Bump astral-sh/setup-uv from 7.6.0 to 9.0.0

Bumps [astral-sh/setup-uv](https://github.com/astral-sh/setup-uv) from 7.6.0 to 9.0.0.
- [Release notes](https://github.com/astral-sh/setup-uv/releases)
- [Commits](https://github.com/astral-sh/setup-uv/compare/37802adc94f370d6bfd71619e3f0bf239e1f3b78...c771a70e6277c0a99b617c7a806ffedaca235ff9)

---
updated-dependencies:
- dependency-name: astral-sh/setup-uv
  dependency-version: 8.3.2
  dependency-type: direct:production
  update-type: version-update:semver-major
...

Signed-off-by: dependabot[bot] <support@github.com>

.buildkite/check-torch-abi.py

@@ -0,0 +1,102 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+"""Audit vLLM compiled libraries for PyTorch stable ABI compliance."""
+
+import fnmatch
+import sys
+from pathlib import Path
+
+from torch_abi_audit import inspect_package
+from torch_abi_audit.report import ExtensionReport, PackageReport
+
+# Temporary allowlist of extensions not yet on the stable ABI.
+# Shrink and remove over time.
+ALLOWED_UNSTABLE_LIBRARIES: tuple[str, ...] = (
+    "vllm_flash_attn/_vllm_fa2_C.abi3.so",
+    "vllm_flash_attn/_vllm_fa3_C.abi3.so",
+    "third_party/deep_gemm/_C*.so",
+)
+
+
+def _relative_path(lib: ExtensionReport, package_root: Path) -> str:
+    try:
+        return lib.path.relative_to(package_root).as_posix()
+    except ValueError:
+        return lib.path.name
+
+
+def _is_torch_unstable(lib: ExtensionReport) -> bool:
+    return lib.error is None and lib.torch.uses_torch and not lib.torch.stable
+
+
+def _matches_allowlist(rel_path: str, patterns: tuple[str, ...]) -> bool:
+    return any(fnmatch.fnmatch(rel_path, pattern) for pattern in patterns)
+
+
+def _iter_libs(report: PackageReport) -> tuple[ExtensionReport, ...]:
+    return (*report.extensions, *report.bundled_libs)
+
+
+def _collect_unstable(report: PackageReport) -> list[str]:
+    return sorted(
+        _relative_path(lib, report.root)
+        for lib in _iter_libs(report)
+        if _is_torch_unstable(lib)
+    )
+
+
+def _find_stale_allowlist_entries(
+    report: PackageReport, patterns: tuple[str, ...]
+) -> list[str]:
+    """Allowlist patterns that match a built library which is no longer unstable."""
+    stale: list[str] = []
+    for pattern in patterns:
+        for lib in _iter_libs(report):
+            if lib.error is not None:
+                continue
+            if not fnmatch.fnmatch(_relative_path(lib, report.root), pattern):
+                continue
+            if not _is_torch_unstable(lib):
+                stale.append(pattern)
+                break
+    return stale
+
+
+def check_torch_abi(
+    package: str = "vllm",
+    patterns: tuple[str, ...] = ALLOWED_UNSTABLE_LIBRARIES,
+) -> int:
+    report = inspect_package(package)
+    if report.error:
+        print(f"error: failed to inspect {package!r}: {report.error}", file=sys.stderr)
+        return 2
+
+    unstable = _collect_unstable(report)
+    unexpected = [
+        rel_path for rel_path in unstable if not _matches_allowlist(rel_path, patterns)
+    ]
+    stale = _find_stale_allowlist_entries(report, patterns)
+
+    if unexpected or stale:
+        if unexpected:
+            print(
+                "Not allowed: torch-unstable libraries outside "
+                f"ALLOWED_UNSTABLE_LIBRARIES: {', '.join(unexpected)}",
+                file=sys.stderr,
+            )
+        if stale:
+            print(
+                "Not allowed: stale ALLOWED_UNSTABLE_LIBRARIES entries: "
+                f"{', '.join(stale)}",
+                file=sys.stderr,
+            )
+        return 1
+
+    print("Torch stable ABI check passed.")
+    return 0
+
+
+if __name__ == "__main__":
+    print(">>> Auditing vLLM extension modules for PyTorch stable ABI compliance")
+    sys.exit(check_torch_abi())

.buildkite/ci_config.yaml

@@ -14,6 +14,7 @@ run_all_patterns:
   - "setup.py"
   - "csrc/"
   - "cmake/"
+  - ".buildkite/check-torch-abi.py"
 run_all_exclude_patterns:
   - "docker/Dockerfile."
   - "csrc/cpu/"

.buildkite/intel_jobs/benchmarks_intel.yaml

@@ -0,0 +1,26 @@
+group: Benchmarks
+depends_on: 
+  - image-build-xpu
+steps:
+- label: Benchmarks CLI Test
+  key: benchmarks-cli-test
+  timeout_in_minutes: 40
+  device: intel_gpu
+  agent_tags:
+    label: production
+    gpu: 1+
+    mem: 16+
+  no_plugin: true
+  working_dir: "."
+  env:
+    REGISTRY: "public.ecr.aws/q9t5s3a7"
+    REPO: "vllm-ci-test-repo"
+    VLLM_TEST_DEVICE: "xpu"
+  source_file_dependencies:
+  - vllm/
+  - tests/benchmarks/
+  commands:
+    - >-
+      bash .buildkite/scripts/hardware_ci/run-intel-test.sh
+      'cd tests &&
+      pytest -v -s benchmarks/'

.buildkite/intel_jobs/engine_intel.yaml

@@ -2,6 +2,44 @@ group: Engine Intel
 depends_on:
   - image-build-xpu
 steps:
+- label: Engine
+  key: engine
+  timeout_in_minutes: 40
+  device: intel_gpu
+  agent_tags:
+    label: production
+    gpu: 1+
+    mem: 16+
+  no_plugin: true
+  working_dir: "."
+  env:
+    REGISTRY: "public.ecr.aws/q9t5s3a7"
+    REPO: "vllm-ci-test-repo"
+    VLLM_TEST_DEVICE: "xpu"
+  source_file_dependencies:
+  - vllm/compilation/
+  - vllm/config/
+  - vllm/engine/
+  - vllm/entrypoints/logger.py
+  - vllm/envs.py
+  - vllm/logger.py
+  - vllm/logging_utils/
+  - vllm/platforms/
+  - vllm/sequence.py
+  - vllm/triton_utils/
+  - vllm/utils/
+  - tests/engine
+  - tests/test_sequence
+  - tests/test_config
+  - tests/test_logger
+  - tests/test_vllm_port
+  - tests/test_jit_monitor.py
+  commands:
+    - >-
+      bash .buildkite/scripts/hardware_ci/run-intel-test.sh
+      'cd tests &&
+      pytest -v -s engine/test_arg_utils.py test_sequence.py test_logger.py test_vllm_port.py test_jit_monitor.py'
+
 - label: Engine (1 GPU)
   timeout_in_minutes: 30
   device: intel_gpu
@@ -23,3 +61,41 @@ steps:
       bash .buildkite/scripts/hardware_ci/run-intel-test.sh
       'cd tests &&
       pytest -v -s v1/engine --ignore v1/engine/test_preprocess_error_handling.py'
+
+- label: V1 e2e (2 GPUs)
+  timeout_in_minutes: 30
+  device: intel_gpu
+  agent_tags:
+    label: production
+    gpu: 2+
+    mem: 16+
+  no_plugin: true
+  working_dir: "."
+  env:
+    REGISTRY: "public.ecr.aws/q9t5s3a7"
+    REPO: "vllm-ci-test-repo"
+    VLLM_TEST_DEVICE: "xpu"
+  source_file_dependencies:
+    - vllm/compilation/
+    - vllm/config/
+    - vllm/distributed/
+    - vllm/engine/
+    - vllm/envs.py
+    - vllm/forward_context.py
+    - vllm/inputs/
+    - vllm/logger.py
+    - vllm/logging_utils/
+    - vllm/model_executor/
+    - vllm/multimodal/
+    - vllm/platforms/
+    - vllm/sampling_params.py
+    - vllm/transformers_utils/
+    - vllm/triton_utils/
+    - vllm/utils/
+    - vllm/v1/
+    - tests/v1/e2e/spec_decode
+  commands:
+  - >-
+      bash .buildkite/scripts/hardware_ci/run-intel-test.sh
+      'cd tests &&
+      pytest -v -s v1/e2e/spec_decode/test_spec_decode.py -k "tensor_parallelism"'

.buildkite/intel_jobs/misc_intel.yaml

@@ -125,13 +125,13 @@ steps:
       pytest -v -s v1/kv_offload &&
       pytest -v -s v1/kv_connector/unit/test_offloading_connector.py'
 
-- label: NixlConnector PD accuracy (2 GPUs)
+- label: NixlConnector PD accuracy (4 GPUs)
   timeout_in_minutes: 60
-  num_devices: 2
+  num_devices: 4
   device: intel_gpu
   agent_tags:
     label: production
-    gpu: 2+
+    gpu: 4+
     mem: 16+
   no_plugin: true
   working_dir: "."
@@ -148,7 +148,10 @@ steps:
     - >-
       bash .buildkite/scripts/hardware_ci/run-intel-test.sh
       'cd tests &&
-      bash v1/kv_connector/nixl_integration/run_xpu_disagg_accuracy_test.sh'
+      bash v1/kv_connector/nixl_integration/run_xpu_disagg_accuracy_test.sh &&
+      PREFILLER_TP_SIZE=2 DECODER_TP_SIZE=1 bash v1/kv_connector/nixl_integration/run_xpu_disagg_accuracy_test.sh &&
+      PREFILLER_TP_SIZE=1 DECODER_TP_SIZE=2 bash v1/kv_connector/nixl_integration/run_xpu_disagg_accuracy_test.sh &&
+      PREFILLER_TP_SIZE=2 DECODER_TP_SIZE=2 bash v1/kv_connector/nixl_integration/run_xpu_disagg_accuracy_test.sh'
 
 - label: Regression
   key: regression
@@ -259,3 +262,25 @@ steps:
     pytest -v -s detokenizer &&
     pytest -v -s -m "not cpu_test" ./multimodal &&
     pytest -v -s utils_ --ignore=utils_/test_mem_utils.py'
+
+- label: Fusion Unit Tests
+  timeout_in_minutes: 30
+  device: intel_gpu
+  agent_tags:
+    label: production
+    gpu: 1+
+    mem: 16+
+  no_plugin: true
+  working_dir: "."
+  env:
+    REGISTRY: "public.ecr.aws/q9t5s3a7"
+    REPO: "vllm-ci-test-repo"
+    VLLM_TEST_DEVICE: "xpu"
+  source_file_dependencies:
+    - vllm/compilation/
+    - tests/compile/passes/test_qk_norm_rope_fusion.py
+  commands:
+    - >-
+      bash .buildkite/scripts/hardware_ci/run-intel-test.sh
+      'cd tests &&
+      pytest -v -s compile/passes/test_qk_norm_rope_fusion.py'

.buildkite/intel_jobs/model_executor_intel.yaml

@@ -0,0 +1,33 @@
+group: Model Executor Intel
+depends_on:
+  - image-build-xpu
+steps:
+- label: Model Executor (Intel)
+  key: model-executor-intel
+  timeout_in_minutes: 45
+  device: intel_gpu
+  agent_tags:
+    label: production
+    gpu: 1+
+    mem: 24+
+  no_plugin: true
+  working_dir: "."
+  env:
+    REGISTRY: "public.ecr.aws/q9t5s3a7"
+    REPO: "vllm-ci-test-repo"
+    VLLM_TEST_DEVICE: "xpu"
+  source_file_dependencies:
+  - vllm/engine/arg_utils.py
+  - vllm/config/model.py
+  - vllm/model_executor
+  - tests/model_executor
+  commands:
+    - >-
+      bash .buildkite/scripts/hardware_ci/run-intel-test.sh
+      'apt-get update && apt-get install -y curl libsodium23 &&
+      pip3 install tensorizer==2.10.1 &&
+      pip3 install runai-model-streamer[s3,gcs,azure]\>=0.15.7 &&
+      export VLLM_WORKER_MULTIPROC_METHOD=spawn &&
+      export PYTHONFAULTHANDLER=1 &&
+      cd tests &&
+      pytest -v -s model_executor -m "not slow_test" --ignore="model_executor/layers/test_rocm_unquantized_gemm.py" --deselect="tests/model_executor/model_loader/test_reload.py::test_kv_scale_reload"'

.buildkite/intel_jobs/model_runner_v2_intel.yaml

@@ -8,7 +8,7 @@ steps:
   agent_tags:
     label: production
     gpu: 2+
-    mem: 16+
+    mem: 24+
   no_plugin: true
   working_dir: "."
   env:
@@ -28,7 +28,9 @@ steps:
       'export VLLM_USE_V2_MODEL_RUNNER=1 &&
       cd tests &&
       pytest -v -s v1/engine/test_llm_engine.py -k "not test_engine_metrics" &&
+      pytest -v -s v1/e2e/general/test_context_length.py &&
       ENFORCE_EAGER=1 pytest -v -s v1/e2e/general/test_async_scheduling.py -k "not ngram" &&
+      pytest -v -s entrypoints/llm/test_struct_output_generate.py -k "xgrammar and not speculative_config6 and not speculative_config7 and not speculative_config8 and not speculative_config0" &&
       pytest -v -s v1/e2e/general/test_min_tokens.py'
 
 - label: Model Runner V2 Examples (Intel)
@@ -60,3 +62,55 @@ steps:
       python3 basic/offline_inference/generate.py --model facebook/opt-125m &&
       python3 generate/multimodal/vision_language_offline.py --seed 0 &&
       python3 features/automatic_prefix_caching/prefix_caching_offline.py'
+
+- label: Model Runner V2 Distributed (2 GPUs)
+  timeout_in_minutes: 50
+  device: intel_gpu
+  agent_tags:
+    label: production
+    gpu: 2+
+    mem: 16+
+  no_plugin: true
+  working_dir: "."
+  env:
+    REGISTRY: "public.ecr.aws/q9t5s3a7"
+    REPO: "vllm-ci-test-repo"
+    VLLM_TEST_DEVICE: "xpu"
+  source_file_dependencies:
+    - vllm/v1/worker/gpu/
+    - vllm/v1/worker/gpu_worker.py
+    - tests/basic_correctness/test_basic_correctness.py
+    - tests/v1/distributed/test_async_llm_dp.py
+    - tests/v1/distributed/test_eagle_dp.py
+  commands:
+    - >-
+      bash .buildkite/scripts/hardware_ci/run-intel-test.sh
+      'export VLLM_USE_V2_MODEL_RUNNER=1 &&
+      cd tests &&
+      TARGET_TEST_SUITE=L4 pytest -v -s basic_correctness/test_basic_correctness.py -m "distributed\(num_gpus=2\)" -k "not ray and not True"'
+
+- label: Model Runner V2 Spec Decode
+  timeout_in_minutes: 50
+  device: intel_gpu
+  agent_tags:
+    label: production
+    gpu: 1+
+    mem: 24+
+  no_plugin: true
+  working_dir: "."
+  env:
+    REGISTRY: "public.ecr.aws/q9t5s3a7"
+    REPO: "vllm-ci-test-repo"
+    VLLM_TEST_DEVICE: "xpu"
+  source_file_dependencies:
+  - vllm/v1/worker/gpu/
+  - vllm/v1/worker/gpu_worker.py
+  - tests/v1/spec_decode/test_max_len.py
+  - tests/v1/spec_decode/test_rejection_sampler_utils.py
+  - tests/v1/e2e/spec_decode/test_spec_decode.py
+  commands:
+    - >-
+      bash .buildkite/scripts/hardware_ci/run-intel-test.sh
+      'export VLLM_USE_V2_MODEL_RUNNER=1 &&
+      cd tests &&
+      pytest -v -s v1/spec_decode/test_synthetic_rejection_sampler_utils.py'

.buildkite/intel_jobs/samplers_intel.yaml

@@ -0,0 +1,29 @@
+group: Samplers Intel
+depends_on:
+  - image-build-xpu
+steps:
+- label: Samplers Test (FlashInfer)
+  key: samplers-test-flashinfer-intel
+  timeout_in_minutes: 40
+  device: intel_gpu
+  agent_tags:
+    label: production
+    gpu: 1+
+    mem: 24+
+  no_plugin: true
+  working_dir: "."
+  env:
+    REGISTRY: "public.ecr.aws/q9t5s3a7"
+    REPO: "vllm-ci-test-repo"
+    VLLM_TEST_DEVICE: "xpu"
+  source_file_dependencies:
+  - vllm/model_executor/layers
+  - vllm/sampling_metadata.py
+  - tests/samplers
+  - tests/conftest.py
+  - vllm/entrypoints/generate/beam_search
+  commands:
+    - >-
+      bash .buildkite/scripts/hardware_ci/run-intel-test.sh
+      'cd tests &&
+      VLLM_USE_FLASHINFER_SAMPLER=1 pytest -v -s samplers'

.buildkite/intel_jobs/test-intel.yaml

@@ -145,7 +145,8 @@ steps:
       - >-
         bash .buildkite/scripts/hardware_ci/run-intel-test.sh
         'cd tests &&
-        pytest -v -s quantization/test_auto_round.py'
+        pytest -v -s quantization/test_auto_round.py &&
+        pytest -v -s quantization/test_online.py'
   - label: "XPU compressed tensors FP8 test"
     depends_on:
       - image-build-xpu
@@ -168,4 +169,4 @@ steps:
       - >-
         bash .buildkite/scripts/hardware_ci/run-intel-test.sh
         'cd tests &&
-        pytest -v -s quantization/test_compressed_tensors.py::test_compressed_tensors_fp8'
+        pytest -v -s quantization/test_compressed_tensors.py::test_compressed_tensors_fp8'

.buildkite/scripts/build-macos-wheel.sh

@@ -7,6 +7,9 @@
 
 set -euo pipefail
 
+# The macmini queue uses persistent checkouts, so refresh tags for setuptools-scm.
+git fetch --tags --force origin
+
 # The Rust frontend build needs protoc.
 if ! command -v protoc >/dev/null 2>&1; then
   brew install protobuf

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

@@ -387,6 +387,7 @@ initialize_native_environment() {
   local job_id="${BUILDKITE_JOB_ID:-${BUILDKITE_PARALLEL_JOB:-local}}"
   local job_id_suffix=""
   local native_root=""
+  local hf_fstype=""
   local hf_mount=""
 
   if [[ "$(id -u)" -ne 0 ]]; then
@@ -405,11 +406,14 @@ initialize_native_environment() {
   VLLM_CACHE_ROOT="${native_root}/cache/vllm"
   XDG_CACHE_HOME="${native_root}/cache/xdg"
   : "${HF_HOME:=/home/buildkite-agent/huggingface}"
+  # datasets uses POSIX locks that are unsupported by the shared HF NFS cache.
+  # Keep processed datasets job-local while retaining the persistent Hub cache.
+  HF_DATASETS_CACHE="${native_root}/cache/huggingface/datasets"
   : "${HF_HUB_DOWNLOAD_TIMEOUT:=300}"
   : "${HF_HUB_ETAG_TIMEOUT:=60}"
   export TMPDIR VLLM_RPC_BASE_PATH
   export TORCHINDUCTOR_CACHE_DIR TRITON_CACHE_DIR VLLM_CACHE_ROOT XDG_CACHE_HOME
-  export HF_HOME HF_HUB_DOWNLOAD_TIMEOUT HF_HUB_ETAG_TIMEOUT
+  export HF_HOME HF_DATASETS_CACHE HF_HUB_DOWNLOAD_TIMEOUT HF_HUB_ETAG_TIMEOUT
   export PYTORCH_ROCM_ARCH=""
 
   mkdir -p "${TMPDIR}" \
@@ -417,7 +421,8 @@ initialize_native_environment() {
     "${TRITON_CACHE_DIR}" \
     "${VLLM_CACHE_ROOT}" \
     "${XDG_CACHE_HOME}" \
-    "${HF_HOME}" || return 1
+    "${HF_HOME}" \
+    "${HF_DATASETS_CACHE}" || return 1
 
   echo "Native compile caches: VLLM_CACHE_ROOT=${VLLM_CACHE_ROOT} TORCHINDUCTOR_CACHE_DIR=${TORCHINDUCTOR_CACHE_DIR}"
 
@@ -432,6 +437,18 @@ initialize_native_environment() {
       return 1
     fi
   fi
+
+  if command -v findmnt >/dev/null 2>&1; then
+    hf_fstype=$(findmnt -n -T "${HF_HOME}" -o FSTYPE 2>/dev/null || true)
+  fi
+  if [[ "${hf_fstype}" == nfs || "${hf_fstype}" == nfs4 ]]; then
+    # Keep hf-xet state local and avoid vectored writes on shared NFS.
+    export HF_XET_CACHE="${native_root}/cache/hf-xet"
+    export HF_XET_HIGH_PERFORMANCE=0
+    export HF_XET_RECONSTRUCTION_USE_VECTORED_WRITE=0
+    mkdir -p "${HF_XET_CACHE}" || return 1
+    echo "Configured hf-xet for shared ${hf_fstype} cache at ${HF_HOME}"
+  fi
 }
 
 run_native_preflight() {

.buildkite/test-amd.yaml

@@ -169,20 +169,6 @@ steps:
   - pip install helion==1.1.0
   - pytest -v -s kernels/helion/
 
-- label: Kernels Mamba Test # TBD
-  timeout_in_minutes: 180
-  mirror_hardwares: [amdexperimental, amdproduction, amdgfx90anightly, amdmi250]
-  agent_pool: mi250_1
-  optional: true
-  working_dir: "/vllm-workspace/tests"
-  source_file_dependencies:
-  - csrc/mamba/
-  - tests/kernels/mamba
-  - vllm/model_executor/layers/mamba/ops
-  - vllm/platforms/rocm.py
-  commands:
-  - pytest -v -s kernels/mamba
-
 #------------------------------------------------------  mi250 · models / basic  -------------------------------------------------------#
 
 - label: Basic Models Test (Other CPU) # TBD
@@ -364,22 +350,6 @@ steps:
   commands:
   - pytest -v -s v1/e2e/spec_decode -k "speculators or mtp_correctness"
 
-- label: V1 attention (H100-MI250) # TBD
-  timeout_in_minutes: 180
-  mirror_hardwares: [amdexperimental, amdproduction, amdgfx90anightly, amdmi250]
-  agent_pool: mi250_1
-  working_dir: "/vllm-workspace/tests"
-  source_file_dependencies:
-  - vllm/config/attention.py
-  - vllm/model_executor/layers/attention
-  - vllm/v1/attention
-  - tests/v1/attention
-  - vllm/_aiter_ops.py
-  - vllm/envs.py
-  - vllm/platforms/rocm.py
-  commands:
-  - pytest -v -s v1/attention
-
 - label: V1 others (CPU) # TBD
   timeout_in_minutes: 180
   mirror_hardwares: [amdexperimental, amdproduction, amdgfx90anightly, amdmi250]
@@ -1577,11 +1547,12 @@ steps:
   commands:
   - pytest -v -s kernels/attention --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT
 
-- label: Kernels Core Operation Test # TBD
+- label: Kernels Core Operation Test %N # TBD
   timeout_in_minutes: 180
   mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
   dind: false
   agent_pool: mi300_1
+  parallelism: 3
   working_dir: "/vllm-workspace/tests"
   source_file_dependencies:
   - csrc/
@@ -1592,7 +1563,7 @@ steps:
   - vllm/_aiter_ops.py
   - vllm/platforms/rocm.py
   commands:
-  - pytest -v -s kernels/core --ignore=kernels/core/test_minimax_reduce_rms.py  kernels/test_concat_mla_q.py kernels/test_top_k_per_row.py
+  - pytest -v -s kernels/core --ignore=kernels/core/test_minimax_reduce_rms.py kernels/test_concat_mla_q.py kernels/test_top_k_per_row.py --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT
 
 - label: Kernels KDA Test # TBD
   timeout_in_minutes: 180
@@ -1610,6 +1581,21 @@ steps:
   commands:
   - pytest -v -s kernels/test_kda.py
 
+- label: Kernels Mamba Test # TBD
+  timeout_in_minutes: 180
+  mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
+  dind: false
+  agent_pool: mi300_1
+  optional: true
+  working_dir: "/vllm-workspace/tests"
+  source_file_dependencies:
+  - csrc/mamba/
+  - tests/kernels/mamba
+  - vllm/model_executor/layers/mamba/ops
+  - vllm/platforms/rocm.py
+  commands:
+  - pytest -v -s kernels/mamba
+
 - label: Kernels MoE Test %N # TBD
   timeout_in_minutes: 180
   mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
@@ -2161,7 +2147,7 @@ steps:
   mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
   dind: false
   agent_pool: mi300_1
-  parallelism: 4
+  parallelism: 8
   optional: true
   working_dir: "/vllm-workspace/"
   source_file_dependencies:
@@ -2694,11 +2680,12 @@ steps:
   - export VLLM_WORKER_MULTIPROC_METHOD=spawn
   - pytest -v -s v1/kv_connector/extract_hidden_states_integration
 
-- label: V1 attention (H100-MI300) # TBD
+- label: V1 attention (H100-MI300) %N # TBD
   timeout_in_minutes: 180
   mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
   dind: false
   agent_pool: mi300_1
+  parallelism: 2
   optional: true
   working_dir: "/vllm-workspace/tests"
   source_file_dependencies:
@@ -2710,7 +2697,7 @@ steps:
   - vllm/envs.py
   - vllm/platforms/rocm.py
   commands:
-  - pytest -v -s v1/attention
+  - pytest -v -s v1/attention --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT
 
 - label: V1 Core + KV + Metrics # TBD
   timeout_in_minutes: 180
@@ -3676,11 +3663,12 @@ steps:
 
 #------------------------------------------------------------  mi355 · v1  -------------------------------------------------------------#
 
-- label: V1 attention (B200-MI355) # TBD
+- label: V1 attention (B200-MI355) %N # TBD
   timeout_in_minutes: 180
   mirror_hardwares: [amdexperimental, amdproduction, amdgfx950nightly, amdmi355]
   dind: false
   agent_pool: mi355_1
+  parallelism: 2
   working_dir: "/vllm-workspace/tests"
   source_file_dependencies:
   - vllm/config/attention.py
@@ -3691,7 +3679,7 @@ steps:
   - vllm/envs.py
   - vllm/platforms/rocm.py
   commands:
-  - pytest -v -s v1/attention
+  - pytest -v -s v1/attention --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT
 
 - label: V1 Core + KV + Metrics # TBD
   timeout_in_minutes: 180

.buildkite/test_areas/cuda.yaml

@@ -16,6 +16,7 @@ steps:
   commands:
     - pytest -v -s cuda/test_cuda_context.py
     - pytest -v -s cuda/test_platform_no_cuda_init.py
+    - pytest -v -s cuda/test_cuda_compatibility_path.py
 
 - label: Cudagraph
   device: h200_35gb

.buildkite/test_areas/disaggregated.yaml

@@ -131,6 +131,22 @@ steps:
         - uv pip install --system -r /vllm-workspace/requirements/kv_connectors_rocm.txt
         - HYBRID_SSM=1 ATTENTION_BACKEND=TRITON_ATTN bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
 
+- label: NixlConnector PD edge case test (2 GPUs)
+  key: nixlconnector-pd-edge-cases-2-gpus
+  timeout_in_minutes: 40
+  working_dir: "/vllm-workspace/tests"
+  num_devices: 2
+  source_file_dependencies:
+    - vllm/distributed/kv_transfer/kv_connector/v1/nixl/
+    - vllm/v1/core/sched/
+    - tests/v1/kv_connector/nixl_integration/
+  env:
+    PREFILL_GPU_ID: "0"
+    DECODE_GPU_ID: "1"
+  commands:
+    - bash /vllm-workspace/.buildkite/scripts/install-kv-connectors.sh
+    - bash v1/kv_connector/nixl_integration/run_edge_case_test.sh
+
 - label: Hybrid SSM NixlConnector PD prefix cache test (2 GPUs)
   key: hybrid-ssm-nixlconnector-pd-prefix-cache-2-gpus
   timeout_in_minutes: 25

.buildkite/test_areas/engine.yaml

@@ -40,9 +40,11 @@ steps:
   source_file_dependencies:
     - vllm/v1/engine/
     - tests/v1/engine/
+    - tests/v1/test_tensor_ipc_queue.py
   commands:
     - pytest -v -s v1/engine/test_preprocess_error_handling.py
     - pytest -v -s v1/engine --ignore v1/engine/test_preprocess_error_handling.py
+    - pytest -v -s v1/test_tensor_ipc_queue.py
   mirror:
     amd:
       device: mi250_1

.buildkite/test_areas/expert_parallelism.yaml

@@ -52,4 +52,5 @@ steps:
   - vllm/compilation/
   - tests/distributed/
   commands:
+  - bash /vllm-workspace/.buildkite/scripts/install-kv-connectors.sh
   - pytest -v -s distributed/test_elastic_ep.py

.buildkite/test_areas/fault_tolerance.yaml

@@ -0,0 +1,26 @@
+group: Fault Tolerance
+depends_on:
+  - image-build
+steps:
+- label: Fault Tolerance E2E (2xH100)
+  key: fault-tolerance-e2e-2xh100
+  timeout_in_minutes: 35
+  device: h100
+  num_devices: 2
+  working_dir: "/vllm-workspace/tests"
+  source_file_dependencies:
+  - vllm/v1/fault_tolerance/
+  - vllm/v1/worker/sentinel/
+  - vllm/entrypoints/serve/fault_tolerance/
+  - vllm/distributed/elastic_ep/
+  - vllm/distributed/device_communicators/
+  - vllm/v1/engine/
+  - vllm/v1/worker/
+  - tests/v1/fault_tolerance/
+  - tests/v1/distributed/test_external_lb_dp.py
+  commands:
+  # Base image has no nixl; install it or has_nixl_ep() skips the tests.
+  - bash /vllm-workspace/.buildkite/scripts/install-kv-connectors.sh
+  # https://github.com/NVIDIA/nccl/issues/1838
+  - export NCCL_CUMEM_HOST_ENABLE=0
+  - pytest -v -s v1/fault_tolerance/test_fault_tolerance_e2e.py

.buildkite/test_areas/kernels.yaml

@@ -61,9 +61,45 @@ steps:
   source_file_dependencies:
   - csrc/fused_deepseek_v4_qnorm_rope_kv_insert_kernel.cu
   - vllm/models/deepseek_v4/common/ops/
+  - vllm/models/deepseek_v4/nvidia/
   - tests/kernels/test_fused_deepseek_v4_qnorm_rope_kv_insert.py
+  - tests/models/test_deepseek_v4_mega_moe.py
   commands:
     - pytest -v -s kernels/test_fused_deepseek_v4_*.py
+    - pytest -v -s models/test_deepseek_v4_mega_moe.py
+
+# Catch-all for test files at the tests/kernels root. This job collects
+# the whole root so new files are wired by default.
+# Files with dedicated jobs elsewhere in this file are excluded via --ignore
+# (test_kda, test_bf16x3_router_gemm_cutedsl and test_ll_bf16_gemm run in
+# their own jobs / Kernels (B200)).
+- label: Kernels Root Misc Test (B200)
+  key: kernels-root-misc-test-b200
+  timeout_in_minutes: 45
+  device: b200-k8s
+  source_file_dependencies:
+  - csrc/
+  - vllm/
+  - tests/kernels/
+  commands:
+    - pytest -v -s kernels/
+        --ignore=kernels/attention
+        --ignore=kernels/core
+        --ignore=kernels/helion
+        --ignore=kernels/ir
+        --ignore=kernels/mamba
+        --ignore=kernels/moe
+        --ignore=kernels/quantization
+        --ignore=kernels/test_concat_mla_q.py
+        --ignore=kernels/test_fused_qk_norm_rope_gate.py
+        --ignore=kernels/test_fused_deepseek_v4_qnorm_rope_kv_insert.py
+        --ignore=kernels/test_top_k_per_row.py
+        --ignore=kernels/test_kda.py
+        --ignore=kernels/test_bf16x3_router_gemm_cutedsl.py
+        --ignore=kernels/test_ll_bf16_gemm.py
+        --ignore=kernels/test_shuffle_rows.py
+    # BROKEN on main, pending kernel fixes (B200):
+    #   test_shuffle_rows.py (1: test_shuffle_rows_edge_cases)
 
 - label: Kernels Attention Test %N
   key: kernels-attention-test

.buildkite/test_areas/lm_eval.yaml

@@ -337,7 +337,7 @@ steps:
 
 - label: LM Eval KV-Offload (2xH100)
   key: kv-offload-medium
-  timeout_in_minutes: 30
+  timeout_in_minutes: 45
   device: h100
   num_devices: 2
   source_file_dependencies:
@@ -347,7 +347,7 @@ steps:
   - vllm/v1/simple_kv_offload/
   - tests/evals/gsm8k/test_gsm8k_offloading.py
   commands:
-  - pytest -s -v evals/gsm8k/test_gsm8k_offloading.py -k "qwen3.5-35b"
+  - pytest -s -v evals/gsm8k/test_gsm8k_offloading.py -k "qwen3.5-35b or deepseek-v2-lite"
 
 - label: LM Eval KV-Offload (4xH100)
   key: kv-offload-large

.buildkite/test_areas/lora.yaml

@@ -16,6 +16,7 @@ steps:
     amd:
       dind: false
       device: mi300_1
+      soft_fail: true
       working_dir: "/vllm-workspace/tests"
       timeout_in_minutes: 85
       source_file_dependencies:

.buildkite/test_areas/misc.yaml

@@ -148,6 +148,7 @@ steps:
     - pytest -v -s -m 'cpu_test' v1/core
     - pytest -v -s v1/structured_output
     - pytest -v -s v1/test_serial_utils.py
+    - pytest -v -s v1/test_kv_cache_spec_registry.py
     - pytest -v -s v1/cudagraph/test_cudagraph_manager.py
     - pytest -v -s -m 'cpu_test' v1/kv_connector/unit
     - pytest -v -s -m 'cpu_test' v1/metrics
@@ -265,6 +266,7 @@ steps:
   - vllm/utils/
   - vllm/v1/
   - tests/v1/tracing
+  - tests/tracing/
   commands:
   - "pip install \
       'opentelemetry-sdk>=1.26.0' \
@@ -272,6 +274,7 @@ steps:
       'opentelemetry-exporter-otlp>=1.26.0' \
       'opentelemetry-semantic-conventions-ai>=0.4.1'"
   - pytest -v -s v1/tracing
+  - pytest -v -s tracing
   mirror:
     amd:
       dind: false
@@ -295,6 +298,7 @@ steps:
     amd:
       device: mi250_1
       timeout_in_minutes: 55
+      soft_fail: true
       depends_on:
       - image-build-amd
       source_file_dependencies:
@@ -394,7 +398,7 @@ steps:
 
 - label: Batch Invariance (A100)
   key: batch-invariance-a100
-  timeout_in_minutes: 40
+  timeout_in_minutes: 60
   device: a100
   source_file_dependencies:
     - vllm/v1/attention
@@ -404,11 +408,11 @@ steps:
     - export VLLM_WORKER_MULTIPROC_METHOD=spawn
     - pip install pytest-timeout pytest-forked
     - pytest -v -s v1/determinism/test_batch_invariance.py
-    - VLLM_TEST_MODEL=deepseek-ai/DeepSeek-V2-Lite-Chat pytest -v -s v1/determinism/test_batch_invariance.py::test_v1_generation_is_deterministic_across_batch_sizes_with_needle[TRITON_MLA]
+    - VLLM_TEST_MODEL=deepseek-ai/DeepSeek-V2-Lite-Chat pytest -v -s v1/determinism/test_batch_invariance.py::test_v1_generation_is_deterministic_across_batch_sizes_with_needle -k TRITON_MLA
 
 - label: Batch Invariance (H100)
   key: batch-invariance-h100
-  timeout_in_minutes: 40
+  timeout_in_minutes: 60
   device: h100
   source_file_dependencies:
     - vllm/v1/attention
@@ -419,12 +423,12 @@ steps:
     - pip install pytest-timeout pytest-forked
     - pytest -v -s v1/determinism/test_batch_invariance.py
     - pytest -v -s v1/determinism/test_rms_norm_batch_invariant.py
-    - VLLM_TEST_MODEL=deepseek-ai/DeepSeek-V2-Lite-Chat pytest -v -s v1/determinism/test_batch_invariance.py::test_v1_generation_is_deterministic_across_batch_sizes_with_needle[TRITON_MLA]
-    - VLLM_TEST_MODEL=Qwen/Qwen3-30B-A3B-Thinking-2507-FP8 pytest -v -s v1/determinism/test_batch_invariance.py::test_v1_generation_is_deterministic_across_batch_sizes_with_needle[FLASH_ATTN]
+    - VLLM_TEST_MODEL=deepseek-ai/DeepSeek-V2-Lite-Chat pytest -v -s v1/determinism/test_batch_invariance.py::test_v1_generation_is_deterministic_across_batch_sizes_with_needle -k TRITON_MLA
+    - VLLM_TEST_MODEL=Qwen/Qwen3-30B-A3B-Thinking-2507-FP8 pytest -v -s v1/determinism/test_batch_invariance.py::test_v1_generation_is_deterministic_across_batch_sizes_with_needle -k FLASH_ATTN
 
 - label: Batch Invariance (B200)
   key: batch-invariance-b200
-  timeout_in_minutes: 35
+  timeout_in_minutes: 45
   device: b200-k8s
   source_file_dependencies:
     - vllm/v1/attention
@@ -435,11 +439,14 @@ steps:
     - pip install pytest-timeout pytest-forked
     - pytest -v -s v1/determinism/test_batch_invariance.py
     - pytest -v -s v1/determinism/test_rms_norm_batch_invariant.py
-    - VLLM_TEST_MODEL=deepseek-ai/DeepSeek-V2-Lite-Chat pytest -v -s v1/determinism/test_batch_invariance.py::test_v1_generation_is_deterministic_across_batch_sizes_with_needle[TRITON_MLA]
-    - VLLM_TEST_MODEL=Qwen/Qwen3-30B-A3B-Thinking-2507-FP8 pytest -v -s v1/determinism/test_batch_invariance.py::test_v1_generation_is_deterministic_across_batch_sizes_with_needle[FLASH_ATTN]
+    - VLLM_TEST_MODEL=deepseek-ai/DeepSeek-V2-Lite-Chat pytest -v -s v1/determinism/test_batch_invariance.py::test_v1_generation_is_deterministic_across_batch_sizes_with_needle -k TRITON_MLA
+    - VLLM_TEST_MODEL=Qwen/Qwen3-30B-A3B-Thinking-2507-FP8 pytest -v -s v1/determinism/test_batch_invariance.py::test_v1_generation_is_deterministic_across_batch_sizes_with_needle -k FLASH_ATTN
     - pytest -v -s v1/determinism/test_nvfp4_batch_invariant.py
     - pytest -v -s v1/determinism/test_nvfp4_batch_invariant_scaled_mm.py
-  
+    - pytest -v -s v1/determinism/test_matmul_batch_invariant.py
+    - pytest -v -s v1/determinism/test_cutlass_batch_invariance.py
+    - pytest -v -s v1/determinism/test_online_batch_invariance.py
+
 - label: Acceptance Length Test (Large Models) # optional
   device: h200_35gb
   key: acceptance-length-test-large-models

.buildkite/test_areas/models_basic.yaml

@@ -61,6 +61,18 @@ steps:
     # FA4 kernel tests require SM100; the suite skips them elsewhere.
     - pytest -v -s models/inkling
 
+- label: Kimi K3 Unit Tests (B200)
+  key: kimi-k3-unit-tests-b200
+  timeout_in_minutes: 40
+  device: b200-k8s
+  source_file_dependencies:
+  - vllm/models/kimi_k3/
+  - csrc/libtorch_stable/kimi_k3/
+  - tests/models/kimi_k3/
+  commands:
+    # The native NVIDIA AttnRes kernel requires the SM100 family.
+    - pytest -v -s models/kimi_k3
+
 - label: Basic Models Test (Other CPU) # 5min
   key: basic-models-test-other-cpu
   depends_on:
@@ -70,7 +82,8 @@ steps:
   - vllm/
   - tests/models/test_utils.py
   - tests/models/test_vision.py
+  - tests/models/test_adapters.py
   - tests/models/transformers/fusers/
   device: cpu-small
   commands:
-    - pytest -v -s models/test_utils.py models/test_vision.py models/transformers/fusers/
+    - pytest -v -s models/test_utils.py models/test_vision.py models/test_adapters.py models/transformers/fusers/

.buildkite/test_areas/spec_decode.yaml

@@ -90,8 +90,10 @@ steps:
     - vllm/v1/spec_decode/
     - vllm/v1/worker/gpu/spec_decode/
     - tests/v1/e2e/spec_decode/
+    - tests/spec_decode/
   commands:
     - pytest -v -s v1/e2e/spec_decode -k "ngram or suffix"
+    - python3 spec_decode/test_custom_proposer.py
   mirror:
     amd:
       dind: false
@@ -188,3 +190,15 @@ steps:
     - tests/v1/e2e/spec_decode/test_mtp_parallel_load.py
   commands:
     - pytest -v -s v1/e2e/spec_decode/test_mtp_parallel_load.py
+
+- label: Spec Decode Acceptance Rates Nightly
+  key: spec-decode-acceptance-rates-nightly
+  timeout_in_minutes: 60
+  device: h200_35gb
+  optional: true
+  source_file_dependencies:
+    - vllm/v1/spec_decode/
+    - vllm/v1/worker/gpu/spec_decode/
+    - tests/v1/e2e/spec_decode/
+  commands:
+    - pytest -v -s v1/e2e/spec_decode/test_spec_decode.py -k "acceptance_rates"

.buildkite/test_areas/torch_abi.yaml

@@ -0,0 +1,14 @@
+group: Torch ABI
+depends_on:
+  - image-build
+steps:
+- label: Torch Stable ABI Audit
+  key: torch-stable-abi-audit
+  timeout_in_minutes: 5
+  source_file_dependencies:
+  - .buildkite/check-torch-abi.py
+  - csrc/
+  - cmake/
+  - setup.py
+  commands:
+  - python3 /vllm-workspace/.buildkite/check-torch-abi.py

.github/mergify.yml

@@ -19,6 +19,7 @@ pull_request_rules:
   description: Comment on PR when pre-commit check fails
   conditions:
     - check-failure=pre-commit
+    - -check-cancelled=pre-commit
     - -closed
     - -draft
     - or:
@@ -232,6 +233,31 @@ pull_request_rules:
       add:
         - gpt-oss
 
+- name: label-kimi
+  description: Automatically apply kimi label
+  conditions:
+    - label != stale
+    - or:
+      - files~=(?i)kimi
+      - files~=(?i)moonshot
+      - title~=(?i)(?:kimi|moonshot)
+  actions:
+    label:
+      add:
+        - kimi
+
+- name: label-k3
+  description: Automatically apply k3 label (launch triage; retire after ramp-down)
+  conditions:
+    - label != stale
+    - or:
+      - files~=(?i)kimi[-_]?k3
+      - title~=(?i)(?:kimi[-\s]?k3|\bk3\b)
+  actions:
+    label:
+      add:
+        - k3
+
 - name: label-nvidia
   description: Automatically apply nvidia label
   conditions:

.github/workflows/issue_autolabel.yml

@@ -130,6 +130,25 @@ jobs:
                   },
                 ],
               },
+              kimi: {
+                keywords: [
+                  { term: "Kimi",     searchIn: "both" },
+                  { term: "Moonshot", searchIn: "both" },
+                ],
+                substrings: [
+                  { term: "moonshotai/", searchIn: "both" },
+                  { term: "kimi",        searchIn: "title" },
+                ],
+              },
+              k3: {
+                keywords: [
+                  { term: "Kimi K3", searchIn: "both" },
+                  { term: "K3",      searchIn: "title" },
+                ],
+                substrings: [
+                  { term: "moonshotai/kimi-k3", searchIn: "both" },
+                ],
+              },
               quantization: {
                 keywords: [
                   {

.github/workflows/macos-smoke-test.yml

@@ -29,7 +29,7 @@ jobs:
     steps:
       - uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0 # v7.0.0
 
-      - uses: astral-sh/setup-uv@37802adc94f370d6bfd71619e3f0bf239e1f3b78 # v7.6.0
+      - uses: astral-sh/setup-uv@c771a70e6277c0a99b617c7a806ffedaca235ff9 # v9.0.0
         with:
           enable-cache: true
           cache-dependency-glob: |

.github/workflows/new_pr_bot.yml

@@ -80,9 +80,9 @@ jobs:
                   '',
                   '\u{1f4ac} Join our developer Slack at https://slack.vllm.ai to discuss your PR in `#pr-reviews`, coordinate on features in `#feat-` channels, or join special interest groups in `#sig-` channels.',
                   '',
-                  'PRs do not trigger a full CI run by default. Once the PR is approved and ready to go, your PR reviewer(s) can run CI to test the changes comprehensively before merging.',
+                  'PRs do not trigger a full CI run by default. Reviewers with write access and configured trusted contributors can comment `/ci run` whenever CI signals are needed.',
                   '',
-                  'To run CI, PR reviewers can either: Add `ready` label to the PR or enable auto-merge.',
+                  'Once the PR is approved or has the `ready` label, the PR author can also use `/ci run` or `/ci retry`. New commits do not start CI automatically.',
                   '',
                   'If you have any questions, please reach out to us on Slack at https://slack.vllm.ai.',
                   '',

.github/workflows/pre-commit.yml

@@ -41,7 +41,7 @@ jobs:
           if (hasReadyLabel || hasVerifiedLabel || mergedCount >= 4) {
             core.info(`Check passed: verified label=${hasVerifiedLabel}, ready label=${hasReadyLabel}, 4+ merged PRs=${mergedCount >= 4}`);
           } else {
-            core.setFailed(`PR must have the 'verified', 'ready', or 'ready-run-all-tests' label (the ready labels also trigger tests) or the author must have at least 4 merged PRs (found ${mergedCount}).`);
+            core.setFailed(`PR must have the 'verified', 'ready', or 'ready-run-all-tests' label to run pre-commit, or the author must have at least 4 merged PRs (found ${mergedCount}).`);
           }
 
   pre-commit:

.github/workflows/run-ci-command.yml

@@ -0,0 +1,40 @@
+name: Run CI from PR comment
+
+on:
+  issue_comment:
+    types: [created]
+
+concurrency:
+  group: run-ci-comment-${{ github.event.issue.number }}
+  cancel-in-progress: false
+
+permissions:
+  contents: read
+  issues: write
+  pull-requests: read
+
+jobs:
+  run-ci-command:
+    if: >-
+      github.event.issue.pull_request &&
+      (github.event.comment.body == '/ci run' ||
+      github.event.comment.body == '/ci retry')
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0 # v7.0.0
+        with:
+          ref: ${{ github.event.repository.default_branch }}
+          persist-credentials: false
+      - uses: astral-sh/setup-uv@c771a70e6277c0a99b617c7a806ffedaca235ff9 # v9.0.0
+        with:
+          python-version: "3.12"
+      - name: Authorize and run CI command
+        run: >-
+          uv run --no-project --python 3.12
+          .github/workflows/scripts/run_ci_command.py
+        env:
+          BUILDKITE_API_TOKEN: ${{ secrets.BUILDKITE_API_TOKEN }}
+          BUILDKITE_ORGANIZATION: vllm
+          BUILDKITE_PIPELINE: ci
+          CI_TRUSTED_USERS: ${{ vars.CI_TRUSTED_USERS }}
+          GH_TOKEN: ${{ github.token }}

.github/workflows/scripts/run_ci_command.py

@@ -0,0 +1,580 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+import json
+import os
+import sys
+import urllib.error
+import urllib.parse
+import urllib.request
+from collections.abc import Mapping, Sequence
+from typing import Any
+
+COMMAND_RUN_CI = "/ci run"
+COMMAND_RETRY_FAILED = "/ci retry"
+READY_LABELS = {"ready", "ready-run-all-tests"}
+TRUSTED_PERMISSIONS = {"admin", "maintain", "write"}
+ACTIVE_BUILD_STATES = {
+    "blocked",
+    "creating",
+    "scheduled",
+    "running",
+    "failing",
+    "canceling",
+    "waiting",
+    "waiting_failed",
+}
+RETRY_STATES = "failed,timed_out,expired"
+
+
+class ApiError(RuntimeError):
+    def __init__(self, status: int | None, message: str) -> None:
+        super().__init__(message)
+        self.status = status
+
+
+class HttpTransport:
+    def request(
+        self,
+        url: str,
+        *,
+        body: Mapping[str, Any] | None = None,
+        headers: Mapping[str, str] | None = None,
+        method: str = "GET",
+    ) -> Any:
+        data = None if body is None else json.dumps(body).encode()
+        request = urllib.request.Request(
+            url,
+            data=data,
+            headers=dict(headers or {}),
+            method=method,
+        )
+        try:
+            with urllib.request.urlopen(request, timeout=30) as response:
+                response_body = response.read().decode()
+        except urllib.error.HTTPError as error:
+            response_body = error.read().decode()
+            message = self._error_message(response_body, error.reason)
+            raise ApiError(
+                error.code,
+                f"API returned {error.code}: {message}",
+            ) from error
+        except urllib.error.URLError as error:
+            raise ApiError(None, f"API request failed: {error.reason}") from error
+
+        if not response_body:
+            return None
+        try:
+            return json.loads(response_body)
+        except json.JSONDecodeError as error:
+            raise ApiError(None, "API returned a non-JSON response.") from error
+
+    @staticmethod
+    def _error_message(response_body: str, fallback: str) -> str:
+        try:
+            parsed = json.loads(response_body)
+        except json.JSONDecodeError:
+            return fallback
+        return str(parsed.get("message", fallback))
+
+
+class GitHubClient:
+    def __init__(
+        self,
+        token: str,
+        repository: str,
+        transport: HttpTransport | None = None,
+    ) -> None:
+        if not token:
+            raise RuntimeError("GH_TOKEN is not set.")
+        self.owner, self.repo = repository.split("/", maxsplit=1)
+        self.transport = transport or HttpTransport()
+        self.headers = {
+            "Accept": "application/vnd.github+json",
+            "Authorization": f"Bearer {token}",
+            "Content-Type": "application/json",
+            "User-Agent": "vllm-ci-command",
+            "X-GitHub-Api-Version": "2022-11-28",
+        }
+
+    def _request(
+        self,
+        path: str,
+        *,
+        body: Mapping[str, Any] | None = None,
+        method: str = "GET",
+    ) -> Any:
+        return self.transport.request(
+            f"https://api.github.com{path}",
+            body=body,
+            headers=self.headers,
+            method=method,
+        )
+
+    def _repo_path(self, suffix: str) -> str:
+        owner = urllib.parse.quote(self.owner, safe="")
+        repo = urllib.parse.quote(self.repo, safe="")
+        return f"/repos/{owner}/{repo}{suffix}"
+
+    def _paginate(self, path: str) -> list[dict[str, Any]]:
+        results: list[dict[str, Any]] = []
+        separator = "&" if "?" in path else "?"
+        for page in range(1, 101):
+            response = self._request(f"{path}{separator}per_page=100&page={page}")
+            if not isinstance(response, list):
+                raise ApiError(None, "GitHub API returned an invalid list response.")
+            results.extend(response)
+            if len(response) < 100:
+                return results
+        raise ApiError(None, "GitHub API pagination exceeded 10,000 results.")
+
+    def get_pr(self, number: int) -> dict[str, Any]:
+        return self._request(self._repo_path(f"/pulls/{number}"))
+
+    def get_permission(self, actor: str) -> str:
+        username = urllib.parse.quote(actor, safe="")
+        try:
+            response = self._request(
+                self._repo_path(f"/collaborators/{username}/permission")
+            )
+        except ApiError as error:
+            if error.status == 404:
+                return "none"
+            raise
+        return str(response["permission"])
+
+    def get_review_decision(self, number: int) -> str | None:
+        query = """
+          query($owner: String!, $repo: String!, $number: Int!) {
+            repository(owner: $owner, name: $repo) {
+              pullRequest(number: $number) {
+                reviewDecision
+              }
+            }
+          }
+        """
+        response = self._request(
+            "/graphql",
+            body={
+                "query": query,
+                "variables": {
+                    "number": number,
+                    "owner": self.owner,
+                    "repo": self.repo,
+                },
+            },
+            method="POST",
+        )
+        return response["data"]["repository"]["pullRequest"]["reviewDecision"]
+
+    def list_reviews(self, number: int) -> list[dict[str, Any]]:
+        return self._paginate(self._repo_path(f"/pulls/{number}/reviews"))
+
+    def list_reactions(self, comment_id: int) -> list[dict[str, Any]]:
+        return self._paginate(
+            self._repo_path(f"/issues/comments/{comment_id}/reactions")
+        )
+
+    def add_reaction(self, comment_id: int, content: str) -> None:
+        self._request(
+            self._repo_path(f"/issues/comments/{comment_id}/reactions"),
+            body={"content": content},
+            method="POST",
+        )
+
+    def add_comment(self, issue_number: int, body: str) -> None:
+        self._request(
+            self._repo_path(f"/issues/{issue_number}/comments"),
+            body={"body": body},
+            method="POST",
+        )
+
+
+class BuildkiteClient:
+    def __init__(
+        self,
+        token: str,
+        organization: str,
+        pipeline: str,
+        transport: HttpTransport | None = None,
+    ) -> None:
+        self.token = token
+        self.transport = transport or HttpTransport()
+        organization = urllib.parse.quote(organization, safe="")
+        pipeline = urllib.parse.quote(pipeline, safe="")
+        self.base_url = (
+            "https://api.buildkite.com/v2/organizations/"
+            f"{organization}/pipelines/{pipeline}/builds"
+        )
+
+    def _request(
+        self,
+        *,
+        body: Mapping[str, Any] | None = None,
+        method: str = "GET",
+        path: str = "",
+        query: Sequence[tuple[str, str]] = (),
+    ) -> Any:
+        if not self.token:
+            raise RuntimeError("The BUILDKITE_API_TOKEN repository secret is not set.")
+        url = f"{self.base_url}{path}"
+        if query:
+            url = f"{url}?{urllib.parse.urlencode(query)}"
+        return self.transport.request(
+            url,
+            body=body,
+            headers={
+                "Authorization": f"Bearer {self.token}",
+                "Content-Type": "application/json",
+                "User-Agent": "vllm-ci-command",
+            },
+            method=method,
+        )
+
+    def list_builds(
+        self,
+        commit: str,
+        *,
+        metadata: tuple[str, str] | None = None,
+    ) -> list[dict[str, Any]]:
+        query = [
+            ("commit", commit),
+            ("exclude_jobs", "true"),
+            ("exclude_pipeline", "true"),
+            ("per_page", "100"),
+        ]
+        if metadata:
+            key, value = metadata
+            query.append((f"meta_data[{key}]", value))
+        response = self._request(query=query)
+        if not isinstance(response, list):
+            raise ApiError(None, "Buildkite API returned an invalid build list.")
+        return response
+
+    def create_build(self, body: Mapping[str, Any]) -> dict[str, Any]:
+        return self._request(body=body, method="POST")
+
+    def retry_failed_jobs(
+        self,
+        build_number: int,
+        states: str,
+    ) -> dict[str, Any]:
+        number = urllib.parse.quote(str(build_number), safe="")
+        return self._request(
+            body={"states": states},
+            method="PUT",
+            path=f"/{number}/retry_failed_jobs",
+        )
+
+
+def parse_command(body: str) -> str | None:
+    if body in {COMMAND_RUN_CI, COMMAND_RETRY_FAILED}:
+        return body
+    return None
+
+
+def parse_trusted_users(value: str = "") -> set[str]:
+    return {
+        user.casefold() for item in value.split(",") for user in item.split() if user
+    }
+
+
+def has_ready_label(pr: Mapping[str, Any]) -> bool:
+    return any(label["name"] in READY_LABELS for label in pr["labels"])
+
+
+def is_trusted_permission(permission: str) -> bool:
+    return permission in TRUSTED_PERMISSIONS
+
+
+def authorize(
+    *,
+    actor: str,
+    permission: str,
+    pr: Mapping[str, Any],
+    trusted_approval: bool = False,
+    trusted_users: set[str] | None = None,
+) -> tuple[bool, str]:
+    trusted_users = trusted_users or set()
+    if is_trusted_permission(permission):
+        return True, f"repository {permission} permission"
+    if actor.casefold() in trusted_users:
+        return True, "configured trusted contributor"
+    if actor.casefold() != pr["user"]["login"].casefold():
+        return (
+            False,
+            "Only reviewers with write access can run CI before it is "
+            "delegated to the PR author.",
+        )
+    if pr["draft"]:
+        return False, "PR authors cannot run CI while the PR is a draft."
+    if has_ready_label(pr):
+        return True, "ready label"
+    if trusted_approval:
+        return True, "approval from a trusted reviewer"
+    return (
+        False,
+        "A reviewer with write access must run `/ci run`, approve the PR, "
+        "or add the `ready` label first.",
+    )
+
+
+def has_trusted_approval(
+    github: GitHubClient,
+    number: int,
+    trusted_users: set[str],
+) -> bool:
+    if github.get_review_decision(number) != "APPROVED":
+        return False
+
+    latest_review_states: dict[str, tuple[str, str]] = {}
+    for review in github.list_reviews(number):
+        user = review.get("user") or {}
+        login = user.get("login")
+        state = review.get("state")
+        if login and state in {"APPROVED", "CHANGES_REQUESTED", "DISMISSED"}:
+            latest_review_states[login.casefold()] = (login, state)
+
+    for login, state in latest_review_states.values():
+        if state != "APPROVED":
+            continue
+        if login.casefold() in trusted_users:
+            return True
+        if is_trusted_permission(github.get_permission(login)):
+            return True
+    return False
+
+
+def is_build_for_pr(build: Mapping[str, Any], pr_number: int) -> bool:
+    pull_request = build.get("pull_request")
+    if isinstance(pull_request, Mapping):
+        build_pr_number = pull_request.get("id", pull_request.get("number"))
+        if build_pr_number is not None:
+            return str(build_pr_number) == str(pr_number)
+    metadata = build.get("meta_data") or {}
+    return str(metadata.get("github-pr-number")) == str(pr_number)
+
+
+def is_active_build(build: Mapping[str, Any]) -> bool:
+    return bool(build.get("blocked")) or build.get("state") in ACTIVE_BUILD_STATES
+
+
+def select_latest_build(
+    builds: Sequence[dict[str, Any]],
+    pr_number: int,
+) -> dict[str, Any] | None:
+    matching = [build for build in builds if is_build_for_pr(build, pr_number)]
+    return max(matching, key=lambda build: build.get("created_at", ""), default=None)
+
+
+def create_build_payload(
+    *,
+    actor: str,
+    comment_id: int,
+    pr: Mapping[str, Any],
+) -> dict[str, Any]:
+    return {
+        "commit": pr["head"]["sha"],
+        "branch": pr["head"]["ref"],
+        "message": f"PR #{pr['number']} {COMMAND_RUN_CI} by @{actor}",
+        "pull_request_id": pr["number"],
+        "pull_request_base_branch": pr["base"]["ref"],
+        "pull_request_repository": pr["head"]["repo"]["clone_url"],
+        "pull_request_labels": [label["name"] for label in pr["labels"]],
+        "env": {
+            "VLLM_CI_GITHUB_COMMENT_ID": str(comment_id),
+            "VLLM_CI_TRIGGERED_BY": actor,
+        },
+        "meta_data": {
+            "github-comment-id": str(comment_id),
+            "github-pr-number": str(pr["number"]),
+            "github-triggered-by": actor,
+        },
+    }
+
+
+def add_reaction_safely(
+    github: GitHubClient,
+    comment_id: int,
+    content: str,
+) -> None:
+    try:
+        github.add_reaction(comment_id, content)
+    except Exception as error:
+        print(f"Could not add {content} reaction: {error}", file=sys.stderr)
+
+
+def is_already_handled(github: GitHubClient, comment_id: int) -> bool:
+    return any(
+        reaction.get("content") in {"rocket", "-1"}
+        and (reaction.get("user") or {}).get("login") == "github-actions[bot]"
+        for reaction in github.list_reactions(comment_id)
+    )
+
+
+def handle_run_ci(
+    *,
+    actor: str,
+    buildkite: BuildkiteClient,
+    comment_id: int,
+    github: GitHubClient,
+    pr: Mapping[str, Any],
+) -> str:
+    duplicate_builds = buildkite.list_builds(
+        pr["head"]["sha"],
+        metadata=("github-comment-id", str(comment_id)),
+    )
+    duplicate = select_latest_build(duplicate_builds, pr["number"])
+    if duplicate:
+        return f"CI was already requested by this comment: {duplicate['web_url']}"
+
+    current_builds = buildkite.list_builds(pr["head"]["sha"])
+    active_build = next(
+        (
+            build
+            for build in current_builds
+            if is_build_for_pr(build, pr["number"]) and is_active_build(build)
+        ),
+        None,
+    )
+    if active_build:
+        return f"CI is already running for this commit: {active_build['web_url']}"
+
+    current_pr = github.get_pr(pr["number"])
+    if current_pr["state"] != "open" or current_pr["head"]["sha"] != pr["head"]["sha"]:
+        return (
+            "The PR head changed while processing the command. Comment `/ci run` again."
+        )
+
+    build = buildkite.create_build(
+        create_build_payload(
+            actor=actor,
+            comment_id=comment_id,
+            pr=current_pr,
+        )
+    )
+    return (
+        f"Triggered [Buildkite CI #{build['number']}]({build['web_url']}) "
+        f"for commit `{current_pr['head']['sha'][:12]}`."
+    )
+
+
+def handle_retry_failed(
+    *,
+    buildkite: BuildkiteClient,
+    pr: Mapping[str, Any],
+) -> str:
+    builds = buildkite.list_builds(pr["head"]["sha"])
+    build = select_latest_build(builds, pr["number"])
+    if not build:
+        return "No CI build exists for the current PR commit. Use `/ci run` first."
+    if not build.get("finished_at") or is_active_build(build):
+        return f"CI is still running for this commit: {build['web_url']}"
+
+    retried = buildkite.retry_failed_jobs(build["number"], RETRY_STATES)
+    if retried["retried_jobs_count"] == 0:
+        return (
+            f"No failed, timed-out, or expired jobs need retrying: {build['web_url']}"
+        )
+    return (
+        f"Queued {retried['retried_jobs_count']} failed job(s) for retry in "
+        f"[Buildkite CI #{build['number']}]({build['web_url']})."
+    )
+
+
+def run(
+    event: Mapping[str, Any],
+    github: GitHubClient,
+    buildkite: BuildkiteClient,
+    trusted_users_value: str = "",
+) -> None:
+    command = parse_command(event["comment"]["body"])
+    if not command or "pull_request" not in event["issue"]:
+        return
+
+    issue_number = event["issue"]["number"]
+    comment_id = event["comment"]["id"]
+    actor = event["comment"]["user"]["login"]
+
+    if is_already_handled(github, comment_id):
+        print(f"Comment {comment_id} was already handled.")
+        return
+    add_reaction_safely(github, comment_id, "eyes")
+
+    try:
+        pr = github.get_pr(issue_number)
+        permission = github.get_permission(actor)
+        if pr["state"] != "open":
+            github.add_comment(issue_number, "CI commands require an open PR.")
+            return
+
+        trusted_users = parse_trusted_users(trusted_users_value)
+        should_check_approval = (
+            not is_trusted_permission(permission)
+            and actor.casefold() not in trusted_users
+            and actor.casefold() == pr["user"]["login"].casefold()
+            and not pr["draft"]
+            and not has_ready_label(pr)
+        )
+        trusted_approval = should_check_approval and has_trusted_approval(
+            github,
+            issue_number,
+            trusted_users,
+        )
+        allowed, reason = authorize(
+            actor=actor,
+            permission=permission,
+            pr=pr,
+            trusted_approval=trusted_approval,
+            trusted_users=trusted_users,
+        )
+        if not allowed:
+            add_reaction_safely(github, comment_id, "-1")
+            github.add_comment(issue_number, f"@{actor}, {reason}")
+            return
+
+        print(f"Authorized @{actor}: {reason}")
+        if command == COMMAND_RUN_CI:
+            message = handle_run_ci(
+                actor=actor,
+                buildkite=buildkite,
+                comment_id=comment_id,
+                github=github,
+                pr=pr,
+            )
+        else:
+            message = handle_retry_failed(buildkite=buildkite, pr=pr)
+        add_reaction_safely(github, comment_id, "rocket")
+        github.add_comment(issue_number, message)
+    except Exception:
+        add_reaction_safely(github, comment_id, "confused")
+        raise
+
+
+def main() -> None:
+    event_path = os.environ["GITHUB_EVENT_PATH"]
+    with open(event_path, encoding="utf-8") as event_file:
+        event = json.load(event_file)
+
+    if not parse_command(event["comment"]["body"]):
+        return
+
+    github = GitHubClient(
+        os.environ.get("GH_TOKEN", ""),
+        os.environ["GITHUB_REPOSITORY"],
+    )
+    buildkite = BuildkiteClient(
+        os.environ.get("BUILDKITE_API_TOKEN", ""),
+        os.environ.get("BUILDKITE_ORGANIZATION", "vllm"),
+        os.environ.get("BUILDKITE_PIPELINE", "ci"),
+    )
+    run(
+        event,
+        github,
+        buildkite,
+        os.environ.get("CI_TRUSTED_USERS", ""),
+    )
+
+
+if __name__ == "__main__":
+    main()

.github/workflows/scripts/test_run_ci_command.py

@@ -0,0 +1,362 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+import unittest
+from typing import Any
+
+from run_ci_command import (
+    COMMAND_RETRY_FAILED,
+    COMMAND_RUN_CI,
+    RETRY_STATES,
+    BuildkiteClient,
+    authorize,
+    create_build_payload,
+    has_trusted_approval,
+    is_active_build,
+    is_build_for_pr,
+    parse_command,
+    parse_trusted_users,
+    run,
+    select_latest_build,
+)
+
+
+def make_pr(**overrides: Any) -> dict[str, Any]:
+    pr = {
+        "base": {"ref": "main"},
+        "draft": False,
+        "head": {
+            "ref": "feature",
+            "repo": {"clone_url": "https://github.com/contributor/vllm.git"},
+            "sha": "0123456789abcdef",
+        },
+        "labels": [],
+        "number": 42,
+        "state": "open",
+        "user": {"login": "author"},
+    }
+    pr.update(overrides)
+    return pr
+
+
+def make_event(command: str, actor: str = "reviewer") -> dict[str, Any]:
+    return {
+        "comment": {
+            "body": command,
+            "id": 99,
+            "user": {"login": actor},
+        },
+        "issue": {
+            "number": 42,
+            "pull_request": {},
+        },
+    }
+
+
+class FakeGitHub:
+    def __init__(
+        self,
+        *,
+        permission: str = "write",
+        permissions: dict[str, str] | None = None,
+        pr: dict[str, Any] | None = None,
+        review_decision: str = "REVIEW_REQUIRED",
+        reviews: list[dict[str, Any]] | None = None,
+    ) -> None:
+        self.comments: list[str] = []
+        self.permission = permission
+        self.permissions = permissions or {}
+        self.pr = pr or make_pr()
+        self.reactions: list[str] = []
+        self.review_decision = review_decision
+        self.reviews = reviews or []
+
+    def get_pr(self, number: int) -> dict[str, Any]:
+        return self.pr
+
+    def get_permission(self, actor: str) -> str:
+        return self.permissions.get(actor, self.permission)
+
+    def get_review_decision(self, number: int) -> str:
+        return self.review_decision
+
+    def list_reviews(self, number: int) -> list[dict[str, Any]]:
+        return self.reviews
+
+    def list_reactions(self, comment_id: int) -> list[dict[str, Any]]:
+        return []
+
+    def add_reaction(self, comment_id: int, content: str) -> None:
+        self.reactions.append(content)
+
+    def add_comment(self, issue_number: int, body: str) -> None:
+        self.comments.append(body)
+
+
+class FakeBuildkite:
+    def __init__(
+        self,
+        build_lists: list[list[dict[str, Any]]] | None = None,
+    ) -> None:
+        self.build_lists = build_lists or []
+        self.created_builds: list[dict[str, Any]] = []
+        self.list_calls: list[tuple[str, tuple[str, str] | None]] = []
+        self.retry_calls: list[tuple[int, str]] = []
+
+    def list_builds(
+        self,
+        commit: str,
+        *,
+        metadata: tuple[str, str] | None = None,
+    ) -> list[dict[str, Any]]:
+        self.list_calls.append((commit, metadata))
+        return self.build_lists.pop(0)
+
+    def create_build(self, body: dict[str, Any]) -> dict[str, Any]:
+        self.created_builds.append(body)
+        return {
+            "number": 123,
+            "web_url": "https://buildkite.example/builds/123",
+        }
+
+    def retry_failed_jobs(
+        self,
+        build_number: int,
+        states: str,
+    ) -> dict[str, Any]:
+        self.retry_calls.append((build_number, states))
+        return {"retried_jobs_count": 3}
+
+
+class FakeTransport:
+    def __init__(self, response: Any) -> None:
+        self.calls: list[dict[str, Any]] = []
+        self.response = response
+
+    def request(self, url: str, **kwargs: Any) -> Any:
+        self.calls.append({"url": url, **kwargs})
+        return self.response
+
+
+class RunCiCommandTest(unittest.TestCase):
+    def test_only_exact_ci_commands_are_accepted(self) -> None:
+        self.assertEqual(parse_command(COMMAND_RUN_CI), COMMAND_RUN_CI)
+        self.assertEqual(
+            parse_command(COMMAND_RETRY_FAILED),
+            COMMAND_RETRY_FAILED,
+        )
+        self.assertIsNone(parse_command("/ci run please"))
+        self.assertIsNone(parse_command(" /ci run"))
+
+    def test_write_access_authorizes_reviewers_and_authors(self) -> None:
+        allowed, _ = authorize(
+            actor="reviewer",
+            permission="write",
+            pr=make_pr(),
+        )
+        self.assertTrue(allowed)
+
+    def test_configured_trusted_contributors_can_run_ci(self) -> None:
+        trusted_users = parse_trusted_users("trusted-one, TRUSTED-TWO")
+        allowed, _ = authorize(
+            actor="trusted-two",
+            permission="read",
+            pr=make_pr(),
+            trusted_users=trusted_users,
+        )
+        self.assertTrue(allowed)
+
+    def test_authors_need_an_approval_or_ready_label(self) -> None:
+        pending, _ = authorize(
+            actor="author",
+            permission="read",
+            pr=make_pr(),
+        )
+        approved, _ = authorize(
+            actor="author",
+            permission="read",
+            pr=make_pr(),
+            trusted_approval=True,
+        )
+        ready, _ = authorize(
+            actor="author",
+            permission="read",
+            pr=make_pr(labels=[{"name": "ready"}]),
+        )
+        self.assertFalse(pending)
+        self.assertTrue(approved)
+        self.assertTrue(ready)
+
+    def test_non_author_contributors_without_write_are_denied(self) -> None:
+        allowed, _ = authorize(
+            actor="contributor",
+            permission="read",
+            pr=make_pr(),
+            trusted_approval=True,
+        )
+        self.assertFalse(allowed)
+
+    def test_authors_cannot_use_ready_state_on_draft_prs(self) -> None:
+        allowed, _ = authorize(
+            actor="author",
+            permission="read",
+            pr=make_pr(draft=True, labels=[{"name": "ready"}]),
+            trusted_approval=True,
+        )
+        self.assertFalse(allowed)
+
+    def test_only_trusted_reviewers_can_delegate_through_approval(self) -> None:
+        approved_review = {
+            "state": "APPROVED",
+            "user": {"login": "reviewer"},
+        }
+        trusted = FakeGitHub(
+            permission="read",
+            permissions={"reviewer": "write"},
+            review_decision="APPROVED",
+            reviews=[approved_review],
+        )
+        untrusted = FakeGitHub(
+            permission="read",
+            review_decision="APPROVED",
+            reviews=[approved_review],
+        )
+        self.assertTrue(has_trusted_approval(trusted, 42, set()))
+        self.assertFalse(has_trusted_approval(untrusted, 42, set()))
+
+    def test_build_matching_is_scoped_to_the_pr(self) -> None:
+        self.assertTrue(is_build_for_pr({"pull_request": {"id": 42}}, 42))
+        self.assertFalse(is_build_for_pr({"pull_request": {"id": 43}}, 42))
+        self.assertTrue(
+            is_build_for_pr(
+                {"meta_data": {"github-pr-number": "42"}},
+                42,
+            )
+        )
+
+    def test_latest_build_selection_ignores_other_prs(self) -> None:
+        latest = select_latest_build(
+            [
+                {
+                    "created_at": "2026-07-28T02:00:00Z",
+                    "number": 3,
+                    "pull_request": {"id": 43},
+                },
+                {
+                    "created_at": "2026-07-28T01:00:00Z",
+                    "number": 2,
+                    "pull_request": {"id": 42},
+                },
+                {
+                    "created_at": "2026-07-28T00:00:00Z",
+                    "number": 1,
+                    "pull_request": {"id": 42},
+                },
+            ],
+            42,
+        )
+        self.assertEqual(latest["number"], 2)
+
+    def test_active_build_states_prevent_duplicate_runs(self) -> None:
+        self.assertTrue(is_active_build({"state": "scheduled"}))
+        self.assertTrue(is_active_build({"state": "running"}))
+        self.assertTrue(is_active_build({"state": "waiting"}))
+        self.assertTrue(is_active_build({"blocked": True, "state": "passed"}))
+        self.assertFalse(is_active_build({"state": "failed"}))
+
+    def test_build_payload_preserves_pr_context(self) -> None:
+        payload = create_build_payload(
+            actor="reviewer",
+            comment_id=99,
+            pr=make_pr(labels=[{"name": "ready"}, {"name": "v1"}]),
+        )
+        self.assertEqual(
+            payload,
+            {
+                "commit": "0123456789abcdef",
+                "branch": "feature",
+                "message": "PR #42 /ci run by @reviewer",
+                "pull_request_id": 42,
+                "pull_request_base_branch": "main",
+                "pull_request_repository": ("https://github.com/contributor/vllm.git"),
+                "pull_request_labels": ["ready", "v1"],
+                "env": {
+                    "VLLM_CI_GITHUB_COMMENT_ID": "99",
+                    "VLLM_CI_TRIGGERED_BY": "reviewer",
+                },
+                "meta_data": {
+                    "github-comment-id": "99",
+                    "github-pr-number": "42",
+                    "github-triggered-by": "reviewer",
+                },
+            },
+        )
+
+    def test_ci_run_dispatches_build_with_current_pr_metadata(self) -> None:
+        github = FakeGitHub()
+        buildkite = FakeBuildkite([[], []])
+        run(make_event(COMMAND_RUN_CI), github, buildkite)
+
+        self.assertEqual(len(buildkite.created_builds), 1)
+        self.assertEqual(
+            buildkite.created_builds[0]["message"],
+            "PR #42 /ci run by @reviewer",
+        )
+        self.assertEqual(github.reactions, ["eyes", "rocket"])
+        self.assertIn("Buildkite CI #123", github.comments[0])
+
+    def test_unapproved_authors_are_denied_without_buildkite(self) -> None:
+        github = FakeGitHub(
+            permission="read",
+            pr=make_pr(),
+            review_decision="REVIEW_REQUIRED",
+        )
+        buildkite = FakeBuildkite()
+        run(make_event(COMMAND_RUN_CI, "author"), github, buildkite)
+
+        self.assertEqual(buildkite.list_calls, [])
+        self.assertEqual(github.reactions, ["eyes", "-1"])
+        self.assertIn("approve the PR", github.comments[0])
+
+    def test_ci_retry_uses_latest_current_sha_build(self) -> None:
+        github = FakeGitHub(
+            permission="read",
+            pr=make_pr(labels=[{"name": "ready"}]),
+        )
+        buildkite = FakeBuildkite(
+            [
+                [
+                    {
+                        "created_at": "2026-07-28T01:00:00Z",
+                        "finished_at": "2026-07-28T02:00:00Z",
+                        "number": 123,
+                        "pull_request": {"id": 42},
+                        "state": "failed",
+                        "web_url": "https://buildkite.example/builds/123",
+                    }
+                ]
+            ]
+        )
+        run(make_event(COMMAND_RETRY_FAILED, "author"), github, buildkite)
+
+        self.assertEqual(buildkite.retry_calls, [(123, RETRY_STATES)])
+        self.assertIn("Queued 3 failed job", github.comments[0])
+
+    def test_buildkite_retry_uses_retry_failed_jobs_endpoint(self) -> None:
+        transport = FakeTransport({"retried_jobs_count": 2})
+        client = BuildkiteClient(
+            "secret",
+            "vllm",
+            "ci",
+            transport=transport,
+        )
+        client.retry_failed_jobs(123, RETRY_STATES)
+
+        call = transport.calls[0]
+        self.assertEqual(call["method"], "PUT")
+        self.assertTrue(call["url"].endswith("/123/retry_failed_jobs"))
+        self.assertEqual(call["body"], {"states": RETRY_STATES})
+
+
+if __name__ == "__main__":
+    unittest.main()

.gitignore

@@ -173,9 +173,6 @@ venv.bak/
 
 # mkdocs documentation
 /site
-docs/argparse
-docs/examples/*
-!docs/examples/README.md
 
 # mypy
 .mypy_cache/

.markdownlint.yaml

@@ -3,6 +3,9 @@ MD007:
 MD013: false
 MD024:
   siblings_only: true
+MD025:
+  # Allow front matter title to be different from the first heading in the document.
+  front_matter_title: ""
 MD031:
   list_items: false
 MD033: false

.pre-commit-config.yaml

@@ -260,10 +260,6 @@ repos:
     files: ^docker/(Dockerfile|versions\.json)$
     pass_filenames: false
     additional_dependencies: [dockerfile-parse]
-  - id: attention-backend-docs
-    name: Check attention backend documentation is up to date
-    entry: python tools/pre_commit/generate_attention_backend_docs.py --check
-    language: python
   - id: check-boolean-context-manager
     name: Check for boolean ops in with-statements
     entry: python tools/pre_commit/check_boolean_context_manager.py

CMakeLists.txt

@@ -219,10 +219,8 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
   # the set of architectures we want to compile for and remove the from the
   # CMAKE_CUDA_FLAGS so that they are not applied globally.
   #
-  # `+PTX` in TORCH_CUDA_ARCH_LIST is not preserved here. It is emitted by torch
-  # as `code=compute_*`, while extract_unique_cuda_archs_ascending() records only
-  # `arch=compute_*`. If a kernel really needs PTX, add `+PTX` to that kernel's
-  # component-specific arch list below.
+  # `+PTX` in TORCH_CUDA_ARCH_LIST is not preserved here. If a kernel really
+  # needs PTX, add `+PTX` to that kernel's component-specific arch list below.
   #
   clear_cuda_arches(CUDA_ARCH_FLAGS)
   extract_unique_cuda_archs_ascending(CUDA_ARCHS "${CUDA_ARCH_FLAGS}")
@@ -232,6 +230,13 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
   cuda_archs_loose_intersection(CUDA_ARCHS
     "${CUDA_SUPPORTED_ARCHS}" "${CUDA_ARCHS}")
   message(STATUS "CUDA supported target architectures: ${CUDA_ARCHS}")
+  if(NOT CUDA_ARCHS)
+    message(FATAL_ERROR
+      "No supported CUDA architectures; the build would produce a binary "
+      "with no usable kernels. Detected gencode flags: ${CUDA_ARCH_FLAGS}; "
+      "supported: ${CUDA_SUPPORTED_ARCHS}. "
+      "Set TORCH_CUDA_ARCH_LIST for your GPU (e.g. 12.0).")
+  endif()
 else()
   #
   # For other GPU targets override the GPU architectures detected by cmake/torch
@@ -1074,6 +1079,24 @@ if(VLLM_GPU_LANG STREQUAL "CUDA" OR VLLM_GPU_LANG STREQUAL "HIP")
     set(MLA_ARCHS)
   endif()
 
+  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
+    cuda_archs_loose_intersection(KIMI_K3_ATTN_RES_ARCHS
+      "10.0f" "${CUDA_ARCHS}")
+  endif()
+  if(KIMI_K3_ATTN_RES_ARCHS)
+    set(KIMI_K3_ATTN_RES_SRC
+      "csrc/libtorch_stable/kimi_k3/attn_res_kernel.cu")
+    set_gencode_flags_for_srcs(
+      SRCS "${KIMI_K3_ATTN_RES_SRC}"
+      CUDA_ARCHS "${KIMI_K3_ATTN_RES_ARCHS}")
+    set_property(SOURCE ${KIMI_K3_ATTN_RES_SRC} APPEND PROPERTY
+      COMPILE_OPTIONS
+      "$<$<COMPILE_LANGUAGE:CUDA>:--expt-relaxed-constexpr;--expt-extended-lambda;--use_fast_math>")
+    list(APPEND VLLM_STABLE_EXT_SRC "${KIMI_K3_ATTN_RES_SRC}")
+    message(STATUS
+      "Building Kimi K3 AttnRes for archs: ${KIMI_K3_ATTN_RES_ARCHS}")
+  endif()
+
   # Hadacore kernels
   cuda_archs_loose_intersection(HADACORE_ARCHS "8.0+PTX;9.0+PTX" "${CUDA_ARCHS}")
   if(HADACORE_ARCHS)
@@ -1115,6 +1138,10 @@ if(VLLM_GPU_LANG STREQUAL "CUDA" OR VLLM_GPU_LANG STREQUAL "HIP")
       target_compile_definitions(_C_stable_libtorch PRIVATE
         VLLM_ENABLE_COOPERATIVE_TOPK=1)
     endif()
+    if(KIMI_K3_ATTN_RES_ARCHS)
+      target_compile_definitions(_C_stable_libtorch PRIVATE
+        VLLM_ENABLE_KIMI_K3_ATTN_RES=1)
+    endif()
     # Needed by CUTLASS kernels
     target_compile_definitions(_C_stable_libtorch PRIVATE
       CUTLASS_ENABLE_DIRECT_CUDA_DRIVER_CALL=1)

benchmarks/attention_benchmarks/benchmark.py

@@ -1358,6 +1358,10 @@ def main():
                             profile_memory=args.profile_memory,
                             warmup_ms=args.warmup_ms,
                             prefill_backend=pb,
+                            kv_lora_rank=args.kv_lora_rank,
+                            qk_nope_head_dim=args.qk_nope_head_dim,
+                            qk_rope_head_dim=args.qk_rope_head_dim,
+                            v_head_dim=args.v_head_dim,
                         )
 
                         result = run_benchmark(config)

benchmarks/kernels/benchmark_inkling_qkvr_prep.py

@@ -0,0 +1,176 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+import statistics
+
+import torch
+from tabulate import tabulate
+
+from vllm.models.inkling.nvidia.ops import qkvr_prep
+from vllm.utils.argparse_utils import FlexibleArgumentParser
+
+
+def make_inputs(tokens: int, tp_size: int, is_local: bool):
+    torch.manual_seed(0)
+    num_q_heads = 64 // tp_size
+    num_kv_heads = (16 if is_local else 8) // tp_size
+    head_dim = 128
+    d_rel = 16
+    rel_extent = 512 if is_local else 1024
+    page_size = 16
+    num_blocks = (tokens + page_size - 1) // page_size
+    q_width = num_q_heads * head_dim
+    kv_width = num_kv_heads * head_dim
+    r_width = num_q_heads * d_rel
+    device = "cuda"
+
+    qkvr = torch.randn(
+        tokens,
+        q_width + 2 * kv_width + r_width,
+        device=device,
+        dtype=torch.bfloat16,
+    )
+    k_weight = torch.randn(kv_width, 4, device=device, dtype=torch.bfloat16)
+    v_weight = torch.randn_like(k_weight)
+    q_norm_weight = torch.randn(head_dim, device=device, dtype=torch.bfloat16)
+    k_norm_weight = torch.randn_like(q_norm_weight)
+    rel_proj = torch.randn(d_rel, rel_extent, device=device, dtype=torch.bfloat16)
+    conv_cache = torch.zeros(
+        num_blocks,
+        num_kv_heads,
+        page_size,
+        2 * head_dim,
+        device=device,
+        dtype=torch.bfloat16,
+    )
+    key_cache = torch.empty(
+        num_blocks,
+        page_size,
+        num_kv_heads,
+        head_dim,
+        device=device,
+        dtype=torch.bfloat16,
+    )
+    value_cache = torch.empty_like(key_cache)
+    positions = torch.arange(tokens, device=device, dtype=torch.int64)
+    block_table = torch.arange(num_blocks, device=device, dtype=torch.int32)[None]
+    seq_idx = torch.zeros(tokens, device=device, dtype=torch.int32)
+    slots = torch.arange(tokens, device=device, dtype=torch.int64)
+    query_start = torch.zeros(tokens, device=device, dtype=torch.int32)
+    log_scaling = None
+    if not is_local:
+        effective_n = (positions + 1).to(torch.float32)
+        log_scaling = 1.0 + 0.1 * torch.log(torch.clamp(effective_n / 128000, min=1.0))
+    return (
+        qkvr,
+        k_weight,
+        v_weight,
+        q_norm_weight,
+        k_norm_weight,
+        rel_proj,
+        1e-6,
+        num_q_heads,
+        num_kv_heads,
+        head_dim,
+        d_rel,
+        conv_cache,
+        key_cache,
+        value_cache,
+        positions,
+        block_table,
+        seq_idx,
+        slots,
+        query_start,
+        slots,
+        0,
+        head_dim,
+        page_size,
+        log_scaling,
+    )
+
+
+def capture(implementation, inputs):
+    outputs = []
+
+    def run():
+        outputs[:] = implementation.fused_qkvr_prep(*inputs)
+
+    stream = torch.cuda.Stream()
+    stream.wait_stream(torch.cuda.current_stream())
+    with torch.cuda.stream(stream):
+        for _ in range(3):
+            run()
+    torch.cuda.current_stream().wait_stream(stream)
+    torch.accelerator.synchronize()
+    graph = torch.cuda.CUDAGraph()
+    with torch.cuda.graph(graph):
+        run()
+    torch.accelerator.synchronize()
+    return graph, outputs
+
+
+def time_graph(graph: torch.cuda.CUDAGraph, warmup: int, repeats: int) -> float:
+    for _ in range(warmup):
+        graph.replay()
+    torch.accelerator.synchronize()
+    start = torch.cuda.Event(enable_timing=True)
+    end = torch.cuda.Event(enable_timing=True)
+    start.record()
+    for _ in range(repeats):
+        graph.replay()
+    end.record()
+    end.synchronize()
+    return start.elapsed_time(end) * 1000 / repeats
+
+
+def benchmark(inputs, args) -> float:
+    graph, _ = capture(qkvr_prep, inputs)
+    return statistics.median(
+        time_graph(graph, args.warmup, args.repeats) for _ in range(args.trials)
+    )
+
+
+@torch.inference_mode()
+def main(args):
+    rows = []
+    for tp_size in args.tp_sizes:
+        for tokens in args.tokens:
+            for is_local in (True, False):
+                triton_us = benchmark(make_inputs(tokens, tp_size, is_local), args)
+                rows.append(
+                    [
+                        tp_size,
+                        tokens,
+                        "local" if is_local else "global",
+                        triton_us,
+                    ]
+                )
+
+    print("Inkling QKVR prep (CUDA graph, median latency)")
+    print(
+        tabulate(
+            rows,
+            headers=[
+                "TP",
+                "tokens",
+                "scope",
+                "Triton (us)",
+            ],
+            floatfmt=("d", "d", "", ".2f"),
+        )
+    )
+
+
+if __name__ == "__main__":
+    parser = FlexibleArgumentParser()
+    parser.add_argument(
+        "--tokens",
+        type=int,
+        nargs="+",
+        default=[1 << power for power in range(15)],
+    )
+    parser.add_argument("--tp-sizes", type=int, nargs="+", default=[4, 8])
+    parser.add_argument("--warmup", type=int, default=20)
+    parser.add_argument("--repeats", type=int, default=200)
+    parser.add_argument("--trials", type=int, default=5)
+    main(parser.parse_args())

benchmarks/kernels/cpu/benchmark_cpu_attn.py

@@ -154,7 +154,7 @@ def main(
                 scale=scale,
                 causal=True,
                 alibi_slopes=None,
-                sliding_window=window_size,
+                sliding_window=window_size if sliding_window is not None else -1,
                 block_table=block_tables,
                 softcap=0,
                 scheduler_metadata=metadata,

cmake/cpu_extension.cmake

@@ -15,6 +15,7 @@ endif()
 #
 set(ENABLE_X86_ISA $ENV{VLLM_CPU_X86})
 set(ENABLE_ARM_BF16 $ENV{VLLM_CPU_ARM_BF16})
+set(ENABLE_ARM_I8MM $ENV{VLLM_CPU_ARM_I8MM})
 set(ENABLE_RVV_BF16 $ENV{VLLM_CPU_RVV_BF16})
 
 include_directories("${CMAKE_SOURCE_DIR}/csrc")
@@ -96,12 +97,14 @@ if (MACOSX_FOUND AND CMAKE_SYSTEM_PROCESSOR STREQUAL "arm64")
     set(ENABLE_NUMA OFF)
     check_sysctl(hw.optional.neon ASIMD_FOUND)
     check_sysctl(hw.optional.arm.FEAT_BF16 ARM_BF16_FOUND)
+    check_sysctl(hw.optional.arm.FEAT_I8MM ARM_I8MM_FOUND)
 else()
     find_isa(${CPUINFO} "Power11" POWER11_FOUND)
     find_isa(${CPUINFO} "POWER10" POWER10_FOUND)
     find_isa(${CPUINFO} "POWER9" POWER9_FOUND)
     find_isa(${CPUINFO} "asimd" ASIMD_FOUND) # Check for ARM NEON support
     find_isa(${CPUINFO} "bf16" ARM_BF16_FOUND) # Check for ARM BF16 support
+    find_isa(${CPUINFO} "i8mm" ARM_I8MM_FOUND) # Check for ARM I8MM support
     find_isa(${CPUINFO} "S390" S390_FOUND)
     find_isa(${CPUINFO} "zvfhmin" RVV_FP16_FOUND) # Check for RISC-V Vector FP16 support
     find_isa(${CPUINFO} "zvfbfmin" RVV_BF16_FOUND) # Check for RISC-V Vector BF16 support
@@ -111,6 +114,11 @@ else()
         set(ARM_BF16_FOUND ON)
         message(STATUS "ARM BF16 support enabled via VLLM_CPU_ARM_BF16 environment variable")
     endif()
+    if (ENABLE_ARM_I8MM)
+        set(ARM_I8MM_FOUND ON)
+        message(STATUS
+            "ARM I8MM support enabled via VLLM_CPU_ARM_I8MM environment variable")
+    endif()
     # Some kernels (e.g. Bianbu on Spacemit X100) do not report zvfbfmin
     # in /proc/cpuinfo despite hardware support. VLLM_CPU_RVV_BF16=1
     # overrides the detection result.
@@ -166,6 +174,11 @@ elseif (ASIMD_FOUND)
         message(WARNING "BF16 functionality is not available")
         set(MARCH_FLAGS "-march=armv8.2-a+dotprod+fp16")  
     endif()
+    if(ARM_I8MM_FOUND)
+        message(STATUS "I8MM extension detected")
+        string(APPEND MARCH_FLAGS "+i8mm")
+        add_compile_definitions(ARM_I8MM_SUPPORT)
+    endif()
     list(APPEND CXX_COMPILE_FLAGS ${MARCH_FLAGS})     
 elseif (S390_FOUND)
     message(STATUS "S390 detected")
@@ -447,8 +460,13 @@ if (ASIMD_FOUND AND NOT APPLE_SILICON_FOUND)
         "csrc/cpu/shm.cpp"
         "csrc/cpu/activation_lut_bf16.cpp"
         "csrc/cpu/cpu_tanhf_neon.hpp"
-        "csrc/cpu/cpu_fused_moe.cpp"
         ${VLLM_EXT_SRC})
+    if (ARM_BF16_FOUND)
+        set(VLLM_EXT_SRC "csrc/cpu/cpu_fused_moe.cpp" ${VLLM_EXT_SRC})
+        if (ARM_I8MM_FOUND)
+            set(VLLM_EXT_SRC "csrc/cpu/cpu_fused_moe_int8.cpp" ${VLLM_EXT_SRC})
+        endif()
+    endif()
 endif()
 
 if (POWER9_FOUND OR POWER10_FOUND OR POWER11_FOUND)	

cmake/utils.cmake

@@ -241,14 +241,15 @@ endmacro()
 # `<major>.<minor>`, dedupes them and then sorts them in ascending order and 
 # stores them in `OUT_ARCHES`.
 #
-# Example:
-#   CUDA_ARCH_FLAGS="-gencode arch=compute_75,code=sm_75;...;-gencode arch=compute_90a,code=sm_90a" 
-#   extract_unique_cuda_archs_ascending(OUT_ARCHES CUDA_ARCH_FLAGS)
-#   OUT_ARCHES="7.5;...;9.0"
+# Prefer `code=sm_*`; fall back to `arch=compute_*` for PTX-only flags.
+# This handles mismatches such as `arch=compute_20,code=sm_121`.
 function(extract_unique_cuda_archs_ascending OUT_ARCHES CUDA_ARCH_FLAGS)
   set(_CUDA_ARCHES)
   foreach(_ARCH ${CUDA_ARCH_FLAGS})
-    string(REGEX MATCH "arch=compute_\([0-9]+[af]?\)" _COMPUTE ${_ARCH})
+    string(REGEX MATCH "code=sm_\([0-9]+[af]?\)" _COMPUTE ${_ARCH})
+    if (NOT _COMPUTE)
+      string(REGEX MATCH "arch=compute_\([0-9]+[af]?\)" _COMPUTE ${_ARCH})
+    endif()
     if (_COMPUTE)
       set(_COMPUTE ${CMAKE_MATCH_1})
     endif()

csrc/cpu/cpu_arch_macros.h

@@ -172,4 +172,15 @@
 
 #endif  // __riscv_v
 
+// Power VSX
+#ifdef __powerpc__
+  // FP32Vec16::exp() in cpu_types_vsx.hpp delegates to FP32Vec8::exp(), which
+  // implements a vectorised 5-term minimax polynomial using VSX intrinsics.
+  #define DEFINE_FAST_EXP                                                     \
+    auto fast_exp = [&](const vec_op::FP32Vec16& vec)                         \
+                        __attribute__((always_inline)) { return vec.exp(); }; \
+    auto fast_exp_f16 = fast_exp;
+
+#endif  // __powerpc__
+
 #endif

csrc/cpu/cpu_fused_moe.cpp

@@ -1,5 +1,6 @@
 #include "cpu/cpu_types.hpp"
 #include "cpu/utils.hpp"
+#include "cpu/cpu_fused_moe_activations.hpp"
 #include "cpu/micro_gemm/cpu_micro_gemm_vec.hpp"
 #include "cpu/cpu_arch_macros.h"
 
@@ -43,193 +44,9 @@
   }()
 
 namespace {
-enum class FusedMOEAct {
-  SiluAndMul,
-  SwigluOAIAndMul,
-  GeluAndMul,
-  GeluTanhAndMul,
-};
 
-FusedMOEAct get_act_type(const std::string& act) {
-  if (act == "silu") {
-    return FusedMOEAct::SiluAndMul;
-  } else if (act == "swigluoai") {
-    return FusedMOEAct::SwigluOAIAndMul;
-  } else if (act == "gelu") {
-    return FusedMOEAct::GeluAndMul;
-  } else if (act == "gelu_tanh") {
-    return FusedMOEAct::GeluTanhAndMul;
-  } else {
-    TORCH_CHECK(false, "Invalid act type: " + act);
-  }
-}
-
-template <typename scalar_t>
-void swigluoai_and_mul(float* __restrict__ input, scalar_t* __restrict__ output,
-                       const int32_t m_size, const int32_t n_size,
-                       const int32_t input_stride,
-                       const int32_t output_stride) {
-  using scalar_vec_t = typename cpu_utils::VecTypeTrait<scalar_t>::vec_t;
-#if !defined(__aarch64__)
-  // For GPT-OSS interleaved gate-up weights
-  alignas(64) static int32_t index[16] = {0,  2,  4,  6,  8,  10, 12, 14,
-                                          16, 18, 20, 22, 24, 26, 28, 30};
-  vec_op::INT32Vec16 index_vec(index);
-#endif
-  vec_op::FP32Vec16 gate_up_max_vec(7.0);
-  vec_op::FP32Vec16 up_min_vec(-7.0);
-  vec_op::FP32Vec16 alpha_vec(1.702);
-  vec_op::FP32Vec16 one_vec(1.0);
-
-  DEFINE_FAST_EXP
-
-  for (int32_t m = 0; m < m_size; ++m) {
-    for (int32_t n = 0; n < n_size; n += 32) {
-      // Note: AdvSIMD does not support gather loads
-#if defined(__aarch64__)
-      vec_op::FP32Vec16 gate_vec(vec_op::uninit);
-      vec_op::FP32Vec16 up_vec(vec_op::uninit);
-      vec_op::FP32Vec16::load_even_odd(input + n, gate_vec, up_vec);
-#else
-      vec_op::FP32Vec16 gate_vec(input + n, index_vec);
-      vec_op::FP32Vec16 up_vec(input + n + 1, index_vec);
-#endif
-      gate_vec = gate_vec.min(gate_up_max_vec);
-      up_vec = up_vec.clamp(up_min_vec, gate_up_max_vec);
-      auto sigmoid_vec = one_vec / (one_vec + fast_exp(-gate_vec * alpha_vec));
-      auto glu = gate_vec * sigmoid_vec;
-      auto gated_output_fp32 = (one_vec + up_vec) * glu;
-      scalar_vec_t gated_output = scalar_vec_t(gated_output_fp32);
-      gated_output.save(output + n / 2);
-    }
-    input += input_stride;
-    output += output_stride;
-  }
-}
-
-template <typename scalar_t>
-void silu_and_mul(float* __restrict__ input, scalar_t* __restrict__ output,
-                  const int32_t m_size, const int32_t n_size,
-                  const int32_t input_stride, const int32_t output_stride) {
-  using scalar_vec_t = typename cpu_utils::VecTypeTrait<scalar_t>::vec_t;
-  const int32_t dim = n_size / 2;
-  float* __restrict__ gate = input;
-  float* __restrict__ up = input + dim;
-  vec_op::FP32Vec16 one_vec(1.0);
-
-  DEFINE_FAST_EXP
-
-  for (int32_t m = 0; m < m_size; ++m) {
-    for (int32_t n = 0; n < dim; n += 16) {
-      vec_op::FP32Vec16 gate_vec(gate + n);
-      vec_op::FP32Vec16 up_vec(up + n);
-      auto sigmoid_vec = one_vec / (one_vec + fast_exp(-gate_vec));
-      auto silu = gate_vec * sigmoid_vec;
-      auto gated_output_fp32 = up_vec * silu;
-      scalar_vec_t gated_output = scalar_vec_t(gated_output_fp32);
-      gated_output.save(output + n);
-    }
-    gate += input_stride;
-    up += input_stride;
-    output += output_stride;
-  }
-}
-
-template <typename scalar_t>
-void gelu_and_mul(float* __restrict__ input, scalar_t* __restrict__ output,
-                  const int32_t m_size, const int32_t n_size,
-                  const int32_t input_stride, const int32_t output_stride) {
-  using scalar_vec_t = typename cpu_utils::VecTypeTrait<scalar_t>::vec_t;
-  const int32_t dim = n_size / 2;
-  float* __restrict__ gate = input;
-  float* __restrict__ up = input + dim;
-  vec_op::FP32Vec16 one_vec(1.0);
-  vec_op::FP32Vec16 w1_vec(M_SQRT1_2);
-  vec_op::FP32Vec16 w2_vec(0.5);
-  alignas(64) float temp[16];
-
-  DEFINE_FAST_EXP
-
-  for (int32_t m = 0; m < m_size; ++m) {
-    for (int32_t n = 0; n < dim; n += 16) {
-      vec_op::FP32Vec16 gate_vec(gate + n);
-      vec_op::FP32Vec16 up_vec(up + n);
-      auto er_input_vec = gate_vec * w1_vec;
-
-      er_input_vec.save(temp);
-      for (int32_t i = 0; i < 16; ++i) {
-        temp[i] = std::erf(temp[i]);
-      }
-      vec_op::FP32Vec16 er_vec(temp);
-      auto gelu = gate_vec * w2_vec * (one_vec + er_vec);
-      auto gated_output_fp32 = up_vec * gelu;
-      scalar_vec_t gated_output = scalar_vec_t(gated_output_fp32);
-      gated_output.save(output + n);
-    }
-    gate += input_stride;
-    up += input_stride;
-    output += output_stride;
-  }
-}
-
-template <typename scalar_t>
-void gelu_tanh_and_mul(float* __restrict__ input, scalar_t* __restrict__ output,
-                       const int32_t m_size, const int32_t n_size,
-                       const int32_t input_stride,
-                       const int32_t output_stride) {
-  using scalar_vec_t = typename cpu_utils::VecTypeTrait<scalar_t>::vec_t;
-  const int32_t dim = n_size / 2;
-  float* __restrict__ gate = input;
-  float* __restrict__ up = input + dim;
-  vec_op::FP32Vec16 one_vec(1.0);
-  vec_op::FP32Vec16 w1_vec(0.7978845608028654);
-  vec_op::FP32Vec16 w2_vec(0.5);
-  vec_op::FP32Vec16 w3_vec(0.044715);
-
-  for (int32_t m = 0; m < m_size; ++m) {
-    for (int32_t n = 0; n < dim; n += 16) {
-      vec_op::FP32Vec16 gate_vec(gate + n);
-      vec_op::FP32Vec16 up_vec(up + n);
-      auto gate_pow3_vec = gate_vec * gate_vec * gate_vec;
-      auto inner_vec = w1_vec * (gate_vec + w3_vec * gate_pow3_vec);
-      // Note: can't use fast_exp form because diffusiongemma will generate
-      // wrong results
-      auto tanh_vec = inner_vec.tanh();
-      auto gelu_tanh = gate_vec * w2_vec * (one_vec + tanh_vec);
-      auto gated_output_fp32 = up_vec * gelu_tanh;
-      scalar_vec_t gated_output = scalar_vec_t(gated_output_fp32);
-      gated_output.save(output + n);
-    }
-    gate += input_stride;
-    up += input_stride;
-    output += output_stride;
-  }
-}
-
-template <typename scalar_t>
-FORCE_INLINE void apply_gated_act(const FusedMOEAct act,
-                                  float* __restrict__ input,
-                                  scalar_t* __restrict__ output,
-                                  const int32_t m, const int32_t n,
-                                  const int32_t input_stride,
-                                  const int32_t output_stride) {
-  switch (act) {
-    case FusedMOEAct::SwigluOAIAndMul:
-      swigluoai_and_mul(input, output, m, n, input_stride, output_stride);
-      return;
-    case FusedMOEAct::SiluAndMul:
-      silu_and_mul(input, output, m, n, input_stride, output_stride);
-      return;
-    case FusedMOEAct::GeluAndMul:
-      gelu_and_mul(input, output, m, n, input_stride, output_stride);
-      return;
-    case FusedMOEAct::GeluTanhAndMul:
-      gelu_tanh_and_mul(input, output, m, n, input_stride, output_stride);
-      return;
-    default:
-      TORCH_CHECK(false, "Unsupported act type.");
-  }
-}
+using cpu_fused_moe_utils::apply_gated_act;
+using cpu_fused_moe_utils::FusedMOEAct;
 
 template <typename scalar_t, typename gemm_t>
 void prepack_moe_weight_impl(scalar_t* __restrict__ weight_ptr,
@@ -817,6 +634,7 @@ void fused_moe_impl(scalar_t* __restrict__ output, scalar_t* __restrict__ input,
     }
   }
 }
+
 }  // namespace
 
 void prepack_moe_weight(
@@ -864,7 +682,7 @@ void cpu_fused_moe(
   const int32_t input_size_2 = w2.size(2);
   const int32_t output_size_2 = w2.size(1);
   const int32_t topk_num = topk_id.size(1);
-  const FusedMOEAct act_type = get_act_type(act);
+  const FusedMOEAct act_type = cpu_fused_moe_utils::get_act_type(act);
   cpu_utils::ISA isa_type = cpu_utils::get_isa(isa);
   TORCH_CHECK(!skip_weighted || topk_num == 1,
               "skip_weighted is only supported for topk=1 on CPU");

csrc/cpu/cpu_fused_moe_activations.hpp

@@ -0,0 +1,204 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+#ifndef CPU_FUSED_MOE_ACTIVATIONS_HPP
+#define CPU_FUSED_MOE_ACTIVATIONS_HPP
+
+#include <cmath>
+#include <cstdint>
+#include <string>
+
+#include "cpu/cpu_arch_macros.h"
+#include "cpu/utils.hpp"
+
+namespace cpu_fused_moe_utils {
+enum class FusedMOEAct {
+  SiluAndMul,
+  SwigluOAIAndMul,
+  GeluAndMul,
+  GeluTanhAndMul,
+};
+
+inline FusedMOEAct get_act_type(const std::string& act) {
+  if (act == "silu") {
+    return FusedMOEAct::SiluAndMul;
+  } else if (act == "swigluoai") {
+    return FusedMOEAct::SwigluOAIAndMul;
+  } else if (act == "gelu") {
+    return FusedMOEAct::GeluAndMul;
+  } else if (act == "gelu_tanh") {
+    return FusedMOEAct::GeluTanhAndMul;
+  } else {
+    TORCH_CHECK(false, "Invalid act type: " + act);
+  }
+}
+
+template <typename scalar_t>
+void swigluoai_and_mul(float* __restrict__ input, scalar_t* __restrict__ output,
+                       const int32_t m_size, const int32_t n_size,
+                       const int32_t input_stride,
+                       const int32_t output_stride) {
+  using scalar_vec_t = typename cpu_utils::VecTypeTrait<scalar_t>::vec_t;
+#if !defined(__aarch64__)
+  // For GPT-OSS interleaved gate-up weights
+  alignas(64) static int32_t index[16] = {0,  2,  4,  6,  8,  10, 12, 14,
+                                          16, 18, 20, 22, 24, 26, 28, 30};
+  vec_op::INT32Vec16 index_vec(index);
+#endif
+  vec_op::FP32Vec16 gate_up_max_vec(7.0);
+  vec_op::FP32Vec16 up_min_vec(-7.0);
+  vec_op::FP32Vec16 alpha_vec(1.702);
+  vec_op::FP32Vec16 one_vec(1.0);
+
+  DEFINE_FAST_EXP
+
+  for (int32_t m = 0; m < m_size; ++m) {
+    for (int32_t n = 0; n < n_size; n += 32) {
+      // Note: AdvSIMD does not support gather loads
+#if defined(__aarch64__)
+      vec_op::FP32Vec16 gate_vec(vec_op::uninit);
+      vec_op::FP32Vec16 up_vec(vec_op::uninit);
+      vec_op::FP32Vec16::load_even_odd(input + n, gate_vec, up_vec);
+#else
+      vec_op::FP32Vec16 gate_vec(input + n, index_vec);
+      vec_op::FP32Vec16 up_vec(input + n + 1, index_vec);
+#endif
+      gate_vec = gate_vec.min(gate_up_max_vec);
+      up_vec = up_vec.clamp(up_min_vec, gate_up_max_vec);
+      auto sigmoid_vec = one_vec / (one_vec + fast_exp(-gate_vec * alpha_vec));
+      auto glu = gate_vec * sigmoid_vec;
+      auto gated_output_fp32 = (one_vec + up_vec) * glu;
+      scalar_vec_t gated_output = scalar_vec_t(gated_output_fp32);
+      gated_output.save(output + n / 2);
+    }
+    input += input_stride;
+    output += output_stride;
+  }
+}
+
+template <typename scalar_t>
+void silu_and_mul(float* __restrict__ input, scalar_t* __restrict__ output,
+                  const int32_t m_size, const int32_t n_size,
+                  const int32_t input_stride, const int32_t output_stride) {
+  using scalar_vec_t = typename cpu_utils::VecTypeTrait<scalar_t>::vec_t;
+  const int32_t dim = n_size / 2;
+  float* __restrict__ gate = input;
+  float* __restrict__ up = input + dim;
+  vec_op::FP32Vec16 one_vec(1.0);
+
+  DEFINE_FAST_EXP
+
+  for (int32_t m = 0; m < m_size; ++m) {
+    for (int32_t n = 0; n < dim; n += 16) {
+      vec_op::FP32Vec16 gate_vec(gate + n);
+      vec_op::FP32Vec16 up_vec(up + n);
+      auto sigmoid_vec = one_vec / (one_vec + fast_exp(-gate_vec));
+      auto silu = gate_vec * sigmoid_vec;
+      auto gated_output_fp32 = up_vec * silu;
+      scalar_vec_t gated_output = scalar_vec_t(gated_output_fp32);
+      gated_output.save(output + n);
+    }
+    gate += input_stride;
+    up += input_stride;
+    output += output_stride;
+  }
+}
+
+template <typename scalar_t>
+void gelu_and_mul(float* __restrict__ input, scalar_t* __restrict__ output,
+                  const int32_t m_size, const int32_t n_size,
+                  const int32_t input_stride, const int32_t output_stride) {
+  using scalar_vec_t = typename cpu_utils::VecTypeTrait<scalar_t>::vec_t;
+  const int32_t dim = n_size / 2;
+  float* __restrict__ gate = input;
+  float* __restrict__ up = input + dim;
+  vec_op::FP32Vec16 one_vec(1.0);
+  vec_op::FP32Vec16 w1_vec(M_SQRT1_2);
+  vec_op::FP32Vec16 w2_vec(0.5);
+  alignas(64) float temp[16];
+
+  DEFINE_FAST_EXP
+
+  for (int32_t m = 0; m < m_size; ++m) {
+    for (int32_t n = 0; n < dim; n += 16) {
+      vec_op::FP32Vec16 gate_vec(gate + n);
+      vec_op::FP32Vec16 up_vec(up + n);
+      auto er_input_vec = gate_vec * w1_vec;
+
+      er_input_vec.save(temp);
+      for (int32_t i = 0; i < 16; ++i) {
+        temp[i] = std::erf(temp[i]);
+      }
+      vec_op::FP32Vec16 er_vec(temp);
+      auto gelu = gate_vec * w2_vec * (one_vec + er_vec);
+      auto gated_output_fp32 = up_vec * gelu;
+      scalar_vec_t gated_output = scalar_vec_t(gated_output_fp32);
+      gated_output.save(output + n);
+    }
+    gate += input_stride;
+    up += input_stride;
+    output += output_stride;
+  }
+}
+
+template <typename scalar_t>
+void gelu_tanh_and_mul(float* __restrict__ input, scalar_t* __restrict__ output,
+                       const int32_t m_size, const int32_t n_size,
+                       const int32_t input_stride,
+                       const int32_t output_stride) {
+  using scalar_vec_t = typename cpu_utils::VecTypeTrait<scalar_t>::vec_t;
+  const int32_t dim = n_size / 2;
+  float* __restrict__ gate = input;
+  float* __restrict__ up = input + dim;
+  vec_op::FP32Vec16 one_vec(1.0);
+  vec_op::FP32Vec16 w1_vec(0.7978845608028654);
+  vec_op::FP32Vec16 w2_vec(0.5);
+  vec_op::FP32Vec16 w3_vec(0.044715);
+
+  for (int32_t m = 0; m < m_size; ++m) {
+    for (int32_t n = 0; n < dim; n += 16) {
+      vec_op::FP32Vec16 gate_vec(gate + n);
+      vec_op::FP32Vec16 up_vec(up + n);
+      auto gate_pow3_vec = gate_vec * gate_vec * gate_vec;
+      auto inner_vec = w1_vec * (gate_vec + w3_vec * gate_pow3_vec);
+      // Note: can't use fast_exp form because diffusiongemma will generate
+      // wrong results
+      auto tanh_vec = inner_vec.tanh();
+      auto gelu_tanh = gate_vec * w2_vec * (one_vec + tanh_vec);
+      auto gated_output_fp32 = up_vec * gelu_tanh;
+      scalar_vec_t gated_output = scalar_vec_t(gated_output_fp32);
+      gated_output.save(output + n);
+    }
+    gate += input_stride;
+    up += input_stride;
+    output += output_stride;
+  }
+}
+
+template <typename scalar_t>
+FORCE_INLINE void apply_gated_act(const FusedMOEAct act,
+                                  float* __restrict__ input,
+                                  scalar_t* __restrict__ output,
+                                  const int32_t m, const int32_t n,
+                                  const int32_t input_stride,
+                                  const int32_t output_stride) {
+  switch (act) {
+    case FusedMOEAct::SwigluOAIAndMul:
+      swigluoai_and_mul(input, output, m, n, input_stride, output_stride);
+      return;
+    case FusedMOEAct::SiluAndMul:
+      silu_and_mul(input, output, m, n, input_stride, output_stride);
+      return;
+    case FusedMOEAct::GeluAndMul:
+      gelu_and_mul(input, output, m, n, input_stride, output_stride);
+      return;
+    case FusedMOEAct::GeluTanhAndMul:
+      gelu_tanh_and_mul(input, output, m, n, input_stride, output_stride);
+      return;
+    default:
+      TORCH_CHECK(false, "Unsupported act type.");
+  }
+}
+}  // namespace cpu_fused_moe_utils
+
+#endif

csrc/cpu/cpu_fused_moe_int8.cpp

@@ -0,0 +1,647 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+#include "cpu/cpu_arch_macros.h"
+
+#include <algorithm>
+#include <cstdint>
+#include <cstring>
+#include <optional>
+#include <string>
+
+#include "cpu/cpu_fused_moe_activations.hpp"
+#include "cpu/cpu_types.hpp"
+#include "cpu/micro_gemm/cpu_micro_gemm_impl.hpp"
+#include "cpu/utils.hpp"
+
+#if defined(ARM_I8MM_SUPPORT) && defined(ARM_BF16_SUPPORT)
+  #include "cpu/micro_gemm/cpu_micro_gemm_int8_neon.hpp"
+  #define NEON_DISPATCH(SCALAR_TYPE, ...)                                   \
+    case cpu_utils::ISA::NEON: {                                            \
+      using gemm_t =                                                        \
+          cpu_micro_gemm::MicroGemmINT8<cpu_utils::ISA::NEON, SCALAR_TYPE>; \
+      return __VA_ARGS__();                                                 \
+    }
+#else
+  #define NEON_DISPATCH(SCALAR_TYPE, ...) case cpu_utils::ISA::NEON:
+#endif
+
+#define CPU_INT8_ISA_DISPATCH_IMPL(ISA_TYPE, SCALAR_TYPE, ...) \
+  [&] {                                                        \
+    switch (ISA_TYPE) {                                        \
+      NEON_DISPATCH(SCALAR_TYPE, __VA_ARGS__)                  \
+      default: {                                               \
+        TORCH_CHECK(false, "Invalid CPU ISA type.");           \
+      }                                                        \
+    }                                                          \
+  }()
+
+namespace {
+using cpu_fused_moe_utils::apply_gated_act;
+using cpu_fused_moe_utils::FusedMOEAct;
+
+template <typename gemm_t>
+void prepack_moe_weight_int8_impl(const int8_t* __restrict__ weight_ptr,
+                                  int8_t* __restrict__ packed_weight_ptr,
+                                  const int32_t expert_num,
+                                  const int32_t output_size,
+                                  const int32_t input_size,
+                                  const int64_t expert_stride) {
+#pragma omp parallel for
+  for (int32_t e_idx = 0; e_idx < expert_num; ++e_idx) {
+    gemm_t::pack_weight(weight_ptr + expert_stride * e_idx,
+                        packed_weight_ptr + expert_stride * e_idx, output_size,
+                        input_size);
+  }
+}
+
+// INT8 MoE kernel, based on the original BF16 kernel in cpu_fused_moe.cpp
+template <typename scalar_t, typename gemm_t>
+void fused_moe_int8_impl(
+    scalar_t* __restrict__ output, const scalar_t* __restrict__ input,
+    const int8_t* __restrict__ w13, const int8_t* __restrict__ w2,
+    const float* __restrict__ w13_scales, const float* __restrict__ w2_scales,
+    scalar_t* __restrict__ w13_bias, scalar_t* __restrict__ w2_bias,
+    const float* __restrict__ topk_weights, const int32_t* __restrict__ topk_id,
+    const FusedMOEAct act_type, const int32_t token_num,
+    const int32_t expert_num, const int32_t topk_num,
+    const int32_t input_size_13, const int32_t output_size_13,
+    const int32_t input_size_2, const int32_t output_size_2,
+    const bool skip_weighted) {
+  using scalar_vec_t = typename cpu_utils::VecTypeTrait<scalar_t>::vec_t;
+  constexpr int32_t gemm_n_tile_size = gemm_t::NSize;
+  constexpr int32_t gemm_m_tile_size = gemm_t::MaxMSize;
+  constexpr int32_t min_w13_n_tile_size = 2 * gemm_n_tile_size;
+
+  TORCH_CHECK_EQ(input_size_13 % gemm_t::K, 0);
+  TORCH_CHECK_EQ(input_size_2 % gemm_t::K, 0);
+  TORCH_CHECK_EQ(output_size_13 % min_w13_n_tile_size, 0);
+  TORCH_CHECK_EQ(output_size_2 % gemm_n_tile_size, 0);
+  TORCH_CHECK_EQ(output_size_13 / 2, input_size_2);
+
+  const int32_t thread_num = cpu_utils::get_max_threads();
+  const int32_t w13_input_buffer_size = cpu_utils::round_up<64>(
+      gemm_m_tile_size * input_size_13 * sizeof(int8_t));
+  const int32_t w2_input_buffer_size =
+      cpu_utils::round_up<64>(gemm_m_tile_size * input_size_2 * sizeof(int8_t));
+
+  const int32_t w13_n_tile_size = [&]() {
+    const int64_t cache_size = cpu_utils::get_available_l2_size();
+    const int32_t n_size_cache_limit =
+        (cache_size - w13_input_buffer_size) /
+        (gemm_m_tile_size * sizeof(float) + input_size_13 * sizeof(int8_t));
+    const int32_t n_size_thread_limit =
+        output_size_13 / std::max(1, thread_num / topk_num);
+    const int32_t n_size = cpu_utils::round_down<min_w13_n_tile_size>(
+        std::min(n_size_cache_limit, n_size_thread_limit));
+    return std::max(n_size, min_w13_n_tile_size);
+  }();
+
+  const int32_t w2_n_tile_size = [&]() {
+    const int64_t cache_size = cpu_utils::get_available_l2_size();
+    const int32_t n_size_cache_limit =
+        (cache_size - w2_input_buffer_size) / (input_size_2 * sizeof(int8_t));
+    const int32_t n_size_thread_limit =
+        output_size_2 / std::max(1, thread_num / topk_num);
+    const int32_t n_size = cpu_utils::round_down<gemm_n_tile_size>(
+        std::min(n_size_cache_limit, n_size_thread_limit));
+    return std::max(n_size, gemm_n_tile_size);
+  }();
+
+  int32_t common_buffer_offset = 0;
+  const int32_t token_num_per_group_buffer_offset = common_buffer_offset;
+  common_buffer_offset += cpu_utils::round_up<64>(expert_num * sizeof(int32_t));
+  const int32_t cu_token_num_per_group_buffer_offset = common_buffer_offset;
+  common_buffer_offset +=
+      cpu_utils::round_up<64>((expert_num + 1) * sizeof(int32_t));
+  const int32_t expanded_token_num = token_num * topk_num;
+  const int32_t expand_token_id_buffer_offset = common_buffer_offset;
+  common_buffer_offset +=
+      cpu_utils::round_up<64>(expanded_token_num * sizeof(int32_t));
+  const int32_t expand_token_id_index_buffer_offset = common_buffer_offset;
+  common_buffer_offset +=
+      cpu_utils::round_up<64>(expanded_token_num * sizeof(int32_t));
+  const int32_t input_quant_buffer_offset = common_buffer_offset;
+  common_buffer_offset +=
+      cpu_utils::round_up<64>(token_num * input_size_13 * sizeof(int8_t));
+  const int32_t input_scale_buffer_offset = common_buffer_offset;
+  common_buffer_offset += cpu_utils::round_up<64>(token_num * sizeof(float));
+  const int32_t w13_gemm_output_buffer_offset = common_buffer_offset;
+  common_buffer_offset += cpu_utils::round_up<64>(
+      expanded_token_num * input_size_2 * sizeof(scalar_t));
+  const int32_t w13_output_scale_buffer_offset = common_buffer_offset;
+  common_buffer_offset +=
+      cpu_utils::round_up<64>(expanded_token_num * sizeof(float));
+  const int32_t w2_gemm_output_buffer_offset = common_buffer_offset;
+  common_buffer_offset += cpu_utils::round_up<64>(
+      expanded_token_num * output_size_2 * sizeof(float));
+
+  int32_t gemm_thread_buffer_offset = 0;
+  const int32_t gemm_input_buffer_offset = gemm_thread_buffer_offset;
+  gemm_thread_buffer_offset +=
+      std::max(w13_input_buffer_size, w2_input_buffer_size);
+  const int32_t gemm_output_buffer_offset = gemm_thread_buffer_offset;
+  gemm_thread_buffer_offset += cpu_utils::round_up<64>(
+      gemm_m_tile_size * std::max(w13_n_tile_size, w2_n_tile_size) *
+      sizeof(int32_t));
+
+  const int32_t ws_output_buffer_offset = 0;
+  const int32_t ws_thread_buffer_size =
+      cpu_utils::round_up<64>(output_size_2 * sizeof(float));
+  const int32_t thread_buffer_size =
+      std::max(gemm_thread_buffer_offset, ws_thread_buffer_size);
+  const int32_t buffer_size =
+      common_buffer_offset + thread_buffer_size * thread_num;
+  cpu_utils::ScratchPadManager::get_scratchpad_manager()->realloc(buffer_size);
+  uint8_t* common_buffer_start =
+      cpu_utils::ScratchPadManager::get_scratchpad_manager()
+          ->get_data<uint8_t>();
+  uint8_t* thread_buffer_start = common_buffer_start + common_buffer_offset;
+
+  int32_t* __restrict__ token_num_per_group_buffer = reinterpret_cast<int32_t*>(
+      common_buffer_start + token_num_per_group_buffer_offset);
+  int32_t* __restrict__ cu_token_num_per_group_buffer =
+      reinterpret_cast<int32_t*>(common_buffer_start +
+                                 cu_token_num_per_group_buffer_offset);
+  int32_t* __restrict__ expand_token_id_buffer = reinterpret_cast<int32_t*>(
+      common_buffer_start + expand_token_id_buffer_offset);
+  int32_t* __restrict__ expand_token_id_index_buffer =
+      reinterpret_cast<int32_t*>(common_buffer_start +
+                                 expand_token_id_index_buffer_offset);
+  int8_t* __restrict__ input_quant_buffer = reinterpret_cast<int8_t*>(
+      common_buffer_start + input_quant_buffer_offset);
+  float* __restrict__ input_scale_buffer =
+      reinterpret_cast<float*>(common_buffer_start + input_scale_buffer_offset);
+
+  std::memset(token_num_per_group_buffer, 0, expert_num * sizeof(int32_t));
+  for (int32_t i = 0; i < expanded_token_num; ++i) {
+    ++token_num_per_group_buffer[topk_id[i]];
+  }
+
+  int32_t token_num_sum = 0;
+  cu_token_num_per_group_buffer[0] = 0;
+  int32_t* token_index_buffer = cu_token_num_per_group_buffer + 1;
+  for (int32_t i = 0; i < expert_num; ++i) {
+    token_index_buffer[i] = token_num_sum;
+    token_num_sum += token_num_per_group_buffer[i];
+  }
+
+  for (int32_t i = 0; i < token_num; ++i) {
+    const int32_t* curr_topk_id = topk_id + i * topk_num;
+    int32_t* curr_index_buffer = expand_token_id_index_buffer + i * topk_num;
+    for (int32_t j = 0; j < topk_num; ++j) {
+      const int32_t curr_expert_id = curr_topk_id[j];
+      const int32_t curr_index = token_index_buffer[curr_expert_id]++;
+      expand_token_id_buffer[curr_index] = i;
+      curr_index_buffer[j] = curr_index;
+    }
+  }
+
+// quantize inputs
+#pragma omp parallel for
+  for (int32_t token_idx = 0; token_idx < token_num; ++token_idx) {
+    gemm_t::quantize_row(input + token_idx * input_size_13,
+                         input_quant_buffer + token_idx * input_size_13,
+                         input_scale_buffer[token_idx], input_size_13);
+  }
+
+  {
+    alignas(64) cpu_utils::Counter counter;
+    cpu_utils::Counter* counter_ptr = &counter;
+
+// w13 GEMM + act
+#pragma omp parallel for schedule(static, 1)
+    for (int32_t thread_id = 0; thread_id < thread_num; ++thread_id) {
+      const int32_t task_num_per_expert =
+          (output_size_13 + w13_n_tile_size - 1) / w13_n_tile_size;
+      const int32_t task_num = task_num_per_expert * expert_num;
+      uint8_t* __restrict__ thread_buffer =
+          thread_buffer_start + thread_id * thread_buffer_size;
+      int8_t* __restrict__ gemm_input_buffer =
+          reinterpret_cast<int8_t*>(thread_buffer + gemm_input_buffer_offset);
+      float* __restrict__ gemm_output_buffer =
+          reinterpret_cast<float*>(thread_buffer + gemm_output_buffer_offset);
+      auto* __restrict__ w13_gemm_output_buffer = reinterpret_cast<scalar_t*>(
+          common_buffer_start + w13_gemm_output_buffer_offset);
+      gemm_t gemm;
+
+      const int32_t w13_n_group_stride =
+          gemm_t::WeightOCGroupSize * input_size_13;
+      const int32_t w13_n_tile_stride = gemm_n_tile_size * input_size_13;
+
+      for (;;) {
+        const int32_t task_id = counter_ptr->acquire_counter();
+        if (task_id >= task_num) {
+          break;
+        }
+        const int32_t curr_expert_id = task_id / task_num_per_expert;
+        const int32_t curr_output_group_id = task_id % task_num_per_expert;
+        const int32_t curr_token_num =
+            token_num_per_group_buffer[curr_expert_id];
+        if (curr_token_num == 0) {
+          continue;
+        }
+
+        const int32_t actual_n_tile_size =
+            std::min(w13_n_tile_size,
+                     output_size_13 - curr_output_group_id * w13_n_tile_size);
+        const int32_t* __restrict__ curr_expand_token_id_buffer =
+            expand_token_id_buffer +
+            cu_token_num_per_group_buffer[curr_expert_id];
+        scalar_t* __restrict__ curr_w13_gemm_output_buffer =
+            w13_gemm_output_buffer +
+            cu_token_num_per_group_buffer[curr_expert_id] * input_size_2 +
+            curr_output_group_id * w13_n_tile_size / 2;
+
+        const int8_t* w13_weight_ptr_0 = nullptr;
+        const int8_t* w13_weight_ptr_1 = nullptr;
+        const float* w13_scale_ptr_0 = nullptr;
+        const float* w13_scale_ptr_1 = nullptr;
+        scalar_t* w13_bias_ptr_0 = nullptr;
+        scalar_t* w13_bias_ptr_1 = nullptr;
+        if (act_type == FusedMOEAct::SwigluOAIAndMul) {
+          const int32_t output_offset = curr_output_group_id * w13_n_tile_size;
+          w13_weight_ptr_0 = w13 +
+                             curr_expert_id * input_size_13 * output_size_13 +
+                             output_offset * input_size_13;
+          w13_weight_ptr_1 =
+              w13_weight_ptr_0 + actual_n_tile_size / 2 * input_size_13;
+          w13_scale_ptr_0 =
+              w13_scales + curr_expert_id * output_size_13 + output_offset;
+          w13_scale_ptr_1 = w13_scale_ptr_0 + actual_n_tile_size / 2;
+          if (w13_bias != nullptr) {
+            w13_bias_ptr_0 =
+                w13_bias + curr_expert_id * output_size_13 + output_offset;
+            w13_bias_ptr_1 = w13_bias_ptr_0 + actual_n_tile_size / 2;
+          }
+        } else {
+          const int32_t output_offset =
+              curr_output_group_id * (w13_n_tile_size / 2);
+          w13_weight_ptr_0 = w13 +
+                             curr_expert_id * input_size_13 * output_size_13 +
+                             output_offset * input_size_13;
+          w13_weight_ptr_1 =
+              w13_weight_ptr_0 + output_size_13 / 2 * input_size_13;
+          w13_scale_ptr_0 =
+              w13_scales + curr_expert_id * output_size_13 + output_offset;
+          w13_scale_ptr_1 = w13_scale_ptr_0 + output_size_13 / 2;
+          if (w13_bias != nullptr) {
+            w13_bias_ptr_0 =
+                w13_bias + curr_expert_id * output_size_13 + output_offset;
+            w13_bias_ptr_1 = w13_bias_ptr_0 + output_size_13 / 2;
+          }
+        }
+
+        for (int32_t token_idx = 0; token_idx < curr_token_num;
+             token_idx += gemm_m_tile_size) {
+          const int32_t actual_token_num =
+              std::min(gemm_m_tile_size, curr_token_num - token_idx);
+          const int8_t* input_rows[gemm_m_tile_size];
+          alignas(64) float input_scales[gemm_m_tile_size];
+          // gather and pack
+          for (int32_t i = 0; i < actual_token_num; ++i) {
+            const int32_t curr_token_id = curr_expand_token_id_buffer[i];
+            input_rows[i] = input_quant_buffer + curr_token_id * input_size_13;
+            input_scales[i] = input_scale_buffer[curr_token_id];
+          }
+          gemm_t::pack_input_from_rows(input_rows, gemm_input_buffer,
+                                       actual_token_num, input_size_13);
+          curr_expand_token_id_buffer += actual_token_num;
+
+          const int8_t* w13_weight_ptr_0_iter = w13_weight_ptr_0;
+          const int8_t* w13_weight_ptr_1_iter = w13_weight_ptr_1;
+          const float* w13_scale_ptr_0_iter = w13_scale_ptr_0;
+          const float* w13_scale_ptr_1_iter = w13_scale_ptr_1;
+          scalar_t* w13_bias_ptr_0_iter = w13_bias_ptr_0;
+          scalar_t* w13_bias_ptr_1_iter = w13_bias_ptr_1;
+          float* w13_output_buffer_0_iter = gemm_output_buffer;
+          float* w13_output_buffer_1_iter =
+              gemm_output_buffer + actual_n_tile_size / 2;
+
+          for (int32_t i = 0; i < actual_n_tile_size;
+               i += min_w13_n_tile_size) {
+            auto* output_0_int32 =
+                reinterpret_cast<int32_t*>(w13_output_buffer_0_iter);
+            gemm.gemm(gemm_input_buffer, w13_weight_ptr_0_iter, output_0_int32,
+                      actual_token_num, input_size_13, w13_n_group_stride,
+                      actual_n_tile_size);
+            gemm_t::dequantize_tile(output_0_int32, w13_output_buffer_0_iter,
+                                    input_scales, w13_scale_ptr_0_iter,
+                                    actual_token_num, gemm_n_tile_size,
+                                    actual_n_tile_size);
+            if (w13_bias != nullptr) {
+              cpu_micro_gemm::add_bias_epilogue<gemm_n_tile_size>(
+                  w13_output_buffer_0_iter, w13_output_buffer_0_iter,
+                  w13_bias_ptr_0_iter, actual_token_num, actual_n_tile_size,
+                  actual_n_tile_size);
+              w13_bias_ptr_0_iter += gemm_n_tile_size;
+            }
+
+            auto* output_1_int32 =
+                reinterpret_cast<int32_t*>(w13_output_buffer_1_iter);
+            gemm.gemm(gemm_input_buffer, w13_weight_ptr_1_iter, output_1_int32,
+                      actual_token_num, input_size_13, w13_n_group_stride,
+                      actual_n_tile_size);
+            gemm_t::dequantize_tile(output_1_int32, w13_output_buffer_1_iter,
+                                    input_scales, w13_scale_ptr_1_iter,
+                                    actual_token_num, gemm_n_tile_size,
+                                    actual_n_tile_size);
+            if (w13_bias != nullptr) {
+              cpu_micro_gemm::add_bias_epilogue<gemm_n_tile_size>(
+                  w13_output_buffer_1_iter, w13_output_buffer_1_iter,
+                  w13_bias_ptr_1_iter, actual_token_num, actual_n_tile_size,
+                  actual_n_tile_size);
+              w13_bias_ptr_1_iter += gemm_n_tile_size;
+            }
+
+            w13_weight_ptr_0_iter += w13_n_tile_stride;
+            w13_weight_ptr_1_iter += w13_n_tile_stride;
+            w13_scale_ptr_0_iter += gemm_n_tile_size;
+            w13_scale_ptr_1_iter += gemm_n_tile_size;
+            w13_output_buffer_0_iter += gemm_n_tile_size;
+            w13_output_buffer_1_iter += gemm_n_tile_size;
+          }
+
+          apply_gated_act(act_type, gemm_output_buffer,
+                          curr_w13_gemm_output_buffer, actual_token_num,
+                          actual_n_tile_size, actual_n_tile_size, input_size_2);
+          curr_w13_gemm_output_buffer += gemm_m_tile_size * input_size_2;
+        }
+      }
+    }
+  }
+
+  auto* __restrict__ w13_gemm_output_buffer = reinterpret_cast<scalar_t*>(
+      common_buffer_start + w13_gemm_output_buffer_offset);
+  float* __restrict__ w13_output_scale_buffer = reinterpret_cast<float*>(
+      common_buffer_start + w13_output_scale_buffer_offset);
+
+// quantize w2 inputs - in place
+#pragma omp parallel for
+  for (int32_t token_idx = 0; token_idx < expanded_token_num; ++token_idx) {
+    scalar_t* input_row = w13_gemm_output_buffer + token_idx * input_size_2;
+    int8_t* output_row = reinterpret_cast<int8_t*>(input_row);
+    gemm_t::quantize_row(input_row, output_row,
+                         w13_output_scale_buffer[token_idx], input_size_2);
+  }
+
+  {
+    alignas(64) cpu_utils::Counter counter;
+    cpu_utils::Counter* counter_ptr = &counter;
+
+// w2 gemm
+#pragma omp parallel for schedule(static, 1)
+    for (int32_t thread_id = 0; thread_id < thread_num; ++thread_id) {
+      const int32_t task_num_per_expert =
+          (output_size_2 + w2_n_tile_size - 1) / w2_n_tile_size;
+      const int32_t task_num = task_num_per_expert * expert_num;
+      uint8_t* __restrict__ thread_buffer =
+          thread_buffer_start + thread_id * thread_buffer_size;
+      int8_t* __restrict__ gemm_input_buffer =
+          reinterpret_cast<int8_t*>(thread_buffer + gemm_input_buffer_offset);
+      float* __restrict__ gemm_output_buffer =
+          reinterpret_cast<float*>(thread_buffer + gemm_output_buffer_offset);
+      float* __restrict__ w2_gemm_output_buffer = reinterpret_cast<float*>(
+          common_buffer_start + w2_gemm_output_buffer_offset);
+      gemm_t gemm;
+
+      const int32_t w2_n_group_stride =
+          gemm_t::WeightOCGroupSize * input_size_2;
+      const int32_t w2_n_tile_stride = gemm_n_tile_size * input_size_2;
+
+      for (;;) {
+        const int32_t task_id = counter_ptr->acquire_counter();
+        if (task_id >= task_num) {
+          break;
+        }
+        const int32_t curr_expert_id = task_id / task_num_per_expert;
+        const int32_t curr_output_group_id = task_id % task_num_per_expert;
+        const int32_t curr_token_num =
+            token_num_per_group_buffer[curr_expert_id];
+        if (curr_token_num == 0) {
+          continue;
+        }
+
+        const int32_t actual_n_tile_size =
+            std::min(w2_n_tile_size,
+                     output_size_2 - curr_output_group_id * w2_n_tile_size);
+        scalar_t* __restrict__ curr_w13_gemm_output_buffer =
+            w13_gemm_output_buffer +
+            cu_token_num_per_group_buffer[curr_expert_id] * input_size_2;
+        float* __restrict__ curr_w13_output_scale_buffer =
+            w13_output_scale_buffer +
+            cu_token_num_per_group_buffer[curr_expert_id];
+        float* __restrict__ curr_w2_gemm_output_buffer =
+            w2_gemm_output_buffer +
+            cu_token_num_per_group_buffer[curr_expert_id] * output_size_2 +
+            curr_output_group_id * w2_n_tile_size;
+        const int8_t* __restrict__ w2_weight_ptr =
+            w2 + curr_expert_id * output_size_2 * input_size_2 +
+            curr_output_group_id * w2_n_tile_size * input_size_2;
+        const float* __restrict__ w2_scale_ptr =
+            w2_scales + curr_expert_id * output_size_2 +
+            curr_output_group_id * w2_n_tile_size;
+        scalar_t* w2_bias_ptr = nullptr;
+        if (w2_bias != nullptr) {
+          w2_bias_ptr = w2_bias + curr_expert_id * output_size_2 +
+                        curr_output_group_id * w2_n_tile_size;
+        }
+
+        for (int32_t token_idx = 0; token_idx < curr_token_num;
+             token_idx += gemm_m_tile_size) {
+          const int32_t actual_token_num =
+              std::min(gemm_m_tile_size, curr_token_num - token_idx);
+          const int8_t* input_rows[gemm_m_tile_size];
+          alignas(64) float input_scales[gemm_m_tile_size];
+          for (int32_t i = 0; i < actual_token_num; ++i) {
+            input_rows[i] = reinterpret_cast<const int8_t*>(
+                curr_w13_gemm_output_buffer + i * input_size_2);
+            input_scales[i] = curr_w13_output_scale_buffer[i];
+          }
+          gemm_t::pack_input_from_rows(input_rows, gemm_input_buffer,
+                                       actual_token_num, input_size_2);
+
+          const int8_t* w2_weight_ptr_iter = w2_weight_ptr;
+          const float* w2_scale_ptr_iter = w2_scale_ptr;
+          scalar_t* w2_bias_ptr_iter = w2_bias_ptr;
+          float* curr_w2_gemm_output_buffer_iter = curr_w2_gemm_output_buffer;
+          for (int32_t i = 0; i < actual_n_tile_size; i += gemm_n_tile_size) {
+            auto* output_int32 = reinterpret_cast<int32_t*>(gemm_output_buffer);
+            gemm.gemm(gemm_input_buffer, w2_weight_ptr_iter, output_int32,
+                      actual_token_num, input_size_2, w2_n_group_stride,
+                      gemm_n_tile_size);
+            gemm_t::dequantize_tile(output_int32, gemm_output_buffer,
+                                    input_scales, w2_scale_ptr_iter,
+                                    actual_token_num, gemm_n_tile_size,
+                                    gemm_n_tile_size);
+            if (w2_bias != nullptr) {
+              cpu_micro_gemm::add_bias_epilogue<gemm_n_tile_size>(
+                  gemm_output_buffer, gemm_output_buffer, w2_bias_ptr_iter,
+                  actual_token_num, gemm_n_tile_size, gemm_n_tile_size);
+              w2_bias_ptr_iter += gemm_n_tile_size;
+            }
+            for (int32_t m_idx = 0; m_idx < actual_token_num; ++m_idx) {
+              std::memcpy(
+                  curr_w2_gemm_output_buffer_iter + m_idx * output_size_2,
+                  gemm_output_buffer + m_idx * gemm_n_tile_size,
+                  gemm_n_tile_size * sizeof(float));
+            }
+
+            w2_weight_ptr_iter += w2_n_tile_stride;
+            w2_scale_ptr_iter += gemm_n_tile_size;
+            curr_w2_gemm_output_buffer_iter += gemm_n_tile_size;
+          }
+
+          curr_w13_gemm_output_buffer += gemm_m_tile_size * input_size_2;
+          curr_w13_output_scale_buffer += gemm_m_tile_size;
+          curr_w2_gemm_output_buffer += gemm_m_tile_size * output_size_2;
+        }
+      }
+    }
+  }
+
+  {
+    alignas(64) cpu_utils::Counter counter;
+    cpu_utils::Counter* counter_ptr = &counter;
+
+#pragma omp parallel for schedule(static, 1)
+    for (int32_t thread_id = 0; thread_id < thread_num; ++thread_id) {
+      uint8_t* __restrict__ thread_buffer =
+          thread_buffer_start + thread_id * thread_buffer_size;
+      float* __restrict__ ws_output_buffer =
+          reinterpret_cast<float*>(thread_buffer + ws_output_buffer_offset);
+      float* __restrict__ w2_gemm_output_buffer = reinterpret_cast<float*>(
+          common_buffer_start + w2_gemm_output_buffer_offset);
+
+      for (;;) {
+        const int32_t token_id = counter_ptr->acquire_counter();
+        if (token_id >= token_num) {
+          break;
+        }
+        int32_t* __restrict__ curr_expand_token_id_index_buffer =
+            expand_token_id_index_buffer + token_id * topk_num;
+        const float* __restrict__ curr_weight =
+            topk_weights + token_id * topk_num;
+        const float first_weight = skip_weighted ? 1.0f : curr_weight[0];
+        scalar_t* __restrict__ curr_output_buffer =
+            output + token_id * output_size_2;
+
+        if (topk_num > 1) {
+          int32_t w2_output_idx = curr_expand_token_id_index_buffer[0];
+          float* w2_output_iter =
+              w2_gemm_output_buffer + w2_output_idx * output_size_2;
+          float* ws_output_buffer_iter = ws_output_buffer;
+          vec_op::FP32Vec16 weight_vec(first_weight);
+          for (int32_t i = 0; i < output_size_2; i += 16) {
+            vec_op::FP32Vec16 vec(w2_output_iter);
+            (vec * weight_vec).save(ws_output_buffer_iter);
+            w2_output_iter += 16;
+            ws_output_buffer_iter += 16;
+          }
+
+          for (int32_t idx = 1; idx < topk_num - 1; ++idx) {
+            w2_output_idx = curr_expand_token_id_index_buffer[idx];
+            w2_output_iter =
+                w2_gemm_output_buffer + w2_output_idx * output_size_2;
+            ws_output_buffer_iter = ws_output_buffer;
+            weight_vec = vec_op::FP32Vec16(curr_weight[idx]);
+            for (int32_t i = 0; i < output_size_2; i += 16) {
+              vec_op::FP32Vec16 vec(w2_output_iter);
+              vec_op::FP32Vec16 sum(ws_output_buffer_iter);
+              (sum + vec * weight_vec).save(ws_output_buffer_iter);
+              w2_output_iter += 16;
+              ws_output_buffer_iter += 16;
+            }
+          }
+
+          const int32_t last_idx = topk_num - 1;
+          w2_output_idx = curr_expand_token_id_index_buffer[last_idx];
+          w2_output_iter =
+              w2_gemm_output_buffer + w2_output_idx * output_size_2;
+          ws_output_buffer_iter = ws_output_buffer;
+          scalar_t* curr_output_buffer_iter = curr_output_buffer;
+          weight_vec = vec_op::FP32Vec16(curr_weight[last_idx]);
+          for (int32_t i = 0; i < output_size_2; i += 16) {
+            vec_op::FP32Vec16 vec(w2_output_iter);
+            vec_op::FP32Vec16 sum(ws_output_buffer_iter);
+            scalar_vec_t(sum + vec * weight_vec).save(curr_output_buffer_iter);
+            w2_output_iter += 16;
+            ws_output_buffer_iter += 16;
+            curr_output_buffer_iter += 16;
+          }
+        } else {
+          const int32_t w2_output_idx = curr_expand_token_id_index_buffer[0];
+          float* w2_output_iter =
+              w2_gemm_output_buffer + w2_output_idx * output_size_2;
+          scalar_t* curr_output_buffer_iter = curr_output_buffer;
+          vec_op::FP32Vec16 weight_vec(first_weight);
+          for (int32_t i = 0; i < output_size_2; i += 16) {
+            vec_op::FP32Vec16 vec(w2_output_iter);
+            scalar_vec_t(vec * weight_vec).save(curr_output_buffer_iter);
+            w2_output_iter += 16;
+            curr_output_buffer_iter += 16;
+          }
+        }
+      }
+    }
+  }
+}
+}  // namespace
+
+void prepack_moe_weight_int8(
+    const torch::Tensor& weight,  // [expert_num, output_size, input_size]
+    torch::Tensor& packed_weight, const std::string& isa) {
+  TORCH_CHECK(weight.is_contiguous());
+  const int32_t expert_num = weight.size(0);
+  const int32_t output_size = weight.size(1);
+  const int32_t input_size = weight.size(2);
+  const int64_t expert_stride = weight.stride(0);
+  const cpu_utils::ISA isa_type = cpu_utils::get_isa(isa);
+  TORCH_CHECK_EQ(output_size % 32, 0);
+
+  CPU_INT8_ISA_DISPATCH_IMPL(isa_type, c10::BFloat16, [&]() {
+    TORCH_CHECK_EQ(input_size % gemm_t::K, 0);
+    prepack_moe_weight_int8_impl<gemm_t>(
+        weight.data_ptr<int8_t>(), packed_weight.data_ptr<int8_t>(), expert_num,
+        output_size, input_size, expert_stride);
+  });
+}
+
+void cpu_fused_moe_int8(torch::Tensor& output, const torch::Tensor& input,
+                        const torch::Tensor& w13, const torch::Tensor& w2,
+                        const torch::Tensor& w13_scale,
+                        const torch::Tensor& w2_scale,
+                        const std::optional<torch::Tensor>& w13_bias,
+                        const std::optional<torch::Tensor>& w2_bias,
+                        const torch::Tensor& topk_weights,
+                        const torch::Tensor& topk_id, const bool skip_weighted,
+                        const std::string& act, const std::string& isa) {
+  const int32_t token_num = input.size(0);
+  const int32_t input_size_13 = input.size(1);
+  const int64_t input_stride = input.stride(0);
+  TORCH_CHECK_EQ(input_stride, input_size_13);
+  const int32_t expert_num = w13.size(0);
+  const int32_t output_size_13 = w13.size(1);
+  const int32_t input_size_2 = w2.size(2);
+  const int32_t output_size_2 = w2.size(1);
+  const int32_t topk_num = topk_id.size(1);
+  const FusedMOEAct act_type = cpu_fused_moe_utils::get_act_type(act);
+  const cpu_utils::ISA isa_type = cpu_utils::get_isa(isa);
+  TORCH_CHECK(!skip_weighted || topk_num == 1,
+              "skip_weighted is only supported for topk=1 on CPU");
+
+  VLLM_DISPATCH_FLOATING_TYPES(
+      input.scalar_type(), "cpu_fused_moe_int8", [&]() {
+        CPU_INT8_ISA_DISPATCH_IMPL(isa_type, scalar_t, [&]() {
+          fused_moe_int8_impl<scalar_t, gemm_t>(
+              output.data_ptr<scalar_t>(), input.data_ptr<scalar_t>(),
+              w13.data_ptr<int8_t>(), w2.data_ptr<int8_t>(),
+              w13_scale.data_ptr<float>(), w2_scale.data_ptr<float>(),
+              w13_bias.has_value() ? w13_bias->data_ptr<scalar_t>() : nullptr,
+              w2_bias.has_value() ? w2_bias->data_ptr<scalar_t>() : nullptr,
+              topk_weights.data_ptr<float>(), topk_id.data_ptr<int32_t>(),
+              act_type, token_num, expert_num, topk_num, input_size_13,
+              output_size_13, input_size_2, output_size_2, skip_weighted);
+        });
+      });
+}

csrc/cpu/cpu_types_vsx.hpp

@@ -287,7 +287,7 @@ struct FP32Vec4 : public Vec<FP32Vec4> {
 
   explicit FP32Vec4(__vector float data) : reg(data) {}
 
-  explicit FP32Vec4(const FP32Vec4& data) : reg(data.reg) {}
+  FP32Vec4(const FP32Vec4& data) : reg(data.reg) {}
 };
 
 struct FP32Vec8 : public Vec<FP32Vec8> {
@@ -316,7 +316,7 @@ struct FP32Vec8 : public Vec<FP32Vec8> {
 
   explicit FP32Vec8(f32x4x2_t data) : reg(data) {}
 
-  explicit FP32Vec8(const FP32Vec8& data) {
+  FP32Vec8(const FP32Vec8& data) {
     reg.val[0] = data.reg.val[0];
     reg.val[1] = data.reg.val[1];
   }
@@ -593,7 +593,7 @@ struct FP32Vec16 : public Vec<FP32Vec16> {
   explicit FP32Vec16(bool, const float* ptr) : FP32Vec16(ptr) {}
   explicit FP32Vec16(f32x4x4_t data) : reg(data) {}
 
-  explicit FP32Vec16(const FP32Vec16& data) {
+  FP32Vec16(const FP32Vec16& data) {
     reg.val[0] = data.reg.val[0];
     reg.val[1] = data.reg.val[1];
     reg.val[2] = data.reg.val[2];
@@ -747,6 +747,15 @@ struct FP32Vec16 : public Vec<FP32Vec16> {
                                 vec_abs(reg.val[2]), vec_abs(reg.val[3])}));
   }
 
+  FP32Vec16 exp() const {
+    FP32Vec8 lo(f32x4x2_t{reg.val[0], reg.val[1]});
+    FP32Vec8 hi(f32x4x2_t{reg.val[2], reg.val[3]});
+    auto lo_e = lo.exp();
+    auto hi_e = hi.exp();
+    return FP32Vec16(f32x4x4_t{lo_e.reg.val[0], lo_e.reg.val[1],
+                               hi_e.reg.val[0], hi_e.reg.val[1]});
+  }
+
   float reduce_max() {
     __vector float max01 = vec_max(reg.val[0], reg.val[1]);
     __vector float max23 = vec_max(reg.val[2], reg.val[3]);

csrc/cpu/micro_gemm/cpu_micro_gemm_impl.hpp

@@ -31,6 +31,9 @@ class MicroGemm {
   }
 };
 
+template <cpu_utils::ISA isa, typename scalar_t>
+class MicroGemmINT8;
+
 template <int32_t n_size, typename scalar_t>
 FORCE_INLINE void default_epilogue(float* __restrict__ c_ptr,
                                    scalar_t* __restrict__ d_ptr,

csrc/cpu/micro_gemm/cpu_micro_gemm_int8_neon.hpp

@@ -0,0 +1,424 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+#ifndef CPU_MICRO_GEMM_INT8_NEON_HPP
+#define CPU_MICRO_GEMM_INT8_NEON_HPP
+
+#include <algorithm>
+#include <cstdint>
+
+#include "cpu/micro_gemm/cpu_micro_gemm_impl.hpp"
+
+#include <arm_bf16.h>
+#include <arm_neon.h>
+#include <c10/util/BFloat16.h>
+#include <c10/util/Exception.h>
+#include <c10/util/Half.h>
+
+namespace cpu_micro_gemm {
+
+namespace neon_smmla {
+
+constexpr int32_t K = 8;
+constexpr int32_t Cols = 2;
+constexpr int32_t TileSize = K * Cols;
+
+FORCE_INLINE float32x4x2_t load_as_f32(const float* input) {
+  float32x4x2_t result;
+  result.val[0] = vld1q_f32(input);
+  result.val[1] = vld1q_f32(input + 4);
+  return result;
+}
+
+FORCE_INLINE float32x4x2_t load_as_f32(const c10::Half* input) {
+  const auto input_vec = vld1q_f16(reinterpret_cast<const float16_t*>(input));
+  float32x4x2_t result;
+  result.val[0] = vcvt_f32_f16(vget_low_f16(input_vec));
+  result.val[1] = vcvt_f32_f16(vget_high_f16(input_vec));
+  return result;
+}
+
+FORCE_INLINE float32x4x2_t load_as_f32(const c10::BFloat16* input) {
+  const auto input_vec = vld1q_bf16(reinterpret_cast<const bfloat16_t*>(input));
+  float32x4x2_t result;
+  result.val[0] = vcvt_f32_bf16(vget_low_bf16(input_vec));
+  result.val[1] = vcvt_f32_bf16(vget_high_bf16(input_vec));
+  return result;
+}
+
+FORCE_INLINE void store_acc_rowpair(const int32x4_t acc01,
+                                    const int32x4_t acc23,
+                                    const int32x4_t acc45,
+                                    const int32x4_t acc67,
+                                    int32_t* __restrict__ c_ptr,
+                                    const int64_t ldc, const int32_t m_rows) {
+  if (m_rows == 0) {
+    return;
+  }
+
+  vst1q_s32(c_ptr, vcombine_s32(vget_low_s32(acc01), vget_low_s32(acc23)));
+  vst1q_s32(c_ptr + 4, vcombine_s32(vget_low_s32(acc45), vget_low_s32(acc67)));
+
+  if (m_rows == 2) {
+    vst1q_s32(c_ptr + ldc,
+              vcombine_s32(vget_high_s32(acc01), vget_high_s32(acc23)));
+    vst1q_s32(c_ptr + ldc + 4,
+              vcombine_s32(vget_high_s32(acc45), vget_high_s32(acc67)));
+  }
+}
+
+FORCE_INLINE void gemm_micro_smmla_8x8_packed_a(
+    const int8_t* __restrict__ a_packed, const int8_t* __restrict__ b_packed,
+    int32_t* __restrict__ c_ptr, const int32_t m, const int32_t k_size,
+    const int64_t ldc) {
+  const int32x4_t zero = vdupq_n_s32(0);
+  int32x4_t acc0101 = zero, acc0123 = zero, acc0145 = zero, acc0167 = zero;
+  int32x4_t acc2301 = zero, acc2323 = zero, acc2345 = zero, acc2367 = zero;
+  int32x4_t acc4501 = zero, acc4523 = zero, acc4545 = zero, acc4567 = zero;
+  int32x4_t acc6701 = zero, acc6723 = zero, acc6745 = zero, acc6767 = zero;
+
+  const int8_t* __restrict__ a_tile = a_packed;
+  const int8_t* __restrict__ b_tile = b_packed;
+
+#pragma GCC unroll 8
+  for (int32_t k_idx = 0; k_idx < k_size; k_idx += K) {
+    const int8x16_t a_tile01 = vld1q_s8(a_tile);
+    const int8x16_t a_tile23 = vld1q_s8(a_tile + TileSize);
+    const int8x16_t a_tile45 = vld1q_s8(a_tile + 2 * TileSize);
+    const int8x16_t a_tile67 = vld1q_s8(a_tile + 3 * TileSize);
+
+    const int8x16_t b_tile01 = vld1q_s8(b_tile);
+    const int8x16_t b_tile23 = vld1q_s8(b_tile + TileSize);
+    const int8x16_t b_tile45 = vld1q_s8(b_tile + 2 * TileSize);
+    const int8x16_t b_tile67 = vld1q_s8(b_tile + 3 * TileSize);
+
+    acc0101 = vmmlaq_s32(acc0101, a_tile01, b_tile01);
+    acc2301 = vmmlaq_s32(acc2301, a_tile23, b_tile01);
+    acc4501 = vmmlaq_s32(acc4501, a_tile45, b_tile01);
+    acc6701 = vmmlaq_s32(acc6701, a_tile67, b_tile01);
+
+    acc0123 = vmmlaq_s32(acc0123, a_tile01, b_tile23);
+    acc2323 = vmmlaq_s32(acc2323, a_tile23, b_tile23);
+    acc4523 = vmmlaq_s32(acc4523, a_tile45, b_tile23);
+    acc6723 = vmmlaq_s32(acc6723, a_tile67, b_tile23);
+
+    acc0145 = vmmlaq_s32(acc0145, a_tile01, b_tile45);
+    acc2345 = vmmlaq_s32(acc2345, a_tile23, b_tile45);
+    acc4545 = vmmlaq_s32(acc4545, a_tile45, b_tile45);
+    acc6745 = vmmlaq_s32(acc6745, a_tile67, b_tile45);
+
+    acc0167 = vmmlaq_s32(acc0167, a_tile01, b_tile67);
+    acc2367 = vmmlaq_s32(acc2367, a_tile23, b_tile67);
+    acc4567 = vmmlaq_s32(acc4567, a_tile45, b_tile67);
+    acc6767 = vmmlaq_s32(acc6767, a_tile67, b_tile67);
+
+    a_tile += 4 * TileSize;
+    b_tile += 4 * TileSize;
+  }
+
+  store_acc_rowpair(acc0101, acc0123, acc0145, acc0167, c_ptr, ldc,
+                    std::min(2, m));
+  store_acc_rowpair(acc2301, acc2323, acc2345, acc2367, c_ptr + 2 * ldc, ldc,
+                    std::min(2, std::max(0, m - 2)));
+  store_acc_rowpair(acc4501, acc4523, acc4545, acc4567, c_ptr + 4 * ldc, ldc,
+                    std::min(2, std::max(0, m - 4)));
+  store_acc_rowpair(acc6701, acc6723, acc6745, acc6767, c_ptr + 6 * ldc, ldc,
+                    std::min(2, std::max(0, m - 6)));
+}
+
+FORCE_INLINE void gemm_micro_smmla_4x16_packed_a(
+    const int8_t* __restrict__ a_packed, const int8_t* __restrict__ b_packed,
+    int32_t* __restrict__ c_ptr, const int32_t m, const int32_t k_size,
+    const int64_t b_n_group_stride, const int64_t ldc) {
+  const int32_t m_rows_01 = std::min(2, m);
+  const int32_t m_rows_23 = std::min(2, std::max(0, m - 2));
+  const int32x4_t zero = vdupq_n_s32(0);
+
+  int32x4_t acc0101 = zero, acc0123 = zero, acc0145 = zero, acc0167 = zero;
+  int32x4_t acc2301 = zero, acc2323 = zero, acc2345 = zero, acc2367 = zero;
+  int32x4_t acc0189 = zero, acc011011 = zero, acc011213 = zero,
+            acc011415 = zero;
+  int32x4_t acc2389 = zero, acc231011 = zero, acc231213 = zero,
+            acc231415 = zero;
+
+  const int8_t* __restrict__ a_tile = a_packed;
+  // note: b packs 8 panels contiguously, so we need 2 b_tile ptrs
+  // for the 4x16 microkernel
+  const int8_t* __restrict__ b_tile0 = b_packed;
+  const int8_t* __restrict__ b_tile1 = b_packed + b_n_group_stride;
+
+#pragma GCC unroll 8
+  for (int32_t k_idx = 0; k_idx < k_size; k_idx += K) {
+    const int8x16_t a_tile01 = vld1q_s8(a_tile);
+    const int8x16_t a_tile23 = vld1q_s8(a_tile + TileSize);
+    const int8x16_t b_tile01 = vld1q_s8(b_tile0);
+    const int8x16_t b_tile23 = vld1q_s8(b_tile0 + TileSize);
+    const int8x16_t b_tile45 = vld1q_s8(b_tile0 + 2 * TileSize);
+    const int8x16_t b_tile67 = vld1q_s8(b_tile0 + 3 * TileSize);
+    const int8x16_t b_tile89 = vld1q_s8(b_tile1);
+    const int8x16_t b_tile1011 = vld1q_s8(b_tile1 + TileSize);
+    const int8x16_t b_tile1213 = vld1q_s8(b_tile1 + 2 * TileSize);
+    const int8x16_t b_tile1415 = vld1q_s8(b_tile1 + 3 * TileSize);
+
+    acc0101 = vmmlaq_s32(acc0101, a_tile01, b_tile01);
+    acc2301 = vmmlaq_s32(acc2301, a_tile23, b_tile01);
+    acc0123 = vmmlaq_s32(acc0123, a_tile01, b_tile23);
+    acc2323 = vmmlaq_s32(acc2323, a_tile23, b_tile23);
+
+    acc0145 = vmmlaq_s32(acc0145, a_tile01, b_tile45);
+    acc2345 = vmmlaq_s32(acc2345, a_tile23, b_tile45);
+    acc0167 = vmmlaq_s32(acc0167, a_tile01, b_tile67);
+    acc2367 = vmmlaq_s32(acc2367, a_tile23, b_tile67);
+
+    acc0189 = vmmlaq_s32(acc0189, a_tile01, b_tile89);
+    acc2389 = vmmlaq_s32(acc2389, a_tile23, b_tile89);
+    acc011011 = vmmlaq_s32(acc011011, a_tile01, b_tile1011);
+    acc231011 = vmmlaq_s32(acc231011, a_tile23, b_tile1011);
+
+    acc011213 = vmmlaq_s32(acc011213, a_tile01, b_tile1213);
+    acc231213 = vmmlaq_s32(acc231213, a_tile23, b_tile1213);
+    acc011415 = vmmlaq_s32(acc011415, a_tile01, b_tile1415);
+    acc231415 = vmmlaq_s32(acc231415, a_tile23, b_tile1415);
+
+    a_tile += 2 * TileSize;
+    b_tile0 += 4 * TileSize;
+    b_tile1 += 4 * TileSize;
+  }
+
+  // rows 0-1, columns 0-7
+  store_acc_rowpair(acc0101, acc0123, acc0145, acc0167, c_ptr, ldc, m_rows_01);
+  // rows 0-1, columns 8-15
+  store_acc_rowpair(acc0189, acc011011, acc011213, acc011415, c_ptr + 8, ldc,
+                    m_rows_01);
+  // rows 2-3, columns 0-7
+  store_acc_rowpair(acc2301, acc2323, acc2345, acc2367, c_ptr + 2 * ldc, ldc,
+                    m_rows_23);
+  // rows 2-3, columns 8-15
+  store_acc_rowpair(acc2389, acc231011, acc231213, acc231415,
+                    c_ptr + 2 * ldc + 8, ldc, m_rows_23);
+}
+
+}  // namespace neon_smmla
+
+template <typename scalar_t>
+class MicroGemmINT8<cpu_utils::ISA::NEON, scalar_t> {
+ public:
+  static constexpr int32_t K = neon_smmla::K;
+  static constexpr int32_t Mr = 8;
+  static constexpr int32_t Nr = 8;
+  static constexpr int32_t NrGemv = 16;
+  static constexpr int32_t MaxMSize = 8;
+  static constexpr int32_t NSize = 32;
+  static constexpr int32_t WeightOCGroupSize = Nr;
+  static_assert(MaxMSize % Mr == 0);
+
+  static FORCE_INLINE void quantize_row(const scalar_t* input, int8_t* output,
+                                        float& scale, const int32_t size) {
+    TORCH_CHECK_EQ(size % K, 0);
+    float32x4_t max_vec = vdupq_n_f32(0.0f);
+
+    for (int32_t i = 0; i < size; i += K) {
+      const float32x4x2_t input_vec = neon_smmla::load_as_f32(input + i);
+      max_vec = vmaxq_f32(max_vec, vabsq_f32(input_vec.val[0]));
+      max_vec = vmaxq_f32(max_vec, vabsq_f32(input_vec.val[1]));
+    }
+
+    const float abs_max = std::max(vmaxvq_f32(max_vec), 1.0e-7f);
+    scale = abs_max / 127.0f;
+    const float32x4_t inv_scale_vec = vdupq_n_f32(127.0f / abs_max);
+
+    for (int32_t i = 0; i < size; i += K) {
+      const float32x4x2_t input_vec = neon_smmla::load_as_f32(input + i);
+      const int32x4_t output_low =
+          vcvtnq_s32_f32(vmulq_f32(input_vec.val[0], inv_scale_vec));
+      const int32x4_t output_high =
+          vcvtnq_s32_f32(vmulq_f32(input_vec.val[1], inv_scale_vec));
+      const int16x8_t output_s16 =
+          vcombine_s16(vqmovn_s32(output_low), vqmovn_s32(output_high));
+      vst1_s8(output + i, vqmovn_s16(output_s16));
+    }
+  }
+
+  // with current code, fusing this into the gemm micro kernel didn't move the
+  // needle
+  static FORCE_INLINE void dequantize_tile(
+      int32_t* input, float* output, const float* __restrict__ input_scales,
+      const float* __restrict__ weight_scales, const int32_t m, const int32_t n,
+      const int32_t stride) {
+    TORCH_CHECK_EQ(n % 4, 0);
+    for (int32_t m_idx = 0; m_idx < m; ++m_idx) {
+      const float32x4_t input_scale_vec = vdupq_n_f32(input_scales[m_idx]);
+      for (int32_t n_idx = 0; n_idx < n; n_idx += 4) {
+        const int32x4_t input_vec = vld1q_s32(input + m_idx * stride + n_idx);
+        const float32x4_t weight_scale_vec = vld1q_f32(weight_scales + n_idx);
+        const float32x4_t output_vec =
+            vmulq_f32(vcvtq_f32_s32(input_vec),
+                      vmulq_f32(input_scale_vec, weight_scale_vec));
+        vst1q_f32(output + m_idx * stride + n_idx, output_vec);
+      }
+    }
+  }
+
+  // physical layout [
+  //  M / (8 or 4); Mr is 8 or 4
+  //  K / 8; K for smmla is 8
+  //  4,   ; 4 row-pairs for each 8 rows
+  //  2,   ; row-pair is 2 rows
+  //  4    ; 4 elements per row
+  // ]
+  static void pack_input_from_rows(const int8_t* const* __restrict__ rows,
+                                   int8_t* __restrict__ a_packed,
+                                   const int32_t m, const int32_t k) {
+    TORCH_CHECK(m > 0 && m <= MaxMSize);
+    TORCH_CHECK(k % K == 0);
+    const int8x8_t zero = vdup_n_s8(0);
+
+    for (int32_t row_base = 0; row_base < m; row_base += Mr) {
+      const int32_t panel_m = std::min(Mr, m - row_base);
+      const int8_t* const* panel_rows = rows + row_base;
+      int8_t* __restrict__ out = a_packed + row_base * k;
+
+      // fast path for full 8-row panels (fast path for 4-row panels didn't move
+      // the needle)
+      if (panel_m == Mr) {
+        const int8_t* __restrict__ row0 = panel_rows[0];
+        const int8_t* __restrict__ row1 = panel_rows[1];
+        const int8_t* __restrict__ row2 = panel_rows[2];
+        const int8_t* __restrict__ row3 = panel_rows[3];
+        const int8_t* __restrict__ row4 = panel_rows[4];
+        const int8_t* __restrict__ row5 = panel_rows[5];
+        const int8_t* __restrict__ row6 = panel_rows[6];
+        const int8_t* __restrict__ row7 = panel_rows[7];
+        int32_t k_idx = 0;
+        for (; k_idx + 2 * K <= k; k_idx += 2 * K) {
+          int8_t* __restrict__ block0 = out;
+          int8_t* __restrict__ block1 = out + 4 * neon_smmla::TileSize;
+
+          int8x16_t a0 = vld1q_s8(row0 + k_idx);
+          int8x16_t a1 = vld1q_s8(row1 + k_idx);
+          vst1q_s8(block0, vcombine_s8(vget_low_s8(a0), vget_low_s8(a1)));
+          vst1q_s8(block1, vcombine_s8(vget_high_s8(a0), vget_high_s8(a1)));
+
+          a0 = vld1q_s8(row2 + k_idx);
+          a1 = vld1q_s8(row3 + k_idx);
+          vst1q_s8(block0 + neon_smmla::TileSize,
+                   vcombine_s8(vget_low_s8(a0), vget_low_s8(a1)));
+          vst1q_s8(block1 + neon_smmla::TileSize,
+                   vcombine_s8(vget_high_s8(a0), vget_high_s8(a1)));
+
+          a0 = vld1q_s8(row4 + k_idx);
+          a1 = vld1q_s8(row5 + k_idx);
+          vst1q_s8(block0 + 2 * neon_smmla::TileSize,
+                   vcombine_s8(vget_low_s8(a0), vget_low_s8(a1)));
+          vst1q_s8(block1 + 2 * neon_smmla::TileSize,
+                   vcombine_s8(vget_high_s8(a0), vget_high_s8(a1)));
+
+          a0 = vld1q_s8(row6 + k_idx);
+          a1 = vld1q_s8(row7 + k_idx);
+          vst1q_s8(block0 + 3 * neon_smmla::TileSize,
+                   vcombine_s8(vget_low_s8(a0), vget_low_s8(a1)));
+          vst1q_s8(block1 + 3 * neon_smmla::TileSize,
+                   vcombine_s8(vget_high_s8(a0), vget_high_s8(a1)));
+
+          out += 8 * neon_smmla::TileSize;
+        }
+
+        for (; k_idx < k; k_idx += K) {
+          int8x8_t a0 = vld1_s8(row0 + k_idx);
+          int8x8_t a1 = vld1_s8(row1 + k_idx);
+          vst1q_s8(out, vcombine_s8(a0, a1));
+
+          a0 = vld1_s8(row2 + k_idx);
+          a1 = vld1_s8(row3 + k_idx);
+          vst1q_s8(out + neon_smmla::TileSize, vcombine_s8(a0, a1));
+
+          a0 = vld1_s8(row4 + k_idx);
+          a1 = vld1_s8(row5 + k_idx);
+          vst1q_s8(out + 2 * neon_smmla::TileSize, vcombine_s8(a0, a1));
+
+          a0 = vld1_s8(row6 + k_idx);
+          a1 = vld1_s8(row7 + k_idx);
+          vst1q_s8(out + 3 * neon_smmla::TileSize, vcombine_s8(a0, a1));
+
+          out += 4 * neon_smmla::TileSize;
+        }
+        continue;
+      }
+
+      const int32_t row_pairs = (panel_m <= 4) ? 2 : Mr / 2;
+      for (int32_t k_idx = 0; k_idx < k; k_idx += K) {
+        for (int32_t pair_idx = 0; pair_idx < row_pairs; ++pair_idx) {
+          const int32_t row_idx = pair_idx * 2;
+          const int8x8_t row0 =
+              (row_idx < panel_m) ? vld1_s8(panel_rows[row_idx] + k_idx) : zero;
+          const int8x8_t row1 = (row_idx + 1 < panel_m)
+                                    ? vld1_s8(panel_rows[row_idx + 1] + k_idx)
+                                    : zero;
+          vst1q_s8(out, vcombine_s8(row0, row1));
+          out += neon_smmla::TileSize;
+        }
+      }
+    }
+  }
+
+  // physical layout [
+  //  N / 8; Nr is 8
+  //  K / 8; K for smmla is 8
+  //  4,   ; 4 col-pairs for each 8 cols
+  //  2,   ; col-pair is 2 cols
+  //  4    ; 4 elements per col
+  // ]
+  static void pack_weight(const int8_t* __restrict__ weight,
+                          int8_t* __restrict__ packed_weight,
+                          const int32_t output_size, const int32_t input_size) {
+    TORCH_CHECK(output_size % NSize == 0);
+    TORCH_CHECK(input_size % K == 0);
+
+    for (int32_t o_idx = 0; o_idx < output_size; o_idx += Nr) {
+      int8_t* __restrict__ dst = packed_weight + o_idx * input_size;
+      for (int32_t k_idx = 0; k_idx < input_size; k_idx += K) {
+        for (int32_t pair_idx = 0; pair_idx < Nr;
+             pair_idx += neon_smmla::Cols) {
+          const int8_t* __restrict__ row0 =
+              weight + (o_idx + pair_idx) * input_size + k_idx;
+          const int8_t* __restrict__ row1 = row0 + input_size;
+          vst1q_s8(dst, vcombine_s8(vld1_s8(row0), vld1_s8(row1)));
+          dst += neon_smmla::TileSize;
+        }
+      }
+    }
+  }
+
+  void gemm(const int8_t* __restrict__ a_packed,
+            const int8_t* __restrict__ b_packed, int32_t* __restrict__ c,
+            const int32_t m, const int32_t k, const int64_t b_n_group_stride,
+            const int64_t ldc) const {
+    TORCH_CHECK(m > 0 && m <= MaxMSize);
+    TORCH_CHECK(k % K == 0);
+
+    for (int32_t n_idx = 0; n_idx < NSize; n_idx += NrGemv) {
+      const int8_t* __restrict__ b_panel = b_packed + n_idx * k;
+
+      for (int32_t row_base = 0; row_base < m; row_base += Mr) {
+        const int32_t panel_m = std::min(Mr, m - row_base);
+        const int8_t* __restrict__ a_panel = a_packed + row_base * k;
+        int32_t* __restrict__ c_panel = c + row_base * ldc + n_idx;
+
+        if (panel_m <= 4) {
+          neon_smmla::gemm_micro_smmla_4x16_packed_a(
+              a_panel, b_panel, c_panel, panel_m, k, b_n_group_stride, ldc);
+        } else {
+          neon_smmla::gemm_micro_smmla_8x8_packed_a(a_panel, b_panel, c_panel,
+                                                    panel_m, k, ldc);
+          neon_smmla::gemm_micro_smmla_8x8_packed_a(
+              a_panel, b_panel + b_n_group_stride, c_panel + Nr, panel_m, k,
+              ldc);
+        }
+      }
+    }
+  }
+};
+
+}  // namespace cpu_micro_gemm
+
+#endif

csrc/cpu/micro_gemm/cpu_micro_gemm_neon.hpp

@@ -1,3 +1,6 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
 #ifndef CPU_MICRO_GEMM_NEON_HPP
 #define CPU_MICRO_GEMM_NEON_HPP
 
@@ -16,9 +19,6 @@ namespace {
 constexpr int32_t K = 4;
 constexpr int32_t Cols = 2;
 constexpr int32_t TileSize = K * Cols;
-constexpr int32_t Mr = 8;
-constexpr int32_t Nr = 8;
-constexpr int32_t Nr_gemv = 16;
 
 // a = [a0, a1, a2, a3], b = [b0, b1, b2, b3] -> [a0, a1, b0, b1]
 FORCE_INLINE float32x4_t zip1_f32x4(const float32x4_t a, const float32x4_t b) {
@@ -132,7 +132,7 @@ FORCE_INLINE void gemm_micro_bfmmla_8x8_packed_a(
     acc6767 = vbfmmlaq_f32(acc6767, a_tile67, b_tile67);
 
     a_tile += 4 * TileSize;
-    b_tile += Nr * K;
+    b_tile += 4 * TileSize;
   }
 
   store_acc_rowpair(acc0101, acc0123, acc0145, acc0167, c_ptr, ldc,
@@ -205,8 +205,8 @@ FORCE_INLINE void gemm_micro_bfmmla_4x16_packed_a(
     acc231415 = vbfmmlaq_f32(acc231415, a_tile23, b_tile1415);
 
     a_tile += 2 * TileSize;
-    b_tile0 += Nr * K;
-    b_tile1 += Nr * K;
+    b_tile0 += 4 * TileSize;
+    b_tile1 += 4 * TileSize;
   }
 
   store_acc_rowpair(acc0101, acc0123, acc0145, acc0167, c_ptr, ldc, m_rows_01);
@@ -223,6 +223,9 @@ FORCE_INLINE void gemm_micro_bfmmla_4x16_packed_a(
 template <typename scalar_t>
 class MicroGemm<cpu_utils::ISA::NEON, scalar_t> {
  public:
+  static constexpr int32_t Mr = 8;
+  static constexpr int32_t Nr = 8;
+  static constexpr int32_t NrGemv = 16;
   static constexpr int32_t MaxMSize = 8;
   static constexpr int32_t NSize = 32;
   static constexpr int32_t WeightOCGroupSize = Nr;
@@ -246,6 +249,9 @@ class MicroGemm<cpu_utils::ISA::NEON, c10::BFloat16> {
  public:
   using scalar_t = c10::BFloat16;
 
+  static constexpr int32_t Mr = 8;
+  static constexpr int32_t Nr = 8;
+  static constexpr int32_t NrGemv = 16;
   static constexpr int32_t MaxMSize = 8;
   static constexpr int32_t NSize = 32;
   static constexpr int32_t WeightOCGroupSize = Nr;
@@ -253,7 +259,7 @@ class MicroGemm<cpu_utils::ISA::NEON, c10::BFloat16> {
 
  public:
   // physical layout [
-  //  M / 8; Mr is 8
+  //  M / (8 or 4); Mr is 8 or 4
   //  K / 4; K for bfmmla is 4
   //  4,   ; 4 row-pairs for each 8 rows
   //  2,   ; row-pair is 2 rows
@@ -439,7 +445,7 @@ class MicroGemm<cpu_utils::ISA::NEON, c10::BFloat16> {
     (void)lda;  // A is packed, so lda is not needed
     TORCH_CHECK_EQ(k % K, 0);
 
-    for (int32_t n_idx = 0; n_idx < NSize; n_idx += Nr_gemv) {
+    for (int32_t n_idx = 0; n_idx < NSize; n_idx += NrGemv) {
       const bfloat16_t* __restrict__ b_panel =
           reinterpret_cast<const bfloat16_t*>(b_ptr) + n_idx * k;
 

csrc/cpu/sgl-kernels/conv.cpp

@@ -451,6 +451,90 @@ void causal_conv1d_update_kernel_impl(
   });
 }
 
+template <typename scalar_t>
+void causal_conv1d_update_multi_kernel_impl(
+    scalar_t* __restrict__ out,
+    const scalar_t* __restrict__ input,
+    scalar_t* __restrict__ conv_states,
+    const scalar_t* __restrict__ weight,
+    const scalar_t* __restrict__ bias,
+    const int32_t* __restrict__ num_accepted_tokens,
+    const int32_t* __restrict__ conv_indices,
+    bool silu_activation,
+    int64_t batch,
+    int64_t dim,
+    int64_t seqlen,
+    int64_t width,
+    int64_t state_len,
+    int64_t conv_state_slot_stride) {
+  constexpr int64_t BLOCK_N = block_size_n() * 2;
+  const int64_t NB = div_up(dim, BLOCK_N);
+
+  AT_DISPATCH_BOOL2(bias != nullptr, has_bias, silu_activation, has_silu, [&] {
+    at::parallel_for(0, batch * NB, 0, [&](int64_t begin, int64_t end) {
+      int64_t bs{0}, nb{0};
+      data_index_init(begin, bs, batch, nb, NB);
+
+      for (int64_t i = begin; i < end; ++i) {
+        const int64_t nb_start = nb * BLOCK_N;
+        const int64_t nb_size = std::min(dim - nb_start, BLOCK_N);
+        const int32_t conv_state_index = conv_indices[bs];
+        const int32_t history_offset = num_accepted_tokens[bs] - 1;
+
+        switch (width << 4 | nb_size >> 4) {
+          case 0x42:
+            tinygemm_kernel<scalar_t, 4, 32, has_bias, has_silu>::apply(
+                input + bs * seqlen * dim + nb_start,
+                weight + nb_start * width,
+                out + bs * seqlen * dim + nb_start,
+                has_bias ? bias + nb_start : nullptr,
+                conv_states + conv_state_index * conv_state_slot_stride +
+                    history_offset * dim + nb_start,
+                true,
+                seqlen,
+                dim,
+                true);
+            break;
+          case 0x44:
+            tinygemm_kernel<scalar_t, 4, 64, has_bias, has_silu>::apply(
+                input + bs * seqlen * dim + nb_start,
+                weight + nb_start * width,
+                out + bs * seqlen * dim + nb_start,
+                has_bias ? bias + nb_start : nullptr,
+                conv_states + conv_state_index * conv_state_slot_stride +
+                    history_offset * dim + nb_start,
+                true,
+                seqlen,
+                dim,
+                true);
+            break;
+          default:
+            TORCH_CHECK(false, "Unexpected block size, ", width, " x ", nb_size);
+        }
+
+        data_index_step(bs, batch, nb, NB);
+      }
+    });
+  });
+
+  at::parallel_for(0, batch, 0, [&](int64_t begin, int64_t end) {
+    for (int64_t bs = begin; bs < end; ++bs) {
+      const int32_t conv_state_index = conv_indices[bs];
+      const int32_t num_accepted = num_accepted_tokens[bs];
+      scalar_t* state = conv_states + conv_state_index * conv_state_slot_stride;
+
+      std::memmove(
+          state,
+          state + num_accepted * dim,
+          (state_len - seqlen) * dim * sizeof(scalar_t));
+      std::memcpy(
+          state + (state_len - seqlen) * dim,
+          input + bs * seqlen * dim,
+          seqlen * dim * sizeof(scalar_t));
+    }
+  });
+}
+
 }  // anonymous namespace
 
 // from [dim, width] or [N, K]
@@ -545,7 +629,7 @@ at::Tensor get_block_indices(const std::optional<at::Tensor>& offsets, int64_t n
 //   query_start_loc: (batch + 1) int32
 //   cache_indices: (batch)  int32
 //   has_initial_state: (batch) bool
-//   conv_states: (..., dim, width - 1) itype
+//   conv_states: (..., dim, state_len) itype, where state_len >= width - 1
 //   activation: either None or "silu" or "swish"
 //   pad_slot_id: int
 //
@@ -586,11 +670,14 @@ at::Tensor causal_conv1d_fwd_cpu(
     CHECK_EQ(conv_states_val.scalar_type(), scalar_type);
     CHECK_GE(padded_batch, batch);
     CHECK_EQ(conv_states_val.size(1), dim);
-    CHECK_EQ(conv_states_val.size(2), width - 1);
+    const int64_t state_len = conv_states_val.size(2);
+    CHECK_GE(state_len, width - 1);
 
     // adjust `conv_states` to be contiguous on `dim`
     // should happen only once
     if (conv_states_val.stride(-2) != 1) {
+      TORCH_CHECK(state_len == width - 1,
+          "causal_conv1d_fwd_cpu: wide conv_states must be contiguous on dim.");
       auto conv_states_copy = conv_states_val.clone();
       conv_states_val.as_strided_({padded_batch, dim, width - 1}, {(width - 1) * dim, 1, dim});
       conv_states_val.copy_(conv_states_copy);
@@ -651,14 +738,14 @@ at::Tensor causal_conv1d_fwd_cpu(
 
 // API aligned with GPUs
 //
-//   x: (batch, dim) or (batch, dim, seqlen)
+//   x: (batch, dim) or (batch, seqlen, dim)
 //   conv_state: (..., dim, state_len), where state_len >= width - 1
 //   weight: (dim, width)
 //   bias: (dim,)
-//   cache_seqlens: (batch,), dtype int32.
+//   num_accepted_tokens: (batch,), dtype int32.
 //   conv_state_indices: (batch,), dtype int32
 //   pad_slot_id: int
-//   out: (batch, dim) or (batch, dim, seqlen)
+//   out: (batch, dim) or (batch, seqlen, dim)
 //
 at::Tensor causal_conv1d_update_cpu(
     const at::Tensor& x,
@@ -666,7 +753,7 @@ at::Tensor causal_conv1d_update_cpu(
     const at::Tensor& weight,
     const std::optional<at::Tensor>& bias,
     bool silu_activation,
-    const std::optional<at::Tensor>& cache_seqlens,
+    const std::optional<at::Tensor>& num_accepted_tokens,
     const std::optional<at::Tensor>& conv_state_indices,
     int64_t pad_slot_id,
     bool is_vnni) {
@@ -674,13 +761,13 @@ at::Tensor causal_conv1d_update_cpu(
   CHECK_CONTIGUOUS(weight);
   auto packed_w = is_vnni ? weight : causal_conv1d_weight_pack(weight);
 
-  // TODO: add multi-token prediction support
-  TORCH_CHECK(x.dim() == 2, "causal_conv1d_update_cpu: expect x to be 2D tensor.");
-  TORCH_CHECK(!cache_seqlens.has_value(), "causal_conv1d_update_cpu: don't support cache_seqlens.");
+  TORCH_CHECK(
+      x.dim() == 2 || x.dim() == 3,
+      "causal_conv1d_update_cpu: expect x to be 2D or 3D tensor.");
 
   int64_t batch = x.size(0);
-  int64_t dim = x.size(1);
-  int64_t seqlen = 1;
+  int64_t dim = x.dim() == 2 ? x.size(1) : x.size(2);
+  int64_t seqlen = x.dim() == 2 ? 1 : x.size(1);
   int64_t width = weight.size(-1);
 
   const auto scalar_type = x.scalar_type();
@@ -690,10 +777,84 @@ at::Tensor causal_conv1d_update_cpu(
 
   CHECK_EQ(conv_states.scalar_type(), scalar_type);
   CHECK_EQ(conv_states.size(1), dim);
-  CHECK_EQ(conv_states.size(2), width - 1);
+  const int64_t state_len = conv_states.size(2);
+  CHECK_GE(state_len, width - 1);
+
+  if (x.dim() == 3) {
+    TORCH_CHECK(
+        num_accepted_tokens.has_value(),
+        "causal_conv1d_update_cpu: num_accepted_tokens is required for 3D x.");
+    TORCH_CHECK(
+        conv_state_indices.has_value(),
+        "causal_conv1d_update_cpu: conv_state_indices is required for 3D x.");
+    CHECK_OPTIONAL_SHAPE_DTYPE(num_accepted_tokens, batch, at::kInt);
+    TORCH_CHECK(
+        width == 4,
+        "causal_conv1d_update_cpu: support only width of 4 for 3D x.");
+    TORCH_CHECK(
+        seqlen > 0,
+        "causal_conv1d_update_cpu: expect non-empty sequence for 3D x.");
+    TORCH_CHECK(
+        state_len >= seqlen,
+        "causal_conv1d_update_cpu: state_len must be >= seqlen for 3D x.");
+    TORCH_CHECK(
+        conv_states.stride(-2) == 1 && conv_states.stride(-1) == dim,
+        "causal_conv1d_update_cpu: 3D x requires SD conv_states layout.");
+
+    const int32_t* accepted_counts =
+        num_accepted_tokens.value().data_ptr<int32_t>();
+    const int32_t* indices = conv_state_indices.value().data_ptr<int32_t>();
+    const int64_t num_slots = conv_states.size(0);
+    for (int64_t bs = 0; bs < batch; ++bs) {
+      const int32_t num_accepted = accepted_counts[bs];
+      const int32_t conv_state_index = indices[bs];
+      TORCH_CHECK(
+          conv_state_index != pad_slot_id,
+          "causal_conv1d_update_cpu: 3D x does not support pad slots.");
+      TORCH_CHECK(
+          conv_state_index >= 0 && conv_state_index < num_slots,
+          "causal_conv1d_update_cpu: conv_state_indices out of range.");
+      TORCH_CHECK(
+          num_accepted >= 1 && num_accepted <= seqlen,
+          "causal_conv1d_update_cpu: num_accepted_tokens must be in [1, "
+          "seqlen].");
+      TORCH_CHECK(
+          num_accepted - 1 + width - 1 <= state_len,
+          "causal_conv1d_update_cpu: history window exceeds conv_states.");
+    }
+
+    int64_t conv_state_slot_stride = conv_states.stride(0);
+    at::Tensor out = at::empty_like(x);
+    AT_DISPATCH_REDUCED_FLOATING_TYPES(
+        scalar_type, "causal_conv1d_update_multi_kernel_impl", [&] {
+          causal_conv1d_update_multi_kernel_impl<scalar_t>(
+              out.data_ptr<scalar_t>(),
+              x.data_ptr<scalar_t>(),
+              conv_states.data_ptr<scalar_t>(),
+              packed_w.data_ptr<scalar_t>(),
+              conditional_data_ptr<scalar_t>(bias),
+              accepted_counts,
+              indices,
+              silu_activation,
+              batch,
+              dim,
+              seqlen,
+              width,
+              state_len,
+              conv_state_slot_stride);
+        });
+    return out;
+  }
+
+  TORCH_CHECK(
+      !num_accepted_tokens.has_value(),
+      "causal_conv1d_update_cpu: num_accepted_tokens is only supported for 3D "
+      "x.");
 
   // adjust `conv_states` to be contiguous on `dim`
   if (conv_states.stride(-2) != 1) {
+    TORCH_CHECK(state_len == width - 1,
+        "causal_conv1d_update_cpu: wide conv_states must be contiguous on dim.");
     int64_t num_cache_lines = conv_states.size(0);
     auto conv_states_copy = conv_states.clone();
     conv_states.as_strided_({num_cache_lines, dim, width - 1}, {(width - 1) * dim, 1, dim});

csrc/cpu/torch_bindings.cpp

@@ -147,7 +147,7 @@ at::Tensor causal_conv1d_fwd_cpu(
 at::Tensor causal_conv1d_update_cpu(
     const at::Tensor& x, const at::Tensor& conv_states,
     const at::Tensor& weight, const std::optional<at::Tensor>& bias,
-    bool silu_activation, const std::optional<at::Tensor>& cache_seqlens,
+    bool silu_activation, const std::optional<at::Tensor>& num_accepted_tokens,
     const std::optional<at::Tensor>& conv_state_indices, int64_t pad_slot_id,
     bool is_vnni);
 
@@ -207,6 +207,20 @@ void cpu_fused_moe(torch::Tensor& output, const torch::Tensor& input,
                    const torch::Tensor& topk_id, const bool skip_weighted,
                    const std::string& act, const std::string& isa);
 
+void prepack_moe_weight_int8(const torch::Tensor& weight,
+                             torch::Tensor& packed_weight,
+                             const std::string& isa);
+
+void cpu_fused_moe_int8(torch::Tensor& output, const torch::Tensor& input,
+                        const torch::Tensor& w13, const torch::Tensor& w2,
+                        const torch::Tensor& w13_scale,
+                        const torch::Tensor& w2_scale,
+                        const std::optional<torch::Tensor>& w13_bias,
+                        const std::optional<torch::Tensor>& w2_bias,
+                        const torch::Tensor& topk_weights,
+                        const torch::Tensor& topk_id, const bool skip_weighted,
+                        const std::string& act, const std::string& isa);
+
 void compute_slot_mapping_kernel_impl(const torch::Tensor query_start_loc,
                                       const torch::Tensor positions,
                                       const torch::Tensor block_table,
@@ -502,7 +516,8 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
   ops.def(
       "causal_conv1d_update_cpu(Tensor x, Tensor(a!) conv_states, Tensor "
       "weight, Tensor? bias, bool silu_activation,"
-      "Tensor? cache_seqlens, Tensor? conv_state_indices, int pad_slot_id, "
+      "Tensor? num_accepted_tokens, Tensor? conv_state_indices, int "
+      "pad_slot_id, "
       "bool is_vnni) -> Tensor");
   ops.impl("causal_conv1d_update_cpu", torch::kCPU, &causal_conv1d_update_cpu);
 #endif
@@ -596,8 +611,7 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
 #endif
 
   // fused moe
-#if defined(__AVX512F__) || \
-    (defined(__aarch64__) && !defined(__APPLE__) && defined(ARM_BF16_SUPPORT))
+#if defined(__AVX512F__) || (defined(ARM_BF16_SUPPORT) && !defined(__APPLE__))
   ops.def(
       "prepack_moe_weight(Tensor weight, Tensor(a1!) packed_weight, str isa) "
       "-> ()");
@@ -608,7 +622,22 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
       "bool skip_weighted, "
       "str act, str isa) -> ()");
   ops.impl("cpu_fused_moe", torch::kCPU, &cpu_fused_moe);
-#endif
+#endif  // #if defined(__AVX512F__) || (defined(ARM_BF16_SUPPORT) &&
+        // !defined(__APPLE__))
+#if defined(ARM_I8MM_SUPPORT) && defined(ARM_BF16_SUPPORT) && \
+    !defined(__APPLE__)
+  ops.def(
+      "prepack_moe_weight_int8(Tensor weight, Tensor(a1!) packed_weight, "
+      "str isa) -> ()");
+  ops.impl("prepack_moe_weight_int8", torch::kCPU, &prepack_moe_weight_int8);
+  ops.def(
+      "cpu_fused_moe_int8(Tensor(a0!) output, Tensor input, Tensor w13, "
+      "Tensor w2, Tensor w13_scale, Tensor w2_scale, Tensor? w13_bias, "
+      "Tensor? w2_bias, Tensor topk_weights, Tensor topk_id, bool "
+      "skip_weighted, str act, str isa) -> ()");
+  ops.impl("cpu_fused_moe_int8", torch::kCPU, &cpu_fused_moe_int8);
+#endif  // #if defined(ARM_I8MM_SUPPORT) && defined(ARM_BF16_SUPPORT) &&
+        // !defined(__APPLE__)
   ops.def(
       "mla_decode_kvcache("
       "   Tensor! out, Tensor query, Tensor kv_cache,"

csrc/fs_io.cpp

@@ -3,19 +3,155 @@
 
 #include <Python.h>
 
+#include <errno.h>
+#include <fcntl.h>
 #include <unistd.h>
 
+#include <filesystem>
+#include <string>
 #include <vector>
 
+#if defined(O_DIRECT)
+constexpr int kODirectFlag = O_DIRECT;
+#else
+constexpr int kODirectFlag = 0;
+#endif
+
 extern "C" {
 
-static void _batch_lookup(const std::vector<const char*>& paths,
+namespace {
+
+// Returns 0 on success, or the std::error_code's POSIX-compatible value on
+// failure, mirroring the errno convention used by the syscalls below.
+inline int ensure_parent_dirs(const std::string& path) {
+  const auto parent = std::filesystem::path(path).parent_path();
+  if (parent.empty()) {
+    return 0;
+  }
+  std::error_code ec;
+  std::filesystem::create_directories(parent, ec);
+  return ec ? ec.value() : 0;
+}
+
+// Core single-block store: src/size are raw pointer + byte count. Returns 0
+// on success, or the errno of the failing step on failure -- captured
+// before any subsequent cleanup call can overwrite it. On failure, the temp
+// file is removed.
+inline int _store_block(const char* tmp_path, const char* dest_path,
+                        const char* src, size_t size, bool use_o_direct) {
+  if (access(dest_path, F_OK) == 0) {
+    return 0;  // Already present.
+  }
+
+  if (const int err = ensure_parent_dirs(dest_path); err != 0) {
+    return err;
+  }
+
+  const int o_direct_flag = use_o_direct ? kODirectFlag : 0;
+  const int fd = open(
+      tmp_path, O_CREAT | O_EXCL | O_WRONLY | O_TRUNC | o_direct_flag, 0644);
+  if (fd < 0) {
+    return errno;
+  }
+
+  const ssize_t written = write(fd, src, size);
+  if (written < 0 || static_cast<size_t>(written) != size) {
+    const int err = written < 0 ? errno : EIO;
+    close(fd);  // Best-effort cleanup; the real error is already captured.
+    unlink(tmp_path);
+    return err;
+  }
+
+  if (close(fd) != 0) {
+    const int err = errno;
+    unlink(tmp_path);
+    return err;
+  }
+
+  if (rename(tmp_path, dest_path) != 0) {
+    const int err = errno;
+    unlink(tmp_path);
+    return err;
+  }
+
+  return 0;
+}
+
+// Core single-block load: dst/size are raw pointer + byte count. Returns 0
+// on success, or the errno of the failing step on failure. On failure,
+// the source file is removed since a partially-read block should not be reused.
+inline int _load_block(const char* source_path, char* dst, size_t size,
+                       bool use_o_direct) {
+  const int o_direct_flag = use_o_direct ? kODirectFlag : 0;
+  const int fd = open(source_path, O_RDONLY | o_direct_flag, 0);
+  if (fd < 0) {
+    const int err = errno;
+    unlink(source_path);
+    return err;
+  }
+
+  const ssize_t bytes_read = read(fd, dst, size);
+  if (bytes_read < 0 || static_cast<size_t>(bytes_read) != size) {
+    const int err = bytes_read < 0 ? errno : EIO;
+    close(fd);
+    unlink(source_path);
+    return err;
+  }
+
+  if (close(fd) != 0) {
+    const int err = errno;
+    unlink(source_path);
+    return err;
+  }
+
+  return 0;
+}
+
+inline void _batch_lookup(const std::vector<const char*>& paths,
                           std::vector<int>& exists_flags) {
   for (size_t i = 0; i < paths.size(); i++) {
     exists_flags[i] = (access(paths[i], F_OK) == 0) ? 1 : 0;
   }
 }
 
+// Helper: extract a list[str] of length n into a vector<const char*>.
+// Returns false and sets a Python exception on error.
+inline bool extract_str_list(PyObject* list, Py_ssize_t n,
+                             std::vector<const char*>& out) {
+  for (Py_ssize_t i = 0; i < n; i++) {
+    out[i] = PyUnicode_AsUTF8AndSize(PyList_GetItem(list, i), nullptr);
+    if (out[i] == nullptr) {
+      return false;
+    }
+  }
+  return true;
+}
+
+// Helper: extract a Py_buffer per element of a list[bytes-like] of length n.
+// On success, `out` holds n acquired buffers (caller must PyBuffer_Release
+// each). On failure, any buffers already acquired are released before
+// returning false, and a Python exception is set.
+inline bool extract_buffer_list(PyObject* list, Py_ssize_t n, int flags,
+                                std::vector<Py_buffer>& out) {
+  for (Py_ssize_t i = 0; i < n; i++) {
+    if (PyObject_GetBuffer(PyList_GetItem(list, i), &out[i], flags) != 0) {
+      for (Py_ssize_t j = 0; j < i; j++) {
+        PyBuffer_Release(&out[j]);
+      }
+      return false;
+    }
+  }
+  return true;
+}
+
+inline void release_buffer_list(std::vector<Py_buffer>& buffers) {
+  for (auto& buf : buffers) {
+    PyBuffer_Release(&buf);
+  }
+}
+
+}  // namespace
+
 /// @brief Check file existence for a batch of paths.
 /// @param paths list[str] – absolute paths to check.
 /// @return list[bool] – True if the corresponding path exists, False otherwise.
@@ -51,11 +187,157 @@ static PyObject* batch_lookup(PyObject* /*self*/, PyObject* args) {
   return result;
 }
 
+/// @brief Store a batch of blocks, each from its own buffer, to disk.
+/// @param tmp_paths    list[str] – one temp path per block.
+/// @param dest_paths   list[str] – one destination path per block.
+/// @param buffers      list[bytes-like] – one source buffer per block.
+/// @param use_o_direct bool – whether to open files with O_DIRECT
+///                     (default True). Ignored where O_DIRECT is unsupported
+///                     by the platform.
+/// @note Releases the GIL for the entire batch. Raises on first error.
+static PyObject* batch_store_block(PyObject* /*self*/, PyObject* args) {
+  PyObject* tmp_paths_obj = nullptr;
+  PyObject* dest_paths_obj = nullptr;
+  PyObject* buffers_obj = nullptr;
+  int use_o_direct = 1;
+
+  if (!PyArg_ParseTuple(args, "O!O!O!|p", &PyList_Type, &tmp_paths_obj,
+                        &PyList_Type, &dest_paths_obj, &PyList_Type,
+                        &buffers_obj, &use_o_direct)) {
+    return nullptr;
+  }
+
+  const Py_ssize_t n = PyList_Size(tmp_paths_obj);
+  if (PyList_Size(dest_paths_obj) != n || PyList_Size(buffers_obj) != n) {
+    PyErr_SetString(
+        PyExc_ValueError,
+        "tmp_paths, dest_paths and buffers must have the same length");
+    return nullptr;
+  }
+
+  std::vector<const char*> tmp_paths(n);
+  std::vector<const char*> dest_paths(n);
+
+  if (!extract_str_list(tmp_paths_obj, n, tmp_paths)) return nullptr;
+  if (!extract_str_list(dest_paths_obj, n, dest_paths)) return nullptr;
+
+  std::vector<Py_buffer> buffers(n);
+  if (!extract_buffer_list(buffers_obj, n, PyBUF_SIMPLE, buffers)) {
+    return nullptr;
+  }
+
+  Py_ssize_t failed_index = -1;
+  int failure_errno = 0;
+
+  {
+    Py_BEGIN_ALLOW_THREADS for (Py_ssize_t i = 0; i < n; i++) {
+      const char* buf = static_cast<const char*>(buffers[i].buf);
+      const int err =
+          _store_block(tmp_paths[i], dest_paths[i], buf,
+                       static_cast<size_t>(buffers[i].len), use_o_direct);
+      if (err != 0) {
+        failed_index = i;
+        failure_errno = err;
+        break;
+      }
+    }
+    Py_END_ALLOW_THREADS
+  }
+
+  release_buffer_list(buffers);
+
+  if (failed_index >= 0) {
+    // PyErr_SetFromErrnoWithFilename() reads the errno to format exception.
+    errno = failure_errno;
+    return PyErr_SetFromErrnoWithFilename(PyExc_OSError,
+                                          dest_paths[failed_index]);
+  }
+
+  Py_RETURN_NONE;
+}
+
+/// @brief Load a batch of blocks from disk, each into its own buffer.
+/// @param source_paths list[str] – one source path per block.
+/// @param buffers      list[writable bytes-like] – one destination buffer
+///                     per block.
+/// @param use_o_direct bool – whether to open files with O_DIRECT
+///                     (default True). Ignored where O_DIRECT is unsupported
+///                     by the platform.
+/// @note Releases the GIL for the entire batch. Raises on first error.
+static PyObject* batch_load_block(PyObject* /*self*/, PyObject* args) {
+  PyObject* source_paths_obj = nullptr;
+  PyObject* buffers_obj = nullptr;
+  int use_o_direct = 1;
+
+  if (!PyArg_ParseTuple(args, "O!O!|p", &PyList_Type, &source_paths_obj,
+                        &PyList_Type, &buffers_obj, &use_o_direct)) {
+    return nullptr;
+  }
+
+  const Py_ssize_t n = PyList_Size(source_paths_obj);
+  if (PyList_Size(buffers_obj) != n) {
+    PyErr_SetString(PyExc_ValueError,
+                    "source_paths and buffers must have the same length");
+    return nullptr;
+  }
+
+  std::vector<const char*> source_paths(n);
+  if (!extract_str_list(source_paths_obj, n, source_paths)) return nullptr;
+
+  std::vector<Py_buffer> buffers(n);
+  if (!extract_buffer_list(buffers_obj, n, PyBUF_WRITABLE, buffers)) {
+    return nullptr;
+  }
+
+  Py_ssize_t failed_index = -1;
+  int failure_errno = 0;
+
+  {
+    Py_BEGIN_ALLOW_THREADS for (Py_ssize_t i = 0; i < n; i++) {
+      char* buf = static_cast<char*>(buffers[i].buf);
+      const int err =
+          _load_block(source_paths[i], buf, static_cast<size_t>(buffers[i].len),
+                      use_o_direct);
+      if (err != 0) {
+        failed_index = i;
+        failure_errno = err;
+        break;
+      }
+    }
+    Py_END_ALLOW_THREADS
+  }
+
+  release_buffer_list(buffers);
+
+  if (failed_index >= 0) {
+    // PyErr_SetFromErrnoWithFilename() reads the errno to format exception.
+    errno = failure_errno;
+    return PyErr_SetFromErrnoWithFilename(PyExc_OSError,
+                                          source_paths[failed_index]);
+  }
+
+  Py_RETURN_NONE;
+}
+
 static PyMethodDef fs_io_C_methods[] = {
     {"batch_lookup", batch_lookup, METH_VARARGS,
      "batch_lookup(paths: list[str]) -> list[bool]\n"
      "\n"
      "Check file existence for a batch of paths."},
+    {"batch_store_block", batch_store_block, METH_VARARGS,
+     "batch_store_block(tmp_paths: list[str], dest_paths: list[str],\n"
+     "                  buffers: list[bytes-like],\n"
+     "                  use_o_direct: bool = True) -> None\n"
+     "\n"
+     "Store a batch of blocks, each from its own buffer, to disk. Raises on "
+     "first error."},
+    {"batch_load_block", batch_load_block, METH_VARARGS,
+     "batch_load_block(source_paths: list[str],\n"
+     "                 buffers: list[writable bytes-like],\n"
+     "                 use_o_direct: bool = True) -> None\n"
+     "\n"
+     "Load a batch of blocks from disk into corresponding buffers. "
+     "Raises on first error."},
     {nullptr, nullptr, 0, nullptr},
 };
 

csrc/libtorch_stable/kimi_k3/attn_res_kernel.cu

@@ -0,0 +1,954 @@
+/*
+ * Copyright (c) 2026, NVIDIA CORPORATION. All rights reserved.
+ */
+
+// Production AttnRes forward for Blackwell (SM100).
+//
+// Warp-specialized online softmax + residual + RMSNorm:
+//   - 1 producer warp issues cp.async.bulk row loads into shared memory.
+//   - 8 consumer warps compute reductions and output.
+//   - Q=res_weight*rms_weight remains in registers across persistent tokens.
+//   - V rows are converted once and cached as FP32 in TMEM between passes.
+//
+// Integration contract: Kimi K3 H=7168, 1<=num_blocks<=8, and token-major
+// block residual storage.
+
+#include "../torch_utils.h"
+
+#include <cfloat>
+#include <cstdint>
+#include <cstdio>
+#include <cuda_runtime.h>
+#include <type_traits>
+
+using bf16_t = __nv_bfloat16;
+
+namespace sm100 {
+namespace fwd_prod_v2 {
+
+constexpr int K_TILE = 1024;
+constexpr int N_CHUNK_DEFAULT = 4;
+constexpr int CHUNK_DEPTH = 2;
+constexpr int BLK = 288;  // 1 producer warp + 8 consumer warps
+constexpr int CONSUMER_THREADS = BLK - 32;  // 256
+constexpr int CONSUMER_WARPS = CONSUMER_THREADS / 32;
+constexpr int CONSUMER_GROUPS = 2;  // two 128-thread consumer groups
+constexpr int CONSUMER_THREADS_PER_GROUP = CONSUMER_THREADS / CONSUMER_GROUPS;
+constexpr int FIRST_USER_NAMED_BARRIER = 8;
+
+__device__ __forceinline__ const bf16_t* residual_addr(
+    const bf16_t* block_res, const bf16_t* layer_res, int source, int N,
+    int token, int block_stride_m, int block_stride_r, int H) {
+  if (source < N - 1) {
+    return block_res + static_cast<long long>(token) * block_stride_m +
+           source * block_stride_r;
+  }
+  return layer_res + static_cast<long long>(token) * H;
+}
+
+__device__ __forceinline__ uint32_t elect_one_sync() {
+  uint32_t pred = 0;
+  uint32_t laneid = 0;
+  asm volatile(
+      "{\n"
+      ".reg .b32 %%rx;\n"
+      ".reg .pred %%px;\n"
+      "     elect.sync %%rx|%%px, %2;\n"
+      "@%%px mov.s32 %1, 1;\n"
+      "     mov.s32 %0, %%rx;\n"
+      "}\n"
+      : "+r"(laneid), "+r"(pred)
+      : "r"(0xffffffff));
+  return pred;
+}
+
+__device__ __forceinline__ void mbarrier_init(uint64_t& barrier,
+                                              int thread_count) {
+  uint32_t const barrier_addr =
+      static_cast<uint32_t>(__cvta_generic_to_shared(&barrier));
+  asm volatile("mbarrier.init.shared::cta.b64 [%0], %1;\n" ::"r"(barrier_addr),
+               "r"(thread_count));
+}
+
+__device__ __forceinline__ void mbarrier_expect_tx(uint64_t& barrier,
+                                                   uint32_t bytes) {
+  uint32_t const barrier_addr =
+      static_cast<uint32_t>(__cvta_generic_to_shared(&barrier));
+  asm volatile("mbarrier.arrive.expect_tx.shared::cta.b64 _, [%0], %1;\n" ::"r"(
+                   barrier_addr),
+               "r"(bytes));
+}
+
+__device__ __forceinline__ void mbarrier_wait(uint64_t& barrier, int phase) {
+  uint32_t const barrier_addr =
+      static_cast<uint32_t>(__cvta_generic_to_shared(&barrier));
+  asm volatile(
+      "{\n"
+      ".reg .pred p;\n"
+      "WAIT:\n"
+      "mbarrier.try_wait.parity.shared::cta.b64 p, [%0], %1;\n"
+      "@p bra DONE;\n"
+      "bra WAIT;\n"
+      "DONE:\n"
+      "}\n" ::"r"(barrier_addr),
+      "r"(phase));
+}
+
+__device__ __forceinline__ void mbarrier_arrive(uint64_t& barrier) {
+  uint32_t const barrier_addr =
+      static_cast<uint32_t>(__cvta_generic_to_shared(&barrier));
+  asm volatile(
+      "{\n"
+      ".reg .b64 state;\n"
+      "mbarrier.arrive.shared::cta.b64 state, [%0];\n"
+      "}\n" ::"r"(barrier_addr));
+}
+
+__device__ __forceinline__ void fence_mbarrier_init() {
+  asm volatile("fence.mbarrier_init.release.cluster;" ::: "memory");
+}
+
+__device__ __forceinline__ void named_barrier_sync(uint32_t num_threads,
+                                                   uint32_t user_barrier_id) {
+  asm volatile(
+      "bar.sync %0, %1;" ::"r"(user_barrier_id + FIRST_USER_NAMED_BARRIER),
+      "r"(num_threads)
+      : "memory");
+}
+
+__device__ __forceinline__ void tmem_allocate(int num_columns, uint32_t* dst) {
+  uint32_t const dst_addr =
+      static_cast<uint32_t>(__cvta_generic_to_shared(dst));
+  asm volatile(
+      "tcgen05.alloc.cta_group::1.sync.aligned.shared::cta.b32 [%0], %1;" ::"r"(
+          dst_addr),
+      "r"(num_columns));
+}
+
+__device__ __forceinline__ void tmem_free(uint32_t tmem_ptr, int num_columns) {
+  asm volatile(
+      "tcgen05.dealloc.cta_group::1.sync.aligned.b32 %0, %1;" ::"r"(tmem_ptr),
+      "r"(num_columns));
+}
+
+__device__ __forceinline__ void tmem_release_allocation_lock() {
+  asm volatile("tcgen05.relinquish_alloc_permit.cta_group::1.sync.aligned;");
+}
+
+__device__ __forceinline__ void tmem_store_wait() {
+  asm volatile("tcgen05.wait::st.sync.aligned;" ::: "memory");
+}
+
+template <int N, typename T>
+__device__ __forceinline__ void tmem_load(uint32_t src_addr, T* dst) {
+  uint32_t* values = reinterpret_cast<uint32_t*>(dst);
+  if constexpr (N == 8) {
+    asm volatile(
+        "tcgen05.ld.sync.aligned.32x32b.x8.b32"
+        "{%0, %1, %2, %3, %4, %5, %6, %7}, [%8];\n"
+        : "=r"(values[0]), "=r"(values[1]), "=r"(values[2]), "=r"(values[3]),
+          "=r"(values[4]), "=r"(values[5]), "=r"(values[6]), "=r"(values[7])
+        : "r"(src_addr));
+  } else {
+    static_assert(N == 4, "AttnRes TMEM helpers support x4 and x8");
+    asm volatile(
+        "tcgen05.ld.sync.aligned.32x32b.x4.b32"
+        "{%0, %1, %2, %3}, [%4];\n"
+        : "=r"(values[0]), "=r"(values[1]), "=r"(values[2]), "=r"(values[3])
+        : "r"(src_addr));
+  }
+}
+
+template <int N, typename T>
+__device__ __forceinline__ void tmem_store(uint32_t dst_addr, T* src) {
+  uint32_t* values = reinterpret_cast<uint32_t*>(src);
+  if constexpr (N == 8) {
+    asm volatile(
+        "tcgen05.st.sync.aligned.32x32b.x8.b32"
+        "[%8], {%0, %1, %2, %3, %4, %5, %6, %7};\n" ::"r"(values[0]),
+        "r"(values[1]), "r"(values[2]), "r"(values[3]), "r"(values[4]),
+        "r"(values[5]), "r"(values[6]), "r"(values[7]), "r"(dst_addr));
+  } else {
+    static_assert(N == 4, "AttnRes TMEM helpers support x4 and x8");
+    asm volatile(
+        "tcgen05.st.sync.aligned.32x32b.x4.b32"
+        "[%4], {%0, %1, %2, %3};\n" ::"r"(values[0]),
+        "r"(values[1]), "r"(values[2]), "r"(values[3]), "r"(dst_addr));
+  }
+}
+
+__device__ __forceinline__ float2 float2_add(const float2& a, const float2& b) {
+  float2 result;
+  asm volatile("add.rn.f32x2 %0, %1, %2;\n"
+               : "=l"(reinterpret_cast<uint64_t&>(result))
+               : "l"(reinterpret_cast<uint64_t const&>(a)),
+                 "l"(reinterpret_cast<uint64_t const&>(b)));
+  return result;
+}
+
+__device__ __forceinline__ float2 float2_mul(const float2& a, const float2& b) {
+  float2 result;
+  asm volatile("mul.f32x2 %0, %1, %2;\n"
+               : "=l"(reinterpret_cast<uint64_t&>(result))
+               : "l"(reinterpret_cast<uint64_t const&>(a)),
+                 "l"(reinterpret_cast<uint64_t const&>(b)));
+  return result;
+}
+
+__device__ __forceinline__ float2 float2_fma(const float2& a, const float2& b,
+                                             const float2& c) {
+  float2 result;
+  asm volatile("fma.rn.f32x2 %0, %1, %2, %3;\n"
+               : "=l"(reinterpret_cast<uint64_t&>(result))
+               : "l"(reinterpret_cast<uint64_t const&>(a)),
+                 "l"(reinterpret_cast<uint64_t const&>(b)),
+                 "l"(reinterpret_cast<uint64_t const&>(c)));
+  return result;
+}
+
+template <int NC>
+struct FwdSmemPlan {
+  alignas(16) uint64_t bar_ready[CHUNK_DEPTH];
+  alignas(16) uint64_t bar_consumed[CHUNK_DEPTH];
+  alignas(16) uint64_t bar_output_norm_ready;
+  alignas(16) float2 ws_stats[CONSUMER_WARPS][NC];
+  uint32_t tmem_base;
+};
+
+__device__ __forceinline__ void cp_async_bulk(void* smem_dst,
+                                              const void* gmem_src, int bytes,
+                                              uint64_t& mbar) {
+  uint32_t const s = static_cast<uint32_t>(__cvta_generic_to_shared(smem_dst));
+  uint32_t const m = static_cast<uint32_t>(__cvta_generic_to_shared(&mbar));
+  asm volatile(
+      "cp.async.bulk.shared::cta.global.mbarrier::complete_tx::bytes [%0], "
+      "[%1], %2, [%3];\n" ::"r"(s),
+      "l"(gmem_src), "r"(bytes), "r"(m)
+      : "memory");
+}
+
+template <int H, int NC = N_CHUNK_DEFAULT, bool RELEASE_TMEM = false,
+          bool HAS_DELTA = false, bool HAS_OUTPUT_NORM = false,
+          bool OUTPUT_NORM_IN_SMEM = false>
+__global__ void __launch_bounds__(BLK, 1) attn_res_fwd_online_v2_kernel(
+    const bf16_t* __restrict__ block_res, bf16_t* __restrict__ layer_res,
+    const bf16_t* __restrict__ delta, const bf16_t* __restrict__ res_w,
+    const bf16_t* __restrict__ rms_w, bf16_t* __restrict__ output, int N, int T,
+    int B, int block_stride_m, int block_stride_r, float rms_eps,
+    const bf16_t* __restrict__ output_norm_weight, float output_norm_eps) {
+#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 1000 && __CUDA_ARCH__ < 1100
+  constexpr float LOG2_E = 1.4426950408889634f;
+  constexpr int N_CHUNK = NC;
+  // The two-source specialization only consumes half of the TMEM columns.
+  constexpr int TMEM_COLS_ALLOC = NC == 2 ? 128 : 256;
+  constexpr int NUM_BUFS = CHUNK_DEPTH * NC;
+  constexpr int NHT = H / K_TILE;
+  constexpr int SLICES_PER_GROUP =
+      (NHT + CONSUMER_GROUPS - 1) / CONSUMER_GROUPS;
+  constexpr int VEC = 8;
+  constexpr int ACC_PER_THREAD = H == 7168 ? 28 : SLICES_PER_GROUP * VEC;
+  constexpr int TMEM_V_COLS_PER_GROUP = SLICES_PER_GROUP * N_CHUNK * VEC;
+  constexpr int TMEM_V_COLS_TOTAL = CONSUMER_GROUPS * TMEM_V_COLS_PER_GROUP;
+  static_assert(TMEM_V_COLS_TOTAL <= TMEM_COLS_ALLOC);
+  static_assert(H >= 4096 && H <= 8192);
+  static_assert(H % K_TILE == 0);
+
+  const int tid = threadIdx.x;
+  const int wid = tid >> 5;
+  const int lane = tid & 31;
+  const int TB = T * B;
+  const int num_ctas = gridDim.x;
+  const int num_chunks = (N + N_CHUNK - 1) / N_CHUNK;
+
+  const int comp_wid = wid - 1;
+  const int comp_tid = tid - 32;
+  const int group = (comp_wid >= 4) ? 1 : 0;
+  const int ct_in_group =
+      (comp_tid >= 0) ? (comp_tid & (CONSUMER_THREADS_PER_GROUP - 1)) : -1;
+  const int k_local = ct_in_group * VEC;
+
+  constexpr size_t V_BYTES = (size_t)NUM_BUFS * H * sizeof(bf16_t);
+  constexpr size_t DELTA_BYTES =
+      HAS_DELTA ? (size_t)CHUNK_DEPTH * H * sizeof(bf16_t) : 0;
+  constexpr size_t OUTPUT_NORM_BYTES =
+      OUTPUT_NORM_IN_SMEM ? (size_t)H * sizeof(bf16_t) : 0;
+  extern __shared__ __align__(16) char smem_raw[];
+  bf16_t* v_bufs = reinterpret_cast<bf16_t*>(smem_raw);  // [NUM_BUFS][H]
+  bf16_t* delta_bufs = reinterpret_cast<bf16_t*>(smem_raw + V_BYTES);
+  bf16_t* output_norm_buf =
+      reinterpret_cast<bf16_t*>(smem_raw + V_BYTES + DELTA_BYTES);
+  FwdSmemPlan<NC>& plan = *reinterpret_cast<FwdSmemPlan<NC>*>(
+      smem_raw + V_BYTES + DELTA_BYTES + OUTPUT_NORM_BYTES);
+
+  auto slot_of = [](long long gci, int n) {
+    return (int)(gci % CHUNK_DEPTH) * N_CHUNK + n;
+  };
+  auto phase_of = [](long long gci) { return (int)((gci / CHUNK_DEPTH) & 1); };
+  auto buf_ptr = [&](int slot) -> bf16_t* { return v_bufs + slot * H; };
+  auto delta_buf_ptr = [&](int chunk_slot) -> bf16_t* {
+    return delta_bufs + chunk_slot * H;
+  };
+
+  if (wid == 0 && elect_one_sync()) {
+  #pragma unroll
+    for (int i = 0; i < CHUNK_DEPTH; i++) {
+      mbarrier_init(plan.bar_ready[i], 1);
+      mbarrier_init(plan.bar_consumed[i], CONSUMER_WARPS);
+    }
+    if constexpr (OUTPUT_NORM_IN_SMEM) {
+      mbarrier_init(plan.bar_output_norm_ready, 1);
+    }
+    fence_mbarrier_init();
+  }
+
+  // gdc wait BEFORE tmem alloc
+  cudaGridDependencySynchronize();
+
+  if (wid == 1) {
+    tmem_allocate(TMEM_COLS_ALLOC, &plan.tmem_base);
+    if constexpr (RELEASE_TMEM) {
+      tmem_release_allocation_lock();
+    }
+  }
+  __syncthreads();
+
+  if constexpr (OUTPUT_NORM_IN_SMEM) {
+    if (wid == 0 && elect_one_sync()) {
+      mbarrier_expect_tx(plan.bar_output_norm_ready, H * (int)sizeof(bf16_t));
+      cp_async_bulk(output_norm_buf, output_norm_weight, H * sizeof(bf16_t),
+                    plan.bar_output_norm_ready);
+    }
+  }
+
+  const uint32_t my_v_tmem =
+      comp_tid >= 0 ? plan.tmem_base + group * TMEM_V_COLS_PER_GROUP : 0;
+  float q_cache[ACC_PER_THREAD];
+  if (comp_tid >= 0) {
+  #pragma unroll
+    for (int si = 0; si < SLICES_PER_GROUP; si++) {
+      if constexpr (H == 7168) {
+        if (si == SLICES_PER_GROUP - 1) {
+          int h_base = 6 * K_TILE + group * (K_TILE / 2) + ct_in_group * 4;
+  #pragma unroll
+          for (int j = 0; j < 4; j++) {
+            int h = h_base + j;
+            q_cache[si * VEC + j] =
+                __bfloat162float(rms_w[h]) * __bfloat162float(res_w[h]);
+          }
+          continue;
+        }
+      }
+      int dt = si * CONSUMER_GROUPS + group;
+      if (dt >= NHT) continue;
+      int h_base = dt * K_TILE + k_local;
+  #pragma unroll
+      for (int j = 0; j < VEC; j++) {
+        int h = h_base + j;
+        q_cache[si * VEC + j] =
+            __bfloat162float(rms_w[h]) * __bfloat162float(res_w[h]);
+      }
+    }
+  }
+
+  if (wid == 0) {
+    if (elect_one_sync()) {
+      long long gci = 0;
+      for (int tb = blockIdx.x; tb < TB; tb += num_ctas) {
+        const int t = tb / B;
+        for (int ci = 0; ci < num_chunks; ci++, gci++) {
+          int ns = ci * N_CHUNK;
+          int an = min(N_CHUNK, N - ns);
+          int chunk_slot = (int)(gci % CHUNK_DEPTH);
+          int pc = phase_of(gci);
+          mbarrier_wait(plan.bar_consumed[chunk_slot], pc ^ 1);
+          int transaction_bytes = an * H * (int)sizeof(bf16_t);
+          if constexpr (HAS_DELTA) {
+            int prefix_n = N - 1 - ns;
+            if (prefix_n >= 0 && prefix_n < an) {
+              transaction_bytes += H * (int)sizeof(bf16_t);
+            }
+          }
+          mbarrier_expect_tx(plan.bar_ready[chunk_slot], transaction_bytes);
+  #pragma unroll
+          for (int n = 0; n < N_CHUNK; n++) {
+            if (n >= an) continue;
+            int slot = slot_of(gci, n);
+            const bf16_t* src =
+                residual_addr(block_res, layer_res, ns + n, N, t,
+                              block_stride_m, block_stride_r, H);
+            cp_async_bulk(buf_ptr(slot), src, H * sizeof(bf16_t),
+                          plan.bar_ready[chunk_slot]);
+          }
+          if constexpr (HAS_DELTA) {
+            int prefix_n = N - 1 - ns;
+            if (prefix_n >= 0 && prefix_n < an) {
+              cp_async_bulk(delta_buf_ptr(chunk_slot),
+                            delta + (long long)tb * H, H * sizeof(bf16_t),
+                            plan.bar_ready[chunk_slot]);
+            }
+          }
+        }
+      }
+    }
+  } else {
+    float acc32[ACC_PER_THREAD] = {};
+    float eps_cache;
+    asm volatile("mov.b32 %0, %1;" : "=f"(eps_cache) : "f"(rms_eps));
+
+    long long gci = 0;
+    for (int tb = blockIdx.x; tb < TB; tb += num_ctas) {
+      float m_running = -FLT_MAX;
+      float s_running = 0.f;
+  #pragma unroll
+      for (int i = 0; i < ACC_PER_THREAD; i++) {
+        acc32[i] = 0.f;
+      }
+
+      for (int ci = 0; ci < num_chunks; ci++, gci++) {
+        int ns = ci * N_CHUNK;
+        int an = min(N_CHUNK, N - ns);
+        int chunk_slot = (int)(gci % CHUNK_DEPTH);
+        int pr = phase_of(gci);
+        mbarrier_wait(plan.bar_ready[chunk_slot], pr);
+
+        float2 sq_local[N_CHUNK] = {};
+        float2 dot_local[N_CHUNK] = {};
+
+        auto pass_A_body = [&](auto AN_TOK) {
+          constexpr int AN = decltype(AN_TOK)::value;
+  #pragma unroll
+          for (int si = 0; si < SLICES_PER_GROUP; si++) {
+            if constexpr (H == 7168) {
+              if (si == SLICES_PER_GROUP - 1) {
+                int h_base =
+                    6 * K_TILE + group * (K_TILE / 2) + ct_in_group * 4;
+                const float* qv = &q_cache[si * VEC];
+  #pragma unroll
+                for (int n = 0; n < AN; n++) {
+                  int slot = slot_of(gci, n);
+                  int2 vp =
+                      *reinterpret_cast<const int2*>(buf_ptr(slot) + h_base);
+                  auto* v2 = reinterpret_cast<__nv_bfloat162*>(&vp);
+                  if constexpr (HAS_DELTA) {
+                    int prefix_n = N - 1 - ns;
+                    if (n == prefix_n) {
+                      const bf16_t* delta_ptr =
+                          delta_buf_ptr(chunk_slot) + h_base;
+  #pragma unroll
+                      for (int j = 0; j < 2; j++) {
+                        auto delta2 = *reinterpret_cast<const __nv_bfloat162*>(
+                            delta_ptr + 2 * j);
+                        v2[j] = __hadd2(v2[j], delta2);
+                      }
+                      *reinterpret_cast<int2*>(layer_res + (long long)tb * H +
+                                               h_base) = vp;
+                    }
+                  }
+                  float2 f[2] = {__bfloat1622float2(v2[0]),
+                                 __bfloat1622float2(v2[1])};
+                  tmem_store<4>(my_v_tmem + (si * N_CHUNK + n) * VEC, f);
+                  sq_local[n] = float2_fma(f[0], f[0], sq_local[n]);
+                  sq_local[n] = float2_fma(f[1], f[1], sq_local[n]);
+                  dot_local[n] =
+                      float2_fma(f[0], make_float2(qv[0], qv[1]), dot_local[n]);
+                  dot_local[n] =
+                      float2_fma(f[1], make_float2(qv[2], qv[3]), dot_local[n]);
+                }
+                continue;
+              }
+            }
+            int dt = si * CONSUMER_GROUPS + group;
+            if (dt >= NHT) continue;
+            int h_base = dt * K_TILE + k_local;
+            const float* qv = &q_cache[si * VEC];
+
+  #pragma unroll
+            for (int n = 0; n < AN; n++) {
+              int slot = slot_of(gci, n);
+              int4 vp = *reinterpret_cast<const int4*>(buf_ptr(slot) + h_base);
+              auto* v2 = reinterpret_cast<__nv_bfloat162*>(&vp);
+              if constexpr (HAS_DELTA) {
+                int prefix_n = N - 1 - ns;
+                if (n == prefix_n) {
+                  const bf16_t* delta_ptr = delta_buf_ptr(chunk_slot) + h_base;
+  #pragma unroll
+                  for (int j = 0; j < VEC / 2; j++) {
+                    auto delta2 = *reinterpret_cast<const __nv_bfloat162*>(
+                        delta_ptr + 2 * j);
+                    v2[j] = __hadd2(v2[j], delta2);
+                  }
+                  *reinterpret_cast<int4*>(layer_res + (long long)tb * H +
+                                           h_base) = vp;
+                }
+              }
+              float2 f[4] = {
+                  __bfloat1622float2(v2[0]), __bfloat1622float2(v2[1]),
+                  __bfloat1622float2(v2[2]), __bfloat1622float2(v2[3])};
+              tmem_store<VEC>(my_v_tmem + (si * N_CHUNK + n) * VEC, f);
+  #pragma unroll
+              for (int j = 0; j < VEC / 2; j++) {
+                sq_local[n] = float2_fma(f[j], f[j], sq_local[n]);
+                dot_local[n] = float2_fma(
+                    f[j], make_float2(qv[2 * j], qv[2 * j + 1]), dot_local[n]);
+              }
+            }
+          }
+        };
+        if constexpr (NC == 4) {
+          switch (an) {
+            case 4:
+              pass_A_body(std::integral_constant<int, 4>{});
+              break;
+            case 3:
+              pass_A_body(std::integral_constant<int, 3>{});
+              break;
+            case 2:
+              pass_A_body(std::integral_constant<int, 2>{});
+              break;
+            case 1:
+              pass_A_body(std::integral_constant<int, 1>{});
+              break;
+            default:
+              __builtin_unreachable();
+          }
+        } else if constexpr (NC == 3) {
+          switch (an) {
+            case 3:
+              pass_A_body(std::integral_constant<int, 3>{});
+              break;
+            case 2:
+              pass_A_body(std::integral_constant<int, 2>{});
+              break;
+            case 1:
+              pass_A_body(std::integral_constant<int, 1>{});
+              break;
+            default:
+              __builtin_unreachable();
+          }
+        } else {
+          static_assert(NC == 2);
+          switch (an) {
+            case 2:
+              pass_A_body(std::integral_constant<int, 2>{});
+              break;
+            case 1:
+              pass_A_body(std::integral_constant<int, 1>{});
+              break;
+            default:
+              __builtin_unreachable();
+          }
+        }
+        if (lane == 0) {
+          mbarrier_arrive(plan.bar_consumed[chunk_slot]);
+        }
+        tmem_store_wait();
+
+        float2 reduce_pair[N_CHUNK];
+  #pragma unroll
+        for (int n = 0; n < N_CHUNK; n++) {
+          reduce_pair[n] = make_float2(sq_local[n].x + sq_local[n].y,
+                                       dot_local[n].x + dot_local[n].y);
+        }
+  #pragma unroll
+        for (int offset = 16; offset > 0; offset >>= 1) {
+  #pragma unroll
+          for (int n = 0; n < N_CHUNK; n++) {
+            uint64_t packed = reinterpret_cast<uint64_t&>(reduce_pair[n]);
+            packed = __shfl_xor_sync(0xffffffff, packed, offset);
+            float2 other = reinterpret_cast<float2&>(packed);
+            reduce_pair[n] = float2_add(reduce_pair[n], other);
+          }
+        }
+        if (lane == 0) {
+  #pragma unroll
+          for (int n = 0; n < N_CHUNK; n++) {
+            plan.ws_stats[comp_wid][n] = reduce_pair[n];
+          }
+        }
+        named_barrier_sync(CONSUMER_THREADS, 0);
+
+        float local_rsig = 0.f;
+        float local_logit = 0.f;
+        int stat_n = lane / CONSUMER_WARPS;
+        int stat_w = lane % CONSUMER_WARPS;
+        float2 totals = {};
+        if (stat_n < N_CHUNK) {
+          totals = plan.ws_stats[stat_w][stat_n];
+        }
+  #pragma unroll
+        for (int offset = CONSUMER_WARPS / 2; offset > 0; offset >>= 1) {
+          totals.x +=
+              __shfl_down_sync(0xffffffff, totals.x, offset, CONSUMER_WARPS);
+          totals.y +=
+              __shfl_down_sync(0xffffffff, totals.y, offset, CONSUMER_WARPS);
+        }
+        if (stat_n < N_CHUNK && stat_w == 0) {
+          local_rsig = rsqrtf(totals.x / H + eps_cache);
+          local_logit = totals.y * local_rsig;
+        }
+        float logit_n[N_CHUNK];
+  #pragma unroll
+        for (int n = 0; n < N_CHUNK; n++) {
+          logit_n[n] = __shfl_sync(0xffffffff, local_logit, n * CONSUMER_WARPS);
+        }
+
+        float m_chunk = -FLT_MAX;
+  #pragma unroll
+        for (int n = 0; n < N_CHUNK; n++) {
+          if (n < an) m_chunk = fmaxf(m_chunk, logit_n[n]);
+        }
+        float m_new = fmaxf(m_running, m_chunk);
+        float corr = exp2f((m_running - m_new) * LOG2_E);
+        float w_n[N_CHUNK] = {};
+        float w_sum = 0.f;
+  #pragma unroll
+        for (int n = 0; n < N_CHUNK; n++) {
+          if (n < an) {
+            w_n[n] = exp2f((logit_n[n] - m_new) * LOG2_E);
+            w_sum += w_n[n];
+          }
+        }
+
+        auto pass_B_body = [&](auto AN_TOK) {
+          constexpr int AN = decltype(AN_TOK)::value;
+  #pragma unroll
+          for (int si = 0; si < SLICES_PER_GROUP; si++) {
+            if constexpr (H == 7168) {
+              if (si == SLICES_PER_GROUP - 1) {
+                float2 corr2 = make_float2(corr, corr);
+                float2 a[2];
+  #pragma unroll
+                for (int j = 0; j < 2; j++) {
+                  float2 old = make_float2(acc32[si * VEC + 2 * j],
+                                           acc32[si * VEC + 2 * j + 1]);
+                  a[j] = float2_mul(old, corr2);
+                }
+                float2 f_cache[AN][2];
+  #pragma unroll
+                for (int n = 0; n < AN; n++) {
+                  tmem_load<4>(my_v_tmem + (si * N_CHUNK + n) * VEC,
+                               f_cache[n]);
+                }
+  #pragma unroll
+                for (int n = 0; n < AN; n++) {
+                  float2 wn = make_float2(w_n[n], w_n[n]);
+  #pragma unroll
+                  for (int j = 0; j < 2; j++) {
+                    a[j] = float2_fma(wn, f_cache[n][j], a[j]);
+                  }
+                }
+  #pragma unroll
+                for (int j = 0; j < 2; j++) {
+                  acc32[si * VEC + 2 * j] = a[j].x;
+                  acc32[si * VEC + 2 * j + 1] = a[j].y;
+                }
+                continue;
+              }
+            }
+            int dt = si * CONSUMER_GROUPS + group;
+            if (dt >= NHT) continue;
+            float2 corr2 = make_float2(corr, corr);
+            float2 a[VEC / 2];
+  #pragma unroll
+            for (int j = 0; j < VEC / 2; j++) {
+              float2 old = make_float2(acc32[si * VEC + 2 * j],
+                                       acc32[si * VEC + 2 * j + 1]);
+              a[j] = float2_mul(old, corr2);
+            }
+            float2 f_cache[AN][VEC / 2];
+  #pragma unroll
+            for (int n = 0; n < AN; n++) {
+              tmem_load<VEC>(my_v_tmem + (si * N_CHUNK + n) * VEC, f_cache[n]);
+            }
+  #pragma unroll
+            for (int n = 0; n < AN; n++) {
+              float2 wn = make_float2(w_n[n], w_n[n]);
+  #pragma unroll
+              for (int j = 0; j < VEC / 2; j++) {
+                a[j] = float2_fma(wn, f_cache[n][j], a[j]);
+              }
+            }
+  #pragma unroll
+            for (int j = 0; j < VEC / 2; j++) {
+              acc32[si * VEC + 2 * j] = a[j].x;
+              acc32[si * VEC + 2 * j + 1] = a[j].y;
+            }
+          }
+        };
+        if constexpr (NC == 4) {
+          switch (an) {
+            case 4:
+              pass_B_body(std::integral_constant<int, 4>{});
+              break;
+            case 3:
+              pass_B_body(std::integral_constant<int, 3>{});
+              break;
+            case 2:
+              pass_B_body(std::integral_constant<int, 2>{});
+              break;
+            case 1:
+              pass_B_body(std::integral_constant<int, 1>{});
+              break;
+            default:
+              __builtin_unreachable();
+          }
+        } else if constexpr (NC == 3) {
+          switch (an) {
+            case 3:
+              pass_B_body(std::integral_constant<int, 3>{});
+              break;
+            case 2:
+              pass_B_body(std::integral_constant<int, 2>{});
+              break;
+            case 1:
+              pass_B_body(std::integral_constant<int, 1>{});
+              break;
+            default:
+              __builtin_unreachable();
+          }
+        } else {
+          static_assert(NC == 2);
+          switch (an) {
+            case 2:
+              pass_B_body(std::integral_constant<int, 2>{});
+              break;
+            case 1:
+              pass_B_body(std::integral_constant<int, 1>{});
+              break;
+            default:
+              __builtin_unreachable();
+          }
+        }
+
+        s_running = s_running * corr + w_sum;
+        m_running = m_new;
+      }
+
+      float inv_s = 1.f / s_running;
+      bf16_t* out_ptr = output + (long long)tb * H;
+      float2 output_sq_pair = {};
+      // When output RMSNorm is fused, the softmax denominator cancels:
+      // (acc / s) * rsqrt(mean((acc / s)^2) + eps)
+      //   = acc * rsqrt(mean(acc^2) + eps * s^2).
+  #pragma unroll
+      for (int si = 0; si < SLICES_PER_GROUP; si++) {
+        if constexpr (H == 7168) {
+          if (si == SLICES_PER_GROUP - 1) {
+            int h_base = 6 * K_TILE + group * (K_TILE / 2) + ct_in_group * 4;
+            uint2 packed;
+            auto* ov2 = reinterpret_cast<__nv_bfloat162*>(&packed);
+            float2 inv2 = make_float2(inv_s, inv_s);
+  #pragma unroll
+            for (int j = 0; j < 2; j++) {
+              float2 old = make_float2(acc32[si * VEC + 2 * j],
+                                       acc32[si * VEC + 2 * j + 1]);
+              if constexpr (HAS_OUTPUT_NORM) {
+                output_sq_pair = float2_fma(old, old, output_sq_pair);
+              } else {
+                float2 mixed = float2_mul(old, inv2);
+                ov2[j] = __float22bfloat162_rn(mixed);
+              }
+            }
+            if constexpr (!HAS_OUTPUT_NORM) {
+              *reinterpret_cast<uint2*>(out_ptr + h_base) = packed;
+            }
+            continue;
+          }
+        }
+        int dt = si * CONSUMER_GROUPS + group;
+        if (dt >= NHT) continue;
+        int h_base = dt * K_TILE + k_local;
+        uint4 packed;
+        auto* ov2 = reinterpret_cast<__nv_bfloat162*>(&packed);
+        float2 inv2 = make_float2(inv_s, inv_s);
+  #pragma unroll
+        for (int j = 0; j < VEC / 2; j++) {
+          float2 old =
+              make_float2(acc32[si * VEC + 2 * j], acc32[si * VEC + 2 * j + 1]);
+          if constexpr (HAS_OUTPUT_NORM) {
+            output_sq_pair = float2_fma(old, old, output_sq_pair);
+          } else {
+            float2 mixed = float2_mul(old, inv2);
+            ov2[j] = __float22bfloat162_rn(mixed);
+          }
+        }
+        if constexpr (!HAS_OUTPUT_NORM) {
+          *reinterpret_cast<uint4*>(out_ptr + h_base) = packed;
+        }
+      }
+
+      if constexpr (HAS_OUTPUT_NORM) {
+        if constexpr (OUTPUT_NORM_IN_SMEM) {
+          // The immutable weight copy is acquired once, at its first use.
+          if (tb == blockIdx.x) {
+            mbarrier_wait(plan.bar_output_norm_ready, 0);
+          }
+        }
+        float output_sq = output_sq_pair.x + output_sq_pair.y;
+  #pragma unroll
+        for (int offset = 16; offset > 0; offset >>= 1) {
+          output_sq += __shfl_xor_sync(0xffffffff, output_sq, offset);
+        }
+        if (lane == 0) {
+          plan.ws_stats[comp_wid][0] = make_float2(output_sq, 0.f);
+        }
+        named_barrier_sync(CONSUMER_THREADS, 0);
+        float total_sq = lane < CONSUMER_WARPS ? plan.ws_stats[lane][0].x : 0.f;
+  #pragma unroll
+        for (int offset = CONSUMER_WARPS / 2; offset > 0; offset >>= 1) {
+          total_sq +=
+              __shfl_down_sync(0xffffffff, total_sq, offset, CONSUMER_WARPS);
+        }
+        if (lane == 0) {
+          total_sq =
+              rsqrtf(total_sq / H + output_norm_eps * s_running * s_running);
+        }
+        float output_rsigma = __shfl_sync(0xffffffff, total_sq, 0);
+  #pragma unroll
+        for (int si = 0; si < SLICES_PER_GROUP; si++) {
+          if constexpr (H == 7168) {
+            if (si == SLICES_PER_GROUP - 1) {
+              int h_base = 6 * K_TILE + group * (K_TILE / 2) + ct_in_group * 4;
+              uint2 packed;
+              auto* values = reinterpret_cast<bf16_t*>(&packed);
+  #pragma unroll
+              for (int j = 0; j < 4; j++) {
+                const bf16_t* weight_ptr =
+                    OUTPUT_NORM_IN_SMEM ? output_norm_buf : output_norm_weight;
+                float weight = __bfloat162float(weight_ptr[h_base + j]);
+                values[j] = __float2bfloat16(acc32[si * VEC + j] *
+                                             output_rsigma * weight);
+              }
+              *reinterpret_cast<uint2*>(out_ptr + h_base) = packed;
+              continue;
+            }
+          }
+          int dt = si * CONSUMER_GROUPS + group;
+          if (dt >= NHT) continue;
+          int h_base = dt * K_TILE + k_local;
+          uint4 packed;
+          auto* values = reinterpret_cast<bf16_t*>(&packed);
+  #pragma unroll
+          for (int j = 0; j < VEC; j++) {
+            const bf16_t* weight_ptr =
+                OUTPUT_NORM_IN_SMEM ? output_norm_buf : output_norm_weight;
+            float weight = __bfloat162float(weight_ptr[h_base + j]);
+            values[j] =
+                __float2bfloat16(acc32[si * VEC + j] * output_rsigma * weight);
+          }
+          *reinterpret_cast<uint4*>(out_ptr + h_base) = packed;
+        }
+      }
+    }
+  }
+
+  cudaTriggerProgrammaticLaunchCompletion();
+  __syncthreads();
+  if (wid == 1) {
+    tmem_free(plan.tmem_base, TMEM_COLS_ALLOC);
+  }
+#else
+  if (threadIdx.x == 0) {
+    printf("attn_res_fwd_online_v2_kernel requires sm_10x\n");
+  }
+#endif
+}
+
+template <int H, int NC = N_CHUNK_DEFAULT, bool RELEASE_TMEM = false,
+          bool HAS_DELTA = false, bool HAS_OUTPUT_NORM = false,
+          bool OUTPUT_NORM_IN_SMEM = false>
+static void launch_fwd(const bf16_t* block_residual, bf16_t* layer_residual,
+                       const bf16_t* delta, const bf16_t* res_weight,
+                       const bf16_t* rms_weight, bf16_t* output, int N, int T,
+                       int B, float rms_eps, int num_sm, cudaStream_t stream,
+                       const bf16_t* output_norm_weight = nullptr,
+                       float output_norm_eps = 0.f, int block_stride_m = 0,
+                       int block_stride_r = 0) {
+  constexpr size_t smem_size =
+      ((size_t)CHUNK_DEPTH * (NC + (HAS_DELTA ? 1 : 0)) * H * sizeof(bf16_t) +
+       (OUTPUT_NORM_IN_SMEM ? (size_t)H * sizeof(bf16_t) : 0) +
+       sizeof(FwdSmemPlan<NC>) + 15) &
+      ~size_t(15);
+  auto kernel =
+      &attn_res_fwd_online_v2_kernel<H, NC, RELEASE_TMEM, HAS_DELTA,
+                                     HAS_OUTPUT_NORM, OUTPUT_NORM_IN_SMEM>;
+  static bool attrs_set = false;
+  if (!attrs_set) {
+    if (smem_size > 48 * 1024) {
+      cudaFuncSetAttribute(kernel, cudaFuncAttributeMaxDynamicSharedMemorySize,
+                           smem_size);
+    }
+    attrs_set = true;
+  }
+  int grid = RELEASE_TMEM ? num_sm * 2 : num_sm;
+  cudaLaunchConfig_t config{};
+  config.gridDim = grid;
+  config.blockDim = BLK;
+  config.dynamicSmemBytes = smem_size;
+  config.stream = stream;
+  cudaLaunchAttribute attrs[1];
+  attrs[0].id = cudaLaunchAttributeProgrammaticStreamSerialization;
+  attrs[0].val.programmaticStreamSerializationAllowed = 1;
+  config.attrs = attrs;
+  config.numAttrs = 1;
+  cudaLaunchKernelEx(&config, kernel, block_residual, layer_residual, delta,
+                     res_weight, rms_weight, output, N, T, B, block_stride_m,
+                     block_stride_r, rms_eps, output_norm_weight,
+                     output_norm_eps);
+}
+
+}  // namespace fwd_prod_v2
+}  // namespace sm100
+
+void kimi_k3_attn_res(torch::stable::Tensor& prefix,
+                      torch::stable::Tensor const& delta,
+                      torch::stable::Tensor const& blocks,
+                      torch::stable::Tensor const& norm_weight,
+                      torch::stable::Tensor const& qk_weight,
+                      torch::stable::Tensor const& output_norm_weight,
+                      torch::stable::Tensor& output, int64_t num_blocks,
+                      double eps, double output_norm_eps) {
+  int const num_tokens = static_cast<int>(prefix.size(0));
+  int const device = prefix.get_device_index();
+  torch::stable::accelerator::DeviceGuard const device_guard(device);
+  cudaDeviceProp const* properties = get_device_prop();
+  STD_TORCH_CHECK(properties->major == 10,
+                  "Kimi K3 AttnRes requires the SM100 family");
+
+  using namespace sm100::fwd_prod_v2;
+  // Two-source chunks and two resident CTAs are beneficial once setup is
+  // amortized by the long, full eight-block prefill workload.
+  if (num_blocks == 8 && num_tokens >= 4096) {
+    launch_fwd<7168, 2, true, true, true, true>(
+        static_cast<bf16_t const*>(blocks.data_ptr()),
+        static_cast<bf16_t*>(prefix.data_ptr()),
+        static_cast<bf16_t const*>(delta.data_ptr()),
+        static_cast<bf16_t const*>(qk_weight.data_ptr()),
+        static_cast<bf16_t const*>(norm_weight.data_ptr()),
+        static_cast<bf16_t*>(output.data_ptr()),
+        static_cast<int>(num_blocks) + 1, num_tokens, 1,
+        static_cast<float>(eps), properties->multiProcessorCount,
+        get_current_cuda_stream(device),
+        static_cast<bf16_t const*>(output_norm_weight.data_ptr()),
+        static_cast<float>(output_norm_eps), static_cast<int>(blocks.stride(0)),
+        static_cast<int>(blocks.stride(1)));
+  } else {
+    launch_fwd<7168, 4, false, true, true, true>(
+        static_cast<bf16_t const*>(blocks.data_ptr()),
+        static_cast<bf16_t*>(prefix.data_ptr()),
+        static_cast<bf16_t const*>(delta.data_ptr()),
+        static_cast<bf16_t const*>(qk_weight.data_ptr()),
+        static_cast<bf16_t const*>(norm_weight.data_ptr()),
+        static_cast<bf16_t*>(output.data_ptr()),
+        static_cast<int>(num_blocks) + 1, num_tokens, 1,
+        static_cast<float>(eps), properties->multiProcessorCount,
+        get_current_cuda_stream(device),
+        static_cast<bf16_t const*>(output_norm_weight.data_ptr()),
+        static_cast<float>(output_norm_eps), static_cast<int>(blocks.stride(0)),
+        static_cast<int>(blocks.stride(1)));
+  }
+  cudaError_t const error = cudaGetLastError();
+  STD_TORCH_CHECK(
+      error == cudaSuccess,
+      "Kimi K3 AttnRes kernel launch failed: ", cudaGetErrorString(error));
+}

csrc/libtorch_stable/layernorm_kernels.cu

@@ -249,7 +249,9 @@ void rms_norm(torch::stable::Tensor& out,    // [..., hidden_size]
   int64_t input_shape_d3 = (num_dims >= 4) ? input.size(-3) : 0;
 
   // For large num_tokens, use smaller blocks to increase SM concurrency.
-  const int max_block_size = (num_tokens < 256) ? 1024 : 256;
+  const bool batch_invariant_launch = vllm::vllm_is_batch_invariant();
+  const int max_block_size =
+      batch_invariant_launch ? 1024 : ((num_tokens < 256) ? 1024 : 256);
   dim3 grid(num_tokens);
   const torch::stable::accelerator::DeviceGuard device_guard(
       input.get_device_index());
@@ -325,8 +327,13 @@ void fused_add_rms_norm(torch::stable::Tensor& input,     // [..., hidden_size]
   /* This kernel is memory-latency bound in many scenarios.
      When num_tokens is large, a smaller block size allows
      for increased block occupancy on CUs and better latency
-     hiding on global mem ops. */
-  const int max_block_size = (num_tokens < 256) ? 1024 : 256;
+     hiding on global mem ops. In batch-invariant mode the block size must
+     not depend on num_tokens, otherwise the same token would use a different
+     reduction width (and thus a different floating-point summation order)
+     across batches; lock it to 1024 to keep results bit-exact. */
+  const bool batch_invariant_launch = vllm::vllm_is_batch_invariant();
+  const int max_block_size =
+      batch_invariant_launch ? 1024 : ((num_tokens < 256) ? 1024 : 256);
   dim3 block(std::min(hidden_size, max_block_size));
   const torch::stable::accelerator::DeviceGuard device_guard(
       input.get_device_index());
@@ -337,7 +344,6 @@ void fused_add_rms_norm(torch::stable::Tensor& input,     // [..., hidden_size]
   auto res_ptr = reinterpret_cast<std::uintptr_t>(residual.data_ptr());
   bool offsets_are_multiple_of_vector_width =
       hidden_size % vector_width == 0 && input_stride % vector_width == 0;
-  bool batch_invariant_launch = vllm::vllm_is_batch_invariant();
   const bool has_weight = weight.has_value();
   if (has_weight) {
     auto wt_ptr = reinterpret_cast<std::uintptr_t>(weight->data_ptr());

csrc/libtorch_stable/layernorm_quant_kernels.cu

@@ -215,7 +215,9 @@ void rms_norm_static_fp8_quant(
   int num_tokens = input.numel() / hidden_size;
 
   // For large num_tokens, use smaller blocks to increase SM concurrency.
-  const int max_block_size = (num_tokens < 256) ? 1024 : 256;
+  const bool batch_invariant_launch = vllm::vllm_is_batch_invariant();
+  const int max_block_size =
+      batch_invariant_launch ? 1024 : ((num_tokens < 256) ? 1024 : 256);
   dim3 grid(num_tokens);
   const torch::stable::accelerator::DeviceGuard device_guard(
       input.get_device_index());
@@ -279,7 +281,9 @@ void fused_add_rms_norm_static_fp8_quant(
      When num_tokens is large, a smaller block size allows
      for increased block occupancy on CUs and better latency
      hiding on global mem ops. */
-  const int max_block_size = (num_tokens < 256) ? 1024 : 256;
+  const bool batch_invariant_launch = vllm::vllm_is_batch_invariant();
+  const int max_block_size =
+      batch_invariant_launch ? 1024 : ((num_tokens < 256) ? 1024 : 256);
   dim3 block(std::min(hidden_size, max_block_size));
   const torch::stable::accelerator::DeviceGuard device_guard(
       input.get_device_index());
@@ -296,7 +300,6 @@ void fused_add_rms_norm_static_fp8_quant(
   auto wt_ptr = reinterpret_cast<std::uintptr_t>(weight.data_ptr());
   bool ptrs_are_aligned =
       inp_ptr % 16 == 0 && res_ptr % 16 == 0 && wt_ptr % 16 == 0;
-  bool batch_invariant_launch = vllm::vllm_is_batch_invariant();
   if (ptrs_are_aligned && hidden_size % 8 == 0 && input_stride % 8 == 0 &&
       !batch_invariant_launch) {
     LAUNCH_FUSED_ADD_RMS_NORM(8);

csrc/libtorch_stable/ops.h

@@ -315,6 +315,17 @@ void fused_minimax_m3_qknorm_rope_kv_insert(
     std::optional<torch::stable::Tensor> index_q_out,
     const std::string& kv_cache_dtype, bool skip_index_branch);
 
+#ifdef VLLM_ENABLE_KIMI_K3_ATTN_RES
+void kimi_k3_attn_res(torch::stable::Tensor& prefix,
+                      torch::stable::Tensor const& delta,
+                      torch::stable::Tensor const& blocks,
+                      torch::stable::Tensor const& norm_weight,
+                      torch::stable::Tensor const& qk_weight,
+                      torch::stable::Tensor const& output_norm_weight,
+                      torch::stable::Tensor& output, int64_t num_blocks,
+                      double eps, double output_norm_eps);
+#endif
+
 // Sampler kernels (shared CUDA/ROCm)
 void apply_repetition_penalties_(
     torch::stable::Tensor& logits, const torch::stable::Tensor& prompt_mask,

csrc/libtorch_stable/quantization/fused_kernels/fused_layernorm_dynamic_per_token_quant.cu

@@ -2,6 +2,7 @@
 #include "../../torch_utils.h"
 
 #include "../../dispatch_utils.h"
+#include "../../../core/batch_invariant.hpp"
 #include "layernorm_utils.cuh"
 #include "quant_conversions.cuh"
 
@@ -231,7 +232,9 @@ void rms_norm_per_block_quant_dispatch(
   auto num_tokens = input.numel() / hidden_size;
 
   dim3 grid(num_tokens);
-  const int max_block_size = (num_tokens <= 256) ? 512 : 256;
+  const bool batch_invariant_launch = vllm::vllm_is_batch_invariant();
+  const int max_block_size =
+      batch_invariant_launch ? 512 : ((num_tokens <= 256) ? 512 : 256);
   dim3 block(std::min(hidden_size, max_block_size));
   const torch::stable::accelerator::DeviceGuard device_guard(
       input.get_device_index());

csrc/libtorch_stable/torch_bindings.cpp

@@ -468,6 +468,14 @@ STABLE_TORCH_LIBRARY_FRAGMENT(_C, ops) {
       "int block_size, Tensor!? q_out, Tensor!? index_q_out, "
       "str kv_cache_dtype, bool skip_index_branch=False) -> ()");
 
+#ifdef VLLM_ENABLE_KIMI_K3_ATTN_RES
+  ops.def(
+      "kimi_k3_attn_res("
+      "Tensor! prefix, Tensor delta, Tensor blocks, Tensor norm_weight, "
+      "Tensor qk_weight, Tensor output_norm_weight, Tensor! output, "
+      "int num_blocks, float eps, float output_norm_eps) -> ()");
+#endif
+
   // Apply repetition penalties to logits in-place.
   ops.def(
       "apply_repetition_penalties_(Tensor! logits, Tensor prompt_mask, "
@@ -693,6 +701,9 @@ STABLE_TORCH_LIBRARY_IMPL(_C, CUDA, ops) {
 #endif
   ops.impl("fused_minimax_m3_qknorm_rope_kv_insert",
            TORCH_BOX(&fused_minimax_m3_qknorm_rope_kv_insert));
+#ifdef VLLM_ENABLE_KIMI_K3_ATTN_RES
+  ops.impl("kimi_k3_attn_res", TORCH_BOX(&kimi_k3_attn_res));
+#endif
 
   // Sampler kernels (shared CUDA/ROCm)
   ops.impl("apply_repetition_penalties_",

csrc/quickreduce/quick_reduce.h

@@ -22,17 +22,25 @@ template <typename AllReduceKernel, typename T>
 __global__ __quickreduce_launch_bounds_two_shot__ static void
 allreduce_prototype_twoshot(T const* A, T* B, uint32_t N, uint32_t num_blocks,
                             int rank, uint8_t** dbuffer_list,
-                            uint32_t data_offset, uint32_t flag_color,
+                            uint32_t data_offset, uint32_t* d_flag_counters,
                             int64_t data_size_per_phase) {
   int block = blockIdx.x;
   int grid = gridDim.x;
 
+  // Load this block's counter from device memory and advance it on-device,
+  // so the color keeps changing across graph replays instead of being frozen.
+  uint32_t flag_color = d_flag_counters[blockIdx.x];
+
   while (block < num_blocks) {
     AllReduceKernel::run(A, B, N, block, rank, dbuffer_list, data_offset,
                          flag_color, data_size_per_phase);
     block += grid;
     flag_color++;
   }
+  // All threads compute the same final value; one writer per block is enough.
+  if (threadIdx.x == 0 && threadIdx.y == 0) {
+    d_flag_counters[blockIdx.x] = flag_color;
+  }
 }
 
 #define TWOSHOT_DISPATCH(__codec)                                           \
@@ -42,21 +50,21 @@ allreduce_prototype_twoshot(T const* A, T* B, uint32_t N, uint32_t num_blocks,
     hipLaunchKernelGGL((allreduce_prototype_twoshot<AllReduceKernel, T>),   \
                        dim3(grid), dim3(kBlockTwoShot), 0, stream, A, B, N, \
                        num_blocks, rank, dbuffer_list, data_offset,         \
-                       flag_color, this->kMaxProblemSize);                  \
+                       d_flag_counters, this->kMaxProblemSize);             \
   } else if (world_size == 4) {                                             \
     using LineCodec = __codec<T, 4>;                                        \
     using AllReduceKernel = AllReduceTwoshot<T, LineCodec, cast_bf2half>;   \
     hipLaunchKernelGGL((allreduce_prototype_twoshot<AllReduceKernel, T>),   \
                        dim3(grid), dim3(kBlockTwoShot), 0, stream, A, B, N, \
                        num_blocks, rank, dbuffer_list, data_offset,         \
-                       flag_color, this->kMaxProblemSize);                  \
+                       d_flag_counters, this->kMaxProblemSize);             \
   } else if (world_size == 8) {                                             \
     using LineCodec = __codec<T, 8>;                                        \
     using AllReduceKernel = AllReduceTwoshot<T, LineCodec, cast_bf2half>;   \
     hipLaunchKernelGGL((allreduce_prototype_twoshot<AllReduceKernel, T>),   \
                        dim3(grid), dim3(kBlockTwoShot), 0, stream, A, B, N, \
                        num_blocks, rank, dbuffer_list, data_offset,         \
-                       flag_color, this->kMaxProblemSize);                  \
+                       d_flag_counters, this->kMaxProblemSize);             \
   }
 
 // INT3 only retains good performance on TP2 (world_size == 2). On TP4/TP8
@@ -69,7 +77,7 @@ allreduce_prototype_twoshot(T const* A, T* B, uint32_t N, uint32_t num_blocks,
     hipLaunchKernelGGL((allreduce_prototype_twoshot<AllReduceKernel, T>),   \
                        dim3(grid), dim3(kBlockTwoShot), 0, stream, A, B, N, \
                        num_blocks, rank, dbuffer_list, data_offset,         \
-                       flag_color, this->kMaxProblemSize);                  \
+                       d_flag_counters, this->kMaxProblemSize);             \
   } else {                                                                  \
     throw std::runtime_error(                                               \
         "INT3 quick all-reduce is only supported for world_size == 2 "      \
@@ -94,7 +102,7 @@ struct DeviceComms {
   static int constexpr kMaxWorldSize = 8;
 
   bool initialized = false;
-  uint32_t flag_color = 1;
+  uint32_t* d_flag_counters = nullptr;
   int world_size;
   int rank;
 
@@ -128,6 +136,16 @@ struct DeviceComms {
     // Clear the flags buffer.
     HIP_CHECK(hipMemset(dbuffer, 0, flags_buffer_size));
 
+    // One flag-color counter per block, advanced by the kernel. Start at 1
+    // to stay clear of the flags buffer we just zeroed.
+    HIP_CHECK(hipMalloc(&d_flag_counters, kMaxNumBlocks * sizeof(uint32_t)));
+    {
+      std::vector<uint32_t> init_color(kMaxNumBlocks, 1u);
+      HIP_CHECK(hipMemcpy(d_flag_counters, init_color.data(),
+                          kMaxNumBlocks * sizeof(uint32_t),
+                          hipMemcpyHostToDevice));
+    }
+
     // Device-side list of IPC buffers.
     buffer_list.resize(world_size);
     HIP_CHECK(hipMalloc(&dbuffer_list, world_size * sizeof(uint8_t*)));
@@ -144,6 +162,12 @@ struct DeviceComms {
   hipIpcMemHandle_t const get_handle() { return buffer_ipc_handle; }
 
   void destroy() {
+    // Allocated before `initialized` flips true, so free it on its own guard
+    // to avoid a leak if init fails partway through.
+    if (d_flag_counters) {
+      HIP_CHECK(hipFree(d_flag_counters));
+      d_flag_counters = nullptr;
+    }
     if (initialized) {
       for (int i = 0; i < world_size; i++) {
         if (i != rank) {
@@ -211,8 +235,6 @@ struct DeviceComms {
         break;
     }
     HIP_CHECK(cudaGetLastError());
-    // Rotate the flag color.
-    flag_color += divceil(N, grid);
   }
 };
 

csrc/spinloop.cpp

@@ -7,7 +7,7 @@ extern "C" {
 
 #if defined(__i386__) || defined(__x86_64__)
   #include <cpuid.h>
-  #include <mwaitxintrin.h>
+  #include <x86intrin.h>
 #endif
 
 #if defined(CLOCK_MONOTONIC_RAW)

docs/.nav.yml

@@ -56,7 +56,9 @@ nav:
   - API Reference:
     - api/README.md
     - api/vllm
-  - CLI Reference: cli
+  - CLI Reference:
+    - cli/README.md
+    - vllm: cli
   - Community:
     - community/*
     - Governance: governance

docs/cli/.nav.yml

@@ -1,10 +1,8 @@
 nav:
-  - README.md
-  - serve.md
-  - chat.md
-  - complete.md
-  - run-batch.md
-  - vllm bench:
-    - bench/**/*.md
-  - vllm launch:
-    - launch/**/*.md
+  - "*.md"
+  - bench:
+    - bench/*.md
+    - sweep:
+      - bench/sweep/*.md
+  - launch:
+    - launch/*.md

docs/cli/bench/latency.md

@@ -1,9 +0,0 @@
-# vllm bench latency
-
-## JSON CLI Arguments
-
---8<-- "docs/cli/json_tip.inc.md"
-
-## Arguments
-
---8<-- "docs/generated/argparse/bench_latency.inc.md"

docs/cli/bench/mm_processor.md

@@ -1,55 +0,0 @@
-# vllm bench mm-processor
-
-## Overview
-
-`vllm bench mm-processor` profiles the multimodal input processor pipeline of
-vision-language models. It measures per-stage latency from the HuggingFace
-processor through to the encoder forward pass, helping you identify
-preprocessing bottlenecks and understand how different image resolutions or
-item counts affect end-to-end request time.
-
-The benchmark supports two data sources: synthetic random multimodal inputs
-(`random-mm`) and HuggingFace datasets (`hf`). Warmup requests are run before
-measurement to ensure stable results.
-
-## Quick Start
-
-```bash
-vllm bench mm-processor \
-  --model Qwen/Qwen2-VL-7B-Instruct \
-  --dataset-name random-mm \
-  --num-prompts 50 \
-  --random-input-len 300 \
-  --random-output-len 40 \
-  --random-mm-base-items-per-request 2 \
-  --random-mm-limit-mm-per-prompt '{"image": 3, "video": 0}' \
-  --random-mm-bucket-config '{(256, 256, 1): 0.7, (720, 1280, 1): 0.3}'
-```
-
-## Measured Stages
-
-| Stage | Description |
-| ----- | ----------- |
-| `get_mm_hashes_secs` | Time spent hashing multimodal inputs |
-| `get_cache_missing_items_secs` | Time spent looking up the processor cache |
-| `apply_hf_processor_secs` | Time spent in the HuggingFace processor |
-| `merge_mm_kwargs_secs` | Time spent merging multimodal kwargs |
-| `apply_prompt_updates_secs` | Time spent updating prompt tokens |
-| `preprocessor_total_secs` | Total preprocessing time |
-| `encoder_forward_secs` | Time spent in the encoder model forward pass |
-| `num_encoder_calls` | Number of encoder invocations per request |
-
-The benchmark also reports end-to-end latency (TTFT + decode time) per
-request. Use `--metric-percentiles` to select which percentiles to report
-(default: p99) and `--output-json` to save results.
-
-For more examples (HF datasets, warmup, JSON output), see
-[Benchmarking CLI — Multimodal Processor Benchmark](../../benchmarking/cli.md#multimodal-processor-benchmark).
-
-## JSON CLI Arguments
-
---8<-- "docs/cli/json_tip.inc.md"
-
-## Arguments
-
---8<-- "docs/generated/argparse/bench_mm_processor.inc.md"

docs/cli/bench/serve.md

@@ -1,9 +0,0 @@
-# vllm bench serve
-
-## JSON CLI Arguments
-
---8<-- "docs/cli/json_tip.inc.md"
-
-## Arguments
-
---8<-- "docs/generated/argparse/bench_serve.inc.md"

docs/cli/bench/sweep/plot.md

@@ -1,9 +0,0 @@
-# vllm bench sweep plot
-
-## JSON CLI Arguments
-
---8<-- "docs/cli/json_tip.inc.md"
-
-## Arguments
-
---8<-- "docs/generated/argparse/bench_sweep_plot.inc.md"

docs/cli/bench/sweep/plot_pareto.md

@@ -1,9 +0,0 @@
-# vllm bench sweep plot_pareto
-
-## JSON CLI Arguments
-
---8<-- "docs/cli/json_tip.inc.md"
-
-## Arguments
-
---8<-- "docs/generated/argparse/bench_sweep_plot_pareto.inc.md"

docs/cli/bench/sweep/serve.md

@@ -1,9 +0,0 @@
-# vllm bench sweep serve
-
-## JSON CLI Arguments
-
---8<-- "docs/cli/json_tip.inc.md"
-
-## Arguments
-
---8<-- "docs/generated/argparse/bench_sweep_serve.inc.md"

docs/cli/bench/sweep/serve_workload.md

@@ -1,9 +0,0 @@
-# vllm bench sweep serve_workload
-
-## JSON CLI Arguments
-
---8<-- "docs/cli/json_tip.inc.md"
-
-## Arguments
-
---8<-- "docs/generated/argparse/bench_sweep_serve_workload.inc.md"

docs/cli/bench/throughput.md

@@ -1,9 +0,0 @@
-# vllm bench throughput
-
-## JSON CLI Arguments
-
---8<-- "docs/cli/json_tip.inc.md"
-
-## Arguments
-
---8<-- "docs/generated/argparse/bench_throughput.inc.md"

docs/cli/chat.md

@@ -1,5 +0,0 @@
-# vllm chat
-
-## Arguments
-
---8<-- "docs/generated/argparse/chat.inc.md"

docs/cli/complete.md

@@ -1,5 +0,0 @@
-# vllm complete
-
-## Arguments
-
---8<-- "docs/generated/argparse/complete.inc.md"

docs/cli/json_tip.inc.md

@@ -1,10 +0,0 @@
-<!-- markdownlint-disable MD041 -->
-When passing JSON CLI arguments, the following sets of arguments are equivalent:
-
-- `--json-arg '{"key1": "value1", "key2": {"key3": "value2"}}'`
-- `--json-arg.key1 value1 --json-arg.key2.key3 value2`
-
-Additionally, list elements can be passed individually using `+`:
-
-- `--json-arg '{"key4": ["value3", "value4", "value5"]}'`
-- `--json-arg.key4+ value3 --json-arg.key4+='value4,value5'`

docs/cli/launch/render.md

@@ -1,22 +0,0 @@
-# vllm launch render
-
-## Overview
-
-`vllm launch render` starts a GPU-less rendering server for preprocessing and
-postprocessing only.
-
-```bash
-vllm launch render meta-llama/Llama-3.2-1B-Instruct --port 8100
-```
-
-This command reuses the standard serving parser, so model, frontend,
-networking, and related CLI options follow the same conventions as
-[`vllm serve`](../serve.md).
-
-## JSON CLI Arguments
-
---8<-- "docs/cli/json_tip.inc.md"
-
-## Arguments
-
---8<-- "docs/generated/argparse/launch_render.inc.md"

docs/cli/run-batch.md

@@ -1,9 +0,0 @@
-# vllm run-batch
-
-## JSON CLI Arguments
-
---8<-- "docs/cli/json_tip.inc.md"
-
-## Arguments
-
---8<-- "docs/generated/argparse/run-batch.inc.md"

docs/cli/serve.md

@@ -1,9 +0,0 @@
-# vllm serve
-
-## JSON CLI Arguments
-
---8<-- "docs/cli/json_tip.inc.md"
-
-## Arguments
-
---8<-- "docs/generated/argparse/serve.inc.md"

docs/configuration/engine_args.md

@@ -11,12 +11,4 @@ Engine arguments control the behavior of the vLLM engine.
 
 The engine argument classes, [EngineArgs][vllm.engine.arg_utils.EngineArgs] and [AsyncEngineArgs][vllm.engine.arg_utils.AsyncEngineArgs], are a combination of the configuration classes defined in [vllm.config][]. Therefore, if you are interested in developer documentation, we recommend looking at these configuration classes as they are the source of truth for types, defaults and docstrings.
 
---8<-- "docs/cli/json_tip.inc.md"
-
-## `EngineArgs`
-
---8<-- "docs/generated/argparse/engine_args.inc.md"
-
-## `AsyncEngineArgs`
-
---8<-- "docs/generated/argparse/async_engine_args.inc.md"
+--8<-- "gen:engine-args"

docs/contributing/README.md

@@ -301,8 +301,10 @@ review process:
   isn't clear or you disagree with a suggestion, feel free to ask for
   clarification or discuss the suggestion.
 - Note that not all CI checks will be executed due to limited computational
-  resources. The reviewer will add `ready` label to the PR when the PR is
-  ready to merge or a full CI run is needed.
+  resources. Reviewers with write access and configured trusted contributors
+  can comment `/ci run` when CI signals are needed before a PR is ready. After
+  the PR is approved or has the `ready` label, the PR author can use `/ci run`
+  or `/ci retry`. New commits do not start CI automatically.
 
 ### Pull Request Limits and Escalation
 

docs/contributing/model/transcription.md

@@ -195,7 +195,7 @@ Provide a fast duration→token estimate to improve streaming usage statistics:
 The API server takes care of basic audio I/O and optional chunking before building prompts:
 
 - Resampling: Input audio is resampled to `SpeechToTextConfig.sample_rate` using `AudioResampler`.
-- Chunking: If `SpeechToTextConfig.allow_audio_chunking` is True and the duration exceeds `max_audio_clip_s`, the server splits the audio into overlapping chunks and generates a prompt per chunk. Overlap is controlled by `overlap_chunk_second`.
+- Chunking: If `SpeechToTextConfig.allow_audio_chunking` is True and the duration exceeds `max_audio_clip_s`, the server splits the audio into chunks and generates a prompt per chunk. There is no overlap between chunks, overlap_chunk_second controls the size of the search window used to find the split point.
 - Energy-aware splitting: When `min_energy_split_window_size` is set, the server finds low-energy regions to minimize cutting within words.
 
 Relevant server logic:

docs/deployment/docker.md

@@ -8,6 +8,26 @@ toc_depth: 2
 
 --8<-- "docs/getting_started/installation/gpu.md:pre-built-images"
 
+## Persist the compile cache across containers
+
+Mounting the Hugging Face cache keeps model weights across containers, but each
+new container still starts with an empty `VLLM_CACHE_ROOT` (default
+`~/.cache/vllm`) and recompiles the model's `torch.compile` artifacts. Mount a
+named volume at that path to reuse the inductor, Triton, and AOT artifacts from
+the second container onward:
+
+```bash
+docker run --rm --gpus all \
+    -v ~/.cache/huggingface:/root/.cache/huggingface \
+    -v vllm-cache:/root/.cache/vllm \
+    -p 8000:8000 \
+    vllm/vllm-openai:latest \
+    meta-llama/Llama-3.1-8B-Instruct
+```
+
+See [Faster Startup](../configuration/optimization.md#faster-startup) for the
+mechanism and for what invalidates the cache.
+
 ## Run as a non-root user
 
 The CUDA `vllm/vllm-openai` image runs as root by default for backward

docs/deployment/integrations/llm-d.md

@@ -1,5 +1,37 @@
 # llm-d
 
-vLLM can be deployed with [llm-d](https://github.com/llm-d/llm-d), a Kubernetes-native distributed inference serving stack providing well-lit paths for anyone to serve large generative AI models at scale. It helps achieve the fastest "time to state-of-the-art (SOTA) performance" for key OSS models across most hardware accelerators and infrastructure providers.
+[llm-d](https://llm-d.ai/) is a Kubernetes-native distributed inference framework for serving large language models at scale, with vLLM as its primary inference engine. llm-d coordinates a fleet of vLLM instances across a cluster so that performance holds up under real production traffic, achieving the fastest "time to state-of-the-art (SOTA) performance" for key OSS models across most hardware accelerators.
 
-You can use vLLM with llm-d directly by following [the official guides](https://llm-d.ai/docs/guides) or via [KServe's LLMInferenceService](https://kserve.github.io/website/docs/model-serving/generative-inference/llmisvc/llmisvc-overview).
+It is a [CNCF Sandbox project](https://www.cncf.io/blog/2026/03/24/welcome-llm-d-to-the-cncf-evolving-kubernetes-into-sota-ai-infrastructure/) founded by Red Hat, Google Cloud, IBM Research, CoreWeave, and NVIDIA.
+
+## What llm-d adds to vLLM
+
+A single vLLM server is fast, but at scale the picture changes: across many replicas, cache locality breaks under round-robin load balancing, long prompts inflate time-to-first-token, and accelerators sit underused. llm-d adds the cluster-level layer that vLLM does not aim to provide on its own:
+
+- **[Prefix-aware routing](https://llm-d.ai/docs/guides/precise-prefix-cache-aware).** Instead of round-robin, llm-d reads vLLM's KV-cache events and routes each request to the replica that already holds its prefix, reusing cache instead of recomputing it.
+- **[Distributed KV-cache management](https://llm-d.ai/docs/guides#advanced-kv-cache-management).** A global index tracks which token blocks live on which replica, and [tiered offloading](https://llm-d.ai/docs/guides/tiered-prefix-cache) spills cache to CPU memory or local SSD, extending the working set beyond accelerator HBM.
+- **[Prefill/decode disaggregation](https://llm-d.ai/docs/guides/pd-disaggregation).** Prompt processing and token generation run on separate vLLM workers, with KV-cache moved over the vLLM [NIXL connector](https://docs.vllm.ai/en/latest/features/nixl_connector_usage/), lowering TTFT and steadying per-token latency on long prompts.
+- **[Wide expert-parallelism](https://llm-d.ai/docs/guides/wide-expert-parallelism).** Serve large Mixture-of-Experts models such as DeepSeek-R1 and GPT-OSS across nodes with combined data and expert parallelism, for more KV-cache capacity and throughput.
+- **SLO-aware [autoscaling](https://llm-d.ai/docs/guides/workload-autoscaling) and [flow control](https://llm-d.ai/docs/guides/flow-control).** Scale vLLM pools on real inference signals (queue depth, true demand) rather than raw GPU utilization, with multi-tenant fairness and priority dispatch.
+
+These are composable. Most teams start by adding prefix-aware routing over an existing vLLM pool, then layer in the rest as specific bottlenecks appear.
+
+## Performance
+
+Representative benchmarked results across accelerators:
+
+- **3x higher output throughput** and **2x faster TTFT** from prefix-aware routing vs round-robin (Llama 3.1 70B, AMD MI300X)
+- **Up to 70% higher tokens/sec** from prefill/decode disaggregation (GPT-OSS, NVIDIA B200)
+- **13.9x throughput** from hierarchical KV offloading at high concurrency vs GPU-only (NVIDIA H100)
+
+See the [full list](https://github.com/llm-d/llm-d#performance-highlights) and reproducible benchmarks on [Prism](https://prism.llm-d.ai/).
+
+## Get started
+
+1. Deploy the [Optimized Baseline](https://llm-d.ai/docs/guides/optimized-baseline) with the [Quickstart](https://llm-d.ai/docs/getting-started/quickstart). It stands up an intelligent router over a vLLM pool on Kubernetes in a tested configuration.
+2. Browse the [well-lit path guides](https://llm-d.ai/docs/guides), each a tested recipe for one of the capabilities above, and add the optimization that fits your workload.
+3. Read the [Introduction](https://llm-d.ai/docs/getting-started) and [Architecture overview](https://llm-d.ai/docs/architecture) to see how the pieces wrap your vLLM deployment.
+
+You can also deploy vLLM with llm-d via [KServe's LLMInferenceService](https://kserve.github.io/website/docs/model-serving/generative-inference/llmisvc/llmisvc-overview).
+
+Questions and contributions are welcome on [GitHub](https://github.com/llm-d/llm-d) and [Slack](https://llm-d.ai/slack).

docs/design/attention_backends.md

@@ -1,15 +1,9 @@
 # Attention Backend Feature Support
 
-This document is auto-generated by `tools/pre_commit/generate_attention_backend_docs.py`.
-It shows the feature support for each registered attention backend
-based on the checks in `AttentionBackend.validate_configuration()`.
-
-**Do not edit this file manually.** Run the following command to
-regenerate it:
-
-```bash
-python tools/pre_commit/generate_attention_backend_docs.py
-```
+The priority and feature tables on this page are auto-generated from the
+attention backend registry by
+`docs/mkdocs/gen_files/generate_attention_backends.py`, based on the checks in
+`AttentionBackend.validate_configuration()`.
 
 ## Setting the Attention Backend
 
@@ -98,40 +92,11 @@ Priority is **1 = highest** (tried first).
 
 ### Standard Attention (MHA, MQA, GQA)
 
-**Blackwell (SM 10.x):**
-
-| Priority | Backend |
-| -------- | ------- |
-| 1 | `FLASHINFER` |
-| 2 | `FLASH_ATTN` |
-| 3 | `TRITON_ATTN` |
-| 4 | `FLEX_ATTENTION` |
-| 5 | `TURBOQUANT` |
-
-**Ampere/Hopper (SM 8.x-9.x):**
-
-| Priority | Backend |
-| -------- | ------- |
-| 1 | `FLASH_ATTN` |
-| 2 | `FLASHINFER` |
-| 3 | `TRITON_ATTN` |
-| 4 | `FLEX_ATTENTION` |
-| 5 | `TURBOQUANT` |
+--8<-- "gen:priority-standard"
 
 ### MLA Attention (DeepSeek-style)
 
-**Blackwell (SM 10.x):**
-
-| Priority | Backend |
-| -------- | ------- |
-| 1 | `FLASHINFER_MLA` |
-| 2 | `TOKENSPEED_MLA` |
-| 3 | `CUTLASS_MLA` |
-| 4 | `FLASH_ATTN_MLA` |
-| 5 | `FLASHMLA` |
-| 6 | `TRITON_MLA` |
-| 7 | `FLASHINFER_MLA_SPARSE`**\*** |
-| 8 | `FLASHMLA_SPARSE` |
+--8<-- "gen:priority-mla"
 
 > **\*** For sparse MLA, FP8 KV cache always prefers `FLASHINFER_MLA_SPARSE`. With BF16 KV cache, `FLASHINFER_MLA_SPARSE` is preferred for low query-head counts (<= 16), while `FLASHMLA_SPARSE` is preferred otherwise.
 >
@@ -157,24 +122,7 @@ Priority is **1 = highest** (tried first).
 
 ## Standard Attention (MHA, MQA, GQA) Backends
 
-| Backend | Version | Dtypes | KV Dtypes | Block Sizes | Head Sizes | Sink | Non-Causal | MM Prefix | DCP | Attention Types | Compute Cap. |
-| ------- | ------- | ------ | --------- | ----------- | ---------- | ---- | ---------- | --------- | --- | --------------- | ------------ |
-| `CPU_ATTN` | | fp16, bf16, fp32 | `auto`, `fp8`, `fp8_e4m3`, `fp8_e5m2` | %16 | 32, 64, 80, 96, 112, 128, 160, 192, 224, 256, 512 | ❌ | ✅ | ❌ | ❌ | All | N/A |
-| `FLASHINFER` | Native† | fp16, bf16 | `auto`, `float16`, `bfloat16`, `fp8`, `fp8_e4m3`, `fp8_e5m2` | 16, 32, 64, 128, 256, 512, 1024 | 64, 128, 256, 512 | ❌ | ✅ | ❌ | ✅ | Decoder | 8.x-9.x |
-| `FLASHINFER` | XQA† | fp16, bf16 | `auto`, `float16`, `bfloat16`, `fp8`, `fp8_e4m3`, `fp8_e5m2` | 16, 32, 64, 128, 256, 512, 1024 | 64, 128, 256, 512 | ❌ | ❌ | ❌ | ✅ | Decoder | 9.0 |
-| `FLASHINFER` | trtllm-gen† | fp16, bf16 | `auto`, `float16`, `bfloat16`, `fp8`, `fp8_e4m3`, `fp8_e5m2`, `nvfp4` | 16, 32, 64, 128, 256, 512, 1024 | 64, 128, 256, 512 | ✅ | ✅ | ❌ | ✅ | Decoder | 10.x |
-| `FLASH_ATTN` | FA2* | fp16, bf16 | `auto`, `float16`, `bfloat16` | %16 | Any | ❌ | ✅ | ❌ | ✅ | All | ≥8.0 |
-| `FLASH_ATTN` | FA3* | fp16, bf16 | `auto`, `float16`, `bfloat16`, `fp8`, `fp8_e4m3` | %16 | Any | ✅ | ✅ | ❌ | ✅ | All | 9.x |
-| `FLASH_ATTN` | FA4* | fp16, bf16 | `auto`, `float16`, `bfloat16`, `fp8`, `fp8_e4m3` | %16 | Any | ✅ | ✅ | ❌ | ✅ | All | ≥10.0 |
-| `FLASH_ATTN_DIFFKV` | | fp16, bf16 | `auto` | Any | Any | ❌ | ❌ | ❌ | ✅ | Decoder | Any |
-| `FLEX_ATTENTION` | | fp16, bf16, fp32 | `auto`, `float16`, `bfloat16` | %16 | Any | ❌ | ✅ | ✅ | ❌ | Decoder, Encoder Only | Any |
-| `HPC_ATTN` | | fp16, bf16 | `auto`, `bfloat16`, `fp8_e4m3` | 64 | 128 | ❌ | ❌ | ❌ | ❌ | Decoder | ≥9.0 |
-| `ROCM_AITER_FA` | | fp16, bf16 | `auto`, `float16`, `bfloat16`, `fp8`, `fp8_e4m3`, `fp8_e5m2` | 16, 32 | 64, 128, 256 | ✅ | ✅ | ❌ | ❌ | Decoder | N/A |
-| `ROCM_AITER_UNIFIED_ATTN` | | fp16, bf16 | `auto`, `float16`, `bfloat16`, `fp8`, `fp8_e4m3`, `fp8_e5m2` | %16 | Any | ✅ | ❌ | ✅ | ❌ | All | N/A |
-| `ROCM_ATTN` | | fp16, bf16, fp32 | `auto`, `float16`, `bfloat16`, `fp8`, `fp8_e4m3`, `fp8_e5m2` | %16 | 32, 64, 80, 96, 128, 160, 192, 224, 256 | ❌ | ✅ | ✅ | ❌ | Decoder, Encoder, Encoder Only | N/A |
-| `TRITON_ATTN` | | fp16, bf16, fp32 | `auto`, `float16`, `bfloat16`, `fp8`, `fp8_e4m3`, `fp8_e5m2`, `int4_per_token_head`, `int8_per_token_head`, `fp8_per_token_head` | %16 | Any | ✅ | ✅ | ✅ | ❌ | All | Any |
-| `TRITON_ATTN_DIFFKV` | | fp16, bf16 | `auto`, `bfloat16` | Any | Any | ❌ | ❌ | ❌ | ❌ | Decoder | Any |
-| `TURBOQUANT` | | fp16, bf16 | `turboquant_k8v4`, `turboquant_4bit_nc`, `turboquant_k3v4_nc`, `turboquant_3bit_nc` | 16, 32, 64, 128 | Any | ❌ | ❌ | ❌ | ❌ | Decoder | Any |
+--8<-- "gen:table-standard"
 
 > **†** FlashInfer Native is the regular FlashInfer path. XQA is the SM90 decode path exposed through FlashInfer's TRTLLM decode API. trtllm-gen is used on SM100 and supports sinks. Disable XQA/trtllm-gen via `--attention-config.use_trtllm_attention=0`.
 >
@@ -188,9 +136,7 @@ automatic priority lists above. A lightning indexer scores KV blocks, the
 top-k blocks (plus fixed init/local blocks) are selected, and attention
 attends only to those blocks; index keys live in a separate side cache.
 
-| Backend | Dtypes | KV Dtypes | Block Sizes | Head Sizes | Sink | Non-Causal | MM Prefix | DCP | Attention Types | Compute Cap. |
-| ------- | ------ | --------- | ----------- | ---------- | ---- | ---------- | --------- | --- | --------------- | ------------ |
-| `MINIMAX_M3_SPARSE` | bf16, fp16 | `bfloat16`, `fp8`, `fp8_e4m3`, `fp8_e5m2` | 128 | 128 | ❌ | ❌ | ❌ | ❌ | Decoder | Any |
+--8<-- "gen:table-minimax"
 
 ## MLA (Multi-head Latent Attention) Backends
 
@@ -203,38 +149,20 @@ To explicitly select a prefill backend, use
 Otherwise, the prefill backend is selected automatically at runtime based on
 hardware and configuration.
 
-| Backend | Description | Dtypes | Compute Cap. | Notes |
-| ------- | ----------- | ------ | ------------ | ----- |
-| `FLASH_ATTN`‡ | FlashAttention varlen (FA2/FA3/FA4) | fp16, bf16 | Any | (qk_nope_head_dim=128, qk_rope_head_dim=64, v_head_dim=128) (FA2/FA3/FA4) or (qk_nope_head_dim=64, qk_rope_head_dim=64, v_head_dim=128) (FA2/FA3/FA4) or (qk_nope_head_dim=192, qk_rope_head_dim=64, v_head_dim=256) (FA2/FA3 only) |
-| `TRTLLM_RAGGED` | TensorRT-LLM ragged attention | fp16, bf16 | 10.x | (qk_nope_head_dim=128, qk_rope_head_dim=64, v_head_dim=128) or (qk_nope_head_dim=192, qk_rope_head_dim=64, v_head_dim=256) only |
-| `FLASHINFER` | FlashInfer CUTLASS backend | fp16, bf16 | 10.x | (qk_nope_head_dim=128, qk_rope_head_dim=64, v_head_dim=128) only |
-| `TOKENSPEED_MLA` | | fp16, bf16 | 10.x | (qk_nope_head_dim=128, qk_rope_head_dim=64, v_head_dim=128) only |
+--8<-- "gen:table-mla-prefill"
 
 > **‡** Automatic selection tries FlashAttention first. On Blackwell
 > (SM100), the fallback order is TRT-LLM Ragged, FlashInfer, then
-> TokenSpeed MLA. On other GPUs, only FlashAttention is considered.
+> TokenSpeed MLA; for (qk_nope_head_dim=192, qk_rope_head_dim=64,
+> v_head_dim=256) TRT-LLM Ragged is tried before FlashAttention. On other
+> GPUs, only FlashAttention is considered.
 
 ### Decode Backends
 
 MLA decode backends are selected using the standard
 `-ac.backend=<BACKEND>` argument (e.g., `FLASHMLA`, `TRITON_MLA`).
 
-| Backend | Dtypes | KV Dtypes | Block Sizes | Head Sizes | Sink | Non-Causal | Sparse | MM Prefix | DCP | Attention Types | Compute Cap. |
-| ------- | ------ | --------- | ----------- | ---------- | ---- | ---------- | ------ | --------- | --- | --------------- | ------------ |
-| `CUTLASS_MLA` | fp16, bf16 | `auto`, `float16`, `bfloat16`, `fp8`, `fp8_e4m3` | 128 | Any | ❌ | ❌ | ❌ | ❌ | ✅ | Decoder | 10.x |
-| `FLASHINFER_MLA` | fp16, bf16 | `auto`, `float16`, `bfloat16`, `fp8`, `fp8_e4m3` | 32, 64 | Any | ❌ | ❌ | ❌ | ❌ | ✅ | Decoder | 10.x |
-| `FLASHINFER_MLA_SPARSE` | fp16, bf16 | `auto`, `float16`, `bfloat16`, `fp8`, `fp8_e4m3` | 32, 64 | Any | ❌ | ❌ | ❌ | ❌ | ✅ | Decoder | 10.x |
-| `FLASHINFER_MLA_SPARSE_SM120` | bf16 | `auto`, `fp8`, `fp8_e4m3`, `fp8_ds_mla` | 64, 256 | Any | ❌ | ❌ | ❌ | ❌ | ❌ | Decoder | 12.x |
-| `FLASHMLA` | fp16, bf16 | `auto`, `float16`, `bfloat16`, `fp8`, `fp8_e4m3` | 64 | Any | ❌ | ❌ | ❌ | ❌ | ✅ | Decoder | 9.x-10.x |
-| `FLASHMLA_SPARSE` | bf16 | `auto`, `bfloat16`, `fp8_ds_mla` | 64 | 576 | ❌ | ❌ | ✅ | ❌ | ❌ | Decoder | 9.x-10.x |
-| `FLASH_ATTN_MLA` | fp16, bf16 | `auto`, `float16`, `bfloat16` | %16 | Any | ❌ | ❌ | ❌ | ❌ | ✅ | Decoder | 9.x |
-| `FLASH_ATTN_MLA_SPARSE` | fp16, bf16 | `auto`, `float16`, `bfloat16` | 64 | Any | ❌ | ❌ | ✅ | ❌ | ❌ | Decoder | 9.x |
-| `ROCM_AITER_MLA` | fp16, bf16 | `auto`, `float16`, `bfloat16`, `fp8`, `fp8_e4m3`, `fp8_e5m2` | %1 | Any | ❌ | ❌ | ❌ | ❌ | ❌ | Decoder | N/A |
-| `ROCM_AITER_MLA_SPARSE` | fp16, bf16 | `auto`, `float16`, `bfloat16`, `fp8`, `fp8_e4m3` | 1, 64 | Any | ❌ | ❌ | ✅ | ❌ | ❌ | Decoder | N/A |
-| `ROCM_AITER_TRITON_MLA` | fp16, bf16 | `auto` | Any | Any | ❌ | ❌ | ❌ | ❌ | ❌ | Decoder | N/A |
-| `TOKENSPEED_MLA` | fp16, bf16 | `fp8`, `fp8_e4m3` | 32, 64 | Any | ❌ | ❌ | ❌ | ❌ | ✅ | Decoder | 10.x |
-| `TRITON_MLA` | fp16, bf16 | `auto`, `float16`, `bfloat16`, `fp8`, `fp8_e4m3` | %16 | Any | ❌ | ❌ | ❌ | ❌ | ✅ | Decoder | Any |
-| `XPU_MLA_SPARSE` | fp16, bf16 | `auto`, `float16`, `bfloat16` | Any | 576 | ❌ | ❌ | ✅ | ❌ | ❌ | Decoder | Any |
+--8<-- "gen:table-mla-decode"
 
 ### DeepSeek V4 Decode Backends
 
@@ -245,8 +173,4 @@ pipeline (compressor + SWA + indexer, 256-token blocks, head 512);
 default on NVIDIA is `FLASHINFER_MLA_SPARSE_DSV4` on SM12x and
 `FLASHMLA_SPARSE_DSV4` on other supported CUDA architectures.
 
-| Backend | Dtypes | KV Dtypes | Block Sizes | Head Sizes | Sink | Non-Causal | Sparse | MM Prefix | DCP | Attention Types | Compute Cap. |
-| ------- | ------ | --------- | ----------- | ---------- | ---- | ---------- | ------ | --------- | --- | --------------- | ------------ |
-| `FLASHINFER_MLA_SPARSE_DSV4` | bf16 | `auto`, `bfloat16`, `fp8`, `fp8_e4m3`, `fp8_ds_mla` | 256 | 512 | ✅ | ❌ | ✅ | ❌ | ❌ | Decoder | 10.x, 12.x |
-| `FLASHMLA_SPARSE_DSV4` | bf16 | `auto`, `fp8_ds_mla`, `fp8` | 256 | 512 | ✅ | ❌ | ✅ | ❌ | ❌ | Decoder | 9.x-10.x |
-| `ROCM_FLASHMLA_SPARSE_DSV4` | fp16, bf16 | `auto` | Any | Any | ❌ | ❌ | ❌ | ❌ | ❌ | Decoder | N/A |
+--8<-- "gen:table-mla-v4-decode"

docs/design/cuda_graphs_multimodal.md

@@ -101,7 +101,7 @@ When `mm_encoder_tp_mode="data"`, the manager distributes images across TP ranks
 Following <https://github.com/vllm-project/vllm/pull/35963> (ViT full CUDA graph support for image inference), <https://github.com/vllm-project/vllm/pull/38061> extends the encoder CUDA graph framework to support video inference for Qwen3-VL. Previously, the CUDA graph capture/replay path only handled image inputs (`pixel_values` + `image_grid_thw`). Video inputs use different keys (`pixel_values_videos` + `video_grid_thw`) and require larger `cu_seqlens` buffers because each video item contributes multiple frames (`T` attention sequences). This PR generalizes the protocol and manager to handle both modalities through a single shared graph manager.
 
 !!! note
-    Video CUDA graphs are automatically disabled when EVS (Efficient Video Sampling) pruning is enabled, since EVS makes the token count data-dependent and incompatible with CUDA graph capture.
+    Video CUDA graphs are automatically disabled when video token pruning (EVS or VidCom2) is enabled, since pruning makes the token count data-dependent and incompatible with CUDA graph capture.
 
     Mixed inputs (image+video) per prompt are also supported now.
 
@@ -129,6 +129,7 @@ Models opt-in to encoder CUDA Graphs by implementing the [SupportsEncoderCudaGra
 | `DeepseekOCRForCausalLM` | `DeepSeek-OCR` | ✅︎ | ❌︎ | ✅︎ |
 | `Gemma3ForConditionalGeneration` | `Gemma3` | ✅︎ | ❌︎ | ❌︎ |
 | `Glm4vForConditionalGeneration` | `GLM-4.1V, GLM-4.6V-Flash` | ✅︎ | ✅︎ | ❌︎ |
+| `Gemma4ForConditionalGeneration` | `Gemma-4` | ✅︎ | ✅︎ | ❌︎ |
 | `InternVLChatModel` | `InternVL3.5`, `InternVL3`, `InternVL2.5`, `InternVL2` | ✅︎ | ✅︎ | ❌︎ |
 | `KimiVLForConditionalGeneration` | `Kimi-VL` | ✅︎ | ❌︎ | ❌︎ |
 | `Llama4ForConditionalGeneration` | `Llama 4` | ✅︎ | ❌︎ | ❌︎ |

docs/features/kv_offloading_usage.md

@@ -157,7 +157,7 @@ Object keys follow the same run-configuration digest scheme as the filesystem ti
 
 The P2P tier (`type: "p2p"`) shares completed KV blocks between vLLM instances over RDMA via NIXL. Each instance binds a control socket on `host:port` and exchanges blocks directly with peers — no shared filesystem required.
 
-PYTHONHASHSEED environment variable must be set to the same fixed value on all nodes.
+The `PYTHONHASHSEED` environment variable must be set to the same fixed value (e.g. `"0"`) on all nodes so that block content hashes match across instances (see [Cross-Process Sharing](#cross-process-sharing)). This is enforced: a P2P instance started without `PYTHONHASHSEED` set fails at startup, and each peer's value is verified during the connect handshake — a peer advertising a different `PYTHONHASHSEED` is rejected.
 
 | Key | Required | Default | Notes |
 | --- | --- | --- | --- |
@@ -176,6 +176,71 @@ Rather than embedding `host`/`port` in each `secondary_tiers` entry, set them on
 - `VLLM_P2P_SIDE_CHANNEL_HOST` (default `localhost`): address the P2P control socket binds to. It is used **verbatim** as both the bind address and the identity peers dial back — there is no auto-detection (this mirrors `VLLM_NIXL_SIDE_CHANNEL_HOST`). The default binds the loopback interface only, so peers on another host cannot reach it. **For any cross-host P2P deployment you must set this explicitly to the node's routable IP** (e.g. the pod IP) before launching `vllm serve` — otherwise remote peers will fail to connect. The NIXL agent name is a separate per-process identifier, so peers sharing a `host:port` never collide.
 - `VLLM_P2P_SIDE_CHANNEL_PORT` (default `5710`): base port for the P2P control socket. The port actually bound is `VLLM_P2P_SIDE_CHANNEL_PORT + data_parallel_index` — one socket per DP replica, matching NIXL (for DP=1 the offset is 0). The peer's port is passed as `remote_port` in `kv_transfer_params`; the router/EPP that selects the DP rank (e.g. via the `X-data-parallel-rank` header) computes `remote_port = base + rank`. The DP-index offset separates replicas *within* one deployment; two co-located *deployments* (a prefiller and a decoder on the same host) still need distinct base ports (e.g. decoder base `5711`) to avoid a bind collision.
 
+#### Orchestration-Layer Protocol
+
+The P2P tier does not decide *which* peer to pull from — that is the orchestration layer's job (the router/EPP and its scheduler). The orchestrator drives every transfer through a request's `kv_transfer_params` dict: it picks the request's role, allocates a unique transaction ID, and supplies the remote peer's address. All block lookup, hash matching, and NIXL transfer happen at the tier level below; the orchestrator only sets the correct role keys and enforces the allowed combinations.
+
+Every vLLM instance is a symmetric **peer**. Per request it acts as a **consumer** (pulls KV blocks from a remote peer's CPU cache instead of computing locally) or a **producer** (serves blocks from its own CPU cache to remote consumers) — or both, on the same session, for different requests. Roles are chosen per request by the keys below; there are no fixed prefiller/decoder processes.
+
+Three role keys are defined, each mapping to a sub-dict. All are optional; a request with none of them uses the tier only as a local CPU cache.
+
+Each key names the **remote counterpart** this peer transfers with (not this
+peer's own role), so the name reads as "the remote ___ I transfer with".
+
+| Key | Set on | Value fields | Meaning |
+| --- | --- | --- | --- |
+| `remote_decoder` | prefill producer request | `kv_request_id` | Peer computes KV and keeps it available in CPU cache for the remote decoder to pull. |
+| `remote_prefiller` | decode consumer request | `kv_request_id`, `remote_host`, `remote_port` | Peer pulls KV from the remote prefiller at the given address (classic P/D disaggregation). |
+| `remote_kv_source` | P2P consumer request | `kv_request_id`, `remote_host`, `remote_port` | Peer looks up and pulls whatever blocks the remote source currently holds in CPU cache. |
+
+Field semantics:
+
+- `kv_request_id` (str): unique transaction ID allocated by the orchestrator and pushed to every peer involved in the transfer; used to correlate the lookup, fetch, and transfer-done messages. The producer is implicit — it serves whatever block hashes it currently holds in its CPU cache for that ID.
+- `remote_host` (str): IP/hostname of the remote peer's control socket to query. Must be the peer's routable node IP (see [Environment Variables](#environment-variables)).
+- `remote_port` (int): the peer's bound control-socket port, i.e. `base + data_parallel_index` for the selected DP rank.
+
+Allowed and forbidden combinations:
+
+- **`remote_decoder` + `remote_kv_source`** is the only legal multi-key combination: a prefill producer may *also* act as a P2P consumer for the same request — skipping prefix prefill by pulling cached blocks from a source while still keeping its own computed blocks available for a downstream decoder.
+- Forbidden: `remote_prefiller` + `remote_decoder` (contradictory roles), `remote_prefiller` + `remote_kv_source` (two competing fetch sources), and all three together.
+
+Minimal examples (values that would appear in the request's `kv_transfer_params`):
+
+```python
+# Prefill producer — compute and keep KV for a remote decoder to pull
+kv_transfer_params = {"remote_decoder": {"kv_request_id": "<unique-transfer-id>"}}
+
+# Decode consumer — pull KV from a specific prefiller (classic P/D)
+kv_transfer_params = {
+    "remote_prefiller": {
+        "kv_request_id": "<unique-transfer-id>",
+        "remote_host": "<prefiller-node-ip>",
+        "remote_port": 5710,
+    }
+}
+
+# P2P consumer — pull whatever the source already has cached
+kv_transfer_params = {
+    "remote_kv_source": {
+        "kv_request_id": "<unique-transfer-id>",
+        "remote_host": "<source-node-ip>",
+        "remote_port": 5710,
+    }
+}
+```
+
+Runtime handshake for a P2P (or P/D) pull, once the orchestrator has set the keys above:
+
+1. Both peers already have listener threads on their control sockets (see [Environment Variables](#environment-variables)).
+2. **Lookup.** The consumer's tiering manager does per-block lookups; in P2P mode the tier returns `None` and registers the key. At `on_schedule_end` the consumer sends one **`LookupMsg`** (`kv_request_id` + block hashes) to the peer, per request, per step.
+3. The producer matches those hashes against its local CPU cache and replies with a **`LookupRespMsg`** carrying the hit block hashes.
+4. **Resolve.** Retried lookups now return hit / miss / in-flight. The consumer calls `submit_load` for hits only, allocating CPU slots only for hits.
+5. The consumer sends a **`FetchMsg`** (`kv_request_id`, block hashes, destination block indexes).
+6. The producer performs the **NIXL WRITE** transfer and sends **`TransferDone`** with a success status.
+7. On `get_finished`, hits are loaded into GPU as ordinary cache hits; misses are recomputed by the engine.
+
+In classic **P/D mode** (`remote_prefiller` set, no `remote_kv_source`), the lookup phase (steps 2–4) is skipped: the decode consumer assumes the prefiller holds all of the request's blocks, so every block `lookup()` returns an immediate hit and the consumer jumps straight to the **`FetchMsg`** in step 5. The `LookupMsg`/`LookupRespMsg` round-trip only happens in P2P mode, where the consumer does not know in advance which blocks the peer has cached.
+
 ## Tuning Tips
 
 - `cpu_bytes_to_use`: a bigger CPU tier means fewer trips to slower secondary tiers and a higher hit rate. The value is total across all workers, not per-worker. Leave headroom for the rest of the host workload.

docs/features/multimodal_inputs.md

@@ -350,6 +350,31 @@ Instead of NumPy arrays, you can also pass `'torch.Tensor'` instances, as shown
 
 Full example: [examples/generate/multimodal/vision_language_offline.py](../../examples/generate/multimodal/vision_language_offline.py)
 
+#### Video Token Pruning
+
+For supported models, vLLM can prune video tokens after the vision encoder to
+reduce prefill time and KV cache usage, at some cost in accuracy. Set
+`--video-pruning-rate <q>` to prune the fraction `q` of video tokens from each
+video, and `--video-pruning-method` to choose the training-free algorithm:
+
+- **`evs`** (Efficient Video Sampling, default): drops the tokens with the
+  lowest temporal dissimilarity to the previous frame. The first frame is
+  always fully retained.
+- **`vidcom2`** (Video Compression Commander): scores tokens by similarity to
+  video-level and frame-level feature centers and gives distinctive frames a
+  larger share of the budget. At least one token per frame is retained.
+
+```bash
+vllm serve Qwen/Qwen3-VL-8B-Instruct \
+    --video-pruning-rate 0.75 --video-pruning-method vidcom2
+```
+
+!!! note
+    `evs` is supported by all models implementing multimodal pruning;
+    `vidcom2` is currently supported by Qwen3-VL only. Unsupported combinations
+    are rejected at startup. Enabling video pruning also disables encoder CUDA
+    graphs, since the retained token count becomes data-dependent.
+
 ### Audio Inputs
 
 You can pass a tuple `(array, sampling_rate)` to the `'audio'` field of the multi-modal dictionary.
@@ -818,16 +843,18 @@ Full example: [examples/generate/multimodal/openai_chat_completion_client_for_mu
 
 #### Video Decoding Backend
 
-vLLM decodes video bytes into frames using a selectable decoding backend. Three
+vLLM decodes video bytes into frames using a selectable decoding backend. Five
 backends are supported:
 
 - `opencv` (default): OpenCV-based decoder.
 - `pyav`: PyAV decoder.
 - `torchcodec`: TorchCodec (PyTorch-native) decoder.
+- `pynvvideocodec`: NVIDIA NVDEC-based decoder.
+- `deepstream`: NVIDIA DeepStream NVDEC-based decoder.
 
-All three backends are ultimately backed by FFmpeg. `torchcodec` lets
-you choose which FFmpeg version is used while `opencv` and `pyav` rely on
-whichever FFmpeg build they were linked against.
+The CPU backends are backed by FFmpeg. `torchcodec` lets you choose which FFmpeg
+version is used while `opencv` and `pyav` rely on whichever FFmpeg build they
+were linked against.
 
 Select the backend by passing the `backend` parameter via `--media-io-kwargs`:
 
@@ -854,6 +881,21 @@ vllm serve Qwen/Qwen3-VL-30B-A3B-Instruct \
   --media-io-kwargs '{"video": {"backend": "torchcodec", "seek_mode": "approximate", "num_ffmpeg_threads": 4}}'
 ```
 
+**PyNvVideoCodec-specific parameters:**
+
+- `hw_decoders`: Maximum number of concurrent hardware decoder slots retained
+  by each API server process. It must be a positive integer and defaults to `2`,
+  which is the recommended starting point for concurrent video workloads.
+  Because vLLM reserves GPU memory for these slots at startup, this value cannot
+  be overridden per request. Benchmark before increasing it because each
+  additional slot increases the GPU memory reservation.
+
+```bash
+# Example: explicitly use the recommended 2 hardware decoders
+vllm serve Qwen/Qwen3-VL-30B-A3B-Instruct \
+  --media-io-kwargs '{"video": {"backend": "pynvvideocodec", "hw_decoders": 2}}'
+```
+
 #### Video Frame Recovery
 
 For improved robustness when processing potentially corrupted or truncated video files, vLLM supports optional frame recovery using a dynamic window forward-scan approach. When enabled, if a target frame fails to load during sequential reading, the next successfully grabbed frame (before the next target frame) will be used in its place.

docs/features/quantization/modelopt.md

@@ -19,6 +19,20 @@ following `quantization.quant_algo` values:
 - `NVFP4`: ModelOpt NVFP4 checkpoints (use `quantization="modelopt_fp4"`).
 - `MXFP8`: ModelOpt MXFP8 checkpoints (use `quantization="modelopt_mxfp8"`).
 
+!!! note
+    For NVFP4 checkpoints, vLLM selects a GEMM kernel automatically at load
+    time from the backends available on the current platform (CUTLASS,
+    FlashInfer, Marlin, and others). On GPUs without a supported native FP4
+    GEMM kernel, vLLM falls back to weight-only (W4A16) execution via Marlin
+    and logs a warning; this may reduce throughput for compute-heavy
+    workloads. Use `--linear-backend` to override the automatic selection
+    (this replaces the deprecated `VLLM_NVFP4_GEMM_BACKEND` environment
+    variable). Values relevant to NVFP4 include `cutlass`,
+    `flashinfer_cutlass`, `flashinfer_trtllm`, `flashinfer_cudnn`, and
+    `marlin`; the full list is documented under `KernelConfig` on the
+    [Engine Arguments](../../configuration/engine_args.md) page and shown by
+    `vllm serve --help=KernelConfig`.
+
 ## Quantizing HuggingFace Models with PTQ
 
 You can quantize HuggingFace models using the example scripts provided in the Model Optimizer repository. The primary script for LLM PTQ is typically found within the `examples/llm_ptq` directory.

docs/mkdocs/gen_files/generate_argparse.py

@@ -2,6 +2,7 @@
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 import importlib.metadata
 import importlib.util
+import inspect
 import logging
 import sys
 import textwrap
@@ -10,17 +11,21 @@ from argparse import SUPPRESS, Action, HelpFormatter
 from collections.abc import Callable, Iterable
 from importlib.machinery import ModuleSpec
 from pathlib import Path
-from typing import TYPE_CHECKING, Literal
+from typing import TYPE_CHECKING
 from unittest.mock import MagicMock, patch
 
+import mkdocs_gen_files
+import regex as re
 from pydantic_core import core_schema
 
 logger = logging.getLogger("mkdocs")
 
 ROOT_DIR = Path(__file__).parent.parent.parent.parent
-ARGPARSE_DOC_DIR = ROOT_DIR / "docs/generated/argparse"
 
 sys.path.insert(0, str(ROOT_DIR))
+sys.path.insert(0, str(Path(__file__).parent))
+
+from generated_content import fill_markers  # noqa: E402
 
 
 def mock_if_no_torch(mock_module: str, mock: MagicMock):
@@ -132,8 +137,8 @@ def auto_mock(module_name: str, attr: str, max_mocks: int = 100):
 
 
 bench_latency = auto_mock("vllm.benchmarks", "latency")
-bench_mm_processor = auto_mock("vllm.benchmarks", "mm_processor")
 bench_serve = auto_mock("vllm.benchmarks", "serve")
+bench_startup = auto_mock("vllm.benchmarks", "startup")
 bench_sweep_plot = auto_mock("vllm.benchmarks.sweep.plot", "SweepPlotArgs")
 bench_sweep_plot_pareto = auto_mock(
     "vllm.benchmarks.sweep.plot_pareto", "SweepPlotParetoArgs"
@@ -142,12 +147,28 @@ bench_sweep_serve = auto_mock("vllm.benchmarks.sweep.serve", "SweepServeArgs")
 bench_sweep_serve_workload = auto_mock(
     "vllm.benchmarks.sweep.serve_workload", "SweepServeWorkloadArgs"
 )
+bench_sweep_startup = auto_mock("vllm.benchmarks.sweep.startup", "SweepStartupArgs")
 bench_throughput = auto_mock("vllm.benchmarks", "throughput")
 AsyncEngineArgs = auto_mock("vllm.engine.arg_utils", "AsyncEngineArgs")
 EngineArgs = auto_mock("vllm.engine.arg_utils", "EngineArgs")
 ChatCommand = auto_mock("vllm.entrypoints.cli.openai", "ChatCommand")
 CompleteCommand = auto_mock("vllm.entrypoints.cli.openai", "CompleteCommand")
+BenchmarkSubcommand = auto_mock(
+    "vllm.entrypoints.cli.benchmark.main", "BenchmarkSubcommand"
+)
+import_bench_subcommands = auto_mock(
+    "vllm.entrypoints.cli.benchmark.main", "_import_bench_subcommand_modules"
+)
+BenchmarkSubcommandBase = auto_mock(
+    "vllm.entrypoints.cli.benchmark.base", "BenchmarkSubcommandBase"
+)
+BenchmarkMMProcessorSubcommand = auto_mock(
+    "vllm.entrypoints.cli.benchmark.mm_processor", "BenchmarkMMProcessorSubcommand"
+)
+LaunchSubcommandBase = auto_mock("vllm.entrypoints.cli.launch", "LaunchSubcommandBase")
+launch_description = auto_mock("vllm.entrypoints.cli.launch", "DESCRIPTION")
 RenderSubcommand = auto_mock("vllm.entrypoints.cli.launch", "RenderSubcommand")
+sweep_subcommands = auto_mock("vllm.benchmarks.sweep.cli", "SUBCOMMANDS")
 openai_cli_args = auto_mock("vllm.entrypoints.openai", "cli_args")
 openai_run_batch = auto_mock("vllm.entrypoints.openai", "run_batch")
 
@@ -179,7 +200,7 @@ class MarkdownFormatter(HelpFormatter):
 
     def add_text(self, text: str):
         if text:
-            self._markdown_output.append(f"{text.strip()}\n\n")
+            self._markdown_output.append(f"{inspect.cleandoc(text)}\n\n")
 
     def add_usage(self, usage, actions, groups, prefix=None):
         pass
@@ -241,49 +262,163 @@ def create_parser(add_cli_args, **kwargs) -> FlexibleArgumentParser:
     return _parser or parser
 
 
-def on_startup(command: Literal["build", "gh-deploy", "serve"], dirty: bool):
-    logger.info("Generating argparse documentation")
-    logger.debug("Root directory: %s", ROOT_DIR.resolve())
-    logger.debug("Output directory: %s", ARGPARSE_DOC_DIR.resolve())
-
-    # Create the ARGPARSE_DOC_DIR if it doesn't exist
-    if not ARGPARSE_DOC_DIR.exists():
-        ARGPARSE_DOC_DIR.mkdir(parents=True)
-
-    # Create parsers to document
-    parsers = {
-        # Engine args
-        "engine_args": create_parser(EngineArgs.add_cli_args),
-        "async_engine_args": create_parser(
-            AsyncEngineArgs.add_cli_args, async_args_only=True
-        ),
-        # CLI
-        "serve": create_parser(openai_cli_args.make_arg_parser),
-        "chat": create_parser(ChatCommand.add_cli_args),
-        "complete": create_parser(CompleteCommand.add_cli_args),
-        "launch_render": create_parser(RenderSubcommand.add_cli_args),
-        "run-batch": create_parser(openai_run_batch.make_arg_parser),
-        # Benchmark CLI
-        "bench_latency": create_parser(bench_latency.add_cli_args),
-        "bench_mm_processor": create_parser(bench_mm_processor.add_cli_args),
-        "bench_serve": create_parser(bench_serve.add_cli_args),
-        "bench_sweep_plot": create_parser(bench_sweep_plot.add_cli_args),
-        "bench_sweep_plot_pareto": create_parser(bench_sweep_plot_pareto.add_cli_args),
-        "bench_sweep_serve": create_parser(bench_sweep_serve.add_cli_args),
-        "bench_sweep_serve_workload": create_parser(
-            bench_sweep_serve_workload.add_cli_args
-        ),
-        "bench_throughput": create_parser(bench_throughput.add_cli_args),
-    }
-
-    # Generate documentation for each parser
-    for stem, parser in parsers.items():
-        doc_path = ARGPARSE_DOC_DIR / f"{stem}.inc.md"
-        # Specify encoding for building on Windows
-        with open(doc_path, "w", encoding="utf-8") as f:
-            f.write(super(type(parser), parser).format_help())
-        logger.info("Argparse generated: %s", doc_path.relative_to(ROOT_DIR))
+def format_help(parser: FlexibleArgumentParser) -> str:
+    """Format a parser's help as markdown using `MarkdownFormatter`."""
+    return super(type(parser), parser).format_help()
 
 
-if __name__ == "__main__":
-    on_startup("build", False)
+# Absolute docs URLs are kept in the help text because they are useful in the
+# terminal. Wrap them as markdown links so the `url_schemes` hook can rewrite
+# them into doc-relative links / cross-references at render time.
+_DOCS_URL = re.compile(r"https://docs\.vllm\.ai/en/[^/\s]+/[^\s)>]+")
+
+
+def linkify_docs_urls(text: str) -> str:
+    """Wrap bare docs.vllm.ai URLs in help text as markdown links."""
+    return _DOCS_URL.sub(lambda m: f"[{m.group()}]({m.group()})", text)
+
+
+logger.info("Generating argparse documentation")
+logger.debug("Root directory: %s", ROOT_DIR.resolve())
+
+# The JSON tip is always rendered immediately before generated argument content,
+# and the generator is its only consumer, so it lives here rather than in a
+# separate snippet file. (The runtime terminal equivalent is
+# `FlexibleArgumentParser._json_tip` in vllm/utils/argparse_utils.py.)
+JSON_TIP = """## JSON CLI Arguments
+
+When passing JSON CLI arguments, the following sets of arguments are equivalent:
+
+- `--json-arg '{"key1": "value1", "key2": {"key3": "value2"}}'`
+- `--json-arg.key1 value1 --json-arg.key2.key3 value2`
+
+Additionally, list elements can be passed individually using `+`:
+
+- `--json-arg '{"key4": ["value3", "value4", "value5"]}'`
+- `--json-arg.key4+ value3 --json-arg.key4+='value4,value5'`
+
+"""
+
+# Argument sections filled into `gen:` markers on handwritten pages
+engine_args = create_parser(EngineArgs.add_cli_args)
+async_engine_args = create_parser(AsyncEngineArgs.add_cli_args, async_args_only=True)
+fill_markers(
+    "configuration/engine_args.md",
+    {
+        "engine-args": (
+            f"{JSON_TIP}## `EngineArgs`\n\n"
+            f"{linkify_docs_urls(format_help(engine_args))}"
+            f"## `AsyncEngineArgs`\n\n"
+            f"{linkify_docs_urls(format_help(async_engine_args))}"
+        )
+    },
+)
+
+# CLI reference pages generated entirely from their parser: page -> (parser, JSON tip)
+pages = {
+    "cli/serve.md": (create_parser(openai_cli_args.make_arg_parser), True),
+    "cli/chat.md": (create_parser(ChatCommand.add_cli_args), False),
+    "cli/complete.md": (create_parser(CompleteCommand.add_cli_args), False),
+    "cli/run-batch.md": (create_parser(openai_run_batch.make_arg_parser), True),
+    "cli/launch/render.md": (create_parser(RenderSubcommand.add_cli_args), True),
+    "cli/bench/latency.md": (create_parser(bench_latency.add_cli_args), True),
+    # URL kept as `mm_processor` for back-compat; command name is `mm-processor`
+    "cli/bench/mm_processor.md": (
+        create_parser(BenchmarkMMProcessorSubcommand.add_cli_args),
+        True,
+    ),
+    "cli/bench/serve.md": (create_parser(bench_serve.add_cli_args), True),
+    "cli/bench/startup.md": (create_parser(bench_startup.add_cli_args), True),
+    "cli/bench/throughput.md": (create_parser(bench_throughput.add_cli_args), True),
+    "cli/bench/sweep/plot.md": (create_parser(bench_sweep_plot.add_cli_args), True),
+    "cli/bench/sweep/plot_pareto.md": (
+        create_parser(bench_sweep_plot_pareto.add_cli_args),
+        True,
+    ),
+    "cli/bench/sweep/serve.md": (create_parser(bench_sweep_serve.add_cli_args), True),
+    "cli/bench/sweep/serve_workload.md": (
+        create_parser(bench_sweep_serve_workload.add_cli_args),
+        True,
+    ),
+    "cli/bench/sweep/startup.md": (
+        create_parser(bench_sweep_startup.add_cli_args),
+        True,
+    ),
+}
+
+# Command name for pages whose file stem differs (URL kept for back-compat).
+COMMAND_NAMES = {"cli/bench/mm_processor.md": "mm-processor"}
+
+for doc_path, (parser, json_tip) in pages.items():
+    segments = Path(doc_path).relative_to("cli").with_suffix("").parts
+    label = COMMAND_NAMES.get(doc_path, segments[-1])
+    command = " ".join([*segments[:-1], label])
+    # `title` frontmatter keeps the nav label to just this command's segment,
+    # while the H1 stays the full `vllm ...` command for the page heading.
+    content = f"---\ntitle: {label}\n---\n\n"
+    content += f"# vllm {command}\n\n"
+    if parser.description:
+        content += f"## Overview\n\n{parser.description}\n\n"
+        # Rendered above instead of at the top of the Arguments section
+        parser.description = None
+    if json_tip:
+        content += JSON_TIP
+    content += f"## Arguments\n\n{linkify_docs_urls(format_help(parser))}"
+    with mkdocs_gen_files.open(doc_path, "w") as f:
+        f.write(content)
+    logger.debug("CLI reference generated: %s", doc_path)
+
+logger.info("Total argparse docs generated: %d", len(pages) + 2)
+
+
+# --- Bare subcommand (group) pages -------------------------------------------
+# Mirror `vllm <group> --help`: an overview plus a table of child subcommands,
+# each linked to its reference page. Children are read from the CLI registries
+# so the listing can never drift from the actual subcommands. Each page is the
+# `README.md` of its command directory so it becomes that section's index and is
+# picked up by the existing nav globs.
+import_bench_subcommands()  # populate BenchmarkSubcommandBase.__subclasses__()
+bench_subcommands = BenchmarkSubcommandBase.__subclasses__()
+bench_children = [(cmd.name, cmd.help) for cmd in bench_subcommands]
+
+groups = {
+    "cli/bench/README.md": (BenchmarkSubcommand.help, bench_children),
+    "cli/launch/README.md": (
+        launch_description,
+        [(cmd.name, cmd.help) for cmd in LaunchSubcommandBase.__subclasses__()],
+    ),
+    "cli/bench/sweep/README.md": (
+        dict(bench_children).get("sweep"),
+        [(args.parser_name, args.parser_help) for args, _ in sweep_subcommands],
+    ),
+}
+
+# Doc paths that exist, so we only link a child that has a reference page.
+existing_pages = set(pages) | set(groups)
+
+
+def child_link(group_doc: str, name: str) -> str | None:
+    group_dir = Path(group_doc).parent  # cli/bench/README.md -> cli/bench
+    for stem in (name, name.replace("-", "_")):
+        # A leaf page (bench/latency.md) or a nested group index (sweep/README.md)
+        for candidate in (group_dir / f"{stem}.md", group_dir / stem / "README.md"):
+            if candidate.as_posix() in existing_pages:
+                return candidate.relative_to(group_dir).as_posix()
+    return None
+
+
+for doc_path, (overview, children) in groups.items():
+    title = "vllm " + Path(doc_path).parent.relative_to("cli").as_posix()
+    lines = [f"# {title.replace('/', ' ')}", ""]
+    if overview:
+        lines += ["## Overview", "", overview.strip(), ""]
+    lines += ["## Subcommands", "", "| Command | Description |", "| --- | --- |"]
+    for name, summary in children:
+        link = child_link(doc_path, name)
+        command = f"[`{name}`]({link})" if link else f"`{name}`"
+        lines.append(f"| {command} | {(summary or '').strip()} |")
+    with mkdocs_gen_files.open(doc_path, "w") as f:
+        f.write("\n".join(lines) + "\n")
+    logger.debug("CLI group reference generated: %s", doc_path)
+
+logger.info("CLI group reference pages generated: %d", len(groups))

docs/mkdocs/gen_files/generate_attention_backends.py

@@ -9,33 +9,28 @@ based on the checks in AttentionBackend.validate_configuration().
 
 This approach avoids requiring CUDA/ROCm/GPU libraries to be installed.
 
-When used as a pre-commit hook, this script receives filenames as arguments
-and only runs the check if any of the relevant files were modified.
+It runs as an mkdocs-gen-files script, so the page is generated at docs build
+time rather than being committed to the repository.
 """
 
-import argparse
 import ast
-import fnmatch
+import logging
 import sys
 from collections.abc import Callable
 from pathlib import Path
 from typing import Any
 
+sys.path.insert(0, str(Path(__file__).parent))
+
+from generated_content import fill_markers  # noqa: E402
+
+logger = logging.getLogger("mkdocs")
+
 # ---------------------------------------------------------------------------
 # Constants and file paths
 # ---------------------------------------------------------------------------
 
-REPO_ROOT = Path(__file__).parent.parent.parent
-
-RELEVANT_PATTERNS = [
-    "vllm/v1/attention/backends/*.py",
-    "vllm/v1/attention/backends/**/*.py",
-    "vllm/models/minimax_m3/common/sparse_attention.py",
-    "vllm/model_executor/layers/attention/mla_attention.py",
-    "vllm/platforms/cuda.py",
-    "tools/pre_commit/generate_attention_backend_docs.py",
-    "docs/design/attention_backends.md",
-]
+REPO_ROOT = Path(__file__).parent.parent.parent.parent
 
 BACKENDS_DIR = REPO_ROOT / "vllm" / "v1" / "attention" / "backends"
 REGISTRY_FILE = BACKENDS_DIR / "registry.py"
@@ -55,19 +50,6 @@ BACKEND_KV_DTYPE_EXCLUDES: dict[str, set[str]] = {
 }
 
 
-def is_relevant_file(filepath: str) -> bool:
-    """Check if a file matches any of the relevant patterns."""
-    path = Path(filepath)
-    if path.is_absolute():
-        try:
-            path = path.relative_to(REPO_ROOT)
-        except ValueError:
-            return False
-    path_str = str(path)
-
-    return any(fnmatch.fnmatch(path_str, pattern) for pattern in RELEVANT_PATTERNS)
-
-
 MLA_PREFILL_DIR = BACKENDS_DIR / "mla" / "prefill"
 MLA_PREFILL_REGISTRY_FILE = MLA_PREFILL_DIR / "registry.py"
 MLA_PREFILL_SELECTOR_FILE = MLA_PREFILL_DIR / "selector.py"
@@ -960,7 +942,7 @@ def analyze_backend(backend_name: str, class_path: str) -> dict[str, Any] | None
     try:
         tree = ast.parse(file_path.read_text())
     except Exception as e:
-        print(f"  Warning: Could not parse {file_path}: {e}", file=sys.stderr)
+        logger.warning("Could not parse %s: %s", file_path, e)
         return None
 
     class_name = class_path.rsplit(".", 1)[1]
@@ -1657,113 +1639,12 @@ def _render_table(
     return lines
 
 
-def generate_markdown_table(
-    backends: list[dict[str, Any]], title: str, is_mla_table: bool = False
-) -> str:
-    """Generate a titled markdown table from backend info."""
-    if not backends:
-        return f"## {title}\n\nNo backends found.\n"
-    has_versions = any(b.get("version") for b in backends)
-    columns = _build_columns(is_mla_table, has_versions)
-    lines = [f"## {title}", ""]
-    lines.extend(_render_table(columns, backends))
-    lines.append("")
-    return "\n".join(lines)
-
-
-# ---------------------------------------------------------------------------
-# Markdown section generators (usage, priority, legend, MLA)
-# ---------------------------------------------------------------------------
-
-
-def generate_usage_section() -> str:
-    """Generate the usage documentation section."""
-    return """## Setting the Attention Backend
-
-### Command Line
-
-There are two ways to specify the backend from the command line:
-
-**Option 1: Using `--attention-backend` (simple)**
-
-```bash
-vllm serve <model> --attention-backend FLASH_ATTN
-```
-
-**Option 2: Using `--attention-config.backend` / `-ac.backend` (structured config)**
-
-```bash
-# Dot notation
-vllm serve <model> --attention-config.backend FLASH_ATTN
-vllm serve <model> -ac.backend FLASH_ATTN
-
-# JSON format
-vllm serve <model> --attention-config '{"backend": "FLASH_ATTN"}'
-vllm serve <model> -ac '{"backend": "FLASH_ATTN"}'
-```
-
-> **Note:** `--attention-backend` and `--attention-config.backend` are mutually
-> exclusive. Use one or the other, not both.
-
-### Python API
-
-Use `AttentionConfig` with the `LLM` class:
-
-```python
-from vllm import LLM
-from vllm.config import AttentionConfig
-from vllm.v1.attention.backends.registry import AttentionBackendEnum
-
-# Method 1: Using AttentionConfig with enum
-llm = LLM(
-    model="Qwen/Qwen3-0.6B",
-    attention_config=AttentionConfig(backend=AttentionBackendEnum.FLASH_ATTN),
-)
-
-# Method 2: Using attention_backend parameter with string
-llm = LLM(
-    model="Qwen/Qwen3-0.6B",
-    attention_backend="FLASH_ATTN",
-)
-```
-
-## Backend Selection Behavior
-
-### Manual Selection
-
-When you explicitly set a backend via `--attention-backend` or `AttentionConfig`:
-
-1. The backend is **validated** against your configuration (model dtype, head
-   size, compute capability, etc.)
-2. If the backend **doesn't support** your configuration, an error is raised
-   with the specific reason
-3. If valid, the backend is used
-
-Example error when selecting an incompatible backend:
-
-```text
-ValueError: Selected backend FLASHMLA is not valid for this configuration.
-Reason: ['compute capability not supported']
-```
-
-### Automatic Selection
-
-When no backend is specified (the default):
-
-1. vLLM iterates through backends in **priority order** (see tables below)
-2. Each backend is validated against your configuration
-3. The **first compatible backend** is selected
-4. If no backend is compatible, an error is raised listing all backends and
-   their incompatibility reasons
-"""
-
-
 def _priority_table(
     title: str,
     backends: list[str],
     annotations: dict[str, str] | None = None,
 ) -> list[str]:
-    """Generate a priority table for a list of backends."""
+    """Render a priority table for a list of backends."""
 
     def _fmt(b: str) -> str:
         suffix = annotations.get(b, "") if annotations else ""
@@ -1779,102 +1660,38 @@ def _priority_table(
     ]
 
 
-def generate_priority_section(priorities: dict[str, list[str]]) -> str:
-    """Generate the priority ranking section."""
-    lines = [
-        "## Backend Priority (CUDA)",
-        "",
-        "When no backend is explicitly selected, vLLM chooses the first",
-        "compatible backend from these priority-ordered lists.",
-        "",
-        "Priority is **1 = highest** (tried first).",
-        "",
-        "### Standard Attention (MHA, MQA, GQA)",
-        "",
-    ]
-
-    sm100 = "Blackwell (SM 10.x)"
-    ampere = "Ampere/Hopper (SM 8.x-9.x)"
-
-    if "standard_sm100" in priorities:
-        lines.extend(_priority_table(sm100, priorities["standard_sm100"]))
-    if "standard_default" in priorities:
-        lines.extend(_priority_table(ampere, priorities["standard_default"]))
-
-    lines.extend(["### MLA Attention (DeepSeek-style)", ""])
-
-    mla_sm100_annotations = {
-        "FLASHINFER_MLA_SPARSE": "**\\***",
-    }
-    if "mla_sm100" in priorities:
-        lines.extend(
-            _priority_table(sm100, priorities["mla_sm100"], mla_sm100_annotations)
-        )
-    if "mla_default" in priorities:
-        lines.extend(_priority_table(ampere, priorities["mla_default"]))
-
-    if "mla_sm100" in priorities:
-        lines.append(
-            "> **\\*** For sparse MLA, FP8 KV cache always prefers "
-            "`FLASHINFER_MLA_SPARSE`. With BF16 KV cache, `FLASHINFER_MLA_SPARSE` "
-            "is preferred for low query-head counts (<= 16), while "
-            "`FLASHMLA_SPARSE` is preferred otherwise."
-        )
-        lines.append(">")
-
-    lines.append(
-        "> **Note:** ROCm and CPU platforms have their own selection logic. "
-        "See the platform-specific documentation for details."
-    )
-    lines.append("")
-
-    return "\n".join(lines)
+_SM100 = "Blackwell (SM 10.x)"
+_AMPERE = "Ampere/Hopper (SM 8.x-9.x)"
 
 
-def generate_legend() -> str:
-    """Generate a legend explaining the table columns."""
-    return """## Legend
-
-| Column | Description |
-| ------ | ----------- |
-| **Dtypes** | Supported model data types (fp16, bf16, fp32) |
-| **KV Dtypes** | Supported KV cache data types (`auto`, `fp8`, `fp8_e4m3`, etc.) |
-| **Block Sizes** | Supported KV cache block sizes (%N means multiples of N) |
-| **Head Sizes** | Supported attention head sizes |
-| **Sink** | Attention sink support (for StreamingLLM) |
-| **Non-Causal** | Non-causal (bidirectional) attention support for decoder models |
-| **Sparse** | Sparse attention support (MLA only) |
-| **MM Prefix** | Multimodal prefix full attention support |
-| **DCP** | Decode Context Parallelism support (`--decode-context-parallel-size`) |
-| **Attention Types** | Supported attention patterns (Decoder, Encoder, Enc-Dec) |
-| **Compute Cap.** | Required CUDA compute capability (N/A for non-CUDA backends) |
-
-**Symbols:** ✅ = Supported, ❌ = Not supported
-"""
-
-
-def generate_mla_section(
-    prefill_backends: list[dict[str, Any]],
-    decode_backends: list[dict[str, Any]],
-    v4_decode_backends: list[dict[str, Any]] | None = None,
+def _priority_block(
+    priorities: dict[str, list[str]],
+    sm100_key: str,
+    default_key: str,
+    sm100_annotations: dict[str, str] | None = None,
 ) -> str:
-    """Generate the complete MLA section with prefill and decode tables."""
+    """Render whichever priority tables exist for one attention category."""
+    lines: list[str] = []
+    if sm100_key in priorities:
+        lines += _priority_table(_SM100, priorities[sm100_key], sm100_annotations)
+    if default_key in priorities:
+        lines += _priority_table(_AMPERE, priorities[default_key])
+    return "\n".join(lines).strip()
+
+
+def _feature_table(backends: list[dict[str, Any]], is_mla: bool) -> str:
+    """Render a backend feature table (header, separator, one row per backend)."""
+    has_versions = any(b.get("version") for b in backends)
+    columns = _build_columns(is_mla, has_versions)
+    return "\n".join(_render_table(columns, backends))
+
+
+def _mla_prefill_table(prefill_backends: list[dict[str, Any]]) -> str:
+    """Render the MLA prefill backend table."""
     lines = [
-        "## MLA (Multi-head Latent Attention) Backends",
-        "",
-        "MLA uses separate backends for prefill and decode phases.",
-        "",
-        "### Prefill Backends",
-        "",
-        "To explicitly select a prefill backend, use",
-        "`-ac.mla_prefill_backend=<BACKEND>` (e.g., `FLASH_ATTN`, `FLASHINFER`).",
-        "Otherwise, the prefill backend is selected automatically at runtime based on",
-        "hardware and configuration.",
-        "",
         "| Backend | Description | Dtypes | Compute Cap. | Notes |",
         "| ------- | ----------- | ------ | ------------ | ----- |",
     ]
-
     for backend in prefill_backends:
         row = "| `{}`{} | {} | {} | {} | {} |".format(
             backend["name"],
@@ -1885,87 +1702,21 @@ def generate_mla_section(
             backend.get("notes", ""),
         )
         lines.append(row.replace("  ", " "))
-
-    lines.extend(
-        [
-            "",
-            "> **‡** Automatic selection tries FlashAttention first. On Blackwell",
-            "> (SM100), the fallback order is TRT-LLM Ragged, FlashInfer, then",
-            "> TokenSpeed MLA. On other GPUs, only FlashAttention is considered.",
-            "",
-            "### Decode Backends",
-            "",
-            "MLA decode backends are selected using the standard",
-            "`-ac.backend=<BACKEND>` argument (e.g., `FLASHMLA`, `TRITON_MLA`).",
-            "",
-        ]
-    )
-
-    # Reuse data-driven table rendering for decode backends
-    columns = _build_columns(is_mla=True, has_versions=False)
-    lines.extend(_render_table(columns, decode_backends))
-
-    if v4_decode_backends:
-        lines.extend(
-            [
-                "",
-                "### DeepSeek V4 Decode Backends",
-                "",
-                "DeepSeek V4 sparse MLA uses its own decode backends, selected via",
-                "`--attention-backend=<BACKEND>` (e.g., `FLASHMLA_SPARSE_DSV4`,",
-                "`FLASHINFER_MLA_SPARSE_DSV4`). They share the V4 sparse-index",
-                "pipeline (compressor + SWA + indexer, 256-token blocks, head 512);",
-                "default on NVIDIA is `FLASHINFER_MLA_SPARSE_DSV4` on SM12x and",
-                "`FLASHMLA_SPARSE_DSV4` on other supported CUDA architectures.",
-                "",
-            ]
-        )
-        lines.extend(_render_table(columns, v4_decode_backends))
-
-    lines.append("")
     return "\n".join(lines)
 
 
-def generate_minimax_section(backends: list[dict[str, Any]]) -> str:
-    """Generate the MiniMax M3 sparse attention section."""
-    lines = [
-        "## MiniMax M3 Sparse Attention Backends",
-        "",
-        'Block-sparse GQA backend used by MiniMax M3 sparse ("lightning indexer")',
-        "layers. It is wired in directly by the model and is not part of the",
-        "automatic priority lists above. A lightning indexer scores KV blocks, the",
-        "top-k blocks (plus fixed init/local blocks) are selected, and attention",
-        "attends only to those blocks; index keys live in a separate side cache.",
-        "",
-    ]
-    columns = _build_columns(is_mla=False, has_versions=False)
-    lines.extend(_render_table(columns, backends))
-    lines.append("")
-    return "\n".join(lines)
+def build_blocks() -> dict[str, str]:
+    """Build the generated table blocks keyed by their `gen:` marker name.
 
-
-# ---------------------------------------------------------------------------
-# Top-level orchestration
-# ---------------------------------------------------------------------------
-
-
-def generate_docs() -> str:
-    """Generate the complete documentation."""
+    Only the tables are generated here; the surrounding prose lives in the
+    handwritten ``docs/design/attention_backends.md`` page.
+    """
     attention_backends_map = parse_registry()
-
-    # Parse priority lists from cuda.py
     priorities = parse_cuda_priority_lists()
-
-    # Parse FlashAttention FA2/FA3 feature differences
     fa_features = parse_flash_attn_features()
-
-    # Parse FlashInfer TRTLLM feature differences (native vs TRTLLM on Blackwell)
     fi_features = parse_flashinfer_trtllm_features()
-
-    # Parse MLA prefill backends
     mla_prefill_backends = parse_mla_prefill_backends()
 
-    # Collect backend info
     all_backends = []
     for backend_name, class_path in attention_backends_map.items():
         if backend_name in SKIP_BACKENDS:
@@ -1973,17 +1724,14 @@ def generate_docs() -> str:
         info = analyze_backend(backend_name, class_path)
         if info:
             all_backends.append(info)
-
-    # Expand backends into version variants
     if fa_features:
         all_backends = _expand_flash_attn_variants(all_backends, fa_features)
     if fi_features:
         all_backends = _expand_flashinfer_variants(all_backends, fi_features)
 
-    # DeepSeek V4 (*_DSV4) decode backends and MiniMax M3 sparse backends each
-    # get their own subsection rather than mixing into the main MLA / standard
-    # tables (the ROCm V4 backend isn't flagged is_mla by the AST heuristic, so
-    # filter purely on the name).
+    # DeepSeek V4 (*_DSV4) and MiniMax M3 sparse backends get their own tables
+    # rather than mixing into the main MLA / standard tables (the ROCm V4 backend
+    # isn't flagged is_mla by the AST heuristic, so filter purely on the name).
     def _is_v4(b: dict[str, Any]) -> bool:
         return b["name"].endswith("_DSV4")
 
@@ -1999,112 +1747,21 @@ def generate_docs() -> str:
         if not b["is_mla"] and not _is_v4(b) and not _is_minimax(b)
     ]
 
-    # Generate documentation
-    script_path = "tools/pre_commit/generate_attention_backend_docs.py"
-    doc_lines = [
-        "# Attention Backend Feature Support",
-        "",
-        f"This document is auto-generated by `{script_path}`.",
-        "It shows the feature support for each registered attention backend",
-        "based on the checks in `AttentionBackend.validate_configuration()`.",
-        "",
-        "**Do not edit this file manually.** Run the following command to",
-        "regenerate it:",
-        "",
-        "```bash",
-        f"python {script_path}",
-        "```",
-        "",
-    ]
-
-    # Add usage documentation
-    doc_lines.append(generate_usage_section())
-
-    # Add priority section
-    doc_lines.append(generate_priority_section(priorities))
-
-    # Add legend and feature tables
-    doc_lines.append(generate_legend())
-    standard_title = "Standard Attention (MHA, MQA, GQA) Backends"
-    doc_lines.append(
-        generate_markdown_table(non_mla_backends, standard_title, is_mla_table=False)
-    )
-    # Add footnotes for version/variant distinctions (in table order)
-    footnotes = []
-    if fi_features:
-        footnotes.append(
-            "> **†** FlashInfer Native is the regular FlashInfer path. XQA is the "
-            "SM90 decode path exposed through FlashInfer's TRTLLM decode API. "
-            "trtllm-gen is used on SM100 and supports sinks. Disable XQA/trtllm-gen "
-            "via `--attention-config.use_trtllm_attention=0`."
-        )
-    if fa_features:
-        footnotes.append(
-            "> **\\*** Specify the FlashAttention version via "
-            "`--attention-config.flash_attn_version=2`, `3`, or `4`. "
-            "Default is FA4 on SM100+ (Blackwell), FA3 on SM90 (Hopper), "
-            "FA2 otherwise."
-        )
-    if footnotes:
-        doc_lines.append("\n>\n".join(footnotes) + "\n")
-
-    # Add MiniMax M3 sparse section (separate category after standard GQA)
-    if minimax_backends:
-        doc_lines.append(generate_minimax_section(minimax_backends))
-
-    # Add MLA section with prefill and decode backends
-    doc_lines.append(
-        generate_mla_section(mla_prefill_backends, mla_backends, v4_decode_backends)
-    )
-
-    return "\n".join(doc_lines)
+    mla_sm100_annotations = {"FLASHINFER_MLA_SPARSE": "**\\***"}
+    return {
+        "priority-standard": _priority_block(
+            priorities, "standard_sm100", "standard_default"
+        ),
+        "priority-mla": _priority_block(
+            priorities, "mla_sm100", "mla_default", mla_sm100_annotations
+        ),
+        "table-standard": _feature_table(non_mla_backends, is_mla=False),
+        "table-minimax": _feature_table(minimax_backends, is_mla=False),
+        "table-mla-prefill": _mla_prefill_table(mla_prefill_backends),
+        "table-mla-decode": _feature_table(mla_backends, is_mla=True),
+        "table-mla-v4-decode": _feature_table(v4_decode_backends, is_mla=True),
+    }
 
 
-def main():
-    parser = argparse.ArgumentParser(
-        description="Generate attention backend documentation table"
-    )
-    parser.add_argument(
-        "--output",
-        "-o",
-        type=str,
-        default=str(REPO_ROOT / "docs" / "design" / "attention_backends.md"),
-        help="Output file path (default: docs/design/attention_backends.md)",
-    )
-    parser.add_argument(
-        "--check",
-        action="store_true",
-        help="Check if the documentation is up to date (for pre-commit)",
-    )
-    parser.add_argument(
-        "files",
-        nargs="*",
-        help="Files to check (passed by pre-commit). If none are relevant, skip.",
-    )
-    args = parser.parse_args()
-
-    if args.files and not any(is_relevant_file(f) for f in args.files):
-        sys.exit(0)
-
-    output_path = Path(args.output)
-    new_content = generate_docs()
-
-    if args.check:
-        needs_update = (
-            not output_path.exists() or output_path.read_text() != new_content
-        )
-        if needs_update:
-            output_path.parent.mkdir(parents=True, exist_ok=True)
-            output_path.write_text(new_content)
-            print(f"🔄 Regenerated: {output_path}")
-            sys.exit(1)
-        print(f"✅ Up to date: {output_path}")
-        sys.exit(0)
-
-    output_path.parent.mkdir(parents=True, exist_ok=True)
-    output_path.write_text(new_content)
-    print(f"Generated: {output_path}")
-
-
-if __name__ == "__main__":
-    main()
+logger.info("Generating attention backend documentation")
+fill_markers("design/attention_backends.md", build_blocks())

docs/mkdocs/gen_files/generate_examples.py

@@ -5,16 +5,15 @@ import logging
 from dataclasses import dataclass
 from functools import cached_property
 from pathlib import Path
-from typing import Literal
 
+import mkdocs_awesome_nav.nav.directory as _nav_dir
+import mkdocs_gen_files
 import regex as re
 
 logger = logging.getLogger("mkdocs")
 
 ROOT_DIR = Path(__file__).parent.parent.parent.parent
-ROOT_DIR_RELATIVE = "../../../../.."
 EXAMPLE_DIR = ROOT_DIR / "examples"
-EXAMPLE_DOC_DIR = ROOT_DIR / "docs/examples"
 
 
 def title(text: str) -> str:
@@ -197,44 +196,38 @@ class Example:
         return content
 
 
-def on_startup(command: Literal["build", "gh-deploy", "serve"], dirty: bool):
-    # Monkey-patch dirname_to_title in awesome-nav so that sub-directory names are
-    # title-cased (e.g. "Offline Inference" instead of "Offline inference").
-    import mkdocs_awesome_nav.nav.directory as _nav_dir
+# Monkey-patch dirname_to_title in awesome-nav so that sub-directory names are
+# title-cased (e.g. "Offline Inference" instead of "Offline inference").
+_nav_dir.dirname_to_title = title
+logger.info("Generating example documentation")
+logger.debug("Root directory: %s", ROOT_DIR.resolve())
+logger.debug("Example directory: %s", EXAMPLE_DIR.resolve())
 
-    _nav_dir.dirname_to_title = title
-    logger.info("Generating example documentation")
-    logger.debug("Root directory: %s", ROOT_DIR.resolve())
-    logger.debug("Example directory: %s", EXAMPLE_DIR.resolve())
-    logger.debug("Example document directory: %s", EXAMPLE_DOC_DIR.resolve())
+categories = sorted(
+    p for p in EXAMPLE_DIR.iterdir() if p.is_dir() and not p.name.startswith(".")
+)
 
-    # Create the EXAMPLE_DOC_DIR if it doesn't exist
-    if not EXAMPLE_DOC_DIR.exists():
-        EXAMPLE_DOC_DIR.mkdir(parents=True)
+examples = []
+glob_patterns = ["*.py", "*.md", "*.sh"]
+# Find categorised examples
+for category in categories:
+    logger.info("Processing category: %s", category.stem)
+    globs = [category.glob(pattern) for pattern in glob_patterns]
+    for path in itertools.chain(*globs):
+        examples.append(Example(path, category.stem))
+    # Find examples in subdirectories
+    globs = [category.glob(f"*/{pattern}") for pattern in glob_patterns]
+    for path in itertools.chain(*globs):
+        examples.append(Example(path.parent, category.stem))
 
-    categories = sorted(p for p in EXAMPLE_DIR.iterdir() if p.is_dir())
-
-    examples = []
-    glob_patterns = ["*.py", "*.md", "*.sh"]
-    # Find categorised examples
-    for category in categories:
-        logger.info("Processing category: %s", category.stem)
-        globs = [category.glob(pattern) for pattern in glob_patterns]
-        for path in itertools.chain(*globs):
-            examples.append(Example(path, category.stem))
-        # Find examples in subdirectories
-        globs = [category.glob(f"*/{pattern}") for pattern in glob_patterns]
-        for path in itertools.chain(*globs):
-            examples.append(Example(path.parent, category.stem))
-
-    # Generate the example documentation
-    for example in sorted(examples, key=lambda e: e.path.stem):
-        example_name = f"{example.path.stem}.md"
-        doc_path = EXAMPLE_DOC_DIR / example.category / example_name
-        if not doc_path.parent.exists():
-            doc_path.parent.mkdir(parents=True)
-        # Specify encoding for building on Windows
-        with open(doc_path, "w+", encoding="utf-8") as f:
-            f.write(example.generate())
-        logger.debug("Example generated: %s", doc_path.relative_to(ROOT_DIR))
-    logger.info("Total examples generated: %d", len(examples))
+# Generate the example documentation
+for example in sorted(examples, key=lambda e: e.path.stem):
+    doc_path = f"examples/{example.category}/{example.path.stem}.md"
+    with mkdocs_gen_files.open(doc_path, "w") as f:
+        f.write(example.generate())
+    if example.main_file is not None:
+        # Point the edit button at the example's source file
+        edit_path = Path("..") / example.main_file.relative_to(ROOT_DIR)
+        mkdocs_gen_files.set_edit_path(doc_path, str(edit_path))
+    logger.debug("Example generated: %s", doc_path)
+logger.info("Total examples generated: %d", len(examples))

docs/mkdocs/gen_files/generate_metrics.py

@@ -2,27 +2,28 @@
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 import ast
 import logging
+import sys
 from pathlib import Path
-from typing import Literal
+
+sys.path.insert(0, str(Path(__file__).parent))
+
+from generated_content import fill_markers  # noqa: E402
 
 logger = logging.getLogger("mkdocs")
 
 ROOT_DIR = Path(__file__).parent.parent.parent.parent
-DOCS_DIR = ROOT_DIR / "docs"
-GENERATED_METRICS_DIR = DOCS_DIR / "generated" / "metrics"
 
-# Files to scan for metric definitions - each will generate a separate table
+# Files to scan for metric definitions - each fills a `gen:` marker in
+# docs/usage/metrics.md with its table (the section heading and any preamble
+# live in the tracked page next to the marker).
 METRIC_SOURCE_FILES = [
-    {"path": "vllm/v1/metrics/loggers.py", "output": "general.inc.md"},
-    {
-        "path": "vllm/v1/spec_decode/metrics.py",
-        "output": "spec_decode.inc.md",
-    },
+    {"path": "vllm/v1/metrics/loggers.py", "key": "metrics-general"},
+    {"path": "vllm/v1/spec_decode/metrics.py", "key": "metrics-spec-decode"},
     {
         "path": "vllm/distributed/kv_transfer/kv_connector/v1/nixl/stats.py",
-        "output": "nixl_connector.inc.md",
+        "key": "metrics-nixl",
     },
-    {"path": "vllm/v1/metrics/perf.py", "output": "perf.inc.md"},
+    {"path": "vllm/v1/metrics/perf.py", "key": "metrics-mfu"},
 ]
 
 
@@ -110,41 +111,27 @@ def generate_markdown_table(metrics: list[dict[str, str]]) -> str:
     return "\n".join(lines) + "\n"
 
 
-def on_startup(command: Literal["build", "gh-deploy", "serve"], dirty: bool):
-    """Generate metrics documentation tables from source files."""
-    logger.info("Generating metrics documentation")
+logger.info("Generating metrics documentation")
 
-    # Create generated directory if it doesn't exist
-    GENERATED_METRICS_DIR.mkdir(parents=True, exist_ok=True)
+blocks = {}
+total_metrics = 0
+for source_config in METRIC_SOURCE_FILES:
+    source_path = source_config["path"]
 
-    total_metrics = 0
-    for source_config in METRIC_SOURCE_FILES:
-        source_path = source_config["path"]
-        output_file = source_config["output"]
+    filepath = ROOT_DIR / source_path
+    if not filepath.exists():
+        raise FileNotFoundError(f"Metrics source file not found: {filepath}")
 
-        filepath = ROOT_DIR / source_path
-        if not filepath.exists():
-            raise FileNotFoundError(f"Metrics source file not found: {filepath}")
+    logger.debug("Extracting metrics from: %s", source_path)
+    metrics = extract_metrics_from_file(filepath)
+    logger.debug("Found %d metrics in %s", len(metrics), source_path)
 
-        logger.debug("Extracting metrics from: %s", source_path)
-        metrics = extract_metrics_from_file(filepath)
-        logger.debug("Found %d metrics in %s", len(metrics), source_path)
+    blocks[source_config["key"]] = generate_markdown_table(metrics).strip()
+    total_metrics += len(metrics)
 
-        # Generate and write the markdown table for this source
-        table_content = generate_markdown_table(metrics)
-        output_path = GENERATED_METRICS_DIR / output_file
-        with open(output_path, "w", encoding="utf-8") as f:
-            f.write(table_content)
-
-        total_metrics += len(metrics)
-        logger.info(
-            "Generated metrics table: %s (%d metrics)",
-            output_path.relative_to(ROOT_DIR),
-            len(metrics),
-        )
-
-    logger.info(
-        "Total metrics generated: %d across %d files",
-        total_metrics,
-        len(METRIC_SOURCE_FILES),
-    )
+fill_markers("usage/metrics.md", blocks)
+logger.info(
+    "Total metrics generated: %d across %d files",
+    total_metrics,
+    len(METRIC_SOURCE_FILES),
+)

docs/mkdocs/gen_files/generated_content.py

@@ -0,0 +1,56 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""Inline build-time generated content into existing docs pages.
+
+Source pages mark where generated content goes with a snippet-style marker,
+`--8<-- "gen:<key>"`, so the insertion point is explicit and readable. The
+substitution happens here (at gen-files time, before mkdocs-gen-files shadows
+the page), not via pymdownx.snippets, so the content can be generated at build
+time without living in a real file on disk.
+
+The `gen:` prefix keeps these markers distinct from real pymdownx.snippets
+includes, and `fill_markers` fails loudly if a marker is missing or left behind
+(pymdownx.snippets would otherwise silently drop an unsubstituted marker).
+"""
+
+from pathlib import Path
+
+import mkdocs_gen_files
+import regex as re
+
+DOCS_DIR = Path(__file__).parent.parent.parent
+
+_MARKER = '--8<-- "gen:{key}"'
+_ANY_MARKER = re.compile(r'--8<-- "gen:[^"]*"')
+
+
+def fill_markers(doc_path: str, blocks: dict[str, str]) -> None:
+    """Replace `--8<-- "gen:<key>"` markers in a docs page with generated content.
+
+    Args:
+        doc_path: Docs-relative path of the source page to fill.
+        blocks: Mapping of marker key to the markdown to insert in its place.
+
+    Raises:
+        FileNotFoundError: If the source page does not exist.
+        ValueError: If an expected marker is missing, or any `gen:` marker is
+            left unsubstituted after filling.
+    """
+    source = DOCS_DIR / doc_path
+    if not source.exists():
+        raise FileNotFoundError(f"Cannot fill markers in missing page: {doc_path}")
+
+    text = source.read_text()
+    for key, content in blocks.items():
+        marker = _MARKER.format(key=key)
+        if marker not in text:
+            raise ValueError(f"{doc_path}: missing marker {marker}")
+        text = text.replace(marker, content)
+
+    if leftover := _ANY_MARKER.search(text):
+        raise ValueError(f"{doc_path}: unsubstituted marker {leftover.group()}")
+
+    with mkdocs_gen_files.open(doc_path, "w") as f:
+        f.write(text)
+    # Keep the edit button pointing at the real source page
+    mkdocs_gen_files.set_edit_path(doc_path, doc_path)

docs/mkdocs/hooks/url_schemes.py

@@ -19,6 +19,7 @@ The on_page_markdown hook passes the current page context to the preprocessor be
 each page is converted.
 """
 
+import posixpath
 from pathlib import Path
 
 import regex as re
@@ -38,18 +39,22 @@ TITLE = r"(?P<title>[^\[\]<>]+?)"
 REPO = r"(?P<repo>.+?/.+?)"
 TYPE = r"(?P<type>issues|pull|projects)"
 NUMBER = r"(?P<number>\d+)"
+VERSION = r"[^/\s]+"
 PATH = r"(?P<path>[^\s]+?)"
 FRAGMENT = r"(?P<fragment>#[^\s]+)?"
-URL = f"https://github.com/{REPO}/{TYPE}/{NUMBER}{FRAGMENT}"
+URL_GITHUB = f"https://github.com/{REPO}/{TYPE}/{NUMBER}{FRAGMENT}"
 RELATIVE = rf"(?!(https?|ftp)://|#){PATH}{FRAGMENT}"
+URL_DOCS = f"https://docs.vllm.ai/en/{VERSION}/{PATH}{FRAGMENT}"
 
 # Common titles to use for GitHub links when none is provided in the link.
 TITLES = {"issues": "Issue ", "pull": "Pull Request ", "projects": "Project "}
 
 # Regex to match GitHub issue, PR, and project links with optional titles.
-github_link = re.compile(rf"(\[{TITLE}\]\(|<){URL}(\)|>)")
+github_link = re.compile(rf"(\[{TITLE}\]\(|<){URL_GITHUB}(\)|>)")
 # Regex to match relative file links with optional titles.
 relative_link = re.compile(rf"\[{TITLE}\]\({RELATIVE}\)")
+# Regex to match absolute docs.vllm.ai links (should only exist in CLI).
+docs_link = re.compile(rf"\[{TITLE}\]\({URL_DOCS}\)")
 
 
 class UrlSchemesPreprocessor(Preprocessor):
@@ -61,7 +66,8 @@ class UrlSchemesPreprocessor(Preprocessor):
 
     def run(self, lines):
         page = self.ext.page
-        if page is None or getattr(page.file, "abs_src_path", None) is None:
+        files = self.ext.files
+        if page is None:
             return lines
 
         def replace_relative_link(match: re.Match) -> str:
@@ -70,7 +76,7 @@ class UrlSchemesPreprocessor(Preprocessor):
             """
             title = match.group("title")
             path = match.group("path")
-            path = (Path(page.file.abs_src_path).parent / path).resolve()
+            path = ((DOC_DIR / page.file.src_uri).parent / path).resolve()
             fragment = match.group("fragment") or ""
 
             # Check if the path exists and is outside the docs dir
@@ -105,9 +111,36 @@ class UrlSchemesPreprocessor(Preprocessor):
             url = f"https://github.com/{repo}/{type}/{number}{fragment}"
             return f"[{gh_icon} {title}]({url})"
 
+        def replace_docs_link(match: re.Match) -> str:
+            """Rewrite absolute docs.vllm.ai links as doc-relative links."""
+            title = match.group("title")
+            path = match.group("path").rstrip("/")
+            fragment = match.group("fragment") or ""
+
+            # vllm.config.<Class> API reference -> mkdocstrings cross-reference
+            if path == "api/vllm/config" and re.fullmatch(
+                r"#vllm\.config\.\w+", fragment
+            ):
+                ident = fragment[1:]
+                return f"[`{ident}`][{ident}]"
+
+            # Other docs pages -> link relative to the current page, but only
+            # when the target is a known docs page (real or generated); leave
+            # unknown/external URLs untouched. This is correct even when the same
+            # docstring is also rendered on its API reference page.
+            src = f"{path.removesuffix('.html')}.md"
+            if files.get_file_from_path(src) is None:
+                return match.group(0)
+            rel = posixpath.relpath(src, posixpath.dirname(page.file.src_uri))
+            # Auto-wrapped bare URLs use the URL as their title; make it readable.
+            if title.startswith("http"):
+                title = path.removesuffix(".html")
+            return f"[{title}]({rel}{fragment})"
+
         markdown = "\n".join(lines)
         markdown = github_link.sub(replace_github_link, markdown)
         markdown = relative_link.sub(replace_relative_link, markdown)
+        markdown = docs_link.sub(replace_docs_link, markdown)
         return markdown.split("\n")
 
 
@@ -116,6 +149,7 @@ class UrlSchemesExtension(Extension):
 
     def __init__(self, **kwargs):
         self.page = None
+        self.files = None
         super().__init__(**kwargs)
 
     def extendMarkdown(self, md):
@@ -138,4 +172,5 @@ def on_page_markdown(
 ) -> str:
     """Pass the current page context to the preprocessor."""
     _ext.page = page
+    _ext.files = files
     return markdown

docs/models/supported_models.md

@@ -19,7 +19,7 @@ vLLM also supports model implementations that are available in Transformers. We
 
 Currently, the Transformers modeling backend works for the following:
 
-- Modalities: embedding models, language models and vision-language models*
+- Modalities: embedding models, language models, vision-language models* and audio-language models
 - Architectures: encoder-only, decoder-only, mixture-of-experts
 - Attention types: full attention and/or sliding attention
 
@@ -427,7 +427,6 @@ th {
 | `OlmoeForCausalLM` | OLMoE | `allenai/OLMoE-1B-7B-0924`, `allenai/OLMoE-1B-7B-0924-Instruct`, etc. | | ✅︎ |
 | `OPTForCausalLM` | OPT, OPT-IML | `facebook/opt-66b`, `facebook/opt-iml-max-30b`, etc. | ✅︎ | ✅︎ |
 | `OrionForCausalLM` | Orion | `OrionStarAI/Orion-14B-Base`, `OrionStarAI/Orion-14B-Chat`, etc. | | ✅︎ |
-| `OuroForCausalLM` | ouro | `ByteDance/Ouro-1.4B`, `ByteDance/Ouro-2.6B`, etc. | ✅︎ | |
 | `PanguEmbeddedForCausalLM` | openPangu-Embedded-7B | `FreedomIntelligence/openPangu-Embedded-7B-V1.1` | ✅︎ | ✅︎ |
 | `PanguProMoEV2ForCausalLM` | openpangu-pro-moe-v2 | | ✅︎ | ✅︎ |
 | `PanguUltraMoEForCausalLM` | openpangu-ultra-moe-718b-model | `FreedomIntelligence/openPangu-Ultra-MoE-718B-V1.1` | ✅︎ | ✅︎ |
@@ -465,6 +464,7 @@ Some models are supported only via the [Transformers modeling backend](#transfor
 | `Olmo2ForCausalLM` | OLMo2 | `allenai/OLMo-2-0425-1B`, etc. | ✅︎ | ✅︎ |
 | `SmolLM3ForCausalLM` | SmolLM3 | `HuggingFaceTB/SmolLM3-3B` | ✅︎ | ✅︎ |
 | `Starcoder2ForCausalLM` | Starcoder2 | `bigcode/starcoder2-3b`, `bigcode/starcoder2-7b`, `bigcode/starcoder2-15b`, etc. | ✅︎ | ✅︎ |
+| `VaultGemmaForCausalLM` | VaultGemma | `google/vaultgemma-1b` | ✅︎ | ✅︎ |
 
 !!! note
     Currently, the ROCm version of vLLM supports Mistral and Mixtral only for context lengths up to 4096.
@@ -607,7 +607,8 @@ Some models are supported only via the [Transformers modeling backend](#transfor
 
 | Architecture | Models | Inputs | Example HF Models | [LoRA](../features/lora.md) | [PP](../serving/parallelism_scaling.md) |
 | ------------ | ------ | ------ | ----------------- | --------------------------- | --------------------------------------- |
-| `Emu3ForConditionalGeneration` | Emu3 | T + I | `BAAI/Emu3-Chat-hf` | ✅︎ | ✅︎ |
+| `Emu3ForConditionalGeneration` | Emu3 | T + I<sup>+</sup> | `BAAI/Emu3-Chat-hf` | ✅︎ | ✅︎ |
+| `VibeVoiceAsrForConditionalGeneration` | VibeVoice-ASR | T + A<sup>+</sup> | `microsoft/VibeVoice-ASR-HF` | ✅︎ | ✅︎ |
 
 <sup>^</sup> You need to set the architecture name via `--hf-overrides` to match the one in vLLM.</br>
 <sup>E</sup> Pre-computed embeddings can be inputted for this modality.</br>

docs/pre_run_check.sh

@@ -28,7 +28,7 @@ DOCS_PATHS=(
   docs/                       # Actual docs content
   examples/                   # Examples are rendered in docs
   vllm/                       # API & CLI reference
-  requirements/test/cuda.txt  # CLI reference (see docs/mkdocs/hooks/generate_argparse.py)
+  requirements/test/cuda.txt  # CLI reference (see docs/mkdocs/gen_files/generate_argparse.py)
   mkdocs.yaml                 # Affects build process
   .readthedocs.yaml           # Affects build process
   requirements/docs.txt       # Affects build process

docs/serving/expert_parallel_deployment.md

@@ -31,9 +31,6 @@ vLLM provides multiple communication backends for EP. Use `--all2all-backend` to
 
 ## Single Node Deployment
 
-!!! warning
-    EP is an experimental feature. Argument names and default values may change in the future.
-
 ### Configuration
 
 Enable EP by setting the `--enable-expert-parallel` flag. The EP size is automatically calculated as:

docs/serving/offline_inference.md

@@ -65,6 +65,8 @@ For further details on Weight Transfer, please refer to [this page](../training/
 - `LLM.start_weight_update` - Starts a new weight update cycle.
 - `LLM.update_weights` - Updates the model weights.
 - `LLM.finish_weight_update` - Finishes the current weight update cycle.
+- `LLM.update_weight_version` - Sets the weight version without updating model weights.
+- `LLM.get_weight_version` - Returns the latest committed weight version.
 
 ## Additional APIs
 

docs/serving/online_serving/README.md

@@ -179,6 +179,8 @@ For further details on Weight Transfer, please refer to [this page](../../traini
 - `/start_weight_update` - Prepares the inference engine for a weight update.
 - `/update_weights` - Update model weights (can alter model behavior)
 - `/finish_weight_update` - Finalizes the weight update
+- `/update_weight_version` - Set the weight version without updating model weights
+- `/weight_info` - Get the latest committed weight version
 - `/get_world_size` - Get distributed world size
 
 ### Collective RPC

docs/training/async_rl.md

@@ -38,11 +38,12 @@ Resumes the scheduler after a pause. Any requests frozen with `mode="keep"` will
 
 ### HTTP Endpoints
 
-When using the vLLM HTTP server, the same functionality is available via:
+With `VLLM_SERVER_DEV_MODE=1`, the vLLM HTTP server exposes the same functionality via:
 
 - `POST /pause?mode=keep` - Pause generation
 - `POST /resume` - Resume generation
 - `POST /abort_requests` - Abort in-flight requests without pausing the scheduler (send `{}` to abort all, or `{"request_ids": [...]}`)
+- `GET /weight_info` - Return the latest committed `weight_version`
 
 !!! note "Data Parallelism"
     When using data parallelism with vLLM's **internal load balancer** (i.e. `data_parallel_backend="ray"`), pause and resume are handled automatically across all DP ranks -- a single call is sufficient. When using an **external load balancer** (i.e. multiple independent vLLM instances behind a proxy), you must send pause and resume requests to **every** engine instance individually before and after the weight update.

docs/training/weight_transfer/README.md

@@ -53,7 +53,9 @@ When running vLLM as an HTTP server, the following endpoints are available for w
 | `/init_weight_transfer_engine` | POST | Initialize the weight transfer engine with backend-specific info |
 | `/start_weight_update` | POST | Start a weight update |
 | `/update_weights` | POST | Transfer a batch of weights with backend-specific metadata |
-| `/finish_weight_update` | POST | Finish the weight update and run post-processing |
+| `/finish_weight_update` | POST | Finish the update and optionally commit its `weight_version` |
+| `/update_weight_version` | POST | Update `weight_version` without changing model weights |
+| `/weight_info` | GET | Get the latest committed weight version |
 | `/pause` | POST | Pause generation before weight sync to handle inflight requests |
 | `/resume` | POST | Resume generation after weight sync |
 | `/get_world_size` | GET | Get the number of inference workers (useful for NCCL world size calculation) |
@@ -79,7 +81,7 @@ EngineClass.trainer_send_weights(
 )
 
 # 4. Finish weight update on inference side
-llm.finish_weight_update()
+llm.finish_weight_update(weight_version="step-42")
 ```
 
 See the [NCCL](nccl.md) and [IPC](ipc.md) pages for backend-specific trainer APIs and full examples.

docs/usage/metrics.md

@@ -35,21 +35,21 @@ The following metrics are exposed:
 
 ## General Metrics
 
---8<-- "docs/generated/metrics/general.inc.md"
+--8<-- "gen:metrics-general"
 
 ## Speculative Decoding Metrics
 
---8<-- "docs/generated/metrics/spec_decode.inc.md"
+--8<-- "gen:metrics-spec-decode"
 
 ## NIXL KV Connector Metrics
 
---8<-- "docs/generated/metrics/nixl_connector.inc.md"
+--8<-- "gen:metrics-nixl"
 
 ## Model Flops Utilization (MFU) Performance Metrics
 
 These metrics are available via `--enable-mfu-metrics`:
 
---8<-- "docs/generated/metrics/perf.inc.md"
+--8<-- "gen:metrics-mfu"
 
 ## Deprecation Policy