.setdefault("VLLM_MOE_USE_DEEP_GEMM", "0")
+
+import re
+import time
+from pathlib import Path
+
+import numpy as np
+import soundfile as sf
+import torch
+
+from vllm_omni import Omni
+
+SAMPLE_RATE = 24_000
+DEFAULT_TEXTS = (
+    "Hello world.",
+    "The quick brown fox jumps over the lazy dog.",
+    "Today is a beautiful day for a walk in the park.",
+)
+
+
+def parse_args():
+    import argparse
+
+    parser = argparse.ArgumentParser(description="Offline higgs-audio v3 inference")
+    parser.add_argument(
+        "--model",
+        type=str,
+        default="bosonai/higgs-audio-v3-tts-4b",
+        help="Stage-0 talker model id or path.",
+    )
+    parser.add_argument(
+        "--texts",
+        type=str,
+        nargs="+",
+        default=list(DEFAULT_TEXTS),
+        help="One or more plain-text prompts to synthesize.",
+    )
+    parser.add_argument(
+        "--output-dir",
+        type=str,
+        default="results/higgs_v3_wavs",
+        help="Directory to write per-prompt WAV files.",
+    )
+    parser.add_argument(
+        "--deploy-config",
+        type=str,
+        default=None,
+        help="Override the deploy config path.",
+    )
+    parser.add_argument(
+        "--max-new-tokens",
+        type=int,
+        default=2048,
+        help="Cap on Stage-0 codec frames per request.",
+    )
+    parser.add_argument(
+        "--ref-audio",
+        type=str,
+        default=None,
+        help="Reference audio for voice clone (WAV/FLAC/MP3 path).",
+    )
+    parser.add_argument(
+        "--ref-text",
+        type=str,
+        default=None,
+        help="Transcript of the reference audio. Optional but improves fidelity.",
+    )
+    return parser.parse_args()
+
+
+def _slugify(text: str) -> str:
+    slug = re.sub(r"\s+", "_", text.strip().lower())
+    slug = re.sub(r"[^a-z0-9_]+", "", slug)
+    return slug[:48] or "prompt"
+
+
+def _extract_pcm(multimodal_output: dict) -> torch.Tensor:
+    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")
+        return torch.cat(valid, dim=0) if len(valid) > 1 else valid[0]
+    return torch.as_tensor(audio).float().cpu().reshape(-1)
+
+
+def _pcm_to_int16(pcm: torch.Tensor) -> np.ndarray:
+    arr = pcm.numpy()
+    if arr.dtype.kind == "f":
+        arr = np.clip(arr, -1.0, 1.0)
+        arr = (arr * 32767.0).astype(np.int16)
+    else:
+        arr = arr.astype(np.int16)
+    return arr
+
+
+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, trust_remote_code=True)
+
+    from transformers import AutoTokenizer
+
+    from vllm_omni.model_executor.models.higgs_audio_v3.higgs_audio_v3_tokenizer import (
+        HiggsAudioV3TokenizerAdapter,
+        apply_delay_pattern,
+        encode_reference_audio,
+    )
+
+    tokenizer = AutoTokenizer.from_pretrained(args.model, trust_remote_code=True)
+    adapter = HiggsAudioV3TokenizerAdapter(tokenizer)
+
+    # Load and encode reference audio once (if voice cloning)
+    ref_codes_delayed: torch.Tensor | None = None
+    if args.ref_audio is not None:
+        ref_wav, ref_sr = sf.read(args.ref_audio, always_2d=False)
+        if ref_wav.ndim == 2:
+            ref_wav = ref_wav.mean(axis=1)
+        ref_codes_raw = encode_reference_audio(ref_wav, int(ref_sr))
+        ref_codes_delayed = apply_delay_pattern(ref_codes_raw)
+        print(f"Reference   : {args.ref_audio}")
+        print(f"Ref codes   : {ref_codes_raw.shape[0]} frames -> {ref_codes_delayed.shape[0]} delayed")
+        if args.ref_text:
+            print(f"Ref text    : {args.ref_text}")
+
+    print(f"Model       : {args.model}")
+    print(f"Prompts     : {len(args.texts)}")
+    print(f"Output dir  : {output_dir}")
+    print(f"Voice clone : {'yes' if ref_codes_delayed is not None else 'no'}")
+
+    total_elapsed = 0.0
+    total_dur = 0.0
+    for text in args.texts:
+        if ref_codes_delayed is not None:
+            prompt_ids = adapter.build_prompt(
+                text,
+                num_ref_tokens=int(ref_codes_delayed.shape[0]),
+                reference_text=args.ref_text,
+            )
+            prompt = {
+                "prompt_token_ids": prompt_ids,
+                "additional_information": {
+                    "audio_input_ids": ref_codes_delayed.to(torch.long),
+                    "audio_input_ids_mask": torch.ones(ref_codes_delayed.shape[0], dtype=torch.bool),
+                },
+            }
+        else:
+            prompt_ids = adapter.build_prompt(text)
+            prompt = {"prompt_token_ids": prompt_ids}
+
+        t_start = time.perf_counter()
+        outputs = engine.generate([prompt])
+        elapsed = time.perf_counter() - t_start
+        total_elapsed += elapsed
+
+        mm = outputs[0].outputs[0].multimodal_output
+        pcm = _extract_pcm(mm)
+        slug = _slugify(text)
+        out_path = output_dir / f"{slug}.wav"
+        sf.write(str(out_path), _pcm_to_int16(pcm), SAMPLE_RATE, format="WAV", subtype="PCM_16")
+        dur = pcm.numel() / SAMPLE_RATE
+        total_dur += dur
+        print(f"  {slug:<50} dur={dur:5.2f}s  -> {out_path}")
+
+    rtf = total_elapsed / total_dur if total_dur > 0 else float("inf")
+    print(f"Total infer : {total_elapsed:.2f}s  total audio: {total_dur:.2f}s  RTF: {rtf:.3f}")
+
+
+if __name__ == "__main__":
+    main()

examples/offline_inference/text_to_speech/ming_tts/README.md

@@ -0,0 +1,248 @@
+# Ming-omni-tts Offline Inference
+
+`end2end.py` runs Ming dense 0.5B end to end with vLLM-Omni. It uses the in-repo Ming prompt assembly helper directly, so the example request shape matches the real integration instead of a simplified wrapper.
+
+## Files
+
+| File | Purpose |
+|---|---|
+| `end2end.py` | Driver: CLI, case loading, prompt construction, orchestration (~150 lines) |
+| `cases.yaml` | All 11 built-in case definitions (prompt, text, instruction, ref-audio flags, flow controls) |
+| `runner.py` | Engine management and audio output (streaming + blocking paths) |
+
+## Model Overview
+
+Ming dense 0.5B is exposed here as a two-stage offline pipeline:
+
+- **Stage 0**: Qwen2-based AR generation with Ming prompt formatting and inline flow controls
+- **Stage 1**: audio VAE decode to mono 44.1 kHz waveform
+
+`config_ming_tts.py` adapts the checkpoint's HuggingFace config fields
+(LLM, DiT, aggregator, AudioVAE, token ids). `vllm_omni/deploy/ming_tts.yaml`
+selects the vLLM-Omni pipeline and stage runtime topology, including
+connectors, async chunking, memory limits, and sampling defaults.
+
+The example supports both:
+
+- **Blocking eager** via `vllm_omni/deploy/ming_tts.yaml`
+- **Async chunk eager** via `vllm_omni/deploy/ming_tts.yaml` (default `async_chunk: true`)
+
+## Setup
+
+Install vLLM-Omni with the platform requirements for your accelerator:
+
+```bash
+uv pip install -e .
+```
+
+The Ming offline example does not require a separate upstream Ming package.
+Reference-audio cases use the repo dependencies for audio loading,
+resampling, and CampPlus speaker extraction, including `soundfile`,
+`torchaudio`, and `onnxruntime-rocm` in the documented ROCm environment.
+
+The tested ROCm environment is summarized in the
+[Ming recipe](https://github.com/vllm-project/vllm-omni/blob/main/recipes/inclusionAI/Ming-omni-tts-0.5B.md).
+
+## Supported Cases
+
+These cases cover the upstream dense 0.5B cookbook surface that maps cleanly onto the current vLLM-Omni example:
+
+- `style`: zero-speaker style-conditioned speech
+- `ip`: zero-speaker IP voice generation
+- `bgm`: music-only generation
+- `tta`: text-to-audio event generation with FlowLoss controls
+- `emotion`: reference-audio speech with emotion control
+- `basic`: reference-audio speech with speed / pitch / volume control
+- `dialect`: reference-audio speech with dialect control
+- `zero_shot`: reference-audio cloning with explicit transcript
+- `podcast`: multi-reference dialogue generation with automatic speaker embedding extraction
+- `speech_bgm`: speech with background music conditioning
+- `speech_sound`: speech with environmental sound conditioning
+
+## Quick Start
+
+Run the zero-speaker style example:
+
+```bash
+python examples/offline_inference/text_to_speech/ming_tts/end2end.py \
+    --case style \
+    --deploy-config vllm_omni/deploy/ming_tts.yaml \
+    --enforce-eager
+```
+
+Run zero-shot cloning with a transcript:
+
+```bash
+python examples/offline_inference/text_to_speech/ming_tts/end2end.py \
+    --case zero_shot \
+    --ref-audio /path/to/10002287-00000094.wav \
+    --ref-text "在此奉劝大家别乱打美白针。" \
+    --deploy-config vllm_omni/deploy/ming_tts.yaml \
+    --enforce-eager
+```
+
+Run emotion-controlled speech:
+
+```bash
+python examples/offline_inference/text_to_speech/ming_tts/end2end.py \
+    --case emotion \
+    --ref-audio /path/to/emotion_prompt.wav \
+    --deploy-config vllm_omni/deploy/ming_tts.yaml \
+    --enforce-eager
+```
+
+Run podcast generation with two reference clips:
+
+```bash
+python examples/offline_inference/text_to_speech/ming_tts/end2end.py \
+    --case podcast \
+    --ref-audio-paths /path/to/CTS-CN-F2F-2019-11-11-423-012-A.wav /path/to/CTS-CN-F2F-2019-11-11-423-012-B.wav \
+    --deploy-config vllm_omni/deploy/ming_tts.yaml \
+    --enforce-eager
+```
+
+The script automatically extracts one 192-d speaker embedding per reference WAV using the Ming model's `campplus.onnx`.
+
+If you already have precomputed multi-speaker embeddings, you can override extraction with:
+
+```bash
+--speaker-embedding /path/to/podcast_speaker_embeddings.json
+```
+
+where the JSON is a list of speaker embeddings, one 192-d vector per speaker.
+
+Run text-to-audio event generation:
+
+```bash
+python examples/offline_inference/text_to_speech/ming_tts/end2end.py \
+    --case tta \
+    --deploy-config vllm_omni/deploy/ming_tts.yaml \
+    --enforce-eager
+```
+
+Use async_chunk streaming:
+
+```bash
+python examples/offline_inference/text_to_speech/ming_tts/end2end.py \
+    --case basic \
+    --ref-audio /path/to/10002287-00000095.wav \
+    --streaming \
+    --deploy-config vllm_omni/deploy/ming_tts.yaml \
+    --enforce-eager
+```
+
+`--streaming` uses `AsyncOmni` and the async_chunk stage config. It currently
+supports one prompt per process invocation; use blocking mode for
+`--num-prompts > 1`.
+
+Collect runtime stats and a manifest:
+
+```bash
+python examples/offline_inference/text_to_speech/ming_tts/end2end.py \
+    --case style \
+    --deploy-config vllm_omni/deploy/ming_tts.yaml \
+    --enforce-eager \
+    --enable-stats \
+    --stats-log-file output_audio/ming_style_pipeline.log \
+    --metadata-json output_audio/ming_style_manifest.json
+```
+
+## Reference Fixtures
+
+The upstream Ming cookbook uses these public audio fixtures from `inclusionAI/Ming-omni-tts/data/wavs`:
+
+- `10002287-00000094.wav` for zero-shot cloning
+- `10002287-00000095.wav` for `basic`
+- `emotion_prompt.wav` for `emotion`
+- `yue_prompt.wav` for `dialect`
+- `00000309-00000300.wav` for `speech_bgm` and `speech_sound`
+- `CTS-CN-F2F-2019-11-11-423-012-A.wav` and `CTS-CN-F2F-2019-11-11-423-012-B.wav` for `podcast`
+
+## Validation Matrix
+
+The repo-facing example is intended to cover the same dense TTS workflows used
+by the local Ming validation script:
+
+| Case | Blocking `deploy/ming_tts.yaml` | Async chunk `deploy/ming_tts.yaml` | Extra inputs |
+|---|---:|---:|---|
+| `style` | Yes | Optional smoke test | none |
+| `ip` | Yes | Optional smoke test | none |
+| `bgm` | Yes | Optional smoke test | none |
+| `tta` | Yes | Optional smoke test | none |
+| `emotion` | Yes | Yes | `--ref-audio emotion_prompt.wav` |
+| `basic` | Yes | Yes | `--ref-audio 10002287-00000095.wav` |
+| `dialect` | Yes | Yes | `--ref-audio yue_prompt.wav` |
+| `zero_shot` | Yes | Yes | `--ref-audio 10002287-00000094.wav --ref-text ...` |
+| `podcast` | Yes | Yes | two `--ref-audio-paths` |
+| `speech_bgm` | Yes | Yes | `--ref-audio 00000309-00000300.wav` |
+| `speech_sound` | Yes | Yes | `--ref-audio 00000309-00000300.wav` |
+
+## Validated Outputs
+
+The following measurements are retained from an earlier L4 CUDA validation;
+they are not ROCm benchmark results. Default async_chunk matched blocking
+output frame counts and Stage-1 patch counts for every case:
+
+| Case | Blocking frames / patches / sec | Async chunk frames / patches / sec |
+|---|---:|---:|
+| `style` | 409248 / 29 / 9.28 | 409248 / 29 / 9.28 |
+| `ip` | 183456 / 13 / 4.16 | 183456 / 13 / 4.16 |
+| `bgm` | 1326528 / 94 / 30.08 | 1326528 / 94 / 30.08 |
+| `tta` | 465696 / 33 / 10.56 | 465696 / 33 / 10.56 |
+| `emotion` | 324576 / 23 / 7.36 | 324576 / 23 / 7.36 |
+| `basic` | 211680 / 15 / 4.80 | 211680 / 15 / 4.80 |
+| `dialect` | 239904 / 17 / 5.44 | 239904 / 17 / 5.44 |
+| `zero_shot` | 409248 / 29 / 9.28 | 409248 / 29 / 9.28 |
+| `podcast` | 437472 / 31 / 9.92 | 437472 / 31 / 9.92 |
+| `speech_bgm` | 296352 / 21 / 6.72 | 296352 / 21 / 6.72 |
+| `speech_sound` | 352800 / 25 / 8.00 | 352800 / 25 / 8.00 |
+
+## Key Arguments
+
+| Argument | Description |
+|---|---|
+| `--model` | Hugging Face repo or local Ming checkpoint path |
+| `--deploy-config` | Deploy config YAML. Use `vllm_omni/deploy/ming_tts.yaml` |
+| `--case` | Built-in demo case |
+| `--ref-audio` | Single reference wav path for cloning-style cases |
+| `--ref-audio-paths` | Multiple reference wav paths, used by `podcast` |
+| `--ref-text` | Reference transcript. Required for `zero_shot` |
+| `--instructions` | Free-form Ming instruction string |
+| `--instruction-json` | Structured Ming instruction JSON |
+| `--speaker-embedding` | JSON file containing a 192-d speaker embedding |
+| `--extract-speaker-embeddings` | Force CampPlus speaker extraction from the provided reference audio paths |
+| `--max-decode-steps` | Override `ming_max_decode_steps` |
+| `--num-prompts` | Repeat the same case N times. Outputs are indexed when `N > 1` |
+| `--streaming` | Use `AsyncOmni` and async_chunk transport |
+| `--enforce-eager` | Recommended for Ming dense; non-eager is out of scope |
+| `--enable-stats` / `--log-stats` | Enable vLLM-Omni per-request stats logging |
+| `--stats-log-file` | Optional path for the stats log |
+| `--metadata-json` | Optional path for the run manifest JSON |
+| `--stage-init-timeout` | Per-stage initialization timeout in seconds |
+| `--init-timeout` | Total initialization timeout in seconds |
+| `--batch-timeout` | Batch timeout in seconds |
+| `--worker-backend` | `multi_process` or `ray` |
+| `--ray-address` | Ray cluster address when using `--worker-backend ray` |
+
+## Output
+
+- The script writes one mono 44.1 kHz WAV file per run
+- Default output directory: `output_audio/`
+- Default filename: `ming_<case>.wav`
+- When `--num-prompts > 1`, outputs are indexed as `ming_<case>_00000.wav`, `..._00001.wav`, etc.
+- When stats are enabled, the script can also write:
+  - a stats log file such as `ming_style_pipeline.log`
+  - a manifest JSON with per-output metadata, stage durations, peak memory info,
+    and streaming client latency metrics when `--streaming` is used
+
+## Notes
+
+- `style` and `ip` are zero-speaker paths and do not require a reference clip
+- `emotion`, `basic`, `dialect`, `speech_bgm`, and `speech_sound` require one reference clip
+- `zero_shot` requires both `--ref-audio` and `--ref-text`
+- `podcast` requires at least two reference clips via `--ref-audio-paths`
+- `podcast` automatically extracts one speaker embedding per reference clip
+- `--speaker-embedding` may contain either one 192-d vector or a list of 192-d vectors
+- `--enforce-eager` was used for the validated runs
+- The earlier L4 validation used SDPA for the Ming audio VAE instead of
+  FlashAttention2, which is the preferred default when available.

examples/offline_inference/text_to_speech/ming_tts/cases.yaml

@@ -0,0 +1,110 @@
+style:
+  prompt: "Please generate speech based on the following description.\n"
+  text: "我会一直在这里陪着你，直到你慢慢、慢慢地沉入那个最温柔的梦里……好吗？"
+  instruction:
+    风格: >-
+      这是一种ASMR耳语，属于一种旨在引发特殊感官体验的创意风格。这个女性使用轻柔的普通话进行耳语，声音气音成分重。音量极低，紧贴麦克风，语速极慢，旨在制造触发听者颅内快感的声学刺激。
+  use_zero_spk_emb: true
+  max_decode_steps: 200
+
+ip:
+  prompt: "Please generate speech based on the following description.\n"
+  text: "这款产品的名字，叫变态坑爹牛肉丸。"
+  instruction:
+    IP: "灵小甄"
+  use_zero_spk_emb: true
+  max_decode_steps: 200
+
+bgm:
+  prompt: "Please generate music based on the following description.\n"
+  text: " Genre: 电子舞曲. Mood: 自信 / 坚定. Instrument: 架子鼓. Theme: 节日. Duration: 30s."
+  instruction: null
+  use_zero_spk_emb: false
+  max_decode_steps: 400
+
+tta:
+  prompt: "Please generate audio events based on given text.\n"
+  text: "Thunder and a gentle rain"
+  instruction: null
+  use_zero_spk_emb: false
+  max_decode_steps: 200
+  cfg: 4.5
+  sigma: 0.3
+  temperature: 2.5
+
+emotion:
+  prompt: "Please generate speech based on the following description.\n"
+  text: "我竟然抢到了陈奕迅的演唱会门票！太棒了！终于可以现场听一听他的歌声了！"
+  instruction:
+    情感: "高兴"
+  requires_ref_audio: true
+  auto_extract_speaker_embeddings: true
+  max_decode_steps: 200
+
+basic:
+  prompt: "Please generate speech based on the following description.\n"
+  text: "简单地说，这相当于惠普把消费领域市场拱手相让了。"
+  instruction:
+    语速: "快速"
+    基频: "中"
+    音量: "中"
+  requires_ref_audio: true
+  auto_extract_speaker_embeddings: true
+  max_decode_steps: 200
+
+dialect:
+  prompt: "Please generate speech based on the following description.\n"
+  text: "我觉得社会企业同个人都有责任"
+  instruction:
+    方言: "广粤话"
+  requires_ref_audio: true
+  auto_extract_speaker_embeddings: true
+  max_decode_steps: 200
+
+zero_shot:
+  prompt: "Please generate speech based on the following description.\n"
+  text: "我们的愿景是构建未来服务业的数字化基础设施，为世界带来更多微小而美好的改变。"
+  instruction: null
+  requires_ref_audio: true
+  requires_ref_text: true
+  auto_extract_speaker_embeddings: true
+  max_decode_steps: 200
+
+podcast:
+  prompt: "Please generate speech based on the following description.\n"
+  text: " speaker_1:你可以说一下，就大概说一下，可能虽然我也不知道，我看过那部电影没有。\n speaker_2:就是那个叫什么，变相一节课的嘛。\n speaker_1:嗯。\n speaker_2:一部搞笑的电影。\n speaker_1:一部搞笑的。\n"
+  instruction: null
+  prompt_text: " speaker_1:并且我们还要进行每个月还要考核 笔试的话还要进行笔试，做个，当服务员还要去笔试了\n speaker_2:对啊，这真的很奇怪，就是 单纯的因，单纯自己工资不高，只是因为可能人家那个店比较出名一点，就对你苛刻要求\n"
+  requires_ref_audio_count: 2
+  auto_extract_speaker_embeddings: true
+  max_decode_steps: 200
+
+speech_bgm:
+  prompt: "Please generate speech based on the following description.\n"
+  text: "此次业绩下滑原因，可归结为企业停止服务某些品牌，而带来的负面影响。"
+  instruction:
+    BGM:
+      Genre: "当代古典音乐."
+      Mood: "温暖 / 友善."
+      Instrument: "电吉他"
+      Theme: "节日."
+      SNR: 10.0
+      ENV: null
+  requires_ref_audio: true
+  auto_extract_speaker_embeddings: true
+  max_decode_steps: 200
+
+speech_sound:
+  prompt: "Please generate speech based on the following description.\n"
+  text: "此次业绩下滑原因，可归结为企业停止服务某些品牌，而带来的负面影响。"
+  instruction:
+    BGM:
+      ENV: "Birds chirping"
+      SNR: 10.0
+      Genre: null
+      Mood: null
+      Instrument: null
+      Theme: null
+  requires_ref_audio: true
+  auto_extract_speaker_embeddings: true
+  max_decode_steps: 200

examples/offline_inference/text_to_speech/ming_tts/end2end.py

@@ -0,0 +1,200 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""Offline inference demo for Ming-omni-tts via vLLM Omni."""
+
+import json
+import os
+from pathlib import Path
+
+import soundfile as sf
+import torch
+import torchaudio
+import yaml
+from transformers import AutoTokenizer
+from vllm.utils.argparse_utils import FlexibleArgumentParser
+
+from vllm_omni.model_executor.models.ming_tts.config_ming_tts import (
+    KEY_CFG,
+    KEY_MAX_DECODE_STEPS,
+    KEY_SIGMA,
+    KEY_TEMPERATURE,
+    SAMPLE_RATE,
+)
+from vllm_omni.model_executor.models.ming_tts.prompt_assembly import build_ming_dense_prompt
+from vllm_omni.model_executor.models.ming_tts.speaker_extractor import MingSpeakerEmbeddingExtractor
+
+try:
+    from .runner import (
+        build_manifest,
+        build_sampling_params,
+        resolve_metadata_json,
+        resolve_stats_log_file,
+        run_generation,
+    )
+except ImportError:
+    from runner import (
+        build_manifest,
+        build_sampling_params,
+        resolve_metadata_json,
+        resolve_stats_log_file,
+        run_generation,
+    )
+
+_DEFAULT_MODEL = "inclusionAI/Ming-omni-tts-0.5B"
+_REPO_ROOT = Path(__file__).resolve().parents[4]
+_DEFAULT_DEPLOY_CONFIG = str(_REPO_ROOT / "vllm_omni/deploy/ming_tts.yaml")
+_CASES_FILE = Path(__file__).with_name("cases.yaml")
+
+CASE_DEFAULTS = yaml.safe_load(_CASES_FILE.read_text(encoding="utf-8")) or {}
+if not CASE_DEFAULTS:
+    raise RuntimeError(f"Empty or missing case definitions in {_CASES_FILE}")
+
+
+def _build_parser() -> FlexibleArgumentParser:
+    p = FlexibleArgumentParser(description="Offline Ming-omni-tts example")
+    p.add_argument("--model", default=_DEFAULT_MODEL, help="Model name or local path")
+    p.add_argument("--deploy-config", default=None, help="Deploy config YAML; auto-selected when omitted")
+    p.add_argument("--case", choices=sorted(CASE_DEFAULTS), default="style", help="Built-in demo case")
+    p.add_argument("--text", default=None, help="Override case text")
+    p.add_argument("--prompt", default=None, help="Override the system prompt prefix")
+    p.add_argument("--instructions", default=None, help="Free-form Ming instruction string")
+    p.add_argument(
+        "--instruction-json", default=None, help='Structured Ming instruction JSON, e.g. \'{"方言":"广粤话"}\''
+    )
+    p.add_argument("--ref-audio", default=None, help="Single reference audio path for cloning-style cases")
+    p.add_argument("--ref-audio-paths", nargs="+", default=None, help="Multiple reference audio paths (podcast)")
+    p.add_argument("--ref-text", default=None, help="Reference transcript; required for zero_shot")
+    p.add_argument("--speaker-embedding", default=None, help="Path to a JSON speaker embedding file")
+    p.add_argument(
+        "--extract-speaker-embeddings",
+        action="store_true",
+        help="Extract speaker embeddings from ref audio via campplus.onnx",
+    )
+    p.add_argument("--max-decode-steps", type=int, default=None, help="Override ming_max_decode_steps")
+    p.add_argument("--output-dir", default="output_audio", help="Directory for output wav files")
+    p.add_argument("--output-name", default=None, help="Output wav filename")
+    p.add_argument("--num-prompts", type=int, default=1, help="Repeat the same prompt N times")
+    p.add_argument("--streaming", action="store_true", help="Use AsyncOmni with async_chunk streaming")
+    p.add_argument("--trust-remote-code", action="store_true")
+    p.add_argument("--enforce-eager", action="store_true")
+    p.add_argument(
+        "--log-stats", "--enable-stats", dest="log_stats", action="store_true", help="Enable Omni stats logging"
+    )
+    p.add_argument("--stats-log-file", default=None, help="Path for the Omni stats log file")
+    p.add_argument("--metadata-json", default=None, help="Path for the run manifest JSON")
+    p.add_argument("--stage-init-timeout", type=int, default=300, help="Per-stage init timeout (s)")
+    p.add_argument("--init-timeout", type=int, default=600, help="Total init timeout (s)")
+    p.add_argument("--batch-timeout", type=int, default=5, help="Batch timeout (s)")
+    p.add_argument("--shm-threshold-bytes", type=int, default=65536)
+    p.add_argument("--worker-backend", default="multi_process", choices=["multi_process", "ray"])
+    p.add_argument("--ray-address", default=None, help="Ray cluster address (--worker-backend ray)")
+    return p
+
+
+def _finalize_args(args) -> None:
+    if args.instructions is not None and args.instruction_json is not None:
+        raise RuntimeError("Use either --instructions or --instruction-json, not both")
+    if args.num_prompts < 1:
+        raise RuntimeError("--num-prompts must be at least 1")
+    if args.streaming and args.num_prompts != 1:
+        raise RuntimeError("--streaming currently supports exactly one prompt")
+    if args.deploy_config is None:
+        args.deploy_config = _DEFAULT_DEPLOY_CONFIG
+
+
+def _load_waveform(path: str) -> torch.Tensor:
+    samples, sr = sf.read(path, dtype="float32")
+    wav = torch.as_tensor(samples, dtype=torch.float32)
+    if wav.ndim == 2:
+        wav = wav.mean(dim=1)
+    wav = wav.reshape(1, -1)
+    if int(sr) != SAMPLE_RATE:
+        wav = torchaudio.functional.resample(wav, int(sr), SAMPLE_RATE)
+    return wav
+
+
+def _load_speaker_embedding(path: str) -> torch.Tensor:
+    return torch.as_tensor(json.loads(Path(path).read_text(encoding="utf-8")), dtype=torch.float32)
+
+
+def _ref_audio_paths(args) -> list[str]:
+    if args.ref_audio is not None and args.ref_audio_paths is not None:
+        raise RuntimeError("Use either --ref-audio or --ref-audio-paths, not both")
+    if args.ref_audio_paths is not None:
+        return list(args.ref_audio_paths)
+    return [args.ref_audio] if args.ref_audio else []
+
+
+def _resolve_reference_inputs(args, case: dict, paths: list[str]):
+    required = int(case.get("requires_ref_audio_count", 0))
+    if required > 0 and len(paths) < required:
+        raise RuntimeError(f"Case '{args.case}' needs {required} ref audio paths via --ref-audio-paths")
+    if required <= 0 and case.get("requires_ref_audio") and not paths:
+        raise RuntimeError(f"--ref-audio required for case '{args.case}'")
+    if not paths:
+        return None
+    wavs = [_load_waveform(p) for p in paths]
+    return wavs[0] if len(wavs) == 1 else wavs
+
+
+def _resolve_speaker_embedding(args, case: dict, paths: list[str]):
+    if args.speaker_embedding:
+        return _load_speaker_embedding(args.speaker_embedding)
+    if not (case.get("auto_extract_speaker_embeddings") or args.extract_speaker_embeddings) or not paths:
+        return None
+    embs = MingSpeakerEmbeddingExtractor(args.model).extract_many(paths)
+    if not embs:
+        raise RuntimeError("Speaker extraction produced no embeddings")
+    return embs[0] if len(embs) == 1 else torch.stack(embs, dim=0)
+
+
+def _build_prompt_payload(tokenizer, args):
+    case = CASE_DEFAULTS[args.case]
+    paths = _ref_audio_paths(args)
+    prompt_text = args.ref_text if args.ref_text is not None else case.get("prompt_text")
+    if case.get("requires_ref_text") and not prompt_text:
+        raise RuntimeError(f"--ref-text required for case '{args.case}'")
+    prompt_waveform = _resolve_reference_inputs(args, case, paths) if prompt_text is not None else None
+    speaker_embedding = _resolve_speaker_embedding(args, case, paths)
+    use_zero_spk_emb = bool(case.get("use_zero_spk_emb")) and prompt_waveform is None and speaker_embedding is None
+    runtime_controls = {KEY_MAX_DECODE_STEPS: args.max_decode_steps or case["max_decode_steps"]}
+    for key, field in [(KEY_CFG, "cfg"), (KEY_SIGMA, "sigma"), (KEY_TEMPERATURE, "temperature")]:
+        if field in case:
+            runtime_controls[key] = case[field]
+    instruction = (
+        json.loads(args.instruction_json) if args.instruction_json else (args.instructions or case.get("instruction"))
+    )
+    return build_ming_dense_prompt(
+        tokenizer,
+        prompt=args.prompt or case["prompt"],
+        text=args.text or case["text"],
+        runtime_controls=runtime_controls,
+        instruction=instruction,
+        prompt_text=prompt_text,
+        prompt_waveform=prompt_waveform,
+        speaker_embedding=speaker_embedding,
+        use_zero_spk_emb=use_zero_spk_emb,
+    )
+
+
+def main():
+    args = _build_parser().parse_args()
+    _finalize_args(args)
+    tokenizer = AutoTokenizer.from_pretrained(args.model, trust_remote_code=False)
+    prompt_payload = _build_prompt_payload(tokenizer, args)
+    case = CASE_DEFAULTS[args.case]
+    sampling_params_list = build_sampling_params(args.max_decode_steps or case["max_decode_steps"])
+    output_dir = Path(args.output_dir)
+    output_dir.mkdir(parents=True, exist_ok=True)
+    stats_log_file = resolve_stats_log_file(args)
+    summaries = run_generation(args, prompt_payload, sampling_params_list, output_dir, stats_log_file)
+    metadata_json = resolve_metadata_json(args)
+    manifest = build_manifest(args, prompt_payload, stats_log_file, summaries)
+    if metadata_json is not None:
+        Path(metadata_json).write_text(json.dumps(manifest, ensure_ascii=False, indent=2), encoding="utf-8")
+        print(f"Saved run manifest to {metadata_json}")
+
+
+if __name__ == "__main__":
+    os.environ.setdefault("VLLM_WORKER_MULTIPROC_METHOD", "spawn")
+    main()

examples/offline_inference/text_to_speech/ming_tts/runner.py

@@ -0,0 +1,277 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+import asyncio
+import json
+import time
+import uuid
+import wave
+from pathlib import Path
+
+import torch
+from vllm import SamplingParams
+
+from vllm_omni import AsyncOmni, Omni
+from vllm_omni.model_executor.models.ming_tts.config_ming_tts import (
+    KEY_SPEAKER_EMBEDDING,
+    SAMPLE_RATE,
+    TEXT_EOS_TOKEN_ID,
+)
+
+
+def coerce_audio_tensor(audio, *, async_chunk: bool) -> torch.Tensor:
+    if isinstance(audio, list):
+        parts = []
+        for item in audio:
+            tensor = torch.as_tensor(item, dtype=torch.float32).reshape(-1)
+            if tensor.numel() > 0:
+                parts.append(tensor)
+        if not parts:
+            return torch.zeros((0,), dtype=torch.float32)
+        return torch.cat(parts, dim=0)
+
+    return torch.as_tensor(audio, dtype=torch.float32).reshape(-1)
+
+
+def resolve_sr(sr) -> int:
+    if isinstance(sr, list):
+        sr = sr[-1]
+    if hasattr(sr, "item"):
+        return int(sr.item())
+    return int(sr)
+
+
+def extract_sample_rate(multimodal_output: dict) -> int:
+    sr = multimodal_output.get("sr")
+    if sr is None:
+        raise RuntimeError("Expected multimodal_output['sr']")
+    return resolve_sr(sr)
+
+
+def write_wav(path: str, audio: torch.Tensor, sample_rate: int) -> None:
+    audio = audio.clamp(-1.0, 1.0)
+    pcm16 = (audio * 32767.0).round().to(torch.int16).cpu().numpy()
+    with wave.open(path, "wb") as wav_file:
+        wav_file.setnchannels(1)
+        wav_file.setsampwidth(2)
+        wav_file.setframerate(int(sample_rate))
+        wav_file.writeframes(pcm16.tobytes())
+
+
+def request_index(request_id: str | None, fallback: int) -> int:
+    try:
+        return int(request_id)
+    except (TypeError, ValueError):
+        if isinstance(request_id, str):
+            head = request_id.split("_", 1)[0]
+            if head.isdigit():
+                return int(head)
+    return fallback
+
+
+def audio_summary(audio: torch.Tensor, sample_rate: int) -> dict:
+    waveform = audio.detach().cpu().reshape(-1).to(torch.float32)
+    return {
+        "sample_rate": int(sample_rate),
+        "num_samples": int(waveform.numel()),
+        "duration_seconds": float(waveform.numel()) / float(sample_rate),
+        "max_abs_amplitude": float(waveform.abs().max().item()) if waveform.numel() > 0 else 0.0,
+    }
+
+
+def resolve_output_name(output_name: str | None, case: str, index: int, total: int) -> str:
+    if total == 1:
+        return output_name or f"ming_{case}.wav"
+    base = Path(output_name or f"ming_{case}.wav")
+    return f"{base.stem}_{index:05d}{base.suffix or '.wav'}"
+
+
+def resolve_stats_log_file(args) -> str | None:
+    if not args.log_stats:
+        return None
+    if args.stats_log_file:
+        return args.stats_log_file
+    base = Path(args.output_name or f"ming_{args.case}.wav").stem
+    return str(Path(args.output_dir) / f"{base}_pipeline.log")
+
+
+def resolve_metadata_json(args) -> str | None:
+    if args.metadata_json:
+        return args.metadata_json
+    if args.log_stats:
+        base = Path(args.output_name or f"ming_{args.case}.wav").stem
+        return str(Path(args.output_dir) / f"{base}_manifest.json")
+    return None
+
+
+def build_manifest(args, prompt_payload, stats_log_file: str | None, outputs: list[dict]) -> dict:
+    additional_information = {}
+    if isinstance(prompt_payload, dict):
+        additional_information = dict(prompt_payload.get("additional_information", {}))
+    return {
+        "model": args.model,
+        "case": args.case,
+        "streaming": bool(args.streaming),
+        "deploy_config": args.deploy_config,
+        "enforce_eager": bool(args.enforce_eager),
+        "num_prompts": int(args.num_prompts),
+        "log_stats": bool(args.log_stats),
+        "stats_log_file": stats_log_file,
+        "prompt_text": additional_information.get("prompt_text"),
+        "instruction": additional_information.get("instruction"),
+        "speaker_embedding_shape": (
+            list(additional_information[KEY_SPEAKER_EMBEDDING].shape)
+            if KEY_SPEAKER_EMBEDDING in additional_information
+            and hasattr(additional_information[KEY_SPEAKER_EMBEDDING], "shape")
+            else None
+        ),
+        "outputs": outputs,
+        "generated_at_unix": time.time(),
+    }
+
+
+def build_engine_kwargs(args, stats_log_file: str | None) -> dict:
+    kwargs = {
+        "model": args.model,
+        "deploy_config": args.deploy_config,
+        "enforce_eager": args.enforce_eager,
+        "trust_remote_code": args.trust_remote_code,
+        "log_stats": args.log_stats,
+        "stage_init_timeout": args.stage_init_timeout,
+        "init_timeout": args.init_timeout,
+        "batch_timeout": args.batch_timeout,
+        "shm_threshold_bytes": args.shm_threshold_bytes,
+        "worker_backend": args.worker_backend,
+    }
+    if stats_log_file is not None:
+        kwargs["log_file"] = stats_log_file
+    if args.ray_address is not None:
+        kwargs["ray_address"] = args.ray_address
+    return kwargs
+
+
+def build_sampling_params(max_decode_steps: int) -> list[SamplingParams]:
+    return [
+        SamplingParams(
+            temperature=0.0,
+            max_tokens=max_decode_steps + 1,
+            stop_token_ids=[int(TEXT_EOS_TOKEN_ID)],
+        ),
+        SamplingParams(temperature=0.0, max_tokens=1),
+    ]
+
+
+async def run_streaming(args, prompt_payload, sampling_params_list, output_dir: Path, stats_log_file: str | None):
+    engine = AsyncOmni(**build_engine_kwargs(args, stats_log_file))
+    try:
+        all_audio_chunks = []
+        accumulated_samples = 0
+        chunk_idx = 0
+        start_time = time.time()
+        chunk_times = []
+        ttfp_seconds = None
+        final_stage_output = None
+        async for stage_output in engine.generate(
+            prompt=prompt_payload,
+            request_id=str(uuid.uuid4()),
+            sampling_params_list=sampling_params_list,
+        ):
+            final_stage_output = stage_output
+            multimodal_output = stage_output.multimodal_output or {}
+            audio = multimodal_output.get("audio")
+            if audio is None:
+                continue
+
+            finished = stage_output.finished
+            if isinstance(audio, torch.Tensor):
+                if finished:
+                    audio_chunk = audio[accumulated_samples:].float().detach().cpu()
+                else:
+                    audio_chunk = audio.float().detach().cpu()
+            elif isinstance(audio, list):
+                audio_chunk = torch.as_tensor(audio[chunk_idx], dtype=torch.float32).reshape(-1).cpu()
+            else:
+                audio_chunk = torch.as_tensor(audio, dtype=torch.float32).reshape(-1).cpu()
+
+            accumulated_samples += int(audio_chunk.numel())
+            chunk_idx += 1
+            if audio_chunk.numel() > 0:
+                now = time.time()
+                if ttfp_seconds is None:
+                    ttfp_seconds = now - start_time
+                chunk_times.append(now)
+                all_audio_chunks.append(audio_chunk)
+
+        if not all_audio_chunks:
+            raise RuntimeError("Streaming Ming example produced no audio chunks")
+
+        waveform = torch.cat(all_audio_chunks, dim=0)
+        output_name = resolve_output_name(args.output_name, args.case, 0, 1)
+        output_path = str(output_dir / output_name)
+        write_wav(output_path, waveform, SAMPLE_RATE)
+        summary = {
+            "request_id": getattr(final_stage_output, "request_id", None),
+            "stage_id": getattr(final_stage_output, "stage_id", None),
+            "output_path": output_path,
+            "stage_durations": getattr(final_stage_output, "stage_durations", {}),
+            "peak_memory_mb": getattr(final_stage_output, "peak_memory_mb", 0.0),
+            "ttfp_seconds": ttfp_seconds,
+            "mean_inter_chunk_seconds": (
+                sum(t1 - t0 for t0, t1 in zip(chunk_times, chunk_times[1:])) / (len(chunk_times) - 1)
+                if len(chunk_times) > 1
+                else None
+            ),
+        }
+        summary.update(audio_summary(waveform, SAMPLE_RATE))
+        print(f"Saved streaming output to {output_path}")
+        print(json.dumps(summary, ensure_ascii=False, indent=2))
+        return [summary]
+    finally:
+        engine.shutdown()
+
+
+def run_non_streaming(args, prompt_payload, sampling_params_list, output_dir: Path, stats_log_file: str | None):
+    engine = Omni(**build_engine_kwargs(args, stats_log_file))
+    try:
+        outputs = engine.generate(
+            prompts=[prompt_payload for _ in range(args.num_prompts)],
+            sampling_params_list=sampling_params_list,
+            py_generator=False,
+        )
+        summaries = []
+        for fallback_index, output in enumerate(outputs):
+            if output.final_output_type != "audio":
+                continue
+            multimodal_output = output.multimodal_output or {}
+            waveform = coerce_audio_tensor(multimodal_output.get("audio"), async_chunk=False)
+            sample_rate = extract_sample_rate(multimodal_output)
+            output_name = resolve_output_name(
+                args.output_name,
+                args.case,
+                request_index(output.request_id, fallback_index),
+                args.num_prompts,
+            )
+            output_path = str(output_dir / output_name)
+            write_wav(output_path, waveform, sample_rate)
+            summary = {
+                "request_id": output.request_id,
+                "stage_id": output.stage_id,
+                "output_path": output_path,
+                "stage_durations": output.stage_durations,
+                "peak_memory_mb": output.peak_memory_mb,
+            }
+            summary.update(audio_summary(waveform, sample_rate))
+            summaries.append(summary)
+            print(f"Saved output to {output_path}")
+            print(json.dumps(summary, ensure_ascii=False, indent=2))
+        if not summaries:
+            raise RuntimeError("Non-streaming Ming example produced no audio outputs")
+        return summaries
+    finally:
+        engine.close()
+
+
+def run_generation(args, prompt_payload, sampling_params_list, output_dir: Path, stats_log_file: str | None):
+    if args.streaming:
+        return asyncio.run(run_streaming(args, prompt_payload, sampling_params_list, output_dir, stats_log_file))
+    return run_non_streaming(args, prompt_payload, sampling_params_list, output_dir, stats_log_file)

examples/offline_inference/text_to_speech/voxtral_tts/end2end.py

@@ -212,7 +212,7 @@ def run_non_streaming(inputs, sampling_params_list, model_name, args, output_dir
     output_audio_dur = 0.0
 
     for batch_idx, o in enumerate(outputs):
-        audio_tensor = torch.cat(o.multimodal_output["audio"])
+        audio_tensor = torch.cat((o.multimodal_output["audio"],))
         audio_array = audio_tensor.tolist()
         output_audio_dur += float(len(audio_array)) / 24000
         if args.write_audio:

examples/online_serving/speech_to_video/README.md

@@ -0,0 +1,115 @@
+# Speech-To-Video
+
+This example demonstrates how to deploy the Wan2.2 speech-to-video (S2V) model for online video generation using vLLM-Omni.
+
+## Supported Models
+
+| Model | Model ID |
+|-------|----------|
+| Wan2.2 S2V (14B) | `Wan-AI/Wan2.2-S2V-14B` |
+
+## Start Server
+
+### Basic Start
+
+```bash
+VLLM_WORKER_MULTIPROC_METHOD=spawn \
+vllm serve Wan-AI/Wan2.2-S2V-14B --omni \
+  --model-class-name WanS2VPipeline \
+  --tensor-parallel-size 2 \
+  --flow-shift 3.0 \
+  --vae-use-slicing --vae-use-tiling \
+  --cache-backend cache_dit \
+  --port 8091
+```
+
+### Start with Script
+
+```bash
+bash run_server.sh
+```
+
+The script allows overriding:
+- `MODEL` (default: `Wan-AI/Wan2.2-S2V-14B`)
+- `PORT` (default: `8091`)
+- `FLOW_SHIFT` (default: `3.0`)
+- `TP` (default: `2`)
+- `CACHE_BACKEND` (default: `cache_dit`)
+
+### Start with 4 GPUs
+
+```bash
+TP=4 bash run_server.sh
+```
+
+## Sync API (Recommended for Testing)
+
+`POST /v1/videos/sync` blocks until generation completes and returns the raw
+video bytes (`video/mp4`) directly in the response body.
+
+```bash
+bash run_curl_speech_to_video.sh
+```
+
+## Async Job API
+
+`POST /v1/videos` creates an asynchronous job and returns immediately.
+
+```bash
+no_proxy=127.0.0.1 \
+create_response=$(curl -sS -X POST http://127.0.0.1:8091/v1/videos \
+  -H "Accept: application/json" \
+  -F "prompt=A person singing" \
+  -F 'image_reference={"image_url": "https://raw.githubusercontent.com/Wan-Video/Wan2.2/main/examples/Five%20Hundred%20Miles.png"}' \
+  -F 'audio_reference={"audio_url": "https://raw.githubusercontent.com/Wan-Video/Wan2.2/main/examples/Five%20Hundred%20Miles.MP3"}' \
+  -F "width=832" -F "height=480" \
+  -F "num_inference_steps=40" \
+  -F "guidance_scale=4.5" \
+  -F "fps=16")
+
+video_id=$(echo "$create_response" | jq -r '.id')
+
+# Poll until complete
+while true; do
+  status=$(curl -s "http://127.0.0.1:8091/v1/videos/${video_id}" | jq -r '.status')
+  if [ "$status" = "completed" ]; then
+    break
+  fi
+  if [ "$status" = "failed" ]; then
+    echo "Video generation failed"
+    exit 1
+  fi
+  sleep 2
+done
+
+# Download
+curl -L "http://127.0.0.1:8091/v1/videos/${video_id}/content" -o s2v_output.mp4
+```
+
+## Request Parameters
+
+| Parameter | Type | Default | Description |
+|-----------|------|---------|-------------|
+| `prompt` | str | - | Text description of the desired video |
+| `image_reference` | JSON str | - | Image reference: `{"image_url": "..."}` — supports HTTP(s) URLs or base64 data URLs |
+| `audio_reference` | JSON str | - | Audio reference: `{"audio_url": "..."}` — supports HTTP(s) URLs or base64 data URLs |
+| `width` | int | None | Video width in pixels |
+| `height` | int | None | Video height in pixels |
+| `fps` | int | None | Output video frame rate |
+| `num_inference_steps` | int | None | Number of denoising steps (40 recommended) |
+| `guidance_scale` | float | None | CFG guidance scale |
+| `seed` | int | None | Random seed for reproducibility |
+
+## Notes
+
+- S2V requires both a reference image (`image_reference`) and an audio reference
+  (`audio_reference`). The generated video will show a person matching the reference
+  image with lip movements synchronized to the audio.
+- `audio_reference` accepts a JSON string: `{"audio_url": "..."}` where the URL can be:
+  - An HTTP/HTTPS URL (e.g., `https://example.com/audio.mp3`)
+  - A base64 data URL (e.g., `data:audio/mp3;base64,...`)
+- `--model-class-name WanS2VPipeline` is required on the server to select the
+  S2V pipeline (distinct from the T2V/I2V pipelines).
+- `--cache-backend cache_dit` enables DiT caching for ~2x speedup on cached steps.
+- Audio is muxed into the output MP4 automatically.
+- Always pass `fps=16` to ensure correct video/audio alignment.

examples/online_serving/speech_to_video/run_curl_speech_to_video.sh

@@ -0,0 +1,44 @@
+#!/bin/bash
+# Wan2.2 S2V (speech-to-video) curl example using the sync video API.
+
+set -euo pipefail
+
+BASE_URL="${BASE_URL:-http://localhost:8091}"
+OUTPUT_PATH="${OUTPUT_PATH:-s2v_480p_serve.mp4}"
+IMAGE_URL="${IMAGE_URL:-https://raw.githubusercontent.com/Wan-Video/Wan2.2/main/examples/Five%20Hundred%20Miles.png}"
+AUDIO_URL="${AUDIO_URL:-https://raw.githubusercontent.com/Wan-Video/Wan2.2/main/examples/Five%20Hundred%20Miles.MP3}"
+PROMPT="${PROMPT:-A person singing}"
+WIDTH="${WIDTH:-832}"
+HEIGHT="${HEIGHT:-480}"
+NUM_INFERENCE_STEPS="${NUM_INFERENCE_STEPS:-40}"
+GUIDANCE_SCALE="${GUIDANCE_SCALE:-4.5}"
+FPS="${FPS:-16}"
+
+echo "Sending S2V request..."
+echo "  Image URL: $IMAGE_URL"
+echo "  Audio URL: $AUDIO_URL"
+echo "  Prompt: $PROMPT"
+echo "  Resolution: ${WIDTH}x${HEIGHT}"
+echo "  Steps: $NUM_INFERENCE_STEPS"
+echo "  FPS: $FPS"
+
+IMAGE_REF_JSON="{\"image_url\": \"${IMAGE_URL}\"}"
+AUDIO_REF_JSON="{\"audio_url\": \"${AUDIO_URL}\"}"
+
+no_proxy=127.0.0.1 \
+curl -X POST "${BASE_URL}/v1/videos/sync" \
+  -F "prompt=${PROMPT}" \
+  -F "image_reference=${IMAGE_REF_JSON}" \
+  -F "audio_reference=${AUDIO_REF_JSON}" \
+  -F "width=${WIDTH}" -F "height=${HEIGHT}" \
+  -F "num_inference_steps=${NUM_INFERENCE_STEPS}" \
+  -F "guidance_scale=${GUIDANCE_SCALE}" \
+  -F "fps=${FPS}" \
+  --output "${OUTPUT_PATH}"
+
+if [ -f "$OUTPUT_PATH" ] && [ -s "$OUTPUT_PATH" ]; then
+    echo "Saved video to ${OUTPUT_PATH} ($(du -h "$OUTPUT_PATH" | cut -f1))"
+else
+    echo "ERROR: Output file is empty or missing"
+    exit 1
+fi

examples/online_serving/speech_to_video/run_server.sh

@@ -0,0 +1,29 @@
+#!/bin/bash
+# Wan2.2 S2V (speech-to-video) online serving startup script
+
+MODEL="${MODEL:-Wan-AI/Wan2.2-S2V-14B}"
+PORT="${PORT:-8091}"
+FLOW_SHIFT="${FLOW_SHIFT:-3.0}"
+TP="${TP:-2}"
+CACHE_BACKEND="${CACHE_BACKEND:-cache_dit}"
+
+echo "Starting Wan2.2 S2V server..."
+echo "Model: $MODEL"
+echo "Port: $PORT"
+echo "Flow shift: $FLOW_SHIFT"
+echo "Tensor parallel size: $TP"
+echo "Cache backend: $CACHE_BACKEND"
+
+CACHE_BACKEND_FLAG=""
+if [ "$CACHE_BACKEND" != "none" ]; then
+    CACHE_BACKEND_FLAG="--cache-backend $CACHE_BACKEND"
+fi
+
+VLLM_WORKER_MULTIPROC_METHOD=spawn \
+vllm serve "$MODEL" --omni \
+    --model-class-name WanS2VPipeline \
+    --tensor-parallel-size "$TP" \
+    --flow-shift "$FLOW_SHIFT" \
+    --vae-use-slicing --vae-use-tiling \
+    --port "$PORT" \
+    $CACHE_BACKEND_FLAG

examples/online_serving/step_audio2/README.md

@@ -0,0 +1,193 @@
+# Step-Audio2 Online Serving
+
+This directory contains examples for running Step-Audio2 with vLLM-Omni's online serving API.
+
+## Installation
+
+Please refer to [README.md](../../../README.md)
+
+## Launch the Server
+
+```bash
+# Async chunk mode (recommended — lower first-packet latency for TTS)
+vllm serve stepfun-ai/Step-Audio2-mini --omni --port 8092 \
+    --stage-configs-path vllm_omni/model_executor/stage_configs/step_audio_2_async_chunk.yaml \
+    --trust-remote-code --enforce-eager
+```
+
+Sequential mode:
+```bash
+vllm serve stepfun-ai/Step-Audio2-mini --omni --port 8092 \
+    --stage-configs-path vllm_omni/model_executor/stage_configs/step_audio_2.yaml \
+    --trust-remote-code --enforce-eager
+```
+
+With local model:
+```bash
+vllm serve /path/to/Step-Audio-2-mini --omni --port 8092 \
+    --trust-remote-code --enforce-eager
+```
+
+## Send Requests
+
+### TTS via `/v1/audio/speech` (Recommended)
+
+```bash
+cd examples/online_serving/step_audio2
+
+# Python client
+python openai_speech_client.py --text "你好世界"
+
+# With custom system prompt
+python openai_speech_client.py --text "Hello, how are you?" \
+    --instructions "You are a friendly assistant."
+
+# Save to specific file
+python openai_speech_client.py --text "你好世界" -o output.wav
+```
+
+Or via curl:
+
+```bash
+curl -X POST http://localhost:8092/v1/audio/speech \
+    -H "Content-Type: application/json" \
+    -d '{"model":"stepfun-ai/Step-Audio2-mini","input":"你好世界","voice":"default"}' \
+    --output output.wav
+```
+
+This endpoint bypasses the chat template and directly triggers TTS mode. It supports async chunk streaming for low first-packet latency.
+
+**Note**: Speaker voice is controlled by `STEP_AUDIO2_DEFAULT_PROMPT_WAV` env var on the server side.
+
+### Chat Completions (ASR / S2ST)
+
+```bash
+# Audio to Text (ASR)
+python openai_chat_completion_client.py --query-type audio_to_text
+
+# Audio to Audio (S2ST)
+python openai_chat_completion_client.py --query-type audio_to_audio --audio-path /path/to/input.wav
+```
+
+| Argument | Description |
+|----------|-------------|
+| `--query-type`, `-q` | Query type: `audio_to_text`, `text_to_audio`, `audio_to_audio` |
+| `--audio-path`, `-a` | Path to input audio file (local or URL) |
+| `--text`, `-t` | Text to synthesize (for TTS mode) |
+| `--prompt`, `-p` | Custom prompt/question |
+| `--output-dir`, `-o` | Output directory for audio files (default: `output_online`) |
+| `--api-base` | API base URL (default: `http://localhost:8092/v1`) |
+
+### Curl (Chat Completions)
+
+```bash
+# Audio to Text
+bash run_curl.sh audio_to_text
+
+# Text to Audio
+bash run_curl.sh text_to_audio
+
+# Audio to Audio
+bash run_curl.sh audio_to_audio
+```
+
+## Query Types
+
+### 1. Audio to Text (ASR)
+
+Transcribe audio to text.
+
+```bash
+python openai_chat_completion_client.py \
+    --query-type audio_to_text \
+    --audio-path /path/to/speech.wav \
+    --prompt "Transcribe this audio."
+```
+
+### 2. Text to Audio (TTS)
+
+Convert text to speech:
+
+```bash
+# Via speech endpoint (recommended, returns WAV directly)
+python openai_speech_client.py --text "Hello, welcome to Step-Audio2."
+
+# Via chat completions
+python openai_chat_completion_client.py \
+    --query-type text_to_audio \
+    --text "Hello, welcome to Step-Audio2."
+```
+
+### 3. Audio to Audio (S2ST)
+
+Process input audio and generate text transcription + audio output.
+
+```bash
+python openai_chat_completion_client.py \
+    --query-type audio_to_audio \
+    --audio-path /path/to/source.wav
+```
+
+## Output
+
+- **Text output**: Printed to console
+- **Audio output**: Saved to `output_online/audio_0.wav` (24kHz WAV)
+
+## API Format
+
+Step-Audio2 uses the OpenAI-compatible chat completions API:
+
+```json
+{
+  "model": "stepfun-ai/Step-Audio2-mini",
+  "messages": [
+    {
+      "role": "system",
+      "content": [{"type": "text", "text": "Transcribe the audio."}]
+    },
+    {
+      "role": "user",
+      "content": [
+        {"type": "audio_url", "audio_url": {"url": "..."}},
+        {"type": "text", "text": "Please transcribe."}
+      ]
+    }
+  ],
+  "sampling_params_list": [
+    {"temperature": 0.7, "max_tokens": 1024},
+    {"temperature": 0.0, "max_tokens": 1}
+  ]
+}
+```
+
+## Performance
+
+### Async Chunk vs Sequential
+
+Benchmark via `/v1/audio/speech` (4x RTX 3090, 10 prompts, concurrency=1):
+
+| Mode | Mean TTFP | Mean E2E | Mean RTF |
+|------|-----------|----------|----------|
+| Sequential | 4316ms | 4316ms | 0.938 |
+| **Async Chunk** | **1437ms** | 4362ms | 0.949 |
+
+Async chunk reduces TTFP by **67%** by streaming audio token chunks from Thinker to Token2Wav as they are generated. RTF < 1 in both modes (real-time capable).
+
+## Troubleshooting
+
+### Server not responding
+- Check if the server is running: `curl http://localhost:8092/v1/models`
+- Verify the port number matches
+
+### FileNotFoundError: prompt_wav file not found
+- Ensure `default_female.wav` exists at `{model_dir}/assets/default_female.wav`
+- Or set `STEP_AUDIO2_DEFAULT_PROMPT_WAV` environment variable when launching the server
+
+### Audio not generated
+- For TTS, use the `/v1/audio/speech` endpoint (recommended) or `openai_speech_client.py`
+- For chat completions TTS, ensure the prompt ends with `<tts_start>`
+- Check server logs for errors
+
+### Out of memory
+- Reduce `gpu_memory_utilization` in stage configs
+- Use a smaller batch size

examples/online_serving/step_audio2/openai_chat_completion_client.py

@@ -0,0 +1,342 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""
+OpenAI-compatible client for Step-Audio2 online serving.
+
+Step-Audio2 is a two-stage model:
+  - Stage 0 (Thinker): Audio understanding -> Text + Audio tokens
+  - Stage 1 (Token2Wav): Audio tokens -> Waveform (24kHz)
+
+Usage examples:
+  # Start server first:
+  vllm serve stepfun-ai/Step-Audio2-mini --omni --port 8092
+
+  # Audio to text (ASR)
+  python openai_chat_completion_client.py --query-type audio_to_text
+
+  # Text to audio (TTS)
+  python openai_chat_completion_client.py --query-type text_to_audio --text "Hello world"
+
+  # Audio to audio (Voice conversion)
+  python openai_chat_completion_client.py --query-type audio_to_audio --audio-path input.wav
+"""
+
+import base64
+import logging
+import os
+
+import requests
+from openai import OpenAI
+from vllm.assets.audio import AudioAsset
+from vllm.utils.argparse_utils import FlexibleArgumentParser
+
+logging.basicConfig(level=logging.INFO, format="%(levelname)s: %(message)s")
+logger = logging.getLogger(__name__)
+
+SEED = 42
+
+
+def encode_base64_content_from_file(file_path: str) -> str:
+    """Encode a local file to base64 format."""
+    with open(file_path, "rb") as f:
+        return base64.b64encode(f.read()).decode("utf-8")
+
+
+def encode_base64_content_from_url(url: str) -> str:
+    """Encode content from URL to base64 format."""
+    with requests.get(url) as response:
+        response.raise_for_status()
+        return base64.b64encode(response.content).decode("utf-8")
+
+
+def get_audio_url(audio_path: str | None = None) -> str:
+    """Convert audio path to URL format for the API.
+
+    Args:
+        audio_path: Local file path or URL. If None, uses default test audio.
+
+    Returns:
+        Audio URL (either original URL or base64 data URL)
+    """
+    if not audio_path:
+        return AudioAsset("mary_had_lamb").url
+
+    if audio_path.startswith(("http://", "https://")):
+        return audio_path
+
+    if not os.path.exists(audio_path):
+        raise FileNotFoundError(f"Audio file not found: {audio_path}")
+
+    # Detect MIME type
+    audio_path_lower = audio_path.lower()
+    if audio_path_lower.endswith(".wav"):
+        mime_type = "audio/wav"
+    elif audio_path_lower.endswith((".mp3", ".mpeg")):
+        mime_type = "audio/mpeg"
+    elif audio_path_lower.endswith(".ogg"):
+        mime_type = "audio/ogg"
+    elif audio_path_lower.endswith(".flac"):
+        mime_type = "audio/flac"
+    else:
+        mime_type = "audio/wav"
+
+    audio_base64 = encode_base64_content_from_file(audio_path)
+    return f"data:{mime_type};base64,{audio_base64}"
+
+
+def get_audio_to_text_query(
+    audio_path: str | None = None,
+    custom_prompt: str | None = None,
+) -> dict:
+    """Build query for Audio-to-Text (ASR) mode."""
+    question = custom_prompt or "Please transcribe the audio content."
+    audio_url = get_audio_url(audio_path)
+
+    return {
+        "role": "user",
+        "content": [
+            {"type": "audio_url", "audio_url": {"url": audio_url}},
+            {"type": "text", "text": question},
+        ],
+    }
+
+
+def get_text_to_audio_query(
+    text: str | None = None,
+    custom_prompt: str | None = None,
+) -> dict:
+    """Build query for Text-to-Audio (TTS) mode.
+
+    Returns a list of messages: user message + assistant message with <tts_start>.
+    The assistant message triggers TTS mode and must be used with
+    continue_final_message=True so the chat template does not append <|im_end|>.
+    """
+    text_to_speak = text or "Hello, this is a test of Step Audio 2 text to speech."
+
+    return [
+        {
+            "role": "user",
+            "content": [{"type": "text", "text": text_to_speak}],
+        },
+        {
+            "role": "assistant",
+            "content": "<tts_start>",
+        },
+    ]
+
+
+def get_audio_to_audio_query(
+    audio_path: str | None = None,
+    custom_prompt: str | None = None,
+) -> dict:
+    """Build query for Audio-to-Audio (S2ST) mode.
+
+    Returns a list of messages: user message (with audio) + assistant message
+    with <tts_start>. Must be used with continue_final_message=True.
+    """
+    question = custom_prompt or "Please listen to this audio and repeat its content."
+    audio_url = get_audio_url(audio_path)
+
+    return [
+        {
+            "role": "user",
+            "content": [
+                {"type": "audio_url", "audio_url": {"url": audio_url}},
+                {"type": "text", "text": question},
+            ],
+        },
+        {
+            "role": "assistant",
+            "content": "<tts_start>",
+        },
+    ]
+
+
+def get_system_prompt(query_type: str) -> dict:
+    """Get system prompt based on query type."""
+    if query_type == "audio_to_text":
+        system_text = "You are a speech recognition assistant. Transcribe the audio accurately."
+    elif query_type == "text_to_audio":
+        system_text = "You are a text-to-speech assistant. Read the text aloud exactly as provided."
+    else:  # audio_to_audio
+        system_text = "You are an audio processing assistant. Listen and repeat the audio content."
+
+    return {
+        "role": "system",
+        "content": [{"type": "text", "text": system_text}],
+    }
+
+
+QUERY_MAP = {
+    "audio_to_text": get_audio_to_text_query,
+    "text_to_audio": get_text_to_audio_query,
+    "audio_to_audio": get_audio_to_audio_query,
+}
+
+
+def run_step_audio2(args) -> None:
+    """Run Step-Audio2 inference via OpenAI-compatible API."""
+
+    # Initialize OpenAI client
+    client = OpenAI(
+        api_key="EMPTY",
+        base_url=args.api_base,
+    )
+
+    model_name = args.model
+
+    # Sampling parameters for each stage
+    thinker_sampling_params = {
+        "temperature": 0.7,
+        "top_p": 0.9,
+        "top_k": -1,
+        "max_tokens": args.max_tokens,
+        "seed": SEED,
+        "detokenize": True,
+        "repetition_penalty": 1.1 if args.query_type != "audio_to_text" else 1.05,
+    }
+
+    # For ASR, stop at EOS token
+    if args.query_type == "audio_to_text":
+        thinker_sampling_params["stop_token_ids"] = [151645]
+
+    token2wav_sampling_params = {
+        "temperature": 0.0,
+        "top_p": 1.0,
+        "top_k": -1,
+        "max_tokens": 1,
+        "seed": SEED,
+        "detokenize": False,
+    }
+
+    sampling_params_list = [
+        thinker_sampling_params,
+        token2wav_sampling_params,
+    ]
+
+    # Build query based on type
+    query_func = QUERY_MAP[args.query_type]
+    if args.query_type == "audio_to_text":
+        user_message = query_func(audio_path=args.audio_path, custom_prompt=args.prompt)
+    elif args.query_type == "text_to_audio":
+        user_message = query_func(text=args.text, custom_prompt=args.prompt)
+    else:  # audio_to_audio
+        user_message = query_func(audio_path=args.audio_path, custom_prompt=args.prompt)
+
+    system_message = get_system_prompt(args.query_type)
+
+    logger.info(f"Query type: {args.query_type}")
+    logger.info(f"Model: {model_name}")
+    logger.info("Sending request to server...")
+
+    # Build messages list.
+    # TTS and S2ST return a list of messages (user + assistant with <tts_start>);
+    # ASR returns a single user message dict.
+    if isinstance(user_message, list):
+        messages = [system_message] + user_message
+    else:
+        messages = [system_message, user_message]
+
+    # For TTS/S2ST, use continue_final_message so <tts_start> in the
+    # assistant turn is not followed by <|im_end|>.
+    extra_body: dict = {"sampling_params_list": sampling_params_list}
+    if args.query_type in ("text_to_audio", "audio_to_audio"):
+        extra_body["continue_final_message"] = True
+        extra_body["add_generation_prompt"] = False
+
+    # Send request
+    chat_completion = client.chat.completions.create(
+        messages=messages,
+        model=model_name,
+        extra_body=extra_body,
+    )
+
+    # Process response
+    os.makedirs(args.output_dir, exist_ok=True)
+
+    audio_count = 0
+    for choice in chat_completion.choices:
+        if choice.message.audio:
+            # Audio response
+            audio_data = base64.b64decode(choice.message.audio.data)
+            audio_file = os.path.join(args.output_dir, f"audio_{audio_count}.wav")
+            with open(audio_file, "wb") as f:
+                f.write(audio_data)
+            logger.info(f"Audio saved to: {audio_file}")
+            audio_count += 1
+
+        if choice.message.content:
+            # Text response
+            logger.info(f"Text output: {choice.message.content}")
+
+    if audio_count == 0 and args.query_type != "audio_to_text":
+        logger.warning("No audio output received. Check server logs for details.")
+
+    logger.info("Done!")
+
+
+def parse_args():
+    parser = FlexibleArgumentParser(description="OpenAI-compatible client for Step-Audio2 online serving")
+
+    parser.add_argument(
+        "--query-type",
+        "-q",
+        type=str,
+        default="audio_to_text",
+        choices=QUERY_MAP.keys(),
+        help="Query type: audio_to_text (ASR), text_to_audio (TTS), audio_to_audio",
+    )
+    parser.add_argument(
+        "--audio-path",
+        "-a",
+        type=str,
+        default=None,
+        help="Path to input audio file (local path or URL)",
+    )
+    parser.add_argument(
+        "--text",
+        "-t",
+        type=str,
+        default=None,
+        help="Text to synthesize (for text_to_audio mode)",
+    )
+    parser.add_argument(
+        "--prompt",
+        "-p",
+        type=str,
+        default=None,
+        help="Custom prompt/question",
+    )
+    parser.add_argument(
+        "--model",
+        "-m",
+        type=str,
+        default="stepfun-ai/Step-Audio2-mini",
+        help="Model name",
+    )
+    parser.add_argument(
+        "--api-base",
+        type=str,
+        default="http://localhost:8092/v1",
+        help="API base URL",
+    )
+    parser.add_argument(
+        "--max-tokens",
+        type=int,
+        default=1024,
+        help="Maximum tokens for Thinker stage",
+    )
+    parser.add_argument(
+        "--output-dir",
+        "-o",
+        type=str,
+        default="output_online",
+        help="Output directory for audio files",
+    )
+
+    return parser.parse_args()
+
+
+if __name__ == "__main__":
+    args = parse_args()
+    run_step_audio2(args)

examples/online_serving/step_audio2/openai_speech_client.py

@@ -0,0 +1,130 @@
+"""OpenAI-compatible client for Step-Audio2 TTS via /v1/audio/speech endpoint.
+
+Examples:
+    # Basic TTS
+    python openai_speech_client.py --text "你好世界"
+
+    # With custom system prompt
+    python openai_speech_client.py --text "Hello, how are you?" \
+        --instructions "You are a friendly assistant."
+
+    # Save to specific file
+    python openai_speech_client.py --text "你好世界" -o output.wav
+"""
+
+import argparse
+
+import httpx
+
+DEFAULT_API_BASE = "http://localhost:8092"
+DEFAULT_API_KEY = "EMPTY"
+
+
+def run_tts_generation(args) -> None:
+    """Run TTS generation via /v1/audio/speech API."""
+    payload = {
+        "model": args.model,
+        "input": args.text,
+        "voice": args.voice,
+        "response_format": args.response_format,
+    }
+
+    if args.instructions:
+        payload["instructions"] = args.instructions
+
+    print(f"Model: {args.model}")
+    print(f"Text: {args.text}")
+    print(f"Voice: {args.voice}")
+    print("Generating audio...")
+
+    api_url = f"{args.api_base}/v1/audio/speech"
+    headers = {
+        "Content-Type": "application/json",
+        "Authorization": f"Bearer {args.api_key}",
+    }
+
+    with httpx.Client(timeout=300.0) as client:
+        response = client.post(api_url, json=payload, headers=headers)
+
+    if response.status_code != 200:
+        print(f"Error: {response.status_code}")
+        print(response.text)
+        return
+
+    try:
+        text = response.content.decode("utf-8")
+        if text.startswith('{"error"'):
+            print(f"Error: {text}")
+            return
+    except UnicodeDecodeError:
+        pass
+
+    output_path = args.output or "tts_output.wav"
+    with open(output_path, "wb") as f:
+        f.write(response.content)
+    print(f"Audio saved to: {output_path}")
+
+
+def parse_args():
+    parser = argparse.ArgumentParser(
+        description="Step-Audio2 TTS client via /v1/audio/speech",
+        formatter_class=argparse.RawDescriptionHelpFormatter,
+        epilog=__doc__,
+    )
+    parser.add_argument(
+        "--api-base",
+        type=str,
+        default=DEFAULT_API_BASE,
+        help=f"API base URL (default: {DEFAULT_API_BASE})",
+    )
+    parser.add_argument(
+        "--api-key",
+        type=str,
+        default=DEFAULT_API_KEY,
+        help="API key (default: EMPTY)",
+    )
+    parser.add_argument(
+        "--model",
+        "-m",
+        type=str,
+        default="stepfun-ai/Step-Audio2-mini",
+        help="Model name/path",
+    )
+    parser.add_argument(
+        "--text",
+        type=str,
+        required=True,
+        help="Text to synthesize",
+    )
+    parser.add_argument(
+        "--voice",
+        type=str,
+        default="default",
+        help="Voice name (default: default)",
+    )
+    parser.add_argument(
+        "--instructions",
+        type=str,
+        default=None,
+        help="System prompt for the Thinker stage",
+    )
+    parser.add_argument(
+        "--response-format",
+        type=str,
+        default="wav",
+        choices=["wav", "pcm"],
+        help="Audio output format (default: wav)",
+    )
+    parser.add_argument(
+        "--output",
+        "-o",
+        type=str,
+        default=None,
+        help="Output audio file path (default: tts_output.wav)",
+    )
+    return parser.parse_args()
+
+
+if __name__ == "__main__":
+    args = parse_args()
+    run_tts_generation(args)

examples/online_serving/step_audio2/run_curl.sh

@@ -0,0 +1,149 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+# Step-Audio2 curl client for online serving
+# Usage: bash run_curl.sh [audio_to_text|text_to_audio|audio_to_audio]
+
+QUERY_TYPE="${1:-audio_to_text}"
+API_BASE="${API_BASE:-http://localhost:8092}"
+
+# Validate query type
+if [[ ! "$QUERY_TYPE" =~ ^(audio_to_text|text_to_audio|audio_to_audio)$ ]]; then
+    echo "Error: Invalid query type '$QUERY_TYPE'"
+    echo "Usage: $0 [audio_to_text|text_to_audio|audio_to_audio]"
+    echo "  audio_to_text: Speech recognition (ASR)"
+    echo "  text_to_audio: Text-to-speech (TTS)"
+    echo "  audio_to_audio: Voice conversion"
+    exit 1
+fi
+
+SEED=42
+
+# Default test audio URL
+MARY_HAD_LAMB_URL="https://vllm-public-assets.s3.us-west-2.amazonaws.com/multimodal_asset/mary_had_lamb.ogg"
+
+# Sampling parameters for Thinker stage
+thinker_sampling_params='{
+  "temperature": 0.7,
+  "top_p": 0.9,
+  "top_k": -1,
+  "max_tokens": 1024,
+  "seed": 42,
+  "detokenize": true,
+  "repetition_penalty": 1.05
+}'
+
+# Sampling parameters for Token2Wav stage
+token2wav_sampling_params='{
+  "temperature": 0.0,
+  "top_p": 1.0,
+  "top_k": -1,
+  "max_tokens": 1,
+  "seed": 42,
+  "detokenize": false
+}'
+
+# Build request based on query type
+case "$QUERY_TYPE" in
+  audio_to_text)
+    system_content='[{"type": "text", "text": "You are a speech recognition assistant. Transcribe the audio accurately."}]'
+    user_content='[
+      {"type": "audio_url", "audio_url": {"url": "'"$MARY_HAD_LAMB_URL"'"}},
+      {"type": "text", "text": "Please transcribe this audio."}
+    ]'
+    # Add stop token for ASR
+    thinker_sampling_params='{
+      "temperature": 0.7,
+      "top_p": 0.9,
+      "top_k": -1,
+      "max_tokens": 1024,
+      "seed": 42,
+      "detokenize": true,
+      "repetition_penalty": 1.05,
+      "stop_token_ids": [151645]
+    }'
+    ;;
+  text_to_audio)
+    system_content='[{"type": "text", "text": "You are a text-to-speech assistant. Read the text aloud exactly as provided."}]'
+    user_content='[
+      {"type": "text", "text": "Hello, this is a test of Step Audio 2 text to speech synthesis.<tts_start>"}
+    ]'
+    thinker_sampling_params='{
+      "temperature": 0.7,
+      "top_p": 0.9,
+      "top_k": -1,
+      "max_tokens": 1024,
+      "seed": 42,
+      "detokenize": true,
+      "repetition_penalty": 1.1
+    }'
+    ;;
+  audio_to_audio)
+    system_content='[{"type": "text", "text": "You are an audio processing assistant. Listen and repeat the audio content."}]'
+    user_content='[
+      {"type": "audio_url", "audio_url": {"url": "'"$MARY_HAD_LAMB_URL"'"}},
+      {"type": "text", "text": "Please listen to this audio and repeat its content.<tts_start>"}
+    ]'
+    thinker_sampling_params='{
+      "temperature": 0.7,
+      "top_p": 0.9,
+      "top_k": -1,
+      "max_tokens": 1024,
+      "seed": 42,
+      "detokenize": true,
+      "repetition_penalty": 1.1
+    }'
+    ;;
+esac
+
+sampling_params_list='[
+  '"$thinker_sampling_params"',
+  '"$token2wav_sampling_params"'
+]'
+
+echo "Query type: $QUERY_TYPE"
+echo "API base: $API_BASE"
+echo "Sending request..."
+echo ""
+
+output=$(curl -sS -X POST "${API_BASE}/v1/chat/completions" \
+    -H "Content-Type: application/json" \
+    -d @- <<EOF
+{
+  "model": "stepfun-ai/Step-Audio2-mini",
+  "sampling_params_list": $sampling_params_list,
+  "messages": [
+    {
+      "role": "system",
+      "content": $system_content
+    },
+    {
+      "role": "user",
+      "content": $user_content
+    }
+  ]
+}
+EOF
+)
+
+# Extract and display text content
+text_content=$(echo "$output" | jq -r '.choices[0].message.content // empty')
+if [[ -n "$text_content" ]]; then
+    echo "Text output: $text_content"
+fi
+
+# Check for audio content
+audio_data=$(echo "$output" | jq -r '.choices[0].message.audio.data // empty')
+if [[ -n "$audio_data" ]]; then
+    echo "Audio output received (base64 encoded)"
+    echo "To save audio, use the Python client or decode the base64 data"
+fi
+
+# Check for errors
+error=$(echo "$output" | jq -r '.error // empty')
+if [[ -n "$error" ]]; then
+    echo "Error: $error"
+fi
+
+echo ""
+echo "Done!"

examples/online_serving/text_to_speech/README.md

@@ -16,6 +16,7 @@ For the full list of supported architectures across all modalities, see
 |---|---|---|---|---|---|
 | Fish Speech S2 Pro | `fishaudio/s2-pro` | ✓ (`ref_audio`+`ref_text`) | ✓ (PCM stream) | — | ✓ |
 | GLM-TTS | `zai-org/GLM-TTS` | ✓ (`ref_audio`+`ref_text`, required) | ✓ (PCM stream) | — | ✓ |
+| Ming-omni-tts | `inclusionAI/Ming-omni-tts-0.5B` | ✓ (`ref_audio` / `speaker_embedding`) | ✓ (PCM stream) | IP labels + structured `instructions` | — |
 | Ming-flash-omni-TTS | `Jonathan1909/Ming-flash-omni-2.0` | — (caption-controlled) | — | caption fields (`instructions`) | — |
 | MOSS-TTS-Nano | `OpenMOSS-Team/MOSS-TTS-Nano` | ✓ (`ref_audio` required) | ✓ (PCM stream) | — | ✓ |
 | OmniVoice | `k2-fsa/OmniVoice` | ✓ | — | — | — |
@@ -202,6 +203,51 @@ python fish_speech/gradio_demo.py --api-base http://localhost:8091  # if server
 
 ---
 
+## Ming-omni-tts
+
+Dense 0.5B two-stage TTS served through `/v1/audio/speech`. Ming uses the standard speech endpoint plus structured controls in `instructions`, `voice`, `language`, `ref_audio`, `ref_text`, and `speaker_embedding`.
+
+### Launch
+```bash
+bash examples/online_serving/text_to_speech/ming_tts/run_server.sh
+```
+Equivalent manual command:
+```bash
+vllm-omni serve inclusionAI/Ming-omni-tts-0.5B \
+    --deploy-config vllm_omni/deploy/ming_tts.yaml \
+    --host 0.0.0.0 --port 8091 \
+    --enforce-eager --omni
+```
+
+### Sending requests
+```bash
+python examples/online_serving/text_to_speech/ming_tts/openai_speech_client.py \
+    --text "你好，这是 Ming 在线语音合成测试。"
+```
+
+Structured dialect control:
+```bash
+python examples/online_serving/text_to_speech/ming_tts/openai_speech_client.py \
+    --text "我觉得社会企业同个人都有责任" \
+    --instruction-json '{"方言":"广粤话"}' \
+    --ref-audio /path/to/yue_prompt.wav
+```
+
+Zero-shot cloning:
+```bash
+python examples/online_serving/text_to_speech/ming_tts/openai_speech_client.py \
+    --text "我们的愿景是构建未来服务业的数字化基础设施，为世界带来更多微小而美好的改变。" \
+    --ref-audio /path/to/10002287-00000094.wav \
+    --ref-text "在此奉劝大家别乱打美白针。"
+```
+
+### Notes
+- `run_curl.sh` keeps a small sanity subset; use the Ming README for the broader request cookbook.
+- Online serving is speech-shaped today; music-only `bgm` and text-to-audio `tta` remain offline examples.
+- Full request details live in [`ming_tts/README.md`](ming_tts/README.md).
+
+---
+
 ## Ming-flash-omni-TTS
 
 Standalone talker-only deployment of Ming-flash-omni-2.0. Voice is controlled through caption text passed via `instructions`.

examples/online_serving/text_to_speech/fish_speech/gradio_demo.py

@@ -51,13 +51,18 @@ def build_payload(
     response_format: str = "wav",
     stream: bool = False,
 ) -> dict:
-    """Build the /v1/audio/speech request payload."""
+    """Build the /v1/audio/speech request payload.
+
+    Fish Speech has no built-in speakers; omit ``voice`` so the model
+    generates with a random timbre (controllable via text tags like
+    ``[cool]``, ``[warm]``, etc.).  When ``ref_audio`` is supplied we
+    enter voice-cloning mode instead.
+    """
     if not text or not text.strip():
         raise gr.Error("Please enter text to synthesize.")
 
     payload: dict = {
         "input": text.strip(),
-        "voice": "default",
         "response_format": "pcm" if stream else response_format,
         "stream": stream,
     }
@@ -193,7 +198,8 @@ def build_interface(api_base: str, stream_chunk_seconds: float = 0.5):
 
                 with gr.Accordion("Voice Cloning (optional)", open=False):
                     gr.Markdown(
-                        "Upload or link a short reference audio (10-30s) and provide its transcript to clone the voice."
+                        "Upload or link a short reference audio (10-30s) and provide its transcript to clone the voice. "
+                        "Leave empty for random timbre; use text tags like `[cool]`, `[warm]`, `[happy]` to steer the output."
                     )
                     ref_audio = gr.Audio(
                         label="Reference Audio",

examples/online_serving/text_to_speech/fish_speech/speech_client.py

@@ -39,7 +39,6 @@ def run_tts(args) -> None:
     payload = {
         "model": args.model,
         "input": args.text,
-        "voice": "default",
         "response_format": args.response_format,
     }
 

examples/online_serving/text_to_speech/higgs_audio_v3/README.md

@@ -0,0 +1,40 @@
+# Higgs-Audio V3 Online Serving
+
+## Start the server
+
+```bash
+# Default: GPU 0, port 8095
+./examples/online_serving/text_to_speech/higgs_audio_v3/run_server.sh
+
+# Custom GPU / port
+PORT=8096 GPUS=0,1 ./examples/online_serving/text_to_speech/higgs_audio_v3/run_server.sh
+```
+
+## Plain text TTS
+
+```bash
+python examples/online_serving/text_to_speech/higgs_audio_v3/batch_speech_client.py \
+    --base-url http://localhost:8095 \
+    --output-dir /tmp/higgs_v3_batch \
+    --prompts "Hello world." "The quick brown fox jumps over the lazy dog."
+```
+
+## Voice clone
+
+```bash
+python examples/online_serving/text_to_speech/higgs_audio_v3/batch_speech_client.py \
+    --base-url http://localhost:8095 \
+    --output-dir /tmp/higgs_v3_clone \
+    --ref-audio path/to/reference.wav \
+    --ref-text "transcript of the reference clip" \
+    --prompts "Text to synthesize in the cloned voice."
+```
+
+## curl example
+
+```bash
+curl -X POST http://localhost:8095/v1/audio/speech \
+    -H "Content-Type: application/json" \
+    -d '{"model": "higgs_audio_v3", "input": "Hello world."}' \
+    --output hello.wav
+```

examples/online_serving/text_to_speech/higgs_audio_v3/batch_speech_client.py

@@ -0,0 +1,115 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""Batch client for the higgs-audio v3 online server.
+
+Sends prompts to ``/v1/audio/speech`` and saves the returned WAV files.
+
+Usage (plain text -> speech):
+
+  python examples/online_serving/text_to_speech/higgs_audio_v3/batch_speech_client.py \
+      --base-url http://localhost:8095 \
+      --output-dir /tmp/higgs_v3_batch \
+      --prompts "Hello world." "The quick brown fox jumps over the lazy dog."
+
+Usage (voice clone):
+
+  python examples/online_serving/text_to_speech/higgs_audio_v3/batch_speech_client.py \
+      --base-url http://localhost:8095 \
+      --output-dir /tmp/higgs_v3_clone \
+      --ref-audio path/to/reference.wav \
+      --ref-text "the transcript of the reference clip" \
+      --prompts "Hello world."
+"""
+
+from __future__ import annotations
+
+import argparse
+import base64
+import sys
+from pathlib import Path
+
+DEFAULT_PROMPTS = (
+    "Hello world.",
+    "The quick brown fox jumps over the lazy dog.",
+    "Today is a beautiful day for a walk in the park.",
+    "Innovation distinguishes between a leader and a follower.",
+)
+
+
+def _slug(text: str) -> str:
+    import re
+
+    s = re.sub(r"\s+", "_", text.strip().lower())
+    return re.sub(r"[^a-z0-9_]+", "", s)[:32] or "prompt"
+
+
+def main() -> int:
+    parser = argparse.ArgumentParser(description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter)
+    parser.add_argument("--base-url", default="http://localhost:8095")
+    parser.add_argument("--model", default="higgs_audio_v3")
+    parser.add_argument("--prompts", nargs="+", default=list(DEFAULT_PROMPTS))
+    parser.add_argument("--output-dir", type=Path, default=Path("/tmp/higgs_v3_batch"))
+    parser.add_argument("--format", choices=("wav", "pcm"), default="wav")
+    parser.add_argument("--max-new-tokens", type=int, default=2048)
+    parser.add_argument("--seed", type=int, default=42)
+    parser.add_argument("--timeout-s", type=float, default=120.0)
+    parser.add_argument(
+        "--ref-audio",
+        type=Path,
+        default=None,
+        help="Reference clip for voice clone (WAV/FLAC/MP3 path). Pair with --ref-text.",
+    )
+    parser.add_argument(
+        "--ref-text",
+        type=str,
+        default=None,
+        help="Transcript of the reference clip. Optional but improves fidelity.",
+    )
+    args = parser.parse_args()
+
+    ref_audio_data_url: str | None = None
+    if args.ref_audio is not None:
+        if not args.ref_audio.exists():
+            print(f"ref-audio file not found: {args.ref_audio}", file=sys.stderr)
+            return 2
+        mime = "audio/wav" if args.ref_audio.suffix.lower() == ".wav" else "audio/mpeg"
+        ref_b64 = base64.b64encode(args.ref_audio.read_bytes()).decode("ascii")
+        ref_audio_data_url = f"data:{mime};base64,{ref_b64}"
+
+    try:
+        import httpx
+    except ImportError:
+        print("this client needs `httpx`. Install with `pip install httpx`.", file=sys.stderr)
+        return 2
+
+    args.output_dir.mkdir(parents=True, exist_ok=True)
+    url = args.base_url.rstrip("/") + "/v1/audio/speech"
+    failures = 0
+    with httpx.Client(timeout=args.timeout_s) as client:
+        for prompt in args.prompts:
+            payload = {
+                "model": args.model,
+                "input": prompt,
+                "response_format": args.format,
+                "max_new_tokens": args.max_new_tokens,
+                "seed": args.seed,
+            }
+            if ref_audio_data_url is not None:
+                payload["ref_audio"] = ref_audio_data_url
+                if args.ref_text:
+                    payload["ref_text"] = args.ref_text
+            resp = client.post(url, json=payload)
+            if resp.status_code != 200:
+                print(f"[FAIL] {prompt!r} -> {resp.status_code}: {resp.text[:200]}", file=sys.stderr)
+                failures += 1
+                continue
+            suffix = ".wav" if args.format == "wav" else ".pcm"
+            out = args.output_dir / f"{_slug(prompt)}{suffix}"
+            out.write_bytes(resp.content)
+            print(f"[ ok ] {prompt!r} -> {out} ({len(resp.content)} bytes)")
+
+    return 1 if failures else 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())

examples/online_serving/text_to_speech/higgs_audio_v3/run_server.sh

@@ -0,0 +1,32 @@
+#!/bin/bash
+# Launch vLLM-Omni server for higgs-audio v3.
+#
+# Supports plain text TTS and voice cloning via /v1/audio/speech.
+#
+# Usage:
+#   ./run_server.sh                 # default port 8095, GPU 0
+#   PORT=8096 GPUS=0,1 ./run_server.sh
+#   MODEL=/path/to/local/checkpoint ./run_server.sh
+
+set -e
+
+MODEL="${MODEL:-bosonai/higgs-audio-v3-tts-4b}"
+PORT="${PORT:-8095}"
+GPUS="${GPUS:-0}"
+GPU_UTIL="${GPU_UTIL:-0.6}"
+
+echo "Starting higgs-audio v3 server"
+echo "  MODEL=$MODEL"
+echo "  PORT=$PORT"
+echo "  CUDA_VISIBLE_DEVICES=$GPUS"
+
+CUDA_VISIBLE_DEVICES="$GPUS" \
+VLLM_USE_DEEP_GEMM=0 \
+VLLM_MOE_USE_DEEP_GEMM=0 \
+vllm-omni serve "$MODEL" \
+    --deploy-config vllm_omni/deploy/higgs_multimodal_qwen3.yaml \
+    --host 0.0.0.0 \
+    --port "$PORT" \
+    --gpu-memory-utilization "$GPU_UTIL" \
+    --trust-remote-code \
+    --omni

examples/online_serving/text_to_speech/ming_tts/README.md

@@ -0,0 +1,113 @@
+# Ming-omni-tts Online Serving
+
+Serve the dense `inclusionAI/Ming-omni-tts-0.5B` two-stage TTS model through
+the OpenAI-compatible `/v1/audio/speech` endpoint.
+
+## Start Server
+
+```bash
+vllm-omni serve inclusionAI/Ming-omni-tts-0.5B \
+    --deploy-config vllm_omni/deploy/ming_tts.yaml \
+    --omni \
+    --port 8091 \
+    --enforce-eager
+```
+
+Or:
+
+```bash
+cd examples/online_serving/text_to_speech/ming_tts
+./run_server.sh
+```
+
+The tested ROCm environment is summarized in the
+[Ming recipe](https://github.com/vllm-project/vllm-omni/blob/main/recipes/inclusionAI/Ming-omni-tts-0.5B.md).
+
+## Send Requests
+
+The Python client targets `http://localhost:8091/v1` with `api_key=EMPTY`; it
+does not call OpenAI's hosted API.
+
+```bash
+python openai_speech_client.py \
+    --text "你好，这是 Ming 在线语音合成测试。" \
+    --max-new-tokens 200
+```
+
+Style or dialect controls can be plain text or Ming JSON. The upstream
+dialect example also uses `yue_prompt.wav` for speaker conditioning:
+
+```bash
+python openai_speech_client.py \
+    --text "我觉得社会企业同个人都有责任" \
+    --instruction-json '{"方言":"广粤话"}' \
+    --ref-audio /path/to/yue_prompt.wav \
+    --max-new-tokens 200
+```
+
+When `--ref-audio` is supplied without `--ref-text`, the server extracts the
+Ming speaker embedding, matching upstream `use_spk_emb=True`, without using the
+audio as a zero-shot prompt.
+
+Reference-audio cloning:
+
+```bash
+python openai_speech_client.py \
+    --text "我们的愿景是构建未来服务业的数字化基础设施，为世界带来更多微小而美好的改变。" \
+    --ref-audio /path/to/10002287-00000094.wav \
+    --ref-text "在此奉劝大家别乱打美白针。" \
+    --max-new-tokens 200
+```
+
+Podcast-style multi-speaker prompt:
+
+```bash
+python openai_speech_client.py \
+    --text " speaker_1:你可以说一下，就大概说一下，可能虽然我也不知道，我看过那部电影没有。
+ speaker_2:就是那个叫什么，变相一节课的嘛。
+ speaker_1:嗯。
+ speaker_2:一部搞笑的电影。
+ speaker_1:一部搞笑的。" \
+    --ref-audio /path/to/CTS-CN-F2F-2019-11-11-423-012-A.wav \
+    --ref-audio /path/to/CTS-CN-F2F-2019-11-11-423-012-B.wav \
+    --ref-text " speaker_1:并且我们还要进行每个月还要考核 笔试的话还要进行笔试，做个，当服务员还要去笔试了
+ speaker_2:对啊，这真的很奇怪，就是 单纯的因，单纯自己工资不高，只是因为可能人家那个店比较出名一点，就对你苛刻要求"
+```
+
+Streaming PCM:
+
+```bash
+python openai_speech_client.py \
+    --text "你好，这是流式输出测试。" \
+    --stream \
+    --output ming_output.pcm
+```
+
+`run_curl.sh` keeps small smoke checks:
+
+```bash
+./run_curl.sh basic
+REF_AUDIO=/path/to/reference.wav REF_TEXT="在此奉劝大家别乱打美白针。" ./run_curl.sh zero_shot
+./run_curl.sh stream
+```
+
+## Request Fields
+
+| Field | Ming meaning |
+|-------|--------------|
+| `input` | target text |
+| `instructions` | plain style text, or JSON object for structured Ming controls |
+| `voice` | Ming IP voice label unless it resolves to an uploaded speaker |
+| `language` | Ming `方言` control |
+| `ref_audio` | speaker reference; with `ref_text`, also supplies the prompt waveform |
+| `ref_text` | transcript enabling zero-shot or podcast prompt-latent conditioning |
+| `speaker_embedding` | 192-d Ming speaker embedding |
+| `max_new_tokens` | Ming `max_decode_steps` |
+
+## Notes
+
+- `ref_audio` accepts local paths through the client, remote URLs, `file://`,
+  or `data:` URLs.
+- Non-streaming responses return WAV bytes; streaming responses return PCM.
+- Music-only `bgm` generation is offline-only until the API exposes Ming
+  prompt-mode selection.

examples/online_serving/text_to_speech/ming_tts/openai_speech_client.py

@@ -0,0 +1,225 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""OpenAI-compatible client for Ming-omni-tts via /v1/audio/speech.
+
+Examples:
+    python openai_speech_client.py --text "你好，世界"
+    python openai_speech_client.py --text "我觉得社会企业同个人都有责任" \
+        --instruction-json '{"方言":"广粤话"}' --ref-audio yue_prompt.wav --max-new-tokens 200
+    python openai_speech_client.py \
+        --text "我们的愿景是构建未来服务业的数字化基础设施，为世界带来更多微小而美好的改变。" \
+        --ref-audio 10002287-00000094.wav --ref-text "在此奉劝大家别乱打美白针。" \
+        --max-new-tokens 200
+    python openai_speech_client.py --text "speaker_1:你好。 speaker_2:你好。" \
+        --ref-audio speaker_1.wav --ref-audio speaker_2.wav --ref-text "speaker_1:你好。 speaker_2:你好。"
+    python openai_speech_client.py --text "你好，这是流式输出测试。" \
+        --stream --output ming_output.pcm
+"""
+
+import argparse
+import base64
+import json
+import os
+
+import httpx
+
+DEFAULT_API_BASE = "http://localhost:8091"
+DEFAULT_API_KEY = "EMPTY"
+DEFAULT_MODEL = "inclusionAI/Ming-omni-tts-0.5B"
+EXPECTED_SPEAKER_EMBEDDING_DIM = 192
+
+
+def encode_audio_to_base64(audio_path: str) -> str:
+    """Encode a local audio file to a base64 data URL."""
+    if not os.path.exists(audio_path):
+        raise FileNotFoundError(f"Audio file not found: {audio_path}")
+    ext = audio_path.lower().rsplit(".", 1)[-1]
+    mime_map = {
+        "wav": "audio/wav",
+        "mp3": "audio/mpeg",
+        "flac": "audio/flac",
+        "ogg": "audio/ogg",
+        "aac": "audio/aac",
+    }
+    mime_type = mime_map.get(ext, "audio/wav")
+    with open(audio_path, "rb") as f:
+        audio_b64 = base64.b64encode(f.read()).decode("utf-8")
+    return f"data:{mime_type};base64,{audio_b64}"
+
+
+def load_speaker_embedding(path: str) -> list[float]:
+    """Load and validate a 192-d Ming speaker embedding JSON file."""
+    with open(path, encoding="utf-8") as f:
+        data = json.load(f)
+
+    if not isinstance(data, list):
+        raise ValueError("speaker_embedding file must contain a JSON list")
+    if len(data) != EXPECTED_SPEAKER_EMBEDDING_DIM:
+        raise ValueError(
+            f"Ming dense speaker_embedding must have {EXPECTED_SPEAKER_EMBEDDING_DIM} values, got {len(data)}"
+        )
+
+    values = []
+    for index, value in enumerate(data):
+        try:
+            values.append(float(value))
+        except (TypeError, ValueError) as exc:
+            raise ValueError(f"speaker_embedding[{index}] must be a number, got {value!r}") from exc
+    return values
+
+
+def build_instruction_payload(args) -> str | None:
+    """Return a string payload for the API `instructions` field."""
+    if args.instructions and args.instruction_json:
+        raise ValueError("Use either --instructions or --instruction-json, not both")
+    if args.instruction_json:
+        parsed = json.loads(args.instruction_json)
+        return json.dumps(parsed, ensure_ascii=False)
+    return args.instructions
+
+
+def validate_args(args) -> None:
+    """Fail fast on invalid combinations before hitting the server."""
+    if args.ref_text and not args.ref_audio:
+        raise ValueError("--ref-audio is required when --ref-text is provided")
+    if args.speaker_embedding and args.ref_audio and len(args.ref_audio) > 1:
+        raise ValueError("--speaker-embedding cannot be combined with multiple --ref-audio values")
+
+
+def run_tts(args) -> None:
+    """Generate speech via the OpenAI-compatible /v1/audio/speech API."""
+    validate_args(args)
+
+    payload = {
+        "model": args.model,
+        "input": args.text,
+        "response_format": args.response_format,
+    }
+
+    if args.voice:
+        payload["voice"] = args.voice
+    if args.task_type:
+        payload["task_type"] = args.task_type
+    if args.dialect:
+        payload["language"] = args.dialect
+
+    instructions = build_instruction_payload(args)
+    if instructions:
+        payload["instructions"] = instructions
+
+    if args.ref_audio:
+        ref_audio = []
+        for audio in args.ref_audio:
+            if audio.startswith(("http://", "https://", "data:")):
+                ref_audio.append(audio)
+            else:
+                ref_audio.append(encode_audio_to_base64(audio))
+        payload["ref_audio"] = ref_audio[0] if len(ref_audio) == 1 else ref_audio
+    if args.ref_text:
+        payload["ref_text"] = args.ref_text
+    if args.speaker_embedding:
+        payload["speaker_embedding"] = load_speaker_embedding(args.speaker_embedding)
+    if args.max_new_tokens:
+        payload["max_new_tokens"] = args.max_new_tokens
+    if args.stream:
+        payload["stream"] = True
+        payload["response_format"] = "pcm"
+
+    api_url = f"{args.api_base}/v1/audio/speech"
+    headers = {
+        "Content-Type": "application/json",
+        "Authorization": f"Bearer {args.api_key}",
+    }
+
+    print(f"Model: {args.model}")
+    print(f"Text: {args.text}")
+    print(f"Payload keys: {sorted(payload)}")
+
+    if args.stream:
+        output_path = args.output or "ming_output.pcm"
+        with httpx.Client(timeout=300.0) as client:
+            with client.stream("POST", api_url, json=payload, headers=headers) as response:
+                if response.status_code != 200:
+                    print(f"Error: {response.status_code}")
+                    print(response.read().decode())
+                    return
+                with open(output_path, "wb") as f:
+                    for chunk in response.iter_bytes():
+                        f.write(chunk)
+        print(f"Streamed PCM audio to: {output_path}")
+        return
+
+    with httpx.Client(timeout=300.0) as client:
+        response = client.post(api_url, json=payload, headers=headers)
+
+    if response.status_code != 200:
+        print(f"Error: {response.status_code}")
+        print(response.text)
+        return
+
+    try:
+        text = response.content.decode("utf-8")
+        if text.startswith('{"error"'):
+            print(f"Error: {text}")
+            return
+    except UnicodeDecodeError:
+        pass
+
+    output_path = args.output or "ming_output.wav"
+    with open(output_path, "wb") as f:
+        f.write(response.content)
+    print(f"Audio saved to: {output_path}")
+
+
+def main():
+    parser = argparse.ArgumentParser(description="OpenAI-compatible client for Ming-omni-tts via /v1/audio/speech")
+    parser.add_argument("--api-base", default=DEFAULT_API_BASE, help="API base URL")
+    parser.add_argument("--api-key", default=DEFAULT_API_KEY, help="API key")
+    parser.add_argument("--model", "-m", default=DEFAULT_MODEL, help="Model name or path")
+    parser.add_argument("--text", required=True, help="Text to synthesize")
+    parser.add_argument(
+        "--task-type",
+        default=None,
+        choices=["CustomVoice", "VoiceDesign", "Base"],
+        help="Optional compatibility task type. Ming accepts the same field but primarily uses prompt metadata.",
+    )
+    parser.add_argument(
+        "--voice",
+        default=None,
+        help="Maps to Ming `IP` when using built-in character voices, or to an uploaded voice sample name",
+    )
+    parser.add_argument("--dialect", default=None, help="Maps to Ming `方言`")
+    parser.add_argument("--instructions", default=None, help="Free-form Ming instruction string")
+    parser.add_argument(
+        "--instruction-json",
+        default=None,
+        help='Structured Ming instruction JSON, for example \'{"情感":"高兴"}\'',
+    )
+    parser.add_argument(
+        "--ref-audio",
+        action="append",
+        default=None,
+        help="Reference audio path, URL, or data URL. Repeat for podcast-style multi-speaker prompts.",
+    )
+    parser.add_argument("--ref-text", default=None, help="Reference transcript for cloning")
+    parser.add_argument(
+        "--speaker-embedding", default=None, help="Path to a JSON file containing a 192-d speaker embedding"
+    )
+    parser.add_argument("--max-new-tokens", type=int, default=None, help="Override ming_max_decode_steps")
+    parser.add_argument("--stream", action="store_true", help="Enable streaming PCM output")
+    parser.add_argument(
+        "--response-format",
+        default="wav",
+        choices=["wav", "mp3", "flac", "pcm", "aac", "opus"],
+        help="Audio format when not streaming",
+    )
+    parser.add_argument("--output", "-o", default=None, help="Output file path")
+    args = parser.parse_args()
+    try:
+        run_tts(args)
+    except (FileNotFoundError, ValueError, json.JSONDecodeError) as exc:
+        raise SystemExit(f"Error: {exc}") from exc
+
+
+if __name__ == "__main__":
+    main()

examples/online_serving/text_to_speech/ming_tts/run_curl.sh

@@ -0,0 +1,76 @@
+#!/bin/bash
+set -euo pipefail
+
+MODE="${1:-basic}"
+HOST="${HOST:-localhost}"
+PORT="${PORT:-8091}"
+MODEL="${MODEL:-inclusionAI/Ming-omni-tts-0.5B}"
+API_URL="http://${HOST}:${PORT}/v1/audio/speech"
+TEXT="${TEXT:-你好，这是 Ming 在线语音合成测试。}"
+OUTPUT="${OUTPUT:-ming_output.wav}"
+STREAM_OUTPUT="${STREAM_OUTPUT:-ming_output.pcm}"
+REF_AUDIO="${REF_AUDIO:-}"
+REF_TEXT="${REF_TEXT:-}"
+
+post_json() {
+    local payload="$1"
+    local output_path="$2"
+    curl -X POST "$API_URL" \
+        -H "Content-Type: application/json" \
+        -H "Authorization: Bearer EMPTY" \
+        -d "$payload" \
+        --output "$output_path"
+}
+
+case "$MODE" in
+    basic)
+        post_json "{
+            \"model\": \"${MODEL}\",
+            \"input\": \"${TEXT}\",
+            \"response_format\": \"wav\"
+        }" "$OUTPUT"
+        ;;
+    zero_shot)
+        if [ -z "$REF_AUDIO" ] || [ -z "$REF_TEXT" ]; then
+            echo "zero_shot requires REF_AUDIO and REF_TEXT" >&2
+            exit 1
+        fi
+        python - <<'PY' > /tmp/ming_zero_shot_payload.json
+import base64
+import json
+import mimetypes
+import os
+from pathlib import Path
+
+path = Path(os.environ["REF_AUDIO"])
+mime_type = mimetypes.guess_type(path.name)[0] or "audio/wav"
+payload = {
+    "model": os.environ["MODEL"],
+    "input": os.environ["TEXT"],
+    "ref_audio": f"data:{mime_type};base64,{base64.b64encode(path.read_bytes()).decode('utf-8')}",
+    "ref_text": os.environ["REF_TEXT"],
+    "response_format": "wav",
+}
+print(json.dumps(payload, ensure_ascii=False))
+PY
+        curl -X POST "$API_URL" \
+            -H "Content-Type: application/json" \
+            -H "Authorization: Bearer EMPTY" \
+            --data-binary @/tmp/ming_zero_shot_payload.json \
+            --output "$OUTPUT"
+        rm -f /tmp/ming_zero_shot_payload.json
+        ;;
+    stream)
+        post_json "{
+            \"model\": \"${MODEL}\",
+            \"input\": \"${TEXT}\",
+            \"stream\": true,
+            \"response_format\": \"pcm\"
+        }" "$STREAM_OUTPUT"
+        ;;
+    *)
+        echo "Unknown mode: $MODE" >&2
+        echo "Supported sanity checks: basic, zero_shot, stream" >&2
+        exit 1
+        ;;
+esac

examples/online_serving/text_to_speech/ming_tts/run_server.sh

@@ -0,0 +1,25 @@
+#!/bin/bash
+# Launch vLLM-Omni server for Ming-omni-tts.
+#
+# Usage:
+#   ./run_server.sh
+#   PORT=8000 ./run_server.sh
+
+set -e
+
+DIR="$(cd "$(dirname "$0")" && pwd)"
+ROOT="$(cd "$DIR/../../../.." && pwd)"
+
+MODEL="${MODEL:-inclusionAI/Ming-omni-tts-0.5B}"
+PORT="${PORT:-8091}"
+DEPLOY_CONFIG="${DEPLOY_CONFIG:-$ROOT/vllm_omni/deploy/ming_tts.yaml}"
+
+echo "Starting Ming-omni-tts server with model: $MODEL"
+echo "Deploy config: $DEPLOY_CONFIG"
+
+vllm-omni serve "$MODEL" \
+    --deploy-config "$DEPLOY_CONFIG" \
+    --host 0.0.0.0 \
+    --port "$PORT" \
+    --enforce-eager \
+    --omni

examples/online_serving/text_to_speech/qwen3_tts/gradio_demo.py

@@ -408,7 +408,10 @@ def create_app(api_base: str):
                 await resp.aclose()
                 await client.aclose()
 
-        return StreamingResponse(relay(), media_type="application/octet-stream")
+        # Use audio/pcm to bypass browsers (e.g. Chromium) MIME type sniffing, which
+        # buffers application/octet-stream responses before delivering chunks to
+        # fetch().body.getReader(), inflating TTFP to total_ms.
+        return StreamingResponse(relay(), media_type="audio/pcm")
 
     # ── Gradio UI ────────────────────────────────────────────────
     voices = fetch_voices(api_base)

examples/quantization/check_modelopt_fp8_export.py

@@ -0,0 +1,247 @@
+#!/usr/bin/env python3
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""Verify a ModelOpt FP8 diffusers checkpoint exported by
+quantize_wan2_2_modelopt_fp8.py / quantize_hunyuanvideo_15_modelopt_fp8.py.
+
+Three checks:
+  A. transformer/config.json has a sane quantization_config block.
+  B. transformer/*.safetensors contains FP8 (float8_e4m3fn) quantized tensors.
+  C. transformer disk size is materially smaller than a BF16 baseline.
+
+Example:
+    python examples/quantization/check_modelopt_fp8_export.py \\
+        --output ./hv15-480p-modelopt-fp8
+
+    # Optional: compare disk size against a local or HF BF16 baseline.
+    python examples/quantization/check_modelopt_fp8_export.py \\
+        --output ./hv15-480p-modelopt-fp8 \\
+        --baseline hunyuanvideo-community/HunyuanVideo-1.5-Diffusers-480p_t2v
+"""
+
+from __future__ import annotations
+
+import argparse
+import json
+import sys
+from collections import Counter
+from pathlib import Path
+
+SUPPORTED_ALGOS = ("FP8",)
+
+
+def _check_config(transformer_dir: Path) -> tuple[int, str | None]:
+    """Returns (status, quant_algo). status: 0 pass, 1 fail, 2 warn."""
+    cfg_path = transformer_dir / "config.json"
+    if not cfg_path.exists():
+        print(f"[FAIL] {cfg_path} missing.")
+        return 1, None
+
+    with cfg_path.open(encoding="utf-8") as f:
+        cfg = json.load(f)
+
+    qc = cfg.get("quantization_config")
+    if not isinstance(qc, dict):
+        print(f"[FAIL] No `quantization_config` block in {cfg_path}.")
+        return 1, None
+
+    print(f"[A] quantization_config from {cfg_path}:")
+    print(json.dumps(qc, indent=2))
+
+    quant_algo = qc.get("quant_algo")
+    issues = []
+    if qc.get("quant_method") != "modelopt":
+        issues.append(f"quant_method={qc.get('quant_method')!r} (expected 'modelopt')")
+    if quant_algo not in SUPPORTED_ALGOS:
+        issues.append(
+            f"quant_algo={quant_algo!r} (expected one of {SUPPORTED_ALGOS} — vllm-omni adapter may not auto-detect)"
+        )
+
+    if issues:
+        print("[A] WARN — config looks incomplete:")
+        for issue in issues:
+            print(f"    - {issue}")
+        return 2, quant_algo
+    print(f"[A] PASS — config looks correct (quant_algo={quant_algo}).")
+    return 0, quant_algo
+
+
+def _read_safetensors_header(path: Path) -> dict:
+    """Read the JSON header of a safetensors file. Bypass-safe — doesn't materialize tensors.
+
+    Returns {tensor_name: {'dtype': 'F8_E4M3', 'shape': [...], 'data_offsets': [...]}}.
+    Header dtype strings: F8_E4M3, F8_E5M2, BF16, F16, F32, F64, I8, I16, I32, I64, BOOL, U8, ...
+    """
+    import struct
+
+    with open(path, "rb") as f:
+        header_len = struct.unpack("<Q", f.read(8))[0]
+        header = json.loads(f.read(header_len))
+    header.pop("__metadata__", None)
+    return header
+
+
+def _classify_weight_scale_granularity(weight_scale_shapes: list[list[int]]) -> str:
+    """Infer per-tensor vs per-channel vs per-block from sample weight_scale shapes.
+
+    ModelOpt block-wise produces shapes like `[16, 1, 16, 1]` (broadcasting dims of 1
+    interleaved with block-count dims). We count "meaningful" dims — ones with size > 1 —
+    and classify: 0 meaningful dims = per-tensor (scalar), 1 = per-channel, 2+ = per-block.
+    """
+    if not weight_scale_shapes:
+        return "no weight_scale tensors found"
+
+    def meaningful_dims(shape: list[int]) -> int:
+        return sum(1 for d in shape if d > 1)
+
+    per_tensor = sum(1 for s in weight_scale_shapes if meaningful_dims(s) == 0)
+    per_channel = sum(1 for s in weight_scale_shapes if meaningful_dims(s) == 1)
+    per_block = sum(1 for s in weight_scale_shapes if meaningful_dims(s) >= 2)
+    total = len(weight_scale_shapes)
+    if per_tensor == total:
+        return "per-tensor (all scalar scales)"
+    if per_channel == total:
+        return "per-channel (1 meaningful dim)"
+    if per_block == total:
+        return "per-block (2+ meaningful dims — e.g. [M//bm, 1, N//bn, 1] for tiles)"
+    return f"mixed: per-tensor={per_tensor}, per-channel={per_channel}, per-block={per_block} of {total}"
+
+
+def _check_safetensors(transformer_dir: Path) -> int:
+    """Returns 0 on pass, 1 on fail. Reads on-disk dtype from the safetensors header."""
+    files = sorted(transformer_dir.glob("*.safetensors"))
+    if not files:
+        print(f"[FAIL] No *.safetensors in {transformer_dir}.")
+        return 1
+
+    header_dtype_counts: Counter[str] = Counter()
+    sample_quant_weight_keys: list[str] = []
+    sample_scale_keys: list[str] = []
+    weight_scale_shapes: list[list[int]] = []
+    sample_weight_scale_entries: list[tuple[str, list[int]]] = []
+
+    for f in files:
+        try:
+            header = _read_safetensors_header(f)
+        except Exception as exc:
+            print(f"[B] WARN — could not parse header of {f}: {exc}")
+            continue
+        for k, info in header.items():
+            dtype = info.get("dtype", "?")
+            header_dtype_counts[dtype] += 1
+            # FP8 stores weights as F8_E4M3 directly.
+            if dtype.startswith("F8") and k.endswith(".weight") and len(sample_quant_weight_keys) < 5:
+                sample_quant_weight_keys.append(k)
+            if k.endswith(("_scale", ".weight_scale", ".input_scale", "_scale_inv")) and len(sample_scale_keys) < 5:
+                sample_scale_keys.append(k)
+            if k.endswith(".weight_scale"):
+                weight_scale_shapes.append(info.get("shape", []))
+                if len(sample_weight_scale_entries) < 5:
+                    sample_weight_scale_entries.append((k, info.get("shape", [])))
+
+    print(f"\n[B] On-disk dtype counts across {len(files)} safetensors file(s) (from header, not get_tensor):")
+    for dtype, count in sorted(header_dtype_counts.items(), key=lambda kv: -kv[1]):
+        marker = "  <-- FP8" if dtype.startswith("F8") else ""
+        print(f"    {dtype:10s} {count:>6d}{marker}")
+
+    quant_count = sum(c for d, c in header_dtype_counts.items() if d.startswith("F8"))
+    if quant_count == 0:
+        print("[B] FAIL — no FP8 tensors on disk. Calibration likely did not actually quantize the weights.")
+        return 1
+
+    print(f"[B] PASS — {quant_count} FP8 tensors stored on disk.")
+    if sample_quant_weight_keys:
+        print(f"    sample quantized weight tensors: {sample_quant_weight_keys[:3]}")
+    if sample_scale_keys:
+        print(f"    sample scale tensors:            {sample_scale_keys[:3]}")
+    print("    (Note: torch's get_tensor() may return these as bf16 views on some versions —")
+    print("     irrelevant; vLLM's loader uses native FP8 ops.)")
+
+    # Weight-scale granularity — per-tensor (scalar) vs per-channel (1-D) vs per-block (N-D).
+    print(f"\n    weight_scale granularity: {_classify_weight_scale_granularity(weight_scale_shapes)}")
+    for key, shape in sample_weight_scale_entries[:3]:
+        print(f"      {key}: shape {shape}")
+    return 0
+
+
+def _disk_size_gib(p: Path) -> float:
+    return sum(f.stat().st_size for f in p.rglob("*") if f.is_file()) / (1024**3)
+
+
+def _check_size_vs_baseline(transformer_dir: Path, baseline: str | None) -> int:
+    """Returns 0 always (informational only)."""
+    quant_size = _disk_size_gib(transformer_dir)
+    print(f"\n[C] FP8 transformer disk size: {quant_size:.2f} GiB")
+
+    if baseline is None:
+        print("[C] SKIP — pass --baseline <path or HF id> to compare against BF16.")
+        return 0
+
+    baseline_path = Path(baseline)
+    if not baseline_path.exists():
+        # Try HF download.
+        try:
+            from huggingface_hub import snapshot_download
+        except ImportError:
+            print("[C] SKIP — huggingface_hub not installed and baseline not a local path.")
+            return 0
+        print(f"    Downloading baseline transformer from HF: {baseline}")
+        baseline_path = Path(snapshot_download(baseline, allow_patterns=["transformer/*"]))
+
+    bf16_dir = baseline_path / "transformer" if (baseline_path / "transformer").exists() else baseline_path
+    bf16_size = _disk_size_gib(bf16_dir)
+    if bf16_size == 0:
+        print(f"[C] WARN — baseline transformer dir empty: {bf16_dir}")
+        return 0
+
+    # Expected reduction for FP8: ~50%.
+    min_reduction = 30
+    reduction = (1 - quant_size / bf16_size) * 100
+    print(f"[C] BF16 baseline transformer disk size: {bf16_size:.2f} GiB ({bf16_dir})")
+    print(f"[C] Disk reduction: {reduction:.1f}%  (FP8 transformer is {quant_size / bf16_size:.0%} of BF16)")
+    if reduction < min_reduction:
+        print(
+            f"[C] WARN — FP8 should typically reduce disk by ~40-50%; <{min_reduction}% suggests partial quantization."
+        )
+    return 0
+
+
+def main() -> None:
+    p = argparse.ArgumentParser(description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter)
+    p.add_argument("--output", required=True, help="Path to the exported ModelOpt FP8 checkpoint root.")
+    p.add_argument(
+        "--baseline",
+        default=None,
+        help="Optional BF16 baseline (local diffusers dir or HF id) for disk-size comparison.",
+    )
+    args = p.parse_args()
+
+    out_root = Path(args.output).expanduser().resolve()
+    transformer_dir = out_root / "transformer"
+    if not transformer_dir.exists():
+        print(f"[FAIL] {transformer_dir} does not exist.")
+        sys.exit(1)
+
+    print(f"Checking: {out_root}\n")
+
+    fail = 0
+    config_status, _quant_algo = _check_config(transformer_dir)
+    fail |= config_status
+    fail |= _check_safetensors(transformer_dir)
+    _check_size_vs_baseline(transformer_dir, args.baseline)
+
+    print()
+    if fail == 0:
+        print("=" * 60)
+        print("ALL CHECKS PASSED — checkpoint looks ready for vllm-omni serving.")
+    elif fail == 1:
+        print("=" * 60)
+        print("FAILURES detected — calibration may need to be re-run.")
+        sys.exit(1)
+    else:
+        print("=" * 60)
+        print("WARNINGS only — checkpoint may serve but with caveats. See [A] above.")
+
+
+if __name__ == "__main__":
+    main()

examples/quantization/quantize_hunyuanvideo_15_modelopt_fp8.py

@@ -0,0 +1,448 @@
+#!/usr/bin/env python3
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""Quantize HunyuanVideo-1.5 (480p T2V) to a ModelOpt FP8 Hugging Face checkpoint.
+
+Calibrates the DiT transformer using a small video prompt set and exports a
+diffusers-style directory whose transformer carries ModelOpt FP8 metadata.
+The exported checkpoint is consumable by vllm-omni's ModelOpt FP8 adapter
+(see vllm_omni/diffusion/model_loader/checkpoint_adapters/modelopt.py).
+
+Layers kept full precision match the #2728 / #2795 pattern: modulation,
+AdaLayerNorm, entry/exit projections, embeddings, the token refiner path,
+and final proj_out. MHA quantizers are off by default; HV-1.5 self-attention
+empirically degrades under FP8 (see #2920 ablation).
+
+Example:
+    python examples/quantization/quantize_hunyuanvideo_15_modelopt_fp8.py \\
+        --model hunyuanvideo-community/HunyuanVideo-1.5-Diffusers-480p_t2v \\
+        --output ./hv15-480p-modelopt-fp8 \\
+        --overwrite
+"""
+
+from __future__ import annotations
+
+import argparse
+import copy
+import json
+import re
+import shutil
+import sys
+from pathlib import Path
+from typing import Any
+
+import torch
+from diffusers import DiffusionPipeline
+
+DEFAULT_PROMPTS = [
+    "A dog running across a field of golden wheat.",
+    "An astronaut riding a horse across the surface of Mars, red dust swirling, cinematic wide shot.",
+    "A hummingbird hovering in front of a vibrant red flower, slow motion, macro shot.",
+    "A crackling campfire at night under a starry sky, sparks rising into the dark.",
+    "An underwater shot of a coral reef with tropical fish swimming by, sun rays piercing the water.",
+    "A close-up of a blooming rose covered in morning dew, soft natural light.",
+    "A peaceful mountain village at dawn, mist rolling over the rooftops, cinematic establishing shot.",
+    "A skateboarder doing a kickflip in an urban plaza, slow motion, golden hour lighting.",
+]
+
+
+def _build_parser() -> argparse.ArgumentParser:
+    p = argparse.ArgumentParser(description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter)
+    p.add_argument("--model", required=True, help="Input HV-1.5 diffusers directory or HF id.")
+    p.add_argument("--output", required=True, help="Output directory for the ModelOpt FP8 checkpoint.")
+    p.add_argument("--dtype", choices=("bfloat16", "float16"), default="bfloat16")
+    p.add_argument("--height", type=int, default=480)
+    p.add_argument("--width", type=int, default=832)
+    p.add_argument(
+        "--num-frames",
+        type=int,
+        default=33,
+        help="Frames per calibration sample. 33 matches the typical short benchmark.",
+    )
+    p.add_argument("--guidance-scale", type=float, default=6.0)
+    p.add_argument(
+        "--calib-steps",
+        type=int,
+        default=10,
+        help="Denoising steps per calibration prompt (10 is enough for amax statistics).",
+    )
+    p.add_argument("--calib-size", type=int, default=8, help="How many prompts to use for calibration.")
+    p.add_argument("--seed", type=int, default=42)
+    p.add_argument(
+        "--prompt",
+        action="append",
+        default=[],
+        help="Custom calibration prompt. Repeat to provide multiple.",
+    )
+    p.add_argument(
+        "--quantize-mha",
+        action="store_true",
+        help="Enable FP8 attention K/V/softmax quantizers. Off by default — empirically degrades HV-1.5 video output.",
+    )
+    p.add_argument(
+        "--weight-block-size",
+        type=str,
+        default=None,
+        help="Per-block weight quantization as 'M,N' (e.g. '128,128' for 128x128 tiles). "
+        "Default: per-tensor (one scale per linear). Block-wise typically gives tighter quality at "
+        "negligible memory cost. Static FP8 is exempt from upstream vLLM's online block-wise gate.",
+    )
+    p.add_argument("--overwrite", action="store_true", help="Replace an existing output directory.")
+    return p
+
+
+def _parse_block_size(spec: str | None) -> list[int] | None:
+    if spec is None:
+        return None
+    parts = [int(x) for x in spec.split(",") if x.strip()]
+    if len(parts) != 2:
+        raise SystemExit(f"--weight-block-size must be 'M,N' (2 ints), got {spec!r}")
+    return parts
+
+
+def _require_modelopt() -> Any:
+    try:
+        import modelopt.torch.quantization as mtq
+    except ModuleNotFoundError as exc:
+        raise SystemExit(
+            "NVIDIA ModelOpt is not installed. Install with:\n"
+            "  pip install 'nvidia-modelopt[all]'\n"
+            f"Original error: {exc}"
+        ) from exc
+    return mtq
+
+
+def _ensure_paths(args: argparse.Namespace) -> tuple[str, Path]:
+    model_path = args.model
+    output_dir = Path(args.output).expanduser().resolve()
+    if output_dir.exists():
+        if not args.overwrite:
+            raise SystemExit(f"Output directory already exists: {output_dir}\nPass --overwrite to replace it.")
+        shutil.rmtree(output_dir)
+    return model_path, output_dir
+
+
+def _select_dtype(name: str) -> torch.dtype:
+    return {"bfloat16": torch.bfloat16, "float16": torch.float16}[name]
+
+
+def _build_prompts(args: argparse.Namespace) -> list[str]:
+    prompts = args.prompt or DEFAULT_PROMPTS
+    if args.calib_size <= 0:
+        raise SystemExit("--calib-size must be positive.")
+    if len(prompts) < args.calib_size:
+        repeats = (args.calib_size + len(prompts) - 1) // len(prompts)
+        prompts = (prompts * repeats)[: args.calib_size]
+    return prompts[: args.calib_size]
+
+
+# Layers to KEEP at full precision (mirror of the #2920 wiring + #2728/#2795 skip pattern).
+# - x_embedder, image_embedder, context_embedder*, time_embed*, cond_type_embed: entry/embedding
+# - norm_out, norm1*.linear, norm1_context*.linear, norm2*, norm2_context*: AdaLayerNorm modulation
+# - proj_out: final output projection
+# - token_refiner*: text-encoder refinement uses diffusers raw nn.Linear
+def _filter_func_hv15(name: str) -> bool:
+    pattern = re.compile(
+        r"(proj_out.*|"
+        r".*(x_embedder|image_embedder|context_embedder|context_embedder_2|"
+        r"time_embed|cond_type_embed|"
+        r"norm_out|norm1\.linear|norm1_context\.linear|norm2|norm2_context|"
+        r"token_refiner).*)"
+    )
+    return pattern.match(name) is not None
+
+
+def _mha_filter_func(name: str) -> bool:
+    pattern = re.compile(
+        r".*(q_bmm_quantizer|k_bmm_quantizer|v_bmm_quantizer|softmax_quantizer|bmm2_output_quantizer).*"
+    )
+    return pattern.match(name) is not None
+
+
+def _disable_known_problematic_quantizers(mtq: Any, backbone: torch.nn.Module, *, quantize_mha: bool) -> None:
+    if not hasattr(mtq, "disable_quantizer"):
+        return
+    mtq.disable_quantizer(backbone, _filter_func_hv15)
+    if not quantize_mha:
+        mtq.disable_quantizer(backbone, _mha_filter_func)
+
+
+def _load_pipeline(model_path: str, dtype: torch.dtype) -> DiffusionPipeline:
+    pipe = DiffusionPipeline.from_pretrained(model_path, torch_dtype=dtype)
+    if hasattr(pipe, "set_progress_bar_config"):
+        pipe.set_progress_bar_config(disable=True)
+    pipe.to("cuda")
+    return pipe
+
+
+def _build_forward_loop(pipe: DiffusionPipeline, args: argparse.Namespace, prompts: list[str]):
+    generator = torch.Generator(device="cuda")
+
+    # Try to set guidance on the pipeline's guider object up front (modern
+    # diffusers HV-1.5 uses a Guider abstraction, not a per-call kwarg). Falls
+    # back silently — calibration uses whatever default the pipeline ships with.
+    guider = getattr(pipe, "guider", None)
+    if guider is not None and hasattr(guider, "guidance_scale"):
+        try:
+            guider.guidance_scale = args.guidance_scale
+        except Exception:
+            pass
+
+    base_kwargs = dict(
+        height=args.height,
+        width=args.width,
+        num_frames=args.num_frames,
+        num_inference_steps=args.calib_steps,
+        output_type="latent",
+    )
+
+    def forward_loop(*_unused_args, **_unused_kwargs) -> None:
+        with torch.inference_mode():
+            for idx, prompt in enumerate(prompts):
+                generator.manual_seed(args.seed + idx)
+                # Try with guidance_scale first; fall back without on TypeError
+                # for pipelines (like HV-1.5) that take CFG via guider config.
+                try:
+                    pipe(prompt=prompt, generator=generator, guidance_scale=args.guidance_scale, **base_kwargs)
+                except TypeError as exc:
+                    if "guidance_scale" not in str(exc):
+                        raise
+                    pipe(prompt=prompt, generator=generator, **base_kwargs)
+
+    return forward_loop
+
+
+def _summarize_export(output_dir: Path) -> None:
+    cfg_path = output_dir / "transformer" / "config.json"
+    if not cfg_path.exists():
+        print(f"[warn] {cfg_path} missing.", file=sys.stderr)
+        return
+    with cfg_path.open(encoding="utf-8") as f:
+        cfg = json.load(f)
+    qc = cfg.get("quantization_config")
+    if not isinstance(qc, dict):
+        print("[warn] No quantization_config in transformer/config.json.", file=sys.stderr)
+        return
+    print("Export summary:")
+    print(f"  quant_method: {qc.get('quant_method')}")
+    print(f"  quant_algo:   {qc.get('quant_algo')}")
+    producer = qc.get("producer")
+    if isinstance(producer, dict):
+        print(f"  producer:     {producer.get('name')} {producer.get('version')}")
+    print(f"  config path:  {cfg_path}")
+
+
+def _force_export_quantized_weights(backbone: torch.nn.Module, dtype: torch.dtype) -> int:
+    """Convert in-memory weights of quantized modules to actual FP8 storage.
+
+    `export_hf_checkpoint` skips this step for unknown model types (HV-1.5 isn't
+    in ModelOpt's recognized-model registry), so we must call the per-weight
+    export helper ourselves. Same workaround as the HunyuanImage-3 calibration
+    helper.
+    """
+    from modelopt.torch.export.quant_utils import (
+        QUANTIZATION_NONE,
+        get_quantization_format,
+        quantizer_attr_names,
+        weight_attr_names,
+    )
+    from modelopt.torch.export.unified_export_hf import _export_quantized_weight
+
+    exported = 0
+    for name, module in backbone.named_modules():
+        try:
+            quantization_format = get_quantization_format(module)
+        except Exception as exc:
+            print(f"[warn] Could not inspect quantization format for {name}: {exc}", file=sys.stderr)
+            continue
+        if quantization_format == QUANTIZATION_NONE:
+            continue
+        for weight_name in weight_attr_names(module):
+            quantizer_attrs = quantizer_attr_names(weight_name)
+            weight_quantizer = getattr(module, quantizer_attrs.weight_quantizer, None)
+            if weight_quantizer is None or not getattr(weight_quantizer, "is_enabled", False):
+                continue
+            _export_quantized_weight(module, dtype, weight_name)
+            exported += 1
+    return exported
+
+
+def _hv15_quant_config_block(weight_block_size: list[int] | None = None) -> dict:
+    """Mirror ModelOpt FP8 metadata expected by vllm-omni's adapter (#2913).
+
+    Same shape as the HunyuanImage-3 author's _hunyuan_quant_config(). When
+    `weight_block_size` is given, advertise block-wise weight quantization in
+    the saved metadata (so consumers know to expect multi-element scale tensors).
+    """
+    weights_cfg: dict = {"dynamic": False, "num_bits": 8, "type": "float"}
+    if weight_block_size is not None:
+        weights_cfg["strategy"] = "block"
+        weights_cfg["block_structure"] = f"{weight_block_size[0]}x{weight_block_size[1]}"
+    return {
+        "config_groups": {
+            "group_0": {
+                "input_activations": {"dynamic": False, "num_bits": 8, "type": "float"},
+                "weights": weights_cfg,
+                "targets": ["Linear"],
+            }
+        },
+        "ignore": [
+            "context_embedder*",
+            "context_embedder_2*",
+            "cond_type_embed*",
+            "image_embedder*",
+            "norm1.linear*",
+            "norm1_context.linear*",
+            "norm2*",
+            "norm2_context*",
+            "norm_out*",
+            "proj_out*",
+            "time_embed*",
+            "token_refiner*",
+            "x_embedder*",
+        ],
+        "producer": {"name": "modelopt"},
+        "quant_algo": "FP8",
+        "quant_method": "modelopt",
+    }
+
+
+def _patch_quant_config(output_dir: Path, weight_block_size: list[int] | None = None) -> None:
+    """Inject quant_algo: FP8 + config_groups into transformer/config.json so
+    vllm-omni's adapter (#2913) recognises the checkpoint as ModelOpt FP8."""
+    cfg_path = output_dir / "transformer" / "config.json"
+    with cfg_path.open(encoding="utf-8") as f:
+        cfg = json.load(f)
+
+    new_qc = _hv15_quant_config_block(weight_block_size=weight_block_size)
+    existing = cfg.get("quantization_config")
+    if isinstance(existing, dict):
+        producer = existing.get("producer")
+        if isinstance(producer, dict):
+            new_qc["producer"] = producer
+
+    cfg["quantization_config"] = new_qc
+    with cfg_path.open("w", encoding="utf-8") as f:
+        json.dump(cfg, f, indent=2)
+
+
+def _save_pipeline_with_fp8_transformer(
+    pipe: DiffusionPipeline,
+    model_path: str,
+    output_dir: Path,
+    max_shard_size: str = "5GB",
+) -> None:
+    """Save the pipeline with the (now FP8) transformer.
+
+    Copies the source directory verbatim except for `transformer/`, then
+    saves the transformer with quantizers hidden so the state dict contains
+    only the FP8 weights + scale tensors.
+    """
+    from modelopt.torch.export.diffusers_utils import hide_quantizers_from_state_dict
+
+    src = Path(model_path)
+    if not src.exists():
+        from huggingface_hub import snapshot_download
+
+        src = Path(snapshot_download(model_path))
+
+    if output_dir.exists():
+        shutil.rmtree(output_dir)
+    shutil.copytree(src, output_dir, ignore=shutil.ignore_patterns("transformer"))
+
+    transformer_out = output_dir / "transformer"
+    # `hide_quantizers_from_state_dict` walks named_modules(); pass the actual
+    # nn.Module (transformer), not the diffusers Pipeline wrapper.
+    with hide_quantizers_from_state_dict(pipe.transformer):
+        pipe.transformer.save_pretrained(
+            str(transformer_out),
+            safe_serialization=True,
+            max_shard_size=max_shard_size,
+        )
+
+
+def main() -> None:
+    args = _build_parser().parse_args()
+    if not torch.cuda.is_available():
+        raise SystemExit("CUDA is required for ModelOpt FP8 quantization.")
+
+    mtq = _require_modelopt()
+    model_path, output_dir = _ensure_paths(args)
+    dtype = _select_dtype(args.dtype)
+    prompts = _build_prompts(args)
+
+    print("Quantization plan:")
+    weight_block_size = _parse_block_size(args.weight_block_size)
+
+    print(f"  input:           {args.model}")
+    print(f"  output:          {output_dir}")
+    print(f"  dtype:           {dtype}")
+    print(f"  height/width:    {args.height}x{args.width}")
+    print(f"  num_frames:      {args.num_frames}")
+    print(f"  calib_size:      {len(prompts)}")
+    print(f"  calib_steps:     {args.calib_steps}")
+    print(f"  quantize_mha:    {args.quantize_mha}")
+    print(
+        f"  weight strategy: {'block-wise ' + str(weight_block_size) if weight_block_size else 'per-tensor (default)'}"
+    )
+
+    pipe = _load_pipeline(model_path, dtype)
+    backbone = pipe.transformer
+
+    quant_config = copy.deepcopy(mtq.FP8_DEFAULT_CFG)
+    if weight_block_size is not None:
+        # Switch from per-tensor (default) to block-wise weight quantization.
+        # ModelOpt's wildcard "*weight_quantizer" matches every linear's weight quantizer.
+        quant_config["quant_cfg"]["*weight_quantizer"] = {
+            "num_bits": (4, 3),  # E4M3 (FP8 weights, same as default)
+            "block_sizes": {-1: weight_block_size[1], -2: weight_block_size[0]},
+        }
+        print(
+            f"  -> overriding weight quantizer with block_sizes={weight_block_size} "
+            f"({weight_block_size[0]}x{weight_block_size[1]} tiles)"
+        )
+
+    forward_loop = _build_forward_loop(pipe, args, prompts)
+    quantized = mtq.quantize(backbone, quant_config, forward_loop)
+    if quantized is not None:
+        pipe.transformer = quantized
+        backbone = quantized
+
+    _disable_known_problematic_quantizers(mtq, backbone, quantize_mha=args.quantize_mha)
+
+    print("\nForcing FP8 weight serialization (HV-1.5 isn't in ModelOpt's recognized-model registry,")
+    print("so we have to call the per-weight export helper ourselves)...")
+    exported = _force_export_quantized_weights(backbone, dtype)
+    print(f"  -> {exported} weights converted to FP8 in memory")
+    if exported == 0:
+        raise SystemExit(
+            "No quantized weights were exported. Calibration may have skipped every layer "
+            "(check the disable_quantizer regex) or `mtq.quantize` did not actually wrap any "
+            "weight quantizers."
+        )
+
+    print("\nSaving pipeline with FP8 transformer...")
+    _save_pipeline_with_fp8_transformer(pipe, model_path, output_dir)
+    _patch_quant_config(output_dir, weight_block_size=weight_block_size)
+    print(f"Saved to: {output_dir}")
+    _summarize_export(output_dir)
+
+    print("\nNext: validate the checkpoint with vllm-omni:")
+    print(
+        "  python examples/offline_inference/text_to_video/text_to_video.py \\\n"
+        f"    --model {output_dir} \\\n"
+        "    --quantization fp8 \\\n"
+        "    --prompt 'A dog running across a field of golden wheat.' \\\n"
+        f"    --height {args.height} --width {args.width} --num-frames {args.num_frames} \\\n"
+        "    --num-inference-steps 30 --guidance-scale 6.0 --seed 42 \\\n"
+        "    --output outputs/hv15_modelopt_fp8.mp4 \\\n"
+        "    --enforce-eager"
+    )
+    print(
+        "\n  (--quantization fp8 is auto-upgraded to ModelOpt FP8 at runtime because the "
+        "checkpoint's config.json has modelopt metadata.)"
+    )
+
+
+if __name__ == "__main__":
+    main()

examples/quantization/quantize_wan2_2_modelopt_fp8.py

@@ -0,0 +1,471 @@
+#!/usr/bin/env python3
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""Quantize Wan2.2 (TI2V-5B, 704x1280 T2V) to a ModelOpt FP8 Hugging Face checkpoint.
+
+Calibrates the DiT transformer using a small video prompt set and exports a
+diffusers-style directory whose transformer carries ModelOpt FP8 metadata.
+The exported checkpoint is consumable by vllm-omni's ModelOpt FP8 adapter
+(see vllm_omni/diffusion/model_loader/checkpoint_adapters/modelopt.py).
+
+Layers kept full precision match the #2728 / #2795 pattern: condition embedder
+(time/text/image), patch embedding, modulation (scale_shift_table), final
+norm + proj_out, and sequence-parallel helpers. All attention + FFN linears
+are quantized — static calibration handles the numerics that online FP8
+couldn't (see #2920 ablation).
+
+Default target is `Wan-AI/Wan2.2-TI2V-5B-Diffusers`, the dense 5B variant that
+fits 80GB BF16. The A14B MoE variants need 2+ GPUs and are out of scope here.
+
+Example:
+    python examples/quantization/quantize_wan2_2_modelopt_fp8.py \\
+        --model Wan-AI/Wan2.2-TI2V-5B-Diffusers \\
+        --output ./wan22-ti2v-modelopt-fp8 \\
+        --overwrite
+"""
+
+from __future__ import annotations
+
+import argparse
+import copy
+import json
+import re
+import shutil
+import sys
+from pathlib import Path
+from typing import Any
+
+import torch
+from diffusers import DiffusionPipeline
+
+DEFAULT_PROMPTS = [
+    "A dog running across a field of golden wheat.",
+    "An astronaut riding a horse across the surface of Mars, red dust swirling, cinematic wide shot.",
+    "A hummingbird hovering in front of a vibrant red flower, slow motion, macro shot.",
+    "A crackling campfire at night under a starry sky, sparks rising into the dark.",
+    "An underwater shot of a coral reef with tropical fish swimming by, sun rays piercing the water.",
+    "A close-up of a blooming rose covered in morning dew, soft natural light.",
+    "A peaceful mountain village at dawn, mist rolling over the rooftops, cinematic establishing shot.",
+    "A skateboarder doing a kickflip in an urban plaza, slow motion, golden hour lighting.",
+]
+
+
+def _build_parser() -> argparse.ArgumentParser:
+    p = argparse.ArgumentParser(description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter)
+    p.add_argument("--model", required=True, help="Input Wan2.2 diffusers directory or HF id.")
+    p.add_argument("--output", required=True, help="Output directory for the ModelOpt FP8 checkpoint.")
+    p.add_argument("--dtype", choices=("bfloat16", "float16"), default="bfloat16")
+    p.add_argument("--height", type=int, default=704, help="Calibration video height (Wan2.2 TI2V-5B native: 704).")
+    p.add_argument("--width", type=int, default=1280, help="Calibration video width (Wan2.2 TI2V-5B native: 1280).")
+    p.add_argument(
+        "--num-frames",
+        type=int,
+        default=49,
+        help="Frames per calibration sample. 49 matches the typical short benchmark; "
+        "use 17 to reduce memory pressure during calibration.",
+    )
+    p.add_argument("--guidance-scale", type=float, default=5.0)
+    p.add_argument(
+        "--calib-steps",
+        type=int,
+        default=10,
+        help="Denoising steps per calibration prompt (10 is enough for amax statistics).",
+    )
+    p.add_argument("--calib-size", type=int, default=8, help="How many prompts to use for calibration.")
+    p.add_argument("--seed", type=int, default=42)
+    p.add_argument(
+        "--prompt",
+        action="append",
+        default=[],
+        help="Custom calibration prompt. Repeat to provide multiple.",
+    )
+    p.add_argument(
+        "--quantize-mha",
+        action="store_true",
+        help="Enable FP8 attention K/V/softmax quantizers. Off by default — Wan2.2's long attention "
+        "sequences amplified FP8 drift in the online ablation (see #2920).",
+    )
+    p.add_argument(
+        "--weight-block-size",
+        type=str,
+        default=None,
+        help="Per-block weight quantization as 'M,N' (e.g. '128,128'). Default per-tensor. "
+        "Note: vllm-omni's ModelOpt adapter may not yet dispatch block-wise scales — check #2924 "
+        "for the HV-1.5 investigation status before relying on this.",
+    )
+    p.add_argument("--overwrite", action="store_true", help="Replace an existing output directory.")
+    return p
+
+
+def _parse_block_size(spec: str | None) -> list[int] | None:
+    if spec is None:
+        return None
+    parts = [int(x) for x in spec.split(",") if x.strip()]
+    if len(parts) != 2:
+        raise SystemExit(f"--weight-block-size must be 'M,N' (2 ints), got {spec!r}")
+    return parts
+
+
+def _require_modelopt() -> Any:
+    try:
+        import modelopt.torch.quantization as mtq
+    except ModuleNotFoundError as exc:
+        raise SystemExit(
+            "NVIDIA ModelOpt is not installed. Install with:\n"
+            "  pip install 'nvidia-modelopt[all]'\n"
+            f"Original error: {exc}"
+        ) from exc
+    return mtq
+
+
+def _ensure_paths(args: argparse.Namespace) -> tuple[str, Path]:
+    model_path = args.model
+    output_dir = Path(args.output).expanduser().resolve()
+    if output_dir.exists():
+        if not args.overwrite:
+            raise SystemExit(f"Output directory already exists: {output_dir}\nPass --overwrite to replace it.")
+        shutil.rmtree(output_dir)
+    return model_path, output_dir
+
+
+def _select_dtype(name: str) -> torch.dtype:
+    return {"bfloat16": torch.bfloat16, "float16": torch.float16}[name]
+
+
+def _build_prompts(args: argparse.Namespace) -> list[str]:
+    prompts = args.prompt or DEFAULT_PROMPTS
+    if args.calib_size <= 0:
+        raise SystemExit("--calib-size must be positive.")
+    if len(prompts) < args.calib_size:
+        repeats = (args.calib_size + len(prompts) - 1) // len(prompts)
+        prompts = (prompts * repeats)[: args.calib_size]
+    return prompts[: args.calib_size]
+
+
+# Layers to KEEP at full precision. Wan2.2's module naming:
+# - condition_embedder: time_embedder, time_proj, text_embedder, image_embedder (I2V)
+# - patch_embedding: Conv3dLayer (already not Linear, belt-and-suspenders skip)
+# - scale_shift_table: nn.Parameter modulation (not Linear, but pattern guard)
+# - norm_out: AdaLayerNorm final
+# - proj_out: final nn.Linear
+# - timestep_proj_prepare / output_scale_shift_prepare: SP helpers
+def _filter_func_wan22(name: str) -> bool:
+    pattern = re.compile(
+        r"(proj_out.*|"
+        r".*(condition_embedder|patch_embedding|"
+        r"norm_out|scale_shift_table|"
+        r"timestep_proj_prepare|output_scale_shift_prepare).*)"
+    )
+    return pattern.match(name) is not None
+
+
+def _mha_filter_func(name: str) -> bool:
+    pattern = re.compile(
+        r".*(q_bmm_quantizer|k_bmm_quantizer|v_bmm_quantizer|softmax_quantizer|bmm2_output_quantizer).*"
+    )
+    return pattern.match(name) is not None
+
+
+def _disable_known_problematic_quantizers(mtq: Any, backbone: torch.nn.Module, *, quantize_mha: bool) -> None:
+    if not hasattr(mtq, "disable_quantizer"):
+        return
+    mtq.disable_quantizer(backbone, _filter_func_wan22)
+    if not quantize_mha:
+        mtq.disable_quantizer(backbone, _mha_filter_func)
+
+
+def _load_pipeline(model_path: str, dtype: torch.dtype) -> DiffusionPipeline:
+    pipe = DiffusionPipeline.from_pretrained(model_path, torch_dtype=dtype)
+    if hasattr(pipe, "set_progress_bar_config"):
+        pipe.set_progress_bar_config(disable=True)
+    pipe.to("cuda")
+    return pipe
+
+
+def _build_forward_loop(pipe: DiffusionPipeline, args: argparse.Namespace, prompts: list[str]):
+    generator = torch.Generator(device="cuda")
+
+    # Try setting guidance on the pipeline's guider if present (newer diffusers APIs).
+    guider = getattr(pipe, "guider", None)
+    if guider is not None and hasattr(guider, "guidance_scale"):
+        try:
+            guider.guidance_scale = args.guidance_scale
+        except Exception:
+            pass
+
+    base_kwargs = dict(
+        height=args.height,
+        width=args.width,
+        num_frames=args.num_frames,
+        num_inference_steps=args.calib_steps,
+        output_type="latent",
+    )
+
+    def forward_loop(*_unused_args, **_unused_kwargs) -> None:
+        with torch.inference_mode():
+            for idx, prompt in enumerate(prompts):
+                generator.manual_seed(args.seed + idx)
+                # Try with guidance_scale first; fall back without on TypeError
+                # for pipelines that take CFG via guider config only.
+                try:
+                    pipe(prompt=prompt, generator=generator, guidance_scale=args.guidance_scale, **base_kwargs)
+                except TypeError as exc:
+                    if "guidance_scale" not in str(exc):
+                        raise
+                    pipe(prompt=prompt, generator=generator, **base_kwargs)
+
+    return forward_loop
+
+
+def _summarize_export(output_dir: Path) -> None:
+    cfg_path = output_dir / "transformer" / "config.json"
+    if not cfg_path.exists():
+        print(f"[warn] {cfg_path} missing.", file=sys.stderr)
+        return
+    with cfg_path.open(encoding="utf-8") as f:
+        cfg = json.load(f)
+    qc = cfg.get("quantization_config")
+    if not isinstance(qc, dict):
+        print("[warn] No quantization_config in transformer/config.json.", file=sys.stderr)
+        return
+    print("Export summary:")
+    print(f"  quant_method: {qc.get('quant_method')}")
+    print(f"  quant_algo:   {qc.get('quant_algo')}")
+    producer = qc.get("producer")
+    if isinstance(producer, dict):
+        print(f"  producer:     {producer.get('name')} {producer.get('version')}")
+    print(f"  config path:  {cfg_path}")
+
+
+def _force_export_quantized_weights(backbone: torch.nn.Module, dtype: torch.dtype) -> int:
+    """Convert in-memory weights of quantized modules to actual FP8 storage.
+
+    `export_hf_checkpoint` skips this step for unknown model types (Wan2.2 isn't
+    in ModelOpt's recognized-model registry), so we must call the per-weight
+    export helper ourselves. Same workaround as the HunyuanVideo-1.5 / HunyuanImage-3
+    calibration helpers.
+    """
+    from modelopt.torch.export.quant_utils import (
+        QUANTIZATION_NONE,
+        get_quantization_format,
+        quantizer_attr_names,
+        weight_attr_names,
+    )
+    from modelopt.torch.export.unified_export_hf import _export_quantized_weight
+
+    exported = 0
+    for name, module in backbone.named_modules():
+        try:
+            quantization_format = get_quantization_format(module)
+        except Exception as exc:
+            print(f"[warn] Could not inspect quantization format for {name}: {exc}", file=sys.stderr)
+            continue
+        if quantization_format == QUANTIZATION_NONE:
+            continue
+        for weight_name in weight_attr_names(module):
+            quantizer_attrs = quantizer_attr_names(weight_name)
+            weight_quantizer = getattr(module, quantizer_attrs.weight_quantizer, None)
+            if weight_quantizer is None or not getattr(weight_quantizer, "is_enabled", False):
+                continue
+            _export_quantized_weight(module, dtype, weight_name)
+            exported += 1
+    return exported
+
+
+def _wan22_quant_config_block(weight_block_size: list[int] | None = None) -> dict:
+    """Mirror ModelOpt FP8 metadata expected by vllm-omni's adapter (#2913)."""
+    weights_cfg: dict = {"dynamic": False, "num_bits": 8, "type": "float"}
+    if weight_block_size is not None:
+        weights_cfg["strategy"] = "block"
+        weights_cfg["block_structure"] = f"{weight_block_size[0]}x{weight_block_size[1]}"
+    return {
+        "config_groups": {
+            "group_0": {
+                "input_activations": {"dynamic": False, "num_bits": 8, "type": "float"},
+                "weights": weights_cfg,
+                "targets": ["Linear"],
+            }
+        },
+        "ignore": [
+            "condition_embedder*",
+            "norm_out*",
+            "output_scale_shift_prepare*",
+            "patch_embedding*",
+            "proj_out*",
+            "scale_shift_table*",
+            "timestep_proj_prepare*",
+        ],
+        "producer": {"name": "modelopt"},
+        "quant_algo": "FP8",
+        "quant_method": "modelopt",
+    }
+
+
+def _patch_quant_config(output_dir: Path, weight_block_size: list[int] | None = None) -> None:
+    """Inject quant_algo: FP8 + config_groups into each transformer's config.json
+    so vllm-omni's adapter (#2913) recognises the checkpoint as ModelOpt FP8.
+
+    Patches both ``transformer`` and ``transformer_2`` (A14B) when present."""
+    for sub in ("transformer", "transformer_2"):
+        cfg_path = output_dir / sub / "config.json"
+        if not cfg_path.exists():
+            continue
+        with cfg_path.open(encoding="utf-8") as f:
+            cfg = json.load(f)
+
+        new_qc = _wan22_quant_config_block(weight_block_size=weight_block_size)
+        existing = cfg.get("quantization_config")
+        if isinstance(existing, dict):
+            producer = existing.get("producer")
+            if isinstance(producer, dict):
+                new_qc["producer"] = producer
+
+        cfg["quantization_config"] = new_qc
+        with cfg_path.open("w", encoding="utf-8") as f:
+            json.dump(cfg, f, indent=2)
+
+
+def _list_transformers(pipe: DiffusionPipeline) -> list[tuple[str, torch.nn.Module]]:
+    """Return [(attr_name, module)] for every DiT on the pipeline.
+
+    Wan2.2 A14B has both ``transformer`` (low-noise) and ``transformer_2``
+    (high-noise). TI2V-5B has only ``transformer``.
+    """
+    out = []
+    for attr in ("transformer", "transformer_2"):
+        mod = getattr(pipe, attr, None)
+        if mod is not None:
+            out.append((attr, mod))
+    return out
+
+
+def _save_pipeline_with_fp8_transformer(
+    pipe: DiffusionPipeline,
+    model_path: str,
+    output_dir: Path,
+    max_shard_size: str = "5GB",
+) -> None:
+    """Copy source dir verbatim minus transformer dirs, then save each quantized DiT.
+
+    Handles A14B's dual transformer (``transformer`` + ``transformer_2``) as well
+    as the single-DiT TI2V-5B.
+    """
+    from modelopt.torch.export.diffusers_utils import hide_quantizers_from_state_dict
+
+    src = Path(model_path)
+    if not src.exists():
+        from huggingface_hub import snapshot_download
+
+        src = Path(snapshot_download(model_path))
+
+    if output_dir.exists():
+        shutil.rmtree(output_dir)
+    shutil.copytree(src, output_dir, ignore=shutil.ignore_patterns("transformer", "transformer_2"))
+
+    for attr, backbone in _list_transformers(pipe):
+        transformer_out = output_dir / attr
+        # Pass the nn.Module (transformer), not the Pipeline wrapper.
+        with hide_quantizers_from_state_dict(backbone):
+            backbone.save_pretrained(
+                str(transformer_out),
+                safe_serialization=True,
+                max_shard_size=max_shard_size,
+            )
+
+
+def main() -> None:
+    args = _build_parser().parse_args()
+    if not torch.cuda.is_available():
+        raise SystemExit("CUDA is required for ModelOpt FP8 quantization.")
+
+    mtq = _require_modelopt()
+    model_path, output_dir = _ensure_paths(args)
+    dtype = _select_dtype(args.dtype)
+    prompts = _build_prompts(args)
+    weight_block_size = _parse_block_size(args.weight_block_size)
+
+    print("Quantization plan:")
+    print(f"  input:           {args.model}")
+    print(f"  output:          {output_dir}")
+    print(f"  dtype:           {dtype}")
+    print(f"  height/width:    {args.height}x{args.width}")
+    print(f"  num_frames:      {args.num_frames}")
+    print(f"  calib_size:      {len(prompts)}")
+    print(f"  calib_steps:     {args.calib_steps}")
+    print(f"  quantize_mha:    {args.quantize_mha}")
+    print(
+        f"  weight strategy: {'block-wise ' + str(weight_block_size) if weight_block_size else 'per-tensor (default)'}"
+    )
+
+    pipe = _load_pipeline(model_path, dtype)
+    transformers = _list_transformers(pipe)
+    if not transformers:
+        raise SystemExit("Pipeline has no transformer or transformer_2 attribute.")
+    print(f"  found {len(transformers)} transformer(s): {', '.join(a for a, _ in transformers)}")
+
+    quant_config = copy.deepcopy(mtq.FP8_DEFAULT_CFG)
+    if weight_block_size is not None:
+        quant_config["quant_cfg"]["*weight_quantizer"] = {
+            "num_bits": (4, 3),
+            "block_sizes": {-1: weight_block_size[1], -2: weight_block_size[0]},
+        }
+        print(
+            f"  -> overriding weight quantizer with block_sizes={weight_block_size} "
+            f"({weight_block_size[0]}x{weight_block_size[1]} tiles)"
+        )
+
+    forward_loop = _build_forward_loop(pipe, args, prompts)
+
+    # Quantize each DiT. The forward_loop runs the full pipeline, so quantizers
+    # on every attached transformer (A14B: transformer + transformer_2) get
+    # calibrated during the same passes.
+    for attr, backbone in transformers:
+        print(f"\nQuantizing {attr}...")
+        quantized = mtq.quantize(backbone, quant_config, forward_loop)
+        if quantized is not None:
+            setattr(pipe, attr, quantized)
+        _disable_known_problematic_quantizers(mtq, getattr(pipe, attr), quantize_mha=args.quantize_mha)
+
+    # Export weights to FP8 only AFTER all transformers are calibrated. Exporting
+    # inside the loop would convert transformer's weights to real FP8, and the
+    # next transformer's forward_loop (full pipeline) would then fake-quantize an
+    # already-FP8 weight -> "fake_e4m3fy not implemented for Float8_e4m3fn".
+    print("\nForcing FP8 weight serialization (Wan2.2 isn't in ModelOpt's recognized-model registry,")
+    print("so we have to call the per-weight export helper ourselves)...")
+    total_exported = 0
+    for attr, backbone in _list_transformers(pipe):
+        exported = _force_export_quantized_weights(backbone, dtype)
+        print(f"  -> {exported} weights converted to FP8 in {attr}")
+        total_exported += exported
+
+    if total_exported == 0:
+        raise SystemExit(
+            "No quantized weights were exported. Calibration may have skipped every layer "
+            "(check the disable_quantizer regex) or `mtq.quantize` did not actually wrap any "
+            "weight quantizers."
+        )
+
+    print("\nSaving pipeline with FP8 transformer(s)...")
+    _save_pipeline_with_fp8_transformer(pipe, model_path, output_dir)
+    _patch_quant_config(output_dir, weight_block_size=weight_block_size)
+    print(f"Saved to: {output_dir}")
+    _summarize_export(output_dir)
+
+    print("\nNext: validate the checkpoint with vllm-omni:")
+    print(
+        "  python examples/offline_inference/text_to_video/text_to_video.py \\\n"
+        f"    --model {output_dir} \\\n"
+        "    --quantization fp8 \\\n"
+        "    --prompt 'A dog running across a field of golden wheat.' \\\n"
+        f"    --height {args.height} --width {args.width} --num-frames {args.num_frames} \\\n"
+        "    --num-inference-steps 30 --guidance-scale 5.0 --seed 42 \\\n"
+        "    --output outputs/wan22_modelopt_fp8.mp4"
+    )
+    print(
+        "\n  (--quantization fp8 is auto-upgraded to ModelOpt FP8 at runtime because the "
+        "checkpoint's config.json has modelopt metadata.)"
+    )
+
+
+if __name__ == "__main__":
+    main()

pyproject.toml

@@ -66,7 +66,8 @@ dev = [
     "voxcpm>=2.0.2", # for VoxCPM2 path-locating in tests
     "torchdiffeq>=0.2.5", # for covo_audio Test
     "FlagEmbedding", # for hunyuan_image3 tests
-    "kernels<0.15.0" # for accuracy tests comparison with diffusers output (0.15.0 incompatible with diffusers 0.38.0)
+    "kernels<0.15.0", # for accuracy tests comparison with diffusers output (0.15.0 incompatible with diffusers 0.38.0)
+    "s3tokenizer>=0.3.0" # for tts test
 ]
 
 demo = [
@@ -90,6 +91,12 @@ minicpmo = [
     "stepaudio2-minicpmo",
 ]
 
+# Optional Blackwell-only fused-bias FP8 GEMM; package is `quack-kernels` (imported
+# as `quack`), not the unrelated `quack`. See docs/user_guide/quantization/fp8.md.
+quack = [
+    "quack-kernels>=0.3.11",
+]
+
 docs = [
     "mkdocs>=1.5.0",
     "mkdocs-api-autonav",
@@ -274,6 +281,7 @@ extend-ignore-identifiers-re = [
     ".*[Oo]no_[Aa]nna.*",
     ".*cann.*",
     ".*[Nn][Dd]3.*",
+    ".*updator.*",
     "byt5.*",
 ]
 extend-ignore-re = [
@@ -286,5 +294,9 @@ fullset = "fullset"
 Vai = "Vai"
 tockens = "tockens"
 CANN = "CANN"
+# Step-Audio reasoning parser streams `</think>` tokens split across chunks
+# (e.g. "</" + "thi" + "nk>"). The fragment "thi" appears literally in its
+# parser logic and tests and is not a typo of "the"/"this".
+thi = "thi"
 ND = "ND"
 byt5 = "byt5"

recipes/BosonAI/Higgs-Audio-V3-TTS.md

@@ -0,0 +1,131 @@
+# Higgs-Audio V3 TTS
+
+> Multilingual text-to-speech with voice cloning on 1×H100
+
+## Summary
+
+- Vendor: Boson AI
+- Model: `bosonai/higgs-audio-v3-tts-4b`
+- Task: Text-to-speech synthesis with optional voice cloning (100+ languages)
+- Mode: Online serving with the OpenAI-compatible `/v1/audio/speech` API; offline batch inference
+- Maintainer: @yuekaiz
+
+## When to use this recipe
+
+Use this recipe to serve `bosonai/higgs-audio-v3-tts-4b` for high-quality
+multilingual TTS. The model generates 24 kHz speech, supports zero-shot voice
+cloning from a reference clip, and handles 100+ languages with inline control
+tokens for emotion, style, and prosody. The architecture is a ~4B Qwen3
+backbone with fused multi-codebook embedding/head (8 codebooks × 1026 vocab,
+MusicGen-style delay pattern).
+
+## References
+
+- Model card: [bosonai/higgs-audio-v3-tts-4b](https://huggingface.co/bosonai/higgs-audio-v3-tts-4b)
+- Offline example: [`examples/offline_inference/text_to_speech/higgs_audio_v3/end2end.py`](../../examples/offline_inference/text_to_speech/higgs_audio_v3/end2end.py)
+- Online example: [`examples/online_serving/text_to_speech/higgs_audio_v3/`](../../examples/online_serving/text_to_speech/higgs_audio_v3/)
+- Benchmark results: see Performance section below
+
+## Hardware Support
+
+## GPU
+
+### 1×H100 80GB
+
+#### Environment
+
+- OS: Linux
+- Python: 3.12+
+- CUDA: 12.x
+- vLLM version: 0.22.0
+- vLLM-Omni version or commit: `36e048fd` (branch `higgs-v3`)
+
+#### Command
+
+**Online serving:**
+
+```bash
+vllm-omni serve bosonai/higgs-audio-v3-tts-4b \
+    --host 0.0.0.0 --port 8095 \
+    --trust-remote-code --omni
+```
+
+The default deploy config `vllm_omni/deploy/higgs_multimodal_qwen3.yaml` is
+loaded automatically by model registry (HF `model_type=higgs_multimodal_qwen3`).
+It matches the high-throughput profile.
+
+**Offline batch inference:**
+
+```bash
+python examples/offline_inference/text_to_speech/higgs_audio_v3/end2end.py \
+    --texts "Hello world." "The quick brown fox jumps over the lazy dog." \
+    --output-dir results/higgs_v3_wavs
+```
+
+**Offline voice clone:**
+
+```bash
+python examples/offline_inference/text_to_speech/higgs_audio_v3/end2end.py \
+    --texts "Text to synthesize in the cloned voice." \
+    --ref-audio path/to/reference.wav \
+    --ref-text "Transcript of the reference clip." \
+    --output-dir results/higgs_v3_clone
+```
+
+#### Verification
+
+Basic TTS via curl:
+
+```bash
+curl -X POST http://localhost:8095/v1/audio/speech \
+    -H "Content-Type: application/json" \
+    -d '{
+        "model": "bosonai/higgs-audio-v3-tts-4b",
+        "input": "Hello, how are you?"
+    }' --output hello.wav
+```
+
+Voice clone via Python client:
+
+```bash
+python examples/online_serving/text_to_speech/higgs_audio_v3/batch_speech_client.py \
+    --base-url http://localhost:8095 \
+    --model bosonai/higgs-audio-v3-tts-4b \
+    --ref-audio path/to/reference.wav \
+    --ref-text "Transcript of the reference." \
+    --prompts "Text to clone."
+```
+
+#### Notes
+
+- Memory usage: Stage 0 (talker, ~4B) uses ~60% GPU memory; Stage 1 (codec decoder) uses ~25%.
+- Key flags: `--trust-remote-code` and `--omni` are required.
+- Output: 24 kHz mono WAV.
+- Voice cloning: `ref_audio` accepts WAV/FLAC/MP3; `ref_text` is optional but improves fidelity.
+- Deploy config: `vllm_omni/deploy/higgs_multimodal_qwen3.yaml` (auto-discovered from `model_type`).
+  - `max_num_seqs=16` for both stages.
+  - Stage 0 and Stage 1 default to the same device (`0`) for single-GPU serving.
+  - Stage 0 intentionally keeps `enforce_eager=true`. This preserves the Higgs-specific local MLP CUDA graph path, which is the current high-throughput default.
+  - Stage 1 remains `enforce_eager=true` for the codec decoder.
+- Deploy profiles:
+  - High throughput: `vllm_omni/deploy/higgs_multimodal_qwen3_high_throughput.yaml`.
+    Use this for medium/high concurrency. The auto-discovered
+    `higgs_multimodal_qwen3.yaml` is kept as a compatibility/default alias for
+    this profile.
+  - Low latency: `vllm_omni/deploy/higgs_multimodal_qwen3_low_latency.yaml`.
+    Use this for low-concurrency serving (for example c1-c4). It sets Stage 0
+    `enforce_eager=false` and enables vLLM `FULL_DECODE_ONLY` CUDA graph through
+    YAML `compilation_config`; no environment variable is required.
+  - Profile details: `vllm_omni/deploy/README_higgs_audio_v3.md`.
+- Performance note:
+  - Do not switch the auto-discovered default Stage 0 profile to vLLM
+    `FULL_DECODE_ONLY` CUDA graph without an end-to-end throughput and
+    audio-quality revalidation. On the H20 SeedTTS c16/full-dataset benchmark,
+    the eager Stage 0 path with Higgs local MLP CUDA graph reproduced
+    ~35 audio_s/s (`1088/1088` OK per run, three runs around 134-136s). A
+    separate H20 low-concurrency smoke after the FULL_DECODE audio-feedback
+    capture fix showed lower c1/c4 latency, so the FULL_DECODE profile is kept
+    as an explicit low-latency option rather than the throughput default.
+- Known limitations:
+  - Stage 1 (code2wav) must use `enforce_eager=true` (`@torch.inference_mode` incompatible with graph capture).
+  - Stage 0 full-decode CUDA graph is experimental; sampler, delay-state updates, staging, and request postprocess remain outside the graph.

recipes/README.md

@@ -33,11 +33,12 @@ recipes/
 | [`fishaudio/Fish-Speech-S2-Pro.md`](./fishaudio/Fish-Speech-S2-Pro.md) | Online serving for TTS | 1x A800 80GB |
 | [`Helios/Helios.md`](./Helios/Helios.md) | Text-to-video, image-to-video, and video-to-video generation | 1x NVIDIA H20 |
 | [`inclusionAI/Ming-flash-omni-2.0.md`](./inclusionAI/Ming-flash-omni-2.0.md) | Online serving for multimodal chat + standalone TTS | 4x H100 / 1x H100 80GB |
+| [`inclusionAI/Ming-omni-tts-0.5B.md`](./inclusionAI/Ming-omni-tts-0.5B.md) | Offline + online dense Ming TTS/audio generation | 1x AMD `gfx942` (ROCm 7.2) |
 | [`LTX/LTX-2.md`](./LTX/LTX-2.md) | Text-to-video and image-to-video serving | 1x H200 141GB |
 | [`LTX/LTX-2.3.md`](./LTX/LTX-2.3.md) | Text-to-video with audio generation (22B) | 1x GPU (96GB VRAM) |
 | [`mistralai/Voxtral-TTS.md`](./mistralai/Voxtral-TTS.md) | Online serving for TTS | 1x RTX 4090 24GB |
-| [`cosmos3/Cosmos3-Nano.md`](./cosmos3/Cosmos3-Nano.md) | Text-to-image, text-to-video, image-to-video generation, text to video with sound, action policy  | 1x H200 141GB / B300 |
-| [`cosmos3/Cosmos3-Super.md`](./cosmos3/Cosmos3-Super.md) | 64B T2I / T2V / I2V generation (+ optional audio) / Action policy | 8x H200/H100/A100 / 2x H200 / B300 |
+| [`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-Super.md`](./cosmos3/Cosmos3-Super.md) | 64B T2I / T2V / I2V / V2V generation (+ optional audio) / Action policy | 8x H200/H100/A100 / 2x H200 / B300 |
 | [`OpenBMB/MiniCPM-o-4_5.md`](./OpenBMB/MiniCPM-o-4_5.md) | Online serving for omni multimodal chat (text / image / audio / video → text + 24 kHz speech) | 2x A100/H100 80GB / 3x mid-tier GPU / 8x RTX 4090 24GB |
 | [`OpenBMB/VoxCPM2.md`](./OpenBMB/VoxCPM2.md) | Online + offline TTS with native AR pipeline (48 kHz, 30+ languages) | 1x RTX 4090 24GB |
 | [`Qwen/Qwen-Image.md`](./Qwen/Qwen-Image.md) | Text-to-image serving with step-wise continuous batching replay and ModelOpt mixed FP8/NVFP4 | 1x A100 80GB / 2x B200 |
@@ -50,9 +51,9 @@ recipes/
 | [`Tencent/HunyuanImage-3.0-Instruct.md`](./Tencent/HunyuanImage-3.0-Instruct.md) | DiT-only text-to-image serving and benchmark, including ModelOpt mixed FP8/NVFP4 | 4x H100/H800 80GB / 2x B200 |
 | [`Wan-AI/Wan2.2-I2V.md`](./Wan-AI/Wan2.2-I2V.md) | Image-to-video serving (Wan2.2 14B) | 8x Ascend NPU (A2/A3) |
 | [`Wan-AI/Wan2.2-S2V.md`](./Wan-AI/Wan2.2-S2V.md) | Speech-to-video serving (Wan2.2 14B) | 2x A100/H100 80GB |
+| [`StabilityAI/Stable-Diffusion-3.5.md`](./StabilityAI/Stable-Diffusion-3.5.md) | Text-to-image serving (SD 3.5-medium and SD 3.5-large) | 1x RTX A6000 48GB |
 | [`zai-org/GLM-TTS.md`](./zai-org/GLM-TTS.md) | Online serving for Chinese/English zero-shot voice-cloned TTS | 1x A40 48GB |
 | [`GLM/GLM-Image.md`](./GLM/GLM-Image.md) | Online serving for image generation | 1x A800 80GB / 2x A800 80GB |
-|
 
 Within a single recipe file, include different hardware support sections such
 as `GPU`, `ROCm`, and `NPU`, and add concrete tested configurations like

recipes/StabilityAI/Stable-Diffusion-3.5.md

@@ -0,0 +1,139 @@
+# Stable Diffusion 3.5
+
+> Text-to-image serving with Stability AI's SD 3.5 family
+
+## Summary
+
+- Vendor: Stability AI
+- Model: `stabilityai/stable-diffusion-3.5-medium`, `stabilityai/stable-diffusion-3.5-large`
+- Task: Text-to-image generation
+- Mode: Online serving with the OpenAI-compatible images API
+- Maintainer: Community
+
+## When to use this recipe
+
+Use this recipe when you want a lightweight, fast text-to-image model that
+fits on a single consumer or workstation GPU. SD 3.5-medium requires only
+~17 GiB VRAM and generates 1024×1024 images in ~4 seconds, making it one of
+the most accessible diffusion models in vLLM-Omni. SD 3.5-large produces
+noticeably sharper images at the cost of ~34 GiB VRAM and ~21 seconds per
+image.
+
+## References
+
+- Model (medium): <https://huggingface.co/stabilityai/stable-diffusion-3.5-medium>
+- Model (large): <https://huggingface.co/stabilityai/stable-diffusion-3.5-large>
+- Related example under `examples/`:
+  [`examples/online_serving/text_to_image/README.md`](../../examples/online_serving/text_to_image/README.md)
+- User guide:
+  [`docs/user_guide/examples/online_serving/text_to_image.md`](../../docs/user_guide/examples/online_serving/text_to_image.md)
+
+## Hardware Support
+
+This recipe documents two CUDA GPU configurations on a single RTX A6000 48 GB,
+one for each model variant. Extend it with more hardware sections as community
+validation lands.
+
+## GPU
+
+### 1x NVIDIA RTX A6000 48GB — stable-diffusion-3.5-medium
+
+#### Environment
+
+- OS: Ubuntu 22.04
+- Python: 3.12
+- Driver / runtime: NVIDIA driver 580.126.09, CUDA 13.0
+- vLLM version: 0.22.0
+- vLLM-Omni version or commit: 0.22.0rc2.dev8+g41a795b5d
+
+#### Command
+
+```bash
+vllm serve stabilityai/stable-diffusion-3.5-medium --omni --port 8091
+```
+
+#### Verification
+
+```bash
+curl -X POST http://localhost:8091/v1/images/generations \
+  -H "Content-Type: application/json" \
+  -d '{
+    "model": "stabilityai/stable-diffusion-3.5-medium",
+    "prompt": "a cute cat sitting on a windowsill, highly detailed, sharp focus",
+    "negative_prompt": "blurry, out of focus",
+    "guidance_scale": 4.5,
+    "n": 1,
+    "size": "1024x1024"
+  }' | python3 -c "
+import json, base64, sys
+data = json.load(sys.stdin)
+with open('output.png', 'wb') as f:
+    f.write(base64.b64decode(data['data'][0]['b64_json']))
+print('saved output.png')
+"
+```
+
+Or use the example client:
+
+```bash
+python examples/online_serving/text_to_image/openai_chat_client.py \
+  --server http://localhost:8091 \
+  --prompt "a cute cat sitting on a windowsill" \
+  --output output.png
+```
+
+#### Notes
+
+- Memory usage: Model weights ~15.6 GiB, peak VRAM ~17.0 GiB (no CFG), ~20.9 GiB with `guidance_scale > 1` (CFG enabled).
+- Generation time: ~4 seconds without CFG, ~8 seconds with `guidance_scale=4.5`.
+- Model loading time: ~55 seconds (including weight download on first run).
+- `negative_prompt` requires `guidance_scale > 1` to take effect; SD3 defaults to 1.0 (no CFG).
+
+### 1x NVIDIA RTX A6000 48GB — stable-diffusion-3.5-large
+
+#### Environment
+
+- OS: Ubuntu 22.04
+- Python: 3.12
+- Driver / runtime: NVIDIA driver 580.126.09, CUDA 13.0
+- vLLM version: 0.22.0
+- vLLM-Omni version or commit: 0.22.0rc2.dev8+g41a795b5d
+
+#### Command
+
+```bash
+vllm serve stabilityai/stable-diffusion-3.5-large --omni --port 8091
+```
+
+> **Note**: `stable-diffusion-3.5-large` is a gated model. Accept the license
+> on [Hugging Face](https://huggingface.co/stabilityai/stable-diffusion-3.5-large)
+> and set `HF_TOKEN` before running.
+
+#### Verification
+
+```bash
+curl -X POST http://localhost:8091/v1/images/generations \
+  -H "Content-Type: application/json" \
+  -d '{
+    "model": "stabilityai/stable-diffusion-3.5-large",
+    "prompt": "a cute cat sitting on a windowsill, highly detailed, sharp focus",
+    "negative_prompt": "blurry, out of focus",
+    "guidance_scale": 4.5,
+    "n": 1,
+    "size": "1024x1024"
+  }' | python3 -c "
+import json, base64, sys
+data = json.load(sys.stdin)
+with open('output.png', 'wb') as f:
+    f.write(base64.b64decode(data['data'][0]['b64_json']))
+print('saved output.png')
+"
+```
+
+#### Notes
+
+- Memory usage: Peak VRAM ~31.6 GiB at 1024×1024, ~45.5 GiB at 2048×2048 with `guidance_scale=4.5`.
+- Generation time: ~19 seconds at 1024×1024, ~102 seconds at 2048×2048.
+- Image quality is noticeably better than medium, with sharper details.
+- `negative_prompt` requires `guidance_scale > 1` to take effect; SD3 defaults to 1.0 (no CFG).
+- 2048×2048 fits within 48 GB but leaves little headroom (~45.5 GiB used out of 49 GiB).

recipes/Wan-AI/Wan2.2-S2V.md

@@ -171,6 +171,34 @@ python examples/offline_inference/speech_to_video/speech_to_video.py \
 ```
 
 
+### Online Serving
+
+#### Server
+
+```bash
+vllm serve Wan-AI/Wan2.2-S2V-14B --omni \
+  --model-class-name WanS2VPipeline \
+  --tensor-parallel-size 2 \
+  --flow-shift 3.0 \
+  --vae-use-slicing --vae-use-tiling \
+  --port 8091
+```
+
+#### Client
+
+```bash
+no_proxy=127.0.0.1 \
+curl -X POST http://127.0.0.1:8091/v1/videos/sync \
+  -F "prompt=A person singing" \
+  -F 'image_reference={"image_url": "https://raw.githubusercontent.com/Wan-Video/Wan2.2/main/examples/Five%20Hundred%20Miles.png"}' \
+  -F 'audio_reference={"audio_url": "https://raw.githubusercontent.com/Wan-Video/Wan2.2/main/examples/Five%20Hundred%20Miles.MP3"}' \
+  -F "width=832" -F "height=480" \
+  -F "num_inference_steps=40" \
+  -F "guidance_scale=4.5" \
+  -F "fps=16" \
+  --output s2v_480p_serve.mp4
+```
+
 #### Verification
 
 The output video should show a talking/singing person matching the reference

recipes/cosmos3/Cosmos3-Nano.md

@@ -1,34 +1,43 @@
 # Cosmos3-Nano
 
-> Text-to-image, text-to-video, and image-to-video serving
+> Text-to-image, text-to-video, image-to-video, and video-to-video serving
 
 ## Summary
 
 - Vendor: NVIDIA
 - Model: `nvidia/Cosmos3-Nano`
-- Task: Text-to-image (T2I), text-to-video (T2V), and image-to-video (I2V) generation, with optional synchronized audio (video + sound), action policy
+- Task: Text-to-image (T2I), text-to-video (T2V), image-to-video (I2V), and video-to-video (V2V) generation, with optional synchronized audio (video + sound), action policy
 - Mode: Online serving with the OpenAI-compatible image/video APIs, plus offline generation via the `Omni` API
 - Maintainer: Community
 
 ## When to use this recipe
 
 Use this recipe to deploy `nvidia/Cosmos3-Nano` for image and video generation.
-A single pipeline class (`Cosmos3OmniDiffusersPipeline`) serves all three modes;
-the mode is selected per request:
+A single pipeline class (`Cosmos3OmniDiffusersPipeline`) serves these modes; the
+mode is selected per request:
 
 - **T2I** — `POST /v1/images/generations` (or a prompt carrying `modalities=["image"]`).
-- **T2V** — `POST /v1/videos/sync` with `num_frames > 1` and no reference image.
+- **T2V** — `POST /v1/videos/sync` with `num_frames > 1` and no reference image/video.
 - **I2V** — `POST /v1/videos/sync` with a reference image (`input_reference` file
   upload, or `image_reference` JSON).
+- **V2V** — `POST /v1/videos/sync` with a reference video (`input_reference`
+  video upload, or `video_reference` JSON). Cosmos3 conditions on selected
+  reference-video latent frames; use `extra_params.condition_frame_indexes_vision`
+  and `extra_params.condition_video_keep` to choose which prefix/tail frames guide
+  generation.
 - **T2VS / I2VS** — add `generate_sound=true` (and optional `sound_duration`) to a
   T2V/I2V `/v1/videos/sync` request to also generate synchronized audio, muxed into
   the mp4 as AAC 48 kHz stereo. See the official model card's "Video + Audio" examples.
 - **Action** — pass `extra_params={"action_mode": ...}` to drive Physical-AI tasks:
-  - `forward_dynamics` — given a first frame **and** an action trajectory, roll out
-    the resulting video. Synchronous: `POST /v1/videos/sync`.
-  - `policy` — given a first frame and a language instruction, **predict** the action
-    trajectory (and a rollout video). Use the async `POST /v1/videos` endpoint and
-    read the predicted action from the top-level `action` field
+  - `forward_dynamics` — given a first frame or video **and** an action trajectory,
+    roll out the resulting video. Synchronous: `POST /v1/videos/sync`.
+  - `policy` — given a first frame or video and a language instruction,
+    **predict** the action trajectory (and a rollout video). Use the async
+    `POST /v1/videos` endpoint and read the predicted action from the top-level
+    `action` field.
+  - `inverse_dynamics` — given a video, **recover** the action trajectory. Use
+    the async `POST /v1/videos` endpoint and read the recovered action from
+    the top-level `action` field
     (`{data, shape, dtype, raw_action_dim, domain_id}`).
 
   Action requests also take `domain_name` (e.g. `av`, `bridge_orig_lerobot`,
@@ -38,8 +47,9 @@ the mode is selected per request:
   **`nvidia/Cosmos3-Nano-Policy-DROID`** is served the same way
   (`domain_name=droid_lerobot`).
 
-  `inverse_dynamics` (recover the action from a given video) is supported by the
-  pipeline; **online inference of inverse dynamics will be added in a follow-up MR.**
+  Action requests can use `input_reference` or `video_reference` for video input.
+  `policy` and `forward_dynamics` can also use an image reference; `inverse_dynamics`
+  requires a video reference.
 
 ## References
 
@@ -142,6 +152,36 @@ curl -sS -X POST http://localhost:8000/v1/videos/sync \
   -F "input_reference=@/path/to/reference.jpg;type=image/jpeg" \
   -o cosmos3_i2v.mp4
 
+# Video-to-video -> /v1/videos/sync with an uploaded reference video.
+# By default Cosmos3 conditions on latent indexes [0, 1]. For the default
+# temporal VAE stride this decodes only the first 5 input frames.
+# The model works best when the prompt describes the actual situation happening in the video.
+# Generic prompts may create sub-standard generations.
+curl -sS -X POST http://localhost:8000/v1/videos/sync \
+  -H "Accept: video/mp4" \
+  -F "model=nvidia/Cosmos3-Nano" \
+  -F "prompt=Continue the same scene with smooth natural motion and consistent subjects." \
+  -F "negative_prompt=blurry, distorted, low quality, jittery, deformed" \
+  -F "size=1280x720" -F "num_frames=189" -F "fps=24" \
+  -F "num_inference_steps=35" -F "guidance_scale=6.0" \
+  -F "max_sequence_length=4096" -F "flow_shift=10.0" \
+  -F 'extra_params={"use_resolution_template":false,"use_duration_template":false,"guardrails":true,"condition_frame_indexes_vision":[0,1],"condition_video_keep":"first"}' \
+  -F "seed=2222" \
+  -F "input_reference=@/path/to/reference.mp4;type=video/mp4" \
+  -o cosmos3_v2v.mp4
+
+# V2V can also use a JSON-safe URL/data-URL video reference. Do not combine
+# video_reference with input_reference or image_reference.
+curl -sS -X POST http://localhost:8000/v1/videos/sync \
+  -H "Accept: video/mp4" \
+  -F "model=nvidia/Cosmos3-Nano" \
+  -F "prompt=Continue the same scene with smooth natural motion and consistent subjects." \
+  -F "size=1280x720" -F "num_frames=189" -F "fps=24" \
+  -F "num_inference_steps=35" -F "guidance_scale=6.0" \
+  -F "max_sequence_length=4096" -F "flow_shift=10.0" \
+  -F 'extra_params={"condition_frame_indexes_vision":[0,1],"condition_video_keep":"last"}' \
+  -F 'video_reference={"video_url":"https://example.com/reference.mp4"}' \
+  -o cosmos3_v2v_from_url.mp4
 
 # Text-to-video-with-sound
 curl -sS -X POST http://localhost:8000/v1/videos/sync \
@@ -190,6 +230,21 @@ VIDEO_ID=$(curl -sS -X POST http://localhost:8000/v1/videos \
 # poll until status == completed, then:
 curl -sS "http://localhost:8000/v1/videos/$VIDEO_ID" | jq '.action | {shape, dtype, raw_action_dim, domain_id}'
 curl -sS -L "http://localhost:8000/v1/videos/$VIDEO_ID/content" -o cosmos3_policy.mp4
+
+# Action — inverse dynamics (video -> recovered action trajectory).
+# Asynchronous: use the job metadata to read the recovered action.
+VIDEO_ID=$(curl -sS -X POST http://localhost:8000/v1/videos \
+  -H "Accept: application/json" \
+  --form-string "model=nvidia/Cosmos3-Nano" \
+  --form-string "prompt=Recover the robot action trajectory from this clip." \
+  -F "input_reference=@motion_clip.mp4;type=video/mp4" \
+  -F "size=640x480" -F "num_frames=17" -F "fps=5" \
+  -F "num_inference_steps=30" -F "guidance_scale=1.0" -F "flow_shift=5.0" \
+  --form-string 'extra_params={"action_mode":"inverse_dynamics","domain_name":"bridge_orig_lerobot","raw_action_dim":10,"action_chunk_size":16}' \
+  -F "seed=0" | jq -r '.id')
+# poll until status == completed, then:
+curl -sS "http://localhost:8000/v1/videos/$VIDEO_ID" | jq '.action | {shape, dtype, raw_action_dim, domain_id}'
+curl -sS -L "http://localhost:8000/v1/videos/$VIDEO_ID/content" -o cosmos3_inverse_dynamics.mp4
 ```
 
 #### Notes
@@ -206,24 +261,27 @@ curl -sS -L "http://localhost:8000/v1/videos/$VIDEO_ID/content" -o cosmos3_polic
 - **Determinism:** identical seed reproduces identical output on the same
   hardware; outputs are not bit-identical across different GPU types.
 - **Supported sizes (per model card):** 256p / 480p / 720p at 16:9, 4:3, 1:1,
-  3:4, 9:16. Defaults: T2I 1024², 50 steps, guidance 7.0; T2V/I2V 1280×720,
-  189 frames, 35 steps, guidance 6.0, `flow_shift=10.0`.
+  3:4, 9:16. Defaults: T2I 1024², 50 steps, guidance 7.0; T2V/I2V/V2V
+  1280×720, 189 frames, 35 steps, guidance 6.0, `flow_shift=10.0`.
 - **Key flags / params:** `--no-guardrails` (server) or
   `extra_params={"guardrails":false}` (per request) toggles safety. The
   per-request flag only takes effect when the server was launched **with**
   guardrails enabled (it cannot re-enable them on a `--no-guardrails` server).
   `use_resolution_template` / `use_duration_template` are off by default and only
   needed when not using upsampled prompts that already encode resolution/duration.
+  For V2V, `condition_frame_indexes_vision` selects the clean conditioned latent
+  frame indexes (default `[0, 1]`), and `condition_video_keep` selects whether the
+  API decodes the first or last needed reference frames (`"first"` by default).
 - **Known limitations:**
   - Guardrails-on requires `cosmos-guardrail` **and** access to the gated
     `nvidia/Cosmos-1.0-Guardrail` repo (accept license + `HF_TOKEN`); otherwise
     the server fails at pipeline build with a gated-repo / safety-checker error.
   - A guardrail-blocked prompt currently returns HTTP 500
     (`"Guardrail blocked prompt"`).
-  - Action `forward_dynamics` (sync `/v1/videos/sync`) and `policy` (async
-    `/v1/videos`, returns the predicted action under the top-level `action`
-    field) are supported online. **Online inference of inverse dynamics will be
-    added in a follow-up MR.**
+  - Action `forward_dynamics`, `policy`, and `inverse_dynamics` are supported
+    online. Use async `POST /v1/videos` when you need the predicted/recovered
+    action payload under the top-level `action` field; sync `/v1/videos/sync`
+    returns raw MP4 bytes and does not expose action metadata in the response body.
 
 ### 1x GPU (Offline generation)
 
@@ -287,8 +345,9 @@ python -c "from PIL import Image; im=Image.open('cosmos3_t2i.png'); print('image
 
 - Same `Cosmos3OmniDiffusersPipeline` as online; mode is chosen by
   `prompt["modalities"]` (`["image"]` → T2I, `["video"]` → T2V) plus
-  `num_frames`/`fps`, and `multi_modal_data={"image": ...}` for I2V. For video,
-  frames are returned in `outputs[0].request_output.images` as an
+  `num_frames`/`fps`, `multi_modal_data={"image": ...}` for I2V, and
+  `multi_modal_data={"video": [<PIL frames>]}` or a video tensor/array for V2V.
+  For video, frames are returned in `outputs[0].request_output.images` as an
   `(B, F, H, W, 3)` array.
 - The offline entry must be guarded by `if __name__ == "__main__":` — the engine
   spawns workers with the `spawn` start method.

recipes/cosmos3/Cosmos3-Super.md

@@ -1,12 +1,12 @@
 # Cosmos3-Super
 
-> Frontier 64B world model: text-to-image, text-to-video, image-to-video (+ optional audio)
+> Frontier 64B world model: text-to-image, text-to-video, image-to-video, video-to-video (+ optional audio)
 
 ## Summary
 
 - Vendor: NVIDIA
 - Model: `nvidia/Cosmos3-Super` (64B; also `Cosmos3-Super-Text2Image`, `Cosmos3-Super-Image2Video`)
-- Task: T2I, T2V, I2V generation, with optional synchronized audio (video + sound)
+- Task: T2I, T2V, I2V, V2V generation, with optional synchronized audio (video + sound)
 - Mode: Online serving with the OpenAI-compatible image/video APIs
 - Maintainer: Community
 
@@ -16,8 +16,8 @@ Use this recipe to deploy the 64B `nvidia/Cosmos3-Super` for the highest-quality
 Cosmos3 generation. It shares the same `Cosmos3OmniDiffusersPipeline` and request
 formats as [Cosmos3-Nano](./Cosmos3-Nano.md) — only the checkpoint size and the
 recommended parallelism differ. Mode is selected per request (T2I →
-`/v1/images/generations`; T2V/I2V → `/v1/videos/sync`; add `generate_sound=true`
-for audio).
+`/v1/images/generations`; T2V/I2V/V2V → `/v1/videos/sync`; add
+`generate_sound=true` for audio).
 
 ## References
 
@@ -62,8 +62,9 @@ disable. `--enable-layerwise-offload` reduces VRAM on smaller GPUs.
 #### Verification
 
 Requests are identical to Nano (see [`Cosmos3-Nano.md`](./Cosmos3-Nano.md) for full
-T2I/T2V/I2V/T2VS curls); official params: `size=1280x720, num_frames=189, fps=24,
-num_inference_steps=35, guidance_scale=6.0, flow_shift=10.0, max_sequence_length=4096`.
+T2I/T2V/I2V/V2V/T2VS curls); official params: `size=1280x720, num_frames=189,
+fps=24, num_inference_steps=35, guidance_scale=6.0, flow_shift=10.0,
+max_sequence_length=4096`.
 
 ```bash
 curl http://localhost:8000/v1/models
@@ -84,6 +85,15 @@ curl -sS -X POST http://localhost:8000/v1/videos/sync -H "Accept: video/mp4" \
   -F "seed=1111" -F "input_reference=@/path/to/reference.jpg;type=image/jpeg" \
   -o cosmos3_super_i2v.mp4
 
+# V2V — add an uploaded reference video. condition_video_keep can be "first" or "last".
+curl -sS -X POST http://localhost:8000/v1/videos/sync -H "Accept: video/mp4" \
+  -F "model=nvidia/Cosmos3-Super" -F "prompt=Continue the same scene with smooth natural motion." \
+  -F "size=1280x720" -F "num_frames=189" -F "fps=24" -F "num_inference_steps=35" \
+  -F "guidance_scale=6.0" -F "max_sequence_length=4096" -F "flow_shift=10.0" \
+  -F 'extra_params={"condition_frame_indexes_vision":[0,1],"condition_video_keep":"first"}' \
+  -F "seed=2222" -F "input_reference=@/path/to/reference.mp4;type=video/mp4" \
+  -o cosmos3_super_v2v.mp4
+
 # T2V + sound — add generate_sound/sound_duration (output muxes AAC 48 kHz stereo)
 curl -sS -X POST http://localhost:8000/v1/videos/sync -H "Accept: video/mp4" \
   -F "model=nvidia/Cosmos3-Super" -F "prompt=A robot arm is cleaning a plate in the kitchen" \
@@ -103,10 +113,12 @@ curl -sS -X POST http://localhost:8000/v1/videos/sync -H "Accept: video/mp4" \
   - T2V + sound (189 frames, 35 steps) → **~198 s**, output muxes **AAC 48 kHz stereo**
   - (NVIDIA's reference: 8×H200 @ 50 steps ≈ 55 s/video; 2×H200 @ 35 steps ≈ 3 min/video.)
 - **Memory:** ~61.5 GiB per GPU when sharded across 2 GPUs (HSDP shard 2); repo ~135 GB on disk.
-- Same generation defaults, supported sizes, and `generate_sound`/`sound_duration`
-  semantics as Nano, including the **action** modality: `forward_dynamics`
-  (sync `/v1/videos/sync`) and `policy` (async `/v1/videos`, predicted action under
-  the top-level `action` field) — see the Cosmos3-Nano recipe for the request shape.
-  Online inference of inverse dynamics will be added in a follow-up MR. Verified on
-  the 64B Super under `--cfg-parallel-size 2`: async `policy` returns the predicted
-  action (`[16, 10]`) and the rollout video reliably.
+- Same generation defaults, supported sizes, V2V reference-video controls
+  (`condition_frame_indexes_vision`, `condition_video_keep`), and
+  `generate_sound`/`sound_duration`
+  semantics as Nano, including the **action** modality: `forward_dynamics`,
+  `policy`, and `inverse_dynamics` — see the Cosmos3-Nano recipe for the request
+  shapes. Use async `/v1/videos` when you need predicted/recovered action metadata
+  under the top-level `action` field. Verified on the 64B Super under
+  `--cfg-parallel-size 2`: async `policy` returns the predicted action (`[16, 10]`)
+  and the rollout video reliably.

recipes/inclusionAI/Ming-omni-tts.md

@@ -0,0 +1,181 @@
+# Ming-omni-tts
+
+> Online serving for voice synthesis, cloning, and design
+
+## Summary
+
+- Vendor: inclusionAI
+- Model: `inclusionAI/Ming-omni-tts-0.5B`
+- Task: Text-to-speech with style, dialect, cloning, and multi-speaker controls
+- Mode: Online serving via the OpenAI-compatible `/v1/audio/speech` API
+- Maintainer: Community
+
+## References
+
+- [Huggingface Model card](https://huggingface.co/inclusionAI/Ming-omni-tts-0.5B)
+- Upstream repository [inclusionAI/Ming-omni-tts](https://github.com/inclusionAI/Ming-omni-tts)
+
+
+## Hardware Support
+
+This recipe documents a validated ROCm configuration and a CUDA configuration for the dense 0.5B two-stage TTS pipeline deployment.
+Other hardware is welcome as community validation lands.
+
+## CUDA
+
+### 1x H100 80GB
+
+#### Environment
+
+- OS: Linux
+- Python: 3.10+
+- CUDA Driver Version: 590.48.01
+- CUDA: 13.0
+- vLLM version: 0.22.0
+- vLLM-Omni version or commit: 0342827d
+
+#### Command
+
+Launch the two-stage talker:
+
+```bash
+vllm-omni serve inclusionAI/Ming-omni-tts-0.5B \
+    --deploy-config vllm_omni/deploy/ming_tts.yaml \
+    --omni \
+    --port 8091 \
+    --enforce-eager
+```
+
+#### Verification
+
+Basic synthesis (save the WAV bytes):
+
+```bash
+curl -X POST http://localhost:8091/v1/audio/speech \
+    -H "Content-Type: application/json" \
+    -d '{
+      "model": "inclusionAI/Ming-omni-tts-0.5B",
+      "input": "你好，这是 Ming 在线语音合成测试。",
+      "response_format": "wav"
+    }' --output ming_tts_basic.wav
+```
+
+The following requests condition on reference clips from upstream audio assets. The server resolves `ref_audio` as an `http(s)://` URL, a `file://` path, or an inline base64 `data:` URL, so the examples point it straight at the public cookbook fixtures in the vendor repo's [`inclusionAI/Ming-omni-tts/tree/main/data/wavs`](https://github.com/inclusionAI/Ming-omni-tts/tree/main/data/wavs) directory:
+
+```bash
+BASE=https://raw.githubusercontent.com/inclusionAI/Ming-omni-tts/main/data/wavs
+```
+
+And note that the server latency will include automatically resolving and downloading input files.
+
+Reference-audio zero-shot cloning:
+
+```bash
+REF_AUDIO="$BASE/10002287-00000094.wav"
+jq -n --arg ref_audio "$REF_AUDIO" \
+    '{
+       model: "inclusionAI/Ming-omni-tts-0.5B",
+       input: "我们的愿景是构建未来服务业的数字化基础设施，为世界带来更多微小而美好的改变。",
+       ref_audio: $ref_audio,
+       ref_text: "在此奉劝大家别乱打美白针。",
+       max_new_tokens: 200,
+       response_format: "wav"
+     }' \
+| curl -X POST http://localhost:8091/v1/audio/speech \
+    -H "Content-Type: application/json" \
+    -d @- --output ming_tts_zero_shot.wav
+```
+
+Podcast-style multi-speaker generation uses one reference clip plus one transcript per speaker, with the `speaker_N` labels in `input`/`ref_text` lined up with the reference order:
+
+```bash
+REF_A="$BASE/CTS-CN-F2F-2019-11-11-423-012-A.wav"
+REF_B="$BASE/CTS-CN-F2F-2019-11-11-423-012-B.wav"
+
+INPUT=' speaker_1:你可以说一下，就大概说一下，可能虽然我也不知道，我看过那部电影没有。
+ speaker_2:就是那个叫什么，变相一节课的嘛。
+ speaker_1:嗯。
+ speaker_2:一部搞笑的电影。
+ speaker_1:一部搞笑的。'
+
+REF_TEXT=' speaker_1:并且我们还要进行每个月还要考核 笔试的话还要进行笔试，做个，当服务员还要去笔试了
+ speaker_2:对啊，这真的很奇怪，就是 单纯的因，单纯自己工资不高，只是因为可能人家那个店比较出名一点，就对你苛刻要求'
+
+jq -n \
+    --arg input "$INPUT" \
+    --arg ref_text "$REF_TEXT" \
+    --arg a "$REF_A" \
+    --arg b "$REF_B" \
+    '{
+       model: "inclusionAI/Ming-omni-tts-0.5B",
+       input: $input,
+       ref_audio: [$a, $b],
+       ref_text: $ref_text,
+       response_format: "wav"
+     }' \
+| curl -X POST http://localhost:8091/v1/audio/speech \
+    -H "Content-Type: application/json" \
+    -d @- --output ming_tts_podcast_style.wav
+```
+
+## ROCm
+
+### 1x AMD MI300X
+
+#### Environment
+
+- OS: Ubuntu 22.04.5 LTS, x86_64
+- Python: 3.12.13
+- ROCm / HIP: 7.2.53211
+- vLLM version: 0.22.0
+- vLLM-Omni version or commit: 0.1.dev1873 / `99c35c410`
+- Docker image: `vllm/vllm-omni-rocm:v0.22.0`
+
+#### Command
+
+From the vLLM-Omni repository root:
+
+```bash
+docker run --rm \
+    --group-add=video \
+    --ipc=host \
+    --cap-add=SYS_PTRACE \
+    --security-opt seccomp=unconfined \
+    --device /dev/kfd \
+    --device /dev/dri \
+    -v ~/.cache/huggingface:/root/.cache/huggingface \
+    -v "$PWD":/app/vllm-omni \
+    -w /app/vllm-omni \
+    -e VLLM_ROCM_USE_AITER=0 \
+    -p 8091:8091 \
+    vllm/vllm-omni-rocm:v0.22.0 \
+    --model inclusionAI/Ming-omni-tts-0.5B \
+    --deploy-config vllm_omni/deploy/ming_tts.yaml \
+    --omni \
+    --port 8091 \
+    --enforce-eager
+```
+
+#### Verification
+
+```bash
+python examples/online_serving/text_to_speech/ming_tts/openai_speech_client.py \
+    --text "我觉得社会企业同个人都有责任" \
+    --instruction-json '{"方言":"广粤话"}' \
+    --ref-audio /path/to/yue_prompt.wav \
+    --max-new-tokens 200 \
+    --output dialect.wav
+```
+
+`--ref-audio` matches upstream `use_spk_emb=True`; do not add `--ref-text`
+for the dialect case.
+
+## Notes
+
+- The official ROCm image includes the platform dependencies.
+- See the [ROCm installation guide](../../docs/getting_started/installation/gpu.md) for interactive and source-build workflows.
+- The reference clips above are a subset of the upstream [`inclusionAI/Ming-omni-tts/tree/main/data/wavs`](https://github.com/inclusionAI/Ming-omni-tts/tree/main/data/wavs) cookbook fixtures.
+- To clone from local audio instead of a URL, pass a `file://` URI (launch the server with `--allowed-local-media-path <dir>`), or base64-encode the clip into a `data:` URL written to a payload file and send it with `curl -d @payload.json`. Putting the base64 directly on the `curl` command line overflows the OS argument limit (`bash: /usr/bin/curl: Argument list too long`).
+- The tested environment uses `--enforce-eager`.
+- Non-streaming responses return WAV bytes; streaming responses return PCM.
+- Output is mono 44.1 kHz audio.

requirements/common.txt

@@ -1,10 +1,15 @@
 # Common dependencies for all platforms
-# transformers: keep aligned with upstream vLLM constraint.
+# 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".
-# 5.9.0 breaks models that define _keys_to_ignore_on_load_unexpected as a list
-# (e.g. OpenMOSS-Team/MOSS-TTS-Nano) because it changed
-# _adjust_missing_and_unexpected_keys from list+list to set|set.
-transformers >= 4.56.0, != 5.0.*, != 5.1.*, != 5.2.*, != 5.3.*, != 5.4.*, != 5.5.0, < 5.9.0
+# The previous extra "< 5.9.0" exclusion was removed when re-aligning with
+# upstream: transformers 5.9 changed _adjust_missing_and_unexpected_keys from
+# 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 >= 4.56.0, != 5.0.*, != 5.1.*, != 5.2.*, != 5.3.*, != 5.4.*, != 5.5.0
 av>=14.0.0
 omegaconf>=2.3.0
 diffusers==0.38.0
@@ -22,4 +27,3 @@ prettytable>=3.8.0
 aenum==3.1.16
 pyzmq>=25.0.0
 janus>=1.0.0
-pydub

requirements/npu.txt

@@ -1,3 +1,3 @@
 -r common.txt
 onnxruntime-cann>=1.23.2
-torchaudio==2.9.0
+torchaudio==2.10.0

requirements/xpu.txt

@@ -1,2 +1,3 @@
 -r common.txt
 onnxruntime>=1.23.2
+auto-round-lib

tests/benchmarks/test_serve_cli.py

@@ -42,10 +42,12 @@ def test_bench_serve_cli_mocks_http_request(tmp_path: Path):
                         yield chunk
 
             class MockResponse:
-                def __init__(self):
-                    self.status = 200
-                    self.reason = "OK"
-                    self.content = _Content(SSE_CHUNKS)
+                def __init__(self, status=200, reason="OK", chunks=None):
+                    self.status = status
+                    self.reason = reason
+                    self.content = _Content(chunks or SSE_CHUNKS)
+                    self._json_data = None
+                    self._text_data = None
 
                 async def __aenter__(self):
                     return self
@@ -53,20 +55,49 @@ def test_bench_serve_cli_mocks_http_request(tmp_path: Path):
                 async def __aexit__(self, exc_type, exc, tb):
                     return False
 
+                async def text(self):
+                    if self._text_data is not None:
+                        return self._text_data
+                    return json.dumps({"tokens": [1, 2, 3, 4, 5]})
+
+                async def json(self):
+                    if self._json_data is not None:
+                        return self._json_data
+                    return {"tokens": [1, 2, 3, 4, 5]}
+
+                def raise_for_status(self):
+                    pass
+
             class MockClientSession:
                 def __init__(self, *args, **kwargs):
                     pass
 
+                async def __aenter__(self):
+                    return self
+
+                async def __aexit__(self, exc_type, exc, tb):
+                    return False
+
                 def post(self, url=None, *args, **kwargs):
                     if url is not None:
                         POST_CALLS.append(url)
-                    return MockResponse()
+                    return MockResponse(status=200)
+
+                async def get(self, url=None, *args, **kwargs):
+                    if url is not None:
+                        POST_CALLS.append(url)
+                    return MockResponse(status=200)
 
                 async def close(self):
                     return None
 
-            import vllm_omni.benchmarks.patch.patch as patch_mod
+            # Patch globally so modules loaded after this also see the mock
+            import aiohttp
+            aiohttp.ClientSession = MockClientSession
+            aiohttp.TCPConnector = lambda *args, **kwargs: object()
 
+            # Also patch in the patch module
+            import vllm_omni.benchmarks.patch.patch as patch_mod
             patch_mod.aiohttp.ClientSession = MockClientSession
             patch_mod.aiohttp.TCPConnector = lambda *args, **kwargs: object()
 
@@ -136,8 +167,13 @@ def test_bench_serve_cli_mocks_http_request(tmp_path: Path):
 
     expected_url = f"http://127.0.0.1:{port}/v1/chat/completions"
     requested_urls = calls["requested_urls"]
-    sent_requests = len(requested_urls)
+    # Count only benchmark requests (to the chat completions endpoint),
+    # excluding tokenize/detokenize alignment calls from the upstream.
+    bench_requests = [url for url in requested_urls if url == expected_url]
+    sent_bench_requests = len(bench_requests)
 
-    assert result["completed"] == sent_requests == num_prompts
-    assert requested_urls
-    assert all(url == expected_url for url in requested_urls), f"Unexpected target URLs: {requested_urls}"
+    assert result["completed"] == sent_bench_requests == num_prompts, (
+        f"completed={result['completed']}, bench_requests={sent_bench_requests}, all_requested_urls={requested_urls}"
+    )
+    assert bench_requests
+    assert all(url == expected_url for url in bench_requests), f"Unexpected target URLs: {bench_requests}"

tests/core/sched/test_omni_generation_scheduler_update_session.py

@@ -20,6 +20,7 @@ from vllm.sampling_params import SamplingParams
 from vllm.v1.engine import EngineCoreEventType
 from vllm.v1.request import Request, RequestStatus, StreamingUpdate
 from vllm_omni.core.sched.omni_generation_scheduler import OmniGenerationScheduler
+from vllm_omni.core.sched.omni_scheduler_mixin import OmniSchedulerMixin
 
 # isort: on
 
@@ -142,3 +143,300 @@ class TestReplaceSessionWithStreamingUpdate:
 
         assert session.events
         assert session.events[-1].type == EngineCoreEventType.QUEUED
+
+
+class _RealignSchedulerStub(OmniSchedulerMixin):
+    """Minimal scheduler surface for ``_realign_request_status_to_queues``.
+
+    Pins ``self.requests`` / ``self.running`` / ``self.waiting`` to the
+    same shapes the upstream ``Scheduler.finish_requests`` reads, so the
+    helper sees realistic state without spinning a real scheduler.
+    """
+
+    def __init__(
+        self,
+        *,
+        requests: dict,
+        running: list,
+        waiting: list,
+    ) -> None:
+        self.requests = requests
+        self.running = running
+        self.waiting = waiting
+
+
+class TestRealignRequestStatusToQueues:
+    """Regression for the residual hang described in
+    https://github.com/vllm-project/vllm-omni/pull/3774 -- chunk-transfer
+    adapter's ``requests_origin_status`` table goes stale on the
+    ``waiting → running`` admit transition, and an abort that lands
+    before the next deque round-trip stomps stale ``WAITING`` onto a
+    request that actually lives in ``self.running``. After
+    ``max_num_seqs`` such aborts every ``input_batch`` slot is leaked and
+    new requests hang at ``chunks=0``.
+    """
+
+    def test_running_with_stale_waiting_status_is_realigned_to_running(
+        self,
+    ) -> None:
+        """The exact race the hang reproduces: a request lives in
+        ``self.running`` but its ``status`` is ``WAITING``. After
+        realign, status must be ``RUNNING`` so upstream
+        ``Scheduler.finish_requests`` removes it from ``self.running``.
+        """
+        req = _make_request(request_id="req-stale")
+        req.status = RequestStatus.WAITING  # stale -- actually in running
+
+        sched = _RealignSchedulerStub(
+            requests={req.request_id: req},
+            running=[req],
+            waiting=[],
+        )
+        sched._realign_request_status_to_queues([req.request_id])
+
+        assert req.status == RequestStatus.RUNNING
+
+    def test_waiting_with_stale_running_status_is_realigned_to_waiting(
+        self,
+    ) -> None:
+        """Symmetric case: request lives in ``self.waiting`` but status
+        is ``RUNNING``. Upstream's else branch should run, so realign to
+        ``WAITING``.
+        """
+        req = _make_request(request_id="req-stale-2")
+        req.status = RequestStatus.RUNNING  # stale -- actually in waiting
+
+        sched = _RealignSchedulerStub(
+            requests={req.request_id: req},
+            running=[],
+            waiting=[req],
+        )
+        sched._realign_request_status_to_queues([req.request_id])
+
+        assert req.status == RequestStatus.WAITING
+
+    def test_already_aligned_status_is_left_unchanged(self) -> None:
+        """Healthy case: status matches actual queue. No status mutation
+        means no spurious side effects.
+        """
+        req_running = _make_request(request_id="req-r")
+        req_running.status = RequestStatus.RUNNING
+        req_waiting = _make_request(request_id="req-w")
+        req_waiting.status = RequestStatus.WAITING
+
+        sched = _RealignSchedulerStub(
+            requests={
+                req_running.request_id: req_running,
+                req_waiting.request_id: req_waiting,
+            },
+            running=[req_running],
+            waiting=[req_waiting],
+        )
+        sched._realign_request_status_to_queues([req_running.request_id, req_waiting.request_id])
+
+        assert req_running.status == RequestStatus.RUNNING
+        assert req_waiting.status == RequestStatus.WAITING
+
+    def test_unknown_request_id_is_skipped_silently(self) -> None:
+        """Aligning ids that aren't in ``self.requests`` (already freed
+        upstream, or never existed) must be a no-op.
+        """
+        sched = _RealignSchedulerStub(requests={}, running=[], waiting=[])
+        sched._realign_request_status_to_queues(["never-existed"])  # no raise
+
+    def test_request_in_neither_queue_is_left_unchanged(self) -> None:
+        """If a tracked request is in neither queue (e.g. parked in a
+        chunk-transfer deque), realign must not invent a status. The
+        adapter / deque purge owns that surface; the realign helper is
+        only for the admit-transition staleness.
+        """
+        req = _make_request(request_id="req-parked")
+        req.status = RequestStatus.WAITING_FOR_CHUNK
+
+        sched = _RealignSchedulerStub(
+            requests={req.request_id: req},
+            running=[],
+            waiting=[],
+        )
+        sched._realign_request_status_to_queues([req.request_id])
+
+        assert req.status == RequestStatus.WAITING_FOR_CHUNK
+
+    def test_request_ids_str_is_treated_as_single_id(self) -> None:
+        """Match upstream ``Scheduler.finish_requests`` resolution: a
+        bare string is treated as one id, not iterated as characters.
+        """
+        req = _make_request(request_id="req-s")
+        req.status = RequestStatus.WAITING
+
+        sched = _RealignSchedulerStub(
+            requests={req.request_id: req},
+            running=[req],
+            waiting=[],
+        )
+        sched._realign_request_status_to_queues(req.request_id)
+
+        assert req.status == RequestStatus.RUNNING
+
+    def test_request_ids_none_aligns_every_known_request(self) -> None:
+        """``request_ids=None`` matches upstream's "all requests" path.
+        Realign must walk every entry in ``self.requests`` and fix any
+        stale status it finds.
+        """
+        stale = _make_request(request_id="req-stale-none")
+        stale.status = RequestStatus.WAITING  # but actually in running
+
+        clean = _make_request(request_id="req-clean-none")
+        clean.status = RequestStatus.RUNNING
+
+        sched = _RealignSchedulerStub(
+            requests={
+                stale.request_id: stale,
+                clean.request_id: clean,
+            },
+            running=[stale, clean],
+            waiting=[],
+        )
+        sched._realign_request_status_to_queues(None)
+
+        assert stale.status == RequestStatus.RUNNING
+        assert clean.status == RequestStatus.RUNNING
+
+    def test_finished_request_is_skipped(self) -> None:
+        """Already-finished requests must not be touched -- they may
+        have legitimate finished statuses (FINISHED_STOPPED etc.) that
+        a status flip would corrupt.
+        """
+        req = _make_request(request_id="req-finished")
+        req.status = RequestStatus.FINISHED_STOPPED
+
+        sched = _RealignSchedulerStub(
+            requests={req.request_id: req},
+            running=[req],  # not realistic, but we want to prove the guard fires
+            waiting=[],
+        )
+        sched._realign_request_status_to_queues([req.request_id])
+
+        assert req.status == RequestStatus.FINISHED_STOPPED
+
+
+class TestPurgeFinishedFromRunning:
+    """Regression for the residual ``self.running`` slot leak surface
+    paired with ``_realign_request_status_to_queues``: even after
+    realign + ``super().finish_requests`` runs, corner cases can leave
+    already-finished or untracked entries in ``self.running`` -- e.g.
+    a connector cleanup that pops from ``self.requests`` without
+    unwinding ``self.running``, or a ``status`` mid-transition when
+    finish ran. The defensive post-finish purge sweeps those residues
+    so the worker's ``input_batch`` slot never pins a freed request.
+
+    See https://github.com/vllm-project/vllm-omni/pull/3774 discussion
+    and the residual-hang reproduction in that PR.
+    """
+
+    def test_finished_request_is_purged_from_running(self) -> None:
+        """``is_finished()`` request lingering in ``self.running`` must
+        be swept so its ``input_batch`` slot is freed."""
+        finished = _make_request(request_id="req-finished")
+        finished.status = RequestStatus.FINISHED_STOPPED  # is_finished() True
+
+        sched = _RealignSchedulerStub(
+            requests={finished.request_id: finished},
+            running=[finished],
+            waiting=[],
+        )
+        sched._purge_finished_from_running()
+
+        assert sched.running == []
+
+    def test_untracked_request_is_purged_from_running(self) -> None:
+        """Request lingering in ``self.running`` but no longer present
+        in ``self.requests`` (already freed by upstream / connector)
+        must be swept."""
+        untracked = _make_request(request_id="req-untracked")
+        untracked.status = RequestStatus.RUNNING
+
+        sched = _RealignSchedulerStub(
+            requests={},  # already deleted from self.requests
+            running=[untracked],
+            waiting=[],
+        )
+        sched._purge_finished_from_running()
+
+        assert sched.running == []
+
+    def test_healthy_running_is_left_unchanged(self) -> None:
+        """Live tracked running requests must not be touched -- the
+        purge is defensive, not aggressive."""
+        alive = _make_request(request_id="req-alive")
+        alive.status = RequestStatus.RUNNING
+
+        sched = _RealignSchedulerStub(
+            requests={alive.request_id: alive},
+            running=[alive],
+            waiting=[],
+        )
+        sched._purge_finished_from_running()
+
+        assert sched.running == [alive]
+
+    def test_empty_running_is_noop(self) -> None:
+        """Empty ``self.running`` must short-circuit cleanly."""
+        sched = _RealignSchedulerStub(requests={}, running=[], waiting=[])
+        sched._purge_finished_from_running()
+
+        assert sched.running == []
+
+    def test_mixed_alive_and_dead_keeps_only_alive_in_order(self) -> None:
+        """Mixed ``self.running`` -- some alive, some finished, some
+        untracked. Sweep keeps only the alive ones and preserves their
+        relative order."""
+        alive_a = _make_request(request_id="req-alive-a")
+        alive_a.status = RequestStatus.RUNNING
+        finished_b = _make_request(request_id="req-finished-b")
+        finished_b.status = RequestStatus.FINISHED_STOPPED
+        untracked_c = _make_request(request_id="req-untracked-c")
+        untracked_c.status = RequestStatus.RUNNING
+        alive_d = _make_request(request_id="req-alive-d")
+        alive_d.status = RequestStatus.RUNNING
+
+        sched = _RealignSchedulerStub(
+            requests={
+                alive_a.request_id: alive_a,
+                finished_b.request_id: finished_b,  # tracked but is_finished
+                # untracked_c is NOT in self.requests
+                alive_d.request_id: alive_d,
+            },
+            running=[alive_a, finished_b, untracked_c, alive_d],
+            waiting=[],
+        )
+        sched._purge_finished_from_running()
+
+        assert sched.running == [alive_a, alive_d]
+
+    def test_in_place_mutation_keeps_helper_local_list_identity(self) -> None:
+        """The slice-assign form (``self.running[:] = ...``) avoids an
+        extra rebind inside this helper. Note that across a full
+        ``finish_requests`` call upstream ``Scheduler.finish_requests``
+        already rebinds ``self.running``, so global identity is not a
+        contract -- this test only pins the local helper behavior so a
+        future refactor away from in-place mutation is intentional.
+        """
+        finished = _make_request(request_id="req-finished")
+        finished.status = RequestStatus.FINISHED_STOPPED
+        alive = _make_request(request_id="req-alive")
+        alive.status = RequestStatus.RUNNING
+
+        sched = _RealignSchedulerStub(
+            requests={
+                finished.request_id: finished,
+                alive.request_id: alive,
+            },
+            running=[finished, alive],
+            waiting=[],
+        )
+        held_ref = sched.running
+        sched._purge_finished_from_running()
+
+        assert held_ref is sched.running
+        assert held_ref == [alive]

tests/core/sched/test_omni_scheduler_finish_requests_purge.py

@@ -0,0 +1,158 @@
+"""Integration tests for the post-finish defensive purge of
+``self.running`` placed in
+``OmniARScheduler.finish_requests`` /
+``OmniGenerationScheduler.finish_requests``.
+
+Pins *placement* — that ``_purge_finished_from_running`` runs AFTER
+``super().finish_requests`` so it catches entries upstream's removal
+missed (e.g. status mid-transition, connector cleanup popping
+``self.requests`` without unwinding ``self.running``). Direct helper
+unit tests live in ``test_omni_scheduler_mixin.py``; these tests
+exercise the full ``finish_requests`` path end-to-end.
+
+Regression for the residual self.running slot leak surface paired
+with ``_realign_request_status_to_queues`` (vllm-omni#3774).
+"""
+
+from __future__ import annotations
+
+import pytest
+from vllm.v1.request import RequestStatus
+
+import vllm_omni.core.sched.omni_ar_scheduler as ar_sched_mod
+import vllm_omni.core.sched.omni_generation_scheduler as gen_sched_mod
+from vllm_omni.core.sched.omni_ar_scheduler import OmniARScheduler
+from vllm_omni.core.sched.omni_generation_scheduler import OmniGenerationScheduler
+
+pytestmark = [pytest.mark.core_model, pytest.mark.cpu]
+
+
+class _StubRequest:
+    """Minimal Request stub with the surface the helpers exercise."""
+
+    def __init__(self, request_id: str, status: RequestStatus) -> None:
+        self.request_id = request_id
+        self.status = status
+
+    def is_finished(self) -> bool:
+        return RequestStatus.is_finished(self.status)
+
+
+_SCHEDULER_PARAMS = [
+    pytest.param(OmniARScheduler, ar_sched_mod, id="ar"),
+    pytest.param(OmniGenerationScheduler, gen_sched_mod, id="generation"),
+]
+
+
+def _make_scheduler(scheduler_cls, *, requests, running, waiting):
+    """Bare scheduler with just the surface ``finish_requests`` reads."""
+    scheduler = scheduler_cls.__new__(scheduler_cls)
+    scheduler.chunk_transfer_adapter = None
+    scheduler.input_coordinator = None
+    scheduler.requests = requests
+    scheduler.running = running
+    scheduler.waiting = waiting
+    return scheduler
+
+
+@pytest.mark.parametrize(("scheduler_cls", "scheduler_mod"), _SCHEDULER_PARAMS)
+def test_finish_requests_purges_finished_request_from_running(
+    monkeypatch: pytest.MonkeyPatch,
+    scheduler_cls,
+    scheduler_mod,
+) -> None:
+    """A request whose ``status`` is a finished value but still lives
+    in ``self.running`` after ``super().finish_requests`` returns must
+    be swept by the post-finish purge. Mocks ``super()`` to simulate
+    the corner case where upstream's status-driven removal missed the
+    entry.
+    """
+    finished = _StubRequest("req-finished", RequestStatus.FINISHED_STOPPED)
+    scheduler = _make_scheduler(
+        scheduler_cls,
+        requests={finished.request_id: finished},
+        running=[finished],
+        waiting=[],
+    )
+
+    def fake_finish_requests(self, request_ids, finished_status):
+        # super() returning the abort tuple but leaving self.running untouched
+        return [(finished.request_id, 0)]
+
+    monkeypatch.setattr(scheduler_mod.VLLMScheduler, "finish_requests", fake_finish_requests)
+
+    result = scheduler_cls.finish_requests(
+        scheduler,
+        [finished.request_id],
+        RequestStatus.FINISHED_STOPPED,
+    )
+
+    assert scheduler.running == []
+    assert result == [(finished.request_id, 0)]
+
+
+@pytest.mark.parametrize(("scheduler_cls", "scheduler_mod"), _SCHEDULER_PARAMS)
+def test_finish_requests_purges_untracked_request_from_running(
+    monkeypatch: pytest.MonkeyPatch,
+    scheduler_cls,
+    scheduler_mod,
+) -> None:
+    """A request that has been deleted from ``self.requests`` (e.g.
+    after upstream's ``_free_request`` ran but a status mismatch
+    caused the ``self.running`` removal branch to no-op) must be
+    swept by the post-finish purge.
+    """
+    untracked = _StubRequest("req-untracked", RequestStatus.RUNNING)
+    scheduler = _make_scheduler(
+        scheduler_cls,
+        requests={},  # already deleted from self.requests
+        running=[untracked],
+        waiting=[],
+    )
+
+    def fake_finish_requests(self, request_ids, finished_status):
+        return [(untracked.request_id, 0)]
+
+    monkeypatch.setattr(scheduler_mod.VLLMScheduler, "finish_requests", fake_finish_requests)
+
+    result = scheduler_cls.finish_requests(
+        scheduler,
+        [untracked.request_id],
+        RequestStatus.FINISHED_STOPPED,
+    )
+
+    assert scheduler.running == []
+    assert result == [(untracked.request_id, 0)]
+
+
+@pytest.mark.parametrize(("scheduler_cls", "scheduler_mod"), _SCHEDULER_PARAMS)
+def test_finish_requests_leaves_healthy_running_intact(
+    monkeypatch: pytest.MonkeyPatch,
+    scheduler_cls,
+    scheduler_mod,
+) -> None:
+    """A live ``RUNNING`` request that is not in the finish set must
+    not be swept by the defensive purge -- this pins that the predicate
+    is conservative and won't aggressively drop healthy entries.
+    """
+    alive = _StubRequest("req-alive", RequestStatus.RUNNING)
+    leaving_id = "req-leaving"
+    scheduler = _make_scheduler(
+        scheduler_cls,
+        requests={alive.request_id: alive},
+        running=[alive],  # leaving already removed by super() (typical happy path)
+        waiting=[],
+    )
+
+    def fake_finish_requests(self, request_ids, finished_status):
+        return [(leaving_id, 0)]
+
+    monkeypatch.setattr(scheduler_mod.VLLMScheduler, "finish_requests", fake_finish_requests)
+
+    scheduler_cls.finish_requests(
+        scheduler,
+        [leaving_id],
+        RequestStatus.FINISHED_STOPPED,
+    )
+
+    assert scheduler.running == [alive]

tests/core/sched/test_omni_scheduler_input_coordinator_cleanup.py

@@ -46,6 +46,9 @@ def test_finish_requests_cleans_input_coordinator_for_finished_ids(
     scheduler = scheduler_cls.__new__(scheduler_cls)
     scheduler.chunk_transfer_adapter = None
     scheduler.input_coordinator = coordinator
+    scheduler.requests = {}
+    scheduler.running = []
+    scheduler.waiting = []
 
     def fake_finish_requests(self, request_ids, finished_status):
         assert request_ids == ["req-a", "req-b"]

tests/dfx/perf/tests/test_higgs_audio_v3.json

@@ -0,0 +1,74 @@
+[
+  {
+    "test_name": "test_higgs_audio_v3",
+    "server_params": {
+      "model": "bosonai/higgs-audio-v3-tts-4b",
+      "stage_overrides": {
+        "0": {
+          "gpu_memory_utilization": 0.3
+        },
+        "1": {
+          "gpu_memory_utilization": 0.2
+        }
+      },
+      "extra_cli_args": [
+        "--deploy-config",
+        "vllm_omni/deploy/higgs_multimodal_qwen3.yaml",
+        "--trust-remote-code"
+      ]
+    },
+    "benchmark_params": [
+      {
+        "task": "default_voice",
+        "eval_phase": "latency",
+        "dataset_name": "seed-tts-text",
+        "backend": "openai-audio-speech",
+        "endpoint": "/v1/audio/speech",
+        "dataset_path": "benchmarks/build_dataset/seed_tts_smoke",
+        "num_prompts": [
+          20
+        ],
+        "max_concurrency": [
+          1
+        ],
+        "seed_tts_locale": "en",
+        "percentile-metrics": "ttft,e2el,audio_rtf,audio_ttfp,audio_duration",
+        "baseline": {
+          "median_audio_ttfp_ms": [
+            3000
+          ],
+          "median_audio_rtf": [
+            0.5
+          ]
+        }
+      },
+      {
+        "task": "default_voice",
+        "eval_phase": "throughput",
+        "dataset_name": "seed-tts-text",
+        "backend": "openai-audio-speech",
+        "endpoint": "/v1/audio/speech",
+        "dataset_path": "benchmarks/build_dataset/seed_tts_smoke",
+        "num_prompts": [
+          500
+        ],
+        "max_concurrency": [
+          16
+        ],
+        "seed_tts_locale": "en",
+        "percentile-metrics": "ttft,e2el,audio_rtf,audio_ttfp,audio_duration",
+        "baseline": {
+          "median_audio_ttfp_ms": [
+            5000
+          ],
+          "median_audio_rtf": [
+            1.2
+          ],
+          "audio_throughput": [
+            30
+          ]
+        }
+      }
+    ]
+  }
+]

tests/dfx/perf/tests/test_tts.json

@@ -339,5 +339,66 @@
         }
       }
     ]
+  },
+  {
+    "test_name": "test_voxtral_tts",
+    "server_params": {
+      "model": "mistralai/Voxtral-4B-TTS-2603"
+    },
+    "benchmark_params": [
+      {
+        "task": "default_voice",
+        "eval_phase": "latency",
+        "dataset_name": "seed-tts-text",
+        "backend": "openai-audio-speech",
+        "endpoint": "/v1/audio/speech",
+        "dataset_path": "benchmarks/build_dataset/seed_tts_smoke",
+        "num_prompts": [20],
+        "max_concurrency": [1],
+        "seed_tts_locale": "en",
+        "extra_body": {"voice": "casual_male"},
+        "percentile-metrics": "ttft,e2el,audio_rtf,audio_ttfp,audio_duration",
+        "baseline": {
+          "median_audio_ttfp_ms": [150],
+          "median_audio_rtf": [0.15]
+        }
+      },
+      {
+        "task": "default_voice",
+        "eval_phase": "throughput",
+        "dataset_name": "seed-tts-text",
+        "backend": "openai-audio-speech",
+        "endpoint": "/v1/audio/speech",
+        "dataset_path": "benchmarks/build_dataset/seed_tts_smoke",
+        "num_prompts": [80],
+        "max_concurrency": [8],
+        "seed_tts_locale": "en",
+        "extra_body": {"voice": "casual_male"},
+        "percentile-metrics": "ttft,e2el,audio_rtf,audio_ttfp,audio_duration",
+        "baseline": {
+          "median_audio_ttfp_ms": [1500],
+          "median_audio_rtf": [0.30],
+          "audio_throughput": [30.0]
+        }
+      },
+      {
+        "task": "default_voice",
+        "eval_phase": "stress",
+        "dataset_name": "seed-tts-text",
+        "backend": "openai-audio-speech",
+        "endpoint": "/v1/audio/speech",
+        "dataset_path": "benchmarks/build_dataset/seed_tts_smoke",
+        "num_prompts": [100],
+        "request_rate": [2.0],
+        "seed_tts_locale": "en",
+        "extra_body": {"voice": "casual_male"},
+        "percentile-metrics": "ttft,e2el,audio_rtf,audio_ttfp,audio_duration",
+        "baseline": {
+          "median_audio_ttfp_ms": [3000],
+          "median_audio_rtf": [0.25],
+          "audio_throughput": [4.0]
+        }
+      }
+    ]
   }
 ]

tests/dfx/reliability/invalid_param_test/test_invalid_omni_chat.py

@@ -198,19 +198,78 @@ def test_video_chat_stream_invalid_requests(
 # WS /v1/realtime
 # ─────────────────────────────────────────────────────────────────────────────
 
+# Resolved in ``test_realtime_invalid_requests`` (skip ``session.created`` handshake).
+_REALTIME_WS_MODEL_MISMATCH = object()
+_REALTIME_WS_INVALID_AUDIO_APPEND = object()
+
 
 @hardware_test(res={"cuda": "H100", "rocm": "MI325"}, num_cards=2)
+@pytest.mark.parametrize(
+    "send_frames_spec, ws_error_code, err_message",
+    [
+        pytest.param("{not-json", "invalid_json", "Invalid JSON", id="invalid_json"),
+        pytest.param(
+            json.dumps({"type": "session.update"}),
+            "invalid_event",
+            ("model", "Missing required field"),
+            id="session_update_missing_model",
+        ),
+        pytest.param(
+            _REALTIME_WS_MODEL_MISMATCH,
+            "model_not_found",
+            ("model", "does not exist"),
+            id="session_update_model_mismatch",
+        ),
+        pytest.param(
+            json.dumps({"type": "input_audio_buffer.commit", "final": False}),
+            "model_not_validated",
+            ("session.update", "validate the model"),
+            id="commit_without_model_validation",
+        ),
+        pytest.param(
+            json.dumps({"type": "unknown.event"}),
+            "unknown_event",
+            "Unknown event type",
+            id="unknown_event_type",
+        ),
+        pytest.param(
+            _REALTIME_WS_INVALID_AUDIO_APPEND,
+            "invalid_audio",
+            "Invalid audio data",
+            id="invalid_audio_append",
+        ),
+    ],
+)
 @pytest.mark.parametrize("omni_server", _QWEN3_OMNI_SERVER, indirect=True)
-def test_realtime_rejected_when_async_chunk_enabled(
+def test_realtime_invalid_requests(
     omni_server: OmniServer,
     openai_client: OpenAIClientHandler,
+    send_frames_spec: Any,
+    ws_error_code: str,
+    err_message: str | tuple[str, ...],
 ) -> None:
+    if send_frames_spec is _REALTIME_WS_MODEL_MISMATCH:
+        send_frames = json.dumps(
+            {
+                "type": "session.update",
+                "model": "this-model-is-not-served-on-this-instance",
+            }
+        )
+    elif send_frames_spec is _REALTIME_WS_INVALID_AUDIO_APPEND:
+        send_frames = [
+            json.dumps({"type": "session.update", "model": omni_server.model}),
+            json.dumps({"type": "input_audio_buffer.append", "audio": "not-valid-base64!!!"}),
+        ]
+    else:
+        send_frames = send_frames_spec
     openai_client.send_realtime_ws_request(
         {
+            "send_frames": send_frames,
             "timeout": 120,
             "ws_max_size": None,
+            "ws_skip_types": ["session.created"],
             "ws_json_type": "error",
-            "ws_error_code": "unsupported",
-            "err_message": "async_chunk",
+            "ws_error_code": ws_error_code,
+            "err_message": err_message,
         }
     )

tests/dfx/reliability/test_reliability_qwen3_omni.py

@@ -319,6 +319,7 @@ QWEN_PARAMS = create_reliability_omni_server_params(RELIABILITY_SCENARIOS, DEPLO
 
 
 @pytest.mark.slow
+@pytest.mark.skip(reason="issue#4285")
 @hardware_test(res={"cuda": "H100"}, num_cards=2)
 @pytest.mark.parametrize("omni_server_function", QWEN_PARAMS, indirect=True)
 def test_reliability_fault_gpu_oom_error_contract_consistent_chat_speech(
@@ -387,6 +388,7 @@ def test_reliability_fault_gpu_oom_error_contract_consistent_chat_speech(
 
 
 @pytest.mark.slow
+@pytest.mark.skip(reason="issue#4285")
 @hardware_test(res={"cuda": "H100"}, num_cards=2)
 @pytest.mark.parametrize("omni_server_function", QWEN_PARAMS, indirect=True)
 def test_reliability_fault_gpu_oom_chat_large_payload_failure(omni_server_function, openai_client_function) -> None:
@@ -429,6 +431,7 @@ def test_reliability_fault_gpu_oom_chat_large_payload_failure(omni_server_functi
 
 
 @pytest.mark.slow
+@pytest.mark.skip(reason="issue#4285")
 @hardware_test(res={"cuda": "H100"}, num_cards=2)
 @pytest.mark.parametrize("omni_server_function", QWEN_PARAMS, indirect=True)
 def test_reliability_fault_gpu_oom_concurrent_pressure_failure(omni_server_function, openai_client_function) -> None:
@@ -771,7 +774,7 @@ def test_reliability_fault_process_kill_tree_with_load_fast_fail_and_cleanup(
 
 
 @pytest.mark.slow
-@pytest.mark.skip(reason="OOM injection causes the serving instance to exit")
+@pytest.mark.skip(reason="issue#4285")
 @hardware_test(res={"cuda": "H100"}, num_cards=2)
 @pytest.mark.parametrize("omni_server_function", QWEN_PARAMS, indirect=True)
 def test_reliability_fault_gpu_oom_state_converges_after_fault_removed(

tests/diffusion/layers/test_norm.py

@@ -578,12 +578,27 @@ def test_rmsnorm_fused_native_parity_2d_input():
 
 
 def test_rmsnorm_fused_native_parity_custom_weight():
-    """Verify fused and native RMSNorm parity with non-default weights."""
+    """Verify fused and native RMSNorm parity with non-default weights.
+
+    Note: bfloat16 requires wider tolerance because the fused CUDA kernel
+    (vllm._custom_ops.rms_norm) computes in bf16 precision internally,
+    while forward_native uses float32 intermediates. With random custom
+    weights (not the default ones-weight), this precision gap exceeds 1e-3.
+
+    The gap surfaced after upstream vLLM routed RMSNorm through the vLLM IR
+    op (ir.ops.rms_norm) — see upstream commit 40bb17502
+    "[vLLM IR] 1/N Implement IR skeleton and rms_norm op (#33825)" — which
+    changed which fused kernel/precision path the layer dispatches to.
+    """
     from vllm_omni.diffusion.layers.norm import RMSNorm
 
     hidden_size = 64
     eps = 1e-6
 
+    # Per-dtype tolerances: bf16 fused kernel has inherent precision limit
+    atol_map = {torch.float32: 1e-3, torch.float16: 1e-3, torch.bfloat16: 2e-2}
+    rtol_map = {torch.float32: 1e-3, torch.float16: 1e-3, torch.bfloat16: 1e-2}
+
     for dtype in [torch.float32, torch.float16, torch.bfloat16]:
         torch.manual_seed(789)
         norm = RMSNorm(hidden_size, eps=eps).cuda()
@@ -594,4 +609,4 @@ def test_rmsnorm_fused_native_parity_custom_weight():
         out_fused = norm._forward_fused(x)
         out_native = norm.forward_native(x)
 
-        torch.testing.assert_close(out_fused, out_native, atol=1e-3, rtol=1e-3)
+        torch.testing.assert_close(out_fused, out_native, atol=atol_map[dtype], rtol=rtol_map[dtype])