optimize allpool forward

Signed-off-by: yewentao256 <zhyanwentao@126.com>
[Docs] Add non-causal support to attention backend docs (#41643 )
2026-05-04 22:47:43 +00:00 · 2026-05-04 20:35:15 +00:00 · 2026-05-04 20:13:37 +00:00 · 2026-05-04 12:45:24 -07:00 · 2026-05-04 11:46:00 -07:00 · 2026-05-04 21:43:07 +03:00
632 changed files with 41908 additions and 8235 deletions
@@ -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: []
@@ -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
@@ -126,5 +126,4 @@ steps:
      '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_qwenvl.py || true) &&
      pytest -v -s lora/test_whisper.py'
@@ -27,7 +27,7 @@ steps:
          - "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 manylinux_2_31"
+          - "bash .buildkite/scripts/upload-nightly-wheels.sh"
        env:
          DOCKER_BUILDKIT: "1"

@@ -37,10 +37,10 @@ steps:
        agents:
          queue: arm64_cpu_queue_release
        commands:
-          - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=13.0.2 --build-arg torch_cuda_arch_list=\"${CUDA_ARCH_AARCH64}\" --build-arg BUILD_BASE_IMAGE=nvidia/cuda:13.0.2-devel-ubuntu22.04  --tag vllm-ci:build-image --target build --progress plain -f docker/Dockerfile ."
+          - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=13.0.2 --build-arg torch_cuda_arch_list=\"${CUDA_ARCH_AARCH64}\" --build-arg BUILD_OS=manylinux --build-arg BUILD_BASE_IMAGE=pytorch/manylinuxaarch64-builder:cuda13.0 --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"

@@ -53,7 +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"

@@ -66,7 +66,7 @@ steps:
          - "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"

@@ -76,10 +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=13.0.2 --build-arg torch_cuda_arch_list=\"${CUDA_ARCH_X86}\" --build-arg BUILD_BASE_IMAGE=nvidia/cuda:13.0.2-devel-ubuntu22.04 --tag vllm-ci:build-image --target build --progress plain -f docker/Dockerfile ."
+          - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=13.0.2 --build-arg torch_cuda_arch_list=\"${CUDA_ARCH_X86}\" --build-arg BUILD_OS=manylinux --build-arg BUILD_BASE_IMAGE=pytorch/manylinux2_28-builder:cuda13.0 --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"

@@ -92,7 +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"

@@ -121,7 +121,19 @@ steps:
          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=13.0.2 --build-arg torch_cuda_arch_list=\"${CUDA_ARCH_X86}\" --build-arg INSTALL_KV_CONNECTORS=true --build-arg BUILD_BASE_IMAGE=nvidia/cuda:13.0.2-devel-ubuntu22.04 --tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$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:$BUILDKITE_COMMIT"
@@ -134,7 +146,19 @@ steps:
          queue: arm64_cpu_queue_release
        commands:
          - "aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin public.ecr.aws/q9t5s3a7"
-          - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=13.0.2 --build-arg torch_cuda_arch_list=\"${CUDA_ARCH_AARCH64}\" --build-arg INSTALL_KV_CONNECTORS=true --build-arg BUILD_BASE_IMAGE=nvidia/cuda:13.0.2-devel-ubuntu22.04 --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 12.9"
@@ -144,7 +168,18 @@ steps:
          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 torch_cuda_arch_list=\"${CUDA_ARCH_X86_CU129}\" --build-arg INSTALL_KV_CONNECTORS=true --tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)-cu129 --target vllm-openai --progress plain -f docker/Dockerfile ."
+          - |
+            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)-cu129 public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-cu129"
@@ -157,7 +192,18 @@ steps:
          queue: arm64_cpu_queue_release
        commands:
          - "aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin public.ecr.aws/q9t5s3a7"
-          - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=12.9.1 --build-arg torch_cuda_arch_list=\"${CUDA_ARCH_AARCH64_CU129}\" --build-arg INSTALL_KV_CONNECTORS=true --tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)-cu129 --target vllm-openai --progress plain -f docker/Dockerfile ."
+          - |
+            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 13.0 - Ubuntu 24.04"
@@ -167,7 +213,21 @@ steps:
          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=13.0.2 --build-arg UBUNTU_VERSION=24.04 --build-arg GDRCOPY_OS_VERSION=Ubuntu24_04 --build-arg torch_cuda_arch_list=\"${CUDA_ARCH_X86}\" --build-arg INSTALL_KV_CONNECTORS=true --build-arg BUILD_BASE_IMAGE=nvidia/cuda:13.0.2-devel-ubuntu24.04 --tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$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"
@@ -179,7 +239,21 @@ steps:
          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=13.0.2 --build-arg UBUNTU_VERSION=24.04 --build-arg GDRCOPY_OS_VERSION=Ubuntu24_04 --build-arg torch_cuda_arch_list=\"${CUDA_ARCH_AARCH64}\" --build-arg INSTALL_KV_CONNECTORS=true --build-arg BUILD_BASE_IMAGE=nvidia/cuda:13.0.2-devel-ubuntu24.04 --tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$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 12.9 - Ubuntu 24.04"
@@ -189,7 +263,20 @@ steps:
          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 torch_cuda_arch_list=\"${CUDA_ARCH_X86_CU129}\" --build-arg INSTALL_KV_CONNECTORS=true --tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)-cu129-ubuntu2404 --target vllm-openai --progress plain -f docker/Dockerfile ."
+          - |
+            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"
@@ -201,7 +288,20 @@ steps:
          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 torch_cuda_arch_list=\"${CUDA_ARCH_AARCH64_CU129}\" --build-arg INSTALL_KV_CONNECTORS=true --tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)-cu129-ubuntu2404 --target vllm-openai --progress plain -f docker/Dockerfile ."
+          - |
+            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"
@@ -623,7 +723,7 @@ steps:
      - "bash tools/vllm-rocm/generate-rocm-wheels-root-index.sh"
    env:
      S3_BUCKET: "vllm-wheels"
-      VARIANT: "rocm721"
+      VARIANT: "rocm722"

  # ROCm Job 6: Build ROCm Release Docker Image
  - label: ":docker: Build release image - x86_64 - ROCm"
@@ -0,0 +1,55 @@
+#!/bin/bash
+# Usage: ./ci-fetch-log.sh <buildkite_job_url> [output_file]
+#        ./ci-fetch-log.sh <build_number> <job_uuid> [output_file]
+#
+# Downloads the raw log for a Buildkite job from the public, unauthenticated
+# /organizations/<org>/pipelines/<pipeline>/builds/<n>/jobs/<uuid>/download
+# endpoint, then strips ANSI/timestamps via ci-clean-log.sh.
+#
+# Find <build_number> and <job_uuid> via:
+#   gh pr checks <PR> --repo vllm-project/vllm
+# Each failing row's URL is .../builds/<build_number>#<job_uuid>.
+
+set -euo pipefail
+
+ORG="vllm"
+PIPELINE="ci"
+
+usage() {
+    echo "Usage: $0 <buildkite_job_url> [output_file]"
+    echo "       $0 <build_number> <job_uuid> [output_file]"
+    exit 1
+}
+
+if [ $# -lt 1 ]; then usage; fi
+
+if [[ "$1" == https://* ]]; then
+    BUILD=$(echo "$1" | sed -nE 's#.*/builds/([0-9]+).*#\1#p')
+    JOB=$(echo "$1" | grep -oE '[0-9a-f]{8}-[0-9a-f-]+' | head -n 1)
+    OUT="${2:-ci-${BUILD}-${JOB:0:8}.log}"
+else
+    if [ $# -lt 2 ]; then usage; fi
+    BUILD="$1"
+    JOB="$2"
+    OUT="${3:-ci-${BUILD}-${JOB:0:8}.log}"
+fi
+
+if [ -z "$BUILD" ] || [ -z "$JOB" ]; then
+    echo "Could not parse build number or job UUID from: $1" >&2
+    usage
+fi
+
+COOKIES=$(mktemp)
+trap 'rm -f "$COOKIES"' EXIT
+
+# Buildkite issues a session cookie on first hit; subsequent /download needs it.
+curl -fsSL -c "$COOKIES" -A "vllm-ci-fetch-log" \
+    "https://buildkite.com/${ORG}/${PIPELINE}/builds/${BUILD}" -o /dev/null
+
+curl -fsSL -b "$COOKIES" -A "vllm-ci-fetch-log" \
+    "https://buildkite.com/organizations/${ORG}/pipelines/${PIPELINE}/builds/${BUILD}/jobs/${JOB}/download" \
+    -o "$OUT"
+
+bash "$(dirname "$0")/ci-clean-log.sh" "$OUT"
+
+echo "$OUT"
@@ -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())
@@ -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}"
@@ -10,20 +10,13 @@ set -ex
 BUCKET="vllm-wheels"
 INDICES_OUTPUT_DIR="indices"
 DEFAULT_VARIANT_ALIAS="cu130" # align with vLLM_MAIN_CUDA_VERSION in vllm/envs.py
-PYTHON="${PYTHON_PROG:-python3}" # try to read from env var, otherwise use python3
 SUBPATH=$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 ========

@@ -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
@@ -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"
+}
@@ -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)"
+}
@@ -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 ----
@@ -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
@@ -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
@@ -388,18 +388,18 @@ steps:
    - python3 basic/offline_inference/embed.py
    - python3 basic/offline_inference/score.py
    # 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
    # Pooling models
    - python3 pooling/embed/vision_embedding_offline.py --seed 0
    # 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 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 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 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

 #----------------------------------------------------------  mi250 · kernels  ----------------------------------------------------------#

@@ -1108,6 +1108,7 @@ steps:
  - export VLLM_TEST_CLEAN_GPU_MEMORY=1
  - VLLM_TEST_CLEAN_GPU_MEMORY=1 pytest -v -s tests/compile/passes/distributed/test_async_tp.py
  - pytest -v -s tests/compile/passes/distributed/test_sequence_parallelism.py
+  - pytest -v -s tests/compile/passes/distributed/test_tp2_ar_rms.py::test_tp2_ar_rms_fusions

 #-----------------------------------------------------------  mi300 · cuda  ------------------------------------------------------------#

@@ -1168,13 +1169,13 @@ steps:
  - vllm/v1/attention/backends/
  - vllm/v1/attention/selector.py
  - tests/distributed/test_context_parallel.py
-  - examples/offline_inference/data_parallel.py
+  - examples/features/data_parallel/data_parallel_offline.py
  - vllm/_aiter_ops.py
  - vllm/platforms/rocm.py
  commands:
  - export TORCH_NCCL_BLOCKING_WAIT=1
  - pytest -v -s tests/distributed/test_context_parallel.py
-  - 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=allgather_reducescatter --disable-nccl-for-dp-synchronization
+  - 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=allgather_reducescatter --disable-nccl-for-dp-synchronization

 - label: Distributed Tests (4xA100-4xMI300) # TBD
  timeout_in_minutes: 180
@@ -1203,7 +1204,7 @@ steps:
  - tests/distributed/test_torchrun_example.py
  - tests/distributed/test_torchrun_example_moe.py
  - examples/rl/
-  - tests/examples/offline_inference/data_parallel.py
+  - tests/examples/features/data_parallel/data_parallel_offline.py
  - vllm/platforms/rocm.py
  commands:
  - export TORCH_NCCL_BLOCKING_WAIT=1
@@ -1213,7 +1214,7 @@ steps:
  - PP_SIZE=2 TP_SIZE=2 torchrun --nproc-per-node=4 distributed/test_torchrun_example_moe.py
  - DP_SIZE=4 ENABLE_EP=1 torchrun --nproc-per-node=4 distributed/test_torchrun_example_moe.py
  - TP_SIZE=2 DP_SIZE=2 ENABLE_EP=1 torchrun --nproc-per-node=4 distributed/test_torchrun_example_moe.py
-  - 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
@@ -1266,7 +1267,7 @@ steps:
  optional: true
  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
@@ -1275,7 +1276,7 @@ steps:
  - vllm/platforms/rocm.py
  commands:
  - export TORCH_NCCL_BLOCKING_WAIT=1
-  - 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

 #--------------------------------------------------------  mi300 · entrypoints  --------------------------------------------------------#

@@ -1647,18 +1648,18 @@ steps:
    - python3 basic/offline_inference/embed.py
    - python3 basic/offline_inference/score.py
    # 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
    # Pooling models
    - python3 pooling/embed/vision_embedding_offline.py --seed 0
    # 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 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 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 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

 #----------------------------------------------------------  mi300 · kernels  ----------------------------------------------------------#

@@ -1951,8 +1952,8 @@ 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
-  - VLLM_WORKER_MULTIPROC_METHOD=spawn python3 examples/offline_inference/audio_language.py --model-type whisper
+  - python3 examples/generate/multimodal/vision_language_offline.py --model-type qwen2_5_vl
+  - VLLM_WORKER_MULTIPROC_METHOD=spawn python3 examples/generate/multimodal/audio_language_offline.py --model-type whisper

 #-------------------------------------------------------  mi300 · quantization  --------------------------------------------------------#

@@ -2302,7 +2303,7 @@ steps:
  commands:
  - export TORCH_NCCL_BLOCKING_WAIT=1
  - VLLM_ALLOW_INSECURE_SERIALIZATION=1 python3 examples/rl/rlhf_async_new_apis.py
-  - 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_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
@@ -2713,7 +2714,7 @@ steps:
  - vllm/v1/attention/selector.py
  - tests/distributed/test_context_parallel.py
  - tests/v1/distributed/test_dbo.py
-  - examples/offline_inference/data_parallel.py
+  - examples/features/data_parallel/data_parallel_offline.py
  - vllm/_aiter_ops.py
  - vllm/platforms/rocm.py
  commands:
@@ -2930,18 +2931,18 @@ steps:
  - python3 basic/offline_inference/embed.py
  - python3 basic/offline_inference/score.py
  # 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
  # Pooling models
  - python3 pooling/embed/vision_embedding_offline.py --seed 0
  # 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 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 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 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

 #----------------------------------------------------------  mi355 · kernels  ----------------------------------------------------------#

@@ -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:
@@ -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:
@@ -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
@@ -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
@@ -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
@@ -3,6 +3,7 @@ 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
@@ -13,6 +14,7 @@ steps:
    - 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
@@ -24,6 +26,7 @@ steps:
    - 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
@@ -35,6 +38,7 @@ steps:
    - 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
@@ -46,6 +50,7 @@ steps:
    - 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
@@ -57,6 +62,7 @@ steps:
    - 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
@@ -71,6 +77,7 @@ steps:
    - 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"
@@ -84,6 +91,7 @@ steps:
    - 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
@@ -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
@@ -222,11 +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"
+    - ./.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
@@ -241,6 +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
@@ -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
@@ -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
@@ -30,6 +33,7 @@ steps:
  - 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
@@ -39,6 +43,7 @@ steps:
  - 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
@@ -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
@@ -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:
@@ -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"
@@ -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
@@ -150,6 +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:
@@ -161,6 +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
@@ -177,6 +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
@@ -186,6 +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
@@ -198,6 +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
@@ -214,6 +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
@@ -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
@@ -126,3 +137,10 @@ steps:
  commands:
    - uv pip install --system 'gpt-oss[eval]==0.0.5'
    - pytest -s -v evals/gpt_oss/test_gpqa_correctness.py --config-list-file=configs/models-b200.txt
+
+- label: MRCR Eval Small Models
+  timeout_in_minutes: 30
+  source_file_dependencies:
+  - tests/evals/mrcr/
+  commands:
+    - pytest -s -v evals/mrcr/test_mrcr_correctness.py --config-list-file=evals/mrcr/configs/models-small.txt
@@ -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:
@@ -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
@@ -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
@@ -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:
@@ -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
@@ -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
@@ -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
@@ -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:
@@ -36,6 +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/
@@ -51,6 +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/
@@ -67,7 +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:
@@ -80,6 +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:
@@ -91,6 +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:
@@ -101,6 +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/
@@ -117,6 +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/
@@ -126,6 +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/
@@ -135,6 +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:
@@ -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
@@ -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
@@ -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
@@ -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.
@@ -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:
@@ -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:
@@ -13,6 +14,7 @@ steps:
    - 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
@@ -24,6 +26,7 @@ steps:
    - 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:
@@ -35,6 +38,7 @@ steps:
    - 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
@@ -47,6 +51,7 @@ steps:
    - 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:
@@ -57,6 +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:
@@ -67,6 +73,7 @@ steps:
    - 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
@@ -78,6 +85,7 @@ steps:
    - 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
@@ -89,3 +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"
@@ -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
@@ -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:
@@ -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
@@ -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
@@ -1053,7 +1053,6 @@ 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/topk_softplus_sqrt_kernels.cu")
 endif()

@@ -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,
+        )
@@ -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:
@@ -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 {
@@ -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);
+            });
      });
 }
@@ -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;
@@ -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);
+    }
+  }
+}
@@ -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                                         \
@@ -1374,6 +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;
@@ -1753,7 +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;
@@ -1823,7 +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);
        }
      }
    }
@@ -1947,8 +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;

@@ -1962,7 +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);
@@ -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
@@ -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;
@@ -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
@@ -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;
@@ -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) {
@@ -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> {
@@ -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; }
 };

@@ -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);
@@ -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

@@ -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"); }
@@ -384,15 +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
@@ -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
  }
 };
@@ -141,7 +141,7 @@ struct cutlass_3x_gemm_sm100 {
              sizeof(typename CollectiveEpilogue::SharedStorage))>,
          KernelSchedule>::CollectiveOp;

-  using GemmKernel = enable_sm100f_only<cutlass::gemm::kernel::GemmUniversal<
+  using GemmKernel = enable_sm100_to_sm120<cutlass::gemm::kernel::GemmUniversal<
      Shape<int, int, int, int>, CollectiveMainloop, CollectiveEpilogue, void>>;
 };

@@ -125,7 +125,7 @@ struct cutlass_3x_gemm_fp8_blockwise {
          MainloopScheduler
      >::CollectiveOp>;

-  using KernelType = enable_sm100f_only<cutlass::gemm::kernel::GemmUniversal<
+  using KernelType = enable_sm100_to_sm120<cutlass::gemm::kernel::GemmUniversal<
      Shape<int, int, int, int>, CollectiveMainloop, CollectiveEpilogue>>;

  struct GemmKernel : public KernelType {};
@@ -92,7 +92,7 @@ struct cutlass_3x_gemm_sm100_fp8 {
  // -----------------------------------------------------------
  // Kernel definition
  // -----------------------------------------------------------
-  using GemmKernel = enable_sm100f_only<cutlass::gemm::kernel::GemmUniversal<
+  using GemmKernel = enable_sm100_to_sm120<cutlass::gemm::kernel::GemmUniversal<
      Shape<int, int, int, int>, CollectiveMainloop, CollectiveEpilogue, void>>;
 };

@@ -236,17 +236,41 @@ void per_token_group_quant_8bit(const torch::stable::Tensor& input,
 #undef LAUNCH_KERNEL
 }

-template <typename T, typename DST_DTYPE>
-__global__ void per_token_group_quant_8bit_packed_kernel(
+// Register-resident fast path for group_size==128.
+//
+// Each thread holds 16 source elements (32 B = uint4 x 2) in registers across
+// the absmax reduce -> scale compute -> quantize pipeline. No shared memory.
+// UE8M0 scale extracted via bit math (bit-exact with exp2f(ceilf(log2f))).
+//
+// Loads two contiguous uint4s (16 B + 16 B = 32 B) per thread; on Blackwell
+// nvcc fuses these into a single 256-bit LDG.E.256.
+//
+// Constraints: GROUP_SIZE % (THREADS_PER_GROUP * VEC_SIZE) == 0; for
+// THREADS_PER_GROUP=8 and bf16/fp16 (VEC_SIZE=16), this means GROUP_SIZE=128.
+template <typename T, typename DST_DTYPE, int GROUP_SIZE>
+__global__ void per_token_group_quant_8bit_packed_register_kernel(
    const T* __restrict__ input, void* __restrict__ output_q,
-    unsigned int* __restrict__ output_s_packed, const int group_size,
-    const int num_groups_padded, const int groups_per_block,
-    const int padded_groups_per_row, const int groups_per_row, const int mn,
-    const int tma_aligned_mn, const int num_scale_elems, const float eps,
+    unsigned int* __restrict__ output_s_packed, const int64_t num_groups_padded,
+    const int groups_per_block, const int padded_groups_per_row,
+    const int groups_per_row, const int mn, const int output_q_mn_extent,
+    const int tma_aligned_mn, const int64_t num_scale_elems, const float eps,
    const float min_8bit, const float max_8bit) {
-  const int threads_per_group = 16;
-  const int64_t local_group_id = threadIdx.x / threads_per_group;
-  const int lane_id = threadIdx.x % threads_per_group;
+  static_assert(GROUP_SIZE == 128, "fast path supports GROUP_SIZE==128");
+  constexpr int THREADS_PER_GROUP = 8;
+  constexpr int VEC_SIZE = 32 / sizeof(T);  // 16 for bf16/fp16
+  static_assert(GROUP_SIZE == THREADS_PER_GROUP * VEC_SIZE,
+                "GROUP_SIZE must equal THREADS_PER_GROUP * VEC_SIZE");
+  // Each group's 8 threads must live in a single warp octet so the
+  // 0xffu << (threadIdx.x & 24u) shuffle mask selects exactly the lanes
+  // that share a group. Requires 32 % THREADS_PER_GROUP == 0 and the host
+  // to launch num_threads as a multiple of THREADS_PER_GROUP (which it does
+  // via num_threads = groups_per_block * THREADS_PER_GROUP).
+  static_assert(32 % THREADS_PER_GROUP == 0,
+                "THREADS_PER_GROUP must divide warp size for the shuffle "
+                "mask to be valid");
+
+  const int local_group_id = threadIdx.x / THREADS_PER_GROUP;
+  const int lane_id = threadIdx.x % THREADS_PER_GROUP;

  const int64_t block_group_id = blockIdx.x * groups_per_block;
  const int64_t global_group_id = block_group_id + local_group_id;
@@ -254,141 +278,207 @@ __global__ void per_token_group_quant_8bit_packed_kernel(
    return;
  }

-  // map flat group id to 2D indices (mn_idx, sf_k_idx)
  const int sf_k_idx =
      static_cast<int>(global_group_id % padded_groups_per_row);
  const int mn_idx = static_cast<int>(global_group_id / padded_groups_per_row);
-
-  // whether it is a valid group (not padding)
  const bool is_valid_group = (mn_idx < mn) && (sf_k_idx < groups_per_row);

-  // shared memory to cache each group's data to avoid double DRAM reads.
-  extern __shared__ __align__(16) char smem_raw[];
-  T* smem = reinterpret_cast<T*>(smem_raw);
-  T* smem_group = smem + local_group_id * group_size;
-
-  // compute scale for valid groups
-  float y_s = 0.f;
+  // Load 16 input elements (32 B) into registers as two adjacent uint4
+  // loads. nvcc keeps these as 2x LDG.E.128 on sm_100; the per-thread cost
+  // is dominated by HBM bandwidth at large MN, so a fused 256-bit load via
+  // inline PTX gave no measurable speedup.
+  // alignas(16) is required so the uint4* reinterpret_cast below is
+  // well-defined for T == bf16/fp16 (default alignof is 2).
+  alignas(16) T regs[VEC_SIZE];
+  float local_absmax = eps;
  if (is_valid_group) {
    const T* group_input =
-        input + static_cast<int64_t>(mn_idx) * groups_per_row * group_size +
-        sf_k_idx * group_size;
-    y_s = ComputeGroupScale<T, true>(group_input, smem_group, group_size,
-                                     lane_id, threads_per_group, eps, max_8bit);
+        input + static_cast<int64_t>(mn_idx) * groups_per_row * GROUP_SIZE +
+        sf_k_idx * GROUP_SIZE + lane_id * VEC_SIZE;
+    uint4* dst = reinterpret_cast<uint4*>(&regs[0]);
+    const uint4* src = reinterpret_cast<const uint4*>(group_input);
+    dst[0] = src[0];
+    dst[1] = src[1];
+#pragma unroll
+    for (int i = 0; i < VEC_SIZE; ++i) {
+      float v = fabsf(static_cast<float>(regs[i]));
+      local_absmax = fmaxf(local_absmax, v);
+    }
  }

-  // pack 4 scales into a uint32 exponent
+  // 8-lane subgroup shuffle reduce (octet of the warp). The mask selects the
+  // 8 lanes within the warp that share a group.
+  unsigned mask = 0xffu << (threadIdx.x & 24u);
+  local_absmax = fmaxf(local_absmax, __shfl_xor_sync(mask, local_absmax, 4));
+  local_absmax = fmaxf(local_absmax, __shfl_xor_sync(mask, local_absmax, 2));
+  local_absmax = fmaxf(local_absmax, __shfl_xor_sync(mask, local_absmax, 1));
+
+  float y_s = local_absmax / max_8bit;
+  y_s = fmaxf(y_s, 1e-10f);
+  uint32_t bits = __float_as_uint(y_s);
+  uint32_t exp_bits = (bits >> 23) & 0xffu;
+  uint32_t mant_bits = bits & 0x7fffffu;
+  uint8_t exp_byte =
+      static_cast<uint8_t>(exp_bits + (mant_bits != 0u ? 1u : 0u));
+
+  // Lane 0 writes the packed scale byte.
  if (lane_id == 0) {
-    // each uint32 in output_s_packed stores 4 packed scales
    const int sf_k_pack_idx = sf_k_idx / 4;
    const int pos = sf_k_idx % 4;
    const int out_idx = sf_k_pack_idx * tma_aligned_mn + mn_idx;
-
    if (is_valid_group) {
-      // reinterpret the UE8M0 scale y_s as IEEE bits, extract the 8-bit
-      // exponent, and place it into the correct byte of the 32-bit word.
-      const unsigned int bits = __float_as_uint(y_s);
-      const uint8_t exponent = static_cast<uint8_t>((bits >> 23u) & 0xffu);
-      reinterpret_cast<uint8_t*>(output_s_packed)[out_idx * 4 + pos] = exponent;
+      reinterpret_cast<uint8_t*>(output_s_packed)[out_idx * 4 + pos] = exp_byte;
    } else if (out_idx < num_scale_elems) {
-      // write zero for padding groups if within bounds of output_s_packed
      reinterpret_cast<uint8_t*>(output_s_packed)[out_idx * 4 + pos] = 0;
    }
  }

-  __syncthreads();
-
-  if (is_valid_group) {
-    DST_DTYPE* group_output =
-        static_cast<DST_DTYPE*>(output_q) +
-        static_cast<int64_t>(mn_idx) * groups_per_row * group_size +
-        sf_k_idx * group_size;
-    QuantizeGroup<T, DST_DTYPE>(smem_group, group_output, group_size, lane_id,
-                                threads_per_group, y_s, min_8bit, max_8bit);
+  // For padded mn rows that fall within output_q's allocated extent, write
+  // a uint4 of zeros to keep the buffer clean for downstream TMA loads.
+  // Skip writes for sf_k padding (those positions don't exist in output_q).
+  if (!is_valid_group) {
+    if (sf_k_idx < groups_per_row && mn_idx >= mn &&
+        mn_idx < output_q_mn_extent) {
+      DST_DTYPE* group_output =
+          static_cast<DST_DTYPE*>(output_q) +
+          static_cast<int64_t>(mn_idx) * groups_per_row * GROUP_SIZE +
+          sf_k_idx * GROUP_SIZE + lane_id * VEC_SIZE;
+      *reinterpret_cast<uint4*>(group_output) = make_uint4(0, 0, 0, 0);
+    }
+    return;
  }
+
+  // Reconstruct y_s as a power-of-2 float and use its reciprocal.
+  float y_s_q = __uint_as_float(static_cast<uint32_t>(exp_byte) << 23);
+  float inv_y = 1.0f / y_s_q;
+
+  // Quantize and pack into 16 fp8/int8 bytes (= uint4). VEC_SIZE==16 so we
+  // fill four 32-bit words, four bytes each.
+  uint32_t packed_lo = 0;
+  uint32_t packed_lo_hi = 0;
+  uint32_t packed_hi_lo = 0;
+  uint32_t packed_hi = 0;
+#pragma unroll
+  for (int i = 0; i < VEC_SIZE; ++i) {
+    float q =
+        fminf(fmaxf(static_cast<float>(regs[i]) * inv_y, min_8bit), max_8bit);
+    DST_DTYPE qb = DST_DTYPE(q);
+    uint8_t byte = *reinterpret_cast<uint8_t*>(&qb);
+    const int shift = (i & 3) * 8;
+    if (i < 4) {
+      packed_lo |= static_cast<uint32_t>(byte) << shift;
+    } else if (i < 8) {
+      packed_lo_hi |= static_cast<uint32_t>(byte) << shift;
+    } else if (i < 12) {
+      packed_hi_lo |= static_cast<uint32_t>(byte) << shift;
+    } else {
+      packed_hi |= static_cast<uint32_t>(byte) << shift;
+    }
+  }
+
+  uint4 packed_out =
+      make_uint4(packed_lo, packed_lo_hi, packed_hi_lo, packed_hi);
+  DST_DTYPE* group_output =
+      static_cast<DST_DTYPE*>(output_q) +
+      static_cast<int64_t>(mn_idx) * groups_per_row * GROUP_SIZE +
+      sf_k_idx * GROUP_SIZE + lane_id * VEC_SIZE;
+  *reinterpret_cast<uint4*>(group_output) = packed_out;
 }

+// Public entry point: register-resident packed quant kernel.
+// Constraints: group_size == 128 and bf16/fp16 input.
 void per_token_group_quant_8bit_packed(const torch::stable::Tensor& input,
                                       torch::stable::Tensor& output_q,
                                       torch::stable::Tensor& output_s_packed,
                                       int64_t group_size, double eps,
                                       double min_8bit, double max_8bit) {
+  STD_TORCH_CHECK(group_size == 128,
+                  "per_token_group_quant_8bit_packed only supports "
+                  "group_size==128, got ",
+                  group_size, ".");
+  const auto in_dtype = input.scalar_type();
+  STD_TORCH_CHECK(
+      in_dtype == torch::headeronly::ScalarType::Half ||
+          in_dtype == torch::headeronly::ScalarType::BFloat16,
+      "per_token_group_quant_8bit_packed only supports bf16/fp16 input.");
+
  STD_TORCH_CHECK(input.is_contiguous());
  STD_TORCH_CHECK(output_q.is_contiguous());

  const int64_t k = input.size(-1);
-  STD_TORCH_CHECK(k % group_size == 0, "Last dimension (", k,
-                  ") must be divisible by group_size (", group_size, ").");
+  STD_TORCH_CHECK(k % group_size == 0, "input last dim k=", k,
+                  " is not divisible by group_size=", group_size, ".");

  const int64_t mn = input.numel() / k;
  const int64_t groups_per_row = k / group_size;
-
-  STD_TORCH_CHECK(output_s_packed.dim() == 2,
-                  "output_s_packed must be 2D, got dim=", output_s_packed.dim(),
-                  ".");
-
  const int64_t k_num_packed_sfk = (groups_per_row + 3) / 4;
  const int64_t tma_aligned_mn = ((mn + 3) / 4) * 4;

+  // output_q may be allocated with extra padded mn rows (e.g.,
+  // (tma_aligned_mn, k)) so the kernel can zero-fill them in-line and the
+  // caller can use torch.empty instead of torch.zeros. The grid only covers
+  // up to tma_aligned_mn, so we cap the extent there.
+  const int64_t output_q_mn_actual = output_q.numel() / k;
+  STD_TORCH_CHECK(output_q_mn_actual >= mn,
+                  "output_q must have at least mn rows; got ",
+                  output_q_mn_actual, " rows for mn=", mn, ".");
+  const int64_t output_q_mn_extent =
+      output_q_mn_actual < tma_aligned_mn ? output_q_mn_actual : tma_aligned_mn;
+
  STD_TORCH_CHECK(
      output_s_packed.scalar_type() == torch::headeronly::ScalarType::Int,
-      "output_s_packed must have dtype int32 for UE8M0-packed scales.");
-  // DeepGEMM expects SFA scales in MN-major form with shape
-  // [mn, ceil_div(K, 128 * 4)] and TMA-aligned stride on the last
-  // dimension.
+      "output_s_packed must be int32 for UE8M0-packed scales.");
  STD_TORCH_CHECK(output_s_packed.size(0) == mn &&
                      output_s_packed.size(1) == k_num_packed_sfk,
                  "output_s_packed shape must be [", mn, ", ", k_num_packed_sfk,
-                  "], but got [", output_s_packed.size(0), ", ",
+                  "]; got [", output_s_packed.size(0), ", ",
                  output_s_packed.size(1), "].");
-  // Verify column-major TMA-aligned layout
  STD_TORCH_CHECK(output_s_packed.stride(0) == 1 &&
                      output_s_packed.stride(1) == tma_aligned_mn,
-                  "output_s_packed must have strides [1, ", tma_aligned_mn,
-                  "], but got [", output_s_packed.stride(0), ", ",
+                  "output_s_packed strides must be [1, ", tma_aligned_mn,
+                  "]; got [", output_s_packed.stride(0), ", ",
                  output_s_packed.stride(1), "].");

  cudaStream_t stream = get_current_cuda_stream();

-  constexpr int THREADS_PER_GROUP = 16;
-
-  // Expand the grid to cover MN and K padding so every byte in
-  // output_s_packed is written (padding bytes get zeroed by the kernel).
+  constexpr int THREADS_PER_GROUP = 8;
  const int64_t padded_groups_per_row = k_num_packed_sfk * 4;
  const int64_t num_groups_padded = tma_aligned_mn * padded_groups_per_row;
-  // Number of elements in output_s_packed.
  const int64_t num_scale_elems = mn + (k_num_packed_sfk - 1) * tma_aligned_mn;
-
  const int groups_per_block = GetGroupsPerBlock(num_groups_padded);

  auto dst_type = output_q.scalar_type();
-  const int num_blocks = num_groups_padded / groups_per_block;
+  const int64_t num_blocks = num_groups_padded / groups_per_block;
  const int num_threads = groups_per_block * THREADS_PER_GROUP;
+  // CUDA caps grid.x at 2^31 - 1; this fits any realistic shape but guard
+  // against pathological inputs.
+  STD_TORCH_CHECK(num_blocks <= static_cast<int64_t>(INT32_MAX),
+                  "per_token_group_quant_8bit_packed grid too large: ",
+                  num_blocks, " blocks (max ", INT32_MAX, ").");

-#define LAUNCH_PACKED_KERNEL(T, DST_DTYPE)                                     \
-  do {                                                                         \
-    dim3 grid(num_blocks);                                                     \
-    dim3 block(num_threads);                                                   \
-    size_t smem_bytes =                                                        \
-        static_cast<size_t>(groups_per_block) * group_size * sizeof(T);        \
-    per_token_group_quant_8bit_packed_kernel<T, DST_DTYPE>                     \
-        <<<grid, block, smem_bytes, stream>>>(                                 \
-            static_cast<const T*>(input.data_ptr()), output_q.data_ptr(),      \
-            reinterpret_cast<unsigned int*>(output_s_packed.data_ptr()),       \
-            static_cast<int>(group_size), static_cast<int>(num_groups_padded), \
-            groups_per_block, static_cast<int>(padded_groups_per_row),         \
-            static_cast<int>(groups_per_row), static_cast<int>(mn),            \
-            static_cast<int>(tma_aligned_mn),                                  \
-            static_cast<int>(num_scale_elems), static_cast<float>(eps),        \
-            static_cast<float>(min_8bit), static_cast<float>(max_8bit));       \
+#define LAUNCH_REG_KERNEL(T, DST_DTYPE)                                   \
+  do {                                                                    \
+    dim3 grid(static_cast<unsigned int>(num_blocks));                     \
+    dim3 block(num_threads);                                              \
+    per_token_group_quant_8bit_packed_register_kernel<T, DST_DTYPE, 128>  \
+        <<<grid, block, 0, stream>>>(                                     \
+            static_cast<const T*>(input.data_ptr()), output_q.data_ptr(), \
+            reinterpret_cast<unsigned int*>(output_s_packed.data_ptr()),  \
+            num_groups_padded, groups_per_block,                          \
+            static_cast<int>(padded_groups_per_row),                      \
+            static_cast<int>(groups_per_row), static_cast<int>(mn),       \
+            static_cast<int>(output_q_mn_extent),                         \
+            static_cast<int>(tma_aligned_mn), num_scale_elems,            \
+            static_cast<float>(eps), static_cast<float>(min_8bit),        \
+            static_cast<float>(max_8bit));                                \
  } while (0)

-  VLLM_STABLE_DISPATCH_FLOATING_TYPES(
-      input.scalar_type(), "per_token_group_quant_8bit_packed", ([&] {
+  VLLM_STABLE_DISPATCH_HALF_TYPES(
+      input.scalar_type(), "per_token_group_quant_8bit_packed_register", ([&] {
        if (dst_type == torch::headeronly::ScalarType::Float8_e4m3fn) {
-          LAUNCH_PACKED_KERNEL(scalar_t, __nv_fp8_e4m3);
+          LAUNCH_REG_KERNEL(scalar_t, __nv_fp8_e4m3);
        } else if (dst_type == torch::headeronly::ScalarType::Char) {
-          LAUNCH_PACKED_KERNEL(scalar_t, int8_t);
+          LAUNCH_REG_KERNEL(scalar_t, int8_t);
        } else {
          STD_TORCH_CHECK(
              false,
@@ -397,7 +487,7 @@ void per_token_group_quant_8bit_packed(const torch::stable::Tensor& input,
        }
      }));

-#undef LAUNCH_PACKED_KERNEL
+#undef LAUNCH_REG_KERNEL
 }

 void per_token_group_quant_fp8(const torch::stable::Tensor& input,
@@ -8,3 +8,13 @@ void per_token_group_quant_8bit(const torch::stable::Tensor& input,
                                torch::stable::Tensor& output_s,
                                int64_t group_size, double eps, double min_8bit,
                                double max_8bit, bool scale_ue8m0 = false);
+
+// Public op: register-resident packed quant for the DeepGEMM Blackwell path.
+// Restricted to group_size == 128 and bf16/fp16 input; other configurations
+// raise STD_TORCH_CHECK. The legacy shared-memory fallback was removed because
+// no production caller (deep_gemm_moe / input_quant_fp8) uses other shapes.
+void per_token_group_quant_8bit_packed(const torch::stable::Tensor& input,
+                                       torch::stable::Tensor& output_q,
+                                       torch::stable::Tensor& output_s_packed,
+                                       int64_t group_size, double eps,
+                                       double min_8bit, double max_8bit);
@@ -67,10 +67,6 @@ void shuffle_rows(const torch::Tensor& input_tensor,
                  torch::Tensor& output_tensor);

 #ifndef USE_ROCM
-// cuBLAS bf16 x bf16 -> fp32 router GEMM (fallback for non-SM90 / batch > 16)
-torch::Tensor router_gemm_bf16_fp32(torch::Tensor const& input,
-                                    torch::Tensor const& weight);
-
 // DeepSeek V3 optimized router GEMM kernel for SM90+
 // Computes output = mat_a @ mat_b.T where:
 //   mat_a: [num_tokens, hidden_dim] in bf16
@@ -1,52 +0,0 @@
-// SPDX-License-Identifier: Apache-2.0
-// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-
-// bf16 x bf16 -> fp32 router GEMM via cuBLAS.
-// Uses CUBLAS_COMPUTE_32F so bf16 operands accumulate into fp32,
-// matching TRT-LLM's cuBLAS fallback behaviour in dsv3RouterGemmOp.
-
-#include <torch/all.h>
-#include <ATen/cuda/CUDAContext.h>
-#include <cublas_v2.h>
-
-// cuBLAS column-major math for row-major PyTorch tensors:
-//   weight[N,K]_row  lda=K  -> cuBLAS sees (K,N) col-major; CUBLAS_OP_T ->
-//   (N,K) input[M,K]_row   ldb=K  -> cuBLAS sees (K,M) col-major; CUBLAS_OP_N
-//   -> (K,M) out[M,N]_row     ldc=N  -> cuBLAS sees (N,M) col-major (written as
-//   output^T)
-// cuBLAS: C(N,M) = weight(N,K) @ input(K,M)  =>  C^T = output[M,N]
-// params: m=N, n=M, k=K, lda=K (weight), ldb=K (input), ldc=N (output)
-
-torch::Tensor router_gemm_bf16_fp32(torch::Tensor const& input,
-                                    torch::Tensor const& weight) {
-  TORCH_CHECK(input.dtype() == torch::kBFloat16,
-              "router_gemm_bf16_fp32: input must be bfloat16");
-  TORCH_CHECK(weight.dtype() == torch::kBFloat16,
-              "router_gemm_bf16_fp32: weight must be bfloat16");
-  TORCH_CHECK(input.dim() == 2 && weight.dim() == 2,
-              "router_gemm_bf16_fp32: input and weight must be 2-D");
-  TORCH_CHECK(input.size(1) == weight.size(1),
-              "router_gemm_bf16_fp32: inner dimensions must match");
-
-  int64_t const M = input.size(0);
-  int64_t const N = weight.size(0);
-  int64_t const K = input.size(1);
-
-  auto out = torch::empty({M, N}, input.options().dtype(torch::kFloat32));
-
-  cublasHandle_t handle = at::cuda::getCurrentCUDABlasHandle();
-  TORCH_CUDABLAS_CHECK(
-      cublasSetStream(handle, at::cuda::getCurrentCUDAStream()));
-
-  float const alpha = 1.0f;
-  float const beta = 0.0f;
-
-  TORCH_CUDABLAS_CHECK(cublasGemmEx(
-      handle, CUBLAS_OP_T, CUBLAS_OP_N, static_cast<int>(N),
-      static_cast<int>(M), static_cast<int>(K), &alpha, weight.data_ptr(),
-      CUDA_R_16BF, static_cast<int>(K), input.data_ptr(), CUDA_R_16BF,
-      static_cast<int>(K), &beta, out.data_ptr(), CUDA_R_32F,
-      static_cast<int>(N), CUBLAS_COMPUTE_32F, CUBLAS_GEMM_DEFAULT));
-
-  return out;
-}
@@ -133,10 +133,6 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, m) {
      "Tensor)");
  m.impl("grouped_topk", torch::kCUDA, &grouped_topk);

-  // cuBLAS bf16 x bf16 -> fp32 router GEMM (fallback for non-SM90 / batch > 16)
-  m.def("router_gemm_bf16_fp32(Tensor input, Tensor weight) -> Tensor");
-  m.impl("router_gemm_bf16_fp32", torch::kCUDA, &router_gemm_bf16_fp32);
-
  // DeepSeek V3 optimized router GEMM for SM90+
  m.def("dsv3_router_gemm(Tensor! output, Tensor mat_a, Tensor mat_b) -> ()");
  // conditionally compiled so impl registration is in source file
@@ -163,6 +163,8 @@ void rotary_embedding(torch::Tensor& positions, torch::Tensor& query,

 void silu_and_mul(torch::Tensor& out, torch::Tensor& input);

+void silu_and_mul_clamp(torch::Tensor& out, torch::Tensor& input, double limit);
+
 void silu_and_mul_quant(torch::Tensor& out, torch::Tensor& input,
                        torch::Tensor& scale);

@@ -887,27 +887,14 @@ __global__ void __launch_bounds__(kThreadsPerBlock, 2)
  uint32_t* shared_ordered =
      reinterpret_cast<uint32_t*>(smem_raw + kFixedSmemLarge);

-  // RadixRowState for multi-CTA cooperative radix
+  // RadixRowState for multi-CTA cooperative radix.
+  // Zero-initialization is done host-side via cudaMemsetAsync in topk.cu
+  // before launch — that gives a stream-ordered happens-before edge for all
+  // CTAs, which the previous in-kernel init (CTA-0 only + intra-CTA
+  // __syncthreads) did not provide and which manifested as a race against
+  // CTA-1+'s first red_release on arrival_counter.
  RadixRowState* state = &params.row_states[group_id];

-  // -- Initialize RadixRowState (only needed if large rows exist) --
-  if (params.max_seq_len > RADIX_THRESHOLD) {
-    if (cta_in_group == 0) {
-      for (uint32_t buf = 0; buf < 3; buf++) {
-        for (uint32_t i = tx; i < RADIX; i += kThreadsPerBlock) {
-          state->histogram[buf][i] = 0;
-        }
-      }
-      if (tx == 0) {
-        state->remaining_k = 0;
-        state->prefix = 0;
-        state->arrival_counter = 0;
-        state->output_counter = 0;
-      }
-    }
-    __syncthreads();
-  }
-
  int barrier_phase = 0;
  const uint32_t total_iters = (params.num_rows + num_groups - 1) / num_groups;

@@ -7,23 +7,23 @@

 namespace vllm {

-template <typename scalar_t, bool IS_NEOX>
+template <typename scalar_t, typename cache_t, bool IS_NEOX>
 inline __device__ void apply_token_rotary_embedding(
-    scalar_t* __restrict__ arr, const float* __restrict__ cos_ptr,
-    const float* __restrict__ sin_ptr, int rot_offset, int embed_dim,
+    scalar_t* __restrict__ arr, const cache_t* __restrict__ cos_ptr,
+    const cache_t* __restrict__ sin_ptr, int rot_offset, int embed_dim,
    const bool inverse) {
  int x_index, y_index;
  float cos_f, sin_f;
  if (IS_NEOX) {
    x_index = rot_offset;
    y_index = embed_dim + rot_offset;
-    cos_f = VLLM_LDG(cos_ptr + x_index);
-    sin_f = VLLM_LDG(sin_ptr + x_index);
+    cos_f = static_cast<float>(VLLM_LDG(cos_ptr + x_index));
+    sin_f = static_cast<float>(VLLM_LDG(sin_ptr + x_index));
  } else {
    x_index = 2 * rot_offset;
    y_index = 2 * rot_offset + 1;
-    cos_f = VLLM_LDG(cos_ptr + x_index / 2);
-    sin_f = VLLM_LDG(sin_ptr + x_index / 2);
+    cos_f = static_cast<float>(VLLM_LDG(cos_ptr + x_index / 2));
+    sin_f = static_cast<float>(VLLM_LDG(sin_ptr + x_index / 2));
  }
  if (inverse) {
    sin_f = -sin_f;
@@ -34,7 +34,7 @@ inline __device__ void apply_token_rotary_embedding(
  arr[y_index] = static_cast<scalar_t>(y_f * cos_f + x_f * sin_f);
 }

-template <typename scalar_t, bool IS_NEOX>
+template <typename scalar_t, typename cache_t, bool IS_NEOX>
 inline __device__ void apply_rotary_embedding(
    scalar_t* __restrict__ query,  // [batch_size, seq_len, num_heads,
                                   // head_size] or [num_tokens, num_heads,
@@ -43,14 +43,14 @@ inline __device__ void apply_rotary_embedding(
                                   // [batch_size, seq_len, num_kv_heads,
                                   // head_size] or [num_tokens, num_kv_heads,
                                   // head_size]
-    const float* cache_ptr, const int head_size, const int num_heads,
+    const cache_t* cache_ptr, const int head_size, const int num_heads,
    const int num_kv_heads, const int rot_dim, const int token_idx,
    const int64_t query_stride, const int64_t key_stride,
    const int64_t head_stride, const int64_t rope_dim_offset,
    const bool inverse) {
  const int embed_dim = rot_dim / 2;
-  const float* cos_ptr = cache_ptr;
-  const float* sin_ptr = cache_ptr + embed_dim;
+  const cache_t* cos_ptr = cache_ptr;
+  const cache_t* sin_ptr = cache_ptr + embed_dim;

  const int nq = num_heads * embed_dim;
  for (int i = threadIdx.x; i < nq; i += blockDim.x) {
@@ -58,7 +58,7 @@ inline __device__ void apply_rotary_embedding(
    const int64_t token_head =
        token_idx * query_stride + head_idx * head_stride + rope_dim_offset;
    const int rot_offset = i % embed_dim;
-    apply_token_rotary_embedding<scalar_t, IS_NEOX>(
+    apply_token_rotary_embedding<scalar_t, cache_t, IS_NEOX>(
        query + token_head, cos_ptr, sin_ptr, rot_offset, embed_dim, inverse);
  }

@@ -69,13 +69,13 @@ inline __device__ void apply_rotary_embedding(
      const int64_t token_head =
          token_idx * key_stride + head_idx * head_stride + rope_dim_offset;
      const int rot_offset = i % embed_dim;
-      apply_token_rotary_embedding<scalar_t, IS_NEOX>(
+      apply_token_rotary_embedding<scalar_t, cache_t, IS_NEOX>(
          key + token_head, cos_ptr, sin_ptr, rot_offset, embed_dim, inverse);
    }
  }
 }

-template <typename scalar_t, bool IS_NEOX>
+template <typename scalar_t, typename cache_t, bool IS_NEOX>
 __global__ void rotary_embedding_kernel(
    const int64_t* __restrict__ positions,  // [batch_size, seq_len] or
                                            // [num_tokens]
@@ -86,15 +86,15 @@ __global__ void rotary_embedding_kernel(
                                 // [batch_size, seq_len, num_kv_heads,
                                 // head_size] or [num_tokens, num_kv_heads,
                                 // head_size]
-    const float* __restrict__ cos_sin_cache,  // [max_position, rot_dim] fp32
+    const cache_t* __restrict__ cos_sin_cache,  // [max_position, rot_dim]
    const int rot_dim, const int64_t query_stride, const int64_t key_stride,
    const int64_t head_stride, const int num_heads, const int num_kv_heads,
    const int head_size, const int64_t rope_dim_offset, const bool inverse) {
  const int token_idx = blockIdx.x;
  int64_t pos = positions[token_idx];
-  const float* cache_ptr = cos_sin_cache + pos * rot_dim;
+  const cache_t* cache_ptr = cos_sin_cache + pos * rot_dim;

-  apply_rotary_embedding<scalar_t, IS_NEOX>(
+  apply_rotary_embedding<scalar_t, cache_t, IS_NEOX>(
      query, key, cache_ptr, head_size, num_heads, num_kv_heads, rot_dim,
      token_idx, query_stride, key_stride, head_stride, rope_dim_offset,
      inverse);
@@ -168,23 +168,28 @@ void rotary_embedding(
  dim3 block(std::min<int64_t>(num_heads * rot_dim / 2, 512));
  const at::cuda::OptionalCUDAGuard device_guard(device_of(query));
  const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
-  auto cache_f32 = cos_sin_cache.to(torch::kFloat32);
  VLLM_DISPATCH_FLOATING_TYPES(query.scalar_type(), "rotary_embedding", [&] {
-    if (is_neox) {
-      vllm::rotary_embedding_kernel<scalar_t, true><<<grid, block, 0, stream>>>(
-          positions.data_ptr<int64_t>(), query.data_ptr<scalar_t>(),
-          key.has_value() ? key->data_ptr<scalar_t>() : nullptr,
-          cache_f32.data_ptr<float>(), rot_dim, query_stride, key_stride,
-          head_stride, num_heads, num_kv_heads, head_size, rope_dim_offset,
-          inverse);
-    } else {
-      vllm::rotary_embedding_kernel<scalar_t, false>
-          <<<grid, block, 0, stream>>>(
-              positions.data_ptr<int64_t>(), query.data_ptr<scalar_t>(),
-              key.has_value() ? key->data_ptr<scalar_t>() : nullptr,
-              cache_f32.data_ptr<float>(), rot_dim, query_stride, key_stride,
-              head_stride, num_heads, num_kv_heads, head_size, rope_dim_offset,
-              inverse);
-    }
+    using query_t = scalar_t;
+    VLLM_DISPATCH_FLOATING_TYPES(
+        cos_sin_cache.scalar_type(), "rotary_embedding_cache", [&] {
+          using cache_t = scalar_t;
+          if (is_neox) {
+            vllm::rotary_embedding_kernel<query_t, cache_t, true>
+                <<<grid, block, 0, stream>>>(
+                    positions.data_ptr<int64_t>(), query.data_ptr<query_t>(),
+                    key.has_value() ? key->data_ptr<query_t>() : nullptr,
+                    cos_sin_cache.data_ptr<cache_t>(), rot_dim, query_stride,
+                    key_stride, head_stride, num_heads, num_kv_heads, head_size,
+                    rope_dim_offset, inverse);
+          } else {
+            vllm::rotary_embedding_kernel<query_t, cache_t, false>
+                <<<grid, block, 0, stream>>>(
+                    positions.data_ptr<int64_t>(), query.data_ptr<query_t>(),
+                    key.has_value() ? key->data_ptr<query_t>() : nullptr,
+                    cos_sin_cache.data_ptr<cache_t>(), rot_dim, query_stride,
+                    key_stride, head_stride, num_heads, num_kv_heads, head_size,
+                    rope_dim_offset, inverse);
+          }
+        });
  });
 }
@@ -82,22 +82,100 @@ void launch_persistent_topk(const torch::Tensor& logits,
    size_t smem_size = P::kFixedSmemLarge + chunk_size * sizeof(uint32_t);
    if (smem_size < P::kSmemMedium) smem_size = P::kSmemMedium;

+    // Query occupancy for the instantiation that will actually launch;
+    // overestimating it deadlocks the cooperative barrier.
    int occupancy = 1;
-    cudaOccupancyMaxActiveBlocksPerMultiprocessor(
-        &occupancy, P::persistent_topk_kernel<TopK, 4>, P::kThreadsPerBlock,
-        smem_size);
+    cudaError_t occ_err = cudaSuccess;
+    if (vec_size == 4) {
+      occ_err = cudaOccupancyMaxActiveBlocksPerMultiprocessor(
+          &occupancy, P::persistent_topk_kernel<TopK, 4>, P::kThreadsPerBlock,
+          smem_size);
+    } else if (vec_size == 2) {
+      occ_err = cudaOccupancyMaxActiveBlocksPerMultiprocessor(
+          &occupancy, P::persistent_topk_kernel<TopK, 2>, P::kThreadsPerBlock,
+          smem_size);
+    } else {
+      occ_err = cudaOccupancyMaxActiveBlocksPerMultiprocessor(
+          &occupancy, P::persistent_topk_kernel<TopK, 1>, P::kThreadsPerBlock,
+          smem_size);
+    }
+    TORCH_CHECK(occ_err == cudaSuccess,
+                "persistent_topk occupancy query failed: ",
+                cudaGetErrorString(occ_err));
    if (occupancy < 1) occupancy = 1;

-    uint32_t max_resident_ctas = static_cast<uint32_t>(num_sms) * occupancy;
+    // The cooperative spin-wait barrier only runs when at least one row hits
+    // the radix path (seq_len > RADIX_THRESHOLD). Below that, non-CTA-0 CTAs
+    // early-exit, so oversubscription can't deadlock and headroom is wasted.
+    const bool needs_cooperative =
+        static_cast<uint32_t>(max_seq_len) > P::RADIX_THRESHOLD;
+
+    const uint32_t hw_resident_cap =
+        static_cast<uint32_t>(num_sms) * static_cast<uint32_t>(occupancy);
+    uint32_t max_resident_ctas = hw_resident_cap;
+    if (needs_cooperative) {
+      // Reserve one CTA per SM when occupancy allows; fall back to a single
+      // CTA when occupancy == 1 (the most deadlock-prone case — any straggler
+      // kernel that takes the only slot on one SM hangs the barrier). Never
+      // drop below one full group's worth.
+      uint32_t headroom = (occupancy > 1) ? static_cast<uint32_t>(num_sms) : 1u;
+      if (max_resident_ctas >= headroom + ctas_per_group) {
+        max_resident_ctas -= headroom;
+      }
+    }
    uint32_t num_groups = std::min(max_resident_ctas / ctas_per_group,
                                   static_cast<uint32_t>(num_rows));
    if (num_groups == 0) num_groups = 1;
    uint32_t total_ctas = num_groups * ctas_per_group;

+    // If the cooperative launch wouldn't fit, fall back to FilteredTopK
+    // instead of deadlocking. Only relevant when needs_cooperative.
+    if (needs_cooperative && total_ctas > hw_resident_cap) {
+      TORCH_CHECK(max_smem_per_block >= 128 * 1024,
+                  "persistent_topk would oversubscribe and the FilteredTopK "
+                  "fallback requires >=128KB smem per block (have ",
+                  max_smem_per_block, "). total_ctas=", total_ctas,
+                  " > num_sms*occupancy=", hw_resident_cap, " (TopK=", TopK,
+                  ", vec_size=", vec_size, ", ctas_per_group=", ctas_per_group,
+                  ", smem=", smem_size, ").");
+      cudaError_t status =
+          vllm::FilteredTopKRaggedTransform<float, int32_t, TopK>(
+              logits.data_ptr<float>(), output.data_ptr<int32_t>(),
+              lengths.data_ptr<int32_t>(), static_cast<uint32_t>(num_rows),
+              static_cast<uint32_t>(TopK), static_cast<uint32_t>(stride),
+              stream);
+      TORCH_CHECK(status == cudaSuccess,
+                  "FilteredTopK fallback failed: ", cudaGetErrorString(status));
+      return;
+    }
+
    size_t state_bytes = num_groups * sizeof(P::RadixRowState);
    TORCH_CHECK(workspace.size(0) >= static_cast<int64_t>(state_bytes),
                "workspace too small, need ", state_bytes, " bytes");

+    // Zero the per-group RadixRowState region before launch — only when the
+    // radix path will actually run (max_seq_len > RADIX_THRESHOLD). The
+    // RadixRowState fields (arrival_counter, histograms) are only touched by
+    // radix_topk; the decode/medium paths inside the persistent kernel
+    // operate purely in shared memory and never read these globals, so a
+    // stale workspace is harmless for them.
+    //
+    // Why we need the memset (when needs_cooperative is true):
+    //   1. arrival_counter accumulates within a launch and is never reset,
+    //      so a prior call leaves it at a large positive value. Without this
+    //      reset, the very first wait_ge in the next call sees counter >>
+    //      target and returns instantly, breaking the barrier.
+    //   2. The previous in-kernel init only ran in CTA-0 with intra-CTA
+    //      __syncthreads(), so it had no happens-before edge to CTA-1+'s
+    //      first red_release. cudaMemsetAsync is stream-ordered: the zero
+    //      is globally visible before any CTA runs.
+    if (needs_cooperative) {
+      cudaError_t mz_err = cudaMemsetAsync(workspace.data_ptr<uint8_t>(), 0,
+                                           state_bytes, stream);
+      TORCH_CHECK(mz_err == cudaSuccess,
+                  "row_states memset failed: ", cudaGetErrorString(mz_err));
+    }
+
    P::PersistentTopKParams params;
    params.input = logits.data_ptr<float>();
    params.output = output.data_ptr<int32_t>();
@@ -106,6 +106,12 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
  ops.def("silu_and_mul(Tensor! result, Tensor input) -> ()");
  ops.impl("silu_and_mul", torch::kCUDA, &silu_and_mul);

+  // SwiGLU activation with input clamping.
+  ops.def(
+      "silu_and_mul_with_clamp(Tensor! result, Tensor input, float limit) "
+      "-> ()");
+  ops.impl("silu_and_mul_with_clamp", torch::kCUDA, &silu_and_mul_clamp);
+
  ops.def(
      "silu_and_mul_quant(Tensor! result, Tensor input, Tensor scale) -> ()");
  ops.impl("silu_and_mul_quant", torch::kCUDA, &silu_and_mul_quant);
@@ -41,6 +41,13 @@ ARG BUILD_BASE_IMAGE=nvidia/cuda:${CUDA_VERSION}-devel-ubuntu22.04
 # Using cuda base image with minimal dependencies necessary for JIT compilation (FlashInfer, DeepGEMM, EP kernels)
 ARG FINAL_BASE_IMAGE=nvidia/cuda:${CUDA_VERSION}-base-ubuntu${UBUNTU_VERSION}

+# OS family of BUILD_BASE_IMAGE. Controls package manager (apt vs dnf) and
+# Python bootstrap. Set to "manylinux" alongside a manylinux build base such
+# as pytorch/manylinux2_28-builder:cuda13.0 to produce wheels with a glibc
+# 2.28 floor (matches PyTorch's own published wheels). Default stays on
+# Ubuntu for backwards compatibility.
+ARG BUILD_OS=ubuntu
+
 # By parameterizing the Deadsnakes repository URL, we allow third-party to use
 # their own mirror. When doing so, we don't benefit from the transparent
 # installation of the GPG key of the PPA, as done by add-apt-repository, so we
@@ -94,35 +101,64 @@ FROM ${BUILD_BASE_IMAGE} AS base

 ARG CUDA_VERSION
 ARG PYTHON_VERSION
+ARG BUILD_OS

 ENV DEBIAN_FRONTEND=noninteractive

-# Install system dependencies including build tools
-RUN apt-get update -y \
-    && apt-get install -y --no-install-recommends \
-        ccache \
-        software-properties-common \
-        git \
-        curl \
-        sudo \
-        python3-pip \
-        libibverbs-dev \
-        # Upgrade to GCC 10 to avoid https://gcc.gnu.org/bugzilla/show_bug.cgi?id=92519
-        # as it was causing spam when compiling the CUTLASS kernels
-        gcc-10 \
-        g++-10 \
-    && update-alternatives --install /usr/bin/gcc gcc /usr/bin/gcc-10 110 --slave /usr/bin/g++ g++ /usr/bin/g++-10 \
-    # Install python dev headers if available (needed for cmake FindPython on Ubuntu 24.04
-    # which ships cmake 3.28 and requires Development.SABIModule; silently skipped on
-    # Ubuntu 20.04/22.04 where python3.x-dev is not available without a PPA)
-    && (apt-get install -y --no-install-recommends python${PYTHON_VERSION}-dev 2>/dev/null || true) \
-    && rm -rf /var/lib/apt/lists/* \
-    && curl -LsSf https://astral.sh/uv/install.sh | sh \
-    && $HOME/.local/bin/uv venv /opt/venv --python ${PYTHON_VERSION} \
+# Install system dependencies including build tools.
+# The Ubuntu path uses apt + deadsnakes-via-uv for Python; the manylinux path
+# (AlmaLinux 8, e.g. pytorch/manylinux2_28-builder) uses dnf and the Python
+# interpreters pre-installed at /opt/python/cpXY-cpXY/.
+RUN if [ "${BUILD_OS}" = "manylinux" ]; then \
+        # rdma-core-devel provides libibverbs headers; ccache lives in EPEL,
+        # which the pytorch manylinux image already enables. git/curl/sudo
+        # are typically pre-installed but listed defensively.
+        dnf install -y --setopt=install_weak_deps=False \
+            ccache \
+            git \
+            curl \
+            sudo \
+            rdma-core-devel \
+        && dnf clean all \
+        && rm -rf /var/cache/dnf; \
+    else \
+        apt-get update -y \
+        && apt-get install -y --no-install-recommends \
+            ccache \
+            software-properties-common \
+            git \
+            curl \
+            sudo \
+            python3-pip \
+            libibverbs-dev \
+            # Upgrade to GCC 10 to avoid https://gcc.gnu.org/bugzilla/show_bug.cgi?id=92519
+            # as it was causing spam when compiling the CUTLASS kernels
+            gcc-10 \
+            g++-10 \
+        && update-alternatives --install /usr/bin/gcc gcc /usr/bin/gcc-10 110 --slave /usr/bin/g++ g++ /usr/bin/g++-10 \
+        # Install python dev headers if available (needed for cmake FindPython on Ubuntu 24.04
+        # which ships cmake 3.28 and requires Development.SABIModule; silently skipped on
+        # Ubuntu 20.04/22.04 where python3.x-dev is not available without a PPA)
+        && (apt-get install -y --no-install-recommends python${PYTHON_VERSION}-dev 2>/dev/null || true) \
+        && rm -rf /var/lib/apt/lists/*; \
+    fi
+
+# Install uv and bootstrap /opt/venv. Both paths converge on /opt/venv so all
+# downstream stages stay distro-agnostic.
+RUN curl -LsSf https://astral.sh/uv/install.sh | sh \
+    && if [ "${BUILD_OS}" = "manylinux" ]; then \
+           # manylinux images ship Python at /opt/python/cpXY-cpXY/; point uv
+           # at the matching interpreter rather than letting it fetch one.
+           PYV_NODOT=$(echo ${PYTHON_VERSION} | tr -d '.') \
+           && MANYLINUX_PY=/opt/python/cp${PYV_NODOT}-cp${PYV_NODOT}/bin/python${PYTHON_VERSION} \
+           && $HOME/.local/bin/uv venv /opt/venv --python "$MANYLINUX_PY"; \
+       else \
+           $HOME/.local/bin/uv venv /opt/venv --python ${PYTHON_VERSION}; \
+       fi \
    && rm -f /usr/bin/python3 /usr/bin/python3-config /usr/bin/pip \
-    && ln -s /opt/venv/bin/python3 /usr/bin/python3 \
-    && ln -s /opt/venv/bin/python3-config /usr/bin/python3-config \
-    && ln -s /opt/venv/bin/pip /usr/bin/pip \
+    && ln -sf /opt/venv/bin/python3 /usr/bin/python3 \
+    && ln -sf /opt/venv/bin/python3-config /usr/bin/python3-config \
+    && ln -sf /opt/venv/bin/pip /usr/bin/pip \
    && python3 --version && python3 -m pip --version

 # Activate virtual environment and add uv to PATH
@@ -433,6 +469,7 @@ FROM base AS dev
 ARG PIP_INDEX_URL UV_INDEX_URL
 ARG PIP_EXTRA_INDEX_URL UV_EXTRA_INDEX_URL
 ARG PYTORCH_CUDA_INDEX_BASE_URL
+ARG BUILD_OS

 # This timeout (in seconds) is necessary when installing some dependencies via uv since it's likely to time out
 # Reference: https://github.com/astral-sh/uv/pull/1694
@@ -442,7 +479,11 @@ ENV UV_INDEX_STRATEGY="unsafe-best-match"
 ENV UV_LINK_MODE=copy

 # Install libnuma-dev, required by fastsafetensors (fixes #20384)
-RUN apt-get update && apt-get install -y --no-install-recommends libnuma-dev && rm -rf /var/lib/apt/lists/*
+RUN if [ "${BUILD_OS}" = "manylinux" ]; then \
+        dnf install -y numactl-devel && dnf clean all && rm -rf /var/cache/dnf; \
+    else \
+        apt-get update && apt-get install -y --no-install-recommends libnuma-dev && rm -rf /var/lib/apt/lists/*; \
+    fi


 # We can specify the standard or nightly build of PyTorch
@@ -538,9 +579,11 @@ RUN CUDA_VERSION_DASH=$(echo $CUDA_VERSION | cut -d. -f1,2 | tr '.' '-') && \
        cuda-nvrtc-${CUDA_VERSION_DASH} \
        cuda-cuobjdump-${CUDA_VERSION_DASH} \
        libcurand-dev-${CUDA_VERSION_DASH} \
-        libcublas-${CUDA_VERSION_DASH} \
+        libcublas-dev-${CUDA_VERSION_DASH} \
        # Required by fastsafetensors (fixes #20384)
-        libnuma-dev && \
+        libnuma-dev \
+        # numactl CLI for NUMA binding at runtime
+        numactl && \
    # Fixes nccl_allocator requiring nccl.h at runtime
    # https://github.com/vllm-project/vllm/blob/1336a1ea244fa8bfd7e72751cabbdb5b68a0c11a/vllm/distributed/device_communicators/pynccl_allocator.py#L22
    # NCCL packages don't use the cuda-MAJOR-MINOR naming convention,
@@ -583,9 +626,7 @@ RUN --mount=type=cache,target=/root/.cache/uv \
 ARG FLASHINFER_VERSION=0.6.8.post1
 RUN --mount=type=cache,target=/root/.cache/uv \
    uv pip install --system flashinfer-jit-cache==${FLASHINFER_VERSION} \
-        --extra-index-url https://flashinfer.ai/whl/cu$(echo $CUDA_VERSION | cut -d. -f1,2 | tr -d '.') \
-    && flashinfer show-config \
-    && flashinfer download-cubin
+        --extra-index-url https://flashinfer.ai/whl/cu$(echo $CUDA_VERSION | cut -d. -f1,2 | tr -d '.')

 # ============================================================
 # OPENAI API SERVER DEPENDENCIES
@@ -667,6 +708,13 @@ RUN --mount=type=bind,from=build,src=/tmp/ep_kernels_workspace/dist,target=/vllm
    uv pip install --system ep_kernels/dist/*.whl --verbose \
        --extra-index-url ${PYTORCH_CUDA_INDEX_BASE_URL}/cu$(echo $CUDA_VERSION | cut -d. -f1,2 | tr -d '.')

+# Download FlashInfer precompiled cubins AFTER all pip installs are done.
+# This must run after the vLLM wheel and EP kernels installs above, because
+# those can reinstall/touch flashinfer packages. Downloading cubins earlier
+# (in the flashinfer-jit-cache layer) causes ~2.5 GB of layer duplication
+# when a later pip install overwrites flashinfer package files.
+RUN flashinfer show-config && flashinfer download-cubin
+
 # CUDA image changed from /usr/local/nvidia to /usr/local/cuda in 12.8 but will
 # return to /usr/local/nvidia in 13.0 to allow container providers to mount drivers
 # consistently from the host (see https://github.com/vllm-project/vllm/issues/18859).
@@ -756,6 +804,10 @@ FROM vllm-base AS vllm-openai-base
 ARG TARGETPLATFORM
 ARG INSTALL_KV_CONNECTORS=false
 ARG CUDA_VERSION
+ARG VLLM_BUILD_COMMIT
+ARG VLLM_BUILD_PIPELINE
+ARG VLLM_BUILD_URL
+ARG VLLM_IMAGE_TAG

 ARG PIP_INDEX_URL UV_INDEX_URL
 ARG PIP_EXTRA_INDEX_URL UV_EXTRA_INDEX_URL
@@ -792,6 +844,18 @@ RUN --mount=type=cache,target=/root/.cache/uv \
    fi

 ENV VLLM_USAGE_SOURCE production-docker-image
+ENV VLLM_BUILD_COMMIT=${VLLM_BUILD_COMMIT:-unknown} \
+    VLLM_BUILD_PIPELINE=${VLLM_BUILD_PIPELINE:-local} \
+    VLLM_BUILD_URL=${VLLM_BUILD_URL:-} \
+    VLLM_IMAGE_TAG=${VLLM_IMAGE_TAG:-local/vllm-openai:dev}
+LABEL org.opencontainers.image.source="https://github.com/vllm-project/vllm" \
+      org.opencontainers.image.revision="${VLLM_BUILD_COMMIT}" \
+      org.opencontainers.image.version="${VLLM_IMAGE_TAG}" \
+      org.opencontainers.image.url="${VLLM_BUILD_URL}" \
+      ai.vllm.build.commit="${VLLM_BUILD_COMMIT}" \
+      ai.vllm.build.pipeline="${VLLM_BUILD_PIPELINE}" \
+      ai.vllm.build.url="${VLLM_BUILD_URL}" \
+      ai.vllm.image.tag="${VLLM_IMAGE_TAG}"

 # define sagemaker first, so it is not default from `docker build`
 FROM vllm-openai-base AS vllm-sagemaker
@@ -192,6 +192,7 @@ ADD ./tests/ ./tests/
 ADD ./examples/ ./examples/
 ADD ./benchmarks/ ./benchmarks/
 ADD ./vllm/collect_env.py .
+ADD ./docker/ ./docker/
 ADD ./.buildkite/ ./.buildkite/

 # install development dependencies (for testing)
@@ -124,9 +124,9 @@ COPY --from=build_vllm ${COMMON_WORKDIR}/vllm/vllm/v1 /vllm_v1

 # RIXL/UCX build stages
 FROM base AS build_rixl
-ARG RIXL_BRANCH="bf4a7214"
+ARG RIXL_BRANCH="39be1de8"
 ARG RIXL_REPO="https://github.com/ROCm/RIXL.git"
-ARG UCX_BRANCH="7009d7a1"
+ARG UCX_BRANCH="bfb51733"
 ARG UCX_REPO="https://github.com/openucx/ucx.git"
 ENV ROCM_PATH=/opt/rocm
 ENV UCX_HOME=/usr/local/ucx
@@ -192,6 +192,7 @@ RUN cd /opt/rixl && \
    sed -i "s/--exclude 'libamdhip64\*'/--exclude 'libamdhip64*' --exclude 'libcore*' --exclude 'libpull*'/" \
        contrib/build-wheel.sh && \
    mkdir -p /app/install && \
+    _ucx_install_dir=${UCX_HOME} \
    ./contrib/build-wheel.sh \
        --output-dir /app/install \
        --rocm-dir ${ROCM_PATH} \
@@ -200,9 +201,9 @@ RUN cd /opt/rixl && \

 # DeepEP build stage
 FROM base AS build_deep
-ARG ROCSHMEM_BRANCH="ba0bf0f3"
+ARG ROCSHMEM_BRANCH="f0acb0c6"
 ARG ROCSHMEM_REPO="https://github.com/ROCm/rocm-systems.git"
-ARG DEEPEP_BRANCH="5d90af8b"
+ARG DEEPEP_BRANCH="a9ea9774"
 ARG DEEPEP_REPO="https://github.com/ROCm/DeepEP.git"
 ARG DEEPEP_NIC="cx7"
 ARG DEEPEP_ROCM_ARCH="gfx942;gfx950"
@@ -213,18 +214,15 @@ RUN git clone ${ROCSHMEM_REPO} \
 && git checkout ${ROCSHMEM_BRANCH} \
 && mkdir -p projects/rocshmem/build \
 && cd projects/rocshmem/build \
- && bash ../scripts/build_configs/all_backends \
-      -DCMAKE_INSTALL_PREFIX="${ROCSHMEM_DIR}" \
-      -DROCM_PATH=/opt/rocm \
-      -DGPU_TARGETS="${DEEPEP_ROCM_ARCH}" \
-      -DUSE_EXTERNAL_MPI=OFF
+ && INSTALL_PREFIX=${ROCSHMEM_DIR} \
+    ../scripts/build_configs/all_backends -DUSE_EXTERNAL_MPI=OFF

 # Build DeepEP wheel.
 # DeepEP looks for rocshmem at ROCSHMEM_DIR.
 RUN git clone ${DEEPEP_REPO} \
 && cd DeepEP \
 && git checkout ${DEEPEP_BRANCH} \
- && python3 setup.py --variant rocm --nic ${DEEPEP_NIC} bdist_wheel --dist-dir=/app/deep_install
+ && python3 setup.py --variant rocm --rocm-explicit-ctx --nic ${DEEPEP_NIC} bdist_wheel --dist-dir=/app/deep_install

 # MoRI runtime dependencies live in Dockerfile.rocm so NIC backend changes do
 # not force users to rebuild the long-lived Dockerfile.rocm_base image.
@@ -388,6 +386,16 @@ RUN --mount=type=bind,from=export_vllm,src=/,target=/install \
 # above are not available once that RUN step completes.
 COPY --from=export_vllm /*.whl /opt/vllm-wheels/

+# Update rdma-core to support latest rocshmem
+ARG DEEPEP_NIC
+RUN if [ "${DEEPEP_NIC}" = "cx7" ] || [ "${DEEPEP_NIC}" = "io" ]; then \
+    git clone --branch v62.0 --depth 1 https://github.com/linux-rdma/rdma-core.git /tmp/rdma-core && \
+    cd /tmp/rdma-core && \
+    mkdir -p build && cd build && \
+    cmake -GNinja -DCMAKE_INSTALL_PREFIX=/usr -DNO_MAN_PAGES=1 .. && \
+    ninja &&     ninja install &&     ldconfig &&     rm -rf /tmp/rdma-core; \
+fi
+
 # Install RIXL wheel
 RUN --mount=type=bind,from=build_rixl,src=/app/install,target=/rixl_install \
    uv pip install --system /rixl_install/*.whl
@@ -1,4 +1,4 @@
-ARG BASE_IMAGE=rocm/dev-ubuntu-22.04:7.2.1-complete
+ARG BASE_IMAGE=rocm/dev-ubuntu-22.04:7.2.2-complete
 ARG TRITON_BRANCH="ba5c1517"
 ARG TRITON_REPO="https://github.com/ROCm/triton.git"
 ARG PYTORCH_BRANCH="8514f051" # release/2.10 as of 3/17
@@ -9,7 +9,7 @@ ARG PYTORCH_AUDIO_BRANCH="v2.9.0"
 ARG PYTORCH_AUDIO_REPO="https://github.com/pytorch/audio.git"
 ARG FA_BRANCH="0e60e394"
 ARG FA_REPO="https://github.com/Dao-AILab/flash-attention.git"
-ARG AITER_BRANCH="v0.1.10.post3"
+ARG AITER_BRANCH="v0.1.12.post2"
 ARG AITER_REPO="https://github.com/ROCm/aiter.git"
 ARG MORI_BRANCH="v1.1.0"
 ARG MORI_REPO="https://github.com/ROCm/mori.git"
@@ -104,6 +104,28 @@ ENV SCCACHE_REGION=${USE_SCCACHE:+${SCCACHE_REGION_NAME}}
 ENV SCCACHE_S3_NO_CREDENTIALS=${USE_SCCACHE:+${SCCACHE_S3_NO_CREDENTIALS}}
 ENV SCCACHE_IDLE_TIMEOUT=${USE_SCCACHE:+0}

+# torch profiler hotfix for 7.2.2: rebuild CLR with https://github.com/ROCm/rocm-systems/pull/5062
+# will be removed once we move to ROCm 7.2.3
+RUN apt-get update && apt-get install -y rocm-llvm-dev
+RUN pip install CppHeaderParser
+RUN git clone --no-checkout --filter=blob:none https://github.com/ROCm/rocm-systems /tmp/rocm-systems \
+    && cd /tmp/rocm-systems \
+    && git sparse-checkout init --cone \
+    && git sparse-checkout set projects/hip projects/clr \
+    && git checkout 35e8c7bf8911862e5389509800e65fdf125412b3 \
+    && export CLR_DIR=/tmp/rocm-systems/projects/clr \
+    && export HIP_DIR=/tmp/rocm-systems/projects/hip \
+    && mkdir -p $CLR_DIR/build && cd $CLR_DIR/build \
+    && cmake \
+        -DHIP_COMMON_DIR=$HIP_DIR \
+        -DCMAKE_PREFIX_PATH="/opt/rocm/" \
+        -DCLR_BUILD_HIP=ON \
+        -DCLR_BUILD_OCL=OFF \
+        -DHIP_PLATFORM=amd \
+        .. \
+    && make -j$(nproc) \
+    && make install \
+    && rm -rf /tmp/rocm-systems

 ###
 ### Triton Build
@@ -153,8 +175,6 @@ RUN git clone ${PYTORCH_REPO} pytorch
 RUN cd pytorch && git checkout ${PYTORCH_BRANCH}
 RUN cd pytorch \
    && pip install -r requirements.txt && git submodule update --init --recursive
-RUN cd pytorch/third_party/kineto \
-    && git remote add rocm https://github.com/ROCm/kineto && git fetch rocm && git checkout 2d73be3 
 RUN cd pytorch && python3 tools/amd_build/build_amd.py \
    && if [ "$USE_SCCACHE" = "1" ]; then \
           export HIP_CLANG_PATH=/opt/sccache-wrappers \
@@ -5,9 +5,6 @@ WORKDIR /workspace/
 ARG PYTHON_VERSION=3.12
 ARG PIP_EXTRA_INDEX_URL="https://download.pytorch.org/whl/xpu"

-RUN wget -O- https://apt.repos.intel.com/intel-gpg-keys/GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB | gpg --dearmor | tee /usr/share/keyrings/oneapi-archive-keyring.gpg > /dev/null && \
-    echo "deb [signed-by=/usr/share/keyrings/oneapi-archive-keyring.gpg] https://apt.repos.intel.com/oneapi all main" | tee /etc/apt/sources.list.d/oneAPI.list
-
 RUN apt clean && apt-get update -y && \
    apt-get install -y --no-install-recommends --fix-missing \
    curl \
@@ -26,8 +23,20 @@ RUN apt clean && apt-get update -y && \
    python3.12-dev \
    python3-pip

-RUN apt update && apt upgrade -y && \
-    apt install -y intel-oneapi-compiler-dpcpp-cpp-2025.3
+# Add oneAPI repo, pin oneAPI to 2025.3, then install pinned packages in one layer.
+RUN wget -O- https://apt.repos.intel.com/intel-gpg-keys/GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB | gpg --dearmor | tee /usr/share/keyrings/oneapi-archive-keyring.gpg > /dev/null && \
+    echo "deb [signed-by=/usr/share/keyrings/oneapi-archive-keyring.gpg] https://apt.repos.intel.com/oneapi all main" | tee /etc/apt/sources.list.d/oneAPI.list && \
+    printf '%s\n' \
+    'Package: intel-oneapi-* intel-deep-learning-essentials* intel-pti*' \
+    'Pin: version 2025.3*' \
+    'Pin-Priority: 1001' \
+    > /etc/apt/preferences.d/oneapi-2025.3.pref && \
+    apt-get update -y && \
+    apt-get install -y --no-install-recommends \
+    intel-oneapi-compiler-dpcpp-cpp-2025.3 \
+    intel-oneapi-mkl-devel-2025.3 \
+    intel-oneapi-dnnl-devel-2025.3 && \
+    rm -rf /var/lib/apt/lists/*

 # Install UMD
 RUN mkdir neo && \
@@ -27,6 +27,22 @@ variable "COMMIT" {
  default = ""
 }

+variable "VLLM_BUILD_COMMIT" {
+  default = "unknown"
+}
+
+variable "VLLM_BUILD_PIPELINE" {
+  default = "local"
+}
+
+variable "VLLM_BUILD_URL" {
+  default = ""
+}
+
+variable "VLLM_IMAGE_TAG" {
+  default = "local/vllm-openai:dev"
+}
+
 # Groups

 group "default" {
@@ -46,6 +62,10 @@ target "_common" {
    max_jobs             = MAX_JOBS
    nvcc_threads         = NVCC_THREADS
    torch_cuda_arch_list = TORCH_CUDA_ARCH_LIST
+    VLLM_BUILD_COMMIT    = VLLM_BUILD_COMMIT != "unknown" ? VLLM_BUILD_COMMIT : (COMMIT != "" ? COMMIT : "unknown")
+    VLLM_BUILD_PIPELINE  = VLLM_BUILD_PIPELINE
+    VLLM_BUILD_URL       = VLLM_BUILD_URL
+    VLLM_IMAGE_TAG       = VLLM_IMAGE_TAG
  }
 }

@@ -56,10 +76,16 @@ target "_labels" {
    "org.opencontainers.image.title"       = "vLLM"
    "org.opencontainers.image.description" = "vLLM: A high-throughput and memory-efficient inference and serving engine for LLMs"
    "org.opencontainers.image.licenses"    = "Apache-2.0"
-    "org.opencontainers.image.revision"    = COMMIT
+    "org.opencontainers.image.revision"    = VLLM_BUILD_COMMIT != "unknown" ? VLLM_BUILD_COMMIT : (COMMIT != "" ? COMMIT : "unknown")
+    "org.opencontainers.image.version"     = VLLM_IMAGE_TAG
+    "org.opencontainers.image.url"         = VLLM_BUILD_URL
+    "ai.vllm.build.commit"                 = VLLM_BUILD_COMMIT != "unknown" ? VLLM_BUILD_COMMIT : (COMMIT != "" ? COMMIT : "unknown")
+    "ai.vllm.build.pipeline"               = VLLM_BUILD_PIPELINE
+    "ai.vllm.build.url"                    = VLLM_BUILD_URL
+    "ai.vllm.image.tag"                    = VLLM_IMAGE_TAG
  }
  annotations = [
-      "index,manifest:org.opencontainers.image.revision=${COMMIT}",
+    "index,manifest:org.opencontainers.image.revision=${VLLM_BUILD_COMMIT != "unknown" ? VLLM_BUILD_COMMIT : (COMMIT != "" ? COMMIT : "unknown")}",
  ]
 }

@@ -16,6 +16,9 @@
    "FINAL_BASE_IMAGE": {
      "default": "nvidia/cuda:13.0.2-base-ubuntu22.04"
    },
+    "BUILD_OS": {
+      "default": "ubuntu"
+    },
    "GET_PIP_URL": {
      "default": "https://bootstrap.pypa.io/get-pip.py"
    },
@@ -163,7 +163,7 @@ Running with a local file:

 ```bash
 vllm run-batch \
-    -i offline_inference/openai_batch/openai_example_batch.jsonl \
+    -i features/openai_batch/openai_example_batch.jsonl \
    -o results.jsonl \
    --model meta-llama/Meta-Llama-3-8B-Instruct
 ```
@@ -172,7 +172,7 @@ Using remote file:

 ```bash
 vllm run-batch \
-    -i https://raw.githubusercontent.com/vllm-project/vllm/main/examples/offline_inference/openai_batch/openai_example_batch.jsonl \
+    -i https://raw.githubusercontent.com/vllm-project/vllm/main/examples/features/openai_batch/openai_example_batch.jsonl \
    -o results.jsonl \
    --model meta-llama/Meta-Llama-3-8B-Instruct
 ```
@@ -23,7 +23,7 @@ llm = LLM(model="ibm-granite/granite-3.1-8b-instruct", tensor_parallel_size=2)
 !!! note
    With tensor parallelism enabled, each process will read the whole model and split it into chunks, which makes the disk reading time even longer (proportional to the size of tensor parallelism).

-    You can convert the model checkpoint to a sharded checkpoint using [examples/offline_inference/save_sharded_state.py](../../examples/offline_inference/save_sharded_state.py). The conversion process might take some time, but later you can load the sharded checkpoint much faster. The model loading time should remain constant regardless of the size of tensor parallelism.
+    You can convert the model checkpoint to a sharded checkpoint using [examples/features/sharded_state/load_sharded_state_offline.py](../../examples/features/sharded_state/load_sharded_state_offline.py). The conversion process might take some time, but later you can load the sharded checkpoint much faster. The model loading time should remain constant regardless of the size of tensor parallelism.

 ## Quantization

@@ -60,9 +60,19 @@ the failure?

 ## Logs Wrangling

-Download the full log file from Buildkite locally.
+Download a job's log (no Buildkite login required):

-Strip timestamps and colorization:
+[.buildkite/scripts/ci-fetch-log.sh](../../../.buildkite/scripts/ci-fetch-log.sh)
+
+```bash
+# Find the failing job. Each row's URL is .../builds/<N>#<job_uuid>:
+gh pr checks <PR> --repo vllm-project/vllm
+
+# Download + strip timestamps/ANSI in one step:
+.buildkite/scripts/ci-fetch-log.sh "https://buildkite.com/vllm/ci/builds/<N>#<job_uuid>"
+```
+
+To clean an already-downloaded log:

 [.buildkite/scripts/ci-clean-log.sh](../../../.buildkite/scripts/ci-clean-log.sh)

@@ -42,7 +42,7 @@ Traces can be visualized using <https://ui.perfetto.dev/>.

 #### Offline Inference

-Refer to [examples/offline_inference/simple_profiling.py](../../examples/offline_inference/simple_profiling.py) for an example.
+Refer to [examples/features/profiling/simple_profiling_offline.py](../../examples/features/profiling/simple_profiling_offline.py) for an example.

 #### OpenAI Server

@@ -155,6 +155,7 @@ Priority is **1 = highest** (tried first).
 | **Block Sizes** | Supported KV cache block sizes (%N means multiples of N) |
 | **Head Sizes** | Supported attention head sizes |
 | **Sink** | Attention sink support (for StreamingLLM) |
+| **Non-Causal** | Non-causal (bidirectional) attention support for decoder models |
 | **Sparse** | Sparse attention support (MLA only) |
 | **MM Prefix** | Multimodal prefix full attention support |
 | **DCP** | Decode Context Parallelism support (`--decode-context-parallel-size`) |
@@ -165,22 +166,22 @@ Priority is **1 = highest** (tried first).

 ## Standard Attention (MHA, MQA, GQA) Backends

-| Backend | Version | Dtypes | KV Dtypes | Block Sizes | Head Sizes | Sink | MM Prefix | DCP | Attention Types | Compute Cap. |
-| ------- | ------- | ------ | --------- | ----------- | ---------- | ---- | --------- | --- | --------------- | ------------ |
-| `CPU_ATTN` | | fp16, bf16, fp32 | `auto` | Any | 32, 64, 80, 96, 112, 128, 160, 192, 224, 256, 512 | ❌ | ❌ | ❌ | All | N/A |
-| `FLASHINFER` | Native† | fp16, bf16 | `auto`, `float16`, `bfloat16`, `fp8`, `fp8_e4m3`, `fp8_e5m2` | 16, 32, 64 | 64, 128, 256 | ❌ | ❌ | ✅ | Decoder | 7.x-9.x |
-| `FLASHINFER` | TRTLLM† | fp16, bf16 | `auto`, `float16`, `bfloat16`, `fp8`, `fp8_e4m3`, `fp8_e5m2` | 16, 32, 64 | 64, 128, 256 | ✅ | ❌ | ✅ | Decoder | 10.x |
-| `FLASH_ATTN` | FA2* | fp16, bf16 | `auto`, `float16`, `bfloat16` | %16 | Any | ❌ | ❌ | ✅ | All | ≥8.0 |
-| `FLASH_ATTN` | FA3* | fp16, bf16 | `auto`, `float16`, `bfloat16`, `fp8`, `fp8_e4m3`, `fp8_e5m2` | %16 | Any | ✅ | ❌ | ✅ | All | 9.x |
-| `FLASH_ATTN` | FA4* | fp16, bf16 | `auto`, `float16`, `bfloat16` | %16 | Any | ✅ | ❌ | ✅ | All | ≥10.0 |
-| `FLASH_ATTN_DIFFKV` | | fp16, bf16 | `auto` | Any | Any | ❌ | ❌ | ✅ | Decoder | Any |
-| `FLEX_ATTENTION` | | fp16, bf16, fp32 | `auto`, `float16`, `bfloat16` | Any | Any | ❌ | ✅ | ❌ | Decoder, Encoder Only | Any |
-| `ROCM_AITER_FA` | | fp16, bf16 | `auto`, `float16`, `bfloat16`, `fp8`, `fp8_e4m3`, `fp8_e5m2` | 16, 32 | 64, 128, 256 | ❌ | ❌ | ❌ | Decoder | N/A |
-| `ROCM_AITER_UNIFIED_ATTN` | | fp16, bf16 | `auto` | %16 | Any | ✅ | ✅ | ❌ | All | N/A |
-| `ROCM_ATTN` | | fp16, bf16, fp32 | `auto`, `float16`, `bfloat16`, `fp8`, `fp8_e4m3`, `fp8_e5m2` | %16 | 32, 64, 80, 96, 128, 160, 192, 224, 256 | ❌ | ✅ | ❌ | Decoder, Encoder, Encoder Only | N/A |
-| `TREE_ATTN` | | fp16, bf16 | `auto`, `float16`, `bfloat16` | %16 | 32, 64, 96, 128, 160, 192, 224, 256 | ❌ | ❌ | ❌ | Decoder | Any |
-| `TRITON_ATTN` | | fp16, bf16, fp32 | `auto`, `float16`, `bfloat16`, `fp8`, `fp8_e4m3`, `fp8_e5m2`, `int8_per_token_head`, `fp8_per_token_head` | %16 | Any | ✅ | ✅ | ❌ | All | Any |
-| `TURBOQUANT` | | fp16, bf16 | `turboquant_k8v4`, `turboquant_4bit_nc`, `turboquant_k3v4_nc`, `turboquant_3bit_nc` | 16, 32, 64, 128 | Any | ❌ | ❌ | ❌ | Decoder | Any |
+| Backend | Version | Dtypes | KV Dtypes | Block Sizes | Head Sizes | Sink | Non-Causal | MM Prefix | DCP | Attention Types | Compute Cap. |
+| ------- | ------- | ------ | --------- | ----------- | ---------- | ---- | ---------- | --------- | --- | --------------- | ------------ |
+| `CPU_ATTN` | | fp16, bf16, fp32 | `auto`, `fp8`, `fp8_e4m3`, `fp8_e5m2` | Any | 32, 64, 80, 96, 112, 128, 160, 192, 224, 256, 512 | ❌ | ❌ | ❌ | ❌ | All | N/A |
+| `FLASHINFER` | Native† | fp16, bf16 | `auto`, `float16`, `bfloat16`, `fp8`, `fp8_e4m3`, `fp8_e5m2` | 16, 32, 64 | 64, 128, 256 | ❌ | ❌ | ❌ | ✅ | Decoder | 7.x-9.x |
+| `FLASHINFER` | TRTLLM† | fp16, bf16 | `auto`, `float16`, `bfloat16`, `fp8`, `fp8_e4m3`, `fp8_e5m2`, `nvfp4` | 16, 32, 64 | 64, 128, 256 | ✅ | ❌ | ❌ | ✅ | Decoder | 10.x |
+| `FLASH_ATTN` | FA2* | fp16, bf16 | `auto`, `float16`, `bfloat16` | %16 | Any | ❌ | ✅ | ❌ | ✅ | All | ≥8.0 |
+| `FLASH_ATTN` | FA3* | fp16, bf16 | `auto`, `float16`, `bfloat16`, `fp8`, `fp8_e4m3`, `fp8_e5m2` | %16 | Any | ✅ | ✅ | ❌ | ✅ | All | 9.x |
+| `FLASH_ATTN` | FA4* | fp16, bf16 | `auto`, `float16`, `bfloat16` | %16 | Any | ✅ | ✅ | ❌ | ✅ | All | ≥10.0 |
+| `FLASH_ATTN_DIFFKV` | | fp16, bf16 | `auto` | Any | Any | ❌ | ❌ | ❌ | ✅ | Decoder | Any |
+| `FLEX_ATTENTION` | | fp16, bf16, fp32 | `auto`, `float16`, `bfloat16` | %16 | Any | ❌ | ✅ | ✅ | ❌ | Decoder, Encoder Only | Any |
+| `ROCM_AITER_FA` | | fp16, bf16 | `auto`, `float16`, `bfloat16`, `fp8`, `fp8_e4m3`, `fp8_e5m2` | 16, 32 | 64, 128, 256 | ❌ | ✅ | ❌ | ❌ | Decoder | N/A |
+| `ROCM_AITER_UNIFIED_ATTN` | | fp16, bf16 | `auto` | %16 | Any | ✅ | ❌ | ✅ | ❌ | All | N/A |
+| `ROCM_ATTN` | | fp16, bf16, fp32 | `auto`, `float16`, `bfloat16`, `fp8`, `fp8_e4m3`, `fp8_e5m2` | %16 | 32, 64, 80, 96, 128, 160, 192, 224, 256 | ❌ | ✅ | ✅ | ❌ | Decoder, Encoder, Encoder Only | N/A |
+| `TREE_ATTN` | | fp16, bf16 | `auto`, `float16`, `bfloat16` | %16 | 32, 64, 96, 128, 160, 192, 224, 256 | ❌ | ❌ | ❌ | ❌ | Decoder | Any |
+| `TRITON_ATTN` | | fp16, bf16, fp32 | `auto`, `float16`, `bfloat16`, `fp8`, `fp8_e4m3`, `fp8_e5m2`, `int8_per_token_head`, `fp8_per_token_head` | %16 | Any | ✅ | ❌ | ✅ | ❌ | All | Any |
+| `TURBOQUANT` | | fp16, bf16 | `turboquant_k8v4`, `turboquant_4bit_nc`, `turboquant_k3v4_nc`, `turboquant_3bit_nc` | 16, 32, 64, 128 | Any | ❌ | ❌ | ❌ | ❌ | Decoder | Any |

 > **†** FlashInfer uses TRTLLM attention on Blackwell (SM100), which supports sinks. Disable via `--attention-config.use_trtllm_attention=0`.
 >
@@ -192,31 +193,35 @@ MLA uses separate backends for prefill and decode phases.

 ### Prefill Backends

-The prefill backend is selected at runtime based on hardware and
-configuration.
+To explicitly select a prefill backend, use
+`-ac.mla_prefill_backend=<BACKEND>` (e.g., `FLASH_ATTN`, `FLASHINFER`).
+Otherwise, the prefill backend is selected automatically at runtime based on
+hardware and configuration.

-| Backend | Description | Compute Cap. | Enable | Disable | Notes |
-| ------- | ----------- | ------------ | ------ | ------- | ----- |
-| TRT-LLM Ragged‡ | TensorRT-LLM ragged attention | 10.x | Default on SM100 | `-ac.use_trtllm_ragged_deepseek_prefill=0` | DeepSeek R1 dims only |
-| FlashInfer | FlashInfer CUTLASS backend | 10.x | `-ac.disable_flashinfer_prefill=0` | `-ac.disable_flashinfer_prefill=1` | DeepSeek R1 dims only |
-| cuDNN | cuDNN-based attention | 10.x | `-ac.use_cudnn_prefill=1` | `-ac.use_cudnn_prefill=0` | |
-| FlashAttention | FlashAttention varlen (FA2/FA3) | Any | Default fallback | Use other backends | FA3 on SM90, FA2 otherwise |
+| Backend | Description | Dtypes | Compute Cap. | Notes |
+| ------- | ----------- | ------ | ------------ | ----- |
+| `FLASH_ATTN`‡ | FlashAttention varlen (FA2/FA3/FA4) | fp16, bf16 | Any | FA4 on SM100+, FA3 on SM90, FA2 otherwise |
+| `TRTLLM_RAGGED` | TensorRT-LLM ragged attention | fp16, bf16 | 10.x | DeepSeek R1 dims only |
+| `FLASHINFER` | FlashInfer CUTLASS backend | fp16, bf16 | 10.x | DeepSeek R1 dims only |

 > **‡** TRT-LLM Ragged is the default on Blackwell (SM100).
 > On other GPUs, FlashAttention is used as the default.

 ### Decode Backends

-| Backend | Dtypes | KV Dtypes | Block Sizes | Head Sizes | Sink | Sparse | MM Prefix | DCP | Attention Types | Compute Cap. |
-| ------- | ------ | --------- | ----------- | ---------- | ---- | ------ | --------- | --- | --------------- | ------------ |
-| `CUTLASS_MLA` | fp16, bf16 | `auto`, `float16`, `bfloat16`, `fp8`, `fp8_e4m3` | 128 | Any | ❌ | ❌ | ❌ | ✅ | Decoder | 10.x |
-| `FLASHINFER_MLA` | fp16, bf16 | `auto`, `float16`, `bfloat16`, `fp8`, `fp8_e4m3` | 32, 64 | Any | ❌ | ❌ | ❌ | ❌ | Decoder | 10.x |
-| `FLASHINFER_MLA_SPARSE` | fp16, bf16 | `auto`, `float16`, `bfloat16`, `fp8`, `fp8_e4m3` | 32, 64 | 576 | ❌ | ✅ | ❌ | ❌ | Decoder | 10.x |
-| `FLASHMLA` | fp16, bf16 | `auto`, `float16`, `bfloat16`, `fp8`, `fp8_e4m3` | 64 | Any | ❌ | ❌ | ❌ | ✅ | Decoder | 9.x-10.x |
-| `FLASHMLA_SPARSE` | bf16 | `auto`, `bfloat16`, `fp8_ds_mla` | 64 | 512, 576 | ❌ | ✅ | ❌ | ❌ | Decoder | 9.x-10.x |
-| `FLASH_ATTN_MLA` | fp16, bf16 | `auto`, `float16`, `bfloat16` | %16 | Any | ❌ | ❌ | ❌ | ✅ | Decoder | 9.x |
-| `ROCM_AITER_MLA` | fp16, bf16 | `auto`, `float16`, `bfloat16`, `fp8`, `fp8_e4m3`, `fp8_e5m2` | %1 | Any | ❌ | ❌ | ❌ | ❌ | Decoder | N/A |
-| `ROCM_AITER_MLA_SPARSE` | fp16, bf16 | `auto`, `float16`, `bfloat16` | 1 | Any | ❌ | ✅ | ❌ | ❌ | Decoder | N/A |
-| `ROCM_AITER_TRITON_MLA` | fp16, bf16 | `auto` | Any | Any | ❌ | ❌ | ❌ | ❌ | Decoder | N/A |
-| `TRITON_MLA` | fp16, bf16 | `auto`, `float16`, `bfloat16`, `fp8`, `fp8_e4m3` | %16 | Any | ❌ | ❌ | ❌ | ✅ | Decoder | Any |
-| `XPU_MLA_SPARSE` | fp16, bf16 | `auto`, `float16`, `bfloat16` | Any | 576 | ❌ | ✅ | ❌ | ❌ | Decoder | Any |
+MLA decode backends are selected using the standard
+`-ac.backend=<BACKEND>` argument (e.g., `FLASHMLA`, `TRITON_MLA`).
+
+| Backend | Dtypes | KV Dtypes | Block Sizes | Head Sizes | Sink | Non-Causal | Sparse | MM Prefix | DCP | Attention Types | Compute Cap. |
+| ------- | ------ | --------- | ----------- | ---------- | ---- | ---------- | ------ | --------- | --- | --------------- | ------------ |
+| `CUTLASS_MLA` | fp16, bf16 | `auto`, `float16`, `bfloat16`, `fp8`, `fp8_e4m3` | 128 | Any | ❌ | ❌ | ❌ | ❌ | ✅ | Decoder | 10.x |
+| `FLASHINFER_MLA` | fp16, bf16 | `auto`, `float16`, `bfloat16`, `fp8`, `fp8_e4m3` | 32, 64 | Any | ❌ | ❌ | ❌ | ❌ | ❌ | Decoder | 10.x |
+| `FLASHINFER_MLA_SPARSE` | fp16, bf16 | `auto`, `float16`, `bfloat16`, `fp8`, `fp8_e4m3` | 32, 64 | 576 | ❌ | ❌ | ✅ | ❌ | ❌ | Decoder | 10.x |
+| `FLASHMLA` | fp16, bf16 | `auto`, `float16`, `bfloat16`, `fp8`, `fp8_e4m3` | 64 | Any | ❌ | ❌ | ❌ | ❌ | ✅ | Decoder | 9.x-10.x |
+| `FLASHMLA_SPARSE` | bf16 | `auto`, `bfloat16`, `fp8_ds_mla` | 64 | 512, 576 | ❌ | ❌ | ✅ | ❌ | ❌ | Decoder | 9.x-10.x |
+| `FLASH_ATTN_MLA` | fp16, bf16 | `auto`, `float16`, `bfloat16` | %16 | Any | ❌ | ❌ | ❌ | ❌ | ✅ | Decoder | 9.x |
+| `ROCM_AITER_MLA` | fp16, bf16 | `auto`, `float16`, `bfloat16`, `fp8`, `fp8_e4m3`, `fp8_e5m2` | %1 | Any | ❌ | ❌ | ❌ | ❌ | ❌ | Decoder | N/A |
+| `ROCM_AITER_MLA_SPARSE` | fp16, bf16 | `auto`, `float16`, `bfloat16`, `fp8`, `fp8_e4m3` | 1, 64 | Any | ❌ | ❌ | ✅ | ❌ | ❌ | Decoder | N/A |
+| `ROCM_AITER_TRITON_MLA` | fp16, bf16 | `auto` | Any | Any | ❌ | ❌ | ❌ | ❌ | ❌ | Decoder | N/A |
+| `TRITON_MLA` | fp16, bf16 | `auto`, `float16`, `bfloat16`, `fp8`, `fp8_e4m3` | %16 | Any | ❌ | ❌ | ❌ | ❌ | ✅ | Decoder | Any |
+| `XPU_MLA_SPARSE` | fp16, bf16 | `auto`, `float16`, `bfloat16` | Any | 576 | ❌ | ❌ | ✅ | ❌ | ❌ | Decoder | Any |
@@ -86,9 +86,11 @@ Models opt-in to encoder CUDA Graphs by implementing the [SupportsEncoderCudaGra
 | Architecture | Models | CG for Image | CG for Video |
 | ------------ | ------ | ------------ | ------------ |
 | `Qwen3VLForConditionalGeneration` | `Qwen3-VL` | ✅︎ | ✅︎ |
+| `Qwen2_5_VLForConditionalGeneration` | `Qwen2.5-VL` | ✅︎ | ✅︎ |

 !!! note
    Encoder CUDA Graphs have currently been tested with `--mm-encoder-attn-backend=FLASH_ATTN` and `--mm-encoder-attn-backend=FLASHINFER` on Blackwell GPUs.
+    For Qwen2.5-VL only FA2 and FA3 has been tested.

 ## Configuration

@@ -5,12 +5,14 @@ TL;DR:
 - use tlparse to acquire torch.compile logs. Include these logs in bug reports and/or support asks.
 - The vLLM-torch.compile integration is multiple pieces. vLLM exposes flags to turn off each piece:

-| Online Flag | Offline Flag | Result |
-| ----------- | ------------ | ------ |
-| --enforce-eager | enforce_eager=True | Turn off torch.compile and CUDAGraphs |
-| -cc.mode=0 | mode=CompilationMode.NONE | Turn off torch.compile only |
-| -cc.cudagraph_mode=NONE | compilation_config=CompilationConfig(cudagraph_mode=CUDAGraphMode.NONE) | Turn off CUDAGraphs only |
-| -cc.backend=eager | compilation_config=CompilationConfig(backend='eager') | Turn off TorchInductor |
+| Online Flag                    | Offline Flag                                                                   | Result                                               |
+|--------------------------------|--------------------------------------------------------------------------------|------------------------------------------------------|
+| --enforce-eager                | enforce_eager=True                                                             | Turn off torch.compile and CUDAGraphs                |
+| -cc.mode=0                     | compilation_config=CompilationConfig(mode=CompilationMode.NONE)                | Turn off torch.compile only                          |
+| -cc.mode=1                     | compilation_config=CompilationConfig(mode=CompilationMode.STOCK_TORCH_COMPILE) | Turn off vLLM-compile modifications to torch.compile |
+| -cc.cudagraph_mode=NONE        | compilation_config=CompilationConfig(cudagraph_mode=CUDAGraphMode.NONE)        | Turn off CUDAGraphs only                             |
+| -cc.backend=eager              | compilation_config=CompilationConfig(backend='eager')                          | Turn off TorchInductor                               |
+| -cc.ir_enable_torch_wrap=False | compilation_config=CompilationConfig(ir_enable_torch_wrap=False)               | Turn off vLLM IR wrapping                            |

 ## vLLM-torch.compile overview

@@ -22,7 +24,7 @@ Most notably, vLLM-compile is NOT torch.compile, it is a custom compiler built u

 - Given a model, we do a full graph capture via TorchDynamo that is dynamic on the batch size (number of tokens)
 - vLLM then optionally splits and/or specializes this graph and then uses TorchInductor to compile each graph into a compiled artifact.
-This step may use vLLM custom Inductor passes to further optimize the graph.
+This step may use vLLM custom Inductor passes to further optimize the graph. This includes vLLM IR lowering to remove dispatch overhead.
 - The compiled artifact is saved to vLLM's compile cache so that it can be loaded in the future.
 - vLLM applies CUDAGraphs to reduce CPU overheads.

@@ -34,6 +36,7 @@ For more details on the design, please see the following resources:

 - [Introduction to vLLM-torch.compile blogpost](https://blog.vllm.ai/2025/08/20/torch-compile.html)
 - [vLLM-torch.compile integration design](./torch_compile.md)
+- [vLLM IR design](./vllm_ir.md)
 - [vLLM Office Hours #26](https://www.youtube.com/live/xLyxc7hxCJc?si=Xulo9pe53C6ywf0V&t=561)
 - [Talk at PyTorch Conference 2025](https://youtu.be/1wV1ESbGrVQ?si=s1GqymUfwiwOrDTg&t=725)

@@ -117,6 +120,21 @@ from vllm.config.compilation import CompilationConfig, CUDAGraphMode
 LLM(model, compilation_config=CompilationConfig(cudagraph_mode=CUDAGraphMode.NONE))
 ```

+vLLM IR makes heavy use of the compilation pipeline, from functionalization, custom fusions, and lowering.
+To turn that off and capture eager-mode dispatching behavior of vLLM IR, run with `ir_enable_torch_wrap=False`.
+IR torch wrap is only enabled by default when using `mode=VLLM_COMPILE` and `backend="inductor"` (default).
+
+```sh
+# Online
+vllm serve -cc.ir_enable_torch_wrap=False
+```
+
+```py
+# Offline
+from vllm.config.compilation import CompilationConfig
+LLM(model, compilation_config=CompilationConfig(ir_enable_torch_wrap=False))
+```
+
 ## Debugging TorchDynamo

 vLLM requires model code be capturable into a full graph via TorchDynamo (torch.compile's frontend).
@@ -36,7 +36,7 @@ th {
 | deepep_high_throughput | standard | fp8 | G(128),A,T<sup>2</sup> | Y | Y | [`DeepEPHTPrepareAndFinalize`][vllm.model_executor.layers.fused_moe.prepare_finalize.deepep_ht.DeepEPHTPrepareAndFinalize] |
 | deepep_low_latency | batched | fp8 | G(128),A,T<sup>3</sup> | Y | Y | [`DeepEPLLPrepareAndFinalize`][vllm.model_executor.layers.fused_moe.prepare_finalize.deepep_ll.DeepEPLLPrepareAndFinalize] |
 | flashinfer_nvlink_two_sided | standard | nvfp4,fp8 | G,A,T | N | N | [`FlashInferNVLinkTwoSidedPrepareAndFinalize`][vllm.model_executor.layers.fused_moe.prepare_finalize.flashinfer_nvlink_two_sided.FlashInferNVLinkTwoSidedPrepareAndFinalize] |
-| flashinfer_nvlink_one_sided | standard | nvfp4 | G,A,T | N | N | [`FlashInferNVLinkOneSidedPrepareAndFinalize`][vllm.model_executor.layers.fused_moe.prepare_finalize.flashinfer_nvlink_one_sided.FlashInferNVLinkOneSidedPrepareAndFinalize] |
+| flashinfer_nvlink_one_sided | standard | nvfp4,bf16,mxfp8 | G,A,T | N | N | [`FlashInferNVLinkOneSidedPrepareAndFinalize`][vllm.model_executor.layers.fused_moe.prepare_finalize.flashinfer_nvlink_one_sided.FlashInferNVLinkOneSidedPrepareAndFinalize] |

 !!! info "Table key"
    1. All types: mxfp4, nvfp4, int4, int8, fp8
@@ -0,0 +1,615 @@
+# vLLM IR: Functional Intermediate Representation
+
+## Motivation
+
+vLLM IR is a **functional intermediate representation (IR)** that fills the gap between
+low-level `torch` ops and vLLM layers like `RMSNorm` and quantization operators,
+By separating operator **semantics** from the **implementation** and **dispatching**,
+vLLM IR simplifies both compilation and kernel registration & dispatching simultaneously.
+It operates as a **dialect** in the torch FX representation, allowing full interoperability
+with “regular” torch ops & custom torch ops/kernels, as well as a piecewise migration from
+the previous `CustomOp` approach.
+
+Key design principles:
+
+- **Eager-compile consistency**: identical behavior (barring minor numerics) in eager and compiled modes
+- **Simple, transparent, yet powerful kernel selection**: good visibility and control allowing easy debugging
+- **Convention over configuration**: near-zero boilerplate required to register ops and implementations
+- **Extensibility**: ops and implementations can be registered anywhere, in-tree or out-of-tree
+- **Interoperability**: fully compatible with “regular” torch ops & custom torch ops/kernels,
+reducing developer friction and allowing piecewise migration
+
+The clean semantics/implementation separation enables a unified and extensible dispatching mechanism,
+allowing multiple kernels per-platform and powerful kernel selection. The separation also facilitates
+cleaner testing and benchmarking, removing much of the boilerplate standard for legacy approaches.
+
+By delaying kernel selection until late in the compilation process, the compiler can operate on
+a higher-level representation, which has the following main benefits:
+
+- Pattern matching in fusion/transformation passes only requires a single, simple pattern per op
+- OOT compiler backends can lower from the higher-level representation (in-progress)
+- The compiler can autotune over available implementations (future feature)
+
+## Quick Overview
+
+### Declaring an IR Operation
+
+IR operations are declared using the `@register_op` decorator with a native PyTorch implementation that defines the op's semantics:
+
+```python
+# vllm/ir/ops/layernorm.py
+from torch import Tensor
+from vllm.ir import register_op
+
+@register_op
+def rms_norm(x: Tensor, weight: Tensor | None, epsilon: float, variance_size: int | None = None) -> Tensor:
+    """Weighted root-mean-square layer normalization"""
+    orig_dtype = x.dtype
+    x = x.to(torch.float32)
+    x_var = x if variance_size is None else x[..., :variance_size]
+    variance = x_var.pow(2).mean(dim=-1, keepdim=True)
+    x = x * torch.rsqrt(variance + epsilon)
+    x = x.to(orig_dtype)
+    if weight is not None:
+        x = x * weight
+    return x
+```
+
+The native implementation serves three purposes:
+
+1. **Semantic definition**: Specifies the exact semantics of the operation, including shapes and strides
+2. **Default implementation**: Used when no other (better) implementation is available
+3. **Reference for testing**: Other implementations must match these semantics
+
+### Registering Implementations
+
+Kernel implementations are registered using the `register_impl` decorator on the IR op object:
+
+```python
+# vllm/kernels/vllm_c.py
+from vllm import ir
+
+rms_norm_no_var = lambda x, weight, epsilon, variance_size=None: variance_size is None
+
+@ir.ops.rms_norm.register_impl("vllm_c", supports_args=rms_norm_no_var, supported=current_platform.is_cuda_alike())
+def rms_norm(x: Tensor, weight: Tensor | None, epsilon: float, variance_size: int | None = None) -> Tensor:
+    output = torch.empty_like(x)
+    torch.ops._C.rms_norm(output, x, weight, epsilon)
+    return output
+```
+
+Implementations can specify:
+
+- `supported`: Static boolean indicating if this implementation is available
+- `supports_args`: Function checking if the implementation supports specific arguments
+- `inplace`: Whether this implementation reuses input memory for outputs
+
+### Using IR Operations in Models
+
+IR operations are imported and called directly in model code:
+
+```python
+# vllm/model_executor/layers/layernorm.py
+from vllm import ir
+
+class RMSNorm(nn.Module):
+    def __init__(self, hidden_size: int, eps: float = 1e-6):
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(hidden_size))
+        self.variance_epsilon = eps
+
+    def forward(self, x: Tensor, residual: Tensor | None = None):
+        if residual is None:
+            return ir.ops.rms_norm(x, self.weight, self.variance_epsilon)
+
+        # Use maybe_inplace overload to allow implementation to reuse input memory for outputs
+        # (using x or residual after this call is undefined behavior)
+        return ir.ops.fused_add_rms_norm.maybe_inplace(
+            x, residual, self.weight, self.variance_epsilon
+        )
+```
+
+### Configuring Kernel Selection
+
+Kernel selection is controlled via priority lists in the configuration.
+Priority lists specify the order in which implementations are considered,
+with the first supported implementation being selected.
+This includes the static support check (`supported=...`) and
+the dynamic arg support check (`supports_args=...`).
+
+#### Command Line Configuration
+
+Use `--ir-op-priority.<op_name>=<provider1>,<provider2>,...`:
+
+```bash
+# CUDA: Use vllm_c implementation for rms_norm
+vllm serve meta-llama/Llama-3.2-1B \
+  --ir-op-priority.rms_norm=vllm_c
+
+# ROCm: Try aiter first, fall back to vllm_c, then native
+vllm serve meta-llama/Llama-3.2-1B \
+  --ir-op-priority.rms_norm=aiter,vllm_c,native
+
+# Configure multiple operations
+vllm serve meta-llama/Llama-3.2-1B \
+  --ir-op-priority.rms_norm=vllm_c \
+  --ir-op-priority.fused_add_rms_norm=vllm_c
+```
+
+#### Python Configuration
+
+```python
+from vllm import LLM
+from vllm.config import VllmConfig, KernelConfig
+
+llm = LLM(
+    model="meta-llama/Llama-3.2-1B",
+    vllm_config=VllmConfig(
+        kernel_config=KernelConfig(
+            ir_op_priority={
+                "rms_norm": ["vllm_c", "native"],
+                "fused_add_rms_norm": ["vllm_c", "native"],
+            }
+        )
+    )
+)
+```
+
+#### Platform Defaults
+
+Each platform provides default priority lists that are automatically applied:
+
+```python
+# CUDA/XPU/ROCm platform defaults (when compiling with Inductor)
+{
+  "rms_norm": ["native"],  # Native torch is default
+  "fused_add_rms_norm": ["native"],
+}
+
+# CUDA platform defaults (eager or Dynamo-only)
+{
+  "rms_norm": ["vllm_c", "native"],
+  "fused_add_rms_norm": ["vllm_c", "native"],
+}
+
+# ROCm platform defaults (future - currently same as CUDA)
+{
+    "rms_norm": ["aiter", "vllm_c", "native"],
+    "fused_add_rms_norm": ["aiter", "vllm_c", "native"],
+}
+
+# XPU platform defaults (eager or Dynamo-only)
+{
+    "rms_norm": ["xpu_kernels", "native"],
+    "fused_add_rms_norm": ["xpu_kernels", "native"],
+}
+```
+
+User-specified priorities are prepended to platform defaults,
+so you only need to specify the out-of-order implementations,
+other implementations are appended automatically.
+
+## Compilation Pipeline
+
+vLLM IR heavily customizes the `torch.compile`-based compilation process to allow custom compile
+passes to operate on high-level IR while still producing efficient low-level code at the end.
+The compilation pipeline consists of several stages:
+
+### 1. Dynamo Tracing
+
+When `torch.compile` traces the model's forward pass, vLLM IR operations appear as custom operations
+in the `vllm_ir` torch library. These operations are opaque to Dynamo, meaning they appear directly
+in the FX graph without decomposition:
+
+```python
+# Python code (epsilon=1e-5)
+x1 = ir.ops.rms_norm(x, weight, epsilon)
+x2, residual_out = ir.ops.fused_add_rms_norm.maybe_inplace(x1, residual, weight, epsilon)
+
+# FX graph after Dynamo tracing
+x1 = torch.ops.vllm_ir.rms_norm.default(x, weight, 1e-5); x = None
+out = torch.ops.vllm_ir.fused_add_rms_norm.maybe_inplace(x1, residual, weight, 1e-5); x1 = residual = None
+x2 = out[0]
+residual_out = out[1]
+```
+
+### 2. AOTAutograd and Functionalization
+
+AOTAutograd functionalizes the graph, converting any mutating operations to functional equivalents.
+For vLLM IR operations with `maybe_inplace` overloads, we perform this manually before AOTAutograd,
+converting them to the functional `default` overload using the pre-grad custom pass hook.
+
+```python
+# After functionalization
+x1 = torch.ops.vllm_ir.rms_norm.default(x, weight, 1e-5); x = None
+out = torch.ops.vllm_ir.fused_add_rms_norm.default(x1, residual, weight, 1e-5); x1 = residual = None
+x2 = out[0]
+residual_out = out[1]
+```
+
+The pass also tracks which inputs were "donated" (passed to `maybe_inplace`),
+storing this information in vLLM's `PassContext` for later use in clone elimination.
+
+### 3. IR Fusion and Transformation Passes
+
+After functionalization, custom vLLM passes operate on the functional FX graph containing high-level IR operations.
+These passes can perform fusion, distribute operations for sequence parallelism, and other transformations:
+
+```python
+# Example: Sequence Parallelism (see SequenceParallelismPass)
+# Before SP pass
+
+all_reduce = torch.ops.vllm.all_reduce(x, "tp:0")
+rms_norm = torch.ops.vllm_ir.rms_norm(all_reduce, weight, 1e-5)
+
+# after SP pass
+reduce_scatter = torch.ops.vllm.reduce_scatter(x, "tp:0")
+rms_norm = torch.ops.vllm_ir.rms_norm(all_reduce, weight, 1e-5)
+all_gather = torch.ops.vllm.all_gather(x, "tp:0")
+```
+
+Fusion passes benefit from the high-level representation: they don't need to match against low-level PyTorch operations,
+handle different kernel implementations separately, or deal with functionalization of custom kernels.
+
+### 4. IR Lowering
+
+The lowering pass (`VllmIRLoweringPass`) replaces each vLLM IR operation with its selected implementation.
+The implementation is chosen based on the priority list and support predicates,
+using the **fake tensors** in the graph's metadata in place of op arguments:
+
+```python
+# Implementation selection, same in eager dispatch and compile lowering
+def dispatch(*args) -> IrOpImpl:
+  for provider in priority_list:  # e.g., ["vllm_c", "native"]
+    impl = ir_op.impls[provider]
+    if not impl.supported:
+      continue
+    if impl.supports_args and not impl.supports_args(*args):
+      continue
+    return impl
+
+# make_fx uses torch.fx.symbolic_trace
+impl_graph = make_fx(selected_impl.impl_fn)
+# Replace IR op node with impl_graph's nodes
+match.replace_by_example(selected_impl.impl_fn, node.args)
+```
+
+For example, lowering `rms_norm` with the `vllm_c` implementation:
+
+```python
+# Before lowering (IR op)
+rms_norm = torch.ops.vllm_ir.rms_norm.default(x, weight, 1e-5)
+
+# After lowering (vllm_c implementation traced)
+# Note: Lowering does not currently functionalize, this will likely change in the future.
+empty =  torch.ops.aten.empty.memory_format(x.shape, ...)
+rms_norm = torch.ops._C.rms_norm(empty, x, weight, 1e-5)
+```
+
+When lowering an implementation that mutates inputs (`inplace=True`),
+the lowering pass inserts clones to preserve functional semantics:
+
+```python
+# vllm_c implementation for fused_add_rms_norm mutates its first two arguments
+# Lowered with clones for safety
+clone_default = torch.ops.aten.clone.default(x)
+clone_default_1 = torch.ops.aten.clone.default(residual)
+fused_add_rms_norm = torch.ops._C.fused_add_rms_norm.default(clone_default, clone_default_1, weight, 1e-5)
+```
+
+### 5. Clone Cleanup
+
+After lowering, the clone elimination pass (`UnsafeCloneEliminationPass`) removes unnecessary clones introduced during lowering.
+This pass is essential for achieving zero-copy behavior when using in-place kernels with `maybe_inplace`.
+The pass removes a clone if:
+
+- the cloned input is created in the graph and not used again in the graph
+- the cloned input is a graph parameter, marked as donated
+
+```python
+# After cleanup (donated inputs, no subsequent uses)
+fused_add_rms_norm = torch.ops._C.fused_add_rms_norm.default(x, residual, weight, 1e-5)
+```
+
+The combination of inplace functionalization (tracking donated inputs) and clone cleanup enables the compiler to safely
+use in-place kernels without adding redundant copies or increasing the memory usage.
+
+### 6. Inductor Optimization and Codegen
+
+After IR lowering and cleanup, the graph contains only standard PyTorch operations and platform-specific custom ops.
+Inductor then performs its standard codegen:
+
+- **Inductor lowering and pointwise fusion**: Fusing element-wise operations, reductions, etc.
+- **Memory planning**: Determining buffer allocation and reuse
+- **Kernel generation**: Generating Triton or C++ code for fused operations
+- **Autotuning**: Selecting the best kernel configurations
+
+### Pipeline Summary
+
+```text
+Model Forward Pass
+    ↓
+[Dynamo Tracing] → FX Graph with vllm_ir.* ops
+    ↓
+[Pre-grad: Inplace Functionalization] → maybe_inplace → default, track donated inputs
+    ↓
+[AOTAutograd] → Functionalization
+    ↓
+[Post-grad: IR Fusion Passes] → Fuse high-level IR ops (e.g., rms_norm + quant)
+    ↓
+[Post-grad: IR Lowering] → vllm_ir.* ops → impl ops (with clones if needed)
+    ↓
+[Post-grad: Clone Cleanup] → Remove unnecessary clones using donated input info
+    ↓
+[Inductor] → Pattern matching, fusion, memory planning, codegen
+    ↓
+Compiled Code
+```
+
+## Core vLLM IR Concepts
+
+### Operation Declaration
+
+Operations are declared with the `@register_op` decorator, which creates an `IrOp` object:
+
+```python
+@register_op(
+    name=None,           # Operation name (defaults to function name)
+    activations=None,    # List of activation parameters (defaults to params starting with 'x')
+    allow_inplace=False, # Whether to create a maybe_inplace overload
+)
+def op_name(...):
+    ...
+```
+
+**Parameters:**
+
+- `activations`: List of parameter names considered "activations" (typically consumed by `maybe_inplace`). Defaults to parameters starting with `x`.
+- `allow_inplace`: Creates a `maybe_inplace` overload for memory-efficient execution (see below).
+
+### The `maybe_inplace` Overload
+
+The `maybe_inplace` overload is a critical feature for memory efficiency in LLM inference.
+It signals that the caller doesn't need to preserve the activation inputs after the operation,
+allowing in-place implementations to reuse input memory for outputs.
+
+#### Semantics and Usage
+
+```python
+# Standard usage: inputs are preserved
+out, res_out = ir.ops.fused_add_rms_norm(x, residual, weight, epsilon)
+# x and residual are unchanged, out and res_out are new tensors
+
+# maybe_inplace: inputs may be modified
+out, res_out = ir.ops.fused_add_rms_norm.maybe_inplace(x, residual, weight, epsilon)
+# x and residual may be modified (undefined behavior to use them after this)
+# out and res_out may alias x and residual
+```
+
+Using an activation input after passing it to `maybe_inplace` is **undefined behavior**:
+
+```python
+# WRONG: Using x after donating it
+out, res_out = ir.ops.fused_add_rms_norm.maybe_inplace(x, residual, weight, epsilon)
+result = out + x  # ERROR: x was donated!
+```
+
+If you need to preserve an input, either use the default overload or clone manually:
+
+```python
+# Option 1: Use default overload
+out, res_out = ir.ops.fused_add_rms_norm(x, residual, weight, epsilon)
+result = out + x  # OK: x is preserved
+
+# Option 2: Clone before maybe_inplace
+out, res_out = ir.ops.fused_add_rms_norm.maybe_inplace(x.clone(), residual, weight, epsilon)
+result = out + x  # OK: x is preserved, clone was donated
+```
+
+#### Compilation Behavior
+
+During compilation, the inplace functionalization pass validates that donated inputs are
+not used again and converts `maybe_inplace` to the functional `default` overload:
+
+```python
+# Inplace functionalization pass (pre-grad)
+for node in graph.nodes:
+    if node.target == torch.ops.vllm_ir.fused_add_rms_norm.maybe_inplace:
+        # Check that activation inputs aren't used after this node
+        for activation_arg in activation_inputs:
+            for user in activation_arg.users:
+                if user appears after node:
+                    raise ValueError(f"Input {activation_arg} donated but used again")
+
+        # Convert to default overload
+        node.target = torch.ops.vllm_ir.fused_add_rms_norm.default
+
+        # Track donated graph inputs for later clone elimination
+        for i, arg in enumerate(node.args):
+            if arg.op == "placeholder" and i in activation_indices:
+                pass_context.donated_input_ids.add(node_to_idx[arg])
+```
+
+The donated input information is then used by the clone cleanup pass to eliminate
+unnecessary copies when in-place kernels are lowered.
+
+#### Eager Mode Behavior
+
+In eager mode (without `torch.compile`), `maybe_inplace` enables **maximally memory-efficient**
+execution by allowing the IR operation to dispatch directly to in-place implementations:
+
+```python
+# Eager dispatch logic for maybe_inplace
+impl: IrOpImpl = ir_op.dispatch(*args)
+return impl.impl_fn(*args)
+
+# Eager dispatch logic for default:
+impl: IrOpImpl = ir_op.dispatch(*args)
+if impl.inplace:
+  args = [
+    arg.clone() if i in ir_op.activations else arg
+    for i, arg in enumerate(args)
+  ]
+return impl.impl_fn(*args)
+```
+
+The combination of `maybe_inplace` in model code and in-place kernel implementations provides optimal memory efficiency
+in both eager and compiled modes, with identical semantics in both cases.
+
+#### Memory Savings Example
+
+Consider a transformer layer with residual connections:
+
+```python
+# Without maybe_inplace (2 allocations per layer)
+hidden_states = self.attention(input)
+normed, residual = ir.ops.fused_add_rms_norm(hidden_states, input, weight, eps)
+# Memory: input (preserved), hidden_states (preserved), normed (new), residual (new)
+
+# With maybe_inplace (0 allocations per layer when using in-place kernel)
+hidden_states = self.attention(input)
+normed, residual = ir.ops.fused_add_rms_norm.maybe_inplace(hidden_states, input, weight, eps)
+# Memory: normed (reuses hidden_states), residual (reuses input)
+```
+
+### Implementation Registration
+
+Implementations are registered using the `register_impl` method:
+
+```python
+@ir.ops.op_name.register_impl(
+    provider="provider_name",  # Unique identifier (e.g., "vllm_c", "aiter", "triton")
+    supported=True,            # Static availability check
+    supports_args=None,        # Dynamic argument support check
+)
+def impl_fn(...):
+    ...
+```
+
+**Provider naming conventions:**
+
+- `native`: Reserved for the native torch implementation (declared with `@register_op`)
+- `vllm_c`: C++/CUDA kernels via `torch.ops._C`
+- `aiter`: AMD AITER library
+- `xpu_kernels`: SYCL/SYCLTLA kernels implemented in `vllm-xpu-kernels`
+- `triton_*`: Triton kernels
+- Platform/library names for other implementations
+
+**Support checking:**
+
+- `supported`: Static boolean, checked once at import time (e.g., `HAS_TRITON`, `is_cuda_alike()`)
+- `supports_args`: Function `(*args, **kwargs) -> bool` checking argument compatibility
+    - Called with **fake tensors** during compilation for zero-cost checking
+    - Called with **real tensors** during eager mode dispatch
+    - Should NOT check batch sizes or add guards based on values
+
+Example support predicate:
+
+```python
+def aiter_rms_norm_supports(x, weight, epsilon, variance_size=None):
+    # Check dtype (OK: doesn't depend on batch size)
+    if x.dtype not in [torch.float16, torch.bfloat16]:
+        return False
+    # Check optional parameter (OK: static check)
+    if variance_size is not None:
+        return False
+    return True
+
+@ir.ops.rms_norm.register_impl("aiter", supports_args=aiter_rms_norm_supports)
+def rms_norm(...):
+    ...
+```
+
+Batch-invariant kernels are automatically selected when `VLLM_BATCH_INVARIANT=1` is set.
+
+### Eager Mode vs Compile Mode
+
+vLLM IR operations behave identically in eager and compile modes:
+
+**Eager mode:**
+
+- Direct dispatch to implementation based on priority list
+- Support checked with real tensor arguments
+- Minimal overhead (can be optimized further if needed)
+
+**Compile mode:**
+
+- IR ops appear in FX graph as `torch.ops.vllm_ir.*` custom ops
+- Lowering selects implementation using fake tensors
+- Full integration with Inductor optimizations
+
+This consistency enables:
+
+- Prototyping in eager mode with confidence
+- Debugging by disabling compilation
+- Gradual migration from eager to compiled execution
+
+## Other Topics
+
+### Out-of-Tree Implementations
+
+External platforms can register implementations without modifying vLLM:
+
+```python
+# In external package
+from vllm import ir
+
+@ir.ops.rms_norm.register_impl("my_platform", supported=is_my_platform())
+def rms_norm(x, weight, epsilon, variance_size=None):
+    return my_platform.rms_norm(x, weight, epsilon)
+```
+
+Then configure priority to use your implementation:
+
+```python
+class MyPlatform(Platform):
+  def get_default_ir_op_priority(self):
+    return IrOpPriorityConfig(rms_norm=['my_platform', 'native'])
+
+# Users can still override priority in the same way
+llm = LLM(ir_op_priority=IrOpPriorityConfig(rms_norm=['custom_oot_kernel']))
+```
+
+### Debugging and Observability
+
+!!! note
+    Please let us know how observability can be improved for your use-case!
+
+Enable debug logging to see kernel selection:
+
+```bash
+VLLM_LOGGING_LEVEL=DEBUG vllm serve ...
+```
+
+This logs:
+
+- Which implementations are selected for each operation
+- Why implementations were rejected (unsupported, args not supported)
+- Compilation cache hits/misses
+- IR lowering statistics
+
+Check selected implementations in compiled graphs:
+
+```python
+# After compilation, inspect the lowering pass
+lowering_pass = backend.lowering_pass
+print(lowering_pass.selected_impls)
+# Output: {'rms_norm': {'node_123': 'vllm_c', 'node_456': 'vllm_c'}}
+```
+
+## Migration from CustomOp
+
+vLLM IR is designed to coexist with and gradually replace `CustomOp`:
+
+1. **Op declaration**: Convert `CustomOp` class `PluggableLayer` and move `forward_native` to `@register_op` function
+2. **Implementation registration**: Use `@ir.ops.op_name.register_impl` instead of overriding methods
+3. **Layer usage**: Replace `self.op(...)` with `ir.ops.op_name(...)`
+4. **Configuration**: Migrate `--compilation-config.custom-ops` to `--ir-op-priority`
+
+The migration can be done incrementally, one operation at a time.
+
+## See Also
+
+- [torch.compile Integration](torch_compile.md) - General compilation infrastructure
+- [Fusions](fusions.md) - Custom fusion and transformation passes in vLLM
+- [Custom Operations](custom_op.md) - Legacy custom op system
@@ -52,10 +52,10 @@ th:not(:first-child) {
 | [mm](multimodal_inputs.md) | ✅ | ✅ | [🟠](https://github.com/vllm-project/vllm/pull/4194)<sup>^</sup> | ❔ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❔ | ✅ | | | |
 | best-of | ✅ | ✅ | ✅ | [❌](https://github.com/vllm-project/vllm/issues/6137) | ✅ | ❌ | ✅ | ✅ | ✅ | ❔ | [❌](https://github.com/vllm-project/vllm/issues/7968) | ✅ | ✅ | | |
 | beam-search | ✅ | ✅ | ✅ | [❌](https://github.com/vllm-project/vllm/issues/6137) | ✅ | ❌ | ✅ | ✅ | ✅ | ❔ | [❌](https://github.com/vllm-project/vllm/issues/7968) | ❔ | ✅ | ✅ | |
-| [prompt-embeds](prompt_embeds.md) | ✅ | ✅ | ✅ | ❌ | ✅ | ❌ | ❌ | ✅ | ❌ | ❔ | ❔ | ❌ | ❔ | ❔ | ✅ |
+| [prompt-embeds](prompt_embeds.md) | ✅ | ✅ | ✅ | ❌ | ✅ | ❌ | ❌ | ✅ | ❌ | ❔ | ❔ | ✅ | ❔ | ❔ | ✅ |

 \* Chunked prefill and prefix caching are only applicable to last-token or all pooling with causal attention.  
-<sup>^</sup> LoRA is only applicable to the language backbone of multimodal models.
+<sup>^</sup> LoRA is only applicable to the language backbone of multimodal models.  

 ### Feature x Hardware

@@ -11,7 +11,7 @@ Automatic Prefix Caching (APC in short) caches the KV cache of existing queries,

 Set `enable_prefix_caching=True` in vLLM engine to enable APC. Here is an example:

-[examples/offline_inference/automatic_prefix_caching.py](../../examples/offline_inference/automatic_prefix_caching.py)
+[examples/features/automatic_prefix_caching/automatic_prefix_caching_offline.py](../../examples/features/automatic_prefix_caching/automatic_prefix_caching_offline.py)

 ## Example workloads

--- a/Show More
+++ b/Show More