deprecate timeout

Signed-off-by: NickLucche <nlucches@redhat.com>
deprecate timeout
2026-04-28 09:03:27 +02:00 · 2026-04-28 09:01:01 +02:00 · 2026-04-27 08:22:35 -07:00 · 2026-04-27 08:19:57 -07:00 · 2026-04-27 13:26:51 +00:00 · 2026-04-27 05:57:13 -07:00
172 changed files with 8104 additions and 2151 deletions
@@ -209,7 +209,6 @@ steps:
        depends_on: ~

      - label: "Build release image - x86_64 - CPU"
-        key: build-cpu-release-image-x86
        depends_on:
          - block-cpu-release-image-build
          - input-release-version
@@ -228,8 +227,7 @@ steps:
        depends_on: ~

      - label: "Build release image - arm64 - CPU"
-        key: build-cpu-release-image-arm64
-        depends_on:
+        depends_on: 
          - block-arm64-cpu-release-image-build
          - input-release-version
        agents:
@@ -338,41 +336,6 @@ steps:
          DOCKER_BUILDKIT: "1"
          DOCKERHUB_USERNAME: "vllmbot"

-      - block: "Publish release images to DockerHub"
-        key: block-publish-release-images
-        depends_on:
-          - create-multi-arch-manifest
-          - create-multi-arch-manifest-cuda-12-9
-          - create-multi-arch-manifest-ubuntu2404
-          - create-multi-arch-manifest-cuda-12-9-ubuntu2404
-          - build-rocm-release-image
-          - input-release-version
-          # Wait for CPU builds if their block steps were unblocked, so publish
-          # doesn't race the in-progress CPU build. allow_failure lets publish
-          # proceed when the operator legitimately leaves the CPU block steps
-          # unblocked or the CPU build fails.
-          - step: build-cpu-release-image-x86
-            allow_failure: true
-          - step: build-cpu-release-image-arm64
-            allow_failure: true
-        if: build.env("NIGHTLY") != "1"
-
-      - label: "Publish release images to DockerHub"
-        depends_on:
-          - block-publish-release-images
-        key: publish-release-images-dockerhub
-        agents:
-          queue: small_cpu_queue_release
-        commands:
-          - "bash .buildkite/scripts/publish-release-images.sh"
-        plugins:
-          - docker-login#v3.0.0:
-              username: vllmbot
-              password-env: DOCKERHUB_TOKEN
-        env:
-          DOCKER_BUILDKIT: "1"
-          DOCKERHUB_USERNAME: "vllmbot"
-
  - group: "Publish wheels"
    key: "publish-wheels"
    steps:
@@ -8,6 +8,8 @@ if [ -z "${RELEASE_VERSION}" ]; then
  RELEASE_VERSION="1.0.0.dev"
 fi

+ROCM_BASE_CACHE_KEY=$(.buildkite/scripts/cache-rocm-base-wheels.sh key)
+
 buildkite-agent annotate --style 'info' --context 'release-workflow' << EOF
 To download the wheel (by commit):
 \`\`\`
@@ -23,5 +25,95 @@ aws s3 cp s3://vllm-wheels/${BUILDKITE_COMMIT}/vllm-${RELEASE_VERSION}+cpu-cp38-
 aws s3 cp s3://vllm-wheels/${BUILDKITE_COMMIT}/vllm-${RELEASE_VERSION}+cpu-cp38-abi3-manylinux_2_35_aarch64.whl .
 \`\`\`

-Docker images are published automatically by the "Publish release images to DockerHub" pipeline step.
+
+To download and upload the image:
+
+\`\`\`
+# Download images:
+
+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:${ROCM_BASE_CACHE_KEY}-rocm-base
+docker pull public.ecr.aws/q9t5s3a7/vllm-release-repo:${BUILDKITE_COMMIT}-rocm
+docker pull public.ecr.aws/q9t5s3a7/vllm-cpu-release-repo:v${RELEASE_VERSION}
+docker pull public.ecr.aws/q9t5s3a7/vllm-arm64-cpu-release-repo:v${RELEASE_VERSION}
+
+# Tag and push images:
+
+## CUDA
+
+docker tag public.ecr.aws/q9t5s3a7/vllm-release-repo:${BUILDKITE_COMMIT}-x86_64 vllm/vllm-openai:x86_64
+docker tag vllm/vllm-openai:x86_64 vllm/vllm-openai:latest-x86_64
+docker tag vllm/vllm-openai:x86_64 vllm/vllm-openai:v${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:${BUILDKITE_COMMIT}-aarch64 vllm/vllm-openai:aarch64
+docker tag vllm/vllm-openai:aarch64 vllm/vllm-openai:latest-aarch64
+docker tag vllm/vllm-openai:aarch64 vllm/vllm-openai:v${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
+
+## ROCm
+
+docker tag public.ecr.aws/q9t5s3a7/vllm-release-repo:${BUILDKITE_COMMIT}-rocm vllm/vllm-openai-rocm:${BUILDKITE_COMMIT}
+docker tag vllm/vllm-openai-rocm:${BUILDKITE_COMMIT} vllm/vllm-openai-rocm:latest
+docker tag vllm/vllm-openai-rocm:${BUILDKITE_COMMIT} vllm/vllm-openai-rocm:v${RELEASE_VERSION}
+docker push vllm/vllm-openai-rocm:latest
+docker push vllm/vllm-openai-rocm:v${RELEASE_VERSION}
+
+docker tag public.ecr.aws/q9t5s3a7/vllm-release-repo:${ROCM_BASE_CACHE_KEY}-rocm-base vllm/vllm-openai-rocm:${BUILDKITE_COMMIT}-base
+docker tag vllm/vllm-openai-rocm:${BUILDKITE_COMMIT}-base vllm/vllm-openai-rocm:latest-base
+docker tag vllm/vllm-openai-rocm:${BUILDKITE_COMMIT}-base vllm/vllm-openai-rocm:v${RELEASE_VERSION}-base
+docker push vllm/vllm-openai-rocm:latest-base
+docker push vllm/vllm-openai-rocm:v${RELEASE_VERSION}-base
+
+## CPU
+
+docker tag public.ecr.aws/q9t5s3a7/vllm-cpu-release-repo:v${RELEASE_VERSION} vllm/vllm-openai-cpu:x86_64
+docker tag vllm/vllm-openai-cpu:x86_64 vllm/vllm-openai-cpu:latest-x86_64
+docker tag vllm/vllm-openai-cpu:x86_64 vllm/vllm-openai-cpu:v${RELEASE_VERSION}-x86_64
+docker push vllm/vllm-openai-cpu:latest-x86_64
+docker push vllm/vllm-openai-cpu:v${RELEASE_VERSION}-x86_64
+
+docker tag public.ecr.aws/q9t5s3a7/vllm-arm64-cpu-release-repo:v${RELEASE_VERSION} vllm/vllm-openai-cpu:arm64
+docker tag vllm/vllm-openai-cpu:arm64 vllm/vllm-openai-cpu:latest-arm64
+docker tag vllm/vllm-openai-cpu:arm64 vllm/vllm-openai-cpu:v${RELEASE_VERSION}-arm64
+docker push vllm/vllm-openai-cpu:latest-arm64
+docker push vllm/vllm-openai-cpu:v${RELEASE_VERSION}-arm64
+
+# Create multi-arch manifest:
+
+docker manifest rm vllm/vllm-openai:latest
+docker manifest create vllm/vllm-openai:latest vllm/vllm-openai:latest-x86_64 vllm/vllm-openai:latest-aarch64
+docker manifest create vllm/vllm-openai:v${RELEASE_VERSION} vllm/vllm-openai:v${RELEASE_VERSION}-x86_64 vllm/vllm-openai:v${RELEASE_VERSION}-aarch64
+docker manifest push vllm/vllm-openai:latest
+docker manifest push vllm/vllm-openai:v${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-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 create vllm/vllm-openai-cpu:v${RELEASE_VERSION} vllm/vllm-openai-cpu:v${RELEASE_VERSION}-x86_64 vllm/vllm-openai-cpu:v${RELEASE_VERSION}-arm64
+docker manifest push vllm/vllm-openai-cpu:latest
+docker manifest push vllm/vllm-openai-cpu:v${RELEASE_VERSION}
+\`\`\`
 EOF
@@ -1,180 +0,0 @@
-#!/bin/bash
-# SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-#
-# Publish release Docker images from ECR to DockerHub.
-# Pulls per-arch images, tags with latest and versioned tags, pushes them,
-# then creates and pushes multi-arch manifests.
-
-set -euo pipefail
-
-RELEASE_VERSION=$(buildkite-agent meta-data get release-version --default "" | sed 's/^v//')
-if [ -z "${RELEASE_VERSION}" ]; then
-  echo "ERROR: release-version metadata not set"
-  exit 1
-fi
-
-COMMIT="$BUILDKITE_COMMIT"
-ROCM_BASE_CACHE_KEY=$(.buildkite/scripts/cache-rocm-base-wheels.sh key)
-
-echo "========================================"
-echo "Publishing release images v${RELEASE_VERSION}"
-echo "  Commit: ${COMMIT}"
-echo "  ROCm base cache key: ${ROCM_BASE_CACHE_KEY}"
-echo "========================================"
-
-# Login to ECR to pull staging images
-aws ecr-public get-login-password --region us-east-1 | \
-  docker login --username AWS --password-stdin public.ecr.aws/q9t5s3a7
-
-# ---- CUDA (default: 13.0) ----
-
-docker pull public.ecr.aws/q9t5s3a7/vllm-release-repo:${COMMIT}-x86_64
-docker pull public.ecr.aws/q9t5s3a7/vllm-release-repo:${COMMIT}-aarch64
-
-docker tag public.ecr.aws/q9t5s3a7/vllm-release-repo:${COMMIT}-x86_64 vllm/vllm-openai:latest-x86_64
-docker tag public.ecr.aws/q9t5s3a7/vllm-release-repo:${COMMIT}-x86_64 vllm/vllm-openai:v${RELEASE_VERSION}-x86_64
-docker push vllm/vllm-openai:latest-x86_64
-docker push vllm/vllm-openai:v${RELEASE_VERSION}-x86_64
-
-docker tag public.ecr.aws/q9t5s3a7/vllm-release-repo:${COMMIT}-aarch64 vllm/vllm-openai:latest-aarch64
-docker tag public.ecr.aws/q9t5s3a7/vllm-release-repo:${COMMIT}-aarch64 vllm/vllm-openai:v${RELEASE_VERSION}-aarch64
-docker push vllm/vllm-openai:latest-aarch64
-docker push vllm/vllm-openai:v${RELEASE_VERSION}-aarch64
-
-docker manifest rm vllm/vllm-openai:latest || true
-docker manifest rm vllm/vllm-openai:v${RELEASE_VERSION} || true
-docker manifest create vllm/vllm-openai:latest vllm/vllm-openai:latest-x86_64 vllm/vllm-openai:latest-aarch64
-docker manifest create vllm/vllm-openai:v${RELEASE_VERSION} vllm/vllm-openai:v${RELEASE_VERSION}-x86_64 vllm/vllm-openai:v${RELEASE_VERSION}-aarch64
-docker manifest push vllm/vllm-openai:latest
-docker manifest push vllm/vllm-openai:v${RELEASE_VERSION}
-
-# ---- CUDA 12.9 ----
-
-docker pull public.ecr.aws/q9t5s3a7/vllm-release-repo:${COMMIT}-x86_64-cu129
-docker pull public.ecr.aws/q9t5s3a7/vllm-release-repo:${COMMIT}-aarch64-cu129
-
-docker tag public.ecr.aws/q9t5s3a7/vllm-release-repo:${COMMIT}-x86_64-cu129 vllm/vllm-openai:latest-x86_64-cu129
-docker tag public.ecr.aws/q9t5s3a7/vllm-release-repo:${COMMIT}-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:${COMMIT}-aarch64-cu129 vllm/vllm-openai:latest-aarch64-cu129
-docker tag public.ecr.aws/q9t5s3a7/vllm-release-repo:${COMMIT}-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 manifest rm vllm/vllm-openai:latest-cu129 || true
-docker manifest rm vllm/vllm-openai:v${RELEASE_VERSION}-cu129 || true
-docker manifest create vllm/vllm-openai:latest-cu129 vllm/vllm-openai:latest-x86_64-cu129 vllm/vllm-openai:latest-aarch64-cu129
-docker manifest create vllm/vllm-openai:v${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${RELEASE_VERSION}-cu129
-
-# ---- Ubuntu 24.04 (CUDA 13.0) ----
-
-docker pull public.ecr.aws/q9t5s3a7/vllm-release-repo:${COMMIT}-x86_64-ubuntu2404
-docker pull public.ecr.aws/q9t5s3a7/vllm-release-repo:${COMMIT}-aarch64-ubuntu2404
-
-docker tag public.ecr.aws/q9t5s3a7/vllm-release-repo:${COMMIT}-x86_64-ubuntu2404 vllm/vllm-openai:latest-x86_64-ubuntu2404
-docker tag public.ecr.aws/q9t5s3a7/vllm-release-repo:${COMMIT}-x86_64-ubuntu2404 vllm/vllm-openai:v${RELEASE_VERSION}-x86_64-ubuntu2404
-docker push vllm/vllm-openai:latest-x86_64-ubuntu2404
-docker push vllm/vllm-openai:v${RELEASE_VERSION}-x86_64-ubuntu2404
-
-docker tag public.ecr.aws/q9t5s3a7/vllm-release-repo:${COMMIT}-aarch64-ubuntu2404 vllm/vllm-openai:latest-aarch64-ubuntu2404
-docker tag public.ecr.aws/q9t5s3a7/vllm-release-repo:${COMMIT}-aarch64-ubuntu2404 vllm/vllm-openai:v${RELEASE_VERSION}-aarch64-ubuntu2404
-docker push vllm/vllm-openai:latest-aarch64-ubuntu2404
-docker push vllm/vllm-openai:v${RELEASE_VERSION}-aarch64-ubuntu2404
-
-docker manifest rm vllm/vllm-openai:latest-ubuntu2404 || true
-docker manifest rm vllm/vllm-openai:v${RELEASE_VERSION}-ubuntu2404 || true
-docker manifest create vllm/vllm-openai:latest-ubuntu2404 vllm/vllm-openai:latest-x86_64-ubuntu2404 vllm/vllm-openai:latest-aarch64-ubuntu2404
-docker manifest create vllm/vllm-openai:v${RELEASE_VERSION}-ubuntu2404 vllm/vllm-openai:v${RELEASE_VERSION}-x86_64-ubuntu2404 vllm/vllm-openai:v${RELEASE_VERSION}-aarch64-ubuntu2404
-docker manifest push vllm/vllm-openai:latest-ubuntu2404
-docker manifest push vllm/vllm-openai:v${RELEASE_VERSION}-ubuntu2404
-
-# ---- Ubuntu 24.04 (CUDA 12.9) ----
-
-docker pull public.ecr.aws/q9t5s3a7/vllm-release-repo:${COMMIT}-x86_64-cu129-ubuntu2404
-docker pull public.ecr.aws/q9t5s3a7/vllm-release-repo:${COMMIT}-aarch64-cu129-ubuntu2404
-
-docker tag public.ecr.aws/q9t5s3a7/vllm-release-repo:${COMMIT}-x86_64-cu129-ubuntu2404 vllm/vllm-openai:latest-x86_64-cu129-ubuntu2404
-docker tag public.ecr.aws/q9t5s3a7/vllm-release-repo:${COMMIT}-x86_64-cu129-ubuntu2404 vllm/vllm-openai:v${RELEASE_VERSION}-x86_64-cu129-ubuntu2404
-docker push vllm/vllm-openai:latest-x86_64-cu129-ubuntu2404
-docker push vllm/vllm-openai:v${RELEASE_VERSION}-x86_64-cu129-ubuntu2404
-
-docker tag public.ecr.aws/q9t5s3a7/vllm-release-repo:${COMMIT}-aarch64-cu129-ubuntu2404 vllm/vllm-openai:latest-aarch64-cu129-ubuntu2404
-docker tag public.ecr.aws/q9t5s3a7/vllm-release-repo:${COMMIT}-aarch64-cu129-ubuntu2404 vllm/vllm-openai:v${RELEASE_VERSION}-aarch64-cu129-ubuntu2404
-docker push vllm/vllm-openai:latest-aarch64-cu129-ubuntu2404
-docker push vllm/vllm-openai:v${RELEASE_VERSION}-aarch64-cu129-ubuntu2404
-
-docker manifest rm vllm/vllm-openai:latest-cu129-ubuntu2404 || true
-docker manifest rm vllm/vllm-openai:v${RELEASE_VERSION}-cu129-ubuntu2404 || true
-docker manifest create vllm/vllm-openai:latest-cu129-ubuntu2404 vllm/vllm-openai:latest-x86_64-cu129-ubuntu2404 vllm/vllm-openai:latest-aarch64-cu129-ubuntu2404
-docker manifest create vllm/vllm-openai:v${RELEASE_VERSION}-cu129-ubuntu2404 vllm/vllm-openai:v${RELEASE_VERSION}-x86_64-cu129-ubuntu2404 vllm/vllm-openai:v${RELEASE_VERSION}-aarch64-cu129-ubuntu2404
-docker manifest push vllm/vllm-openai:latest-cu129-ubuntu2404
-docker manifest push vllm/vllm-openai:v${RELEASE_VERSION}-cu129-ubuntu2404
-
-# ---- ROCm ----
-
-docker pull public.ecr.aws/q9t5s3a7/vllm-release-repo:${COMMIT}-rocm
-docker pull public.ecr.aws/q9t5s3a7/vllm-release-repo:${ROCM_BASE_CACHE_KEY}-rocm-base
-
-docker tag public.ecr.aws/q9t5s3a7/vllm-release-repo:${COMMIT}-rocm vllm/vllm-openai-rocm:latest
-docker tag public.ecr.aws/q9t5s3a7/vllm-release-repo:${COMMIT}-rocm vllm/vllm-openai-rocm:v${RELEASE_VERSION}
-docker push vllm/vllm-openai-rocm:latest
-docker push vllm/vllm-openai-rocm:v${RELEASE_VERSION}
-
-docker tag public.ecr.aws/q9t5s3a7/vllm-release-repo:${ROCM_BASE_CACHE_KEY}-rocm-base vllm/vllm-openai-rocm:latest-base
-docker tag public.ecr.aws/q9t5s3a7/vllm-release-repo:${ROCM_BASE_CACHE_KEY}-rocm-base vllm/vllm-openai-rocm:v${RELEASE_VERSION}-base
-docker push vllm/vllm-openai-rocm:latest-base
-docker push vllm/vllm-openai-rocm:v${RELEASE_VERSION}-base
-
-# ---- CPU ----
-# CPU images are behind separate block steps and may not have been built.
-# All-or-nothing: inspect both arches first, then either publish everything
-# (per-arch + multi-arch manifest) or skip everything. Publishing only one
-# arch would leave `:latest-x86_64` pointing at the new release while the
-# `:latest` multi-arch manifest still resolves to the previous release.
-
-CPU_X86_TAG=public.ecr.aws/q9t5s3a7/vllm-cpu-release-repo:v${RELEASE_VERSION}
-CPU_ARM_TAG=public.ecr.aws/q9t5s3a7/vllm-arm64-cpu-release-repo:v${RELEASE_VERSION}
-
-CPU_X86_AVAILABLE=false
-CPU_ARM_AVAILABLE=false
-docker manifest inspect "${CPU_X86_TAG}" >/dev/null 2>&1 && CPU_X86_AVAILABLE=true
-docker manifest inspect "${CPU_ARM_TAG}" >/dev/null 2>&1 && CPU_ARM_AVAILABLE=true
-
-if [ "$CPU_X86_AVAILABLE" = "true" ] && [ "$CPU_ARM_AVAILABLE" = "true" ]; then
-  docker pull "${CPU_X86_TAG}"
-  docker tag "${CPU_X86_TAG}" vllm/vllm-openai-cpu:latest-x86_64
-  docker tag "${CPU_X86_TAG}" vllm/vllm-openai-cpu:v${RELEASE_VERSION}-x86_64
-  docker push vllm/vllm-openai-cpu:latest-x86_64
-  docker push vllm/vllm-openai-cpu:v${RELEASE_VERSION}-x86_64
-
-  docker pull "${CPU_ARM_TAG}"
-  docker tag "${CPU_ARM_TAG}" vllm/vllm-openai-cpu:latest-arm64
-  docker tag "${CPU_ARM_TAG}" vllm/vllm-openai-cpu:v${RELEASE_VERSION}-arm64
-  docker push vllm/vllm-openai-cpu:latest-arm64
-  docker push vllm/vllm-openai-cpu:v${RELEASE_VERSION}-arm64
-
-  docker manifest rm vllm/vllm-openai-cpu:latest || true
-  docker manifest rm vllm/vllm-openai-cpu:v${RELEASE_VERSION} || true
-  docker manifest create vllm/vllm-openai-cpu:latest vllm/vllm-openai-cpu:latest-x86_64 vllm/vllm-openai-cpu:latest-arm64
-  docker manifest create vllm/vllm-openai-cpu:v${RELEASE_VERSION} vllm/vllm-openai-cpu:v${RELEASE_VERSION}-x86_64 vllm/vllm-openai-cpu:v${RELEASE_VERSION}-arm64
-  docker manifest push vllm/vllm-openai-cpu:latest
-  docker manifest push vllm/vllm-openai-cpu:v${RELEASE_VERSION}
-elif [ "$CPU_X86_AVAILABLE" = "false" ] && [ "$CPU_ARM_AVAILABLE" = "false" ]; then
-  echo "WARNING: Neither CPU image found in ECR, skipping CPU publish (ensure block-cpu-release-image-build and block-arm64-cpu-release-image-build were unblocked and the builds finished pushing)"
-else
-  # Partial state: one arch built, the other did not. Fail loudly rather than
-  # ship a Docker Hub state where `:latest-${arch}` and `:latest` (multi-arch)
-  # disagree on which release they point at.
-  echo "ERROR: Partial CPU build detected (x86_64=${CPU_X86_AVAILABLE}, arm64=${CPU_ARM_AVAILABLE})."
-  echo "       Refusing to publish to avoid split-tag drift between per-arch and multi-arch tags."
-  echo "       Re-run the missing CPU build and retry, or manually publish if a single-arch release is intended."
-  exit 1
-fi
-
-echo ""
-echo "Successfully published release images for v${RELEASE_VERSION}"
@@ -39,11 +39,10 @@ fi

 set -x # avoid printing secrets above

-# install twine and sdist build prerequisites from pypi
+# install twine from pypi
 python3 -m venv /tmp/vllm-release-env
 source /tmp/vllm-release-env/bin/activate
 pip install twine
-pip install -r requirements/build/cuda.txt
 python3 -m twine --version

 # copy release wheels to local directory
@@ -388,10 +388,10 @@ steps:
    - python3 basic/offline_inference/embed.py
    - python3 basic/offline_inference/score.py
    # Multi-modal models
-    - python3 offline_inference/audio_language.py --seed 0
-    - python3 offline_inference/vision_language.py --seed 0
-    - python3 offline_inference/vision_language_multi_image.py --seed 0
-    - python3 offline_inference/encoder_decoder_multimodal.py --model-type whisper --seed 0
+    - python3 generate/multimodal/audio_language_offline.py --seed 0
+    - python3 generate/multimodal/vision_language_offline.py --seed 0
+    - python3 generate/multimodal/vision_language_multi_image_offline.py --seed 0
+    - python3 generate/multimodal/encoder_decoder_multimodal_offline.py --model-type whisper --seed 0
    # Pooling models
    - python3 pooling/embed/vision_embedding_offline.py --seed 0
    # Features demo
@@ -1647,10 +1647,10 @@ steps:
    - python3 basic/offline_inference/embed.py
    - python3 basic/offline_inference/score.py
    # Multi-modal models
-    - python3 offline_inference/audio_language.py --seed 0
-    - python3 offline_inference/vision_language.py --seed 0
-    - python3 offline_inference/vision_language_multi_image.py --seed 0
-    - python3 offline_inference/encoder_decoder_multimodal.py --model-type whisper --seed 0
+    - python3 generate/multimodal/audio_language_offline.py --seed 0
+    - python3 generate/multimodal/vision_language_offline.py --seed 0
+    - python3 generate/multimodal/vision_language_multi_image_offline.py --seed 0
+    - python3 generate/multimodal/encoder_decoder_multimodal_offline.py --model-type whisper --seed 0
    # Pooling models
    - python3 pooling/embed/vision_embedding_offline.py --seed 0
    # Features demo
@@ -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/offline_inference/vision_language.py --model-type qwen2_5_vl
-  - VLLM_WORKER_MULTIPROC_METHOD=spawn python3 examples/offline_inference/audio_language.py --model-type whisper
+  - python3 examples/generate/multimodal/vision_language_offline.py --model-type qwen2_5_vl
+  - VLLM_WORKER_MULTIPROC_METHOD=spawn python3 examples/generate/multimodal/audio_language_offline.py --model-type whisper

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

@@ -2930,10 +2930,10 @@ steps:
  - python3 basic/offline_inference/embed.py
  - python3 basic/offline_inference/score.py
  # Multi-modal models
-  - python3 offline_inference/audio_language.py --seed 0
-  - python3 offline_inference/vision_language.py --seed 0
-  - python3 offline_inference/vision_language_multi_image.py --seed 0
-  - python3 offline_inference/encoder_decoder_multimodal.py --model-type whisper --seed 0
+  - python3 generate/multimodal/audio_language_offline.py --seed 0
+  - python3 generate/multimodal/vision_language_offline.py --seed 0
+  - python3 generate/multimodal/vision_language_multi_image_offline.py --seed 0
+  - python3 generate/multimodal/encoder_decoder_multimodal_offline.py --model-type whisper --seed 0
  # Pooling models
  - python3 pooling/embed/vision_embedding_offline.py --seed 0
  # Features demo
@@ -95,11 +95,13 @@ 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 offline_inference/audio_language.py --seed 0
-    - python3 offline_inference/vision_language.py --seed 0
-    - python3 offline_inference/vision_language_multi_image.py --seed 0
-    - python3 offline_inference/encoder_decoder_multimodal.py --model-type whisper --seed 0
+    - python3 generate/multimodal/audio_language_offline.py --seed 0
+    - python3 generate/multimodal/vision_language_offline.py --seed 0
+    - python3 generate/multimodal/vision_language_multi_image_offline.py --seed 0
+    - python3 generate/multimodal/encoder_decoder_multimodal_offline.py --model-type whisper --seed 0
     # for pooling models
    - python3 pooling/embed/vision_embedding_offline.py --seed 0
    # for features demo
@@ -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 offline_inference/audio_language.py --seed 0
-    - python3 offline_inference/vision_language.py --seed 0
-    - python3 offline_inference/vision_language_multi_image.py --seed 0
-    - python3 offline_inference/encoder_decoder_multimodal.py --model-type whisper --seed 0
+    - python3 generate/multimodal/audio_language_offline.py --seed 0
+    - python3 generate/multimodal/vision_language_offline.py --seed 0
+    - python3 generate/multimodal/vision_language_multi_image_offline.py --seed 0
+    - python3 generate/multimodal/encoder_decoder_multimodal_offline.py --model-type whisper --seed 0
    # for pooling models
    - python3 pooling/embed/vision_embedding_offline.py --seed 0
    # for features demo
@@ -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/offline_inference/vision_language.py --model-type qwen2_5_vl
+    - python3 examples/generate/multimodal/vision_language_offline.py --model-type qwen2_5_vl
    # Whisper needs spawn method to avoid deadlock
-    - VLLM_WORKER_MULTIPROC_METHOD=spawn python3 examples/offline_inference/audio_language.py --model-type whisper
+    - VLLM_WORKER_MULTIPROC_METHOD=spawn python3 examples/generate/multimodal/audio_language_offline.py --model-type whisper

 - label: Transformers Backward Compatibility Models Test
  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/offline_inference/basic/chat.py
-    - python3 examples/offline_inference/vision_language.py --model-type qwen2_5_vl
+    - python3 examples/basic/offline_inference/chat.py
+    - python3 examples/generate/multimodal/vision_language_offline.py --model-type qwen2_5_vl
    # Whisper needs spawn method to avoid deadlock
-    - VLLM_WORKER_MULTIPROC_METHOD=spawn python3 examples/offline_inference/audio_language.py --model-type whisper
+    - VLLM_WORKER_MULTIPROC_METHOD=spawn python3 examples/generate/multimodal/audio_language_offline.py --model-type whisper
@@ -389,11 +389,7 @@ pull_request_rules:
      - files~=^tests/entrypoints/anthropic/.*tool.*
      - files~=^vllm/tool_parsers/
      - files=docs/features/tool_calling.md
-      - files~=^examples/tool_chat_*
-      - files=examples/offline_inference/chat_with_tools.py
-      - files=examples/online_serving/openai_chat_completion_client_with_tools_required.py
-      - files=examples/online_serving/openai_chat_completion_tool_calls_with_reasoning.py
-      - files=examples/online_serving/openai_chat_completion_client_with_tools.py
+      - files~=^examples/tool_calling/
  actions:
    label:
      add:
@@ -0,0 +1,324 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+# Benchmarks FP8 vs BF16 ViT attention via FlashInfer cuDNN backend.
+#
+# == Usage Examples ==
+#
+# Benchmark mode (default, FlashInfer CUDAGraph Bench)
+#   python3 benchmark_vit_fp8_attn.py
+#
+# Profile mode (PyTorch profiler, saves TensorBoard traces):
+#   python3 benchmark_vit_fp8_attn.py --profile
+#   python3 benchmark_vit_fp8_attn.py --profile --profile-output-dir ./profile_traces
+#
+# Custom seq_lens:
+#   python3 benchmark_vit_fp8_attn.py --seq-lens 4096 8192 16384
+
+from functools import partial
+
+import numpy as np
+import torch
+from torch.profiler import ProfilerActivity, profile, record_function
+
+from vllm.utils.argparse_utils import FlexibleArgumentParser
+
+# Qwen3-VL defaults
+NUM_HEADS = 16
+HEAD_DIM = 72
+DEFAULT_SEQ_LENS = [2304, 4096, 8192, 16384]
+
+
+def _setup_fp8_attention(num_heads: int, head_dim: int) -> tuple:
+    """Create FP8 and BF16 attention modules + workspace."""
+    from types import SimpleNamespace
+    from unittest.mock import patch
+
+    from vllm.config import VllmConfig, set_current_vllm_config
+    from vllm.config.multimodal import MultiModalConfig
+    from vllm.model_executor.layers.attention.mm_encoder_attention import (
+        MMEncoderAttention,
+        _get_flashinfer_workspace_buffer,
+    )
+    from vllm.v1.attention.backends.registry import AttentionBackendEnum
+
+    old_dtype = torch.get_default_dtype()
+    torch.set_default_dtype(torch.bfloat16)
+
+    backend_patch = patch(
+        "vllm.model_executor.layers.attention.mm_encoder_attention"
+        ".get_vit_attn_backend",
+        return_value=AttentionBackendEnum.FLASHINFER,
+    )
+
+    # FP8 attention
+    mm_config_fp8 = MultiModalConfig(mm_encoder_attn_dtype="fp8")
+    vllm_config_fp8 = VllmConfig()
+    vllm_config_fp8.model_config = SimpleNamespace(multimodal_config=mm_config_fp8)
+    with set_current_vllm_config(vllm_config_fp8), backend_patch:
+        attn_fp8 = MMEncoderAttention(
+            num_heads=num_heads,
+            head_size=head_dim,
+            prefix="visual.blocks.0.attn",
+        ).to("cuda")
+
+    # BF16 attention (no FP8)
+    with set_current_vllm_config(VllmConfig()), backend_patch:
+        attn_bf16 = MMEncoderAttention(
+            num_heads=num_heads,
+            head_size=head_dim,
+            prefix="visual.blocks.0.attn",
+        ).to("cuda")
+
+    torch.set_default_dtype(old_dtype)
+
+    workspace = _get_flashinfer_workspace_buffer()
+    return attn_fp8, attn_bf16, workspace
+
+
+def _build_meta(
+    seq_len: int,
+    num_heads: int,
+    head_dim: int,
+    fp8: bool,
+):
+    """Build cu_seqlens, max_seqlen, sequence_lengths."""
+    from vllm.model_executor.layers.attention.mm_encoder_attention import (
+        MMEncoderAttention,
+    )
+    from vllm.utils.math_utils import round_up
+    from vllm.v1.attention.backends.registry import AttentionBackendEnum
+
+    cu_np = np.array([0, seq_len], dtype=np.int32)
+    fp8_padded = num_heads * round_up(head_dim, 16) if fp8 else None
+
+    seq_lengths = MMEncoderAttention.maybe_compute_seq_lens(
+        AttentionBackendEnum.FLASHINFER, cu_np, torch.device("cuda")
+    )
+    max_seqlen = torch.tensor(
+        MMEncoderAttention.compute_max_seqlen(AttentionBackendEnum.FLASHINFER, cu_np),
+        dtype=torch.int32,
+    )
+    cu_seqlens = MMEncoderAttention.maybe_recompute_cu_seqlens(
+        AttentionBackendEnum.FLASHINFER,
+        cu_np,
+        num_heads * head_dim,
+        1,
+        torch.device("cuda"),
+        fp8_padded_hidden_size=fp8_padded,
+    )
+    return cu_seqlens, max_seqlen, seq_lengths
+
+
+def run_benchmark(
+    seq_lens: list[int],
+    num_heads: int,
+    head_dim: int,
+    method: str,
+):
+    """Benchmark FP8 vs BF16 attention across seq_lens.
+
+    Uses FlashInfer GPU-level timing to measure pure kernel time,
+    excluding CPU launch overhead.
+    """
+    if method == "cupti":
+        from flashinfer.testing import bench_gpu_time_with_cupti as bench_fn
+
+        bench_fn = partial(bench_fn, use_cuda_graph=True, cold_l2_cache=False)
+    elif method == "cudagraph":
+        from flashinfer.testing import (
+            bench_gpu_time_with_cudagraph as bench_fn,
+        )
+
+        bench_fn = partial(bench_fn, cold_l2_cache=False)
+    else:
+        raise ValueError(f"Invalid method: {method}")
+
+    attn_fp8, attn_bf16, workspace = _setup_fp8_attention(num_heads, head_dim)
+
+    print(f"Timing method: {method}")
+    print(f"{'seq_len':>8} {'BF16 (us)':>12} {'FP8 (us)':>12} {'Speedup':>10}")
+    print("-" * 46)
+
+    for seq_len in seq_lens:
+        torch.manual_seed(42)
+
+        q = torch.randn(
+            seq_len,
+            num_heads,
+            head_dim,
+            device="cuda",
+            dtype=torch.bfloat16,
+        )
+        k = torch.randn_like(q)
+        v = torch.randn_like(q)
+
+        cu_fp8, max_s, seq_l = _build_meta(seq_len, num_heads, head_dim, fp8=True)
+        # we can reuse cu_fp8 for cu_bf16 since q, k, and v are contiguous
+        cu_bf16 = cu_fp8.clone()
+
+        def bf16_fn(q=q, k=k, v=v, cu=cu_bf16, ms=max_s, sl=seq_l):
+            attn_bf16._forward_flashinfer(q, k, v, cu, ms, sl)
+
+        def fp8_fn(q=q, k=k, v=v, cu=cu_fp8, ms=max_s, sl=seq_l):
+            attn_fp8._forward_flashinfer(q, k, v, cu, ms, sl)
+
+        # bench_fn returns List[float] of per-iteration times in ms
+        bf16_times = bench_fn(bf16_fn)
+        fp8_times = bench_fn(fp8_fn)
+
+        bf16_us = np.median(bf16_times) * 1e3  # ms -> us
+        fp8_us = np.median(fp8_times) * 1e3
+        speedup = bf16_us / fp8_us if fp8_us > 0 else float("inf")
+
+        print(f"{seq_len:>8} {bf16_us:>12.1f} {fp8_us:>12.1f} {speedup:>9.2f}x")
+
+
+def _make_trace_handler(output_dir: str, worker_name: str, label: str):
+    """Create a trace handler that saves to TensorBoard and prints summary."""
+
+    def handler(prof):
+        torch.profiler.tensorboard_trace_handler(output_dir, worker_name)(prof)
+        print(f"\n{'=' * 80}")
+        print(label)
+        print(f"{'=' * 80}")
+        print(prof.key_averages().table(sort_by="cuda_time_total", row_limit=20))
+
+    return handler
+
+
+def run_profile(
+    seq_len: int,
+    num_heads: int,
+    head_dim: int,
+    warmup: int,
+    output_dir: str,
+):
+    """Profile FP8 vs BF16 attention with PyTorch profiler."""
+    attn_fp8, attn_bf16, workspace = _setup_fp8_attention(num_heads, head_dim)
+
+    torch.manual_seed(42)
+    q = torch.randn(
+        seq_len,
+        num_heads,
+        head_dim,
+        device="cuda",
+        dtype=torch.bfloat16,
+    )
+    k = torch.randn_like(q)
+    v = torch.randn_like(q)
+
+    cu_fp8, max_s, seq_l = _build_meta(seq_len, num_heads, head_dim, fp8=True)
+    # we can reuse cu_fp8 for cu_bf16 since q, k, and v are contiguous
+    cu_bf16 = cu_fp8.clone()
+
+    sched = torch.profiler.schedule(wait=0, warmup=warmup, active=1)
+
+    # Profile BF16 (warmup handled by profiler schedule)
+    with profile(
+        activities=[ProfilerActivity.CPU, ProfilerActivity.CUDA],
+        schedule=sched,
+        on_trace_ready=_make_trace_handler(
+            output_dir,
+            f"bf16_h{head_dim}_s{seq_len}",
+            f"BF16 Attention (seq_len={seq_len}, heads={num_heads}, "
+            f"head_dim={head_dim})",
+        ),
+    ) as prof_bf16:
+        for _ in range(warmup + 1):
+            with record_function("bf16_attention"):
+                attn_bf16._forward_flashinfer(
+                    q.clone(), k.clone(), v.clone(), cu_bf16, max_s, seq_l
+                )
+                torch.accelerator.synchronize()
+            prof_bf16.step()
+
+    # Profile FP8 (warmup handled by profiler schedule)
+    with profile(
+        activities=[ProfilerActivity.CPU, ProfilerActivity.CUDA],
+        schedule=sched,
+        on_trace_ready=_make_trace_handler(
+            output_dir,
+            f"fp8_h{head_dim}_s{seq_len}",
+            f"FP8 Attention (seq_len={seq_len}, heads={num_heads}, "
+            f"head_dim={head_dim})",
+        ),
+    ) as prof_fp8:
+        for _ in range(warmup + 1):
+            with record_function("fp8_attention"):
+                attn_fp8._forward_flashinfer(
+                    q.clone(), k.clone(), v.clone(), cu_fp8, max_s, seq_l
+                )
+                torch.accelerator.synchronize()
+            prof_fp8.step()
+
+    print(f"\nTensorBoard traces saved to: {output_dir}")
+    print(f"View with: tensorboard --logdir={output_dir}")
+
+
+if __name__ == "__main__":
+    parser = FlexibleArgumentParser(description="Benchmark FP8 vs BF16 ViT attention.")
+    parser.add_argument(
+        "--seq-lens",
+        type=int,
+        nargs="+",
+        default=DEFAULT_SEQ_LENS,
+        help="Sequence lengths to benchmark",
+    )
+    parser.add_argument(
+        "--num-heads",
+        type=int,
+        default=NUM_HEADS,
+    )
+    parser.add_argument(
+        "--head-dim",
+        type=int,
+        default=HEAD_DIM,
+    )
+    parser.add_argument(
+        "--method",
+        choices=["cupti", "cudagraph"],
+        default="cudagraph",
+        help="GPU timing method: cupti (CUPTI kernel timing) or "
+        "cudagraph (CUDA graph capture/replay). Default: cudagraph",
+    )
+    parser.add_argument(
+        "--warmup",
+        type=int,
+        default=10,
+        help="Warmup iterations (profile mode only)",
+    )
+    parser.add_argument(
+        "--profile",
+        action="store_true",
+        help="Run PyTorch profiler instead of benchmark",
+    )
+    parser.add_argument(
+        "--profile-seq-len",
+        type=int,
+        default=8192,
+        help="Sequence length for profiling (default: 8192)",
+    )
+    parser.add_argument(
+        "--profile-output-dir",
+        type=str,
+        default="./profile_traces",
+        help="Output directory for TensorBoard traces (default: ./profile_traces)",
+    )
+    args = parser.parse_args()
+
+    if args.profile:
+        run_profile(
+            args.profile_seq_len,
+            args.num_heads,
+            args.head_dim,
+            args.warmup,
+            args.profile_output_dir,
+        )
+    else:
+        run_benchmark(
+            args.seq_lens,
+            args.num_heads,
+            args.head_dim,
+            args.method,
+        )
@@ -887,14 +887,27 @@ __global__ void __launch_bounds__(kThreadsPerBlock, 2)
  uint32_t* shared_ordered =
      reinterpret_cast<uint32_t*>(smem_raw + kFixedSmemLarge);

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

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

@@ -82,109 +82,22 @@ void launch_persistent_topk(const torch::Tensor& logits,
    size_t smem_size = P::kFixedSmemLarge + chunk_size * sizeof(uint32_t);
    if (smem_size < P::kSmemMedium) smem_size = P::kSmemMedium;

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

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

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

-    // Zero the per-group RadixRowState region before launch.
-    //
-    // Issued UNCONDITIONALLY so the memset is captured as its own node in
-    // the cudagraph (a separate cudaMemsetAsync node, sequenced before the
-    // persistent_topk_kernel launch on the same stream). The previous
-    // host-side guard `if (needs_cooperative)` was evaluated at capture time;
-    // when capture-time max_seq_len <= RADIX_THRESHOLD (always true under
-    // FULL_DECODE_ONLY with max_model_len < 32 K) the memset would NOT be
-    // captured, leaving the workspace state to accumulate across replays.
-    // That's a latent correctness bug if the runtime data ever takes the
-    // radix path, and removes one variable while debugging hangs in the
-    // decode/medium paths.
-    //
-    // Cost is sub-microsecond: state_bytes = num_groups * sizeof(RadixRowState)
-    // is ~3 KB per group, ~100 KB for the largest grids on this hardware.
-    //
-    // Why the memset is required (regardless of which path the kernel takes):
-    //   1. arrival_counter accumulates within a launch and is never reset,
-    //      so a prior call leaves it at a large positive value. Without this
-    //      reset, the very first wait_ge in the next call sees counter >>
-    //      target and returns instantly, breaking the barrier.
-    //   2. The previous in-kernel init only ran in CTA-0 with intra-CTA
-    //      __syncthreads(), so it had no happens-before edge to CTA-1+'s
-    //      first red_release. cudaMemsetAsync is stream-ordered: the zero
-    //      is globally visible before any CTA runs.
-    {
-      cudaError_t mz_err = cudaMemsetAsync(workspace.data_ptr<uint8_t>(), 0,
-                                           state_bytes, stream);
-      TORCH_CHECK(mz_err == cudaSuccess,
-                  "row_states memset failed: ", cudaGetErrorString(mz_err));
-    }
-
    P::PersistentTopKParams params;
    params.input = logits.data_ptr<float>();
    params.output = output.data_ptr<int32_t>();
@@ -490,46 +490,41 @@ WORKDIR /vllm-workspace
 RUN PYTHON_VERSION_STR=$(echo ${PYTHON_VERSION} | sed 's/\.//g') && \
    echo "export PYTHON_VERSION_STR=${PYTHON_VERSION_STR}" >> /etc/environment

-# Install Python from source and system dependencies
+# Install Python and system dependencies
 RUN apt-get update -y \
    && apt-get install -y --no-install-recommends \
+        software-properties-common \
        curl \
        sudo \
        ffmpeg \
        libsm6 \
        libxext6 \
        libgl1 \
+    && if [ ! -z ${DEADSNAKES_MIRROR_URL} ] ; then \
+        if [ ! -z "${DEADSNAKES_GPGKEY_URL}" ] ; then \
+            mkdir -p -m 0755 /etc/apt/keyrings ; \
+            curl -L ${DEADSNAKES_GPGKEY_URL} | gpg --dearmor > /etc/apt/keyrings/deadsnakes.gpg ; \
+            sudo chmod 644 /etc/apt/keyrings/deadsnakes.gpg ; \
+            echo "deb [signed-by=/etc/apt/keyrings/deadsnakes.gpg] ${DEADSNAKES_MIRROR_URL} $(lsb_release -cs) main" > /etc/apt/sources.list.d/deadsnakes.list ; \
+        fi ; \
+    else \
+        for i in 1 2 3; do \
+            add-apt-repository -y ppa:deadsnakes/ppa && break || \
+            { echo "Attempt $i failed, retrying in 5s..."; sleep 5; }; \
+        done ; \
+    fi \
+    && apt-get update -y \
+    && apt-get install -y --no-install-recommends \
+        python${PYTHON_VERSION} \
+        python${PYTHON_VERSION}-dev \
+        python${PYTHON_VERSION}-venv \
        libibverbs-dev \
-        build-essential \
-        libssl-dev \
-        libffi-dev \
-        zlib1g-dev \
-        libbz2-dev \
-        libreadline-dev \
-        libsqlite3-dev \
-        libncurses-dev \
-        liblzma-dev \
-        libgdbm-dev \
-        uuid-dev \
-        tk-dev \
-    && PYTHON_MAJOR_MINOR=${PYTHON_VERSION} \
-    && PYTHON_FULL_VERSION=$(curl -s https://www.python.org/ftp/python/ \
-        | grep -oE "${PYTHON_MAJOR_MINOR}\.[0-9]+" \
-        | sort -t. -k3 -n | uniq \
-        | tail -1) \
-    && echo "Building Python ${PYTHON_FULL_VERSION} from source..." \
-    && curl -fSL https://www.python.org/ftp/python/${PYTHON_FULL_VERSION}/Python-${PYTHON_FULL_VERSION}.tgz -o /tmp/python.tgz \
-    && tar -xzf /tmp/python.tgz -C /tmp \
-    && cd /tmp/Python-${PYTHON_FULL_VERSION} \
-    && ./configure --enable-optimizations --with-ensurepip=install --prefix=/usr/local \
-    && make -j$(nproc) \
-    && make install \
-    && cd / \
-    && rm -rf /tmp/python.tgz /tmp/Python-${PYTHON_FULL_VERSION} \
-    && ln -sf /usr/local/bin/python${PYTHON_MAJOR_MINOR} /usr/bin/python3 \
-    && ln -sf /usr/local/bin/python${PYTHON_MAJOR_MINOR}-config /usr/bin/python3-config \
-    && ln -sf /usr/local/bin/pip${PYTHON_MAJOR_MINOR} /usr/bin/pip \
    && rm -rf /var/lib/apt/lists/* \
+    && update-alternatives --install /usr/bin/python3 python3 /usr/bin/python${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

 # Install CUDA development tools for runtime JIT compilation
@@ -543,11 +538,9 @@ RUN CUDA_VERSION_DASH=$(echo $CUDA_VERSION | cut -d. -f1,2 | tr '.' '-') && \
        cuda-nvrtc-${CUDA_VERSION_DASH} \
        cuda-cuobjdump-${CUDA_VERSION_DASH} \
        libcurand-dev-${CUDA_VERSION_DASH} \
-        libcublas-${CUDA_VERSION_DASH} \
+        libcublas-dev-${CUDA_VERSION_DASH} \
        # Required by fastsafetensors (fixes #20384)
-        libnuma-dev \
-        # numactl CLI for NUMA binding at runtime
-        numactl && \
+        libnuma-dev && \
    # Fixes nccl_allocator requiring nccl.h at runtime
    # https://github.com/vllm-project/vllm/blob/1336a1ea244fa8bfd7e72751cabbdb5b68a0c11a/vllm/distributed/device_communicators/pynccl_allocator.py#L22
    # NCCL packages don't use the cuda-MAJOR-MINOR naming convention,
@@ -36,7 +36,7 @@ th {
 | deepep_high_throughput | standard | fp8 | G(128),A,T<sup>2</sup> | Y | Y | [`DeepEPHTPrepareAndFinalize`][vllm.model_executor.layers.fused_moe.prepare_finalize.deepep_ht.DeepEPHTPrepareAndFinalize] |
 | deepep_low_latency | batched | fp8 | G(128),A,T<sup>3</sup> | Y | Y | [`DeepEPLLPrepareAndFinalize`][vllm.model_executor.layers.fused_moe.prepare_finalize.deepep_ll.DeepEPLLPrepareAndFinalize] |
 | flashinfer_nvlink_two_sided | standard | nvfp4,fp8 | G,A,T | N | N | [`FlashInferNVLinkTwoSidedPrepareAndFinalize`][vllm.model_executor.layers.fused_moe.prepare_finalize.flashinfer_nvlink_two_sided.FlashInferNVLinkTwoSidedPrepareAndFinalize] |
-| flashinfer_nvlink_one_sided | standard | nvfp4,bf16,mxfp8 | G,A,T | N | N | [`FlashInferNVLinkOneSidedPrepareAndFinalize`][vllm.model_executor.layers.fused_moe.prepare_finalize.flashinfer_nvlink_one_sided.FlashInferNVLinkOneSidedPrepareAndFinalize] |
+| flashinfer_nvlink_one_sided | standard | nvfp4 | G,A,T | N | N | [`FlashInferNVLinkOneSidedPrepareAndFinalize`][vllm.model_executor.layers.fused_moe.prepare_finalize.flashinfer_nvlink_one_sided.FlashInferNVLinkOneSidedPrepareAndFinalize] |

 !!! info "Table key"
    1. All types: mxfp4, nvfp4, int4, int8, fp8
@@ -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/offline_inference/vision_language.py](../../examples/offline_inference/vision_language.py)
+Full example: [examples/generate/multimodal/vision_language_offline.py](../../examples/generate/multimodal/vision_language_offline.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/offline_inference/vision_language_multi_image.py](../../examples/offline_inference/vision_language_multi_image.py)
+Full example: [examples/generate/multimodal/vision_language_multi_image_offline.py](../../examples/generate/multimodal/vision_language_multi_image_offline.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/offline_inference/vision_language.py](../../examples/offline_inference/vision_language.py)
+Full example: [examples/generate/multimodal/vision_language_offline.py](../../examples/generate/multimodal/vision_language_offline.py)

 ### Audio Inputs

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

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

 !!! note
    By default, the timeout for fetching videos through HTTP URL is `30` seconds.
@@ -958,7 +958,7 @@ Alternatively, you can pass `audio_url`, which is the audio counterpart of `imag
    print("Chat completion output from audio url:", result)
    ```

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

 !!! note
    By default, the timeout for fetching audios through HTTP URL is `10` seconds.
@@ -20,6 +20,7 @@ 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

@@ -0,0 +1,109 @@
+# 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/online_serving/openai_chat_completion_tool_calls_with_reasoning.py](../../examples/online_serving/openai_chat_completion_tool_calls_with_reasoning.py).
+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).

 ## Server-Level Default Chat Template Kwargs

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

 ### Score task

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

 ### Pooling multitask support

-Pooling multitask support is deprecated and will be removed in v0.20. When the default pooling task is not what you want,
+Pooling multitask support has been removed in v0.21. When the default pooling task is not what you want,
 you need to manually specify it via `PoolerConfig(task=<task>)` offline or `--pooler-config.task <task>` online.
@@ -439,6 +439,7 @@ 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. | | |
@@ -590,6 +591,7 @@ 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. | | ✅︎ |
@@ -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/online_serving/openai_responses_client_with_tools.py)
+Code example: [examples/online_serving/openai_responses_client_with_tools.py](../../examples/tool_calling/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/online_serving/openai_transcription_client.py](../../examples/online_serving/openai_transcription_client.py)
+Code example: [examples/speech_to_text/openai/openai_transcription_client.py](../../examples/speech_to_text/openai/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/online_serving/openai_translation_client.py](../../examples/online_serving/openai_translation_client.py)
+Code example: [examples/speech_to_text/openai/openai_translation_client.py](../../examples/speech_to_text/openai/openai_translation_client.py)

 #### Extra Parameters

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

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

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

 # Multiple audios
-python examples/offline_inference/qwen2_5_omni/only_thinker.py \
+python examples/generate/multimodal/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/offline_inference/audio_language.py \
+python examples/generate/multimodal/audio_language_offline.py \
    --model-type qwen2_5_omni

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

 # Process video inputs
-python examples/offline_inference/vision_language.py \
+python examples/generate/multimodal/vision_language_offline.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.py` # noqa
+    # If you want to use `MiniCPM-o-2_6` with audio inputs, check `audio_language_offline.py` # noqa

    # 2.0
    # The official repo doesn't work yet, so we need to use a fork for now
@@ -4,68 +4,74 @@
 import torch

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

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

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

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

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

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


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

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


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

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

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

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

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

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

 import argparse
+import pprint

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


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

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

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

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

 if __name__ == "__main__":
    args = parse_args()
@@ -34,10 +34,7 @@ def _run_vllm(vllm_runner):
            mode=CompilationMode.VLLM_COMPILE,
            cudagraph_mode=CUDAGraphMode.NONE,
        ),
-        # Phi-tiny-MoE uses SWA, whose admission cap is `cdiv(L, block_size) + 1`
-        # at default block_size=16 — i.e. 17 blocks for max_model_len=256. Use
-        # 32 for headroom.
-        num_gpu_blocks_override=32,
+        num_gpu_blocks_override=8,
    ):
        pass

@@ -193,7 +190,7 @@ def _run_model(vllm_runner, spec: ModelStartupSpec):
            cudagraph_mode=CUDAGraphMode.NONE,
            pass_config=PassConfig(fuse_allreduce_rms=False),
        ),
-        num_gpu_blocks_override=16,
+        num_gpu_blocks_override=8,
    ):
        pass

@@ -405,9 +405,6 @@ def test_should_split():
        (None, 0, 1, False, 2048, CUDAGraphMode.NONE, 0),
        # truncated to nearest multiple of 8 or 16
        (None, 257, 1, False, 2048, CUDAGraphMode.FULL_AND_PIECEWISE, 256),
-        # max_num_batched_tokens <= max_cudagraph_capture_size should always be
-        # captured even if not landing on a 16-stride step
-        (None, 2048, 1, False, 257, CUDAGraphMode.FULL_AND_PIECEWISE, 257),
        # max from list
        ([1, 2, 4, 15], None, 1, False, 2048, CUDAGraphMode.FULL_AND_PIECEWISE, 15),
        # SP forces full-graph compilation, sizes are filtered by TP
@@ -41,3 +41,21 @@ def test_language_model_only_affects_model_hash():
    base_hash = ModelConfig(model).compute_hash()
    lm_only_hash = ModelConfig(model, language_model_only=True).compute_hash()
    assert base_hash != lm_only_hash
+
+
+def test_mm_encoder_fp8_scale_path_requires_fp8():
+    with pytest.raises(ValueError, match="mm_encoder_attn_dtype"):
+        MultiModalConfig(mm_encoder_fp8_scale_path="/tmp/scales.json")
+
+
+def test_mm_encoder_attn_dtype_hash_updates(tmp_path):
+    scale_file = tmp_path / "scales.json"
+    scale_file.write_text("{}")
+    base_hash = MultiModalConfig().compute_hash()
+    fp8_hash = MultiModalConfig(mm_encoder_attn_dtype="fp8").compute_hash()
+    fp8_static_hash = MultiModalConfig(
+        mm_encoder_attn_dtype="fp8",
+        mm_encoder_fp8_scale_path=str(scale_file),
+    ).compute_hash()
+    assert base_hash != fp8_hash
+    assert fp8_hash != fp8_static_hash
@@ -0,0 +1,76 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""Unit tests for ``system_fingerprint`` construction."""
+
+from types import SimpleNamespace
+
+import pytest
+
+from vllm.entrypoints.openai import fingerprint as fp
+
+
+def _cfg(tp=1, pp=1, dp=1, ep=False, digest="a3b21f94deadbeef"):
+    c = SimpleNamespace(
+        parallel_config=SimpleNamespace(
+            tensor_parallel_size=tp,
+            pipeline_parallel_size=pp,
+            data_parallel_size=dp,
+            enable_expert_parallel=ep,
+        )
+    )
+    c.compute_hash = lambda: digest  # type: ignore[attr-defined]
+    return c
+
+
+@pytest.fixture(autouse=True)
+def _reset():
+    fp.set_default_fingerprint_mode("full")
+    yield
+    fp.set_default_fingerprint_mode("full")
+
+
+def test_four_modes_produce_expected_shapes():
+    from vllm import __version__ as v
+
+    cfg = _cfg(tp=8, ep=True)
+
+    assert fp.build_system_fingerprint(cfg, "full") == (f"vllm-{v}-tp8-ep-a3b21f94")
+    assert fp.build_system_fingerprint(cfg, "hash") == f"vllm-{v}-a3b21f94"
+    assert fp.build_system_fingerprint(cfg, "custom", "my-fp") == "my-fp"
+    assert fp.build_system_fingerprint(cfg, "none") is None
+
+
+def test_full_mode_emits_only_non_trivial_parallelism():
+    from vllm import __version__ as v
+
+    # Single-GPU: nothing between version and hash.
+    assert fp.build_system_fingerprint(_cfg(), "full") == f"vllm-{v}-a3b21f94"
+    # All parallelism axes.
+    assert (
+        fp.build_system_fingerprint(_cfg(tp=8, pp=2, dp=4, ep=True), "full")
+        == f"vllm-{v}-tp8-pp2-dp4-ep-a3b21f94"
+    )
+
+
+def test_get_respects_set_default():
+    cfg = _cfg(tp=8)
+    full = fp.get_system_fingerprint(cfg)
+    assert full == fp.get_system_fingerprint(cfg)
+
+    fp.set_default_fingerprint_mode("hash")
+    hashed = fp.get_system_fingerprint(cfg)
+    assert hashed != full
+    assert "tp8" not in hashed
+
+    fp.set_default_fingerprint_mode("custom", "deploy-42")
+    assert fp.get_system_fingerprint(cfg) == "deploy-42"
+
+    fp.set_default_fingerprint_mode("none")
+    assert fp.get_system_fingerprint(cfg) is None
+
+
+def test_compute_hash_failure_does_not_raise():
+    cfg = _cfg()
+    cfg.compute_hash = lambda: (_ for _ in ()).throw(RuntimeError("boom"))
+    assert fp.build_system_fingerprint(cfg, "full").endswith("-nohash")
+    assert fp.build_system_fingerprint(cfg, "hash").endswith("-nohash")
@@ -1,13 +1,12 @@
 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-import logging
 import weakref

 import pytest
 import torch

 from tests.models.utils import softmax
-from vllm import LLM, ClassificationRequestOutput, PoolingParams, PoolingRequestOutput
+from vllm import LLM, ClassificationRequestOutput, PoolingParams
 from vllm.distributed import cleanup_dist_env_and_memory
 from vllm.tasks import PoolingTask

@@ -66,18 +65,6 @@ def test_list_prompts(llm: LLM):
        assert len(outputs[i].outputs.probs) == num_labels


-@pytest.mark.skip_global_cleanup
-def test_token_classify(llm: LLM, caplog_vllm):
-    with caplog_vllm.at_level(level=logging.WARNING, logger="vllm"):
-        outputs = llm.encode(prompt, pooling_task="token_classify", use_tqdm=False)
-        assert "deprecated" in caplog_vllm.text
-
-    assert len(outputs) == 1
-    assert isinstance(outputs[0], PoolingRequestOutput)
-    assert outputs[0].prompt_token_ids == prompt_token_ids
-    assert outputs[0].outputs.data.shape == (len(prompt_token_ids), num_labels)
-
-
@pytest.mark.skip_global_cleanup
 def test_pooling_params(llm: LLM):
    def get_outputs(use_activation):
@@ -110,10 +97,12 @@ def test_score_api(llm: LLM):
        llm.score("ping", "pong", use_tqdm=False)


-@pytest.mark.parametrize("task", ["embed", "token_embed", "plugin"])
+@pytest.mark.parametrize("task", ["embed", "token_embed", "token_classify", "plugin"])
 def test_unsupported_tasks(llm: LLM, task: PoolingTask):
    if task == "plugin":
        err_msg = "No IOProcessor plugin installed."
+    elif task == "token_classify":
+        err_msg = "Try switching the model's pooling_task via.+"
    else:
        err_msg = "Embedding API is not supported by this model.+"
    with pytest.raises(ValueError, match=err_msg):
@@ -436,26 +436,7 @@ async def test_pooling_classify(server: RemoteOpenAIServer, model_name: str):

@pytest.mark.asyncio
@pytest.mark.parametrize("model_name", [MODEL_NAME])
-async def test_pooling_token_classify(server: RemoteOpenAIServer, model_name: str):
-    task = "token_classify"
-    response = requests.post(
-        server.url_for("pooling"),
-        json={
-            "model": model_name,
-            "input": input_text,
-            "encoding_format": "float",
-            "task": task,
-        },
-    )
-    poolings = PoolingResponse.model_validate(response.json())
-    assert len(poolings.data) == 1
-    assert len(poolings.data[0].data) == 8
-    assert len(poolings.data[0].data[0]) == 2
-
-
-@pytest.mark.asyncio
-@pytest.mark.parametrize("model_name", [MODEL_NAME])
-@pytest.mark.parametrize("task", ["embed", "token_embed", "plugin"])
+@pytest.mark.parametrize("task", ["embed", "token_embed", "token_classify", "plugin"])
 async def test_pooling_not_supported(
    server: RemoteOpenAIServer, model_name: str, task: str
 ):
@@ -469,8 +450,11 @@ async def test_pooling_not_supported(
        },
    )
    assert response.json()["error"]["type"] == "BadRequestError"
+
    if task == "plugin":
        err_msg = "No IOProcessor plugin installed."
+    elif task == "token_classify":
+        err_msg = "Try switching the model's pooling_task via"
    else:
        err_msg = f"Unsupported task: {task!r}"
    assert response.json()["error"]["message"].startswith(err_msg)
@@ -1,6 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-import logging
 import weakref

 import pytest
@@ -38,11 +37,11 @@ def llm():
        seed=0,
        attention_config=attention_config,
    )
+    assert embedding_size == llm.model_config.embedding_size

    yield weakref.proxy(llm)

    del llm
-
    cleanup_dist_env_and_memory()


@@ -74,16 +73,6 @@ def test_list_prompts(llm: LLM):
        assert len(outputs[i].outputs.embedding) == embedding_size


-@pytest.mark.skip_global_cleanup
-def test_token_embed(llm: LLM, caplog_vllm):
-    with caplog_vllm.at_level(level=logging.WARNING, logger="vllm"):
-        outputs = llm.encode(prompt, pooling_task="token_embed", use_tqdm=False)
-        assert "deprecated" in caplog_vllm.text
-
-    multi_vector = outputs[0].outputs.data
-    assert multi_vector.shape == (11, 384)
-
-
@pytest.mark.skip_global_cleanup
 def test_pooling_params(llm: LLM):
    def get_outputs(normalize):
@@ -107,10 +96,14 @@ def test_pooling_params(llm: LLM):
    )


-@pytest.mark.parametrize("task", ["token_classify", "classify", "plugin"])
+@pytest.mark.parametrize(
+    "task", ["token_classify", "classify", "token_embed", "plugin"]
+)
 def test_unsupported_tasks(llm: LLM, task: PoolingTask):
    if task == "plugin":
        err_msg = "No IOProcessor plugin installed."
+    elif task == "token_embed":
+        err_msg = "Try switching the model's pooling_task via.+"
    else:
        err_msg = "Classification API is not supported by this model.+"
    with pytest.raises(ValueError, match=err_msg):
@@ -732,28 +732,9 @@ async def test_pooling_embed(server: RemoteOpenAIServer, model_name: str):

@pytest.mark.asyncio
@pytest.mark.parametrize("model_name", [MODEL_NAME])
-async def test_pooling_token_embed(server: RemoteOpenAIServer, model_name: str):
-    task = "token_embed"
-    response = requests.post(
-        server.url_for("pooling"),
-        json={
-            "model": model_name,
-            "input": input_text,
-            "encoding_format": "float",
-            "task": task,
-        },
-    )
-
-    poolings = PoolingResponse.model_validate(response.json())
-
-    assert len(poolings.data) == 1
-    assert len(poolings.data[0].data) == len(input_tokens)
-    assert len(poolings.data[0].data[0]) == 384
-
-
-@pytest.mark.asyncio
-@pytest.mark.parametrize("model_name", [MODEL_NAME])
-@pytest.mark.parametrize("task", ["classify", "token_classify", "plugin"])
+@pytest.mark.parametrize(
+    "task", ["classify", "token_classify", "token_embed", "plugin"]
+)
 async def test_pooling_not_supported(
    server: RemoteOpenAIServer, model_name: str, task: str
 ):
@@ -769,6 +750,8 @@ async def test_pooling_not_supported(
    assert response.json()["error"]["type"] == "BadRequestError"
    if task == "plugin":
        err_msg = "No IOProcessor plugin installed."
+    elif task == "token_embed":
+        err_msg = "Try switching the model's pooling_task via"
    else:
        err_msg = f"Unsupported task: {task!r}"
    assert response.json()["error"]["message"].startswith(err_msg)
@@ -452,25 +452,6 @@ async def test_pooling_classify(server: RemoteOpenAIServer):
    assert len(poolings.data[0].data) == 1


-@pytest.mark.asyncio
-async def test_pooling_token_classify(server: RemoteOpenAIServer):
-    response = requests.post(
-        server.url_for("pooling"),
-        json={
-            "model": MODEL_NAME,
-            "task": "token_classify",
-            "input": input_text,
-            "encoding_format": "float",
-        },
-    )
-
-    poolings = PoolingResponse.model_validate(response.json())
-
-    assert len(poolings.data) == 1
-    assert len(poolings.data[0].data) == len(input_tokens)
-    assert len(poolings.data[0].data[0]) == 1
-
-
@pytest.mark.asyncio
 async def test_rerank_max_tokens_per_doc(
    server: RemoteOpenAIServer,
@@ -544,7 +525,7 @@ async def test_rerank_max_tokens_per_doc_validation(


@pytest.mark.asyncio
-@pytest.mark.parametrize("task", ["embed", "token_embed", "plugin"])
+@pytest.mark.parametrize("task", ["embed", "token_embed", "token_classify", "plugin"])
 async def test_pooling_not_supported(server: RemoteOpenAIServer, task: str):
    response = requests.post(
        server.url_for("pooling"),
@@ -558,6 +539,8 @@ async def test_pooling_not_supported(server: RemoteOpenAIServer, task: str):
    assert response.json()["error"]["type"] == "BadRequestError"
    if task == "plugin":
        err_msg = "No IOProcessor plugin installed."
+    elif task == "token_classify":
+        err_msg = "Try switching the model's pooling_task via"
    else:
        err_msg = f"Unsupported task: {task!r}"
    assert response.json()["error"]["message"].startswith(err_msg)
@@ -1,6 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-import logging
 import weakref

 import pytest
@@ -60,22 +59,19 @@ def test_token_ids_prompts(llm: LLM):

@pytest.mark.skip_global_cleanup
 def test_score_api(llm: LLM):
-    err_msg = "Scoring API is only enabled for num_labels == 1."
+    err_msg = "This model does not support the Scoring API."
    with pytest.raises(ValueError, match=err_msg):
        llm.score("ping", "pong", use_tqdm=False)


@pytest.mark.parametrize("task", ["classify", "embed", "token_embed", "plugin"])
 def test_unsupported_tasks(llm: LLM, task: PoolingTask, caplog_vllm):
-    if task == "classify":
-        with caplog_vllm.at_level(level=logging.WARNING, logger="vllm"):
-            llm.encode(prompt, pooling_task=task, use_tqdm=False)
-        assert "deprecated" in caplog_vllm.text
+    if task == "plugin":
+        err_msg = "No IOProcessor plugin installed."
+    elif task == "classify":
+        err_msg = "Try switching the model's pooling_task via.+"
    else:
-        if task == "plugin":
-            err_msg = "No IOProcessor plugin installed."
-        else:
-            err_msg = "Embedding API is not supported by this model.+"
+        err_msg = "Embedding API is not supported by this model.+"

-        with pytest.raises(ValueError, match=err_msg):
-            llm.encode(prompt, pooling_task=task, use_tqdm=False)
+    with pytest.raises(ValueError, match=err_msg):
+        llm.encode(prompt, pooling_task=task, use_tqdm=False)
@@ -50,7 +50,7 @@ async def test_pooling_token_classify(server: RemoteOpenAIServer, model_name: st

@pytest.mark.asyncio
@pytest.mark.parametrize("model_name", [MODEL_NAME])
-@pytest.mark.parametrize("task", ["embed", "token_embed", "plugin"])
+@pytest.mark.parametrize("task", ["classify", "embed", "token_embed", "plugin"])
 async def test_pooling_not_supported(
    server: RemoteOpenAIServer, model_name: str, task: str
 ):
@@ -63,9 +63,12 @@ async def test_pooling_not_supported(
            "task": task,
        },
    )
+    assert response.json()["error"]["type"] == "BadRequestError"

    if task == "plugin":
        err_msg = "No IOProcessor plugin installed."
+    elif task == "classify":
+        err_msg = "Try switching the model's pooling_task via"
    else:
        err_msg = f"Unsupported task: {task!r}"
    assert response.json()["error"]["message"].startswith(err_msg)
@@ -1,6 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-import logging
 import weakref

 import pytest
@@ -64,15 +63,12 @@ def test_token_ids_prompts(llm: LLM):

@pytest.mark.parametrize("task", ["embed", "classify", "token_classify", "plugin"])
 def test_unsupported_tasks(llm: LLM, task: PoolingTask, caplog_vllm):
-    if task == "embed":
-        with caplog_vllm.at_level(level=logging.WARNING, logger="vllm"):
-            llm.encode(prompt, pooling_task=task, use_tqdm=False)
-        assert "deprecated" in caplog_vllm.text
+    if task == "plugin":
+        err_msg = "No IOProcessor plugin installed."
+    elif task == "embed":
+        err_msg = "Try switching the model's pooling_task via.+"
    else:
-        if task == "plugin":
-            err_msg = "No IOProcessor plugin installed."
-        else:
-            err_msg = "Classification API is not supported by this model.+"
+        err_msg = "Classification API is not supported by this model.+"

-        with pytest.raises(ValueError, match=err_msg):
-            llm.encode(prompt, pooling_task=task, use_tqdm=False)
+    with pytest.raises(ValueError, match=err_msg):
+        llm.encode(prompt, pooling_task=task, use_tqdm=False)
@@ -73,7 +73,7 @@ async def test_pooling_token_embed(server: RemoteOpenAIServer, model_name: str):

@pytest.mark.asyncio
@pytest.mark.parametrize("model_name", [MODEL_NAME])
-@pytest.mark.parametrize("task", ["classify", "token_classify", "plugin"])
+@pytest.mark.parametrize("task", ["embed", "classify", "token_classify", "plugin"])
 async def test_pooling_not_supported(
    server: RemoteOpenAIServer, model_name: str, task: str
 ):
@@ -86,9 +86,12 @@ async def test_pooling_not_supported(
            "task": task,
        },
    )
+    assert response.json()["error"]["type"] == "BadRequestError"

    if task == "plugin":
        err_msg = "No IOProcessor plugin installed."
+    elif task == "embed":
+        err_msg = "Try switching the model's pooling_task via"
    else:
        err_msg = f"Unsupported task: {task!r}"
    assert response.json()["error"]["message"].startswith(err_msg)
@@ -0,0 +1,279 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""Tests for the full FP8 ViT attention path (quantize -> cuDNN -> un-pad)."""
+
+import contextlib
+
+import pytest
+import torch
+
+from vllm.triton_utils import HAS_TRITON
+from vllm.utils.flashinfer import (
+    is_flashinfer_cudnn_fp8_prefill_attn_supported,
+)
+from vllm.v1.attention.backends.registry import AttentionBackendEnum
+
+
+def _has_flashinfer_cudnn() -> bool:
+    """Check if FlashInfer cuDNN backend is available."""
+    try:
+        from flashinfer.prefill import (
+            cudnn_batch_prefill_with_kv_cache,  # noqa: F401
+        )
+
+        return True
+    except ImportError:
+        return False
+
+
+HEAD_DIMS = [72, 80]
+SEQ_LENS = [256]
+NUM_HEADS = [16]
+
+
+@pytest.fixture
+def _fp8_attention():
+    """Create FP8-enabled MMEncoderAttention via config."""
+    from types import SimpleNamespace
+    from unittest.mock import patch
+
+    from vllm.config import VllmConfig, set_current_vllm_config
+    from vllm.config.multimodal import MultiModalConfig
+
+    if not is_flashinfer_cudnn_fp8_prefill_attn_supported():
+        pytest.skip("FlashInfer cuDNN FP8 prefill attention not supported")
+
+    mm_config = MultiModalConfig(mm_encoder_attn_dtype="fp8")
+    vllm_config = VllmConfig()
+    vllm_config.model_config = SimpleNamespace(multimodal_config=mm_config)
+
+    # MMEncoderAttention reads torch.get_default_dtype() during init
+    # to determine the output dtype. In real model loading this is bf16.
+    old_dtype = torch.get_default_dtype()
+    torch.set_default_dtype(torch.bfloat16)
+
+    with (
+        set_current_vllm_config(vllm_config),
+        patch(
+            "vllm.model_executor.layers.attention.mm_encoder_attention"
+            ".get_vit_attn_backend",
+            return_value=AttentionBackendEnum.FLASHINFER,
+        ),
+    ):
+        yield
+
+    torch.set_default_dtype(old_dtype)
+
+
+def _build_cu_seqlens_and_meta(
+    seq_len: int,
+    num_heads: int,
+    head_dim: int,
+    fp8_padded_hidden_size: int | None = None,
+) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+    """Build cu_seqlens, max_seqlen, sequence_lengths for a single sequence."""
+    import numpy as np
+
+    from vllm.model_executor.layers.attention.mm_encoder_attention import (
+        MMEncoderAttention,
+    )
+
+    cu_seqlens_np = np.array([0, seq_len], dtype=np.int32)
+
+    sequence_lengths = MMEncoderAttention.maybe_compute_seq_lens(
+        AttentionBackendEnum.FLASHINFER,
+        cu_seqlens_np,
+        torch.device("cuda"),
+    )
+
+    max_seqlen = torch.tensor(
+        MMEncoderAttention.compute_max_seqlen(
+            AttentionBackendEnum.FLASHINFER, cu_seqlens_np
+        ),
+        dtype=torch.int32,
+    )
+
+    cu_seqlens = MMEncoderAttention.maybe_recompute_cu_seqlens(
+        AttentionBackendEnum.FLASHINFER,
+        cu_seqlens_np,
+        num_heads * head_dim,
+        1,  # tp_size
+        torch.device("cuda"),
+        fp8_padded_hidden_size=fp8_padded_hidden_size,
+    )
+
+    return cu_seqlens, max_seqlen, sequence_lengths
+
+
+@pytest.mark.skipif(
+    not (HAS_TRITON and _has_flashinfer_cudnn()),
+    reason="Triton and FlashInfer cuDNN required",
+)
+@pytest.mark.parametrize("head_dim", HEAD_DIMS)
+@pytest.mark.parametrize("seq_len", SEQ_LENS)
+@pytest.mark.parametrize("num_heads", NUM_HEADS)
+def test_fp8_attn_output_shape(
+    head_dim: int,
+    seq_len: int,
+    num_heads: int,
+    _fp8_attention,
+) -> None:
+    """Verify FP8 attention produces correct output shape after un-padding."""
+    from vllm.model_executor.layers.attention.mm_encoder_attention import (
+        MMEncoderAttention,
+    )
+    from vllm.utils.math_utils import round_up
+
+    attn = None
+    with contextlib.suppress(ValueError, ImportError):
+        attn = MMEncoderAttention(
+            num_heads=num_heads,
+            head_size=head_dim,
+            prefix="visual.blocks.0.attn",
+        ).to("cuda")
+
+    if attn is None or not attn.fp8_enabled:
+        pytest.skip("FP8 MMEncoderAttention not available")
+    assert attn is not None  # mypy narrowing
+
+    # FP8 always needs fp8_padded_hidden_size for correct cu_seqlens
+    fp8_padded_hidden_size = num_heads * round_up(head_dim, 16)
+
+    cu_seqlens, max_seqlen, sequence_lengths = _build_cu_seqlens_and_meta(
+        seq_len, num_heads, head_dim, fp8_padded_hidden_size=fp8_padded_hidden_size
+    )
+
+    q = torch.randn(
+        seq_len,
+        num_heads,
+        head_dim,
+        device="cuda",
+        dtype=torch.bfloat16,
+    )
+    k = torch.randn_like(q)
+    v = torch.randn_like(q)
+
+    output = attn._forward_flashinfer(q, k, v, cu_seqlens, max_seqlen, sequence_lengths)
+
+    # Output should have original head_dim (un-padded)
+    assert output.shape[-1] == head_dim
+    assert output.dtype == torch.bfloat16
+
+
+@pytest.mark.skipif(
+    not (HAS_TRITON and _has_flashinfer_cudnn()),
+    reason="Triton and FlashInfer cuDNN required",
+)
+@pytest.mark.parametrize("head_dim", HEAD_DIMS)
+@pytest.mark.parametrize("seq_len", SEQ_LENS)
+@pytest.mark.parametrize("num_heads", NUM_HEADS)
+def test_fp8_vs_bf16_close(
+    head_dim: int, seq_len: int, num_heads: int, _fp8_attention
+) -> None:
+    """FP8 attention output should be reasonably close to BF16 baseline."""
+    from vllm.model_executor.layers.attention.mm_encoder_attention import (
+        MMEncoderAttention,
+    )
+    from vllm.utils.math_utils import round_up
+
+    torch.manual_seed(42)
+    q = torch.randn(
+        1,
+        seq_len,
+        num_heads,
+        head_dim,
+        device="cuda",
+        dtype=torch.bfloat16,
+    )
+    k = torch.randn_like(q)
+    v = torch.randn_like(q)
+
+    # FP8 path
+    attn_fp8 = None
+    with contextlib.suppress(ValueError, ImportError):
+        attn_fp8 = MMEncoderAttention(
+            num_heads=num_heads,
+            head_size=head_dim,
+            prefix="visual.blocks.0.attn",
+        ).to("cuda")
+
+    if attn_fp8 is None or not attn_fp8.fp8_enabled:
+        pytest.skip("FP8 MMEncoderAttention not available")
+    assert attn_fp8 is not None  # mypy narrowing
+
+    fp8_padded_hidden_size = num_heads * round_up(head_dim, 16)
+    cu_seqlens, max_seqlen, seq_lengths = _build_cu_seqlens_and_meta(
+        seq_len,
+        num_heads,
+        head_dim,
+        fp8_padded_hidden_size=fp8_padded_hidden_size,
+    )
+
+    out_fp8 = attn_fp8._forward_flashinfer(
+        q.clone(),
+        k.clone(),
+        v.clone(),
+        cu_seqlens,
+        max_seqlen,
+        seq_lengths,
+    )
+
+    # BF16 baseline (create non-FP8 attention by using scale=attn_fp8.scale
+    # and calling the wrapper directly without FP8 quantization)
+    from vllm.model_executor.layers.attention.mm_encoder_attention import (
+        _get_flashinfer_workspace_buffer,
+    )
+    from vllm.v1.attention.ops.vit_attn_wrappers import (
+        vit_flashinfer_wrapper,
+    )
+
+    out_bf16 = vit_flashinfer_wrapper(
+        q=q.clone(),
+        k=k.clone(),
+        v=v.clone(),
+        scale=attn_fp8.scale,
+        workspace_buffer=_get_flashinfer_workspace_buffer(),
+        cu_seqlens=cu_seqlens,
+        max_seqlen=max_seqlen,
+        sequence_lengths=seq_lengths,
+    )
+
+    out_fp8_f = out_fp8.float()
+    out_bf16_f = out_bf16.float()
+
+    abs_diff = (out_fp8_f - out_bf16_f).abs()
+    abs_diff_flat = abs_diff.flatten()
+
+    # Relative diff (avoid division by zero)
+    denom = out_bf16_f.abs().clamp(min=1e-6)
+    rel_diff_flat = (abs_diff / denom).flatten()
+
+    cosine_sim = torch.nn.functional.cosine_similarity(
+        out_fp8_f.flatten().unsqueeze(0),
+        out_bf16_f.flatten().unsqueeze(0),
+    ).item()
+
+    pcts = [50, 90, 95, 99, 99.9]
+    abs_pct = {p: torch.quantile(abs_diff_flat, p / 100).item() for p in pcts}
+    rel_pct = {p: torch.quantile(rel_diff_flat, p / 100).item() for p in pcts}
+
+    print(f"\nFP8 vs BF16 (head_dim={head_dim}, seq_len={seq_len}):")
+    print(f"  cosine_sim={cosine_sim:.6f}")
+    print(
+        f"  abs_diff: max={abs_diff_flat.max().item():.6f}, "
+        f"mean={abs_diff_flat.mean().item():.6f}, "
+        + ", ".join(f"p{p}={abs_pct[p]:.6f}" for p in pcts)
+    )
+    print(
+        f"  rel_diff: max={rel_diff_flat.max().item():.6f}, "
+        f"mean={rel_diff_flat.mean().item():.6f}, "
+        + ", ".join(f"p{p}={rel_pct[p]:.6f}" for p in pcts)
+    )
+
+    assert abs_diff_flat.max().item() < 0.3, (
+        f"FP8 vs BF16 max abs diff too large: {abs_diff_flat.max().item()}"
+    )
+    assert abs_diff_flat.mean().item() < 0.03, (
+        f"FP8 vs BF16 mean abs diff too large: {abs_diff_flat.mean().item()}"
+    )
+    assert cosine_sim > 0.99, f"Cosine similarity too low: {cosine_sim:.6f}"
@@ -0,0 +1,124 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""Tests for the stride-aware FP8 quantization kernel with head_dim padding."""
+
+import pytest
+import torch
+
+from vllm.platforms import current_platform
+from vllm.triton_utils import HAS_TRITON
+
+if HAS_TRITON:
+    from vllm.kernels.triton.qkv_padded_fp8_quant import (
+        quantize_fp8_pad_head_dim_triton,
+    )
+
+HEAD_DIMS = [72, 80, 128]
+SEQ_LENS = [64, 256]
+NUM_HEADS = [16]
+SCALES = [0.01, 0.1, 1.0]
+
+
+def _naive_fp8_quantize(
+    tensor: torch.Tensor, scale: torch.Tensor, skip_scale: bool
+) -> torch.Tensor:
+    """Reference FP8 quantization in PyTorch."""
+    fp8_dtype = current_platform.fp8_dtype()
+    fp8_max = torch.finfo(fp8_dtype).max
+    fp8_min = -fp8_max
+
+    x = tensor.float()
+    if not skip_scale:
+        x = x / scale.item()
+    x = x.clamp(fp8_min, fp8_max)
+    return x.to(fp8_dtype)
+
+
+@pytest.mark.skipif(not HAS_TRITON, reason="Triton not available")
+@pytest.mark.parametrize("head_dim", HEAD_DIMS)
+@pytest.mark.parametrize("seq_len", SEQ_LENS)
+@pytest.mark.parametrize("num_heads", NUM_HEADS)
+@pytest.mark.parametrize("scale_val", SCALES)
+def test_quantize_contiguous(
+    head_dim: int, seq_len: int, num_heads: int, scale_val: float
+) -> None:
+    """Test quantization of contiguous 3D tensors."""
+    torch.manual_seed(42)
+    tensor = torch.randn(
+        seq_len, num_heads, head_dim, device="cuda", dtype=torch.bfloat16
+    )
+    scale = torch.tensor([scale_val], dtype=torch.float32, device="cuda").view(
+        1, 1, 1, 1
+    )
+
+    result = quantize_fp8_pad_head_dim_triton(tensor, scale)
+
+    padded_dim = (head_dim + 15) // 16 * 16
+    assert result.shape == (seq_len, num_heads, padded_dim)
+    assert result.is_contiguous()
+    assert result.dtype == current_platform.fp8_dtype()
+
+    # Compare unpadded portion against reference
+    ref = _naive_fp8_quantize(tensor, scale, skip_scale=False)
+    torch.testing.assert_close(result[:, :, :head_dim].float(), ref.float())
+
+    # Padded region should be zero
+    if padded_dim > head_dim:
+        assert (result[:, :, head_dim:].float() == 0).all()
+
+
+@pytest.mark.skipif(not HAS_TRITON, reason="Triton not available")
+@pytest.mark.parametrize("head_dim", [72, 80])
+def test_quantize_non_contiguous(head_dim: int) -> None:
+    """Test quantization from non-contiguous QKV views (interleaved buffer)."""
+    seq_len, num_heads = 64, 16
+    # Simulate interleaved QKV buffer: shape (seq_len, 3 * num_heads, head_dim)
+    qkv = torch.randn(
+        seq_len, 3 * num_heads, head_dim, device="cuda", dtype=torch.bfloat16
+    )
+    # Q is every 3rd head slice - non-contiguous view
+    q = qkv[:, 0::3, :]
+    assert not q.is_contiguous()
+
+    scale = torch.tensor([0.1], dtype=torch.float32, device="cuda").view(1, 1, 1, 1)
+    result = quantize_fp8_pad_head_dim_triton(q, scale)
+
+    padded_dim = (head_dim + 15) // 16 * 16
+    assert result.shape == (seq_len, num_heads, padded_dim)
+    assert result.is_contiguous()
+
+    # Compare against contiguous reference
+    ref = _naive_fp8_quantize(q.contiguous(), scale, skip_scale=False)
+    torch.testing.assert_close(result[:, :, :head_dim].float(), ref.float())
+
+
+@pytest.mark.skipif(not HAS_TRITON, reason="Triton not available")
+def test_skip_scale() -> None:
+    """Test skip_scale=True produces cast-only output (no division)."""
+    seq_len, num_heads, head_dim = 32, 8, 80
+    tensor = torch.randn(
+        seq_len, num_heads, head_dim, device="cuda", dtype=torch.bfloat16
+    )
+    scale = torch.tensor([0.5], dtype=torch.float32, device="cuda").view(1, 1, 1, 1)
+
+    result_skip = quantize_fp8_pad_head_dim_triton(tensor, scale, skip_scale=True)
+    result_noskip = quantize_fp8_pad_head_dim_triton(tensor, scale, skip_scale=False)
+
+    # skip_scale should just cast, not divide
+    ref_cast = _naive_fp8_quantize(tensor, scale, skip_scale=True)
+    torch.testing.assert_close(result_skip[:, :, :head_dim].float(), ref_cast.float())
+
+    # With scale != 1.0, skip and no-skip should differ
+    assert not torch.equal(result_skip.float(), result_noskip.float())
+
+
+@pytest.mark.skipif(not HAS_TRITON, reason="Triton not available")
+def test_4d_input() -> None:
+    """Test that 4D input (B, S, H, D) is handled correctly."""
+    B, S, H, D = 2, 32, 8, 72
+    tensor = torch.randn(B, S, H, D, device="cuda", dtype=torch.bfloat16)
+    scale = torch.tensor([0.1], dtype=torch.float32, device="cuda").view(1, 1, 1, 1)
+
+    result = quantize_fp8_pad_head_dim_triton(tensor, scale)
+    padded_dim = (D + 15) // 16 * 16
+    assert result.shape == (B, S, H, padded_dim)
@@ -0,0 +1,251 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""Tests for FP8 scaling (dynamic and static) in MMEncoderAttention."""
+
+import contextlib
+import json
+from types import SimpleNamespace
+from unittest.mock import patch
+
+import pytest
+import torch
+
+from vllm.model_executor.layers.attention.mm_encoder_attention import (
+    _FP8_AMAX_HISTORY_LEN,
+    _FP8_MAX,
+)
+from vllm.utils.flashinfer import (
+    is_flashinfer_cudnn_fp8_prefill_attn_supported,
+)
+
+LAYER_0 = "visual.blocks.0.attn.attn"
+LAYER_1 = "visual.blocks.1.attn.attn"
+NUM_HEADS = 16
+HEAD_DIM = 72
+
+
+@contextlib.contextmanager
+def _build_attention(mm_config):
+    """Yield an MMEncoderAttention with the given multimodal config.
+
+    The VllmConfig context stays active while the test runs so that
+    ``get_multimodal_config()`` calls during the forward path resolve. Also
+    invokes ``process_weights_after_loading`` to simulate the model loader's
+    auto-scan. Yields ``None`` if FlashInfer cuDNN is not available.
+    """
+    from vllm.config import VllmConfig, set_current_vllm_config
+    from vllm.model_executor.layers.attention.mm_encoder_attention import (
+        MMEncoderAttention,
+    )
+    from vllm.v1.attention.backends.registry import AttentionBackendEnum
+
+    if not is_flashinfer_cudnn_fp8_prefill_attn_supported():
+        yield None
+        return
+
+    vllm_config = VllmConfig()
+    vllm_config.model_config = SimpleNamespace(multimodal_config=mm_config)
+
+    with (
+        set_current_vllm_config(vllm_config),
+        patch(
+            "vllm.model_executor.layers.attention.mm_encoder_attention"
+            ".get_vit_attn_backend",
+            return_value=AttentionBackendEnum.FLASHINFER,
+        ),
+    ):
+        attn = MMEncoderAttention(
+            num_heads=NUM_HEADS,
+            head_size=HEAD_DIM,
+            prefix=LAYER_0,
+        )
+        attn.process_weights_after_loading(torch.bfloat16)
+        yield attn
+
+
+@pytest.fixture
+def _make_attention():
+    """Create an MMEncoderAttention with dynamic FP8 scaling."""
+    from vllm.config.multimodal import MultiModalConfig
+
+    with _build_attention(MultiModalConfig(mm_encoder_attn_dtype="fp8")) as attn:
+        yield attn
+
+
+@pytest.fixture
+def _make_static_attention(tmp_path):
+    """Create an MMEncoderAttention with static FP8 scales from a file."""
+    from vllm.config.multimodal import MultiModalConfig
+
+    scale_file = tmp_path / "scales.json"
+    scale_file.write_text(
+        json.dumps(
+            {
+                LAYER_0: {"q": 224.0, "k": 198.0, "v": 210.0},
+                LAYER_1: {"q": 100.0, "k": 110.0, "v": 120.0},
+            }
+        )
+    )
+    with _build_attention(
+        MultiModalConfig(
+            mm_encoder_attn_dtype="fp8",
+            mm_encoder_fp8_scale_path=str(scale_file),
+        )
+    ) as attn:
+        yield attn
+
+
+def test_dynamic_scaling_updates_scales(_make_attention) -> None:
+    """Verify that _record_amax_and_update_scales updates scale buffers."""
+    attn = _make_attention
+    if attn is None or not attn.fp8_enabled:
+        pytest.skip("FP8 attention not available (FlashInfer backend required)")
+
+    attn = attn.to("cuda")
+
+    S, H, D = 32, NUM_HEADS, HEAD_DIM
+    q = torch.full((S, H, D), 2.0, device="cuda", dtype=torch.bfloat16)
+    k = torch.full((S, H, D), 3.0, device="cuda", dtype=torch.bfloat16)
+    v = torch.full((S, H, D), 4.0, device="cuda", dtype=torch.bfloat16)
+
+    attn._record_amax_and_update_scales(q, k, v)
+
+    expected_q_scale = 2.0 / _FP8_MAX
+    expected_k_scale = 3.0 / _FP8_MAX
+    expected_v_scale = 4.0 / _FP8_MAX
+
+    torch.testing.assert_close(attn._fp8_q_scale.item(), expected_q_scale)
+    torch.testing.assert_close(attn._fp8_k_scale.item(), expected_k_scale)
+    torch.testing.assert_close(attn._fp8_v_scale.item(), expected_v_scale)
+
+
+def test_circular_buffer_wraps(_make_attention) -> None:
+    """Verify the amax circular buffer wraps at HISTORY_LEN."""
+    attn = _make_attention
+    if attn is None or not attn.fp8_enabled:
+        pytest.skip("FP8 attention not available (FlashInfer backend required)")
+
+    attn = attn.to("cuda")
+    S, H, D = 16, NUM_HEADS, HEAD_DIM
+
+    for i in range(_FP8_AMAX_HISTORY_LEN + 2):
+        mag = float(i + 1)
+        q = torch.full((S, H, D), mag, device="cuda", dtype=torch.bfloat16)
+        k = torch.full((S, H, D), mag, device="cuda", dtype=torch.bfloat16)
+        v = torch.full((S, H, D), mag, device="cuda", dtype=torch.bfloat16)
+        attn._record_amax_and_update_scales(q, k, v)
+
+    assert attn._fp8_amax_pos == 2
+
+    expected_max = float(_FP8_AMAX_HISTORY_LEN + 2)
+    expected_scale = expected_max / _FP8_MAX
+    torch.testing.assert_close(attn._fp8_q_scale.item(), expected_scale)
+
+
+def test_static_scales_loaded(_make_static_attention) -> None:
+    """Verify static scales are loaded from the JSON file."""
+    attn = _make_static_attention
+    if attn is None or not attn.fp8_enabled:
+        pytest.skip("FP8 attention not available (FlashInfer backend required)")
+
+    assert attn.fp8_enabled
+    assert not attn._fp8_dynamic_scale
+
+    # Layer 0 scales (the layer this attention was created with).
+    assert attn._fp8_q_scale.item() == 224.0
+    assert attn._fp8_k_scale.item() == 198.0
+    assert attn._fp8_v_scale.item() == 210.0
+
+    assert not attn.skip_scale_q
+    assert not attn.skip_scale_k
+    assert not attn.skip_scale_v
+
+    # No amax history buffers for static scaling.
+    assert not hasattr(attn, "_fp8_q_amax")
+
+
+def test_static_scales_missing_layer(tmp_path) -> None:
+    """Verify error when requested layer is not in the scale file."""
+    from vllm.config import VllmConfig, set_current_vllm_config
+    from vllm.config.multimodal import MultiModalConfig
+    from vllm.v1.attention.backends.registry import AttentionBackendEnum
+
+    if not is_flashinfer_cudnn_fp8_prefill_attn_supported():
+        pytest.skip("FlashInfer cuDNN not available")
+
+    scale_file = tmp_path / "wrong_layer.json"
+    scale_file.write_text(
+        json.dumps({"visual.blocks.99.attn": {"q": 1.0, "k": 1.0, "v": 1.0}})
+    )
+    mm_config = MultiModalConfig(
+        mm_encoder_attn_dtype="fp8",
+        mm_encoder_fp8_scale_path=str(scale_file),
+    )
+    vllm_config = VllmConfig()
+    vllm_config.model_config = SimpleNamespace(multimodal_config=mm_config)
+
+    from vllm.model_executor.layers.attention.mm_encoder_attention import (
+        MMEncoderAttention,
+    )
+
+    with (
+        set_current_vllm_config(vllm_config),
+        patch(
+            "vllm.model_executor.layers.attention.mm_encoder_attention"
+            ".get_vit_attn_backend",
+            return_value=AttentionBackendEnum.FLASHINFER,
+        ),
+    ):
+        attn = MMEncoderAttention(
+            num_heads=NUM_HEADS,
+            head_size=HEAD_DIM,
+            prefix=LAYER_0,
+        )
+        with pytest.raises(ValueError, match="scales not found for layer"):
+            attn.process_weights_after_loading(torch.bfloat16)
+
+
+def test_dynamic_scales_auto_save(tmp_path) -> None:
+    """Verify scales are saved to disk after the amax buffer fills."""
+    import vllm.model_executor.layers.attention.mm_encoder_attention as _mod
+    from vllm.config.multimodal import MultiModalConfig
+
+    if not is_flashinfer_cudnn_fp8_prefill_attn_supported():
+        pytest.skip("FlashInfer cuDNN not available")
+
+    # Reset module-level state between runs (other tests may have left
+    # state behind after triggering a save).
+    _mod._fp8_scale_save_path = None
+    _mod._fp8_saved_scale_refs.clear()
+
+    save_file = tmp_path / "auto_scales.json"
+    with _build_attention(
+        MultiModalConfig(
+            mm_encoder_attn_dtype="fp8",
+            mm_encoder_fp8_scale_save_path=str(save_file),
+        )
+    ) as attn:
+        if attn is None or not attn.fp8_enabled:
+            pytest.skip("FP8 attention not available")
+
+        attn = attn.to("cuda")
+        S, H, D = 16, NUM_HEADS, HEAD_DIM
+
+        # Run exactly _FP8_AMAX_HISTORY_LEN forward passes.
+        for i in range(_FP8_AMAX_HISTORY_LEN):
+            mag = float(i + 1)
+            q = torch.full((S, H, D), mag, device="cuda", dtype=torch.bfloat16)
+            k = torch.full((S, H, D), mag * 0.5, device="cuda", dtype=torch.bfloat16)
+            v = torch.full((S, H, D), mag * 0.3, device="cuda", dtype=torch.bfloat16)
+            attn._record_amax_and_update_scales(q, k, v)
+
+    # File should have been written on the 16th call (buffer wrap).
+    assert save_file.is_file(), "Scale file was not saved"
+    scales = json.loads(save_file.read_text())
+    assert LAYER_0 in scales
+    assert set(scales[LAYER_0].keys()) == {"q", "k", "v"}
+    for val in scales[LAYER_0].values():
+        assert isinstance(val, float) and val > 0
+
+    # Path is cleared after the one-shot save fires.
+    assert _mod._fp8_scale_save_path is None
@@ -6,6 +6,7 @@ from transformers import AutoModel

 from tests.models.utils import check_embeddings_close
 from vllm import TokensPrompt
+from vllm.config import PoolerConfig


@pytest.mark.parametrize(
@@ -21,6 +22,7 @@ def test_embed_models(hf_runner, vllm_runner, model: str):
    with vllm_runner(
        model,
        runner="pooling",
+        pooler_config=PoolerConfig(task="token_embed"),
        max_model_len=128,
        max_num_batched_tokens=chunk_size,
        enforce_eager=True,
@@ -3,7 +3,6 @@
 import httpx
 import openai
 import pytest
-import pytest_asyncio
 import torch

 from ....utils import RemoteOpenAIServer
@@ -25,29 +24,42 @@ sentences_2 = [
 similarity_reference = [[0.6259, 0.3474], [0.3309, 0.6734]]
 lexical_score_reference = [0.19554901123046875, 0.0]
 colbert_score_reference = [0.7797, 0.4620]
+SUPPORTED_TASKS = ["embed", "token_embed", "token_classify"]
+
+
+@pytest.fixture(scope="module", params=SUPPORTED_TASKS)
+def pooling_task(request):
+    yield request.param


@pytest.fixture(scope="module")
-def server():
+def server(pooling_task):
    args = [
        "--max-model-len",
        str(MAX_MODEL_LEN),
        "--hf-overrides",
        '{"architectures": ["BgeM3EmbeddingModel"]}',
+        "--pooler-config.task",
+        pooling_task,
    ]

    with RemoteOpenAIServer(MODEL_NAME, args) as remote_server:
        yield remote_server


-@pytest_asyncio.fixture
-async def client(server):
-    async with server.get_async_client() as async_client:
-        yield async_client
-
-
@pytest.mark.asyncio
-async def test_bge_m3_api_server_embedding(client: openai.AsyncOpenAI):
+async def test_bge_m3_api_server_embedding(server, pooling_task):
+    client = server.get_async_client()
+
+    if pooling_task != "embed":
+        with pytest.raises(openai.InternalServerError):
+            await run_client_embeddings(
+                client,
+                MODEL_NAME,
+                sentences_1,
+            )
+        return
+
    embeddings_list_1 = await run_client_embeddings(
        client,
        MODEL_NAME,
@@ -117,7 +129,14 @@ def compute_lexical_matching_score(


@pytest.mark.asyncio
-async def test_bge_m3_api_server_sparse_embedding(client: openai.AsyncOpenAI):
+async def test_bge_m3_api_server_sparse_embedding(server, pooling_task):
+    client = server.get_async_client()
+
+    if pooling_task != "token_classify":
+        with pytest.raises(openai.BadRequestError):
+            await sparse_embeddings(client, sentences_1)
+        return
+
    embeddings_1 = await sparse_embeddings(client, sentences_1)
    embeddings_2 = await sparse_embeddings(client, sentences_2)

@@ -137,9 +156,11 @@ async def test_bge_m3_api_server_sparse_embedding(client: openai.AsyncOpenAI):


@pytest.mark.asyncio
-async def test_bge_m3_api_server_sparse_embedding_corner_case(
-    client: openai.AsyncOpenAI,
-):
+async def test_bge_m3_api_server_sparse_embedding_corner_case(server, pooling_task):
+    if pooling_task != "token_classify":
+        return
+
+    client = server.get_async_client()
    embeddings = await sparse_embeddings(client, ["Hi"])
    assert len(embeddings) == 1
    assert 2673 in embeddings[0]
@@ -155,7 +176,18 @@ def colbert_score(q_reps: torch.Tensor, p_reps: torch.Tensor) -> torch.Tensor:


@pytest.mark.asyncio
-async def test_bge_m3_api_server_multi_vector(client: openai.AsyncOpenAI):
+async def test_bge_m3_api_server_multi_vector(server, pooling_task):
+    client = server.get_async_client()
+
+    if pooling_task != "token_embed":
+        with pytest.raises(openai.BadRequestError):
+            await client.post(
+                "../pooling",
+                body={"model": MODEL_NAME, "input": sentences_1, "task": "token_embed"},
+                cast_to=httpx.Response,
+            )
+        return
+
    result_1 = await client.post(
        "../pooling",
        body={"model": MODEL_NAME, "input": sentences_1, "task": "token_embed"},
@@ -4,6 +4,7 @@ import pytest
 import torch

 from vllm import TokensPrompt
+from vllm.config import PoolerConfig


@pytest.mark.parametrize(
@@ -20,6 +21,7 @@ def test_extract_hidden_states(hf_runner, vllm_runner, model: str):
        max_model_len=128,
        enforce_eager=True,
        runner="pooling",
+        pooler_config=PoolerConfig(task="token_embed"),
        enable_prefix_caching=True,
    ) as vllm_model:
        pooling_outputs = vllm_model.llm.encode(
@@ -44,14 +46,3 @@ def test_extract_hidden_states(hf_runner, vllm_runner, model: str):
            assert len(output.prompt_token_ids) == n
            assert len(output.outputs.data) == n
            assert output.num_cached_tokens == 0
-
-        # skip_reading_prefix_cache can still write to cache
-        # to accelerate following requests
-        pooling_outputs = vllm_model.llm.encode(
-            [TokensPrompt(prompt_token_ids=t) for t in token_prompts],
-            pooling_task="embed",
-        )
-
-        for n, output in zip(n_prompt_tokens, pooling_outputs):
-            assert len(output.prompt_token_ids) == n
-            assert output.num_cached_tokens > 0
@@ -5,6 +5,7 @@ import torch
 from transformers import AutoModel

 from tests.models.utils import check_embeddings_close
+from vllm.config import PoolerConfig


@pytest.mark.parametrize(
@@ -17,6 +18,7 @@ def test_embed_models(hf_runner, vllm_runner, example_prompts, model: str, dtype
    with vllm_runner(
        model,
        runner="pooling",
+        pooler_config=PoolerConfig(task="token_embed"),
        max_model_len=None,
    ) as vllm_model:
        vllm_outputs = vllm_model.token_embed(example_prompts)
@@ -146,7 +146,7 @@ def test_multi_vector_retrieval_models_using_normalize(
        model,
        max_model_len=512,
        dtype=dtype,
-        pooler_config=PoolerConfig(use_activation=False),
+        pooler_config=PoolerConfig(use_activation=False, task="token_embed"),
    ) as vllm_model:
        wo_normalize = vllm_model.token_embed(example_prompts)

@@ -154,7 +154,7 @@ def test_multi_vector_retrieval_models_using_normalize(
        model,
        max_model_len=512,
        dtype=dtype,
-        pooler_config=PoolerConfig(use_activation=True),
+        pooler_config=PoolerConfig(use_activation=True, task="token_embed"),
    ) as vllm_model:
        w_normalize = vllm_model.token_embed(example_prompts)

@@ -594,6 +594,9 @@ _TEXT_GENERATION_EXAMPLE_MODELS = {
    "MiMoV2FlashForCausalLM": _HfExamplesInfo(
        "XiaomiMiMo/MiMo-V2-Flash", trust_remote_code=True
    ),
+    "MiMoV2ProForCausalLM": _HfExamplesInfo(
+        "XiaomiMiMo/MiMo-V2.5-Pro", trust_remote_code=True, is_available_online=False
+    ),
    "Dots1ForCausalLM": _HfExamplesInfo("rednote-hilab/dots.llm1.inst"),
 }

@@ -961,6 +964,18 @@ _MULTIMODAL_EXAMPLE_MODELS = {
        "PerceptronAI/Isaac-0.1",
        trust_remote_code=True,
        extras={"0.2-2B-Preview": "PerceptronAI/Isaac-0.2-2B-Preview"},
+        max_transformers_version="4.57",
+        transformers_version_reason={
+            "vllm": (
+                "Custom Isaac code is not compatible with Transformers v5. "
+                "The model should be upstreamed to Transformers for "
+                "long-term support."
+            ),
+            "hf": (
+                "Isaac's remote model and processor code import or configure "
+                "APIs that changed in Transformers v5."
+            ),
+        },
    ),
    "InternS1ForConditionalGeneration": _HfExamplesInfo(
        "internlm/Intern-S1",
@@ -1057,6 +1072,9 @@ _MULTIMODAL_EXAMPLE_MODELS = {
    "MiDashengLMModel": _HfExamplesInfo(
        "mispeech/midashenglm-7b", trust_remote_code=True
    ),
+    "MiMoV2OmniForCausalLM": _HfExamplesInfo(
+        "XiaomiMiMo/MiMo-V2.5-Omni", trust_remote_code=True, is_available_online=False
+    ),
    "MiniCPMO": _HfExamplesInfo(
        "openbmb/MiniCPM-o-2_6",
        trust_remote_code=True,
@@ -1540,6 +1558,18 @@ _SPECULATIVE_DECODING_EXAMPLE_MODELS = {
        trust_remote_code=True,
        speculative_model="XiaomiMiMo/MiMo-7B-RL",
    ),
+    "MiMoV2MTPModel": _HfExamplesInfo(
+        "XiaomiMiMo/MiMo-V2.5-Pro",
+        trust_remote_code=True,
+        speculative_model="XiaomiMiMo/MiMo-V2.5-Pro",
+        is_available_online=False,
+    ),
+    "MiMoV2OmniMTPModel": _HfExamplesInfo(
+        "XiaomiMiMo/MiMo-V2.5-Omni",
+        trust_remote_code=True,
+        speculative_model="XiaomiMiMo/MiMo-V2.5-Omni",
+        is_available_online=False,
+    ),
    "NemotronHMTPModel": _HfExamplesInfo(
        "nvidia/Nemotron-Super-Placeholder",
        speculative_model="nvidia/Nemotron-Super-Placeholder",
@@ -5,6 +5,8 @@ Unit tests for MultiModalRegistry.supports_multimodal_inputs and
 Qwen2.5-VL visual component loading behavior.
 """

+from types import SimpleNamespace
+
 import pytest

 from vllm.multimodal import MULTIMODAL_REGISTRY
@@ -32,3 +34,17 @@ def test_supports_multimodal_inputs(model_id, limit_mm_per_prompt, expected):
        limit_mm_per_prompt=limit_mm_per_prompt,
    )
    assert MULTIMODAL_REGISTRY.supports_multimodal_inputs(ctx.model_config) is expected
+
+
+def test_create_processor_error_uses_served_model_name():
+    model_config = SimpleNamespace(
+        is_multimodal_model=False,
+        model="/path/to/model/weights",
+        served_model_name="friendly-model-name",
+    )
+
+    with pytest.raises(
+        ValueError,
+        match="friendly-model-name is not a multimodal model",
+    ):
+        MULTIMODAL_REGISTRY.create_processor(model_config)
@@ -0,0 +1,185 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""Tests for Cutlass W4A16 (Machete) kernel on Hopper.
+
+Verifies that W4A16 quantized models loaded through vllm select the
+MacheteLinearKernel on sm_90 GPUs, that weights are correctly repacked,
+and that inference produces valid output.
+
+Run `pytest tests/quantization/test_cutlass_w4a16.py`.
+"""
+
+import pytest
+import torch
+
+from vllm.platforms import current_platform
+
+if not current_platform.has_device_capability(90):
+    pytest.skip(
+        "Machete W4A16 requires Hopper (sm_90).",
+        allow_module_level=True,
+    )
+
+from vllm.model_executor.kernels.linear import (
+    MPLinearLayerConfig,
+    choose_mp_linear_kernel,
+)
+from vllm.model_executor.kernels.linear.mixed_precision import (
+    MacheteLinearKernel,
+)
+from vllm.model_executor.layers.quantization.compressed_tensors.compressed_tensors import (  # noqa: E501
+    CompressedTensorsLinearMethod,
+    CompressedTensorsWNA16,
+)
+from vllm.scalar_type import scalar_types
+
+
+@pytest.fixture(scope="function", autouse=True)
+def enable_pickle(monkeypatch):
+    """`LLM.apply_model` requires pickling a function."""
+    monkeypatch.setenv("VLLM_ALLOW_INSECURE_SERIALIZATION", "1")
+
+
+@pytest.mark.parametrize(
+    "act_type,weight_type,group_size,zero_points",
+    [
+        (torch.float16, scalar_types.uint4b8, 128, False),
+        (torch.bfloat16, scalar_types.uint4b8, 128, False),
+        (torch.float16, scalar_types.uint4, 128, True),
+        (torch.float16, scalar_types.uint4b8, -1, False),
+    ],
+    ids=[
+        "fp16-gptq-g128",
+        "bf16-gptq-g128",
+        "fp16-awq-g128",
+        "fp16-channelwise",
+    ],
+)
+def test_machete_kernel_selected(act_type, weight_type, group_size, zero_points):
+    """Verify choose_mp_linear_kernel picks MacheteLinearKernel."""
+    config = MPLinearLayerConfig(
+        full_weight_shape=(4096, 4096),
+        partition_weight_shape=(4096, 4096),
+        act_type=act_type,
+        weight_type=weight_type,
+        group_size=group_size,
+        zero_points=zero_points,
+        has_g_idx=False,
+    )
+    kernel = choose_mp_linear_kernel(config)
+    assert kernel is MacheteLinearKernel, (
+        f"Expected MacheteLinearKernel, got {kernel.__name__}"
+    )
+
+
+@pytest.mark.parametrize(
+    "full_shape,part_shape,weight_type,group_size,has_g_idx,expected_reason",
+    [
+        ((4096, 4096), (2048, 4096), scalar_types.uint4b8, 128, True, "Act reordering"),
+        (
+            (4096, 4096),
+            (4096, 4096),
+            scalar_types.float6_e3m2f,
+            128,
+            False,
+            "Quant type",
+        ),
+        ((4096, 4096), (4096, 4096), scalar_types.uint4b8, 32, False, "Group size"),
+    ],
+    ids=["partitioned-g_idx", "unsupported-quant-type", "unsupported-group-size"],
+)
+def test_machete_rejects_invalid_config(
+    full_shape, part_shape, weight_type, group_size, has_g_idx, expected_reason
+):
+    """Verify Machete rejects unsupported configurations."""
+    config = MPLinearLayerConfig(
+        full_weight_shape=full_shape,
+        partition_weight_shape=part_shape,
+        act_type=torch.float16,
+        weight_type=weight_type,
+        group_size=group_size,
+        zero_points=False,
+        has_g_idx=has_g_idx,
+    )
+    can_impl, reason = MacheteLinearKernel.can_implement(config)
+    assert not can_impl
+    assert expected_reason in reason
+
+
+def test_kernel_selection_with_disabled_machete(monkeypatch):
+    """Verify kernel selection falls back when Machete is disabled."""
+    monkeypatch.setattr("vllm.envs.VLLM_DISABLED_KERNELS", ["MacheteLinearKernel"])
+
+    config = MPLinearLayerConfig(
+        full_weight_shape=(4096, 4096),
+        partition_weight_shape=(4096, 4096),
+        act_type=torch.float16,
+        weight_type=scalar_types.uint4b8,
+        group_size=128,
+        zero_points=False,
+        has_g_idx=False,
+    )
+    kernel = choose_mp_linear_kernel(config)
+    assert kernel is not MacheteLinearKernel, "MacheteLinearKernel should be disabled"
+
+
+@pytest.mark.parametrize(
+    "model_name",
+    [
+        "nm-testing/tinyllama-oneshot-w4a16-channel-v2",
+        "nm-testing/TinyLlama-1.1B-Chat-v1.0-W4A16-G128-Asym-Updated-ActOrder",
+    ],
+)
+def test_w4a16_machete_e2e(vllm_runner, model_name):
+    """Load a W4A16 model, verify Machete kernel is used, and generate."""
+    with vllm_runner(model_name, enforce_eager=True, gpu_memory_utilization=0.5) as llm:
+
+        def check_model(model):
+            layer = model.model.layers[0]
+            qkv_proj = layer.self_attn.qkv_proj
+
+            assert isinstance(qkv_proj.quant_method, CompressedTensorsLinearMethod)
+            assert isinstance(qkv_proj.scheme, CompressedTensorsWNA16)
+            assert isinstance(qkv_proj.scheme.kernel, MacheteLinearKernel), (
+                f"Expected MacheteLinearKernel on Hopper, "
+                f"got {type(qkv_proj.scheme.kernel).__name__}"
+            )
+
+            assert hasattr(qkv_proj, "weight_packed")
+            assert hasattr(qkv_proj, "weight_scale")
+            assert qkv_proj.weight_packed.dtype == torch.int32
+
+        llm.apply_model(check_model)
+
+        output = llm.generate_greedy("Hello my name is", max_tokens=10)
+        assert output
+        assert len(output[0][1]) > 0
+
+
+def test_w4a16_machete_bfloat16_deterministic(vllm_runner):
+    """Verify Machete works with bf16 activations and is deterministic."""
+    model_name = "nm-testing/tinyllama-oneshot-w4a16-channel-v2"
+    prompt = "The capital of France is"
+
+    with vllm_runner(
+        model_name,
+        enforce_eager=True,
+        dtype="bfloat16",
+        gpu_memory_utilization=0.5,
+    ) as llm:
+
+        def check_kernel_type(model):
+            layer = model.model.layers[0]
+            scheme = layer.self_attn.qkv_proj.scheme
+            assert isinstance(scheme.kernel, MacheteLinearKernel), (
+                f"Expected MacheteLinearKernel with bf16, "
+                f"got {type(scheme.kernel).__name__}"
+            )
+
+        llm.apply_model(check_kernel_type)
+
+        out1 = llm.generate_greedy(prompt, max_tokens=10)
+        out2 = llm.generate_greedy(prompt, max_tokens=10)
+        assert out1[0][1] == out2[0][1], (
+            f"Non-deterministic: '{out1[0][1]}' vs '{out2[0][1]}'"
+        )
@@ -188,30 +188,6 @@ class TestExtractToolCalls:
            "location": "NYC"
        }

-    def test_type_conversion_in_non_streaming(self):
-        """Non-streaming extraction must convert params using the tool schema."""
-        tool = ChatCompletionToolsParam(
-            function=FunctionDefinition(
-                name="toggle",
-                parameters={
-                    "type": "object",
-                    "properties": {
-                        "enabled": {"type": "boolean"},
-                        "count": {"type": "integer"},
-                    },
-                },
-            ),
-        )
-        parser = make_parser(tools=[tool])
-        model_output = build_tool_call("toggle", {"enabled": "true", "count": "42"})
-        result = parser.extract_tool_calls(model_output, None)
-        assert result.tools_called
-        assert len(result.tool_calls) == 1
-        args = json.loads(result.tool_calls[0].function.arguments)
-        assert args == {"enabled": True, "count": 42}
-        assert isinstance(args["enabled"], bool)
-        assert isinstance(args["count"], int)
-

 # ---------------------------------------------------------------------------
 # Tests: extract_tool_calls_streaming
@@ -2074,54 +2074,6 @@ def test_auto_fit_max_model_len_not_triggered():
    assert vllm_config.model_config.max_model_len == 16


-def test_auto_fit_max_model_len_respects_num_gpu_blocks_override():
-    """Auto-fit must size max_model_len against the override-clamped pool, not
-    the raw `available_memory`. Without this, auto-fit could pick a
-    max_model_len that no longer fits once `num_gpu_blocks_override` is applied.
-    """
-    model_config = ModelConfig(max_model_len=16384)
-    model_config.original_max_model_len = -1  # request auto-fit
-    vllm_config = VllmConfig(model_config=model_config)
-    # Cap the cache to 32 blocks regardless of available memory.
-    vllm_config.cache_config.num_gpu_blocks_override = 32
-
-    mem_per_block_per_layer = 16 * 2 * 64 * 4 * 2
-    kv_cache_specs = {
-        "layer_1": new_kv_cache_spec(),  # block_size=16
-        "layer_2": new_kv_cache_spec(),
-    }
-    # Plenty of raw memory (1024 blocks per layer would fit max_model_len=16384).
-    large_available_memory = mem_per_block_per_layer * 2 * 1024
-
-    get_kv_cache_configs(vllm_config, [kv_cache_specs], [large_available_memory])
-
-    # 32 blocks * block_size 16 = 512 token slots, so max_model_len must
-    # auto-fit at or below that.
-    assert 0 < vllm_config.model_config.max_model_len <= 32 * 16
-
-
-def test_check_enough_kv_cache_memory_respects_num_gpu_blocks_override():
-    """Admission check must use the override-clamped pool size, not raw
-    `available_memory`. Without this, startup could accept a max_model_len
-    that does not actually fit in `num_gpu_blocks_override` blocks.
-    """
-    model_config = ModelConfig(max_model_len=16384)
-    vllm_config = VllmConfig(model_config=model_config)
-    # 32 blocks is far too small for max_model_len=16384 (would need 1024).
-    vllm_config.cache_config.num_gpu_blocks_override = 32
-
-    mem_per_block_per_layer = 16 * 2 * 64 * 4 * 2
-    kv_cache_specs = {
-        "layer_1": new_kv_cache_spec(),
-        "layer_2": new_kv_cache_spec(),
-    }
-    # Plenty of raw memory: a bytes-only check against this would pass.
-    large_available_memory = mem_per_block_per_layer * 2 * 1024
-
-    with pytest.raises(ValueError, match="max seq len"):
-        get_kv_cache_configs(vllm_config, [kv_cache_specs], [large_available_memory])
-
-
 def test_unify_hybrid_kv_cache_specs():
    # 1. has_full_attention and has_sliding_window
    before_spec_1 = new_kv_cache_spec()
@@ -432,52 +432,3 @@ def test_chunked_local_attention_get_num_blocks_to_allocate():
        )
        == 15
    )
-
-
-def test_predictor_matches_allocator_blocks_calculation_with_admission_cap():
-    """In forward steps, `get_num_blocks_to_allocate` must return exactly what
-    `allocate_new_blocks` will pull; otherwise `block_pool.get_new_blocks`
-    raises `ValueError: Cannot get N free blocks from the pool`.
-    """
-    block_size = 2
-    sliding_window = 8  # 4-block live window
-    cap = sliding_window // block_size
-
-    spec = SlidingWindowSpec(
-        block_size=block_size,
-        num_kv_heads=1,
-        head_size=1,
-        dtype=torch.float32,
-        sliding_window=sliding_window,
-    )
-    block_pool = BlockPool(
-        num_gpu_blocks=100, enable_caching=True, hash_block_size=block_size
-    )
-    manager = SlidingWindowManager(
-        spec,
-        block_pool=block_pool,
-        enable_caching=False,
-        kv_cache_group_id=0,
-        max_admission_blocks_per_request=cap,
-    )
-
-    request_id = "req"
-    total_computed = 0
-    # Walk through request forward steps. Check num_blocks returned by
-    # `get_num_blocks_to_allocate` matches what `allocate_new_blocks` pulls
-    for num_tokens in (4, 8, 12, 16):
-        predicted = manager.get_num_blocks_to_allocate(
-            request_id=request_id,
-            num_tokens=num_tokens,
-            new_computed_blocks=[],
-            total_computed_tokens=total_computed,
-            num_tokens_main_model=num_tokens,
-        )
-        new_blocks = manager.allocate_new_blocks(
-            request_id, num_tokens=num_tokens, num_tokens_main_model=num_tokens
-        )
-        assert predicted == len(new_blocks), (
-            f"num_tokens={num_tokens}: predictor returned {predicted} "
-            f"but allocator pulled {len(new_blocks)}"
-        )
-        total_computed = num_tokens
@@ -324,13 +324,10 @@ def run_test(
 ):
    spec_decoding = spec_config is not None
    cache_arg: dict[str, Any] = (
-        # Force preemptions: with 32 blocks the cache holds at most a single
-        # max-length request, so the ~34 concurrent prompts contend and trigger
-        # preemption. (Prompts here are << max_model_len, so dropping
-        # max_model_len from 4096 to 512 doesn't change generation behavior.)
-        dict(num_gpu_blocks_override=32, max_model_len=512)
+        # Force preemptions
+        dict(num_gpu_blocks_override=32)
        if test_preemption
-        else dict(gpu_memory_utilization=0.9, max_model_len=4096)
+        else dict(gpu_memory_utilization=0.9)
    )
    spec_mml = (spec_config or {}).get("max_model_len")
    spec_method = (spec_config or {}).get("method", "none")
@@ -346,6 +343,7 @@ def run_test(

    with VllmRunner(
        model,
+        max_model_len=4096,
        enable_chunked_prefill=test_prefill_chunking,
        # Force prefill chunking
        max_num_batched_tokens=48 if test_prefill_chunking else None,
@@ -8,16 +8,11 @@ from unittest.mock import Mock
 import pytest

 from vllm.config import ModelConfig, SchedulerConfig, VllmConfig
+from vllm.reasoning import ReasoningParser
 from vllm.v1.request import Request
 from vllm.v1.structured_output import StructuredOutputManager


-class MockReasoner:
-    def __init__(self, tokenizer):
-        self.is_reasoning_end = Mock(return_value=False)
-        self.is_reasoning_end_streaming = Mock(return_value=False)
-
-
 class TestReasoningStructuredOutput:
    """Test reasoning-aware structured output functionality."""

@@ -55,6 +50,13 @@ class TestReasoningStructuredOutput:
        config.speculative_config = None
        return config

+    @pytest.fixture
+    def mock_reasoning_parser(self):
+        """Create a mock ReasoningParser."""
+        parser = Mock(spec=ReasoningParser)
+        parser.is_reasoning_end = Mock(return_value=False)
+        return parser
+
    @pytest.fixture
    def mock_request_with_structured_output(self):
        """Create a mock request with structured output."""
@@ -62,8 +64,6 @@ class TestReasoningStructuredOutput:
        request.structured_output_request = Mock()
        request.structured_output_request.reasoning_ended = None
        request.structured_output_request.grammar = Mock()
-        request.structured_output_request.reasoning_parser_kwargs = None
-        request.structured_output_request.reasoner = None
        request.structured_output_request.grammar.is_terminated = Mock(
            return_value=False
        )
@@ -74,13 +74,6 @@ class TestReasoningStructuredOutput:
        request.num_output_placeholders = 0
        return request

-    @pytest.fixture
-    def manager_with_reasoner(self, mock_vllm_config):
-        manager = StructuredOutputManager(mock_vllm_config)
-        manager.reasoner_cls = MockReasoner
-        manager.tokenizer = Mock()
-        return manager
-
    def test_should_fill_bitmask_with_enable_in_reasoning(
        self, mock_vllm_config, mock_request_with_structured_output
    ):
@@ -96,17 +89,22 @@ class TestReasoningStructuredOutput:

    def test_should_fill_bitmask_without_enable_in_reasoning(
        self,
-        manager_with_reasoner,
+        mock_vllm_config,
        mock_request_with_structured_output,
+        mock_reasoning_parser,
    ):
        """Test should_fill_bitmask when enable_in_reasoning is False."""
        # Keep enable_in_reasoning as False (default)
-        config = manager_with_reasoner.vllm_config.structured_outputs_config
+        config = mock_vllm_config.structured_outputs_config
        assert config.enable_in_reasoning is False

-        result = manager_with_reasoner.should_fill_bitmask(
-            mock_request_with_structured_output
-        )
+        manager = StructuredOutputManager(mock_vllm_config)
+        manager.reasoner = mock_reasoning_parser
+
+        # Mock reasoning not ended
+        mock_reasoning_parser.is_reasoning_end.return_value = False
+
+        result = manager.should_fill_bitmask(mock_request_with_structured_output)

        # Should set reasoning_ended and return its value
        assert (
@@ -120,92 +118,68 @@ class TestReasoningStructuredOutput:
    ):
        """Test should_fill_bitmask when no reasoner is configured."""
        manager = StructuredOutputManager(mock_vllm_config)
+        manager.reasoner = None

        result = manager.should_fill_bitmask(mock_request_with_structured_output)

        # Should default to True when no reasoner
        assert result is True

-    def test_should_fill_bitmask_uses_request_reasoning_parser_kwargs(
-        self, mock_vllm_config, mock_request_with_structured_output
-    ):
-        """Test request-level parser kwargs override the default reasoner."""
-
-        class KwargReasoner:
-            def __init__(self, tokenizer, chat_template_kwargs=None):
-                self.chat_template_kwargs = chat_template_kwargs or {}
-
-            def is_reasoning_end(self, input_ids):
-                return not self.chat_template_kwargs.get("enable_thinking", False)
-
-        manager = StructuredOutputManager(mock_vllm_config)
-        manager.reasoner_cls = KwargReasoner
-        manager.tokenizer = Mock()
-
-        structured_req = mock_request_with_structured_output.structured_output_request
-        structured_req.reasoning_parser_kwargs = {
-            "chat_template_kwargs": {"enable_thinking": True}
-        }
-
-        result = manager.should_fill_bitmask(mock_request_with_structured_output)
-
-        assert result is False
-        assert (
-            mock_request_with_structured_output.structured_output_request.reasoner
-            is not None
-        )
-
    def test_should_advance_with_enable_in_reasoning(
        self,
-        manager_with_reasoner,
+        mock_vllm_config,
        mock_request_with_structured_output,
+        mock_reasoning_parser,
    ):
        """Test should_advance when enable_in_reasoning is True."""
        # Enable enable_in_reasoning
-        manager_with_reasoner.enable_in_reasoning = True
+        mock_vllm_config.structured_outputs_config.enable_in_reasoning = True
+
+        manager = StructuredOutputManager(mock_vllm_config)
+        manager.reasoner = mock_reasoning_parser

        # Should always return True when enable_in_reasoning is enabled
-        result = manager_with_reasoner.should_advance(
-            mock_request_with_structured_output
-        )
+        result = manager.should_advance(mock_request_with_structured_output)
        assert result is True

    def test_should_advance_reasoning_not_ended(
        self,
-        manager_with_reasoner,
+        mock_vllm_config,
        mock_request_with_structured_output,
+        mock_reasoning_parser,
    ):
        """Test should_advance when reasoning has not ended."""
+        manager = StructuredOutputManager(mock_vllm_config)
+        manager.reasoner = mock_reasoning_parser
+
        # Set reasoning as not ended
        (
            mock_request_with_structured_output.structured_output_request
        ).reasoning_ended = False
+        mock_reasoning_parser.is_reasoning_end.return_value = False

-        result = manager_with_reasoner.should_advance(
-            mock_request_with_structured_output
-        )
+        result = manager.should_advance(mock_request_with_structured_output)

        # Should return False since reasoning hasn't ended
        assert result is False

    def test_should_advance_reasoning_just_ended(
        self,
-        manager_with_reasoner,
+        mock_vllm_config,
        mock_request_with_structured_output,
+        mock_reasoning_parser,
    ):
        """Test should_advance when reasoning ends in current step."""
+        manager = StructuredOutputManager(mock_vllm_config)
+        manager.reasoner = mock_reasoning_parser
+
        # Set reasoning as not ended initially, but ends in this step
        (
            mock_request_with_structured_output.structured_output_request
        ).reasoning_ended = False
-        reasoner = MockReasoner(tokenizer=Mock())
-        reasoner.is_reasoning_end_streaming.return_value = True
-        structured_req = mock_request_with_structured_output.structured_output_request
-        structured_req.reasoner = reasoner
+        mock_reasoning_parser.is_reasoning_end.return_value = True

-        result = manager_with_reasoner.should_advance(
-            mock_request_with_structured_output
-        )
+        result = manager.should_advance(mock_request_with_structured_output)

        # Should set reasoning_ended to True but return False for this step
        assert (
@@ -216,18 +190,20 @@ class TestReasoningStructuredOutput:

    def test_should_advance_reasoning_already_ended(
        self,
-        manager_with_reasoner,
+        mock_vllm_config,
        mock_request_with_structured_output,
+        mock_reasoning_parser,
    ):
        """Test should_advance when reasoning has already ended."""
+        manager = StructuredOutputManager(mock_vllm_config)
+        manager.reasoner = mock_reasoning_parser
+
        # Set reasoning as already ended
        (
            mock_request_with_structured_output.structured_output_request
        ).reasoning_ended = True

-        result = manager_with_reasoner.should_advance(
-            mock_request_with_structured_output
-        )
+        result = manager.should_advance(mock_request_with_structured_output)

        # Should return True since reasoning has ended
        assert result is True
@@ -326,6 +326,10 @@ class ModelConfig:
    mm_encoder_only: InitVar[bool | None] = None
    mm_encoder_tp_mode: InitVar[MMEncoderTPMode | None] = None
    mm_encoder_attn_backend: InitVar[AttentionBackendEnum | str | None] = None
+    mm_encoder_attn_dtype: InitVar[str | None] = None
+    mm_encoder_fp8_scale_path: InitVar[str | None] = None
+    mm_encoder_fp8_scale_save_path: InitVar[str | None] = None
+    mm_encoder_fp8_scale_save_margin: InitVar[float | None] = None
    interleave_mm_strings: InitVar[bool | None] = None
    skip_mm_profiling: InitVar[bool | None] = None
    video_pruning_rate: InitVar[float | None] = None
@@ -447,6 +451,10 @@ class ModelConfig:
        mm_encoder_only: bool | None,
        mm_encoder_tp_mode: MMEncoderTPMode | None,
        mm_encoder_attn_backend: AttentionBackendEnum | str | None,
+        mm_encoder_attn_dtype: str | None,
+        mm_encoder_fp8_scale_path: str | None,
+        mm_encoder_fp8_scale_save_path: str | None,
+        mm_encoder_fp8_scale_save_margin: float | None,
        interleave_mm_strings: bool | None,
        skip_mm_profiling: bool | None,
        video_pruning_rate: float | None,
@@ -513,6 +521,7 @@ class ModelConfig:
        if dict_overrides:
            self._apply_dict_overrides(hf_config, dict_overrides)
        self.hf_text_config = get_hf_text_config(self.hf_config)
+        self.model_arch_config = self.get_model_arch_config()
        self.attention_chunk_size = getattr(
            self.hf_text_config, "attention_chunk_size", None
        )
@@ -520,7 +529,6 @@ class ModelConfig:
        self.hf_image_processor_config = get_hf_image_processor_config(
            self.model, hf_token=self.hf_token, revision=self.revision
        )
-        self.model_arch_config = self.get_model_arch_config()

        architectures = self.architectures
        registry = self.registry
@@ -643,6 +651,10 @@ class ModelConfig:
                mm_encoder_only=mm_encoder_only,
                mm_encoder_tp_mode=mm_encoder_tp_mode,
                mm_encoder_attn_backend=mm_encoder_attn_backend,
+                mm_encoder_attn_dtype=mm_encoder_attn_dtype,
+                mm_encoder_fp8_scale_path=mm_encoder_fp8_scale_path,
+                mm_encoder_fp8_scale_save_path=mm_encoder_fp8_scale_save_path,
+                mm_encoder_fp8_scale_save_margin=mm_encoder_fp8_scale_save_margin,
                interleave_mm_strings=interleave_mm_strings,
                skip_mm_profiling=skip_mm_profiling,
                video_pruning_rate=video_pruning_rate,
@@ -2,6 +2,7 @@
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project

 from collections.abc import Mapping
+from pathlib import Path
 from typing import Any, Literal, TypeAlias, TypedDict, final

 from pydantic import ConfigDict, Field, field_validator, model_validator
@@ -158,6 +159,24 @@ class MultiModalConfig:
    """Optional override for the multi-modal encoder attention backend when
    using vision transformers. Accepts any value from
    `vllm.v1.attention.backends.registry.AttentionBackendEnum` (e.g. `FLASH_ATTN`)."""
+    mm_encoder_attn_dtype: Literal["fp8"] | None = None
+    """Optional dtype override for ViT encoder attention. Set to `"fp8"` to
+    enable FP8 quantization via the FlashInfer cuDNN backend. When set to
+    `"fp8"` without a scale file, dynamic scaling is used automatically.
+    See docs/features/quantization/fp8_vit_attn.md for details."""
+    mm_encoder_fp8_scale_path: str | None = None
+    """Path to a JSON file containing per-layer FP8 Q/K/V scales for ViT
+    encoder attention. When provided (with `mm_encoder_attn_dtype="fp8"`),
+    static scaling is used. When omitted, dynamic scaling is used."""
+    mm_encoder_fp8_scale_save_path: str | None = None
+    """When set with dynamic FP8 scaling (`mm_encoder_attn_dtype="fp8"`
+    and no `mm_encoder_fp8_scale_path`), saves the calibrated scales to
+    this file after the amax history buffer is full. The saved file can
+    then be used as `mm_encoder_fp8_scale_path` in subsequent runs."""
+    mm_encoder_fp8_scale_save_margin: float = Field(default=1.5, gt=0.0)
+    """Safety margin multiplied onto scales when auto-saving. A value > 1
+    leaves headroom so that inputs with larger activations than the
+    calibration set do not overflow FP8 range. Default 1.5."""
    interleave_mm_strings: bool = False
    """Enable fully interleaved support for multimodal prompts, while using
    --chat-template-content-format=string."""
@@ -233,6 +252,36 @@ class MultiModalConfig:
                "'mm_shm_cache_max_object_size_mb' should only be set when "
                "'mm_processor_cache_type' is 'shm'."
            )
+        # Validate FP8 scale path combinations.
+        if self.mm_encoder_attn_dtype != "fp8" and (
+            self.mm_encoder_fp8_scale_path is not None
+            or self.mm_encoder_fp8_scale_save_path is not None
+        ):
+            raise ValueError(
+                "'mm_encoder_fp8_scale_path' and "
+                "'mm_encoder_fp8_scale_save_path' require "
+                "'mm_encoder_attn_dtype' to be 'fp8'."
+            )
+        if (
+            self.mm_encoder_fp8_scale_path is not None
+            and self.mm_encoder_fp8_scale_save_path is not None
+        ):
+            raise ValueError(
+                "'mm_encoder_fp8_scale_save_path' cannot be used with "
+                "'mm_encoder_fp8_scale_path' (saving requires dynamic scaling)."
+            )
+
+        # Validate file paths exist.
+        if self.mm_encoder_fp8_scale_path is not None:
+            scale_path = Path(self.mm_encoder_fp8_scale_path)
+            if not scale_path.is_file():
+                raise FileNotFoundError(f"FP8 scale file not found: {scale_path}")
+        if self.mm_encoder_fp8_scale_save_path is not None:
+            save_parent = Path(self.mm_encoder_fp8_scale_save_path).parent
+            if not save_parent.is_dir():
+                raise FileNotFoundError(
+                    f"Parent directory for FP8 scale save path not found: {save_parent}"
+                )
        return self

    def compute_hash(self) -> str:
@@ -252,6 +301,8 @@ class MultiModalConfig:
            if self.mm_encoder_attn_backend is not None
            else None,
            self.mm_encoder_tp_mode,
+            self.mm_encoder_attn_dtype,
+            self.mm_encoder_fp8_scale_path,
        ]
        hash_str = safe_hash(str(factors).encode(), usedforsecurity=False).hexdigest()
        return hash_str
@@ -34,6 +34,7 @@ logger = init_logger(__name__)
 MTPModelTypes = Literal[
    "deepseek_mtp",
    "mimo_mtp",
+    "mimo_v2_mtp",
    "glm4_moe_mtp",
    "glm4_moe_lite_mtp",
    "glm_ocr_mtp",
@@ -63,7 +64,8 @@ SpeculativeMethod = Literal[
    EagleModelTypes,
    NgramGPUTypes,
 ]
-RejectionSampleMethod = Literal["strict", "probabilistic", "synthetic"]
+RejectionSampleMethod = Literal["standard", "synthetic"]
+DraftSampleMethod = Literal["greedy", "gumbel"]


@config
@@ -183,11 +185,11 @@ class SpeculativeConfig:
    """Load config for the draft model. If not specified, will use the load
    config from the target model."""

-    rejection_sample_method: RejectionSampleMethod = "strict"
-    """Whether to use strict (target and draft sampled tokens match exactly)
-    or probabilistic rejection sampling. Both respect the target model
-    distribution, but the latter yields a higher acceptance rate at the cost
-    of more memory to cache draft logits."""
+    rejection_sample_method: RejectionSampleMethod = "standard"
+    """The rejection sampling method to use. 'standard' uses probabilistic
+    rejection sampling (with or without cached draft logits, controlled by
+    draft_sample_method). 'synthetic' accepts draft tokens with a decaying
+    probability calibrated to synthetic_acceptance_rate."""

    synthetic_acceptance_rates: list[float] | None = None
    """Per-position *unconditional* acceptance rates for synthetic rejection
@@ -248,6 +250,14 @@ class SpeculativeConfig:
            )
        return SpeculativeConfig._acceptance_length_to_rates(length, n)

+    draft_sample_method: DraftSampleMethod = "greedy"
+    """How the draft model samples tokens. 'greedy' always picks the argmax
+    token, and the draft probabilities are treated as one-hot during rejection
+    sampling. 'gumbel' adds Gumbel noise for stochastic sampling, and the full
+    draft logits are used for the probability ratio test during rejection
+    sampling. This comes at the cost of additional GPU memory usage. This
+    parameter currently only applies to Model Runner V2."""
+
    def compute_hash(self) -> str:
        """
        WARNING: Whenever a new field is added to this config,
@@ -323,6 +333,48 @@ class SpeculativeConfig:
                }
            )

+        if (arch := hf_config.architectures[0]) in (
+            "MiMoV2ProForCausalLM",
+            "MiMoV2OmniForCausalLM",
+        ):
+            from vllm.model_executor.models.mimo_v2_mtp import (
+                _MIMO_V2_PRO_NUM_MTP_LAYERS,
+            )
+
+            mtp_arch_maps = {
+                "MiMoV2ProForCausalLM": "MiMoV2MTPModel",
+                "MiMoV2OmniForCausalLM": "MiMoV2OmniMTPModel",
+            }
+
+            hf_config.model_type = "mimo_v2_mtp"
+            # vLLM currently supports only the first MiMo-V2 MTP layer.
+            n_predict = _MIMO_V2_PRO_NUM_MTP_LAYERS
+            hf_config.update(
+                {
+                    "num_hidden_layers": 0,
+                    "n_predict": n_predict,
+                    "num_nextn_predict_layers": n_predict,
+                    "architectures": [mtp_arch_maps[arch]],
+                }
+            )
+
+        if hf_config.architectures[0] == "MiMoV2FlashForCausalLM":
+            from vllm.model_executor.models.mimo_v2_mtp import (
+                _MIMO_V2_FLASH_NUM_MTP_LAYERS,
+            )
+
+            hf_config.model_type = "mimo_v2_mtp"
+            # vLLM currently supports only the first MiMo-V2 MTP layer.
+            n_predict = _MIMO_V2_FLASH_NUM_MTP_LAYERS
+            hf_config.update(
+                {
+                    "num_hidden_layers": 0,
+                    "n_predict": n_predict,
+                    "num_nextn_predict_layers": n_predict,
+                    "architectures": ["MiMoV2MTPModel"],
+                }
+            )
+
        if hf_config.architectures[0] == "Glm4MoeForCausalLM":
            hf_config.model_type = "glm4_moe_mtp"
            n_predict = getattr(hf_config, "num_nextn_predict_layers", None)
@@ -1432,10 +1432,6 @@ class VllmConfig:
        cudagraph_capture_sizes = [1, 2, 4] + list(range(8, 256, 8)) + list(
            range(256, max_graph_size + 1, 16))

-        `max_num_batched_tokens` is also appended to the list if it fits
-        within `max_cudagraph_capture_size`, so the max batch size is captured
-        even when off-stride.
-
        In the end, `vllm_config.compilation_config.cudagraph_capture_sizes`
        will be the final sizes to capture cudagraph (in ascending order).

@@ -1524,12 +1520,6 @@ class VllmConfig:
                    cudagraph_capture_sizes += list(
                        range(256, max_cudagraph_capture_size + 1, 16)
                    )
-                # ensure max_num_tokens is captured if within max capture size
-                if (
-                    max_num_tokens <= max_cudagraph_capture_size
-                    and max_num_tokens not in cudagraph_capture_sizes
-                ):
-                    cudagraph_capture_sizes.append(max_num_tokens)
                # de-duplicate and sort the sizes
                cudagraph_capture_sizes = sorted(set(cudagraph_capture_sizes))

@@ -584,8 +584,6 @@ class FlashInferNVLinkOneSidedManager(All2AllManagerBase):
        top_k: int,
        num_experts: int,
        hidden_size: int,
-        dispatch_dtype_bytes_per_elem: int = 0,
-        dispatch_scale_bytes_per_token: int = 0,
    ):
        """Initialize the MoeAlltoAll workspace."""
        if self.initialized:
@@ -616,13 +614,9 @@ class FlashInferNVLinkOneSidedManager(All2AllManagerBase):
        ep_config = MnnvlConfig(
            comm_backend=CustomCommunicator(self.cpu_group),
        )
-        if dispatch_dtype_bytes_per_elem == 0:
-            hidden_bytes = hidden_size // 2
-        else:
-            hidden_bytes = hidden_size * dispatch_dtype_bytes_per_elem
        total_dispatch_payload_size_per_token = (
-            hidden_bytes
-            + dispatch_scale_bytes_per_token
+            hidden_size // 2  # nvfp4 hidden states
+            + hidden_size // 16  # fp8 scaling factors
            + top_k * 4  # int32 topks ids
            + top_k * 4  # float32 topk weights
        )
@@ -2,6 +2,7 @@
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 """Scheduler-side logic for the NIXL connector."""

+import os
 import threading
 import time
 from typing import TYPE_CHECKING, Any
@@ -88,6 +89,12 @@ class NixlConnectorScheduler:
        if vllm_config.scheduler_config.disable_hybrid_kv_cache_manager:
            logger.info("Hybrid Memory Allocator is enabled with NIXL")

+        if os.environ.get("VLLM_NIXL_ABORT_REQUEST_TIMEOUT") is not None:
+            logger.warning(
+                "VLLM_NIXL_ABORT_REQUEST_TIMEOUT is deprecated and will be "
+                "removed in release 0.22.0."
+            )
+
        # Background thread for handling new handshake requests.
        self._nixl_handshake_listener_t: threading.Thread | None = None
        self._stop_event = threading.Event()
@@ -542,6 +542,14 @@ class EngineArgs:
    mm_encoder_attn_backend: AttentionBackendEnum | str | None = (
        MultiModalConfig.mm_encoder_attn_backend
    )
+    mm_encoder_attn_dtype: str | None = MultiModalConfig.mm_encoder_attn_dtype
+    mm_encoder_fp8_scale_path: str | None = MultiModalConfig.mm_encoder_fp8_scale_path
+    mm_encoder_fp8_scale_save_path: str | None = (
+        MultiModalConfig.mm_encoder_fp8_scale_save_path
+    )
+    mm_encoder_fp8_scale_save_margin: float = (
+        MultiModalConfig.mm_encoder_fp8_scale_save_margin
+    )
    io_processor_plugin: str | None = None
    renderer_num_workers: int = 1
    skip_mm_profiling: bool = MultiModalConfig.skip_mm_profiling
@@ -1179,6 +1187,22 @@ class EngineArgs:
            "--mm-encoder-attn-backend",
            **multimodal_kwargs["mm_encoder_attn_backend"],
        )
+        multimodal_group.add_argument(
+            "--mm-encoder-attn-dtype",
+            **multimodal_kwargs["mm_encoder_attn_dtype"],
+        )
+        multimodal_group.add_argument(
+            "--mm-encoder-fp8-scale-path",
+            **multimodal_kwargs["mm_encoder_fp8_scale_path"],
+        )
+        multimodal_group.add_argument(
+            "--mm-encoder-fp8-scale-save-path",
+            **multimodal_kwargs["mm_encoder_fp8_scale_save_path"],
+        )
+        multimodal_group.add_argument(
+            "--mm-encoder-fp8-scale-save-margin",
+            **multimodal_kwargs["mm_encoder_fp8_scale_save_margin"],
+        )
        multimodal_group.add_argument(
            "--interleave-mm-strings", **multimodal_kwargs["interleave_mm_strings"]
        )
@@ -1517,6 +1541,10 @@ class EngineArgs:
            mm_encoder_only=self.mm_encoder_only,
            mm_encoder_tp_mode=self.mm_encoder_tp_mode,
            mm_encoder_attn_backend=self.mm_encoder_attn_backend,
+            mm_encoder_attn_dtype=self.mm_encoder_attn_dtype,
+            mm_encoder_fp8_scale_path=self.mm_encoder_fp8_scale_path,
+            mm_encoder_fp8_scale_save_path=self.mm_encoder_fp8_scale_save_path,
+            mm_encoder_fp8_scale_save_margin=self.mm_encoder_fp8_scale_save_margin,
            pooler_config=self.pooler_config,
            generation_config=self.generation_config,
            override_generation_config=self.override_generation_config,
@@ -78,7 +78,6 @@ class EngineClient(ABC):
        priority: int = 0,
        data_parallel_rank: int | None = None,
        reasoning_ended: bool | None = None,
-        reasoning_parser_kwargs: dict[str, Any] | None = None,
    ) -> AsyncGenerator[RequestOutput, None]:
        """Generate outputs for a request."""
        ...
@@ -79,7 +79,7 @@ from vllm.renderers.inputs.preprocess import (
    prompt_to_seq,
 )
 from vllm.sampling_params import BeamSearchParams, RequestOutputKind, SamplingParams
-from vllm.tasks import PoolingTask
+from vllm.tasks import SCORE_TYPE_MAP, PoolingTask
 from vllm.tokenizers import TokenizerLike
 from vllm.usage.usage_lib import UsageContext
 from vllm.utils.counter import Counter
@@ -1204,12 +1204,9 @@ class LLM:
                    f"Supported tasks: {self.supported_tasks}"
                )
            else:
-                logger.warning_once(
-                    "Pooling multitask support is deprecated and will "
-                    "be removed in v0.20. When the default pooling task is "
-                    "not what you want, you need to manually specify it "
-                    'via PoolerConfig(task="%s"). ',
-                    pooling_task,
+                raise ValueError(
+                    f"Try switching the model's pooling_task "
+                    f'via `PoolerConfig(task="{pooling_task}")`'
                )

        if pooling_task == "plugin" and "plugin" not in self.pooling_io_processors:
@@ -1412,7 +1409,7 @@ class LLM:
                "pooling model."
            )

-        score_type = self.model_config.score_type
+        score_type: str | None = SCORE_TYPE_MAP.get(self.pooling_task, None)  # type: ignore[arg-type]
        if (
            score_type == "cross-encoder"
            and getattr(self.model_config.hf_config, "num_labels", 0) != 1
@@ -317,4 +317,5 @@ class OpenAIServingChatBatch(OpenAIServingChat):
            model=model_name,
            choices=choices,
            usage=usage,
+            system_fingerprint=self.system_fingerprint,
        )
@@ -129,6 +129,9 @@ class ChatCompletionStreamResponse(OpenAIBaseModel):
    model: str
    choices: list[ChatCompletionResponseStreamChoice]
    usage: UsageInfo | None = Field(default=None)
+    # Set only on the final chunk of a stream to mirror non-streaming responses
+    # without the per-chunk serialization overhead.
+    system_fingerprint: str | None = None
    # not part of the OpenAI spec but for tracing the tokens
    prompt_token_ids: list[int] | None = None

@@ -347,11 +347,6 @@ class OpenAIServingChat(OpenAIServing):
                    priority=request.priority,
                    data_parallel_rank=data_parallel_rank,
                    reasoning_ended=reasoning_ended,
-                    reasoning_parser_kwargs={
-                        "chat_template_kwargs": chat_template_kwargs,
-                    }
-                    if reasoning_parser
-                    else None,
                )

            generators.append(generator)
@@ -1200,6 +1195,16 @@ class OpenAIServingChat(OpenAIServing):
                        choices=[choice_data],
                        model=model_name,
                    )
+                    # Stamp the fingerprint on terminal chunks only (those with
+                    # finish_reason set). When ``include_usage`` is on, the
+                    # trailing usage chunk below overrides this as the true
+                    # final message.
+                    if (
+                        not include_usage
+                        and self.system_fingerprint is not None
+                        and choice_data.finish_reason is not None
+                    ):
+                        chunk.system_fingerprint = self.system_fingerprint

                    # handle usage stats if requested & if continuous
                    if include_continuous_usage:
@@ -1234,6 +1239,7 @@ class OpenAIServingChat(OpenAIServing):
                    choices=[],
                    model=model_name,
                    usage=final_usage,
+                    system_fingerprint=self.system_fingerprint,
                )
                final_usage_data = final_usage_chunk.model_dump_json(
                    exclude_unset=True, exclude_none=True
@@ -1642,6 +1648,7 @@ class OpenAIServingChat(OpenAIServing):
            model=model_name,
            choices=choices,
            usage=usage,
+            system_fingerprint=self.system_fingerprint,
            prompt_logprobs=clamp_prompt_logprobs(final_res.prompt_logprobs),
            prompt_token_ids=(
                final_res.prompt_token_ids if request.return_token_ids else None
@@ -153,9 +153,21 @@ class BaseFrontendArgs:
    """If set to True, log the stack trace of error responses"""
    tokens_only: bool = False
    """
-    If set to True, only enable the Tokens In<>Out endpoint. 
+    If set to True, only enable the Tokens In<>Out endpoint.
    This is intended for use in a Disaggregated Everything setup.
    """
+    fingerprint_mode: Literal["full", "hash", "custom", "none"] = "full"
+    """Controls the ``system_fingerprint`` field on responses.
+
+    - ``full`` (default): ``vllm-<version>[-<parallelism>]-<hash8>``. Encodes
+      server version, non-trivial parallelism degrees (tp/pp/dp/ep), and an
+      8-char config hash.
+    - ``hash``: ``vllm-<version>-<hash8>``. Parallelism stripped.
+    - ``custom``: emits the literal string from ``--fingerprint-value``.
+    - ``none``: the field is omitted (serialized as ``null``).
+    """
+    fingerprint_value: str | None = None
+    """Literal fingerprint string used when ``--fingerprint-mode=custom``."""

    @classmethod
    def _customize_cli_kwargs(
@@ -512,3 +512,6 @@ class CompletionStreamResponse(OpenAIBaseModel):
    model: str
    choices: list[CompletionResponseStreamChoice]
    usage: UsageInfo | None = Field(default=None)
+    # Set only on the final chunk of a stream to mirror non-streaming responses
+    # without the per-chunk serialization overhead.
+    system_fingerprint: str | None = None
@@ -383,6 +383,7 @@ class OpenAIServingCompletion(OpenAIServing):

                    chunk = CompletionStreamResponse(
                        id=request_id,
+                        object="text_completion",
                        created=created_time,
                        model=model_name,
                        choices=[
@@ -401,6 +402,14 @@ class OpenAIServingCompletion(OpenAIServing):
                            )
                        ],
                    )
+                    # Stamp on terminal chunk only when no trailing usage chunk
+                    # will follow (that one is the true final message).
+                    if (
+                        not include_usage
+                        and self.system_fingerprint is not None
+                        and finish_reason is not None
+                    ):
+                        chunk.system_fingerprint = self.system_fingerprint
                    if include_continuous_usage:
                        prompt_tokens = num_prompt_tokens[prompt_idx]
                        completion_tokens = previous_num_tokens[i]
@@ -410,7 +419,7 @@ class OpenAIServingCompletion(OpenAIServing):
                            total_tokens=prompt_tokens + completion_tokens,
                        )

-                    response_json = chunk.model_dump_json(exclude_unset=False)
+                    response_json = chunk.model_dump_json(exclude_unset=True)
                    yield f"data: {response_json}\n\n"

            total_prompt_tokens = sum(num_prompt_tokens)
@@ -433,6 +442,7 @@ class OpenAIServingCompletion(OpenAIServing):
                    model=model_name,
                    choices=[],
                    usage=final_usage_info,
+                    system_fingerprint=self.system_fingerprint,
                )
                final_usage_data = final_usage_chunk.model_dump_json(
                    exclude_unset=False, exclude_none=True
@@ -562,6 +572,7 @@ class OpenAIServingCompletion(OpenAIServing):
            model=model_name,
            choices=choices,
            usage=usage,
+            system_fingerprint=self.system_fingerprint,
            kv_transfer_params=kv_transfer_params,
        )

@@ -157,6 +157,19 @@ class OpenAIServing:
        self.renderer = engine_client.renderer
        self.input_processor = engine_client.input_processor

+        # Computed once at startup (cached by ``vllm_config`` identity) and
+        # stamped on non-streaming responses. Streaming chunks deliberately
+        # omit it to avoid per-chunk overhead.
+        from vllm.entrypoints.openai.fingerprint import get_system_fingerprint
+
+        try:
+            self.system_fingerprint: str | None = get_system_fingerprint(
+                engine_client.vllm_config
+            )
+        except Exception:
+            # Never fail server startup over the fingerprint.
+            self.system_fingerprint = None
+
    async def beam_search(
        self,
        prompt: EngineInput,
@@ -0,0 +1,84 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""Build the ``system_fingerprint`` string returned by the OpenAI-compatible
+server.
+
+Four modes, configured via ``--fingerprint-mode``:
+
+* ``full`` (default): ``vllm-<version>[-<parallelism>]-<hash8>`` — encodes
+  server version, any non-trivial parallelism degree (tp/pp/dp/ep), and an
+  8-char prefix of ``vllm_config.compute_hash()`` (covers model identity,
+  quant config, speculative, attention backend, etc.).
+* ``hash``: ``vllm-<version>-<hash8>`` — parallelism stripped.
+* ``custom``: user-provided literal via ``--fingerprint-value``.
+* ``none``: the field is omitted (serialized as ``null``).
+
+``get_system_fingerprint`` is only called at serving-class init (a handful
+of times per server); each subclass caches the returned string on
+``self.system_fingerprint``, so per-request cost is one attribute read.
+"""
+
+from __future__ import annotations
+
+from typing import Any, Literal
+
+FingerprintMode = Literal["full", "hash", "custom", "none"]
+
+_DEFAULT_MODE: FingerprintMode = "full"
+_CUSTOM_VALUE: str | None = None
+
+
+def set_default_fingerprint_mode(
+    mode: FingerprintMode,
+    custom_value: str | None = None,
+) -> None:
+    """Configure the fingerprint mode for subsequent ``get_system_fingerprint``
+    calls. Called once at server startup."""
+    global _DEFAULT_MODE, _CUSTOM_VALUE
+    _DEFAULT_MODE = mode
+    _CUSTOM_VALUE = custom_value
+
+
+def get_system_fingerprint(vllm_config: Any) -> str | None:
+    """Return the fingerprint for ``vllm_config`` using the mode configured by
+    ``set_default_fingerprint_mode``."""
+    return build_system_fingerprint(vllm_config, _DEFAULT_MODE, _CUSTOM_VALUE)
+
+
+def build_system_fingerprint(
+    vllm_config: Any,
+    mode: FingerprintMode = "full",
+    custom_value: str | None = None,
+) -> str | None:
+    if mode == "none":
+        return None
+    if mode == "custom":
+        return custom_value
+
+    from vllm import __version__ as vllm_version
+
+    try:
+        hash8 = vllm_config.compute_hash()[:8]
+    except Exception:
+        hash8 = "nohash"
+
+    if mode == "hash":
+        return f"vllm-{vllm_version}-{hash8}"
+
+    # mode == "full"
+    parts: list[str] = [f"vllm-{vllm_version}"]
+    pc = getattr(vllm_config, "parallel_config", None)
+    if pc is not None:
+        tp = getattr(pc, "tensor_parallel_size", 1)
+        if tp > 1:
+            parts.append(f"tp{tp}")
+        pp = getattr(pc, "pipeline_parallel_size", 1)
+        if pp > 1:
+            parts.append(f"pp{pp}")
+        dp = getattr(pc, "data_parallel_size", 1)
+        if dp > 1:
+            parts.append(f"dp{dp}")
+        if getattr(pc, "enable_expert_parallel", False):
+            parts.append("ep")
+    parts.append(hash8)
+    return "-".join(parts)
@@ -61,9 +61,17 @@ async def init_generate_state(
    )
    from vllm.entrypoints.openai.chat_completion.serving import OpenAIServingChat
    from vllm.entrypoints.openai.completion.serving import OpenAIServingCompletion
+    from vllm.entrypoints.openai.fingerprint import set_default_fingerprint_mode
    from vllm.entrypoints.openai.responses.serving import OpenAIServingResponses
    from vllm.entrypoints.serve.disagg.serving import ServingTokens

+    # Applied before any serving class is constructed so that each one picks
+    # up the chosen mode on its first cache miss.
+    set_default_fingerprint_mode(
+        getattr(args, "fingerprint_mode", "full"),
+        getattr(args, "fingerprint_value", None),
+    )
+
    if args.tool_server == "demo":
        tool_server: ToolServer | None = DemoToolServer()
        assert isinstance(tool_server, DemoToolServer)
--- a/Show More
+++ b/Show More
Author	SHA1	Message	Date
NickLucche	e0dd86f6c3	deprecate timeout Signed-off-by: NickLucche <nlucches@redhat.com>	2026-04-28 09:03:27 +02:00
NickLucche	01f19dee7f	deprecate timeout Signed-off-by: NickLucche <nlucches@redhat.com>	2026-04-28 09:01:01 +02:00
Moritz SanftandGitHub	2c06cf3486	[Bugfix] use `served_model_name` for multimodal error message (#41003 ) Signed-off-by: Moritz Sanft <58110325+msanft@users.noreply.github.com>	2026-04-27 08:22:35 -07:00
Harry Mellor GitHubmergify[bot] <37929162+mergify[bot]@users.noreply.github.com>	e6f710a87f	Deprecate support for Transformers v4 (#40389 ) Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com> Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>	2026-04-27 08:19:57 -07:00
Isotr0py GitHub Jee Jee Li zjy0516 zjy0516 yasong Claude Opus 4.6 Copilot	c245d35ff4	[Model] Add MiMo-V2.5 support (#40967 ) Signed-off-by: Jee Jee Li <pandaleefree@gmail.com> Signed-off-by: Isotr0py <mozf@mail2.sysu.edu.cn> Signed-off-by: Isotr0py <Isotr0py@outlook.com> Signed-off-by: zjy0516 <riverclouds.zhu@qq.com> Co-authored-by: Jee Jee Li <pandaleefree@gmail.com> Co-authored-by: zjy0516 <riverclouds.zhu@qq.com> Co-authored-by: zjy0516 <zhujiangyun@inferact.ai> Co-authored-by: yasong <yasong.wang@inferact.ai> Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com> Co-authored-by: Copilot <copilot@github.com>	2026-04-27 13:26:51 +00:00
Xiaoshuang Wang GitHubmergify[bot] <37929162+mergify[bot]@users.noreply.github.com>	f8ac0c7cf0	[Bugfix] Fix k_norm weight sharding in MiniMaxM2Attention when total_num_kv_heads < tp_size (#38191 ) Signed-off-by: wxsIcey <1790571317@qq.com> Signed-off-by: Icey <1790571317@qq.com> Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>	2026-04-27 05:57:13 -07:00
Simon Mo GitHub Claude	ebf862c351	Add system_fingerprint field to OpenAI-compatible API responses (#40537 ) Co-authored-by: Claude <noreply@anthropic.com>	2026-04-27 16:17:52 +08:00
wang.yuqiandGitHub	8d8062d0a7	[Examples] Resettle generate examples. (#36464 ) Signed-off-by: wang.yuqi <yuqi.wang@daocloud.io>	2026-04-27 07:48:37 +00:00
Roy Wang GitHub Roger Wang	985961345a	[Bugfix] Install libcublas-dev in Dockerfile for FlashInfer CuTe DSL JIT (#39855 ) Signed-off-by: esmeetu <jasonailu87@gmail.com> Co-authored-by: Roger Wang <hey@rogerw.io>	2026-04-27 15:47:39 +08:00
Yongye ZhuandGitHub	706a04d34b	[DSV4] Add silu clamp limit to shared expert (#40950 ) Signed-off-by: Yongye Zhu <zyy1102000@gmail.com>	2026-04-27 00:37:43 -07:00
Isotr0pyandGitHub	22631f80a0	[Bugfix] Remove invalid deepstack boundary check for Qwen3-VL (#40932 ) Signed-off-by: Isotr0py <mozf@mail2.sysu.edu.cn>	2026-04-27 07:27:06 +00:00
BhoomitandGitHub	2cc008e7b4	[Attention][TurboQuant] Share dequant buffers, eliminate float16_copy (#40941 ) Signed-off-by: Bhoomit Vasani <bhoomit.2010@gmail.com> Signed-off-by: Vasani Bhoomit <bhoomit.2010@gmail.com>	2026-04-27 13:48:36 +08:00
Zhanda Zhu GitHub Yubo Gao	5d5c776444	[Perf] FP8 FlashInfer Attn for ViT (#38065 ) Signed-off-by: Zhanda Zhu <zhandazhu@gmail.com> Co-authored-by: Yubo Gao <ybgao-nvidia@users.noreply.github.com>	2026-04-27 13:44:15 +08:00
ojhaanshikaandGitHub	592ae6805c	Cutlass W4A16 (Machete) Tests (#35450 ) Signed-off-by: Anshika Ojha <anshikao@nvidia.com>	2026-04-27 05:15:29 +00:00
Dao007forever GitHub Claude Nick Hill	7b1bc0a3eb	[Bugfix] Cap SWA/chunked-local runtime admission to startup pool-sizing bound (#40946 ) Signed-off-by: Dao Le <Dao007forever@gmail.com> Signed-off-by: Nick Hill <nickhill123@gmail.com> Co-authored-by: Claude <noreply@anthropic.com> Co-authored-by: Nick Hill <nickhill123@gmail.com>	2026-04-27 04:33:13 +00:00
Silu PandaandGitHub	c0879d9483	[Tests] Gate Isaac under Transformers v5 (#40907 ) Signed-off-by: Silu Panda <31051721+SiluPanda@users.noreply.github.com>	2026-04-26 19:26:51 -07:00
Giancarlo DelfinandGitHub	f5f9878514	[Model Runner V2] Fix rejection sampling acceptance rate gap vs MRV1 (#40651 ) Signed-off-by: Giancarlo Delfin <gdelfin@inferact.ai>	2026-04-26 19:12:08 -07:00
youkaichao GitHub Claudegemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>	2ce95a761b	Auto-disable expandable_segments around cumem memory pool (#40812 ) Signed-off-by: youkaichao <youkaichao@gmail.com> Co-authored-by: Claude <noreply@anthropic.com> Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>	2026-04-27 09:37:22 +08:00