remove torch 2.9, 2.10 workround

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

.buildkite/hardware_tests/cpu.yaml

@@ -69,11 +69,11 @@ steps:
       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"
       
-- label: CPU-Distributed Tests
+- label: CPU-Distributed Tests (PP+TP)
   depends_on: []
   device: intel_cpu
   no_plugin: true
-  source_file_dependencies:
+  source_file_dependencies: &cpu_distributed_deps
   - csrc/cpu/shm.cpp
   - vllm/v1/worker/cpu_worker.py
   - vllm/v1/worker/gpu_worker.py
@@ -82,10 +82,21 @@ steps:
   - vllm/platforms/cpu.py
   - vllm/distributed/parallel_state.py
   - vllm/distributed/device_communicators/cpu_communicator.py
+  - .buildkite/scripts/hardware_ci/run-cpu-distributed-smoke-test.sh
   commands:
     - |
       bash .buildkite/scripts/hardware_ci/run-cpu-test.sh 10m "
-      bash .buildkite/scripts/hardware_ci/run-cpu-distributed-smoke-test.sh"
+      bash .buildkite/scripts/hardware_ci/run-cpu-distributed-smoke-test.sh tp_pp"
+
+- label: CPU-Distributed Tests (DP+TP)
+  depends_on: []
+  device: intel_cpu
+  no_plugin: true
+  source_file_dependencies: *cpu_distributed_deps
+  commands:
+    - |
+      bash .buildkite/scripts/hardware_ci/run-cpu-test.sh 10m "
+      bash .buildkite/scripts/hardware_ci/run-cpu-distributed-smoke-test.sh dp_tp"
 
 - label: CPU-Multi-Modal Model Tests %N
   depends_on: []

.buildkite/image_build/image_build.sh

@@ -92,8 +92,8 @@ check_and_skip_if_image_exists() {
 }
 
 ecr_login() {
-    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
 }
 
 prepare_cache_tags() {
@@ -192,6 +192,7 @@ export BUILDKITE_COMMIT
 export PARENT_COMMIT
 export IMAGE_TAG
 export IMAGE_TAG_LATEST
+export COMMIT="${COMMIT:-${BUILDKITE_COMMIT}}"
 export CACHE_FROM
 export CACHE_FROM_BASE_BRANCH
 export CACHE_FROM_MAIN

.buildkite/image_build/image_build_cpu.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"-cpu) ]]; then

.buildkite/image_build/image_build_cpu_arm64.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"-arm64-cpu) ]]; then

.buildkite/image_build/image_build_torch_nightly.sh

@@ -46,7 +46,7 @@ echo "Image not found, proceeding with build..."
 
 # --- CUDA 13.0 for nightly builds ---
 # Nightly CI uses CUDA 13.0 while regular CI stays on CUDA 12.9
-NIGHTLY_CUDA_VERSION="13.0.0"
+NIGHTLY_CUDA_VERSION="13.0.2"
 NIGHTLY_BUILD_BASE_IMAGE="nvidia/cuda:${NIGHTLY_CUDA_VERSION}-devel-ubuntu22.04"
 NIGHTLY_FINAL_BASE_IMAGE="nvidia/cuda:${NIGHTLY_CUDA_VERSION}-base-ubuntu22.04"
 

.buildkite/intel_jobs/engine_intel.yaml

@@ -0,0 +1,21 @@
+group: Engine Intel
+depends_on:
+  - image-build-xpu
+steps:
+- label: Engine (1 GPU)
+  timeout_in_minutes: 30
+  device: intel_gpu
+  no_plugin: true
+  working_dir: "."
+  env:
+    REGISTRY: "public.ecr.aws/q9t5s3a7"
+    REPO: "vllm-ci-test-repo"
+    VLLM_TEST_DEVICE: "xpu"
+  source_file_dependencies:
+    - vllm/v1/engine/
+    - tests/v1/engine/
+  commands:
+    - >-
+      bash .buildkite/scripts/hardware_ci/run-intel-test.sh
+      'cd tests &&
+      pytest -v -s v1/engine --ignore v1/engine/test_preprocess_error_handling.py'

.buildkite/intel_jobs/kernels_intel.yaml

@@ -0,0 +1,21 @@
+group: Kernels Intel
+depends_on: 
+  - image-build-xpu
+steps:
+- label: vLLM IR Tests
+  timeout_in_minutes: 30
+  device: intel_gpu
+  no_plugin: true
+  working_dir: "."
+  env:
+    REGISTRY: "public.ecr.aws/q9t5s3a7"
+    REPO: "vllm-ci-test-repo"
+    VLLM_TEST_DEVICE: "xpu"
+  source_file_dependencies:
+    - vllm/ir
+    - vllm/kernels
+  commands:
+    - >-
+      bash .buildkite/scripts/hardware_ci/run-intel-test.sh
+      'cd tests &&
+      pytest -v -s kernels/ir'

.buildkite/intel_jobs/lora_intel.yaml

@@ -0,0 +1,129 @@
+group: LoRA Intel
+depends_on:
+  - image-build-xpu
+steps:
+- label: LoRA Runtime + Utils
+  timeout_in_minutes: 45
+  device: intel_gpu
+  no_plugin: true
+  working_dir: "."
+  env:
+    REGISTRY: "public.ecr.aws/q9t5s3a7"
+    REPO: "vllm-ci-test-repo"
+    VLLM_TEST_DEVICE: "xpu"
+  source_file_dependencies:
+    - vllm/lora
+    - tests/lora
+  commands:
+    - >-
+      bash .buildkite/scripts/hardware_ci/run-intel-test.sh
+      'cd tests &&
+      pytest -v -s lora/test_layers.py &&
+      pytest -v -s lora/test_lora_checkpoints.py &&
+      (pytest -v -s lora/test_lora_functions.py --deselect="tests/lora/test_lora_functions.py::test_lora_functions_sync" --deselect="tests/lora/test_lora_functions.py::test_lora_functions_async" || true) &&
+      pytest -v -s lora/test_lora_huggingface.py &&
+      pytest -v -s lora/test_lora_manager.py &&
+      pytest -v -s lora/test_lora_utils.py &&
+      pytest -v -s lora/test_peft_helper.py &&
+      pytest -v -s lora/test_resolver.py &&
+      pytest -v -s lora/test_utils.py &&
+      (pytest -v -s lora/test_add_lora.py --deselect="tests/lora/test_add_lora.py::test_add_lora" || true) &&
+      (pytest -v -s lora/test_worker.py --deselect="tests/lora/test_worker.py::test_worker_apply_lora" || true)'
+
+- label: LoRA Fused/MoE Kernels
+  timeout_in_minutes: 45
+  device: intel_gpu
+  no_plugin: true
+  working_dir: "."
+  env:
+    REGISTRY: "public.ecr.aws/q9t5s3a7"
+    REPO: "vllm-ci-test-repo"
+    VLLM_TEST_DEVICE: "xpu"
+  source_file_dependencies:
+    - vllm/lora
+    - tests/lora
+  commands:
+    - >-
+      bash .buildkite/scripts/hardware_ci/run-intel-test.sh
+      'cd tests &&
+      pytest -v -s lora/test_fused_moe_lora_kernel.py && 
+      pytest -v -s lora/test_moe_lora_align_sum.py'
+
+- label: LoRA Punica Kernels
+  timeout_in_minutes: 45
+  device: intel_gpu
+  no_plugin: true
+  working_dir: "."
+  env:
+    REGISTRY: "public.ecr.aws/q9t5s3a7"
+    REPO: "vllm-ci-test-repo"
+    VLLM_TEST_DEVICE: "xpu"
+  source_file_dependencies:
+    - vllm/lora
+    - tests/lora
+  commands:
+    - >-
+      bash .buildkite/scripts/hardware_ci/run-intel-test.sh
+      'cd tests &&
+      set -o pipefail &&
+      pytest -v -s lora/test_punica_ops.py --deselect="tests/lora/test_punica_ops.py::test_kernels[shrink-0-xpu:0-dtype0-2-2049-64-32-32]" --deselect="tests/lora/test_punica_ops.py::test_kernels_hidden_size[expand-0-xpu:0-dtype1-2-64000-32-4-4]" --deselect="tests/lora/test_punica_ops.py::test_kernels[shrink-0-xpu:0-dtype0-1-2049-128-1-32]" --deselect="tests/lora/test_punica_ops.py::test_kernels[shrink-0-xpu:0-dtype0-1-2049-256-1-4]" --deselect="tests/lora/test_punica_ops.py::test_kernels[shrink-0-xpu:0-dtype0-1-2049-256-8-4]" --deselect="tests/lora/test_punica_ops.py::test_kernels[expand-0-xpu:0-dtype0-3-2049-128-8-16]" --deselect="tests/lora/test_punica_ops.py::test_kernels[shrink-0-xpu:0-dtype0-1-2049-128-8-32]" --deselect="tests/lora/test_punica_ops.py::test_kernels[expand-0-xpu:0-dtype1-1-2049-256-128-32]" --deselect="tests/lora/test_punica_ops.py::test_kernels_hidden_size[shrink-0-xpu:0-dtype0-3-64256-32-4-4]" --deselect="tests/lora/test_punica_ops.py::test_kernels_hidden_size[shrink-0-xpu:0-dtype1-2-29696-32-4-4]" --deselect="tests/lora/test_punica_ops.py::test_kernels_hidden_size[shrink-0-xpu:0-dtype1-3-49408-32-4-4]" --deselect="tests/lora/test_punica_ops.py::test_kernels_hidden_size[shrink-0-xpu:0-dtype0-2-16384-32-4-4]" --deselect="tests/lora/test_punica_ops.py::test_kernels_hidden_size[expand-0-xpu:0-dtype0-2-51328-32-4-4]" --deselect="tests/lora/test_punica_ops.py::test_kernels_hidden_size[expand-0-xpu:0-dtype1-1-102656-32-4-4]"'
+
+- label: LoRA Punica FP8/XPU Ops
+  timeout_in_minutes: 45
+  device: intel_gpu
+  no_plugin: true
+  working_dir: "."
+  env:
+    REGISTRY: "public.ecr.aws/q9t5s3a7"
+    REPO: "vllm-ci-test-repo"
+    VLLM_TEST_DEVICE: "xpu"
+  source_file_dependencies:
+    - vllm/lora
+    - tests/lora
+  commands:
+    - >-
+      bash .buildkite/scripts/hardware_ci/run-intel-test.sh
+      'cd tests &&
+      pytest -v -s lora/test_punica_ops_fp8.py &&
+      pytest -v -s lora/test_punica_xpu_ops.py'
+
+- label: LoRA Models
+  timeout_in_minutes: 45
+  device: intel_gpu
+  no_plugin: true
+  working_dir: "."
+  env:
+    REGISTRY: "public.ecr.aws/q9t5s3a7"
+    REPO: "vllm-ci-test-repo"
+    VLLM_TEST_DEVICE: "xpu"
+  source_file_dependencies:
+    - vllm/lora
+    - tests/lora
+  commands:
+    - >-
+      bash .buildkite/scripts/hardware_ci/run-intel-test.sh
+      'cd tests &&
+      (pytest -v -s lora/test_mixtral.py --deselect="tests/lora/test_mixtral.py::test_mixtral_lora[4]" || true) &&
+      pytest -v -s lora/test_quant_model.py --deselect="tests/lora/test_quant_model.py::test_quant_model_lora[model0]" --deselect="tests/lora/test_quant_model.py::test_quant_model_lora[model1]" --deselect="tests/lora/test_quant_model.py::test_quant_model_tp_equality[model0]" &&
+      pytest -v -s lora/test_qwen35_densemodel_lora.py &&
+      pytest -v -s lora/test_transformers_model.py'
+
+- label: LoRA Multimodal
+  timeout_in_minutes: 45
+  device: intel_gpu
+  no_plugin: true
+  working_dir: "."
+  env:
+    REGISTRY: "public.ecr.aws/q9t5s3a7"
+    REPO: "vllm-ci-test-repo"
+    VLLM_TEST_DEVICE: "xpu"
+  source_file_dependencies:
+    - vllm/lora
+    - tests/lora
+  commands:
+    - >-
+      bash .buildkite/scripts/hardware_ci/run-intel-test.sh
+      'cd tests &&
+      pytest -v -s lora/test_default_mm_loras.py && 
+      (pytest -v -s lora/test_qwen3_unembed.py || true) &&
+      pytest -v -s lora/test_whisper.py'

.buildkite/intel_jobs/misc_intel.yaml

@@ -0,0 +1,55 @@
+group: Miscellaneous Intel
+depends_on:
+  - image-build-xpu
+steps:
+- label: V1 Core + KV + Metrics
+  timeout_in_minutes: 30
+  device: intel_gpu
+  no_plugin: true
+  working_dir: "."
+  env:
+    REGISTRY: "public.ecr.aws/q9t5s3a7"
+    REPO: "vllm-ci-test-repo"
+    VLLM_TEST_DEVICE: "xpu"
+  source_file_dependencies:
+    - vllm/
+    - tests/v1/core
+    - tests/v1/executor
+    - tests/v1/kv_offload
+    - tests/v1/worker
+    - tests/v1/kv_connector/unit
+    - tests/v1/metrics
+    - tests/entrypoints/openai/correctness/test_lmeval.py
+  commands:
+    - >-
+      bash .buildkite/scripts/hardware_ci/run-intel-test.sh
+      'pip install -r requirements/kv_connectors.txt &&
+      export VLLM_WORKER_MULTIPROC_METHOD=spawn &&
+      cd tests &&
+      pytest -v -s v1/executor'
+
+- label: V1 Sample + Logits
+  timeout_in_minutes: 30
+  device: intel_gpu
+  no_plugin: true
+  working_dir: "."
+  env:
+    REGISTRY: "public.ecr.aws/q9t5s3a7"
+    REPO: "vllm-ci-test-repo"
+    VLLM_TEST_DEVICE: "xpu"
+  source_file_dependencies:
+    - vllm/
+    - tests/v1/sample
+    - tests/v1/logits_processors
+    - tests/v1/test_oracle.py
+    - tests/v1/test_request.py
+    - tests/v1/test_outputs.py
+  commands:
+    - >-
+      bash .buildkite/scripts/hardware_ci/run-intel-test.sh
+      'export VLLM_WORKER_MULTIPROC_METHOD=spawn &&
+      cd tests &&
+      pytest -v -s v1/logits_processors &&
+      pytest -v -s v1/test_oracle.py &&
+      pytest -v -s v1/test_request.py &&
+      pytest -v -s v1/test_outputs.py'

.buildkite/release-pipeline.yaml

@@ -1,3 +1,13 @@
+# CUDA architecture lists — following PyTorch RELEASE.md
+# (https://github.com/pytorch/pytorch/blob/main/RELEASE.md)
+# SM86 included for broader Ampere coverage; SM89 for marlin fp8 support
+env:
+  CUDA_ARCH_X86: "7.5 8.0 8.6 8.9 9.0 10.0 12.0+PTX"
+  # aarch64 only architectures: 8.7 for Orin, 11.0 for Thor (since CUDA 13)
+  CUDA_ARCH_AARCH64: "8.0 8.7 8.9 9.0 10.0 11.0 12.0+PTX"
+  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"
+
 steps:
   - input: "Provide Release version here"
     id: input-release-version
@@ -14,9 +24,7 @@ steps:
         agents:
           queue: arm64_cpu_queue_release
         commands:
-          # #NOTE: torch_cuda_arch_list is derived from upstream PyTorch build files here:
-          # https://github.com/pytorch/pytorch/blob/main/.ci/aarch64_linux/aarch64_ci_build.sh#L7
-          - "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='8.7 8.9 9.0 10.0+PTX 12.0' --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}\" --tag vllm-ci:build-image --target build --progress plain -f docker/Dockerfile ."
           - "mkdir artifacts"
           - "docker run --rm -v $(pwd)/artifacts:/artifacts_host vllm-ci:build-image bash -c 'cp -r dist /artifacts_host && chmod -R a+rw /artifacts_host'"
           - "bash .buildkite/scripts/upload-nightly-wheels.sh"
@@ -29,12 +37,10 @@ steps:
         agents:
           queue: arm64_cpu_queue_release
         commands:
-          # #NOTE: torch_cuda_arch_list is derived from upstream PyTorch build files here:
-          # https://github.com/pytorch/pytorch/blob/main/.ci/aarch64_linux/aarch64_ci_build.sh#L7
-          - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=13.0.1 --build-arg torch_cuda_arch_list='8.7 8.9 9.0 10.0+PTX 12.0' --build-arg BUILD_BASE_IMAGE=nvidia/cuda:13.0.1-devel-ubuntu22.04  --tag vllm-ci:build-image --target build --progress plain -f docker/Dockerfile ."
+          - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=13.0.2 --build-arg torch_cuda_arch_list=\"${CUDA_ARCH_AARCH64}\" --build-arg BUILD_BASE_IMAGE=nvidia/cuda:13.0.2-devel-ubuntu22.04  --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 manylinux_2_35"
+          - "bash .buildkite/scripts/upload-nightly-wheels.sh"
         env:
           DOCKER_BUILDKIT: "1"
 
@@ -47,7 +53,7 @@ steps:
           - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg GIT_REPO_CHECK=1 --build-arg VLLM_BUILD_ACL=ON --tag vllm-ci:build-image --target vllm-build --progress plain -f docker/Dockerfile.cpu ."
           - "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 manylinux_2_35"
+          - "bash .buildkite/scripts/upload-nightly-wheels.sh"
         env:
           DOCKER_BUILDKIT: "1"
 
@@ -57,10 +63,10 @@ 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 --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}\" --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 manylinux_2_31"
+          - "bash .buildkite/scripts/upload-nightly-wheels.sh"
         env:
           DOCKER_BUILDKIT: "1"
 
@@ -70,10 +76,10 @@ 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=13.0.1 --build-arg BUILD_BASE_IMAGE=nvidia/cuda:13.0.1-devel-ubuntu22.04 --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=13.0.2 --build-arg torch_cuda_arch_list=\"${CUDA_ARCH_X86}\" --build-arg BUILD_BASE_IMAGE=nvidia/cuda:13.0.2-devel-ubuntu22.04 --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 manylinux_2_35"
+          - "bash .buildkite/scripts/upload-nightly-wheels.sh"
         env:
           DOCKER_BUILDKIT: "1"
 
@@ -86,7 +92,7 @@ steps:
           - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg GIT_REPO_CHECK=1 --build-arg VLLM_CPU_X86=true --tag vllm-ci:build-image --target vllm-build --progress plain -f docker/Dockerfile.cpu ."
           - "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 manylinux_2_35"
+          - "bash .buildkite/scripts/upload-nightly-wheels.sh"
         env:
           DOCKER_BUILDKIT: "1"
 
@@ -108,96 +114,195 @@ steps:
     depends_on: block-build-release-images
     allow_dependency_failure: true
     steps:
-      - label: "Build release image - x86_64 - CUDA 12.9"
+      - label: "Build release image - x86_64 - CUDA 13.0"
         depends_on: ~
         id: build-release-image-x86
         agents:
           queue: cpu_queue_release
         commands:
           - "aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin public.ecr.aws/q9t5s3a7"
-          - "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 FLASHINFER_AOT_COMPILE=true --build-arg INSTALL_KV_CONNECTORS=true --tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m) --target vllm-openai --progress plain -f docker/Dockerfile ."
+          - |
+            DOCKER_BUILDKIT=1 docker build \
+              $(bash .buildkite/scripts/docker-build-metadata-args.sh) \
+              --build-arg max_jobs=16 \
+              --build-arg USE_SCCACHE=1 \
+              --build-arg GIT_REPO_CHECK=1 \
+              --build-arg CUDA_VERSION=13.0.2 \
+              --build-arg torch_cuda_arch_list="${CUDA_ARCH_X86}" \
+              --build-arg INSTALL_KV_CONNECTORS=true \
+              --build-arg BUILD_BASE_IMAGE=nvidia/cuda:13.0.2-devel-ubuntu22.04 \
+              --target vllm-openai \
+              --progress plain \
+              -f docker/Dockerfile .
           - "docker push public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)"
           # re-tag to default image tag and push, just in case arm64 build fails
           - "docker tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m) public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT"
           - "docker push public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT"
 
-      - label: "Build release image - aarch64 - CUDA 12.9"
+      - label: "Build release image - aarch64 - CUDA 13.0"
         depends_on: ~
         id: build-release-image-arm64
         agents:
           queue: arm64_cpu_queue_release
         commands:
           - "aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin public.ecr.aws/q9t5s3a7"
-          - "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 FLASHINFER_AOT_COMPILE=true --build-arg torch_cuda_arch_list='8.7 8.9 9.0 10.0+PTX 12.0' --build-arg INSTALL_KV_CONNECTORS=true --tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m) --target vllm-openai --progress plain -f docker/Dockerfile ."
+          - |
+            DOCKER_BUILDKIT=1 docker build \
+              $(bash .buildkite/scripts/docker-build-metadata-args.sh) \
+              --build-arg max_jobs=16 \
+              --build-arg USE_SCCACHE=1 \
+              --build-arg GIT_REPO_CHECK=1 \
+              --build-arg CUDA_VERSION=13.0.2 \
+              --build-arg torch_cuda_arch_list="${CUDA_ARCH_AARCH64}" \
+              --build-arg INSTALL_KV_CONNECTORS=true \
+              --build-arg BUILD_BASE_IMAGE=nvidia/cuda:13.0.2-devel-ubuntu22.04 \
+              --target vllm-openai \
+              --progress plain \
+              -f docker/Dockerfile .
           - "docker push public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)"
 
-      - label: "Build release image - x86_64 - CUDA 13.0"
+      - label: "Build release image - x86_64 - CUDA 12.9"
         depends_on: ~
-        id: build-release-image-x86-cuda-13-0
+        id: build-release-image-x86-cuda-12-9
         agents:
           queue: cpu_queue_release
         commands:
           - "aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin public.ecr.aws/q9t5s3a7"
-          - "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=13.0.1 --build-arg INSTALL_KV_CONNECTORS=true --build-arg BUILD_BASE_IMAGE=nvidia/cuda:13.0.1-devel-ubuntu22.04 --tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)-cu130 --target vllm-openai --progress plain -f docker/Dockerfile ."
-          - "docker push public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)-cu130"
+          - |
+            DOCKER_BUILDKIT=1 docker build \
+              $(bash .buildkite/scripts/docker-build-metadata-args.sh cu129) \
+              --build-arg max_jobs=16 \
+              --build-arg USE_SCCACHE=1 \
+              --build-arg GIT_REPO_CHECK=1 \
+              --build-arg CUDA_VERSION=12.9.1 \
+              --build-arg torch_cuda_arch_list="${CUDA_ARCH_X86_CU129}" \
+              --build-arg INSTALL_KV_CONNECTORS=true \
+              --target vllm-openai \
+              --progress plain \
+              -f docker/Dockerfile .
+          - "docker push public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)-cu129"
           # re-tag to default image tag and push, just in case arm64 build fails
-          - "docker tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)-cu130 public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-cu130"
-          - "docker push public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-cu130"
+          - "docker tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)-cu129 public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-cu129"
+          - "docker push public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-cu129"
 
-      - label: "Build release image - aarch64 - CUDA 13.0"
+      - label: "Build release image - aarch64 - CUDA 12.9"
         depends_on: ~
-        id: build-release-image-arm64-cuda-13-0
+        id: build-release-image-arm64-cuda-12-9
         agents:
           queue: arm64_cpu_queue_release
         commands:
           - "aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin public.ecr.aws/q9t5s3a7"
-          # compute capability 12.0 for RTX-50 series / RTX PRO 6000 Blackwell, 12.1 for DGX Spark
-          - "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=13.0.1 --build-arg torch_cuda_arch_list='8.7 8.9 9.0 10.0+PTX 12.0 12.1' --build-arg INSTALL_KV_CONNECTORS=true --build-arg BUILD_BASE_IMAGE=nvidia/cuda:13.0.1-devel-ubuntu22.04 --tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)-cu130 --target vllm-openai --progress plain -f docker/Dockerfile ."
-          - "docker push public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)-cu130"
+          - |
+            DOCKER_BUILDKIT=1 docker build \
+              $(bash .buildkite/scripts/docker-build-metadata-args.sh cu129) \
+              --build-arg max_jobs=16 \
+              --build-arg USE_SCCACHE=1 \
+              --build-arg GIT_REPO_CHECK=1 \
+              --build-arg CUDA_VERSION=12.9.1 \
+              --build-arg torch_cuda_arch_list="${CUDA_ARCH_AARCH64_CU129}" \
+              --build-arg INSTALL_KV_CONNECTORS=true \
+              --target vllm-openai \
+              --progress plain \
+              -f docker/Dockerfile .
+          - "docker push public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)-cu129"
 
-      - label: "Build release image - x86_64 - CUDA 12.9 - Ubuntu 24.04"
+      - label: "Build release image - x86_64 - CUDA 13.0 - Ubuntu 24.04"
         depends_on: ~
         id: build-release-image-x86-ubuntu2404
         agents:
           queue: cpu_queue_release
         commands:
           - "aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin public.ecr.aws/q9t5s3a7"
-          - "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 UBUNTU_VERSION=24.04 --build-arg GDRCOPY_OS_VERSION=Ubuntu24_04 --build-arg FLASHINFER_AOT_COMPILE=true --build-arg torch_cuda_arch_list='8.7 8.9 9.0 10.0+PTX 12.0' --build-arg INSTALL_KV_CONNECTORS=true --tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)-ubuntu2404 --target vllm-openai --progress plain -f docker/Dockerfile ."
+          - |
+            DOCKER_BUILDKIT=1 docker build \
+              $(bash .buildkite/scripts/docker-build-metadata-args.sh ubuntu2404) \
+              --build-arg max_jobs=16 \
+              --build-arg USE_SCCACHE=1 \
+              --build-arg GIT_REPO_CHECK=1 \
+              --build-arg CUDA_VERSION=13.0.2 \
+              --build-arg UBUNTU_VERSION=24.04 \
+              --build-arg GDRCOPY_OS_VERSION=Ubuntu24_04 \
+              --build-arg torch_cuda_arch_list="${CUDA_ARCH_X86}" \
+              --build-arg INSTALL_KV_CONNECTORS=true \
+              --build-arg BUILD_BASE_IMAGE=nvidia/cuda:13.0.2-devel-ubuntu24.04 \
+              --target vllm-openai \
+              --progress plain \
+              -f docker/Dockerfile .
           - "docker push public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)-ubuntu2404"
           - "docker tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)-ubuntu2404 public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-ubuntu2404"
           - "docker push public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-ubuntu2404"
 
-      - label: "Build release image - aarch64 - CUDA 12.9 - Ubuntu 24.04"
+      - label: "Build release image - aarch64 - CUDA 13.0 - Ubuntu 24.04"
         depends_on: ~
         id: build-release-image-arm64-ubuntu2404
         agents:
           queue: arm64_cpu_queue_release
         commands:
           - "aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin public.ecr.aws/q9t5s3a7"
-          - "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 UBUNTU_VERSION=24.04 --build-arg GDRCOPY_OS_VERSION=Ubuntu24_04 --build-arg FLASHINFER_AOT_COMPILE=true --build-arg torch_cuda_arch_list='8.7 8.9 9.0 10.0+PTX 12.0' --build-arg INSTALL_KV_CONNECTORS=true --tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)-ubuntu2404 --target vllm-openai --progress plain -f docker/Dockerfile ."
+          - |
+            DOCKER_BUILDKIT=1 docker build \
+              $(bash .buildkite/scripts/docker-build-metadata-args.sh ubuntu2404) \
+              --build-arg max_jobs=16 \
+              --build-arg USE_SCCACHE=1 \
+              --build-arg GIT_REPO_CHECK=1 \
+              --build-arg CUDA_VERSION=13.0.2 \
+              --build-arg UBUNTU_VERSION=24.04 \
+              --build-arg GDRCOPY_OS_VERSION=Ubuntu24_04 \
+              --build-arg torch_cuda_arch_list="${CUDA_ARCH_AARCH64}" \
+              --build-arg INSTALL_KV_CONNECTORS=true \
+              --build-arg BUILD_BASE_IMAGE=nvidia/cuda:13.0.2-devel-ubuntu24.04 \
+              --target vllm-openai \
+              --progress plain \
+              -f docker/Dockerfile .
           - "docker push public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)-ubuntu2404"
 
-      - label: "Build release image - x86_64 - CUDA 13.0 - Ubuntu 24.04"
+      - label: "Build release image - x86_64 - CUDA 12.9 - Ubuntu 24.04"
         depends_on: ~
-        id: build-release-image-x86-cuda-13-0-ubuntu2404
+        id: build-release-image-x86-cuda-12-9-ubuntu2404
         agents:
           queue: cpu_queue_release
         commands:
           - "aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin public.ecr.aws/q9t5s3a7"
-          - "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=13.0.1 --build-arg UBUNTU_VERSION=24.04 --build-arg GDRCOPY_OS_VERSION=Ubuntu24_04 --build-arg FLASHINFER_AOT_COMPILE=true --build-arg torch_cuda_arch_list='8.7 8.9 9.0 10.0+PTX 12.0 12.1' --build-arg INSTALL_KV_CONNECTORS=true --build-arg BUILD_BASE_IMAGE=nvidia/cuda:13.0.1-devel-ubuntu24.04 --tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)-cu130-ubuntu2404 --target vllm-openai --progress plain -f docker/Dockerfile ."
-          - "docker push public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)-cu130-ubuntu2404"
-          - "docker tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)-cu130-ubuntu2404 public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-cu130-ubuntu2404"
-          - "docker push public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-cu130-ubuntu2404"
+          - |
+            DOCKER_BUILDKIT=1 docker build \
+              $(bash .buildkite/scripts/docker-build-metadata-args.sh cu129-ubuntu2404) \
+              --build-arg max_jobs=16 \
+              --build-arg USE_SCCACHE=1 \
+              --build-arg GIT_REPO_CHECK=1 \
+              --build-arg CUDA_VERSION=12.9.1 \
+              --build-arg UBUNTU_VERSION=24.04 \
+              --build-arg GDRCOPY_OS_VERSION=Ubuntu24_04 \
+              --build-arg torch_cuda_arch_list="${CUDA_ARCH_X86_CU129}" \
+              --build-arg INSTALL_KV_CONNECTORS=true \
+              --target vllm-openai \
+              --progress plain \
+              -f docker/Dockerfile .
+          - "docker push public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)-cu129-ubuntu2404"
+          - "docker tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)-cu129-ubuntu2404 public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-cu129-ubuntu2404"
+          - "docker push public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-cu129-ubuntu2404"
 
-      - label: "Build release image - aarch64 - CUDA 13.0 - Ubuntu 24.04"
+      - label: "Build release image - aarch64 - CUDA 12.9 - Ubuntu 24.04"
         depends_on: ~
-        id: build-release-image-arm64-cuda-13-0-ubuntu2404
+        id: build-release-image-arm64-cuda-12-9-ubuntu2404
         agents:
           queue: arm64_cpu_queue_release
         commands:
           - "aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin public.ecr.aws/q9t5s3a7"
-          - "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=13.0.1 --build-arg UBUNTU_VERSION=24.04 --build-arg GDRCOPY_OS_VERSION=Ubuntu24_04 --build-arg FLASHINFER_AOT_COMPILE=true --build-arg torch_cuda_arch_list='8.7 8.9 9.0 10.0+PTX 12.0 12.1' --build-arg INSTALL_KV_CONNECTORS=true --build-arg BUILD_BASE_IMAGE=nvidia/cuda:13.0.1-devel-ubuntu24.04 --tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)-cu130-ubuntu2404 --target vllm-openai --progress plain -f docker/Dockerfile ."
-          - "docker push public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)-cu130-ubuntu2404"
+          - |
+            DOCKER_BUILDKIT=1 docker build \
+              $(bash .buildkite/scripts/docker-build-metadata-args.sh cu129-ubuntu2404) \
+              --build-arg max_jobs=16 \
+              --build-arg USE_SCCACHE=1 \
+              --build-arg GIT_REPO_CHECK=1 \
+              --build-arg CUDA_VERSION=12.9.1 \
+              --build-arg UBUNTU_VERSION=24.04 \
+              --build-arg GDRCOPY_OS_VERSION=Ubuntu24_04 \
+              --build-arg torch_cuda_arch_list="${CUDA_ARCH_AARCH64_CU129}" \
+              --build-arg INSTALL_KV_CONNECTORS=true \
+              --target vllm-openai \
+              --progress plain \
+              -f docker/Dockerfile .
+          - "docker push public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)-cu129-ubuntu2404"
 
       - block: "Build release image for x86_64 CPU"
         key: block-cpu-release-image-build
@@ -238,7 +343,7 @@ steps:
   - group: "Publish release images"
     key: "publish-release-images"
     steps:
-      - label: "Create multi-arch manifest - CUDA 12.9"
+      - label: "Create multi-arch manifest - CUDA 13.0"
         depends_on:
           - build-release-image-x86
           - build-release-image-arm64
@@ -250,7 +355,7 @@ steps:
           - "docker manifest create public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-x86_64 public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-aarch64 --amend"
           - "docker manifest push public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT"
 
-      - label: "Annotate release workflow - CUDA 12.9"
+      - label: "Annotate release workflow - CUDA 13.0"
         depends_on:
           - create-multi-arch-manifest
         id: annotate-release-workflow
@@ -259,19 +364,19 @@ steps:
         commands:
           - "bash .buildkite/scripts/annotate-release.sh"
 
-      - label: "Create multi-arch manifest - CUDA 13.0"
+      - label: "Create multi-arch manifest - CUDA 12.9"
         depends_on:
-          - build-release-image-x86-cuda-13-0
-          - build-release-image-arm64-cuda-13-0
-        id: create-multi-arch-manifest-cuda-13-0
+          - build-release-image-x86-cuda-12-9
+          - build-release-image-arm64-cuda-12-9
+        id: create-multi-arch-manifest-cuda-12-9
         agents:
           queue: small_cpu_queue_release
         commands:
           - "aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin public.ecr.aws/q9t5s3a7"
-          - "docker manifest create public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-cu130 public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-x86_64-cu130 public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-aarch64-cu130 --amend"
-          - "docker manifest push public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-cu130"
+          - "docker manifest create public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-cu129 public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-x86_64-cu129 public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-aarch64-cu129 --amend"
+          - "docker manifest push public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-cu129"
 
-      - label: "Create multi-arch manifest - CUDA 12.9 - Ubuntu 24.04"
+      - label: "Create multi-arch manifest - CUDA 13.0 - Ubuntu 24.04"
         depends_on:
           - build-release-image-x86-ubuntu2404
           - build-release-image-arm64-ubuntu2404
@@ -283,17 +388,17 @@ steps:
           - "docker manifest create public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-ubuntu2404 public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-x86_64-ubuntu2404 public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-aarch64-ubuntu2404 --amend"
           - "docker manifest push public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-ubuntu2404"
 
-      - label: "Create multi-arch manifest - CUDA 13.0 - Ubuntu 24.04"
+      - label: "Create multi-arch manifest - CUDA 12.9 - Ubuntu 24.04"
         depends_on:
-          - build-release-image-x86-cuda-13-0-ubuntu2404
-          - build-release-image-arm64-cuda-13-0-ubuntu2404
-        id: create-multi-arch-manifest-cuda-13-0-ubuntu2404
+          - build-release-image-x86-cuda-12-9-ubuntu2404
+          - build-release-image-arm64-cuda-12-9-ubuntu2404
+        id: create-multi-arch-manifest-cuda-12-9-ubuntu2404
         agents:
           queue: small_cpu_queue_release
         commands:
           - "aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin public.ecr.aws/q9t5s3a7"
-          - "docker manifest create public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-cu130-ubuntu2404 public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-x86_64-cu130-ubuntu2404 public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-aarch64-cu130-ubuntu2404 --amend"
-          - "docker manifest push public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-cu130-ubuntu2404"
+          - "docker manifest create public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-cu129-ubuntu2404 public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-x86_64-cu129-ubuntu2404 public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-aarch64-cu129-ubuntu2404 --amend"
+          - "docker manifest push public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-cu129-ubuntu2404"
 
       - label: "Publish nightly multi-arch image to DockerHub"
         depends_on:
@@ -313,16 +418,16 @@ steps:
           DOCKER_BUILDKIT: "1"
           DOCKERHUB_USERNAME: "vllmbot"
 
-      - label: "Publish nightly multi-arch image to DockerHub - CUDA 13.0"
+      - label: "Publish nightly multi-arch image to DockerHub - CUDA 12.9"
         depends_on:
-          - create-multi-arch-manifest-cuda-13-0
+          - create-multi-arch-manifest-cuda-12-9
         if: build.env("NIGHTLY") == "1"
         agents:
           queue: small_cpu_queue_release
         commands:
-          - "bash .buildkite/scripts/push-nightly-builds.sh cu130"
+          - "bash .buildkite/scripts/push-nightly-builds.sh cu129"
           # Clean up old nightly builds (keep only last 14)
-          - "bash .buildkite/scripts/cleanup-nightly-builds.sh cu130-nightly-"
+          - "bash .buildkite/scripts/cleanup-nightly-builds.sh cu129-nightly-"
         plugins:
           - docker-login#v3.0.0:
               username: vllmbot

.buildkite/scripts/annotate-release.sh

@@ -13,12 +13,12 @@ ROCM_BASE_CACHE_KEY=$(.buildkite/scripts/cache-rocm-base-wheels.sh key)
 buildkite-agent annotate --style 'info' --context 'release-workflow' << EOF
 To download the wheel (by commit):
 \`\`\`
-aws s3 cp s3://vllm-wheels/${BUILDKITE_COMMIT}/vllm-${RELEASE_VERSION}-cp38-abi3-manylinux_2_31_x86_64.whl .
-aws s3 cp s3://vllm-wheels/${BUILDKITE_COMMIT}/vllm-${RELEASE_VERSION}-cp38-abi3-manylinux_2_31_aarch64.whl .
+aws s3 cp s3://vllm-wheels/${BUILDKITE_COMMIT}/vllm-${RELEASE_VERSION}-cp38-abi3-manylinux_2_35_x86_64.whl .
+aws s3 cp s3://vllm-wheels/${BUILDKITE_COMMIT}/vllm-${RELEASE_VERSION}-cp38-abi3-manylinux_2_35_aarch64.whl .
 
-(Optional) For CUDA 13.0:
-aws s3 cp s3://vllm-wheels/${BUILDKITE_COMMIT}/vllm-${RELEASE_VERSION}+cu130-cp38-abi3-manylinux_2_35_x86_64.whl .
-aws s3 cp s3://vllm-wheels/${BUILDKITE_COMMIT}/vllm-${RELEASE_VERSION}+cu130-cp38-abi3-manylinux_2_35_aarch64.whl .
+(Optional) For CUDA 12.9:
+aws s3 cp s3://vllm-wheels/${BUILDKITE_COMMIT}/vllm-${RELEASE_VERSION}+cu129-cp38-abi3-manylinux_2_31_x86_64.whl .
+aws s3 cp s3://vllm-wheels/${BUILDKITE_COMMIT}/vllm-${RELEASE_VERSION}+cu129-cp38-abi3-manylinux_2_31_aarch64.whl .
 
 (Optional) For CPU:
 aws s3 cp s3://vllm-wheels/${BUILDKITE_COMMIT}/vllm-${RELEASE_VERSION}+cpu-cp38-abi3-manylinux_2_35_x86_64.whl .
@@ -33,8 +33,8 @@ To download and upload the image:
 
 docker pull public.ecr.aws/q9t5s3a7/vllm-release-repo:${BUILDKITE_COMMIT}-x86_64
 docker pull public.ecr.aws/q9t5s3a7/vllm-release-repo:${BUILDKITE_COMMIT}-aarch64
-docker pull public.ecr.aws/q9t5s3a7/vllm-release-repo:${BUILDKITE_COMMIT}-x86_64-cu130
-docker pull public.ecr.aws/q9t5s3a7/vllm-release-repo:${BUILDKITE_COMMIT}-aarch64-cu130
+docker pull public.ecr.aws/q9t5s3a7/vllm-release-repo:${BUILDKITE_COMMIT}-x86_64-cu129
+docker pull public.ecr.aws/q9t5s3a7/vllm-release-repo:${BUILDKITE_COMMIT}-aarch64-cu129
 docker pull public.ecr.aws/q9t5s3a7/vllm-release-repo:${ROCM_BASE_CACHE_KEY}-rocm-base
 docker pull public.ecr.aws/q9t5s3a7/vllm-release-repo:${BUILDKITE_COMMIT}-rocm
 docker pull public.ecr.aws/q9t5s3a7/vllm-cpu-release-repo:v${RELEASE_VERSION}
@@ -50,11 +50,11 @@ docker tag vllm/vllm-openai:x86_64 vllm/vllm-openai:v${RELEASE_VERSION}-x86_64
 docker push vllm/vllm-openai:latest-x86_64
 docker push vllm/vllm-openai:v${RELEASE_VERSION}-x86_64
 
-docker tag public.ecr.aws/q9t5s3a7/vllm-release-repo:${BUILDKITE_COMMIT}-x86_64-cu130 vllm/vllm-openai:x86_64-cu130
-docker tag vllm/vllm-openai:x86_64-cu130 vllm/vllm-openai:latest-x86_64-cu130
-docker tag vllm/vllm-openai:x86_64-cu130 vllm/vllm-openai:v${RELEASE_VERSION}-x86_64-cu130
-docker push vllm/vllm-openai:latest-x86_64-cu130
-docker push vllm/vllm-openai:v${RELEASE_VERSION}-x86_64-cu130
+docker tag public.ecr.aws/q9t5s3a7/vllm-release-repo:${BUILDKITE_COMMIT}-x86_64-cu129 vllm/vllm-openai:x86_64-cu129
+docker tag vllm/vllm-openai:x86_64-cu129 vllm/vllm-openai:latest-x86_64-cu129
+docker tag vllm/vllm-openai:x86_64-cu129 vllm/vllm-openai:v${RELEASE_VERSION}-x86_64-cu129
+docker push vllm/vllm-openai:latest-x86_64-cu129
+docker push vllm/vllm-openai:v${RELEASE_VERSION}-x86_64-cu129
 
 docker tag public.ecr.aws/q9t5s3a7/vllm-release-repo:${BUILDKITE_COMMIT}-aarch64 vllm/vllm-openai:aarch64
 docker tag vllm/vllm-openai:aarch64 vllm/vllm-openai:latest-aarch64
@@ -62,11 +62,11 @@ docker tag vllm/vllm-openai:aarch64 vllm/vllm-openai:v${RELEASE_VERSION}-aarch64
 docker push vllm/vllm-openai:latest-aarch64
 docker push vllm/vllm-openai:v${RELEASE_VERSION}-aarch64
 
-docker tag public.ecr.aws/q9t5s3a7/vllm-release-repo:${BUILDKITE_COMMIT}-aarch64-cu130 vllm/vllm-openai:aarch64-cu130
-docker tag vllm/vllm-openai:aarch64-cu130 vllm/vllm-openai:latest-aarch64-cu130
-docker tag vllm/vllm-openai:aarch64-cu130 vllm/vllm-openai:v${RELEASE_VERSION}-aarch64-cu130
-docker push vllm/vllm-openai:latest-aarch64-cu130
-docker push vllm/vllm-openai:v${RELEASE_VERSION}-aarch64-cu130
+docker tag public.ecr.aws/q9t5s3a7/vllm-release-repo:${BUILDKITE_COMMIT}-aarch64-cu129 vllm/vllm-openai:aarch64-cu129
+docker tag vllm/vllm-openai:aarch64-cu129 vllm/vllm-openai:latest-aarch64-cu129
+docker tag vllm/vllm-openai:aarch64-cu129 vllm/vllm-openai:v${RELEASE_VERSION}-aarch64-cu129
+docker push vllm/vllm-openai:latest-aarch64-cu129
+docker push vllm/vllm-openai:v${RELEASE_VERSION}-aarch64-cu129
 
 ## ROCm
 
@@ -104,11 +104,11 @@ docker manifest create vllm/vllm-openai:v${RELEASE_VERSION} vllm/vllm-openai:v${
 docker manifest push vllm/vllm-openai:latest
 docker manifest push vllm/vllm-openai:v${RELEASE_VERSION}
 
-docker manifest rm vllm/vllm-openai:latest-cu130
-docker manifest create vllm/vllm-openai:latest-cu130 vllm/vllm-openai:latest-x86_64-cu130 vllm/vllm-openai:latest-aarch64-cu130
-docker manifest create vllm/vllm-openai:v${RELEASE_VERSION}-cu130 vllm/vllm-openai:v${RELEASE_VERSION}-x86_64-cu130 vllm/vllm-openai:v${RELEASE_VERSION}-aarch64-cu130
-docker manifest push vllm/vllm-openai:latest-cu130
-docker manifest push vllm/vllm-openai:v${RELEASE_VERSION}-cu130
+docker manifest rm vllm/vllm-openai:latest-cu129
+docker manifest create vllm/vllm-openai:latest-cu129 vllm/vllm-openai:latest-x86_64-cu129 vllm/vllm-openai:latest-aarch64-cu129
+docker manifest create vllm/vllm-openai:v${RELEASE_VERSION}-cu129 vllm/vllm-openai:v${RELEASE_VERSION}-x86_64-cu129 vllm/vllm-openai:v${RELEASE_VERSION}-aarch64-cu129
+docker manifest push vllm/vllm-openai:latest-cu129
+docker manifest push vllm/vllm-openai:v${RELEASE_VERSION}-cu129
 
 docker manifest rm vllm/vllm-openai-cpu:latest || true
 docker manifest create vllm/vllm-openai-cpu:latest vllm/vllm-openai-cpu:latest-x86_64 vllm/vllm-openai-cpu:latest-arm64

.buildkite/scripts/check-ray-compatibility.sh

@@ -29,7 +29,7 @@ if python3 -c "import torch; assert torch.version.hip" 2>/dev/null; then
         TORCH_INDEX_URL=""
     fi
 else
-    TORCH_INDEX_URL="https://download.pytorch.org/whl/cu129"
+    TORCH_INDEX_URL="https://download.pytorch.org/whl/cu130"
 fi
 echo ">>> Using PyTorch index: ${TORCH_INDEX_URL:-PyPI default}"
 

.buildkite/scripts/detect-manylinux-tag.py

@@ -0,0 +1,142 @@
+#!/usr/bin/env python3
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""Detect the manylinux platform tag for a wheel and rename it in place.
+
+vLLM's build images produce wheels with the generic ``linux_<arch>`` platform
+tag, which installers like ``pip`` won't accept off PyPI/our index. We need to
+rewrite the platform tag to the appropriate ``manylinux_<major>_<minor>_<arch>``
+before uploading.
+
+Historically the tag was hard-coded per build (``manylinux_2_31`` for the
+Ubuntu 20.04-based image, ``manylinux_2_35`` for the Ubuntu 22.04-based
+images). That is brittle: bumping the base image silently produces wheels
+labelled with the wrong glibc requirement. This script asks ``auditwheel``
+to derive the tag from the symbol versions actually referenced by the
+binaries inside the wheel, so the label tracks reality.
+
+We can't simply call ``auditwheel repair`` -- it tries to graft external
+shared libraries into the wheel and fails on vLLM's CUDA/cuBLAS dependencies.
+Instead we use ``auditwheel.wheel_abi.analyze_wheel_abi`` directly, which is
+the same call that powers ``auditwheel show``, and read off
+``winfo.sym_policy.name``.
+
+Usage:
+    detect-manylinux-tag.py <wheel_path>
+
+The wheel is renamed in place; the new path is printed on stdout. All
+diagnostics go to stderr so callers can capture stdout safely.
+"""
+
+from __future__ import annotations
+
+import argparse
+import sys
+from pathlib import Path
+
+from auditwheel.error import (
+    AuditwheelError,
+    NonPlatformWheelError,
+    WheelToolsError,
+)
+from auditwheel.wheel_abi import analyze_wheel_abi
+from auditwheel.wheeltools import get_wheel_architecture, get_wheel_libc
+
+
+def detect_platform_tag(wheel_path: Path) -> str:
+    """Return the most precise platform tag the wheel is consistent with.
+
+    Mirrors ``auditwheel show`` but returns ``sym_policy`` rather than
+    ``overall_policy``: we only care about the glibc symbol versions used,
+    not about other policy axes (ISA extensions, blacklist, etc.) that
+    ``overall_policy`` folds in.
+    """
+    fn = wheel_path.name
+
+    try:
+        arch = get_wheel_architecture(fn)
+    except (WheelToolsError, NonPlatformWheelError):
+        # Architecture isn't deducible from the filename; let auditwheel
+        # infer it from the ELF binaries inside the wheel.
+        arch = None
+
+    try:
+        libc = get_wheel_libc(fn)
+    except WheelToolsError:
+        # An unrepaired wheel uses ``linux_<arch>``, which doesn't encode
+        # libc. Let auditwheel infer it from the ELF binaries.
+        libc = None
+
+    winfo = analyze_wheel_abi(
+        libc,
+        arch,
+        wheel_path,
+        frozenset(),
+        disable_isa_ext_check=False,
+        allow_graft=False,
+    )
+    return winfo.sym_policy.name
+
+
+def rename_wheel(wheel_path: Path, new_platform_tag: str) -> Path:
+    """Rename the wheel in place, replacing only its platform tag."""
+    # Wheel filename per PEP 427:
+    #   {distribution}-{version}(-{build})?-{python}-{abi}-{platform}.whl
+    # The platform tag is always the last ``-``-separated token before
+    # ``.whl``. Compound tags like ``manylinux_2_31_x86_64`` use ``_`` as the
+    # internal separator, so ``-``-splitting is unambiguous.
+    parts = wheel_path.stem.split("-")
+    if len(parts) < 5:
+        raise ValueError(f"Unrecognised wheel filename: {wheel_path.name}")
+    parts[-1] = new_platform_tag
+    new_path = wheel_path.with_name("-".join(parts) + ".whl")
+    if new_path != wheel_path:
+        wheel_path.rename(new_path)
+    return new_path
+
+
+def main() -> int:
+    parser = argparse.ArgumentParser(
+        description="Detect a wheel's manylinux platform tag with "
+        "auditwheel and rename the wheel in place."
+    )
+    parser.add_argument(
+        "wheel",
+        type=Path,
+        help="Path to the wheel to inspect and rename.",
+    )
+    args = parser.parse_args()
+
+    wheel_path: Path = args.wheel
+    if not wheel_path.is_file():
+        print(f"error: {wheel_path} is not a file", file=sys.stderr)
+        return 1
+
+    # Catch the things that ``analyze_wheel_abi`` and ``rename_wheel`` can
+    # raise: any subclass of ``AuditwheelError`` (pure-Python wheels,
+    # invalid libc, malformed wheels), filesystem errors, or our own
+    # ``ValueError`` for an unrecognised wheel filename. Print a single
+    # ``ERROR_TYPE: message`` line to stderr instead of a Python
+    # traceback, which is much friendlier in CI logs.
+    try:
+        new_tag = detect_platform_tag(wheel_path)
+        print(f"detected platform tag: {new_tag}", file=sys.stderr)
+        new_path = rename_wheel(wheel_path, new_tag)
+    except (AuditwheelError, ValueError, OSError) as e:
+        print(
+            f"error: failed to retag {wheel_path.name}: {type(e).__name__}: {e}",
+            file=sys.stderr,
+        )
+        return 2
+
+    if new_path != wheel_path:
+        print(f"renamed {wheel_path.name} -> {new_path.name}", file=sys.stderr)
+    else:
+        print(f"wheel already tagged {new_tag}", file=sys.stderr)
+
+    print(new_path)
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())

.buildkite/scripts/docker-build-metadata-args.sh

@@ -0,0 +1,54 @@
+#!/bin/bash
+# Emit docker build flags for release image provenance metadata.
+# Keep this helper best-effort: missing Buildkite metadata should fall back to
+# local/default values instead of blocking the Docker build.
+
+# Variant examples: "", "cu129", "ubuntu2404", "cu129-ubuntu2404".
+variant="${1:-}"
+variant_suffix="${variant:+-${variant}}"
+
+image_name="${VLLM_DOCKER_IMAGE_NAME:-vllm/vllm-openai}"
+staging_repo="${VLLM_STAGING_IMAGE_REPO:-public.ecr.aws/q9t5s3a7/vllm-release-repo}"
+build_commit="${VLLM_BUILD_COMMIT:-${BUILDKITE_COMMIT:-unknown}}"
+build_pipeline="${VLLM_BUILD_PIPELINE:-${BUILDKITE_PIPELINE_ID:-${BUILDKITE_PIPELINE_SLUG:-local}}}"
+build_url="${VLLM_BUILD_URL:-${BUILDKITE_BUILD_URL:-}}"
+tag_commit="${BUILDKITE_COMMIT:-${build_commit}}"
+
+if [[ -n "${BUILDKITE:-}" || -n "${BUILDKITE_COMMIT:-}" ]]; then
+  release_version="${RELEASE_VERSION:-}"
+  if command -v buildkite-agent >/dev/null 2>&1; then
+    release_version="${release_version:-$(buildkite-agent meta-data get release-version 2>/dev/null)}"
+  fi
+  release_version="${release_version#v}"
+  release_version="${release_version:-${tag_commit}}"
+
+  staging_image_ref="${staging_repo}:${tag_commit}-$(uname -m)${variant_suffix}"
+
+  if [[ "${NIGHTLY:-}" == "1" ]]; then
+    if [[ -z "${variant}" ]]; then
+      image_tag="${image_name}:nightly-${tag_commit}"
+    elif [[ "${variant}" == cu* ]]; then
+      cuda_variant="${variant%%-*}"
+      remaining_variant="${variant#${cuda_variant}}"
+      image_tag="${image_name}:${cuda_variant}-nightly-${tag_commit}${remaining_variant}"
+    else
+      image_tag="${image_name}:nightly-${tag_commit}${variant_suffix}"
+    fi
+  else
+    image_tag="${image_name}:v${release_version}${variant_suffix}"
+  fi
+else
+  image_tag="${VLLM_IMAGE_TAG:-local/vllm-openai:dev}"
+  staging_image_ref="${image_tag}"
+fi
+
+emit_arg() {
+  printf -- "--build-arg %s=%s " "$1" "$2"
+}
+
+emit_arg VLLM_BUILD_COMMIT "${build_commit}"
+emit_arg VLLM_BUILD_PIPELINE "${build_pipeline}"
+emit_arg VLLM_BUILD_URL "${build_url}"
+# This is the intended public tag. The final digest is only known after push.
+emit_arg VLLM_IMAGE_TAG "${image_tag}"
+printf -- "--tag %s " "${staging_image_ref}"

.buildkite/scripts/generate-and-upload-nightly-index.sh

@@ -9,21 +9,14 @@ set -ex
 
 BUCKET="vllm-wheels"
 INDICES_OUTPUT_DIR="indices"
-DEFAULT_VARIANT_ALIAS="cu129" # align with vLLM_MAIN_CUDA_VERSION in vllm/envs.py
-PYTHON="${PYTHON_PROG:-python3}" # try to read from env var, otherwise use python3
+DEFAULT_VARIANT_ALIAS="cu130" # align with vLLM_MAIN_CUDA_VERSION in vllm/envs.py
 SUBPATH=$BUILDKITE_COMMIT
 S3_COMMIT_PREFIX="s3://$BUCKET/$SUBPATH/"
 
-# detect if python3.12+ is available
-has_new_python=$($PYTHON -c "print(1 if __import__('sys').version_info >= (3,12) else 0)")
-if [[ "$has_new_python" -eq 0 ]]; then
-    # use new python from docker
-    docker pull python:3-slim
-    PYTHON="docker run --rm -u $(id -u):$(id -g) -v $(pwd):/app -w /app python:3-slim python3"
-fi
-
-echo "Using python interpreter: $PYTHON"
-echo "Python version: $($PYTHON --version)"
+# Select python3 (>= 3.12) -- local if available, else a docker fallback.
+# shellcheck source=lib/select-python.sh
+source .buildkite/scripts/lib/select-python.sh
+select_python
 
 # ======== generate and upload indices ========
 

.buildkite/scripts/hardware_ci/run-cpu-distributed-smoke-test.sh

@@ -3,42 +3,37 @@ set -euox pipefail
 export VLLM_CPU_CI_ENV=0
 export VLLM_CPU_KVCACHE_SPACE=1 # avoid OOM
 
-echo "--- PP+TP"
-vllm serve meta-llama/Llama-3.2-3B-Instruct -tp=2 -pp=2 --max-model-len=4096 &
-server_pid=$!
-timeout 600 bash -c "until curl localhost:8000/v1/models > /dev/null 2>&1; do sleep 1; done" || exit 1
-vllm bench serve \
-    --backend vllm \
-    --dataset-name random \
-    --model meta-llama/Llama-3.2-3B-Instruct \
-    --num-prompts 20 \
-    --result-dir ./test_results \
-    --result-filename tp_pp.json \
-    --save-result \
-    --endpoint /v1/completions
-kill -s SIGTERM $server_pid; wait $server_pid || true
-failed_req=$(jq '.failed' ./test_results/tp_pp.json)
-if [ "$failed_req" -ne 0 ]; then
-  echo "Some requests were failed!"
-  exit 1
-fi
+MODE=${1:-all}
 
-echo "--- DP+TP"
-vllm serve meta-llama/Llama-3.2-3B-Instruct -tp=2 -dp=2 --max-model-len=4096 &
-server_pid=$!
-timeout 600 bash -c "until curl localhost:8000/v1/models > /dev/null 2>&1; do sleep 1; done" || exit 1
-vllm bench serve \
-   --backend vllm \
-   --dataset-name random \
-   --model meta-llama/Llama-3.2-3B-Instruct \
-   --num-prompts 20 \
-   --result-dir ./test_results \
-   --result-filename dp_pp.json \
-   --save-result \
-   --endpoint /v1/completions
-kill -s SIGTERM $server_pid; wait $server_pid || true
-failed_req=$(jq '.failed' ./test_results/dp_pp.json)
-if [ "$failed_req" -ne 0 ]; then
- echo "Some requests were failed!"
- exit 1
-fi
+run_scenario() {
+    local label="$1" result_file="$2"
+    shift 2
+    echo "--- $label"
+    vllm serve meta-llama/Llama-3.2-3B-Instruct "$@" --max-model-len=4096 &
+    local server_pid=$!
+    timeout 600 bash -c "until curl localhost:8000/v1/models > /dev/null 2>&1; do sleep 1; done" || exit 1
+    vllm bench serve \
+        --backend vllm \
+        --dataset-name random \
+        --model meta-llama/Llama-3.2-3B-Instruct \
+        --num-prompts 20 \
+        --result-dir ./test_results \
+        --result-filename "$result_file" \
+        --save-result \
+        --endpoint /v1/completions
+    kill -s SIGTERM "$server_pid"; wait "$server_pid" || true
+    if [ "$(jq '.failed' "./test_results/$result_file")" -ne 0 ]; then
+        echo "Some requests were failed in $label!"
+        exit 1
+    fi
+}
+
+case "$MODE" in
+    tp_pp) run_scenario "PP+TP" tp_pp.json -tp=2 -pp=2 ;;
+    dp_tp) run_scenario "DP+TP" dp_tp.json -tp=2 -dp=2 ;;
+    all)
+        run_scenario "PP+TP" tp_pp.json -tp=2 -pp=2
+        run_scenario "DP+TP" dp_tp.json -tp=2 -dp=2
+        ;;
+    *) echo "ERROR: unknown mode '$MODE' (expected: tp_pp | dp_tp | all)" >&2; exit 1 ;;
+esac

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

@@ -25,22 +25,100 @@ export PYTHONPATH=".."
 ###############################################################################
 
 cleanup_docker() {
+  # Share the same lock with image pull to avoid cleanup/pull races on one node.
+  local docker_lock="/tmp/docker-pull.lock"
+  exec 9>"$docker_lock"
+  flock 9
+
   docker_root=$(docker info -f '{{.DockerRootDir}}')
   if [ -z "$docker_root" ]; then
     echo "Failed to determine Docker root directory." >&2
-    exit 1
+    flock -u 9
+    return 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."
+    echo "Disk usage is above $threshold%. Running aggressive CI image cleanup..."
+    cleanup_old_ci_images "${REGISTRY}/${REPO}" "${image_name}" "${DOCKER_IMAGE_CLEANUP_HOURS:-72}" 1
   else
-    echo "Disk usage is below $threshold%. No cleanup needed."
+    echo "Disk usage is below $threshold%. Checking old CI images anyway."
+    cleanup_old_ci_images "${REGISTRY}/${REPO}" "${image_name}" "${DOCKER_IMAGE_CLEANUP_HOURS:-72}" 0
+  fi
+  echo "Old CI image cleanup completed."
+
+  flock -u 9
+}
+
+cleanup_old_ci_images() {
+  local repo_prefix="$1"
+  local current_image_ref="$2"
+  local ttl_hours="$3"
+  local aggressive_cleanup="$4"
+
+  if [[ -z "$repo_prefix" || "$repo_prefix" == "/" ]]; then
+    echo "Skip old-image cleanup: invalid repo prefix '${repo_prefix}'"
+    return 0
+  fi
+
+  if ! [[ "$ttl_hours" =~ ^[0-9]+$ ]]; then
+    echo "Invalid DOCKER_IMAGE_CLEANUP_HOURS='${ttl_hours}', fallback to 72"
+    ttl_hours=72
+  fi
+
+  local now_epoch cutoff_epoch
+  now_epoch=$(date +%s)
+  cutoff_epoch=$((now_epoch - ttl_hours * 3600))
+
+  local -a used_image_ids
+  mapfile -t used_image_ids < <(docker ps -aq | xargs -r docker inspect --format '{{.Image}}' | sort -u)
+
+  local removed_count=0
+  local examined_count=0
+  declare -A seen_ids=()
+
+  while read -r image_ref image_id; do
+    [[ -z "$image_ref" || -z "$image_id" ]] && continue
+    ((examined_count++))
+
+    # Keep the image this job is going to use.
+    if [[ "$image_ref" == "$current_image_ref" ]]; then
+      continue
+    fi
+
+    # Avoid duplicate deletes when multiple tags point to same image id.
+    if [[ -n "${seen_ids[$image_id]:-}" ]]; then
+      continue
+    fi
+    seen_ids[$image_id]=1
+
+    # Never delete images that are used by any container on this node.
+    if printf '%s\n' "${used_image_ids[@]}" | grep -qx "$image_id"; then
+      continue
+    fi
+
+    local created created_epoch
+    created=$(docker image inspect -f '{{.Created}}' "$image_id" 2>/dev/null || true)
+    [[ -z "$created" ]] && continue
+    created_epoch=$(date -d "$created" +%s 2>/dev/null || true)
+    [[ -z "$created_epoch" ]] && continue
+
+    if (( created_epoch < cutoff_epoch )) || [[ "$aggressive_cleanup" == "1" ]]; then
+      if docker image rm -f "$image_id" >/dev/null 2>&1; then
+        ((removed_count++))
+      fi
+    fi
+  done < <(docker image ls --no-trunc "$repo_prefix" --format '{{.Repository}}:{{.Tag}} {{.ID}}')
+
+  # Also trim old dangling layers; this is safe and does not remove referenced images.
+  docker image prune -f --filter "until=${ttl_hours}h" >/dev/null 2>&1 || true
+
+  if [[ "$aggressive_cleanup" == "1" ]]; then
+    echo "Examined ${examined_count} images under ${repo_prefix}, removed ${removed_count} unused images under disk pressure."
+  else
+    echo "Examined ${examined_count} images under ${repo_prefix}, removed ${removed_count} old images (>${ttl_hours}h)."
   fi
 }
 
@@ -240,7 +318,6 @@ fi
 cleanup_docker
 
 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
 
 # --- Build or pull test image ---
 IMAGE="${IMAGE_TAG_XPU:-${image_name}}"
@@ -266,8 +343,6 @@ fi
 
 remove_docker_container() {
   docker rm -f "${container_name}" || true
-  docker image rm -f "${image_name}" || true
-  docker system prune -f || true
 }
 trap remove_docker_container EXIT
 
@@ -283,6 +358,7 @@ docker run \
     --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:-}" \

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

@@ -12,9 +12,7 @@ docker build -t "${image_name}" -f docker/Dockerfile.xpu .
 
 # Setup cleanup
 remove_docker_container() {
-  docker rm -f "${container_name}" || true;
-  docker image rm -f "${image_name}" || true;
-  docker system prune -f || true;
+  docker rm -f "${container_name}" || true
 }
 trap remove_docker_container EXIT
 

.buildkite/scripts/lib/manylinux.sh

@@ -0,0 +1,127 @@
+#!/usr/bin/env bash
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+#
+# Shared helper for rewriting a wheel's platform tag from the generic
+# ``linux_<arch>`` to the correct ``manylinux_<major>_<minor>_<arch>``.
+# After sourcing, call ``apply_manylinux_tag <wheel>`` on each wheel
+# that still carries the generic tag; the renamed path is printed on
+# stdout (logs go to stderr).
+#
+# Why a pinned Docker container instead of using whatever Python
+# happens to be on the agent:
+#   - vLLM's release agents are heterogeneous -- they don't agree on
+#     a Python minor version, and we can't rely on a particular
+#     ``auditwheel`` being installed.
+#   - ``detect-manylinux-tag.py`` reads ``auditwheel.wheel_abi`` and
+#     ``Policy.sym_policy``, which are *internal* APIs without a
+#     stability promise. Pinning both Python and auditwheel makes the
+#     detected tag a function of the inputs alone, and shifts version
+#     bumps from "implicit drift" to "deliberate, retested change".
+#   - Other release scripts (``generate-and-upload-nightly-index.sh``,
+#     ``upload-rocm-wheels.sh``) already use the python:3-slim image
+#     when the agent's interpreter is too old; this is the same idea
+#     made stricter.
+#
+# To keep the per-wheel cost down (the ROCm upload retags ~10 wheels
+# each run), we install auditwheel into a long-lived helper container
+# once on source, then ``docker exec`` into it for each call.
+#
+# Trap behaviour:
+# - Sourcing installs an EXIT trap that calls ``manylinux_cleanup`` to
+#   tear down the helper container. Any EXIT trap that was already in
+#   place when this file was sourced is captured and run AFTER our
+#   cleanup, so we don't silently clobber it.
+# - If a caller sets a new EXIT trap *after* sourcing, that trap will
+#   replace ours; in that case the caller should call
+#   ``manylinux_cleanup`` from their own handler.
+
+if [[ -n "${_MANYLINUX_LIB_SOURCED:-}" ]]; then
+    return 0
+fi
+_MANYLINUX_LIB_SOURCED=1
+
+# Pin both sides. Bump these deliberately and re-run a representative
+# wheel from each build target through the detection.
+_MANYLINUX_PYTHON_IMAGE="python:3.12-slim"
+_MANYLINUX_AUDITWHEEL_VERSION="6.6.0"
+
+# Resolve our own directory (and the sibling detect script) using the
+# canonical, symlink-resolved path. The container mounts cwd at the
+# same absolute path on both sides, so all paths we hand to it -- the
+# script, the wheel -- must canonicalise to a location under cwd.
+_MANYLINUX_LIB_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd -P)"
+_MANYLINUX_DETECT_SCRIPT="$(cd "${_MANYLINUX_LIB_DIR}/.." && pwd -P)/detect-manylinux-tag.py"
+_MANYLINUX_CWD="$(pwd -P)"
+
+docker pull --quiet "$_MANYLINUX_PYTHON_IMAGE" >/dev/null
+
+# Spin up a long-lived helper container so we install auditwheel once
+# and then ``docker exec`` into it for each wheel.
+#
+# The container runs as root so ``pip install`` can write into the
+# system site-packages; individual ``docker exec`` calls below pin
+# themselves to the host UID so any file rename happens with host
+# ownership, not root.
+_MANYLINUX_CONTAINER="$(docker run -d --rm \
+    -v "$_MANYLINUX_CWD:$_MANYLINUX_CWD" \
+    -w "$_MANYLINUX_CWD" \
+    "$_MANYLINUX_PYTHON_IMAGE" \
+    sleep infinity)"
+docker exec "$_MANYLINUX_CONTAINER" \
+    pip install --quiet --disable-pip-version-check \
+    --root-user-action=ignore \
+    "auditwheel==${_MANYLINUX_AUDITWHEEL_VERSION}"
+
+# Public cleanup -- safe to call multiple times.
+manylinux_cleanup() {
+    if [[ -n "${_MANYLINUX_CONTAINER:-}" ]]; then
+        docker rm -f "$_MANYLINUX_CONTAINER" >/dev/null 2>&1 || true
+        _MANYLINUX_CONTAINER=""
+    fi
+}
+
+# Capture any EXIT trap that was already in place so we can chain to
+# it rather than overwrite it. ``trap -p EXIT`` prints the handler in
+# eval-able form (``trap -- 'CMD' EXIT``) or nothing if unset; we
+# strip the wrapper to recover ``CMD``. Handles the common case --
+# CMDs without embedded single quotes -- and degrades gracefully (we
+# still run our own cleanup) for the pathological case.
+_manylinux_prev_exit_trap_cmd=""
+_manylinux_existing_exit_trap="$(trap -p EXIT)"
+if [[ -n "$_manylinux_existing_exit_trap" ]]; then
+    _tmp="${_manylinux_existing_exit_trap#trap -- \'}"
+    _manylinux_prev_exit_trap_cmd="${_tmp%\' EXIT}"
+    unset _tmp
+fi
+unset _manylinux_existing_exit_trap
+
+_manylinux_run_exit_chain() {
+    manylinux_cleanup
+    if [[ -n "$_manylinux_prev_exit_trap_cmd" ]]; then
+        eval "$_manylinux_prev_exit_trap_cmd"
+    fi
+}
+trap _manylinux_run_exit_chain EXIT
+
+# Detect the manylinux platform tag for a single wheel and rename it
+# in place, printing the renamed wheel path on stdout. Returns
+# non-zero on failure (which under ``set -e`` propagates to caller).
+#
+# The wheel must be reachable via a path under the host cwd so it's
+# visible inside the helper container; in CI the wheels always live
+# under ``artifacts/`` so this is fine.
+apply_manylinux_tag() {
+    local wheel="$1"
+    local abs_wheel
+    abs_wheel="$(realpath "$wheel")"
+    local new_wheel
+    new_wheel="$(docker exec -u "$(id -u):$(id -g)" \
+        "$_MANYLINUX_CONTAINER" \
+        python "$_MANYLINUX_DETECT_SCRIPT" "$abs_wheel")"
+    if [[ -z "$new_wheel" || ! -f "$new_wheel" ]]; then
+        echo "apply_manylinux_tag: detect-manylinux-tag.py did not produce a valid wheel path for $wheel" >&2
+        return 1
+    fi
+    printf '%s\n' "$new_wheel"
+}

.buildkite/scripts/lib/select-python.sh

@@ -0,0 +1,41 @@
+#!/usr/bin/env bash
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+#
+# Pick a Python interpreter for buildkite scripts: prefer a local
+# ``python3`` if it is recent enough (>= 3.12), otherwise fall back to
+# a one-shot Docker container running ``python:3-slim``. After
+# ``select_python`` returns, ``$PYTHON`` is set in the caller's shell
+# and is safe to use as a command (e.g. ``$PYTHON some_script.py``).
+#
+# The 3.12 threshold matches what the existing nightly-index work
+# expects -- typing features used by ``generate-nightly-index.py``.
+# This helper does not pin the *minor* version; if you need stricter
+# reproducibility (e.g. relying on auditwheel internals), invoke
+# Docker yourself with a pinned tag rather than calling this.
+
+if [[ -n "${_SELECT_PYTHON_LIB_SOURCED:-}" ]]; then
+    return 0
+fi
+_SELECT_PYTHON_LIB_SOURCED=1
+
+# Sets ``PYTHON`` in the caller's shell and exports it. Idempotent --
+# calling twice is safe and the second call simply re-runs the probe.
+select_python() {
+    local py="${PYTHON_PROG:-python3}"
+    local has_new_python
+    has_new_python=$("$py" -c \
+        "print(1 if __import__('sys').version_info >= (3,12) else 0)" \
+        2>/dev/null || echo 0)
+    if [[ "$has_new_python" -eq 0 ]]; then
+        # ``-u $(id -u):$(id -g)`` so files created via the container
+        # end up owned by the host user, not root.
+        docker pull python:3-slim
+        PYTHON="docker run --rm -u $(id -u):$(id -g) -v $(pwd):/app -w /app python:3-slim python3"
+    else
+        PYTHON="$py"
+    fi
+    export PYTHON
+    echo "Using python interpreter: $PYTHON"
+    echo "Python version: $($PYTHON --version)"
+}

.buildkite/scripts/scheduled_integration_test/qwen30b_a3b_fp8_dp4_async_eplb.sh

@@ -0,0 +1,55 @@
+#!/usr/bin/env bash
+set -euxo pipefail
+
+# args: [THRESHOLD] [NUM_QUESTIONS] [START_PORT]
+THRESHOLD=${1:-0.8}
+NUM_Q=${2:-1319}
+PORT=${3:-8050}
+OUT_DIR=${OUT_DIR:-/tmp/vllm-scheduled}
+mkdir -p "${OUT_DIR}"
+
+wait_for_server() {
+  local port=$1
+  timeout 600 bash -c '
+    until curl -sf "http://127.0.0.1:'"$port"'/health" > /dev/null; do
+      sleep 1
+    done'
+}
+
+MODEL="Qwen/Qwen3-30B-A3B-FP8"
+BACK="allgather_reducescatter"
+
+cleanup() {
+  if [[ -n "${SERVER_PID:-}" ]] && kill -0 "${SERVER_PID}" 2>/dev/null; then
+    kill "${SERVER_PID}" 2>/dev/null || true
+    for _ in {1..20}; do
+      kill -0 "${SERVER_PID}" 2>/dev/null || break
+      sleep 0.5
+    done
+    kill -9 "${SERVER_PID}" 2>/dev/null || true
+  fi
+}
+trap cleanup EXIT
+
+VLLM_DEEP_GEMM_WARMUP=skip \
+vllm serve "$MODEL" \
+--enforce-eager \
+--data-parallel-size 4 \
+--enable-expert-parallel \
+--enable-eplb \
+--all2all-backend "$BACK" \
+--eplb-config '{"window_size":20, "step_interval":100, "use_async":true}' \
+--trust-remote-code \
+--max-model-len 2048 \
+--port "$PORT" &
+SERVER_PID=$!
+wait_for_server "$PORT"
+
+TAG=$(echo "$MODEL" | tr '/: \\n' '_____')
+OUT="${OUT_DIR}/${TAG}_${BACK}.json"
+python3 tests/evals/gsm8k/gsm8k_eval.py --host http://127.0.0.1 --port "$PORT" --num-questions "${NUM_Q}" --save-results "${OUT}"
+python3 - <<PY
+import json; acc=json.load(open('${OUT}'))['accuracy']
+print(f"${MODEL} ${BACK}: accuracy {acc:.3f}")
+assert acc >= ${THRESHOLD}, f"${MODEL} ${BACK} accuracy {acc}"
+PY

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

@@ -28,6 +28,7 @@
 #   BFCL_MAX_MODEL_LEN  - Max model length (default: 4096)
 #   BFCL_PORT           - Server port (default: 8000)
 #   BFCL_REASONING_PARSER - Reasoning parser name (default: disabled)
+#   BFCL_TEMPERATURE    - Temperature (default: 0.0)
 #   BFCL_EXTRA_ARGS     - Additional vLLM server args
 
 set -euo pipefail
@@ -43,6 +44,7 @@ TP_SIZE="${BFCL_TP_SIZE:-1}"
 MAX_MODEL_LEN="${BFCL_MAX_MODEL_LEN:-4096}"
 PORT="${BFCL_PORT:-8000}"
 REASONING_PARSER="${BFCL_REASONING_PARSER:-}"
+TEMPERATURE="${BFCL_TEMPERATURE:-0.0}"
 EXTRA_ARGS="${BFCL_EXTRA_ARGS:-}"
 
 # Set up output directory
@@ -139,7 +141,7 @@ echo "vLLM server is ready. (started in ${SECONDS_WAITED}s)"
 # be patched in-process so BFCL knows to use the OpenAI-compatible handler
 # against our local vLLM server.
 bfcl_exit_code=0
-python3 - "$MODEL" "$TEST_CATEGORY" "$NUM_THREADS" "$PORT" "$API_TYPE" "$OUTPUT_DIR" << 'PYEOF' || bfcl_exit_code=$?
+python3 - "$MODEL" "$TEST_CATEGORY" "$NUM_THREADS" "$PORT" "$API_TYPE" "$TEMPERATURE" "$OUTPUT_DIR" << 'PYEOF' || bfcl_exit_code=$?
 import os
 import sys
 
@@ -148,7 +150,8 @@ test_category = sys.argv[2]
 num_threads = int(sys.argv[3])
 port = sys.argv[4]
 api_type = sys.argv[5]
-output_dir = sys.argv[6] if len(sys.argv) > 6 and sys.argv[6] else os.getcwd()
+temperature = float(sys.argv[6])
+output_dir = sys.argv[7] if len(sys.argv) > 7 and sys.argv[7] else os.getcwd()
 
 os.environ["OPENAI_BASE_URL"] = f"http://localhost:{port}/v1"
 os.environ["OPENAI_API_KEY"] = "dummy"
@@ -204,6 +207,7 @@ gen_kwargs["model"] = [model]
 gen_kwargs["test_category"] = [c.strip() for c in test_category.split(",")]
 gen_kwargs["skip_server_setup"] = True
 gen_kwargs["num_threads"] = num_threads
+gen_kwargs["temperature"] = temperature
 generate(**gen_kwargs)
 
 # ---- evaluate ----

.buildkite/scripts/upload-nightly-wheels.sh

@@ -2,14 +2,18 @@
 
 set -ex
 
-# Upload a single wheel to S3 (rename linux -> manylinux).
+# Upload a single wheel to S3, after detecting and applying the appropriate
+# manylinux platform tag with auditwheel.
 # Index generation is handled separately by generate-and-upload-nightly-index.sh.
 
+# shellcheck source=lib/manylinux.sh
+source .buildkite/scripts/lib/manylinux.sh
+
 BUCKET="vllm-wheels"
 SUBPATH=$BUILDKITE_COMMIT
 S3_COMMIT_PREFIX="s3://$BUCKET/$SUBPATH/"
 
-# ========= collect, rename & upload the wheel ==========
+# ========= locate the wheel ==========
 
 # Assume wheels are in artifacts/dist/*.whl
 wheel_files=(artifacts/dist/*.whl)
@@ -21,19 +25,9 @@ if [[ ${#wheel_files[@]} -ne 1 ]]; then
 fi
 wheel="${wheel_files[0]}"
 
-# default build image uses ubuntu 20.04, which corresponds to manylinux_2_31
-# we also accept params as manylinux tag
-# refer to https://github.com/mayeut/pep600_compliance?tab=readme-ov-file#acceptable-distros-to-build-wheels
-manylinux_version="${1:-manylinux_2_31}"
+# ========= detect manylinux tag and rename ==========
 
-# Rename 'linux' to the appropriate manylinux version in the wheel filename
-if [[ "$wheel" != *"linux"* ]]; then
-  echo "Error: Wheel filename does not contain 'linux': $wheel"
-  exit 1
-fi
-new_wheel="${wheel/linux/$manylinux_version}"
-mv -- "$wheel" "$new_wheel"
-wheel="$new_wheel"
+wheel="$(apply_manylinux_tag "$wheel")"
 echo "Renamed wheel to: $wheel"
 
 # Extract the version from the wheel

.buildkite/scripts/upload-rocm-wheels.sh

@@ -20,10 +20,6 @@ BUCKET="${S3_BUCKET:-vllm-wheels}"
 ROCM_SUBPATH="rocm/${BUILDKITE_COMMIT}"
 S3_COMMIT_PREFIX="s3://$BUCKET/$ROCM_SUBPATH/"
 INDICES_OUTPUT_DIR="rocm-indices"
-PYTHON="${PYTHON_PROG:-python3}"
-
-# ROCm uses manylinux_2_35 (Ubuntu 22.04 based)
-MANYLINUX_VERSION="manylinux_2_35"
 
 echo "========================================"
 echo "ROCm Wheel Upload Configuration"
@@ -34,19 +30,21 @@ echo "Commit: $BUILDKITE_COMMIT"
 echo "Branch: $BUILDKITE_BRANCH"
 echo "========================================"
 
-# ======== Part 0: Setup Python ========
+# ======== Part 0: Setup Python and helpers ========
 
-# Detect if python3.12+ is available
-has_new_python=$($PYTHON -c "print(1 if __import__('sys').version_info >= (3,12) else 0)" 2>/dev/null || echo 0)
-if [[ "$has_new_python" -eq 0 ]]; then
-    # Use new python from docker
-    # Use --user to ensure files are created with correct ownership (not root)
-    docker pull python:3-slim
-    PYTHON="docker run --rm --user $(id -u):$(id -g) -v $(pwd):/app -w /app python:3-slim python3"
-fi
+# Pick a Python interpreter for index generation -- local if recent
+# enough, else a one-shot docker fallback.
+# shellcheck source=lib/select-python.sh
+source .buildkite/scripts/lib/select-python.sh
+select_python
 
-echo "Using python interpreter: $PYTHON"
-echo "Python version: $($PYTHON --version)"
+# Set up auditwheel-in-a-container for the manylinux retagging step.
+# Distinct from select_python: ``manylinux.sh`` deliberately pins both
+# the Python and auditwheel versions (the script reads auditwheel
+# internals) and so always runs in a known-good container regardless
+# of what's on the agent.
+# shellcheck source=lib/manylinux.sh
+source .buildkite/scripts/lib/manylinux.sh
 
 # ======== Part 1: Collect and prepare wheels ========
 
@@ -63,11 +61,18 @@ if [ "$WHEEL_COUNT" -eq 0 ]; then
     exit 1
 fi
 
-# Rename linux to manylinux in wheel filenames
+# Detect the appropriate manylinux platform tag for any wheel that still
+# carries the generic ``linux_<arch>`` tag, and rename it in place. We use
+# auditwheel via ``apply_manylinux_tag`` (see lib/manylinux.sh) rather than
+# a hard-coded ``manylinux_2_35`` string so that the label tracks the actual
+# glibc symbol versions used by the binaries (and stays correct if the
+# rocm_base image is rebased).
+#
+# The ``linux``/``manylinux`` filter below skips both pre-tagged wheels
+# (e.g. upstream torch) and pure-Python ``-any.whl`` wheels.
 for wheel in all-rocm-wheels/*.whl; do
     if [[ "$wheel" == *"linux"* ]] && [[ "$wheel" != *"manylinux"* ]]; then
-        new_wheel="${wheel/linux/$MANYLINUX_VERSION}"
-        mv -- "$wheel" "$new_wheel"
+        new_wheel="$(apply_manylinux_tag "$wheel")"
         echo "Renamed: $(basename "$wheel") -> $(basename "$new_wheel")"
     fi
 done

.buildkite/test_areas/attention.yaml

@@ -3,6 +3,7 @@ depends_on:
   - image-build
 steps:
 - label: V1 attention (H100)
+  key: v1-attention-h100
   timeout_in_minutes: 30
   device: h100
   source_file_dependencies:
@@ -14,6 +15,7 @@ steps:
     - pytest -v -s v1/attention
 
 - label: V1 attention (B200)
+  key: v1-attention-b200
   timeout_in_minutes: 30
   device: b200
   source_file_dependencies:

.buildkite/test_areas/basic_correctness.yaml

@@ -3,6 +3,7 @@ depends_on:
   - image-build
 steps:
 - label: Basic Correctness
+  key: basic-correctness
   timeout_in_minutes: 30
   device: h200_18gb
   source_file_dependencies:

.buildkite/test_areas/benchmarks.yaml

@@ -3,6 +3,7 @@ depends_on:
   - image-build
 steps:
 - label: Benchmarks CLI Test
+  key: benchmarks-cli-test
   timeout_in_minutes: 20
   device: h200_18gb
   source_file_dependencies:
@@ -12,6 +13,7 @@ steps:
   - pytest -v -s benchmarks/
 
 - label: Attention Benchmarks Smoke Test (B200)
+  key: attention-benchmarks-smoke-test-b200
   device: b200
   num_gpus: 2
   optional: true

.buildkite/test_areas/compile.yaml

@@ -3,6 +3,7 @@ depends_on:
   - image-build
 steps:
 - label: Sequence Parallel Correctness Tests (2 GPUs)
+  key: sequence-parallel-correctness-tests-2-gpus
   timeout_in_minutes: 50
   working_dir: "/vllm-workspace/"
   num_devices: 2
@@ -17,6 +18,7 @@ steps:
   - pytest -v -s tests/compile/correctness_e2e/test_sequence_parallel.py
 
 - label: Sequence Parallel Correctness Tests (2xH100)
+  key: sequence-parallel-correctness-tests-2xh100
   timeout_in_minutes: 50
   working_dir: "/vllm-workspace/"
   device: h100
@@ -27,6 +29,7 @@ steps:
   - pytest -v -s tests/compile/correctness_e2e/test_sequence_parallel.py
 
 - label: AsyncTP Correctness Tests (2xH100)
+  key: asynctp-correctness-tests-2xh100
   timeout_in_minutes: 50
   working_dir: "/vllm-workspace/"
   device: h100
@@ -37,6 +40,7 @@ steps:
   - pytest -v -s tests/compile/correctness_e2e/test_async_tp.py
 
 - label: AsyncTP Correctness Tests (B200)
+  key: asynctp-correctness-tests-b200
   timeout_in_minutes: 50
   working_dir: "/vllm-workspace/"
   device: b200
@@ -47,6 +51,7 @@ steps:
   - pytest -v -s tests/compile/correctness_e2e/test_async_tp.py
 
 - label: Distributed Compile Unit Tests (2xH100)
+  key: distributed-compile-unit-tests-2xh100
   timeout_in_minutes: 20
   working_dir: "/vllm-workspace/"
   device: h100
@@ -60,6 +65,7 @@ steps:
   - pytest -s -v tests/compile/passes/distributed
 
 - label: Fusion and Compile Unit Tests (2xB200)
+  key: fusion-and-compile-unit-tests-2xb200
   timeout_in_minutes: 20
   working_dir: "/vllm-workspace/"
   device: b200
@@ -89,6 +95,7 @@ steps:
     - pytest -v -s tests/compile/fullgraph/test_full_graph.py::test_fp8_kv_scale_compile
 
 - label: Fusion E2E Quick (H100)
+  key: fusion-e2e-quick-h100
   timeout_in_minutes: 15
   working_dir: "/vllm-workspace/"
   device: h100
@@ -107,6 +114,7 @@ steps:
     - pytest -v -s tests/compile/fusions_e2e/test_tp1_quant.py -k "inductor_partition and not +rms_norm and +quant_fp8 and (qwen3 or deepseek)"
 
 - label: Fusion E2E Config Sweep (H100)
+  key: fusion-e2e-config-sweep-h100
   timeout_in_minutes: 30
   working_dir: "/vllm-workspace/"
   device: h100
@@ -126,6 +134,7 @@ steps:
     - pytest -v -s tests/compile/fusions_e2e/test_tp1_quant.py -k "llama-3"
 
 - label: Fusion E2E Config Sweep (B200)
+  key: fusion-e2e-config-sweep-b200
   timeout_in_minutes: 30
   working_dir: "/vllm-workspace/"
   device: b200
@@ -139,6 +148,7 @@ steps:
     - pytest -v -s tests/compile/fusions_e2e/test_tp1_quant.py -k "inductor_partition and (FLASHINFER and not +rms_norm and (not +quant_fp8 or +quant_fp8 and (qwen3 or deepseek)) or llama-3)"
 
 - label: Fusion E2E TP2 Quick (H100)
+  key: fusion-e2e-tp2-quick-h100
   timeout_in_minutes: 20
   working_dir: "/vllm-workspace/"
   device: h100
@@ -156,6 +166,7 @@ steps:
     - pytest -v -s tests/compile/fusions_e2e/test_tp2_async_tp.py -k "inductor_partition and not +rms_norm and (not +quant_fp8 or +quant_fp8 and (qwen3 or deepseek))"
 
 - label: Fusion E2E TP2 AR-RMS Config Sweep (H100)
+  key: fusion-e2e-tp2-ar-rms-config-sweep-h100
   timeout_in_minutes: 40
   working_dir: "/vllm-workspace/"
   device: h100
@@ -175,6 +186,7 @@ steps:
     - pytest -v -s tests/compile/fusions_e2e/test_tp2_ar_rms.py -k "llama-3"
 
 - label: Fusion E2E TP2 AsyncTP Config Sweep (H100)
+  key: fusion-e2e-tp2-asynctp-config-sweep-h100
   timeout_in_minutes: 40
   working_dir: "/vllm-workspace/"
   device: h100
@@ -194,6 +206,7 @@ steps:
     - pytest -v -s tests/compile/fusions_e2e/test_tp2_async_tp.py -k "llama-3"
 
 - label: Fusion E2E TP2 (B200)
+  key: fusion-e2e-tp2-b200
   timeout_in_minutes: 20
   working_dir: "/vllm-workspace/"
   device: b200

.buildkite/test_areas/cuda.yaml

@@ -3,6 +3,7 @@ depends_on:
   - image-build
 steps:
 - label: Platform Tests (CUDA)
+  key: platform-tests-cuda
   timeout_in_minutes: 15
   device: h200_18gb
   source_file_dependencies:
@@ -13,6 +14,7 @@ steps:
     - pytest -v -s cuda/test_platform_no_cuda_init.py
 
 - label: Cudagraph
+  key: cudagraph
   timeout_in_minutes: 20
   source_file_dependencies:
   - tests/v1/cudagraph

.buildkite/test_areas/disaggregated.yaml

@@ -0,0 +1,106 @@
+group: Disaggregated
+depends_on: 
+  - image-build
+steps:
+- label: Distributed NixlConnector PD accuracy (4 GPUs)
+  key: distributed-nixlconnector-pd-accuracy-4-gpus
+  timeout_in_minutes: 30
+  working_dir: "/vllm-workspace/tests"
+  num_devices: 4
+  source_file_dependencies:
+    - vllm/distributed/kv_transfer/kv_connector/v1/nixl_connector.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/config_sweep_accuracy_test.sh
+- label: Distributed FlashInfer NixlConnector PD accuracy (4 GPUs)
+  key: distributed-flashinfer-nixlconnector-pd-accuracy-4-gpus
+  timeout_in_minutes: 30
+  working_dir: "/vllm-workspace/tests"
+  num_devices: 4
+  source_file_dependencies:
+    - vllm/distributed/kv_transfer/kv_connector/v1/nixl_connector.py
+    - tests/v1/kv_connector/nixl_integration/
+  commands:
+    - uv pip install --system -r /vllm-workspace/requirements/kv_connectors.txt
+    - FLASHINFER=1 bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
+
+- label: DP EP Distributed NixlConnector PD accuracy tests (4 GPUs)
+  key: dp-ep-distributed-nixlconnector-pd-accuracy-tests-4-gpus
+  timeout_in_minutes: 30
+  working_dir: "/vllm-workspace/tests"
+  num_devices: 4
+  source_file_dependencies:
+    - vllm/distributed/kv_transfer/kv_connector/v1/nixl_connector.py
+    - tests/v1/kv_connector/nixl_integration/
+  commands:
+    - uv pip install --system -r /vllm-workspace/requirements/kv_connectors.txt
+    - 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)
+  key: crosslayer-kv-layout-distributed-nixlconnector-pd-accuracy-tests-4-gpus
+  timeout_in_minutes: 30
+  working_dir: "/vllm-workspace/tests"
+  num_devices: 4
+  source_file_dependencies:
+    - vllm/distributed/kv_transfer/kv_connector/v1/nixl_connector.py
+    - tests/v1/kv_connector/nixl_integration/
+  commands:
+    - uv pip install --system -r /vllm-workspace/requirements/kv_connectors.txt
+    - CROSS_LAYERS_BLOCKS=True bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
+
+- label: Hybrid SSM NixlConnector PD accuracy tests (4 GPUs)
+  key: hybrid-ssm-nixlconnector-pd-accuracy-tests-4-gpus
+  timeout_in_minutes: 20
+  working_dir: "/vllm-workspace/tests"
+  num_devices: 4
+  source_file_dependencies:
+    - vllm/distributed/kv_transfer/kv_connector/v1/nixl_connector.py
+    - tests/v1/kv_connector/nixl_integration/
+  commands:
+    - uv pip install --system -r /vllm-workspace/requirements/kv_connectors.txt
+    - HYBRID_SSM=1 bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
+
+- label: MultiConnector (Nixl+Offloading) PD accuracy (2 GPUs)
+  key: multiconnector-nixl-offloading-pd-accuracy-2-gpus
+  timeout_in_minutes: 30
+  working_dir: "/vllm-workspace/tests"
+  num_devices: 2
+  source_file_dependencies:
+    - vllm/distributed/kv_transfer/kv_connector/v1/nixl_connector.py
+    - vllm/distributed/kv_transfer/kv_connector/v1/multi_connector.py
+    - vllm/distributed/kv_transfer/kv_connector/v1/offloading_connector.py
+    - 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_accuracy_test.sh
+
+- label: NixlConnector PD + Spec Decode acceptance (2 GPUs)
+  key: nixlconnector-pd-spec-decode-acceptance-2-gpus
+  timeout_in_minutes: 30
+  device: a100
+  working_dir: "/vllm-workspace/tests"
+  num_devices: 2
+  source_file_dependencies:
+    - vllm/distributed/kv_transfer/kv_connector/v1/nixl_connector.py
+    - 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
+
+- label: MultiConnector (Nixl+Offloading) PD edge cases (2 GPUs)
+  key: multiconnector-nixl-offloading-pd-edge-cases-2-gpus
+  timeout_in_minutes: 30
+  working_dir: "/vllm-workspace/tests"
+  num_devices: 2
+  source_file_dependencies:
+    - vllm/distributed/kv_transfer/kv_connector/v1/nixl_connector.py
+    - vllm/distributed/kv_transfer/kv_connector/v1/multi_connector.py
+    - vllm/distributed/kv_transfer/kv_connector/v1/offloading_connector.py
+    - 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

.buildkite/test_areas/distributed.yaml

@@ -3,6 +3,7 @@ depends_on:
   - image-build
 steps:
 - label: Distributed Comm Ops
+  key: distributed-comm-ops
   timeout_in_minutes: 20
   working_dir: "/vllm-workspace/tests"
   num_devices: 2
@@ -16,6 +17,7 @@ steps:
   - pytest -v -s distributed/test_shm_storage.py
 
 - label: Distributed DP Tests (2 GPUs)
+  key: distributed-dp-tests-2-gpus
   timeout_in_minutes: 20
   working_dir: "/vllm-workspace/tests"
   num_devices: 2
@@ -37,6 +39,7 @@ steps:
   - DP_SIZE=2 pytest -v -s entrypoints/openai/test_multi_api_servers.py
 
 - label: Distributed Compile + RPC Tests (2 GPUs)
+  key: distributed-compile-rpc-tests-2-gpus
   timeout_in_minutes: 20
   working_dir: "/vllm-workspace/tests"
   num_devices: 2
@@ -59,6 +62,7 @@ steps:
   - pytest -v -s ./compile/test_wrapper.py
 
 - label: Distributed Torchrun + Shutdown Tests (2 GPUs)
+  key: distributed-torchrun-shutdown-tests-2-gpus
   timeout_in_minutes: 20
   working_dir: "/vllm-workspace/tests"
   num_devices: 2
@@ -81,6 +85,7 @@ steps:
   - pytest -v -s v1/worker/test_worker_memory_snapshot.py
 
 - label: Distributed Torchrun + Examples (4 GPUs)
+  key: distributed-torchrun-examples-4-gpus
   timeout_in_minutes: 30
   working_dir: "/vllm-workspace"
   num_devices: 4
@@ -88,9 +93,8 @@ steps:
   - vllm/distributed/
   - tests/distributed/test_torchrun_example.py
   - tests/distributed/test_torchrun_example_moe.py
-  - examples/offline_inference/rlhf_colocate.py
   - examples/rl/
-  - tests/examples/offline_inference/data_parallel.py
+  - tests/examples/features/data_parallel/data_parallel_offline.py
   commands:
   # https://github.com/NVIDIA/nccl/issues/1838
   - export NCCL_CUMEM_HOST_ENABLE=0
@@ -107,12 +111,13 @@ steps:
   # test with torchrun tp=2 and dp=2 with ep
   - TP_SIZE=2 DP_SIZE=2 ENABLE_EP=1 torchrun --nproc-per-node=4 tests/distributed/test_torchrun_example_moe.py
   # test with internal dp
-  - python3 examples/offline_inference/data_parallel.py --enforce-eager
+  - python3 examples/features/data_parallel/data_parallel_offline.py --enforce-eager
   # rlhf examples
   - VLLM_ALLOW_INSECURE_SERIALIZATION=1 python3 examples/rl/rlhf_nccl.py
   - VLLM_ALLOW_INSECURE_SERIALIZATION=1 python3 examples/rl/rlhf_ipc.py
 
 - label: Distributed DP Tests (4 GPUs)
+  key: distributed-dp-tests-4-gpus
   timeout_in_minutes: 30
   working_dir: "/vllm-workspace/tests"
   num_devices: 4
@@ -133,6 +138,7 @@ steps:
   - pytest -v -s distributed/test_utils.py
 
 - label: Distributed Compile + Comm (4 GPUs)
+  key: distributed-compile-comm-4-gpus
   timeout_in_minutes: 30
   working_dir: "/vllm-workspace/tests"
   num_devices: 4
@@ -154,24 +160,28 @@ steps:
   - pytest -v -s distributed/test_multiproc_executor.py::test_multiproc_executor_multi_node
 
 - label: Distributed Tests (8 GPUs)(H100)
+  key: distributed-tests-8-gpus-h100
   timeout_in_minutes: 10
   device: h100
   num_devices: 8
   working_dir: "/vllm-workspace/tests"
   source_file_dependencies:
-  - examples/offline_inference/torchrun_dp_example.py
+  - examples/features/torchrun/torchrun_dp_example_offline.py
   - vllm/config/parallel.py
   - vllm/distributed/
   - vllm/v1/engine/llm_engine.py
   - vllm/v1/executor/uniproc_executor.py
   - vllm/v1/worker/gpu_worker.py
+  - tests/distributed/test_mnnvl_alltoall.py
+
   commands:
   # https://github.com/NVIDIA/nccl/issues/1838
   - export NCCL_CUMEM_HOST_ENABLE=0
   # test with torchrun tp=2 and dp=4 with ep
-  - torchrun --nproc-per-node=8 ../examples/offline_inference/torchrun_dp_example.py --tp-size=2 --pp-size=1 --dp-size=4 --enable-ep
+  - torchrun --nproc-per-node=8 ../examples/features/torchrun/torchrun_dp_example_offline.py --tp-size=2 --pp-size=1 --dp-size=4 --enable-ep
 
 - label: Distributed Tests (4 GPUs)(A100)
+  key: distributed-tests-4-gpus-a100
   device: a100
   optional: true
   num_devices: 4
@@ -186,6 +196,7 @@ steps:
   - pytest -v -s -x lora/test_mixtral.py
 
 - label: Distributed Tests (2 GPUs)(H100)
+  key: distributed-tests-2-gpus-h100
   timeout_in_minutes: 15
   device: h100
   optional: true
@@ -194,12 +205,13 @@ steps:
   commands:
     - pytest -v -s tests/distributed/test_context_parallel.py
     - VLLM_ALLOW_INSECURE_SERIALIZATION=1 python3 examples/rl/rlhf_async_new_apis.py
-    - VLLM_USE_DEEP_GEMM=1 VLLM_LOGGING_LEVEL=DEBUG python3 examples/offline_inference/data_parallel.py --model=Qwen/Qwen1.5-MoE-A2.7B -tp=1 -dp=2 --max-model-len=2048 --all2all-backend=deepep_high_throughput
+    - VLLM_USE_DEEP_GEMM=1 VLLM_LOGGING_LEVEL=DEBUG python3 examples/features/data_parallel/data_parallel_offline.py --model=Qwen/Qwen1.5-MoE-A2.7B -tp=1 -dp=2 --max-model-len=2048 --all2all-backend=deepep_high_throughput
     - pytest -v -s tests/v1/distributed/test_dbo.py
     - VLLM_ALLOW_INSECURE_SERIALIZATION=1 pytest -v -s tests/distributed/test_weight_transfer.py
     - pytest -v -s tests/distributed/test_packed_tensor.py
 
 - label: Distributed Tests (2 GPUs)(B200)
+  key: distributed-tests-2-gpus-b200
   device: b200
   optional: true
   working_dir: "/vllm-workspace/"
@@ -208,8 +220,12 @@ steps:
     - pytest -v -s tests/distributed/test_context_parallel.py
     - pytest -v -s tests/distributed/test_nccl_symm_mem_allreduce.py
     - pytest -v -s tests/v1/distributed/test_dbo.py
+    - pytest -v -s tests/distributed/test_mnnvl_alltoall.py
+
+    
 
 - label: 2 Node Test (4 GPUs)
+  key: 2-node-test-4-gpus
   timeout_in_minutes: 30
   working_dir: "/vllm-workspace/tests"
   num_devices: 2
@@ -222,96 +238,12 @@ steps:
   - vllm/executor/
   - vllm/model_executor/models/
   - tests/distributed/
-  - tests/examples/offline_inference/data_parallel.py
+  - tests/examples/features/data_parallel/data_parallel_offline.py
   commands:
-    - ./.buildkite/scripts/run-multi-node-test.sh /vllm-workspace/tests 2 2 $IMAGE_TAG "VLLM_TEST_SAME_HOST=0 torchrun --nnodes 2 --nproc-per-node=2 --rdzv_backend=c10d --rdzv_endpoint=192.168.10.10 distributed/test_same_node.py | grep 'Same node test passed' && NUM_NODES=2 torchrun --nnodes 2 --nproc-per-node=2 --rdzv_backend=c10d --rdzv_endpoint=192.168.10.10 distributed/test_node_count.py | grep 'Node count test passed' && python3 ../examples/offline_inference/data_parallel.py -dp=2 -tp=1 --dp-num-nodes=2 --dp-node-rank=0 --dp-master-addr=192.168.10.10 --dp-master-port=12345 --enforce-eager --trust-remote-code && VLLM_MULTI_NODE=1 pytest -v -s distributed/test_multi_node_assignment.py && VLLM_MULTI_NODE=1 pytest -v -s distributed/test_pipeline_parallel.py" "VLLM_TEST_SAME_HOST=0 torchrun --nnodes 2 --nproc-per-node=2 --rdzv_backend=c10d --rdzv_endpoint=192.168.10.10 distributed/test_same_node.py | grep 'Same node test passed' && NUM_NODES=2 torchrun --nnodes 2 --nproc-per-node=2 --rdzv_backend=c10d --rdzv_endpoint=192.168.10.10 distributed/test_node_count.py | grep 'Node count test passed' && python3 ../examples/offline_inference/data_parallel.py -dp=2 -tp=1 --dp-num-nodes=2 --dp-node-rank=1 --dp-master-addr=192.168.10.10 --dp-master-port=12345 --enforce-eager --trust-remote-code"
-
-- label: Distributed NixlConnector PD accuracy (4 GPUs)
-  timeout_in_minutes: 30
-  working_dir: "/vllm-workspace/tests"
-  num_devices: 4
-  source_file_dependencies:
-    - vllm/distributed/kv_transfer/kv_connector/v1/nixl_connector.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/config_sweep_accuracy_test.sh
-
-- label: DP EP Distributed NixlConnector PD accuracy tests (4 GPUs)
-  timeout_in_minutes: 30
-  working_dir: "/vllm-workspace/tests"
-  num_devices: 4
-  source_file_dependencies:
-    - vllm/distributed/kv_transfer/kv_connector/v1/nixl_connector.py
-    - tests/v1/kv_connector/nixl_integration/
-  commands:
-    - uv pip install --system -r /vllm-workspace/requirements/kv_connectors.txt
-    - 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)
-  timeout_in_minutes: 30
-  working_dir: "/vllm-workspace/tests"
-  num_devices: 4
-  source_file_dependencies:
-    - vllm/distributed/kv_transfer/kv_connector/v1/nixl_connector.py
-    - tests/v1/kv_connector/nixl_integration/
-  commands:
-    - uv pip install --system -r /vllm-workspace/requirements/kv_connectors.txt
-    - CROSS_LAYERS_BLOCKS=True bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
-
-- label: Hyrbid SSM NixlConnector PD accuracy tests (4 GPUs)
-  timeout_in_minutes: 20
-  working_dir: "/vllm-workspace/tests"
-  num_devices: 4
-  source_file_dependencies:
-    - vllm/distributed/kv_transfer/kv_connector/v1/nixl_connector.py
-    - tests/v1/kv_connector/nixl_integration/
-  commands:
-    - uv pip install --system -r /vllm-workspace/requirements/kv_connectors.txt
-    - HYBRID_SSM=1 bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
-
-- label: MultiConnector (Nixl+Offloading) PD accuracy (2 GPUs)
-  timeout_in_minutes: 30
-  working_dir: "/vllm-workspace/tests"
-  num_devices: 2
-  source_file_dependencies:
-    - vllm/distributed/kv_transfer/kv_connector/v1/nixl_connector.py
-    - vllm/distributed/kv_transfer/kv_connector/v1/multi_connector.py
-    - vllm/distributed/kv_transfer/kv_connector/v1/offloading_connector.py
-    - 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_accuracy_test.sh
-
-- label: NixlConnector PD + Spec Decode acceptance (2 GPUs)
-  timeout_in_minutes: 30
-  device: a100
-  working_dir: "/vllm-workspace/tests"
-  num_devices: 2
-  source_file_dependencies:
-    - vllm/distributed/kv_transfer/kv_connector/v1/nixl_connector.py
-    - 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
-
-- label: MultiConnector (Nixl+Offloading) PD edge cases (2 GPUs)
-  timeout_in_minutes: 30
-  working_dir: "/vllm-workspace/tests"
-  num_devices: 2
-  source_file_dependencies:
-    - vllm/distributed/kv_transfer/kv_connector/v1/nixl_connector.py
-    - vllm/distributed/kv_transfer/kv_connector/v1/multi_connector.py
-    - vllm/distributed/kv_transfer/kv_connector/v1/offloading_connector.py
-    - 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
+    - ./.buildkite/scripts/run-multi-node-test.sh /vllm-workspace/tests 2 2 $IMAGE_TAG "VLLM_TEST_SAME_HOST=0 torchrun --nnodes 2 --nproc-per-node=2 --rdzv_backend=c10d --rdzv_endpoint=192.168.10.10 distributed/test_same_node.py | grep 'Same node test passed' && NUM_NODES=2 torchrun --nnodes 2 --nproc-per-node=2 --rdzv_backend=c10d --rdzv_endpoint=192.168.10.10 distributed/test_node_count.py | grep 'Node count test passed' && python3 ../examples/features/data_parallel/data_parallel_offline.py -dp=2 -tp=1 --dp-num-nodes=2 --dp-node-rank=0 --dp-master-addr=192.168.10.10 --dp-master-port=12345 --enforce-eager --trust-remote-code && VLLM_MULTI_NODE=1 pytest -v -s distributed/test_multi_node_assignment.py && VLLM_MULTI_NODE=1 pytest -v -s distributed/test_pipeline_parallel.py" "VLLM_TEST_SAME_HOST=0 torchrun --nnodes 2 --nproc-per-node=2 --rdzv_backend=c10d --rdzv_endpoint=192.168.10.10 distributed/test_same_node.py | grep 'Same node test passed' && NUM_NODES=2 torchrun --nnodes 2 --nproc-per-node=2 --rdzv_backend=c10d --rdzv_endpoint=192.168.10.10 distributed/test_node_count.py | grep 'Node count test passed' && python3 ../examples/features/data_parallel/data_parallel_offline.py -dp=2 -tp=1 --dp-num-nodes=2 --dp-node-rank=1 --dp-master-addr=192.168.10.10 --dp-master-port=12345 --enforce-eager --trust-remote-code"
 
 - label: Pipeline + Context Parallelism (4 GPUs)
+  key: pipeline-context-parallelism-4-gpus
   timeout_in_minutes: 60
   working_dir: "/vllm-workspace/tests"
   num_devices: 4
@@ -326,6 +258,7 @@ steps:
   - pytest -v -s distributed/test_pipeline_parallel.py
 
 - label: RayExecutorV2 (4 GPUs)
+  key: rayexecutorv2-4-gpus
   timeout_in_minutes: 60
   working_dir: "/vllm-workspace/tests"
   num_devices: 4

.buildkite/test_areas/docker.yaml

@@ -0,0 +1,16 @@
+group: Docker
+depends_on:
+  - image-build-cpu
+steps:
+- label: Docker Build Metadata
+  timeout_in_minutes: 10
+  device: cpu-small
+  source_file_dependencies:
+    - .buildkite/release-pipeline.yaml
+    - .buildkite/scripts/docker-build-metadata-args.sh
+    - docker/Dockerfile
+    - docker/Dockerfile.cpu
+    - docker/docker-bake.hcl
+    - tests/tools/test_docker_build_metadata_args.py
+  commands:
+    - pytest -v -s tools/test_docker_build_metadata_args.py

.buildkite/test_areas/e2e_integration.yaml

@@ -3,6 +3,7 @@ depends_on:
   - image-build
 steps:
 - label: DeepSeek V2-Lite Accuracy
+  key: deepseek-v2-lite-accuracy
   timeout_in_minutes: 60
   device: h100
   optional: true
@@ -12,6 +13,7 @@ steps:
   - 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
   timeout_in_minutes: 60
   device: h100
   optional: true
@@ -21,6 +23,7 @@ steps:
   - 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
   timeout_in_minutes: 60
   device: b200
   optional: true
@@ -29,7 +32,18 @@ steps:
   commands:
   - bash .buildkite/scripts/scheduled_integration_test/qwen30b_a3b_fp8_block_ep_eplb.sh 0.8 200 8020 2 1
 
+- label: Qwen3-30B-A3B-FP8 DP4 Async EPLB Accuracy
+  key: qwen3-30b-a3b-fp8-dp4-async-eplb-accuracy
+  timeout_in_minutes: 60
+  device: h100
+  optional: true
+  num_devices: 4
+  working_dir: "/vllm-workspace"
+  commands:
+  - bash .buildkite/scripts/scheduled_integration_test/qwen30b_a3b_fp8_dp4_async_eplb.sh 0.8 200 8050
+
 - label: DeepSeek V2-Lite Prefetch Offload Accuracy (H100)
+  key: deepseek-v2-lite-prefetch-offload-accuracy-h100
   timeout_in_minutes: 60
   device: h100
   optional: true

.buildkite/test_areas/engine.yaml

@@ -3,6 +3,7 @@ depends_on:
   - image-build
 steps:
 - label: Engine
+  key: engine
   timeout_in_minutes: 15
   device: h200_18gb
   source_file_dependencies:
@@ -16,6 +17,7 @@ steps:
   - pytest -v -s engine test_sequence.py test_config.py test_logger.py test_vllm_port.py
 
 - label: Engine (1 GPU)
+  key: engine-1-gpu
   timeout_in_minutes: 30
   source_file_dependencies:
     - vllm/v1/engine/
@@ -25,6 +27,7 @@ steps:
     - pytest -v -s v1/engine --ignore v1/engine/test_preprocess_error_handling.py
 
 - label: e2e Scheduling (1 GPU)
+  key: e2e-scheduling-1-gpu
   timeout_in_minutes: 30
   device: h200_18gb
   source_file_dependencies:
@@ -34,6 +37,7 @@ steps:
     - pytest -v -s v1/e2e/general/test_async_scheduling.py
 
 - label: e2e Core (1 GPU)
+  key: e2e-core-1-gpu
   timeout_in_minutes: 30
   source_file_dependencies:
     - vllm/v1/
@@ -42,6 +46,7 @@ steps:
     - pytest -v -s v1/e2e/general --ignore v1/e2e/general/test_async_scheduling.py
 
 - label: V1 e2e (2 GPUs)
+  key: v1-e2e-2-gpus
   timeout_in_minutes: 60 # TODO: Fix timeout after we have more confidence in the test stability
   optional: true
   num_devices: 2
@@ -58,6 +63,7 @@ steps:
       - image-build-amd
 
 - label: V1 e2e (4 GPUs)
+  key: v1-e2e-4-gpus
   timeout_in_minutes: 60 # TODO: Fix timeout after we have more confidence in the test stability
   optional: true
   num_devices: 4
@@ -74,6 +80,7 @@ steps:
       - image-build-amd
 
 - label: V1 e2e (4xH100)
+  key: v1-e2e-4xh100
   timeout_in_minutes: 60
   device: h100
   num_devices: 4

.buildkite/test_areas/entrypoints.yaml

@@ -2,7 +2,8 @@ group: Entrypoints
 depends_on: 
   - image-build
 steps:
-- label: Entrypoints Unit Tests  
+- label: Entrypoints Unit Tests
+  key: entrypoints-unit-tests
   timeout_in_minutes: 10
   working_dir: "/vllm-workspace/tests"
   source_file_dependencies:
@@ -13,6 +14,7 @@ steps:
   - 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
 
 - label: Entrypoints Integration (LLM)
+  key: entrypoints-integration-llm
   timeout_in_minutes: 40
   working_dir: "/vllm-workspace/tests"
   source_file_dependencies:
@@ -26,6 +28,7 @@ steps:
   - pytest -v -s entrypoints/offline_mode # Needs to avoid interference with other tests
 
 - label: Entrypoints Integration (API Server openai - Part 1)
+  key: entrypoints-integration-api-server-openai-part-1
   timeout_in_minutes: 50
   working_dir: "/vllm-workspace/tests"
   source_file_dependencies:
@@ -43,6 +46,7 @@ steps:
 
 
 - label: Entrypoints Integration (API Server openai - Part 2)
+  key: entrypoints-integration-api-server-openai-part-2
   timeout_in_minutes: 50
   working_dir: "/vllm-workspace/tests"
   source_file_dependencies:
@@ -60,6 +64,7 @@ steps:
       - image-build-amd
 
 - label: Entrypoints Integration (API Server openai - Part 3)
+  key: entrypoints-integration-api-server-openai-part-3
   timeout_in_minutes: 50
   device: h200_18gb
   working_dir: "/vllm-workspace/tests"
@@ -72,6 +77,7 @@ steps:
   - pytest -v -s entrypoints/openai --ignore=entrypoints/openai/chat_completion --ignore=entrypoints/openai/completion --ignore=entrypoints/openai/speech_to_text/ --ignore=entrypoints/openai/correctness/ --ignore=entrypoints/openai/tool_parsers/ --ignore=entrypoints/openai/responses --ignore=entrypoints/openai/test_multi_api_servers.py
 
 - label: Entrypoints Integration (API Server 2)
+  key: entrypoints-integration-api-server-2
   timeout_in_minutes: 130
   working_dir: "/vllm-workspace/tests"
   source_file_dependencies:
@@ -86,6 +92,7 @@ steps:
   - pytest -v -s tool_use
 
 - label: Entrypoints Integration (Pooling)
+  key: entrypoints-integration-pooling
   timeout_in_minutes: 50
   working_dir: "/vllm-workspace/tests"
   source_file_dependencies:
@@ -96,6 +103,7 @@ steps:
   - pytest -v -s entrypoints/pooling
 
 - label: Entrypoints Integration (Responses API)
+  key: entrypoints-integration-responses-api
   timeout_in_minutes: 50
   working_dir: "/vllm-workspace/tests"
   source_file_dependencies:
@@ -105,6 +113,7 @@ steps:
   - pytest -v -s entrypoints/openai/responses
 
 - label: OpenAI API Correctness
+  key: openai-api-correctness
   timeout_in_minutes: 30
   device: h200_18gb
   source_file_dependencies:

.buildkite/test_areas/expert_parallelism.yaml

@@ -3,6 +3,7 @@ depends_on:
   - image-build
 steps:
 - label: EPLB Algorithm
+  key: eplb-algorithm
   timeout_in_minutes: 15
   device: h200_18gb
   working_dir: "/vllm-workspace/tests"
@@ -15,6 +16,7 @@ steps:
   - pytest -v -s distributed/test_eplb_utils.py
 
 - label: EPLB Execution # 17min
+  key: eplb-execution
   timeout_in_minutes: 27
   working_dir: "/vllm-workspace/tests"
   num_devices: 4
@@ -26,6 +28,7 @@ steps:
   - pytest -v -s distributed/test_eplb_spec_decode.py
 
 - label: Elastic EP Scaling Test
+  key: elastic-ep-scaling-test
   timeout_in_minutes: 20
   device: h100
   working_dir: "/vllm-workspace/tests"

.buildkite/test_areas/kernels.yaml

@@ -3,6 +3,7 @@ depends_on:
   - image-build
 steps:
 - label: vLLM IR Tests
+  key: vllm-ir-tests
   timeout_in_minutes: 10
   device: h200_18gb
   working_dir: "/vllm-workspace/"
@@ -14,6 +15,7 @@ steps:
     - pytest -v -s tests/kernels/ir
 
 - label: Kernels Core Operation Test
+  key: kernels-core-operation-test
   timeout_in_minutes: 75
   source_file_dependencies:
   - csrc/
@@ -23,6 +25,7 @@ steps:
     - pytest -v -s kernels/core --ignore=kernels/core/test_minimax_reduce_rms.py  kernels/test_concat_mla_q.py
 
 - label: Kernels MiniMax Reduce RMS Test (2 GPUs)
+  key: kernels-minimax-reduce-rms-test-2-gpus
   timeout_in_minutes: 15
   num_devices: 2
   device: h100
@@ -36,6 +39,7 @@ steps:
     - pytest -v -s kernels/core/test_minimax_reduce_rms.py
 
 - label: Kernels Attention Test %N
+  key: kernels-attention-test
   timeout_in_minutes: 35
   source_file_dependencies:
   - csrc/attention/
@@ -49,6 +53,7 @@ steps:
   parallelism: 2
 
 - label: Kernels Quantization Test %N
+  key: kernels-quantization-test
   timeout_in_minutes: 90
   source_file_dependencies:
   - csrc/quantization/
@@ -59,6 +64,7 @@ steps:
   parallelism: 2
 
 - label: Kernels MoE Test %N
+  key: kernels-moe-test
   timeout_in_minutes: 25
   source_file_dependencies:
   - csrc/quantization/cutlass_w8a8/moe/
@@ -74,6 +80,7 @@ steps:
   parallelism: 5
 
 - label: Kernels Mamba Test
+  key: kernels-mamba-test
   timeout_in_minutes: 45
   source_file_dependencies:
   - csrc/mamba/
@@ -82,7 +89,18 @@ steps:
   commands:
     - pytest -v -s kernels/mamba
 
+- label: Kernels KDA Test
+  timeout_in_minutes: 20
+  source_file_dependencies:
+  - vllm/model_executor/layers/fla/ops/kda.py
+  - vllm/model_executor/layers/fla/ops/chunk_delta_h.py
+  - vllm/model_executor/layers/fla/ops/l2norm.py
+  - tests/kernels/test_kda.py
+  commands:
+    - pytest -v -s kernels/test_kda.py
+
 - label: Kernels DeepGEMM Test (H100)
+  key: kernels-deepgemm-test-h100
   timeout_in_minutes: 45
   device: h100
   num_devices: 1
@@ -95,13 +113,16 @@ steps:
   - tests/kernels/moe/test_deepgemm.py
   - tests/kernels/moe/test_batched_deepgemm.py
   - tests/kernels/attention/test_deepgemm_attention.py
+  - tests/quantization/test_cutlass_w4a16.py
   commands:
     - pytest -v -s kernels/quantization/test_block_fp8.py
     - pytest -v -s kernels/moe/test_deepgemm.py
     - pytest -v -s kernels/moe/test_batched_deepgemm.py
     - pytest -v -s kernels/attention/test_deepgemm_attention.py
+    - pytest -v -s quantization/test_cutlass_w4a16.py
 
 - label: Kernels (B200)
+  key: kernels-b200
   timeout_in_minutes: 30
   working_dir: "/vllm-workspace/"
   device: b200
@@ -141,6 +162,7 @@ steps:
     - pytest -v -s tests/kernels/quantization/test_nvfp4_qutlass.py
     - pytest -v -s tests/kernels/quantization/test_mxfp4_qutlass.py
     - pytest -v -s tests/kernels/moe/test_nvfp4_moe.py
+    - pytest -v -s tests/kernels/moe/test_mxfp4_moe.py
     - 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
@@ -149,6 +171,7 @@ steps:
     - pytest -v -s tests/models/quantization/test_nvfp4.py
 
 - label: Kernels Helion Test
+  key: kernels-helion-test
   timeout_in_minutes: 30
   device: h100
   source_file_dependencies:
@@ -160,6 +183,7 @@ steps:
 
  
 - label: Kernels FP8 MoE Test (1 H100)
+  key: kernels-fp8-moe-test-1-h100
   timeout_in_minutes: 90
   device: h100
   num_devices: 1
@@ -176,6 +200,7 @@ steps:
     - pytest -v -s kernels/moe/test_triton_moe_ptpc_fp8.py
 
 - label: Kernels FP8 MoE Test (2 H100s)
+  key: kernels-fp8-moe-test-2-h100s
   timeout_in_minutes: 90
   device: h100
   num_devices: 2
@@ -185,6 +210,7 @@ steps:
     - pytest -v -s kernels/moe/test_deepep_moe.py
 
 - label: Kernels Fp4 MoE Test (B200)
+  key: kernels-fp4-moe-test-b200
   timeout_in_minutes: 60
   device: b200
   num_devices: 1
@@ -197,6 +223,7 @@ steps:
 
 
 - label: Kernels FusedMoE Layer Test (2 H100s)
+  key: kernels-fusedmoe-layer-test-2-h100s
   timeout_in_minutes: 90
   device: h100
   num_devices: 2
@@ -213,6 +240,7 @@ steps:
 
 
 - label: Kernels FusedMoE Layer Test (2 B200s)
+  key: kernels-fusedmoe-layer-test-2-b200s
   timeout_in_minutes: 90
   device: b200
   num_devices: 2

.buildkite/test_areas/lm_eval.yaml

@@ -3,6 +3,7 @@ depends_on:
   - image-build
 steps:
 - label: LM Eval Small Models
+  key: lm-eval-small-models
   timeout_in_minutes: 75
   source_file_dependencies:
   - csrc/
@@ -24,6 +25,7 @@ steps:
 #   - pytest -s -v test_lm_eval_correctness.py --config-list-file=configs/models-large.txt --tp-size=4
 
 - label: LM Eval Large Models (4 GPUs)(H100)
+  key: lm-eval-large-models-4-gpus-h100
   device: h100
   optional: true
   num_devices: 4
@@ -36,6 +38,7 @@ steps:
     - pytest -s -v test_lm_eval_correctness.py --config-list-file=configs/models-large-hopper.txt --tp-size=4
 
 - label: LM Eval Small Models (B200)
+  key: lm-eval-small-models-b200
   timeout_in_minutes: 120
   device: b200
   optional: true
@@ -46,6 +49,7 @@ steps:
   - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-blackwell.txt
 
 - label: LM Eval Qwen3.5 Models (B200)
+  key: lm-eval-qwen3-5-models-b200
   timeout_in_minutes: 120
   device: b200
   optional: true
@@ -62,6 +66,7 @@ steps:
   - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-qwen35-blackwell.txt
 
 - label: LM Eval Large Models (H200)
+  key: lm-eval-large-models-h200
   timeout_in_minutes: 60
   device: h200
   optional: true
@@ -70,6 +75,7 @@ steps:
     - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-h200.txt
 
 - label: MoE Refactor Integration Test (H100 - TEMPORARY)
+  key: moe-refactor-integration-test-h100-temporary
   device: h100
   optional: true
   num_devices: 2
@@ -77,6 +83,7 @@ steps:
     - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=evals/gsm8k/configs/moe-refactor/config-h100.txt
   
 - label: MoE Refactor Integration Test (B200 - TEMPORARY)
+  key: moe-refactor-integration-test-b200-temporary
   device: b200
   optional: true
   num_devices: 2
@@ -84,6 +91,7 @@ steps:
     - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=evals/gsm8k/configs/moe-refactor/config-b200.txt
 
 - label: MoE Refactor Integration Test (B200 DP - TEMPORARY)
+  key: moe-refactor-integration-test-b200-dp-temporary
   device: b200
   optional: true
   num_devices: 2
@@ -92,6 +100,7 @@ steps:
 
 
 - label: LM Eval TurboQuant KV Cache
+  key: lm-eval-turboquant-kv-cache
   timeout_in_minutes: 75
   source_file_dependencies:
   - vllm/model_executor/layers/quantization/turboquant/
@@ -102,6 +111,7 @@ steps:
   - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=evals/gsm8k/configs/models-turboquant.txt
 
 - label: GPQA Eval (GPT-OSS) (H100)
+  key: gpqa-eval-gpt-oss-h100
   timeout_in_minutes: 120
   device: h100
   optional: true
@@ -115,6 +125,7 @@ steps:
     - pytest -s -v evals/gpt_oss/test_gpqa_correctness.py --config-list-file=configs/models-h100.txt
 
 - label: GPQA Eval (GPT-OSS) (B200)
+  key: gpqa-eval-gpt-oss-b200
   timeout_in_minutes: 120
   device: b200
   optional: true

.buildkite/test_areas/lora.yaml

@@ -3,6 +3,7 @@ depends_on:
   - image-build
 steps:
 - label: LoRA %N
+  key: lora
   timeout_in_minutes: 30
   source_file_dependencies:
   - vllm/lora
@@ -13,6 +14,7 @@ steps:
 
 
 - label: LoRA TP (Distributed)
+  key: lora-tp-distributed
   timeout_in_minutes: 30
   num_devices: 4
   source_file_dependencies:

.buildkite/test_areas/misc.yaml

@@ -3,6 +3,7 @@ depends_on:
   - image-build
 steps:
 - label: V1 Spec Decode
+  key: v1-spec-decode
   timeout_in_minutes: 30
   source_file_dependencies:
     - vllm/
@@ -18,6 +19,7 @@ steps:
       - image-build-amd
 
 - label: V1 Sample + Logits
+  key: v1-sample-logits
   timeout_in_minutes: 30
   device: h200_18gb
   source_file_dependencies:
@@ -41,6 +43,7 @@ steps:
       - image-build-amd
 
 - label: V1 Core + KV + Metrics
+  key: v1-core-kv-metrics
   timeout_in_minutes: 30
   source_file_dependencies:
     - vllm/
@@ -71,6 +74,7 @@ steps:
       - image-build-amd
 
 - label: V1 Others (CPU)
+  key: v1-others-cpu
   depends_on:
     - image-build-cpu
   source_file_dependencies:
@@ -86,6 +90,7 @@ steps:
     - pytest -v -s -m 'cpu_test' v1/metrics
 
 - label: Regression
+  key: regression
   timeout_in_minutes: 20
   device: h200_18gb
   source_file_dependencies:
@@ -97,6 +102,7 @@ steps:
   working_dir: "/vllm-workspace/tests" # optional
 
 - label: Examples
+  key: examples
   timeout_in_minutes: 45
   working_dir: "/vllm-workspace/examples"
   source_file_dependencies:
@@ -113,21 +119,22 @@ steps:
     - python3 basic/offline_inference/embed.py
     - python3 basic/offline_inference/score.py
     # for multi-modal models
-    - python3 offline_inference/audio_language.py --seed 0
-    - python3 offline_inference/vision_language.py --seed 0
-    - python3 offline_inference/vision_language_multi_image.py --seed 0
-    - python3 offline_inference/encoder_decoder_multimodal.py --model-type whisper --seed 0
+    - python3 generate/multimodal/audio_language_offline.py --seed 0
+    - python3 generate/multimodal/vision_language_offline.py --seed 0
+    - python3 generate/multimodal/vision_language_multi_image_offline.py --seed 0
+    - python3 generate/multimodal/encoder_decoder_multimodal_offline.py --model-type whisper --seed 0
      # for pooling models
     - python3 pooling/embed/vision_embedding_offline.py --seed 0
     # for features demo
-    - python3 offline_inference/prefix_caching.py
+    - python3 features/automatic_prefix_caching/prefix_caching_offline.py
     - python3 offline_inference/llm_engine_example.py
     - python3 others/tensorize_vllm_model.py --model facebook/opt-125m serialize --serialized-directory /tmp/ --suffix v1 && python3 others/tensorize_vllm_model.py --model facebook/opt-125m deserialize --path-to-tensors /tmp/vllm/facebook/opt-125m/v1/model.tensors
-    - python3 offline_inference/spec_decode.py --test --method eagle --num_spec_tokens 3 --dataset-name hf --dataset-path philschmid/mt-bench --num-prompts 80 --temp 0 --top-p 1.0 --top-k -1 --tp 1 --enable-chunked-prefill --max-model-len 2048
+    - python3 features/speculative_decoding/spec_decode_offline.py --test --method eagle --num_spec_tokens 3 --dataset-name hf --dataset-path philschmid/mt-bench --num-prompts 80 --temp 0 --top-p 1.0 --top-k -1 --tp 1 --enable-chunked-prefill --max-model-len 2048
     # https://github.com/vllm-project/vllm/pull/26682 uses slightly more memory in PyTorch 2.9+ causing this test to OOM in 1xL4 GPU
-    - python3 offline_inference/spec_decode.py --test --method eagle3 --num_spec_tokens 3 --dataset-name hf --dataset-path philschmid/mt-bench --num-prompts 80 --temp 0 --top-p 1.0 --top-k -1 --tp 1 --enable-chunked-prefill --max-model-len 1536
+    - python3 features/speculative_decoding/spec_decode_offline.py --test --method eagle3 --num_spec_tokens 3 --dataset-name hf --dataset-path philschmid/mt-bench --num-prompts 80 --temp 0 --top-p 1.0 --top-k -1 --tp 1 --enable-chunked-prefill --max-model-len 1536
 
 - label: Metrics, Tracing (2 GPUs)
+  key: metrics-tracing-2-gpus
   timeout_in_minutes: 20
   num_devices: 2
   source_file_dependencies:
@@ -142,6 +149,7 @@ steps:
   - pytest -v -s v1/tracing
 
 - label: Python-only Installation
+  key: python-only-installation
   depends_on: ~
   timeout_in_minutes: 20
   source_file_dependencies:
@@ -151,6 +159,7 @@ steps:
   - bash standalone_tests/python_only_compile.sh
 
 - label: Async Engine, Inputs, Utils, Worker
+  key: async-engine-inputs-utils-worker
   timeout_in_minutes: 50
   source_file_dependencies:
   - vllm/
@@ -163,7 +172,8 @@ steps:
   - pytest -v -s utils_
 
 - label: Async Engine, Inputs, Utils, Worker, Config (CPU)
-  depends_on: 
+  key: async-engine-inputs-utils-worker-config-cpu
+  depends_on:
   - image-build-cpu
   timeout_in_minutes: 30
   source_file_dependencies:
@@ -196,6 +206,7 @@ steps:
   - pytest -v -s config
 
 - label: Batch Invariance (H100)
+  key: batch-invariance-h100
   timeout_in_minutes: 30
   device: h100
   source_file_dependencies:
@@ -211,6 +222,7 @@ steps:
     - 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]
 
 - label: Batch Invariance (B200)
+  key: batch-invariance-b200
   timeout_in_minutes: 30
   device: b200
   source_file_dependencies:
@@ -227,6 +239,7 @@ steps:
     - pytest -v -s v1/determinism/test_nvfp4_batch_invariant.py
   
 - label: Acceptance Length Test (Large Models) # optional
+  key: acceptance-length-test-large-models
   timeout_in_minutes: 25
   gpu: h100
   optional: true

.buildkite/test_areas/model_executor.yaml

@@ -3,6 +3,7 @@ depends_on:
   - image-build
 steps:
 - label: Model Executor
+  key: model-executor
   timeout_in_minutes: 35
   source_file_dependencies:
   - vllm/engine/arg_utils.py

.buildkite/test_areas/model_runner_v2.yaml

@@ -3,6 +3,7 @@ depends_on:
   - image-build
 steps:
 - label: Model Runner V2 Core Tests
+  key: model-runner-v2-core-tests
   timeout_in_minutes: 45
   source_file_dependencies:
   - vllm/v1/worker/gpu/
@@ -25,14 +26,16 @@ 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
+  key: model-runner-v2-examples
   timeout_in_minutes: 45
   working_dir: "/vllm-workspace/examples"
   source_file_dependencies:
     - vllm/v1/worker/gpu/
     - vllm/v1/core/sched/
     - vllm/v1/worker/gpu_worker.py
-    - examples/offline_inference/
     - examples/basic/offline_inference/
+    - examples/generate/multimodal/
+    - examples/features/
     - examples/pooling/embed/vision_embedding_offline.py
     - examples/others/tensorize_vllm_model.py
   commands:
@@ -44,21 +47,22 @@ steps:
     #- python3 basic/offline_inference/generate.py --model meta-llama/Llama-2-13b-chat-hf --cpu-offload-gb 10  # TODO
     #- python3 basic/offline_inference/embed.py   # TODO
     # for multi-modal models
-    - python3 offline_inference/audio_language.py --seed 0
-    - python3 offline_inference/vision_language.py --seed 0
-    - python3 offline_inference/vision_language_multi_image.py --seed 0
-    - python3 offline_inference/encoder_decoder_multimodal.py --model-type whisper --seed 0
+    - python3 generate/multimodal/audio_language_offline.py --seed 0
+    - python3 generate/multimodal/vision_language_offline.py --seed 0
+    - python3 generate/multimodal/vision_language_multi_image_offline.py --seed 0
+    - python3 generate/multimodal/encoder_decoder_multimodal_offline.py --model-type whisper --seed 0
     # for pooling models
     - python3 pooling/embed/vision_embedding_offline.py --seed 0
     # for features demo
-    - python3 offline_inference/prefix_caching.py
+    - python3 features/automatic_prefix_caching/prefix_caching_offline.py
     - python3 offline_inference/llm_engine_example.py
     - python3 others/tensorize_vllm_model.py --model facebook/opt-125m serialize --serialized-directory /tmp/ --suffix v1 && python3 others/tensorize_vllm_model.py --model facebook/opt-125m deserialize --path-to-tensors /tmp/vllm/facebook/opt-125m/v1/model.tensors
-    - python3 offline_inference/spec_decode.py --test --method eagle --num_spec_tokens 3 --dataset-name hf --dataset-path philschmid/mt-bench --num-prompts 80 --temp 0 --top-p 1.0 --top-k -1 --tp 1 --enable-chunked-prefill --max-model-len 2048
+    - python3 features/speculative_decoding/spec_decode_offline.py --test --method eagle --num_spec_tokens 3 --dataset-name hf --dataset-path philschmid/mt-bench --num-prompts 80 --temp 0 --top-p 1.0 --top-k -1 --tp 1 --enable-chunked-prefill --max-model-len 2048
     # https://github.com/vllm-project/vllm/pull/26682 uses slightly more memory in PyTorch 2.9+ causing this test to OOM in 1xL4 GPU
-    - python3 offline_inference/spec_decode.py --test --method eagle3 --num_spec_tokens 3 --dataset-name hf --dataset-path philschmid/mt-bench --num-prompts 80 --temp 0 --top-p 1.0 --top-k -1 --tp 1 --enable-chunked-prefill --max-model-len 1536
+    - python3 features/speculative_decoding/spec_decode_offline.py --test --method eagle3 --num_spec_tokens 3 --dataset-name hf --dataset-path philschmid/mt-bench --num-prompts 80 --temp 0 --top-p 1.0 --top-k -1 --tp 1 --enable-chunked-prefill --max-model-len 1536
 
 - label: Model Runner V2 Distributed (2 GPUs)
+  key: model-runner-v2-distributed-2-gpus
   timeout_in_minutes: 45
   working_dir: "/vllm-workspace/tests"
   num_devices: 2
@@ -79,6 +83,7 @@ steps:
     - TP_SIZE=1 DP_SIZE=2 pytest -v -s v1/distributed/test_eagle_dp.py
 
 - label: Model Runner V2 Pipeline Parallelism (4 GPUs)
+  key: model-runner-v2-pipeline-parallelism-4-gpus
   timeout_in_minutes: 60
   working_dir: "/vllm-workspace/tests"
   num_devices: 4
@@ -94,6 +99,7 @@ steps:
     - pytest -v -s distributed/test_pp_cudagraph.py -k "not ray"
 
 - label: Model Runner V2 Spec Decode
+  key: model-runner-v2-spec-decode
   timeout_in_minutes: 30
   working_dir: "/vllm-workspace/tests"
   source_file_dependencies:

.buildkite/test_areas/models_basic.yaml

@@ -3,6 +3,7 @@ depends_on:
   - image-build
 steps:
 - label: Basic Models Tests (Initialization)
+  key: basic-models-tests-initialization
   timeout_in_minutes: 45
   torch_nightly: true
   source_file_dependencies:
@@ -16,6 +17,7 @@ steps:
     torch_nightly: {}
 
 - label: Basic Models Tests (Extra Initialization) %N
+  key: basic-models-tests-extra-initialization
   timeout_in_minutes: 45
   source_file_dependencies:
   - vllm/model_executor/models/
@@ -31,6 +33,7 @@ steps:
     torch_nightly: {}
 
 - label: Basic Models Tests (Other)
+  key: basic-models-tests-other
   timeout_in_minutes: 45
   source_file_dependencies:
   - vllm/
@@ -47,6 +50,7 @@ steps:
 
 
 - label: Basic Models Test (Other CPU) # 5min
+  key: basic-models-test-other-cpu
   depends_on:
   - image-build-cpu
   timeout_in_minutes: 10
@@ -59,6 +63,7 @@ steps:
     - 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
@@ -69,11 +74,12 @@ steps:
     - 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/offline_inference/vision_language.py --model-type qwen2_5_vl
+    - 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/offline_inference/audio_language.py --model-type whisper
+    - 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
@@ -83,7 +89,7 @@ steps:
     - 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/offline_inference/basic/chat.py
-    - python3 examples/offline_inference/vision_language.py --model-type qwen2_5_vl
+    - 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/offline_inference/audio_language.py --model-type whisper
+    - VLLM_WORKER_MULTIPROC_METHOD=spawn python3 examples/generate/multimodal/audio_language_offline.py --model-type whisper

.buildkite/test_areas/models_distributed.yaml

@@ -3,6 +3,7 @@ depends_on:
   - image-build
 steps:
 - label: Distributed Model Tests (2 GPUs)
+  key: distributed-model-tests-2-gpus
   timeout_in_minutes: 50
   working_dir: "/vllm-workspace/tests"
   num_devices: 2

.buildkite/test_areas/models_language.yaml

@@ -3,6 +3,7 @@ depends_on:
   - image-build
 steps:
 - label: Language Models Tests (Standard)
+  key: language-models-tests-standard
   timeout_in_minutes: 25
   source_file_dependencies:
   - vllm/
@@ -15,6 +16,7 @@ steps:
     torch_nightly: {}
 
 - label: Language Models Tests (Extra Standard) %N
+  key: language-models-tests-extra-standard
   timeout_in_minutes: 45
   source_file_dependencies:
   - vllm/model_executor/models/
@@ -31,6 +33,7 @@ steps:
     torch_nightly: {}
 
 - label: Language Models Tests (Hybrid) %N
+  key: language-models-tests-hybrid
   timeout_in_minutes: 75
   source_file_dependencies:
   - vllm/
@@ -47,6 +50,7 @@ steps:
     torch_nightly: {}
 
 - label: Language Models Test (Extended Generation) # 80min
+  key: language-models-test-extended-generation
   timeout_in_minutes: 110
   optional: true
   source_file_dependencies:
@@ -69,6 +73,7 @@ steps:
       - pytest -v -s models/language/generation -m '(not core_model) and (not hybrid_model)'
 
 - label: Language Models Test (PPL)
+  key: language-models-test-ppl
   timeout_in_minutes: 110
   device: h200_18gb
   optional: true
@@ -79,6 +84,7 @@ steps:
     - pytest -v -s models/language/generation_ppl_test
 
 - label: Language Models Test (Extended Pooling)  # 36min
+  key: language-models-test-extended-pooling
   timeout_in_minutes: 50
   optional: true
   source_file_dependencies:
@@ -93,6 +99,7 @@ steps:
       - image-build-amd
 
 - label: Language Models Test (MTEB)
+  key: language-models-test-mteb
   timeout_in_minutes: 110
   device: h200_18gb
   optional: true

.buildkite/test_areas/models_multimodal.yaml

@@ -3,6 +3,7 @@ depends_on:
   - image-build
 steps:
 - label: "Multi-Modal Models (Standard) 1: qwen2"
+  key: multi-modal-models-standard-1-qwen2
   timeout_in_minutes: 45
   device: h200_18gb
   source_file_dependencies:
@@ -19,6 +20,7 @@ steps:
       - image-build-amd
 
 - label: "Multi-Modal Models (Standard) 2: qwen3 + gemma"
+  key: multi-modal-models-standard-2-qwen3-gemma
   timeout_in_minutes: 45
   device: h200_18gb
   source_file_dependencies:
@@ -28,6 +30,7 @@ steps:
     - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
     - pytest -v -s models/multimodal/generation/test_common.py -m core_model -k "qwen3 or gemma"
     - pytest -v -s models/multimodal/generation/test_qwen2_5_vl.py -m core_model
+    - pytest -v -s models/multimodal/generation/test_vit_cudagraph.py -m core_model
   mirror:
     amd:
       device: mi325_1
@@ -35,6 +38,7 @@ steps:
       - image-build-amd
 
 - label: "Multi-Modal Models (Standard) 3: llava + qwen2_vl"
+  key: multi-modal-models-standard-3-llava-qwen2-vl
   timeout_in_minutes: 45
   source_file_dependencies:
   - vllm/
@@ -50,6 +54,7 @@ steps:
       - image-build-amd
 
 - label: "Multi-Modal Models (Standard) 4: other + whisper"
+  key: multi-modal-models-standard-4-other-whisper
   timeout_in_minutes: 45
   source_file_dependencies:
   - vllm/
@@ -66,7 +71,8 @@ steps:
       - image-build-amd
 
 - label: Multi-Modal Processor (CPU)
-  depends_on: 
+  key: multi-modal-processor-cpu
+  depends_on:
   - image-build-cpu
   timeout_in_minutes: 60
   source_file_dependencies:
@@ -79,6 +85,7 @@ steps:
     - pytest -v -s models/multimodal/processing --ignore models/multimodal/processing/test_tensor_schema.py
 
 - label: Multi-Modal Processor # 44min
+  key: multi-modal-processor
   timeout_in_minutes: 60
   device: h200_18gb
   source_file_dependencies:
@@ -90,6 +97,7 @@ steps:
     - pytest -v -s models/multimodal/processing/test_tensor_schema.py
 
 - label: Multi-Modal Accuracy Eval (Small Models) # 50min
+  key: multi-modal-accuracy-eval-small-models
   timeout_in_minutes: 70
   working_dir: "/vllm-workspace/.buildkite/lm-eval-harness"
   source_file_dependencies:
@@ -100,6 +108,7 @@ steps:
   - pytest -s -v test_lm_eval_correctness.py --config-list-file=configs/models-mm-small.txt --tp-size=1
 
 - label: Multi-Modal Models (Extended Generation 1)
+  key: multi-modal-models-extended-generation-1
   optional: true
   source_file_dependencies:
   - vllm/
@@ -116,6 +125,7 @@ steps:
       - image-build-amd
 
 - label: Multi-Modal Models (Extended Generation 2)
+  key: multi-modal-models-extended-generation-2
   optional: true
   source_file_dependencies:
   - vllm/
@@ -125,6 +135,7 @@ 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)
+  key: multi-modal-models-extended-generation-3
   optional: true
   source_file_dependencies:
   - vllm/
@@ -134,6 +145,7 @@ steps:
     - pytest -v -s models/multimodal/generation/test_common.py -m 'split(group=1) and not core_model'
 
 - label: Multi-Modal Models (Extended Pooling)
+  key: multi-modal-models-extended-pooling
   optional: true
   device: h200_18gb
   source_file_dependencies:

.buildkite/test_areas/plugins.yaml

@@ -3,6 +3,7 @@ depends_on:
   - image-build
 steps:
 - label: Plugin Tests (2 GPUs)
+  key: plugin-tests-2-gpus
   timeout_in_minutes: 60
   working_dir: "/vllm-workspace/tests"
   num_devices: 2

.buildkite/test_areas/pytorch.yaml

@@ -3,6 +3,7 @@ depends_on:
   - image-build
 steps:
 - label: PyTorch Compilation Unit Tests
+  key: pytorch-compilation-unit-tests
   timeout_in_minutes: 10
   source_file_dependencies:
     - vllm/
@@ -18,6 +19,7 @@ steps:
   - "find compile/ -maxdepth 1 -name 'test_*.py' -print0 | xargs -0 -n1 -I{} pytest -s -v '{}'"
 
 - label: PyTorch Compilation Unit Tests (H100)
+  key: pytorch-compilation-unit-tests-h100
   timeout_in_minutes: 30
   device: h100
   num_devices: 1
@@ -28,6 +30,7 @@ steps:
   - "find compile/h100/ -name 'test_*.py' -print0 | xargs -0 -n1 -I{} pytest -s -v '{}'"
 
 - label: PyTorch Compilation Passes Unit Tests
+  key: pytorch-compilation-passes-unit-tests
   timeout_in_minutes: 20
   source_file_dependencies:
     - vllm/
@@ -36,6 +39,7 @@ steps:
   - pytest -s -v compile/passes --ignore compile/passes/distributed
 
 - label: PyTorch Fullgraph Smoke Test
+  key: pytorch-fullgraph-smoke-test
   timeout_in_minutes: 35
   source_file_dependencies:
   - vllm/
@@ -48,6 +52,7 @@ steps:
   - "find compile/fullgraph/ -name 'test_*.py' -not -name 'test_full_graph.py' -print0 | xargs -0 -n1 -I{} pytest -s -v '{}'"
 
 - label: PyTorch Fullgraph
+  key: pytorch-fullgraph
   timeout_in_minutes: 30
   device: h200_18gb
   source_file_dependencies:
@@ -58,6 +63,7 @@ steps:
   - pytest -v -s compile/fullgraph/test_full_graph.py -k 'not test_fp8_kv_scale_compile'
 
 - label: Pytorch Nightly Dependency Override Check # 2min
+  key: pytorch-nightly-dependency-override-check
   # if this test fails, it means the nightly torch version is not compatible with some
   # of the dependencies. Please check the error message and add the package to whitelist
   # in /vllm/tools/pre_commit/generate_nightly_torch_test.py

.buildkite/test_areas/quantization.yaml

@@ -3,6 +3,7 @@ depends_on:
   - image-build
 steps:
 - label: Quantization
+  key: quantization
   timeout_in_minutes: 90
   source_file_dependencies:
   - csrc/
@@ -21,6 +22,7 @@ steps:
   - VLLM_TEST_FORCE_LOAD_FORMAT=auto pytest -v -s quantization/ --ignore quantization/test_blackwell_moe.py
 
 - label: Quantized MoE Test (B200)
+  key: quantized-moe-test-b200
   timeout_in_minutes: 60
   working_dir: "/vllm-workspace/"
   device: b200
@@ -38,6 +40,7 @@ steps:
     - pytest -s -v tests/quantization/test_blackwell_moe.py
 
 - label: Quantized Models Test
+  key: quantized-models-test
   timeout_in_minutes: 60
   source_file_dependencies:
   - vllm/model_executor/layers/quantization

.buildkite/test_areas/ray_compat.yaml

@@ -3,6 +3,7 @@ depends_on:
   - image-build
 steps:
 - label: Ray Dependency Compatibility Check
+  key: ray-dependency-compatibility-check
   # Informational only — does not block the pipeline.
   # If this fails, it means the PR introduces a dependency that
   # conflicts with Ray's dependency constraints.

.buildkite/test_areas/samplers.yaml

@@ -3,6 +3,7 @@ depends_on:
   - image-build
 steps:
 - label: Samplers Test
+  key: samplers-test
   timeout_in_minutes: 75
   source_file_dependencies:
   - vllm/model_executor/layers
@@ -10,7 +11,9 @@ steps:
   - tests/samplers
   - tests/conftest.py
   commands:
-    - pytest -v -s samplers
+    # VLLM_USE_FLASHINFER_SAMPLER defaults to 1 now, so we need to pin both
+    # values explicitly to still cover the PyTorch-native (Triton) path.
+    - VLLM_USE_FLASHINFER_SAMPLER=0 pytest -v -s samplers
     - VLLM_USE_FLASHINFER_SAMPLER=1 pytest -v -s samplers
   mirror:
     amd:

.buildkite/test_areas/spec_decode.yaml

@@ -3,6 +3,7 @@ depends_on:
   - image-build
 steps:
 - label: Spec Decode Eagle
+  key: spec-decode-eagle
   timeout_in_minutes: 30
   device: h200_18gb
   source_file_dependencies:
@@ -12,7 +13,20 @@ steps:
   commands:
     - pytest -v -s v1/e2e/spec_decode -k "eagle_correctness"
 
+- label: Spec Decode Eagle Nightly B200
+  key: spec-decode-eagle-nightly-b200
+  timeout_in_minutes: 30
+  device: b200
+  optional: true
+  source_file_dependencies:
+    - vllm/v1/spec_decode/
+    - vllm/v1/worker/gpu/spec_decode/
+    - tests/v1/e2e/spec_decode/
+  commands:
+    - pytest -v -s v1/e2e/spec_decode -k "eagle_correctness"
+
 - label: Spec Decode Speculators + MTP
+  key: spec-decode-speculators-mtp
   timeout_in_minutes: 30
   device: h200_18gb
   source_file_dependencies:
@@ -23,7 +37,21 @@ steps:
   commands:
     - pytest -v -s v1/e2e/spec_decode -k "speculators or mtp_correctness"
 
+- label: Spec Decode Speculators + MTP Nightly B200
+  key: spec-decode-speculators-mtp-nightly-b200
+  timeout_in_minutes: 30
+  device: b200
+  optional: true
+  source_file_dependencies:
+    - vllm/v1/spec_decode/
+    - vllm/v1/worker/gpu/spec_decode/
+    - vllm/transformers_utils/configs/speculators/
+    - tests/v1/e2e/spec_decode/
+  commands:
+    - pytest -v -s v1/e2e/spec_decode -k "speculators or mtp_correctness"
+  
 - label: Spec Decode Ngram + Suffix
+  key: spec-decode-ngram-suffix
   timeout_in_minutes: 30
   device: h200_18gb
   source_file_dependencies:
@@ -34,6 +62,7 @@ steps:
     - pytest -v -s v1/e2e/spec_decode -k "ngram or suffix"
 
 - label: Spec Decode Draft Model
+  key: spec-decode-draft-model
   timeout_in_minutes: 30
   device: h200_18gb
   source_file_dependencies:
@@ -43,7 +72,20 @@ steps:
   commands:
     - pytest -v -s v1/e2e/spec_decode -k "draft_model or no_sync or batch_inference"
 
+- label: Spec Decode Draft Model Nightly B200
+  key: spec-decode-draft-model-nightly-b200
+  timeout_in_minutes: 30
+  device: b200
+  optional: true
+  source_file_dependencies:
+    - vllm/v1/spec_decode/
+    - vllm/v1/worker/gpu/spec_decode/
+    - tests/v1/e2e/spec_decode/
+  commands:
+    - pytest -v -s v1/e2e/spec_decode -k "draft_model or no_sync or batch_inference"
+
 - label: DFlash Speculators Correctness
+  key: dflash-speculators-correctness
   timeout_in_minutes: 30
   device: h100
   optional: true
@@ -55,3 +97,16 @@ steps:
   commands:
     - export VLLM_ALLOW_INSECURE_SERIALIZATION=1
     - pytest -v -s v1/spec_decode/test_speculators_dflash.py -m slow_test
+
+- label: Spec Decode MTP hybrid (B200)
+  timeout_in_minutes: 30
+  device: b200
+  optional: true
+  source_file_dependencies:
+    - vllm/v1/spec_decode/
+    - vllm/v1/worker/gpu/spec_decode/
+    - vllm/model_executor/models/qwen3_5.py
+    - vllm/model_executor/models/qwen3_5_mtp.py
+    - tests/v1/e2e/spec_decode/
+  commands:
+    - pytest -v -s v1/e2e/spec_decode -k "qwen3_5-hybrid"

.buildkite/test_areas/weight_loading.yaml

@@ -3,6 +3,7 @@ depends_on:
   - image-build
 steps:
 - label: Weight Loading Multiple GPU  # 33min
+  key: weight-loading-multiple-gpu
   timeout_in_minutes: 45
   working_dir: "/vllm-workspace/tests"
   num_devices: 2

.github/mergify.yml

@@ -262,7 +262,7 @@ pull_request_rules:
       - files~=^docker/Dockerfile.xpu
       - files~=^\\.buildkite/intel_jobs/
       - files=\.buildkite/ci_config_intel.yaml
-      - files=vllm/model_executor/layers/fused_moe/xpu_fused_moe.py
+      - files=vllm/model_executor/layers/fused_moe/experts/xpu_moe.py
       - files=vllm/model_executor/kernels/linear/mixed_precision/xpu.py
       - files=vllm/model_executor/kernels/linear/mxfp8/xpu.py
       - files=vllm/model_executor/kernels/linear/scaled_mm/xpu.py
@@ -308,8 +308,7 @@ pull_request_rules:
       - files=benchmarks/benchmark_serving_structured_output.py
       - files=benchmarks/run_structured_output_benchmark.sh
       - files=docs/features/structured_outputs.md
-      - files=examples/offline_inference/structured_outputs.py
-      - files=examples/online_serving/structured_outputs/structured_outputs.py
+      - files=^examples/features/structured_outputs/
       - files~=^tests/v1/structured_output/
       - files=tests/entrypoints/llm/test_struct_output_generate.py
       - files~=^vllm/v1/structured_output/
@@ -325,7 +324,7 @@ pull_request_rules:
     - or:
       - files~=^vllm/v1/spec_decode/
       - files~=^tests/v1/spec_decode/
-      - files~=^examples/.*(spec_decode|mlpspeculator|eagle|speculation).*\.py
+      - files=^examples/features/speculative_decoding/
       - files~=^vllm/model_executor/models/.*eagle.*\.py
       - files=vllm/model_executor/models/mlp_speculator.py
       - files~=^vllm/transformers_utils/configs/(eagle|medusa|mlp_speculator)\.py
@@ -389,11 +388,7 @@ pull_request_rules:
       - files~=^tests/entrypoints/anthropic/.*tool.*
       - files~=^vllm/tool_parsers/
       - files=docs/features/tool_calling.md
-      - files~=^examples/tool_chat_*
-      - files=examples/offline_inference/chat_with_tools.py
-      - files=examples/online_serving/openai_chat_completion_client_with_tools_required.py
-      - files=examples/online_serving/openai_chat_completion_tool_calls_with_reasoning.py
-      - files=examples/online_serving/openai_chat_completion_client_with_tools.py
+      - files~=^examples/tool_calling/
   actions:
     label:
       add:

.github/workflows/pre-commit.yml

@@ -16,11 +16,7 @@ permissions:
 
 jobs:
   pre-run-check:
-    if: >-
-      github.event_name == 'pull_request' &&
-      (github.event.action != 'labeled' ||
-       github.event.label.name == 'ready' ||
-       github.event.label.name == 'verified')
+    if: github.event_name == 'pull_request'
     runs-on: ubuntu-latest
     steps:
     - name: Check PR label and author merge count
@@ -49,12 +45,7 @@ jobs:
 
   pre-commit:
     needs: pre-run-check
-    if: >-
-      always() &&
-      (github.event.action != 'labeled' ||
-       github.event.label.name == 'ready' ||
-       github.event.label.name == 'verified') &&
-      (needs.pre-run-check.result == 'success' || needs.pre-run-check.result == 'skipped')
+    if: always() && (needs.pre-run-check.result == 'success' || needs.pre-run-check.result == 'skipped')
     runs-on: ubuntu-latest
     steps:
     - uses: actions/checkout@8e8c483db84b4bee98b60c0593521ed34d9990e8 # v6.0.1

.github/workflows/scripts/build.sh

@@ -14,7 +14,7 @@ $python_executable -m pip install -r requirements/build/cuda.txt -r requirements
 # Limit the number of parallel jobs to avoid OOM
 export MAX_JOBS=1
 # Make sure release wheels are built for the following architectures
-export TORCH_CUDA_ARCH_LIST="7.0 7.5 8.0 8.6 8.9 9.0+PTX"
+export TORCH_CUDA_ARCH_LIST="7.5 8.0 8.6 8.9 9.0 10.0 12.0+PTX"
 
 bash tools/check_repo.sh
 

.gitignore

@@ -237,6 +237,7 @@ ep_kernels_workspace/
 
 # Allow tracked library source folders under submodules (e.g., benchmarks/lib)
 !vllm/benchmarks/lib/
+!.buildkite/scripts/lib/
 
 # Generated gRPC protobuf files (compiled at build time from vllm_engine.proto)
 vllm/grpc/vllm_engine_pb2.py

CMakeLists.txt

@@ -34,10 +34,10 @@ install(CODE "set(CMAKE_INSTALL_LOCAL_ONLY TRUE)" ALL_COMPONENTS)
 # Supported python versions.  These versions will be searched in order, the
 # first match will be selected.  These should be kept in sync with setup.py.
 #
-set(PYTHON_SUPPORTED_VERSIONS "3.10" "3.11" "3.12" "3.13")
+set(PYTHON_SUPPORTED_VERSIONS "3.10" "3.11" "3.12" "3.13" "3.14")
 
 # Supported AMD GPU architectures.
-set(HIP_SUPPORTED_ARCHS "gfx906;gfx908;gfx90a;gfx942;gfx950;gfx1030;gfx1100;gfx1101;gfx1150;gfx1151;gfx1152;gfx1153;gfx1200;gfx1201")
+set(HIP_SUPPORTED_ARCHS "gfx906;gfx908;gfx90a;gfx942;gfx950;gfx1030;gfx1100;gfx1101;gfx1102;gfx1103;gfx1150;gfx1151;gfx1152;gfx1153;gfx1200;gfx1201")
 
 # ROCm installation prefix. Default to /opt/rocm but allow override via
 # -DROCM_PATH=/your/rocm/path when invoking cmake.
@@ -94,12 +94,15 @@ find_package(Torch REQUIRED)
 # This check must happen after find_package(Torch) because that's when CMAKE_CUDA_COMPILER_VERSION gets defined
 if(DEFINED CMAKE_CUDA_COMPILER_VERSION AND
    CMAKE_CUDA_COMPILER_VERSION VERSION_GREATER_EQUAL 13.0)
-  set(CUDA_SUPPORTED_ARCHS "7.5;8.0;8.6;8.7;8.9;9.0;10.0;11.0;12.0;12.1")
+   # starting from CUDA 12.9 and Blackwell (10.0), we use family-specific targets (10.0f, 12.0f, etc)
+   # to support the whole generation without specifying all sub-architectures
+   # see: https://developer.nvidia.com/blog/nvidia-blackwell-and-nvidia-cuda-12-9-introduce-family-specific-architecture-features/
+  set(CUDA_SUPPORTED_ARCHS "7.5;8.0;8.6;8.7;8.9;9.0;10.0;11.0;12.0")
 elseif(DEFINED CMAKE_CUDA_COMPILER_VERSION AND
    CMAKE_CUDA_COMPILER_VERSION VERSION_GREATER_EQUAL 12.8)
-  set(CUDA_SUPPORTED_ARCHS "7.0;7.2;7.5;8.0;8.6;8.7;8.9;9.0;10.0;10.1;12.0;12.1")
+  set(CUDA_SUPPORTED_ARCHS "7.5;8.0;8.6;8.7;8.9;9.0;10.0;10.1;10.3;12.0;12.1")
 else()
-  set(CUDA_SUPPORTED_ARCHS "7.0;7.2;7.5;8.0;8.6;8.7;8.9;9.0")
+  set(CUDA_SUPPORTED_ARCHS "7.0;7.5;8.0;8.6;8.7;8.9;9.0")
 endif()
 
 #
@@ -307,7 +310,9 @@ set(VLLM_EXT_SRC
   "csrc/torch_bindings.cpp")
 
 if(VLLM_GPU_LANG STREQUAL "CUDA")
-  list(APPEND VLLM_EXT_SRC "csrc/minimax_reduce_rms_kernel.cu")
+  list(APPEND VLLM_EXT_SRC
+    "csrc/minimax_reduce_rms_kernel.cu"
+    "csrc/fused_deepseek_v4_qnorm_rope_kv_insert_kernel.cu")
 
   SET(CUTLASS_ENABLE_HEADERS_ONLY ON CACHE BOOL "Enable only the header library")
 
@@ -952,7 +957,9 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
       "csrc/libtorch_stable/quantization/fp4/activation_nvfp4_quant_fusion_kernels.cu"
       "csrc/libtorch_stable/quantization/fp4/nvfp4_experts_quant.cu"
       "csrc/libtorch_stable/quantization/fp4/nvfp4_scaled_mm_kernels.cu"
-      "csrc/libtorch_stable/quantization/fp4/nvfp4_blockwise_moe_kernel.cu")
+      "csrc/libtorch_stable/quantization/fp4/nvfp4_blockwise_moe_kernel.cu"
+      "csrc/libtorch_stable/quantization/fp4/mxfp4_experts_quant.cu"
+      "csrc/libtorch_stable/quantization/fp4/mxfp4_blockwise_moe_kernel.cu")
     set_gencode_flags_for_srcs(
       SRCS "${SRCS}"
       CUDA_ARCHS "${FP4_ARCHS}")
@@ -1046,7 +1053,8 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
   list(APPEND VLLM_MOE_EXT_SRC
     "csrc/moe/moe_wna16.cu"
     "csrc/moe/grouped_topk_kernels.cu"
-    "csrc/moe/router_gemm.cu")
+    "csrc/moe/router_gemm.cu"
+    "csrc/moe/topk_softplus_sqrt_kernels.cu")
 endif()
 
 if(VLLM_GPU_LANG STREQUAL "CUDA")

README.md

@@ -14,7 +14,7 @@ Easy, fast, and cheap LLM serving for everyone
 | <a href="https://docs.vllm.ai"><b>Documentation</b></a> | <a href="https://blog.vllm.ai/"><b>Blog</b></a> | <a href="https://arxiv.org/abs/2309.06180"><b>Paper</b></a> | <a href="https://x.com/vllm_project"><b>Twitter/X</b></a> | <a href="https://discuss.vllm.ai"><b>User Forum</b></a> | <a href="https://slack.vllm.ai"><b>Developer Slack</b></a> |
 </p>
 
-🔥 We have built a vllm website to help you get started with vllm. Please visit [vllm.ai](https://vllm.ai) to learn more.
+🔥 We have built a vLLM website to help you get started with vLLM. Please visit [vllm.ai](https://vllm.ai) to learn more.
 For events, please visit [vllm.ai/events](https://vllm.ai/events) to join us.
 
 ---
@@ -50,7 +50,7 @@ vLLM is flexible and easy to use with:
 - Efficient multi-LoRA support for dense and MoE layers
 - Support for NVIDIA GPUs, AMD GPUs, and x86/ARM/PowerPC CPUs. Additionally, diverse hardware plugins such as Google TPUs, Intel Gaudi, IBM Spyre, Huawei Ascend, Rebellions NPU, Apple Silicon, MetaX GPU, and more.
 
-vLLM seamlessly supports 200+ model architectures on HuggingFace, including:
+vLLM seamlessly supports 200+ model architectures on Hugging Face, including:
 
 - Decoder-only LLMs (e.g., Llama, Qwen, Gemma)
 - Mixture-of-Expert LLMs (e.g., Mixtral, DeepSeek-V3, Qwen-MoE, GPT-OSS)

benchmarks/attention_benchmarks/mla_runner.py

@@ -404,6 +404,7 @@ def _build_attention_metadata(
         query_start_loc=q_start_gpu,
         query_start_loc_cpu=q_start_cpu,
         seq_lens=seq_lens_gpu,
+        seq_lens_cpu_upper_bound=seq_lens_cpu,
         _seq_lens_cpu=seq_lens_cpu,
         _num_computed_tokens_cpu=num_computed_tokens_cpu,
         slot_mapping=slot_mapping,

benchmarks/kernels/benchmark_cutlass_moe_fp8.py

@@ -16,7 +16,7 @@ from vllm.model_executor.layers.fused_moe.all2all_utils import (
     maybe_make_prepare_finalize,
 )
 from vllm.model_executor.layers.fused_moe.config import fp8_w8a8_moe_quant_config
-from vllm.model_executor.layers.fused_moe.cutlass_moe import CutlassExpertsFp8
+from vllm.model_executor.layers.fused_moe.experts.cutlass_moe import CutlassExpertsFp8
 from vllm.model_executor.layers.fused_moe.fused_moe import fused_experts, fused_topk
 from vllm.platforms import current_platform
 from vllm.utils.argparse_utils import FlexibleArgumentParser

benchmarks/kernels/benchmark_cutlass_moe_nvfp4.py

@@ -22,7 +22,7 @@ from vllm.model_executor.layers.fused_moe.config import (
     fp8_w8a8_moe_quant_config,
     nvfp4_moe_quant_config,
 )
-from vllm.model_executor.layers.fused_moe.cutlass_moe import (
+from vllm.model_executor.layers.fused_moe.experts.cutlass_moe import (
     CutlassExpertsFp4,
 )
 from vllm.model_executor.layers.fused_moe.fused_moe import fused_experts, fused_topk

benchmarks/kernels/benchmark_grouped_gemm_cutlass.py

@@ -13,7 +13,7 @@ from vllm.model_executor.layers.fused_moe.all2all_utils import (
     maybe_make_prepare_finalize,
 )
 from vllm.model_executor.layers.fused_moe.config import fp8_w8a8_moe_quant_config
-from vllm.model_executor.layers.fused_moe.cutlass_moe import CutlassExpertsFp8
+from vllm.model_executor.layers.fused_moe.experts.cutlass_moe import CutlassExpertsFp8
 from vllm.model_executor.layers.fused_moe.fused_moe import (
     fused_experts,
     fused_topk,

benchmarks/kernels/benchmark_vit_fp8_attn.py

@@ -0,0 +1,324 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+# Benchmarks FP8 vs BF16 ViT attention via FlashInfer cuDNN backend.
+#
+# == Usage Examples ==
+#
+# Benchmark mode (default, FlashInfer CUDAGraph Bench)
+#   python3 benchmark_vit_fp8_attn.py
+#
+# Profile mode (PyTorch profiler, saves TensorBoard traces):
+#   python3 benchmark_vit_fp8_attn.py --profile
+#   python3 benchmark_vit_fp8_attn.py --profile --profile-output-dir ./profile_traces
+#
+# Custom seq_lens:
+#   python3 benchmark_vit_fp8_attn.py --seq-lens 4096 8192 16384
+
+from functools import partial
+
+import numpy as np
+import torch
+from torch.profiler import ProfilerActivity, profile, record_function
+
+from vllm.utils.argparse_utils import FlexibleArgumentParser
+
+# Qwen3-VL defaults
+NUM_HEADS = 16
+HEAD_DIM = 72
+DEFAULT_SEQ_LENS = [2304, 4096, 8192, 16384]
+
+
+def _setup_fp8_attention(num_heads: int, head_dim: int) -> tuple:
+    """Create FP8 and BF16 attention modules + workspace."""
+    from types import SimpleNamespace
+    from unittest.mock import patch
+
+    from vllm.config import VllmConfig, set_current_vllm_config
+    from vllm.config.multimodal import MultiModalConfig
+    from vllm.model_executor.layers.attention.mm_encoder_attention import (
+        MMEncoderAttention,
+        _get_flashinfer_workspace_buffer,
+    )
+    from vllm.v1.attention.backends.registry import AttentionBackendEnum
+
+    old_dtype = torch.get_default_dtype()
+    torch.set_default_dtype(torch.bfloat16)
+
+    backend_patch = patch(
+        "vllm.model_executor.layers.attention.mm_encoder_attention"
+        ".get_vit_attn_backend",
+        return_value=AttentionBackendEnum.FLASHINFER,
+    )
+
+    # FP8 attention
+    mm_config_fp8 = MultiModalConfig(mm_encoder_attn_dtype="fp8")
+    vllm_config_fp8 = VllmConfig()
+    vllm_config_fp8.model_config = SimpleNamespace(multimodal_config=mm_config_fp8)
+    with set_current_vllm_config(vllm_config_fp8), backend_patch:
+        attn_fp8 = MMEncoderAttention(
+            num_heads=num_heads,
+            head_size=head_dim,
+            prefix="visual.blocks.0.attn",
+        ).to("cuda")
+
+    # BF16 attention (no FP8)
+    with set_current_vllm_config(VllmConfig()), backend_patch:
+        attn_bf16 = MMEncoderAttention(
+            num_heads=num_heads,
+            head_size=head_dim,
+            prefix="visual.blocks.0.attn",
+        ).to("cuda")
+
+    torch.set_default_dtype(old_dtype)
+
+    workspace = _get_flashinfer_workspace_buffer()
+    return attn_fp8, attn_bf16, workspace
+
+
+def _build_meta(
+    seq_len: int,
+    num_heads: int,
+    head_dim: int,
+    fp8: bool,
+):
+    """Build cu_seqlens, max_seqlen, sequence_lengths."""
+    from vllm.model_executor.layers.attention.mm_encoder_attention import (
+        MMEncoderAttention,
+    )
+    from vllm.utils.math_utils import round_up
+    from vllm.v1.attention.backends.registry import AttentionBackendEnum
+
+    cu_np = np.array([0, seq_len], dtype=np.int32)
+    fp8_padded = num_heads * round_up(head_dim, 16) if fp8 else None
+
+    seq_lengths = MMEncoderAttention.maybe_compute_seq_lens(
+        AttentionBackendEnum.FLASHINFER, cu_np, torch.device("cuda")
+    )
+    max_seqlen = torch.tensor(
+        MMEncoderAttention.compute_max_seqlen(AttentionBackendEnum.FLASHINFER, cu_np),
+        dtype=torch.int32,
+    )
+    cu_seqlens = MMEncoderAttention.maybe_recompute_cu_seqlens(
+        AttentionBackendEnum.FLASHINFER,
+        cu_np,
+        num_heads * head_dim,
+        1,
+        torch.device("cuda"),
+        fp8_padded_hidden_size=fp8_padded,
+    )
+    return cu_seqlens, max_seqlen, seq_lengths
+
+
+def run_benchmark(
+    seq_lens: list[int],
+    num_heads: int,
+    head_dim: int,
+    method: str,
+):
+    """Benchmark FP8 vs BF16 attention across seq_lens.
+
+    Uses FlashInfer GPU-level timing to measure pure kernel time,
+    excluding CPU launch overhead.
+    """
+    if method == "cupti":
+        from flashinfer.testing import bench_gpu_time_with_cupti as bench_fn
+
+        bench_fn = partial(bench_fn, use_cuda_graph=True, cold_l2_cache=False)
+    elif method == "cudagraph":
+        from flashinfer.testing import (
+            bench_gpu_time_with_cudagraph as bench_fn,
+        )
+
+        bench_fn = partial(bench_fn, cold_l2_cache=False)
+    else:
+        raise ValueError(f"Invalid method: {method}")
+
+    attn_fp8, attn_bf16, workspace = _setup_fp8_attention(num_heads, head_dim)
+
+    print(f"Timing method: {method}")
+    print(f"{'seq_len':>8} {'BF16 (us)':>12} {'FP8 (us)':>12} {'Speedup':>10}")
+    print("-" * 46)
+
+    for seq_len in seq_lens:
+        torch.manual_seed(42)
+
+        q = torch.randn(
+            seq_len,
+            num_heads,
+            head_dim,
+            device="cuda",
+            dtype=torch.bfloat16,
+        )
+        k = torch.randn_like(q)
+        v = torch.randn_like(q)
+
+        cu_fp8, max_s, seq_l = _build_meta(seq_len, num_heads, head_dim, fp8=True)
+        # we can reuse cu_fp8 for cu_bf16 since q, k, and v are contiguous
+        cu_bf16 = cu_fp8.clone()
+
+        def bf16_fn(q=q, k=k, v=v, cu=cu_bf16, ms=max_s, sl=seq_l):
+            attn_bf16._forward_flashinfer(q, k, v, cu, ms, sl)
+
+        def fp8_fn(q=q, k=k, v=v, cu=cu_fp8, ms=max_s, sl=seq_l):
+            attn_fp8._forward_flashinfer(q, k, v, cu, ms, sl)
+
+        # bench_fn returns List[float] of per-iteration times in ms
+        bf16_times = bench_fn(bf16_fn)
+        fp8_times = bench_fn(fp8_fn)
+
+        bf16_us = np.median(bf16_times) * 1e3  # ms -> us
+        fp8_us = np.median(fp8_times) * 1e3
+        speedup = bf16_us / fp8_us if fp8_us > 0 else float("inf")
+
+        print(f"{seq_len:>8} {bf16_us:>12.1f} {fp8_us:>12.1f} {speedup:>9.2f}x")
+
+
+def _make_trace_handler(output_dir: str, worker_name: str, label: str):
+    """Create a trace handler that saves to TensorBoard and prints summary."""
+
+    def handler(prof):
+        torch.profiler.tensorboard_trace_handler(output_dir, worker_name)(prof)
+        print(f"\n{'=' * 80}")
+        print(label)
+        print(f"{'=' * 80}")
+        print(prof.key_averages().table(sort_by="cuda_time_total", row_limit=20))
+
+    return handler
+
+
+def run_profile(
+    seq_len: int,
+    num_heads: int,
+    head_dim: int,
+    warmup: int,
+    output_dir: str,
+):
+    """Profile FP8 vs BF16 attention with PyTorch profiler."""
+    attn_fp8, attn_bf16, workspace = _setup_fp8_attention(num_heads, head_dim)
+
+    torch.manual_seed(42)
+    q = torch.randn(
+        seq_len,
+        num_heads,
+        head_dim,
+        device="cuda",
+        dtype=torch.bfloat16,
+    )
+    k = torch.randn_like(q)
+    v = torch.randn_like(q)
+
+    cu_fp8, max_s, seq_l = _build_meta(seq_len, num_heads, head_dim, fp8=True)
+    # we can reuse cu_fp8 for cu_bf16 since q, k, and v are contiguous
+    cu_bf16 = cu_fp8.clone()
+
+    sched = torch.profiler.schedule(wait=0, warmup=warmup, active=1)
+
+    # Profile BF16 (warmup handled by profiler schedule)
+    with profile(
+        activities=[ProfilerActivity.CPU, ProfilerActivity.CUDA],
+        schedule=sched,
+        on_trace_ready=_make_trace_handler(
+            output_dir,
+            f"bf16_h{head_dim}_s{seq_len}",
+            f"BF16 Attention (seq_len={seq_len}, heads={num_heads}, "
+            f"head_dim={head_dim})",
+        ),
+    ) as prof_bf16:
+        for _ in range(warmup + 1):
+            with record_function("bf16_attention"):
+                attn_bf16._forward_flashinfer(
+                    q.clone(), k.clone(), v.clone(), cu_bf16, max_s, seq_l
+                )
+                torch.accelerator.synchronize()
+            prof_bf16.step()
+
+    # Profile FP8 (warmup handled by profiler schedule)
+    with profile(
+        activities=[ProfilerActivity.CPU, ProfilerActivity.CUDA],
+        schedule=sched,
+        on_trace_ready=_make_trace_handler(
+            output_dir,
+            f"fp8_h{head_dim}_s{seq_len}",
+            f"FP8 Attention (seq_len={seq_len}, heads={num_heads}, "
+            f"head_dim={head_dim})",
+        ),
+    ) as prof_fp8:
+        for _ in range(warmup + 1):
+            with record_function("fp8_attention"):
+                attn_fp8._forward_flashinfer(
+                    q.clone(), k.clone(), v.clone(), cu_fp8, max_s, seq_l
+                )
+                torch.accelerator.synchronize()
+            prof_fp8.step()
+
+    print(f"\nTensorBoard traces saved to: {output_dir}")
+    print(f"View with: tensorboard --logdir={output_dir}")
+
+
+if __name__ == "__main__":
+    parser = FlexibleArgumentParser(description="Benchmark FP8 vs BF16 ViT attention.")
+    parser.add_argument(
+        "--seq-lens",
+        type=int,
+        nargs="+",
+        default=DEFAULT_SEQ_LENS,
+        help="Sequence lengths to benchmark",
+    )
+    parser.add_argument(
+        "--num-heads",
+        type=int,
+        default=NUM_HEADS,
+    )
+    parser.add_argument(
+        "--head-dim",
+        type=int,
+        default=HEAD_DIM,
+    )
+    parser.add_argument(
+        "--method",
+        choices=["cupti", "cudagraph"],
+        default="cudagraph",
+        help="GPU timing method: cupti (CUPTI kernel timing) or "
+        "cudagraph (CUDA graph capture/replay). Default: cudagraph",
+    )
+    parser.add_argument(
+        "--warmup",
+        type=int,
+        default=10,
+        help="Warmup iterations (profile mode only)",
+    )
+    parser.add_argument(
+        "--profile",
+        action="store_true",
+        help="Run PyTorch profiler instead of benchmark",
+    )
+    parser.add_argument(
+        "--profile-seq-len",
+        type=int,
+        default=8192,
+        help="Sequence length for profiling (default: 8192)",
+    )
+    parser.add_argument(
+        "--profile-output-dir",
+        type=str,
+        default="./profile_traces",
+        help="Output directory for TensorBoard traces (default: ./profile_traces)",
+    )
+    args = parser.parse_args()
+
+    if args.profile:
+        run_profile(
+            args.profile_seq_len,
+            args.num_heads,
+            args.head_dim,
+            args.warmup,
+            args.profile_output_dir,
+        )
+    else:
+        run_benchmark(
+            args.seq_lens,
+            args.num_heads,
+            args.head_dim,
+            args.method,
+        )

benchmarks/kernels/ir/bench_ir_ops.py

@@ -0,0 +1,378 @@
+#!/usr/bin/env python3
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""
+Generic benchmark harness for vLLM IR ops.
+
+Usage:
+    python benchmarks/kernels/ir/bench_ir_ops.py
+    python benchmarks/kernels/ir/bench_ir_ops.py --ops rms_norm
+    python benchmarks/kernels/ir/bench_ir_ops.py --ops rms_norm,silu_mul
+    python benchmarks/kernels/ir/bench_ir_ops.py --no-cuda-graph
+    python benchmarks/kernels/ir/bench_ir_ops.py --ops rms_norm --save-path ./results/
+"""
+
+import argparse
+import contextlib
+import csv
+import dataclasses
+import datetime
+import math
+import os
+import subprocess
+import sys
+import tempfile
+
+# Ensure repo root is on sys.path so `benchmarks` is importable as a package.
+_REPO_ROOT = os.path.abspath(os.path.join(os.path.dirname(__file__), "../../.."))
+if _REPO_ROOT not in sys.path:
+    sys.path.insert(0, _REPO_ROOT)
+
+# Suppress noisy C++ warnings from vllm kernel registration (written to fd 2
+# directly by the dynamic linker, so Python-level sys.stderr redirect won't
+# catch them).
+_saved_fd = os.dup(2)
+try:
+    with open(os.devnull, "w") as _devnull:
+        os.dup2(_devnull.fileno(), 2)
+        import torch
+
+        import vllm.kernels  # noqa: E402, F401
+finally:
+    os.dup2(_saved_fd, 2)
+    os.close(_saved_fd)
+
+from tqdm import tqdm  # noqa: E402
+
+from benchmarks.kernels.ir.shapes import SHAPE_CONFIGS  # noqa: E402  # isort: skip
+from vllm.ir.op import IrOp  # noqa: E402
+from vllm.platforms import current_platform  # noqa: E402
+from vllm.triton_utils import triton  # noqa: E402
+
+
+@dataclasses.dataclass(frozen=True)
+class BenchConfig:
+    use_cuda_graph: bool = True
+    warmup: int = 25
+    rep: int = 100
+
+
+def _pkg_version(name: str) -> str:
+    from importlib.metadata import PackageNotFoundError, version
+
+    with contextlib.suppress(PackageNotFoundError):
+        return version(name)
+    return "not installed"
+
+
+_METADATA_LABELS = {
+    "timestamp": "Timestamp",
+    "git_commit": "Git commit",
+    "vllm": "vLLM",
+    "pytorch": "PyTorch",
+    "cuda_runtime": "CUDA runtime",
+    "triton": "Triton",
+    "cutlass": "CUTLASS",
+    "helion": "Helion",
+    "device": "Device",
+    "bench_mode": "Bench mode",
+    "warmup": "Warmup",
+    "rep": "Repetitions",
+}
+
+
+def collect_env_metadata(cfg: BenchConfig) -> dict[str, str]:
+    from vllm.collect_env import get_env_info
+
+    env = get_env_info()
+
+    git_sha = "unknown"
+    with contextlib.suppress(subprocess.CalledProcessError, FileNotFoundError):
+        git_sha = (
+            subprocess.check_output(
+                ["git", "rev-parse", "--short", "HEAD"], stderr=subprocess.DEVNULL
+            )
+            .decode()
+            .strip()
+        )
+
+    device_name = current_platform.get_device_name()
+
+    warmup_note = " ms" if not cfg.use_cuda_graph else " ms (ignored)"
+    rep_note = " replays" if cfg.use_cuda_graph else " ms"
+
+    return {
+        "timestamp": datetime.datetime.now().strftime("%Y-%m-%d %H:%M:%S"),
+        "git_commit": git_sha,
+        "vllm": str(env.vllm_version),
+        "pytorch": str(env.torch_version),
+        "cuda_runtime": str(env.cuda_runtime_version),
+        "triton": triton.__version__,
+        "cutlass": _pkg_version("nvidia-cutlass-dsl"),
+        "helion": _pkg_version("helion"),
+        "device": device_name,
+        "bench_mode": "cuda_graph" if cfg.use_cuda_graph else "eager",
+        "warmup": f"{cfg.warmup}{warmup_note}",
+        "rep": f"{cfg.rep}{rep_note}",
+    }
+
+
+def print_metadata(metadata: dict[str, str]):
+    print("=" * 60)
+    for key, val in metadata.items():
+        print(f"{_METADATA_LABELS.get(key, key) + ':':<16}{val}")
+    print("=" * 60)
+
+
+def _clone_args(args: tuple) -> tuple:
+    return tuple(a.clone() if isinstance(a, torch.Tensor) else a for a in args)
+
+
+# TODO(gmagogsfm): When the `maybe_inplace` PR lands, ops marked as
+# inplace=True will mutate bench_args across iterations. Both CUDA graph
+# and eager modes will accumulate drift from repeated in-place mutation.
+# We need to re-clone inputs per iteration for inplace ops.
+def _bench_one(fn, args, cfg: BenchConfig) -> float:
+    bench_args = _clone_args(args)
+    bench_fn = lambda: fn(*bench_args)
+
+    if cfg.use_cuda_graph:
+        ms = triton.testing.do_bench_cudagraph(bench_fn, rep=cfg.rep, quantiles=[0.5])
+    else:
+        ms = triton.testing.do_bench(
+            bench_fn, warmup=cfg.warmup, rep=cfg.rep, quantiles=[0.5]
+        )
+    return ms * 1000
+
+
+# TODO(gmagogsfm): Once compiled native implementation lands (#38775),
+# the benchmark baseline should be the compiled native (what vLLM runs by
+# default) rather than the uncompiled native implementation.
+def collect_timings(
+    op: IrOp, shape_configs: list[dict], cfg: BenchConfig
+) -> tuple[list[str], list[str], dict[str, dict[str, float]]]:
+    def fmt(v) -> str:
+        return str(v).split(".")[-1] if isinstance(v, torch.dtype) else str(v)
+
+    case_names = [
+        "_".join(f"{k}={fmt(v)}" for k, v in kwargs.items()) for kwargs in shape_configs
+    ]
+    providers = [n for n, impl in op.impls.items() if impl.supported]
+
+    results: dict[str, dict[str, float]] = {c: {} for c in case_names}
+    for provider in providers:
+        impl = op.impls[provider]
+        desc = f"{op.name} / {provider}"
+        for case_name, kwargs in tqdm(
+            zip(case_names, shape_configs),
+            desc=desc,
+            total=len(case_names),
+            unit=" cases",
+        ):
+            args = op.generate_inputs(**kwargs)
+            if impl.supports_args(*args):
+                results[case_name][provider] = _bench_one(impl.impl_fn, args, cfg)
+            else:
+                results[case_name][provider] = float("nan")
+
+    return case_names, providers, results
+
+
+def analyze_results(
+    op_name: str,
+    case_names: list[str],
+    providers: list[str],
+    results: dict[str, dict[str, float]],
+) -> tuple[list[dict[str, str]], list[dict[str, str]], list[str]]:
+    native_col = "native"
+    non_native = [p for p in providers if p != native_col]
+
+    header_cols = ["case"]
+    for p in providers:
+        header_cols.append(f"{p} (us)")
+    for p in non_native:
+        header_cols.append(f"{p} speedup")
+
+    detail_rows: list[dict[str, str]] = []
+    speedup_data: dict[str, list[tuple[float, str]]] = {p: [] for p in non_native}
+
+    for case_name in case_names:
+        timings = results[case_name]
+        row: dict[str, str] = {"case": case_name}
+
+        for p in providers:
+            val = timings.get(p, float("nan"))
+            row[f"{p} (us)"] = f"{val:.2f}" if not math.isnan(val) else "n/a"
+
+        native_us = timings.get(native_col, float("nan"))
+        for p in non_native:
+            p_us = timings.get(p, float("nan"))
+            if not math.isnan(native_us) and not math.isnan(p_us) and p_us > 0:
+                speedup = native_us / p_us
+                row[f"{p} speedup"] = f"{speedup:.2f}x"
+                speedup_data[p].append((speedup, case_name))
+            else:
+                row[f"{p} speedup"] = "n/a"
+
+        detail_rows.append(row)
+
+    summary_rows: list[dict[str, str]] = []
+    for p in non_native:
+        entries = speedup_data[p]
+        if not entries:
+            continue
+        speedups = [s for s, _ in entries]
+        geomean = math.exp(sum(math.log(s) for s in speedups) / len(speedups))
+        best_val, best_case = max(entries)
+        worst_val, worst_case = min(entries)
+        wins = sum(1 for s in speedups if s > 1.0)
+        losses = sum(1 for s in speedups if s < 1.0)
+        total = len(speedups)
+
+        print(f"\n{p} vs native ({wins}/{total} faster, {losses}/{total} slower):")
+        print(f"  geomean speedup: {geomean:.2f}x")
+        print(f"  best:            {best_val:.2f}x  ({best_case})")
+        print(f"  worst:           {worst_val:.2f}x  ({worst_case})")
+
+        summary_rows.append(
+            {
+                "op": op_name,
+                "provider": p,
+                "geomean_speedup": f"{geomean:.2f}",
+                "best_speedup": f"{best_val:.2f}",
+                "best_case": best_case,
+                "worst_speedup": f"{worst_val:.2f}",
+                "worst_case": worst_case,
+                "wins": str(wins),
+                "losses": str(losses),
+                "total": str(total),
+            }
+        )
+
+    return detail_rows, summary_rows, header_cols
+
+
+def write_csv(path: str, rows: list[dict[str, str]], fieldnames: list[str]):
+    with open(path, "w", newline="") as f:
+        writer = csv.DictWriter(f, fieldnames=fieldnames)
+        writer.writeheader()
+        writer.writerows(rows)
+
+
+def save_results(
+    save_dir: str,
+    op_name: str,
+    detail_rows: list[dict[str, str]],
+    header_cols: list[str],
+    all_summary_rows: list[dict[str, str]],
+    metadata: dict[str, str],
+):
+    write_csv(
+        os.path.join(save_dir, f"{op_name}_detail.csv"),
+        detail_rows,
+        header_cols,
+    )
+    if all_summary_rows:
+        write_csv(
+            os.path.join(save_dir, "summary.csv"),
+            all_summary_rows,
+            list(all_summary_rows[0].keys()),
+        )
+    write_csv(
+        os.path.join(save_dir, "metadata.csv"),
+        [metadata],
+        list(metadata.keys()),
+    )
+
+
+def parse_args():
+    parser = argparse.ArgumentParser(description="Benchmark vLLM IR ops")
+    parser.add_argument(
+        "--ops",
+        type=str,
+        default=None,
+        help="Comma-separated list of op names to benchmark (substring match)",
+    )
+    parser.add_argument(
+        "--no-cuda-graph",
+        action="store_true",
+        help="Disable CUDA graph; use do_bench with L2 cache flushing instead",
+    )
+    parser.add_argument(
+        "--warmup",
+        type=int,
+        default=25,
+        help="Warmup time in ms (do_bench) or ignored with CUDA graph (default: 25)",
+    )
+    parser.add_argument(
+        "--rep",
+        type=int,
+        default=100,
+        help="Repetition time in ms (do_bench) or number of graph replays "
+        "(do_bench_cudagraph) (default: 100)",
+    )
+    parser.add_argument(
+        "--save-path",
+        type=str,
+        default=None,
+        help="Directory to save results (default: auto-created temp dir)",
+    )
+    return parser.parse_args()
+
+
+def main():
+    args = parse_args()
+    cfg = BenchConfig(
+        use_cuda_graph=not args.no_cuda_graph,
+        warmup=args.warmup,
+        rep=args.rep,
+    )
+
+    torch.set_default_device(current_platform.device_type)
+
+    metadata = collect_env_metadata(cfg)
+    print_metadata(metadata)
+
+    timestamp = datetime.datetime.now().strftime("%Y%m%d_%H%M%S")
+    save_dir = args.save_path or os.path.join(
+        tempfile.gettempdir(), f"vllm_ir_bench_{timestamp}"
+    )
+    os.makedirs(save_dir, exist_ok=True)
+
+    op_filters = [f.strip() for f in args.ops.split(",")] if args.ops else None
+    all_summary_rows: list[dict[str, str]] = []
+
+    for op in IrOp.registry.values():
+        if op_filters and not any(f in op.name for f in op_filters):
+            continue
+        if not op.has_input_generator:
+            print(f"Skipping op '{op.name}': no input generator registered")
+            continue
+        if op.name not in SHAPE_CONFIGS:
+            raise RuntimeError(
+                f"No benchmark shape config for op '{op.name}'. "
+                f"Add it to benchmarks/kernels/ir/shapes.py"
+            )
+
+        case_names, providers, results = collect_timings(
+            op, SHAPE_CONFIGS[op.name], cfg
+        )
+        detail_rows, summary_rows, header_cols = analyze_results(
+            op.name, case_names, providers, results
+        )
+        all_summary_rows.extend(summary_rows)
+
+        save_results(
+            save_dir,
+            op.name,
+            detail_rows,
+            header_cols,
+            all_summary_rows,
+            metadata,
+        )
+
+    print(f"\nResults saved to: {save_dir}")
+
+
+if __name__ == "__main__":
+    main()

benchmarks/kernels/ir/shapes.py

@@ -0,0 +1,29 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""
+Shape configurations for IR op benchmarks.
+"""
+
+import torch
+
+NUM_TOKENS = [1, 2, 4, 16, 64, 256, 1024, 4096, 16384]
+COMMON_HIDDEN_SIZES = [
+    2048,  # Llama 3.2 1B, Qwen 3 MoE 30B-A3B, Gemma 3n
+    3072,  # Gemma 7B/9B
+    4096,  # Llama 3 8B, Qwen 3 8B, Mistral 7B
+    5120,  # Llama 4 Scout 17B-16E
+    7168,  # DeepSeek V3
+    8192,  # Llama 3 70B
+    16384,  # Llama 3 405B
+]
+
+# Each entry maps an op name to a list of kwarg dicts that will be passed
+# to that op's registered input generator via op.generate_inputs(**kwargs).
+SHAPE_CONFIGS: dict[str, list[dict]] = {
+    "rms_norm": [
+        {"num_tokens": n, "hidden_size": d, "dtype": dtype}
+        for dtype in [torch.float16, torch.bfloat16, torch.float32]
+        for d in COMMON_HIDDEN_SIZES
+        for n in NUM_TOKENS
+    ],
+}

benchmarks/multi_turn/benchmark_serving_multi_turn.py

@@ -217,6 +217,7 @@ async def send_request(
     min_tokens: int | None = None,
     max_tokens: int | None = None,
     timeout_sec: int = 120,
+    conversation_id: str | None = None,
 ) -> ServerResponse:
     payload = {
         "model": model,
@@ -225,6 +226,9 @@ async def send_request(
         "temperature": 0.0,
     }
 
+    if conversation_id is not None:
+        payload["conversation_id"] = conversation_id
+
     if stream:
         payload["stream"] = True
         payload["stream_options"] = {"include_usage": False}
@@ -419,6 +423,7 @@ async def send_turn(
         min_tokens,
         max_tokens,
         req_args.timeout_sec,
+        conversation_id=conv_id,
     )
 
     if response.valid is False:

cmake/cpu_extension.cmake

@@ -161,16 +161,49 @@ elseif (S390_FOUND)
         "-mtune=native")
 elseif (CMAKE_SYSTEM_PROCESSOR MATCHES "riscv64")
     message(STATUS "RISC-V detected")
-    if(RVV_BF16_FOUND)
-        message(STATUS "BF16 extension detected")
-        set(MARCH_FLAGS -march=rv64gcv_zvfh_zfbfmin_zvfbfmin_zvl128b -mrvv-vector-bits=zvl -mabi=lp64d)
-        add_compile_definitions(RISCV_BF16_SUPPORT)
-    elseif (RVV_FP16_FOUND)
-        message(WARNING "BF16 functionality is not available")
-        set(MARCH_FLAGS -march=rv64gcv_zvfh_zvl128b -mrvv-vector-bits=zvl -mabi=lp64d)
+    # VLLM_RVV_VLEN selects the target VLEN. Auto-detected from /proc/cpuinfo
+    # by default; override with -DVLLM_RVV_VLEN=128 or -DVLLM_RVV_VLEN=256.
+    if(NOT DEFINED VLLM_RVV_VLEN)
+        # Auto-detect: find the largest zvl<N>b in /proc/cpuinfo isa line.
+        if(EXISTS /proc/cpuinfo)
+            file(READ /proc/cpuinfo _cpuinfo)
+            set(_best 0)
+            foreach(_n IN ITEMS 128 256 512 1024)
+                if(_cpuinfo MATCHES "zvl${_n}b")
+                    set(_best ${_n})
+                endif()
+            endforeach()
+            if(_best GREATER 0)
+                set(VLLM_RVV_VLEN ${_best})
+            endif()
+        endif()
+        # If auto-detect failed (no /proc/cpuinfo or no zvl<N>b reported)
+        # but the compiler supports RVV, require explicit specification.
+        if(NOT DEFINED VLLM_RVV_VLEN AND (RVV_FP16_FOUND OR RVV_BF16_FOUND))
+            message(FATAL_ERROR
+                "RISC-V RVV is available but VLEN could not be auto-detected. "
+                "Please specify VLEN explicitly:\n"
+                "  -DVLLM_RVV_VLEN=128   (for VLEN=128 hardware)\n"
+                "  -DVLLM_RVV_VLEN=256   (for VLEN=256 hardware, e.g. Spacemit X100)\n"
+                "  -DVLLM_RVV_VLEN=0     (force scalar, no RVV)")
+        endif()
+    endif()
+    if(VLLM_RVV_VLEN AND VLLM_RVV_VLEN GREATER 0)
+        message(STATUS "RISC-V RVV VLEN=${VLLM_RVV_VLEN}")
+        if(RVV_BF16_FOUND)
+            message(STATUS "BF16 extension detected")
+            set(MARCH_FLAGS -march=rv64gcv_zvfh_zfbfmin_zvfbfmin_zvl${VLLM_RVV_VLEN}b -mrvv-vector-bits=zvl -mabi=lp64d)
+            add_compile_definitions(RISCV_BF16_SUPPORT)
+        elseif(RVV_FP16_FOUND)
+            message(WARNING "BF16 functionality is not available")
+            set(MARCH_FLAGS -march=rv64gcv_zvfh_zvl${VLLM_RVV_VLEN}b -mrvv-vector-bits=zvl -mabi=lp64d)
+        else()
+            message(STATUS "compile riscv with scalar (no FP16/BF16)")
+            set(MARCH_FLAGS -march=rv64gc)
+        endif()
     else()
         message(STATUS "compile riscv with scalar")
-        list(APPEND CXX_COMPILE_FLAGS "-march=rv64gc")
+        set(MARCH_FLAGS -march=rv64gc)
     endif()
     list(APPEND CXX_COMPILE_FLAGS ${MARCH_FLAGS})
 else()

cmake/external_projects/deepgemm.cmake

@@ -20,7 +20,7 @@ else()
   FetchContent_Declare(
     deepgemm
     GIT_REPOSITORY https://github.com/deepseek-ai/DeepGEMM.git
-    GIT_TAG 477618cd51baffca09c4b0b87e97c03fe827ef03
+    GIT_TAG 891d57b4db1071624b5c8fa0d1e51cb317fa709f
     GIT_SUBMODULES "third-party/cutlass" "third-party/fmt"
     GIT_PROGRESS TRUE
     CONFIGURE_COMMAND ""
@@ -120,6 +120,11 @@ if(DEEPGEMM_ARCHS)
     COMPONENT _deep_gemm_C
     FILES_MATCHING PATTERN "*.py")
 
+  install(DIRECTORY "${deepgemm_SOURCE_DIR}/deep_gemm/mega/"
+    DESTINATION vllm/third_party/deep_gemm/mega
+    COMPONENT _deep_gemm_C
+    FILES_MATCHING PATTERN "*.py")
+
   # Generate envs.py (normally generated by DeepGEMM's setup.py build step)
   file(WRITE "${CMAKE_CURRENT_BINARY_DIR}/deep_gemm_envs.py"
     "# Pre-installed environment variables\npersistent_envs = dict()\n")

cmake/external_projects/flashmla.cmake

@@ -19,7 +19,7 @@ else()
   FetchContent_Declare(
         flashmla
         GIT_REPOSITORY https://github.com/vllm-project/FlashMLA
-        GIT_TAG 692917b1cda61b93ac9ee2d846ec54e75afe87b1
+        GIT_TAG a6ec2ba7bd0a7dff98b3f4d3e6b52b159c48d78b
         GIT_PROGRESS TRUE
         CONFIGURE_COMMAND ""
         BUILD_COMMAND ""

csrc/activation_kernels.cu

@@ -11,29 +11,74 @@
 namespace vllm {
 
 template <typename scalar_t, scalar_t (*ACT_FN)(const scalar_t&),
-          bool act_first>
+          bool act_first, bool HAS_CLAMP>
 __device__ __forceinline__ scalar_t compute(const scalar_t& x,
-                                            const scalar_t& y) {
-  return act_first ? ACT_FN(x) * y : x * ACT_FN(y);
+                                            const scalar_t& y,
+                                            const float limit) {
+  if constexpr (act_first) {
+    scalar_t gate = x;
+    scalar_t up = y;
+    if constexpr (HAS_CLAMP) {
+      gate = (scalar_t)fminf((float)gate, limit);
+      up = (scalar_t)fmaxf(fminf((float)up, limit), -limit);
+    }
+    return ACT_FN(gate) * up;
+  } else {
+    scalar_t gate = x;
+    scalar_t up = y;
+    if constexpr (HAS_CLAMP) {
+      gate = (scalar_t)fmaxf(fminf((float)gate, limit), -limit);
+      up = (scalar_t)fminf((float)up, limit);
+    }
+    return gate * ACT_FN(up);
+  }
 }
 
 template <typename packed_t, packed_t (*PACKED_ACT_FN)(const packed_t&),
-          bool act_first>
+          bool act_first, bool HAS_CLAMP>
 __device__ __forceinline__ packed_t packed_compute(const packed_t& x,
-                                                   const packed_t& y) {
-  return act_first ? packed_mul(PACKED_ACT_FN(x), y)
-                   : packed_mul(x, PACKED_ACT_FN(y));
+                                                   const packed_t& y,
+                                                   const float limit) {
+  if constexpr (act_first) {
+    packed_t gate = x;
+    packed_t up = y;
+    if constexpr (HAS_CLAMP) {
+      float2 g = cast_to_float2(gate);
+      float2 u = cast_to_float2(up);
+      g.x = fminf(g.x, limit);
+      g.y = fminf(g.y, limit);
+      u.x = fmaxf(fminf(u.x, limit), -limit);
+      u.y = fmaxf(fminf(u.y, limit), -limit);
+      gate = cast_to_packed<packed_t>(g);
+      up = cast_to_packed<packed_t>(u);
+    }
+    return packed_mul(PACKED_ACT_FN(gate), up);
+  } else {
+    packed_t gate = x;
+    packed_t up = y;
+    if constexpr (HAS_CLAMP) {
+      float2 g = cast_to_float2(gate);
+      float2 u = cast_to_float2(up);
+      g.x = fmaxf(fminf(g.x, limit), -limit);
+      g.y = fmaxf(fminf(g.y, limit), -limit);
+      u.x = fminf(u.x, limit);
+      u.y = fminf(u.y, limit);
+      gate = cast_to_packed<packed_t>(g);
+      up = cast_to_packed<packed_t>(u);
+    }
+    return packed_mul(gate, PACKED_ACT_FN(up));
+  }
 }
 
 // Activation and gating kernel template.
 template <typename scalar_t, typename packed_t,
           scalar_t (*ACT_FN)(const scalar_t&),
           packed_t (*PACKED_ACT_FN)(const packed_t&), bool act_first,
-          bool use_vec, bool use_256b = false>
+          bool use_vec, bool HAS_CLAMP, bool use_256b = false>
 __global__ void act_and_mul_kernel(
     scalar_t* __restrict__ out,          // [..., d]
     const scalar_t* __restrict__ input,  // [..., 2, d]
-    const int d) {
+    const int d, const float limit) {
   const scalar_t* x_ptr = input + blockIdx.x * 2 * d;
   const scalar_t* y_ptr = x_ptr + d;
   scalar_t* out_ptr = out + blockIdx.x * d;
@@ -58,8 +103,9 @@ __global__ void act_and_mul_kernel(
       }
 #pragma unroll
       for (int j = 0; j < pvec_t::NUM_ELTS; j++) {
-        x.elts[j] = packed_compute<packed_t, PACKED_ACT_FN, act_first>(
-            x.elts[j], y.elts[j]);
+        x.elts[j] =
+            packed_compute<packed_t, PACKED_ACT_FN, act_first, HAS_CLAMP>(
+                x.elts[j], y.elts[j], limit);
       }
       if constexpr (use_256b) {
         st256(x, &out_vec[i]);
@@ -72,7 +118,8 @@ __global__ void act_and_mul_kernel(
     for (int64_t idx = threadIdx.x; idx < d; idx += blockDim.x) {
       const scalar_t x = VLLM_LDG(&x_ptr[idx]);
       const scalar_t y = VLLM_LDG(&y_ptr[idx]);
-      out_ptr[idx] = compute<scalar_t, ACT_FN, act_first>(x, y);
+      out_ptr[idx] =
+          compute<scalar_t, ACT_FN, act_first, HAS_CLAMP>(x, y, limit);
     }
   }
 }
@@ -151,8 +198,11 @@ packed_gelu_tanh_kernel(const packed_t& val) {
 
 // Launch activation and gating kernel.
 // Use ACT_FIRST (bool) indicating whether to apply the activation function
-// first.
-#define LAUNCH_ACTIVATION_GATE_KERNEL(KERNEL, PACKED_KERNEL, ACT_FIRST)        \
+// first. HAS_CLAMP (bool) enables pre-activation clamping: gate input is
+// clamped (max only) and up input is clamped (both sides) before the
+// activation function is applied.
+#define LAUNCH_ACTIVATION_GATE_KERNEL(KERNEL, PACKED_KERNEL, ACT_FIRST,        \
+                                      HAS_CLAMP, LIMIT)                        \
   auto dtype = input.scalar_type();                                            \
   int d = input.size(-1) / 2;                                                  \
   int64_t num_tokens = input.numel() / input.size(-1);                         \
@@ -177,8 +227,8 @@ packed_gelu_tanh_kernel(const packed_t& val) {
             scalar_t, typename vllm::PackedTypeConverter<scalar_t>::Type,      \
             KERNEL<scalar_t>,                                                  \
             PACKED_KERNEL<typename vllm::PackedTypeConverter<scalar_t>::Type>, \
-            ACT_FIRST, true, true><<<grid, block, 0, stream>>>(                \
-            out.data_ptr<scalar_t>(), input.data_ptr<scalar_t>(), d);          \
+            ACT_FIRST, true, HAS_CLAMP, true><<<grid, block, 0, stream>>>(     \
+            out.data_ptr<scalar_t>(), input.data_ptr<scalar_t>(), d, LIMIT);   \
       });                                                                      \
     } else {                                                                   \
       VLLM_DISPATCH_FLOATING_TYPES(dtype, "act_and_mul_kernel", [&] {          \
@@ -186,8 +236,8 @@ packed_gelu_tanh_kernel(const packed_t& val) {
             scalar_t, typename vllm::PackedTypeConverter<scalar_t>::Type,      \
             KERNEL<scalar_t>,                                                  \
             PACKED_KERNEL<typename vllm::PackedTypeConverter<scalar_t>::Type>, \
-            ACT_FIRST, true, false><<<grid, block, 0, stream>>>(               \
-            out.data_ptr<scalar_t>(), input.data_ptr<scalar_t>(), d);          \
+            ACT_FIRST, true, HAS_CLAMP, false><<<grid, block, 0, stream>>>(    \
+            out.data_ptr<scalar_t>(), input.data_ptr<scalar_t>(), d, LIMIT);   \
       });                                                                      \
     }                                                                          \
   } else {                                                                     \
@@ -197,8 +247,8 @@ packed_gelu_tanh_kernel(const packed_t& val) {
           scalar_t, typename vllm::PackedTypeConverter<scalar_t>::Type,        \
           KERNEL<scalar_t>,                                                    \
           PACKED_KERNEL<typename vllm::PackedTypeConverter<scalar_t>::Type>,   \
-          ACT_FIRST, false><<<grid, block, 0, stream>>>(                       \
-          out.data_ptr<scalar_t>(), input.data_ptr<scalar_t>(), d);            \
+          ACT_FIRST, false, HAS_CLAMP><<<grid, block, 0, stream>>>(            \
+          out.data_ptr<scalar_t>(), input.data_ptr<scalar_t>(), d, LIMIT);     \
     });                                                                        \
   }
 
@@ -206,7 +256,14 @@ void silu_and_mul(torch::Tensor& out,    // [..., d]
                   torch::Tensor& input)  // [..., 2 * d]
 {
   LAUNCH_ACTIVATION_GATE_KERNEL(vllm::silu_kernel, vllm::packed_silu_kernel,
-                                true);
+                                true, false, 0.0f);
+}
+
+void silu_and_mul_clamp(torch::Tensor& out,    // [..., d]
+                        torch::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]
@@ -215,21 +272,21 @@ void mul_and_silu(torch::Tensor& out,    // [..., 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.
   LAUNCH_ACTIVATION_GATE_KERNEL(vllm::silu_kernel, vllm::packed_silu_kernel,
-                                false);
+                                false, false, 0.0f);
 }
 
 void gelu_and_mul(torch::Tensor& out,    // [..., d]
                   torch::Tensor& input)  // [..., 2 * d]
 {
   LAUNCH_ACTIVATION_GATE_KERNEL(vllm::gelu_kernel, vllm::packed_gelu_kernel,
-                                true);
+                                true, false, 0.0f);
 }
 
 void gelu_tanh_and_mul(torch::Tensor& out,    // [..., d]
                        torch::Tensor& input)  // [..., 2 * d]
 {
-  LAUNCH_ACTIVATION_GATE_KERNEL(vllm::gelu_tanh_kernel,
-                                vllm::packed_gelu_tanh_kernel, true);
+  LAUNCH_ACTIVATION_GATE_KERNEL(
+      vllm::gelu_tanh_kernel, vllm::packed_gelu_tanh_kernel, true, false, 0.0f);
 }
 
 namespace vllm {

csrc/cache_kernels.cu

@@ -599,6 +599,11 @@ __global__ void cp_gather_indexer_k_quant_cache_kernel(
   const int head_idx = (blockIdx.y * blockDim.x + threadIdx.x) * VEC_SIZE;
   // Find batch index within a block
   __shared__ int batch_idx[BLOCK_Y_SIZE];
+  if (threadIdx.x == 0) {
+    batch_idx[threadIdx.y] = -1;
+  }
+  __syncthreads();
+
   for (int iter = 0; iter < cuda_utils::ceil_div(batch_size, int(blockDim.x));
        iter++) {
     int tid = iter * blockDim.x + threadIdx.x;
@@ -611,16 +616,18 @@ __global__ void cp_gather_indexer_k_quant_cache_kernel(
     }
   }
 
-#ifndef USE_ROCM
-  __syncwarp();
-#endif
+  __syncthreads();
 
-  if (head_idx >= head_dim || token_idx >= num_tokens) {
+  // num_tokens may be an allocation upper bound when Python avoids a D2H sync.
+  // Only tokens covered by the exact device-side cu_seq_lens are valid to
+  // gather.
+  const int batch = batch_idx[threadIdx.y];
+  if (head_idx >= head_dim || token_idx >= num_tokens || batch < 0) {
     return;
   }
-  const int inbatch_seq_idx = token_idx - cu_seq_lens[batch_idx[threadIdx.y]];
-  const int block_idx = block_table[batch_idx[threadIdx.y] * num_blocks +
-                                    inbatch_seq_idx / cache_block_size];
+  const int inbatch_seq_idx = token_idx - cu_seq_lens[batch];
+  const int block_idx =
+      block_table[batch * num_blocks + inbatch_seq_idx / cache_block_size];
   const int64_t src_block_offset = block_idx * block_stride;
   const int64_t cache_inblock_offset =
       (inbatch_seq_idx % cache_block_size) * head_dim + head_idx;
@@ -1490,6 +1497,9 @@ void concat_mla_q(torch::Tensor& ql_nope,  // [num_tokens, num_heads, nope_dim]
   TORCH_CHECK(ql_nope.stride(2) == 1, "ql_nope must have stride 1 in dim 2");
   TORCH_CHECK(q_pe.stride(2) == 1, "q_pe must have stride 1 in dim 2");
   TORCH_CHECK(q_out.stride(2) == 1, "q_out must have stride 1 in dim 2");
+  TORCH_CHECK(ql_nope.scalar_type() == at::ScalarType::Half ||
+                  ql_nope.scalar_type() == at::ScalarType::BFloat16,
+              "ql_nope must be float16 or bfloat16 dtype");
 
   if (num_tokens == 0) return;
 
@@ -1501,7 +1511,7 @@ void concat_mla_q(torch::Tensor& ql_nope,  // [num_tokens, num_heads, nope_dim]
   const at::cuda::OptionalCUDAGuard device_guard(device_of(ql_nope));
   const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
 
-  VLLM_DISPATCH_FLOATING_TYPES(ql_nope.scalar_type(), "concat_mla_q", [&] {
+  VLLM_DISPATCH_HALF_TYPES(ql_nope.scalar_type(), "concat_mla_q", [&] {
     vllm::ConcatMLAQKernel<scalar_t, 512><<<grid_size, block_size, 0, stream>>>(
         q_out.data_ptr<scalar_t>(), ql_nope.data_ptr<scalar_t>(),
         q_pe.data_ptr<scalar_t>(), num_tokens, num_heads, q_out.stride(0),

csrc/cpu/cpu_arch_macros.h

@@ -61,8 +61,23 @@
 #endif
 
 #ifdef __aarch64__
-  // Implementation copied from Arm Optimized Routines (expf AdvSIMD)
+  // Implementation of neon_expf copied from Arm Optimized Routines (expf
+  // AdvSIMD)
   // https://github.com/ARM-software/optimized-routines/blob/master/math/aarch64/advsimd/expf.c
+  //
+  // Additional fast exponential intended for cases where outputs will be
+  // downcasted to FP16 / BF16 (e.g. attention softmax). Accurate within 1 ULP
+  // for FP16 Accurate within 1 ULP for BF16 for inputs in [-87.683, 88.376] &
+  // clamps inputs outside this range to 0 / inf. Implementation is similar to
+  // exp_u20, but:
+  // - uses a third degree polynomial approximation for exp(r) instead of a
+  // fifth degree one, with coefficients re-tuned.
+  // - does not split natural log (ln) into high / low parts
+  // - clamps exp(x) to 0 for x < -87.683113f and inf for x > 88.3762589f
+  // exp(x) = 2^n (exp(r))
+  // r = x - n*ln2, with n = round(x/ln2)
+  // exp(r) ~ poly(r) = 1 + r + r^2 * (c3 + c2 * r)
+  // n = round(x / ln2), r = x - n*ln2
   #include <limits>
   #define DEFINE_FAST_EXP                                                      \
     const float32x4_t inv_ln2 = vdupq_n_f32(0x1.715476p+0f);                   \
@@ -106,7 +121,38 @@
                           result.val[2] = neon_expf(vec.reg.val[2]);           \
                           result.val[3] = neon_expf(vec.reg.val[3]);           \
                           return vec_op::FP32Vec16(result);                    \
-                        };
+                        };                                                     \
+    const float32x4_t lower_bound = vdupq_n_f32(-0x1.5ebb82p+6f);              \
+    const float32x4_t upper_bound = vdupq_n_f32(0x1.61814ap+6f);               \
+    constexpr float ln2 = 0x1.62e43p-1f;                                       \
+    constexpr float f_c2 = 0x1.5592ecp-3f;                                     \
+    const float32x4_t f_c3 = vdupq_n_f32(0x1.017d34p-1f);                      \
+    auto neon_expf_f16 = [&](float32x4_t values) __attribute__((               \
+                             always_inline)) {                                 \
+      const uint32x4_t lt_lower = vcltq_f32(values, lower_bound);              \
+      const uint32x4_t gt_upper = vcgtq_f32(values, upper_bound);              \
+      float32x4_t n = vrndaq_f32(vmulq_f32(values, inv_ln2));                  \
+      float32x4_t r = vfmsq_n_f32(values, n, ln2);                             \
+      uint32x4_t e = vshlq_n_u32(vreinterpretq_u32_s32(vcvtq_s32_f32(n)), 23); \
+      float32x4_t r2 = vmulq_f32(r, r);                                        \
+      float32x4_t q = vfmaq_n_f32(f_c3, r, f_c2);                              \
+      float32x4_t s = vaddq_f32(vdupq_n_f32(1.0f), r);                         \
+      float32x4_t p = vfmaq_f32(s, q, r2);                                     \
+      float32x4_t y =                                                          \
+          vreinterpretq_f32_u32(vaddq_u32(vreinterpretq_u32_f32(p), e));       \
+      y = vbslq_f32(lt_lower, vdupq_n_f32(0.0f), y);                           \
+      y = vbslq_f32(gt_upper, vdupq_n_f32(INFINITY), y);                       \
+      return y;                                                                \
+    };                                                                         \
+    auto fast_exp_f16 = [&](const vec_op::FP32Vec16& vec)                      \
+                            __attribute__((always_inline)) {                   \
+                              float32x4x4_t result;                            \
+                              result.val[0] = neon_expf_f16(vec.reg.val[0]);   \
+                              result.val[1] = neon_expf_f16(vec.reg.val[1]);   \
+                              result.val[2] = neon_expf_f16(vec.reg.val[2]);   \
+                              result.val[3] = neon_expf_f16(vec.reg.val[3]);   \
+                              return vec_op::FP32Vec16(result);                \
+                            };
 
 #endif  // __aarch64__
 

csrc/cpu/cpu_attn.cpp

@@ -1,5 +1,16 @@
 #include "cpu_attn_dispatch_generated.h"
 
+// Maps kv_cache_dtype string to Fp8KVCacheDataType enum.
+// "auto" -> kAuto(0); "fp8"/"fp8_e4m3" -> kFp8E4M3; "fp8_e5m2" -> kFp8E5M2.
+static inline cpu_attention::Fp8KVCacheDataType parse_fp8_kv_dtype(
+    const std::string& kv_cache_dtype) {
+  if (kv_cache_dtype == "fp8_e5m2")
+    return cpu_attention::Fp8KVCacheDataType::kFp8E5M2;
+  if (kv_cache_dtype == "fp8_e4m3" || kv_cache_dtype == "fp8")
+    return cpu_attention::Fp8KVCacheDataType::kFp8E4M3;
+  return cpu_attention::Fp8KVCacheDataType::kAuto;
+}
+
 torch::Tensor get_scheduler_metadata(
     const int64_t num_req, const int64_t num_heads_q,
     const int64_t num_heads_kv, const int64_t head_dim,
@@ -49,7 +60,7 @@ torch::Tensor get_scheduler_metadata(
   input.enable_kv_split = enable_kv_split;
 
   VLLM_DISPATCH_FLOATING_TYPES(dtype, "get_scheduler_metadata", [&]() {
-    CPU_ATTN_DISPATCH(head_dim, isa, [&]() {
+    CPU_ATTN_DISPATCH(head_dim, isa, 0, [&]() {
       input.elem_size = sizeof(scalar_t);
       input.q_buffer_elem_size = sizeof(attn_impl::q_buffer_t);
       input.logits_buffer_elem_size = sizeof(attn_impl::logits_buffer_t);
@@ -72,7 +83,9 @@ void cpu_attn_reshape_and_cache(
         key_cache,  // [num_blocks, num_kv_heads, block_size, head_size]
     torch::Tensor&
         value_cache,  // [num_blocks, num_kv_heads, block_size, head_size]
-    const torch::Tensor& slot_mapping, const std::string& isa) {
+    const torch::Tensor& slot_mapping, const std::string& isa,
+    const double k_scale = 1.0, const double v_scale = 1.0,
+    const std::string& kv_cache_dtype = "auto") {
   TORCH_CHECK_EQ(key.dim(), 3);
   TORCH_CHECK_EQ(value.dim(), 3);
   TORCH_CHECK_EQ(key_cache.dim(), 4);
@@ -80,18 +93,30 @@ void cpu_attn_reshape_and_cache(
   TORCH_CHECK_EQ(key.stride(2), 1);
   TORCH_CHECK_EQ(value.stride(2), 1);
 
+  const int64_t kv_cache_idx =
+      static_cast<int64_t>(parse_fp8_kv_dtype(kv_cache_dtype));
+  const bool is_fp8 = (kv_cache_idx != 0);
+
+  if (is_fp8) {
+    TORCH_CHECK(key_cache.scalar_type() == at::ScalarType::Byte,
+                "key_cache must be uint8 for FP8 path");
+    TORCH_CHECK(value_cache.scalar_type() == at::ScalarType::Byte,
+                "value_cache must be uint8 for FP8 path");
+    TORCH_CHECK(k_scale > 0, "k_scale must be positive for FP8 path");
+    TORCH_CHECK(v_scale > 0, "v_scale must be positive for FP8 path");
+  }
+
+  const float k_inv = is_fp8 ? 1.0f / static_cast<float>(k_scale) : 0.0f;
+  const float v_inv = is_fp8 ? 1.0f / static_cast<float>(v_scale) : 0.0f;
+
   const int64_t token_num = key.size(0);
-  const int64_t key_token_num_stride = key.stride(0);
-  const int64_t value_token_num_stride = value.stride(0);
-  const int64_t head_num = value.size(1);
-  const int64_t key_head_num_stride = key.stride(1);
-  const int64_t value_head_num_stride = value.stride(1);
+  const int64_t head_num = key.size(1);
+  const int64_t head_dim = key.size(2);
   const int64_t num_blocks = key_cache.size(0);
   const int64_t num_blocks_stride = key_cache.stride(0);
   const int64_t cache_head_num_stride = key_cache.stride(1);
   const int64_t block_size = key_cache.size(2);
   const int64_t block_size_stride = key_cache.stride(2);
-  const int64_t head_dim = key.size(-1);
 
   cpu_attention::ISA isa_tag = [&]() {
     if (isa == "amx") {
@@ -109,16 +134,24 @@ void cpu_attn_reshape_and_cache(
     }
   }();
 
+  if (is_fp8) {
+    TORCH_CHECK(isa_tag == cpu_attention::ISA::AMX ||
+                    isa_tag == cpu_attention::ISA::VEC,
+                "FP8 KV cache is only supported on x86 (AMX/VEC) ISA");
+  }
+
   VLLM_DISPATCH_FLOATING_TYPES(
       key.scalar_type(), "cpu_attn_reshape_and_cache", [&]() {
-        CPU_ATTN_DISPATCH(head_dim, isa_tag, [&]() {
+        CPU_ATTN_DISPATCH(head_dim, isa_tag, kv_cache_idx, [&]() {
+          using kv_t = typename attn_impl::kv_cache_t;
           attn_impl::reshape_and_cache(
               key.data_ptr<scalar_t>(), value.data_ptr<scalar_t>(),
-              key_cache.data_ptr<scalar_t>(), value_cache.data_ptr<scalar_t>(),
-              slot_mapping.data_ptr<int64_t>(), token_num, key_token_num_stride,
-              value_token_num_stride, head_num, key_head_num_stride,
-              value_head_num_stride, num_blocks, num_blocks_stride,
-              cache_head_num_stride, block_size, block_size_stride);
+              reinterpret_cast<kv_t*>(key_cache.data_ptr()),
+              reinterpret_cast<kv_t*>(value_cache.data_ptr()),
+              slot_mapping.data_ptr<int64_t>(), token_num, key.stride(0),
+              value.stride(0), head_num, key.stride(1), value.stride(1),
+              num_blocks, num_blocks_stride, cache_head_num_stride, block_size,
+              block_size_stride, k_inv, v_inv);
         });
       });
 }
@@ -137,13 +170,26 @@ void cpu_attention_with_kv_cache(
     const int64_t sliding_window_left, const int64_t sliding_window_right,
     const torch::Tensor& block_table,  // [num_tokens, max_block_num]
     const double softcap, const torch::Tensor& scheduler_metadata,
-    const std::optional<torch::Tensor>& s_aux  // [num_heads]
-) {
+    const std::optional<torch::Tensor>& s_aux,  // [num_heads]
+    const double k_scale = 1.0, const double v_scale = 1.0,
+    const std::string& kv_cache_dtype = "auto") {
   TORCH_CHECK_EQ(query.dim(), 3);
   TORCH_CHECK_EQ(query.stride(2), 1);
   TORCH_CHECK_EQ(key_cache.dim(), 4);
   TORCH_CHECK_EQ(value_cache.dim(), 4);
 
+  const int64_t kv_cache_idx =
+      static_cast<int64_t>(parse_fp8_kv_dtype(kv_cache_dtype));
+  const bool is_fp8 = (kv_cache_idx != 0);
+  if (is_fp8) {
+    TORCH_CHECK(key_cache.scalar_type() == at::ScalarType::Byte,
+                "key_cache must be uint8 for FP8 path");
+    TORCH_CHECK(value_cache.scalar_type() == at::ScalarType::Byte,
+                "value_cache must be uint8 for FP8 path");
+    TORCH_CHECK(k_scale > 0, "k_scale must be positive for FP8 path");
+    TORCH_CHECK(v_scale > 0, "v_scale must be positive for FP8 path");
+  }
+
   cpu_attention::AttentionInput input;
   input.metadata = reinterpret_cast<cpu_attention::AttentionMetadata*>(
       scheduler_metadata.data_ptr());
@@ -165,25 +211,32 @@ void cpu_attention_with_kv_cache(
   input.block_table = block_table.data_ptr<int32_t>();
   input.alibi_slopes =
       alibi_slopes.has_value() ? alibi_slopes->data_ptr<float>() : nullptr;
-  // For now sink must be bf16
   input.s_aux = s_aux.has_value() ? s_aux->data_ptr<c10::BFloat16>() : nullptr;
   input.scale = scale;
   input.causal = causal;
   input.sliding_window_left = sliding_window_left;
   input.sliding_window_right = sliding_window_right;
   if (input.causal) {
-    // to make boundary calculation easier
     input.sliding_window_right = 0;
   }
-  float softcap_fp32 = softcap;
-  input.softcap = softcap_fp32;
+  input.softcap = static_cast<float>(softcap);
+
+  if (is_fp8) {
+    input.k_scale_fp8 = static_cast<float>(k_scale);
+    input.v_scale_fp8 = static_cast<float>(v_scale);
+    TORCH_CHECK(input.metadata->isa == cpu_attention::ISA::AMX ||
+                    input.metadata->isa == cpu_attention::ISA::VEC,
+                "FP8 KV cache is only supported on x86 (AMX/VEC) ISA");
+  }
 
   VLLM_DISPATCH_FLOATING_TYPES(
       query.scalar_type(), "cpu_attention_with_kv_cache", [&]() {
-        CPU_ATTN_DISPATCH(query.size(2), input.metadata->isa, [&]() {
-          TORCH_CHECK_EQ(input.block_size % attn_impl::BlockSizeAlignment, 0);
-          cpu_attention::AttentionMainLoop<attn_impl> mainloop;
-          mainloop(&input);
-        });
+        CPU_ATTN_DISPATCH(
+            query.size(2), input.metadata->isa, kv_cache_idx, [&]() {
+              TORCH_CHECK_EQ(input.block_size % attn_impl::BlockSizeAlignment,
+                             0);
+              cpu_attention::AttentionMainLoop<attn_impl> mainloop;
+              mainloop(&input);
+            });
       });
 }

csrc/cpu/cpu_attn_amx.hpp

@@ -1,6 +1,7 @@
 #ifndef CPU_ATTN_AMX_HPP
 #define CPU_ATTN_AMX_HPP
 
+#include "cpu_attn_fp8.hpp"
 #include "cpu_attn_impl.hpp"
 
 namespace cpu_attention {
@@ -21,9 +22,10 @@ typedef struct __tile_config {
 // 2-2-4 pattern, for 16 < m <= 32
 // TILE 0, 1: load A matrix, row num should be 16, m - 16
 // TILE 2, 3: load B matrix, row num should be 16
-// TILE 4, 5, 6, 7: store results C matrix, row num should be 16, 16, m - 16, m
-// - 16
-template <typename kv_cache_t>
+// TILE 4, 5, 6, 7: store results C matrix, row num should be 16, 16,
+// m - 16, m - 16
+// q_buffer_t: A (Q/P) tile type; kv_cache_t: B (K/V cache) tile type.
+template <typename q_buffer_t, typename kv_cache_t>
 class TileGemm224 {
  public:
   template <AttentionGemmPhase phase, int32_t k_size>
@@ -42,13 +44,56 @@ class TileGemm224 {
   }
 };
 
-template <>
-class TileGemm224<c10::BFloat16> {
+// Dequantize one FP8 tile (AMX_TILE_ROW_NUM rows x 32 cols) to BF16.
+template <typename kv_cache_t>
+FORCE_INLINE void deq_tile_amx(const uint8_t* src, c10::BFloat16* dst) {
+  for (int r = 0; r < AMX_TILE_ROW_NUM; ++r) {
+    if constexpr (std::is_same_v<kv_cache_t, c10::Float8_e4m3fn>) {
+      vec_op::BF16Vec32(src + r * 32, vec_op::fp8_bf16_e4m3_tag{})
+          .save(dst + r * 32);
+    } else {
+      vec_op::BF16Vec32(src + r * 32, vec_op::fp8_bf16_e5m2_tag{})
+          .save(dst + r * 32);
+    }
+  }
+}
+
+// For FP8: dequant src into scratch and return scratch.
+// For BF16: return src directly (scratch is unused; the compiler elides it).
+template <typename kv_cache_t>
+FORCE_INLINE const c10::BFloat16* prepare_b_tile(const kv_cache_t* src,
+                                                 c10::BFloat16* scratch) {
+  if constexpr (std::is_same_v<kv_cache_t, c10::Float8_e4m3fn> ||
+                std::is_same_v<kv_cache_t, c10::Float8_e5m2>) {
+    deq_tile_amx<kv_cache_t>(reinterpret_cast<const uint8_t*>(src), scratch);
+    return scratch;
+  } else {
+    return reinterpret_cast<const c10::BFloat16*>(src);
+  }
+}
+
+// Handles both BF16 and FP8 KV cache (2-2-4 pattern).
+template <typename kv_cache_t>
+class TileGemm224<c10::BFloat16, kv_cache_t> {
+  static_assert(std::is_same_v<kv_cache_t, c10::BFloat16> ||
+                    std::is_same_v<kv_cache_t, c10::Float8_e4m3fn> ||
+                    std::is_same_v<kv_cache_t, c10::Float8_e5m2>,
+                "kv_cache_t must be BFloat16, Float8_e4m3fn, or Float8_e5m2");
+
+  static constexpr bool fp8_kv =
+      std::is_same_v<kv_cache_t, c10::Float8_e4m3fn> ||
+      std::is_same_v<kv_cache_t, c10::Float8_e5m2>;
+
+  static constexpr int64_t tile_elems = AMX_TILE_BYTES / sizeof(c10::BFloat16);
+  // BF16 path: scratch_elems=1 so the scratch array is eliminated by the
+  // compiler.
+  static constexpr int64_t scratch_elems = fp8_kv ? tile_elems : 1;
+
  public:
   template <AttentionGemmPhase phase, int32_t k_size>
   FORCE_INLINE static void gemm(const int32_t m_size,
                                 c10::BFloat16* __restrict__ a_tile,
-                                c10::BFloat16* __restrict__ b_tile,
+                                kv_cache_t* __restrict__ b_tile,
                                 float* __restrict__ c_tile, const int64_t lda,
                                 const int64_t ldb, const int64_t ldc,
                                 const int32_t block_size,
@@ -56,6 +101,7 @@ class TileGemm224<c10::BFloat16> {
                                 const bool accum_c) {
     const int32_t k_times =
         dynamic_k_size / (AMX_TILE_ROW_NUM * 4 / sizeof(c10::BFloat16));
+
     c10::BFloat16* __restrict__ a_tile_0 = a_tile;
     c10::BFloat16* __restrict__ a_tile_1 = a_tile + lda * AMX_TILE_ROW_NUM;
     const int64_t a_tile_stride = [&]() {
@@ -70,8 +116,8 @@ class TileGemm224<c10::BFloat16> {
       }
     }();
 
-    c10::BFloat16* __restrict__ b_tile_2 = b_tile;
-    c10::BFloat16* __restrict__ b_tile_3 = [&]() {
+    kv_cache_t* __restrict__ b_tile_2 = b_tile;
+    kv_cache_t* __restrict__ b_tile_3 = [&]() {
       if constexpr (phase == AttentionGemmPhase::QK) {
         // k_cache is prepacked
         return b_tile + (k_size * AMX_TILE_ROW_BYTES / 4);
@@ -106,11 +152,16 @@ class TileGemm224<c10::BFloat16> {
       _tile_zero(7);
     }
 
+    alignas(64) c10::BFloat16 scratch_2[scratch_elems];
+    alignas(64) c10::BFloat16 scratch_3[scratch_elems];
     for (int32_t k = 0; k < k_times; ++k) {
+      const c10::BFloat16* load_2 = prepare_b_tile(b_tile_2, scratch_2);
+      const c10::BFloat16* load_3 = prepare_b_tile(b_tile_3, scratch_3);
+
       _tile_loadd(0, a_tile_0, a_tile_stride);
-      _tile_stream_loadd(2, b_tile_2, b_tile_stride);
+      _tile_stream_loadd(2, const_cast<c10::BFloat16*>(load_2), b_tile_stride);
       _tile_dpbf16ps(4, 0, 2);
-      _tile_stream_loadd(3, b_tile_3, b_tile_stride);
+      _tile_stream_loadd(3, const_cast<c10::BFloat16*>(load_3), b_tile_stride);
       _tile_dpbf16ps(5, 0, 3);
       _tile_loadd(1, a_tile_1, a_tile_stride);
       _tile_dpbf16ps(6, 1, 2);
@@ -154,13 +205,13 @@ class TileGemm224<c10::BFloat16> {
 };
 
 // 1-2-2 pattern, for 0 < m <= 16
-// TILE 0, (1): load A matrix, use extra 1 tile for prefetch, row num should be
-// m, m
-// TILE 2, 3, (4, 5): load B matrix, use extra 2 tiles for prefetch, row
-// num should be 16
-// TILE 6, 7, (6, 7): store results C matrix, row num should be
-// m
-template <typename kv_cache_t>
+// TILE 0, (1): load A matrix, use extra 1 tile for prefetch, row num should
+// be m, m
+// TILE 2, 3, (4, 5): load B matrix, use extra 2 tiles for prefetch, row num
+// should be 16
+// TILE 6, 7: store results C matrix, row num should be m
+// q_buffer_t: A (Q/P) tile type; kv_cache_t: B (K/V cache) tile type.
+template <typename q_buffer_t, typename kv_cache_t>
 class TileGemm122 {
  public:
   template <AttentionGemmPhase phase, int32_t k_size>
@@ -179,13 +230,26 @@ class TileGemm122 {
   }
 };
 
-template <>
-class TileGemm122<c10::BFloat16> {
+// Handles both BF16 and FP8 KV cache (1-2-2 pattern).
+template <typename kv_cache_t>
+class TileGemm122<c10::BFloat16, kv_cache_t> {
+  static_assert(std::is_same_v<kv_cache_t, c10::BFloat16> ||
+                    std::is_same_v<kv_cache_t, c10::Float8_e4m3fn> ||
+                    std::is_same_v<kv_cache_t, c10::Float8_e5m2>,
+                "kv_cache_t must be BFloat16, Float8_e4m3fn, or Float8_e5m2");
+
+  static constexpr bool fp8_kv =
+      std::is_same_v<kv_cache_t, c10::Float8_e4m3fn> ||
+      std::is_same_v<kv_cache_t, c10::Float8_e5m2>;
+
+  static constexpr int64_t tile_elems = AMX_TILE_BYTES / sizeof(c10::BFloat16);
+  static constexpr int64_t scratch_elems = fp8_kv ? tile_elems : 1;
+
  public:
   template <AttentionGemmPhase phase, int32_t k_size>
   FORCE_INLINE static void gemm(const int32_t m_size,
                                 c10::BFloat16* __restrict__ a_tile,
-                                c10::BFloat16* __restrict__ b_tile,
+                                kv_cache_t* __restrict__ b_tile,
                                 float* __restrict__ c_tile, const int64_t lda,
                                 const int64_t ldb, const int64_t ldc,
                                 const int32_t block_size,
@@ -215,21 +279,19 @@ class TileGemm122<c10::BFloat16> {
       }
     }();
 
-    c10::BFloat16* __restrict__ b_tile_2 = b_tile;
-    c10::BFloat16* __restrict__ b_tile_3 = [&]() {
+    kv_cache_t* __restrict__ b_tile_2 = b_tile;
+    kv_cache_t* __restrict__ b_tile_3 = [&]() {
       if constexpr (phase == AttentionGemmPhase::QK) {
-        // k_cache is prepacked
         return b_tile + (k_size * AMX_TILE_ROW_BYTES / 4);
       } else if constexpr (phase == AttentionGemmPhase::PV) {
-        // v_cache is prepacked
         return b_tile + (block_size * AMX_TILE_ROW_BYTES / 4);
       } else {
         TORCH_CHECK(false, "Unreachable");
       }
     }();
-    c10::BFloat16* __restrict__ b_tile_4 =
+    kv_cache_t* __restrict__ b_tile_4 =
         b_tile_2 + AMX_TILE_BYTES / sizeof(c10::BFloat16);
-    c10::BFloat16* __restrict__ b_tile_5 =
+    kv_cache_t* __restrict__ b_tile_5 =
         b_tile_3 + AMX_TILE_BYTES / sizeof(c10::BFloat16);
     int64_t b_stride = AMX_TILE_ROW_BYTES;
 
@@ -250,16 +312,25 @@ class TileGemm122<c10::BFloat16> {
       _tile_zero(7);
     }
 
+    alignas(64) c10::BFloat16 scratch_2[scratch_elems];
+    alignas(64) c10::BFloat16 scratch_3[scratch_elems];
+    alignas(64) c10::BFloat16 scratch_4[scratch_elems];
+    alignas(64) c10::BFloat16 scratch_5[scratch_elems];
     for (int32_t k = 0; k < k_group_times; ++k) {
+      const c10::BFloat16* load_2 = prepare_b_tile(b_tile_2, scratch_2);
+      const c10::BFloat16* load_3 = prepare_b_tile(b_tile_3, scratch_3);
+      const c10::BFloat16* load_4 = prepare_b_tile(b_tile_4, scratch_4);
+      const c10::BFloat16* load_5 = prepare_b_tile(b_tile_5, scratch_5);
+
       _tile_loadd(0, a_tile_0, a_tile_stride);
-      _tile_stream_loadd(2, b_tile_2, b_stride);
+      _tile_stream_loadd(2, const_cast<c10::BFloat16*>(load_2), b_stride);
       _tile_dpbf16ps(6, 0, 2);
-      _tile_stream_loadd(3, b_tile_3, b_stride);
+      _tile_stream_loadd(3, const_cast<c10::BFloat16*>(load_3), b_stride);
       _tile_dpbf16ps(7, 0, 3);
       _tile_loadd(1, a_tile_1, a_tile_stride);
-      _tile_stream_loadd(4, b_tile_4, b_stride);
+      _tile_stream_loadd(4, const_cast<c10::BFloat16*>(load_4), b_stride);
       _tile_dpbf16ps(6, 1, 4);
-      _tile_stream_loadd(5, b_tile_5, b_stride);
+      _tile_stream_loadd(5, const_cast<c10::BFloat16*>(load_5), b_stride);
       _tile_dpbf16ps(7, 1, 5);
 
       // update ptrs
@@ -279,10 +350,13 @@ class TileGemm122<c10::BFloat16> {
     }
 
     if (has_tail) {
+      const c10::BFloat16* load_2 = prepare_b_tile(b_tile_2, scratch_2);
+      const c10::BFloat16* load_3 = prepare_b_tile(b_tile_3, scratch_3);
+
       _tile_loadd(0, a_tile_0, a_tile_stride);
-      _tile_stream_loadd(2, b_tile_2, b_stride);
+      _tile_stream_loadd(2, const_cast<c10::BFloat16*>(load_2), b_stride);
       _tile_dpbf16ps(6, 0, 2);
-      _tile_stream_loadd(3, b_tile_3, b_stride);
+      _tile_stream_loadd(3, const_cast<c10::BFloat16*>(load_3), b_stride);
       _tile_dpbf16ps(7, 0, 3);
     }
 
@@ -302,21 +376,25 @@ class TileGemm122<c10::BFloat16> {
     _tile_loadconfig(&config);
   }
 };
+
 }  // namespace
 
-template <typename scalar_t, int64_t head_dim>
-class AttentionImpl<ISA::AMX, scalar_t, head_dim> {
+template <typename scalar_t, int64_t head_dim, typename kv_cache_scalar_t>
+class AttentionImpl<ISA::AMX, scalar_t, head_dim, kv_cache_scalar_t> {
+  static constexpr bool fp8_kv =
+      std::is_same_v<kv_cache_scalar_t, c10::Float8_e4m3fn> ||
+      std::is_same_v<kv_cache_scalar_t, c10::Float8_e5m2>;
+
  public:
   using query_t = scalar_t;
   using q_buffer_t = scalar_t;
-  using kv_cache_t = scalar_t;
+  using kv_cache_t = kv_cache_scalar_t;
   using logits_buffer_t = float;
   using partial_output_buffer_t = float;
   using prob_buffer_t = scalar_t;
 
   constexpr static int64_t BlockSizeAlignment =
-      AMX_TILE_ROW_BYTES /
-      sizeof(kv_cache_t);  // KV token num unit of QK and PV phases
+      32;  // AMX_TILE_ROW_NUM = 16 tokens/tile; 32 = 2 tiles
   constexpr static int64_t HeadDimAlignment =
       2 * (AMX_TILE_ROW_BYTES / 4);  // headdim num unit of PV phase
   constexpr static int64_t MaxQHeadNumPerIteration = 32;
@@ -324,6 +402,9 @@ class AttentionImpl<ISA::AMX, scalar_t, head_dim> {
   constexpr static ISA ISAType = ISA::AMX;
   constexpr static bool scale_on_logits = true;
 
+  float k_scale = 1.0f;
+  float v_scale = 1.0f;
+
  public:
   AttentionImpl() : current_q_head_num_(0) {
     // Use all columns in AMX tiles
@@ -332,21 +413,50 @@ class AttentionImpl<ISA::AMX, scalar_t, head_dim> {
 
   ~AttentionImpl() { _tile_release(); }
 
+  void init_from_input(const AttentionInput* input) {
+    if constexpr (fp8_kv) {
+      k_scale = input->k_scale_fp8;
+      v_scale = input->v_scale_fp8;
+    }
+  }
+
+  float get_output_v_scale() const noexcept {
+    if constexpr (fp8_kv) {
+      // AMX dequant places FP8 payload into a BF16 field (exponent bias 127).
+      // Correction = 2^(127 - FP8_bias): E4M3 bias=7 → 2^120, E5M2 bias=15 →
+      // 2^112.
+      constexpr float bias =
+          std::is_same_v<kv_cache_t, c10::Float8_e5m2> ? 0x1p112f : 0x1p120f;
+      return v_scale * bias;
+    }
+    return 1.0f;
+  }
+
   template <template <typename tile_gemm_t> typename attention>
   FORCE_INLINE void execute_attention(DEFINE_CPU_ATTENTION_PARAMS) {
+    if constexpr (fp8_kv) {
+      // Same bias correction as get_output_v_scale: AMX FP8→BF16 dequant
+      // shifts the exponent bias from FP8 to BF16 (127), so we multiply by
+      // 2^(127-FP8_bias) to recover the true value. E4M3: 2^120, E5M2: 2^112.
+      const float bias =
+          std::is_same_v<kv_cache_t, c10::Float8_e5m2> ? 0x1p112f : 0x1p120f;
+      scale *= k_scale * bias;
+    }
     if (q_head_num > AMX_TILE_ROW_NUM) {
       if (q_head_num != current_q_head_num_) {
         current_q_head_num_ = q_head_num;
-        TileGemm224<kv_cache_t>::init_tile_config(q_head_num, amx_tile_config_);
+        TileGemm224<q_buffer_t, kv_cache_t>::init_tile_config(q_head_num,
+                                                              amx_tile_config_);
       }
-      attention<TileGemm224<kv_cache_t>> attention_iteration;
+      attention<TileGemm224<q_buffer_t, kv_cache_t>> attention_iteration;
       attention_iteration(CPU_ATTENTION_PARAMS);
     } else {
       if (q_head_num != current_q_head_num_) {
         current_q_head_num_ = q_head_num;
-        TileGemm122<kv_cache_t>::init_tile_config(q_head_num, amx_tile_config_);
+        TileGemm122<q_buffer_t, kv_cache_t>::init_tile_config(q_head_num,
+                                                              amx_tile_config_);
       }
-      attention<TileGemm122<kv_cache_t>> attention_iteration;
+      attention<TileGemm122<q_buffer_t, kv_cache_t>> attention_iteration;
       attention_iteration(CPU_ATTENTION_PARAMS);
     }
   }
@@ -411,13 +521,26 @@ class AttentionImpl<ISA::AMX, scalar_t, head_dim> {
   // reshape KV to AMX friendly layout
   static void reshape_and_cache(
       const scalar_t* __restrict__ key, const scalar_t* __restrict__ value,
-      scalar_t* __restrict__ key_cache, scalar_t* __restrict__ value_cache,
+      kv_cache_t* __restrict__ key_cache, kv_cache_t* __restrict__ value_cache,
       const int64_t* __restrict__ slot_mapping, const int64_t token_num,
       const int64_t key_token_num_stride, const int64_t value_token_num_stride,
       const int64_t head_num, const int64_t key_head_num_stride,
       const int64_t value_head_num_stride, const int64_t num_blocks,
       const int64_t num_blocks_stride, const int64_t cache_head_num_stride,
-      const int64_t block_size, const int64_t block_size_stride) {
+      const int64_t block_size, const int64_t block_size_stride,
+      const float k_inv = 0.0f, const float v_inv = 0.0f) {
+    if constexpr (fp8_kv) {
+      constexpr auto qfn = select_fp8_quant_fn<kv_cache_t>();
+      reshape_and_cache_fp8_amx_impl<scalar_t, qfn>(
+          key, value, reinterpret_cast<uint8_t*>(key_cache),
+          reinterpret_cast<uint8_t*>(value_cache), slot_mapping, token_num,
+          head_num, head_dim, block_size, key_token_num_stride,
+          key_head_num_stride, value_token_num_stride, value_head_num_stride,
+          num_blocks_stride, cache_head_num_stride, num_blocks_stride,
+          cache_head_num_stride, k_inv, v_inv);
+      return;
+    }
+
     // For AMX 2D tiles, size of each line is 64 bytes
     constexpr int64_t amx_tile_row_size = AMX_TILE_ROW_BYTES;
     // For AMX B matrix, N always is 16
@@ -426,6 +549,9 @@ class AttentionImpl<ISA::AMX, scalar_t, head_dim> {
     // For now suppose block_size is divisible by amx_tile_column_num
     TORCH_CHECK_EQ(block_size % amx_b_tile_k_size, 0);
 
+    scalar_t* __restrict__ kc = reinterpret_cast<scalar_t*>(key_cache);
+    scalar_t* __restrict__ vc = reinterpret_cast<scalar_t*>(value_cache);
+
 #pragma omp parallel for collapse(2)
     for (int64_t token_idx = 0; token_idx < token_num; ++token_idx) {
       for (int64_t head_idx = 0; head_idx < head_num; ++head_idx) {
@@ -453,8 +579,7 @@ class AttentionImpl<ISA::AMX, scalar_t, head_dim> {
           constexpr int64_t quadword_num_per_group =
               token_num_per_group * quadword_num;
           int32_t* key_cache_start_ptr =
-              reinterpret_cast<int32_t*>(key_cache +
-                                         block_idx * num_blocks_stride +
+              reinterpret_cast<int32_t*>(kc + block_idx * num_blocks_stride +
                                          head_idx * cache_head_num_stride) +
               group_idx * quadword_num_per_group + group_offset;
 
@@ -483,7 +608,7 @@ class AttentionImpl<ISA::AMX, scalar_t, head_dim> {
                                             token_idx * value_token_num_stride +
                                             head_idx * value_head_num_stride;
           scalar_t* value_cache_start_ptr =
-              value_cache + block_idx * num_blocks_stride +
+              vc + block_idx * num_blocks_stride +
               head_idx * cache_head_num_stride +
               sub_group_idx * token_num_per_sub_group * amx_b_tile_n_size +
               sub_group_offset;

csrc/cpu/cpu_attn_fp8.hpp

@@ -0,0 +1,214 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+#pragma once
+#include <algorithm>
+#include <cmath>
+#include <cstdint>
+#include <limits>
+#include <type_traits>
+
+#include "cpu/utils.hpp"
+
+typedef uint32_t __attribute__((__may_alias__)) u32_alias_t;
+typedef uint16_t __attribute__((__may_alias__)) u16_alias_t;
+typedef float __attribute__((__may_alias__)) f32_alias_t;
+
+// Reference scalar dequant — used to verify vectorized AMX dequant.
+inline float fp8e4m3_to_float_scalar(uint8_t b, float scale) noexcept {
+  // NaN encoding in E4M3
+  if ((b & 0x7F) == 0x7F) return std::numeric_limits<float>::quiet_NaN();
+  uint32_t b_u32 = static_cast<uint32_t>(b);
+  uint32_t sign = (b_u32 & 0x80) << 24;
+  uint32_t payload = (b_u32 & 0x7F) << 20;
+  uint32_t bits = sign | payload;
+  float b_f32_unscaled = *reinterpret_cast<const f32_alias_t*>(&bits);
+  float b_f32_scaled = b_f32_unscaled * scale * 0x1p120f;
+  return b_f32_scaled;
+}
+
+inline uint8_t float_to_fp8e4m3_scalar(float v, float inv_scale) noexcept {
+  v *= inv_scale;
+  constexpr float fp8_max = 448.0f;
+  v = std::max(-fp8_max, std::min(fp8_max, v));
+  if (v == 0.0f) return 0;
+
+  // Inverse mapping of fp8e4m3_to_float_scalar: shift the effective exponent
+  // bias from fp32 (127) back to fp8 e4m3 (7), then pack sign|payload.
+  float v_f32_unscaled = v * 0x1p-120f;
+  uint32_t bits = *reinterpret_cast<const u32_alias_t*>(&v_f32_unscaled);
+  uint8_t sign = static_cast<uint8_t>((bits >> 24) & 0x80);
+  uint8_t payload = static_cast<uint8_t>((bits >> 20) & 0x7F);
+  if (payload == 0) return sign;
+  payload = std::min<uint8_t>(payload, 0x7E);  // keep 0x7F as NaN encoding
+  return static_cast<uint8_t>(sign | payload);
+}
+
+// ---------------------------------------------------------------------------
+// AMX reshape impl — parameterised on the quantisation function.
+// Writes key/value into uint8 FP8 KV cache using the AMX tile-friendly layout.
+// K: halfword-packed (2 FP8 per uint16, token_num_per_group=16).
+// V: sub-group packing (token_num_per_sub_group=2, head_elems_per_group=16).
+// block_size must be divisible by 32.
+// ---------------------------------------------------------------------------
+template <typename scalar_t, uint8_t (*quant_fn)(float, float)>
+inline void reshape_and_cache_fp8_amx_impl(
+    const scalar_t* key_ptr, const scalar_t* value_ptr, uint8_t* key_cache_ptr,
+    uint8_t* value_cache_ptr, const int64_t* slot_ptr, int64_t token_num,
+    int64_t head_num, int64_t head_dim, int64_t block_size, int64_t k_stride0,
+    int64_t k_stride1, int64_t v_stride0, int64_t v_stride1, int64_t kc_stride0,
+    int64_t kc_stride1, int64_t vc_stride0, int64_t vc_stride1, float k_inv,
+    float v_inv) {
+  constexpr int64_t token_num_per_group = 16;  // AMX_TILE_ROW_NUM
+  const int64_t halfword_num = head_dim / 2;   // 2 FP8 per uint16
+  const int64_t halfword_num_per_group = token_num_per_group * halfword_num;
+  constexpr int64_t head_elems_per_group = 16;
+  constexpr int64_t token_num_per_sub_group = 2;  // = 4 / sizeof(BF16)
+  const int64_t group_num = head_dim / head_elems_per_group;
+  const int64_t group_size = block_size * head_elems_per_group;
+
+#pragma omp parallel for collapse(2) schedule(static)
+  for (int64_t tok = 0; tok < token_num; ++tok) {
+    for (int64_t h = 0; h < head_num; ++h) {
+      const int64_t slot = slot_ptr[tok];
+      if (slot < 0) continue;
+      const int64_t block_idx = slot / block_size;
+      const int64_t block_offset = slot % block_size;
+
+      // Key: halfword-packed, 2 FP8 per uint16
+      {
+        const scalar_t* ksrc = key_ptr + tok * k_stride0 + h * k_stride1;
+        const int64_t group_idx = block_offset / token_num_per_group;
+        const int64_t group_offset = block_offset % token_num_per_group;
+        uint16_t* kdst =
+            reinterpret_cast<uint16_t*>(key_cache_ptr + block_idx * kc_stride0 +
+                                        h * kc_stride1) +
+            group_idx * halfword_num_per_group + group_offset;
+        for (int64_t j = 0; j < halfword_num; ++j) {
+          uint8_t fp8_0 = quant_fn(static_cast<float>(ksrc[j * 2]), k_inv);
+          uint8_t fp8_1 = quant_fn(static_cast<float>(ksrc[j * 2 + 1]), k_inv);
+          uint8_t bytes[2] = {fp8_0, fp8_1};
+          uint16_t hw = *reinterpret_cast<const u16_alias_t*>(bytes);
+          kdst[j * token_num_per_group] = hw;
+        }
+      }
+
+      // Value: sub-group packing (token_num_per_sub_group = 2)
+      {
+        const scalar_t* vsrc = value_ptr + tok * v_stride0 + h * v_stride1;
+        const int64_t sub_group_idx = block_offset / token_num_per_sub_group;
+        const int64_t sub_group_offset = block_offset % token_num_per_sub_group;
+        uint8_t* vdst =
+            value_cache_ptr + block_idx * vc_stride0 + h * vc_stride1 +
+            sub_group_idx * token_num_per_sub_group * head_elems_per_group +
+            sub_group_offset;
+        for (int64_t i = 0; i < group_num; ++i) {
+          for (int64_t j = 0; j < head_elems_per_group; ++j)
+            vdst[j * token_num_per_sub_group] =
+                quant_fn(static_cast<float>(vsrc[j]), v_inv);
+          vsrc += head_elems_per_group;
+          vdst += group_size;
+        }
+      }
+    }
+  }
+}
+
+// ---------------------------------------------------------------------------
+// FP8 E5M2 scalar helpers
+// ---------------------------------------------------------------------------
+
+// Reference scalar dequant — used to verify vectorized AMX dequant.
+// FP8 E5M2: s[7] e[6:2] m[1:0], exponent bias = 15 (same as FP16).
+// Byte b → FP16 bits = b << 8 (no bias correction needed).
+inline float fp8e5m2_to_float_scalar(uint8_t b, float scale) noexcept {
+  const uint8_t exp_bits = (b >> 2) & 0x1F;
+  const uint8_t mant_bits = b & 0x03;
+  // NaN: exp=11111, mant!=00
+  if (exp_bits == 0x1F && mant_bits != 0)
+    return std::numeric_limits<float>::quiet_NaN();
+  const uint32_t sign = static_cast<uint32_t>(b & 0x80) << 24;
+  if (exp_bits == 0x1F)
+    return sign ? -std::numeric_limits<float>::infinity()
+                : std::numeric_limits<float>::infinity();
+  if (exp_bits == 0) {  // subnormal: (-1)^s * 2^-14 * mant/4
+    if (mant_bits == 0) return 0.0f;
+    float v = mant_bits * 0x1p-16f;
+    return (sign ? -v : v) * scale;
+  }
+  // Normal: FP32 exp = exp5 - 15 + 127, mantissa top 2 bits
+  uint32_t fp32_bits = sign |
+                       ((static_cast<uint32_t>(exp_bits) - 15 + 127) << 23) |
+                       (static_cast<uint32_t>(mant_bits) << 21);
+  float val = *reinterpret_cast<const f32_alias_t*>(&fp32_bits);
+  return val * scale;
+}
+
+inline uint8_t float_to_fp8e5m2_scalar(float v, float inv_scale) noexcept {
+  v *= inv_scale;
+  constexpr float fp8_e5m2_max = 57344.0f;
+  v = std::max(-fp8_e5m2_max, std::min(fp8_e5m2_max, v));
+  if (v == 0.0f) return 0;
+  uint32_t bits = *reinterpret_cast<const u32_alias_t*>(&v);
+  const uint8_t sign = static_cast<uint8_t>((bits >> 24) & 0x80);
+  const int32_t exp_fp32 = static_cast<int32_t>((bits >> 23) & 0xFF) - 127;
+  const uint8_t mant2 = static_cast<uint8_t>((bits >> 21) & 0x03);
+  if (exp_fp32 < -14) {  // subnormal in E5M2
+    const int shift = -14 - exp_fp32;
+    if (shift + 21 >= 32)
+      return sign;  // underflow: too small for E5M2 subnormal
+    const uint32_t m = (0x800000u | (bits & 0x7FFFFFu)) >> (shift + 21);
+    return sign | static_cast<uint8_t>(std::min<uint32_t>(m, 3u));
+  }
+  const uint8_t exp5 = static_cast<uint8_t>(exp_fp32 + 15);
+  return sign | (exp5 << 2) | mant2;
+}
+
+// ---------------------------------------------------------------------------
+// Select the FP8 quant function at compile time based on kv_cache_t.
+// ---------------------------------------------------------------------------
+template <typename kv_cache_t>
+constexpr auto select_fp8_quant_fn() {
+  if constexpr (std::is_same_v<kv_cache_t, c10::Float8_e5m2>)
+    return float_to_fp8e5m2_scalar;
+  else
+    return float_to_fp8e4m3_scalar;
+}
+
+// ---------------------------------------------------------------------------
+// VEC reshape impl — parameterised on the quantisation function.
+// Writes key (column-major) and value (row-major) into uint8 FP8 KV cache.
+// The pragma omp must live outside VLLM_DISPATCH_FLOATING_TYPES because
+// #pragma cannot appear inside variadic macro arguments.
+// ---------------------------------------------------------------------------
+template <typename scalar_t, uint8_t (*quant_fn)(float, float)>
+inline void reshape_and_cache_fp8_vec_impl(
+    const scalar_t* key_ptr, const scalar_t* value_ptr, uint8_t* key_cache_ptr,
+    uint8_t* value_cache_ptr, const int64_t* slot_ptr, int64_t token_num,
+    int64_t head_num, int64_t head_dim, int64_t block_size, int64_t k_stride0,
+    int64_t k_stride1, int64_t v_stride0, int64_t v_stride1, int64_t kc_stride0,
+    int64_t kc_stride1, int64_t vc_stride0, int64_t vc_stride1, float k_inv,
+    float v_inv) {
+#pragma omp parallel for collapse(2) schedule(static)
+  for (int64_t tok = 0; tok < token_num; ++tok) {
+    for (int64_t h = 0; h < head_num; ++h) {
+      const int64_t slot = slot_ptr[tok];
+      if (slot < 0) continue;
+      const int64_t block_idx = slot / block_size;
+      const int64_t block_offset = slot % block_size;
+
+      // Key layout: column-major within block
+      const scalar_t* ksrc = key_ptr + tok * k_stride0 + h * k_stride1;
+      uint8_t* kdst = key_cache_ptr + block_idx * kc_stride0 + h * kc_stride1 +
+                      block_offset;
+      for (int64_t i = 0; i < head_dim; ++i)
+        kdst[i * block_size] = quant_fn(static_cast<float>(ksrc[i]), k_inv);
+
+      // Value layout: row-major within block (contiguous head_dim bytes)
+      const scalar_t* vsrc = value_ptr + tok * v_stride0 + h * v_stride1;
+      uint8_t* vdst = value_cache_ptr + block_idx * vc_stride0 +
+                      h * vc_stride1 + block_offset * head_dim;
+      for (int64_t i = 0; i < head_dim; ++i)
+        vdst[i] = quant_fn(static_cast<float>(vsrc[i]), v_inv);
+    }
+  }
+}

csrc/cpu/cpu_attn_impl.hpp

@@ -14,8 +14,22 @@
 namespace cpu_attention {
 enum class ISA { AMX, VEC, VEC16, NEON, VXE };
 
-template <ISA isa, typename scalar_t, int64_t head_dim>
-class AttentionImpl {};
+// Mirrors csrc/attention/dtype_fp8.cuh Fp8KVCacheDataType exactly.
+enum class Fp8KVCacheDataType {
+  kAuto = 0,
+  kFp8E4M3 = 1,
+  kFp8E5M2 = 2,
+};
+
+struct AttentionInput;
+
+template <ISA isa, typename scalar_t, int64_t head_dim,
+          typename kv_cache_scalar_t = scalar_t>
+class AttentionImpl {
+ public:
+  void init_from_input(const AttentionInput*) {}
+  float get_output_v_scale() const noexcept { return 1.0f; }
+};
 
 struct AttentionWorkItemGroup {
   int32_t req_id;
@@ -780,6 +794,9 @@ struct AttentionInput {
   int32_t sliding_window_left;
   int32_t sliding_window_right;
   float softcap;
+  // FP8 KV cache scales (used by FP8 attention implementations)
+  float k_scale_fp8 = 1.0f;
+  float v_scale_fp8 = 1.0f;
 };
 
 #define DEFINE_CPU_ATTENTION_PARAMS                                         \
@@ -1152,7 +1169,11 @@ class AttentionMainLoop {
                        bool use_sink) {
 #ifdef DEFINE_FAST_EXP
       DEFINE_FAST_EXP
+      bool constexpr IsReducedPrecision =
+          std::is_same_v<query_t, c10::BFloat16> ||
+          std::is_same_v<query_t, c10::Half>;
 #endif
+
       using prob_buffer_vec_t = typename VecTypeTrait<prob_buffer_t>::vec_t;
       static_assert(sizeof(prob_buffer_t) <= sizeof(logits_buffer_t));
 
@@ -1201,8 +1222,17 @@ class AttentionMainLoop {
             vec = vec - max_vec;
 
             // compute exp
-#ifdef DEFINE_FAST_EXP
-            vec = fast_exp(vec);
+
+#if defined(DEFINE_FAST_EXP)
+  #ifdef __aarch64__
+            if constexpr (IsReducedPrecision) {
+              vec = fast_exp_f16(vec);
+            } else
+  #endif
+            {
+              vec = fast_exp(vec);
+            }
+
             prob_buffer_vec_t output_vec(vec);
             output_vec.save(curr_prob_buffer_iter);
 #else
@@ -1258,7 +1288,11 @@ class AttentionMainLoop {
                        int32_t kv_tile_token_num, float softcap_scale) {
 #ifdef DEFINE_FAST_EXP
       DEFINE_FAST_EXP
+      bool constexpr IsReducedPrecision =
+          std::is_same_v<query_t, c10::BFloat16> ||
+          std::is_same_v<query_t, c10::Half>;
 #endif
+
       float inv_softcap_scale = 1.0 / softcap_scale;
       vec_op::FP32Vec16 softcap_scale_vec(softcap_scale);
       vec_op::FP32Vec16 inv_softcap_scale_vec(inv_softcap_scale);
@@ -1272,8 +1306,15 @@ class AttentionMainLoop {
           vec_op::FP32Vec16 vec(curr_logits_buffer_iter);
           vec = vec * inv_softcap_scale_vec;
 
-#ifdef DEFINE_FAST_EXP
-          vec = fast_exp(vec);
+#if defined(DEFINE_FAST_EXP)
+  #ifdef __aarch64__
+          if constexpr (IsReducedPrecision) {
+            vec = fast_exp_f16(vec);
+          } else
+  #endif
+          {
+            vec = fast_exp(vec);
+          }
           vec_op::FP32Vec16 inv_vec = ones_vec / vec;
           vec = (vec - inv_vec) / (vec + inv_vec);
 #else
@@ -1350,6 +1391,13 @@ class AttentionMainLoop {
       }
 
       attention_impl_t attn_impl;
+      constexpr bool fp8_kv = std::is_same_v<kv_cache_t, c10::Float8_e4m3fn> ||
+                              std::is_same_v<kv_cache_t, c10::Float8_e5m2>;
+      float output_v_scale = 1.0f;
+      if constexpr (fp8_kv) {
+        attn_impl.init_from_input(input);
+        output_v_scale = attn_impl.get_output_v_scale();
+      }
 
       // general information
       const int32_t q_head_num = input->num_heads;
@@ -1729,7 +1777,7 @@ class AttentionMainLoop {
                                reinterpret_cast<query_t*>(input->output) +
                                    output_buffer_offset,
                                sum_buffer, actual_q_heads_per_kv,
-                               actual_q_token_num, q_head_num);
+                               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;
@@ -1799,7 +1847,7 @@ class AttentionMainLoop {
               split_output_buffer,
               reinterpret_cast<query_t*>(input->output) + output_buffer_offset,
               split_sum_buffer, actual_q_heads_per_kv, curr_output_token_num,
-              q_head_num);
+              q_head_num, output_v_scale);
         }
       }
     }
@@ -1923,8 +1971,8 @@ class AttentionMainLoop {
                     query_t* __restrict__ curr_output_buffer,
                     float* __restrict__ sum_buffer,
                     const int32_t q_heads_per_kv,
-                    const int32_t actual_q_token_num,
-                    const int32_t q_head_num) {
+                    const int32_t actual_q_token_num, const int32_t q_head_num,
+                    const float v_scale = 1.0f) {
     // final output
     using output_vec_t = typename VecTypeTrait<query_t>::vec_t;
 
@@ -1938,7 +1986,7 @@ class AttentionMainLoop {
           curr_partial_output_buffer;
       query_t* __restrict__ curr_output_buffer_iter = curr_output_buffer;
       for (int32_t head_idx = 0; head_idx < q_heads_per_kv; ++head_idx) {
-        vec_op::FP32Vec16 inv_sum_scale_vec(1.0 / *curr_sum_buffer);
+        vec_op::FP32Vec16 inv_sum_scale_vec(v_scale / *curr_sum_buffer);
 
         for (int32_t i = 0; i < group_num_per_head; ++i) {
           vec_op::FP32Vec16 vec(curr_partial_output_buffer_iter);

csrc/cpu/cpu_attn_neon.hpp

@@ -248,8 +248,8 @@ class TileGemmNeonFMLA {
 }  // namespace
 
 // this is similar to "ISA::VEC" at the moment
-template <typename scalar_t, int64_t head_dim>
-class AttentionImpl<ISA::NEON, scalar_t, head_dim> {
+template <typename scalar_t, int64_t head_dim, typename kv_cache_scalar_t>
+class AttentionImpl<ISA::NEON, scalar_t, head_dim, kv_cache_scalar_t> {
  public:
   using query_t = scalar_t;
   using q_buffer_t = float;
@@ -343,7 +343,8 @@ class AttentionImpl<ISA::NEON, scalar_t, head_dim> {
       const int64_t head_num, const int64_t key_head_num_stride,
       const int64_t value_head_num_stride, const int64_t num_blocks,
       const int64_t num_blocks_stride, const int64_t cache_head_num_stride,
-      const int64_t block_size, const int64_t block_size_stride) {
+      const int64_t block_size, const int64_t block_size_stride,
+      const float /*k_inv*/ = 0.0f, const float /*v_inv*/ = 0.0f) {
 #pragma omp parallel for collapse(2)
     for (int64_t token_idx = 0; token_idx < token_num; ++token_idx) {
       for (int64_t head_idx = 0; head_idx < head_num; ++head_idx) {
@@ -388,7 +389,7 @@ class AttentionImpl<ISA::NEON, scalar_t, head_dim> {
 #ifdef ARM_BF16_SUPPORT
 // For BF16 on Arm, reuse the BFMMLA kernels with 32-token alignment.
 template <int64_t head_dim>
-class AttentionImpl<ISA::NEON, c10::BFloat16, head_dim>
+class AttentionImpl<ISA::NEON, c10::BFloat16, head_dim, c10::BFloat16>
     : public AttentionImplNEONBFMMLA<BLOCK_SIZE_ALIGNMENT, ISA::NEON,
                                      head_dim> {};
 #endif

csrc/cpu/cpu_attn_neon_bfmmla.hpp

@@ -602,7 +602,8 @@ class AttentionImplNEONBFMMLA {
       [[maybe_unused]] const int64_t num_blocks,
       const int64_t num_blocks_stride, const int64_t cache_head_num_stride,
       const int64_t block_size,
-      [[maybe_unused]] const int64_t block_size_stride) {
+      [[maybe_unused]] const int64_t block_size_stride,
+      const float /*k_inv*/ = 0.0f, const float /*v_inv*/ = 0.0f) {
     const int64_t k_block_stride = (head_dim / TILE_K) * K_INNER_STRIDE;
     const int64_t v_pair_stride =
         (block_size / V_TOKENS_PER_ROW_BLOCK) * V_INNER_STRIDE;

csrc/cpu/cpu_attn_vec.hpp

@@ -1,11 +1,37 @@
 #ifndef CPU_ATTN_VEC_HPP
 #define CPU_ATTN_VEC_HPP
 
+#include "cpu_attn_fp8.hpp"
 #include "cpu_attn_impl.hpp"
 
 namespace cpu_attention {
 
 namespace {
+
+// Load 32 kv_cache_t elements starting at ptr and return them as two FP32Vec16s
+// covering the lower 16 and upper 16 positions.
+// For FP8: both halves come from a single BF16Vec32 dequant of 32 bytes.
+// For BF16/FP16/FP32: two separate vector loads at ptr and ptr+16.
+template <typename kv_cache_t>
+FORCE_INLINE std::pair<vec_op::FP32Vec16, vec_op::FP32Vec16> load_b_pair_vec(
+    const kv_cache_t* ptr) {
+  if constexpr (std::is_same_v<kv_cache_t, c10::Float8_e4m3fn>) {
+    // BF16 container, but values are in the FP16 exponent range (bias 15 not
+    // 127).
+    vec_op::BF16Vec32 bf16_b_reg(reinterpret_cast<const uint8_t*>(ptr),
+                                 vec_op::fp8_e4m3_tag{});
+    return {vec_op::FP32Vec16(bf16_b_reg, 0), vec_op::FP32Vec16(bf16_b_reg, 1)};
+  } else if constexpr (std::is_same_v<kv_cache_t, c10::Float8_e5m2>) {
+    vec_op::BF16Vec32 bf16_b_reg(reinterpret_cast<const uint8_t*>(ptr),
+                                 vec_op::fp8_e5m2_tag{});
+    return {vec_op::FP32Vec16(bf16_b_reg, 0), vec_op::FP32Vec16(bf16_b_reg, 1)};
+  } else {
+    using load_vec_t = typename VecTypeTrait<kv_cache_t>::vec_t;
+    return {vec_op::FP32Vec16(load_vec_t(ptr)),
+            vec_op::FP32Vec16(load_vec_t(ptr + 16))};
+  }
+}
+
 // 8-2-16 pattern, 8 regs for A, 2 regs for B, 16 regs for C, [8, K] @ [k, 32]
 template <typename kv_cache_t>
 class TileGemm82 {
@@ -54,10 +80,7 @@ class TileGemm82 {
                          const int32_t block_size, const int32_t dynamic_k_size,
                          const bool accum_c) {
     static_assert(0 < M && M <= 8);
-    using load_vec_t = typename VecTypeTrait<kv_cache_t>::vec_t;
 
-    kv_cache_t* __restrict__ curr_b_0 = b_tile;
-    kv_cache_t* __restrict__ curr_b_1 = b_tile + 16;
     float* __restrict__ curr_c_0 = c_tile;
     float* __restrict__ curr_c_1 = c_tile + 16;
 
@@ -76,16 +99,14 @@ class TileGemm82 {
     }
 
     float* __restrict__ curr_a = a_tile;
+    kv_cache_t* __restrict__ curr_b = b_tile;
+
     for (int32_t k = 0; k < dynamic_k_size; ++k) {
-      load_vec_t b_0_reg(curr_b_0);
-      vec_op::FP32Vec16 fp32_b_0_reg(b_0_reg);
-      load_vec_t b_1_reg(curr_b_1);
-      vec_op::FP32Vec16 fp32_b_1_reg(b_1_reg);
+      auto [fp32_b_0_reg, fp32_b_1_reg] = load_b_pair_vec(curr_b);
 
       float* __restrict__ curr_m_a = curr_a;
       vec_op::unroll_loop<int32_t, M>([&](int32_t i) {
-        float v = *curr_m_a;
-        vec_op::FP32Vec16 a_reg(v);
+        vec_op::FP32Vec16 a_reg(*curr_m_a);
         c_regs[i * 2] = c_regs[i * 2] + a_reg * fp32_b_0_reg;
         c_regs[i * 2 + 1] = c_regs[i * 2 + 1] + a_reg * fp32_b_1_reg;
 
@@ -95,8 +116,7 @@ class TileGemm82 {
 
       // update
       curr_a += 1;
-      curr_b_0 += ldb;
-      curr_b_1 += ldb;
+      curr_b += ldb;
     }
 
     vec_op::unroll_loop<int32_t, M>([&](int32_t i) {
@@ -109,15 +129,20 @@ class TileGemm82 {
     });
   }
 };
+
 }  // namespace
 
 // This is a general but naive implementation based on vector instructions
-template <typename scalar_t, int64_t head_dim>
-class AttentionImpl<ISA::VEC, scalar_t, head_dim> {
+template <typename scalar_t, int64_t head_dim, typename kv_cache_scalar_t>
+class AttentionImpl<ISA::VEC, scalar_t, head_dim, kv_cache_scalar_t> {
+  static constexpr bool fp8_kv =
+      std::is_same_v<kv_cache_scalar_t, c10::Float8_e4m3fn> ||
+      std::is_same_v<kv_cache_scalar_t, c10::Float8_e5m2>;
+
  public:
   using query_t = scalar_t;
   using q_buffer_t = float;
-  using kv_cache_t = scalar_t;
+  using kv_cache_t = kv_cache_scalar_t;
   using logits_buffer_t = float;
   using partial_output_buffer_t = float;
   using prob_buffer_t = float;
@@ -129,11 +154,45 @@ class AttentionImpl<ISA::VEC, scalar_t, head_dim> {
   constexpr static int64_t MaxQHeadNumPerIteration = 8;
   constexpr static int64_t HeadDim = head_dim;
   constexpr static ISA ISAType = ISA::VEC;
-  constexpr static bool scale_on_logits = false;  // apply scale on q_buffer
+  constexpr static bool scale_on_logits = fp8_kv;
+
+  float k_scale = 1.0f;
+  float v_scale = 1.0f;
 
  public:
+  void init_from_input(const AttentionInput* input) {
+    if constexpr (fp8_kv) {
+      k_scale = input->k_scale_fp8;
+      v_scale = input->v_scale_fp8;
+    }
+  }
+
+  float get_output_v_scale() const noexcept {
+    if constexpr (fp8_kv) {
+      // VEC dequant unpacks FP8 into a pseudo-FP16 layout (exponent bias 15).
+      // E4M3 (bias=7) needs correction 2^(15-7) = 2^8; E5M2 bias matches FP16
+      // so no correction.
+      if constexpr (std::is_same_v<kv_cache_t, c10::Float8_e5m2>) {
+        return v_scale;
+      } else {
+        return v_scale * 0x1p8f;
+      }
+    }
+    return 1.0f;
+  }
+
   template <template <typename tile_gemm_t> typename attention>
   FORCE_INLINE void execute_attention(DEFINE_CPU_ATTENTION_PARAMS) {
+    if constexpr (fp8_kv) {
+      // Same bias correction as get_output_v_scale: VEC FP8→pseudo-FP16 dequant
+      // uses bias 15; E4M3 (bias=7) needs ×2^8, E5M2 (bias=15) needs no
+      // correction.
+      if constexpr (std::is_same_v<kv_cache_t, c10::Float8_e5m2>) {
+        scale *= k_scale;
+      } else {
+        scale *= k_scale * 0x1p8f;
+      }
+    }
     attention<TileGemm82<kv_cache_t>> attention_iteration;
     attention_iteration(CPU_ATTENTION_PARAMS);
   }
@@ -161,17 +220,19 @@ class AttentionImpl<ISA::VEC, scalar_t, head_dim> {
                               // row-major
   }
 
-  // Copy q to q_buffer and cast it to fp32
-  static void copy_q_heads_tile(
-      scalar_t* __restrict__ src,  // [q_num, q_heads_per_kv, head_size]
-      float* __restrict__ q_buffer, const int32_t q_num,
-      const int32_t q_heads_per_kv, const int64_t q_num_stride,
-      const int64_t q_head_stride, float scale) {
+  // Copy q to q_buffer and cast it to fp32.
+  // FP8: QK scale is folded into execute_attention; copy Q unscaled here.
+  void copy_q_heads_tile(scalar_t* __restrict__ src,
+                         float* __restrict__ q_buffer, const int32_t q_num,
+                         const int32_t q_heads_per_kv,
+                         const int64_t q_num_stride,
+                         const int64_t q_head_stride, float scale) {
     static_assert(head_dim % 16 == 0);
     constexpr int32_t unroll_size = head_dim / 16;
     using load_vec_t = typename VecTypeTrait<scalar_t>::vec_t;
 
-    vec_op::FP32Vec16 scale_vec(scale);
+    const float effective_scale = fp8_kv ? 1.0f : scale;
+    vec_op::FP32Vec16 scale_vec(effective_scale);
     for (int32_t q_num_idx = 0; q_num_idx < q_num; ++q_num_idx) {
       for (int32_t q_head_idx = 0; q_head_idx < q_heads_per_kv; ++q_head_idx) {
         scalar_t* __restrict__ curr_q =
@@ -196,13 +257,26 @@ class AttentionImpl<ISA::VEC, scalar_t, head_dim> {
   // reshape K as column-major and V as row-major
   static void reshape_and_cache(
       const scalar_t* __restrict__ key, const scalar_t* __restrict__ value,
-      scalar_t* __restrict__ key_cache, scalar_t* __restrict__ value_cache,
+      kv_cache_t* __restrict__ key_cache, kv_cache_t* __restrict__ value_cache,
       const int64_t* __restrict__ slot_mapping, const int64_t token_num,
       const int64_t key_token_num_stride, const int64_t value_token_num_stride,
       const int64_t head_num, const int64_t key_head_num_stride,
       const int64_t value_head_num_stride, const int64_t num_blocks,
       const int64_t num_blocks_stride, const int64_t cache_head_num_stride,
-      const int64_t block_size, const int64_t block_size_stride) {
+      const int64_t block_size, const int64_t block_size_stride,
+      const float k_inv = 0.0f, const float v_inv = 0.0f) {
+    if constexpr (fp8_kv) {
+      constexpr auto qfn = select_fp8_quant_fn<kv_cache_t>();
+      reshape_and_cache_fp8_vec_impl<scalar_t, qfn>(
+          key, value, reinterpret_cast<uint8_t*>(key_cache),
+          reinterpret_cast<uint8_t*>(value_cache), slot_mapping, token_num,
+          head_num, head_dim, block_size, key_token_num_stride,
+          key_head_num_stride, value_token_num_stride, value_head_num_stride,
+          num_blocks_stride, cache_head_num_stride, num_blocks_stride,
+          cache_head_num_stride, k_inv, v_inv);
+      return;
+    }
+
 #pragma omp parallel for collapse(2)
     for (int64_t token_idx = 0; token_idx < token_num; ++token_idx) {
       for (int64_t head_idx = 0; head_idx < head_num; ++head_idx) {
@@ -220,8 +294,9 @@ class AttentionImpl<ISA::VEC, scalar_t, head_dim> {
                                           token_idx * key_token_num_stride +
                                           head_idx * key_head_num_stride;
           scalar_t* key_cache_start_ptr =
-              key_cache + block_idx * num_blocks_stride +
-              head_idx * cache_head_num_stride + block_offset;
+              reinterpret_cast<scalar_t*>(key_cache) +
+              block_idx * num_blocks_stride + head_idx * cache_head_num_stride +
+              block_offset;
 
 #pragma GCC unroll 8
           for (int64_t i = 0, j = 0; i < head_dim; ++i, j += block_size) {
@@ -234,8 +309,9 @@ class AttentionImpl<ISA::VEC, scalar_t, head_dim> {
                                             token_idx * value_token_num_stride +
                                             head_idx * value_head_num_stride;
           scalar_t* value_cache_start_ptr =
-              value_cache + block_idx * num_blocks_stride +
-              head_idx * cache_head_num_stride + block_offset * head_dim;
+              reinterpret_cast<scalar_t*>(value_cache) +
+              block_idx * num_blocks_stride + head_idx * cache_head_num_stride +
+              block_offset * head_dim;
           std::memcpy(value_cache_start_ptr, value_start_ptr,
                       sizeof(scalar_t) * head_dim);
         }
@@ -243,6 +319,7 @@ class AttentionImpl<ISA::VEC, scalar_t, head_dim> {
     }
   }
 };
+
 }  // namespace cpu_attention
 
 #endif

csrc/cpu/cpu_attn_vec16.hpp

@@ -116,9 +116,9 @@ class TileGemm161 {
 }  // namespace
 
 // This is a general but naive implementation based on vector instructions
-template <typename scalar_t, int64_t head_dim>
-class AttentionImpl<ISA::VEC16, scalar_t, head_dim>
-    : public AttentionImpl<ISA::VEC, scalar_t, head_dim> {
+template <typename scalar_t, int64_t head_dim, typename kv_cache_scalar_t>
+class AttentionImpl<ISA::VEC16, scalar_t, head_dim, kv_cache_scalar_t>
+    : public AttentionImpl<ISA::VEC, scalar_t, head_dim, kv_cache_scalar_t> {
  public:
   using query_t = scalar_t;
   using q_buffer_t = float;

csrc/cpu/cpu_attn_vxe.hpp

@@ -244,8 +244,8 @@ class TileGemmS390X {
 
 }  // namespace
 
-template <typename scalar_t, int64_t head_dim>
-class AttentionImpl<ISA::VXE, scalar_t, head_dim> {
+template <typename scalar_t, int64_t head_dim, typename kv_cache_scalar_t>
+class AttentionImpl<ISA::VXE, scalar_t, head_dim, kv_cache_scalar_t> {
  public:
   using query_t = scalar_t;
   using q_buffer_t = float;
@@ -342,7 +342,8 @@ class AttentionImpl<ISA::VXE, scalar_t, head_dim> {
       const int64_t head_num, const int64_t key_head_num_stride,
       const int64_t value_head_num_stride, const int64_t num_blocks,
       const int64_t num_blocks_stride, const int64_t cache_head_num_stride,
-      const int64_t block_size, const int64_t block_size_stride) {
+      const int64_t block_size, const int64_t block_size_stride,
+      const float /*k_inv*/ = 0.0f, const float /*v_inv*/ = 0.0f) {
 #pragma omp parallel for collapse(2)
     for (int64_t token_idx = 0; token_idx < token_num; ++token_idx) {
       for (int64_t head_idx = 0; head_idx < head_num; ++head_idx) {

csrc/cpu/cpu_types_arm.hpp

@@ -15,6 +15,9 @@ using namespace at::vec;
 
 namespace vec_op {
 
+struct fp8_e4m3_tag {};
+struct fp8_e5m2_tag {};
+
 #define VLLM_DISPATCH_CASE_FLOATING_TYPES(...)         \
   AT_DISPATCH_CASE(at::ScalarType::Float, __VA_ARGS__) \
   AT_DISPATCH_CASE(at::ScalarType::Half, __VA_ARGS__)  \
@@ -322,6 +325,9 @@ struct BF16Vec32 : public VectorizedRegWrapper<BF16Vec32, 4, c10::BFloat16> {
     reg.val[2] = vec8_data.reg.val[0];
     reg.val[3] = vec8_data.reg.val[0];
   };
+
+  explicit BF16Vec32(const uint8_t*, fp8_e4m3_tag) : Base() {}
+  explicit BF16Vec32(const uint8_t*, fp8_e5m2_tag) : Base() {}
 };
 
 struct FP32Vec4 : public VectorizedRegWrapper<FP32Vec4, 1, float> {

csrc/cpu/cpu_types_riscv.hpp

@@ -1,832 +1,25 @@
 #ifndef CPU_TYPES_RISCV_HPP
 #define CPU_TYPES_RISCV_HPP
 
-#include <algorithm>
-#include <cmath>
-#include <cstring>
-#include <iostream>
-#include <limits>
-#include <riscv_vector.h>
-#include <torch/all.h>
+// RISC-V Vector (RVV) CPU type definitions for vLLM.
+//
+// Supports multiple VLENs via compile-time dispatch. The compiler defines
+// __riscv_v_min_vlen from the zvl<N>b extension in -march. The defs header
+// maps VLEN to the correct LMUL suffixes, and the impl header provides
+// VLEN-independent class implementations.
+//
+// To add support for a new VLEN, add the LMUL mapping in
+// cpu_types_riscv_defs.hpp (the impl header needs no changes).
 
-// ============================================================================
-// Vector Register Type Definitions (VLEN=128 bits)
-// ============================================================================
-
-typedef vfloat16m1_t fixed_vfloat16m1_t
-    __attribute__((riscv_rvv_vector_bits(128)));
-typedef vfloat16m2_t fixed_vfloat16m2_t
-    __attribute__((riscv_rvv_vector_bits(256)));
-
-typedef vfloat32m1_t fixed_vfloat32m1_t
-    __attribute__((riscv_rvv_vector_bits(128)));
-typedef vfloat32m2_t fixed_vfloat32m2_t
-    __attribute__((riscv_rvv_vector_bits(256)));
-typedef vfloat32m4_t fixed_vfloat32m4_t
-    __attribute__((riscv_rvv_vector_bits(512)));
-typedef vfloat32m8_t fixed_vfloat32m8_t
-    __attribute__((riscv_rvv_vector_bits(1024)));
-
-typedef vint32m2_t fixed_vint32m2_t __attribute__((riscv_rvv_vector_bits(256)));
-typedef vint32m4_t fixed_vint32m4_t __attribute__((riscv_rvv_vector_bits(512)));
-
-typedef vuint16m1_t fixed_vuint16m1_t
-    __attribute__((riscv_rvv_vector_bits(128)));
-typedef vuint16m2_t fixed_vuint16m2_t
-    __attribute__((riscv_rvv_vector_bits(256)));
-typedef vuint16m4_t fixed_vuint16m4_t
-    __attribute__((riscv_rvv_vector_bits(512)));
-
-#ifdef RISCV_BF16_SUPPORT
-typedef vbfloat16m1_t fixed_vbfloat16m1_t
-    __attribute__((riscv_rvv_vector_bits(128)));
-typedef vbfloat16m2_t fixed_vbfloat16m2_t
-    __attribute__((riscv_rvv_vector_bits(256)));
-typedef vbfloat16m4_t fixed_vbfloat16m4_t
-    __attribute__((riscv_rvv_vector_bits(512)));
+#ifndef __riscv_vector
+  #error "cpu_types_riscv.hpp included in a non-RVV translation unit"
 #endif
 
-namespace vec_op {
-
-#ifdef RISCV_BF16_SUPPORT
-  #define VLLM_DISPATCH_CASE_FLOATING_TYPES(...)         \
-    AT_DISPATCH_CASE(at::ScalarType::Float, __VA_ARGS__) \
-    AT_DISPATCH_CASE(at::ScalarType::Half, __VA_ARGS__)  \
-    AT_DISPATCH_CASE(at::ScalarType::BFloat16, __VA_ARGS__)
-#else
-  #define VLLM_DISPATCH_CASE_FLOATING_TYPES(...)         \
-    AT_DISPATCH_CASE(at::ScalarType::Float, __VA_ARGS__) \
-    AT_DISPATCH_CASE(at::ScalarType::Half, __VA_ARGS__)
+#ifndef __riscv_v_min_vlen
+  #error "compiler did not define __riscv_v_min_vlen; pass -march=...zvl<N>b"
 #endif
 
-#define VLLM_DISPATCH_FLOATING_TYPES(TYPE, NAME, ...) \
-  AT_DISPATCH_SWITCH(TYPE, NAME, VLLM_DISPATCH_CASE_FLOATING_TYPES(__VA_ARGS__))
+#include "cpu_types_riscv_defs.hpp"
+#include "cpu_types_riscv_impl.hpp"
 
-#define FORCE_INLINE __attribute__((always_inline)) inline
-
-namespace {
-template <typename T, T... indexes, typename F>
-constexpr void unroll_loop_item(std::integer_sequence<T, indexes...>, F&& f) {
-  (f(std::integral_constant<T, indexes>{}), ...);
-};
-}  // namespace
-
-template <typename T, T count, typename F,
-          typename = std::enable_if_t<std::is_invocable_v<F, T>>>
-constexpr void unroll_loop(F&& f) {
-  unroll_loop_item(std::make_integer_sequence<T, count>{}, std::forward<F>(f));
-}
-
-template <typename T>
-struct Vec {
-  constexpr static int get_elem_num() { return T::VEC_ELEM_NUM; };
-};
-
-struct FP32Vec8;
-struct FP32Vec16;
-
-// ============================================================================
-// FP16 Implementation
-// ============================================================================
-
-struct FP16Vec8 : public Vec<FP16Vec8> {
-  constexpr static int VEC_ELEM_NUM = 8;
-  fixed_vfloat16m1_t reg;
-
-  explicit FP16Vec8(const void* ptr)
-      : reg(__riscv_vle16_v_f16m1(static_cast<const _Float16*>(ptr),
-                                  VEC_ELEM_NUM)) {};
-
-  explicit FP16Vec8(const FP32Vec8&);
-
-  void save(void* ptr) const {
-    __riscv_vse16_v_f16m1(static_cast<_Float16*>(ptr), reg, VEC_ELEM_NUM);
-  }
-  void save(void* ptr, int elem_num) const {
-    __riscv_vse16_v_f16m1(static_cast<_Float16*>(ptr), reg, elem_num);
-  }
-  void save_strided(void* ptr, ptrdiff_t stride) const {
-    ptrdiff_t byte_stride = stride * sizeof(_Float16);
-    __riscv_vsse16_v_f16m1(static_cast<_Float16*>(ptr), byte_stride, reg,
-                           VEC_ELEM_NUM);
-  }
-};
-
-struct FP16Vec16 : public Vec<FP16Vec16> {
-  constexpr static int VEC_ELEM_NUM = 16;
-  fixed_vfloat16m2_t reg;
-
-  explicit FP16Vec16(const void* ptr)
-      : reg(__riscv_vle16_v_f16m2(static_cast<const _Float16*>(ptr),
-                                  VEC_ELEM_NUM)) {};
-
-  explicit FP16Vec16(const FP32Vec16& vec);
-
-  void save(void* ptr) const {
-    __riscv_vse16_v_f16m2(static_cast<_Float16*>(ptr), reg, VEC_ELEM_NUM);
-  }
-  void save(void* ptr, int elem_num) const {
-    __riscv_vse16_v_f16m2(static_cast<_Float16*>(ptr), reg, elem_num);
-  }
-  void save_strided(void* ptr, ptrdiff_t stride) const {
-    ptrdiff_t byte_stride = stride * sizeof(_Float16);
-    __riscv_vsse16_v_f16m2(static_cast<_Float16*>(ptr), byte_stride, reg,
-                           VEC_ELEM_NUM);
-  }
-};
-
-// ============================================================================
-// BF16 Implementation
-// ============================================================================
-
-#ifdef RISCV_BF16_SUPPORT
-
-FORCE_INLINE fixed_vuint16m1_t bf16_to_u16(fixed_vbfloat16m1_t v) {
-  return __riscv_vreinterpret_v_bf16m1_u16m1(v);
-}
-FORCE_INLINE fixed_vuint16m2_t bf16_to_u16(fixed_vbfloat16m2_t v) {
-  return __riscv_vreinterpret_v_bf16m2_u16m2(v);
-}
-FORCE_INLINE fixed_vuint16m4_t bf16_to_u16(fixed_vbfloat16m4_t v) {
-  return __riscv_vreinterpret_v_bf16m4_u16m4(v);
-}
-
-struct BF16Vec8 : public Vec<BF16Vec8> {
-  constexpr static int VEC_ELEM_NUM = 8;
-  fixed_vbfloat16m1_t reg;
-
-  explicit BF16Vec8(const void* ptr)
-      : reg(__riscv_vreinterpret_v_u16m1_bf16m1(__riscv_vle16_v_u16m1(
-            reinterpret_cast<const uint16_t*>(ptr), VEC_ELEM_NUM))) {};
-
-  explicit BF16Vec8(fixed_vbfloat16m1_t data) : reg(data) {};
-  explicit BF16Vec8(const FP32Vec8&);
-
-  void save(void* ptr) const {
-    __riscv_vse16_v_u16m1(reinterpret_cast<uint16_t*>(ptr), bf16_to_u16(reg),
-                          VEC_ELEM_NUM);
-  }
-  void save(void* ptr, int elem_num) const {
-    __riscv_vse16_v_u16m1(reinterpret_cast<uint16_t*>(ptr), bf16_to_u16(reg),
-                          elem_num);
-  }
-  void save_strided(void* ptr, ptrdiff_t stride) const {
-    ptrdiff_t byte_stride = stride * sizeof(uint16_t);
-    __riscv_vsse16_v_u16m1(reinterpret_cast<uint16_t*>(ptr), byte_stride,
-                           bf16_to_u16(reg), VEC_ELEM_NUM);
-  }
-};
-
-struct BF16Vec16 : public Vec<BF16Vec16> {
-  constexpr static int VEC_ELEM_NUM = 16;
-  fixed_vbfloat16m2_t reg;
-
-  explicit BF16Vec16(const void* ptr)
-      : reg(__riscv_vreinterpret_v_u16m2_bf16m2(__riscv_vle16_v_u16m2(
-            reinterpret_cast<const uint16_t*>(ptr), VEC_ELEM_NUM))) {};
-
-  explicit BF16Vec16(fixed_vbfloat16m2_t data) : reg(data) {};
-  explicit BF16Vec16(const FP32Vec16&);
-
-  void save(void* ptr) const {
-    __riscv_vse16_v_u16m2(reinterpret_cast<uint16_t*>(ptr), bf16_to_u16(reg),
-                          VEC_ELEM_NUM);
-  }
-  void save(void* ptr, int elem_num) const {
-    __riscv_vse16_v_u16m2(reinterpret_cast<uint16_t*>(ptr), bf16_to_u16(reg),
-                          elem_num);
-  }
-  void save_strided(void* ptr, ptrdiff_t stride) const {
-    ptrdiff_t byte_stride = stride * sizeof(uint16_t);
-    __riscv_vsse16_v_u16m2(reinterpret_cast<uint16_t*>(ptr), byte_stride,
-                           bf16_to_u16(reg), VEC_ELEM_NUM);
-  }
-};
-
-struct BF16Vec32 : public Vec<BF16Vec32> {
-  constexpr static int VEC_ELEM_NUM = 32;
-  fixed_vbfloat16m4_t reg;
-
-  explicit BF16Vec32(const void* ptr)
-      : reg(__riscv_vreinterpret_v_u16m4_bf16m4(__riscv_vle16_v_u16m4(
-            reinterpret_cast<const uint16_t*>(ptr), VEC_ELEM_NUM))) {};
-
-  explicit BF16Vec32(fixed_vbfloat16m4_t data) : reg(data) {};
-
-  explicit BF16Vec32(const BF16Vec8& v) {
-    fixed_vuint16m1_t u16_val = bf16_to_u16(v.reg);
-    fixed_vuint16m4_t u16_combined =
-        __riscv_vcreate_v_u16m1_u16m4(u16_val, u16_val, u16_val, u16_val);
-    reg = __riscv_vreinterpret_v_u16m4_bf16m4(u16_combined);
-  };
-
-  void save(void* ptr) const {
-    __riscv_vse16_v_u16m4(reinterpret_cast<uint16_t*>(ptr), bf16_to_u16(reg),
-                          VEC_ELEM_NUM);
-  }
-  void save(void* ptr, int elem_num) const {
-    __riscv_vse16_v_u16m4(reinterpret_cast<uint16_t*>(ptr), bf16_to_u16(reg),
-                          elem_num);
-  }
-  void save_strided(void* ptr, ptrdiff_t stride) const {
-    ptrdiff_t byte_stride = stride * sizeof(uint16_t);
-    __riscv_vsse16_v_u16m4(reinterpret_cast<uint16_t*>(ptr), byte_stride,
-                           bf16_to_u16(reg), VEC_ELEM_NUM);
-  }
-};
-
-#else
-// ============================================================================
-// BF16 Fallback Implementation (FP32 Simulation)
-// ============================================================================
-
-struct BF16Vec8 : public Vec<BF16Vec8> {
-  constexpr static int VEC_ELEM_NUM = 8;
-  fixed_vfloat32m2_t reg_fp32;
-  explicit BF16Vec8(const void* ptr) {
-    const uint16_t* u16 = static_cast<const uint16_t*>(ptr);
-    float tmp[8];
-    for (int i = 0; i < 8; ++i) {
-      uint32_t v = static_cast<uint32_t>(u16[i]) << 16;
-      std::memcpy(&tmp[i], &v, 4);
-    }
-    reg_fp32 = __riscv_vle32_v_f32m2(tmp, 8);
-  }
-  explicit BF16Vec8(const FP32Vec8&);
-  void save(void* ptr) const {
-    float tmp[8];
-    __riscv_vse32_v_f32m2(tmp, reg_fp32, 8);
-    uint16_t* u16 = static_cast<uint16_t*>(ptr);
-    for (int i = 0; i < 8; ++i) {
-      uint32_t v;
-      std::memcpy(&v, &tmp[i], 4);
-      u16[i] = static_cast<uint16_t>(v >> 16);
-    }
-  }
-  void save(void* ptr, int elem_num) const {
-    float tmp[8];
-    __riscv_vse32_v_f32m2(tmp, reg_fp32, 8);
-    uint16_t* u16 = static_cast<uint16_t*>(ptr);
-    for (int i = 0; i < elem_num; ++i) {
-      uint32_t v;
-      std::memcpy(&v, &tmp[i], 4);
-      u16[i] = static_cast<uint16_t>(v >> 16);
-    }
-  }
-  void save_strided(void* ptr, ptrdiff_t stride) const {
-    float tmp[8];
-    __riscv_vse32_v_f32m2(tmp, reg_fp32, 8);
-    uint8_t* u8 = static_cast<uint8_t*>(ptr);
-    ptrdiff_t byte_stride = stride * sizeof(uint16_t);
-    for (int i = 0; i < 8; ++i) {
-      uint32_t v;
-      std::memcpy(&v, &tmp[i], 4);
-      uint16_t val = static_cast<uint16_t>(v >> 16);
-      *reinterpret_cast<uint16_t*>(u8 + i * byte_stride) = val;
-    }
-  }
-};
-
-struct BF16Vec16 : public Vec<BF16Vec16> {
-  constexpr static int VEC_ELEM_NUM = 16;
-  fixed_vfloat32m4_t reg_fp32;
-  explicit BF16Vec16(const void* ptr) {
-    const uint16_t* u16 = static_cast<const uint16_t*>(ptr);
-    float tmp[16];
-    for (int i = 0; i < 16; ++i) {
-      uint32_t v = static_cast<uint32_t>(u16[i]) << 16;
-      std::memcpy(&tmp[i], &v, 4);
-    }
-    reg_fp32 = __riscv_vle32_v_f32m4(tmp, 16);
-  }
-  explicit BF16Vec16(const FP32Vec16&);
-  void save(void* ptr) const {
-    float tmp[16];
-    __riscv_vse32_v_f32m4(tmp, reg_fp32, 16);
-    uint16_t* u16 = static_cast<uint16_t*>(ptr);
-    for (int i = 0; i < 16; ++i) {
-      uint32_t v;
-      std::memcpy(&v, &tmp[i], 4);
-      u16[i] = static_cast<uint16_t>(v >> 16);
-    }
-  }
-  void save(void* ptr, int elem_num) const {
-    float tmp[16];
-    __riscv_vse32_v_f32m4(tmp, reg_fp32, 16);
-    uint16_t* u16 = static_cast<uint16_t*>(ptr);
-    for (int i = 0; i < elem_num; ++i) {
-      uint32_t v;
-      std::memcpy(&v, &tmp[i], 4);
-      u16[i] = static_cast<uint16_t>(v >> 16);
-    }
-  }
-  void save_strided(void* ptr, ptrdiff_t stride) const {
-    float tmp[16];
-    __riscv_vse32_v_f32m4(tmp, reg_fp32, 16);
-    uint8_t* u8 = static_cast<uint8_t*>(ptr);
-    ptrdiff_t byte_stride = stride * sizeof(uint16_t);
-    for (int i = 0; i < 16; ++i) {
-      uint32_t v;
-      std::memcpy(&v, &tmp[i], 4);
-      uint16_t val = static_cast<uint16_t>(v >> 16);
-      *reinterpret_cast<uint16_t*>(u8 + i * byte_stride) = val;
-    }
-  }
-};
-
-struct BF16Vec32 : public Vec<BF16Vec32> {
-  constexpr static int VEC_ELEM_NUM = 32;
-  fixed_vfloat32m8_t reg_fp32;
-
-  explicit BF16Vec32(const void* ptr) {
-    const uint16_t* u16 = static_cast<const uint16_t*>(ptr);
-    float tmp[32];
-    for (int i = 0; i < 32; ++i) {
-      uint32_t v = static_cast<uint32_t>(u16[i]) << 16;
-      std::memcpy(&tmp[i], &v, 4);
-    }
-    reg_fp32 = __riscv_vle32_v_f32m8(tmp, 32);
-  }
-
-  explicit BF16Vec32(const BF16Vec8& v) {
-    float tmp_small[8];
-    __riscv_vse32_v_f32m2(tmp_small, v.reg_fp32, 8);
-    float tmp_large[32];
-    for (int i = 0; i < 4; ++i) {
-      std::memcpy(tmp_large + (i * 8), tmp_small, 8 * sizeof(float));
-    }
-    reg_fp32 = __riscv_vle32_v_f32m8(tmp_large, 32);
-  }
-
-  void save(void* ptr) const {
-    float tmp[32];
-    __riscv_vse32_v_f32m8(tmp, reg_fp32, 32);
-    uint16_t* u16 = static_cast<uint16_t*>(ptr);
-    for (int i = 0; i < 32; ++i) {
-      uint32_t v;
-      std::memcpy(&v, &tmp[i], 4);
-      u16[i] = static_cast<uint16_t>(v >> 16);
-    }
-  }
-
-  void save(void* ptr, int elem_num) const {
-    float tmp[32];
-    __riscv_vse32_v_f32m8(tmp, reg_fp32, 32);
-    uint16_t* u16 = static_cast<uint16_t*>(ptr);
-    for (int i = 0; i < elem_num; ++i) {
-      uint32_t v;
-      std::memcpy(&v, &tmp[i], 4);
-      u16[i] = static_cast<uint16_t>(v >> 16);
-    }
-  }
-
-  void save_strided(void* ptr, ptrdiff_t stride) const {
-    float tmp[32];
-    __riscv_vse32_v_f32m8(tmp, reg_fp32, 32);
-    uint8_t* u8 = static_cast<uint8_t*>(ptr);
-    ptrdiff_t byte_stride = stride * sizeof(uint16_t);
-    for (int i = 0; i < 32; ++i) {
-      uint32_t v;
-      std::memcpy(&v, &tmp[i], 4);
-      uint16_t val = static_cast<uint16_t>(v >> 16);
-      *reinterpret_cast<uint16_t*>(u8 + i * byte_stride) = val;
-    }
-  }
-};
-#endif
-
-// ============================================================================
-// FP32 Implementation
-// ============================================================================
-
-struct FP32Vec4 : public Vec<FP32Vec4> {
-  constexpr static int VEC_ELEM_NUM = 4;
-  fixed_vfloat32m1_t reg;
-  explicit FP32Vec4(float v) : reg(__riscv_vfmv_v_f_f32m1(v, VEC_ELEM_NUM)) {};
-  explicit FP32Vec4() : reg(__riscv_vfmv_v_f_f32m1(0.0f, VEC_ELEM_NUM)) {};
-  explicit FP32Vec4(const float* ptr)
-      : reg(__riscv_vle32_v_f32m1(ptr, VEC_ELEM_NUM)) {};
-  explicit FP32Vec4(fixed_vfloat32m1_t data) : reg(data) {};
-  explicit FP32Vec4(const FP32Vec4& data) : reg(data.reg) {};
-  void save(float* ptr) const { __riscv_vse32_v_f32m1(ptr, reg, VEC_ELEM_NUM); }
-  void save(float* ptr, int elem_num) const {
-    __riscv_vse32_v_f32m1(ptr, reg, elem_num);
-  }
-};
-
-struct FP32Vec8 : public Vec<FP32Vec8> {
-  constexpr static int VEC_ELEM_NUM = 8;
-  fixed_vfloat32m2_t reg;
-
-  explicit FP32Vec8(float v) : reg(__riscv_vfmv_v_f_f32m2(v, VEC_ELEM_NUM)) {};
-  explicit FP32Vec8() : reg(__riscv_vfmv_v_f_f32m2(0.0f, VEC_ELEM_NUM)) {};
-  explicit FP32Vec8(const float* ptr)
-      : reg(__riscv_vle32_v_f32m2(ptr, VEC_ELEM_NUM)) {};
-  explicit FP32Vec8(fixed_vfloat32m2_t data) : reg(data) {};
-  explicit FP32Vec8(const FP32Vec8& data) : reg(data.reg) {};
-  explicit FP32Vec8(const FP16Vec8& v)
-      : reg(__riscv_vfwcvt_f_f_v_f32m2(v.reg, VEC_ELEM_NUM)) {};
-  explicit FP32Vec8(fixed_vfloat16m1_t v)
-      : reg(__riscv_vfwcvt_f_f_v_f32m2(v, VEC_ELEM_NUM)) {};
-
-#ifdef RISCV_BF16_SUPPORT
-  explicit FP32Vec8(fixed_vbfloat16m1_t v)
-      : reg(__riscv_vfwcvtbf16_f_f_v_f32m2(v, VEC_ELEM_NUM)) {};
-  explicit FP32Vec8(const BF16Vec8& v)
-      : reg(__riscv_vfwcvtbf16_f_f_v_f32m2(v.reg, VEC_ELEM_NUM)) {};
-#else
-  explicit FP32Vec8(const BF16Vec8& v) : reg(v.reg_fp32) {};
-#endif
-
-  float reduce_sum() const {
-    fixed_vfloat32m1_t scalar = __riscv_vfmv_s_f_f32m1(0.0f, 1);
-    scalar = __riscv_vfredusum_vs_f32m2_f32m1(reg, scalar, VEC_ELEM_NUM);
-    return __riscv_vfmv_f_s_f32m1_f32(scalar);
-  }
-
-  FP32Vec8 operator*(const FP32Vec8& b) const {
-    return FP32Vec8(__riscv_vfmul_vv_f32m2(reg, b.reg, VEC_ELEM_NUM));
-  }
-  FP32Vec8 operator+(const FP32Vec8& b) const {
-    return FP32Vec8(__riscv_vfadd_vv_f32m2(reg, b.reg, VEC_ELEM_NUM));
-  }
-  FP32Vec8 operator-(const FP32Vec8& b) const {
-    return FP32Vec8(__riscv_vfsub_vv_f32m2(reg, b.reg, VEC_ELEM_NUM));
-  }
-  FP32Vec8 operator/(const FP32Vec8& b) const {
-    return FP32Vec8(__riscv_vfdiv_vv_f32m2(reg, b.reg, VEC_ELEM_NUM));
-  }
-
-  FP32Vec8 min(const FP32Vec8& b) const {
-    return FP32Vec8(__riscv_vfmin_vv_f32m2(reg, b.reg, VEC_ELEM_NUM));
-  }
-  FP32Vec8 max(const FP32Vec8& b) const {
-    return FP32Vec8(__riscv_vfmax_vv_f32m2(reg, b.reg, VEC_ELEM_NUM));
-  }
-  FP32Vec8 abs() const {
-    return FP32Vec8(__riscv_vfabs_v_f32m2(reg, VEC_ELEM_NUM));
-  }
-
-  FP32Vec8 min(const FP32Vec8& b, int elem_num) const {
-    return FP32Vec8(__riscv_vfmin_vv_f32m2(reg, b.reg, elem_num));
-  }
-  FP32Vec8 max(const FP32Vec8& b, int elem_num) const {
-    return FP32Vec8(__riscv_vfmax_vv_f32m2(reg, b.reg, elem_num));
-  }
-
-  FP32Vec8 clamp(const FP32Vec8& min_v, const FP32Vec8& max_v) const {
-    fixed_vfloat32m2_t temp =
-        __riscv_vfmax_vv_f32m2(min_v.reg, reg, VEC_ELEM_NUM);
-    return FP32Vec8(__riscv_vfmin_vv_f32m2(max_v.reg, temp, VEC_ELEM_NUM));
-  }
-
-  void save(float* ptr) const { __riscv_vse32_v_f32m2(ptr, reg, VEC_ELEM_NUM); }
-  void save(float* ptr, int elem_num) const {
-    __riscv_vse32_v_f32m2(ptr, reg, elem_num);
-  }
-  void save_strided(float* ptr, ptrdiff_t stride) const {
-    ptrdiff_t byte_stride = stride * sizeof(float);
-    __riscv_vsse32_v_f32m2(ptr, byte_stride, reg, VEC_ELEM_NUM);
-  }
-
-  FP32Vec8 exp() const {
-    const float inv_ln2 = 1.44269504088896341f;
-    fixed_vfloat32m2_t x_scaled =
-        __riscv_vfmul_vf_f32m2(reg, inv_ln2, VEC_ELEM_NUM);
-    fixed_vint32m2_t n_int = __riscv_vfcvt_x_f_v_i32m2(x_scaled, VEC_ELEM_NUM);
-    fixed_vfloat32m2_t n_float = __riscv_vfcvt_f_x_v_f32m2(n_int, VEC_ELEM_NUM);
-
-    fixed_vfloat32m2_t r =
-        __riscv_vfsub_vv_f32m2(x_scaled, n_float, VEC_ELEM_NUM);
-
-    fixed_vfloat32m2_t poly =
-        __riscv_vfmv_v_f_f32m2(0.001333355810164f, VEC_ELEM_NUM);
-    poly = __riscv_vfmul_vv_f32m2(poly, r, VEC_ELEM_NUM);
-    poly = __riscv_vfadd_vf_f32m2(poly, 0.009618129107628f, VEC_ELEM_NUM);
-    poly = __riscv_vfmul_vv_f32m2(poly, r, VEC_ELEM_NUM);
-    poly = __riscv_vfadd_vf_f32m2(poly, 0.055504108664821f, VEC_ELEM_NUM);
-    poly = __riscv_vfmul_vv_f32m2(poly, r, VEC_ELEM_NUM);
-    poly = __riscv_vfadd_vf_f32m2(poly, 0.240226506959101f, VEC_ELEM_NUM);
-    poly = __riscv_vfmul_vv_f32m2(poly, r, VEC_ELEM_NUM);
-    poly = __riscv_vfadd_vf_f32m2(poly, 0.693147180559945f, VEC_ELEM_NUM);
-    poly = __riscv_vfmul_vv_f32m2(poly, r, VEC_ELEM_NUM);
-    poly = __riscv_vfadd_vf_f32m2(poly, 1.0f, VEC_ELEM_NUM);
-
-    fixed_vint32m2_t biased_exp =
-        __riscv_vadd_vx_i32m2(n_int, 127, VEC_ELEM_NUM);
-    biased_exp = __riscv_vmax_vx_i32m2(biased_exp, 0, VEC_ELEM_NUM);
-    fixed_vint32m2_t exponent_bits =
-        __riscv_vsll_vx_i32m2(biased_exp, 23, VEC_ELEM_NUM);
-    fixed_vfloat32m2_t scale =
-        __riscv_vreinterpret_v_i32m2_f32m2(exponent_bits);
-
-    return FP32Vec8(__riscv_vfmul_vv_f32m2(poly, scale, VEC_ELEM_NUM));
-  }
-
-  FP32Vec8 tanh() const {
-    fixed_vfloat32m2_t x_clamped = __riscv_vfmin_vf_f32m2(
-        __riscv_vfmax_vf_f32m2(reg, -9.0f, VEC_ELEM_NUM), 9.0f, VEC_ELEM_NUM);
-    fixed_vfloat32m2_t x2 =
-        __riscv_vfmul_vf_f32m2(x_clamped, 2.0f, VEC_ELEM_NUM);
-    FP32Vec8 exp_val = FP32Vec8(x2).exp();
-    fixed_vfloat32m2_t num =
-        __riscv_vfsub_vf_f32m2(exp_val.reg, 1.0f, VEC_ELEM_NUM);
-    fixed_vfloat32m2_t den =
-        __riscv_vfadd_vf_f32m2(exp_val.reg, 1.0f, VEC_ELEM_NUM);
-    return FP32Vec8(__riscv_vfdiv_vv_f32m2(num, den, VEC_ELEM_NUM));
-  }
-
-  FP32Vec8 er() const {
-    const float p = 0.3275911f, a1 = 0.254829592f, a2 = -0.284496736f,
-                a3 = 1.421413741f, a4 = -1.453152027f, a5 = 1.061405429f;
-    fixed_vfloat32m2_t abs_x = __riscv_vfabs_v_f32m2(reg, VEC_ELEM_NUM);
-
-    fixed_vfloat32m2_t t = __riscv_vfadd_vf_f32m2(
-        __riscv_vfmul_vf_f32m2(abs_x, p, VEC_ELEM_NUM), 1.0f, VEC_ELEM_NUM);
-    t = __riscv_vfrdiv_vf_f32m2(t, 1.0f, VEC_ELEM_NUM);
-
-    fixed_vfloat32m2_t poly = __riscv_vfmv_v_f_f32m2(a5, VEC_ELEM_NUM);
-    poly = __riscv_vfadd_vf_f32m2(__riscv_vfmul_vv_f32m2(poly, t, VEC_ELEM_NUM),
-                                  a4, VEC_ELEM_NUM);
-    poly = __riscv_vfadd_vf_f32m2(__riscv_vfmul_vv_f32m2(poly, t, VEC_ELEM_NUM),
-                                  a3, VEC_ELEM_NUM);
-    poly = __riscv_vfadd_vf_f32m2(__riscv_vfmul_vv_f32m2(poly, t, VEC_ELEM_NUM),
-                                  a2, VEC_ELEM_NUM);
-    poly = __riscv_vfadd_vf_f32m2(__riscv_vfmul_vv_f32m2(poly, t, VEC_ELEM_NUM),
-                                  a1, VEC_ELEM_NUM);
-    poly = __riscv_vfmul_vv_f32m2(poly, t, VEC_ELEM_NUM);
-
-    fixed_vfloat32m2_t exp_val =
-        FP32Vec8(__riscv_vfneg_v_f32m2(
-                     __riscv_vfmul_vv_f32m2(abs_x, abs_x, VEC_ELEM_NUM),
-                     VEC_ELEM_NUM))
-            .exp()
-            .reg;
-    fixed_vfloat32m2_t res = __riscv_vfrsub_vf_f32m2(
-        __riscv_vfmul_vv_f32m2(poly, exp_val, VEC_ELEM_NUM), 1.0f,
-        VEC_ELEM_NUM);
-
-    vbool16_t mask = __riscv_vmflt_vf_f32m2_b16(reg, 0.0f, VEC_ELEM_NUM);
-    return FP32Vec8(__riscv_vfneg_v_f32m2_m(mask, res, VEC_ELEM_NUM));
-  }
-};
-
-struct FP32Vec16 : public Vec<FP32Vec16> {
-  constexpr static int VEC_ELEM_NUM = 16;
-  fixed_vfloat32m4_t reg;
-
-  explicit FP32Vec16(float v) : reg(__riscv_vfmv_v_f_f32m4(v, VEC_ELEM_NUM)) {};
-  explicit FP32Vec16() : reg(__riscv_vfmv_v_f_f32m4(0.0f, VEC_ELEM_NUM)) {};
-  explicit FP32Vec16(const float* ptr)
-      : reg(__riscv_vle32_v_f32m4(ptr, VEC_ELEM_NUM)) {};
-  explicit FP32Vec16(fixed_vfloat32m4_t data) : reg(data) {};
-  explicit FP32Vec16(const FP32Vec8& data)
-      : reg(__riscv_vcreate_v_f32m2_f32m4(data.reg, data.reg)) {};
-  explicit FP32Vec16(const FP32Vec16& data) : reg(data.reg) {};
-  explicit FP32Vec16(const FP16Vec16& v);
-
-#ifdef RISCV_BF16_SUPPORT
-  explicit FP32Vec16(fixed_vbfloat16m2_t v)
-      : reg(__riscv_vfwcvtbf16_f_f_v_f32m4(v, VEC_ELEM_NUM)) {};
-  explicit FP32Vec16(const BF16Vec16& v)
-      : reg(__riscv_vfwcvtbf16_f_f_v_f32m4(v.reg, VEC_ELEM_NUM)) {};
-#else
-  explicit FP32Vec16(const BF16Vec16& v) : reg(v.reg_fp32) {};
-#endif
-
-  FP32Vec16 operator+(const FP32Vec16& b) const {
-    return FP32Vec16(__riscv_vfadd_vv_f32m4(reg, b.reg, VEC_ELEM_NUM));
-  }
-  FP32Vec16 operator-(const FP32Vec16& b) const {
-    return FP32Vec16(__riscv_vfsub_vv_f32m4(reg, b.reg, VEC_ELEM_NUM));
-  }
-  FP32Vec16 operator*(const FP32Vec16& b) const {
-    return FP32Vec16(__riscv_vfmul_vv_f32m4(reg, b.reg, VEC_ELEM_NUM));
-  }
-  FP32Vec16 operator/(const FP32Vec16& b) const {
-    return FP32Vec16(__riscv_vfdiv_vv_f32m4(reg, b.reg, VEC_ELEM_NUM));
-  }
-
-  FP32Vec16 fma(const FP32Vec16& a, const FP32Vec16& b) const {
-    return FP32Vec16(__riscv_vfmacc_vv_f32m4(reg, a.reg, b.reg, VEC_ELEM_NUM));
-  }
-
-  float reduce_sum() const {
-    fixed_vfloat32m1_t scalar = __riscv_vfmv_s_f_f32m1(0.0f, 1);
-    scalar = __riscv_vfredusum_vs_f32m4_f32m1(reg, scalar, VEC_ELEM_NUM);
-    return __riscv_vfmv_f_s_f32m1_f32(scalar);
-  }
-
-  float reduce_max() const {
-    fixed_vfloat32m1_t scalar =
-        __riscv_vfmv_s_f_f32m1(std::numeric_limits<float>::lowest(), 1);
-    scalar = __riscv_vfredmax_vs_f32m4_f32m1(reg, scalar, VEC_ELEM_NUM);
-    return __riscv_vfmv_f_s_f32m1_f32(scalar);
-  }
-
-  float reduce_min() const {
-    fixed_vfloat32m1_t scalar =
-        __riscv_vfmv_s_f_f32m1(std::numeric_limits<float>::max(), 1);
-    scalar = __riscv_vfredmin_vs_f32m4_f32m1(reg, scalar, VEC_ELEM_NUM);
-    return __riscv_vfmv_f_s_f32m1_f32(scalar);
-  }
-
-  template <int group_size>
-  float reduce_sub_sum(int idx) {
-    static_assert(VEC_ELEM_NUM % group_size == 0);
-    const int start = idx * group_size;
-    vuint32m4_t indices = __riscv_vid_v_u32m4(VEC_ELEM_NUM);
-    vbool8_t mask = __riscv_vmand_mm_b8(
-        __riscv_vmsgeu_vx_u32m4_b8(indices, start, VEC_ELEM_NUM),
-        __riscv_vmsltu_vx_u32m4_b8(indices, start + group_size, VEC_ELEM_NUM),
-        VEC_ELEM_NUM);
-    fixed_vfloat32m1_t scalar = __riscv_vfmv_s_f_f32m1(0.0f, 1);
-    scalar =
-        __riscv_vfredusum_vs_f32m4_f32m1_m(mask, reg, scalar, VEC_ELEM_NUM);
-    return __riscv_vfmv_f_s_f32m1_f32(scalar);
-  };
-
-  FP32Vec16 max(const FP32Vec16& b) const {
-    return FP32Vec16(__riscv_vfmax_vv_f32m4(reg, b.reg, VEC_ELEM_NUM));
-  }
-  FP32Vec16 min(const FP32Vec16& b) const {
-    return FP32Vec16(__riscv_vfmin_vv_f32m4(reg, b.reg, VEC_ELEM_NUM));
-  }
-  FP32Vec16 abs() const {
-    return FP32Vec16(__riscv_vfabs_v_f32m4(reg, VEC_ELEM_NUM));
-  }
-
-  FP32Vec16 clamp(const FP32Vec16& min_v, const FP32Vec16& max_v) const {
-    return FP32Vec16(__riscv_vfmin_vv_f32m4(
-        max_v.reg, __riscv_vfmax_vv_f32m4(min_v.reg, reg, VEC_ELEM_NUM),
-        VEC_ELEM_NUM));
-  }
-
-  void save(float* ptr) const { __riscv_vse32_v_f32m4(ptr, reg, VEC_ELEM_NUM); }
-  void save(float* ptr, int elem_num) const {
-    __riscv_vse32_v_f32m4(ptr, reg, elem_num);
-  }
-  void save_strided(float* ptr, ptrdiff_t stride) const {
-    ptrdiff_t byte_stride = stride * sizeof(float);
-    __riscv_vsse32_v_f32m4(ptr, byte_stride, reg, VEC_ELEM_NUM);
-  }
-
-  FP32Vec16 exp() const {
-    const float inv_ln2 = 1.44269504088896341f;
-    fixed_vfloat32m4_t x_scaled =
-        __riscv_vfmul_vf_f32m4(reg, inv_ln2, VEC_ELEM_NUM);
-    fixed_vint32m4_t n_int = __riscv_vfcvt_x_f_v_i32m4(x_scaled, VEC_ELEM_NUM);
-    fixed_vfloat32m4_t n_float = __riscv_vfcvt_f_x_v_f32m4(n_int, VEC_ELEM_NUM);
-    fixed_vfloat32m4_t r =
-        __riscv_vfsub_vv_f32m4(x_scaled, n_float, VEC_ELEM_NUM);
-
-    fixed_vfloat32m4_t poly =
-        __riscv_vfmv_v_f_f32m4(0.001333355810164f, VEC_ELEM_NUM);
-    poly = __riscv_vfadd_vf_f32m4(__riscv_vfmul_vv_f32m4(poly, r, VEC_ELEM_NUM),
-                                  0.009618129107628f, VEC_ELEM_NUM);
-    poly = __riscv_vfadd_vf_f32m4(__riscv_vfmul_vv_f32m4(poly, r, VEC_ELEM_NUM),
-                                  0.055504108664821f, VEC_ELEM_NUM);
-    poly = __riscv_vfadd_vf_f32m4(__riscv_vfmul_vv_f32m4(poly, r, VEC_ELEM_NUM),
-                                  0.240226506959101f, VEC_ELEM_NUM);
-    poly = __riscv_vfadd_vf_f32m4(__riscv_vfmul_vv_f32m4(poly, r, VEC_ELEM_NUM),
-                                  0.693147180559945f, VEC_ELEM_NUM);
-    poly = __riscv_vfadd_vf_f32m4(__riscv_vfmul_vv_f32m4(poly, r, VEC_ELEM_NUM),
-                                  1.0f, VEC_ELEM_NUM);
-
-    fixed_vint32m4_t biased_exp = __riscv_vmax_vx_i32m4(
-        __riscv_vadd_vx_i32m4(n_int, 127, VEC_ELEM_NUM), 0, VEC_ELEM_NUM);
-    fixed_vfloat32m4_t scale = __riscv_vreinterpret_v_i32m4_f32m4(
-        __riscv_vsll_vx_i32m4(biased_exp, 23, VEC_ELEM_NUM));
-
-    return FP32Vec16(__riscv_vfmul_vv_f32m4(poly, scale, VEC_ELEM_NUM));
-  }
-
-  FP32Vec16 tanh() const {
-    fixed_vfloat32m4_t x_clamped = __riscv_vfmin_vf_f32m4(
-        __riscv_vfmax_vf_f32m4(reg, -9.0f, VEC_ELEM_NUM), 9.0f, VEC_ELEM_NUM);
-    FP32Vec16 exp_val =
-        FP32Vec16(__riscv_vfmul_vf_f32m4(x_clamped, 2.0f, VEC_ELEM_NUM)).exp();
-    return FP32Vec16(__riscv_vfdiv_vv_f32m4(
-        __riscv_vfsub_vf_f32m4(exp_val.reg, 1.0f, VEC_ELEM_NUM),
-        __riscv_vfadd_vf_f32m4(exp_val.reg, 1.0f, VEC_ELEM_NUM), VEC_ELEM_NUM));
-  }
-
-  FP32Vec16 er() const {
-    const float p = 0.3275911f, a1 = 0.254829592f, a2 = -0.284496736f,
-                a3 = 1.421413741f, a4 = -1.453152027f, a5 = 1.061405429f;
-    fixed_vfloat32m4_t abs_x = __riscv_vfabs_v_f32m4(reg, VEC_ELEM_NUM);
-    fixed_vfloat32m4_t t = __riscv_vfrdiv_vf_f32m4(
-        __riscv_vfadd_vf_f32m4(__riscv_vfmul_vf_f32m4(abs_x, p, VEC_ELEM_NUM),
-                               1.0f, VEC_ELEM_NUM),
-        1.0f, VEC_ELEM_NUM);
-
-    fixed_vfloat32m4_t poly = __riscv_vfmv_v_f_f32m4(a5, VEC_ELEM_NUM);
-    poly = __riscv_vfadd_vf_f32m4(__riscv_vfmul_vv_f32m4(poly, t, VEC_ELEM_NUM),
-                                  a4, VEC_ELEM_NUM);
-    poly = __riscv_vfadd_vf_f32m4(__riscv_vfmul_vv_f32m4(poly, t, VEC_ELEM_NUM),
-                                  a3, VEC_ELEM_NUM);
-    poly = __riscv_vfadd_vf_f32m4(__riscv_vfmul_vv_f32m4(poly, t, VEC_ELEM_NUM),
-                                  a2, VEC_ELEM_NUM);
-    poly = __riscv_vfadd_vf_f32m4(__riscv_vfmul_vv_f32m4(poly, t, VEC_ELEM_NUM),
-                                  a1, VEC_ELEM_NUM);
-    poly = __riscv_vfmul_vv_f32m4(poly, t, VEC_ELEM_NUM);
-
-    fixed_vfloat32m4_t exp_val =
-        FP32Vec16(__riscv_vfneg_v_f32m4(
-                      __riscv_vfmul_vv_f32m4(abs_x, abs_x, VEC_ELEM_NUM),
-                      VEC_ELEM_NUM))
-            .exp()
-            .reg;
-    fixed_vfloat32m4_t res = __riscv_vfrsub_vf_f32m4(
-        __riscv_vfmul_vv_f32m4(poly, exp_val, VEC_ELEM_NUM), 1.0f,
-        VEC_ELEM_NUM);
-
-    vbool8_t mask = __riscv_vmflt_vf_f32m4_b8(reg, 0.0f, VEC_ELEM_NUM);
-    return FP32Vec16(__riscv_vfneg_v_f32m4_m(mask, res, VEC_ELEM_NUM));
-  }
-};
-
-// ============================================================================
-// Type Traits & Global Helpers
-// ============================================================================
-
-template <typename T>
-struct VecType {
-  using vec_type = void;
-  using vec_t = void;
-};
-
-template <typename T>
-using vec_t = typename VecType<T>::vec_type;
-
-template <>
-struct VecType<float> {
-  using vec_type = FP32Vec8;
-  using vec_t = FP32Vec8;
-};
-template <>
-struct VecType<c10::Half> {
-  using vec_type = FP16Vec8;
-  using vec_t = FP16Vec8;
-};
-template <>
-struct VecType<c10::BFloat16> {
-  using vec_type = BF16Vec8;
-  using vec_t = BF16Vec8;
-};
-
-template <typename T>
-void storeFP32(float v, T* ptr) {
-  *ptr = v;
-}
-template <>
-inline void storeFP32<c10::Half>(float v, c10::Half* ptr) {
-  *reinterpret_cast<_Float16*>(ptr) = static_cast<_Float16>(v);
-}
-
-inline FP16Vec16::FP16Vec16(const FP32Vec16& v) {
-  reg = __riscv_vfncvt_f_f_w_f16m2(v.reg, VEC_ELEM_NUM);
-}
-inline FP16Vec8::FP16Vec8(const FP32Vec8& v) {
-  reg = __riscv_vfncvt_f_f_w_f16m1(v.reg, VEC_ELEM_NUM);
-}
-inline FP32Vec16::FP32Vec16(const FP16Vec16& v) {
-  reg = __riscv_vfwcvt_f_f_v_f32m4(v.reg, VEC_ELEM_NUM);
-}
-inline void fma(FP32Vec16& acc, const FP32Vec16& a, const FP32Vec16& b) {
-  acc = acc.fma(a, b);
-}
-
-#ifdef RISCV_BF16_SUPPORT
-template <>
-inline void storeFP32<c10::BFloat16>(float v, c10::BFloat16* ptr) {
-  *ptr = static_cast<__bf16>(v);
-};
-inline BF16Vec8::BF16Vec8(const FP32Vec8& v)
-    : reg(__riscv_vfncvtbf16_f_f_w_bf16m1(v.reg, VEC_ELEM_NUM)) {};
-inline BF16Vec16::BF16Vec16(const FP32Vec16& v)
-    : reg(__riscv_vfncvtbf16_f_f_w_bf16m2(v.reg, VEC_ELEM_NUM)) {};
-#else
-template <>
-inline void storeFP32<c10::BFloat16>(float v, c10::BFloat16* ptr) {
-  uint32_t val;
-  std::memcpy(&val, &v, 4);
-  *reinterpret_cast<uint16_t*>(ptr) = static_cast<uint16_t>(val >> 16);
-}
-inline BF16Vec8::BF16Vec8(const FP32Vec8& v) : reg_fp32(v.reg) {}
-inline BF16Vec16::BF16Vec16(const FP32Vec16& v) : reg_fp32(v.reg) {}
-#endif
-
-inline void prefetch(const void* addr) { __builtin_prefetch(addr, 0, 1); }
-
-}  // namespace vec_op
-
-#ifndef CPU_KERNEL_GUARD_IN
-  #define CPU_KERNEL_GUARD_IN(NAME)
-#endif
-
-#ifndef CPU_KERNEL_GUARD_OUT
-  #define CPU_KERNEL_GUARD_OUT(NAME)
-#endif
-
-#endif  // CPU_TYPES_RISCV_HPP
+#endif  // CPU_TYPES_RISCV_HPP

csrc/cpu/cpu_types_riscv_defs.hpp

@@ -0,0 +1,98 @@
+#ifndef CPU_TYPES_RISCV_DEFS_HPP
+#define CPU_TYPES_RISCV_DEFS_HPP
+
+// VLEN-to-LMUL mapping for RISC-V Vector extension.
+//
+// LMUL_<N> expands to the LMUL suffix giving N total bits of vector data:
+//   VLEN=128: LMUL_128=m1,  LMUL_256=m2,  LMUL_512=m4,  LMUL_1024=m8
+//   VLEN=256: LMUL_128=mf2, LMUL_256=m1,  LMUL_512=m2,  LMUL_1024=m4
+
+#include <riscv_vector.h>
+
+#if __riscv_v_min_vlen == 128
+  #define LMUL_128 m1
+  #define LMUL_256 m2
+  #define LMUL_512 m4
+  #define LMUL_1024 m8
+  #define BOOL_256 b16
+  #define BOOL_512 b8
+#elif __riscv_v_min_vlen == 256
+  #define LMUL_128 mf2
+  #define LMUL_256 m1
+  #define LMUL_512 m2
+  #define LMUL_1024 m4
+  #define BOOL_256 b32
+  #define BOOL_512 b16
+#else
+  #error "cpu_types_riscv_defs.hpp: unsupported __riscv_v_min_vlen"
+#endif
+
+// Token-paste helpers.
+#define _RVV_P2(a, b) a##b
+#define _RVV_P3(a, b, c) a##b##c
+#define _RVV_P4(a, b, c, d) a##b##c##d
+#define RVVTYPE(base, lmul, suffix) _RVV_P3(base, lmul, suffix)
+#define RVVI(base, lmul) _RVV_P2(base, lmul)
+#define RVVI3(base, lmul, suffix) _RVV_P3(base, lmul, suffix)
+#define RVVI4(a, b, c, d) _RVV_P4(a, b, c, d)
+// For mask intrinsics: RVVIB(base, LMUL_256, BOOL_256) → base##m2##_##b16
+#define _RVV_PB(base, lmul, btype) base##lmul##_##btype
+#define RVVIB(base, lmul, btype) _RVV_PB(base, lmul, btype)
+
+// ---- Semantic fixed-vector typedefs (named by element count) ----
+
+// float16
+typedef RVVTYPE(vfloat16, LMUL_128, _t) fixed_fp16x8_t
+    __attribute__((riscv_rvv_vector_bits(128)));
+typedef RVVTYPE(vfloat16, LMUL_256, _t) fixed_fp16x16_t
+    __attribute__((riscv_rvv_vector_bits(256)));
+
+// float32
+typedef RVVTYPE(vfloat32, LMUL_128, _t) fixed_fp32x4_t
+    __attribute__((riscv_rvv_vector_bits(128)));
+typedef RVVTYPE(vfloat32, LMUL_256, _t) fixed_fp32x8_t
+    __attribute__((riscv_rvv_vector_bits(256)));
+typedef RVVTYPE(vfloat32, LMUL_512, _t) fixed_fp32x16_t
+    __attribute__((riscv_rvv_vector_bits(512)));
+typedef RVVTYPE(vfloat32, LMUL_1024, _t) fixed_fp32x32_t
+    __attribute__((riscv_rvv_vector_bits(1024)));
+
+// int32
+typedef RVVTYPE(vint32, LMUL_256, _t) fixed_i32x8_t
+    __attribute__((riscv_rvv_vector_bits(256)));
+typedef RVVTYPE(vint32, LMUL_512, _t) fixed_i32x16_t
+    __attribute__((riscv_rvv_vector_bits(512)));
+
+// uint16
+typedef RVVTYPE(vuint16, LMUL_128, _t) fixed_u16x8_t
+    __attribute__((riscv_rvv_vector_bits(128)));
+typedef RVVTYPE(vuint16, LMUL_256, _t) fixed_u16x16_t
+    __attribute__((riscv_rvv_vector_bits(256)));
+typedef RVVTYPE(vuint16, LMUL_512, _t) fixed_u16x32_t
+    __attribute__((riscv_rvv_vector_bits(512)));
+
+// bfloat16
+#ifdef RISCV_BF16_SUPPORT
+typedef RVVTYPE(vbfloat16, LMUL_128, _t) fixed_bf16x8_t
+    __attribute__((riscv_rvv_vector_bits(128)));
+typedef RVVTYPE(vbfloat16, LMUL_256, _t) fixed_bf16x16_t
+    __attribute__((riscv_rvv_vector_bits(256)));
+typedef RVVTYPE(vbfloat16, LMUL_512, _t) fixed_bf16x32_t
+    __attribute__((riscv_rvv_vector_bits(512)));
+#endif
+
+// ---- Reduction accumulator type (always m1 = one register of f32) ----
+// Used for scalar reductions; only element [0] is meaningful.
+typedef vfloat32m1_t rvv_f32_accum_t
+    __attribute__((riscv_rvv_vector_bits(__riscv_v_min_vlen)));
+
+// ---- Mask types for f32 elements ----
+#if __riscv_v_min_vlen == 128
+typedef vbool16_t rvv_mask_f32x8_t;
+typedef vbool8_t rvv_mask_f32x16_t;
+#elif __riscv_v_min_vlen == 256
+typedef vbool32_t rvv_mask_f32x8_t;
+typedef vbool16_t rvv_mask_f32x16_t;
+#endif
+
+#endif  // CPU_TYPES_RISCV_DEFS_HPP

csrc/cpu/cpu_types_riscv_impl.hpp

@@ -0,0 +1,905 @@
+#ifndef CPU_TYPES_RISCV_IMPL_HPP
+#define CPU_TYPES_RISCV_IMPL_HPP
+
+// Shared implementation of RVV vector-type wrapper classes.
+// This file is VLEN-independent: it uses the semantic type names and
+// RVVI() intrinsic macros from cpu_types_riscv_defs.hpp.
+//
+// DO NOT include this file directly; include cpu_types_riscv.hpp instead.
+
+#include <algorithm>
+#include <cmath>
+#include <cstring>
+#include <iostream>
+#include <limits>
+#include <torch/all.h>
+namespace vec_op {
+
+// BFloat16 is always supported on RISC-V: natively when RISCV_BF16_SUPPORT
+// is defined, otherwise via the FP32-simulation fallback path.
+#define VLLM_DISPATCH_CASE_FLOATING_TYPES(...)         \
+  AT_DISPATCH_CASE(at::ScalarType::Float, __VA_ARGS__) \
+  AT_DISPATCH_CASE(at::ScalarType::Half, __VA_ARGS__)  \
+  AT_DISPATCH_CASE(at::ScalarType::BFloat16, __VA_ARGS__)
+
+#define VLLM_DISPATCH_FLOATING_TYPES(TYPE, NAME, ...) \
+  AT_DISPATCH_SWITCH(TYPE, NAME, VLLM_DISPATCH_CASE_FLOATING_TYPES(__VA_ARGS__))
+
+#define FORCE_INLINE __attribute__((always_inline)) inline
+
+namespace {
+template <typename T, T... indexes, typename F>
+constexpr void unroll_loop_item(std::integer_sequence<T, indexes...>, F&& f) {
+  (f(std::integral_constant<T, indexes>{}), ...);
+};
+}  // namespace
+
+template <typename T, T count, typename F,
+          typename = std::enable_if_t<std::is_invocable_v<F, T>>>
+constexpr void unroll_loop(F&& f) {
+  unroll_loop_item(std::make_integer_sequence<T, count>{}, std::forward<F>(f));
+}
+
+template <typename T>
+struct Vec {
+  constexpr static int get_elem_num() { return T::VEC_ELEM_NUM; };
+};
+
+struct FP32Vec8;
+struct FP32Vec16;
+
+// ============================================================================
+// FP16 Implementation
+// ============================================================================
+
+struct FP16Vec8 : public Vec<FP16Vec8> {
+  constexpr static int VEC_ELEM_NUM = 8;
+  fixed_fp16x8_t reg;
+
+  explicit FP16Vec8(const void* ptr)
+      : reg(RVVI(__riscv_vle16_v_f16, LMUL_128)(
+            static_cast<const _Float16*>(ptr), VEC_ELEM_NUM)) {};
+
+  explicit FP16Vec8(const FP32Vec8&);
+
+  void save(void* ptr) const {
+    RVVI(__riscv_vse16_v_f16, LMUL_128)(static_cast<_Float16*>(ptr), reg,
+                                        VEC_ELEM_NUM);
+  }
+  void save(void* ptr, int elem_num) const {
+    RVVI(__riscv_vse16_v_f16, LMUL_128)(static_cast<_Float16*>(ptr), reg,
+                                        elem_num);
+  }
+  void save_strided(void* ptr, ptrdiff_t stride) const {
+    ptrdiff_t byte_stride = stride * sizeof(_Float16);
+    RVVI(__riscv_vsse16_v_f16, LMUL_128)(static_cast<_Float16*>(ptr),
+                                         byte_stride, reg, VEC_ELEM_NUM);
+  }
+};
+
+struct FP16Vec16 : public Vec<FP16Vec16> {
+  constexpr static int VEC_ELEM_NUM = 16;
+  fixed_fp16x16_t reg;
+
+  explicit FP16Vec16(const void* ptr)
+      : reg(RVVI(__riscv_vle16_v_f16, LMUL_256)(
+            static_cast<const _Float16*>(ptr), VEC_ELEM_NUM)) {};
+
+  explicit FP16Vec16(const FP32Vec16& vec);
+
+  void save(void* ptr) const {
+    RVVI(__riscv_vse16_v_f16, LMUL_256)(static_cast<_Float16*>(ptr), reg,
+                                        VEC_ELEM_NUM);
+  }
+  void save(void* ptr, int elem_num) const {
+    RVVI(__riscv_vse16_v_f16, LMUL_256)(static_cast<_Float16*>(ptr), reg,
+                                        elem_num);
+  }
+  void save_strided(void* ptr, ptrdiff_t stride) const {
+    ptrdiff_t byte_stride = stride * sizeof(_Float16);
+    RVVI(__riscv_vsse16_v_f16, LMUL_256)(static_cast<_Float16*>(ptr),
+                                         byte_stride, reg, VEC_ELEM_NUM);
+  }
+};
+
+// ============================================================================
+// BF16 Implementation
+// ============================================================================
+
+#ifdef RISCV_BF16_SUPPORT
+
+FORCE_INLINE fixed_u16x8_t bf16_to_u16(fixed_bf16x8_t v) {
+  return RVVI4(__riscv_vreinterpret_v_bf16, LMUL_128, _u16, LMUL_128)(v);
+}
+FORCE_INLINE fixed_u16x16_t bf16_to_u16(fixed_bf16x16_t v) {
+  return RVVI4(__riscv_vreinterpret_v_bf16, LMUL_256, _u16, LMUL_256)(v);
+}
+FORCE_INLINE fixed_u16x32_t bf16_to_u16(fixed_bf16x32_t v) {
+  return RVVI4(__riscv_vreinterpret_v_bf16, LMUL_512, _u16, LMUL_512)(v);
+}
+
+struct BF16Vec8 : public Vec<BF16Vec8> {
+  constexpr static int VEC_ELEM_NUM = 8;
+  fixed_bf16x8_t reg;
+
+  explicit BF16Vec8(const void* ptr)
+      : reg(RVVI4(__riscv_vreinterpret_v_u16, LMUL_128, _bf16,
+                  LMUL_128)(RVVI(__riscv_vle16_v_u16, LMUL_128)(
+            reinterpret_cast<const uint16_t*>(ptr), VEC_ELEM_NUM))) {};
+
+  explicit BF16Vec8(fixed_bf16x8_t data) : reg(data) {};
+  explicit BF16Vec8(const FP32Vec8&);
+
+  void save(void* ptr) const {
+    RVVI(__riscv_vse16_v_u16, LMUL_128)(reinterpret_cast<uint16_t*>(ptr),
+                                        bf16_to_u16(reg), VEC_ELEM_NUM);
+  }
+  void save(void* ptr, int elem_num) const {
+    RVVI(__riscv_vse16_v_u16, LMUL_128)(reinterpret_cast<uint16_t*>(ptr),
+                                        bf16_to_u16(reg), elem_num);
+  }
+  void save_strided(void* ptr, ptrdiff_t stride) const {
+    ptrdiff_t byte_stride = stride * sizeof(uint16_t);
+    RVVI(__riscv_vsse16_v_u16, LMUL_128)(reinterpret_cast<uint16_t*>(ptr),
+                                         byte_stride, bf16_to_u16(reg),
+                                         VEC_ELEM_NUM);
+  }
+};
+
+struct BF16Vec16 : public Vec<BF16Vec16> {
+  constexpr static int VEC_ELEM_NUM = 16;
+  fixed_bf16x16_t reg;
+
+  explicit BF16Vec16(const void* ptr)
+      : reg(RVVI4(__riscv_vreinterpret_v_u16, LMUL_256, _bf16,
+                  LMUL_256)(RVVI(__riscv_vle16_v_u16, LMUL_256)(
+            reinterpret_cast<const uint16_t*>(ptr), VEC_ELEM_NUM))) {};
+
+  explicit BF16Vec16(fixed_bf16x16_t data) : reg(data) {};
+  explicit BF16Vec16(const FP32Vec16&);
+
+  void save(void* ptr) const {
+    RVVI(__riscv_vse16_v_u16, LMUL_256)(reinterpret_cast<uint16_t*>(ptr),
+                                        bf16_to_u16(reg), VEC_ELEM_NUM);
+  }
+  void save(void* ptr, int elem_num) const {
+    RVVI(__riscv_vse16_v_u16, LMUL_256)(reinterpret_cast<uint16_t*>(ptr),
+                                        bf16_to_u16(reg), elem_num);
+  }
+  void save_strided(void* ptr, ptrdiff_t stride) const {
+    ptrdiff_t byte_stride = stride * sizeof(uint16_t);
+    RVVI(__riscv_vsse16_v_u16, LMUL_256)(reinterpret_cast<uint16_t*>(ptr),
+                                         byte_stride, bf16_to_u16(reg),
+                                         VEC_ELEM_NUM);
+  }
+};
+
+struct BF16Vec32 : public Vec<BF16Vec32> {
+  constexpr static int VEC_ELEM_NUM = 32;
+  fixed_bf16x32_t reg;
+
+  explicit BF16Vec32(const void* ptr)
+      : reg(RVVI4(__riscv_vreinterpret_v_u16, LMUL_512, _bf16,
+                  LMUL_512)(RVVI(__riscv_vle16_v_u16, LMUL_512)(
+            reinterpret_cast<const uint16_t*>(ptr), VEC_ELEM_NUM))) {};
+
+  explicit BF16Vec32(fixed_bf16x32_t data) : reg(data) {};
+
+  explicit BF16Vec32(const BF16Vec8& v) {
+    fixed_u16x8_t u16_val = bf16_to_u16(v.reg);
+    fixed_u16x32_t u16_combined =
+        RVVI4(__riscv_vcreate_v_u16, LMUL_128, _u16, LMUL_512)(
+            u16_val, u16_val, u16_val, u16_val);
+    reg = RVVI4(__riscv_vreinterpret_v_u16, LMUL_512, _bf16,
+                LMUL_512)(u16_combined);
+  };
+
+  void save(void* ptr) const {
+    RVVI(__riscv_vse16_v_u16, LMUL_512)(reinterpret_cast<uint16_t*>(ptr),
+                                        bf16_to_u16(reg), VEC_ELEM_NUM);
+  }
+  void save(void* ptr, int elem_num) const {
+    RVVI(__riscv_vse16_v_u16, LMUL_512)(reinterpret_cast<uint16_t*>(ptr),
+                                        bf16_to_u16(reg), elem_num);
+  }
+  void save_strided(void* ptr, ptrdiff_t stride) const {
+    ptrdiff_t byte_stride = stride * sizeof(uint16_t);
+    RVVI(__riscv_vsse16_v_u16, LMUL_512)(reinterpret_cast<uint16_t*>(ptr),
+                                         byte_stride, bf16_to_u16(reg),
+                                         VEC_ELEM_NUM);
+  }
+};
+
+#else
+// ============================================================================
+// BF16 Fallback Implementation (FP32 Simulation)
+// ============================================================================
+
+struct BF16Vec8 : public Vec<BF16Vec8> {
+  constexpr static int VEC_ELEM_NUM = 8;
+  fixed_fp32x8_t reg_fp32;
+  explicit BF16Vec8(const void* ptr) {
+    const uint16_t* u16 = static_cast<const uint16_t*>(ptr);
+    float tmp[8];
+    for (int i = 0; i < 8; ++i) {
+      uint32_t v = static_cast<uint32_t>(u16[i]) << 16;
+      std::memcpy(&tmp[i], &v, 4);
+    }
+    reg_fp32 = RVVI(__riscv_vle32_v_f32, LMUL_256)(tmp, 8);
+  }
+  explicit BF16Vec8(const FP32Vec8&);
+  void save(void* ptr) const {
+    float tmp[8];
+    RVVI(__riscv_vse32_v_f32, LMUL_256)(tmp, reg_fp32, 8);
+    uint16_t* u16 = static_cast<uint16_t*>(ptr);
+    for (int i = 0; i < 8; ++i) {
+      uint32_t v;
+      std::memcpy(&v, &tmp[i], 4);
+      u16[i] = static_cast<uint16_t>(v >> 16);
+    }
+  }
+  void save(void* ptr, int elem_num) const {
+    float tmp[8];
+    RVVI(__riscv_vse32_v_f32, LMUL_256)(tmp, reg_fp32, 8);
+    uint16_t* u16 = static_cast<uint16_t*>(ptr);
+    for (int i = 0; i < elem_num; ++i) {
+      uint32_t v;
+      std::memcpy(&v, &tmp[i], 4);
+      u16[i] = static_cast<uint16_t>(v >> 16);
+    }
+  }
+  void save_strided(void* ptr, ptrdiff_t stride) const {
+    float tmp[8];
+    RVVI(__riscv_vse32_v_f32, LMUL_256)(tmp, reg_fp32, 8);
+    uint8_t* u8 = static_cast<uint8_t*>(ptr);
+    ptrdiff_t byte_stride = stride * sizeof(uint16_t);
+    for (int i = 0; i < 8; ++i) {
+      uint32_t v;
+      std::memcpy(&v, &tmp[i], 4);
+      uint16_t val = static_cast<uint16_t>(v >> 16);
+      *reinterpret_cast<uint16_t*>(u8 + i * byte_stride) = val;
+    }
+  }
+};
+
+struct BF16Vec16 : public Vec<BF16Vec16> {
+  constexpr static int VEC_ELEM_NUM = 16;
+  fixed_fp32x16_t reg_fp32;
+  explicit BF16Vec16(const void* ptr) {
+    const uint16_t* u16 = static_cast<const uint16_t*>(ptr);
+    float tmp[16];
+    for (int i = 0; i < 16; ++i) {
+      uint32_t v = static_cast<uint32_t>(u16[i]) << 16;
+      std::memcpy(&tmp[i], &v, 4);
+    }
+    reg_fp32 = RVVI(__riscv_vle32_v_f32, LMUL_512)(tmp, 16);
+  }
+  explicit BF16Vec16(const FP32Vec16&);
+  void save(void* ptr) const {
+    float tmp[16];
+    RVVI(__riscv_vse32_v_f32, LMUL_512)(tmp, reg_fp32, 16);
+    uint16_t* u16 = static_cast<uint16_t*>(ptr);
+    for (int i = 0; i < 16; ++i) {
+      uint32_t v;
+      std::memcpy(&v, &tmp[i], 4);
+      u16[i] = static_cast<uint16_t>(v >> 16);
+    }
+  }
+  void save(void* ptr, int elem_num) const {
+    float tmp[16];
+    RVVI(__riscv_vse32_v_f32, LMUL_512)(tmp, reg_fp32, 16);
+    uint16_t* u16 = static_cast<uint16_t*>(ptr);
+    for (int i = 0; i < elem_num; ++i) {
+      uint32_t v;
+      std::memcpy(&v, &tmp[i], 4);
+      u16[i] = static_cast<uint16_t>(v >> 16);
+    }
+  }
+  void save_strided(void* ptr, ptrdiff_t stride) const {
+    float tmp[16];
+    RVVI(__riscv_vse32_v_f32, LMUL_512)(tmp, reg_fp32, 16);
+    uint8_t* u8 = static_cast<uint8_t*>(ptr);
+    ptrdiff_t byte_stride = stride * sizeof(uint16_t);
+    for (int i = 0; i < 16; ++i) {
+      uint32_t v;
+      std::memcpy(&v, &tmp[i], 4);
+      uint16_t val = static_cast<uint16_t>(v >> 16);
+      *reinterpret_cast<uint16_t*>(u8 + i * byte_stride) = val;
+    }
+  }
+};
+
+struct BF16Vec32 : public Vec<BF16Vec32> {
+  constexpr static int VEC_ELEM_NUM = 32;
+  fixed_fp32x32_t reg_fp32;
+
+  explicit BF16Vec32(const void* ptr) {
+    const uint16_t* u16 = static_cast<const uint16_t*>(ptr);
+    float tmp[32];
+    for (int i = 0; i < 32; ++i) {
+      uint32_t v = static_cast<uint32_t>(u16[i]) << 16;
+      std::memcpy(&tmp[i], &v, 4);
+    }
+    reg_fp32 = RVVI(__riscv_vle32_v_f32, LMUL_1024)(tmp, 32);
+  }
+
+  explicit BF16Vec32(const BF16Vec8& v) {
+    float tmp_small[8];
+    RVVI(__riscv_vse32_v_f32, LMUL_256)(tmp_small, v.reg_fp32, 8);
+    float tmp_large[32];
+    for (int i = 0; i < 4; ++i) {
+      std::memcpy(tmp_large + (i * 8), tmp_small, 8 * sizeof(float));
+    }
+    reg_fp32 = RVVI(__riscv_vle32_v_f32, LMUL_1024)(tmp_large, 32);
+  }
+
+  void save(void* ptr) const {
+    float tmp[32];
+    RVVI(__riscv_vse32_v_f32, LMUL_1024)(tmp, reg_fp32, 32);
+    uint16_t* u16 = static_cast<uint16_t*>(ptr);
+    for (int i = 0; i < 32; ++i) {
+      uint32_t v;
+      std::memcpy(&v, &tmp[i], 4);
+      u16[i] = static_cast<uint16_t>(v >> 16);
+    }
+  }
+
+  void save(void* ptr, int elem_num) const {
+    float tmp[32];
+    RVVI(__riscv_vse32_v_f32, LMUL_1024)(tmp, reg_fp32, 32);
+    uint16_t* u16 = static_cast<uint16_t*>(ptr);
+    for (int i = 0; i < elem_num; ++i) {
+      uint32_t v;
+      std::memcpy(&v, &tmp[i], 4);
+      u16[i] = static_cast<uint16_t>(v >> 16);
+    }
+  }
+
+  void save_strided(void* ptr, ptrdiff_t stride) const {
+    float tmp[32];
+    RVVI(__riscv_vse32_v_f32, LMUL_1024)(tmp, reg_fp32, 32);
+    uint8_t* u8 = static_cast<uint8_t*>(ptr);
+    ptrdiff_t byte_stride = stride * sizeof(uint16_t);
+    for (int i = 0; i < 32; ++i) {
+      uint32_t v;
+      std::memcpy(&v, &tmp[i], 4);
+      uint16_t val = static_cast<uint16_t>(v >> 16);
+      *reinterpret_cast<uint16_t*>(u8 + i * byte_stride) = val;
+    }
+  }
+};
+#endif
+
+// ============================================================================
+// FP32 Implementation
+// ============================================================================
+
+struct FP32Vec4 : public Vec<FP32Vec4> {
+  constexpr static int VEC_ELEM_NUM = 4;
+  fixed_fp32x4_t reg;
+  explicit FP32Vec4(float v)
+      : reg(RVVI(__riscv_vfmv_v_f_f32, LMUL_128)(v, VEC_ELEM_NUM)) {};
+  explicit FP32Vec4()
+      : reg(RVVI(__riscv_vfmv_v_f_f32, LMUL_128)(0.0f, VEC_ELEM_NUM)) {};
+  explicit FP32Vec4(const float* ptr)
+      : reg(RVVI(__riscv_vle32_v_f32, LMUL_128)(ptr, VEC_ELEM_NUM)) {};
+  explicit FP32Vec4(fixed_fp32x4_t data) : reg(data) {};
+  explicit FP32Vec4(const FP32Vec4& data) : reg(data.reg) {};
+  void save(float* ptr) const {
+    RVVI(__riscv_vse32_v_f32, LMUL_128)(ptr, reg, VEC_ELEM_NUM);
+  }
+  void save(float* ptr, int elem_num) const {
+    RVVI(__riscv_vse32_v_f32, LMUL_128)(ptr, reg, elem_num);
+  }
+};
+
+struct FP32Vec8 : public Vec<FP32Vec8> {
+  constexpr static int VEC_ELEM_NUM = 8;
+  fixed_fp32x8_t reg;
+
+  explicit FP32Vec8(float v)
+      : reg(RVVI(__riscv_vfmv_v_f_f32, LMUL_256)(v, VEC_ELEM_NUM)) {};
+  explicit FP32Vec8()
+      : reg(RVVI(__riscv_vfmv_v_f_f32, LMUL_256)(0.0f, VEC_ELEM_NUM)) {};
+  explicit FP32Vec8(const float* ptr)
+      : reg(RVVI(__riscv_vle32_v_f32, LMUL_256)(ptr, VEC_ELEM_NUM)) {};
+  explicit FP32Vec8(fixed_fp32x8_t data) : reg(data) {};
+  explicit FP32Vec8(const FP32Vec8& data) : reg(data.reg) {};
+  explicit FP32Vec8(const FP16Vec8& v)
+      : reg(RVVI(__riscv_vfwcvt_f_f_v_f32, LMUL_256)(v.reg, VEC_ELEM_NUM)) {};
+  explicit FP32Vec8(fixed_fp16x8_t v)
+      : reg(RVVI(__riscv_vfwcvt_f_f_v_f32, LMUL_256)(v, VEC_ELEM_NUM)) {};
+
+#ifdef RISCV_BF16_SUPPORT
+  explicit FP32Vec8(fixed_bf16x8_t v)
+      : reg(RVVI(__riscv_vfwcvtbf16_f_f_v_f32, LMUL_256)(v, VEC_ELEM_NUM)) {};
+  explicit FP32Vec8(const BF16Vec8& v)
+      : reg(RVVI(__riscv_vfwcvtbf16_f_f_v_f32, LMUL_256)(v.reg, VEC_ELEM_NUM)) {
+        };
+#else
+  explicit FP32Vec8(const BF16Vec8& v) : reg(v.reg_fp32) {};
+#endif
+
+  float reduce_sum() const {
+    rvv_f32_accum_t scalar = __riscv_vfmv_s_f_f32m1(0.0f, 1);
+    scalar = RVVI3(__riscv_vfredusum_vs_f32, LMUL_256, _f32m1)(reg, scalar,
+                                                               VEC_ELEM_NUM);
+    return __riscv_vfmv_f_s_f32m1_f32(scalar);
+  }
+
+  FP32Vec8 operator*(const FP32Vec8& b) const {
+    return FP32Vec8(
+        RVVI(__riscv_vfmul_vv_f32, LMUL_256)(reg, b.reg, VEC_ELEM_NUM));
+  }
+  FP32Vec8 operator+(const FP32Vec8& b) const {
+    return FP32Vec8(
+        RVVI(__riscv_vfadd_vv_f32, LMUL_256)(reg, b.reg, VEC_ELEM_NUM));
+  }
+  FP32Vec8 operator-(const FP32Vec8& b) const {
+    return FP32Vec8(
+        RVVI(__riscv_vfsub_vv_f32, LMUL_256)(reg, b.reg, VEC_ELEM_NUM));
+  }
+  FP32Vec8 operator/(const FP32Vec8& b) const {
+    return FP32Vec8(
+        RVVI(__riscv_vfdiv_vv_f32, LMUL_256)(reg, b.reg, VEC_ELEM_NUM));
+  }
+
+  FP32Vec8 min(const FP32Vec8& b) const {
+    return FP32Vec8(
+        RVVI(__riscv_vfmin_vv_f32, LMUL_256)(reg, b.reg, VEC_ELEM_NUM));
+  }
+  FP32Vec8 max(const FP32Vec8& b) const {
+    return FP32Vec8(
+        RVVI(__riscv_vfmax_vv_f32, LMUL_256)(reg, b.reg, VEC_ELEM_NUM));
+  }
+  FP32Vec8 abs() const {
+    return FP32Vec8(RVVI(__riscv_vfabs_v_f32, LMUL_256)(reg, VEC_ELEM_NUM));
+  }
+
+  FP32Vec8 min(const FP32Vec8& b, int elem_num) const {
+    return FP32Vec8(RVVI(__riscv_vfmin_vv_f32, LMUL_256)(reg, b.reg, elem_num));
+  }
+  FP32Vec8 max(const FP32Vec8& b, int elem_num) const {
+    return FP32Vec8(RVVI(__riscv_vfmax_vv_f32, LMUL_256)(reg, b.reg, elem_num));
+  }
+
+  FP32Vec8 clamp(const FP32Vec8& min_v, const FP32Vec8& max_v) const {
+    fixed_fp32x8_t temp =
+        RVVI(__riscv_vfmax_vv_f32, LMUL_256)(min_v.reg, reg, VEC_ELEM_NUM);
+    return FP32Vec8(
+        RVVI(__riscv_vfmin_vv_f32, LMUL_256)(max_v.reg, temp, VEC_ELEM_NUM));
+  }
+
+  void save(float* ptr) const {
+    RVVI(__riscv_vse32_v_f32, LMUL_256)(ptr, reg, VEC_ELEM_NUM);
+  }
+  void save(float* ptr, int elem_num) const {
+    RVVI(__riscv_vse32_v_f32, LMUL_256)(ptr, reg, elem_num);
+  }
+  void save_strided(float* ptr, ptrdiff_t stride) const {
+    ptrdiff_t byte_stride = stride * sizeof(float);
+    RVVI(__riscv_vsse32_v_f32, LMUL_256)(ptr, byte_stride, reg, VEC_ELEM_NUM);
+  }
+
+  FP32Vec8 exp() const {
+    // Clamp input to prevent NaN: exp(-inf) must return 0, not NaN.
+    // Without clamping, -inf * 0.0 = NaN in the final poly * scale step.
+    // Matches the clamping strategy used by x86 AVX-512 and ARM NEON.
+    constexpr float exp_lo = -87.3365447505f;  // ln(FLT_MIN)
+    constexpr float exp_hi = 88.7228391117f;   // ln(FLT_MAX)
+    fixed_fp32x8_t x = RVVI(__riscv_vfmin_vf_f32, LMUL_256)(
+        RVVI(__riscv_vfmax_vf_f32, LMUL_256)(reg, exp_lo, VEC_ELEM_NUM), exp_hi,
+        VEC_ELEM_NUM);
+
+    const float inv_ln2 = 1.44269504088896341f;
+    fixed_fp32x8_t x_scaled =
+        RVVI(__riscv_vfmul_vf_f32, LMUL_256)(x, inv_ln2, VEC_ELEM_NUM);
+    fixed_i32x8_t n_int =
+        RVVI(__riscv_vfcvt_x_f_v_i32, LMUL_256)(x_scaled, VEC_ELEM_NUM);
+    fixed_fp32x8_t n_float =
+        RVVI(__riscv_vfcvt_f_x_v_f32, LMUL_256)(n_int, VEC_ELEM_NUM);
+
+    fixed_fp32x8_t r =
+        RVVI(__riscv_vfsub_vv_f32, LMUL_256)(x_scaled, n_float, VEC_ELEM_NUM);
+
+    fixed_fp32x8_t poly =
+        RVVI(__riscv_vfmv_v_f_f32, LMUL_256)(0.001333355810164f, VEC_ELEM_NUM);
+    poly = RVVI(__riscv_vfmul_vv_f32, LMUL_256)(poly, r, VEC_ELEM_NUM);
+    poly = RVVI(__riscv_vfadd_vf_f32, LMUL_256)(poly, 0.009618129107628f,
+                                                VEC_ELEM_NUM);
+    poly = RVVI(__riscv_vfmul_vv_f32, LMUL_256)(poly, r, VEC_ELEM_NUM);
+    poly = RVVI(__riscv_vfadd_vf_f32, LMUL_256)(poly, 0.055504108664821f,
+                                                VEC_ELEM_NUM);
+    poly = RVVI(__riscv_vfmul_vv_f32, LMUL_256)(poly, r, VEC_ELEM_NUM);
+    poly = RVVI(__riscv_vfadd_vf_f32, LMUL_256)(poly, 0.240226506959101f,
+                                                VEC_ELEM_NUM);
+    poly = RVVI(__riscv_vfmul_vv_f32, LMUL_256)(poly, r, VEC_ELEM_NUM);
+    poly = RVVI(__riscv_vfadd_vf_f32, LMUL_256)(poly, 0.693147180559945f,
+                                                VEC_ELEM_NUM);
+    poly = RVVI(__riscv_vfmul_vv_f32, LMUL_256)(poly, r, VEC_ELEM_NUM);
+    poly = RVVI(__riscv_vfadd_vf_f32, LMUL_256)(poly, 1.0f, VEC_ELEM_NUM);
+
+    fixed_i32x8_t biased_exp =
+        RVVI(__riscv_vadd_vx_i32, LMUL_256)(n_int, 127, VEC_ELEM_NUM);
+    biased_exp =
+        RVVI(__riscv_vmax_vx_i32, LMUL_256)(biased_exp, 0, VEC_ELEM_NUM);
+    fixed_i32x8_t exponent_bits =
+        RVVI(__riscv_vsll_vx_i32, LMUL_256)(biased_exp, 23, VEC_ELEM_NUM);
+    fixed_fp32x8_t scale = RVVI4(__riscv_vreinterpret_v_i32, LMUL_256, _f32,
+                                 LMUL_256)(exponent_bits);
+
+    return FP32Vec8(
+        RVVI(__riscv_vfmul_vv_f32, LMUL_256)(poly, scale, VEC_ELEM_NUM));
+  }
+
+  FP32Vec8 tanh() const {
+    fixed_fp32x8_t x_clamped = RVVI(__riscv_vfmin_vf_f32, LMUL_256)(
+        RVVI(__riscv_vfmax_vf_f32, LMUL_256)(reg, -9.0f, VEC_ELEM_NUM), 9.0f,
+        VEC_ELEM_NUM);
+    fixed_fp32x8_t x2 =
+        RVVI(__riscv_vfmul_vf_f32, LMUL_256)(x_clamped, 2.0f, VEC_ELEM_NUM);
+    FP32Vec8 exp_val = FP32Vec8(x2).exp();
+    fixed_fp32x8_t num =
+        RVVI(__riscv_vfsub_vf_f32, LMUL_256)(exp_val.reg, 1.0f, VEC_ELEM_NUM);
+    fixed_fp32x8_t den =
+        RVVI(__riscv_vfadd_vf_f32, LMUL_256)(exp_val.reg, 1.0f, VEC_ELEM_NUM);
+    return FP32Vec8(
+        RVVI(__riscv_vfdiv_vv_f32, LMUL_256)(num, den, VEC_ELEM_NUM));
+  }
+
+  FP32Vec8 er() const {
+    const float p = 0.3275911f, a1 = 0.254829592f, a2 = -0.284496736f,
+                a3 = 1.421413741f, a4 = -1.453152027f, a5 = 1.061405429f;
+    fixed_fp32x8_t abs_x =
+        RVVI(__riscv_vfabs_v_f32, LMUL_256)(reg, VEC_ELEM_NUM);
+
+    fixed_fp32x8_t t = RVVI(__riscv_vfadd_vf_f32, LMUL_256)(
+        RVVI(__riscv_vfmul_vf_f32, LMUL_256)(abs_x, p, VEC_ELEM_NUM), 1.0f,
+        VEC_ELEM_NUM);
+    t = RVVI(__riscv_vfrdiv_vf_f32, LMUL_256)(t, 1.0f, VEC_ELEM_NUM);
+
+    fixed_fp32x8_t poly =
+        RVVI(__riscv_vfmv_v_f_f32, LMUL_256)(a5, VEC_ELEM_NUM);
+    poly = RVVI(__riscv_vfadd_vf_f32, LMUL_256)(
+        RVVI(__riscv_vfmul_vv_f32, LMUL_256)(poly, t, VEC_ELEM_NUM), a4,
+        VEC_ELEM_NUM);
+    poly = RVVI(__riscv_vfadd_vf_f32, LMUL_256)(
+        RVVI(__riscv_vfmul_vv_f32, LMUL_256)(poly, t, VEC_ELEM_NUM), a3,
+        VEC_ELEM_NUM);
+    poly = RVVI(__riscv_vfadd_vf_f32, LMUL_256)(
+        RVVI(__riscv_vfmul_vv_f32, LMUL_256)(poly, t, VEC_ELEM_NUM), a2,
+        VEC_ELEM_NUM);
+    poly = RVVI(__riscv_vfadd_vf_f32, LMUL_256)(
+        RVVI(__riscv_vfmul_vv_f32, LMUL_256)(poly, t, VEC_ELEM_NUM), a1,
+        VEC_ELEM_NUM);
+    poly = RVVI(__riscv_vfmul_vv_f32, LMUL_256)(poly, t, VEC_ELEM_NUM);
+
+    fixed_fp32x8_t exp_val = FP32Vec8(RVVI(__riscv_vfneg_v_f32, LMUL_256)(
+                                          RVVI(__riscv_vfmul_vv_f32, LMUL_256)(
+                                              abs_x, abs_x, VEC_ELEM_NUM),
+                                          VEC_ELEM_NUM))
+                                 .exp()
+                                 .reg;
+    fixed_fp32x8_t res = RVVI(__riscv_vfrsub_vf_f32, LMUL_256)(
+        RVVI(__riscv_vfmul_vv_f32, LMUL_256)(poly, exp_val, VEC_ELEM_NUM), 1.0f,
+        VEC_ELEM_NUM);
+
+    rvv_mask_f32x8_t mask = RVVIB(__riscv_vmflt_vf_f32, LMUL_256, BOOL_256)(
+        reg, 0.0f, VEC_ELEM_NUM);
+    return FP32Vec8(
+        RVVI3(__riscv_vfneg_v_f32, LMUL_256, _m)(mask, res, VEC_ELEM_NUM));
+  }
+};
+
+struct FP32Vec16 : public Vec<FP32Vec16> {
+  constexpr static int VEC_ELEM_NUM = 16;
+  fixed_fp32x16_t reg;
+
+  explicit FP32Vec16(float v)
+      : reg(RVVI(__riscv_vfmv_v_f_f32, LMUL_512)(v, VEC_ELEM_NUM)) {};
+  explicit FP32Vec16()
+      : reg(RVVI(__riscv_vfmv_v_f_f32, LMUL_512)(0.0f, VEC_ELEM_NUM)) {};
+  explicit FP32Vec16(const float* ptr)
+      : reg(RVVI(__riscv_vle32_v_f32, LMUL_512)(ptr, VEC_ELEM_NUM)) {};
+  explicit FP32Vec16(fixed_fp32x16_t data) : reg(data) {};
+  explicit FP32Vec16(const FP32Vec8& data)
+      : 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(const FP16Vec16& v);
+
+#ifdef RISCV_BF16_SUPPORT
+  explicit FP32Vec16(fixed_bf16x16_t v)
+      : reg(RVVI(__riscv_vfwcvtbf16_f_f_v_f32, LMUL_512)(v, VEC_ELEM_NUM)) {};
+  explicit FP32Vec16(const BF16Vec16& v)
+      : reg(RVVI(__riscv_vfwcvtbf16_f_f_v_f32, LMUL_512)(v.reg, VEC_ELEM_NUM)) {
+        };
+#else
+  explicit FP32Vec16(const BF16Vec16& v) : reg(v.reg_fp32) {};
+#endif
+
+  FP32Vec16 operator+(const FP32Vec16& b) const {
+    return FP32Vec16(
+        RVVI(__riscv_vfadd_vv_f32, LMUL_512)(reg, b.reg, VEC_ELEM_NUM));
+  }
+  FP32Vec16 operator-(const FP32Vec16& b) const {
+    return FP32Vec16(
+        RVVI(__riscv_vfsub_vv_f32, LMUL_512)(reg, b.reg, VEC_ELEM_NUM));
+  }
+  FP32Vec16 operator*(const FP32Vec16& b) const {
+    return FP32Vec16(
+        RVVI(__riscv_vfmul_vv_f32, LMUL_512)(reg, b.reg, VEC_ELEM_NUM));
+  }
+  FP32Vec16 operator/(const FP32Vec16& b) const {
+    return FP32Vec16(
+        RVVI(__riscv_vfdiv_vv_f32, LMUL_512)(reg, b.reg, VEC_ELEM_NUM));
+  }
+
+  FP32Vec16 fma(const FP32Vec16& a, const FP32Vec16& b) const {
+    return FP32Vec16(
+        RVVI(__riscv_vfmacc_vv_f32, LMUL_512)(reg, a.reg, b.reg, VEC_ELEM_NUM));
+  }
+
+  float reduce_sum() const {
+    rvv_f32_accum_t scalar = __riscv_vfmv_s_f_f32m1(0.0f, 1);
+    scalar = RVVI3(__riscv_vfredusum_vs_f32, LMUL_512, _f32m1)(reg, scalar,
+                                                               VEC_ELEM_NUM);
+    return __riscv_vfmv_f_s_f32m1_f32(scalar);
+  }
+
+  float reduce_max() const {
+    rvv_f32_accum_t scalar =
+        __riscv_vfmv_s_f_f32m1(std::numeric_limits<float>::lowest(), 1);
+    scalar = RVVI3(__riscv_vfredmax_vs_f32, LMUL_512, _f32m1)(reg, scalar,
+                                                              VEC_ELEM_NUM);
+    return __riscv_vfmv_f_s_f32m1_f32(scalar);
+  }
+
+  float reduce_min() const {
+    rvv_f32_accum_t scalar =
+        __riscv_vfmv_s_f_f32m1(std::numeric_limits<float>::max(), 1);
+    scalar = RVVI3(__riscv_vfredmin_vs_f32, LMUL_512, _f32m1)(reg, scalar,
+                                                              VEC_ELEM_NUM);
+    return __riscv_vfmv_f_s_f32m1_f32(scalar);
+  }
+
+  template <int group_size>
+  float reduce_sub_sum(int idx) {
+    static_assert(VEC_ELEM_NUM % group_size == 0);
+    const int start = idx * group_size;
+    auto indices = RVVI(__riscv_vid_v_u32, LMUL_512)(VEC_ELEM_NUM);
+    rvv_mask_f32x16_t mask = RVVI(__riscv_vmand_mm_, BOOL_512)(
+        RVVIB(__riscv_vmsgeu_vx_u32, LMUL_512, BOOL_512)(indices, start,
+                                                         VEC_ELEM_NUM),
+        RVVIB(__riscv_vmsltu_vx_u32, LMUL_512, BOOL_512)(
+            indices, start + group_size, VEC_ELEM_NUM),
+        VEC_ELEM_NUM);
+    rvv_f32_accum_t scalar = __riscv_vfmv_s_f_f32m1(0.0f, 1);
+    scalar = RVVI3(__riscv_vfredusum_vs_f32, LMUL_512, _f32m1_m)(
+        mask, reg, scalar, VEC_ELEM_NUM);
+    return __riscv_vfmv_f_s_f32m1_f32(scalar);
+  };
+
+  FP32Vec16 max(const FP32Vec16& b) const {
+    return FP32Vec16(
+        RVVI(__riscv_vfmax_vv_f32, LMUL_512)(reg, b.reg, VEC_ELEM_NUM));
+  }
+  FP32Vec16 min(const FP32Vec16& b) const {
+    return FP32Vec16(
+        RVVI(__riscv_vfmin_vv_f32, LMUL_512)(reg, b.reg, VEC_ELEM_NUM));
+  }
+  FP32Vec16 abs() const {
+    return FP32Vec16(RVVI(__riscv_vfabs_v_f32, LMUL_512)(reg, VEC_ELEM_NUM));
+  }
+
+  FP32Vec16 clamp(const FP32Vec16& min_v, const FP32Vec16& max_v) const {
+    return FP32Vec16(RVVI(__riscv_vfmin_vv_f32, LMUL_512)(
+        max_v.reg,
+        RVVI(__riscv_vfmax_vv_f32, LMUL_512)(min_v.reg, reg, VEC_ELEM_NUM),
+        VEC_ELEM_NUM));
+  }
+
+  void save(float* ptr) const {
+    RVVI(__riscv_vse32_v_f32, LMUL_512)(ptr, reg, VEC_ELEM_NUM);
+  }
+  void save(float* ptr, int elem_num) const {
+    RVVI(__riscv_vse32_v_f32, LMUL_512)(ptr, reg, elem_num);
+  }
+  void save_strided(float* ptr, ptrdiff_t stride) const {
+    ptrdiff_t byte_stride = stride * sizeof(float);
+    RVVI(__riscv_vsse32_v_f32, LMUL_512)(ptr, byte_stride, reg, VEC_ELEM_NUM);
+  }
+
+  FP32Vec16 exp() const {
+    // Clamp input to prevent NaN: exp(-inf) must return 0, not NaN.
+    // Without clamping, -inf * 0.0 = NaN in the final poly * scale step.
+    // Matches the clamping strategy used by x86 AVX-512 and ARM NEON.
+    constexpr float exp_lo = -87.3365447505f;  // ln(FLT_MIN)
+    constexpr float exp_hi = 88.7228391117f;   // ln(FLT_MAX)
+    fixed_fp32x16_t x = RVVI(__riscv_vfmin_vf_f32, LMUL_512)(
+        RVVI(__riscv_vfmax_vf_f32, LMUL_512)(reg, exp_lo, VEC_ELEM_NUM), exp_hi,
+        VEC_ELEM_NUM);
+
+    const float inv_ln2 = 1.44269504088896341f;
+    fixed_fp32x16_t x_scaled =
+        RVVI(__riscv_vfmul_vf_f32, LMUL_512)(x, inv_ln2, VEC_ELEM_NUM);
+    fixed_i32x16_t n_int =
+        RVVI(__riscv_vfcvt_x_f_v_i32, LMUL_512)(x_scaled, VEC_ELEM_NUM);
+    fixed_fp32x16_t n_float =
+        RVVI(__riscv_vfcvt_f_x_v_f32, LMUL_512)(n_int, VEC_ELEM_NUM);
+    fixed_fp32x16_t r =
+        RVVI(__riscv_vfsub_vv_f32, LMUL_512)(x_scaled, n_float, VEC_ELEM_NUM);
+
+    fixed_fp32x16_t poly =
+        RVVI(__riscv_vfmv_v_f_f32, LMUL_512)(0.001333355810164f, VEC_ELEM_NUM);
+    poly = RVVI(__riscv_vfadd_vf_f32, LMUL_512)(
+        RVVI(__riscv_vfmul_vv_f32, LMUL_512)(poly, r, VEC_ELEM_NUM),
+        0.009618129107628f, VEC_ELEM_NUM);
+    poly = RVVI(__riscv_vfadd_vf_f32, LMUL_512)(
+        RVVI(__riscv_vfmul_vv_f32, LMUL_512)(poly, r, VEC_ELEM_NUM),
+        0.055504108664821f, VEC_ELEM_NUM);
+    poly = RVVI(__riscv_vfadd_vf_f32, LMUL_512)(
+        RVVI(__riscv_vfmul_vv_f32, LMUL_512)(poly, r, VEC_ELEM_NUM),
+        0.240226506959101f, VEC_ELEM_NUM);
+    poly = RVVI(__riscv_vfadd_vf_f32, LMUL_512)(
+        RVVI(__riscv_vfmul_vv_f32, LMUL_512)(poly, r, VEC_ELEM_NUM),
+        0.693147180559945f, VEC_ELEM_NUM);
+    poly = RVVI(__riscv_vfadd_vf_f32, LMUL_512)(
+        RVVI(__riscv_vfmul_vv_f32, LMUL_512)(poly, r, VEC_ELEM_NUM), 1.0f,
+        VEC_ELEM_NUM);
+
+    fixed_i32x16_t biased_exp = RVVI(__riscv_vmax_vx_i32, LMUL_512)(
+        RVVI(__riscv_vadd_vx_i32, LMUL_512)(n_int, 127, VEC_ELEM_NUM), 0,
+        VEC_ELEM_NUM);
+    fixed_fp32x16_t scale =
+        RVVI4(__riscv_vreinterpret_v_i32, LMUL_512, _f32, LMUL_512)(
+            RVVI(__riscv_vsll_vx_i32, LMUL_512)(biased_exp, 23, VEC_ELEM_NUM));
+
+    return FP32Vec16(
+        RVVI(__riscv_vfmul_vv_f32, LMUL_512)(poly, scale, VEC_ELEM_NUM));
+  }
+
+  FP32Vec16 tanh() const {
+    fixed_fp32x16_t x_clamped = RVVI(__riscv_vfmin_vf_f32, LMUL_512)(
+        RVVI(__riscv_vfmax_vf_f32, LMUL_512)(reg, -9.0f, VEC_ELEM_NUM), 9.0f,
+        VEC_ELEM_NUM);
+    FP32Vec16 exp_val = FP32Vec16(RVVI(__riscv_vfmul_vf_f32, LMUL_512)(
+                                      x_clamped, 2.0f, VEC_ELEM_NUM))
+                            .exp();
+    return FP32Vec16(RVVI(__riscv_vfdiv_vv_f32, LMUL_512)(
+        RVVI(__riscv_vfsub_vf_f32, LMUL_512)(exp_val.reg, 1.0f, VEC_ELEM_NUM),
+        RVVI(__riscv_vfadd_vf_f32, LMUL_512)(exp_val.reg, 1.0f, VEC_ELEM_NUM),
+        VEC_ELEM_NUM));
+  }
+
+  FP32Vec16 er() const {
+    const float p = 0.3275911f, a1 = 0.254829592f, a2 = -0.284496736f,
+                a3 = 1.421413741f, a4 = -1.453152027f, a5 = 1.061405429f;
+    fixed_fp32x16_t abs_x =
+        RVVI(__riscv_vfabs_v_f32, LMUL_512)(reg, VEC_ELEM_NUM);
+    fixed_fp32x16_t t = RVVI(__riscv_vfrdiv_vf_f32, LMUL_512)(
+        RVVI(__riscv_vfadd_vf_f32, LMUL_512)(
+            RVVI(__riscv_vfmul_vf_f32, LMUL_512)(abs_x, p, VEC_ELEM_NUM), 1.0f,
+            VEC_ELEM_NUM),
+        1.0f, VEC_ELEM_NUM);
+
+    fixed_fp32x16_t poly =
+        RVVI(__riscv_vfmv_v_f_f32, LMUL_512)(a5, VEC_ELEM_NUM);
+    poly = RVVI(__riscv_vfadd_vf_f32, LMUL_512)(
+        RVVI(__riscv_vfmul_vv_f32, LMUL_512)(poly, t, VEC_ELEM_NUM), a4,
+        VEC_ELEM_NUM);
+    poly = RVVI(__riscv_vfadd_vf_f32, LMUL_512)(
+        RVVI(__riscv_vfmul_vv_f32, LMUL_512)(poly, t, VEC_ELEM_NUM), a3,
+        VEC_ELEM_NUM);
+    poly = RVVI(__riscv_vfadd_vf_f32, LMUL_512)(
+        RVVI(__riscv_vfmul_vv_f32, LMUL_512)(poly, t, VEC_ELEM_NUM), a2,
+        VEC_ELEM_NUM);
+    poly = RVVI(__riscv_vfadd_vf_f32, LMUL_512)(
+        RVVI(__riscv_vfmul_vv_f32, LMUL_512)(poly, t, VEC_ELEM_NUM), a1,
+        VEC_ELEM_NUM);
+    poly = RVVI(__riscv_vfmul_vv_f32, LMUL_512)(poly, t, VEC_ELEM_NUM);
+
+    fixed_fp32x16_t exp_val =
+        FP32Vec16(RVVI(__riscv_vfneg_v_f32, LMUL_512)(
+                      RVVI(__riscv_vfmul_vv_f32, LMUL_512)(abs_x, abs_x,
+                                                           VEC_ELEM_NUM),
+                      VEC_ELEM_NUM))
+            .exp()
+            .reg;
+    fixed_fp32x16_t res = RVVI(__riscv_vfrsub_vf_f32, LMUL_512)(
+        RVVI(__riscv_vfmul_vv_f32, LMUL_512)(poly, exp_val, VEC_ELEM_NUM), 1.0f,
+        VEC_ELEM_NUM);
+
+    rvv_mask_f32x16_t mask = RVVIB(__riscv_vmflt_vf_f32, LMUL_512, BOOL_512)(
+        reg, 0.0f, VEC_ELEM_NUM);
+    return FP32Vec16(
+        RVVI3(__riscv_vfneg_v_f32, LMUL_512, _m)(mask, res, VEC_ELEM_NUM));
+  }
+};
+
+// ============================================================================
+// Type Traits & Global Helpers
+// ============================================================================
+
+template <typename T>
+struct VecType {
+  using vec_type = void;
+  using vec_t = void;
+};
+
+template <typename T>
+using vec_t = typename VecType<T>::vec_type;
+
+template <>
+struct VecType<float> {
+  using vec_type = FP32Vec8;
+  using vec_t = FP32Vec8;
+};
+template <>
+struct VecType<c10::Half> {
+  using vec_type = FP16Vec8;
+  using vec_t = FP16Vec8;
+};
+template <>
+struct VecType<c10::BFloat16> {
+  using vec_type = BF16Vec8;
+  using vec_t = BF16Vec8;
+};
+
+template <typename T>
+void storeFP32(float v, T* ptr) {
+  *ptr = v;
+}
+template <>
+inline void storeFP32<c10::Half>(float v, c10::Half* ptr) {
+  *reinterpret_cast<_Float16*>(ptr) = static_cast<_Float16>(v);
+}
+
+inline FP16Vec16::FP16Vec16(const FP32Vec16& v) {
+  reg = RVVI(__riscv_vfncvt_f_f_w_f16, LMUL_256)(v.reg, VEC_ELEM_NUM);
+}
+inline FP16Vec8::FP16Vec8(const FP32Vec8& v) {
+  reg = RVVI(__riscv_vfncvt_f_f_w_f16, LMUL_128)(v.reg, VEC_ELEM_NUM);
+}
+inline FP32Vec16::FP32Vec16(const FP16Vec16& v) {
+  reg = RVVI(__riscv_vfwcvt_f_f_v_f32, LMUL_512)(v.reg, VEC_ELEM_NUM);
+}
+inline void fma(FP32Vec16& acc, const FP32Vec16& a, const FP32Vec16& b) {
+  acc = acc.fma(a, b);
+}
+
+#ifdef RISCV_BF16_SUPPORT
+template <>
+inline void storeFP32<c10::BFloat16>(float v, c10::BFloat16* ptr) {
+  *ptr = static_cast<__bf16>(v);
+};
+inline BF16Vec8::BF16Vec8(const FP32Vec8& v)
+    : reg(RVVI(__riscv_vfncvtbf16_f_f_w_bf16, LMUL_128)(v.reg, VEC_ELEM_NUM)) {
+      };
+inline BF16Vec16::BF16Vec16(const FP32Vec16& v)
+    : reg(RVVI(__riscv_vfncvtbf16_f_f_w_bf16, LMUL_256)(v.reg, VEC_ELEM_NUM)) {
+      };
+#else
+template <>
+inline void storeFP32<c10::BFloat16>(float v, c10::BFloat16* ptr) {
+  uint32_t val;
+  std::memcpy(&val, &v, 4);
+  *reinterpret_cast<uint16_t*>(ptr) = static_cast<uint16_t>(val >> 16);
+}
+inline BF16Vec8::BF16Vec8(const FP32Vec8& v) : reg_fp32(v.reg) {}
+inline BF16Vec16::BF16Vec16(const FP32Vec16& v) : reg_fp32(v.reg) {}
+#endif
+
+inline void prefetch(const void* addr) { __builtin_prefetch(addr, 0, 1); }
+
+}  // namespace vec_op
+
+#ifndef CPU_KERNEL_GUARD_IN
+  #define CPU_KERNEL_GUARD_IN(NAME)
+#endif
+
+#ifndef CPU_KERNEL_GUARD_OUT
+  #define CPU_KERNEL_GUARD_OUT(NAME)
+#endif
+
+#endif  // CPU_TYPES_RISCV_IMPL_HPP

csrc/cpu/cpu_types_vxe.hpp

@@ -8,6 +8,9 @@
 #include <torch/all.h>
 namespace vec_op {
 
+struct fp8_e4m3_tag {};
+struct fp8_e5m2_tag {};
+
 #define vec_neg(a) (-(a))
 #define vec_add(a, b) ((a) + (b))
 #define vec_sub(a, b) ((a) - (b))
@@ -241,6 +244,9 @@ struct BF16Vec32 : public Vec<BF16Vec32> {
   explicit BF16Vec32(const BF16Vec8& vec8_data)
       : reg({vec8_data.reg, vec8_data.reg, vec8_data.reg, vec8_data.reg}) {}
 
+  explicit BF16Vec32(const uint8_t*, fp8_e4m3_tag) : reg{} {}
+  explicit BF16Vec32(const uint8_t*, fp8_e5m2_tag) : reg{} {}
+
   void save(void* ptr) const { *reinterpret_cast<ss16x8x4_t*>(ptr) = reg; }
 };
 

csrc/cpu/cpu_types_x86.hpp

@@ -11,6 +11,17 @@ static_assert(false, "AVX2 must be supported for the current implementation.");
 
 namespace vec_op {
 
+// Tags for FP8 BF16Vec32 constructors (avoid overload collision with
+// BF16Vec32(void*)).
+// VEC path (FP8 → pseudo-FP16 layout, scale correction applied later):
+struct fp8_e4m3_tag {};  // E4M3 → pseudo-FP16; BF16 value = true_E4M3 * 2^-8
+struct fp8_e5m2_tag {};  // E5M2 → FP16 bits directly (same exponent bias=15)
+// AMX path (FP8 → unscaled BF16, no FP32 round-trip):
+// BF16 value = true_E4M3 * 2^-120 (E4M3) or true_E5M2 * 2^-112 (E5M2).
+// Exponent rebiasing is folded into k/v scales by the caller.
+struct fp8_bf16_e4m3_tag {};
+struct fp8_bf16_e5m2_tag {};
+
 #define VLLM_DISPATCH_CASE_FLOATING_TYPES(...)            \
   AT_DISPATCH_CASE(at::ScalarType::Float, __VA_ARGS__)    \
   AT_DISPATCH_CASE(at::ScalarType::BFloat16, __VA_ARGS__) \
@@ -176,6 +187,50 @@ struct BF16Vec32 : public Vec<BF16Vec32> {
                                (__m128i)vec8_data.reg, 2),
             (__m128i)vec8_data.reg, 3)) {}
 
+  // Decode 32 FP8-E4M3 bytes to pseudo-FP16 layout (stored in the BF16
+  // register).  Result = true_E4M3 * 2^-8; caller applies scale * 2^8.
+  explicit BF16Vec32(const uint8_t* ptr, fp8_e4m3_tag) {
+    __m256i b8 = _mm256_loadu_si256(reinterpret_cast<const __m256i*>(ptr));
+    __m512i b16 = _mm512_cvtepu8_epi16(b8);
+    __m512i sign =
+        _mm512_slli_epi16(_mm512_and_si512(b16, _mm512_set1_epi16(0x80)), 8);
+    __m512i payload =
+        _mm512_slli_epi16(_mm512_and_si512(b16, _mm512_set1_epi16(0x7F)), 7);
+    reg = _mm512_or_si512(sign, payload);
+  }
+
+  // Decode 32 FP8-E5M2 bytes to FP16 layout.
+  // E5M2 and FP16 share the same 5-bit exponent bias (15), so FP8 byte b maps
+  // directly to FP16 bits by shifting left 8 — no sign/payload reconstruction.
+  explicit BF16Vec32(const uint8_t* ptr, fp8_e5m2_tag) {
+    __m256i b8 = _mm256_loadu_si256(reinterpret_cast<const __m256i*>(ptr));
+    reg = _mm512_slli_epi16(_mm512_cvtepu8_epi16(b8), 8);
+  }
+
+  // Direct FP8-E4M3 → unscaled BF16 for AMX (no FP32 round-trip).
+  // BF16 value = true_E4M3 * 2^-120; exponent rebiasing folded into k/v scales.
+  explicit BF16Vec32(const uint8_t* ptr, fp8_bf16_e4m3_tag) {
+    __m256i b8 = _mm256_loadu_si256(reinterpret_cast<const __m256i*>(ptr));
+    __m512i b16 = _mm512_cvtepu8_epi16(b8);
+    __m512i sign =
+        _mm512_slli_epi16(_mm512_and_si512(b16, _mm512_set1_epi16(0x80)), 8);
+    __m512i payload =
+        _mm512_slli_epi16(_mm512_and_si512(b16, _mm512_set1_epi16(0x7F)), 4);
+    reg = _mm512_or_si512(sign, payload);
+  }
+
+  // Direct FP8-E5M2 → unscaled BF16 for AMX (no FP32 round-trip).
+  // BF16 value = true_E5M2 * 2^-112; exponent rebiasing folded into k/v scales.
+  explicit BF16Vec32(const uint8_t* ptr, fp8_bf16_e5m2_tag) {
+    __m256i b8 = _mm256_loadu_si256(reinterpret_cast<const __m256i*>(ptr));
+    __m512i b16 = _mm512_cvtepu8_epi16(b8);
+    __m512i sign =
+        _mm512_slli_epi16(_mm512_and_si512(b16, _mm512_set1_epi16(0x80)), 8);
+    __m512i payload =
+        _mm512_slli_epi16(_mm512_and_si512(b16, _mm512_set1_epi16(0x7F)), 5);
+    reg = _mm512_or_si512(sign, payload);
+  }
+
   void save(void* ptr) const { *reinterpret_cast<__m512i*>(ptr) = reg; }
 };
 #else
@@ -200,6 +255,77 @@ struct BF16Vec32 : public Vec<BF16Vec32> {
             _mm256_castsi128_si256((__m128i)vec8_data.reg),
             (__m128i)vec8_data.reg, 1)) {}
 
+  // E4M3 decode (AVX2 path) — same bit-layout trick as the AVX512 variant
+  // above.  Result = true_E4M3 * 2^-8; caller applies scale * 2^8.
+  explicit BF16Vec32(const uint8_t* ptr, fp8_e4m3_tag) {
+    __m256i b8 = _mm256_loadu_si256(reinterpret_cast<const __m256i*>(ptr));
+    __m128i b8_low = _mm256_extracti128_si256(b8, 0);
+    __m128i b8_high = _mm256_extracti128_si256(b8, 1);
+    __m256i b16_low = _mm256_cvtepu8_epi16(b8_low);
+    __m256i b16_high = _mm256_cvtepu8_epi16(b8_high);
+
+    __m256i sign_low = _mm256_slli_epi16(
+        _mm256_and_si256(b16_low, _mm256_set1_epi16(0x80)), 8);
+    __m256i payload_low = _mm256_slli_epi16(
+        _mm256_and_si256(b16_low, _mm256_set1_epi16(0x7F)), 7);
+    __m256i sign_high = _mm256_slli_epi16(
+        _mm256_and_si256(b16_high, _mm256_set1_epi16(0x80)), 8);
+    __m256i payload_high = _mm256_slli_epi16(
+        _mm256_and_si256(b16_high, _mm256_set1_epi16(0x7F)), 7);
+    reg_low = _mm256_or_si256(sign_low, payload_low);
+    reg_high = _mm256_or_si256(sign_high, payload_high);
+  }
+
+  // E5M2 decode (AVX2 path) — b << 8 maps to FP16 bits; see AVX512 variant
+  // above.
+  explicit BF16Vec32(const uint8_t* ptr, fp8_e5m2_tag) {
+    __m256i b8 = _mm256_loadu_si256(reinterpret_cast<const __m256i*>(ptr));
+    __m128i b8_low = _mm256_extracti128_si256(b8, 0);
+    __m128i b8_high = _mm256_extracti128_si256(b8, 1);
+    reg_low = _mm256_slli_epi16(_mm256_cvtepu8_epi16(b8_low), 8);
+    reg_high = _mm256_slli_epi16(_mm256_cvtepu8_epi16(b8_high), 8);
+  }
+
+  // Direct FP8-E4M3 → unscaled BF16 for AMX (AVX2 path, no FP32 round-trip).
+  // BF16 value = true_E4M3 * 2^-120; exponent rebiasing folded into k/v scales.
+  explicit BF16Vec32(const uint8_t* ptr, fp8_bf16_e4m3_tag) {
+    __m256i b8 = _mm256_loadu_si256(reinterpret_cast<const __m256i*>(ptr));
+    __m128i b8_low = _mm256_extracti128_si256(b8, 0);
+    __m128i b8_high = _mm256_extracti128_si256(b8, 1);
+    __m256i b16_low = _mm256_cvtepu8_epi16(b8_low);
+    __m256i b16_high = _mm256_cvtepu8_epi16(b8_high);
+    reg_low = _mm256_or_si256(
+        _mm256_slli_epi16(_mm256_and_si256(b16_low, _mm256_set1_epi16(0x80)),
+                          8),
+        _mm256_slli_epi16(_mm256_and_si256(b16_low, _mm256_set1_epi16(0x7F)),
+                          4));
+    reg_high = _mm256_or_si256(
+        _mm256_slli_epi16(_mm256_and_si256(b16_high, _mm256_set1_epi16(0x80)),
+                          8),
+        _mm256_slli_epi16(_mm256_and_si256(b16_high, _mm256_set1_epi16(0x7F)),
+                          4));
+  }
+
+  // Direct FP8-E5M2 → unscaled BF16 for AMX (AVX2 path, no FP32 round-trip).
+  // BF16 value = true_E5M2 * 2^-112; exponent rebiasing folded into k/v scales.
+  explicit BF16Vec32(const uint8_t* ptr, fp8_bf16_e5m2_tag) {
+    __m256i b8 = _mm256_loadu_si256(reinterpret_cast<const __m256i*>(ptr));
+    __m128i b8_low = _mm256_extracti128_si256(b8, 0);
+    __m128i b8_high = _mm256_extracti128_si256(b8, 1);
+    __m256i b16_low = _mm256_cvtepu8_epi16(b8_low);
+    __m256i b16_high = _mm256_cvtepu8_epi16(b8_high);
+    reg_low = _mm256_or_si256(
+        _mm256_slli_epi16(_mm256_and_si256(b16_low, _mm256_set1_epi16(0x80)),
+                          8),
+        _mm256_slli_epi16(_mm256_and_si256(b16_low, _mm256_set1_epi16(0x7F)),
+                          5));
+    reg_high = _mm256_or_si256(
+        _mm256_slli_epi16(_mm256_and_si256(b16_high, _mm256_set1_epi16(0x80)),
+                          8),
+        _mm256_slli_epi16(_mm256_and_si256(b16_high, _mm256_set1_epi16(0x7F)),
+                          5));
+  }
+
   void save(void* ptr) const {
     _mm256_storeu_si256((__m256i*)ptr, reg_low);
     _mm256_storeu_si256((__m256i*)ptr + 1, reg_high);
@@ -390,6 +516,11 @@ struct FP32Vec16 : public Vec<FP32Vec16> {
       : reg(_mm512_castsi512_ps(
             _mm512_bslli_epi128(_mm512_cvtepu16_epi32(v.reg), 2))) {}
 
+  explicit FP32Vec16(const BF16Vec32& v, int upper) {
+    __m256i v_half_i = _mm512_extracti32x8_epi32(v.reg, upper);
+    reg = _mm512_cvtph_ps(v_half_i);
+  }
+
   explicit FP32Vec16(const FP16Vec16& v) : reg(_mm512_cvtph_ps(v.reg)) {}
 
   explicit FP32Vec16(const FP16Vec8& v) : FP32Vec16(FP32Vec8(v)) {}
@@ -494,6 +625,14 @@ struct FP32Vec16 : public Vec<FP32Vec16> {
   explicit FP32Vec16(const FP32Vec8& data)
       : reg_low(data.reg), reg_high(data.reg) {}
 
+  explicit FP32Vec16(const BF16Vec32& v, int upper) {
+    const __m256i& half = upper ? v.reg_high : v.reg_low;
+    __m128i lo = _mm256_extractf128_si256(half, 0);
+    __m128i hi = _mm256_extractf128_si256(half, 1);
+    reg_low = _mm256_cvtph_ps(lo);
+    reg_high = _mm256_cvtph_ps(hi);
+  }
+
   explicit FP32Vec16(const FP16Vec16& v) {
     __m128i low = _mm256_extractf128_si256(v.reg, 0);
     __m128i high = _mm256_extractf128_si256(v.reg, 1);

csrc/cpu/generate_cpu_attn_dispatch.py

@@ -22,71 +22,95 @@ ISA_TYPES = {
     "VXE": 4,
 }
 
+# KV cache index: 0 = auto (same as scalar_t), 1 = fp8_e4m3, 2 = fp8_e5m2
+KV_CACHE_IDX = {
+    "auto": 0,
+    "fp8_e4m3": 1,
+    "fp8_e5m2": 2,
+}
+
+# C++ type for each kv_cache index
+KV_CACHE_CPP_TYPES = {
+    "auto": "scalar_t",
+    "fp8_e4m3": "c10::Float8_e4m3fn",
+    "fp8_e5m2": "c10::Float8_e5m2",
+}
+
 # ISAs supported for head_dims divisible by 32
 ISA_FOR_32 = ["AMX", "NEON", "VEC", "VEC16", "VXE"]
 
 # ISAs supported for head_dims divisible by 16 only
 ISA_FOR_16 = ["VEC16"]
 
+# ISAs that support FP8 KV cache (x86 AVX2/AVX-512 required)
+ISA_FOR_FP8 = ["AMX", "VEC"]
 
-def encode_params(head_dim: int, isa_type: str) -> int:
-    """Encode head_dim and ISA type into a single int64_t."""
+
+def encode_params(head_dim: int, isa_type: str, kv_cache: str = "auto") -> int:
+    """Encode head_dim, ISA type, and KV cache type into a single int64_t."""
     isa_val = ISA_TYPES[isa_type]
-    # Encoding: (head_dim << 8) | isa_type
-    # This allows head_dim up to 2^56 - 1 and 256 ISA types
-    return (head_dim << 8) | isa_val
+    kv_val = KV_CACHE_IDX[kv_cache]
+    # Encoding: (head_dim << 16) | (kv_cache_idx << 8) | isa_type
+    # This allows head_dim up to 2^48 - 1, 256 KV cache types, and 256 ISA types
+    return (head_dim << 16) | (kv_val << 8) | isa_val
 
 
-def generate_cases_for_isa_group(isa_list: list[str]) -> str:
+def _make_case(
+    head_dim: int, isa: str, kv_cache: str = "auto", isa_override: str | None = None
+) -> str:
+    """Generate a single switch case line."""
+    encoded = encode_params(head_dim, isa, kv_cache)
+    actual_isa = isa_override if isa_override else isa
+    cpp_type = KV_CACHE_CPP_TYPES[kv_cache]
+    attn_impl = (
+        f"cpu_attention::AttentionImpl<"
+        f"cpu_attention::ISA::{actual_isa}, \\\n"
+        f"                                                       "
+        f"scalar_t, head_dim, {cpp_type}>"
+    )
+    comment = (
+        f"head_dim={head_dim}, isa={isa}"
+        if kv_cache == "auto"
+        else f"head_dim={head_dim}, isa={isa}, kv_cache={kv_cache}"
+    )
+    return (
+        f"""      case {encoded}LL: {{ """
+        f"""/* {comment} */ \\"""
+        f"""
+        constexpr size_t head_dim = {head_dim}; \\"""
+        f"""
+        using attn_impl = {attn_impl}; \\"""
+        f"""
+        return __VA_ARGS__(); \\"""
+        f"""
+      }} \\"""
+    )
+
+
+def generate_cases_for_isa_group(isa_list: list[str], include_fp8: bool = False) -> str:
     """Generate switch cases for a specific ISA group."""
     cases = []
 
-    # Generate cases for head_dims divisible by 32
+    # Non-FP8 cases for head_dims divisible by 32
     for head_dim in HEAD_DIMS_32:
         for isa in isa_list:
             if isa not in ISA_FOR_32:
                 continue
-            encoded = encode_params(head_dim, isa)
-            case_str = (
-                f"""      case {encoded}LL: {{ """
-                f"""/* head_dim={head_dim}, isa={isa} */ \\"""
-                f"""
-        constexpr size_t head_dim = {head_dim}; \\"""
-                f"""
-        using attn_impl = cpu_attention::AttentionImpl<"""
-                f"""cpu_attention::ISA::{isa}, \\"""
-                f"""
-                                                       """
-                f"""scalar_t, head_dim>; \\"""
-                f"""
-        return __VA_ARGS__(); \\"""
-                f"""
-      }} \\"""
-            )
-            cases.append(case_str)
+            cases.append(_make_case(head_dim, isa, "auto"))
 
-    # Generate cases for head_dims divisible by 16 only
+    # Non-FP8 cases for head_dims divisible by 16 only
     for head_dim in HEAD_DIMS_16:
         for isa in isa_list:
-            encoded = encode_params(head_dim, isa)
-            case_str = (
-                f"""      case {encoded}LL: {{ """
-                f"""/* head_dim={head_dim}, isa={isa} """
-                f"""(using VEC16) */ \\"""
-                f"""
-        constexpr size_t head_dim = {head_dim}; \\"""
-                f"""
-        using attn_impl = cpu_attention::AttentionImpl<"""
-                f"""cpu_attention::ISA::VEC16, \\"""
-                f"""
-                                                       """
-                f"""scalar_t, head_dim>; \\"""
-                f"""
-        return __VA_ARGS__(); \\"""
-                f"""
-      }} \\"""
-            )
-            cases.append(case_str)
+            cases.append(_make_case(head_dim, isa, "auto", isa_override="VEC16"))
+
+    # FP8 cases: only AMX and VEC, only head_dims divisible by 32
+    if include_fp8:
+        for fp8_type in ("fp8_e4m3", "fp8_e5m2"):
+            for head_dim in HEAD_DIMS_32:
+                for isa in isa_list:
+                    if isa not in ISA_FOR_FP8:
+                        continue
+                    cases.append(_make_case(head_dim, isa, fp8_type))
 
     return "\n".join(cases)
 
@@ -94,8 +118,9 @@ def generate_cases_for_isa_group(isa_list: list[str]) -> str:
 def generate_helper_function() -> str:
     """Generate helper function to encode parameters."""
     return """
-inline int64_t encode_cpu_attn_params(int64_t head_dim, cpu_attention::ISA isa) {
-  return (head_dim << 8) | static_cast<int64_t>(isa);
+inline int64_t encode_cpu_attn_params(int64_t head_dim, cpu_attention::ISA isa,
+                                      int64_t kv_cache_idx = 0) {
+  return (head_dim << 16) | (kv_cache_idx << 8) | static_cast<int64_t>(isa);
 }
 """
 
@@ -129,87 +154,78 @@ def generate_header_file() -> str:
 
     # Generate dispatch macro with conditional compilation for different ISA sets
     header += """
-// Dispatch macro using encoded parameters
+// Dispatch macro using encoded parameters.
+// KV_CACHE_IDX: Fp8KVCacheDataType enum value (kAuto=0, kFp8E4M3=1, kFp8E5M2=2).
+// FP8 cases (kv_cache_idx != 0) are generated on x86 platforms with AVX2 or
+// AVX-512: BF16Vec32 FP8 constructors have both AVX-512 and AVX2 implementations
+// in cpu_types_x86.hpp. Non-x86 platforms (#else fallback) have fp8=False.
 """
 
-    # x86_64 with AMX
-    header += """#if defined(CPU_CAPABILITY_AMXBF16)
-#define CPU_ATTN_DISPATCH(HEAD_DIM, ISA_TYPE, ...) \\
-  [&] { \\
-    int64_t encoded_params = encode_cpu_attn_params(HEAD_DIM, ISA_TYPE); \\
-    switch (encoded_params) { \\
-"""
-    header += generate_cases_for_isa_group(["AMX", "VEC", "VEC16"])
-    header += """
-      default: { \\
-        TORCH_CHECK(false, "Unsupported CPU attention configuration: head_dim=" + \\
-                    std::to_string(HEAD_DIM) + " isa=" + \\
-                    std::to_string(static_cast<int>(ISA_TYPE))); \\
-      } \\
-    } \\
-  }()
+    def _macro_block(guard: str, isa_list: list[str], fp8: bool) -> str:
+        """Return one CPU_ATTN_DISPATCH macro block for a given guard."""
+        enc = (
+            "    int64_t encoded_params = encode_cpu_attn_params("
+            "HEAD_DIM, ISA_TYPE, KV_CACHE_IDX); \\"
+        )
+        cases = generate_cases_for_isa_group(isa_list, include_fp8=fp8)
+        tail = (
+            "\n"
+            "      default: { \\\n"
+            "        TORCH_CHECK(false, "
+            '"Unsupported CPU attention configuration: head_dim=" + \\\n'
+            '                    std::to_string(HEAD_DIM) + " isa=" + \\\n'
+            "                    std::to_string(static_cast<int>(ISA_TYPE))"
+            " + \\\n"
+            '                    " kv_cache_idx=" + '
+            "std::to_string(KV_CACHE_IDX)); \\\n"
+            "      } \\\n"
+            "    } \\\n"
+            "  }()\n\n"
+        )
+        return (
+            f"{guard}\n"
+            "#define CPU_ATTN_DISPATCH(HEAD_DIM, ISA_TYPE, KV_CACHE_IDX, ...) \\\n"
+            "  [&] { \\\n"
+            f"{enc}\n"
+            "    switch (encoded_params) { \\\n"
+            f"{cases}"
+            f"{tail}"
+        )
 
-"""
-
-    # ARM64 with NEON
-    header += """#elif defined(__aarch64__)
-#define CPU_ATTN_DISPATCH(HEAD_DIM, ISA_TYPE, ...) \\
-  [&] { \\
-    int64_t encoded_params = encode_cpu_attn_params(HEAD_DIM, ISA_TYPE); \\
-    switch (encoded_params) { \\
-"""
-    header += generate_cases_for_isa_group(["NEON", "VEC", "VEC16"])
-    header += """
-      default: { \\
-        TORCH_CHECK(false, "Unsupported CPU attention configuration: head_dim=" + \\
-                    std::to_string(HEAD_DIM) + " isa=" + \\
-                    std::to_string(static_cast<int>(ISA_TYPE))); \\
-      } \\
-    } \\
-  }()
-
-"""
-
-    # s390x with VXE
-    header += """#elif defined(__s390x__)
-#define CPU_ATTN_DISPATCH(HEAD_DIM, ISA_TYPE, ...) \\
-  [&] { \\
-    int64_t encoded_params = encode_cpu_attn_params(HEAD_DIM, ISA_TYPE); \\
-    switch (encoded_params) { \\
-"""
-    header += generate_cases_for_isa_group(["VXE", "VEC", "VEC16"])
-    header += """
-      default: { \\
-        TORCH_CHECK(false, "Unsupported CPU attention configuration: head_dim=" + \\
-                    std::to_string(HEAD_DIM) + " isa=" + \\
-                    std::to_string(static_cast<int>(ISA_TYPE))); \\
-      } \\
-    } \\
-  }()
-
-"""
-
-    # Fallback: VEC and VEC16 only
-    header += """#else
-#define CPU_ATTN_DISPATCH(HEAD_DIM, ISA_TYPE, ...) \\
-  [&] { \\
-    int64_t encoded_params = encode_cpu_attn_params(HEAD_DIM, ISA_TYPE); \\
-    switch (encoded_params) { \\
-"""
-    header += generate_cases_for_isa_group(["VEC", "VEC16"])
-    header += """
-      default: { \\
-        TORCH_CHECK(false, "Unsupported CPU attention configuration: head_dim=" + \\
-                    std::to_string(HEAD_DIM) + " isa=" + \\
-                    std::to_string(static_cast<int>(ISA_TYPE))); \\
-      } \\
-    } \\
-  }()
-
-#endif  /* CPU_CAPABILITY_AMXBF16 / __aarch64__ / __s390x__ */
-
-#endif  // CPU_ATTN_DISPATCH_GENERATED_H
-"""
+    header += _macro_block(
+        "#if defined(CPU_CAPABILITY_AMXBF16)",
+        ["AMX", "VEC", "VEC16"],
+        fp8=True,
+    )
+    header += _macro_block(
+        "#elif defined(__aarch64__)",
+        ["NEON", "VEC", "VEC16"],
+        fp8=False,
+    )
+    header += _macro_block(
+        "#elif defined(__s390x__)",
+        ["VXE", "VEC", "VEC16"],
+        fp8=False,
+    )
+    header += _macro_block(
+        "#elif defined(__AVX512F__)",
+        ["VEC", "VEC16"],
+        fp8=True,
+    )
+    header += _macro_block(
+        "#elif defined(__AVX2__)",
+        ["VEC", "VEC16"],
+        fp8=False,
+    )
+    header += _macro_block(
+        "#else",
+        ["VEC", "VEC16"],
+        fp8=False,
+    )
+    header += (
+        "#endif  /* CPU_CAPABILITY_AMXBF16 / __aarch64__ / __s390x__ */\n\n"
+        "#endif  // CPU_ATTN_DISPATCH_GENERATED_H\n"
+    )
 
     return header
 

csrc/cpu/pos_encoding.cpp

@@ -178,7 +178,12 @@ void rotary_embedding_gptj_impl(
 
 void rotary_embedding(torch::Tensor& positions, torch::Tensor& query,
                       std::optional<torch::Tensor> key, int64_t head_size,
-                      torch::Tensor& cos_sin_cache, bool is_neox) {
+                      torch::Tensor& cos_sin_cache, bool is_neox,
+                      int64_t rope_dim_offset, bool inverse) {
+  TORCH_CHECK(rope_dim_offset == 0,
+              "rope_dim_offset != 0 is not supported on CPU");
+  TORCH_CHECK(!inverse, "inverse rotary embedding is not supported on CPU");
+
   int num_tokens = positions.numel();
   int rot_dim = cos_sin_cache.size(1);
   int num_heads = query.size(-1) / head_size;

csrc/cpu/torch_bindings.cpp

@@ -101,7 +101,9 @@ void cpu_attn_reshape_and_cache(const torch::Tensor& key,
                                 torch::Tensor& key_cache,
                                 torch::Tensor& value_cache,
                                 const torch::Tensor& slot_mapping,
-                                const std::string& isa);
+                                const std::string& isa, const double k_scale,
+                                const double v_scale,
+                                const std::string& kv_cache_dtype);
 
 void cpu_attention_with_kv_cache(
     const torch::Tensor& query, const torch::Tensor& key_cache,
@@ -112,7 +114,8 @@ void cpu_attention_with_kv_cache(
     const int64_t sliding_window_left, const int64_t sliding_window_right,
     const torch::Tensor& block_table, const double softcap,
     const torch::Tensor& scheduler_metadata,
-    const std::optional<torch::Tensor>& s_aux);
+    const std::optional<torch::Tensor>& s_aux, const double k_scale,
+    const double v_scale, const std::string& kv_cache_dtype);
 
 // Note: just for avoiding importing errors
 void placeholder_op() { TORCH_CHECK(false, "Unimplemented"); }
@@ -263,7 +266,8 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
   ops.def(
       "rotary_embedding(Tensor positions, Tensor! query,"
       "                 Tensor!? key, int head_size,"
-      "                 Tensor cos_sin_cache, bool is_neox) -> ()");
+      "                 Tensor cos_sin_cache, bool is_neox, int "
+      "rope_dim_offset=0, bool inverse=False) -> ()");
   ops.impl("rotary_embedding", torch::kCPU, &rotary_embedding);
 
   // Quantization
@@ -383,15 +387,18 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
       &get_scheduler_metadata);
   ops.def(
       "cpu_attn_reshape_and_cache(Tensor key, Tensor value, Tensor(a2!) "
-      "key_cache, Tensor(a3!) value_cache, Tensor slot_mapping, str "
-      "isa) -> ()",
+      "key_cache, Tensor(a3!) value_cache, Tensor slot_mapping, str isa, "
+      "float k_scale=1.0, float v_scale=1.0, str kv_cache_dtype=\"auto\") -> "
+      "()",
       &cpu_attn_reshape_and_cache);
   ops.def(
       "cpu_attention_with_kv_cache(Tensor query, Tensor key_cache, Tensor "
       "value_cache, Tensor(a3!) output, Tensor query_start_loc, Tensor "
       "seq_lens, float scale, bool causal, Tensor? alibi_slopes, SymInt "
       "sliding_window_left, SymInt sliding_window_right, Tensor block_table, "
-      "float softcap, Tensor scheduler_metadata, Tensor? s_aux) -> ()",
+      "float softcap, Tensor scheduler_metadata, Tensor? s_aux, "
+      "float k_scale=1.0, float v_scale=1.0, str kv_cache_dtype=\"auto\") -> "
+      "()",
       &cpu_attention_with_kv_cache);
 
   // placeholders

csrc/cutlass_extensions/common.hpp

@@ -96,44 +96,14 @@ struct enable_sm90_or_later : Kernel {
 };
 
 template <typename Kernel>
-struct enable_sm90_only : Kernel {
+struct enable_sm100_to_sm120 : Kernel {
   template <typename... Args>
   CUTLASS_DEVICE void operator()(Args&&... args) {
 #if defined __CUDA_ARCH__
-  #if __CUDA_ARCH__ == 900
+  #if (__CUDA_ARCH__ >= 1000 && __CUDA_ARCH__ < 1200)
     Kernel::operator()(std::forward<Args>(args)...);
   #else
-    printf("This kernel only supports sm90.\n");
-    asm("trap;");
-  #endif
-#endif
-  }
-};
-
-template <typename Kernel>
-struct enable_sm100f_only : Kernel {
-  template <typename... Args>
-  CUTLASS_DEVICE void operator()(Args&&... args) {
-#if defined __CUDA_ARCH__
-  #if __CUDA_ARCH__ == 1000 || __CUDA_ARCH__ == 1030
-    Kernel::operator()(std::forward<Args>(args)...);
-  #else
-    printf("This kernel only supports sm100f.\n");
-    asm("trap;");
-  #endif
-#endif
-  }
-};
-
-template <typename Kernel>
-struct enable_sm100a_only : Kernel {
-  template <typename... Args>
-  CUTLASS_DEVICE void operator()(Args&&... args) {
-#if defined __CUDA_ARCH__
-  #if __CUDA_ARCH__ == 1000
-    Kernel::operator()(std::forward<Args>(args)...);
-  #else
-    printf("This kernel only supports sm100a.\n");
+    printf("This kernel only supports sm[100, 120).\n");
     asm("trap;");
   #endif
 #endif
@@ -148,7 +118,7 @@ struct enable_sm120_only : Kernel {
   #if __CUDA_ARCH__ == 1200
     Kernel::operator()(std::forward<Args>(args)...);
   #else
-    printf("This kernel only supports sm120.\n");
+    printf("This kernel only supports sm120a.\n");
     asm("trap;");
   #endif
 #endif
@@ -160,8 +130,13 @@ template <typename Kernel>
 struct enable_sm120_family : Kernel {
   template <typename... Args>
   CUTLASS_DEVICE void operator()(Args&&... args) {
-#if defined __CUDA_ARCH__ && (__CUDA_ARCH__ >= 1200 && __CUDA_ARCH__ < 1300)
+#if defined __CUDA_ARCH__
+  #if (__CUDA_ARCH__ >= 1200 && __CUDA_ARCH__ < 1300)
     Kernel::operator()(std::forward<Args>(args)...);
+  #else
+    printf("This kernel only supports sm120f.\n");
+    asm("trap;");
+  #endif
 #endif
   }
 };

csrc/fused_deepseek_v4_qnorm_rope_kv_insert_kernel.cu

@@ -0,0 +1,477 @@
+/*
+ * SPDX-License-Identifier: Apache-2.0
+ * SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+ *
+ * Horizontally-fused DeepseekV4-MLA kernel:
+ *   - Q side:  per-head RMSNorm (no weight) + GPT-J RoPE on last ROPE_DIM
+ *   - KV side: GPT-J RoPE on last ROPE_DIM + UE8M0 FP8 quant on NoPE + paged
+ *              cache insert
+ *
+ * Structured after `applyMLARopeAndAssignQKVKernelGeneration` in
+ * TensorRT-LLM's mlaKernels.cu: one kernel, one grid, with head-slot
+ * dispatch choosing Q vs KV work per warp.  The per-warp RMSNorm/RoPE
+ * skeleton is adapted from vllm-deepseek_v4's existing
+ * `fusedQKNormRopeKernel` (csrc/fused_qknorm_rope_kernel.cu).
+ *
+ * Assumptions (hard-coded for DeepseekV4 attention):
+ *   HEAD_DIM  = 512
+ *   ROPE_DIM  = 64   (RoPE applied to dims [NOPE_DIM, HEAD_DIM))
+ *   NOPE_DIM  = 448
+ *   QUANT_BLOCK = 64 (UE8M0 FP8 quant block)
+ *   FP8_MAX   = 448.0f
+ *   is_neox=false (GPT-J interleaved pairs)
+ *   cos_sin_cache layout [max_pos, rope_dim] = cos || sin (cos first, sin
+ *     second along last dim; each half is rope_dim/2 = 32 values)
+ *
+ * Cache layout per paged-cache block (block_size tokens):
+ *   [0,            bs*576):          token data, 448 fp8 + 128 bf16 each
+ *   [bs*576,       bs*576 + bs*8):   UE8M0 scales, 7 real + 1 pad per token
+ */
+
+#include <cmath>
+#include <cuda_fp8.h>
+#include <cuda_runtime.h>
+#include <type_traits>
+
+#include <ATen/cuda/CUDAContext.h>
+#include <c10/cuda/CUDAGuard.h>
+#include <torch/cuda.h>
+
+#include "cuda_compat.h"
+#include "dispatch_utils.h"
+#include "type_convert.cuh"
+
+#ifndef FINAL_MASK
+  #define FINAL_MASK 0xffffffffu
+#endif
+
+namespace vllm {
+namespace deepseek_v4_fused_ops {
+
+namespace {
+inline int getSMVersion() {
+  auto* props = at::cuda::getCurrentDeviceProperties();
+  return props->major * 10 + props->minor;
+}
+}  // namespace
+
+// ────────────────────────────────────────────────────────────────────────────
+// Constants
+// ────────────────────────────────────────────────────────────────────────────
+constexpr int kHeadDim = 512;
+constexpr int kRopeDim = 64;
+constexpr int kNopeDim = kHeadDim - kRopeDim;  // 448
+constexpr int kQuantBlock = 64;
+constexpr int kNumQuantBlocks = kNopeDim / kQuantBlock;   // 7
+constexpr int kScaleBytesPerToken = kNumQuantBlocks + 1;  // 8 (7 real + 1 pad)
+constexpr int kTokenDataBytes = kNopeDim + kRopeDim * 2;  // 448 + 128 = 576
+constexpr float kFp8Max = 448.0f;
+
+// Per-warp layout:  32 lanes × 16 elems/lane = 512 elems = HEAD_DIM.
+constexpr int kNumLanes = 32;
+constexpr int kElemsPerLane = kHeadDim / kNumLanes;  // 16
+
+// ────────────────────────────────────────────────────────────────────────────
+// Small inline helpers
+// ────────────────────────────────────────────────────────────────────────────
+__device__ __forceinline__ float warp4MaxAbs(float val) {
+  // Reduce absolute max across 4 consecutive lanes (lane id & 3 group).
+  float peer = __shfl_xor_sync(FINAL_MASK, val, 1);
+  val = fmaxf(val, peer);
+  peer = __shfl_xor_sync(FINAL_MASK, val, 2);
+  val = fmaxf(val, peer);
+  return val;
+}
+
+template <typename T>
+__device__ __forceinline__ float warpSum(float val) {
+#pragma unroll
+  for (int mask = 16; mask > 0; mask >>= 1) {
+    val += __shfl_xor_sync(FINAL_MASK, val, mask, 32);
+  }
+  return val;
+}
+
+// ────────────────────────────────────────────────────────────────────────────
+// Kernel
+// ────────────────────────────────────────────────────────────────────────────
+//
+// Grid: 1D, gridDim.x = ceil(num_tokens_full * (num_heads_q + 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)
+//
+// 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.
+//
+template <typename scalar_t_in>
+__global__ void fusedDeepseekV4QNormRopeKVRopeQuantInsertKernel(
+    scalar_t_in* __restrict__ q_inout,         // [N, H, 512] bf16, in place
+    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
+    int64_t const* __restrict__ position_ids,  // [N] i64
+    float const* __restrict__ cos_sin_cache,   // [max_pos, 64] fp32
+    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 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)
+  // BF16 _typeConvert specialization is unavailable on pre-Ampere.  The
+  // DeepseekV4 kernel only runs with bf16 inputs in practice, so compile a
+  // no-op stub for sm_70/sm_75 to keep multi-arch builds happy.
+  if constexpr (std::is_same_v<scalar_t_in, c10::BFloat16>) {
+    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;
+    if (tokenIdx >= num_tokens_full) return;
+
+    bool const isKV = (slotIdx == num_heads_q);
+    // KV branch: skip DP-padded tokens (no slot reserved for them).
+    if (isKV && tokenIdx >= num_tokens_insert) return;
+
+    // PDL: wait for predecessor kernel (upstream q/kv producer) to signal
+    // before touching any global memory.  No-op when PDL is not enabled on
+    // the launch.  The CUDA runtime wrapper emits the griddepcontrol.wait
+    // PTX with the required memory clobber internally.
+#if defined(__CUDA_ARCH__) && (__CUDA_ARCH__ >= 900)
+    cudaGridDependencySynchronize();
+#endif
+
+    // 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;
+    }
+
+    // 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);
+
+    {
+      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();
+#endif
+      return;
+    }
+
+    // ═══════════════════════════════════════════════════════════════════════
+    // KV branch.
+    // ═══════════════════════════════════════════════════════════════════════
+    int64_t const slot_id = slot_mapping[tokenIdx];
+    if (slot_id < 0) {
+#if defined(__CUDA_ARCH__) && (__CUDA_ARCH__ >= 900)
+      cudaTriggerProgrammaticLaunchCompletion();
+#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;
+
+    // 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]));
+    }
+
+    // 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 (!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);
+        __nv_fp8_storage_t s =
+            __nv_cvt_float_to_fp8(scaled, __NV_SATFINITE, __NV_E4M3);
+        out_bytes[i] = static_cast<uint8_t>(s);
+      }
+      // One 16-byte STG per lane.
+      *reinterpret_cast<uint4*>(token_fp8_ptr + dim_base) =
+          *reinterpret_cast<uint4 const*>(out_bytes);
+
+      // 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
+#if (!defined(__CUDA_ARCH__) || __CUDA_ARCH__ < 800) && !defined(USE_ROCM)
+  }
+#endif
+}
+
+// ────────────────────────────────────────────────────────────────────────────
+// 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,
+    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,
+    cudaStream_t stream) {
+  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);
+  int const grid =
+      static_cast<int>((total_warps + kWarpsPerBlock - 1) / kWarpsPerBlock);
+
+  // PDL: enable programmatic stream serialization whenever the hardware
+  // supports it (SM90+).  On pre-Hopper GPUs the attribute is unavailable,
+  // so leave numAttrs = 0 and launch as a regular kernel.
+  static int const sm_version = getSMVersion();
+  // Host-side guard: the device kernel body is compiled as a no-op for
+  // bf16 on pre-Ampere (sm_70/sm_75) because _typeConvert<BFloat16> is
+  // unavailable there.  Refuse the launch loudly instead of silently
+  // skipping the work.
+  TORCH_CHECK(
+      sm_version >= 80,
+      "fused_deepseek_v4_qnorm_rope_kv_rope_quant_insert requires sm_80+ "
+      "(Ampere or newer); got sm_",
+      sm_version);
+  cudaLaunchConfig_t config;
+  config.gridDim = dim3(grid);
+  config.blockDim = dim3(kBlockSize);
+  config.dynamicSmemBytes = 0;
+  config.stream = stream;
+  cudaLaunchAttribute attrs[1];
+  attrs[0].id = cudaLaunchAttributeProgrammaticStreamSerialization;
+  attrs[0].val.programmaticStreamSerializationAllowed = 1;
+  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);
+}
+
+}  // 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 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
+    double eps, int64_t cache_block_size) {
+  TORCH_CHECK(q.is_cuda() && q.is_contiguous(), "q 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,
+              "slot_mapping must be int64 CUDA");
+  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(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(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]");
+  TORCH_CHECK(cos_sin_cache.dtype() == torch::kFloat32,
+              "cos_sin_cache must be float32");
+
+  // 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_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 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));
+  auto stream = at::cuda::getCurrentCUDAStream();
+
+  VLLM_DISPATCH_HALF_TYPES(
+      q.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*>(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);
+      });
+}

csrc/layernorm_kernels.cu

@@ -77,7 +77,8 @@ __global__ void rms_norm_kernel(
 #pragma unroll
     for (int j = 0; j < VEC_SIZE; j++) {
       float x = static_cast<float>(src1.val[j]);
-      dst.val[j] = ((scalar_t)(x * s_variance)) * src2.val[j];
+      float w = static_cast<float>(src2.val[j]);
+      dst.val[j] = static_cast<scalar_t>(x * s_variance * w);
     }
     v_out[i] = dst;
   }
@@ -134,10 +135,17 @@ fused_add_rms_norm_kernel(
   for (int idx = threadIdx.x; idx < vec_hidden_size; idx += blockDim.x) {
     int id = blockIdx.x * vec_hidden_size + idx;
     int64_t strided_id = blockIdx.x * vec_input_stride + idx;
-    _f16Vec<scalar_t, width> temp = residual_v[id];
-    temp *= s_variance;
-    temp *= weight_v[idx];
-    input_v[strided_id] = temp;
+    _f16Vec<scalar_t, width> res = residual_v[id];
+    _f16Vec<scalar_t, width> w = weight_v[idx];
+    _f16Vec<scalar_t, width> out;
+    using Converter = _typeConvert<scalar_t>;
+#pragma unroll
+    for (int j = 0; j < width; ++j) {
+      float x = Converter::convert(res.data[j]);
+      float wf = Converter::convert(w.data[j]);
+      out.data[j] = Converter::convert(x * s_variance * wf);
+    }
+    input_v[strided_id] = out;
   }
 }
 
@@ -174,8 +182,8 @@ fused_add_rms_norm_kernel(
 
   for (int idx = threadIdx.x; idx < hidden_size; idx += blockDim.x) {
     float x = (float)residual[blockIdx.x * hidden_size + idx];
-    input[blockIdx.x * input_stride + idx] =
-        ((scalar_t)(x * s_variance)) * weight[idx];
+    float w = (float)weight[idx];
+    input[blockIdx.x * input_stride + idx] = (scalar_t)(x * s_variance * w);
   }
 }
 

csrc/layernorm_quant_kernels.cu

@@ -65,9 +65,16 @@ __global__ void rms_norm_static_fp8_quant_kernel(
 #pragma unroll
     for (int j = 0; j < VEC_SIZE; j++) {
       float x = static_cast<float>(src1.val[j]);
-      float const out_norm = ((scalar_t)(x * s_variance)) * src2.val[j];
+      float w = static_cast<float>(src2.val[j]);
+      // Round normalized result through scalar_t to match the precision of the
+      // unfused composite (rms_norm writes scalar_t, then
+      // static_scaled_fp8_quant re-loads it as float before FP8 conversion).
+      // Without this round, the fused path is strictly more accurate and
+      // disagrees with the composite at exact E4M3 quantization tie boundaries.
+      scalar_t out_norm = static_cast<scalar_t>(x * s_variance * w);
       out[blockIdx.x * hidden_size + idx * VEC_SIZE + j] =
-          scaled_fp8_conversion<true, fp8_type>(out_norm, scale_inv);
+          scaled_fp8_conversion<true, fp8_type>(static_cast<float>(out_norm),
+                                                scale_inv);
     }
   }
 }
@@ -127,13 +134,21 @@ fused_add_rms_norm_static_fp8_quant_kernel(
 
   for (int idx = threadIdx.x; idx < vec_hidden_size; idx += blockDim.x) {
     int id = blockIdx.x * vec_hidden_size + idx;
-    _f16Vec<scalar_t, width> temp = residual_v[id];
-    temp *= s_variance;
-    temp *= weight_v[idx];
+    _f16Vec<scalar_t, width> res = residual_v[id];
+    _f16Vec<scalar_t, width> w = weight_v[idx];
+    using Converter = _typeConvert<scalar_t>;
+    using HipT = typename Converter::hip_type;
 #pragma unroll
     for (int i = 0; i < width; ++i) {
-      out[id * width + i] =
-          scaled_fp8_conversion<true, fp8_type>(float(temp.data[i]), scale_inv);
+      float x = Converter::convert(res.data[i]);
+      float wf = Converter::convert(w.data[i]);
+      // See note in rms_norm_static_fp8_quant_kernel: round through scalar_t
+      // to match the unfused composite path at FP8 boundaries. We use the
+      // backend's hip_type for the intermediate since c10::Half/BFloat16 has
+      // ambiguous conversions on CUDA and no implicit conversion on ROCm.
+      HipT out_norm_h = Converter::convert(x * s_variance * wf);
+      out[id * width + i] = scaled_fp8_conversion<true, fp8_type>(
+          Converter::convert(out_norm_h), scale_inv);
     }
   }
 }
@@ -176,9 +191,12 @@ fused_add_rms_norm_static_fp8_quant_kernel(
 
   for (int idx = threadIdx.x; idx < hidden_size; idx += blockDim.x) {
     float x = (float)residual[blockIdx.x * hidden_size + idx];
-    float const out_norm = ((scalar_t)(x * s_variance)) * weight[idx];
-    out[blockIdx.x * hidden_size + idx] =
-        scaled_fp8_conversion<true, fp8_type>(out_norm, scale_inv);
+    float w = (float)weight[idx];
+    // See note in rms_norm_static_fp8_quant_kernel: round through scalar_t
+    // to match the unfused composite path at FP8 boundaries.
+    scalar_t out_norm = static_cast<scalar_t>(x * s_variance * w);
+    out[blockIdx.x * hidden_size + idx] = scaled_fp8_conversion<true, fp8_type>(
+        static_cast<float>(out_norm), scale_inv);
   }
 }
 

csrc/libtorch_stable/ops.h

@@ -134,4 +134,13 @@ void silu_and_mul_nvfp4_quant(torch::stable::Tensor& out,
                               torch::stable::Tensor& input,
                               torch::stable::Tensor& input_global_scale);
 
+void cutlass_mxfp4_group_mm(torch::stable::Tensor& output,
+                            const torch::stable::Tensor& a,
+                            const torch::stable::Tensor& b,
+                            const torch::stable::Tensor& a_blockscale,
+                            const torch::stable::Tensor& b_blockscales,
+                            const torch::stable::Tensor& problem_sizes,
+                            const torch::stable::Tensor& expert_offsets,
+                            const torch::stable::Tensor& sf_offsets);
+
 #endif