[Bugfix][MLA] Change default SM100 MLA prefill backend back to TRT-LLM (#38562)

Signed-off-by: Matthew Bonanni <mbonanni@redhat.com>
(cherry picked from commit 2c734ed0e0)

.buildkite/release-pipeline.yaml

@@ -12,7 +12,7 @@ steps:
         depends_on: ~
         id: build-wheel-arm64-cuda-12-9
         agents:
-          queue: arm64_cpu_queue_postmerge
+          queue: arm64_cpu_queue_release
         commands:
           # #NOTE: torch_cuda_arch_list is derived from upstream PyTorch build files here:
           # https://github.com/pytorch/pytorch/blob/main/.ci/aarch64_linux/aarch64_ci_build.sh#L7
@@ -27,7 +27,7 @@ steps:
         depends_on: ~
         id: build-wheel-arm64-cuda-13-0
         agents:
-          queue: arm64_cpu_queue_postmerge
+          queue: arm64_cpu_queue_release
         commands:
           # #NOTE: torch_cuda_arch_list is derived from upstream PyTorch build files here:
           # https://github.com/pytorch/pytorch/blob/main/.ci/aarch64_linux/aarch64_ci_build.sh#L7
@@ -42,7 +42,7 @@ steps:
         depends_on: ~
         id: build-wheel-arm64-cpu
         agents:
-          queue: arm64_cpu_queue_postmerge
+          queue: arm64_cpu_queue_release
         commands:
           - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg GIT_REPO_CHECK=1 --build-arg VLLM_BUILD_ACL=ON --tag vllm-ci:build-image --target vllm-build --progress plain -f docker/Dockerfile.cpu ."
           - "mkdir artifacts"
@@ -55,7 +55,7 @@ steps:
         depends_on: ~
         id: build-wheel-x86-cuda-12-9
         agents:
-          queue: cpu_queue_postmerge
+          queue: cpu_queue_release
         commands:
           - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=12.9.1 --tag vllm-ci:build-image --target build --progress plain -f docker/Dockerfile ."
           - "mkdir artifacts"
@@ -68,7 +68,7 @@ steps:
         depends_on: ~
         id: build-wheel-x86-cuda-13-0
         agents:
-          queue: cpu_queue_postmerge
+          queue: cpu_queue_release
         commands:
           - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=13.0.1 --build-arg BUILD_BASE_IMAGE=nvidia/cuda:13.0.1-devel-ubuntu22.04 --tag vllm-ci:build-image --target build --progress plain -f docker/Dockerfile ."
           - "mkdir artifacts"
@@ -81,7 +81,7 @@ steps:
         depends_on: ~
         id: build-wheel-x86-cpu
         agents:
-          queue: cpu_queue_postmerge
+          queue: cpu_queue_release
         commands:
           - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg GIT_REPO_CHECK=1 --build-arg VLLM_CPU_X86=true --tag vllm-ci:build-image --target vllm-build --progress plain -f docker/Dockerfile.cpu ."
           - "mkdir artifacts"
@@ -97,7 +97,7 @@ steps:
         depends_on: ~
         id: build-release-image-x86
         agents:
-          queue: cpu_queue_postmerge
+          queue: cpu_queue_release
         commands:
           - "aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin public.ecr.aws/q9t5s3a7"
           - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=12.9.1 --build-arg FLASHINFER_AOT_COMPILE=true --build-arg INSTALL_KV_CONNECTORS=true --tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m) --target vllm-openai --progress plain -f docker/Dockerfile ."
@@ -110,7 +110,7 @@ steps:
         depends_on: ~
         id: build-release-image-arm64
         agents:
-          queue: arm64_cpu_queue_postmerge
+          queue: arm64_cpu_queue_release
         commands:
           - "aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin public.ecr.aws/q9t5s3a7"
           - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=12.9.1 --build-arg FLASHINFER_AOT_COMPILE=true --build-arg torch_cuda_arch_list='8.7 8.9 9.0 10.0+PTX 12.0' --build-arg INSTALL_KV_CONNECTORS=true --tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m) --target vllm-openai --progress plain -f docker/Dockerfile ."
@@ -120,7 +120,7 @@ steps:
         depends_on: ~
         id: build-release-image-x86-cuda-13-0
         agents:
-          queue: cpu_queue_postmerge
+          queue: cpu_queue_release
         commands:
           - "aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin public.ecr.aws/q9t5s3a7"
           - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=13.0.1 --build-arg INSTALL_KV_CONNECTORS=true --build-arg BUILD_BASE_IMAGE=nvidia/cuda:13.0.1-devel-ubuntu22.04 --tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)-cu130 --target vllm-openai --progress plain -f docker/Dockerfile ."
@@ -133,13 +133,57 @@ steps:
         depends_on: ~
         id: build-release-image-arm64-cuda-13-0
         agents:
-          queue: arm64_cpu_queue_postmerge
+          queue: arm64_cpu_queue_release
         commands:
           - "aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin public.ecr.aws/q9t5s3a7"
           # compute capability 12.0 for RTX-50 series / RTX PRO 6000 Blackwell, 12.1 for DGX Spark
           - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=13.0.1 --build-arg torch_cuda_arch_list='8.7 8.9 9.0 10.0+PTX 12.0 12.1' --build-arg INSTALL_KV_CONNECTORS=true --build-arg BUILD_BASE_IMAGE=nvidia/cuda:13.0.1-devel-ubuntu22.04 --tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)-cu130 --target vllm-openai --progress plain -f docker/Dockerfile ."
           - "docker push public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)-cu130"
 
+      - 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=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 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=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 13.0 - Ubuntu 24.04"
+        depends_on: ~
+        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=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 13.0 - Ubuntu 24.04"
+        depends_on: ~
+        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=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
         depends_on: ~
@@ -148,8 +192,9 @@ steps:
         depends_on:
           - block-cpu-release-image-build
           - input-release-version
+        id: build-release-image-x86-cpu
         agents:
-          queue: cpu_queue_postmerge
+          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 GIT_REPO_CHECK=1 --build-arg VLLM_CPU_X86=true --tag public.ecr.aws/q9t5s3a7/vllm-cpu-release-repo:$(buildkite-agent meta-data get release-version) --tag public.ecr.aws/q9t5s3a7/vllm-cpu-release-repo:latest --progress plain --target vllm-openai -f docker/Dockerfile.cpu ."
@@ -163,11 +208,12 @@ steps:
         depends_on: ~
 
       - label: "Build release image - arm64 - CPU"
-        depends_on: 
+        depends_on:
           - block-arm64-cpu-release-image-build
           - input-release-version
+        id: build-release-image-arm64-cpu
         agents:
-          queue: arm64_cpu_queue_postmerge
+          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 GIT_REPO_CHECK=1 --tag public.ecr.aws/q9t5s3a7/vllm-arm64-cpu-release-repo:$(buildkite-agent meta-data get release-version) --tag public.ecr.aws/q9t5s3a7/vllm-arm64-cpu-release-repo:latest --progress plain --target vllm-openai -f docker/Dockerfile.cpu ."
@@ -185,7 +231,7 @@ steps:
           - build-release-image-arm64
         id: create-multi-arch-manifest
         agents:
-          queue: small_cpu_queue_postmerge
+          queue: small_cpu_queue_release
         commands:
           - "aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin public.ecr.aws/q9t5s3a7"
           - "docker manifest create public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-x86_64 public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-aarch64 --amend"
@@ -196,7 +242,7 @@ steps:
           - create-multi-arch-manifest
         id: annotate-release-workflow
         agents:
-          queue: small_cpu_queue_postmerge
+          queue: small_cpu_queue_release
         commands:
           - "bash .buildkite/scripts/annotate-release.sh"
 
@@ -206,18 +252,67 @@ steps:
           - build-release-image-arm64-cuda-13-0
         id: create-multi-arch-manifest-cuda-13-0
         agents:
-          queue: small_cpu_queue_postmerge
+          queue: small_cpu_queue_release
         commands:
           - "aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin public.ecr.aws/q9t5s3a7"
           - "docker manifest create public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-cu130 public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-x86_64-cu130 public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-aarch64-cu130 --amend"
           - "docker manifest push public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-cu130"
 
+      - label: "Create multi-arch manifest - CUDA 12.9 - Ubuntu 24.04"
+        depends_on:
+          - build-release-image-x86-ubuntu2404
+          - build-release-image-arm64-ubuntu2404
+        id: create-multi-arch-manifest-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:$BUILDKITE_COMMIT-ubuntu2404 public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-x86_64-ubuntu2404 public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-aarch64-ubuntu2404 --amend"
+          - "docker manifest push public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-ubuntu2404"
+
+      - label: "Create multi-arch manifest - CUDA 13.0 - Ubuntu 24.04"
+        depends_on:
+          - build-release-image-x86-cuda-13-0-ubuntu2404
+          - build-release-image-arm64-cuda-13-0-ubuntu2404
+        id: create-multi-arch-manifest-cuda-13-0-ubuntu2404
+        agents:
+          queue: small_cpu_queue_release
+        commands:
+          - "aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin public.ecr.aws/q9t5s3a7"
+          - "docker manifest create public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-cu130-ubuntu2404 public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-x86_64-cu130-ubuntu2404 public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-aarch64-cu130-ubuntu2404 --amend"
+          - "docker manifest push public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-cu130-ubuntu2404"
+
+      - block: "Confirm publishing release images to DockerHub"
+        key: block-publish-release-images-dockerhub
+        depends_on:
+          - create-multi-arch-manifest
+          - create-multi-arch-manifest-cuda-13-0
+          - build-release-image-x86-cpu
+          - build-release-image-arm64-cpu
+          - build-rocm-release-image
+
+      - label: "Publish release images to DockerHub"
+        key: publish-release-images-dockerhub
+        depends_on:
+          - block-publish-release-images-dockerhub
+        agents:
+          queue: small_cpu_queue_release
+        commands:
+          - "bash .buildkite/scripts/push-release-builds.sh"
+        plugins:
+          - docker-login#v3.0.0:
+              username: vllmbot
+              password-env: DOCKERHUB_TOKEN
+        env:
+          DOCKER_BUILDKIT: "1"
+          DOCKERHUB_USERNAME: "vllmbot"
+
       - label: "Publish nightly multi-arch image to DockerHub"
         depends_on:
           - create-multi-arch-manifest
         if: build.env("NIGHTLY") == "1"
         agents:
-          queue: small_cpu_queue_postmerge
+          queue: small_cpu_queue_release
         commands:
           - "bash .buildkite/scripts/push-nightly-builds.sh"
           # Clean up old nightly builds (keep only last 14)
@@ -235,7 +330,7 @@ steps:
           - create-multi-arch-manifest-cuda-13-0
         if: build.env("NIGHTLY") == "1"
         agents:
-          queue: small_cpu_queue_postmerge
+          queue: small_cpu_queue_release
         commands:
           - "bash .buildkite/scripts/push-nightly-builds.sh cu130"
           # Clean up old nightly builds (keep only last 14)
@@ -262,7 +357,7 @@ steps:
           - block-upload-release-wheels
         id: upload-release-wheels
         agents:
-          queue: small_cpu_queue_postmerge
+          queue: small_cpu_queue_release
         commands:
           - "bash .buildkite/scripts/upload-release-wheels-pypi.sh"
 
@@ -323,7 +418,7 @@ steps:
       - step: input-rocm-config
         allow_failure: true  # Allow failure so non-UI builds can proceed (input step is skipped)
     agents:
-      queue: cpu_queue_postmerge
+      queue: cpu_queue_release
     commands:
       # Set configuration and check cache
       - |
@@ -465,7 +560,7 @@ steps:
       - step: build-rocm-base-wheels
         allow_failure: false
     agents:
-      queue: cpu_queue_postmerge
+      queue: cpu_queue_release
     timeout_in_minutes: 180
     commands:
       # Download artifacts and prepare Docker image
@@ -575,7 +670,7 @@ steps:
       - step: build-rocm-vllm-wheel
         allow_failure: false
     agents:
-      queue: cpu_queue_postmerge
+      queue: cpu_queue_release
     timeout_in_minutes: 60
     commands:
       # Download all wheel artifacts and run upload
@@ -624,7 +719,7 @@ steps:
       - step: input-release-version
         allow_failure: true
     agents:
-      queue: cpu_queue_postmerge
+      queue: cpu_queue_release
     commands:
       - "bash .buildkite/scripts/annotate-rocm-release.sh"
     env:
@@ -641,7 +736,7 @@ steps:
     depends_on: block-generate-root-index-rocm-wheels
     id: generate-root-index-rocm-wheels
     agents:
-      queue: cpu_queue_postmerge
+      queue: cpu_queue_release
     commands:
       - "bash tools/vllm-rocm/generate-rocm-wheels-root-index.sh"
     env:
@@ -655,7 +750,7 @@ steps:
       - step: build-rocm-base-wheels
         allow_failure: false
     agents:
-      queue: cpu_queue_postmerge
+      queue: cpu_queue_release
     timeout_in_minutes: 60
     commands:
       - |

.buildkite/scripts/push-release-builds.sh

@@ -0,0 +1,113 @@
+#!/bin/bash
+
+set -euo pipefail
+
+# Ensure git tags are up-to-date (Buildkite's default fetch doesn't always include tags)
+echo "Fetching latest tags from origin..."
+git fetch --tags --force origin
+
+# Derive release version from the git tag on the current commit.
+# The pipeline must be triggered on a tagged commit (e.g. v0.18.1).
+RELEASE_VERSION=$(git describe --exact-match --tags "${BUILDKITE_COMMIT}" 2>/dev/null || true)
+if [ -z "${RELEASE_VERSION}" ]; then
+    echo "[FATAL] Commit ${BUILDKITE_COMMIT} has no exact git tag. " \
+         "Release images must be published from a tagged commit."
+    exit 1
+fi
+
+# Strip leading 'v' for use in Docker tags (e.g. v0.18.1 -> 0.18.1)
+PURE_VERSION="${RELEASE_VERSION#v}"
+
+echo "========================================"
+echo "Publishing release images"
+echo "  Commit:          ${BUILDKITE_COMMIT}"
+echo "  Release version: ${RELEASE_VERSION}"
+echo "========================================"
+
+set -x
+
+# ---- CUDA (default, CUDA 12.9) ----
+docker pull "public.ecr.aws/q9t5s3a7/vllm-release-repo:${BUILDKITE_COMMIT}-x86_64"
+docker pull "public.ecr.aws/q9t5s3a7/vllm-release-repo:${BUILDKITE_COMMIT}-aarch64"
+
+docker tag "public.ecr.aws/q9t5s3a7/vllm-release-repo:${BUILDKITE_COMMIT}-x86_64" "vllm/vllm-openai:latest-x86_64"
+docker tag "public.ecr.aws/q9t5s3a7/vllm-release-repo:${BUILDKITE_COMMIT}-x86_64" "vllm/vllm-openai:v${PURE_VERSION}-x86_64"
+docker push "vllm/vllm-openai:latest-x86_64"
+docker push "vllm/vllm-openai:v${PURE_VERSION}-x86_64"
+
+docker tag "public.ecr.aws/q9t5s3a7/vllm-release-repo:${BUILDKITE_COMMIT}-aarch64" "vllm/vllm-openai:latest-aarch64"
+docker tag "public.ecr.aws/q9t5s3a7/vllm-release-repo:${BUILDKITE_COMMIT}-aarch64" "vllm/vllm-openai:v${PURE_VERSION}-aarch64"
+docker push "vllm/vllm-openai:latest-aarch64"
+docker push "vllm/vllm-openai:v${PURE_VERSION}-aarch64"
+
+docker manifest rm "vllm/vllm-openai:latest" || true
+docker manifest create "vllm/vllm-openai:latest" "vllm/vllm-openai:latest-x86_64" "vllm/vllm-openai:latest-aarch64"
+docker manifest push "vllm/vllm-openai:latest"
+
+docker manifest rm "vllm/vllm-openai:v${PURE_VERSION}" || true
+docker manifest create "vllm/vllm-openai:v${PURE_VERSION}" "vllm/vllm-openai:v${PURE_VERSION}-x86_64" "vllm/vllm-openai:v${PURE_VERSION}-aarch64"
+docker manifest push "vllm/vllm-openai:v${PURE_VERSION}"
+
+# ---- CUDA 13.0 ----
+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 tag "public.ecr.aws/q9t5s3a7/vllm-release-repo:${BUILDKITE_COMMIT}-x86_64-cu130" "vllm/vllm-openai:latest-x86_64-cu130"
+docker tag "public.ecr.aws/q9t5s3a7/vllm-release-repo:${BUILDKITE_COMMIT}-x86_64-cu130" "vllm/vllm-openai:v${PURE_VERSION}-x86_64-cu130"
+docker push "vllm/vllm-openai:latest-x86_64-cu130"
+docker push "vllm/vllm-openai:v${PURE_VERSION}-x86_64-cu130"
+
+docker tag "public.ecr.aws/q9t5s3a7/vllm-release-repo:${BUILDKITE_COMMIT}-aarch64-cu130" "vllm/vllm-openai:latest-aarch64-cu130"
+docker tag "public.ecr.aws/q9t5s3a7/vllm-release-repo:${BUILDKITE_COMMIT}-aarch64-cu130" "vllm/vllm-openai:v${PURE_VERSION}-aarch64-cu130"
+docker push "vllm/vllm-openai:latest-aarch64-cu130"
+docker push "vllm/vllm-openai:v${PURE_VERSION}-aarch64-cu130"
+
+docker manifest rm "vllm/vllm-openai:latest-cu130" || true
+docker manifest create "vllm/vllm-openai:latest-cu130" "vllm/vllm-openai:latest-x86_64-cu130" "vllm/vllm-openai:latest-aarch64-cu130"
+docker manifest push "vllm/vllm-openai:latest-cu130"
+
+docker manifest rm "vllm/vllm-openai:v${PURE_VERSION}-cu130" || true
+docker manifest create "vllm/vllm-openai:v${PURE_VERSION}-cu130" "vllm/vllm-openai:v${PURE_VERSION}-x86_64-cu130" "vllm/vllm-openai:v${PURE_VERSION}-aarch64-cu130"
+docker manifest push "vllm/vllm-openai:v${PURE_VERSION}-cu130"
+
+# ---- ROCm ----
+docker pull "public.ecr.aws/q9t5s3a7/vllm-release-repo:${BUILDKITE_COMMIT}-rocm"
+docker pull "public.ecr.aws/q9t5s3a7/vllm-release-repo:${BUILDKITE_COMMIT}-rocm-base"
+
+docker tag "public.ecr.aws/q9t5s3a7/vllm-release-repo:${BUILDKITE_COMMIT}-rocm" "vllm/vllm-openai-rocm:latest"
+docker tag "public.ecr.aws/q9t5s3a7/vllm-release-repo:${BUILDKITE_COMMIT}-rocm" "vllm/vllm-openai-rocm:v${PURE_VERSION}"
+docker push "vllm/vllm-openai-rocm:latest"
+docker push "vllm/vllm-openai-rocm:v${PURE_VERSION}"
+
+docker tag "public.ecr.aws/q9t5s3a7/vllm-release-repo:${BUILDKITE_COMMIT}-rocm-base" "vllm/vllm-openai-rocm:latest-base"
+docker tag "public.ecr.aws/q9t5s3a7/vllm-release-repo:${BUILDKITE_COMMIT}-rocm-base" "vllm/vllm-openai-rocm:v${PURE_VERSION}-base"
+docker push "vllm/vllm-openai-rocm:latest-base"
+docker push "vllm/vllm-openai-rocm:v${PURE_VERSION}-base"
+
+# ---- CPU ----
+# CPU images in ECR are tagged with the full version including 'v' (e.g. v0.18.1),
+# matching the value from the Buildkite release-version metadata input.
+docker pull "public.ecr.aws/q9t5s3a7/vllm-cpu-release-repo:${RELEASE_VERSION}"
+docker pull "public.ecr.aws/q9t5s3a7/vllm-arm64-cpu-release-repo:${RELEASE_VERSION}"
+
+docker tag "public.ecr.aws/q9t5s3a7/vllm-cpu-release-repo:${RELEASE_VERSION}" "vllm/vllm-openai-cpu:latest-x86_64"
+docker tag "public.ecr.aws/q9t5s3a7/vllm-cpu-release-repo:${RELEASE_VERSION}" "vllm/vllm-openai-cpu:v${PURE_VERSION}-x86_64"
+docker push "vllm/vllm-openai-cpu:latest-x86_64"
+docker push "vllm/vllm-openai-cpu:v${PURE_VERSION}-x86_64"
+
+docker tag "public.ecr.aws/q9t5s3a7/vllm-arm64-cpu-release-repo:${RELEASE_VERSION}" "vllm/vllm-openai-cpu:latest-arm64"
+docker tag "public.ecr.aws/q9t5s3a7/vllm-arm64-cpu-release-repo:${RELEASE_VERSION}" "vllm/vllm-openai-cpu:v${PURE_VERSION}-arm64"
+docker push "vllm/vllm-openai-cpu:latest-arm64"
+docker push "vllm/vllm-openai-cpu:v${PURE_VERSION}-arm64"
+
+docker manifest rm "vllm/vllm-openai-cpu:latest" || true
+docker manifest create "vllm/vllm-openai-cpu:latest" "vllm/vllm-openai-cpu:latest-x86_64" "vllm/vllm-openai-cpu:latest-arm64"
+docker manifest push "vllm/vllm-openai-cpu:latest"
+
+docker manifest rm "vllm/vllm-openai-cpu:v${PURE_VERSION}" || true
+docker manifest create "vllm/vllm-openai-cpu:v${PURE_VERSION}" "vllm/vllm-openai-cpu:v${PURE_VERSION}-x86_64" "vllm/vllm-openai-cpu:v${PURE_VERSION}-arm64"
+docker manifest push "vllm/vllm-openai-cpu:v${PURE_VERSION}"
+
+echo "========================================"
+echo "Successfully published release images for ${RELEASE_VERSION}"
+echo "========================================"

.buildkite/test_areas/lm_eval.yaml

@@ -45,6 +45,22 @@ steps:
   commands:
   - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-blackwell.txt
 
+- label: LM Eval Qwen3.5 Models (B200)
+  timeout_in_minutes: 120
+  device: b200
+  optional: true
+  num_devices: 2
+  source_file_dependencies:
+  - vllm/model_executor/models/qwen3_5.py
+  - vllm/model_executor/models/qwen3_5_mtp.py
+  - vllm/transformers_utils/configs/qwen3_5.py
+  - vllm/transformers_utils/configs/qwen3_5_moe.py
+  - vllm/model_executor/models/qwen3_next.py
+  - vllm/model_executor/models/qwen3_next_mtp.py
+  - vllm/model_executor/layers/fla/ops/
+  commands:
+    - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-qwen35-blackwell.txt
+
 - label: LM Eval Large Models (H200)
   timeout_in_minutes: 60
   device: h200

docker/Dockerfile

@@ -24,6 +24,7 @@
 
 ARG CUDA_VERSION=12.9.1
 ARG PYTHON_VERSION=3.12
+ARG UBUNTU_VERSION=22.04
 
 # By parameterizing the base images, we allow third-party to use their own
 # base images. One use case is hermetic builds with base images stored in
@@ -38,7 +39,7 @@ ARG PYTHON_VERSION=3.12
 # version are not backwards compatible with OSes that use an earlier version.
 ARG BUILD_BASE_IMAGE=nvidia/cuda:${CUDA_VERSION}-devel-ubuntu20.04
 # Using cuda base image with minimal dependencies necessary for JIT compilation (FlashInfer, DeepGEMM, EP kernels)
-ARG FINAL_BASE_IMAGE=nvidia/cuda:${CUDA_VERSION}-base-ubuntu22.04
+ARG FINAL_BASE_IMAGE=nvidia/cuda:${CUDA_VERSION}-base-ubuntu${UBUNTU_VERSION}
 
 # By parameterizing the Deadsnakes repository URL, we allow third-party to use
 # their own mirror. When doing so, we don't benefit from the transparent
@@ -111,6 +112,10 @@ RUN apt-get update -y \
         gcc-10 \
         g++-10 \
     && update-alternatives --install /usr/bin/gcc gcc /usr/bin/gcc-10 110 --slave /usr/bin/g++ g++ /usr/bin/g++-10 \
+    # Install python dev headers if available (needed for cmake FindPython on Ubuntu 24.04
+    # which ships cmake 3.28 and requires Development.SABIModule; silently skipped on
+    # Ubuntu 20.04/22.04 where python3.x-dev is not available without a PPA)
+    && (apt-get install -y --no-install-recommends python${PYTHON_VERSION}-dev 2>/dev/null || true) \
     && rm -rf /var/lib/apt/lists/* \
     && curl -LsSf https://astral.sh/uv/install.sh | sh \
     && $HOME/.local/bin/uv venv /opt/venv --python ${PYTHON_VERSION} \
@@ -507,7 +512,6 @@ RUN apt-get update -y \
         software-properties-common \
         curl \
         sudo \
-        python3-pip \
         ffmpeg \
         libsm6 \
         libxext6 \
@@ -535,6 +539,7 @@ RUN apt-get update -y \
     && update-alternatives --install /usr/bin/python3 python3 /usr/bin/python${PYTHON_VERSION} 1 \
     && update-alternatives --set python3 /usr/bin/python${PYTHON_VERSION} \
     && ln -sf /usr/bin/python${PYTHON_VERSION}-config /usr/bin/python3-config \
+    && rm -f /usr/lib/python${PYTHON_VERSION}/EXTERNALLY-MANAGED \
     && curl -sS ${GET_PIP_URL} | python${PYTHON_VERSION} \
     && python3 --version && python3 -m pip --version
 
@@ -593,6 +598,25 @@ RUN --mount=type=cache,target=/root/.cache/uv \
         --extra-index-url https://flashinfer.ai/whl/cu$(echo $CUDA_VERSION | cut -d. -f1,2 | tr -d '.') \
     && flashinfer show-config
 
+# Pre-download FlashInfer TRTLLM BMM headers for air-gapped environments.
+# At runtime, MoE JIT compilation downloads these from edge.urm.nvidia.com
+# which fails without internet. This step caches them at build time.
+RUN python3 <<'PYEOF'
+from flashinfer.jit import env as jit_env
+from flashinfer.jit.cubin_loader import download_trtllm_headers, get_cubin
+from flashinfer.artifacts import ArtifactPath, CheckSumHash
+
+download_trtllm_headers(
+    'bmm',
+    jit_env.FLASHINFER_CUBIN_DIR / 'flashinfer' / 'trtllm' / 'batched_gemm' / 'trtllmGen_bmm_export',
+    f'{ArtifactPath.TRTLLM_GEN_BMM}/include/trtllmGen_bmm_export',
+    ArtifactPath.TRTLLM_GEN_BMM,
+    get_cubin(f'{ArtifactPath.TRTLLM_GEN_BMM}/checksums.txt', CheckSumHash.TRTLLM_GEN_BMM),
+)
+
+print('FlashInfer TRTLLM BMM headers downloaded successfully')
+PYEOF
+
 # ============================================================
 # OPENAI API SERVER DEPENDENCIES
 # Pre-install these to avoid reinstalling on every vLLM wheel rebuild

docker/docker-bake.hcl

@@ -33,6 +33,10 @@ group "default" {
   targets = ["openai"]
 }
 
+group "all" {
+  targets = ["openai", "openai-ubuntu2404"]
+}
+
 # Base targets
 
 target "_common" {
@@ -74,3 +78,29 @@ target "openai" {
   tags     = ["vllm:openai"]
   output   = ["type=docker"]
 }
+
+# Ubuntu 24.04 targets
+
+target "test-ubuntu2404" {
+  inherits = ["_common", "_labels"]
+  target   = "test"
+  tags     = ["vllm:test-ubuntu24.04"]
+  args = {
+    UBUNTU_VERSION          = "24.04"
+    GDRCOPY_OS_VERSION      = "Ubuntu24_04"
+    FLASHINFER_AOT_COMPILE  = "true"
+  }
+  output = ["type=docker"]
+}
+
+target "openai-ubuntu2404" {
+  inherits = ["_common", "_labels"]
+  target   = "vllm-openai"
+  tags     = ["vllm:openai-ubuntu24.04"]
+  args = {
+    UBUNTU_VERSION          = "24.04"
+    GDRCOPY_OS_VERSION      = "Ubuntu24_04"
+    FLASHINFER_AOT_COMPILE  = "true"
+  }
+  output = ["type=docker"]
+}

docker/versions.json

@@ -7,6 +7,9 @@
     "PYTHON_VERSION": {
       "default": "3.12"
     },
+    "UBUNTU_VERSION": {
+      "default": "22.04"
+    },
     "BUILD_BASE_IMAGE": {
       "default": "nvidia/cuda:12.9.1-devel-ubuntu20.04"
     },

tests/evals/gsm8k/configs/Qwen3.5-35B-A3B-DEP2.yaml

@@ -0,0 +1,9 @@
+model_name: "Qwen/Qwen3.5-35B-A3B"
+accuracy_threshold: 0.84
+tolerance: 0.03
+num_questions: 1319
+num_fewshot: 5
+server_args: >-
+  --max-model-len 4096
+  --data-parallel-size 2
+  --enable-expert-parallel

tests/evals/gsm8k/configs/Qwen3.5-35B-A3B-FP8-DEP2.yaml

@@ -0,0 +1,10 @@
+model_name: "Qwen/Qwen3.5-35B-A3B-FP8"
+accuracy_threshold: 0.79
+tolerance: 0.03
+num_questions: 1319
+num_fewshot: 5
+server_args: >-
+  --max-model-len 4096
+  --data-parallel-size 2
+  --enable-expert-parallel
+  --kv-cache-dtype fp8

tests/evals/gsm8k/configs/Qwen3.5-397B-A17B-NVFP4-DEP2.yaml

@@ -0,0 +1,9 @@
+model_name: "nvidia/Qwen3.5-397B-A17B-NVFP4"
+accuracy_threshold: 0.88
+tolerance: 0.03
+num_questions: 1319
+num_fewshot: 5
+server_args: >-
+  --max-model-len 4096
+  --data-parallel-size 2
+  --enable-expert-parallel

tests/evals/gsm8k/configs/models-qwen35-blackwell.txt

@@ -0,0 +1,3 @@
+Qwen3.5-35B-A3B-DEP2.yaml
+Qwen3.5-35B-A3B-FP8-DEP2.yaml
+Qwen3.5-397B-A17B-NVFP4-DEP2.yaml

tests/evals/gsm8k/test_gsm8k_correctness.py

@@ -19,8 +19,6 @@ from vllm.platforms import current_platform
 
 from .gsm8k_eval import evaluate_gsm8k
 
-TOL = 0.08  # Absolute tolerance for accuracy comparison
-
 
 def run_gsm8k_eval(eval_config: dict, server_url: str) -> dict:
     """Run GSM8K evaluation using our isolated script."""
@@ -99,20 +97,20 @@ def test_gsm8k_correctness(config_filename):
 
         measured_metric = results["accuracy"]
         expected_metric = eval_config["accuracy_threshold"]
+        tol = eval_config.get("tolerance", 0.08)
 
         print(f"GSM8K Results for {eval_config['model_name']}:")
         print(f"  Measured metric: {measured_metric:.4f}")
         print(f"  Expected metric: {expected_metric:.4f}")
-        print(f"  Tolerance: {TOL:.4f}")
+        print(f"  Tolerance: {tol:.4f}")
         print(f"  Questions: {results['num_questions']}")
         print(f"  Invalid rate: {results['invalid_rate']:.3f}")
         print(f"  Latency: {results['latency']:.1f}s")
         print(f"  QPS: {results['questions_per_second']:.1f}")
 
-        # Verify metric is within tolerance
-        assert measured_metric >= expected_metric - TOL, (
+        assert measured_metric >= expected_metric - tol, (
             f"GSM8K metric too low: {measured_metric:.4f} < "
-            f"{expected_metric:.4f} - {TOL:.4f} = {expected_metric - TOL:.4f}"
+            f"{expected_metric:.4f} - {tol:.4f} = {expected_metric - tol:.4f}"
         )
 
         print(f"✅ GSM8K test passed for {eval_config['model_name']}")

tests/models/multimodal/generation/test_common.py

@@ -777,6 +777,7 @@ VLM_TEST_SETTINGS = {
         max_model_len=8192,
         max_num_seqs=2,
         auto_cls=AutoModelForCausalLM,
+        patch_hf_runner=model_utils.paddleocr_vl_patch_hf_runner,
         image_size_factors=[(0.25,)],
         marks=[
             pytest.mark.skipif(

tests/models/multimodal/generation/vlm_utils/model_utils.py

@@ -24,6 +24,7 @@ from transformers import (
     GenerationConfig,
     GenerationMixin,
 )
+from transformers.masking_utils import create_causal_mask
 from transformers.video_utils import VideoMetadata
 
 from vllm.logprobs import SampleLogprobs
@@ -679,10 +680,14 @@ def isaac_patch_hf_runner(hf_model: HfRunner) -> HfRunner:
         sin = sin.to(inputs_embeds.dtype)
 
         # Prepare attention mask
-        if attention_mask is not None:
-            attention_mask = self._update_causal_mask(
-                attention_mask, inputs_embeds, cache_position, past_key_values, False
-            )
+        attention_mask = create_causal_mask(
+            config=self.config,
+            input_embeds=inputs_embeds,
+            attention_mask=attention_mask,
+            past_key_values=past_key_values,
+            position_ids=position_ids,
+            cache_position=cache_position,
+        )
 
         # Initialize and collect hidden states
         hidden_states = inputs_embeds
@@ -1149,6 +1154,31 @@ def ovis2_5_patch_hf_runner(hf_model: HfRunner) -> HfRunner:
     return hf_model
 
 
+def paddleocr_vl_patch_hf_runner(hf_model: HfRunner) -> HfRunner:
+    """Patches the HfRunner to fix create_causal_mask API mismatch.
+
+    The PaddleOCR-VL HF model passes `inputs_embeds` to create_causal_mask,
+    but transformers renamed this parameter to `input_embeds`.
+    """
+    import sys
+
+    model_module = sys.modules.get(type(hf_model.model.model).__module__)
+    if model_module is None:
+        return hf_model
+
+    original_create_causal_mask = getattr(model_module, "create_causal_mask", None)
+    if original_create_causal_mask is None:
+        return hf_model
+
+    def patched_create_causal_mask(*args, **kwargs):
+        if "inputs_embeds" in kwargs:
+            kwargs["input_embeds"] = kwargs.pop("inputs_embeds")
+        return original_create_causal_mask(*args, **kwargs)
+
+    model_module.create_causal_mask = patched_create_causal_mask  # type: ignore[attr-defined]
+    return hf_model
+
+
 def qwen2_5_omni_patch_hf_runner(hf_model: HfRunner) -> HfRunner:
     """Patches and returns an instance of the HfRunner for Qwen2.5-Omni."""
     thinker = hf_model.model.thinker

tests/models/registry.py

@@ -780,6 +780,7 @@ _MULTIMODAL_EXAMPLE_MODELS = {
     "Ernie4_5_VLMoeForConditionalGeneration": _HfExamplesInfo(
         "baidu/ERNIE-4.5-VL-28B-A3B-PT",
         trust_remote_code=True,
+        revision="refs/pr/17",
     ),
     "FireRedASR2ForConditionalGeneration": _HfExamplesInfo(
         "allendou/FireRedASR2-LLM-vllm",

vllm/compilation/compiler_interface.py

@@ -373,8 +373,15 @@ class InductorStandaloneAdaptor(CompilerInterface):
                 break
 
         if input_fake_mode is not None:
-            fake_mode_ctx: Any = patch(
-                "torch._inductor.standalone_compile.FakeTensorMode",
+            # Use patch.object on the actual module from sys.modules
+            # because in Python <=3.10 the string-based patch() resolves
+            # torch._inductor.standalone_compile to the wrapper function
+            # (defined in __init__.py) instead of the module.
+            import sys
+
+            fake_mode_ctx: Any = patch.object(
+                sys.modules["torch._inductor.standalone_compile"],
+                "FakeTensorMode",
                 lambda *a, **kw: input_fake_mode,
             )
         else:

vllm/config/attention.py

@@ -30,7 +30,7 @@ class AttentionConfig:
     use_cudnn_prefill: bool = False
     """Whether to use cudnn prefill."""
 
-    use_trtllm_ragged_deepseek_prefill: bool = False
+    use_trtllm_ragged_deepseek_prefill: bool = True
     """Whether to use TRTLLM ragged deepseek prefill."""
 
     use_trtllm_attention: bool | None = None

vllm/config/vllm.py

@@ -682,6 +682,27 @@ class VllmConfig:
                 self.model_config, self.load_config
             )
 
+        if (
+            self.quant_config is not None
+            and self.model_config is not None
+            and hasattr(self.quant_config, "use_deep_gemm")
+            and self.quant_config.use_deep_gemm is None
+        ):
+            from vllm.utils.deep_gemm import should_auto_disable_deep_gemm
+
+            model_type = getattr(self.model_config.hf_text_config, "model_type", None)
+            if should_auto_disable_deep_gemm(model_type):
+                self.quant_config.use_deep_gemm = False
+                logger.warning_once(
+                    "Auto-disabled DeepGemm for model_type=%s on Blackwell. "
+                    "DeepGemm E8M0 scale format causes accuracy degradation "
+                    "for this architecture. Falling back to CUTLASS. "
+                    "To disable DeepGemm globally, set VLLM_USE_DEEP_GEMM=0.",
+                    model_type,
+                )
+
+        from vllm.v1.executor.abstract import Executor
+
         executor_backend = self.parallel_config.distributed_executor_backend
         executor_supports_async_sched = executor_backend in (
             "mp",

vllm/model_executor/layers/fused_moe/experts/trtllm_fp8_moe.py

@@ -253,23 +253,25 @@ class TrtLlmFp8ExpertsMonolithic(TrtLlmFp8ExpertsBase, mk.FusedMoEExpertsMonolit
         weight_key: QuantKey | None,
         activation_key: QuantKey | None,
     ) -> bool:
-        """Monolithic kernels need to express router support."""
+        """Monolithic kernels need to express router support.
+        Renormalize/RenormalizeNaive are excluded: the monolithic kernel's
+        internal routing for these methods produces output uncorrelated
+        with the modular kernel's output and with Triton kernel's output
+        for Qwen3.5-35B-A3B-FP8.
+        See: https://github.com/vllm-project/vllm/issues/37591
+        """
         # NOTE(dbari): TopK routing could also be enabled, but need to validate models
         # NOTE(dbari): Default is not implemented and should not be enabled until it is
         if (weight_key, activation_key) == (kFp8Static128BlockSym, kFp8Dynamic128Sym):
             # NOTE(rob): potentially allow others here. This is a conservative list.
             return routing_method in [
                 RoutingMethodType.DeepSeekV3,
-                RoutingMethodType.Renormalize,
-                RoutingMethodType.RenormalizeNaive,
             ]
         elif (weight_key, activation_key) == (kFp8StaticTensorSym, kFp8StaticTensorSym):
             # NOTE(dbari): as above, potentially allow others here.
             return routing_method in [
                 RoutingMethodType.DeepSeekV3,
                 RoutingMethodType.Llama4,
-                RoutingMethodType.Renormalize,
-                RoutingMethodType.RenormalizeNaive,
             ]
         else:
             raise ValueError("Unsupported quantization scheme.")

vllm/model_executor/layers/quantization/fp8.py

@@ -135,6 +135,7 @@ class Fp8Config(QuantizationConfig):
                     f"{activation_scheme} activation scheme."
                 )
         self.weight_block_size = weight_block_size
+        self.use_deep_gemm: bool | None = None
 
     @classmethod
     def get_name(cls) -> QuantizationMethods:
@@ -291,7 +292,10 @@ class Fp8LinearMethod(LinearMethodBase):
             self.use_marlin = False
 
         self.use_aiter_and_is_supported = rocm_aiter_ops.is_linear_fp8_enabled()
-        self.use_deep_gemm = is_deep_gemm_supported()
+        if self.quant_config.use_deep_gemm is not None:
+            self.use_deep_gemm = self.quant_config.use_deep_gemm
+        else:
+            self.use_deep_gemm = is_deep_gemm_supported()
 
         self.weight_block_size = self.quant_config.weight_block_size
         self.block_quant = self.weight_block_size is not None
@@ -305,6 +309,7 @@ class Fp8LinearMethod(LinearMethodBase):
                 act_quant_group_shape=GroupShape(1, self.weight_block_size[0]),
                 cutlass_block_fp8_supported=self.cutlass_block_fp8_supported,
                 use_aiter_and_is_supported=self.use_aiter_and_is_supported,
+                use_deep_gemm=self.use_deep_gemm,
             )
         else:
             # Use per-token quantization for better perf if dynamic and cutlass
@@ -440,7 +445,7 @@ class Fp8LinearMethod(LinearMethodBase):
             del layer.input_scale
             return
 
-        if self.block_quant:
+        if self.block_quant and self.use_deep_gemm:
             maybe_post_process_fp8_weight_block(layer)
 
     def apply(

vllm/model_executor/layers/quantization/input_quant_fp8.py

@@ -91,6 +91,7 @@ class QuantFP8(CustomOp):
 
         if (
             self.is_group_quant
+            and self.use_ue8m0
             and self.use_deep_gemm_supported
             and (DeepGemmQuantScaleFMT.from_oracle() == DeepGemmQuantScaleFMT.UE8M0)
         ):

vllm/model_executor/layers/quantization/utils/fp8_utils.py

@@ -356,10 +356,14 @@ class W8A8BlockFp8LinearOp:
         act_quant_group_shape: GroupShape,
         cutlass_block_fp8_supported: bool = CUTLASS_BLOCK_FP8_SUPPORTED,
         use_aiter_and_is_supported: bool = False,
+        use_deep_gemm: bool | None = None,
     ):
         self.weight_group_shape = weight_group_shape
         self.act_quant_group_shape = act_quant_group_shape
-        self.is_deep_gemm_supported = is_deep_gemm_supported()
+        if use_deep_gemm is not None:
+            self.is_deep_gemm_supported = use_deep_gemm
+        else:
+            self.is_deep_gemm_supported = is_deep_gemm_supported()
         self.is_hopper = current_platform.is_device_capability(90)
         self.use_deep_gemm_e8m0 = is_deep_gemm_e8m0_used()
         self.is_flashinfer_supported = is_flashinfer_fp8_blockscale_gemm_supported()

vllm/utils/deep_gemm.py

@@ -23,6 +23,24 @@ from vllm.platforms import current_platform
 from vllm.utils.import_utils import has_deep_gemm
 from vllm.utils.math_utils import cdiv
 
+_DEEPGEMM_BLACKWELL_EXCLUDED_MODEL_TYPES: set[str] = {
+    "qwen3_5_text",
+    "qwen3_5_moe_text",
+}
+
+
+def should_auto_disable_deep_gemm(model_type: str | None) -> bool:
+    """Check if DeepGemm should be auto-disabled for this model on Blackwell.
+
+    Returns True if the model is known to have accuracy degradation with
+    DeepGemm's E8M0 scale format on Blackwell GPUs (SM100+).
+    """
+    if model_type is None:
+        return False
+    if not current_platform.is_device_capability_family(100):
+        return False
+    return model_type in _DEEPGEMM_BLACKWELL_EXCLUDED_MODEL_TYPES
+
 
 class DeepGemmQuantScaleFMT(Enum):
     # Float32 scales in Float32 tensor

vllm/v1/attention/backends/mla/cutlass_mla.py

@@ -162,6 +162,11 @@ class CutlassMLAImpl(MLACommonImpl[MLACommonMetadata]):
         # Share workspace buffer across all executions
         self._workspace = g_sm100_workspace
 
+        # Pre-allocated output buffer, lazily sized on first call.
+        # Zero-init once to prevent NaN in padding slots (seq_lens=0)
+        # from contaminating downstream per-tensor reductions.
+        self._decode_out: torch.Tensor | None = None
+
     def _sm100_cutlass_mla_decode(
         self,
         q_nope: torch.Tensor,
@@ -218,7 +223,15 @@ class CutlassMLAImpl(MLACommonImpl[MLACommonMetadata]):
             if is_quantized_kv_cache(self.kv_cache_dtype)
             else q_nope.dtype
         )
-        out = q_nope.new_empty((B_q, MAX_HEADS, D_latent), dtype=dtype)
+        # Reuse pre-allocated zero-init output buffer to avoid a memset
+        # kernel on every CUDA graph replay.
+        if (
+            self._decode_out is None
+            or self._decode_out.shape[0] < B_q
+            or self._decode_out.dtype != dtype
+        ):
+            self._decode_out = q_nope.new_zeros((B_q, MAX_HEADS, D_latent), dtype=dtype)
+        out = self._decode_out[:B_q]
         lse = (
             torch.empty((B_q, MAX_HEADS), dtype=torch.float32, device=q_nope.device)
             if self.need_to_return_lse_for_decode

vllm/v1/attention/backends/mla/flashinfer_mla.py

@@ -21,6 +21,7 @@ from vllm.v1.attention.backend import (
     AttentionLayer,
     AttentionType,
     MultipleOf,
+    is_quantized_kv_cache,
 )
 from vllm.v1.attention.backends.utils import KVCacheLayoutType
 
@@ -151,6 +152,11 @@ class FlashInferMLAImpl(MLACommonImpl[MLACommonMetadata]):
         self.bmm1_scale: float | None = None
         self.bmm2_scale: float | None = None
 
+        # Pre-allocated output buffer, lazily sized on first call.
+        # Zero-init once to prevent NaN in padding slots (seq_lens=0)
+        # from contaminating downstream per-tensor reductions.
+        self._decode_out: torch.Tensor | None = None
+
     def forward_mqa(
         self,
         q: torch.Tensor | tuple[torch.Tensor, torch.Tensor],
@@ -181,6 +187,37 @@ class FlashInferMLAImpl(MLACommonImpl[MLACommonMetadata]):
         if self.bmm2_scale is None:
             self.bmm2_scale = layer._v_scale_float
 
+        # Reuse pre-allocated zero-init output buffer to avoid a memset
+        # kernel on every CUDA graph replay.
+        # q is 4D: (batch, q_len_per_req, num_heads, head_dim)
+        # FlashInfer has a bug where out= validation hardcodes 3D shape
+        # (batch, num_heads, kv_lora_rank), but the kernel writes 4D
+        # (batch, q_len, num_heads, kv_lora_rank) when q_len > 1.
+        # So we can only pass out= for single-token decode (q_len == 1).
+        # For q_len > 1, we zero padding slots after the kernel returns.
+        # TODO: upstream fix to FlashInfer
+        B, q_len_per_req = q.shape[0], q.shape[1]
+        out_kwargs: dict[str, torch.Tensor] = {}
+        if q_len_per_req == 1:
+            dtype = (
+                torch.bfloat16
+                if is_quantized_kv_cache(self.kv_cache_dtype)
+                else q.dtype
+            )
+            if (
+                self._decode_out is None
+                or self._decode_out.shape[0] < B
+                or self._decode_out.dtype != dtype
+            ):
+                self._decode_out = torch.zeros(
+                    B,
+                    q.shape[2],
+                    self.kv_lora_rank,
+                    dtype=dtype,
+                    device=q.device,
+                )
+            out_kwargs["out"] = self._decode_out[:B]
+
         o = trtllm_batch_decode_with_kv_cache_mla(
             query=q,
             kv_cache=kv_c_and_k_pe_cache.unsqueeze(1),
@@ -193,8 +230,15 @@ class FlashInferMLAImpl(MLACommonImpl[MLACommonMetadata]):
             max_seq_len=attn_metadata.max_seq_len,
             bmm1_scale=self.bmm1_scale,
             bmm2_scale=self.bmm2_scale,
+            **out_kwargs,
         )
 
+        # For q_len > 1, we can't pass out= so we work around by zeroing padding slots
+        if not out_kwargs:
+            num_real = attn_metadata.num_decodes
+            if num_real < o.shape[0]:
+                o[num_real:] = 0
+
         # Flatten the output for consistent shape
         o = o.view(-1, o.shape[-2], o.shape[-1])
 

vllm/v1/worker/gpu_worker.py

@@ -392,8 +392,10 @@ class Worker(WorkerBase):
             )
 
             # Profile CUDA graph memory if graphs will be captured.
+            # Skip on ROCm/HIP as graph pool handles and mem_get_info behave
+            # differently and can produce incorrect/negative estimates.
             cudagraph_memory_estimate = 0
-            if not self.model_config.enforce_eager:
+            if not self.model_config.enforce_eager and not current_platform.is_rocm():
                 cudagraph_memory_estimate = self.model_runner.profile_cudagraph_memory()
 
         # Use the pre-cudagraph torch peak to avoid double-counting.
@@ -406,6 +408,8 @@ class Worker(WorkerBase):
             + profile_result.weights_memory
         )
 
+        # On ROCm, cudagraph_memory_estimate is always 0 so this is a no-op.
+        # On CUDA, respect the opt-in flag as originally designed.
         cudagraph_memory_estimate_applied = (
             cudagraph_memory_estimate
             if envs.VLLM_MEMORY_PROFILER_ESTIMATE_CUDAGRAPHS
@@ -517,7 +521,6 @@ class Worker(WorkerBase):
 
     def update_max_model_len(self, max_model_len: int) -> None:
         """Update max_model_len after auto-fit to GPU memory.
-
         This is called when max_model_len=-1 is used and the engine
         automatically determines the maximum context length that fits
         in GPU memory. Workers need to update their cached max_model_len