type ignore

Signed-off-by: yewentao256 <zhyanwentao@126.com>
assert not None
2026-04-21 15:06:50 -04:00 · 2026-04-21 14:59:44 -04:00 · 2026-04-21 14:47:35 -04:00
645 changed files with 8241 additions and 47387 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

@@ -388,10 +388,10 @@ steps:
    - python3 basic/offline_inference/embed.py
    - python3 basic/offline_inference/score.py
    # Multi-modal models
-    - python3 generate/multimodal/audio_language_offline.py --seed 0
-    - python3 generate/multimodal/vision_language_offline.py --seed 0
-    - python3 generate/multimodal/vision_language_multi_image_offline.py --seed 0
-    - python3 generate/multimodal/encoder_decoder_multimodal_offline.py --model-type whisper --seed 0
+    - python3 offline_inference/audio_language.py --seed 0
+    - python3 offline_inference/vision_language.py --seed 0
+    - python3 offline_inference/vision_language_multi_image.py --seed 0
+    - python3 offline_inference/encoder_decoder_multimodal.py --model-type whisper --seed 0
    # Pooling models
    - python3 pooling/embed/vision_embedding_offline.py --seed 0
    # Features demo
@@ -1647,10 +1647,10 @@ steps:
    - python3 basic/offline_inference/embed.py
    - python3 basic/offline_inference/score.py
    # Multi-modal models
-    - python3 generate/multimodal/audio_language_offline.py --seed 0
-    - python3 generate/multimodal/vision_language_offline.py --seed 0
-    - python3 generate/multimodal/vision_language_multi_image_offline.py --seed 0
-    - python3 generate/multimodal/encoder_decoder_multimodal_offline.py --model-type whisper --seed 0
+    - python3 offline_inference/audio_language.py --seed 0
+    - python3 offline_inference/vision_language.py --seed 0
+    - python3 offline_inference/vision_language_multi_image.py --seed 0
+    - python3 offline_inference/encoder_decoder_multimodal.py --model-type whisper --seed 0
    # Pooling models
    - python3 pooling/embed/vision_embedding_offline.py --seed 0
    # Features demo
@@ -1951,8 +1951,8 @@ steps:
  - pytest -v -s tests/models/multimodal/processing/
  - pytest -v -s tests/models/multimodal/test_mapping.py
  - python3 examples/basic/offline_inference/chat.py
-  - python3 examples/generate/multimodal/vision_language_offline.py --model-type qwen2_5_vl
-  - VLLM_WORKER_MULTIPROC_METHOD=spawn python3 examples/generate/multimodal/audio_language_offline.py --model-type whisper
+  - python3 examples/offline_inference/vision_language.py --model-type qwen2_5_vl
+  - VLLM_WORKER_MULTIPROC_METHOD=spawn python3 examples/offline_inference/audio_language.py --model-type whisper

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

@@ -2930,10 +2930,10 @@ steps:
  - python3 basic/offline_inference/embed.py
  - python3 basic/offline_inference/score.py
  # Multi-modal models
-  - python3 generate/multimodal/audio_language_offline.py --seed 0
-  - python3 generate/multimodal/vision_language_offline.py --seed 0
-  - python3 generate/multimodal/vision_language_multi_image_offline.py --seed 0
-  - python3 generate/multimodal/encoder_decoder_multimodal_offline.py --model-type whisper --seed 0
+  - python3 offline_inference/audio_language.py --seed 0
+  - python3 offline_inference/vision_language.py --seed 0
+  - python3 offline_inference/vision_language_multi_image.py --seed 0
+  - python3 offline_inference/encoder_decoder_multimodal.py --model-type whisper --seed 0
  # Pooling models
  - python3 pooling/embed/vision_embedding_offline.py --seed 0
  # Features demo
@@ -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"
@@ -95,13 +95,11 @@ steps:
  - tests/kernels/moe/test_deepgemm.py
  - tests/kernels/moe/test_batched_deepgemm.py
  - tests/kernels/attention/test_deepgemm_attention.py
-  - tests/quantization/test_cutlass_w4a16.py
  commands:
    - pytest -v -s kernels/quantization/test_block_fp8.py
    - pytest -v -s kernels/moe/test_deepgemm.py
    - pytest -v -s kernels/moe/test_batched_deepgemm.py
    - pytest -v -s kernels/attention/test_deepgemm_attention.py
-    - pytest -v -s quantization/test_cutlass_w4a16.py

 - label: Kernels (B200)
  timeout_in_minutes: 30
@@ -113,10 +113,10 @@ steps:
    - python3 basic/offline_inference/embed.py
    - python3 basic/offline_inference/score.py
    # for multi-modal models
-    - python3 generate/multimodal/audio_language_offline.py --seed 0
-    - python3 generate/multimodal/vision_language_offline.py --seed 0
-    - python3 generate/multimodal/vision_language_multi_image_offline.py --seed 0
-    - python3 generate/multimodal/encoder_decoder_multimodal_offline.py --model-type whisper --seed 0
+    - python3 offline_inference/audio_language.py --seed 0
+    - python3 offline_inference/vision_language.py --seed 0
+    - python3 offline_inference/vision_language_multi_image.py --seed 0
+    - python3 offline_inference/encoder_decoder_multimodal.py --model-type whisper --seed 0
     # for pooling models
    - python3 pooling/embed/vision_embedding_offline.py --seed 0
    # for features demo
@@ -44,10 +44,10 @@ steps:
    #- python3 basic/offline_inference/generate.py --model meta-llama/Llama-2-13b-chat-hf --cpu-offload-gb 10  # TODO
    #- python3 basic/offline_inference/embed.py   # TODO
    # for multi-modal models
-    - python3 generate/multimodal/audio_language_offline.py --seed 0
-    - python3 generate/multimodal/vision_language_offline.py --seed 0
-    - python3 generate/multimodal/vision_language_multi_image_offline.py --seed 0
-    - python3 generate/multimodal/encoder_decoder_multimodal_offline.py --model-type whisper --seed 0
+    - python3 offline_inference/audio_language.py --seed 0
+    - python3 offline_inference/vision_language.py --seed 0
+    - python3 offline_inference/vision_language_multi_image.py --seed 0
+    - python3 offline_inference/encoder_decoder_multimodal.py --model-type whisper --seed 0
    # for pooling models
    - python3 pooling/embed/vision_embedding_offline.py --seed 0
    # for features demo
@@ -69,9 +69,9 @@ steps:
    - pytest -v -s tests/models/multimodal/processing/
    - pytest -v -s tests/models/multimodal/test_mapping.py
    - python3 examples/basic/offline_inference/chat.py
-    - python3 examples/generate/multimodal/vision_language_offline.py --model-type qwen2_5_vl
+    - python3 examples/offline_inference/vision_language.py --model-type qwen2_5_vl
    # Whisper needs spawn method to avoid deadlock
-    - VLLM_WORKER_MULTIPROC_METHOD=spawn python3 examples/generate/multimodal/audio_language_offline.py --model-type whisper
+    - VLLM_WORKER_MULTIPROC_METHOD=spawn python3 examples/offline_inference/audio_language.py --model-type whisper

 - label: Transformers Backward Compatibility Models Test
  working_dir: "/vllm-workspace/"
@@ -83,7 +83,7 @@ steps:
    - pytest -v -s tests/models/test_transformers.py
    - pytest -v -s tests/models/multimodal/processing/
    - pytest -v -s tests/models/multimodal/test_mapping.py
-    - python3 examples/basic/offline_inference/chat.py
-    - python3 examples/generate/multimodal/vision_language_offline.py --model-type qwen2_5_vl
+    - python3 examples/offline_inference/basic/chat.py
+    - python3 examples/offline_inference/vision_language.py --model-type qwen2_5_vl
    # Whisper needs spawn method to avoid deadlock
-    - VLLM_WORKER_MULTIPROC_METHOD=spawn python3 examples/generate/multimodal/audio_language_offline.py --model-type whisper
+    - VLLM_WORKER_MULTIPROC_METHOD=spawn python3 examples/offline_inference/audio_language.py --model-type whisper
@@ -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
@@ -389,7 +389,11 @@ pull_request_rules:
      - files~=^tests/entrypoints/anthropic/.*tool.*
      - files~=^vllm/tool_parsers/
      - files=docs/features/tool_calling.md
-      - files~=^examples/tool_calling/
+      - files~=^examples/tool_chat_*
+      - files=examples/offline_inference/chat_with_tools.py
+      - files=examples/online_serving/openai_chat_completion_client_with_tools_required.py
+      - files=examples/online_serving/openai_chat_completion_tool_calls_with_reasoning.py
+      - files=examples/online_serving/openai_chat_completion_client_with_tools.py
  actions:
    label:
      add:
@@ -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,
@@ -1,324 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-
-# Benchmarks FP8 vs BF16 ViT attention via FlashInfer cuDNN backend.
-#
-# == Usage Examples ==
-#
-# Benchmark mode (default, FlashInfer CUDAGraph Bench)
-#   python3 benchmark_vit_fp8_attn.py
-#
-# Profile mode (PyTorch profiler, saves TensorBoard traces):
-#   python3 benchmark_vit_fp8_attn.py --profile
-#   python3 benchmark_vit_fp8_attn.py --profile --profile-output-dir ./profile_traces
-#
-# Custom seq_lens:
-#   python3 benchmark_vit_fp8_attn.py --seq-lens 4096 8192 16384
-
-from functools import partial
-
-import numpy as np
-import torch
-from torch.profiler import ProfilerActivity, profile, record_function
-
-from vllm.utils.argparse_utils import FlexibleArgumentParser
-
-# Qwen3-VL defaults
-NUM_HEADS = 16
-HEAD_DIM = 72
-DEFAULT_SEQ_LENS = [2304, 4096, 8192, 16384]
-
-
-def _setup_fp8_attention(num_heads: int, head_dim: int) -> tuple:
-    """Create FP8 and BF16 attention modules + workspace."""
-    from types import SimpleNamespace
-    from unittest.mock import patch
-
-    from vllm.config import VllmConfig, set_current_vllm_config
-    from vllm.config.multimodal import MultiModalConfig
-    from vllm.model_executor.layers.attention.mm_encoder_attention import (
-        MMEncoderAttention,
-        _get_flashinfer_workspace_buffer,
-    )
-    from vllm.v1.attention.backends.registry import AttentionBackendEnum
-
-    old_dtype = torch.get_default_dtype()
-    torch.set_default_dtype(torch.bfloat16)
-
-    backend_patch = patch(
-        "vllm.model_executor.layers.attention.mm_encoder_attention"
-        ".get_vit_attn_backend",
-        return_value=AttentionBackendEnum.FLASHINFER,
-    )
-
-    # FP8 attention
-    mm_config_fp8 = MultiModalConfig(mm_encoder_attn_dtype="fp8")
-    vllm_config_fp8 = VllmConfig()
-    vllm_config_fp8.model_config = SimpleNamespace(multimodal_config=mm_config_fp8)
-    with set_current_vllm_config(vllm_config_fp8), backend_patch:
-        attn_fp8 = MMEncoderAttention(
-            num_heads=num_heads,
-            head_size=head_dim,
-            prefix="visual.blocks.0.attn",
-        ).to("cuda")
-
-    # BF16 attention (no FP8)
-    with set_current_vllm_config(VllmConfig()), backend_patch:
-        attn_bf16 = MMEncoderAttention(
-            num_heads=num_heads,
-            head_size=head_dim,
-            prefix="visual.blocks.0.attn",
-        ).to("cuda")
-
-    torch.set_default_dtype(old_dtype)
-
-    workspace = _get_flashinfer_workspace_buffer()
-    return attn_fp8, attn_bf16, workspace
-
-
-def _build_meta(
-    seq_len: int,
-    num_heads: int,
-    head_dim: int,
-    fp8: bool,
-):
-    """Build cu_seqlens, max_seqlen, sequence_lengths."""
-    from vllm.model_executor.layers.attention.mm_encoder_attention import (
-        MMEncoderAttention,
-    )
-    from vllm.utils.math_utils import round_up
-    from vllm.v1.attention.backends.registry import AttentionBackendEnum
-
-    cu_np = np.array([0, seq_len], dtype=np.int32)
-    fp8_padded = num_heads * round_up(head_dim, 16) if fp8 else None
-
-    seq_lengths = MMEncoderAttention.maybe_compute_seq_lens(
-        AttentionBackendEnum.FLASHINFER, cu_np, torch.device("cuda")
-    )
-    max_seqlen = torch.tensor(
-        MMEncoderAttention.compute_max_seqlen(AttentionBackendEnum.FLASHINFER, cu_np),
-        dtype=torch.int32,
-    )
-    cu_seqlens = MMEncoderAttention.maybe_recompute_cu_seqlens(
-        AttentionBackendEnum.FLASHINFER,
-        cu_np,
-        num_heads * head_dim,
-        1,
-        torch.device("cuda"),
-        fp8_padded_hidden_size=fp8_padded,
-    )
-    return cu_seqlens, max_seqlen, seq_lengths
-
-
-def run_benchmark(
-    seq_lens: list[int],
-    num_heads: int,
-    head_dim: int,
-    method: str,
-):
-    """Benchmark FP8 vs BF16 attention across seq_lens.
-
-    Uses FlashInfer GPU-level timing to measure pure kernel time,
-    excluding CPU launch overhead.
-    """
-    if method == "cupti":
-        from flashinfer.testing import bench_gpu_time_with_cupti as bench_fn
-
-        bench_fn = partial(bench_fn, use_cuda_graph=True, cold_l2_cache=False)
-    elif method == "cudagraph":
-        from flashinfer.testing import (
-            bench_gpu_time_with_cudagraph as bench_fn,
-        )
-
-        bench_fn = partial(bench_fn, cold_l2_cache=False)
-    else:
-        raise ValueError(f"Invalid method: {method}")
-
-    attn_fp8, attn_bf16, workspace = _setup_fp8_attention(num_heads, head_dim)
-
-    print(f"Timing method: {method}")
-    print(f"{'seq_len':>8} {'BF16 (us)':>12} {'FP8 (us)':>12} {'Speedup':>10}")
-    print("-" * 46)
-
-    for seq_len in seq_lens:
-        torch.manual_seed(42)
-
-        q = torch.randn(
-            seq_len,
-            num_heads,
-            head_dim,
-            device="cuda",
-            dtype=torch.bfloat16,
-        )
-        k = torch.randn_like(q)
-        v = torch.randn_like(q)
-
-        cu_fp8, max_s, seq_l = _build_meta(seq_len, num_heads, head_dim, fp8=True)
-        # we can reuse cu_fp8 for cu_bf16 since q, k, and v are contiguous
-        cu_bf16 = cu_fp8.clone()
-
-        def bf16_fn(q=q, k=k, v=v, cu=cu_bf16, ms=max_s, sl=seq_l):
-            attn_bf16._forward_flashinfer(q, k, v, cu, ms, sl)
-
-        def fp8_fn(q=q, k=k, v=v, cu=cu_fp8, ms=max_s, sl=seq_l):
-            attn_fp8._forward_flashinfer(q, k, v, cu, ms, sl)
-
-        # bench_fn returns List[float] of per-iteration times in ms
-        bf16_times = bench_fn(bf16_fn)
-        fp8_times = bench_fn(fp8_fn)
-
-        bf16_us = np.median(bf16_times) * 1e3  # ms -> us
-        fp8_us = np.median(fp8_times) * 1e3
-        speedup = bf16_us / fp8_us if fp8_us > 0 else float("inf")
-
-        print(f"{seq_len:>8} {bf16_us:>12.1f} {fp8_us:>12.1f} {speedup:>9.2f}x")
-
-
-def _make_trace_handler(output_dir: str, worker_name: str, label: str):
-    """Create a trace handler that saves to TensorBoard and prints summary."""
-
-    def handler(prof):
-        torch.profiler.tensorboard_trace_handler(output_dir, worker_name)(prof)
-        print(f"\n{'=' * 80}")
-        print(label)
-        print(f"{'=' * 80}")
-        print(prof.key_averages().table(sort_by="cuda_time_total", row_limit=20))
-
-    return handler
-
-
-def run_profile(
-    seq_len: int,
-    num_heads: int,
-    head_dim: int,
-    warmup: int,
-    output_dir: str,
-):
-    """Profile FP8 vs BF16 attention with PyTorch profiler."""
-    attn_fp8, attn_bf16, workspace = _setup_fp8_attention(num_heads, head_dim)
-
-    torch.manual_seed(42)
-    q = torch.randn(
-        seq_len,
-        num_heads,
-        head_dim,
-        device="cuda",
-        dtype=torch.bfloat16,
-    )
-    k = torch.randn_like(q)
-    v = torch.randn_like(q)
-
-    cu_fp8, max_s, seq_l = _build_meta(seq_len, num_heads, head_dim, fp8=True)
-    # we can reuse cu_fp8 for cu_bf16 since q, k, and v are contiguous
-    cu_bf16 = cu_fp8.clone()
-
-    sched = torch.profiler.schedule(wait=0, warmup=warmup, active=1)
-
-    # Profile BF16 (warmup handled by profiler schedule)
-    with profile(
-        activities=[ProfilerActivity.CPU, ProfilerActivity.CUDA],
-        schedule=sched,
-        on_trace_ready=_make_trace_handler(
-            output_dir,
-            f"bf16_h{head_dim}_s{seq_len}",
-            f"BF16 Attention (seq_len={seq_len}, heads={num_heads}, "
-            f"head_dim={head_dim})",
-        ),
-    ) as prof_bf16:
-        for _ in range(warmup + 1):
-            with record_function("bf16_attention"):
-                attn_bf16._forward_flashinfer(
-                    q.clone(), k.clone(), v.clone(), cu_bf16, max_s, seq_l
-                )
-                torch.accelerator.synchronize()
-            prof_bf16.step()
-
-    # Profile FP8 (warmup handled by profiler schedule)
-    with profile(
-        activities=[ProfilerActivity.CPU, ProfilerActivity.CUDA],
-        schedule=sched,
-        on_trace_ready=_make_trace_handler(
-            output_dir,
-            f"fp8_h{head_dim}_s{seq_len}",
-            f"FP8 Attention (seq_len={seq_len}, heads={num_heads}, "
-            f"head_dim={head_dim})",
-        ),
-    ) as prof_fp8:
-        for _ in range(warmup + 1):
-            with record_function("fp8_attention"):
-                attn_fp8._forward_flashinfer(
-                    q.clone(), k.clone(), v.clone(), cu_fp8, max_s, seq_l
-                )
-                torch.accelerator.synchronize()
-            prof_fp8.step()
-
-    print(f"\nTensorBoard traces saved to: {output_dir}")
-    print(f"View with: tensorboard --logdir={output_dir}")
-
-
-if __name__ == "__main__":
-    parser = FlexibleArgumentParser(description="Benchmark FP8 vs BF16 ViT attention.")
-    parser.add_argument(
-        "--seq-lens",
-        type=int,
-        nargs="+",
-        default=DEFAULT_SEQ_LENS,
-        help="Sequence lengths to benchmark",
-    )
-    parser.add_argument(
-        "--num-heads",
-        type=int,
-        default=NUM_HEADS,
-    )
-    parser.add_argument(
-        "--head-dim",
-        type=int,
-        default=HEAD_DIM,
-    )
-    parser.add_argument(
-        "--method",
-        choices=["cupti", "cudagraph"],
-        default="cudagraph",
-        help="GPU timing method: cupti (CUPTI kernel timing) or "
-        "cudagraph (CUDA graph capture/replay). Default: cudagraph",
-    )
-    parser.add_argument(
-        "--warmup",
-        type=int,
-        default=10,
-        help="Warmup iterations (profile mode only)",
-    )
-    parser.add_argument(
-        "--profile",
-        action="store_true",
-        help="Run PyTorch profiler instead of benchmark",
-    )
-    parser.add_argument(
-        "--profile-seq-len",
-        type=int,
-        default=8192,
-        help="Sequence length for profiling (default: 8192)",
-    )
-    parser.add_argument(
-        "--profile-output-dir",
-        type=str,
-        default="./profile_traces",
-        help="Output directory for TensorBoard traces (default: ./profile_traces)",
-    )
-    args = parser.parse_args()
-
-    if args.profile:
-        run_profile(
-            args.profile_seq_len,
-            args.num_heads,
-            args.head_dim,
-            args.warmup,
-            args.profile_output_dir,
-        )
-    else:
-        run_benchmark(
-            args.seq_lens,
-            args.num_heads,
-            args.head_dim,
-            args.method,
-        )
@@ -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
@@ -15,12 +15,16 @@
 #include <torch/all.h>
 namespace vec_op {

-// BFloat16 is always supported on RISC-V: natively when RISCV_BF16_SUPPORT
-// is defined, otherwise via the FP32-simulation fallback path.
-#define VLLM_DISPATCH_CASE_FLOATING_TYPES(...)         \
-  AT_DISPATCH_CASE(at::ScalarType::Float, __VA_ARGS__) \
-  AT_DISPATCH_CASE(at::ScalarType::Half, __VA_ARGS__)  \
-  AT_DISPATCH_CASE(at::ScalarType::BFloat16, __VA_ARGS__)
+#ifdef RISCV_BF16_SUPPORT
+  #define VLLM_DISPATCH_CASE_FLOATING_TYPES(...)         \
+    AT_DISPATCH_CASE(at::ScalarType::Float, __VA_ARGS__) \
+    AT_DISPATCH_CASE(at::ScalarType::Half, __VA_ARGS__)  \
+    AT_DISPATCH_CASE(at::ScalarType::BFloat16, __VA_ARGS__)
+#else
+  #define VLLM_DISPATCH_CASE_FLOATING_TYPES(...)         \
+    AT_DISPATCH_CASE(at::ScalarType::Float, __VA_ARGS__) \
+    AT_DISPATCH_CASE(at::ScalarType::Half, __VA_ARGS__)
+#endif

 #define VLLM_DISPATCH_FLOATING_TYPES(TYPE, NAME, ...) \
  AT_DISPATCH_SWITCH(TYPE, NAME, VLLM_DISPATCH_CASE_FLOATING_TYPES(__VA_ARGS__))
@@ -482,18 +486,9 @@ struct FP32Vec8 : public Vec<FP32Vec8> {
  }

  FP32Vec8 exp() const {
-    // Clamp input to prevent NaN: exp(-inf) must return 0, not NaN.
-    // Without clamping, -inf * 0.0 = NaN in the final poly * scale step.
-    // Matches the clamping strategy used by x86 AVX-512 and ARM NEON.
-    constexpr float exp_lo = -87.3365447505f;  // ln(FLT_MIN)
-    constexpr float exp_hi = 88.7228391117f;   // ln(FLT_MAX)
-    fixed_fp32x8_t x = RVVI(__riscv_vfmin_vf_f32, LMUL_256)(
-        RVVI(__riscv_vfmax_vf_f32, LMUL_256)(reg, exp_lo, VEC_ELEM_NUM), exp_hi,
-        VEC_ELEM_NUM);
-
    const float inv_ln2 = 1.44269504088896341f;
    fixed_fp32x8_t x_scaled =
-        RVVI(__riscv_vfmul_vf_f32, LMUL_256)(x, inv_ln2, VEC_ELEM_NUM);
+        RVVI(__riscv_vfmul_vf_f32, LMUL_256)(reg, inv_ln2, VEC_ELEM_NUM);
    fixed_i32x8_t n_int =
        RVVI(__riscv_vfcvt_x_f_v_i32, LMUL_256)(x_scaled, VEC_ELEM_NUM);
    fixed_fp32x8_t n_float =
@@ -711,18 +706,9 @@ struct FP32Vec16 : public Vec<FP32Vec16> {
  }

  FP32Vec16 exp() const {
-    // Clamp input to prevent NaN: exp(-inf) must return 0, not NaN.
-    // Without clamping, -inf * 0.0 = NaN in the final poly * scale step.
-    // Matches the clamping strategy used by x86 AVX-512 and ARM NEON.
-    constexpr float exp_lo = -87.3365447505f;  // ln(FLT_MIN)
-    constexpr float exp_hi = 88.7228391117f;   // ln(FLT_MAX)
-    fixed_fp32x16_t x = RVVI(__riscv_vfmin_vf_f32, LMUL_512)(
-        RVVI(__riscv_vfmax_vf_f32, LMUL_512)(reg, exp_lo, VEC_ELEM_NUM), exp_hi,
-        VEC_ELEM_NUM);
-
    const float inv_ln2 = 1.44269504088896341f;
    fixed_fp32x16_t x_scaled =
-        RVVI(__riscv_vfmul_vf_f32, LMUL_512)(x, inv_ln2, VEC_ELEM_NUM);
+        RVVI(__riscv_vfmul_vf_f32, LMUL_512)(reg, inv_ln2, VEC_ELEM_NUM);
    fixed_i32x16_t n_int =
        RVVI(__riscv_vfcvt_x_f_v_i32, LMUL_512)(x_scaled, VEC_ELEM_NUM);
    fixed_fp32x16_t n_float =
@@ -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);
  }
 }

@@ -277,9 +277,7 @@ void quant_impl(void* output, void* output_scale, void* input,
      (totalWorkSize + block.x * grid.x - 1) / (block.x * grid.x);
  if (blockRepeat > 1) {
    size_t shared_mem_size = (n_experts + 1) * sizeof(uint32_t);
-    // The shared-memory vectorized offset load only handles full 4-expert
-    // chunks. Use the scalar specialization for the remainder cases.
-    if (n_experts >= 4 && n_experts % 4 == 0) {
+    if (n_experts >= 4) {
      cvt_fp16_to_fp4<T, FUSE_SILU_MUL, false, false>
          <<<grid, block, shared_mem_size, stream>>>(
              m_topk, k, reinterpret_cast<T*>(input),
@@ -301,9 +299,7 @@ void quant_impl(void* output, void* output_scale, void* input,
              n_experts);
    }
  } else {
-    // The low-latency vectorized expert lookup only handles full 16-expert
-    // chunks. Fall back to the scalar lookup path for the remainder cases.
-    if (n_experts >= 16 && n_experts % 16 == 0) {
+    if (n_experts >= 16) {
      cvt_fp16_to_fp4<T, FUSE_SILU_MUL, false, false>
          <<<grid, block, 0, stream>>>(
              m_topk, k, reinterpret_cast<T*>(input),
@@ -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 =
@@ -84,7 +101,7 @@ void launch_persistent_topk(const torch::Tensor& logits,

    int occupancy = 1;
    cudaOccupancyMaxActiveBlocksPerMultiprocessor(
-        &occupancy, P::persistent_topk_kernel<TopK, 4>, P::kThreadsPerBlock,
+        &occupancy, P::persistent_topk_kernel<4>, P::kThreadsPerBlock,
        smem_size);
    if (occupancy < 1) occupancy = 1;

@@ -104,16 +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,                                         \
@@ -122,11 +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
  }
@@ -134,46 +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 ####################

@@ -538,7 +538,7 @@ RUN CUDA_VERSION_DASH=$(echo $CUDA_VERSION | cut -d. -f1,2 | tr '.' '-') && \
        cuda-nvrtc-${CUDA_VERSION_DASH} \
        cuda-cuobjdump-${CUDA_VERSION_DASH} \
        libcurand-dev-${CUDA_VERSION_DASH} \
-        libcublas-dev-${CUDA_VERSION_DASH} \
+        libcublas-${CUDA_VERSION_DASH} \
        # Required by fastsafetensors (fixes #20384)
        libnuma-dev && \
    # Fixes nccl_allocator requiring nccl.h at runtime
@@ -765,7 +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} \
@@ -312,30 +307,6 @@ RUN --mount=type=bind,source=.git,target=vllm/.git \
    && echo "Detected vLLM version: ${VLLM_VERSION}" \
    && echo "${VLLM_VERSION}" > /tmp/vllm_version.txt

-# Fail if git-based package dependencies are found in requirements files
-# (uv doesn't handle git+ URLs well, and packages should be distributed on PyPI)
-# Extra notes: pip install is able to handle git+ URLs, but uv doesn't.
-RUN echo "Checking for git-based packages in requirements files..." \
-    && echo "Checking common.txt for git-based packages:" \
-    && if grep -q 'git+' ${COMMON_WORKDIR}/vllm/requirements/common.txt; then \
-         echo "ERROR: Git-based packages found in common.txt:"; \
-         grep 'git+' ${COMMON_WORKDIR}/vllm/requirements/common.txt; \
-         echo "Please publish these packages to PyPI instead of using git dependencies."; \
-         exit 1; \
-       else \
-         echo "  ✓ No git-based packages found in common.txt"; \
-       fi \
-    && echo "Checking rocm.txt for git-based packages:" \
-    && if grep -q 'git+' ${COMMON_WORKDIR}/vllm/requirements/rocm.txt; then \
-         echo "ERROR: Git-based packages found in rocm.txt:"; \
-         grep 'git+' ${COMMON_WORKDIR}/vllm/requirements/rocm.txt; \
-         echo "Please publish these packages to PyPI instead of using git dependencies."; \
-         exit 1; \
-       else \
-         echo "  ✓ No git-based packages found in rocm.txt"; \
-       fi \
-    && echo "All requirements files are clean - no git-based packages found"
-
 # Pin vLLM dependencies to exact versions of custom ROCm wheels
 # This ensures 'pip install vllm' automatically installs correct torch/triton/torchvision/amdsmi
 COPY tools/vllm-rocm/pin_rocm_dependencies.py /tmp/pin_rocm_dependencies.py
@@ -436,10 +407,6 @@ COPY --from=export_vllm /vllm_v1 /usr/local/lib/python${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.
@@ -66,7 +66,7 @@ This is for controlling general behavior of the API when serving your model:

 See [Audio preprocessing and chunking](#audio-preprocessing-and-chunking) for what each field controls.

-Implement the prompt construction via [get_generation_prompt][vllm.model_executor.models.interfaces.SupportsTranscription.get_generation_prompt]. The server builds a [SpeechToTextParams][vllm.config.speech_to_text.SpeechToTextParams] object that bundles the resampled waveform, task parameters, and request-specific options. Your model receives this single object and returns a valid [PromptType][vllm.inputs.llm.PromptType]. There are two common patterns:
+Implement the prompt construction via [get_generation_prompt][vllm.model_executor.models.interfaces.SupportsTranscription.get_generation_prompt]. The server passes you the resampled waveform and task parameters; you return a valid [PromptType][vllm.inputs.llm.PromptType]. There are two common patterns:

 #### Multimodal LLM with audio embeddings (e.g., Voxtral, Gemma3n)

@@ -75,20 +75,21 @@ Return a dict containing `multi_modal_data` with the audio, and either a `prompt
 ??? code "get_generation_prompt()"

    ```python
-    from vllm.config.speech_to_text import SpeechToTextParams
-
    class YourASRModel(nn.Module, SupportsTranscription):
        ...

        @classmethod
        def get_generation_prompt(
            cls,
-            stt_params: SpeechToTextParams,
+            audio: np.ndarray,
+            stt_config: SpeechToTextConfig,
+            model_config: ModelConfig,
+            language: str | None,
+            task_type: Literal["transcribe", "translate"],
+            request_prompt: str,
+            to_language: str | None,
        ) -> PromptType:
-            audio = stt_params.audio
-            stt_config = stt_params.stt_config
-            task_type = stt_params.task_type
-
+            # Example with a free-form instruction prompt
            task_word = "Transcribe" if task_type == "transcribe" else "Translate"
            prompt = (
                "<start_of_turn>user\n"
@@ -111,22 +112,20 @@ Return a dict with separate `encoder_prompt` and `decoder_prompt` entries:
 ??? code "get_generation_prompt()"

    ```python
-    from vllm.config.speech_to_text import SpeechToTextParams
-
    class YourASRModel(nn.Module, SupportsTranscription):
        ...

        @classmethod
        def get_generation_prompt(
            cls,
-            stt_params: SpeechToTextParams,
+            audio: np.ndarray,
+            stt_config: SpeechToTextConfig,
+            model_config: ModelConfig,
+            language: str | None,
+            task_type: Literal["transcribe", "translate"],
+            request_prompt: str,
+            to_language: str | None,
        ) -> PromptType:
-            audio = stt_params.audio
-            stt_config = stt_params.stt_config
-            language = stt_params.language
-            task_type = stt_params.task_type
-            request_prompt = stt_params.request_prompt
-
            if language is None:
                raise ValueError("Language must be specified")

@@ -214,13 +213,15 @@ Relevant server logic:
        chunks = [y] if not do_split_audio else self._split_audio(y, int(sr))
        prompts = []
        for chunk in chunks:
-            stt_params = request.build_stt_params(
+            prompt = self.model_cls.get_generation_prompt(
                audio=chunk,
                stt_config=self.asr_config,
                model_config=self.model_config,
+                language=language,
                task_type=self.task_type,
+                request_prompt=request.prompt,
+                to_language=to_language,
            )
-            prompt = self.model_cls.get_generation_prompt(stt_params)
            prompts.append(prompt)
        return prompts, duration
    ```
@@ -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 |
@@ -44,7 +44,7 @@ The table below lists the quantization schemes supported by each fusion on each
 | `fuse_allreduce_rms`         | FP16/BF16, FP8 static, NVFP4             | FP16/BF16, FP8 static                    | —                                        | —             | —                                        |
 | `fuse_minimax_qk_norm`\*     | FP16/BF16                                | FP16/BF16                                | FP16/BF16                                | FP16/BF16     | —                                        |
 | `fuse_attn_quant`\*          | FP8 static\*, NVFP4\*                    | FP8 static\*                             | FP8 static\*                             | —             | FP8 static\*                             |
-| `fuse_attn_quant` (MLA)\*    | FP8 static\*, FP8 per-group\*, NVFP4\*   | FP8 static\*, FP8 per-group\*            | FP8 static\*, FP8 per-group\*            | —             | FP8 static\* (untested)                  |
+| `fuse_attn_quant` (MLA)\*    | FP8 static\*, NVFP4\*                    | FP8 static\*                             | FP8 static\*                             | —             | FP8 static(untested)\*                   |
 | `fuse_rope_kvcache`          | —                                        | —                                        | —                                        | —             | FP16/BF16                                |
 | `enable_qk_norm_rope_fusion` | FP16/BF16                                | FP16/BF16                                | FP16/BF16†                               | FP16/BF16†    | —                                        |
 | `enable_sp`                  | FP16/BF16, FP8 static†                   | FP16/BF16, FP8 static                    | FP16/BF16†                               | FP16/BF16†    | —                                        |
@@ -152,7 +152,7 @@ standard `Attention` and `MLAAttention` (used by DeepSeek-V2/V3/R1 models). Patt

 - `FLASHINFER`: CUDA sm100+ with FlashInfer installed

-`MLAAttention → FP8 static, FP8 per-group, NVFP4 dynamic quant`
+`MLAAttention → FP8 static quant` / `MLAAttention → NVFP4 dynamic quant`:

 The MLA fusion operates at the graph level on the `unified_mla_attention_with_output` op and works
 with all MLA decode and prefill backend combinations. Unlike standard `Attention` backends (where
@@ -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.
@@ -68,7 +68,7 @@ You can pass a single image to the `'image'` field of the multi-modal dictionary
        print(generated_text)
    ```

-Full example: [examples/generate/multimodal/vision_language_offline.py](../../examples/generate/multimodal/vision_language_offline.py)
+Full example: [examples/offline_inference/vision_language.py](../../examples/offline_inference/vision_language.py)

 To substitute multiple images inside the same text prompt, you can pass in a list of images instead:

@@ -101,7 +101,7 @@ To substitute multiple images inside the same text prompt, you can pass in a lis
        print(generated_text)
    ```

-Full example: [examples/generate/multimodal/vision_language_multi_image_offline.py](../../examples/generate/multimodal/vision_language_multi_image_offline.py)
+Full example: [examples/offline_inference/vision_language_multi_image.py](../../examples/offline_inference/vision_language_multi_image.py)

 If using the [LLM.chat](../models/generative_models.md#llmchat) method, you can pass images directly in the message content using various formats: image URLs, PIL Image objects, or pre-computed embeddings:

@@ -287,13 +287,13 @@ Instead of NumPy arrays, you can also pass `'torch.Tensor'` instances, as shown
    !!! note
        'process_vision_info' is only applicable to Qwen2.5-VL and similar models.

-Full example: [examples/generate/multimodal/vision_language_offline.py](../../examples/generate/multimodal/vision_language_offline.py)
+Full example: [examples/offline_inference/vision_language.py](../../examples/offline_inference/vision_language.py)

 ### Audio Inputs

 You can pass a tuple `(array, sampling_rate)` to the `'audio'` field of the multi-modal dictionary.

-Full example: [examples/generate/multimodal/audio_language_offline.py](../../examples/generate/multimodal/audio_language_offline.py)
+Full example: [examples/offline_inference/audio_language.py](../../examples/offline_inference/audio_language.py)

 #### Chunking Long Audio for Transcription

@@ -674,7 +674,7 @@ Then, you can use the OpenAI client as follows:
    print("Chat completion output:", chat_response.choices[0].message.content)
    ```

-Full example: [examples/generate/multimodal/openai_chat_completion_client_for_multimodal.py](../../examples/generate/multimodal/openai_chat_completion_client_for_multimodal.py)
+Full example: [examples/online_serving/openai_chat_completion_client_for_multimodal.py](../../examples/online_serving/openai_chat_completion_client_for_multimodal.py)

 !!! tip
    Loading from local file paths is also supported on vLLM: You can specify the allowed local media path via `--allowed-local-media-path` when launching the API server/engine,
@@ -745,7 +745,7 @@ Then, you can use the OpenAI client as follows:
    print("Chat completion output from image url:", result)
    ```

-Full example: [examples/generate/multimodal/openai_chat_completion_client_for_multimodal.py](../../examples/generate/multimodal/openai_chat_completion_client_for_multimodal.py)
+Full example: [examples/online_serving/openai_chat_completion_client_for_multimodal.py](../../examples/online_serving/openai_chat_completion_client_for_multimodal.py)

 !!! note
    By default, the timeout for fetching videos through HTTP URL is `30` seconds.
@@ -780,70 +780,6 @@ vllm serve Qwen/Qwen3-VL-30B-A3B-Instruct \

 Works with common video formats like MP4 when using OpenCV backends.

-#### Pre-extracted Frame Sequences with `media_io_kwargs`
-
-When you extract video frames on the client side and send them as `video/jpeg` (base64-concatenated JPEG frames), you can preserve the original video metadata by using `media_io_kwargs` in your request. This enables more accurate video understanding by preserving temporal information that would otherwise be lost during client-side frame extraction.
-
-**Supported Parameters:**
-
-| Parameter | Type | Description |
-| --------- | ---- | ----------- |
-| `fps` | float | Frame rate of the original video |
-| `frames_indices` | list[int] | Indices of the actually sampled frames |
-| `total_num_frames` | int | Total frame count of the original video |
-| `duration` | float | Duration of the original video in seconds |
-| `do_sample_frames` | bool | Whether to perform frame sampling |
-
-??? code
-
-    ```python
-    from openai import OpenAI
-
-    client = OpenAI(base_url="http://localhost:8000/v1", api_key="EMPTY")
-
-    # Client-side frame extraction
-    frames = extract_frames(video_path, num_frames=32)
-    frames_b64 = ",".join([encode_image(f) for f in frames])
-    video_url = f"data:video/jpeg;base64,{frames_b64}"
-
-    # Pass video metadata via media_io_kwargs
-    response = client.chat.completions.create(
-        model="your-multimodal-model",
-        messages=[{
-            "role": "user",
-            "content": [
-                {"type": "video_url", "video_url": {"url": video_url}},
-                {"type": "text", "text": "Describe what happens in this video."}
-            ]
-        }],
-        extra_body={
-            "media_io_kwargs": {
-                "video": {
-                    "fps": 30.0,
-                    "frames_indices": [0, 10, 20, 30, 40, 50, 60, 70, 80, 90,
-                                       100, 110, 120, 130, 140, 150, 160, 170,
-                                       180, 190, 200, 210, 220, 230, 240, 250,
-                                       260, 270, 280, 290, 300, 310],
-                    "total_num_frames": 900,
-                    "duration": 30.0,
-                }
-            }
-        },
-    )
-
-    print(response.choices[0].message.content)
-    ```
-
-**Why use `media_io_kwargs`?**
-
-When extracting frames client-side, the server loses important context about the original video:
-
- **Temporal information**: Which frames were sampled and their positions in the original timeline
- **Video duration**: How long the original video was
- **Frame rate**: The original playback speed
-
-By passing this metadata, the model can better understand the temporal distribution of the sampled frames and whether important moments might have been skipped.
-
 #### Custom RGBA Background Color

 To use a custom background color for RGBA images, pass the `rgba_background_color` parameter via `--media-io-kwargs`:
@@ -958,7 +894,7 @@ Alternatively, you can pass `audio_url`, which is the audio counterpart of `imag
    print("Chat completion output from audio url:", result)
    ```

-Full example: [examples/generate/multimodal/openai_chat_completion_client_for_multimodal.py](../../examples/generate/multimodal/openai_chat_completion_client_for_multimodal.py)
+Full example: [examples/online_serving/openai_chat_completion_client_for_multimodal.py](../../examples/online_serving/openai_chat_completion_client_for_multimodal.py)

 !!! note
    By default, the timeout for fetching audios through HTTP URL is `10` seconds.
@@ -20,7 +20,6 @@ The following are the supported quantization formats for vLLM:
 - [AMD Quark](quark.md)
 - [Quantized KV Cache](quantized_kvcache.md)
 - [TorchAO](torchao.md)
- [FP8 ViT Encoder Attention](fp8_vit_attn.md)

 ## Supported Hardware

@@ -1,109 +0,0 @@
-# FP8 ViT Encoder Attention
-
-For visual understanding workloads with large images (e.g. QHD, 4K) and relatively
-short text prompts/generation, the ViT encoder attention can become a significant
-bottleneck, especially when the text model is quantized (e.g. NVFP4). vLLM
-supports optional FP8 quantization for the ViT encoder attention via the
-FlashInfer cuDNN backend. Q/K/V are quantized on-the-fly to FP8 before the
-cuDNN attention call.
-
-!!! note
-    - Currently supports Qwen3-VL family models only (`qwen3_vl`, `qwen3_vl_moe`,
-      `qwen3_5`, `qwen3_5_moe`, and other models using Qwen3 ViT).
-    - Dynamic scaling is not compatible with ViT full CUDA graphs.
-    - Performance gains are mostly visible at QHD/4K resolutions or multi-image
-      requests. Smaller images may see no speedup due to quantization overhead
-      (3 quantization kernel launches + un-padding).
-    - FP8 tensor-core speedup is more pronounced on GB300 than GB200.
-
-## Requirements
-
- FlashInfer cuDNN backend with cuDNN >= 9.17.1.
-
-## Usage
-
-Enable FP8 ViT attention by passing `--mm-encoder-attn-dtype fp8` together
-with `--mm-encoder-attn-backend FLASHINFER`:
-
-```bash
-vllm serve $MODEL \
-    --mm-encoder-attn-backend FLASHINFER \
-    --mm-encoder-attn-dtype fp8
-```
-
-By default (no scale file), **dynamic scaling** is used: a 16-entry circular
-buffer of observed Q/K/V amax values drives per-forward scale updates. This
-matches BF16 accuracy without any calibration but adds a small per-forward
-overhead.
-
-## Calibrate-Once, Reuse Workflow (Recommended)
-
-For production, calibrate static scales on a representative dataset once and
-reuse them to avoid the dynamic overhead:
-
-```bash
-# Step 1: calibrate and save scales (runs dynamic scaling for 16 passes,
-# then dumps the learned scales to JSON).
-vllm bench mm-processor \
-    --model $MODEL --mm-encoder-attn-backend FLASHINFER \
-    --mm-encoder-attn-dtype fp8 \
-    --mm-encoder-fp8-scale-save-path /path/to/scales.json \
-    --dataset-name hf --dataset-path lmarena-ai/VisionArena-Chat \
-    --num-prompts 100
-
-# Step 2: serve with static scales (no dynamic overhead).
-vllm serve $MODEL \
-    --mm-encoder-attn-backend FLASHINFER \
-    --mm-encoder-attn-dtype fp8 \
-    --mm-encoder-fp8-scale-path /path/to/scales.json
-```
-
-Saved scales are multiplied by `--mm-encoder-fp8-scale-save-margin` (default
-`1.5`) to leave headroom against activation outliers not present in the
-calibration set. The default has been validated to generalize across datasets
-(e.g. VisionArena-Chat calibration maintains BF16 accuracy on ChartQA).
-
-## Scale File Format
-
-```json
-{
-    "visual.blocks.0.attn.attn": {"q": 224.0, "k": 198.0, "v": 210.0},
-    "visual.blocks.1.attn.attn": {"q": 218.0, "k": 195.0, "v": 207.0}
-}
-```
-
-Keys `q_scale` / `k_scale` / `v_scale` are accepted as aliases.
-
-## Performance
-
-**Core cuDNN attention kernel** (PyTorch profiler, `cudnn_generated_fort_native_sdpa_sm100_flash_fprop`, head_dim=128, seq_len=8192):
-
-| Hardware | BF16 | FP8 | Speedup |
-| -------- | ---- | ---- | ------- |
-| GB200 | 350 us | 312 us | **1.12x** |
-| GB300 | 300 us | 211 us | **1.42x** |
-
-**End-to-end encoder forward time** (Qwen3-VL-30B-A3B-Instruct on GB200, 3 images/request):
-
-| Resolution | BF16 median | FP8 median | Speedup |
-| ---------- | ----------- | ---------- | ------- |
-| HD (720x1280) | 31.77 ms | 36.39 ms | 0.87x |
-| FullHD (1080x1920) | 57.99 ms | 58.73 ms | ~same |
-| QHD (1440x2560) | 131.83 ms | 122.30 ms | **1.08x** |
-| 4K (2160x3840) | 543.44 ms | 460.31 ms | **1.18x** |
-
-Crossover is around FullHD with 3 images/request. At QHD and above, FP8 wins.
-
-## Accuracy
-
-ChartQA, Qwen3-VL-8B-Instruct, 500 samples. FP8 static uses scales calibrated
-on VisionArena-Chat (with default 1.5x margin):
-
-| Metric | BF16 | FP8 dynamic | FP8 static |
-| ------ | ---- | ----------- | ---------- |
-| relaxed_accuracy | 0.780 | 0.776 | 0.780 |
-| anywhere_accuracy | 0.806 | 0.816 | 0.814 |
-| exact_match | 0.584 | 0.582 | 0.578 |
-
-All three configurations match within statistical noise, confirming that
-static scales calibrated on one dataset generalize to another.
@@ -202,7 +202,7 @@ The reasoning content is also available when both tool calling and the reasoning
    print(f"Arguments: {tool_call.arguments}")
    ```

-For more examples, please refer to [examples/reasoning/openai_chat_completion_tool_calls_with_reasoning.py](../../examples/reasoning/openai_chat_completion_tool_calls_with_reasoning.py).
+For more examples, please refer to [examples/online_serving/openai_chat_completion_tool_calls_with_reasoning.py](../../examples/online_serving/openai_chat_completion_tool_calls_with_reasoning.py).

 ## Server-Level Default Chat Template Kwargs

@@ -35,99 +35,6 @@ For reproducible measurements in your environment, use
 [`examples/offline_inference/spec_decode.py`](../../../examples/offline_inference/spec_decode.py)
 or the [benchmark CLI guide](../../benchmarking/cli.md).

-## `--speculative-config` schema
-
-Use `--speculative-config` to pass speculative decoding settings as a JSON
-object on the CLI:
-
-```bash
-vllm serve <target-model> \
-  --speculative-config '{
-    "method": "draft_model",
-    "model": "<draft-model>",
-    "num_speculative_tokens": 5
-  }'
-```
-
-The same keys are accepted from Python via `LLM(..., speculative_config={...})`.
-The tables below highlight common user-facing keys accepted in this JSON
-object; they are not an exhaustive schema reference.
-For more details, see the generated [engine arguments reference](../../configuration/engine_args.md)
-and the API docs for [vllm.config.SpeculativeConfig][].
-
-### Common keys
-
-These keys are commonly used across speculative decoding setups, though some
-only apply to model-based methods such as `draft_model`, `mtp`, `eagle3`, and
-`dflash`.
-
-| Key | Type | Default | Allowed values / meaning |
-| --- | --- | --- | --- |
-| `method` | `string` | `None` | Speculation method. Common values include `draft_model`, `ngram`, `suffix`, `mtp`, `eagle3`, and `dflash`. If omitted, vLLM infers the method from the provided configuration when possible. |
-| `model` | `string` | `None` | Draft model, EAGLE head, or auxiliary model identifier. For `ngram`, `ngram_gpu`, `suffix`, and `mtp`, this can often be omitted. |
-| `num_speculative_tokens` | `integer > 0` | `None` | Number of speculative tokens to propose per step. Required for methods that do not infer it from model metadata. |
-| `draft_tensor_parallel_size` | `integer >= 1` | `None` | Tensor parallel size for the draft model. |
-| `max_model_len` | `integer >= 1` | `None` | Maximum context length for the draft model. |
-| `parallel_drafting` | `boolean` | `false` | Enable parallel draft token generation. Only compatible with EAGLE and draft-model methods. |
-| `rejection_sample_method` | `string` | `strict` | `strict`, `probabilistic`, or `synthetic`. |
-| `synthetic_acceptance_rate` | `float` | `None` | Average acceptance rate to target when `rejection_sample_method` is `synthetic`. Valid range is `[0, 1]`. |
-
-### Method-specific keys
-
-#### N-gram
-
-| Key | Type | Default | Meaning |
-| --- | --- | --- | --- |
-| `prompt_lookup_max` | `integer >= 1` | `5` if both lookup bounds are omitted; otherwise mirrors `prompt_lookup_min` when omitted | Maximum n-gram window size. |
-| `prompt_lookup_min` | `integer >= 1` | `5` if both lookup bounds are omitted; otherwise mirrors `prompt_lookup_max` when omitted | Minimum n-gram window size. |
-
-Example:
-
-```bash
-vllm serve <target-model> \
-  --speculative-config '{
-    "method": "ngram",
-    "num_speculative_tokens": 4,
-    "prompt_lookup_min": 2,
-    "prompt_lookup_max": 5
-  }'
-```
-
-#### Suffix decoding
-
-| Key | Type | Default | Meaning |
-| --- | --- | --- | --- |
-| `suffix_decoding_max_tree_depth` | `integer` | `24` | Maximum combined prefix-match and speculation tree depth. |
-| `suffix_decoding_max_cached_requests` | `integer` | `10000` | Maximum number of requests cached in the global suffix tree. Set `0` to disable the global cache. |
-| `suffix_decoding_max_spec_factor` | `float` | `1.0` | Caps speculative length as a multiple of prefix-match length. |
-| `suffix_decoding_min_token_prob` | `float` | `0.1` | Minimum estimated token probability required to speculate a token. |
-
-Example:
-
-```bash
-vllm serve <target-model> \
-  --speculative-config '{
-    "method": "suffix",
-    "num_speculative_tokens": 8,
-    "suffix_decoding_max_tree_depth": 24,
-    "suffix_decoding_max_cached_requests": 10000,
-    "suffix_decoding_max_spec_factor": 1.0,
-    "suffix_decoding_min_token_prob": 0.1
-  }'
-```
-
-### Notes
-
- `--speculative-config` expects a JSON object on the CLI. In YAML config
-  files, use a nested mapping instead of an escaped JSON string.
- `tensor_parallel_size` is not a valid key in `speculative_config`. Use
-  `draft_tensor_parallel_size` instead.
- Keys such as `temperature` and `top_p` are sampling parameters, not
-  `--speculative-config` fields.
- Internal fields such as `target_model_config`, `draft_model_config`,
-  `target_parallel_config`, `draft_parallel_config`, and `draft_load_config`
-  are populated by vLLM and are not intended to be set by users.
-
 ## Lossless guarantees of Speculative Decoding

 In vLLM, speculative decoding aims to enhance inference efficiency while maintaining accuracy. This section addresses the lossless guarantees of
@@ -33,9 +33,9 @@ vllm serve Qwen/Qwen3-4B-Thinking-2507 \
    --port 8000 \
    --seed 42 \
    -tp 1 \
-    --max-model-len 2048 \
-    --gpu-memory-utilization 0.8 \
-    --speculative-config '{"model": "Qwen/Qwen3-0.6B", "num_speculative_tokens": 5, "method": "draft_model"}'
+    --max_model_len 2048 \
+    --gpu_memory_utilization 0.8 \
+    --speculative_config '{"model": "Qwen/Qwen3-0.6B", "num_speculative_tokens": 5, "method": "draft_model"}'
 ```

 The code used to request as completions as a client remains unchanged:
@@ -77,8 +77,4 @@ The code used to request as completions as a client remains unchanged:
    ```

 !!! warning
-    Note: Please use `--speculative-config` to set all configurations related
-    to speculative decoding. The previous method of specifying the model
-    through `--speculative-model` and adding related parameters such as
-    `--num-speculative-tokens` separately has been deprecated. For supported
-    keys and examples, see the [`--speculative-config` schema](README.md#--speculative-config-schema).
+    Note: Please use `--speculative_config` to set all configurations related to speculative decoding. The previous method of specifying the model through `--speculative_model` and adding related parameters (e.g., `--num_speculative_tokens`) separately has been deprecated.
@@ -38,7 +38,7 @@ for output in outputs:
 ```bash
 vllm serve XiaomiMiMo/MiMo-7B-Base \
    --tensor-parallel-size 1 \
-    --speculative-config '{"method":"mtp","num_speculative_tokens":1}'
+    --speculative_config '{"method":"mtp","num_speculative_tokens":1}'
 ```

 ## Notes
@@ -36,9 +36,9 @@ vllm serve Qwen/Qwen3-4B \
    --port 8000 \
    --seed 42 \
    -tp 1 \
-    --max-model-len 2048 \
-    --gpu-memory-utilization 0.8 \
-    --speculative-config '{"model": "amd/PARD-Qwen3-0.6B", "num_speculative_tokens": 12, "method": "draft_model", "parallel_drafting": true}'
+    --max_model_len 2048 \
+    --gpu_memory_utilization 0.8 \
+    --speculative_config '{"model": "amd/PARD-Qwen3-0.6B", "num_speculative_tokens": 12, "method": "draft_model", "parallel_drafting": true}'
 ```

 ## Pre-trained PARD weights
@@ -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

@@ -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. | ✅︎ | ✅︎ |
@@ -439,7 +437,6 @@ th {
 | `Mamba2ForCausalLM` | Mamba2 | `mistralai/Mamba-Codestral-7B-v0.1`, etc. | | ✅︎ |
 | `MiMoForCausalLM` | MiMo | `XiaomiMiMo/MiMo-7B-RL`, etc. | ✅︎ | ✅︎ |
 | `MiMoV2FlashForCausalLM` | MiMoV2Flash | `XiaomiMiMo/MiMo-V2-Flash`, etc. | | ✅︎ |
-| `MiMoV2ProForCausalLM` | MiMoV2Pro | `XiaomiMiMo/MiMo-V2.5-Pro`, etc. | | ✅︎ |
 | `MiniCPMForCausalLM` | MiniCPM | `openbmb/MiniCPM-2B-sft-bf16`, `openbmb/MiniCPM-2B-dpo-bf16`, `openbmb/MiniCPM-S-1B-sft`, etc. | ✅︎ | ✅︎ |
 | `MiniCPM3ForCausalLM` | MiniCPM3 | `openbmb/MiniCPM3-4B`, etc. | ✅︎ | ✅︎ |
 | `MiniMaxForCausalLM` | MiniMax-Text | `MiniMaxAI/MiniMax-Text-01-hf`, etc. | | |
@@ -475,7 +472,6 @@ th {
 | `Qwen3MoeForCausalLM` | Qwen3MoE | `Qwen/Qwen3-30B-A3B`, etc. | ✅︎ | ✅︎ |
 | `Qwen3NextForCausalLM` | Qwen3NextMoE | `Qwen/Qwen3-Next-80B-A3B-Instruct`, etc. | ✅︎ | ✅︎ |
 | `RWForCausalLM` | Falcon RW | `tiiuae/falcon-40b`, etc. | | ✅︎ |
-| `Rnj1ForCausalLM` | Rnj1 | `EssentialAI/rnj-1-instruct`, etc. | | |
 | `SarvamMoEForCausalLM` | Sarvam 2 | `sarvamai/sarvam2-30b-a3b`, etc. | ✅︎ | ✅︎ |
 | `SarvamMLAForCausalLM` | Sarvam 2 | `sarvamai/sarvam2-105b-a9b`, etc. | | ✅︎ |
 | `SeedOssForCausalLM` | SeedOss | `ByteDance-Seed/Seed-OSS-36B-Instruct`, etc. | ✅︎ | ✅︎ |
@@ -591,7 +587,6 @@ These models primarily accept the [`LLM.generate`](./generative_models.md#llmgen
 | `LlavaNextVideoForConditionalGeneration` | LLaVA-NeXT-Video | T + V | `llava-hf/LLaVA-NeXT-Video-7B-hf`, etc. | | ✅︎ |
 | `LlavaOnevisionForConditionalGeneration` | LLaVA-Onevision | T + I<sup>+</sup> + V<sup>+</sup> | `llava-hf/llava-onevision-qwen2-7b-ov-hf`, `llava-hf/llava-onevision-qwen2-0.5b-ov-hf`, etc. | | ✅︎ |
 | `MiDashengLMModel` | MiDashengLM | T + A<sup>+</sup> | `mispeech/midashenglm-7b` | | ✅︎ |
-| `MiMoV2OmniForCausalLM` | MiMo-V2.5-Omni | T + I<sup>E+</sup> + V<sup>E+</sup> + A<sup>+</sup> | `XiaomiMiMo/MiMo-V2.5-Omni` | | ✅︎ |
 | `MiniCPMO` | MiniCPM-O | T + I<sup>E+</sup> + V<sup>E+</sup> + A<sup>E+</sup> | `openbmb/MiniCPM-o-2_6`, etc. | ✅︎ | ✅︎ |
 | `MiniCPMV` | MiniCPM-V | T + I<sup>E+</sup> + V<sup>E+</sup> | `openbmb/MiniCPM-V-2` (see note), `openbmb/MiniCPM-Llama3-V-2_5`, `openbmb/MiniCPM-V-2_6`, `openbmb/MiniCPM-V-4`, `openbmb/MiniCPM-V-4_5`, etc. | ✅︎ | |
 | `MiniMaxVL01ForConditionalGeneration` | MiniMax-VL | T + I<sup>E+</sup> | `MiniMaxAI/MiniMax-VL-01`, etc. | | ✅︎ |
@@ -646,10 +641,10 @@ 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
@@ -251,7 +251,7 @@ The following extra parameters are supported:
 Our Responses API is compatible with [OpenAI's Responses API](https://platform.openai.com/docs/api-reference/responses);
 you can use the [official OpenAI Python client](https://github.com/openai/openai-python) to interact with it.

-Code example: [examples/online_serving/openai_responses_client_with_tools.py](../../examples/tool_calling/openai_responses_client_with_tools.py)
+Code example: [examples/online_serving/openai_responses_client_with_tools.py](../../examples/online_serving/openai_responses_client_with_tools.py)

 #### Extra parameters

@@ -279,7 +279,7 @@ you can use the [official OpenAI Python client](https://github.com/openai/openai
 !!! note
    To use the Transcriptions API, please install with extra audio dependencies using `pip install vllm[audio]`.

-Code example: [examples/speech_to_text/openai/openai_transcription_client.py](../../examples/speech_to_text/openai/openai_transcription_client.py)
+Code example: [examples/online_serving/openai_transcription_client.py](../../examples/online_serving/openai_transcription_client.py)

 NOTE: beam search is currently supported in the transcriptions endpoint for encoder-decoder multimodal models, e.g., whisper, but highly inefficient as work for handling the encoder/decoder cache is actively ongoing. This is an active point of ongoing optimization and will be handled properly in the very near future.

@@ -397,7 +397,7 @@ Please mind that the popular `openai/whisper-large-v3-turbo` model does not supp
 !!! note
    To use the Translation API, please install with extra audio dependencies using `pip install vllm[audio]`.

-Code example: [examples/speech_to_text/openai/openai_translation_client.py](../../examples/speech_to_text/openai/openai_translation_client.py)
+Code example: [examples/online_serving/openai_translation_client.py](../../examples/online_serving/openai_translation_client.py)

 #### Extra Parameters

@@ -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)


@@ -6,15 +6,15 @@ This folder provides several example scripts on how to inference Qwen2.5-Omni of

 ```bash
 # Audio + image + video
-python examples/generate/multimodal/qwen2_5_omni/only_thinker.py \
+python examples/offline_inference/qwen2_5_omni/only_thinker.py \
    -q mixed_modalities

 # Read vision and audio inputs from a single video file
-python examples/generate/multimodal/qwen2_5_omni/only_thinker.py \
+python examples/offline_inference/qwen2_5_omni/only_thinker.py \
    -q use_audio_in_video

 # Multiple audios
-python examples/generate/multimodal/qwen2_5_omni/only_thinker.py \
+python examples/offline_inference/qwen2_5_omni/only_thinker.py \
    -q multi_audios
 ```

@@ -24,16 +24,16 @@ You can also test Qwen2.5-Omni on a single modality:

 ```bash
 # Process audio inputs
-python examples/generate/multimodal/audio_language_offline.py \
+python examples/offline_inference/audio_language.py \
    --model-type qwen2_5_omni

 # Process image inputs
-python examples/generate/multimodal/vision_language_offline.py \
+python examples/offline_inference/vision_language.py \
    --modality image \
    --model-type qwen2_5_omni

 # Process video inputs
-python examples/generate/multimodal/vision_language_offline.py \
+python examples/offline_inference/vision_language.py \
    --modality video \
    --model-type qwen2_5_omni
 ```
@@ -1402,7 +1402,7 @@ def run_mantis(questions: list[str], modality: str) -> ModelRequestData:
 # MiniCPM-V
 def run_minicpmv_base(questions: list[str], modality: str, model_name):
    assert modality in ["image", "video", "image+video"]
-    # If you want to use `MiniCPM-o-2_6` with audio inputs, check `audio_language_offline.py` # noqa
+    # If you want to use `MiniCPM-o-2_6` with audio inputs, check `audio_language.py` # noqa

    # 2.0
    # The official repo doesn't work yet, so we need to use a fork for now
@@ -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`.
@@ -10,8 +10,9 @@ import uuid

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

 from vllm.distributed.kv_transfer.kv_connector.v1.moriio.moriio_common import (
    MoRIIOConstants,
@@ -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
@@ -170,13 +139,10 @@ async def send_request_to_prefill(
                return await response.json()

            else:
-                error_message = (
-                    f"send_request_to_prefill response ={response},"
-                    f"reason={response.reason}, status={response.status},"
-                    f"method={response.method}, url={response.url},"
-                    f"real_url={response.real_url}"
+                raise RuntimeError(
+                    "send_request_to_prefill response.status != 200response.status = ",
+                    response.status,
                )
-                raise RuntimeError(error_message)


 async def start_decode_request(endpoint, req_data, request_id):
@@ -197,13 +163,9 @@ async def stream_decode_response(session, response, request_id):
            async for chunk_bytes in response.content.iter_chunked(1024):
                yield chunk_bytes
        else:
-            error_message = (
-                f"stream_decode_response response ={response},"
-                f"reason={response.reason}, status={response.status},"
-                f"method={response.method}, url={response.url},"
-                f"real_url={response.real_url}"
+            raise RuntimeError(
+                f"decode response.status != 200, status = {response.status}"
            )
-            raise RuntimeError(error_message)
    finally:
        await session.close()

@@ -213,22 +175,21 @@ def example_round_robin_dp_loader(request_number, dp_size):


@app.route("/v1/completions", methods=["POST"])
-async def handle_completions_request():
-    return await handle_request("/completions", request)
-
-
@app.route("/v1/chat/completions", methods=["POST"])
-async def handle_chat_completions_request():
-    return await handle_request("/chat/completions", request)
-
-
-async def handle_request(api: str, request: Request):
+async def handle_request():
    try:
        with _list_lock:
            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 +215,7 @@ 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())}"

@@ -276,36 +230,40 @@ async def handle_request(api: str, request: Request):
        )
        req_data_to_prefill["kv_transfer_params"]["transfer_id"] = transfer_id

-        prefill_request_url = prefill_instance_endpoint["request_address"] + api
        send_prefill_task = asyncio.create_task(
            send_request_to_prefill(
-                prefill_request_url,
+                prefill_instance_endpoint["request_address"],
                req_data_to_prefill,
                request_id,
+                decode_instance_endpoint,
+                dip,
+                dport,
                selected_prefill_dp_rank,
            )
        )
+        ip, port = extract_ip_port_fast(prefill_instance_endpoint["request_address"])

        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"
@@ -316,10 +274,12 @@ 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(
-            start_decode_request(decode_request_url, req_data, request_id)
+            start_decode_request(
+                decode_instance_endpoint["request_address"], req_data, request_id
+            )
        )

        session, decode_response = await decode_request_task
--- a/Show More
+++ b/Show More
Author	SHA1	Message	Date
yewentao256	b547993038	type ignore Signed-off-by: yewentao256 <zhyanwentao@126.com>	2026-04-21 15:06:50 -04:00
yewentao256	9026d5ec05	assert not None Signed-off-by: yewentao256 <zhyanwentao@126.com>	2026-04-21 14:59:44 -04:00
yewentao256	ee778e2cd8	update batch invariant docstring Signed-off-by: yewentao256 <zhyanwentao@126.com>	2026-04-21 14:47:35 -04:00