[Test] Add tests for n parameter in chat completions API (#35283)

Signed-off-by: KrxGu <krishom70@gmail.com>

.buildkite/hardware_tests/amd.yaml

@@ -10,7 +10,7 @@ steps:
       docker build
       --build-arg max_jobs=16
       --build-arg REMOTE_VLLM=1
-      --build-arg ARG_PYTORCH_ROCM_ARCH='gfx90a;gfx942'
+      --build-arg ARG_PYTORCH_ROCM_ARCH='gfx942;gfx950'
       --build-arg VLLM_BRANCH=$BUILDKITE_COMMIT
       --tag "rocm/vllm-ci:${BUILDKITE_COMMIT}"
       -f docker/Dockerfile.rocm

.buildkite/image_build/image_build_cpu_arm64.sh

@@ -14,7 +14,7 @@ BUILDKITE_COMMIT=$3
 aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin "$REGISTRY"
 
 # skip build if image already exists
-if [[ -z $(docker manifest inspect "$REGISTRY"/"$REPO":"$BUILDKITE_COMMIT"-cpu) ]]; then
+if [[ -z $(docker manifest inspect "$REGISTRY"/"$REPO":"$BUILDKITE_COMMIT"-arm64-cpu) ]]; then
   echo "Image not found, proceeding with build..."
 else
   echo "Image found"
@@ -25,9 +25,9 @@ fi
 docker build --file docker/Dockerfile.cpu \
   --build-arg max_jobs=16 \
   --build-arg buildkite_commit="$BUILDKITE_COMMIT" \
-  --tag "$REGISTRY"/"$REPO":"$BUILDKITE_COMMIT"-cpu \
+  --tag "$REGISTRY"/"$REPO":"$BUILDKITE_COMMIT"-arm64-cpu \
   --target vllm-test \
   --progress plain .
 
 # push
-docker push "$REGISTRY"/"$REPO":"$BUILDKITE_COMMIT"-cpu
+docker push "$REGISTRY"/"$REPO":"$BUILDKITE_COMMIT"-arm64-cpu

.buildkite/release-pipeline.yaml

@@ -31,7 +31,7 @@ steps:
         commands:
           # #NOTE: torch_cuda_arch_list is derived from upstream PyTorch build files here:
           # https://github.com/pytorch/pytorch/blob/main/.ci/aarch64_linux/aarch64_ci_build.sh#L7
-          - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=13.0.1 --build-arg BUILDER_CUDA_VERSION=13.0.1 --build-arg torch_cuda_arch_list='8.7 8.9 9.0 10.0+PTX 12.0' --build-arg BUILD_BASE_IMAGE=nvidia/cuda:13.0.1-devel-ubuntu22.04  --tag vllm-ci:build-image --target build --progress plain -f docker/Dockerfile ."
+          - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=13.0.1 --build-arg torch_cuda_arch_list='8.7 8.9 9.0 10.0+PTX 12.0' --build-arg BUILD_BASE_IMAGE=nvidia/cuda:13.0.1-devel-ubuntu22.04  --tag vllm-ci:build-image --target build --progress plain -f docker/Dockerfile ."
           - "mkdir artifacts"
           - "docker run --rm -v $(pwd)/artifacts:/artifacts_host vllm-ci:build-image bash -c 'cp -r dist /artifacts_host && chmod -R a+rw /artifacts_host'"
           - "bash .buildkite/scripts/upload-nightly-wheels.sh manylinux_2_35"
@@ -70,7 +70,7 @@ steps:
         agents:
           queue: cpu_queue_postmerge
         commands:
-          - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=13.0.1 --build-arg BUILDER_CUDA_VERSION=13.0.1 --build-arg BUILD_BASE_IMAGE=nvidia/cuda:13.0.1-devel-ubuntu22.04 --tag vllm-ci:build-image --target build --progress plain -f docker/Dockerfile ."
+          - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=13.0.1 --build-arg BUILD_BASE_IMAGE=nvidia/cuda:13.0.1-devel-ubuntu22.04 --tag vllm-ci:build-image --target build --progress plain -f docker/Dockerfile ."
           - "mkdir artifacts"
           - "docker run --rm -v $(pwd)/artifacts:/artifacts_host vllm-ci:build-image bash -c 'cp -r dist /artifacts_host && chmod -R a+rw /artifacts_host'"
           - "bash .buildkite/scripts/upload-nightly-wheels.sh manylinux_2_35"
@@ -123,7 +123,7 @@ steps:
           queue: cpu_queue_postmerge
         commands:
           - "aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin public.ecr.aws/q9t5s3a7"
-          - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=13.0.1 --build-arg BUILDER_CUDA_VERSION=13.0.1 --build-arg INSTALL_KV_CONNECTORS=true --build-arg BUILD_BASE_IMAGE=nvidia/cuda:13.0.1-devel-ubuntu22.04 --tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)-cu130 --target vllm-openai --progress plain -f docker/Dockerfile ."
+          - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=13.0.1 --build-arg INSTALL_KV_CONNECTORS=true --build-arg BUILD_BASE_IMAGE=nvidia/cuda:13.0.1-devel-ubuntu22.04 --tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)-cu130 --target vllm-openai --progress plain -f docker/Dockerfile ."
           - "docker push public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)-cu130"
           # re-tag to default image tag and push, just in case arm64 build fails
           - "docker tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)-cu130 public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-cu130"
@@ -137,7 +137,7 @@ steps:
         commands:
           - "aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin public.ecr.aws/q9t5s3a7"
           # compute capability 12.0 for RTX-50 series / RTX PRO 6000 Blackwell, 12.1 for DGX Spark
-          - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=13.0.1 --build-arg BUILDER_CUDA_VERSION=13.0.1 --build-arg torch_cuda_arch_list='8.7 8.9 9.0 10.0+PTX 12.0 12.1' --build-arg INSTALL_KV_CONNECTORS=true --build-arg BUILD_BASE_IMAGE=nvidia/cuda:13.0.1-devel-ubuntu22.04 --tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)-cu130 --target vllm-openai --progress plain -f docker/Dockerfile ."
+          - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=13.0.1 --build-arg torch_cuda_arch_list='8.7 8.9 9.0 10.0+PTX 12.0 12.1' --build-arg INSTALL_KV_CONNECTORS=true --build-arg BUILD_BASE_IMAGE=nvidia/cuda:13.0.1-devel-ubuntu22.04 --tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)-cu130 --target vllm-openai --progress plain -f docker/Dockerfile ."
           - "docker push public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)-cu130"
 
       - block: "Build release image for x86_64 CPU"

.buildkite/scripts/hardware_ci/run-amd-test.sh

@@ -1,25 +1,37 @@
 #!/bin/bash
 
-# This script runs test inside the corresponding ROCm docker container.
+# This script runs tests inside the corresponding ROCm docker container.
+# It handles both single-node and multi-node test configurations.
+#
+# Multi-node detection: Instead of matching on fragile group names, we detect
+# multi-node jobs structurally by looking for the bracket command syntax
+# "[node0_cmds] && [node1_cmds]" or via the NUM_NODES environment variable.
 set -o pipefail
 
 # Export Python path
 export PYTHONPATH=".."
 
-# Print ROCm version
-echo "--- Confirming Clean Initial State"
-while true; do
-        sleep 3
-        if grep -q clean /opt/amdgpu/etc/gpu_state; then
-                echo "GPUs state is \"clean\""
-                break
-        fi
-done
+###############################################################################
+# Helper Functions
+###############################################################################
 
-echo "--- ROCm info"
-rocminfo
+wait_for_clean_gpus() {
+  local timeout=${1:-300}
+  local start=$SECONDS
+  echo "--- Waiting for clean GPU state (timeout: ${timeout}s)"
+  while true; do
+    if grep -q clean /opt/amdgpu/etc/gpu_state; then
+      echo "GPUs state is \"clean\""
+      return
+    fi
+    if (( SECONDS - start >= timeout )); then
+      echo "Error: GPUs did not reach clean state within ${timeout}s" >&2
+      exit 1
+    fi
+    sleep 3
+  done
+}
 
-# cleanup older docker images
 cleanup_docker() {
   # Get Docker's root directory
   docker_root=$(docker info -f '{{.DockerRootDir}}')
@@ -28,15 +40,12 @@ cleanup_docker() {
     exit 1
   fi
   echo "Docker root directory: $docker_root"
-  # Check disk usage of the filesystem where Docker's root directory is located
+
   disk_usage=$(df "$docker_root" | tail -1 | awk '{print $5}' | sed 's/%//')
-  # Define the threshold
   threshold=70
   if [ "$disk_usage" -gt "$threshold" ]; then
     echo "Disk usage is above $threshold%. Cleaning up Docker images and volumes..."
-    # Remove dangling images (those that are not tagged and not used by any container)
     docker image prune -f
-    # Remove unused volumes / force the system prune for old images as well.
     docker volume prune -f && docker system prune --force --filter "until=72h" --all
     echo "Docker images and volumes cleanup completed."
   else
@@ -45,193 +54,258 @@ cleanup_docker() {
 }
 
 cleanup_network() {
-  for node in $(seq 0 $((NUM_NODES-1))); do
-    if docker pr -a -q -f name="node${node}" | grep -q .; then
-      docker stop "node${node}"
+  local max_nodes=${NUM_NODES:-2}
+  for node in $(seq 0 $((max_nodes - 1))); do
+    if docker ps -a -q -f name="node${node}" | grep -q .; then
+      docker stop "node${node}" || true
     fi
   done
-  if docker network ls | grep docker-net; then
-    docker network rm docker-net
+  if docker network ls | grep -q docker-net; then
+    docker network rm docker-net || true
   fi
 }
 
-# Call the cleanup docker function
+is_multi_node() {
+  local cmds="$1"
+  # Primary signal: NUM_NODES environment variable set by the pipeline
+  if [[ "${NUM_NODES:-1}" -gt 1 ]]; then
+    return 0
+  fi
+  # Fallback: detect the bracket syntax structurally
+  # Pattern: [...] && [...] (per-node command arrays)
+  if [[ "$cmds" =~ \[.*\].*\&\&.*\[.*\] ]]; then
+    return 0
+  fi
+  return 1
+}
+
+###############################################################################
+# Pytest marker re-quoting
+#
+# When commands are passed through Buildkite -> shell -> $* -> bash -c,
+# quotes around pytest -m marker expressions get stripped:
+#   pytest -v -s -m 'not cpu_test' v1/core
+# becomes:
+#   pytest -v -s -m not cpu_test v1/core
+#
+# pytest then interprets "cpu_test" as a file path, not part of the marker.
+# This function detects unquoted multi-word marker expressions and re-quotes
+# them so they survive the final bash -c expansion.
+###############################################################################
+
+re_quote_pytest_markers() {
+  local cmds="$1"
+  # Pattern: -m not <identifier>  ->  -m 'not <identifier>'
+  # Handles the common cases: 'not cpu_test', 'not slow_test', etc.
+  cmds=$(echo "$cmds" | sed -E "s/-m not ([a-zA-Z_][a-zA-Z0-9_]*)/-m 'not \1'/g")
+  echo "$cmds"
+}
+
+###############################################################################
+# ROCm-specific pytest command rewrites
+#
+# These apply ignore flags and environment overrides for tests that are not
+# yet supported or behave differently on ROCm hardware. Kept as a single
+# function so new exclusions are easy to add in one place.
+###############################################################################
+
+apply_rocm_test_overrides() {
+  local cmds="$1"
+
+  # --- Model registry filter ---
+  if [[ $cmds == *"pytest -v -s models/test_registry.py"* ]]; then
+    cmds=${cmds//"pytest -v -s models/test_registry.py"/"pytest -v -s models/test_registry.py -k 'not BambaForCausalLM and not GritLM and not Mamba2ForCausalLM and not Zamba2ForCausalLM'"}
+  fi
+
+  # --- LoRA: disable custom paged attention ---
+  if [[ $cmds == *"pytest -v -s lora"* ]]; then
+    cmds=${cmds//"pytest -v -s lora"/"VLLM_ROCM_CUSTOM_PAGED_ATTN=0 pytest -v -s lora"}
+  fi
+
+  # --- Kernel ignores ---
+  if [[ $cmds == *" kernels/core"* ]]; then
+    cmds="${cmds} \
+    --ignore=kernels/core/test_fused_quant_layernorm.py \
+    --ignore=kernels/core/test_permute_cols.py"
+  fi
+
+  if [[ $cmds == *" kernels/attention"* ]]; then
+    cmds="${cmds} \
+    --ignore=kernels/attention/test_attention_selector.py \
+    --ignore=kernels/attention/test_encoder_decoder_attn.py \
+    --ignore=kernels/attention/test_flash_attn.py \
+    --ignore=kernels/attention/test_flashinfer.py \
+    --ignore=kernels/attention/test_prefix_prefill.py \
+    --ignore=kernels/attention/test_cascade_flash_attn.py \
+    --ignore=kernels/attention/test_mha_attn.py \
+    --ignore=kernels/attention/test_lightning_attn.py \
+    --ignore=kernels/attention/test_attention.py"
+  fi
+
+  if [[ $cmds == *" kernels/quantization"* ]]; then
+    cmds="${cmds} \
+    --ignore=kernels/quantization/test_int8_quant.py \
+    --ignore=kernels/quantization/test_machete_mm.py \
+    --ignore=kernels/quantization/test_block_fp8.py \
+    --ignore=kernels/quantization/test_block_int8.py \
+    --ignore=kernels/quantization/test_marlin_gemm.py \
+    --ignore=kernels/quantization/test_cutlass_scaled_mm.py \
+    --ignore=kernels/quantization/test_int8_kernel.py"
+  fi
+
+  if [[ $cmds == *" kernels/mamba"* ]]; then
+    cmds="${cmds} \
+    --ignore=kernels/mamba/test_mamba_mixer2.py \
+    --ignore=kernels/mamba/test_causal_conv1d.py \
+    --ignore=kernels/mamba/test_mamba_ssm_ssd.py"
+  fi
+
+  if [[ $cmds == *" kernels/moe"* ]]; then
+    cmds="${cmds} \
+    --ignore=kernels/moe/test_moe.py \
+    --ignore=kernels/moe/test_cutlass_moe.py \
+    --ignore=kernels/moe/test_triton_moe_ptpc_fp8.py"
+  fi
+
+  # --- Entrypoint ignores ---
+  if [[ $cmds == *" entrypoints/openai "* ]]; then
+    cmds=${cmds//" entrypoints/openai "/" entrypoints/openai \
+    --ignore=entrypoints/openai/test_audio.py \
+    --ignore=entrypoints/openai/test_shutdown.py \
+    --ignore=entrypoints/openai/test_completion.py \
+    --ignore=entrypoints/openai/test_models.py \
+    --ignore=entrypoints/openai/test_lora_adapters.py \
+    --ignore=entrypoints/openai/test_return_tokens_as_ids.py \
+    --ignore=entrypoints/openai/test_root_path.py \
+    --ignore=entrypoints/openai/test_tokenization.py \
+    --ignore=entrypoints/openai/test_prompt_validation.py "}
+  fi
+
+  if [[ $cmds == *" entrypoints/llm "* ]]; then
+    cmds=${cmds//" entrypoints/llm "/" entrypoints/llm \
+    --ignore=entrypoints/llm/test_chat.py \
+    --ignore=entrypoints/llm/test_accuracy.py \
+    --ignore=entrypoints/llm/test_init.py \
+    --ignore=entrypoints/llm/test_prompt_validation.py "}
+  fi
+
+  # Clean up escaped newlines from --ignore appends
+  cmds=$(echo "$cmds" | sed 's/ \\ / /g')
+
+  echo "$cmds"
+}
+
+###############################################################################
+# Main
+###############################################################################
+
+# --- GPU initialization ---
+echo "--- Confirming Clean Initial State"
+wait_for_clean_gpus
+
+echo "--- ROCm info"
+rocminfo
+
+# --- Docker housekeeping ---
 cleanup_docker
 
 echo "--- Resetting GPUs"
-
 echo "reset" > /opt/amdgpu/etc/gpu_state
+wait_for_clean_gpus
 
-while true; do
-        sleep 3
-        if grep -q clean /opt/amdgpu/etc/gpu_state; then
-                echo "GPUs state is \"clean\""
-                break
-        fi
-done
-
+# --- Pull test image ---
 echo "--- Pulling container"
 image_name="rocm/vllm-ci:${BUILDKITE_COMMIT}"
 container_name="rocm_${BUILDKITE_COMMIT}_$(tr -dc A-Za-z0-9 < /dev/urandom | head -c 10; echo)"
 docker pull "${image_name}"
 
 remove_docker_container() {
-   docker rm -f "${container_name}" || docker image rm -f "${image_name}" || true
+  docker rm -f "${container_name}" || docker image rm -f "${image_name}" || true
 }
 trap remove_docker_container EXIT
 
+# --- Prepare commands ---
 echo "--- Running container"
 
 HF_CACHE="$(realpath ~)/huggingface"
 mkdir -p "${HF_CACHE}"
 HF_MOUNT="/root/.cache/huggingface"
 
-commands=$@
+commands="$*"
 echo "Raw commands: $commands"
 
-commands=${commands//"pytest -v -s basic_correctness/test_basic_correctness.py"/"pytest -v -s basic_correctness/test_basic_correctness.py"}
-
-if [[ $commands == *"pytest -v -s models/test_registry.py"* ]]; then
-  commands=${commands//"pytest -v -s models/test_registry.py"/"pytest -v -s models/test_registry.py -k 'not BambaForCausalLM and not GritLM and not Mamba2ForCausalLM and not Zamba2ForCausalLM'"}
-fi
-
-commands=${commands//"pytest -v -s compile/test_basic_correctness.py"/"pytest -v -s compile/test_basic_correctness.py"}
-
-if [[ $commands == *"pytest -v -s lora"* ]]; then
-  commands=${commands//"pytest -v -s lora"/"VLLM_ROCM_CUSTOM_PAGED_ATTN=0 pytest -v -s lora"}
-fi
-
-#ignore certain kernels tests
-if [[ $commands == *" kernels/core"* ]]; then
-  commands="${commands} \
-  --ignore=kernels/core/test_fused_quant_layernorm.py \
-  --ignore=kernels/core/test_permute_cols.py"
-fi
-
-if [[ $commands == *" kernels/attention"* ]]; then
-  commands="${commands} \
-  --ignore=kernels/attention/test_attention_selector.py \
-  --ignore=kernels/attention/test_encoder_decoder_attn.py \
-  --ignore=kernels/attention/test_flash_attn.py \
-  --ignore=kernels/attention/test_flashinfer.py \
-  --ignore=kernels/attention/test_prefix_prefill.py \
-  --ignore=kernels/attention/test_cascade_flash_attn.py \
-  --ignore=kernels/attention/test_mha_attn.py \
-  --ignore=kernels/attention/test_lightning_attn.py \
-  --ignore=kernels/attention/test_attention.py"
-fi
-
-if [[ $commands == *" kernels/quantization"* ]]; then
-  commands="${commands} \
-  --ignore=kernels/quantization/test_int8_quant.py \
-  --ignore=kernels/quantization/test_machete_mm.py \
-  --ignore=kernels/quantization/test_block_fp8.py \
-  --ignore=kernels/quantization/test_block_int8.py \
-  --ignore=kernels/quantization/test_marlin_gemm.py \
-  --ignore=kernels/quantization/test_cutlass_scaled_mm.py \
-  --ignore=kernels/quantization/test_int8_kernel.py"
-fi
-
-if [[ $commands == *" kernels/mamba"* ]]; then
-  commands="${commands} \
-  --ignore=kernels/mamba/test_mamba_mixer2.py \
-  --ignore=kernels/mamba/test_causal_conv1d.py \
-  --ignore=kernels/mamba/test_mamba_ssm_ssd.py"
-fi
-
-if [[ $commands == *" kernels/moe"* ]]; then
-  commands="${commands} \
-  --ignore=kernels/moe/test_moe.py \
-  --ignore=kernels/moe/test_cutlass_moe.py \
-  --ignore=kernels/moe/test_triton_moe_ptpc_fp8.py"
-fi
-
-#ignore certain Entrypoints/openai tests
-if [[ $commands == *" entrypoints/openai "* ]]; then
-  commands=${commands//" entrypoints/openai "/" entrypoints/openai \
-  --ignore=entrypoints/openai/test_audio.py \
-  --ignore=entrypoints/openai/test_shutdown.py \
-  --ignore=entrypoints/openai/test_completion.py \
-  --ignore=entrypoints/openai/test_models.py \
-  --ignore=entrypoints/openai/test_lora_adapters.py \
-  --ignore=entrypoints/openai/test_return_tokens_as_ids.py \
-  --ignore=entrypoints/openai/test_root_path.py \
-  --ignore=entrypoints/openai/test_tokenization.py \
-  --ignore=entrypoints/openai/test_prompt_validation.py "}
-fi
-
-#ignore certain Entrypoints/llm tests
-if [[ $commands == *" entrypoints/llm "* ]]; then
-  commands=${commands//" entrypoints/llm "/" entrypoints/llm \
-  --ignore=entrypoints/llm/test_chat.py \
-  --ignore=entrypoints/llm/test_accuracy.py \
-  --ignore=entrypoints/llm/test_init.py \
-  --ignore=entrypoints/llm/test_prompt_validation.py "}
-fi
-
-commands=$(echo "$commands" | sed 's/ \\ / /g')
+# Fix quoting before ROCm overrides (so overrides see correct structure)
+commands=$(re_quote_pytest_markers "$commands")
+commands=$(apply_rocm_test_overrides "$commands")
 echo "Final commands: $commands"
 
-# --ignore=entrypoints/openai/test_encoder_decoder.py \
-# --ignore=entrypoints/openai/test_embedding.py \
-# --ignore=entrypoints/openai/test_oot_registration.py
-# --ignore=entrypoints/openai/test_accuracy.py \
-# --ignore=entrypoints/openai/test_models.py <= Fails on MI250 but passes on MI300 as of 2025-03-13
-
-
 MYPYTHONPATH=".."
 
-# Test that we're launching on the machine that has
-# proper access to GPUs
+# Verify GPU access
 render_gid=$(getent group render | cut -d: -f3)
 if [[ -z "$render_gid" ]]; then
   echo "Error: 'render' group not found. This is required for GPU access." >&2
   exit 1
 fi
 
-if [[ $commands == *"VLLM_TEST_GROUP_NAME=mi325_4-2-node-tests-4-gpus-in-total"* ]]; then
-
+# --- Route: multi-node vs single-node ---
+if is_multi_node "$commands"; then
+  echo "--- Multi-node job detected"
   export DCKR_VER=$(docker --version | sed 's/Docker version \(.*\), build .*/\1/')
 
-  if [[ "$commands" =~ ^(.*)"["(.*)"] && ["(.*)"]"$ ]]; then
-      prefix=$( echo "${BASH_REMATCH[1]}" | sed 's/;//g')
-      echo "PREFIX: ${prefix}"
-      export composite_command="(command rocm-smi || true)"
-      myIFS=$IFS
-      IFS=','
-      read -ra node0 <<< ${BASH_REMATCH[2]}
-      read -ra node1 <<< ${BASH_REMATCH[3]}
-      IFS=$myIFS
-      for i in "${!node0[@]}";do 
-        command_node_0=$(echo ${node0[i]} | sed 's/\"//g')
-        command_node_1=$(echo ${node1[i]} | sed 's/\"//g')
-        
-        export commands="./.buildkite/scripts/run-multi-node-test.sh /vllm-workspace/tests 2 2 ${image_name} '${command_node_0}' '${command_node_1}'"
-        echo "COMMANDS: ${commands}"
-        composite_command=$(echo "${composite_command} && ${commands}")
-      done
-      /bin/bash -c "${composite_command}"
-      cleanup_network
+  # Parse the bracket syntax:  prefix ; [node0_cmds] && [node1_cmds]
+  #   BASH_REMATCH[1] = prefix (everything before first bracket)
+  #   BASH_REMATCH[2] = comma-separated node0 commands
+  #   BASH_REMATCH[3] = comma-separated node1 commands
+  if [[ "$commands" =~ ^(.*)\[(.*)"] && ["(.*)\]$ ]]; then
+    prefix=$(echo "${BASH_REMATCH[1]}" | sed 's/;//g')
+    echo "PREFIX: ${prefix}"
+
+    export composite_command="(command rocm-smi || true)"
+    saved_IFS=$IFS
+    IFS=','
+    read -ra node0 <<< "${BASH_REMATCH[2]}"
+    read -ra node1 <<< "${BASH_REMATCH[3]}"
+    IFS=$saved_IFS
+
+    if [[ ${#node0[@]} -ne ${#node1[@]} ]]; then
+      echo "Warning: node0 has ${#node0[@]} commands, node1 has ${#node1[@]}. They will be paired by index."
+    fi
+
+    for i in "${!node0[@]}"; do
+      command_node_0=$(echo "${node0[i]}" | sed 's/\"//g')
+      command_node_1=$(echo "${node1[i]}" | sed 's/\"//g')
+
+      step_cmd="./.buildkite/scripts/run-multi-node-test.sh /vllm-workspace/tests 2 2 ${image_name} '${command_node_0}' '${command_node_1}'"
+      echo "COMMANDS: ${step_cmd}"
+      composite_command="${composite_command} && ${step_cmd}"
+    done
+
+    /bin/bash -c "${composite_command}"
+    cleanup_network
   else
-      echo "Failed to parse node commands! Exiting."
-      cleanup_network
-      exit 111
+    echo "Multi-node job detected but failed to parse bracket command syntax."
+    echo "Expected format: prefix ; [node0_cmd1, node0_cmd2] && [node1_cmd1, node1_cmd2]"
+    echo "Got: $commands"
+    cleanup_network
+    exit 111
   fi
 else
+  echo "--- Single-node job"
   echo "Render devices: $BUILDKITE_AGENT_META_DATA_RENDER_DEVICES"
   docker run \
-          --device /dev/kfd $BUILDKITE_AGENT_META_DATA_RENDER_DEVICES \
-          --network=host \
-          --shm-size=16gb \
-          --group-add "$render_gid" \
-          --rm \
-          -e HF_TOKEN \
-          -e AWS_ACCESS_KEY_ID \
-          -e AWS_SECRET_ACCESS_KEY \
-          -v "${HF_CACHE}:${HF_MOUNT}" \
-          -e "HF_HOME=${HF_MOUNT}" \
-          -e "PYTHONPATH=${MYPYTHONPATH}" \
-          --name "${container_name}" \
-          "${image_name}" \
-          /bin/bash -c "${commands}"
+    --device /dev/kfd $BUILDKITE_AGENT_META_DATA_RENDER_DEVICES \
+    --network=host \
+    --shm-size=16gb \
+    --group-add "$render_gid" \
+    --rm \
+    -e HF_TOKEN \
+    -e AWS_ACCESS_KEY_ID \
+    -e AWS_SECRET_ACCESS_KEY \
+    -v "${HF_CACHE}:${HF_MOUNT}" \
+    -e "HF_HOME=${HF_MOUNT}" \
+    -e "PYTHONPATH=${MYPYTHONPATH}" \
+    --name "${container_name}" \
+    "${image_name}" \
+    /bin/bash -c "${commands}"
 fi

.buildkite/scripts/run-multi-node-test.sh

@@ -67,7 +67,7 @@ start_nodes() {
         # 3. map the huggingface cache directory to the container
         # 3. assign ip addresses to the containers (head node: 192.168.10.10, worker nodes:
         #    starting from 192.168.10.11)
-        docker run -d "$GPU_DEVICES" --shm-size=10.24gb -e HF_TOKEN \
+        docker run -d $GPU_DEVICES --shm-size=10.24gb -e HF_TOKEN \
             -v ~/.cache/huggingface:/root/.cache/huggingface --name "node$node" \
             --network docker-net --ip 192.168.10.$((10 + $node)) --rm "$DOCKER_IMAGE" \
             /bin/bash -c "tail -f /dev/null"

.buildkite/scripts/run-prime-rl-test.sh

@@ -1,64 +0,0 @@
-#!/bin/bash
-# SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-
-# Setup script for Prime-RL integration tests
-# This script prepares the environment for running Prime-RL tests with nightly vLLM
-
-set -euo pipefail
-
-SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
-REPO_ROOT="$(cd "${SCRIPT_DIR}/../.." && pwd)"
-PRIME_RL_REPO="https://github.com/PrimeIntellect-ai/prime-rl.git"
-PRIME_RL_DIR="${REPO_ROOT}/prime-rl"
-
-if command -v rocm-smi &> /dev/null || command -v rocminfo &> /dev/null; then
-    echo "AMD GPU detected. Prime-RL currently only supports NVIDIA. Skipping..."
-    exit 0
-fi
-
-echo "Setting up Prime-RL integration test environment..."
-
-# Clean up any existing Prime-RL directory
-if [ -d "${PRIME_RL_DIR}" ]; then
-    echo "Removing existing Prime-RL directory..."
-    rm -rf "${PRIME_RL_DIR}"
-fi
-
-# Install UV if not available
-if ! command -v uv &> /dev/null; then
-    echo "Installing UV package manager..."
-    curl -LsSf https://astral.sh/uv/install.sh | sh
-    source "$HOME"/.local/bin/env
-fi
-
-# Clone Prime-RL repository at specific branch for reproducible tests
-PRIME_RL_BRANCH="integ-vllm-main"
-echo "Cloning Prime-RL repository at branch: ${PRIME_RL_BRANCH}..."
-git clone --branch "${PRIME_RL_BRANCH}" --single-branch "${PRIME_RL_REPO}" "${PRIME_RL_DIR}"
-cd "${PRIME_RL_DIR}"
-
-echo "Setting up UV project environment..."
-export UV_PROJECT_ENVIRONMENT=/usr/local
-ln -s /usr/bin/python3 /usr/local/bin/python
-
-# Remove vllm pin from pyproject.toml
-echo "Removing vllm pin from pyproject.toml..."
-sed -i '/vllm==/d' pyproject.toml
-
-# Sync Prime-RL dependencies
-echo "Installing Prime-RL dependencies..."
-uv sync --inexact && uv sync --inexact --all-extras
-
-# Verify installation
-echo "Verifying installations..."
-uv run python -c "import vllm; print(f'vLLM version: {vllm.__version__}')"
-uv run python -c "import prime_rl; print('Prime-RL imported successfully')"
-
-echo "Prime-RL integration test environment setup complete!"
-
-echo "Running Prime-RL integration tests..."
-export WANDB_MODE=offline # this makes this test not require a WANDB_API_KEY
-uv run pytest -vs tests/integration/test_rl.py -m gpu
-
-echo "Prime-RL integration tests completed!"

.buildkite/scripts/scheduled_integration_test/deepseek_v2_lite_prefetch_offload.sh

@@ -0,0 +1,57 @@
+#!/usr/bin/env bash
+set -euxo pipefail
+
+# Nightly e2e test for prefetch offloading with a MoE model.
+# Runs DeepSeek-V2-Lite with prefetch offloading of MoE expert weights
+# and validates GSM8K accuracy matches baseline (no offloading).
+#
+# args: [THRESHOLD] [NUM_QUESTIONS] [START_PORT]
+THRESHOLD=${1:-0.25}
+NUM_Q=${2:-1319}
+PORT=${3:-8030}
+OUT_DIR=${OUT_DIR:-/tmp/vllm-scheduled}
+mkdir -p "${OUT_DIR}"
+
+wait_for_server() {
+  local port=$1
+  timeout 600 bash -c '
+    until curl -sf "http://127.0.0.1:'"$port"'/health" > /dev/null; do
+      sleep 1
+    done'
+}
+
+MODEL="deepseek-ai/DeepSeek-V2-Lite"
+
+cleanup() {
+  if [[ -n "${SERVER_PID:-}" ]] && kill -0 "${SERVER_PID}" 2>/dev/null; then
+    kill "${SERVER_PID}" 2>/dev/null || true
+    for _ in {1..20}; do
+      kill -0 "${SERVER_PID}" 2>/dev/null || break
+      sleep 0.5
+    done
+    kill -9 "${SERVER_PID}" 2>/dev/null || true
+  fi
+}
+trap cleanup EXIT
+
+vllm serve "$MODEL" \
+  --max-model-len 2048 \
+  --offload-group-size 8 \
+  --offload-num-in-group 2 \
+  --offload-prefetch-step 1 \
+  --offload-params w13_weight w2_weight \
+  --port "$PORT" &
+SERVER_PID=$!
+wait_for_server "$PORT"
+
+TAG=$(echo "$MODEL" | tr '/: \\n' '_____')
+OUT="${OUT_DIR}/${TAG}_prefetch_offload.json"
+python3 tests/evals/gsm8k/gsm8k_eval.py --host http://127.0.0.1 --port "$PORT" --num-questions "${NUM_Q}" --save-results "${OUT}"
+python3 - <<PY
+import json; acc=json.load(open('${OUT}'))['accuracy']
+print(f"${MODEL} prefetch_offload: accuracy {acc:.3f}")
+assert acc >= ${THRESHOLD}, f"${MODEL} prefetch_offload accuracy {acc}"
+PY
+
+cleanup
+SERVER_PID=

.buildkite/test-amd.yaml

@@ -55,9 +55,11 @@ steps:
   grade: Blocking
   source_file_dependencies:
   - vllm/
+  - tests/detokenizer
   - tests/multimodal
   - tests/utils_
   commands:
+  - pytest -v -s detokenizer
   - pytest -v -s -m 'not cpu_test' multimodal
   - pytest -v -s utils_
 
@@ -65,7 +67,7 @@ steps:
   timeout_in_minutes: 30
   mirror_hardwares: [amdexperimental, amdproduction, amdtentative]
   agent_pool: mi325_1
-  grade: Blocking
+  # grade: Blocking
   source_file_dependencies:
   - vllm/
   - tests/test_inputs.py
@@ -547,7 +549,7 @@ steps:
   - tests/samplers
   - tests/conftest.py
   commands:
-    - pytest -v -s -m 'not skip_v1' samplers
+    - pytest -v -s samplers
 
 - label: LoRA Test %N # 20min each
   timeout_in_minutes: 30
@@ -1105,18 +1107,6 @@ steps:
   commands:
     - pytest -v -s models/quantization
 
-# This test is used only in PR development phase to test individual models and should never run on main
-- label: Custom Models Test
-  mirror_hardwares: [amdexperimental, amdproduction]
-  agent_pool: mi325_1
-  # grade: Blocking
-  optional: true
-  commands:
-    - echo 'Testing custom models...'
-    # PR authors can temporarily add commands below to test individual models
-    # e.g. pytest -v -s models/encoder_decoder/vision_language/test_mllama.py
-    # *To avoid merge conflicts, remember to REMOVE (not just comment out) them before merging the PR*
-
 - label: Transformers Nightly Models Test
   mirror_hardwares: [amdexperimental]
   agent_pool: mi325_1
@@ -1632,21 +1622,6 @@ steps:
     - uv pip install --system 'gpt-oss[eval]==0.0.5'
     - VLLM_ROCM_USE_AITER_MHA=0 VLLM_ROCM_USE_AITER=1 VLLM_USE_AITER_UNIFIED_ATTENTION=1 pytest -s -v tests/evals/gpt_oss/test_gpqa_correctness.py --model openai/gpt-oss-20b --metric 0.58
 
-##### RL Integration Tests #####
-- label: Prime-RL Integration Test # 15min
-  mirror_hardwares: [amdexperimental]
-  agent_pool: mi325_2
-  # grade: Blocking
-  timeout_in_minutes: 30
-  optional: true
-  num_gpus: 2
-  working_dir: "/vllm-workspace"
-  source_file_dependencies:
-  - vllm/
-  - .buildkite/scripts/run-prime-rl-test.sh
-  commands:
-    - bash .buildkite/scripts/run-prime-rl-test.sh
-
 ##### EPLB Accuracy Tests #####
 - label: DeepSeek V2-Lite Accuracy
   mirror_hardwares: [amdexperimental, amdproduction]
@@ -1707,7 +1682,6 @@ steps:
   # in /vllm/tools/pre_commit/generate_nightly_torch_test.py
   mirror_hardwares: [amdexperimental, amdproduction, amdtentative]
   agent_pool: mi355_1
-  grade: Blocking
   soft_fail: true
   source_file_dependencies:
   - requirements/nightly_torch_test.txt
@@ -1718,7 +1692,6 @@ steps:
   timeout_in_minutes: 15
   mirror_hardwares: [amdexperimental, amdproduction, amdtentative]
   agent_pool: mi355_1
-  grade: Blocking
   source_file_dependencies:
   - vllm/
   - tests/multimodal
@@ -1731,7 +1704,6 @@ steps:
   timeout_in_minutes: 30
   mirror_hardwares: [amdexperimental, amdproduction, amdtentative]
   agent_pool: mi355_1
-  grade: Blocking
   source_file_dependencies:
   - vllm/
   - tests/test_inputs.py
@@ -1761,7 +1733,6 @@ steps:
   timeout_in_minutes: 20
   mirror_hardwares: [amdexperimental]
   agent_pool: mi355_1
-  # grade: Blocking
   source_file_dependencies:
   - tests/standalone_tests/python_only_compile.sh
   - setup.py
@@ -1772,7 +1743,6 @@ steps:
   timeout_in_minutes: 30
   mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi355_1
-  # grade: Blocking
   fast_check: true
   torch_nightly: true
   source_file_dependencies:
@@ -1789,7 +1759,6 @@ steps:
 - label: Entrypoints Unit Tests # 5min
   mirror_hardwares: [amdexperimental, amdproduction, amdtentative]
   agent_pool: mi355_1
-  grade: Blocking
   timeout_in_minutes: 10
   working_dir: "/vllm-workspace/tests"
   fast_check: true
@@ -1804,7 +1773,6 @@ steps:
   timeout_in_minutes: 40
   mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi355_1
-  # grade: Blocking
   working_dir: "/vllm-workspace/tests"
   fast_check: true
   torch_nightly: true
@@ -1822,7 +1790,6 @@ steps:
   timeout_in_minutes: 130
   mirror_hardwares: [amdexperimental]
   agent_pool: mi355_1
-  # grade: Blocking
   working_dir: "/vllm-workspace/tests"
   fast_check: true
   torch_nightly: true
@@ -1839,7 +1806,6 @@ steps:
   timeout_in_minutes: 50
   mirror_hardwares: [amdexperimental]
   agent_pool: mi355_1
-  # grade: Blocking
   working_dir: "/vllm-workspace/tests"
   fast_check: true
   torch_nightly: true
@@ -1858,7 +1824,6 @@ steps:
   timeout_in_minutes: 50
   mirror_hardwares: [amdexperimental]
   agent_pool: mi355_1
-  # grade: Blocking
   working_dir: "/vllm-workspace/tests"
   fast_check: true
   torch_nightly: true
@@ -1873,7 +1838,6 @@ steps:
   timeout_in_minutes: 50
   mirror_hardwares: [amdexperimental]
   agent_pool: mi355_1
-  # grade: Blocking
   working_dir: "/vllm-workspace/tests"
   fast_check: true
   torch_nightly: true
@@ -1888,7 +1852,6 @@ steps:
   timeout_in_minutes: 50
   mirror_hardwares: [amdexperimental]
   agent_pool: mi355_4
-  # grade: Blocking
   working_dir: "/vllm-workspace/tests"
   num_gpus: 4
   source_file_dependencies:
@@ -1950,7 +1913,6 @@ steps:
   timeout_in_minutes: 10
   mirror_hardwares: [amdexperimental]
   agent_pool: mi355_8
-  # grade: Blocking
   gpu: h100
   num_gpus: 8
   working_dir: "/vllm-workspace/tests"
@@ -1971,7 +1933,6 @@ steps:
 - label: EPLB Algorithm Test # 5min
   mirror_hardwares: [amdexperimental, amdproduction, amdtentative]
   agent_pool: mi355_1
-  grade: Blocking
   timeout_in_minutes: 15
   working_dir: "/vllm-workspace/tests"
   source_file_dependencies:
@@ -1983,7 +1944,6 @@ steps:
 - label: EPLB Execution Test # 10min
   mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi355_4
-  # grade: Blocking
   timeout_in_minutes: 20
   working_dir: "/vllm-workspace/tests"
   num_gpus: 4
@@ -1998,7 +1958,6 @@ steps:
   timeout_in_minutes: 20
   mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi355_2
-  # grade: Blocking
   num_gpus: 2
   source_file_dependencies:
   - vllm/
@@ -2018,7 +1977,6 @@ steps:
   timeout_in_minutes: 20
   mirror_hardwares: [amdexperimental, amdproduction, amdtentative]
   agent_pool: mi355_1
-  grade: Blocking
   source_file_dependencies:
   - vllm/
   - tests/test_regression
@@ -2031,7 +1989,6 @@ steps:
   timeout_in_minutes: 15
   mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi355_1
-  # grade: Blocking
   source_file_dependencies:
   - vllm/
   - tests/engine
@@ -2048,7 +2005,6 @@ steps:
   # The test uses 4 GPUs, but we schedule it on 8-GPU machines for stability.
   # See discussion here: https://github.com/vllm-project/vllm/pull/31040
   agent_pool: mi355_8
-  # grade: Blocking
   source_file_dependencies:
     - vllm/
     - tests/v1
@@ -2062,7 +2018,6 @@ steps:
   timeout_in_minutes: 50
   mirror_hardwares: [amdexperimental, amdproduction, amdtentative]
   agent_pool: mi355_1
-  grade: Blocking
   source_file_dependencies:
     - vllm/
     - tests/v1
@@ -2073,7 +2028,6 @@ steps:
   timeout_in_minutes: 60
   mirror_hardwares: [amdexperimental]
   agent_pool: mi355_1
-  # grade: Blocking
   source_file_dependencies:
     - vllm/
     - tests/v1
@@ -2101,7 +2055,6 @@ steps:
 - label: V1 Test attention (H100) # 10min
   mirror_hardwares: [amdexperimental]
   agent_pool: mi355_1
-  # grade: Blocking
   timeout_in_minutes: 30
   gpu: h100
   source_file_dependencies:
@@ -2141,7 +2094,6 @@ steps:
 - label: V1 Test others (CPU) # 5 mins
   mirror_hardwares: [amdexperimental, amdproduction, amdtentative]
   agent_pool: mi355_1
-  grade: Blocking
   source_file_dependencies:
     - vllm/
     - tests/v1
@@ -2159,7 +2111,6 @@ steps:
   timeout_in_minutes: 45
   mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi355_1
-  # grade: Blocking
   working_dir: "/vllm-workspace/examples"
   source_file_dependencies:
   - vllm/entrypoints
@@ -2194,7 +2145,6 @@ steps:
   timeout_in_minutes: 15
   mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi355_1
-  # grade: Blocking
   source_file_dependencies:
   - vllm/
   - tests/cuda
@@ -2206,20 +2156,18 @@ steps:
   timeout_in_minutes: 75
   mirror_hardwares: [amdexperimental]
   agent_pool: mi355_1
-  # grade: Blocking
   source_file_dependencies:
   - vllm/model_executor/layers
   - vllm/sampling_metadata.py
   - tests/samplers
   - tests/conftest.py
   commands:
-    - pytest -v -s -m 'not skip_v1' samplers
+    - pytest -v -s samplers
 
 - label: LoRA Test %N # 20min each
   timeout_in_minutes: 30
   mirror_hardwares: [amdexperimental]
   agent_pool: mi355_1
-  # grade: Blocking
   source_file_dependencies:
   - vllm/lora
   - tests/lora
@@ -2240,7 +2188,6 @@ steps:
   timeout_in_minutes: 30
   mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi355_1
-  # grade: Blocking
   torch_nightly: true
   source_file_dependencies:
     - vllm/
@@ -2257,7 +2204,6 @@ steps:
   timeout_in_minutes: 30
   mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi355_1
-  # grade: Blocking
   torch_nightly: true
   source_file_dependencies:
   - vllm/
@@ -2303,7 +2249,6 @@ steps:
   timeout_in_minutes: 75
   mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi355_1
-  # grade: Blocking
   source_file_dependencies:
   - csrc/
   - tests/kernels/core
@@ -2315,7 +2260,6 @@ steps:
   timeout_in_minutes: 35
   mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi355_1
-  # grade: Blocking
   source_file_dependencies:
   - csrc/attention/
   - vllm/v1/attention
@@ -2330,7 +2274,6 @@ steps:
   timeout_in_minutes: 90
   mirror_hardwares: [amdexperimental]
   agent_pool: mi355_1
-  # grade: Blocking
   source_file_dependencies:
   - csrc/quantization/
   - vllm/model_executor/layers/quantization
@@ -2343,7 +2286,6 @@ steps:
   timeout_in_minutes: 60
   mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi355_1
-  # grade: Blocking
   source_file_dependencies:
   - csrc/quantization/cutlass_w8a8/moe/
   - csrc/moe/
@@ -2360,7 +2302,6 @@ steps:
   timeout_in_minutes: 45
   mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi355_1
-  # grade: Blocking
   source_file_dependencies:
   - csrc/mamba/
   - tests/kernels/mamba
@@ -2404,7 +2345,6 @@ steps:
   torch_nightly: true
   mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi355_1
-  # grade: Blocking
   source_file_dependencies:
   - vllm/engine/arg_utils.py
   - vllm/config/model.py
@@ -2421,7 +2361,6 @@ steps:
   timeout_in_minutes: 20
   mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi355_1
-  # grade: Blocking
   working_dir: "/vllm-workspace/.buildkite"
   source_file_dependencies:
   - benchmarks/
@@ -2432,7 +2371,6 @@ steps:
   timeout_in_minutes: 20
   mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi355_1
-  # grade: Blocking
   source_file_dependencies:
   - vllm/
   - tests/benchmarks/
@@ -2443,7 +2381,6 @@ steps:
   timeout_in_minutes: 90
   mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi355_1
-  # grade: Blocking
   source_file_dependencies:
   - csrc/
   - vllm/model_executor/layers/quantization
@@ -2464,7 +2401,6 @@ steps:
   timeout_in_minutes: 75
   mirror_hardwares: [amdexperimental]
   agent_pool: mi355_1
-  # grade: Blocking
   source_file_dependencies:
   - csrc/
   - vllm/model_executor/layers/quantization
@@ -2476,7 +2412,6 @@ steps:
   timeout_in_minutes: 15
   mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi355_1
-  # grade: Blocking
   source_file_dependencies:
   - csrc/
   - vllm/entrypoints/openai/
@@ -2493,7 +2428,6 @@ steps:
   timeout_in_minutes: 45
   mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi355_1
-  # grade: Blocking
   torch_nightly: true
   source_file_dependencies:
   - vllm/
@@ -2506,7 +2440,6 @@ steps:
   timeout_in_minutes: 45
   mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi355_1
-  # grade: Blocking
   torch_nightly: true
   source_file_dependencies:
   - vllm/model_executor/models/
@@ -2526,7 +2459,6 @@ steps:
   timeout_in_minutes: 45
   mirror_hardwares: [amdexperimental]
   agent_pool: mi355_1
-  # grade: Blocking
   torch_nightly: true
   source_file_dependencies:
   - vllm/
@@ -2539,7 +2471,6 @@ steps:
 - label: Basic Models Test (Other CPU) # 5min
   mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi355_1
-  # grade: Blocking
   timeout_in_minutes: 10
   torch_nightly: true
   source_file_dependencies:
@@ -2554,7 +2485,6 @@ steps:
   timeout_in_minutes: 25
   mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi355_1
-  # grade: Blocking
   torch_nightly: true
   source_file_dependencies:
   - vllm/
@@ -2568,7 +2498,6 @@ steps:
   timeout_in_minutes: 45
   mirror_hardwares: [amdexperimental]
   agent_pool: mi355_1
-  # grade: Blocking
   torch_nightly: true
   source_file_dependencies:
   - vllm/model_executor/models/
@@ -2589,7 +2518,6 @@ steps:
   timeout_in_minutes: 75
   mirror_hardwares: [amdexperimental]
   agent_pool: mi355_1
-  # grade: Blocking
   torch_nightly: true
   source_file_dependencies:
   - vllm/
@@ -2610,7 +2538,6 @@ steps:
   timeout_in_minutes: 110
   mirror_hardwares: [amdexperimental]
   agent_pool: mi355_1
-  # grade: Blocking
   optional: true
   source_file_dependencies:
   - vllm/
@@ -2626,7 +2553,6 @@ steps:
   timeout_in_minutes: 110
   mirror_hardwares: [amdexperimental]
   agent_pool: mi355_1
-  # grade: Blocking
   optional: true
   source_file_dependencies:
   - vllm/
@@ -2638,7 +2564,6 @@ steps:
   timeout_in_minutes: 50
   mirror_hardwares: [amdexperimental]
   agent_pool: mi355_1
-  # grade: Blocking
   optional: true
   source_file_dependencies:
   - vllm/
@@ -2674,7 +2599,6 @@ steps:
   timeout_in_minutes: 60
   mirror_hardwares: [amdexperimental]
   agent_pool: mi355_1
-  # grade: Blocking
   source_file_dependencies:
   - vllm/
   - tests/models/multimodal
@@ -2686,7 +2610,6 @@ steps:
   timeout_in_minutes: 100
   mirror_hardwares: [amdexperimental]
   agent_pool: mi355_1
-  # grade: Blocking
   torch_nightly: true
   source_file_dependencies:
   - vllm/
@@ -2704,7 +2627,6 @@ steps:
   timeout_in_minutes: 10
   mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi355_1
-  # grade: Blocking
   working_dir: "/vllm-workspace/.buildkite/lm-eval-harness"
   source_file_dependencies:
   - vllm/multimodal/
@@ -2719,7 +2641,6 @@ steps:
   timeout_in_minutes: 120
   mirror_hardwares: [amdexperimental]
   agent_pool: mi355_1
-  # grade: Blocking
   optional: true
   source_file_dependencies:
   - vllm/
@@ -2734,7 +2655,6 @@ steps:
   timeout_in_minutes: 120
   mirror_hardwares: [amdexperimental]
   agent_pool: mi355_1
-  # grade: Blocking
   optional: true
   source_file_dependencies:
   - vllm/
@@ -2749,7 +2669,6 @@ steps:
   timeout_in_minutes: 150
   mirror_hardwares: [amdexperimental]
   agent_pool: mi355_1
-  # grade: Blocking
   optional: true
   source_file_dependencies:
   - vllm/
@@ -2764,29 +2683,15 @@ steps:
   timeout_in_minutes: 60
   mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi355_1
-  # grade: Blocking
   source_file_dependencies:
   - vllm/model_executor/layers/quantization
   - tests/models/quantization
   commands:
     - pytest -v -s models/quantization
 
-# This test is used only in PR development phase to test individual models and should never run on main
-- label: Custom Models Test
-  mirror_hardwares: [amdexperimental, amdproduction]
-  agent_pool: mi355_1
-  # grade: Blocking
-  optional: true
-  commands:
-    - echo 'Testing custom models...'
-    # PR authors can temporarily add commands below to test individual models
-    # e.g. pytest -v -s models/encoder_decoder/vision_language/test_mllama.py
-    # *To avoid merge conflicts, remember to REMOVE (not just comment out) them before merging the PR*
-
 - label: Transformers Nightly Models Test
   mirror_hardwares: [amdexperimental]
   agent_pool: mi355_1
-  # grade: Blocking
   working_dir: "/vllm-workspace/"
   optional: true
   commands:
@@ -2925,7 +2830,6 @@ steps:
   timeout_in_minutes: 20
   mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi355_2
-  # grade: Blocking
   working_dir: "/vllm-workspace/tests"
   num_gpus: 2
   source_file_dependencies:
@@ -2941,7 +2845,6 @@ steps:
   timeout_in_minutes: 30
   mirror_hardwares: [amdexperimental, amdmultinode]
   agent_pool: mi355_4
-  # grade: Blocking
   working_dir: "/vllm-workspace/tests"
   num_gpus: 2
   num_nodes: 2
@@ -2968,7 +2871,6 @@ steps:
   timeout_in_minutes: 90
   mirror_hardwares: [amdexperimental]
   agent_pool: mi355_2
-  # grade: Blocking
   working_dir: "/vllm-workspace/tests"
   num_gpus: 2
   source_file_dependencies:
@@ -3008,7 +2910,6 @@ steps:
   timeout_in_minutes: 50
   mirror_hardwares: [amdexperimental]
   agent_pool: mi355_2
-  # grade: Blocking
   working_dir: "/vllm-workspace/tests"
   num_gpus: 2
   source_file_dependencies:
@@ -3030,7 +2931,6 @@ steps:
   timeout_in_minutes: 60
   mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi355_2
-  # grade: Blocking
   working_dir: "/vllm-workspace/tests"
   num_gpus: 2
   source_file_dependencies:
@@ -3064,7 +2964,6 @@ steps:
   timeout_in_minutes: 60
   mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi355_4
-  # grade: Blocking
   working_dir: "/vllm-workspace/tests"
   num_gpus: 4
   source_file_dependencies:
@@ -3081,7 +2980,6 @@ steps:
   timeout_in_minutes: 30
   mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi355_4
-  # grade: Blocking
   num_gpus: 4
   source_file_dependencies:
   - vllm/lora
@@ -3106,7 +3004,6 @@ steps:
   timeout_in_minutes: 45
   mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi355_2
-  # grade: Blocking
   working_dir: "/vllm-workspace/tests"
   num_gpus: 2
   optional: true
@@ -3119,7 +3016,6 @@ steps:
 - label: Weight Loading Multiple GPU Test - Large Models # optional
   mirror_hardwares: [amdexperimental]
   agent_pool: mi355_2
-  # grade: Blocking
   working_dir: "/vllm-workspace/tests"
   num_gpus: 2
   optional: true
@@ -3132,7 +3028,6 @@ steps:
 - label: NixlConnector PD accuracy tests (Distributed) # 30min
   mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi355_4
-  # grade: Blocking
   timeout_in_minutes: 30
   working_dir: "/vllm-workspace/tests"
   num_gpus: 4
@@ -3146,7 +3041,6 @@ steps:
 - label: DP EP NixlConnector PD accuracy tests (Distributed) # 15min
   mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi355_4
-  # grade: Blocking
   timeout_in_minutes: 15
   working_dir: "/vllm-workspace/tests"
   num_gpus: 4
@@ -3163,7 +3057,6 @@ steps:
 - label: Distributed Tests (A100) # optional
   mirror_hardwares: [amdexperimental]
   agent_pool: mi355_4
-  # grade: Blocking
   gpu: a100
   optional: true
   num_gpus: 4
@@ -3186,7 +3079,6 @@ steps:
   optional: true
   mirror_hardwares: [amdexperimental]
   agent_pool: mi355_4
-  # grade: Blocking
   num_gpus: 4
   working_dir: "/vllm-workspace/.buildkite/lm-eval-harness"
   source_file_dependencies:
@@ -3202,7 +3094,6 @@ steps:
   optional: true
   mirror_hardwares: [amdexperimental]
   agent_pool: mi355_4
-  # grade: Blocking
   num_gpus: 4
   working_dir: "/vllm-workspace/.buildkite/lm-eval-harness"
   source_file_dependencies:
@@ -3217,7 +3108,6 @@ steps:
 - label: Distributed Tests (H200) # optional
   mirror_hardwares: [amdexperimental]
   agent_pool: mi355_2
-  # grade: Blocking
   gpu: h200
   optional: true
   working_dir: "/vllm-workspace/"
@@ -3252,7 +3142,6 @@ steps:
   timeout_in_minutes: 20
   mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi355_1
-  # grade: Blocking
   source_file_dependencies:
   - csrc/
   - vllm/model_executor/layers/quantization
@@ -3262,7 +3151,6 @@ steps:
 - label: LM Eval Large Models (4 Card)
   mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi355_4
-  # grade: Blocking
   gpu: a100
   optional: true
   num_gpus: 4
@@ -3298,26 +3186,10 @@ steps:
     - uv pip install --system 'gpt-oss[eval]==0.0.5'
     - VLLM_ROCM_USE_AITER_MHA=0 VLLM_ROCM_USE_AITER=1 VLLM_USE_AITER_UNIFIED_ATTENTION=1 pytest -s -v tests/evals/gpt_oss/test_gpqa_correctness.py --model openai/gpt-oss-20b --metric 0.58
 
-##### RL Integration Tests #####
-- label: Prime-RL Integration Test # 15min
-  mirror_hardwares: [amdexperimental]
-  agent_pool: mi355_2
-  # grade: Blocking
-  timeout_in_minutes: 30
-  optional: true
-  num_gpus: 2
-  working_dir: "/vllm-workspace"
-  source_file_dependencies:
-  - vllm/
-  - .buildkite/scripts/run-prime-rl-test.sh
-  commands:
-    - bash .buildkite/scripts/run-prime-rl-test.sh
-
 ##### EPLB Accuracy Tests #####
 - label: DeepSeek V2-Lite Accuracy
   mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi355_4
-  # grade: Blocking
   timeout_in_minutes: 60
   gpu: h100
   optional: true
@@ -3329,7 +3201,6 @@ steps:
 - label: Qwen3-30B-A3B-FP8-block Accuracy (H100)
   mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi355_4
-  # grade: Blocking
   timeout_in_minutes: 60
   gpu: h100
   optional: true
@@ -3352,7 +3223,6 @@ steps:
   timeout_in_minutes: 60
   mirror_hardwares: [amdexperimental]
   agent_pool: mi355_4
-  # grade: Blocking
   optional: true
   num_gpus: 4
   working_dir: "/vllm-workspace"

.buildkite/test_areas/distributed.yaml

@@ -104,7 +104,6 @@ steps:
   # NEW rlhf examples
   - cd new_weight_syncing
   - VLLM_ALLOW_INSECURE_SERIALIZATION=1 python3 rlhf.py
-  - VLLM_ALLOW_INSECURE_SERIALIZATION=1 python3 rlhf_async_new_apis.py
 
 - label: Distributed Tests (8 GPUs)(H100)
   timeout_in_minutes: 10
@@ -146,6 +145,7 @@ steps:
   num_devices: 2
   commands:
     - pytest -v -s tests/distributed/test_context_parallel.py
+    - VLLM_ALLOW_INSECURE_SERIALIZATION=1 python3 examples/offline_inference/new_weight_syncing/rlhf_async_new_apis.py
     - VLLM_USE_DEEP_GEMM=1 VLLM_LOGGING_LEVEL=DEBUG python3 examples/offline_inference/data_parallel.py --model=Qwen/Qwen1.5-MoE-A2.7B -tp=1 -dp=2 --max-model-len=2048 --all2all-backend=deepep_high_throughput
     - pytest -v -s tests/v1/distributed/test_dbo.py
 
@@ -165,6 +165,7 @@ steps:
   num_devices: 2
   num_nodes: 2
   no_plugin: true
+  optional: true # TODO: revert once infra issue solved
   source_file_dependencies:
   - vllm/distributed/
   - vllm/engine/
@@ -208,7 +209,7 @@ steps:
     - uv pip install --system -r /vllm-workspace/requirements/kv_connectors.txt
     - CROSS_LAYERS_BLOCKS=True bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
 
-- label: Pipeline + Context Parallelism (4 GPUs))
+- label: Pipeline + Context Parallelism (4 GPUs)
   timeout_in_minutes: 60
   working_dir: "/vllm-workspace/tests"
   num_devices: 4

.buildkite/test_areas/e2e_integration.yaml

@@ -29,15 +29,11 @@ steps:
   commands:
   - bash .buildkite/scripts/scheduled_integration_test/qwen30b_a3b_fp8_block_ep_eplb.sh 0.8 200 8020 2 1
 
-- label: Prime-RL Integration (2 GPUs)
-  timeout_in_minutes: 30
+- label: DeepSeek V2-Lite Prefetch Offload Accuracy (H100)
+  timeout_in_minutes: 60
+  device: h100
   optional: true
-  soft_fail: true
-  num_devices: 2
+  num_devices: 1
   working_dir: "/vllm-workspace"
-  source_file_dependencies:
-  - vllm/
-  - .buildkite/scripts/run-prime-rl-test.sh
   commands:
-    - nvidia-smi
-    - bash .buildkite/scripts/run-prime-rl-test.sh
+  - bash .buildkite/scripts/scheduled_integration_test/deepseek_v2_lite_prefetch_offload.sh 0.25 200 8030

.buildkite/test_areas/engine.yaml

@@ -28,3 +28,11 @@ steps:
     - pytest -v -s v1/engine/test_preprocess_error_handling.py
     # Run the rest of v1/engine tests
     - pytest -v -s v1/engine --ignore v1/engine/test_preprocess_error_handling.py
+  mirror:
+    amd:
+      device: mi325_1
+      depends_on:
+      - image-build-amd
+      commands:
+      - pytest -v -s v1/e2e
+      - pytest -v -s v1/engine

.buildkite/test_areas/entrypoints.yaml

@@ -65,6 +65,11 @@ steps:
   commands:
   - export VLLM_WORKER_MULTIPROC_METHOD=spawn
   - pytest -v -s entrypoints/pooling
+  mirror:
+    amd:
+      device: mi325_1
+      depends_on:
+      - image-build-amd
 
 - label: Entrypoints Integration (Responses API)
   timeout_in_minutes: 50

.buildkite/test_areas/kernels.yaml

@@ -115,6 +115,7 @@ steps:
     - pytest -v -s tests/kernels/moe/test_nvfp4_moe.py
     - pytest -v -s tests/kernels/moe/test_ocp_mx_moe.py
     - pytest -v -s tests/kernels/moe/test_flashinfer.py
+    - pytest -v -s tests/kernels/moe/test_flashinfer_moe.py
     - pytest -v -s tests/kernels/moe/test_cutedsl_moe.py
     # e2e
     - pytest -v -s tests/models/quantization/test_nvfp4.py
@@ -156,14 +157,3 @@ steps:
     - pytest -v -s kernels/moe/test_deepep_moe.py
     - pytest -v -s kernels/moe/test_pplx_cutlass_moe.py
     # - pytest -v -s kernels/moe/test_pplx_moe.py - failing on main
-  
-- label: Kernels Fp4 MoE Test (B200)
-  timeout_in_minutes: 60
-  device: b200
-  num_devices: 1
-  optional: true
-  commands:
-    - pytest -v -s kernels/moe/test_cutedsl_moe.py
-    - pytest -v -s kernels/moe/test_flashinfer_moe.py
-    - pytest -v -s kernels/moe/test_nvfp4_moe.py
-    - pytest -v -s kernels/moe/test_ocp_mx_moe.py

.buildkite/test_areas/lm_eval.yaml

@@ -73,3 +73,29 @@ steps:
   num_devices: 2
   commands:
     - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=evals/gsm8k/configs/moe-refactor-dp-ep/config-b200.txt
+
+- label: GPQA Eval (GPT-OSS) (H100)
+  timeout_in_minutes: 120
+  device: h100
+  optional: true
+  num_devices: 2
+  source_file_dependencies:
+  - csrc/
+  - vllm/model_executor/layers/quantization
+  - tests/evals/gpt_oss/
+  commands:
+    - uv pip install --system 'gpt-oss[eval]==0.0.5'
+    - pytest -s -v evals/gpt_oss/test_gpqa_correctness.py --config-list-file=configs/models-h100.txt
+
+- label: GPQA Eval (GPT-OSS) (B200)
+  timeout_in_minutes: 120
+  device: b200
+  optional: true
+  num_devices: 2
+  source_file_dependencies:
+  - csrc/
+  - vllm/model_executor/layers/quantization
+  - tests/evals/gpt_oss/
+  commands:
+    - uv pip install --system 'gpt-oss[eval]==0.0.5'
+    - pytest -s -v evals/gpt_oss/test_gpqa_correctness.py --config-list-file=configs/models-b200.txt

.buildkite/test_areas/misc.yaml

@@ -16,6 +16,7 @@ steps:
     - pytest -v -s v1/sample
     - pytest -v -s v1/logits_processors
     - pytest -v -s v1/worker
+    # TODO: create another `optional` test group for slow tests
     - pytest -v -s -m 'not slow_test' v1/spec_decode
     - pytest -v -s -m 'not cpu_test' v1/kv_connector/unit
     - pytest -v -s -m 'not cpu_test' v1/metrics
@@ -25,6 +26,11 @@ steps:
     # Integration test for streaming correctness (requires special branch).
     - pip install -U git+https://github.com/robertgshaw2-redhat/lm-evaluation-harness.git@streaming-api
     - pytest -v -s entrypoints/openai/correctness/test_lmeval.py::test_lm_eval_accuracy_v1_engine
+  mirror:
+    amd:
+      device: mi325_1
+      depends_on:
+      - image-build-amd
 
 - label: V1 Others (CPU)
   depends_on:
@@ -108,9 +114,11 @@ steps:
   timeout_in_minutes: 50
   source_file_dependencies:
   - vllm/
+  - tests/detokenizer
   - tests/multimodal
   - tests/utils_
   commands:
+  - pytest -v -s detokenizer
   - pytest -v -s -m 'not cpu_test' multimodal
   - pytest -v -s utils_
 
@@ -145,20 +153,6 @@ steps:
   - pytest -v -s transformers_utils
   - pytest -v -s config
 
-- label: GPT-OSS Eval (B200)
-  timeout_in_minutes: 60
-  working_dir: "/vllm-workspace/"
-  device: b200
-  optional: true
-  source_file_dependencies:
-  - tests/evals/gpt_oss
-  - vllm/model_executor/models/gpt_oss.py
-  - vllm/model_executor/layers/quantization/mxfp4.py
-  - vllm/v1/attention/backends/flashinfer.py
-  commands:
-    - uv pip install --system 'gpt-oss[eval]==0.0.5'
-    - pytest -s -v tests/evals/gpt_oss/test_gpqa_correctness.py --model openai/gpt-oss-20b --metric 0.58
-
 - label: Batch Invariance (H100)
   timeout_in_minutes: 25
   device: h100

.buildkite/test_areas/models_language.yaml

@@ -55,6 +55,15 @@ steps:
     - uv pip install --system --no-build-isolation 'git+https://github.com/state-spaces/mamba@v2.3.0'
     - uv pip install --system --no-build-isolation 'git+https://github.com/Dao-AILab/causal-conv1d@v1.5.2'
     - pytest -v -s models/language/generation -m '(not core_model) and (not hybrid_model)'
+  mirror:
+    amd:
+      device: mi325_1
+      depends_on:
+      - image-build-amd
+      commands:
+      - uv pip install --system --no-build-isolation 'git+https://github.com/AndreasKaratzas/mamba@fix-rocm-7.0-warp-size-constexpr'
+      - uv pip install --system --no-build-isolation 'git+https://github.com/Dao-AILab/causal-conv1d@v1.5.2'
+      - pytest -v -s models/language/generation -m '(not core_model) and (not hybrid_model)'
 
 - label: Language Models Test (PPL)
   timeout_in_minutes: 110
@@ -73,6 +82,11 @@ steps:
   - tests/models/language/pooling
   commands:
     - pytest -v -s models/language/pooling -m 'not core_model'
+  mirror:
+    amd:
+      device: mi325_1
+      depends_on:
+      - image-build-amd
 
 - label: Language Models Test (MTEB)
   timeout_in_minutes: 110

.buildkite/test_areas/samplers.yaml

@@ -18,4 +18,4 @@ steps:
       depends_on:
       - image-build-amd
       commands:
-      - pytest -v -s -m 'not skip_v1' samplers
+      - pytest -v -s samplers

.github/CODEOWNERS

@@ -5,14 +5,14 @@
 /vllm/compilation @zou3519 @youkaichao @ProExpertProg
 /vllm/distributed/kv_transfer @NickLucche @ApostaC @orozery
 /vllm/lora @jeejeelee
-/vllm/model_executor/layers/attention @LucasWilkinson
+/vllm/model_executor/layers/attention @LucasWilkinson @MatthewBonanni
 /vllm/model_executor/layers/fused_moe @mgoin @pavanimajety
 /vllm/model_executor/layers/quantization @mgoin @robertgshaw2-redhat @tlrmchlsmth @yewentao256 @pavanimajety
 /vllm/model_executor/layers/mamba @tdoublep
 /vllm/model_executor/model_loader @22quinn
 /vllm/model_executor/layers/batch_invariant.py @yewentao256 
 /vllm/multimodal @DarkLight1337 @ywang96 @NickLucche @tjtanaa
-/vllm/vllm_flash_attn @LucasWilkinson
+/vllm/vllm_flash_attn @LucasWilkinson @MatthewBonanni
 CMakeLists.txt @tlrmchlsmth @LucasWilkinson
 
 # Any change to the VllmConfig changes can have a large user-facing impact,
@@ -43,19 +43,19 @@ CMakeLists.txt @tlrmchlsmth @LucasWilkinson
 /vllm/tool_parsers @aarnphm @chaunceyjiang
 
 # vLLM V1
-/vllm/v1/attention @LucasWilkinson
+/vllm/v1/attention @LucasWilkinson @MatthewBonanni
 /vllm/v1/attention/backend.py @WoosukKwon @zhuohan123 @youkaichao @alexm-redhat @njhill
 /vllm/v1/attention/backends/mla @pavanimajety
 /vllm/v1/attention/backends/flashinfer.py @mgoin @pavanimajety
 /vllm/v1/attention/backends/triton_attn.py @tdoublep
 /vllm/v1/core @WoosukKwon @robertgshaw2-redhat @njhill @ywang96 @alexm-redhat @heheda12345 @ApostaC @orozery
 /vllm/v1/sample @22quinn @houseroad @njhill
-/vllm/v1/spec_decode @benchislett @luccafong
+/vllm/v1/spec_decode @benchislett @luccafong @MatthewBonanni
 /vllm/v1/structured_output @mgoin @russellb @aarnphm @benchislett
 /vllm/v1/kv_cache_interface.py @heheda12345
 /vllm/v1/kv_offload @ApostaC @orozery
 /vllm/v1/worker/gpu/kv_connector.py @orozery
-/vllm/v1/worker/kv_connector_model_runner_mixin.py @orozery
+/vllm/v1/worker/kv_connector_model_runner_mixin.py @orozery @NickLucche
 
 # Model runner V2
 /vllm/v1/worker/gpu @WoosukKwon

.pre-commit-config.yaml

@@ -38,7 +38,7 @@ repos:
   rev: 0.9.1
   hooks:
     - id: pip-compile
-      args: [requirements/test.in, -o, requirements/test.txt, --index-strategy, unsafe-best-match, --torch-backend, cu128, --python-platform, x86_64-manylinux_2_28, --python-version, "3.12"]
+      args: [requirements/test.in, -o, requirements/test.txt, --index-strategy, unsafe-best-match, --torch-backend, cu129, --python-platform, x86_64-manylinux_2_28, --python-version, "3.12"]
       files: ^requirements/test\.(in|txt)$
 - repo: local
   hooks:

CMakeLists.txt

@@ -771,6 +771,24 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
     endif()
   endif()
 
+  # DeepSeek V3 fused A GEMM kernel (requires SM 9.0+, Hopper and later)
+  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
+    cuda_archs_loose_intersection(DSV3_FUSED_A_GEMM_ARCHS "9.0a;10.0f;11.0f" "${CUDA_ARCHS}")
+  else()
+    cuda_archs_loose_intersection(DSV3_FUSED_A_GEMM_ARCHS "9.0a;10.0a;10.1a;10.3a" "${CUDA_ARCHS}")
+  endif()
+  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.0 AND DSV3_FUSED_A_GEMM_ARCHS)
+    set(DSV3_FUSED_A_GEMM_SRC "csrc/dsv3_fused_a_gemm.cu")
+    set_gencode_flags_for_srcs(
+      SRCS "${DSV3_FUSED_A_GEMM_SRC}"
+      CUDA_ARCHS "${DSV3_FUSED_A_GEMM_ARCHS}")
+    list(APPEND VLLM_EXT_SRC ${DSV3_FUSED_A_GEMM_SRC})
+    message(STATUS "Building dsv3_fused_a_gemm for archs: ${DSV3_FUSED_A_GEMM_ARCHS}")
+  else()
+    message(STATUS "Not building dsv3_fused_a_gemm as no compatible archs found "
+                   "in CUDA target architectures.")
+  endif()
+
   # moe_data.cu is used by all CUTLASS MoE kernels.
   if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
     cuda_archs_loose_intersection(CUTLASS_MOE_DATA_ARCHS "9.0a;10.0f;11.0f;12.0f" "${CUDA_ARCHS}")
@@ -1082,6 +1100,27 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
     message(STATUS "Not building Marlin MOE kernels as no compatible archs found"
                    " in CUDA target architectures")
   endif()
+
+  # DeepSeek V3 router GEMM kernel - requires SM90+
+  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
+    cuda_archs_loose_intersection(DSV3_ROUTER_GEMM_ARCHS "9.0a;10.0f;11.0f" "${CUDA_ARCHS}")
+  else()
+    cuda_archs_loose_intersection(DSV3_ROUTER_GEMM_ARCHS "9.0a;10.0a;10.1a;10.3a" "${CUDA_ARCHS}")
+  endif()
+  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.0 AND DSV3_ROUTER_GEMM_ARCHS)
+    set(DSV3_ROUTER_GEMM_SRC
+      "csrc/moe/dsv3_router_gemm_entry.cu"
+      "csrc/moe/dsv3_router_gemm_float_out.cu"
+      "csrc/moe/dsv3_router_gemm_bf16_out.cu")
+    set_gencode_flags_for_srcs(
+      SRCS "${DSV3_ROUTER_GEMM_SRC}"
+      CUDA_ARCHS "${DSV3_ROUTER_GEMM_ARCHS}")
+    list(APPEND VLLM_MOE_EXT_SRC "${DSV3_ROUTER_GEMM_SRC}")
+    message(STATUS "Building DSV3 router GEMM kernel for archs: ${DSV3_ROUTER_GEMM_ARCHS}")
+  else()
+    message(STATUS "Not building DSV3 router GEMM kernel as no compatible archs found"
+                   " (requires SM90+ and CUDA >= 12.0)")
+  endif()
 endif()
 
 message(STATUS "Enabling moe extension.")

benchmarks/fused_kernels/layernorm_rms_benchmarks.py

@@ -13,6 +13,7 @@ from torch.utils.benchmark import Measurement as TMeasurement
 from tqdm import tqdm
 
 import vllm._custom_ops as ops
+from vllm.benchmarks.lib.utils import default_vllm_config
 from vllm.model_executor.layers.layernorm import RMSNorm
 from vllm.model_executor.layers.quantization.utils.fp8_utils import (
     per_token_group_quant_fp8,
@@ -291,6 +292,7 @@ def print_timers(timers: Iterable[TMeasurement]):
     compare.print()
 
 
+@default_vllm_config()
 def main():
     torch.set_default_device("cuda")
     bench_params = get_bench_params()

benchmarks/kernels/benchmark_activation.py

@@ -7,6 +7,7 @@ import itertools
 import torch
 
 import vllm.model_executor.layers.activation  # noqa F401
+from vllm.benchmarks.lib.utils import default_vllm_config
 from vllm.model_executor.custom_op import op_registry
 from vllm.triton_utils import triton
 from vllm.utils.argparse_utils import FlexibleArgumentParser
@@ -18,6 +19,7 @@ intermediate_size = [3072, 9728, 12288]
 configs = list(itertools.product(batch_size_range, seq_len_range, intermediate_size))
 
 
+@default_vllm_config()
 def benchmark_activation(
     batch_size: int,
     seq_len: int,

benchmarks/kernels/benchmark_block_fp8_gemm.py

@@ -8,6 +8,7 @@ os.environ["VLLM_USE_DEEP_GEMM"] = "0"
 
 import torch
 
+from vllm.benchmarks.lib.utils import default_vllm_config
 from vllm.model_executor.layers.quantization.utils.fp8_utils import (
     W8A8BlockFp8LinearOp,
 )
@@ -40,6 +41,7 @@ DEEPSEEK_V3_SHAPES = [
 ]
 
 
+@default_vllm_config()
 def build_w8a8_block_fp8_runner(M, N, K, block_size, device, use_cutlass):
     """Build runner function for w8a8 block fp8 matmul."""
     factor_for_scale = 1e-2

benchmarks/kernels/benchmark_device_communicators.py

@@ -30,6 +30,9 @@ import torch.distributed as dist
 from torch.distributed import ProcessGroup
 
 from vllm.distributed.device_communicators.custom_all_reduce import CustomAllreduce
+from vllm.distributed.device_communicators.flashinfer_all_reduce import (
+    FlashInferAllReduce,
+)
 from vllm.distributed.device_communicators.pynccl import (
     PyNcclCommunicator,
     register_nccl_symmetric_ops,
@@ -44,7 +47,7 @@ from vllm.utils.argparse_utils import FlexibleArgumentParser
 logger = init_logger(__name__)
 
 # Default sequence lengths to benchmark
-DEFAULT_SEQUENCE_LENGTHS = [128, 512, 1024, 2048, 4096, 8192]
+DEFAULT_SEQUENCE_LENGTHS = [16, 64, 128, 512, 1024, 2048, 4096, 8192]
 
 # Fixed hidden size and dtype for all benchmarks
 HIDDEN_SIZE = 8192
@@ -81,6 +84,7 @@ class CommunicatorBenchmark:
         self.symm_mem_comm = None
         self.symm_mem_comm_multimem = None
         self.symm_mem_comm_two_shot = None
+        self.fi_ar_comm = None
 
         self._init_communicators()
 
@@ -161,6 +165,22 @@ class CommunicatorBenchmark:
             )
             self.symm_mem_comm_two_shot = None
 
+        try:
+            self.fi_ar_comm = FlashInferAllReduce(
+                group=self.cpu_group,
+                device=self.device,
+            )
+            if not self.fi_ar_comm.disabled:
+                logger.info("Rank %s: FlashInferAllReduce initialized", self.rank)
+            else:
+                logger.info("Rank %s: FlashInferAllReduce disabled", self.rank)
+                self.fi_ar_comm = None
+        except Exception as e:
+            logger.warning(
+                "Rank %s: Failed to initialize FlashInferAllReduce: %s", self.rank, e
+            )
+            self.fi_ar_comm = None
+
     def benchmark_allreduce(
         self, sequence_length: int, num_warmup: int, num_trials: int
     ) -> dict[str, float]:
@@ -180,7 +200,8 @@ class CommunicatorBenchmark:
                     lambda t, c=comm: c.custom_all_reduce(t),
                     lambda t, c=comm: c.should_custom_ar(t),
                     comm.capture(),
-                    "1stage",  # env variable value
+                    {"VLLM_CUSTOM_ALLREDUCE_ALGO": "1stage"},
+                    None,  # no destroy function
                 )
             )
             # CustomAllreduce two-shot
@@ -190,7 +211,8 @@ class CommunicatorBenchmark:
                     lambda t, c=comm: c.custom_all_reduce(t),
                     lambda t, c=comm: c.should_custom_ar(t),
                     comm.capture(),
-                    "2stage",  # env variable value
+                    {"VLLM_CUSTOM_ALLREDUCE_ALGO": "2stage"},
+                    None,  # no destroy function
                 )
             )
 
@@ -202,7 +224,8 @@ class CommunicatorBenchmark:
                     lambda t, c=comm: c.all_reduce(t),
                     lambda t: True,  # Always available if initialized
                     nullcontext(),
-                    None,  # no env variable needed
+                    {},  # no env variable needed
+                    None,  # no destroy function
                 )
             )
             communicators.append(
@@ -211,7 +234,8 @@ class CommunicatorBenchmark:
                     lambda t: torch.ops.vllm.all_reduce_symmetric_with_copy(t),
                     lambda t: True,  # Always available if initialized
                     nullcontext(),
-                    None,  # no env variable needed
+                    {},  # no env variable needed
+                    None,  # no destroy function
                 )
             )
 
@@ -223,7 +247,8 @@ class CommunicatorBenchmark:
                     lambda t, c=comm: c.all_reduce(t),
                     lambda t, c=comm: c.should_use_symm_mem(t),
                     nullcontext(),
-                    None,  # no env variable needed
+                    {},  # no env variable needed
+                    None,  # no destroy function
                 )
             )
 
@@ -235,29 +260,67 @@ class CommunicatorBenchmark:
                     lambda t, c=comm: c.all_reduce(t),
                     lambda t, c=comm: c.should_use_symm_mem(t),
                     nullcontext(),
-                    None,  # no env variable needed
+                    {},  # no env variable needed
+                    None,  # no destroy function needed
+                )
+            )
+
+        if self.fi_ar_comm is not None:
+            comm = self.fi_ar_comm
+            communicators.append(
+                (
+                    "flashinfer_trtllm",
+                    lambda t, c=comm: c.all_reduce(t),
+                    lambda t, c=comm: c.should_use_fi_ar(t),
+                    nullcontext(),
+                    {"VLLM_FLASHINFER_ALLREDUCE_BACKEND": "trtllm"},
+                    lambda c=comm: c.destroy(),
+                )
+            )
+            communicators.append(
+                (
+                    "flashinfer_mnnvl",
+                    lambda t, c=comm: c.all_reduce(t),
+                    lambda t, c=comm: c.should_use_fi_ar(t),
+                    nullcontext(),
+                    {"VLLM_FLASHINFER_ALLREDUCE_BACKEND": "mnnvl"},
+                    lambda c=comm: c.destroy(),
                 )
             )
 
         # Benchmark each communicator
-        for name, allreduce_fn, should_use_fn, context, env_var in communicators:
-            # Set environment variable if needed
-            if env_var is not None:
-                os.environ["VLLM_CUSTOM_ALLREDUCE_ALGO"] = env_var
-            else:
-                # Clear the environment variable to avoid interference
-                os.environ.pop("VLLM_CUSTOM_ALLREDUCE_ALGO", None)
-
-            latency = self.benchmark_allreduce_single(
-                sequence_length,
-                allreduce_fn,
-                should_use_fn,
-                context,
-                num_warmup,
-                num_trials,
-            )
-            if latency is not None:
-                results[name] = latency
+        for (
+            name,
+            allreduce_fn,
+            should_use_fn,
+            context,
+            env_dict,
+            destroy_fn,
+        ) in communicators:
+            # Save original values and apply new environment variables
+            saved_env = {key: os.environ.get(key) for key in env_dict}
+            for key, value in env_dict.items():
+                os.environ[key] = value
+            try:
+                latency = self.benchmark_allreduce_single(
+                    sequence_length,
+                    allreduce_fn,
+                    should_use_fn,
+                    context,
+                    num_warmup,
+                    num_trials,
+                )
+                if latency is not None:
+                    results[name] = latency
+            finally:
+                if destroy_fn is not None:
+                    destroy_fn()
+                # Restore environment variables to their original state
+                for key, original_value in saved_env.items():
+                    if original_value is None:
+                        os.environ.pop(key, None)
+                    else:
+                        os.environ[key] = original_value
 
         return results
 

benchmarks/kernels/benchmark_fused_collective.py

@@ -5,8 +5,11 @@
 Benchmark for FlashInfer fused collective operations vs standard operations.
 
 This benchmark compares:
-1. FlashInfer's allreduce_fusion (fused allreduce + rmsnorm + optional quant)
-2. Standard tensor_model_parallel_all_reduce + separate rmsnorm/quant operations
+1. FlashInfer's allreduce_fusion with trtllm backend
+   (fused allreduce + rmsnorm + optional FP8/FP4 quant)
+2. FlashInfer's allreduce_fusion with mnnvl backend
+   (fused allreduce + rmsnorm only, no quantization support)
+3. Standard tensor_model_parallel_all_reduce + separate rmsnorm/quant operations
 
 Usage with torchrun:
     torchrun --nproc_per_node=2 benchmark_fused_collective.py
@@ -48,8 +51,12 @@ SCALED_FP4_QUANT_OP = torch.ops._C.scaled_fp4_quant
 logger = init_logger(__name__)
 
 # Try to import FlashInfer
+TorchDistBackend = None
 try:
     import flashinfer.comm as flashinfer_comm  # type: ignore
+    from flashinfer.comm.mnnvl import (  # type: ignore
+        TorchDistBackend,
+    )
 
     if not (
         hasattr(flashinfer_comm, "allreduce_fusion")
@@ -74,11 +81,15 @@ _FI_MAX_SIZES = {
     8: 64 * MiB,  # 64MB
 }
 
-# Global workspace tensor for FlashInfer
-_FI_WORKSPACE = None
+# Global workspace tensors for FlashInfer (keyed by backend name)
+_FI_WORKSPACES: dict = {}
+
+# Backends to benchmark
+FLASHINFER_BACKENDS = ["trtllm", "mnnvl"]
 
 
 def setup_flashinfer_workspace(
+    backend: str,
     world_size: int,
     rank: int,
     hidden_dim: int,
@@ -86,41 +97,54 @@ def setup_flashinfer_workspace(
     dtype: torch.dtype,
 ):
     """Setup FlashInfer workspace for fused allreduce operations."""
-    global _FI_WORKSPACE
+    global FI_WORKSPACES
 
     if flashinfer_comm is None:
-        return None, None
+        return None
 
     if world_size not in _FI_MAX_SIZES:
         logger.warning("FlashInfer not supported for world size %s", world_size)
-        return None, None
+        return None
 
     try:
+        kwargs = {}
+        if TorchDistBackend is not None:
+            kwargs["comm_backend"] = TorchDistBackend(group=dist.group.WORLD)
+
         workspace = flashinfer_comm.create_allreduce_fusion_workspace(
-            backend="trtllm",
+            backend=backend,
             world_size=world_size,
             rank=rank,
             max_token_num=max_token_num,
             hidden_dim=hidden_dim,
             dtype=dtype,
+            **kwargs,
         )
 
-        _FI_WORKSPACE = workspace
+        _FI_WORKSPACES[backend] = workspace
         return workspace
     except Exception as e:
-        logger.error("Failed to setup FlashInfer workspace: %s", e)
+        logger.error(
+            "Failed to setup FlashInfer workspace (backend=%s): %s", backend, e
+        )
         return None
 
 
-def cleanup_flashinfer_workspace(workspace):
-    """Cleanup FlashInfer workspace."""
-    if flashinfer_comm is None or workspace is None:
+def cleanup_flashinfer_workspaces():
+    """Cleanup all FlashInfer workspaces."""
+    if flashinfer_comm is None:
         return
 
-    try:
-        workspace.destroy()
-    except Exception as e:
-        logger.error("Failed to cleanup FlashInfer workspace: %s", e)
+    for backend, workspace in _FI_WORKSPACES.items():
+        try:
+            workspace.destroy()
+        except Exception as e:
+            logger.error(
+                "Failed to cleanup FlashInfer workspace (backend=%s): %s",
+                backend,
+                e,
+            )
+    _FI_WORKSPACES.clear()
 
 
 class FlashInferFusedAllReduceParams:
@@ -134,7 +158,7 @@ class FlashInferFusedAllReduceParams:
         self.fp32_acc = True
         self.max_token_num = max_token_num
 
-    def get_trtllm_fused_allreduce_kwargs(self):
+    def get_flashinfer_fused_allreduce_kwargs(self):
         return {
             "launch_with_pdl": self.launch_with_pdl,
             "fp32_acc": self.fp32_acc,
@@ -147,11 +171,12 @@ def flashinfer_fused_allreduce_rmsnorm(
     rms_gamma: torch.Tensor,
     rms_eps: float,
     allreduce_params: "FlashInferFusedAllReduceParams",
+    workspace: object,
     use_oneshot: bool,
     norm_out: torch.Tensor | None = None,
 ):
     """FlashInfer fused allreduce + rmsnorm operation."""
-    if flashinfer_comm is None or _FI_WORKSPACE is None:
+    if flashinfer_comm is None or workspace is None:
         raise RuntimeError("FlashInfer not available or workspace not initialized")
 
     if norm_out is None:
@@ -160,9 +185,13 @@ def flashinfer_fused_allreduce_rmsnorm(
     else:
         residual_out = input_tensor
 
+    layout_code = None
+    if workspace.backend == "trtllm":
+        layout_code = flashinfer_comm.QuantizationSFLayout.SWIZZLED_128x4
+
     flashinfer_comm.allreduce_fusion(
         input=input_tensor,
-        workspace=_FI_WORKSPACE,
+        workspace=workspace,
         pattern=flashinfer_comm.AllReduceFusionPattern.kARResidualRMSNorm,
         residual_in=residual,
         residual_out=residual_out,
@@ -171,10 +200,10 @@ def flashinfer_fused_allreduce_rmsnorm(
         rms_eps=rms_eps,
         quant_out=None,
         scale_out=None,
-        layout_code=flashinfer_comm.QuantizationSFLayout.SWIZZLED_128x4,
+        layout_code=layout_code,
         scale_factor=None,
         use_oneshot=use_oneshot,
-        **allreduce_params.get_trtllm_fused_allreduce_kwargs(),
+        **allreduce_params.get_flashinfer_fused_allreduce_kwargs(),
     )
 
 
@@ -185,12 +214,16 @@ def flashinfer_fused_allreduce_rmsnorm_fp8_quant(
     rms_eps: float,
     scale_factor: torch.Tensor,
     allreduce_params: FlashInferFusedAllReduceParams,
+    workspace: object,
     use_oneshot: bool = True,
     norm_out: torch.Tensor | None = None,
     quant_out: torch.Tensor | None = None,
 ):
-    """FlashInfer fused allreduce + rmsnorm + FP8 quantization."""
-    if flashinfer_comm is None or _FI_WORKSPACE is None:
+    """FlashInfer fused allreduce + rmsnorm + FP8 quantization.
+
+    Note: Only supported by the trtllm backend.
+    """
+    if flashinfer_comm is None or workspace is None:
         raise RuntimeError("FlashInfer not available or workspace not initialized")
 
     if norm_out is None:
@@ -201,7 +234,7 @@ def flashinfer_fused_allreduce_rmsnorm_fp8_quant(
 
     flashinfer_comm.allreduce_fusion(
         input=input_tensor,
-        workspace=_FI_WORKSPACE,
+        workspace=workspace,
         pattern=flashinfer_comm.AllReduceFusionPattern.kARResidualRMSNormFP8Quant,
         residual_in=residual,
         residual_out=residual_out,
@@ -213,7 +246,7 @@ def flashinfer_fused_allreduce_rmsnorm_fp8_quant(
         layout_code=flashinfer_comm.QuantizationSFLayout.SWIZZLED_128x4,
         scale_factor=scale_factor,
         use_oneshot=use_oneshot,
-        **allreduce_params.get_trtllm_fused_allreduce_kwargs(),
+        **allreduce_params.get_flashinfer_fused_allreduce_kwargs(),
     )
 
 
@@ -224,13 +257,17 @@ def flashinfer_fused_allreduce_rmsnorm_fp4_quant(
     rms_eps: float,
     input_global_scale: torch.Tensor,
     allreduce_params: FlashInferFusedAllReduceParams,
+    workspace: object,
     quant_out: torch.Tensor,
     use_oneshot: bool,
     output_scale: torch.Tensor,
     norm_out: torch.Tensor | None = None,
 ):
-    """FlashInfer fused allreduce + rmsnorm + FP4 quantization."""
-    if flashinfer_comm is None or _FI_WORKSPACE is None:
+    """FlashInfer fused allreduce + rmsnorm + FP4 quantization.
+
+    Note: Only supported by the trtllm backend.
+    """
+    if flashinfer_comm is None or workspace is None:
         raise RuntimeError("FlashInfer not available or workspace not initialized")
 
     if norm_out is None:
@@ -241,7 +278,7 @@ def flashinfer_fused_allreduce_rmsnorm_fp4_quant(
 
     flashinfer_comm.allreduce_fusion(
         input=input_tensor,
-        workspace=_FI_WORKSPACE,
+        workspace=workspace,
         pattern=flashinfer_comm.AllReduceFusionPattern.kARResidualRMSNormFP4Quant,
         residual_in=residual,
         residual_out=residual_out,
@@ -253,7 +290,7 @@ def flashinfer_fused_allreduce_rmsnorm_fp4_quant(
         layout_code=flashinfer_comm.QuantizationSFLayout.SWIZZLED_128x4,
         scale_factor=input_global_scale,
         use_oneshot=use_oneshot,
-        **allreduce_params.get_trtllm_fused_allreduce_kwargs(),
+        **allreduce_params.get_flashinfer_fused_allreduce_kwargs(),
     )
 
 
@@ -386,13 +423,16 @@ def run_benchmarks(
     dtype: torch.dtype,
     use_residual: bool,
     allreduce_params: FlashInferFusedAllReduceParams | None,
+    workspaces: dict,
     quant_modes: set[str],
     no_oneshot: bool,
 ):
     """Run all benchmarks for given configuration.
 
     Args:
-        quant_mode: "none", "fp8_only", "fp4_only", or "all"
+        allreduce_params: Shared parameters for FlashInfer fused allreduce.
+        workspaces: Dict mapping backend name ("trtllm", "mnnvl") to workspace.
+        quant_modes: Set of quantization modes: "none", "fp8", "fp4".
     """
     (
         input_tensor,
@@ -408,18 +448,18 @@ def run_benchmarks(
 
     rms_eps = 1e-6
     results = {}
-    vllm_fused_allreduce = VllmFusedAllreduce(hidden_dim, dtype)
     use_oneshot_options = [False] if no_oneshot else [True, False]
 
-    # Create RMSNorm and QuantFP8 layers once for native benchmarks
-
     if "none" in quant_modes:
         # Standard AllReduce + RMSNorm
+        # Re-create VllmFusedAllreduce per config so CustomOp binds the
+        # correct forward method (native vs custom kernel).
         for custom_op in ["-rms_norm", "+rms_norm"]:
             with set_current_vllm_config(
                 VllmConfig(compilation_config=CompilationConfig(custom_ops=[custom_op]))
             ):
                 try:
+                    vllm_fused_allreduce = VllmFusedAllreduce(hidden_dim, dtype)
                     suffix = (
                         "_custom_rms_norm" if "+" in custom_op else "_native_rms_norm"
                     )
@@ -438,6 +478,7 @@ def run_benchmarks(
             VllmConfig(compilation_config=CompilationConfig(custom_ops=["-rms_norm"]))
         ):
             try:
+                vllm_fused_allreduce = VllmFusedAllreduce(hidden_dim, dtype)
                 standard_allreduce_rmsnorm_native_compiled = torch.compile(
                     vllm_fused_allreduce.allreduce_rmsnorm,
                     fullgraph=True,
@@ -453,10 +494,11 @@ def run_benchmarks(
                 logger.error("Standard AllReduce+RMSNorm Native Compiled failed: %s", e)
                 results["standard_allreduce_rmsnorm_native_compiled"] = float("inf")
 
-        # FlashInfer Fused AllReduce + RMSNorm Oneshot/Twoshot
-        if flashinfer_comm is not None and allreduce_params is not None:
+        # FlashInfer Fused AllReduce + RMSNorm (all backends)
+        for backend, workspace in workspaces.items():
             for use_oneshot in use_oneshot_options:
                 suffix = "_oneshot" if use_oneshot else "_twoshot"
+                key = f"flashinfer_{backend}_fused_allreduce_rmsnorm{suffix}"
                 try:
                     time_ms = benchmark_operation(
                         flashinfer_fused_allreduce_rmsnorm,
@@ -466,14 +508,17 @@ def run_benchmarks(
                         rms_gamma=rms_gamma,
                         rms_eps=rms_eps,
                         allreduce_params=allreduce_params,
+                        workspace=workspace,
                         use_oneshot=use_oneshot,
                     )
-                    results[f"flashinfer_fused_allreduce_rmsnorm{suffix}"] = time_ms
+                    results[key] = time_ms
                 except Exception as e:
-                    logger.error("FlashInfer Fused AllReduce+RMSNorm failed: %s", e)
-                    results[f"flashinfer_fused_allreduce_rmsnorm{suffix}"] = float(
-                        "inf"
+                    logger.error(
+                        "FlashInfer (%s) Fused AllReduce+RMSNorm failed: %s",
+                        backend,
+                        e,
                     )
+                    results[key] = float("inf")
 
     if "fp8" in quant_modes:
         # Standard AllReduce + RMSNorm + FP8 Quant
@@ -482,7 +527,7 @@ def run_benchmarks(
                 "_custom_rms_norm" if "+" in rms_norm_custom_op else "_native_rms_norm"
             )
             for quant_fp8_custom_op in ["-quant_fp8", "+quant_fp8"]:
-                suffix += (
+                op_suffix = suffix + (
                     "_custom_quant_fp8"
                     if "+" in quant_fp8_custom_op
                     else "_native_quant_fp8"
@@ -495,16 +540,17 @@ def run_benchmarks(
                     )
                 ):
                     try:
+                        vllm_fused_allreduce = VllmFusedAllreduce(hidden_dim, dtype)
                         time_ms = benchmark_operation(
                             vllm_fused_allreduce.allreduce_rmsnorm_fp8_quant,
                             input_tensor,
                             residual=residual,
                             scale_factor=scale_fp8,
                         )
-                        results[f"standard_allreduce{suffix}"] = time_ms
+                        results[f"standard_allreduce{op_suffix}"] = time_ms
                     except Exception as e:
                         logger.error("Standard AllReduce+RMSNorm+FP8 failed: %s", e)
-                        results[f"standard_allreduce{suffix}"] = float("inf")
+                        results[f"standard_allreduce{op_suffix}"] = float("inf")
 
         # Standard AllReduce + RMSNorm + FP8 Quant Native Compiled
         with set_current_vllm_config(
@@ -515,6 +561,7 @@ def run_benchmarks(
             )
         ):
             try:
+                vllm_fused_allreduce = VllmFusedAllreduce(hidden_dim, dtype)
                 standard_allreduce_rmsnorm_fp8_quant_native_compiled = torch.compile(
                     vllm_fused_allreduce.allreduce_rmsnorm_fp8_quant,
                     fullgraph=True,
@@ -537,10 +584,12 @@ def run_benchmarks(
                     "inf"
                 )
 
-        # FlashInfer Fused AllReduce + RMSNorm + FP8 Quant Oneshot
-        if flashinfer_comm is not None and allreduce_params is not None:
+        # FlashInfer Fused AllReduce + RMSNorm + FP8 Quant (trtllm only)
+        if "trtllm" in workspaces:
+            trtllm_ws = workspaces["trtllm"]
             for use_oneshot in use_oneshot_options:
                 suffix = "_oneshot" if use_oneshot else "_twoshot"
+                key = f"flashinfer_trtllm_fused_allreduce_rmsnorm_fp8_quant{suffix}"
                 try:
                     time_ms = benchmark_operation(
                         flashinfer_fused_allreduce_rmsnorm_fp8_quant,
@@ -552,19 +601,16 @@ def run_benchmarks(
                         scale_factor=scale_fp8,
                         quant_out=quant_out_fp8,
                         allreduce_params=allreduce_params,
+                        workspace=trtllm_ws,
                         use_oneshot=use_oneshot,
                     )
-                    results[f"flashinfer_fused_allreduce_rmsnorm_fp8_quant{suffix}"] = (
-                        time_ms
-                    )
+                    results[key] = time_ms
                 except Exception as e:
                     logger.error(
-                        "FlashInfer Fused AllReduce+RMSNorm+FP8 Oneshot failed: %s",
+                        "FlashInfer (trtllm) Fused AllReduce+RMSNorm+FP8 failed: %s",
                         e,
                     )
-                    results[f"flashinfer_fused_allreduce_rmsnorm_fp8_quant{suffix}"] = (
-                        float("inf")
-                    )
+                    results[key] = float("inf")
 
     if "fp4" in quant_modes and current_platform.has_device_capability(100):
         # Standard AllReduce + RMSNorm + FP4 Quant
@@ -580,6 +626,7 @@ def run_benchmarks(
                 )
             ):
                 try:
+                    vllm_fused_allreduce = VllmFusedAllreduce(hidden_dim, dtype)
                     time_ms = benchmark_operation(
                         vllm_fused_allreduce.allreduce_rmsnorm_fp4_quant,
                         input_tensor,
@@ -598,6 +645,7 @@ def run_benchmarks(
             VllmConfig(compilation_config=CompilationConfig(custom_ops=["-rms_norm"]))
         ):
             try:
+                vllm_fused_allreduce = VllmFusedAllreduce(hidden_dim, dtype)
                 standard_allreduce_rmsnorm_fp4_quant_native_compiled = torch.compile(
                     vllm_fused_allreduce.allreduce_rmsnorm_fp4_quant,
                     fullgraph=True,
@@ -622,10 +670,12 @@ def run_benchmarks(
                     "inf"
                 )
 
-        # FlashInfer Fused AllReduce + RMSNorm + FP4 Quant Oneshot
-        if flashinfer_comm is not None and allreduce_params is not None:
+        # FlashInfer Fused AllReduce + RMSNorm + FP4 Quant (trtllm only)
+        if "trtllm" in workspaces:
+            trtllm_ws = workspaces["trtllm"]
             for use_oneshot in use_oneshot_options:
                 suffix = "_oneshot" if use_oneshot else "_twoshot"
+                key = f"flashinfer_trtllm_fused_allreduce_rmsnorm_fp4_quant{suffix}"
                 try:
                     time_ms = benchmark_operation(
                         flashinfer_fused_allreduce_rmsnorm_fp4_quant,
@@ -636,49 +686,18 @@ def run_benchmarks(
                         rms_eps=rms_eps,
                         input_global_scale=scale_fp4,
                         allreduce_params=allreduce_params,
+                        workspace=trtllm_ws,
                         quant_out=fp4_quant_out,
                         output_scale=fp4_output_scale,
                         use_oneshot=use_oneshot,
                     )
-                    results[f"flashinfer_fused_allreduce_rmsnorm_fp4_quant{suffix}"] = (
-                        time_ms
-                    )
+                    results[key] = time_ms
                 except Exception as e:
                     logger.error(
-                        "FlashInfer Fused AllReduce+RMSNorm+FP4 Oneshot failed: %s",
+                        "FlashInfer (trtllm) Fused AllReduce+RMSNorm+FP4 failed: %s",
                         e,
                     )
-                    results[f"flashinfer_fused_allreduce_rmsnorm_fp4_quant{suffix}"] = (
-                        float("inf")
-                    )
-
-        # FlashInfer Fused AllReduce + RMSNorm + FP4 Quant Two-shot
-        if flashinfer_comm is not None and allreduce_params is not None:
-            try:
-                time_ms = benchmark_operation(
-                    flashinfer_fused_allreduce_rmsnorm_fp4_quant,
-                    input_tensor,
-                    residual=residual,
-                    norm_out=norm_out,
-                    rms_gamma=rms_gamma,
-                    rms_eps=rms_eps,
-                    input_global_scale=scale_fp4,
-                    allreduce_params=allreduce_params,
-                    quant_out=fp4_quant_out,
-                    output_scale=fp4_output_scale,
-                    use_oneshot=False,
-                )
-                results["flashinfer_fused_allreduce_rmsnorm_fp4_quant_twoshot"] = (
-                    time_ms
-                )
-            except Exception as e:
-                logger.error(
-                    "FlashInfer Fused AllReduce+RMSNorm+FP4 Two-shot failed: %s",
-                    e,
-                )
-                results["flashinfer_fused_allreduce_rmsnorm_fp4_quant_twoshot"] = float(
-                    "inf"
-                )
+                    results[key] = float("inf")
 
     return results
 
@@ -1016,8 +1035,7 @@ def main():
 
     configs = list(itertools.product(args.num_tokens, dtypes, residual_options))
 
-    # Setup FlashInfer workspace if available
-    workspace = None
+    # Setup FlashInfer workspaces for all backends
     allreduce_params = None
 
     if flashinfer_comm is not None:
@@ -1032,15 +1050,17 @@ def main():
             args.hidden_dim * max_element_size
         )
 
-        workspace = setup_flashinfer_workspace(
-            world_size,
-            rank,
-            args.hidden_dim,
-            max_num_token,
-            dtype=workspace_dtype,
-        )
+        for backend in FLASHINFER_BACKENDS:
+            setup_flashinfer_workspace(
+                backend=backend,
+                world_size=world_size,
+                rank=rank,
+                hidden_dim=args.hidden_dim,
+                max_token_num=max_num_token,
+                dtype=workspace_dtype,
+            )
 
-        if workspace is not None:
+        if _FI_WORKSPACES:
             allreduce_params = FlashInferFusedAllReduceParams(
                 max_token_num=max_num_token,
             )
@@ -1066,6 +1086,7 @@ def main():
                 dtype,
                 use_residual,
                 allreduce_params,
+                workspaces=_FI_WORKSPACES,
                 quant_modes=quant_modes,
                 no_oneshot=args.no_oneshot,
             )
@@ -1104,11 +1125,13 @@ def main():
 
     finally:
         # Cleanup
-        if workspace is not None:
-            cleanup_flashinfer_workspace(workspace)
+        cleanup_flashinfer_workspaces()
 
         dist.barrier()
 
 
 if __name__ == "__main__":
-    main()
+    from vllm.config import VllmConfig, set_current_vllm_config
+
+    with set_current_vllm_config(VllmConfig()):
+        main()

benchmarks/kernels/benchmark_layernorm.py

@@ -5,12 +5,14 @@ import time
 
 import torch
 
+from vllm.benchmarks.lib.utils import default_vllm_config
 from vllm.model_executor.layers.layernorm import RMSNorm
 from vllm.utils.argparse_utils import FlexibleArgumentParser
 from vllm.utils.torch_utils import STR_DTYPE_TO_TORCH_DTYPE, set_random_seed
 
 
 @torch.inference_mode()
+@default_vllm_config()
 def main(
     num_tokens: int,
     hidden_size: int,

benchmarks/kernels/benchmark_moe_defaults.py

@@ -0,0 +1,278 @@
+#!/usr/bin/env python3
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""
+Benchmark comparing old vs new default fused MoE configs.
+
+Runs the triton fused_moe kernel with three configurations for each scenario:
+  1. Tuned config (from JSON file, if available) — the target to match
+  2. Old default (the hardcoded defaults before this change)
+  3. New default (the improved defaults)
+
+Usage:
+    python benchmarks/kernels/benchmark_moe_defaults.py
+
+Produces a table showing kernel time (us) and speedup of new vs old defaults.
+"""
+
+import torch
+
+from vllm.model_executor.layers.fused_moe import fused_topk, override_config
+from vllm.model_executor.layers.fused_moe.config import FusedMoEQuantConfig
+from vllm.model_executor.layers.fused_moe.fused_moe import (
+    fused_experts,
+    get_default_config,
+    get_moe_configs,
+)
+from vllm.platforms import current_platform
+from vllm.triton_utils import triton
+from vllm.utils.torch_utils import set_random_seed
+
+FP8_DTYPE = current_platform.fp8_dtype()
+
+
+def old_default_config(M, E, N, K, topk, dtype=None, block_shape=None):
+    """The original defaults before https://github.com/vllm-project/vllm/pull/34846,
+    for comparison."""
+    if dtype == "fp8_w8a8" and block_shape is not None:
+        return {
+            "BLOCK_SIZE_M": 64,
+            "BLOCK_SIZE_N": block_shape[0],
+            "BLOCK_SIZE_K": block_shape[1],
+            "GROUP_SIZE_M": 32,
+            "SPLIT_K": 1,
+            "num_warps": 4,
+            "num_stages": 3 if not current_platform.is_rocm() else 2,
+        }
+    elif M <= E:
+        return {
+            "BLOCK_SIZE_M": 16,
+            "BLOCK_SIZE_N": 32,
+            "BLOCK_SIZE_K": 64,
+            "GROUP_SIZE_M": 1,
+            "SPLIT_K": 1,
+        }
+    else:
+        return {
+            "BLOCK_SIZE_M": 64,
+            "BLOCK_SIZE_N": 64,
+            "BLOCK_SIZE_K": 32,
+            "GROUP_SIZE_M": 8,
+            "SPLIT_K": 1,
+        }
+
+
+def benchmark_config(
+    config,
+    M,
+    E,
+    N,
+    K,
+    topk,
+    dtype,
+    use_fp8=False,
+    block_shape=None,
+    num_iters=100,
+):
+    """Time a single kernel config. Returns kernel time in microseconds."""
+    init_dtype = torch.float16 if use_fp8 else dtype
+
+    a = torch.randn(M, K, device="cuda", dtype=init_dtype) / 10
+    w1 = torch.randn(E, 2 * N, K, device="cuda", dtype=init_dtype) / 10
+    w2 = torch.randn(E, K, N, device="cuda", dtype=init_dtype) / 10
+
+    w1_scale = None
+    w2_scale = None
+    a1_scale = None
+    a2_scale = None
+    if use_fp8:
+        if block_shape is not None:
+            bsn, bsk = block_shape
+            n_tiles_w1 = triton.cdiv(2 * N, bsn)
+            k_tiles_w1 = triton.cdiv(K, bsk)
+            n_tiles_w2 = triton.cdiv(K, bsn)
+            k_tiles_w2 = triton.cdiv(N, bsk)
+            w1_scale = torch.rand(
+                E, n_tiles_w1, k_tiles_w1, device="cuda", dtype=torch.float32
+            )
+            w2_scale = torch.rand(
+                E, n_tiles_w2, k_tiles_w2, device="cuda", dtype=torch.float32
+            )
+        else:
+            w1_scale = torch.rand(E, device="cuda", dtype=torch.float32)
+            w2_scale = torch.rand(E, device="cuda", dtype=torch.float32)
+        a1_scale = torch.rand(1, device="cuda", dtype=torch.float32)
+        a2_scale = torch.rand(1, device="cuda", dtype=torch.float32)
+        # Only weights are stored in fp8; activations stay in bf16/fp16
+        # and get dynamically quantized inside the kernel.
+        w1 = w1.to(FP8_DTYPE)
+        w2 = w2.to(FP8_DTYPE)
+
+    quant_config = FusedMoEQuantConfig.make(
+        quant_dtype=torch.float8_e4m3fn if use_fp8 else None,
+        w1_scale=w1_scale,
+        w2_scale=w2_scale,
+        a1_scale=a1_scale,
+        a2_scale=a2_scale,
+        block_shape=block_shape,
+    )
+
+    gating = torch.randn(M, E, device="cuda", dtype=torch.float32)
+
+    # Warmup
+    for _ in range(20):
+        with override_config(config):
+            topk_weights, topk_ids, _ = fused_topk(a, gating, topk, renormalize=True)
+            fused_experts(
+                a,
+                w1,
+                w2,
+                topk_weights,
+                topk_ids,
+                quant_config=quant_config,
+            )
+    torch.cuda.synchronize()
+
+    # Benchmark
+    start = torch.cuda.Event(enable_timing=True)
+    end = torch.cuda.Event(enable_timing=True)
+    start.record()
+    for _ in range(num_iters):
+        with override_config(config):
+            topk_weights, topk_ids, _ = fused_topk(a, gating, topk, renormalize=True)
+            fused_experts(
+                a,
+                w1,
+                w2,
+                topk_weights,
+                topk_ids,
+                quant_config=quant_config,
+            )
+    end.record()
+    torch.cuda.synchronize()
+    return start.elapsed_time(end) / num_iters * 1000  # ms -> us
+
+
+# Model configurations: (name, E, N, K, topk, dtype_str, use_fp8, block_shape)
+# N = moe_intermediate_size // tp_size (the value used in config file lookup)
+MODELS = [
+    # --- Few experts ---
+    ("Mixtral bf16", 8, 7168, 4096, 2, None, False, None),
+    ("Mixtral fp8", 8, 7168, 4096, 2, "fp8_w8a8", True, None),
+    # --- Many experts: real model shapes at tp=1 ---
+    # Qwen2-MoE-57B: E=60, topk=4, N=1408, K=2048
+    ("Qwen2-MoE bf16", 60, 1408, 2048, 4, None, False, None),
+    # DeepSeek-V2: E=64, topk=6, N=1407, K=4096
+    # (use 1408 to avoid odd alignment; real model is 1407)
+    ("DeepSeek-V2 bf16", 64, 1408, 4096, 6, None, False, None),
+    # OLMoE-7B: E=64, topk=8, N=2048, K=2048
+    ("OLMoE bf16", 64, 2048, 2048, 8, None, False, None),
+    # GLM-4-100B-A10B: E=128, topk=8, N=1408, K=4096
+    ("GLM-4-MoE bf16", 128, 1408, 4096, 8, None, False, None),
+    # Qwen3-30B-A3B: E=128, topk=8, N=768, K=2048
+    ("Qwen3-MoE bf16", 128, 768, 2048, 8, None, False, None),
+    # DeepSeek-V3 / MiMo-V2-Flash: E=256, topk=8, N=2048, K=7168
+    ("DeepSeek-V3 bf16", 256, 2048, 7168, 8, None, False, None),
+    # Qwen3.5-70B-A22B (Qwen3-Next): E=512, topk=10, N=512, K=2048
+    ("Qwen3-Next bf16", 512, 512, 2048, 10, None, False, None),
+    # E=128 N=1856 bf16
+    ("E128 N1856 bf16", 128, 1856, 4096, 8, None, False, None),
+    # E=256 N=512 bf16 (DS-V3 tp=4)
+    ("DS-V3 tp4 bf16", 256, 512, 7168, 8, None, False, None),
+    # E=512 N=512 bf16 (Qwen3-Next tp=1)
+    ("Qwen3-Next bf16", 512, 512, 2048, 10, None, False, None),
+    # E=512 N=256 bf16 (Qwen3-Next tp=2)
+    ("Qwen3-Next tp2", 512, 256, 2048, 10, None, False, None),
+    # --- FP8 block quant (many experts) ---
+    # DS-V3 tp=4: E=256, N=512, fp8 block
+    ("DS-V3 tp4 fp8blk", 256, 512, 7168, 8, "fp8_w8a8", True, [128, 128]),
+    # DS-V3 tp=8: E=256, N=256, fp8 block
+    ("DS-V3 tp8 fp8blk", 256, 256, 7168, 8, "fp8_w8a8", True, [128, 128]),
+    # Qwen3-Next tp=2 fp8 block
+    ("Qwen3-Next tp2 fp8blk", 512, 256, 2048, 10, "fp8_w8a8", True, [128, 128]),
+]
+
+BATCH_SIZES = [1, 4, 8, 16, 32, 64, 128, 256, 512, 1024, 2048, 4096]
+
+
+def main():
+    set_random_seed(0)
+    torch.set_default_device("cuda")
+    dtype = torch.bfloat16
+
+    for name, E, N, K, topk, dtype_str, use_fp8, block_shape in MODELS:
+        print(f"\n{'=' * 90}")
+        print(f"  {name}  (E={E}, N={N}, K={K}, topk={topk})")
+        print(f"{'=' * 90}")
+
+        # Try to load tuned config
+        block_n = block_shape[0] if block_shape else None
+        block_k = block_shape[1] if block_shape else None
+        tuned = get_moe_configs(E, N, dtype_str, block_n, block_k)
+        has_tuned = tuned is not None
+        print(f"  Tuned config available: {has_tuned}")
+
+        hdr = (
+            f"{'Batch':>6} | {'Tuned (us)':>11} | {'Old (us)':>11} | "
+            f"{'New (us)':>11} | {'New/Old':>8} | {'New/Tuned':>10}"
+        )
+        print(f"  {hdr}")
+        print(f"  {'-' * len(hdr)}")
+
+        for M in BATCH_SIZES:
+            old_cfg = old_default_config(M, E, N, K, topk, dtype_str, block_shape)
+            new_cfg = get_default_config(M, E, N, K, topk, dtype_str, block_shape)
+
+            if has_tuned:
+                tuned_cfg = tuned[min(tuned.keys(), key=lambda x: abs(x - M))]
+                t_tuned = benchmark_config(
+                    tuned_cfg,
+                    M,
+                    E,
+                    N,
+                    K,
+                    topk,
+                    dtype,
+                    use_fp8=use_fp8,
+                    block_shape=block_shape,
+                )
+            else:
+                t_tuned = None
+
+            t_old = benchmark_config(
+                old_cfg,
+                M,
+                E,
+                N,
+                K,
+                topk,
+                dtype,
+                use_fp8=use_fp8,
+                block_shape=block_shape,
+            )
+            t_new = benchmark_config(
+                new_cfg,
+                M,
+                E,
+                N,
+                K,
+                topk,
+                dtype,
+                use_fp8=use_fp8,
+                block_shape=block_shape,
+            )
+
+            ratio_new_old = t_new / t_old
+            tuned_str = f"{t_tuned:11.2f}" if t_tuned else f"{'N/A':>11}"
+            ratio_tuned = f"{t_new / t_tuned:10.2f}x" if t_tuned else f"{'N/A':>10}"
+            # flag regressions where new default is >5% slower than old
+            marker = " <--" if ratio_new_old > 1.05 else ""
+
+            print(
+                f"  {M:>6} | {tuned_str} | {t_old:11.2f} | {t_new:11.2f} "
+                f"| {ratio_new_old:7.2f}x | {ratio_tuned}{marker}"
+            )
+
+
+if __name__ == "__main__":
+    main()

benchmarks/kernels/benchmark_mrope.py

@@ -36,6 +36,7 @@ from typing import Any
 import numpy as np
 import torch
 
+from vllm.benchmarks.lib.utils import default_vllm_config
 from vllm.model_executor.layers.rotary_embedding import get_rope
 from vllm.transformers_utils.config import get_config
 from vllm.utils.argparse_utils import FlexibleArgumentParser
@@ -78,6 +79,7 @@ def calculate_stats(times: list[float]) -> dict[str, float]:
     }
 
 
+@default_vllm_config()
 def benchmark_mrope(
     model_name: str,
     num_tokens: int,

benchmarks/kernels/benchmark_per_token_quant_fp8.py

@@ -7,6 +7,7 @@ from unittest.mock import patch
 import pandas as pd
 import torch
 
+from vllm.benchmarks.lib.utils import default_vllm_config
 from vllm.model_executor.layers.quantization.input_quant_fp8 import QuantFP8
 from vllm.model_executor.layers.quantization.utils.quant_utils import GroupShape
 from vllm.triton_utils import triton
@@ -84,6 +85,7 @@ def calculate_diff(
 configs = []
 
 
+@default_vllm_config()
 def benchmark_quantization(
     batch_size,
     hidden_size,

benchmarks/kernels/benchmark_rope.py

@@ -5,6 +5,7 @@ import itertools
 
 import torch
 
+from vllm.benchmarks.lib.utils import default_vllm_config
 from vllm.model_executor.layers.rotary_embedding import get_rope
 from vllm.triton_utils import triton
 from vllm.utils.argparse_utils import FlexibleArgumentParser
@@ -29,6 +30,7 @@ def get_benchmark(head_size, rotary_dim, is_neox_style, device):
             args={},
         )
     )
+    @default_vllm_config()
     def benchmark(batch_size, seq_len, num_heads, provider):
         dtype = torch.bfloat16
         max_position = 8192

csrc/activation_kernels.cu

@@ -14,7 +14,8 @@ struct alignas(32) u32x8_t {
 };
 
 __device__ __forceinline__ void ld256(u32x8_t& val, const u32x8_t* ptr) {
-#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 1000
+#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 1000 && \
+    defined(CUDA_VERSION) && CUDA_VERSION >= 12090
   asm volatile("ld.global.nc.v8.u32 {%0,%1,%2,%3,%4,%5,%6,%7}, [%8];\n"
                : "=r"(val.u0), "=r"(val.u1), "=r"(val.u2), "=r"(val.u3),
                  "=r"(val.u4), "=r"(val.u5), "=r"(val.u6), "=r"(val.u7)
@@ -35,7 +36,8 @@ __device__ __forceinline__ void ld256(u32x8_t& val, const u32x8_t* ptr) {
 }
 
 __device__ __forceinline__ void st256(u32x8_t& val, u32x8_t* ptr) {
-#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 1000
+#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 1000 && \
+    defined(CUDA_VERSION) && CUDA_VERSION >= 12090
   asm volatile("st.global.v8.u32 [%0], {%1,%2,%3,%4,%5,%6,%7,%8};\n"
                :
                : "l"(ptr), "r"(val.u0), "r"(val.u1), "r"(val.u2), "r"(val.u3),

csrc/cache_kernels.cu

@@ -1305,7 +1305,8 @@ void indexer_k_quant_and_cache(
   const at::cuda::OptionalCUDAGuard device_guard(device_of(k));
   const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
 
-  DISPATCH_BY_KV_CACHE_DTYPE(k.dtype(), "fp8_e4m3",
+  static const std::string kv_cache_dtype = "fp8_e4m3";
+  DISPATCH_BY_KV_CACHE_DTYPE(k.dtype(), kv_cache_dtype,
                              CALL_INDEXER_K_QUANT_AND_CACHE);
 }
 

csrc/cpu/cpu_attn.cpp

@@ -16,6 +16,8 @@ torch::Tensor get_scheduler_metadata(
     isa = cpu_attention::ISA::VEC16;
   } else if (isa_hint == "neon") {
     isa = cpu_attention::ISA::NEON;
+  } else if (isa_hint == "vxe") {
+    isa = cpu_attention::ISA::VXE;
   } else {
     TORCH_CHECK(false, "Unsupported CPU attention ISA hint: " + isa_hint);
   }
@@ -100,6 +102,8 @@ void cpu_attn_reshape_and_cache(
       return cpu_attention::ISA::VEC16;
     } else if (isa == "neon") {
       return cpu_attention::ISA::NEON;
+    } else if (isa == "vxe") {
+      return cpu_attention::ISA::VXE;
     } else {
       TORCH_CHECK(false, "Invalid ISA type: " + isa);
     }

csrc/cpu/cpu_attn_impl.hpp

@@ -12,7 +12,7 @@
 #include "cpu/utils.hpp"
 
 namespace cpu_attention {
-enum class ISA { AMX, VEC, VEC16, NEON };
+enum class ISA { AMX, VEC, VEC16, NEON, VXE };
 
 template <ISA isa, typename scalar_t, int64_t head_dim>
 class AttentionImpl {};

csrc/cpu/cpu_attn_vxe.hpp

@@ -0,0 +1,386 @@
+#ifndef CPU_ATTN_VXE_HPP
+#define CPU_ATTN_VXE_HPP
+
+#include "cpu_attn_impl.hpp"
+#include <vecintrin.h>
+#include <type_traits>
+
+namespace cpu_attention {
+
+namespace {
+
+// s390x Vector = 16 bytes (128 bits)
+#define BLOCK_SIZE_ALIGNMENT 32
+#define HEAD_SIZE_ALIGNMENT 32
+#define MAX_Q_HEAD_NUM_PER_ITER 16
+
+template <typename kv_cache_t>
+FORCE_INLINE void load_row8_B_as_f32(const kv_cache_t* p, __vector float& b0,
+                                     __vector float& b1);
+
+// [1] Float Specialization
+template <>
+FORCE_INLINE void load_row8_B_as_f32<float>(const float* p, __vector float& b0,
+                                            __vector float& b1) {
+  // Explicitly cast to long long for offset, and float* for pointer
+  b0 = vec_xl((long long)0, const_cast<float*>(p));
+  b1 = vec_xl((long long)0, const_cast<float*>(p + 4));
+}
+
+// [2] BFloat16 Specialization (Big Endian Fix)
+template <>
+FORCE_INLINE void load_row8_B_as_f32<c10::BFloat16>(const c10::BFloat16* p,
+                                                    __vector float& b0,
+                                                    __vector float& b1) {
+  // 1. Load 8 BF16s (16 bytes) into one vector
+  // Explicit cast to unsigned short* for vec_xl to return vector unsigned short
+  __vector unsigned short raw = vec_xl((long long)0, (unsigned short*)p);
+
+  // 2. Prepare Zero vector
+  __vector unsigned short zeros = vec_splat_u16(0);
+
+  // 3. Merge High/Low to expand BF16 -> Float32
+  // On Big Endian, a float is [BF16_bits | 16_zero_bits]
+  b0 = (__vector float)vec_mergeh(raw, zeros);
+  b1 = (__vector float)vec_mergel(raw, zeros);
+}
+
+template <>
+FORCE_INLINE void load_row8_B_as_f32<c10::Half>(const c10::Half* p,
+                                                __vector float& b0,
+                                                __vector float& b1) {
+  alignas(16) float tmp[8];
+
+  // Manual unroll / conversion
+  tmp[0] = static_cast<float>(p[0]);
+  tmp[1] = static_cast<float>(p[1]);
+  tmp[2] = static_cast<float>(p[2]);
+  tmp[3] = static_cast<float>(p[3]);
+  tmp[4] = static_cast<float>(p[4]);
+  tmp[5] = static_cast<float>(p[5]);
+  tmp[6] = static_cast<float>(p[6]);
+  tmp[7] = static_cast<float>(p[7]);
+
+  // Explicit arguments for intrinsic: (long long offset, float* ptr)
+  b0 = vec_xl((long long)0, (float*)tmp);
+  b1 = vec_xl((long long)0, (float*)(tmp + 4));
+}
+
+template <int32_t M, typename kv_cache_t>
+FORCE_INLINE void gemm_micro_s390x_Mx8_Ku4(
+    const float* __restrict A,       // [M x K]
+    const kv_cache_t* __restrict B,  // [K x 8]
+    float* __restrict C,             // [M x 8]
+    int64_t lda, int64_t ldb, int64_t ldc, int32_t K, bool accumulate) {
+  static_assert(1 <= M && M <= 8, "M must be in [1,8]");
+
+// Helper macros to unroll codegen for M rows
+#define ROWS_APPLY(OP) OP(0) OP(1) OP(2) OP(3) OP(4) OP(5) OP(6) OP(7)
+#define IF_M(i) if constexpr (M > (i))
+
+  // 1. Define A pointers
+#define DECL_A(i) const float* a##i = A + (i) * lda;
+  ROWS_APPLY(DECL_A)
+#undef DECL_A
+
+  // 2. Define Accumulators (2 vectors covers 8 columns)
+#define DECL_ACC(i) __vector float acc##i##_0, acc##i##_1;
+  ROWS_APPLY(DECL_ACC)
+#undef DECL_ACC
+
+  // 3. Initialize Accumulators (Load C or Zero)
+#define INIT_ACC(i)                                                    \
+  IF_M(i) {                                                            \
+    if (accumulate) {                                                  \
+      acc##i##_0 =                                                     \
+          vec_xl((long long)0, const_cast<float*>(C + (i) * ldc + 0)); \
+      acc##i##_1 =                                                     \
+          vec_xl((long long)0, const_cast<float*>(C + (i) * ldc + 4)); \
+    } else {                                                           \
+      acc##i##_0 = vec_splats(0.0f);                                   \
+      acc##i##_1 = vec_splats(0.0f);                                   \
+    }                                                                  \
+  }
+  ROWS_APPLY(INIT_ACC)
+#undef INIT_ACC
+
+  int32_t k = 0;
+
+  for (; k + 3 < K; k += 4) {
+    // Load 4 values of A for each Row M: A[k...k+3]
+#define LOAD_A4(i)        \
+  __vector float a##i##v; \
+  IF_M(i) a##i##v = vec_xl((long long)0, const_cast<float*>(a##i + k));
+    ROWS_APPLY(LOAD_A4)
+#undef LOAD_A4
+
+    // Helper: FMA for specific lane L of A
+    // s390x: vec_madd(b, vec_splat(a, lane), acc)
+#define FMAS_LANE(i, aiv, L)                        \
+  IF_M(i) {                                         \
+    __vector float a_broad = vec_splat(aiv, L);     \
+    acc##i##_0 = vec_madd(b0, a_broad, acc##i##_0); \
+    acc##i##_1 = vec_madd(b1, a_broad, acc##i##_1); \
+  }
+
+    // Unroll K=0..3
+    {
+      __vector float b0, b1;
+      load_row8_B_as_f32<kv_cache_t>(B + (int64_t)(k + 0) * ldb, b0, b1);
+#define STEP_K0(i) FMAS_LANE(i, a##i##v, 0)
+      ROWS_APPLY(STEP_K0)
+#undef STEP_K0
+    }
+    {
+      __vector float b0, b1;
+      load_row8_B_as_f32<kv_cache_t>(B + (int64_t)(k + 1) * ldb, b0, b1);
+#define STEP_K1(i) FMAS_LANE(i, a##i##v, 1)
+      ROWS_APPLY(STEP_K1)
+#undef STEP_K1
+    }
+    {
+      __vector float b0, b1;
+      load_row8_B_as_f32<kv_cache_t>(B + (int64_t)(k + 2) * ldb, b0, b1);
+#define STEP_K2(i) FMAS_LANE(i, a##i##v, 2)
+      ROWS_APPLY(STEP_K2)
+#undef STEP_K2
+    }
+
+    {
+      __vector float b0, b1;
+      load_row8_B_as_f32<kv_cache_t>(B + (int64_t)(k + 3) * ldb, b0, b1);
+#define STEP_K3(i) FMAS_LANE(i, a##i##v, 3)
+      ROWS_APPLY(STEP_K3)
+#undef STEP_K3
+    }
+#undef FMAS_LANE
+  }
+
+  for (; k < K; ++k) {
+    __vector float b0, b1;
+    load_row8_B_as_f32<kv_cache_t>(B + (int64_t)k * ldb, b0, b1);
+#define TAIL_ROW(i)                              \
+  IF_M(i) {                                      \
+    __vector float ai = vec_splats(*(a##i + k)); \
+    acc##i##_0 = vec_madd(b0, ai, acc##i##_0);   \
+    acc##i##_1 = vec_madd(b1, ai, acc##i##_1);   \
+  }
+    ROWS_APPLY(TAIL_ROW)
+#undef TAIL_ROW
+  }
+
+#define STORE_ROW(i)                           \
+  IF_M(i) {                                    \
+    vec_xst(acc##i##_0, 0, C + (i) * ldc + 0); \
+    vec_xst(acc##i##_1, 0, C + (i) * ldc + 4); \
+  }
+  ROWS_APPLY(STORE_ROW)
+#undef STORE_ROW
+
+#undef ROWS_APPLY
+#undef IF_M
+}
+
+template <int32_t N, typename kv_cache_t>
+FORCE_INLINE void gemm_macro_s390x_Mx8_Ku4(const float* __restrict A,
+                                           const kv_cache_t* __restrict B,
+                                           float* __restrict C, int32_t M,
+                                           int32_t K, int64_t lda, int64_t ldb,
+                                           int64_t ldc, bool accumulate) {
+  static_assert(N % 8 == 0, "N must be a multiple of 8");
+  for (int32_t m = 0; m < M;) {
+    int32_t mb = (M - m >= 8) ? 8 : (M - m >= 4) ? 4 : (M - m >= 2) ? 2 : 1;
+    const float* Ab = A + m * lda;
+    float* Cb = C + m * ldc;
+
+    for (int32_t n = 0; n < N; n += 8) {
+      const kv_cache_t* Bn = B + n;
+      float* Cn = Cb + n;
+      switch (mb) {
+        case 8:
+          gemm_micro_s390x_Mx8_Ku4<8, kv_cache_t>(Ab, Bn, Cn, lda, ldb, ldc, K,
+                                                  accumulate);
+          break;
+        case 4:
+          gemm_micro_s390x_Mx8_Ku4<4, kv_cache_t>(Ab, Bn, Cn, lda, ldb, ldc, K,
+                                                  accumulate);
+          break;
+        case 2:
+          gemm_micro_s390x_Mx8_Ku4<2, kv_cache_t>(Ab, Bn, Cn, lda, ldb, ldc, K,
+                                                  accumulate);
+          break;
+        default:
+          gemm_micro_s390x_Mx8_Ku4<1, kv_cache_t>(Ab, Bn, Cn, lda, ldb, ldc, K,
+                                                  accumulate);
+          break;
+      }
+    }
+    m += mb;
+  }
+}
+
+template <typename kv_cache_t>
+class TileGemmS390X {
+ public:
+  template <AttentionGemmPhase phase, int32_t k_size>
+  FORCE_INLINE static void gemm(const int32_t m_size,
+                                float* __restrict__ a_tile,
+                                kv_cache_t* __restrict__ b_tile,
+                                float* __restrict__ c_tile, const int64_t lda,
+                                const int64_t ldb, const int64_t ldc,
+                                const int32_t block_size,
+                                const int32_t dynamic_k_size,
+                                const bool accum_c) {
+    if constexpr (phase == AttentionGemmPhase::QK) {
+      gemm_macro_s390x_Mx8_Ku4<BLOCK_SIZE_ALIGNMENT, kv_cache_t>(
+          a_tile, b_tile, c_tile, m_size, k_size, lda, ldb, ldc, accum_c);
+    } else {
+      gemm_macro_s390x_Mx8_Ku4<HEAD_SIZE_ALIGNMENT, kv_cache_t>(
+          a_tile, b_tile, c_tile, m_size, dynamic_k_size, lda, ldb, ldc,
+          accum_c);
+    }
+  }
+};
+
+}  // namespace
+
+template <typename scalar_t, int64_t head_dim>
+class AttentionImpl<ISA::VXE, scalar_t, head_dim> {
+ public:
+  using query_t = scalar_t;
+  using q_buffer_t = float;
+  using kv_cache_t = scalar_t;
+  using logits_buffer_t = float;
+  using partial_output_buffer_t = float;
+  using prob_buffer_t = float;
+
+  constexpr static int64_t BlockSizeAlignment = BLOCK_SIZE_ALIGNMENT;
+  constexpr static int64_t HeadDimAlignment = HEAD_SIZE_ALIGNMENT;
+  constexpr static int64_t MaxQHeadNumPerIteration = MAX_Q_HEAD_NUM_PER_ITER;
+  constexpr static int64_t HeadDim = head_dim;
+  constexpr static ISA ISAType = ISA::VXE;
+  constexpr static bool scale_on_logits =
+      false;  // Scale is applied to Q during copy
+
+ public:
+  AttentionImpl() {}
+
+  template <template <typename tile_gemm_t> typename attention>
+  FORCE_INLINE void execute_attention(DEFINE_CPU_ATTENTION_PARAMS) {
+    attention<TileGemmS390X<kv_cache_t>> attention_iteration;
+    attention_iteration(CPU_ATTENTION_PARAMS);
+  }
+
+  // Strides for Memory Layout
+  constexpr static int64_t k_cache_token_group_stride(
+      const int32_t block_size) {
+    return BlockSizeAlignment;  // [head_dim, block_size] layout
+  }
+
+  constexpr static int64_t v_cache_token_group_stride(
+      const int32_t block_size) {
+    return head_dim * BlockSizeAlignment;
+  }
+
+  constexpr static int64_t v_cache_head_group_stride(const int32_t block_size) {
+    return HeadDimAlignment;
+  }
+
+  static void copy_q_heads_tile(scalar_t* __restrict__ src,
+                                float* __restrict__ q_buffer,
+                                const int32_t q_num,
+                                const int32_t q_heads_per_kv,
+                                const int64_t q_num_stride,
+                                const int64_t q_head_stride, float scale) {
+    __vector float scale_vec = vec_splats(scale);
+    constexpr bool is_bf16 = std::is_same<scalar_t, c10::BFloat16>::value;
+
+    // Process 8 elements at a time (32 bytes of float output)
+    for (int32_t i = 0; i < q_num; ++i) {
+      for (int32_t h = 0; h < q_heads_per_kv; ++h) {
+        scalar_t* curr_src = src + i * q_num_stride + h * q_head_stride;
+        float* curr_dst =
+            q_buffer + i * q_heads_per_kv * head_dim + h * head_dim;
+
+        int32_t d = 0;
+        for (; d <= head_dim - 8; d += 8) {
+          if constexpr (is_bf16) {
+            __vector float v0, v1;
+            // Reuse our Big-Endian-Safe loader
+            load_row8_B_as_f32<scalar_t>(curr_src + d, v0, v1);
+
+            v0 = vec_mul(v0, scale_vec);
+            v1 = vec_mul(v1, scale_vec);
+
+            vec_xst(v0, 0, curr_dst + d);
+            vec_xst(v1, 0, curr_dst + d + 4);
+          } else {
+            __vector float v0 = vec_xl((long long)0, (float*)curr_src + d);
+            __vector float v1 = vec_xl((long long)0, (float*)curr_src + d + 4);
+
+            v0 = vec_mul(v0, scale_vec);
+            v1 = vec_mul(v1, scale_vec);
+
+            vec_xst(v0, 0, curr_dst + d);
+            vec_xst(v1, 0, curr_dst + d + 4);
+          }
+        }
+
+        for (; d < head_dim; ++d) {
+          float val = static_cast<float>(curr_src[d]);
+          curr_dst[d] = val * scale;
+        }
+      }
+    }
+  }
+
+  static void reshape_and_cache(
+      const scalar_t* __restrict__ key, const scalar_t* __restrict__ value,
+      scalar_t* __restrict__ key_cache, scalar_t* __restrict__ value_cache,
+      const int64_t* __restrict__ slot_mapping, const int64_t token_num,
+      const int64_t key_token_num_stride, const int64_t value_token_num_stride,
+      const int64_t head_num, const int64_t key_head_num_stride,
+      const int64_t value_head_num_stride, const int64_t num_blocks,
+      const int64_t num_blocks_stride, const int64_t cache_head_num_stride,
+      const int64_t block_size, const int64_t block_size_stride) {
+#pragma omp parallel for collapse(2)
+    for (int64_t token_idx = 0; token_idx < token_num; ++token_idx) {
+      for (int64_t head_idx = 0; head_idx < head_num; ++head_idx) {
+        const int64_t pos = slot_mapping[token_idx];
+        if (pos < 0) continue;
+
+        const int64_t block_idx = pos / block_size;
+        const int64_t block_offset = pos % block_size;
+
+        {
+          const scalar_t* key_src = key + token_idx * key_token_num_stride +
+                                    head_idx * key_head_num_stride;
+          scalar_t* key_dst = key_cache + block_idx * num_blocks_stride +
+                              head_idx * cache_head_num_stride + block_offset;
+
+          for (int64_t i = 0, j = 0; i < head_dim; ++i, j += block_size) {
+            key_dst[j] = key_src[i];
+          }
+        }
+
+        {
+          const scalar_t* val_src = value + token_idx * value_token_num_stride +
+                                    head_idx * value_head_num_stride;
+          scalar_t* val_dst = value_cache + block_idx * num_blocks_stride +
+                              head_idx * cache_head_num_stride +
+                              block_offset * head_dim;
+
+          std::memcpy(val_dst, val_src, sizeof(scalar_t) * head_dim);
+        }
+      }
+    }
+  }
+};
+
+}  // namespace cpu_attention
+
+#undef BLOCK_SIZE_ALIGNMENT
+#undef HEAD_SIZE_ALIGNMENT
+#undef MAX_Q_HEAD_NUM_PER_ITER
+
+#endif

csrc/cpu/generate_cpu_attn_dispatch.py

@@ -19,10 +19,11 @@ ISA_TYPES = {
     "VEC": 1,
     "VEC16": 2,
     "NEON": 3,
+    "VXE": 4,
 }
 
 # ISAs supported for head_dims divisible by 32
-ISA_FOR_32 = ["AMX", "NEON", "VEC", "VEC16"]
+ISA_FOR_32 = ["AMX", "NEON", "VEC", "VEC16", "VXE"]
 
 # ISAs supported for head_dims divisible by 16 only
 ISA_FOR_16 = ["VEC16"]
@@ -118,6 +119,10 @@ def generate_header_file() -> str:
   #include "cpu_attn_neon.hpp"
 #endif
 
+#ifdef __s390x__
+  #include "cpu_attn_vxe.hpp"
+#endif
+
 """
 
     header += generate_helper_function()
@@ -163,6 +168,25 @@ def generate_header_file() -> str:
     } \\
   }()
 
+"""
+
+    # s390x with VXE
+    header += """#elif defined(__s390x__)
+#define CPU_ATTN_DISPATCH(HEAD_DIM, ISA_TYPE, ...) \\
+  [&] { \\
+    int64_t encoded_params = encode_cpu_attn_params(HEAD_DIM, ISA_TYPE); \\
+    switch (encoded_params) { \\
+"""
+    header += generate_cases_for_isa_group(["VXE", "VEC", "VEC16"])
+    header += """
+      default: { \\
+        TORCH_CHECK(false, "Unsupported CPU attention configuration: head_dim=" + \\
+                    std::to_string(HEAD_DIM) + " isa=" + \\
+                    std::to_string(static_cast<int>(ISA_TYPE))); \\
+      } \\
+    } \\
+  }()
+
 """
 
     # Fallback: VEC and VEC16 only
@@ -182,7 +206,7 @@ def generate_header_file() -> str:
     } \\
   }()
 
-#endif  /* CPU_CAPABILITY_AMXBF16 / __aarch64__ */
+#endif  /* CPU_CAPABILITY_AMXBF16 / __aarch64__ / __s390x__ */
 
 #endif  // CPU_ATTN_DISPATCH_GENERATED_H
 """

csrc/dsv3_fused_a_gemm.cu

@@ -0,0 +1,751 @@
+/*
+ * Adapted from
+ * https://github.com/sgl-project/sglang/blob/main/sgl-kernel/csrc/gemm/dsv3_fused_a_gemm.cu
+ * which was adapted from
+ * https://github.com/NVIDIA/TensorRT-LLM/blob/619709fc33bd5dc268f19d6a741fe7ed51c0f8f5/cpp/tensorrt_llm/kernels/dsv3MinLatencyKernels/dsv3FusedAGemm.cu
+ *
+ * Copyright (c) 2019-2024, NVIDIA CORPORATION.  All rights reserved.
+ * Copyright (c) 2021, NAVER Corp.  Authored by CLOVA.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include <ATen/ATen.h>
+#include <ATen/cuda/CUDAContext.h>
+#include <cuda_bf16.h>
+#include <cuda_runtime.h>
+#include <torch/all.h>
+
+#include "core/registration.h"
+
+#include <cstdlib>
+#include <mutex>
+
+namespace {
+
+inline int getSMVersion() {
+  auto* props = at::cuda::getCurrentDeviceProperties();
+  return props->major * 10 + props->minor;
+}
+
+inline bool getEnvEnablePDL() {
+  static std::once_flag flag;
+  static bool enablePDL = false;
+  std::call_once(flag, [&]() {
+    if (getSMVersion() >= 90) {
+      char const* env = std::getenv("TRTLLM_ENABLE_PDL");
+      enablePDL = env && env[0] == '1' && env[1] == '\0';
+    }
+  });
+  return enablePDL;
+}
+
+}  // namespace
+
+using bf16_t = __nv_bfloat16;
+
+__device__ void hmma_16_8_16_f32acc_bf16ab(float (&d_reg)[4],
+                                           const bf16_t (&a_reg)[8],
+                                           const bf16_t (&b_reg)[4],
+                                           float const (&c_reg)[4]) {
+#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 900
+  uint32_t a0 = *reinterpret_cast<uint32_t const*>(a_reg + 0);
+  uint32_t a1 = *reinterpret_cast<uint32_t const*>(a_reg + 2);
+  uint32_t a2 = *reinterpret_cast<uint32_t const*>(a_reg + 4);
+  uint32_t a3 = *reinterpret_cast<uint32_t const*>(a_reg + 6);
+  uint32_t b0 = *reinterpret_cast<uint32_t const*>(b_reg + 0);
+  uint32_t b1 = *reinterpret_cast<uint32_t const*>(b_reg + 2);
+  asm volatile(
+      "mma.sync.aligned.m16n8k16.row.col.f32.bf16.bf16.f32 "
+      "{%0,  %1,  %2,  %3},"
+      "{%4,  %5,  %6,  %7},"
+      "{%8,  %9},"
+      "{%10, %11, %12, %13};\n"
+      : "=f"(d_reg[0]), "=f"(d_reg[1]), "=f"(d_reg[2]), "=f"(d_reg[3])
+      : "r"(a0), "r"(a1), "r"(a2), "r"(a3), "r"(b0), "r"(b1), "f"(d_reg[0]),
+        "f"(d_reg[1]), "f"(d_reg[2]), "f"(d_reg[3]));
+#endif
+}
+
+extern "C" {
+__device__ uint32_t __nvvm_get_smem_pointer(void*);
+}
+
+__device__ void ldgsts_128(void const* gPtr, void* sPtr, uint32_t pred) {
+#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 900
+  if (pred) {
+    uint32_t smemPtrAsUint32 = __nvvm_get_smem_pointer(sPtr);
+    asm volatile("cp.async.cg.shared.global.L2::128B [%0], [%1], %2;\n" ::"r"(
+                     smemPtrAsUint32),
+                 "l"(gPtr), "n"(16));
+  }
+#endif
+}
+
+__device__ void ldsm_x4(void* smem_ptr, uint32_t* reg_ptr) {
+#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 900
+  asm volatile(
+      "ldmatrix.sync.aligned.x4.m8n8.shared.b16 {%0, %1, %2, %3}, [%4];\n"
+      : "=r"(reg_ptr[0]), "=r"(reg_ptr[1]), "=r"(reg_ptr[2]), "=r"(reg_ptr[3])
+      : "r"(__nvvm_get_smem_pointer(smem_ptr)));
+#endif
+}
+
+template <class Type>
+__device__ int apply_swizzle_343_on_elem_row_col(int row_idx_, int col_idx_) {
+  uint32_t row_idx = *reinterpret_cast<uint32_t*>(&row_idx_);
+  uint32_t col_idx = *reinterpret_cast<uint32_t*>(&col_idx_);
+  row_idx = row_idx % 8;
+  row_idx = row_idx * (16 / sizeof(Type));
+  col_idx = col_idx ^ row_idx;
+  return *reinterpret_cast<int*>(&col_idx);
+}
+
+__device__ void initialize_barrier(
+    uint64_t* smem_barrier,  // 64 bits user-manged barrier in smem
+    int thread_count =
+        1)  // Thread count expected to arrive/wait on this barrier
+{
+#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 900
+  uint32_t smem_int_ptr = __nvvm_get_smem_pointer(smem_barrier);
+  asm volatile("mbarrier.init.shared::cta.b64 [%0], %1;\n" ::"r"(smem_int_ptr),
+               "r"(thread_count));
+#endif
+}
+
+// Barrier wait
+__device__ void wait_barrier(
+    uint64_t* smem_barrier,  // 64 bits user-manged barrier in smem
+    int phase_bit)           // Current phase bit the barrier waiting to flip
+{
+#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 900
+  uint32_t smem_int_ptr = __nvvm_get_smem_pointer(smem_barrier);
+  asm volatile(
+      "{\n"
+      ".reg .pred                P1;\n"
+      "LAB_WAIT:\n"
+      "mbarrier.try_wait.parity.shared::cta.b64 P1, [%0], %1;\n"
+      "@P1                       bra DONE;\n"
+      "bra                   LAB_WAIT;\n"
+      "DONE:\n"
+      "}\n" ::"r"(smem_int_ptr),
+      "r"(phase_bit));
+#endif
+}
+
+__device__ bool try_wait_barrier(uint64_t* smem_ptr, int phase_bit) {
+#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 900
+  uint32_t wait_complete;
+  uint32_t smem_int_ptr = __nvvm_get_smem_pointer(smem_ptr);
+  asm volatile(
+      "{\n\t"
+      ".reg .pred P1; \n\t"
+      "mbarrier.try_wait.parity.shared::cta.b64 P1, [%1], %2; \n\t"
+      "selp.b32 %0, 1, 0, P1; \n\t"
+      "}"
+      : "=r"(wait_complete)
+      : "r"(smem_int_ptr), "r"(phase_bit));
+  return static_cast<bool>(wait_complete);
+#endif
+  return false;
+}
+
+// Barrier arrive
+__device__ void arrive_barrier(
+    uint64_t* smem_barrier)  // 64 bits user-manged barrier in smem
+{
+#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 900
+  uint32_t smem_int_ptr = __nvvm_get_smem_pointer(smem_barrier);
+  asm volatile(
+      "{\n"
+      ".reg .b64 state; \n"
+      "mbarrier.arrive.shared::cta.b64   state, [%0];\n"
+      "}\n" ::"r"(smem_int_ptr));
+#endif
+}
+
+__device__ void ldgsts_arrive(uint64_t* smem_barrier) {
+#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 900
+  uint32_t smem_int_ptr = __nvvm_get_smem_pointer(smem_barrier);
+  asm volatile("cp.async.mbarrier.arrive.noinc.shared.b64 [%0];"
+               :
+               : "r"(smem_int_ptr));
+#endif
+}
+
+template <int gemm_k, int tile_m, int tile_k, int stage_cnt>
+struct GmemLoaderA {
+  static constexpr int elem_bytes = 2;
+  static constexpr int vec_bytes = 16;
+  static constexpr int vec_elems = vec_bytes / elem_bytes;
+  static constexpr int thread_cnt = 64;
+  static_assert((tile_m * tile_k) % (vec_elems * thread_cnt) == 0);
+  static constexpr int a_inst_cnt_per_iter =
+      (tile_m * tile_k) / (vec_elems * thread_cnt);
+  static_assert(gemm_k % tile_k == 0);
+  static constexpr int k_iter_cnt = gemm_k / tile_k;
+
+  // Extra params to keep the order of k reduction...
+  static constexpr int mma_warp_cnt = 4;
+  static constexpr int per_mma_warp_k = tile_k / mma_warp_cnt;
+  static constexpr int k_each_chunk = gemm_k / mma_warp_cnt;
+
+ private:
+  __device__ int k_project(int tile_k_idx) {
+    return (tile_k_idx / per_mma_warp_k * k_each_chunk) +
+           (tile_k_idx % per_mma_warp_k);
+  }
+
+ public:
+  __device__ GmemLoaderA(bf16_t const* gmem_a_local_, bf16_t* smem_a_,
+                         uint64_t* smem_barrier_)
+      : gmem_a(gmem_a_local_),
+        smem_a(smem_a_),
+        smem_barrier(smem_barrier_),
+        local_tid(threadIdx.x % thread_cnt) {}
+
+  __device__ void prepare() {
+#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 900
+  // swizzle, that's what we want.
+  #pragma unroll
+    for (int i = 0; i < a_inst_cnt_per_iter; i++) {
+      int linear_idx = local_tid * vec_elems + i * thread_cnt * vec_elems;
+      int m_idx = linear_idx / tile_k;
+      int k_idx = linear_idx % tile_k;
+      k_idx = apply_swizzle_343_on_elem_row_col<bf16_t>(m_idx, k_idx);
+      a_smem_offsets[i] = m_idx * tile_k + k_idx;
+    }
+#endif
+  }
+
+  __device__ void issue_mainloop() {
+#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 900
+  #pragma unroll 1
+    for (int loop_idx = 0; loop_idx < k_iter_cnt; loop_idx++) {
+      if (need_wait) {
+        wait_barrier(smem_barrier + 1 + stage_idx * 2, phase_bit);
+      }
+      int next_stage_idx = stage_idx + 1;
+      int next_phase_bit =
+          next_stage_idx == stage_cnt ? phase_bit ^ 1 : phase_bit;
+      next_stage_idx = next_stage_idx == stage_cnt ? 0 : next_stage_idx;
+      if (loop_idx != k_iter_cnt - 1) {
+        need_wait = !try_wait_barrier(smem_barrier + 1 + next_stage_idx * 2,
+                                      next_phase_bit);
+      }
+
+  #pragma unroll
+      for (int i = 0; i < a_inst_cnt_per_iter; i++) {
+        int smem_offset = a_smem_offsets[i];
+        bf16_t* smem_ptr_this_iter =
+            smem_a + stage_idx * tile_m * tile_k + smem_offset;
+        int linear_idx = local_tid * vec_elems + i * thread_cnt * vec_elems;
+        int m_idx = linear_idx / tile_k;
+        int k_idx = linear_idx % tile_k;
+        int gmem_offset = m_idx * gemm_k + k_project(k_idx);
+        bf16_t const* gmem_ptr_this_iter = gmem_a + gmem_offset;
+        ldgsts_128(gmem_ptr_this_iter, smem_ptr_this_iter, true);
+      }
+      ldgsts_arrive(smem_barrier + stage_idx * 2);
+
+      stage_idx = next_stage_idx;
+      phase_bit = next_phase_bit;
+      gmem_a += per_mma_warp_k;
+    }
+#endif
+  }
+
+  bf16_t const* gmem_a;
+  bf16_t* smem_a;
+  uint64_t* smem_barrier;
+  int local_tid;
+  int stage_idx = 0;
+  int phase_bit = 1;
+  bool need_wait = true;
+
+  // per smem_stage, store with swizzle information
+  int a_smem_offsets[a_inst_cnt_per_iter];
+};
+
+template <int gemm_k, int tile_n, int tile_k, int stage_cnt>
+struct GmemLoaderB {
+  static constexpr int elem_bytes = 2;
+  static constexpr int vec_bytes = 16;
+  static constexpr int vec_elems = vec_bytes / elem_bytes;
+  static constexpr int thread_cnt = 64;
+  static_assert((tile_n * tile_k) % (vec_elems * thread_cnt) == 0);
+  static constexpr int b_inst_cnt_per_iter =
+      (tile_n * tile_k) / (vec_elems * thread_cnt);
+  static_assert(gemm_k % tile_k == 0);
+  static constexpr int k_iter_cnt = gemm_k / tile_k;
+
+  // Extra params to keep the order of k reduction...
+  static constexpr int mma_warp_cnt = 4;
+  static constexpr int per_mma_warp_k = tile_k / mma_warp_cnt;
+  static constexpr int k_each_chunk = gemm_k / mma_warp_cnt;
+
+ private:
+  __device__ int k_project(int tile_k_idx) {
+    return (tile_k_idx / per_mma_warp_k * k_each_chunk) +
+           (tile_k_idx % per_mma_warp_k);
+  }
+
+ public:
+  __device__ GmemLoaderB(bf16_t const* gmem_b_local_, bf16_t* smem_b_,
+                         uint64_t* smem_barrier_, int gemm_n_)
+      : gmem_b(gmem_b_local_),
+        smem_b(smem_b_),
+        smem_barrier(smem_barrier_),
+        gemm_n(gemm_n_),
+        local_tid(threadIdx.x % thread_cnt) {}
+
+  __device__ void prepare() {
+#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 900
+  // swizzle, that's what we want.
+  #pragma unroll
+    for (int i = 0; i < b_inst_cnt_per_iter; i++) {
+      int linear_idx = local_tid * vec_elems + i * thread_cnt * vec_elems;
+      int n_idx = linear_idx / tile_k;
+      int k_idx = linear_idx % tile_k;
+      k_idx = apply_swizzle_343_on_elem_row_col<bf16_t>(n_idx, k_idx);
+      b_smem_offsets[i] = n_idx * tile_k + k_idx;
+      preds[i] = n_idx < gemm_n;
+    }
+#endif
+  }
+
+  __device__ void issue_mainloop() {
+#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 900
+    asm volatile("griddepcontrol.wait;");
+  #pragma unroll 1
+    for (int loop_idx = 0; loop_idx < k_iter_cnt; loop_idx++) {
+      if (need_wait) {
+        wait_barrier(smem_barrier + 1 + stage_idx * 2, phase_bit);
+      }
+      int next_stage_idx = stage_idx + 1;
+      int next_phase_bit =
+          next_stage_idx == stage_cnt ? phase_bit ^ 1 : phase_bit;
+      next_stage_idx = next_stage_idx == stage_cnt ? 0 : next_stage_idx;
+      if (loop_idx != k_iter_cnt - 1) {
+        need_wait = !try_wait_barrier(smem_barrier + 1 + next_stage_idx * 2,
+                                      next_phase_bit);
+      }
+  #pragma unroll
+      for (int i = 0; i < b_inst_cnt_per_iter; i++) {
+        int smem_offset = b_smem_offsets[i];
+        bf16_t* smem_ptr_this_iter =
+            smem_b + stage_idx * tile_n * tile_k + smem_offset;
+        int linear_idx = local_tid * vec_elems + i * thread_cnt * vec_elems;
+        int n_idx = linear_idx / tile_k;
+        int k_idx = linear_idx % tile_k;
+        int gmem_offset = n_idx * gemm_k + k_project(k_idx);
+        bf16_t const* gmem_ptr_this_iter = gmem_b + gmem_offset;
+        ldgsts_128(gmem_ptr_this_iter, smem_ptr_this_iter, preds[i]);
+      }
+      ldgsts_arrive(smem_barrier + stage_idx * 2);
+
+      stage_idx = next_stage_idx;
+      phase_bit = next_phase_bit;
+      gmem_b += per_mma_warp_k;
+    }
+#endif
+  }
+
+  bf16_t const* gmem_b;
+  bf16_t* smem_b;
+  uint64_t* smem_barrier;
+  int gemm_n;
+  int local_tid;
+  int stage_idx = 0;
+  int phase_bit = 1;
+  bool need_wait = true;
+
+  // per smem_stage, store with swizzle information
+  int b_smem_offsets[b_inst_cnt_per_iter];
+  uint32_t preds[b_inst_cnt_per_iter];
+};
+
+template <int gemm_m, int gemm_k, int tile_m, int tile_n, int tile_k,
+          int stage_cnt>
+struct MmaComputer {
+  static constexpr int elem_bytes = 2;
+  static constexpr int thread_cnt = 128;
+  static_assert(gemm_k % tile_k == 0);
+  static_assert(tile_k % (thread_cnt / 32) == 0);
+  static constexpr int per_warp_tile_k = tile_k / (thread_cnt / 32);
+  static constexpr int k_iter_cnt = gemm_k / tile_k;
+  static constexpr int k_phase_cnt = per_warp_tile_k / 16;
+  static constexpr int m_iter_cnt = (tile_m + 15) / 16;
+  static constexpr int n_iter_cnt =
+      (tile_n + 7) /
+      8;  // Possible to have non-1 n_iter_cnt for ab_swap m16 case.
+  static_assert(m_iter_cnt == 1);
+  static_assert(n_iter_cnt == 1 || n_iter_cnt == 2);
+
+  __device__ MmaComputer(bf16_t* gmem_c_local_, bf16_t* smem_a_,
+                         bf16_t* smem_b_, uint64_t* smem_barrier_,
+                         int warp_idx_, int gemm_n_)
+      : gmem_c(gmem_c_local_),
+        smem_a(smem_a_),
+        smem_b(smem_b_),
+        smem_barrier(smem_barrier_),
+        warp_idx(warp_idx_ - (thread_cnt / 32)),
+        gemm_n(gemm_n_) {}
+
+ private:
+  __device__ constexpr int internal_b_atom_func(int tid) {
+    if constexpr (tile_n < 8) {
+      return (tid % tile_n) + ((tid % 8) / tile_n * 0) + tid / 8 * 8 * tile_n;
+    } else {
+      return (tid % 8) + ((tid % 32) / 8 * (tile_n * 8));
+    }
+  }
+
+ public:
+  __device__ void prepare() {
+#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 900
+  #pragma unroll
+    for (int i = 0; i < k_phase_cnt; i++) {
+      int linear_idx = (lane_idx % 16) + (lane_idx / 16) * 128 + i * 256;
+      int m_idx = linear_idx % tile_m;
+      int k_idx = linear_idx / tile_m + warp_k_offset_in_tile_k;
+      k_idx = apply_swizzle_343_on_elem_row_col<bf16_t>(m_idx, k_idx);
+      a_smem_offsets[0][i] = m_idx * tile_k + k_idx;
+    }
+  #pragma unroll
+    for (int n_iter_idx = 0; n_iter_idx < n_iter_cnt; n_iter_idx++) {
+  #pragma unroll
+      for (int i = 0; i < k_phase_cnt; i += 2) {  // Special i+=2 for B.
+        int linear_idx =
+            internal_b_atom_func(lane_idx) + i * tile_n * 16 + n_iter_idx * 8;
+        int n_idx = linear_idx % tile_n;
+        int k_idx = linear_idx / tile_n + warp_k_offset_in_tile_k;
+        k_idx = apply_swizzle_343_on_elem_row_col<bf16_t>(n_idx, k_idx);
+        b_smem_offsets[n_iter_idx][i] = n_idx * tile_k + k_idx;
+      }
+    }
+#endif
+  }
+
+  __device__ void issue_mainloop() {
+#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 900
+  #pragma unroll 1
+    for (int loop_idx = 0; loop_idx < k_iter_cnt; loop_idx++) {
+      wait_barrier(smem_barrier + 0 + stage_idx * 2, phase_bit);
+
+  #pragma unroll
+      for (int i = 0; i < k_phase_cnt; i++) {
+        int smem_offset = a_smem_offsets[0][i];
+        bf16_t* smem_ptr_this_iter =
+            smem_a + stage_idx * tile_m * tile_k + smem_offset;
+        ldsm_x4(smem_ptr_this_iter, reinterpret_cast<uint32_t*>(a_reg[0][i]));
+      }
+
+  #pragma unroll
+      for (int n_iter_idx = 0; n_iter_idx < n_iter_cnt; n_iter_idx++) {
+  #pragma unroll
+        for (int i = 0; i < k_phase_cnt; i += 2) {
+          int smem_offset = b_smem_offsets[n_iter_idx][i];
+          bf16_t* smem_ptr_this_iter =
+              smem_b + stage_idx * tile_n * tile_k + smem_offset;
+          ldsm_x4(smem_ptr_this_iter,
+                  reinterpret_cast<uint32_t*>(b_reg[n_iter_idx][i]));
+        }
+      }
+
+  #pragma unroll
+      for (int k_iter_idx = 0; k_iter_idx < k_phase_cnt; k_iter_idx++) {
+  #pragma unroll
+        for (int n_iter_idx = 0; n_iter_idx < n_iter_cnt; n_iter_idx++) {
+          hmma_16_8_16_f32acc_bf16ab(
+              acc_reg[0][n_iter_idx], a_reg[0][k_iter_idx],
+              b_reg[n_iter_idx][k_iter_idx], acc_reg[0][n_iter_idx]);
+        }
+      }
+      ::arrive_barrier(smem_barrier + 1 + stage_idx * 2);
+      stage_idx += 1;
+      phase_bit = stage_idx == stage_cnt ? phase_bit ^ 1 : phase_bit;
+      stage_idx = stage_idx == stage_cnt ? 0 : stage_idx;
+    }
+#endif
+  }
+
+  __device__ void epi() {
+#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 900
+    asm volatile("bar.sync %0, %1;" : : "r"(1), "r"(thread_cnt));
+    // reorganize the acc_reg
+    constexpr int thread_m = 2;
+    constexpr int thread_n = 2 * n_iter_cnt;
+    constexpr int cta_mma_n = n_iter_cnt * 8;
+    float acc_reg_reorg[thread_m][thread_n];
+
+    for (int i = 0; i < thread_m; i++) {
+      for (int j = 0; j < thread_n; j++) {
+        acc_reg_reorg[i][j] = acc_reg[0][j / 2][(j % 2) + (i * 2)];
+      }
+    }
+
+    // 4 x cosize(smem_c_layout)
+    float* smem_c = reinterpret_cast<float*>(smem_a);
+    // coord -> index
+    auto smem_c_index_func = [&](int m_idx, int n_idx) {
+      int group_rows = 32 / cta_mma_n;
+      int group_cnt = 2;
+      return (m_idx % group_rows * cta_mma_n) +
+             (m_idx / group_rows * (32 + group_cnt)) + n_idx;
+    };
+    constexpr int cosize_smem_c = ((tile_m * cta_mma_n) / 32) * (32 + 2);
+
+  // This should be optimized to STS.64 but can not be STS.128 due to the bank
+  // index.
+  #pragma unroll
+    for (int m_idx_thread = 0; m_idx_thread < thread_m; m_idx_thread++) {
+  #pragma unroll
+      for (int n_idx_thread = 0; n_idx_thread < thread_n; n_idx_thread++) {
+        int m_idx = (lane_idx / 4) + m_idx_thread * 8;
+        int n_idx =
+            ((lane_idx % 4) * 2) + (n_idx_thread % 2) + (n_idx_thread / 2) * 8;
+        smem_c[cosize_smem_c * warp_idx + smem_c_index_func(m_idx, n_idx)] =
+            acc_reg_reorg[m_idx_thread][n_idx_thread];
+      }
+    }
+    asm volatile("bar.sync %0, %1;" : : "r"(1), "r"(thread_cnt));
+
+    if (warp_idx == 0) {
+      constexpr int final_acc_reg_cnt = (tile_m * tile_n + 31) / 32;
+      float acc_final[final_acc_reg_cnt]{};
+
+  #pragma unroll
+      for (int reg_idx = 0; reg_idx < final_acc_reg_cnt; reg_idx++) {
+        int linear_idx = reg_idx * 32 + lane_idx;
+        int m_idx = linear_idx % tile_m;
+        int n_idx = linear_idx / tile_m;
+        acc_final[reg_idx] +=
+            smem_c[smem_c_index_func(m_idx, n_idx) + 0 * cosize_smem_c] +
+            smem_c[smem_c_index_func(m_idx, n_idx) + 1 * cosize_smem_c] +
+            smem_c[smem_c_index_func(m_idx, n_idx) + 2 * cosize_smem_c] +
+            smem_c[smem_c_index_func(m_idx, n_idx) + 3 * cosize_smem_c];
+      }
+
+  #pragma unroll
+      for (int reg_idx = 0; reg_idx < final_acc_reg_cnt; reg_idx++) {
+        int linear_idx = reg_idx * 32 + lane_idx;
+        int m_idx = linear_idx % tile_m;
+        int n_idx = linear_idx / tile_m;
+        if (m_idx < tile_m && n_idx < gemm_n) {
+          gmem_c[n_idx * gemm_m + m_idx] = acc_final[reg_idx];
+        }
+      }
+    }
+#endif
+  }
+
+  bf16_t* gmem_c;
+  bf16_t* smem_a;
+  bf16_t* smem_b;
+  uint64_t* smem_barrier;
+  int warp_idx;
+  int gemm_n;
+  int stage_idx = 0;
+  int phase_bit = 0;
+  int lane_idx = threadIdx.x % 32;
+  int warp_k_offset_in_tile_k = warp_idx * per_warp_tile_k;
+
+  int a_smem_offsets[m_iter_cnt][k_phase_cnt];
+  int b_smem_offsets[n_iter_cnt][k_phase_cnt];
+
+  bf16_t a_reg[m_iter_cnt][k_phase_cnt][8];
+  bf16_t b_reg[n_iter_cnt][k_phase_cnt][4];
+  float acc_reg[m_iter_cnt][n_iter_cnt][4]{};
+};
+
+// AB swapped, kernel is k-major, k-major, m-major
+template <int batch_size, int gemm_m, int gemm_k, int tile_m, int tile_n,
+          int tile_k, int stage_cnt>
+__global__ __launch_bounds__(256, 1) void fused_a_gemm_kernel(
+    bf16_t* output, bf16_t const* mat_a, bf16_t const* mat_b, int gemm_n) {
+#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 900
+  constexpr int load_thread_cnt = 128;
+  constexpr int compute_thread_cnt = 128;
+  constexpr int thread_cnt = load_thread_cnt + compute_thread_cnt;
+  (void)thread_cnt;
+  static_assert(gemm_m % 16 == 0);
+  static_assert(gemm_k % tile_k == 0);
+  static_assert(gemm_m % tile_m == 0);
+  static_assert(
+      tile_k == 128 || tile_k == 256 || tile_k == 512 ||
+      tile_k == 1024);  // tile_k must be larger than 64 since 4 warp splitK.
+  static_assert(tile_m == 16);
+  constexpr int g2s_vec_bytes = 16;
+  constexpr int a_elem_bytes = 2;
+  constexpr int b_elem_bytes = 2;
+  static_assert((tile_m * a_elem_bytes + tile_n * b_elem_bytes) * tile_k *
+                    stage_cnt <=
+                225 * 1024);
+  static_assert((tile_m * tile_k * a_elem_bytes) %
+                    (load_thread_cnt * g2s_vec_bytes) ==
+                0);
+  static_assert((tile_n * tile_k * b_elem_bytes) %
+                    (load_thread_cnt * g2s_vec_bytes) ==
+                0);
+
+  extern __shared__ char smem[];
+  uint64_t* smem_barrier = reinterpret_cast<uint64_t*>(
+      smem);  // producer,consumer; producer,consumer; ...
+  bf16_t* smem_a = reinterpret_cast<bf16_t*>(smem + (stage_cnt * 8 * 2 + 1024) /
+                                                        1024 * 1024);
+  bf16_t* smem_b = smem_a + tile_m * tile_k * stage_cnt;
+
+  int cta_m_idx = tile_m * blockIdx.x;
+  int cta_n_idx = tile_n * blockIdx.y;
+  bf16_t const* gmem_a_local = mat_a + cta_m_idx * gemm_k;
+  bf16_t const* gmem_b_local = mat_b + cta_n_idx * gemm_k;
+  bf16_t* gmem_c_local = output + cta_n_idx * gemm_m + cta_m_idx;
+
+  int warp_idx = __shfl_sync(0xffffffff, threadIdx.x / 32, 0);
+
+  if (warp_idx == 4) {
+    for (int i = 0; i < stage_cnt; i++) {
+      initialize_barrier(smem_barrier + i * 2 + 0,
+                         load_thread_cnt);  // producer
+      initialize_barrier(smem_barrier + i * 2 + 1,
+                         compute_thread_cnt);  // consumer
+    }
+  }
+  __syncthreads();
+
+  if (warp_idx < 2) {
+    GmemLoaderA<gemm_k, tile_m, tile_k, stage_cnt> a_loader(
+        gmem_a_local, smem_a, smem_barrier);
+    a_loader.prepare();
+    a_loader.issue_mainloop();
+  } else if (warp_idx < 4) {
+    GmemLoaderB<gemm_k, tile_n, tile_k, stage_cnt> b_loader(
+        gmem_b_local, smem_b, smem_barrier, gemm_n);
+    b_loader.prepare();
+    b_loader.issue_mainloop();
+  } else {
+    MmaComputer<gemm_m, gemm_k, tile_m, tile_n, tile_k, stage_cnt> mma_computer(
+        gmem_c_local, smem_a, smem_b, smem_barrier, warp_idx, gemm_n);
+    mma_computer.prepare();
+    mma_computer.issue_mainloop();
+    mma_computer.epi();
+  }
+  asm volatile("griddepcontrol.launch_dependents;");
+#endif
+}
+
+template <typename T, int kHdIn, int kHdOut, int kTileN>
+void invokeFusedAGemm(T* output, T const* mat_a, T const* mat_b, int num_tokens,
+                      cudaStream_t const stream) {
+  constexpr int gemm_m = kHdOut;  // 2112
+  int const gemm_n = num_tokens;  // 1-16
+  constexpr int gemm_k = kHdIn;   // 7168
+  constexpr int batch_size = 1;
+  std::swap(mat_a, mat_b);
+  constexpr int tile_m = 16;
+  constexpr int tile_n = kTileN;                        // 8 or 16
+  constexpr int tile_k = std::max(256, 1024 / tile_n);  // 256
+  constexpr int max_stage_cnt =
+      1024 * 192 / ((tile_m + tile_n) * tile_k * sizeof(bf16_t));
+  constexpr int k_iter_cnt = gemm_k / tile_k;
+  constexpr int stage_cnt =
+      k_iter_cnt > max_stage_cnt ? max_stage_cnt : k_iter_cnt;
+  int cta_m_cnt = gemm_m / tile_m;
+  int cta_n_cnt = (gemm_n + tile_n - 1) / tile_n;
+  constexpr int barrier_bytes = (stage_cnt * 16 + 1023) / 1024 * 1024;
+  constexpr int smem_bytes =
+      ((tile_m * 2 + tile_n * 2) * tile_k * stage_cnt + barrier_bytes + 1023) /
+      1024 * 1024;
+
+  dim3 grid(cta_m_cnt, cta_n_cnt, 1);
+  dim3 block_size(256);
+  cudaLaunchConfig_t config;
+  config.gridDim = grid;
+  config.blockDim = block_size;
+  config.dynamicSmemBytes = smem_bytes;
+  config.stream = stream;
+  cudaLaunchAttribute attrs[1];
+  attrs[0].id = cudaLaunchAttributeProgrammaticStreamSerialization;
+  attrs[0].val.programmaticStreamSerializationAllowed = getEnvEnablePDL();
+  config.numAttrs = 1;
+  config.attrs = attrs;
+  if (smem_bytes >= (48 * 1024)) {
+    cudaFuncSetAttribute(fused_a_gemm_kernel<batch_size, gemm_m, gemm_k, tile_m,
+                                             tile_n, tile_k, stage_cnt>,
+                         cudaFuncAttributeMaxDynamicSharedMemorySize,
+                         smem_bytes);
+  }
+  cudaLaunchKernelEx(&config,
+                     fused_a_gemm_kernel<batch_size, gemm_m, gemm_k, tile_m,
+                                         tile_n, tile_k, stage_cnt>,
+                     output, mat_a, mat_b, gemm_n);
+}
+
+template void invokeFusedAGemm<__nv_bfloat16, 7168, 2112, 8>(
+    __nv_bfloat16*, __nv_bfloat16 const*, __nv_bfloat16 const*, int num_tokens,
+    cudaStream_t);
+
+template void invokeFusedAGemm<__nv_bfloat16, 7168, 2112, 16>(
+    __nv_bfloat16*, __nv_bfloat16 const*, __nv_bfloat16 const*, int num_tokens,
+    cudaStream_t);
+
+void dsv3_fused_a_gemm(torch::Tensor& output, torch::Tensor const& mat_a,
+                       torch::Tensor const& mat_b) {
+  TORCH_CHECK(mat_a.dim() == 2 && mat_b.dim() == 2 && output.dim() == 2);
+  int const num_tokens = mat_a.size(0);
+  int const hd_in = mat_a.size(1);
+  int const hd_out = mat_b.size(1);
+
+  constexpr int kHdIn = 7168;
+  constexpr int kHdOut = 2112;
+  TORCH_CHECK(num_tokens >= 1 && num_tokens <= 16,
+              "required 1 <= mat_a.shape[0] <= 16")
+  TORCH_CHECK(hd_in == kHdIn, "required mat_a.shape[1] == 7168")
+  TORCH_CHECK(hd_out == kHdOut, "required mat_b.shape[1] == 2112")
+  TORCH_CHECK(output.size(0) == num_tokens,
+              "required output.shape[0] == mat_a.shape[0]")
+  TORCH_CHECK(output.size(1) == hd_out,
+              "required output.shape[1] == mat_b.shape[1]")
+
+  TORCH_CHECK(mat_a.stride(1) == 1, "mat_a must be a row major tensor");
+  TORCH_CHECK(output.stride(1) == 1, "output must be a row major tensor");
+  TORCH_CHECK(mat_b.stride(0) == 1, "mat_b must be a column major tensor");
+
+  TORCH_CHECK(mat_a.scalar_type() == torch::kBFloat16 &&
+                  mat_b.scalar_type() == torch::kBFloat16,
+              "Only BFloat16 input dtype is supported")
+  TORCH_CHECK(output.scalar_type() == torch::kBFloat16,
+              "Only BFloat16 output dtype is supported")
+
+  TORCH_CHECK(getSMVersion() >= 90, "required CUDA ARCH >= SM_90");
+
+  auto stream = at::cuda::getCurrentCUDAStream(mat_a.get_device());
+  if (num_tokens <= 8) {
+    invokeFusedAGemm<__nv_bfloat16, kHdIn, kHdOut, 8>(
+        reinterpret_cast<__nv_bfloat16*>(output.mutable_data_ptr()),
+        reinterpret_cast<__nv_bfloat16 const*>(mat_a.data_ptr()),
+        reinterpret_cast<__nv_bfloat16 const*>(mat_b.data_ptr()), num_tokens,
+        stream);
+  } else {
+    invokeFusedAGemm<__nv_bfloat16, kHdIn, kHdOut, 16>(
+        reinterpret_cast<__nv_bfloat16*>(output.mutable_data_ptr()),
+        reinterpret_cast<__nv_bfloat16 const*>(mat_a.data_ptr()),
+        reinterpret_cast<__nv_bfloat16 const*>(mat_b.data_ptr()), num_tokens,
+        stream);
+  }
+}
+
+TORCH_LIBRARY_IMPL_EXPAND(TORCH_EXTENSION_NAME, CUDA, m) {
+  m.impl("dsv3_fused_a_gemm", &dsv3_fused_a_gemm);
+}

csrc/moe/dsv3_router_gemm_bf16_out.cu

@@ -0,0 +1,291 @@
+/*
+ * Adapted from SGLang's sgl-kernel implementation, which was adapted from
+ * https://github.com/NVIDIA/TensorRT-LLM/blob/main/cpp/tensorrt_llm/kernels/dsv3MinLatencyKernels/dsv3RouterGemm.cu
+ * https://github.com/NVIDIA/TensorRT-LLM/blob/main/cpp/tensorrt_llm/thop/dsv3RouterGemmOp.cpp
+ *
+ * Copyright (c) 2019-2023, NVIDIA CORPORATION.  All rights reserved.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include <ATen/ATen.h>
+#include <ATen/cuda/CUDAContext.h>
+
+#include <cuda_bf16.h>
+#include <cuda_runtime.h>
+
+#include "dsv3_router_gemm_utils.h"
+
+// Custom FMA implementation using PTX assembly instructions
+__device__ __forceinline__ void fma(float2& d, float2 const& a, float2 const& b,
+                                    float2 const& c) {
+  asm volatile("fma.rn.f32x2 %0, %1, %2, %3;\n"
+               : "=l"(reinterpret_cast<uint64_t&>(d))
+               : "l"(reinterpret_cast<uint64_t const&>(a)),
+                 "l"(reinterpret_cast<uint64_t const&>(b)),
+                 "l"(reinterpret_cast<uint64_t const&>(c)));
+}
+
+// Convert 8 bfloat16 values from a uint4 to float array - optimized conversion
+template <int VPT>
+__device__ __forceinline__ void bf16_uint4_to_float8(uint4 const& vec,
+                                                     float* dst) {
+  __nv_bfloat16* bf16_ptr =
+      reinterpret_cast<__nv_bfloat16*>(const_cast<uint4*>(&vec));
+
+#pragma unroll
+  for (int i = 0; i < VPT; i++) {
+    dst[i] = __bfloat162float(bf16_ptr[i]);
+  }
+}
+
+template <typename T, int kBlockSize, int VPT, int kNumTokens, int kNumExperts,
+          int kHiddenDim>
+__global__ __launch_bounds__(128, 1) void router_gemm_kernel_bf16_output(
+    __nv_bfloat16* out, T const* mat_a, T const* mat_b) {
+  // Each block handles one expert column
+  int const n_idx = blockIdx.x;
+  int const tid = threadIdx.x;
+  constexpr int kWarpSize = 32;
+  constexpr int kNumWarps = kBlockSize / kWarpSize;
+  // Constants for this kernel
+  constexpr int k_elems_per_k_iteration = VPT * kBlockSize;
+  constexpr int k_iterations =
+      kHiddenDim / k_elems_per_k_iteration;  // Total K iterations
+
+  // Initialize accumulators for all M rows
+  float acc[kNumTokens] = {};
+
+  // Shared memory for warp-level reduction
+  __shared__ float sm_reduction[kNumTokens][kNumWarps];  // kNumWarps
+
+  // B matrix is in column-major order, so we can directly load a column for the
+  // n_idx expert
+  T const* b_col = mat_b + n_idx * kHiddenDim;
+
+  // Pre-compute k_base values for each iteration to help compiler optimize
+  int k_bases[k_iterations];
+#pragma unroll
+  for (int ki = 0; ki < k_iterations; ki++) {
+    k_bases[ki] = ki * k_elems_per_k_iteration + tid * VPT;
+  }
+
+#if (defined(__CUDA_ARCH__) && (__CUDA_ARCH__ >= 900))
+  asm volatile("griddepcontrol.wait;");
+#endif
+
+  // Process the GEMM in chunks
+  for (int ki = 0; ki < k_iterations; ki++) {
+    int const k_base = k_bases[ki];
+
+    // Load B matrix values using vector load (8 bf16 values)
+    uint4 b_vec = *reinterpret_cast<uint4 const*>(b_col + k_base);
+
+    // Convert B values to float
+    float b_float[VPT];
+    bf16_uint4_to_float8<VPT>(b_vec, b_float);
+
+// Process each token
+#pragma unroll
+    for (int m_idx = 0; m_idx < kNumTokens; m_idx++) {
+      // Load both rows of A matrix using vector loads
+      uint4 a_vec = *reinterpret_cast<uint4 const*>(
+          mat_a + (m_idx * kHiddenDim) + k_base);
+
+      // Convert A values to float
+      float a_float[VPT];
+      bf16_uint4_to_float8<VPT>(a_vec, a_float);
+
+// Process elements in this chunk
+#pragma unroll
+      for (int k = 0; k < VPT; k++) {
+        float a = a_float[k];
+        float b = b_float[k];
+        acc[m_idx] += a * b;
+      }
+    }
+  }
+
+  // Perform warp-level reduction
+  int const warpSize = 32;
+  int const warpId = tid / warpSize;
+  int const laneId = tid % warpSize;
+
+  // Register for warp-level reduction results
+  float warp_result[kNumTokens];
+
+#pragma unroll
+  for (int m_idx = 0; m_idx < kNumTokens; m_idx++) {
+    warp_result[m_idx] = acc[m_idx];
+  }
+
+// Perform warp-level reduction using optimized butterfly pattern
+#pragma unroll
+  for (int m = 0; m < kNumTokens; m++) {
+    float sum = warp_result[m];
+
+    // Butterfly reduction pattern
+    sum += __shfl_xor_sync(0xffffffff, sum, 16);
+    sum += __shfl_xor_sync(0xffffffff, sum, 8);
+    sum += __shfl_xor_sync(0xffffffff, sum, 4);
+    sum += __shfl_xor_sync(0xffffffff, sum, 2);
+    sum += __shfl_xor_sync(0xffffffff, sum, 1);
+
+    // Only the first thread in each warp stores to shared memory
+    if (laneId == 0) {
+      sm_reduction[m][warpId] = sum;
+    }
+  }
+
+  __syncthreads();
+
+  // Final reduction across warps (only first thread)
+  if (tid == 0) {
+#pragma unroll
+    for (int m = 0; m < kNumTokens; m++) {
+      float final_sum = 0.0f;
+
+// Sum across the kNumWarps
+#pragma unroll
+      for (int w = 0; w < kNumWarps; w++) {
+        final_sum += sm_reduction[m][w];
+      }
+
+      // Write final result
+      out[m * kNumExperts + n_idx] = __float2bfloat16(final_sum);
+    }
+  }
+#if (defined(__CUDA_ARCH__) && (__CUDA_ARCH__ >= 900))
+  asm volatile("griddepcontrol.launch_dependents;");
+#endif
+}
+
+template <typename T, int kNumTokens, int kNumExperts, int kHiddenDim>
+void invokeRouterGemmBf16Output(__nv_bfloat16* output, T const* mat_a,
+                                T const* mat_b, cudaStream_t stream) {
+  constexpr int VPT = 16 / sizeof(T);
+  constexpr int kBlockSize = 128;
+  cudaLaunchConfig_t config;
+  config.gridDim = kNumExperts;
+  config.blockDim = kBlockSize;
+  config.dynamicSmemBytes = 0;
+  config.stream = stream;
+  cudaLaunchAttribute attrs[1];
+  attrs[0].id = cudaLaunchAttributeProgrammaticStreamSerialization;
+  attrs[0].val.programmaticStreamSerializationAllowed = getEnvEnablePDL();
+  config.numAttrs = 1;
+  config.attrs = attrs;
+  cudaLaunchKernelEx(
+      &config,
+      router_gemm_kernel_bf16_output<T, kBlockSize, VPT, kNumTokens,
+                                     kNumExperts, kHiddenDim>,
+      output, mat_a, mat_b);
+}
+
+// Template instantiations for DEFAULT_NUM_EXPERTS experts
+template void invokeRouterGemmBf16Output<__nv_bfloat16, 1, 256, 7168>(
+    __nv_bfloat16*, __nv_bfloat16 const*, __nv_bfloat16 const*, cudaStream_t);
+
+template void invokeRouterGemmBf16Output<__nv_bfloat16, 2, 256, 7168>(
+    __nv_bfloat16*, __nv_bfloat16 const*, __nv_bfloat16 const*, cudaStream_t);
+
+template void invokeRouterGemmBf16Output<__nv_bfloat16, 3, 256, 7168>(
+    __nv_bfloat16*, __nv_bfloat16 const*, __nv_bfloat16 const*, cudaStream_t);
+
+template void invokeRouterGemmBf16Output<__nv_bfloat16, 4, 256, 7168>(
+    __nv_bfloat16*, __nv_bfloat16 const*, __nv_bfloat16 const*, cudaStream_t);
+
+template void invokeRouterGemmBf16Output<__nv_bfloat16, 5, 256, 7168>(
+    __nv_bfloat16*, __nv_bfloat16 const*, __nv_bfloat16 const*, cudaStream_t);
+
+template void invokeRouterGemmBf16Output<__nv_bfloat16, 6, 256, 7168>(
+    __nv_bfloat16*, __nv_bfloat16 const*, __nv_bfloat16 const*, cudaStream_t);
+
+template void invokeRouterGemmBf16Output<__nv_bfloat16, 7, 256, 7168>(
+    __nv_bfloat16*, __nv_bfloat16 const*, __nv_bfloat16 const*, cudaStream_t);
+
+template void invokeRouterGemmBf16Output<__nv_bfloat16, 8, 256, 7168>(
+    __nv_bfloat16*, __nv_bfloat16 const*, __nv_bfloat16 const*, cudaStream_t);
+
+template void invokeRouterGemmBf16Output<__nv_bfloat16, 9, 256, 7168>(
+    __nv_bfloat16*, __nv_bfloat16 const*, __nv_bfloat16 const*, cudaStream_t);
+
+template void invokeRouterGemmBf16Output<__nv_bfloat16, 10, 256, 7168>(
+    __nv_bfloat16*, __nv_bfloat16 const*, __nv_bfloat16 const*, cudaStream_t);
+
+template void invokeRouterGemmBf16Output<__nv_bfloat16, 11, 256, 7168>(
+    __nv_bfloat16*, __nv_bfloat16 const*, __nv_bfloat16 const*, cudaStream_t);
+
+template void invokeRouterGemmBf16Output<__nv_bfloat16, 12, 256, 7168>(
+    __nv_bfloat16*, __nv_bfloat16 const*, __nv_bfloat16 const*, cudaStream_t);
+
+template void invokeRouterGemmBf16Output<__nv_bfloat16, 13, 256, 7168>(
+    __nv_bfloat16*, __nv_bfloat16 const*, __nv_bfloat16 const*, cudaStream_t);
+
+template void invokeRouterGemmBf16Output<__nv_bfloat16, 14, 256, 7168>(
+    __nv_bfloat16*, __nv_bfloat16 const*, __nv_bfloat16 const*, cudaStream_t);
+
+template void invokeRouterGemmBf16Output<__nv_bfloat16, 15, 256, 7168>(
+    __nv_bfloat16*, __nv_bfloat16 const*, __nv_bfloat16 const*, cudaStream_t);
+
+template void invokeRouterGemmBf16Output<__nv_bfloat16, 16, 256, 7168>(
+    __nv_bfloat16*, __nv_bfloat16 const*, __nv_bfloat16 const*, cudaStream_t);
+
+// Template instantiations for KIMI_K2_NUM_EXPERTS experts
+template void invokeRouterGemmBf16Output<__nv_bfloat16, 1, 384, 7168>(
+    __nv_bfloat16*, __nv_bfloat16 const*, __nv_bfloat16 const*, cudaStream_t);
+
+template void invokeRouterGemmBf16Output<__nv_bfloat16, 2, 384, 7168>(
+    __nv_bfloat16*, __nv_bfloat16 const*, __nv_bfloat16 const*, cudaStream_t);
+
+template void invokeRouterGemmBf16Output<__nv_bfloat16, 3, 384, 7168>(
+    __nv_bfloat16*, __nv_bfloat16 const*, __nv_bfloat16 const*, cudaStream_t);
+
+template void invokeRouterGemmBf16Output<__nv_bfloat16, 4, 384, 7168>(
+    __nv_bfloat16*, __nv_bfloat16 const*, __nv_bfloat16 const*, cudaStream_t);
+
+template void invokeRouterGemmBf16Output<__nv_bfloat16, 5, 384, 7168>(
+    __nv_bfloat16*, __nv_bfloat16 const*, __nv_bfloat16 const*, cudaStream_t);
+
+template void invokeRouterGemmBf16Output<__nv_bfloat16, 6, 384, 7168>(
+    __nv_bfloat16*, __nv_bfloat16 const*, __nv_bfloat16 const*, cudaStream_t);
+
+template void invokeRouterGemmBf16Output<__nv_bfloat16, 7, 384, 7168>(
+    __nv_bfloat16*, __nv_bfloat16 const*, __nv_bfloat16 const*, cudaStream_t);
+
+template void invokeRouterGemmBf16Output<__nv_bfloat16, 8, 384, 7168>(
+    __nv_bfloat16*, __nv_bfloat16 const*, __nv_bfloat16 const*, cudaStream_t);
+
+template void invokeRouterGemmBf16Output<__nv_bfloat16, 9, 384, 7168>(
+    __nv_bfloat16*, __nv_bfloat16 const*, __nv_bfloat16 const*, cudaStream_t);
+
+template void invokeRouterGemmBf16Output<__nv_bfloat16, 10, 384, 7168>(
+    __nv_bfloat16*, __nv_bfloat16 const*, __nv_bfloat16 const*, cudaStream_t);
+
+template void invokeRouterGemmBf16Output<__nv_bfloat16, 11, 384, 7168>(
+    __nv_bfloat16*, __nv_bfloat16 const*, __nv_bfloat16 const*, cudaStream_t);
+
+template void invokeRouterGemmBf16Output<__nv_bfloat16, 12, 384, 7168>(
+    __nv_bfloat16*, __nv_bfloat16 const*, __nv_bfloat16 const*, cudaStream_t);
+
+template void invokeRouterGemmBf16Output<__nv_bfloat16, 13, 384, 7168>(
+    __nv_bfloat16*, __nv_bfloat16 const*, __nv_bfloat16 const*, cudaStream_t);
+
+template void invokeRouterGemmBf16Output<__nv_bfloat16, 14, 384, 7168>(
+    __nv_bfloat16*, __nv_bfloat16 const*, __nv_bfloat16 const*, cudaStream_t);
+
+template void invokeRouterGemmBf16Output<__nv_bfloat16, 15, 384, 7168>(
+    __nv_bfloat16*, __nv_bfloat16 const*, __nv_bfloat16 const*, cudaStream_t);
+
+template void invokeRouterGemmBf16Output<__nv_bfloat16, 16, 384, 7168>(
+    __nv_bfloat16*, __nv_bfloat16 const*, __nv_bfloat16 const*, cudaStream_t);

csrc/moe/dsv3_router_gemm_entry.cu

@@ -0,0 +1,169 @@
+/*
+ * Adapted from SGLang's sgl-kernel implementation, which was adapted from
+ * https://github.com/NVIDIA/TensorRT-LLM/blob/main/cpp/tensorrt_llm/kernels/dsv3MinLatencyKernels/dsv3RouterGemm.cu
+ * https://github.com/NVIDIA/TensorRT-LLM/blob/main/cpp/tensorrt_llm/thop/dsv3RouterGemmOp.cpp
+ *
+ * Copyright (c) 2019-2023, NVIDIA CORPORATION.  All rights reserved.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include <ATen/ATen.h>
+#include <ATen/cuda/CUDAContext.h>
+#include <torch/all.h>
+
+#include <cuda_bf16.h>
+#include <cuda_runtime.h>
+
+#include "core/registration.h"
+#include "dsv3_router_gemm_utils.h"
+
+static constexpr int DEFAULT_NUM_EXPERTS = 256;
+static constexpr int KIMI_K2_NUM_EXPERTS = 384;
+static constexpr int DEFAULT_HIDDEN_DIM = 7168;
+
+template <typename T, int kNumTokens, int kNumExperts, int kHiddenDim>
+void invokeRouterGemmFloatOutput(float* output, T const* mat_a, T const* mat_b,
+                                 cudaStream_t stream);
+
+template <typename T, int kNumTokens, int kNumExperts, int kHiddenDim>
+void invokeRouterGemmBf16Output(__nv_bfloat16* output, T const* mat_a,
+                                T const* mat_b, cudaStream_t stream);
+
+template <int kBegin, int kEnd, int kNumExperts, int kHiddenDim>
+struct LoopUnroller {
+  static void unroll_float_output(int num_tokens, float* output,
+                                  __nv_bfloat16 const* input,
+                                  __nv_bfloat16 const* weights,
+                                  cudaStream_t stream) {
+    if (num_tokens == kBegin) {
+      invokeRouterGemmFloatOutput<__nv_bfloat16, kBegin, kNumExperts,
+                                  kHiddenDim>(output, input, weights, stream);
+    } else {
+      LoopUnroller<kBegin + 1, kEnd, kNumExperts,
+                   kHiddenDim>::unroll_float_output(num_tokens, output, input,
+                                                    weights, stream);
+    }
+  }
+
+  static void unroll_bf16_output(int num_tokens, __nv_bfloat16* output,
+                                 __nv_bfloat16 const* input,
+                                 __nv_bfloat16 const* weights,
+                                 cudaStream_t stream) {
+    if (num_tokens == kBegin) {
+      invokeRouterGemmBf16Output<__nv_bfloat16, kBegin, kNumExperts,
+                                 kHiddenDim>(output, input, weights, stream);
+    } else {
+      LoopUnroller<kBegin + 1, kEnd, kNumExperts,
+                   kHiddenDim>::unroll_bf16_output(num_tokens, output, input,
+                                                   weights, stream);
+    }
+  }
+};
+
+template <int kEnd, int kNumExperts, int kHiddenDim>
+struct LoopUnroller<kEnd, kEnd, kNumExperts, kHiddenDim> {
+  static void unroll_float_output(int num_tokens, float* output,
+                                  __nv_bfloat16 const* input,
+                                  __nv_bfloat16 const* weights,
+                                  cudaStream_t stream) {
+    if (num_tokens == kEnd) {
+      invokeRouterGemmFloatOutput<__nv_bfloat16, kEnd, kNumExperts, kHiddenDim>(
+          output, input, weights, stream);
+    } else {
+      throw std::invalid_argument("Invalid num_tokens, only supports 1 to 16");
+    }
+  }
+
+  static void unroll_bf16_output(int num_tokens, __nv_bfloat16* output,
+                                 __nv_bfloat16 const* input,
+                                 __nv_bfloat16 const* weights,
+                                 cudaStream_t stream) {
+    if (num_tokens == kEnd) {
+      invokeRouterGemmBf16Output<__nv_bfloat16, kEnd, kNumExperts, kHiddenDim>(
+          output, input, weights, stream);
+    } else {
+      throw std::invalid_argument("Invalid num_tokens, only supports 1 to 16");
+    }
+  }
+};
+
+void dsv3_router_gemm(at::Tensor& output,       // [num_tokens, num_experts]
+                      const at::Tensor& mat_a,  // [num_tokens, hidden_dim]
+                      const at::Tensor& mat_b   // [num_experts, hidden_dim]
+) {
+  TORCH_CHECK(output.dim() == 2 && mat_a.dim() == 2 && mat_b.dim() == 2);
+
+  const int num_tokens = mat_a.size(0);
+  const int num_experts = mat_b.size(0);
+  const int hidden_dim = mat_a.size(1);
+
+  TORCH_CHECK(mat_a.size(1) == mat_b.size(1),
+              "mat_a and mat_b must have the same hidden_dim");
+  TORCH_CHECK(hidden_dim == DEFAULT_HIDDEN_DIM,
+              "Expected hidden_dim=", DEFAULT_HIDDEN_DIM,
+              ", but got hidden_dim=", hidden_dim);
+  TORCH_CHECK(
+      num_experts == DEFAULT_NUM_EXPERTS || num_experts == KIMI_K2_NUM_EXPERTS,
+      "Expected num_experts=", DEFAULT_NUM_EXPERTS,
+      " or num_experts=", KIMI_K2_NUM_EXPERTS,
+      ", but got num_experts=", num_experts);
+  TORCH_CHECK(num_tokens >= 1 && num_tokens <= 16,
+              "currently num_tokens must be less than or equal to 16 for "
+              "router_gemm");
+  TORCH_CHECK(mat_a.dtype() == at::kBFloat16, "mat_a must be bf16");
+  TORCH_CHECK(mat_b.dtype() == at::kBFloat16, "mat_b must be bf16");
+  TORCH_CHECK(output.dtype() == at::kFloat || output.dtype() == at::kBFloat16,
+              "output must be float32 or bf16");
+
+  auto const sm = getSMVersion();
+  TORCH_CHECK(sm >= 90 && sm <= 103, "required SM_103 >= CUDA ARCH >= SM_90");
+
+  const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
+
+  if (output.dtype() == at::kFloat) {
+    if (num_experts == DEFAULT_NUM_EXPERTS) {
+      LoopUnroller<1, 16, DEFAULT_NUM_EXPERTS, DEFAULT_HIDDEN_DIM>::
+          unroll_float_output(
+              num_tokens, reinterpret_cast<float*>(output.mutable_data_ptr()),
+              reinterpret_cast<__nv_bfloat16 const*>(mat_a.data_ptr()),
+              reinterpret_cast<__nv_bfloat16 const*>(mat_b.data_ptr()), stream);
+    } else if (num_experts == KIMI_K2_NUM_EXPERTS) {
+      LoopUnroller<1, 16, KIMI_K2_NUM_EXPERTS, DEFAULT_HIDDEN_DIM>::
+          unroll_float_output(
+              num_tokens, reinterpret_cast<float*>(output.mutable_data_ptr()),
+              reinterpret_cast<__nv_bfloat16 const*>(mat_a.data_ptr()),
+              reinterpret_cast<__nv_bfloat16 const*>(mat_b.data_ptr()), stream);
+    }
+  } else if (output.dtype() == at::kBFloat16) {
+    if (num_experts == DEFAULT_NUM_EXPERTS) {
+      LoopUnroller<1, 16, DEFAULT_NUM_EXPERTS, DEFAULT_HIDDEN_DIM>::
+          unroll_bf16_output(
+              num_tokens,
+              reinterpret_cast<__nv_bfloat16*>(output.mutable_data_ptr()),
+              reinterpret_cast<__nv_bfloat16 const*>(mat_a.data_ptr()),
+              reinterpret_cast<__nv_bfloat16 const*>(mat_b.data_ptr()), stream);
+    } else if (num_experts == KIMI_K2_NUM_EXPERTS) {
+      LoopUnroller<1, 16, KIMI_K2_NUM_EXPERTS, DEFAULT_HIDDEN_DIM>::
+          unroll_bf16_output(
+              num_tokens,
+              reinterpret_cast<__nv_bfloat16*>(output.mutable_data_ptr()),
+              reinterpret_cast<__nv_bfloat16 const*>(mat_a.data_ptr()),
+              reinterpret_cast<__nv_bfloat16 const*>(mat_b.data_ptr()), stream);
+    }
+  }
+}
+
+TORCH_LIBRARY_IMPL_EXPAND(TORCH_EXTENSION_NAME, CUDA, m) {
+  m.impl("dsv3_router_gemm", &dsv3_router_gemm);
+}

csrc/moe/dsv3_router_gemm_float_out.cu

@@ -0,0 +1,291 @@
+/*
+ * Adapted from SGLang's sgl-kernel implementation, which was adapted from
+ * https://github.com/NVIDIA/TensorRT-LLM/blob/main/cpp/tensorrt_llm/kernels/dsv3MinLatencyKernels/dsv3RouterGemm.cu
+ * https://github.com/NVIDIA/TensorRT-LLM/blob/main/cpp/tensorrt_llm/thop/dsv3RouterGemmOp.cpp
+ *
+ * Copyright (c) 2019-2023, NVIDIA CORPORATION.  All rights reserved.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include <ATen/ATen.h>
+#include <ATen/cuda/CUDAContext.h>
+
+#include <cuda_bf16.h>
+#include <cuda_runtime.h>
+
+#include "dsv3_router_gemm_utils.h"
+
+// Custom FMA implementation using PTX assembly instructions
+__device__ __forceinline__ void fma(float2& d, float2 const& a, float2 const& b,
+                                    float2 const& c) {
+  asm volatile("fma.rn.f32x2 %0, %1, %2, %3;\n"
+               : "=l"(reinterpret_cast<uint64_t&>(d))
+               : "l"(reinterpret_cast<uint64_t const&>(a)),
+                 "l"(reinterpret_cast<uint64_t const&>(b)),
+                 "l"(reinterpret_cast<uint64_t const&>(c)));
+}
+
+// Convert 8 bfloat16 values from a uint4 to float array - optimized conversion
+template <int VPT>
+__device__ __forceinline__ void bf16_uint4_to_float8(uint4 const& vec,
+                                                     float* dst) {
+  __nv_bfloat16* bf16_ptr =
+      reinterpret_cast<__nv_bfloat16*>(const_cast<uint4*>(&vec));
+
+#pragma unroll
+  for (int i = 0; i < VPT; i++) {
+    dst[i] = __bfloat162float(bf16_ptr[i]);
+  }
+}
+
+template <typename T, int kBlockSize, int VPT, int kNumTokens, int kNumExperts,
+          int kHiddenDim>
+__global__ __launch_bounds__(128, 1) void router_gemm_kernel_float_output(
+    float* out, T const* mat_a, T const* mat_b) {
+  // Each block handles one expert column
+  int const n_idx = blockIdx.x;
+  int const tid = threadIdx.x;
+  constexpr int kWarpSize = 32;
+  constexpr int kNumWarps = kBlockSize / kWarpSize;
+  // Constants for this kernel
+  constexpr int k_elems_per_k_iteration = VPT * kBlockSize;
+  constexpr int k_iterations =
+      kHiddenDim / k_elems_per_k_iteration;  // Total K iterations
+
+  // Initialize accumulators for all M rows
+  float acc[kNumTokens] = {};
+
+  // Shared memory for warp-level reduction
+  __shared__ float sm_reduction[kNumTokens][kNumWarps];  // kNumWarps
+
+  // B matrix is in column-major order, so we can directly load a column for the
+  // n_idx expert
+  T const* b_col = mat_b + n_idx * kHiddenDim;
+
+  // Pre-compute k_base values for each iteration to help compiler optimize
+  int k_bases[k_iterations];
+#pragma unroll
+  for (int ki = 0; ki < k_iterations; ki++) {
+    k_bases[ki] = ki * k_elems_per_k_iteration + tid * VPT;
+  }
+
+#if (defined(__CUDA_ARCH__) && (__CUDA_ARCH__ >= 900))
+  asm volatile("griddepcontrol.wait;");
+#endif
+
+  // Process the GEMM in chunks
+  for (int ki = 0; ki < k_iterations; ki++) {
+    int const k_base = k_bases[ki];
+
+    // Load B matrix values using vector load (8 bf16 values)
+    uint4 b_vec = *reinterpret_cast<uint4 const*>(b_col + k_base);
+
+    // Convert B values to float
+    float b_float[VPT];
+    bf16_uint4_to_float8<VPT>(b_vec, b_float);
+
+// Process each token
+#pragma unroll
+    for (int m_idx = 0; m_idx < kNumTokens; m_idx++) {
+      // Load both rows of A matrix using vector loads
+      uint4 a_vec = *reinterpret_cast<uint4 const*>(
+          mat_a + (m_idx * kHiddenDim) + k_base);
+
+      // Convert A values to float
+      float a_float[VPT];
+      bf16_uint4_to_float8<VPT>(a_vec, a_float);
+
+// Process elements in this chunk
+#pragma unroll
+      for (int k = 0; k < VPT; k++) {
+        float a = a_float[k];
+        float b = b_float[k];
+        acc[m_idx] += a * b;
+      }
+    }
+  }
+
+  // Perform warp-level reduction
+  int const warpSize = 32;
+  int const warpId = tid / warpSize;
+  int const laneId = tid % warpSize;
+
+  // Register for warp-level reduction results
+  float warp_result[kNumTokens];
+
+#pragma unroll
+  for (int m_idx = 0; m_idx < kNumTokens; m_idx++) {
+    warp_result[m_idx] = acc[m_idx];
+  }
+
+// Perform warp-level reduction using optimized butterfly pattern
+#pragma unroll
+  for (int m = 0; m < kNumTokens; m++) {
+    float sum = warp_result[m];
+
+    // Butterfly reduction pattern
+    sum += __shfl_xor_sync(0xffffffff, sum, 16);
+    sum += __shfl_xor_sync(0xffffffff, sum, 8);
+    sum += __shfl_xor_sync(0xffffffff, sum, 4);
+    sum += __shfl_xor_sync(0xffffffff, sum, 2);
+    sum += __shfl_xor_sync(0xffffffff, sum, 1);
+
+    // Only the first thread in each warp stores to shared memory
+    if (laneId == 0) {
+      sm_reduction[m][warpId] = sum;
+    }
+  }
+
+  __syncthreads();
+
+  // Final reduction across warps (only first thread)
+  if (tid == 0) {
+#pragma unroll
+    for (int m = 0; m < kNumTokens; m++) {
+      float final_sum = 0.0f;
+
+// Sum across the kNumWarps
+#pragma unroll
+      for (int w = 0; w < kNumWarps; w++) {
+        final_sum += sm_reduction[m][w];
+      }
+
+      // Write final result
+      out[m * kNumExperts + n_idx] = final_sum;
+    }
+  }
+#if (defined(__CUDA_ARCH__) && (__CUDA_ARCH__ >= 900))
+  asm volatile("griddepcontrol.launch_dependents;");
+#endif
+}
+
+template <typename T, int kNumTokens, int kNumExperts, int kHiddenDim>
+void invokeRouterGemmFloatOutput(float* output, T const* mat_a, T const* mat_b,
+                                 cudaStream_t stream) {
+  constexpr int VPT = 16 / sizeof(T);
+  constexpr int kBlockSize = 128;
+  cudaLaunchConfig_t config;
+  config.gridDim = kNumExperts;
+  config.blockDim = kBlockSize;
+  config.dynamicSmemBytes = 0;
+  config.stream = stream;
+  cudaLaunchAttribute attrs[1];
+  attrs[0].id = cudaLaunchAttributeProgrammaticStreamSerialization;
+  attrs[0].val.programmaticStreamSerializationAllowed = getEnvEnablePDL();
+  config.numAttrs = 1;
+  config.attrs = attrs;
+  cudaLaunchKernelEx(
+      &config,
+      router_gemm_kernel_float_output<T, kBlockSize, VPT, kNumTokens,
+                                      kNumExperts, kHiddenDim>,
+      output, mat_a, mat_b);
+}
+
+// Template instantiations for DEFAULT_NUM_EXPERTS experts
+template void invokeRouterGemmFloatOutput<__nv_bfloat16, 1, 256, 7168>(
+    float*, __nv_bfloat16 const*, __nv_bfloat16 const*, cudaStream_t);
+
+template void invokeRouterGemmFloatOutput<__nv_bfloat16, 2, 256, 7168>(
+    float*, __nv_bfloat16 const*, __nv_bfloat16 const*, cudaStream_t);
+
+template void invokeRouterGemmFloatOutput<__nv_bfloat16, 3, 256, 7168>(
+    float*, __nv_bfloat16 const*, __nv_bfloat16 const*, cudaStream_t);
+
+template void invokeRouterGemmFloatOutput<__nv_bfloat16, 4, 256, 7168>(
+    float*, __nv_bfloat16 const*, __nv_bfloat16 const*, cudaStream_t);
+
+template void invokeRouterGemmFloatOutput<__nv_bfloat16, 5, 256, 7168>(
+    float*, __nv_bfloat16 const*, __nv_bfloat16 const*, cudaStream_t);
+
+template void invokeRouterGemmFloatOutput<__nv_bfloat16, 6, 256, 7168>(
+    float*, __nv_bfloat16 const*, __nv_bfloat16 const*, cudaStream_t);
+
+template void invokeRouterGemmFloatOutput<__nv_bfloat16, 7, 256, 7168>(
+    float*, __nv_bfloat16 const*, __nv_bfloat16 const*, cudaStream_t);
+
+template void invokeRouterGemmFloatOutput<__nv_bfloat16, 8, 256, 7168>(
+    float*, __nv_bfloat16 const*, __nv_bfloat16 const*, cudaStream_t);
+
+template void invokeRouterGemmFloatOutput<__nv_bfloat16, 9, 256, 7168>(
+    float*, __nv_bfloat16 const*, __nv_bfloat16 const*, cudaStream_t);
+
+template void invokeRouterGemmFloatOutput<__nv_bfloat16, 10, 256, 7168>(
+    float*, __nv_bfloat16 const*, __nv_bfloat16 const*, cudaStream_t);
+
+template void invokeRouterGemmFloatOutput<__nv_bfloat16, 11, 256, 7168>(
+    float*, __nv_bfloat16 const*, __nv_bfloat16 const*, cudaStream_t);
+
+template void invokeRouterGemmFloatOutput<__nv_bfloat16, 12, 256, 7168>(
+    float*, __nv_bfloat16 const*, __nv_bfloat16 const*, cudaStream_t);
+
+template void invokeRouterGemmFloatOutput<__nv_bfloat16, 13, 256, 7168>(
+    float*, __nv_bfloat16 const*, __nv_bfloat16 const*, cudaStream_t);
+
+template void invokeRouterGemmFloatOutput<__nv_bfloat16, 14, 256, 7168>(
+    float*, __nv_bfloat16 const*, __nv_bfloat16 const*, cudaStream_t);
+
+template void invokeRouterGemmFloatOutput<__nv_bfloat16, 15, 256, 7168>(
+    float*, __nv_bfloat16 const*, __nv_bfloat16 const*, cudaStream_t);
+
+template void invokeRouterGemmFloatOutput<__nv_bfloat16, 16, 256, 7168>(
+    float*, __nv_bfloat16 const*, __nv_bfloat16 const*, cudaStream_t);
+
+// Template instantiations for KIMI_K2_NUM_EXPERTS experts
+template void invokeRouterGemmFloatOutput<__nv_bfloat16, 1, 384, 7168>(
+    float*, __nv_bfloat16 const*, __nv_bfloat16 const*, cudaStream_t);
+
+template void invokeRouterGemmFloatOutput<__nv_bfloat16, 2, 384, 7168>(
+    float*, __nv_bfloat16 const*, __nv_bfloat16 const*, cudaStream_t);
+
+template void invokeRouterGemmFloatOutput<__nv_bfloat16, 3, 384, 7168>(
+    float*, __nv_bfloat16 const*, __nv_bfloat16 const*, cudaStream_t);
+
+template void invokeRouterGemmFloatOutput<__nv_bfloat16, 4, 384, 7168>(
+    float*, __nv_bfloat16 const*, __nv_bfloat16 const*, cudaStream_t);
+
+template void invokeRouterGemmFloatOutput<__nv_bfloat16, 5, 384, 7168>(
+    float*, __nv_bfloat16 const*, __nv_bfloat16 const*, cudaStream_t);
+
+template void invokeRouterGemmFloatOutput<__nv_bfloat16, 6, 384, 7168>(
+    float*, __nv_bfloat16 const*, __nv_bfloat16 const*, cudaStream_t);
+
+template void invokeRouterGemmFloatOutput<__nv_bfloat16, 7, 384, 7168>(
+    float*, __nv_bfloat16 const*, __nv_bfloat16 const*, cudaStream_t);
+
+template void invokeRouterGemmFloatOutput<__nv_bfloat16, 8, 384, 7168>(
+    float*, __nv_bfloat16 const*, __nv_bfloat16 const*, cudaStream_t);
+
+template void invokeRouterGemmFloatOutput<__nv_bfloat16, 9, 384, 7168>(
+    float*, __nv_bfloat16 const*, __nv_bfloat16 const*, cudaStream_t);
+
+template void invokeRouterGemmFloatOutput<__nv_bfloat16, 10, 384, 7168>(
+    float*, __nv_bfloat16 const*, __nv_bfloat16 const*, cudaStream_t);
+
+template void invokeRouterGemmFloatOutput<__nv_bfloat16, 11, 384, 7168>(
+    float*, __nv_bfloat16 const*, __nv_bfloat16 const*, cudaStream_t);
+
+template void invokeRouterGemmFloatOutput<__nv_bfloat16, 12, 384, 7168>(
+    float*, __nv_bfloat16 const*, __nv_bfloat16 const*, cudaStream_t);
+
+template void invokeRouterGemmFloatOutput<__nv_bfloat16, 13, 384, 7168>(
+    float*, __nv_bfloat16 const*, __nv_bfloat16 const*, cudaStream_t);
+
+template void invokeRouterGemmFloatOutput<__nv_bfloat16, 14, 384, 7168>(
+    float*, __nv_bfloat16 const*, __nv_bfloat16 const*, cudaStream_t);
+
+template void invokeRouterGemmFloatOutput<__nv_bfloat16, 15, 384, 7168>(
+    float*, __nv_bfloat16 const*, __nv_bfloat16 const*, cudaStream_t);
+
+template void invokeRouterGemmFloatOutput<__nv_bfloat16, 16, 384, 7168>(
+    float*, __nv_bfloat16 const*, __nv_bfloat16 const*, cudaStream_t);

csrc/moe/dsv3_router_gemm_utils.h

@@ -0,0 +1,43 @@
+/*
+ * Adapted from SGLang's sgl-kernel implementation, which was adapted from
+ * https://github.com/NVIDIA/TensorRT-LLM/blob/main/cpp/tensorrt_llm/kernels/dsv3MinLatencyKernels/dsv3RouterGemm.cu
+ * https://github.com/NVIDIA/TensorRT-LLM/blob/main/cpp/tensorrt_llm/thop/dsv3RouterGemmOp.cpp
+ *
+ * Copyright (c) 2019-2023, NVIDIA CORPORATION.  All rights reserved.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#pragma once
+
+#include <ATen/cuda/CUDAContext.h>
+
+#include <cstdlib>
+#include <mutex>
+
+inline int getSMVersion() {
+  auto* props = at::cuda::getCurrentDeviceProperties();
+  return props->major * 10 + props->minor;
+}
+
+inline bool getEnvEnablePDL() {
+  static std::once_flag flag;
+  static bool enablePDL = false;
+  std::call_once(flag, [&]() {
+    if (getSMVersion() >= 90) {
+      const char* env = std::getenv("TRTLLM_ENABLE_PDL");
+      enablePDL = env && env[0] == '1' && env[1] == '\0';
+    }
+  });
+  return enablePDL;
+}

csrc/moe/grouped_topk_kernels.cu

@@ -1,6 +1,6 @@
 /*
  * Adapted from
- * https://github.com/NVIDIA/TensorRT-LLM/blob/v0.21.0/cpp/tensorrt_llm/kernels/noAuxTcKernels.cu
+ * https://github.com/NVIDIA/TensorRT-LLM/blob/v1.3.0rc2/cpp/tensorrt_llm/kernels/noAuxTcKernels.cu
  * Copyright (c) 2025, The vLLM team.
  * SPDX-FileCopyrightText: Copyright (c) 1993-2024 NVIDIA CORPORATION &
  * AFFILIATES. All rights reserved. SPDX-License-Identifier: Apache-2.0
@@ -17,8 +17,10 @@
  * See the License for the specific language governing permissions and
  * limitations under the License.
  */
+#include "moeTopKFuncs.cuh"
 #include <c10/cuda/CUDAStream.h>
 #include <torch/all.h>
+#include <cmath>
 #include <cuda_fp16.h>
 #include <cuda_bf16.h>
 #include <cuda/std/limits>
@@ -30,7 +32,17 @@ namespace vllm {
 namespace moe {
 
 constexpr unsigned FULL_WARP_MASK = 0xffffffff;
-constexpr int32_t WARP_SIZE = 32;
+static constexpr int WARP_SIZE = 32;
+static constexpr int NumNemotronExperts = 512;
+static constexpr int NumKimiK2Experts = 384;
+static constexpr int NumDeepseekExperts = 256;
+static constexpr int MaxSupportedExpertCount =
+    std::max({NumNemotronExperts, NumKimiK2Experts, NumDeepseekExperts});
+static constexpr int MaxNumExpertsUnit = 128;
+static constexpr int NumTopGroupScores = 2;
+static constexpr int DefaultMaxNumTopExperts = 8;
+static constexpr int MaxSupportedTopExperts = 22;
+static constexpr int MaxNumTopGroups = 4;
 
 namespace warp_topk {
 
@@ -657,76 +669,335 @@ __global__ void grouped_topk_fused_kernel(
 #endif
 }
 
-template <typename T, typename BiasT, typename IdxT>
+template <typename T, typename BiasT, typename IdxT, ScoringFunc SF,
+          int MaxNumExperts, bool UseGroups,
+          int MaxNumTopExperts = DefaultMaxNumTopExperts>
+__global__ void grouped_topk_fused_small_expert_count_kernel(
+    T* scores, float* topkValues, IdxT* topkIndices, BiasT const* routingBias,
+    int64_t const numTokens, int64_t const numGroup, int64_t const topkGroup,
+    int64_t const topk, int64_t const numExperts,
+    int64_t const numExpertsPerGroup, bool const renormalize,
+    double const routedScalingFactor) {
+#if (defined(__CUDA_ARCH__) && (__CUDA_ARCH__ >= 900))
+  cudaGridDependencySynchronize();
+#endif
+  // declare shared memory structure
+  // number of experts is bounded by number of threads
+  __shared__ float __attribute((aligned(128))) smemScoreSigmoid[MaxNumExperts];
+  __shared__ float __attribute((aligned(128))) smemScoreBias[MaxNumExperts];
+  // number of expert groups is bounded by number of warps
+  int constexpr NumWarps = MaxNumExperts / WARP_SIZE;
+  __shared__ float __attribute((aligned(128))) smemGroupScores[NumWarps];
+
+  // needed for warp reduce
+  auto block = cg::this_thread_block();
+  auto warp = cg::tiled_partition<WARP_SIZE>(block);
+
+  // for the final reduction of weight norm, only some lanes need to participate
+  int32_t laneIdx = threadIdx.x % WARP_SIZE;
+  int32_t warpIdx = __shfl_sync(0xffffffff, threadIdx.x / WARP_SIZE, 0);
+
+  if constexpr (UseGroups) {
+    if (warpIdx >= numGroup) {
+      return;
+    }
+  }
+  // note that for invalid scores, we simply use a negative value:
+  // they work well even with the compacted format used in topK, and
+  // sigmoid / bias activated scores cannot be negative
+  const float invalidScoreFloat = float{-INFINITY};
+
+  // load bias already; each warp represents one expert group
+  auto threadExpert = threadIdx.x;
+  bool expertSelected = threadExpert < numExperts;
+  if constexpr (UseGroups) {
+    threadExpert = warpIdx * numExpertsPerGroup + laneIdx;
+    expertSelected = laneIdx < numExpertsPerGroup;
+  }
+
+  auto scoreIdx = int64_t{blockIdx.x} * int64_t{numExperts} + threadExpert;
+  auto biasVal = expertSelected ? static_cast<float>(routingBias[threadExpert])
+                                : invalidScoreFloat;
+  topkValues += blockIdx.x * topk;
+  topkIndices += blockIdx.x * topk;
+
+  // get our assigned thread score; each warp represents one expert group
+  float score =
+      expertSelected ? static_cast<float>(scores[scoreIdx]) : invalidScoreFloat;
+  auto scoreSigmoid = apply_scoring<SF>(score);
+  // write the sigmoid score to shared for later use
+  if (expertSelected) {
+    smemScoreSigmoid[threadExpert] = scoreSigmoid;
+  }
+
+  // get the score with bias
+  // note that with invalid values, because sigmoid is < 1 and bias is -1,
+  // we must get a negative value, which is smaller than any valid value
+  auto scoreBias = float{scoreSigmoid + float{biasVal}};
+
+  if (expertSelected) {
+    smemScoreBias[threadExpert] = scoreBias;
+  }
+
+  // registers for top group score reduction
+  float topExpGroupScores[NumTopGroupScores];
+  [[maybe_unused]] int32_t topExpGroupIdx[NumTopGroupScores];
+  float topGroups[MaxNumTopGroups];  // bound of numGroup
+  int32_t topGroupIdx[MaxNumTopGroups];
+  float expertScoreGroup[MaxNumTopGroups];
+  int32_t expertIdxGroup[MaxNumTopGroups];
+  float topScores[MaxNumTopExperts];  // bound of topk
+  int32_t topExperts[MaxNumTopExperts];
+
+  if constexpr (UseGroups) {
+    reduce_topk::reduceTopK(warp, topExpGroupScores, topExpGroupIdx, scoreBias,
+                            threadExpert,
+                            /* minValue */ invalidScoreFloat);
+
+    // get the final group score and write it to shared
+    if (warp.thread_rank() == 0) {
+      auto groupScore = topExpGroupScores[0] + topExpGroupScores[1];
+      smemGroupScores[warpIdx] = groupScore;
+    }
+  }
+
+  // make group scores available to all warps
+  __syncthreads();
+
+  if constexpr (UseGroups) {
+    if (warpIdx == 0) {
+      // a single warp performs the selection of top groups, and goes on to
+      // select the final experts
+      float groupScore =
+          laneIdx < numGroup ? smemGroupScores[laneIdx] : invalidScoreFloat;
+
+      reduce_topk::reduceTopK(warp, topGroups, topGroupIdx, groupScore, laneIdx,
+                              /* minValue */ invalidScoreFloat);
+      // final expert selection: get relevant indexes and scores from shared
+#pragma unroll
+      for (int ii = 0; ii < MaxNumTopGroups; ++ii) {  // bound of numGroup
+        auto groupIdx = topGroupIdx[ii];
+        expertIdxGroup[ii] = groupIdx * numExpertsPerGroup + laneIdx;
+
+        expertScoreGroup[ii] = (ii < topkGroup) && expertSelected
+                                   ? smemScoreBias[expertIdxGroup[ii]]
+                                   : invalidScoreFloat;
+      }
+
+      reduce_topk::reduceTopK(warp, topScores, topExperts, expertScoreGroup,
+                              expertIdxGroup, /* minValue */ invalidScoreFloat,
+                              topk);
+    }
+  } else if constexpr (MaxNumExperts > MaxNumExpertsUnit) {
+    // without groups, and the expert number is larger than MaxNumExpertsUnit,
+    // we need to use multiple warps to calculate the intermediate topk results
+
+    int constexpr NumExpertWarps = (MaxNumExperts - 1) / MaxNumExpertsUnit + 1;
+    int constexpr NumInterTopK = NumExpertWarps * MaxNumTopExperts;
+    __shared__ float
+        __attribute((aligned(128))) smemInterTopScores[NumInterTopK];
+    __shared__ int32_t
+        __attribute((aligned(128))) smemInterTopExperts[NumInterTopK];
+    if (warpIdx < NumExpertWarps) {
+      int offset = warpIdx * WARP_SIZE * MaxNumTopGroups;
+#pragma unroll
+      for (int ii = 0; ii < MaxNumTopGroups; ++ii) {
+        auto expertIdx = ii * WARP_SIZE + laneIdx;
+        expertIdxGroup[ii] = offset + expertIdx;
+        expertScoreGroup[ii] = offset + expertIdx < numExperts
+                                   ? smemScoreBias[offset + expertIdx]
+                                   : invalidScoreFloat;
+      }
+      reduce_topk::reduceTopK(warp, topScores, topExperts, expertScoreGroup,
+                              expertIdxGroup,
+                              /* minValue */ invalidScoreFloat, topk);
+
+      if (laneIdx < topk) {
+        smemInterTopScores[warpIdx * MaxNumTopExperts + laneIdx] =
+            topScores[laneIdx];
+        smemInterTopExperts[warpIdx * MaxNumTopExperts + laneIdx] =
+            topExperts[laneIdx];
+      } else if (laneIdx >= topk && laneIdx < MaxNumTopExperts) {
+        smemInterTopScores[warpIdx * MaxNumTopExperts + laneIdx] =
+            invalidScoreFloat;
+        smemInterTopExperts[warpIdx * MaxNumTopExperts + laneIdx] =
+            MaxNumExperts - 1;
+      }
+    }
+    __syncthreads();
+    if (warpIdx == 0) {
+      int constexpr NumInterTopKPerThread = (NumInterTopK - 1) / WARP_SIZE + 1;
+      float intermediateScore[NumInterTopKPerThread];
+      int32_t intermediateExpert[NumInterTopKPerThread];
+      for (int i = laneIdx; i < NumInterTopKPerThread * WARP_SIZE;
+           i += WARP_SIZE) {
+        int ii = i / WARP_SIZE;
+        if (i < NumInterTopK) {
+          intermediateScore[ii] = smemInterTopScores[i];
+          intermediateExpert[ii] = smemInterTopExperts[i];
+        } else {
+          intermediateScore[ii] = invalidScoreFloat;
+          intermediateExpert[ii] = MaxNumExperts - 1;
+        }
+      }
+      reduce_topk::reduceTopK(warp, topScores, topExperts, intermediateScore,
+                              intermediateExpert,
+                              /* minValue */ invalidScoreFloat, topk);
+    }
+  } else {
+    // without groups, and the expert number is smaller than MaxNumExpertsUnit
+    // each thread just takes `MaxNumTopGroups` experts
+    if (warpIdx == 0) {
+#pragma unroll
+      for (int ii = 0; ii < MaxNumTopGroups; ++ii) {
+        auto expertIdx = ii * WARP_SIZE + laneIdx;
+        expertIdxGroup[ii] = expertIdx;
+        expertScoreGroup[ii] = expertIdx < numExperts ? smemScoreBias[expertIdx]
+                                                      : invalidScoreFloat;
+      }
+      reduce_topk::reduceTopK(warp, topScores, topExperts, expertScoreGroup,
+                              expertIdxGroup,
+                              /* minValue */ invalidScoreFloat, topk);
+    }
+  }
+
+  if (warpIdx == 0) {
+    // determine our lane's expert index and write to output
+    int32_t expertIdx =
+        laneIdx < topk ? topExperts[laneIdx] : MaxNumExperts - 1;
+    float scoreNorm = laneIdx < topk ? smemScoreSigmoid[expertIdx] : 0.F;
+    float finalScore = static_cast<float>(scoreNorm * routedScalingFactor);
+    // norm the value
+    if (renormalize) {
+      auto redNorm = cg::reduce(warp, scoreNorm, cg::plus<float>{});
+      finalScore /= (redNorm + 1e-20);
+    }
+    // store the topk scores and experts to output
+    if (laneIdx < topk) {
+      topkValues[laneIdx] = finalScore;
+      topkIndices[laneIdx] = expertIdx;
+    }
+  }
+
+#if (defined(__CUDA_ARCH__) && (__CUDA_ARCH__ >= 900))
+  cudaTriggerProgrammaticLaunchCompletion();
+#endif
+}
+
+template <typename T, typename BiasT, typename IdxT, ScoringFunc SF>
 void invokeNoAuxTc(T* scores, float* topk_values, IdxT* topk_indices,
                    BiasT const* bias, int64_t const num_tokens,
                    int64_t const num_experts, int64_t const n_group,
                    int64_t const topk_group, int64_t const topk,
                    bool const renormalize, double const routed_scaling_factor,
-                   int const scoring_func, bool enable_pdl = false,
-                   cudaStream_t const stream = 0) {
+                   bool enable_pdl = false, cudaStream_t const stream = 0) {
   cudaLaunchConfig_t config;
-  // One block per token; one warp per group.
-  config.gridDim = static_cast<uint32_t>(num_tokens);
-  config.blockDim = static_cast<uint32_t>(n_group) * WARP_SIZE;
-  // Dynamic shared memory: WarpSelect staging + per-group topk buffers.
-  int32_t const num_warps = static_cast<int32_t>(n_group);
-  size_t const val_bytes =
-      static_cast<size_t>(num_warps) * WARP_SIZE * sizeof(T);
-  size_t const val_bytes_aligned =
-      warp_topk::round_up_to_multiple_of<256>(val_bytes);
-  size_t const idx_bytes =
-      static_cast<size_t>(num_warps) * WARP_SIZE * sizeof(int32_t);
-  size_t const internal_bytes = val_bytes_aligned + idx_bytes;
-  size_t const extra_bytes = 16 + static_cast<size_t>(n_group) * sizeof(T);
-  config.dynamicSmemBytes = internal_bytes + extra_bytes;
   config.stream = stream;
   cudaLaunchAttribute attrs[1];
   attrs[0].id = cudaLaunchAttributeProgrammaticStreamSerialization;
   attrs[0].val.programmaticStreamSerializationAllowed = enable_pdl;
   config.numAttrs = 1;
   config.attrs = attrs;
-  auto const sf = static_cast<ScoringFunc>(scoring_func);
-  switch (sf) {
-    case SCORING_NONE: {
-      auto* kernel_instance =
-          &grouped_topk_fused_kernel<T, BiasT, IdxT, SCORING_NONE>;
-      cudaLaunchKernelEx(&config, kernel_instance, scores, topk_values,
-                         topk_indices, bias, num_tokens, num_experts, n_group,
-                         topk_group, topk, renormalize, routed_scaling_factor);
-      return;
+
+  // Check if we can use the optimized
+  // grouped_topk_fused_small_expert_count_kernel
+  bool const is_single_group =
+      (n_group == 1) && (topk_group == 1) &&
+      (num_experts <= MaxSupportedExpertCount) &&
+      (topk <= DefaultMaxNumTopExperts || topk == MaxSupportedTopExperts);
+
+  int64_t const experts_per_group = num_experts / n_group;
+  bool const is_multi_group =
+      (n_group > 1) && (num_experts <= NumDeepseekExperts) &&
+      (experts_per_group <= WARP_SIZE) &&
+      (experts_per_group * topk_group <= MaxNumExpertsUnit) &&
+      (topk <= DefaultMaxNumTopExperts) && (topk_group <= MaxNumTopGroups);
+
+  if (is_single_group || is_multi_group) {
+    auto* kernel_instance =
+        &grouped_topk_fused_small_expert_count_kernel<T, BiasT, IdxT, SF,
+                                                      NumDeepseekExperts, true>;
+    int num_threads = NumDeepseekExperts;
+    if (is_single_group) {
+      // Special case for Nemotron, which selects top 22 from 512 experts, and 1
+      // group only.
+      if (num_experts == NumNemotronExperts && n_group == 1 &&
+          topk == MaxSupportedTopExperts) {
+        kernel_instance = &grouped_topk_fused_small_expert_count_kernel<
+            T, BiasT, IdxT, SF, NumNemotronExperts, false,
+            MaxSupportedTopExperts>;
+        num_threads = NumNemotronExperts;
+      } else if (num_experts > NumKimiK2Experts &&
+                 num_experts <= MaxSupportedExpertCount) {
+        kernel_instance = &grouped_topk_fused_small_expert_count_kernel<
+            T, BiasT, IdxT, SF, MaxSupportedExpertCount, false>;
+        num_threads = MaxSupportedExpertCount;
+      } else if (num_experts > MaxNumExpertsUnit &&
+                 num_experts <= NumKimiK2Experts) {
+        kernel_instance = &grouped_topk_fused_small_expert_count_kernel<
+            T, BiasT, IdxT, SF, NumKimiK2Experts, false>;
+        num_threads = NumKimiK2Experts;
+      } else {
+        kernel_instance = &grouped_topk_fused_small_expert_count_kernel<
+            T, BiasT, IdxT, SF, MaxNumExpertsUnit, false>;
+        num_threads = MaxNumExpertsUnit;
+      }
     }
-    case SCORING_SIGMOID: {
-      auto* kernel_instance =
-          &grouped_topk_fused_kernel<T, BiasT, IdxT, SCORING_SIGMOID>;
-      cudaLaunchKernelEx(&config, kernel_instance, scores, topk_values,
-                         topk_indices, bias, num_tokens, num_experts, n_group,
-                         topk_group, topk, renormalize, routed_scaling_factor);
-      return;
-    }
-    default:
-      // should be guarded by higher level checks.
-      TORCH_CHECK(false, "Unsupported scoring_func in invokeNoAuxTc");
+    config.gridDim = num_tokens;
+    config.blockDim = num_threads;
+    config.dynamicSmemBytes = 0;
+    cudaLaunchKernelEx(&config, kernel_instance, scores, topk_values,
+                       topk_indices, bias, num_tokens, n_group, topk_group,
+                       topk, num_experts, num_experts / n_group, renormalize,
+                       routed_scaling_factor);
+  } else {
+    auto* kernel_instance = &grouped_topk_fused_kernel<T, BiasT, IdxT, SF>;
+    // One block per token; one warp per group.
+    config.gridDim = static_cast<uint32_t>(num_tokens);
+    config.blockDim = static_cast<uint32_t>(n_group) * WARP_SIZE;
+    // Dynamic shared memory: WarpSelect staging + per-group topk buffers.
+    int32_t const num_warps = static_cast<int32_t>(n_group);
+    size_t const val_bytes =
+        static_cast<size_t>(num_warps) * WARP_SIZE * sizeof(T);
+    size_t const val_bytes_aligned =
+        warp_topk::round_up_to_multiple_of<256>(val_bytes);
+    size_t const idx_bytes =
+        static_cast<size_t>(num_warps) * WARP_SIZE * sizeof(int32_t);
+    size_t const internal_bytes = val_bytes_aligned + idx_bytes;
+    size_t const extra_bytes = 16 + static_cast<size_t>(n_group) * sizeof(T);
+    config.dynamicSmemBytes = internal_bytes + extra_bytes;
+    cudaLaunchKernelEx(&config, kernel_instance, scores, topk_values,
+                       topk_indices, bias, num_tokens, num_experts, n_group,
+                       topk_group, topk, renormalize, routed_scaling_factor);
   }
 }
 
-#define INSTANTIATE_NOAUX_TC(T, BiasT, IdxT)                                 \
-  template void invokeNoAuxTc<T, BiasT, IdxT>(                               \
+#define INSTANTIATE_NOAUX_TC(T, BiasT, IdxT, SF)                             \
+  template void invokeNoAuxTc<T, BiasT, IdxT, SF>(                           \
       T * scores, float* topk_values, IdxT* topk_indices, BiasT const* bias, \
       int64_t const num_tokens, int64_t const num_experts,                   \
       int64_t const n_group, int64_t const topk_group, int64_t const topk,   \
       bool const renormalize, double const routed_scaling_factor,            \
-      int const scoring_func, bool enable_pdl, cudaStream_t const stream);
+      bool enable_pdl, cudaStream_t const stream);
 
-INSTANTIATE_NOAUX_TC(float, float, int32_t);
-INSTANTIATE_NOAUX_TC(float, half, int32_t);
-INSTANTIATE_NOAUX_TC(float, __nv_bfloat16, int32_t);
-INSTANTIATE_NOAUX_TC(half, float, int32_t);
-INSTANTIATE_NOAUX_TC(half, half, int32_t);
-INSTANTIATE_NOAUX_TC(half, __nv_bfloat16, int32_t);
-INSTANTIATE_NOAUX_TC(__nv_bfloat16, float, int32_t);
-INSTANTIATE_NOAUX_TC(__nv_bfloat16, half, int32_t);
-INSTANTIATE_NOAUX_TC(__nv_bfloat16, __nv_bfloat16, int32_t);
+INSTANTIATE_NOAUX_TC(float, float, int32_t, SCORING_SIGMOID);
+INSTANTIATE_NOAUX_TC(float, half, int32_t, SCORING_SIGMOID);
+INSTANTIATE_NOAUX_TC(float, __nv_bfloat16, int32_t, SCORING_SIGMOID);
+INSTANTIATE_NOAUX_TC(half, float, int32_t, SCORING_SIGMOID);
+INSTANTIATE_NOAUX_TC(half, half, int32_t, SCORING_SIGMOID);
+INSTANTIATE_NOAUX_TC(half, __nv_bfloat16, int32_t, SCORING_SIGMOID);
+INSTANTIATE_NOAUX_TC(__nv_bfloat16, float, int32_t, SCORING_SIGMOID);
+INSTANTIATE_NOAUX_TC(__nv_bfloat16, half, int32_t, SCORING_SIGMOID);
+INSTANTIATE_NOAUX_TC(__nv_bfloat16, __nv_bfloat16, int32_t, SCORING_SIGMOID);
+INSTANTIATE_NOAUX_TC(float, float, int32_t, SCORING_NONE);
+INSTANTIATE_NOAUX_TC(float, half, int32_t, SCORING_NONE);
+INSTANTIATE_NOAUX_TC(float, __nv_bfloat16, int32_t, SCORING_NONE);
+INSTANTIATE_NOAUX_TC(half, float, int32_t, SCORING_NONE);
+INSTANTIATE_NOAUX_TC(half, half, int32_t, SCORING_NONE);
+INSTANTIATE_NOAUX_TC(half, __nv_bfloat16, int32_t, SCORING_NONE);
+INSTANTIATE_NOAUX_TC(__nv_bfloat16, float, int32_t, SCORING_NONE);
+INSTANTIATE_NOAUX_TC(__nv_bfloat16, half, int32_t, SCORING_NONE);
+INSTANTIATE_NOAUX_TC(__nv_bfloat16, __nv_bfloat16, int32_t, SCORING_NONE);
 }  // end namespace moe
 }  // namespace vllm
 
@@ -762,46 +1033,53 @@ std::tuple<torch::Tensor, torch::Tensor> grouped_topk(
       {num_tokens, topk}, torch::dtype(torch::kInt32).device(torch::kCUDA));
 
   auto stream = c10::cuda::getCurrentCUDAStream(scores.get_device());
+  auto const sf = static_cast<vllm::moe::ScoringFunc>(scoring_func);
 
-#define LAUNCH_KERNEL(T, IdxT)                                               \
-  do {                                                                       \
-    switch (bias_type) {                                                     \
-      case torch::kFloat16:                                                  \
-        vllm::moe::invokeNoAuxTc<T, half, IdxT>(                             \
-            reinterpret_cast<T*>(scores.mutable_data_ptr()),                 \
-            reinterpret_cast<float*>(topk_values.mutable_data_ptr()),        \
-            reinterpret_cast<IdxT*>(topk_indices.mutable_data_ptr()),        \
-            reinterpret_cast<half const*>(bias.data_ptr()), num_tokens,      \
-            num_experts, n_group, topk_group, topk, renormalize,             \
-            routed_scaling_factor, static_cast<int>(scoring_func), false,    \
-            stream);                                                         \
-        break;                                                               \
-      case torch::kFloat32:                                                  \
-        vllm::moe::invokeNoAuxTc<T, float, IdxT>(                            \
-            reinterpret_cast<T*>(scores.mutable_data_ptr()),                 \
-            reinterpret_cast<float*>(topk_values.mutable_data_ptr()),        \
-            reinterpret_cast<IdxT*>(topk_indices.mutable_data_ptr()),        \
-            reinterpret_cast<float const*>(bias.data_ptr()), num_tokens,     \
-            num_experts, n_group, topk_group, topk, renormalize,             \
-            routed_scaling_factor, static_cast<int>(scoring_func), false,    \
-            stream);                                                         \
-        break;                                                               \
-      case torch::kBFloat16:                                                 \
-        vllm::moe::invokeNoAuxTc<T, __nv_bfloat16, IdxT>(                    \
-            reinterpret_cast<T*>(scores.mutable_data_ptr()),                 \
-            reinterpret_cast<float*>(topk_values.mutable_data_ptr()),        \
-            reinterpret_cast<IdxT*>(topk_indices.mutable_data_ptr()),        \
-            reinterpret_cast<__nv_bfloat16 const*>(bias.data_ptr()),         \
-            num_tokens, num_experts, n_group, topk_group, topk, renormalize, \
-            routed_scaling_factor, static_cast<int>(scoring_func), false,    \
-            stream);                                                         \
-        break;                                                               \
-      default:                                                               \
-        throw std::invalid_argument(                                         \
-            "Invalid bias dtype, only supports float16, float32, and "       \
-            "bfloat16");                                                     \
-        break;                                                               \
-    }                                                                        \
+#define LAUNCH_KERNEL_SF(T, BiasT, IdxT)                                      \
+  do {                                                                        \
+    switch (sf) {                                                             \
+      case vllm::moe::SCORING_NONE:                                           \
+        vllm::moe::invokeNoAuxTc<T, BiasT, IdxT, vllm::moe::SCORING_NONE>(    \
+            reinterpret_cast<T*>(scores.mutable_data_ptr()),                  \
+            reinterpret_cast<float*>(topk_values.mutable_data_ptr()),         \
+            reinterpret_cast<IdxT*>(topk_indices.mutable_data_ptr()),         \
+            reinterpret_cast<BiasT const*>(bias.data_ptr()), num_tokens,      \
+            num_experts, n_group, topk_group, topk, renormalize,              \
+            routed_scaling_factor, false, stream);                            \
+        break;                                                                \
+      case vllm::moe::SCORING_SIGMOID:                                        \
+        vllm::moe::invokeNoAuxTc<T, BiasT, IdxT, vllm::moe::SCORING_SIGMOID>( \
+            reinterpret_cast<T*>(scores.mutable_data_ptr()),                  \
+            reinterpret_cast<float*>(topk_values.mutable_data_ptr()),         \
+            reinterpret_cast<IdxT*>(topk_indices.mutable_data_ptr()),         \
+            reinterpret_cast<BiasT const*>(bias.data_ptr()), num_tokens,      \
+            num_experts, n_group, topk_group, topk, renormalize,              \
+            routed_scaling_factor, false, stream);                            \
+        break;                                                                \
+      default:                                                                \
+        throw std::invalid_argument("Unsupported scoring_func");              \
+        break;                                                                \
+    }                                                                         \
+  } while (0)
+
+#define LAUNCH_KERNEL(T, IdxT)                                         \
+  do {                                                                 \
+    switch (bias_type) {                                               \
+      case torch::kFloat16:                                            \
+        LAUNCH_KERNEL_SF(T, half, IdxT);                               \
+        break;                                                         \
+      case torch::kFloat32:                                            \
+        LAUNCH_KERNEL_SF(T, float, IdxT);                              \
+        break;                                                         \
+      case torch::kBFloat16:                                           \
+        LAUNCH_KERNEL_SF(T, __nv_bfloat16, IdxT);                      \
+        break;                                                         \
+      default:                                                         \
+        throw std::invalid_argument(                                   \
+            "Invalid bias dtype, only supports float16, float32, and " \
+            "bfloat16");                                               \
+        break;                                                         \
+    }                                                                  \
   } while (0)
 
   switch (data_type) {
@@ -824,5 +1102,6 @@ std::tuple<torch::Tensor, torch::Tensor> grouped_topk(
       break;
   }
 #undef LAUNCH_KERNEL
+#undef LAUNCH_KERNEL_SF
   return {topk_values, topk_indices};
 }

csrc/moe/moeTopKFuncs.cuh

@@ -0,0 +1,257 @@
+/*
+ * Adapted from
+ * https://github.com/NVIDIA/TensorRT-LLM/blob/v1.3.0rc2/cpp/tensorrt_llm/kernels/moeTopKFuncs.cuh
+ * Copyright (c) 2026, The vLLM team.
+ * SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION. All rights
+ * reserved. SPDX-License-Identifier: Apache-2.0
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#pragma once
+
+#include <cooperative_groups.h>
+#include <cooperative_groups/reduce.h>
+#include <cub/cub.cuh>
+
+namespace vllm {
+namespace moe {
+namespace reduce_topk {
+namespace cg = cooperative_groups;
+static constexpr int kWARP_SIZE = 32;
+
+template <typename T_>
+struct TopKRedType {
+  using T = T_;
+  static_assert(
+      std::is_same_v<T, float> || std::is_same_v<T, half> ||
+          std::is_same_v<T, __nv_bfloat16> || std::is_same_v<T, int>,
+      "Top K reduction only implemented for int, float, float16 and bfloat16");
+
+  using TypeCmp = std::conditional_t<sizeof(T) == 4, uint64_t, uint32_t>;
+  using IdxT = std::conditional_t<sizeof(T) == 4, int32_t, int16_t>;
+
+  static constexpr int kMoveBits = (sizeof(T) == 4) ? 32 : 16;
+  static constexpr int kMaxIdx = 65535;
+  TypeCmp compValIdx;
+
+  static __host__ __device__ inline TypeCmp makeCmpVal(T val, int32_t idx = 0) {
+    auto valueBits = cub::Traits<T>::TwiddleIn(
+        reinterpret_cast<typename cub::Traits<T>::UnsignedBits&>(val));
+    TypeCmp compactTmp = valueBits;
+    compactTmp = (compactTmp << kMoveBits) | (0xFFFF & (kMaxIdx - idx));
+    // Use 65535 minus idx to give higher priority to elements with smaller
+    // indices.
+    return compactTmp;
+  }
+
+  static __host__ __device__ void unpack(T& value, int32_t& index,
+                                         TypeCmp cmp) {
+    // Since “65535-idx” is always smaller than 65536 and positive, we can
+    // directly use it as the lower 16 bits
+    index = kMaxIdx - static_cast<int32_t>((cmp & 0xFFFF));
+
+    auto compactTmp = cmp >> kMoveBits;
+    auto valueBits = cub::Traits<T>::TwiddleOut(
+        reinterpret_cast<typename cub::Traits<T>::UnsignedBits&>(compactTmp));
+    value = reinterpret_cast<T&>(valueBits);
+  }
+
+  __host__ __device__ TopKRedType() = default;
+
+  __host__ __device__ TopKRedType(T val, int32_t idx)
+      : compValIdx(makeCmpVal(val, idx)) {}
+
+  __host__ __device__ operator TypeCmp() const noexcept { return compValIdx; }
+
+  __device__ inline TypeCmp reduce(
+      cg::thread_block_tile<kWARP_SIZE> const& warp) {
+    return cg::reduce(warp, compValIdx, cg::greater<TypeCmp>{});
+  }
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template <int K_, bool Enable_>
+struct TopKIdx {
+  // by default, empty
+};
+
+template <int K_>
+struct TopKIdx<K_, true> {
+  static constexpr int K = K_;
+  int32_t val[K];
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+#define TOPK_SWAP(I, J)                                         \
+  {                                                             \
+    auto pairMin = min(topK[I].compValIdx, topK[J].compValIdx); \
+    auto pairMax = max(topK[I].compValIdx, topK[J].compValIdx); \
+    topK[I].compValIdx = pairMax;                               \
+    topK[J].compValIdx = pairMin;                               \
+  }
+
+template <int N, typename RedType>
+struct Sort;
+
+template <typename RedType>
+struct Sort<1, RedType> {
+  static __device__ void run(RedType* topK) {}
+};
+
+template <typename RedType>
+struct Sort<2, RedType> {
+  static __device__ void run(RedType* topK) { TOPK_SWAP(0, 1); }
+};
+
+template <typename RedType>
+struct Sort<3, RedType> {
+  static __device__ void run(RedType* topK) {
+    TOPK_SWAP(0, 1);
+    TOPK_SWAP(1, 2);
+    TOPK_SWAP(0, 1);
+  }
+};
+
+template <typename RedType>
+struct Sort<4, RedType> {
+  static __device__ void run(RedType* topK) {
+    TOPK_SWAP(0, 2);
+    TOPK_SWAP(1, 3);
+    TOPK_SWAP(0, 1);
+    TOPK_SWAP(2, 3);
+    TOPK_SWAP(1, 2);
+  }
+};
+
+template <int K, typename Type>
+__forceinline__ __device__ void reduceTopK(
+    cg::thread_block_tile<kWARP_SIZE> const& warp, Type (&out)[K],
+    int32_t (&outIdx)[K], Type value, int32_t idx, Type const minValue,
+    int actualK = K) {
+  static_assert(K > 0, "Top K must have K > 0");
+  static_assert(K < kWARP_SIZE, "Top K must have K < kWARP_SIZE");
+  using RedType = TopKRedType<Type>;
+  RedType topK{value, idx};
+  typename RedType::TypeCmp packedMax{};
+#pragma unroll
+  for (int kk = 0; kk < actualK; ++kk) {
+    topK =
+        kk > 0 && packedMax == topK.compValIdx ? RedType{minValue, idx} : topK;
+    // get the next largest value
+    packedMax = topK.reduce(warp);
+    RedType::unpack(out[kk], outIdx[kk], packedMax);
+  }
+};
+
+template <int K, typename Type, int N, bool IsSorted = false>
+__device__ void reduceTopKFunc(cg::thread_block_tile<kWARP_SIZE> const& warp,
+                               Type (&out)[K], int32_t (&outIdx)[K],
+                               Type (&value)[N], int32_t (&idx)[N],
+                               Type minValue, int actualK = K) {
+  static_assert(K > 0, "Top K must have K > 0");
+  static_assert(K < kWARP_SIZE, "Top K must have K < kWARP_SIZE");
+  static_assert(N > 0, "Top K must have N > 0");
+  static_assert(N < 5,
+                "Only support candidates number less than or equal to 128");
+  using RedType = TopKRedType<Type>;
+  RedType topK[N];
+#pragma unroll
+  for (int nn = 0; nn < N; ++nn) {
+    topK[nn] = RedType{value[nn], idx[nn]};
+  }
+
+  if constexpr (!IsSorted) {
+    Sort<N, RedType>::run(topK);
+  }
+  typename RedType::TypeCmp packedMax{};
+#pragma unroll
+  for (int kk = 0; kk < actualK; ++kk) {
+    bool update = kk > 0 && packedMax == topK[0].compValIdx;
+#pragma unroll
+    for (int nn = 0; nn < N; ++nn) {
+      topK[nn] = update && nn == N - 1 ? RedType{minValue, idx[nn]}
+                 : update              ? topK[nn + 1]
+                                       : topK[nn];
+    }
+    // get the next largest value
+    packedMax = topK[0].reduce(warp);
+    RedType::unpack(out[kk], outIdx[kk], packedMax);
+  }
+};
+
+template <int K, typename Type, int N>
+__forceinline__ __device__ void reduceTopK(
+    cg::thread_block_tile<kWARP_SIZE> const& warp, Type (&out)[K],
+    int32_t (&outIdx)[K], Type (&value)[N], int32_t (&idx)[N],
+    Type const minValue, int actualK = K) {
+  static_assert(K > 0, "Top K must have K > 0");
+  static_assert(K < kWARP_SIZE, "Top K must have K < kWARP_SIZE");
+  static_assert(N > 0, "Top K must have N > 0");
+  static_assert(
+      N <= 16,
+      "Only support candidates number less than or equal to 16*32=512");
+  static_assert(N <= 4 || N % 4 == 0,
+                "Only support candidates number is a multiple of 4*32=128 or "
+                "less than or equal to 4");
+  using RedType = TopKRedType<Type>;
+
+  if constexpr (N <= 4) {
+    reduceTopKFunc<K, Type, N>(warp, out, outIdx, value, idx, minValue,
+                               actualK);
+  } else {
+    constexpr int numLoops = N / 4;
+    constexpr int numResults = (numLoops * K - 1) / kWARP_SIZE + 1;
+
+    Type topKBufferValue[numResults];
+    int32_t topKBufferIdx[numResults];
+    int32_t laneIdx = threadIdx.x % kWARP_SIZE;
+
+    for (int ii = 0; ii < numResults; ++ii) {
+      topKBufferValue[ii] = minValue;
+      topKBufferIdx[ii] = ii * kWARP_SIZE - 1;
+    }
+    for (int loop = 0; loop < numLoops; ++loop) {
+      int start = loop * 4;
+      Type topKValue[K];
+      int32_t topKIdx[K];
+      Type inValue[4];
+      int32_t inIdx[4];
+      for (int i = 0; i < 4; ++i) {
+        inValue[i] = value[start + i];
+        inIdx[i] = idx[start + i];
+      }
+      reduceTopKFunc<K, Type, 4>(warp, topKValue, topKIdx, inValue, inIdx,
+                                 minValue, actualK);
+      int inOffset = laneIdx % K;
+      if (laneIdx >= loop * K && laneIdx < (loop + 1) * K) {
+        topKBufferValue[0] = topKValue[inOffset];
+        topKBufferIdx[0] = topKIdx[inOffset];
+      }
+      if (loop == numLoops - 1 && (laneIdx < (numLoops * K - kWARP_SIZE))) {
+        topKBufferValue[1] = topKValue[inOffset];
+        topKBufferIdx[1] = topKIdx[inOffset];
+      }
+    }
+
+    reduceTopKFunc<K, Type, numResults>(warp, out, outIdx, topKBufferValue,
+                                        topKBufferIdx, minValue, actualK);
+  }
+};
+
+#undef TOPK_SWAP
+
+}  // namespace reduce_topk
+}  // namespace moe
+}  // namespace vllm

csrc/moe/moe_align_sum_kernels.cu

@@ -172,7 +172,7 @@ __device__ void _moe_align_block_size(
     }
   }
 
-  // Fill remaining expert_ids with 0
+  // Fill remaining expert_ids with -1
   const size_t fill_start_idx =
       cumsum[cumsum_offset + num_experts] / block_size + threadIdx.x;
   for (size_t i = fill_start_idx; i < max_num_m_blocks; i += blockDim.x) {
@@ -265,7 +265,7 @@ __device__ void _moe_align_block_size_small_batch_expert(
     }
   }
 
-  // Fill remaining expert_ids with 0
+  // Fill remaining expert_ids with -1
   const size_t fill_start_idx = cumsum[num_experts] / block_size + tid;
   for (size_t i = fill_start_idx; i < max_num_m_blocks; i += stride) {
     expert_ids[expert_ids_offset + i] = inactive_expert_id;
@@ -332,7 +332,7 @@ __global__ void moe_align_block_size_kernel(
       topk_ids, sorted_token_ids, expert_ids, total_tokens_post_pad, expert_map,
       num_experts, padded_num_experts, experts_per_warp, block_size, numel,
       cumsum, max_num_tokens_padded, CEILDIV(max_num_tokens_padded, block_size),
-      0, 0, topk_num, nullptr, has_expert_map);
+      0, -1, topk_num, nullptr, has_expert_map);
 }
 
 template <typename scalar_t>
@@ -373,7 +373,7 @@ __global__ void moe_align_block_size_small_batch_expert_kernel(
   _moe_align_block_size_small_batch_expert<scalar_t, fill_threads>(
       topk_ids, sorted_token_ids, expert_ids, total_tokens_post_pad, expert_map,
       num_experts, block_size, numel, max_num_tokens_padded,
-      CEILDIV(max_num_tokens_padded, block_size), 0, 0, topk_num, nullptr,
+      CEILDIV(max_num_tokens_padded, block_size), -1, 0, topk_num, nullptr,
       has_expert_map);
 }
 

csrc/moe/moe_ops.h

@@ -55,4 +55,15 @@ bool moe_permute_unpermute_supported();
 
 void shuffle_rows(const torch::Tensor& input_tensor,
                   const torch::Tensor& dst2src_map,
-                  torch::Tensor& output_tensor);
+                  torch::Tensor& output_tensor);
+
+#ifndef USE_ROCM
+// DeepSeek V3 optimized router GEMM kernel for SM90+
+// Computes output = mat_a @ mat_b.T where:
+//   mat_a: [num_tokens, hidden_dim] in bf16
+//   mat_b: [num_experts, hidden_dim] in bf16
+//   output: [num_tokens, num_experts] in bf16 or fp32
+// Supports num_tokens in [1, 16], num_experts in {256, 384}, hidden_dim = 7168
+void dsv3_router_gemm(torch::Tensor& output, const torch::Tensor& mat_a,
+                      const torch::Tensor& mat_b);
+#endif

csrc/moe/torch_bindings.cpp

@@ -124,6 +124,10 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, m) {
       "routed_scaling_factor, Tensor bias, int scoring_func) -> (Tensor, "
       "Tensor)");
   m.impl("grouped_topk", torch::kCUDA, &grouped_topk);
+
+  // DeepSeek V3 optimized router GEMM for SM90+
+  m.def("dsv3_router_gemm(Tensor! output, Tensor mat_a, Tensor mat_b) -> ()");
+  // conditionally compiled so impl registration is in source file
 #endif
 }
 

csrc/ops.h

@@ -315,7 +315,9 @@ void silu_and_mul_scaled_fp4_experts_quant(
 void per_token_group_quant_fp8(const torch::Tensor& input,
                                torch::Tensor& output_q, torch::Tensor& output_s,
                                int64_t group_size, double eps, double fp8_min,
-                               double fp8_max, bool scale_ue8m0);
+                               double fp8_max, bool scale_ue8m0,
+                               bool dummy_is_scale_transposed,
+                               bool dummy_is_tma_aligned);
 
 void per_token_group_quant_int8(const torch::Tensor& input,
                                 torch::Tensor& output_q,
@@ -408,3 +410,8 @@ void qr_all_reduce(fptr_t _fa, torch::Tensor& inp, torch::Tensor& out,
                    int64_t quant_level, bool cast_bf2half = false);
 int64_t qr_max_size();
 #endif
+
+#ifndef USE_ROCM
+void dsv3_fused_a_gemm(torch::Tensor& output, torch::Tensor const& mat_a,
+                       torch::Tensor const& mat_b);
+#endif

csrc/quantization/fp4/activation_nvfp4_quant_fusion_kernels.cu

@@ -107,7 +107,9 @@ __global__ void __launch_bounds__(512, VLLM_BLOCKS_PER_SM(512))
               (uint64_t(out_val.hi) << 32) | uint64_t(out_val.lo);
           reinterpret_cast<uint64_t*>(out)[outOffset >> 1] = packed64;
         } else {
-          out[inOffset] = out_val;
+          int64_t outOffset =
+              rowIdx * (numCols / CVT_FP4_ELTS_PER_THREAD) + colIdx;
+          out[outOffset] = out_val;
         }
       }
     }
@@ -140,7 +142,7 @@ void silu_and_mul_nvfp4_quant_sm1xxa(torch::Tensor& output,  // [..., d]
   int const numBlocksPerSM =
       vllm_runtime_blocks_per_sm(static_cast<int>(block.x));
 
-  int sf_n_unpadded = int(n / CVT_FP4_SF_VEC_SIZE);
+  int sf_n_unpadded = int(n / CVT_FP4_ELTS_PER_THREAD);
 
   int grid_y = vllm::div_round_up(sf_n_unpadded, static_cast<int>(block.x));
   int grid_x = std::min(

csrc/quantization/fp4/nvfp4_quant_kernels.cu

@@ -109,7 +109,8 @@ __global__ void __launch_bounds__(512, VLLM_BLOCKS_PER_SM(512))
 template <class Type, bool UE8M0_SF = false>
 __global__ void __launch_bounds__(512, VLLM_BLOCKS_PER_SM(512))
     cvt_fp16_to_fp4_sf_major(int32_t numRows, int32_t numCols,
-                             int32_t sf_n_unpadded, Type const* __restrict__ in,
+                             int32_t sf_n_unpadded, int32_t num_packed_cols,
+                             Type const* __restrict__ in,
                              float const* __restrict__ SFScale,
                              uint32_t* __restrict__ out,
                              uint32_t* __restrict__ SFout) {
@@ -131,7 +132,7 @@ __global__ void __launch_bounds__(512, VLLM_BLOCKS_PER_SM(512))
   // Iterate over all rows and cols including padded ones -
   //  ensures we visit every single scale factor address to initialize it.
   for (int rowIdx = blockIdx.x; rowIdx < numRows; rowIdx += gridDim.x) {
-    if (colIdx < sf_n_unpadded) {
+    if (colIdx < num_packed_cols) {
       PackedVec in_vec;
       int64_t inOffset = rowIdx * (numCols / CVT_FP4_ELTS_PER_THREAD) + colIdx;
 
@@ -222,7 +223,8 @@ void scaled_fp4_quant_sm1xxa(torch::Tensor const& output,
           reinterpret_cast<uint32_t*>(sf_out));
     });
   } else {
-    int grid_y = vllm::div_round_up(sf_n_unpadded, static_cast<int>(block.x));
+    int num_packed_cols = n / CVT_FP4_ELTS_PER_THREAD;
+    int grid_y = vllm::div_round_up(num_packed_cols, static_cast<int>(block.x));
     int grid_x = std::min(
         m, std::max(1, (multiProcessorCount * numBlocksPerSM) / grid_y));
     dim3 grid(grid_x, grid_y);
@@ -232,8 +234,8 @@ void scaled_fp4_quant_sm1xxa(torch::Tensor const& output,
       auto input_ptr = static_cast<cuda_type const*>(input.data_ptr());
       // NOTE: We don't support e8m0 scales at this moment.
       vllm::cvt_fp16_to_fp4_sf_major<cuda_type, false>
-          <<<grid, block, 0, stream>>>(m, n, sf_n_unpadded, input_ptr,
-                                       input_sf_ptr,
+          <<<grid, block, 0, stream>>>(m, n, sf_n_unpadded, num_packed_cols,
+                                       input_ptr, input_sf_ptr,
                                        reinterpret_cast<uint32_t*>(output_ptr),
                                        reinterpret_cast<uint32_t*>(sf_out));
     });

csrc/quantization/fused_kernels/fused_layernorm_dynamic_per_token_quant.cu

@@ -97,7 +97,7 @@ __global__ void rms_norm_per_block_quant_kernel(
     scalar_t const* __restrict__ input,   // [..., hidden_size]
     scalar_t const* __restrict__ weight,  // [hidden_size]
     float const* scale_ub, float const var_epsilon, int32_t const hidden_size,
-    scalar_t* __restrict__ residual = nullptr) {
+    scalar_t* __restrict__ residual = nullptr, int64_t outer_scale_stride = 1) {
   float rms;
   // Compute RMS
   // Always able to vectorize due to constraints on hidden_size
@@ -108,7 +108,8 @@ __global__ void rms_norm_per_block_quant_kernel(
   // Always able to vectorize due to constraints on hidden_size and group_size
   vllm::vectorized::compute_dynamic_per_token_scales<
       scalar_t, scalar_out_t, has_residual, is_scale_transposed, group_size>(
-      nullptr, scales, input, weight, rms, scale_ub, hidden_size, residual);
+      nullptr, scales, input, weight, rms, scale_ub, hidden_size, residual,
+      outer_scale_stride);
 
   // RMS Norm + Quant
   // Always able to vectorize due to constraints on hidden_size
@@ -119,7 +120,8 @@ __global__ void rms_norm_per_block_quant_kernel(
   vllm::vectorized::norm_and_quant<
       scalar_t, scalar_out_t, std::is_same_v<scalar_out_t, int8_t>,
       has_residual, is_scale_transposed, group_size>(
-      out, input, weight, rms, scales, hidden_size, residual);
+      out, input, weight, rms, scales, hidden_size, residual,
+      outer_scale_stride);
 }
 
 }  // namespace vllm
@@ -225,7 +227,8 @@ void rms_norm_per_block_quant_dispatch(
                                                  : nullptr,
                             var_epsilon, hidden_size,
                             has_residual ? residual->data_ptr<scalar_in_t>()
-                                         : nullptr);
+                                         : nullptr,
+                            scales.stride(1));
                   });
             });
           });
@@ -257,6 +260,11 @@ void rms_norm_per_block_quant(torch::Tensor& out, torch::Tensor const& input,
   TORCH_CHECK(group_size == 128 || group_size == 64,
               "Unsupported group size: ", group_size);
 
+  if (scales.stride(1) > 1) {
+    TORCH_CHECK(is_scale_transposed,
+                "Outer scale stride must be 1 when scales are not transposed");
+  }
+
   rms_norm_per_block_quant_dispatch(out, input, weight, scales, group_size,
                                     var_epsilon, scale_ub, residual,
                                     is_scale_transposed);

csrc/quantization/fused_kernels/layernorm_utils.cuh

@@ -74,7 +74,7 @@ __device__ void compute_dynamic_per_token_scales(
     scalar_t const* __restrict__ input, scalar_t const* __restrict__ weight,
     float const rms, float const* __restrict__ scale_ub,
     int32_t const hidden_size, scalar_t const* __restrict__ residual = nullptr,
-    int32_t const group_size = 0) {
+    int32_t const group_size = 0, int64_t outer_scale_stride = 1) {
   float block_absmax_val_maybe = 0.0f;
   constexpr scalar_out_t qmax{quant_type_max_v<scalar_out_t>};
   __syncthreads();
@@ -133,7 +133,9 @@ __device__ void compute_dynamic_per_token_scales(
       scale = max(scale / qmax, min_scaling_factor<scalar_out_t>::val());
       // Global output store
       if constexpr (is_scale_transposed) {
-        all_token_scales[(threadIdx.x / threads_per_group) * gridDim.x +
+        int64_t const scale_rows = (gridDim.x + outer_scale_stride - 1) /
+                                   outer_scale_stride * outer_scale_stride;
+        all_token_scales[(threadIdx.x / threads_per_group) * scale_rows +
                          blockIdx.x] = scale;
       } else {
         all_token_scales[blockIdx.x * num_groups +
@@ -180,13 +182,11 @@ __device__ void compute_dynamic_per_token_scales(
 
 template <typename scalar_t, typename scalar_out_t, bool is_scale_inverted,
           bool has_residual = false, bool is_scale_transposed = false>
-__device__ void norm_and_quant(scalar_out_t* __restrict__ output,
-                               scalar_t const* __restrict__ input,
-                               scalar_t const* __restrict__ weight,
-                               float const rms, float* const scale,
-                               int32_t const hidden_size,
-                               scalar_t* __restrict__ residual = nullptr,
-                               int32_t const group_size = 0) {
+__device__ void norm_and_quant(
+    scalar_out_t* __restrict__ output, scalar_t const* __restrict__ input,
+    scalar_t const* __restrict__ weight, float const rms, float* const scale,
+    int32_t const hidden_size, scalar_t* __restrict__ residual = nullptr,
+    int32_t const group_size = 0, int64_t outer_scale_stride = 1) {
   int64_t const token_offset = blockIdx.x * static_cast<int64_t>(hidden_size);
 
   for (auto i = threadIdx.x; i < hidden_size; i += blockDim.x) {
@@ -202,7 +202,9 @@ __device__ void norm_and_quant(scalar_out_t* __restrict__ output,
     int64_t scale_idx = 0;
     if (group_size > 0) {
       if constexpr (is_scale_transposed) {
-        scale_idx = (i / group_size) * gridDim.x + blockIdx.x;
+        int64_t const scale_rows = (gridDim.x + outer_scale_stride - 1) /
+                                   outer_scale_stride * outer_scale_stride;
+        scale_idx = (i / group_size) * scale_rows + blockIdx.x;
       } else {
         scale_idx = blockIdx.x * (hidden_size / group_size) + i / group_size;
       }
@@ -286,8 +288,8 @@ __device__ void compute_dynamic_per_token_scales(
     float* __restrict__ token_scale, float* __restrict__ all_token_scales,
     scalar_t const* __restrict__ input, scalar_t const* __restrict__ weight,
     float const rms, float const* __restrict__ scale_ub,
-    int32_t const hidden_size,
-    scalar_t const* __restrict__ residual = nullptr) {
+    int32_t const hidden_size, scalar_t const* __restrict__ residual = nullptr,
+    int64_t outer_scale_stride = 1) {
   constexpr scalar_out_t qmax{quant_type_max_v<scalar_out_t>};
 
   const int VEC_SIZE = 4;
@@ -382,7 +384,9 @@ __device__ void compute_dynamic_per_token_scales(
       scale = max(scale / qmax, min_scaling_factor<scalar_out_t>::val());
       // Global output store
       if constexpr (is_scale_transposed) {
-        all_token_scales[(threadIdx.x / threads_per_group) * gridDim.x +
+        int64_t const scale_rows = (gridDim.x + outer_scale_stride - 1) /
+                                   outer_scale_stride * outer_scale_stride;
+        all_token_scales[(threadIdx.x / threads_per_group) * scale_rows +
                          blockIdx.x] = scale;
       } else {
         all_token_scales[blockIdx.x * num_groups +
@@ -463,7 +467,8 @@ __device__ void norm_and_quant(scalar_out_t* __restrict__ output,
                                scalar_t const* __restrict__ weight,
                                float const rms, float* const scale,
                                int32_t const hidden_size,
-                               scalar_t* __restrict__ residual = nullptr) {
+                               scalar_t* __restrict__ residual = nullptr,
+                               int64_t outer_scale_stride = 1) {
   int64_t const token_offset = blockIdx.x * static_cast<int64_t>(hidden_size);
 
   // Vectorized input/output/weight/residual to better utilize memory bandwidth.
@@ -516,7 +521,9 @@ __device__ void norm_and_quant(scalar_out_t* __restrict__ output,
       int64_t const num_groups = hidden_size / group_size;
       int64_t scale_idx = 0;
       if constexpr (is_scale_transposed) {
-        scale_idx = (i * VEC_SIZE / group_size) * gridDim.x + blockIdx.x;
+        int64_t const scale_rows = (gridDim.x + outer_scale_stride - 1) /
+                                   outer_scale_stride * outer_scale_stride;
+        scale_idx = (i * VEC_SIZE / group_size) * scale_rows + blockIdx.x;
       } else {
         scale_idx = blockIdx.x * num_groups + i * VEC_SIZE / group_size;
       }

csrc/quantization/w8a8/cutlass/c3x/scaled_mm_sm120_fp8_dispatch.cuh

@@ -12,6 +12,68 @@ namespace vllm {
 
 using c3x::cutlass_gemm_caller;
 
+// Custom wrapper to allow specifying EpilogueTile for small M
+template <typename ElementAB_, typename ElementD_,
+          template <typename, typename, typename> typename Epilogue_,
+          typename TileShape, typename ClusterShape, typename KernelSchedule,
+          typename EpilogueSchedule, typename EpilogueTile>
+struct cutlass_3x_gemm_sm120_custom {
+  using ElementAB = ElementAB_;
+  using LayoutA = cutlass::layout::RowMajor;
+  static constexpr int AlignmentA =
+      128 / cutlass::sizeof_bits<ElementAB>::value;
+
+  using LayoutB = cutlass::layout::ColumnMajor;
+  static constexpr int AlignmentB =
+      128 / cutlass::sizeof_bits<ElementAB>::value;
+
+  using ElementC = void;
+  using LayoutC = cutlass::layout::RowMajor;
+  static constexpr int AlignmentC =
+      128 / cutlass::sizeof_bits<ElementD_>::value;
+
+  using ElementD = ElementD_;
+  using LayoutD = cutlass::layout::RowMajor;
+  static constexpr int AlignmentD = AlignmentC;
+
+  using ElementAcc =
+      typename std::conditional<std::is_same_v<ElementAB, int8_t>, int32_t,
+                                float>::type;
+  using Epilogue = Epilogue_<ElementAcc, ElementD, TileShape>;
+
+  // MMA type
+  using ElementAccumulator = float;
+
+  // Epilogue types
+  using ElementBias = cutlass::half_t;
+  using ElementCompute = float;
+  using ElementAux = ElementD;
+  using LayoutAux = LayoutD;
+  using ElementAmax = float;
+
+  using EVTCompute = typename Epilogue::EVTCompute;
+
+  using CollectiveEpilogue =
+      typename cutlass::epilogue::collective::CollectiveBuilder<
+          cutlass::arch::Sm120, cutlass::arch::OpClassTensorOp, TileShape,
+          ClusterShape, EpilogueTile,  // Use custom EpilogueTile
+          ElementAccumulator, ElementCompute, ElementC, LayoutC, AlignmentC,
+          ElementD, LayoutD, AlignmentD, EpilogueSchedule,
+          EVTCompute>::CollectiveOp;
+
+  using CollectiveMainloop =
+      typename cutlass::gemm::collective::CollectiveBuilder<
+          cutlass::arch::Sm120, cutlass::arch::OpClassTensorOp, ElementAB,
+          LayoutA, AlignmentA, ElementAB, LayoutB, AlignmentB,
+          ElementAccumulator, TileShape, ClusterShape,
+          cutlass::gemm::collective::StageCountAutoCarveout<static_cast<int>(
+              sizeof(typename CollectiveEpilogue::SharedStorage))>,
+          KernelSchedule, void>::CollectiveOp;
+
+  using GemmKernel = enable_sm120_only<cutlass::gemm::kernel::GemmUniversal<
+      Shape<int, int, int, int>, CollectiveMainloop, CollectiveEpilogue, void>>;
+};
+
 template <typename InType, typename OutType,
           template <typename, typename, typename> typename Epilogue>
 struct sm120_fp8_config_default {
@@ -25,6 +87,54 @@ struct sm120_fp8_config_default {
                             KernelSchedule, EpilogueSchedule>;
 };
 
+template <typename InType, typename OutType,
+          template <typename, typename, typename> typename Epilogue>
+struct sm120_fp8_config_M64 {
+  static_assert(std::is_same<InType, cutlass::float_e4m3_t>());
+  // SM120 Cooperative kernel requires Tile M >= 128.
+  // For M=64 tile, we use Pingpong schedule which is more flexible with small
+  // tiles.
+  using KernelSchedule = cutlass::gemm::KernelTmaWarpSpecializedPingpong;
+  using EpilogueSchedule = cutlass::epilogue::collective::EpilogueScheduleAuto;
+  using TileShape = Shape<_64, _64, _128>;
+  // CUTLASS 3.x on SM120 currently restricts programmatic multicast (Cluster >
+  // 1) for certain schedules/types. Reverting to 1x1x1 to ensure compilation.
+  using ClusterShape = Shape<_1, _1, _1>;
+  using Cutlass3xGemm =
+      cutlass_3x_gemm_sm120<InType, OutType, Epilogue, TileShape, ClusterShape,
+                            KernelSchedule, EpilogueSchedule>;
+};
+
+template <typename InType, typename OutType,
+          template <typename, typename, typename> typename Epilogue>
+struct sm120_fp8_config_M32 {
+  static_assert(std::is_same<InType, cutlass::float_e4m3_t>());
+  using KernelSchedule = cutlass::gemm::KernelTmaWarpSpecializedPingpong;
+  using EpilogueSchedule = cutlass::epilogue::collective::EpilogueScheduleAuto;
+  using TileShape = Shape<_32, _64, _128>;
+  using ClusterShape = Shape<_1, _1, _1>;
+  // Use custom gemm to specify EpilogueTile M=32
+  using Cutlass3xGemm =
+      cutlass_3x_gemm_sm120_custom<InType, OutType, Epilogue, TileShape,
+                                   ClusterShape, KernelSchedule,
+                                   EpilogueSchedule, Shape<_32, _32>>;
+};
+
+template <typename InType, typename OutType,
+          template <typename, typename, typename> typename Epilogue>
+struct sm120_fp8_config_M16 {
+  static_assert(std::is_same<InType, cutlass::float_e4m3_t>());
+  using KernelSchedule = cutlass::gemm::KernelTmaWarpSpecializedPingpong;
+  using EpilogueSchedule = cutlass::epilogue::collective::EpilogueScheduleAuto;
+  using TileShape = Shape<_16, _64, _128>;
+  using ClusterShape = Shape<_1, _1, _1>;
+  // Use custom gemm to specify EpilogueTile M=16
+  using Cutlass3xGemm =
+      cutlass_3x_gemm_sm120_custom<InType, OutType, Epilogue, TileShape,
+                                   ClusterShape, KernelSchedule,
+                                   EpilogueSchedule, Shape<_16, _32>>;
+};
+
 template <typename InType, typename OutType,
           template <typename, typename, typename> typename Epilogue,
           typename... EpilogueArgs>
@@ -36,6 +146,28 @@ inline void cutlass_gemm_sm120_fp8_dispatch(torch::Tensor& out,
   TORCH_CHECK(a.dtype() == torch::kFloat8_e4m3fn);
   TORCH_CHECK(b.dtype() == torch::kFloat8_e4m3fn);
 
+  int M = a.size(0);
+
+  if (M <= 16) {
+    using Cutlass3xGemmM16 =
+        typename sm120_fp8_config_M16<InType, OutType, Epilogue>::Cutlass3xGemm;
+    return cutlass_gemm_caller<Cutlass3xGemmM16>(
+        out, a, b, std::forward<EpilogueArgs>(args)...);
+  }
+  if (M <= 32) {
+    using Cutlass3xGemmM32 =
+        typename sm120_fp8_config_M32<InType, OutType, Epilogue>::Cutlass3xGemm;
+    return cutlass_gemm_caller<Cutlass3xGemmM32>(
+        out, a, b, std::forward<EpilogueArgs>(args)...);
+  }
+
+  if (M <= 256) {
+    using Cutlass3xGemmM64 =
+        typename sm120_fp8_config_M64<InType, OutType, Epilogue>::Cutlass3xGemm;
+    return cutlass_gemm_caller<Cutlass3xGemmM64>(
+        out, a, b, std::forward<EpilogueArgs>(args)...);
+  }
+
   using Cutlass3xGemmDefault =
       typename sm120_fp8_config_default<InType, OutType,
                                         Epilogue>::Cutlass3xGemm;
@@ -64,4 +196,4 @@ void cutlass_scaled_mm_sm120_fp8_epilogue(torch::Tensor& out,
   }
 }
 
-}  // namespace vllm
+}  // namespace vllm

csrc/quantization/w8a8/fp8/per_token_group_quant.cu

@@ -379,7 +379,9 @@ void per_token_group_quant_8bit_packed(const torch::Tensor& input,
 void per_token_group_quant_fp8(const torch::Tensor& input,
                                torch::Tensor& output_q, torch::Tensor& output_s,
                                int64_t group_size, double eps, double fp8_min,
-                               double fp8_max, bool scale_ue8m0) {
+                               double fp8_max, bool scale_ue8m0,
+                               bool dummy_is_scale_transposed = false,
+                               bool dummy_is_tma_aligned = false) {
   per_token_group_quant_8bit(input, output_q, output_s, group_size, eps,
                              fp8_min, fp8_max, scale_ue8m0);
 }

csrc/rocm/skinny_gemms.cu

@@ -1902,7 +1902,7 @@ __global__ void __launch_bounds__(WvPrGrp* THRDS)
   float sB = *s_B;
 
   while (m < M) {
-    floatx16 sum[N][YTILE] = {};
+    scalar8 sum[N][YTILE] = {};
     for (uint32_t k1 = 0; k1 < K; k1 += THRDS * A_CHUNK * UNRL) {
       bigType bigA[N][UNRL] = {};
       bigType bigB[YTILE][UNRL];
@@ -1936,7 +1936,7 @@ __global__ void __launch_bounds__(WvPrGrp* THRDS)
         for (uint32_t n = 0; n < N; n++) {
           for (int i = 0; i < A_CHUNK; i += 8) {
             for (int y = 0; y < YTILE; ++y) {
-              sum[n][y] = __builtin_amdgcn_mfma_f32_32x32x16_fp8_fp8(
+              sum[n][y] = __builtin_amdgcn_mfma_f32_16x16x32_fp8_fp8(
                   bigA[n][k2].l[i / 8], bigB[y][k2].l[i / 8], sum[n][y], 0, 0,
                   0);
             }
@@ -1949,31 +1949,15 @@ __global__ void __launch_bounds__(WvPrGrp* THRDS)
     for (int n = 0; n < N; n++) {
       for (int y = 0; y < YTILE; y++) {
         float accm0 = sum[n][y][0];
-        float accm16 = sum[n][y][8];
         accm0 += __builtin_amdgcn_mov_dpp(sum[n][y][1], 0x101, 0xf, 0xf,
                                           1);  // row_shl1
-        accm16 += __builtin_amdgcn_mov_dpp(sum[n][y][9], 0x101, 0xf, 0xf, 1);
         accm0 += __builtin_amdgcn_mov_dpp(sum[n][y][2], 0x102, 0xf, 0xf,
                                           1);  // row_shl2
-        accm16 += __builtin_amdgcn_mov_dpp(sum[n][y][10], 0x102, 0xf, 0xf, 1);
         accm0 += __builtin_amdgcn_mov_dpp(sum[n][y][3], 0x103, 0xf, 0xf,
                                           1);  // row_shl3
-        accm16 += __builtin_amdgcn_mov_dpp(sum[n][y][11], 0x103, 0xf, 0xf, 1);
-        accm0 += __builtin_amdgcn_mov_dpp(sum[n][y][4], 0x108, 0xf, 0xf,
-                                          1);  // row_shl8
-        accm16 += __builtin_amdgcn_mov_dpp(sum[n][y][12], 0x108, 0xf, 0xf, 1);
-        accm0 += __builtin_amdgcn_mov_dpp(sum[n][y][5], 0x109, 0xf, 0xf,
-                                          1);  // row_shl9
-        accm16 += __builtin_amdgcn_mov_dpp(sum[n][y][13], 0x109, 0xf, 0xf, 1);
-        accm0 += __builtin_amdgcn_mov_dpp(sum[n][y][6], 0x10a, 0xf, 0xf,
-                                          1);  // row_shl10
-        accm16 += __builtin_amdgcn_mov_dpp(sum[n][y][14], 0x10a, 0xf, 0xf, 1);
-        accm0 += __builtin_amdgcn_mov_dpp(sum[n][y][7], 0x10b, 0xf, 0xf,
-                                          1);  // row_shl11
-        accm16 += __builtin_amdgcn_mov_dpp(sum[n][y][15], 0x10b, 0xf, 0xf, 1);
-        accm0 += __shfl(accm0, 36);
-        accm16 += __shfl(accm16, 52);
-        sum[n][y][0] = accm0 + __shfl(accm16, 16);
+        accm0 += __shfl_down(accm0, 20);
+        accm0 += __shfl_down(accm0, 40);
+        sum[n][y][0] = accm0;
       }
     }
 
@@ -2064,7 +2048,7 @@ __global__ void __launch_bounds__(WvPrGrp* THRDS)
   float sB = *s_B;
 
   while (m < M) {
-    floatx16 sum[N][YTILE] = {};
+    scalar8 sum[N][YTILE] = {};
     for (uint32_t k1 = 0; k1 < K; k1 += THRDS * A_CHUNK * UNRL) {
       bigType bigA[N][UNRL] = {};
       bigType bigB[YTILE][UNRL];
@@ -2100,7 +2084,7 @@ __global__ void __launch_bounds__(WvPrGrp* THRDS)
         for (uint32_t n = 0; n < N; n++) {
           for (int i = 0; i < A_CHUNK; i += 8) {
             for (int y = 0; y < YTILE; ++y) {
-              sum[n][y] = __builtin_amdgcn_mfma_f32_32x32x16_fp8_fp8(
+              sum[n][y] = __builtin_amdgcn_mfma_f32_16x16x32_fp8_fp8(
                   bigA[n][k2].l[i / 8], bigB[y][k2].l[i / 8], sum[n][y], 0, 0,
                   0);
             }
@@ -2113,31 +2097,15 @@ __global__ void __launch_bounds__(WvPrGrp* THRDS)
     for (int n = 0; n < N; n++) {
       for (int y = 0; y < YTILE; y++) {
         float accm0 = sum[n][y][0];
-        float accm16 = sum[n][y][8];
         accm0 += __builtin_amdgcn_mov_dpp(sum[n][y][1], 0x101, 0xf, 0xf,
                                           1);  // row_shl1
-        accm16 += __builtin_amdgcn_mov_dpp(sum[n][y][9], 0x101, 0xf, 0xf, 1);
         accm0 += __builtin_amdgcn_mov_dpp(sum[n][y][2], 0x102, 0xf, 0xf,
                                           1);  // row_shl2
-        accm16 += __builtin_amdgcn_mov_dpp(sum[n][y][10], 0x102, 0xf, 0xf, 1);
         accm0 += __builtin_amdgcn_mov_dpp(sum[n][y][3], 0x103, 0xf, 0xf,
                                           1);  // row_shl3
-        accm16 += __builtin_amdgcn_mov_dpp(sum[n][y][11], 0x103, 0xf, 0xf, 1);
-        accm0 += __builtin_amdgcn_mov_dpp(sum[n][y][4], 0x108, 0xf, 0xf,
-                                          1);  // row_shl8
-        accm16 += __builtin_amdgcn_mov_dpp(sum[n][y][12], 0x108, 0xf, 0xf, 1);
-        accm0 += __builtin_amdgcn_mov_dpp(sum[n][y][5], 0x109, 0xf, 0xf,
-                                          1);  // row_shl9
-        accm16 += __builtin_amdgcn_mov_dpp(sum[n][y][13], 0x109, 0xf, 0xf, 1);
-        accm0 += __builtin_amdgcn_mov_dpp(sum[n][y][6], 0x10a, 0xf, 0xf,
-                                          1);  // row_shl10
-        accm16 += __builtin_amdgcn_mov_dpp(sum[n][y][14], 0x10a, 0xf, 0xf, 1);
-        accm0 += __builtin_amdgcn_mov_dpp(sum[n][y][7], 0x10b, 0xf, 0xf,
-                                          1);  // row_shl11
-        accm16 += __builtin_amdgcn_mov_dpp(sum[n][y][15], 0x10b, 0xf, 0xf, 1);
-        accm0 += __shfl(accm0, 36);
-        accm16 += __shfl(accm16, 52);
-        sum[n][y][0] = accm0 + __shfl(accm16, 16);
+        accm0 += __shfl_down(accm0, 20);
+        accm0 += __shfl_down(accm0, 40);
+        sum[n][y][0] = accm0;
       }
     }
 
@@ -2242,16 +2210,16 @@ void wvSplitKQ(const at::Tensor& in_b, const at::Tensor& in_a,
                           : nullptr;
       switch (N_in) {
         case 1:
-          WVSPLITKQ(12, 2, 2, 2, 2, 1)
+          WVSPLITKQ(16, 2, 2, 2, 2, 1)
           break;
         case 2:
-          WVSPLITKQ(12, 2, 2, 2, 2, 2)
+          WVSPLITKQ(16, 2, 2, 2, 2, 2)
           break;
         case 3:
-          WVSPLITKQ(8, 2, 2, 1, 1, 3)
+          WVSPLITKQ(16, 2, 2, 2, 2, 3)
           break;
         case 4:
-          WVSPLITKQ(4, 2, 2, 1, 1, 4)
+          WVSPLITKQ(16, 2, 2, 2, 2, 4)
           break;
         default:
           throw std::runtime_error(

csrc/torch_bindings.cpp

@@ -239,6 +239,11 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
 
   // Quantization ops
 #ifndef USE_ROCM
+  // DeepSeek V3 fused A GEMM (SM 9.0+, bf16 only, 1-16 tokens).
+  ops.def(
+      "dsv3_fused_a_gemm(Tensor! output, Tensor mat_a, Tensor mat_b) -> ()");
+  // conditionally compiled so impl registration is in source file
+
   // Quantized GEMM for AWQ.
   ops.def(
       "awq_gemm(Tensor _in_feats, Tensor _kernel, Tensor _scaling_factors, "
@@ -643,11 +648,13 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
 
 #ifndef USE_ROCM
   // Compute per-token-group FP8 quantized tensor and scaling factor.
+  // The dummy arguments are here so we can correctly fuse with RMSNorm.
   ops.def(
       "per_token_group_fp8_quant(Tensor input, Tensor! output_q, Tensor! "
       "output_s, "
       "int group_size, float eps, float fp8_min, float fp8_max, bool "
-      "scale_ue8m0) -> ()");
+      "scale_ue8m0, bool dummy_is_scale_transposed, bool dummy_is_tma_aligned "
+      ") -> ()");
   ops.impl("per_token_group_fp8_quant", torch::kCUDA,
            &per_token_group_quant_fp8);
 

docker/Dockerfile

@@ -22,12 +22,7 @@
 #   docker buildx bake -f docker/docker-bake.hcl -f docker/versions.json
 # =============================================================================
 
-ARG CUDA_VERSION=12.8.1
-# BUILDER_CUDA_VERSION controls the CUDA toolkit used to compile csrc/ and
-# extensions (DeepGEMM, EP kernels).  It can differ from CUDA_VERSION, which
-# is the CUDA version shipped in the final runtime image and used to select
-# the matching PyTorch wheel.
-ARG BUILDER_CUDA_VERSION=12.9.1
+ARG CUDA_VERSION=12.9.1
 ARG PYTHON_VERSION=3.12
 
 # By parameterizing the base images, we allow third-party to use their own
@@ -41,7 +36,7 @@ ARG PYTHON_VERSION=3.12
 # compatibility with other Linux OSes. The main reason for this is that the
 # glibc version is baked into the distro, and binaries built with one glibc
 # version are not backwards compatible with OSes that use an earlier version.
-ARG BUILD_BASE_IMAGE=nvidia/cuda:${BUILDER_CUDA_VERSION}-devel-ubuntu20.04
+ARG BUILD_BASE_IMAGE=nvidia/cuda:${CUDA_VERSION}-devel-ubuntu20.04
 # Using cuda base image with minimal dependencies necessary for JIT compilation (FlashInfer, DeepGEMM, EP kernels)
 ARG FINAL_BASE_IMAGE=nvidia/cuda:${CUDA_VERSION}-base-ubuntu22.04
 
@@ -97,13 +92,8 @@ ARG INSTALL_KV_CONNECTORS=false
 FROM ${BUILD_BASE_IMAGE} AS base
 
 ARG CUDA_VERSION
-ARG BUILDER_CUDA_VERSION
 ARG PYTHON_VERSION
 
-# Override the CUDA_VERSION env var inherited from the nvidia base image
-# (which equals BUILDER_CUDA_VERSION) so that $CUDA_VERSION in RUN commands
-# resolves to the runtime CUDA version used for PyTorch wheel selection.
-ENV CUDA_VERSION=${CUDA_VERSION}
 ENV DEBIAN_FRONTEND=noninteractive
 
 # Install system dependencies including build tools
@@ -142,8 +132,10 @@ ENV UV_LINK_MODE=copy
 # Verify GCC version
 RUN gcc --version
 
-# Ensure CUDA compatibility library is loaded
-RUN echo "/usr/local/cuda-$(echo "$BUILDER_CUDA_VERSION" | cut -d. -f1,2)/compat/" > /etc/ld.so.conf.d/cuda-compat.conf && ldconfig
+# Enable CUDA forward compatibility by setting '-e VLLM_ENABLE_CUDA_COMPATIBILITY=1'
+# Only needed for datacenter/professional GPUs with older drivers.
+# See: https://docs.nvidia.com/deploy/cuda-compatibility/
+ENV VLLM_ENABLE_CUDA_COMPATIBILITY=0
 
 # ============================================================
 # SLOW-CHANGING DEPENDENCIES BELOW
@@ -319,7 +311,7 @@ RUN --mount=type=cache,target=/root/.cache/ccache \
 # Build DeepGEMM, pplx-kernels, DeepEP - runs in PARALLEL with csrc-build
 # This stage is independent and doesn't affect csrc cache
 FROM base AS extensions-build
-ARG BUILDER_CUDA_VERSION
+ARG CUDA_VERSION
 
 # This timeout (in seconds) is necessary when installing some dependencies via uv since it's likely to time out
 ENV UV_HTTP_TIMEOUT=500
@@ -335,7 +327,7 @@ COPY tools/install_deepgemm.sh /tmp/install_deepgemm.sh
 RUN --mount=type=cache,target=/root/.cache/uv \
     mkdir -p /tmp/deepgemm/dist && \
     VLLM_DOCKER_BUILD_CONTEXT=1 TORCH_CUDA_ARCH_LIST="9.0a 10.0a" /tmp/install_deepgemm.sh \
-        --cuda-version "${BUILDER_CUDA_VERSION}" \
+        --cuda-version "${CUDA_VERSION}" \
         ${DEEPGEMM_GIT_REF:+--ref "$DEEPGEMM_GIT_REF"} \
         --wheel-dir /tmp/deepgemm/dist || \
     echo "DeepGEMM build skipped (CUDA version requirement not met)"
@@ -570,8 +562,10 @@ ENV UV_HTTP_TIMEOUT=500
 ENV UV_INDEX_STRATEGY="unsafe-best-match"
 ENV UV_LINK_MODE=copy
 
-# Ensure CUDA compatibility library is loaded
-RUN echo "/usr/local/cuda-$(echo "$CUDA_VERSION" | cut -d. -f1,2)/compat/" > /etc/ld.so.conf.d/cuda-compat.conf && ldconfig
+# Enable CUDA forward compatibility by setting '-e VLLM_ENABLE_CUDA_COMPATIBILITY=1'
+# Only needed for datacenter/professional GPUs with older drivers.
+# See: https://docs.nvidia.com/deploy/cuda-compatibility/
+ENV VLLM_ENABLE_CUDA_COMPATIBILITY=0
 
 # ============================================================
 # SLOW-CHANGING DEPENDENCIES BELOW
@@ -592,7 +586,7 @@ RUN --mount=type=cache,target=/root/.cache/uv \
 # This is ~1.1GB and only changes when FlashInfer version bumps
 # https://docs.flashinfer.ai/installation.html
 # From versions.json: .flashinfer.version
-ARG FLASHINFER_VERSION=0.6.3
+ARG FLASHINFER_VERSION=0.6.4
 RUN --mount=type=cache,target=/root/.cache/uv \
     uv pip install --system flashinfer-cubin==${FLASHINFER_VERSION} \
     && uv pip install --system flashinfer-jit-cache==${FLASHINFER_VERSION} \

docker/Dockerfile.nightly_torch

@@ -217,13 +217,13 @@ RUN pip install setuptools==75.6.0 packaging==23.2 ninja==1.11.1.3 build==1.2.2.
 
 
 # build flashinfer for torch nightly from source around 10 mins
-# release version: v0.6.3
+# release version: v0.6.4
 # todo(elainewy): cache flashinfer build result for faster build
 ENV CCACHE_DIR=/root/.cache/ccache
 RUN --mount=type=cache,target=/root/.cache/ccache \
     --mount=type=cache,target=/root/.cache/uv \
     echo "git clone flashinfer..." \
-    && git clone --depth 1 --branch v0.6.3 --recursive https://github.com/flashinfer-ai/flashinfer.git \
+    && git clone --depth 1 --branch v0.6.4 --recursive https://github.com/flashinfer-ai/flashinfer.git \
     && cd flashinfer \
     && git submodule update --init --recursive \
     && echo "finish git clone flashinfer..." \

docker/Dockerfile.rocm

@@ -305,6 +305,14 @@ RUN --mount=type=bind,from=export_vllm,src=/,target=/install \
 RUN --mount=type=bind,from=build_rixl,src=/app/install,target=/rixl_install \
     uv pip install --system /rixl_install/*.whl
 
+# RIXL/MoRIIO runtime dependencies (RDMA userspace libraries)
+RUN apt-get update -q -y && apt-get install -q -y \
+    librdmacm1 \
+    libibverbs1 \
+    ibverbs-providers \
+    ibverbs-utils \
+    && rm -rf /var/lib/apt/lists/*
+
 WORKDIR /vllm-workspace
 ARG COMMON_WORKDIR
 COPY --from=build_vllm ${COMMON_WORKDIR}/vllm /vllm-workspace
@@ -330,6 +338,11 @@ RUN bash /tmp/install_torchcodec.sh \
 # Copy in the v1 package (for python-only install test group)
 COPY --from=export_vllm /vllm_v1 /usr/local/lib/python${PYTHON_VERSION}/dist-packages/vllm/v1
 
+# Set MIOPEN ENVS to resolve performance regressions in MIOpen 3D convolution kernel
+# See: https://github.com/pytorch/pytorch/issues/169857
+ENV MIOPEN_DEBUG_CONV_DIRECT=0
+ENV MIOPEN_DEBUG_CONV_GEMM=0
+
 # Source code is used in the `python_only_compile.sh` test
 # We hide it inside `src/` so that this source code
 # will not be imported by other tests

docker/versions.json

@@ -2,9 +2,6 @@
   "_comment": "Auto-generated from Dockerfile ARGs. Do not edit manually. Run: python tools/generate_versions_json.py",
   "variable": {
     "CUDA_VERSION": {
-      "default": "12.8.1"
-    },
-    "BUILDER_CUDA_VERSION": {
       "default": "12.9.1"
     },
     "PYTHON_VERSION": {
@@ -14,7 +11,7 @@
       "default": "nvidia/cuda:12.9.1-devel-ubuntu20.04"
     },
     "FINAL_BASE_IMAGE": {
-      "default": "nvidia/cuda:12.8.1-base-ubuntu22.04"
+      "default": "nvidia/cuda:12.9.1-base-ubuntu22.04"
     },
     "GET_PIP_URL": {
       "default": "https://bootstrap.pypa.io/get-pip.py"
@@ -71,7 +68,7 @@
       "default": "true"
     },
     "FLASHINFER_VERSION": {
-      "default": "0.6.3"
+      "default": "0.6.4"
     },
     "GDRCOPY_CUDA_VERSION": {
       "default": "12.8"

docs/benchmarking/cli.md

@@ -4,6 +4,11 @@ This section guides you through running benchmark tests with the extensive datas
 
 It's a living document, updated as new features and datasets become available.
 
+!!! tip
+    The benchmarks described on this page are mainly for evaluating specific vLLM features as well as regression testing.
+
+    For benchmarking production vLLM servers, we recommend [GuideLLM](https://github.com/vllm-project/guidellm), an established performance benchmarking framework with live progress updates and automatic report generation. It is also more flexible than `vllm bench serve` in terms of dataset loading, request formatting, and workload patterns.
+
 ## Dataset Overview
 
 <style>

docs/benchmarking/sweeps.md

@@ -1,10 +1,15 @@
 # Parameter Sweeps
 
+`vllm bench sweep` is a suite of commands designed to run benchmarks across multiple configurations and compare them by visualizing the results.
+
 ## Online Benchmark
 
 ### Basic
 
-`vllm bench sweep serve` automatically starts `vllm serve` and runs `vllm bench serve` to evaluate vLLM over multiple configurations.
+`vllm bench sweep serve` starts `vllm serve` and iteratively runs `vllm bench serve` for each server configuration.
+
+!!! tip
+    If you only need to run benchmarks for a single server configuration, consider using [GuideLLM](https://github.com/vllm-project/guidellm), an established performance benchmarking framework with live progress updates and automatic report generation. It is also more flexible than `vllm bench serve` in terms of dataset loading, request formatting, and workload patterns.
 
 Follow these steps to run the script:
 
@@ -50,14 +55,17 @@ Follow these steps to run the script:
     ```json
     [
         {
+            "_benchmark_name": "scenario_A",
             "random_input_len": 128,
             "random_output_len": 32
         },
         {
+            "_benchmark_name": "scenario_B",
             "random_input_len": 256,
             "random_output_len": 64
         },
         {
+            "_benchmark_name": "scenario_C",
             "random_input_len": 512,
             "random_output_len": 128
         }
@@ -77,6 +85,8 @@ vllm bench sweep serve \
     -o benchmarks/results
 ```
 
+By default, each parameter combination is benchmarked 3 times to make the results more reliable. You can adjust the number of runs by setting `--num-runs`.
+
 !!! important
     If both `--serve-params` and `--bench-params` are passed, the script will iterate over the Cartesian product between them.
     You can use `--dry-run` to preview the commands to be run.
@@ -86,60 +96,40 @@ vllm bench sweep serve \
     In case you are using a custom `--serve-cmd`, you can override the commands used for resetting the state by setting `--after-bench-cmd`.
 
 !!! note
-    By default, each parameter combination is run 3 times to make the results more reliable. You can adjust the number of runs by setting `--num-runs`.
+    You should set `_benchmark_name` to provide a human-readable name for parameter combinations involving many variables.
+    This becomes mandatory if the file name would otherwise exceed the maximum path length allowed by the filesystem.
 
 !!! tip
-    You can use the `--resume` option to continue the parameter sweep if one of the runs failed.
-  
-### SLA auto-tuner
+    You can use the `--resume` option to continue the parameter sweep if an unexpected error occurs, e.g., timeout when connecting to HF Hub.
 
-`vllm bench sweep serve_sla` is a wrapper over `vllm bench sweep serve` that tunes either the request rate or concurrency (choose using `--sla-variable`) in order to satisfy the SLA constraints given by `--sla-params`.
+### SLA Scanner
 
-For example, to ensure E2E latency within different target values for 99% of requests:
-
-```json
-[
-    {
-        "p99_e2el_ms": "<=200"
-    },
-    {
-        "p99_e2el_ms": "<=500"
-    },
-    {
-        "p99_e2el_ms": "<=1000"
-    },
-    {
-        "p99_e2el_ms": "<=2000"
-    }
-]
-```
+`vllm bench sweep serve_sla` is a variant of `vllm bench sweep serve` that scans through values of request rate or concurrency (choose using `--sla-variable`) in order to find the tradeoff between latency and throughput. The results can then be [visualized](#visualization) to determine the feasible SLAs.
 
 Example command:
 
 ```bash
 vllm bench sweep serve_sla \
     --serve-cmd 'vllm serve meta-llama/Llama-2-7b-chat-hf' \
-    --bench-cmd 'vllm bench serve --model meta-llama/Llama-2-7b-chat-hf --backend vllm --endpoint /v1/completions --dataset-name sharegpt --dataset-path benchmarks/ShareGPT_V3_unfiltered_cleaned_split.json' \
+    --bench-cmd 'vllm bench serve --model meta-llama/Llama-2-7b-chat-hf --backend vllm --endpoint /v1/completions --dataset-name sharegpt --dataset-path benchmarks/ShareGPT_V3_unfiltered_cleaned_split.json --num-prompts 100' \
     --serve-params benchmarks/serve_hparams.json \
-    --bench-params benchmarks/bench_hparams.json \
-    --sla-params benchmarks/sla_hparams.json \
-    --sla-variable max_concurrency \
+    --bench-params benchmarks/bench_hparams.json
     -o benchmarks/results
 ```
 
-The algorithm for adjusting the SLA variable is as follows:
+The algorithm for scanning through different values of `sla_variable` can be summarized as follows:
 
-1. Run the benchmark once with maximum possible QPS, and once with minimum possible QPS. For each run, calculate the distance of the SLA metrics from their targets, resulting in data points of QPS vs SLA distance.
-2. Perform spline interpolation between the data points to estimate the QPS that results in zero SLA distance.
-3. Run the benchmark with the estimated QPS and add the resulting data point to the history.
-4. Repeat Steps 2 and 3 until the maximum QPS that passes SLA and the minimum QPS that fails SLA in the history are close enough to each other.
+1. Run the benchmark once with `sla_variable = 1` to simulate serial inference. This results in the lowest possible latency and throughput.
+2. Run the benchmark once with `sla_variable = num_prompts` to simulate batch inference over the whole dataset. This results in the highest possible latency and throughput.
+3. Estimate the maximum value of `sla_variable` that can be supported by the server without oversaturating it.
+4. Run the benchmark over intermediate values of `sla_variable` uniformly using the remaining iterations.
 
-!!! important
-    SLA tuning is applied over each combination of `--serve-params`, `--bench-params`, and `--sla-params`.
+You can override the number of iterations in the algorithm by setting `--sla-iters`.
 
-    For a given combination of `--serve-params` and `--bench-params`, we share the benchmark results across `--sla-params` to avoid rerunning benchmarks with the same SLA variable value.
+!!! tip
+    This is our equivalent of [GuideLLM's `--profile sweep`](https://github.com/vllm-project/guidellm/blob/v0.5.3/src/guidellm/benchmark/profiles.py#L575).
 
-### Startup
+## Startup Benchmark
 
 `vllm bench sweep startup` runs `vllm bench startup` across parameter combinations to compare cold/warm startup time for different engine settings.
 
@@ -202,15 +192,28 @@ vllm bench sweep startup \
 
 `vllm bench sweep plot` can be used to plot performance curves from parameter sweep results.
 
-Example command:
+Control the variables to plot via `--var-x` and `--var-y`, optionally applying `--filter-by` and `--bin-by` to the values. The plot is organized according to `--fig-by`, `--row-by`, `--col-by`, and `--curve-by`.
+
+Example commands for visualizing [SLA Scanner](#sla-scanner) results:
 
 ```bash
+# Latency increases as the request rate increases
 vllm bench sweep plot benchmarks/results/<timestamp> \
-    --var-x max_concurrency \
+    --var-x request_rate \
+    --var-y p99_ttft_ms \
     --row-by random_input_len \
     --col-by random_output_len \
-    --curve-by api_server_count,max_num_batched_tokens \
-    --filter-by 'max_concurrency<=1024'
+    --curve-by max_num_seqs,max_num_batched_tokens \
+    --filter-by 'request_rate<=128'
+
+# Tradeoff between latency and throughput
+vllm bench sweep plot benchmarks/results/<timestamp> \
+    --var-x request_throughput \
+    --var-y median_ttft_ms \
+    --row-by random_input_len \
+    --col-by random_output_len \
+    --curve-by max_num_seqs,max_num_batched_tokens \
+    --filter-by 'request_rate<=128'
 ```
 
 !!! tip
@@ -233,3 +236,6 @@ Example:
 vllm bench sweep plot_pareto benchmarks/results/<timestamp> \
   --label-by max_concurrency,tensor_parallel_size,pipeline_parallel_size
 ```
+
+!!! tip
+    You can use `--dry-run` to preview the figures to be plotted.

docs/contributing/README.md

@@ -49,7 +49,13 @@ If you are developing vLLM's Python and CUDA/C++ code, install Pytorch first:
 uv pip install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu129
 ```
 
-then install vLLM using:
+Then install the necessary build dependencies from `requirements/build.txt`, skipping `torch` as it was installed in the previous step:
+
+```bash
+grep -v '^torch==' requirements/build.txt | uv pip install -r -
+```
+
+Finally install vLLM using:
 
 ```bash
 uv pip install -e . --no-build-isolation

docs/contributing/model/multimodal.md

@@ -293,21 +293,22 @@ Assuming that the memory usage increases with the number of tokens, the dummy in
             self,
             seq_len: int,
             mm_counts: Mapping[str, int],
-            mm_options: Mapping[str, BaseDummyOptions] | None = None,
+            mm_options: Mapping[str, BaseDummyOptions],
         ) -> MultiModalDataDict:
             num_images = mm_counts.get("image", 0)
 
             target_width, target_height = \
                 self.info.get_image_size_with_most_features()
 
-            image_overrides = mm_options.get("image") if mm_options else None
+            image_overrides = mm_options.get("image")
 
             return {
-                "image":
-                self._get_dummy_images(width=target_width,
-                                    height=target_height,
-                                    num_images=num_images,
-                                    overrides=image_overrides)
+                "image": self._get_dummy_images(
+                    width=target_width,
+                    height=target_height,
+                    num_images=num_images,
+                    overrides=image_overrides,
+                )
             }
         ```
 
@@ -479,17 +480,16 @@ Assuming that the memory usage increases with the number of tokens, the dummy in
             self,
             seq_len: int,
             mm_counts: Mapping[str, int],
-            mm_options: Optional[Mapping[str, BaseDummyOptions]] = None,
+            mm_options: Mapping[str, BaseDummyOptions],
         ) -> MultiModalDataDict:
             target_width, target_height = \
                 self.info.get_image_size_with_most_features()
             num_images = mm_counts.get("image", 0)
 
-            image_overrides = mm_options.get("image") if mm_options else None
+            image_overrides = mm_options.get("image")
 
             return {
-                "image":
-                self._get_dummy_images(
+                "image": self._get_dummy_images(
                     width=target_width,
                     height=target_height,
                     num_images=num_images,

docs/design/arch_overview.md

@@ -208,9 +208,7 @@ configurations affect the class we ultimately get.
 
 The following figure shows the class hierarchy of vLLM:
 
-> <figure markdown="span">
->   ![](../assets/design/hierarchy.png){ align="center" alt="query" width="100%" }
-> </figure>
+![Class Hierarchy](../assets/design/hierarchy.png)
 
 There are several important design choices behind this class hierarchy:
 

docs/design/moe_kernel_features.md

@@ -32,10 +32,10 @@ th {
 
 | Backend | Output act. format | Quant. types | Quant. format | Async | Apply Weight On Input | Subclass |
 |---------|--------------------|--------------|---------------|-------|-----------------------|-----------|
-| naive | standard | all<sup>1</sup> | G,A,T | N | <sup>6</sup> | [layer.py][vllm.model_executor.layers.fused_moe.layer.FusedMoE |
+| naive | standard | all<sup>1</sup> | G,A,T | N | <sup>6</sup> | [layer.py][vllm.model_executor.layers.fused_moe.layer.FusedMoE] |
 | pplx | batched | fp8,int8 | G,A,T | Y | Y | [`PplxPrepareAndFinalize`][vllm.model_executor.layers.fused_moe.pplx_prepare_finalize.PplxPrepareAndFinalize] |
-| deepep_high_throughput | standard | fp8 | G(128),A,T<sup>2</sup> | Y | Y | [`DeepEPLLPrepareAndFinalize`][vllm.model_executor.layers.fused_moe.deepep_ll_prepare_finalize.DeepEPLLPrepareAndFinalize] |
-| deepep_low_latency | batched | fp8 | G(128),A,T<sup>3</sup> | Y | Y | [`DeepEPHTPrepareAndFinalize`][vllm.model_executor.layers.fused_moe.deepep_ht_prepare_finalize.DeepEPHTPrepareAndFinalize] |
+| deepep_high_throughput | standard | fp8 | G(128),A,T<sup>2</sup> | Y | Y | [`DeepEPHTPrepareAndFinalize`][vllm.model_executor.layers.fused_moe.deepep_ht_prepare_finalize.DeepEPHTPrepareAndFinalize] |
+| deepep_low_latency | batched | fp8 | G(128),A,T<sup>3</sup> | Y | Y | [`DeepEPLLPrepareAndFinalize`][vllm.model_executor.layers.fused_moe.deepep_ll_prepare_finalize.DeepEPLLPrepareAndFinalize] |
 | flashinfer_all2allv | standard | nvfp4,fp8 | G,A,T | N | N | [`FlashInferA2APrepareAndFinalize`][vllm.model_executor.layers.fused_moe.flashinfer_a2a_prepare_finalize.FlashInferA2APrepareAndFinalize] |
 | MoEPrepareAndFinalizeNoEP<sup>5</sup> | standard | fp8,int8 | G,A,T | N | Y | [`MoEPrepareAndFinalizeNoEP`][vllm.model_executor.layers.fused_moe.prepare_finalize.MoEPrepareAndFinalizeNoEP] |
 | BatchedPrepareAndFinalize<sup>5</sup> | batched | fp8,int8 | G,A,T | N | Y | [`BatchedPrepareAndFinalize`][vllm.model_executor.layers.fused_moe.fused_batched_moe.BatchedPrepareAndFinalize] |

docs/design/plugin_system.md

@@ -155,3 +155,4 @@ The interface for the model/module may change during vLLM's development. If you
     - `use_v1` parameter in `Platform.get_attn_backend_cls` is deprecated. It has been removed in v0.13.0.
     - `_Backend` in `vllm.attention` is deprecated. It has been removed in v0.13.0. Please use `vllm.v1.attention.backends.registry.register_backend` to add new attention backend to `AttentionBackendEnum` instead.
     - `seed_everything` platform interface is deprecated. It has been removed in v0.16.0. Please use `vllm.utils.torch_utils.set_random_seed` instead.
+    - `prompt` in `Platform.validate_request` is deprecated and will be removed in v0.18.0.

docs/features/README.md

@@ -36,12 +36,12 @@ th:not(:first-child) {
 }
 </style>
 
-| Feature | [CP](../configuration/optimization.md#chunked-prefill) | [APC](automatic_prefix_caching.md) | [LoRA](lora.md) | [SD](spec_decode/README.md) | CUDA graph | [pooling](../models/pooling_models.md) | <abbr title="Encoder-Decoder Models">enc-dec</abbr> | <abbr title="Logprobs">logP</abbr> | <abbr title="Prompt Logprobs">prmpt logP</abbr> | <abbr title="Async Output Processing">async output</abbr> | multi-step | <abbr title="Multimodal Inputs">mm</abbr> | best-of | beam-search | [prompt-embeds](prompt_embeds.md) |
+| Feature | [CP](../configuration/optimization.md#chunked-prefill) | [APC](automatic_prefix_caching.md) | [LoRA](lora.md) | [SD](speculative_decoding/README.md) | CUDA graph | [pooling](../models/pooling_models.md) | <abbr title="Encoder-Decoder Models">enc-dec</abbr> | <abbr title="Logprobs">logP</abbr> | <abbr title="Prompt Logprobs">prmpt logP</abbr> | <abbr title="Async Output Processing">async output</abbr> | multi-step | <abbr title="Multimodal Inputs">mm</abbr> | best-of | beam-search | [prompt-embeds](prompt_embeds.md) |
 |---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
 | [CP](../configuration/optimization.md#chunked-prefill) | ✅ | | | | | | | | | | | | | | |
 | [APC](automatic_prefix_caching.md) | ✅ | ✅ | | | | | | | | | | | | | |
 | [LoRA](lora.md) | ✅ | ✅ | ✅ | | | | | | | | | | | | |
-| [SD](spec_decode/README.md) | ✅ | ✅ | ❌ | ✅ | | | | | | | | | | | |
+| [SD](speculative_decoding/README.md) | ✅ | ✅ | ❌ | ✅ | | | | | | | | | | | |
 | CUDA graph | ✅ | ✅ | ✅ | ✅ | ✅ | | | | | | | | | | |
 | [pooling](../models/pooling_models.md) | 🟠\* | 🟠\* | ✅ | ❌ | ✅ | ✅ | | | | | | | | | |
 | <abbr title="Encoder-Decoder Models">enc-dec</abbr> | ❌ | [❌](https://github.com/vllm-project/vllm/issues/7366) | ❌ | [❌](https://github.com/vllm-project/vllm/issues/7366) | ✅ | ✅ | ✅ | | | | | | | | |
@@ -64,7 +64,7 @@ th:not(:first-child) {
 | [CP](../configuration/optimization.md#chunked-prefill)                                     | [❌](https://github.com/vllm-project/vllm/issues/2729) | ✅        | ✅        | ✅     | ✅        | ✅                  | ✅     | ✅        |
 | [APC](automatic_prefix_caching.md)                        | [❌](https://github.com/vllm-project/vllm/issues/3687) | ✅        | ✅        | ✅     | ✅        | ✅                  | ✅     | ✅        |
 | [LoRA](lora.md)                                           | ✅                  | ✅        | ✅        | ✅     | ✅        | ✅                  | ✅     | ✅        |
-| [SD](spec_decode/README.md)                        | ✅                  | ✅        | ✅        | ✅     | ✅        | ❌                  | ✅     | ✅        |
+| [SD](speculative_decoding/README.md)                        | ✅                  | ✅        | ✅        | ✅     | ✅        | ❌                  | ✅     | ✅        |
 | CUDA graph                                                | ✅                  | ✅        | ✅        | ✅     | ✅        | ❌                  | ✅     | [❌](https://github.com/vllm-project/vllm/issues/26970)        |
 | [pooling](../models/pooling_models.md)                    | ✅                  | ✅        | ✅        | ✅     | ✅        | ✅                  | ✅     | ✅        |
 | <abbr title="Encoder-Decoder Models">enc-dec</abbr>       | ✅                  | ✅        | ✅        | ✅     | ✅        | ✅                  | ❌     | ✅        |

docs/features/multimodal_inputs.md

@@ -295,6 +295,51 @@ You can pass a tuple `(array, sampling_rate)` to the `'audio'` field of the mult
 
 Full example: [examples/offline_inference/audio_language.py](../../examples/offline_inference/audio_language.py)
 
+#### Chunking Long Audio for Transcription
+
+Speech-to-text models like Whisper have a maximum audio length they can process (typically 30 seconds). For longer audio files, vLLM provides a utility to intelligently split audio into chunks at quiet points to minimize cutting through speech.
+
+```python
+import librosa
+from vllm import LLM, SamplingParams
+from vllm.multimodal.audio import split_audio
+
+# Load long audio file
+audio, sr = librosa.load("long_audio.wav", sr=16000)
+
+# Split into chunks at low-energy (quiet) regions
+chunks = split_audio(
+    audio_data=audio,
+    sample_rate=sr,
+    max_clip_duration_s=30.0,      # Maximum chunk length in seconds
+    overlap_duration_s=1.0,         # Search window for finding quiet split points
+    min_energy_window_size=1600,    # Window size for energy calculation (~100ms at 16kHz)
+)
+
+# Initialize Whisper model
+llm = LLM(model="openai/whisper-large-v3-turbo")
+sampling_params = SamplingParams(temperature=0, max_tokens=256)
+
+# Transcribe each chunk
+transcriptions = []
+for chunk in chunks:
+    outputs = llm.generate({
+        "prompt": "<|startoftranscript|><|en|><|transcribe|><|notimestamps|>",
+        "multi_modal_data": {"audio": (chunk, sr)},
+    }, sampling_params)
+    transcriptions.append(outputs[0].outputs[0].text)
+
+# Combine results
+full_transcription = " ".join(transcriptions)
+```
+
+The `split_audio` function:
+
+- Splits audio at quiet points to avoid cutting through speech
+- Uses RMS energy to find low-amplitude regions within the overlap window
+- Preserves all audio samples (no data loss)
+- Supports any sample rate
+
 #### Automatic Audio Channel Normalization
 
 vLLM automatically normalizes audio channels for models that require specific audio formats. When loading audio with libraries like `torchaudio`, stereo files return shape `[channels, time]`, but many audio models (particularly Whisper-based models) expect mono audio with shape `[time]`.

docs/features/nixl_connector_usage.md

@@ -197,8 +197,8 @@ For multi-host DP deployment, only need to provide the host/port of the head ins
 
 The `kv_load_failure_policy` setting controls how the system handles failures when the decoder instance loads KV cache blocks from the prefiller instance:
 
-- **fail** (recommended): Immediately fail the request with an error when KV load fails. This prevents performance degradation by avoiding recomputation of prefill work on the decode instance.
-- **recompute** (default): Recompute failed blocks locally on the decode instance. This may cause performance _jitter_ on decode instances as the scheduled prefill will delay and interfere with other decodes. Furthermore, decode instances are typically configured with low-latency optimizations.
+- **fail** (default): Immediately fail the request with an error when KV load fails. This prevents performance degradation by avoiding recomputation of prefill work on the decode instance.
+- **recompute**: Recompute failed blocks locally on the decode instance. This may cause performance _jitter_ on decode instances as the scheduled prefill will delay and interfere with other decodes. Furthermore, decode instances are typically configured with low-latency optimizations.
 
 !!! warning
     Using `kv_load_failure_policy="recompute"` can lead to performance degradation in production deployments. When KV loads fail, the decode instance will execute prefill work with decode-optimized configurations, which is inefficient and defeats the purpose of disaggregated prefilling. This also increases tail latency for other ongoing decode requests.

docs/features/quantization/bnb.md

@@ -7,7 +7,7 @@ Compared to other quantization methods, BitsAndBytes eliminates the need for cal
 Below are the steps to utilize BitsAndBytes with vLLM.
 
 ```bash
-pip install bitsandbytes>=0.46.1
+pip install bitsandbytes>=0.49.2
 ```
 
 vLLM reads the model's config file and supports both in-flight quantization and pre-quantized checkpoint.

docs/features/spec_decode/README.md

@@ -1,330 +0,0 @@
-# Speculative Decoding
-
-!!! warning
-    Currently, speculative decoding in vLLM is not compatible with pipeline parallelism.
-
-This document shows how to use [Speculative Decoding](https://x.com/karpathy/status/1697318534555336961) with vLLM.
-Speculative decoding is a technique which improves inter-token latency in memory-bound LLM inference.
-
-!!! tip
-    To train your own draft models for speculative decoding, see [Speculators](speculators.md), a library for training draft models that integrates seamlessly with vLLM.
-
-## Speculating with a draft model
-
-The following code configures vLLM in an offline mode to use speculative decoding with a draft model, speculating 5 tokens at a time.
-
-!!! warning
-    In vllm v0.10.0, speculative decoding with a draft model is not supported.
-    If you use the following code, you will get a `NotImplementedError`.
-
-??? code
-
-    ```python
-    from vllm import LLM, SamplingParams
-
-    prompts = [
-        "The future of AI is",
-    ]
-    sampling_params = SamplingParams(temperature=0.8, top_p=0.95)
-
-    llm = LLM(
-        model="facebook/opt-6.7b",
-        tensor_parallel_size=1,
-        speculative_config={
-            "model": "facebook/opt-125m",
-            "num_speculative_tokens": 5,
-        },
-    )
-    outputs = llm.generate(prompts, sampling_params)
-
-    for output in outputs:
-        prompt = output.prompt
-        generated_text = output.outputs[0].text
-        print(f"Prompt: {prompt!r}, Generated text: {generated_text!r}")
-    ```
-
-To perform the same with an online mode launch the server:
-
-```bash
-vllm serve facebook/opt-6.7b \
-    --host 0.0.0.0 \
-    --port 8000 \
-    --seed 42 \
-    -tp 1 \
-    --gpu_memory_utilization 0.8 \
-    --speculative_config '{"model": "facebook/opt-125m", "num_speculative_tokens": 5}'
-```
-
-!!! warning
-    Note: Please use `--speculative_config` to set all configurations related to speculative decoding. The previous method of specifying the model through `--speculative_model` and adding related parameters (e.g., `--num_speculative_tokens`) separately has been deprecated now.
-
-Then use a client:
-
-??? code
-
-    ```python
-    from openai import OpenAI
-
-    # Modify OpenAI's API key and API base to use vLLM's API server.
-    openai_api_key = "EMPTY"
-    openai_api_base = "http://localhost:8000/v1"
-
-    client = OpenAI(
-        # defaults to os.environ.get("OPENAI_API_KEY")
-        api_key=openai_api_key,
-        base_url=openai_api_base,
-    )
-
-    models = client.models.list()
-    model = models.data[0].id
-
-    # Completion API
-    stream = False
-    completion = client.completions.create(
-        model=model,
-        prompt="The future of AI is",
-        echo=False,
-        n=1,
-        stream=stream,
-    )
-
-    print("Completion results:")
-    if stream:
-        for c in completion:
-            print(c)
-    else:
-        print(completion)
-    ```
-
-## Speculating by matching n-grams in the prompt
-
-The following code configures vLLM to use speculative decoding where proposals are generated by
-matching n-grams in the prompt. For more information read [this thread.](https://x.com/joao_gante/status/1747322413006643259)
-
-??? code
-
-    ```python
-    from vllm import LLM, SamplingParams
-
-    prompts = [
-        "The future of AI is",
-    ]
-    sampling_params = SamplingParams(temperature=0.8, top_p=0.95)
-
-    llm = LLM(
-        model="facebook/opt-6.7b",
-        tensor_parallel_size=1,
-        speculative_config={
-            "method": "ngram",
-            "num_speculative_tokens": 5,
-            "prompt_lookup_max": 4,
-        },
-    )
-    outputs = llm.generate(prompts, sampling_params)
-
-    for output in outputs:
-        prompt = output.prompt
-        generated_text = output.outputs[0].text
-        print(f"Prompt: {prompt!r}, Generated text: {generated_text!r}")
-    ```
-
-## Speculating using Suffix Decoding
-
-The following code configures vLLM to use speculative decoding where proposals are generated using Suffix Decoding ([technical report](https://arxiv.org/abs/2411.04975)).
-
-Like n-gram, Suffix Decoding can generate draft tokens by pattern-matching using the last `n` generated tokens. Unlike n-gram, Suffix Decoding (1) can pattern-match against both the prompt and previous generations, (2) uses frequency counts to propose the most likely continuations, and (3) speculates an adaptive number of tokens for each request at each iteration to get better acceptance rates.
-
-Suffix Decoding can achieve better performance for tasks with high repetition, such as code-editing, agentic loops (e.g. self-reflection, self-consistency), and RL rollouts.
-
-!!! tip "Install Arctic Inference"
-    Suffix Decoding requires [Arctic Inference](https://github.com/snowflakedb/ArcticInference). You can install it with `pip install arctic-inference`.
-
-!!! tip "Suffix Decoding Speculative Tokens"
-    Suffix Decoding will speculate a dynamic number of tokens for each request at each decoding step, so the `num_speculative_tokens` configuration specifies the *maximum* number of speculative tokens. It is suggested to use a high number such as `16` or `32` (default).
-
-??? code
-
-    ```python
-    from vllm import LLM, SamplingParams
-
-    prompts = [
-        "The future of AI is",
-    ]
-    sampling_params = SamplingParams(temperature=0.8, top_p=0.95)
-
-    llm = LLM(
-        model="facebook/opt-6.7b",
-        tensor_parallel_size=1,
-        speculative_config={
-            "method": "suffix",
-            "num_speculative_tokens": 32,
-        },
-    )
-    outputs = llm.generate(prompts, sampling_params)
-
-    for output in outputs:
-        prompt = output.prompt
-        generated_text = output.outputs[0].text
-        print(f"Prompt: {prompt!r}, Generated text: {generated_text!r}")
-    ```
-
-## Speculating using MLP speculators
-
-The following code configures vLLM to use speculative decoding where proposals are generated by
-draft models that condition draft predictions on both context vectors and sampled tokens.
-For more information see [this blog](https://pytorch.org/blog/hitchhikers-guide-speculative-decoding/) or
-[this technical report](https://arxiv.org/abs/2404.19124).
-
-??? code
-
-    ```python
-    from vllm import LLM, SamplingParams
-
-    prompts = [
-        "The future of AI is",
-    ]
-    sampling_params = SamplingParams(temperature=0.8, top_p=0.95)
-
-    llm = LLM(
-        model="meta-llama/Meta-Llama-3.1-70B-Instruct",
-        tensor_parallel_size=4,
-        speculative_config={
-            "model": "ibm-ai-platform/llama3-70b-accelerator",
-            "draft_tensor_parallel_size": 1,
-        },
-    )
-    outputs = llm.generate(prompts, sampling_params)
-
-    for output in outputs:
-        prompt = output.prompt
-        generated_text = output.outputs[0].text
-        print(f"Prompt: {prompt!r}, Generated text: {generated_text!r}")
-    ```
-
-Note that these speculative models currently need to be run without tensor parallelism, although
-it is possible to run the main model using tensor parallelism (see example above). Since the
-speculative models are relatively small, we still see significant speedups. However, this
-limitation will be fixed in a future release.
-
-A variety of speculative models of this type are available on HF hub:
-
-- [llama-13b-accelerator](https://huggingface.co/ibm-ai-platform/llama-13b-accelerator)
-- [llama3-8b-accelerator](https://huggingface.co/ibm-ai-platform/llama3-8b-accelerator)
-- [codellama-34b-accelerator](https://huggingface.co/ibm-ai-platform/codellama-34b-accelerator)
-- [llama2-70b-accelerator](https://huggingface.co/ibm-ai-platform/llama2-70b-accelerator)
-- [llama3-70b-accelerator](https://huggingface.co/ibm-ai-platform/llama3-70b-accelerator)
-- [granite-3b-code-instruct-accelerator](https://huggingface.co/ibm-granite/granite-3b-code-instruct-accelerator)
-- [granite-8b-code-instruct-accelerator](https://huggingface.co/ibm-granite/granite-8b-code-instruct-accelerator)
-- [granite-7b-instruct-accelerator](https://huggingface.co/ibm-granite/granite-7b-instruct-accelerator)
-- [granite-20b-code-instruct-accelerator](https://huggingface.co/ibm-granite/granite-20b-code-instruct-accelerator)
-
-## Speculating using EAGLE based draft models
-
-The following code configures vLLM to use speculative decoding where proposals are generated by
-an [EAGLE (Extrapolation Algorithm for Greater Language-model Efficiency)](https://arxiv.org/pdf/2401.15077) based draft model. A more detailed example for offline mode, including how to extract request level acceptance rate, can be found in [examples/offline_inference/spec_decode.py](../../../examples/offline_inference/spec_decode.py)
-
-??? code
-
-    ```python
-    from vllm import LLM, SamplingParams
-
-    prompts = [
-        "The future of AI is",
-    ]
-    sampling_params = SamplingParams(temperature=0.8, top_p=0.95)
-
-    llm = LLM(
-        model="meta-llama/Meta-Llama-3-8B-Instruct",
-        tensor_parallel_size=4,
-        speculative_config={
-            "model": "yuhuili/EAGLE-LLaMA3-Instruct-8B",
-            "draft_tensor_parallel_size": 1,
-            "num_speculative_tokens": 2,
-            "method": "eagle",
-        },
-    )
-
-    outputs = llm.generate(prompts, sampling_params)
-
-    for output in outputs:
-        prompt = output.prompt
-        generated_text = output.outputs[0].text
-        print(f"Prompt: {prompt!r}, Generated text: {generated_text!r}")
-
-    ```
-
-A few important things to consider when using the EAGLE based draft models:
-
-1. The EAGLE draft models available in the [HF repository for EAGLE models](https://huggingface.co/yuhuili) should
-   be able to be loaded and used directly by vLLM after <https://github.com/vllm-project/vllm/pull/12304>.
-   If you are using vllm version before <https://github.com/vllm-project/vllm/pull/12304>, please use the
-   [script](https://gist.github.com/abhigoyal1997/1e7a4109ccb7704fbc67f625e86b2d6d) to convert the speculative model,
-   and specify `"model": "path/to/modified/eagle/model"` in `speculative_config`. If weight-loading problems still occur when using the latest version of vLLM, please leave a comment or raise an issue.
-
-2. The EAGLE based draft models need to be run without tensor parallelism
-   (i.e. draft_tensor_parallel_size is set to 1 in `speculative_config`), although
-   it is possible to run the main model using tensor parallelism (see example above).
-
-3. When using EAGLE-based speculators with vLLM, the observed speedup is lower than what is
-   reported in the reference implementation [here](https://github.com/SafeAILab/EAGLE). This issue is under
-   investigation and tracked here: <https://github.com/vllm-project/vllm/issues/9565>.
-
-4. When using EAGLE-3 based draft model, option "method" must be set to "eagle3".
-   That is, to specify `"method": "eagle3"` in `speculative_config`.
-
-A variety of EAGLE draft models are available on the Hugging Face hub:
-
-| Base Model                                                           | EAGLE on Hugging Face                     | # EAGLE Parameters |
-|---------------------------------------------------------------------|-------------------------------------------|--------------------|
-| Vicuna-7B-v1.3                                                       | yuhuili/EAGLE-Vicuna-7B-v1.3             | 0.24B              |
-| Vicuna-13B-v1.3                                                      | yuhuili/EAGLE-Vicuna-13B-v1.3            | 0.37B              |
-| Vicuna-33B-v1.3                                                      | yuhuili/EAGLE-Vicuna-33B-v1.3            | 0.56B              |
-| LLaMA2-Chat 7B                                                       | yuhuili/EAGLE-llama2-chat-7B             | 0.24B              |
-| LLaMA2-Chat 13B                                                      | yuhuili/EAGLE-llama2-chat-13B            | 0.37B              |
-| LLaMA2-Chat 70B                                                      | yuhuili/EAGLE-llama2-chat-70B            | 0.99B              |
-| Mixtral-8x7B-Instruct-v0.1                                           | yuhuili/EAGLE-mixtral-instruct-8x7B      | 0.28B              |
-| LLaMA3-Instruct 8B                                                   | yuhuili/EAGLE-LLaMA3-Instruct-8B         | 0.25B              |
-| LLaMA3-Instruct 70B                                                  | yuhuili/EAGLE-LLaMA3-Instruct-70B        | 0.99B              |
-| Qwen2-7B-Instruct                                                    | yuhuili/EAGLE-Qwen2-7B-Instruct          | 0.26B              |
-| Qwen2-72B-Instruct                                                   | yuhuili/EAGLE-Qwen2-72B-Instruct         | 1.05B              |
-
-## Lossless guarantees of Speculative Decoding
-
-In vLLM, speculative decoding aims to enhance inference efficiency while maintaining accuracy. This section addresses the lossless guarantees of
-speculative decoding, breaking down the guarantees into three key areas:
-
-1. **Theoretical Losslessness**
-   \- Speculative decoding sampling is theoretically lossless up to the precision limits of hardware numerics. Floating-point errors might
-   cause slight variations in output distributions, as discussed
-   in [Accelerating Large Language Model Decoding with Speculative Sampling](https://arxiv.org/pdf/2302.01318)
-
-2. **Algorithmic Losslessness**
-   \- vLLM’s implementation of speculative decoding is algorithmically validated to be lossless. Key validation tests include:
-
-    > - **Rejection Sampler Convergence**: Ensures that samples from vLLM’s rejection sampler align with the target
-    >   distribution. [View Test Code](https://github.com/vllm-project/vllm/blob/47b65a550866c7ffbd076ecb74106714838ce7da/tests/samplers/test_rejection_sampler.py#L252)
-    > - **Greedy Sampling Equality**: Confirms that greedy sampling with speculative decoding matches greedy sampling
-    >   without it. This verifies that vLLM's speculative decoding framework, when integrated with the vLLM forward pass and the vLLM rejection sampler,
-    >   provides a lossless guarantee. Almost all of the tests in [tests/spec_decode/e2e](../../tests/spec_decode/e2e).
-    >   verify this property using [this assertion implementation](https://github.com/vllm-project/vllm/blob/b67ae00cdbbe1a58ffc8ff170f0c8d79044a684a/tests/spec_decode/e2e/conftest.py#L291)
-
-3. **vLLM Logprob Stability**
-   \- vLLM does not currently guarantee stable token log probabilities (logprobs). This can result in different outputs for the
-   same request across runs. For more details, see the FAQ section
-   titled *Can the output of a prompt vary across runs in vLLM?* in the [FAQs](../../usage/faq.md).
-
-While vLLM strives to ensure losslessness in speculative decoding, variations in generated outputs with and without speculative decoding
-can occur due to following factors:
-
-- **Floating-Point Precision**: Differences in hardware numerical precision may lead to slight discrepancies in the output distribution.
-- **Batch Size and Numerical Stability**: Changes in batch size may cause variations in logprobs and output probabilities, potentially
-  due to non-deterministic behavior in batched operations or numerical instability.
-
-For mitigation strategies, please refer to the FAQ entry *Can the output of a prompt vary across runs in vLLM?* in the [FAQs](../../usage/faq.md).
-
-## Resources for vLLM contributors
-
-- [A Hacker's Guide to Speculative Decoding in vLLM](https://www.youtube.com/watch?v=9wNAgpX6z_4)
-- [What is Lookahead Scheduling in vLLM?](https://docs.google.com/document/d/1Z9TvqzzBPnh5WHcRwjvK2UEeFeq5zMZb5mFE8jR0HCs/edit#heading=h.1fjfb0donq5a)
-- [Information on batch expansion](https://docs.google.com/document/d/1T-JaS2T1NRfdP51qzqpyakoCXxSXTtORppiwaj5asxA/edit#heading=h.kk7dq05lc6q8)
-- [Dynamic speculative decoding](https://github.com/vllm-project/vllm/issues/4565)

docs/features/speculative_decoding/README.md

@@ -0,0 +1,62 @@
+# Speculative Decoding
+
+This document shows how to use [Speculative Decoding](https://arxiv.org/pdf/2302.01318) with vLLM to reduce inter-token latency under medium-to-low QPS (query per second), memory-bound workloads.
+
+To train your own draft models for optimized speculative decoding, see [vllm-project/speculators](speculators.md) for seamless training and integration with vLLM.
+
+## vLLM Speculation Methods
+
+vLLM supports a variety of methods of speculative decoding. Model-based methods such as EAGLE, draft models, and mlp provide the best latency reduction, while simpler methods such as n-gram and and suffix decoding provide modest speedups without increasing workload during peak traffic.
+
+- [EAGLE](eagle.md)
+- [Draft Model](draft_model.md)
+- [Multi-Layer Perceptron](mlp.md)
+- [N-Gram](n_gram.md)
+- [Suffix Decoding](suffix.md)
+
+## Lossless guarantees of Speculative Decoding
+
+In vLLM, speculative decoding aims to enhance inference efficiency while maintaining accuracy. This section addresses the lossless guarantees of
+speculative decoding, breaking down the guarantees into three key areas:
+
+1. **Theoretical Losslessness**
+   \- Speculative decoding sampling is theoretically lossless up to the precision limits of hardware numerics. Floating-point errors might
+   cause slight variations in output distributions, as discussed
+   in [Accelerating Large Language Model Decoding with Speculative Sampling](https://arxiv.org/pdf/2302.01318)
+
+2. **Algorithmic Losslessness**
+   \- vLLM’s implementation of speculative decoding is algorithmically validated to be lossless. Key validation tests include:
+
+    > - **Rejection Sampler Convergence**: Ensures that samples from vLLM’s rejection sampler align with the target
+    >   distribution. [View Test Code](https://github.com/vllm-project/vllm/blob/47b65a550866c7ffbd076ecb74106714838ce7da/tests/samplers/test_rejection_sampler.py#L252)
+    > - **Greedy Sampling Equality**: Confirms that greedy sampling with speculative decoding matches greedy sampling
+    >   without it. This verifies that vLLM's speculative decoding framework, when integrated with the vLLM forward pass and the vLLM rejection sampler,
+    >   provides a lossless guarantee. Almost all of the tests in [tests/spec_decode/e2e](/tests/v1/spec_decode).
+    >   verify this property using [this assertion implementation](https://github.com/vllm-project/vllm/blob/b67ae00cdbbe1a58ffc8ff170f0c8d79044a684a/tests/spec_decode/e2e/conftest.py#L291)
+
+3. **vLLM Logprob Stability**
+   \- vLLM does not currently guarantee stable token log probabilities (logprobs). This can result in different outputs for the
+   same request across runs. For more details, see the FAQ section
+   titled *Can the output of a prompt vary across runs in vLLM?* in the [FAQs](../../usage/faq.md).
+
+While vLLM strives to ensure losslessness in speculative decoding, variations in generated outputs with and without speculative decoding
+can occur due to following factors:
+
+- **Floating-Point Precision**: Differences in hardware numerical precision may lead to slight discrepancies in the output distribution.
+- **Batch Size and Numerical Stability**: Changes in batch size may cause variations in logprobs and output probabilities, potentially
+  due to non-deterministic behavior in batched operations or numerical instability.
+
+For mitigation strategies, please refer to the FAQ entry *Can the output of a prompt vary across runs in vLLM?* in the [FAQs](../../usage/faq.md).
+
+## Known Feature Incompatibility
+
+1. Pipeline parallelism is not composible with speculative decoding as of `vllm<=0.15.0`
+2. Speculative decoding with a draft models is not supported in `vllm<=0.10.0`
+
+## Resources for vLLM contributors
+
+- [[vLLM Office Hours #40] Intro to Speculators](https://www.youtube.com/watch?v=2ISAr_JVGLs)
+- [A Hacker's Guide to Speculative Decoding in vLLM](https://www.youtube.com/watch?v=9wNAgpX6z_4)
+- [What is Lookahead Scheduling in vLLM?](https://docs.google.com/document/d/1Z9TvqzzBPnh5WHcRwjvK2UEeFeq5zMZb5mFE8jR0HCs/edit#heading=h.1fjfb0donq5a)
+- [Information on batch expansion](https://docs.google.com/document/d/1T-JaS2T1NRfdP51qzqpyakoCXxSXTtORppiwaj5asxA/edit#heading=h.kk7dq05lc6q8)
+- [Dynamic speculative decoding](https://github.com/vllm-project/vllm/issues/4565)

docs/features/speculative_decoding/draft_model.md

@@ -0,0 +1,80 @@
+# Draft Models
+
+The following code configures vLLM in an offline mode to use speculative decoding with a draft model, speculating 5 tokens at a time.
+
+```python
+from vllm import LLM, SamplingParams
+
+prompts = ["The future of AI is"]
+sampling_params = SamplingParams(temperature=0.8, top_p=0.95)
+
+llm = LLM(
+    model="Qwen/Qwen3-8B",
+    tensor_parallel_size=1,
+    speculative_config={
+        "model": "Qwen/Qwen3-0.6B",
+        "num_speculative_tokens": 5,
+        "method": "draft_model",
+    },
+)
+outputs = llm.generate(prompts, sampling_params)
+
+for output in outputs:
+    prompt = output.prompt
+    generated_text = output.outputs[0].text
+    print(f"Prompt: {prompt!r}, Generated text: {generated_text!r}")
+```
+
+To perform the equivalent launch in online mode, use the following server-side code:
+
+```bash
+vllm serve Qwen/Qwen3-4B-Thinking-2507 \
+    --host 0.0.0.0 \
+    --port 8000 \
+    --seed 42 \
+    -tp 1 \
+    --max_model_len 2048 \
+    --gpu_memory_utilization 0.8 \
+    --speculative_config '{"model": "Qwen/Qwen3-0.6B", "num_speculative_tokens": 5, "method": "draft_model"}'
+```
+
+The code used to request as completions as a client remains unchanged:
+
+??? code
+
+    ```python
+    from openai import OpenAI
+
+    # Modify OpenAI's API key and API base to use vLLM's API server.
+    openai_api_key = "EMPTY"
+    openai_api_base = "http://localhost:8000/v1"
+
+    client = OpenAI(
+        # defaults to os.environ.get("OPENAI_API_KEY")
+        api_key=openai_api_key,
+        base_url=openai_api_base,
+    )
+
+    models = client.models.list()
+    model = models.data[0].id
+
+    # Completion API
+    stream = False
+    completion = client.completions.create(
+        model=model,
+        prompt="The future of AI is",
+        echo=False,
+        n=1,
+        stream=stream,
+    )
+
+    print("Completion results:")
+    if stream:
+        for c in completion:
+            print(c)
+    else:
+        print(completion)
+    ```
+
+!!! warning
+    Note: Please use `--speculative_config` to set all configurations related to speculative decoding. The previous method of specifying the model through `--speculative_model` and adding related parameters (e.g., `--num_speculative_tokens`) separately has been deprecated.

docs/features/speculative_decoding/eagle.md

@@ -0,0 +1,67 @@
+# EAGLE Draft Models
+
+The following code configures vLLM to use speculative decoding where proposals are generated by an [EAGLE (Extrapolation Algorithm for Greater Language-model Efficiency)](https://arxiv.org/pdf/2401.15077) based draft model. A more detailed example for offline mode, including how to extract request level acceptance rate, can be found in [examples/offline_inference/spec_decode.py](../../../examples/offline_inference/spec_decode.py)
+
+## Eagle Drafter Example
+
+```python
+from vllm import LLM, SamplingParams
+
+prompts = ["The future of AI is"]
+sampling_params = SamplingParams(temperature=0.8, top_p=0.95)
+
+llm = LLM(
+    model="meta-llama/Meta-Llama-3-8B-Instruct",
+    tensor_parallel_size=4,
+    speculative_config={
+        "model": "yuhuili/EAGLE-LLaMA3-Instruct-8B",
+        "draft_tensor_parallel_size": 1,
+        "num_speculative_tokens": 2,
+        "method": "eagle",
+    },
+)
+
+outputs = llm.generate(prompts, sampling_params)
+
+for output in outputs:
+    prompt = output.prompt
+    generated_text = output.outputs[0].text
+    print(f"Prompt: {prompt!r}, Generated text: {generated_text!r}")
+```
+
+## Eagle3 Drafter Example
+
+```python
+from vllm import LLM, SamplingParams
+
+prompts = ["The future of AI is"]
+sampling_params = SamplingParams(temperature=0.8, top_p=0.95)
+
+llm = LLM(
+    model="meta-llama/Meta-Llama-3-8B-Instruct",
+    tensor_parallel_size=2,
+    speculative_config={
+        "model": "RedHatAI/Llama-3.1-8B-Instruct-speculator.eagle3",
+        "draft_tensor_parallel_size": 2,
+        "num_speculative_tokens": 2,
+        "method": "eagle3",
+    },
+)
+
+outputs = llm.generate(prompts, sampling_params)
+
+for output in outputs:
+    prompt = output.prompt
+    generated_text = output.outputs[0].text
+    print(f"Prompt: {prompt!r}, Generated text: {generated_text!r}")
+```
+
+## Pre-Trained Eagle Draft Models
+
+A variety of EAGLE draft models are available on the Hugging Face hub:
+
+* [RedHatAI/speculator-models](https://huggingface.co/collections/RedHatAI/speculator-models)
+* [yuhuili/models](https://huggingface.co/yuhuili/models?search=eagle)
+
+!!! warning
+    If you are using `vllm<0.7.0`, please use [this script](https://gist.github.com/abhigoyal1997/1e7a4109ccb7704fbc67f625e86b2d6d) to convert the speculative model and specify `"model": "path/to/modified/eagle/model"` in `speculative_config`.

docs/features/speculative_decoding/mlp.md

@@ -0,0 +1,42 @@
+# MLP Draft Models
+
+The following code configures vLLM to use speculative decoding where proposals are generated by draft models that condition draft predictions on both context vectors and sampled tokens. For more information see [The Hitchhiker's Guide to Speculative Decoding](https://pytorch.org/blog/hitchhikers-guide-speculative-decoding/) and [IBM Research's Technical Report](https://arxiv.org/abs/2404.19124).
+
+## MLP Drafter Example
+
+```python
+from vllm import LLM, SamplingParams
+
+prompts = ["The future of AI is"]
+sampling_params = SamplingParams(temperature=0.8, top_p=0.95)
+
+llm = LLM(
+    model="meta-llama/Meta-Llama-3.1-70B-Instruct",
+    tensor_parallel_size=4,
+    speculative_config={
+        "model": "ibm-ai-platform/llama3-70b-accelerator",
+        "draft_tensor_parallel_size": 1,
+        "method": "mlp_speculator",
+    },
+)
+outputs = llm.generate(prompts, sampling_params)
+
+for output in outputs:
+    prompt = output.prompt
+    generated_text = output.outputs[0].text
+    print(f"Prompt: {prompt!r}, Generated text: {generated_text!r}")
+```
+
+## Pre-Trained MLP Drafter Models
+
+A variety of speculative models of this type are available on HF hub:
+
+- [llama-13b-accelerator](https://huggingface.co/ibm-ai-platform/llama-13b-accelerator)
+- [llama3-8b-accelerator](https://huggingface.co/ibm-ai-platform/llama3-8b-accelerator)
+- [codellama-34b-accelerator](https://huggingface.co/ibm-ai-platform/codellama-34b-accelerator)
+- [llama2-70b-accelerator](https://huggingface.co/ibm-ai-platform/llama2-70b-accelerator)
+- [llama3-70b-accelerator](https://huggingface.co/ibm-ai-platform/llama3-70b-accelerator)
+- [granite-3b-code-instruct-accelerator](https://huggingface.co/ibm-granite/granite-3b-code-instruct-accelerator)
+- [granite-8b-code-instruct-accelerator](https://huggingface.co/ibm-granite/granite-8b-code-instruct-accelerator)
+- [granite-7b-instruct-accelerator](https://huggingface.co/ibm-granite/granite-7b-instruct-accelerator)
+- [granite-20b-code-instruct-accelerator](https://huggingface.co/ibm-granite/granite-20b-code-instruct-accelerator)

docs/features/speculative_decoding/n_gram.md

@@ -0,0 +1,27 @@
+# N-Gram Speculation
+
+The following code configures vLLM to use speculative decoding where proposals are generated by
+matching n-grams in the prompt. For more information read [this thread.](https://x.com/joao_gante/status/1747322413006643259)
+
+```python
+from vllm import LLM, SamplingParams
+
+prompts = ["The future of AI is"]
+sampling_params = SamplingParams(temperature=0.8, top_p=0.95)
+
+llm = LLM(
+    model="Qwen/Qwen3-8B",
+    tensor_parallel_size=1,
+    speculative_config={
+        "method": "ngram",
+        "num_speculative_tokens": 5,
+        "prompt_lookup_max": 4,
+    },
+)
+outputs = llm.generate(prompts, sampling_params)
+
+for output in outputs:
+    prompt = output.prompt
+    generated_text = output.outputs[0].text
+    print(f"Prompt: {prompt!r}, Generated text: {generated_text!r}")
+```

docs/features/speculative_decoding/speculators.md

@@ -1,4 +1,7 @@
-# Speculators
+# vLLM-Project/Speculators
+
+![User Flow Light](../../assets/features/speculative_decoding/speculators-user-flow-light.svg#only-light)
+![User Flow Dark](../../assets/features/speculative_decoding/speculators-user-flow-dark.svg#only-dark)
 
 [Speculators](https://docs.vllm.ai/projects/speculators/en/latest/) is a library for accelerating LLM inference through speculative decoding, providing efficient draft model training that integrates seamlessly with vLLM to reduce latency and improve throughput.
 

docs/features/speculative_decoding/suffix.md

@@ -0,0 +1,35 @@
+# Suffix Decoding
+
+The following code configures vLLM to use speculative decoding where proposals are generated using Suffix Decoding ([technical report](https://arxiv.org/abs/2411.04975)).
+
+Like n-gram, Suffix Decoding can generate draft tokens by pattern-matching using the last `n` generated tokens. Unlike n-gram, Suffix Decoding (1) can pattern-match against both the prompt and previous generations, (2) uses frequency counts to propose the most likely continuations, and (3) speculates an adaptive number of tokens for each request at each iteration to get better acceptance rates.
+
+Suffix Decoding can achieve better performance for tasks with high repetition, such as code-editing, agentic loops (e.g. self-reflection, self-consistency), and RL rollouts.
+
+!!! tip "Install Arctic Inference"
+    Suffix Decoding requires [Arctic Inference](https://github.com/snowflakedb/ArcticInference). You can install it with `pip install arctic-inference`.
+
+!!! tip "Suffix Decoding Speculative Tokens"
+    Suffix Decoding will speculate a dynamic number of tokens for each request at each decoding step, so the `num_speculative_tokens` configuration specifies the *maximum* number of speculative tokens. It is suggested to use a high number such as `16` or `32` (default).
+
+```python
+from vllm import LLM, SamplingParams
+
+prompts = ["The future of AI is"]
+sampling_params = SamplingParams(temperature=0.8, top_p=0.95)
+
+llm = LLM(
+    model="Qwen/Qwen3-8B",
+    tensor_parallel_size=1,
+    speculative_config={
+        "method": "suffix",
+        "num_speculative_tokens": 32,
+    },
+)
+outputs = llm.generate(prompts, sampling_params)
+
+for output in outputs:
+    prompt = output.prompt
+    generated_text = output.outputs[0].text
+    print(f"Prompt: {prompt!r}, Generated text: {generated_text!r}")
+```

docs/getting_started/installation/gpu.cuda.inc.md

@@ -297,6 +297,23 @@ You can add any other [engine-args](https://docs.vllm.ai/en/latest/configuration
     RUN uv pip install --system git+https://github.com/huggingface/transformers.git
     ```
 
+#### Running on Systems with Older CUDA Drivers
+
+vLLM's Docker image comes with [CUDA compatibility libraries](https://docs.nvidia.com/deploy/cuda-compatibility/index.html) pre-installed. This allows you to run vLLM on systems with NVIDIA drivers that are older than the CUDA Toolkit version used in the image, but only supports select professional and datacenter NVIDIA GPUs.
+
+To enable this feature, set the `VLLM_ENABLE_CUDA_COMPATIBILITY` environment variable to `1` or `true` when running the container:
+
+```bash
+docker run --runtime nvidia --gpus all \
+    -v ~/.cache/huggingface:/root/.cache/huggingface \
+    -p 8000:8000 \
+    --env "HF_TOKEN=<secret>" \
+    --env "VLLM_ENABLE_CUDA_COMPATIBILITY=1" \
+    vllm/vllm-openai <args...>
+```
+
+This will automatically configure `LD_LIBRARY_PATH` to point to the compatibility libraries before loading PyTorch and other dependencies.
+
 # --8<-- [end:pre-built-images]
 # --8<-- [start:build-image-from-source]
 

docs/getting_started/installation/python_env_setup.inc.md

@@ -1,6 +1,6 @@
 It's recommended to use [uv](https://docs.astral.sh/uv/), a very fast Python environment manager, to create and manage Python environments. Please follow the [documentation](https://docs.astral.sh/uv/#getting-started) to install `uv`. After installing `uv`, you can create a new Python environment using the following commands:
 
 ```bash
-uv venv --python 3.12 --seed
+uv venv --python 3.12 --seed --managed-python
 source .venv/bin/activate
 ```

docs/governance/process.md

@@ -79,13 +79,15 @@ Specially, committers are almost all area owners. They author subsystems, review
 
 For a full list of committers and their respective areas, see the [committers](./committers.md) page.
 
-#### Nomination Process
+#### Committer Proposal Process
 
-Any committer can nominate candidates via our private mailing list:
+Any committer can nominate candidates via our private committer mailing list. The process runs as follows:
 
-1. **Nominate**: Any committer may nominate a candidate by email to the private maintainers’ list, citing evidence mapped to the pre‑existing standards with links to PRs, reviews, RFCs, issues, benchmarks, and adoption evidence.
-2. **Vote**: The lead maintainers will group voices support or concerns. Shared concerns can stop the process. The vote typically last 3 working days. For concerns, committers group discuss the clear criteria for such person to be nominated again. The lead maintainers will make the final decision.
-3. **Confirm**: The lead maintainers send invitation, update CODEOWNERS, assign permissions, add to communications channels (mailing list and Slack).
+1. **Nominate**: A committer sends email to the committer group to nominate a candidate, highlighting the candidate’s contributions (e.g., links to PRs, reviews, RFCs, issues, benchmarks, and adoption evidence) and how they map to the standards below.
+2. **Discuss and vote**: The committer group discusses the nomination, votes, and voices concerns if needed. Shared concerns can stop the process. For concerns, the group discusses clear criteria for the person to be nominated again. Most cases are decided by consensus; in contentious cases, the lead maintainers resolve conflicts and make the decision.
+3. **Feedback period**: After a two-week feedback period (allowing time for any last input or concerns), if no blocking concerns arise and the nominator confirms with lead maintainer group to move forward (via the mailing list or committers slack channel), the nominator sends an invitation to the candidate asking them to open a PR to update their code ownership (e.g., CODEOWNERS and committers list).
+4. **Permissions and onboarding**: In parallel, the lead maintainers assign the necessary permissions in GitHub and add the new member to the committer mailing list, the committer-only Slack channel, and other communications channels as appropriate.
+5. **Finalize**: Once the CODEOWNERS/committer PR is ready and permissions are in place, the PR is merged and the new committer is welcomed.
 
 Committership is highly selective and merit based. The selection criteria requires:
 

docs/mkdocs/hooks/generate_metrics.py

@@ -22,6 +22,7 @@ METRIC_SOURCE_FILES = [
         "path": "vllm/distributed/kv_transfer/kv_connector/v1/nixl_connector.py",
         "output": "nixl_connector.inc.md",
     },
+    {"path": "vllm/v1/metrics/perf.py", "output": "perf.inc.md"},
 ]
 
 

docs/models/pooling_models.md

@@ -382,6 +382,7 @@ ColQwen3 is based on [ColPali](https://arxiv.org/abs/2407.01449), which extends
 |---|---|---|
 | `ColQwen3` | Qwen3-VL | `TomoroAI/tomoro-colqwen3-embed-4b`, `TomoroAI/tomoro-colqwen3-embed-8b` |
 | `OpsColQwen3Model` | Qwen3-VL | `OpenSearch-AI/Ops-Colqwen3-4B`, `OpenSearch-AI/Ops-Colqwen3-8B` |
+| `Qwen3VLNemotronEmbedModel` | Qwen3-VL | `nvidia/nemotron-colembed-vl-4b-v2`, `nvidia/nemotron-colembed-vl-8b-v2` |
 
 Start the server:
 
@@ -389,7 +390,9 @@ Start the server:
 vllm serve TomoroAI/tomoro-colqwen3-embed-4b --max-model-len 4096
 ```
 
-Then you can use the rerank endpoint:
+#### Text-only scoring and reranking
+
+Use the `/rerank` endpoint:
 
 ```shell
 curl -s http://localhost:8000/rerank -H "Content-Type: application/json" -d '{
@@ -403,7 +406,7 @@ curl -s http://localhost:8000/rerank -H "Content-Type: application/json" -d '{
 }'
 ```
 
-Or the score endpoint:
+Or the `/score` endpoint:
 
 ```shell
 curl -s http://localhost:8000/score -H "Content-Type: application/json" -d '{
@@ -413,7 +416,57 @@ curl -s http://localhost:8000/score -H "Content-Type: application/json" -d '{
 }'
 ```
 
-You can also get the raw token embeddings using the pooling endpoint with `token_embed` task:
+#### Multi-modal scoring and reranking (text query × image documents)
+
+The `/score` and `/rerank` endpoints also accept multi-modal inputs directly.
+Pass image documents using the `data_1`/`data_2` (for `/score`) or `documents` (for `/rerank`) fields
+with a `content` list containing `image_url` and `text` parts — the same format used by the
+OpenAI chat completion API:
+
+Score a text query against image documents:
+
+```shell
+curl -s http://localhost:8000/score -H "Content-Type: application/json" -d '{
+    "model": "TomoroAI/tomoro-colqwen3-embed-4b",
+    "data_1": "Retrieve the city of Beijing",
+    "data_2": [
+        {
+            "content": [
+                {"type": "image_url", "image_url": {"url": "data:image/png;base64,<BASE64>"}},
+                {"type": "text", "text": "Describe the image."}
+            ]
+        }
+    ]
+}'
+```
+
+Rerank image documents by a text query:
+
+```shell
+curl -s http://localhost:8000/rerank -H "Content-Type: application/json" -d '{
+    "model": "TomoroAI/tomoro-colqwen3-embed-4b",
+    "query": "Retrieve the city of Beijing",
+    "documents": [
+        {
+            "content": [
+                {"type": "image_url", "image_url": {"url": "data:image/png;base64,<BASE64_1>"}},
+                {"type": "text", "text": "Describe the image."}
+            ]
+        },
+        {
+            "content": [
+                {"type": "image_url", "image_url": {"url": "data:image/png;base64,<BASE64_2>"}},
+                {"type": "text", "text": "Describe the image."}
+            ]
+        }
+    ],
+    "top_n": 2
+}'
+```
+
+#### Raw token embeddings
+
+You can also get the raw token embeddings using the `/pooling` endpoint with `token_embed` task:
 
 ```shell
 curl -s http://localhost:8000/pooling -H "Content-Type: application/json" -d '{
@@ -423,7 +476,7 @@ curl -s http://localhost:8000/pooling -H "Content-Type: application/json" -d '{
 }'
 ```
 
-For **image inputs**, use the chat-style `messages` field so that the vLLM multimodal processor handles them correctly:
+For **image inputs** via the pooling endpoint, use the chat-style `messages` field:
 
 ```shell
 curl -s http://localhost:8000/pooling -H "Content-Type: application/json" -d '{
@@ -440,10 +493,10 @@ curl -s http://localhost:8000/pooling -H "Content-Type: application/json" -d '{
 }'
 ```
 
-Examples can be found here:
+#### Examples
 
 - Multi-vector retrieval: [examples/pooling/token_embed/colqwen3_token_embed_online.py](../../examples/pooling/token_embed/colqwen3_token_embed_online.py)
-- Reranking: [examples/pooling/score/colqwen3_rerank_online.py](../../examples/pooling/score/colqwen3_rerank_online.py)
+- Reranking (text + multi-modal): [examples/pooling/score/colqwen3_rerank_online.py](../../examples/pooling/score/colqwen3_rerank_online.py)
 
 ### BAAI/bge-m3
 

docs/models/supported_models.md

@@ -682,7 +682,7 @@ These models primarily accept the [`LLM.generate`](./generative_models.md#llmgen
 | `Blip2ForConditionalGeneration` | BLIP-2 | T + I<sup>E</sup> | `Salesforce/blip2-opt-2.7b`, `Salesforce/blip2-opt-6.7b`, etc. | ✅︎ | ✅︎ |
 | `ChameleonForConditionalGeneration` | Chameleon | T + I | `facebook/chameleon-7b`, etc. | | ✅︎ |
 | `Cohere2VisionForConditionalGeneration` | Command A Vision | T + I<sup>+</sup> | `CohereLabs/command-a-vision-07-2025`, etc. | | ✅︎ |
-| `DeepseekVLV2ForCausalLM`<sup>^</sup> | DeepSeek-VL2 | T + I<sup>+</sup> | `deepseek-ai/deepseek-vl2-tiny`, `deepseek-ai/deepseek-vl2-small`, `deepseek-ai/deepseek-vl2`, etc. | | ✅︎ |
+| `DeepseekVLV2ForCausalLM` | DeepSeek-VL2 | T + I<sup>+</sup> | `deepseek-ai/deepseek-vl2-tiny`, `deepseek-ai/deepseek-vl2-small`, `deepseek-ai/deepseek-vl2`, etc. | | ✅︎ |
 | `DeepseekOCRForCausalLM` | DeepSeek-OCR | T + I<sup>+</sup> | `deepseek-ai/DeepSeek-OCR`, etc. | ✅︎ | ✅︎ |
 | `DeepseekOCR2ForCausalLM` | DeepSeek-OCR-2 | T + I<sup>+</sup> | `deepseek-ai/DeepSeek-OCR-2`, etc. | ✅︎ | ✅︎ |
 | `Eagle2_5_VLForConditionalGeneration` | Eagle2.5-VL | T + I<sup>E+</sup> | `nvidia/Eagle2.5-8B`, etc. | ✅︎ | ✅︎ |
@@ -762,10 +762,8 @@ Some models are supported only via the [Transformers modeling backend](#transfor
 |--------------|--------|--------|-------------------|-----------------------------|-----------------------------------------|
 | `Emu3ForConditionalGeneration` | Emu3 | T + I | `BAAI/Emu3-Chat-hf` | ✅︎ | ✅︎ |
 
-<sup>^</sup> You need to set the architecture name via `--hf-overrides` to match the one in vLLM.
-&nbsp;&nbsp;&nbsp;&nbsp;• For example, to use DeepSeek-VL2 series models:
-&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;`--hf-overrides '{"architectures": ["DeepseekVLV2ForCausalLM"]}'`
-<sup>E</sup> Pre-computed embeddings can be inputted for this modality.
+<sup>^</sup> You need to set the architecture name via `--hf-overrides` to match the one in vLLM.</br>
+<sup>E</sup> Pre-computed embeddings can be inputted for this modality.</br>
 <sup>+</sup> Multiple items can be inputted per text prompt for this modality.
 
 !!! note
@@ -821,6 +819,7 @@ The following table lists those that are tested in vLLM.
 | Architecture | Models | Inputs | Example HF Models | [LoRA](../features/lora.md) | [PP](../serving/parallelism_scaling.md) |
 |--------------|--------|--------|-------------------|----------------------|---------------------------|
 | `CLIPModel` | CLIP | T / I | `openai/clip-vit-base-patch32`, `openai/clip-vit-large-patch14`, etc. | | |
+| `ColModernVBertForRetrieval` | ColModernVBERT | T / I | `ModernVBERT/colmodernvbert-merged` | | |
 | `LlavaNextForConditionalGeneration`<sup>C</sup> | LLaVA-NeXT-based | T / I | `royokong/e5-v` | | ✅︎ |
 | `Phi3VForCausalLM`<sup>C</sup> | Phi-3-Vision-based | T + I | `TIGER-Lab/VLM2Vec-Full` | | ✅︎ |
 | `Qwen3VLForConditionalGeneration`<sup>C</sup> | Qwen3-VL | T + I + V | `Qwen/Qwen3-VL-Embedding-2B`, etc. | ✅︎ | ✅︎ |

docs/usage/metrics.md

@@ -45,6 +45,12 @@ The following metrics are exposed:
 
 --8<-- "docs/generated/metrics/nixl_connector.inc.md"
 
+## Model Flops Utilization (MFU) Performance Metrics
+
+These metrics are available via `--enable-mfu-metrics`:
+
+--8<-- "docs/generated/metrics/perf.inc.md"
+
 ## Deprecation Policy
 
 Note: when metrics are deprecated in version `X.Y`, they are hidden in version `X.Y+1`

docs/usage/troubleshooting.md

@@ -318,7 +318,32 @@ This indicates vLLM failed to initialize the NCCL communicator, possibly due to
 
 ## CUDA error: the provided PTX was compiled with an unsupported toolchain
 
-If you see an error like `RuntimeError: CUDA error: the provided PTX was compiled with an unsupported toolchain.`, it means that the CUDA PTX in vLLM's wheels was compiled with a toolchain unsupported by your system. The released vLLM wheels have to be compiled with a specific version of CUDA toolkit, and the compiled code might fail to run on lower versions of CUDA drivers. Read [cuda compatibility](https://docs.nvidia.com/deploy/cuda-compatibility/) for more details. The solution is to install `cuda-compat` package from your package manager. For example, on Ubuntu, you can run `sudo apt-get install cuda-compat-12-9`, and then add `export LD_LIBRARY_PATH=/usr/local/cuda-12.9/compat:$LD_LIBRARY_PATH` to your `.bashrc` file. When successfully installed, you should see that the output of `nvidia-smi` will show `CUDA Version: 12.9`. Note that we use CUDA 12.9 as an example here, you may want to install a higher version of cuda-compat package in case vLLM's default CUDA version goes higher.
+If you see an error like `RuntimeError: CUDA error: the provided PTX was compiled with an unsupported toolchain`, it means that the CUDA PTX in vLLM's wheels was compiled with a toolchain unsupported by your system. This section also applies if you get the error `RuntimeError: The NVIDIA driver on your system is too old`.
+
+The released vLLM wheels are compiled with a specific version of CUDA toolkit, and the compiled code might fail to run on lower versions of CUDA drivers. Read [CUDA compatibility](https://docs.nvidia.com/deploy/cuda-compatibility/) for more details. **This is only supported on select professional and datacenter NVIDIA GPUs.**
+
+If you are using the vLLM official Docker image, you can solve this by adding `-e VLLM_ENABLE_CUDA_COMPATIBILITY=1` to your `docker run` command. This will enable the pre-installed CUDA forward compatibility libraries.
+
+If you are running vLLM outside of Docker, the solution is to install the `cuda-compat` package from your package manager with the [CUDA repository](https://docs.nvidia.com/cuda/cuda-installation-guide-linux/) enabled. For example, on Ubuntu, you can run `sudo apt-get install cuda-compat-12-9`, and then set `export VLLM_ENABLE_CUDA_COMPATIBILITY=1` and `export VLLM_CUDA_COMPATIBILITY_PATH="/usr/local/cuda-12.9/compat"`.
+
+On Conda, you can install the `conda-forge::cuda-compat` package (e.g., `conda install -c conda-forge cuda-compat=12.9`), then after activating the environment, set `export VLLM_ENABLE_CUDA_COMPATIBILITY=1` and `export VLLM_CUDA_COMPATIBILITY_PATH="${CONDA_PREFIX}/cuda-compat"`.
+
+You can verify the configuration works by running a minimal Python script that initializes CUDA via vLLM:
+
+```bash
+export VLLM_ENABLE_CUDA_COMPATIBILITY=1
+export VLLM_CUDA_COMPATIBILITY_PATH="/usr/local/cuda-12.9/compat"
+
+python3 - << 'EOF'
+import vllm
+import torch
+
+print(f"CUDA available: {torch.cuda.is_available()}")
+print(f"CUDA device count: {torch.cuda.device_count()}")
+EOF
+```
+
+Note that we use CUDA 12.9 as an example here, and you may want to install a higher version of cuda-compat package in case vLLM's default CUDA version goes higher.
 
 ## ptxas fatal: Value 'sm_110a' is not defined for option 'gpu-name'
 

examples/offline_inference/kv_load_failure_recovery/decode_example.py

@@ -42,6 +42,7 @@ def main():
                 "async_load": args.async_load,
             },
             kv_connector_module_path="load_recovery_example_connector",
+            kv_load_failure_policy="recompute",
         )
         out_file = (
             "async_decode_recovered_output.txt"

examples/offline_inference/new_weight_syncing/rlhf_async_new_apis.py

@@ -26,14 +26,12 @@ workloads. Residual GPU activity interferes with vLLM memory profiling and
 causes unexpected behavior.
 """
 
-import os
+import asyncio
 import uuid
 from dataclasses import asdict
 
 import ray
 import torch
-from ray.util.placement_group import placement_group
-from ray.util.scheduling_strategies import PlacementGroupSchedulingStrategy
 from transformers import AutoModelForCausalLM, AutoTokenizer
 
 import vllm
@@ -51,14 +49,15 @@ from vllm.distributed.weight_transfer.nccl_engine import (
 from vllm.utils.network_utils import get_ip, get_open_port
 from vllm.v1.executor import Executor
 
-MODEL_NAME = "facebook/opt-125m"
+MODEL_NAME_V1 = "Qwen/Qwen3-1.7B-Base"
+MODEL_NAME_V2 = "Qwen/Qwen3-1.7B"
+PAUSE_TOKEN_THRESHOLD = 10
 
 
 class MyLLM(vllm.AsyncLLMEngine):
     """Configure the vLLM worker for Ray placement group execution."""
 
     def __init__(self, **kwargs):
-        os.environ["VLLM_RAY_BUNDLE_INDICES"] = "0,1"
         engine_args = vllm.AsyncEngineArgs(**kwargs)
         vllm_config = engine_args.create_engine_config()
         executor_class = Executor.get_class(vllm_config)
@@ -68,26 +67,44 @@ class MyLLM(vllm.AsyncLLMEngine):
             log_requests=engine_args.enable_log_requests,
             log_stats=not engine_args.disable_log_stats,
         )
+        self._generation_paused = False
+        self._request_pause_flag = False
 
-    async def generate_with_retry(
+    async def do_generate(
         self, prompt_token_ids: list[int], sampling_params: vllm.SamplingParams
-    ) -> vllm.RequestOutput:
-        finish_reason = "abort"
-        while finish_reason == "abort":
-            async for request_output in self.generate(
-                {"prompt_token_ids": prompt_token_ids},
-                sampling_params,
-                request_id=str(uuid.uuid4()),
+    ) -> tuple[vllm.RequestOutput, int]:
+        """Generate a single request, setting the request pause flag once the
+        token count reaches the threshold.
+
+        Returns (output, pause_token_index). pause_token_index is the number
+        of tokens generated before the weight change, or -1 if no pause.
+        """
+        pause_token_index = -1
+        prev_token_count = 0
+        async for request_output in self.generate(
+            {"prompt_token_ids": prompt_token_ids},
+            sampling_params,
+            request_id=str(uuid.uuid4()),
+        ):
+            output = request_output
+            cur_token_count = len(output.outputs[0].token_ids)
+            if (
+                cur_token_count >= PAUSE_TOKEN_THRESHOLD
+                and not self._request_pause_flag
             ):
-                output = request_output
-            finish_reason = output.outputs[0].finish_reason
-            if finish_reason == "abort":
-                print(
-                    f"ABORT, prompt_token_ids: {prompt_token_ids}, "
-                    f"generated token_ids: {list(output.outputs[0].token_ids)}"
-                )
-            prompt_token_ids = prompt_token_ids + list(output.outputs[0].token_ids)
-        return output
+                self._request_pause_flag = True
+            if self._generation_paused and pause_token_index == -1:
+                pause_token_index = prev_token_count
+            prev_token_count = cur_token_count
+        return output, pause_token_index
+
+    async def pause_after_n_tokens(self):
+        """Wait for any request to set the pause flag, then pause."""
+        while not self._request_pause_flag:
+            await asyncio.sleep(0)
+        await super().pause_generation(mode="keep")
+        await asyncio.sleep(0.2)
+        self._generation_paused = True
 
 
 @ray.remote(num_gpus=1)
@@ -95,6 +112,14 @@ class TrainModel:
     """Ray actor that wraps the training model on a dedicated GPU."""
 
     def __init__(self, model_name: str):
+        from vllm.model_executor.layers.batch_invariant import (
+            init_batch_invariance,
+        )
+        from vllm.v1.attention.backends.registry import AttentionBackendEnum
+
+        # need to init all env vars for batch invariance which affect nccl ops
+        init_batch_invariance(AttentionBackendEnum.FLASH_ATTN)
+
         self.model = AutoModelForCausalLM.from_pretrained(
             model_name, dtype=torch.bfloat16
         ).to("cuda:0")
@@ -133,70 +158,80 @@ class TrainModel:
             packed=packed,
         )
 
+    @torch.inference_mode()
+    def generate(self, token_ids: list[int], max_new_tokens: int) -> list[int]:
+        """Greedy-decode max_new_tokens from the given context."""
+        input_ids = torch.tensor([token_ids], device="cuda:0")
+        output = self.model.generate(
+            input_ids,
+            max_new_tokens=max_new_tokens,
+            do_sample=False,
+        )
+        new_token_ids = output[0, len(token_ids) :].tolist()
+        return new_token_ids
 
-# Initialize Ray and set the visible devices. The vLLM engine will
-# be placed on GPUs 1 and 2.
-ray.init()
+
+ray.init(
+    runtime_env={
+        "env_vars": {
+            # enable batch invariance for deterministic outputs
+            "VLLM_BATCH_INVARIANT": "1",
+            # prevent ray from setting CUDA_VISIBLE_DEVICES
+            "RAY_EXPERIMENTAL_NOSET_CUDA_ENV_VAR": "1",
+        }
+    }
+)
 
 # Launch the training model actor. Ray's resource scheduler will allocate
 # 1 GPU (via num_gpus=1 in the decorator), ensuring pg_inference gets different GPUs.
-train_model = TrainModel.remote(MODEL_NAME)
-
-# Create a placement group that reserves GPU 1–2 for the vLLM inference engine.
-# Learn more about Ray placement groups:
-# https://docs.ray.io/en/latest/placement-groups.html
-
-pg_inference = placement_group([{"GPU": 1, "CPU": 0}] * 2)
-ray.get(pg_inference.ready())
-scheduling_inference = PlacementGroupSchedulingStrategy(
-    placement_group=pg_inference,
-    placement_group_capture_child_tasks=True,
-    placement_group_bundle_index=0,
-)
+train_model = TrainModel.remote(MODEL_NAME_V2)
 
 # Launch the vLLM inference engine. The `enforce_eager` flag reduces
 # start-up latency.
-# Note: Weight transfer APIs (init_weight_transfer_engine, update_weights)
-# are now native to vLLM workers.
+# With data_parallel_backend="ray", vLLM's CoreEngineActorManager creates
+# its own placement groups internally for each DP rank, so we must NOT
+# create an outer placement group (it would reserve GPUs and hide them
+# from the internal DP resource check).
 llm = ray.remote(
     num_cpus=0,
     num_gpus=0,
-    scheduling_strategy=scheduling_inference,
 )(MyLLM).remote(
-    model=MODEL_NAME,
+    model=MODEL_NAME_V1,
     enforce_eager=True,
-    tensor_parallel_size=2,
+    max_model_len=8192,
     distributed_executor_backend="ray",
-    load_format="dummy",
+    attention_backend="FLASH_ATTN",
+    gpu_memory_utilization=0.75,
     weight_transfer_config=WeightTransferConfig(backend="nccl"),
 )
 
-# Generate text from the prompts.
-prompts = [
-    "My name is",
+PROMPTS = [
     "The president of the United States is",
     "The capital of France is",
-    "The future of AI is",
+    "The largest ocean on Earth is",
+    "The speed of light in a vacuum is",
+    "The chemical formula for water is",
+    "The tallest mountain in the world is",
+    "The first person to walk on the moon was",
+    "The Great Wall of China was built to",
+    "Photosynthesis is the process by which",
+    "The theory of general relativity was proposed by",
+    "The boiling point of water at sea level is",
+    "The largest planet in our solar system is",
+    "DNA stands for deoxyribonucleic acid and it",
 ]
 
-# Tokenize prompts to token IDs
-tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
-prompt_token_ids_list = [
-    tokenizer.encode(prompt, add_special_tokens=False) for prompt in prompts
+tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME_V1)
+batch_prompt_token_ids = [
+    tokenizer.encode(prompt, add_special_tokens=False) for prompt in PROMPTS
 ]
 
-sampling_params = [
-    SamplingParams(temperature=0, max_tokens=2),
-    SamplingParams(temperature=0, max_tokens=32),
-    SamplingParams(temperature=0, max_tokens=32),
-    SamplingParams(temperature=0, max_tokens=32),
-]
 
 # Set up the communication channel between the training process and the
 # inference engine.
 master_address, master_port = ray.get(train_model.get_master_address_and_port.remote())
 
-world_size = 3  # 1 trainer + 2 inference workers (tensor_parallel_size=2)
+world_size = 2  # 1 trainer + 1 inference worker
 inference_handle = llm.init_weight_transfer_engine.remote(
     WeightTransferInitRequest(
         init_info=asdict(
@@ -215,22 +250,28 @@ train_handle = train_model.init_weight_transfer_group.remote(world_size)
 ray.get([train_handle, inference_handle])
 
 
-generation_futures = [
-    llm.generate_with_retry.remote(prompt_token_ids, params)
-    for prompt_token_ids, params in zip(prompt_token_ids_list, sampling_params)
-]
+N_NEW_TOKENS = 100
 
-finished, pending = ray.wait(generation_futures, num_returns=1)
-
-# Pause generation in preparation for weight sync
-ray.get(llm.pause_generation.remote(wait_for_inflight_requests=False))
-
-# Synchronize the updated weights to the inference engine using batched API.
-# Collect all weight metadata from the training actor
+# Collect weight metadata once
 names, dtype_names, shapes = ray.get(train_model.get_weight_metadata.remote())
 
-# Issue update_weights call with NCCL-specific update info
-# packed=True enables efficient batched tensor broadcasting
+# ── Phase 1: concurrent requests with weight sync ───────────────────
+print(f"\n{'=' * 50}")
+print(f"Prompts ({len(PROMPTS)}):")
+for p in PROMPTS:
+    print(f"  - {p!r}")
+print(f"{'=' * 50}")
+
+sampling_params = SamplingParams(
+    temperature=0, max_tokens=PAUSE_TOKEN_THRESHOLD + N_NEW_TOKENS
+)
+
+gen_futures = [
+    llm.do_generate.remote(ptids, sampling_params) for ptids in batch_prompt_token_ids
+]
+
+ray.get(llm.pause_after_n_tokens.remote())
+
 inference_handle = llm.update_weights.remote(
     WeightTransferUpdateRequest(
         update_info=asdict(
@@ -243,41 +284,76 @@ inference_handle = llm.update_weights.remote(
         )
     )
 )
-
-# Broadcast all weights from trainer using the weight transfer API
 train_handle = train_model.broadcast_weights.remote(packed=True)
 ray.get([train_handle, inference_handle])
 
-# Resume generation since weight sync is complete
 ray.get(llm.resume_generation.remote())
+results = ray.get(gen_futures)
 
-# Get outputs separately - finished completed before pause, pending were paused/resumed
-finished_outputs = ray.get(finished)
-pending_outputs = ray.get(pending)
+for i, (output, pause_idx) in enumerate(results):
+    all_token_ids = list(output.outputs[0].token_ids)
+    before_text = tokenizer.decode(all_token_ids[:pause_idx])
+    after_text = tokenizer.decode(all_token_ids[pause_idx:])
+    print(f"\n  Request {i} ({PROMPTS[i]!r}):")
+    print(f"    Old weights ({pause_idx} tokens): {before_text!r}")
+    n_after = len(all_token_ids) - pause_idx
+    print(f"    New weights ({n_after} tokens): {after_text!r}")
 
-# Requests that finished before the pause: all generation used original weights
-print("-" * 50)
-print("Requests that completed BEFORE weight change:")
-print("-" * 50)
-for output in finished_outputs:
-    prompt_text = tokenizer.decode(output.prompt_token_ids)
-    print(f"Prompt: {prompt_text!r}")
-    print(f"Generated (with original weights): {output.outputs[0].text!r}")
-    print("-" * 50)
+# ── Phase 2: validate with a fresh V2 vLLM instance ────────────────
+print(f"\n{'=' * 50}")
+print("VALIDATION: comparing weight-synced vLLM with fresh V2 instance")
+print(f"{'=' * 50}")
 
-# Requests that were paused mid-generation: some text before, some after weight change
-print("Requests that were PAUSED and RESUMED after weight change:")
-print("-" * 50)
-for output in pending_outputs:
-    # Decode the full prompt token IDs (original + generated before pause)
-    full_prompt_text = tokenizer.decode(output.prompt_token_ids)
-    # Find the original prompt by checking which one this output started with
-    original_prompt = next(p for p in prompts if full_prompt_text.startswith(p))
-    # output.prompt_token_ids contains original prompt + tokens generated before pause
-    # output.outputs[0].text is what was generated after resuming with new weights
-    text_before_pause = full_prompt_text[len(original_prompt) :]
-    text_after_pause = output.outputs[0].text
-    print(f"Original prompt: {original_prompt!r}")
-    print(f"Generated before weight change: {text_before_pause!r}")
-    print(f"Generated after weight change: {text_after_pause!r}")
-    print("-" * 50)
+ray.get(llm.shutdown.remote())
+ray.kill(llm)
+ray.kill(train_model)
+
+llm_v2 = ray.remote(
+    num_cpus=0,
+    num_gpus=0,
+)(MyLLM).remote(
+    model=MODEL_NAME_V2,
+    enforce_eager=True,
+    max_model_len=8192,
+    gpu_memory_utilization=0.75,
+    distributed_executor_backend="ray",
+    attention_backend="FLASH_ATTN",
+)
+
+val_futures = [
+    llm_v2.do_generate.remote(
+        list(output.prompt_token_ids) + list(output.outputs[0].token_ids)[:pause_idx],
+        SamplingParams(
+            temperature=0, max_tokens=len(output.outputs[0].token_ids) - pause_idx
+        ),
+    )
+    for output, pause_idx in results
+]
+val_results = ray.get(val_futures)
+
+all_pass = True
+for i, ((output, pause_idx), (val_output, _)) in enumerate(zip(results, val_results)):
+    expected = list(output.outputs[0].token_ids)[pause_idx:]
+    actual = list(val_output.outputs[0].token_ids)
+    match = actual == expected
+
+    if match:
+        print(f"  [PASS] {PROMPTS[i]!r}")
+    else:
+        all_pass = False
+        print(f"  [FAIL] {PROMPTS[i]!r}")
+        print(f"         weight-synced vLLM: {tokenizer.decode(expected)!r}")
+        print(f"         V2 vLLM:           {tokenizer.decode(actual)!r}")
+        for j, (e, a) in enumerate(zip(expected, actual)):
+            if e != a:
+                print(
+                    f"         first divergence at output token {j}: "
+                    f"expected {e} ({tokenizer.decode([e])!r}) vs "
+                    f"actual {a} ({tokenizer.decode([a])!r})"
+                )
+                break
+
+ray.get(llm_v2.shutdown.remote())
+ray.kill(llm_v2)
+assert all_pass, "Some prompts failed validation, see above for details"
+print("=" * 50)

examples/online_serving/kv_events_subscriber.py

@@ -37,6 +37,12 @@ class BlockStored(KVCacheEvent):
     medium: str | None
     lora_name: str | None
 
+    extra_keys: list[tuple[Any, ...] | None] | None = None
+    """Extra keys used in block hash computation, one entry per block in
+    block_hashes. Each entry contains MM identifiers, LoRA name, cache_salt,
+    prompt embeddings data, etc. for that specific block.
+    """
+
 
 class BlockRemoved(KVCacheEvent):
     block_hashes: list[ExternalBlockHash]

examples/online_serving/opentelemetry/README.md

@@ -1,14 +1,6 @@
 # Setup OpenTelemetry POC
 
-1. Install OpenTelemetry packages:
-
-    ```bash
-    pip install \
-      'opentelemetry-sdk>=1.26.0,<1.27.0' \
-      'opentelemetry-api>=1.26.0,<1.27.0' \
-      'opentelemetry-exporter-otlp>=1.26.0,<1.27.0' \
-      'opentelemetry-semantic-conventions-ai>=0.4.1,<0.5.0'
-    ```
+> **Note:** The core OpenTelemetry packages (`opentelemetry-sdk`, `opentelemetry-api`, `opentelemetry-exporter-otlp`, `opentelemetry-semantic-conventions-ai`) are bundled with vLLM. Manual installation is not required.
 
 1. Start Jaeger in a docker container:
 

examples/pooling/score/colmodernvbert_rerank_online.py

@@ -0,0 +1,166 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""
+Example of using ColModernVBERT late interaction model for reranking.
+
+ColModernVBERT is a multi-modal ColBERT-style model combining a SigLIP
+vision encoder with a ModernBERT text encoder. It produces per-token
+embeddings and uses MaxSim scoring for retrieval and reranking.
+Supports both text and image inputs.
+
+Start the server with:
+    vllm serve ModernVBERT/colmodernvbert-merged --max-model-len 8192
+
+Then run this script:
+    python colmodernvbert_rerank_online.py
+"""
+
+import requests
+
+MODEL = "ModernVBERT/colmodernvbert-merged"
+BASE_URL = "http://127.0.0.1:8000"
+
+headers = {"accept": "application/json", "Content-Type": "application/json"}
+
+IMAGE_URL = "https://upload.wikimedia.org/wikipedia/commons/thumb/4/47/PNG_transparency_demonstration_1.png/300px-PNG_transparency_demonstration_1.png"  # noqa: E501
+
+
+def rerank_text():
+    """Text-only reranking via /rerank endpoint."""
+    print("=" * 60)
+    print("1. Text reranking (/rerank)")
+    print("=" * 60)
+
+    data = {
+        "model": MODEL,
+        "query": "What is machine learning?",
+        "documents": [
+            "Machine learning is a subset of artificial intelligence.",
+            "Python is a programming language.",
+            "Deep learning uses neural networks for complex tasks.",
+            "The weather today is sunny.",
+        ],
+    }
+
+    response = requests.post(f"{BASE_URL}/rerank", headers=headers, json=data)
+
+    if response.status_code == 200:
+        result = response.json()
+        print("\n  Ranked documents (most relevant first):")
+        for item in result["results"]:
+            doc_idx = item["index"]
+            score = item["relevance_score"]
+            print(f"    [{score:.4f}] {data['documents'][doc_idx]}")
+    else:
+        print(f"  Request failed: {response.status_code}")
+        print(f"  {response.text[:300]}")
+
+
+def score_text():
+    """Text-only scoring via /score endpoint."""
+    print()
+    print("=" * 60)
+    print("2. Text scoring (/score)")
+    print("=" * 60)
+
+    query = "What is the capital of France?"
+    documents = [
+        "The capital of France is Paris.",
+        "Berlin is the capital of Germany.",
+        "Python is a programming language.",
+    ]
+
+    data = {
+        "model": MODEL,
+        "text_1": query,
+        "text_2": documents,
+    }
+
+    response = requests.post(f"{BASE_URL}/score", headers=headers, json=data)
+
+    if response.status_code == 200:
+        result = response.json()
+        print(f"\n  Query: {query}\n")
+        for item in result["data"]:
+            idx = item["index"]
+            score = item["score"]
+            print(f"    Doc {idx} (score={score:.4f}): {documents[idx]}")
+    else:
+        print(f"  Request failed: {response.status_code}")
+        print(f"  {response.text[:300]}")
+
+
+def score_text_top_n():
+    """Text reranking with top_n filtering via /rerank endpoint."""
+    print()
+    print("=" * 60)
+    print("3. Text reranking with top_n=2 (/rerank)")
+    print("=" * 60)
+
+    data = {
+        "model": MODEL,
+        "query": "What is the capital of France?",
+        "documents": [
+            "The capital of France is Paris.",
+            "Berlin is the capital of Germany.",
+            "Python is a programming language.",
+            "The Eiffel Tower is in Paris.",
+        ],
+        "top_n": 2,
+    }
+
+    response = requests.post(f"{BASE_URL}/rerank", headers=headers, json=data)
+
+    if response.status_code == 200:
+        result = response.json()
+        print(f"\n  Top {data['top_n']} results:")
+        for item in result["results"]:
+            doc_idx = item["index"]
+            score = item["relevance_score"]
+            print(f"    [{score:.4f}] {data['documents'][doc_idx]}")
+    else:
+        print(f"  Request failed: {response.status_code}")
+        print(f"  {response.text[:300]}")
+
+
+def rerank_multimodal():
+    """Multimodal reranking with text and image documents via /rerank."""
+    print()
+    print("=" * 60)
+    print("4. Multimodal reranking: text query vs image document (/rerank)")
+    print("=" * 60)
+
+    data = {
+        "model": MODEL,
+        "query": "A colorful logo with transparency",
+        "documents": [
+            {"content": [{"type": "image_url", "image_url": {"url": IMAGE_URL}}]},
+            "Python is a programming language.",
+            "The weather today is sunny.",
+        ],
+    }
+
+    response = requests.post(f"{BASE_URL}/rerank", headers=headers, json=data)
+
+    if response.status_code == 200:
+        result = response.json()
+        print("\n  Ranked documents (most relevant first):")
+        labels = ["[image]", "Python doc", "Weather doc"]
+        for item in result["results"]:
+            doc_idx = item["index"]
+            score = item["relevance_score"]
+            print(f"    [{score:.4f}] {labels[doc_idx]}")
+    else:
+        print(f"  Request failed: {response.status_code}")
+        print(f"  {response.text[:300]}")
+
+
+def main():
+    rerank_text()
+    score_text()
+    score_text_top_n()
+    rerank_multimodal()
+
+
+if __name__ == "__main__":
+    main()