Bump actions/setup-python from 6.1.0 to 6.2.0

Bumps [actions/setup-python](https://github.com/actions/setup-python) from 6.1.0 to 6.2.0.
- [Release notes](https://github.com/actions/setup-python/releases)
- [Commits](https://github.com/actions/setup-python/compare/83679a892e2d95755f2dac6acb0bfd1e9ac5d548...a309ff8b426b58ec0e2a45f0f869d46889d02405)

---
updated-dependencies:
- dependency-name: actions/setup-python
  dependency-version: 6.2.0
  dependency-type: direct:production
  update-type: version-update:semver-minor
...

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

.buildkite/hardware_tests/amd.yaml

@@ -17,6 +17,26 @@ steps:
       --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}"
     env:
       DOCKER_BUILDKIT: "1"

.buildkite/hardware_tests/cpu.yaml

@@ -12,19 +12,21 @@ steps:
   - vllm/_custom_ops.py
   - tests/kernels/attention/test_cpu_attn.py
   - tests/kernels/moe/test_cpu_fused_moe.py
-  - tests/kernels/moe/test_cpu_fp8_fused_moe.py
+  - tests/kernels/moe/test_cpu_quant_fused_moe.py
   - tests/kernels/test_onednn.py
   - tests/kernels/test_awq_int4_to_int8.py
   - tests/kernels/quantization/test_cpu_fp8_scaled_mm.py
+  - tests/kernels/mamba/cpu/test_cpu_gdn_ops.py
   commands:
     - |
       bash .buildkite/scripts/hardware_ci/run-cpu-test.sh 30m "
       pytest -x -v -s tests/kernels/attention/test_cpu_attn.py
       pytest -x -v -s tests/kernels/moe/test_cpu_fused_moe.py
-      pytest -x -v -s tests/kernels/moe/test_cpu_fp8_fused_moe.py
+      pytest -x -v -s tests/kernels/moe/test_cpu_quant_fused_moe.py
       pytest -x -v -s tests/kernels/test_onednn.py
       pytest -x -v -s tests/kernels/test_awq_int4_to_int8.py
-      pytest -x -v -s tests/kernels/quantization/test_cpu_fp8_scaled_mm.py"
+      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: []
@@ -54,23 +56,41 @@ steps:
       pytest -x -v -s tests/models/language/generation -m cpu_model
       pytest -x -v -s tests/models/language/pooling -m cpu_model"
 
+- label: CPU-ModelRunnerV2 Tests
+  depends_on: []
+  device: intel_cpu
+  no_plugin: true
+  soft_fail: true
+  source_file_dependencies:
+  - vllm/v1/worker/cpu/
+  - vllm/v1/worker/gpu/
+  - vllm/v1/sample/ops/topk_topp_triton.py
+  - vllm/v1/sample/ops/topk_topp_sampler.py
+  - tests/v1/sample/test_topk_topp_sampler.py
+  commands:
+    - |
+      bash .buildkite/scripts/hardware_ci/run-cpu-test.sh 45m "
+      uv pip install git+https://github.com/triton-lang/triton-cpu.git@270e696d
+      VLLM_USE_V2_MODEL_RUNNER=1 pytest -x -v -s tests/models/language/generation/test_granite.py -m cpu_model
+      # TODO: move to CPU-Kernel Tests once triton-cpu has a pre-built wheel
+      pytest -x -v -s tests/v1/sample/test_topk_topp_sampler.py::TestTritonTopkTopp"
+
 - label: CPU-Quantization Model Tests
   depends_on: []
   device: intel_cpu
   no_plugin: true
   source_file_dependencies:
   - csrc/cpu/
-  - vllm/model_executor/layers/quantization/cpu_wna16.py
   - vllm/model_executor/layers/quantization/auto_gptq.py
   - vllm/model_executor/layers/quantization/compressed_tensors/schemes/compressed_tensors_w8a8_int8.py
-  - vllm/model_executor/layers/quantization/kernels/scaled_mm/cpu.py
-  - vllm/model_executor/layers/quantization/kernels/mixed_precision/cpu.py
+  - vllm/model_executor/kernels/linear/mixed_precision/cpu.py
+  - vllm/model_executor/kernels/linear/scaled_mm/cpu.py
   - vllm/model_executor/layers/fused_moe/experts/cpu_moe.py
   - tests/quantization/test_compressed_tensors.py
   - tests/quantization/test_cpu_wna16.py
   commands:
     - |
-      bash .buildkite/scripts/hardware_ci/run-cpu-test.sh 20m "
+      bash .buildkite/scripts/hardware_ci/run-cpu-test.sh 30m "
       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.yaml

@@ -8,10 +8,3 @@ steps:
     commands: 
     - bash .buildkite/scripts/hardware_ci/run-hpu-test.sh
 
-  - label: "Intel GPU Test"
-    depends_on: []
-    soft_fail: true
-    device: intel_gpu
-    no_plugin: true
-    commands: 
-    - bash .buildkite/scripts/hardware_ci/run-xpu-test.sh

.buildkite/image_build/image_build.yaml

@@ -13,6 +13,60 @@ steps:
         - exit_status: -10  # Agent was lost
           limit: 2
 
+  - label: ":docker: :smoking: Non-root smoke tests"
+    key: image-build-smoke-test
+    depends_on:
+      - image-build
+    commands:
+    # Smoke 1: the default (root) image must still be importable
+    # under a non-root UID via `--user 2000:0`. Validates the `vllm` passwd
+    # entry + group-0-writable /home/vllm + uv path cleanup from #31959.
+    # Uses `import vllm` rather than `vllm serve --help` because the latter
+    # instantiates `VllmConfig` which requires a GPU attached to the
+    # container.
+    - docker run --rm --user 2000:0 --entrypoint python3 "$IMAGE_TAG" -c "import vllm; print(vllm.__version__)"
+    # Smoke 2: assert the non-root enabling invariants are baked
+    # into the image. Runs as UID 2000:0 via a shell so we can verify
+    # filesystem perms + passwd/group file state + wrapper presence without
+    # triggering vLLM's GPU-requiring config-init path. The opt-in
+    # `vllm-openai-nonroot` target adds only `USER vllm`, `WORKDIR
+    # /home/vllm`, and an `ENTRYPOINT` override on top of these invariants;
+    # its build correctness is reviewed at the Dockerfile level. Wrapper
+    # logic is covered separately by the pre-commit hook
+    # `test-nonroot-entrypoint` (see .pre-commit-config.yaml).
+    - |
+      docker run --rm --user 2000:0 --entrypoint /bin/sh "$IMAGE_TAG" -ec '
+        if ! getent passwd 2000 | grep -q ^vllm:; then
+          echo FAIL: UID 2000 != vllm
+          exit 1
+        fi
+        if ! id -gn 2>/dev/null | grep -qx root; then
+          echo FAIL: GID 0 not root group
+          exit 1
+        fi
+        touch /home/vllm/.smoke && rm /home/vllm/.smoke
+        touch /opt/uv/cache/.smoke && rm /opt/uv/cache/.smoke
+        if ! test -x /usr/local/bin/vllm-nonroot-entrypoint.sh; then
+          echo FAIL: wrapper missing
+          exit 1
+        fi
+        if ! test -w /etc/passwd; then
+          echo FAIL: /etc/passwd not group-writable
+          exit 1
+        fi
+        if ! test -w /etc/group; then
+          echo FAIL: /etc/group not group-writable
+          exit 1
+        fi
+        echo non-root invariants OK
+      '
+    retry:
+      automatic:
+        - exit_status: -1  # Agent was lost
+          limit: 2
+        - exit_status: -10  # Agent was lost
+          limit: 2
+
   - label: ":docker: Build CPU image"
     key: image-build-cpu
     depends_on: []
@@ -56,3 +110,21 @@ steps:
           limit: 2
         - exit_status: -10  # Agent was lost
           limit: 2
+
+  - label: ":docker: Build arm64 image"
+    key: arm64-image-build
+    depends_on: []
+    source_file_dependencies:
+      - ".buildkite/image_build/image_build.yaml"
+      - ".buildkite/image_build/image_build_arm64.sh"
+      - "docker/Dockerfile"
+    commands:
+    - .buildkite/image_build/image_build_arm64.sh $REGISTRY $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

.buildkite/image_build/image_build_arm64.sh

@@ -0,0 +1,37 @@
+#!/bin/bash
+set -e
+
+if [[ $# -lt 3 ]]; then
+  echo "Usage: $0 <registry> <repo> <commit>"
+  exit 1
+fi
+
+REGISTRY=$1
+REPO=$2
+BUILDKITE_COMMIT=$3
+
+# authenticate with AWS ECR
+aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin "$REGISTRY" || true
+
+# skip build if image already exists
+if [[ -z $(docker manifest inspect "$REGISTRY"/"$REPO":"$BUILDKITE_COMMIT"-arm64) ]]; then
+  echo "Image not found, proceeding with build..."
+else
+  echo "Image found"
+  exit 0
+fi
+
+# build (Grace/GH200 is the arm64 GPU target; sm_90)
+docker build --file docker/Dockerfile \
+  --platform linux/arm64 \
+  --build-arg max_jobs=16 \
+  --build-arg nvcc_threads=4 \
+  --build-arg torch_cuda_arch_list="9.0" \
+  --build-arg USE_SCCACHE=1 \
+  --build-arg buildkite_commit="$BUILDKITE_COMMIT" \
+  --tag "$REGISTRY"/"$REPO":"$BUILDKITE_COMMIT"-arm64 \
+  --target test \
+  --progress plain .
+
+# push
+docker push "$REGISTRY"/"$REPO":"$BUILDKITE_COMMIT"-arm64

.buildkite/image_build/image_build_hpu.sh

@@ -11,7 +11,7 @@ REPO=$2
 BUILDKITE_COMMIT=$3
 
 # authenticate with AWS ECR
-aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin "$REGISTRY"
+aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin "$REGISTRY" || true
 
 # skip build if image already exists
 if [[ -z $(docker manifest inspect "$REGISTRY"/"$REPO":"$BUILDKITE_COMMIT"-hpu) ]]; then

.buildkite/image_build/image_build_xpu.sh

@@ -11,8 +11,8 @@ REPO=$2
 BUILDKITE_COMMIT=$3
 
 # authenticate with AWS ECR
-aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin "$REGISTRY"
-aws ecr get-login-password --region us-east-1 | docker login --username AWS --password-stdin 936637512419.dkr.ecr.us-east-1.amazonaws.com
+aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin "$REGISTRY" || true
+aws ecr get-login-password --region us-east-1 | docker login --username AWS --password-stdin 936637512419.dkr.ecr.us-east-1.amazonaws.com || true
 
 # skip build if image already exists
 if ! docker manifest inspect "$REGISTRY"/"$REPO":"$BUILDKITE_COMMIT"-xpu &> /dev/null; then

.buildkite/intel_jobs/lora_intel.yaml

@@ -49,7 +49,7 @@ steps:
       'cd tests &&
       export VLLM_WORKER_MULTIPROC_METHOD=spawn &&
       pytest -v -s lora/test_fused_moe_lora_kernel.py && 
-      pytest -v -s lora/test_moe_lora_align_sum.py'
+      pytest -v -s lora/test_moe_lora_align_sum.py --deselect="tests/lora/test_moe_lora_align_sum.py::test_moe_lora_align_block_size_mixed_base_and_lora[1]"'
 
 - label: LoRA Punica Kernels
   timeout_in_minutes: 45

.buildkite/intel_jobs/test-intel.yaml

@@ -39,7 +39,9 @@ steps:
         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 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
+        '
   - label: "XPU V1 test"
     depends_on:
       - image-build-xpu
@@ -64,3 +66,22 @@ steps:
         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'
+  - label: "XPU server 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
+        'pip install av &&
+        cd tests &&
+        pytest -v -s entrypoints/openai/chat_completion/test_audio_in_video.py &&
+        pytest -v -s benchmarks/test_serve_cli.py'

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

@@ -1,77 +1,112 @@
-[
-    {
-        "test_name": "serving_llama8B_tp1_sharegpt",
-        "qps_list": [1, 4, 16, "inf"],
-        "server_parameters": {
-            "model": "meta-llama/Meta-Llama-3.1-8B-Instruct",
-            "tensor_parallel_size": 1,
-            "disable_log_stats": "",
-            "load_format": "dummy"
-        },
-        "client_parameters": {
-            "model": "meta-llama/Meta-Llama-3.1-8B-Instruct",
-            "backend": "vllm",
-            "dataset_name": "sharegpt",
-            "dataset_path": "./ShareGPT_V3_unfiltered_cleaned_split.json",
-            "temperature": 0,
-            "num_prompts": 200
-        }
+{
+  "defaults": {
+    "qps_list": [
+      "inf"
+    ],
+    "max_concurrency_list": [12, 16, 24, 32, 64, 128, 200],
+    "server_parameters": {
+      "model": "meta-llama/Llama-3.1-8B-Instruct",
+      "tensor_parallel_size": 1,
+      "dtype": "bfloat16"
     },
-    {
-        "test_name": "serving_llama70B_tp4_sharegpt",
-        "qps_list": [1, 4, 16, "inf"],
-        "server_parameters": {
-            "model": "meta-llama/Meta-Llama-3.1-70B-Instruct",
-            "tensor_parallel_size": 4,
-            "disable_log_stats": "",
-            "load_format": "dummy"
-        },
-        "client_parameters": {
-            "model": "meta-llama/Meta-Llama-3.1-70B-Instruct",
-            "backend": "vllm",
-            "dataset_name": "sharegpt",
-            "dataset_path": "./ShareGPT_V3_unfiltered_cleaned_split.json",
-            "temperature": 0,
-            "num_prompts": 200
-        }
-    },
-    {
-        "test_name": "serving_mixtral8x7B_tp2_sharegpt",
-        "qps_list": [1, 4, 16, "inf"],
-        "server_parameters": {
-            "model": "mistralai/Mixtral-8x7B-Instruct-v0.1",
-            "tensor_parallel_size": 2,
-            "disable_log_stats": "",
-            "load_format": "dummy"
-        },
-        "client_parameters": {
-            "model": "mistralai/Mixtral-8x7B-Instruct-v0.1",
-            "backend": "vllm",
-            "dataset_name": "sharegpt",
-            "dataset_path": "./ShareGPT_V3_unfiltered_cleaned_split.json",
-            "temperature": 0,
-            "num_prompts": 200
-        }
-    },
-    {
-        "test_name": "serving_llama70B_tp4_sharegpt_specdecode",
-        "qps_list": [2],
-        "server_parameters": {
-            "model": "meta-llama/Meta-Llama-3.1-70B-Instruct", 
-            "tensor_parallel_size": 4,
-            "speculative_config": {
-                "model": "turboderp/Qwama-0.5B-Instruct",
-                "num_speculative_tokens": 4,
-                "draft_tensor_parallel_size": 1
-            }
-        },
-        "client_parameters": {
-            "model": "meta-llama/Meta-Llama-3.1-70B-Instruct",
-            "backend": "vllm",
-            "dataset_name": "sharegpt",
-            "dataset_path": "./ShareGPT_V3_unfiltered_cleaned_split.json",
-            "temperature": 0,
-            "num_prompts": 200 
-        }
+    "client_parameters": {
+      "model": "meta-llama/Llama-3.1-8B-Instruct",
+      "backend": "vllm",
+      "ignore-eos": "",
+      "temperature": 0,
+      "num_prompts": 200
     }
-]
+  },
+  "tests": [
+    {
+      "test_name": "serving_llama8B_tp1_sharegpt",
+      "server_parameters": {
+        "tensor_parallel_size": 1
+      },
+      "client_parameters": {
+        "dataset_name": "sharegpt",
+        "dataset_path": "./ShareGPT_V3_unfiltered_cleaned_split.json"
+      }
+    },
+    {
+        "dataset_name": "sharegpt",
+        "dataset_path": "./ShareGPT_V3_unfiltered_cleaned_split.json"
+      }
+    },
+    {
+      "test_name": "serving_llama8B_tp1_random_128_128",
+      "server_parameters": {
+        "tensor_parallel_size": 1
+      },
+      "client_parameters": {
+        "dataset_name": "random",
+        "random-input-len": 128,
+        "random-output-len": 128
+      }
+    },
+    {
+      "test_name": "serving_llama8B_tp1_random_128_2048",
+      "server_parameters": {
+        "tensor_parallel_size": 1
+      },
+      "client_parameters": {
+        "dataset_name": "random",
+        "random-input-len": 128,
+        "random-output-len": 2048
+      }
+    },
+    {
+      "test_name": "serving_llama8B_tp1_random_2048_128",
+      "server_parameters": {
+        "tensor_parallel_size": 1
+      },
+      "client_parameters": {
+        "dataset_name": "random",
+        "random-input-len": 2048,
+        "random-output-len": 128
+      }
+    },
+    {
+      "test_name": "serving_llama8B_tp1_random_2048_2048",
+      "server_parameters": {
+        "tensor_parallel_size": 1
+      },
+      "client_parameters": {
+        "dataset_name": "random",
+        "random-input-len": 2048,
+        "random-output-len": 2048
+      }
+    },
+    {
+      "test_name": "serving_llama70B_tp4_random_128_128",
+      "server_parameters": {
+        "model": "meta-llama/Llama-3.3-70B-Instruct",
+        "async_scheduling": "",
+	"no_enable_prefix_caching": "",
+        "max_num_batched_tokens": 8192
+      },
+      "client_parameters": {
+        "model": "meta-llama/Llama-3.3-70B-Instruct",
+        "dataset_name": "random",
+        "random-input-len": 128,
+        "random-output-len": 128
+      }
+    },
+    {
+      "test_name": "serving_gemma4-e4b_tp1_random_128_128",
+      "server_parameters": {
+        "model": "google/gemma-4-E4B-it",
+	"enable_auto_tool_choice": "",
+	"tool_call_parser": "gemma4",
+	"chat_template": "examples/tool_chat_template_gemma4.jinja",
+	"reasoning_parser": "gemma4"
+      },
+      "client_parameters": {
+        "model": "google/gemma-4-E4B-it",
+        "dataset_name": "random",
+        "random-input-len": 128,
+        "random-output-len": 128
+      }
+    }
+  ]
+}

.buildkite/release-pipeline.yaml

@@ -7,6 +7,9 @@ env:
   CUDA_ARCH_AARCH64: "8.0 8.7 8.9 9.0 10.0 11.0 12.0+PTX"
   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"
+  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"
 
 steps:
   - input: "Provide Release version here"
@@ -24,7 +27,7 @@ steps:
         agents:
           queue: arm64_cpu_queue_release
         commands:
-          - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=12.9.1 --build-arg torch_cuda_arch_list=\"${CUDA_ARCH_AARCH64_CU129}\" --tag vllm-ci:build-image --target build --progress plain -f docker/Dockerfile ."
+          - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=12.9.1 --build-arg torch_cuda_arch_list=\"${CUDA_ARCH_AARCH64_CU129}\" --build-arg BUILD_OS=manylinux --build-arg BUILD_BASE_IMAGE=pytorch/manylinuxaarch64-builder:cuda12.9 --tag vllm-ci:build-image --target build --progress plain -f docker/Dockerfile ."
           - "mkdir artifacts"
           - "docker run --rm -v $(pwd)/artifacts:/artifacts_host vllm-ci:build-image bash -c 'cp -r dist /artifacts_host && chmod -R a+rw /artifacts_host'"
           - "bash .buildkite/scripts/upload-nightly-wheels.sh"
@@ -66,7 +69,7 @@ steps:
         agents:
           queue: cpu_queue_release
         commands:
-          - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=12.9.1 --build-arg torch_cuda_arch_list=\"${CUDA_ARCH_X86_CU129}\" --tag vllm-ci:build-image --target build --progress plain -f docker/Dockerfile ."
+          - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=12.9.1 --build-arg torch_cuda_arch_list=\"${CUDA_ARCH_X86_CU129}\" --build-arg BUILD_OS=manylinux --build-arg BUILD_BASE_IMAGE=pytorch/manylinux2_28-builder:cuda12.9 --tag vllm-ci:build-image --target build --progress plain -f docker/Dockerfile ."
           - "mkdir artifacts"
           - "docker run --rm -v $(pwd)/artifacts:/artifacts_host vllm-ci:build-image bash -c 'cp -r dist /artifacts_host && chmod -R a+rw /artifacts_host'"
           - "bash .buildkite/scripts/upload-nightly-wheels.sh"
@@ -136,6 +139,8 @@ steps:
               --build-arg CUDA_VERSION=13.0.2 \
               --build-arg torch_cuda_arch_list="${CUDA_ARCH_X86}" \
               --build-arg INSTALL_KV_CONNECTORS=true \
+              --build-arg MOONCAKE_WHEEL_AARCH64="${MOONCAKE_WHEEL_AARCH64_2_35}" \
+              --build-arg MOONCAKE_WHEEL_X86_64="${MOONCAKE_WHEEL_X86_64}" \
               --build-arg BUILD_BASE_IMAGE=nvidia/cuda:13.0.2-devel-ubuntu22.04 \
               --target vllm-openai \
               --progress plain \
@@ -162,6 +167,8 @@ steps:
               --build-arg CUDA_VERSION=13.0.2 \
               --build-arg torch_cuda_arch_list="${CUDA_ARCH_AARCH64}" \
               --build-arg INSTALL_KV_CONNECTORS=true \
+              --build-arg MOONCAKE_WHEEL_AARCH64="${MOONCAKE_WHEEL_AARCH64_2_35}" \
+              --build-arg MOONCAKE_WHEEL_X86_64="${MOONCAKE_WHEEL_X86_64}" \
               --build-arg BUILD_BASE_IMAGE=nvidia/cuda:13.0.2-devel-ubuntu22.04 \
               --target vllm-openai \
               --progress plain \
@@ -185,6 +192,8 @@ steps:
               --build-arg CUDA_VERSION=12.9.1 \
               --build-arg torch_cuda_arch_list="${CUDA_ARCH_X86_CU129}" \
               --build-arg INSTALL_KV_CONNECTORS=true \
+              --build-arg MOONCAKE_WHEEL_AARCH64="${MOONCAKE_WHEEL_AARCH64_2_35}" \
+              --build-arg MOONCAKE_WHEEL_X86_64="${MOONCAKE_WHEEL_X86_64}" \
               --target vllm-openai \
               --progress plain \
               -f docker/Dockerfile .
@@ -210,6 +219,8 @@ steps:
               --build-arg CUDA_VERSION=12.9.1 \
               --build-arg torch_cuda_arch_list="${CUDA_ARCH_AARCH64_CU129}" \
               --build-arg INSTALL_KV_CONNECTORS=true \
+              --build-arg MOONCAKE_WHEEL_AARCH64="${MOONCAKE_WHEEL_AARCH64_2_35}" \
+              --build-arg MOONCAKE_WHEEL_X86_64="${MOONCAKE_WHEEL_X86_64}" \
               --target vllm-openai \
               --progress plain \
               -f docker/Dockerfile .
@@ -234,6 +245,8 @@ steps:
               --build-arg GDRCOPY_OS_VERSION=Ubuntu24_04 \
               --build-arg torch_cuda_arch_list="${CUDA_ARCH_X86}" \
               --build-arg INSTALL_KV_CONNECTORS=true \
+              --build-arg MOONCAKE_WHEEL_AARCH64="${MOONCAKE_WHEEL_AARCH64_2_39}" \
+              --build-arg MOONCAKE_WHEEL_X86_64="${MOONCAKE_WHEEL_X86_64}" \
               --build-arg BUILD_BASE_IMAGE=nvidia/cuda:13.0.2-devel-ubuntu24.04 \
               --target vllm-openai \
               --progress plain \
@@ -261,6 +274,8 @@ steps:
               --build-arg GDRCOPY_OS_VERSION=Ubuntu24_04 \
               --build-arg torch_cuda_arch_list="${CUDA_ARCH_AARCH64}" \
               --build-arg INSTALL_KV_CONNECTORS=true \
+              --build-arg MOONCAKE_WHEEL_AARCH64="${MOONCAKE_WHEEL_AARCH64_2_39}" \
+              --build-arg MOONCAKE_WHEEL_X86_64="${MOONCAKE_WHEEL_X86_64}" \
               --build-arg BUILD_BASE_IMAGE=nvidia/cuda:13.0.2-devel-ubuntu24.04 \
               --target vllm-openai \
               --progress plain \
@@ -286,6 +301,8 @@ steps:
               --build-arg GDRCOPY_OS_VERSION=Ubuntu24_04 \
               --build-arg torch_cuda_arch_list="${CUDA_ARCH_X86_CU129}" \
               --build-arg INSTALL_KV_CONNECTORS=true \
+              --build-arg MOONCAKE_WHEEL_AARCH64="${MOONCAKE_WHEEL_AARCH64_2_39}" \
+              --build-arg MOONCAKE_WHEEL_X86_64="${MOONCAKE_WHEEL_X86_64}" \
               --target vllm-openai \
               --progress plain \
               -f docker/Dockerfile .
@@ -312,6 +329,8 @@ steps:
               --build-arg GDRCOPY_OS_VERSION=Ubuntu24_04 \
               --build-arg torch_cuda_arch_list="${CUDA_ARCH_AARCH64_CU129}" \
               --build-arg INSTALL_KV_CONNECTORS=true \
+              --build-arg MOONCAKE_WHEEL_AARCH64="${MOONCAKE_WHEEL_AARCH64_2_39}" \
+              --build-arg MOONCAKE_WHEEL_X86_64="${MOONCAKE_WHEEL_X86_64}" \
               --target vllm-openai \
               --progress plain \
               -f docker/Dockerfile .
@@ -718,7 +737,7 @@ steps:
       - "bash tools/vllm-rocm/generate-rocm-wheels-root-index.sh"
     env:
       S3_BUCKET: "vllm-wheels"
-      VARIANT: "rocm722"
+      VARIANT: "rocm723"
 
   # ROCm Job 6: Build ROCm Release Docker Image
   - label: ":docker: Build release image - x86_64 - ROCm"

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

@@ -9,6 +9,13 @@
 # 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.
 
 set -euo pipefail
 
@@ -26,12 +33,12 @@ if [ $# -lt 1 ]; then usage; fi
 if [[ "$1" == https://* ]]; then
     BUILD=$(echo "$1" | sed -nE 's#.*/builds/([0-9]+).*#\1#p')
     JOB=$(echo "$1" | grep -oE '[0-9a-f]{8}-[0-9a-f-]+' | head -n 1)
-    OUT="${2:-ci-${BUILD}-${JOB:0:8}.log}"
+    OUT="${2:-}"
 else
     if [ $# -lt 2 ]; then usage; fi
     BUILD="$1"
     JOB="$2"
-    OUT="${3:-ci-${BUILD}-${JOB:0:8}.log}"
+    OUT="${3:-}"
 fi
 
 if [ -z "$BUILD" ] || [ -z "$JOB" ]; then
@@ -39,6 +46,18 @@ if [ -z "$BUILD" ] || [ -z "$JOB" ]; then
     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 $OUT (set CI_FETCH_LOG_FORCE=1 or pass an explicit output path)." >&2
+    exit 1
+fi
+
 COOKIES=$(mktemp)
 trap 'rm -f "$COOKIES"' EXIT
 

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

@@ -35,25 +35,9 @@ export PYTHONPATH=".."
 # Helper Functions
 ###############################################################################
 
-cleanup_docker() {
-  # Get Docker's root directory
-  docker_root=$(docker info -f '{{.DockerRootDir}}')
-  if [ -z "$docker_root" ]; then
-    echo "Failed to determine Docker root directory."
-    exit 1
-  fi
-  echo "Docker root directory: $docker_root"
-
-  disk_usage=$(df "$docker_root" | tail -1 | awk '{print $5}' | sed 's/%//')
-  threshold=70
-  if [ "$disk_usage" -gt "$threshold" ]; then
-    echo "Disk usage is above $threshold%. Cleaning up Docker images and volumes..."
-    docker image prune -f
-    docker volume prune -f && docker system prune --force --filter "until=72h" --all
-    echo "Docker images and volumes cleanup completed."
-  else
-    echo "Disk usage is below $threshold%. No cleanup needed."
-  fi
+report_docker_usage() {
+  echo "--- Docker usage"
+  docker system df || true
 }
 
 cleanup_network() {
@@ -114,8 +98,7 @@ handle_pytest_exit() {
 # unquoted since they have no spaces and work fine.
 #
 # Already-quoted expressions (containing literal single quotes) are passed
-# through untouched to avoid double-quoting values injected by
-# apply_rocm_test_overrides.
+# through untouched to avoid double-quoting well-formed shell fragments.
 #
 # NOTE: This ONLY fixes -m/-k flags. It cannot recover arbitrary inner
 # double-quotes stripped by the calling shell (see header comment).
@@ -247,102 +230,6 @@ re_quote_pytest_markers() {
   echo "${output% }"
 }
 
-###############################################################################
-# ROCm-specific pytest command rewrites
-#
-# These apply ignore flags and environment overrides for tests that are not
-# yet supported or behave differently on ROCm hardware. Kept as a single
-# function so new exclusions are easy to add in one place.
-###############################################################################
-
-apply_rocm_test_overrides() {
-  local cmds="$1"
-
-  # --- Model registry filter ---
-  if [[ $cmds == *"pytest -v -s models/test_registry.py"* ]]; then
-    cmds=${cmds//"pytest -v -s models/test_registry.py"/"pytest -v -s models/test_registry.py -k 'not BambaForCausalLM and not GritLM and not Mamba2ForCausalLM and not Zamba2ForCausalLM'"}
-  fi
-
-  # --- LoRA: disable custom paged attention ---
-  if [[ $cmds == *"pytest -v -s lora"* ]]; then
-    cmds=${cmds//"pytest -v -s lora"/"pytest -v -s lora"}
-  fi
-
-  # --- Kernel ignores ---
-  if [[ $cmds == *" kernels/core"* ]]; then
-    cmds="${cmds} \
-    --ignore=kernels/core/test_fused_quant_layernorm.py \
-    --ignore=kernels/core/test_permute_cols.py"
-  fi
-
-  if [[ $cmds == *" kernels/attention"* ]]; then
-    cmds="${cmds} \
-    --ignore=kernels/attention/test_attention_selector.py \
-    --ignore=kernels/attention/test_encoder_decoder_attn.py \
-    --ignore=kernels/attention/test_flash_attn.py \
-    --ignore=kernels/attention/test_flashinfer.py \
-    --ignore=kernels/attention/test_prefix_prefill.py \
-    --ignore=kernels/attention/test_cascade_flash_attn.py \
-    --ignore=kernels/attention/test_mha_attn.py \
-    --ignore=kernels/attention/test_lightning_attn.py \
-    --ignore=kernels/attention/test_attention.py"
-  fi
-
-  if [[ $cmds == *" kernels/quantization"* ]]; then
-    cmds="${cmds} \
-    --ignore=kernels/quantization/test_int8_quant.py \
-    --ignore=kernels/quantization/test_machete_mm.py \
-    --ignore=kernels/quantization/test_block_fp8.py \
-    --ignore=kernels/quantization/test_block_int8.py \
-    --ignore=kernels/quantization/test_marlin_gemm.py \
-    --ignore=kernels/quantization/test_cutlass_scaled_mm.py \
-    --ignore=kernels/quantization/test_int8_kernel.py"
-  fi
-
-  if [[ $cmds == *" kernels/mamba"* ]]; then
-    cmds="${cmds} \
-    --ignore=kernels/mamba/test_mamba_mixer2.py \
-    --ignore=kernels/mamba/test_causal_conv1d.py \
-    --ignore=kernels/mamba/test_mamba_ssm_ssd.py"
-  fi
-
-  if [[ $cmds == *" kernels/moe"* ]]; then
-    cmds="${cmds} \
-    --ignore=kernels/moe/test_moe.py \
-    --ignore=kernels/moe/test_cutlass_moe.py"
-  fi
-
-  # --- Entrypoint ignores ---
-  if [[ $cmds == *" entrypoints/openai "* ]]; then
-    cmds=${cmds//" entrypoints/openai "/" entrypoints/openai \
-    --ignore=entrypoints/openai/chat_completion/test_audio.py \
-    --ignore=entrypoints/openai/completion/test_shutdown.py \
-    --ignore=entrypoints/openai/test_completion.py \
-    --ignore=entrypoints/openai/models/test_models.py \
-    --ignore=entrypoints/openai/test_return_tokens_as_ids.py \
-    --ignore=entrypoints/openai/chat_completion/test_root_path.py \
-    --ignore=entrypoints/openai/completion/test_prompt_validation.py "}
-  fi
-
-  if [[ $cmds == *" entrypoints/serve"* ]]; then
-    cmds="${cmds} \
-    --ignore=entrypoints/serve/lora/test_lora_adapters.py"
-  fi
-
-  if [[ $cmds == *" entrypoints/llm "* ]]; then
-    cmds=${cmds//" entrypoints/llm "/" entrypoints/llm \
-    --ignore=entrypoints/llm/test_chat.py \
-    --ignore=entrypoints/llm/test_accuracy.py \
-    --ignore=entrypoints/llm/test_init.py \
-    --ignore=entrypoints/llm/test_prompt_validation.py "}
-  fi
-
-  # Clean up escaped newlines from --ignore appends
-  cmds=$(echo "$cmds" | sed 's/ \\ / /g')
-
-  echo "$cmds"
-}
-
 ###############################################################################
 # Main
 ###############################################################################
@@ -351,8 +238,8 @@ apply_rocm_test_overrides() {
 echo "--- ROCm info"
 rocminfo
 
-# --- Docker housekeeping ---
-cleanup_docker
+# --- Docker status ---
+report_docker_usage
 
 # --- Pull test image ---
 echo "--- Pulling container"
@@ -361,9 +248,17 @@ container_name="rocm_${BUILDKITE_COMMIT}_$(tr -dc A-Za-z0-9 < /dev/urandom | hea
 docker pull "${image_name}"
 
 remove_docker_container() {
-  docker rm -f "${container_name}" || docker image rm -f "${image_name}" || true
+  # 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 "${container_name}" >/dev/null 2>&1 || true
 }
-trap remove_docker_container EXIT
+
+on_exit() {
+  local exit_code=$?
+  remove_docker_container
+  exit "$exit_code"
+}
+trap on_exit EXIT
 
 # --- Prepare commands ---
 echo "--- Running container"
@@ -409,7 +304,6 @@ else
   echo "Skipping re-quoting for VLLM_TEST_COMMANDS input"
 fi
 
-commands=$(apply_rocm_test_overrides "$commands")
 echo "Final commands: $commands"
 
 MYPYTHONPATH=".."

.buildkite/scripts/hardware_ci/run-cpu-test-arm.sh

@@ -37,7 +37,8 @@ function cpu_tests() {
     pytest -x -v -s tests/kernels/test_onednn.py
     pytest -x -v -s tests/kernels/attention/test_cpu_attn.py
     pytest -x -v -s tests/kernels/core/test_cpu_activation.py
-    pytest -x -v -s tests/kernels/moe/test_moe.py -k test_cpu_fused_moe_basic"
+    pytest -x -v -s tests/kernels/moe/test_moe.py -k test_cpu_fused_moe_basic
+    pytest -x -v -s tests/kernels/mamba/cpu/test_cpu_gdn_ops.py"
 
   # skip tests requiring model downloads if HF_TOKEN is not set
   # due to rate-limits

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

@@ -352,17 +352,31 @@ if [[ -z "${ZE_AFFINITY_MASK:-}" ]]; then
   echo "Warning: ZE_AFFINITY_MASK is not set. Proceeding without device affinity." >&2
 fi
 
-docker run \
+export CMDS="${commands}"
+export HF_TOKEN ZE_AFFINITY_MASK
+
+{
+  flock 9
+  if ! docker image inspect "${IMAGE}" >/dev/null 2>&1; then
+    echo 'Image missing before container creation, pulling again...'
+    timeout 900 docker pull "${IMAGE}"
+  fi
+
+  docker create \
     --device /dev/dri:/dev/dri \
     --net=host \
     --ipc=host \
     --privileged \
     -v /dev/dri/by-path:/dev/dri/by-path \
-    -v ${HOME}/.cache/huggingface:/root/.cache/huggingface \
-    --entrypoint="" \
-    -e "HF_TOKEN=${HF_TOKEN:-}" \
-    -e "ZE_AFFINITY_MASK=${ZE_AFFINITY_MASK:-}" \
-    -e "CMDS=${commands}" \
+    -v "${HOME}/.cache/huggingface:/root/.cache/huggingface" \
+    --entrypoint='' \
+    -e HF_TOKEN \
+    -e ZE_AFFINITY_MASK \
+    -e CMDS \
     --name "${container_name}" \
-    "${image_name}" \
-    bash -c 'set -e; echo "ZE_AFFINITY_MASK is ${ZE_AFFINITY_MASK:-}"; eval "$CMDS"'
+    "${IMAGE}" \
+    bash -c 'set -e; echo "ZE_AFFINITY_MASK is ${ZE_AFFINITY_MASK:-}"; eval "$CMDS"' \
+    >/dev/null
+} 9>/tmp/docker-pull.lock
+
+docker start -a "${container_name}"

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

@@ -1,53 +0,0 @@
-#!/bin/bash
-
-# This script build the CPU docker image and run the offline inference inside the container.
-# It serves a sanity check for compilation and basic model usage.
-set -ex
-
-image_name="xpu/vllm-ci:${BUILDKITE_COMMIT}"
-container_name="xpu_${BUILDKITE_COMMIT}_$(tr -dc A-Za-z0-9 < /dev/urandom | head -c 10; echo)"
-
-# Try building the docker image
-docker build -t "${image_name}" -f docker/Dockerfile.xpu .
-
-# Setup cleanup
-remove_docker_container() {
-  docker rm -f "${container_name}" || true
-}
-trap remove_docker_container EXIT
-
-# Run the image and test offline inference/tensor parallel
-docker run \
-    --device /dev/dri:/dev/dri \
-    --net=host \
-    --ipc=host \
-    --privileged \
-    -v /dev/dri/by-path:/dev/dri/by-path \
-    --entrypoint="" \
-    -e "HF_TOKEN=${HF_TOKEN}" \
-    -e "ZE_AFFINITY_MASK=${ZE_AFFINITY_MASK}" \
-    --name "${container_name}" \
-    "${image_name}" \
-    bash -c '
-    set -e
-    echo $ZE_AFFINITY_MASK
-    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 ray
-    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 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 OPEA/Qwen2.5-0.5B-Instruct-int4-sym-inc --block-size 64 --enforce-eager --max-model-len 8192
-    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
-    pytest -v -s v1/sample --ignore=v1/sample/test_logprobs.py --ignore=v1/sample/test_logprobs_e2e.py
-    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/spec_decode --ignore=v1/spec_decode/test_max_len.py --ignore=v1/spec_decode/test_tree_attention.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
-    pytest -v -s v1/test_serial_utils.py
-'

.buildkite/scripts/install-kv-connectors.sh

@@ -0,0 +1,39 @@
+#!/bin/bash
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+set -euo pipefail
+
+REQUIREMENTS_FILE="${KV_CONNECTORS_REQUIREMENTS:-/vllm-workspace/requirements/kv_connectors.txt}"
+
+uv pip install --system -r "${REQUIREMENTS_FILE}"
+
+NIXL_METADATA=$(python3 - <<'PY'
+import importlib.metadata as metadata
+
+import torch
+
+cuda_version = torch.version.cuda
+if cuda_version is None:
+    raise SystemExit("torch.version.cuda is not set")
+
+print(cuda_version.split(".", 1)[0], metadata.version("nixl"))
+PY
+)
+read -r CUDA_MAJOR NIXL_VERSION <<<"${NIXL_METADATA}"
+
+# nixl>=1.1.0 can install multiple CUDA wheel variants. Keep only the variant
+# matching this CI image so nixl_ep_cpp links against the available libcudart.
+uv pip uninstall --system nixl-cu12 nixl-cu13 2>/dev/null || true
+uv pip install --system --no-deps "nixl-cu${CUDA_MAJOR}==${NIXL_VERSION}"
+
+python3 - <<'PY'
+import importlib.metadata as metadata
+
+for package_name in ("nixl", "nixl-cu12", "nixl-cu13"):
+    try:
+        version = metadata.version(package_name)
+    except metadata.PackageNotFoundError:
+        version = "not installed"
+    print(f"{package_name}: {version}")
+PY

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

@@ -0,0 +1,156 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+MODE="${1:-}"
+
+if [[ "$MODE" != "style-clippy" && "$MODE" != "test" ]]; then
+  echo "Usage: $0 {style-clippy|test}" >&2
+  exit 2
+fi
+
+ROOT_DIR="$(git rev-parse --show-toplevel)"
+cd "$ROOT_DIR"
+
+export CARGO_TERM_COLOR="${CARGO_TERM_COLOR:-always}"
+export CARGO_HOME="${CARGO_HOME:-$HOME/.cargo}"
+export RUSTUP_HOME="${RUSTUP_HOME:-$HOME/.rustup}"
+export PATH="$CARGO_HOME/bin:$PATH"
+
+log_section() {
+  echo "--- $*"
+}
+
+install_protoc() {
+  if command -v protoc >/dev/null 2>&1; then
+    return
+  fi
+
+  local version="${PROTOC_VERSION:-31.1}"
+  local arch
+  case "$(uname -m)" in
+    x86_64)
+      arch="x86_64"
+      ;;
+    aarch64|arm64)
+      arch="aarch_64"
+      ;;
+    *)
+      echo "Unsupported protoc architecture: $(uname -m)" >&2
+      return 1
+      ;;
+  esac
+
+  local url="https://github.com/protocolbuffers/protobuf/releases/download/v${version}/protoc-${version}-linux-${arch}.zip"
+  local tmp_dir
+  tmp_dir="$(mktemp -d)"
+
+  log_section "Installing protoc ${version}"
+  curl -L --proto '=https' --tlsv1.2 -sSf "$url" -o "$tmp_dir/protoc.zip"
+  mkdir -p "$CARGO_HOME/bin"
+  unzip -q "$tmp_dir/protoc.zip" bin/protoc 'include/*' -d "$CARGO_HOME"
+  chmod +x "$CARGO_HOME/bin/protoc"
+  rm -rf "$tmp_dir"
+}
+
+rust_toolchain() {
+  awk -F '"' '/channel[[:space:]]*=/ { print $2; exit }' rust-toolchain.toml
+}
+
+install_rust_toolchain() {
+  log_section "Installing Rust toolchain"
+  if ! command -v rustup >/dev/null 2>&1; then
+    curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs \
+      | sh -s -- -y --profile minimal --default-toolchain none
+  fi
+
+  local toolchain
+  toolchain="$(rust_toolchain)"
+  rustup toolchain install "$toolchain" --profile minimal --component rustfmt,clippy
+  rustup component add --toolchain "$toolchain" rustfmt clippy
+}
+
+install_cargo_binstall() {
+  if command -v cargo-binstall >/dev/null 2>&1; then
+    return
+  fi
+
+  log_section "Installing cargo-binstall"
+  curl -L --proto '=https' --tlsv1.2 -sSf \
+    https://raw.githubusercontent.com/cargo-bins/cargo-binstall/main/install-from-binstall-release.sh \
+    | bash
+}
+
+install_cargo_sort() {
+  if command -v cargo-sort >/dev/null 2>&1; then
+    return
+  fi
+
+  log_section "Installing cargo-sort"
+  install_cargo_binstall
+  cargo binstall --no-confirm cargo-sort
+}
+
+install_cargo_nextest() {
+  if command -v cargo-nextest >/dev/null 2>&1; then
+    return
+  fi
+
+  log_section "Installing cargo-nextest"
+  install_cargo_binstall
+  cargo binstall --no-confirm --secure cargo-nextest
+}
+
+install_uv() {
+  if command -v uv >/dev/null 2>&1; then
+    return
+  fi
+
+  log_section "Installing uv"
+  curl -LsSf --proto '=https' --tlsv1.2 https://astral.sh/uv/install.sh \
+    | env UV_INSTALL_DIR="$CARGO_HOME/bin" sh
+}
+
+run_style_clippy() {
+  install_cargo_sort
+
+  log_section "Checking Rust formatting"
+  cargo fmt --manifest-path rust/Cargo.toml --all -- --check
+
+  log_section "Checking Cargo.toml ordering"
+  cargo sort --workspace --check rust
+
+  log_section "Running clippy"
+  cargo clippy \
+    --manifest-path rust/Cargo.toml \
+    --workspace \
+    --all-targets \
+    --all-features \
+    --locked \
+    -- \
+    -D warnings
+}
+
+run_tests() {
+  install_uv
+  install_cargo_nextest
+
+  log_section "Running cargo nextest"
+  cargo nextest run \
+    --manifest-path rust/Cargo.toml \
+    --workspace \
+    --all-features \
+    --locked \
+    --no-fail-fast
+}
+
+install_protoc
+install_rust_toolchain
+
+case "$MODE" in
+  style-clippy)
+    run_style_clippy
+    ;;
+  test)
+    run_tests
+    ;;
+esac

.buildkite/scripts/scheduled_integration_test/deepseek_v2_lite_ep_eplb.sh

@@ -49,6 +49,7 @@ for BACK in "${BACKENDS[@]}"; do
     --data-parallel-size 2 \
     --enable-expert-parallel \
     --enable-eplb \
+    --eplb-config '{"use_async": false}' \
     --trust-remote-code \
     --max-model-len 2048 \
     --all2all-backend "$BACK" \

.buildkite/scripts/scheduled_integration_test/qwen30b_a3b_fp8_block_ep_eplb.sh

@@ -48,7 +48,7 @@ for BACK in "${BACKENDS[@]}"; do
     --enforce-eager \
     --enable-eplb \
     --all2all-backend "$BACK" \
-    --eplb-config '{"window_size":10, "step_interval":100, "num_redundant_experts":0, "log_balancedness":true}' \
+    --eplb-config '{"window_size":10, "step_interval":100, "num_redundant_experts":0, "log_balancedness":true, "use_async":false}' \
     --tensor-parallel-size "${TENSOR_PARALLEL_SIZE}" \
     --data-parallel-size "${DATA_PARALLEL_SIZE}" \
     --enable-expert-parallel \

.buildkite/scripts/tool_call/run-bfcl-eval.sh

@@ -70,7 +70,7 @@ echo "============================================"
 # ---- Install bfcl-eval if missing ----
 if ! python3 -c "import bfcl_eval" 2>/dev/null; then
     echo "Installing bfcl-eval..."
-    pip install "bfcl-eval>=2025.10.20.1,<2026"
+    uv pip install "bfcl-eval>=2025.10.20.1,<2026"
 fi
 
 # ---- Cleanup handler ----
@@ -100,7 +100,7 @@ SERVE_ARGS=(
     --tensor-parallel-size "$TP_SIZE"
     --max-model-len "$MAX_MODEL_LEN"
     --enforce-eager
-    --no-enable-prefix-caching
+    --enable-prefix-caching
 )
 
 # Append reasoning parser if specified

.buildkite/test-amd.yaml

@@ -139,19 +139,6 @@ steps:
   - pytest models/multimodal -v -s -m 'distributed(num_gpus=2)' --ignore models/multimodal/generation/test_whisper.py
   - VLLM_WORKER_MULTIPROC_METHOD=spawn pytest models/multimodal/generation/test_whisper.py -v -s -m 'distributed(num_gpus=2)'
 
-#--------------------------------------------------------  mi250 · benchmarks  ---------------------------------------------------------#
-
-- label: Benchmarks # TBD
-  timeout_in_minutes: 180
-  mirror_hardwares: [amdexperimental, amdproduction, amdgfx90anightly, amdmi250]
-  agent_pool: mi250_1
-  working_dir: "/vllm-workspace/.buildkite"
-  source_file_dependencies:
-  - benchmarks/
-  - vllm/platforms/rocm.py
-  commands:
-  - bash scripts/run-benchmarks.sh
-
 #----------------------------------------------------------  mi250 · compile  ----------------------------------------------------------#
 
 - label: PyTorch Compilation Unit Tests # TBD
@@ -485,7 +472,7 @@ steps:
   - pytest -v -s model_executor -m '(not slow_test)'
   - pytest -v -s entrypoints/openai/completion/test_tensorizer_entrypoint.py
 
-#----------------------------------------------------------  mi250 · models  -----------------------------------------------------------#
+#------------------------------------------------------  mi250 · models / basic  -------------------------------------------------------#
 
 - label: Basic Models Test (Other CPU) # TBD
   timeout_in_minutes: 180
@@ -546,6 +533,8 @@ steps:
   commands:
   - pytest -v -s models/test_terratorch.py models/test_transformers.py models/test_registry.py
 
+#-----------------------------------------------------  mi250 · models / language  -----------------------------------------------------#
+
 - label: Language Models Test (MTEB) # TBD
   timeout_in_minutes: 180
   mirror_hardwares: [amdexperimental, amdproduction, amdgfx90anightly, amdmi250]
@@ -590,6 +579,8 @@ steps:
   - pip freeze | grep -E 'torch'
   - pytest -v -s models/language -m 'core_model and slow_test' --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --shard-id=$$BUILDKITE_PARALLEL_JOB
 
+#----------------------------------------------------  mi250 · models / multimodal  ----------------------------------------------------#
+
 - label: Multi-Modal Models (Extended Generation 2) # TBD
   timeout_in_minutes: 180
   mirror_hardwares: [amdexperimental, amdproduction, amdgfx90anightly, amdmi250]
@@ -976,18 +967,6 @@ steps:
 
 #--------------------------------------------------------  mi300 · benchmarks  ---------------------------------------------------------#
 
-- label: Benchmarks # TBD
-  timeout_in_minutes: 180
-  mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
-  agent_pool: mi300_1
-  optional: true
-  working_dir: "/vllm-workspace/.buildkite"
-  source_file_dependencies:
-  - benchmarks/
-  - vllm/platforms/rocm.py
-  commands:
-  - bash scripts/run-benchmarks.sh
-
 - label: Benchmarks CLI Test # TBD
   timeout_in_minutes: 180
   mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
@@ -1259,14 +1238,11 @@ steps:
   working_dir: "/vllm-workspace/tests"
   source_file_dependencies:
   - vllm/
-  - tests/entrypoints/rpc
-  - tests/entrypoints/serve/instrumentator
-  - tests/tool_use
+  - tests/entrypoints/serve
   commands:
   - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-  - pytest -v -s entrypoints/serve/instrumentator
-  - PYTHONPATH=/vllm-workspace pytest -v -s entrypoints/rpc
-  - pytest -v -s tool_use
+  - 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
   timeout_in_minutes: 180
@@ -1282,7 +1258,7 @@ 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_chat_with_tool_reasoning.py --ignore=entrypoints/openai/chat_completion/test_oot_registration.py
+  - pytest -v -s entrypoints/openai/chat_completion --ignore=entrypoints/openai/chat_completion/test_oot_registration.py
 
 - label: Entrypoints Integration (API Server openai - Part 2) # TBD
   timeout_in_minutes: 180
@@ -1296,10 +1272,14 @@ 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/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
   timeout_in_minutes: 180
@@ -1389,7 +1369,7 @@ steps:
   - vllm/platforms/rocm.py
   commands:
   - pytest -v -s entrypoints/openai/tool_parsers
-  - pytest -v -s entrypoints/ --ignore=entrypoints/llm --ignore=entrypoints/rpc --ignore=entrypoints/sleep --ignore=entrypoints/serve/instrumentator --ignore=entrypoints/openai --ignore=entrypoints/offline_mode --ignore=entrypoints/test_chat_utils.py  --ignore=entrypoints/pooling
+  - 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
 
 - label: OpenAI API correctness # TBD
   timeout_in_minutes: 180
@@ -1505,7 +1485,7 @@ steps:
   commands:
   - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-mi3xx-fp8-and-mixed.txt
 
-- label: DeepSeek V2-Lite Accuracy (4xH100-4xMI300) # TBD
+- label: DeepSeek V2-Lite Sync EPLB Accuracy (4xH100-4xMI300) # TBD
   timeout_in_minutes: 180
   mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
   agent_pool: mi300_4
@@ -1547,7 +1527,7 @@ steps:
   - export VLLM_WORKER_MULTIPROC_METHOD=spawn
   - pytest -s -v test_lm_eval_correctness.py --config-list-file=configs/models-large.txt --tp-size=4
 
-- label: Qwen3-30B-A3B-FP8-block Accuracy (4xH100-4xMI300) # TBD
+- label: Qwen3-30B-A3B-FP8-block Sync EPLB Accuracy (4xH100-4xMI300) # TBD
   timeout_in_minutes: 180
   mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
   agent_pool: mi300_4
@@ -1759,7 +1739,7 @@ steps:
   - pytest -v -s -x lora/test_gptoss_tp.py
   - pytest -v -s -x lora/test_qwen35_densemodel_lora.py
 
-#----------------------------------------------------------  mi300 · models  -----------------------------------------------------------#
+#-----------------------------------------------------  mi300 · models / language  -----------------------------------------------------#
 
 - label: Language Models Test (Extended Pooling)  # TBD
   timeout_in_minutes: 180
@@ -1787,6 +1767,8 @@ steps:
   - pip freeze | grep -E 'torch'
   - pytest -v -s models/language -m 'core_model and (not slow_test)'
 
+#----------------------------------------------------  mi300 · models / multimodal  ----------------------------------------------------#
+
 - label: Multi-Modal Models (Extended Generation 1) # TBD
   timeout_in_minutes: 180
   mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
@@ -1892,10 +1874,11 @@ steps:
   - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
   - pytest -v -s models/multimodal/processing/test_tensor_schema.py
 
-- label: Multi-Modal Processor (CPU) # TBD
+- label: Multi-Modal Processor (CPU) %N # TBD
   timeout_in_minutes: 180
   mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
   agent_pool: mi300_1
+  parallelism: 4
   no_gpu: true
   optional: true
   working_dir: "/vllm-workspace/tests"
@@ -1905,7 +1888,9 @@ steps:
   - tests/models/registry.py
   commands:
   - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
-  - pytest -v -s models/multimodal/processing --ignore models/multimodal/processing/test_tensor_schema.py
+  - pytest -v -s models/multimodal/processing --ignore models/multimodal/processing/test_tensor_schema.py --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --shard-id=$$BUILDKITE_PARALLEL_JOB
+
+#-----------------------------------------------------  mi300 · models / quantized  -----------------------------------------------------#
 
 - label: Quantized Models Test # TBD
   timeout_in_minutes: 180
@@ -1921,7 +1906,31 @@ steps:
   commands:
   - pytest -v -s models/quantization
 
-- label: Transformers Nightly Models # TBD
+#--------------------------------------------------  mi300 · models / transformers  ---------------------------------------------------#
+
+- label: Transformers Nightly Models (Shardable) %N # TBD
+  timeout_in_minutes: 180
+  mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
+  agent_pool: mi300_1
+  parallelism: 4
+  optional: true
+  working_dir: "/vllm-workspace/"
+  source_file_dependencies:
+  - vllm/model_executor/models/
+  - vllm/model_executor/model_loader/
+  - vllm/multimodal/
+  - vllm/model_executor/layers/
+  - vllm/v1/attention/backends/
+  - vllm/v1/attention/selector.py
+  - vllm/_aiter_ops.py
+  - vllm/platforms/rocm.py
+  - tests/models/
+  commands:
+  - pip install --upgrade git+https://github.com/huggingface/transformers
+  - pytest -v -s tests/models/test_initialization.py --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --shard-id=$$BUILDKITE_PARALLEL_JOB
+  - pytest -v -s tests/models/multimodal/processing/ --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --shard-id=$$BUILDKITE_PARALLEL_JOB
+
+- label: Transformers Nightly Models (Single) # TBD
   timeout_in_minutes: 180
   mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
   agent_pool: mi300_1
@@ -1940,9 +1949,7 @@ steps:
   - examples/
   commands:
   - pip install --upgrade git+https://github.com/huggingface/transformers
-  - pytest -v -s tests/models/test_initialization.py
   - pytest -v -s tests/models/test_transformers.py
-  - pytest -v -s tests/models/multimodal/processing/
   - pytest -v -s tests/models/multimodal/test_mapping.py
   - python3 examples/basic/offline_inference/chat.py
   - python3 examples/generate/multimodal/vision_language_offline.py --model-type qwen2_5_vl
@@ -2003,7 +2010,7 @@ steps:
   - vllm/model_executor/layers
   - vllm/sampling_metadata.py
   - vllm/v1/sample/
-  - vllm/beam_search.py
+  - vllm/entrypoints/generate/beam_search/
   - tests/samplers
   - tests/conftest.py
   - vllm/_aiter_ops.py
@@ -2391,7 +2398,7 @@ steps:
   - 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
 
-- label: Hyrbid SSM NixlConnector PD accuracy tests (4 GPUs) # TBD
+- label: Hybrid SSM NixlConnector PD accuracy tests (4 GPUs) # TBD
   timeout_in_minutes: 180
   mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
   agent_pool: mi300_4
@@ -2593,7 +2600,7 @@ steps:
   - export VLLM_WORKER_MULTIPROC_METHOD=spawn
   - pytest -s -v test_lm_eval_correctness.py --config-list-file=configs/models-large-rocm.txt --tp-size=8
 
-#----------------------------------------------------------  mi325 · models  -----------------------------------------------------------#
+#-----------------------------------------------------  mi325 · models / language  -----------------------------------------------------#
 
 - label: Language Models Test (Extended Generation) # TBD
   timeout_in_minutes: 180
@@ -2624,6 +2631,8 @@ steps:
   - 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
 
+#----------------------------------------------------  mi325 · models / multimodal  ----------------------------------------------------#
+
 - label: Multi-Modal Models (Extended Pooling) # TBD
   timeout_in_minutes: 180
   mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi325]
@@ -2695,19 +2704,35 @@ steps:
   optional: true
   working_dir: "/vllm-workspace/"
   source_file_dependencies:
+  - csrc/custom_quickreduce.cu
+  - csrc/ops.h
+  - csrc/torch_bindings.cpp
   - vllm/distributed/
-  - vllm/v1/distributed/
+  - vllm/model_executor/layers/
+  - vllm/entrypoints/llm.py
+  - vllm/config/parallel.py
   - vllm/model_executor/layers/fused_moe/
+  - vllm/v1/engine/
+  - vllm/v1/executor/
+  - vllm/v1/worker/
+  - vllm/v1/distributed/
   - vllm/v1/attention/backends/
   - vllm/v1/attention/selector.py
-  - tests/distributed/test_context_parallel.py
-  - tests/v1/distributed/test_dbo.py
-  - examples/features/data_parallel/data_parallel_offline.py
   - vllm/_aiter_ops.py
+  - vllm/_custom_ops.py
   - vllm/platforms/rocm.py
+  - vllm/envs.py
+  - examples/offline_inference/data_parallel.py
+  - tests/distributed/test_context_parallel.py
+  - tests/distributed/test_rocm_quick_reduce.py
+  - tests/distributed/test_quick_all_reduce.py
+  - tests/v1/distributed/test_dbo.py
+  - tests/utils.py
   commands:
   - pytest -v -s tests/distributed/test_context_parallel.py
   - pytest -v -s tests/v1/distributed/test_dbo.py
+  - pytest -v -s tests/distributed/test_rocm_quick_reduce.py
+  - pytest -v -s tests/distributed/test_quick_all_reduce.py
 
 #--------------------------------------------------------  mi355 · entrypoints  --------------------------------------------------------#
 
@@ -2721,14 +2746,11 @@ steps:
   working_dir: "/vllm-workspace/tests"
   source_file_dependencies:
   - vllm/
-  - tests/entrypoints/rpc
-  - tests/entrypoints/serve/instrumentator
-  - tests/tool_use
+  - tests/entrypoints/serve
   commands:
   - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-  - pytest -v -s entrypoints/serve/instrumentator
-  - PYTHONPATH=/vllm-workspace pytest -v -s entrypoints/rpc
-  - pytest -v -s tool_use
+  - 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
   timeout_in_minutes: 180
@@ -2744,7 +2766,7 @@ 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_chat_with_tool_reasoning.py --ignore=entrypoints/openai/chat_completion/test_oot_registration.py
+  - pytest -v -s entrypoints/openai/chat_completion --ignore=entrypoints/openai/chat_completion/test_oot_registration.py
 
 - label: Entrypoints Integration (API Server openai - Part 2) # TBD
   timeout_in_minutes: 180
@@ -2758,10 +2780,14 @@ 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/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
   timeout_in_minutes: 180
@@ -2871,7 +2897,7 @@ steps:
   commands:
   - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-mi3xx-fp8-and-mixed.txt
 
-- label: Qwen3-30B-A3B-FP8-block Accuracy (B200-MI355) # TBD
+- label: Qwen3-30B-A3B-FP8-block Sync EPLB Accuracy (B200-MI355) # TBD
   timeout_in_minutes: 180
   mirror_hardwares: [amdexperimental, amdproduction, amdgfx950nightly, amdmi355]
   agent_pool: mi355_2
@@ -3043,7 +3069,7 @@ steps:
   commands:
     - pytest -v -s kernels/moe/test_deepep_moe.py
 
-#----------------------------------------------------------  mi355 · models  -----------------------------------------------------------#
+#-----------------------------------------------------  mi355 · models / language  -----------------------------------------------------#
 
 - label: Language Models Test (Extended Generation) # TBD
   timeout_in_minutes: 180
@@ -3111,6 +3137,8 @@ steps:
   - pip freeze | grep -E 'torch'
   - pytest -v -s models/language -m 'core_model and (not slow_test)'
 
+#----------------------------------------------------  mi355 · models / multimodal  ----------------------------------------------------#
+
 - label: Multi-Modal Models (Extended Generation 1) # TBD
   timeout_in_minutes: 180
   mirror_hardwares: [amdexperimental, amdproduction, amdgfx950nightly, amdmi355]
@@ -3182,6 +3210,8 @@ steps:
   - pytest -v -s 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
 
+#-----------------------------------------------------  mi355 · models / quantized  -----------------------------------------------------#
+
 - label: Quantized Models Test # TBD
   timeout_in_minutes: 180
   mirror_hardwares: [amdexperimental, amdproduction, amdgfx950nightly, amdmi355]

.buildkite/test_areas/cuda.yaml

@@ -27,4 +27,5 @@ steps:
   - vllm/compilation
   commands:
     - pytest -v -s v1/cudagraph/test_cudagraph_dispatch.py
-    - pytest -v -s v1/cudagraph/test_cudagraph_mode.py
+    - pytest -v -s v1/cudagraph/test_cudagraph_mode.py
+    - pytest -v -s v1/cudagraph/test_breakable_cudagraph.py

.buildkite/test_areas/disaggregated.yaml

@@ -11,7 +11,7 @@ steps:
     - vllm/distributed/kv_transfer/kv_connector/v1/nixl/
     - tests/v1/kv_connector/nixl_integration/
   commands:
-    - uv pip install --system -r /vllm-workspace/requirements/kv_connectors.txt
+    - bash /vllm-workspace/.buildkite/scripts/install-kv-connectors.sh
     - bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
 - label: Distributed FlashInfer NixlConnector PD accuracy (4 GPUs)
   key: distributed-flashinfer-nixlconnector-pd-accuracy-4-gpus
@@ -22,7 +22,7 @@ steps:
     - vllm/distributed/kv_transfer/kv_connector/v1/nixl/
     - tests/v1/kv_connector/nixl_integration/
   commands:
-    - uv pip install --system -r /vllm-workspace/requirements/kv_connectors.txt
+    - bash /vllm-workspace/.buildkite/scripts/install-kv-connectors.sh
     - FLASHINFER=1 bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
 
 - label: DP EP Distributed NixlConnector PD accuracy tests (4 GPUs)
@@ -34,7 +34,7 @@ steps:
     - vllm/distributed/kv_transfer/kv_connector/v1/nixl/
     - tests/v1/kv_connector/nixl_integration/
   commands:
-    - uv pip install --system -r /vllm-workspace/requirements/kv_connectors.txt
+    - bash /vllm-workspace/.buildkite/scripts/install-kv-connectors.sh
     - DP_EP=1 bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
 
 - label: CrossLayer KV layout Distributed NixlConnector PD accuracy tests (4 GPUs)
@@ -46,7 +46,7 @@ steps:
     - vllm/distributed/kv_transfer/kv_connector/v1/nixl/
     - tests/v1/kv_connector/nixl_integration/
   commands:
-    - uv pip install --system -r /vllm-workspace/requirements/kv_connectors.txt
+    - bash /vllm-workspace/.buildkite/scripts/install-kv-connectors.sh
     - CROSS_LAYERS_BLOCKS=True bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
 
 - label: Hybrid SSM NixlConnector PD accuracy tests (4 GPUs)
@@ -58,7 +58,7 @@ steps:
     - vllm/distributed/kv_transfer/kv_connector/v1/nixl/
     - tests/v1/kv_connector/nixl_integration/
   commands:
-    - uv pip install --system -r /vllm-workspace/requirements/kv_connectors.txt
+    - bash /vllm-workspace/.buildkite/scripts/install-kv-connectors.sh
     - HYBRID_SSM=1 bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
 
 - label: MultiConnector (Nixl+Offloading) PD accuracy (2 GPUs)
@@ -73,7 +73,7 @@ steps:
     - vllm/distributed/kv_transfer/kv_connector/v1/offloading/
     - tests/v1/kv_connector/nixl_integration/
   commands:
-    - uv pip install --system -r /vllm-workspace/requirements/kv_connectors.txt
+    - bash /vllm-workspace/.buildkite/scripts/install-kv-connectors.sh
     - bash v1/kv_connector/nixl_integration/run_multi_connector_accuracy_test.sh
 
 - label: NixlConnector PD + Spec Decode acceptance (2 GPUs)
@@ -87,8 +87,8 @@ steps:
     - vllm/v1/worker/kv_connector_model_runner_mixin.py
     - tests/v1/kv_connector/nixl_integration/
   commands:
-    - uv pip install --system -r /vllm-workspace/requirements/kv_connectors.txt
-    - bash v1/kv_connector/nixl_integration/spec_decode_acceptance_test.sh
+    - bash /vllm-workspace/.buildkite/scripts/install-kv-connectors.sh
+    - bash v1/kv_connector/nixl_integration/config_sweep_spec_decode_test.sh
 
 - label: MultiConnector (Nixl+Offloading) PD edge cases (2 GPUs)
   key: multiconnector-nixl-offloading-pd-edge-cases-2-gpus
@@ -102,5 +102,5 @@ steps:
     - vllm/distributed/kv_transfer/kv_connector/v1/offloading/
     - tests/v1/kv_connector/nixl_integration/
   commands:
-    - uv pip install --system -r /vllm-workspace/requirements/kv_connectors.txt
-    - bash v1/kv_connector/nixl_integration/run_multi_connector_edge_case_test.sh
+    - bash /vllm-workspace/.buildkite/scripts/install-kv-connectors.sh
+    - bash v1/kv_connector/nixl_integration/run_multi_connector_edge_case_test.sh

.buildkite/test_areas/e2e_integration.yaml

@@ -2,8 +2,8 @@ group: E2E Integration
 depends_on: 
   - image-build
 steps:
-- label: DeepSeek V2-Lite Accuracy
-  key: deepseek-v2-lite-accuracy
+- label: DeepSeek V2-Lite Sync EPLB Accuracy
+  key: deepseek-v2-lite-sync-eplb-accuracy
   timeout_in_minutes: 60
   device: h100
   optional: true
@@ -12,8 +12,8 @@ steps:
   commands:
   - bash .buildkite/scripts/scheduled_integration_test/deepseek_v2_lite_ep_eplb.sh 0.25 200 8010
 
-- label: Qwen3-30B-A3B-FP8-block Accuracy
-  key: qwen3-30b-a3b-fp8-block-accuracy
+- label: Qwen3-30B-A3B-FP8-block Sync EPLB Accuracy
+  key: qwen3-30b-a3b-fp8-block-sync-eplb-accuracy
   timeout_in_minutes: 60
   device: h100
   optional: true
@@ -22,8 +22,8 @@ steps:
   commands:
   - bash .buildkite/scripts/scheduled_integration_test/qwen30b_a3b_fp8_block_ep_eplb.sh 0.8 200 8020
 
-- label: Qwen3-30B-A3B-FP8-block Accuracy (B200)
-  key: qwen3-30b-a3b-fp8-block-accuracy-b200
+- label: Qwen3-30B-A3B-FP8-block Sync EPLB Accuracy (B200)
+  key: qwen3-30b-a3b-fp8-block-sync-eplb-accuracy-b200
   timeout_in_minutes: 60
   device: b200-k8s
   optional: true

.buildkite/test_areas/engine.yaml

@@ -38,7 +38,7 @@ steps:
     - pytest -v -s v1/engine --ignore v1/engine/test_preprocess_error_handling.py
   mirror:
     amd:
-      device: mi300_1
+      device: mi325_1
       timeout_in_minutes: 40
       depends_on:
       - image-build-amd
@@ -60,6 +60,7 @@ steps:
       - image-build-amd
 
 - label: e2e Core (1 GPU)
+  device: h200_35gb
   key: e2e-core-1-gpu
   timeout_in_minutes: 30
   source_file_dependencies:

.buildkite/test_areas/entrypoints.yaml

@@ -11,7 +11,7 @@ steps:
   - tests/entrypoints/
   commands:
   - pytest -v -s entrypoints/openai/tool_parsers
-  - pytest -v -s entrypoints/ --ignore=entrypoints/llm --ignore=entrypoints/rpc --ignore=entrypoints/sleep --ignore=entrypoints/serve/instrumentator --ignore=entrypoints/openai --ignore=entrypoints/offline_mode --ignore=entrypoints/test_chat_utils.py  --ignore=entrypoints/pooling --ignore=entrypoints/speech_to_text
+  - 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
 
 - label: Entrypoints Integration (LLM)
   key: entrypoints-integration-llm
@@ -28,7 +28,8 @@ steps:
   - pytest -v -s entrypoints/offline_mode # Needs to avoid interference with other tests
   mirror:
     amd:
-      device: mi300_1
+      device: mi325_1
+      soft_fail: true
       depends_on:
       - image-build-amd
 
@@ -42,10 +43,11 @@ 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_chat_with_tool_reasoning.py --ignore=entrypoints/openai/chat_completion/test_oot_registration.py
+  - pytest -v -s entrypoints/openai/chat_completion --ignore=entrypoints/openai/chat_completion/test_oot_registration.py
   mirror:
     amd:
-      device: mi300_1
+      device: mi325_1
+      soft_fail: true
       timeout_in_minutes: 80
       depends_on:
       - image-build-amd
@@ -58,12 +60,17 @@ steps:
   - 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: mi300_1
+      device: mi325_1
+      soft_fail: true
       timeout_in_minutes: 60
       depends_on:
       - image-build-amd
@@ -82,32 +89,33 @@ steps:
   - 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: mi300_1
+      device: mi325_1
+      soft_fail: true
       timeout_in_minutes: 60
       depends_on:
       - image-build-amd
 
 - label: Entrypoints Integration (API Server 2)
+  device: h200_35gb
   key: entrypoints-integration-api-server-2
   timeout_in_minutes: 130
   working_dir: "/vllm-workspace/tests"
   source_file_dependencies:
   - vllm/
-  - tests/entrypoints/rpc
-  - tests/entrypoints/serve/instrumentator
-  - tests/tool_use
+  - tests/entrypoints/serve
   commands:
   - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-  - pytest -v -s entrypoints/serve/instrumentator
-  - PYTHONPATH=/vllm-workspace pytest -v -s entrypoints/rpc
-  - pytest -v -s tool_use
+  - pytest -v -s entrypoints/serve --ignore=entrypoints/serve/dev/rpc
+  - PYTHONPATH=/vllm-workspace pytest -v -s entrypoints/serve/dev/rpc
   mirror:
     amd:
-      device: mi300_1
+      device: mi325_1
+      soft_fail: true
       depends_on:
       - image-build-amd
 
 - label: Entrypoints Integration (Speech to Text)
+  device: h200_35gb
   key: entrypoints-integration-speech_to_text
   timeout_in_minutes: 50
   working_dir: "/vllm-workspace/tests"
@@ -146,6 +154,5 @@ steps:
   source_file_dependencies:
   - csrc/
   - vllm/entrypoints/openai/
-  - vllm/model_executor/models/whisper.py
   commands: # LMEval
   - pytest -s entrypoints/openai/correctness/

.buildkite/test_areas/kernels.yaml

@@ -38,6 +38,28 @@ steps:
   commands:
     - pytest -v -s kernels/core/test_minimax_reduce_rms.py
 
+- label: Deepseek V4 Kernel Test (H100)
+  key: deepseek-v4-kernel-test-h100
+  timeout_in_minutes: 15
+  device: h100
+  source_file_dependencies:
+  - csrc/fused_deepseek_v4_qnorm_rope_kv_insert_kernel.cu
+  - vllm/models/deepseek_v4/common/ops/
+  - tests/kernels/test_fused_deepseek_v4_qnorm_rope_kv_insert.py
+  commands:
+    - pytest -v -s kernels/test_fused_deepseek_v4_*.py
+
+- label: Deepseek V4 Kernel Test (B200)
+  key: deepseek-v4-kernel-test-b200
+  timeout_in_minutes: 15
+  device: b200-k8s
+  source_file_dependencies:
+  - csrc/fused_deepseek_v4_qnorm_rope_kv_insert_kernel.cu
+  - vllm/models/deepseek_v4/common/ops/
+  - tests/kernels/test_fused_deepseek_v4_qnorm_rope_kv_insert.py
+  commands:
+    - pytest -v -s kernels/test_fused_deepseek_v4_*.py
+
 - label: Kernels Attention Test %N
   key: kernels-attention-test
   timeout_in_minutes: 35
@@ -64,7 +86,7 @@ steps:
   parallelism: 2
   mirror:
     amd:
-      device: mi300_1
+      device: mi325_1
       source_file_dependencies:
       - csrc/quantization/
       - vllm/model_executor/layers/quantization
@@ -150,9 +172,12 @@ steps:
   - csrc/quantization/fp4/
   - csrc/attention/mla/
   - csrc/quantization/cutlass_w8a8/moe/
-  - vllm/model_executor/layers/fused_moe/cutlass_moe.py
-  - vllm/model_executor/layers/fused_moe/flashinfer_cutlass_moe.py
-  - vllm/model_executor/layers/fused_moe/flashinfer_a2a_prepare_finalize.py
+  - vllm/model_executor/layers/fused_moe/experts/cutlass_moe.py
+  - vllm/model_executor/layers/fused_moe/experts/flashinfer_cutlass_moe.py
+  - vllm/model_executor/layers/fused_moe/experts/trtllm_nvfp4_moe.py
+  - vllm/model_executor/layers/fused_moe/oracle/nvfp4.py
+  - vllm/model_executor/layers/fused_moe/prepare_finalize/flashinfer_nvlink_one_sided.py
+  - vllm/model_executor/layers/fused_moe/prepare_finalize/flashinfer_nvlink_two_sided.py
   - vllm/model_executor/layers/quantization/utils/flashinfer_utils.py
   - vllm/v1/attention/backends/flashinfer.py
   - vllm/v1/attention/backends/mla/cutlass_mla.py
@@ -185,6 +210,7 @@ steps:
     - pytest -v -s tests/kernels/moe/test_ocp_mx_moe.py
     - pytest -v -s tests/kernels/moe/test_flashinfer.py
     - pytest -v -s tests/kernels/moe/test_flashinfer_moe.py
+    - pytest -v -s tests/kernels/moe/test_trtllm_nvfp4_moe.py
     - pytest -v -s tests/kernels/moe/test_cutedsl_moe.py
     # e2e
     - pytest -v -s tests/models/quantization/test_nvfp4.py

.buildkite/test_areas/lm_eval.yaml

@@ -3,6 +3,7 @@ depends_on:
   - image-build
 steps:
 - label: LM Eval Small Models
+  device: h200_35gb
   key: lm-eval-small-models
   timeout_in_minutes: 75
   source_file_dependencies:
@@ -152,6 +153,7 @@ steps:
     - pytest -s -v evals/gpt_oss/test_gpqa_correctness.py --config-list-file=configs/models-b200.txt
 
 - label: MRCR Eval Small Models
+  device: h200_35gb
   timeout_in_minutes: 30
   source_file_dependencies:
   - tests/evals/mrcr/

.buildkite/test_areas/lora.yaml

@@ -3,6 +3,7 @@ depends_on:
   - image-build
 steps:
 - label: LoRA %N
+  device: h200_35gb
   key: lora
   timeout_in_minutes: 30
   source_file_dependencies:

.buildkite/test_areas/misc.yaml

@@ -3,6 +3,7 @@ depends_on:
   - image-build
 steps:
 - label: V1 Spec Decode
+  device: h200_35gb
   key: v1-spec-decode
   timeout_in_minutes: 30
   source_file_dependencies:
@@ -51,7 +52,7 @@ steps:
     - pytest -v -s v1/test_outputs.py
   mirror:
     amd:
-      device: mi300_1
+      device: mi325_1
       depends_on:
       - image-build-amd
 
@@ -85,7 +86,7 @@ steps:
     - tests/v1/metrics
     - tests/entrypoints/openai/correctness/test_lmeval.py
   commands:
-    - uv pip install --system -r /vllm-workspace/requirements/kv_connectors.txt
+    - bash /vllm-workspace/.buildkite/scripts/install-kv-connectors.sh
     - export VLLM_WORKER_MULTIPROC_METHOD=spawn
     # split the test to avoid interference
     - pytest -v -s -m 'not cpu_test' v1/core
@@ -165,6 +166,7 @@ steps:
   working_dir: "/vllm-workspace/tests" # optional
 
 - label: Examples
+  device: h200_35gb
   key: examples
   timeout_in_minutes: 45
   working_dir: "/vllm-workspace/examples"
@@ -235,6 +237,7 @@ steps:
   - bash standalone_tests/python_only_compile.sh
 
 - label: Async Engine, Inputs, Utils, Worker
+  device: h200_35gb
   key: async-engine-inputs-utils-worker
   timeout_in_minutes: 50
   source_file_dependencies:
@@ -364,8 +367,10 @@ steps:
     - VLLM_TEST_MODEL=deepseek-ai/DeepSeek-V2-Lite-Chat pytest -v -s v1/determinism/test_batch_invariance.py::test_v1_generation_is_deterministic_across_batch_sizes_with_needle[TRITON_MLA]
     - VLLM_TEST_MODEL=Qwen/Qwen3-30B-A3B-Thinking-2507-FP8 pytest -v -s v1/determinism/test_batch_invariance.py::test_v1_generation_is_deterministic_across_batch_sizes_with_needle[FLASH_ATTN]
     - pytest -v -s v1/determinism/test_nvfp4_batch_invariant.py
+    - pytest -v -s v1/determinism/test_nvfp4_batch_invariant_scaled_mm.py
   
 - label: Acceptance Length Test (Large Models) # optional
+  device: h200_35gb
   key: acceptance-length-test-large-models
   timeout_in_minutes: 25
   gpu: h100

.buildkite/test_areas/model_executor.yaml

@@ -14,5 +14,12 @@ steps:
   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
+    # Dump tracebacks of all threads if a test hangs, so a wedged GPU/CUDA
+    # init surfaces a stack instead of silently stalling.
+    - export PYTHONFAULTHANDLER=1
+    # Per-test watchdog: a single hung test (e.g. stuck during engine/CUDA
+    # init) fails fast with a traceback instead of running until the global
+    # build timeout. The `thread` method also handles hangs inside C/CUDA
+    # calls that the signal method cannot interrupt.
+    - pytest -v -s model_executor -m '(not slow_test)' --timeout=900 --timeout-method=thread
+    - pytest -v -s entrypoints/openai/completion/test_tensorizer_entrypoint.py --timeout=900 --timeout-method=thread

.buildkite/test_areas/model_runner_v2.yaml

@@ -3,6 +3,7 @@ depends_on:
   - image-build
 steps:
 - label: Model Runner V2 Core Tests
+  device: h200_35gb
   key: model-runner-v2-core-tests
   timeout_in_minutes: 45
   source_file_dependencies:
@@ -26,6 +27,7 @@ steps:
   - pytest -v -s entrypoints/llm/test_struct_output_generate.py -k "xgrammar and not speculative_config6 and not speculative_config7 and not speculative_config8 and not speculative_config0"
 
 - label: Model Runner V2 Examples
+  device: h200_35gb
   key: model-runner-v2-examples
   timeout_in_minutes: 45
   working_dir: "/vllm-workspace/examples"
@@ -99,6 +101,7 @@ steps:
     - pytest -v -s distributed/test_pp_cudagraph.py -k "not ray"
 
 - label: Model Runner V2 Spec Decode
+  device: h200_35gb
   key: model-runner-v2-spec-decode
   timeout_in_minutes: 30
   working_dir: "/vllm-workspace/tests"

.buildkite/test_areas/models_basic.yaml

@@ -18,6 +18,7 @@ steps:
     torch_nightly: {}
 
 - label: Basic Models Tests (Extra Initialization) %N
+  device: h200_35gb
   key: basic-models-tests-extra-initialization
   timeout_in_minutes: 45
   source_file_dependencies:
@@ -34,6 +35,7 @@ steps:
     torch_nightly: {}
 
 - label: Basic Models Tests (Other)
+  device: h200_35gb
   key: basic-models-tests-other
   timeout_in_minutes: 45
   source_file_dependencies:
@@ -56,35 +58,3 @@ steps:
   device: cpu-small
   commands:
     - pytest -v -s models/test_utils.py models/test_vision.py
-
-- label: Transformers Nightly Models
-  key: transformers-nightly-models
-  working_dir: "/vllm-workspace/"
-  optional: true
-  soft_fail: true
-  commands:
-    - pip install --upgrade git+https://github.com/huggingface/transformers
-    - pytest -v -s tests/models/test_initialization.py
-    - pytest -v -s tests/models/test_transformers.py
-    - pytest -v -s tests/models/multimodal/processing/
-    - pytest -v -s tests/models/multimodal/test_mapping.py
-    - python3 examples/basic/offline_inference/chat.py
-    - python3 examples/generate/multimodal/vision_language_offline.py --model-type qwen2_5_vl
-    # Whisper needs spawn method to avoid deadlock
-    - VLLM_WORKER_MULTIPROC_METHOD=spawn python3 examples/generate/multimodal/audio_language_offline.py --model-type whisper
-
-- label: Transformers Backward Compatibility Models Test
-  key: transformers-backward-compatibility-models-test
-  working_dir: "/vllm-workspace/"
-  optional: true
-  soft_fail: true
-  commands:
-    - pip install transformers==4.57.5
-    - pytest -v -s tests/models/test_initialization.py
-    - pytest -v -s tests/models/test_transformers.py
-    - pytest -v -s tests/models/multimodal/processing/
-    - pytest -v -s tests/models/multimodal/test_mapping.py
-    - python3 examples/basic/offline_inference/chat.py
-    - python3 examples/generate/multimodal/vision_language_offline.py --model-type qwen2_5_vl
-    # Whisper needs spawn method to avoid deadlock
-    - VLLM_WORKER_MULTIPROC_METHOD=spawn python3 examples/generate/multimodal/audio_language_offline.py --model-type whisper

.buildkite/test_areas/models_language.yaml

@@ -50,7 +50,7 @@ steps:
   mirror:
     torch_nightly: {}
     amd:
-      device: mi300_1
+      device: mi325_1
       depends_on:
       - image-build-amd
       commands:
@@ -59,6 +59,7 @@ steps:
       - pytest -v -s models/language/generation -m hybrid_model --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --shard-id=$$BUILDKITE_PARALLEL_JOB
 
 - label: Language Models Test (Extended Generation) # 80min
+  device: h200_35gb
   key: language-models-test-extended-generation
   timeout_in_minutes: 110
   optional: true
@@ -84,6 +85,7 @@ steps:
     - pytest -v -s models/language/generation_ppl_test
 
 - label: Language Models Test (Extended Pooling)  # 36min
+  device: h200_35gb
   key: language-models-test-extended-pooling
   timeout_in_minutes: 50
   optional: true
@@ -94,7 +96,7 @@ steps:
     - pytest -v -s models/language/pooling -m 'not core_model'
   mirror:
     amd:
-      device: mi300_1
+      device: mi325_1
       timeout_in_minutes: 100
       depends_on:
       - image-build-amd

.buildkite/test_areas/models_multimodal.yaml

@@ -15,7 +15,7 @@ steps:
     - pytest -v -s models/multimodal/generation/test_ultravox.py -m core_model
   mirror:
     amd:
-      device: mi300_1
+      device: mi325_1
       depends_on:
       - image-build-amd
 
@@ -33,11 +33,12 @@ steps:
     - pytest -v -s models/multimodal/generation/test_vit_cudagraph.py -m core_model
   mirror:
     amd:
-      device: mi300_1
+      device: mi325_1
       depends_on:
       - image-build-amd
 
 - label: "Multi-Modal Models (Standard) 3: llava + qwen2_vl"
+  device: h200_35gb
   key: multi-modal-models-standard-3-llava-qwen2-vl
   timeout_in_minutes: 45
   source_file_dependencies:
@@ -49,11 +50,12 @@ steps:
     - pytest -v -s models/multimodal/generation/test_qwen2_vl.py -m core_model
   mirror:
     amd:
-      device: mi300_1
+      device: mi325_1
       depends_on:
       - image-build-amd
 
 - label: "Multi-Modal Models (Standard) 4: other + whisper"
+  device: h200_35gb
   key: multi-modal-models-standard-4-other-whisper
   timeout_in_minutes: 45
   source_file_dependencies:
@@ -92,6 +94,7 @@ steps:
     - pytest -v -s models/multimodal/processing/test_tensor_schema.py
 
 - label: Multi-Modal Accuracy Eval (Small Models) # 50min
+  device: h200_35gb
   key: multi-modal-accuracy-eval-small-models
   timeout_in_minutes: 70
   working_dir: "/vllm-workspace/.buildkite/lm-eval-harness"
@@ -115,11 +118,12 @@ steps:
     - pytest -v -s models/multimodal/test_mapping.py
   mirror:
     amd:
-      device: mi300_1
+      device: mi325_1
       depends_on:
       - image-build-amd
 
 - label: Multi-Modal Models (Extended Generation 2)
+  device: h200_35gb
   key: multi-modal-models-extended-generation-2
   optional: true
   source_file_dependencies:
@@ -130,6 +134,7 @@ steps:
     - pytest -v -s models/multimodal/generation/test_common.py -m 'split(group=0) and not core_model'
 
 - label: Multi-Modal Models (Extended Generation 3)
+  device: h200_35gb
   key: multi-modal-models-extended-generation-3
   optional: true
   source_file_dependencies:

.buildkite/test_areas/pytorch.yaml

@@ -3,6 +3,7 @@ depends_on:
   - image-build
 steps:
 - label: PyTorch Compilation Unit Tests
+  device: h200_35gb
   key: pytorch-compilation-unit-tests
   timeout_in_minutes: 10
   source_file_dependencies:

.buildkite/test_areas/rust_frontend.yaml

@@ -0,0 +1,107 @@
+group: Rust Frontend E2E
+depends_on:
+  - image-build
+steps:
+- label: Rust Frontend OpenAI Coverage
+  timeout_in_minutes: 90
+  device: h200_18gb
+  working_dir: "/vllm-workspace/tests"
+  source_file_dependencies:
+  - rust/
+  - vllm/benchmarks/
+  - vllm/entrypoints/openai/
+  - vllm/entrypoints/serve/
+  - vllm/v1/sample/
+  - tests/utils.py
+  - tests/benchmarks/test_serve_cli.py
+  - tests/entrypoints/openai/chat_completion/test_chat_completion.py
+  # - tests/entrypoints/openai/chat_completion/test_chat_logit_bias_validation.py
+
+  # - tests/entrypoints/openai/completion/test_prompt_validation.py
+  - tests/entrypoints/openai/completion/test_shutdown.py
+  # - tests/entrypoints/openai/test_return_token_ids.py
+  # - tests/entrypoints/openai/test_uds.py
+  - tests/v1/sample/test_logprobs_e2e.py
+  commands:
+  - export VLLM_USE_RUST_FRONTEND=1
+  - export VLLM_WORKER_MULTIPROC_METHOD=spawn
+  - pytest -v -s benchmarks/test_serve_cli.py -k "not insecure and not (test_bench_serve and not test_bench_serve_chat)"
+  - pytest -v -s entrypoints/openai/chat_completion/test_chat_completion.py
+  # - pytest -v -s entrypoints/openai/chat_completion/test_chat_logit_bias_validation.py -k "not invalid"
+
+  # - pytest -v -s entrypoints/openai/completion/test_prompt_validation.py -k "not prompt_embeds"
+  - pytest -v -s entrypoints/openai/completion/test_shutdown.py -k "not engine_failure and not test_abort_timeout_exits_quickly"
+  # - pytest -v -s entrypoints/openai/test_return_token_ids.py
+  # - pytest -v -s entrypoints/openai/test_uds.py
+  - pytest -v -s v1/sample/test_logprobs_e2e.py -k "test_prompt_logprobs_e2e_server"
+
+- label: Rust Frontend Serve/Admin Coverage
+  timeout_in_minutes: 60
+  device: h200_18gb
+  working_dir: "/vllm-workspace/tests"
+  source_file_dependencies:
+  - rust/
+  - vllm/entrypoints/openai/
+  - vllm/entrypoints/serve/
+  - vllm/v1/engine/
+  - tests/utils.py
+  # - tests/entrypoints/serve/dev/rpc/test_collective_rpc.py
+  - tests/entrypoints/serve/disagg/test_serving_tokens.py
+  - tests/entrypoints/serve/instrumentator/test_basic.py
+  - tests/entrypoints/serve/instrumentator/test_metrics.py
+  # - tests/entrypoints/serve/dev/test_sleep.py
+  commands:
+  - export VLLM_USE_RUST_FRONTEND=1
+  - export VLLM_WORKER_MULTIPROC_METHOD=spawn
+  # - pytest -v -s entrypoints/serve/dev/rpc/test_collective_rpc.py
+  - pytest -v -s entrypoints/serve/instrumentator/test_basic.py -k "not show_version and not server_load"
+  - pytest -v -s entrypoints/serve/disagg/test_serving_tokens.py -k "not stream and not lora and not test_generate_logprobs and not stop_string_workflow"
+  - pytest -v -s entrypoints/serve/instrumentator/test_metrics.py -k "text and not show and not run_batch and not test_metrics_counts and not test_metrics_exist"
+  # - pytest -v -s entrypoints/serve/dev/test_sleep.py
+
+- label: Rust Frontend Core Correctness
+  timeout_in_minutes: 30
+  device: h200_18gb
+  working_dir: "/vllm-workspace/tests"
+  source_file_dependencies:
+  - rust/
+  - vllm/entrypoints/openai/
+  - tests/utils.py
+  - tests/entrypoints/openai/correctness/test_lmeval.py
+  commands:
+  - export VLLM_USE_RUST_FRONTEND=1
+  - export VLLM_WORKER_MULTIPROC_METHOD=spawn
+  - pytest -s entrypoints/openai/correctness/test_lmeval.py::test_lm_eval_accuracy_v1_engine
+
+- label: Rust Frontend Tool Use
+  timeout_in_minutes: 60
+  working_dir: "/vllm-workspace/tests"
+  source_file_dependencies:
+  - rust/
+  - vllm/entrypoints/openai/
+  - vllm/tool_parsers/
+  - tests/utils.py
+  - tests/tool_use/
+  commands:
+  - export VLLM_USE_RUST_FRONTEND=1
+  - export VLLM_WORKER_MULTIPROC_METHOD=spawn
+  - pytest -v -s tool_use --ignore=tool_use/mistral --models llama3.2 -k "not test_response_format_with_tool_choice_required and not test_parallel_tool_calls_false and not test_tool_call_and_choice"
+
+- label: Rust Frontend Distributed
+  timeout_in_minutes: 30
+  num_devices: 4
+  working_dir: "/vllm-workspace/tests"
+  source_file_dependencies:
+  - rust/
+  - vllm/distributed/
+  - vllm/engine/
+  - vllm/executor/
+  - vllm/v1/engine/
+  - vllm/v1/worker/
+  - tests/utils.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"

.buildkite/test_areas/rust_frontend_cargo.yaml

@@ -0,0 +1,30 @@
+group: Rust Frontend Cargo
+depends_on: []
+steps:
+- label: Rust Frontend Cargo Style + Clippy
+  key: rust-frontend-cargo-style-clippy
+  depends_on: []
+  timeout_in_minutes: 30
+  device: cpu-medium
+  no_plugin: true
+  source_file_dependencies:
+  - rust/
+  - rust-toolchain.toml
+  - .buildkite/test_areas/rust_frontend_cargo.yaml
+  - .buildkite/scripts/run-rust-frontend-cargo-ci.sh
+  commands:
+  - .buildkite/scripts/run-rust-frontend-cargo-ci.sh style-clippy
+
+- label: Rust Frontend Cargo Tests
+  key: rust-frontend-cargo-tests
+  depends_on: []
+  timeout_in_minutes: 30
+  device: cpu-medium
+  no_plugin: true
+  source_file_dependencies:
+  - rust/
+  - rust-toolchain.toml
+  - .buildkite/test_areas/rust_frontend_cargo.yaml
+  - .buildkite/scripts/run-rust-frontend-cargo-ci.sh
+  commands:
+  - .buildkite/scripts/run-rust-frontend-cargo-ci.sh test

.buildkite/test_areas/samplers.yaml

@@ -3,6 +3,7 @@ depends_on:
   - image-build
 steps:
 - label: Samplers Test
+  device: h200_35gb
   key: samplers-test
   timeout_in_minutes: 75
   source_file_dependencies:
@@ -10,6 +11,7 @@ steps:
   - vllm/sampling_metadata.py
   - tests/samplers
   - tests/conftest.py
+  - vllm/entrypoints/generate/beam_search
   commands:
     # VLLM_USE_FLASHINFER_SAMPLER defaults to 1 now, so we need to pin both
     # values explicitly to still cover the PyTorch-native (Triton) path.

.buildkite/test_areas/spec_decode.yaml

@@ -32,6 +32,7 @@ steps:
   source_file_dependencies:
     - vllm/v1/spec_decode/
     - vllm/v1/worker/gpu/spec_decode/
+    - vllm/v1/attention/backends/
     - vllm/transformers_utils/configs/speculators/
     - tests/v1/e2e/spec_decode/
   commands:

.dockerignore

@@ -2,6 +2,7 @@
 /build
 dist
 vllm/*.so
+vllm/vllm-rs
 
 # Byte-compiled / optimized / DLL files
 __pycache__/
@@ -31,3 +32,4 @@ share/python-wheels/
 .installed.cfg
 *.egg
 MANIFEST
+rust/target/

.github/CODEOWNERS

@@ -40,6 +40,12 @@
 /vllm/entrypoints/chat_utils.py @DarkLight1337
 /vllm/entrypoints/llm.py @DarkLight1337
 
+# Rust Frontend
+/rust/ @BugenZhao @njhill
+/build_rust.sh @BugenZhao @njhill
+/rust-toolchain.toml @BugenZhao @njhill
+/.buildkite/test_areas/rust* @BugenZhao @njhill
+
 # Input/Output Processing
 /vllm/sampling_params.py @njhill @NickLucche
 /vllm/pooling_params.py @noooop @DarkLight1337
@@ -72,21 +78,23 @@
 /vllm/v1/worker/gpu/kv_connector.py @orozery
 
 # CI & building
-/.buildkite @Harry-Chen
-/docker/Dockerfile @Harry-Chen
+/.buildkite @Harry-Chen @khluu
+/docker/Dockerfile @Harry-Chen @khluu
+/pyproject.toml @khluu
+/setup.py @khluu
 
 # Test ownership
-/.buildkite/lm-eval-harness @mgoin 
+/.buildkite/lm-eval-harness @mgoin
 /tests/distributed/test_multi_node_assignment.py @youkaichao
 /tests/distributed/test_pipeline_parallel.py @youkaichao
 /tests/distributed/test_same_node.py @youkaichao
-/tests/entrypoints @DarkLight1337 @robertgshaw2-redhat @aarnphm @NickLucche
-/tests/evals @mgoin @vadiklyutiy
-/tests/kernels @mgoin @tlrmchlsmth @WoosukKwon @yewentao256 @zyongye
+/tests/entrypoints @DarkLight1337 @robertgshaw2-redhat @aarnphm @NickLucche @AndreasKaratzas
+/tests/evals @mgoin @vadiklyutiy @AndreasKaratzas
+/tests/kernels @mgoin @tlrmchlsmth @WoosukKwon @yewentao256 @zyongye @AndreasKaratzas
 /tests/kernels/ir @ProExpertProg @tjtanaa
-/tests/models @DarkLight1337 @ywang96
+/tests/models @DarkLight1337 @ywang96 @AndreasKaratzas
 /tests/multimodal @DarkLight1337 @ywang96 @NickLucche
-/tests/quantization @mgoin @robertgshaw2-redhat @yewentao256 @pavanimajety @zyongye
+/tests/quantization @mgoin @robertgshaw2-redhat @yewentao256 @pavanimajety @zyongye @AndreasKaratzas
 /tests/test_inputs.py @DarkLight1337 @ywang96
 /tests/entrypoints/llm/test_struct_output_generate.py @mgoin @russellb @aarnphm
 /tests/v1/structured_output @mgoin @russellb @aarnphm
@@ -153,9 +161,7 @@ mkdocs.yaml @hmellor
 /vllm/model_executor/models/deepseek_mtp.py @luccafong
 
 # DeepseekV4-specific files
-/vllm/v1/attention/ops/deepseek_v4_ops @zyongye
-/vllm/model_executor/layers/deepseek_compressor.py @zyongye
-/vllm/model_executor/layers/deepseek_v4_attention.py @zyongye
+/vllm/models/deepseek_v4 @zyongye
 /vllm/model_executor/layers/sparse_attn_indexer.py @zyongye
 
 # Mistral-specific files
@@ -172,21 +178,21 @@ mkdocs.yaml @hmellor
 /vllm/model_executor/layers/fla @ZJY0516 @vadiklyutiy
 
 # ROCm related: specify owner with write access to notify AMD folks for careful code review
-/vllm/**/*rocm* @tjtanaa
-/docker/Dockerfile.rocm* @gshtras @tjtanaa
-/vllm/v1/attention/backends/rocm*.py @gshtras @tjtanaa
-/vllm/v1/attention/backends/mla/rocm*.py @gshtras @tjtanaa
-/vllm/v1/attention/ops/rocm*.py @gshtras @tjtanaa
-/vllm/model_executor/layers/fused_moe/rocm*.py @gshtras @tjtanaa
-/csrc/rocm @gshtras @tjtanaa
-/requirements/*rocm* @tjtanaa
-/tests/**/*rocm* @tjtanaa
+/vllm/**/*rocm* @tjtanaa @dllehr-amd
+/docker/Dockerfile.rocm* @tjtanaa @dllehr-amd @AndreasKaratzas
+/vllm/v1/attention/backends/rocm*.py @tjtanaa @dllehr-amd
+/vllm/v1/attention/backends/mla/rocm*.py @tjtanaa @dllehr-amd
+/vllm/v1/attention/ops/rocm*.py @tjtanaa @dllehr-amd
+/vllm/model_executor/layers/fused_moe/rocm*.py @tjtanaa @dllehr-amd
+/csrc/rocm @tjtanaa @dllehr-amd
+/requirements/*rocm* @tjtanaa @AndreasKaratzas
+/tests/**/*rocm* @tjtanaa @AndreasKaratzas
 /docs/**/*rocm* @tjtanaa
 /vllm/**/*quark* @tjtanaa
-/tests/**/*quark* @tjtanaa
+/tests/**/*quark* @tjtanaa @AndreasKaratzas
 /docs/**/*quark* @tjtanaa
-/vllm/**/*aiter* @tjtanaa
-/tests/**/*aiter* @tjtanaa
+/vllm/**/*aiter* @tjtanaa @AndreasKaratzas
+/tests/**/*aiter* @tjtanaa @AndreasKaratzas
 
 # TPU
 /vllm/v1/worker/tpu* @NickLucche

.github/mergify.yml

@@ -103,6 +103,19 @@ pull_request_rules:
       add:
         - frontend
 
+- name: label-rust
+  description: Automatically apply rust label
+  conditions:
+    - label != stale
+    - or:
+      - files~=(?i)rust
+      - title~=(?i)rust
+      - title~=(?i)vllm-rs
+  actions:
+    label:
+      add:
+        - rust
+
 - name: label-llama
   description: Automatically apply llama label
   conditions:

.github/workflows/add_label_automerge.yml

@@ -10,7 +10,7 @@ jobs:
         runs-on: ubuntu-latest
         steps:
             -   name: Add label
-                uses: actions/github-script@ed597411d8f924073f98dfc5c65a23a2325f34cd # v8.0.0
+                uses: actions/github-script@3a2844b7e9c422d3c10d287c895573f7108da1b3 # v9.0.0
                 with:
                     script: |
                         github.rest.issues.addLabels({

.github/workflows/issue_autolabel.yml

@@ -14,7 +14,7 @@ jobs:
     steps:
       - name: Label issues based on keywords
         id: label-step
-        uses: actions/github-script@ed597411d8f924073f98dfc5c65a23a2325f34cd  # v8.0.0
+        uses: actions/github-script@3a2844b7e9c422d3c10d287c895573f7108da1b3  # v9.0.0
         with:
           script: |
             // Configuration: Add new labels and keywords here
@@ -315,7 +315,7 @@ jobs:
 
       - name: CC users for labeled issues
         if: steps.label-step.outputs.labels_added != '[]'
-        uses: actions/github-script@ed597411d8f924073f98dfc5c65a23a2325f34cd  # v8.0.0
+        uses: actions/github-script@3a2844b7e9c422d3c10d287c895573f7108da1b3  # v9.0.0
         with:
           script: |
             // Configuration: Map labels to GitHub users to CC
@@ -392,7 +392,7 @@ jobs:
 
       - name: Request missing ROCm info from issue author
         if: contains(steps.label-step.outputs.labels_added, 'rocm') && contains(toJSON(github.event.issue.labels.*.name), 'bug')
-        uses: actions/github-script@ed597411d8f924073f98dfc5c65a23a2325f34cd  # v8.0.0
+        uses: actions/github-script@3a2844b7e9c422d3c10d287c895573f7108da1b3  # v9.0.0
         with:
           script: |
             const body = (context.payload.issue.body || '').toLowerCase();

.github/workflows/new_pr_bot.yml

@@ -12,7 +12,7 @@ jobs:
     runs-on: ubuntu-latest
     steps:
       - name: Update PR description
-        uses: actions/github-script@ed597411d8f924073f98dfc5c65a23a2325f34cd # v8.0.0
+        uses: actions/github-script@3a2844b7e9c422d3c10d287c895573f7108da1b3 # v9.0.0
         with:
           script: |
             const { owner, repo } = context.repo;
@@ -55,7 +55,7 @@ jobs:
     runs-on: ubuntu-latest
     steps:
       - name: Post welcome comment for first-time contributors
-        uses: actions/github-script@ed597411d8f924073f98dfc5c65a23a2325f34cd # v8.0.0
+        uses: actions/github-script@3a2844b7e9c422d3c10d287c895573f7108da1b3 # v9.0.0
         with:
           script: |
             const { owner, repo } = context.repo;

.github/workflows/pre-commit.yml

@@ -20,7 +20,7 @@ jobs:
     runs-on: ubuntu-latest
     steps:
     - name: Check PR label and author merge count
-      uses: actions/github-script@ed597411d8f924073f98dfc5c65a23a2325f34cd # v8.0.0
+      uses: actions/github-script@3a2844b7e9c422d3c10d287c895573f7108da1b3 # v9.0.0
       with:
         script: |
           const { data: pr } = await github.rest.pulls.get({
@@ -49,7 +49,7 @@ jobs:
     runs-on: ubuntu-latest
     steps:
     - uses: actions/checkout@8e8c483db84b4bee98b60c0593521ed34d9990e8 # v6.0.1
-    - uses: actions/setup-python@83679a892e2d95755f2dac6acb0bfd1e9ac5d548 # v6.1.0
+    - uses: actions/setup-python@a309ff8b426b58ec0e2a45f0f869d46889d02405 # v6.2.0
       with:
         python-version: "3.12"
     - run: echo "::add-matcher::.github/workflows/matchers/actionlint.json"

.gitignore

@@ -26,6 +26,9 @@ __pycache__/
 # C extensions
 *.so
 
+# Rust binaries
+vllm/vllm-rs
+
 # Distribution / packaging
 .Python
 build/

.pre-commit-config.yaml

@@ -21,7 +21,7 @@ repos:
   rev: v21.1.2
   hooks:
   - id: clang-format
-    exclude: 'csrc/(moe/topk_softmax_kernels.cu|quantization/gguf/(ggml-common.h|dequantize.cuh|vecdotq.cuh|mmq.cuh|mmvq.cuh))|vllm/third_party/.*'
+    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/.*'
     types_or: [c++, cuda]
     args: [--style=file, --verbose]
 - repo: https://github.com/DavidAnson/markdownlint-cli2
@@ -222,6 +222,12 @@ repos:
     name: Update Dockerfile dependency graph
     entry: tools/pre_commit/update-dockerfile-graph.sh
     language: script
+  - id: test-nonroot-entrypoint
+    name: Test non-root entrypoint wrapper
+    entry: bash docker/entrypoints/test_vllm_nonroot_entrypoint.sh
+    language: system
+    pass_filenames: false
+    files: ^docker/entrypoints/(vllm-nonroot-entrypoint|test_vllm_nonroot_entrypoint)\.sh$
   - id: check-forbidden-imports
     name: Check for forbidden imports
     entry: python tools/pre_commit/check_forbidden_imports.py
@@ -256,6 +262,32 @@ repos:
     entry: python tools/pre_commit/check_boolean_context_manager.py
     language: python
     types: [python]
+  # Rust hooks. These shell out to `cargo`; tools/pre_commit/rust-check.sh
+  # skips with a warning when cargo is not installed.
+  - id: rust-cargo-autoinherit
+    name: Rust - Normalize Cargo manifests with autoinherit
+    entry: tools/pre_commit/rust-check.sh autoinherit --prefer-simple-dotted
+    language: script
+    pass_filenames: false
+    require_serial: true
+    stages: [pre-commit] # Only run locally as Buildkite will cover this
+    files: ^rust/(Cargo\.toml|src/.*/Cargo\.toml)$
+  - id: rust-cargo-sort
+    name: Rust - Sort Cargo manifest sections
+    entry: tools/pre_commit/rust-check.sh sort --workspace
+    language: script
+    pass_filenames: false
+    require_serial: true
+    stages: [pre-commit] # Only run locally as Buildkite will cover this
+    files: ^rust/(Cargo\.toml|src/.*/Cargo\.toml)$
+  - id: rust-cargo-fmt
+    name: Rust - Format code
+    entry: tools/pre_commit/rust-check.sh fmt
+    language: script
+    pass_filenames: false
+    require_serial: true
+    stages: [pre-commit] # Only run locally as Buildkite will cover this
+    files: ^rust/.*(\.rs|Cargo\.toml|rustfmt\.toml)$
   # Keep `suggestion` last
   - id: suggestion
     name: Suggestion

.readthedocs.yaml

@@ -9,45 +9,8 @@ build:
     python: "3.12"
   jobs:
     post_checkout:
-      - |
-        if [ "$READTHEDOCS_VERSION_TYPE" = "external" ]; then
-          MAX_WAIT=300
-          INTERVAL=60
-          ELAPSED=0
-          while :; do
-            RAW=$(curl -sS -w "\n%{http_code}" "https://api.github.com/repos/vllm-project/vllm/commits/${READTHEDOCS_GIT_COMMIT_HASH}/check-runs?check_name=pre-run-check&filter=latest")
-            HTTP_CODE=$(printf %s "$RAW" | tail -n1)
-            BODY=$(printf %s "$RAW" | head -n -1)
-            if [ "$HTTP_CODE" != "200" ]; then
-              echo "GitHub API returned HTTP $HTTP_CODE (likely rate-limited); skipping pre-run-check gate."
-              break
-            fi
-            STATUS=$(printf %s "$BODY" | python3 -c "import sys, json; r=json.load(sys.stdin).get(\"check_runs\",[]); print((r[0].get(\"status\") or \"\") if r else \"none\")")
-            CONCLUSION=$(printf %s "$BODY" | python3 -c "import sys, json; r=json.load(sys.stdin).get(\"check_runs\",[]); print((r[0].get(\"conclusion\") or \"\") if r else \"\")")
-            if [ "$STATUS" = "none" ]; then
-              echo "no pre-run-check found for this commit; skipping gate."
-              break
-            fi
-            if [ -n "$CONCLUSION" ]; then
-              echo "pre-run-check conclusion: $CONCLUSION"
-              if [ "$CONCLUSION" = "failure" ] || [ "$CONCLUSION" = "cancelled" ] || [ "$CONCLUSION" = "timed_out" ]; then
-                echo "pre-run-check did not pass; failing docs build."
-                exit 1
-              fi
-              break
-            fi
-            if [ "$ELAPSED" -ge "$MAX_WAIT" ]; then
-              echo "pre-run-check status=$STATUS after ${MAX_WAIT}s; skipping gate."
-              break
-            fi
-            echo "pre-run-check status=$STATUS; waiting ${INTERVAL}s..."
-            sleep "$INTERVAL"
-            ELAPSED=$((ELAPSED + INTERVAL))
-          done
-        else
-          echo "Not a PR build (version type=$READTHEDOCS_VERSION_TYPE); skipping pre-run-check gate."
-        fi
       - git fetch origin main --unshallow --no-tags --filter=blob:none || true
+      - bash docs/pre_run_check.sh
     pre_create_environment:
       - pip install uv
     create_environment:

AGENTS.md

@@ -101,6 +101,8 @@ pre-commit run ruff-check --all-files
 pre-commit run mypy-3.10 --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.
+
 ### Commit messages
 
 Add attribution using commit trailers such as `Co-authored-by:` (other projects use `Assisted-by:` or `Generated-by:`). For example:

CMakeLists.txt

@@ -144,14 +144,14 @@ endif()
 # Set up GPU language and check the torch version and warn if it isn't
 # what is expected.
 #
-if (NOT HIP_FOUND AND CUDA_FOUND)
+if (NOT HIP_FOUND AND NOT PYTORCH_FOUND_HIP AND CUDA_FOUND)
   set(VLLM_GPU_LANG "CUDA")
 
   if (NOT Torch_VERSION VERSION_EQUAL ${TORCH_SUPPORTED_VERSION_CUDA})
     message(WARNING "Pytorch version ${TORCH_SUPPORTED_VERSION_CUDA} "
       "expected for CUDA build, saw ${Torch_VERSION} instead.")
   endif()
-elseif(HIP_FOUND)
+elseif(HIP_FOUND OR PYTORCH_FOUND_HIP)
   set(VLLM_GPU_LANG "HIP")
 
   # Importing torch recognizes and sets up some HIP/ROCm configuration but does
@@ -305,27 +305,8 @@ endif()
 #
 
 set(VLLM_EXT_SRC
-  "csrc/mamba/mamba_ssm/selective_scan_fwd.cu"
-  "csrc/cache_kernels.cu"
-  "csrc/cache_kernels_fused.cu"
-  "csrc/attention/paged_attention_v1.cu"
-  "csrc/attention/paged_attention_v2.cu"
-  "csrc/attention/merge_attn_states.cu"
-  "csrc/attention/vertical_slash_index.cu"
-  "csrc/pos_encoding_kernels.cu"
-  "csrc/activation_kernels.cu"
-  "csrc/layernorm_kernels.cu"
-  "csrc/fused_qknorm_rope_kernel.cu"
-  "csrc/layernorm_quant_kernels.cu"
-  "csrc/sampler.cu"
-  "csrc/topk.cu"
   "csrc/cuda_view.cu"
-  "csrc/quantization/gptq/q_gemm.cu"
-  "csrc/quantization/w8a8/int8/scaled_quant.cu"
-  "csrc/quantization/w8a8/fp8/common.cu"
-  "csrc/quantization/fused_kernels/fused_layernorm_dynamic_per_token_quant.cu"
   "csrc/quantization/fused_kernels/fused_silu_mul_block_quant.cu"
-  "csrc/quantization/gguf/gguf_kernel.cu"
   "csrc/quantization/activation_kernels.cu"
   "csrc/cuda_utils_kernels.cu"
   "csrc/custom_all_reduce.cu"
@@ -380,16 +361,30 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
   # are not supported by Machete yet.
 
   # marlin arches for fp16 output
-  cuda_archs_loose_intersection(MARLIN_ARCHS "8.0+PTX" "${CUDA_ARCHS}")
+  # Family-conditional 12.0f (one cubin for SM12x family) requires CUDA >= 13.0;
+  # fall back to architecture-specific 12.0a;12.1a on CUDA < 13.0 (e.g. 12.8).
+  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
+    cuda_archs_loose_intersection(MARLIN_ARCHS "8.0+PTX;12.0f" "${CUDA_ARCHS}")
+  else()
+    cuda_archs_loose_intersection(MARLIN_ARCHS "8.0+PTX;12.0a;12.1a" "${CUDA_ARCHS}")
+  endif()
   # marlin has limited support for turing
   cuda_archs_loose_intersection(MARLIN_SM75_ARCHS "7.5" "${CUDA_ARCHS}")
   # marlin arches for bf16 output (we need 9.0 for bf16 atomicAdd PTX)
-  cuda_archs_loose_intersection(MARLIN_BF16_ARCHS "8.0+PTX;9.0+PTX" "${CUDA_ARCHS}")
+  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
+    cuda_archs_loose_intersection(MARLIN_BF16_ARCHS "8.0+PTX;9.0+PTX;12.0f" "${CUDA_ARCHS}")
+  else()
+    cuda_archs_loose_intersection(MARLIN_BF16_ARCHS "8.0+PTX;9.0+PTX;12.0a;12.1a" "${CUDA_ARCHS}")
+  endif()
   # marlin arches for fp8 input
   # - sm80 doesn't support fp8 computation
   # - sm90 and sm100 don't support QMMA.16832.F32.E4M3.E4M3 SAAS instruction
   # so we only enable fp8 computation for SM89 (e.g. RTX 40x0)  and 12.0 (e.g. RTX 50x0)
-  cuda_archs_loose_intersection(MARLIN_FP8_ARCHS "8.9;12.0;12.1" "${CUDA_ARCHS}")
+  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
+    cuda_archs_loose_intersection(MARLIN_FP8_ARCHS "8.9;12.0f" "${CUDA_ARCHS}")
+  else()
+    cuda_archs_loose_intersection(MARLIN_FP8_ARCHS "8.9;12.0a;12.1a" "${CUDA_ARCHS}")
+  endif()
   # marlin arches for other files
   cuda_archs_loose_intersection(MARLIN_OTHER_ARCHS "7.5;8.0+PTX" "${CUDA_ARCHS}")
 
@@ -629,33 +624,46 @@ define_extension_target(
 # Setting this variable sidesteps the issue by calling the driver directly.
 target_compile_definitions(_C PRIVATE CUTLASS_ENABLE_DIRECT_CUDA_DRIVER_CALL=1)
 
-# add OR VLLM_GPU_LANG STREQUAL "HIP" here once
-# https://github.com/vllm-project/vllm/issues/35163 is resolved
-if(VLLM_GPU_LANG STREQUAL "CUDA")
+if(VLLM_GPU_LANG STREQUAL "CUDA" OR VLLM_GPU_LANG STREQUAL "HIP")
   #
   # _C_stable_libtorch extension (ops registered via STABLE_TORCH_LIBRARY)
   #
   set(VLLM_STABLE_EXT_SRC
     "csrc/libtorch_stable/torch_bindings.cpp"
-    "csrc/cutlass_extensions/common.cpp"
-    "csrc/cuda_utils_kernels.cu"
-    "csrc/libtorch_stable/quantization/w8a8/cutlass/scaled_mm_entry.cu"
-    "csrc/libtorch_stable/quantization/fp4/nvfp4_quant_entry.cu"
-    "csrc/libtorch_stable/quantization/fp4/nvfp4_scaled_mm_entry.cu")
+    "csrc/libtorch_stable/activation_kernels.cu"
+    "csrc/libtorch_stable/quantization/w8a8/int8/scaled_quant.cu"
+    "csrc/libtorch_stable/quantization/w8a8/fp8/common.cu"
+    "csrc/libtorch_stable/quantization/gptq/q_gemm.cu"
+    "csrc/libtorch_stable/quantization/gguf/gguf_kernel.cu"
+    "csrc/libtorch_stable/pos_encoding_kernels.cu"
+    "csrc/libtorch_stable/fused_qknorm_rope_kernel.cu"
+    "csrc/libtorch_stable/layernorm_kernels.cu"
+    "csrc/libtorch_stable/layernorm_quant_kernels.cu"
+    "csrc/libtorch_stable/quantization/fused_kernels/fused_layernorm_dynamic_per_token_quant.cu"
+    "csrc/libtorch_stable/attention/merge_attn_states.cu"
+    "csrc/libtorch_stable/sampler.cu"
+    "csrc/libtorch_stable/topk.cu"
+    "csrc/libtorch_stable/mamba/selective_scan_fwd.cu"
+    "csrc/libtorch_stable/attention/paged_attention_v1.cu"
+    "csrc/libtorch_stable/attention/paged_attention_v2.cu"
+    "csrc/libtorch_stable/cache_kernels.cu"
+    "csrc/libtorch_stable/cache_kernels_fused.cu")
 
   if(VLLM_GPU_LANG STREQUAL "CUDA")
     list(APPEND VLLM_STABLE_EXT_SRC
+      "csrc/cuda_utils_kernels.cu"
+      "csrc/cutlass_extensions/common.cpp"
+      "csrc/libtorch_stable/quantization/w8a8/cutlass/scaled_mm_entry.cu"
+      "csrc/libtorch_stable/quantization/fp4/nvfp4_quant_entry.cu"
+      "csrc/libtorch_stable/quantization/fp4/nvfp4_scaled_mm_entry.cu"
       "csrc/libtorch_stable/permute_cols.cu"
       "csrc/libtorch_stable/quantization/w8a8/fp8/per_token_group_quant.cu"
       "csrc/libtorch_stable/quantization/w8a8/int8/per_token_group_quant.cu"
       "csrc/libtorch_stable/quantization/awq/gemm_kernels.cu")
-  endif()
 
-  if(VLLM_GPU_LANG STREQUAL "CUDA")
     set_gencode_flags_for_srcs(
       SRCS "${VLLM_STABLE_EXT_SRC}"
       CUDA_ARCHS "${CUDA_ARCHS}")
-  endif()
 
   # DeepSeek V3 fused A GEMM kernel (requires SM 9.0+, Hopper and later)
   if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
@@ -675,6 +683,22 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
                    "in CUDA target architectures.")
   endif()
 
+  # FP32 router GEMM (H=3072, E=256, M<=32). Requires SM90+ and CUDA >= 12.0.
+  cuda_archs_sm90plus(FP32_ROUTER_GEMM_ARCHS "${CUDA_ARCHS}")
+  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.0 AND FP32_ROUTER_GEMM_ARCHS)
+    set(SRCS
+      "csrc/libtorch_stable/fp32_router_gemm_entry.cu"
+      "csrc/libtorch_stable/fp32_router_gemm.cu")
+    set_gencode_flags_for_srcs(
+      SRCS "${SRCS}"
+      CUDA_ARCHS "${FP32_ROUTER_GEMM_ARCHS}")
+    list(APPEND VLLM_STABLE_EXT_SRC "${SRCS}")
+    message(STATUS "Building fp32_router_gemm for archs: ${FP32_ROUTER_GEMM_ARCHS}")
+  else()
+    message(STATUS "Not building fp32_router_gemm as no compatible archs found "
+                   "(requires SM90+ and CUDA >= 12.0).")
+  endif()
+
   # Only build AllSpark kernels if we are building for at least some compatible archs.
   cuda_archs_loose_intersection(ALLSPARK_ARCHS "8.0;8.6;8.7;8.9" "${CUDA_ARCHS}")
   if (ALLSPARK_ARCHS)
@@ -916,13 +940,11 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
       SRCS "${SRCS}"
       CUDA_ARCHS "${FP4_ARCHS}")
     list(APPEND VLLM_STABLE_EXT_SRC "${SRCS}")
-    # nvfp4_kv_cache_kernels uses non-stable torch API and is called directly
-    # from cache_kernels.cu, so it belongs in _C rather than _C_stable.
-    set(NVFP4_KV_SRC "csrc/nvfp4_kv_cache_kernels.cu")
+    set(NVFP4_KV_SRC "csrc/libtorch_stable/nvfp4_kv_cache_kernels.cu")
     set_gencode_flags_for_srcs(
       SRCS "${NVFP4_KV_SRC}"
       CUDA_ARCHS "${FP4_ARCHS}")
-    target_sources(_C PRIVATE ${NVFP4_KV_SRC})
+    list(APPEND VLLM_STABLE_EXT_SRC "${NVFP4_KV_SRC}")
     target_compile_definitions(_C PRIVATE ENABLE_NVFP4_SM120=1)
     list(APPEND VLLM_GPU_FLAGS "-DENABLE_NVFP4_SM120=1")
     list(APPEND VLLM_GPU_FLAGS "-DENABLE_CUTLASS_MOE_SM120=1")
@@ -952,11 +974,11 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
       SRCS "${SRCS}"
       CUDA_ARCHS "${FP4_ARCHS}")
     list(APPEND VLLM_STABLE_EXT_SRC "${SRCS}")
-    set(NVFP4_KV_SRC "csrc/nvfp4_kv_cache_kernels.cu")
+    set(NVFP4_KV_SRC "csrc/libtorch_stable/nvfp4_kv_cache_kernels.cu")
     set_gencode_flags_for_srcs(
       SRCS "${NVFP4_KV_SRC}"
       CUDA_ARCHS "${FP4_ARCHS}")
-    target_sources(_C PRIVATE ${NVFP4_KV_SRC})
+    list(APPEND VLLM_STABLE_EXT_SRC "${NVFP4_KV_SRC}")
     target_compile_definitions(_C PRIVATE ENABLE_NVFP4_SM100=1)
     list(APPEND VLLM_GPU_FLAGS "-DENABLE_NVFP4_SM100=1")
     list(APPEND VLLM_GPU_FLAGS "-DENABLE_CUTLASS_MOE_SM100=1")
@@ -1035,6 +1057,9 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
     message(STATUS "Building hadacore")
   endif()
 
+  # if CUDA endif
+  endif()
+
   message(STATUS "Enabling C_stable extension.")
   define_extension_target(
     _C_stable_libtorch
@@ -1054,13 +1079,34 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
   target_compile_definitions(_C_stable_libtorch PRIVATE
     TORCH_TARGET_VERSION=0x020A000000000000ULL)
 
-  # Needed to use cuda APIs from C-shim
-  target_compile_definitions(_C_stable_libtorch PRIVATE
-    USE_CUDA)
+  # Needed to use cuda/hip APIs from C-shim
+  if(VLLM_GPU_LANG STREQUAL "CUDA")
+    target_compile_definitions(_C_stable_libtorch PRIVATE USE_CUDA)
+    # Needed by CUTLASS kernels
+    target_compile_definitions(_C_stable_libtorch PRIVATE
+      CUTLASS_ENABLE_DIRECT_CUDA_DRIVER_CALL=1)
+  elseif(VLLM_GPU_LANG STREQUAL "HIP")
+    target_compile_definitions(_C_stable_libtorch PRIVATE USE_ROCM)
+  endif()
 
-  # Needed by CUTLASS kernels
-  target_compile_definitions(_C_stable_libtorch PRIVATE
-    CUTLASS_ENABLE_DIRECT_CUDA_DRIVER_CALL=1)
+  # On ROCm, _C_stable_libtorch calls raw HIP APIs (e.g. hipGetDevice in
+  # get_device_prop()) which must resolve to the same libamdhip64.so that
+  # PyTorch uses.  When PyTorch bundles its own copy (pip/conda wheels),
+  # the raw HIP calls would otherwise resolve to the system ROCm copy,
+  # initializing a second HIP runtime that corrupts device state (wrong
+  # device on DeviceGuard, core dumps on multi-GPU tests).
+  #
+  # If PyTorch doesn't bundle libamdhip64 (built from source against system
+  # ROCm), there is only one copy in the process and no action is needed —
+  # the HIP compiler already links the system libamdhip64 automatically.
+  if(VLLM_GPU_LANG STREQUAL "HIP")
+    find_library(_STABLE_TORCH_AMDHIP64 amdhip64
+      PATHS "${TORCH_INSTALL_PREFIX}/lib" NO_DEFAULT_PATH)
+    if(_STABLE_TORCH_AMDHIP64)
+      message(STATUS "Found PyTorch-bundled libamdhip64 at ${_STABLE_TORCH_AMDHIP64}")
+      target_link_libraries(_C_stable_libtorch PRIVATE ${_STABLE_TORCH_AMDHIP64})
+    endif()
+  endif()
 endif()
 
 #
@@ -1103,7 +1149,11 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
   # moe marlin arches
   # note that we always set `use_atomic_add=False` for moe marlin now,
   # so we don't need 9.0 for bf16 atomicAdd PTX
-  cuda_archs_loose_intersection(MARLIN_MOE_ARCHS "8.0+PTX" "${CUDA_ARCHS}")
+  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
+    cuda_archs_loose_intersection(MARLIN_MOE_ARCHS "8.0+PTX;12.0f" "${CUDA_ARCHS}")
+  else()
+    cuda_archs_loose_intersection(MARLIN_MOE_ARCHS "8.0+PTX;12.0a;12.1a" "${CUDA_ARCHS}")
+  endif()
   # moe marlin has limited support for turing
   cuda_archs_loose_intersection(MARLIN_MOE_SM75_ARCHS "7.5" "${CUDA_ARCHS}")
   # moe marlin arches for fp8 input
@@ -1206,34 +1256,22 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
                    " in CUDA target architectures")
   endif()
 
-  # DeepSeek V3 router GEMM kernel - requires SM90+
-  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
-    cuda_archs_loose_intersection(DSV3_ROUTER_GEMM_ARCHS "9.0a;10.0f;11.0f" "${CUDA_ARCHS}")
-  else()
-    cuda_archs_loose_intersection(DSV3_ROUTER_GEMM_ARCHS "9.0a;10.0a;10.1a;10.3a" "${CUDA_ARCHS}")
-  endif()
-  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.0 AND DSV3_ROUTER_GEMM_ARCHS)
+  # DeepSeek V3 router GEMM kernel requires SM90+ and CUDA >= 12.0.
+  # (fp32_router_gemm has been migrated to _C_stable_libtorch above.)
+  cuda_archs_sm90plus(SM90PLUS_ROUTER_GEMM_ARCHS "${CUDA_ARCHS}")
+  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.0 AND SM90PLUS_ROUTER_GEMM_ARCHS)
     set(DSV3_ROUTER_GEMM_SRC
       "csrc/moe/dsv3_router_gemm_entry.cu"
       "csrc/moe/dsv3_router_gemm_float_out.cu"
       "csrc/moe/dsv3_router_gemm_bf16_out.cu")
     set_gencode_flags_for_srcs(
       SRCS "${DSV3_ROUTER_GEMM_SRC}"
-      CUDA_ARCHS "${DSV3_ROUTER_GEMM_ARCHS}")
+      CUDA_ARCHS "${SM90PLUS_ROUTER_GEMM_ARCHS}")
     list(APPEND VLLM_MOE_EXT_SRC "${DSV3_ROUTER_GEMM_SRC}")
-    message(STATUS "Building DSV3 router GEMM kernel for archs: ${DSV3_ROUTER_GEMM_ARCHS}")
 
-    # DeepSeek V4 fused RMSNorm + router GEMV - same arch gating as DSV3.
-    set(DSV4_NORM_ROUTER_GEMM_SRC
-      "csrc/moe/dsv4_norm_router_gemm_entry.cu"
-      "csrc/moe/dsv4_norm_router_gemm_kernel.cu")
-    set_gencode_flags_for_srcs(
-      SRCS "${DSV4_NORM_ROUTER_GEMM_SRC}"
-      CUDA_ARCHS "${DSV3_ROUTER_GEMM_ARCHS}")
-    list(APPEND VLLM_MOE_EXT_SRC "${DSV4_NORM_ROUTER_GEMM_SRC}")
-    message(STATUS "Building DSV4 norm+router GEMV kernel for archs: ${DSV3_ROUTER_GEMM_ARCHS}")
+    message(STATUS "Building DSV3 router GEMM kernels for archs: ${SM90PLUS_ROUTER_GEMM_ARCHS}")
   else()
-    message(STATUS "Not building DSV3 router GEMM kernel as no compatible archs found"
+    message(STATUS "Not building DSV3 router GEMM kernels as no compatible archs found"
                    " (requires SM90+ and CUDA >= 12.0)")
   endif()
 endif()
@@ -1260,6 +1298,14 @@ if(VLLM_GPU_LANG STREQUAL "HIP")
     "csrc/rocm/skinny_gemms.cu"
     "csrc/rocm/attention.cu")
 
+  set(VLLM_ROCM_HAS_GFX1100 OFF)
+  if(VLLM_GPU_ARCHES MATCHES "gfx1100")
+    set(VLLM_ROCM_HAS_GFX1100 ON)
+    list(APPEND VLLM_ROCM_EXT_SRC
+      "csrc/rocm/q_gemm_rdna3.cu"
+      "csrc/rocm/q_gemm_rdna3_wmma.cu")
+  endif()
+
   define_extension_target(
     _rocm_C
     DESTINATION vllm
@@ -1269,6 +1315,16 @@ if(VLLM_GPU_LANG STREQUAL "HIP")
     ARCHITECTURES ${VLLM_GPU_ARCHES}
     USE_SABI 3
     WITH_SOABI)
+
+  if(VLLM_ROCM_HAS_GFX1100)
+    target_compile_definitions(_rocm_C PRIVATE VLLM_ROCM_GFX1100)
+  endif()
+endif()
+
+# Must run after the last HIP `define_extension_target` so every extension
+# has registered its sources.
+if (VLLM_GPU_LANG STREQUAL "HIP")
+  vllm_finalize_hipify_target()
 endif()
 
 # For CUDA and HIP builds also build the triton_kernels external package.

benchmarks/benchmark_hidden_state_extraction.py

@@ -0,0 +1,415 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""
+Benchmark hidden state extraction throughput.
+
+Measures two modes:
+  1. Baseline: bulk inference with max_tokens=1, no extraction.
+  2. Extract:  async hidden state extraction via ExampleHiddenStatesConnector
+               with N concurrent clients, each consuming hidden states as
+               soon as their request finishes (overlapping I/O with generation).
+
+Reports tokens/s and prompts/s for each mode.
+
+Usage:
+  python benchmarks/benchmark_hidden_state_extraction.py \
+      --model Qwen/Qwen3-0.6B \
+      --num-prompts 64 \
+      --num-clients 8 \
+      --prompt-len 8192 \
+      --layers 1 2 3 4
+"""
+
+import argparse
+import asyncio
+import time
+from concurrent.futures import ThreadPoolExecutor
+
+import torch
+from transformers import AutoConfig
+
+from vllm import LLM, SamplingParams
+from vllm.config.kv_transfer import KVTransferConfig
+from vllm.distributed.kv_transfer.kv_connector.v1 import (
+    example_hidden_states_connector,
+)
+from vllm.engine.arg_utils import AsyncEngineArgs
+from vllm.sampling_params import RequestOutputKind
+from vllm.v1.engine.async_llm import AsyncLLM
+
+
+def _make_profiler_config(profile_dir: str) -> dict:
+    """Build a profiler_config dict for torch profiling."""
+    return {
+        "profiler": "torch",
+        "torch_profiler_dir": profile_dir,
+        "torch_profiler_with_stack": True,
+    }
+
+
+def make_random_prompts(
+    num_prompts: int, prompt_len: int, vocab_size: int, seed: int = 42
+) -> list[list[int]]:
+    """Generate lists of random token IDs."""
+    # Set seed for reproducibility
+    torch.manual_seed(seed)
+    return [
+        torch.randint(0, vocab_size, (prompt_len,)).tolist() for _ in range(num_prompts)
+    ]
+
+
+def consume_hidden_states(path: str) -> float:
+    """Load hidden states from disk and compute per-position mean.
+
+    Returns a single float: the grand mean of all hidden state values.
+    This forces the benchmark to actually read and reduce the data.
+
+    Uses :func:`load_hidden_states` which acquires a shared flock,
+    blocking (without polling) until the async writer releases its
+    exclusive lock.
+    """
+    obj = example_hidden_states_connector.load_hidden_states(path)
+    hs = obj["hidden_states"]
+    total = hs.mean().item()
+
+    example_hidden_states_connector.cleanup_hidden_states(path)
+
+    return total
+
+
+def run_baseline(
+    model: str,
+    prompts: list[list[int]],
+    extra_args: dict,
+    profile_dir: str | None = None,
+) -> dict:
+    """Baseline: bulk inference, no hidden state extraction."""
+    if profile_dir:
+        extra_args = {
+            **extra_args,
+            "profiler_config": _make_profiler_config(profile_dir),
+        }
+    llm = LLM(
+        model=model,
+        enable_prefix_caching=False,
+        enable_chunked_prefill=False,
+        **extra_args,
+    )
+    sampling_params = SamplingParams(max_tokens=1)
+    prompt_inputs = [{"prompt_token_ids": p} for p in prompts]
+
+    # Warmup
+    llm.generate(prompt_inputs[:4], sampling_params, use_tqdm=False)
+
+    if profile_dir:
+        llm.start_profile()
+
+    t0 = time.perf_counter()
+    outputs = llm.generate(prompt_inputs, sampling_params, use_tqdm=True)
+    elapsed = time.perf_counter() - t0
+
+    if profile_dir:
+        llm.stop_profile()
+
+    total_prompt_tokens = sum(len(o.prompt_token_ids) for o in outputs)
+    num_prompts = len(outputs)
+
+    del llm
+    torch.accelerator.empty_cache()
+
+    return {
+        "mode": "baseline",
+        "elapsed_s": elapsed,
+        "num_prompts": num_prompts,
+        "total_prompt_tokens": total_prompt_tokens,
+        "tokens_per_s": total_prompt_tokens / elapsed,
+        "prompts_per_s": num_prompts / elapsed,
+    }
+
+
+# ---- Async extraction benchmark ----
+
+
+async def _client_loop(
+    engine: AsyncLLM,
+    prompt_queue: asyncio.Queue,
+    consume_pool: ThreadPoolExecutor,
+    results: list[dict],
+    client_id: int,
+):
+    """A single async client: pulls prompts, submits to engine, consumes
+    hidden states as soon as each request finishes."""
+    loop = asyncio.get_event_loop()
+    while True:
+        item = await prompt_queue.get()
+        if item is None:
+            prompt_queue.task_done()
+            break
+        idx, token_ids = item
+
+        request_id = f"req-{idx}"
+        sampling_params = SamplingParams(
+            max_tokens=1,
+            output_kind=RequestOutputKind.FINAL_ONLY,
+        )
+
+        final_output = None
+        async for output in engine.generate(
+            request_id=request_id,
+            prompt={"prompt_token_ids": token_ids},
+            sampling_params=sampling_params,
+        ):
+            if output.finished:
+                final_output = output
+
+        # Consume hidden states on a thread (disk I/O)
+        path = final_output.kv_transfer_params["hidden_states_path"]
+        mean_val = await loop.run_in_executor(consume_pool, consume_hidden_states, path)
+        num_tokens = len(final_output.prompt_token_ids)
+
+        results.append(
+            {
+                "request_id": request_id,
+                "num_prompt_tokens": num_tokens,
+                "mean_hidden_value": mean_val,
+            }
+        )
+        prompt_queue.task_done()
+
+
+async def _run_extraction_async(
+    model: str,
+    prompts: list[list[int]],
+    num_clients: int,
+    layers: list[int],
+    tmpdir: str,
+    extra_args: dict,
+    profile_dir: str | None = None,
+) -> dict:
+    if profile_dir:
+        extra_args = {
+            **extra_args,
+            "profiler_config": _make_profiler_config(profile_dir),
+        }
+    engine_args = AsyncEngineArgs(
+        model=model,
+        enable_prefix_caching=False,
+        enable_chunked_prefill=False,
+        max_num_batched_tokens=40960,
+        max_model_len=40960,
+        speculative_config={
+            "method": "extract_hidden_states",
+            "num_speculative_tokens": 1,
+            "draft_model_config": {
+                "hf_config": {
+                    "eagle_aux_hidden_state_layer_ids": layers,
+                },
+            },
+        },
+        kv_transfer_config=KVTransferConfig(
+            kv_connector="ExampleHiddenStatesConnector",
+            kv_role="kv_producer",
+            kv_connector_extra_config={
+                "shared_storage_path": tmpdir,
+            },
+        ),
+        **extra_args,
+    )
+    engine = AsyncLLM.from_engine_args(engine_args)
+
+    try:
+        # Warmup: run a few prompts sequentially, cleaning up generated files
+        for i in range(min(4, len(prompts))):
+            sp = SamplingParams(max_tokens=1, output_kind=RequestOutputKind.FINAL_ONLY)
+            final_output = None
+            async for output in engine.generate(
+                request_id=f"warmup-{i}",
+                prompt={"prompt_token_ids": prompts[i]},
+                sampling_params=sp,
+            ):
+                if output.finished:
+                    final_output = output
+            if final_output and final_output.kv_transfer_params:
+                path = final_output.kv_transfer_params.get("hidden_states_path")
+                if path:
+                    example_hidden_states_connector.cleanup_hidden_states(path)
+
+        if profile_dir:
+            await engine.start_profile()
+
+        # Fill prompt queue
+        prompt_queue: asyncio.Queue = asyncio.Queue()
+        for idx, token_ids in enumerate(prompts):
+            prompt_queue.put_nowait((idx, token_ids))
+        # Sentinel per client
+        for _ in range(num_clients):
+            prompt_queue.put_nowait(None)
+
+        results: list[dict] = []
+        consume_pool = ThreadPoolExecutor(max_workers=num_clients)
+
+        t0 = time.perf_counter()
+        tasks = [
+            asyncio.create_task(
+                _client_loop(engine, prompt_queue, consume_pool, results, i)
+            )
+            for i in range(num_clients)
+        ]
+        await asyncio.gather(*tasks)
+        elapsed = time.perf_counter() - t0
+
+        consume_pool.shutdown(wait=True)
+
+        if profile_dir:
+            await engine.stop_profile()
+
+        total_prompt_tokens = sum(r["num_prompt_tokens"] for r in results)
+        num_prompts = len(results)
+        mean_hidden = sum(r["mean_hidden_value"] for r in results) / max(
+            len(results), 1
+        )
+
+        return {
+            "mode": "extract",
+            "elapsed_s": elapsed,
+            "num_prompts": num_prompts,
+            "total_prompt_tokens": total_prompt_tokens,
+            "tokens_per_s": total_prompt_tokens / elapsed,
+            "prompts_per_s": num_prompts / elapsed,
+            "mean_hidden_value": mean_hidden,
+        }
+    finally:
+        engine.shutdown()
+
+
+def run_extraction(
+    model: str,
+    prompts: list[list[int]],
+    num_clients: int,
+    layers: list[int],
+    extra_args: dict,
+    profile_dir: str | None = None,
+) -> dict:
+    return asyncio.run(
+        _run_extraction_async(
+            model,
+            prompts,
+            num_clients,
+            layers,
+            "/dev/shm",
+            extra_args,
+            profile_dir=profile_dir,
+        )
+    )
+
+
+def print_results(results: dict):
+    mode = results["mode"]
+    print(f"\n{'=' * 60}")
+    print(f"  {mode.upper()} RESULTS")
+    print(f"{'=' * 60}")
+    print(f"  Prompts:             {results['num_prompts']}")
+    print(f"  Total prompt tokens: {results['total_prompt_tokens']:,}")
+    print(f"  Wall time:           {results['elapsed_s']:.2f}s")
+    print(f"  Tokens/s:            {results['tokens_per_s']:,.0f}")
+    print(f"  Prompts/s:           {results['prompts_per_s']:.2f}")
+    if mode == "extract":
+        print(f"  Mean hidden value:   {results['mean_hidden_value']:.6f}")
+    print(f"{'=' * 60}\n")
+
+
+def main():
+    parser = argparse.ArgumentParser(
+        description="Benchmark hidden state extraction throughput"
+    )
+    parser.add_argument("--model", type=str, required=True)
+    parser.add_argument("--num-prompts", type=int, default=64)
+    parser.add_argument("--num-clients", type=int, default=8)
+    parser.add_argument("--prompt-len", type=int, default=8192)
+    parser.add_argument("--layers", type=int, nargs="+", default=[1, 2, 3, 4])
+    parser.add_argument("--skip-baseline", action="store_true")
+    parser.add_argument("--skip-extract", action="store_true")
+    parser.add_argument("--gpu-memory-utilization", type=float, default=0.9)
+    parser.add_argument("--max-num-batched-tokens", type=int, default=None)
+    parser.add_argument("--max-cudagraph-capture-size", type=int, default=None)
+    parser.add_argument("--max-model-len", type=int, default=None)
+    parser.add_argument("--enforce-eager", action="store_true")
+    parser.add_argument("--load-format", type=str, default=None)
+    parser.add_argument(
+        "--profile",
+        action="store_true",
+        help="Enable torch profiler for both baseline and extraction runs.",
+    )
+    parser.add_argument(
+        "--torch-profiler-dir",
+        type=str,
+        default="./vllm_profile",
+        help="Directory to save torch profiler traces (default: ./vllm_profile).",
+    )
+    parser.add_argument(
+        "--enable-flashinfer-autotune",
+        action="store_true",
+        default=False,
+        help="Enable FlashInfer autotuning (can be slow).",
+    )
+    args = parser.parse_args()
+
+    extra_args = {
+        "gpu_memory_utilization": args.gpu_memory_utilization,
+    }
+    if args.max_model_len is not None:
+        extra_args["max_model_len"] = args.max_model_len
+    if args.max_num_batched_tokens is not None:
+        extra_args["max_num_batched_tokens"] = args.max_num_batched_tokens
+        if args.max_model_len and args.max_num_batched_tokens < args.max_model_len:
+            raise ValueError(
+                "max_num_batched_tokens must be >= max_model_len since chunked prefill"
+                " is not supported by hidden state extraction."
+            )
+    if args.enforce_eager:
+        extra_args["enforce_eager"] = True
+    if args.load_format is not None:
+        extra_args["load_format"] = args.load_format
+    if args.max_cudagraph_capture_size is not None:
+        extra_args["max_cudagraph_capture_size"] = args.max_cudagraph_capture_size
+    extra_args["enable_flashinfer_autotune"] = args.enable_flashinfer_autotune
+
+    # Get vocab size from HF config without loading the full model
+    hf_config = AutoConfig.from_pretrained(args.model, trust_remote_code=True)
+    vocab_size = hf_config.vocab_size
+    prompts = make_random_prompts(args.num_prompts, args.prompt_len, vocab_size)
+    print(
+        f"Generated {args.num_prompts} prompts, "
+        f"{args.prompt_len} tokens each (vocab {vocab_size})"
+    )
+
+    profile_dir = args.torch_profiler_dir if args.profile else None
+    if profile_dir:
+        print(f"Torch profiler enabled, traces will be saved to {profile_dir}/")
+
+    if not args.skip_baseline:
+        baseline_profile_dir = f"{profile_dir}/baseline" if profile_dir else None
+        baseline = run_baseline(
+            args.model, prompts, extra_args, profile_dir=baseline_profile_dir
+        )
+        print_results(baseline)
+
+    if not args.skip_extract:
+        extract_profile_dir = f"{profile_dir}/extract" if profile_dir else None
+        extract = run_extraction(
+            args.model,
+            prompts,
+            args.num_clients,
+            args.layers,
+            extra_args,
+            profile_dir=extract_profile_dir,
+        )
+        print_results(extract)
+
+    if not args.skip_baseline and not args.skip_extract:
+        slowdown = baseline["tokens_per_s"] / extract["tokens_per_s"]
+        print("Extraction slowdown factor: {:.2f}x".format(slowdown))
+
+
+if __name__ == "__main__":
+    main()

benchmarks/benchmark_serving_structured_output.py

@@ -115,6 +115,39 @@ class SampleRequest:
 def sample_requests(
     tokenizer: PreTrainedTokenizerBase, args: argparse.Namespace
 ) -> list[SampleRequest]:
+    def _apply_random_prefix(
+        tokenizer: PreTrainedTokenizerBase,
+        requests: list[SampleRequest],
+        prefix_len: int,
+        seed: int,
+    ) -> list[SampleRequest]:
+        if prefix_len <= 0:
+            return requests
+        rng = np.random.default_rng(seed)
+        vocab_size = tokenizer.vocab_size
+        prohibited = getattr(tokenizer, "all_special_ids", None) or []
+        allowed = np.array([i for i in range(vocab_size) if i not in prohibited])
+        if len(allowed) == 0:
+            return requests
+        prefix_ids = rng.integers(0, len(allowed), size=prefix_len)
+        prefix_token_ids = allowed[prefix_ids].tolist()
+        out = []
+        for req in requests:
+            prompt_ids = tokenizer(req.prompt, add_special_tokens=False).input_ids
+            full_ids = prefix_token_ids + prompt_ids
+            full_prompt = tokenizer.decode(full_ids, skip_special_tokens=False)
+            out.append(
+                SampleRequest(
+                    prompt=full_prompt,
+                    prompt_len=len(tokenizer(full_prompt).input_ids),
+                    expected_output_len=req.expected_output_len,
+                    schema=req.schema,
+                    structure_type=req.structure_type,
+                    completion=req.completion,
+                )
+            )
+        return out
+
     if args.dataset == "json" or args.dataset == "json-unique":
         if args.json_schema_path is None:
             dir_path = os.path.dirname(os.path.realpath(__file__))
@@ -261,6 +294,9 @@ def sample_requests(
                 )
             )
 
+    requests = _apply_random_prefix(
+        tokenizer, requests, args.random_prefix_len, args.seed
+    )
     return requests
 
 
@@ -945,6 +981,15 @@ def create_argument_parser():
         "results in a more uniform arrival of requests.",
     )
     parser.add_argument("--seed", type=int, default=0)
+    parser.add_argument(
+        "--random-prefix-len",
+        type=int,
+        default=0,
+        help=(
+            "Number of prefix tokens to prepend to every prompt. "
+            "The same prefix is used for all prompts to enable prefix caching."
+        ),
+    )
     parser.add_argument(
         "--trust-remote-code",
         action="store_true",

benchmarks/kernels/benchmark_fused_moe_lora_one_shot.py

@@ -0,0 +1,465 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""Benchmark the fused MoE-LoRA fast path (one-shot) vs two-kernel baseline.
+
+The "one_shot" provider goes through `vllm.lora.ops.triton_ops.fused_moe_lora`
+which dispatches to the single-kernel one-shot implementation when
+fully_sharded=False (the prefill default).
+
+The "two_kernel" provider drives `fused_moe_lora_shrink` + `fused_moe_lora_expand`
+directly, bypassing the dispatch and matching the legacy two-kernel path's
+work distribution. This isolates the win from kernel fusion.
+
+Run:
+    .venv/bin/python -m benchmarks.kernels.benchmark_fused_moe_lora_one_shot
+    .venv/bin/python -m benchmarks.kernels.benchmark_fused_moe_lora_one_shot \\
+        --model qwen3moe
+"""
+
+from __future__ import annotations
+
+import argparse
+import os
+import random
+
+import torch
+
+from vllm import _custom_ops as ops
+from vllm.lora.ops.triton_ops import (
+    fused_moe_lora,
+    fused_moe_lora_expand,
+    fused_moe_lora_shrink,
+)
+from vllm.triton_utils import triton
+
+DTYPE = torch.bfloat16
+DEVICE = "cuda"
+
+
+# ----- input fabrication -----------------------------------------------------
+
+
+def _round_up(x: int, base: int) -> int:
+    return ((x + base - 1) // base) * base
+
+
+def _ceildiv(x: int, y: int) -> int:
+    return (x + y - 1) // y
+
+
+def _assign_loras(num_tokens: int, num_sequences: int, max_loras: int) -> torch.Tensor:
+    tokens_per_seq = num_tokens // num_sequences
+    rem = num_tokens % num_sequences
+    out = torch.empty(num_tokens, dtype=torch.int32)
+    start = 0
+    for i in range(num_sequences):
+        end = start + tokens_per_seq + (1 if i < rem else 0)
+        out[start:end] = random.randint(0, max_loras - 1)
+        start = end
+    return out
+
+
+def _assign_experts(num_tokens: int, num_experts: int, top_k: int):
+    expert_indices = torch.empty((num_tokens, top_k), dtype=torch.int32)
+    for i in range(num_tokens):
+        expert_indices[i] = torch.randperm(num_experts)[:top_k]
+    weights = torch.rand((num_tokens, top_k), dtype=torch.float32)
+    weights = weights / weights.sum(dim=1, keepdim=True)
+    return expert_indices, weights
+
+
+def _make_inputs(
+    M: int,
+    K: int,
+    N_per_slice: int,
+    rank: int,
+    num_experts: int,
+    top_k: int,
+    max_loras: int,
+    num_slices: int,
+    block_size_m: int,
+):
+    """Mirrors the production caller's tensor layout."""
+    torch.manual_seed(0)
+    random.seed(0)
+
+    num_sequences = max(1, min(M, 8))
+    topk_ids_cpu, topk_weights_cpu = _assign_experts(M, num_experts, top_k)
+    token_lora_cpu = _assign_loras(M, num_sequences, max_loras)
+    lora_ids_cpu = torch.full((max_loras + 1,), -1, dtype=torch.int32)
+    uniq = torch.unique(token_lora_cpu, sorted=True)
+    lora_ids_cpu[: uniq.size(0)].copy_(uniq)
+
+    topk_ids = topk_ids_cpu.to(DEVICE)
+    topk_weights = topk_weights_cpu.to(device=DEVICE, dtype=DTYPE)
+    token_lora_mapping = token_lora_cpu.to(DEVICE)
+    lora_ids = lora_ids_cpu.to(DEVICE)
+    adapter_enabled = torch.ones(max_loras + 1, dtype=torch.int32, device=DEVICE)
+
+    lora_a = [
+        torch.randn((max_loras, num_experts, rank, K), dtype=DTYPE, device=DEVICE)
+        / max(K, 1) ** 0.5
+        for _ in range(num_slices)
+    ]
+    lora_b = [
+        torch.randn(
+            (max_loras, num_experts, N_per_slice, rank),
+            dtype=DTYPE,
+            device=DEVICE,
+        )
+        / max(rank, 1) ** 0.5
+        for _ in range(num_slices)
+    ]
+    hidden = torch.randn((M, K), dtype=DTYPE, device=DEVICE)
+    out_template = torch.zeros(
+        (M, top_k, num_slices * N_per_slice), dtype=DTYPE, device=DEVICE
+    )
+
+    # Sorted-path metadata (the prefill default).
+    max_pad = topk_ids.numel() + num_experts * (block_size_m - 1)
+    max_pad = _round_up(max_pad, block_size_m)
+    max_blocks = _ceildiv(max_pad, block_size_m)
+    sorted_token_ids = torch.empty(
+        (max_loras * max_pad,), dtype=torch.int32, device=DEVICE
+    )
+    expert_ids = torch.empty(
+        (max_loras * max_blocks,), dtype=torch.int32, device=DEVICE
+    )
+    num_post = torch.empty((max_loras,), dtype=torch.int32, device=DEVICE)
+    ops.moe_lora_align_block_size(
+        topk_ids,
+        token_lora_mapping,
+        num_experts,
+        block_size_m,
+        max_loras,
+        max_pad,
+        max_blocks,
+        sorted_token_ids,
+        expert_ids,
+        num_post,
+        adapter_enabled,
+        lora_ids,
+    )
+    expert_ids = expert_ids.view(max_loras, -1).contiguous()
+    sorted_token_ids = sorted_token_ids.view(max_loras, -1).contiguous()
+    num_active = torch.tensor([max_loras + 1], dtype=torch.int32, device="cpu")
+
+    return dict(
+        hidden=hidden,
+        lora_a=lora_a,
+        lora_b=lora_b,
+        topk_weights=topk_weights,
+        sorted_token_ids=sorted_token_ids,
+        expert_ids=expert_ids,
+        num_post=num_post,
+        token_lora_mapping=token_lora_mapping,
+        lora_ids=lora_ids,
+        num_active=num_active,
+        adapter_enabled=adapter_enabled,
+        out_template=out_template,
+        # bookkeeping
+        M=M,
+        K=K,
+        N_per_slice=N_per_slice,
+        rank=rank,
+        num_experts=num_experts,
+        top_k=top_k,
+        max_loras=max_loras,
+        num_slices=num_slices,
+        block_size_m=block_size_m,
+    )
+
+
+# ----- providers -------------------------------------------------------------
+
+
+def _run_one_shot(inp: dict):
+    """Drive `fused_moe_lora` with fully_sharded=False -> one-shot fast path."""
+    out = inp["out_template"].clone()
+    fused_moe_lora(
+        out,
+        inp["hidden"],
+        inp["lora_a"],
+        inp["lora_b"],
+        inp["topk_weights"],
+        inp["sorted_token_ids"],
+        inp["expert_ids"],
+        inp["num_post"],
+        inp["token_lora_mapping"],
+        inp["rank"],
+        inp["top_k"],
+        inp["lora_ids"],
+        inp["num_active"],
+        inp["adapter_enabled"],
+        inp["block_size_m"],
+        64,
+        32,
+        8,
+        4,
+        3,
+        1,
+        inp["block_size_m"],
+        64,
+        32,
+        8,
+        4,
+        3,
+        1,
+        False,
+        False,
+        0,
+    )
+    return out
+
+
+def _run_two_kernel(inp: dict):
+    """Drive `fused_moe_lora_shrink` + `fused_moe_lora_expand` directly,
+    bypassing the dispatch. Matches the legacy two-kernel work distribution.
+    """
+    M = inp["M"]
+    top_k = inp["top_k"]
+    rank = inp["rank"]
+    num_slices = inp["num_slices"]
+    N_per_slice = inp["N_per_slice"]
+    K = inp["K"]
+    num_experts = inp["num_experts"]
+    block_m = inp["block_size_m"]
+
+    intermediate = torch.zeros((num_slices, M, top_k, rank), dtype=DTYPE, device=DEVICE)
+    out = inp["out_template"].clone()
+    EM = inp["sorted_token_ids"].shape[1]
+    num_tokens = M * top_k
+
+    fused_moe_lora_shrink(
+        intermediate,
+        inp["hidden"],
+        inp["lora_a"],
+        inp["topk_weights"],
+        inp["sorted_token_ids"],
+        inp["expert_ids"],
+        inp["num_post"],
+        inp["token_lora_mapping"],
+        top_k,
+        inp["lora_ids"],
+        inp["adapter_enabled"],
+        torch.device(DEVICE),
+        rank,
+        M,
+        EM,
+        K,
+        num_tokens,
+        num_experts,
+        num_slices,
+        block_m,
+        64,
+        32,
+        8,
+        4,
+        3,
+        1,
+        inp["num_active"],
+        False,
+    )
+    fused_moe_lora_expand(
+        out,
+        intermediate,
+        inp["lora_b"],
+        inp["topk_weights"],
+        inp["sorted_token_ids"],
+        inp["expert_ids"],
+        inp["num_post"],
+        inp["token_lora_mapping"],
+        top_k,
+        inp["lora_ids"],
+        inp["adapter_enabled"],
+        torch.device(DEVICE),
+        rank,
+        M,
+        EM,
+        K,
+        num_tokens,
+        num_experts,
+        num_slices,
+        rank,
+        N_per_slice,
+        block_m,
+        64,
+        32,
+        8,
+        4,
+        3,
+        1,
+        inp["num_active"],
+        False,
+        0,
+    )
+    return out
+
+
+PROVIDER_FNS = {
+    "one_shot": _run_one_shot,
+    "two_kernel": _run_two_kernel,
+}
+
+
+# ----- model presets ---------------------------------------------------------
+
+
+MODEL_PRESETS: dict[str, dict] = {
+    # Mixtral-8x7B style: E=8, top_k=2, hidden=4096, intermediate=14336
+    "mixtral": dict(
+        K=4096,
+        N_per_slice=7168,
+        num_experts=8,
+        top_k=2,
+        max_loras=4,
+        num_slices=2,
+        block_size_m=64,
+    ),
+    # Qwen3-MoE / DeepSeek-V2 style: E=64, top_k=8, hidden=2048, inter=1408
+    "qwen3moe": dict(
+        K=2048,
+        N_per_slice=1408,
+        num_experts=64,
+        top_k=8,
+        max_loras=4,
+        num_slices=2,
+        block_size_m=64,
+    ),
+    # GLM-5.1 (zai-org/GLM-5.1-FP8): E=256, top_k=8, hidden=6144,
+    # moe_intermediate=2048
+    "glm5_1": dict(
+        K=6144,
+        N_per_slice=2048,
+        num_experts=256,
+        top_k=8,
+        max_loras=4,
+        num_slices=2,
+        block_size_m=64,
+    ),
+}
+
+
+M_RANGE = [16, 64, 256, 1024, 4096, 16384]
+RANK_RANGE = [8, 16, 32, 64]
+
+
+def get_benchmark(model: str, max_loras: int | None = None):
+    preset = dict(MODEL_PRESETS[model])
+    if max_loras is not None:
+        preset["max_loras"] = max_loras
+
+    @triton.testing.perf_report(
+        triton.testing.Benchmark(
+            x_names=["M", "rank"],
+            x_vals=[(M, R) for M in M_RANGE for R in RANK_RANGE],
+            line_arg="provider",
+            line_vals=list(PROVIDER_FNS.keys()),
+            line_names=["one_shot (fused)", "two_kernel (legacy)"],
+            styles=[("red", "-"), ("blue", "-")],
+            ylabel="ms",
+            plot_name=f"fused_moe_lora-{model}-loras{preset['max_loras']}",
+            args={"preset": preset},
+        )
+    )
+    def benchmark(M, rank, provider, preset):
+        inp = _make_inputs(
+            M=M,
+            K=preset["K"],
+            N_per_slice=preset["N_per_slice"],
+            rank=rank,
+            num_experts=preset["num_experts"],
+            top_k=preset["top_k"],
+            max_loras=preset["max_loras"],
+            num_slices=preset["num_slices"],
+            block_size_m=preset["block_size_m"],
+        )
+        fn = PROVIDER_FNS[provider]
+        quantiles = [0.5, 0.2, 0.8]
+        ms, min_ms, max_ms = triton.testing.do_bench(
+            lambda: fn(inp), quantiles=quantiles
+        )
+        return ms, max_ms, min_ms
+
+    return benchmark
+
+
+# ----- correctness sanity ---------------------------------------------------
+
+
+def calculate_diff(model: str, M: int, rank: int, max_loras: int | None = None):
+    preset = dict(MODEL_PRESETS[model])
+    if max_loras is not None:
+        preset["max_loras"] = max_loras
+    inp = _make_inputs(
+        M=M,
+        K=preset["K"],
+        N_per_slice=preset["N_per_slice"],
+        rank=rank,
+        num_experts=preset["num_experts"],
+        top_k=preset["top_k"],
+        max_loras=preset["max_loras"],
+        num_slices=preset["num_slices"],
+        block_size_m=preset["block_size_m"],
+    )
+    out_one = _run_one_shot(inp)
+    out_two = _run_two_kernel(inp)
+    max_abs = (out_one.float() - out_two.float()).abs().max().item()
+    print(
+        f"  model={model:<9} M={M:<6} rank={rank:<3}  "
+        f"max|one_shot - two_kernel|={max_abs:.4g}  "
+        f"ref|max|={out_two.float().abs().max().item():.3g}"
+    )
+    if max_abs <= 5e-2:
+        print("  ✅ outputs match within bf16 tolerance")
+    else:
+        print("  ❌ outputs differ beyond expected bf16 noise")
+
+
+# ----- main ------------------------------------------------------------------
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--model",
+        type=str,
+        default="mixtral",
+        choices=list(MODEL_PRESETS.keys()),
+        help="Model preset to sweep",
+    )
+    parser.add_argument(
+        "--save-path",
+        type=str,
+        default="./configs/fused_moe_lora_one_shot/",
+        help="Directory to save benchmark results",
+    )
+    parser.add_argument(
+        "--check-only",
+        action="store_true",
+        help="Run correctness sanity check only, no perf sweep",
+    )
+    parser.add_argument(
+        "--max-loras",
+        type=int,
+        default=None,
+        help="Override max_loras in the model preset (number of LoRA adapters "
+        "active in the batch). Defaults to the preset's value.",
+    )
+    args = parser.parse_args()
+
+    print(f"Correctness check ({args.model}):")
+    calculate_diff(args.model, M=256, rank=32, max_loras=args.max_loras)
+    if args.check_only:
+        raise SystemExit(0)
+
+    effective_max_loras = (
+        args.max_loras
+        if args.max_loras is not None
+        else MODEL_PRESETS[args.model]["max_loras"]
+    )
+    print(f"\nGPU: {torch.cuda.get_device_name()}")
+    print(f"Model preset: {args.model}  max_loras={effective_max_loras}\n")
+    benchmark = get_benchmark(args.model, max_loras=args.max_loras)
+    os.makedirs(args.save_path, exist_ok=True)
+    benchmark.run(print_data=True, save_path=args.save_path)

benchmarks/kernels/benchmark_moe_permute_unpermute.py

@@ -10,6 +10,7 @@ from transformers import AutoConfig
 
 from vllm.model_executor.layers.fused_moe import fused_topk
 from vllm.model_executor.layers.fused_moe.moe_permute_unpermute import (
+    MoEPermuteScratch,
     moe_permute,
     moe_unpermute,
 )
@@ -54,6 +55,15 @@ def benchmark_permute(
     topk_weights, topk_ids, token_expert_indices = fused_topk(
         qhidden_states, input_gating, topk, False
     )
+    scratch = MoEPermuteScratch(
+        max_num_tokens=num_tokens,
+        topk=topk,
+        num_experts=num_experts,
+        num_local_experts=num_experts,
+        device=qhidden_states.device,
+        hidden_size=hidden_size,
+        hidden_dtype=qhidden_states.dtype,
+    )
 
     def prepare(i: int):
         input_gating.copy_(gating_output[i])
@@ -65,6 +75,7 @@ def benchmark_permute(
             topk_ids=topk_ids,
             n_expert=num_experts,
             expert_map=None,
+            scratch=scratch,
         )
 
     # JIT compilation & warmup
@@ -123,6 +134,15 @@ def benchmark_unpermute(
     topk_weights, topk_ids, token_expert_indices = fused_topk(
         qhidden_states, input_gating, topk, False
     )
+    scratch = MoEPermuteScratch(
+        max_num_tokens=num_tokens,
+        topk=topk,
+        num_experts=num_experts,
+        num_local_experts=num_experts,
+        device=qhidden_states.device,
+        hidden_size=hidden_size,
+        hidden_dtype=qhidden_states.dtype,
+    )
 
     def prepare():
         (
@@ -137,6 +157,7 @@ def benchmark_unpermute(
             topk_ids=topk_ids,
             n_expert=num_experts,
             expert_map=None,
+            scratch=scratch,
         )
         # convert to fp16/bf16 as gemm output
         return (

benchmarks/kernels/benchmark_norm_router_gemm.py

@@ -1,183 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-"""Benchmark and correctness check for ``ops.dsv4_norm_router_gemm``.
-
-Two implementations are compared:
-
-  1. ``unfused``   — ``vllm_ops.rms_norm`` then ``ops.dsv3_router_gemm``,
-                     i.e. the current vLLM hot path (two kernel launches).
-  2. ``fused``     — ``ops.dsv4_norm_router_gemm``, the new single-kernel
-                     fused path.
-
-Both produce ``(normed_x: bf16, router_logits: fp32)``.  The correctness
-check verifies that ``fused`` and ``unfused`` agree to within ~1 bf16
-ULP — that is the precision floor for this op.
-"""
-
-import argparse
-
-import torch
-
-from vllm import _custom_ops as vllm_ops
-from vllm.triton_utils import triton
-
-# The fused dsv4_norm_router_gemm kernel is templated only for DSV4-Pro
-# (hidden_size=7168, num_experts=384).  Other shapes fall back to the
-# unfused path on the Python side (NormGatedLinear), so benchmark only
-# the configuration that the fused kernel actually targets.
-HIDDEN_SIZE = 7168
-NUM_EXPERTS_CHOICES = (384,)
-RMS_EPS = 1e-6
-
-
-def unfused_norm_router_gemm(
-    x: torch.Tensor,
-    norm_weight: torch.Tensor,
-    gate_weight: torch.Tensor,
-    eps: float,
-) -> tuple[torch.Tensor, torch.Tensor]:
-    # Call ``_C::rms_norm`` directly (mirroring ``_dsv4_pro_norm_gate``'s
-    # fallback path) so the benchmarked baseline doesn't inherit any
-    # Python wrapper overhead or risk falling through to the native
-    # eager-primitive ``RMSNorm.forward_native`` path.
-    normed = torch.empty_like(x)
-    torch.ops._C.rms_norm(normed, x, norm_weight, eps)
-    logits = vllm_ops.dsv3_router_gemm(normed, gate_weight, torch.float32)
-    return normed, logits
-
-
-def fused_norm_router_gemm(
-    x: torch.Tensor,
-    norm_weight: torch.Tensor,
-    gate_weight: torch.Tensor,
-    eps: float,
-) -> tuple[torch.Tensor, torch.Tensor]:
-    return vllm_ops.dsv4_norm_router_gemm(x, norm_weight, gate_weight, eps)
-
-
-def _make_inputs(num_tokens: int, num_experts: int, hidden_size: int, seed: int = 0):
-    torch.manual_seed(seed)
-    device = "cuda"
-    x = torch.randn(num_tokens, hidden_size, dtype=torch.bfloat16, device=device)
-    norm_w = torch.randn(hidden_size, dtype=torch.bfloat16, device=device)
-    gate_w = torch.randn(num_experts, hidden_size, dtype=torch.bfloat16, device=device)
-    # Down-scale gate_w so the GEMV output stays in a representable range.
-    gate_w = gate_w / float(hidden_size) ** 0.5
-    norm_w = (norm_w * 0.1) + 1.0
-    return x, norm_w, gate_w
-
-
-def calculate_diff(
-    num_tokens: int,
-    num_experts: int,
-    hidden_size: int = HIDDEN_SIZE,
-    normed_atol: float = 2e-3,
-    logits_atol: float = 1e-2,
-    rtol: float = 1e-2,
-) -> None:
-    x, norm_w, gate_w = _make_inputs(num_tokens, num_experts, hidden_size)
-
-    normed_unfused, logits_unfused = unfused_norm_router_gemm(
-        x.clone(), norm_w, gate_w, RMS_EPS
-    )
-    normed_fused, logits_fused = fused_norm_router_gemm(
-        x.clone(), norm_w, gate_w, RMS_EPS
-    )
-
-    def _max_abs(a, b):
-        return (a.float() - b.float()).abs().max().item()
-
-    print(f"\n=== M={num_tokens} E={num_experts} H={hidden_size} ===")
-    print(f"normed_x  |fused - unfused| = {_max_abs(normed_fused, normed_unfused):.3e}")
-    print(f"logits    |fused - unfused| = {_max_abs(logits_fused, logits_unfused):.3e}")
-
-    ok_normed = torch.allclose(
-        normed_fused.float(),
-        normed_unfused.float(),
-        atol=normed_atol,
-        rtol=rtol,
-    )
-    ok_logits = torch.allclose(
-        logits_fused.float(),
-        logits_unfused.float(),
-        atol=logits_atol,
-        rtol=rtol,
-    )
-    if ok_normed and ok_logits:
-        print(
-            f"OK   fused vs unfused within "
-            f"normed_atol={normed_atol:.0e} logits_atol={logits_atol:.0e} "
-            f"rtol={rtol:.0e}"
-        )
-    else:
-        print(
-            f"FAIL normed_ok={ok_normed} logits_ok={ok_logits}; "
-            f"see max-abs values above"
-        )
-
-
-def get_benchmark():
-    # Only num_tokens varies (DSV4-Pro hard-codes E=384); single-axis
-    # sweep yields a clean line plot with M on the x-axis.
-    num_experts = NUM_EXPERTS_CHOICES[0]
-
-    @triton.testing.perf_report(
-        triton.testing.Benchmark(
-            x_names=["num_tokens"],
-            x_vals=list(range(1, 17)),
-            line_arg="provider",
-            line_vals=["unfused", "fused"],
-            line_names=["unfused (rms+dsv3)", "fused (dsv4)"],
-            styles=[("green", "-"), ("red", "-")],
-            ylabel="us",
-            plot_name=f"norm-router-gemm-E{num_experts}-H{HIDDEN_SIZE}",
-            args={},
-        )
-    )
-    def benchmark(num_tokens, provider):
-        x, norm_w, gate_w = _make_inputs(num_tokens, num_experts, HIDDEN_SIZE)
-
-        quantiles = [0.5, 0.2, 0.8]
-        if provider == "unfused":
-            fn = lambda: unfused_norm_router_gemm(  # noqa: E731
-                x, norm_w, gate_w, RMS_EPS
-            )
-        else:
-            fn = lambda: fused_norm_router_gemm(  # noqa: E731
-                x, norm_w, gate_w, RMS_EPS
-            )
-
-        ms, min_ms, max_ms = triton.testing.do_bench(fn, quantiles=quantiles)
-        return 1000 * ms, 1000 * max_ms, 1000 * min_ms
-
-    return benchmark
-
-
-def main() -> None:
-    parser = argparse.ArgumentParser()
-    parser.add_argument(
-        "--save-path",
-        type=str,
-        default="./configs/norm_router_gemm/",
-    )
-    parser.add_argument(
-        "--skip-bench",
-        action="store_true",
-        help="Run only the correctness check, not the perf sweep.",
-    )
-    args = parser.parse_args()
-
-    # Correctness sweep over the full fast-path range M=1..16.
-    for m in range(1, 17):
-        for e in NUM_EXPERTS_CHOICES:
-            calculate_diff(num_tokens=m, num_experts=e, hidden_size=HIDDEN_SIZE)
-
-    if args.skip_bench:
-        return
-
-    benchmark = get_benchmark()
-    benchmark.run(print_data=True, save_path=args.save_path)
-
-
-if __name__ == "__main__":
-    main()

benchmarks/kernels/benchmark_router_gemm.py

@@ -0,0 +1,154 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+import torch
+import torch.nn.functional as F
+
+from vllm import _custom_ops as ops
+from vllm.platforms import current_platform
+from vllm.transformers_utils.config import get_config
+from vllm.triton_utils import triton
+from vllm.utils.argparse_utils import FlexibleArgumentParser
+
+# Dimensions supported by the DSV3 specialized kernel
+DSV3_SUPPORTED_NUM_EXPERTS = [256, 384]
+DSV3_SUPPORTED_HIDDEN_SIZES = [7168]
+
+# Dimensions supported by the gpt-oss specialized kernel
+GPT_OSS_SUPPORTED_NUM_EXPERTS = [32, 128]
+GPT_OSS_SUPPORTED_HIDDEN_SIZES = [2880]
+
+# Dimensions supported by the fp32 specialized kernel (MiniMax-M2)
+FP32_SUPPORTED_NUM_EXPERTS = [256]
+FP32_SUPPORTED_HIDDEN_SIZES = [3072]
+FP32_MAX_TOKENS = 32
+
+
+def get_batch_size_range(max_batch_size):
+    return [2**x for x in range(14) if 2**x <= max_batch_size]
+
+
+def get_model_params(config):
+    if config.architectures[0] in (
+        "DeepseekV2ForCausalLM",
+        "DeepseekV3ForCausalLM",
+        "DeepseekV32ForCausalLM",
+    ):
+        num_experts = config.n_routed_experts
+        hidden_size = config.hidden_size
+    elif config.architectures[0] in ("GptOssForCausalLM",) or config.architectures[
+        0
+    ] in ("MiniMaxM2ForCausalLM",):
+        num_experts = config.num_local_experts
+        hidden_size = config.hidden_size
+    else:
+        raise ValueError(f"Unsupported architecture: {config.architectures}")
+    return num_experts, hidden_size
+
+
+def get_benchmark(model, max_batch_size, trust_remote_code):
+    @triton.testing.perf_report(
+        triton.testing.Benchmark(
+            x_names=["batch_size"],
+            x_vals=get_batch_size_range(max_batch_size),
+            x_log=False,
+            line_arg="provider",
+            line_vals=[
+                "torch",
+                "vllm",
+            ],
+            line_names=["PyTorch", "vLLM"],
+            styles=([("blue", "-"), ("red", "-")]),
+            ylabel="TFLOPs",
+            plot_name=f"{model} router gemm throughput",
+            args={},
+        )
+    )
+    def benchmark(batch_size, provider):
+        config = get_config(model=model, trust_remote_code=trust_remote_code)
+        num_experts, hidden_size = get_model_params(config)
+
+        is_hopper_or_blackwell = current_platform.is_device_capability(
+            90
+        ) or current_platform.is_device_capability_family(100)
+        allow_dsv3_router_gemm = (
+            is_hopper_or_blackwell
+            and num_experts in DSV3_SUPPORTED_NUM_EXPERTS
+            and hidden_size in DSV3_SUPPORTED_HIDDEN_SIZES
+        )
+        allow_gpt_oss_router_gemm = (
+            is_hopper_or_blackwell
+            and num_experts in GPT_OSS_SUPPORTED_NUM_EXPERTS
+            and hidden_size in GPT_OSS_SUPPORTED_HIDDEN_SIZES
+        )
+        is_fp32_router_model = (
+            is_hopper_or_blackwell
+            and num_experts in FP32_SUPPORTED_NUM_EXPERTS
+            and hidden_size in FP32_SUPPORTED_HIDDEN_SIZES
+        )
+        allow_fp32_router_gemm = is_fp32_router_model and batch_size <= FP32_MAX_TOKENS
+
+        # Weight dtype: fp32 kernel requires fp32 weights; others use bf16.
+        weight_dtype = torch.float32 if is_fp32_router_model else torch.bfloat16
+        mat_a = torch.randn(
+            (batch_size, hidden_size), dtype=torch.bfloat16, device="cuda"
+        ).contiguous()
+        mat_b = torch.randn(
+            (num_experts, hidden_size), dtype=weight_dtype, device="cuda"
+        ).contiguous()
+        bias = torch.randn(
+            num_experts, dtype=torch.bfloat16, device="cuda"
+        ).contiguous()
+
+        has_bias = allow_gpt_oss_router_gemm
+
+        quantiles = [0.5, 0.2, 0.8]
+
+        if provider == "torch":
+
+            def runner():
+                if allow_fp32_router_gemm:
+                    F.linear(mat_a.float(), mat_b)
+                elif has_bias:
+                    F.linear(mat_a, mat_b, bias)
+                else:
+                    F.linear(mat_a, mat_b)
+        elif provider == "vllm":
+
+            def runner():
+                if allow_dsv3_router_gemm:
+                    ops.dsv3_router_gemm(mat_a, mat_b, torch.bfloat16)
+                elif allow_fp32_router_gemm:
+                    ops.fp32_router_gemm(mat_a, mat_b)
+                elif allow_gpt_oss_router_gemm:
+                    ops.gpt_oss_router_gemm(mat_a, mat_b, bias)
+                elif is_fp32_router_model:
+                    # batch_size > FP32_MAX_TOKENS: fall back to F.linear
+                    F.linear(mat_a.float(), mat_b)
+                else:
+                    F.linear(mat_a, mat_b)
+
+        ms, min_ms, max_ms = triton.testing.do_bench_cudagraph(
+            runner, quantiles=quantiles
+        )
+
+        def tflops(t_ms):
+            flops = 2 * batch_size * hidden_size * num_experts
+            return flops / (t_ms * 1e-3) / 1e12
+
+        return tflops(ms), tflops(max_ms), tflops(min_ms)
+
+    return benchmark
+
+
+if __name__ == "__main__":
+    parser = FlexibleArgumentParser()
+    parser.add_argument("--model", type=str, default="openai/gpt-oss-20b")
+    parser.add_argument("--max-batch-size", default=16, type=int)
+    parser.add_argument("--trust-remote-code", action="store_true")
+    args = parser.parse_args()
+
+    # Get the benchmark function
+    benchmark = get_benchmark(args.model, args.max_batch_size, args.trust_remote_code)
+    # Run performance benchmark
+    benchmark.run(print_data=True)

benchmarks/kernels/benchmark_selective_state_update.py

@@ -0,0 +1,774 @@
+#!/usr/bin/env python3
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""
+Benchmark and tuning script for the Mamba selective_state_update kernel.
+
+Mirrors the fused MoE tuning workflow: sweeps (BLOCK_SIZE_M, num_warps) across
+an effective_batch grid for a given (headdim, dstate, ngroups, cache_dtype) and
+saves the best config per effective_batch to JSON. Generated configs are picked
+up by selective_state_update at runtime.
+
+Usage:
+    python -m benchmarks.kernels.benchmark_selective_state_update \
+        --all-dstates --save-configs --compare
+"""
+
+import argparse
+import json
+import os
+import sys
+from io import StringIO
+from itertools import product
+from typing import Any
+
+import torch
+
+from tests.kernels.mamba.utils import selective_state_update_ref
+from vllm.model_executor.layers.mamba.ops.mamba_ssm import (
+    _CONFIGS_DIR,
+    _canonical_cache_dtype,
+    _get_default_ssm_launch_config,
+    get_ssm_config_file_name,
+    get_ssm_device_name,
+    override_ssm_config,
+    selective_state_update,
+)
+from vllm.triton_utils import triton
+
+# bf16 shares configs with fp16 - same bit width.
+_SSM_CACHE_DTYPE_MAP: dict[str, torch.dtype] = {
+    "float32": torch.float32,
+    "float16": torch.float16,
+    "bfloat16": torch.float16,
+}
+
+_RESULTS_DIR = os.path.dirname(os.path.realpath(__file__))
+
+# ---------------------------------------------------------------------------
+# Tuning search space
+# ---------------------------------------------------------------------------
+
+_BSM_CHOICES_ALL = [4, 8, 16, 32, 64, 128, 256]
+
+NUM_WARPS_CHOICES = [1, 2, 4, 8]
+
+
+def _block_size_m_choices(headdim: int) -> list[int]:
+    """BLOCK_SIZE_M candidates worth sweeping for a given headdim.
+
+    BLOCK_SIZE_M > next_pow2(headdim) wastes >=50% of each tile via masking
+    (offs_m >= dim rows are zeroed out), so we cap the sweep there.
+    """
+    ceiling = 1
+    while ceiling < headdim:
+        ceiling <<= 1
+    return [b for b in _BSM_CHOICES_ALL if b <= ceiling]
+
+
+# Default deployment shapes. effective_batch = batch * nheads scales the
+# kernel grid, so configs transfer across (model, TP) combos sharing
+# (headdim, dstate, cache_dtype).
+DEFAULT_BATCH_SIZES = [1, 8, 16, 32, 64, 128, 256, 512, 1024, 1536, 2048]
+DEFAULT_NHEADS = [128, 256]
+
+ALL_DSTATES = [16, 32, 64, 128, 256]
+
+# Default tuning shape — matches Nemotron-3-Super and Nemotron-3-Nano Mamba layers.
+# Override with CLI flags for other architectures.
+DEFAULT_HEADDIM = 64
+DEFAULT_NGROUPS = 8
+
+
+# ---------------------------------------------------------------------------
+# Benchmark helper
+# ---------------------------------------------------------------------------
+
+
+def _make_inputs(
+    batch: int,
+    nheads: int,
+    dim: int,
+    dstate: int,
+    ngroups: int,
+    dtype: torch.dtype,
+    state_dtype: torch.dtype | None = None,
+    device: str = "cuda",
+):
+    if state_dtype is None:
+        state_dtype = dtype
+    state = torch.randn(batch, nheads, dim, dstate, dtype=state_dtype, device=device)
+    x = torch.randn(batch, nheads, dim, dtype=dtype, device=device)
+    dt = torch.randn(batch, nheads, dim, dtype=dtype, device=device)
+    A = -torch.rand(nheads, dim, dstate, dtype=torch.float32, device=device)
+    B = torch.randn(batch, ngroups, dstate, dtype=dtype, device=device)
+    C = torch.randn(batch, ngroups, dstate, dtype=dtype, device=device)
+    D = torch.randn(nheads, dim, dtype=dtype, device=device)
+    dt_bias = torch.randn(nheads, dim, dtype=dtype, device=device)
+    out = torch.zeros(batch, nheads, dim, dtype=dtype, device=device)
+    return state, x, dt, A, B, C, D, dt_bias, out
+
+
+def benchmark_config(
+    batch: int,
+    nheads: int,
+    dim: int,
+    dstate: int,
+    ngroups: int,
+    block_size_m: int,
+    num_warps_val: int,
+    dtype: torch.dtype,
+    state_dtype: torch.dtype | None = None,
+    num_iters: int = 100,
+    num_warmup: int = 20,
+    graph_batch_size: int = 10,
+) -> float | None:
+    """
+    Time one (BLOCK_SIZE_M, num_warps) config for selective_state_update.
+    Returns elapsed time in microseconds, or None on error.
+
+    Uses CUDA graph capture-and-replay to isolate kernel time from Python
+    eager-mode dispatch / kwarg-resolution overhead, mirroring the timing
+    methodology in benchmarks/kernels/benchmark_moe.py.
+    """
+    state, x, dt, A, B, C, D, dt_bias, out = _make_inputs(
+        batch, nheads, dim, dstate, ngroups, dtype, state_dtype=state_dtype
+    )
+
+    def _call_kernel() -> None:
+        selective_state_update(
+            state,
+            x,
+            dt,
+            A,
+            B,
+            C,
+            D=D,
+            z=None,
+            dt_bias=dt_bias,
+            dt_softplus=True,
+            out=out,
+        )
+
+    try:
+        with override_ssm_config((block_size_m, num_warps_val)):
+            # Eager-mode warmup: triggers Triton autotune / JIT, primes caches.
+            for _ in range(num_warmup):
+                _call_kernel()
+            torch.accelerator.synchronize()
+
+            # Capture graph_batch_size invocations into a CUDA graph so the
+            # timed region runs without Python dispatch overhead per call.
+            graph = torch.cuda.CUDAGraph()
+            with torch.cuda.graph(graph):
+                for _ in range(graph_batch_size):
+                    _call_kernel()
+            torch.accelerator.synchronize()
+
+            # Warmup graph replays (let the runtime stabilize).
+            for _ in range(5):
+                graph.replay()
+            torch.accelerator.synchronize()
+
+            start = torch.cuda.Event(enable_timing=True)
+            end = torch.cuda.Event(enable_timing=True)
+            latencies: list[float] = []
+            for _ in range(num_iters):
+                start.record()
+                graph.replay()
+                end.record()
+                end.synchronize()
+                latencies.append(start.elapsed_time(end))
+            graph.reset()
+        # elapsed_time returns ms; each replay runs graph_batch_size kernels,
+        # so divide by (num_iters * graph_batch_size) and convert ms -> us.
+        return sum(latencies) / (num_iters * graph_batch_size) * 1000
+    except Exception as e:
+        if "OutOfResources" not in str(e):
+            print(
+                f"    Warning: config M={block_size_m},w={num_warps_val} "
+                f"raised {type(e).__name__}: {e}"
+            )
+        return None
+
+
+# ---------------------------------------------------------------------------
+# Tuning loop
+# ---------------------------------------------------------------------------
+
+
+# CUDA grid Y/Z dim limit — both `batch` and `nheads` must fit individually.
+_CUDA_MAX_GRID_DIM = 65535
+
+# Above this, kernel state-offset arithmetic (batch * nheads * headdim * dstate)
+# overflows int32 and the launch raises cudaErrorIllegalAddress.
+# 262144 covers Nemotron Super TP1 BS=2048.
+_MAX_EFFECTIVE_BATCH = 262144
+
+
+def expand_batch_x_nheads(
+    batch_sizes: list[int],
+    nheads_list: list[int],
+    ngroups: int,
+) -> list[tuple[int, int, int]]:
+    """Cross-product batch_sizes × nheads_list → sorted [(effective_batch,
+    batch, nheads)], deduped by effective_batch. Filters pairs that exceed
+    the CUDA grid dim limit, the effective_batch ceiling, or where nheads is
+    not a positive multiple of ngroups.
+    """
+    seen: dict[int, tuple[int, int]] = {}
+    skipped_grid: list[tuple[int, int]] = []
+    skipped_ngroups: list[tuple[int, int]] = []
+    skipped_eb: list[tuple[int, int]] = []
+    for b, n in product(batch_sizes, nheads_list):
+        if b <= 0 or n <= 0:
+            continue
+        if b > _CUDA_MAX_GRID_DIM or n > _CUDA_MAX_GRID_DIM:
+            skipped_grid.append((b, n))
+            continue
+        if n % ngroups != 0:
+            skipped_ngroups.append((b, n))
+            continue
+        if b * n > _MAX_EFFECTIVE_BATCH:
+            skipped_eb.append((b, n))
+            continue
+        seen.setdefault(b * n, (b, n))
+    if skipped_grid:
+        print(
+            f"  Note: skipping (batch, nheads) pairs exceeding CUDA grid dim "
+            f"{_CUDA_MAX_GRID_DIM}: {skipped_grid}"
+        )
+    if skipped_ngroups:
+        print(
+            f"  Note: skipping (batch, nheads) pairs where nheads % ngroups != 0 "
+            f"for ngroups={ngroups}: {skipped_ngroups}"
+        )
+    if skipped_eb:
+        print(
+            f"  Note: skipping (batch, nheads) pairs whose effective_batch "
+            f"exceeds {_MAX_EFFECTIVE_BATCH}: {skipped_eb}"
+        )
+    return sorted((eb, b, n) for eb, (b, n) in seen.items())
+
+
+def tune_dstate(
+    dstate: int,
+    headdim: int,
+    ngroups: int,
+    dtype: torch.dtype,
+    num_iters: int,
+    verbose: bool,
+    active: list[tuple[int, int, int]],
+    state_dtype: torch.dtype | None = None,
+) -> tuple[dict[int, dict], dict[int, dict[tuple[int, int], float]]]:
+    """For each (effective_batch, batch, nheads) in *active*, sweep
+    (BLOCK_SIZE_M, num_warps) and return
+    ({effective_batch: best_config}, {effective_batch: {(bsm, nw): us}}).
+    The second map is the full timing grid, used downstream so we don't
+    re-measure the same config in the comparison phase.
+    """
+    best_per_eb: dict[int, dict] = {}
+    timings: dict[int, dict[tuple[int, int], float]] = {}
+
+    print(f"\n{'=' * 74}")
+    effective_state_dtype = state_dtype if state_dtype is not None else dtype
+    print(
+        f"Tuning  headdim={headdim}  dstate={dstate}  ngroups={ngroups}  "
+        f"dtype={dtype}  ssm_cache_dtype={effective_state_dtype}"
+    )
+    print(f"{'=' * 74}")
+
+    bsm_choices = _block_size_m_choices(headdim)
+    print(f"BSM candidates (capped at next_pow2(headdim={headdim})): {bsm_choices}")
+
+    hdr = f"{'EffBatch':>8} | {'BLOCK_M':>7} | {'warps':>5} | {'us':>10} | note"
+    print(hdr)
+    print("-" * 52)
+
+    for eb, batch, nheads in active:
+        best_time = float("inf")
+        best_cfg: dict = {}
+        eb_timings: dict[tuple[int, int], float] = {}
+
+        for bsm, nw in product(bsm_choices, NUM_WARPS_CHOICES):
+            t = benchmark_config(
+                batch=batch,
+                nheads=nheads,
+                dim=headdim,
+                dstate=dstate,
+                ngroups=ngroups,
+                block_size_m=bsm,
+                num_warps_val=nw,
+                dtype=dtype,
+                state_dtype=state_dtype,
+                num_iters=num_iters,
+            )
+            if t is None:
+                continue
+            eb_timings[(bsm, nw)] = t
+            is_best = t < best_time
+            if is_best:
+                best_time = t
+                best_cfg = {"BLOCK_SIZE_M": bsm, "num_warps": nw}
+            if verbose:
+                marker = " <-- best" if is_best else ""
+                print(f"{eb:>8} | {bsm:>7} | {nw:>5} | {t:>10.2f} |{marker}")
+
+        timings[eb] = eb_timings
+
+        if not best_cfg:
+            print(
+                f"{eb:>8} | {'-':>7} | {'-':>5} | {'-':>10} | "
+                f"no working config (skipped)"
+            )
+            continue
+
+        if not verbose:
+            print(
+                f"{eb:>8} | {best_cfg['BLOCK_SIZE_M']:>7} | "
+                f"{best_cfg['num_warps']:>5} | {best_time:>10.2f} | best"
+            )
+
+        best_per_eb[eb] = best_cfg
+
+    return best_per_eb, timings
+
+
+# ---------------------------------------------------------------------------
+# Correctness validation
+# ---------------------------------------------------------------------------
+
+
+def validate_configs(
+    dstate: int,
+    headdim: int,
+    ngroups: int,
+    tuned: dict[int, dict],
+    active: list[tuple[int, int, int]],
+    dtype: torch.dtype,
+    atol: float = 1e-2,
+    rtol: float = 1e-2,
+    state_dtype: torch.dtype | None = None,
+) -> dict[int, bool]:
+    """
+    For every (effective_batch, batch, nheads) in *active* that has a tuned
+    config, run the kernel with that config and compare against the reference.
+    Returns {effective_batch: passed}.
+    """
+    # Disable TF32 so the reference's matmul matches the Triton kernel's
+    # fp32 accumulation; otherwise large ebs show bf16 rounding mismatches.
+    torch.set_float32_matmul_precision("highest")
+
+    print(f"\n{'=' * 74}")
+    effective_state_dtype = state_dtype if state_dtype is not None else dtype
+    print(
+        f"Validation  headdim={headdim}  dstate={dstate}  ngroups={ngroups}  "
+        f"dtype={dtype}  ssm_cache_dtype={effective_state_dtype}  atol={atol}"
+    )
+    print(f"{'=' * 74}")
+    print(f"{'EffBatch':>8} | {'MaxAbsErr':>12} | {'Status':>8}")
+    print("-" * 36)
+
+    results: dict[int, bool] = {}
+
+    for eb, batch, nheads in active:
+        cfg = tuned.get(eb)
+        if cfg is None:
+            continue
+        state, x, dt, A, B, C, D, dt_bias, out = _make_inputs(
+            batch=batch,
+            nheads=nheads,
+            dim=headdim,
+            dstate=dstate,
+            ngroups=ngroups,
+            dtype=dtype,
+            state_dtype=state_dtype,
+        )
+        # Clone state before GPU kernel modifies it in-place
+        state_ref = state.clone()
+
+        with override_ssm_config((cfg["BLOCK_SIZE_M"], cfg["num_warps"])):
+            selective_state_update(
+                state,
+                x,
+                dt,
+                A,
+                B,
+                C,
+                D=D,
+                z=None,
+                dt_bias=dt_bias,
+                dt_softplus=True,
+                out=out,
+            )
+        torch.accelerator.synchronize()
+        gpu_out = out.detach().cpu()
+
+        # Reference uses the original (unmodified) state
+        # Upcast to fp32 so the reference sums in fp32 (matches the Triton
+        # kernel); summing in bf16 over `dstate` blows up the error.
+        ref_out = (
+            selective_state_update_ref(
+                state_ref.float(),
+                x.float(),
+                dt.float(),
+                A.float(),
+                B.float(),
+                C.float(),
+                D=D.float(),
+                dt_bias=dt_bias.float(),
+                dt_softplus=True,
+            )
+            .to(out.dtype)
+            .cpu()
+        )
+
+        passed = torch.allclose(gpu_out.float(), ref_out.float(), atol=atol, rtol=rtol)
+        max_err = (gpu_out.float() - ref_out.float()).abs().max().item()
+        status = "PASS" if passed else "FAIL"
+        results[eb] = passed
+        print(f"{eb:>8} | {max_err:>12.6f} | {status:>8}")
+
+    n_pass = sum(results.values())
+    n_total = len(results)
+    print(f"\n  {n_pass}/{n_total} configs passed validation for dstate={dstate}")
+    return results
+
+
+# ---------------------------------------------------------------------------
+# Save configs
+# ---------------------------------------------------------------------------
+
+
+def save_configs(
+    headdim: int,
+    dstate: int,
+    cache_dtype: str,
+    configs: dict[int, dict],
+    save_dir: str | None = None,
+) -> str:
+    # bf16 shares configs with fp16, use common filename for both
+    cache_dtype = _canonical_cache_dtype(cache_dtype)
+
+    base_dir = save_dir if save_dir else _CONFIGS_DIR
+    os.makedirs(base_dir, exist_ok=True)
+    file_path = os.path.join(
+        base_dir,
+        get_ssm_config_file_name(headdim, dstate, cache_dtype, get_ssm_device_name()),
+    )
+    # triton_version is informational only, the loader ignores it
+    payload: dict[str, Any] = {
+        "triton_version": triton.__version__,
+        **{str(k): v for k, v in sorted(configs.items())},
+    }
+    with open(file_path, "w") as f:
+        json.dump(payload, f, indent=4)
+    return file_path
+
+
+# ---------------------------------------------------------------------------
+# Comparison table
+# ---------------------------------------------------------------------------
+
+
+def current_heuristic(dstate: int, is_blackwell: bool = False) -> dict:
+    """Return the current hard-coded BLOCK_SIZE_M / num_warps for dstate."""
+    bsm, nw = _get_default_ssm_launch_config(dstate, is_blackwell)
+    return {"BLOCK_SIZE_M": bsm, "num_warps": nw}
+
+
+def compare_heuristic_vs_tuned(
+    dstate: int,
+    headdim: int,
+    ngroups: int,
+    tuned: dict[int, dict],
+    timings: dict[int, dict[tuple[int, int], float]],
+    active: list[tuple[int, int, int]],
+    dtype: torch.dtype,
+    num_iters: int,
+    is_blackwell: bool,
+    state_dtype: torch.dtype | None = None,
+):
+    heur_cfg = current_heuristic(dstate, is_blackwell)
+    heur_key = (heur_cfg["BLOCK_SIZE_M"], heur_cfg["num_warps"])
+
+    print(f"\n{'=' * 74}")
+    print(
+        f"Comparison  headdim={headdim}  dstate={dstate}  "
+        f"ngroups={ngroups}  —  heuristic vs tuned"
+    )
+    print(
+        f"Heuristic: BLOCK_SIZE_M={heur_cfg['BLOCK_SIZE_M']}, "
+        f"num_warps={heur_cfg['num_warps']}"
+    )
+    print(f"{'=' * 74}")
+    hdr = (
+        f"{'EffBatch':>8} | {'Heur(us)':>10} | {'Tuned(us)':>10} | "
+        f"{'Speedup':>8} | Best config"
+    )
+    print(hdr)
+    print("-" * len(hdr))
+
+    for eb, batch, nheads in active:
+        eb_timings = timings.get(eb, {})
+
+        # Heuristic timing: reuse the tuning measurement if the heuristic
+        # config was in the swept grid; otherwise measure it once.
+        t_h = eb_timings.get(heur_key)
+        if t_h is None:
+            t_h = benchmark_config(
+                batch=batch,
+                nheads=nheads,
+                dim=headdim,
+                dstate=dstate,
+                ngroups=ngroups,
+                block_size_m=heur_cfg["BLOCK_SIZE_M"],
+                num_warps_val=heur_cfg["num_warps"],
+                dtype=dtype,
+                state_dtype=state_dtype,
+                num_iters=num_iters,
+            )
+
+        # `tuned[eb]` may be missing if all configs failed in tune_dstate;
+        # in that case fall back to the heuristic so the table still prints.
+        best = tuned.get(eb) or heur_cfg
+        t_t = eb_timings.get((best["BLOCK_SIZE_M"], best["num_warps"]))
+
+        if t_h is None or t_t is None:
+            print(f"{eb:>8} | {'N/A':>10} | {'N/A':>10} | {'N/A':>8} |")
+            continue
+        speedup = t_h / t_t
+        marker = " <--" if speedup > 1.05 else ""
+        print(
+            f"{eb:>8} | {t_h:>10.2f} | {t_t:>10.2f} | "
+            f"{speedup:>7.2f}x | "
+            f"M={best['BLOCK_SIZE_M']},w={best['num_warps']}{marker}"
+        )
+
+
+# ---------------------------------------------------------------------------
+# CLI
+# ---------------------------------------------------------------------------
+
+
+def save_results(device_name: str, output: str, results_file: str | None = None) -> str:
+    """Save the full benchmark output to a results text file."""
+    if results_file is None:
+        results_file = os.path.join(
+            _RESULTS_DIR, f"ssm_benchmark_results_{device_name}.txt"
+        )
+    with open(results_file, "w") as f:
+        f.write(output)
+    return results_file
+
+
+def main():
+    parser = argparse.ArgumentParser(
+        description="Tune selective_state_update kernel for Mamba SSM"
+    )
+    parser.add_argument(
+        "--dstate",
+        type=int,
+        default=128,
+        help="SSM state size to tune for (default: 128)",
+    )
+    parser.add_argument(
+        "--all-dstates",
+        action="store_true",
+        help="Tune all common dstate values: " + str(ALL_DSTATES),
+    )
+    parser.add_argument(
+        "--dtype",
+        type=str,
+        default="bfloat16",
+        choices=["float16", "bfloat16"],
+        help="Activation / input data type (default: bfloat16)",
+    )
+    parser.add_argument(
+        "--mamba-ssm-cache-dtype",
+        type=str,
+        default="float32",
+        choices=list(_SSM_CACHE_DTYPE_MAP.keys()),
+        help="SSM state cache dtype (default: float32)",
+    )
+    parser.add_argument(
+        "--num-iters",
+        type=int,
+        default=100,
+        help="Number of timing iterations (default: 100)",
+    )
+    parser.add_argument(
+        "--save-configs",
+        action="store_true",
+        help=f"Save best configs to JSON in {_CONFIGS_DIR}",
+    )
+    parser.add_argument(
+        "--compare",
+        action="store_true",
+        help="Show comparison table: heuristic vs tuned",
+    )
+    parser.add_argument(
+        "--verbose",
+        action="store_true",
+        help="Print every (BLOCK_SIZE_M, num_warps) result, not just best",
+    )
+    parser.add_argument(
+        "--results-file",
+        type=str,
+        default=None,
+        help="Path to save the benchmark results text file "
+        "(default: ssm_benchmark_results_<device>.txt alongside this script)",
+    )
+    parser.add_argument(
+        "--save-dir",
+        type=str,
+        default=None,
+        help=f"Directory to save JSON configs (default: {_CONFIGS_DIR})",
+    )
+    parser.add_argument(
+        "--headdim",
+        type=int,
+        default=DEFAULT_HEADDIM,
+        help=f"Per-head feature dim (default: {DEFAULT_HEADDIM})",
+    )
+    parser.add_argument(
+        "--ngroups",
+        type=int,
+        default=DEFAULT_NGROUPS,
+        help=f"Number of B/C groups (default: {DEFAULT_NGROUPS})",
+    )
+    parser.add_argument(
+        "--batch-sizes",
+        type=int,
+        nargs="+",
+        default=DEFAULT_BATCH_SIZES,
+        metavar="B",
+        help=f"Decoder batch sizes to sweep (default: {DEFAULT_BATCH_SIZES})",
+    )
+    parser.add_argument(
+        "--nheads",
+        type=int,
+        nargs="+",
+        default=DEFAULT_NHEADS,
+        metavar="N",
+        help=f"Number of heads per rank to sweep (default: {DEFAULT_NHEADS}). "
+        "effective_batch = batch * nheads; cross-product is deduped by eb.",
+    )
+    parser.add_argument(
+        "--validate",
+        action="store_true",
+        help="After tuning, verify each best config against a CPU reference "
+        "implementation. Configs that fail are flagged in the output.",
+    )
+    parser.add_argument(
+        "--atol",
+        type=float,
+        default=1e-2,
+        help="Absolute tolerance for --validate (default: 1e-2)",
+    )
+    args = parser.parse_args()
+
+    dtype = torch.bfloat16 if args.dtype == "bfloat16" else torch.float16
+    state_dtype = _SSM_CACHE_DTYPE_MAP[args.mamba_ssm_cache_dtype]
+    device_name = get_ssm_device_name()
+    cap = torch.cuda.get_device_capability()
+    is_blackwell = cap[0] >= 10
+
+    # Mirror all output to a results file (like Unix tee).
+    buf = StringIO()
+
+    class _Tee:
+        """Writes to both the original stdout and an in-memory buffer."""
+
+        def write(self, s):
+            buf.write(s)
+            sys.__stdout__.write(s)
+
+        def flush(self):
+            sys.__stdout__.flush()
+
+    sys.stdout = _Tee()  # type: ignore[assignment]
+
+    try:
+        print(f"Device : {device_name}  (sm_{cap[0]}{cap[1]})")
+        print(f"Blackwell: {is_blackwell}")
+        print(f"dtype  : {args.dtype}")
+        print(f"ssm_cache_dtype: {args.mamba_ssm_cache_dtype}")
+        print(f"headdim: {args.headdim}")
+        print(f"ngroups: {args.ngroups}")
+        print(f"triton : {triton.__version__}")
+
+        dstates = ALL_DSTATES if args.all_dstates else [args.dstate]
+        active = expand_batch_x_nheads(args.batch_sizes, args.nheads, args.ngroups)
+
+        for dstate in dstates:
+            tuned, timings = tune_dstate(
+                dstate=dstate,
+                headdim=args.headdim,
+                ngroups=args.ngroups,
+                dtype=dtype,
+                num_iters=args.num_iters,
+                verbose=args.verbose,
+                active=active,
+                state_dtype=state_dtype,
+            )
+
+            if args.compare:
+                compare_heuristic_vs_tuned(
+                    dstate=dstate,
+                    headdim=args.headdim,
+                    ngroups=args.ngroups,
+                    tuned=tuned,
+                    timings=timings,
+                    active=active,
+                    dtype=dtype,
+                    num_iters=args.num_iters,
+                    is_blackwell=is_blackwell,
+                    state_dtype=state_dtype,
+                )
+
+            if args.validate:
+                validity = validate_configs(
+                    dstate=dstate,
+                    headdim=args.headdim,
+                    ngroups=args.ngroups,
+                    tuned=tuned,
+                    active=active,
+                    dtype=dtype,
+                    atol=args.atol,
+                    state_dtype=state_dtype,
+                )
+                # Filter out any configs that failed correctness check
+                failed = [eb for eb, ok in validity.items() if not ok]
+                if failed:
+                    print(
+                        f"\n  WARNING: {len(failed)} config(s) failed validation "
+                        f"for dstate={dstate}: effective_batches {failed}"
+                    )
+                    print("  These will NOT be saved even with --save-configs.")
+                    tuned = {
+                        eb: cfg for eb, cfg in tuned.items() if validity.get(eb, True)
+                    }
+
+            if args.save_configs:
+                path = save_configs(
+                    headdim=args.headdim,
+                    dstate=dstate,
+                    cache_dtype=args.mamba_ssm_cache_dtype,
+                    configs=tuned,
+                    save_dir=args.save_dir,
+                )
+                print(f"\nSaved: {path}")
+            else:
+                print(f"\nBest configs for dstate={dstate}:")
+                for eb, cfg in sorted(tuned.items()):
+                    print(f"  effective_batch={eb:>6}: {cfg}")
+                print("\n(Re-run with --save-configs to persist to JSON)")
+    finally:
+        sys.stdout = sys.__stdout__
+        results_path = save_results(device_name, buf.getvalue(), args.results_file)
+        print(f"\nResults saved to: {results_path}")
+
+
+if __name__ == "__main__":
+    main()

build_rust.sh

@@ -0,0 +1,44 @@
+#!/bin/bash
+# Build the vllm-rs Rust frontend binary and install it into the vllm package.
+# Usage: ./build_rust.sh [--debug]
+#
+# By default builds in release mode. Pass --debug for faster compile times
+# during development.
+
+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/')
+
+# Ensure rustup and the required toolchain are available.
+if ! command -v rustup &>/dev/null; then
+    echo "rustup not found, installing..."
+    curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh -s -- -y --default-toolchain none
+    source "$HOME/.cargo/env"
+fi
+
+if ! rustup run "$TOOLCHAIN" rustc --version &>/dev/null; then
+    echo "Installing Rust toolchain: $TOOLCHAIN"
+    rustup toolchain install "$TOOLCHAIN"
+fi
+
+if [[ "${1:-}" == "--debug" ]]; then
+    PROFILE_ARGS=()
+    PROFILE_DIR="debug"
+else
+    PROFILE_ARGS=(--release)
+    PROFILE_DIR="release"
+fi
+
+cargo +"$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"

cmake/cpu_extension.cmake

@@ -369,6 +369,18 @@ else()
     add_compile_definitions(-DVLLM_NUMA_DISABLED)
 endif()
 
+# check if the pytorch wheel ships libopenblas.so.
+set(VLLM_OPENBLAS_LIB "")
+if (NOT ENABLE_X86_ISA)
+    file(GLOB _VLLM_TORCH_OPENBLAS_LIBS
+        "${TORCH_INSTALL_PREFIX}/lib/libopenblas*.so*")
+    # Note: we don't link openblas directly to _C extension, as it's available through libtorch.so 
+    if (_VLLM_TORCH_OPENBLAS_LIBS)
+        list(GET _VLLM_TORCH_OPENBLAS_LIBS 0 VLLM_OPENBLAS_LIB)
+        message(STATUS "CPU OpenBLAS library: ${VLLM_OPENBLAS_LIB}")
+    endif()
+endif()
+
 #
 # Generate CPU attention dispatch header
 #
@@ -387,6 +399,7 @@ endif()
 #
 set(VLLM_EXT_SRC
     "csrc/cpu/activation.cpp"
+    "csrc/cpu/sgl-kernels/fla.cpp"
     "csrc/cpu/utils.cpp"
     "csrc/cpu/spec_decode_utils.cpp"
     "csrc/cpu/layernorm.cpp"
@@ -396,6 +409,13 @@ set(VLLM_EXT_SRC
     "csrc/cpu/cpu_attn.cpp"
     "csrc/cpu/torch_bindings.cpp")
 
+if (CMAKE_SYSTEM_PROCESSOR MATCHES "riscv64" AND VLLM_RVV_VLEN AND
+        VLLM_RVV_VLEN GREATER 0 AND (RVV_FP16_FOUND OR RVV_BF16_FOUND))
+    set(VLLM_EXT_SRC
+        "csrc/cpu/cpu_wna16.cpp"
+        ${VLLM_EXT_SRC})
+endif()
+
 if (ASIMD_FOUND AND NOT APPLE_SILICON_FOUND)
     set(VLLM_EXT_SRC
         "csrc/cpu/shm.cpp"
@@ -403,6 +423,12 @@ if (ASIMD_FOUND AND NOT APPLE_SILICON_FOUND)
         ${VLLM_EXT_SRC})
 endif()
 
+if (POWER9_FOUND OR POWER10_FOUND OR POWER11_FOUND)	
+    set(VLLM_EXT_SRC
+        "csrc/cpu/shm.cpp"
+        ${VLLM_EXT_SRC})
+endif()
+
 if(USE_ONEDNN)
     set(VLLM_EXT_SRC
         "csrc/cpu/dnnl_kernels.cpp"
@@ -411,7 +437,6 @@ endif()
 
 if (ENABLE_X86_ISA)
     set(VLLM_EXT_SRC_SGL
-        "csrc/cpu/sgl-kernels/fla.cpp"
         "csrc/cpu/sgl-kernels/conv.cpp"
         "csrc/cpu/sgl-kernels/gemm.cpp"
         "csrc/cpu/sgl-kernels/gemm_int8.cpp"
@@ -423,6 +448,7 @@ if (ENABLE_X86_ISA)
         "csrc/cpu/sgl-kernels/moe_fp8.cpp")
 
     set(VLLM_EXT_SRC_AVX512
+        "csrc/cpu/sgl-kernels/fla.cpp"
         "csrc/cpu/shm.cpp"
         "csrc/cpu/cpu_wna16.cpp"
         "csrc/cpu/cpu_fused_moe.cpp"
@@ -439,6 +465,7 @@ if (ENABLE_X86_ISA)
         "csrc/moe/dynamic_4bit_int_moe_cpu.cpp") 
 
     set(VLLM_EXT_SRC_AVX2
+        "csrc/cpu/sgl-kernels/fla.cpp"
         "csrc/cpu/utils.cpp"
         "csrc/cpu/spec_decode_utils.cpp"
         "csrc/cpu/cpu_attn.cpp"
@@ -512,6 +539,9 @@ else()
         USE_SABI 3
         WITH_SOABI
     )
+    if (VLLM_OPENBLAS_LIB)
+        target_compile_definitions(_C PRIVATE VLLM_HAS_OPENBLAS)
+    endif()
 endif()
 
 message(STATUS "Enabling C extension.")

cmake/external_projects/triton_kernels.cmake

@@ -1,6 +1,6 @@
 # Install OpenAI triton_kernels from https://github.com/triton-lang/triton/tree/main/python/triton_kernels
 
-set(DEFAULT_TRITON_KERNELS_TAG "v3.6.0")
+set(DEFAULT_TRITON_KERNELS_TAG "v3.5.1")
 
 # Set TRITON_KERNELS_SRC_DIR for use with local development with vLLM. We expect TRITON_KERNELS_SRC_DIR to
 # be directly set to the triton_kernels python directory.

cmake/external_projects/vllm_flash_attn.cmake

@@ -31,7 +31,7 @@ endif()
 
 if(VLLM_FLASH_ATTN_SRC_DIR)
   FetchContent_Declare(
-          vllm-flash-attn SOURCE_DIR 
+          vllm-flash-attn SOURCE_DIR
           ${VLLM_FLASH_ATTN_SRC_DIR}
           BINARY_DIR ${CMAKE_BINARY_DIR}/vllm-flash-attn
   )
@@ -39,7 +39,7 @@ else()
   FetchContent_Declare(
           vllm-flash-attn
           GIT_REPOSITORY https://github.com/vllm-project/flash-attention.git
-          GIT_TAG bce29425653ec0fbc579d329883030e832d15ada
+          GIT_TAG dd62dac706b1cf7895bd99b18c6cb7e7e117ee25
           GIT_PROGRESS TRUE
           # Don't share the vllm-flash-attn build between build types
           BINARY_DIR ${CMAKE_BINARY_DIR}/vllm-flash-attn

cmake/utils.cmake

@@ -47,12 +47,17 @@ macro (append_cmake_prefix_path PKG EXPR)
   list(APPEND CMAKE_PREFIX_PATH ${_PREFIX_PATH})
 endmacro()
 
-#
-# Add a target named `hipify${NAME}` that runs the hipify preprocessor on a set
-# of CUDA source files. The names of the corresponding "hipified" sources are
-# stored in `OUT_SRCS`.
-#
+# Resolve hipified output paths for `NAME` into `OUT_SRCS` and register the
+# `.cu` sources with the shared `hipify_all` target. Per-extension hipify
+# targets are unsafe to run in parallel against a shared csrc/ output dir, so
+# accumulation here is paired with a single finalize step.
 function (hipify_sources_target OUT_SRCS NAME ORIG_SRCS)
+  if (TARGET hipify_all)
+    message(FATAL_ERROR
+      "hipify_sources_target(${NAME}) called after vllm_finalize_hipify_target. "
+      "Add the new HIP extension before the finalizer call in CMakeLists.txt.")
+  endif()
+
   #
   # Split into C++ and non-C++ (i.e. CUDA) sources.
   #
@@ -73,19 +78,41 @@ function (hipify_sources_target OUT_SRCS NAME ORIG_SRCS)
     list(APPEND HIP_SRCS "${CMAKE_CURRENT_BINARY_DIR}/${SRC}")
   endforeach()
 
-  set(CSRC_BUILD_DIR ${CMAKE_CURRENT_BINARY_DIR}/csrc)
-  add_custom_target(
-    hipify${NAME}
-    COMMAND ${Python_EXECUTABLE} ${CMAKE_SOURCE_DIR}/cmake/hipify.py -p ${CMAKE_SOURCE_DIR}/csrc -o ${CSRC_BUILD_DIR} ${SRCS}
-    DEPENDS ${CMAKE_SOURCE_DIR}/cmake/hipify.py ${SRCS}
-    BYPRODUCTS ${HIP_SRCS}
-    COMMENT "Running hipify on ${NAME} extension source files.")
+  set_property(GLOBAL APPEND PROPERTY VLLM_HIPIFY_ALL_SRCS ${SRCS})
+  set_property(GLOBAL APPEND PROPERTY VLLM_HIPIFY_ALL_BYPRODUCTS ${HIP_SRCS})
 
   # Swap out original extension sources with hipified sources.
   list(APPEND HIP_SRCS ${CXX_SRCS})
   set(${OUT_SRCS} ${HIP_SRCS} PARENT_SCOPE)
 endfunction()
 
+# Define the single shared `hipify_all` custom target that runs hipify once
+# on the union of every HIP extension's sources. Call after the last HIP
+# `define_extension_target`.
+function (vllm_finalize_hipify_target)
+  if (TARGET hipify_all)
+    return()
+  endif()
+
+  get_property(ALL_SRCS GLOBAL PROPERTY VLLM_HIPIFY_ALL_SRCS)
+  get_property(ALL_BYPRODUCTS GLOBAL PROPERTY VLLM_HIPIFY_ALL_BYPRODUCTS)
+
+  if (NOT ALL_SRCS)
+    return()
+  endif()
+
+  list(REMOVE_DUPLICATES ALL_SRCS)
+  list(REMOVE_DUPLICATES ALL_BYPRODUCTS)
+
+  set(CSRC_BUILD_DIR ${CMAKE_CURRENT_BINARY_DIR}/csrc)
+  add_custom_target(
+    hipify_all
+    COMMAND ${Python_EXECUTABLE} ${CMAKE_SOURCE_DIR}/cmake/hipify.py -p ${CMAKE_SOURCE_DIR}/csrc -o ${CSRC_BUILD_DIR} ${ALL_SRCS}
+    DEPENDS ${CMAKE_SOURCE_DIR}/cmake/hipify.py ${ALL_SRCS}
+    BYPRODUCTS ${ALL_BYPRODUCTS}
+    COMMENT "Running hipify on all extension source files.")
+endfunction()
+
 #
 # Get additional GPU compiler flags from torch.
 #
@@ -449,6 +476,16 @@ function(cuda_archs_loose_intersection OUT_CUDA_ARCHS SRC_CUDA_ARCHS TGT_CUDA_AR
   set(${OUT_CUDA_ARCHS} ${_CUDA_ARCHS} PARENT_SCOPE)
 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}")
+  else()
+    cuda_archs_loose_intersection(_archs "9.0a;10.0a;10.1a;10.3a" "${TGT_CUDA_ARCHS}")
+  endif()
+  set(${OUT_CUDA_ARCHS} ${_archs} PARENT_SCOPE)
+endfunction()
+
 #
 # Override the GPU architectures detected by cmake/torch and filter them by
 # `GPU_SUPPORTED_ARCHES`. Sets the final set of architectures in
@@ -551,7 +588,7 @@ function (define_extension_target MOD_NAME)
 
   if (ARG_LANGUAGE STREQUAL "HIP")
     # Make this target dependent on the hipify preprocessor step.
-    add_dependencies(${MOD_NAME} hipify${MOD_NAME})
+    add_dependencies(${MOD_NAME} hipify_all)
     # Make sure we include the hipified versions of the headers, and avoid conflicts with the ones in the original source folder
     target_include_directories(${MOD_NAME} PRIVATE ${CMAKE_CURRENT_BINARY_DIR}/csrc
       ${ARG_INCLUDE_DIRECTORIES})

csrc/attention/dtype_fp8.cuh

@@ -1,6 +1,7 @@
 #pragma once
 
 #include "attention_generic.cuh"
+#include "torch_utils.h"
 
 #include <stdint.h>
 #ifdef ENABLE_FP8
@@ -30,7 +31,7 @@ inline Fp8KVCacheDataType get_fp8_kv_cache_data_type(
   } else if (dtype_str == "fp8_e5m2") {
     return Fp8KVCacheDataType::kFp8E5M2;
   }
-  TORCH_CHECK(false, "Unsupported fp8 kv cache data type: ", dtype_str);
+  TORCH_UTILS_CHECK(false, "Unsupported fp8 kv cache data type: ", dtype_str);
 }
 
 // fp8 vector types for quantization of kv cache

csrc/attention/vertical_slash_index.cu

@@ -1,401 +0,0 @@
-// Copyright (c) Microsoft Corporation.
-// Licensed under the MIT license.
-
-#include <assert.h>
-
-#include <cuda.h>
-
-#include <torch/all.h>
-
-__device__ int64_t save_blocks(int* block_offset, int64_t range_start,
-                               int64_t range_end, int64_t block_size,
-                               int64_t input_block_count, int64_t kv_seqlen) {
-  if (range_start >= kv_seqlen) {
-    return input_block_count;
-  }
-  if (range_end > kv_seqlen) {
-    range_end = kv_seqlen;
-  }
-  int64_t current_block_count = input_block_count;
-  for (int idx = range_start; idx < range_end; idx += block_size) {
-    block_offset[current_block_count++] = idx;
-  }
-  return current_block_count;
-}
-
-__global__ void convert_vertical_slash_indexes_kernel(
-    const int* q_seqlens,         // [BATCH, ]
-    const int* kv_seqlens,        // [BATCH, ]
-    const int* vertical_indexes,  // [BATCH, N_HEADS, NNZ_V]
-    const int* slash_indexes,     // [BATCH, N_HEADS, NNZ_S]
-    int* block_count,             // [BATCH, N_HEADS, cdiv(N_CTX, BLOCK_SIZE_M)]
-    int* block_offset,  // [BATCH, N_HEADS, cdiv(N_CTX, BLOCK_SIZE_M), NNZ_S]
-    int* column_count,  // [BATCH, N_HEADS, cdiv(N_CTX, BLOCK_SIZE_M)]
-    int* column_index,  // [BATCH, N_HEADS, cdiv(N_CTX, BLOCK_SIZE_M), NNZ_V]
-    int64_t N_HEADS, int64_t N_ROWS, int64_t BLOCK_SIZE_M, int64_t BLOCK_SIZE_N,
-    int64_t NNZ_V, int64_t NNZ_S,
-    bool causal  // True for intra, False for succ
-) {
-  const int batch_idx = blockIdx.y;
-  const int head_idx = blockIdx.x;
-  const int group_idx = blockIdx.z;
-
-  int64_t q_seqlen = q_seqlens[batch_idx];
-  int64_t kv_seqlen = kv_seqlens[batch_idx];
-  int64_t block_idx_m = group_idx * blockDim.x + threadIdx.x;
-  int64_t start_m = block_idx_m * BLOCK_SIZE_M;
-  if (start_m >= q_seqlen) {
-    return;
-  }
-  int64_t end_m = start_m + BLOCK_SIZE_M;
-  vertical_indexes += (batch_idx * N_HEADS + head_idx) * NNZ_V;
-  slash_indexes += (batch_idx * N_HEADS + head_idx) * NNZ_S;
-  int64_t row_offset = (batch_idx * N_HEADS + head_idx) * N_ROWS + block_idx_m;
-  block_count += row_offset;
-  block_offset += row_offset * NNZ_S;
-  column_count += row_offset;
-  column_index += row_offset * NNZ_V;
-
-  bool has_slash = true;
-  int64_t tmp_col_cnt = 0, tmp_blk_cnt = 0;
-  int64_t s = 0, v = 0;
-  int64_t v_idx = vertical_indexes[v++];
-  int64_t s_idx = slash_indexes[s++];
-  if (causal) {
-    while (s_idx >= end_m + (kv_seqlen - q_seqlen) && s < NNZ_S) {
-      s_idx = slash_indexes[s++];
-    }
-    if (s_idx > end_m + (kv_seqlen - q_seqlen)) has_slash = false;
-    s_idx = max((kv_seqlen - q_seqlen) + end_m - s_idx, BLOCK_SIZE_M);
-  } else {
-    while (s_idx >= end_m + kv_seqlen && s < NNZ_S) {
-      s_idx = slash_indexes[s++];
-    }
-    if (s_idx > end_m + kv_seqlen) has_slash = false;
-    s_idx = max(kv_seqlen + end_m - s_idx, BLOCK_SIZE_M);
-  }
-
-  int64_t range_start = s_idx - BLOCK_SIZE_M, range_end = s_idx;
-  if (!has_slash) {
-    if (causal) {
-      range_start = (kv_seqlen - q_seqlen) + end_m;
-      range_end = (kv_seqlen - q_seqlen) + end_m + BLOCK_SIZE_N;
-    } else {
-      range_start = kv_seqlen;
-      range_end = kv_seqlen + BLOCK_SIZE_N;
-    }
-  }
-
-  bool slash_finished = false;
-  while (1) {
-    if (v_idx < range_end) {
-      if (v_idx < range_start) {
-        column_index[tmp_col_cnt++] = v_idx;
-      }
-      if (v < NNZ_V) {
-        v_idx = vertical_indexes[v++];
-      } else {
-        if (causal)
-          v_idx = end_m + BLOCK_SIZE_N + (kv_seqlen - q_seqlen);
-        else
-          v_idx = end_m + BLOCK_SIZE_N + kv_seqlen;
-      }
-    } else {
-      if ((s < NNZ_S && causal) ||
-          (s < NNZ_S && !causal && slash_indexes[s] >= start_m)) {
-        if (causal)
-          s_idx = max((kv_seqlen - q_seqlen) + end_m - slash_indexes[s++],
-                      BLOCK_SIZE_M);
-        else
-          s_idx = max(kv_seqlen + end_m - slash_indexes[s++], BLOCK_SIZE_M);
-      } else {
-        if (v == NNZ_V || (v_idx > range_start && causal)) {
-          // add the last vertical if no more slash
-          if (v == NNZ_V && !causal && v_idx < kv_seqlen) {
-            column_index[tmp_col_cnt++] = v_idx;
-          }
-          tmp_blk_cnt = save_blocks(block_offset, range_start, range_end,
-                                    BLOCK_SIZE_N, tmp_blk_cnt, kv_seqlen);
-          break;
-        } else {
-          if (causal) {
-            range_start = (kv_seqlen - q_seqlen) + end_m;
-            range_end = (kv_seqlen - q_seqlen) + end_m + BLOCK_SIZE_N;
-          } else {
-            // if slash_finished but there are vertical left, save current
-            // blocks
-            tmp_blk_cnt = save_blocks(block_offset, range_start, range_end,
-                                      BLOCK_SIZE_N, tmp_blk_cnt, kv_seqlen);
-            range_start = kv_seqlen;
-            range_end = kv_seqlen + BLOCK_SIZE_N;
-          }
-          slash_finished = true;
-        }
-      }
-      if (!slash_finished) {
-        if (s_idx > range_end + BLOCK_SIZE_M) {
-          tmp_blk_cnt = save_blocks(block_offset, range_start, range_end,
-                                    BLOCK_SIZE_N, tmp_blk_cnt, kv_seqlen);
-          range_start = s_idx - BLOCK_SIZE_M;
-          range_end = s_idx;
-        } else if (s_idx > range_end) {
-          range_end += BLOCK_SIZE_M;
-        }
-      }
-    }
-  }
-
-  block_count[0] = tmp_blk_cnt;
-  column_count[0] = tmp_col_cnt;
-}
-
-void convert_vertical_slash_indexes_64x64(
-    const int* q_seqlens,         // [BATCH, ]
-    const int* kv_seqlens,        // [BATCH, ]
-    const int* vertical_indexes,  // [BATCH, N_HEADS, NNZ_V]
-    const int* slash_indexes,     // [BATCH, N_HEADS, NNZ_S]
-    int* block_count,             // [BATCH, N_HEADS, cdiv(N_CTX, BLOCK_SIZE_M)]
-    int* block_offset,  // [BATCH, N_HEADS, cdiv(N_CTX, BLOCK_SIZE_M), NNZ_S]
-    int* column_count,  // [BATCH, N_HEADS, cdiv(N_CTX, BLOCK_SIZE_M)]
-    int* column_index,  // [BATCH, N_HEADS, cdiv(N_CTX, BLOCK_SIZE_M), NNZ_V]
-    int64_t BATCH_SIZE, int64_t N_HEADS, int64_t N_ROWS, int64_t BLOCK_SIZE_M,
-    int64_t BLOCK_SIZE_N, int64_t NNZ_V, int64_t NNZ_S, bool causal) {
-  const int N_THREADS = 64;
-  const dim3 dimBlock(N_THREADS);
-  const dim3 dimGrid(N_HEADS, BATCH_SIZE, (N_ROWS + N_THREADS - 1) / N_THREADS);
-  convert_vertical_slash_indexes_kernel<<<dimGrid, dimBlock>>>(
-      q_seqlens, kv_seqlens, vertical_indexes, slash_indexes, block_count,
-      block_offset, column_count, column_index, N_HEADS, N_ROWS, BLOCK_SIZE_M,
-      BLOCK_SIZE_N, NNZ_V, NNZ_S, causal);
-}
-
-/**
- * Implements the Algorithm 4 in paper https://arxiv.org/abs/2407.02490.
- *
- * This function builds the index of each row of blocks from vertical indices
- * and slash indices. The vertical indices are treated as points, while the
- * slash indices are converted as ranges. The output consists of the merged
- * ranges and separate column indices, where the ranges are represented by
- * block indices.
- *
- * The implementation is referenced from the original MInference repo:
- * https://github.com/microsoft/MInference/blob/main/csrc/vertical_slash_index.cu.
- */
-void convert_vertical_slash_indexes(
-    torch::Tensor& block_count,      // [BATCH, N_HEADS, NUM_ROWS]
-    torch::Tensor& block_offset,     // [BATCH, N_HEADS, NUM_ROWS, NNZ_S]
-    torch::Tensor& column_count,     // [BATCH, N_HEADS, NUM_ROWS]
-    torch::Tensor& column_index,     // [BATCH, N_HEADS, NUM_ROWS, NNZ_V]
-    torch::Tensor q_seqlens,         // [BATCH, ]
-    torch::Tensor kv_seqlens,        // [BATCH, ]
-    torch::Tensor vertical_indexes,  // [BATCH, N_HEADS, NNZ_V]
-    torch::Tensor slash_indexes,     // [BATCH, N_HEADS, NNZ_S]
-    int64_t context_size, int64_t block_size_M, int64_t block_size_N,
-    bool causal) {
-  cudaSetDevice(q_seqlens.get_device());
-
-  int batch_size = slash_indexes.size(0);
-  int num_heads = slash_indexes.size(1);
-  int nnz_slash = slash_indexes.size(2);
-  int nnz_vertical = vertical_indexes.size(2);
-  int num_rows = (context_size + block_size_M - 1) / block_size_M;
-
-  convert_vertical_slash_indexes_64x64(
-      q_seqlens.data_ptr<int>(), kv_seqlens.data_ptr<int>(),
-      vertical_indexes.data_ptr<int>(), slash_indexes.data_ptr<int>(),
-      block_count.data_ptr<int>(), block_offset.data_ptr<int>(),
-      column_count.data_ptr<int>(), column_index.data_ptr<int>(), batch_size,
-      num_heads, num_rows, block_size_M, block_size_N, nnz_vertical, nnz_slash,
-      causal);
-}
-
-__global__ void convert_vertical_slash_indexes_kernel_mergehead(
-    const int* q_seqlens,         // [BATCH, ]
-    const int* kv_seqlens,        // [BATCH, ]
-    const int* vertical_indexes,  // [BATCH, N_HEADS, NNZ_V]
-    const int* slash_indexes,     // [BATCH, N_HEADS, NNZ_S]
-    const int* per_head_vertical_topkv, const int* per_head_slash_topkv,
-    int* block_count,   // [BATCH, N_HEADS, cdiv(N_CTX, BLOCK_SIZE_M)]
-    int* block_offset,  // [BATCH, N_HEADS, cdiv(N_CTX, BLOCK_SIZE_M), NNZ_S]
-    int* column_count,  // [BATCH, N_HEADS, cdiv(N_CTX, BLOCK_SIZE_M)]
-    int* column_index,  // [BATCH, N_HEADS, cdiv(N_CTX, BLOCK_SIZE_M), NNZ_V]
-    int64_t N_HEADS, int64_t N_ROWS, int64_t BLOCK_SIZE_M, int64_t BLOCK_SIZE_N,
-    int64_t NNZ_V, int64_t NNZ_S,
-    bool causal  // True for intra, False for succ
-) {
-  const int batch_idx = blockIdx.y;
-  const int head_idx = blockIdx.x;
-  const int group_idx = blockIdx.z;
-
-  int64_t q_seqlen = q_seqlens[batch_idx];
-  int64_t kv_seqlen = kv_seqlens[batch_idx];
-  int64_t block_idx_m = group_idx * blockDim.x + threadIdx.x;
-  int64_t start_m = block_idx_m * BLOCK_SIZE_M;
-  if (start_m >= q_seqlen) {
-    return;
-  }
-  int64_t end_m = start_m + BLOCK_SIZE_M;
-  vertical_indexes += (batch_idx * N_HEADS + head_idx) * NNZ_V;
-  slash_indexes += (batch_idx * N_HEADS + head_idx) * NNZ_S;
-  int64_t row_offset = (batch_idx * N_HEADS + head_idx) * N_ROWS + block_idx_m;
-  block_count += row_offset;
-  block_offset += row_offset * NNZ_S;
-  column_count += row_offset;
-  column_index += row_offset * NNZ_V;
-
-  // MergeHead: each head has it's unique max topk NNZ_V，NNZ_S. (NNZ_V，NNZ_S
-  // above is buffer size, use to compute offset)
-  NNZ_S = per_head_slash_topkv[head_idx];
-  NNZ_V = per_head_vertical_topkv[head_idx];
-
-  bool has_slash = true;
-  int64_t tmp_col_cnt = 0, tmp_blk_cnt = 0;
-  int64_t s = 0, v = 0;
-  int64_t v_idx = vertical_indexes[v++];
-  int64_t s_idx = slash_indexes[s++];
-  if (causal) {
-    while (s_idx >= end_m + (kv_seqlen - q_seqlen) && s < NNZ_S) {
-      s_idx = slash_indexes[s++];
-    }
-    if (s_idx > end_m + (kv_seqlen - q_seqlen)) has_slash = false;
-    s_idx = max((kv_seqlen - q_seqlen) + end_m - s_idx, BLOCK_SIZE_M);
-  } else {
-    while (s_idx >= end_m + kv_seqlen && s < NNZ_S) {
-      s_idx = slash_indexes[s++];
-    }
-    if (s_idx > end_m + kv_seqlen) has_slash = false;
-    s_idx = max(kv_seqlen + end_m - s_idx, BLOCK_SIZE_M);
-  }
-
-  int64_t range_start = s_idx - BLOCK_SIZE_M, range_end = s_idx;
-  if (!has_slash) {
-    if (causal) {
-      range_start = (kv_seqlen - q_seqlen) + end_m;
-      range_end = (kv_seqlen - q_seqlen) + end_m + BLOCK_SIZE_N;
-    } else {
-      range_start = kv_seqlen;
-      range_end = kv_seqlen + BLOCK_SIZE_N;
-    }
-  }
-
-  bool slash_finished = false;
-  while (1) {
-    if (v_idx < range_end) {
-      if (v_idx < range_start) {
-        column_index[tmp_col_cnt++] = v_idx;
-      }
-      if (v < NNZ_V) {
-        v_idx = vertical_indexes[v++];
-      } else {
-        if (causal)
-          v_idx = end_m + BLOCK_SIZE_N + (kv_seqlen - q_seqlen);
-        else
-          v_idx = end_m + BLOCK_SIZE_N + kv_seqlen;
-      }
-    } else {
-      if ((s < NNZ_S && causal) ||
-          (s < NNZ_S && !causal && slash_indexes[s] >= start_m)) {
-        if (causal)
-          s_idx = max((kv_seqlen - q_seqlen) + end_m - slash_indexes[s++],
-                      BLOCK_SIZE_M);
-        else
-          s_idx = max(kv_seqlen + end_m - slash_indexes[s++], BLOCK_SIZE_M);
-      } else {
-        if (v == NNZ_V || (v_idx > range_start && causal)) {
-          // add the last vertical if no more slash
-          if (v == NNZ_V && !causal && v_idx < kv_seqlen) {
-            column_index[tmp_col_cnt++] = v_idx;
-          }
-          tmp_blk_cnt = save_blocks(block_offset, range_start, range_end,
-                                    BLOCK_SIZE_N, tmp_blk_cnt, kv_seqlen);
-          break;
-        } else {
-          if (causal) {
-            range_start = (kv_seqlen - q_seqlen) + end_m;
-            range_end = (kv_seqlen - q_seqlen) + end_m + BLOCK_SIZE_N;
-          } else {
-            // if slash_finished but there are vertical left, save current
-            // blocks
-            tmp_blk_cnt = save_blocks(block_offset, range_start, range_end,
-                                      BLOCK_SIZE_N, tmp_blk_cnt, kv_seqlen);
-            range_start = kv_seqlen;
-            range_end = kv_seqlen + BLOCK_SIZE_N;
-          }
-          slash_finished = true;
-        }
-      }
-      if (!slash_finished) {
-        if (s_idx > range_end + BLOCK_SIZE_M) {
-          tmp_blk_cnt = save_blocks(block_offset, range_start, range_end,
-                                    BLOCK_SIZE_N, tmp_blk_cnt, kv_seqlen);
-          range_start = s_idx - BLOCK_SIZE_M;
-          range_end = s_idx;
-        } else if (s_idx > range_end) {
-          range_end += BLOCK_SIZE_M;
-        }
-      }
-    }
-  }
-
-  block_count[0] = tmp_blk_cnt;
-  column_count[0] = tmp_col_cnt;
-}
-
-void convert_vertical_slash_indexes_64x64_mergehead(
-    const int* q_seqlens,         // [BATCH, ]
-    const int* kv_seqlens,        // [BATCH, ]
-    const int* vertical_indexes,  // [BATCH, N_HEADS, NNZ_V]
-    const int* slash_indexes,     // [BATCH, N_HEADS, NNZ_S]
-    int* per_head_vertical_topkv, int* per_head_slash_topkv,
-    int* block_count,   // [BATCH, N_HEADS, cdiv(N_CTX, BLOCK_SIZE_M)]
-    int* block_offset,  // [BATCH, N_HEADS, cdiv(N_CTX, BLOCK_SIZE_M), NNZ_S]
-    int* column_count,  // [BATCH, N_HEADS, cdiv(N_CTX, BLOCK_SIZE_M)]
-    int* column_index,  // [BATCH, N_HEADS, cdiv(N_CTX, BLOCK_SIZE_M), NNZ_V]
-    int64_t BATCH_SIZE, int64_t N_HEADS, int64_t N_ROWS, int64_t BLOCK_SIZE_M,
-    int64_t BLOCK_SIZE_N, int64_t NNZ_V, int64_t NNZ_S, bool causal) {
-  const int N_THREADS = 64;
-  const dim3 dimBlock(N_THREADS);
-  const dim3 dimGrid(N_HEADS, BATCH_SIZE, (N_ROWS + N_THREADS - 1) / N_THREADS);
-  convert_vertical_slash_indexes_kernel_mergehead<<<dimGrid, dimBlock>>>(
-      q_seqlens, kv_seqlens, vertical_indexes, slash_indexes,
-      per_head_vertical_topkv, per_head_slash_topkv, block_count, block_offset,
-      column_count, column_index, N_HEADS, N_ROWS, BLOCK_SIZE_M, BLOCK_SIZE_N,
-      NNZ_V, NNZ_S, causal);
-}
-
-/**
- * Implements the Algorithm 4 in paper https://arxiv.org/abs/2407.02490.
- *
- * Like the above convert_vertical_slash_indexes, but with
- * pre-computed vertical and slash counts.
- */
-void convert_vertical_slash_indexes_mergehead(
-    torch::Tensor& block_count,            // [BATCH, N_HEADS, NUM_ROWS]
-    torch::Tensor& block_offset,           // [BATCH, N_HEADS, NUM_ROWS, NNZ_S]
-    torch::Tensor& column_count,           // [BATCH, N_HEADS, NUM_ROWS]
-    torch::Tensor& column_index,           // [BATCH, N_HEADS, NUM_ROWS, NNZ_V]
-    torch::Tensor q_seqlens,               // [BATCH, ]
-    torch::Tensor kv_seqlens,              // [BATCH, ]
-    torch::Tensor vertical_indexes,        // [BATCH, N_HEADS, NNZ_V]
-    torch::Tensor slash_indexes,           // [BATCH, N_HEADS, NNZ_S]
-    torch::Tensor vertical_indices_count,  // [N_HEADS, ]
-    torch::Tensor slash_indices_count,     // [N_HEADS, ]
-    int64_t context_size, int64_t block_size_M, int64_t block_size_N,
-    bool causal) {
-  cudaSetDevice(q_seqlens.get_device());
-
-  int batch_size = slash_indexes.size(0);
-  int num_heads = slash_indexes.size(1);
-  int nnz_slash = slash_indexes.size(2);
-  int nnz_vertical = vertical_indexes.size(2);
-  int num_rows = (context_size + block_size_M - 1) / block_size_M;
-
-  convert_vertical_slash_indexes_64x64_mergehead(
-      q_seqlens.data_ptr<int>(), kv_seqlens.data_ptr<int>(),
-      vertical_indexes.data_ptr<int>(), slash_indexes.data_ptr<int>(),
-      vertical_indices_count.data_ptr<int>(),
-      slash_indices_count.data_ptr<int>(), block_count.data_ptr<int>(),
-      block_offset.data_ptr<int>(), column_count.data_ptr<int>(),
-      column_index.data_ptr<int>(), batch_size, num_heads, num_rows,
-      block_size_M, block_size_N, nnz_vertical, nnz_slash, causal);
-}

csrc/cpu/cpu_attn_impl.hpp

@@ -408,9 +408,19 @@ class AttentionScheduler {
     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;
     const int32_t kv_len_alignment = input.kv_block_alignment;
+    bool has_decode_request = false;
+    bool decode_only_batch = true;
+    for (int32_t req_id = 0; req_id < input.num_reqs; ++req_id) {
+      const int32_t q_token_num =
+          input.query_start_loc[req_id + 1] - input.query_start_loc[req_id];
+      has_decode_request = has_decode_request || (q_token_num == 1);
+      decode_only_batch = decode_only_batch && (q_token_num == 1);
+    }
     int32_t q_head_per_kv = input.num_heads_q / input.num_heads_kv;
-    const bool use_gqa = (max_num_q_per_iter % q_head_per_kv == 0);
-    if (!use_gqa) {
+    const bool supports_gqa = q_head_per_kv <= max_num_q_per_iter;
+    const bool use_gqa_fast_path = supports_gqa && decode_only_batch;
+    const bool use_gqa_scratchpad = supports_gqa && has_decode_request;
+    if (!use_gqa_scratchpad) {
       q_head_per_kv = 1;  // fallback to MHA
     }
     const int32_t min_split_kv_len =
@@ -459,7 +469,7 @@ class AttentionScheduler {
     const int64_t kv_len_per_thread =
         (((total_kv_len / thread_num) + kv_len_alignment - 1) /
          kv_len_alignment) *
-        kv_len_alignment * (use_gqa ? input.num_heads_kv : input.num_heads_q);
+        kv_len_alignment;
     std::vector<AttentionWorkItemGroup> workitems;
     std::vector<ReductionWorkItemGroup> reduce_workitems;
     workitems.reserve(1024);
@@ -680,7 +690,7 @@ class AttentionScheduler {
         metadata_ptr->attention_scratchpad_size_per_thread *
             metadata_ptr->thread_num +
         metadata_ptr->reduction_scratchpad_size_per_kv_head *
-            (use_gqa ? input.num_heads_kv : input.num_heads_q);
+            (use_gqa_fast_path ? input.num_heads_kv : input.num_heads_q);
     cpu_utils::ScratchPadManager::get_scratchpad_manager()->realloc(
         scratchpad_size);
 
@@ -1409,13 +1419,24 @@ class AttentionMainLoop {
       const int32_t q_head_num = input->num_heads;
       const int32_t kv_head_num = input->num_kv_heads;
       const int32_t q_heads_per_kv = q_head_num / kv_head_num;
-      const bool use_gqa =
-          (max_q_head_num_per_iter % q_heads_per_kv == 0) ? true : false;
-      const int32_t actual_kv_head_num = use_gqa ? kv_head_num : q_head_num;
-      const int32_t actual_q_heads_per_kv = use_gqa ? q_heads_per_kv : 1;
+      AttentionWorkItemGroup* const workitem_groups =
+          metadata.workitem_groups_ptr;
+      const int32_t* cu_workitem_num_per_thread =
+          metadata.cu_workitem_num_per_thread;
+      ReductionWorkItemGroup* const reduction_items =
+          metadata.reduction_items_ptr;
+      const bool supports_gqa = q_heads_per_kv <= max_q_head_num_per_iter;
+      bool decode_only_batch = true;
+      for (int32_t i = 0; i < metadata.workitem_group_num; ++i) {
+        decode_only_batch =
+            decode_only_batch && (workitem_groups[i].q_token_num == 1);
+      }
+      const bool use_gqa_fast_path = supports_gqa && decode_only_batch;
+      const int32_t actual_kv_head_num =
+          use_gqa_fast_path ? kv_head_num : q_head_num;
+      const int32_t actual_q_heads_per_kv =
+          use_gqa_fast_path ? q_heads_per_kv : 1;
       TORCH_CHECK_LE(actual_q_heads_per_kv, max_q_head_num_per_iter);
-      const int32_t max_q_token_num_per_iter =
-          max_q_head_num_per_iter / actual_q_heads_per_kv;
       const int64_t q_token_num_stride = input->query_num_tokens_stride;
       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;
@@ -1461,15 +1482,6 @@ class AttentionMainLoop {
               sizeof(q_buffer_t), sizeof(logits_buffer_t),
               sizeof(partial_output_buffer_t), max_q_head_num_per_iter,
               max_q_head_num_per_iter);
-      const int32_t default_q_tile_token_num =
-          default_tile_size / actual_q_heads_per_kv;
-
-      AttentionWorkItemGroup* const workitem_groups =
-          metadata.workitem_groups_ptr;
-      const int32_t* cu_workitem_num_per_thread =
-          metadata.cu_workitem_num_per_thread;
-      ReductionWorkItemGroup* const reduction_items =
-          metadata.reduction_items_ptr;
 
       const int32_t effective_thread_num = metadata.effective_thread_num;
       const int32_t reduction_item_num = metadata.reduction_item_num;
@@ -1513,8 +1525,6 @@ class AttentionMainLoop {
               cu_workitem_num_per_thread[thread_offset + 1] -
               cu_workitem_num_per_thread[thread_offset];
 
-          const int32_t q_head_start_idx = kv_head_idx * actual_q_heads_per_kv;
-
           for (int32_t workitem_group_idx = 0;
                workitem_group_idx < curr_workitem_groups_num;
                ++workitem_group_idx) {
@@ -1529,6 +1539,21 @@ class AttentionMainLoop {
             const int32_t q_token_id_start =
                 current_workitem_group->q_token_id_start;
             const int32_t q_token_num = current_workitem_group->q_token_num;
+            const bool curr_use_gqa =
+                use_gqa_fast_path || (supports_gqa && q_token_num == 1);
+            if (!use_gqa_fast_path && curr_use_gqa &&
+                kv_head_idx % q_heads_per_kv != 0) {
+              continue;
+            }
+            const int32_t curr_q_heads_per_kv =
+                curr_use_gqa ? q_heads_per_kv : 1;
+            const int32_t curr_max_q_token_num_per_iter =
+                max_q_head_num_per_iter / curr_q_heads_per_kv;
+            const int32_t curr_default_q_tile_token_num =
+                default_tile_size / curr_q_heads_per_kv;
+            const int32_t q_head_start_idx =
+                use_gqa_fast_path ? (kv_head_idx * q_heads_per_kv)
+                                  : kv_head_idx;
 
             // taskgroup general information
             const int32_t q_end = input->query_start_loc[current_group_idx + 1];
@@ -1542,7 +1567,7 @@ class AttentionMainLoop {
                              current_workitem_group->local_split_id == 0);
 
             for (int32_t q_token_offset = 0; q_token_offset < q_token_num;
-                 q_token_offset += default_q_tile_token_num) {
+                 q_token_offset += curr_default_q_tile_token_num) {
               bool first_iter_flag[AttentionScheduler::MaxQTileIterNum];
               for (int32_t i = 0; i < AttentionScheduler::MaxQTileIterNum;
                    ++i) {
@@ -1552,9 +1577,9 @@ class AttentionMainLoop {
               const int32_t q_token_start_idx =
                   q_start + q_token_offset + q_token_id_start;
               const int32_t actual_q_token_num = std::min(
-                  default_q_tile_token_num, q_token_num - q_token_offset);
+                  curr_default_q_tile_token_num, q_token_num - q_token_offset);
               const int32_t q_head_tile_size =
-                  actual_q_token_num * actual_q_heads_per_kv;
+                  actual_q_token_num * curr_q_heads_per_kv;
               const int32_t rounded_q_head_tile_size =
                   ((q_head_tile_size + max_q_head_num_per_iter - 1) /
                    max_q_head_num_per_iter) *
@@ -1591,10 +1616,9 @@ class AttentionMainLoop {
                   AttentionScheduler::align_kv_tile_pos(
                       kv_tile_start_pos, kv_tile_end_pos, blocksize_alignment);
 
-              int32_t curr_kv_head_idx =
-                  use_gqa ? kv_head_idx
-                          : (kv_head_idx /
-                             q_heads_per_kv);  // for GQA disabled case
+              const int32_t curr_kv_head_idx =
+                  use_gqa_fast_path ? kv_head_idx
+                                    : (kv_head_idx / q_heads_per_kv);
 
               // std::printf("thread_id: %d, req_id: %d, q_token_start: %d,
               // q_token_end: %d, q_head_start: %d, q_head_end: %d, kv_head_idx:
@@ -1629,12 +1653,12 @@ class AttentionMainLoop {
                   (s_aux != nullptr ? s_aux + q_head_start_idx : nullptr);
 
               // copy the Q tile to q_buffer, the logical layout of q_buffer is
-              // [actual_q_token_num, actual_q_heads_per_kv, head_dim]
+              // [actual_q_token_num, curr_q_heads_per_kv, head_dim]
               {
                 attn_impl.copy_q_heads_tile(
                     q_tile_ptr, q_buffer, actual_q_token_num,
-                    actual_q_heads_per_kv, q_token_num_stride,
-                    q_head_num_stride, scale);
+                    curr_q_heads_per_kv, q_token_num_stride, q_head_num_stride,
+                    scale);
               }
 
               if (use_sink) {
@@ -1648,29 +1672,29 @@ class AttentionMainLoop {
                 float* __restrict__ curr_max_buffer = max_buffer;
                 for (int32_t token_idx = 0; token_idx < actual_q_token_num;
                      ++token_idx) {
-                  for (int32_t head_idx = 0; head_idx < actual_q_heads_per_kv;
+                  for (int32_t head_idx = 0; head_idx < curr_q_heads_per_kv;
                        ++head_idx) {
                     curr_sum_buffer[head_idx] = 1.0f;
                     curr_max_buffer[head_idx] = s_aux_fp32[head_idx];
                   }
 
-                  curr_sum_buffer += actual_q_heads_per_kv;
-                  curr_max_buffer += actual_q_heads_per_kv;
+                  curr_sum_buffer += curr_q_heads_per_kv;
+                  curr_max_buffer += curr_q_heads_per_kv;
                 }
               } else {
                 float* __restrict__ curr_sum_buffer = sum_buffer;
                 float* __restrict__ curr_max_buffer = max_buffer;
                 for (int32_t token_idx = 0; token_idx < actual_q_token_num;
                      ++token_idx) {
-                  for (int32_t head_idx = 0; head_idx < actual_q_heads_per_kv;
+                  for (int32_t head_idx = 0; head_idx < curr_q_heads_per_kv;
                        ++head_idx) {
                     curr_sum_buffer[head_idx] = 0.0f;
                     curr_max_buffer[head_idx] =
                         std::numeric_limits<float>::lowest();
                   }
 
-                  curr_sum_buffer += actual_q_heads_per_kv;
-                  curr_max_buffer += actual_q_heads_per_kv;
+                  curr_sum_buffer += curr_q_heads_per_kv;
+                  curr_max_buffer += curr_q_heads_per_kv;
                 }
               }
 
@@ -1683,16 +1707,17 @@ class AttentionMainLoop {
                     kv_tile_pos_left + kv_tile_size, rounded_kv_tile_end_pos);
                 for (int32_t q_head_tile_token_offset = 0;
                      q_head_tile_token_offset < actual_q_token_num;
-                     q_head_tile_token_offset += max_q_token_num_per_iter) {
+                     q_head_tile_token_offset +=
+                     curr_max_q_token_num_per_iter) {
                   const int32_t q_tile_pos_left =
                       q_tile_start_pos + q_head_tile_token_offset;
                   const int32_t q_tile_token_num =
-                      std::min(max_q_token_num_per_iter,
+                      std::min(curr_max_q_token_num_per_iter,
                                actual_q_token_num - q_head_tile_token_offset);
                   const int32_t q_tile_head_offset =
-                      q_head_tile_token_offset * actual_q_heads_per_kv;
+                      q_head_tile_token_offset * curr_q_heads_per_kv;
                   const int32_t q_tile_head_num =
-                      q_tile_token_num * actual_q_heads_per_kv;
+                      q_tile_token_num * curr_q_heads_per_kv;
                   const int32_t q_tile_pos_right =
                       q_tile_pos_left + q_tile_token_num;
                   const auto [actual_kv_tile_pos_left,
@@ -1702,7 +1727,7 @@ class AttentionMainLoop {
                           q_tile_pos_right, sliding_window_left,
                           sliding_window_right);
                   const int32_t q_iter_idx =
-                      q_head_tile_token_offset / max_q_token_num_per_iter;
+                      q_head_tile_token_offset / curr_max_q_token_num_per_iter;
 
                   if (actual_kv_tile_pos_right <= actual_kv_tile_pos_left) {
                     continue;
@@ -1768,7 +1793,7 @@ class AttentionMainLoop {
                       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,
-                      q_tile_token_num, q_tile_pos_left, actual_q_heads_per_kv,
+                      q_tile_token_num, q_tile_pos_left, curr_q_heads_per_kv,
                       block_size, sliding_window_left, sliding_window_right,
                       scale, softcap_scale, curr_alibi_slopes,
                       first_iter_flag[q_iter_idx], use_sink, debug_info);
@@ -1782,11 +1807,11 @@ class AttentionMainLoop {
                   final_output(partial_q_buffer,
                                reinterpret_cast<query_t*>(input->output) +
                                    output_buffer_offset,
-                               sum_buffer, actual_q_heads_per_kv,
+                               sum_buffer, curr_q_heads_per_kv,
                                actual_q_token_num, q_head_num, output_v_scale);
                 } else {
                   const int32_t stride =
-                      actual_q_heads_per_kv * split_kv_q_token_num_threshold;
+                      curr_q_heads_per_kv * split_kv_q_token_num_threshold;
                   buffer_manager.update(kv_head_idx, total_reduction_split_num,
                                         head_dim, stride, sizeof(float));
                   volatile bool* split_flag_buffer =
@@ -1822,18 +1847,26 @@ class AttentionMainLoop {
           const int32_t curr_split_id = curr_workitem_groups->split_start_id;
           const int32_t curr_split_num = curr_workitem_groups->split_num;
           const int32_t current_group_idx = curr_workitem_groups->req_id;
+          const bool curr_use_gqa =
+              use_gqa_fast_path || (supports_gqa && curr_output_token_num == 1);
+          if (!use_gqa_fast_path && curr_use_gqa &&
+              kv_head_idx % q_heads_per_kv != 0) {
+            continue;
+          }
+          const int32_t curr_q_heads_per_kv = curr_use_gqa ? q_heads_per_kv : 1;
           const int32_t curr_output_head_num =
-              curr_output_token_num * actual_q_heads_per_kv;
+              curr_output_token_num * curr_q_heads_per_kv;
 
           const int32_t q_start = input->query_start_loc[current_group_idx];
           const int32_t q_token_start_idx = q_start + curr_output_token_idx;
-          const int32_t q_head_start_idx = kv_head_idx * actual_q_heads_per_kv;
+          const int32_t q_head_start_idx =
+              use_gqa_fast_path ? (kv_head_idx * q_heads_per_kv) : kv_head_idx;
           size_t output_buffer_offset =
               q_token_start_idx * q_head_num * head_dim +
               q_head_start_idx * head_dim;
 
           const int32_t stride =
-              actual_q_heads_per_kv * split_kv_q_token_num_threshold;
+              curr_q_heads_per_kv * split_kv_q_token_num_threshold;
           buffer_manager.update(kv_head_idx, total_reduction_split_num,
                                 head_dim, stride, sizeof(float));
           volatile bool* split_flag_buffer =
@@ -1852,7 +1885,7 @@ class AttentionMainLoop {
           final_output(
               split_output_buffer,
               reinterpret_cast<query_t*>(input->output) + output_buffer_offset,
-              split_sum_buffer, actual_q_heads_per_kv, curr_output_token_num,
+              split_sum_buffer, curr_q_heads_per_kv, curr_output_token_num,
               q_head_num, output_v_scale);
         }
       }

csrc/cpu/cpu_attn_rvv.hpp

@@ -4,17 +4,18 @@
 #ifndef CPU_ATTN_RVV_HPP
 #define CPU_ATTN_RVV_HPP
 
-// This kernel is currently hardcoded to VLEN=128 (m1/m2 intrinsics, vl=8).
-// The fixed-width typedefs below use `riscv_rvv_vector_bits(128)`, which
-// only matches `vfloat16m1_t`/`vuint16m1_t` register layout when VLEN==128;
-// at VLEN>=256 those typedefs fail to compile.  Scalar RISC-V builds
-// (-march=rv64gc) additionally don't have <riscv_vector.h>.  For both
-// cases we omit the file entirely and let the dispatcher fall back to the
-// scalar VEC / VEC16 implementations.  TODO: migrate to RVVI() macros +
-// semantic names in cpu_types_riscv_defs.hpp to support VLEN>=256 natively.
-#if defined(__riscv_v_min_vlen) && __riscv_v_min_vlen == 128
+// RVV attention kernel using VLEN-agnostic RVVI() macros from
+// cpu_types_riscv_defs.hpp.  The Mx8 tile GEMM uses 8 FP32 elements
+// per vector (LMUL_256 bits of FP32 data), which maps to:
+//   VLEN=128: m2 (256 bits = 8 x FP32)
+//   VLEN=256: m1 (256 bits = 8 x FP32)
+// Only VLEN=128 and VLEN=256 are supported; other VLENs (512, 1024)
+// and scalar RISC-V builds fall back to VEC/VEC16.
+#if defined(__riscv_v_min_vlen) && \
+    (__riscv_v_min_vlen == 128 || __riscv_v_min_vlen == 256)
 
   #include "cpu_attn_impl.hpp"
+  #include "cpu_types_riscv_defs.hpp"
   #include <riscv_vector.h>
   #include <type_traits>
 
@@ -22,73 +23,50 @@ namespace cpu_attention {
 
 namespace {
 
-// File-local concrete-LMUL typedefs.  The shared _defs.hpp exposes
-// VLEN-independent semantic names (fixed_fp32x8_t, fixed_fp16x8_t, ...),
-// but this kernel is currently hardcoded to VLEN=128 (m1/m2 intrinsics),
-// so keep the legacy concrete aliases scoped to this file.
-typedef vfloat16m1_t fixed_vfloat16m1_t
-    __attribute__((riscv_rvv_vector_bits(128)));
-typedef vfloat32m2_t fixed_vfloat32m2_t
-    __attribute__((riscv_rvv_vector_bits(256)));
-typedef vuint16m1_t fixed_vuint16m1_t
-    __attribute__((riscv_rvv_vector_bits(128)));
-typedef vuint32m2_t fixed_vuint32m2_t
-    __attribute__((riscv_rvv_vector_bits(256)));
-  #ifdef __riscv_zvfbfmin
-typedef vbfloat16m1_t fixed_vbfloat16m1_t
-    __attribute__((riscv_rvv_vector_bits(128)));
-  #endif
-
   #define BLOCK_SIZE_ALIGNMENT 32
   #define HEAD_SIZE_ALIGNMENT 32
   #define MAX_Q_HEAD_NUM_PER_ITER 16
 
 // ============================================================================
-// B-matrix row loading: load 8 elements as FP32 (using m2 LMUL at VLEN=128)
+// B-matrix row loading: load 8 elements as FP32
 // ============================================================================
 
 template <typename kv_cache_t>
-FORCE_INLINE fixed_vfloat32m2_t load_row8_B_as_f32(const kv_cache_t* p);
+FORCE_INLINE fixed_fp32x8_t load_row8_B_as_f32(const kv_cache_t* p);
 
 template <>
-FORCE_INLINE fixed_vfloat32m2_t load_row8_B_as_f32<float>(const float* p) {
-  return __riscv_vle32_v_f32m2(p, 8);
+FORCE_INLINE fixed_fp32x8_t load_row8_B_as_f32<float>(const float* p) {
+  return RVVI(__riscv_vle32_v_f32, LMUL_256)(p, 8);
 }
 
 template <>
-FORCE_INLINE fixed_vfloat32m2_t
-load_row8_B_as_f32<c10::Half>(const c10::Half* p) {
+FORCE_INLINE fixed_fp32x8_t load_row8_B_as_f32<c10::Half>(const c10::Half* p) {
   #ifdef __riscv_zvfh
-  fixed_vfloat16m1_t h =
-      __riscv_vle16_v_f16m1(reinterpret_cast<const _Float16*>(p), 8);
-  return __riscv_vfwcvt_f_f_v_f32m2(h, 8);
+  fixed_fp16x8_t h = RVVI(__riscv_vle16_v_f16, LMUL_128)(
+      reinterpret_cast<const _Float16*>(p), 8);
+  return RVVI(__riscv_vfwcvt_f_f_v_f32, LMUL_256)(h, 8);
   #else
-  // Fallback for hardware without Zvfh: scalar half->float conversion.
-  // c10::Half provides operator float() so this is correct on any RVV CPU
-  // that has only the base V extension.  Slower than the Zvfh path, but
-  // keeps the kernel buildable on Zvfhmin-only / no-fp16 hardware.
   alignas(16) float tmp[8];
   for (int i = 0; i < 8; ++i) {
     tmp[i] = static_cast<float>(p[i]);
   }
-  return __riscv_vle32_v_f32m2(tmp, 8);
+  return RVVI(__riscv_vle32_v_f32, LMUL_256)(tmp, 8);
   #endif
 }
 
 template <>
-FORCE_INLINE fixed_vfloat32m2_t
+FORCE_INLINE fixed_fp32x8_t
 load_row8_B_as_f32<c10::BFloat16>(const c10::BFloat16* p) {
   #ifdef __riscv_zvfbfmin
-  fixed_vbfloat16m1_t bf =
-      __riscv_vle16_v_bf16m1(reinterpret_cast<const __bf16*>(p), 8);
-  return __riscv_vfwcvtbf16_f_f_v_f32m2(bf, 8);
+  fixed_bf16x8_t bf = RVVI(__riscv_vle16_v_bf16, LMUL_128)(
+      reinterpret_cast<const __bf16*>(p), 8);
+  return RVVI(__riscv_vfwcvtbf16_f_f_v_f32, LMUL_256)(bf, 8);
   #else
-  // Fallback: load as uint16, zero-extend to uint32, shift left by 16
-  fixed_vuint16m1_t raw =
-      __riscv_vle16_v_u16m1(reinterpret_cast<const uint16_t*>(p), 8);
-  fixed_vuint32m2_t wide = __riscv_vzext_vf2_u32m2(raw, 8);
-  fixed_vuint32m2_t shifted = __riscv_vsll_vx_u32m2(wide, 16, 8);
-  return __riscv_vreinterpret_v_u32m2_f32m2(shifted);
+  fixed_u16x8_t raw = RVVI(__riscv_vle16_v_u16, LMUL_128)(
+      reinterpret_cast<const uint16_t*>(p), 8);
+  fixed_u32x8_t wide = RVVI(__riscv_vzext_vf2_u32, LMUL_256)(raw, 8);
+  fixed_u32x8_t shifted = RVVI(__riscv_vsll_vx_u32, LMUL_256)(wide, 16, 8);
+  return RVVI4(__riscv_vreinterpret_v_u32, LMUL_256, _f32, LMUL_256)(shifted);
   #endif
 }
 
@@ -96,14 +74,12 @@ load_row8_B_as_f32<c10::BFloat16>(const c10::BFloat16* p) {
 // Micro kernel: Mx8 tile, K unrolled by 4, RVV scalar-broadcast FMA
 // ============================================================================
 //
-// NEON uses vfmaq_laneq_f32 (lane-indexed FMA from a preloaded A vector).
 // RVV has no lane-indexed FMA; instead we load A elements as scalars and
-// use __riscv_vfmacc_vf (scalar * vector + accumulator), which is equally
-// efficient and avoids the need for vrgather/vslidedown.
+// use vfmacc_vf (scalar * vector + accumulator).
 //
-// At VLEN=128, m2 holds 8 x FP32, matching the 8-column tile width.
-// Register budget: M accumulators (m2 each) + 1 B temp = 2M+2 regs.
-// M=8 => 18 regs out of 32 available — no spills.
+// The 8-column tile uses LMUL_256 bits of FP32 data:
+//   VLEN=128: m2 (2 regs per accumulator), M=8 => 18 of 32 regs
+//   VLEN=256: m1 (1 reg  per accumulator), M=8 =>  9 of 32 regs
 
 template <int32_t M, typename kv_cache_t>
 FORCE_INLINE void gemm_micro_rvv_fma_Mx8_Ku4(
@@ -115,94 +91,90 @@ FORCE_INLINE void gemm_micro_rvv_fma_Mx8_Ku4(
 
   constexpr size_t vl = 8;
 
-  // helpers for per-M codegen
   #define ROWS_APPLY(OP) OP(0) OP(1) OP(2) OP(3) OP(4) OP(5) OP(6) OP(7)
   #define IF_M(i) if constexpr (M > (i))
 
-    // A row base pointers
   #define DECL_A(i) const float* a##i = A + (i) * lda;
   ROWS_APPLY(DECL_A)
   #undef DECL_A
 
-    // declare one m2 accumulator per row
-  #define DECL_ACC(i) fixed_vfloat32m2_t acc##i;
+  #define DECL_ACC(i) fixed_fp32x8_t acc##i;
   ROWS_APPLY(DECL_ACC)
   #undef DECL_ACC
 
-    // initialize accumulators
-  #define INIT_ACC(i)                                      \
-    IF_M(i) {                                              \
-      if (accumulate) {                                    \
-        acc##i = __riscv_vle32_v_f32m2(C + (i) * ldc, vl); \
-      } else {                                             \
-        acc##i = __riscv_vfmv_v_f_f32m2(0.f, vl);          \
-      }                                                    \
+  #define INIT_ACC(i)                                                    \
+    IF_M(i) {                                                            \
+      if (accumulate) {                                                  \
+        acc##i = RVVI(__riscv_vle32_v_f32, LMUL_256)(C + (i) * ldc, vl); \
+      } else {                                                           \
+        acc##i = RVVI(__riscv_vfmv_v_f_f32, LMUL_256)(0.f, vl);          \
+      }                                                                  \
     }
   ROWS_APPLY(INIT_ACC)
   #undef INIT_ACC
 
   int32_t k = 0;
 
-  // K unrolled by 4
   for (; k + 3 < K; k += 4) {
-    // k + 0
     {
-      fixed_vfloat32m2_t b =
+      fixed_fp32x8_t b =
           load_row8_B_as_f32<kv_cache_t>(B + (int64_t)(k + 0) * ldb);
-  #define STEP_K0(i)                                                    \
-    IF_M(i) {                                                           \
-      acc##i = __riscv_vfmacc_vf_f32m2(acc##i, *(a##i + k + 0), b, vl); \
+  #define STEP_K0(i)                                                          \
+    IF_M(i) {                                                                 \
+      acc##i = RVVI(__riscv_vfmacc_vf_f32, LMUL_256)(acc##i, *(a##i + k + 0), \
+                                                     b, vl);                  \
     }
       ROWS_APPLY(STEP_K0)
   #undef STEP_K0
     }
-    // k + 1
     {
-      fixed_vfloat32m2_t b =
+      fixed_fp32x8_t b =
           load_row8_B_as_f32<kv_cache_t>(B + (int64_t)(k + 1) * ldb);
-  #define STEP_K1(i)                                                    \
-    IF_M(i) {                                                           \
-      acc##i = __riscv_vfmacc_vf_f32m2(acc##i, *(a##i + k + 1), b, vl); \
+  #define STEP_K1(i)                                                          \
+    IF_M(i) {                                                                 \
+      acc##i = RVVI(__riscv_vfmacc_vf_f32, LMUL_256)(acc##i, *(a##i + k + 1), \
+                                                     b, vl);                  \
     }
       ROWS_APPLY(STEP_K1)
   #undef STEP_K1
     }
-    // k + 2
     {
-      fixed_vfloat32m2_t b =
+      fixed_fp32x8_t b =
           load_row8_B_as_f32<kv_cache_t>(B + (int64_t)(k + 2) * ldb);
-  #define STEP_K2(i)                                                    \
-    IF_M(i) {                                                           \
-      acc##i = __riscv_vfmacc_vf_f32m2(acc##i, *(a##i + k + 2), b, vl); \
+  #define STEP_K2(i)                                                          \
+    IF_M(i) {                                                                 \
+      acc##i = RVVI(__riscv_vfmacc_vf_f32, LMUL_256)(acc##i, *(a##i + k + 2), \
+                                                     b, vl);                  \
     }
       ROWS_APPLY(STEP_K2)
   #undef STEP_K2
     }
-    // k + 3
     {
-      fixed_vfloat32m2_t b =
+      fixed_fp32x8_t b =
           load_row8_B_as_f32<kv_cache_t>(B + (int64_t)(k + 3) * ldb);
-  #define STEP_K3(i)                                                    \
-    IF_M(i) {                                                           \
-      acc##i = __riscv_vfmacc_vf_f32m2(acc##i, *(a##i + k + 3), b, vl); \
+  #define STEP_K3(i)                                                          \
+    IF_M(i) {                                                                 \
+      acc##i = RVVI(__riscv_vfmacc_vf_f32, LMUL_256)(acc##i, *(a##i + k + 3), \
+                                                     b, vl);                  \
     }
       ROWS_APPLY(STEP_K3)
   #undef STEP_K3
     }
   }
 
-  // K tail
   for (; k < K; ++k) {
-    fixed_vfloat32m2_t b = load_row8_B_as_f32<kv_cache_t>(B + (int64_t)k * ldb);
-  #define TAIL_ROW(i) \
-    IF_M(i) { acc##i = __riscv_vfmacc_vf_f32m2(acc##i, *(a##i + k), b, vl); }
+    fixed_fp32x8_t b = load_row8_B_as_f32<kv_cache_t>(B + (int64_t)k * ldb);
+  #define TAIL_ROW(i)                                                        \
+    IF_M(i) {                                                                \
+      acc##i =                                                               \
+          RVVI(__riscv_vfmacc_vf_f32, LMUL_256)(acc##i, *(a##i + k), b, vl); \
+    }
     ROWS_APPLY(TAIL_ROW)
   #undef TAIL_ROW
   }
 
-    // store accumulators to C
   #define STORE_ROW(i) \
-    IF_M(i) { __riscv_vse32_v_f32m2(C + (i) * ldc, acc##i, vl); }
+    IF_M(i) { RVVI(__riscv_vse32_v_f32, LMUL_256)(C + (i) * ldc, acc##i, vl); }
   ROWS_APPLY(STORE_ROW)
   #undef STORE_ROW
 
@@ -381,7 +353,6 @@ class AttentionImpl<ISA::RVV, scalar_t, head_dim, kv_cache_scalar_t> {
         const int64_t block_idx = pos / block_size;
         const int64_t block_offset = pos % block_size;
         {
-          // Write Key (transpose to column-major: [head_dim, block_size])
           const scalar_t* key_start_ptr = key +
                                           token_idx * key_token_num_stride +
                                           head_idx * key_head_num_stride;
@@ -389,8 +360,6 @@ class AttentionImpl<ISA::RVV, scalar_t, head_dim, kv_cache_scalar_t> {
               key_cache + block_idx * num_blocks_stride +
               head_idx * cache_head_num_stride + block_offset;
 
-          // Strided vector store for efficient transpose.
-          // Load contiguous key elements, store with stride = block_size.
           {
             const ptrdiff_t byte_stride = block_size * sizeof(scalar_t);
             int64_t i = 0;
@@ -405,7 +374,6 @@ class AttentionImpl<ISA::RVV, scalar_t, head_dim, kv_cache_scalar_t> {
                                              i * block_size),
                     byte_stride, v, vl);
               } else {
-                // Half and BFloat16 are both 16-bit types
                 vl = __riscv_vsetvl_e16m1(head_dim - i);
                 vuint16m1_t v = __riscv_vle16_v_u16m1(
                     reinterpret_cast<const uint16_t*>(key_start_ptr + i), vl);
@@ -419,7 +387,6 @@ class AttentionImpl<ISA::RVV, scalar_t, head_dim, kv_cache_scalar_t> {
           }
         }
         {
-          // Write Value (row-major: [block_size, head_dim])
           const scalar_t* value_start_ptr = value +
                                             token_idx * value_token_num_stride +
                                             head_idx * value_head_num_stride;
@@ -440,6 +407,6 @@ class AttentionImpl<ISA::RVV, scalar_t, head_dim, kv_cache_scalar_t> {
   #undef HEAD_SIZE_ALIGNMENT
   #undef MAX_Q_HEAD_NUM_PER_ITER
 
-#endif  // __riscv_v_min_vlen == 128
+#endif  // __riscv_v_min_vlen == 128 || 256
 
 #endif  // CPU_ATTN_RVV_HPP

csrc/cpu/cpu_types_riscv_defs.hpp

@@ -71,6 +71,10 @@ typedef RVVTYPE(vuint16, LMUL_256, _t) fixed_u16x16_t
 typedef RVVTYPE(vuint16, LMUL_512, _t) fixed_u16x32_t
     __attribute__((riscv_rvv_vector_bits(512)));
 
+// uint32
+typedef RVVTYPE(vuint32, LMUL_256, _t) fixed_u32x8_t
+    __attribute__((riscv_rvv_vector_bits(256)));
+
 // bfloat16
 #ifdef __riscv_zvfbfmin
 typedef RVVTYPE(vbfloat16, LMUL_128, _t) fixed_bf16x8_t

csrc/cpu/cpu_types_riscv_impl.hpp

@@ -94,6 +94,10 @@ struct FP16Vec16 : public Vec<FP16Vec16> {
       : reg(RVVI(__riscv_vle16_v_f16, LMUL_256)(
             static_cast<const _Float16*>(ptr), VEC_ELEM_NUM)) {};
 
+  explicit FP16Vec16(const c10::Half v)
+      : reg(RVVI4(__riscv_vreinterpret_v_u16, LMUL_256, _f16, LMUL_256)(
+            RVVI(__riscv_vmv_v_x_u16, LMUL_256)(v.x, VEC_ELEM_NUM))) {};
+
   explicit FP16Vec16(const FP32Vec16& vec);
 
   void save(void* ptr) const {
@@ -165,6 +169,9 @@ struct BF16Vec16 : public Vec<BF16Vec16> {
             reinterpret_cast<const uint16_t*>(ptr), VEC_ELEM_NUM))) {};
 
   explicit BF16Vec16(fixed_bf16x16_t data) : reg(data) {};
+  explicit BF16Vec16(const c10::BFloat16 v)
+      : reg(RVVI4(__riscv_vreinterpret_v_u16, LMUL_256, _bf16, LMUL_256)(
+            RVVI(__riscv_vmv_v_x_u16, LMUL_256)(v.x, VEC_ELEM_NUM))) {};
   explicit BF16Vec16(const FP32Vec16&);
 
   void save(void* ptr) const {
@@ -290,6 +297,9 @@ struct BF16Vec16 : public Vec<BF16Vec16> {
     }
     reg_fp32 = RVVI(__riscv_vle32_v_f32, LMUL_512)(tmp, 16);
   }
+  explicit BF16Vec16(const c10::BFloat16 v)
+      : reg_fp32(RVVI(__riscv_vfmv_v_f_f32, LMUL_512)(static_cast<float>(v),
+                                                      VEC_ELEM_NUM)) {}
   explicit BF16Vec16(const FP32Vec16&);
   void save(void* ptr) const {
     float tmp[16];
@@ -629,6 +639,19 @@ struct FP32Vec16 : public Vec<FP32Vec16> {
       : reg(RVVI4(__riscv_vcreate_v_f32, LMUL_256, _f32, LMUL_512)(
             data.reg, data.reg)) {};
   explicit FP32Vec16(const FP32Vec16& data) : reg(data.reg) {};
+  explicit FP32Vec16(int64_t value, const FP32Vec16& lut) {
+    const uint64_t q_values = static_cast<uint64_t>(value);
+    auto packed = RVVI(__riscv_vmv_v_x_u64, LMUL_1024)(q_values, VEC_ELEM_NUM);
+    auto lane_ids = RVVI(__riscv_vid_v_u64, LMUL_1024)(VEC_ELEM_NUM);
+    auto shifts =
+        RVVI(__riscv_vsll_vx_u64, LMUL_1024)(lane_ids, 2, VEC_ELEM_NUM);
+    auto shifted =
+        RVVI(__riscv_vsrl_vv_u64, LMUL_1024)(packed, shifts, VEC_ELEM_NUM);
+    auto idx64 =
+        RVVI(__riscv_vand_vx_u64, LMUL_1024)(shifted, 0xF, VEC_ELEM_NUM);
+    auto idx32 = RVVI(__riscv_vnsrl_wx_u32, LMUL_512)(idx64, 0, VEC_ELEM_NUM);
+    reg = RVVI(__riscv_vrgather_vv_f32, LMUL_512)(lut.reg, idx32, VEC_ELEM_NUM);
+  }
   explicit FP32Vec16(const FP16Vec16& v);
 
 #ifdef __riscv_zvfbfmin
@@ -641,6 +664,10 @@ struct FP32Vec16 : public Vec<FP32Vec16> {
   explicit FP32Vec16(const BF16Vec16& v) : reg(v.reg_fp32) {};
 #endif
 
+  // FP8 stub: dead code on RISC-V (fp8 KV cache is x86-only), needed for
+  // load_b_pair_vec template to compile on all platforms.
+  explicit FP32Vec16(const BF16Vec32&, int) : FP32Vec16() {}
+
   FP32Vec16 operator+(const FP32Vec16& b) const {
     return FP32Vec16(
         RVVI(__riscv_vfadd_vv_f32, LMUL_512)(reg, b.reg, VEC_ELEM_NUM));
@@ -891,6 +918,30 @@ inline void fma(FP32Vec16& acc, const FP32Vec16& a, const FP32Vec16& b) {
   acc = acc.fma(a, b);
 }
 
+template <typename VecT>
+static void interleave_save_16b(const VecT& vec0, const VecT& vec1, void* ptr) {
+  alignas(64) uint16_t values0[VecT::VEC_ELEM_NUM];
+  alignas(64) uint16_t values1[VecT::VEC_ELEM_NUM];
+  vec0.save(values0);
+  vec1.save(values1);
+
+  auto* packed = reinterpret_cast<uint32_t*>(ptr);
+  for (int32_t i = 0; i < VecT::VEC_ELEM_NUM; ++i) {
+    packed[i] = static_cast<uint32_t>(values0[i]) |
+                (static_cast<uint32_t>(values1[i]) << 16);
+  }
+}
+
+static void interleave_save(const FP16Vec16& vec0, const FP16Vec16& vec1,
+                            void* ptr) {
+  interleave_save_16b(vec0, vec1, ptr);
+}
+
+static void interleave_save(const BF16Vec16& vec0, const BF16Vec16& vec1,
+                            void* ptr) {
+  interleave_save_16b(vec0, vec1, ptr);
+}
+
 #ifdef __riscv_zvfbfmin
 template <>
 inline void storeFP32<c10::BFloat16>(float v, c10::BFloat16* ptr) {

csrc/cpu/cpu_types_vsx.hpp

@@ -89,6 +89,35 @@ struct BF16Vec8 : public Vec<BF16Vec8> {
   }
 };
 
+struct FP16Vec16 : public Vec<FP16Vec16> {
+  constexpr static int VEC_ELEM_NUM = 16;
+  ss16x8x2_t reg;
+
+  explicit FP16Vec16(const void* ptr) {
+    reg.val[0] = (__vector signed short)vec_xl(0, (signed short*)ptr);
+    reg.val[1] = (__vector signed short)vec_xl(16, (signed short*)ptr);
+  }
+
+  explicit FP16Vec16(bool, const void* ptr) : FP16Vec16(ptr) {}
+
+  explicit FP16Vec16(const FP32Vec16&);
+
+  void save(void* ptr) const {
+    vec_xst(reg.val[0], 0, (signed short*)ptr);
+    vec_xst(reg.val[1], 16, (signed short*)ptr);
+  }
+
+  void save(void* ptr, int elem_num) const {
+    int num = std::max(0, std::min(elem_num, VEC_ELEM_NUM));
+    if (num <= 8) {
+      vec_xst_len(reg.val[0], (signed short*)ptr, num * 2);
+    } else {
+      vec_xst(reg.val[0], 0, (signed short*)ptr);
+      vec_xst_len(reg.val[1], (signed short*)ptr + 8, (num - 8) * 2);
+    }
+  }
+};
+
 struct BF16Vec16 : public Vec<BF16Vec16> {
   constexpr static int VEC_ELEM_NUM = 16;
 
@@ -100,6 +129,8 @@ struct BF16Vec16 : public Vec<BF16Vec16> {
     reg.val[1] = (__vector signed short)vec_xl(16, (signed short*)ptr);
   }
 
+  explicit BF16Vec16(bool, const void* ptr) : BF16Vec16(ptr) {}
+
   explicit BF16Vec16(const FP32Vec16&);
 
   void save(void* ptr) const {
@@ -379,6 +410,8 @@ struct FP32Vec16 : public Vec<FP32Vec16> {
     reg.val[3] = vec_xl(48, ptr);
   }
 
+  explicit FP32Vec16(bool, const float* ptr) : FP32Vec16(ptr) {}
+
   explicit FP32Vec16(f32x4x4_t data) : reg(data) {}
 
   explicit FP32Vec16(const FP32Vec16& data) {
@@ -402,6 +435,7 @@ struct FP32Vec16 : public Vec<FP32Vec16> {
     reg.val[3] = data.reg.val[1];
   }
 
+  explicit FP32Vec16(const FP16Vec16& v);
   explicit FP32Vec16(const BF16Vec16& v) {
     reg.val[0] = (__vector float)vec_mergeh(zero, v.reg.val[0]);
     reg.val[1] = (__vector float)vec_mergel(zero, v.reg.val[0]);
@@ -735,6 +769,40 @@ inline BF16Vec8::BF16Vec8(const FP32Vec8& v) {
 #endif
 }
 
+inline FP16Vec16::FP16Vec16(const FP32Vec16& v) {
+  alignas(16) float temp_fp32[16];
+  alignas(16) c10::Half temp_fp16[16];
+
+  vec_xst(v.reg.val[0], 0, temp_fp32);
+  vec_xst(v.reg.val[1], 16, temp_fp32);
+  vec_xst(v.reg.val[2], 32, temp_fp32);
+  vec_xst(v.reg.val[3], 48, temp_fp32);
+
+  for (int i = 0; i < 16; i++) {
+    temp_fp16[i] = c10::Half(temp_fp32[i]);
+  }
+
+  reg.val[0] = (__vector signed short)vec_xl(0, (signed short*)temp_fp16);
+  reg.val[1] = (__vector signed short)vec_xl(16, (signed short*)temp_fp16);
+}
+
+inline FP32Vec16::FP32Vec16(const FP16Vec16& v) {
+  alignas(16) c10::Half temp_fp16[16];
+  alignas(16) float temp_fp32[16];
+
+  vec_xst(v.reg.val[0], 0, (signed short*)temp_fp16);
+  vec_xst(v.reg.val[1], 16, (signed short*)temp_fp16);
+
+  for (int i = 0; i < 16; i++) {
+    temp_fp32[i] = float(temp_fp16[i]);
+  }
+
+  reg.val[0] = vec_xl(0, temp_fp32);
+  reg.val[1] = vec_xl(16, temp_fp32);
+  reg.val[2] = vec_xl(32, temp_fp32);
+  reg.val[3] = vec_xl(48, temp_fp32);
+}
+
 inline BF16Vec16::BF16Vec16(const FP32Vec16& v) {
 #ifdef _ARCH_PWR10
   __vector signed short ret[4];
@@ -794,6 +862,43 @@ inline void prefetch(const void* addr) {
   __asm__ __volatile__("dcbt 0, %0" : : "r"(addr) : "memory");
 }
 
-};  // namespace vec_op
+struct INT8Vec64 {
+  __vector signed char data[4];
+
+  INT8Vec64() = default;
+
+  explicit INT8Vec64(const int8_t* ptr) {
+    data[0] = vec_xl(0, ptr);
+    data[1] = vec_xl(16, ptr);
+    data[2] = vec_xl(32, ptr);
+    data[3] = vec_xl(48, ptr);
+  }
+
+  explicit INT8Vec64(bool, const int8_t* ptr) : INT8Vec64(ptr) {}
+
+  void save(int8_t* ptr) const {
+    vec_xst(data[0], 0, ptr);
+    vec_xst(data[1], 16, ptr);
+    vec_xst(data[2], 32, ptr);
+    vec_xst(data[3], 48, ptr);
+  }
+
+  void save(int8_t* ptr, int elem_num) const {
+    if (elem_num <= 0) return;
+
+    int full_vecs = elem_num / 16;
+    for (int i = 0; i < full_vecs && i < 4; i++) {
+      vec_xst(data[i], i * 16, ptr);
+    }
+
+    int remaining = elem_num % 16;
+    if (remaining > 0 && full_vecs < 4) {
+      vec_xst_len(data[full_vecs], ptr + full_vecs * 16, remaining);
+    }
+  }
+
+  void nt_save(int8_t* ptr) const { save(ptr); }
+};
+}  // namespace vec_op
 
 #endif

csrc/cpu/generate_cpu_attn_dispatch.py

@@ -150,12 +150,10 @@ def generate_header_file() -> str:
   #include "cpu_attn_vxe.hpp"
 #endif
 
-// cpu_attn_rvv.hpp is hardcoded to VLEN==128 (m1/m2 intrinsics, vl=8) and
-// itself includes <riscv_vector.h>, which is unavailable on scalar
-// (-march=rv64gc) builds. Gate the include the same way as the dispatch
-// macro below, so non-128 / scalar RISC-V builds skip it entirely.
+// cpu_attn_rvv.hpp supports VLEN=128 and VLEN=256 via RVVI() macros.
+// Other VLENs and scalar RISC-V builds skip it entirely.
 #if defined(__riscv) && defined(__riscv_v_min_vlen) && \
-    __riscv_v_min_vlen == 128
+    (__riscv_v_min_vlen == 128 || __riscv_v_min_vlen == 256)
   #include "cpu_attn_rvv.hpp"
 #endif
 
@@ -222,15 +220,12 @@ def generate_header_file() -> str:
         ["VXE", "VEC", "VEC16"],
         fp8=False,
     )
-    # RISC-V with RVV.  cpu_attn_rvv.hpp is hardcoded to VLEN==128
-    # (riscv_rvv_vector_bits(128) typedefs + vl=8 m1/m2 intrinsics), so
-    # we split the dispatch into two top-level branches: VLEN==128 builds
-    # get the full RVV+VEC+VEC16 case set, other VLEN builds get a
-    # VEC/VEC16-only fallback.  Preprocessor directives cannot appear
-    # inside a #define body, so this duplication is necessary.
+    # RISC-V with RVV.  cpu_attn_rvv.hpp supports VLEN=128 and VLEN=256
+    # via RVVI() macros.  Builds with a supported VLEN get
+    # RVV+VEC+VEC16; other RISC-V builds fall back to VEC/VEC16 only.
     header += _macro_block(
         "#elif defined(__riscv) && defined(__riscv_v_min_vlen) "
-        "&& __riscv_v_min_vlen == 128",
+        "&& (__riscv_v_min_vlen == 128 || __riscv_v_min_vlen == 256)",
         ["RVV", "VEC", "VEC16"],
         fp8=False,
     )

csrc/cpu/sgl-kernels/blas_gemm.h

@@ -0,0 +1,82 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+#include <ATen/native/CPUBlas.h>
+
+// Unlike brgemm, PyTorch does not publicly expose at::native::cpublas::gemm
+// If OpenBLS is available in the PyTorch wheel, we rely on it for fast
+// bf16:bf16->fp32 GEMMs Otherwise, we fall back to PyTorch reference BLAS path.
+#if defined(VLLM_HAS_OPENBLAS)
+extern "C" void sbgemm_(char* transa, char* transb, int* m, int* n, int* k,
+                        float* alpha, const at::BFloat16* a, int* lda,
+                        const at::BFloat16* b, int* ldb, float* beta, float* c,
+                        int* ldc);
+
+extern "C" void sgemm_(char* transa, char* transb, int* m, int* n, int* k,
+                       float* alpha, const float* a, int* lda, const float* b,
+                       int* ldb, float* beta, float* c, int* ldc);
+
+inline char blas_transpose(at::native::TransposeType trans) {
+  switch (trans) {
+    case at::native::TransposeType::NoTranspose:
+      return 'n';
+    case at::native::TransposeType::Transpose:
+      return 't';
+    case at::native::TransposeType::ConjTranspose:
+      return 'c';
+  }
+  return 'n';
+}
+
+inline void blas_gemm(at::native::TransposeType transa,
+                      at::native::TransposeType transb, int64_t m, int64_t n,
+                      int64_t k, float alpha, const at::BFloat16* a,
+                      int64_t lda, const at::BFloat16* b, int64_t ldb,
+                      float beta, float* c, int64_t ldc) {
+  char transa_ = blas_transpose(transa);
+  char transb_ = blas_transpose(transb);
+  int m_ = static_cast<int>(m);
+  int n_ = static_cast<int>(n);
+  int k_ = static_cast<int>(k);
+  int lda_ = static_cast<int>(lda);
+  int ldb_ = static_cast<int>(ldb);
+  int ldc_ = static_cast<int>(ldc);
+  sbgemm_(&transa_, &transb_, &m_, &n_, &k_, &alpha, a, &lda_, b, &ldb_, &beta,
+          c, &ldc_);
+}
+
+inline void blas_gemm(at::native::TransposeType transa,
+                      at::native::TransposeType transb, int64_t m, int64_t n,
+                      int64_t k, float alpha, const float* a, int64_t lda,
+                      const float* b, int64_t ldb, float beta, float* c,
+                      int64_t ldc) {
+  char transa_ = blas_transpose(transa);
+  char transb_ = blas_transpose(transb);
+  int m_ = static_cast<int>(m);
+  int n_ = static_cast<int>(n);
+  int k_ = static_cast<int>(k);
+  int lda_ = static_cast<int>(lda);
+  int ldb_ = static_cast<int>(ldb);
+  int ldc_ = static_cast<int>(ldc);
+  sgemm_(&transa_, &transb_, &m_, &n_, &k_, &alpha, a, &lda_, b, &ldb_, &beta,
+         c, &ldc_);
+}
+
+inline void blas_gemm(at::native::TransposeType, at::native::TransposeType,
+                      int64_t, int64_t, int64_t, float, const at::Half*,
+                      int64_t, const at::Half*, int64_t, float, float*,
+                      int64_t) {
+  TORCH_CHECK(false, "CPU OpenBLAS hgemm is not available.");
+}
+#else
+template <typename scalar_t>
+inline void blas_gemm(at::native::TransposeType transa,
+                      at::native::TransposeType transb, int64_t m, int64_t n,
+                      int64_t k, float alpha, const scalar_t* a, int64_t lda,
+                      const scalar_t* b, int64_t ldb, float beta, float* c,
+                      int64_t ldc) {
+  auto gemm = at::native::cpublas::gemm_no_downcast_stub.DEFAULT;
+  gemm(c10::CppTypeToScalarType<scalar_t>::value, transa, transb, m, n, k,
+       at::Scalar(alpha), a, lda, b, ldb, at::Scalar(beta), c, ldc);
+}
+#endif

csrc/cpu/sgl-kernels/common.h

@@ -348,6 +348,13 @@ inline int get_cache_blocks<at::Float8_e4m3fn>(int chunk_size) {
   return std::min(MAX_CACHE_BLOCK_SIZE, cache_block_size);
 }
 
+template <>
+inline int get_cache_blocks<uint8_t>(int chunk_size) {
+  // mxfp4 uses bf16 as accumulate type
+  int cache_block_size = get_cache_blocks<at::BFloat16>(chunk_size);
+  return std::min(MAX_CACHE_BLOCK_SIZE, cache_block_size);
+}
+
 // 2d sequential loop in range : [mb0, mb1), [nb0, nb1)
 template <typename T, typename func_t>
 inline void loop_2d(int64_t mb0, int64_t mb1, int64_t nb0, int64_t nb1, int64_t chunk_size, const func_t& f) {

csrc/cpu/sgl-kernels/conv.cpp

@@ -213,7 +213,7 @@ struct tinygemm_kernel<at::BFloat16, K, BLOCK_N, has_bias, has_silu> {
       weight + nb_start * width,                                                             \
       out + bs * seqlen * dim + mb_start * dim + nb_start,                                   \
       has_bias ? bias + nb_start : nullptr,                                                  \
-      has_conv_states ? conv_states + conv_state_index * (K - 1) * dim + nb_start : nullptr, \
+      has_conv_states ? conv_states + conv_state_index * conv_state_slot_stride + nb_start : nullptr, \
       has_initial_states_value,                                                              \
       mb_size,                                                                               \
       dim,                                                                                   \
@@ -233,7 +233,8 @@ void causal_conv1d_fwd_kernel_impl(
     int64_t dim,
     int64_t seqlen,
     int64_t width,
-    int64_t num_seq_blocks) {
+    int64_t num_seq_blocks,
+    int64_t conv_state_slot_stride) {
   // handle 32 x 64 per block
   constexpr int64_t BLOCK_M = block_size_m();
   constexpr int64_t BLOCK_N = block_size_n() * 2;
@@ -282,7 +283,7 @@ void causal_conv1d_fwd_kernel_impl(
     at::parallel_for(0, batch, 0, [&](int64_t begin, int64_t end) {
       for (int64_t bs = begin; bs < end; ++bs) {
         update_conv_state(
-            conv_states + bs * (width - 1) * dim, input + bs * seqlen * dim, width, dim, seqlen, has_initial_state[bs]);
+            conv_states + bs * conv_state_slot_stride, input + bs * seqlen * dim, width, dim, seqlen, has_initial_state[bs]);
       }
     });
   }
@@ -316,7 +317,8 @@ void causal_conv1d_fwd_varlen_kernel_impl(
     int64_t batch,
     int64_t dim,
     int64_t width,
-    int64_t num_seq_blocks) {
+    int64_t num_seq_blocks,
+    int64_t conv_state_slot_stride) {
   // handle 32 x 64 per block
   constexpr int64_t BLOCK_M = block_size_m();
   constexpr int64_t BLOCK_N = block_size_n() * 2;
@@ -366,7 +368,7 @@ void causal_conv1d_fwd_varlen_kernel_impl(
         int32_t seqlen = query_start_loc[bs + 1] - query_start_loc[bs];
         int32_t batch_offset = query_start_loc[bs];
         update_conv_state(
-            conv_states + conv_state_index * (width - 1) * dim,
+            conv_states + conv_state_index * conv_state_slot_stride,
             input + batch_offset * dim,
             width,
             dim,
@@ -389,7 +391,8 @@ void causal_conv1d_update_kernel_impl(
     int64_t batch,
     int64_t dim,
     int64_t seqlen,
-    int64_t width) {
+    int64_t width,
+    int64_t conv_state_slot_stride) {
   // handle 32 x 64 per block
   constexpr int64_t BLOCK_M = block_size_m();
   constexpr int64_t BLOCK_N = block_size_n() * 2;
@@ -430,7 +433,7 @@ void causal_conv1d_update_kernel_impl(
     });
   });
 
-#define CONV_STATE_INDEXR(w) conv_states + conv_state_index*(width - 1) * dim + (w) * dim
+#define CONV_STATE_INDEXR(w) conv_states + conv_state_index*conv_state_slot_stride + (w) * dim
 
   // update conv_states
   at::parallel_for(0, batch, 0, [&](int64_t begin, int64_t end) {
@@ -592,6 +595,9 @@ at::Tensor causal_conv1d_fwd_cpu(
     }
   }
 
+// IMPORTANT: To make the kernal compatible with vLLM KV cache layout 
+  int64_t conv_state_slot_stride = conv_states->stride(0);
+
   // block size for sequence blocks, 32
   constexpr int64_t BLOCK_M = block_size_m();
 
@@ -618,7 +624,8 @@ at::Tensor causal_conv1d_fwd_cpu(
           batch,
           dim,
           width,
-          num_seq_blocks);
+          num_seq_blocks,
+          conv_state_slot_stride);
     } else {
       causal_conv1d_fwd_kernel_impl<scalar_t>(
           out.data_ptr<scalar_t>(),
@@ -633,7 +640,8 @@ at::Tensor causal_conv1d_fwd_cpu(
           dim,
           seqlen,
           width,
-          num_seq_blocks);
+          num_seq_blocks,
+          conv_state_slot_stride);
     }
   });
   return out;
@@ -690,6 +698,8 @@ at::Tensor causal_conv1d_update_cpu(
     conv_states.copy_(conv_states_copy);
   }
 
+  // IMPORTANT: To make the kernal compatible with vLLM KV cache layout 
+  int64_t conv_state_slot_stride = conv_states.stride(0);
   at::Tensor out = at::empty_like(x);
   AT_DISPATCH_REDUCED_FLOATING_TYPES(scalar_type, "causal_conv1d_update_kernel_impl", [&] {
     causal_conv1d_update_kernel_impl<scalar_t>(
@@ -703,7 +713,8 @@ at::Tensor causal_conv1d_update_cpu(
         batch,
         dim,
         seqlen,
-        width);
+        width,
+        conv_state_slot_stride);
   });
   return out;
 }

csrc/cpu/sgl-kernels/fla.cpp

@@ -301,25 +301,42 @@ void chunk_gated_delta_rule_kernel_impl(
       // attn = k_beta @ key.transpose(-1, -2)
       // attn: [B, HV, num_chunk, chunk_size, chunk_size]
       // transpose and pack for key
-      pack_vnni<scalar_t>(
-          /*    dst */ k_transpose,
-          /*    src */ curr_k_pad,
-          /*     N  */ chunk_size,
-          /*     K  */ qk_head_size,
-          /* ld_src */ qk_head_size,
-          /* ld_dst */ chunk_size);
-      // k_beta @ key.transpose(-1, -2)
-      at::native::cpublas::brgemm(
-          /*     M */ chunk_size,
-          /*     N */ chunk_size,
-          /*     K */ qk_head_size,
-          /*   lda */ qk_head_size,
-          /*   ldb */ chunk_size,
-          /*   ldc */ chunk_size,
-          /* add_C */ false,
-          /*     A */ curr_k_beta,
-          /*     B */ k_transpose,
-          /*     C */ curr_attn);
+      if constexpr (brgemm_supported()) {
+        pack_vnni<scalar_t>(
+            /*    dst */ k_transpose,
+            /*    src */ curr_k_pad,
+            /*     N  */ chunk_size,
+            /*     K  */ qk_head_size,
+            /* ld_src */ qk_head_size,
+            /* ld_dst */ chunk_size);
+        // k_beta @ key.transpose(-1, -2)
+        at::native::cpublas::brgemm(
+            /*     M */ chunk_size,
+            /*     N */ chunk_size,
+            /*     K */ qk_head_size,
+            /*   lda */ qk_head_size,
+            /*   ldb */ chunk_size,
+            /*   ldc */ chunk_size,
+            /* add_C */ false,
+            /*     A */ curr_k_beta,
+            /*     B */ k_transpose,
+            /*     C */ curr_attn);
+      } else {
+        blas_gemm(
+            at::native::TransposeType::Transpose,
+            at::native::TransposeType::NoTranspose,
+            chunk_size,
+            chunk_size,
+            qk_head_size,
+            1.0f,
+            curr_k_pad,
+            qk_head_size,
+            curr_k_beta,
+            qk_head_size,
+            0.0f,
+            curr_attn,
+            chunk_size);
+      }
       // attn = attn * decay_mask
       for (int64_t m = 0; m < chunk_size; m++) {
         at::vec::map2<float>(
@@ -413,25 +430,42 @@ void chunk_gated_delta_rule_kernel_impl(
       // k_beta_g = k_beta * g: [B, HV, num_chunk, chunk_size, EK]
       // k_cumdecay: [B, HV, num_chunk, chunk_size, EK]
       // pack for value
-      pack_vnni2<scalar_t>(
-          /*    dst */ v_pack,
-          /*    src */ curr_v_beta,
-          /*     N  */ chunk_size,
-          /*     K  */ v_head_size,
-          /* ld_src */ v_head_size,
-          /* ld_dst */ v_head_size);
-      // value = attn @ v_beta
-      at::native::cpublas::brgemm(
-          /*     M */ chunk_size,
-          /*     N */ v_head_size,
-          /*     K */ chunk_size,
-          /*   lda */ chunk_size,
-          /*   ldb */ v_head_size,
-          /*   ldc */ v_head_size,
-          /* add_C */ false,
-          /*     A */ curr_attn_reduced,
-          /*     B */ v_pack,
-          /*     C */ curr_value);
+      if constexpr (brgemm_supported()) {
+        pack_vnni2<scalar_t>(
+            /*    dst */ v_pack,
+            /*    src */ curr_v_beta,
+            /*     N  */ chunk_size,
+            /*     K  */ v_head_size,
+            /* ld_src */ v_head_size,
+            /* ld_dst */ v_head_size);
+        // value = attn @ v_beta
+        at::native::cpublas::brgemm(
+            /*     M */ chunk_size,
+            /*     N */ v_head_size,
+            /*     K */ chunk_size,
+            /*   lda */ chunk_size,
+            /*   ldb */ v_head_size,
+            /*   ldc */ v_head_size,
+            /* add_C */ false,
+            /*     A */ curr_attn_reduced,
+            /*     B */ v_pack,
+            /*     C */ curr_value);
+      } else {
+        blas_gemm(
+            at::native::TransposeType::NoTranspose,
+            at::native::TransposeType::NoTranspose,
+            v_head_size,
+            chunk_size,
+            chunk_size,
+            1.0f,
+            curr_v_beta,
+            v_head_size,
+            curr_attn_reduced,
+            chunk_size,
+            0.0f,
+            curr_value,
+            v_head_size);
+      }
       // k_beta_g = k_beta * g.exp().unsqueeze(-1)
       for (int64_t j = 0; j < chunk_size; j++) {
         int64_t i = 0;
@@ -445,25 +479,42 @@ void chunk_gated_delta_rule_kernel_impl(
         }
       }
       // pack for k_beta_g
-      pack_vnni2<scalar_t>(
-          /*    dst */ k_beta_g_pack,
-          /*    src */ k_beta_g,
-          /*     N  */ chunk_size,
-          /*     K  */ qk_head_size,
-          /* ld_src */ qk_head_size,
-          /* ld_dst */ qk_head_size);
-      // k_cumdecay = attn @ k_beta_g
-      at::native::cpublas::brgemm(
-          /*     M */ chunk_size,
-          /*     N */ qk_head_size,
-          /*     K */ chunk_size,
-          /*   lda */ chunk_size,
-          /*   ldb */ qk_head_size,
-          /*   ldc */ qk_head_size,
-          /* add_C */ false,
-          /*     A */ curr_attn_reduced,
-          /*     B */ k_beta_g_pack,
-          /*     C */ k_cumdecay);
+      if constexpr (brgemm_supported()) {
+        pack_vnni2<scalar_t>(
+            /*    dst */ k_beta_g_pack,
+            /*    src */ k_beta_g,
+            /*     N  */ chunk_size,
+            /*     K  */ qk_head_size,
+            /* ld_src */ qk_head_size,
+            /* ld_dst */ qk_head_size);
+        // k_cumdecay = attn @ k_beta_g
+        at::native::cpublas::brgemm(
+            /*     M */ chunk_size,
+            /*     N */ qk_head_size,
+            /*     K */ chunk_size,
+            /*   lda */ chunk_size,
+            /*   ldb */ qk_head_size,
+            /*   ldc */ qk_head_size,
+            /* add_C */ false,
+            /*     A */ curr_attn_reduced,
+            /*     B */ k_beta_g_pack,
+            /*     C */ k_cumdecay);
+      } else {
+        blas_gemm(
+            at::native::TransposeType::NoTranspose,
+            at::native::TransposeType::NoTranspose,
+            qk_head_size,
+            chunk_size,
+            chunk_size,
+            1.0f,
+            k_beta_g,
+            qk_head_size,
+            curr_attn_reduced,
+            chunk_size,
+            0.0f,
+            k_cumdecay,
+            qk_head_size);
+      }
       for (int i = 0; i < chunk_size; i++) {
         at::vec::map<scalar_t>(
             [](fVec x) { return x; },
@@ -551,25 +602,42 @@ void chunk_gated_delta_rule_kernel_impl(
 
         // attn_i = (q_i @ k_i.transpose(-1, -2) * decay_mask[:, :, i]).masked_fill_(mask, 0)
         // k_transpose_i = k_i.transpose(-1, -2)
-        pack_vnni<scalar_t>(
-            /*    dst */ k_transpose_i,
-            /*    src */ k_i,
-            /*     N  */ chunk_size,
-            /*     K  */ qk_head_size,
-            /* ld_src */ qk_head_size,
-            /* ld_dst */ chunk_size);
-        // attn_i = q_i @ k_transpose_i
-        at::native::cpublas::brgemm(
-            /* M */ chunk_size,
-            /* N */ chunk_size,
-            /* K */ qk_head_size,
-            /* lda */ qk_head_size,
-            /* ldb */ chunk_size,
-            /* ldc */ chunk_size,
-            /* add_C */ false,
-            /* A */ q_i,
-            /* B */ k_transpose_i,
-            /* C */ attn_i);
+        if constexpr (brgemm_supported()) {
+          pack_vnni<scalar_t>(
+              /*    dst */ k_transpose_i,
+              /*    src */ k_i,
+              /*     N  */ chunk_size,
+              /*     K  */ qk_head_size,
+              /* ld_src */ qk_head_size,
+              /* ld_dst */ chunk_size);
+          // attn_i = q_i @ k_transpose_i
+          at::native::cpublas::brgemm(
+              /* M */ chunk_size,
+              /* N */ chunk_size,
+              /* K */ qk_head_size,
+              /* lda */ qk_head_size,
+              /* ldb */ chunk_size,
+              /* ldc */ chunk_size,
+              /* add_C */ false,
+              /* A */ q_i,
+              /* B */ k_transpose_i,
+              /* C */ attn_i);
+        } else {
+          blas_gemm(
+              at::native::TransposeType::Transpose,
+              at::native::TransposeType::NoTranspose,
+              chunk_size,
+              chunk_size,
+              qk_head_size,
+              1.0f,
+              k_i,
+              qk_head_size,
+              q_i,
+              qk_head_size,
+              0.0f,
+              attn_i,
+              chunk_size);
+        }
         // attn_i = attn_i * decay_mask_i
         for (int64_t m = 0; m < chunk_size; m++) {
           auto attn_i_m = attn_i + m * chunk_size;
@@ -609,28 +677,45 @@ void chunk_gated_delta_rule_kernel_impl(
         }
 
         // pack for curr_last_recurrent_state
-        pack_vnni2<scalar_t>(
-            /*    dst */ curr_last_recurrent_state_pack_reduced,
-            /*    src */ curr_last_recurrent_state_reduced,
-            /*     N  */ qk_head_size,
-            /*     K  */ v_head_size,
-            /* ld_src */ v_head_size,
-            /* ld_dst */ v_head_size);
+        if constexpr (brgemm_supported()) {
+          pack_vnni2<scalar_t>(
+              /*    dst */ curr_last_recurrent_state_pack_reduced,
+              /*    src */ curr_last_recurrent_state_reduced,
+              /*     N  */ qk_head_size,
+              /*     K  */ v_head_size,
+              /* ld_src */ v_head_size,
+              /* ld_dst */ v_head_size);
 
-        // v_prime = k_cumdecay_i @ curr_last_recurrent_state: [chunk_size, EV]
-        // k_cumdecay_i: [chunk_size, EK]
-        // curr_last_recurrent_state: [EK, EV]
-        at::native::cpublas::brgemm(
-            /*     M */ chunk_size,
-            /*     N */ v_head_size,
-            /*     K */ qk_head_size,
-            /*   lda */ qk_head_size,
-            /*   ldb */ v_head_size,
-            /*   ldc */ v_head_size,
-            /* add_C */ false,
-            /*     A */ k_cumdecay_i_reduced,
-            /*     B */ curr_last_recurrent_state_pack_reduced,
-            /*     C */ v_prime);
+          // v_prime = k_cumdecay_i @ curr_last_recurrent_state: [chunk_size, EV]
+          // k_cumdecay_i: [chunk_size, EK]
+          // curr_last_recurrent_state: [EK, EV]
+          at::native::cpublas::brgemm(
+              /*     M */ chunk_size,
+              /*     N */ v_head_size,
+              /*     K */ qk_head_size,
+              /*   lda */ qk_head_size,
+              /*   ldb */ v_head_size,
+              /*   ldc */ v_head_size,
+              /* add_C */ false,
+              /*     A */ k_cumdecay_i_reduced,
+              /*     B */ curr_last_recurrent_state_pack_reduced,
+              /*     C */ v_prime);
+        } else {
+          blas_gemm(
+              at::native::TransposeType::NoTranspose,
+              at::native::TransposeType::NoTranspose,
+              v_head_size,
+              chunk_size,
+              qk_head_size,
+              1.0f,
+              curr_last_recurrent_state_reduced,
+              v_head_size,
+              k_cumdecay_i_reduced,
+              qk_head_size,
+              0.0f,
+              v_prime,
+              v_head_size);
+        }
 
         // v_new = v_prime = v_i - v_prime
         // v_i: [chunk_size, EV]
@@ -663,41 +748,75 @@ void chunk_gated_delta_rule_kernel_impl(
         }
         // attn_inter = qg @ curr_last_recurrent_state: [chunk_size, EV]
         // curr_last_recurrent_state: [EK, EV]
-        at::native::cpublas::brgemm(
-            /* M */ chunk_size,
-            /* N */ v_head_size,
-            /* K */ qk_head_size,
-            /* lda */ qk_head_size,
-            /* ldb */ v_head_size,
-            /* ldc */ v_head_size,
-            /* add_C */ false,
-            /* A */ qg,
-            /* B */ curr_last_recurrent_state_pack_reduced,
-            /* C */ attn_inter);
+        if constexpr (brgemm_supported()) {
+          at::native::cpublas::brgemm(
+              /* M */ chunk_size,
+              /* N */ v_head_size,
+              /* K */ qk_head_size,
+              /* lda */ qk_head_size,
+              /* ldb */ v_head_size,
+              /* ldc */ v_head_size,
+              /* add_C */ false,
+              /* A */ qg,
+              /* B */ curr_last_recurrent_state_pack_reduced,
+              /* C */ attn_inter);
+        } else {
+          blas_gemm(
+              at::native::TransposeType::NoTranspose,
+              at::native::TransposeType::NoTranspose,
+              v_head_size,
+              chunk_size,
+              qk_head_size,
+              1.0f,
+              curr_last_recurrent_state_reduced,
+              v_head_size,
+              qg,
+              qk_head_size,
+              0.0f,
+              attn_inter,
+              v_head_size);
+        }
 
         // core_attn_out[:, :, i] = attn_inter + attn_i @ v_new
         // pack for v_prime
-        pack_vnni2<scalar_t>(
-            /*    dst */ v_prime_pack_reduced,
-            /*    src */ v_prime_reduced,
-            /*     N  */ chunk_size,
-            /*     K  */ v_head_size,
-            /* ld_src */ v_head_size,
-            /* ld_dst */ v_head_size);
-        // attn_inter = attn_inter + attn_i @ v_new: [chunk_size, EV]
-        // attn_i: [chunk_size, chunk_size]
-        // v_new: [chunk_size, EV]
-        at::native::cpublas::brgemm(
-            /* M */ chunk_size,
-            /* N */ v_head_size,
-            /* K */ chunk_size,
-            /* lda */ chunk_size,
-            /* ldb */ v_head_size,
-            /* ldc */ v_head_size,
-            /* add_C */ true,
-            /* A */ attn_i_reduced,
-            /* B */ v_prime_pack_reduced,
-            /* C */ attn_inter);
+        if constexpr (brgemm_supported()) {
+          pack_vnni2<scalar_t>(
+              /*    dst */ v_prime_pack_reduced,
+              /*    src */ v_prime_reduced,
+              /*     N  */ chunk_size,
+              /*     K  */ v_head_size,
+              /* ld_src */ v_head_size,
+              /* ld_dst */ v_head_size);
+          // attn_inter = attn_inter + attn_i @ v_new: [chunk_size, EV]
+          // attn_i: [chunk_size, chunk_size]
+          // v_new: [chunk_size, EV]
+          at::native::cpublas::brgemm(
+              /* M */ chunk_size,
+              /* N */ v_head_size,
+              /* K */ chunk_size,
+              /* lda */ chunk_size,
+              /* ldb */ v_head_size,
+              /* ldc */ v_head_size,
+              /* add_C */ true,
+              /* A */ attn_i_reduced,
+              /* B */ v_prime_pack_reduced,
+              /* C */ attn_inter);
+        } else {
+          blas_gemm(
+              at::native::TransposeType::NoTranspose,
+              at::native::TransposeType::NoTranspose,
+              v_head_size,
+              chunk_size,
+              chunk_size,
+              1.0f,
+              v_prime_reduced,
+              v_head_size,
+              attn_i_reduced,
+              chunk_size,
+              1.0f,
+              attn_inter,
+              v_head_size);
+        }
 
         // core_attn_out[:, :, i] = attn_inter
         for (int64_t m = 0; m < chunk_size; m++) {
@@ -762,17 +881,34 @@ void chunk_gated_delta_rule_kernel_impl(
             /* ld_dst */ chunk_size);
         // kgv = kg.transpose(-1, -2) @ v_new
         // v_new: [chunk_size, EV]
-        at::native::cpublas::brgemm(
-            /* M */ qk_head_size,
-            /* N */ v_head_size,
-            /* K */ chunk_size,
-            /* lda */ chunk_size,
-            /* ldb */ v_head_size,
-            /* ldc */ v_head_size,
-            /* add_C */ false,
-            /* A */ kg_transpose,
-            /* B */ v_prime_pack_reduced,
-            /* C */ kgv);
+        if constexpr (brgemm_supported()) {
+          at::native::cpublas::brgemm(
+              /* M */ qk_head_size,
+              /* N */ v_head_size,
+              /* K */ chunk_size,
+              /* lda */ chunk_size,
+              /* ldb */ v_head_size,
+              /* ldc */ v_head_size,
+              /* add_C */ false,
+              /* A */ kg_transpose,
+              /* B */ v_prime_pack_reduced,
+              /* C */ kgv);
+        } else {
+          blas_gemm(
+              at::native::TransposeType::NoTranspose,
+              at::native::TransposeType::NoTranspose,
+              v_head_size,
+              qk_head_size,
+              chunk_size,
+              1.0f,
+              v_prime_reduced,
+              v_head_size,
+              kg_transpose,
+              chunk_size,
+              0.0f,
+              kgv,
+              v_head_size);
+        }
         // last_recurrent_state = 1) + 2)
         for (int64_t m = 0; m < qk_head_size; m++) {
           at::vec::map2<float>(
@@ -847,6 +983,7 @@ void fused_sigmoid_gating_delta_rule_update_kernel_impl(
     int64_t v_strideB,
     int64_t v_strideS,
     int64_t v_strideH,
+    int64_t state_slot_stride,
     bool use_qk_l2norm_in_kernel,
     double softplus_threshold) {
   using bVec = at::vec::Vectorized<scalar_t>;
@@ -907,7 +1044,7 @@ void fused_sigmoid_gating_delta_rule_update_kernel_impl(
     data_index_init(begin, bi, batch_size, si, seq_len, ni, v_num_heads);
     for (int64_t i = begin; i < end; ++i) {
       int64_t cache_index = indices_ptr[bi];
-      int64_t state_offset = (cache_index * v_num_heads + ni) * head_dim * v_head_dim;
+      int64_t state_offset = cache_index * state_slot_stride + ni * head_dim * v_head_dim;
       float g_val = -std::exp(float(A_log_ptr[ni])) *
                     softplus(float(a_ptr[bi * v_num_heads + ni]) + float(dt_bias_ptr[ni]), softplus_threshold);
       float g_val_exp = std::exp(g_val);
@@ -920,7 +1057,8 @@ void fused_sigmoid_gating_delta_rule_update_kernel_impl(
       float k_scale = use_qk_l2norm_in_kernel ? qk_scale_buf[k_scale_offset] : 1.0f;
       int64_t v_offset = si * v_strideS + bi * v_strideB + ni * v_strideH;
       int64_t o_offset = ((bi * seq_len + si) * v_num_heads + ni) * v_head_dim;
-      float beta_val = 1 / (1 + std::exp(-b_ptr[ni]));
+      // See: https://github.com/sgl-project/sglang/pull/26634
+      float beta_val = 1 / (1 + std::exp(-b_ptr[bi * v_num_heads + ni]));
       fVec beta_vec = fVec(beta_val);
       int64_t dvi = 0;
       for (; dvi <= v_head_dim - VecSize; dvi += VecSize) {
@@ -1321,6 +1459,8 @@ at::Tensor fused_sigmoid_gating_delta_rule_update_cpu(
   int64_t v_strideB = v.stride(1);
   int64_t v_strideS = v.stride(0);
   int64_t v_strideH = v.stride(2);
+  // IMPORTANT: To make the kernal compatible with vLLM KV cache layout 
+  int64_t state_slot_stride = initial_state_source.stride(0);
   at::Tensor core_attn_out = at::empty({batch_size, seq_len, v_num_heads, v_head_dim}, q.options());
   at::Tensor qk_scale_buf = at::empty({2 * batch_size, seq_len, num_heads}, at::kFloat);
 
@@ -1353,6 +1493,7 @@ at::Tensor fused_sigmoid_gating_delta_rule_update_cpu(
             v_strideB,
             v_strideS,
             v_strideH,
+            state_slot_stride,
             use_qk_l2norm_in_kernel,
             softplus_threshold);
       });

csrc/cpu/sgl-kernels/gemm.h

@@ -4,9 +4,12 @@
 // clang-format off
 
 #pragma once
-#include <ATen/native/CPUBlas.h>
-
 #include "common.h"
+#include "blas_gemm.h"
+
+#if defined(__AVX512F__) && defined(__AVX512BF16__) && defined(__AMX_BF16__)
+#define CPU_CAPABILITY_AVX512
+#endif
 
 // amx-bf16
 #define TILE_M 16
@@ -21,31 +24,39 @@ constexpr int block_size_n() {
   return 2 * TILE_N;
 }
 
+constexpr bool brgemm_supported() {
+#if defined(CPU_CAPABILITY_AVX512)
+  return true;
+#else
+  return false;
+#endif
+}
+
 // define threshold using brgemm (intel AMX)
 template <typename T>
 inline bool can_use_brgemm(int M);
 template <>
 inline bool can_use_brgemm<at::BFloat16>(int M) {
-  return M > 4;
+  return brgemm_supported() && M > 4;
 }
 template <>
 inline bool can_use_brgemm<at::Half>(int M) {
-  return true;
+  return brgemm_supported();
 }
 // this requires PyTorch 2.7 or above
 template <>
 inline bool can_use_brgemm<int8_t>(int M) {
-  return M > 4;
+  return brgemm_supported() && M > 4;
 }
 
 template <>
 inline bool can_use_brgemm<uint8_t>(int M) {
-  return M > 4;
+  return brgemm_supported() && M > 4;
 }
 
 template <>
 inline bool can_use_brgemm<at::Float8_e4m3fn>(int M) {
-  return M > 4;
+  return brgemm_supported() && M > 4;
 }
 
 // work around compiler internal error
@@ -72,7 +83,17 @@ inline int64_t get_row_size(int64_t K, bool use_int8_w8a8) {
   return use_int8_w8a8 ? K + sizeof(int32_t) : K;
 }
 
-enum class CPUQuantMethod : int64_t { BF16 = 0, INT8_W8A8 = 1, FP8_W8A16 = 2, INT4_W4A8 = 3 };
+enum class CPUAcTMethod : int { silu_and_mul = 0, swiglu = 1 };
+
+constexpr bool operator==(CPUAcTMethod a, int b) {
+  return static_cast<int>(a) == b;
+}
+
+constexpr bool operator==(int a, CPUAcTMethod b) {
+  return a == static_cast<int>(b);
+}
+
+enum class CPUQuantMethod : int64_t { BF16 = 0, INT8_W8A8 = 1, FP8_W8A16 = 2, INT4_W4A8 = 3, MXFP4 = 4 };
 
 constexpr bool operator==(CPUQuantMethod a, int64_t b) {
   return static_cast<int64_t>(a) == b;
@@ -99,6 +120,9 @@ inline int64_t get_4bit_block_k_size(int64_t group_size) {
 // pack weight to vnni format
 at::Tensor convert_weight_packed(at::Tensor& weight);
 
+// pack scale to blocked format for mxfp4
+at::Tensor convert_scale_packed(at::Tensor& scale);
+
 // pack weight to vnni format for int4
 std::tuple<at::Tensor, at::Tensor, at::Tensor>
 convert_weight_packed_scale_zp(at::Tensor qweight, at::Tensor qzeros, at::Tensor scales);
@@ -129,9 +153,9 @@ void fused_experts_int8_kernel_impl(
     int64_t topk,
     int64_t num_tokens_post_pad);
 
-// moe implementations for fp8 w8a16
-template <typename scalar_t>
-void fused_experts_fp8_kernel_impl(
+// moe implementations for fp8 w8a16 and mxfp4
+template <typename scalar_t, typename packed_t, typename param_t, bool is_mxfp4>
+void fused_experts_fp_kernel_impl(
     scalar_t* __restrict__ output,
     scalar_t* __restrict__ ic0,
     scalar_t* __restrict__ ic1,
@@ -140,10 +164,12 @@ void fused_experts_fp8_kernel_impl(
     scalar_t* __restrict__ B_tmp,
     float* __restrict__ C_tmp,
     const scalar_t* __restrict__ input,
-    const at::Float8_e4m3fn* __restrict__ packed_w1,
-    const at::Float8_e4m3fn* __restrict__ packed_w2,
-    const float* __restrict__ w1s,
-    const float* __restrict__ w2s,
+    const packed_t* __restrict__ packed_w1,
+    const packed_t* __restrict__ packed_w2,
+    const float* __restrict__ w1_bias,
+    const float* __restrict__ w2_bias,
+    const param_t* __restrict__ w1s,
+    const param_t* __restrict__ w2s,
     int64_t block_size_N,
     int64_t block_size_K,
     const float* __restrict__ topk_weights,
@@ -155,7 +181,11 @@ void fused_experts_fp8_kernel_impl(
     int64_t K,
     int64_t E,
     int64_t topk,
-    int64_t num_tokens_post_pad);
+    int64_t num_tokens_post_pad,
+    float alpha,
+    float limit,
+    CPUAcTMethod act_func,
+    bool with_bias);
 
 // shared expert implementation for int8 w8a8
 template <typename scalar_t>
@@ -266,6 +296,7 @@ void tinygemm_kernel(
     scalar_t* __restrict__ C,
     scalar_t* __restrict__ Btmp,
     float* __restrict__ Ctmp,
+    const float* __restrict__ Bbias,
     const float* __restrict__ scale,
     int64_t M,
     int64_t N,
@@ -294,6 +325,26 @@ void tinygemm_kernel(
     int64_t ldc,
     bool brg);
 
+// mxfp4
+template <typename scalar_t>
+void tinygemm_kernel(
+    const scalar_t* __restrict__ A,
+    const uint8_t* __restrict__ B,
+    scalar_t* __restrict__ C,
+    scalar_t* __restrict__ Btmp,
+    float* __restrict__ Ctmp,
+    const float* __restrict__ Bbias,
+    const uint8_t* __restrict__ scale,
+    int64_t M,
+    int64_t N,
+    int64_t K,
+    int64_t lda,
+    int64_t ldb,
+    int64_t ldc,
+    bool brg,
+    int64_t block_size_K,
+    bool do_unpack = true);
+
 template <typename scalar_t>
 void tinygemm_kernel(
     scalar_t* C,

csrc/cpu/sgl-kernels/gemm_fp8.cpp

@@ -67,6 +67,23 @@ inline void copy_mul_stub(scalar_t* __restrict__ out, const float* __restrict__
   }
 }
 
+template <>
+inline void
+copy_add_stub(float* __restrict__ out, const float* __restrict__ input, const float* __restrict__ bias, int64_t size) {
+  using fVec = at::vec::Vectorized<float>;
+  constexpr int kVecSize = fVec::size();
+
+  int64_t d;
+#pragma GCC unroll 4
+  for (d = 0; d <= size - kVecSize; d += kVecSize) {
+    fVec data = fVec::loadu(input + d) + fVec::loadu(bias + d);
+    data.store(out + d);
+  }
+  for (; d < size; ++d) {
+    out[d] = input[d] + bias[d];
+  }
+}
+
 inline void unpack_B(
     at::BFloat16* __restrict__ Btmp,
     const at::Float8_e4m3fn* __restrict__ packed_B,
@@ -343,7 +360,6 @@ struct tinygemm_kernel_nn<at::BFloat16, at::Float8_e4m3fn, float, has_bias, BLOC
     Unroll<ROWS * COLS>{}(storec);
   }
 };
-
 template <int BLOCK_M, int BLOCK_N>
 struct tinygemm_kernel_nn2<at::BFloat16, BLOCK_M, BLOCK_N> {
   static inline void apply(
@@ -918,6 +934,7 @@ void tinygemm_kernel(
     scalar_t* __restrict__ C,
     scalar_t* __restrict__ Btmp,
     float* __restrict__ Ctmp,
+    const float* __restrict__ Bbias,
     const float* __restrict__ scale,
     int64_t M,
     int64_t N,
@@ -928,6 +945,11 @@ void tinygemm_kernel(
     bool brg,
     int64_t block_size_K,
     bool do_unpack) {
+  if (Bbias != nullptr) {
+    tinygemm_kernel<scalar_t, at::Float8_e4m3fn, float, true>(
+        A, B, C, Btmp, Ctmp, scale, Bbias, M, N, K, lda, ldb, ldc, brg, block_size_K, do_unpack);
+    return;
+  }
   tinygemm_kernel<scalar_t, at::Float8_e4m3fn, float, false>(
       A, B, C, Btmp, Ctmp, scale, nullptr, M, N, K, lda, ldb, ldc, brg, block_size_K, do_unpack);
 }
@@ -949,7 +971,6 @@ void tinygemm_kernel(
     bool brg) {
   tinygemm_kernel2<scalar_t>(A, B, C, Btmp, Ctmp, scale, M, N, K, lda, ldb, ldc, brg);
 }
-
 template <typename scalar_t>
 void tinygemm_kernel(
     const scalar_t* __restrict__ A,
@@ -957,6 +978,7 @@ void tinygemm_kernel(
     scalar_t* __restrict__ C,
     scalar_t* __restrict__ Btmp,
     float* __restrict__ Ctmp,
+    const float* __restrict__ Bbias,
     const uint8_t* __restrict__ scale,
     int64_t M,
     int64_t N,
@@ -967,10 +989,68 @@ void tinygemm_kernel(
     bool brg,
     int64_t block_size_K,
     bool do_unpack) {
+  if (Bbias != nullptr) {
+    tinygemm_kernel<scalar_t, uint8_t, uint8_t, true>(
+        A, B, C, Btmp, Ctmp, scale, Bbias, M, N, K, lda, ldb, ldc, brg, block_size_K, do_unpack);
+    return;
+  }
   tinygemm_kernel<scalar_t, uint8_t, uint8_t, false>(
       A, B, C, Btmp, Ctmp, scale, nullptr, M, N, K, lda, ldb, ldc, brg, block_size_K, do_unpack);
 }
 
+// tinygemm interface
+template <typename scalar_t>
+void tinygemm_kernel(
+    const scalar_t* __restrict__ A,
+    const at::Float8_e4m3fn* __restrict__ B,
+    float* __restrict__ C,
+    scalar_t* __restrict__ Btmp,
+    const float* __restrict__ Bbias,
+    const float* __restrict__ scale,
+    int64_t M,
+    int64_t N,
+    int64_t K,
+    int64_t lda,
+    int64_t ldb,
+    int64_t ldc,
+    bool brg,
+    int64_t block_size_K,
+    bool do_unpack) {
+  if (Bbias != nullptr) {
+    tinygemm_kernel<scalar_t, at::Float8_e4m3fn, float, true>(
+        A, B, C, Btmp, scale, Bbias, M, N, K, lda, ldb, ldc, brg, block_size_K, do_unpack);
+    return;
+  }
+  tinygemm_kernel<scalar_t, at::Float8_e4m3fn, float, false>(
+      A, B, C, Btmp, scale, nullptr, M, N, K, lda, ldb, ldc, brg, block_size_K, do_unpack);
+}
+
+template <typename scalar_t>
+void tinygemm_kernel(
+    const scalar_t* __restrict__ A,
+    const uint8_t* __restrict__ B,
+    float* __restrict__ C,
+    scalar_t* __restrict__ Btmp,
+    const float* __restrict__ Bbias,
+    const uint8_t* __restrict__ scale,
+    int64_t M,
+    int64_t N,
+    int64_t K,
+    int64_t lda,
+    int64_t ldb,
+    int64_t ldc,
+    bool brg,
+    int64_t block_size_K,
+    bool do_unpack) {
+  if (Bbias != nullptr) {
+    tinygemm_kernel<scalar_t, uint8_t, uint8_t, true>(
+        A, B, C, Btmp, scale, Bbias, M, N, K, lda, ldb, ldc, brg, block_size_K, do_unpack);
+    return;
+  }
+  tinygemm_kernel<scalar_t, uint8_t, uint8_t, false>(
+      A, B, C, Btmp, scale, nullptr, M, N, K, lda, ldb, ldc, brg, block_size_K, do_unpack);
+}
+
 #define INSTANTIATE_TINYGEMM_TEMPLATE(TYPE_A, TYPE_B, TYPE_S) \
   template void tinygemm_kernel<TYPE_A>(                      \
       const TYPE_A* __restrict__ A,                           \
@@ -978,6 +1058,7 @@ void tinygemm_kernel(
       TYPE_A* __restrict__ C,                                 \
       TYPE_A* __restrict__ Btmp,                              \
       float* __restrict__ Ctmp,                               \
+      const float* __restrict__ Bbias,                        \
       const TYPE_S* __restrict__ scale,                       \
       int64_t M,                                              \
       int64_t N,                                              \
@@ -1020,7 +1101,6 @@ inline const float* get_bias_data(const std::optional<at::Tensor>& bias, int64_t
   }
   return nullptr;
 }
-
 // FP8 and MXFP4 WoQ uses the same pattern:
 //   Btmp : [T, BLOCK_N * K]
 //   Ctmp : [T, BLOCK_M * BLOCK_N]

csrc/cpu/sgl-kernels/moe.cpp

@@ -812,6 +812,7 @@ void shared_expert_kernel_impl(
 static inline void check_moe_scales(
     bool use_int8_w8a8,
     bool use_fp8_w8a16,
+    bool use_mxfp4,
     const std::optional<at::Tensor>& w1_scale,
     const std::optional<at::Tensor>& w2_scale,
     const std::optional<std::vector<int64_t>> block_size) {
@@ -825,6 +826,12 @@ static inline void check_moe_scales(
     TORCH_CHECK(block_size.has_value(), "missing block_size for fp8 w8a16.");
     TORCH_CHECK(block_size.value().size() == 2, "expect block_size.size() to be 2.");
   }
+  if (use_mxfp4) {
+    TORCH_CHECK(w1_scale.has_value(), "missing w1_scale for mxfp4.");
+    TORCH_CHECK(w2_scale.has_value(), "missing w2_scale for mxfp4.");
+    TORCH_CHECK(w1_scale.value().scalar_type() == at::kByte, "expect w1_scale to be uint8.");
+    TORCH_CHECK(w2_scale.value().scalar_type() == at::kByte, "expect w2_scale to be uint8.");
+  }
 }
 
 #define CHECK_MOE_SCALES_FP8(DIM0, DIM1)                      \
@@ -839,8 +846,8 @@ static inline void check_moe_scales(
   TORCH_CHECK(w2s.size(DIM1) == div_up(N, block_size_K))
 
 // hidden_states: [M, K]
-// w1: [E, 2N, K]
-// w2: [E, K, N]
+// w1: [E, 2N, K] or [E, 2N, K / 2] for uint8
+// w2: [E, K, N] or [E, K, N / 2] for uint8
 // topk_weights: [M, topk]
 // topk_ids: [M, topk] (int32_t)
 //
@@ -858,6 +865,10 @@ at::Tensor fused_experts_cpu(
     const std::optional<at::Tensor>& w1_zero,
     const std::optional<at::Tensor>& w2_zero,
     const std::optional<std::vector<int64_t>> block_size,
+    const std::optional<at::Tensor>& w1_bias,
+    const std::optional<at::Tensor>& w2_bias,
+    const std::optional<double>& alpha,
+    const std::optional<double>& limit,
     bool is_vnni) {
   auto packed_w1 = is_vnni ? w1 : convert_weight_packed(w1);
   auto packed_w2 = is_vnni ? w2 : convert_weight_packed(w2);
@@ -897,8 +908,12 @@ at::Tensor fused_experts_cpu(
   int64_t topk = topk_weights_.size(1);
 
   // we use int32_t compensation for int8 w8a8
-  int64_t packed_K = get_row_size(K, moe_comp_method == CPUQuantMethod::INT8_W8A8);
-  int64_t packed_N = get_row_size(N, moe_comp_method == CPUQuantMethod::INT8_W8A8);
+  int64_t packed_K = moe_comp_method == CPUQuantMethod::MXFP4
+                         ? get_row_size<uint8_t>(K)
+                         : get_row_size(K, moe_comp_method == CPUQuantMethod::INT8_W8A8);
+  int64_t packed_N = moe_comp_method == CPUQuantMethod::MXFP4
+                         ? get_row_size<uint8_t>(N)
+                         : get_row_size(N, moe_comp_method == CPUQuantMethod::INT8_W8A8);
 
   // check weight shapes
   CHECK_EQ(w2.size(0), E);
@@ -911,6 +926,7 @@ at::Tensor fused_experts_cpu(
   check_moe_scales(
       moe_comp_method == CPUQuantMethod::INT8_W8A8,
       moe_comp_method == CPUQuantMethod::FP8_W8A16,
+      moe_comp_method == CPUQuantMethod::MXFP4,
       w1_scale,
       w2_scale,
       block_size);
@@ -965,7 +981,7 @@ at::Tensor fused_experts_cpu(
   //   5. Aq_tmp : [M, K] or [M * topk, N]
   //   6. As_tmp : [M * topk]
   //
-  // for fp8 w8a16:
+  // for fp8 w8a16 and mxfp4:
   //   7. intermediate_cache0 : [M * topk, 2N]
   //   8. B_tmp : [T, MAX_CACHE_BLOCK_SIZE, BLOCK_N, std::max(K, N)]
   //
@@ -978,7 +994,7 @@ at::Tensor fused_experts_cpu(
   if (moe_comp_method == CPUQuantMethod::INT8_W8A8) {
     buffer_size_nbytes += std::max(M * K, M * topk * N) + M * topk * sizeof(float);
   }
-  if (moe_comp_method == CPUQuantMethod::FP8_W8A16) {
+  if (moe_comp_method == CPUQuantMethod::FP8_W8A16 || moe_comp_method == CPUQuantMethod::MXFP4) {
     buffer_size_nbytes += M * topk * 2 * N * 2 + num_threads * MAX_CACHE_BLOCK_SIZE * BLOCK_N * std::max(K, N) * 2;
   }
   if (moe_comp_method == CPUQuantMethod::INT4_W4A8) {
@@ -1031,9 +1047,11 @@ at::Tensor fused_experts_cpu(
       float* __restrict__ C_tmp = (float*)((void*)(A_tmp + num_threads * BLOCK_M * K));
       scalar_t* __restrict__ intermediate_cache0 = (scalar_t*)((void*)(C_tmp + num_threads * 2 * BLOCK_M * BLOCK_N));
       scalar_t* __restrict__ B_tmp = (scalar_t*)((void*)(intermediate_cache0 + M * topk * 2 * N));
+      bool with_bias = w1_bias.has_value();
+      auto act_func = alpha.has_value() && limit.has_value() ? CPUAcTMethod::swiglu : CPUAcTMethod::silu_and_mul;
 
       CHECK_MOE_SCALES_FP8(1, 2);
-      fused_experts_fp8_kernel_impl(
+      fused_experts_fp_kernel_impl<scalar_t, at::Float8_e4m3fn, float, false>(
           out_hidden_states.data_ptr<scalar_t>(),
           intermediate_cache0,
           intermediate_cache1,
@@ -1044,6 +1062,8 @@ at::Tensor fused_experts_cpu(
           hidden_states.data_ptr<scalar_t>(),
           packed_w1.data_ptr<at::Float8_e4m3fn>(),
           packed_w2.data_ptr<at::Float8_e4m3fn>(),
+          with_bias ? w1_bias.value().data_ptr<float>() : nullptr,
+          with_bias ? w2_bias.value().data_ptr<float>() : nullptr,
           w1s.data_ptr<float>(),
           w2s.data_ptr<float>(),
           block_size_N,
@@ -1057,7 +1077,56 @@ at::Tensor fused_experts_cpu(
           K,
           E,
           topk,
-          num_tokens_post_pad);
+          num_tokens_post_pad,
+          alpha.has_value() ? float(alpha.value()) : 0,
+          limit.has_value() ? float(limit.value()) : 0,
+          act_func,
+          with_bias);
+    } else if (moe_comp_method == CPUQuantMethod::MXFP4) {
+      scalar_t* __restrict__ A_tmp = (scalar_t*)((void*)(intermediate_cache2 + M * topk * K));
+      float* __restrict__ C_tmp = (float*)((void*)(A_tmp + num_threads * BLOCK_M * K));
+      scalar_t* __restrict__ intermediate_cache0 = (scalar_t*)((void*)(C_tmp + num_threads * 2 * BLOCK_M * BLOCK_N));
+      scalar_t* __restrict__ B_tmp = (scalar_t*)((void*)(intermediate_cache0 + M * topk * 2 * N));
+      bool with_bias = w1_bias.has_value();
+      auto act_func = alpha.has_value() && limit.has_value() ? CPUAcTMethod::swiglu : CPUAcTMethod::silu_and_mul;
+
+      // mxfp4 supports only group size of 32 (2^5)
+      constexpr int64_t group_size = 32;
+      auto w1s = w1_scale.value();
+      auto w2s = w2_scale.value();
+      TORCH_CHECK(w1s.numel() == E * 2 * N * K / group_size, "w1_scale size mismatch");
+      TORCH_CHECK(w2s.numel() == E * K * N / group_size, "w2_scale size mismatch");
+      fused_experts_fp_kernel_impl<scalar_t, uint8_t, uint8_t, true>(
+          out_hidden_states.data_ptr<scalar_t>(),
+          intermediate_cache0,
+          intermediate_cache1,
+          intermediate_cache2,
+          A_tmp,
+          B_tmp,
+          C_tmp,
+          hidden_states.data_ptr<scalar_t>(),
+          packed_w1.data_ptr<uint8_t>(),
+          packed_w2.data_ptr<uint8_t>(),
+          with_bias ? w1_bias.value().data_ptr<float>() : nullptr,
+          with_bias ? w2_bias.value().data_ptr<float>() : nullptr,
+          w1s.data_ptr<uint8_t>(),
+          w2s.data_ptr<uint8_t>(),
+          /*block_size_N*/ 1,
+          /*block_size_K*/ group_size,
+          topk_weights_.data_ptr<float>(),
+          sorted_ids,
+          expert_ids,
+          offsets,
+          M,
+          N,
+          K,
+          E,
+          topk,
+          num_tokens_post_pad,
+          alpha.has_value() ? float(alpha.value()) : 0,
+          limit.has_value() ? float(limit.value()) : 0,
+          act_func,
+          with_bias);
     } else if (moe_comp_method == CPUQuantMethod::INT4_W4A8) {
       uint8_t* __restrict__ A_tmp = (uint8_t*)((void*)(intermediate_cache2 + M * topk * K));
       float* __restrict__ C_tmp = (float*)((void*)(A_tmp + num_threads * BLOCK_M * K));
@@ -1185,7 +1254,7 @@ at::Tensor shared_expert_cpu(
   CHECK_EQ(packed_w2.size(1), packed_N);
 
   // check scales
-  check_moe_scales(use_int8_w8a8, use_fp8_w8a16, w1_scale, w2_scale, block_size);
+  check_moe_scales(use_int8_w8a8, use_fp8_w8a16, false, w1_scale, w2_scale, block_size);
 
   at::Tensor out_hidden_states = inplace ? hidden_states : at::empty_like(hidden_states);
 

csrc/cpu/sgl-kernels/moe.h

@@ -176,3 +176,109 @@ inline void silu_and_mul_stub(
     out_vec.store(out + d);
   }
 }
+
+template <typename scalar_t>
+inline void copy_mul_stub(scalar_t* __restrict__ out, const float* __restrict__ input, float weight, int64_t size) {
+  using bVec = at::vec::Vectorized<scalar_t>;
+  using fVec = at::vec::Vectorized<float>;
+  constexpr int kVecSize = bVec::size();
+  const fVec weight_vec = fVec(weight);
+  int64_t d;
+#pragma GCC unroll 4
+  for (d = 0; d <= size - kVecSize; d += kVecSize) {
+    fVec data0 = fVec::loadu(input + d) * weight_vec;
+    fVec data1 = fVec::loadu(input + d + fVec::size()) * weight_vec;
+    bVec out_vec = convert_from_float_ext<scalar_t>(data0, data1);
+    out_vec.store(out + d);
+  }
+  for (; d < size; ++d) {
+    out[d] = static_cast<scalar_t>(input[d] * weight);
+  }
+}
+
+// input = input + input2
+inline void add_bias_stub(float* __restrict__ input, const float* __restrict__ input2, int64_t size) {
+  using fVec = at::vec::Vectorized<float>;
+  constexpr int kVecSize = fVec::size();
+  int64_t d;
+#pragma GCC unroll 4
+  for (d = 0; d <= size - kVecSize; d += kVecSize) {
+    fVec x_fvec = fVec::loadu(input + d);
+    fVec y_fvec = fVec::loadu(input2 + d);
+    x_fvec = x_fvec + y_fvec;
+    x_fvec.store(input + d);
+  }
+  for (; d < size; ++d) {
+    input[d] = input[d] + input2[d];
+  }
+}
+
+template <typename scalar_t>
+inline void copy_mul_stub(scalar_t* __restrict__ out, const scalar_t* __restrict__ input, float weight, int64_t size) {
+  using bVec = at::vec::Vectorized<scalar_t>;
+  using fVec = at::vec::Vectorized<float>;
+  constexpr int kVecSize = bVec::size();
+  const fVec weight_vec = fVec(weight);
+  int64_t d;
+#pragma GCC unroll 4
+  for (d = 0; d <= size - kVecSize; d += kVecSize) {
+    bVec x = bVec::loadu(input + d);
+    fVec x0, x1;
+    std::tie(x0, x1) = at::vec::convert_to_float(x);
+    x0 = x0 * weight_vec;
+    x1 = x1 * weight_vec;
+    bVec out_vec = convert_from_float_ext<scalar_t>(x0, x1);
+    out_vec.store(out + d);
+  }
+  for (; d < size; ++d) {
+    out[d] = static_cast<scalar_t>(input[d] * weight);
+  }
+}
+
+template <typename scalar_t>
+inline void clamp_sigmoid_and_mul_stub(
+    scalar_t* __restrict__ out,
+    const scalar_t* __restrict__ input,
+    int64_t size,
+    const float alpha,
+    const float limit) {
+  using bVec = at::vec::Vectorized<scalar_t>;
+  using fVec = at::vec::Vectorized<float>;
+  const fVec one = fVec(1.f);
+  const fVec zero = fVec(0.f);
+  const fVec limit_v = fVec(limit);
+  const fVec nlimit_v = fVec(-limit);
+  const fVec alpha_v = fVec(alpha);
+
+  // no remainder
+#pragma GCC unroll 4
+  for (int64_t d = 0; d < size; d += bVec::size()) {
+    bVec x = bVec::loadu(input + d);
+    fVec x0_, y0_;
+    std::tie(x0_, y0_) = at::vec::convert_to_float(x);
+    float tmp_buffer[fVec::size() * 2];  // 32
+    float tmp_glu[fVec::size()];         // 16
+    float tmp_linear[fVec::size()];      // 16
+    x0_.store(tmp_buffer);
+    y0_.store(tmp_buffer + fVec::size());
+    // interleaved: x[2i] = glu, x[2i+1] = linear
+    for (int j = 0; j < fVec::size(); ++j) {
+      // x0 [0,2,..30]
+      tmp_glu[j] = tmp_buffer[j * 2];
+      // y0 [1,3,...31]
+      tmp_linear[j] = tmp_buffer[j * 2 + 1];
+    }
+    fVec x0 = fVec::loadu(tmp_glu);
+    fVec y0 = fVec::loadu(tmp_linear);
+
+    // clamp
+    x0 = at::vec::minimum(x0, limit_v);
+    y0 = at::vec::minimum(limit_v, at::vec::maximum(nlimit_v, y0));
+    // x * sigmoid(x * alpha)
+    x0 = x0 / (one + (x0 * alpha_v).neg().exp_u20());
+    // (y + 1) * x
+    y0 = y0 + one;
+    x0 = x0 * y0;
+    convert_from_float_and_store<scalar_t>(out + d / 2, x0);
+  }
+}

csrc/cpu/sgl-kernels/moe_fp8.cpp

@@ -7,8 +7,8 @@
 #include "gemm.h"
 #include "moe.h"
 
-template <typename scalar_t>
-void fused_experts_fp8_kernel_impl(
+template <typename scalar_t, typename packed_t, typename param_t, bool is_mxfp4>
+void fused_experts_fp_kernel_impl(
     scalar_t* __restrict__ output,
     scalar_t* __restrict__ ic0,
     scalar_t* __restrict__ ic1,
@@ -17,10 +17,12 @@ void fused_experts_fp8_kernel_impl(
     scalar_t* __restrict__ B_tmp,
     float* __restrict__ C_tmp,
     const scalar_t* __restrict__ input,
-    const at::Float8_e4m3fn* __restrict__ packed_w1,
-    const at::Float8_e4m3fn* __restrict__ packed_w2,
-    const float* __restrict__ w1s,
-    const float* __restrict__ w2s,
+    const packed_t* __restrict__ packed_w1,
+    const packed_t* __restrict__ packed_w2,
+    const float* __restrict__ w1_bias,
+    const float* __restrict__ w2_bias,
+    const param_t* __restrict__ w1s,
+    const param_t* __restrict__ w2s,
     int64_t block_size_N,
     int64_t block_size_K,
     const float* __restrict__ topk_weights,
@@ -32,7 +34,11 @@ void fused_experts_fp8_kernel_impl(
     int64_t K,
     int64_t E,
     int64_t topk,
-    int64_t num_tokens_post_pad) {
+    int64_t num_tokens_post_pad,
+    float alpha,
+    float limit,
+    CPUAcTMethod act_func,
+    bool with_bias) {
   constexpr int64_t BLOCK_M = block_size_m();
   constexpr int64_t BLOCK_N = block_size_n();
 
@@ -42,12 +48,20 @@ void fused_experts_fp8_kernel_impl(
   int64_t scale_size_N = div_up(2 * N, block_size_N);
   int64_t scale_size_K = div_up(K, block_size_K);
   int64_t blocks_n_per_group = block_size_N / BLOCK_N;
+  std::function<int64_t(int64_t)> scale_offset_per_block;
+  if constexpr (is_mxfp4) {
+    scale_offset_per_block = [&](int64_t a) { return a * BLOCK_N; };
+  } else {
+    scale_offset_per_block = [&](int64_t a) { return a / blocks_n_per_group; };
+  }
 
-  const int64_t stride_e = 2 * N * K;
-  const int64_t stride_n = K;
+  const int64_t packed_K = get_row_size<packed_t>(K);
+
+  const int64_t stride_e = 2 * N * packed_K;
+  const int64_t stride_n = packed_K;
 
   int64_t avg_M = std::max(int64_t(1), M * topk / E);
-  const bool use_brgemm = can_use_brgemm<at::Float8_e4m3fn>(avg_M);
+  const bool use_brgemm = can_use_brgemm<packed_t>(avg_M);
 
   int64_t B_tmp_size_per_thread = MAX_CACHE_BLOCK_SIZE * BLOCK_N * std::max(K, N);
 
@@ -57,14 +71,15 @@ void fused_experts_fp8_kernel_impl(
     int tid = get_thread_num();
     scalar_t* __restrict__ A = A_tmp + tid * BLOCK_M * K;
 
-    loop_2d<at::Float8_e4m3fn>(mb0, mb1, nb0, nb1, BLOCK_N * K, [&](int64_t mb, int64_t nb, int64_t nb_offset) {
+    loop_2d<packed_t>(mb0, mb1, nb0, nb1, BLOCK_N * K, [&](int64_t mb, int64_t nb, int64_t nb_offset) {
       int64_t n_size = std::min(2 * N - nb * BLOCK_N, BLOCK_N);
 
       // B shape [K, n_size] in vnni format
       int32_t expert_id = expert_ids[mb];
-      const at::Float8_e4m3fn* __restrict__ B = packed_w1 + expert_id * stride_e + nb * BLOCK_N * stride_n;
-      const float* __restrict__ Bs =
-          w1s + expert_id * scale_size_N * scale_size_K + (nb / blocks_n_per_group) * scale_size_K;
+      const packed_t* __restrict__ B = packed_w1 + expert_id * stride_e + nb * BLOCK_N * stride_n;
+      const param_t* __restrict__ Bs =
+          w1s + expert_id * scale_size_N * scale_size_K + scale_offset_per_block(nb) * scale_size_K;
+      const float* __restrict__ B_bias = with_bias ? w1_bias + expert_id * 2 * N + nb * BLOCK_N : nullptr;
 
       // do unpacking for the first row or a new expert
       int32_t pre_expert_id = mb == 0 ? -1 : expert_ids[mb - 1];
@@ -88,6 +103,7 @@ void fused_experts_fp8_kernel_impl(
           /*   C            */ ic0 + offset * 2 * N + nb * BLOCK_N,
           /*   Btmp         */ B_tmp + tid * B_tmp_size_per_thread + nb_offset * BLOCK_N * K,
           /*   Ctmp         */ C_tmp + tid * 2 * BLOCK_M * BLOCK_N,
+          /*   Bbias        */ B_bias,
           /*   scale        */ Bs,
           /*   M            */ m_size,
           /*   N            */ n_size,
@@ -106,12 +122,20 @@ void fused_experts_fp8_kernel_impl(
   });
 
   // stage 1.5: intermediate_cache1 = silu(intermediate_cache0)
-  at::parallel_for(0, M * topk, 0, [&](int64_t begin, int64_t end) {
-    for (int64_t m = begin; m < end; ++m) {
-      silu_and_mul_stub(ic1 + m * N, ic0 + m * 2 * N, ic0 + m * 2 * N + N, N);
-    }
-  });
-
+  if (act_func == CPUAcTMethod::silu_and_mul) {
+    at::parallel_for(0, M * topk, 0, [&](int64_t begin, int64_t end) {
+      for (int64_t m = begin; m < end; ++m) {
+        silu_and_mul_stub(ic1 + m * N, ic0 + m * 2 * N, ic0 + m * 2 * N + N, N);
+      }
+    });
+  } else if (act_func == CPUAcTMethod::swiglu) {
+    at::parallel_for(0, M * topk, 0, [&](int64_t begin, int64_t end) {
+      for (int64_t m = begin; m < end; ++m) {
+        clamp_sigmoid_and_mul_stub(ic1 + m * N, ic0 + m * 2 * N, N, alpha, limit);
+        clamp_sigmoid_and_mul_stub(ic1 + m * N + N / 2, ic0 + m * 2 * N + N, N, alpha, limit);
+      }
+    });
+  }
   // stage 2: intermediate_cache2 = intermediate_cache1 @ w2
   //   w2 : [E, K, N] as [E, OC, IC]
   const int64_t OC = K;  // rename K as OC
@@ -120,15 +144,16 @@ void fused_experts_fp8_kernel_impl(
   const int64_t NB2 = div_up(OC, BLOCK_N);
   scale_size_N = div_up(K, block_size_N);
   scale_size_K = div_up(N, block_size_K);
-  const int64_t stride_e2 = OC * IC;
-  const int64_t stride_oc = IC;
+  const int64_t packed_IC = get_row_size<packed_t>(IC);
+  const int64_t stride_e2 = OC * packed_IC;
+  const int64_t stride_oc = packed_IC;
 
   // parallel on [MB2, NB2]
   parallel_2d(MB2, NB2, [&](int64_t mb0, int64_t mb1, int64_t nb0, int64_t nb1) {
     int tid = get_thread_num();
     alignas(64) scalar_t C[BLOCK_M * BLOCK_K];
 
-    loop_2d<at::Float8_e4m3fn>(mb0, mb1, nb0, nb1, BLOCK_N * IC, [&](int64_t mb, int64_t nb, int64_t nb_offset) {
+    loop_2d<packed_t>(mb0, mb1, nb0, nb1, BLOCK_N * IC, [&](int64_t mb, int64_t nb, int64_t nb_offset) {
       int64_t m_size = offsets[mb + 1] - offsets[mb];
       int64_t n_size = std::min(OC - nb * BLOCK_N, BLOCK_N);
 
@@ -139,9 +164,10 @@ void fused_experts_fp8_kernel_impl(
 
       // B shape [IC, n_size] in vnni format
       int32_t expert_id = expert_ids[mb];
-      const at::Float8_e4m3fn* __restrict__ B = packed_w2 + expert_id * stride_e2 + nb * BLOCK_N * stride_oc;
-      const float* __restrict__ Bs =
-          w2s + expert_id * scale_size_N * scale_size_K + (nb / blocks_n_per_group) * scale_size_K;
+      const packed_t* __restrict__ B = packed_w2 + expert_id * stride_e2 + nb * BLOCK_N * stride_oc;
+      const param_t* __restrict__ Bs =
+          w2s + expert_id * scale_size_N * scale_size_K + scale_offset_per_block(nb) * scale_size_K;
+      const float* __restrict__ B_bias = with_bias ? w2_bias + expert_id * OC + nb * BLOCK_N : nullptr;
 
       // do unpacking for the first row or a new expert
       int32_t pre_expert_id = mb == 0 ? -1 : expert_ids[mb - 1];
@@ -153,6 +179,7 @@ void fused_experts_fp8_kernel_impl(
           /*   C            */ C,
           /*   Btmp         */ B_tmp + tid * B_tmp_size_per_thread + nb_offset * BLOCK_N * IC,
           /*   Ctmp         */ C_tmp + tid * 2 * BLOCK_M * BLOCK_N,
+          /*   Bbias        */ B_bias,
           /*   scale        */ Bs,
           /*   M            */ m_size,
           /*   N            */ n_size,
@@ -177,7 +204,6 @@ void fused_experts_fp8_kernel_impl(
       at::native::cpublas::brgemm_release();
     }
   });
-
   // stage 3: out = intermediate_cache2.sum(dim=1)
   //   from [M, topk, K] to [M, K]
   at::parallel_for(0, M, 0, [&](int64_t begin, int64_t end) {
@@ -187,35 +213,43 @@ void fused_experts_fp8_kernel_impl(
   });
 }
 
-#define INSTANTIATE_MOE_FP8_TEMPLATE(TYPE)             \
-  template void fused_experts_fp8_kernel_impl<TYPE>(   \
-      TYPE* __restrict__ output,                       \
-      TYPE* __restrict__ ic0,                          \
-      TYPE* __restrict__ ic1,                          \
-      TYPE* __restrict__ ic2,                          \
-      TYPE* __restrict__ A_tmp,                        \
-      TYPE* __restrict__ B_tmp,                        \
-      float* __restrict__ C_tmp,                       \
-      const TYPE* __restrict__ input,                  \
-      const at::Float8_e4m3fn* __restrict__ packed_w1, \
-      const at::Float8_e4m3fn* __restrict__ packed_w2, \
-      const float* __restrict__ w1s,                   \
-      const float* __restrict__ w2s,                   \
-      int64_t block_size_N,                            \
-      int64_t block_size_K,                            \
-      const float* __restrict__ topk_weights,          \
-      const int32_t* __restrict__ sorted_ids,          \
-      const int32_t* __restrict__ expert_ids,          \
-      const int32_t* __restrict__ offsets,             \
-      int64_t M,                                       \
-      int64_t N,                                       \
-      int64_t K,                                       \
-      int64_t E,                                       \
-      int64_t topk,                                    \
-      int64_t num_tokens_post_pad)
+#define INSTANTIATE_MOE_FP_TEMPLATE(TYPE1, TYPE2, TYPE3, IS_MXFP4)           \
+  template void fused_experts_fp_kernel_impl<TYPE1, TYPE2, TYPE3, IS_MXFP4>( \
+      TYPE1* __restrict__ output,                                            \
+      TYPE1* __restrict__ ic0,                                               \
+      TYPE1* __restrict__ ic1,                                               \
+      TYPE1* __restrict__ ic2,                                               \
+      TYPE1* __restrict__ A_tmp,                                             \
+      TYPE1* __restrict__ B_tmp,                                             \
+      float* __restrict__ C_tmp,                                             \
+      const TYPE1* __restrict__ input,                                       \
+      const TYPE2* __restrict__ packed_w1,                                   \
+      const TYPE2* __restrict__ packed_w2,                                   \
+      const float* __restrict__ w1_bias,                                     \
+      const float* __restrict__ w2_bias,                                     \
+      const TYPE3* __restrict__ w1s,                                         \
+      const TYPE3* __restrict__ w2s,                                         \
+      int64_t block_size_N,                                                  \
+      int64_t block_size_K,                                                  \
+      const float* __restrict__ topk_weights,                                \
+      const int32_t* __restrict__ sorted_ids,                                \
+      const int32_t* __restrict__ expert_ids,                                \
+      const int32_t* __restrict__ offsets,                                   \
+      int64_t M,                                                             \
+      int64_t N,                                                             \
+      int64_t K,                                                             \
+      int64_t E,                                                             \
+      int64_t topk,                                                          \
+      int64_t num_tokens_post_pad,                                           \
+      float alpha,                                                           \
+      float limit,                                                           \
+      CPUAcTMethod act_func,                                                 \
+      bool with_bias)
 
-INSTANTIATE_MOE_FP8_TEMPLATE(at::BFloat16);
-INSTANTIATE_MOE_FP8_TEMPLATE(at::Half);
+INSTANTIATE_MOE_FP_TEMPLATE(at::BFloat16, at::Float8_e4m3fn, float, false);
+INSTANTIATE_MOE_FP_TEMPLATE(at::Half, at::Float8_e4m3fn, float, false);
+INSTANTIATE_MOE_FP_TEMPLATE(at::BFloat16, uint8_t, uint8_t, true);
+INSTANTIATE_MOE_FP_TEMPLATE(at::Half, uint8_t, uint8_t, true);
 
 template <typename scalar_t>
 void shared_expert_fp8_kernel_impl(
@@ -266,6 +300,7 @@ void shared_expert_fp8_kernel_impl(
           /*   C            */ ic0 + mb * BLOCK_M * 2 * N + nb * BLOCK_N,
           /*   Btmp         */ B_tmp + tid * B_tmp_size_per_thread + nb_offset * BLOCK_N * K,
           /*   Ctmp         */ C_tmp + tid * 2 * BLOCK_M * BLOCK_N,
+          /*   Bbias        */ nullptr,
           /*   scale        */ w1s + (nb / blocks_n_per_group) * scale_size_K,
           /*   M            */ m_size,
           /*   N            */ n_size,
@@ -317,6 +352,7 @@ void shared_expert_fp8_kernel_impl(
           /*   C            */ C,
           /*   Btmp         */ B_tmp + tid * B_tmp_size_per_thread + nb_offset * BLOCK_N * IC,
           /*   Ctmp         */ C_tmp + tid * 2 * BLOCK_M * BLOCK_N,
+          /*   Bbias        */ nullptr,
           /*   scale        */ w2s + (nb / blocks_n_per_group) * scale_size_K,
           /*   M            */ m_size,
           /*   N            */ n_size,

csrc/cpu/sgl-kernels/vec.h

@@ -11,7 +11,9 @@
 
 #include <ATen/cpu/vec/functional.h>
 #include <ATen/cpu/vec/vec.h>
-
+#if defined(CPU_CAPABILITY_AVX512)
+#include <immintrin.h>
+#endif
 namespace {
 
 using namespace at::vec;
@@ -21,6 +23,15 @@ inline Vectorized<scalar_t> convert_from_float_ext(const Vectorized<float>& a, c
   return at::vec::convert_from_float<scalar_t>(a, b);
 }
 
+template <typename scalar_t>
+inline void convert_from_float_and_store(scalar_t* out, const Vectorized<float>& a) {
+  float out_buffer[at::vec::Vectorized<float>::size()];
+  a.store(out_buffer);
+  for (int i = 0; i < 16; i++) {
+    out[i] = (scalar_t)out_buffer[i];
+  }
+}
+
 // allow f16, bf16
 template <typename scalar_t, typename std::enable_if_t<is_reduced_floating_point_v<scalar_t>, int> = 1>
 inline std::tuple<Vectorized<float>, Vectorized<float>> load_float_vec2(const scalar_t* __restrict__ data) {
@@ -50,6 +61,11 @@ convert_from_float_ext<at::BFloat16>(const Vectorized<float>& a, const Vectorize
   return (__m512i)(_mm512_cvtne2ps_pbh(__m512(b), __m512(a)));
 }
 
+template <>
+inline void convert_from_float_and_store<at::BFloat16>(at::BFloat16* out, const Vectorized<float>& a) {
+  _mm256_storeu_si256((__m256i*)out, (__m256i)(_mm512_cvtneps_pbh(__m512(a))));
+}
+
 #define CVT_BF16_TO_FP32(a) _mm512_castsi512_ps(_mm512_slli_epi32(_mm512_cvtepu16_epi32(a), 16))
 
 #define CVT_FP16_TO_FP32(a) _mm512_cvtph_ps(a)
@@ -125,7 +141,6 @@ inline __m512bh CVT_FP8_TO_BF16(__m256i a) {
   return cvt_e4m3_bf16_intrinsic_with_denorm(a);
 #endif
 }
-
 // faster version of float8_e4m3fn conversion to bfloat16
 //
 // we mapped cuda implementation from below link and vectorized with avx512:

csrc/cpu/shm.cpp

@@ -5,7 +5,7 @@
 #include <sys/stat.h>
 #include <unistd.h>
 
-#ifdef __aarch64__
+#if defined(__aarch64__) || defined(__powerpc64__)
   #include <atomic>
 #endif
 
@@ -38,7 +38,7 @@ struct KernelVecType<c10::Half> {
 };
 
 struct ThreadSHMContext {
-#ifdef __aarch64__
+#if defined(__aarch64__) || defined(__powerpc64__)
   // memory model is weaker on AArch64, so we use atomic variables for
   // consumer (load-acquire) and producer (store-release) to make sure
   // that a stamp cannot be ready before the corresponding data is ready.
@@ -75,7 +75,7 @@ struct ThreadSHMContext {
     TORCH_CHECK(group_size <= MAX_SHM_RANK_NUM);
     TORCH_CHECK((size_t)this % 64 == 0);
     TORCH_CHECK((size_t)thread_shm_ptr % 64 == 0);
-#ifdef __aarch64__
+#if defined(__aarch64__) || defined(__powerpc64__)
     _curr_thread_stamp[0].store(1, std::memory_order_relaxed);
     _curr_thread_stamp[1].store(1, std::memory_order_relaxed);
     _ready_thread_stamp[0].store(0, std::memory_order_relaxed);
@@ -124,7 +124,7 @@ struct ThreadSHMContext {
   }
 
   char get_curr_stamp(int idx) const {
-#ifdef __aarch64__
+#if defined(__aarch64__) || defined(__powerpc64__)
     return _curr_thread_stamp[idx].load(std::memory_order_acquire);
 #else
     return _curr_thread_stamp[idx];
@@ -132,7 +132,7 @@ struct ThreadSHMContext {
   }
 
   char get_ready_stamp(int idx) const {
-#ifdef __aarch64__
+#if defined(__aarch64__) || defined(__powerpc64__)
     return _ready_thread_stamp[idx].load(std::memory_order_acquire);
 #else
     return _ready_thread_stamp[idx];
@@ -140,7 +140,7 @@ struct ThreadSHMContext {
   }
 
   void next_stamp() {
-#ifdef __aarch64__
+#if defined(__aarch64__) || defined(__powerpc64__)
     _curr_thread_stamp[local_stamp_buffer_idx].fetch_add(
         1, std::memory_order_release);
 #else
@@ -150,7 +150,7 @@ struct ThreadSHMContext {
   }
 
   void commit_ready_stamp() {
-#ifdef __aarch64__
+#if defined(__aarch64__) || defined(__powerpc64__)
     _ready_thread_stamp[local_stamp_buffer_idx].store(
         _curr_thread_stamp[local_stamp_buffer_idx].load(
             std::memory_order_relaxed),
@@ -186,8 +186,10 @@ struct ThreadSHMContext {
         break;
       }
       ++_spinning_count;
-#ifdef __aarch64__
+#if defined(__aarch64__)
       __asm__ __volatile__("yield");
+#elif defined(__powerpc64__)
+      __asm__ __volatile__("or 1,1,1");
 #else
       _mm_pause();
 #endif  // __aarch64__

csrc/cpu/torch_bindings.cpp

@@ -64,6 +64,8 @@ at::Tensor weight_packed_linear(at::Tensor& mat1, at::Tensor& mat2,
 
 at::Tensor convert_weight_packed(at::Tensor& weight);
 
+at::Tensor convert_scale_packed(at::Tensor& scale);
+
 at::Tensor fused_experts_cpu(
     at::Tensor& hidden_states, at::Tensor& w1, at::Tensor& w2,
     at::Tensor& topk_weights, at::Tensor& topk_ids, bool inplace,
@@ -71,7 +73,11 @@ at::Tensor fused_experts_cpu(
     const std::optional<at::Tensor>& w2_scale,
     const std::optional<at::Tensor>& w1_zero,
     const std::optional<at::Tensor>& w2_zero,
-    const std::optional<std::vector<int64_t>> block_size, bool is_vnni);
+    const std::optional<std::vector<int64_t>> block_size,
+    const std::optional<at::Tensor>& w1_bias,
+    const std::optional<at::Tensor>& w2_bias,
+    const std::optional<double>& alpha, const std::optional<double>& limit,
+    bool is_vnni);
 
 at::Tensor int8_scaled_mm_with_quant(at::Tensor& mat1, at::Tensor& mat2,
                                      at::Tensor& scales2,
@@ -98,6 +104,45 @@ at::Tensor int4_scaled_mm_cpu(at::Tensor& x, at::Tensor& w, at::Tensor& w_zeros,
                               at::Tensor& w_scales,
                               std::optional<at::Tensor> bias);
 
+// Adapted from sglang: GDN
+std::tuple<at::Tensor, at::Tensor> chunk_gated_delta_rule_cpu(
+    const at::Tensor& query, const at::Tensor& key, const at::Tensor& value,
+    const at::Tensor& g, const at::Tensor& beta,
+    const at::Tensor& initial_state, bool output_final_state,
+    const at::Tensor& cu_seqlens, bool head_first, bool use_qk_l2norm_in_kernel,
+    double eps = 1e-5);
+
+at::Tensor fused_sigmoid_gating_delta_rule_update_cpu(
+    const at::Tensor& A_log, const at::Tensor& dt_bias, const at::Tensor& q,
+    const at::Tensor& k, const at::Tensor& v, const at::Tensor& a,
+    const at::Tensor& b, at::Tensor& initial_state_source,
+    const at::Tensor& initial_state_indices, const at::Tensor& cu_seqlens,
+    bool use_qk_l2norm_in_kernel, double softplus_beta = 1.0,
+    double softplus_threshold = 20.0);
+
+std::tuple<at::Tensor, at::Tensor> fused_gdn_gating_cpu(
+    const at::Tensor& A_log, const at::Tensor& a, const at::Tensor& b,
+    const at::Tensor& dt_bias);
+
+// Adapted from sglang: casual_conv1d kernels
+at::Tensor causal_conv1d_weight_pack(const at::Tensor& weight);
+
+at::Tensor causal_conv1d_fwd_cpu(
+    const at::Tensor& x, const at::Tensor& weight,
+    const std::optional<at::Tensor>& bias,
+    const std::optional<at::Tensor>& conv_states,
+    const std::optional<at::Tensor>& query_start_loc,
+    const std::optional<at::Tensor>& cache_indices,
+    const std::optional<at::Tensor>& has_initial_state, bool silu_activation,
+    int64_t pad_slot_id, bool is_vnni);
+
+at::Tensor causal_conv1d_update_cpu(
+    const at::Tensor& x, const at::Tensor& conv_states,
+    const at::Tensor& weight, const std::optional<at::Tensor>& bias,
+    bool silu_activation, const std::optional<at::Tensor>& cache_seqlens,
+    const std::optional<at::Tensor>& conv_state_indices, int64_t pad_slot_id,
+    bool is_vnni);
+
 void activation_lut_bf16(torch::Tensor& out, torch::Tensor& input,
                          const std::string& activation);
 
@@ -333,7 +378,8 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
 #endif
 
 // SHM CCL
-#if defined(__AVX512F__) || (defined(__aarch64__) && !defined(__APPLE__))
+#if defined(__AVX512F__) || (defined(__aarch64__) && !defined(__APPLE__)) || \
+    defined(__powerpc64__)
   ops.def(
       "init_shm_manager(str name, int group_size, int rank, int thread_num) -> "
       "int",
@@ -365,11 +411,15 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
   ops.impl("weight_packed_linear", torch::kCPU, &weight_packed_linear);
   ops.def("convert_weight_packed(Tensor! weight) -> Tensor");
   ops.impl("convert_weight_packed", torch::kCPU, &convert_weight_packed);
+  ops.def("convert_scale_packed(Tensor! scale) -> Tensor");
+  ops.impl("convert_scale_packed", torch::kCPU, &convert_scale_packed);
   ops.def(
       "fused_experts_cpu(Tensor hidden_states, Tensor w1, Tensor w2, Tensor "
       "topk_weights, Tensor topk_ids, bool "
       "inplace, int moe_comp_method, Tensor? w1_scale, Tensor? w2_scale, "
-      "Tensor? w1_zero, Tensor? w2_zero, int[]? block_size, bool is_vnni) -> "
+      "Tensor? w1_zero, Tensor? w2_zero, int[]? block_size, "
+      "Tensor? w1_bias, Tensor? w2_bias, float? alpha, float? limit, "
+      "bool is_vnni) -> "
       "Tensor");
   ops.impl("fused_experts_cpu", torch::kCPU, &fused_experts_cpu);
   ops.def(
@@ -397,8 +447,49 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
       "scales2, SymInt[] block_size, Tensor? bias, ScalarType out_dtype, "
       "bool is_vnni) -> Tensor");
   ops.impl("fp8_scaled_mm_cpu", torch::kCPU, &fp8_scaled_mm_cpu);
+
+  // Adapted from sglang: casual_conv1d kernels
+  ops.def("causal_conv1d_weight_pack(Tensor weight) -> Tensor");
+  ops.impl("causal_conv1d_weight_pack", torch::kCPU,
+           &causal_conv1d_weight_pack);
+  ops.def(
+      "causal_conv1d_fwd_cpu(Tensor x, Tensor weight, Tensor? bias, Tensor? "
+      "conv_states, Tensor? query_start_loc,"
+      "Tensor? cache_indices, Tensor? has_initial_state, bool silu_activation, "
+      "int pad_slot_id, bool is_vnni) -> "
+      "Tensor");
+  ops.impl("causal_conv1d_fwd_cpu", torch::kCPU, &causal_conv1d_fwd_cpu);
+  ops.def(
+      "causal_conv1d_update_cpu(Tensor x, Tensor(a!) conv_states, Tensor "
+      "weight, Tensor? bias, bool silu_activation,"
+      "Tensor? cache_seqlens, Tensor? conv_state_indices, int pad_slot_id, "
+      "bool is_vnni) -> Tensor");
+  ops.impl("causal_conv1d_update_cpu", torch::kCPU, &causal_conv1d_update_cpu);
 #endif
 
+  // Adapted from sglang: GDN kernels
+  ops.def(
+      "chunk_gated_delta_rule_cpu(Tensor query, Tensor key, Tensor value, "
+      "Tensor g, Tensor beta, "
+      "Tensor initial_state, bool output_final_state, Tensor cu_seqlens, bool "
+      "head_first, "
+      "bool use_qk_l2norm_in_kernel, float eps=1e-5) -> (Tensor, Tensor)");
+  ops.impl("chunk_gated_delta_rule_cpu", torch::kCPU,
+           &chunk_gated_delta_rule_cpu);
+  ops.def(
+      "fused_sigmoid_gating_delta_rule_update_cpu(Tensor A_log, Tensor "
+      "dt_bias, Tensor q, Tensor k, Tensor v, Tensor "
+      "a, Tensor b, Tensor(a!) initial_state_source, Tensor "
+      "initial_state_indices, Tensor cu_seqlens, bool "
+      "use_qk_l2norm_in_kernel, float softplus_beta=1.0, float "
+      "softplus_threshold=20.0) -> Tensor");
+  ops.impl("fused_sigmoid_gating_delta_rule_update_cpu", torch::kCPU,
+           &fused_sigmoid_gating_delta_rule_update_cpu);
+  ops.def(
+      "fused_gdn_gating_cpu(Tensor A_log, Tensor a, Tensor b, Tensor dt_bias) "
+      "-> (Tensor, Tensor)");
+  ops.impl("fused_gdn_gating_cpu", torch::kCPU, &fused_gdn_gating_cpu);
+
   // CPU attention kernels
   ops.def(
       "get_scheduler_metadata(int num_req, int num_heads_q, int num_heads_kv, "
@@ -428,7 +519,7 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
   ops.def("dynamic_per_token_scaled_fp8_quant() -> ()", placeholder_op);
 
   // WNA16
-#if defined(__AVX512F__)
+#if defined(__AVX512F__) || defined(__riscv_v)
   ops.def(
       "cpu_gemm_wna16(Tensor input, Tensor q_weight, Tensor(a2!) output, "
       "Tensor scales, Tensor? zeros, Tensor? g_idx, Tensor? bias, SymInt "

csrc/cutlass_extensions/torch_utils.hpp

@@ -1,5 +1,7 @@
 #pragma once
 
+#include "torch_utils.h"
+
 // This header is shared between _C (unstable ABI, used by machete) and
 // _C_stable_libtorch (stable ABI, used by W4A8/sparse). TORCH_TARGET_VERSION
 // is defined only for the stable target, so we switch includes and types
@@ -8,13 +10,9 @@
   #include <torch/csrc/stable/tensor.h>
   #include <torch/headeronly/util/BFloat16.h>
   #include <torch/headeronly/util/Half.h>
-  #include <torch/headeronly/util/shim_utils.h>  // for STD_TORCH_CHECK
 using TorchTensor = torch::stable::Tensor;
-  #define TORCH_UTILS_CHECK STD_TORCH_CHECK
 #else
-  #include <torch/all.h>
 using TorchTensor = torch::Tensor;
-  #define TORCH_UTILS_CHECK TORCH_CHECK
 #endif
 
 #include "cute/layout.hpp"

csrc/fused_deepseek_v4_qnorm_rope_kv_insert_kernel.cu

@@ -87,6 +87,12 @@ constexpr int kScaleBytesPerToken = kNumQuantBlocks + 1;  // 8 (7 real + 1 pad)
 constexpr int kTokenDataBytes = kNopeDim + kRopeDim * 2;  // 448 + 128 = 576
 constexpr float kFp8Max = 448.0f;
 
+#ifndef USE_ROCM
+// When num_tokens is less than this threshold,
+// run the reduced grid variant on cuda
+constexpr float NUM_TOKEN_CUTOFF = 1024;
+#endif
+
 // Per-warp layout:  32 lanes × 16 elems/lane = 512 elems = HEAD_DIM.
 constexpr int kNumLanes = 32;
 constexpr int kElemsPerLane = kHeadDim / kNumLanes;  // 16
@@ -112,24 +118,255 @@ __device__ __forceinline__ float warpSum(float val) {
   return val;
 }
 
+// ────────────────────────────────────────────────────────────────────────────
+// Per-slot inner pipeline
+// ────────────────────────────────────────────────────────────────────────────
+// Shared by both kernel variants: 1 CTA per (token, head) pair vs. 1 CTA per
+// token.  Templated on `kNumHeadsQPadded` so the KV-sentinel comparison and
+// q_out stride fold to compile-time constants.
+//
+// Slot layout (per token):
+//   slot < num_heads_q                          → live-Q   (RMSNorm + RoPE,
+//                                                           read q_in →
+//                                                           write q_out)
+//   num_heads_q <= slot < kNumHeadsQPadded      → pad-Q    (zero-fill q_out;
+//                                                           v0/v1 unused)
+//   slot == kNumHeadsQPadded                    → KV       (RoPE + UE8M0 quant
+//                                                           + paged-cache
+//                                                           insert)
+template <typename scalar_t_in, int kNumHeadsQPadded>
+__device__ __forceinline__ void processDeepseekV4Slot(
+    uint4 v0, uint4 v1, int const tokenIdx, int const slotIdx,
+    int const dim_base, int const laneId, int const num_heads_q,
+    float const eps, scalar_t_in* __restrict__ q_out,
+    uint8_t* __restrict__ k_cache, int64_t const* __restrict__ slot_mapping,
+    int64_t const* __restrict__ position_ids,
+    float const* __restrict__ cos_sin_cache, int const cache_block_size,
+    int const kv_block_stride) {
+  using Converter = vllm::_typeConvert<scalar_t_in>;
+  bool const isKV = (slotIdx == kNumHeadsQPadded);
+  bool const isPadQ = !isKV && (slotIdx >= num_heads_q);
+
+  // ── Pad-Q branch: write 32 B of zeros and exit. ─────────────────────────
+  // FlashMLA reads these slots; bf16 +0.0 is bit pattern 0x0000, so a uint4
+  // zero literal is correct.  Matches the live-Q branch's vectorized store.
+  if (isPadQ) {
+    scalar_t_in* dst =
+        q_out +
+        (static_cast<int64_t>(tokenIdx) * kNumHeadsQPadded + slotIdx) *
+            kHeadDim +
+        dim_base;
+    uint4 const zero4 = {0u, 0u, 0u, 0u};
+    *reinterpret_cast<uint4*>(dst) = zero4;
+    *reinterpret_cast<uint4*>(dst + 8) = zero4;
+    return;
+  }
+
+  // ── Decode the bf16 → 16 fp32 registers ─────────────────────────────
+  float elements[kElemsPerLane];
+  {
+    typename Converter::packed_hip_type const* p0 =
+        reinterpret_cast<typename Converter::packed_hip_type const*>(&v0);
+    typename Converter::packed_hip_type const* p1 =
+        reinterpret_cast<typename Converter::packed_hip_type const*>(&v1);
+#pragma unroll
+    for (int i = 0; i < 4; i++) {
+      float2 f2 = Converter::convert(p0[i]);
+      elements[2 * i] = f2.x;
+      elements[2 * i + 1] = f2.y;
+    }
+#pragma unroll
+    for (int i = 0; i < 4; i++) {
+      float2 f2 = Converter::convert(p1[i]);
+      elements[8 + 2 * i] = f2.x;
+      elements[8 + 2 * i + 1] = f2.y;
+    }
+  }
+
+  // ── Q branch: RMSNorm (no weight) ───────────────────────────────────
+  if (!isKV) {
+    float sumOfSquares = 0.0f;
+#pragma unroll
+    for (int i = 0; i < kElemsPerLane; i++) {
+      sumOfSquares += elements[i] * elements[i];
+    }
+    sumOfSquares = warpSum<float>(sumOfSquares);
+    float const rms_rcp =
+        rsqrtf(sumOfSquares / static_cast<float>(kHeadDim) + eps);
+#pragma unroll
+    for (int i = 0; i < kElemsPerLane; i++) {
+      elements[i] = elements[i] * rms_rcp;
+    }
+  }
+
+  // ── GPT-J RoPE on dims [NOPE_DIM, HEAD_DIM) ─────────────────────────────
+  // All math in fp32.  cos_sin_cache is loaded as fp32 (its native storage).
+  bool const is_rope_lane = dim_base >= kNopeDim;
+  if (is_rope_lane) {
+    int64_t const pos = position_ids[tokenIdx];
+    constexpr int kHalfRope = kRopeDim / 2;
+    float const* cos_ptr = cos_sin_cache + pos * kRopeDim;
+    float const* sin_ptr = cos_ptr + kHalfRope;
+
+    int const rope_local_base = dim_base - kNopeDim;
+    int const half_base = rope_local_base >> 1;
+
+    // Load phase: 4 vectorized LDGs issue back-to-back.
+    float4 const c0 = *reinterpret_cast<float4 const*>(cos_ptr + half_base);
+    float4 const c1 = *reinterpret_cast<float4 const*>(cos_ptr + half_base + 4);
+    float4 const s0 = *reinterpret_cast<float4 const*>(sin_ptr + half_base);
+    float4 const s1 = *reinterpret_cast<float4 const*>(sin_ptr + half_base + 4);
+    float const cos_arr[8] = {c0.x, c0.y, c0.z, c0.w, c1.x, c1.y, c1.z, c1.w};
+    float const sin_arr[8] = {s0.x, s0.y, s0.z, s0.w, s1.x, s1.y, s1.z, s1.w};
+
+#pragma unroll
+    for (int p = 0; p < kElemsPerLane / 2; p++) {
+      float const x_even = elements[2 * p];
+      float const x_odd = elements[2 * p + 1];
+      elements[2 * p] = x_even * cos_arr[p] - x_odd * sin_arr[p];
+      elements[2 * p + 1] = x_even * sin_arr[p] + x_odd * cos_arr[p];
+    }
+  }
+
+  // ═══════════════════════════════════════════════════════════════════
+  // Q / KV branch dispatch. Restructured as if/else (no early `return`)
+  // so every code path lands at the same exit point — callers own PDL
+  // triggering and per-iteration buffer rotation.
+  // ═══════════════════════════════════════════════════════════════════
+  if (!isKV) {
+    // ── Live-Q: cast back to bf16 and store into the padded q_out. ─────
+    uint4 out0, out1;
+    typename Converter::packed_hip_type* po0 =
+        reinterpret_cast<typename Converter::packed_hip_type*>(&out0);
+    typename Converter::packed_hip_type* po1 =
+        reinterpret_cast<typename Converter::packed_hip_type*>(&out1);
+#pragma unroll
+    for (int i = 0; i < 4; i++) {
+      po0[i] =
+          Converter::convert(make_float2(elements[2 * i], elements[2 * i + 1]));
+    }
+#pragma unroll
+    for (int i = 0; i < 4; i++) {
+      po1[i] = Converter::convert(
+          make_float2(elements[8 + 2 * i], elements[8 + 2 * i + 1]));
+    }
+    scalar_t_in* dst =
+        q_out +
+        (static_cast<int64_t>(tokenIdx) * kNumHeadsQPadded + slotIdx) *
+            kHeadDim +
+        dim_base;
+    *reinterpret_cast<uint4*>(dst) = out0;
+    *reinterpret_cast<uint4*>(dst + 8) = out1;
+  } else {
+    // ── KV: FP8 quant on NoPE + bf16 store on RoPE + cache insert.
+    int64_t const slot_id = slot_mapping[tokenIdx];
+    if (slot_id >= 0) {
+      int64_t const block_idx = slot_id / cache_block_size;
+      int64_t const pos_in_block = slot_id % cache_block_size;
+      uint8_t* block_base =
+          k_cache + block_idx * static_cast<int64_t>(kv_block_stride);
+      uint8_t* token_fp8_ptr = block_base + pos_in_block * kTokenDataBytes;
+      uint8_t* token_bf16_ptr = token_fp8_ptr + kNopeDim;
+      uint8_t* token_scale_ptr =
+          block_base +
+          static_cast<int64_t>(cache_block_size) * kTokenDataBytes +
+          pos_in_block * kScaleBytesPerToken;
+
+#pragma unroll
+      for (int i = 0; i < kElemsPerLane; i++) {
+        elements[i] = Converter::convert(Converter::convert(elements[i]));
+      }
+
+      float local_absmax = 0.0f;
+#pragma unroll
+      for (int i = 0; i < kElemsPerLane; i++) {
+        local_absmax = fmaxf(local_absmax, fabsf(elements[i]));
+      }
+      float const absmax = fmaxf(warp4MaxAbs(local_absmax), 1e-4f);
+      float const exponent = ceilf(log2f(absmax / kFp8Max));
+      float const inv_scale = exp2f(-exponent);
+
+      if (!is_rope_lane) {
+        uint8_t out_bytes[kElemsPerLane];
+#pragma unroll
+        for (int i = 0; i < kElemsPerLane; i++) {
+          float scaled = elements[i] * inv_scale;
+          scaled = fminf(fmaxf(scaled, -kFp8Max), kFp8Max);
+#ifndef USE_ROCM
+          __nv_fp8_storage_t s =
+              __nv_cvt_float_to_fp8(scaled, __NV_SATFINITE, __NV_E4M3);
+          out_bytes[i] = static_cast<uint8_t>(s);
+#else
+          out_bytes[i] = rocm_cvt_float_to_fp8_e4m3(scaled);
+#endif
+        }
+        *reinterpret_cast<uint4*>(token_fp8_ptr + dim_base) =
+            *reinterpret_cast<uint4 const*>(out_bytes);
+
+        if ((laneId & 3) == 0) {
+          int const q_block_idx = laneId >> 2;
+          float encoded = fmaxf(fminf(exponent + 127.0f, 255.0f), 0.0f);
+          token_scale_ptr[q_block_idx] = static_cast<uint8_t>(encoded);
+        }
+        if (laneId == 0) {
+          token_scale_ptr[kNumQuantBlocks] = 0;
+        }
+      } else {
+        uint4 out0, out1;
+        typename Converter::packed_hip_type* po0 =
+            reinterpret_cast<typename Converter::packed_hip_type*>(&out0);
+        typename Converter::packed_hip_type* po1 =
+            reinterpret_cast<typename Converter::packed_hip_type*>(&out1);
+#pragma unroll
+        for (int i = 0; i < 4; i++) {
+          po0[i] = Converter::convert(
+              make_float2(elements[2 * i], elements[2 * i + 1]));
+        }
+#pragma unroll
+        for (int i = 0; i < 4; i++) {
+          po1[i] = Converter::convert(
+              make_float2(elements[8 + 2 * i], elements[8 + 2 * i + 1]));
+        }
+        int const rope_local_base = dim_base - kNopeDim;
+        scalar_t_in* bf16_dst =
+            reinterpret_cast<scalar_t_in*>(token_bf16_ptr) + rope_local_base;
+        *reinterpret_cast<uint4*>(bf16_dst) = out0;
+        *reinterpret_cast<uint4*>(bf16_dst + 8) = out1;
+      }
+    }
+  }
+}
+
 // ────────────────────────────────────────────────────────────────────────────
 // Kernel
 // ────────────────────────────────────────────────────────────────────────────
 //
-// Grid: 1D, gridDim.x = ceil(num_tokens_full * (num_heads_q + 1) /
+// Grid: 1D, gridDim.x = ceil(num_tokens_full * (kNumHeadsQPadded + 1) /
 // warps_per_block) Block: blockDim.x = 256 threads (8 warps per block) Each
-// warp handles one (token, head_slot) pair. head_slot < num_heads_q          →
-// Q branch (RMSNorm + RoPE, in place) head_slot == num_heads_q         → KV
-// branch (RoPE + UE8M0 quant + insert)
+// warp handles one (token, head_slot) pair.
+//   slot < num_heads_q                              → live-Q branch
+//                                                     (RMSNorm + RoPE,
+//                                                      read q_in → write q_out)
+//   num_heads_q <= slot < kNumHeadsQPadded          → pad-Q branch
+//                                                     (zero-fill q_out)
+//   slot == kNumHeadsQPadded                        → KV branch
+//                                                     (RoPE + UE8M0 quant +
+//                                                      paged-cache insert)
+//
+// `kNumHeadsQPadded` is a template parameter (compile-time constant) so the
+// divisions in the grid math and the KV-sentinel comparison fold to fast
+// constant operations.  The launch wrapper dispatches the runtime value to
+// the matching instantiation.
 //
 // With DP padding, q/kv/position_ids can have more rows than slot_mapping.
-// The Q branch covers all `num_tokens_full` rows (downstream attention uses
-// them).  The KV branch only inserts the first `num_tokens_insert` tokens
-// (= slot_mapping length) into the paged cache.
+// The live-Q and pad-Q branches cover all `num_tokens_full` rows (downstream
+// attention uses them).  The KV branch only inserts the first
+// `num_tokens_insert` tokens (= slot_mapping length) into the paged cache.
 //
-template <typename scalar_t_in>
+template <typename scalar_t_in, int kNumHeadsQPadded>
 __global__ void fusedDeepseekV4QNormRopeKVRopeQuantInsertKernel(
-    scalar_t_in* __restrict__ q_inout,         // [N, H, 512] bf16, in place
+    scalar_t_in const* __restrict__ q_in,      // [N, num_heads_q,      512]
+    scalar_t_in* __restrict__ q_out,           // [N, kNumHeadsQPadded, 512]
     scalar_t_in const* __restrict__ kv_in,     // [N, 512] bf16
     uint8_t* __restrict__ k_cache,             // [num_blocks, block_stride]
     int64_t const* __restrict__ slot_mapping,  // [num_tokens_insert] i64
@@ -138,7 +375,7 @@ __global__ void fusedDeepseekV4QNormRopeKVRopeQuantInsertKernel(
     float const eps,
     int const num_tokens_full,    // = q.size(0) = kv.size(0)
     int const num_tokens_insert,  // = slot_mapping.size(0), ≤ num_tokens_full
-    int const num_heads_q,        // H
+    int const num_heads_q,        // live Q heads (input layout)
     int const cache_block_size,   // tokens per paged-cache block
     int const kv_block_stride) {  // bytes per paged-cache block
 #if (!defined(__CUDA_ARCH__) || __CUDA_ARCH__ < 800) && !defined(USE_ROCM)
@@ -149,19 +386,18 @@ __global__ void fusedDeepseekV4QNormRopeKVRopeQuantInsertKernel(
     return;
   } else {
 #endif
-    using Converter = vllm::_typeConvert<scalar_t_in>;
-
     int const warpsPerBlock = blockDim.x / 32;
     int const warpId = threadIdx.x / 32;
     int const laneId = threadIdx.x % 32;
     int const globalWarpIdx = blockIdx.x * warpsPerBlock + warpId;
 
-    int const total_slots_per_token = num_heads_q + 1;
-    int const tokenIdx = globalWarpIdx / total_slots_per_token;
-    int const slotIdx = globalWarpIdx % total_slots_per_token;
+    constexpr int kTotalSlotsPerToken = kNumHeadsQPadded + 1;
+    int const tokenIdx = globalWarpIdx / kTotalSlotsPerToken;
+    int const slotIdx = globalWarpIdx % kTotalSlotsPerToken;
     if (tokenIdx >= num_tokens_full) return;
 
-    bool const isKV = (slotIdx == num_heads_q);
+    bool const isKV = (slotIdx == kNumHeadsQPadded);
+    bool const isPadQ = !isKV && (slotIdx >= num_heads_q);
     // KV branch: skip DP-padded tokens (no slot reserved for them).
     if (isKV && tokenIdx >= num_tokens_insert) return;
 
@@ -176,209 +412,123 @@ __global__ void fusedDeepseekV4QNormRopeKVRopeQuantInsertKernel(
     // Dim range this lane owns within the 512-wide head.
     int const dim_base = laneId * kElemsPerLane;  // in [0, 512) step 16
 
-    // ── Load 16 bf16 → 16 fp32 registers (one 16-byte + one 16-byte LDG) ────
-    float elements[kElemsPerLane];
-    float sumOfSquares = 0.0f;
-
-    scalar_t_in const* src_ptr;
-    if (isKV) {
-      src_ptr = kv_in + static_cast<int64_t>(tokenIdx) * kHeadDim + dim_base;
-    } else {
-      int64_t const q_row_offset =
-          (static_cast<int64_t>(tokenIdx) * num_heads_q + slotIdx) * kHeadDim +
-          dim_base;
-      src_ptr = q_inout + q_row_offset;
+    // Load only for live-Q and KV slots; pad-Q skips the read (q_in beyond
+    // num_heads_q is out of bounds) and the helper zero-fills its output.
+    uint4 v0, v1;
+    if (!isPadQ) {
+      scalar_t_in const* src_ptr;
+      if (isKV) {
+        src_ptr = kv_in + static_cast<int64_t>(tokenIdx) * kHeadDim + dim_base;
+      } else {
+        int64_t const q_row_offset =
+            (static_cast<int64_t>(tokenIdx) * num_heads_q + slotIdx) *
+                kHeadDim +
+            dim_base;
+        src_ptr = q_in + q_row_offset;
+      }
+      v0 = *reinterpret_cast<uint4 const*>(src_ptr);
+      v1 = *reinterpret_cast<uint4 const*>(src_ptr + 8);
     }
 
-    // Two 16-byte loads per thread (8 bf16 each).  Use uint4 as the vector
-    // type and bitcast to scalar_t_in packed pairs for conversion.
-    uint4 v0 = *reinterpret_cast<uint4 const*>(src_ptr);
-    uint4 v1 = *reinterpret_cast<uint4 const*>(src_ptr + 8);
+    processDeepseekV4Slot<scalar_t_in, kNumHeadsQPadded>(
+        v0, v1, tokenIdx, slotIdx, dim_base, laneId, num_heads_q, eps, q_out,
+        k_cache, slot_mapping, position_ids, cos_sin_cache, cache_block_size,
+        kv_block_stride);
 
-    {
-      typename Converter::packed_hip_type const* p0 =
-          reinterpret_cast<typename Converter::packed_hip_type const*>(&v0);
-      typename Converter::packed_hip_type const* p1 =
-          reinterpret_cast<typename Converter::packed_hip_type const*>(&v1);
-// Each packed_hip_type holds 2 bf16 → 4 packed = 8 elems per uint4.
-#pragma unroll
-      for (int i = 0; i < 4; i++) {
-        float2 f2 = Converter::convert(p0[i]);
-        elements[2 * i] = f2.x;
-        elements[2 * i + 1] = f2.y;
-      }
-#pragma unroll
-      for (int i = 0; i < 4; i++) {
-        float2 f2 = Converter::convert(p1[i]);
-        elements[8 + 2 * i] = f2.x;
-        elements[8 + 2 * i + 1] = f2.y;
-      }
-    }
-
-    // ── Q branch: RMSNorm with no weight (has_weight=False) ─────────────────
-    // Variance + rsqrt + multiply all in fp32, no intermediate bf16 round.
-    // The downstream bf16 round only happens at the final store.
-    if (!isKV) {
-#pragma unroll
-      for (int i = 0; i < kElemsPerLane; i++) {
-        sumOfSquares += elements[i] * elements[i];
-      }
-      sumOfSquares = warpSum<float>(sumOfSquares);
-      float const rms_rcp =
-          rsqrtf(sumOfSquares / static_cast<float>(kHeadDim) + eps);
-#pragma unroll
-      for (int i = 0; i < kElemsPerLane; i++) {
-        elements[i] = elements[i] * rms_rcp;
-      }
-    }
-
-    // ── GPT-J RoPE on dims [NOPE_DIM, HEAD_DIM) ─────────────────────────────
-    // All math in fp32.  cos_sin_cache is loaded as fp32 (its native storage).
-    bool const is_rope_lane = dim_base >= kNopeDim;
-    if (is_rope_lane) {
-      int64_t const pos = position_ids[tokenIdx];
-      constexpr int kHalfRope = kRopeDim / 2;  // 32
-      float const* cos_ptr = cos_sin_cache + pos * kRopeDim;
-      float const* sin_ptr = cos_ptr + kHalfRope;
-
-      int const rope_local_base = dim_base - kNopeDim;  // in [0, 64) step 16
-#pragma unroll
-      for (int p = 0; p < kElemsPerLane / 2; p++) {
-        int const pair_dim = rope_local_base + 2 * p;
-        int const half_idx = pair_dim / 2;
-        float const cos_v = VLLM_LDG(cos_ptr + half_idx);
-        float const sin_v = VLLM_LDG(sin_ptr + half_idx);
-        float const x_even = elements[2 * p];
-        float const x_odd = elements[2 * p + 1];
-        elements[2 * p] = x_even * cos_v - x_odd * sin_v;
-        elements[2 * p + 1] = x_even * sin_v + x_odd * cos_v;
-      }
-    }
-
-    // ═══════════════════════════════════════════════════════════════════════
-    // Q branch: cast to bf16 and store back in place.
-    // ═══════════════════════════════════════════════════════════════════════
-    if (!isKV) {
-      uint4 out0, out1;
-      typename Converter::packed_hip_type* po0 =
-          reinterpret_cast<typename Converter::packed_hip_type*>(&out0);
-      typename Converter::packed_hip_type* po1 =
-          reinterpret_cast<typename Converter::packed_hip_type*>(&out1);
-#pragma unroll
-      for (int i = 0; i < 4; i++) {
-        po0[i] = Converter::convert(
-            make_float2(elements[2 * i], elements[2 * i + 1]));
-      }
-#pragma unroll
-      for (int i = 0; i < 4; i++) {
-        po1[i] = Converter::convert(
-            make_float2(elements[8 + 2 * i], elements[8 + 2 * i + 1]));
-      }
-      scalar_t_in* dst =
-          q_inout +
-          (static_cast<int64_t>(tokenIdx) * num_heads_q + slotIdx) * kHeadDim +
-          dim_base;
-      *reinterpret_cast<uint4*>(dst) = out0;
-      *reinterpret_cast<uint4*>(dst + 8) = out1;
 #if defined(__CUDA_ARCH__) && (__CUDA_ARCH__ >= 900)
-      cudaTriggerProgrammaticLaunchCompletion();
+    cudaTriggerProgrammaticLaunchCompletion();
 #endif
-      return;
-    }
+#if (!defined(__CUDA_ARCH__) || __CUDA_ARCH__ < 800) && !defined(USE_ROCM)
+  }
+#endif
+}
+
+// ────────────────────────────────────────────────────────────────────────────
+// Kernel
+// ────────────────────────────────────────────────────────────────────────────
+//
+// Grid: 1D, gridDim.x = num_tokens_full
+// Block: blockDim.x = 256 threads (8 warps per block) Each
+// warp handles one token, iterating over each head.
+// Q branch (RMSNorm + RoPE, in place) head_slot == num_heads_q
+// KV branch (RoPE + UE8M0 quant + insert)
+//
+template <typename scalar_t_in, int kNumHeadsQPadded>
+__global__ void fusedDeepseekV4QNormRopeKVRopeQuantInsertKernelReducedGrid(
+    scalar_t_in const* __restrict__ q_in, scalar_t_in* __restrict__ q_out,
+    scalar_t_in const* __restrict__ kv_in, uint8_t* __restrict__ k_cache,
+    int64_t const* __restrict__ slot_mapping,
+    int64_t const* __restrict__ position_ids,
+    float const* __restrict__ cos_sin_cache, float const eps,
+    int const num_tokens_full, int const num_tokens_insert,
+    int const num_heads_q, int const cache_block_size,
+    int const kv_block_stride) {
+#if (!defined(__CUDA_ARCH__) || __CUDA_ARCH__ < 800) && !defined(USE_ROCM)
+  if constexpr (std::is_same_v<scalar_t_in, c10::BFloat16>) {
+    return;
+  } else {
+#endif
+    int const warpsPerBlock = blockDim.x / 32;
+    int const warpId = threadIdx.x / 32;
+    int const laneId = threadIdx.x % 32;
+
+    int const tokenIdx = blockIdx.x;
+    if (tokenIdx >= num_tokens_full) return;
 
-    // ═══════════════════════════════════════════════════════════════════════
-    // KV branch.
-    // ═══════════════════════════════════════════════════════════════════════
-    int64_t const slot_id = slot_mapping[tokenIdx];
-    if (slot_id < 0) {
 #if defined(__CUDA_ARCH__) && (__CUDA_ARCH__ >= 900)
-      cudaTriggerProgrammaticLaunchCompletion();
+    cudaGridDependencySynchronize();
 #endif
-      return;
-    }
 
-    int64_t const block_idx = slot_id / cache_block_size;
-    int64_t const pos_in_block = slot_id % cache_block_size;
-    uint8_t* block_base =
-        k_cache + block_idx * static_cast<int64_t>(kv_block_stride);
-    uint8_t* token_fp8_ptr = block_base + pos_in_block * kTokenDataBytes;
-    uint8_t* token_bf16_ptr = token_fp8_ptr + kNopeDim;
-    uint8_t* token_scale_ptr =
-        block_base + static_cast<int64_t>(cache_block_size) * kTokenDataBytes +
-        pos_in_block * kScaleBytesPerToken;
+    int const dim_base = laneId * kElemsPerLane;  // in [0, 512) step 16
+    // Slot enumeration: live-Q + pad-Q + (KV if this token has a slot).
+    int const slot_end = (tokenIdx >= num_tokens_insert)
+                             ? kNumHeadsQPadded
+                             : (kNumHeadsQPadded + 1);
 
-    // Round K to bf16 first, matching the unfused reference path where K is
-    // materialized as bf16 before K quantization.  absmax, clamp, and FP8
-    // quant below all run on these bf16-rounded values.
-#pragma unroll
-    for (int i = 0; i < kElemsPerLane; i++) {
-      elements[i] = Converter::convert(Converter::convert(elements[i]));
-    }
+    auto load_slot = [&](int s, uint4& va, uint4& vb) {
+      // pad-Q slots skip the load — q_in beyond num_heads_q is OOB.
+      if (s >= num_heads_q && s < kNumHeadsQPadded) return;
+      scalar_t_in const* src;
+      if (s == kNumHeadsQPadded) {
+        src = kv_in + static_cast<int64_t>(tokenIdx) * kHeadDim + dim_base;
+      } else {
+        src = q_in +
+              (static_cast<int64_t>(tokenIdx) * num_heads_q +
+               static_cast<int64_t>(s)) *
+                  kHeadDim +
+              dim_base;
+      }
+      va = *reinterpret_cast<uint4 const*>(src);
+      vb = *reinterpret_cast<uint4 const*>(src + 8);
+    };
 
-    // Per-quant-block absmax must be computed by ALL 32 lanes (warp-collective
-    // shuffle requires full participation).  RoPE lanes contribute garbage,
-    // but their values are gated out below via `!is_rope_lane`.
-    float local_absmax = 0.0f;
-#pragma unroll
-    for (int i = 0; i < kElemsPerLane; i++) {
-      local_absmax = fmaxf(local_absmax, fabsf(elements[i]));
-    }
-    float const absmax = fmaxf(warp4MaxAbs(local_absmax), 1e-4f);
-    float const exponent = ceilf(log2f(absmax / kFp8Max));
-    float const inv_scale = exp2f(-exponent);
+    if (warpId < slot_end) {
+      int curr_slot = warpId;
+      uint4 v0_curr, v1_curr;
+      load_slot(curr_slot, v0_curr, v1_curr);
 
-    if (!is_rope_lane) {
-      // ── NoPE lane: UE8M0 FP8 quant ───────────────────────────────────────
-      uint8_t out_bytes[kElemsPerLane];
-#pragma unroll
-      for (int i = 0; i < kElemsPerLane; i++) {
-        float scaled = elements[i] * inv_scale;
-        scaled = fminf(fmaxf(scaled, -kFp8Max), kFp8Max);
-#ifndef USE_ROCM
-        __nv_fp8_storage_t s =
-            __nv_cvt_float_to_fp8(scaled, __NV_SATFINITE, __NV_E4M3);
-        out_bytes[i] = static_cast<uint8_t>(s);
-#else
-      out_bytes[i] = rocm_cvt_float_to_fp8_e4m3(scaled);
-#endif
-      }
-      // One 16-byte STG per lane.
-      *reinterpret_cast<uint4*>(token_fp8_ptr + dim_base) =
-          *reinterpret_cast<uint4 const*>(out_bytes);
+      while (curr_slot < slot_end) {
+        int const next_slot = curr_slot + warpsPerBlock;
+        bool const has_next = (next_slot < slot_end);
+
+        // Prefetch src for the next slot
+        uint4 v0_next, v1_next;
+        if (has_next) {
+          load_slot(next_slot, v0_next, v1_next);
+        }
+
+        processDeepseekV4Slot<scalar_t_in, kNumHeadsQPadded>(
+            v0_curr, v1_curr, tokenIdx, curr_slot, dim_base, laneId,
+            num_heads_q, eps, q_out, k_cache, slot_mapping, position_ids,
+            cos_sin_cache, cache_block_size, kv_block_stride);
+
+        // ── Buffer rotation: hand the prefetched LDGs to the next iter.
+        v0_curr = v0_next;
+        v1_curr = v1_next;
+        curr_slot = next_slot;
+      }  // while
+    }  // if (warpId < slot_end)
 
-      // Lane (4k) of each 4-lane group writes the scale byte for block k<7.
-      if ((laneId & 3) == 0) {
-        int const q_block_idx = laneId >> 2;  // 0..6 for NoPE lanes
-        float encoded = fmaxf(fminf(exponent + 127.0f, 255.0f), 0.0f);
-        token_scale_ptr[q_block_idx] = static_cast<uint8_t>(encoded);
-      }
-      // Lane 0 also writes the padding byte at index 7.
-      if (laneId == 0) {
-        token_scale_ptr[kNumQuantBlocks] = 0;  // pad
-      }
-    } else {
-      // ── RoPE lane: cast back to bf16 and store to cache bf16 tail ────────
-      uint4 out0, out1;
-      typename Converter::packed_hip_type* po0 =
-          reinterpret_cast<typename Converter::packed_hip_type*>(&out0);
-      typename Converter::packed_hip_type* po1 =
-          reinterpret_cast<typename Converter::packed_hip_type*>(&out1);
-#pragma unroll
-      for (int i = 0; i < 4; i++) {
-        po0[i] = Converter::convert(
-            make_float2(elements[2 * i], elements[2 * i + 1]));
-      }
-#pragma unroll
-      for (int i = 0; i < 4; i++) {
-        po1[i] = Converter::convert(
-            make_float2(elements[8 + 2 * i], elements[8 + 2 * i + 1]));
-      }
-      int const rope_local_base = dim_base - kNopeDim;  // in [0, 64)
-      scalar_t_in* bf16_dst =
-          reinterpret_cast<scalar_t_in*>(token_bf16_ptr) + rope_local_base;
-      *reinterpret_cast<uint4*>(bf16_dst) = out0;
-      *reinterpret_cast<uint4*>(bf16_dst + 8) = out1;
-    }
 #if defined(__CUDA_ARCH__) && (__CUDA_ARCH__ >= 900)
     cudaTriggerProgrammaticLaunchCompletion();
 #endif
@@ -390,10 +540,10 @@ __global__ void fusedDeepseekV4QNormRopeKVRopeQuantInsertKernel(
 // ────────────────────────────────────────────────────────────────────────────
 // Launch wrapper
 // ────────────────────────────────────────────────────────────────────────────
-template <typename scalar_t_in>
-void launchFusedDeepseekV4QNormRopeKVRopeQuantInsert(
-    scalar_t_in* q_inout, scalar_t_in const* kv_in, uint8_t* k_cache,
-    int64_t const* slot_mapping, int64_t const* position_ids,
+template <typename scalar_t_in, int kNumHeadsQPadded>
+static void launchFusedDeepseekV4Templated(
+    scalar_t_in const* q_in, scalar_t_in* q_out, scalar_t_in const* kv_in,
+    uint8_t* k_cache, int64_t const* slot_mapping, int64_t const* position_ids,
     float const* cos_sin_cache, float const eps, int const num_tokens_full,
     int const num_tokens_insert, int const num_heads_q,
     int const cache_block_size, int const kv_block_stride,
@@ -401,7 +551,7 @@ void launchFusedDeepseekV4QNormRopeKVRopeQuantInsert(
   constexpr int kBlockSize = 256;
   constexpr int kWarpsPerBlock = kBlockSize / 32;
   int64_t const total_warps =
-      static_cast<int64_t>(num_tokens_full) * (num_heads_q + 1);
+      static_cast<int64_t>(num_tokens_full) * (kNumHeadsQPadded + 1);
   int const grid =
       static_cast<int>((total_warps + kWarpsPerBlock - 1) / kWarpsPerBlock);
 
@@ -430,37 +580,87 @@ void launchFusedDeepseekV4QNormRopeKVRopeQuantInsert(
   config.attrs = attrs;
   config.numAttrs = (sm_version >= 90) ? 1 : 0;
 
-  cudaLaunchKernelEx(
-      &config, fusedDeepseekV4QNormRopeKVRopeQuantInsertKernel<scalar_t_in>,
-      q_inout, kv_in, k_cache, slot_mapping, position_ids, cos_sin_cache, eps,
-      num_tokens_full, num_tokens_insert, num_heads_q, cache_block_size,
-      kv_block_stride);
+  if (num_tokens_full < NUM_TOKEN_CUTOFF) {
+    cudaLaunchKernelEx(
+        &config,
+        fusedDeepseekV4QNormRopeKVRopeQuantInsertKernel<scalar_t_in,
+                                                        kNumHeadsQPadded>,
+        q_in, q_out, kv_in, k_cache, slot_mapping, position_ids, cos_sin_cache,
+        eps, num_tokens_full, num_tokens_insert, num_heads_q, cache_block_size,
+        kv_block_stride);
+  } else {
+    config.gridDim = dim3(num_tokens_full);
+    cudaLaunchKernelEx(
+        &config,
+        fusedDeepseekV4QNormRopeKVRopeQuantInsertKernelReducedGrid<
+            scalar_t_in, kNumHeadsQPadded>,
+        q_in, q_out, kv_in, k_cache, slot_mapping, position_ids, cos_sin_cache,
+        eps, num_tokens_full, num_tokens_insert, num_heads_q, cache_block_size,
+        kv_block_stride);
+  }
 #else
   // ROCm: use standard kernel launch syntax (no PDL/stream serialization)
   // clang-format off
-  fusedDeepseekV4QNormRopeKVRopeQuantInsertKernel<scalar_t_in>
+  fusedDeepseekV4QNormRopeKVRopeQuantInsertKernel<scalar_t_in, kNumHeadsQPadded>
       <<<grid, kBlockSize, 0, stream>>>(
-          q_inout, kv_in, k_cache, slot_mapping, position_ids, cos_sin_cache,
-          eps, num_tokens_full, num_tokens_insert, num_heads_q,
+          q_in, q_out, kv_in, k_cache, slot_mapping, position_ids,
+          cos_sin_cache, eps, num_tokens_full, num_tokens_insert, num_heads_q,
           cache_block_size, kv_block_stride);
 #endif
 }
 
+// Runtime dispatch into one of the precompiled `kNumHeadsQPadded`
+// instantiations.  Supported padded head counts: 8, 16, 32, 64, 128.
+template <typename scalar_t_in>
+void launchFusedDeepseekV4QNormRopeKVRopeQuantInsert(
+    scalar_t_in const* q_in, scalar_t_in* q_out, scalar_t_in const* kv_in,
+    uint8_t* k_cache, int64_t const* slot_mapping,
+    int64_t const* position_ids, float const* cos_sin_cache, float const eps,
+    int const num_tokens_full, int const num_tokens_insert,
+    int const num_heads_q, int const num_heads_q_padded,
+    int const cache_block_size, int const kv_block_stride,
+    cudaStream_t stream) {
+#define DISPATCH(N)                                                         \
+  case N:                                                                   \
+    launchFusedDeepseekV4Templated<scalar_t_in, N>(                         \
+        q_in, q_out, kv_in, k_cache, slot_mapping, position_ids,            \
+        cos_sin_cache, eps, num_tokens_full, num_tokens_insert, num_heads_q, \
+        cache_block_size, kv_block_stride, stream);                         \
+    return;
+
+  switch (num_heads_q_padded) {
+    DISPATCH(8)
+    DISPATCH(16)
+    DISPATCH(32)
+    DISPATCH(64)
+    DISPATCH(128)
+    default:
+      TORCH_CHECK(false,
+                  "fused_deepseek_v4_qnorm_rope_kv_rope_quant_insert: "
+                  "unsupported num_heads_q_padded=",
+                  num_heads_q_padded,
+                  " (compiled instantiations: 8, 16, 32, 64, 128).");
+  }
+#undef DISPATCH
+}
+
 }  // namespace deepseek_v4_fused_ops
 }  // namespace vllm
 
 // ────────────────────────────────────────────────────────────────────────────
 // Torch op wrapper
 // ────────────────────────────────────────────────────────────────────────────
-void fused_deepseek_v4_qnorm_rope_kv_rope_quant_insert(
-    torch::Tensor& q,                    // [N, H, 512] bf16, in place
+torch::Tensor fused_deepseek_v4_qnorm_rope_kv_rope_quant_insert(
+    torch::Tensor const& q_in,           // [N, num_heads_q, 512] bf16
     torch::Tensor const& kv,             // [N, 512] bf16 (read-only)
     torch::Tensor& k_cache,              // [num_blocks, block_bytes] uint8
     torch::Tensor const& slot_mapping,   // [N] int64
     torch::Tensor const& position_ids,   // [N] int64
     torch::Tensor const& cos_sin_cache,  // [max_pos, rope_dim] bf16
+    int64_t q_head_padded,               // padded Q head count for output
     double eps, int64_t cache_block_size) {
-  TORCH_CHECK(q.is_cuda() && q.is_contiguous(), "q must be contiguous CUDA");
+  TORCH_CHECK(q_in.is_cuda() && q_in.is_contiguous(),
+              "q_in must be contiguous CUDA");
   TORCH_CHECK(kv.is_cuda() && kv.is_contiguous(), "kv must be contiguous CUDA");
   TORCH_CHECK(k_cache.is_cuda(), "k_cache must be CUDA");
   TORCH_CHECK(slot_mapping.is_cuda() && slot_mapping.dtype() == torch::kInt64,
@@ -468,9 +668,12 @@ void fused_deepseek_v4_qnorm_rope_kv_rope_quant_insert(
   TORCH_CHECK(position_ids.is_cuda() && position_ids.dtype() == torch::kInt64,
               "position_ids must be int64 CUDA");
   TORCH_CHECK(cos_sin_cache.is_cuda(), "cos_sin_cache must be CUDA");
-  TORCH_CHECK(q.dim() == 3 && q.size(2) == 512, "q shape [N, H, 512]");
+  TORCH_CHECK(q_in.dim() == 3 && q_in.size(2) == 512,
+              "q_in shape [N, num_heads_q, 512]");
   TORCH_CHECK(kv.dim() == 2 && kv.size(1) == 512, "kv shape [N, 512]");
-  TORCH_CHECK(q.dtype() == kv.dtype(), "q and kv dtype must match");
+  TORCH_CHECK(q_in.dtype() == kv.dtype(), "q_in and kv dtype must match");
+  TORCH_CHECK(q_head_padded >= q_in.size(1),
+              "q_head_padded must be >= q_in.size(1) (num_heads_q)");
   TORCH_CHECK(k_cache.dtype() == torch::kUInt8, "k_cache must be uint8");
   TORCH_CHECK(cos_sin_cache.dim() == 2 && cos_sin_cache.size(1) == 64,
               "cos_sin_cache shape [max_pos, 64]");
@@ -480,32 +683,41 @@ void fused_deepseek_v4_qnorm_rope_kv_rope_quant_insert(
   // With DP padding, slot_mapping can be shorter than q/kv/positions.
   // Q-norm+RoPE runs on all q.size(0) rows (downstream attention uses them);
   // KV quant+insert runs only on the first slot_mapping.size(0) rows.
-  int const num_tokens_full = static_cast<int>(q.size(0));
+  int const num_tokens_full = static_cast<int>(q_in.size(0));
   int const num_tokens_insert = static_cast<int>(slot_mapping.size(0));
   TORCH_CHECK(static_cast<int>(kv.size(0)) == num_tokens_full &&
                   static_cast<int>(position_ids.size(0)) == num_tokens_full,
               "q/kv/position_ids row counts must match");
   TORCH_CHECK(num_tokens_insert <= num_tokens_full,
               "slot_mapping must not exceed q row count");
-  int const num_heads_q = static_cast<int>(q.size(1));
+  int const num_heads_q = static_cast<int>(q_in.size(1));
+  int const num_heads_q_padded = static_cast<int>(q_head_padded);
   int const cache_block_size_i = static_cast<int>(cache_block_size);
   int const kv_block_stride = static_cast<int>(k_cache.stride(0));
 
-  at::cuda::OptionalCUDAGuard device_guard(device_of(q));
+  at::cuda::OptionalCUDAGuard device_guard(device_of(q_in));
   auto stream = at::cuda::getCurrentCUDAStream();
 
+  // Allocate the padded q output.  The kernel writes every element (live
+  // region gets RMSNorm+RoPE; pad region gets zeros), so `empty` is safe.
+  torch::Tensor q_out = torch::empty(
+      {q_in.size(0), q_head_padded, q_in.size(2)}, q_in.options());
+
   VLLM_DISPATCH_HALF_TYPES(
-      q.scalar_type(), "fused_deepseek_v4_qnorm_rope_kv_insert", [&] {
+      q_in.scalar_type(), "fused_deepseek_v4_qnorm_rope_kv_insert", [&] {
         using qkv_scalar_t = scalar_t;
         vllm::deepseek_v4_fused_ops::
             launchFusedDeepseekV4QNormRopeKVRopeQuantInsert<qkv_scalar_t>(
-                reinterpret_cast<qkv_scalar_t*>(q.data_ptr()),
+                reinterpret_cast<qkv_scalar_t const*>(q_in.data_ptr()),
+                reinterpret_cast<qkv_scalar_t*>(q_out.data_ptr()),
                 reinterpret_cast<qkv_scalar_t const*>(kv.data_ptr()),
                 reinterpret_cast<uint8_t*>(k_cache.data_ptr()),
                 reinterpret_cast<int64_t const*>(slot_mapping.data_ptr()),
                 reinterpret_cast<int64_t const*>(position_ids.data_ptr()),
                 cos_sin_cache.data_ptr<float>(), static_cast<float>(eps),
                 num_tokens_full, num_tokens_insert, num_heads_q,
-                cache_block_size_i, kv_block_stride, stream);
+                num_heads_q_padded, cache_block_size_i, kv_block_stride,
+                stream);
       });
+  return q_out;
 }

csrc/libtorch_stable/activation_kernels.cu

@@ -1,12 +1,12 @@
-#include <ATen/cuda/CUDAContext.h>
-#include <torch/all.h>
-#include <c10/cuda/CUDAGuard.h>
+#include <cuda.h>
+#include <torch/csrc/stable/tensor.h>
 
 #include <cmath>
 
-#include "cuda_compat.h"
+#include "../cuda_compat.h"
 #include "cuda_vec_utils.cuh"
 #include "dispatch_utils.h"
+#include "torch_utils.h"
 
 namespace vllm {
 
@@ -210,64 +210,68 @@ packed_gelu_tanh_kernel(const packed_t& val) {
     return;                                                                    \
   }                                                                            \
   dim3 grid(num_tokens);                                                       \
-  int cc_major = at::cuda::getCurrentDeviceProperties()->major;                \
+  int cc_major = get_device_prop()->major;                                     \
   int support_vec =                                                            \
       (CUDA_VERSION >= 12090 && cc_major >= 10 && num_tokens > 128)            \
           ? vllm::VecTraits<true>::ARCH_MAX_VEC_SIZE                           \
           : vllm::VecTraits<false>::ARCH_MAX_VEC_SIZE;                         \
-  int vec_size = support_vec / at::elementSize(dtype);                         \
+  int vec_size = support_vec / input.element_size();                           \
   const bool use_vec = (d % vec_size == 0);                                    \
-  const at::cuda::OptionalCUDAGuard device_guard(device_of(input));            \
-  const cudaStream_t stream = at::cuda::getCurrentCUDAStream();                \
+  const torch::stable::accelerator::DeviceGuard device_guard(                  \
+      input.get_device_index());                                               \
+  const cudaStream_t stream = get_current_cuda_stream();                       \
   if (use_vec) {                                                               \
     dim3 block(std::min(d / vec_size, 1024));                                  \
     if (CUDA_VERSION >= 12090 && cc_major >= 10 && num_tokens > 128) {         \
-      VLLM_DISPATCH_FLOATING_TYPES(dtype, "act_and_mul_kernel", [&] {          \
+      VLLM_STABLE_DISPATCH_FLOATING_TYPES(dtype, "act_and_mul_kernel", [&] {   \
         vllm::act_and_mul_kernel<                                              \
             scalar_t, typename vllm::PackedTypeConverter<scalar_t>::Type,      \
             KERNEL<scalar_t>,                                                  \
             PACKED_KERNEL<typename vllm::PackedTypeConverter<scalar_t>::Type>, \
             ACT_FIRST, true, HAS_CLAMP, true><<<grid, block, 0, stream>>>(     \
-            out.data_ptr<scalar_t>(), input.data_ptr<scalar_t>(), d, LIMIT);   \
+            out.mutable_data_ptr<scalar_t>(),                                  \
+            input.const_data_ptr<scalar_t>(), d, LIMIT);                       \
       });                                                                      \
     } else {                                                                   \
-      VLLM_DISPATCH_FLOATING_TYPES(dtype, "act_and_mul_kernel", [&] {          \
+      VLLM_STABLE_DISPATCH_FLOATING_TYPES(dtype, "act_and_mul_kernel", [&] {   \
         vllm::act_and_mul_kernel<                                              \
             scalar_t, typename vllm::PackedTypeConverter<scalar_t>::Type,      \
             KERNEL<scalar_t>,                                                  \
             PACKED_KERNEL<typename vllm::PackedTypeConverter<scalar_t>::Type>, \
             ACT_FIRST, true, HAS_CLAMP, false><<<grid, block, 0, stream>>>(    \
-            out.data_ptr<scalar_t>(), input.data_ptr<scalar_t>(), d, LIMIT);   \
+            out.mutable_data_ptr<scalar_t>(),                                  \
+            input.const_data_ptr<scalar_t>(), d, LIMIT);                       \
       });                                                                      \
     }                                                                          \
   } else {                                                                     \
     dim3 block(std::min(d, 1024));                                             \
-    VLLM_DISPATCH_FLOATING_TYPES(dtype, "act_and_mul_kernel", [&] {            \
+    VLLM_STABLE_DISPATCH_FLOATING_TYPES(dtype, "act_and_mul_kernel", [&] {     \
       vllm::act_and_mul_kernel<                                                \
           scalar_t, typename vllm::PackedTypeConverter<scalar_t>::Type,        \
           KERNEL<scalar_t>,                                                    \
           PACKED_KERNEL<typename vllm::PackedTypeConverter<scalar_t>::Type>,   \
           ACT_FIRST, false, HAS_CLAMP><<<grid, block, 0, stream>>>(            \
-          out.data_ptr<scalar_t>(), input.data_ptr<scalar_t>(), d, LIMIT);     \
+          out.mutable_data_ptr<scalar_t>(), input.const_data_ptr<scalar_t>(),  \
+          d, LIMIT);                                                           \
     });                                                                        \
   }
 
-void silu_and_mul(torch::Tensor& out,    // [..., d]
-                  torch::Tensor& input)  // [..., 2 * d]
+void silu_and_mul(torch::stable::Tensor& out,    // [..., d]
+                  torch::stable::Tensor& input)  // [..., 2 * d]
 {
   LAUNCH_ACTIVATION_GATE_KERNEL(vllm::silu_kernel, vllm::packed_silu_kernel,
                                 true, false, 0.0f);
 }
 
-void silu_and_mul_clamp(torch::Tensor& out,    // [..., d]
-                        torch::Tensor& input,  // [..., 2 * d]
+void silu_and_mul_clamp(torch::stable::Tensor& out,    // [..., d]
+                        torch::stable::Tensor& input,  // [..., 2 * d]
                         double limit) {
   LAUNCH_ACTIVATION_GATE_KERNEL(vllm::silu_kernel, vllm::packed_silu_kernel,
                                 true, true, (float)limit);
 }
 
-void mul_and_silu(torch::Tensor& out,    // [..., d]
-                  torch::Tensor& input)  // [..., 2 * d]
+void mul_and_silu(torch::stable::Tensor& out,    // [..., d]
+                  torch::stable::Tensor& input)  // [..., 2 * d]
 {
   // The difference between mul_and_silu and silu_and_mul is that mul_and_silu
   // applies the silu to the latter half of the input.
@@ -275,15 +279,15 @@ void mul_and_silu(torch::Tensor& out,    // [..., d]
                                 false, false, 0.0f);
 }
 
-void gelu_and_mul(torch::Tensor& out,    // [..., d]
-                  torch::Tensor& input)  // [..., 2 * d]
+void gelu_and_mul(torch::stable::Tensor& out,    // [..., d]
+                  torch::stable::Tensor& input)  // [..., 2 * d]
 {
   LAUNCH_ACTIVATION_GATE_KERNEL(vllm::gelu_kernel, vllm::packed_gelu_kernel,
                                 true, false, 0.0f);
 }
 
-void gelu_tanh_and_mul(torch::Tensor& out,    // [..., d]
-                       torch::Tensor& input)  // [..., 2 * d]
+void gelu_tanh_and_mul(torch::stable::Tensor& out,    // [..., d]
+                       torch::stable::Tensor& input)  // [..., 2 * d]
 {
   LAUNCH_ACTIVATION_GATE_KERNEL(
       vllm::gelu_tanh_kernel, vllm::packed_gelu_tanh_kernel, true, false, 0.0f);
@@ -434,19 +438,20 @@ __global__ void swigluoai_and_mul_kernel(
     return;                                                                    \
   }                                                                            \
   dim3 grid(num_tokens);                                                       \
-  int cc_major = at::cuda::getCurrentDeviceProperties()->major;                \
+  int cc_major = get_device_prop()->major;                                     \
   int support_vec =                                                            \
       (CUDA_VERSION >= 12090 && cc_major >= 10 && num_tokens > 128)            \
           ? vllm::VecTraits<true>::ARCH_MAX_VEC_SIZE                           \
           : vllm::VecTraits<false>::ARCH_MAX_VEC_SIZE;                         \
-  int vec_size = support_vec / at::elementSize(dtype);                         \
+  int vec_size = support_vec / input.element_size();                           \
   const bool use_vec = (d % vec_size == 0);                                    \
-  const at::cuda::OptionalCUDAGuard device_guard(device_of(input));            \
-  const cudaStream_t stream = at::cuda::getCurrentCUDAStream();                \
+  const torch::stable::accelerator::DeviceGuard device_guard(                  \
+      input.get_device_index());                                               \
+  const cudaStream_t stream = get_current_cuda_stream();                       \
   if (use_vec) {                                                               \
     dim3 block(std::min(d / vec_size, 1024));                                  \
     if (CUDA_VERSION >= 12090 && cc_major >= 10 && num_tokens > 128) {         \
-      VLLM_DISPATCH_FLOATING_TYPES(                                            \
+      VLLM_STABLE_DISPATCH_FLOATING_TYPES(                                     \
           dtype, "act_and_mul_kernel_with_param", [&] {                        \
             vllm::act_and_mul_kernel_with_param<                               \
                 scalar_t, typename vllm::PackedTypeConverter<scalar_t>::Type,  \
@@ -454,11 +459,11 @@ __global__ void swigluoai_and_mul_kernel(
                 PACKED_KERNEL<                                                 \
                     typename vllm::PackedTypeConverter<scalar_t>::Type>,       \
                 true, true><<<grid, block, 0, stream>>>(                       \
-                out.data_ptr<scalar_t>(), input.data_ptr<scalar_t>(), d,       \
-                PARAM);                                                        \
+                out.mutable_data_ptr<scalar_t>(),                              \
+                input.const_data_ptr<scalar_t>(), d, PARAM);                   \
           });                                                                  \
     } else {                                                                   \
-      VLLM_DISPATCH_FLOATING_TYPES(                                            \
+      VLLM_STABLE_DISPATCH_FLOATING_TYPES(                                     \
           dtype, "act_and_mul_kernel_with_param", [&] {                        \
             vllm::act_and_mul_kernel_with_param<                               \
                 scalar_t, typename vllm::PackedTypeConverter<scalar_t>::Type,  \
@@ -466,45 +471,49 @@ __global__ void swigluoai_and_mul_kernel(
                 PACKED_KERNEL<                                                 \
                     typename vllm::PackedTypeConverter<scalar_t>::Type>,       \
                 true, false><<<grid, block, 0, stream>>>(                      \
-                out.data_ptr<scalar_t>(), input.data_ptr<scalar_t>(), d,       \
-                PARAM);                                                        \
+                out.mutable_data_ptr<scalar_t>(),                              \
+                input.const_data_ptr<scalar_t>(), d, PARAM);                   \
           });                                                                  \
     }                                                                          \
   } else {                                                                     \
     dim3 block(std::min(d, 1024));                                             \
-    VLLM_DISPATCH_FLOATING_TYPES(dtype, "act_and_mul_kernel_with_param", [&] { \
-      vllm::act_and_mul_kernel_with_param<                                     \
-          scalar_t, typename vllm::PackedTypeConverter<scalar_t>::Type,        \
-          KERNEL<scalar_t>,                                                    \
-          PACKED_KERNEL<typename vllm::PackedTypeConverter<scalar_t>::Type>,   \
-          false><<<grid, block, 0, stream>>>(                                  \
-          out.data_ptr<scalar_t>(), input.data_ptr<scalar_t>(), d, PARAM);     \
-    });                                                                        \
+    VLLM_STABLE_DISPATCH_FLOATING_TYPES(                                       \
+        dtype, "act_and_mul_kernel_with_param", [&] {                          \
+          vllm::act_and_mul_kernel_with_param<                                 \
+              scalar_t, typename vllm::PackedTypeConverter<scalar_t>::Type,    \
+              KERNEL<scalar_t>,                                                \
+              PACKED_KERNEL<                                                   \
+                  typename vllm::PackedTypeConverter<scalar_t>::Type>,         \
+              false><<<grid, block, 0, stream>>>(                              \
+              out.mutable_data_ptr<scalar_t>(),                                \
+              input.const_data_ptr<scalar_t>(), d, PARAM);                     \
+        });                                                                    \
   }
 
-#define LAUNCH_SIGLUOAI_AND_MUL(KERNEL, ALPHA, LIMIT)                          \
-  int d = input.size(-1) / 2;                                                  \
-  int64_t num_tokens = input.numel() / input.size(-1);                         \
-  dim3 grid(num_tokens);                                                       \
-  dim3 block(std::min(d, 1024));                                               \
-  const at::cuda::OptionalCUDAGuard device_guard(device_of(input));            \
-  const cudaStream_t stream = at::cuda::getCurrentCUDAStream();                \
-  VLLM_DISPATCH_FLOATING_TYPES(                                                \
-      input.scalar_type(), "clamp_swiglu_kernel_with_params", [&] {            \
-        vllm::swigluoai_and_mul_kernel<scalar_t, KERNEL<scalar_t>>             \
-            <<<grid, block, 0, stream>>>(out.data_ptr<scalar_t>(),             \
-                                         input.data_ptr<scalar_t>(), d, ALPHA, \
-                                         LIMIT);                               \
+#define LAUNCH_SIGLUOAI_AND_MUL(KERNEL, ALPHA, LIMIT)                         \
+  int d = input.size(-1) / 2;                                                 \
+  int64_t num_tokens = input.numel() / input.size(-1);                        \
+  dim3 grid(num_tokens);                                                      \
+  dim3 block(std::min(d, 1024));                                              \
+  const torch::stable::accelerator::DeviceGuard device_guard(                 \
+      input.get_device_index());                                              \
+  const cudaStream_t stream = get_current_cuda_stream();                      \
+  VLLM_STABLE_DISPATCH_FLOATING_TYPES(                                        \
+      input.scalar_type(), "clamp_swiglu_kernel_with_params", [&] {           \
+        vllm::swigluoai_and_mul_kernel<scalar_t, KERNEL<scalar_t>>            \
+            <<<grid, block, 0, stream>>>(out.mutable_data_ptr<scalar_t>(),    \
+                                         input.const_data_ptr<scalar_t>(), d, \
+                                         ALPHA, LIMIT);                       \
       });
 
-void fatrelu_and_mul(torch::Tensor& out,    // [..., d],
-                     torch::Tensor& input,  // [..., 2 * d]
+void fatrelu_and_mul(torch::stable::Tensor& out,    // [..., d],
+                     torch::stable::Tensor& input,  // [..., 2 * d]
                      double threshold) {
   LAUNCH_ACTIVATION_GATE_KERNEL_WITH_PARAM(
       vllm::fatrelu_kernel, vllm::packed_fatrelu_kernel, threshold);
 }
-void swigluoai_and_mul(torch::Tensor& out,    // [..., d]
-                       torch::Tensor& input,  // [..., 2 * d]
+void swigluoai_and_mul(torch::stable::Tensor& out,    // [..., d]
+                       torch::stable::Tensor& input,  // [..., 2 * d]
                        double alpha, double limit) {
   LAUNCH_SIGLUOAI_AND_MUL(vllm::swigluoai_and_mul, alpha, limit);
 }
@@ -559,45 +568,46 @@ __global__ void activation_kernel(
 }  // namespace vllm
 
 // Launch element-wise activation kernel.
-#define LAUNCH_ACTIVATION_KERNEL(KERNEL)                                 \
-  auto dtype = input.scalar_type();                                      \
-  int d = input.size(-1);                                                \
-  int64_t num_tokens = input.numel() / input.size(-1);                   \
-  if (num_tokens == 0) {                                                 \
-    return;                                                              \
-  }                                                                      \
-  dim3 grid(num_tokens);                                                 \
-  int cc_major = at::cuda::getCurrentDeviceProperties()->major;          \
-  int support_vec =                                                      \
-      (CUDA_VERSION >= 12090 && cc_major >= 10 && num_tokens > 128)      \
-          ? vllm::VecTraits<true>::ARCH_MAX_VEC_SIZE                     \
-          : vllm::VecTraits<false>::ARCH_MAX_VEC_SIZE;                   \
-  int vec_size = support_vec / at::elementSize(dtype);                   \
-  const bool use_vec = (d % vec_size == 0);                              \
-  const at::cuda::OptionalCUDAGuard device_guard(device_of(input));      \
-  const cudaStream_t stream = at::cuda::getCurrentCUDAStream();          \
-  if (use_vec) {                                                         \
-    dim3 block(std::min(d / vec_size, 1024));                            \
-    if (CUDA_VERSION >= 12090 && cc_major >= 10 && num_tokens > 128) {   \
-      VLLM_DISPATCH_FLOATING_TYPES(dtype, "activation_kernel", [&] {     \
-        vllm::activation_kernel<scalar_t, KERNEL<scalar_t>, true, true>  \
-            <<<grid, block, 0, stream>>>(out.data_ptr<scalar_t>(),       \
-                                         input.data_ptr<scalar_t>(), d); \
-      });                                                                \
-    } else {                                                             \
-      VLLM_DISPATCH_FLOATING_TYPES(dtype, "activation_kernel", [&] {     \
-        vllm::activation_kernel<scalar_t, KERNEL<scalar_t>, true, false> \
-            <<<grid, block, 0, stream>>>(out.data_ptr<scalar_t>(),       \
-                                         input.data_ptr<scalar_t>(), d); \
-      });                                                                \
-    }                                                                    \
-  } else {                                                               \
-    dim3 block(std::min(d, 1024));                                       \
-    VLLM_DISPATCH_FLOATING_TYPES(dtype, "activation_kernel", [&] {       \
-      vllm::activation_kernel<scalar_t, KERNEL<scalar_t>, false>         \
-          <<<grid, block, 0, stream>>>(out.data_ptr<scalar_t>(),         \
-                                       input.data_ptr<scalar_t>(), d);   \
-    });                                                                  \
+#define LAUNCH_ACTIVATION_KERNEL(KERNEL)                                       \
+  auto dtype = input.scalar_type();                                            \
+  int d = input.size(-1);                                                      \
+  int64_t num_tokens = input.numel() / input.size(-1);                         \
+  if (num_tokens == 0) {                                                       \
+    return;                                                                    \
+  }                                                                            \
+  dim3 grid(num_tokens);                                                       \
+  int cc_major = get_device_prop()->major;                                     \
+  int support_vec =                                                            \
+      (CUDA_VERSION >= 12090 && cc_major >= 10 && num_tokens > 128)            \
+          ? vllm::VecTraits<true>::ARCH_MAX_VEC_SIZE                           \
+          : vllm::VecTraits<false>::ARCH_MAX_VEC_SIZE;                         \
+  int vec_size = support_vec / input.element_size();                           \
+  const bool use_vec = (d % vec_size == 0);                                    \
+  const torch::stable::accelerator::DeviceGuard device_guard(                  \
+      input.get_device_index());                                               \
+  const cudaStream_t stream = get_current_cuda_stream();                       \
+  if (use_vec) {                                                               \
+    dim3 block(std::min(d / vec_size, 1024));                                  \
+    if (CUDA_VERSION >= 12090 && cc_major >= 10 && num_tokens > 128) {         \
+      VLLM_STABLE_DISPATCH_FLOATING_TYPES(dtype, "activation_kernel", [&] {    \
+        vllm::activation_kernel<scalar_t, KERNEL<scalar_t>, true, true>        \
+            <<<grid, block, 0, stream>>>(out.mutable_data_ptr<scalar_t>(),     \
+                                         input.const_data_ptr<scalar_t>(), d); \
+      });                                                                      \
+    } else {                                                                   \
+      VLLM_STABLE_DISPATCH_FLOATING_TYPES(dtype, "activation_kernel", [&] {    \
+        vllm::activation_kernel<scalar_t, KERNEL<scalar_t>, true, false>       \
+            <<<grid, block, 0, stream>>>(out.mutable_data_ptr<scalar_t>(),     \
+                                         input.const_data_ptr<scalar_t>(), d); \
+      });                                                                      \
+    }                                                                          \
+  } else {                                                                     \
+    dim3 block(std::min(d, 1024));                                             \
+    VLLM_STABLE_DISPATCH_FLOATING_TYPES(dtype, "activation_kernel", [&] {      \
+      vllm::activation_kernel<scalar_t, KERNEL<scalar_t>, false>               \
+          <<<grid, block, 0, stream>>>(out.mutable_data_ptr<scalar_t>(),       \
+                                       input.const_data_ptr<scalar_t>(), d);   \
+    });                                                                        \
   }
 
 namespace vllm {
@@ -625,20 +635,20 @@ __device__ __forceinline__ T gelu_quick_kernel(const T& x) {
 
 }  // namespace vllm
 
-void gelu_new(torch::Tensor& out,    // [..., d]
-              torch::Tensor& input)  // [..., d]
+void gelu_new(torch::stable::Tensor& out,    // [..., d]
+              torch::stable::Tensor& input)  // [..., d]
 {
   LAUNCH_ACTIVATION_KERNEL(vllm::gelu_new_kernel);
 }
 
-void gelu_fast(torch::Tensor& out,    // [..., d]
-               torch::Tensor& input)  // [..., d]
+void gelu_fast(torch::stable::Tensor& out,    // [..., d]
+               torch::stable::Tensor& input)  // [..., d]
 {
   LAUNCH_ACTIVATION_KERNEL(vllm::gelu_fast_kernel);
 }
 
-void gelu_quick(torch::Tensor& out,    // [..., d]
-                torch::Tensor& input)  // [..., d]
+void gelu_quick(torch::stable::Tensor& out,    // [..., d]
+                torch::stable::Tensor& input)  // [..., d]
 {
   LAUNCH_ACTIVATION_KERNEL(vllm::gelu_quick_kernel);
 }

csrc/libtorch_stable/attention/attention_kernels.cuh

@@ -17,21 +17,18 @@
  * limitations under the License.
  */
 
-#include <torch/all.h>
-#include <ATen/cuda/CUDAContext.h>
-#include <c10/cuda/CUDAGuard.h>
 #include <algorithm>
 
-#include "attention_dtypes.h"
+#include "../../attention/attention_dtypes.h"
 #include "attention_utils.cuh"
-#include "../cuda_compat.h"
+#include "../../cuda_compat.h"
 
 #ifdef USE_ROCM
   #include <hip/hip_bf16.h>
-  #include "../quantization/w8a8/fp8/amd/quant_utils.cuh"
+  #include "../../quantization/w8a8/fp8/amd/quant_utils.cuh"
 typedef __hip_bfloat16 __nv_bfloat16;
 #else
-  #include "../quantization/w8a8/fp8/nvidia/quant_utils.cuh"
+  #include "../../quantization/w8a8/fp8/nvidia/quant_utils.cuh"
 #endif
 
 #define MAX(a, b) ((a) > (b) ? (a) : (b))

csrc/libtorch_stable/attention/attention_utils.cuh

@@ -18,8 +18,8 @@
  */
 #pragma once
 
-#include "../cuda_compat.h"
-#include "attention_dtypes.h"
+#include "../../cuda_compat.h"
+#include "../../attention/attention_dtypes.h"
 
 #include <float.h>
 #include <type_traits>

csrc/libtorch_stable/attention/merge_attn_states.cu

@@ -1,14 +1,14 @@
 #include <optional>
-#include <torch/all.h>
-#include <ATen/cuda/CUDAContext.h>
-#include <c10/cuda/CUDAGuard.h>
 #include <algorithm>
 #include <limits>
 
-#include "attention_dtypes.h"
-#include "attention_utils.cuh"
-#include "../quantization/w8a8/fp8/common.cuh"
+#include "../torch_utils.h"
 #include "../dispatch_utils.h"
+#include <torch/headeronly/core/ScalarType.h>
+
+#include "../../attention/attention_dtypes.h"
+#include "attention_utils.cuh"
+#include "../../quantization/w8a8/fp8/common.cuh"
 
 namespace vllm {
 
@@ -196,17 +196,17 @@ __global__ void merge_attn_states_kernel(
 // The following macro is used to dispatch the conversion function based on
 // the output data type. The FN is a macro that calls a function with
 // template<typename scalar_t>.
-#define DISPATCH_BY_SCALAR_DTYPE(scalar_dtype, fn)                      \
-  {                                                                     \
-    if (scalar_dtype == at::ScalarType::Float) {                        \
-      fn(float);                                                        \
-    } else if (scalar_dtype == at::ScalarType::Half) {                  \
-      fn(uint16_t);                                                     \
-    } else if (scalar_dtype == at::ScalarType::BFloat16) {              \
-      fn(__nv_bfloat16);                                                \
-    } else {                                                            \
-      TORCH_CHECK(false, "Unsupported data type of O: ", scalar_dtype); \
-    }                                                                   \
+#define DISPATCH_BY_SCALAR_DTYPE(scalar_dtype, fn)                          \
+  {                                                                         \
+    if (scalar_dtype == torch::headeronly::ScalarType::Float) {             \
+      fn(float);                                                            \
+    } else if (scalar_dtype == torch::headeronly::ScalarType::Half) {       \
+      fn(uint16_t);                                                         \
+    } else if (scalar_dtype == torch::headeronly::ScalarType::BFloat16) {   \
+      fn(__nv_bfloat16);                                                    \
+    } else {                                                                \
+      STD_TORCH_CHECK(false, "Unsupported data type of O: ", scalar_dtype); \
+    }                                                                       \
   }
 
 #define LAUNCH_MERGE_ATTN_STATES(scalar_t, output_t, NUM_THREADS,           \
@@ -245,11 +245,14 @@ __global__ void merge_attn_states_kernel(
  */
 template <typename scalar_t>
 void merge_attn_states_launcher(
-    torch::Tensor& output, std::optional<torch::Tensor> output_lse,
-    const torch::Tensor& prefix_output, const torch::Tensor& prefix_lse,
-    const torch::Tensor& suffix_output, const torch::Tensor& suffix_lse,
+    torch::stable::Tensor& output,
+    std::optional<torch::stable::Tensor> output_lse,
+    const torch::stable::Tensor& prefix_output,
+    const torch::stable::Tensor& prefix_lse,
+    const torch::stable::Tensor& suffix_output,
+    const torch::stable::Tensor& suffix_lse,
     const std::optional<int64_t> prefill_tokens_with_context,
-    const std::optional<torch::Tensor>& output_scale) {
+    const std::optional<torch::stable::Tensor>& output_scale) {
   constexpr uint NUM_THREADS = 128;
   const uint num_tokens = output.size(0);
   const uint num_heads = output.size(1);
@@ -258,23 +261,23 @@ void merge_attn_states_launcher(
   const uint output_head_stride = output.stride(1);
   // Thread mapping is based on input BF16 pack_size
   const uint pack_size = 16 / sizeof(scalar_t);
-  TORCH_CHECK(head_size % pack_size == 0,
-              "headsize must be multiple of pack_size:", pack_size);
+  STD_TORCH_CHECK(head_size % pack_size == 0,
+                  "headsize must be multiple of pack_size:", pack_size);
 
   const uint prefix_num_tokens =
       prefill_tokens_with_context.has_value()
           ? static_cast<uint>(prefill_tokens_with_context.value())
           : num_tokens;
-  TORCH_CHECK(prefix_num_tokens <= num_tokens,
-              "prefix_num_tokens must be <= num_tokens");
+  STD_TORCH_CHECK(prefix_num_tokens <= num_tokens,
+                  "prefix_num_tokens must be <= num_tokens");
 
   float* output_lse_ptr = nullptr;
   if (output_lse.has_value()) {
-    output_lse_ptr = output_lse.value().data_ptr<float>();
+    output_lse_ptr = output_lse.value().mutable_data_ptr<float>();
   }
   float* output_scale_ptr = nullptr;
   if (output_scale.has_value()) {
-    output_scale_ptr = output_scale.value().data_ptr<float>();
+    output_scale_ptr = output_scale.value().mutable_data_ptr<float>();
   }
   // Process one pack elements per thread. for float, the
   // pack_size is 4 for half/bf16, the pack_size is 8.
@@ -284,14 +287,15 @@ void merge_attn_states_launcher(
   dim3 block(NUM_THREADS);
   dim3 grid((total_threads + NUM_THREADS - 1) / NUM_THREADS);
 
-  const c10::cuda::OptionalCUDAGuard device_guard(prefix_output.device());
-  auto stream = at::cuda::getCurrentCUDAStream();
+  const torch::stable::accelerator::DeviceGuard device_guard(
+      prefix_output.get_device_index());
+  auto stream = get_current_cuda_stream();
 
   if (output_scale.has_value()) {
     // FP8 output path - dispatch on output FP8 type
-    VLLM_DISPATCH_FP8_TYPES(output.scalar_type(), "merge_attn_states_fp8", [&] {
-      LAUNCH_MERGE_ATTN_STATES(scalar_t, fp8_t, NUM_THREADS, true);
-    });
+    VLLM_STABLE_DISPATCH_FP8_TYPES(
+        output.scalar_type(), "merge_attn_states_fp8",
+        [&] { LAUNCH_MERGE_ATTN_STATES(scalar_t, fp8_t, NUM_THREADS, true); });
   } else {
     // Original BF16/FP16/FP32 output path
     LAUNCH_MERGE_ATTN_STATES(scalar_t, scalar_t, NUM_THREADS, false);
@@ -305,26 +309,29 @@ void merge_attn_states_launcher(
         suffix_lse, prefill_tokens_with_context, output_scale);       \
   }
 
-void merge_attn_states(torch::Tensor& output,
-                       std::optional<torch::Tensor> output_lse,
-                       const torch::Tensor& prefix_output,
-                       const torch::Tensor& prefix_lse,
-                       const torch::Tensor& suffix_output,
-                       const torch::Tensor& suffix_lse,
-                       std::optional<int64_t> prefill_tokens_with_context,
-                       const std::optional<torch::Tensor>& output_scale) {
+void merge_attn_states(
+    torch::stable::Tensor& output,
+    std::optional<torch::stable::Tensor> output_lse,
+    const torch::stable::Tensor& prefix_output,
+    const torch::stable::Tensor& prefix_lse,
+    const torch::stable::Tensor& suffix_output,
+    const torch::stable::Tensor& suffix_lse,
+    const std::optional<int64_t> prefill_tokens_with_context,
+    const std::optional<torch::stable::Tensor>& output_scale) {
   if (output_scale.has_value()) {
-    TORCH_CHECK(output.scalar_type() == at::ScalarType::Float8_e4m3fn ||
-                    output.scalar_type() == at::ScalarType::Float8_e4m3fnuz,
-                "output must be FP8 when output_scale is provided, got: ",
-                output.scalar_type());
+    STD_TORCH_CHECK(
+        output.scalar_type() == torch::headeronly::ScalarType::Float8_e4m3fn ||
+            output.scalar_type() ==
+                torch::headeronly::ScalarType::Float8_e4m3fnuz,
+        "output must be FP8 when output_scale is provided, got: ",
+        output.scalar_type());
   } else {
-    TORCH_CHECK(output.scalar_type() == prefix_output.scalar_type(),
-                "output dtype (", output.scalar_type(),
-                ") must match prefix_output dtype (",
-                prefix_output.scalar_type(), ") when output_scale is not set");
+    STD_TORCH_CHECK(
+        output.scalar_type() == prefix_output.scalar_type(), "output dtype (",
+        output.scalar_type(), ") must match prefix_output dtype (",
+        prefix_output.scalar_type(), ") when output_scale is not set");
   }
   // Always dispatch on prefix_output (input) dtype
-  DISPATCH_BY_SCALAR_DTYPE(prefix_output.dtype(),
+  DISPATCH_BY_SCALAR_DTYPE(prefix_output.scalar_type(),
                            CALL_MERGE_ATTN_STATES_LAUNCHER);
 }

csrc/libtorch_stable/attention/paged_attention_v1.cu

@@ -16,8 +16,9 @@
  * See the License for the specific language governing permissions and
  * limitations under the License.
  */
+#include "../torch_utils.h"
 #include "attention_kernels.cuh"
-#include "../cuda_compat.h"
+#include "../../cuda_compat.h"
 
 #define MAX(a, b) ((a) > (b) ? (a) : (b))
 #define MIN(a, b) ((a) < (b) ? (a) : (b))
@@ -44,13 +45,15 @@ template <typename T, typename CACHE_T, int BLOCK_SIZE,
           vllm::Fp8KVCacheDataType KV_DTYPE, bool IS_BLOCK_SPARSE,
           int NUM_THREADS = 128>
 void paged_attention_v1_launcher(
-    torch::Tensor& out, torch::Tensor& query, torch::Tensor& key_cache,
-    torch::Tensor& value_cache, int num_kv_heads, float scale,
-    torch::Tensor& block_tables, torch::Tensor& seq_lens, int max_seq_len,
-    const std::optional<torch::Tensor>& alibi_slopes, torch::Tensor& k_scale,
-    torch::Tensor& v_scale, const int tp_rank,
-    const int blocksparse_local_blocks, const int blocksparse_vert_stride,
-    const int blocksparse_block_size, const int blocksparse_head_sliding_step) {
+    torch::stable::Tensor& out, torch::stable::Tensor& query,
+    torch::stable::Tensor& key_cache, torch::stable::Tensor& value_cache,
+    int num_kv_heads, float scale, torch::stable::Tensor& block_tables,
+    torch::stable::Tensor& seq_lens, int max_seq_len,
+    const std::optional<torch::stable::Tensor>& alibi_slopes,
+    torch::stable::Tensor& k_scale, torch::stable::Tensor& v_scale,
+    const int tp_rank, const int blocksparse_local_blocks,
+    const int blocksparse_vert_stride, const int blocksparse_block_size,
+    const int blocksparse_head_sliding_step) {
   int num_seqs = query.size(0);
   int num_heads = query.size(1);
   int head_size = query.size(2);
@@ -69,8 +72,8 @@ void paged_attention_v1_launcher(
   T* query_ptr = reinterpret_cast<T*>(query.data_ptr());
   CACHE_T* key_cache_ptr = reinterpret_cast<CACHE_T*>(key_cache.data_ptr());
   CACHE_T* value_cache_ptr = reinterpret_cast<CACHE_T*>(value_cache.data_ptr());
-  int* block_tables_ptr = block_tables.data_ptr<int>();
-  int* seq_lens_ptr = seq_lens.data_ptr<int>();
+  int* block_tables_ptr = block_tables.mutable_data_ptr<int>();
+  int* seq_lens_ptr = seq_lens.mutable_data_ptr<int>();
   const float* k_scale_ptr = reinterpret_cast<const float*>(k_scale.data_ptr());
   const float* v_scale_ptr = reinterpret_cast<const float*>(v_scale.data_ptr());
 
@@ -85,8 +88,9 @@ void paged_attention_v1_launcher(
 
   dim3 grid(num_heads, num_seqs, 1);
   dim3 block(NUM_THREADS);
-  const at::cuda::OptionalCUDAGuard device_guard(device_of(query));
-  const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
+  const torch::stable::accelerator::DeviceGuard device_guard(
+      query.get_device_index());
+  const cudaStream_t stream = get_current_cuda_stream();
   switch (head_size) {
     // NOTE(woosuk): To reduce the compilation time, we only compile for the
     // head sizes that we use in the model. However, we can easily extend this
@@ -119,7 +123,7 @@ void paged_attention_v1_launcher(
       LAUNCH_PAGED_ATTENTION_V1(256);
       break;
     default:
-      TORCH_CHECK(false, "Unsupported head size: ", head_size);
+      STD_TORCH_CHECK(false, "Unsupported head size: ", head_size);
       break;
   }
 }
@@ -141,43 +145,43 @@ void paged_attention_v1_launcher(
 
 // NOTE(woosuk): To reduce the compilation time, we omitted block sizes
 // 1, 2, 4, 64, 128, 256.
-#define CALL_V1_LAUNCHER_BLOCK_SIZE(T, CACHE_T, KV_DTYPE)         \
-  switch (block_size) {                                           \
-    case 8:                                                       \
-      CALL_V1_LAUNCHER_SPARSITY(T, CACHE_T, 8, KV_DTYPE);         \
-      break;                                                      \
-    case 16:                                                      \
-      CALL_V1_LAUNCHER_SPARSITY(T, CACHE_T, 16, KV_DTYPE);        \
-      break;                                                      \
-    case 32:                                                      \
-      CALL_V1_LAUNCHER_SPARSITY(T, CACHE_T, 32, KV_DTYPE);        \
-      break;                                                      \
-    default:                                                      \
-      TORCH_CHECK(false, "Unsupported block size: ", block_size); \
-      break;                                                      \
+#define CALL_V1_LAUNCHER_BLOCK_SIZE(T, CACHE_T, KV_DTYPE)             \
+  switch (block_size) {                                               \
+    case 8:                                                           \
+      CALL_V1_LAUNCHER_SPARSITY(T, CACHE_T, 8, KV_DTYPE);             \
+      break;                                                          \
+    case 16:                                                          \
+      CALL_V1_LAUNCHER_SPARSITY(T, CACHE_T, 16, KV_DTYPE);            \
+      break;                                                          \
+    case 32:                                                          \
+      CALL_V1_LAUNCHER_SPARSITY(T, CACHE_T, 32, KV_DTYPE);            \
+      break;                                                          \
+    default:                                                          \
+      STD_TORCH_CHECK(false, "Unsupported block size: ", block_size); \
+      break;                                                          \
   }
 
 void paged_attention_v1(
-    torch::Tensor& out,    // [num_seqs, num_heads, head_size]
-    torch::Tensor& query,  // [num_seqs, num_heads, head_size]
-    torch::Tensor&
+    torch::stable::Tensor& out,    // [num_seqs, num_heads, head_size]
+    torch::stable::Tensor& query,  // [num_seqs, num_heads, head_size]
+    torch::stable::Tensor&
         key_cache,  // [num_blocks, num_heads, head_size/x, block_size, x]
-    torch::Tensor&
+    torch::stable::Tensor&
         value_cache,       // [num_blocks, num_heads, head_size, block_size]
     int64_t num_kv_heads,  // [num_heads]
     double scale,
-    torch::Tensor& block_tables,  // [num_seqs, max_num_blocks_per_seq]
-    torch::Tensor& seq_lens,      // [num_seqs]
+    torch::stable::Tensor& block_tables,  // [num_seqs, max_num_blocks_per_seq]
+    torch::stable::Tensor& seq_lens,      // [num_seqs]
     int64_t block_size, int64_t max_seq_len,
-    const std::optional<torch::Tensor>& alibi_slopes,
-    const std::string& kv_cache_dtype, torch::Tensor& k_scale,
-    torch::Tensor& v_scale, const int64_t tp_rank,
+    const std::optional<torch::stable::Tensor>& alibi_slopes,
+    const std::string& kv_cache_dtype, torch::stable::Tensor& k_scale,
+    torch::stable::Tensor& v_scale, const int64_t tp_rank,
     const int64_t blocksparse_local_blocks,
     const int64_t blocksparse_vert_stride, const int64_t blocksparse_block_size,
     const int64_t blocksparse_head_sliding_step) {
   const bool is_block_sparse = (blocksparse_vert_stride > 1);
 
-  DISPATCH_BY_KV_CACHE_DTYPE(query.dtype(), kv_cache_dtype,
+  DISPATCH_BY_KV_CACHE_DTYPE(query.scalar_type(), kv_cache_dtype,
                              CALL_V1_LAUNCHER_BLOCK_SIZE)
 }
 

csrc/libtorch_stable/attention/paged_attention_v2.cu

@@ -16,8 +16,9 @@
  * See the License for the specific language governing permissions and
  * limitations under the License.
  */
+#include "../torch_utils.h"
 #include "attention_kernels.cuh"
-#include "../cuda_compat.h"
+#include "../../cuda_compat.h"
 
 #define MAX(a, b) ((a) > (b) ? (a) : (b))
 #define MIN(a, b) ((a) < (b) ? (a) : (b))
@@ -44,14 +45,16 @@ template <typename T, typename CACHE_T, int BLOCK_SIZE,
           vllm::Fp8KVCacheDataType KV_DTYPE, bool IS_BLOCK_SPARSE,
           int NUM_THREADS = 128, int PARTITION_SIZE = 512>
 void paged_attention_v2_launcher(
-    torch::Tensor& out, torch::Tensor& exp_sums, torch::Tensor& max_logits,
-    torch::Tensor& tmp_out, torch::Tensor& query, torch::Tensor& key_cache,
-    torch::Tensor& value_cache, int num_kv_heads, float scale,
-    torch::Tensor& block_tables, torch::Tensor& seq_lens, int max_seq_len,
-    const std::optional<torch::Tensor>& alibi_slopes, torch::Tensor& k_scale,
-    torch::Tensor& v_scale, const int tp_rank,
-    const int blocksparse_local_blocks, const int blocksparse_vert_stride,
-    const int blocksparse_block_size, const int blocksparse_head_sliding_step) {
+    torch::stable::Tensor& out, torch::stable::Tensor& exp_sums,
+    torch::stable::Tensor& max_logits, torch::stable::Tensor& tmp_out,
+    torch::stable::Tensor& query, torch::stable::Tensor& key_cache,
+    torch::stable::Tensor& value_cache, int num_kv_heads, float scale,
+    torch::stable::Tensor& block_tables, torch::stable::Tensor& seq_lens,
+    int max_seq_len, const std::optional<torch::stable::Tensor>& alibi_slopes,
+    torch::stable::Tensor& k_scale, torch::stable::Tensor& v_scale,
+    const int tp_rank, const int blocksparse_local_blocks,
+    const int blocksparse_vert_stride, const int blocksparse_block_size,
+    const int blocksparse_head_sliding_step) {
   int num_seqs = query.size(0);
   int num_heads = query.size(1);
   int head_size = query.size(2);
@@ -73,8 +76,8 @@ void paged_attention_v2_launcher(
   T* query_ptr = reinterpret_cast<T*>(query.data_ptr());
   CACHE_T* key_cache_ptr = reinterpret_cast<CACHE_T*>(key_cache.data_ptr());
   CACHE_T* value_cache_ptr = reinterpret_cast<CACHE_T*>(value_cache.data_ptr());
-  int* block_tables_ptr = block_tables.data_ptr<int>();
-  int* seq_lens_ptr = seq_lens.data_ptr<int>();
+  int* block_tables_ptr = block_tables.mutable_data_ptr<int>();
+  int* seq_lens_ptr = seq_lens.mutable_data_ptr<int>();
   const float* k_scale_ptr = reinterpret_cast<const float*>(k_scale.data_ptr());
   const float* v_scale_ptr = reinterpret_cast<const float*>(v_scale.data_ptr());
 
@@ -91,8 +94,9 @@ void paged_attention_v2_launcher(
   int reduce_shared_mem_size = 2 * max_num_partitions * sizeof(float);
 
   dim3 block(NUM_THREADS);
-  const at::cuda::OptionalCUDAGuard device_guard(device_of(query));
-  const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
+  const torch::stable::accelerator::DeviceGuard device_guard(
+      query.get_device_index());
+  const cudaStream_t stream = get_current_cuda_stream();
   switch (head_size) {
     // NOTE(woosuk): To reduce the compilation time, we only compile for the
     // head sizes that we use in the model. However, we can easily extend this
@@ -125,7 +129,7 @@ void paged_attention_v2_launcher(
       LAUNCH_PAGED_ATTENTION_V2(256);
       break;
     default:
-      TORCH_CHECK(false, "Unsupported head size: ", head_size);
+      STD_TORCH_CHECK(false, "Unsupported head size: ", head_size);
       break;
   }
 }
@@ -148,46 +152,48 @@ void paged_attention_v2_launcher(
 
 // NOTE(woosuk): To reduce the compilation time, we omitted block sizes
 // 1, 2, 4, 64, 128, 256.
-#define CALL_V2_LAUNCHER_BLOCK_SIZE(T, CACHE_T, KV_DTYPE)         \
-  switch (block_size) {                                           \
-    case 8:                                                       \
-      CALL_V2_LAUNCHER_SPARSITY(T, CACHE_T, 8, KV_DTYPE);         \
-      break;                                                      \
-    case 16:                                                      \
-      CALL_V2_LAUNCHER_SPARSITY(T, CACHE_T, 16, KV_DTYPE);        \
-      break;                                                      \
-    case 32:                                                      \
-      CALL_V2_LAUNCHER_SPARSITY(T, CACHE_T, 32, KV_DTYPE);        \
-      break;                                                      \
-    default:                                                      \
-      TORCH_CHECK(false, "Unsupported block size: ", block_size); \
-      break;                                                      \
+#define CALL_V2_LAUNCHER_BLOCK_SIZE(T, CACHE_T, KV_DTYPE)             \
+  switch (block_size) {                                               \
+    case 8:                                                           \
+      CALL_V2_LAUNCHER_SPARSITY(T, CACHE_T, 8, KV_DTYPE);             \
+      break;                                                          \
+    case 16:                                                          \
+      CALL_V2_LAUNCHER_SPARSITY(T, CACHE_T, 16, KV_DTYPE);            \
+      break;                                                          \
+    case 32:                                                          \
+      CALL_V2_LAUNCHER_SPARSITY(T, CACHE_T, 32, KV_DTYPE);            \
+      break;                                                          \
+    default:                                                          \
+      STD_TORCH_CHECK(false, "Unsupported block size: ", block_size); \
+      break;                                                          \
   }
 
 void paged_attention_v2(
-    torch::Tensor& out,         // [num_seqs, num_heads, head_size]
-    torch::Tensor& exp_sums,    // [num_seqs, num_heads, max_num_partitions]
-    torch::Tensor& max_logits,  // [num_seqs, num_heads, max_num_partitions]
-    torch::Tensor&
+    torch::stable::Tensor& out,  // [num_seqs, num_heads, head_size]
+    torch::stable::Tensor&
+        exp_sums,  // [num_seqs, num_heads, max_num_partitions]
+    torch::stable::Tensor&
+        max_logits,  // [num_seqs, num_heads, max_num_partitions]
+    torch::stable::Tensor&
         tmp_out,  // [num_seqs, num_heads, max_num_partitions, head_size]
-    torch::Tensor& query,  // [num_seqs, num_heads, head_size]
-    torch::Tensor&
+    torch::stable::Tensor& query,  // [num_seqs, num_heads, head_size]
+    torch::stable::Tensor&
         key_cache,  // [num_blocks, num_heads, head_size/x, block_size, x]
-    torch::Tensor&
+    torch::stable::Tensor&
         value_cache,       // [num_blocks, num_heads, head_size, block_size]
     int64_t num_kv_heads,  // [num_heads]
     double scale,
-    torch::Tensor& block_tables,  // [num_seqs, max_num_blocks_per_seq]
-    torch::Tensor& seq_lens,      // [num_seqs]
+    torch::stable::Tensor& block_tables,  // [num_seqs, max_num_blocks_per_seq]
+    torch::stable::Tensor& seq_lens,      // [num_seqs]
     int64_t block_size, int64_t max_seq_len,
-    const std::optional<torch::Tensor>& alibi_slopes,
-    const std::string& kv_cache_dtype, torch::Tensor& k_scale,
-    torch::Tensor& v_scale, const int64_t tp_rank,
+    const std::optional<torch::stable::Tensor>& alibi_slopes,
+    const std::string& kv_cache_dtype, torch::stable::Tensor& k_scale,
+    torch::stable::Tensor& v_scale, const int64_t tp_rank,
     const int64_t blocksparse_local_blocks,
     const int64_t blocksparse_vert_stride, const int64_t blocksparse_block_size,
     const int64_t blocksparse_head_sliding_step) {
   const bool is_block_sparse = (blocksparse_vert_stride > 1);
-  DISPATCH_BY_KV_CACHE_DTYPE(query.dtype(), kv_cache_dtype,
+  DISPATCH_BY_KV_CACHE_DTYPE(query.scalar_type(), kv_cache_dtype,
                              CALL_V2_LAUNCHER_BLOCK_SIZE)
 }
 

csrc/libtorch_stable/cache_kernels_fused.cu

@@ -1,15 +1,13 @@
-#include <torch/all.h>
-#include <ATen/cuda/CUDAContext.h>
-#include <c10/cuda/CUDAGuard.h>
-
-#include "cuda_compat.h"
+#include "torch_utils.h"
 #include "dispatch_utils.h"
 
-#include "quantization/w8a8/fp8/common.cuh"
+#include "../cuda_compat.h"
+
+#include "../quantization/w8a8/fp8/common.cuh"
 #ifdef USE_ROCM
-  #include "quantization/w8a8/fp8/amd/quant_utils.cuh"
+  #include "../quantization/w8a8/fp8/amd/quant_utils.cuh"
 #else
-  #include "quantization/w8a8/fp8/nvidia/quant_utils.cuh"
+  #include "../quantization/w8a8/fp8/nvidia/quant_utils.cuh"
 #endif
 
 #ifdef USE_ROCM
@@ -164,43 +162,52 @@ __global__ void concat_and_cache_mla_rope_fused_kernel(
 
 }  // namespace vllm
 
-#define CALL_CONCAT_AND_CACHE_MLA_ROPE_FUSED(RAW_KV_T, CACHE_T, KV_DTYPE)     \
-  do {                                                                        \
-    VLLM_DISPATCH_FLOATING_TYPES(q_pe.scalar_type(), "qk_scalar_type", [&] {  \
-      using qk_t = scalar_t;                                                  \
-      VLLM_DISPATCH_FLOATING_TYPES(                                           \
-          rope_cos_sin_cache.scalar_type(), "rope_cos_sin_cache_scalar_type", \
-          [&] {                                                               \
-            using cos_sin_t = scalar_t;                                       \
-            if (rope_is_neox) {                                               \
-              vllm::concat_and_cache_mla_rope_fused_kernel<                   \
-                  qk_t, cos_sin_t, true, RAW_KV_T, CACHE_T, KV_DTYPE>         \
-                  <<<grid, block, 0, stream>>>(                               \
-                      positions.data_ptr<int64_t>(), q_pe.data_ptr<qk_t>(),   \
-                      k_pe.data_ptr<qk_t>(), kv_c.data_ptr<qk_t>(),           \
-                      rope_cos_sin_cache.data_ptr<cos_sin_t>(), rot_dim,      \
-                      q_pe_stride_token, q_pe_stride_head, k_pe_stride,       \
-                      kv_c_stride, num_q_heads,                               \
-                      reinterpret_cast<CACHE_T*>(kv_cache.data_ptr()),        \
-                      slot_mapping.data_ptr<int64_t>(), block_stride,         \
-                      entry_stride, kv_lora_rank, block_size,                 \
-                      kv_cache_quant_scale.data_ptr<float>());                \
-            } else {                                                          \
-              vllm::concat_and_cache_mla_rope_fused_kernel<                   \
-                  qk_t, cos_sin_t, false, RAW_KV_T, CACHE_T, KV_DTYPE>        \
-                  <<<grid, block, 0, stream>>>(                               \
-                      positions.data_ptr<int64_t>(), q_pe.data_ptr<qk_t>(),   \
-                      k_pe.data_ptr<qk_t>(), kv_c.data_ptr<qk_t>(),           \
-                      rope_cos_sin_cache.data_ptr<cos_sin_t>(), rot_dim,      \
-                      q_pe_stride_token, q_pe_stride_head, k_pe_stride,       \
-                      kv_c_stride, num_q_heads,                               \
-                      reinterpret_cast<CACHE_T*>(kv_cache.data_ptr()),        \
-                      slot_mapping.data_ptr<int64_t>(), block_stride,         \
-                      entry_stride, kv_lora_rank, block_size,                 \
-                      kv_cache_quant_scale.data_ptr<float>());                \
-            }                                                                 \
-          });                                                                 \
-    });                                                                       \
+#define CALL_CONCAT_AND_CACHE_MLA_ROPE_FUSED(RAW_KV_T, CACHE_T, KV_DTYPE)  \
+  do {                                                                     \
+    VLLM_STABLE_DISPATCH_FLOATING_TYPES(                                   \
+        q_pe.scalar_type(), "qk_scalar_type", [&] {                        \
+          using qk_t = scalar_t;                                           \
+          VLLM_STABLE_DISPATCH_FLOATING_TYPES(                             \
+              rope_cos_sin_cache.scalar_type(),                            \
+              "rope_cos_sin_cache_scalar_type", [&] {                      \
+                using cos_sin_t = scalar_t;                                \
+                if (rope_is_neox) {                                        \
+                  vllm::concat_and_cache_mla_rope_fused_kernel<            \
+                      qk_t, cos_sin_t, true, RAW_KV_T, CACHE_T, KV_DTYPE>  \
+                      <<<grid, block, 0, stream>>>(                        \
+                          positions.const_data_ptr<int64_t>(),             \
+                          q_pe.mutable_data_ptr<qk_t>(),                   \
+                          k_pe.mutable_data_ptr<qk_t>(),                   \
+                          kv_c.const_data_ptr<qk_t>(),                     \
+                          rope_cos_sin_cache.const_data_ptr<cos_sin_t>(),  \
+                          rot_dim, q_pe_stride_token, q_pe_stride_head,    \
+                          k_pe_stride, kv_c_stride, num_q_heads,           \
+                          reinterpret_cast<CACHE_T*>(                      \
+                              kv_cache.mutable_data_ptr()),                \
+                          slot_mapping.const_data_ptr<int64_t>(),          \
+                          block_stride, entry_stride, kv_lora_rank,        \
+                          block_size,                                      \
+                          kv_cache_quant_scale.const_data_ptr<float>());   \
+                } else {                                                   \
+                  vllm::concat_and_cache_mla_rope_fused_kernel<            \
+                      qk_t, cos_sin_t, false, RAW_KV_T, CACHE_T, KV_DTYPE> \
+                      <<<grid, block, 0, stream>>>(                        \
+                          positions.const_data_ptr<int64_t>(),             \
+                          q_pe.mutable_data_ptr<qk_t>(),                   \
+                          k_pe.mutable_data_ptr<qk_t>(),                   \
+                          kv_c.const_data_ptr<qk_t>(),                     \
+                          rope_cos_sin_cache.const_data_ptr<cos_sin_t>(),  \
+                          rot_dim, q_pe_stride_token, q_pe_stride_head,    \
+                          k_pe_stride, kv_c_stride, num_q_heads,           \
+                          reinterpret_cast<CACHE_T*>(                      \
+                              kv_cache.mutable_data_ptr()),                \
+                          slot_mapping.const_data_ptr<int64_t>(),          \
+                          block_stride, entry_stride, kv_lora_rank,        \
+                          block_size,                                      \
+                          kv_cache_quant_scale.const_data_ptr<float>());   \
+                }                                                          \
+              });                                                          \
+        });                                                                \
   } while (false)
 
 // Executes RoPE on q_pe and k_pe, then writes k_pe and kv_c in the kv cache.
@@ -208,64 +215,69 @@ __global__ void concat_and_cache_mla_rope_fused_kernel(
 // Replaces DeepseekScalingRotaryEmbedding.self.rotary_emb and
 // concat_and_cache_mla.
 void concat_and_cache_mla_rope_fused(
-    torch::Tensor& positions,           // [num_tokens]
-    torch::Tensor& q_pe,                // [num_tokens, num_q_heads, rot_dim]
-    torch::Tensor& k_pe,                // [num_tokens, rot_dim]
-    torch::Tensor& kv_c,                // [num_tokens, kv_lora_rank]
-    torch::Tensor& rope_cos_sin_cache,  // [max_position, rot_dim]
+    torch::stable::Tensor& positions,  // [num_tokens]
+    torch::stable::Tensor& q_pe,       // [num_tokens, num_q_heads, rot_dim]
+    torch::stable::Tensor& k_pe,       // [num_tokens, rot_dim]
+    torch::stable::Tensor& kv_c,       // [num_tokens, kv_lora_rank]
+    torch::stable::Tensor& rope_cos_sin_cache,  // [max_position, rot_dim]
     bool rope_is_neox,
-    torch::Tensor& slot_mapping,  // [num_tokens] or [num_actual_tokens]
-    torch::Tensor&
+    torch::stable::Tensor& slot_mapping,  // [num_tokens] or [num_actual_tokens]
+    torch::stable::Tensor&
         kv_cache,  // [num_blocks, block_size, (kv_lora_rank + rot_dim)]
-    const std::string& kv_cache_dtype, torch::Tensor& kv_cache_quant_scale) {
+    const std::string& kv_cache_dtype,
+    torch::stable::Tensor& kv_cache_quant_scale) {
   // NOTE(woosuk): In vLLM V1, query/key/position.size(0) can be different from
   // slot_mapping.size(0) because of padding for CUDA graphs.
-  // In vLLM V0, key.size(0) is always equal to slot_mapping.size(0) because
-  // both include padding.
-  // In vLLM V1, however, key.size(0) can be larger than slot_mapping.size(0)
-  // since key includes padding for CUDA graphs, while slot_mapping does not.
-  // In this case, slot_mapping.size(0) represents the actual number of tokens
+  // In vLLM V0, key.size(0) is always equal to slot_mapping.size(0)
+  // because both include padding.
+  // In vLLM V1, however, key.size(0) can be larger than
+  // slot_mapping.size(0) since key includes padding for CUDA graphs,
+  // while slot_mapping does not. In this case,
+  // slot_mapping.size(0) represents the actual number of tokens
   // before padding.
-  // For compatibility with both cases, we use slot_mapping.size(0) as the
-  // number of tokens.
-  int num_tokens = slot_mapping.size(0);
-  int num_padded_tokens = q_pe.size(0);
-  TORCH_CHECK_GE(num_padded_tokens, num_tokens);
+  // For compatibility with both cases, we use slot_mapping.size(0) as
+  // the number of tokens.
+  const int64_t num_tokens = slot_mapping.size(0);
+  const int64_t num_padded_tokens = q_pe.size(0);
+  STD_TORCH_CHECK(num_padded_tokens >= num_tokens);
 
   const int num_q_heads = q_pe.size(1);
   const int rot_dim = q_pe.size(2);
   const int kv_lora_rank = kv_c.size(1);
 
-  TORCH_CHECK_EQ(positions.size(0), num_padded_tokens);
-  TORCH_CHECK_EQ(positions.dim(), 1);
-  TORCH_CHECK_EQ(positions.scalar_type(), c10::ScalarType::Long);
+  STD_TORCH_CHECK(positions.size(0) == num_padded_tokens);
+  STD_TORCH_CHECK(positions.dim() == 1);
+  STD_TORCH_CHECK(positions.scalar_type() ==
+                  torch::headeronly::ScalarType::Long);
 
-  TORCH_CHECK_EQ(q_pe.dim(), 3);
-  TORCH_CHECK_EQ(q_pe.size(0), num_padded_tokens);
-  TORCH_CHECK_EQ(q_pe.size(1), num_q_heads);
-  TORCH_CHECK_EQ(q_pe.size(2), rot_dim);
+  STD_TORCH_CHECK(q_pe.dim() == 3);
+  STD_TORCH_CHECK(q_pe.size(0) == num_padded_tokens);
+  STD_TORCH_CHECK(q_pe.size(1) == num_q_heads);
+  STD_TORCH_CHECK(q_pe.size(2) == rot_dim);
 
-  TORCH_CHECK_EQ(k_pe.dim(), 2);
-  TORCH_CHECK_EQ(k_pe.size(0), num_padded_tokens);
-  TORCH_CHECK_EQ(k_pe.size(1), rot_dim);
-  TORCH_CHECK_EQ(k_pe.scalar_type(), q_pe.scalar_type());
+  STD_TORCH_CHECK(k_pe.dim() == 2);
+  STD_TORCH_CHECK(k_pe.size(0) == num_padded_tokens);
+  STD_TORCH_CHECK(k_pe.size(1) == rot_dim);
+  STD_TORCH_CHECK(k_pe.scalar_type() == q_pe.scalar_type());
 
-  TORCH_CHECK_EQ(kv_c.dim(), 2);
-  TORCH_CHECK_EQ(kv_c.size(0), num_padded_tokens);
-  TORCH_CHECK_EQ(kv_c.size(1), kv_lora_rank);
-  TORCH_CHECK_EQ(kv_c.scalar_type(), q_pe.scalar_type());
-  TORCH_CHECK_EQ(kv_c.dtype(), q_pe.dtype());
+  STD_TORCH_CHECK(kv_c.dim() == 2);
+  STD_TORCH_CHECK(kv_c.size(0) == num_padded_tokens);
+  STD_TORCH_CHECK(kv_c.size(1) == kv_lora_rank);
+  STD_TORCH_CHECK(kv_c.scalar_type() == q_pe.scalar_type());
 
-  TORCH_CHECK_EQ(rope_cos_sin_cache.size(1), rot_dim);
+  STD_TORCH_CHECK(rope_cos_sin_cache.size(1) == rot_dim);
+  STD_TORCH_CHECK(rope_cos_sin_cache.scalar_type() == q_pe.scalar_type());
 
-  TORCH_CHECK_EQ(slot_mapping.size(0), num_tokens);
-  TORCH_CHECK_EQ(slot_mapping.scalar_type(), c10::ScalarType::Long);
+  STD_TORCH_CHECK(slot_mapping.size(0) == num_tokens);
+  STD_TORCH_CHECK(slot_mapping.scalar_type() ==
+                  torch::headeronly::ScalarType::Long);
 
-  TORCH_CHECK_EQ(kv_cache.size(2), kv_lora_rank + rot_dim);
-  TORCH_CHECK_EQ(kv_cache.dim(), 3);
+  STD_TORCH_CHECK(kv_cache.size(2) == kv_lora_rank + rot_dim);
+  STD_TORCH_CHECK(kv_cache.dim() == 3);
 
-  TORCH_CHECK_EQ(kv_cache_quant_scale.numel(), 1);
-  TORCH_CHECK_EQ(kv_cache_quant_scale.scalar_type(), c10::ScalarType::Float);
+  STD_TORCH_CHECK(kv_cache_quant_scale.numel() == 1);
+  STD_TORCH_CHECK(kv_cache_quant_scale.scalar_type() ==
+                  torch::headeronly::ScalarType::Float);
 
   int64_t q_pe_stride_token = q_pe.stride(0);
   int64_t q_pe_stride_head = q_pe.stride(1);
@@ -286,9 +298,10 @@ void concat_and_cache_mla_rope_fused(
   dim3 grid(num_tokens, 1, 1);
   dim3 block(thread_block_size, 1, 1);
 
-  const at::cuda::OptionalCUDAGuard device_guard(device_of(positions));
-  const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
+  const torch::stable::accelerator::DeviceGuard device_guard(
+      positions.get_device_index());
+  const cudaStream_t stream = get_current_cuda_stream();
 
-  DISPATCH_BY_KV_CACHE_DTYPE(kv_c.dtype(), kv_cache_dtype,
+  DISPATCH_BY_KV_CACHE_DTYPE(kv_c.scalar_type(), kv_cache_dtype,
                              CALL_CONCAT_AND_CACHE_MLA_ROPE_FUSED);
 }

csrc/libtorch_stable/cuda_vec_utils.cuh

@@ -9,6 +9,8 @@
 
 #ifdef USE_ROCM
   #include <hip/hip_runtime.h>
+  #include <hip/hip_bf16.h>
+  #include <hip/hip_fp16.h>
 #else
   #include <cuda_bf16.h>
   #include <cuda_fp16.h>