Fix ci

Signed-off-by: yewentao256 <zhyanwentao@126.com>
2026-04-22 18:26:48 +00:00
523 changed files with 7572 additions and 37929 deletions
@@ -46,7 +46,7 @@ echo "Image not found, proceeding with build..."

 # --- CUDA 13.0 for nightly builds ---
 # Nightly CI uses CUDA 13.0 while regular CI stays on CUDA 12.9
-NIGHTLY_CUDA_VERSION="13.0.2"
+NIGHTLY_CUDA_VERSION="13.0.0"
 NIGHTLY_BUILD_BASE_IMAGE="nvidia/cuda:${NIGHTLY_CUDA_VERSION}-devel-ubuntu22.04"
 NIGHTLY_FINAL_BASE_IMAGE="nvidia/cuda:${NIGHTLY_CUDA_VERSION}-base-ubuntu22.04"

@@ -20,7 +20,7 @@ steps:
      'cd tests &&
      pytest -v -s lora/test_layers.py &&
      pytest -v -s lora/test_lora_checkpoints.py &&
-      (pytest -v -s lora/test_lora_functions.py --deselect="tests/lora/test_lora_functions.py::test_lora_functions_sync" --deselect="tests/lora/test_lora_functions.py::test_lora_functions_async" || true) &&
+      pytest -v -s lora/test_lora_functions.py &&
      pytest -v -s lora/test_lora_huggingface.py &&
      pytest -v -s lora/test_lora_manager.py &&
      pytest -v -s lora/test_lora_utils.py &&
@@ -125,6 +125,6 @@ steps:
      bash .buildkite/scripts/hardware_ci/run-intel-test.sh
      'cd tests &&
      pytest -v -s lora/test_default_mm_loras.py && 
-      (pytest -v -s lora/test_qwen3_unembed.py || true) &&
-      (pytest -v -s lora/test_qwenvl.py || true) &&
+      pytest -v -s lora/test_qwen3_unembed.py &&
+      pytest -v -s lora/test_qwenvl.py &&
      pytest -v -s lora/test_whisper.py'
@@ -1,13 +1,3 @@
-# CUDA architecture lists — following PyTorch RELEASE.md
-# (https://github.com/pytorch/pytorch/blob/main/RELEASE.md)
-# SM86 included for broader Ampere coverage; SM89 for marlin fp8 support
-env:
-  CUDA_ARCH_X86: "7.5 8.0 8.6 8.9 9.0 10.0 12.0+PTX"
-  # aarch64 only architectures: 8.7 for Orin, 11.0 for Thor (since CUDA 13)
-  CUDA_ARCH_AARCH64: "8.0 8.7 8.9 9.0 10.0 11.0 12.0+PTX"
-  CUDA_ARCH_X86_CU129: "7.5 8.0 8.6 8.9 9.0 10.0 12.0"
-  CUDA_ARCH_AARCH64_CU129: "8.0 8.7 8.9 9.0 10.0 12.0"
-
 steps:
  - input: "Provide Release version here"
    id: input-release-version
@@ -24,10 +14,12 @@ steps:
        agents:
          queue: arm64_cpu_queue_release
        commands:
-          - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=12.9.1 --build-arg torch_cuda_arch_list=\"${CUDA_ARCH_AARCH64_CU129}\" --tag vllm-ci:build-image --target build --progress plain -f docker/Dockerfile ."
+          # #NOTE: torch_cuda_arch_list is derived from upstream PyTorch build files here:
+          # https://github.com/pytorch/pytorch/blob/main/.ci/aarch64_linux/aarch64_ci_build.sh#L7
+          - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=12.9.1 --build-arg torch_cuda_arch_list='8.7 8.9 9.0 10.0+PTX 12.0' --tag vllm-ci:build-image --target build --progress plain -f docker/Dockerfile ."
          - "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"

@@ -37,7 +29,9 @@ 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 ."
+          # #NOTE: torch_cuda_arch_list is derived from upstream PyTorch build files here:
+          # https://github.com/pytorch/pytorch/blob/main/.ci/aarch64_linux/aarch64_ci_build.sh#L7
+          - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=13.0.1 --build-arg torch_cuda_arch_list='8.7 8.9 9.0 10.0+PTX 12.0' --build-arg BUILD_BASE_IMAGE=nvidia/cuda:13.0.1-devel-ubuntu22.04  --tag vllm-ci:build-image --target build --progress plain -f docker/Dockerfile ."
          - "mkdir artifacts"
          - "docker run --rm -v $(pwd)/artifacts:/artifacts_host vllm-ci:build-image bash -c 'cp -r dist /artifacts_host && chmod -R a+rw /artifacts_host'"
          - "bash .buildkite/scripts/upload-nightly-wheels.sh manylinux_2_35"
@@ -63,7 +57,7 @@ steps:
        agents:
          queue: cpu_queue_release
        commands:
-          - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=12.9.1 --build-arg torch_cuda_arch_list=\"${CUDA_ARCH_X86_CU129}\" --tag vllm-ci:build-image --target build --progress plain -f docker/Dockerfile ."
+          - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=12.9.1 --tag vllm-ci:build-image --target build --progress plain -f docker/Dockerfile ."
          - "mkdir artifacts"
          - "docker run --rm -v $(pwd)/artifacts:/artifacts_host vllm-ci:build-image bash -c 'cp -r dist /artifacts_host && chmod -R a+rw /artifacts_host'"
          - "bash .buildkite/scripts/upload-nightly-wheels.sh manylinux_2_31"
@@ -76,7 +70,7 @@ steps:
        agents:
          queue: cpu_queue_release
        commands:
-          - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=13.0.2 --build-arg torch_cuda_arch_list=\"${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=13.0.1 --build-arg BUILD_BASE_IMAGE=nvidia/cuda:13.0.1-devel-ubuntu22.04 --tag vllm-ci:build-image --target build --progress plain -f docker/Dockerfile ."
          - "mkdir artifacts"
          - "docker run --rm -v $(pwd)/artifacts:/artifacts_host vllm-ci:build-image bash -c 'cp -r dist /artifacts_host && chmod -R a+rw /artifacts_host'"
          - "bash .buildkite/scripts/upload-nightly-wheels.sh manylinux_2_35"
@@ -114,95 +108,96 @@ steps:
    depends_on: block-build-release-images
    allow_dependency_failure: true
    steps:
-      - label: "Build release image - x86_64 - CUDA 13.0"
+      - label: "Build release image - x86_64 - CUDA 12.9"
        depends_on: ~
        id: build-release-image-x86
        agents:
          queue: cpu_queue_release
        commands:
          - "aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin public.ecr.aws/q9t5s3a7"
-          - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=13.0.2 --build-arg torch_cuda_arch_list=\"${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 --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=12.9.1 --build-arg FLASHINFER_AOT_COMPILE=true --build-arg INSTALL_KV_CONNECTORS=true --tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m) --target vllm-openai --progress plain -f docker/Dockerfile ."
          - "docker push public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)"
          # re-tag to default image tag and push, just in case arm64 build fails
          - "docker tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m) public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT"
          - "docker push public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT"

-      - label: "Build release image - aarch64 - CUDA 13.0"
+      - label: "Build release image - aarch64 - CUDA 12.9"
        depends_on: ~
        id: build-release-image-arm64
        agents:
          queue: arm64_cpu_queue_release
        commands:
          - "aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin public.ecr.aws/q9t5s3a7"
-          - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=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-$(uname -m) --target vllm-openai --progress plain -f docker/Dockerfile ."
+          - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=12.9.1 --build-arg FLASHINFER_AOT_COMPILE=true --build-arg torch_cuda_arch_list='8.7 8.9 9.0 10.0+PTX 12.0' --build-arg INSTALL_KV_CONNECTORS=true --tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m) --target vllm-openai --progress plain -f docker/Dockerfile ."
          - "docker push public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)"

-      - label: "Build release image - x86_64 - CUDA 12.9"
+      - label: "Build release image - x86_64 - CUDA 13.0"
        depends_on: ~
-        id: build-release-image-x86-cuda-12-9
+        id: build-release-image-x86-cuda-13-0
        agents:
          queue: cpu_queue_release
        commands:
          - "aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin public.ecr.aws/q9t5s3a7"
-          - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=12.9.1 --build-arg torch_cuda_arch_list=\"${CUDA_ARCH_X86_CU129}\" --build-arg INSTALL_KV_CONNECTORS=true --tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)-cu129 --target vllm-openai --progress plain -f docker/Dockerfile ."
-          - "docker push public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)-cu129"
+          - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=13.0.1 --build-arg INSTALL_KV_CONNECTORS=true --build-arg BUILD_BASE_IMAGE=nvidia/cuda:13.0.1-devel-ubuntu22.04 --tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)-cu130 --target vllm-openai --progress plain -f docker/Dockerfile ."
+          - "docker push public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)-cu130"
          # re-tag to default image tag and push, just in case arm64 build fails
-          - "docker tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)-cu129 public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-cu129"
-          - "docker push public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-cu129"
+          - "docker tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)-cu130 public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-cu130"
+          - "docker push public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-cu130"

-      - label: "Build release image - aarch64 - CUDA 12.9"
+      - label: "Build release image - aarch64 - CUDA 13.0"
        depends_on: ~
-        id: build-release-image-arm64-cuda-12-9
+        id: build-release-image-arm64-cuda-13-0
        agents:
          queue: arm64_cpu_queue_release
        commands:
          - "aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin public.ecr.aws/q9t5s3a7"
-          - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=12.9.1 --build-arg torch_cuda_arch_list=\"${CUDA_ARCH_AARCH64_CU129}\" --build-arg INSTALL_KV_CONNECTORS=true --tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)-cu129 --target vllm-openai --progress plain -f docker/Dockerfile ."
-          - "docker push public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)-cu129"
+          # compute capability 12.0 for RTX-50 series / RTX PRO 6000 Blackwell, 12.1 for DGX Spark
+          - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=13.0.1 --build-arg torch_cuda_arch_list='8.7 8.9 9.0 10.0+PTX 12.0 12.1' --build-arg INSTALL_KV_CONNECTORS=true --build-arg BUILD_BASE_IMAGE=nvidia/cuda:13.0.1-devel-ubuntu22.04 --tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)-cu130 --target vllm-openai --progress plain -f docker/Dockerfile ."
+          - "docker push public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)-cu130"

-      - label: "Build release image - x86_64 - CUDA 13.0 - Ubuntu 24.04"
+      - label: "Build release image - x86_64 - CUDA 12.9 - Ubuntu 24.04"
        depends_on: ~
        id: build-release-image-x86-ubuntu2404
        agents:
          queue: cpu_queue_release
        commands:
          - "aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin public.ecr.aws/q9t5s3a7"
-          - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=13.0.2 --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-$(uname -m)-ubuntu2404 --target vllm-openai --progress plain -f docker/Dockerfile ."
+          - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=12.9.1 --build-arg UBUNTU_VERSION=24.04 --build-arg GDRCOPY_OS_VERSION=Ubuntu24_04 --build-arg FLASHINFER_AOT_COMPILE=true --build-arg torch_cuda_arch_list='8.7 8.9 9.0 10.0+PTX 12.0' --build-arg INSTALL_KV_CONNECTORS=true --tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)-ubuntu2404 --target vllm-openai --progress plain -f docker/Dockerfile ."
          - "docker push public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)-ubuntu2404"
          - "docker tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)-ubuntu2404 public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-ubuntu2404"
          - "docker push public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-ubuntu2404"

-      - label: "Build release image - aarch64 - CUDA 13.0 - Ubuntu 24.04"
+      - label: "Build release image - aarch64 - CUDA 12.9 - Ubuntu 24.04"
        depends_on: ~
        id: build-release-image-arm64-ubuntu2404
        agents:
          queue: arm64_cpu_queue_release
        commands:
          - "aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin public.ecr.aws/q9t5s3a7"
-          - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=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-$(uname -m)-ubuntu2404 --target vllm-openai --progress plain -f docker/Dockerfile ."
+          - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=12.9.1 --build-arg UBUNTU_VERSION=24.04 --build-arg GDRCOPY_OS_VERSION=Ubuntu24_04 --build-arg FLASHINFER_AOT_COMPILE=true --build-arg torch_cuda_arch_list='8.7 8.9 9.0 10.0+PTX 12.0' --build-arg INSTALL_KV_CONNECTORS=true --tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)-ubuntu2404 --target vllm-openai --progress plain -f docker/Dockerfile ."
          - "docker push public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)-ubuntu2404"

-      - label: "Build release image - x86_64 - CUDA 12.9 - Ubuntu 24.04"
+      - label: "Build release image - x86_64 - CUDA 13.0 - Ubuntu 24.04"
        depends_on: ~
-        id: build-release-image-x86-cuda-12-9-ubuntu2404
+        id: build-release-image-x86-cuda-13-0-ubuntu2404
        agents:
          queue: cpu_queue_release
        commands:
          - "aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin public.ecr.aws/q9t5s3a7"
-          - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=12.9.1 --build-arg UBUNTU_VERSION=24.04 --build-arg GDRCOPY_OS_VERSION=Ubuntu24_04 --build-arg torch_cuda_arch_list=\"${CUDA_ARCH_X86_CU129}\" --build-arg INSTALL_KV_CONNECTORS=true --tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)-cu129-ubuntu2404 --target vllm-openai --progress plain -f docker/Dockerfile ."
-          - "docker push public.ecr.aws/q9t5s3a7/vllm-release-repo:$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:$BUILDKITE_COMMIT-cu129-ubuntu2404"
+          - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=13.0.1 --build-arg UBUNTU_VERSION=24.04 --build-arg GDRCOPY_OS_VERSION=Ubuntu24_04 --build-arg FLASHINFER_AOT_COMPILE=true --build-arg torch_cuda_arch_list='8.7 8.9 9.0 10.0+PTX 12.0 12.1' --build-arg INSTALL_KV_CONNECTORS=true --build-arg BUILD_BASE_IMAGE=nvidia/cuda:13.0.1-devel-ubuntu24.04 --tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)-cu130-ubuntu2404 --target vllm-openai --progress plain -f docker/Dockerfile ."
+          - "docker push public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)-cu130-ubuntu2404"
+          - "docker tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)-cu130-ubuntu2404 public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-cu130-ubuntu2404"
+          - "docker push public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-cu130-ubuntu2404"

-      - label: "Build release image - aarch64 - CUDA 12.9 - Ubuntu 24.04"
+      - label: "Build release image - aarch64 - CUDA 13.0 - Ubuntu 24.04"
        depends_on: ~
-        id: build-release-image-arm64-cuda-12-9-ubuntu2404
+        id: build-release-image-arm64-cuda-13-0-ubuntu2404
        agents:
          queue: arm64_cpu_queue_release
        commands:
          - "aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin public.ecr.aws/q9t5s3a7"
-          - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=12.9.1 --build-arg UBUNTU_VERSION=24.04 --build-arg GDRCOPY_OS_VERSION=Ubuntu24_04 --build-arg torch_cuda_arch_list=\"${CUDA_ARCH_AARCH64_CU129}\" --build-arg INSTALL_KV_CONNECTORS=true --tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)-cu129-ubuntu2404 --target vllm-openai --progress plain -f docker/Dockerfile ."
-          - "docker push public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)-cu129-ubuntu2404"
+          - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=13.0.1 --build-arg UBUNTU_VERSION=24.04 --build-arg GDRCOPY_OS_VERSION=Ubuntu24_04 --build-arg FLASHINFER_AOT_COMPILE=true --build-arg torch_cuda_arch_list='8.7 8.9 9.0 10.0+PTX 12.0 12.1' --build-arg INSTALL_KV_CONNECTORS=true --build-arg BUILD_BASE_IMAGE=nvidia/cuda:13.0.1-devel-ubuntu24.04 --tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)-cu130-ubuntu2404 --target vllm-openai --progress plain -f docker/Dockerfile ."
+          - "docker push public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)-cu130-ubuntu2404"

      - block: "Build release image for x86_64 CPU"
        key: block-cpu-release-image-build
@@ -243,7 +238,7 @@ steps:
  - group: "Publish release images"
    key: "publish-release-images"
    steps:
-      - label: "Create multi-arch manifest - CUDA 13.0"
+      - label: "Create multi-arch manifest - CUDA 12.9"
        depends_on:
          - build-release-image-x86
          - build-release-image-arm64
@@ -255,7 +250,7 @@ steps:
          - "docker manifest create public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-x86_64 public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-aarch64 --amend"
          - "docker manifest push public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT"

-      - label: "Annotate release workflow - CUDA 13.0"
+      - label: "Annotate release workflow - CUDA 12.9"
        depends_on:
          - create-multi-arch-manifest
        id: annotate-release-workflow
@@ -264,19 +259,19 @@ steps:
        commands:
          - "bash .buildkite/scripts/annotate-release.sh"

-      - label: "Create multi-arch manifest - CUDA 12.9"
+      - label: "Create multi-arch manifest - CUDA 13.0"
        depends_on:
-          - build-release-image-x86-cuda-12-9
-          - build-release-image-arm64-cuda-12-9
-        id: create-multi-arch-manifest-cuda-12-9
+          - build-release-image-x86-cuda-13-0
+          - build-release-image-arm64-cuda-13-0
+        id: create-multi-arch-manifest-cuda-13-0
        agents:
          queue: small_cpu_queue_release
        commands:
          - "aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin public.ecr.aws/q9t5s3a7"
-          - "docker manifest create public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-cu129 public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-x86_64-cu129 public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-aarch64-cu129 --amend"
-          - "docker manifest push public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-cu129"
+          - "docker manifest create public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-cu130 public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-x86_64-cu130 public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-aarch64-cu130 --amend"
+          - "docker manifest push public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-cu130"

-      - label: "Create multi-arch manifest - CUDA 13.0 - Ubuntu 24.04"
+      - label: "Create multi-arch manifest - CUDA 12.9 - Ubuntu 24.04"
        depends_on:
          - build-release-image-x86-ubuntu2404
          - build-release-image-arm64-ubuntu2404
@@ -288,17 +283,17 @@ steps:
          - "docker manifest create public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-ubuntu2404 public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-x86_64-ubuntu2404 public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-aarch64-ubuntu2404 --amend"
          - "docker manifest push public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-ubuntu2404"

-      - label: "Create multi-arch manifest - CUDA 12.9 - Ubuntu 24.04"
+      - label: "Create multi-arch manifest - CUDA 13.0 - Ubuntu 24.04"
        depends_on:
-          - build-release-image-x86-cuda-12-9-ubuntu2404
-          - build-release-image-arm64-cuda-12-9-ubuntu2404
-        id: create-multi-arch-manifest-cuda-12-9-ubuntu2404
+          - build-release-image-x86-cuda-13-0-ubuntu2404
+          - build-release-image-arm64-cuda-13-0-ubuntu2404
+        id: create-multi-arch-manifest-cuda-13-0-ubuntu2404
        agents:
          queue: small_cpu_queue_release
        commands:
          - "aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin public.ecr.aws/q9t5s3a7"
-          - "docker manifest create public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-cu129-ubuntu2404 public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-x86_64-cu129-ubuntu2404 public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-aarch64-cu129-ubuntu2404 --amend"
-          - "docker manifest push public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-cu129-ubuntu2404"
+          - "docker manifest create public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-cu130-ubuntu2404 public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-x86_64-cu130-ubuntu2404 public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-aarch64-cu130-ubuntu2404 --amend"
+          - "docker manifest push public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-cu130-ubuntu2404"

      - label: "Publish nightly multi-arch image to DockerHub"
        depends_on:
@@ -318,16 +313,16 @@ steps:
          DOCKER_BUILDKIT: "1"
          DOCKERHUB_USERNAME: "vllmbot"

-      - label: "Publish nightly multi-arch image to DockerHub - CUDA 12.9"
+      - label: "Publish nightly multi-arch image to DockerHub - CUDA 13.0"
        depends_on:
-          - create-multi-arch-manifest-cuda-12-9
+          - create-multi-arch-manifest-cuda-13-0
        if: build.env("NIGHTLY") == "1"
        agents:
          queue: small_cpu_queue_release
        commands:
-          - "bash .buildkite/scripts/push-nightly-builds.sh cu129"
+          - "bash .buildkite/scripts/push-nightly-builds.sh cu130"
          # Clean up old nightly builds (keep only last 14)
-          - "bash .buildkite/scripts/cleanup-nightly-builds.sh cu129-nightly-"
+          - "bash .buildkite/scripts/cleanup-nightly-builds.sh cu130-nightly-"
        plugins:
          - docker-login#v3.0.0:
              username: vllmbot
@@ -13,12 +13,12 @@ ROCM_BASE_CACHE_KEY=$(.buildkite/scripts/cache-rocm-base-wheels.sh key)
 buildkite-agent annotate --style 'info' --context 'release-workflow' << EOF
 To download the wheel (by commit):
 \`\`\`
-aws s3 cp s3://vllm-wheels/${BUILDKITE_COMMIT}/vllm-${RELEASE_VERSION}-cp38-abi3-manylinux_2_35_x86_64.whl .
-aws s3 cp s3://vllm-wheels/${BUILDKITE_COMMIT}/vllm-${RELEASE_VERSION}-cp38-abi3-manylinux_2_35_aarch64.whl .
+aws s3 cp s3://vllm-wheels/${BUILDKITE_COMMIT}/vllm-${RELEASE_VERSION}-cp38-abi3-manylinux_2_31_x86_64.whl .
+aws s3 cp s3://vllm-wheels/${BUILDKITE_COMMIT}/vllm-${RELEASE_VERSION}-cp38-abi3-manylinux_2_31_aarch64.whl .

-(Optional) For CUDA 12.9:
-aws s3 cp s3://vllm-wheels/${BUILDKITE_COMMIT}/vllm-${RELEASE_VERSION}+cu129-cp38-abi3-manylinux_2_31_x86_64.whl .
-aws s3 cp s3://vllm-wheels/${BUILDKITE_COMMIT}/vllm-${RELEASE_VERSION}+cu129-cp38-abi3-manylinux_2_31_aarch64.whl .
+(Optional) For CUDA 13.0:
+aws s3 cp s3://vllm-wheels/${BUILDKITE_COMMIT}/vllm-${RELEASE_VERSION}+cu130-cp38-abi3-manylinux_2_35_x86_64.whl .
+aws s3 cp s3://vllm-wheels/${BUILDKITE_COMMIT}/vllm-${RELEASE_VERSION}+cu130-cp38-abi3-manylinux_2_35_aarch64.whl .

 (Optional) For CPU:
 aws s3 cp s3://vllm-wheels/${BUILDKITE_COMMIT}/vllm-${RELEASE_VERSION}+cpu-cp38-abi3-manylinux_2_35_x86_64.whl .
@@ -33,8 +33,8 @@ To download and upload the image:

 docker pull public.ecr.aws/q9t5s3a7/vllm-release-repo:${BUILDKITE_COMMIT}-x86_64
 docker pull public.ecr.aws/q9t5s3a7/vllm-release-repo:${BUILDKITE_COMMIT}-aarch64
-docker pull public.ecr.aws/q9t5s3a7/vllm-release-repo:${BUILDKITE_COMMIT}-x86_64-cu129
-docker pull public.ecr.aws/q9t5s3a7/vllm-release-repo:${BUILDKITE_COMMIT}-aarch64-cu129
+docker pull public.ecr.aws/q9t5s3a7/vllm-release-repo:${BUILDKITE_COMMIT}-x86_64-cu130
+docker pull public.ecr.aws/q9t5s3a7/vllm-release-repo:${BUILDKITE_COMMIT}-aarch64-cu130
 docker pull public.ecr.aws/q9t5s3a7/vllm-release-repo:${ROCM_BASE_CACHE_KEY}-rocm-base
 docker pull public.ecr.aws/q9t5s3a7/vllm-release-repo:${BUILDKITE_COMMIT}-rocm
 docker pull public.ecr.aws/q9t5s3a7/vllm-cpu-release-repo:v${RELEASE_VERSION}
@@ -50,11 +50,11 @@ docker tag vllm/vllm-openai:x86_64 vllm/vllm-openai:v${RELEASE_VERSION}-x86_64
 docker push vllm/vllm-openai:latest-x86_64
 docker push vllm/vllm-openai:v${RELEASE_VERSION}-x86_64

-docker tag public.ecr.aws/q9t5s3a7/vllm-release-repo:${BUILDKITE_COMMIT}-x86_64-cu129 vllm/vllm-openai:x86_64-cu129
-docker tag vllm/vllm-openai:x86_64-cu129 vllm/vllm-openai:latest-x86_64-cu129
-docker tag vllm/vllm-openai:x86_64-cu129 vllm/vllm-openai:v${RELEASE_VERSION}-x86_64-cu129
-docker push vllm/vllm-openai:latest-x86_64-cu129
-docker push vllm/vllm-openai:v${RELEASE_VERSION}-x86_64-cu129
+docker tag public.ecr.aws/q9t5s3a7/vllm-release-repo:${BUILDKITE_COMMIT}-x86_64-cu130 vllm/vllm-openai:x86_64-cu130
+docker tag vllm/vllm-openai:x86_64-cu130 vllm/vllm-openai:latest-x86_64-cu130
+docker tag vllm/vllm-openai:x86_64-cu130 vllm/vllm-openai:v${RELEASE_VERSION}-x86_64-cu130
+docker push vllm/vllm-openai:latest-x86_64-cu130
+docker push vllm/vllm-openai:v${RELEASE_VERSION}-x86_64-cu130

 docker tag public.ecr.aws/q9t5s3a7/vllm-release-repo:${BUILDKITE_COMMIT}-aarch64 vllm/vllm-openai:aarch64
 docker tag vllm/vllm-openai:aarch64 vllm/vllm-openai:latest-aarch64
@@ -62,11 +62,11 @@ docker tag vllm/vllm-openai:aarch64 vllm/vllm-openai:v${RELEASE_VERSION}-aarch64
 docker push vllm/vllm-openai:latest-aarch64
 docker push vllm/vllm-openai:v${RELEASE_VERSION}-aarch64

-docker tag public.ecr.aws/q9t5s3a7/vllm-release-repo:${BUILDKITE_COMMIT}-aarch64-cu129 vllm/vllm-openai:aarch64-cu129
-docker tag vllm/vllm-openai:aarch64-cu129 vllm/vllm-openai:latest-aarch64-cu129
-docker tag vllm/vllm-openai:aarch64-cu129 vllm/vllm-openai:v${RELEASE_VERSION}-aarch64-cu129
-docker push vllm/vllm-openai:latest-aarch64-cu129
-docker push vllm/vllm-openai:v${RELEASE_VERSION}-aarch64-cu129
+docker tag public.ecr.aws/q9t5s3a7/vllm-release-repo:${BUILDKITE_COMMIT}-aarch64-cu130 vllm/vllm-openai:aarch64-cu130
+docker tag vllm/vllm-openai:aarch64-cu130 vllm/vllm-openai:latest-aarch64-cu130
+docker tag vllm/vllm-openai:aarch64-cu130 vllm/vllm-openai:v${RELEASE_VERSION}-aarch64-cu130
+docker push vllm/vllm-openai:latest-aarch64-cu130
+docker push vllm/vllm-openai:v${RELEASE_VERSION}-aarch64-cu130

 ## ROCm

@@ -104,11 +104,11 @@ docker manifest create vllm/vllm-openai:v${RELEASE_VERSION} vllm/vllm-openai:v${
 docker manifest push vllm/vllm-openai:latest
 docker manifest push vllm/vllm-openai:v${RELEASE_VERSION}

-docker manifest rm vllm/vllm-openai:latest-cu129
-docker manifest create vllm/vllm-openai:latest-cu129 vllm/vllm-openai:latest-x86_64-cu129 vllm/vllm-openai:latest-aarch64-cu129
-docker manifest create vllm/vllm-openai:v${RELEASE_VERSION}-cu129 vllm/vllm-openai:v${RELEASE_VERSION}-x86_64-cu129 vllm/vllm-openai:v${RELEASE_VERSION}-aarch64-cu129
-docker manifest push vllm/vllm-openai:latest-cu129
-docker manifest push vllm/vllm-openai:v${RELEASE_VERSION}-cu129
+docker manifest rm vllm/vllm-openai:latest-cu130
+docker manifest create vllm/vllm-openai:latest-cu130 vllm/vllm-openai:latest-x86_64-cu130 vllm/vllm-openai:latest-aarch64-cu130
+docker manifest create vllm/vllm-openai:v${RELEASE_VERSION}-cu130 vllm/vllm-openai:v${RELEASE_VERSION}-x86_64-cu130 vllm/vllm-openai:v${RELEASE_VERSION}-aarch64-cu130
+docker manifest push vllm/vllm-openai:latest-cu130
+docker manifest push vllm/vllm-openai:v${RELEASE_VERSION}-cu130

 docker manifest rm vllm/vllm-openai-cpu:latest || true
 docker manifest create vllm/vllm-openai-cpu:latest vllm/vllm-openai-cpu:latest-x86_64 vllm/vllm-openai-cpu:latest-arm64
@@ -29,7 +29,7 @@ if python3 -c "import torch; assert torch.version.hip" 2>/dev/null; then
        TORCH_INDEX_URL=""
    fi
 else
-    TORCH_INDEX_URL="https://download.pytorch.org/whl/cu130"
+    TORCH_INDEX_URL="https://download.pytorch.org/whl/cu129"
 fi
 echo ">>> Using PyTorch index: ${TORCH_INDEX_URL:-PyPI default}"

@@ -9,7 +9,7 @@ set -ex

 BUCKET="vllm-wheels"
 INDICES_OUTPUT_DIR="indices"
-DEFAULT_VARIANT_ALIAS="cu130" # align with vLLM_MAIN_CUDA_VERSION in vllm/envs.py
+DEFAULT_VARIANT_ALIAS="cu129" # align with vLLM_MAIN_CUDA_VERSION in vllm/envs.py
 PYTHON="${PYTHON_PROG:-python3}" # try to read from env var, otherwise use python3
 SUBPATH=$BUILDKITE_COMMIT
 S3_COMMIT_PREFIX="s3://$BUCKET/$SUBPATH/"
@@ -25,100 +25,22 @@ export PYTHONPATH=".."
 ###############################################################################

 cleanup_docker() {
-  # Share the same lock with image pull to avoid cleanup/pull races on one node.
-  local docker_lock="/tmp/docker-pull.lock"
-  exec 9>"$docker_lock"
-  flock 9
-
  docker_root=$(docker info -f '{{.DockerRootDir}}')
  if [ -z "$docker_root" ]; then
    echo "Failed to determine Docker root directory." >&2
-    flock -u 9
-    return 1
+    exit 1
  fi
  echo "Docker root directory: $docker_root"

  disk_usage=$(df "$docker_root" | tail -1 | awk '{print $5}' | sed 's/%//')
  threshold=70
  if [ "$disk_usage" -gt "$threshold" ]; then
-    echo "Disk usage is above $threshold%. Running aggressive CI image cleanup..."
-    cleanup_old_ci_images "${REGISTRY}/${REPO}" "${image_name}" "${DOCKER_IMAGE_CLEANUP_HOURS:-72}" 1
+    echo "Disk usage is above $threshold%. Cleaning up Docker images and volumes..."
+    docker image prune -f
+    docker volume prune -f && docker system prune --force --filter "until=72h" --all
+    echo "Docker images and volumes cleanup completed."
  else
-    echo "Disk usage is below $threshold%. Checking old CI images anyway."
-    cleanup_old_ci_images "${REGISTRY}/${REPO}" "${image_name}" "${DOCKER_IMAGE_CLEANUP_HOURS:-72}" 0
-  fi
-  echo "Old CI image cleanup completed."
-
-  flock -u 9
-}
-
-cleanup_old_ci_images() {
-  local repo_prefix="$1"
-  local current_image_ref="$2"
-  local ttl_hours="$3"
-  local aggressive_cleanup="$4"
-
-  if [[ -z "$repo_prefix" || "$repo_prefix" == "/" ]]; then
-    echo "Skip old-image cleanup: invalid repo prefix '${repo_prefix}'"
-    return 0
-  fi
-
-  if ! [[ "$ttl_hours" =~ ^[0-9]+$ ]]; then
-    echo "Invalid DOCKER_IMAGE_CLEANUP_HOURS='${ttl_hours}', fallback to 72"
-    ttl_hours=72
-  fi
-
-  local now_epoch cutoff_epoch
-  now_epoch=$(date +%s)
-  cutoff_epoch=$((now_epoch - ttl_hours * 3600))
-
-  local -a used_image_ids
-  mapfile -t used_image_ids < <(docker ps -aq | xargs -r docker inspect --format '{{.Image}}' | sort -u)
-
-  local removed_count=0
-  local examined_count=0
-  declare -A seen_ids=()
-
-  while read -r image_ref image_id; do
-    [[ -z "$image_ref" || -z "$image_id" ]] && continue
-    ((examined_count++))
-
-    # Keep the image this job is going to use.
-    if [[ "$image_ref" == "$current_image_ref" ]]; then
-      continue
-    fi
-
-    # Avoid duplicate deletes when multiple tags point to same image id.
-    if [[ -n "${seen_ids[$image_id]:-}" ]]; then
-      continue
-    fi
-    seen_ids[$image_id]=1
-
-    # Never delete images that are used by any container on this node.
-    if printf '%s\n' "${used_image_ids[@]}" | grep -qx "$image_id"; then
-      continue
-    fi
-
-    local created created_epoch
-    created=$(docker image inspect -f '{{.Created}}' "$image_id" 2>/dev/null || true)
-    [[ -z "$created" ]] && continue
-    created_epoch=$(date -d "$created" +%s 2>/dev/null || true)
-    [[ -z "$created_epoch" ]] && continue
-
-    if (( created_epoch < cutoff_epoch )) || [[ "$aggressive_cleanup" == "1" ]]; then
-      if docker image rm -f "$image_id" >/dev/null 2>&1; then
-        ((removed_count++))
-      fi
-    fi
-  done < <(docker image ls --no-trunc "$repo_prefix" --format '{{.Repository}}:{{.Tag}} {{.ID}}')
-
-  # Also trim old dangling layers; this is safe and does not remove referenced images.
-  docker image prune -f --filter "until=${ttl_hours}h" >/dev/null 2>&1 || true
-
-  if [[ "$aggressive_cleanup" == "1" ]]; then
-    echo "Examined ${examined_count} images under ${repo_prefix}, removed ${removed_count} unused images under disk pressure."
-  else
-    echo "Examined ${examined_count} images under ${repo_prefix}, removed ${removed_count} old images (>${ttl_hours}h)."
+    echo "Disk usage is below $threshold%. No cleanup needed."
  fi
 }

@@ -343,6 +265,8 @@ fi

 remove_docker_container() {
  docker rm -f "${container_name}" || true
+  docker image rm -f "${image_name}" || true
+  docker system prune -f || true
 }
 trap remove_docker_container EXIT

@@ -12,7 +12,9 @@ docker build -t "${image_name}" -f docker/Dockerfile.xpu .

 # Setup cleanup
 remove_docker_container() {
-  docker rm -f "${container_name}" || true
+  docker rm -f "${container_name}" || true;
+  docker image rm -f "${image_name}" || true;
+  docker system prune -f || true;
 }
 trap remove_docker_container EXIT

@@ -1,98 +0,0 @@
-group: Disaggregated
-depends_on: 
-  - image-build
-steps:
- label: Distributed NixlConnector PD accuracy (4 GPUs)
-  timeout_in_minutes: 30
-  working_dir: "/vllm-workspace/tests"
-  num_devices: 4
-  source_file_dependencies:
-    - vllm/distributed/kv_transfer/kv_connector/v1/nixl_connector.py
-    - tests/v1/kv_connector/nixl_integration/
-  commands:
-    - uv pip install --system -r /vllm-workspace/requirements/kv_connectors.txt
-    - bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
- label: Distributed FlashInfer NixlConnector PD accuracy (4 GPUs)
-  timeout_in_minutes: 30
-  working_dir: "/vllm-workspace/tests"
-  num_devices: 4
-  source_file_dependencies:
-    - vllm/distributed/kv_transfer/kv_connector/v1/nixl_connector.py
-    - tests/v1/kv_connector/nixl_integration/
-  commands:
-    - uv pip install --system -r /vllm-workspace/requirements/kv_connectors.txt
-    - FLASHINFER=1 bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
-
- label: DP EP Distributed NixlConnector PD accuracy tests (4 GPUs)
-  timeout_in_minutes: 30
-  working_dir: "/vllm-workspace/tests"
-  num_devices: 4
-  source_file_dependencies:
-    - vllm/distributed/kv_transfer/kv_connector/v1/nixl_connector.py
-    - tests/v1/kv_connector/nixl_integration/
-  commands:
-    - uv pip install --system -r /vllm-workspace/requirements/kv_connectors.txt
-    - DP_EP=1 bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
-
- label: CrossLayer KV layout Distributed NixlConnector PD accuracy tests (4 GPUs)
-  timeout_in_minutes: 30
-  working_dir: "/vllm-workspace/tests"
-  num_devices: 4
-  source_file_dependencies:
-    - vllm/distributed/kv_transfer/kv_connector/v1/nixl_connector.py
-    - tests/v1/kv_connector/nixl_integration/
-  commands:
-    - uv pip install --system -r /vllm-workspace/requirements/kv_connectors.txt
-    - CROSS_LAYERS_BLOCKS=True bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
-
- label: Hybrid SSM NixlConnector PD accuracy tests (4 GPUs)
-  timeout_in_minutes: 20
-  working_dir: "/vllm-workspace/tests"
-  num_devices: 4
-  source_file_dependencies:
-    - vllm/distributed/kv_transfer/kv_connector/v1/nixl_connector.py
-    - tests/v1/kv_connector/nixl_integration/
-  commands:
-    - uv pip install --system -r /vllm-workspace/requirements/kv_connectors.txt
-    - HYBRID_SSM=1 bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
-
- label: MultiConnector (Nixl+Offloading) PD accuracy (2 GPUs)
-  timeout_in_minutes: 30
-  working_dir: "/vllm-workspace/tests"
-  num_devices: 2
-  source_file_dependencies:
-    - vllm/distributed/kv_transfer/kv_connector/v1/nixl_connector.py
-    - vllm/distributed/kv_transfer/kv_connector/v1/multi_connector.py
-    - vllm/distributed/kv_transfer/kv_connector/v1/offloading_connector.py
-    - vllm/distributed/kv_transfer/kv_connector/v1/offloading/
-    - tests/v1/kv_connector/nixl_integration/
-  commands:
-    - uv pip install --system -r /vllm-workspace/requirements/kv_connectors.txt
-    - bash v1/kv_connector/nixl_integration/run_multi_connector_accuracy_test.sh
-
- label: NixlConnector PD + Spec Decode acceptance (2 GPUs)
-  timeout_in_minutes: 30
-  device: a100
-  working_dir: "/vllm-workspace/tests"
-  num_devices: 2
-  source_file_dependencies:
-    - vllm/distributed/kv_transfer/kv_connector/v1/nixl_connector.py
-    - vllm/v1/worker/kv_connector_model_runner_mixin.py
-    - tests/v1/kv_connector/nixl_integration/
-  commands:
-    - uv pip install --system -r /vllm-workspace/requirements/kv_connectors.txt
-    - bash v1/kv_connector/nixl_integration/spec_decode_acceptance_test.sh
-
- label: MultiConnector (Nixl+Offloading) PD edge cases (2 GPUs)
-  timeout_in_minutes: 30
-  working_dir: "/vllm-workspace/tests"
-  num_devices: 2
-  source_file_dependencies:
-    - vllm/distributed/kv_transfer/kv_connector/v1/nixl_connector.py
-    - vllm/distributed/kv_transfer/kv_connector/v1/multi_connector.py
-    - vllm/distributed/kv_transfer/kv_connector/v1/offloading_connector.py
-    - vllm/distributed/kv_transfer/kv_connector/v1/offloading/
-    - tests/v1/kv_connector/nixl_integration/
-  commands:
-    - uv pip install --system -r /vllm-workspace/requirements/kv_connectors.txt
-    - bash v1/kv_connector/nixl_integration/run_multi_connector_edge_case_test.sh
@@ -226,6 +226,91 @@ steps:
  commands:
    - ./.buildkite/scripts/run-multi-node-test.sh /vllm-workspace/tests 2 2 $IMAGE_TAG "VLLM_TEST_SAME_HOST=0 torchrun --nnodes 2 --nproc-per-node=2 --rdzv_backend=c10d --rdzv_endpoint=192.168.10.10 distributed/test_same_node.py | grep 'Same node test passed' && NUM_NODES=2 torchrun --nnodes 2 --nproc-per-node=2 --rdzv_backend=c10d --rdzv_endpoint=192.168.10.10 distributed/test_node_count.py | grep 'Node count test passed' && python3 ../examples/offline_inference/data_parallel.py -dp=2 -tp=1 --dp-num-nodes=2 --dp-node-rank=0 --dp-master-addr=192.168.10.10 --dp-master-port=12345 --enforce-eager --trust-remote-code && VLLM_MULTI_NODE=1 pytest -v -s distributed/test_multi_node_assignment.py && VLLM_MULTI_NODE=1 pytest -v -s distributed/test_pipeline_parallel.py" "VLLM_TEST_SAME_HOST=0 torchrun --nnodes 2 --nproc-per-node=2 --rdzv_backend=c10d --rdzv_endpoint=192.168.10.10 distributed/test_same_node.py | grep 'Same node test passed' && NUM_NODES=2 torchrun --nnodes 2 --nproc-per-node=2 --rdzv_backend=c10d --rdzv_endpoint=192.168.10.10 distributed/test_node_count.py | grep 'Node count test passed' && python3 ../examples/offline_inference/data_parallel.py -dp=2 -tp=1 --dp-num-nodes=2 --dp-node-rank=1 --dp-master-addr=192.168.10.10 --dp-master-port=12345 --enforce-eager --trust-remote-code"

+- label: Distributed NixlConnector PD accuracy (4 GPUs)
+  timeout_in_minutes: 30
+  working_dir: "/vllm-workspace/tests"
+  num_devices: 4
+  source_file_dependencies:
+    - vllm/distributed/kv_transfer/kv_connector/v1/nixl_connector.py
+    - tests/v1/kv_connector/nixl_integration/
+  commands:
+    - uv pip install --system -r /vllm-workspace/requirements/kv_connectors.txt
+    - bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
+
+- label: DP EP Distributed NixlConnector PD accuracy tests (4 GPUs)
+  timeout_in_minutes: 30
+  working_dir: "/vllm-workspace/tests"
+  num_devices: 4
+  source_file_dependencies:
+    - vllm/distributed/kv_transfer/kv_connector/v1/nixl_connector.py
+    - tests/v1/kv_connector/nixl_integration/
+  commands:
+    - uv pip install --system -r /vllm-workspace/requirements/kv_connectors.txt
+    - DP_EP=1 bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
+
+- label: CrossLayer KV layout Distributed NixlConnector PD accuracy tests (4 GPUs)
+  timeout_in_minutes: 30
+  working_dir: "/vllm-workspace/tests"
+  num_devices: 4
+  source_file_dependencies:
+    - vllm/distributed/kv_transfer/kv_connector/v1/nixl_connector.py
+    - tests/v1/kv_connector/nixl_integration/
+  commands:
+    - uv pip install --system -r /vllm-workspace/requirements/kv_connectors.txt
+    - CROSS_LAYERS_BLOCKS=True bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
+
+- label: Hyrbid SSM NixlConnector PD accuracy tests (4 GPUs)
+  timeout_in_minutes: 20
+  working_dir: "/vllm-workspace/tests"
+  num_devices: 4
+  source_file_dependencies:
+    - vllm/distributed/kv_transfer/kv_connector/v1/nixl_connector.py
+    - tests/v1/kv_connector/nixl_integration/
+  commands:
+    - uv pip install --system -r /vllm-workspace/requirements/kv_connectors.txt
+    - HYBRID_SSM=1 bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
+
+- label: MultiConnector (Nixl+Offloading) PD accuracy (2 GPUs)
+  timeout_in_minutes: 30
+  working_dir: "/vllm-workspace/tests"
+  num_devices: 2
+  source_file_dependencies:
+    - vllm/distributed/kv_transfer/kv_connector/v1/nixl_connector.py
+    - vllm/distributed/kv_transfer/kv_connector/v1/multi_connector.py
+    - vllm/distributed/kv_transfer/kv_connector/v1/offloading_connector.py
+    - vllm/distributed/kv_transfer/kv_connector/v1/offloading/
+    - tests/v1/kv_connector/nixl_integration/
+  commands:
+    - uv pip install --system -r /vllm-workspace/requirements/kv_connectors.txt
+    - bash v1/kv_connector/nixl_integration/run_multi_connector_accuracy_test.sh
+
+- label: NixlConnector PD + Spec Decode acceptance (2 GPUs)
+  timeout_in_minutes: 30
+  device: a100
+  working_dir: "/vllm-workspace/tests"
+  num_devices: 2
+  source_file_dependencies:
+    - vllm/distributed/kv_transfer/kv_connector/v1/nixl_connector.py
+    - vllm/v1/worker/kv_connector_model_runner_mixin.py
+    - tests/v1/kv_connector/nixl_integration/
+  commands:
+    - uv pip install --system -r /vllm-workspace/requirements/kv_connectors.txt
+    - bash v1/kv_connector/nixl_integration/spec_decode_acceptance_test.sh
+
+- label: MultiConnector (Nixl+Offloading) PD edge cases (2 GPUs)
+  timeout_in_minutes: 30
+  working_dir: "/vllm-workspace/tests"
+  num_devices: 2
+  source_file_dependencies:
+    - vllm/distributed/kv_transfer/kv_connector/v1/nixl_connector.py
+    - vllm/distributed/kv_transfer/kv_connector/v1/multi_connector.py
+    - vllm/distributed/kv_transfer/kv_connector/v1/offloading_connector.py
+    - vllm/distributed/kv_transfer/kv_connector/v1/offloading/
+    - tests/v1/kv_connector/nixl_integration/
+  commands:
+    - uv pip install --system -r /vllm-workspace/requirements/kv_connectors.txt
+    - bash v1/kv_connector/nixl_integration/run_multi_connector_edge_case_test.sh
+
 - label: Pipeline + Context Parallelism (4 GPUs)
  timeout_in_minutes: 60
  working_dir: "/vllm-workspace/tests"
@@ -28,7 +28,6 @@ steps:
    - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
    - pytest -v -s models/multimodal/generation/test_common.py -m core_model -k "qwen3 or gemma"
    - pytest -v -s models/multimodal/generation/test_qwen2_5_vl.py -m core_model
-    - pytest -v -s models/multimodal/generation/test_vit_cudagraph.py -m core_model
  mirror:
    amd:
      device: mi325_1
@@ -262,7 +262,7 @@ pull_request_rules:
      - files~=^docker/Dockerfile.xpu
      - files~=^\\.buildkite/intel_jobs/
      - files=\.buildkite/ci_config_intel.yaml
-      - files=vllm/model_executor/layers/fused_moe/experts/xpu_moe.py
+      - files=vllm/model_executor/layers/fused_moe/xpu_fused_moe.py
      - files=vllm/model_executor/kernels/linear/mixed_precision/xpu.py
      - files=vllm/model_executor/kernels/linear/mxfp8/xpu.py
      - files=vllm/model_executor/kernels/linear/scaled_mm/xpu.py
@@ -14,7 +14,7 @@ $python_executable -m pip install -r requirements/build/cuda.txt -r requirements
 # Limit the number of parallel jobs to avoid OOM
 export MAX_JOBS=1
 # Make sure release wheels are built for the following architectures
-export TORCH_CUDA_ARCH_LIST="7.5 8.0 8.6 8.9 9.0 10.0 12.0+PTX"
+export TORCH_CUDA_ARCH_LIST="7.0 7.5 8.0 8.6 8.9 9.0+PTX"

 bash tools/check_repo.sh

@@ -34,10 +34,10 @@ install(CODE "set(CMAKE_INSTALL_LOCAL_ONLY TRUE)" ALL_COMPONENTS)
 # Supported python versions.  These versions will be searched in order, the
 # first match will be selected.  These should be kept in sync with setup.py.
 #
-set(PYTHON_SUPPORTED_VERSIONS "3.10" "3.11" "3.12" "3.13" "3.14")
+set(PYTHON_SUPPORTED_VERSIONS "3.10" "3.11" "3.12" "3.13")

 # Supported AMD GPU architectures.
-set(HIP_SUPPORTED_ARCHS "gfx906;gfx908;gfx90a;gfx942;gfx950;gfx1030;gfx1100;gfx1101;gfx1102;gfx1103;gfx1150;gfx1151;gfx1152;gfx1153;gfx1200;gfx1201")
+set(HIP_SUPPORTED_ARCHS "gfx906;gfx908;gfx90a;gfx942;gfx950;gfx1030;gfx1100;gfx1101;gfx1150;gfx1151;gfx1152;gfx1153;gfx1200;gfx1201")

 # ROCm installation prefix. Default to /opt/rocm but allow override via
 # -DROCM_PATH=/your/rocm/path when invoking cmake.
@@ -94,15 +94,12 @@ find_package(Torch REQUIRED)
 # This check must happen after find_package(Torch) because that's when CMAKE_CUDA_COMPILER_VERSION gets defined
 if(DEFINED CMAKE_CUDA_COMPILER_VERSION AND
   CMAKE_CUDA_COMPILER_VERSION VERSION_GREATER_EQUAL 13.0)
-   # starting from CUDA 12.9 and Blackwell (10.0), we use family-specific targets (10.0f, 12.0f, etc)
-   # to support the whole generation without specifying all sub-architectures
-   # see: https://developer.nvidia.com/blog/nvidia-blackwell-and-nvidia-cuda-12-9-introduce-family-specific-architecture-features/
-  set(CUDA_SUPPORTED_ARCHS "7.5;8.0;8.6;8.7;8.9;9.0;10.0;11.0;12.0")
+  set(CUDA_SUPPORTED_ARCHS "7.5;8.0;8.6;8.7;8.9;9.0;10.0;11.0;12.0;12.1")
 elseif(DEFINED CMAKE_CUDA_COMPILER_VERSION AND
   CMAKE_CUDA_COMPILER_VERSION VERSION_GREATER_EQUAL 12.8)
-  set(CUDA_SUPPORTED_ARCHS "7.5;8.0;8.6;8.7;8.9;9.0;10.0;10.1;10.3;12.0;12.1")
+  set(CUDA_SUPPORTED_ARCHS "7.0;7.2;7.5;8.0;8.6;8.7;8.9;9.0;10.0;10.1;12.0;12.1")
 else()
-  set(CUDA_SUPPORTED_ARCHS "7.0;7.5;8.0;8.6;8.7;8.9;9.0")
+  set(CUDA_SUPPORTED_ARCHS "7.0;7.2;7.5;8.0;8.6;8.7;8.9;9.0")
 endif()

 #
@@ -310,9 +307,7 @@ set(VLLM_EXT_SRC
  "csrc/torch_bindings.cpp")

 if(VLLM_GPU_LANG STREQUAL "CUDA")
-  list(APPEND VLLM_EXT_SRC
-    "csrc/minimax_reduce_rms_kernel.cu"
-    "csrc/fused_deepseek_v4_qnorm_rope_kv_insert_kernel.cu")
+  list(APPEND VLLM_EXT_SRC "csrc/minimax_reduce_rms_kernel.cu")

  SET(CUTLASS_ENABLE_HEADERS_ONLY ON CACHE BOOL "Enable only the header library")

@@ -1053,8 +1048,7 @@ 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")
+    "csrc/moe/router_gemm.cu")
 endif()

 if(VLLM_GPU_LANG STREQUAL "CUDA")
@@ -14,7 +14,7 @@ Easy, fast, and cheap LLM serving for everyone
 | <a href="https://docs.vllm.ai"><b>Documentation</b></a> | <a href="https://blog.vllm.ai/"><b>Blog</b></a> | <a href="https://arxiv.org/abs/2309.06180"><b>Paper</b></a> | <a href="https://x.com/vllm_project"><b>Twitter/X</b></a> | <a href="https://discuss.vllm.ai"><b>User Forum</b></a> | <a href="https://slack.vllm.ai"><b>Developer Slack</b></a> |
 </p>

-🔥 We have built a vLLM website to help you get started with vLLM. Please visit [vllm.ai](https://vllm.ai) to learn more.
+🔥 We have built a vllm website to help you get started with vllm. Please visit [vllm.ai](https://vllm.ai) to learn more.
 For events, please visit [vllm.ai/events](https://vllm.ai/events) to join us.

 ---
@@ -50,7 +50,7 @@ vLLM is flexible and easy to use with:
 - Efficient multi-LoRA support for dense and MoE layers
 - Support for NVIDIA GPUs, AMD GPUs, and x86/ARM/PowerPC CPUs. Additionally, diverse hardware plugins such as Google TPUs, Intel Gaudi, IBM Spyre, Huawei Ascend, Rebellions NPU, Apple Silicon, MetaX GPU, and more.

-vLLM seamlessly supports 200+ model architectures on Hugging Face, including:
+vLLM seamlessly supports 200+ model architectures on HuggingFace, including:

 - Decoder-only LLMs (e.g., Llama, Qwen, Gemma)
 - Mixture-of-Expert LLMs (e.g., Mixtral, DeepSeek-V3, Qwen-MoE, GPT-OSS)
@@ -404,7 +404,6 @@ def _build_attention_metadata(
        query_start_loc=q_start_gpu,
        query_start_loc_cpu=q_start_cpu,
        seq_lens=seq_lens_gpu,
-        seq_lens_cpu_upper_bound=seq_lens_cpu,
        _seq_lens_cpu=seq_lens_cpu,
        _num_computed_tokens_cpu=num_computed_tokens_cpu,
        slot_mapping=slot_mapping,
@@ -16,7 +16,7 @@ from vllm.model_executor.layers.fused_moe.all2all_utils import (
    maybe_make_prepare_finalize,
 )
 from vllm.model_executor.layers.fused_moe.config import fp8_w8a8_moe_quant_config
-from vllm.model_executor.layers.fused_moe.experts.cutlass_moe import CutlassExpertsFp8
+from vllm.model_executor.layers.fused_moe.cutlass_moe import CutlassExpertsFp8
 from vllm.model_executor.layers.fused_moe.fused_moe import fused_experts, fused_topk
 from vllm.platforms import current_platform
 from vllm.utils.argparse_utils import FlexibleArgumentParser
@@ -22,7 +22,7 @@ from vllm.model_executor.layers.fused_moe.config import (
    fp8_w8a8_moe_quant_config,
    nvfp4_moe_quant_config,
 )
-from vllm.model_executor.layers.fused_moe.experts.cutlass_moe import (
+from vllm.model_executor.layers.fused_moe.cutlass_moe import (
    CutlassExpertsFp4,
 )
 from vllm.model_executor.layers.fused_moe.fused_moe import fused_experts, fused_topk
@@ -13,7 +13,7 @@ from vllm.model_executor.layers.fused_moe.all2all_utils import (
    maybe_make_prepare_finalize,
 )
 from vllm.model_executor.layers.fused_moe.config import fp8_w8a8_moe_quant_config
-from vllm.model_executor.layers.fused_moe.experts.cutlass_moe import CutlassExpertsFp8
+from vllm.model_executor.layers.fused_moe.cutlass_moe import CutlassExpertsFp8
 from vllm.model_executor.layers.fused_moe.fused_moe import (
    fused_experts,
    fused_topk,
@@ -20,7 +20,7 @@ else()
  FetchContent_Declare(
    deepgemm
    GIT_REPOSITORY https://github.com/deepseek-ai/DeepGEMM.git
-    GIT_TAG 891d57b4db1071624b5c8fa0d1e51cb317fa709f
+    GIT_TAG 477618cd51baffca09c4b0b87e97c03fe827ef03
    GIT_SUBMODULES "third-party/cutlass" "third-party/fmt"
    GIT_PROGRESS TRUE
    CONFIGURE_COMMAND ""
@@ -120,11 +120,6 @@ if(DEEPGEMM_ARCHS)
    COMPONENT _deep_gemm_C
    FILES_MATCHING PATTERN "*.py")

-  install(DIRECTORY "${deepgemm_SOURCE_DIR}/deep_gemm/mega/"
-    DESTINATION vllm/third_party/deep_gemm/mega
-    COMPONENT _deep_gemm_C
-    FILES_MATCHING PATTERN "*.py")
-
  # Generate envs.py (normally generated by DeepGEMM's setup.py build step)
  file(WRITE "${CMAKE_CURRENT_BINARY_DIR}/deep_gemm_envs.py"
    "# Pre-installed environment variables\npersistent_envs = dict()\n")
@@ -19,7 +19,7 @@ else()
  FetchContent_Declare(
        flashmla
        GIT_REPOSITORY https://github.com/vllm-project/FlashMLA
-        GIT_TAG a6ec2ba7bd0a7dff98b3f4d3e6b52b159c48d78b
+        GIT_TAG 692917b1cda61b93ac9ee2d846ec54e75afe87b1
        GIT_PROGRESS TRUE
        CONFIGURE_COMMAND ""
        BUILD_COMMAND ""
@@ -11,74 +11,29 @@
 namespace vllm {

 template <typename scalar_t, scalar_t (*ACT_FN)(const scalar_t&),
-          bool act_first, bool HAS_CLAMP>
+          bool act_first>
 __device__ __forceinline__ scalar_t compute(const scalar_t& x,
-                                            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);
-  }
+                                            const scalar_t& y) {
+  return act_first ? ACT_FN(x) * y : x * ACT_FN(y);
 }

 template <typename packed_t, packed_t (*PACKED_ACT_FN)(const packed_t&),
-          bool act_first, bool HAS_CLAMP>
+          bool act_first>
 __device__ __forceinline__ packed_t packed_compute(const packed_t& x,
-                                                   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));
-  }
+                                                   const packed_t& y) {
+  return act_first ? packed_mul(PACKED_ACT_FN(x), y)
+                   : packed_mul(x, PACKED_ACT_FN(y));
 }

 // 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 HAS_CLAMP, bool use_256b = false>
+          bool use_vec, 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 float limit) {
+    const int d) {
  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;
@@ -103,9 +58,8 @@ __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, HAS_CLAMP>(
-                x.elts[j], y.elts[j], limit);
+        x.elts[j] = packed_compute<packed_t, PACKED_ACT_FN, act_first>(
+            x.elts[j], y.elts[j]);
      }
      if constexpr (use_256b) {
        st256(x, &out_vec[i]);
@@ -118,8 +72,7 @@ __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, HAS_CLAMP>(x, y, limit);
+      out_ptr[idx] = compute<scalar_t, ACT_FN, act_first>(x, y);
    }
  }
 }
@@ -198,11 +151,8 @@ 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. 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)                        \
+// first.
+#define LAUNCH_ACTIVATION_GATE_KERNEL(KERNEL, PACKED_KERNEL, ACT_FIRST)        \
  auto dtype = input.scalar_type();                                            \
  int d = input.size(-1) / 2;                                                  \
  int64_t num_tokens = input.numel() / input.size(-1);                         \
@@ -227,8 +177,8 @@ 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, HAS_CLAMP, true><<<grid, block, 0, stream>>>(     \
-            out.data_ptr<scalar_t>(), input.data_ptr<scalar_t>(), d, LIMIT);   \
+            ACT_FIRST, true, true><<<grid, block, 0, stream>>>(                \
+            out.data_ptr<scalar_t>(), input.data_ptr<scalar_t>(), d);          \
      });                                                                      \
    } else {                                                                   \
      VLLM_DISPATCH_FLOATING_TYPES(dtype, "act_and_mul_kernel", [&] {          \
@@ -236,8 +186,8 @@ 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, HAS_CLAMP, false><<<grid, block, 0, stream>>>(    \
-            out.data_ptr<scalar_t>(), input.data_ptr<scalar_t>(), d, LIMIT);   \
+            ACT_FIRST, true, false><<<grid, block, 0, stream>>>(               \
+            out.data_ptr<scalar_t>(), input.data_ptr<scalar_t>(), d);          \
      });                                                                      \
    }                                                                          \
  } else {                                                                     \
@@ -247,8 +197,8 @@ 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, HAS_CLAMP><<<grid, block, 0, stream>>>(            \
-          out.data_ptr<scalar_t>(), input.data_ptr<scalar_t>(), d, LIMIT);     \
+          ACT_FIRST, false><<<grid, block, 0, stream>>>(                       \
+          out.data_ptr<scalar_t>(), input.data_ptr<scalar_t>(), d);            \
    });                                                                        \
  }

@@ -256,14 +206,7 @@ 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, 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);
+                                true);
 }

 void mul_and_silu(torch::Tensor& out,    // [..., d]
@@ -272,21 +215,21 @@ 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, 0.0f);
+                                false);
 }

 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, false, 0.0f);
+                                true);
 }

 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, false, 0.0f);
+  LAUNCH_ACTIVATION_GATE_KERNEL(vllm::gelu_tanh_kernel,
+                                vllm::packed_gelu_tanh_kernel, true);
 }

 namespace vllm {
@@ -599,11 +599,6 @@ __global__ void cp_gather_indexer_k_quant_cache_kernel(
  const int head_idx = (blockIdx.y * blockDim.x + threadIdx.x) * VEC_SIZE;
  // Find batch index within a block
  __shared__ int batch_idx[BLOCK_Y_SIZE];
-  if (threadIdx.x == 0) {
-    batch_idx[threadIdx.y] = -1;
-  }
-  __syncthreads();
-
  for (int iter = 0; iter < cuda_utils::ceil_div(batch_size, int(blockDim.x));
       iter++) {
    int tid = iter * blockDim.x + threadIdx.x;
@@ -616,18 +611,16 @@ __global__ void cp_gather_indexer_k_quant_cache_kernel(
    }
  }

-  __syncthreads();
+#ifndef USE_ROCM
+  __syncwarp();
+#endif

-  // num_tokens may be an allocation upper bound when Python avoids a D2H sync.
-  // Only tokens covered by the exact device-side cu_seq_lens are valid to
-  // gather.
-  const int batch = batch_idx[threadIdx.y];
-  if (head_idx >= head_dim || token_idx >= num_tokens || batch < 0) {
+  if (head_idx >= head_dim || token_idx >= num_tokens) {
    return;
  }
-  const int inbatch_seq_idx = token_idx - cu_seq_lens[batch];
-  const int block_idx =
-      block_table[batch * num_blocks + inbatch_seq_idx / cache_block_size];
+  const int inbatch_seq_idx = token_idx - cu_seq_lens[batch_idx[threadIdx.y]];
+  const int block_idx = block_table[batch_idx[threadIdx.y] * num_blocks +
+                                    inbatch_seq_idx / cache_block_size];
  const int64_t src_block_offset = block_idx * block_stride;
  const int64_t cache_inblock_offset =
      (inbatch_seq_idx % cache_block_size) * head_dim + head_idx;
@@ -1497,9 +1490,6 @@ void concat_mla_q(torch::Tensor& ql_nope,  // [num_tokens, num_heads, nope_dim]
  TORCH_CHECK(ql_nope.stride(2) == 1, "ql_nope must have stride 1 in dim 2");
  TORCH_CHECK(q_pe.stride(2) == 1, "q_pe must have stride 1 in dim 2");
  TORCH_CHECK(q_out.stride(2) == 1, "q_out must have stride 1 in dim 2");
-  TORCH_CHECK(ql_nope.scalar_type() == at::ScalarType::Half ||
-                  ql_nope.scalar_type() == at::ScalarType::BFloat16,
-              "ql_nope must be float16 or bfloat16 dtype");

  if (num_tokens == 0) return;

@@ -1511,7 +1501,7 @@ void concat_mla_q(torch::Tensor& ql_nope,  // [num_tokens, num_heads, nope_dim]
  const at::cuda::OptionalCUDAGuard device_guard(device_of(ql_nope));
  const cudaStream_t stream = at::cuda::getCurrentCUDAStream();

-  VLLM_DISPATCH_HALF_TYPES(ql_nope.scalar_type(), "concat_mla_q", [&] {
+  VLLM_DISPATCH_FLOATING_TYPES(ql_nope.scalar_type(), "concat_mla_q", [&] {
    vllm::ConcatMLAQKernel<scalar_t, 512><<<grid_size, block_size, 0, stream>>>(
        q_out.data_ptr<scalar_t>(), ql_nope.data_ptr<scalar_t>(),
        q_pe.data_ptr<scalar_t>(), num_tokens, num_heads, q_out.stride(0),
@@ -61,23 +61,8 @@
 #endif

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

 #endif  // __aarch64__

@@ -1152,11 +1152,7 @@ class AttentionMainLoop {
                       bool use_sink) {
 #ifdef DEFINE_FAST_EXP
      DEFINE_FAST_EXP
-      bool constexpr IsReducedPrecision =
-          std::is_same_v<query_t, c10::BFloat16> ||
-          std::is_same_v<query_t, c10::Half>;
 #endif
-
      using prob_buffer_vec_t = typename VecTypeTrait<prob_buffer_t>::vec_t;
      static_assert(sizeof(prob_buffer_t) <= sizeof(logits_buffer_t));

@@ -1205,17 +1201,8 @@ class AttentionMainLoop {
            vec = vec - max_vec;

            // compute exp
-
-#if defined(DEFINE_FAST_EXP)
-  #ifdef __aarch64__
-            if constexpr (IsReducedPrecision) {
-              vec = fast_exp_f16(vec);
-            } else
-  #endif
-            {
-              vec = fast_exp(vec);
-            }
-
+#ifdef DEFINE_FAST_EXP
+            vec = fast_exp(vec);
            prob_buffer_vec_t output_vec(vec);
            output_vec.save(curr_prob_buffer_iter);
 #else
@@ -1271,11 +1258,7 @@ class AttentionMainLoop {
                       int32_t kv_tile_token_num, float softcap_scale) {
 #ifdef DEFINE_FAST_EXP
      DEFINE_FAST_EXP
-      bool constexpr IsReducedPrecision =
-          std::is_same_v<query_t, c10::BFloat16> ||
-          std::is_same_v<query_t, c10::Half>;
 #endif
-
      float inv_softcap_scale = 1.0 / softcap_scale;
      vec_op::FP32Vec16 softcap_scale_vec(softcap_scale);
      vec_op::FP32Vec16 inv_softcap_scale_vec(inv_softcap_scale);
@@ -1289,15 +1272,8 @@ class AttentionMainLoop {
          vec_op::FP32Vec16 vec(curr_logits_buffer_iter);
          vec = vec * inv_softcap_scale_vec;

-#if defined(DEFINE_FAST_EXP)
-  #ifdef __aarch64__
-          if constexpr (IsReducedPrecision) {
-            vec = fast_exp_f16(vec);
-          } else
-  #endif
-          {
-            vec = fast_exp(vec);
-          }
+#ifdef DEFINE_FAST_EXP
+          vec = fast_exp(vec);
          vec_op::FP32Vec16 inv_vec = ones_vec / vec;
          vec = (vec - inv_vec) / (vec + inv_vec);
 #else
@@ -178,12 +178,7 @@ void rotary_embedding_gptj_impl(

 void rotary_embedding(torch::Tensor& positions, torch::Tensor& query,
                      std::optional<torch::Tensor> key, int64_t head_size,
-                      torch::Tensor& cos_sin_cache, bool is_neox,
-                      int64_t rope_dim_offset, bool inverse) {
-  TORCH_CHECK(rope_dim_offset == 0,
-              "rope_dim_offset != 0 is not supported on CPU");
-  TORCH_CHECK(!inverse, "inverse rotary embedding is not supported on CPU");
-
+                      torch::Tensor& cos_sin_cache, bool is_neox) {
  int num_tokens = positions.numel();
  int rot_dim = cos_sin_cache.size(1);
  int num_heads = query.size(-1) / head_size;
@@ -263,8 +263,7 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
  ops.def(
      "rotary_embedding(Tensor positions, Tensor! query,"
      "                 Tensor!? key, int head_size,"
-      "                 Tensor cos_sin_cache, bool is_neox, int "
-      "rope_dim_offset=0, bool inverse=False) -> ()");
+      "                 Tensor cos_sin_cache, bool is_neox) -> ()");
  ops.impl("rotary_embedding", torch::kCPU, &rotary_embedding);

  // Quantization
@@ -1,477 +0,0 @@
-/*
- * SPDX-License-Identifier: Apache-2.0
- * SPDX-FileCopyrightText: Copyright contributors to the vLLM project
- *
- * Horizontally-fused DeepseekV4-MLA kernel:
- *   - Q side:  per-head RMSNorm (no weight) + GPT-J RoPE on last ROPE_DIM
- *   - KV side: GPT-J RoPE on last ROPE_DIM + UE8M0 FP8 quant on NoPE + paged
- *              cache insert
- *
- * Structured after `applyMLARopeAndAssignQKVKernelGeneration` in
- * TensorRT-LLM's mlaKernels.cu: one kernel, one grid, with head-slot
- * dispatch choosing Q vs KV work per warp.  The per-warp RMSNorm/RoPE
- * skeleton is adapted from vllm-deepseek_v4's existing
- * `fusedQKNormRopeKernel` (csrc/fused_qknorm_rope_kernel.cu).
- *
- * Assumptions (hard-coded for DeepseekV4 attention):
- *   HEAD_DIM  = 512
- *   ROPE_DIM  = 64   (RoPE applied to dims [NOPE_DIM, HEAD_DIM))
- *   NOPE_DIM  = 448
- *   QUANT_BLOCK = 64 (UE8M0 FP8 quant block)
- *   FP8_MAX   = 448.0f
- *   is_neox=false (GPT-J interleaved pairs)
- *   cos_sin_cache layout [max_pos, rope_dim] = cos || sin (cos first, sin
- *     second along last dim; each half is rope_dim/2 = 32 values)
- *
- * Cache layout per paged-cache block (block_size tokens):
- *   [0,            bs*576):          token data, 448 fp8 + 128 bf16 each
- *   [bs*576,       bs*576 + bs*8):   UE8M0 scales, 7 real + 1 pad per token
- */
-
-#include <cmath>
-#include <cuda_fp8.h>
-#include <cuda_runtime.h>
-#include <type_traits>
-
-#include <ATen/cuda/CUDAContext.h>
-#include <c10/cuda/CUDAGuard.h>
-#include <torch/cuda.h>
-
-#include "cuda_compat.h"
-#include "dispatch_utils.h"
-#include "type_convert.cuh"
-
-#ifndef FINAL_MASK
-  #define FINAL_MASK 0xffffffffu
-#endif
-
-namespace vllm {
-namespace deepseek_v4_fused_ops {
-
-namespace {
-inline int getSMVersion() {
-  auto* props = at::cuda::getCurrentDeviceProperties();
-  return props->major * 10 + props->minor;
-}
-}  // namespace
-
-// ────────────────────────────────────────────────────────────────────────────
-// Constants
-// ────────────────────────────────────────────────────────────────────────────
-constexpr int kHeadDim = 512;
-constexpr int kRopeDim = 64;
-constexpr int kNopeDim = kHeadDim - kRopeDim;  // 448
-constexpr int kQuantBlock = 64;
-constexpr int kNumQuantBlocks = kNopeDim / kQuantBlock;   // 7
-constexpr int kScaleBytesPerToken = kNumQuantBlocks + 1;  // 8 (7 real + 1 pad)
-constexpr int kTokenDataBytes = kNopeDim + kRopeDim * 2;  // 448 + 128 = 576
-constexpr float kFp8Max = 448.0f;
-
-// Per-warp layout:  32 lanes × 16 elems/lane = 512 elems = HEAD_DIM.
-constexpr int kNumLanes = 32;
-constexpr int kElemsPerLane = kHeadDim / kNumLanes;  // 16
-
-// ────────────────────────────────────────────────────────────────────────────
-// Small inline helpers
-// ────────────────────────────────────────────────────────────────────────────
-__device__ __forceinline__ float warp4MaxAbs(float val) {
-  // Reduce absolute max across 4 consecutive lanes (lane id & 3 group).
-  float peer = __shfl_xor_sync(FINAL_MASK, val, 1);
-  val = fmaxf(val, peer);
-  peer = __shfl_xor_sync(FINAL_MASK, val, 2);
-  val = fmaxf(val, peer);
-  return val;
-}
-
-template <typename T>
-__device__ __forceinline__ float warpSum(float val) {
-#pragma unroll
-  for (int mask = 16; mask > 0; mask >>= 1) {
-    val += __shfl_xor_sync(FINAL_MASK, val, mask, 32);
-  }
-  return val;
-}
-
-// ────────────────────────────────────────────────────────────────────────────
-// Kernel
-// ────────────────────────────────────────────────────────────────────────────
-//
-// Grid: 1D, gridDim.x = ceil(num_tokens_full * (num_heads_q + 1) /
-// warps_per_block) Block: blockDim.x = 256 threads (8 warps per block) Each
-// warp handles one (token, head_slot) pair. head_slot < num_heads_q          →
-// Q branch (RMSNorm + RoPE, in place) head_slot == num_heads_q         → KV
-// branch (RoPE + UE8M0 quant + insert)
-//
-// With DP padding, q/kv/position_ids can have more rows than slot_mapping.
-// The Q branch covers all `num_tokens_full` rows (downstream attention uses
-// them).  The KV branch only inserts the first `num_tokens_insert` tokens
-// (= slot_mapping length) into the paged cache.
-//
-template <typename scalar_t_in>
-__global__ void fusedDeepseekV4QNormRopeKVRopeQuantInsertKernel(
-    scalar_t_in* __restrict__ q_inout,         // [N, H, 512] bf16, in place
-    scalar_t_in const* __restrict__ kv_in,     // [N, 512] bf16
-    uint8_t* __restrict__ k_cache,             // [num_blocks, block_stride]
-    int64_t const* __restrict__ slot_mapping,  // [num_tokens_insert] i64
-    int64_t const* __restrict__ position_ids,  // [N] i64
-    float const* __restrict__ cos_sin_cache,   // [max_pos, 64] fp32
-    float const eps,
-    int const num_tokens_full,    // = q.size(0) = kv.size(0)
-    int const num_tokens_insert,  // = slot_mapping.size(0), ≤ num_tokens_full
-    int const num_heads_q,        // H
-    int const cache_block_size,   // tokens per paged-cache block
-    int const kv_block_stride) {  // bytes per paged-cache block
-#if (!defined(__CUDA_ARCH__) || __CUDA_ARCH__ < 800) && !defined(USE_ROCM)
-  // BF16 _typeConvert specialization is unavailable on pre-Ampere.  The
-  // DeepseekV4 kernel only runs with bf16 inputs in practice, so compile a
-  // no-op stub for sm_70/sm_75 to keep multi-arch builds happy.
-  if constexpr (std::is_same_v<scalar_t_in, c10::BFloat16>) {
-    return;
-  } else {
-#endif
-    using Converter = vllm::_typeConvert<scalar_t_in>;
-
-    int const warpsPerBlock = blockDim.x / 32;
-    int const warpId = threadIdx.x / 32;
-    int const laneId = threadIdx.x % 32;
-    int const globalWarpIdx = blockIdx.x * warpsPerBlock + warpId;
-
-    int const total_slots_per_token = num_heads_q + 1;
-    int const tokenIdx = globalWarpIdx / total_slots_per_token;
-    int const slotIdx = globalWarpIdx % total_slots_per_token;
-    if (tokenIdx >= num_tokens_full) return;
-
-    bool const isKV = (slotIdx == num_heads_q);
-    // KV branch: skip DP-padded tokens (no slot reserved for them).
-    if (isKV && tokenIdx >= num_tokens_insert) return;
-
-    // PDL: wait for predecessor kernel (upstream q/kv producer) to signal
-    // before touching any global memory.  No-op when PDL is not enabled on
-    // the launch.  The CUDA runtime wrapper emits the griddepcontrol.wait
-    // PTX with the required memory clobber internally.
-#if defined(__CUDA_ARCH__) && (__CUDA_ARCH__ >= 900)
-    cudaGridDependencySynchronize();
-#endif
-
-    // Dim range this lane owns within the 512-wide head.
-    int const dim_base = laneId * kElemsPerLane;  // in [0, 512) step 16
-
-    // ── Load 16 bf16 → 16 fp32 registers (one 16-byte + one 16-byte LDG) ────
-    float elements[kElemsPerLane];
-    float sumOfSquares = 0.0f;
-
-    scalar_t_in const* src_ptr;
-    if (isKV) {
-      src_ptr = kv_in + static_cast<int64_t>(tokenIdx) * kHeadDim + dim_base;
-    } else {
-      int64_t const q_row_offset =
-          (static_cast<int64_t>(tokenIdx) * num_heads_q + slotIdx) * kHeadDim +
-          dim_base;
-      src_ptr = q_inout + q_row_offset;
-    }
-
-    // Two 16-byte loads per thread (8 bf16 each).  Use uint4 as the vector
-    // type and bitcast to scalar_t_in packed pairs for conversion.
-    uint4 v0 = *reinterpret_cast<uint4 const*>(src_ptr);
-    uint4 v1 = *reinterpret_cast<uint4 const*>(src_ptr + 8);
-
-    {
-      typename Converter::packed_hip_type const* p0 =
-          reinterpret_cast<typename Converter::packed_hip_type const*>(&v0);
-      typename Converter::packed_hip_type const* p1 =
-          reinterpret_cast<typename Converter::packed_hip_type const*>(&v1);
-// Each packed_hip_type holds 2 bf16 → 4 packed = 8 elems per uint4.
-#pragma unroll
-      for (int i = 0; i < 4; i++) {
-        float2 f2 = Converter::convert(p0[i]);
-        elements[2 * i] = f2.x;
-        elements[2 * i + 1] = f2.y;
-      }
-#pragma unroll
-      for (int i = 0; i < 4; i++) {
-        float2 f2 = Converter::convert(p1[i]);
-        elements[8 + 2 * i] = f2.x;
-        elements[8 + 2 * i + 1] = f2.y;
-      }
-    }
-
-    // ── Q branch: RMSNorm with no weight (has_weight=False) ─────────────────
-    // Variance + rsqrt + multiply all in fp32, no intermediate bf16 round.
-    // The downstream bf16 round only happens at the final store.
-    if (!isKV) {
-#pragma unroll
-      for (int i = 0; i < kElemsPerLane; i++) {
-        sumOfSquares += elements[i] * elements[i];
-      }
-      sumOfSquares = warpSum<float>(sumOfSquares);
-      float const rms_rcp =
-          rsqrtf(sumOfSquares / static_cast<float>(kHeadDim) + eps);
-#pragma unroll
-      for (int i = 0; i < kElemsPerLane; i++) {
-        elements[i] = elements[i] * rms_rcp;
-      }
-    }
-
-    // ── GPT-J RoPE on dims [NOPE_DIM, HEAD_DIM) ─────────────────────────────
-    // All math in fp32.  cos_sin_cache is loaded as fp32 (its native storage).
-    bool const is_rope_lane = dim_base >= kNopeDim;
-    if (is_rope_lane) {
-      int64_t const pos = position_ids[tokenIdx];
-      constexpr int kHalfRope = kRopeDim / 2;  // 32
-      float const* cos_ptr = cos_sin_cache + pos * kRopeDim;
-      float const* sin_ptr = cos_ptr + kHalfRope;
-
-      int const rope_local_base = dim_base - kNopeDim;  // in [0, 64) step 16
-#pragma unroll
-      for (int p = 0; p < kElemsPerLane / 2; p++) {
-        int const pair_dim = rope_local_base + 2 * p;
-        int const half_idx = pair_dim / 2;
-        float const cos_v = VLLM_LDG(cos_ptr + half_idx);
-        float const sin_v = VLLM_LDG(sin_ptr + half_idx);
-        float const x_even = elements[2 * p];
-        float const x_odd = elements[2 * p + 1];
-        elements[2 * p] = x_even * cos_v - x_odd * sin_v;
-        elements[2 * p + 1] = x_even * sin_v + x_odd * cos_v;
-      }
-    }
-
-    // ═══════════════════════════════════════════════════════════════════════
-    // Q branch: cast to bf16 and store back in place.
-    // ═══════════════════════════════════════════════════════════════════════
-    if (!isKV) {
-      uint4 out0, out1;
-      typename Converter::packed_hip_type* po0 =
-          reinterpret_cast<typename Converter::packed_hip_type*>(&out0);
-      typename Converter::packed_hip_type* po1 =
-          reinterpret_cast<typename Converter::packed_hip_type*>(&out1);
-#pragma unroll
-      for (int i = 0; i < 4; i++) {
-        po0[i] = Converter::convert(
-            make_float2(elements[2 * i], elements[2 * i + 1]));
-      }
-#pragma unroll
-      for (int i = 0; i < 4; i++) {
-        po1[i] = Converter::convert(
-            make_float2(elements[8 + 2 * i], elements[8 + 2 * i + 1]));
-      }
-      scalar_t_in* dst =
-          q_inout +
-          (static_cast<int64_t>(tokenIdx) * num_heads_q + slotIdx) * kHeadDim +
-          dim_base;
-      *reinterpret_cast<uint4*>(dst) = out0;
-      *reinterpret_cast<uint4*>(dst + 8) = out1;
-#if defined(__CUDA_ARCH__) && (__CUDA_ARCH__ >= 900)
-      cudaTriggerProgrammaticLaunchCompletion();
-#endif
-      return;
-    }
-
-    // ═══════════════════════════════════════════════════════════════════════
-    // KV branch.
-    // ═══════════════════════════════════════════════════════════════════════
-    int64_t const slot_id = slot_mapping[tokenIdx];
-    if (slot_id < 0) {
-#if defined(__CUDA_ARCH__) && (__CUDA_ARCH__ >= 900)
-      cudaTriggerProgrammaticLaunchCompletion();
-#endif
-      return;
-    }
-
-    int64_t const block_idx = slot_id / cache_block_size;
-    int64_t const pos_in_block = slot_id % cache_block_size;
-    uint8_t* block_base =
-        k_cache + block_idx * static_cast<int64_t>(kv_block_stride);
-    uint8_t* token_fp8_ptr = block_base + pos_in_block * kTokenDataBytes;
-    uint8_t* token_bf16_ptr = token_fp8_ptr + kNopeDim;
-    uint8_t* token_scale_ptr =
-        block_base + static_cast<int64_t>(cache_block_size) * kTokenDataBytes +
-        pos_in_block * kScaleBytesPerToken;
-
-    // Round K to bf16 first, matching the unfused reference path where K is
-    // materialized as bf16 before K quantization.  absmax, clamp, and FP8
-    // quant below all run on these bf16-rounded values.
-#pragma unroll
-    for (int i = 0; i < kElemsPerLane; i++) {
-      elements[i] = Converter::convert(Converter::convert(elements[i]));
-    }
-
-    // Per-quant-block absmax must be computed by ALL 32 lanes (warp-collective
-    // shuffle requires full participation).  RoPE lanes contribute garbage,
-    // but their values are gated out below via `!is_rope_lane`.
-    float local_absmax = 0.0f;
-#pragma unroll
-    for (int i = 0; i < kElemsPerLane; i++) {
-      local_absmax = fmaxf(local_absmax, fabsf(elements[i]));
-    }
-    float const absmax = fmaxf(warp4MaxAbs(local_absmax), 1e-4f);
-    float const exponent = ceilf(log2f(absmax / kFp8Max));
-    float const inv_scale = exp2f(-exponent);
-
-    if (!is_rope_lane) {
-      // ── NoPE lane: UE8M0 FP8 quant ───────────────────────────────────────
-      uint8_t out_bytes[kElemsPerLane];
-#pragma unroll
-      for (int i = 0; i < kElemsPerLane; i++) {
-        float scaled = elements[i] * inv_scale;
-        scaled = fminf(fmaxf(scaled, -kFp8Max), kFp8Max);
-        __nv_fp8_storage_t s =
-            __nv_cvt_float_to_fp8(scaled, __NV_SATFINITE, __NV_E4M3);
-        out_bytes[i] = static_cast<uint8_t>(s);
-      }
-      // One 16-byte STG per lane.
-      *reinterpret_cast<uint4*>(token_fp8_ptr + dim_base) =
-          *reinterpret_cast<uint4 const*>(out_bytes);
-
-      // Lane (4k) of each 4-lane group writes the scale byte for block k<7.
-      if ((laneId & 3) == 0) {
-        int const q_block_idx = laneId >> 2;  // 0..6 for NoPE lanes
-        float encoded = fmaxf(fminf(exponent + 127.0f, 255.0f), 0.0f);
-        token_scale_ptr[q_block_idx] = static_cast<uint8_t>(encoded);
-      }
-      // Lane 0 also writes the padding byte at index 7.
-      if (laneId == 0) {
-        token_scale_ptr[kNumQuantBlocks] = 0;  // pad
-      }
-    } else {
-      // ── RoPE lane: cast back to bf16 and store to cache bf16 tail ────────
-      uint4 out0, out1;
-      typename Converter::packed_hip_type* po0 =
-          reinterpret_cast<typename Converter::packed_hip_type*>(&out0);
-      typename Converter::packed_hip_type* po1 =
-          reinterpret_cast<typename Converter::packed_hip_type*>(&out1);
-#pragma unroll
-      for (int i = 0; i < 4; i++) {
-        po0[i] = Converter::convert(
-            make_float2(elements[2 * i], elements[2 * i + 1]));
-      }
-#pragma unroll
-      for (int i = 0; i < 4; i++) {
-        po1[i] = Converter::convert(
-            make_float2(elements[8 + 2 * i], elements[8 + 2 * i + 1]));
-      }
-      int const rope_local_base = dim_base - kNopeDim;  // in [0, 64)
-      scalar_t_in* bf16_dst =
-          reinterpret_cast<scalar_t_in*>(token_bf16_ptr) + rope_local_base;
-      *reinterpret_cast<uint4*>(bf16_dst) = out0;
-      *reinterpret_cast<uint4*>(bf16_dst + 8) = out1;
-    }
-#if defined(__CUDA_ARCH__) && (__CUDA_ARCH__ >= 900)
-    cudaTriggerProgrammaticLaunchCompletion();
-#endif
-#if (!defined(__CUDA_ARCH__) || __CUDA_ARCH__ < 800) && !defined(USE_ROCM)
-  }
-#endif
-}
-
-// ────────────────────────────────────────────────────────────────────────────
-// Launch wrapper
-// ────────────────────────────────────────────────────────────────────────────
-template <typename scalar_t_in>
-void launchFusedDeepseekV4QNormRopeKVRopeQuantInsert(
-    scalar_t_in* q_inout, scalar_t_in const* kv_in, uint8_t* k_cache,
-    int64_t const* slot_mapping, int64_t const* position_ids,
-    float const* cos_sin_cache, float const eps, int const num_tokens_full,
-    int const num_tokens_insert, int const num_heads_q,
-    int const cache_block_size, int const kv_block_stride,
-    cudaStream_t stream) {
-  constexpr int kBlockSize = 256;
-  constexpr int kWarpsPerBlock = kBlockSize / 32;
-  int64_t const total_warps =
-      static_cast<int64_t>(num_tokens_full) * (num_heads_q + 1);
-  int const grid =
-      static_cast<int>((total_warps + kWarpsPerBlock - 1) / kWarpsPerBlock);
-
-  // PDL: enable programmatic stream serialization whenever the hardware
-  // supports it (SM90+).  On pre-Hopper GPUs the attribute is unavailable,
-  // so leave numAttrs = 0 and launch as a regular kernel.
-  static int const sm_version = getSMVersion();
-  // Host-side guard: the device kernel body is compiled as a no-op for
-  // bf16 on pre-Ampere (sm_70/sm_75) because _typeConvert<BFloat16> is
-  // unavailable there.  Refuse the launch loudly instead of silently
-  // skipping the work.
-  TORCH_CHECK(
-      sm_version >= 80,
-      "fused_deepseek_v4_qnorm_rope_kv_rope_quant_insert requires sm_80+ "
-      "(Ampere or newer); got sm_",
-      sm_version);
-  cudaLaunchConfig_t config;
-  config.gridDim = dim3(grid);
-  config.blockDim = dim3(kBlockSize);
-  config.dynamicSmemBytes = 0;
-  config.stream = stream;
-  cudaLaunchAttribute attrs[1];
-  attrs[0].id = cudaLaunchAttributeProgrammaticStreamSerialization;
-  attrs[0].val.programmaticStreamSerializationAllowed = 1;
-  config.attrs = attrs;
-  config.numAttrs = (sm_version >= 90) ? 1 : 0;
-
-  cudaLaunchKernelEx(
-      &config, fusedDeepseekV4QNormRopeKVRopeQuantInsertKernel<scalar_t_in>,
-      q_inout, kv_in, k_cache, slot_mapping, position_ids, cos_sin_cache, eps,
-      num_tokens_full, num_tokens_insert, num_heads_q, cache_block_size,
-      kv_block_stride);
-}
-
-}  // namespace deepseek_v4_fused_ops
-}  // namespace vllm
-
-// ────────────────────────────────────────────────────────────────────────────
-// Torch op wrapper
-// ────────────────────────────────────────────────────────────────────────────
-void fused_deepseek_v4_qnorm_rope_kv_rope_quant_insert(
-    torch::Tensor& q,                    // [N, H, 512] bf16, in place
-    torch::Tensor const& kv,             // [N, 512] bf16 (read-only)
-    torch::Tensor& k_cache,              // [num_blocks, block_bytes] uint8
-    torch::Tensor const& slot_mapping,   // [N] int64
-    torch::Tensor const& position_ids,   // [N] int64
-    torch::Tensor const& cos_sin_cache,  // [max_pos, rope_dim] bf16
-    double eps, int64_t cache_block_size) {
-  TORCH_CHECK(q.is_cuda() && q.is_contiguous(), "q must be contiguous CUDA");
-  TORCH_CHECK(kv.is_cuda() && kv.is_contiguous(), "kv must be contiguous CUDA");
-  TORCH_CHECK(k_cache.is_cuda(), "k_cache must be CUDA");
-  TORCH_CHECK(slot_mapping.is_cuda() && slot_mapping.dtype() == torch::kInt64,
-              "slot_mapping must be int64 CUDA");
-  TORCH_CHECK(position_ids.is_cuda() && position_ids.dtype() == torch::kInt64,
-              "position_ids must be int64 CUDA");
-  TORCH_CHECK(cos_sin_cache.is_cuda(), "cos_sin_cache must be CUDA");
-  TORCH_CHECK(q.dim() == 3 && q.size(2) == 512, "q shape [N, H, 512]");
-  TORCH_CHECK(kv.dim() == 2 && kv.size(1) == 512, "kv shape [N, 512]");
-  TORCH_CHECK(q.dtype() == kv.dtype(), "q and kv dtype must match");
-  TORCH_CHECK(k_cache.dtype() == torch::kUInt8, "k_cache must be uint8");
-  TORCH_CHECK(cos_sin_cache.dim() == 2 && cos_sin_cache.size(1) == 64,
-              "cos_sin_cache shape [max_pos, 64]");
-  TORCH_CHECK(cos_sin_cache.dtype() == torch::kFloat32,
-              "cos_sin_cache must be float32");
-
-  // With DP padding, slot_mapping can be shorter than q/kv/positions.
-  // Q-norm+RoPE runs on all q.size(0) rows (downstream attention uses them);
-  // KV quant+insert runs only on the first slot_mapping.size(0) rows.
-  int const num_tokens_full = static_cast<int>(q.size(0));
-  int const num_tokens_insert = static_cast<int>(slot_mapping.size(0));
-  TORCH_CHECK(static_cast<int>(kv.size(0)) == num_tokens_full &&
-                  static_cast<int>(position_ids.size(0)) == num_tokens_full,
-              "q/kv/position_ids row counts must match");
-  TORCH_CHECK(num_tokens_insert <= num_tokens_full,
-              "slot_mapping must not exceed q row count");
-  int const num_heads_q = static_cast<int>(q.size(1));
-  int const cache_block_size_i = static_cast<int>(cache_block_size);
-  int const kv_block_stride = static_cast<int>(k_cache.stride(0));
-
-  at::cuda::OptionalCUDAGuard device_guard(device_of(q));
-  auto stream = at::cuda::getCurrentCUDAStream();
-
-  VLLM_DISPATCH_HALF_TYPES(
-      q.scalar_type(), "fused_deepseek_v4_qnorm_rope_kv_insert", [&] {
-        using qkv_scalar_t = scalar_t;
-        vllm::deepseek_v4_fused_ops::
-            launchFusedDeepseekV4QNormRopeKVRopeQuantInsert<qkv_scalar_t>(
-                reinterpret_cast<qkv_scalar_t*>(q.data_ptr()),
-                reinterpret_cast<qkv_scalar_t const*>(kv.data_ptr()),
-                reinterpret_cast<uint8_t*>(k_cache.data_ptr()),
-                reinterpret_cast<int64_t const*>(slot_mapping.data_ptr()),
-                reinterpret_cast<int64_t const*>(position_ids.data_ptr()),
-                cos_sin_cache.data_ptr<float>(), static_cast<float>(eps),
-                num_tokens_full, num_tokens_insert, num_heads_q,
-                cache_block_size_i, kv_block_stride, stream);
-      });
-}
@@ -77,8 +77,7 @@ __global__ void rms_norm_kernel(
 #pragma unroll
    for (int j = 0; j < VEC_SIZE; j++) {
      float x = static_cast<float>(src1.val[j]);
-      float w = static_cast<float>(src2.val[j]);
-      dst.val[j] = static_cast<scalar_t>(x * s_variance * w);
+      dst.val[j] = ((scalar_t)(x * s_variance)) * src2.val[j];
    }
    v_out[i] = dst;
  }
@@ -135,17 +134,10 @@ fused_add_rms_norm_kernel(
  for (int idx = threadIdx.x; idx < vec_hidden_size; idx += blockDim.x) {
    int id = blockIdx.x * vec_hidden_size + idx;
    int64_t strided_id = blockIdx.x * vec_input_stride + idx;
-    _f16Vec<scalar_t, width> res = residual_v[id];
-    _f16Vec<scalar_t, width> w = weight_v[idx];
-    _f16Vec<scalar_t, width> out;
-    using Converter = _typeConvert<scalar_t>;
-#pragma unroll
-    for (int j = 0; j < width; ++j) {
-      float x = Converter::convert(res.data[j]);
-      float wf = Converter::convert(w.data[j]);
-      out.data[j] = Converter::convert(x * s_variance * wf);
-    }
-    input_v[strided_id] = out;
+    _f16Vec<scalar_t, width> temp = residual_v[id];
+    temp *= s_variance;
+    temp *= weight_v[idx];
+    input_v[strided_id] = temp;
  }
 }

@@ -182,8 +174,8 @@ fused_add_rms_norm_kernel(

  for (int idx = threadIdx.x; idx < hidden_size; idx += blockDim.x) {
    float x = (float)residual[blockIdx.x * hidden_size + idx];
-    float w = (float)weight[idx];
-    input[blockIdx.x * input_stride + idx] = (scalar_t)(x * s_variance * w);
+    input[blockIdx.x * input_stride + idx] =
+        ((scalar_t)(x * s_variance)) * weight[idx];
  }
 }

@@ -65,16 +65,9 @@ __global__ void rms_norm_static_fp8_quant_kernel(
 #pragma unroll
    for (int j = 0; j < VEC_SIZE; j++) {
      float x = static_cast<float>(src1.val[j]);
-      float w = static_cast<float>(src2.val[j]);
-      // Round normalized result through scalar_t to match the precision of the
-      // unfused composite (rms_norm writes scalar_t, then
-      // static_scaled_fp8_quant re-loads it as float before FP8 conversion).
-      // Without this round, the fused path is strictly more accurate and
-      // disagrees with the composite at exact E4M3 quantization tie boundaries.
-      scalar_t out_norm = static_cast<scalar_t>(x * s_variance * w);
+      float const out_norm = ((scalar_t)(x * s_variance)) * src2.val[j];
      out[blockIdx.x * hidden_size + idx * VEC_SIZE + j] =
-          scaled_fp8_conversion<true, fp8_type>(static_cast<float>(out_norm),
-                                                scale_inv);
+          scaled_fp8_conversion<true, fp8_type>(out_norm, scale_inv);
    }
  }
 }
@@ -134,21 +127,13 @@ fused_add_rms_norm_static_fp8_quant_kernel(

  for (int idx = threadIdx.x; idx < vec_hidden_size; idx += blockDim.x) {
    int id = blockIdx.x * vec_hidden_size + idx;
-    _f16Vec<scalar_t, width> res = residual_v[id];
-    _f16Vec<scalar_t, width> w = weight_v[idx];
-    using Converter = _typeConvert<scalar_t>;
-    using HipT = typename Converter::hip_type;
+    _f16Vec<scalar_t, width> temp = residual_v[id];
+    temp *= s_variance;
+    temp *= weight_v[idx];
 #pragma unroll
    for (int i = 0; i < width; ++i) {
-      float x = Converter::convert(res.data[i]);
-      float wf = Converter::convert(w.data[i]);
-      // See note in rms_norm_static_fp8_quant_kernel: round through scalar_t
-      // to match the unfused composite path at FP8 boundaries. We use the
-      // backend's hip_type for the intermediate since c10::Half/BFloat16 has
-      // ambiguous conversions on CUDA and no implicit conversion on ROCm.
-      HipT out_norm_h = Converter::convert(x * s_variance * wf);
-      out[id * width + i] = scaled_fp8_conversion<true, fp8_type>(
-          Converter::convert(out_norm_h), scale_inv);
+      out[id * width + i] =
+          scaled_fp8_conversion<true, fp8_type>(float(temp.data[i]), scale_inv);
    }
  }
 }
@@ -191,12 +176,9 @@ fused_add_rms_norm_static_fp8_quant_kernel(

  for (int idx = threadIdx.x; idx < hidden_size; idx += blockDim.x) {
    float x = (float)residual[blockIdx.x * hidden_size + idx];
-    float w = (float)weight[idx];
-    // See note in rms_norm_static_fp8_quant_kernel: round through scalar_t
-    // to match the unfused composite path at FP8 boundaries.
-    scalar_t out_norm = static_cast<scalar_t>(x * s_variance * w);
-    out[blockIdx.x * hidden_size + idx] = scaled_fp8_conversion<true, fp8_type>(
-        static_cast<float>(out_norm), scale_inv);
+    float const out_norm = ((scalar_t)(x * s_variance)) * weight[idx];
+    out[blockIdx.x * hidden_size + idx] =
+        scaled_fp8_conversion<true, fp8_type>(out_norm, scale_inv);
  }
 }

@@ -12,15 +12,6 @@ void topk_sigmoid(torch::Tensor& topk_weights, torch::Tensor& topk_indices,
                  torch::Tensor& gating_output, bool renormalize,
                  std::optional<torch::Tensor> bias);

-void topk_softplus_sqrt(torch::Tensor& topk_weights,
-                        torch::Tensor& topk_indices,
-                        torch::Tensor& token_expert_indices,
-                        torch::Tensor& gating_output, bool renormalize,
-                        double routed_scaling_factor,
-                        const c10::optional<torch::Tensor>& correction_bias,
-                        const c10::optional<torch::Tensor>& input_ids,
-                        const c10::optional<torch::Tensor>& tid2eid);
-
 void moe_sum(torch::Tensor& input, torch::Tensor& output);

 void moe_align_block_size(torch::Tensor topk_ids, int64_t num_experts,
@@ -1,715 +0,0 @@
-/*
- * Adapted from
- * https://github.com/NVIDIA/TensorRT-LLM/blob/v0.7.1/cpp/tensorrt_llm/kernels/mixtureOfExperts/moe_kernels.cu
- * Copyright (c) 2024, The vLLM team.
- * SPDX-FileCopyrightText: Copyright (c) 1993-2023 NVIDIA CORPORATION &
- * AFFILIATES. All rights reserved. SPDX-License-Identifier: Apache-2.0
- *
- * Licensed under the Apache License, Version 2.0 (the "License");
- * you may not use this file except in compliance with the License.
- * You may obtain a copy of the License at
- *
- * http://www.apache.org/licenses/LICENSE-2.0
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS,
- * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
- * See the License for the specific language governing permissions and
- * limitations under the License.
- */
-#include <type_traits>
-#include <torch/all.h>
-#include <ATen/cuda/CUDAContext.h>
-#include <c10/cuda/CUDAGuard.h>
-#include "../cuda_compat.h"
-#include "../cub_helpers.h"
-#ifndef USE_ROCM
-  #include <cuda_bf16.h>
-  #include <cuda_fp16.h>
-#else
-  #include <hip/hip_bf16.h>
-  #include <hip/hip_fp16.h>
-typedef __hip_bfloat16 __nv_bfloat16;
-typedef __hip_bfloat162 __nv_bfloat162;
-#endif
-
-#define MAX(a, b) ((a) > (b) ? (a) : (b))
-#define MIN(a, b) ((a) < (b) ? (a) : (b))
-
-namespace vllm {
-namespace moe {
-
-/// Aligned array type
-template <typename T,
-          /// Number of elements in the array
-          int N,
-          /// Alignment requirement in bytes
-          int Alignment = sizeof(T) * N>
-struct alignas(Alignment) AlignedArray {
-  T data[N];
-};
-
-template <typename T>
-__device__ __forceinline__ float toFloat(T value) {
-  if constexpr (std::is_same_v<T, float>) {
-    return value;
-  } else if constexpr (std::is_same_v<T, __nv_bfloat16>) {
-    return __bfloat162float(value);
-  } else if constexpr (std::is_same_v<T, __half>) {
-    return __half2float(value);
-  }
-}
-
-#define FINAL_MASK 0xffffffff
-template <typename T>
-__inline__ __device__ T warpReduceSum(T val) {
-#pragma unroll
-  for (int mask = 16; mask > 0; mask >>= 1)
-    val += __shfl_xor_sync(FINAL_MASK, val, mask, 32);
-  return val;
-}
-
-// ====================== TopK softplus_sqrt things
-// ===============================
-
-/*
-  A Top-K gating softplus_sqrt written to exploit when the number of experts in
-  the MoE layers are a small power of 2. This allows us to cleanly share the
-  rows among the threads in a single warp and eliminate communication between
-  warps (so no need to use shared mem).
-
-  It fuses the sigmoid, max and argmax into a single kernel.
-
-  Limitations:
-  1) This implementation is optimized for when the number of experts is a small
-  power of 2. Additionally it also supports when number of experts is multiple
-  of 64 which is still faster than the computing sigmoid and topK separately
-  (only tested on CUDA yet). 2) This implementation assumes k is small, but will
-  work for any k.
-*/
-
-template <int VPT, int NUM_EXPERTS, int WARPS_PER_CTA, int BYTES_PER_LDG,
-          int WARP_SIZE_PARAM, bool USE_HASH, typename IndType,
-          typename InputType = float>
-__launch_bounds__(WARPS_PER_CTA* WARP_SIZE_PARAM) __global__
-    void topkGatingSoftplusSqrt(
-        const InputType* input, const bool* finished, float* output,
-        const int num_rows, IndType* indices, int* source_rows, const int k,
-        const int start_expert, const int end_expert, const bool renormalize,
-        double routed_scaling_factor, const float* correction_bias,
-        const IndType* input_ids, const IndType* tid2eid) {
-  static_assert(std::is_same_v<InputType, float> ||
-                    std::is_same_v<InputType, __nv_bfloat16> ||
-                    std::is_same_v<InputType, __half>,
-                "InputType must be float, __nv_bfloat16, or __half");
-
-  // We begin by enforcing compile time assertions and setting up compile time
-  // constants.
-  static_assert(BYTES_PER_LDG == (BYTES_PER_LDG & -BYTES_PER_LDG),
-                "BYTES_PER_LDG must be power of 2");
-  static_assert(BYTES_PER_LDG <= 16, "BYTES_PER_LDG must be leq 16");
-
-  // Number of bytes each thread pulls in per load
-  static constexpr int ELTS_PER_LDG = BYTES_PER_LDG / sizeof(InputType);
-  static constexpr int ELTS_PER_ROW = NUM_EXPERTS;
-  static constexpr int THREADS_PER_ROW = ELTS_PER_ROW / VPT;
-  static constexpr int LDG_PER_THREAD = VPT / ELTS_PER_LDG;
-
-  if constexpr (std::is_same_v<InputType, __nv_bfloat16> ||
-                std::is_same_v<InputType, __half>) {
-    static_assert(ELTS_PER_LDG == 1 || ELTS_PER_LDG % 2 == 0,
-                  "ELTS_PER_LDG must be 1 or even for 16-bit conversion");
-  }
-
-  // Restrictions based on previous section.
-  static_assert(
-      VPT % ELTS_PER_LDG == 0,
-      "The elements per thread must be a multiple of the elements per ldg");
-  static_assert(WARP_SIZE_PARAM % THREADS_PER_ROW == 0,
-                "The threads per row must cleanly divide the threads per warp");
-  static_assert(THREADS_PER_ROW == (THREADS_PER_ROW & -THREADS_PER_ROW),
-                "THREADS_PER_ROW must be power of 2");
-  static_assert(THREADS_PER_ROW <= WARP_SIZE_PARAM,
-                "THREADS_PER_ROW can be at most warp size");
-
-  // We have NUM_EXPERTS elements per row. We specialize for small #experts
-  static constexpr int ELTS_PER_WARP = WARP_SIZE_PARAM * VPT;
-  static constexpr int ROWS_PER_WARP = ELTS_PER_WARP / ELTS_PER_ROW;
-  static constexpr int ROWS_PER_CTA = WARPS_PER_CTA * ROWS_PER_WARP;
-
-  // Restrictions for previous section.
-  static_assert(ELTS_PER_WARP % ELTS_PER_ROW == 0,
-                "The elts per row must cleanly divide the total elt per warp");
-
-  // ===================== From this point, we finally start computing run-time
-  // variables. ========================
-
-  // Compute CTA and warp rows. We pack multiple rows into a single warp, and a
-  // block contains WARPS_PER_CTA warps. This, each block processes a chunk of
-  // rows. We start by computing the start row for each block.
-  const int cta_base_row = blockIdx.x * ROWS_PER_CTA;
-
-  // Now, using the base row per thread block, we compute the base row per warp.
-  const int warp_base_row = cta_base_row + threadIdx.y * ROWS_PER_WARP;
-
-  // The threads in a warp are split into sub-groups that will work on a row.
-  // We compute row offset for each thread sub-group
-  const int thread_row_in_warp = threadIdx.x / THREADS_PER_ROW;
-  const int thread_row = warp_base_row + thread_row_in_warp;
-
-  // Threads with indices out of bounds should early exit here.
-  if (thread_row >= num_rows) {
-    return;
-  }
-  const bool row_is_active = finished ? !finished[thread_row] : true;
-
-  // We finally start setting up the read pointers for each thread. First, each
-  // thread jumps to the start of the row it will read.
-  const InputType* thread_row_ptr = input + thread_row * ELTS_PER_ROW;
-
-  // Now, we compute the group each thread belong to in order to determine the
-  // first column to start loads.
-  const int thread_group_idx = threadIdx.x % THREADS_PER_ROW;
-  const int first_elt_read_by_thread = thread_group_idx * ELTS_PER_LDG;
-  const InputType* thread_read_ptr = thread_row_ptr + first_elt_read_by_thread;
-
-  // Finally, we pull in the data from global mem
-  float row_chunk[VPT];
-
-#if (defined(__CUDA_ARCH__) && (__CUDA_ARCH__ >= 900))
-  asm volatile("griddepcontrol.wait;");
-#endif
-
-  // NOTE(zhuhaoran): dispatch different input types loading, BF16/FP16 convert
-  // to float
-  if constexpr (std::is_same_v<InputType, float>) {
-    using VecType = AlignedArray<float, ELTS_PER_LDG>;
-    VecType* row_chunk_vec_ptr = reinterpret_cast<VecType*>(&row_chunk);
-    const VecType* vec_thread_read_ptr =
-        reinterpret_cast<const VecType*>(thread_read_ptr);
-#pragma unroll
-    for (int ii = 0; ii < LDG_PER_THREAD; ++ii) {
-      row_chunk_vec_ptr[ii] = vec_thread_read_ptr[ii * THREADS_PER_ROW];
-    }
-  } else if constexpr (std::is_same_v<InputType, __nv_bfloat16>) {
-    if constexpr (ELTS_PER_LDG >= 2) {
-      using VecType = AlignedArray<__nv_bfloat16, ELTS_PER_LDG>;
-      float2* row_chunk_f2 = reinterpret_cast<float2*>(row_chunk);
-      const VecType* vec_thread_read_ptr =
-          reinterpret_cast<const VecType*>(thread_read_ptr);
-#pragma unroll
-      for (int ii = 0; ii < LDG_PER_THREAD; ++ii) {
-        VecType vec = vec_thread_read_ptr[ii * THREADS_PER_ROW];
-        int base_idx_f2 = ii * ELTS_PER_LDG / 2;
-#pragma unroll
-        for (int jj = 0; jj < ELTS_PER_LDG / 2; ++jj) {
-          row_chunk_f2[base_idx_f2 + jj] = __bfloat1622float2(
-              *reinterpret_cast<const __nv_bfloat162*>(vec.data + jj * 2));
-        }
-      }
-    } else {  // ELTS_PER_LDG == 1
-#pragma unroll
-      for (int ii = 0; ii < LDG_PER_THREAD; ++ii) {
-        const __nv_bfloat16* scalar_ptr =
-            thread_read_ptr + ii * THREADS_PER_ROW;
-        row_chunk[ii] = __bfloat162float(*scalar_ptr);
-      }
-    }
-  } else if constexpr (std::is_same_v<InputType, __half>) {
-    if constexpr (ELTS_PER_LDG >= 2) {
-      using VecType = AlignedArray<__half, ELTS_PER_LDG>;
-      float2* row_chunk_f2 = reinterpret_cast<float2*>(row_chunk);
-      const VecType* vec_thread_read_ptr =
-          reinterpret_cast<const VecType*>(thread_read_ptr);
-#pragma unroll
-      for (int ii = 0; ii < LDG_PER_THREAD; ++ii) {
-        VecType vec = vec_thread_read_ptr[ii * THREADS_PER_ROW];
-        int base_idx_f2 = ii * ELTS_PER_LDG / 2;
-#pragma unroll
-        for (int jj = 0; jj < ELTS_PER_LDG / 2; ++jj) {
-          row_chunk_f2[base_idx_f2 + jj] = __half22float2(
-              *reinterpret_cast<const __half2*>(vec.data + jj * 2));
-        }
-      }
-    } else {  // ELTS_PER_LDG == 1
-#pragma unroll
-      for (int ii = 0; ii < LDG_PER_THREAD; ++ii) {
-        const __half* scalar_ptr = thread_read_ptr + ii * THREADS_PER_ROW;
-        row_chunk[ii] = __half2float(*scalar_ptr);
-      }
-    }
-  }
-  constexpr float threshold = 20.0f;
-  constexpr float beta = 1.0f;
-
-  // Hash MoE path: indices are predetermined from lookup table
-  if constexpr (USE_HASH) {
-    const IndType token_id = input_ids[thread_row];
-    const IndType* expert_indices_for_token = tid2eid + token_id * k;
-#pragma unroll
-    for (int ii = 0; ii < VPT; ++ii) {
-      float val = row_chunk[ii];
-      float val_b = val * beta;
-      val = (val_b > threshold) ? val : (__logf(1.0f + __expf(val_b))) / beta;
-      row_chunk[ii] = sqrtf(val);
-    }
-    float selected_sum = 0.f;
-#pragma unroll
-    for (int k_idx = 0; k_idx < k; ++k_idx) {
-      const int expert = expert_indices_for_token[k_idx];
-      const int idx = k * thread_row + k_idx;
-      for (int ii = 0; ii < VPT; ++ii) {
-        const int group_id = ii / ELTS_PER_LDG;
-        const int local_id = ii % ELTS_PER_LDG;
-        const int expert_idx = first_elt_read_by_thread +
-                               group_id * THREADS_PER_ROW * ELTS_PER_LDG +
-                               local_id;
-        if (expert == expert_idx) {
-          indices[idx] = expert;
-          selected_sum += row_chunk[ii];
-          break;
-        }
-      }
-    }
-    // Compute per-thread scale (using warp reduction when renormalizing).
-    if (renormalize) {
-      selected_sum = warpReduceSum(selected_sum);
-    }
-    float scale = static_cast<float>(routed_scaling_factor);
-    if (renormalize) {
-      const float denom = selected_sum > 0.f ? selected_sum : 1.f;
-      scale /= denom;
-    }
-
-#pragma unroll
-    for (int k_idx = 0; k_idx < k; ++k_idx) {
-      const int expert = expert_indices_for_token[k_idx];
-      const int idx = k * thread_row + k_idx;
-      for (int ii = 0; ii < VPT; ++ii) {
-        const int group_id = ii / ELTS_PER_LDG;
-        const int local_id = ii % ELTS_PER_LDG;
-        const int expert_idx = first_elt_read_by_thread +
-                               group_id * THREADS_PER_ROW * ELTS_PER_LDG +
-                               local_id;
-        if (expert == expert_idx) {
-          output[idx] = row_chunk[ii] * scale;
-          break;
-        }
-      }
-    }
-#if (defined(__CUDA_ARCH__) && (__CUDA_ARCH__ >= 900))
-    asm volatile("griddepcontrol.launch_dependents;");
-#endif
-    return;
-  }
-
-#pragma unroll
-  for (int ii = 0; ii < VPT; ++ii) {
-    float val = row_chunk[ii];
-    float val_b = val * beta;
-    // Compute softplus: log(1 + exp(val)) with numerical stability
-    // When val > threshold, softplus(x) ≈ x to avoid exp overflow
-    val = (val_b > threshold) ? val : (__logf(1.0f + __expf(val_b))) / beta;
-    val = sqrtf(val);
-    if (correction_bias) {
-      const int group_id = ii / ELTS_PER_LDG;
-      const int local_id = ii % ELTS_PER_LDG;
-      const int expert_idx = first_elt_read_by_thread +
-                             group_id * THREADS_PER_ROW * ELTS_PER_LDG +
-                             local_id;
-      val = val + correction_bias[expert_idx];
-    }
-    row_chunk[ii] = val;
-  }
-
-  // Original TopK path: find top-k experts by score
-  // Now, sigmoid_res contains the sigmoid of the row chunk. Now, I want to find
-  // the topk elements in each row, along with the max index.
-  int start_col = first_elt_read_by_thread;
-  static constexpr int COLS_PER_GROUP_LDG = ELTS_PER_LDG * THREADS_PER_ROW;
-
-  float selected_sum = 0.f;
-  for (int k_idx = 0; k_idx < k; ++k_idx) {
-    // First, each thread does the local argmax
-    float max_val = row_chunk[0];
-    int expert = start_col;
-#pragma unroll
-    for (int ldg = 0, col = start_col; ldg < LDG_PER_THREAD;
-         ++ldg, col += COLS_PER_GROUP_LDG) {
-#pragma unroll
-      for (int ii = 0; ii < ELTS_PER_LDG; ++ii) {
-        float val = row_chunk[ldg * ELTS_PER_LDG + ii];
-
-        // No check on the experts here since columns with the smallest index
-        // are processed first and only updated if > (not >=)
-        if (val > max_val) {
-          max_val = val;
-          expert = col + ii;
-        }
-      }
-    }
-
-// Now, we perform the argmax reduce. We use the butterfly pattern so threads
-// reach consensus about the max. This will be useful for K > 1 so that the
-// threads can agree on "who" had the max value. That thread can then blank out
-// their max with -inf and the warp can run more iterations...
-#pragma unroll
-    for (int mask = THREADS_PER_ROW / 2; mask > 0; mask /= 2) {
-      float other_max =
-          VLLM_SHFL_XOR_SYNC_WIDTH(max_val, mask, THREADS_PER_ROW);
-      int other_expert =
-          VLLM_SHFL_XOR_SYNC_WIDTH(expert, mask, THREADS_PER_ROW);
-
-      // We want lower indices to "win" in every thread so we break ties this
-      // way
-      if (other_max > max_val ||
-          (other_max == max_val && other_expert < expert)) {
-        max_val = other_max;
-        expert = other_expert;
-      }
-    }
-
-    // Write the max for this k iteration to global memory.
-    if (thread_group_idx == 0) {
-      // Add a guard to ignore experts not included by this node
-      const bool node_uses_expert =
-          expert >= start_expert && expert < end_expert;
-      const bool should_process_row = row_is_active && node_uses_expert;
-
-      // The lead thread from each sub-group will write out the final results to
-      // global memory. (This will be a single) thread per row of the
-      // input/output matrices.
-      const int idx = k * thread_row + k_idx;
-      if (correction_bias != nullptr) {
-        max_val -= correction_bias[expert];
-      }
-      output[idx] = max_val;
-      indices[idx] = should_process_row ? (expert - start_expert) : NUM_EXPERTS;
-      source_rows[idx] = k_idx * num_rows + thread_row;
-      if (renormalize) {
-        selected_sum += max_val;
-      }
-    }
-
-    // Finally, we clear the value in the thread with the current max if there
-    // is another iteration to run.
-    if (k_idx + 1 < k) {
-      const int ldg_group_for_expert = expert / COLS_PER_GROUP_LDG;
-      const int thread_to_clear_in_group =
-          (expert / ELTS_PER_LDG) % THREADS_PER_ROW;
-
-      // Only the thread in the group which produced the max will reset the
-      // "winning" value to -inf.
-      if (thread_group_idx == thread_to_clear_in_group) {
-        const int offset_for_expert = expert % ELTS_PER_LDG;
-        // Safe to set to any negative value since row_chunk values must be
-        // between 0 and 1.
-        row_chunk[ldg_group_for_expert * ELTS_PER_LDG + offset_for_expert] =
-            -10000.f;
-      }
-    }
-  }
-
-  // Apply renormalization and routed scaling factor to final weights.
-  if (thread_group_idx == 0) {
-    float scale = static_cast<float>(routed_scaling_factor);
-    if (renormalize) {
-      const float denom = selected_sum > 0.f ? selected_sum : 1.f;
-      scale /= denom;
-    }
-    for (int k_idx = 0; k_idx < k; ++k_idx) {
-      const int idx = k * thread_row + k_idx;
-      output[idx] = output[idx] * scale;
-    }
-  }
-#if (defined(__CUDA_ARCH__) && (__CUDA_ARCH__ >= 900))
-  asm volatile("griddepcontrol.launch_dependents;");
-#endif
-}
-
-namespace detail {
-// Constructs some constants needed to partition the work across threads at
-// compile time.
-template <int EXPERTS, int BYTES_PER_LDG, int WARP_SIZE_PARAM,
-          typename InputType>
-struct TopkConstants {
-  static constexpr int ELTS_PER_LDG = BYTES_PER_LDG / sizeof(InputType);
-  static_assert(EXPERTS / (ELTS_PER_LDG * WARP_SIZE_PARAM) == 0 ||
-                    EXPERTS % (ELTS_PER_LDG * WARP_SIZE_PARAM) == 0,
-                "");
-  static constexpr int VECs_PER_THREAD =
-      MAX(1, EXPERTS / (ELTS_PER_LDG * WARP_SIZE_PARAM));
-  static constexpr int VPT = VECs_PER_THREAD * ELTS_PER_LDG;
-  static constexpr int THREADS_PER_ROW = EXPERTS / VPT;
-  static const int ROWS_PER_WARP = WARP_SIZE_PARAM / THREADS_PER_ROW;
-};
-}  // namespace detail
-
-#define DISPATCH_HASH(use_hash, USE_HASH, ...)                                 \
-  if (use_hash) {                                                              \
-    const bool USE_HASH = true;                                                \
-    static_assert(USE_HASH == true, "USE_HASH must be compile-time constant"); \
-    __VA_ARGS__                                                                \
-  } else {                                                                     \
-    const bool USE_HASH = false;                                               \
-    static_assert(USE_HASH == false,                                           \
-                  "USE_HASH must be compile-time constant");                   \
-    __VA_ARGS__                                                                \
-  }
-
-template <int EXPERTS, int WARPS_PER_TB, int WARP_SIZE_PARAM,
-          int MAX_BYTES_PER_LDG, typename IndType, typename InputType>
-void topkGatingSoftplusSqrtLauncherHelper(
-    const InputType* input, const bool* finished, float* output,
-    IndType* indices, int* source_row, const int num_rows, const int k,
-    const int start_expert, const int end_expert, const bool renormalize,
-    double routed_scaling_factor, const float* correction_bias,
-    const bool use_hash, const IndType* input_ids, const IndType* tid2eid,
-    cudaStream_t stream) {
-  static constexpr int BYTES_PER_LDG =
-      MIN(MAX_BYTES_PER_LDG, sizeof(InputType) * EXPERTS);
-  using Constants =
-      detail::TopkConstants<EXPERTS, BYTES_PER_LDG, WARP_SIZE_PARAM, InputType>;
-  static constexpr int VPT = Constants::VPT;
-  static constexpr int ROWS_PER_WARP = Constants::ROWS_PER_WARP;
-  const int num_warps = (num_rows + ROWS_PER_WARP - 1) / ROWS_PER_WARP;
-  const int num_blocks = (num_warps + WARPS_PER_TB - 1) / WARPS_PER_TB;
-  dim3 block_dim(WARP_SIZE_PARAM, WARPS_PER_TB);
-  DISPATCH_HASH(use_hash, USE_HASH, {
-    auto* kernel =
-        &topkGatingSoftplusSqrt<VPT, EXPERTS, WARPS_PER_TB, BYTES_PER_LDG,
-                                WARP_SIZE_PARAM, USE_HASH, IndType, InputType>;
-#ifndef USE_ROCM
-    cudaLaunchConfig_t config = {};
-    config.gridDim = num_blocks;
-    config.blockDim = block_dim;
-    config.dynamicSmemBytes = 0;
-    config.stream = stream;
-    cudaLaunchAttribute attrs[1];
-    attrs[0].id = cudaLaunchAttributeProgrammaticStreamSerialization;
-    attrs[0].val.programmaticStreamSerializationAllowed = 1;
-    config.numAttrs = 1;
-    config.attrs = attrs;
-    cudaLaunchKernelEx(&config, kernel, input, finished, output, num_rows,
-                       indices, source_row, k, start_expert, end_expert,
-                       renormalize, routed_scaling_factor, correction_bias,
-                       input_ids, tid2eid);
-#else
-    kernel<<<num_blocks, block_dim, 0, stream>>>(
-        input, finished, output, num_rows, indices, source_row, k, start_expert,
-        end_expert, renormalize, routed_scaling_factor, correction_bias,
-        input_ids, tid2eid);
-#endif
-  })
-}
-
-#ifndef USE_ROCM
-  #define LAUNCH_SOFTPLUS_SQRT(NUM_EXPERTS, WARPS_PER_TB, MAX_BYTES)           \
-    static_assert(WARP_SIZE == 32,                                             \
-                  "Unsupported warp size. Only 32 is supported for CUDA");     \
-    topkGatingSoftplusSqrtLauncherHelper<NUM_EXPERTS, WARPS_PER_TB, WARP_SIZE, \
-                                         MAX_BYTES>(                           \
-        gating_output, nullptr, topk_weights, topk_indices,                    \
-        token_expert_indices, num_tokens, topk, 0, num_experts, renormalize,   \
-        routed_scaling_factor, correction_bias, use_hash, input_ids, tid2eid,  \
-        stream);
-#else
-  #define LAUNCH_SOFTPLUS_SQRT(NUM_EXPERTS, WARPS_PER_TB, MAX_BYTES)           \
-    if (WARP_SIZE == 64) {                                                     \
-      topkGatingSoftplusSqrtLauncherHelper<NUM_EXPERTS, WARPS_PER_TB, 64,      \
-                                           MAX_BYTES>(                         \
-          gating_output, nullptr, topk_weights, topk_indices,                  \
-          token_expert_indices, num_tokens, topk, 0, num_experts, renormalize, \
-          routed_scaling_factor, correction_bias, use_hash, input_ids,         \
-          tid2eid, stream);                                                    \
-    } else if (WARP_SIZE == 32) {                                              \
-      topkGatingSoftplusSqrtLauncherHelper<NUM_EXPERTS, WARPS_PER_TB, 32,      \
-                                           MAX_BYTES>(                         \
-          gating_output, nullptr, topk_weights, topk_indices,                  \
-          token_expert_indices, num_tokens, topk, 0, num_experts, renormalize, \
-          routed_scaling_factor, correction_bias, use_hash, input_ids,         \
-          tid2eid, stream);                                                    \
-    } else {                                                                   \
-      assert(false &&                                                          \
-             "Unsupported warp size. Only 32 and 64 are supported for ROCm");  \
-    }
-#endif
-
-template <typename IndType, typename InputType>
-void topkGatingSoftplusSqrtKernelLauncher(
-    const InputType* gating_output, float* topk_weights, IndType* topk_indices,
-    int* token_expert_indices, const int num_tokens, const int num_experts,
-    const int topk, const bool renormalize, double routed_scaling_factor,
-    const float* correction_bias, const bool use_hash, const IndType* input_ids,
-    const IndType* tid2eid, cudaStream_t stream) {
-  static constexpr int WARPS_PER_TB = 4;
-  static constexpr int BYTES_PER_LDG_POWER_OF_2 = 16;
-#ifndef USE_ROCM
-  // for bfloat16 dtype, we need 4 bytes loading to make sure num_experts
-  // elements can be loaded by a warp
-  static constexpr int BYTES_PER_LDG_MULTIPLE_64 =
-      (std::is_same_v<InputType, __nv_bfloat16> ||
-       std::is_same_v<InputType, __half>)
-          ? 4
-          : 8;
-#endif
-  switch (num_experts) {
-    case 1:
-      LAUNCH_SOFTPLUS_SQRT(1, WARPS_PER_TB, BYTES_PER_LDG_POWER_OF_2);
-      break;
-    case 2:
-      LAUNCH_SOFTPLUS_SQRT(2, WARPS_PER_TB, BYTES_PER_LDG_POWER_OF_2);
-      break;
-    case 4:
-      LAUNCH_SOFTPLUS_SQRT(4, WARPS_PER_TB, BYTES_PER_LDG_POWER_OF_2);
-      break;
-    case 8:
-      LAUNCH_SOFTPLUS_SQRT(8, WARPS_PER_TB, BYTES_PER_LDG_POWER_OF_2);
-      break;
-    case 16:
-      LAUNCH_SOFTPLUS_SQRT(16, WARPS_PER_TB, BYTES_PER_LDG_POWER_OF_2);
-      break;
-    case 32:
-      LAUNCH_SOFTPLUS_SQRT(32, WARPS_PER_TB, BYTES_PER_LDG_POWER_OF_2);
-      break;
-    case 64:
-      LAUNCH_SOFTPLUS_SQRT(64, WARPS_PER_TB, BYTES_PER_LDG_POWER_OF_2);
-      break;
-    case 128:
-      LAUNCH_SOFTPLUS_SQRT(128, WARPS_PER_TB, BYTES_PER_LDG_POWER_OF_2);
-      break;
-    case 256:
-      LAUNCH_SOFTPLUS_SQRT(256, WARPS_PER_TB, BYTES_PER_LDG_POWER_OF_2);
-      break;
-    case 512:
-      LAUNCH_SOFTPLUS_SQRT(512, WARPS_PER_TB, BYTES_PER_LDG_POWER_OF_2);
-      break;
-      // (CUDA only) support multiples of 64 when num_experts is not power of 2.
-      // ROCm uses WARP_SIZE 64 so 8 bytes loading won't fit for some of
-      // num_experts, alternatively we can test 4 bytes loading and enable it in
-      // future.
-#ifndef USE_ROCM
-    case 192:
-      LAUNCH_SOFTPLUS_SQRT(192, WARPS_PER_TB, BYTES_PER_LDG_MULTIPLE_64);
-      break;
-    case 320:
-      LAUNCH_SOFTPLUS_SQRT(320, WARPS_PER_TB, BYTES_PER_LDG_MULTIPLE_64);
-      break;
-    case 384:
-      LAUNCH_SOFTPLUS_SQRT(384, WARPS_PER_TB, BYTES_PER_LDG_MULTIPLE_64);
-      break;
-    case 448:
-      LAUNCH_SOFTPLUS_SQRT(448, WARPS_PER_TB, BYTES_PER_LDG_MULTIPLE_64);
-      break;
-    case 576:
-      LAUNCH_SOFTPLUS_SQRT(576, WARPS_PER_TB, BYTES_PER_LDG_MULTIPLE_64);
-      break;
-#endif
-    default: {
-      TORCH_CHECK(false, "Unsupported expert number: ", num_experts);
-    }
-  }
-}
-
-}  // namespace moe
-}  // namespace vllm
-
-template <typename ComputeType>
-void dispatch_topk_softplus_sqrt_launch(
-    const ComputeType* gating_output, torch::Tensor& topk_weights,
-    torch::Tensor& topk_indices, torch::Tensor& token_expert_indices,
-    int num_tokens, int num_experts, int topk, bool renormalize,
-    double routed_scaling_factor,
-    const c10::optional<torch::Tensor>& correction_bias,
-    const c10::optional<torch::Tensor>& input_ids,
-    const c10::optional<torch::Tensor>& tid2eid, cudaStream_t stream) {
-  const float* bias_ptr = nullptr;
-  if (correction_bias.has_value()) {
-    bias_ptr = correction_bias.value().data_ptr<float>();
-  }
-  bool use_hash = false;
-  if (tid2eid.has_value()) {
-    TORCH_CHECK(input_ids.has_value(), "input_ids is required for hash MoE");
-    use_hash = true;
-  }
-  if (topk_indices.scalar_type() == at::ScalarType::Int) {
-    const int* input_ids_ptr = nullptr;
-    const int* tid2eid_ptr = nullptr;
-    if (tid2eid.has_value()) {
-      input_ids_ptr = input_ids.value().data_ptr<int>();
-      tid2eid_ptr = tid2eid.value().data_ptr<int>();
-    }
-
-    vllm::moe::topkGatingSoftplusSqrtKernelLauncher<int, ComputeType>(
-        gating_output, topk_weights.data_ptr<float>(),
-        topk_indices.data_ptr<int>(), token_expert_indices.data_ptr<int>(),
-        num_tokens, num_experts, topk, renormalize, routed_scaling_factor,
-        bias_ptr, use_hash, input_ids_ptr, tid2eid_ptr, stream);
-  } else if (topk_indices.scalar_type() == at::ScalarType::UInt32) {
-    const uint32_t* input_ids_ptr = nullptr;
-    const uint32_t* tid2eid_ptr = nullptr;
-    if (tid2eid.has_value()) {
-      input_ids_ptr = input_ids.value().data_ptr<uint32_t>();
-      tid2eid_ptr = tid2eid.value().data_ptr<uint32_t>();
-    }
-    vllm::moe::topkGatingSoftplusSqrtKernelLauncher<uint32_t, ComputeType>(
-        gating_output, topk_weights.data_ptr<float>(),
-        topk_indices.data_ptr<uint32_t>(), token_expert_indices.data_ptr<int>(),
-        num_tokens, num_experts, topk, renormalize, routed_scaling_factor,
-        bias_ptr, use_hash, input_ids_ptr, tid2eid_ptr, stream);
-  } else {
-    TORCH_CHECK(topk_indices.scalar_type() == at::ScalarType::Long);
-
-    const int64_t* input_ids_ptr = nullptr;
-    const int64_t* tid2eid_ptr = nullptr;
-    if (tid2eid.has_value()) {
-      input_ids_ptr = input_ids.value().data_ptr<int64_t>();
-      tid2eid_ptr = tid2eid.value().data_ptr<int64_t>();
-    }
-
-    vllm::moe::topkGatingSoftplusSqrtKernelLauncher<int64_t, ComputeType>(
-        gating_output, topk_weights.data_ptr<float>(),
-        topk_indices.data_ptr<int64_t>(), token_expert_indices.data_ptr<int>(),
-        num_tokens, num_experts, topk, renormalize, routed_scaling_factor,
-        bias_ptr, use_hash, input_ids_ptr, tid2eid_ptr, stream);
-  }
-}
-
-void topk_softplus_sqrt(
-    torch::Tensor& topk_weights,          // [num_tokens, topk]
-    torch::Tensor& topk_indices,          // [num_tokens, topk]
-    torch::Tensor& token_expert_indices,  // [num_tokens, topk]
-    torch::Tensor& gating_output,         // [num_tokens, num_experts]
-    bool renormalize, double routed_scaling_factor,
-    const c10::optional<torch::Tensor>& correction_bias,
-    const c10::optional<torch::Tensor>& input_ids,
-    const c10::optional<torch::Tensor>& tid2eid) {
-  const int num_experts = gating_output.size(-1);
-  const auto num_tokens = gating_output.numel() / num_experts;
-  const int topk = topk_weights.size(-1);
-  const at::cuda::OptionalCUDAGuard device_guard(device_of(gating_output));
-  const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
-
-  if (gating_output.scalar_type() == at::ScalarType::Float) {
-    dispatch_topk_softplus_sqrt_launch<float>(
-        gating_output.data_ptr<float>(), topk_weights, topk_indices,
-        token_expert_indices, num_tokens, num_experts, topk, renormalize,
-        routed_scaling_factor, correction_bias, input_ids, tid2eid, stream);
-  } else if (gating_output.scalar_type() == at::ScalarType::Half) {
-    dispatch_topk_softplus_sqrt_launch<__half>(
-        reinterpret_cast<const __half*>(gating_output.data_ptr<at::Half>()),
-        topk_weights, topk_indices, token_expert_indices, num_tokens,
-        num_experts, topk, renormalize, routed_scaling_factor, correction_bias,
-        input_ids, tid2eid, stream);
-  } else if (gating_output.scalar_type() == at::ScalarType::BFloat16) {
-    dispatch_topk_softplus_sqrt_launch<__nv_bfloat16>(
-        reinterpret_cast<const __nv_bfloat16*>(
-            gating_output.data_ptr<at::BFloat16>()),
-        topk_weights, topk_indices, token_expert_indices, num_tokens,
-        num_experts, topk, renormalize, routed_scaling_factor, correction_bias,
-        input_ids, tid2eid, stream);
-  } else {
-    TORCH_CHECK(false, "Unsupported gating_output data type: ",
-                gating_output.scalar_type());
-  }
-}
@@ -16,14 +16,6 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, m) {
      "bias) -> ()");
  m.impl("topk_sigmoid", torch::kCUDA, &topk_sigmoid);

-#ifndef USE_ROCM
-  m.def(
-      "topk_softplus_sqrt(Tensor! topk_weights, Tensor! topk_indices, Tensor! "
-      "token_expert_indices, Tensor gating_output, bool renormalize, float "
-      "routed_scaling_factor, Tensor? "
-      "bias, Tensor? input_ids, Tensor? tid2eid) -> ()");
-  m.impl("topk_softplus_sqrt", torch::kCUDA, &topk_softplus_sqrt);
-#endif
  // Calculate the result of moe by summing up the partial results
  // from all selected experts.
  m.def("moe_sum(Tensor input, Tensor! output) -> ()");
@@ -100,11 +100,6 @@ void fused_qk_norm_rope(torch::Tensor& qkv, int64_t num_heads_q,
                        bool is_neox, torch::Tensor& position_ids,
                        int64_t forced_token_heads_per_warp);

-void fused_deepseek_v4_qnorm_rope_kv_rope_quant_insert(
-    torch::Tensor& q, torch::Tensor const& kv, torch::Tensor& k_cache,
-    torch::Tensor const& slot_mapping, torch::Tensor const& position_ids,
-    torch::Tensor const& cos_sin_cache, double eps, int64_t cache_block_size);
-
 void apply_repetition_penalties_(torch::Tensor& logits,
                                 const torch::Tensor& prompt_mask,
                                 const torch::Tensor& output_mask,
@@ -158,13 +153,10 @@ void silu_and_mul_per_block_quant(torch::Tensor& out,

 void rotary_embedding(torch::Tensor& positions, torch::Tensor& query,
                      std::optional<torch::Tensor> key, int64_t head_size,
-                      torch::Tensor& cos_sin_cache, bool is_neox,
-                      int64_t rope_dim_offset, bool inverse);
+                      torch::Tensor& cos_sin_cache, bool is_neox);

 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);

@@ -18,6 +18,7 @@ namespace persistent {
 // Constants
 // ============================================================================

+constexpr int TopK = 2048;
 constexpr int kThreadsPerBlock = 1024;
 constexpr int RADIX = 256;

@@ -127,12 +128,11 @@ struct RadixRowState {

 struct PersistentTopKParams {
  const float* __restrict__ input;  // [num_rows, stride]
-  int32_t* __restrict__ output;     // [num_rows, top_k]
+  int32_t* __restrict__ output;     // [num_rows, TopK]
  int32_t* __restrict__ lengths;    // [num_rows]
  RadixRowState* row_states;        // large path: per-group state
  uint32_t num_rows;
  uint32_t stride;
-  uint32_t top_k;           // actual k value for output stride
  uint32_t chunk_size;      // large path: elements per CTA
  uint32_t ctas_per_group;  // 1=medium, >1=large
  uint32_t max_seq_len;     // max seq_len across all rows (for early CTA exit)
@@ -154,7 +154,6 @@ __device__ __forceinline__ uint32_t decode_bin(float x) {
  return key >> 5;
 }

-template <int TopK>
 __device__ __noinline__ void histogram_2048_topk(
    const float* __restrict__ logits, int32_t* __restrict__ output_indices,
    int32_t seq_len) {
@@ -419,7 +418,6 @@ __device__ __noinline__ void histogram_2048_topk(
 // by: DarkSharpness
 // which at the same time is an optimized topk kernel copied from tilelang
 // kernel
-template <int TopK>
 __device__ __noinline__ void histogram_256_topk(
    const float* __restrict__ logits, int* __restrict__ output_indices,
    int logits_offset, int seq_len) {
@@ -651,7 +649,7 @@ __device__ __forceinline__ void wait_ge(int* ptr, int target_val,
 // Adapted from https://github.com/flashinfer-ai/flashinfer/pull/2215
 // ============================================================================

-template <int TopK, uint32_t VEC_SIZE>
+template <uint32_t VEC_SIZE>
 __device__ void radix_topk(const float* __restrict__ row_input,
                           int32_t* __restrict__ row_output, uint32_t seq_len,
                           uint32_t my_chunk_start, uint32_t chunk_size,
@@ -859,7 +857,7 @@ __device__ void radix_topk(const float* __restrict__ row_input,
 // see filtered_topk.cuh)
 // ============================================================================

-template <int TopK = 2048, uint32_t VEC_SIZE = 1>
+template <uint32_t VEC_SIZE = 1>
 __global__ void __launch_bounds__(kThreadsPerBlock, 2)
    persistent_topk_kernel(PersistentTopKParams params) {
  const uint32_t tx = threadIdx.x;
@@ -917,7 +915,7 @@ __global__ void __launch_bounds__(kThreadsPerBlock, 2)
    if (row_idx >= params.num_rows) break;

    const uint32_t seq_len = params.lengths[row_idx];
-    int32_t* row_output = params.output + row_idx * params.top_k;
+    int32_t* row_output = params.output + row_idx * TopK;
    const float* row_input = params.input + row_idx * params.stride;

    if (seq_len <= RADIX_THRESHOLD) {
@@ -929,19 +927,19 @@ __global__ void __launch_bounds__(kThreadsPerBlock, 2)
            row_output[i] = (i < seq_len) ? static_cast<int32_t>(i) : -1;
          }
        } else if (seq_len <= static_cast<uint32_t>(HIST2048_THRESHOLD)) {
-          histogram_2048_topk<TopK>(row_input, row_output, seq_len);
+          histogram_2048_topk(row_input, row_output, seq_len);
        } else {
-          histogram_256_topk<TopK>(row_input, row_output, 0, seq_len);
+          histogram_256_topk(row_input, row_output, 0, seq_len);
        }
      }
      continue;
    }

    const uint32_t my_chunk_start = cta_in_group * chunk_size;
-    radix_topk<TopK, VEC_SIZE>(
-        row_input, row_output, seq_len, my_chunk_start, chunk_size,
-        local_histogram, suffix_sum, shared_scalars, shared_ordered, state,
-        cta_in_group, ctas_per_group, barrier_phase, iter, tx);
+    radix_topk<VEC_SIZE>(row_input, row_output, seq_len, my_chunk_start,
+                         chunk_size, local_histogram, suffix_sum,
+                         shared_scalars, shared_ordered, state, cta_in_group,
+                         ctas_per_group, barrier_phase, iter, tx);
  }
 }

@@ -1013,6 +1011,7 @@ struct FilteredTopKTraits<float> {
  }
 };

+constexpr uint32_t FILTERED_TOPK_MAX_K = 2048;
 constexpr uint32_t FILTERED_TOPK_BLOCK_THREADS = 1024;
 constexpr uint32_t FILTERED_TOPK_SMEM_INPUT_SIZE =
    16 * 1024;  // 16K indices per buffer
@@ -1026,7 +1025,7 @@ constexpr size_t FILTERED_TOPK_SMEM_DYNAMIC =
 * \tparam IdType Index type (int32_t)
 * \tparam VEC_SIZE Vector size for input loads (1, 2, 4, or 8)
 */
-template <typename DType, typename IdType, int VEC_SIZE, uint32_t MAX_K = 2048>
+template <typename DType, typename IdType, int VEC_SIZE>
 __global__ void __launch_bounds__(FILTERED_TOPK_BLOCK_THREADS)
    FilteredTopKUnifiedKernel(const DType* __restrict__ input,
                              IdType* __restrict__ output,
@@ -1060,7 +1059,7 @@ __global__ void __launch_bounds__(FILTERED_TOPK_BLOCK_THREADS)
  alignas(128) __shared__ int s_counter;
  alignas(128) __shared__ int s_threshold_bin_id;
  alignas(128) __shared__ int s_num_input[2];
-  alignas(128) __shared__ int s_indices[MAX_K];
+  alignas(128) __shared__ int s_indices[FILTERED_TOPK_MAX_K];

  auto& s_histogram = s_histogram_buf[0];

@@ -1281,7 +1280,7 @@ constexpr int ComputeFilteredTopKVecSize(uint32_t max_len) {
  return static_cast<int>(g);
 }

-template <typename DType, typename IdType, uint32_t MAX_K = 2048>
+template <typename DType, typename IdType>
 cudaError_t FilteredTopKRaggedTransform(DType* input, IdType* output_indices,
                                        IdType* lengths, uint32_t num_rows,
                                        uint32_t top_k_val, uint32_t max_len,
@@ -1298,7 +1297,7 @@ cudaError_t FilteredTopKRaggedTransform(DType* input, IdType* output_indices,

 #define DISPATCH_VEC_SIZE(VS)                                               \
  if (vec_size == VS) {                                                     \
-    auto kernel = FilteredTopKUnifiedKernel<DType, IdType, VS, MAX_K>;      \
+    auto kernel = FilteredTopKUnifiedKernel<DType, IdType, VS>;             \
    FLASHINFER_CUDA_CALL(cudaFuncSetAttribute(                              \
        kernel, cudaFuncAttributeMaxDynamicSharedMemorySize, smem_size));   \
    FLASHINFER_CUDA_CALL(cudaLaunchKernel((void*)kernel, grid, block, args, \
@@ -9,29 +9,28 @@ namespace vllm {

 template <typename scalar_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,
-    const bool inverse) {
+    scalar_t* __restrict__ arr, const scalar_t* __restrict__ cos_ptr,
+    const scalar_t* __restrict__ sin_ptr, int rot_offset, int embed_dim) {
  int x_index, y_index;
-  float cos_f, sin_f;
+  scalar_t cos, sin;
  if (IS_NEOX) {
+    // GPT-NeoX style rotary embedding.
    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 = VLLM_LDG(cos_ptr + x_index);
+    sin = VLLM_LDG(sin_ptr + x_index);
  } else {
+    // GPT-J style rotary embedding.
    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 = VLLM_LDG(cos_ptr + x_index / 2);
+    sin = VLLM_LDG(sin_ptr + x_index / 2);
  }
-  if (inverse) {
-    sin_f = -sin_f;
-  }
-  const float x_f = static_cast<float>(arr[x_index]);
-  const float y_f = static_cast<float>(arr[y_index]);
-  arr[x_index] = static_cast<scalar_t>(x_f * cos_f - y_f * sin_f);
-  arr[y_index] = static_cast<scalar_t>(y_f * cos_f + x_f * sin_f);
+
+  const scalar_t x = arr[x_index];
+  const scalar_t y = arr[y_index];
+  arr[x_index] = x * cos - y * sin;
+  arr[y_index] = y * cos + x * sin;
 }

 template <typename scalar_t, bool IS_NEOX>
@@ -43,23 +42,22 @@ 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 scalar_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 int64_t head_stride) {
  const int embed_dim = rot_dim / 2;
-  const float* cos_ptr = cache_ptr;
-  const float* sin_ptr = cache_ptr + embed_dim;
+  const scalar_t* cos_ptr = cache_ptr;
+  const scalar_t* sin_ptr = cache_ptr + embed_dim;

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

  if (key != nullptr) {
@@ -67,10 +65,10 @@ inline __device__ void apply_rotary_embedding(
    for (int i = threadIdx.x; i < nk; i += blockDim.x) {
      const int head_idx = i / embed_dim;
      const int64_t token_head =
-          token_idx * key_stride + head_idx * head_stride + rope_dim_offset;
+          token_idx * key_stride + head_idx * head_stride;
      const int rot_offset = i % embed_dim;
      apply_token_rotary_embedding<scalar_t, IS_NEOX>(
-          key + token_head, cos_ptr, sin_ptr, rot_offset, embed_dim, inverse);
+          key + token_head, cos_ptr, sin_ptr, rot_offset, embed_dim);
    }
  }
 }
@@ -86,18 +84,19 @@ __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 scalar_t* __restrict__ cos_sin_cache,  // [max_position, 2, rot_dim //
+                                                 // 2]
    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 head_size) {
+  // Each thread block is responsible for one token.
  const int token_idx = blockIdx.x;
  int64_t pos = positions[token_idx];
-  const float* cache_ptr = cos_sin_cache + pos * rot_dim;
+  const scalar_t* cache_ptr = cos_sin_cache + pos * rot_dim;

  apply_rotary_embedding<scalar_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);
+      token_idx, query_stride, key_stride, head_stride);
 }

 }  // namespace vllm
@@ -116,7 +115,7 @@ void rotary_embedding(
    // [num_tokens, num_heads, head_size]
    int64_t head_size,
    torch::Tensor& cos_sin_cache,  // [max_position, rot_dim]
-    bool is_neox, int64_t rope_dim_offset, bool inverse) {
+    bool is_neox) {
  // num_tokens = batch_size * seq_len
  int64_t num_tokens = positions.numel();
  int positions_ndim = positions.dim();
@@ -155,8 +154,6 @@ void rotary_embedding(
  int seq_dim_idx = positions_ndim - 1;
  int64_t query_stride = query.stride(seq_dim_idx);
  int64_t key_stride = key.has_value() ? key->stride(seq_dim_idx) : 0;
-
-  TORCH_CHECK((rot_dim + rope_dim_offset) <= head_size);
  // Determine head stride: for [*, heads, head_size] use stride of last dim;
  // for flat [*, heads*head_size], heads blocks are contiguous of size
  // head_size
@@ -168,23 +165,20 @@ 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);
+          cos_sin_cache.data_ptr<scalar_t>(), rot_dim, query_stride, key_stride,
+          head_stride, num_heads, num_kv_heads, head_size);
    } 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);
+              cos_sin_cache.data_ptr<scalar_t>(), rot_dim, query_stride,
+              key_stride, head_stride, num_heads, num_kv_heads, head_size);
    }
  });
 }
@@ -40,6 +40,15 @@ using __hip_fp8_e5m2 = __hip_fp8_e5m2_fnuz;
  #define __HIP__FP8MFMA__
 #endif

+#if defined(__HIPCC__) && (defined(__gfx1100__) || defined(__gfx1101__) || \
+                           defined(__gfx1150__) || defined(__gfx1151__))
+  #define __HIP__GFX11__
+#endif
+
+#if defined(__HIPCC__) && (defined(__gfx1200__) || defined(__gfx1201__))
+  #define __HIP__GFX12__
+#endif
+
 #if defined(NDEBUG)
  #undef NDEBUG
  #include <assert.h>
@@ -1620,7 +1629,7 @@ __launch_bounds__(NUM_THREADS) void paged_attention_ll4mi_reduce_kernel(
  }
 }

-#elif defined(__GFX11__)
+#elif defined(__HIP__GFX11__)

 using floatx8 = __attribute__((__vector_size__(8 * sizeof(float)))) float;

@@ -2379,7 +2388,7 @@ __launch_bounds__(NUM_THREADS) void paged_attention_ll4mi_reduce_kernel(
  out_ptr[threadIdx.x] = from_float<scalar_t>(acc);
 }

-#elif defined(__GFX12__)
+#elif defined(__HIP__GFX12__)

 using floatx8 = __attribute__((__vector_size__(8 * sizeof(float)))) float;

@@ -26,11 +26,16 @@
  #define __HIP__GFX9__
 #endif

-// Combined RDNA macro (gfx11 + gfx12) - both use 32-wide wavefronts
-#if defined(__GFX11__) || defined(__GFX12__)
+#if defined(__HIPCC__) &&                                                    \
+    (defined(__gfx1100__) || defined(__gfx1101__) || defined(__gfx1150__) || \
+     defined(__gfx1151__) || defined(__gfx1200__) || defined(__gfx1201__))
  #define __HIP__GFX1X__
 #endif

+#if defined(__HIPCC__) && (defined(__gfx1200__) || defined(__gfx1201__))
+  #define __HIP__GFX12__
+#endif
+
 #if defined(__HIPCC__) && (defined(__gfx942__) || defined(__gfx950__))
  #define __HIP__MI3XX__
 #endif
@@ -1840,7 +1845,7 @@ torch::Tensor wvSplitKrc(const at::Tensor& in_a, const at::Tensor& in_b,
  return out_c;
 }

-#if defined(__HIP__MI3XX__) || defined(__GFX12__)
+#if defined(__HIP__MI3XX__) || defined(__HIP__GFX12__)
 template <typename scalar_t, typename fp8_t, int THRDS, int YTILE, int WvPrGrp,
          int A_CHUNK, int UNRL, int N>
 __global__ void __launch_bounds__(WvPrGrp* THRDS)
@@ -1888,7 +1893,7 @@ __global__ void __launch_bounds__(WvPrGrp* THRDS)
  float sB = *s_B;

  while (m < M) {
-  #ifdef __GFX12__
+  #ifdef __HIP__GFX12__
    // gfx12: per-lane scalar accumulation via v_dot4_f32_fp8_fp8
    float sum[N][YTILE] = {};
  #else
@@ -1926,7 +1931,7 @@ __global__ void __launch_bounds__(WvPrGrp* THRDS)
  #pragma unroll
      for (uint32_t k2 = 0; k2 < UNRL; k2++) {
        for (uint32_t n = 0; n < N; n++) {
-  #ifdef __GFX12__
+  #ifdef __HIP__GFX12__
          // gfx12: 4 x dot4 per A_CHUNK=16 bytes (4 FP8 per dot4)
          for (int y = 0; y < YTILE; ++y) {
    #pragma unroll
@@ -1950,7 +1955,7 @@ __global__ void __launch_bounds__(WvPrGrp* THRDS)
    }

    // Final reduction
-  #ifdef __GFX12__
+  #ifdef __HIP__GFX12__
    // gfx12 wave32: DPP row_shr within 16-lane rows + cross-row shuffle
    for (int n = 0; n < N; n++) {
      for (int y = 0; y < YTILE; y++) {
@@ -1988,7 +1993,7 @@ __global__ void __launch_bounds__(WvPrGrp* THRDS)
  #endif

    const bool writeback_lane =
-  #ifdef __GFX12__
+  #ifdef __HIP__GFX12__
        threadIdx.x == (THRDS - 1);
  #else
        threadIdx.x == 0;
@@ -2004,7 +2009,7 @@ __global__ void __launch_bounds__(WvPrGrp* THRDS)
      for (int n = 0; n < N; n++) {
        for (int y = 0; y < YTILE; y++) {
          if (y + m >= M) break;  // To avoid mem access fault.
-  #ifdef __GFX12__
+  #ifdef __HIP__GFX12__
          float result = sum[n][y] * sA * sB;
  #else
          float result = sum[n][y][0] * sA * sB;
@@ -2022,7 +2027,7 @@ __global__ void __launch_bounds__(WvPrGrp* THRDS)
    m += CuCount * _WvPrGrp * YTILE;
  }
 }
-#else   // !defined(__HIP__MI3XX__) && !defined(__GFX12__)
+#else   // !defined(__HIP__MI3XX__) && !defined(__HIP__GFX12__)
 template <typename scalar_t, typename fp8_t, int THRDS, int YTILE, int WvPrGrp,
          int A_CHUNK, int UNRL, int N>
 __global__ void wvSplitKQ_hf_sml_(const int K, const int Kap, const int Kbp,
@@ -2034,9 +2039,9 @@ __global__ void wvSplitKQ_hf_sml_(const int K, const int Kap, const int Kbp,
                                  const int _WvPrGrp, const int CuCount) {
  UNREACHABLE_CODE
 }
-#endif  // defined(__HIP__MI3XX__) || defined(__GFX12__)
+#endif  // defined(__HIP__MI3XX__) || defined(__HIP__GFX12__)

-#if defined(__HIP__MI3XX__) || defined(__GFX12__)
+#if defined(__HIP__MI3XX__) || defined(__HIP__GFX12__)
 template <typename scalar_t, typename fp8_t, int THRDS, int YTILE, int WvPrGrp,
          int A_CHUNK, int UNRL, int N>
 __global__ void __launch_bounds__(WvPrGrp* THRDS)
@@ -2083,7 +2088,7 @@ __global__ void __launch_bounds__(WvPrGrp* THRDS)
  float sB = *s_B;

  while (m < M) {
-  #ifdef __GFX12__
+  #ifdef __HIP__GFX12__
    // gfx12: per-lane scalar accumulation via v_dot4_f32_fp8_fp8
    float sum[N][YTILE] = {};
  #else
@@ -2123,7 +2128,7 @@ __global__ void __launch_bounds__(WvPrGrp* THRDS)
  #pragma unroll
      for (uint32_t k2 = 0; k2 < UNRL; k2++) {
        for (uint32_t n = 0; n < N; n++) {
-  #ifdef __GFX12__
+  #ifdef __HIP__GFX12__
          // gfx12: 4 x dot4 per A_CHUNK=16 bytes (4 FP8 per dot4)
          for (int y = 0; y < YTILE; ++y) {
    #pragma unroll
@@ -2147,7 +2152,7 @@ __global__ void __launch_bounds__(WvPrGrp* THRDS)
    }

    // Final reduction
-  #ifdef __GFX12__
+  #ifdef __HIP__GFX12__
    // gfx12 wave32: DPP row_shr within 16-lane rows + cross-row shuffle
    for (int n = 0; n < N; n++) {
      for (int y = 0; y < YTILE; y++) {
@@ -2185,7 +2190,7 @@ __global__ void __launch_bounds__(WvPrGrp* THRDS)
  #endif

    const bool writeback_lane =
-  #ifdef __GFX12__
+  #ifdef __HIP__GFX12__
        threadIdx.x == (THRDS - 1);
  #else
        threadIdx.x == 0;
@@ -2201,7 +2206,7 @@ __global__ void __launch_bounds__(WvPrGrp* THRDS)
      for (int n = 0; n < N; n++) {
        for (int y = 0; y < YTILE; y++) {
          if (y + m >= M) break;  // To avoid mem access fault.
-  #ifdef __GFX12__
+  #ifdef __HIP__GFX12__
          float result = sum[n][y] * sA * sB;
  #else
          float result = sum[n][y][0] * sA * sB;
@@ -2219,7 +2224,7 @@ __global__ void __launch_bounds__(WvPrGrp* THRDS)
    m += CuCount * _WvPrGrp * YTILE;
  }
 }
-#else   // !defined(__HIP__MI3XX__) && !defined(__GFX12__)
+#else   // !defined(__HIP__MI3XX__) && !defined(__HIP__GFX12__)
 template <typename scalar_t, typename fp8_t, int THRDS, int YTILE, int WvPrGrp,
          int A_CHUNK, int UNRL, int N>
 __global__ void wvSplitKQ_hf_(const int K, const int Kap, const int Kbp,
@@ -2231,7 +2236,7 @@ __global__ void wvSplitKQ_hf_(const int K, const int Kap, const int Kbp,
                              const int CuCount) {
  UNREACHABLE_CODE
 }
-#endif  // defined(__HIP__MI3XX__) || defined(__GFX12__)
+#endif  // defined(__HIP__MI3XX__) || defined(__HIP__GFX12__)

 void wvSplitKQ(const at::Tensor& in_b, const at::Tensor& in_a,
               const std::optional<at::Tensor>& in_bias, at::Tensor& out_c,
@@ -258,13 +258,7 @@ __device__ bool processHistogramStep(
  auto processBins = [&](float logit, int idx) {
    if (isPartialMatch<patternShift>(logit, logitPattern)) {
      uint32_t binIdx = extractBinIdx<step>(logit);
-      // Only write elements with binIdx < thresholdBinIdx when:
-      // 1. This is step 0 and the threshold bin is small enough (no step 1)
-      // 2. This is step >= 1 (where pattern matching filters correctly)
-      // This prevents duplicates when step 0 and step 1 both run.
-      bool shouldWriteDirectly =
-          (step == 0 && smemFinalBinSize[0] <= kNumFinalItems) || (step >= 1);
-      if (binIdx < thresholdBinIdx && shouldWriteDirectly) {
+      if (binIdx < thresholdBinIdx) {
        // The element is part of the top-k selection
        int dstIdx = atomicAdd(&smemFoundTopKValues[0], 1);

@@ -10,17 +10,33 @@
  #include "persistent_topk.cuh"
 #endif

-namespace {
-
+void persistent_topk(const torch::Tensor& logits, const torch::Tensor& lengths,
+                     torch::Tensor& output, torch::Tensor& workspace, int64_t k,
+                     int64_t max_seq_len) {
 #ifndef USE_ROCM
-template <int TopK>
-void launch_persistent_topk(const torch::Tensor& logits,
-                            const torch::Tensor& lengths, torch::Tensor& output,
-                            torch::Tensor& workspace, int64_t max_seq_len) {
-  namespace P = vllm::persistent;
+  TORCH_CHECK(logits.is_cuda(), "logits must be CUDA tensor");
+  TORCH_CHECK(lengths.is_cuda(), "lengths must be CUDA tensor");
+  TORCH_CHECK(output.is_cuda(), "output must be CUDA tensor");
+  TORCH_CHECK(logits.dtype() == torch::kFloat32, "Only float32 supported");
+  TORCH_CHECK(lengths.dtype() == torch::kInt32, "lengths must be int32");
+  TORCH_CHECK(output.dtype() == torch::kInt32, "output must be int32");
+  TORCH_CHECK(logits.dim() == 2, "logits must be 2D");
+  TORCH_CHECK(lengths.dim() == 1 || lengths.dim() == 2,
+              "lengths must be 1D or 2D");
+  TORCH_CHECK(lengths.is_contiguous(), "lengths must be contiguous");
+  TORCH_CHECK(output.dim() == 2, "output must be 2D");

  const int64_t num_rows = logits.size(0);
  const int64_t stride = logits.size(1);
+
+  TORCH_CHECK(lengths.numel() == num_rows, "lengths size mismatch");
+  TORCH_CHECK(output.size(0) == num_rows && output.size(1) == k,
+              "output size mismatch");
+  namespace P = vllm::persistent;
+
+  TORCH_CHECK(k == P::TopK, "k must be 2048");
+  TORCH_CHECK(k <= stride, "k out of range");
+
  cudaStream_t stream = at::cuda::getCurrentCUDAStream();

  static int num_sms = 0;
@@ -34,17 +50,18 @@ void launch_persistent_topk(const torch::Tensor& logits,
  }

  if (num_rows > 32 && max_smem_per_block >= 128 * 1024) {
-    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);
+    cudaError_t status = vllm::FilteredTopKRaggedTransform<float, int32_t>(
+        logits.data_ptr<float>(), output.data_ptr<int32_t>(),
+        lengths.data_ptr<int32_t>(), static_cast<uint32_t>(num_rows),
+        static_cast<uint32_t>(k), static_cast<uint32_t>(stride), stream);
    TORCH_CHECK(status == cudaSuccess,
                "FilteredTopK failed: ", cudaGetErrorString(status));
  } else {
    TORCH_CHECK(workspace.is_cuda(), "workspace must be CUDA tensor");
    TORCH_CHECK(workspace.dtype() == torch::kUInt8, "workspace must be uint8");

+    // Smem cap: smaller smem → more CTAs/group → more per-row parallelism for
+    // large path. Empirically tuned.
    int effective_max_smem;
    if (num_rows <= 4) {
      effective_max_smem =
@@ -82,73 +99,18 @@ 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;
-    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));
+    cudaOccupancyMaxActiveBlocksPerMultiprocessor(
+        &occupancy, P::persistent_topk_kernel<4>, P::kThreadsPerBlock,
+        smem_size);
    if (occupancy < 1) occupancy = 1;

-    // 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 max_resident_ctas = static_cast<uint32_t>(num_sms) * occupancy;
    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");
@@ -159,16 +121,15 @@ void launch_persistent_topk(const torch::Tensor& logits,
    params.lengths = lengths.data_ptr<int32_t>();
    params.num_rows = static_cast<uint32_t>(num_rows);
    params.stride = static_cast<uint32_t>(stride);
-    params.top_k = static_cast<uint32_t>(TopK);
    params.chunk_size = chunk_size;
    params.row_states =
        reinterpret_cast<P::RadixRowState*>(workspace.data_ptr<uint8_t>());
    params.ctas_per_group = ctas_per_group;
    params.max_seq_len = static_cast<uint32_t>(max_seq_len);

-  #define LAUNCH_PERSISTENT(TOPK_VAL, VS)                                     \
+  #define LAUNCH_PERSISTENT(VS)                                               \
    do {                                                                      \
-      auto kernel = &P::persistent_topk_kernel<TOPK_VAL, VS>;                 \
+      auto kernel = &P::persistent_topk_kernel<VS>;                           \
      cudaError_t err = cudaFuncSetAttribute(                                 \
          kernel, cudaFuncAttributeMaxDynamicSharedMemorySize, smem_size);    \
      TORCH_CHECK(err == cudaSuccess,                                         \
@@ -177,11 +138,11 @@ void launch_persistent_topk(const torch::Tensor& logits,
    } while (0)

    if (vec_size == 4) {
-      LAUNCH_PERSISTENT(TopK, 4);
+      LAUNCH_PERSISTENT(4);
    } else if (vec_size == 2) {
-      LAUNCH_PERSISTENT(TopK, 2);
+      LAUNCH_PERSISTENT(2);
    } else {
-      LAUNCH_PERSISTENT(TopK, 1);
+      LAUNCH_PERSISTENT(1);
    }
  #undef LAUNCH_PERSISTENT
  }
@@ -189,46 +150,6 @@ void launch_persistent_topk(const torch::Tensor& logits,
  cudaError_t err = cudaGetLastError();
  TORCH_CHECK(err == cudaSuccess,
              "persistent_topk failed: ", cudaGetErrorString(err));
-}
-#endif
-
-}  // anonymous namespace
-
-void persistent_topk(const torch::Tensor& logits, const torch::Tensor& lengths,
-                     torch::Tensor& output, torch::Tensor& workspace, int64_t k,
-                     int64_t max_seq_len) {
-#ifndef USE_ROCM
-  TORCH_CHECK(logits.is_cuda(), "logits must be CUDA tensor");
-  TORCH_CHECK(lengths.is_cuda(), "lengths must be CUDA tensor");
-  TORCH_CHECK(output.is_cuda(), "output must be CUDA tensor");
-  TORCH_CHECK(logits.dtype() == torch::kFloat32, "Only float32 supported");
-  TORCH_CHECK(lengths.dtype() == torch::kInt32, "lengths must be int32");
-  TORCH_CHECK(output.dtype() == torch::kInt32, "output must be int32");
-  TORCH_CHECK(logits.dim() == 2, "logits must be 2D");
-  TORCH_CHECK(lengths.dim() == 1 || lengths.dim() == 2,
-              "lengths must be 1D or 2D");
-  TORCH_CHECK(lengths.is_contiguous(), "lengths must be contiguous");
-  TORCH_CHECK(output.dim() == 2, "output must be 2D");
-
-  const int64_t num_rows = logits.size(0);
-  const int64_t stride = logits.size(1);
-
-  TORCH_CHECK(lengths.numel() == num_rows, "lengths size mismatch");
-  TORCH_CHECK(output.size(0) == num_rows && output.size(1) == k,
-              "output size mismatch");
-  TORCH_CHECK(k == 512 || k == 1024 || k == 2048,
-              "persistent_topk supports k=512, k=1024, or k=2048, got k=", k);
-
-  if (k == 512) {
-    launch_persistent_topk<512>(logits, lengths, output, workspace,
-                                max_seq_len);
-  } else if (k == 1024) {
-    launch_persistent_topk<1024>(logits, lengths, output, workspace,
-                                 max_seq_len);
-  } else {
-    launch_persistent_topk<2048>(logits, lengths, output, workspace,
-                                 max_seq_len);
-  }
 #else
  TORCH_CHECK(false, "persistent_topk is not supported on ROCm");
 #endif
@@ -106,12 +106,6 @@ 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);
@@ -183,19 +177,6 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
      "int forced_token_heads_per_warp=-1) -> ()");
  ops.impl("fused_qk_norm_rope", torch::kCUDA, &fused_qk_norm_rope);

-#ifndef USE_ROCM
-  // Horizontally-fused DeepseekV4-MLA: per-head RMSNorm + GPT-J RoPE for Q, and
-  // GPT-J RoPE + UE8M0 FP8 quant + paged cache insert for KV, all in one
-  // kernel launch.
-  ops.def(
-      "fused_deepseek_v4_qnorm_rope_kv_rope_quant_insert("
-      "Tensor! q, Tensor kv, Tensor! k_cache, "
-      "Tensor slot_mapping, Tensor position_ids, Tensor cos_sin_cache, "
-      "float eps, int cache_block_size) -> ()");
-  ops.impl("fused_deepseek_v4_qnorm_rope_kv_rope_quant_insert", torch::kCUDA,
-           &fused_deepseek_v4_qnorm_rope_kv_rope_quant_insert);
-#endif
-
  // Apply repetition penalties to logits in-place
  ops.def(
      "apply_repetition_penalties_(Tensor! logits, Tensor prompt_mask, "
@@ -259,8 +240,7 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
  ops.def(
      "rotary_embedding(Tensor positions, Tensor! query,"
      "                 Tensor!? key, int head_size,"
-      "                 Tensor cos_sin_cache, bool is_neox, int "
-      "rope_dim_offset=0, bool inverse=False) -> ()");
+      "                 Tensor cos_sin_cache, bool is_neox) -> ()");
  ops.impl("rotary_embedding", torch::kCUDA, &rotary_embedding);

  // Quantization ops
@@ -22,7 +22,7 @@
 #   docker buildx bake -f docker/docker-bake.hcl -f docker/versions.json
 # =============================================================================

-ARG CUDA_VERSION=13.0.2
+ARG CUDA_VERSION=13.0.0
 ARG PYTHON_VERSION=3.12
 ARG UBUNTU_VERSION=22.04

@@ -188,7 +188,7 @@ RUN --mount=type=cache,target=/root/.cache/uv \
 # Explicitly set the list to avoid issues with torch 2.2
 # See https://github.com/pytorch/pytorch/pull/123243
 # From versions.json: .torch.cuda_arch_list
-ARG torch_cuda_arch_list='7.5 8.0 8.6 8.9 9.0 10.0 11.0 12.0+PTX'
+ARG torch_cuda_arch_list='7.0 7.5 8.0 8.9 9.0 10.0 12.0'
 ENV TORCH_CUDA_ARCH_LIST=${torch_cuda_arch_list}
 #################### BUILD BASE IMAGE ####################

@@ -478,6 +478,9 @@ FROM ${FINAL_BASE_IMAGE} AS vllm-base

 ARG CUDA_VERSION
 ARG PYTHON_VERSION
+ARG DEADSNAKES_MIRROR_URL
+ARG DEADSNAKES_GPGKEY_URL
+ARG GET_PIP_URL

 ENV DEBIAN_FRONTEND=noninteractive
 WORKDIR /vllm-workspace
@@ -487,35 +490,43 @@ WORKDIR /vllm-workspace
 RUN PYTHON_VERSION_STR=$(echo ${PYTHON_VERSION} | sed 's/\.//g') && \
    echo "export PYTHON_VERSION_STR=${PYTHON_VERSION_STR}" >> /etc/environment

-# Install Python (via uv / python-build-standalone) and system dependencies.
-# This replaces the deadsnakes PPA, removing the build-time dependency on
-# Launchpad and matching how the build-stage (`base`) installs Python.
-# python-build-standalone bundles dev headers, the venv module, and
-# python3-config, so the python3.X-dev / python3.X-venv apt packages
-# are not needed.
+# Install Python and system dependencies
 RUN apt-get update -y \
    && apt-get install -y --no-install-recommends \
+        software-properties-common \
        curl \
        sudo \
        ffmpeg \
        libsm6 \
        libxext6 \
        libgl1 \
+    && if [ ! -z ${DEADSNAKES_MIRROR_URL} ] ; then \
+        if [ ! -z "${DEADSNAKES_GPGKEY_URL}" ] ; then \
+            mkdir -p -m 0755 /etc/apt/keyrings ; \
+            curl -L ${DEADSNAKES_GPGKEY_URL} | gpg --dearmor > /etc/apt/keyrings/deadsnakes.gpg ; \
+            sudo chmod 644 /etc/apt/keyrings/deadsnakes.gpg ; \
+            echo "deb [signed-by=/etc/apt/keyrings/deadsnakes.gpg] ${DEADSNAKES_MIRROR_URL} $(lsb_release -cs) main" > /etc/apt/sources.list.d/deadsnakes.list ; \
+        fi ; \
+    else \
+        for i in 1 2 3; do \
+            add-apt-repository -y ppa:deadsnakes/ppa && break || \
+            { echo "Attempt $i failed, retrying in 5s..."; sleep 5; }; \
+        done ; \
+    fi \
+    && apt-get update -y \
+    && apt-get install -y --no-install-recommends \
+        python${PYTHON_VERSION} \
+        python${PYTHON_VERSION}-dev \
+        python${PYTHON_VERSION}-venv \
        libibverbs-dev \
    && rm -rf /var/lib/apt/lists/* \
-    && curl -LsSf https://astral.sh/uv/install.sh | sh \
-    && $HOME/.local/bin/uv venv /opt/venv --python ${PYTHON_VERSION} \
-    && rm -f /usr/bin/python3 /usr/bin/python3-config /usr/bin/pip \
-    && ln -s /opt/venv/bin/python3 /usr/bin/python3 \
-    && ln -s /opt/venv/bin/python${PYTHON_VERSION} /usr/bin/python${PYTHON_VERSION} \
-    && ln -s /opt/venv/bin/python3-config /usr/bin/python3-config \
-    && ln -s /opt/venv/bin/pip /usr/bin/pip \
+    && update-alternatives --install /usr/bin/python3 python3 /usr/bin/python${PYTHON_VERSION} 1 \
+    && update-alternatives --set python3 /usr/bin/python${PYTHON_VERSION} \
+    && ln -sf /usr/bin/python${PYTHON_VERSION}-config /usr/bin/python3-config \
+    && rm -f /usr/lib/python${PYTHON_VERSION}/EXTERNALLY-MANAGED \
+    && curl -sS ${GET_PIP_URL} | python${PYTHON_VERSION} \
    && python3 --version && python3 -m pip --version

-# Activate virtual environment and add uv to PATH
-ENV PATH="/opt/venv/bin:/root/.local/bin:$PATH"
-ENV VIRTUAL_ENV="/opt/venv"
-
 # Install CUDA development tools for runtime JIT compilation
 # (FlashInfer, DeepGEMM, EP kernels all require compilation at runtime)
 RUN CUDA_VERSION_DASH=$(echo $CUDA_VERSION | cut -d. -f1,2 | tr '.' '-') && \
@@ -529,9 +540,7 @@ RUN CUDA_VERSION_DASH=$(echo $CUDA_VERSION | cut -d. -f1,2 | tr '.' '-') && \
        libcurand-dev-${CUDA_VERSION_DASH} \
        libcublas-${CUDA_VERSION_DASH} \
        # Required by fastsafetensors (fixes #20384)
-        libnuma-dev \
-        # numactl CLI for NUMA binding at runtime
-        numactl && \
+        libnuma-dev && \
    # Fixes nccl_allocator requiring nccl.h at runtime
    # https://github.com/vllm-project/vllm/blob/1336a1ea244fa8bfd7e72751cabbdb5b68a0c11a/vllm/distributed/device_communicators/pynccl_allocator.py#L22
    # NCCL packages don't use the cuda-MAJOR-MINOR naming convention,
@@ -540,6 +549,9 @@ RUN CUDA_VERSION_DASH=$(echo $CUDA_VERSION | cut -d. -f1,2 | tr '.' '-') && \
    apt-get install -y --no-install-recommends --allow-change-held-packages libnccl-dev=${NCCL_VER} libnccl2=${NCCL_VER} && \
    rm -rf /var/lib/apt/lists/*

+# Install uv for faster pip installs
+RUN python3 -m pip install uv
+
 # Environment for uv
 ENV UV_HTTP_TIMEOUT=500
 ENV UV_INDEX_STRATEGY="unsafe-best-match"
@@ -729,7 +741,7 @@ ENV HF_XET_HIGH_PERFORMANCE 1
 ENV HF_HUB_DOWNLOAD_TIMEOUT 60

 # Copy in the v1 package for testing (it isn't distributed yet)
-COPY vllm/v1 /opt/venv/lib/python${PYTHON_VERSION}/site-packages/vllm/v1
+COPY vllm/v1 /usr/local/lib/python${PYTHON_VERSION}/dist-packages/vllm/v1

 # Source code is used in the `python_only_compile.sh` test
 # We hide it inside `src/` so that this source code
@@ -753,7 +765,7 @@ ARG PIP_EXTRA_INDEX_URL UV_EXTRA_INDEX_URL
 ENV UV_HTTP_TIMEOUT=500

 # install kv_connectors if requested
-ARG torch_cuda_arch_list='7.5 8.0 8.6 8.9 9.0 10.0 11.0 12.0+PTX'
+ARG torch_cuda_arch_list='7.0 7.5 8.0 8.9 9.0 10.0 12.0'
 ENV TORCH_CUDA_ARCH_LIST=${torch_cuda_arch_list}
 RUN --mount=type=cache,target=/root/.cache/uv \
    --mount=type=bind,source=requirements/kv_connectors.txt,target=/tmp/kv_connectors.txt,ro \
@@ -77,7 +77,7 @@ RUN --mount=type=cache,target=/root/.cache/uv \
    uv pip install --system $pkgs --index-url https://download.pytorch.org/whl/nightly/cu128

 RUN --mount=type=cache,target=/root/.cache/uv \
-    uv pip install --system numba==0.65.0
+    uv pip install --system numba==0.61.2

 RUN --mount=type=cache,target=/root/.cache/uv \
 uv pip install --system -r requirements/common.txt
@@ -124,10 +124,10 @@ 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="f33a5599"
 ARG RIXL_REPO="https://github.com/ROCm/RIXL.git"
-ARG UCX_BRANCH="7009d7a1"
-ARG UCX_REPO="https://github.com/openucx/ucx.git"
+ARG UCX_BRANCH="da3fac2a"
+ARG UCX_REPO="https://github.com/ROCm/ucx.git"
 ENV ROCM_PATH=/opt/rocm
 ENV UCX_HOME=/usr/local/ucx
 ENV RIXL_HOME=/usr/local/rixl
@@ -165,7 +165,7 @@ RUN cd /usr/local/src && \
        --disable-doxygen-doc \
        --enable-optimizations \
        --enable-devel-headers \
-        --with-rocm=${ROCM_PATH} \
+        --with-rocm=/opt/rocm \
        --with-verbs \
        --with-dm \
        --enable-mt && \
@@ -186,12 +186,7 @@ RUN git clone ${RIXL_REPO} /opt/rixl && \
    ninja install

 # Generate RIXL wheel
-# Exclude libcore and libpull from auditwheel: transitive dependencies
-# that are not shipped in the wheel and vary across base images.
-RUN cd /opt/rixl && \
-    sed -i "s/--exclude 'libamdhip64\*'/--exclude 'libamdhip64*' --exclude 'libcore*' --exclude 'libpull*'/" \
-        contrib/build-wheel.sh && \
-    mkdir -p /app/install && \
+RUN cd /opt/rixl && mkdir -p /app/install && \
    ./contrib/build-wheel.sh \
        --output-dir /app/install \
        --rocm-dir ${ROCM_PATH} \
@@ -436,10 +431,6 @@ COPY --from=export_vllm /vllm_v1 /usr/local/lib/python${PYTHON_VERSION}/dist-pac
 ENV MIOPEN_DEBUG_CONV_DIRECT=0
 ENV MIOPEN_DEBUG_CONV_GEMM=0

-# Use legacy IPC mode for HSA to avoid GPU memory pinning issues with UCX rocm_ipc
-# See: https://github.com/ROCm/rocm-libraries/issues/6266
-ENV HSA_ENABLE_IPC_MODE_LEGACY=1
-
 # Source code is used in the `python_only_compile.sh` test
 # We hide it inside `src/` so that this source code
 # will not be imported by other tests
@@ -50,9 +50,9 @@ RUN curl -LsSf https://astral.sh/uv/install.sh | sh
 RUN uv venv --python ${PYTHON_VERSION} --seed ${VIRTUAL_ENV}
 ENV PATH="$VIRTUAL_ENV/bin:$PATH"

-# This oneccl contains the BMG support which is not the case for default version of oneapi 2025.3.
-ARG ONECCL_INSTALLER="intel-oneccl-2021.15.9.14_offline.sh"
-RUN wget "https://github.com/uxlfoundation/oneCCL/releases/download/2021.15.9/${ONECCL_INSTALLER}" && \
+# This oneccl contains the BMG support which is not the case for default version of oneapi 2025.2.
+ARG ONECCL_INSTALLER="intel-oneccl-2021.15.7.8_offline.sh"
+RUN wget "https://github.com/uxlfoundation/oneCCL/releases/download/2021.15.7/${ONECCL_INSTALLER}" && \
    bash "${ONECCL_INSTALLER}" -a --silent --eula accept && \
    rm "${ONECCL_INSTALLER}" && \
    echo "source /opt/intel/oneapi/setvars.sh --force" >> /root/.bashrc && \
@@ -164,7 +164,7 @@ RUN --mount=type=cache,target=/root/.cache/uv \
 # FIX triton
 RUN --mount=type=cache,target=/root/.cache/uv \
    uv pip uninstall triton triton-xpu && \
-    uv pip install triton-xpu==3.7.0
+    uv pip install triton-xpu==3.6.0

 # remove torch bundled oneccl to avoid conflicts
 RUN --mount=type=cache,target=/root/.cache/uv \
@@ -20,7 +20,7 @@ variable "NVCC_THREADS" {
 }

 variable "TORCH_CUDA_ARCH_LIST" {
-  default = "8.0 8.9 9.0 10.0 11.0 12.0"
+  default = "8.0 8.9 9.0 10.0"
 }

 variable "COMMIT" {
@@ -88,6 +88,7 @@ target "test-ubuntu2404" {
  args = {
    UBUNTU_VERSION          = "24.04"
    GDRCOPY_OS_VERSION      = "Ubuntu24_04"
+    FLASHINFER_AOT_COMPILE  = "true"
  }
  output = ["type=docker"]
 }
@@ -99,6 +100,7 @@ target "openai-ubuntu2404" {
  args = {
    UBUNTU_VERSION          = "24.04"
    GDRCOPY_OS_VERSION      = "Ubuntu24_04"
+    FLASHINFER_AOT_COMPILE  = "true"
  }
  output = ["type=docker"]
 }
@@ -2,7 +2,7 @@
  "_comment": "Auto-generated from Dockerfile ARGs. Do not edit manually. Run: python tools/generate_versions_json.py",
  "variable": {
    "CUDA_VERSION": {
-      "default": "13.0.2"
+      "default": "13.0.0"
    },
    "PYTHON_VERSION": {
      "default": "3.12"
@@ -11,10 +11,10 @@
      "default": "22.04"
    },
    "BUILD_BASE_IMAGE": {
-      "default": "nvidia/cuda:13.0.2-devel-ubuntu22.04"
+      "default": "nvidia/cuda:13.0.0-devel-ubuntu22.04"
    },
    "FINAL_BASE_IMAGE": {
-      "default": "nvidia/cuda:13.0.2-base-ubuntu22.04"
+      "default": "nvidia/cuda:13.0.0-base-ubuntu22.04"
    },
    "GET_PIP_URL": {
      "default": "https://bootstrap.pypa.io/get-pip.py"
@@ -32,7 +32,7 @@
      "default": "false"
    },
    "TORCH_CUDA_ARCH_LIST": {
-      "default": "7.5 8.0 8.6 8.9 9.0 10.0 11.0 12.0+PTX"
+      "default": "7.0 7.5 8.0 8.9 9.0 10.0 12.0"
    },
    "MAX_JOBS": {
      "default": "2"
@@ -108,38 +108,6 @@ P99 ITL (ms):                            8.39
 ==================================================
 ```

-#### Results Visualization
-
-The `--plot-timeline` and `--plot-dataset-stats` can be used to generate respectively the requests completion timeline and dataset prompt and output tokens statistics, which can be useful for debugging purpose or for deeper analysis.
-
-```bash
-vllm bench serve \
-    --backend vllm \
-    --model meta-llama/Llama-3.1-8B-Instruct \
-    --endpoint /v1/completions \
-    --dataset-name sharegpt \
-    --dataset-path <your data path>/ShareGPT_V3_unfiltered_cleaned_split.json \
-    --num-prompts 100 \
-    --plot-timeline \
-    --timeline-itl-thresholds 2,5 \
-    --plot-dataset-stats \
-    --save-result
-```
-
-##### Interactive Timeline
-
-The generated timeline is an interactive visualization in the form of an HTML file that can be rendered in most browsers. To customize the ITL color thresholds, one can use `--timeline-itl-thresholds` flag (default: 25ms, 50ms)
-
-Example output:
-
-<iframe src="../../assets/contributing/vllm_bench_serve_timeline.html" width="100%" height="600" frameborder="0"></iframe>
-
-##### Dataset statistics
-
-The generated figure shows the input prompt and output tokens distribution.
-
-Example output: ![Dataset Statistics](../assets/contributing/vllm_bench_serve_dataset_stats.png)
-
 #### Custom Dataset

 If the dataset you want to benchmark is not supported yet in vLLM, even then you can benchmark on it using `CustomDataset`. Your data needs to be in `.jsonl` format and needs to have "prompt" field per entry, e.g., data.jsonl
@@ -155,9 +155,9 @@ switch to `--physcpubind=<cpu-list> --membind=<node>`.

 These `--numa-bind*` options only apply to GPU execution processes. They do not
 configure the CPU backend's separate thread-affinity controls. Automatic
-GPU-to-NUMA detection is currently implemented for CUDA/NVML-based as well as
-ROCM-based platforms; other GPU backends must provide explicit binding lists if
-they use these options.
+GPU-to-NUMA detection is currently implemented for CUDA/NVML-based platforms;
+other GPU backends must provide explicit binding lists if they use these
+options.

 `--numa-bind-nodes` takes one non-negative NUMA node index per visible GPU, in
 the same order as the GPU indices.
@@ -4,7 +4,6 @@ Deploying vLLM on Kubernetes is a scalable and efficient way to serve machine le

 - [Deployment with CPUs](#deployment-with-cpus)
 - [Deployment with GPUs](#deployment-with-gpus)
- [Serving with gRPC](#serving-with-grpc)
 - [Troubleshooting](#troubleshooting)
    - [Startup Probe or Readiness Probe Failure, container log contains "KeyboardInterrupt: terminated"](#startup-probe-or-readiness-probe-failure-container-log-contains-keyboardinterrupt-terminated)
 - [Conclusion](#conclusion)
@@ -388,49 +387,6 @@ INFO:     Uvicorn running on http://0.0.0.0:8000 (Press CTRL+C to quit)

      If the service is correctly deployed, you should receive a response from the vLLM model.

-## Serving with gRPC
-
-vLLM can serve models over gRPC instead of HTTP by passing the `--grpc` flag. This requires the optional gRPC dependencies:
-
-```bash
-pip install vllm[grpc]
-```
-
-When using `--grpc`, the server exposes the standard [gRPC Health Checking Protocol](https://github.com/grpc/grpc/blob/master/doc/health-checking.md) (`grpc.health.v1.Health`), which integrates with Kubernetes [native gRPC probes](https://kubernetes.io/docs/tasks/configure-pod-container/configure-liveness-readiness-startup-probes/#define-a-grpc-liveness-probe) (available since Kubernetes 1.24).
-
-To deploy with gRPC, change the `vllm serve` command to include `--grpc` and replace `httpGet` probes with `grpc` probes:
-
-```yaml
-containers:
- name: mistral-7b
-  image: vllm/vllm-openai:latest
-  command: ["/bin/sh", "-c"]
-  args: [
-    "pip install vllm[grpc] && vllm serve mistralai/Mistral-7B-Instruct-v0.3 --grpc --port 50051 --trust-remote-code"
-  ]
-  ports:
-  - containerPort: 50051
-  livenessProbe:
-    grpc:
-      port: 50051
-    initialDelaySeconds: 120
-    periodSeconds: 10
-  readinessProbe:
-    grpc:
-      port: 50051
-    initialDelaySeconds: 120
-    periodSeconds: 5
-```
-
-!!! note
-    The gRPC health service checks the engine status on every probe. If the engine is unhealthy or the server is shutting down, the probe returns `NOT_SERVING`.
-
-You can also verify the health service manually with `grpcurl`:
-
-```bash
-grpcurl -plaintext localhost:50051 grpc.health.v1.Health/Check
-```
-
 ## Troubleshooting

 ### Startup Probe or Readiness Probe Failure, container log contains "KeyboardInterrupt: terminated"
@@ -172,7 +172,7 @@ Priority is **1 = highest** (tried first).
 | `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` | 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 |
@@ -213,7 +213,7 @@ configuration.
 | `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 |
+| `FLASHMLA_SPARSE` | bf16 | `auto`, `bfloat16`, `fp8_ds_mla` | 64 | 576 | ❌ | ✅ | ❌ | ❌ | Decoder | 9.x-10.x |
 | `FLASH_ATTN_MLA` | fp16, bf16 | `auto`, `float16`, `bfloat16` | %16 | Any | ❌ | ❌ | ❌ | ✅ | Decoder | 9.x |
 | `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 |
@@ -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,bf16,mxfp8 | 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 | 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
@@ -83,8 +83,8 @@ To be used with a particular `FusedMoEPrepareAndFinalizeModular` subclass, MoE k
 | triton | standard | all<sup>1</sup> | G,A,T | silu, gelu,</br>swigluoai,</br>silu_no_mul,</br>gelu_no_mul | Y | Y | [`fused_experts`][vllm.model_executor.layers.fused_moe.fused_moe.fused_experts],</br>[`TritonExperts`][vllm.model_executor.layers.fused_moe.fused_moe.TritonExperts] |
 | triton (batched) | batched | all<sup>1</sup> | G,A,T | silu, gelu | <sup>6</sup> | Y | [`BatchedTritonExperts`][vllm.model_executor.layers.fused_moe.fused_batched_moe.BatchedTritonExperts] |
 | deep gemm | standard,</br>batched | fp8 | G(128),A,T | silu, gelu | <sup>6</sup> | Y | </br>[`DeepGemmExperts`][vllm.model_executor.layers.fused_moe.experts.deep_gemm_moe.DeepGemmExperts],</br>[`BatchedDeepGemmExperts`][vllm.model_executor.layers.fused_moe.experts.batched_deep_gemm_moe.BatchedDeepGemmExperts] |
-| cutlass_fp4 | standard,</br>batched | nvfp4 | A,T | silu | Y | Y | [`CutlassExpertsFp4`][vllm.model_executor.layers.fused_moe.experts.cutlass_moe.CutlassExpertsFp4] |
-| cutlass_fp8 | standard,</br>batched | fp8 | A,T | silu, gelu | Y | Y | [`CutlassExpertsFp8`][vllm.model_executor.layers.fused_moe.experts.cutlass_moe.CutlassExpertsFp8],</br>[`CutlasBatchedExpertsFp8`][vllm.model_executor.layers.fused_moe.experts.cutlass_moe.CutlassBatchedExpertsFp8] |
+| cutlass_fp4 | standard,</br>batched | nvfp4 | A,T | silu | Y | Y | [`CutlassExpertsFp4`][vllm.model_executor.layers.fused_moe.cutlass_moe.CutlassExpertsFp4] |
+| cutlass_fp8 | standard,</br>batched | fp8 | A,T | silu, gelu | Y | Y | [`CutlassExpertsFp8`][vllm.model_executor.layers.fused_moe.cutlass_moe.CutlassExpertsFp8],</br>[`CutlasBatchedExpertsFp8`][vllm.model_executor.layers.fused_moe.cutlass_moe.CutlassBatchedExpertsFp8] |
 | flashinfer | standard | nvfp4,</br>fp8 | T | <sup>5</sup> | N | Y | [`FlashInferExperts`][vllm.model_executor.layers.fused_moe.flashinfer_cutlass_moe.FlashInferExperts] |
 | gpt oss triton | standard | N/A | N/A | <sup>5</sup> | Y | Y | [`triton_kernel_fused_experts`][vllm.model_executor.layers.fused_moe.experts.gpt_oss_triton_kernels_moe.triton_kernel_fused_experts],</br>[`OAITritonExperts`][vllm.model_executor.layers.fused_moe.experts.gpt_oss_triton_kernels_moe.OAITritonExperts] |
 | marlin | standard,</br>batched | <sup>3</sup> / N/A | <sup>3</sup> / N/A | silu,</br>swigluoai | Y | Y | [`fused_marlin_moe`][vllm.model_executor.layers.fused_moe.fused_marlin_moe.fused_marlin_moe],</br>[`MarlinExperts`][vllm.model_executor.layers.fused_moe.fused_marlin_moe.MarlinExperts],</br>[`BatchedMarlinExperts`][vllm.model_executor.layers.fused_moe.fused_marlin_moe.BatchedMarlinExperts] |
@@ -1,70 +0,0 @@
-# Context Extension
-
-!!! note
-    The `--rope-scaling` parameter used in older versions of vLLM is no longer supported. Please use the `--hf-overrides` method with `rope_parameters` instead.
-This directory contains examples for extending the context length of models using vLLM.
-
-## Offline Inference Example
-
-The [`context_extension.py`](../../examples/offline_inference/context_extension) script demonstrates how to extend the context length of a Qwen model using the YARN method (rope_parameters) and run a simple chat example.
-
-### Usage
-
-```bash
-python examples/offline_inference/context_extension.py
-```
-
-## OpenAI Online Method
-
-You can also use vLLM's OpenAI-compatible API to serve models with extended context length.
-
-### Usage
-
-Run the vLLM server with the following command to extend the context length using YARN:
-
-```bash
-vllm serve Qwen/Qwen3-0.6B \
-  --hf-overrides '{"rope_parameters": {"factor": 4.0, "original_max_position_embeddings": 32768, "rope_theta": 1000000, "rope_type": "yarn"}}' \
-  --max-model-len 131072
-```
-
-### Client Example
-
-After starting the server, you can use the OpenAI Python client to interact with it:
-
-```python
-from openai import OpenAI
-
-client = OpenAI(
-    base_url="http://localhost:8000/v1",
-    api_key="token-abc123"  # Dummy API key, required by the client
-)
-
-response = client.chat.completions.create(
-    model="Qwen/Qwen3-0.6B",
-    messages=[
-        {"role": "system", "content": "You are a helpful assistant"},
-        {"role": "user", "content": "Hello"}
-    ],
-    max_tokens=128,
-    temperature=0.8,
-    top_p=0.95
-)
-
-print(response.choices[0].message.content)
-```
-
-### Key Parameters
-
-The available parameters depend on the `rope_type` you choose. For detailed information about all supported RoPE types and their specific parameters, please refer to the [Hugging Face Transformers RoPE documentation](https://huggingface.co/docs/transformers/main/en/internal/rope_utils#transformers.RopeParameters).
-
-Common parameters include:
-
- `rope_type`: The type of RoPE implementation (e.g., "yarn", "linear", "dynamic")
- `factor`: The factor by which to extend the context length
- `original_max_position_embeddings`: The original maximum position embeddings of the model
-
-The following parameters are specific to vLLM:
-
- `max_model_len`: The new maximum sequence length after extension (original * factor).
-  Used for KV cache pre‑allocation and request limit at serving time.
@@ -375,8 +375,8 @@ For (G)B300, we recommend using CUDA 13, as shown in the following command.

    ```bash
    DOCKER_BUILDKIT=1 docker build \
-    --build-arg CUDA_VERSION=13.0.2 \
-    --build-arg BUILD_BASE_IMAGE=nvidia/cuda:13.0.2-devel-ubuntu22.04 \
+    --build-arg CUDA_VERSION=13.0.1 \
+    --build-arg BUILD_BASE_IMAGE=nvidia/cuda:13.0.1-devel-ubuntu22.04 \
    --build-arg max_jobs=256 \
    --build-arg nvcc_threads=2 \
    --build-arg RUN_WHEEL_CHECK=false \
@@ -46,7 +46,7 @@ pip install -v -r requirements/xpu.txt

    !!! note
        - `triton` (without suffix) is for NVIDIA GPUs only. On XPU, using it instead of `triton-xpu` can cause correctness or runtime issues.
-        - For torch 2.11 (the version used in `requirements/xpu.txt`), the matching package is `triton-xpu==3.7.0`. If you use a different version of torch, check the corresponding `triton-xpu` version in [docker/Dockerfile.xpu](https://github.com/vllm-project/vllm/blob/main/docker/Dockerfile.xpu).
+        - For torch 2.10 (the version used in `requirements/xpu.txt`), the matching package is `triton-xpu==3.6.0`. If you use a different version of torch, check the corresponding `triton-xpu` version in [docker/Dockerfile.xpu](https://github.com/vllm-project/vllm/blob/main/docker/Dockerfile.xpu).

 - Finally, build and install vLLM XPU backend:

@@ -78,7 +78,7 @@ The scoring models is designed to compute similarity scores between two input pr
 |-----------------------|---------------|----------------------------------------------|--------------------|--------------------------|
 | `classify` (see note) | Sequence-wise | reranker score for each sequence             | `cross-encoder`    | linear classifier        |
 | `embed`               | Sequence-wise | vector representations for each sequence     | `bi-encoder`       | cosine similarity        |
-| `token_classify`      | Token-wise    | probability vector of classes for each token | N/A                | N/A                      |
+| `token_classify`      | Token-wise    | probability vector of classes for each token | nan                | nan                      |
 | `token_embed`         | Token-wise    | vector representations for each token        | `late-interaction` | late interaction(MaxSim) |

 !!! note
@@ -86,15 +86,14 @@ The scoring models is designed to compute similarity scores between two input pr

 ### Pooling Usages

-| Pooling Usages              | Description                                                                                                                                               |
-|-----------------------------|-----------------------------------------------------------------------------------------------------------------------------------------------------------|
-| Classification Usages       | Predicting which predefined category, class, or label best corresponds to a given input.                                                                  |
-| Embedding Usages            | Converts unstructured data (text, images, audio, etc.) into structured numerical vectors (embeddings).                                                    |
-| Token Classification Usages | Token-wise classification                                                                                                                                 |
-| Token Embedding Usages      | Token-wise embedding                                                                                                                                      |
-| Reward Usages               | Evaluates the quality of outputs generated by a language model, acting as a proxy for human preferences.                                                  |
-| Scoring Usages              | Computes similarity scores between two inputs. It supports three model types (aka `score_type`): `cross-encoder`, `late-interaction`, and `bi-encoder`.   |
-| Plugins Usages              | Allow users to customize input and output processors. For more information, please refer to [IO Processor Plugins](../../design/io_processor_plugins.md). |
+| Pooling Usages              | Description                                                                                                                                             |
+|-----------------------------|---------------------------------------------------------------------------------------------------------------------------------------------------------|
+| Classification Usages       | Predicting which predefined category, class, or label best corresponds to a given input.                                                                |
+| Embedding Usages            | Converts unstructured data (text, images, audio, etc.) into structured numerical vectors (embeddings).                                                  |
+| Token Classification Usages | Token-wise classification                                                                                                                               |
+| Token Embedding Usages      | Token-wise embedding                                                                                                                                    |
+| Scoring Usages              | Computes similarity scores between two inputs. It supports three model types (aka `score_type`): `cross-encoder`, `late-interaction`, and `bi-encoder`. |
+| Reward Usages               | Evaluates the quality of outputs generated by a language model, acting as a proxy for human preferences.                                                |

 We also have some special models that support multiple pooling tasks, or have specific usage scenarios, or support special inputs and outputs.

@@ -102,9 +101,9 @@ For more detailed information, please refer to the link below.

 - [Classification Usages](classify.md)
 - [Embedding Usages](embed.md)
+- [Reward Usages](reward.md)
 - [Token Classification Usages](token_classify.md)
 - [Token Embedding Usages](token_embed.md)
- [Reward Usages](reward.md)
 - [Scoring Usages](scoring.md)
 - [Specific Model Examples](specific_models.md)

@@ -114,17 +113,15 @@ Each pooling model in vLLM supports one or more of these tasks according to
 [Pooler.get_supported_tasks][vllm.model_executor.layers.pooler.Pooler.get_supported_tasks],
 enabling the corresponding APIs.

-### Offline APIs corresponding to pooling usages
+### Offline APIs corresponding to pooling tasks

-| Pooling Usages              | Dedicated API       | Pooling task for `LLM.encode` API | Score Types                | scoring function         |
-|-----------------------------|---------------------|-----------------------------------|----------------------------|--------------------------|
-| Classification Usages       | `LLM.classify(...)` | `classify`                        | `cross-encoder` (see note) | linear classifier        |
-| Embedding Usages            | `LLM.embed(...)`    | `embed`                           | `bi-encoder`               | cosine similarity        |
-| Token Classification Usages | N/A                 | `token_classify`                  | N/A                        | N/A                      |
-| Token Embedding Usages      | N/A                 | `token_embed`                     | `late-interaction`         | late interaction(MaxSim) |
-| Reward Usages               | N/A                 | `classify` & `token_classify`     | N/A                        | N/A                      |
-| Scoring Usages              | `LLM.score(...)`    | N/A                               | N/A                        | N/A                      |
-| Plugins Usages              | N/A                 | `plugin`                          | N/A                        | N/A                      |
+| Task             | APIs                                                                                  |
+|------------------|---------------------------------------------------------------------------------------|
+| `embed`          | `LLM.embed(...)`, `LLM.encode(..., pooling_task="embed")`, `LLM.score(...)`(see note) |
+| `classify`       | `LLM.classify(...)`, `LLM.encode(..., pooling_task="classify")`, `LLM.score(...)`     |
+| `token_classify` | `LLM.reward(...)`, `LLM.encode(..., pooling_task="token_classify")`                   |
+| `token_embed`    | `LLM.encode(..., pooling_task="token_embed")`, `LLM.score(...)`                       |
+| `plugin`         | `LLM.encode(..., pooling_task="plugin")`                                              |

 !!! note
    Only when a classification model outputs num_labels equal to 1 can it be used as a scoring model and have its scoring API enabled.
@@ -150,7 +147,7 @@ It is primarily designed for [score models](scoring.md).

 The [encode][vllm.LLM.encode] method is available to all pooling models in vLLM.

-Please use one of the more specific methods or set the task directly when using `LLM.encode`, refer to the [table above](#offline-apis-corresponding-to-pooling-usages).
+Please use one of the more specific methods or set the task directly when using `LLM.encode`, refer to the [table above](#offline-apis-corresponding-to-pooling-tasks).

 ### Examples

@@ -186,12 +183,9 @@ Our Pooling API (`/pooling`) is similar to `LLM.encode`, being applicable to all

 The input format is the same as [Embeddings API](embed.md#openai-compatible-embeddings-api), but the output data can contain an arbitrary nested list, not just a 1-D list of floats.

-Please use one of the more specific APIs or set the task directly when using the Pooling API, refer to the [table above](#offline-apis-corresponding-to-pooling-usages).
+Please use one of the more specific APIs or set the task directly when using the Pooling API, refer to the [table above](#offline-apis-corresponding-to-pooling-tasks).

-Code examples:
-
- [Online example](../../../examples/pooling/reward/token_reward_online.py)
- [Offline example](../../../examples/pooling/reward/token_reward_offline.py)
+Code example: [examples/pooling/pooling/pooling_online.py](../../../examples/pooling/pooling/pooling_online.py)

 ### Examples

@@ -292,10 +286,10 @@ Pooling models now support token-wise task.

 ### Score task

-`score` task is deprecated and will be removed in v0.20. Please use `classify` instead. Only when a
-classification model outputs num_labels equal to 1 can it be used as a scoring model and have its scoring API enabled.
+`score` task have has been removed in v0.21, use `classify` instead. Only when a classification model outputs num_labels
+equal to 1 can it be used as a scoring model and have its scoring API enabled.

 ### Pooling multitask support

-Pooling multitask support is deprecated and will be removed in v0.20. When the default pooling task is not what you want,
+Pooling multitask support has been removed in v0.21. When the default pooling task is not what you want,
 you need to manually specify it via `PoolerConfig(task=<task>)` offline or `--pooler-config.task <task>` online.
@@ -134,13 +134,3 @@ print(f"Data: {data!r}")
 ## Online Serving

 Please refer to the [pooling API](README.md#pooling-api). Pooling task corresponding to reward model types refer to the [table above](#summary).
-
-## More examples
-
-More examples can be found here: [examples/pooling/reward](../../../examples/pooling/reward)
-
-## Deprecated Features
-
-### `LLM.reward`
-
-`llm.reward` api is deprecated and will be removed in v0.23. Please use `LLM.encode` with `pooling_task="classify"` or `pooling_task="token_classify"` instead.
@@ -384,7 +384,6 @@ th {
 | `DeepseekForCausalLM` | DeepSeek | `deepseek-ai/deepseek-llm-67b-base`, `deepseek-ai/deepseek-llm-7b-chat`, etc. | ✅︎ | ✅︎ |
 | `DeepseekV2ForCausalLM` | DeepSeek-V2 | `deepseek-ai/DeepSeek-V2`, `deepseek-ai/DeepSeek-V2-Chat`, etc. | ✅︎ | ✅︎ |
 | `DeepseekV3ForCausalLM` | DeepSeek-V3 | `deepseek-ai/DeepSeek-V3`, `deepseek-ai/DeepSeek-R1`, `deepseek-ai/DeepSeek-V3.1`, etc. | ✅︎ | ✅︎ |
-| `DeepseekV4ForCausalLM` | DeepSeek-V4 | `deepseek-ai/DeepSeek-V4-Flash`, `deepseek-ai/DeepSeek-V4-Pro`, etc. | | |
 | `Dots1ForCausalLM` | dots.llm1 | `rednote-hilab/dots.llm1.base`, `rednote-hilab/dots.llm1.inst`, etc. | | ✅︎ |
 | `DotsOCRForCausalLM` | dots_ocr | `rednote-hilab/dots.ocr` | ✅︎ | ✅︎ |
 | `Ernie4_5ForCausalLM` | Ernie4.5 | `baidu/ERNIE-4.5-0.3B-PT`, etc. | ✅︎ | ✅︎ |
@@ -420,7 +419,6 @@ th {
 | `Grok1ForCausalLM` | Grok2 | `xai-org/grok-2` | ✅︎ | ✅︎ |
 | `HunYuanDenseV1ForCausalLM` | Hunyuan Dense | `tencent/Hunyuan-7B-Instruct` | ✅︎ | ✅︎ |
 | `HunYuanMoEV1ForCausalLM` | Hunyuan-A13B | `tencent/Hunyuan-A13B-Instruct`, `tencent/Hunyuan-A13B-Pretrain`, `tencent/Hunyuan-A13B-Instruct-FP8`, etc. | ✅︎ | ✅︎ |
-| `HYV3ForCausalLM` | HY3 | `tencent/Hy3-preview-Base`, `tencent/Hy3-preview` | ✅︎ | ✅︎ |
 | `HyperCLOVAXForCausalLM` | HyperCLOVAX-SEED-Think-14B | `naver-hyperclovax/HyperCLOVAX-SEED-Think-14B` | ✅︎ | ✅︎ |
 | `InternLMForCausalLM` | InternLM | `internlm/internlm-7b`, `internlm/internlm-chat-7b`, etc. | ✅︎ | ✅︎ |
 | `InternLM2ForCausalLM` | InternLM2 | `internlm/internlm2-7b`, `internlm/internlm2-chat-7b`, etc. | ✅︎ | ✅︎ |
@@ -644,10 +642,10 @@ Some models are supported only via the [Transformers modeling backend](#transfor
 !!! note
    `Gemma3nForConditionalGeneration` is only supported on V1 due to shared KV caching and it depends on `timm>=1.0.17` to make use of its
    MobileNet-v5 vision backbone.
-
+  
    Performance is not yet fully optimized mainly due to:
-
-    - Both audio and vision MM encoders use `transformers.AutoModel` implementation.
+  
+    - Both audio and vision MM encoders use `transformers.AutoModel` implementation.  
    - There's no PLE caching or out-of-memory swapping support, as described in [Google's blog](https://developers.googleblog.com/en/introducing-gemma-3n/). These features might be too model-specific for vLLM, and swapping in particular may be better suited for constrained setups.

 !!! note
@@ -1,15 +1,6 @@
 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project

-"""
-Example offline usage of token reward models.
-
-The key distinction between sequence classification and token classification
-lies in their output granularity: sequence classification produces a single
-result for an entire input sequence, whereas token classification yields a
-result for each individual token within the sequence.
-"""
-
 from argparse import Namespace

 from vllm import LLM, EngineArgs
@@ -45,14 +36,14 @@ def main(args: Namespace):
    llm = LLM(**vars(args))

    # Generate rewards. The output is a list of PoolingRequestOutput.
-    outputs = llm.encode(prompts, pooling_task="token_classify")
+    outputs = llm.reward(prompts)

    # Print the outputs.
    print("\nGenerated Outputs:\n" + "-" * 60)
    for prompt, output in zip(prompts, outputs):
        rewards = output.outputs.data
        print(f"Prompt: {prompt!r}")
-        print_embeddings(rewards.tolist(), prefix="Reward")
+        print_embeddings(rewards, prefix="Reward")
        print("-" * 60)


@@ -2463,12 +2463,6 @@ MODELS_NEED_VIDEO_METADATA = [
 ]


-MODELS_SUPPORT_VIT_CUDA_GRAPH = [
-    "qwen3_vl",
-    "qwen3_vl_moe",
-]
-
-
 def get_multi_modal_input(args):
    """
    return {
@@ -2581,29 +2575,6 @@ def apply_image_repeat(
    return inputs, inputs_with_empty_media


-def maybe_add_vit_cuda_graph_compilation_config(args, engine_args):
-    model = args.model_type
-    modality = args.modality
-    enable_vit_cuda_graph = args.enable_vit_cuda_graph
-
-    if enable_vit_cuda_graph and model in MODELS_SUPPORT_VIT_CUDA_GRAPH:
-        if modality == "image" or modality == "video":
-            vision_items_per_batch = 1
-        elif modality == "image+video":
-            vision_items_per_batch = 2
-        else:
-            raise ValueError(
-                f"modality={modality} is not supported for vit cuda graph."
-            )
-
-        engine_args.compilation_config = {
-            "cudagraph_mm_encoder": True,
-            "encoder_cudagraph_max_vision_items_per_batch": vision_items_per_batch,
-        }
-
-    return engine_args
-
-
@contextmanager
 def time_counter(enable: bool):
    if enable:
@@ -2654,28 +2625,33 @@ def parse_args():
        default=0,
        help="Set the seed when initializing `vllm.LLM`.",
    )
+
    parser.add_argument(
        "--image-repeat-prob",
        type=float,
        default=None,
        help="Simulates the hit-ratio for multi-modal preprocessor cache (if enabled)",
    )
+
    parser.add_argument(
        "--disable-mm-processor-cache",
        action="store_true",
        help="If True, disables caching of multi-modal processor.",
    )
+
    parser.add_argument(
        "--time-generate",
        action="store_true",
        help="If True, then print the total generate() call time",
    )
+
    parser.add_argument(
        "--use-different-prompt-per-request",
        action="store_true",
        help="If True, then use different prompt (with the same multi-modal "
        "data) for each request.",
    )
+
    parser.add_argument(
        "--verify-mm-cache-hit-with-uuids",
        action="store_true",
@@ -2689,11 +2665,6 @@ def parse_args():
        default=None,
        help="Tensor parallel size to override the model's default setting. ",
    )
-    parser.add_argument(
-        "--enable-vit-cuda-graph",
-        action="store_true",
-        help="If True, will enable vit cuda graph capture and replay for the model.",
-    )
    return parser.parse_args()


@@ -2727,7 +2698,6 @@ def main(args):
    engine_args.mm_processor_cache_gb = mm_processor_cache_gb
    if args.tensor_parallel_size is not None:
        engine_args.tensor_parallel_size = args.tensor_parallel_size
-    engine_args = maybe_add_vit_cuda_graph_compilation_config(args, engine_args)
    llm = LLM.from_engine_args(engine_args)

    # Don't want to check the flag multiple times, so just hijack `prompts`.
@@ -10,6 +10,7 @@ import uuid

 import aiohttp
 import msgpack
+import regex as re
 import zmq
 from quart import Quart, Request, make_response, request

@@ -24,10 +25,32 @@ decode_instances: list[dict] = []
 request_nums = 0
 app = Quart(__name__)

+IP_PORT_PATTERN = re.compile(r"//(\d{1,3}\.\d{1,3}\.\d{1,3}\.\d{1,3}):(\d+)")
+

 TRANSFER_TYPE = None


+def _append_whole_dict_unique(target_list, data_dict):
+    new_filtered = {k: v for k, v in data_dict.items() if k != "index"}
+    for existed in target_list:
+        existed_filtered = {k: v for k, v in existed.items() if k != "index"}
+        if existed_filtered == new_filtered:
+            return False
+    print("!!APPEND!!", data_dict)
+    target_list.append(data_dict)
+    transfer_mode = data_dict.get("transfer_mode", "unknown")
+    global TRANSFER_TYPE
+
+    if TRANSFER_TYPE is None:
+        TRANSFER_TYPE = transfer_mode
+        logger.info("SET TRANSFER TYPE TO %s", TRANSFER_TYPE)
+    elif transfer_mode != TRANSFER_TYPE:
+        raise ValueError(f"mismatched transfer mode {TRANSFER_TYPE} vs {transfer_mode}")
+
+    return True
+
+
 _list_lock = threading.RLock()


@@ -45,81 +68,23 @@ def _listen_for_register(hostname, port):
        if router_socket in socks:
            remote_addr, msg = router_socket.recv_multipart()
            data = msgpack.loads(msg)
-            if data.get("type") == "HELLO":
+            if data["type"] == "HELLO":
                pass
-            elif data.get("type") in ("P", "D"):
-                role = data["type"]
-                required_keys = {
-                    "http_address",
-                    "zmq_address",
-                    "dp_size",
-                    "tp_size",
-                    "transfer_mode",
-                }
-                missing = required_keys - data.keys()
-                if missing:
-                    logger.error(
-                        "Registration message missing required keys %s; skipping",
-                        missing,
-                    )
-                    continue
-                # Derive request_address from http_address
-                # api path suffix is appended at request time
-                instance = {
-                    "role": role,
-                    "request_address": f"http://{data['http_address']}/v1",
-                    "http_address": data["http_address"],
-                    "zmq_address": data["zmq_address"],
-                    "dp_size": data["dp_size"],
-                    "tp_size": data["tp_size"],
-                    "transfer_mode": data["transfer_mode"],
-                }
-                # zmq_address format: "host:IP,handshake:PORT,notify:PORT"
-                # Stored verbatim; embedded into the request_id by handle_request.
-
-                global TRANSFER_TYPE
-                transfer_mode = instance["transfer_mode"]
-                target_list = prefill_instances if role == "P" else decode_instances
+            elif (
+                data["type"] == "register"
+                and data["role"] == "P"
+                and data["request_address"] not in prefill_instances
+            ):
                with _list_lock:
-                    if TRANSFER_TYPE is None:
-                        TRANSFER_TYPE = transfer_mode
-                        logger.info("SET TRANSFER TYPE TO %s", TRANSFER_TYPE)
-                    elif transfer_mode != TRANSFER_TYPE:
-                        logger.error(
-                            "Mismatched transfer mode: expected %s, got %s;"
-                            " skipping registration of %s",
-                            TRANSFER_TYPE,
-                            transfer_mode,
-                            data["http_address"],
-                        )
-                        continue
-                    existing_idx = next(
-                        (
-                            idx
-                            for idx, i in enumerate(target_list)
-                            if i.get("http_address") == data["http_address"]
-                        ),
-                        None,
-                    )
-                    if existing_idx is not None:
-                        target_list[existing_idx] = instance
-                        logger.info(
-                            "Updated existing %s instance: %s",
-                            "Prefill" if role == "P" else "Decode",
-                            instance,
-                        )
-                    else:
-                        target_list.append(instance)
-                        logger.info(
-                            "Registered %s instance: %s",
-                            "Prefill" if role == "P" else "Decode",
-                            instance,
-                        )
-            else:
-                logger.warning(
-                    "Received message with unrecognized type %r; ignoring",
-                    data.get("type"),
-                )
+                    _append_whole_dict_unique(prefill_instances, data)
+
+            elif (
+                data["type"] == "register"
+                and data["role"] == "D"
+                and data["request_address"] not in decode_instances
+            ):
+                with _list_lock:
+                    _append_whole_dict_unique(decode_instances, data)


 def start_service_discovery(hostname, port):
@@ -136,7 +101,7 @@ def start_service_discovery(hostname, port):


 async def send_request_to_prefill(
-    endpoint, req_data, request_id, selected_prefill_dp_rank
+    endpoint, req_data, request_id, d_endpoint, dip, dport, selected_prefill_dp_rank
 ):
    req_data_copy = req_data

@@ -144,8 +109,12 @@ async def send_request_to_prefill(
        {
            "do_remote_decode": True,
            "do_remote_prefill": False,
+            "remote_handshake_port": d_endpoint["handshake_port"],
+            "remote_notify_port": d_endpoint["notify_port"],
            "remote_engine_id": None,
            "remote_block_ids": None,
+            "remote_host": dip,
+            "remote_port": dport,
        }
    )
    req_data_copy["stream"] = False
@@ -228,7 +197,14 @@ async def handle_request(api: str, request: Request):
            global request_nums
            request_nums += 1

+        def extract_ip_port_fast(url):
+            match = IP_PORT_PATTERN.search(url)
+            if not match:
+                raise ValueError(f"Invalid URL format: {url}")
+            return match.groups()
+
        req_data = await request.get_json()
+        request_id = str(uuid.uuid4())

        prefill_instance_endpoint = None
        decode_instance_endpoint = None
@@ -254,14 +230,7 @@ async def handle_request(api: str, request: Request):
                prefill_instance_endpoint["dp_size"],
            )

-        # Embed both zmq_addresses in the request_id so the connector can parse
-        # the peer's host/ports from it, similar to P2P-NCCL
-        uid = str(uuid.uuid4()).replace("-", "")
-        request_id = (
-            f"___prefill_addr_{prefill_instance_endpoint['zmq_address']}"
-            f"___decode_addr_{decode_instance_endpoint['zmq_address']}"
-            f"_{uid}"
-        )
+        dip, dport = extract_ip_port_fast(decode_instance_endpoint["request_address"])

        transfer_id = f"{MoRIIOConstants.TRANSFER_PREFIX}-{str(uuid.uuid4())}"

@@ -282,30 +251,35 @@ async def handle_request(api: str, request: Request):
                prefill_request_url,
                req_data_to_prefill,
                request_id,
+                decode_instance_endpoint,
+                dip,
+                dport,
                selected_prefill_dp_rank,
            )
        )
+        ip, port = extract_ip_port_fast(prefill_request_url)

        req_data["max_tokens"] -= 1

        req_data["kv_transfer_params"] = {
            "do_remote_decode": False,
            "do_remote_prefill": True,
+            "remote_handshake_port": prefill_instance_endpoint["handshake_port"],
+            "remote_notify_port": prefill_instance_endpoint["notify_port"],
            "remote_engine_id": None,
            "remote_block_ids": None,
-            "transfer_id": transfer_id,
+            "remote_host": ip,
+            "remote_port": port,
        }
        if TRANSFER_TYPE == "READ":
            # In read mode, prefill and decode are executed serially.
            prefill_response = await send_prefill_task
-            prefill_kv = prefill_response["kv_transfer_params"]
-            req_data["kv_transfer_params"]["remote_engine_id"] = prefill_kv[
-                "remote_engine_id"
-            ]
-            req_data["kv_transfer_params"]["remote_block_ids"] = prefill_kv[
-                "remote_block_ids"
-            ]
-            req_data["kv_transfer_params"]["transfer_id"] = prefill_kv["transfer_id"]
+            req_data["kv_transfer_params"]["remote_engine_id"] = prefill_response[
+                "kv_transfer_params"
+            ]["remote_engine_id"]
+            req_data["kv_transfer_params"]["remote_block_ids"] = prefill_response[
+                "kv_transfer_params"
+            ]["remote_block_ids"]

        req_data["kv_transfer_params"]["remote_dp_size"] = prefill_instance_endpoint[
            "dp_size"
@@ -316,6 +290,7 @@ async def handle_request(api: str, request: Request):

        if selected_prefill_dp_rank is not None:
            req_data["kv_transfer_params"]["remote_dp_rank"] = selected_prefill_dp_rank
+        req_data["kv_transfer_params"]["transfer_id"] = transfer_id

        decode_request_url = decode_instance_endpoint["request_address"] + api
        decode_request_task = asyncio.create_task(
@@ -1,17 +1,12 @@
 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 """
-Example online usage of token reward models.
+Example online usage of Pooling API.

 Run `vllm serve <model> --runner pooling`
 to start up the server in vLLM. e.g.

 vllm serve internlm/internlm2-1_8b-reward --trust-remote-code
-
-The key distinction between sequence classification and token classification
-lies in their output granularity: sequence classification produces a single
-result for an entire input sequence, whereas token classification yields a
-result for each individual token within the sequence.
 """

 import argparse
@@ -1,62 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-
-"""
-Example offline usage of sequence reward models.
-
-The key distinction between sequence classification and token classification
-lies in their output granularity: sequence classification produces a single
-result for an entire input sequence, whereas token classification yields a
-result for each individual token within the sequence.
-"""
-
-from argparse import Namespace
-
-from vllm import LLM, EngineArgs
-from vllm.utils.argparse_utils import FlexibleArgumentParser
-from vllm.utils.print_utils import print_embeddings
-
-
-def parse_args():
-    parser = FlexibleArgumentParser()
-    parser = EngineArgs.add_cli_args(parser)
-    # Set example specific arguments
-    parser.set_defaults(
-        model="Skywork/Skywork-Reward-V2-Qwen3-0.6B",
-        runner="pooling",
-        enforce_eager=True,
-        max_model_len=1024,
-        trust_remote_code=True,
-    )
-    return parser.parse_args()
-
-
-def main(args: Namespace):
-    # Sample prompts.
-    prompts = [
-        "Hello, my name is",
-        "The president of the United States is",
-        "The capital of France is",
-        "The future of AI is",
-    ]
-
-    # Create an LLM.
-    # You should pass runner="pooling" for reward models
-    llm = LLM(**vars(args))
-
-    # Generate rewards. The output is a list of PoolingRequestOutput.
-    # Use pooling_task="classify" for sequence reward models.
-    outputs = llm.encode(prompts, pooling_task="classify")
-
-    # Print the outputs.
-    print("\nGenerated Outputs:\n" + "-" * 60)
-    for prompt, output in zip(prompts, outputs):
-        rewards = output.outputs.data
-        print(f"Prompt: {prompt!r}")
-        print_embeddings(rewards.tolist(), prefix="Reward")
-        print("-" * 60)
-
-
-if __name__ == "__main__":
-    args = parse_args()
-    main(args)
@@ -1,71 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-"""
-Example online usage of sequence reward models.
-
-Run `vllm serve <model> --runner pooling`
-to start up the server in vLLM. e.g.
-
-vllm serve Skywork/Skywork-Reward-V2-Qwen3-0.6B
-
-The key distinction between sequence classification and token classification
-lies in their output granularity: sequence classification produces a single
-result for an entire input sequence, whereas token classification yields a
-result for each individual token within the sequence.
-"""
-
-import argparse
-import pprint
-
-import requests
-
-
-def post_http_request(prompt: dict, api_url: str) -> requests.Response:
-    headers = {"User-Agent": "Test Client"}
-    response = requests.post(api_url, headers=headers, json=prompt)
-    return response
-
-
-def parse_args():
-    parser = argparse.ArgumentParser()
-    parser.add_argument("--host", type=str, default="localhost")
-    parser.add_argument("--port", type=int, default=8000)
-
-    return parser.parse_args()
-
-
-def main(args):
-    base_url = f"http://{args.host}:{args.port}"
-    models_url = base_url + "/v1/models"
-    pooing_url = base_url + "/pooling"
-
-    response = requests.get(models_url)
-    model = response.json()["data"][0]["id"]
-
-    # Input like Completions API
-    prompt = {"model": model, "input": "vLLM is great!"}
-    pooling_response = post_http_request(prompt=prompt, api_url=pooing_url)
-    print("-" * 50)
-    print("Pooling Response:")
-    pprint.pprint(pooling_response.json())
-    print("-" * 50)
-
-    # Input like Chat API
-    prompt = {
-        "model": model,
-        "messages": [
-            {
-                "role": "user",
-                "content": [{"type": "text", "text": "vLLM is great!"}],
-            }
-        ],
-    }
-    pooling_response = post_http_request(prompt=prompt, api_url=pooing_url)
-    print("Pooling Response:")
-    pprint.pprint(pooling_response.json())
-    print("-" * 50)
-
-
-if __name__ == "__main__":
-    args = parse_args()
-    main(args)
@@ -4,68 +4,74 @@
 import torch

 from vllm import LLM
+from vllm.config import PoolerConfig
 from vllm.inputs import TextPrompt
 from vllm.multimodal.utils import fetch_image

-# Initialize model
-model = LLM(
-    model="jinaai/jina-embeddings-v4-vllm-text-matching",
-    runner="pooling",
-    max_model_len=1024,
-    gpu_memory_utilization=0.8,
-)

-# Create text prompts
-text1 = "Ein wunderschöner Sonnenuntergang am Strand"
-text1_prompt = TextPrompt(prompt=f"Query: {text1}")
+def main():
+    # Initialize model
+    model = LLM(
+        model="jinaai/jina-embeddings-v4-vllm-text-matching",
+        pooler_config=PoolerConfig(task="token_embed"),
+        runner="pooling",
+        max_model_len=1024,
+        gpu_memory_utilization=0.8,
+    )

-text2 = "浜辺に沈む美しい夕日"
-text2_prompt = TextPrompt(prompt=f"Query: {text2}")
+    # Create text prompts
+    text1 = "Ein wunderschöner Sonnenuntergang am Strand"
+    text1_prompt = TextPrompt(prompt=f"Query: {text1}")

-# Create image prompt
-image = fetch_image(
-    "https://vllm-public-assets.s3.us-west-2.amazonaws.com/multimodal_asset/eskimo.jpg"  # noqa: E501
-)
-image_prompt = TextPrompt(
-    prompt="<|im_start|>user\n<|vision_start|><|image_pad|><|vision_end|>Describe the image.<|im_end|>\n",  # noqa: E501
-    multi_modal_data={"image": image},
-)
+    text2 = "浜辺に沈む美しい夕日"
+    text2_prompt = TextPrompt(prompt=f"Query: {text2}")

-# Encode all prompts
-prompts = [text1_prompt, text2_prompt, image_prompt]
-outputs = model.encode(prompts, pooling_task="token_embed")
+    # Create image prompt
+    image = fetch_image(
+        "https://vllm-public-assets.s3.us-west-2.amazonaws.com/multimodal_asset/eskimo.jpg"  # noqa: E501
+    )
+    image_prompt = TextPrompt(
+        prompt="<|im_start|>user\n<|vision_start|><|image_pad|><|vision_end|>Describe the image.<|im_end|>\n",  # noqa: E501
+        multi_modal_data={"image": image},
+    )
+
+    # Encode all prompts
+    prompts = [text1_prompt, text2_prompt, image_prompt]
+    outputs = model.encode(prompts, pooling_task="token_embed")
+
+    def get_embeddings(outputs):
+        VISION_START_TOKEN_ID, VISION_END_TOKEN_ID = 151652, 151653
+
+        embeddings = []
+        for output in outputs:
+            if VISION_START_TOKEN_ID in output.prompt_token_ids:
+                # Gather only vision tokens
+                img_start_pos = torch.where(
+                    torch.tensor(output.prompt_token_ids) == VISION_START_TOKEN_ID
+                )[0][0]
+                img_end_pos = torch.where(
+                    torch.tensor(output.prompt_token_ids) == VISION_END_TOKEN_ID
+                )[0][0]
+                embeddings_tensor = output.outputs.data.detach().clone()[
+                    img_start_pos : img_end_pos + 1
+                ]
+            else:
+                # Use all tokens for text-only prompts
+                embeddings_tensor = output.outputs.data.detach().clone()
+
+            # Pool and normalize embeddings
+            pooled_output = (
+                embeddings_tensor.sum(dim=0, dtype=torch.float32)
+                / embeddings_tensor.shape[0]
+            )
+            embeddings.append(torch.nn.functional.normalize(pooled_output, dim=-1))
+        return embeddings
+
+    embeddings = get_embeddings(outputs)
+
+    for embedding in embeddings:
+        print(embedding.shape)


-def get_embeddings(outputs):
-    VISION_START_TOKEN_ID, VISION_END_TOKEN_ID = 151652, 151653
-
-    embeddings = []
-    for output in outputs:
-        if VISION_START_TOKEN_ID in output.prompt_token_ids:
-            # Gather only vision tokens
-            img_start_pos = torch.where(
-                torch.tensor(output.prompt_token_ids) == VISION_START_TOKEN_ID
-            )[0][0]
-            img_end_pos = torch.where(
-                torch.tensor(output.prompt_token_ids) == VISION_END_TOKEN_ID
-            )[0][0]
-            embeddings_tensor = output.outputs.data.detach().clone()[
-                img_start_pos : img_end_pos + 1
-            ]
-        else:
-            # Use all tokens for text-only prompts
-            embeddings_tensor = output.outputs.data.detach().clone()
-
-        # Pool and normalize embeddings
-        pooled_output = (
-            embeddings_tensor.sum(dim=0, dtype=torch.float32)
-            / embeddings_tensor.shape[0]
-        )
-        embeddings.append(torch.nn.functional.normalize(pooled_output, dim=-1))
-    return embeddings
-
-
-embeddings = get_embeddings(outputs)
-
-for embedding in embeddings:
-    print(embedding.shape)
+if __name__ == "__main__":
+    main()
@@ -4,6 +4,7 @@
 from argparse import Namespace

 from vllm import LLM, EngineArgs
+from vllm.config import PoolerConfig
 from vllm.utils.argparse_utils import FlexibleArgumentParser


@@ -13,6 +14,7 @@ def parse_args():
    # Set example specific arguments
    parser.set_defaults(
        model="BAAI/bge-m3",
+        pooler_config=PoolerConfig(task="token_embed"),
        runner="pooling",
        enforce_eager=True,
    )
@@ -32,15 +34,6 @@ def main(args: Namespace):
    # You should pass runner="pooling" for embedding models
    llm = LLM(**vars(args))

-    # Generate embedding. The output is a list of EmbeddingRequestOutputs.
-    outputs = llm.embed(prompts)
-
-    # Print the outputs.
-    print("\nGenerated Outputs:\n" + "-" * 60)
-    for prompt, output in zip(prompts, outputs):
-        embeds = output.outputs.embedding
-        print(len(embeds))
-
    # Generate embedding for each token. The output is a list of PoolingRequestOutput.
    outputs = llm.encode(prompts, pooling_task="token_embed")

@@ -50,6 +43,20 @@ def main(args: Namespace):
        multi_vector = output.outputs.data
        print(multi_vector.shape)

+    query = "What is the capital of France?"
+    documents = [
+        "The capital of Brazil is Brasilia.",
+        "The capital of France is Paris.",
+    ]
+    # Generate scores.
+    outputs = llm.score(query, documents)
+    # Print the outputs.
+    print("\nGenerated Outputs:\n" + "-" * 60)
+    for document, output in zip(documents, outputs):
+        score = output.outputs.score
+        print(f"Pair: {[query, document]!r} \nScore: {score}")
+        print("-" * 60)
+

 if __name__ == "__main__":
    args = parse_args()
@@ -7,10 +7,11 @@ Example online usage of Pooling API for multi vector retrieval.
 Run `vllm serve <model> --runner pooling`
 to start up the server in vLLM. e.g.

-vllm serve BAAI/bge-m3
+vllm serve BAAI/bge-m3 --pooler-config.task token_embed
 """

 import argparse
+import pprint

 import requests
 import torch
@@ -32,7 +33,8 @@ def parse_args():


 def main(args):
-    api_url = f"http://{args.host}:{args.port}/pooling"
+    pooling_url = f"http://{args.host}:{args.port}/pooling"
+    score_url = f"http://{args.host}:{args.port}/score"
    model_name = args.model

    prompts = [
@@ -43,11 +45,23 @@ def main(args):
    ]
    prompt = {"model": model_name, "input": prompts}

-    pooling_response = post_http_request(prompt=prompt, api_url=api_url)
+    pooling_response = post_http_request(prompt=prompt, api_url=pooling_url)
    for output in pooling_response.json()["data"]:
        multi_vector = torch.tensor(output["data"])
        print(multi_vector.shape)

+    queries = "What is the capital of France?"
+    documents = [
+        "The capital of Brazil is Brasilia.",
+        "The capital of France is Paris.",
+    ]
+    prompt = {"model": model_name, "queries": queries, "documents": documents}
+    score_response = post_http_request(prompt=prompt, api_url=score_url)
+    print("\nPrompt when queries is string and documents is a list:")
+    pprint.pprint(prompt)
+    print("\nScore Response:")
+    pprint.pprint(score_response.json())
+

 if __name__ == "__main__":
    args = parse_args()
@@ -131,9 +131,16 @@ class TrainModel:
        from vllm.model_executor.layers.batch_invariant import (
            init_batch_invariance,
        )
+        from vllm.platforms import current_platform
+        from vllm.v1.attention.backends.registry import AttentionBackendEnum

        # need to init all env vars for batch invariance which affect nccl ops
-        init_batch_invariance()
+        attn_backend = (
+            AttentionBackendEnum.TRITON_ATTN
+            if current_platform.is_rocm()
+            else AttentionBackendEnum.FLASH_ATTN
+        )
+        init_batch_invariance(attn_backend)

        self.model = AutoModelForCausalLM.from_pretrained(
            model_name, dtype=torch.bfloat16
@@ -24,14 +24,13 @@ classifiers = [
    "Programming Language :: Python :: 3.11",
    "Programming Language :: Python :: 3.12",
    "Programming Language :: Python :: 3.13",
-    "Programming Language :: Python :: 3.14",
    "Intended Audience :: Developers",
    "Intended Audience :: Information Technology",
    "Intended Audience :: Science/Research",
    "Topic :: Scientific/Engineering :: Artificial Intelligence",
    "Topic :: Scientific/Engineering :: Information Analysis",
 ]
-requires-python = ">=3.10,<3.15"
+requires-python = ">=3.10,<3.14"
 dynamic = [ "version", "dependencies", "optional-dependencies"]

 [project.urls]
@@ -124,8 +123,7 @@ extend-exclude = ["tests/models/fixtures/*", "tests/prompts/*", "tests/tokenizer
    "benchmarks/sonnet.txt", "tests/lora/data/*", "build/*",
    "examples/pooling/token_embed/*", "tests/models/language/pooling/*",
    "vllm/third_party/*", "vllm/entrypoints/serve/instrumentator/static/*", "tests/entrypoints/openai/speech_to_text/test_transcription_validation.py",
-    "docs/governance/process.md", "docs/assets/contributing/vllm_bench_serve_timeline.html", 
-    "tests/v1/engine/test_fast_incdec_prefix_err.py", ".git/*"]
+    "docs/governance/process.md", "tests/v1/engine/test_fast_incdec_prefix_err.py", ".git/*"]
 ignore-hidden = false

 [tool.typos.default]
@@ -20,7 +20,7 @@ prometheus-fastapi-instrumentator >= 7.0.0
 tiktoken >= 0.6.0  # Required for DBRX tokenizer
 lm-format-enforcer == 0.11.3
 llguidance >= 1.3.0, < 1.4.0; platform_machine == "x86_64" or platform_machine == "arm64" or platform_machine == "aarch64" or platform_machine == "ppc64le"
-outlines_core == 0.2.14
+outlines_core == 0.2.11
 # required for outlines backend disk cache
 diskcache == 5.6.3
 lark == 1.2.2
@@ -4,7 +4,7 @@

 setuptools==77.0.3 # this version can reuse CMake build dir

-numba == 0.65.0; platform_machine != "s390x" # Required for N-gram speculative decoding
+numba == 0.61.2; platform_machine != "s390x" # Required for N-gram speculative decoding

 # Dependencies for CPUs
 torch==2.11.0+cpu; platform_machine == "x86_64" or platform_machine == "s390x" or platform_machine == "aarch64"
@@ -1,7 +1,7 @@
 # Common dependencies
 -r common.txt

-numba == 0.65.0 # Required for N-gram speculative decoding
+numba == 0.61.2 # Required for N-gram speculative decoding

 # Dependencies for NVIDIA GPUs
 torch==2.11.0
@@ -11,8 +11,6 @@ torchvision==0.26.0 # Required for phi3v processor. See https://github.com/pytor
 # FlashInfer should be updated together with the Dockerfile
 flashinfer-python==0.6.8.post1
 flashinfer-cubin==0.6.8.post1
-apache-tvm-ffi==0.1.9
-tilelang==0.1.9
 # Cap nvidia-cudnn-frontend (transitive dep of flashinfer) due to
 # breaking changes in 1.19.0
 nvidia-cudnn-frontend>=1.13.0,<1.19.0
@@ -5,7 +5,7 @@
 grpcio==1.78.0
 grpcio-reflection==1.78.0

-numba == 0.65.0 # Required for N-gram speculative decoding
+numba == 0.61.2 # Required for N-gram speculative decoding

 # Dependencies for AMD GPUs
 datasets
@@ -20,4 +20,4 @@ conch-triton-kernels==1.2.1
 timm>=1.0.17
 # amd-quark: required for Quark quantization on ROCm 
 # To be consistent with test_quark.py
-amd-quark>=0.8.99
+amd-quark>=0.8.99
@@ -54,7 +54,7 @@ grpcio==1.78.0
 grpcio-reflection==1.78.0

 arctic-inference == 0.1.1 # Required for suffix decoding test
-numba == 0.65.0 # Required for N-gram speculative decoding
+numba == 0.61.2 # Required for N-gram speculative decoding
 numpy
 runai-model-streamer[s3,gcs,azure]==0.15.7
 fastsafetensors>=0.2.2 # 0.2.2 contains important fixes for multi-GPU mem usage
@@ -479,7 +479,7 @@ lightning-utilities==0.14.3
    #   lightning
    #   pytorch-lightning
    #   torchmetrics
-llvmlite==0.47.0
+llvmlite==0.44.0
    # via numba
 lm-eval==0.4.11
    # via -r requirements/test/cuda.in
@@ -550,7 +550,7 @@ nltk==3.9.1
    # via rouge-score
 num2words==0.5.14
    # via -r requirements/test/cuda.in
-numba==0.65.0
+numba==0.61.2
    # via
    #   -c requirements/cuda.txt
    #   -r requirements/test/cuda.in
@@ -40,7 +40,7 @@ buildkite-test-collector==0.1.9
 genai_perf>=0.0.8
 tritonclient>=2.51.0

-numba == 0.65.0 # Required for N-gram speculative decoding
+numba == 0.61.2 # Required for N-gram speculative decoding
 numpy
 runai-model-streamer[s3,gcs,azure]==0.15.7
 fastsafetensors>=0.2.2
@@ -52,7 +52,7 @@ grpcio==1.78.0
 grpcio-reflection==1.78.0

 arctic-inference==0.1.1 # Required for suffix decoding test
-numba==0.65.0 # Required for N-gram speculative decoding
+numba==0.61.2 # Required for N-gram speculative decoding
 numpy
 runai-model-streamer[s3,gcs,azure]==0.15.7
 fastsafetensors @ git+https://github.com/foundation-model-stack/fastsafetensors.git@0.2.2 # PyPI only ships CUDA wheels
@@ -559,7 +559,7 @@ llguidance==1.3.0
    # via
    #   -c requirements/common.txt
    #   -r requirements/test/../common.txt
-llvmlite==0.47.0
+llvmlite==0.44.0
    # via numba
 lm-eval==0.4.11
    # via -r requirements/test/rocm.in
@@ -655,7 +655,7 @@ nltk==3.9.3
    # via rouge-score
 num2words==0.5.14
    # via -r requirements/test/rocm.in
-numba==0.65.0
+numba==0.61.2
    # via
    #   -c requirements/rocm.txt
    #   -r requirements/test/rocm.in
@@ -811,7 +811,7 @@ orjson==3.11.7
    # via
    #   genai-perf
    #   kaleido
-outlines-core==0.2.14
+outlines-core==0.2.11
    # via
    #   -c requirements/common.txt
    #   -r requirements/test/../common.txt
@@ -93,7 +93,7 @@ docker==7.1.0
    # via gpt-oss
 docopt==0.6.2
    # via num2words
-dpcpp-cpp-rt==2025.3.2
+dpcpp-cpp-rt==2025.3.1
    # via
    #   onemkl-sycl-blas
    #   onemkl-sycl-dft
@@ -172,27 +172,27 @@ idna==3.11
    #   yarl
 imageio==2.37.3
    # via scikit-image
-impi-rt==2021.17.2
+impi-rt==2021.17.0
    # via
    #   oneccl
    #   torch
 iniconfig==2.3.0
    # via pytest
-intel-cmplr-lib-rt==2025.3.2
+intel-cmplr-lib-rt==2025.3.1
    # via
    #   intel-sycl-rt
    #   torch
-intel-cmplr-lib-ur==2025.3.2
+intel-cmplr-lib-ur==2025.3.1
    # via
    #   intel-openmp
    #   intel-sycl-rt
    #   torch
-intel-cmplr-lic-rt==2025.3.2
+intel-cmplr-lic-rt==2025.3.1
    # via
    #   intel-opencl-rt
    #   intel-sycl-rt
    #   torch
-intel-opencl-rt==2025.3.2
+intel-opencl-rt==2025.3.1
    # via
    #   dpcpp-cpp-rt
    #   onemkl-sycl-blas
@@ -201,14 +201,14 @@ intel-opencl-rt==2025.3.2
    #   onemkl-sycl-rng
    #   onemkl-sycl-sparse
    #   torch
-intel-openmp==2025.3.2
+intel-openmp==2025.3.1
    # via
    #   dpcpp-cpp-rt
    #   mkl
    #   torch
-intel-pti==0.16.0
+intel-pti==0.15.0
    # via torch
-intel-sycl-rt==2025.3.2
+intel-sycl-rt==2025.3.1
    # via
    #   dpcpp-cpp-rt
    #   oneccl
@@ -244,7 +244,7 @@ lazy-loader==0.5
    #   scikit-image
 librosa==0.10.2.post1
    # via -r requirements/test/xpu.in
-llvmlite==0.47.0
+llvmlite==0.44.0
    # via numba
 lm-eval==0.4.11
    # via -r requirements/test/xpu.in
@@ -270,7 +270,7 @@ mistral-common==1.11.0
    # via
    #   -c requirements/common.txt
    #   -r requirements/test/xpu.in
-mkl==2025.3.1
+mkl==2025.3.0
    # via
    #   onemkl-sycl-blas
    #   onemkl-sycl-dft
@@ -304,7 +304,7 @@ nltk==3.9.4
    # via rouge-score
 num2words==0.5.14
    # via -r requirements/test/xpu.in
-numba==0.65.0
+numba==0.61.2
    # via
    #   -c requirements/xpu.txt
    #   librosa
@@ -335,28 +335,28 @@ numpy==2.2.6
    #   tifffile
    #   torchvision
    #   transformers
-oneccl==2021.17.2
+oneccl==2021.17.1
    # via
    #   oneccl-devel
    #   torch
-oneccl-devel==2021.17.2
+oneccl-devel==2021.17.1
    # via torch
-onemkl-license==2025.3.1
+onemkl-license==2025.3.0
    # via
    #   mkl
    #   torch
-onemkl-sycl-blas==2025.3.1
+onemkl-sycl-blas==2025.3.0
    # via
    #   onemkl-sycl-lapack
    #   onemkl-sycl-sparse
    #   torch
-onemkl-sycl-dft==2025.3.1
+onemkl-sycl-dft==2025.3.0
    # via torch
-onemkl-sycl-lapack==2025.3.1
+onemkl-sycl-lapack==2025.3.0
    # via torch
-onemkl-sycl-rng==2025.3.1
+onemkl-sycl-rng==2025.3.0
    # via torch
-onemkl-sycl-sparse==2025.3.1
+onemkl-sycl-sparse==2025.3.0
    # via torch
 openai-harmony==0.0.8
    # via
@@ -608,7 +608,7 @@ tabledata==1.3.4
    # via pytablewriter
 tabulate==0.10.0
    # via sacrebleu
-tbb==2022.3.1
+tbb==2022.3.0
    # via
    #   intel-opencl-rt
    #   mkl
@@ -645,7 +645,7 @@ tokenizers==0.22.2
    # via
    #   -c requirements/common.txt
    #   transformers
-torch==2.11.0+xpu
+torch==2.10.0+xpu
    # via
    #   -c requirements/xpu.txt
    #   accelerate
@@ -653,7 +653,7 @@ torch==2.11.0+xpu
    #   sentence-transformers
    #   timm
    #   torchvision
-torchvision==0.26.0+xpu
+torchvision==0.25.0+xpu
    # via timm
 tqdm==4.67.3
    # via
@@ -671,7 +671,7 @@ transformers==5.5.3
    # via
    #   -c requirements/common.txt
    #   sentence-transformers
-triton-xpu==3.7.0
+triton-xpu==3.6.0
    # via torch
 typepy==1.3.4
    # via
@@ -710,7 +710,7 @@ typing-inspection==0.4.2
    # via
    #   fastapi
    #   pydantic
-umf==1.0.3
+umf==1.0.2
    # via
    #   intel-cmplr-lib-ur
    #   torch
@@ -9,9 +9,9 @@ setuptools>=77.0.3,<81.0.0
 wheel
 jinja2>=3.1.6
 datasets # for benchmark scripts
-numba == 0.65.0 # Required for N-gram speculative decoding
+numba == 0.61.2 # Required for N-gram speculative decoding
 --extra-index-url=https://download.pytorch.org/whl/xpu
-torch==2.11.0+xpu
+torch==2.10.0+xpu
 torchaudio
 torchvision

@@ -927,9 +927,7 @@ def get_vllm_version() -> str:
    elif _is_tpu():
        version += f"{sep}tpu"
    elif _is_cpu():
-        # Check the local VLLM_TARGET_DEVICE (may be set by auto-detect above),
-        # not envs.VLLM_TARGET_DEVICE, so CPU-only hosts still get `+cpu`.
-        if VLLM_TARGET_DEVICE == "cpu":
+        if envs.VLLM_TARGET_DEVICE == "cpu":
            version += f"{sep}cpu"
    elif _is_xpu():
        version += f"{sep}xpu"
@@ -1109,7 +1107,7 @@ setup(
        #   - .buildkite/test-amd.yaml
        "helion": ["helion==1.0.0"],
        # Optional deps for gRPC server (vllm serve --grpc)
-        "grpc": ["smg-grpc-servicer[vllm] >= 0.5.2"],
+        "grpc": ["smg-grpc-servicer[vllm] >= 0.5.0"],
        # Optional deps for OpenTelemetry tracing
        "otel": [
            "opentelemetry-sdk>=1.26.0",
@@ -261,8 +261,6 @@ def _compare_sp(
        },
        "use_inductor_graph_partition": use_inductor_graph_partition,
    }
-    if not use_inductor_graph_partition:
-        compilation_config["splitting_ops"] = []

    tp_sp_args = [
        *common_args,
@@ -116,11 +116,6 @@ def run_e2e_fusion_test(monkeypatch, caplog_mp_spawn):
        model_kwargs["attention_config"] = {"backend": attn_backend.backend.name}
        model_kwargs["tensor_parallel_size"] = tp_size

-        # Cap warmup memory: tests use small max_model_len (1024) but the
-        # engine default max_num_batched_tokens is 16384. Warming up large
-        # models (e.g. Llama-4-Scout-FP8) at 16384 tokens may trigger OOM.
-        model_kwargs.setdefault("max_num_batched_tokens", 8192)
-
        # Sparse MLA models (DSv3.2) hit an over-strict inductor assertion in
        # decompose_auto_functionalized when +rotary_embedding is forced into
        # the compile graph. Disable qk_norm+rope fusion (which auto-enables
@@ -194,7 +189,7 @@ def run_e2e_fusion_test(monkeypatch, caplog_mp_spawn):
            # TODO: Remove log counting in unit tests
            # once all matchers implement VllmFusionPatternMatcherPass
            n_expected = tp_size * num_ranges_activated
-            if match_name not in ("attn_quant_fusion", "act_quant_fusion"):
+            if match_name != "attn_quant_fusion":
                assert len(log_matches) == n_expected, (
                    f"Could not find {n_expected} {match_name} "
                    f"(found {len(log_matches)}) in:\n {log_holder.text}"
@@ -255,12 +250,6 @@ def run_e2e_fusion_test(monkeypatch, caplog_mp_spawn):
                    f"entries (SP took precedence), found: {log_matches}"
                )

-            elif match_name == "act_quant_fusion":
-                actual_match = match_table.get("activation_quant_fusion_pass", 0)
-                assert actual_match == expected_matches * n_expected, (
-                    f"Could not find {expected_matches * n_expected} "
-                    f"{match_name} (found {actual_match})."
-                )
            elif match_name == "attn_quant_fusion":
                actual_match = match_table.get(
                    "attn_quant_fusion", 0
@@ -34,10 +34,7 @@ def _run_vllm(vllm_runner):
            mode=CompilationMode.VLLM_COMPILE,
            cudagraph_mode=CUDAGraphMode.NONE,
        ),
-        # Phi-tiny-MoE uses SWA, whose admission cap is `cdiv(L, block_size) + 1`
-        # at default block_size=16 — i.e. 17 blocks for max_model_len=256. Use
-        # 32 for headroom.
-        num_gpu_blocks_override=32,
+        num_gpu_blocks_override=8,
    ):
        pass

@@ -59,7 +56,6 @@ def _cold_start(vllm_runner):
 def test_moe_startup(monkeypatch, vllm_runner, fresh_vllm_cache, mega_aot_artifact):
    monkeypatch.setenv("VLLM_ENABLE_V1_MULTIPROCESSING", "0")
    monkeypatch.setenv("VLLM_USE_MEGA_AOT_ARTIFACT", mega_aot_artifact)
-    monkeypatch.setenv("VLLM_DEEP_GEMM_WARMUP", "skip")

    # Cold start in a forked child (must fork before CUDA init).
    # This model has 32 identical transformer layers which produce
@@ -139,9 +135,10 @@ MODEL_SPECS = [
            model="deepseek-ai/DeepSeek-V3.2",
            hf_overrides=_SMALL_MOE_OVERRIDES,
            cold_artifacts_saved=4,
-            # https://github.com/vllm-project/vllm/issues/38051
-            warm_artifacts_saved=0 if is_torch_equal_or_newer("2.12.0") else 4,
-            warm_artifacts_loaded=4 if is_torch_equal_or_newer("2.12.0") else 0,
+            # TODO: https://github.com/vllm-project/vllm/issues/38051
+            # We shouldn't be saving any artifacts on warm start.
+            warm_artifacts_saved=4,
+            warm_artifacts_loaded=0,
        ),
        id="deepseek_v3.2",
    ),
@@ -150,9 +147,10 @@ MODEL_SPECS = [
            model="moonshotai/Kimi-K2.5",
            hf_overrides={"text_config": _SMALL_MOE_OVERRIDES},
            cold_artifacts_saved=4,
-            # https://github.com/vllm-project/vllm/issues/38051
-            warm_artifacts_saved=0 if is_torch_equal_or_newer("2.12.0") else 4,
-            warm_artifacts_loaded=4 if is_torch_equal_or_newer("2.12.0") else 0,
+            # TODO: https://github.com/vllm-project/vllm/issues/38051
+            # We shouldn't be saving any artifacts on warm start.
+            warm_artifacts_saved=4,
+            warm_artifacts_loaded=0,
        ),
        id="kimi_k2.5",
    ),
@@ -193,7 +191,7 @@ def _run_model(vllm_runner, spec: ModelStartupSpec):
            cudagraph_mode=CUDAGraphMode.NONE,
            pass_config=PassConfig(fuse_allreduce_rms=False),
        ),
-        num_gpu_blocks_override=16,
+        num_gpu_blocks_override=8,
    ):
        pass

@@ -239,7 +237,6 @@ def _cold_start_model(vllm_runner, spec: ModelStartupSpec):
@fork_new_process_for_each_test
 def test_model_startup(monkeypatch, vllm_runner, fresh_vllm_cache, spec):
    monkeypatch.setenv("VLLM_ENABLE_V1_MULTIPROCESSING", "0")
-    monkeypatch.setenv("VLLM_DEEP_GEMM_WARMUP", "skip")

    # Cold start in a forked child (must fork before CUDA init).
    ctx = mp.get_context("fork")
@@ -19,7 +19,6 @@ from vllm.config import (
    VllmConfig,
    set_current_vllm_config,
 )
-from vllm.config.utils import Range
 from vllm.distributed import (
    tensor_model_parallel_all_gather,
    tensor_model_parallel_reduce_scatter,
@@ -289,22 +288,6 @@ def test_async_tp_pass_replace(
    run_torch_spawn(async_tp_pass_on_test_model, num_processes)


-def test_async_tp_pass_requires_full_graph_compilation():
-    vllm_config = VllmConfig()
-    vllm_config.compilation_config.use_inductor_graph_partition = False
-    vllm_config.compilation_config.splitting_ops = [
-        "vllm::unified_attention_with_output"
-    ]
-
-    async_tp_pass = object.__new__(AsyncTPPass)
-    async_tp_pass.compilation_config = vllm_config.compilation_config
-
-    with pytest.raises(
-        AssertionError, match="AsyncTPPass requires full-graph compilation"
-    ):
-        async_tp_pass.is_applicable_for_range(Range(start=8, end=8))
-
-
 def async_tp_pass_on_test_model(
    local_rank: int,
    world_size: int,
@@ -22,7 +22,6 @@ from vllm.config import (
    get_current_vllm_config,
    set_current_vllm_config,
 )
-from vllm.config.utils import Range
 from vllm.distributed import tensor_model_parallel_all_reduce
 from vllm.distributed.parallel_state import (
    init_distributed_environment,
@@ -217,24 +216,6 @@ def test_sequence_parallelism_pass(
    run_torch_spawn(sequence_parallelism_pass_on_test_model, num_processes)


-def test_sequence_parallelism_pass_requires_full_graph_compilation():
-    vllm_config = VllmConfig()
-    vllm_config.compilation_config.use_inductor_graph_partition = False
-    vllm_config.compilation_config.splitting_ops = [
-        "vllm::unified_attention_with_output"
-    ]
-
-    sequence_parallelism_pass = object.__new__(SequenceParallelismPass)
-    sequence_parallelism_pass.compilation_config = vllm_config.compilation_config
-    sequence_parallelism_pass.min_token_num = 1
-
-    with pytest.raises(
-        AssertionError,
-        match="SequenceParallelismPass requires full-graph compilation",
-    ):
-        sequence_parallelism_pass.is_applicable_for_range(Range(start=8, end=8))
-
-
 def sequence_parallelism_pass_on_test_model(
    local_rank: int,
    world_size: int,
@@ -117,9 +117,9 @@ class TestFusedAddRMSNorm(torch.nn.Module):
        else:
            return norm_output, residual_output

-    def example_inputs(self, batch_size=8, seq_len=16):
-        hidden_states = torch.randn((batch_size * seq_len, self.hidden_size))
-        residual = torch.randn((batch_size * seq_len, self.intermediate_size))
+    def example_inputs(self, batch_size=8, hidden_size=16, seq_len=16):
+        hidden_states = torch.randn((batch_size * seq_len, hidden_size))
+        residual = torch.randn((batch_size * seq_len, hidden_size))
        return (hidden_states, residual)

    def ops_in_model(self, do_fusion):
@@ -51,7 +51,6 @@ from vllm.model_executor.layers.quantization.utils.w8a8_utils import (
 )
 from vllm.platforms import current_platform
 from vllm.utils.deep_gemm import (
-    is_deep_gemm_e8m0_used,
    is_deep_gemm_supported,
 )

@@ -318,26 +317,6 @@ def test_fusion_rmsnorm_quant(
    ):
        pytest.skip("Unsupported group shape 64 for CUTLASS/DeepGemm")

-    # TODO(quant-rms-fusion): DeepGEMM UE8M0 activation quant on B200 lowers
-    # to a packed int32-scale op (per_token_group_quant_fp8_packed_for_deepgemm),
-    # but the rms+quant fusion pattern only matches the fp32-scale variant, so
-    # the fused output gets a mismatched scale layout and produces NaN. Only
-    # reproduces on bf16 (DeepGEMM UE8M0 on B200 is bf16-only).
-    # To re-enable: make rms_norm_per_block_quant emit packed UE8M0 scales
-    # and extend the fusion pattern to rewrite the packed activation quant.
-    deepgemm_kernels = (
-        DeepGemmFp8BlockScaledMMKernel,
-        FlashInferFp8DeepGEMMDynamicBlockScaledKernel,
-    )
-    if (
-        dtype == torch.bfloat16
-        and force_kernel in deepgemm_kernels
-        and is_deep_gemm_e8m0_used()
-    ):
-        pytest.skip(
-            "rms+quant fusion does not yet match the packed UE8M0 DeepGEMM path"
-        )
-
    custom_ops = []
    if enable_rms_norm_custom_op:
        custom_ops.append("+rms_norm")
@@ -10,7 +10,7 @@ import vllm.envs as envs
 from tests.compile.backend import TestBackend
 from tests.kernels.quantization.nvfp4_utils import quant_nvfp4_tensor
 from tests.utils import TestFP8Layer
-from vllm._aiter_ops import IS_AITER_FOUND, rocm_aiter_ops
+from vllm._aiter_ops import IS_AITER_FOUND
 from vllm._custom_ops import cutlass_scaled_fp4_mm, scaled_fp4_quant
 from vllm.compilation.passes.fusion.act_quant_fusion import (
    FUSED_OPS,
@@ -157,27 +157,23 @@ class TestSiluMulGroupFp8QuantModel(torch.nn.Module):
            activation_quant_key=self.act_quant_key,
            input_dtype=dtype,
        )
+        self.w = torch.rand(hidden_size, hidden_size).to(dtype=FP8_DTYPE).t()

-        if not current_platform.is_fp8_fnuz():
-            kernel = self.w8a8_block_fp8_linear.kernel
-            orig_quant = kernel.quant_fp8
-            kernel.quant_fp8 = lambda *a, use_triton=False, **kw: orig_quant(
-                *a, use_triton=True, **kw
-            )
+        scale_hidden_size = (hidden_size + 128 - 1) // 128
+        self.wscale = torch.rand(
+            (scale_hidden_size, scale_hidden_size), dtype=torch.float32
+        )

        self.enable_silu_mul_custom_op = self.silu_and_mul.enabled()

    def forward(self, x):
        y = self.silu_and_mul(x)
-        x2 = self.w8a8_block_fp8_linear(y)
+        x2 = self.w8a8_block_fp8_linear(y, self.w, self.wscale)
        return x2

    def ops_in_model_before(self):
        return [
            SILU_MUL_OP if self.enable_silu_mul_custom_op else torch.ops.aten.mul,
-            rocm_aiter_ops.get_group_quant_op()
-            if current_platform.is_fp8_fnuz()
-            else torch.ops.vllm.triton_per_token_group_quant_fp8.default,
        ]

    def ops_in_model_after(self):
@@ -328,6 +324,7 @@ def test_fusion_silu_and_mul_quant(
    with set_current_vllm_config(config), monkeypatch.context() as m:
        fusion_passes = [ActivationQuantFusionPass(config)]
        if IS_AITER_FOUND and model_class is TestSiluMulGroupFp8QuantModel:
+            from vllm._aiter_ops import rocm_aiter_ops
            from vllm.compilation.passes.fusion.rocm_aiter_fusion import (
                RocmAiterSiluMulFp8GroupQuantFusionPass,
            )
@@ -355,16 +352,10 @@ def test_fusion_silu_and_mul_quant(
            atol, rtol = 1e-3, 1e-3
        elif isinstance(model, TestSiluMulNvfp4QuantModel):
            atol, rtol = 1e-1, 1e-1
-        elif isinstance(model, TestSiluMulGroupFp8QuantModel):
+        elif isinstance(
+            model, (TestSiluMulGroupFp8QuantModel, TestSiluMulBlockQuantModel)
+        ):
            atol, rtol = 5e-2, 5e-2
-        elif isinstance(model, TestSiluMulBlockQuantModel):
-            if current_platform.is_rocm():
-                atol, rtol = 1e-3, 1e-3
-            else:
-                # CUDA fused kernel computes silu*mul in fp32 while the reference
-                # goes through bf16/fp16 storage, so group maxima (and thus scales)
-                # can shift by one FP8-e4m3 code (~1/8 relative step).
-                atol, rtol = 5e-2, 5e-2

        torch.testing.assert_close(
            result[0].to(dtype=dtype), result2[0].to(dtype=dtype), atol=atol, rtol=rtol
@@ -205,22 +205,6 @@ def test_enforce_eager(vllm_runner, monkeypatch):
        pass


-@pytest.mark.forked
-def test_torch_compile_disable(vllm_runner, monkeypatch):
-    monkeypatch.setenv("VLLM_ENABLE_V1_MULTIPROCESSING", "0")
-    monkeypatch.setenv("TORCH_COMPILE_DISABLE", "1")
-    monkeypatch.setenv("VLLM_DISABLE_COMPILE_CACHE", "1")
-
-    with (
-        compilation_counter.expect(num_graphs_seen=0, stock_torch_compile_count=0),
-        vllm_runner(
-            "facebook/opt-125m",
-            gpu_memory_utilization=0.4,
-        ) as _,
-    ):
-        pass
-
-
 def test_splitting_ops_dynamic():
    # Default config
    config = VllmConfig()
@@ -405,12 +389,9 @@ def test_should_split():
        (None, 0, 1, False, 2048, CUDAGraphMode.NONE, 0),
        # truncated to nearest multiple of 8 or 16
        (None, 257, 1, False, 2048, CUDAGraphMode.FULL_AND_PIECEWISE, 256),
-        # max_num_batched_tokens <= max_cudagraph_capture_size should always be
-        # captured even if not landing on a 16-stride step
-        (None, 2048, 1, False, 257, CUDAGraphMode.FULL_AND_PIECEWISE, 257),
        # max from list
        ([1, 2, 4, 15], None, 1, False, 2048, CUDAGraphMode.FULL_AND_PIECEWISE, 15),
-        # SP forces full-graph compilation, sizes are filtered by TP
+        # filtered out 15 due to SP
        ([1, 2, 4, 15], None, 2, True, 2048, CUDAGraphMode.FULL_AND_PIECEWISE, 4),
        # limited by the max_tokens
        ([1, 2, 4, 15], None, 1, False, 8, CUDAGraphMode.FULL_AND_PIECEWISE, 4),
@@ -468,123 +449,6 @@ def test_cudagraph_sizes_post_init(
        )


-@pytest.mark.skipif(
-    not current_platform.support_static_graph_mode(),
-    reason="Skip if not cudagraph mode supported",
-)
-@pytest.mark.parametrize(
-    (
-        "cudagraph_mode",
-        "use_inductor_graph_partition",
-        "expected_enable_sp",
-        "expected_cudagraph_mode",
-        "expected_piecewise_compile",
-        "expected_capture_sizes",
-        "expected_max_size",
-    ),
-    [
-        (CUDAGraphMode.PIECEWISE, False, True, CUDAGraphMode.FULL, False, [2, 4], 4),
-        (
-            CUDAGraphMode.FULL_DECODE_ONLY,
-            False,
-            True,
-            CUDAGraphMode.FULL_DECODE_ONLY,
-            False,
-            [2, 4],
-            4,
-        ),
-        (
-            CUDAGraphMode.FULL_AND_PIECEWISE,
-            False,
-            True,
-            CUDAGraphMode.FULL,
-            False,
-            [2, 4],
-            4,
-        ),
-        (
-            CUDAGraphMode.FULL_AND_PIECEWISE,
-            True,
-            True,
-            CUDAGraphMode.FULL_AND_PIECEWISE,
-            True,
-            [2, 4],
-            4,
-        ),
-    ],
-)
-def test_sequence_parallelism_requires_full_graph_compilation(
-    cudagraph_mode: CUDAGraphMode,
-    use_inductor_graph_partition: bool,
-    expected_enable_sp: bool,
-    expected_cudagraph_mode: CUDAGraphMode,
-    expected_piecewise_compile: bool,
-    expected_capture_sizes: list[int],
-    expected_max_size: int,
-):
-    with patch.object(current_platform, "device_count", return_value=2):
-        vllm_config = VllmConfig(
-            parallel_config=ParallelConfig(tensor_parallel_size=2),
-            scheduler_config=SchedulerConfig(
-                max_num_seqs=128,
-                max_num_batched_tokens=2048,
-                max_model_len=2048,
-                is_encoder_decoder=False,
-            ),
-        )
-        vllm_config.model_config = MagicMock(
-            dtype=torch.float16,
-            enforce_eager=False,
-            is_moe=False,
-            disable_cascade_attn=False,
-            get_hidden_size=MagicMock(return_value=4096),
-        )
-        vllm_config.compilation_config = CompilationConfig(
-            mode=CompilationMode.VLLM_COMPILE,
-            cudagraph_capture_sizes=[1, 2, 4, 15],
-            max_cudagraph_capture_size=None,
-            compile_sizes=["cudagraph_capture_sizes"],
-            use_inductor_graph_partition=use_inductor_graph_partition,
-            pass_config=PassConfig(
-                enable_sp=True,
-                fuse_gemm_comms=True,
-                fuse_norm_quant=True,
-                fuse_act_quant=True,
-                eliminate_noops=True,
-                sp_min_token_num=512,
-            ),
-            cudagraph_mode=cudagraph_mode,
-        )
-        vllm_config.compilation_config.set_splitting_ops_for_v1(
-            all2all_backend=vllm_config.parallel_config.all2all_backend,
-            data_parallel_size=1,
-        )
-        vllm_config._set_compile_ranges()
-        vllm_config._set_cudagraph_sizes()
-
-    assert (
-        vllm_config.compilation_config.use_inductor_graph_partition
-        == use_inductor_graph_partition
-    )
-    assert (
-        bool(vllm_config.compilation_config.splitting_ops) == expected_piecewise_compile
-    )
-    assert vllm_config.compilation_config.pass_config.enable_sp == expected_enable_sp
-    assert (
-        vllm_config.compilation_config.pass_config.fuse_gemm_comms == expected_enable_sp
-    )
-    assert vllm_config.compilation_config.cudagraph_mode == expected_cudagraph_mode
-    assert (
-        vllm_config.compilation_config.cudagraph_capture_sizes == expected_capture_sizes
-    )
-    assert (
-        vllm_config.compilation_config.max_cudagraph_capture_size == expected_max_size
-    )
-    assert (
-        511 in vllm_config.compilation_config.compile_ranges_endpoints
-    ) == expected_enable_sp
-
-
 def test_cached_compilation_config(default_vllm_config):
    import torch
    from torch._inductor.utils import run_and_get_code
@@ -753,24 +617,6 @@ def test_inductor_asserts_enabled_in_debug(monkeypatch):
        assert config.inductor_compile_config.get("scalar_asserts") is True


-def test_get_inductor_factors_includes_configs():
-    """Changing inductor or functorch config must change the cache key factors."""
-    from torch._functorch import config as functorch_config
-    from torch._inductor import config as inductor_config
-
-    from vllm.compilation.compiler_interface import get_inductor_factors
-
-    baseline = get_inductor_factors()
-
-    with inductor_config.patch("max_autotune", not inductor_config.max_autotune):
-        patched = get_inductor_factors()
-    assert baseline != patched, "inductor config change was not reflected"
-
-    with functorch_config.patch("donated_buffer", not functorch_config.donated_buffer):
-        patched = get_inductor_factors()
-    assert baseline != patched, "functorch config change was not reflected"
-
-
 def test_inductor_asserts_user_override(monkeypatch):
    """Test that explicit inductor_compile_config overrides the
    debug-logging default."""
@@ -28,7 +28,7 @@ def get_test_models():
        "Qwen/Qwen2-7B-Instruct",
        "meta-llama/Llama-3.1-8B",
    ]
-    if is_torch_equal_or_newer("2.12.0.dev"):
+    if is_torch_equal_or_newer("2.12.0"):
        models.append("Qwen/Qwen3-4B-Instruct-2507")
    return models

@@ -1183,7 +1183,7 @@ class VllmRunner:
        return [req_output.outputs.data for req_output in req_outputs]

    def reward(self, prompts: list[str]) -> list[list[float]]:
-        req_outputs = self.llm.encode(prompts, pooling_task="token_classify")
+        req_outputs = self.llm.reward(prompts)
        return [req_output.outputs.data for req_output in req_outputs]

    def score(
@@ -1,6 +1,7 @@
 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project

+import asyncio
 import random

 import pytest
@@ -360,14 +361,16 @@ def _test_async_transfer_layer_without_mtp_worker(
        communicator.set_stream(cuda_stream)

        for layer_idx in range(num_layers):
-            transfer_metadata = transfer_layer(
-                old_layer_indices=old_indices_cpu[layer_idx],
-                new_layer_indices=new_indices_cpu[layer_idx],
-                expert_weights=expert_weights[layer_idx],
-                expert_weights_buffer=expert_buffer,
-                ep_group=ep_group,
-                communicator=communicator,
-                cuda_stream=cuda_stream,
+            transfer_metadata = asyncio.run(
+                transfer_layer(
+                    old_layer_indices=old_indices_cpu[layer_idx],
+                    new_layer_indices=new_indices_cpu[layer_idx],
+                    expert_weights=expert_weights[layer_idx],
+                    expert_weights_buffer=expert_buffer,
+                    ep_group=ep_group,
+                    communicator=communicator,
+                    cuda_stream=cuda_stream,
+                )
            )
            cuda_stream.synchronize()
            move_from_buffer(
@@ -10,8 +10,7 @@ import torch.distributed as dist
 from vllm import LLM, SamplingParams
 from vllm.distributed.parallel_state import get_world_group

-# Let PyTorch choose the WORLD backend for the current device type.
-dist.init_process_group()
+dist.init_process_group(backend="gloo")

 # Create prompts
 prompts = [
@@ -10,8 +10,7 @@ import torch.distributed as dist
 from vllm import LLM, SamplingParams
 from vllm.distributed.parallel_state import get_tp_group, get_world_group

-# Let PyTorch choose the WORLD backend for the current device type.
-dist.init_process_group()
+dist.init_process_group(backend="gloo")

 # Create prompts
 prompts = [
@@ -12,7 +12,6 @@ from pydantic import Field
 from vllm.config import AttentionConfig, CompilationConfig, config
 from vllm.engine.arg_utils import (
    EngineArgs,
-    _expand_json_human_readable_numbers,
    contains_type,
    get_kwargs,
    get_type,
@@ -564,32 +563,3 @@ def test_ir_op_priority():
            ir_op_priority=ir_op_priority,
            kernel_config=KernelConfig(ir_op_priority=ir_op_priority),
        ).create_engine_config()
-
-
-@pytest.mark.parametrize(
-    ("input_json", "expected_json"),
-    [
-        # Decimal suffixes (lowercase)
-        ('{"x": 80g}', '{"x": 80000000000}'),
-        ('{"x": 1k}', '{"x": 1000}'),
-        ('{"x": 5m}', '{"x": 5000000}'),
-        ('{"x": 2t}', '{"x": 2000000000000}'),
-        # Binary suffixes (uppercase)
-        ('{"x": 1K}', f'{{"x": {2**10}}}'),
-        ('{"x": 1G}', f'{{"x": {2**30}}}'),
-        # Decimal values
-        ('{"x": 1.5g}', '{"x": 1500000000}'),
-        # Quoted strings must NOT be modified
-        ('{"my_key": 80g}', '{"my_key": 80000000000}'),
-        ('{"name": "80g"}', '{"name": "80g"}'),
-        ('{"model_name": "foo_bar"}', '{"model_name": "foo_bar"}'),
-        # Multiple values
-        ('{"a": 1k, "b": 2m}', '{"a": 1000, "b": 2000000}'),
-        # Plain numbers are untouched
-        ('{"x": 42}', '{"x": 42}'),
-        # Nested JSON
-        ('{"outer": {"inner": 10g}}', '{"outer": {"inner": 10000000000}}'),
-    ],
-)
-def test_expand_json_human_readable_numbers(input_json, expected_json):
-    assert _expand_json_human_readable_numbers(input_json) == expected_json
@@ -164,58 +164,6 @@ async def test_chat_error_non_stream():
        await serving_chat.create_chat_completion(request)


-@pytest.mark.asyncio
-async def test_openai_chat_keeps_mm_cache_for_engine_execution():
-    mock_engine = MagicMock(spec=AsyncLLM)
-    mock_engine.errored = False
-    mock_engine.model_config = MockModelConfig()
-    mock_engine.input_processor = MagicMock()
-    mock_engine.renderer = _build_renderer(mock_engine.model_config)
-
-    serving_chat = _build_serving_chat(mock_engine)
-
-    request = ChatCompletionRequest(
-        model=MODEL_NAME,
-        messages=[{"role": "user", "content": "Test prompt"}],
-    )
-
-    result = await serving_chat.render_chat_request(request)
-
-    assert isinstance(result, tuple)
-    assert (
-        serving_chat.openai_serving_render.preprocess_chat.call_args.kwargs[
-            "skip_mm_cache"
-        ]
-        is False
-    )
-
-
-@pytest.mark.asyncio
-async def test_renderer_only_chat_request_skips_mm_cache():
-    mock_engine = MagicMock(spec=AsyncLLM)
-    mock_engine.errored = False
-    mock_engine.model_config = MockModelConfig()
-    mock_engine.input_processor = MagicMock()
-    mock_engine.renderer = _build_renderer(mock_engine.model_config)
-
-    serving_chat = _build_serving_chat(mock_engine)
-
-    request = ChatCompletionRequest(
-        model=MODEL_NAME,
-        messages=[{"role": "user", "content": "Test prompt"}],
-    )
-
-    result = await serving_chat.openai_serving_render.render_chat_request(request)
-
-    assert result.token_ids == [1, 2, 3]
-    assert (
-        serving_chat.openai_serving_render.preprocess_chat.call_args.kwargs[
-            "skip_mm_cache"
-        ]
-        is True
-    )
-
-
@pytest.mark.asyncio
 async def test_chat_error_stream():
    """test finish_reason='error' returns 500 InternalServerError (streaming)"""
--- a/Show More
+++ b/Show More