Merge branch 'main' into codex/rocm-artifact-tensorizer

Signed-off-by: yewentao256 <zhyanwentao@126.com>

.buildkite/ci_config_rocm.yaml

@@ -0,0 +1,23 @@
+name: vllm_rocm_ci
+job_dirs:
+  - ".buildkite/hardware_tests"
+run_all_patterns:
+  - "docker/Dockerfile.rocm"
+  - "docker/Dockerfile.rocm_base"
+  - "docker/ci-rocm.hcl"
+  - "docker/docker-bake-rocm.hcl"
+  - ".buildkite/hardware_tests/amd.yaml"
+  - ".buildkite/scripts/ci-bake-rocm.sh"
+  - ".buildkite/scripts/hardware_ci/run-amd-test.py"
+  - ".buildkite/scripts/hardware_ci/run-amd-test.sh"
+  - "CMakeLists.txt"
+  - "requirements/common.txt"
+  - "requirements/rocm.txt"
+  - "requirements/build/rocm.txt"
+  - "requirements/test/rocm.txt"
+  - "setup.py"
+  - "csrc/"
+  - "cmake/"
+run_all_exclude_patterns:
+  - "csrc/cpu/"
+  - "cmake/cpu_extension.cmake"

.buildkite/hardware_tests/amd.yaml

@@ -1,42 +1,73 @@
-group: Hardware - AMD Build 
+group: Hardware - AMD Build
 steps:
-  - label: "AMD: :docker: build image"
-    key: image-build-amd
+  # Ensure ci_base is up-to-date before building the test image.
+  # Compares a content hash of ci_base-affecting files against the remote
+  # image label. If hashes match the build is skipped (< 30 s); if they
+  # differ ci_base is rebuilt and pushed automatically.
+  - label: "AMD: :docker: ensure ci_base"
+    key: ensure-ci-base-amd
     depends_on: []
     device: amd_cpu
     no_plugin: true
     commands:
-    - >
-      docker build
-      --build-arg max_jobs=16
-      --build-arg REMOTE_VLLM=1
-      --build-arg ARG_PYTORCH_ROCM_ARCH='gfx90a;gfx942;gfx950'
-      --build-arg VLLM_BRANCH=$BUILDKITE_COMMIT
-      --tag "rocm/vllm-ci:${BUILDKITE_COMMIT}"
-      -f docker/Dockerfile.rocm
-      --target test
-      --no-cache
-      --progress plain .
-    - |
-      docker run --rm --network=none --entrypoint /bin/bash "rocm/vllm-ci:${BUILDKITE_COMMIT}" -ec '
-        if [ ! -d /vllm-workspace ]; then echo Missing directory: /vllm-workspace >&2; exit 1; fi
-        if [ ! -d /vllm-workspace/tests ]; then echo Missing directory: /vllm-workspace/tests >&2; exit 1; fi
-        if [ ! -d /vllm-workspace/src/vllm ]; then echo Missing directory: /vllm-workspace/src/vllm >&2; exit 1; fi
-        if [ ! -x /vllm-workspace/src/vllm/vllm-rs ]; then echo Missing executable: /vllm-workspace/src/vllm/vllm-rs >&2; exit 1; fi
-        command -v python3
-        command -v uv
-        command -v pytest
-        if ! command -v amd-smi >/dev/null 2>&1 && ! command -v rocminfo >/dev/null 2>&1; then
-          echo No ROCm CLI found in image >&2
-          exit 1
-        fi
-        python3 - <<PY
-      import torch, vllm
-      print(torch.__version__)
-      print(vllm.__version__)
-      PY
-        echo AMD image smoke OK
-      '
-    - docker push "rocm/vllm-ci:${BUILDKITE_COMMIT}"
+      - bash .buildkite/scripts/ci-bake-rocm.sh ci-base-rocm-ci-with-deps
     env:
       DOCKER_BUILDKIT: "1"
+      VLLM_BAKE_FILE: "docker/docker-bake-rocm.hcl"
+      PYTORCH_ROCM_ARCH: "gfx90a;gfx942;gfx950"
+      REMOTE_VLLM: "1"
+      VLLM_BRANCH: "$BUILDKITE_COMMIT"
+    retry:
+      automatic:
+        - exit_status: -1  # Agent was lost
+          limit: 1
+        - exit_status: -10  # Agent was lost
+          limit: 1
+
+  - label: "AMD: :docker: build test image and artifacts"
+    key: image-build-amd
+    depends_on:
+      - ensure-ci-base-amd
+    device: amd_cpu
+    no_plugin: true
+    commands:
+      - |
+        if [[ "${ROCM_CI_ARTIFACT_ONLY:-0}" == "1" ]]; then
+          echo "ROCM_CI_ARTIFACT_ONLY=1; building ROCm wheel artifact only"
+          IMAGE_TAG="" bash .buildkite/scripts/ci-bake-rocm.sh test-rocm-ci-with-artifacts
+        else
+          bash .buildkite/scripts/ci-bake-rocm.sh test-rocm-ci-with-wheel
+        fi
+      - |
+        docker run --rm --network=none --entrypoint /bin/bash "rocm/vllm-ci:${BUILDKITE_COMMIT}" -ec '
+          if [ ! -d /vllm-workspace ]; then echo Missing directory: /vllm-workspace >&2; exit 1; fi
+          if [ ! -d /vllm-workspace/tests ]; then echo Missing directory: /vllm-workspace/tests >&2; exit 1; fi
+          if [ ! -d /vllm-workspace/src/vllm ]; then echo Missing directory: /vllm-workspace/src/vllm >&2; exit 1; fi
+          if [ ! -x /vllm-workspace/src/vllm/vllm-rs ]; then echo Missing executable: /vllm-workspace/src/vllm/vllm-rs >&2; exit 1; fi
+          command -v python3
+          command -v uv
+          command -v pytest
+          if ! command -v amd-smi >/dev/null 2>&1 && ! command -v rocminfo >/dev/null 2>&1; then
+            echo No ROCm CLI found in image >&2
+            exit 1
+          fi
+          python3 - <<PY
+        import torch, vllm
+        print(torch.__version__)
+        print(vllm.__version__)
+        PY
+          echo AMD image smoke OK
+        '
+    env:
+      DOCKER_BUILDKIT: "1"
+      VLLM_BAKE_FILE: "docker/docker-bake-rocm.hcl"
+      PYTORCH_ROCM_ARCH: "gfx90a;gfx942;gfx950"
+      IMAGE_TAG: "rocm/vllm-ci:$BUILDKITE_COMMIT"
+      REMOTE_VLLM: "1"
+      VLLM_BRANCH: "$BUILDKITE_COMMIT"
+    retry:
+      automatic:
+        - exit_status: -1  # Agent was lost
+          limit: 1
+        - exit_status: -10  # Agent was lost
+          limit: 1

.buildkite/hardware_tests/cpu.yaml

@@ -28,18 +28,19 @@ steps:
       pytest -x -v -s tests/kernels/quantization/test_cpu_fp8_scaled_mm.py
       pytest -x -v -s tests/kernels/mamba/cpu/test_cpu_gdn_ops.py"
 
-- label: CPU-Compatibility Tests
-  depends_on: []
-  device: intel_cpu
-  no_plugin: true
-  source_file_dependencies:
-  - cmake/cpu_extension.cmake
-  - setup.py
-  - vllm/platforms/cpu.py
-  commands:
-    - |
-      bash .buildkite/scripts/hardware_ci/run-cpu-test.sh 20m "
-      bash .buildkite/scripts/hardware_ci/run-cpu-compatibility-test.sh"
+# Note: SDE can't be downloaded from CI host because of AWS WAF
+# - label: CPU-Compatibility Tests
+#   depends_on: []
+#   device: intel_cpu
+#   no_plugin: true
+#   source_file_dependencies:
+#   - cmake/cpu_extension.cmake
+#   - setup.py
+#   - vllm/platforms/cpu.py
+#   commands:
+#     - |
+#       bash .buildkite/scripts/hardware_ci/run-cpu-test.sh 20m "
+#       bash .buildkite/scripts/hardware_ci/run-cpu-compatibility-test.sh"
 
 - label: CPU-Language Generation and Pooling Model Tests
   depends_on: []
@@ -90,7 +91,7 @@ steps:
   - tests/quantization/test_cpu_wna16.py
   commands:
     - |
-      bash .buildkite/scripts/hardware_ci/run-cpu-test.sh 30m "
+      bash .buildkite/scripts/hardware_ci/run-cpu-test.sh 45m "
       pytest -x -v -s tests/quantization/test_compressed_tensors.py::test_compressed_tensors_w8a8_logprobs
       pytest -x -v -s tests/quantization/test_cpu_wna16.py"
       

.buildkite/hardware_tests/intel_xpu_ci/test-intel.yaml

@@ -0,0 +1,68 @@
+group: Intel
+steps:
+  - label: ":docker: Build XPU image"
+    soft_fail: true
+    optional: true
+    depends_on: []
+    key: image-build-xpu
+    commands:
+      - bash -lc '.buildkite/image_build/image_build_xpu.sh "public.ecr.aws/q9t5s3a7" "vllm-ci-test-repo" "$BUILDKITE_COMMIT"'
+    env:
+      DOCKER_BUILDKIT: "1"
+    retry:
+      automatic:
+        - exit_status: -1  # Agent was lost
+          limit: 2
+        - exit_status: -10  # Agent was lost
+          limit: 2
+  - label: "XPU example Test"
+    depends_on:
+      - image-build-xpu
+    timeout_in_minutes: 30
+    optional: true
+    device: intel_gpu
+    no_plugin: true
+    env:
+      REGISTRY: "public.ecr.aws/q9t5s3a7"
+      REPO: "vllm-ci-test-repo"
+    source_file_dependencies:
+      - .buildkite/hardware_tests/intel_xpu_ci/test-intel.yaml
+      - .buildkite/scripts/hardware_ci/run-intel-ci-test.sh
+    commands:
+      - >-
+        bash .buildkite/scripts/hardware_ci/run-intel-test.sh
+        'bash .buildkite/scripts/hardware_ci/run-intel-ci-test.sh example'
+  - label: "XPU V1 test"
+    depends_on:
+      - image-build-xpu
+    timeout_in_minutes: 30
+    optional: true
+    device: intel_gpu
+    no_plugin: true
+    env:
+      REGISTRY: "public.ecr.aws/q9t5s3a7"
+      REPO: "vllm-ci-test-repo"
+    source_file_dependencies:
+      - .buildkite/hardware_tests/intel_xpu_ci/test-intel.yaml
+      - .buildkite/scripts/hardware_ci/run-intel-ci-test.sh
+    commands:
+      - >-
+        bash .buildkite/scripts/hardware_ci/run-intel-test.sh
+        'bash .buildkite/scripts/hardware_ci/run-intel-ci-test.sh v1'
+  - label: "XPU server test"
+    depends_on:
+      - image-build-xpu
+    timeout_in_minutes: 30
+    optional: true
+    device: intel_gpu
+    no_plugin: true
+    env:
+      REGISTRY: "public.ecr.aws/q9t5s3a7"
+      REPO: "vllm-ci-test-repo"
+    source_file_dependencies:
+      - .buildkite/hardware_tests/intel_xpu_ci/test-intel.yaml
+      - .buildkite/scripts/hardware_ci/run-intel-ci-test.sh
+    commands:
+      - >-
+        bash .buildkite/scripts/hardware_ci/run-intel-test.sh
+        'bash .buildkite/scripts/hardware_ci/run-intel-ci-test.sh server'

.buildkite/intel_jobs/basic_correctness.yaml

@@ -0,0 +1,22 @@
+group: Basic Correctness
+depends_on:
+  - image-build-xpu
+steps:
+- label: XPU Sleep Mode
+  timeout_in_minutes: 30
+  device: intel_gpu
+  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/basic_correctness/test_cumem.py
+  commands:
+    - >-
+      bash .buildkite/scripts/hardware_ci/run-intel-test.sh
+      'cd tests &&
+      export VLLM_WORKER_MULTIPROC_METHOD=spawn &&
+      pytest -v -s basic_correctness/test_mem.py::test_end_to_end'

.buildkite/intel_jobs/expert_parallelism_intel.yaml

@@ -0,0 +1,23 @@
+group: Expert Parallelism
+depends_on: 
+  - image-build-xpu
+steps:
+- label: EPLB Algorithm
+  key: eplb-algorithm
+  timeout_in_minutes: 45
+  device: intel_gpu
+  no_plugin: true
+  working_dir: "."
+  env:
+    REGISTRY: "public.ecr.aws/q9t5s3a7"
+    REPO: "vllm-ci-test-repo"
+    VLLM_TEST_DEVICE: "xpu"
+  source_file_dependencies:
+  - vllm/distributed/eplb
+  - tests/distributed/test_eplb_algo.py
+  - tests/distributed/test_eplb_utils.py
+  commands:
+    - >-
+      bash .buildkite/scripts/hardware_ci/run-intel-test.sh
+      'cd tests &&
+      pytest -v -s distributed/test_eplb_algo.py'

.buildkite/intel_jobs/misc_intel.yaml

@@ -38,7 +38,17 @@ steps:
     REPO: "vllm-ci-test-repo"
     VLLM_TEST_DEVICE: "xpu"
   source_file_dependencies:
-    - vllm/
+    - vllm/config/
+    - vllm/distributed/
+    - vllm/engine/
+    - vllm/inputs/
+    - vllm/logger.py
+    - vllm/model_executor/
+    - vllm/platforms/
+    - vllm/sampling_params.py
+    - vllm/transformers_utils/
+    - vllm/utils/
+    - vllm/v1/
     - tests/v1/sample
     - tests/v1/logits_processors
     - tests/v1/test_oracle.py
@@ -47,9 +57,134 @@ steps:
   commands:
     - >-
       bash .buildkite/scripts/hardware_ci/run-intel-test.sh
-      'export VLLM_WORKER_MULTIPROC_METHOD=spawn &&
+      'pip install lm_eval[api]>=0.4.12 &&
+      export VLLM_WORKER_MULTIPROC_METHOD=spawn &&
       cd tests &&
       pytest -v -s v1/logits_processors --ignore=v1/logits_processors/test_custom_online.py --ignore=v1/logits_processors/test_custom_offline.py &&
       pytest -v -s v1/test_oracle.py &&
       pytest -v -s v1/test_request.py &&
-      pytest -v -s v1/test_outputs.py'
+      pytest -v -s v1/test_outputs.py &&
+      pytest -v -s v1/sample/test_topk_topp_sampler.py &&
+      pytest -v -s v1/sample/test_logprobs.py &&
+      pytest -v -s v1/sample/test_logprobs_e2e.py'
+
+- label: XPU CPU Offload
+  timeout_in_minutes: 60
+  device: intel_gpu
+  no_plugin: true
+  working_dir: "."
+  env:
+    REGISTRY: "public.ecr.aws/q9t5s3a7"
+    REPO: "vllm-ci-test-repo"
+    VLLM_TEST_DEVICE: "xpu"
+  source_file_dependencies:
+    - vllm/
+    - vllm/v1/kv_offload/
+    - vllm/v1/kv_connector/
+    - tests/v1/kv_offload/
+    - tests/v1/kv_connector/unit/test_offloading_connector.py
+  commands:
+    - >-
+      bash .buildkite/scripts/hardware_ci/run-intel-test.sh
+      'export VLLM_WORKER_MULTIPROC_METHOD=spawn &&
+      cd tests &&
+      pytest -v -s v1/kv_offload &&
+      pytest -v -s v1/kv_connector/unit/test_offloading_connector.py'
+
+- label: Regression
+  key: regression
+  timeout_in_minutes: 30
+  device: intel_gpu
+  no_plugin: true
+  working_dir: "."
+  env:
+    REGISTRY: "public.ecr.aws/q9t5s3a7"
+    REPO: "vllm-ci-test-repo"
+    VLLM_TEST_DEVICE: "xpu"
+  source_file_dependencies:
+  - vllm/config/
+  - vllm/distributed/
+  - vllm/engine/
+  - vllm/inputs/
+  - vllm/model_executor/
+  - vllm/multimodal/
+  - vllm/platforms/
+  - vllm/sampling_params.py
+  - vllm/transformers_utils/
+  - vllm/utils/
+  - vllm/v1/
+  - tests/test_regression
+  commands:
+  - >-
+      bash .buildkite/scripts/hardware_ci/run-intel-test.sh
+      'pip install modelscope &&
+      cd tests &&
+      pytest -v -s test_regression.py'
+
+- label: Metrics, Tracing (2 GPUs)
+  key: metrics-tracing-2-gpus
+  timeout_in_minutes: 30
+  num_devices: 2
+  device: intel_gpu
+  no_plugin: true
+  working_dir: "."
+  env:
+    REGISTRY: "public.ecr.aws/q9t5s3a7"
+    REPO: "vllm-ci-test-repo"
+    VLLM_TEST_DEVICE: "xpu"
+  source_file_dependencies:
+  - vllm/config/
+  - vllm/distributed/
+  - vllm/engine/
+  - vllm/inputs/
+  - vllm/model_executor/
+  - vllm/multimodal/
+  - vllm/platforms/
+  - vllm/sampling_params.py
+  - vllm/tracing/
+  - vllm/transformers_utils/
+  - vllm/utils/
+  - vllm/v1/
+  - tests/v1/tracing
+  commands:
+  - >-
+    bash .buildkite/scripts/hardware_ci/run-intel-test.sh
+    'pip install opentelemetry-sdk\>=1.26.0 opentelemetry-api\>=1.26.0 opentelemetry-exporter-otlp\>=1.26.0 opentelemetry-semantic-conventions-ai\>=0.4.1 &&
+    cd tests &&
+    pytest -v -s v1/tracing'
+
+- label: Async Engine, Inputs, Utils, Worker
+  key: async-engine-inputs-utils-worker
+  timeout_in_minutes: 30
+  device: intel_gpu
+  no_plugin: true
+  working_dir: "."
+  env:
+    REGISTRY: "public.ecr.aws/q9t5s3a7"
+    REPO: "vllm-ci-test-repo"
+    VLLM_TEST_DEVICE: "xpu"
+  source_file_dependencies:
+  - vllm/assets/
+  - vllm/config/
+  - vllm/distributed/
+  - vllm/engine/
+  - vllm/inputs/
+  - vllm/model_executor/
+  - vllm/multimodal/
+  - vllm/platforms/
+  - vllm/sampling_params.py
+  - vllm/tokenizers/
+  - vllm/transformers_utils/
+  - vllm/utils/
+  - vllm/v1/
+  - tests/detokenizer
+  - tests/multimodal
+  - tests/utils_
+  commands:
+  - >-
+    bash .buildkite/scripts/hardware_ci/run-intel-test.sh
+    'cd tests &&
+    pip install av &&
+    pytest -v -s detokenizer &&
+    pytest -v -s -m "not cpu_test" ./multimodal &&
+    pytest -v -s utils_ --ignore=utils_/test_mem_utils.py'

.buildkite/intel_jobs/model_runner_v2_intel.yaml

@@ -0,0 +1,54 @@
+group: Model Runner V2 Intel
+depends_on:
+  - image-build-xpu
+steps:
+- label: Model Runner V2 Core Tests (Intel)
+  timeout_in_minutes: 45
+  device: intel_gpu
+  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
+    - vllm/v1/core/sched/
+    - vllm/v1/attention/
+    - tests/v1/engine/test_llm_engine.py
+    - tests/v1/e2e/
+  commands:
+    - >-
+      bash .buildkite/scripts/hardware_ci/run-intel-test.sh
+      'export VLLM_USE_V2_MODEL_RUNNER=1 &&
+      cd tests &&
+      pytest -v -s v1/engine/test_llm_engine.py -k "not test_engine_metrics" &&
+      ENFORCE_EAGER=1 pytest -v -s v1/e2e/general/test_async_scheduling.py -k "not ngram" &&
+      pytest -v -s v1/e2e/general/test_min_tokens.py'
+
+- label: Model Runner V2 Examples (Intel)
+  timeout_in_minutes: 45
+  device: intel_gpu
+  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/core/sched/
+    - vllm/v1/worker/gpu_worker.py
+    - examples/basic/offline_inference/
+    - examples/generate/multimodal/
+    - examples/features/
+  commands:
+    - >-
+      bash .buildkite/scripts/hardware_ci/run-intel-test.sh
+      'export VLLM_USE_V2_MODEL_RUNNER=1 &&
+      cd examples &&
+      python3 basic/offline_inference/chat.py &&
+      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'

.buildkite/intel_jobs/models_multimodal_intel.yaml

@@ -0,0 +1,111 @@
+group: Models - Multimodal
+depends_on: 
+  - image-build-xpu
+steps:
+- label: "Multi-Modal Models (Standard) 1: qwen2"
+  key: multi-modal-models-standard-1-qwen2
+  timeout_in_minutes: 45
+  device: intel_gpu
+  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/models/multimodal
+  commands:
+    - >-
+      bash .buildkite/scripts/hardware_ci/run-intel-test.sh
+      'pip install av git+https://github.com/TIGER-AI-Lab/Mantis.git &&
+      cd tests &&
+      pytest -v -s models/multimodal/generation/test_common.py -m core_model -k "qwen2" &&
+      pytest -v -s models/multimodal/generation/test_ultravox.py -m core_model'
+
+- label: "Multi-Modal Models (Standard) 2: qwen3 + gemma"
+  key: multi-modal-models-standard-2-qwen3-gemma
+  timeout_in_minutes: 45
+  device: intel_gpu
+  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/models/multimodal
+  commands:
+    - >-
+      bash .buildkite/scripts/hardware_ci/run-intel-test.sh
+      'pip install git+https://github.com/TIGER-AI-Lab/Mantis.git &&
+      cd tests &&
+      pytest -v -s models/multimodal/generation/test_qwen2_5_vl.py -m core_model'
+
+- label: "Multi-Modal Models (Standard) 3: llava + qwen2_vl"
+  key: multi-modal-models-standard-3-llava-qwen2-vl
+  timeout_in_minutes: 45
+  device: intel_gpu
+  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/models/multimodal
+  commands:
+    - >-
+      bash .buildkite/scripts/hardware_ci/run-intel-test.sh
+      'pip install git+https://github.com/TIGER-AI-Lab/Mantis.git &&
+      cd tests &&
+      pytest -v -s models/multimodal/generation/test_common.py -m core_model -k "not qwen2 and not qwen3 and not gemma" &&
+      pytest -v -s models/multimodal/generation/test_qwen2_vl.py -m core_model'
+
+- label: "Multi-Modal Models (Standard) 4: other + whisper"
+  key: multi-modal-models-standard-4-other-whisper
+  timeout_in_minutes: 45
+  device: intel_gpu
+  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/models/multimodal
+  commands:
+    - >-
+      bash .buildkite/scripts/hardware_ci/run-intel-test.sh
+      'pip install av git+https://github.com/TIGER-AI-Lab/Mantis.git &&
+      cd tests &&
+      pytest -v -s models/multimodal -m core_model --ignore models/multimodal/generation/test_common.py --ignore models/multimodal/generation/test_ultravox.py --ignore models/multimodal/generation/test_qwen2_5_vl.py --ignore models/multimodal/generation/test_qwen2_vl.py --ignore models/multimodal/generation/test_whisper.py  --ignore models/multimodal/generation/test_memory_leak.py --ignore models/multimodal/processing'
+
+- label: Multi-Modal Processor # 44min
+  key: multi-modal-processor
+  timeout_in_minutes: 45
+  device: intel_gpu
+  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/models/multimodal
+  - tests/models/registry.py
+  commands:
+    - >-
+      bash .buildkite/scripts/hardware_ci/run-intel-test.sh
+      'pip install av matplotlib ftfy git+https://github.com/TIGER-AI-Lab/Mantis.git &&
+      pip install open-clip-torch --no-deps &&
+      cd tests &&
+      pytest -v -s models/multimodal/processing/test_tensor_schema.py
+      --deselect "tests/models/multimodal/processing/test_tensor_schema.py::test_model_tensor_schema[mistralai/Mistral-Large-3-675B-Instruct-2512-NVFP4]"
+      --deselect "tests/models/multimodal/processing/test_tensor_schema.py::test_model_tensor_schema[Qwen/Qwen2.5-Omni-7B-AWQ]"
+      --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --shard-id=$$BUILDKITE_PARALLEL_JOB'
+  parallelism: 4

.buildkite/intel_jobs/test-intel.yaml

@@ -40,7 +40,9 @@ steps:
         python3 examples/basic/offline_inference/generate.py --model superjob/Qwen3-4B-Instruct-2507-GPTQ-Int4 --block-size 64 --enforce-eager --max-model-len 8192 &&
         python3 examples/basic/offline_inference/generate.py --model ibm-research/PowerMoE-3b --block-size 64 --enforce-eager -tp 2 &&
         python3 examples/basic/offline_inference/generate.py --model ibm-research/PowerMoE-3b --block-size 64 --enforce-eager -tp 2 --enable-expert-parallel &&
-        python3 examples/basic/offline_inference/generate.py --model superjob/Qwen3-4B-Instruct-2507-GPTQ-Int4 --max-model-len 8192
+        python3 examples/basic/offline_inference/generate.py --model superjob/Qwen3-4B-Instruct-2507-GPTQ-Int4 --max-model-len 8192 &&
+        VLLM_XPU_FUSED_MOE_USE_REF=1 python3 examples/basic/offline_inference/generate.py --model Qwen/Qwen3-30B-A3B-Instruct-2507-FP8 --enforce-eager -tp 2 --max-model-len 8192 &&
+        python3 examples/basic/offline_inference/generate.py --model INCModel/Qwen3-30B-A3B-Instruct-2507-MXFP4-LLMC --enforce-eager -tp 2 --max-model-len 8192
         '
   - label: "XPU V1 test"
     depends_on:
@@ -58,6 +60,7 @@ steps:
       - >-
         bash .buildkite/scripts/hardware_ci/run-intel-test.sh
         'cd tests &&
+        bash v1/kv_connector/nixl_integration/run_xpu_disagg_accuracy_test.sh &&
         pytest -v -s v1/core --ignore=v1/core/test_reset_prefix_cache_e2e.py --ignore=v1/core/test_scheduler_e2e.py &&
         pytest -v -s v1/engine --ignore=v1/engine/test_output_processor.py &&
         pytest -v -s v1/sample --ignore=v1/sample/test_logprobs.py --ignore=v1/sample/test_logprobs_e2e.py -k "not test_topk_only and not test_topp_only and not test_topk_and_topp" &&
@@ -83,5 +86,22 @@ steps:
         bash .buildkite/scripts/hardware_ci/run-intel-test.sh
         'pip install av &&
         cd tests &&
-        pytest -v -s entrypoints/openai/chat_completion/test_audio_in_video.py &&
+        pytest -v -s entrypoints/multimodal/openai/chat_completion/test_audio_in_video.py &&
         pytest -v -s benchmarks/test_serve_cli.py'
+  - label: "XPU quantization test"
+    depends_on:
+      - image-build-xpu
+    timeout_in_minutes: 30
+    device: intel_gpu
+    no_plugin: true
+    env:
+      REGISTRY: "public.ecr.aws/q9t5s3a7"
+      REPO: "vllm-ci-test-repo"
+    source_file_dependencies:
+      - vllm/
+      - .buildkite/intel_jobs/test-intel.yaml 
+    commands:
+      - >-
+        bash .buildkite/scripts/hardware_ci/run-intel-test.sh
+        'cd tests &&
+        pytest -v -s quantization/test_auto_round.py'

.buildkite/lm-eval-harness/configs/NVIDIA-Nemotron-3-Nano-30B-A3B-FP8.yaml

@@ -6,9 +6,7 @@ tasks:
     value: 0.7142
   - name: "exact_match,flexible-extract"
     value: 0.4579
-env_vars:
-  VLLM_USE_FLASHINFER_MOE_FP8: "1"
-  VLLM_FLASHINFER_MOE_BACKEND: "throughput"
+moe_backend: "flashinfer_cutlass"
 limit: 1319
 num_fewshot: 5
 max_model_len: 262144

.buildkite/lm-eval-harness/test_lm_eval_correctness.py

@@ -68,6 +68,10 @@ def launch_lm_eval(eval_config, tp_size):
     if current_platform.is_rocm() and "Nemotron-3" in eval_config["model_name"]:
         model_args += "attention_backend=TRITON_ATTN"
 
+    moe_backend = eval_config.get("moe_backend", None)
+    if moe_backend is not None:
+        model_args += f"moe_backend={moe_backend},"
+
     env_vars = eval_config.get("env_vars", None)
     with scoped_env_vars(env_vars):
         results = lm_eval.simple_evaluate(

.buildkite/performance-benchmarks/tests/latency-tests-arm64-cpu.json

@@ -2,7 +2,6 @@
     {
         "test_name": "latency_llama8B_tp1",
         "environment_variables": {
-            "VLLM_RPC_TIMEOUT": 100000,
             "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
             "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
             "VLLM_CPU_KVCACHE_SPACE": 40

.buildkite/performance-benchmarks/tests/latency-tests-cpu.json

@@ -2,7 +2,6 @@
     {
         "test_name": "latency_llama8B_tp2",
         "environment_variables": {
-            "VLLM_RPC_TIMEOUT": 100000,
 	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
 	    "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
 	    "VLLM_CPU_SGL_KERNEL": 1,

.buildkite/performance-benchmarks/tests/serving-tests-arm64-cpu.json

@@ -13,7 +13,6 @@
       200
     ],
     "server_environment_variables": {
-      "VLLM_RPC_TIMEOUT": 100000,
       "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
       "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
       "VLLM_CPU_SGL_KERNEL": 1,

.buildkite/performance-benchmarks/tests/serving-tests-cpu-asr.json

@@ -5,7 +5,6 @@
     ],
     "max_concurrency_list": [12, 16, 24, 32, 64, 128, 200],
     "server_environment_variables": {
-      "VLLM_RPC_TIMEOUT": 100000,
       "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120
     },
     "server_parameters": {

.buildkite/performance-benchmarks/tests/serving-tests-cpu-embed.json

@@ -9,7 +9,6 @@
       128
     ],
     "server_environment_variables": {
-      "VLLM_RPC_TIMEOUT": 100000,
       "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
       "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
       "VLLM_CPU_SGL_KERNEL": 1,

.buildkite/performance-benchmarks/tests/serving-tests-cpu-text.json

@@ -5,7 +5,6 @@
     ],
     "max_concurrency_list": [12, 16, 24, 32, 64, 128, 200],
     "server_environment_variables": {
-      "VLLM_RPC_TIMEOUT": 100000,
       "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
       "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
       "VLLM_CPU_SGL_KERNEL": 1,

.buildkite/performance-benchmarks/tests/serving-tests-cpu.json

@@ -5,7 +5,6 @@
     ],
     "max_concurrency_list": [12, 16, 24, 32, 64, 128, 200],
     "server_environment_variables": {
-      "VLLM_RPC_TIMEOUT": 100000,
       "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
       "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
       "VLLM_CPU_SGL_KERNEL": 1,

.buildkite/performance-benchmarks/tests/throughput-tests-arm64-cpu.json

@@ -2,7 +2,6 @@
     {
         "test_name": "throughput_llama8B_tp1",
         "environment_variables": {
-            "VLLM_RPC_TIMEOUT": 100000,
             "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
             "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
             "VLLM_CPU_KVCACHE_SPACE": 40

.buildkite/performance-benchmarks/tests/throughput-tests-cpu.json

@@ -2,7 +2,6 @@
     {
         "test_name": "throughput_llama8B_tp2",
         "environment_variables": {
-            "VLLM_RPC_TIMEOUT": 100000,
 	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
 	    "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
 	    "VLLM_CPU_SGL_KERNEL": 1,

.buildkite/release-pipeline.yaml

@@ -1,12 +1,25 @@
 # CUDA architecture lists — following PyTorch RELEASE.md
 # (https://github.com/pytorch/pytorch/blob/main/RELEASE.md)
 # SM86 included for broader Ampere coverage; SM89 for marlin fp8 support
+# These requested arches are filtered by CMake's CUDA_SUPPORTED_ARCHS before
+# per-kernel arch selection. Do not add +PTX here: top-level +PTX is stripped
+# during that filtering, so kernels that need PTX must request it locally.
 env:
-  CUDA_ARCH_X86: "7.5 8.0 8.6 8.9 9.0 10.0 12.0+PTX"
-  # aarch64 only architectures: 8.7 for Orin, 11.0 for Thor (since CUDA 13)
-  CUDA_ARCH_AARCH64: "8.0 8.7 8.9 9.0 10.0 11.0 12.0+PTX"
+  # for CUDA >=13, sm_100+ targets have family specifiers (see CMakeLists.txt)
+  # so targets like 10.3 and 12.1 are automatically supported with this list
+  CUDA_ARCH_X86: "7.5 8.0 8.6 8.9 9.0 10.0 12.0"
+  # aarch64-only targets: Orin (8.7), Thor (11.0, CUDA 13+)
+  CUDA_ARCH_AARCH64: "8.0 8.7 8.9 9.0 10.0 11.0 12.0"
+
+  # for CUDA <13, we need to specify all needed targets
+  # some targets (10.3, 12.1) are skipped to limit the wheel size (< 500MB)
+  # please use CUDA 13 wheels or compile yourself on these new devices
   CUDA_ARCH_X86_CU129: "7.5 8.0 8.6 8.9 9.0 10.0 12.0"
   CUDA_ARCH_AARCH64_CU129: "8.0 8.7 8.9 9.0 10.0 12.0"
+  
+  # pre-built mooncake wheels
+  # the manylinux_2_35 wheel has compatibility issue on Ubuntu 24.04
+  # so we use different wheels for the time being
   MOONCAKE_WHEEL_AARCH64_2_35: "https://vllm-wheels.s3.amazonaws.com/mooncake/mooncake_transfer_engine-0.3.10.post2-0da9dfea3-cp312-cp312-manylinux_2_35_aarch64.whl"
   MOONCAKE_WHEEL_AARCH64_2_39: "https://vllm-wheels.s3.amazonaws.com/mooncake/mooncake_transfer_engine-0.3.10.post2-0da9dfea3-cp312-cp312-manylinux_2_39_aarch64.whl"
   MOONCAKE_WHEEL_X86_64: "https://vllm-wheels.s3.amazonaws.com/mooncake/mooncake_transfer_engine-0.3.10.post2-0da9dfea3-cp312-cp312-manylinux_2_35_x86_64.whl"
@@ -846,7 +859,6 @@ steps:
         allow_failure: true
       - step: build-cpu-release-image-arm64
         allow_failure: true
-    if: build.env("NIGHTLY") != "1"
 
   - label: "Publish release images to DockerHub"
     depends_on:

.buildkite/scripts/ci-clean-log.sh

@@ -13,5 +13,8 @@ INPUT_FILE="$1"
 # Strip timestamps
 sed -i 's/^\[[0-9]\{4\}-[0-9]\{2\}-[0-9]\{2\}T[0-9]\{2\}:[0-9]\{2\}:[0-9]\{2\}Z\] //' "$INPUT_FILE"
 
+# Strip Buildkite inline timestamp markers (ESC _bk;t=<ms> BEL)
+sed -i 's/\x1B_bk;t=[0-9]*\x07//g' "$INPUT_FILE"
+
 # Strip colorization
 sed -i -r 's/\x1B\[[0-9;]*[mK]//g' "$INPUT_FILE"

.buildkite/scripts/ci-fetch-log.sh

@@ -1,74 +1,178 @@
 #!/bin/bash
-# Usage: ./ci-fetch-log.sh <buildkite_job_url> [output_file]
-#        ./ci-fetch-log.sh <build_number> <job_uuid> [output_file]
+# Fetch vLLM Buildkite CI logs (public; no login required).
 #
-# Downloads the raw log for a Buildkite job from the public, unauthenticated
-# /organizations/<org>/pipelines/<pipeline>/builds/<n>/jobs/<uuid>/download
-# endpoint, then strips ANSI/timestamps via ci-clean-log.sh.
+# Usage:
+#   ci-fetch-log.sh [--soft|--all] --pr [<PR>]  failed jobs in the PR's latest
+#                                               build (current branch if omitted)
+#   ci-fetch-log.sh [--soft|--all] <build_url>  failed jobs in that build
+#   ci-fetch-log.sh <job_url> [output]          one job; both #<job_uuid> and
+#                                               ?sid=<id> URL forms work
+#   ci-fetch-log.sh <build> <job_uuid> [output]
 #
-# Find <build_number> and <job_uuid> via:
-#   gh pr checks <PR> --repo vllm-project/vllm
-# Each failing row's URL is .../builds/<build_number>#<job_uuid>.
-#
-# Default output path: ci-<build>-<uuid_first_13_chars>.log (e.g.
-# ci-68478-019e6b07-daae.log). Jobs in the same build share the UUID's
-# first 8 chars, so the second segment is needed for uniqueness when
-# fetching multiple jobs in parallel. The script refuses to overwrite an
-# existing output file; pass an explicit path or set CI_FETCH_LOG_FORCE=1
-# to override.
+# --soft also fetches soft-failed jobs; --all fetches every finished job.
+# Saves each log as ci-<build>-<job-name>.log (ANSI/timestamps stripped) and
+# prints "<file>\t<job name>" per job. [output] is single-job only; "-"
+# streams to stdout. Existing files are kept; CI_FETCH_LOG_FORCE=1 refetches.
 
 set -euo pipefail
 
 ORG="vllm"
 PIPELINE="ci"
+UA="vllm-ci-fetch-log"
+UUID_RE='[0-9a-f]{8}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{12}'
 
 usage() {
-    echo "Usage: $0 <buildkite_job_url> [output_file]"
-    echo "       $0 <build_number> <job_uuid> [output_file]"
+    sed -n '2,15p' "$0" | sed 's/^# \{0,1\}//'
     exit 1
 }
 
-if [ $# -lt 1 ]; then usage; fi
+die() {
+    echo "$1" >&2
+    exit 1
+}
 
-if [[ "$1" == https://* ]]; then
+BUILD="" JOB="" SID="" OUT=""
+SCOPE="failed"
+
+while :; do
+    case "${1:-}" in
+    --soft) SCOPE="soft" ;;
+    --all) SCOPE="all" ;;
+    *) break ;;
+    esac
+    shift
+done
+
+case "${1:-}" in
+--pr)
+    PR="${2:-}"
+    # gh pr checks exits non-zero when checks are failing; that is the
+    # expected case here.
+    URL=$(gh pr checks ${PR:+"$PR"} --repo vllm-project/vllm 2>/dev/null |
+        grep -oE "https://buildkite.com/${ORG}/${PIPELINE}/builds/[0-9]+" |
+        sort -t/ -k7 -n | tail -1 || true)
+    [ -n "$URL" ] || die "No Buildkite build found via: gh pr checks ${PR:-<current branch>}"
+    BUILD="${URL##*/}"
+    ;;
+https://*)
     BUILD=$(echo "$1" | sed -nE 's#.*/builds/([0-9]+).*#\1#p')
-    JOB=$(echo "$1" | grep -oE '[0-9a-f]{8}-[0-9a-f-]+' | head -n 1)
+    JOB=$(echo "$1" | grep -oE "#${UUID_RE}" | head -n 1 | cut -c2- || true)
+    SID=$(echo "$1" | grep -oE "[?&]sid=${UUID_RE}" | head -n 1 | sed 's/.*sid=//' || true)
     OUT="${2:-}"
-else
-    if [ $# -lt 2 ]; then usage; fi
+    [ -n "$BUILD" ] || die "Could not parse build number from: $1"
+    ;;
+[0-9]*)
+    [ $# -ge 2 ] || usage
     BUILD="$1"
     JOB="$2"
     OUT="${3:-}"
-fi
-
-if [ -z "$BUILD" ] || [ -z "$JOB" ]; then
-    echo "Could not parse build number or job UUID from: $1" >&2
+    ;;
+*)
     usage
-fi
-
-# Jobs in the same build share the UUID's first segment, so include the
-# second segment (chars 9-13, e.g. "019e6b07-daae") to keep default filenames
-# unique when fetching multiple jobs from one build in parallel.
-if [ -z "$OUT" ]; then
-    OUT="ci-${BUILD}-${JOB:0:13}.log"
-fi
-
-if [ -e "$OUT" ] && [ -z "${CI_FETCH_LOG_FORCE:-}" ]; then
-    echo "Refusing to overwrite existing $OUT (set CI_FETCH_LOG_FORCE=1 or pass an explicit output path)." >&2
-    exit 1
-fi
+    ;;
+esac
 
 COOKIES=$(mktemp)
-trap 'rm -f "$COOKIES"' EXIT
+JOBS_TSV=$(mktemp)
+trap 'rm -f "$COOKIES" "$JOBS_TSV"' EXIT
 
-# Buildkite issues a session cookie on first hit; subsequent /download needs it.
-curl -fsSL -c "$COOKIES" -A "vllm-ci-fetch-log" \
+# Buildkite issues a session cookie on first hit; later requests need it.
+curl -fsSL -c "$COOKIES" -A "$UA" \
     "https://buildkite.com/${ORG}/${PIPELINE}/builds/${BUILD}" -o /dev/null
 
-curl -fsSL -b "$COOKIES" -A "vllm-ci-fetch-log" \
-    "https://buildkite.com/organizations/${ORG}/pipelines/${PIPELINE}/builds/${BUILD}/jobs/${JOB}/download" \
-    -o "$OUT"
+# The build's job list (id, step uuid, state, name) is served as JSON from
+# the user-facing /data/jobs endpoint. Flatten it to TSV for easy filtering:
+#   job_id  step_uuid  failed  soft_failed  finished  slug  name
+curl -fsSL -b "$COOKIES" -A "$UA" \
+    "https://buildkite.com/${ORG}/${PIPELINE}/builds/${BUILD}/data/jobs" |
+    python3 -c '
+import json, re, sys
 
-bash "$(dirname "$0")/ci-clean-log.sh" "$OUT"
+data = json.load(sys.stdin)
+if data.get("has_next_page"):
+    print("warning: job list is paginated; some jobs not shown", file=sys.stderr)
+for r in data["records"]:
+    if r.get("type") != "script":
+        continue
+    name = (r.get("name") or "").replace("\t", " ").replace("\n", " ")
+    slug = re.sub(r"[^a-z0-9]+", "-", name.lower()).strip("-")[:60]
+    print("\t".join([
+        r["id"],
+        r.get("step_uuid") or "",
+        str(r.get("passed") is False),
+        str(bool(r.get("soft_failed"))),
+        str(bool(r.get("finished_at"))),
+        slug,
+        name,
+    ]))
+' >"$JOBS_TSV" || die "Could not list jobs for build ${BUILD}"
 
-echo "$OUT"
+if [ -n "$SID" ] && [ -z "$JOB" ]; then
+    # The ?sid= in builds/<N>/list URLs is the *step* uuid, not the job uuid.
+    JOB=$(awk -F'\t' -v s="$SID" '$1 == s || $2 == s {print $1; exit}' "$JOBS_TSV")
+    [ -n "$JOB" ] || die "No job matching sid=${SID} in build ${BUILD}"
+fi
+
+fetch_job() { # <job_uuid> <output_file>
+    curl -fsSL -b "$COOKIES" -A "$UA" \
+        "https://buildkite.com/organizations/${ORG}/pipelines/${PIPELINE}/builds/${BUILD}/jobs/$1/download" \
+        -o "$2"
+    bash "$(dirname "$0")/ci-clean-log.sh" "$2"
+}
+
+if [ -n "$JOB" ]; then
+    # Single-job mode.
+    NAME=$(awk -F'\t' -v j="$JOB" '$1 == j {print $7; exit}' "$JOBS_TSV")
+    SLUG=$(awk -F'\t' -v j="$JOB" '$1 == j {print $6; exit}' "$JOBS_TSV")
+    [ -n "$OUT" ] || OUT="ci-${BUILD}-${SLUG:-${JOB:0:13}}.log"
+    if [ "$OUT" = "-" ]; then
+        TMP=$(mktemp)
+        fetch_job "$JOB" "$TMP"
+        cat "$TMP"
+        rm -f "$TMP"
+        exit 0
+    fi
+    if [ -e "$OUT" ] && [ -z "${CI_FETCH_LOG_FORCE:-}" ]; then
+        die "Refusing to overwrite existing ${OUT} (set CI_FETCH_LOG_FORCE=1 or pass an output path)."
+    fi
+    fetch_job "$JOB" "$OUT"
+    printf '%s\t%s\n' "$OUT" "${NAME:-$JOB}"
+    exit 0
+fi
+
+# Build-wide mode: fetch finished jobs matching $SCOPE.
+[ -z "$OUT" ] || die "[output_file] is only valid when fetching a single job."
+
+case "$SCOPE" in
+failed) FILTER='$3 == "True" && $4 == "False" && $5 == "True"' ;;
+soft) FILTER='$3 == "True" && $5 == "True"' ;;
+all) FILTER='$5 == "True"' ;;
+esac
+
+if [ "$SCOPE" = "failed" ]; then
+    SOFT=$(awk -F'\t' '$3 == "True" && $4 == "True"' "$JOBS_TSV" | wc -l)
+    [ "$SOFT" -eq 0 ] || echo "Skipping ${SOFT} soft-failed job(s); use --soft to include them." >&2
+fi
+
+FOUND=0
+EMITTED=" "
+while IFS=$'\t' read -r job_id _ _ _ _ slug name; do
+    FOUND=$((FOUND + 1))
+    out="ci-${BUILD}-${slug:-${job_id:0:13}}.log"
+    # Retries share a name with the original job; disambiguate by uuid.
+    case "$EMITTED" in
+    *" $out "*) out="ci-${BUILD}-${slug:-job}-${job_id:0:13}.log" ;;
+    esac
+    EMITTED="${EMITTED}${out} "
+    if [ -e "$out" ] && [ -z "${CI_FETCH_LOG_FORCE:-}" ]; then
+        echo "Keeping existing ${out} (set CI_FETCH_LOG_FORCE=1 to refetch)." >&2
+    elif ! fetch_job "$job_id" "$out"; then
+        echo "Failed to download log for job ${job_id} (${name})." >&2
+        continue
+    fi
+    printf '%s\t%s\n' "$out" "$name"
+done < <(awk -F'\t' "$FILTER" "$JOBS_TSV")
+
+if [ "$FOUND" -eq 0 ]; then
+    echo "No matching jobs in build ${BUILD} (scope: ${SCOPE})." >&2
+fi

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

@@ -28,8 +28,10 @@
 ###############################################################################
 set -o pipefail
 
-# Export Python path
-export PYTHONPATH=".."
+# Export Python path for commands that run directly on the host. Containerized
+# tests set this to /vllm-workspace below so spawned Python processes do not
+# depend on their current working directory.
+export PYTHONPATH="${PYTHONPATH:-..}"
 
 ###############################################################################
 # Helper Functions
@@ -52,6 +54,108 @@ cleanup_network() {
   fi
 }
 
+prepare_artifact_image() {
+  if [[ "${VLLM_CI_USE_ARTIFACTS:-0}" != "1" ]]; then
+    return 1
+  fi
+  if ! command -v buildkite-agent >/dev/null 2>&1; then
+    echo "buildkite-agent not found; cannot download ROCm wheel artifact"
+    return 1
+  fi
+
+  local artifact_glob="${VLLM_CI_ARTIFACT_GLOB:-artifacts/vllm-rocm-install/vllm-rocm-install.tar.gz}"
+  local archive=""
+  local metadata_file=""
+  local base_image="${VLLM_CI_BASE_IMAGE:-rocm/vllm-dev:ci_base}"
+  local artifact_image=""
+  local artifact_key=""
+  local base_digest=""
+  local wheel_dir=""
+  local context_dir=""
+  local workspace_dir=""
+
+  artifact_work_dir=$(mktemp -d -t vllm-rocm-artifact.XXXXXX)
+  wheel_dir="${artifact_work_dir}/wheels"
+  context_dir="${artifact_work_dir}/context"
+  workspace_dir="${context_dir}/workspace"
+  mkdir -p "${wheel_dir}" "${context_dir}/wheels" "${workspace_dir}"
+
+  echo "--- Downloading ROCm wheel artifact"
+  if ! buildkite-agent artifact download "${artifact_glob}" "${artifact_work_dir}"; then
+    echo "Failed to download ${artifact_glob}"
+    return 1
+  fi
+  buildkite-agent artifact download \
+    "artifacts/vllm-rocm-install/ci-base-image.txt" \
+    "${artifact_work_dir}" >/dev/null 2>&1 || true
+
+  archive=$(find "${artifact_work_dir}" -name "vllm-rocm-install.tar.gz" -type f | head -1)
+  if [[ -z "${archive}" || ! -f "${archive}" ]]; then
+    echo "ROCm wheel artifact archive was not found"
+    return 1
+  fi
+
+  metadata_file=$(find "${artifact_work_dir}" -name "ci-base-image.txt" -type f | head -1)
+  if [[ -n "${metadata_file}" && -s "${metadata_file}" ]]; then
+    base_image=$(tr -d '[:space:]' < "${metadata_file}")
+  fi
+
+  echo "--- Preparing local ROCm test image"
+  echo "Base image: ${base_image}"
+  docker pull "${base_image}" || return 1
+  base_digest=$(
+    docker image inspect \
+      --format='{{if .RepoDigests}}{{index .RepoDigests 0}}{{else}}{{.Id}}{{end}}' \
+      "${base_image}" 2>/dev/null || printf '%s' "${base_image}"
+  )
+
+  artifact_key=$(
+    {
+      printf 'base-image:%s\n' "${base_digest}"
+      sha256sum "${archive}"
+    } | sha256sum | cut -c1-24
+  )
+  artifact_image="rocm/vllm-ci-artifact:${artifact_key}"
+
+  if docker image inspect "${artifact_image}" >/dev/null 2>&1; then
+    echo "Using existing local ROCm artifact image: ${artifact_image}"
+    image_name="${artifact_image}"
+    return 0
+  fi
+
+  tar -xzf "${archive}" -C "${wheel_dir}" || return 1
+  if ! ls "${wheel_dir}"/*.whl >/dev/null 2>&1; then
+    echo "ROCm wheel artifact did not contain a wheel"
+    return 1
+  fi
+  if [[ ! -d "${wheel_dir}/tests" ]]; then
+    echo "ROCm wheel artifact did not contain the test workspace"
+    return 1
+  fi
+
+  cp "${wheel_dir}"/*.whl "${context_dir}/wheels/" || return 1
+  tar -C "${wheel_dir}" --exclude='*.whl' -cf - . \
+    | tar -C "${workspace_dir}" -xf - || return 1
+  cat > "${context_dir}/Dockerfile" <<'EOF'
+ARG BASE_IMAGE
+FROM ${BASE_IMAGE}
+COPY wheels/ /tmp/vllm-wheels/
+COPY workspace/ /vllm-workspace/
+RUN python3 -m pip install --no-deps --force-reinstall /tmp/vllm-wheels/*.whl \
+    && rm -rf /tmp/vllm-wheels
+WORKDIR /vllm-workspace
+EOF
+
+  echo "--- Building local ROCm test image"
+  docker build \
+    --pull=false \
+    --build-arg "BASE_IMAGE=${base_image}" \
+    -t "${artifact_image}" \
+    "${context_dir}" || return 1
+  image_name="${artifact_image}"
+  return 0
+}
+
 is_multi_node() {
   local cmds="$1"
   # Primary signal: NUM_NODES environment variable set by the pipeline
@@ -243,22 +347,30 @@ report_docker_usage
 
 # --- Pull test image ---
 echo "--- Pulling container"
-image_name="rocm/vllm-ci:${BUILDKITE_COMMIT}"
+image_name="${VLLM_CI_FALLBACK_IMAGE:-rocm/vllm-ci:${BUILDKITE_COMMIT:-local}}"
+artifact_work_dir=""
 container_name="rocm_${BUILDKITE_COMMIT}_$(tr -dc A-Za-z0-9 < /dev/urandom | head -c 10; echo)"
-docker pull "${image_name}"
 
 remove_docker_container() {
-  # docker run uses --rm, so the container is normally already gone when the
-  # EXIT trap runs. Cleanup is best-effort and must not affect the test result.
-  docker rm -f "${container_name}" >/dev/null 2>&1 || true
+  if docker container inspect "${container_name}" >/dev/null 2>&1; then
+    docker rm -f "${container_name}" || true
+  fi
+  if [[ "${VLLM_CI_REMOVE_TEST_IMAGE:-0}" == "1" ]]; then
+    docker image rm -f "${image_name}" || true
+  else
+    # Keep images by default so later jobs on the same AMD node can reuse layers.
+    echo "Keeping ROCm test image locally: ${image_name}"
+  fi
+  if [[ -n "${artifact_work_dir}" ]]; then
+    rm -rf "${artifact_work_dir}"
+  fi
 }
+trap remove_docker_container EXIT
 
-on_exit() {
-  local exit_code=$?
-  remove_docker_container
-  exit "$exit_code"
-}
-trap on_exit EXIT
+if ! prepare_artifact_image; then
+  echo "Using full ROCm CI image: ${image_name}"
+  docker pull "${image_name}" || exit 1
+fi
 
 # --- Prepare commands ---
 echo "--- Running container"
@@ -267,6 +379,14 @@ HF_CACHE="$(realpath ~)/huggingface"
 mkdir -p "${HF_CACHE}"
 HF_MOUNT="/root/.cache/huggingface"
 
+# Hugging Face Hub defaults to 10s request/download timeouts, while the ROCm
+# CI image currently raises downloads to 60s. AMD model-test jobs routinely
+# start from a cold or partially-populated shared cache, and the 60s read cap
+# has still timed out before pytest reached the vLLM behavior under test.
+# Keep the CI default explicit and overridable from the Buildkite environment.
+: "${HF_HUB_DOWNLOAD_TIMEOUT:=300}"
+: "${HF_HUB_ETAG_TIMEOUT:=60}"
+
 # ---- Command source selection ----
 # Prefer VLLM_TEST_COMMANDS (preserves all inner quoting intact).
 # Fall back to $* for backward compatibility, but warn that inner
@@ -306,7 +426,14 @@ fi
 
 echo "Final commands: $commands"
 
-MYPYTHONPATH=".."
+MYPYTHONPATH="/vllm-workspace"
+
+container_job_id="${BUILDKITE_JOB_ID:-${BUILDKITE_PARALLEL_JOB:-0}}"
+container_job_id="${container_job_id//[^A-Za-z0-9_.-]/_}"
+container_job_id_short="${container_job_id:0:8}"
+CONTAINER_TMPDIR="/tmp/vllm-${container_job_id_short}"
+CONTAINER_CACHE_ROOT="/tmp/vllm-buildkite-${container_job_id}/cache"
+CONTAINER_PREFLIGHT="mkdir -p \"\$TMPDIR\" \"\$TORCHINDUCTOR_CACHE_DIR\" \"\$TRITON_CACHE_DIR\" \"\$VLLM_CACHE_ROOT\" \"\$XDG_CACHE_HOME\" && python -c \"import encodings, importlib.metadata as im, importlib.util as iu; [im.version(d) for d in ('transformers', 'torch', 'ray', 'sympy', 'markupsafe', 'vllm')]; missing=[m for m in ('torch.utils.model_zoo', 'transformers.models.nomic_bert', 'ray.dag', 'sympy.physics', 'markupsafe._speedups') if iu.find_spec(m) is None]; assert not missing, missing\""
 
 # Verify GPU access
 render_gid=$(getent group render | cut -d: -f3)
@@ -383,6 +510,8 @@ else
     --group-add "$render_gid" \
     --rm \
     -e HF_TOKEN \
+    -e "HF_HUB_DOWNLOAD_TIMEOUT=${HF_HUB_DOWNLOAD_TIMEOUT}" \
+    -e "HF_HUB_ETAG_TIMEOUT=${HF_HUB_ETAG_TIMEOUT}" \
     -e AWS_ACCESS_KEY_ID \
     -e AWS_SECRET_ACCESS_KEY \
     -e BUILDKITE_PARALLEL_JOB \
@@ -390,10 +519,15 @@ else
     -v "${HF_CACHE}:${HF_MOUNT}" \
     -e "HF_HOME=${HF_MOUNT}" \
     -e "PYTHONPATH=${MYPYTHONPATH}" \
+    -e "TMPDIR=${CONTAINER_TMPDIR}/tmp" \
+    -e "TORCHINDUCTOR_CACHE_DIR=${CONTAINER_CACHE_ROOT}/torchinductor" \
+    -e "TRITON_CACHE_DIR=${CONTAINER_CACHE_ROOT}/triton" \
+    -e "VLLM_CACHE_ROOT=${CONTAINER_CACHE_ROOT}/vllm" \
+    -e "XDG_CACHE_HOME=${CONTAINER_CACHE_ROOT}/xdg" \
     -e "PYTORCH_ROCM_ARCH=" \
     --name "${container_name}" \
     "${image_name}" \
-    /bin/bash -c "${commands}"
+    /bin/bash -c "${CONTAINER_PREFLIGHT} && ${commands}"
 
   exit_code=$?
   handle_pytest_exit "$exit_code"

.buildkite/scripts/hardware_ci/run-intel-ci-test.sh

@@ -0,0 +1,51 @@
+#!/bin/bash
+
+set -euo pipefail
+
+test_suite="${1:-}"
+
+if [[ -z "${test_suite}" ]]; then
+  echo "Usage: $0 <example|v1|server>" >&2
+  exit 1
+fi
+
+case "${test_suite}" in
+  example)
+    pip install tblib==3.1.0
+
+    python3 examples/basic/offline_inference/generate.py --model facebook/opt-125m --block-size 64 --enforce-eager
+    python3 examples/basic/offline_inference/generate.py --model facebook/opt-125m --block-size 64 -O3 -cc.cudagraph_mode=NONE
+    python3 examples/basic/offline_inference/generate.py --model facebook/opt-125m --block-size 64 --enforce-eager -tp 2 --distributed-executor-backend mp
+    python3 examples/basic/offline_inference/generate.py --model facebook/opt-125m --block-size 64 --enforce-eager --attention-backend=TRITON_ATTN
+    python3 examples/basic/offline_inference/generate.py --model facebook/opt-125m --block-size 64 --enforce-eager --quantization fp8
+    python3 examples/basic/offline_inference/generate.py --model facebook/opt-125m --block-size 64 --enforce-eager --kv-cache-dtype fp8
+    python3 examples/basic/offline_inference/generate.py --model nvidia/Llama-3.1-8B-Instruct-FP8 --block-size 64 --enforce-eager --quantization modelopt --kv-cache-dtype fp8 --attention-backend TRITON_ATTN --max-model-len 4096
+    python3 examples/basic/offline_inference/generate.py --model superjob/Qwen3-4B-Instruct-2507-GPTQ-Int4 --block-size 64 --enforce-eager --max-model-len 8192
+    python3 examples/basic/offline_inference/generate.py --model ibm-research/PowerMoE-3b --block-size 64 --enforce-eager -tp 2
+    python3 examples/basic/offline_inference/generate.py --model ibm-research/PowerMoE-3b --block-size 64 --enforce-eager -tp 2 --enable-expert-parallel
+    python3 examples/basic/offline_inference/generate.py --model superjob/Qwen3-4B-Instruct-2507-GPTQ-Int4 --max-model-len 8192
+    ;;
+  v1)
+    cd tests
+
+    pytest -v -s v1/core --ignore=v1/core/test_reset_prefix_cache_e2e.py --ignore=v1/core/test_scheduler_e2e.py
+    pytest -v -s v1/engine --ignore=v1/engine/test_output_processor.py
+    pytest -v -s v1/sample --ignore=v1/sample/test_logprobs.py --ignore=v1/sample/test_logprobs_e2e.py -k "not test_topk_only and not test_topp_only and not test_topk_and_topp"
+    pytest -v -s v1/worker --ignore=v1/worker/test_gpu_model_runner.py --ignore=v1/worker/test_worker_memory_snapshot.py
+    pytest -v -s v1/structured_output
+    pytest -v -s v1/test_serial_utils.py
+    pytest -v -s v1/spec_decode --ignore=v1/spec_decode/test_max_len.py --ignore=v1/spec_decode/test_speculators_eagle3.py --ignore=v1/spec_decode/test_acceptance_length.py
+    pytest -v -s v1/kv_connector/unit --ignore=v1/kv_connector/unit/test_multi_connector.py --ignore=v1/kv_connector/unit/test_example_connector.py --ignore=v1/kv_connector/unit/test_lmcache_integration.py --ignore=v1/kv_connector/unit/test_hf3fs_client.py --ignore=v1/kv_connector/unit/test_hf3fs_connector.py --ignore=v1/kv_connector/unit/test_hf3fs_metadata_server.py --ignore=v1/kv_connector/unit/test_offloading_connector.py
+    ;;
+  server)
+    pip install av
+    cd tests
+
+    pytest -v -s entrypoints/multimodal/openai/chat_completion/test_audio_in_video.py
+    pytest -v -s benchmarks/test_serve_cli.py
+    ;;
+  *)
+    echo "Unknown Intel test suite: ${test_suite}" >&2
+    exit 1
+    ;;
+esac

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

@@ -243,8 +243,10 @@ container_name="xpu_${BUILDKITE_COMMIT}_$(tr -dc A-Za-z0-9 < /dev/urandom | head
 
 # ---- Command source selection ----
 commands=""
+commands_source=""
 if [[ -n "${VLLM_TEST_COMMANDS:-}" ]]; then
   commands="${VLLM_TEST_COMMANDS}"
+  commands_source="env"
   echo "Commands sourced from VLLM_TEST_COMMANDS (quoting preserved)"
 elif [[ $# -gt 0 ]]; then
   all_yaml=true
@@ -303,8 +305,12 @@ if [[ -z "$commands" ]]; then
 fi
 
 echo "Raw commands: $commands"
-commands=$(re_quote_pytest_markers "$commands")
-echo "After re-quoting: $commands"
+if [[ "$commands_source" != "env" ]]; then
+  commands=$(re_quote_pytest_markers "$commands")
+  echo "After re-quoting: $commands"
+else
+  echo "Skipping re-quoting for VLLM_TEST_COMMANDS input"
+fi
 commands=$(apply_intel_test_overrides "$commands")
 echo "Final commands: $commands"
 
@@ -324,23 +330,6 @@ IMAGE="${IMAGE_TAG_XPU:-${image_name}}"
 
 echo "Using image: ${IMAGE}"
 
-if docker image inspect "${IMAGE}" >/dev/null 2>&1; then
-  echo "Image already exists locally, skipping pull"
-else
-  echo "Image not found locally, waiting for lock..."
-
-  flock /tmp/docker-pull.lock bash -c "
-    if docker image inspect '${IMAGE}' >/dev/null 2>&1; then
-      echo 'Image already pulled by another runner'
-    else
-      echo 'Pulling image...'
-      timeout 900 docker pull '${IMAGE}'
-    fi
-  "
-
-  echo "Pull step completed"
-fi
-
 remove_docker_container() {
   docker rm -f "${container_name}" || true
 }
@@ -357,9 +346,12 @@ export HF_TOKEN ZE_AFFINITY_MASK
 
 {
   flock 9
-  if ! docker image inspect "${IMAGE}" >/dev/null 2>&1; then
-    echo 'Image missing before container creation, pulling again...'
+  if docker image inspect "${IMAGE}" >/dev/null 2>&1; then
+    echo "Image already exists locally, skipping pull"
+  else
+    echo "Image not found locally, pulling image..."
     timeout 900 docker pull "${IMAGE}"
+    echo "Pull step completed"
   fi
 
   docker create \
@@ -372,6 +364,8 @@ export HF_TOKEN ZE_AFFINITY_MASK
     --entrypoint='' \
     -e HF_TOKEN \
     -e ZE_AFFINITY_MASK \
+    -e BUILDKITE_PARALLEL_JOB \
+    -e BUILDKITE_PARALLEL_JOB_COUNT \
     -e CMDS \
     --name "${container_name}" \
     "${IMAGE}" \

.buildkite/scripts/run-rust-frontend-cargo-ci.sh

@@ -110,6 +110,36 @@ install_uv() {
     | env UV_INSTALL_DIR="$CARGO_HOME/bin" sh
 }
 
+setup_pyo3_python() {
+  local python_version="${PYO3_PYTHON_VERSION:-3.12}"
+
+  log_section "Installing Python ${python_version} for PyO3 tests"
+  uv python install "$python_version"
+  PYO3_PYTHON="$(uv python find \
+    --managed-python \
+    --no-project \
+    --resolve-links \
+    "$python_version")"
+  export PYO3_PYTHON
+
+  local python_libdir
+  python_libdir="$("$PYO3_PYTHON" - <<'PY'
+import pathlib
+import sysconfig
+
+libdir = pathlib.Path(sysconfig.get_config_var("LIBDIR"))
+ldlibrary = sysconfig.get_config_var("LDLIBRARY")
+assert sysconfig.get_config_var("Py_ENABLE_SHARED") == 1
+assert ldlibrary
+assert (libdir / ldlibrary).exists(), libdir / ldlibrary
+print(libdir)
+PY
+)"
+
+  export LD_LIBRARY_PATH="${python_libdir}:${LD_LIBRARY_PATH:-}"
+  export LIBRARY_PATH="${python_libdir}:${LIBRARY_PATH:-}"
+}
+
 run_style_clippy() {
   install_cargo_sort
 
@@ -132,6 +162,7 @@ run_style_clippy() {
 
 run_tests() {
   install_uv
+  setup_pyo3_python
   install_cargo_nextest
 
   log_section "Running cargo nextest"

.buildkite/test-amd.yaml

@@ -88,16 +88,16 @@
 #     - Do NOT remove `VLLM_WORKER_MULTIPROC_METHOD=spawn` setting as ROCm requires this for certain models to function.            #
 #   * [Transformers Nightly Models]: Whisper needs `VLLM_WORKER_MULTIPROC_METHOD=spawn` to avoid deadlock.                          #
 #   * [Plugin Tests (2 GPUs)]:                                                                                                      #
-#     - {`pytest -v -s entrypoints/openai/test_oot_registration.py`}: It needs a clean process                                      #
-#     - {`pytest -v -s models/test_oot_registration.py`}: It needs a clean process                                                  #
-#     - {`pytest -v -s plugins/lora_resolvers`}: Unit tests for in-tree lora resolver plugins                                       #
+#     - {`pytest -v -s plugins_tests/test_oot_registration_online.py`}: It needs a clean process                                    #
+#     - {`pytest -v -s plugins_tests/test_oot_registration_offline.py`}: It needs a clean process                                   #
+#     - {`pytest -v -s plugins_tests/lora_resolvers`}: Unit tests for in-tree lora resolver plugins                                 #
 #   * [LoRA TP (Distributed)]:                                                                                                      #
 #     - There is some Tensor Parallelism related processing logic in LoRA that requires multi-GPU testing for validation.           #
 #     - {`pytest -v -s -x lora/test_gptoss_tp.py`}: Disabled for now because MXFP4 backend on non-cuda platform doesn't support     #
 #                                                   LoRA yet.                                                                       #
 #   * [Distributed Tests (NxGPUs)(HW-TAG)]: Don't test llama model here, it seems hf implementation is buggy. See:                  #
 #                                           https://github.com/vllm-project/vllm/pull/5689                                          #
-#   * [Distributed Tests (NxGPUs)(HW-TAG)]: Some old E2E tests were removed in https://github.com/vllm-project/vllm/pull/33293     #
+#   * [Distributed Tests (NxGPUs)(HW-TAG)]: Some old E2E tests were removed in https://github.com/vllm-project/vllm/pull/33293      #
 #                                           in favor of new tests in fusions_e2e. We avoid replicating the new jobs in              #
 #                                           this file as it's deprecated.                                                           #
 #                                                                                                                                   #
@@ -315,24 +315,6 @@ steps:
   - pytest -v -s distributed/test_pp_cudagraph.py
   - pytest -v -s distributed/test_pipeline_parallel.py
 
-#----------------------------------------------------------  mi250 · engine  -----------------------------------------------------------#
-
-- label: Engine # 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:
-  - vllm/
-  - tests/engine
-  - tests/test_sequence
-  - tests/test_config
-  - tests/test_logger
-  - tests/test_vllm_port
-  commands:
-  - pytest -v -s engine test_sequence.py test_config.py test_logger.py test_vllm_port.py
-
 #-----------------------------------------------------------  mi250 · evals  -----------------------------------------------------------#
 
 - label: Multi-Modal Accuracy Eval (Small Models) # TBD
@@ -416,7 +398,7 @@ steps:
   - tests/kernels/helion/
   - vllm/platforms/rocm.py
   commands:
-  - pip install helion==1.0.0
+  - pip install helion==1.1.0
   - pytest -v -s kernels/helion/
 
 - label: Kernels Mamba Test # TBD
@@ -433,45 +415,6 @@ steps:
   commands:
   - pytest -v -s kernels/mamba
 
-#-----------------------------------------------------------  mi250 · lora  ------------------------------------------------------------#
-
-- label: LoRA %N # TBD
-  timeout_in_minutes: 180
-  mirror_hardwares: [amdexperimental, amdproduction, amdgfx90anightly, amdmi250]
-  agent_pool: mi250_1
-  parallelism: 4
-  optional: true
-  working_dir: "/vllm-workspace/tests"
-  source_file_dependencies:
-  - vllm/lora
-  - tests/lora
-  - vllm/platforms/rocm.py
-  commands:
-  - pytest -v -s lora --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --ignore=lora/test_chatglm3_tp.py --ignore=lora/test_llama_tp.py --ignore=lora/test_qwen3_with_multi_loras.py --ignore=lora/test_olmoe_tp.py --ignore=lora/test_deepseekv2_tp.py --ignore=lora/test_gptoss_tp.py --ignore=lora/test_qwen3moe_tp.py --ignore=lora/test_qwen35_densemodel_lora.py
-
-#------------------------------------------------------  mi250 · model_executor  -------------------------------------------------------#
-
-- label: Model Executor # TBD
-  timeout_in_minutes: 180
-  mirror_hardwares: [amdexperimental, amdproduction, amdgfx90anightly, amdmi250]
-  agent_pool: mi250_1
-  optional: true
-  torch_nightly: true
-  working_dir: "/vllm-workspace/tests"
-  source_file_dependencies:
-  - vllm/engine/arg_utils.py
-  - vllm/config/model.py
-  - vllm/model_executor
-  - tests/model_executor
-  - tests/entrypoints/openai/completion/test_tensorizer_entrypoint.py
-  - vllm/_aiter_ops.py
-  - vllm/platforms/rocm.py
-  commands:
-  - apt-get update && apt-get install -y curl libsodium23
-  - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-  - pytest -v -s model_executor -m '(not slow_test)'
-  - pytest -v -s entrypoints/openai/completion/test_tensorizer_entrypoint.py
-
 #------------------------------------------------------  mi250 · models / basic  -------------------------------------------------------#
 
 - label: Basic Models Test (Other CPU) # TBD
@@ -649,6 +592,11 @@ steps:
   - pytest -v -s plugins_tests/test_bge_m3_sparse_io_processor_plugins.py
   - pip uninstall bge_m3_sparse_plugin -y
   # END: `bge_m3_sparse io_processor` test
+  # BEGIN: `colbert_query io_processor` test
+  - pip install -e ./plugins/colbert_query_plugin
+  - pytest -v -s plugins_tests/test_colbert_query_io_processor_plugins.py
+  - pip uninstall colbert_query_plugin -y
+  # END: `colbert_query io_processor` test
   # BEGIN: `stat_logger` plugins test
   - pip install -e ./plugins/vllm_add_dummy_stat_logger
   - pytest -v -s plugins_tests/test_stats_logger_plugins.py
@@ -658,9 +606,9 @@ steps:
   - pytest -v -s plugins_tests/test_scheduler_plugins.py
   - pip install -e ./plugins/vllm_add_dummy_model
   - pytest -v -s distributed/test_distributed_oot.py
-  - pytest -v -s entrypoints/openai/chat_completion/test_oot_registration.py
-  - pytest -v -s models/test_oot_registration.py
-  - pytest -v -s plugins/lora_resolvers
+  - pytest -v -s plugins_tests/test_oot_registration_online.py # it needs a clean process
+  - pytest -v -s plugins_tests/test_oot_registration_offline.py # it needs a clean process
+  - pytest -v -s plugins_tests/lora_resolvers # unit tests for in-tree lora resolver plugins
 
 #------------------------------------------------------------  mi250 · v1  -------------------------------------------------------------#
 
@@ -845,7 +793,7 @@ steps:
   - vllm/platforms/rocm.py
   commands:
   - uv pip install --system -r /vllm-workspace/requirements/kv_connectors_rocm.txt
-  - ATTENTION_BACKEND=ROCM_ATTN bash v1/kv_connector/nixl_integration/spec_decode_acceptance_test.sh
+  - ATTENTION_BACKEND=TRITON_ATTN bash v1/kv_connector/nixl_integration/spec_decode_acceptance_test.sh
 
 - label: V1 e2e (2 GPUs) # TBD
   timeout_in_minutes: 180
@@ -871,7 +819,7 @@ steps:
   - vllm/platforms/rocm.py
   commands:
   - uv pip install --system -r /vllm-workspace/requirements/kv_connectors_rocm.txt
-  - ROCM_ATTN=1 bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
+  - ATTENTION_BACKEND=TRITON_ATTN bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
 
 #-------------------------------------------------------------  mi250 · misc  ------------------------------------------------------------#
 
@@ -931,10 +879,10 @@ steps:
   - vllm/
   - tests/basic_correctness/test_basic_correctness
   - tests/basic_correctness/test_cpu_offload
-  - tests/basic_correctness/test_cumem.py
+  - tests/basic_correctness/test_mem.py
   commands:
   - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-  - pytest -v -s basic_correctness/test_cumem.py
+  - pytest -v -s basic_correctness/test_mem.py
   - pytest -v -s basic_correctness/test_basic_correctness.py
   - pytest -v -s basic_correctness/test_cpu_offload.py
 
@@ -1228,7 +1176,39 @@ steps:
 
 #--------------------------------------------------------  mi300 · entrypoints  --------------------------------------------------------#
 
-- label: Entrypoints Integration (API Server 2) # TBD
+- label: Entrypoints Unit Tests # TBD
+  timeout_in_minutes: 180
+  mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
+  agent_pool: mi300_1
+  fast_check: true
+  working_dir: "/vllm-workspace/tests"
+  source_file_dependencies:
+  - vllm/entrypoints
+  - tests/entrypoints/unit_tests
+  - tests/entrypoints/weight_transfer
+  - vllm/platforms/rocm.py
+  commands:
+  - pytest -v -s entrypoints/unit_tests
+  - pytest -v -s entrypoints/weight_transfer
+
+- label: Entrypoints Integration (LLM) # TBD
+  timeout_in_minutes: 180
+  mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
+  agent_pool: mi300_1
+  optional: true
+  fast_check: true
+  torch_nightly: true
+  working_dir: "/vllm-workspace/tests"
+  source_file_dependencies:
+  - vllm/
+  - tests/entrypoints/llm
+  commands:
+  - export VLLM_WORKER_MULTIPROC_METHOD=spawn
+  - pytest -v -s entrypoints/llm --ignore=entrypoints/llm/test_generate.py --ignore=entrypoints/llm/test_collective_rpc.py --ignore=entrypoints/llm/offline_mode
+  - pytest -v -s entrypoints/llm/test_generate.py # it needs a clean process
+  - pytest -v -s entrypoints/llm/offline_mode # Needs to avoid interference with other tests
+
+- label: Entrypoints Integration (API Server) # TBD
   timeout_in_minutes: 180
   mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
   agent_pool: mi300_1
@@ -1244,7 +1224,7 @@ steps:
   - pytest -v -s entrypoints/serve --ignore=entrypoints/serve/dev/rpc
   - PYTHONPATH=/vllm-workspace pytest -v -s entrypoints/serve/dev/rpc
 
-- label: Entrypoints Integration (API Server openai - Part 1) # TBD
+- label: Entrypoints Integration (API Server OpenAI - Part 1) # TBD
   timeout_in_minutes: 180
   mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
   agent_pool: mi300_1
@@ -1258,90 +1238,45 @@ steps:
   - tests/entrypoints/test_chat_utils
   commands:
   - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-  - pytest -v -s entrypoints/openai/chat_completion --ignore=entrypoints/openai/chat_completion/test_oot_registration.py
+  - pytest -v -s entrypoints/openai/ --ignore=entrypoints/openai/completion --ignore=entrypoints/openai/chat_completion --ignore=entrypoints/openai/responses --ignore=entrypoints/openai/correctness
 
-- label: Entrypoints Integration (API Server openai - Part 2) # TBD
+- label: Entrypoints Integration (API Server OpenAI - Part 2) # TBD
   timeout_in_minutes: 180
   mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
   agent_pool: mi300_1
-  optional: true
   fast_check: true
   torch_nightly: true
+  optional: true
   working_dir: "/vllm-workspace/tests"
   source_file_dependencies:
   - vllm/
   - tests/entrypoints/openai
   - tests/entrypoints/test_chat_utils
-  - tests/entrypoints/generate
-  - tests/tool_use
   commands:
   - export VLLM_WORKER_MULTIPROC_METHOD=spawn
+  - pytest -v -s entrypoints/openai/chat_completion
   - pytest -v -s entrypoints/openai/completion --ignore=entrypoints/openai/completion/test_tensorizer_entrypoint.py
-  - pytest -v -s entrypoints/test_chat_utils.py
-  - pytest -v -s entrypoints/generate
+
+- label: Entrypoints Integration (API Server Generate) # TBD
+  timeout_in_minutes: 180
+  mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
+  agent_pool: mi300_1
+  optional: true
+  fast_check: true
+  torch_nightly: true
+  working_dir: "/vllm-workspace/tests"
+  source_file_dependencies:
+  - vllm/
+  - tests/tool_use
+  - tests/entrypoints/tool_parsers
+  - tests/entrypoints/anthropic
+  - tests/entrypoints/generate
+  commands:
+  - export VLLM_WORKER_MULTIPROC_METHOD=spawn
   - pytest -v -s tool_use
-
-- label: Entrypoints Integration (API Server openai - Part 3) # TBD
-  timeout_in_minutes: 180
-  mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
-  agent_pool: mi300_1
-  optional: true
-  fast_check: true
-  torch_nightly: true
-  working_dir: "/vllm-workspace/tests"
-  source_file_dependencies:
-  - vllm/
-  - tests/entrypoints/openai
-  - tests/entrypoints/test_chat_utils
-  commands:
-  - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-  - pytest -v -s entrypoints/openai --ignore=entrypoints/openai/chat_completion --ignore=entrypoints/openai/completion --ignore=entrypoints/openai/correctness/ --ignore=entrypoints/openai/tool_parsers/ --ignore=entrypoints/openai/responses --ignore=entrypoints/openai/test_multi_api_servers.py
-
-- label: Entrypoints Integration (Speech to Text) # TBD
-  timeout_in_minutes: 180
-  mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
-  agent_pool: mi300_1
-  fast_check: true
-  torch_nightly: true
-  working_dir: "/vllm-workspace/tests"
-  source_file_dependencies:
-  - vllm/
-  - tests/entrypoints/speech_to_text
-  commands:
-  - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-  - pytest -v -s entrypoints/speech_to_text
-
-- label: Entrypoints Integration (LLM) # TBD
-  timeout_in_minutes: 180
-  mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
-  agent_pool: mi300_1
-  optional: true
-  fast_check: true
-  torch_nightly: true
-  working_dir: "/vllm-workspace/tests"
-  source_file_dependencies:
-  - vllm/
-  - tests/entrypoints/llm
-  - tests/entrypoints/offline_mode
-  commands:
-  - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-  - pytest -v -s entrypoints/llm --ignore=entrypoints/llm/test_generate.py --ignore=entrypoints/llm/test_collective_rpc.py
-  - pytest -v -s entrypoints/llm/test_generate.py
-  - pytest -v -s entrypoints/offline_mode
-
-- label: Entrypoints Integration (Pooling) # TBD
-  timeout_in_minutes: 180
-  mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
-  agent_pool: mi300_1
-  fast_check: true
-  torch_nightly: true
-  working_dir: "/vllm-workspace/tests"
-  source_file_dependencies:
-  - vllm/
-  - tests/entrypoints/pooling
-  commands:
-  - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-  - pytest -v -s entrypoints/pooling
+  - pytest -v -s entrypoints/tool_parsers
+  - pytest -v -s entrypoints/generate
+  - pytest -v -s entrypoints/anthropic
 
 - label: Entrypoints Integration (Responses API) # TBD
   timeout_in_minutes: 180
@@ -1357,29 +1292,57 @@ steps:
   - export VLLM_WORKER_MULTIPROC_METHOD=spawn
   - pytest -v -s entrypoints/openai/responses
 
-- label: Entrypoints Unit Tests # TBD
+- label: Entrypoints Integration (Speech to Text) # TBD
   timeout_in_minutes: 180
   mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
   agent_pool: mi300_1
   fast_check: true
+  torch_nightly: true
   working_dir: "/vllm-workspace/tests"
   source_file_dependencies:
-  - vllm/entrypoints
-  - tests/entrypoints/
-  - vllm/platforms/rocm.py
+  - vllm/
+  - tests/entrypoints/speech_to_text
   commands:
-  - pytest -v -s entrypoints/openai/tool_parsers
-  - pytest -v -s entrypoints/ --ignore=entrypoints/llm --ignore=entrypoints/offline_mode --ignore=entrypoints/openai --ignore=entrypoints/serve --ignore=entrypoints/test_chat_utils.py  --ignore=entrypoints/pooling --ignore=entrypoints/speech_to_text --ignore=tests/entrypoints/generate
+  - export VLLM_WORKER_MULTIPROC_METHOD=spawn
+  - pytest -v -s entrypoints/speech_to_text
+
+- label: Entrypoints Integration (Multimodal)
+  timeout_in_minutes: 180
+  mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
+  agent_pool: mi300_1
+  fast_check: true
+  torch_nightly: true
+  working_dir: "/vllm-workspace/tests"
+  source_file_dependencies:
+  - vllm/
+  - tests/entrypoints/multimodal
+  commands:
+  - export VLLM_WORKER_MULTIPROC_METHOD=spawn
+  - pytest -v -s entrypoints/multimodal
+
+- label: Entrypoints Integration (Pooling) # TBD
+  timeout_in_minutes: 180
+  mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
+  agent_pool: mi300_1
+  fast_check: true
+  torch_nightly: true
+  working_dir: "/vllm-workspace/tests"
+  source_file_dependencies:
+  - vllm/
+  - tests/entrypoints/pooling
+  commands:
+  - export VLLM_WORKER_MULTIPROC_METHOD=spawn
+  - pytest -v -s entrypoints/pooling
 
 - label: OpenAI API correctness # TBD
   timeout_in_minutes: 180
   mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
   agent_pool: mi300_1
+  optional: true
   working_dir: "/vllm-workspace/tests"
   source_file_dependencies:
   - csrc/
   - vllm/entrypoints/openai/
-  - vllm/model_executor/models/whisper.py
   - vllm/model_executor/layers/
   - vllm/v1/attention/backends/
   - vllm/v1/attention/selector.py
@@ -1433,6 +1396,7 @@ steps:
   timeout_in_minutes: 180
   mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
   agent_pool: mi300_1
+  optional: true
   working_dir: "/vllm-workspace/.buildkite/lm-eval-harness"
   source_file_dependencies:
   - csrc/
@@ -1719,6 +1683,20 @@ steps:
 
 #-----------------------------------------------------------  mi300 · lora  ------------------------------------------------------------#
 
+- label: LoRA %N # TBD
+  timeout_in_minutes: 180
+  mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
+  agent_pool: mi300_1
+  parallelism: 4
+  optional: true
+  working_dir: "/vllm-workspace/tests"
+  source_file_dependencies:
+  - vllm/lora
+  - tests/lora
+  - vllm/platforms/rocm.py
+  commands:
+  - pytest -v -s lora --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --ignore=lora/test_chatglm3_tp.py --ignore=lora/test_llama_tp.py --ignore=lora/test_qwen3_with_multi_loras.py --ignore=lora/test_olmoe_tp.py --ignore=lora/test_deepseekv2_tp.py --ignore=lora/test_gptoss_tp.py --ignore=lora/test_qwen3moe_tp.py --ignore=lora/test_qwen35_densemodel_lora.py
+
 - label: LoRA TP (Distributed) # TBD
   timeout_in_minutes: 180
   mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
@@ -1739,6 +1717,29 @@ steps:
   - pytest -v -s -x lora/test_gptoss_tp.py
   - pytest -v -s -x lora/test_qwen35_densemodel_lora.py
 
+#------------------------------------------------------  mi300 · model_executor  -------------------------------------------------------#
+
+- label: Model Executor # TBD
+  timeout_in_minutes: 180
+  mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
+  agent_pool: mi300_1
+  optional: true
+  torch_nightly: true
+  working_dir: "/vllm-workspace/tests"
+  source_file_dependencies:
+  - vllm/engine/arg_utils.py
+  - vllm/config/model.py
+  - vllm/model_executor
+  - tests/model_executor
+  - tests/entrypoints/openai/completion/test_tensorizer_entrypoint.py
+  - vllm/_aiter_ops.py
+  - vllm/platforms/rocm.py
+  commands:
+  - apt-get update && apt-get install -y curl libsodium23
+  - export VLLM_WORKER_MULTIPROC_METHOD=spawn
+  - pytest -v -s model_executor -m '(not slow_test)'
+  - pytest -v -s entrypoints/openai/completion/test_tensorizer_entrypoint.py
+
 #-----------------------------------------------------  mi300 · models / language  -----------------------------------------------------#
 
 - label: Language Models Test (Extended Pooling)  # TBD
@@ -1780,10 +1781,9 @@ steps:
   - tests/models/multimodal/generation
   - tests/models/multimodal/test_mapping.py
   commands:
-  - uv pip install --system --no-build-isolation 'git+https://github.com/AndreasKaratzas/mamba@rocm-7.0-v2.3.0'
-  - uv pip install --system --no-build-isolation 'git+https://github.com/Dao-AILab/causal-conv1d@v1.6.0'
-  - pytest -v -s models/language/generation -m hybrid_model --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --shard-id=$$BUILDKITE_PARALLEL_JOB
-
+  - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
+  - pytest -v -s models/multimodal/generation -m 'not core_model' --ignore models/multimodal/generation/test_common.py
+  - pytest -v -s models/multimodal/test_mapping.py
 
 - label: Multi-Modal Models (Extended Generation 2) # TBD
   timeout_in_minutes: 180
@@ -1795,9 +1795,8 @@ steps:
   - vllm/
   - tests/models/multimodal/generation
   commands:
-  - uv pip install --system --no-build-isolation 'git+https://github.com/AndreasKaratzas/mamba@rocm-7.0-v2.3.0'
-  - uv pip install --system --no-build-isolation 'git+https://github.com/Dao-AILab/causal-conv1d@v1.6.0'
-  - pytest -v -s models/language/generation -m '(not core_model) and (not hybrid_model)'
+  - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
+  - pytest -v -s models/multimodal/generation/test_common.py -m 'split(group=0) and not core_model'
 
 
 - label: Multi-Modal Models (Extended Generation 3) # TBD
@@ -2184,10 +2183,72 @@ steps:
   commands:
   - pytest -v -s v1/e2e/spec_decode -k "speculators or mtp_correctness"
 
+- label: Speculators Correctness # TBD
+  timeout_in_minutes: 180
+  mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
+  agent_pool: mi300_1
+  optional: true
+  working_dir: "/vllm-workspace/tests"
+  source_file_dependencies:
+  - vllm/config/speculative.py
+  - vllm/engine/arg_utils.py
+  - vllm/transformers_utils/config.py
+  - vllm/transformers_utils/configs/speculators/
+  - vllm/v1/spec_decode/
+  - vllm/v1/worker/gpu/spec_decode/
+  - vllm/v1/worker/gpu_model_runner.py
+  - vllm/v1/sample/
+  - vllm/v1/attention/backends/
+  - vllm/v1/attention/selector.py
+  - vllm/model_executor/model_loader/
+  - vllm/model_executor/layers/
+  - vllm/model_executor/models/llama_eagle3.py
+  - vllm/model_executor/models/qwen3.py
+  - vllm/model_executor/models/qwen3_dflash.py
+  - vllm/model_executor/models/registry.py
+  - vllm/_aiter_ops.py
+  - tests/evals/gsm8k/
+  - tests/v1/spec_decode/test_speculators_correctness.py
+  - vllm/platforms/rocm.py
+  commands:
+  - export VLLM_ALLOW_INSECURE_SERIALIZATION=1
+  - pytest -v -s v1/spec_decode/test_speculators_correctness.py -m slow_test
+
+- label: Extract Hidden States Integration # TBD
+  timeout_in_minutes: 180
+  mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
+  agent_pool: mi300_1
+  working_dir: "/vllm-workspace/tests"
+  source_file_dependencies:
+  - vllm/config/speculative.py
+  - vllm/distributed/kv_transfer/kv_connector/
+  - vllm/model_executor/layers/attention/
+  - vllm/model_executor/layers/mamba/
+  - vllm/model_executor/model_loader/
+  - vllm/model_executor/models/extract_hidden_states.py
+  - vllm/model_executor/models/llama.py
+  - vllm/model_executor/models/qwen3_5.py
+  - vllm/model_executor/models/qwen3_next.py
+  - vllm/model_executor/models/registry.py
+  - vllm/transformers_utils/configs/extract_hidden_states.py
+  - vllm/transformers_utils/configs/qwen3_5.py
+  - vllm/v1/attention/backends/
+  - vllm/v1/attention/selector.py
+  - vllm/v1/kv_cache_interface.py
+  - vllm/v1/spec_decode/extract_hidden_states.py
+  - vllm/v1/worker/gpu_model_runner.py
+  - vllm/_aiter_ops.py
+  - tests/v1/kv_connector/extract_hidden_states_integration/
+  - vllm/platforms/rocm.py
+  commands:
+  - export VLLM_WORKER_MULTIPROC_METHOD=spawn
+  - pytest -v -s v1/kv_connector/extract_hidden_states_integration
+
 - label: V1 attention (H100-MI300) # TBD
   timeout_in_minutes: 180
   mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
   agent_pool: mi300_1
+  optional: true
   working_dir: "/vllm-workspace/tests"
   source_file_dependencies:
   - vllm/config/attention.py
@@ -2346,7 +2407,7 @@ steps:
   - vllm/platforms/rocm.py
   commands:
   - uv pip install --system -r /vllm-workspace/requirements/kv_connectors_rocm.txt
-  - CROSS_LAYERS_BLOCKS=True ROCM_ATTN=1 bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
+  - CROSS_LAYERS_BLOCKS=True ATTENTION_BACKEND=TRITON_ATTN bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
 
 - label: Distributed DP Tests (4 GPUs) # TBD
   timeout_in_minutes: 180
@@ -2382,7 +2443,7 @@ steps:
   - vllm/platforms/rocm.py
   commands:
   - uv pip install --system -r /vllm-workspace/requirements/kv_connectors_rocm.txt
-  - ROCM_ATTN=1 bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
+  - ATTENTION_BACKEND=TRITON_ATTN bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
 
 - label: DP EP Distributed NixlConnector PD accuracy tests (4 GPUs) # TBD
   timeout_in_minutes: 180
@@ -2396,7 +2457,7 @@ steps:
   - vllm/platforms/rocm.py
   commands:
   - uv pip install --system -r /vllm-workspace/requirements/kv_connectors_rocm.txt
-  - DP_EP=1 ROCM_ATTN=1 bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
+  - DP_EP=1 ATTENTION_BACKEND=TRITON_ATTN bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
 
 - label: Hybrid SSM NixlConnector PD accuracy tests (4 GPUs) # TBD
   timeout_in_minutes: 180
@@ -2410,7 +2471,7 @@ steps:
   - vllm/platforms/rocm.py
   commands:
   - uv pip install --system -r /vllm-workspace/requirements/kv_connectors_rocm.txt
-  - HYBRID_SSM=1 ROCM_ATTN=1 bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
+  - HYBRID_SSM=1 ATTENTION_BACKEND=TRITON_ATTN bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
 
 - label: V1 e2e (4 GPUs) # TBD
   timeout_in_minutes: 180
@@ -2546,6 +2607,7 @@ steps:
   timeout_in_minutes: 180
   mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi325]
   agent_pool: mi325_1
+  optional: true
   working_dir: "/vllm-workspace/tests"
   source_file_dependencies:
   - vllm/
@@ -2555,7 +2617,7 @@ steps:
   - tests/test_logger
   - tests/test_vllm_port
   commands:
-  - pytest -v -s engine test_sequence.py test_config.py test_logger.py test_vllm_port.py
+  - pytest -v -s engine test_sequence.py test_config.py test_logger.py test_vllm_port.py test_jit_monitor.py
 
 #-----------------------------------------------------------  mi325 · evals  -----------------------------------------------------------#
 
@@ -2637,6 +2699,7 @@ steps:
   timeout_in_minutes: 180
   mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi325]
   agent_pool: mi325_1
+  optional: true
   working_dir: "/vllm-workspace/tests"
   source_file_dependencies:
   - vllm/
@@ -2692,7 +2755,7 @@ steps:
   - vllm/_aiter_ops.py
   - vllm/platforms/rocm.py
   commands:
-  - python3 benchmarks/attention_benchmarks/benchmark.py --backends ROCM_ATTN ROCM_AITER_FA ROCM_AITER_UNIFIED_ATTN --batch-specs "8q1s1k" --repeats 1 --warmup-iters 1
+  - python3 benchmarks/attention_benchmarks/benchmark.py --backends ROCM_ATTN ROCM_AITER_FA ROCM_AITER_UNIFIED_ATTN --batch-specs "8q1s1k"
 
 #--------------------------------------------------------  mi355 · distributed  --------------------------------------------------------#
 
@@ -2736,7 +2799,7 @@ steps:
 
 #--------------------------------------------------------  mi355 · entrypoints  --------------------------------------------------------#
 
-- label: Entrypoints Integration (API Server 2) # TBD
+- label: Entrypoints Integration (API Server) # TBD
   timeout_in_minutes: 180
   mirror_hardwares: [amdexperimental, amdproduction, amdgfx950nightly, amdmi355]
   agent_pool: mi355_1
@@ -2752,7 +2815,7 @@ steps:
   - pytest -v -s entrypoints/serve --ignore=entrypoints/serve/dev/rpc
   - PYTHONPATH=/vllm-workspace pytest -v -s entrypoints/serve/dev/rpc
 
-- label: Entrypoints Integration (API Server openai - Part 1) # TBD
+- label: Entrypoints Integration (API Server OpenAI - Part 1) # TBD
   timeout_in_minutes: 180
   mirror_hardwares: [amdexperimental, amdproduction, amdgfx950nightly, amdmi355]
   agent_pool: mi355_1
@@ -2766,9 +2829,9 @@ steps:
   - tests/entrypoints/test_chat_utils
   commands:
   - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-  - pytest -v -s entrypoints/openai/chat_completion --ignore=entrypoints/openai/chat_completion/test_oot_registration.py
+  - pytest -v -s entrypoints/openai --ignore=entrypoints/openai/completion --ignore=entrypoints/openai/chat_completion --ignore=entrypoints/openai/responses --ignore=entrypoints/openai/correctness
 
-- label: Entrypoints Integration (API Server openai - Part 2) # TBD
+- label: Entrypoints Integration (API Server OpenAI - Part 2) # TBD
   timeout_in_minutes: 180
   mirror_hardwares: [amdexperimental, amdproduction, amdgfx950nightly, amdmi355]
   agent_pool: mi355_1
@@ -2780,29 +2843,31 @@ steps:
   - vllm/
   - tests/entrypoints/openai
   - tests/entrypoints/test_chat_utils
-  - tests/entrypoints/generate
-  - tests/tool_use
   commands:
   - export VLLM_WORKER_MULTIPROC_METHOD=spawn
+  - pytest -v -s entrypoints/openai/chat_completion
   - pytest -v -s entrypoints/openai/completion --ignore=entrypoints/openai/completion/test_tensorizer_entrypoint.py
-  - pytest -v -s entrypoints/test_chat_utils.py
-  - pytest -v -s entrypoints/generate
-  - pytest -v -s tool_use
 
-- label: Entrypoints Integration (API Server openai - Part 3) # TBD
+- label: Entrypoints Integration (API Server Generate) # TBD
   timeout_in_minutes: 180
   mirror_hardwares: [amdexperimental, amdproduction, amdgfx950nightly, amdmi355]
   agent_pool: mi355_1
   fast_check: true
   torch_nightly: true
+  optional: true
   working_dir: "/vllm-workspace/tests"
   source_file_dependencies:
   - vllm/
-  - tests/entrypoints/openai
-  - tests/entrypoints/test_chat_utils
+  - tests/tool_use
+  - tests/entrypoints/tool_parsers
+  - tests/entrypoints/anthropic
+  - tests/entrypoints/generate
   commands:
   - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-  - pytest -v -s entrypoints/openai --ignore=entrypoints/openai/chat_completion --ignore=entrypoints/openai/completion --ignore=entrypoints/openai/correctness/ --ignore=entrypoints/openai/tool_parsers/ --ignore=entrypoints/openai/responses --ignore=entrypoints/openai/test_multi_api_servers.py
+  - pytest -v -s tool_use
+  - pytest -v -s entrypoints/tool_parsers
+  - pytest -v -s entrypoints/generate
+  - pytest -v -s entrypoints/anthropic
 
 - label: Entrypoints Integration (Speech to Text) # TBD
   timeout_in_minutes: 180
@@ -2818,6 +2883,20 @@ steps:
   - export VLLM_WORKER_MULTIPROC_METHOD=spawn
   - pytest -v -s entrypoints/speech_to_text
 
+- label: Entrypoints Integration (Multimodal)
+  timeout_in_minutes: 180
+  mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi355]
+  agent_pool: mi355_1
+  fast_check: true
+  torch_nightly: true
+  working_dir: "/vllm-workspace/tests"
+  source_file_dependencies:
+  - vllm/
+  - tests/entrypoints/multimodal
+  commands:
+  - export VLLM_WORKER_MULTIPROC_METHOD=spawn
+  - pytest -v -s entrypoints/multimodal
+
 - label: Entrypoints Integration (Pooling) # TBD
   timeout_in_minutes: 180
   mirror_hardwares: [amdexperimental, amdproduction, amdgfx950nightly, amdmi355]
@@ -2868,7 +2947,6 @@ steps:
   - vllm/model_executor/models/qwen3_5_mtp.py
   - vllm/transformers_utils/configs/qwen3_5.py
   - vllm/transformers_utils/configs/qwen3_5_moe.py
-  - vllm/model_executor/models/qwen.py
   - vllm/model_executor/models/qwen2.py
   - vllm/model_executor/models/qwen3.py
   - vllm/model_executor/models/qwen3_next.py
@@ -2995,7 +3073,7 @@ steps:
   - vllm/_aiter_ops.py
   commands:
   - rocm-smi
-  - python3 examples/basic/offline_inference/chat.py
+  - python3 examples/basic/offline_inference/chat.py --attention-backend TRITON_ATTN
   - pytest -v -s tests/kernels/attention/test_attention_selector.py
 
 - label: Kernels Attention Test %N # TBD
@@ -3106,7 +3184,6 @@ steps:
   - vllm/model_executor/models/qwen3_5_mtp.py
   - vllm/transformers_utils/configs/qwen3_5.py
   - vllm/transformers_utils/configs/qwen3_5_moe.py
-  - vllm/model_executor/models/qwen.py
   - vllm/model_executor/models/qwen2.py
   - vllm/model_executor/models/qwen3.py
   - vllm/model_executor/models/qwen3_next.py
@@ -3359,7 +3436,7 @@ steps:
   - vllm/platforms/rocm.py
   commands:
   - uv pip install --system -r /vllm-workspace/requirements/kv_connectors_rocm.txt
-  - ATTENTION_BACKEND=ROCM_ATTN bash v1/kv_connector/nixl_integration/spec_decode_acceptance_test.sh
+  - ATTENTION_BACKEND=TRITON_ATTN bash v1/kv_connector/nixl_integration/spec_decode_acceptance_test.sh
 
 - label: Distributed NixlConnector PD accuracy (4 GPUs) # TBD
   timeout_in_minutes: 180
@@ -3374,7 +3451,7 @@ steps:
   - vllm/platforms/rocm.py
   commands:
   - uv pip install --system -r /vllm-workspace/requirements/kv_connectors_rocm.txt
-  - ROCM_ATTN=1 bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
+  - ATTENTION_BACKEND=TRITON_ATTN bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
 
 - label: DP EP Distributed NixlConnector PD accuracy tests (4 GPUs) # TBD
   timeout_in_minutes: 180
@@ -3389,7 +3466,7 @@ steps:
   - vllm/platforms/rocm.py
   commands:
   - uv pip install --system -r /vllm-workspace/requirements/kv_connectors_rocm.txt
-  - DP_EP=1 ROCM_ATTN=1 bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
+  - DP_EP=1 ATTENTION_BACKEND=TRITON_ATTN bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
 
 #------------------------------------------------------  mi355 · weight_loading  -------------------------------------------------------#
 

.buildkite/test_areas/attention.yaml

@@ -2,8 +2,8 @@ group: Attention
 depends_on: 
   - image-build
 steps:
-- label: V1 attention (H100)
-  key: v1-attention-h100
+- label: V1 attention (H100-MI300)
+  key: v1-attention-h100-mi300
   timeout_in_minutes: 30
   device: h100
   source_file_dependencies:
@@ -13,6 +13,20 @@ steps:
     - tests/v1/attention
   commands:
     - pytest -v -s v1/attention
+  mirror:
+    amd:
+      device: mi325_1
+      timeout_in_minutes: 70
+      depends_on:
+      - image-build-amd
+      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
 
 - label: V1 attention (B200)
   key: v1-attention-b200

.buildkite/test_areas/basic_correctness.yaml

@@ -10,9 +10,9 @@ steps:
   - vllm/
   - tests/basic_correctness/test_basic_correctness
   - tests/basic_correctness/test_cpu_offload
-  - tests/basic_correctness/test_cumem.py
+  - tests/basic_correctness/test_mem.py
   commands:
   - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-  - pytest -v -s basic_correctness/test_cumem.py
+  - pytest -v -s basic_correctness/test_mem.py
   - pytest -v -s basic_correctness/test_basic_correctness.py
   - pytest -v -s basic_correctness/test_cpu_offload.py

.buildkite/test_areas/benchmarks.yaml

@@ -23,4 +23,4 @@ steps:
   - benchmarks/attention_benchmarks/
   - vllm/v1/attention/
   commands:
-  - python3 benchmarks/attention_benchmarks/benchmark.py --backends flash flashinfer --batch-specs "8q1s1k" --repeats 1 --warmup-iters 1
+  - python3 benchmarks/attention_benchmarks/benchmark.py --backends flash flashinfer --batch-specs "8q1s1k"

.buildkite/test_areas/disaggregated.yaml

@@ -61,6 +61,20 @@ steps:
     - bash /vllm-workspace/.buildkite/scripts/install-kv-connectors.sh
     - HYBRID_SSM=1 bash v1/kv_connector/nixl_integration/config_sweep_accuracy_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
+  working_dir: "/vllm-workspace/tests"
+  num_devices: 2
+  source_file_dependencies:
+    - vllm/distributed/kv_transfer/kv_connector/v1/nixl/
+    - vllm/v1/core/sched/
+    - vllm/v1/core/kv_cache_coordinator.py
+    - tests/v1/kv_connector/nixl_integration/
+  commands:
+    - bash /vllm-workspace/.buildkite/scripts/install-kv-connectors.sh
+    - bash v1/kv_connector/nixl_integration/run_mamba_prefix_cache_test.sh
+
 - label: MultiConnector (Nixl+Offloading) PD accuracy (2 GPUs)
   key: multiconnector-nixl-offloading-pd-accuracy-2-gpus
   timeout_in_minutes: 30

.buildkite/test_areas/engine.yaml

@@ -26,6 +26,12 @@ steps:
   - tests/test_jit_monitor.py
   commands:
   - pytest -v -s engine test_sequence.py test_config.py test_logger.py test_vllm_port.py test_jit_monitor.py
+  mirror:
+    amd:
+      device: mi325_1
+      timeout_in_minutes: 60
+      depends_on:
+      - image-build-amd
 
 - label: Engine (1 GPU)
   key: engine-1-gpu

.buildkite/test_areas/entrypoints.yaml

@@ -8,10 +8,11 @@ steps:
   working_dir: "/vllm-workspace/tests"
   source_file_dependencies:
   - vllm/entrypoints
-  - tests/entrypoints/
+  - tests/entrypoints/unit_tests
+  - tests/entrypoints/weight_transfer
   commands:
-  - pytest -v -s entrypoints/openai/tool_parsers
-  - pytest -v -s entrypoints/ --ignore=entrypoints/llm --ignore=entrypoints/offline_mode --ignore=entrypoints/openai --ignore=entrypoints/serve --ignore=entrypoints/test_chat_utils.py  --ignore=entrypoints/pooling --ignore=entrypoints/speech_to_text --ignore=tests/entrypoints/generate
+  - pytest -v -s entrypoints/unit_tests
+  - pytest -v -s entrypoints/weight_transfer
 
 - label: Entrypoints Integration (LLM)
   key: entrypoints-integration-llm
@@ -20,84 +21,20 @@ steps:
   source_file_dependencies:
   - vllm/
   - tests/entrypoints/llm
-  - tests/entrypoints/offline_mode
   commands:
   - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-  - pytest -v -s entrypoints/llm --ignore=entrypoints/llm/test_generate.py --ignore=entrypoints/llm/test_collective_rpc.py
+  - pytest -v -s entrypoints/llm --ignore=entrypoints/llm/test_generate.py --ignore=entrypoints/llm/test_collective_rpc.py --ignore=entrypoints/llm/offline_mode
   - pytest -v -s entrypoints/llm/test_generate.py # it needs a clean process
-  - pytest -v -s entrypoints/offline_mode # Needs to avoid interference with other tests
+  - pytest -v -s entrypoints/llm/offline_mode # Needs to avoid interference with other tests
   mirror:
     amd:
       device: mi325_1
-      soft_fail: true
       depends_on:
       - image-build-amd
 
-- label: Entrypoints Integration (API Server openai - Part 1)
-  key: entrypoints-integration-api-server-openai-part-1
-  timeout_in_minutes: 50
-  working_dir: "/vllm-workspace/tests"
-  source_file_dependencies:
-  - vllm/
-  - tests/entrypoints/openai
-  - tests/entrypoints/test_chat_utils
-  commands:
-  - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-  - pytest -v -s entrypoints/openai/chat_completion --ignore=entrypoints/openai/chat_completion/test_oot_registration.py
-  mirror:
-    amd:
-      device: mi325_1
-      soft_fail: true
-      timeout_in_minutes: 80
-      depends_on:
-      - image-build-amd
-
-- label: Entrypoints Integration (API Server openai - Part 2)
-  key: entrypoints-integration-api-server-openai-part-2
-  timeout_in_minutes: 50
-  working_dir: "/vllm-workspace/tests"
-  source_file_dependencies:
-  - vllm/
-  - tests/entrypoints/openai
-  - tests/entrypoints/test_chat_utils
-  - tests/entrypoints/generate
-  - tests/tool_use
-  commands:
-  - pytest -v -s entrypoints/openai/completion --ignore=entrypoints/openai/completion/test_tensorizer_entrypoint.py
-  - pytest -v -s entrypoints/test_chat_utils.py
-  - pytest -v -s entrypoints/generate
-  - pytest -v -s tool_use
-  mirror:
-    amd:
-      device: mi325_1
-      soft_fail: true
-      timeout_in_minutes: 60
-      depends_on:
-      - image-build-amd
-
-- label: Entrypoints Integration (API Server openai - Part 3)
-  key: entrypoints-integration-api-server-openai-part-3
-  timeout_in_minutes: 50
-  device: h200_18gb
-  working_dir: "/vllm-workspace/tests"
-  source_file_dependencies:
-  - vllm/
-  - tests/entrypoints/openai
-  - tests/entrypoints/test_chat_utils
-  commands:
-  - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-  - pytest -v -s entrypoints/openai --ignore=entrypoints/openai/chat_completion --ignore=entrypoints/openai/completion --ignore=entrypoints/openai/correctness/ --ignore=entrypoints/openai/tool_parsers/ --ignore=entrypoints/openai/responses --ignore=entrypoints/openai/test_multi_api_servers.py
-  mirror:
-    amd:
-      device: mi325_1
-      soft_fail: true
-      timeout_in_minutes: 60
-      depends_on:
-      - image-build-amd
-
-- label: Entrypoints Integration (API Server 2)
+- label: Entrypoints Integration (API Server)
+  key: entrypoints-integration-api-server
   device: h200_35gb
-  key: entrypoints-integration-api-server-2
   timeout_in_minutes: 130
   working_dir: "/vllm-workspace/tests"
   source_file_dependencies:
@@ -110,10 +47,78 @@ steps:
   mirror:
     amd:
       device: mi325_1
-      soft_fail: true
       depends_on:
       - image-build-amd
 
+- label: Entrypoints Integration (API Server OpenAI - Part 1)
+  key: entrypoints-integration-api-server-openai-part-1
+  timeout_in_minutes: 50
+  working_dir: "/vllm-workspace/tests"
+  source_file_dependencies:
+  - vllm/
+  - tests/entrypoints/openai
+  - tests/entrypoints/test_chat_utils
+  commands:
+  - export VLLM_WORKER_MULTIPROC_METHOD=spawn
+  - pytest -v -s entrypoints/openai --ignore=entrypoints/openai/completion --ignore=entrypoints/openai/chat_completion --ignore=entrypoints/openai/responses --ignore=entrypoints/openai/correctness
+  mirror:
+    amd:
+      device: mi325_1
+      timeout_in_minutes: 80
+      depends_on:
+      - image-build-amd
+
+- label: Entrypoints Integration (API Server OpenAI - Part 2)
+  key: entrypoints-integration-api-server-openai-part-2
+  timeout_in_minutes: 50
+  working_dir: "/vllm-workspace/tests"
+  source_file_dependencies:
+  - vllm/
+  - tests/entrypoints/openai
+  - tests/entrypoints/test_chat_utils
+  commands:
+  - export VLLM_WORKER_MULTIPROC_METHOD=spawn
+  - pytest -v -s entrypoints/openai/chat_completion
+  - pytest -v -s entrypoints/openai/completion --ignore=entrypoints/openai/completion/test_tensorizer_entrypoint.py
+  mirror:
+    amd:
+      device: mi325_1
+      timeout_in_minutes: 80
+      depends_on:
+      - image-build-amd
+
+- label: Entrypoints Integration (API Server Generate)
+  key: entrypoints-integration-api-server-generate
+  timeout_in_minutes: 50
+  working_dir: "/vllm-workspace/tests"
+  source_file_dependencies:
+  - vllm/
+  - tests/tool_use
+  - tests/entrypoints/tool_parsers
+  - tests/entrypoints/anthropic
+  - tests/entrypoints/generate
+  commands:
+  - pytest -v -s tool_use
+  - pytest -v -s entrypoints/tool_parsers
+  - pytest -v -s entrypoints/generate
+  - pytest -v -s entrypoints/anthropic
+  mirror:
+    amd:
+      device: mi325_1
+      timeout_in_minutes: 60
+      depends_on:
+      - image-build-amd
+
+- label: Entrypoints Integration (Responses API)
+  key: entrypoints-integration-responses-api
+  timeout_in_minutes: 50
+  working_dir: "/vllm-workspace/tests"
+  source_file_dependencies:
+  - vllm/
+  - tests/entrypoints/openai/responses
+  commands:
+  - pytest -v -s entrypoints/openai/responses
+
 - label: Entrypoints Integration (Speech to Text)
   device: h200_35gb
   key: entrypoints-integration-speech_to_text
@@ -126,6 +131,18 @@ steps:
   - export VLLM_WORKER_MULTIPROC_METHOD=spawn
   - pytest -v -s entrypoints/speech_to_text
 
+- label: Entrypoints Integration (Multimodal)
+  device: h200_35gb
+  key: entrypoints-integration-multimodal
+  timeout_in_minutes: 50
+  working_dir: "/vllm-workspace/tests"
+  source_file_dependencies:
+  - vllm/
+  - tests/entrypoints/multimodal
+  commands:
+  - export VLLM_WORKER_MULTIPROC_METHOD=spawn
+  - pytest -v -s entrypoints/multimodal
+
 - label: Entrypoints Integration (Pooling)
   key: entrypoints-integration-pooling
   timeout_in_minutes: 50
@@ -137,16 +154,6 @@ steps:
   - export VLLM_WORKER_MULTIPROC_METHOD=spawn
   - pytest -v -s entrypoints/pooling
 
-- label: Entrypoints Integration (Responses API)
-  key: entrypoints-integration-responses-api
-  timeout_in_minutes: 50
-  working_dir: "/vllm-workspace/tests"
-  source_file_dependencies:
-  - vllm/
-  - tests/entrypoints/openai/responses
-  commands:
-  - pytest -v -s entrypoints/openai/responses
-
 - label: OpenAI API Correctness
   key: openai-api-correctness
   timeout_in_minutes: 30
@@ -156,3 +163,20 @@ steps:
   - vllm/entrypoints/openai/
   commands: # LMEval
   - pytest -s entrypoints/openai/correctness/
+  mirror:
+    amd:
+      device: mi325_1
+      depends_on:
+      - image-build-amd
+      source_file_dependencies:
+      - csrc/
+      - vllm/entrypoints/openai/
+      - vllm/model_executor/layers/
+      - vllm/v1/attention/backends/
+      - vllm/v1/attention/selector.py
+      - vllm/_aiter_ops.py
+      - vllm/platforms/rocm.py
+      - vllm/model_executor/model_loader/
+      commands:
+      - bash ../tools/install_torchcodec_rocm.sh || exit 1
+      - pytest -s entrypoints/openai/correctness/

.buildkite/test_areas/kernels.yaml

@@ -21,8 +21,9 @@ steps:
   - csrc/
   - tests/kernels/core
   - tests/kernels/test_concat_mla_q.py
+  - tests/kernels/test_fused_qk_norm_rope_gate.py
   commands:
-    - pytest -v -s kernels/core --ignore=kernels/core/test_minimax_reduce_rms.py  kernels/test_concat_mla_q.py
+    - pytest -v -s kernels/core --ignore=kernels/core/test_minimax_reduce_rms.py  kernels/test_concat_mla_q.py kernels/test_fused_qk_norm_rope_gate.py
 
 - label: Kernels MiniMax Reduce RMS Test (2 GPUs)
   key: kernels-minimax-reduce-rms-test-2-gpus
@@ -74,6 +75,19 @@ steps:
     - pytest -v -s kernels/attention --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT
   parallelism: 2
 
+- label: Kernels Attention DiffKV Test (H100)
+  key: kernels-attention-diffkv-test-h100
+  timeout_in_minutes: 20
+  device: h100
+  num_devices: 1
+  source_file_dependencies:
+  - vllm/v1/attention/ops/triton_unified_attention_diffkv.py
+  - vllm/v1/attention/backends/triton_attn_diffkv.py
+  - vllm/v1/attention/backends/flash_attn_diffkv.py
+  - tests/kernels/attention/test_triton_unified_attention_diffkv.py
+  commands:
+    - pytest -v -s kernels/attention/test_triton_unified_attention_diffkv.py
+
 - label: Kernels Quantization Test %N
   key: kernels-quantization-test
   timeout_in_minutes: 90
@@ -223,7 +237,7 @@ steps:
   - vllm/utils/import_utils.py
   - tests/kernels/helion/
   commands:
-    - pip install helion==1.0.0
+    - pip install helion==1.1.0
     - pytest -v -s kernels/helion/
 
  
@@ -299,3 +313,4 @@ steps:
   - vllm/config
   commands:
     - pytest -v -s kernels/moe/test_moe_layer.py
+    - pytest -v -s kernels/moe/test_deepep_v2_moe.py

.buildkite/test_areas/lm_eval.yaml

@@ -12,6 +12,21 @@ steps:
   autorun_on_main: true
   commands:
   - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-small.txt
+  mirror:
+    amd:
+      device: mi325_1
+      timeout_in_minutes: 55
+      depends_on:
+      - image-build-amd
+      source_file_dependencies:
+      - csrc/
+      - vllm/model_executor/layers/quantization
+      - vllm/model_executor/models/
+      - vllm/model_executor/model_loader/
+      - vllm/v1/attention/backends/
+      - vllm/v1/attention/selector.py
+      - vllm/_aiter_ops.py
+      - vllm/platforms/rocm.py
 
 # - label: LM Eval Large Models (4 GPUs)(A100)
 #   device: a100

.buildkite/test_areas/lora.yaml

@@ -12,6 +12,17 @@ steps:
   commands:
     - pytest -v -s lora --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --ignore=lora/test_chatglm3_tp.py --ignore=lora/test_llama_tp.py --ignore=lora/test_qwen3_with_multi_loras.py --ignore=lora/test_olmoe_tp.py --ignore=lora/test_deepseekv2_tp.py --ignore=lora/test_gptoss_tp.py --ignore=lora/test_qwen3moe_tp.py --ignore=lora/test_qwen35_densemodel_lora.py 
   parallelism: 4
+  mirror:
+    amd:
+      device: mi325_1
+      working_dir: "/vllm-workspace/tests"
+      timeout_in_minutes: 60
+      source_file_dependencies:
+      - vllm/lora
+      - tests/lora
+      - vllm/platforms/rocm.py
+      depends_on:
+      - image-build-amd
 
 
 - label: LoRA TP (Distributed)

.buildkite/test_areas/misc.yaml

@@ -21,6 +21,12 @@ steps:
     - export VLLM_WORKER_MULTIPROC_METHOD=spawn
     # TODO: create another `optional` test group for slow tests
     - pytest -v -s -m 'not slow_test' v1/spec_decode
+  mirror:
+    amd:
+      device: mi300_1
+      timeout_in_minutes: 65
+      depends_on:
+      - image-build-amd
 
 - label: V1 Sample + Logits
   key: v1-sample-logits
@@ -138,11 +144,26 @@ steps:
     - vllm/v1/spec_decode/extract_hidden_states.py
     - vllm/model_executor/models/extract_hidden_states.py
     - vllm/transformers_utils/configs/extract_hidden_states.py
+    - vllm/distributed/kv_transfer/kv_connector/v1/example_hidden_states_connector.py
     - tests/v1/kv_connector/extract_hidden_states_integration
   commands:
     - export VLLM_WORKER_MULTIPROC_METHOD=spawn
     - pytest -v -s v1/kv_connector/extract_hidden_states_integration
 
+- label: Extract Hidden States Integration (2 GPUs)
+  key: extract-hidden-states-integration-2-gpus
+  timeout_in_minutes: 20
+  num_devices: 2
+  source_file_dependencies:
+    - vllm/v1/spec_decode/extract_hidden_states.py
+    - vllm/model_executor/models/extract_hidden_states.py
+    - vllm/transformers_utils/configs/extract_hidden_states.py
+    - vllm/distributed/kv_transfer/kv_connector/v1/example_hidden_states_connector.py
+    - tests/v1/kv_connector/extract_hidden_states_integration
+  commands:
+    - export VLLM_WORKER_MULTIPROC_METHOD=spawn
+    - pytest -v -s -m 'distributed' v1/kv_connector/extract_hidden_states_integration
+
 - label: Regression
   key: regression
   timeout_in_minutes: 20
@@ -281,6 +302,7 @@ steps:
   - vllm/model_executor/layers/quantization/quark/
   - vllm/multimodal/
   - vllm/outputs.py
+  - vllm/parser/
   - vllm/platforms/
   - vllm/pooling_params.py
   - vllm/ray/
@@ -292,6 +314,7 @@ steps:
   - vllm/transformers_utils/
   - vllm/utils/
   - vllm/v1/
+  - tests/test_envs.py
   - tests/test_inputs.py
   - tests/test_outputs.py
   - tests/test_pooling_params.py
@@ -299,24 +322,25 @@ steps:
   - tests/multimodal
   - tests/renderers
   - tests/standalone_tests/lazy_imports.py
-  - tests/tokenizers_
   - tests/reasoning
   - tests/tool_parsers
+  - tests/tokenizers_
   - tests/parser
   - tests/transformers_utils
   - tests/config
   device: cpu-small
   commands:
   - python3 standalone_tests/lazy_imports.py
+  - pytest -v -s test_envs.py
   - pytest -v -s test_inputs.py
   - pytest -v -s test_outputs.py
   - pytest -v -s test_pooling_params.py
   - pytest -v -s test_ray_env.py
   - pytest -v -s -m 'cpu_test' multimodal
   - pytest -v -s renderers
-  - pytest -v -s tokenizers_
   - pytest -v -s reasoning --ignore=reasoning/test_seedoss_reasoning_parser.py --ignore=reasoning/test_glm4_moe_reasoning_parser.py
   - pytest -v -s tool_parsers
+  - pytest -v -s tokenizers_
   - pytest -v -s parser
   - pytest -v -s transformers_utils
   - pytest -v -s config

.buildkite/test_areas/model_runner_v2.yaml

@@ -94,11 +94,13 @@ steps:
     - vllm/v1/worker/gpu_worker.py
     - tests/distributed/test_pipeline_parallel.py
     - tests/distributed/test_pp_cudagraph.py
+    - tests/v1/distributed/test_pp_dp_v2.py
   commands:
     - set -x
     - export VLLM_USE_V2_MODEL_RUNNER=1
     - pytest -v -s distributed/test_pipeline_parallel.py -k "not ray and not Jamba"
     - pytest -v -s distributed/test_pp_cudagraph.py -k "not ray"
+    - pytest -v -s v1/distributed/test_pp_dp_v2.py
 
 - label: Model Runner V2 Spec Decode
   device: h200_35gb

.buildkite/test_areas/models_language.yaml

@@ -51,6 +51,7 @@ steps:
     torch_nightly: {}
     amd:
       device: mi325_1
+      timeout_in_minutes: 90
       depends_on:
       - image-build-amd
       commands:

.buildkite/test_areas/models_multimodal.yaml

@@ -30,7 +30,6 @@ steps:
     - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
     - pytest -v -s models/multimodal/generation/test_common.py -m core_model -k "qwen3 or gemma"
     - pytest -v -s models/multimodal/generation/test_qwen2_5_vl.py -m core_model
-    - pytest -v -s models/multimodal/generation/test_vit_cudagraph.py -m core_model
   mirror:
     amd:
       device: mi325_1
@@ -63,9 +62,16 @@ steps:
   - tests/models/multimodal
   commands:
     - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
-    - pytest -v -s models/multimodal -m core_model --ignore models/multimodal/generation/test_common.py --ignore models/multimodal/generation/test_ultravox.py --ignore models/multimodal/generation/test_qwen2_5_vl.py --ignore models/multimodal/generation/test_qwen2_vl.py --ignore models/multimodal/generation/test_whisper.py  --ignore models/multimodal/generation/test_memory_leak.py --ignore models/multimodal/processing
+    - pytest -v -s models/multimodal -m core_model --ignore models/multimodal/generation/test_common.py --ignore models/multimodal/generation/test_ultravox.py --ignore models/multimodal/generation/test_qwen2_5_vl.py --ignore models/multimodal/generation/test_qwen2_vl.py --ignore models/multimodal/generation/test_whisper.py  --ignore models/multimodal/generation/test_memory_leak.py --ignore models/multimodal/generation/test_vit_cudagraph.py --ignore models/multimodal/processing
+    - pytest -v -s models/multimodal/generation/test_vit_cudagraph.py -m core_model
     - pytest models/multimodal/generation/test_memory_leak.py -m core_model
     - cd .. && VLLM_WORKER_MULTIPROC_METHOD=spawn pytest -v -s tests/models/multimodal/generation/test_whisper.py -m core_model  # Otherwise, mp_method="spawn" doesn't work
+  mirror:
+    amd:
+      soft_fail: true
+      device: mi325_1
+      depends_on:
+      - image-build-amd
 
 - label: Multi-Modal Processor (CPU)
   key: multi-modal-processor-cpu
@@ -153,3 +159,12 @@ steps:
   - tests/models/multimodal/pooling
   commands:
     - pytest -v -s models/multimodal/pooling -m 'not core_model'
+  mirror:
+    amd:
+      device: mi325_1
+      timeout_in_minutes: 60
+      depends_on:
+      - image-build-amd
+      source_file_dependencies:
+      - vllm/
+      - tests/models/multimodal/pooling

.buildkite/test_areas/plugins.yaml

@@ -27,6 +27,10 @@ steps:
   - pip install -e ./plugins/bge_m3_sparse_plugin
   - pytest -v -s plugins_tests/test_bge_m3_sparse_io_processor_plugins.py
   - pip uninstall bge_m3_sparse_plugin -y
+  # test colbert_query io_processor plugin
+  - pip install -e ./plugins/colbert_query_plugin
+  - pytest -v -s plugins_tests/test_colbert_query_io_processor_plugins.py
+  - pip uninstall colbert_query_plugin -y
   # end io_processor plugins test
   # begin stat_logger plugins test
   - pip install -e ./plugins/vllm_add_dummy_stat_logger
@@ -37,6 +41,20 @@ steps:
   - pytest -v -s plugins_tests/test_scheduler_plugins.py
   - pip install -e ./plugins/vllm_add_dummy_model
   - pytest -v -s distributed/test_distributed_oot.py
-  - pytest -v -s entrypoints/openai/chat_completion/test_oot_registration.py # it needs a clean process
-  - pytest -v -s models/test_oot_registration.py # it needs a clean process
-  - pytest -v -s plugins/lora_resolvers # unit tests for in-tree lora resolver plugins
+  - pytest -v -s plugins_tests/test_oot_registration_online.py # it needs a clean process
+  - pytest -v -s plugins_tests/test_oot_registration_offline.py # it needs a clean process
+  - pytest -v -s plugins_tests/lora_resolvers # unit tests for in-tree lora resolver plugins
+
+
+- label: GGUF Plugin
+  key: gguf-plugin
+  device: h200_18gb
+  timeout_in_minutes: 30
+  soft_fail: true
+  optional: true
+  source_file_dependencies:
+  - vllm/model_executor/layers/quantization
+  - tests/plugins_tests/test_gguf_plugin.py
+  commands:
+    - pip install "vllm-gguf-plugin >= 0.0.2"
+    - pytest -v -s plugins_tests/gguf

.buildkite/test_areas/quantization.yaml

@@ -21,6 +21,18 @@ steps:
   - uv pip install --system conch-triton-kernels
   - VLLM_TEST_FORCE_LOAD_FORMAT=auto pytest -v -s quantization/ --ignore quantization/test_blackwell_moe.py
 
+- label: Quantized Fusions
+  key: quantized-fusions
+  timeout_in_minutes: 30
+  source_file_dependencies:
+  - tests/fusion
+  - vllm/model_executor/layers/fusion
+  - vllm/model_executor/kernels/linear
+  - vllm/model_executor/layers/quantization/compressed_tensors
+  - vllm/model_executor/layers/quantization/modelopt.py
+  commands:
+    - pytest -v -s fusion/
+
 - label: Quantized MoE Test (B200)
   key: quantized-moe-test-b200
   timeout_in_minutes: 60

.buildkite/test_areas/rust_frontend.yaml

@@ -99,9 +99,13 @@ steps:
   - vllm/v1/engine/
   - vllm/v1/worker/
   - tests/utils.py
+  - tests/v1/distributed/test_external_lb_dp.py
+  - tests/v1/distributed/test_hybrid_lb_dp.py
   - tests/v1/distributed/test_internal_lb_dp.py
   commands:
   - export VLLM_USE_RUST_FRONTEND=1
   - export VLLM_WORKER_MULTIPROC_METHOD=spawn
   - export NCCL_CUMEM_HOST_ENABLE=0
   - TP_SIZE=1 DP_SIZE=4 pytest -v -s v1/distributed/test_internal_lb_dp.py -k "not 4 and not server_info"
+  - TP_SIZE=1 DP_SIZE=2 pytest -v -s v1/distributed/test_external_lb_dp.py -k "not 4 and not server_info"
+  - TP_SIZE=1 DP_SIZE=4 pytest -v -s v1/distributed/test_hybrid_lb_dp.py -k "not 4 and not server_info"

.buildkite/test_areas/spec_decode.yaml

@@ -37,6 +37,21 @@ steps:
     - tests/v1/e2e/spec_decode/
   commands:
     - pytest -v -s v1/e2e/spec_decode -k "speculators or mtp_correctness"
+  mirror:
+    amd:
+      device: mi325_1
+      timeout_in_minutes: 65
+      depends_on:
+      - image-build-amd
+      source_file_dependencies:
+      - vllm/v1/spec_decode/
+      - vllm/v1/worker/gpu/spec_decode/
+      - vllm/model_executor/model_loader/
+      - vllm/v1/sample/
+      - vllm/model_executor/layers/
+      - vllm/transformers_utils/configs/speculators/
+      - tests/v1/e2e/spec_decode/
+      - vllm/platforms/rocm.py
 
 - label: Spec Decode Speculators + MTP Nightly B200
   key: spec-decode-speculators-mtp-nightly-b200
@@ -61,6 +76,20 @@ steps:
     - tests/v1/e2e/spec_decode/
   commands:
     - pytest -v -s v1/e2e/spec_decode -k "ngram or suffix"
+  mirror:
+    amd:
+      device: mi325_1
+      timeout_in_minutes: 65
+      depends_on:
+      - image-build-amd
+      source_file_dependencies:
+      - vllm/v1/spec_decode/
+      - vllm/v1/worker/gpu/spec_decode/
+      - vllm/model_executor/model_loader/
+      - vllm/v1/sample/
+      - vllm/model_executor/layers/
+      - tests/v1/e2e/spec_decode/
+      - vllm/platforms/rocm.py
 
 - label: Spec Decode Draft Model
   key: spec-decode-draft-model
@@ -72,6 +101,20 @@ steps:
     - tests/v1/e2e/spec_decode/
   commands:
     - pytest -v -s v1/e2e/spec_decode -k "draft_model or no_sync or batch_inference"
+  mirror:
+    amd:
+      device: mi325_1
+      timeout_in_minutes: 50
+      depends_on:
+      - image-build-amd
+      source_file_dependencies:
+      - vllm/v1/spec_decode/
+      - vllm/v1/worker/gpu/spec_decode/
+      - vllm/model_executor/model_loader/
+      - vllm/v1/sample/
+      - vllm/model_executor/layers/
+      - tests/v1/e2e/spec_decode/
+      - vllm/platforms/rocm.py
 
 - label: Spec Decode Draft Model Nightly B200
   key: spec-decode-draft-model-nightly-b200

.github/CODEOWNERS

@@ -23,9 +23,14 @@
 
 # Any change to the VllmConfig changes can have a large user-facing impact,
 # so spam a lot of people
-/vllm/config @WoosukKwon @youkaichao @robertgshaw2-redhat @mgoin @tlrmchlsmth @houseroad @hmellor @yewentao256 @ProExpertProg
+/vllm/config @WoosukKwon @youkaichao @robertgshaw2-redhat @mgoin @tlrmchlsmth @houseroad @yewentao256 @ProExpertProg
 /vllm/config/cache.py @heheda12345
 
+# Config utils
+/vllm/config/utils.py @hmellor
+/vllm/engine/arg_utils.py @hmellor
+/vllm/utils/argparse_utils.py
+
 # Entrypoints
 /vllm/entrypoints/anthropic @mgoin @DarkLight1337
 /vllm/entrypoints/cli @hmellor @mgoin @DarkLight1337 @russellb
@@ -34,10 +39,11 @@
 /vllm/entrypoints/speech_to_text/realtime @njhill
 /vllm/entrypoints/speech_to_text @NickLucche
 /vllm/entrypoints/pooling @noooop
-/vllm/entrypoints/sagemaker @DarkLight1337
+/vllm/entrypoints/serve/sagemaker @DarkLight1337
 /vllm/entrypoints/serve @njhill
 /vllm/entrypoints/*.py @njhill
 /vllm/entrypoints/chat_utils.py @DarkLight1337
+/vllm/entrypoints/offline_utils.py @DarkLight1337
 /vllm/entrypoints/llm.py @DarkLight1337
 
 # Rust Frontend
@@ -74,7 +80,7 @@
 /vllm/v1/worker/kv_connector_model_runner_mixin.py @orozery @NickLucche
 
 # Model runner V2
-/vllm/v1/worker/gpu @WoosukKwon @njhill
+/vllm/v1/worker/gpu @WoosukKwon @njhill @yewentao256
 /vllm/v1/worker/gpu/kv_connector.py @orozery
 
 # CI & building
@@ -114,16 +120,6 @@
 /vllm/model_executor/models/transformers @hmellor
 /tests/models/test_transformers.py @hmellor
 
-# Observability
-/vllm/config/observability.py @markmc
-/vllm/v1/metrics @markmc
-/tests/v1/metrics @markmc
-/vllm/tracing.py @markmc
-/tests/v1/tracing/test_tracing.py @markmc
-/vllm/config/kv_events.py @markmc
-/vllm/distributed/kv_events.py @markmc
-/tests/distributed/test_events.py @markmc
-
 # Docs
 /docs/mkdocs @hmellor
 /docs/**/*.yml @hmellor

.github/actionlint.yaml

@@ -0,0 +1,5 @@
+# Custom self-hosted runner labels (e.g. the autoscaling vllm-runners pool) so
+# actionlint doesn't flag them as unknown in `runs-on`.
+self-hosted-runner:
+  labels:
+    - vllm-runners

.github/dependabot.yml

@@ -21,7 +21,6 @@ updates:
       - dependency-name: "torchvision"
       - dependency-name: "xformers"
       - dependency-name: "lm-format-enforcer"
-      - dependency-name: "gguf"
       - dependency-name: "compressed-tensors"
       - dependency-name: "ray[cgraph]" # Ray Compiled Graph
       - dependency-name: "lm-eval"

.github/mergify.yml

@@ -21,6 +21,9 @@ pull_request_rules:
     - check-failure=pre-commit
     - -closed
     - -draft
+    - or:
+      - label=ready
+      - label=verified
   actions:
     comment:
       message: |
@@ -36,18 +39,6 @@ pull_request_rules:
 
         For future commits, `pre-commit` will run automatically on changed files before each commit.
 
-        > [!TIP]
-        > <details>
-        > <summary>Is <code>mypy</code> failing?</summary>
-        > <br/>
-        > <code>mypy</code> is run differently in CI. If the failure is related to this check, please use the following command to run it locally:
-        >
-        > ```bash
-        > # For mypy (substitute "3.10" with the failing version if needed)
-        > pre-commit run --hook-stage manual mypy-3.10
-        > ```
-        > </details>
-
 - name: comment-dco-failure
   description: Comment on PR when DCO check fails
   conditions:
@@ -153,12 +144,12 @@ pull_request_rules:
     - label != stale
     - or:
       - files~=^examples/.*mistral.*\.py
-      - files~=^tests/.*mistral.*\.py
-      - files~=^vllm/model_executor/models/.*mistral.*\.py
+      - files~=^tests/.*(?:mistral|voxtral|mixtral|pixtral).*\.py
+      - files~=^vllm/model_executor/models/.*(?:mistral|voxtral|mixtral|pixtral).*\.py
       - files~=^vllm/reasoning/.*mistral.*\.py
       - files~=^vllm/tool_parsers/.*mistral.*\.py
-      - files~=^vllm/transformers_utils/.*mistral.*\.py
-      - title~=(?i)Mistral
+      - files~=^vllm/transformers_utils/.*(?:mistral|voxtral|pixtral).*\.py
+      - title~=(?i)(?:mistral|ministral|voxtral|mixtral|pixtral)
   actions:
     label:
       add:
@@ -397,9 +388,13 @@ pull_request_rules:
     - or:
       - files~=^tests/tool_use/
       - files~=^tests/tool_parsers/
+      - files~=^tests/parser/
+      - files~=^tests/reasoning/
       - files~=^tests/entrypoints/openai/.*tool.*
       - files~=^tests/entrypoints/anthropic/.*tool.*
       - files~=^vllm/tool_parsers/
+      - files~=^vllm/parser/
+      - files~=^vllm/reasoning/
       - files=docs/features/tool_calling.md
       - files~=^examples/tool_calling/
   actions:

.github/workflows/pre-commit.yml

@@ -46,12 +46,16 @@ jobs:
   pre-commit:
     needs: pre-run-check
     if: always() && (needs.pre-run-check.result == 'success' || needs.pre-run-check.result == 'skipped')
-    runs-on: ubuntu-latest
+    runs-on: [self-hosted, linux, x64, vllm-runners]
     steps:
     - uses: actions/checkout@8e8c483db84b4bee98b60c0593521ed34d9990e8 # v6.0.1
-    - uses: actions/setup-python@a309ff8b426b58ec0e2a45f0f869d46889d02405 # v6.2.0
+    - uses: actions/setup-python@83679a892e2d95755f2dac6acb0bfd1e9ac5d548 # v6.1.0
       with:
         python-version: "3.12"
+    # Provide shellcheck on PATH so tools/pre_commit/shellcheck.sh skips its
+    # wget + tar -xJ self-download, which the self-hosted runner image lacks
+    # (no wget/xz). Pinned to shellcheck 0.10.0 to match the script's "stable".
+    - run: python -m pip install shellcheck-py==0.10.0.1
     - run: echo "::add-matcher::.github/workflows/matchers/actionlint.json"
     - run: echo "::add-matcher::.github/workflows/matchers/markdownlint.json"
     - run: echo "::add-matcher::.github/workflows/matchers/mypy.json"

.github/workflows/scripts/build.sh

@@ -9,7 +9,7 @@ PATH=${cuda_home}/bin:$PATH
 LD_LIBRARY_PATH=${cuda_home}/lib64:$LD_LIBRARY_PATH
 
 # Install requirements
-if [ "$(echo $2 | cut -d. -f1)" = "12" ]; then
+if [ "$(echo "$2" | cut -d. -f1)" = "12" ]; then
     sed -i 's/^nvidia-cutlass-dsl\[cu13\]>=/nvidia-cutlass-dsl>=/' requirements/cuda.txt
 fi
 $python_executable -m pip install -r requirements/build/cuda.txt -r requirements/cuda.txt
@@ -17,7 +17,10 @@ $python_executable -m pip install -r requirements/build/cuda.txt -r requirements
 # Limit the number of parallel jobs to avoid OOM
 export MAX_JOBS=1
 # Make sure release wheels are built for the following architectures
-export TORCH_CUDA_ARCH_LIST="7.5 8.0 8.6 8.9 9.0 10.0 12.0+PTX"
+# Do not add +PTX here: vLLM filters torch's top-level PTX flag when it
+# converts global gencode flags into per-kernel arch lists. If a specific
+# kernel needs PTX, add +PTX to that kernel's CMake arch list instead.
+export TORCH_CUDA_ARCH_LIST="7.5 8.0 8.6 8.9 9.0 10.0 12.0"
 
 bash tools/check_repo.sh
 

.github/workflows/stale.yml

@@ -15,7 +15,7 @@ jobs:
       actions: write
     runs-on: ubuntu-latest
     steps:
-      - uses: actions/stale@997185467fa4f803885201cee163a9f38240193d # v10.1.1
+      - uses: actions/stale@eb5cf3af3ac0a1aa4c9c45633dd1ae542a27a899 # v10.3.0
         with:
           # Increasing this value ensures that changes to this workflow
           # propagate to all issues and PRs in days rather than months

.gitignore

@@ -15,6 +15,9 @@ vllm/third_party/flashmla/flash_mla_interface.py
 # DeepGEMM vendored package built from source
 vllm/third_party/deep_gemm/
 
+# fmha_sm100 vendored package built from source
+vllm/third_party/fmha_sm100/
+
 # triton jit
 .triton
 
@@ -233,7 +236,7 @@ actionlint
 shellcheck*/
 
 # Ignore moe/marlin_moe gen code
-csrc/moe/marlin_moe_wna16/kernel_*
+csrc/libtorch_stable/moe/marlin_moe_wna16/kernel_*
 
 # Ignore ep_kernels_workspace folder
 ep_kernels_workspace/

.pre-commit-config.yaml

@@ -21,7 +21,7 @@ repos:
   rev: v21.1.2
   hooks:
   - id: clang-format
-    exclude: 'csrc/(moe/topk_softmax_kernels.cu|libtorch_stable/quantization/gguf/(ggml-common.h|dequantize.cuh|vecdotq.cuh|mmq.cuh|mmvq.cuh))|vllm/third_party/.*'
+    exclude: 'csrc/libtorch_stable/moe/topk_softmax_kernels.cu|vllm/third_party/.*'
     types_or: [c++, cuda]
     args: [--style=file, --verbose]
 - repo: https://github.com/DavidAnson/markdownlint-cli2
@@ -148,33 +148,27 @@ repos:
     language: python
     entry: python tools/pre_commit/generate_nightly_torch_test.py
     files: ^requirements/test/cuda\.(in|txt)$
-  - id: mypy-local
-    name: Run mypy locally for lowest supported Python version
-    entry: python tools/pre_commit/mypy.py 0 "3.10"
-    stages: [pre-commit] # Don't run in CI
+  - id: mypy-3.10 # TODO: Use https://github.com/pre-commit/mirrors-mypy when mypy setup is less awkward
+    name: Run mypy for Python 3.10
+    entry: python tools/pre_commit/mypy.py "3.10"
     <<: &mypy_common
       language: python
       types_or: [python, pyi]
       require_serial: true
-      additional_dependencies: ["mypy[faster-cache]==1.19.1", regex, types-cachetools, types-setuptools, types-PyYAML, types-requests, types-torch, pydantic]
-  - id: mypy-3.10 # TODO: Use https://github.com/pre-commit/mirrors-mypy when mypy setup is less awkward
-    name: Run mypy for Python 3.10
-    entry: python tools/pre_commit/mypy.py 1 "3.10"
-    <<: *mypy_common
-    stages: [manual] # Only run in CI
+      additional_dependencies: ["mypy==1.20.2", regex, types-cachetools, types-setuptools, types-PyYAML, types-requests, types-torch, pydantic]
   - id: mypy-3.11 # TODO: Use https://github.com/pre-commit/mirrors-mypy when mypy setup is less awkward
     name: Run mypy for Python 3.11
-    entry: python tools/pre_commit/mypy.py 1 "3.11"
+    entry: python tools/pre_commit/mypy.py "3.11"
     <<: *mypy_common
     stages: [manual] # Only run in CI
   - id: mypy-3.12 # TODO: Use https://github.com/pre-commit/mirrors-mypy when mypy setup is less awkward
     name: Run mypy for Python 3.12
-    entry: python tools/pre_commit/mypy.py 1 "3.12"
+    entry: python tools/pre_commit/mypy.py "3.12"
     <<: *mypy_common
     stages: [manual] # Only run in CI
   - id: mypy-3.13 # TODO: Use https://github.com/pre-commit/mirrors-mypy when mypy setup is less awkward
     name: Run mypy for Python 3.13
-    entry: python tools/pre_commit/mypy.py 1 "3.13"
+    entry: python tools/pre_commit/mypy.py "3.13"
     <<: *mypy_common
     stages: [manual] # Only run in CI
   - id: shellcheck

AGENTS.md

@@ -98,11 +98,33 @@ pre-commit run --all-files
 pre-commit run ruff-check --all-files
 
 # Run mypy as it is in CI:
-pre-commit run mypy-3.10 --all-files --hook-stage manual
+pre-commit run mypy-3.12 --all-files --hook-stage manual
 ```
 
 The line length limit for Python code is 88 characters. If you are not sure, use pre-commit to check.
 
+Use [Google-style docstrings](https://google.github.io/styleguide/pyguide.html#38-comments-and-docstrings) (`Args:`/`Returns:`/`Raises:` sections), not reStructuredText/Sphinx fields (`:param:`, `:return:`, `:rtype:`).
+
+### Coding style guidelines
+
+Follow these rules for all code changes in this repository:
+
+- Try to match existing code style.
+- Code should be self-documenting and self-explanatory.
+- Keep comments and docstrings minimal and concise.
+- Assume the reader is familiar with vLLM.
+
+### Diagnosing CI failures
+
+Buildkite logs are public; no login needed. Details: [docs/contributing/ci/failures.md](docs/contributing/ci/failures.md).
+
+```bash
+# All failed-job logs for a PR's latest build (current branch's PR if omitted):
+.buildkite/scripts/ci-fetch-log.sh --pr <PR>
+# Any Buildkite build or job URL also works:
+.buildkite/scripts/ci-fetch-log.sh "<buildkite_url>"
+```
+
 ### Commit messages
 
 Add attribution using commit trailers such as `Co-authored-by:` (other projects use `Assisted-by:` or `Generated-by:`). For example:

MANIFEST.in

@@ -4,6 +4,7 @@ include requirements/cuda.txt
 include requirements/rocm.txt
 include requirements/cpu.txt
 include CMakeLists.txt
+include tools/build_rust.py
 
 recursive-include cmake *
 recursive-include csrc *

SECURITY.md

@@ -34,6 +34,15 @@ Vulnerabilities that cause denial of service or partial disruption, but do not a
 
 Minor issues such as informational disclosures, logging errors, non-exploitable flaws, or weaknesses that require local or high-privilege access and offer negligible impact. Examples include side channel attacks or hash collisions. These issues often have CVSS scores less than 4.0
 
+## Fix disclosure policy
+
+When a security report is accepted, the fix process depends on the severity:
+
+* **CRITICAL and HIGH severity**: Fixes are developed in a private security fork and coordinated with the prenotification group before public disclosure.
+* **MODERATE and LOW severity**: Fixes are developed and submitted as public pull requests. These issues do not require embargo since they do not enable arbitrary code execution or significant data breach, and public visibility accelerates community review and adoption of the fix.
+
+The vulnerability management team reserves the right to adjust the disclosure approach on a case-by-case basis, taking into account factors such as active exploitation, unusual attack surface, or coordination requirements with downstream vendors.
+
 ## Prenotification policy
 
 For certain security issues of CRITICAL, HIGH, or MODERATE severity level, we may prenotify certain organizations or vendors that ship vLLM. The purpose of this prenotification is to allow for a coordinated release of fixes for severe issues.

benchmarks/attention_benchmarks/README.md

@@ -108,7 +108,6 @@ python benchmark.py \
     --backends flash triton flashinfer \
     --batch-specs "q2k" "8q1s1k" "2q2k_32q1s1k" \
     --num-layers 10 \
-    --repeats 5 \
     --output-csv results.csv
 ```
 
@@ -164,14 +163,17 @@ python benchmark.py \
 # Model configuration
 --num-layers N                      # Number of layers
 --head-dim N                        # Head dimension
+--v-head-dim N                      # Value head dimension (defaults to --head-dim)
 --num-q-heads N                     # Query heads
 --num-kv-heads N                    # KV heads
 --block-size N                      # Block size
+--kv-lora-rank N                    # MLA KV LoRA rank
+--qk-nope-head-dim N                # MLA non-RoPE QK head dim
+--qk-rope-head-dim N                # MLA RoPE QK head dim
 
 # Benchmark settings
 --device DEVICE                     # Device (default: cuda:0)
---repeats N                         # Repetitions
---warmup-iters N                    # Warmup iterations
+--warmup-ms N                       # Warmup window in ms for triton do_bench
 --profile-memory                    # Profile memory usage
 
 # Parameter sweeps
@@ -211,8 +213,6 @@ config = BenchmarkConfig(
     num_kv_heads=1,
     block_size=128,
     device="cuda:0",
-    repeats=5,
-    warmup_iters=3,
 )
 
 # CUTLASS MLA with specific num_kv_splits
@@ -253,14 +253,10 @@ formatter.save_json(results, "output.json")
 
 ## Tips
 
-**1. Warmup matters** - Use `--warmup-iters 10` for stable results
+**1. Save results** - Always use `--output-csv` or `--output-json`
 
-**2. Multiple repeats** - Use `--repeats 20` for low variance
+**2. Test incrementally** - Start with `--num-layers 1`
 
-**3. Save results** - Always use `--output-csv` or `--output-json`
+**3. Extended grammar** - Leverage spec decode, chunked prefill patterns
 
-**4. Test incrementally** - Start with `--num-layers 1 --repeats 1`
-
-**5. Extended grammar** - Leverage spec decode, chunked prefill patterns
-
-**6. Parameter sweeps** - Use `--sweep-param` and `--sweep-values` to find optimal values
+**4. Parameter sweeps** - Use `--sweep-param` and `--sweep-values` to find optimal values

benchmarks/attention_benchmarks/benchmark.py

@@ -26,6 +26,9 @@ Examples:
 """
 
 import argparse
+import os
+import shutil
+import subprocess
 import sys
 from dataclasses import replace
 from pathlib import Path
@@ -50,6 +53,16 @@ from common import (
 from vllm.v1.worker.workspace import init_workspace_manager
 
 
+def _str2bool(v) -> bool:
+    if isinstance(v, bool):
+        return v
+    if v.lower() in ("true", "1", "yes", "t"):
+        return True
+    if v.lower() in ("false", "0", "no", "f"):
+        return False
+    raise argparse.ArgumentTypeError(f"expected a boolean, got {v!r}")
+
+
 def run_standard_attention_benchmark(config: BenchmarkConfig) -> BenchmarkResult:
     """Run standard attention benchmark (Flash/Triton/FlashInfer)."""
     from runner import run_attention_benchmark
@@ -83,13 +96,15 @@ def run_benchmark(config: BenchmarkConfig, **kwargs) -> BenchmarkResult:
         else:
             return run_standard_attention_benchmark(config)
     except Exception as e:
+        error_msg = str(e) or repr(e)
         return BenchmarkResult(
             config=config,
             mean_time=float("inf"),
+            median_time=float("inf"),
             std_time=0,
             min_time=float("inf"),
             max_time=float("inf"),
-            error=str(e),
+            error=error_msg,
         )
 
 
@@ -115,9 +130,12 @@ def run_model_parameter_sweep(
     """
     all_results = []
 
-    console.print(
-        f"[yellow]Model sweep mode: testing {sweep.param_name} = {sweep.values}[/]"
+    sweep_desc = (
+        f"{sweep.param_name} = {sweep.values}"
+        if sweep.param_name
+        else f"{len(sweep.values)} configurations"
     )
+    console.print(f"[yellow]Model sweep mode: testing {sweep_desc}[/]")
 
     total = len(backends) * len(batch_specs) * len(sweep.values)
 
@@ -125,9 +143,9 @@ def run_model_parameter_sweep(
         for backend in backends:
             for spec in batch_specs:
                 for value in sweep.values:
-                    # Create config with modified model parameter
+                    # Create config with modified model parameter(s)
                     config_args = base_config_args.copy()
-                    config_args[sweep.param_name] = value
+                    sweep.apply(config_args, value)
 
                     # Create config with original backend for running
                     clean_config = BenchmarkConfig(
@@ -144,13 +162,21 @@ def run_model_parameter_sweep(
                     all_results.append(result)
 
                     if not result.success:
+                        err_label = (
+                            f"{sweep.param_name}={value}"
+                            if sweep.param_name
+                            else f"{value}"
+                        )
                         console.print(
-                            f"[red]Error {backend} {spec} {sweep.param_name}="
-                            f"{value}: {result.error}[/]"
+                            f"[red]Error {backend} {spec} {err_label}"
+                            f": {result.error}[/]"
                         )
 
                     pbar.update(1)
 
+    if base_config_args.get("ncu_profile"):
+        return all_results
+
     # Display sweep results - create separate table for each parameter value
     console.print("\n[bold green]Model Parameter Sweep Results:[/]")
     formatter = ResultsFormatter(console)
@@ -184,7 +210,10 @@ def run_model_parameter_sweep(
     )
 
     for param_value in sorted_param_values:
-        console.print(f"\n[bold cyan]{sweep.param_name} = {param_value}[/]")
+        label = (
+            f"{sweep.param_name} = {param_value}" if sweep.param_name else param_value
+        )
+        console.print(f"\n[bold cyan]{label}[/]")
         param_results = by_param_value[param_value]
 
         # Create modified results with original backend names
@@ -200,8 +229,9 @@ def run_model_parameter_sweep(
         formatter.print_table(modified_results, backends, compare_to_fastest=True)
 
     # Show optimal backend for each (param_value, batch_spec) combination
+    sweep_name = sweep.param_name or "config"
     console.print(
-        f"\n[bold cyan]Optimal backend for each ({sweep.param_name}, batch_spec):[/]"
+        f"\n[bold cyan]Optimal backend for each ({sweep_name}, batch_spec):[/]"
     )
 
     # Group by (param_value, batch_spec)
@@ -236,7 +266,10 @@ def run_model_parameter_sweep(
     for param_value, spec in sorted_keys:
         # Print header when param value changes
         if param_value != current_param_value:
-            console.print(f"\n  [bold]{sweep.param_name}={param_value}:[/]")
+            header = (
+                f"{sweep.param_name}={param_value}" if sweep.param_name else param_value
+            )
+            console.print(f"\n  [bold]{header}:[/]")
             current_param_value = param_value
 
         results = by_param_and_spec[(param_value, spec)]
@@ -322,6 +355,9 @@ def run_parameter_sweep(
 
                     pbar.update(1)
 
+    if base_config_args.get("ncu_profile"):
+        return all_results
+
     # Display sweep results
     console.print("\n[bold green]Sweep Results:[/]")
     backend_labels = [sweep.get_label(b, v) for b in backends for v in sweep_values]
@@ -459,6 +495,20 @@ def main():
         help="Prefill backends to compare (fa2, fa3, fa4). "
         "Uses the first decode backend for impl construction.",
     )
+    parser.add_argument(
+        "--fp8-output-scale",
+        type=float,
+        help="Static per-tensor scale enabling the MLA prefill FP8-output "
+        "comparison on FA4 (fused write vs standalone post-quant).",
+    )
+    parser.add_argument(
+        "--fuse-quant-op",
+        nargs="+",
+        type=_str2bool,
+        help="FP8-output write path(s) to run: false = bf16 attention + "
+        "standalone static-FP8 quant, true = FA4 writes FP8 directly. "
+        "Default: both.",
+    )
 
     # Batch specifications
     parser.add_argument(
@@ -474,11 +524,35 @@ def main():
     parser.add_argument("--num-q-heads", type=int, default=32, help="Query heads")
     parser.add_argument("--num-kv-heads", type=int, default=8, help="KV heads")
     parser.add_argument("--block-size", type=int, default=16, help="Block size")
+    parser.add_argument(
+        "--v-head-dim",
+        type=int,
+        default=None,
+        help="Value head dimension (defaults to --head-dim if unset)",
+    )
+
+    # MLA-specific model dimensions
+    parser.add_argument(
+        "--kv-lora-rank", type=int, default=None, help="MLA KV LoRA rank"
+    )
+    parser.add_argument(
+        "--qk-nope-head-dim", type=int, default=None, help="MLA non-RoPE QK head dim"
+    )
+    parser.add_argument(
+        "--qk-rope-head-dim", type=int, default=None, help="MLA RoPE QK head dim"
+    )
 
     # Benchmark settings
     parser.add_argument("--device", default="cuda:0", help="Device")
-    parser.add_argument("--repeats", type=int, default=1, help="Repetitions")
-    parser.add_argument("--warmup-iters", type=int, default=3, help="Warmup iterations")
+    parser.add_argument(
+        "--warmup-ms",
+        type=int,
+        default=None,
+        help=(
+            "Warmup window in ms for triton's do_bench (default: triton's own). "
+            "Has no effect with CUDA graphs; pass --no-cuda-graphs to use it."
+        ),
+    )
     parser.add_argument("--profile-memory", action="store_true", help="Profile memory")
     parser.add_argument(
         "--kv-cache-dtype",
@@ -491,10 +565,33 @@ def main():
         action=argparse.BooleanOptionalAction,
         default=True,
         help=(
-            "Launch kernels with CUDA graphs to eliminate CPU overhead"
-            "in measurements (default: True)"
+            "Use triton do_bench_cudagraph (True) or do_bench (False) "
+            "for timing. CUDA graphs eliminate CPU launch overhead "
+            "(default: True)"
         ),
     )
+    parser.add_argument(
+        "--num-splits",
+        type=int,
+        default=None,
+        help="FlashAttention split-K factor (0=auto heuristic, 1=disabled, >1=force N)",
+    )
+    parser.add_argument(
+        "--ncu-profile",
+        action="store_true",
+        default=False,
+        help=(
+            "Enable Nsight Compute profiling mode. Automatically wraps the "
+            "script with ncu, capturing a profile with source correlation. "
+            "Use --ncu-output to set the output file name."
+        ),
+    )
+    parser.add_argument(
+        "--ncu-output",
+        type=str,
+        default="profile",
+        help="Output file name for ncu profile (default: 'profile').",
+    )
 
     # Parameter sweep (use YAML config for advanced sweeps)
     parser.add_argument(
@@ -545,6 +642,12 @@ def main():
         # Prefill backends (e.g., ["fa3", "fa4"])
         args.prefill_backends = yaml_config.get("prefill_backends", None)
 
+        # FP8 output benchmark knobs; CLI wins.
+        if args.fp8_output_scale is None:
+            args.fp8_output_scale = yaml_config.get("fp8_output_scale", None)
+        if args.fuse_quant_op is None:
+            args.fuse_quant_op = yaml_config.get("fuse_quant_op", None)
+
         # Check for special modes
         args.mode = yaml_config.get("mode", None)
 
@@ -576,23 +679,28 @@ def main():
             model = yaml_config["model"]
             args.num_layers = model.get("num_layers", args.num_layers)
             args.head_dim = model.get("head_dim", args.head_dim)
+            args.v_head_dim = model.get("v_head_dim", args.v_head_dim)
             args.num_q_heads = model.get("num_q_heads", args.num_q_heads)
             args.num_kv_heads = model.get("num_kv_heads", args.num_kv_heads)
             args.block_size = model.get("block_size", args.block_size)
+            # MLA-specific dimensions
+            args.kv_lora_rank = model.get("kv_lora_rank", args.kv_lora_rank)
+            args.qk_nope_head_dim = model.get("qk_nope_head_dim", args.qk_nope_head_dim)
+            args.qk_rope_head_dim = model.get("qk_rope_head_dim", args.qk_rope_head_dim)
 
         # Benchmark settings (top-level keys)
         if "device" in yaml_config:
             args.device = yaml_config["device"]
-        if "repeats" in yaml_config:
-            args.repeats = yaml_config["repeats"]
-        if "warmup_iters" in yaml_config:
-            args.warmup_iters = yaml_config["warmup_iters"]
+        if "warmup_ms" in yaml_config:
+            args.warmup_ms = yaml_config["warmup_ms"]
         if "profile_memory" in yaml_config:
             args.profile_memory = yaml_config["profile_memory"]
         if "kv_cache_dtype" in yaml_config:
             args.kv_cache_dtype = yaml_config["kv_cache_dtype"]
         if "cuda_graphs" in yaml_config:
             args.cuda_graphs = yaml_config["cuda_graphs"]
+        if "ncu_profile" in yaml_config:
+            args.ncu_profile = yaml_config["ncu_profile"]
 
         # Parameter sweep configuration
         if "parameter_sweep" in yaml_config:
@@ -612,7 +720,7 @@ def main():
         if "model_parameter_sweep" in yaml_config:
             sweep_config = yaml_config["model_parameter_sweep"]
             args.model_parameter_sweep = ModelParameterSweep(
-                param_name=sweep_config["param_name"],
+                param_name=sweep_config.get("param_name"),
                 values=sweep_config["values"],
                 label_format=sweep_config.get(
                     "label_format", "{backend}_{param_name}_{value}"
@@ -631,6 +739,32 @@ def main():
 
         console.print()
 
+    # Re-exec under ncu if --ncu-profile and not already inside ncu. This runs
+    # after YAML processing so ncu_profile set via config file is honored.
+    if args.ncu_profile and "_NCU_INNER" not in os.environ:
+        ncu = shutil.which("ncu")
+        if ncu is None:
+            print("Error: 'ncu' not found in PATH", file=sys.stderr)
+            sys.exit(1)
+        cmd = [
+            ncu,
+            "--profile-from-start",
+            "off",
+            "--set",
+            "full",
+            "--import-source",
+            "yes",
+            "-o",
+            args.ncu_output,
+            sys.executable,
+            *sys.argv,
+        ]
+        env = os.environ.copy()
+        env["CUTE_DSL_LINEINFO"] = "1"
+        env["_NCU_INNER"] = "1"
+        print(f"Launching: {' '.join(cmd)}")
+        sys.exit(subprocess.call(cmd, env=env))
+
     # Handle CLI-based parameter sweep (if not from YAML)
     if (
         (not hasattr(args, "parameter_sweep") or args.parameter_sweep is None)
@@ -655,6 +789,18 @@ def main():
     console.print(f"Batch specs: {', '.join(args.batch_specs)}")
     console.print(f"KV cache dtype: {args.kv_cache_dtype}")
     console.print(f"CUDA graphs: {args.cuda_graphs}")
+    if args.warmup_ms is not None and args.cuda_graphs:
+        console.print(
+            "[yellow]Warning: --warmup-ms is ignored with CUDA graphs "
+            "(do_bench_cudagraph warms up internally). Pass --no-cuda-graphs "
+            "to use it.[/]"
+        )
+    if args.num_splits == 0 and args.cuda_graphs:
+        console.print(
+            "[yellow]Warning: --num-splits 0 (FA3 heuristic) is not CUDA-graph "
+            "compatible and may fail or fall back. Pass --no-cuda-graphs or use "
+            "--num-splits >=1.[/]"
+        )
     console.print()
 
     init_workspace_manager(args.device)
@@ -662,8 +808,68 @@ def main():
     # Run benchmarks
     all_results = []
 
+    # Under ncu profiling the kernels run only to be captured by the profiler;
+    # timings are placeholder zeros, so the result tables and saved metrics are
+    # skipped. The Nsight Compute report (--ncu-output) holds the real data.
+    if args.ncu_profile:
+        console.print(
+            "[dim]ncu profiling enabled: result tables and saved metrics are "
+            "skipped (timings are placeholder zeros).[/]"
+        )
+
+    # FA4 fused FP8 output vs standalone post-quant, on the same fa4 kernel:
+    # the delta is the post-quant kernel the fused path removes.
+    fp8_output_scale = getattr(args, "fp8_output_scale", None)
+    if fp8_output_scale is not None:
+        decode_backend = backends[0]
+        fuse_variants = args.fuse_quant_op or [False, True]
+        label_of = {False: "post_quant", True: "fused"}
+        console.print(
+            f"[yellow]FP8 output comparison @ scale={fp8_output_scale} "
+            f"(prefill=fa4, decode impl={decode_backend})[/]"
+        )
+        fp8_results = []
+        total = len(fuse_variants) * len(args.batch_specs)
+        with tqdm(total=total, desc="FP8 output benchmarking") as pbar:
+            for spec in args.batch_specs:
+                for fuse in fuse_variants:
+                    config = BenchmarkConfig(
+                        backend=decode_backend,
+                        batch_spec=spec,
+                        num_layers=args.num_layers,
+                        head_dim=args.head_dim,
+                        num_q_heads=args.num_q_heads,
+                        num_kv_heads=args.num_kv_heads,
+                        block_size=args.block_size,
+                        device=args.device,
+                        repeats=args.repeats,
+                        warmup_iters=args.warmup_iters,
+                        profile_memory=args.profile_memory,
+                        kv_cache_dtype=args.kv_cache_dtype,
+                        use_cuda_graphs=args.cuda_graphs,
+                        prefill_backend="fa4",
+                    )
+                    result = run_benchmark(
+                        config, output_scale=fp8_output_scale, fuse_quant_op=fuse
+                    )
+                    label = label_of[fuse]
+                    labeled_config = replace(result.config, backend=label)
+                    result = replace(result, config=labeled_config)
+                    fp8_results.append(result)
+
+                    if not result.success:
+                        console.print(f"[red]Error {label} {spec}: {result.error}[/]")
+
+                    pbar.update(1)
+
+        console.print("\n[bold green]FP8 Output Results:[/]")
+        formatter = ResultsFormatter(console)
+        labels = [label_of[f] for f in fuse_variants]
+        formatter.print_table(fp8_results, labels, compare_to_fastest=True)
+        all_results = fp8_results
+
     # Handle special mode: decode_vs_prefill comparison
-    if hasattr(args, "mode") and args.mode == "decode_vs_prefill":
+    elif hasattr(args, "mode") and args.mode == "decode_vs_prefill":
         console.print("[yellow]Mode: Decode vs Prefill pipeline comparison[/]")
         console.print(
             "[dim]For each query length, testing both decode and prefill pipelines[/]"
@@ -708,11 +914,11 @@ def main():
                         num_kv_heads=args.num_kv_heads,
                         block_size=args.block_size,
                         device=args.device,
-                        repeats=args.repeats,
-                        warmup_iters=args.warmup_iters,
                         profile_memory=args.profile_memory,
                         kv_cache_dtype=args.kv_cache_dtype,
                         use_cuda_graphs=args.cuda_graphs,
+                        ncu_profile=args.ncu_profile,
+                        warmup_ms=args.warmup_ms,
                     )
 
                     # Add decode pipeline config
@@ -749,6 +955,7 @@ def main():
                         result = BenchmarkResult(
                             config=config,
                             mean_time=timing["mean"],
+                            median_time=timing.get("median", timing["mean"]),
                             std_time=timing["std"],
                             min_time=timing["min"],
                             max_time=timing["max"],
@@ -770,6 +977,7 @@ def main():
                         result = BenchmarkResult(
                             config=config,
                             mean_time=float("inf"),
+                            median_time=float("inf"),
                             std_time=0,
                             min_time=float("inf"),
                             max_time=float("inf"),
@@ -779,6 +987,9 @@ def main():
 
                 pbar.update(1)
 
+        if args.ncu_profile:
+            return
+
         # Display decode vs prefill results
         console.print("\n[bold green]Decode vs Prefill Results:[/]")
 
@@ -858,15 +1069,20 @@ def main():
         base_config_args = {
             "num_layers": args.num_layers,
             "head_dim": args.head_dim,
+            "v_head_dim": args.v_head_dim,
             "num_q_heads": args.num_q_heads,
             "num_kv_heads": args.num_kv_heads,
             "block_size": args.block_size,
             "device": args.device,
-            "repeats": args.repeats,
-            "warmup_iters": args.warmup_iters,
             "profile_memory": args.profile_memory,
             "kv_cache_dtype": args.kv_cache_dtype,
             "use_cuda_graphs": args.cuda_graphs,
+            "ncu_profile": args.ncu_profile,
+            "warmup_ms": args.warmup_ms,
+            "num_splits": args.num_splits,
+            "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,
         }
         all_results = run_model_parameter_sweep(
             backends,
@@ -882,15 +1098,17 @@ def main():
         base_config_args = {
             "num_layers": args.num_layers,
             "head_dim": args.head_dim,
+            "v_head_dim": args.v_head_dim,
             "num_q_heads": args.num_q_heads,
             "num_kv_heads": args.num_kv_heads,
             "block_size": args.block_size,
             "device": args.device,
-            "repeats": args.repeats,
-            "warmup_iters": args.warmup_iters,
             "profile_memory": args.profile_memory,
             "kv_cache_dtype": args.kv_cache_dtype,
             "use_cuda_graphs": args.cuda_graphs,
+            "ncu_profile": args.ncu_profile,
+            "warmup_ms": args.warmup_ms,
+            "num_splits": args.num_splits,
         }
         all_results = run_parameter_sweep(
             backends, args.batch_specs, base_config_args, args.parameter_sweep, console
@@ -914,15 +1132,17 @@ def main():
                             batch_spec=spec,
                             num_layers=args.num_layers,
                             head_dim=args.head_dim,
+                            v_head_dim=getattr(args, "v_head_dim", None),
                             num_q_heads=args.num_q_heads,
                             num_kv_heads=args.num_kv_heads,
                             block_size=args.block_size,
                             device=args.device,
-                            repeats=args.repeats,
-                            warmup_iters=args.warmup_iters,
                             profile_memory=args.profile_memory,
                             kv_cache_dtype=args.kv_cache_dtype,
                             use_cuda_graphs=args.cuda_graphs,
+                            ncu_profile=args.ncu_profile,
+                            warmup_ms=args.warmup_ms,
+                            num_splits=args.num_splits,
                         )
 
                         result = run_benchmark(config)
@@ -935,9 +1155,10 @@ def main():
 
                         pbar.update(1)
 
-            console.print("\n[bold green]Results:[/]")
-            formatter = ResultsFormatter(console)
-            formatter.print_table(decode_results, backends)
+            if not args.ncu_profile:
+                console.print("\n[bold green]Results:[/]")
+                formatter = ResultsFormatter(console)
+                formatter.print_table(decode_results, backends)
 
         # Run prefill backend comparison
         if prefill_backends:
@@ -962,9 +1183,8 @@ def main():
                             num_kv_heads=args.num_kv_heads,
                             block_size=args.block_size,
                             device=args.device,
-                            repeats=args.repeats,
-                            warmup_iters=args.warmup_iters,
                             profile_memory=args.profile_memory,
+                            warmup_ms=args.warmup_ms,
                             prefill_backend=pb,
                         )
 
@@ -980,16 +1200,17 @@ def main():
 
                         pbar.update(1)
 
-            console.print("\n[bold green]Prefill Backend Results:[/]")
-            formatter = ResultsFormatter(console)
-            formatter.print_table(
-                prefill_results, prefill_backends, compare_to_fastest=True
-            )
+            if not args.ncu_profile:
+                console.print("\n[bold green]Prefill Backend Results:[/]")
+                formatter = ResultsFormatter(console)
+                formatter.print_table(
+                    prefill_results, prefill_backends, compare_to_fastest=True
+                )
 
         all_results = decode_results + prefill_results
 
-    # Save results
-    if all_results:
+    # Save results (skip ncu profiling runs: timings are placeholder zeros)
+    if all_results and not args.ncu_profile:
         formatter = ResultsFormatter(console)
         if args.output_csv:
             formatter.save_csv(all_results, args.output_csv)

benchmarks/attention_benchmarks/common.py

@@ -15,6 +15,8 @@ from batch_spec import get_batch_type, parse_batch_spec
 from rich.console import Console
 from rich.table import Table
 
+from vllm.triton_utils import triton
+
 
 def batch_spec_sort_key(spec: str) -> tuple[int, int, int]:
     """
@@ -34,6 +36,30 @@ def batch_spec_sort_key(spec: str) -> tuple[int, int, int]:
         return (0, 0, 0)
 
 
+def run_do_bench(
+    benchmark_fn,
+    use_cuda_graphs: bool,
+    warmup_ms: int | None = None,
+) -> list[float]:
+    kwargs: dict[str, Any] = {"return_mode": "all"}
+    if use_cuda_graphs:
+        result = triton.testing.do_bench_cudagraph(benchmark_fn, **kwargs)
+    else:
+        if warmup_ms is not None:
+            kwargs["warmup"] = warmup_ms
+        result = triton.testing.do_bench(benchmark_fn, **kwargs)
+    return result
+
+
+def run_ncu_profile(benchmark_fn) -> None:
+    benchmark_fn()
+    torch.accelerator.synchronize()
+    torch.cuda.cudart().cudaProfilerStart()
+    benchmark_fn()
+    torch.accelerator.synchronize()
+    torch.cuda.cudart().cudaProfilerStop()
+
+
 # Mock classes for vLLM attention infrastructure
 
 
@@ -182,18 +208,37 @@ class ParameterSweep:
 
 @dataclass
 class ModelParameterSweep:
-    """Configuration for sweeping a model configuration parameter."""
+    """Configuration for sweeping model configuration parameter(s).
 
-    param_name: str  # Name of the model config parameter to sweep (e.g., "num_q_heads")
-    values: list[Any]  # List of values to test
-    label_format: str = "{backend}_{param_name}_{value}"  # Result label template
+    Supports two modes:
+    - Single param: param_name="head_dim", values=[128, 256, 512]
+    - Multi param: values=[{head_dim: 192, v_head_dim: 128}, {head_dim: 256}]
+      When values are dicts, each dict's keys are applied as config overrides.
+    """
+
+    param_name: str | None = None
+    values: list[Any] | None = None
+    label_format: str = "{backend}_{param_name}_{value}"
 
     def get_label(self, backend: str, value: Any) -> str:
         """Generate a label for a specific parameter value."""
+        if isinstance(value, dict):
+            return self.label_format.format(
+                backend=backend, param_name=self.param_name, value=value, **value
+            )
         return self.label_format.format(
             backend=backend, param_name=self.param_name, value=value
         )
 
+    def apply(self, config_args: dict, value: Any) -> None:
+        """Apply a sweep value to config args."""
+        if isinstance(value, dict):
+            config_args.update(value)
+        elif self.param_name is not None:
+            config_args[self.param_name] = value
+        else:
+            raise ValueError("param_name must be set if sweep values are not dicts")
+
 
 @dataclass
 class BenchmarkConfig:
@@ -208,10 +253,10 @@ class BenchmarkConfig:
     block_size: int
     device: str
     dtype: torch.dtype = torch.float16
-    repeats: int = 1
-    warmup_iters: int = 3
     profile_memory: bool = False
     use_cuda_graphs: bool = False
+    ncu_profile: bool = False
+    warmup_ms: int | None = None
 
     # "auto" or "fp8"
     kv_cache_dtype: str = "auto"
@@ -226,6 +271,7 @@ class BenchmarkConfig:
     # Backend-specific tuning
     num_kv_splits: int | None = None  # CUTLASS MLA
     reorder_batch_threshold: int | None = None  # FlashAttn MLA, FlashMLA
+    num_splits: int | None = None  # FlashAttention split-K (0=auto, 1=disabled)
 
 
 @dataclass
@@ -234,6 +280,7 @@ class BenchmarkResult:
 
     config: BenchmarkConfig
     mean_time: float  # seconds
+    median_time: float  # seconds
     std_time: float  # seconds
     min_time: float  # seconds
     max_time: float  # seconds
@@ -252,6 +299,7 @@ class BenchmarkResult:
         return {
             "config": asdict(self.config),
             "mean_time": self.mean_time,
+            "median_time": self.median_time,
             "std_time": self.std_time,
             "min_time": self.min_time,
             "max_time": self.max_time,

benchmarks/attention_benchmarks/configs/mla_decode.yaml

@@ -56,8 +56,6 @@ backends:
   - TOKENSPEED_MLA  # Blackwell + R1 dims + FP8 KV (use --kv-cache-dtype fp8)
 
 device: "cuda:0"
-repeats: 100
-warmup_iters: 10
 profile_memory: true
 
 # Backend-specific tuning

benchmarks/attention_benchmarks/configs/mla_fa4_fp8_output.yaml

@@ -0,0 +1,44 @@
+# MLA prefill FP8-output microbenchmark (FA4).
+# Compares the fused FP8 write against bf16 attention + a standalone static-FP8
+# quant; the delta is the post-quant kernel the fused path removes.
+# DeepSeek-Coder-V2-Lite dims; FA4 needs SM100/110.
+#
+# Usage:
+#   python benchmark.py --config configs/mla_fa4_fp8_output.yaml
+
+description: "MLA prefill FA4 fused-FP8 output vs post-quant"
+
+model:
+  name: "deepseek-v2-lite"
+  num_layers: 27
+  num_q_heads: 16
+  num_kv_heads: 1
+  head_dim: 576
+  kv_lora_rank: 512
+  qk_nope_head_dim: 128
+  qk_rope_head_dim: 64
+  v_head_dim: 128
+  block_size: 128
+
+# Pure prefill (q_len == kv_len) so every token goes through forward_mha.
+batch_specs:
+  - "q512"
+  - "q1k"
+  - "q2k"
+  - "q4k"
+  - "q8k"
+  - "2q4k"
+  - "4q4k"
+  - "8q4k"
+
+# Only used to construct the MLA impl; the pure-prefill specs skip decode.
+decode_backends:
+  - CUTLASS_MLA
+
+# Sweep the two FP8 write paths (prefill backend is fixed to fa4).
+fp8_output_scale: 0.1
+fuse_quant_op: [false, true]
+
+device: "cuda:0"
+repeats: 50
+warmup_iters: 10

benchmarks/attention_benchmarks/configs/mla_mixed_batch.yaml

@@ -51,8 +51,6 @@ backends:
   - FLASHMLA         # Hopper only
 
 device: "cuda:0"
-repeats: 5
-warmup_iters: 3
 profile_memory: true
 
 # Analyze chunked prefill workspace size impact

benchmarks/attention_benchmarks/configs/mla_prefill.yaml

@@ -124,5 +124,3 @@ prefill_backends:
   - tokenspeed
 
 device: "cuda:0"
-repeats: 20
-warmup_iters: 5

benchmarks/attention_benchmarks/configs/mla_sparse_decode.yaml

@@ -53,6 +53,4 @@ backends:
   - FLASHINFER_MLA_SPARSE
 
 device: "cuda:0"
-repeats: 100
-warmup_iters: 10
 profile_memory: true

benchmarks/attention_benchmarks/configs/mla_sparse_prefill.yaml

@@ -57,6 +57,4 @@ backends:
   - FLASHINFER_MLA_SPARSE
 
 device: "cuda:0"
-repeats: 10
-warmup_iters: 3
 profile_memory: true

benchmarks/attention_benchmarks/configs/reorder_threshold.yaml

@@ -63,8 +63,6 @@ model:
 
 # Benchmark settings
 device: "cuda:0"
-repeats: 15          # More repeats for spec decode variance
-warmup_iters: 5
 profile_memory: false
 
 # Output

benchmarks/attention_benchmarks/configs/speculative_decode.yaml

@@ -49,8 +49,6 @@ backends:
 
 # Benchmark settings
 device: "cuda:0"
-repeats: 10  # More repeats for statistical significance
-warmup_iters: 5
 profile_memory: false
 
 # Test these threshold values for optimization

benchmarks/attention_benchmarks/configs/standard_attention.yaml

@@ -43,6 +43,4 @@ backends:
   - FLASHINFER
 
 device: "cuda:0"
-repeats: 5
-warmup_iters: 3
 profile_memory: false

benchmarks/attention_benchmarks/configs/standard_decode.yaml

@@ -0,0 +1,142 @@
+# Standard attention decode benchmark configuration
+# Sweeps num_q_heads and num_kv_heads to isolate effects of:
+#   1. GQA ratio (fixed num_q_heads=32, vary num_kv_heads)
+#   2. Absolute head count (fixed 4:1 ratio, vary scale)
+
+model:
+  num_layers: 32
+  num_q_heads: 32 # Base value, overridden by sweep
+  num_kv_heads: 8 # Base value, overridden by sweep
+  head_dim: 128
+  block_size: 16
+
+# Head count sweep: each entry overrides num_q_heads, num_kv_heads, and
+# head_dim where it differs from the base (128). Head counts are per-GPU
+# (i.e. after TP sharding).
+#
+# Group A — vary GQA ratio (fixed q=32, head_dim=128):
+#   32:32 (MHA), 32:8 (GQA 4:1), 32:4 (GQA 8:1), 32:1 (MQA)
+#
+# Groups B-E — real model configs at various TP degrees:
+#   Model              head_dim  Full    TP2      TP4      TP8
+#   Llama 3 8B           128    32:8   16:4      8:2      4:1
+#   Llama 3 70B          128    64:8   32:4     16:2      8:1
+#   GPT-OSS 120B          64    64:8   32:4     16:2      8:1
+#   Llama 3 405B         128   128:8   64:4     32:2     16:1
+model_parameter_sweep:
+  values:
+    # --- head_dim=128 (Llama 3 family) ---
+    - { num_q_heads: 32, num_kv_heads: 32, head_dim: 128 } # MHA 1:1
+    - { num_q_heads: 32, num_kv_heads: 1, head_dim: 128 } # MQA 32:1
+    - { num_q_heads: 4, num_kv_heads: 1, head_dim: 128 } # Llama 3 8B TP8
+    - { num_q_heads: 8, num_kv_heads: 2, head_dim: 128 } # Llama 3 8B TP4
+    - { num_q_heads: 16, num_kv_heads: 4, head_dim: 128 } # Llama 3 8B TP2
+    - { num_q_heads: 32, num_kv_heads: 8, head_dim: 128 } # Llama 3 8B TP1 / GQA 4:1
+    - { num_q_heads: 8, num_kv_heads: 1, head_dim: 128 } # Llama 3 70B TP8
+    - { num_q_heads: 16, num_kv_heads: 2, head_dim: 128 } # Llama 3 70B TP4
+    - { num_q_heads: 32, num_kv_heads: 4, head_dim: 128 } # Llama 3 70B TP2 / GQA 8:1
+    - { num_q_heads: 64, num_kv_heads: 8, head_dim: 128 } # Llama 3 70B TP1
+    - { num_q_heads: 16, num_kv_heads: 1, head_dim: 128 } # Llama 3 405B TP8
+    - { num_q_heads: 32, num_kv_heads: 2, head_dim: 128 } # Llama 3 405B TP4
+    - { num_q_heads: 64, num_kv_heads: 4, head_dim: 128 } # Llama 3 405B TP2
+    - { num_q_heads: 128, num_kv_heads: 8, head_dim: 128 } # Llama 3 405B TP1
+    # --- head_dim=64 (GPT-OSS 120B) ---
+    - { num_q_heads: 8, num_kv_heads: 1, head_dim: 64 } # GPT-OSS 120B TP8
+    - { num_q_heads: 16, num_kv_heads: 2, head_dim: 64 } # GPT-OSS 120B TP4
+    - { num_q_heads: 32, num_kv_heads: 4, head_dim: 64 } # GPT-OSS 120B TP2
+    - { num_q_heads: 64, num_kv_heads: 8, head_dim: 64 } # GPT-OSS 120B TP1
+  label_format: "{backend}_q{num_q_heads}kv{num_kv_heads}d{head_dim}"
+
+batch_specs:
+  # ---- batch_size x seq_len grid (decode: q_len=1) ----
+  # Small grid for quick iteration. Uncomment for full sweep.
+
+  # Batch size 1
+  - "q1s1k"
+  - "q1s512"
+  - "q1s2k"
+  - "q1s4k"
+  - "q1s8k"
+  - "q1s16k"
+  - "q1s32k"
+
+  # Batch size 2
+  - "2q1s512"
+  - "2q1s1k"
+  - "2q1s2k"
+  - "2q1s4k"
+  - "2q1s8k"
+  - "2q1s16k"
+  - "2q1s32k"
+
+  # Batch size 4
+  - "4q1s512"
+  - "4q1s1k"
+  - "4q1s2k"
+  - "4q1s4k"
+  - "4q1s8k"
+  - "4q1s16k"
+  - "4q1s32k"
+
+  # Batch size 8
+  - "8q1s1k"
+  - "8q1s512"
+  - "8q1s2k"
+  - "8q1s4k"
+  - "8q1s8k"
+  - "8q1s16k"
+  - "8q1s32k"
+
+  # Batch size 16
+  - "16q1s512"
+  - "16q1s1k"
+  - "16q1s2k"
+  - "16q1s4k"
+  - "16q1s8k"
+  - "16q1s16k"
+  - "16q1s32k"
+
+  # Batch size 32
+  - "32q1s512"
+  - "32q1s1k"
+  - "32q1s2k"
+  - "32q1s4k"
+  - "32q1s8k"
+  - "32q1s16k"
+  - "32q1s32k"
+
+  # Batch size 64
+  - "64q1s1k"
+  - "64q1s512"
+  - "64q1s2k"
+  - "64q1s4k"
+  - "64q1s8k"
+  - "64q1s16k"
+  - "64q1s32k"
+
+  # Batch size 128
+  - "128q1s512"
+  - "128q1s1k"
+  - "128q1s2k"
+  - "128q1s4k"
+  - "128q1s8k"
+  - "128q1s16k"
+  - "128q1s32k"
+
+  # Batch size 256
+  - "256q1s1k"
+  - "256q1s512"
+  - "256q1s2k"
+  - "256q1s4k"
+  - "256q1s8k"
+  - "256q1s16k"
+  - "256q1s32k"
+
+# Available backends: FLASH_ATTN, TRITON_ATTN, FLASHINFER
+backends:
+  - FLASH_ATTN
+  - TRITON_ATTN
+  - FLASHINFER
+
+device: "cuda:0"
+profile_memory: false

benchmarks/attention_benchmarks/configs/standard_prefill.yaml

@@ -0,0 +1,108 @@
+# Standard attention prefill benchmark configuration
+# Sweeps num_q_heads and num_kv_heads to isolate effects of:
+#   1. GQA ratio (fixed num_q_heads=32, vary num_kv_heads)
+#   2. Absolute head count (fixed 4:1 ratio, vary scale)
+
+model:
+  num_layers: 32
+  num_q_heads: 32 # Base value, overridden by sweep
+  num_kv_heads: 8 # Base value, overridden by sweep
+  head_dim: 128
+  block_size: 16
+
+# Head count sweep: each entry overrides num_q_heads, num_kv_heads, and
+# head_dim where it differs from the base (128). Head counts are per-GPU
+# (i.e. after TP sharding).
+#
+# Group A — vary GQA ratio (fixed q=32, head_dim=128):
+#   32:32 (MHA), 32:8 (GQA 4:1), 32:4 (GQA 8:1), 32:1 (MQA)
+#
+# Groups B-E — real model configs at various TP degrees:
+#   Model              head_dim  Full    TP2      TP4      TP8
+#   Llama 3 8B           128    32:8   16:4      8:2      4:1
+#   Llama 3 70B          128    64:8   32:4     16:2      8:1
+#   GPT-OSS 120B          64    64:8   32:4     16:2      8:1
+#   Llama 3 405B         128   128:8   64:4     32:2     16:1
+model_parameter_sweep:
+  values:
+    # --- head_dim=128 (Llama 3 family) ---
+    - { num_q_heads: 32, num_kv_heads: 32, head_dim: 128 } # MHA 1:1
+    - { num_q_heads: 32, num_kv_heads: 1, head_dim: 128 } # MQA 32:1
+    - { num_q_heads: 4, num_kv_heads: 1, head_dim: 128 } # Llama 3 8B TP8
+    - { num_q_heads: 8, num_kv_heads: 2, head_dim: 128 } # Llama 3 8B TP4
+    - { num_q_heads: 16, num_kv_heads: 4, head_dim: 128 } # Llama 3 8B TP2
+    - { num_q_heads: 32, num_kv_heads: 8, head_dim: 128 } # Llama 3 8B TP1 / GQA 4:1
+    - { num_q_heads: 8, num_kv_heads: 1, head_dim: 128 } # Llama 3 70B TP8
+    - { num_q_heads: 16, num_kv_heads: 2, head_dim: 128 } # Llama 3 70B TP4
+    - { num_q_heads: 32, num_kv_heads: 4, head_dim: 128 } # Llama 3 70B TP2 / GQA 8:1
+    - { num_q_heads: 64, num_kv_heads: 8, head_dim: 128 } # Llama 3 70B TP1
+    - { num_q_heads: 16, num_kv_heads: 1, head_dim: 128 } # Llama 3 405B TP8
+    - { num_q_heads: 32, num_kv_heads: 2, head_dim: 128 } # Llama 3 405B TP4
+    - { num_q_heads: 64, num_kv_heads: 4, head_dim: 128 } # Llama 3 405B TP2
+    - { num_q_heads: 128, num_kv_heads: 8, head_dim: 128 } # Llama 3 405B TP1
+    # --- head_dim=64 (GPT-OSS 120B) ---
+    - { num_q_heads: 8, num_kv_heads: 1, head_dim: 64 } # GPT-OSS 120B TP8
+    - { num_q_heads: 16, num_kv_heads: 2, head_dim: 64 } # GPT-OSS 120B TP4
+    - { num_q_heads: 32, num_kv_heads: 4, head_dim: 64 } # GPT-OSS 120B TP2
+    - { num_q_heads: 64, num_kv_heads: 8, head_dim: 64 } # GPT-OSS 120B TP1
+  label_format: "{backend}_q{num_q_heads}kv{num_kv_heads}d{head_dim}"
+
+batch_specs:
+  # ---- batch_size x prefill_len grid (prefill: q_len == seq_len) ----
+  # Total tokens = batch_size * prefill_len, and prefill compute scales with
+  # prefill_len^2, so the largest cells are expensive. Trim batch sizes or
+  # lengths for quick iteration.
+
+  # Batch size 1
+  - "q512"
+  - "q1k"
+  - "q2k"
+  - "q4k"
+  - "q8k"
+  - "q16k"
+  - "q32k"
+
+  # Batch size 2
+  - "2q512"
+  - "2q1k"
+  - "2q2k"
+  - "2q4k"
+  - "2q8k"
+  - "2q16k"
+  - "2q32k"
+
+  # Batch size 4
+  - "4q512"
+  - "4q1k"
+  - "4q2k"
+  - "4q4k"
+  - "4q8k"
+  - "4q16k"
+  - "4q32k"
+
+  # Batch size 8
+  - "8q512"
+  - "8q1k"
+  - "8q2k"
+  - "8q4k"
+  - "8q8k"
+  - "8q16k"
+  - "8q32k"
+
+  # Batch size 16
+  - "16q512"
+  - "16q1k"
+  - "16q2k"
+  - "16q4k"
+  - "16q8k"
+  - "16q16k"
+  - "16q32k"
+
+# Available backends: FLASH_ATTN, TRITON_ATTN, FLASHINFER
+backends:
+  - FLASH_ATTN
+  - TRITON_ATTN
+  - FLASHINFER
+
+device: "cuda:0"
+profile_memory: false

benchmarks/attention_benchmarks/mla_runner.py

@@ -8,6 +8,8 @@ This module provides helpers for running MLA backends without
 needing full VllmConfig integration.
 """
 
+import statistics
+
 import numpy as np
 import torch
 from batch_spec import parse_batch_spec
@@ -17,6 +19,8 @@ from common import (
     MockIndexer,
     MockKVBProj,
     MockLayer,
+    run_do_bench,
+    run_ncu_profile,
     setup_mla_dims,
 )
 
@@ -704,6 +708,8 @@ def _run_single_benchmark(
     device: torch.device,
     indexer=None,
     kv_cache_dtype: str | None = None,
+    output_scale: float | None = None,
+    fuse_quant_op: bool = False,
 ) -> BenchmarkResult:
     """
     Run a single benchmark iteration.
@@ -717,6 +723,11 @@ def _run_single_benchmark(
         mla_dims: MLA dimension configuration
         device: Target device
         indexer: Optional MockIndexer for sparse backends
+        output_scale: Static per-tensor FP8 scale for prefill output. None
+            keeps the plain bf16 output (no quantization).
+        fuse_quant_op: With output_scale set, True lets the prefill kernel write
+            FP8 directly; False runs bf16 attention then a standalone static-FP8
+            quant. The delta isolates the saved post-quant kernel.
 
     Returns:
         BenchmarkResult with timing statistics
@@ -820,63 +831,86 @@ def _run_single_benchmark(
             num_prefill, mla_dims, query_fmt, device, torch.bfloat16
         )
 
-    # Build forward function
+    # Prefill FP8 output: fused (kernel writes e4m3) vs separate post-quant.
+    prefill_fp8_output = None
+    prefill_output_scale = None
+    prefill_quant_op = None
+    if has_prefill and output_scale is not None:
+        from vllm.platforms import current_platform
+
+        prefill_output_scale = torch.tensor(
+            [output_scale], device=device, dtype=torch.float32
+        )
+        if fuse_quant_op:
+            prefill_fp8_output = torch.empty_like(
+                prefill_inputs["output"], dtype=current_platform.fp8_dtype()
+            )
+        else:
+            from vllm.model_executor.layers.quantization.input_quant_fp8 import (
+                QuantFP8,
+            )
+            from vllm.model_executor.layers.quantization.utils.quant_utils import (
+                GroupShape,
+            )
+
+            prefill_quant_op = QuantFP8(static=True, group_shape=GroupShape.PER_TENSOR)
+
+    fused_output = output_scale is not None and fuse_quant_op
+
+    # Build forward function (runs a single decode/prefill pass)
     def forward_fn():
         results = []
         if has_decode:
             results.append(impl.forward_mqa(decode_inputs, kv_cache, metadata, layer))
         if has_prefill:
-            results.append(
-                impl.forward_mha(
-                    prefill_inputs["q"],
-                    prefill_inputs["k_c_normed"],
-                    prefill_inputs["k_pe"],
-                    kv_cache,
-                    metadata,
-                    prefill_inputs["k_scale"],
-                    prefill_inputs["output"],
-                )
+            out = impl.forward_mha(
+                prefill_inputs["q"],
+                prefill_inputs["k_c_normed"],
+                prefill_inputs["k_pe"],
+                kv_cache,
+                metadata,
+                prefill_inputs["k_scale"],
+                prefill_fp8_output if fused_output else prefill_inputs["output"],
+                prefill_output_scale if fused_output else None,
             )
+            if fused_output:
+                out = prefill_fp8_output
+            elif prefill_quant_op is not None:
+                out, _ = prefill_quant_op(
+                    prefill_inputs["output"], prefill_output_scale
+                )
+            results.append(out)
         return results[0] if len(results) == 1 else tuple(results)
 
-    # Warmup
-    for _ in range(config.warmup_iters):
-        forward_fn()
-    torch.accelerator.synchronize()
-
-    # Optionally capture a CUDA graph after warmup.
-    # Graph replay eliminates CPU launch overhead so timings reflect pure
-    # kernel time.
-    if config.use_cuda_graphs:
-        graph = torch.cuda.CUDAGraph()
-        with torch.cuda.graph(graph):
-            forward_fn()
-        benchmark_fn = graph.replay
-    else:
-        benchmark_fn = forward_fn
-
-    # Benchmark
-    times = []
-    for _ in range(config.repeats):
-        start = torch.cuda.Event(enable_timing=True)
-        end = torch.cuda.Event(enable_timing=True)
-
-        start.record()
+    def benchmark_fn():
         for _ in range(config.num_layers):
-            benchmark_fn()
-        end.record()
+            forward_fn()
 
-        torch.accelerator.synchronize()
-        elapsed_ms = start.elapsed_time(end)
-        times.append(elapsed_ms / 1000.0 / config.num_layers)
+    if config.ncu_profile:
+        run_ncu_profile(benchmark_fn)
+        return BenchmarkResult(
+            config=config,
+            mean_time=0.0,
+            median_time=0.0,
+            std_time=0.0,
+            min_time=0.0,
+            max_time=0.0,
+            throughput_tokens_per_sec=0.0,
+        )
+
+    all_ms = run_do_bench(benchmark_fn, config.use_cuda_graphs, config.warmup_ms)
+
+    # Convert ms to seconds per layer
+    times = [t / 1000.0 / config.num_layers for t in all_ms]
+    mean_time = statistics.mean(times)
 
-    mean_time = float(np.mean(times))
     return BenchmarkResult(
         config=config,
         mean_time=mean_time,
-        std_time=float(np.std(times)),
-        min_time=float(np.min(times)),
-        max_time=float(np.max(times)),
+        median_time=statistics.median(times),
+        std_time=statistics.stdev(times) if len(times) > 1 else 0.0,
+        min_time=min(times),
+        max_time=max(times),
         throughput_tokens_per_sec=total_q / mean_time if mean_time > 0 else 0,
     )
 
@@ -886,6 +920,8 @@ def _run_mla_benchmark_batched(
     configs_with_params: list[tuple],  # [(config, threshold, num_splits), ...]
     index_topk: int = 2048,
     prefill_backend: str | None = None,
+    output_scale: float | None = None,
+    fuse_quant_op: bool = False,
 ) -> list[BenchmarkResult]:
     """
     Unified batched MLA benchmark runner for all backends.
@@ -1025,6 +1061,8 @@ def _run_mla_benchmark_batched(
                     device,
                     indexer=indexer,
                     kv_cache_dtype=kv_cache_dtype,
+                    output_scale=output_scale,
+                    fuse_quant_op=fuse_quant_op,
                 )
                 results.append(result)
 
@@ -1052,6 +1090,8 @@ def run_mla_benchmark(
     num_kv_splits: int | None = None,
     index_topk: int = 2048,
     prefill_backend: str | None = None,
+    output_scale: float | None = None,
+    fuse_quant_op: bool = False,
 ) -> BenchmarkResult | list[BenchmarkResult]:
     """
     Unified MLA benchmark runner for all backends.
@@ -1071,6 +1111,9 @@ def run_mla_benchmark(
         index_topk: Topk value for sparse MLA backends (default 2048)
         prefill_backend: Prefill backend name (e.g., "fa3", "fa4").
             When set, forces the specified FlashAttention version for prefill.
+        output_scale: Static per-tensor FP8 scale for prefill output (None = bf16).
+        fuse_quant_op: With output_scale set, fuse the FP8 write into the prefill
+            kernel vs a standalone post-quant kernel. See _run_single_benchmark.
 
     Returns:
         BenchmarkResult (single mode) or list of BenchmarkResult (batched mode)
@@ -1095,7 +1138,12 @@ def run_mla_benchmark(
 
     # Use unified batched execution
     results = _run_mla_benchmark_batched(
-        backend, configs_with_params, index_topk, prefill_backend=prefill_backend
+        backend,
+        configs_with_params,
+        index_topk,
+        prefill_backend=prefill_backend,
+        output_scale=output_scale,
+        fuse_quant_op=fuse_quant_op,
     )
 
     # Return single result or list based on input

benchmarks/attention_benchmarks/runner.py

@@ -9,13 +9,20 @@ This module provides helpers for running standard attention backends
 """
 
 import logging
+import statistics
 import types
 from contextlib import contextmanager
 
-import numpy as np
 import torch
 from batch_spec import parse_batch_spec, reorder_for_flashinfer
-from common import BenchmarkConfig, BenchmarkResult, MockLayer, get_attention_scale
+from common import (
+    BenchmarkConfig,
+    BenchmarkResult,
+    MockLayer,
+    get_attention_scale,
+    run_do_bench,
+    run_ncu_profile,
+)
 
 from vllm.config import (
     CacheConfig,
@@ -208,6 +215,13 @@ def _create_backend_impl(
 
     scale = get_attention_scale(config.head_dim)
 
+    # Set v_head_dim for diff-headdim backends. Always reset (defaulting to
+    # head_dim) so a prior run's value doesn't leak into this one via the
+    # backend's class-level state.
+    if hasattr(backend_class, "set_head_size_v"):
+        v_dim = config.v_head_dim if config.v_head_dim is not None else config.head_dim
+        backend_class.set_head_size_v(v_dim)
+
     impl = backend_class.get_impl_cls()(
         num_heads=config.num_q_heads,
         head_size=config.head_dim,
@@ -300,6 +314,7 @@ def _create_input_tensors(
         from vllm.platforms import current_platform
 
         q_dtype = current_platform.fp8_dtype()
+    v_dim = config.v_head_dim if config.v_head_dim is not None else config.head_dim
     q_list = [
         torch.randn(
             total_q, config.num_q_heads, config.head_dim, device=device, dtype=dtype
@@ -313,9 +328,7 @@ def _create_input_tensors(
         for _ in range(config.num_layers)
     ]
     v_list = [
-        torch.randn(
-            total_q, config.num_kv_heads, config.head_dim, device=device, dtype=dtype
-        )
+        torch.randn(total_q, config.num_kv_heads, v_dim, device=device, dtype=dtype)
         for _ in range(config.num_layers)
     ]
     return q_list, k_list, v_list
@@ -389,14 +402,17 @@ def _run_single_benchmark(
     device: torch.device,
     dtype: torch.dtype,
 ) -> tuple:
-    """Run single benchmark iteration with warmup and timing loop."""
-    total_q = q_list[0].shape[0]
-    out = torch.empty(
-        total_q, config.num_q_heads, config.head_dim, device=device, dtype=dtype
-    )
+    """Run single benchmark using triton's do_bench_cudagraph/do_bench.
 
-    # Warmup
-    for _ in range(config.warmup_iters):
+    Returns:
+        (timing_stats, mem_stats) where timing_stats is a dict with
+        mean/std/min/max in seconds per layer.
+    """
+    total_q = q_list[0].shape[0]
+    v_dim = config.v_head_dim if config.v_head_dim is not None else config.head_dim
+    out = torch.empty(total_q, config.num_q_heads, v_dim, device=device, dtype=dtype)
+
+    def benchmark_fn():
         for i in range(config.num_layers):
             impl.forward(
                 layer,
@@ -407,52 +423,22 @@ def _run_single_benchmark(
                 attn_metadata,
                 output=out,
             )
-    torch.accelerator.synchronize()
 
-    # Optionally capture a CUDA graph after warmup.
-    # Graph replay eliminates CPU launch overhead so timings reflect pure
-    # kernel time.
-    if config.use_cuda_graphs:
-        graph = torch.cuda.CUDAGraph()
-        with torch.cuda.graph(graph):
-            for i in range(config.num_layers):
-                impl.forward(
-                    layer,
-                    q_list[i],
-                    k_list[i],
-                    v_list[i],
-                    cache_list[i],
-                    attn_metadata,
-                    output=out,
-                )
-        benchmark_fn = graph.replay
+    if config.ncu_profile:
+        run_ncu_profile(benchmark_fn)
+        timing_stats = dict.fromkeys(("mean", "median", "std", "min", "max"), 0.0)
     else:
+        all_ms = run_do_bench(benchmark_fn, config.use_cuda_graphs, config.warmup_ms)
 
-        def benchmark_fn():
-            for i in range(config.num_layers):
-                impl.forward(
-                    layer,
-                    q_list[i],
-                    k_list[i],
-                    v_list[i],
-                    cache_list[i],
-                    attn_metadata,
-                    output=out,
-                )
-
-    # Benchmark
-    times = []
-    for _ in range(config.repeats):
-        start = torch.cuda.Event(enable_timing=True)
-        end = torch.cuda.Event(enable_timing=True)
-
-        start.record()
-        benchmark_fn()
-        end.record()
-
-        torch.accelerator.synchronize()
-        elapsed_ms = start.elapsed_time(end)
-        times.append(elapsed_ms / 1000.0 / config.num_layers)  # seconds per layer
+        # Convert ms to seconds per layer
+        times = [t / 1000.0 / config.num_layers for t in all_ms]
+        timing_stats = {
+            "mean": statistics.mean(times),
+            "std": statistics.stdev(times) if len(times) > 1 else 0.0,
+            "min": min(times),
+            "max": max(times),
+            "median": statistics.median(times),
+        }
 
     mem_stats = {}
     if config.profile_memory:
@@ -461,7 +447,7 @@ def _run_single_benchmark(
             "reserved_mb": torch.accelerator.memory_reserved(device) / 1024**2,
         }
 
-    return times, mem_stats
+    return timing_stats, mem_stats
 
 
 # ============================================================================
@@ -541,6 +527,12 @@ def run_attention_benchmark(config: BenchmarkConfig) -> BenchmarkResult:
                 common_attn_metadata=common_metadata,
             )
 
+            # Override num_splits for split-K testing (FlashAttention only)
+            if config.num_splits is not None and hasattr(
+                attn_metadata, "max_num_splits"
+            ):
+                attn_metadata.max_num_splits = config.num_splits
+
             # Only quantize queries when the impl supports it
             quantize_query = config.kv_cache_dtype.startswith("fp8") and getattr(
                 impl, "supports_quant_query_input", False
@@ -553,7 +545,7 @@ def run_attention_benchmark(config: BenchmarkConfig) -> BenchmarkResult:
                 config, max_num_blocks, backend_class, device, dtype
             )
 
-            times, mem_stats = _run_single_benchmark(
+            timing_stats, mem_stats = _run_single_benchmark(
                 config,
                 impl,
                 layer,
@@ -566,15 +558,16 @@ def run_attention_benchmark(config: BenchmarkConfig) -> BenchmarkResult:
                 dtype,
             )
 
-    mean_time = np.mean(times)
+    mean_time = timing_stats["mean"]
     throughput = total_q / mean_time if mean_time > 0 else 0
 
     return BenchmarkResult(
         config=config,
         mean_time=mean_time,
-        std_time=np.std(times),
-        min_time=np.min(times),
-        max_time=np.max(times),
+        median_time=timing_stats["median"],
+        std_time=timing_stats["std"],
+        min_time=timing_stats["min"],
+        max_time=timing_stats["max"],
         throughput_tokens_per_sec=throughput,
         memory_allocated_mb=mem_stats.get("allocated_mb"),
         memory_reserved_mb=mem_stats.get("reserved_mb"),

benchmarks/benchmark_hidden_state_extraction.py

@@ -92,7 +92,6 @@ def run_baseline(
     llm = LLM(
         model=model,
         enable_prefix_caching=False,
-        enable_chunked_prefill=False,
         **extra_args,
     )
     sampling_params = SamplingParams(max_tokens=1)
@@ -194,7 +193,6 @@ async def _run_extraction_async(
     engine_args = AsyncEngineArgs(
         model=model,
         enable_prefix_caching=False,
-        enable_chunked_prefill=False,
         max_num_batched_tokens=40960,
         max_model_len=40960,
         speculative_config={

benchmarks/benchmark_pin_memory.py

@@ -0,0 +1,358 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""Benchmark and regression-test pinned (page-locked) CPU memory for vLLM.
+
+Verifies that enabling pinned memory does not regress throughput or latency
+compared to unpinned memory.  Each condition runs in an isolated ``spawn``
+subprocess so both start from a cold CUDA context, giving an unbiased
+comparison.
+
+Usage
+-----
+Run all tests with the default model::
+
+    python benchmarks/benchmark_pin_memory.py -v
+
+Override the model and optional max-model-len::
+
+    python benchmarks/benchmark_pin_memory.py --model unsloth/Qwen3-1.7B -v
+    python benchmarks/benchmark_pin_memory.py --model unsloth/Qwen3-1.7B \
+        --max-model-len 8192 -v
+
+Run only throughput or latency tests::
+
+    python benchmarks/benchmark_pin_memory.py -v -k test_throughput
+    python benchmarks/benchmark_pin_memory.py -v -k test_latency
+
+Run only the v1 or v2 runner variant::
+
+    python benchmarks/benchmark_pin_memory.py -v -k v1
+    python benchmarks/benchmark_pin_memory.py -v -k v2
+
+Note: on WSL2, v1 runner tests are skipped because pin memory is not available
+for the v1 runner without cpu_offload_gb.  Run on other platforms to exercise v1.
+"""
+
+import argparse
+import json
+import multiprocessing
+import sys
+import tempfile
+
+import pytest
+
+# Allow up to 2% degradation.  Both benchmark runs start from an identical
+# cold CUDA context (separate spawn subprocesses), so the measured difference
+# reflects the genuine pin_memory overhead rather than cold/warm ordering bias.
+_THROUGHPUT_TOLERANCE = 0.98
+_THROUGHPUT_NUM_REQUESTS = 200
+_THROUGHPUT_INPUT_LEN = 128
+_THROUGHPUT_OUTPUT_LEN = 512
+_THROUGHPUT_MAX_NUM_SEQS = 128
+
+# Latency benchmark constants — match latency.py defaults.
+_LATENCY_TOLERANCE = 1.02  # Allow up to 2% latency regression.
+_LATENCY_BATCH_SIZE = 64
+_LATENCY_INPUT_LEN = 32
+_LATENCY_OUTPUT_LEN = 128
+_LATENCY_WARMUP_ITERS = 5
+_LATENCY_BENCH_ITERS = 15
+
+_DEFAULT_MODEL = "unsloth/Qwen3-1.7B"
+_DEFAULT_MAX_MODEL_LEN = 16384
+
+
+def _benchmark_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(add_help=False)
+    parser.add_argument("--model", default=_DEFAULT_MODEL)
+    parser.add_argument("--max-model-len", type=int, default=_DEFAULT_MAX_MODEL_LEN)
+    args, _ = parser.parse_known_args()
+    return args
+
+
+@pytest.fixture
+def model() -> str:
+    return _benchmark_args().model
+
+
+@pytest.fixture
+def max_model_len() -> int:
+    return _benchmark_args().max_model_len
+
+
+def _skip_if_pin_memory_not_available(engine_args_kwargs: dict) -> None:
+    """Skip the current pytest test if pin_memory is unavailable for this config."""
+    import vllm.utils.platform_utils as pu
+    from vllm.config import set_current_vllm_config
+    from vllm.engine.arg_utils import EngineArgs
+
+    vllm_config = EngineArgs(**engine_args_kwargs).create_engine_config()
+    with set_current_vllm_config(vllm_config):
+        pu.is_pin_memory_available.cache_clear()
+        if not pu.is_pin_memory_available():
+            import os
+
+            runner = "v2" if os.environ.get("VLLM_USE_V2_MODEL_RUNNER") == "1" else "v1"
+            model = engine_args_kwargs.get("model", "unknown")
+            print(
+                f"\033[33mSKIP: pin_memory not available for "
+                f"{runner} runner, model={model}\033[0m"
+            )
+            pytest.skip("pin_memory not available for this configuration")
+
+
+def _throughput_worker(
+    pin: bool,
+    engine_args_kwargs: dict,
+    q: "multiprocessing.Queue[float]",
+    v2_mode: bool = False,
+) -> None:
+    """Run throughput benchmark in a fresh spawn subprocess.
+
+    Delegates to vllm/benchmarks/throughput.py main() using the random dataset,
+    so the methodology matches the official benchmark.  Results are written to a
+    temp JSON file and forwarded through the queue as tokens/s.
+
+    v2_mode: when True, monkeypatches is_uva_available() to always return True
+    so the v2 model runner's UVA buffers remain functional even when pin=False.
+    This isolates the non-UVA pin_memory paths in v2.
+    """
+    import vllm.utils.platform_utils as pu
+    from vllm.platforms import current_platform
+
+    pu.is_pin_memory_available.cache_clear()
+    pu.is_uva_available.cache_clear()
+    type(current_platform).is_pin_memory_available = classmethod(lambda cls: pin)
+    if v2_mode:
+        pu.is_uva_available = lambda: True
+
+    from vllm.benchmarks.throughput import add_cli_args
+    from vllm.benchmarks.throughput import main as throughput_main
+
+    parser = argparse.ArgumentParser()
+    add_cli_args(parser)
+    args = parser.parse_args([])
+
+    for key, val in engine_args_kwargs.items():
+        setattr(args, key, val)
+    args.max_num_seqs = _THROUGHPUT_MAX_NUM_SEQS
+    args.dataset_name = "random"
+    args.input_len = _THROUGHPUT_INPUT_LEN
+    args.output_len = _THROUGHPUT_OUTPUT_LEN
+    # Nullify defaults that conflict with explicit input/output_len.
+    args.random_input_len = None
+    args.random_output_len = None
+    args.random_prefix_len = None
+    args.num_prompts = _THROUGHPUT_NUM_REQUESTS
+    args.seed = 0
+    args.disable_detokenize = True
+
+    with tempfile.NamedTemporaryFile(mode="w", suffix=".json", delete=False) as f:
+        tmp_path = f.name
+    args.output_json = tmp_path
+
+    throughput_main(args)
+
+    with open(tmp_path) as f:
+        results = json.load(f)
+    q.put(results["tokens_per_second"])
+
+
+def _run_throughput_benchmark(
+    pin: bool,
+    engine_args_kwargs: dict,
+    v2_mode: bool = False,
+) -> float:
+    ctx = multiprocessing.get_context("spawn")
+    q = ctx.Queue()
+    p = ctx.Process(
+        target=_throughput_worker,
+        args=(pin, engine_args_kwargs, q, v2_mode),
+    )
+    p.start()
+    p.join()
+    if p.exitcode != 0:
+        raise RuntimeError(
+            f"Throughput benchmark subprocess (pin={pin}) exited with code {p.exitcode}"
+        )
+    return q.get()
+
+
+def _latency_worker(
+    pin: bool,
+    engine_args_kwargs: dict,
+    q: "multiprocessing.Queue[dict]",
+    v2_mode: bool = False,
+) -> None:
+    """Run latency benchmark in a fresh spawn subprocess.
+
+    Follows latency.py methodology: fixed batch of dummy token IDs, warmup
+    iterations to reach steady state, then timed iterations reduced to avg
+    and percentiles.  Results are written to a temp JSON file by latency_main
+    and forwarded through the queue.
+    """
+    import vllm.utils.platform_utils as pu
+    from vllm.platforms import current_platform
+
+    pu.is_pin_memory_available.cache_clear()
+    pu.is_uva_available.cache_clear()
+    type(current_platform).is_pin_memory_available = classmethod(lambda cls: pin)
+    if v2_mode:
+        pu.is_uva_available = lambda: True
+
+    from vllm.benchmarks.latency import add_cli_args
+    from vllm.benchmarks.latency import main as latency_main
+
+    parser = argparse.ArgumentParser()
+    add_cli_args(parser)
+    args = parser.parse_args([])
+
+    for key, val in engine_args_kwargs.items():
+        setattr(args, key, val)
+    args.input_len = _LATENCY_INPUT_LEN
+    args.output_len = _LATENCY_OUTPUT_LEN
+    args.batch_size = _LATENCY_BATCH_SIZE
+    args.num_iters_warmup = _LATENCY_WARMUP_ITERS
+    args.num_iters = _LATENCY_BENCH_ITERS
+    args.profile = False
+    args.disable_detokenize = True
+
+    with tempfile.NamedTemporaryFile(mode="w", suffix=".json", delete=False) as f:
+        tmp_path = f.name
+    args.output_json = tmp_path
+
+    latency_main(args)
+
+    with open(tmp_path) as f:
+        results = json.load(f)
+    q.put(results)
+
+
+def _run_latency_benchmark(
+    pin: bool,
+    engine_args_kwargs: dict,
+    v2_mode: bool = False,
+) -> dict:
+    ctx = multiprocessing.get_context("spawn")
+    q = ctx.Queue()
+    p = ctx.Process(
+        target=_latency_worker,
+        args=(pin, engine_args_kwargs, q, v2_mode),
+    )
+    p.start()
+    p.join()
+    if p.exitcode != 0:
+        raise RuntimeError(
+            f"Latency benchmark subprocess (pin={pin}) exited with code {p.exitcode}"
+        )
+    return q.get()
+
+
+@pytest.mark.parametrize(
+    "test_v2_runner",
+    [
+        pytest.param(False, id="v1"),
+        pytest.param(True, id="v2"),
+    ],
+)
+class TestPinnedMemory:
+    """Verify pinned memory yields >= throughput vs unpinned via real vLLM inference."""
+
+    def test_throughput(self, monkeypatch, test_v2_runner, model, max_model_len):
+        """Benchmark throughput with pin_memory forced on then off.
+
+        Delegates to vllm/benchmarks/throughput.py main() with the random
+        dataset.  Each condition runs in an isolated spawn subprocess so both
+        start from a cold CUDA context, giving an unbiased comparison.
+        """
+        monkeypatch.setenv("VLLM_ENABLE_V1_MULTIPROCESSING", "0")
+        monkeypatch.setenv("VLLM_USE_V2_MODEL_RUNNER", "1" if test_v2_runner else "0")
+
+        engine_args_kwargs = dict(
+            model=model,
+            gpu_memory_utilization=0.88,
+            max_model_len=max_model_len,
+            enable_prefix_caching=False,
+        )
+
+        _skip_if_pin_memory_not_available(engine_args_kwargs)
+
+        unpinned_tps = _run_throughput_benchmark(
+            False, engine_args_kwargs, v2_mode=test_v2_runner
+        )
+        pinned_tps = _run_throughput_benchmark(
+            True, engine_args_kwargs, v2_mode=test_v2_runner
+        )
+
+        pct_diff = (pinned_tps - unpinned_tps) / unpinned_tps * 100
+        runner = "v2" if test_v2_runner else "v1"
+        print(
+            f"\n=== Throughput results ({runner} runner, {model}) ==="
+            f"\npin_memory=True:  {pinned_tps:.1f} tok/s"
+            f"\npin_memory=False: {unpinned_tps:.1f} tok/s"
+            f"\nDifference: {pct_diff:+.1f}% (pinned vs unpinned)"
+        )
+
+        assert pinned_tps >= unpinned_tps * _THROUGHPUT_TOLERANCE, (
+            f"Pinned throughput ({pinned_tps:.1f} tok/s) fell more than "
+            f"{(1.0 - _THROUGHPUT_TOLERANCE) * 100:.1f}% below "
+            f"unpinned ({unpinned_tps:.1f} tok/s)."
+        )
+
+    def test_latency(self, monkeypatch, test_v2_runner, model, max_model_len):
+        """Benchmark per-batch latency with pin_memory forced on then off.
+
+        Follows vllm/benchmarks/latency.py: fixed dummy-token batch, warmup
+        iterations to reach steady state, then timed iterations reduced to avg
+        and percentiles.  Subprocesses run serially so each gets a cold CUDA
+        context without GPU memory pressure from the other run.
+        """
+        monkeypatch.setenv("VLLM_ENABLE_V1_MULTIPROCESSING", "0")
+        monkeypatch.setenv("VLLM_USE_V2_MODEL_RUNNER", "1" if test_v2_runner else "0")
+
+        engine_args_kwargs = dict(
+            model=model,
+            gpu_memory_utilization=0.88,
+            max_model_len=max_model_len,
+            enable_prefix_caching=False,
+        )
+
+        _skip_if_pin_memory_not_available(engine_args_kwargs)
+
+        unpinned = _run_latency_benchmark(
+            False, engine_args_kwargs, v2_mode=test_v2_runner
+        )
+        pinned = _run_latency_benchmark(
+            True, engine_args_kwargs, v2_mode=test_v2_runner
+        )
+
+        pct_diff = (
+            (pinned["avg_latency"] - unpinned["avg_latency"])
+            / unpinned["avg_latency"]
+            * 100
+        )
+        runner = "v2" if test_v2_runner else "v1"
+        print(
+            f"\n=== Latency results ({runner} runner, {model}) ==="
+            f"\npin_memory=True:  avg={pinned['avg_latency']:.3f}s"
+            f"  p50={pinned['percentiles']['50']:.3f}s"
+            f"  p99={pinned['percentiles']['99']:.3f}s"
+            f"\npin_memory=False: avg={unpinned['avg_latency']:.3f}s"
+            f"  p50={unpinned['percentiles']['50']:.3f}s"
+            f"  p99={unpinned['percentiles']['99']:.3f}s"
+            f"\nDifference: {pct_diff:+.1f}% (pinned vs unpinned)"
+        )
+
+        assert pinned["avg_latency"] <= unpinned["avg_latency"] * _LATENCY_TOLERANCE, (
+            f"Pinned avg latency ({pinned['avg_latency']:.3f}s) exceeded "
+            f"unpinned ({unpinned['avg_latency']:.3f}s) by more than "
+            f"{(_LATENCY_TOLERANCE - 1.0) * 100:.1f}%."
+        )
+
+
+if __name__ == "__main__":
+    _parser = argparse.ArgumentParser(add_help=False)
+    _parser.add_argument("--model", default=_DEFAULT_MODEL)
+    _parser.add_argument("--max-model-len", type=int, default=_DEFAULT_MAX_MODEL_LEN)
+    _, _remaining = _parser.parse_known_args()
+    sys.exit(pytest.main([__file__] + _remaining))

benchmarks/disagg_benchmarks/disagg_overhead_benchmark.sh

@@ -1,143 +0,0 @@
-#!/bin/bash
-
-# benchmark the overhead of disaggregated prefill.
-# methodology:
-# - send all request to prefill vLLM instance. It will buffer KV cache.
-# - then send all request to decode instance.
-# - The TTFT of decode instance is the overhead.
-
-set -ex
-
-kill_gpu_processes() {
-  # kill all processes on GPU.
-  pgrep pt_main_thread | xargs -r kill -9
-  pgrep python3 | xargs -r kill -9
-  # vLLM now names the process with VLLM prefix after https://github.com/vllm-project/vllm/pull/21445
-  pgrep VLLM | xargs -r kill -9
-  sleep 10
-
-  # remove vllm config file
-  rm -rf ~/.config/vllm
-
-  # Print the GPU memory usage
-  # so that we know if all GPU processes are killed.
-  gpu_memory_usage=$(nvidia-smi --query-gpu=memory.used --format=csv,noheader,nounits -i 0)
-  # The memory usage should be 0 MB.
-  echo "GPU 0 Memory Usage: $gpu_memory_usage MB"
-}
-
-wait_for_server() {
-  # wait for vllm server to start
-  # return 1 if vllm server crashes
-  local port=$1
-  timeout 1200 bash -c "
-    until curl -s localhost:${port}/v1/completions > /dev/null; do
-      sleep 1
-    done" && return 0 || return 1
-}
-
-
-benchmark() {
-
-  export VLLM_LOGGING_LEVEL=DEBUG
-  export VLLM_HOST_IP=$(hostname -I | awk '{print $1}')
-
-  # compare chunked prefill with disaggregated prefill
-
-  results_folder="./results"
-  model="meta-llama/Meta-Llama-3.1-8B-Instruct"
-  dataset_name="sonnet"
-  dataset_path="../sonnet_4x.txt"
-  num_prompts=10
-  qps=$1
-  prefix_len=50
-  input_len=2048
-  output_len=$2
-
-
-  CUDA_VISIBLE_DEVICES=0 vllm serve $model \
-    --port 8100 \
-    --max-model-len 10000 \
-    --gpu-memory-utilization 0.6 \
-    --kv-transfer-config \
-    '{"kv_connector":"P2pNcclConnector","kv_role":"kv_producer","kv_rank":0,"kv_parallel_size":2,"kv_buffer_size":5e9}' &
-
-
-  CUDA_VISIBLE_DEVICES=1 vllm serve $model \
-    --port 8200 \
-    --max-model-len 10000 \
-    --gpu-memory-utilization 0.6 \
-    --kv-transfer-config \
-    '{"kv_connector":"P2pNcclConnector","kv_role":"kv_consumer","kv_rank":1,"kv_parallel_size":2,"kv_buffer_size":5e9}' &
-
-  wait_for_server 8100
-  wait_for_server 8200
-
-  # let the prefill instance finish prefill
-  vllm bench serve \
-    --backend vllm \
-    --model $model \
-    --dataset-name $dataset_name \
-    --dataset-path $dataset_path \
-    --sonnet-input-len $input_len \
-    --sonnet-output-len "$output_len" \
-    --sonnet-prefix-len $prefix_len \
-    --num-prompts $num_prompts \
-    --port 8100 \
-    --save-result \
-    --result-dir $results_folder \
-    --result-filename disagg_prefill_tp1.json \
-    --request-rate "inf"
-
-
-  # send the request to decode.
-  # The TTFT of this command will be the overhead of disagg prefill impl.
-  vllm bench serve \
-    --backend vllm \
-    --model $model \
-    --dataset-name $dataset_name \
-    --dataset-path $dataset_path \
-    --sonnet-input-len $input_len \
-    --sonnet-output-len "$output_len" \
-    --sonnet-prefix-len $prefix_len \
-    --num-prompts $num_prompts \
-    --port 8200 \
-    --save-result \
-    --result-dir $results_folder \
-    --result-filename disagg_prefill_tp1_overhead.json \
-    --request-rate "$qps"
-  kill_gpu_processes
-
-}
-
-
-main() {
-
-  (which wget && which curl) || (apt-get update && apt-get install -y wget curl)
-  (which jq) || (apt-get -y install jq)
-  (which socat) || (apt-get -y install socat)
-
-  pip install quart httpx datasets
-
-  cd "$(dirname "$0")"
-
-  cd ..
-  # create sonnet-4x.txt
-  echo "" > sonnet_4x.txt
-  for _ in {1..4}
-  do
-    cat sonnet.txt >> sonnet_4x.txt
-  done
-  cd disagg_benchmarks
-
-  rm -rf results
-  mkdir results
-
-  default_qps=1
-  default_output_len=1
-  benchmark $default_qps $default_output_len
-
-}
-
-
-main "$@"

benchmarks/disagg_benchmarks/disagg_performance_benchmark.sh

@@ -1,157 +0,0 @@
-#!/bin/bash
-
-# Requirement: 2x GPUs.
-
-
-# Model: meta-llama/Meta-Llama-3.1-8B-Instruct
-# Query: 1024 input tokens, 6 output tokens, QPS 2/4/6/8, 100 requests
-# Resource: 2x GPU
-# Approaches:
-# 2. Chunked prefill: 2 vllm instance with tp=4, equivalent to 1 tp=4 instance with QPS 4
-# 3. Disaggregated prefill: 1 prefilling instance and 1 decoding instance
-# Prefilling instance: max_output_token=1
-# Decoding instance: force the input tokens be the same across requests to bypass prefilling
-
-set -ex
-
-kill_gpu_processes() {
-  # kill all processes on GPU.
-  pgrep pt_main_thread | xargs -r kill -9
-  pgrep python3 | xargs -r kill -9
-  # vLLM now names the process with VLLM prefix after https://github.com/vllm-project/vllm/pull/21445
-  pgrep VLLM | xargs -r kill -9
-  for port in 8000 8100 8200; do lsof -t -i:$port | xargs -r kill -9; done
-  sleep 1
-}
-
-wait_for_server() {
-  # wait for vllm server to start
-  # return 1 if vllm server crashes
-  local port=$1
-  timeout 1200 bash -c "
-    until curl -s localhost:${port}/v1/completions > /dev/null; do
-      sleep 1
-    done" && return 0 || return 1
-}
-
-
-launch_chunked_prefill() {
-  model="meta-llama/Meta-Llama-3.1-8B-Instruct"
-  # disagg prefill
-  CUDA_VISIBLE_DEVICES=0 vllm serve $model \
-    --port 8100 \
-    --max-model-len 10000 \
-    --enable-chunked-prefill \
-    --gpu-memory-utilization 0.6 &
-  CUDA_VISIBLE_DEVICES=1 vllm serve $model \
-    --port 8200 \
-    --max-model-len 10000 \
-    --enable-chunked-prefill \
-    --gpu-memory-utilization 0.6 &
-  wait_for_server 8100
-  wait_for_server 8200
-  python3 round_robin_proxy.py &
-  sleep 1
-}
-
-
-launch_disagg_prefill() {
-  model="meta-llama/Meta-Llama-3.1-8B-Instruct"
-  # disagg prefill
-  CUDA_VISIBLE_DEVICES=0 vllm serve $model \
-    --port 8100 \
-    --max-model-len 10000 \
-    --gpu-memory-utilization 0.6 \
-    --kv-transfer-config \
-    '{"kv_connector":"P2pNcclConnector","kv_role":"kv_producer","kv_rank":0,"kv_parallel_size":2,"kv_buffer_size":5e9}' &
-
-  CUDA_VISIBLE_DEVICES=1 vllm serve $model \
-    --port 8200 \
-    --max-model-len 10000 \
-    --gpu-memory-utilization 0.6 \
-    --kv-transfer-config \
-    '{"kv_connector":"P2pNcclConnector","kv_role":"kv_consumer","kv_rank":1,"kv_parallel_size":2,"kv_buffer_size":5e9}' &
-
-  wait_for_server 8100
-  wait_for_server 8200
-  python3 disagg_prefill_proxy_server.py &
-  sleep 1
-}
-
-
-benchmark() {
-  results_folder="./results"
-  model="meta-llama/Meta-Llama-3.1-8B-Instruct"
-  dataset_name="sonnet"
-  dataset_path="../sonnet_4x.txt"
-  num_prompts=100
-  qps=$1
-  prefix_len=50
-  input_len=1024
-  output_len=$2
-  tag=$3
-
-  vllm bench serve \
-    --backend vllm \
-    --model $model \
-    --dataset-name $dataset_name \
-    --dataset-path $dataset_path \
-    --sonnet-input-len $input_len \
-    --sonnet-output-len "$output_len" \
-    --sonnet-prefix-len $prefix_len \
-    --num-prompts $num_prompts \
-    --port 8000 \
-    --save-result \
-    --result-dir $results_folder \
-    --result-filename "$tag"-qps-"$qps".json \
-    --request-rate "$qps"
-
-  sleep 2
-}
-
-
-main() {
-
-  (which wget && which curl) || (apt-get update && apt-get install -y wget curl)
-  (which jq) || (apt-get -y install jq)
-  (which socat) || (apt-get -y install socat)
-  (which lsof) || (apt-get -y install lsof)
-
-  pip install quart httpx matplotlib aiohttp datasets
-
-  cd "$(dirname "$0")"
-
-  cd ..
-  # create sonnet-4x.txt so that we can sample 2048 tokens for input
-  echo "" > sonnet_4x.txt
-  for _ in {1..4}
-  do
-    cat sonnet.txt >> sonnet_4x.txt
-  done
-  cd disagg_benchmarks
-
-  rm -rf results
-  mkdir results
-
-  default_output_len=6
-
-  export VLLM_HOST_IP=$(hostname -I | awk '{print $1}')
-
-  launch_chunked_prefill
-  for qps in 2 4 6 8; do
-  benchmark $qps $default_output_len chunked_prefill
-  done
-  kill_gpu_processes
-
-  launch_disagg_prefill
-  for qps in 2 4 6 8; do
-  benchmark $qps $default_output_len disagg_prefill
-  done
-  kill_gpu_processes
-
-  python3 visualize_benchmark_results.py
-
-}
-
-
-main "$@"

benchmarks/disagg_benchmarks/disagg_prefill_proxy_server.py

@@ -1,260 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-
-import argparse
-import asyncio
-import logging
-import os
-import time
-import uuid
-from urllib.parse import urlparse
-
-import aiohttp
-from quart import Quart, Response, make_response, request
-
-# Configure logging
-logging.basicConfig(level=logging.INFO)
-logger = logging.getLogger(__name__)
-
-
-def parse_args():
-    """parse command line arguments"""
-    parser = argparse.ArgumentParser(description="vLLM P/D disaggregation proxy server")
-
-    # Add args
-    parser.add_argument(
-        "--timeout",
-        type=float,
-        default=6 * 60 * 60,
-        help="Timeout for backend service requests in seconds (default: 21600)",
-    )
-    parser.add_argument(
-        "--port",
-        type=int,
-        default=8000,
-        help="Port to run the server on (default: 8000)",
-    )
-    parser.add_argument(
-        "--prefill-url",
-        type=str,
-        default="http://localhost:8100",
-        help="Prefill service base URL (protocol + host[:port])",
-    )
-    parser.add_argument(
-        "--decode-url",
-        type=str,
-        default="http://localhost:8200",
-        help="Decode service base URL (protocol + host[:port])",
-    )
-    parser.add_argument(
-        "--kv-host",
-        type=str,
-        default="localhost",
-        help="Hostname or IP used by KV transfer (default: localhost)",
-    )
-    parser.add_argument(
-        "--prefill-kv-port",
-        type=int,
-        default=14579,
-        help="Prefill KV port (default: 14579)",
-    )
-    parser.add_argument(
-        "--decode-kv-port",
-        type=int,
-        default=14580,
-        help="Decode KV port (default: 14580)",
-    )
-
-    return parser.parse_args()
-
-
-def main():
-    """parse command line arguments"""
-    args = parse_args()
-
-    # Initialize configuration using command line parameters
-    AIOHTTP_TIMEOUT = aiohttp.ClientTimeout(total=args.timeout)
-    PREFILL_SERVICE_URL = args.prefill_url
-    DECODE_SERVICE_URL = args.decode_url
-    PORT = args.port
-
-    PREFILL_KV_ADDR = f"{args.kv_host}:{args.prefill_kv_port}"
-    DECODE_KV_ADDR = f"{args.kv_host}:{args.decode_kv_port}"
-
-    logger.info(
-        "Proxy resolved KV addresses -> prefill: %s, decode: %s",
-        PREFILL_KV_ADDR,
-        DECODE_KV_ADDR,
-    )
-
-    app = Quart(__name__)
-
-    # Attach the configuration object to the application instance so helper
-    # coroutines can read the resolved backend URLs and timeouts without using
-    # globals.
-    app.config.update(
-        {
-            "AIOHTTP_TIMEOUT": AIOHTTP_TIMEOUT,
-            "PREFILL_SERVICE_URL": PREFILL_SERVICE_URL,
-            "DECODE_SERVICE_URL": DECODE_SERVICE_URL,
-            "PREFILL_KV_ADDR": PREFILL_KV_ADDR,
-            "DECODE_KV_ADDR": DECODE_KV_ADDR,
-        }
-    )
-
-    def _normalize_base_url(url: str) -> str:
-        """Remove any trailing slash so path joins behave predictably."""
-        return url.rstrip("/")
-
-    def _get_host_port(url: str) -> str:
-        """Return the hostname:port portion for logging and KV headers."""
-        parsed = urlparse(url)
-        host = parsed.hostname or "localhost"
-        port = parsed.port
-        if port is None:
-            port = 80 if parsed.scheme == "http" else 443
-        return f"{host}:{port}"
-
-    PREFILL_BASE = _normalize_base_url(PREFILL_SERVICE_URL)
-    DECODE_BASE = _normalize_base_url(DECODE_SERVICE_URL)
-    KV_TARGET = _get_host_port(DECODE_SERVICE_URL)
-
-    def _build_headers(request_id: str) -> dict[str, str]:
-        """Construct the headers expected by vLLM's P2P disagg connector."""
-        headers: dict[str, str] = {"X-Request-Id": request_id, "X-KV-Target": KV_TARGET}
-        api_key = os.environ.get("OPENAI_API_KEY")
-        if api_key:
-            headers["Authorization"] = f"Bearer {api_key}"
-        return headers
-
-    async def _run_prefill(
-        request_path: str,
-        payload: dict,
-        headers: dict[str, str],
-        request_id: str,
-    ):
-        url = f"{PREFILL_BASE}{request_path}"
-        start_ts = time.perf_counter()
-        logger.info("[prefill] start request_id=%s url=%s", request_id, url)
-        try:
-            async with (
-                aiohttp.ClientSession(timeout=AIOHTTP_TIMEOUT) as session,
-                session.post(url=url, json=payload, headers=headers) as resp,
-            ):
-                if resp.status != 200:
-                    error_text = await resp.text()
-                    raise RuntimeError(
-                        f"Prefill backend error {resp.status}: {error_text}"
-                    )
-                await resp.read()
-                logger.info(
-                    "[prefill] done request_id=%s status=%s elapsed=%.2fs",
-                    request_id,
-                    resp.status,
-                    time.perf_counter() - start_ts,
-                )
-        except asyncio.TimeoutError as exc:
-            raise RuntimeError(f"Prefill service timeout at {url}") from exc
-        except aiohttp.ClientError as exc:
-            raise RuntimeError(f"Prefill service unavailable at {url}") from exc
-
-    async def _stream_decode(
-        request_path: str,
-        payload: dict,
-        headers: dict[str, str],
-        request_id: str,
-    ):
-        url = f"{DECODE_BASE}{request_path}"
-        # Stream tokens from the decode service once the prefill stage has
-        # materialized KV caches on the target workers.
-        logger.info("[decode] start request_id=%s url=%s", request_id, url)
-        try:
-            async with (
-                aiohttp.ClientSession(timeout=AIOHTTP_TIMEOUT) as session,
-                session.post(url=url, json=payload, headers=headers) as resp,
-            ):
-                if resp.status != 200:
-                    error_text = await resp.text()
-                    logger.error(
-                        "Decode backend error %s - %s", resp.status, error_text
-                    )
-                    err_msg = (
-                        '{"error": "Decode backend error ' + str(resp.status) + '"}'
-                    )
-                    yield err_msg.encode()
-                    return
-                logger.info(
-                    "[decode] streaming response request_id=%s status=%s",
-                    request_id,
-                    resp.status,
-                )
-                async for chunk_bytes in resp.content.iter_chunked(1024):
-                    yield chunk_bytes
-                logger.info("[decode] finished streaming request_id=%s", request_id)
-        except asyncio.TimeoutError:
-            logger.error("Decode service timeout at %s", url)
-            yield b'{"error": "Decode service timeout"}'
-        except aiohttp.ClientError as exc:
-            logger.error("Decode service error at %s: %s", url, exc)
-            yield b'{"error": "Decode service unavailable"}'
-
-    async def process_request():
-        """Process a single request through prefill and decode stages"""
-        try:
-            original_request_data = await request.get_json()
-
-            # Create prefill request (max_tokens=1)
-            prefill_request = original_request_data.copy()
-            prefill_request["max_tokens"] = 1
-            if "max_completion_tokens" in prefill_request:
-                prefill_request["max_completion_tokens"] = 1
-
-            # Execute prefill stage
-            # The request id encodes both KV socket addresses so the backend can
-            # shuttle tensors directly via NCCL once the prefill response
-            # completes.
-            request_id = (
-                f"___prefill_addr_{PREFILL_KV_ADDR}___decode_addr_"
-                f"{DECODE_KV_ADDR}_{uuid.uuid4().hex}"
-            )
-
-            headers = _build_headers(request_id)
-            await _run_prefill(request.path, prefill_request, headers, request_id)
-
-            # Execute decode stage and stream response
-            # Pass the unmodified user request so the decode phase can continue
-            # sampling with the already-populated KV cache.
-            generator = _stream_decode(
-                request.path, original_request_data, headers, request_id
-            )
-            response = await make_response(generator)
-            response.timeout = None  # Disable timeout for streaming response
-            return response
-
-        except Exception:
-            logger.exception("Error processing request")
-            return Response(
-                response=b'{"error": "Internal server error"}',
-                status=500,
-                content_type="application/json",
-            )
-
-    @app.route("/v1/completions", methods=["POST"])
-    async def handle_request():
-        """Handle incoming API requests with concurrency and rate limiting"""
-        try:
-            return await process_request()
-        except asyncio.CancelledError:
-            logger.warning("Request cancelled")
-            return Response(
-                response=b'{"error": "Request cancelled"}',
-                status=503,
-                content_type="application/json",
-            )
-
-    # Start the Quart server with host can be set to 0.0.0.0
-    app.run(port=PORT)
-
-
-if __name__ == "__main__":
-    main()

benchmarks/disagg_benchmarks/round_robin_proxy.py

@@ -1,63 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-
-import asyncio
-import itertools
-
-import aiohttp
-from aiohttp import web
-
-
-class RoundRobinProxy:
-    def __init__(self, target_ports):
-        self.target_ports = target_ports
-        self.port_cycle = itertools.cycle(self.target_ports)
-
-    async def handle_request(self, request):
-        target_port = next(self.port_cycle)
-        target_url = f"http://localhost:{target_port}{request.path_qs}"
-
-        async with aiohttp.ClientSession() as session:
-            try:
-                # Forward the request
-                async with session.request(
-                    method=request.method,
-                    url=target_url,
-                    headers=request.headers,
-                    data=request.content,
-                ) as response:
-                    # Start sending the response
-                    resp = web.StreamResponse(
-                        status=response.status, headers=response.headers
-                    )
-                    await resp.prepare(request)
-
-                    # Stream the response content
-                    async for chunk in response.content.iter_any():
-                        await resp.write(chunk)
-
-                    await resp.write_eof()
-                    return resp
-
-            except Exception as e:
-                return web.Response(text=f"Error: {str(e)}", status=500)
-
-
-async def main():
-    proxy = RoundRobinProxy([8100, 8200])
-    app = web.Application()
-    app.router.add_route("*", "/{path:.*}", proxy.handle_request)
-
-    runner = web.AppRunner(app)
-    await runner.setup()
-    site = web.TCPSite(runner, "localhost", 8000)
-    await site.start()
-
-    print("Proxy server started on http://localhost:8000")
-
-    # Keep the server running
-    await asyncio.Event().wait()
-
-
-if __name__ == "__main__":
-    asyncio.run(main())

benchmarks/disagg_benchmarks/visualize_benchmark_results.py

@@ -1,47 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-
-import json
-
-import matplotlib.pyplot as plt
-import pandas as pd
-
-if __name__ == "__main__":
-    data = []
-    for name in ["disagg_prefill", "chunked_prefill"]:
-        for qps in [2, 4, 6, 8]:
-            with open(f"results/{name}-qps-{qps}.json") as f:
-                x = json.load(f)
-                x["name"] = name
-                x["qps"] = qps
-                data.append(x)
-
-    df = pd.DataFrame.from_dict(data)
-    dis_df = df[df["name"] == "disagg_prefill"]
-    chu_df = df[df["name"] == "chunked_prefill"]
-
-    plt.style.use("bmh")
-    plt.rcParams["font.size"] = 20
-
-    for key in [
-        "mean_ttft_ms",
-        "median_ttft_ms",
-        "p99_ttft_ms",
-        "mean_itl_ms",
-        "median_itl_ms",
-        "p99_itl_ms",
-    ]:
-        fig, ax = plt.subplots(figsize=(11, 7))
-        plt.plot(
-            dis_df["qps"], dis_df[key], label="disagg_prefill", marker="o", linewidth=4
-        )
-        plt.plot(
-            chu_df["qps"], chu_df[key], label="chunked_prefill", marker="o", linewidth=4
-        )
-        ax.legend()
-
-        ax.set_xlabel("QPS")
-        ax.set_ylabel(key)
-        ax.set_ylim(bottom=0)
-        fig.savefig(f"results/{key}.png")
-        plt.close(fig)

benchmarks/kernels/benchmark_flydsl_moe_w4a16.py

@@ -0,0 +1,277 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# Copyright (c) 2025 FlyDSL Project Contributors
+
+import json
+import os
+
+import torch
+from aiter.test_common import run_perftest
+
+from vllm.model_executor.layers.fused_moe import fused_experts
+from vllm.model_executor.layers.fused_moe.activation import MoEActivation
+from vllm.model_executor.layers.fused_moe.config import (
+    int4_w4a16_moe_quant_config,
+)
+from vllm.model_executor.layers.fused_moe.fused_flydsl_moe import fused_flydsl_moe
+from vllm.model_executor.layers.quantization.compressed_tensors.compressed_tensors_moe import (  # noqa: E501
+    compressed_tensors_moe_w4a16_flydsl,
+)
+from vllm.platforms import current_platform
+
+RoutingBuffers = tuple[
+    torch.Tensor,  # sorted_token_ids
+    torch.Tensor,  # sorted_weights
+    torch.Tensor,  # sorted_expert_ids
+    torch.Tensor,  # num_valid_ids (shape [1], i32)
+    int,  # sorted_size
+    int,  # blocks
+]
+
+MODEL_PARAMS_TO_TUNE = [
+    # (num_experts, inter_dim, hidden_size, topk)
+    (384, 256, 7168, 8),  # Kimi K2.5 TP=8
+    (384, 512, 7168, 8),  # Kimi K2.5 TP=4
+]
+
+NUM_TOKENS_TO_TUNE = [
+    1,
+    2,
+    4,
+    8,
+    16,
+    24,
+    32,
+    48,
+    64,
+    128,
+    256,
+    512,
+    1024,
+    2048,
+    4096,
+    8192,
+]
+
+TILE_M_SEARCH_SPACE = [16, 32, 64, 128, 256]
+TILE_N_SEARCH_SPACE = [16, 32, 64, 128, 256]
+TILE_K_SEARCH_SPACE = [16, 32, 64, 128, 256, 512]
+TILE_N2_SEARCH_SPACE = [16, 32, 64, 128, 256]
+TILE_K2_SEARCH_SPACE = [16, 32, 64, 128, 256, 512]
+
+TILE_CONFIGS = []
+for tile_m in TILE_M_SEARCH_SPACE:
+    for tile_n in TILE_N_SEARCH_SPACE:
+        for tile_k in TILE_K_SEARCH_SPACE:
+            for tile_n2 in TILE_N2_SEARCH_SPACE:
+                for tile_k2 in TILE_K2_SEARCH_SPACE:
+                    TILE_CONFIGS.append(
+                        {
+                            "tile_m": tile_m,
+                            "tile_n": tile_n,
+                            "tile_k": tile_k,
+                            "tile_n2": tile_n2,
+                            "tile_k2": tile_k2,
+                        }
+                    )
+
+
+def tune_flydsl_moe_w4a16(
+    device: str = "cuda", num_iters: int = 100, num_warmup: int = 10
+):
+    packed_factor = 8
+    w13_num_shards = 2
+    params_dtype = torch.bfloat16
+    group_size = 32
+    scale_factor = 0.01
+
+    for model_params in MODEL_PARAMS_TO_TUNE:
+        num_experts = model_params[0]
+        inter_dim = model_params[1]
+        hidden_size = model_params[2]
+        topk = model_params[3]
+        print(
+            f"\nTuning: num_experts={num_experts}, inter_dim={inter_dim}, "
+            f"hidden_size={hidden_size}, topk={topk}...\n"
+        )
+
+        w2_scales_size = inter_dim
+        num_groups_w2 = w2_scales_size // group_size
+        num_groups_w13 = hidden_size // group_size
+
+        w13_weight = torch.randint(
+            0,
+            255,
+            (num_experts, hidden_size // packed_factor, w13_num_shards * inter_dim),
+            dtype=torch.int32,
+            device=device,
+        )
+
+        w2_weight = torch.randint(
+            0,
+            255,
+            (num_experts, inter_dim // packed_factor, hidden_size),
+            dtype=torch.int32,
+            device=device,
+        )
+        w13_scale = scale_factor * torch.randn(
+            num_experts,
+            num_groups_w13,
+            w13_num_shards * inter_dim,
+            dtype=params_dtype,
+            device=device,
+        )
+        w2_scale = scale_factor * torch.randn(
+            num_experts, num_groups_w2, hidden_size, dtype=params_dtype, device=device
+        )
+
+        w13 = w13_weight
+        w13 = compressed_tensors_moe_w4a16_flydsl._gptq_int32_to_flydsl_packed(w13)
+        w13 = w13.view(-1).contiguous()
+
+        w2 = w2_weight
+        w2 = compressed_tensors_moe_w4a16_flydsl._gptq_int32_to_flydsl_packed(w2)
+        w2 = w2.view(-1).contiguous()
+
+        w13_scale_flydsl = w13_scale
+        w2_scale_flydsl = w2_scale
+
+        if group_size > 0 and w13_scale.dim() == 3 and w13_scale.shape[1] > 1:
+            E, G, N = w13_scale.shape
+            w13_scale_flydsl = (
+                w13_scale_flydsl.view(E, G // 2, 2, N)
+                .permute(0, 1, 3, 2)
+                .contiguous()
+                .view(-1)
+                .contiguous()
+            )
+        elif w13_scale.dim() == 3 and w13_scale.shape[1] == 1:
+            w13_scale_flydsl = w13_scale_flydsl.squeeze(1)
+
+        if group_size > 0 and w2_scale.dim() == 3 and w2_scale.shape[1] > 1:
+            E, G, N = w2_scale.shape
+            w2_scale_flydsl = (
+                w2_scale_flydsl.view(E, G // 2, 2, N)
+                .permute(0, 1, 3, 2)
+                .contiguous()
+                .view(-1)
+                .contiguous()
+            )
+        elif w2_scale.dim() == 3 and w2_scale.shape[1] == 1:
+            w2_scale_flydsl = w2_scale_flydsl.squeeze(1)
+
+        w13_scale_flydsl = w13_scale_flydsl.contiguous()
+        w2_scale_flydsl = w2_scale_flydsl.contiguous()
+
+        w13.is_shuffled = True
+        w2.is_shuffled = True
+
+        w13_weight_scale = w13_scale.transpose(1, 2).contiguous()
+        w2_weight_scale = w2_scale.transpose(1, 2).contiguous()
+        w13_weight_packed = w13_weight.transpose(1, 2).contiguous().view(torch.uint8)
+        w2_weight_packed = w2_weight.transpose(1, 2).contiguous().view(torch.uint8)
+
+        moe_quant_config = int4_w4a16_moe_quant_config(
+            w1_scale=w13_weight_scale,
+            w2_scale=w2_weight_scale,
+            w1_zp=None,
+            w2_zp=None,
+            block_shape=[0, group_size],
+        )
+
+        tuned_config = {}
+
+        for num_tokens in NUM_TOKENS_TO_TUNE:
+            score = torch.rand(
+                (num_tokens, num_experts), device=device, dtype=torch.float32
+            )
+            topk_vals, topk_ids = torch.topk(score, k=topk, dim=1)
+            topk_weights = torch.softmax(topk_vals, dim=1).to(torch.float32)
+            x = torch.randn(
+                (num_tokens, hidden_size), dtype=torch.bfloat16, device=device
+            )
+            us_best = float("inf")
+            for tile_config in TILE_CONFIGS:
+                try:
+                    tile_m = tile_config["tile_m"]
+                    tile_n = tile_config["tile_n"]
+                    tile_k = tile_config["tile_k"]
+                    tile_n2 = tile_config["tile_n2"]
+                    tile_k2 = tile_config["tile_k2"]
+
+                    model_dim = x.shape[1]
+                    assert model_dim % 64 == 0
+                    assert model_dim % tile_k == 0
+                    assert inter_dim % tile_n == 0
+                    assert model_dim % tile_n2 == 0
+                    assert inter_dim % tile_k2 == 0
+                    assert ((tile_m * tile_k2) % 256) == 0
+                    bytes_per_thread_x = (tile_m * tile_k2) // 256
+                    assert (bytes_per_thread_x % 4) == 0
+
+                    out, _us = run_perftest(
+                        fused_flydsl_moe,
+                        x,
+                        w13,
+                        w2,
+                        num_experts,
+                        inter_dim,
+                        topk_weights,
+                        topk_ids,
+                        num_iters=num_iters,
+                        num_warmup=num_warmup,
+                        w1_scale=w13_scale_flydsl,
+                        w2_scale=w2_scale_flydsl,
+                        topk=topk_weights.shape[-1],
+                        group_size=group_size,
+                        doweight_stage1=False,
+                        scale_is_bf16=True,
+                        config=tile_config,
+                    )
+                    torch.accelerator.synchronize()
+                except Exception:
+                    torch.accelerator.synchronize()
+                    continue
+                else:
+                    us = _us.item()
+                    if us < us_best:
+                        out_ref = fused_experts(
+                            x,
+                            w13_weight_packed,
+                            w2_weight_packed,
+                            topk_weights=topk_weights,
+                            topk_ids=topk_ids,
+                            activation=MoEActivation.SILU,
+                            apply_router_weight_on_input=False,
+                            global_num_experts=num_experts,
+                            expert_map=None,
+                            quant_config=moe_quant_config,
+                        )
+                        try:
+                            assert torch.allclose(out, out_ref, atol=0.5, rtol=0.1)
+                        except Exception:
+                            continue
+                        else:
+                            print(
+                                f"For [num_tokens={num_tokens}, num_experts={num_experts}, "  # noqa: E501
+                                f"inter_dim={inter_dim}] found new best "  # noqa: E501
+                                f"config={tile_config}, us={us:0.3f}"
+                            )
+                            us_best = us
+                            tuned_config[str(num_tokens)] = tile_config
+            device_name = current_platform.get_device_name().replace(" ", "_")
+            tuned_config_file_name = (
+                f"E={num_experts},N={inter_dim},device_name={device_name},"
+                f"dtype=int4_w4a16,backend=flydsl.json"
+            )
+            tuner_dir_path = os.path.dirname(os.path.realpath(__file__))
+            store_path = os.path.join(tuner_dir_path, tuned_config_file_name)
+            with open(store_path, "w") as f:
+                json.dump(tuned_config, f, indent=4)
+            print(
+                f"\nTuned config for num_tokens={num_tokens} was stored at {store_path}\n"  # noqa: E501
+            )
+
+
+if __name__ == "__main__":
+    tune_flydsl_moe_w4a16(device="cuda")

benchmarks/kernels/benchmark_moe.py

@@ -250,7 +250,7 @@ def benchmark_config(
                     num_experts=num_experts,
                     experts_per_token=topk,
                     hidden_dim=hidden_size,
-                    intermediate_size_per_partition=shard_intermediate_size,
+                    intermediate_size=shard_intermediate_size,
                     num_local_experts=num_experts,
                     num_logical_experts=num_experts,
                     activation=MoEActivation.SILU,
@@ -271,7 +271,6 @@ def benchmark_config(
                     moe_config=moe_config,
                     quant_config=quant_config,
                 ),
-                inplace=not disable_inplace(),
             )
 
         with override_config(config):
@@ -279,7 +278,6 @@ def benchmark_config(
                 x, input_gating, topk, renormalize=not use_deep_gemm
             )
 
-            inplace = not disable_inplace()
             if use_deep_gemm:
                 return deep_gemm_experts.apply(
                     x,
@@ -298,7 +296,6 @@ def benchmark_config(
                 w2,
                 topk_weights,
                 topk_ids,
-                inplace=inplace,
                 quant_config=quant_config,
             )
 
@@ -795,6 +792,12 @@ def get_model_params(config):
         topk = text_config.num_experts_per_tok
         intermediate_size = text_config.moe_intermediate_size
         hidden_size = text_config.hidden_size
+    elif architecture == "DiffusionGemmaForBlockDiffusion":
+        text_config = config.get_text_config()
+        E = text_config.num_experts
+        topk = text_config.top_k_experts
+        intermediate_size = text_config.moe_intermediate_size
+        hidden_size = text_config.hidden_size
     elif architecture == "HunYuanMoEV1ForCausalLM":
         E = config.num_experts
         topk = config.moe_topk[0]

benchmarks/kv_cache_watermark.sh

@@ -0,0 +1,248 @@
+#!/bin/bash
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+#
+# Reproducible demonstration of the KV cache watermark (`--watermark`) for
+# reducing preemption thrashing.
+#
+# The watermark is the fraction of total KV cache blocks the scheduler keeps
+# free when admitting a waiting/preempted request into the running queue.
+#
+# Why this workload triggers thrashing:
+#   Requests are admitted based on the KV cache they need *at admission time*.
+#   With `--scheduler-reserve-full-isl` (default) the input length is reserved up
+#   front, but the *output* length is unknown and unreserved. A decode-heavy
+#   workload (output >> input) at high concurrency therefore over-admits while
+#   requests are short, then runs out of KV cache as they all grow during decode
+#   -> the scheduler preempts (recompute) recently-admitted requests, re-prefills
+#   them later, and repeats. The watermark keeps a block of KV cache free so
+#   running requests can grow into it instead of triggering this churn.
+#
+# This script launches `vllm serve` under a deliberately KV-constrained config
+# and a decode-heavy workload, sweeping the watermark across several values, and
+# reports the preemption count (scraped from /metrics), throughput, and latency
+# percentiles for each. It then plots the results.
+#
+# Default workload: concurrency 200, input ~300 tokens, output ~4000 tokens
+# (+/- 20% variance), sized to run each config for ~5 minutes.
+#
+# Usage:
+#   benchmarks/kv_cache_watermark.sh
+#   MODEL=Qwen/Qwen2.5-14B-Instruct TP=2 benchmarks/kv_cache_watermark.sh
+#
+# Run inside the vLLM virtualenv (so `vllm` and `python` resolve to it).
+set -euo pipefail
+
+# ---- Config (override via environment) -------------------------------------
+MODEL=${MODEL:-Qwen/Qwen2.5-7B-Instruct}
+TP=${TP:-1}
+PORT=${PORT:-8000}
+URL="http://127.0.0.1:${PORT}"
+# Constrain the KV cache to a *near-critical* size: large enough that the engine
+# can run stably, but small enough that greedy over-admission tips it into
+# preemption thrashing. (Independent of GPU size, so the demo is reproducible.)
+# At the default workload this fits ~1.5x the mean concurrent KV demand.
+KV_CACHE_MEMORY_GB=${KV_CACHE_MEMORY_GB:-16}
+MAX_MODEL_LEN=${MAX_MODEL_LEN:-8192}
+MAX_NUM_SEQS=${MAX_NUM_SEQS:-256}
+# Optional weight loader (e.g. fastsafetensors on the GCP cluster).
+LOAD_FORMAT=${LOAD_FORMAT:-auto}
+# Decode-heavy workload: moderate input, long output, with length variance. The
+# long output means preempted requests have generated a lot before eviction, so
+# resuming them re-prefills a long sequence (high recomputation cost).
+INPUT_LEN=${INPUT_LEN:-1000}
+OUTPUT_LEN=${OUTPUT_LEN:-5000}
+RANGE_RATIO=${RANGE_RATIO:-0.2}
+CONCURRENCY=${CONCURRENCY:-128}
+# Enough prompts to keep each config saturated for ~5+ minutes.
+NUM_PROMPTS=${NUM_PROMPTS:-450}
+OUTDIR=${OUTDIR:-./watermark_bench_results}
+# Watermark fractions compared. "label value" per line; value=0 disables it.
+CONFIGS=${CONFIGS:-"off    0
+w0.02  0.02
+w0.05  0.05
+w0.10  0.10
+w0.15  0.15"}
+
+KV_CACHE_MEMORY_BYTES=$((KV_CACHE_MEMORY_GB * 1024 * 1024 * 1024))
+mkdir -p "$OUTDIR"
+
+SERVER_PID=""
+cleanup() { [[ -n "$SERVER_PID" ]] && kill "$SERVER_PID" 2>/dev/null || true; }
+trap cleanup EXIT
+
+scrape_preemptions() {
+  # Sum the vllm:num_preemptions_total counter across engines.
+  python - "${URL}/metrics" <<'PY'
+import sys, urllib.request
+total = 0.0
+try:
+    body = urllib.request.urlopen(sys.argv[1], timeout=10).read().decode("utf-8", "replace")
+    for line in body.splitlines():
+        if line.startswith("vllm:num_preemptions_total"):
+            total += float(line.rsplit(" ", 1)[-1])
+except Exception as e:  # noqa: BLE001
+    print(f"scrape error: {e}", file=sys.stderr)
+print(int(total))
+PY
+}
+
+wait_for_server() {
+  for _ in $(seq 1 300); do
+    if curl -s "${URL}/health" >/dev/null 2>&1; then return 0; fi
+    if ! kill -0 "$SERVER_PID" 2>/dev/null; then
+      echo "ERROR: server process exited during startup" >&2; return 1
+    fi
+    sleep 5
+  done
+  echo "ERROR: server did not become ready" >&2; return 1
+}
+
+run_one() {
+  local label=$1 watermark=$2
+  echo
+  echo "==================== watermark: ${label} (${watermark}) ===================="
+  vllm serve "$MODEL" \
+    --tensor-parallel-size "$TP" \
+    --load-format "$LOAD_FORMAT" \
+    --kv-cache-memory-bytes "$KV_CACHE_MEMORY_BYTES" \
+    --max-model-len "$MAX_MODEL_LEN" \
+    --max-num-seqs "$MAX_NUM_SEQS" \
+    --no-enable-prefix-caching \
+    --watermark "$watermark" \
+    --port "$PORT" >"${OUTDIR}/serve_${label}.log" 2>&1 &
+  SERVER_PID=$!
+  wait_for_server
+  sleep 5
+
+  local pre post
+  pre=$(scrape_preemptions)
+  vllm bench serve \
+    --backend vllm \
+    --base-url "$URL" \
+    --model "$MODEL" \
+    --dataset-name random \
+    --random-input-len "$INPUT_LEN" \
+    --random-output-len "$OUTPUT_LEN" \
+    --random-range-ratio "$RANGE_RATIO" \
+    --ignore-eos \
+    --num-prompts "$NUM_PROMPTS" \
+    --max-concurrency "$CONCURRENCY" \
+    --percentile-metrics "ttft,tpot,itl,e2el" \
+    --metric-percentiles "50,90,99" \
+    --save-result \
+    --result-dir "$OUTDIR" \
+    --result-filename "bench_${label}.json"
+  post=$(scrape_preemptions)
+  echo "${label} ${watermark} $((post - pre))" >>"${OUTDIR}/preemptions.txt"
+
+  kill "$SERVER_PID" 2>/dev/null || true
+  for _ in $(seq 1 60); do curl -s "${URL}/health" >/dev/null 2>&1 || break; sleep 2; done
+  SERVER_PID=""
+  sleep 10
+}
+
+: >"${OUTDIR}/preemptions.txt"
+while read -r label watermark; do
+  [[ -z "${label:-}" ]] && continue
+  run_one "$label" "$watermark"
+done <<<"$CONFIGS"
+
+echo
+echo "==================== summary ===================="
+python - "$OUTDIR" <<'PY'
+import json, os, sys
+outdir = sys.argv[1]
+pre = {}
+order = []
+for line in open(os.path.join(outdir, "preemptions.txt")):
+    label, watermark, n = line.split()
+    pre[label] = (float(watermark), int(n))
+    order.append(label)
+
+def g(d, *names):
+    for n in names:
+        if d.get(n) is not None:
+            return d[n]
+    return float("nan")
+
+cols = ["watermark", "frac", "preempt", "out_tok/s", "req/s",
+        "TTFT_p50", "TTFT_p99", "ITL_p99", "E2EL_p50"]
+print("  ".join(f"{c:>10}" for c in cols))
+rows = []
+for label in order:
+    watermark, n = pre[label]
+    d = json.load(open(os.path.join(outdir, f"bench_{label}.json")))
+    rows.append(dict(
+        label=label, watermark=watermark, preempt=n,
+        out_tok_s=g(d, "output_throughput"),
+        req_s=g(d, "request_throughput"),
+        ttft_p50=g(d, "p50_ttft_ms", "median_ttft_ms"),
+        ttft_p99=g(d, "p99_ttft_ms"),
+        itl_p99=g(d, "p99_itl_ms"),
+        e2el_p50=g(d, "p50_e2el_ms", "median_e2el_ms"),
+    ))
+    print("  ".join(f"{str(v):>10}" for v in [
+        label, watermark, n,
+        f"{rows[-1]['out_tok_s']:.0f}",
+        f"{rows[-1]['req_s']:.3f}",
+        f"{rows[-1]['ttft_p50']/1000:.2f}",
+        f"{rows[-1]['ttft_p99']/1000:.2f}",
+        f"{rows[-1]['itl_p99']:.2f}",
+        f"{rows[-1]['e2el_p50']/1000:.1f}",
+    ]))
+print("\n(TTFT/E2EL in seconds; ITL in ms. Lower preempt is better.)")
+
+# ---- Plot -------------------------------------------------------------------
+try:
+    import matplotlib
+    matplotlib.use("Agg")
+    import matplotlib.pyplot as plt
+except Exception as e:  # noqa: BLE001
+    print(f"\n(skip plot: matplotlib unavailable: {e})")
+    sys.exit(0)
+
+x = [r["watermark"] for r in rows]
+xt = [f"{r['watermark']:g}\n({r['label']})" for r in rows]
+idx = list(range(len(rows)))
+
+fig, axes = plt.subplots(2, 2, figsize=(12, 8))
+fig.suptitle(
+    f"KV cache watermark sweep — {os.path.basename(os.path.abspath(outdir))}",
+    fontsize=12,
+)
+
+ax = axes[0][0]
+ax.bar(idx, [r["preempt"] for r in rows], color="tab:red")
+ax.set_title("Preemptions (lower is better)")
+ax.set_ylabel("preemptions")
+ax.set_xticks(idx); ax.set_xticklabels(xt)
+
+ax = axes[0][1]
+ax.plot(idx, [r["out_tok_s"] for r in rows], "o-", color="tab:green")
+ax.set_title("Output throughput (higher is better)")
+ax.set_ylabel("tokens/s")
+ax.set_xticks(idx); ax.set_xticklabels(xt)
+
+ax = axes[1][0]
+ax.plot(idx, [r["itl_p99"] for r in rows], "o-", color="tab:blue")
+ax.set_title("Inter-token latency p99 (lower is better)")
+ax.set_ylabel("ITL p99 (ms)")
+ax.set_xlabel("watermark fraction")
+ax.set_xticks(idx); ax.set_xticklabels(xt)
+
+ax = axes[1][1]
+ax.plot(idx, [r["ttft_p50"] / 1000 for r in rows], "o-", label="TTFT p50")
+ax.plot(idx, [r["ttft_p99"] / 1000 for r in rows], "o-", label="TTFT p99")
+ax.plot(idx, [r["e2el_p50"] / 1000 for r in rows], "o-", label="E2EL p50")
+ax.set_title("Latency (lower is better)")
+ax.set_ylabel("seconds")
+ax.set_xlabel("watermark fraction")
+ax.set_xticks(idx); ax.set_xticklabels(xt)
+ax.legend()
+
+fig.tight_layout(rect=(0, 0, 1, 0.95))
+out_png = os.path.join(outdir, "watermark_results.png")
+fig.savefig(out_png, dpi=120)
+print(f"\nWrote plot: {out_png}")
+PY

benchmarks/multi_turn/benchmark_serving_multi_turn.py

@@ -65,6 +65,32 @@ class RequestArgs(NamedTuple):
     limit_min_tokens: int  # Use negative value for no limit
     limit_max_tokens: int  # Use negative value for no limit
     timeout_sec: int
+    send_conversation_id: bool
+    headers: dict[str, str]
+
+
+def parse_custom_header(header: str) -> tuple[str, str]:
+    separators = (":", "=")
+    for separator in separators:
+        if separator in header:
+            key, value = header.split(separator, 1)
+            key = key.strip()
+            value = value.strip()
+            if key:
+                return key, value
+            break
+    raise argparse.ArgumentTypeError(
+        "Headers must be provided as 'Header-Name: value' or 'Header-Name=value'"
+    )
+
+
+def build_request_headers(
+    api_key: str | None, custom_headers: list[tuple[str, str]] | None
+) -> dict[str, str]:
+    headers = dict(custom_headers or [])
+    if api_key:
+        headers["Authorization"] = f"Bearer {api_key}"
+    return headers
 
 
 class BenchmarkArgs(NamedTuple):
@@ -218,12 +244,11 @@ async def send_request(
     max_tokens: int | None = None,
     timeout_sec: int = 120,
     conversation_id: str | None = None,
+    headers: dict[str, str] | None = None,
 ) -> ServerResponse:
     payload = {
         "model": model,
         "messages": messages,
-        "seed": 0,
-        "temperature": 0.0,
     }
 
     if conversation_id is not None:
@@ -233,13 +258,17 @@ async def send_request(
         payload["stream"] = True
         payload["stream_options"] = {"include_usage": False}
 
-    if min_tokens is not None:
-        payload["min_tokens"] = min_tokens
+    # if min_tokens is not None:
+    #     payload["min_tokens"] = min_tokens
 
     if max_tokens is not None:
         payload["max_tokens"] = max_tokens
 
-    headers = {"Content-Type": "application/json"}
+    request_headers = {"Content-Type": "application/json"}
+    if conversation_id is not None:
+        request_headers["X-Session-ID"] = str(conversation_id)
+    if headers is not None:
+        request_headers.update(headers)
 
     # Calculate the timeout for the request
     if max_tokens is not None:
@@ -265,7 +294,7 @@ async def send_request(
     most_recent_timestamp: int = start_time
 
     async with session.post(
-        url=chat_url, json=payload, headers=headers, timeout=timeout
+        url=chat_url, json=payload, headers=request_headers, timeout=timeout
     ) as response:
         http_status = HTTPStatus(response.status)
         if http_status == HTTPStatus.OK:
@@ -317,6 +346,8 @@ async def send_request(
             latency = time.perf_counter_ns() - start_time
 
     if ttft is None:
+        if stream:
+            valid_response = False
         # The response was a single chunk
         ttft = latency
 
@@ -423,7 +454,8 @@ async def send_turn(
         min_tokens,
         max_tokens,
         req_args.timeout_sec,
-        conversation_id=conv_id,
+        conversation_id=conv_id if req_args.send_conversation_id else None,
+        headers=req_args.headers,
     )
 
     if response.valid is False:
@@ -872,6 +904,7 @@ def get_client_config(
     # Arguments for API requests
     chat_url = f"{args.url}/v1/chat/completions"
     model_name = args.served_model_name if args.served_model_name else args.model
+    headers = build_request_headers(args.api_key, args.header)
 
     req_args = RequestArgs(
         chat_url=chat_url,
@@ -880,6 +913,8 @@ def get_client_config(
         limit_min_tokens=args.limit_min_tokens,
         limit_max_tokens=args.limit_max_tokens,
         timeout_sec=args.request_timeout_sec,
+        send_conversation_id=args.send_conversation_id,
+        headers=headers,
     )
 
     return client_args, req_args
@@ -1245,19 +1280,19 @@ def process_statistics(
         )
 
 
-async def get_server_info(url: str) -> None:
+async def get_server_info(url: str, headers: dict[str, str] | None = None) -> None:
     logger.info(f"{Color.BLUE}Collecting information from server: {url}{Color.RESET}")
     async with aiohttp.ClientSession() as session:
         # Get server version (not mandatory, "version" endpoint may not exist)
         url_version = f"{url}/version"
-        async with session.get(url_version) as response:
+        async with session.get(url_version, headers=headers) as response:
             if HTTPStatus(response.status) == HTTPStatus.OK:
                 text = await response.text()
                 logger.info(f"{Color.BLUE}Server version: {text}{Color.RESET}")
 
         # Get available models
         url_models = f"{url}/v1/models"
-        async with session.get(url_models) as response:
+        async with session.get(url_models, headers=headers) as response:
             if HTTPStatus(response.status) == HTTPStatus.OK:
                 text = await response.text()
                 logger.info(f"{Color.BLUE}Models:{Color.RESET}")
@@ -1323,6 +1358,22 @@ async def main() -> None:
         help="Base URL for the LLM API server",
     )
 
+    parser.add_argument(
+        "--api-key",
+        type=str,
+        default=None,
+        help="API key to send as an Authorization bearer token",
+    )
+    parser.add_argument(
+        "--header",
+        action="append",
+        type=parse_custom_header,
+        default=None,
+        metavar="KEY=VALUE",
+        help="Custom request header. Can be specified multiple times. "
+        "Accepts 'Header-Name: value' or 'Header-Name=value'.",
+    )
+
     parser.add_argument(
         "-p",
         "--num-clients",
@@ -1437,6 +1488,22 @@ async def main() -> None:
         help="Disable stream/streaming mode (set 'stream' to False in the API request)",
     )
 
+    parser.add_argument(
+        "--send-conversation-id",
+        default=False,
+        action="store_true",
+        help=(
+            "Inject a `conversation_id` field into each Chat Completions "
+            "payload. This is a non-standard OpenAI extension consumed by "
+            "vLLM's disaggregated multi-turn proxy "
+            "(examples/disaggregated/disaggregated_serving/"
+            "disagg_proxy_multiturn.py) to key cross-turn KV cache reuse. "
+            "Leave disabled (default) when targeting strict "
+            "OpenAI-compatible endpoints; enable when benchmarking the "
+            "disaggregated proxy."
+        ),
+    )
+
     parser.add_argument(
         "-e",
         "--excel-output",
@@ -1525,7 +1592,8 @@ async def main() -> None:
         args.model, trust_remote_code=args.trust_remote_code
     )
 
-    await get_server_info(args.url)
+    headers = build_request_headers(args.api_key, args.header)
+    await get_server_info(args.url, headers=headers)
 
     # Load the input file (either conversations of configuration file)
     logger.info(f"Reading input file: {args.input_file}")

build_rust.sh

@@ -1,5 +1,5 @@
 #!/bin/bash
-# Build the vllm-rs Rust frontend binary and install it into the vllm package.
+# Build vLLM Rust artifacts and install them into the vllm package.
 # Usage: ./build_rust.sh [--debug]
 #
 # By default builds in release mode. Pass --debug for faster compile times
@@ -8,8 +8,6 @@
 set -euo pipefail
 
 REPO_ROOT="$(cd "$(dirname "$0")" && pwd)"
-RUST_DIR="$REPO_ROOT/rust"
-TARGET_PATH="${VLLM_RS_TARGET_PATH:-$REPO_ROOT/vllm/vllm-rs}"
 
 # Read the required toolchain from rust-toolchain.toml.
 TOOLCHAIN=$(grep '^channel' "$REPO_ROOT/rust-toolchain.toml" | sed 's/.*= *"\(.*\)"/\1/')
@@ -27,18 +25,9 @@ if ! rustup run "$TOOLCHAIN" rustc --version &>/dev/null; then
 fi
 
 if [[ "${1:-}" == "--debug" ]]; then
-    PROFILE_ARGS=()
-    PROFILE_DIR="debug"
+    PROFILE_ARG="--debug"
 else
-    PROFILE_ARGS=(--release)
-    PROFILE_DIR="release"
+    PROFILE_ARG="--release"
 fi
 
-cargo +"$TOOLCHAIN" build "${PROFILE_ARGS[@]}" \
-    --manifest-path "$RUST_DIR/Cargo.toml" \
-    --bin vllm-rs \
-    --features native-tls-vendored
-
-mkdir -p "$(dirname "$TARGET_PATH")"
-cp "$RUST_DIR/target/$PROFILE_DIR/vllm-rs" "$TARGET_PATH"
-echo "Installed vllm-rs to $TARGET_PATH"
+python3 "$REPO_ROOT/tools/build_rust.py" "$PROFILE_ARG"

cmake/cpu_extension.cmake

@@ -166,6 +166,10 @@ elseif (S390_FOUND)
         "-mtune=native")
 elseif (CMAKE_SYSTEM_PROCESSOR MATCHES "riscv64")
     message(STATUS "RISC-V detected")
+    if(DEFINED VLLM_RVV_VLEN AND NOT VLLM_RVV_VLEN GREATER 0)
+        message(FATAL_ERROR
+            "VLLM_RVV_VLEN must be a positive integer; got '${VLLM_RVV_VLEN}'")
+    endif()
     # VLLM_RVV_VLEN selects the target VLEN. Auto-detected from /proc/cpuinfo
     # by default; override with -DVLLM_RVV_VLEN=128 or -DVLLM_RVV_VLEN=256.
     if(NOT DEFINED VLLM_RVV_VLEN)
@@ -189,8 +193,7 @@ elseif (CMAKE_SYSTEM_PROCESSOR MATCHES "riscv64")
                 "RISC-V RVV is available but VLEN could not be auto-detected. "
                 "Please specify VLEN explicitly:\n"
                 "  -DVLLM_RVV_VLEN=128   (for VLEN=128 hardware)\n"
-                "  -DVLLM_RVV_VLEN=256   (for VLEN=256 hardware, e.g. Spacemit X100)\n"
-                "  -DVLLM_RVV_VLEN=0     (force scalar, no RVV)")
+                "  -DVLLM_RVV_VLEN=256   (for VLEN=256 hardware, e.g. Spacemit X100)")
         endif()
     endif()
     if(VLLM_RVV_VLEN AND VLLM_RVV_VLEN GREATER 0)
@@ -219,7 +222,7 @@ endif()
 
 
 # Build oneDNN for GEMM kernels
-if (ENABLE_X86_ISA OR (ASIMD_FOUND AND NOT APPLE_SILICON_FOUND) OR POWER9_FOUND OR POWER10_FOUND OR POWER11_FOUND)
+if (ENABLE_X86_ISA OR (ASIMD_FOUND AND NOT APPLE_SILICON_FOUND) OR POWER9_FOUND OR POWER10_FOUND OR POWER11_FOUND OR RVV_FP16_FOUND OR RVV_BF16_FOUND)
     # Fetch and build Arm Compute Library (ACL) as oneDNN's backend for AArch64
     # TODO [fadara01]: remove this once ACL can be fetched and built automatically as a dependency of oneDNN
     set(ONEDNN_AARCH64_USE_ACL OFF CACHE BOOL "")
@@ -435,6 +438,12 @@ if(USE_ONEDNN)
         ${VLLM_EXT_SRC})
 endif()
 
+if (CMAKE_SYSTEM_PROCESSOR MATCHES "riscv64")
+    set(VLLM_EXT_SRC
+        "csrc/cpu/sgl-kernels/gemm_int4.cpp"
+        ${VLLM_EXT_SRC})
+endif()
+
 if (ENABLE_X86_ISA)
     set(VLLM_EXT_SRC_SGL
         "csrc/cpu/sgl-kernels/conv.cpp"

cmake/external_projects/fmha_sm100.cmake

@@ -0,0 +1,48 @@
+include(FetchContent)
+
+# If FMHA_SM100_SRC_DIR is set, fmha_sm100 is installed from that directory
+# instead of downloading. This is useful for local MSA development.
+if(DEFINED ENV{FMHA_SM100_SRC_DIR})
+  set(FMHA_SM100_SRC_DIR $ENV{FMHA_SM100_SRC_DIR})
+endif()
+
+if(FMHA_SM100_SRC_DIR)
+  FetchContent_Declare(
+    fmha_sm100
+    SOURCE_DIR ${FMHA_SM100_SRC_DIR}
+    CONFIGURE_COMMAND ""
+    BUILD_COMMAND ""
+  )
+else()
+  FetchContent_Declare(
+    fmha_sm100
+    GIT_REPOSITORY https://github.com/vllm-project/MSA.git
+    GIT_TAG 544eee5e09ae2dfa774d5b06739013f9b7402c57
+    GIT_PROGRESS TRUE
+    CONFIGURE_COMMAND ""
+    BUILD_COMMAND ""
+  )
+endif()
+
+FetchContent_GetProperties(fmha_sm100)
+if(NOT fmha_sm100_POPULATED)
+  FetchContent_Populate(fmha_sm100)
+endif()
+message(STATUS "fmha_sm100 is available at ${fmha_sm100_SOURCE_DIR}")
+
+add_custom_target(fmha_sm100)
+
+set(FMHA_SM100_PY_ROOT "${fmha_sm100_SOURCE_DIR}/python/fmha_sm100")
+
+install(FILES
+  "${FMHA_SM100_PY_ROOT}/__init__.py"
+  "${FMHA_SM100_PY_ROOT}/sparse.py"
+  DESTINATION vllm/third_party/fmha_sm100
+  COMPONENT fmha_sm100)
+
+install(DIRECTORY "${FMHA_SM100_PY_ROOT}/cute/"
+  DESTINATION vllm/third_party/fmha_sm100/cute
+  COMPONENT fmha_sm100
+  PATTERN "__pycache__" EXCLUDE
+  PATTERN "*.pyc" EXCLUDE
+  PATTERN ".git*" EXCLUDE)

cmake/external_projects/qutlass.cmake

@@ -32,21 +32,33 @@ endif()
 message(STATUS "[QUTLASS] QuTLASS is available at ${qutlass_SOURCE_DIR}")
 
 if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
-  cuda_archs_loose_intersection(QUTLASS_ARCHS "10.0f;12.0f" "${CUDA_ARCHS}")
+  cuda_archs_loose_intersection(QUTLASS_SM120_ARCHS "12.0f" "${CUDA_ARCHS}")
+  cuda_archs_loose_intersection(QUTLASS_SM100_ARCHS "10.0f" "${CUDA_ARCHS}")
 else()
-  cuda_archs_loose_intersection(QUTLASS_ARCHS "12.0a;12.1a;10.0a;10.3a" "${CUDA_ARCHS}")
+  cuda_archs_loose_intersection(QUTLASS_SM120_ARCHS "12.0a;12.1a" "${CUDA_ARCHS}")
+  cuda_archs_loose_intersection(QUTLASS_SM100_ARCHS "10.0a;10.3a" "${CUDA_ARCHS}")
+endif()
+
+# QUTLASS uses TARGET_CUDA_ARCH as a single preprocessor selector for all its
+# sources. Do not compile a mixed SM100/SM120 arch list with one selector; prefer
+# SM100 when both families are requested because that is the primary deployed
+# target for this extension today.
+if(QUTLASS_SM100_ARCHS)
+  set(QUTLASS_ARCHS "${QUTLASS_SM100_ARCHS}")
+  set(QUTLASS_TARGET_CC 100)
+  if(QUTLASS_SM120_ARCHS)
+    message(WARNING
+      "[QUTLASS] Both SM100 and SM120 archs were requested; selecting SM100 "
+      "because TARGET_CUDA_ARCH is a single compile-time selector.")
+  endif()
+elseif(QUTLASS_SM120_ARCHS)
+  set(QUTLASS_ARCHS "${QUTLASS_SM120_ARCHS}")
+  set(QUTLASS_TARGET_CC 120)
+else()
+  set(QUTLASS_ARCHS)
 endif()
 
 if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND QUTLASS_ARCHS)
-
-  if(QUTLASS_ARCHS MATCHES "10\\.(0a|3a|0f)")
-    set(QUTLASS_TARGET_CC 100)
-  elseif(QUTLASS_ARCHS MATCHES "12\\.[01][af]?")
-    set(QUTLASS_TARGET_CC 120)
-  else()
-    message(FATAL_ERROR "[QUTLASS] internal error parsing CUDA_ARCHS='${QUTLASS_ARCHS}'.")
-  endif()
-
   set(QUTLASS_SOURCES
     ${qutlass_SOURCE_DIR}/qutlass/csrc/bindings.cpp
     ${qutlass_SOURCE_DIR}/qutlass/csrc/gemm.cu

cmake/external_projects/vllm_flash_attn.cmake

@@ -39,7 +39,7 @@ else()
   FetchContent_Declare(
           vllm-flash-attn
           GIT_REPOSITORY https://github.com/vllm-project/flash-attention.git
-          GIT_TAG dd62dac706b1cf7895bd99b18c6cb7e7e117ee25
+          GIT_TAG 803020a8fa15407871341d41eba4919ade2ee1ee
           GIT_PROGRESS TRUE
           # Don't share the vllm-flash-attn build between build types
           BINARY_DIR ${CMAKE_BINARY_DIR}/vllm-flash-attn

cmake/hipify.py

@@ -14,7 +14,18 @@ import argparse
 import os
 import shutil
 
-from torch.utils.hipify.hipify_python import hipify
+from torch.utils.hipify.hipify_python import get_hip_file_path, hipify
+
+
+def _expected_hip_build_path(source_abs: str, output_directory: str) -> str:
+    """Match torch.utils.hipify.hipify_python.preprocessor fout_path naming."""
+    rel = os.path.relpath(source_abs, output_directory)
+    return os.path.abspath(
+        os.path.join(
+            output_directory, get_hip_file_path(rel, is_pytorch_extension=True)
+        )
+    )
+
 
 if __name__ == "__main__":
     parser = argparse.ArgumentParser()
@@ -53,7 +64,11 @@ if __name__ == "__main__":
     hipify_result = hipify(
         project_directory=args.project_dir,
         output_directory=args.output_dir,
-        header_include_dirs=[],
+        # Hipify resolves quoted includes next to the including file first; vLLM
+        # uses paths relative to csrc/ (e.g. "libtorch_stable/torch_utils.h"
+        # from quantization/w8a8/fp8/*.cu). Without an include root here, those
+        # headers are never found and are not hipified or rewritten in dependents.
+        header_include_dirs=["."],
         includes=includes,
         extra_files=extra_files,
         show_detailed=True,
@@ -64,14 +79,20 @@ if __name__ == "__main__":
     hipified_sources = []
     for source in args.sources:
         s_abs = os.path.abspath(source)
-        hipified_s_abs = (
-            hipify_result[s_abs].hipified_path
-            if (
-                s_abs in hipify_result
-                and hipify_result[s_abs].hipified_path is not None
-            )
-            else s_abs
-        )
+        if s_abs in hipify_result and hipify_result[s_abs].hipified_path is not None:
+            path = hipify_result[s_abs].hipified_path
+            # PyTorch skips writing when is_pytorch_extension and text unchanged;
+            # hipified_path then stays *.cu. CMake expects *.hip under output_dir.
+            if s_abs.endswith(".cu") and path.endswith(".cu"):
+                dest = _expected_hip_build_path(s_abs, args.output_dir)
+                if os.path.normpath(path) != os.path.normpath(dest):
+                    os.makedirs(os.path.dirname(dest), exist_ok=True)
+                    shutil.copy2(path, dest)
+                hipified_s_abs = dest
+            else:
+                hipified_s_abs = path
+        else:
+            hipified_s_abs = s_abs
         hipified_sources.append(hipified_s_abs)
 
     assert len(hipified_sources) == len(args.sources)

cmake/utils.cmake

@@ -81,6 +81,14 @@ function (hipify_sources_target OUT_SRCS NAME ORIG_SRCS)
   set_property(GLOBAL APPEND PROPERTY VLLM_HIPIFY_ALL_SRCS ${SRCS})
   set_property(GLOBAL APPEND PROPERTY VLLM_HIPIFY_ALL_BYPRODUCTS ${HIP_SRCS})
 
+  # Chain hipify targets so they run sequentially. Parallel hipify
+  # invocations race on shutil.copytree, overwriting .hip files
+  # produced by another target back to .cu originals.
+  if (DEFINED _VLLM_LAST_HIPIFY_TARGET)
+    add_dependencies(hipify${NAME} ${_VLLM_LAST_HIPIFY_TARGET})
+  endif()
+  set(_VLLM_LAST_HIPIFY_TARGET "hipify${NAME}" PARENT_SCOPE)
+
   # Swap out original extension sources with hipified sources.
   list(APPEND HIP_SRCS ${CXX_SRCS})
   set(${OUT_SRCS} ${HIP_SRCS} PARENT_SCOPE)
@@ -479,9 +487,9 @@ endfunction()
 
 function(cuda_archs_sm90plus OUT_CUDA_ARCHS TGT_CUDA_ARCHS)
   if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
-    cuda_archs_loose_intersection(_archs "9.0a;10.0f;11.0f" "${TGT_CUDA_ARCHS}")
+    cuda_archs_loose_intersection(_archs "9.0a;10.0f;11.0f;12.0f" "${TGT_CUDA_ARCHS}")
   else()
-    cuda_archs_loose_intersection(_archs "9.0a;10.0a;10.1a;10.3a" "${TGT_CUDA_ARCHS}")
+    cuda_archs_loose_intersection(_archs "9.0a;10.0a;10.1a;10.3a;12.0a;12.1a" "${TGT_CUDA_ARCHS}")
   endif()
   set(${OUT_CUDA_ARCHS} ${_archs} PARENT_SCOPE)
 endfunction()

csrc/cpu/cpu_attn.cpp

@@ -11,13 +11,25 @@ static inline cpu_attention::Fp8KVCacheDataType parse_fp8_kv_dtype(
   return cpu_attention::Fp8KVCacheDataType::kAuto;
 }
 
+bool cpu_attn_has_isa(const std::string& isa) {
+  if (isa == "rvv") {
+#if defined(__riscv) && defined(__riscv_v_min_vlen) && __riscv_v_min_vlen == 128
+    return true;
+#else
+    return false;
+#endif
+  }
+  return false;
+}
+
 torch::Tensor get_scheduler_metadata(
     const int64_t num_req, const int64_t num_heads_q,
     const int64_t num_heads_kv, const int64_t head_dim,
     const torch::Tensor& seq_lens, at::ScalarType dtype,
-    const torch::Tensor& query_start_loc, const bool casual,
+    const torch::Tensor& query_start_loc, const bool causal,
     const int64_t window_size, const std::string& isa_hint,
-    const bool enable_kv_split) {
+    const bool enable_kv_split,
+    const std::optional<torch::Tensor>& dynamic_causal) {
   cpu_attention::ISA isa;
   if (isa_hint == "amx") {
     isa = cpu_attention::ISA::AMX;
@@ -44,24 +56,13 @@ torch::Tensor get_scheduler_metadata(
   input.head_dim = head_dim;
   input.query_start_loc = query_start_loc.data_ptr<int32_t>();
   input.seq_lens = seq_lens.data_ptr<int32_t>();
-  if (window_size != -1) {
-    input.left_sliding_window_size = window_size - 1;
-    if (casual) {
-      input.right_sliding_window_size = 0;
-    } else {
-      input.right_sliding_window_size = window_size - 1;
-    }
-  } else {
-    input.left_sliding_window_size = -1;
-    if (casual) {
-      input.right_sliding_window_size = 0;
-    } else {
-      input.right_sliding_window_size = -1;
-    }
-  }
-  input.casual = casual;
+
+  input.sliding_window_size = window_size;
+  input.causal = causal;
   input.isa = isa;
   input.enable_kv_split = enable_kv_split;
+  input.dynamic_causal =
+      dynamic_causal.has_value() ? dynamic_causal->data_ptr<bool>() : nullptr;
 
   VLLM_DISPATCH_FLOATING_TYPES(dtype, "get_scheduler_metadata", [&]() {
     CPU_ATTN_DISPATCH(head_dim, isa, 0, [&]() {
@@ -175,10 +176,11 @@ void cpu_attention_with_kv_cache(
     const torch::Tensor& seq_lens,         // [num_tokens]
     const double scale, const bool causal,
     const std::optional<torch::Tensor>& alibi_slopes,  // [num_heads]
-    const int64_t sliding_window_left, const int64_t sliding_window_right,
+    const int64_t sliding_window,
     const torch::Tensor& block_table,  // [num_tokens, max_block_num]
     const double softcap, const torch::Tensor& scheduler_metadata,
-    const std::optional<torch::Tensor>& s_aux,  // [num_heads]
+    const std::optional<torch::Tensor>& s_aux,           // [num_heads]
+    const std::optional<torch::Tensor>& dynamic_causal,  // [num_reqs]
     const double k_scale = 1.0, const double v_scale = 1.0,
     const std::string& kv_cache_dtype = "auto") {
   TORCH_CHECK_EQ(query.dim(), 3);
@@ -220,13 +222,11 @@ void cpu_attention_with_kv_cache(
   input.alibi_slopes =
       alibi_slopes.has_value() ? alibi_slopes->data_ptr<float>() : nullptr;
   input.s_aux = s_aux.has_value() ? s_aux->data_ptr<c10::BFloat16>() : nullptr;
+  input.dynamic_causal =
+      dynamic_causal.has_value() ? dynamic_causal->data_ptr<bool>() : nullptr;
   input.scale = scale;
   input.causal = causal;
-  input.sliding_window_left = sliding_window_left;
-  input.sliding_window_right = sliding_window_right;
-  if (input.causal) {
-    input.sliding_window_right = 0;
-  }
+  input.sliding_window_size = sliding_window;
   input.softcap = static_cast<float>(softcap);
 
   if (is_fp8) {

csrc/cpu/cpu_attn_impl.hpp

@@ -388,13 +388,13 @@ class AttentionScheduler {
     int32_t head_dim;
     int32_t* query_start_loc;
     int32_t* seq_lens;
-    int32_t left_sliding_window_size;
-    int32_t right_sliding_window_size;
-    bool casual;
+    int32_t sliding_window_size;
+    bool causal;
     cpu_attention::ISA isa;
     int32_t max_num_q_per_iter;  // max Q head num can be hold in registers
     int32_t kv_block_alignment;  // context length alignment requirement
     bool enable_kv_split;
+    bool* dynamic_causal;
   };
 
   static constexpr int32_t MaxQTileIterNum = 128;
@@ -403,7 +403,8 @@ class AttentionScheduler {
       : available_cache_size_(cpu_utils::get_available_l2_size()) {}
 
   torch::Tensor schedule(const ScheduleInput& input) const {
-    const bool casual = input.casual;
+    const bool causal = input.causal;
+    const bool is_dynamic_causal = input.dynamic_causal != nullptr;
     const int32_t thread_num = omp_get_max_threads();
     const int64_t cache_size = cpu_utils::get_available_l2_size();
     const int32_t max_num_q_per_iter = input.max_num_q_per_iter;
@@ -434,8 +435,7 @@ class AttentionScheduler {
     const int32_t default_tile_token_num = default_tile_size / q_head_per_kv;
     const int32_t split_kv_q_token_num_threshold =
         input.enable_kv_split ? 1 : 0;
-    const int32_t left_sliding_window_size = input.left_sliding_window_size;
-    const int32_t right_sliding_window_size = input.right_sliding_window_size;
+    const int32_t sliding_window_size = input.sliding_window_size;
     TORCH_CHECK_LE(split_kv_q_token_num_threshold * q_head_per_kv, 16);
 
     // get total kv len
@@ -444,7 +444,9 @@ class AttentionScheduler {
       const int32_t seq_len = input.seq_lens[req_id];
       const int32_t q_token_num =
           input.query_start_loc[req_id + 1] - input.query_start_loc[req_id];
-      const int32_t q_start_pos = (casual ? (seq_len - q_token_num) : 0);
+      const bool req_causal =
+          is_dynamic_causal ? input.dynamic_causal[req_id] : causal;
+      const int32_t q_start_pos = seq_len - q_token_num;
       const int32_t kv_start_pos = 0;
       const int32_t kv_end_pos = seq_len;
 
@@ -456,7 +458,7 @@ class AttentionScheduler {
         const int32_t q_tile_pos_right = q_tile_pos_left + q_tile_token_num;
         const auto [kv_tile_pos_left, kv_tile_pos_right] = calcu_kv_tile_pos(
             kv_start_pos, kv_end_pos, q_tile_pos_left, q_tile_pos_right,
-            left_sliding_window_size, right_sliding_window_size);
+            sliding_window_size, req_causal);
         const auto [aligned_kv_tile_pos_left, aligned_kv_tile_pos_right] =
             align_kv_tile_pos(kv_tile_pos_left, kv_tile_pos_right,
                               kv_len_alignment);
@@ -484,7 +486,9 @@ class AttentionScheduler {
       const int32_t seq_len = input.seq_lens[req_id];
       const int32_t q_token_num =
           input.query_start_loc[req_id + 1] - input.query_start_loc[req_id];
-      const int32_t q_start_pos = (casual ? (seq_len - q_token_num) : 0);
+      const bool req_causal =
+          is_dynamic_causal ? input.dynamic_causal[req_id] : causal;
+      const int32_t q_start_pos = seq_len - q_token_num;
       const int32_t kv_start_pos = 0;
       const int32_t kv_end_pos = seq_len;
       int32_t local_split_id = 0;
@@ -498,7 +502,7 @@ class AttentionScheduler {
         const int32_t q_tile_pos_right = q_tile_pos_left + q_tile_token_num;
         const auto [kv_tile_pos_left, kv_tile_pos_right] = calcu_kv_tile_pos(
             kv_start_pos, kv_end_pos, q_tile_pos_left, q_tile_pos_right,
-            left_sliding_window_size, right_sliding_window_size);
+            sliding_window_size, req_causal);
         const auto [aligned_kv_tile_pos_left, aligned_kv_tile_pos_right] =
             align_kv_tile_pos(kv_tile_pos_left, kv_tile_pos_right,
                               kv_len_alignment);
@@ -708,15 +712,41 @@ class AttentionScheduler {
     return metadata_tensor;
   }
 
+  FORCE_INLINE static std::pair<int32_t, int32_t> calcu_sliding_window_size(
+      int32_t window_size, bool causal) {
+    int32_t left_sliding_window_size, right_sliding_window_size;
+    if (window_size != -1) {
+      left_sliding_window_size = window_size - 1;
+      if (causal) {
+        right_sliding_window_size = 0;
+      } else {
+        right_sliding_window_size = window_size - 1;
+      }
+    } else {
+      left_sliding_window_size = -1;
+      if (causal) {
+        right_sliding_window_size = 0;
+      } else {
+        right_sliding_window_size = -1;
+      }
+    }
+
+    return {left_sliding_window_size, right_sliding_window_size};
+  }
+
   FORCE_INLINE static std::pair<int32_t, int32_t> calcu_kv_tile_pos(
       int32_t kv_left_pos, int32_t kv_right_pos, int32_t q_left_pos,
-      int32_t q_right_pos, int32_t sliding_window_left,
-      int32_t sliding_window_right) {
-    if (sliding_window_left != -1) {
-      kv_left_pos = std::max(kv_left_pos, q_left_pos - sliding_window_left);
+      int32_t q_right_pos, int32_t window_size, bool causal) {
+    auto [left_sliding_window_size, right_sliding_window_size] =
+        calcu_sliding_window_size(window_size, causal);
+
+    if (left_sliding_window_size != -1) {
+      kv_left_pos =
+          std::max(kv_left_pos, q_left_pos - left_sliding_window_size);
     }
-    if (sliding_window_right != -1) {
-      kv_right_pos = std::min(kv_right_pos, q_right_pos + sliding_window_right);
+    if (right_sliding_window_size != -1) {
+      kv_right_pos =
+          std::min(kv_right_pos, q_right_pos + right_sliding_window_size);
     }
     return {kv_left_pos, kv_right_pos};
   }
@@ -805,10 +835,10 @@ struct AttentionInput {
   int32_t* block_table;
   float* alibi_slopes;
   c10::BFloat16* s_aux;
+  bool* dynamic_causal;
   float scale;
   bool causal;
-  int32_t sliding_window_left;
-  int32_t sliding_window_right;
+  int32_t sliding_window_size;
   float softcap;
   // FP8 KV cache scales (used by FP8 attention implementations)
   float k_scale_fp8 = 1.0f;
@@ -822,8 +852,8 @@ struct AttentionInput {
       logits_buffer_t *__restrict__ logits_buffer,                          \
       float *__restrict__ partial_q_buffer, float *__restrict__ max_buffer, \
       float *__restrict__ sum_buffer, int32_t *__restrict__ block_table,    \
-      const int32_t kv_tile_start_pos, const int32_t kv_tile_end_pos,       \
-      const int32_t kv_tile_token_num,                                      \
+      const int32_t kv_end_pos, const int32_t kv_tile_start_pos,            \
+      const int32_t kv_tile_end_pos, const int32_t kv_tile_token_num,       \
       const int64_t kv_cache_num_blocks_stride, const int32_t q_head_num,   \
       const int32_t q_token_num, const int32_t q_tile_start_pos,            \
       const int32_t q_heads_per_kv, const int32_t block_size,               \
@@ -834,7 +864,7 @@ struct AttentionInput {
 
 #define CPU_ATTENTION_PARAMS                                                  \
   q_heads_buffer, k_head_cache_ptr, v_head_cache_ptr, logits_buffer,          \
-      partial_q_buffer, max_buffer, sum_buffer, block_table,                  \
+      partial_q_buffer, max_buffer, sum_buffer, block_table, kv_end_pos,      \
       kv_tile_start_pos, kv_tile_end_pos, kv_tile_token_num,                  \
       kv_cache_num_blocks_stride, q_head_num, q_token_num, q_tile_start_pos,  \
       q_heads_per_kv, block_size, left_window_size, right_window_size, scale, \
@@ -917,6 +947,7 @@ class AttentionMainLoop {
     //  - max_buffer: [MaxQHeadNumPerIteration, 1], store max logits
     //  - sum_buffer: [MaxQHeadNumPerIteration, 1], store sum of exp
     //  - block_table
+    //  - kv_end_pos: un-aligned end position of KV cache
     //  - kv_tile_start_pos: start position of KV cache, aligned to
     //  BlockSizeAlignment
     //  - kv_tile_end_pos: end position of KV cache, aligned to
@@ -1043,7 +1074,7 @@ class AttentionMainLoop {
         }
 
         apply_mask(logits_buffer, kv_tile_token_num, q_tile_start_pos,
-                   kv_tile_start_pos, kv_tile_end_pos, q_token_num,
+                   kv_end_pos, kv_tile_start_pos, kv_tile_end_pos, q_token_num,
                    q_heads_per_kv, left_window_size, right_window_size);
 
         // if (debug_info){
@@ -1126,7 +1157,7 @@ class AttentionMainLoop {
 
     void apply_mask(logits_buffer_t* __restrict__ logits_buffer,
                     const int64_t logits_buffer_stride,
-                    const int32_t q_tile_start_pos,
+                    const int32_t q_tile_start_pos, const int32_t kv_end_pos,
                     const int32_t kv_tile_start_pos,
                     const int32_t kv_tile_end_pos, const int32_t q_token_num,
                     const int32_t q_heads_per_kv,
@@ -1154,7 +1185,7 @@ class AttentionMainLoop {
                            std::max(kv_tile_start_pos,
                                     curr_token_pos + sliding_window_right + 1));
           }
-          return pos;
+          return std::min(pos, kv_end_pos);
         }();
 
         int32_t left_invalid_token_num = left_kv_pos - kv_tile_start_pos;
@@ -1441,15 +1472,16 @@ class AttentionMainLoop {
       const int64_t q_head_num_stride = input->query_num_heads_stride;
       const int64_t kv_cache_head_num_stride = input->cache_num_kv_heads_stride;
       const int64_t kv_cache_block_num_stride = input->cache_num_blocks_stride;
-      const int32_t sliding_window_left = input->sliding_window_left;
-      const int32_t sliding_window_right = input->sliding_window_right;
+      const int32_t sliding_window_size = input->sliding_window_size;
       const int32_t block_size = input->block_size;
       const float scale = input->scale;
       const float softcap_scale = input->softcap;
       const float* alibi_slopes = input->alibi_slopes;
       const c10::BFloat16* s_aux = input->s_aux;
+      const bool* dynamic_causal = input->dynamic_causal;
+      const bool is_dynamic_causal = dynamic_causal != nullptr;
 
-      const bool casual = input->causal;
+      const bool causal = input->causal;
       int32_t* const block_table = input->block_table;
       const int64_t block_table_stride = input->blt_num_tokens_stride;
 
@@ -1532,6 +1564,11 @@ class AttentionMainLoop {
                 &curr_workitem_groups[workitem_group_idx];
 
             const int32_t current_group_idx = current_workitem_group->req_id;
+            const int32_t current_group_causal =
+                is_dynamic_causal ? dynamic_causal[current_group_idx] : causal;
+            auto [sliding_window_left, sliding_window_right] =
+                AttentionScheduler::calcu_sliding_window_size(
+                    sliding_window_size, current_group_causal);
             const int32_t kv_start_pos =
                 current_workitem_group->kv_split_pos_start;
             const int32_t kv_end_pos = current_workitem_group->kv_split_pos_end;
@@ -1559,8 +1596,7 @@ class AttentionMainLoop {
             const int32_t q_end = input->query_start_loc[current_group_idx + 1];
             const int32_t q_start = input->query_start_loc[current_group_idx];
             const int32_t seq_len = input->seq_lens[current_group_idx];
-            const int32_t q_start_pos =
-                (casual ? seq_len - (q_end - q_start) : 0);
+            const int32_t q_start_pos = seq_len - (q_end - q_start);
             const int32_t block_num = (seq_len + block_size - 1) / block_size;
             // Only apply sink for the first KV split
             bool use_sink = (s_aux != nullptr &&
@@ -1610,8 +1646,8 @@ class AttentionMainLoop {
               const auto [kv_tile_start_pos, kv_tile_end_pos] =
                   AttentionScheduler::calcu_kv_tile_pos(
                       kv_start_pos, kv_end_pos, q_tile_start_pos,
-                      q_tile_end_pos, sliding_window_left,
-                      sliding_window_right);
+                      q_tile_end_pos, sliding_window_size,
+                      current_group_causal);
               const auto [rounded_kv_tile_start_pos, rounded_kv_tile_end_pos] =
                   AttentionScheduler::align_kv_tile_pos(
                       kv_tile_start_pos, kv_tile_end_pos, blocksize_alignment);
@@ -1724,8 +1760,8 @@ class AttentionMainLoop {
                               actual_kv_tile_pos_right] =
                       AttentionScheduler::calcu_kv_tile_pos(
                           kv_tile_pos_left, kv_tile_pos_right, q_tile_pos_left,
-                          q_tile_pos_right, sliding_window_left,
-                          sliding_window_right);
+                          q_tile_pos_right, sliding_window_size,
+                          current_group_causal);
                   const int32_t q_iter_idx =
                       q_head_tile_token_offset / curr_max_q_token_num_per_iter;
 
@@ -1789,7 +1825,7 @@ class AttentionMainLoop {
                   attn_impl.template execute_attention<Attention>(
                       curr_q_heads_buffer, curr_k_cache, curr_v_cache,
                       logits_buffer, curr_partial_q_buffer, curr_max_buffer,
-                      curr_sum_buffer, curr_block_table,
+                      curr_sum_buffer, curr_block_table, kv_end_pos,
                       aligned_actual_kv_tile_pos_left,
                       aligned_actual_kv_tile_pos_right, actual_kv_token_num,
                       kv_cache_block_num_stride, q_tile_head_num,

csrc/cpu/cpu_fused_moe.cpp

@@ -1,3 +1,5 @@
+#include <sleef.h>
+
 #include "cpu/cpu_types.hpp"
 #include "cpu/utils.hpp"
 #include "cpu/micro_gemm/cpu_micro_gemm_vec.hpp"
@@ -30,7 +32,12 @@
   }()
 
 namespace {
-enum class FusedMOEAct { SiluAndMul, SwigluOAIAndMul, GeluAndMul };
+enum class FusedMOEAct {
+  SiluAndMul,
+  SwigluOAIAndMul,
+  GeluAndMul,
+  GeluTanhAndMul,
+};
 
 FusedMOEAct get_act_type(const std::string& act) {
   if (act == "silu") {
@@ -39,6 +46,8 @@ FusedMOEAct get_act_type(const std::string& act) {
     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);
   }
@@ -143,6 +152,40 @@ void gelu_and_mul(float* __restrict__ input, scalar_t* __restrict__ output,
   }
 }
 
+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
+      vec_op::FP32Vec16 tanh_vec(Sleef_tanhf16_u10(inner_vec.reg));
+      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,
@@ -160,6 +203,9 @@ FORCE_INLINE void apply_gated_act(const FusedMOEAct act,
     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.");
   }

csrc/cpu/cpu_types_riscv_defs.hpp

@@ -57,6 +57,10 @@ typedef RVVTYPE(vfloat32, LMUL_512, _t) fixed_fp32x16_t
 typedef RVVTYPE(vfloat32, LMUL_1024, _t) fixed_fp32x32_t
     __attribute__((riscv_rvv_vector_bits(1024)));
 
+// int8
+typedef RVVTYPE(vint8, LMUL_128, _t) fixed_i8x16_t
+    __attribute__((riscv_rvv_vector_bits(128)));
+
 // int32
 typedef RVVTYPE(vint32, LMUL_256, _t) fixed_i32x8_t
     __attribute__((riscv_rvv_vector_bits(256)));

csrc/cpu/cpu_types_riscv_impl.hpp

@@ -9,10 +9,14 @@
 
 #include <algorithm>
 #include <cmath>
+#include <cstdint>
 #include <cstring>
 #include <iostream>
 #include <limits>
 #include <torch/all.h>
+
+#include "float_convert.hpp"
+
 namespace vec_op {
 
 // FP8 KV cache is not supported on RISC-V. These tag types and the
@@ -245,8 +249,7 @@ struct BF16Vec8 : public Vec<BF16Vec8> {
     const uint16_t* u16 = static_cast<const uint16_t*>(ptr);
     float tmp[8];
     for (int i = 0; i < 8; ++i) {
-      uint32_t v = static_cast<uint32_t>(u16[i]) << 16;
-      std::memcpy(&tmp[i], &v, 4);
+      tmp[i] = bf16_to_float(u16[i]);
     }
     reg_fp32 = RVVI(__riscv_vle32_v_f32, LMUL_256)(tmp, 8);
   }
@@ -256,9 +259,7 @@ struct BF16Vec8 : public Vec<BF16Vec8> {
     RVVI(__riscv_vse32_v_f32, LMUL_256)(tmp, reg_fp32, 8);
     uint16_t* u16 = static_cast<uint16_t*>(ptr);
     for (int i = 0; i < 8; ++i) {
-      uint32_t v;
-      std::memcpy(&v, &tmp[i], 4);
-      u16[i] = static_cast<uint16_t>(v >> 16);
+      u16[i] = float_to_bf16(tmp[i]);
     }
   }
   void save(void* ptr, int elem_num) const {
@@ -266,9 +267,7 @@ struct BF16Vec8 : public Vec<BF16Vec8> {
     RVVI(__riscv_vse32_v_f32, LMUL_256)(tmp, reg_fp32, 8);
     uint16_t* u16 = static_cast<uint16_t*>(ptr);
     for (int i = 0; i < elem_num; ++i) {
-      uint32_t v;
-      std::memcpy(&v, &tmp[i], 4);
-      u16[i] = static_cast<uint16_t>(v >> 16);
+      u16[i] = float_to_bf16(tmp[i]);
     }
   }
   void save_strided(void* ptr, ptrdiff_t stride) const {
@@ -277,10 +276,8 @@ struct BF16Vec8 : public Vec<BF16Vec8> {
     uint8_t* u8 = static_cast<uint8_t*>(ptr);
     ptrdiff_t byte_stride = stride * sizeof(uint16_t);
     for (int i = 0; i < 8; ++i) {
-      uint32_t v;
-      std::memcpy(&v, &tmp[i], 4);
-      uint16_t val = static_cast<uint16_t>(v >> 16);
-      *reinterpret_cast<uint16_t*>(u8 + i * byte_stride) = val;
+      *reinterpret_cast<uint16_t*>(u8 + i * byte_stride) =
+          float_to_bf16(tmp[i]);
     }
   }
 };
@@ -292,8 +289,7 @@ struct BF16Vec16 : public Vec<BF16Vec16> {
     const uint16_t* u16 = static_cast<const uint16_t*>(ptr);
     float tmp[16];
     for (int i = 0; i < 16; ++i) {
-      uint32_t v = static_cast<uint32_t>(u16[i]) << 16;
-      std::memcpy(&tmp[i], &v, 4);
+      tmp[i] = bf16_to_float(u16[i]);
     }
     reg_fp32 = RVVI(__riscv_vle32_v_f32, LMUL_512)(tmp, 16);
   }
@@ -306,9 +302,7 @@ struct BF16Vec16 : public Vec<BF16Vec16> {
     RVVI(__riscv_vse32_v_f32, LMUL_512)(tmp, reg_fp32, 16);
     uint16_t* u16 = static_cast<uint16_t*>(ptr);
     for (int i = 0; i < 16; ++i) {
-      uint32_t v;
-      std::memcpy(&v, &tmp[i], 4);
-      u16[i] = static_cast<uint16_t>(v >> 16);
+      u16[i] = float_to_bf16(tmp[i]);
     }
   }
   void save(void* ptr, int elem_num) const {
@@ -316,9 +310,7 @@ struct BF16Vec16 : public Vec<BF16Vec16> {
     RVVI(__riscv_vse32_v_f32, LMUL_512)(tmp, reg_fp32, 16);
     uint16_t* u16 = static_cast<uint16_t*>(ptr);
     for (int i = 0; i < elem_num; ++i) {
-      uint32_t v;
-      std::memcpy(&v, &tmp[i], 4);
-      u16[i] = static_cast<uint16_t>(v >> 16);
+      u16[i] = float_to_bf16(tmp[i]);
     }
   }
   void save_strided(void* ptr, ptrdiff_t stride) const {
@@ -327,10 +319,8 @@ struct BF16Vec16 : public Vec<BF16Vec16> {
     uint8_t* u8 = static_cast<uint8_t*>(ptr);
     ptrdiff_t byte_stride = stride * sizeof(uint16_t);
     for (int i = 0; i < 16; ++i) {
-      uint32_t v;
-      std::memcpy(&v, &tmp[i], 4);
-      uint16_t val = static_cast<uint16_t>(v >> 16);
-      *reinterpret_cast<uint16_t*>(u8 + i * byte_stride) = val;
+      *reinterpret_cast<uint16_t*>(u8 + i * byte_stride) =
+          float_to_bf16(tmp[i]);
     }
   }
 };
@@ -343,8 +333,7 @@ struct BF16Vec32 : public Vec<BF16Vec32> {
     const uint16_t* u16 = static_cast<const uint16_t*>(ptr);
     float tmp[32];
     for (int i = 0; i < 32; ++i) {
-      uint32_t v = static_cast<uint32_t>(u16[i]) << 16;
-      std::memcpy(&tmp[i], &v, 4);
+      tmp[i] = bf16_to_float(u16[i]);
     }
     reg_fp32 = RVVI(__riscv_vle32_v_f32, LMUL_1024)(tmp, 32);
   }
@@ -371,9 +360,7 @@ struct BF16Vec32 : public Vec<BF16Vec32> {
     RVVI(__riscv_vse32_v_f32, LMUL_1024)(tmp, reg_fp32, 32);
     uint16_t* u16 = static_cast<uint16_t*>(ptr);
     for (int i = 0; i < 32; ++i) {
-      uint32_t v;
-      std::memcpy(&v, &tmp[i], 4);
-      u16[i] = static_cast<uint16_t>(v >> 16);
+      u16[i] = float_to_bf16(tmp[i]);
     }
   }
 
@@ -382,9 +369,7 @@ struct BF16Vec32 : public Vec<BF16Vec32> {
     RVVI(__riscv_vse32_v_f32, LMUL_1024)(tmp, reg_fp32, 32);
     uint16_t* u16 = static_cast<uint16_t*>(ptr);
     for (int i = 0; i < elem_num; ++i) {
-      uint32_t v;
-      std::memcpy(&v, &tmp[i], 4);
-      u16[i] = static_cast<uint16_t>(v >> 16);
+      u16[i] = float_to_bf16(tmp[i]);
     }
   }
 
@@ -394,10 +379,8 @@ struct BF16Vec32 : public Vec<BF16Vec32> {
     uint8_t* u8 = static_cast<uint8_t*>(ptr);
     ptrdiff_t byte_stride = stride * sizeof(uint16_t);
     for (int i = 0; i < 32; ++i) {
-      uint32_t v;
-      std::memcpy(&v, &tmp[i], 4);
-      uint16_t val = static_cast<uint16_t>(v >> 16);
-      *reinterpret_cast<uint16_t*>(u8 + i * byte_stride) = val;
+      *reinterpret_cast<uint16_t*>(u8 + i * byte_stride) =
+          float_to_bf16(tmp[i]);
     }
   }
 };
@@ -734,10 +717,18 @@ struct FP32Vec16 : public Vec<FP32Vec16> {
     return FP32Vec16(
         RVVI(__riscv_vfmax_vv_f32, LMUL_512)(reg, b.reg, VEC_ELEM_NUM));
   }
+  FP32Vec16 max(const FP32Vec16& b, const int elem_num) const {
+    return FP32Vec16(
+        RVVI(__riscv_vfmax_vv_f32, LMUL_512)(reg, b.reg, elem_num));
+  }
   FP32Vec16 min(const FP32Vec16& b) const {
     return FP32Vec16(
         RVVI(__riscv_vfmin_vv_f32, LMUL_512)(reg, b.reg, VEC_ELEM_NUM));
   }
+  FP32Vec16 min(const FP32Vec16& b, const int elem_num) const {
+    return FP32Vec16(
+        RVVI(__riscv_vfmin_vv_f32, LMUL_512)(reg, b.reg, elem_num));
+  }
   FP32Vec16 abs() const {
     return FP32Vec16(RVVI(__riscv_vfabs_v_f32, LMUL_512)(reg, VEC_ELEM_NUM));
   }
@@ -867,6 +858,27 @@ struct FP32Vec16 : public Vec<FP32Vec16> {
   }
 };
 
+struct INT8Vec16 : public Vec<INT8Vec16> {
+  constexpr static int VEC_ELEM_NUM = 16;
+  fixed_i8x16_t reg;
+
+  explicit INT8Vec16(const FP32Vec16& vec) {
+    auto i32_vec =
+        RVVI(__riscv_vfcvt_x_f_v_i32, LMUL_512)(vec.reg, VEC_ELEM_NUM);
+    auto i16_vec = RVVI(__riscv_vnclip_wx_i16, LMUL_256)(
+        i32_vec, 0, __RISCV_VXRM_RNU, VEC_ELEM_NUM);
+    reg = RVVI(__riscv_vnclip_wx_i8, LMUL_128)(i16_vec, 0, __RISCV_VXRM_RNU,
+                                               VEC_ELEM_NUM);
+  }
+
+  void save(int8_t* ptr) const {
+    RVVI(__riscv_vse8_v_i8, LMUL_128)(ptr, reg, VEC_ELEM_NUM);
+  }
+  void save(int8_t* ptr, int elem_num) const {
+    RVVI(__riscv_vse8_v_i8, LMUL_128)(ptr, reg, elem_num);
+  }
+};
+
 // ============================================================================
 // Type Traits & Global Helpers
 // ============================================================================
@@ -956,9 +968,7 @@ inline BF16Vec16::BF16Vec16(const FP32Vec16& v)
 #else
 template <>
 inline void storeFP32<c10::BFloat16>(float v, c10::BFloat16* ptr) {
-  uint32_t val;
-  std::memcpy(&val, &v, 4);
-  *reinterpret_cast<uint16_t*>(ptr) = static_cast<uint16_t>(val >> 16);
+  *reinterpret_cast<uint16_t*>(ptr) = float_to_bf16(v);
 }
 inline BF16Vec8::BF16Vec8(const FP32Vec8& v) : reg_fp32(v.reg) {}
 inline BF16Vec16::BF16Vec16(const FP32Vec16& v) : reg_fp32(v.reg) {}

csrc/cpu/cpu_types_vxe.hpp

@@ -3,7 +3,9 @@
 #define CPU_TYPES_VXE_HPP
 
 #include <vecintrin.h>
+#include <bit>
 #include <cmath>
+#include <cstdint>
 #include <limits>
 #include <torch/all.h>
 namespace vec_op {
@@ -817,8 +819,7 @@ inline void storeFP32<::c10::Half>(float v, ::c10::Half* ptr) {
   // intrinsics for FP32 to FP16 conversion does not use IEEE rounding and can
   // produce incorrect results for some inputs. Process each of the 4 vectors
   // separately.
-  uint32_t in;
-  std::memcpy(&in, &v, sizeof(in));
+  uint32_t in = std::bit_cast<uint32_t>(v);
 
   uint32_t s = (in & 0x80000000) >> 16;  // Sign
   uint32_t e = (in & 0x7F800000) >> 23;  // Exponent