fix build

Signed-off-by: Lucas Wilkinson <lwilkinson@neuralmagic.com>

.buildkite/check-wheel-size.py

@@ -1,5 +1,4 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 import os
 import sys

.buildkite/generate_index.py

@@ -1,5 +1,4 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 import argparse
 import os

.buildkite/lm-eval-harness/conftest.py

@@ -1,5 +1,4 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 from pathlib import Path
 
 import pytest

.buildkite/lm-eval-harness/test_lm_eval_correctness.py

@@ -1,5 +1,4 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 """
 LM eval harness on model to compare vs HF baseline computed offline.
 Configs are found in configs/$MODEL.yaml

.buildkite/nightly-benchmarks/README.md

@@ -113,7 +113,7 @@ WARNING: The benchmarking script will save json results by itself, so please do
 
 ### Visualizing the results
 
-The `convert-results-json-to-markdown.py` helps you put the benchmarking results inside a markdown table, by formatting [descriptions.md](performance-benchmarks-descriptions.md) with real benchmarking results.
+The `convert-results-json-to-markdown.py` helps you put the benchmarking results inside a markdown table, by formatting [descriptions.md](tests/descriptions.md) with real benchmarking results.
 You can find the result presented as a table inside the `buildkite/performance-benchmark` job page.
 If you do not see the table, please wait till the benchmark finish running.
 The json version of the table (together with the json version of the benchmark) will be also attached to the markdown file.

.buildkite/nightly-benchmarks/nightly-annotation.md

@@ -16,7 +16,7 @@ Please download the visualization scripts in the post
   - Download `nightly-benchmarks.zip`.
   - In the same folder, run the following code:
 
-  ```bash
+  ```console
   export HF_TOKEN=<your HF token>
   apt update
   apt install -y git

.buildkite/nightly-benchmarks/scripts/convert-results-json-to-markdown.py

@@ -1,5 +1,4 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 import json
 import os

.buildkite/nightly-benchmarks/scripts/download-tokenizer.py

@@ -1,5 +1,4 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 import argparse
 

.buildkite/nightly-benchmarks/scripts/generate-nightly-markdown.py

@@ -1,5 +1,4 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 import argparse
 import json

.buildkite/nightly-benchmarks/scripts/get-lmdeploy-modelname.py

@@ -1,5 +1,4 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 from lmdeploy.serve.openai.api_client import APIClient
 

.buildkite/nightly-benchmarks/scripts/summary-nightly-results.py

@@ -1,5 +1,4 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 import datetime
 import json

.buildkite/release-pipeline.yaml

@@ -1,6 +1,5 @@
 steps:
   - label: "Build wheel - CUDA 12.8"
-    id: build-wheel-cuda-12-8
     agents:
       queue: cpu_queue_postmerge
     commands:
@@ -12,7 +11,6 @@ steps:
       DOCKER_BUILDKIT: "1"
 
   - label: "Build wheel - CUDA 12.6"
-    id: build-wheel-cuda-12-6
     agents:
       queue: cpu_queue_postmerge
     commands:
@@ -30,7 +28,6 @@ steps:
 
   - label: "Build wheel - CUDA 11.8"
     # depends_on: block-build-cu118-wheel
-    id: build-wheel-cuda-11-8
     agents:
       queue: cpu_queue_postmerge
     commands:
@@ -47,7 +44,6 @@ steps:
 
   - label: "Build release image"
     depends_on: block-release-image-build
-    id: build-release-image
     agents:
       queue: cpu_queue_postmerge
     commands:
@@ -55,18 +51,6 @@ steps:
       - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=12.8.1 --tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT --target vllm-openai --progress plain -f docker/Dockerfile ."
       - "docker push public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT"
 
-  - label: "Annotate release workflow"
-    depends_on:
-      - build-release-image
-      - build-wheel-cuda-12-8
-      - build-wheel-cuda-12-6
-      - build-wheel-cuda-11-8
-    id: annotate-release-workflow
-    agents:
-      queue: cpu_queue_postmerge
-    commands:
-      - "bash .buildkite/scripts/annotate-release.sh"
-
   - label: "Build and publish TPU release image"
     depends_on: ~
     if: build.env("NIGHTLY") == "1"
@@ -86,10 +70,9 @@ steps:
       DOCKER_BUILDKIT: "1"
 
   - input: "Provide Release version here"
-    id: input-release-version
     fields:
       - text: "What is the release version?"
-        key: release-version
+        key: "release-version"
 
   - block: "Build CPU release image"
     key: block-cpu-release-image-build
@@ -102,7 +85,6 @@ steps:
     commands:
       - "aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin public.ecr.aws/q9t5s3a7"
       - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg GIT_REPO_CHECK=1 --tag public.ecr.aws/q9t5s3a7/vllm-cpu-release-repo:$(buildkite-agent meta-data get release-version) --tag public.ecr.aws/q9t5s3a7/vllm-cpu-release-repo:latest --progress plain --target vllm-openai -f docker/Dockerfile.cpu ."
-      - "docker push public.ecr.aws/q9t5s3a7/vllm-cpu-release-repo:latest"
       - "docker push public.ecr.aws/q9t5s3a7/vllm-cpu-release-repo:$(buildkite-agent meta-data get release-version)"
     env:
       DOCKER_BUILDKIT: "1"
@@ -118,7 +100,6 @@ steps:
     commands:
       - "aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin public.ecr.aws/q9t5s3a7"
       - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg GIT_REPO_CHECK=1 --tag public.ecr.aws/q9t5s3a7/vllm-neuron-release-repo:$(buildkite-agent meta-data get release-version) --tag public.ecr.aws/q9t5s3a7/vllm-neuron-release-repo:latest --progress plain -f docker/Dockerfile.neuron ."
-      - "docker push public.ecr.aws/q9t5s3a7/vllm-neuron-release-repo:latest"
       - "docker push public.ecr.aws/q9t5s3a7/vllm-neuron-release-repo:$(buildkite-agent meta-data get release-version)"
     env:
       DOCKER_BUILDKIT: "1"

.buildkite/scripts/annotate-release.sh

@@ -1,31 +0,0 @@
-#!/bin/bash
-
-set -ex
-
-# Get release version and strip leading 'v' if present
-RELEASE_VERSION=$(buildkite-agent meta-data get release-version | sed 's/^v//')
-
-if [ -z "$RELEASE_VERSION" ]; then
-  echo "Error: RELEASE_VERSION is empty. 'release-version' metadata might not be set or is invalid."
-  exit 1
-fi
-
-buildkite-agent annotate --style 'info' --context 'release-workflow' << EOF
-To download the wheel:
-\`\`\`
-aws s3 cp s3://vllm-wheels/${RELEASE_VERSION}/vllm-${RELEASE_VERSION}-cp38-abi3-manylinux1_x86_64.whl .
-aws s3 cp s3://vllm-wheels/${RELEASE_VERSION}+cu126/vllm-${RELEASE_VERSION}+cu126-cp38-abi3-manylinux1_x86_64.whl .
-aws s3 cp s3://vllm-wheels/${RELEASE_VERSION}+cu118/vllm-${RELEASE_VERSION}+cu118-cp38-abi3-manylinux1_x86_64.whl . 
-\`\`\`
-
-To download and upload the image:
-
-\`\`\`
-docker pull public.ecr.aws/q9t5s3a7/vllm-release-repo:${BUILDKITE_COMMIT}
-docker tag public.ecr.aws/q9t5s3a7/vllm-release-repo:${BUILDKITE_COMMIT} vllm/vllm-openai
-docker tag vllm/vllm-openai vllm/vllm-openai:latest
-docker tag vllm/vllm-openai vllm/vllm-openai:v${RELEASE_VERSION}
-docker push vllm/vllm-openai:latest
-docker push vllm/vllm-openai:v${RELEASE_VERSION}
-\`\`\`
-EOF 

.buildkite/scripts/ci-clean-log.sh

@@ -1,17 +0,0 @@
-#!/bin/bash
-# Usage: ./ci_clean_log.sh ci.log
-# This script strips timestamps and color codes from CI log files.
-
-# Check if argument is given
-if [ $# -lt 1 ]; then
-    echo "Usage: $0 ci.log"
-    exit 1
-fi
-
-INPUT_FILE="$1"
-
-# Strip timestamps
-sed -i 's/^\[[0-9]\{4\}-[0-9]\{2\}-[0-9]\{2\}T[0-9]\{2\}:[0-9]\{2\}:[0-9]\{2\}Z\] //' "$INPUT_FILE"
-
-# Strip colorization
-sed -i -r 's/\x1B\[[0-9;]*[mK]//g' "$INPUT_FILE"

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

@@ -94,10 +94,6 @@ if [[ $commands == *"pytest -v -s compile/test_basic_correctness.py"* ]]; then
   commands=${commands//"pytest -v -s compile/test_basic_correctness.py"/"VLLM_USE_TRITON_FLASH_ATTN=0 pytest -v -s compile/test_basic_correctness.py"}
 fi
 
-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} \

.buildkite/scripts/hardware_ci/run-cpu-test-ppc64le.sh

@@ -7,7 +7,6 @@ set -ex
 # Setup cleanup
 remove_docker_container() {
   if [[ -n "$container_id" ]]; then
-      podman stop --all -t0
       podman rm -f "$container_id" || true
   fi
   podman system prune -f
@@ -38,7 +37,7 @@ function cpu_tests() {
     pytest -v -s tests/models/language/generation/test_common.py::test_models[False-5-32-facebook/opt-125m]
     pytest -v -s tests/models/language/generation/test_common.py::test_models[False-5-32-google/gemma-1.1-2b-it]
     pytest -v -s tests/models/language/pooling/test_classification.py::test_models[float-jason9693/Qwen2.5-1.5B-apeach]
-    pytest -v -s tests/models/language/pooling/test_embedding.py -m cpu_model"
+    pytest -v -s tests/models/language/pooling/test_embedding.py::test_models[half-BAAI/bge-base-en-v1.5]"
 }
 
 # All of CPU tests are expected to be finished less than 40 mins.

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

@@ -6,82 +6,75 @@ set -ex
 
 # allow to bind to different cores
 CORE_RANGE=${CORE_RANGE:-48-95}
-OMP_CORE_RANGE=${OMP_CORE_RANGE:-48-95}
 NUMA_NODE=${NUMA_NODE:-1}
 
-export CMAKE_BUILD_PARALLEL_LEVEL=32
-
 # Setup cleanup
 remove_docker_container() { 
     set -e; 
-    docker rm -f cpu-test-"$NUMA_NODE" cpu-test-"$NUMA_NODE"-avx2 || true; 
+    docker rm -f cpu-test-"$BUILDKITE_BUILD_NUMBER"-"$NUMA_NODE" cpu-test-"$BUILDKITE_BUILD_NUMBER"-avx2-"$NUMA_NODE" || true; 
+    docker image rm cpu-test-"$BUILDKITE_BUILD_NUMBER" cpu-test-"$BUILDKITE_BUILD_NUMBER"-avx2 || true; 
 }
 trap remove_docker_container EXIT
 remove_docker_container
 
 # Try building the docker image
-numactl -C "$CORE_RANGE" -N "$NUMA_NODE" docker build --tag cpu-test-"$NUMA_NODE" --target vllm-test -f docker/Dockerfile.cpu .
-numactl -C "$CORE_RANGE" -N "$NUMA_NODE" docker build --build-arg VLLM_CPU_DISABLE_AVX512="true" --tag cpu-test-"$NUMA_NODE"-avx2 --target vllm-test -f docker/Dockerfile.cpu .
+numactl -C "$CORE_RANGE" -N "$NUMA_NODE" docker build --tag cpu-test-"$BUILDKITE_BUILD_NUMBER" --target vllm-test -f docker/Dockerfile.cpu .
+numactl -C "$CORE_RANGE" -N "$NUMA_NODE" docker build --build-arg VLLM_CPU_DISABLE_AVX512="true" --tag cpu-test-"$BUILDKITE_BUILD_NUMBER"-avx2 --target vllm-test -f docker/Dockerfile.cpu .
 
 # Run the image, setting --shm-size=4g for tensor parallel.
-docker run -itd --cpuset-cpus="$CORE_RANGE" --cpuset-mems="$NUMA_NODE" --entrypoint /bin/bash -v ~/.cache/huggingface:/root/.cache/huggingface --privileged=true -e HF_TOKEN --env VLLM_CPU_KVCACHE_SPACE=4 --env VLLM_CPU_OMP_THREADS_BIND="$OMP_CORE_RANGE" --env VLLM_CPU_CI_ENV=1 --shm-size=4g --name cpu-test-"$NUMA_NODE" cpu-test-"$NUMA_NODE"
-docker run -itd --cpuset-cpus="$CORE_RANGE" --cpuset-mems="$NUMA_NODE" --entrypoint /bin/bash -v ~/.cache/huggingface:/root/.cache/huggingface --privileged=true -e HF_TOKEN --env VLLM_CPU_KVCACHE_SPACE=4 --env VLLM_CPU_OMP_THREADS_BIND="$OMP_CORE_RANGE" --env VLLM_CPU_CI_ENV=1 --shm-size=4g --name cpu-test-"$NUMA_NODE"-avx2 cpu-test-"$NUMA_NODE"-avx2
+docker run -itd --entrypoint /bin/bash -v ~/.cache/huggingface:/root/.cache/huggingface --cpuset-cpus="$CORE_RANGE"  \
+ --cpuset-mems="$NUMA_NODE" --privileged=true -e HF_TOKEN --env VLLM_CPU_KVCACHE_SPACE=4 --shm-size=4g --name cpu-test-"$BUILDKITE_BUILD_NUMBER"-"$NUMA_NODE" cpu-test-"$BUILDKITE_BUILD_NUMBER"
+docker run -itd --entrypoint /bin/bash -v ~/.cache/huggingface:/root/.cache/huggingface --cpuset-cpus="$CORE_RANGE" \
+ --cpuset-mems="$NUMA_NODE" --privileged=true -e HF_TOKEN --env VLLM_CPU_KVCACHE_SPACE=4 --shm-size=4g --name cpu-test-"$BUILDKITE_BUILD_NUMBER"-avx2-"$NUMA_NODE" cpu-test-"$BUILDKITE_BUILD_NUMBER"-avx2
 
 function cpu_tests() {
   set -e
   export NUMA_NODE=$2
-
-  # list packages
-  docker exec cpu-test-"$NUMA_NODE"-avx2 bash -c "
-    set -e
-    pip list"
-
-  docker exec cpu-test-"$NUMA_NODE" bash -c "
-    set -e
-    pip list"
+  export BUILDKITE_BUILD_NUMBER=$3
 
   # offline inference
-  docker exec cpu-test-"$NUMA_NODE"-avx2 bash -c "
+  docker exec cpu-test-"$BUILDKITE_BUILD_NUMBER"-avx2-"$NUMA_NODE" bash -c "
     set -e
     python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m"
 
   # Run basic model test
-  docker exec cpu-test-"$NUMA_NODE" bash -c "
+  docker exec cpu-test-"$BUILDKITE_BUILD_NUMBER"-"$NUMA_NODE" bash -c "
     set -e
-    pytest -v -s tests/kernels/attention/test_cache.py -m cpu_model
-    pytest -v -s tests/kernels/attention/test_mla_decode_cpu.py -m cpu_model
-    pytest -v -s tests/models/language/generation -m cpu_model
-    pytest -v -s tests/models/language/pooling -m cpu_model
-    pytest -v -s tests/models/multimodal/generation \
-                --ignore=tests/models/multimodal/generation/test_mllama.py \
-                --ignore=tests/models/multimodal/generation/test_pixtral.py \
-                -m cpu_model"
+    pytest -v -s tests/kernels/test_cache.py -m cpu_model
+    pytest -v -s tests/kernels/test_mla_decode_cpu.py -m cpu_model
+    pytest -v -s tests/models/decoder_only/language -m cpu_model
+    pytest -v -s tests/models/embedding/language -m cpu_model
+    pytest -v -s tests/models/encoder_decoder/language -m cpu_model
+    pytest -v -s tests/models/decoder_only/audio_language -m cpu_model
+    pytest -v -s tests/models/decoder_only/vision_language -m cpu_model"
 
   # Run compressed-tensor test
-  docker exec cpu-test-"$NUMA_NODE" bash -c "
+  docker exec cpu-test-"$BUILDKITE_BUILD_NUMBER"-"$NUMA_NODE" bash -c "
     set -e
     pytest -s -v \
     tests/quantization/test_compressed_tensors.py::test_compressed_tensors_w8a8_static_setup \
     tests/quantization/test_compressed_tensors.py::test_compressed_tensors_w8a8_dynamic_per_token"
 
   # Run AWQ test
-  docker exec cpu-test-"$NUMA_NODE" bash -c "
+  docker exec cpu-test-"$BUILDKITE_BUILD_NUMBER"-"$NUMA_NODE" bash -c "
     set -e
-    VLLM_USE_V1=0 pytest -s -v \
+    pytest -s -v \
     tests/quantization/test_ipex_quant.py"
 
   # Run chunked-prefill and prefix-cache test
-  docker exec cpu-test-"$NUMA_NODE" bash -c "
+  docker exec cpu-test-"$BUILDKITE_BUILD_NUMBER"-"$NUMA_NODE" bash -c "
     set -e
     pytest -s -v -k cpu_model \
     tests/basic_correctness/test_chunked_prefill.py"  
 
   # online serving
-  docker exec cpu-test-"$NUMA_NODE" bash -c "
+  docker exec cpu-test-"$BUILDKITE_BUILD_NUMBER"-"$NUMA_NODE" bash -c "
     set -e
+    export VLLM_CPU_KVCACHE_SPACE=10 
+    export VLLM_CPU_OMP_THREADS_BIND=$1
     python3 -m vllm.entrypoints.openai.api_server --model facebook/opt-125m --dtype half & 
     timeout 600 bash -c 'until curl localhost:8000/v1/models; do sleep 1; done' || exit 1
-    VLLM_CPU_CI_ENV=0 python3 benchmarks/benchmark_serving.py \
+    python3 benchmarks/benchmark_serving.py \
       --backend vllm \
       --dataset-name random \
       --model facebook/opt-125m \
@@ -90,7 +83,7 @@ function cpu_tests() {
       --tokenizer facebook/opt-125m"
 
   # Run multi-lora tests
-  docker exec cpu-test-"$NUMA_NODE" bash -c "
+  docker exec cpu-test-"$BUILDKITE_BUILD_NUMBER"-"$NUMA_NODE" bash -c "
     set -e
     pytest -s -v \
     tests/lora/test_qwen2vl.py"
@@ -98,4 +91,4 @@ function cpu_tests() {
 
 # All of CPU tests are expected to be finished less than 40 mins.
 export -f cpu_tests
-timeout 1h bash -c "cpu_tests $CORE_RANGE $NUMA_NODE"
+timeout 40m bash -c "cpu_tests $CORE_RANGE $NUMA_NODE $BUILDKITE_BUILD_NUMBER"

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

@@ -54,11 +54,10 @@ docker run --rm -it --device=/dev/neuron0 --network bridge \
        --name "${container_name}" \
        ${image_name} \
        /bin/bash -c "
-            set -e; # Exit on first error
             python3 /workspace/vllm/examples/offline_inference/neuron.py;
             python3 -m pytest /workspace/vllm/tests/neuron/1_core/ -v --capture=tee-sys;
             for f in /workspace/vllm/tests/neuron/2_core/*.py; do
-                echo \"Running test file: \$f\";
+                echo 'Running test file: '$f;
                 python3 -m pytest \$f -v --capture=tee-sys;
             done
        "

.buildkite/scripts/hardware_ci/run-tpu-v1-test.sh

@@ -2,184 +2,102 @@
 
 set -xu
 
-
-remove_docker_container() { 
-    docker rm -f tpu-test || true; 
-    docker rm -f vllm-tpu || true;
-}
-
-trap remove_docker_container EXIT
-
-# Remove the container that might not be cleaned up in the previous run.
-remove_docker_container
-
 # Build the docker image.
 docker build -f docker/Dockerfile.tpu -t vllm-tpu .
 
 # Set up cleanup.
-cleanup_docker() {
-  # Get Docker's root directory
-  docker_root=$(docker info -f '{{.DockerRootDir}}')
-  if [ -z "$docker_root" ]; then
-    echo "Failed to determine Docker root directory."
-    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
-    echo "Disk usage is below $threshold%. No cleanup needed."
-  fi
-}
-cleanup_docker
+remove_docker_container() { docker rm -f tpu-test || true; }
+trap remove_docker_container EXIT
+# Remove the container that might not be cleaned up in the previous run.
+remove_docker_container
 
 # For HF_TOKEN.
 source /etc/environment
-
+# Run a simple end-to-end example.
 docker run --privileged --net host --shm-size=16G -it \
     -e "HF_TOKEN=$HF_TOKEN" --name tpu-test \
-    vllm-tpu /bin/bash -c '
-set -e # Exit immediately if a command exits with a non-zero status.
-set -u # Treat unset variables as an error.
+    vllm-tpu /bin/bash -c "python3 -m pip install git+https://github.com/thuml/depyf.git \
+    && python3 -m pip install pytest pytest-asyncio tpu-info \
+    && python3 -m pip install lm_eval[api]==0.4.4 \
+    && export VLLM_XLA_CACHE_PATH= \
+    && export VLLM_USE_V1=1 \
+    && export VLLM_XLA_CHECK_RECOMPILATION=1 \
+    && echo HARDWARE \
+    && tpu-info \
+    && { \
+        echo TEST_0: Running test_perf.py; \
+        python3 -m pytest -s -v /workspace/vllm/tests/tpu/test_perf.py; \
+        echo TEST_0_EXIT_CODE: \$?; \
+    } & \
+    { \
+        echo TEST_1: Running test_compilation.py; \
+        python3 -m pytest -s -v /workspace/vllm/tests/tpu/test_compilation.py; \
+        echo TEST_1_EXIT_CODE: \$?; \
+    } & \
+    { \
+        echo TEST_2: Running test_basic.py; \
+        python3 -m pytest -s -v /workspace/vllm/tests/v1/tpu/test_basic.py; \
+        echo TEST_2_EXIT_CODE: \$?; \
+    } & \
+    { \
+        echo TEST_3: Running test_accuracy.py::test_lm_eval_accuracy_v1_engine; \
+        python3 -m pytest -s -v /workspace/vllm/tests/entrypoints/llm/test_accuracy.py::test_lm_eval_accuracy_v1_engine; \
+        echo TEST_3_EXIT_CODE: \$?; \
+    } & \
+    { \
+        echo TEST_4: Running test_quantization_accuracy.py; \
+        python3 -m pytest -s -v /workspace/vllm/tests/tpu/test_quantization_accuracy.py; \
+        echo TEST_4_EXIT_CODE: \$?; \
+    } & \
+    { \
+        echo TEST_5: Running examples/offline_inference/tpu.py; \
+        python3 /workspace/vllm/examples/offline_inference/tpu.py; \
+        echo TEST_5_EXIT_CODE: \$?; \
+    } & \
+    { \
+        echo TEST_6: Running test_tpu_model_runner.py; \
+        python3 -m pytest -s -v /workspace/vllm/tests/tpu/worker/test_tpu_model_runner.py; \
+        echo TEST_6_EXIT_CODE: \$?; \
+    } & \
+    { \
+        echo TEST_7: Running test_sampler.py; \
+        python3 -m pytest -s -v /workspace/vllm/tests/v1/tpu/test_sampler.py; \
+        echo TEST_7_EXIT_CODE: \$?; \
+    } & \
+    { \
+        echo TEST_8: Running test_topk_topp_sampler.py; \
+        python3 -m pytest -s -v /workspace/vllm/tests/v1/tpu/test_topk_topp_sampler.py; \
+        echo TEST_8_EXIT_CODE: \$?; \
+    } & \
+    { \
+        echo TEST_9: Running test_multimodal.py; \
+        python3 -m pytest -s -v /workspace/vllm/tests/v1/tpu/test_multimodal.py; \
+        echo TEST_9_EXIT_CODE: \$?; \
+    } & \
+    { \
+        echo TEST_10: Running test_pallas.py; \
+        python3 -m pytest -s -v /workspace/vllm/tests/v1/tpu/test_pallas.py; \
+        echo TEST_10_EXIT_CODE: \$?; \
+    } & \
+    { \
+        echo TEST_11: Running test_struct_output_generate.py; \
+        python3 -m pytest -s -v /workspace/vllm/tests/v1/entrypoints/llm/test_struct_output_generate.py; \
+        echo TEST_11_EXIT_CODE: \$?; \
+    } & \
+    { \
+        echo TEST_12: Running test_moe_pallas.py; \
+        python3 -m pytest -s -v /workspace/vllm/tests/tpu/test_moe_pallas.py; \
+        echo TEST_12_EXIT_CODE: \$?; \
+    } & \
+    # Disable the TPU LoRA tests until the feature is activated
+    # & { \
+    #     echo TEST_13: Running test_moe_pallas.py; \
+    #     python3 -m pytest -s -v /workspace/vllm/tests/tpu/lora/; \
+    #     echo TEST_13_EXIT_CODE: \$?; \
+    # } & \
+    wait \
+    && echo 'All tests have attempted to run. Check logs for individual test statuses and exit codes.' \
+"
 
-echo "--- Starting script inside Docker container ---"
-
-# Create results directory
-RESULTS_DIR=$(mktemp -d)
-# If mktemp fails, set -e will cause the script to exit.
-echo "Results will be stored in: $RESULTS_DIR"
-
-# Install dependencies
-echo "--- Installing Python dependencies ---"
-python3 -m pip install --progress-bar off git+https://github.com/thuml/depyf.git \
-    && python3 -m pip install --progress-bar off pytest pytest-asyncio tpu-info \
-    && python3 -m pip install --progress-bar off lm_eval[api]==0.4.4
-echo "--- Python dependencies installed ---"
-export VLLM_USE_V1=1
-export VLLM_XLA_CHECK_RECOMPILATION=1
-export VLLM_XLA_CACHE_PATH=
-echo "Using VLLM V1"
-
-echo "--- Hardware Information ---"
-tpu-info
-echo "--- Starting Tests ---"
-set +e
-overall_script_exit_code=0
-
-# --- Test Definitions ---
-# If a test fails, this function will print logs and will not cause the main script to exit.
-run_test() {
-    local test_num=$1
-    local test_name=$2
-    local test_command=$3
-    local log_file="$RESULTS_DIR/test_${test_num}.log"
-    local actual_exit_code
-
-    echo "--- TEST_$test_num: Running $test_name ---"
-    
-    # Execute the test command.
-    eval "$test_command" > >(tee -a "$log_file") 2> >(tee -a "$log_file" >&2)
-    actual_exit_code=$?
-
-    echo "TEST_${test_num}_COMMAND_EXIT_CODE: $actual_exit_code" # This goes to main log
-    echo "TEST_${test_num}_COMMAND_EXIT_CODE: $actual_exit_code" >> "$log_file" # Also to per-test log
-
-    if [ "$actual_exit_code" -ne 0 ]; then
-        echo "TEST_$test_num ($test_name) FAILED with exit code $actual_exit_code." >&2
-        echo "--- Log for failed TEST_$test_num ($test_name) ---" >&2
-        if [ -f "$log_file" ]; then
-            cat "$log_file" >&2
-        else
-            echo "Log file $log_file not found for TEST_$test_num ($test_name)." >&2
-        fi
-        echo "--- End of log for TEST_$test_num ($test_name) ---" >&2
-        return "$actual_exit_code" # Return the failure code
-    else
-        echo "TEST_$test_num ($test_name) PASSED."
-        return 0 # Return success
-    fi
-}
-
-# Helper function to call run_test and update the overall script exit code
-run_and_track_test() {
-    local test_num_arg="$1"
-    local test_name_arg="$2"
-    local test_command_arg="$3"
-
-    # Run the test
-    run_test "$test_num_arg" "$test_name_arg" "$test_command_arg"
-    local test_specific_exit_code=$?
-
-    # If the test failed, set the overall script exit code to 1
-    if [ "$test_specific_exit_code" -ne 0 ]; then
-        # No need for extra echo here, run_test already logged the failure.
-        overall_script_exit_code=1
-    fi
-}
-
-# --- Actual Test Execution ---
-run_and_track_test 0 "test_perf.py" \
-    "python3 -m pytest -s -v /workspace/vllm/tests/v1/tpu/test_perf.py"
-run_and_track_test 1 "test_compilation.py" \
-    "python3 -m pytest -s -v /workspace/vllm/tests/tpu/test_compilation.py"
-run_and_track_test 2 "test_basic.py" \
-    "python3 -m pytest -s -v /workspace/vllm/tests/v1/tpu/test_basic.py"
-run_and_track_test 3 "test_accuracy.py::test_lm_eval_accuracy_v1_engine" \
-    "python3 -m pytest -s -v /workspace/vllm/tests/entrypoints/llm/test_accuracy.py::test_lm_eval_accuracy_v1_engine"
-run_and_track_test 4 "test_quantization_accuracy.py" \
-    "python3 -m pytest -s -v /workspace/vllm/tests/tpu/test_quantization_accuracy.py"
-run_and_track_test 5 "examples/offline_inference/tpu.py" \
-    "python3 /workspace/vllm/examples/offline_inference/tpu.py"
-run_and_track_test 6 "test_tpu_model_runner.py" \
-    "python3 -m pytest -s -v /workspace/vllm/tests/v1/tpu/worker/test_tpu_model_runner.py"
-run_and_track_test 7 "test_sampler.py" \
-    "python3 -m pytest -s -v /workspace/vllm/tests/v1/tpu/test_sampler.py"
-run_and_track_test 8 "test_topk_topp_sampler.py" \
-    "python3 -m pytest -s -v /workspace/vllm/tests/v1/tpu/test_topk_topp_sampler.py"
-run_and_track_test 9 "test_multimodal.py" \
-    "python3 -m pytest -s -v /workspace/vllm/tests/v1/tpu/test_multimodal.py"
-run_and_track_test 10 "test_pallas.py" \
-    "python3 -m pytest -s -v /workspace/vllm/tests/v1/tpu/test_pallas.py"
-run_and_track_test 11 "test_struct_output_generate.py" \
-    "python3 -m pytest -s -v /workspace/vllm/tests/v1/entrypoints/llm/test_struct_output_generate.py -k \"not test_structured_output_with_reasoning_matrices\""
-run_and_track_test 12 "test_moe_pallas.py" \
-    "python3 -m pytest -s -v /workspace/vllm/tests/tpu/test_moe_pallas.py"
-run_and_track_test 13 "test_lora.py" \
-    "VLLM_XLA_CHECK_RECOMPILATION=0 python3 -m pytest -s -v /workspace/vllm/tests/tpu/lora/test_lora.py"
-run_and_track_test 14 "test_tpu_qkv_linear.py" \
-    "python3 -m pytest -s -v /workspace/vllm/tests/v1/tpu/test_tpu_qkv_linear.py"
-run_and_track_test 15 "test_spmd_model_weight_loading.py" \
-    "python3 -m pytest -s -v /workspace/vllm/tests/v1/tpu/test_spmd_model_weight_loading.py"
-
-# After all tests have been attempted, exit with the overall status.
-if [ "$overall_script_exit_code" -ne 0 ]; then
-    echo "--- One or more tests FAILED. Overall script exiting with failure code 1. ---"
-else
-    echo "--- All tests have completed and PASSED. Overall script exiting with success code 0. ---"
-fi
-exit "$overall_script_exit_code"
-' # IMPORTANT: This is the closing single quote for the bash -c "..." command. Ensure it is present and correct.
-
-# Capture the exit code of the docker run command
-DOCKER_RUN_EXIT_CODE=$?
-
-# The trap will run for cleanup.
-# Exit the main script with the Docker run command's exit code.
-if [ "$DOCKER_RUN_EXIT_CODE" -ne 0 ]; then
-    echo "Docker run command failed with exit code $DOCKER_RUN_EXIT_CODE."
-    exit "$DOCKER_RUN_EXIT_CODE"
-else
-    echo "Docker run command completed successfully."
-    exit 0
-fi
 # TODO: This test fails because it uses RANDOM_SEED sampling
-# pytest -v -s /workspace/vllm/tests/tpu/test_custom_dispatcher.py \
+# && VLLM_USE_V1=1 pytest -v -s /workspace/vllm/tests/tpu/test_custom_dispatcher.py \

.buildkite/scripts/rerun-test.sh

@@ -1,18 +0,0 @@
-#!/bin/bash
-
-# Usage: ./rerun_test.sh path/to/test.py::test_name
-
-# Check if argument is given
-if [ $# -lt 1 ]; then
-    echo "Usage: $0 path/to/test.py::test_name"
-    echo "Example: $0 tests/v1/engine/test_engine_core_client.py::test_kv_cache_events[True-tcp]"
-    exit 1
-fi
-
-TEST=$1
-COUNT=1
-
-while pytest -sv "$TEST"; do
-    COUNT=$((COUNT + 1))
-    echo "RUN NUMBER ${COUNT}"
-done

.buildkite/scripts/tpu/cleanup_docker.sh

@@ -1,24 +0,0 @@
-#!/bin/bash
-
-set -euo pipefail
-
-docker_root=$(docker info -f '{{.DockerRootDir}}')
-if [ -z "$docker_root" ]; then
-  echo "Failed to determine Docker root directory."
-  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
-  echo "Disk usage is below $threshold%. No cleanup needed."
-fi

.buildkite/scripts/tpu/config_v6e_1.env

@@ -1,14 +0,0 @@
-# Environment config
-TEST_NAME=llama8b
-CONTAINER_NAME=vllm-tpu
-
-# vllm config
-MODEL=meta-llama/Llama-3.1-8B-Instruct
-MAX_NUM_SEQS=256
-MAX_NUM_BATCHED_TOKENS=1024
-TENSOR_PARALLEL_SIZE=1
-MAX_MODEL_LEN=2048
-DOWNLOAD_DIR=/mnt/disks/persist
-EXPECTED_THROUGHPUT=8.0
-INPUT_LEN=1800
-OUTPUT_LEN=128

.buildkite/scripts/tpu/docker_run_bm.sh

@@ -1,102 +0,0 @@
-#!/bin/bash
-
-if [ ! -f "$1" ]; then
-  echo "Error: The env file '$1' does not exist."
-  exit 1  # Exit the script with a non-zero status to indicate an error
-fi
-
-ENV_FILE=$1
-
-# For testing on local vm, use `set -a` to export all variables
-source /etc/environment
-source $ENV_FILE
-
-remove_docker_container() { 
-    docker rm -f tpu-test || true; 
-    docker rm -f vllm-tpu || true;
-    docker rm -f $CONTAINER_NAME || true;
-}
-
-trap remove_docker_container EXIT
-
-# Remove the container that might not be cleaned up in the previous run.
-remove_docker_container
-
-# Build docker image.
-# TODO: build the image outside the script and share the image with other
-# tpu test if building time is too long.
-DOCKER_BUILDKIT=1 docker build \
-  --build-arg max_jobs=16 \
-  --build-arg USE_SCCACHE=1 \
-  --build-arg GIT_REPO_CHECK=0 \
-  --tag vllm/vllm-tpu-bm \
-  --progress plain -f docker/Dockerfile.tpu .
-
-LOG_ROOT=$(mktemp -d)
-# If mktemp fails, set -e will cause the script to exit.
-echo "Results will be stored in: $LOG_ROOT"
-
-if [ -z "$HF_TOKEN" ]; then
-  echo "Error: HF_TOKEN is not set or is empty."  
-  exit 1
-fi
-
-# Make sure mounted disk or dir exists
-if [ ! -d "$DOWNLOAD_DIR" ]; then
-    echo "Error: Folder $DOWNLOAD_DIR does not exist. This is useually a mounted drive. If no mounted drive, just create a folder."
-    exit 1
-fi
-
-echo "Run model $MODEL"
-echo
-
-echo "starting docker...$CONTAINER_NAME"
-echo    
-docker run \
- -v $DOWNLOAD_DIR:$DOWNLOAD_DIR \
- --env-file $ENV_FILE \
- -e HF_TOKEN="$HF_TOKEN" \
- -e TARGET_COMMIT=$BUILDKITE_COMMIT \
- -e MODEL=$MODEL \
- -e WORKSPACE=/workspace \
- --name $CONTAINER_NAME \
- -d \
- --privileged \
- --network host \
- -v /dev/shm:/dev/shm \
- vllm/vllm-tpu-bm tail -f /dev/null
-
-echo "run script..."
-echo
-docker exec "$CONTAINER_NAME" /bin/bash -c ".buildkite/scripts/hardware_ci/run_bm.sh"
-
-echo "copy result back..."
-VLLM_LOG="$LOG_ROOT/$TEST_NAME"_vllm_log.txt
-BM_LOG="$LOG_ROOT/$TEST_NAME"_bm_log.txt
-docker cp "$CONTAINER_NAME:/workspace/vllm_log.txt" "$VLLM_LOG" 
-docker cp "$CONTAINER_NAME:/workspace/bm_log.txt" "$BM_LOG"
-
-throughput=$(grep "Request throughput (req/s):" "$BM_LOG" | sed 's/[^0-9.]//g')
-echo "throughput for $TEST_NAME at $BUILDKITE_COMMIT: $throughput"
-
-if [ "$BUILDKITE" = "true" ]; then
-  echo "Running inside Buildkite"
-  buildkite-agent artifact upload "$VLLM_LOG" 
-  buildkite-agent artifact upload "$BM_LOG"
-else
-  echo "Not running inside Buildkite"
-fi
-
-#
-# compare the throughput with EXPECTED_THROUGHPUT 
-# and assert meeting the expectation
-# 
-if [[ -z "$throughput" || ! "$throughput" =~ ^[0-9]+([.][0-9]+)?$ ]]; then
-  echo "Failed to get the throughput"
-  exit 1
-fi
-
-if (( $(echo "$throughput < $EXPECTED_THROUGHPUT" | bc -l) )); then
-  echo "Error: throughput($throughput) is less than expected($EXPECTED_THROUGHPUT)"
-  exit 1
-fi

.buildkite/scripts/tpu/run_bm.sh

@@ -1,94 +0,0 @@
-#!/bin/bash
-
-set -euo pipefail
-
-VLLM_LOG="$WORKSPACE/vllm_log.txt"
-BM_LOG="$WORKSPACE/bm_log.txt"
-
-if [ -n "$TARGET_COMMIT" ]; then
-  head_hash=$(git rev-parse HEAD)
-  if [ "$TARGET_COMMIT" != "$head_hash" ]; then
-    echo "Error: target commit $TARGET_COMMIT does not match HEAD: $head_hash"
-    exit 1
-  fi
-fi
-
-echo "model: $MODEL"
-echo
-
-#
-# create a log folder
-#
-mkdir "$WORKSPACE/log"
-
-# TODO: Move to image building.
-pip install pandas
-pip install datasets
-
-#
-# create sonnet_4x
-#
-echo "Create sonnet_4x.txt"
-echo "" > benchmarks/sonnet_4x.txt
-for _ in {1..4}
- do
-  cat benchmarks/sonnet.txt >> benchmarks/sonnet_4x.txt
-done
-
-#
-# start vllm service in backend
-#
-echo "lanching vllm..."
-echo "logging to $VLLM_LOG"
-echo
-
-VLLM_USE_V1=1 vllm serve $MODEL \
- --seed 42 \
- --disable-log-requests \
- --max-num-seqs $MAX_NUM_SEQS \
- --max-num-batched-tokens $MAX_NUM_BATCHED_TOKENS \
- --tensor-parallel-size $TENSOR_PARALLEL_SIZE \
- --no-enable-prefix-caching \
- --download_dir $DOWNLOAD_DIR \
- --max-model-len $MAX_MODEL_LEN > "$VLLM_LOG" 2>&1 &
-
-
-echo "wait for 20 minutes.."
-echo
-# sleep 1200
-# wait for 10 minutes...
-for i in {1..120}; do
-    # TODO: detect other type of errors.
-    if grep -Fq "raise RuntimeError" "$VLLM_LOG"; then
-        echo "Detected RuntimeError, exiting."
-        exit 1
-    elif grep -Fq "Application startup complete" "$VLLM_LOG"; then
-        echo "Application started"
-        break
-    else
-        echo "wait for 10 seconds..."
-        sleep 10
-    fi
-done
-
-#
-# run test
-#
-echo "run benchmark test..."
-echo "logging to $BM_LOG"
-echo
-python benchmarks/benchmark_serving.py \
-    --backend vllm \
-    --model $MODEL  \
-    --dataset-name sonnet \
-    --dataset-path benchmarks/sonnet_4x.txt \
-    --sonnet-input-len $INPUT_LEN \
-    --sonnet-output-len $OUTPUT_LEN \
-    --ignore-eos > "$BM_LOG"
-
-echo "completed..."
-echo
-
-throughput=$(grep "Request throughput (req/s):" "$BM_LOG" | sed 's/[^0-9.]//g')
-echo "throughput: $throughput"
-echo

.buildkite/test-pipeline.yaml

@@ -89,7 +89,7 @@ steps:
   - VLLM_TEST_ENABLE_ARTIFICIAL_PREEMPT=1 pytest -v -s basic_correctness/test_preemption.py
 
 - label: Chunked Prefill Test
-  mirror_hardwares: [amdexperimental, amdproduction]
+  mirror_hardwares: [amdexperimental]
   source_file_dependencies:
   - vllm/
   - tests/basic_correctness/test_chunked_prefill
@@ -145,7 +145,6 @@ steps:
   - examples/offline_inference/rlhf_colocate.py
   - tests/examples/offline_inference/data_parallel.py
   - tests/v1/test_async_llm_dp.py
-  - tests/v1/engine/test_engine_core_client.py
   commands:
   # test with tp=2 and external_dp=2
   - VLLM_USE_V1=0 torchrun --nproc-per-node=4 distributed/test_torchrun_example.py
@@ -155,7 +154,6 @@ steps:
   # test with internal dp
   - python3 ../examples/offline_inference/data_parallel.py
   - TP_SIZE=2 DP_SIZE=2 pytest -v -s v1/test_async_llm_dp.py
-  - pytest -v -s v1/engine/test_engine_core_client.py::test_kv_cache_events_dp
   - pytest -v -s distributed/test_utils.py
   - pytest -v -s compile/test_basic_correctness.py
   - pytest -v -s distributed/test_pynccl.py
@@ -177,11 +175,6 @@ steps:
   - tests/tracing
   commands:
   - pytest -v -s metrics
-  - "pip install \
-      'opentelemetry-sdk>=1.26.0' \
-      'opentelemetry-api>=1.26.0' \
-      'opentelemetry-exporter-otlp>=1.26.0' \
-      'opentelemetry-semantic-conventions-ai>=0.4.1'"
   - pytest -v -s tracing
 
 ##### fast check tests  #####
@@ -206,9 +199,8 @@ steps:
   - tests/test_sequence
   - tests/test_config
   - tests/test_logger
-  - tests/test_vllm_port
   commands:
-  - pytest -v -s engine test_sequence.py test_config.py test_logger.py test_vllm_port.py
+  - pytest -v -s engine test_sequence.py test_config.py test_logger.py
   # OOM in the CI unless we run this separately
   - pytest -v -s tokenization
 
@@ -271,15 +263,6 @@ steps:
   commands:
     - pytest -v -s prefix_caching
 
-
-- label: Platform Tests (CUDA)
-  mirror_hardwares: [amdexperimental]
-  source_file_dependencies:
-  - vllm/
-  - tests/cuda
-  commands:
-    - pytest -v -s cuda/test_cuda_context.py
-
 - label: Samplers Test # 36min
   mirror_hardwares: [amdexperimental]
   source_file_dependencies:
@@ -291,6 +274,17 @@ steps:
     - pytest -v -s samplers
     - VLLM_USE_FLASHINFER_SAMPLER=1 pytest -v -s samplers
 
+- label: LogitsProcessor Test # 5min
+  mirror_hardwares: [amdexperimental, amdproduction]
+  source_file_dependencies:
+  - vllm/model_executor/layers
+  - vllm/model_executor/guided_decoding
+  - tests/test_logits_processor
+  - tests/model_executor/test_guided_processors
+  commands:
+    - pytest -v -s test_logits_processor.py
+    - pytest -v -s model_executor/test_guided_processors.py
+
 - label: Speculative decoding tests # 40min
   mirror_hardwares: [amdexperimental]
   source_file_dependencies:
@@ -303,7 +297,7 @@ steps:
     - pytest -v -s spec_decode/e2e/test_eagle_correctness.py
 
 - label: LoRA Test %N # 15min each
-  mirror_hardwares: [amdexperimental, amdproduction]
+  mirror_hardwares: [amdexperimental]
   source_file_dependencies:
   - vllm/lora
   - tests/lora
@@ -319,7 +313,6 @@ steps:
   commands:
     - pytest -v -s compile/test_pass_manager.py
     - pytest -v -s compile/test_fusion.py
-    - pytest -v -s compile/test_fusion_attn.py
     - pytest -v -s compile/test_silu_mul_quant_fusion.py
     - pytest -v -s compile/test_sequence_parallelism.py
     - pytest -v -s compile/test_async_tp.py
@@ -335,7 +328,6 @@ steps:
   # these tests need to be separated, cannot combine
   - pytest -v -s compile/piecewise/test_simple.py
   - pytest -v -s compile/piecewise/test_toy_llama.py
-  - pytest -v -s compile/piecewise/test_full_cudagraph.py
 
 - label: PyTorch Fullgraph Test # 18min
   mirror_hardwares: [amdexperimental, amdproduction]
@@ -405,17 +397,6 @@ steps:
     - pytest -v -s tensorizer_loader
     - pytest -v -s entrypoints/openai/test_tensorizer_entrypoint.py
 
-- label: Model Executor Test
-  mirror_hardwares: [amdexperimental, amdproduction]
-  soft_fail: true
-  source_file_dependencies:
-  - vllm/model_executor
-  - tests/model_executor
-  commands:
-    - apt-get update && apt-get install -y curl libsodium23
-    - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-    - pytest -v -s model_executor
-
 - label: Benchmarks # 9min
   mirror_hardwares: [amdexperimental, amdproduction]
   working_dir: "/vllm-workspace/.buildkite"
@@ -439,9 +420,6 @@ steps:
   - vllm/model_executor/layers/quantization
   - tests/quantization
   commands:
-  # temporary install here since we need nightly, will move to requirements/test.in
-  # after torchao 0.12 release
-  - pip install --pre torchao --index-url https://download.pytorch.org/whl/nightly/cu126
   - VLLM_TEST_FORCE_LOAD_FORMAT=auto pytest -v -s quantization
 
 - label: LM Eval Small Models # 53min
@@ -639,11 +617,9 @@ steps:
   - vllm/worker/model_runner.py
   - entrypoints/llm/test_collective_rpc.py
   - tests/v1/test_async_llm_dp.py
-  - tests/v1/entrypoints/openai/test_multi_api_servers.py
   - vllm/v1/engine/
   commands:
   - TP_SIZE=1 DP_SIZE=2 pytest -v -s v1/test_async_llm_dp.py
-  - DP_SIZE=2 pytest -v -s v1/entrypoints/openai/test_multi_api_servers.py
   - pytest -v -s entrypoints/llm/test_collective_rpc.py
   - pytest -v -s ./compile/test_basic_correctness.py
   - pytest -v -s ./compile/test_wrapper.py
@@ -684,7 +660,7 @@ steps:
   - pytest -v -s plugins/lora_resolvers # unit tests for in-tree lora resolver plugins
 
 - label: Multi-step Tests (4 GPUs) # 36min
-  mirror_hardwares: [amdexperimental, amdproduction]
+  mirror_hardwares: [amdexperimental]
   working_dir: "/vllm-workspace/tests"
   num_gpus: 4
   source_file_dependencies:

.github/CODEOWNERS

@@ -10,17 +10,15 @@
 /vllm/worker/worker.py @zhuohan123 @youkaichao @alexm-redhat @comaniac @njhill
 /vllm/model_executor/layers/sampler.py @zhuohan123 @youkaichao @alexm-redhat @comaniac @njhill
 /vllm/model_executor/layers/quantization @mgoin @robertgshaw2-redhat @tlrmchlsmth
-/vllm/model_executor/guided_decoding @mgoin @russellb @aarnphm
+/vllm/model_executor/guided_decoding @mgoin @russellb
 /vllm/multimodal @DarkLight1337 @ywang96
 /vllm/vllm_flash_attn @LucasWilkinson
 /vllm/lora @jeejeelee
-/vllm/reasoning @aarnphm
-/vllm/entrypoints @aarnphm
 CMakeLists.txt @tlrmchlsmth
 
 # vLLM V1
 /vllm/v1 @WoosukKwon @robertgshaw2-redhat @njhill @ywang96 @comaniac @alexm-redhat
-/vllm/v1/structured_output @mgoin @russellb @aarnphm
+/vllm/v1/structured_output @mgoin @russellb
 
 # Test ownership
 /.buildkite/lm-eval-harness @mgoin @simon-mo
@@ -29,8 +27,8 @@ CMakeLists.txt @tlrmchlsmth
 /tests/distributed/test_multi_node_assignment.py @youkaichao
 /tests/distributed/test_pipeline_parallel.py @youkaichao
 /tests/distributed/test_same_node.py @youkaichao
-/tests/entrypoints @DarkLight1337 @robertgshaw2-redhat @simon-mo @aarnphm
-/tests/entrypoints/llm/test_guided_generate.py @mgoin @russellb @aarnphm
+/tests/entrypoints @DarkLight1337 @robertgshaw2-redhat @simon-mo
+/tests/entrypoints/llm/test_guided_generate.py @mgoin @russellb
 /tests/kernels @tlrmchlsmth @WoosukKwon
 /tests/model_executor/test_guided_processors.py @mgoin @russellb
 /tests/models @DarkLight1337 @ywang96
@@ -40,11 +38,11 @@ CMakeLists.txt @tlrmchlsmth
 /tests/quantization @mgoin @robertgshaw2-redhat
 /tests/spec_decode @njhill @LiuXiaoxuanPKU
 /tests/test_inputs.py @DarkLight1337 @ywang96
-/tests/v1/entrypoints/llm/test_struct_output_generate.py @mgoin @russellb @aarnphm
-/tests/v1/structured_output @mgoin @russellb @aarnphm
+/tests/v1/entrypoints/llm/test_struct_output_generate.py @mgoin @russellb
+/tests/v1/structured_output @mgoin @russellb
 /tests/weight_loading @mgoin @youkaichao
 /tests/lora @jeejeelee
 
 # Docs
 /docs @hmellor
-mkdocs.yaml @hmellor
+mkdocs.yaml @hmellor

.github/ISSUE_TEMPLATE/400-bug-report.yml

@@ -8,16 +8,6 @@ body:
   attributes:
     value: >
       #### Before submitting an issue, please make sure the issue hasn't been already addressed by searching through [the existing and past issues](https://github.com/vllm-project/vllm/issues?q=is%3Aissue+sort%3Acreated-desc+).
-- type: markdown
-  attributes:
-    value: |
-      ⚠️ **SECURITY WARNING:** Please review any text you paste to ensure it does not contain sensitive information such as:
-      - API tokens or keys (e.g., Hugging Face tokens, OpenAI API keys)
-      - Passwords or authentication credentials
-      - Private URLs or endpoints
-      - Personal or confidential data
-      
-      Consider redacting or replacing sensitive values with placeholders like `<YOUR_TOKEN_HERE>` when sharing configuration or code examples.
 - type: textarea
   attributes:
     label: Your current environment

.github/PULL_REQUEST_TEMPLATE.md

@@ -1,18 +1,6 @@
-## Essential Elements of an Effective PR Description Checklist
-- [ ] The purpose of the PR, such as "Fix some issue (link existing issues this PR will resolve)".
-- [ ] The test plan, such as providing test command.
-- [ ] The test results, such as pasting the results comparison before and after, or e2e results
-- [ ] (Optional) The necessary documentation update, such as updating `supported_models.md` and `examples` for a new model.
+FILL IN THE PR DESCRIPTION HERE
 
-PLEASE FILL IN THE PR DESCRIPTION HERE ENSURING ALL CHECKLIST ITEMS ABOVE HAVE BEEN CONSIDERED.
-
-## Purpose
-
-## Test Plan
-
-## Test Result
-
-## (Optional) Documentation Update
+FIX #xxxx (*link existing issues this PR will resolve*)
 
 <!--- pyml disable-next-line no-emphasis-as-heading -->
 **BEFORE SUBMITTING, PLEASE READ <https://docs.vllm.ai/en/latest/contributing>** (anything written below this line will be removed by GitHub Actions)

.github/mergify.yml

@@ -36,21 +36,6 @@ pull_request_rules:
       add:
         - frontend
 
-- name: label-llama
-  description: Automatically apply llama label
-  conditions:
-    - or:
-      - files~=^examples/.*llama.*\.py
-      - files~=^tests/.*llama.*\.py
-      - files~=^vllm/entrypoints/openai/tool_parsers/llama.*\.py
-      - files~=^vllm/model_executor/models/.*llama.*\.py
-      - files~=^vllm/transformers_utils/configs/.*llama.*\.py
-      - title~=(?i)llama
-  actions:
-    label:
-      add:
-        - llama
-
 - name: label-multi-modality
   description: Automatically apply multi-modality label
   conditions:
@@ -66,53 +51,6 @@ pull_request_rules:
       add:
         - multi-modality
 
-- name: label-performance
-  description: Automatically apply performance label
-  conditions:
-    - or:
-      - files~=^benchmarks/
-      - files~=^vllm/benchmarks/
-      - files~=^tests/benchmarks/
-      - files~=^\.buildkite/nightly-benchmarks/
-  actions:
-    label:
-      add:
-        - performance
-
-- name: label-qwen
-  description: Automatically apply qwen label
-  conditions:
-    - or:
-      - files~=^examples/.*qwen.*\.py
-      - files~=^tests/.*qwen.*\.py
-      - files~=^vllm/model_executor/models/.*qwen.*\.py
-      - files~=^vllm/reasoning/.*qwen.*\.py
-      - title~=(?i)Qwen
-  actions:
-    label:
-      add:
-        - qwen
-
-- name: label-rocm
-  description: Automatically apply rocm label
-  conditions:
-    - or:
-      - files~=^csrc/rocm/
-      - files~=^docker/Dockerfile.rocm
-      - files~=^requirements/rocm.*\.txt
-      - files~=^vllm/attention/backends/rocm.*\.py
-      - files~=^vllm/attention/ops/rocm.*\.py
-      - files~=^vllm/model_executor/layers/fused_moe/rocm.*\.py
-      - files~=^vllm/v1/attention/backends/mla/rocm.*\.py
-      - files~=^tests/kernels/.*_rocm.*\.py
-      - files=vllm/platforms/rocm.py
-      - title~=(?i)AMD
-      - title~=(?i)ROCm
-  actions:
-    label:
-      add:
-        - rocm
-
 - name: label-structured-output
   description: Automatically apply structured-output label
   conditions:

.gitignore

@@ -200,5 +200,5 @@ benchmarks/**/*.json
 actionlint
 shellcheck*/
 
-# Ignore moe/marlin_moe gen code
+# Ingore moe/marlin_moe gen code
 csrc/moe/marlin_moe_wna16/kernel_*

.pre-commit-config.yaml

@@ -11,8 +11,6 @@ repos:
   hooks:
   - id: yapf
     args: [--in-place, --verbose]
-    # Keep the same list from yapfignore here to avoid yapf failing without any inputs
-    exclude: '(.buildkite|benchmarks|build|examples)/.*'
 - repo: https://github.com/astral-sh/ruff-pre-commit
   rev: v0.11.7
   hooks:
@@ -20,10 +18,12 @@ repos:
     args: [--output-format, github, --fix]
   - id: ruff-format
     files: ^(.buildkite|benchmarks|examples)/.*
-- repo: https://github.com/crate-ci/typos
-  rev: v1.32.0
+- repo: https://github.com/codespell-project/codespell
+  rev: v2.4.1
   hooks:
-  - id: typos
+  - id: codespell
+    additional_dependencies: ['tomli']
+    args: ['--toml', 'pyproject.toml']
 - repo: https://github.com/PyCQA/isort
   rev: 6.0.1
   hooks:
@@ -58,7 +58,7 @@ repos:
     entry: tools/mypy.sh 0 "local"
     language: python
     types: [python]
-    additional_dependencies: &mypy_deps [mypy==1.11.1, types-cachetools, types-setuptools, types-PyYAML, types-requests, pydantic]
+    additional_dependencies: &mypy_deps [mypy==1.11.1, types-cachetools, types-setuptools, types-PyYAML, types-requests]
     stages: [pre-commit] # Don't run in CI
   - id: mypy-3.9 # TODO: Use https://github.com/pre-commit/mirrors-mypy when mypy setup is less awkward
     name: Run mypy for Python 3.9
@@ -115,11 +115,6 @@ repos:
     entry: python tools/check_spdx_header.py
     language: python
     types: [python]
-  - id: check-root-lazy-imports
-    name: Check root lazy imports
-    entry: python tools/check_init_lazy_imports.py
-    language: python
-    types: [python]
   - id: check-filenames
     name: Check for spaces in all filenames
     entry: bash
@@ -148,13 +143,6 @@ repos:
     types: [python]
     pass_filenames: false
     additional_dependencies: [regex]
-  - id: check-pickle-imports
-    name: Prevent new pickle/cloudpickle imports
-    entry: python tools/check_pickle_imports.py
-    language: python
-    types: [python]
-    pass_filenames: false
-    additional_dependencies: [pathspec, regex]
   # Keep `suggestion` last
   - id: suggestion
     name: Suggestion

CMakeLists.txt

@@ -23,9 +23,6 @@ include(${CMAKE_CURRENT_LIST_DIR}/cmake/utils.cmake)
 # Suppress potential warnings about unused manually-specified variables
 set(ignoreMe "${VLLM_PYTHON_PATH}")
 
-# Prevent installation of dependencies (cutlass) by default.
-install(CODE "set(CMAKE_INSTALL_LOCAL_ONLY TRUE)" ALL_COMPONENTS)
-
 #
 # Supported python versions.  These versions will be searched in order, the
 # first match will be selected.  These should be kept in sync with setup.py.
@@ -182,6 +179,9 @@ include(FetchContent)
 file(MAKE_DIRECTORY ${FETCHCONTENT_BASE_DIR}) # Ensure the directory exists
 message(STATUS "FetchContent base directory: ${FETCHCONTENT_BASE_DIR}")
 
+#
+# Set rocm version dev int.
+#
 if(VLLM_GPU_LANG STREQUAL "HIP")
   #
   # Overriding the default -O set up by cmake, adding ggdb3 for the most verbose devug info
@@ -189,6 +189,7 @@ if(VLLM_GPU_LANG STREQUAL "HIP")
   set(CMAKE_${VLLM_GPU_LANG}_FLAGS_DEBUG "${CMAKE_${VLLM_GPU_LANG}_FLAGS_DEBUG} -O0 -ggdb3")
   set(CMAKE_CXX_FLAGS_DEBUG "${CMAKE_CXX_FLAGS_DEBUG} -O0 -ggdb3")
 
+
   #
   # Certain HIP functions are marked as [[nodiscard]], yet vllm ignores the result which generates
   # a lot of warnings that always mask real issues. Suppressing until this is properly addressed.
@@ -242,7 +243,6 @@ set(VLLM_EXT_SRC
   "csrc/activation_kernels.cu"
   "csrc/layernorm_kernels.cu"
   "csrc/layernorm_quant_kernels.cu"
-  "csrc/sampler.cu"
   "csrc/cuda_view.cu"
   "csrc/quantization/gptq/q_gemm.cu"
   "csrc/quantization/compressed_tensors/int8_quant_kernels.cu"
@@ -308,333 +308,142 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
   # Keep building Marlin for 9.0 as there are some group sizes and shapes that
   # are not supported by Machete yet.
   # 9.0 for latest bf16 atomicAdd PTX
-  cuda_archs_loose_intersection(MARLIN_ARCHS "8.0;8.7;9.0+PTX" "${CUDA_ARCHS}")
-  if (MARLIN_ARCHS)
+  # Marlin kernels: generate and build for supported architectures
+  optional_cuda_sources(
+    NAME Marlin
+    ARCHS "8.0;9.0+PTX"
+    GEN_SCRIPT "${CMAKE_CURRENT_SOURCE_DIR}/csrc/quantization/gptq_marlin/generate_kernels.py"
+    GEN_GLOB "csrc/quantization/gptq_marlin/kernel_*.cu"
+    SRCS
+      "csrc/quantization/marlin/dense/marlin_cuda_kernel.cu"
+      "csrc/quantization/marlin/sparse/marlin_24_cuda_kernel.cu"
+      "csrc/quantization/marlin/qqq/marlin_qqq_gemm_kernel.cu"
+      "csrc/quantization/gptq_marlin/gptq_marlin.cu"
+      "csrc/quantization/gptq_marlin/gptq_marlin_repack.cu"
+      "csrc/quantization/gptq_marlin/awq_marlin_repack.cu"
+  )
 
-    #
-    # For the Marlin kernels we automatically generate sources for various
-    # preselected input type pairs and schedules.
-    # Generate sources:
-    set(MARLIN_GEN_SCRIPT
-      ${CMAKE_CURRENT_SOURCE_DIR}/csrc/quantization/gptq_marlin/generate_kernels.py)
-    file(MD5 ${MARLIN_GEN_SCRIPT} MARLIN_GEN_SCRIPT_HASH)
+  # AllSpark kernels
+  optional_cuda_sources(
+    NAME AllSpark
+    ARCHS "8.0;8.6;8.7;8.9"
+    SRCS
+      "csrc/quantization/gptq_allspark/allspark_repack.cu"
+      "csrc/quantization/gptq_allspark/allspark_qgemm_w8a16.cu"
+  )
 
-    message(STATUS "Marlin generation script hash: ${MARLIN_GEN_SCRIPT_HASH}")
-    message(STATUS "Last run Marlin generate script hash: $CACHE{MARLIN_GEN_SCRIPT_HASH}")
+  # The cutlass_scaled_mm kernels for Hopper (c3x, i.e. CUTLASS 3.x) require CUDA 12.0 or later
+  optional_cuda_sources(
+    NAME scaled_mm_c3x_sm90
+    MIN_VERSION 12.0
+    ARCHS "9.0a"
+    SRCS
+      "csrc/quantization/cutlass_w8a8/scaled_mm_c3x_sm90.cu"
+      "csrc/quantization/cutlass_w8a8/c3x/scaled_mm_sm90_fp8.cu"
+      "csrc/quantization/cutlass_w8a8/c3x/scaled_mm_sm90_int8.cu"
+      "csrc/quantization/cutlass_w8a8/c3x/scaled_mm_azp_sm90_int8.cu"
+      "csrc/quantization/cutlass_w8a8/c3x/scaled_mm_blockwise_sm90_fp8.cu"
+    FLAGS "-DENABLE_SCALED_MM_SM90=1"
+    VERSION_MSG
+      "Not building scaled_mm_c3x_sm90: CUDA Compiler version is not >= 12.0."
+      "Please upgrade to CUDA 12.0 or later to run FP8 quantized models on Hopper."
+  )
 
-    if (NOT DEFINED CACHE{MARLIN_GEN_SCRIPT_HASH}
-        OR NOT $CACHE{MARLIN_GEN_SCRIPT_HASH} STREQUAL ${MARLIN_GEN_SCRIPT_HASH})
-      execute_process(
-        COMMAND ${CMAKE_COMMAND} -E env
-        PYTHONPATH=$PYTHONPATH
-          ${Python_EXECUTABLE} ${MARLIN_GEN_SCRIPT}
-        RESULT_VARIABLE marlin_generation_result
-        OUTPUT_VARIABLE marlin_generation_result
-        OUTPUT_FILE ${CMAKE_CURRENT_BINARY_DIR}/marlin_generation.log
-        ERROR_FILE ${CMAKE_CURRENT_BINARY_DIR}/marlin_generation.log
-      )
-
-      if (NOT marlin_generation_result EQUAL 0)
-        message(FATAL_ERROR "Marlin generation failed."
-                            " Result: \"${marlin_generation_result}\""
-                            "\nCheck the log for details: "
-                            "${CMAKE_CURRENT_BINARY_DIR}/marlin_generation.log")
-      else()
-        set(MARLIN_GEN_SCRIPT_HASH ${MARLIN_GEN_SCRIPT_HASH}
-            CACHE STRING "Last run Marlin generate script hash" FORCE)
-        message(STATUS "Marlin generation completed successfully.")
-      endif()
-    else()
-      message(STATUS "Marlin generation script has not changed, skipping generation.")
-    endif()
-
-    file(GLOB MARLIN_TEMPLATE_KERNEL_SRC "csrc/quantization/gptq_marlin/kernel_*.cu")
-    set_gencode_flags_for_srcs(
-      SRCS "${MARLIN_TEMPLATE_KERNEL_SRC}"
-      CUDA_ARCHS "${MARLIN_ARCHS}")
-
-    list(APPEND VLLM_EXT_SRC ${MARLIN_TEMPLATE_KERNEL_SRC})
-
-    set(MARLIN_SRCS
-       "csrc/quantization/marlin/dense/marlin_cuda_kernel.cu"
-       "csrc/quantization/marlin/sparse/marlin_24_cuda_kernel.cu"
-       "csrc/quantization/marlin/qqq/marlin_qqq_gemm_kernel.cu"
-       "csrc/quantization/gptq_marlin/gptq_marlin.cu"
-       "csrc/quantization/gptq_marlin/gptq_marlin_repack.cu"
-       "csrc/quantization/gptq_marlin/awq_marlin_repack.cu")
-    set_gencode_flags_for_srcs(
-      SRCS "${MARLIN_SRCS}"
-      CUDA_ARCHS "${MARLIN_ARCHS}")
-    list(APPEND VLLM_EXT_SRC "${MARLIN_SRCS}")
-    message(STATUS "Building Marlin kernels for archs: ${MARLIN_ARCHS}")
-  else()
-    message(STATUS "Not building Marlin kernels as no compatible archs found"
-                   " in CUDA target architectures")
-  endif()
-
-  # Only build AllSpark kernels if we are building for at least some compatible archs.
-  cuda_archs_loose_intersection(ALLSPARK_ARCHS "8.0;8.6;8.7;8.9" "${CUDA_ARCHS}")
-  if (ALLSPARK_ARCHS)
-    set(ALLSPARK_SRCS
-       "csrc/quantization/gptq_allspark/allspark_repack.cu"
-       "csrc/quantization/gptq_allspark/allspark_qgemm_w8a16.cu")
-    set_gencode_flags_for_srcs(
-      SRCS "${ALLSPARK_SRCS}"
-      CUDA_ARCHS "${ALLSPARK_ARCHS}")
-    list(APPEND VLLM_EXT_SRC "${ALLSPARK_SRCS}")
-    message(STATUS "Building AllSpark kernels for archs: ${ALLSPARK_ARCHS}")
-  else()
-    message(STATUS "Not building AllSpark kernels as no compatible archs found"
-                   " in CUDA target architectures")
-  endif()
-
-
-  set(SCALED_MM_3X_ARCHS)
-  # The cutlass_scaled_mm kernels for Hopper (c3x, i.e. CUTLASS 3.x) require
-  # CUDA 12.0 or later
-  cuda_archs_loose_intersection(SCALED_MM_ARCHS "9.0a;" "${CUDA_ARCHS}")
-  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.0 AND SCALED_MM_ARCHS)
-    set(SRCS
-       "csrc/quantization/cutlass_w8a8/scaled_mm_c3x_sm90.cu"
-       "csrc/quantization/cutlass_w8a8/c3x/scaled_mm_sm90_fp8.cu"
-       "csrc/quantization/cutlass_w8a8/c3x/scaled_mm_sm90_int8.cu"
-       "csrc/quantization/cutlass_w8a8/c3x/scaled_mm_azp_sm90_int8.cu"
-       "csrc/quantization/cutlass_w8a8/c3x/scaled_mm_blockwise_sm90_fp8.cu")
-    set_gencode_flags_for_srcs(
-      SRCS "${SRCS}"
-      CUDA_ARCHS "${SCALED_MM_ARCHS}")
-    list(APPEND VLLM_EXT_SRC "${SRCS}")
-    list(APPEND VLLM_GPU_FLAGS "-DENABLE_SCALED_MM_SM90=1")
-    # Let scaled_mm_c2x know it doesn't need to build these arches
-    list(APPEND SCALED_MM_3X_ARCHS "${SCALED_MM_ARCHS}")
-    message(STATUS "Building scaled_mm_c3x_sm90 for archs: ${SCALED_MM_ARCHS}")
-  else()
-    if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.0 AND SCALED_MM_ARCHS)
-      message(STATUS "Not building scaled_mm_c3x_sm90 as CUDA Compiler version is "
-                     "not >= 12.0, we recommend upgrading to CUDA 12.0 or "
-                     "later if you intend on running FP8 quantized models on "
-                     "Hopper.")
-    else()
-      message(STATUS "Not building scaled_mm_c3x_sm90 as no compatible archs found "
-                     "in CUDA target architectures")
-    endif()
-  endif()
-
-  # The cutlass_scaled_mm kernels for Blackwell SM100 (c3x, i.e. CUTLASS 3.x)
-  # require CUDA 12.8 or later
-  cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0a;10.1a" "${CUDA_ARCHS}")
-  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.8 AND SCALED_MM_ARCHS)
-    set(SRCS
+  # The cutlass_scaled_mm kernels for Blackwell (c3x, i.e. CUTLASS 3.x) require CUDA 12.8 or later
+  optional_cuda_sources(
+    NAME scaled_mm_c3x_sm100
+    MIN_VERSION 12.8
+    ARCHS "10.0a;10.1a;12.0a"
+    SRCS
       "csrc/quantization/cutlass_w8a8/scaled_mm_c3x_sm100.cu"
       "csrc/quantization/cutlass_w8a8/c3x/scaled_mm_sm100_fp8.cu"
       "csrc/quantization/cutlass_w8a8/c3x/scaled_mm_blockwise_sm100_fp8.cu"
-    )
-    set_gencode_flags_for_srcs(
-      SRCS "${SRCS}"
-      CUDA_ARCHS "${SCALED_MM_ARCHS}")
-    list(APPEND VLLM_EXT_SRC "${SRCS}")
-    list(APPEND VLLM_GPU_FLAGS "-DENABLE_SCALED_MM_SM100=1")
-    # Let scaled_mm_c2x know it doesn't need to build these arches
-    list(APPEND SCALED_MM_3X_ARCHS "${SCALED_MM_ARCHS}")
-    message(STATUS "Building scaled_mm_c3x_sm100 for archs: ${SCALED_MM_ARCHS}")
-  else()
-    if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.8 AND SCALED_MM_ARCHS)
-      message(STATUS "Not building scaled_mm_c3x_sm100 as CUDA Compiler version is "
-                     "not >= 12.8, we recommend upgrading to CUDA 12.8 or "
-                     "later if you intend on running FP8 quantized models on "
-                     "Blackwell.")
-    else()
-      message(STATUS "Not building scaled_mm_c3x_100 as no compatible archs found "
-                     "in CUDA target architectures")
-    endif()
-  endif()
+    FLAGS "-DENABLE_SCALED_MM_SM100=1"
+    VERSION_MSG
+      "Not building scaled_mm_c3x_sm100: CUDA Compiler version is not >= 12.8."
+      "Please upgrade to CUDA 12.8 or later to run FP8 quantized models on Blackwell."
+  )
 
-  #
-  # For the cutlass_scaled_mm kernels we want to build the c2x (CUTLASS 2.x)
-  # kernels for the remaining archs that are not already built for 3x.
-  # (Build 8.9 for FP8)
-  cuda_archs_loose_intersection(SCALED_MM_2X_ARCHS
-    "7.5;8.0;8.7;8.9+PTX" "${CUDA_ARCHS}")
-  # subtract out the archs that are already built for 3x
-  list(REMOVE_ITEM SCALED_MM_2X_ARCHS ${SCALED_MM_3X_ARCHS})
-  if (SCALED_MM_2X_ARCHS)
-    set(SRCS "csrc/quantization/cutlass_w8a8/scaled_mm_c2x.cu")
-    set_gencode_flags_for_srcs(
-      SRCS "${SRCS}"
-      CUDA_ARCHS "${SCALED_MM_2X_ARCHS}")
-    list(APPEND VLLM_EXT_SRC "${SRCS}")
-    list(APPEND VLLM_GPU_FLAGS "-DENABLE_SCALED_MM_C2X=1")
-    message(STATUS "Building scaled_mm_c2x for archs: ${SCALED_MM_2X_ARCHS}")
-  else()
-    if (SCALED_MM_3X_ARCHS)
-      message(STATUS "Not building scaled_mm_c2x as all archs are already built"
-                     " for and covered by scaled_mm_c3x")
-    else()
-      message(STATUS "Not building scaled_mm_c2x as no compatible archs found "
-                    "in CUDA target architectures")
-    endif()
-  endif()
+  # For the cutlass_scaled_mm kernels for Pre-hopper (c2x, i.e. CUTLASS 2.x)
+  optional_cuda_sources(
+    NAME scaled_mm_c2x
+    ARCHS "7.5;8.0;8.9+PTX"
+    SRCS "csrc/quantization/cutlass_w8a8/scaled_mm_c2x.cu"
+    FLAGS "-DENABLE_SCALED_MM_C2X=1"
+  )
 
   #
   # 2:4 Sparse Kernels
-
-  # The 2:4 sparse kernels cutlass_scaled_sparse_mm and cutlass_compressor
-  # require CUDA 12.2 or later (and only work on Hopper).
-  cuda_archs_loose_intersection(SCALED_MM_ARCHS "9.0a;" "${CUDA_ARCHS}")
-  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.2 AND SCALED_MM_ARCHS)
-    set(SRCS "csrc/sparse/cutlass/sparse_scaled_mm_c3x.cu")
-    set_gencode_flags_for_srcs(
-      SRCS "${SRCS}"
-      CUDA_ARCHS "${SCALED_MM_ARCHS}")
-    list(APPEND VLLM_EXT_SRC "${SRCS}")
-    list(APPEND VLLM_GPU_FLAGS "-DENABLE_SPARSE_SCALED_MM_C3X=1")
-    message(STATUS "Building sparse_scaled_mm_c3x for archs: ${SCALED_MM_ARCHS}")
-  else()
-    if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.2 AND SCALED_MM_ARCHS)
-      message(STATUS "Not building sparse_scaled_mm_c3x kernels as CUDA Compiler version is "
-                     "not >= 12.2, we recommend upgrading to CUDA 12.2 or later "
-                     "if you intend on running FP8 sparse quantized models on Hopper.")
-    else()
-      message(STATUS "Not building sparse_scaled_mm_c3x as no compatible archs found "
-                     "in CUDA target architectures")
-    endif()
-  endif()
+  optional_cuda_sources(
+    NAME sparse_scaled_mm_c3x
+    MIN_VERSION 12.2
+    ARCHS "9.0a;"
+    SRCS "csrc/sparse/cutlass/sparse_scaled_mm_c3x.cu"
+    FLAGS "-DENABLE_SPARSE_SCALED_MM_C3X=1"
+    VERSION_MSG
+      "Not building sparse_scaled_mm_c3x: CUDA Compiler version is not >= 12.2."
+      "Please upgrade to CUDA 12.2 or later to run FP8 sparse quantized models on Hopper."
+  )
 
   # FP4 Archs and flags
-  cuda_archs_loose_intersection(FP4_ARCHS "10.0a" "${CUDA_ARCHS}")
-  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.8 AND FP4_ARCHS)
-    set(SRCS
+  optional_cuda_sources(
+    NAME NVFP4
+    MIN_VERSION 12.8
+    ARCHS "10.0a"
+    SRCS
       "csrc/quantization/fp4/nvfp4_quant_kernels.cu"
       "csrc/quantization/fp4/nvfp4_experts_quant.cu"
       "csrc/quantization/fp4/nvfp4_scaled_mm_kernels.cu"
-      "csrc/quantization/fp4/nvfp4_blockwise_moe_kernel.cu")
-    set_gencode_flags_for_srcs(
-      SRCS "${SRCS}"
-      CUDA_ARCHS "${FP4_ARCHS}")
-    list(APPEND VLLM_EXT_SRC "${SRCS}")
-    list(APPEND VLLM_GPU_FLAGS "-DENABLE_NVFP4=1")
-    message(STATUS "Building NVFP4 for archs: ${FP4_ARCHS}")
-  else()
-    message(STATUS "Not building NVFP4 as no compatible archs were found.")
-    # clear FP4_ARCHS
-    set(FP4_ARCHS)
-  endif()
+      "csrc/quantization/fp4/nvfp4_blockwise_moe_kernel.cu"
+    FLAGS "-DENABLE_NVFP4=1"
+  )
 
   # CUTLASS MLA Archs and flags
-  cuda_archs_loose_intersection(MLA_ARCHS "10.0a" "${CUDA_ARCHS}")
-  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.8 AND MLA_ARCHS)
-    set(SRCS
-      "csrc/attention/mla/cutlass_mla_kernels.cu")
-    set_gencode_flags_for_srcs(
-      SRCS "${SRCS}"
-      CUDA_ARCHS "${MLA_ARCHS}")
-    list(APPEND VLLM_EXT_SRC "${SRCS}")
-    list(APPEND VLLM_GPU_FLAGS "-DENABLE_CUTLASS_MLA=1")
-    # Add MLA-specific include directories only to MLA source files
-    set_source_files_properties(${SRCS}
-      PROPERTIES INCLUDE_DIRECTORIES "${CUTLASS_DIR}/examples/77_blackwell_fmha;${CUTLASS_DIR}/examples/common")
-    message(STATUS "Building CUTLASS MLA for archs: ${MLA_ARCHS}")
-  else()
-    message(STATUS "Not building CUTLASS MLA as no compatible archs were found.")
-    # clear MLA_ARCHS
-    set(MLA_ARCHS)
-  endif()
+  optional_cuda_sources(
+    NAME CUTLASS_MLA
+    MIN_VERSION 12.8
+    ARCHS "10.0a"
+    SRCS "csrc/attention/mla/cutlass_mla_kernels.cu"
+    FLAGS "-DENABLE_CUTLASS_MLA=1"
+  )
+  # Add MLA-specific include directories only to MLA source files
+  set_source_files_properties(
+    "csrc/attention/mla/cutlass_mla_kernels.cu"
+    PROPERTIES INCLUDE_DIRECTORIES "${CUTLASS_DIR}/examples/77_blackwell_fmha;${CUTLASS_DIR}/examples/common"
+  )
 
   # CUTLASS MoE kernels
-
-  # The MoE kernel cutlass_moe_mm requires CUDA 12.3 or later (and ONLY works
-  # on Hopper). get_cutlass_(pplx_)moe_mm_data should only be compiled
-  # if it's possible to compile MoE kernels that use its output.
-  cuda_archs_loose_intersection(SCALED_MM_ARCHS "9.0a" "${CUDA_ARCHS}")
-  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.3 AND SCALED_MM_ARCHS)
-    set(SRCS "csrc/quantization/cutlass_w8a8/moe/grouped_mm_c3x.cu"
-             "csrc/quantization/cutlass_w8a8/moe/moe_data.cu")
-    set_gencode_flags_for_srcs(
-      SRCS "${SRCS}"
-      CUDA_ARCHS "${SCALED_MM_ARCHS}")
-    list(APPEND VLLM_EXT_SRC "${SRCS}")
-    list(APPEND VLLM_GPU_FLAGS "-DENABLE_CUTLASS_MOE_SM90=1")
-    message(STATUS "Building grouped_mm_c3x for archs: ${SCALED_MM_ARCHS}")
-  else()
-    if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.3 AND SCALED_MM_ARCHS)
-      message(STATUS "Not building grouped_mm_c3x kernels as CUDA Compiler version is "
-                     "not >= 12.3, we recommend upgrading to CUDA 12.3 or later "
-                     "if you intend on running FP8 quantized MoE models on Hopper.")
-    else()
-      message(STATUS "Not building grouped_mm_c3x as no compatible archs found "
-                     "in CUDA target architectures")
-    endif()
-  endif()
+  optional_cuda_sources(
+    NAME grouped_mm_c3x
+    MIN_VERSION 12.3
+    ARCHS "9.0a;10.0a"
+    SRCS
+      "csrc/quantization/cutlass_w8a8/moe/grouped_mm_c3x.cu"
+      "csrc/quantization/cutlass_w8a8/moe/moe_data.cu"
+    FLAGS "-DENABLE_CUTLASS_MOE_SM90=1"
+    VERSION_MSG 
+      "Not building grouped_mm_c3x kernels as CUDA Compiler is less than 12.3."
+      "We recommend upgrading to CUDA 12.3 or later if you intend on running FP8 quantized MoE models on Hopper."
+  )
 
   #
   # Machete kernels
 
-  # The machete kernels only work on hopper and require CUDA 12.0 or later.
-  # Only build Machete kernels if we are building for something compatible with sm90a
+  # Machete kernels: generate and build for supported architectures
   cuda_archs_loose_intersection(MACHETE_ARCHS "9.0a" "${CUDA_ARCHS}")
-  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.0 AND MACHETE_ARCHS)
-    #
-    # For the Machete kernels we automatically generate sources for various
-    # preselected input type pairs and schedules.
-    # Generate sources:
-    set(MACHETE_GEN_SCRIPT
-      ${CMAKE_CURRENT_SOURCE_DIR}/csrc/quantization/machete/generate.py)
-    file(MD5 ${MACHETE_GEN_SCRIPT} MACHETE_GEN_SCRIPT_HASH)
-
-    message(STATUS "Machete generation script hash: ${MACHETE_GEN_SCRIPT_HASH}")
-    message(STATUS "Last run machete generate script hash: $CACHE{MACHETE_GEN_SCRIPT_HASH}")
-
-    if (NOT DEFINED CACHE{MACHETE_GEN_SCRIPT_HASH}
-        OR NOT $CACHE{MACHETE_GEN_SCRIPT_HASH} STREQUAL ${MACHETE_GEN_SCRIPT_HASH})
-      execute_process(
-        COMMAND ${CMAKE_COMMAND} -E env
-        PYTHONPATH=${CMAKE_CURRENT_SOURCE_DIR}/csrc/cutlass_extensions/:${CUTLASS_DIR}/python/:${VLLM_PYTHON_PATH}:$PYTHONPATH
-          ${Python_EXECUTABLE} ${MACHETE_GEN_SCRIPT}
-        RESULT_VARIABLE machete_generation_result
-        OUTPUT_VARIABLE machete_generation_output
-        OUTPUT_FILE ${CMAKE_CURRENT_BINARY_DIR}/machete_generation.log
-        ERROR_FILE ${CMAKE_CURRENT_BINARY_DIR}/machete_generation.log
-      )
-
-      if (NOT machete_generation_result EQUAL 0)
-        message(FATAL_ERROR "Machete generation failed."
-                            " Result: \"${machete_generation_result}\""
-                            "\nCheck the log for details: "
-                            "${CMAKE_CURRENT_BINARY_DIR}/machete_generation.log")
-      else()
-        set(MACHETE_GEN_SCRIPT_HASH ${MACHETE_GEN_SCRIPT_HASH}
-            CACHE STRING "Last run machete generate script hash" FORCE)
-        message(STATUS "Machete generation completed successfully.")
-      endif()
-    else()
-      message(STATUS "Machete generation script has not changed, skipping generation.")
-    endif()
-
-    # Add machete generated sources
-    file(GLOB MACHETE_GEN_SOURCES "csrc/quantization/machete/generated/*.cu")
-    list(APPEND VLLM_EXT_SRC ${MACHETE_GEN_SOURCES})
-
-    # forward compatible
-    set_gencode_flags_for_srcs(
-      SRCS "${MACHETE_GEN_SOURCES}"
-      CUDA_ARCHS "${MACHETE_ARCHS}")
-
-    list(APPEND VLLM_EXT_SRC
-      csrc/quantization/machete/machete_pytorch.cu)
-
-    message(STATUS "Building Machete kernels for archs: ${MACHETE_ARCHS}")
-  else()
-    if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.0
-        AND MACHETE_ARCHS)
-      message(STATUS "Not building Machete kernels as CUDA Compiler version is "
-                     "not >= 12.0, we recommend upgrading to CUDA 12.0 or "
-                     "later if you intend on running w4a16 quantized models on "
-                     "Hopper.")
-    else()
-      message(STATUS "Not building Machete kernels as no compatible archs "
-                     "found in CUDA target architectures")
-    endif()
-  endif()
+  optional_cuda_sources(
+    NAME Machete
+    MIN_VERSION 12.0
+    ARCHS "${MACHETE_ARCHS}"
+    GEN_SCRIPT "${CMAKE_CURRENT_SOURCE_DIR}/csrc/quantization/machete/generate.py"
+    GEN_PYTHONPATH_PREPEND 
+      "${CMAKE_CURRENT_SOURCE_DIR}/csrc/cutlass_extensions/:${CUTLASS_DIR}/python/"
+    GEN_GLOB "csrc/quantization/machete/generated/*.cu"
+    SRCS "csrc/quantization/machete/machete_pytorch.cu"
+    VERSION_MSG
+      "Not building Machete kernels as CUDA Compiler version is less than 12.0."
+      "We recommend upgrading to CUDA 12.0 or later to run w4a16 quantized models on Hopper."
+  )
 # if CUDA endif
 endif()
 
@@ -670,72 +479,20 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
   list(APPEND VLLM_MOE_EXT_SRC "csrc/moe/moe_wna16.cu")
 endif()
 
+# Apply gencode flags to base MOE extension sources
 set_gencode_flags_for_srcs(
   SRCS "${VLLM_MOE_EXT_SRC}"
   CUDA_ARCHS "${CUDA_ARCHS}")
 
+## Marlin MOE kernels: generate and include for supported architectures
 if(VLLM_GPU_LANG STREQUAL "CUDA")
-  set(VLLM_MOE_WNA16_SRC
-    "csrc/moe/moe_wna16.cu")
-
-  set_gencode_flags_for_srcs(
-    SRCS "${VLLM_MOE_WNA16_SRC}"
-    CUDA_ARCHS "${CUDA_ARCHS}")
-
-  list(APPEND VLLM_MOE_EXT_SRC "${VLLM_MOE_WNA16_SRC}")
-  # 9.0 for latest bf16 atomicAdd PTX
-  cuda_archs_loose_intersection(MARLIN_MOE_ARCHS "8.0;8.7;9.0+PTX" "${CUDA_ARCHS}")
-  if (MARLIN_MOE_ARCHS)
-
-    #
-    # For the Marlin MOE kernels we automatically generate sources for various
-    # preselected input type pairs and schedules.
-    # Generate sources:
-    set(MOE_MARLIN_GEN_SCRIPT
-      ${CMAKE_CURRENT_SOURCE_DIR}/csrc/moe/marlin_moe_wna16/generate_kernels.py)
-    file(MD5 ${MOE_MARLIN_GEN_SCRIPT} MOE_MARLIN_GEN_SCRIPT_HASH)
-
-    message(STATUS "Marlin MOE generation script hash: ${MOE_MARLIN_GEN_SCRIPT_HASH}")
-    message(STATUS "Last run Marlin MOE generate script hash: $CACHE{MOE_MARLIN_GEN_SCRIPT_HASH}")
-
-    if (NOT DEFINED CACHE{MOE_MARLIN_GEN_SCRIPT_HASH}
-        OR NOT $CACHE{MOE_MARLIN_GEN_SCRIPT_HASH} STREQUAL ${MOE_MARLIN_GEN_SCRIPT_HASH})
-      execute_process(
-        COMMAND ${CMAKE_COMMAND} -E env
-        PYTHONPATH=$PYTHONPATH
-          ${Python_EXECUTABLE} ${MOE_MARLIN_GEN_SCRIPT}
-        RESULT_VARIABLE moe_marlin_generation_result
-        OUTPUT_VARIABLE moe_marlin_generation_output
-        OUTPUT_FILE ${CMAKE_CURRENT_BINARY_DIR}/moe_marlin_generation.log
-        ERROR_FILE ${CMAKE_CURRENT_BINARY_DIR}/moe_marlin_generation.log
-      )
-
-      if (NOT moe_marlin_generation_result EQUAL 0)
-        message(FATAL_ERROR "Marlin MOE generation failed."
-                            " Result: \"${moe_marlin_generation_result}\""
-                            "\nCheck the log for details: "
-                            "${CMAKE_CURRENT_BINARY_DIR}/moe_marlin_generation.log")
-      else()
-        set(MOE_MARLIN_GEN_SCRIPT_HASH ${MOE_MARLIN_GEN_SCRIPT_HASH}
-            CACHE STRING "Last run Marlin MOE generate script hash" FORCE)
-        message(STATUS "Marlin MOE generation completed successfully.")
-      endif()
-    else()
-      message(STATUS "Marlin MOE generation script has not changed, skipping generation.")
-    endif()
-
-    file(GLOB MOE_WNAA16_MARLIN_SRC "csrc/moe/marlin_moe_wna16/*.cu")
-    set_gencode_flags_for_srcs(
-      SRCS "${MOE_WNAA16_MARLIN_SRC}"
-      CUDA_ARCHS "${MARLIN_MOE_ARCHS}")
-
-    list(APPEND VLLM_MOE_EXT_SRC ${MOE_WNAA16_MARLIN_SRC})
-
-    message(STATUS "Building Marlin MOE kernels for archs: ${MARLIN_MOE_ARCHS}")
-  else()
-    message(STATUS "Not building Marlin MOE kernels as no compatible archs found"
-                   " in CUDA target architectures")
-  endif()
+  optional_cuda_sources(
+    NAME "Marlin MOE"
+    ARCHS "8.0;9.0+PTX"
+    GEN_SCRIPT "${CMAKE_CURRENT_SOURCE_DIR}/csrc/moe/marlin_moe_wna16/generate_kernels.py"
+    GEN_GLOB "csrc/moe/marlin_moe_wna16/*.cu"
+    OUT_SRCS_VAR VLLM_MOE_EXT_SRC
+  )
 endif()
 
 if(VLLM_GPU_LANG STREQUAL "CUDA")
@@ -785,7 +542,5 @@ endif()
 # For CUDA we also build and ship some external projects.
 if (VLLM_GPU_LANG STREQUAL "CUDA")
     include(cmake/external_projects/flashmla.cmake)
-
-    # vllm-flash-attn should be last as it overwrites some CMake functions
     include(cmake/external_projects/vllm_flash_attn.cmake)
 endif ()

README.md

@@ -58,8 +58,8 @@ vLLM is fast with:
 - Efficient management of attention key and value memory with [**PagedAttention**](https://blog.vllm.ai/2023/06/20/vllm.html)
 - Continuous batching of incoming requests
 - Fast model execution with CUDA/HIP graph
-- Quantizations: [GPTQ](https://arxiv.org/abs/2210.17323), [AWQ](https://arxiv.org/abs/2306.00978), [AutoRound](https://arxiv.org/abs/2309.05516), INT4, INT8, and FP8
-- Optimized CUDA kernels, including integration with FlashAttention and FlashInfer
+- Quantizations: [GPTQ](https://arxiv.org/abs/2210.17323), [AWQ](https://arxiv.org/abs/2306.00978), [AutoRound](https://arxiv.org/abs/2309.05516),INT4, INT8, and FP8.
+- Optimized CUDA kernels, including integration with FlashAttention and FlashInfer.
 - Speculative decoding
 - Chunked prefill
 
@@ -72,14 +72,14 @@ vLLM is flexible and easy to use with:
 - Tensor parallelism and pipeline parallelism support for distributed inference
 - Streaming outputs
 - OpenAI-compatible API server
-- Support NVIDIA GPUs, AMD CPUs and GPUs, Intel CPUs and GPUs, PowerPC CPUs, TPU, and AWS Neuron
+- Support NVIDIA GPUs, AMD CPUs and GPUs, Intel CPUs and GPUs, PowerPC CPUs, TPU, and AWS Neuron.
 - Prefix caching support
 - Multi-LoRA support
 
 vLLM seamlessly supports most popular open-source models on HuggingFace, including:
 - Transformer-like LLMs (e.g., Llama)
 - Mixture-of-Expert LLMs (e.g., Mixtral, Deepseek-V2 and V3)
-- Embedding Models (e.g., E5-Mistral)
+- Embedding Models (e.g. E5-Mistral)
 - Multi-modal LLMs (e.g., LLaVA)
 
 Find the full list of supported models [here](https://docs.vllm.ai/en/latest/models/supported_models.html).
@@ -154,14 +154,12 @@ If you use vLLM for your research, please cite our [paper](https://arxiv.org/abs
 
 ## Contact Us
 
-<!-- --8<-- [start:contact-us] -->
 - For technical questions and feature requests, please use GitHub [Issues](https://github.com/vllm-project/vllm/issues) or [Discussions](https://github.com/vllm-project/vllm/discussions)
 - For discussing with fellow users, please use the [vLLM Forum](https://discuss.vllm.ai)
-- For coordinating contributions and development, please use [Slack](https://slack.vllm.ai)
+- coordinating contributions and development, please use [Slack](https://slack.vllm.ai)
 - For security disclosures, please use GitHub's [Security Advisories](https://github.com/vllm-project/vllm/security/advisories) feature
 - For collaborations and partnerships, please contact us at [vllm-questions@lists.berkeley.edu](mailto:vllm-questions@lists.berkeley.edu)
-<!-- --8<-- [end:contact-us] -->
 
 ## Media Kit
 
-- If you wish to use vLLM's logo, please refer to [our media kit repo](https://github.com/vllm-project/media-kit)
+- If you wish to use vLLM's logo, please refer to [our media kit repo](https://github.com/vllm-project/media-kit).

SECURITY.md

@@ -8,6 +8,4 @@ Please report security issues privately using [the vulnerability submission form
 
 ---
 
-Please see the [Security Guide in the vLLM documentation](https://docs.vllm.ai/en/latest/usage/security.html) for more information on vLLM's security assumptions and recommendations.
-
 Please see [PyTorch's Security Policy](https://github.com/pytorch/pytorch/blob/main/SECURITY.md) for more information and recommendations on how to securely interact with models.

benchmarks/README.md

@@ -64,12 +64,6 @@ become available.
       <td style="text-align: center;">✅</td>
       <td><code>lmms-lab/LLaVA-OneVision-Data</code>, <code>Aeala/ShareGPT_Vicuna_unfiltered</code></td>
     </tr>
-    <tr>
-      <td><strong>Custom</strong></td>
-      <td style="text-align: center;">✅</td>
-      <td style="text-align: center;">✅</td>
-      <td>Local file: <code>data.jsonl</code></td>
-    </tr>
   </tbody>
 </table>
 
@@ -130,38 +124,6 @@ P99 ITL (ms):                            8.39
 ==================================================
 ```
 
-### Custom Dataset
-If the dataset you want to benchmark is not supported yet in vLLM, even then you can benchmark on it using `CustomDataset`. Your data needs to be in `.jsonl` format and needs to have "prompt" field per entry, e.g., data.jsonl
-
-```
-{"prompt": "What is the capital of India?"}
-{"prompt": "What is the capital of Iran?"}
-{"prompt": "What is the capital of China?"}
-``` 
-
-```bash
-# start server
-VLLM_USE_V1=1 vllm serve meta-llama/Llama-3.1-8B-Instruct --disable-log-requests
-```
-
-```bash
-# run benchmarking script
-python3 benchmarks/benchmark_serving.py --port 9001 --save-result --save-detailed \
-  --backend vllm \
-  --model meta-llama/Llama-3.1-8B-Instruct \
-  --endpoint /v1/completions \
-  --dataset-name custom \
-  --dataset-path <path-to-your-data-jsonl> \
-  --custom-skip-chat-template \
-  --num-prompts 80 \
-  --max-concurrency 1 \
-  --temperature=0.3 \
-  --top-p=0.75 \
-  --result-dir "./log/"
-```
-
-You can skip applying chat template if your data already has it by using `--custom-skip-chat-template`.
-
 ### VisionArena Benchmark for Vision Language Models
 
 ```bash
@@ -184,9 +146,9 @@ python3 vllm/benchmarks/benchmark_serving.py \
 
 ``` bash
 VLLM_USE_V1=1 vllm serve meta-llama/Meta-Llama-3-8B-Instruct \
-    --speculative-config $'{"method": "ngram",
-    "num_speculative_tokens": 5, "prompt_lookup_max": 5,
-    "prompt_lookup_min": 2}'
+    --ngram_prompt_lookup_min 2 \
+    --ngram-prompt-lookup-max 5 \
+    --speculative_config '{"model": "[ngram]", "num_speculative_tokens": 5}
 ```
 
 ``` bash
@@ -241,16 +203,6 @@ python3 vllm/benchmarks/benchmark_serving.py \
     --seed 42
 ```
 
-**`philschmid/mt-bench`**
-
-``` bash
-python3 vllm/benchmarks/benchmark_serving.py \
-    --model Qwen/QwQ-32B \
-    --dataset-name hf \
-    --dataset-path philschmid/mt-bench \
-    --num-prompts 80
-```
-
 ### Running With Sampling Parameters
 
 When using OpenAI-compatible backends such as `vllm`, optional sampling
@@ -321,9 +273,9 @@ python3 vllm/benchmarks/benchmark_throughput.py \
     --output-len=100 \
     --num-prompts=2048 \
     --async-engine \
-    --speculative-config $'{"method": "ngram",
-    "num_speculative_tokens": 5, "prompt_lookup_max": 5,
-    "prompt_lookup_min": 2}'
+    --ngram_prompt_lookup_min=2 \
+    --ngram-prompt-lookup-max=5 \
+    --speculative_config '{"model": "[ngram]", "num_speculative_tokens": 5}
 ```
 
 ```
@@ -387,178 +339,3 @@ python3 vllm/benchmarks/benchmark_throughput.py \
   --enable-lora \
   --lora-path yard1/llama-2-7b-sql-lora-test
   ```
-
----
-## Example - Structured Output Benchmark
-
-Benchmark the performance of structured output generation (JSON, grammar, regex).
-
-### Server Setup
-
-```bash
-vllm serve NousResearch/Hermes-3-Llama-3.1-8B --disable-log-requests
-```
-
-### JSON Schema Benchmark
-
-```bash
-python3 benchmarks/benchmark_serving_structured_output.py \
-  --backend vllm \
-  --model NousResearch/Hermes-3-Llama-3.1-8B \
-  --dataset json \
-  --structured-output-ratio 1.0 \
-  --request-rate 10 \
-  --num-prompts 1000
-```
-
-### Grammar-based Generation Benchmark
-
-```bash
-python3 benchmarks/benchmark_serving_structured_output.py \
-  --backend vllm \
-  --model NousResearch/Hermes-3-Llama-3.1-8B \
-  --dataset grammar \
-  --structure-type grammar \
-  --request-rate 10 \
-  --num-prompts 1000
-```
-
-### Regex-based Generation Benchmark
-
-```bash
-python3 benchmarks/benchmark_serving_structured_output.py \
-  --backend vllm \
-  --model NousResearch/Hermes-3-Llama-3.1-8B \
-  --dataset regex \
-  --request-rate 10 \
-  --num-prompts 1000
-```
-
-### Choice-based Generation Benchmark
-
-```bash
-python3 benchmarks/benchmark_serving_structured_output.py \
-  --backend vllm \
-  --model NousResearch/Hermes-3-Llama-3.1-8B \
-  --dataset choice \
-  --request-rate 10 \
-  --num-prompts 1000
-```
-
-### XGrammar Benchmark Dataset
-
-```bash
-python3 benchmarks/benchmark_serving_structured_output.py \
-  --backend vllm \
-  --model NousResearch/Hermes-3-Llama-3.1-8B \
-  --dataset xgrammar_bench \
-  --request-rate 10 \
-  --num-prompts 1000
-```
-
----
-## Example - Long Document QA Throughput Benchmark
-
-Benchmark the performance of long document question-answering with prefix caching.
-
-### Basic Long Document QA Test
-
-```bash
-python3 benchmarks/benchmark_long_document_qa_throughput.py \
-  --model meta-llama/Llama-2-7b-chat-hf \
-  --enable-prefix-caching \
-  --num-documents 16 \
-  --document-length 2000 \
-  --output-len 50 \
-  --repeat-count 5
-```
-
-### Different Repeat Modes
-
-```bash
-# Random mode (default) - shuffle prompts randomly
-python3 benchmarks/benchmark_long_document_qa_throughput.py \
-  --model meta-llama/Llama-2-7b-chat-hf \
-  --enable-prefix-caching \
-  --num-documents 8 \
-  --document-length 3000 \
-  --repeat-count 3 \
-  --repeat-mode random
-
-# Tile mode - repeat entire prompt list in sequence
-python3 benchmarks/benchmark_long_document_qa_throughput.py \
-  --model meta-llama/Llama-2-7b-chat-hf \
-  --enable-prefix-caching \
-  --num-documents 8 \
-  --document-length 3000 \
-  --repeat-count 3 \
-  --repeat-mode tile
-
-# Interleave mode - repeat each prompt consecutively
-python3 benchmarks/benchmark_long_document_qa_throughput.py \
-  --model meta-llama/Llama-2-7b-chat-hf \
-  --enable-prefix-caching \
-  --num-documents 8 \
-  --document-length 3000 \
-  --repeat-count 3 \
-  --repeat-mode interleave
-```
-
----
-## Example - Prefix Caching Benchmark
-
-Benchmark the efficiency of automatic prefix caching.
-
-### Fixed Prompt with Prefix Caching
-
-```bash
-python3 benchmarks/benchmark_prefix_caching.py \
-  --model meta-llama/Llama-2-7b-chat-hf \
-  --enable-prefix-caching \
-  --num-prompts 1 \
-  --repeat-count 100 \
-  --input-length-range 128:256
-```
-
-### ShareGPT Dataset with Prefix Caching
-
-```bash
-# download dataset
-# wget https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/resolve/main/ShareGPT_V3_unfiltered_cleaned_split.json
-
-python3 benchmarks/benchmark_prefix_caching.py \
-  --model meta-llama/Llama-2-7b-chat-hf \
-  --dataset-path /path/ShareGPT_V3_unfiltered_cleaned_split.json \
-  --enable-prefix-caching \
-  --num-prompts 20 \
-  --repeat-count 5 \
-  --input-length-range 128:256
-```
-
----
-## Example - Request Prioritization Benchmark
-
-Benchmark the performance of request prioritization in vLLM.
-
-### Basic Prioritization Test
-
-```bash
-python3 benchmarks/benchmark_prioritization.py \
-  --model meta-llama/Llama-2-7b-chat-hf \
-  --input-len 128 \
-  --output-len 64 \
-  --num-prompts 100 \
-  --scheduling-policy priority
-```
-
-### Multiple Sequences per Prompt
-
-```bash
-python3 benchmarks/benchmark_prioritization.py \
-  --model meta-llama/Llama-2-7b-chat-hf \
-  --input-len 128 \
-  --output-len 64 \
-  --num-prompts 100 \
-  --scheduling-policy priority \
-  --n 2
-```

benchmarks/auto_tune.sh

@@ -10,16 +10,11 @@
 # 3. Set variables (ALL REQUIRED)
 #   BASE: your directory for vllm repo
 #   MODEL: the model served by vllm
-#   SYSTEM: the hardware, choice TPU or GPU, for other systems, "get best profile" might not support.
-#   TP: ways of tensor parallelism
 #   DOWNLOAD_DIR: directory to download and load model weights.
 #   INPUT_LEN: request input len
 #   OUTPUT_LEN: request output len
 #   MIN_CACHE_HIT_PCT: prefix cache rate
 #   MAX_LATENCY_ALLOWED_MS: (e2e) latency requirement. If there's no latency requirement, set it to a large number like 1000000000
-#   NUM_SEQS_LIST: a list of `max-num-seqs` you want to loop with.
-#   NUM_BATCHED_TOKENS_LIST: a list of `max-num-batched-tokens` you want to loop with.
-#   Note that the default NUM_SEQS_LIST and NUM_BATCHED_TOKENS_LIST are set for medium size input/output len, for extra short context (such as 20:20), you might need to include larger numbers in NUM_SEQS_LIST.
 # 4. Run the script, it might take a long time, you can use tmux to avoid the script stop if disconnection happens.
 # 5. The final result will be saved in RESULT file. 
 
@@ -35,31 +30,31 @@
 TAG=$(date +"%Y_%m_%d_%H_%M")
 BASE=""
 MODEL="meta-llama/Llama-3.1-8B-Instruct"
-SYSTEM="TPU"
-TP=1
 DOWNLOAD_DIR=""
 INPUT_LEN=4000
 OUTPUT_LEN=16
-MIN_CACHE_HIT_PCT=0
+MIN_CACHE_HIT_PCT_PCT=0
 MAX_LATENCY_ALLOWED_MS=100000000000
-NUM_SEQS_LIST="128 256"
-NUM_BATCHED_TOKENS_LIST="512 1024 2048 4096"
 
 LOG_FOLDER="$BASE/auto-benchmark/$TAG"
 RESULT="$LOG_FOLDER/result.txt"
-PROFILE_PATH="$LOG_FOLDER/profile"
 
-echo "result file: $RESULT"
+echo "result file$ $RESULT"
 echo "model: $MODEL"
+echo
 
 rm -rf $LOG_FOLDER
-rm -rf $PROFILE_PATH
 mkdir -p $LOG_FOLDER
-mkdir -p $PROFILE_PATH
 
 cd "$BASE/vllm"
+# create sonnet-4x.txt so that we can sample 2048 tokens for input
+echo "" > benchmarks/sonnet_4x.txt
+for _ in {1..4}
+do
+cat benchmarks/sonnet.txt >> benchmarks/sonnet_4x.txt
+done
 
-pip install -q datasets
+pip install datasets
 
 current_hash=$(git rev-parse HEAD)
 echo "hash:$current_hash" >> "$RESULT"
@@ -69,88 +64,53 @@ best_throughput=0
 best_max_num_seqs=0
 best_num_batched_tokens=0
 best_goodput=0
-
-start_server() {
-    local gpu_memory_utilization=$1
-    local max_num_seqs=$2
-    local max_num_batched_tokens=$3
-    local vllm_log=$4
-    local profile_dir=$5
-    
-    pkill -f vllm
-
-    VLLM_USE_V1=1 VLLM_SERVER_DEV_MODE=1 VLLM_TORCH_PROFILER_DIR=$profile_dir vllm serve $MODEL \
-        --disable-log-requests \
-        --port 8004 \
-        --gpu-memory-utilization $gpu_memory_utilization \
-        --max-num-seqs $max_num_seqs \
-        --max-num-batched-tokens $max_num_batched_tokens \
-        --tensor-parallel-size $TP \
-        --enable-prefix-caching \
-        --load-format dummy \
-        --download-dir "$DOWNLOAD_DIR" \
-        --max-model-len $(( INPUT_LEN+OUTPUT_LEN )) > "$vllm_log" 2>&1 &
-
-    # wait for 10 minutes...
-    server_started=0
-    for i in {1..60}; do  
-        RESPONSE=$(curl -s -X GET "http://0.0.0.0:8004/health" -w "%{http_code}" -o /dev/stdout)
-        STATUS_CODE=$(echo "$RESPONSE" | tail -n 1) 
-        if [[ "$STATUS_CODE" -eq 200 ]]; then
-            server_started=1
-            break
-        else
-            sleep 10
-        fi
-    done
-    if (( ! server_started )); then
-        echo "server did not start within 10 minutes. Please check server log at $vllm_log".
-        return 1
-    else
-        return 0
-    fi
-}
-
-update_best_profile() {
-    local profile_dir=$1
-    local profile_index=$2
-    sorted_paths=($(find "$profile_dir" -maxdepth 1 -not -path "$profile_dir" | sort))
-    selected_profile_file=
-    if [[ "$SYSTEM" == "TPU" ]]; then
-        selected_profile_file="${sorted_paths[$profile_index]}/*.xplane.pb"
-    fi 
-    if [[ "$SYSTEM" == "GPU" ]]; then
-        selected_profile_file="${sorted_paths[$profile_index]}"
-    fi 
-    rm -f $PROFILE_PATH/*
-    cp $selected_profile_file $PROFILE_PATH
-}
-
 run_benchmark() {
     local max_num_seqs=$1
     local max_num_batched_tokens=$2
-    local gpu_memory_utilization=$3
     echo "max_num_seq: $max_num_seqs, max_num_batched_tokens: $max_num_batched_tokens"
     local vllm_log="$LOG_FOLDER/vllm_log_${max_num_seqs}_${max_num_batched_tokens}.txt"
-    local profile_dir="$LOG_FOLDER/profile_${max_num_seqs}_${max_num_batched_tokens}"
     echo "vllm_log: $vllm_log"
     echo
     rm -f $vllm_log
-    mkdir -p $profile_dir
-    pkill -f vllm
-    local profile_index=0
 
-    echo "starting server..."
-    start_server $gpu_memory_utilization $max_num_seqs $max_num_batched_tokens $vllm_log $profile_dir
-    result=$?
-    if [[ "$result" -eq 1 ]]; then
-        echo "server failed to start. gpu_memory_utilization:$gpu_memory_utilization, max_num_seqs:$max_num_seqs, max_num_batched_tokens: $max_num_batched_tokens"
-    else
-        echo "server started."
-    fi
+    # start the server
+    VLLM_USE_V1=1 VLLM_SERVER_DEV_MODE=1 vllm serve $MODEL \
+        --disable-log-requests \
+        --port 8004 \
+        --gpu-memory-utilization 0.98 \
+        --max-num-seqs $max_num_seqs \
+        --max-num-batched-tokens $max_num_batched_tokens \
+        --tensor-parallel-size 1 \
+        --enable-prefix-caching \
+        --load-format dummy \
+        --download-dir $DOWNLOAD_DIR \
+        --max-model-len $(( INPUT_LEN+OUTPUT_LEN )) > "$vllm_log" 2>&1 &
+    echo "wait for 10 minutes.."
     echo
+    # wait for 10 minutes...
+    server_started=0
+    for i in {1..60}; do        
+        if grep -Fq "Application startup complete" "$vllm_log"; then
+            echo "Application started"
+            server_started=1
+            break
+        else
+            # echo "wait for 10 seconds..."
+            sleep 10
+        fi
+    done
+ 
+    if (( ! server_started )); then
+        echo "server did not start within 10 minutes, terminate the benchmarking. Please check server log at $vllm_log"
+        echo "pkill -f vllm"
+        echo
+        pkill vllm
+        sleep 10
+        return 1
+    fi
     
     echo "run benchmark test..."
+    echo
     meet_latency_requirement=0
     # get a basic qps by using request-rate inf
     bm_log="$LOG_FOLDER/bm_log_${max_num_seqs}_${max_num_batched_tokens}_requestrate_inf.txt"
@@ -158,32 +118,30 @@ run_benchmark() {
     python benchmarks/benchmark_serving.py \
         --backend vllm \
         --model $MODEL  \
-        --dataset-name random \
-        --random-input-len $INPUT_LEN \
-        --random-output-len $OUTPUT_LEN \
+        --dataset-name sonnet \
+        --dataset-path benchmarks/sonnet_4x.txt \
+        --sonnet-input-len $INPUT_LEN \
+        --sonnet-output-len $OUTPUT_LEN \
         --ignore-eos \
         --disable-tqdm \
         --request-rate inf \
         --percentile-metrics ttft,tpot,itl,e2el \
         --goodput e2el:$MAX_LATENCY_ALLOWED_MS \
-        --num-prompts 1000 \
-        --random-prefix-len $prefix_len \
-        --port 8004 \
-        --profile &> "$bm_log"
-    throughput=$(grep "Request throughput (req/s):" "$bm_log" | sed 's/[^0-9.]//g')
+        --num-prompts 100 \
+        --sonnet-prefix-len $prefix_len \
+        --port 8004 > "$bm_log"
+    through_put=$(grep "Request throughput (req/s):" "$bm_log" | sed 's/[^0-9.]//g')
     e2el=$(grep "P99 E2EL (ms):" "$bm_log" | awk '{print $NF}')
     goodput=$(grep "Request goodput (req/s):" "$bm_log" | sed 's/[^0-9.]//g')
 
     if (( $(echo "$e2el <= $MAX_LATENCY_ALLOWED_MS" | bc -l) )); then
         meet_latency_requirement=1
-        request_rate=inf
     fi
 
     if (( ! meet_latency_requirement )); then
-    # start from request-rate as int(throughput) + 1
-        request_rate=$((${throughput%.*} + 1))
+    # start from request-rate as int(through_put) + 1
+        request_rate=$((${through_put%.*} + 1))
         while ((request_rate > 0)); do
-            profile_index=$((profile_index+1))
             # clear prefix cache
             curl -X POST http://0.0.0.0:8004/reset_prefix_cache
             sleep 5
@@ -191,18 +149,19 @@ run_benchmark() {
             python benchmarks/benchmark_serving.py \
                 --backend vllm \
                 --model $MODEL  \
-                --dataset-name random \
-                --random-input-len $INPUT_LEN \
-                --random-output-len $OUTPUT_LEN \
-                --ignore-eos \
+                --dataset-name sonnet \
+                --dataset-path benchmarks/sonnet_4x.txt \
+                --sonnet-input-len $INPUT_LEN \
+                --sonnet-output-len $OUTPUT_LEN \
+                --ignore_eos \
                 --disable-tqdm \
                 --request-rate $request_rate \
                 --percentile-metrics ttft,tpot,itl,e2el \
                 --goodput e2el:$MAX_LATENCY_ALLOWED_MS \
                 --num-prompts 100 \
-                --random-prefix-len $prefix_len \
-                --port 8004 &> "$bm_log"
-            throughput=$(grep "Request throughput (req/s):" "$bm_log" | sed 's/[^0-9.]//g')
+                --sonnet-prefix-len $prefix_len \
+                --port 8004 > "$bm_log"
+            through_put=$(grep "Request throughput (req/s):" "$bm_log" | sed 's/[^0-9.]//g')
             e2el=$(grep "P99 E2EL (ms):" "$bm_log" | awk '{print $NF}')
             goodput=$(grep "Request goodput (req/s):" "$bm_log" | sed 's/[^0-9.]//g')
             if (( $(echo "$e2el <= $MAX_LATENCY_ALLOWED_MS" | bc -l) )); then
@@ -214,19 +173,13 @@ run_benchmark() {
     fi
     # write the results and update the best result.
     if ((meet_latency_requirement)); then
-        echo "max_num_seqs: $max_num_seqs, max_num_batched_tokens: $max_num_batched_tokens, request_rate: $request_rate, e2el: $e2el, throughput: $throughput, goodput: $goodput"
-        echo "max_num_seqs: $max_num_seqs, max_num_batched_tokens: $max_num_batched_tokens, request_rate: $request_rate, e2el: $e2el, throughput: $throughput, goodput: $goodput" >> "$RESULT"
-        if (( $(echo "$throughput > $best_throughput" | bc -l) )); then
-            best_throughput=$throughput
+        echo "max_num_seqs: $max_num_seqs, max_num_batched_tokens: $max_num_batched_tokens, request_rate: $request_rate, e2el: $e2el, through put: $through_put, goodput: $goodput"
+        echo "max_num_seqs: $max_num_seqs, max_num_batched_tokens: $max_num_batched_tokens, request_rate: $request_rate, e2el: $e2el, through put: $through_put, goodput: $goodput" >> "$RESULT"
+        if (( $(echo "$through_put > $best_throughput" | bc -l) )); then
+            best_throughput=$through_put
             best_max_num_seqs=$max_num_seqs
             best_num_batched_tokens=$max_num_batched_tokens
             best_goodput=$goodput
-            if [[ "$SYSTEM" == "TPU" ]]; then
-                update_best_profile "$profile_dir/plugins/profile" $profile_index
-            fi
-            if [[ "$SYSTEM" == "GPU" ]]; then
-                update_best_profile "$profile_dir" $profile_index
-            fi
         fi
     else
         echo "max_num_seqs: $max_num_seqs, max_num_batched_tokens: $max_num_batched_tokens does not meet latency requirement ${MAX_LATENCY_ALLOWED_MS}"
@@ -235,42 +188,25 @@ run_benchmark() {
 
     echo "best_max_num_seqs: $best_max_num_seqs, best_num_batched_tokens: $best_num_batched_tokens, best_throughput: $best_throughput"
 
+    echo "pkill -f vllm"
+    echo
     pkill vllm
     sleep 10
+    rm -f $vllm_log
     printf '=%.0s' $(seq 1 20)
     return 0
 }
 
-read -r -a num_seqs_list <<< "$NUM_SEQS_LIST"
-read -r -a num_batched_tokens_list <<< "$NUM_BATCHED_TOKENS_LIST"
 
-# first find out the max gpu-memory-utilization without HBM OOM.
-gpu_memory_utilization=0.98
-find_gpu_memory_utilization=0
-while (( $(echo "$gpu_memory_utilization >= 0.9" | bc -l) )); do
-    start_server $gpu_memory_utilization "${num_seqs_list[-1]}" "${num_batched_tokens_list[-1]}" "$LOG_FOLDER/vllm_log_gpu_memory_utilization_$gpu_memory_utilization.log"
-    result=$?
-    if [[ "$result" -eq 0 ]]; then
-        find_gpu_memory_utilization=1
-        break
-    else
-        gpu_memory_utilization=$(echo "$gpu_memory_utilization - 0.01" | bc)
-    fi
-done
-
-if [[ "$find_gpu_memory_utilization" -eq 1 ]]; then
-    echo "Using gpu_memory_utilization=$gpu_memory_utilization to serve model."
-else
-    echo "Cannot find a proper gpu_memory_utilization over 0.9 to serve the model, please check logs in $LOG_FOLDER."
-    exit 1
-fi
-
-for num_seqs in "${num_seqs_list[@]}"; do
-    for num_batched_tokens in "${num_batched_tokens_list[@]}"; do
-        run_benchmark $num_seqs $num_batched_tokens $gpu_memory_utilization
+num_seqs_list="128 256"
+num_batched_tokens_list="512 1024 2048 4096"
+for num_seqs in $num_seqs_list; do
+    for num_batched_tokens in $num_batched_tokens_list; do
+        run_benchmark $num_seqs $num_batched_tokens
+        exit 0
     done
 done
 echo "finish permutations"
-echo "best_max_num_seqs: $best_max_num_seqs, best_num_batched_tokens: $best_num_batched_tokens, best_throughput: $best_throughput, profile saved in: $PROFILE_PATH"
-echo "best_max_num_seqs: $best_max_num_seqs, best_num_batched_tokens: $best_num_batched_tokens, best_throughput: $best_throughput, profile saved in: $PROFILE_PATH" >> "$RESULT"
+echo "best_max_num_seqs: $best_max_num_seqs, best_num_batched_tokens: $best_num_batched_tokens, best_throughput: $best_throughput"
+echo "best_max_num_seqs: $best_max_num_seqs, best_num_batched_tokens: $best_num_batched_tokens, best_throughput: $best_throughput" >> "$RESULT"
 

benchmarks/backend_request_func.py

@@ -1,5 +1,4 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 import io
 import json
@@ -325,7 +324,7 @@ async def async_request_openai_completions(
 
                                 most_recent_timestamp = timestamp
                                 generated_text += text or ""
-                            if usage := data.get("usage"):
+                            elif usage := data.get("usage"):
                                 output.output_tokens = usage.get("completion_tokens")
                     if first_chunk_received:
                         output.success = True
@@ -404,14 +403,8 @@ async def async_request_openai_chat_completions(
                         chunk_bytes = chunk_bytes.strip()
                         if not chunk_bytes:
                             continue
-                        chunk_bytes = chunk_bytes.decode("utf-8")
-                        # NOTE: SSE comments (often used as pings) start with a colon.
-                        # These are not JSON data payload and should be skipped.
-                        if chunk_bytes.startswith(":"):
-                            continue
-
-                        chunk = chunk_bytes.removeprefix("data: ")
 
+                        chunk = chunk_bytes.decode("utf-8").removeprefix("data: ")
                         if chunk != "[DONE]":
                             timestamp = time.perf_counter()
                             data = json.loads(chunk)
@@ -618,7 +611,6 @@ ASYNC_REQUEST_FUNCS = {
     "tensorrt-llm": async_request_trt_llm,
     "scalellm": async_request_openai_completions,
     "sglang": async_request_openai_completions,
-    "llama.cpp": async_request_openai_completions,
 }
 
 OPENAI_COMPATIBLE_BACKENDS = [

benchmarks/benchmark_dataset.py

@@ -1,5 +1,4 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 """
 This module defines a framework for sampling benchmark requests from various
 datasets. Each dataset subclass of BenchmarkDataset must implement sample
@@ -10,6 +9,9 @@ generation. Supported dataset types include:
   - BurstGPT
   - HuggingFace
   - VisionArena
+
+TODO: Implement CustomDataset to parse a JSON file and convert its contents into
+SampleRequest instances, similar to the approach used in ShareGPT.
 """
 
 import base64
@@ -353,7 +355,7 @@ class RandomDataset(BenchmarkDataset):
                 : input_lens[i]
             ]
             prompt = tokenizer.decode(re_encoded_sequence)
-            total_input_len = len(re_encoded_sequence)
+            total_input_len = prefix_len + int(input_lens[i])
             requests.append(
                 SampleRequest(
                     prompt=prompt,
@@ -440,97 +442,6 @@ class ShareGPTDataset(BenchmarkDataset):
         return samples
 
 
-# -----------------------------------------------------------------------------
-# Custom Dataset Implementation
-# -----------------------------------------------------------------------------
-
-
-class CustomDataset(BenchmarkDataset):
-    """
-    Implements the Custom dataset.  Loads data from a JSONL file and generates
-    sample requests based on conversation turns. E.g.,
-    ```
-    {"prompt": "What is the capital of India?"}
-    {"prompt": "What is the capital of Iran?"}
-    {"prompt": "What is the capital of China?"}
-    ```
-    """
-
-    def __init__(self, **kwargs) -> None:
-        super().__init__(**kwargs)
-        self.load_data()
-
-    def load_data(self) -> None:
-        if self.dataset_path is None:
-            raise ValueError("dataset_path must be provided for loading data.")
-
-        # self.data will be a list of dictionaries
-        # e.g., [{"prompt": "What is the capital of India?"}, ...]
-        # This will be the standardized format which load_data()
-        # has to convert into depending on the filetype of dataset_path.
-        # sample() will assume this standardized format of self.data
-        self.data = []
-
-        # Load the JSONL file
-        if self.dataset_path.endswith(".jsonl"):
-            jsonl_data = pd.read_json(path_or_buf=self.dataset_path, lines=True)
-
-            # check if the JSONL file has a 'prompt' column
-            if "prompt" not in jsonl_data.columns:
-                raise ValueError("JSONL file must contain a 'prompt' column.")
-
-            # Convert each row to a dictionary and append to self.data
-            # This will convert the DataFrame to a list of dictionaries
-            # where each dictionary corresponds to a row in the DataFrame.
-            # This is the standardized format we want for self.data
-            for _, row in jsonl_data.iterrows():
-                self.data.append(row.to_dict())
-        else:
-            raise NotImplementedError(
-                "Only JSONL format is supported for CustomDataset."
-            )
-
-        random.seed(self.random_seed)
-        random.shuffle(self.data)
-
-    def sample(
-        self,
-        tokenizer: PreTrainedTokenizerBase,
-        num_requests: int,
-        lora_path: Optional[str] = None,
-        max_loras: Optional[int] = None,
-        output_len: Optional[int] = None,
-        enable_multimodal_chat: bool = False,
-        skip_chat_template: bool = False,
-        **kwargs,
-    ) -> list:
-        sampled_requests = []
-        for item in self.data:
-            if len(sampled_requests) >= num_requests:
-                break
-            prompt = item["prompt"]
-
-            # apply template
-            if not skip_chat_template:
-                prompt = tokenizer.apply_chat_template(
-                    [{"role": "user", "content": prompt}],
-                    add_generation_prompt=True,
-                    tokenize=False,
-                )
-
-            prompt_len = len(tokenizer(prompt).input_ids)
-            sampled_requests.append(
-                SampleRequest(
-                    prompt=prompt,
-                    prompt_len=prompt_len,
-                    expected_output_len=output_len,
-                )
-            )
-        self.maybe_oversample_requests(sampled_requests, num_requests)
-
-        return sampled_requests
-
-
 # -----------------------------------------------------------------------------
 # Sonnet Dataset Implementation
 # -----------------------------------------------------------------------------
@@ -865,15 +776,7 @@ class InstructCoderDataset(HuggingFaceDataset):
         for item in self.data:
             if len(sampled_requests) >= num_requests:
                 break
-            prompt = f"{item['input']}\n\n{item['instruction']} Just output \
-            the code, do not include any explanation."
-
-            # apply template
-            prompt = tokenizer.apply_chat_template(
-                [{"role": "user", "content": prompt}],
-                add_generation_prompt=True,
-                tokenize=False,
-            )
+            prompt = f"{item['instruction']}:\n{item['input']}"
             prompt_len = len(tokenizer(prompt).input_ids)
             sampled_requests.append(
                 SampleRequest(

benchmarks/benchmark_latency.py

@@ -1,5 +1,4 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 """Benchmark the latency of processing a single batch of requests."""
 
 import argparse
@@ -7,12 +6,13 @@ import dataclasses
 import json
 import os
 import time
+from pathlib import Path
 from typing import Any, Optional
 
 import numpy as np
+import torch
 from tqdm import tqdm
 
-import vllm.envs as envs
 from benchmark_utils import convert_to_pytorch_benchmark_format, write_to_json
 from vllm import LLM, SamplingParams
 from vllm.engine.arg_utils import EngineArgs
@@ -80,9 +80,17 @@ def main(args: argparse.Namespace):
 
     def run_to_completion(profile_dir: Optional[str] = None):
         if profile_dir:
-            llm.start_profile()
-            llm_generate()
-            llm.stop_profile()
+            with torch.profiler.profile(
+                activities=[
+                    torch.profiler.ProfilerActivity.CPU,
+                    torch.profiler.ProfilerActivity.CUDA,
+                ],
+                on_trace_ready=torch.profiler.tensorboard_trace_handler(
+                    str(profile_dir)
+                ),
+            ) as p:
+                llm_generate()
+            print(p.key_averages().table(sort_by="self_cuda_time_total"))
         else:
             start_time = time.perf_counter()
             llm_generate()
@@ -95,7 +103,11 @@ def main(args: argparse.Namespace):
         run_to_completion(profile_dir=None)
 
     if args.profile:
-        profile_dir = envs.VLLM_TORCH_PROFILER_DIR
+        profile_dir = args.profile_result_dir
+        if not profile_dir:
+            profile_dir = (
+                Path(".") / "vllm_benchmark_result" / f"latency_result_{time.time()}"
+            )
         print(f"Profiling (results will be saved to '{profile_dir}')...")
         run_to_completion(profile_dir=profile_dir)
         return
@@ -123,7 +135,7 @@ def main(args: argparse.Namespace):
         save_to_pytorch_benchmark_format(args, results)
 
 
-def create_argument_parser():
+if __name__ == "__main__":
     parser = FlexibleArgumentParser(
         description="Benchmark the latency of processing a single batch of "
         "requests till completion."
@@ -152,6 +164,15 @@ def create_argument_parser():
         action="store_true",
         help="profile the generation process of a single batch",
     )
+    parser.add_argument(
+        "--profile-result-dir",
+        type=str,
+        default=None,
+        help=(
+            "path to save the pytorch profiler output. Can be visualized "
+            "with ui.perfetto.dev or Tensorboard."
+        ),
+    )
     parser.add_argument(
         "--output-json",
         type=str,
@@ -171,16 +192,5 @@ def create_argument_parser():
     # V1 enables prefix caching by default which skews the latency
     # numbers. We need to disable prefix caching by default.
     parser.set_defaults(enable_prefix_caching=False)
-
-    return parser
-
-
-if __name__ == "__main__":
-    parser = create_argument_parser()
     args = parser.parse_args()
-    if args.profile and not envs.VLLM_TORCH_PROFILER_DIR:
-        raise OSError(
-            "The environment variable 'VLLM_TORCH_PROFILER_DIR' is not set. "
-            "Please set it to a valid path to use torch profiler."
-        )
     main(args)

benchmarks/benchmark_long_document_qa_throughput.py

@@ -1,5 +1,4 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 """
 Offline benchmark to test the long document QA throughput.
 
@@ -142,7 +141,7 @@ def main(args):
     )
 
 
-def create_argument_parser():
+if __name__ == "__main__":
     parser = FlexibleArgumentParser(
         description="Benchmark the performance with or "
         "without automatic prefix caching."
@@ -192,11 +191,5 @@ def create_argument_parser():
     )
 
     parser = EngineArgs.add_cli_args(parser)
-
-    return parser
-
-
-if __name__ == "__main__":
-    parser = create_argument_parser()
     args = parser.parse_args()
     main(args)

benchmarks/benchmark_prefix_caching.py

@@ -1,5 +1,4 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 """
 Benchmark the efficiency of prefix caching.
 
@@ -218,7 +217,7 @@ def main(args):
     )
 
 
-def create_argument_parser():
+if __name__ == "__main__":
     parser = FlexibleArgumentParser(
         description="Benchmark the performance with or without "
         "automatic prefix caching."
@@ -268,11 +267,5 @@ def create_argument_parser():
     )
 
     parser = EngineArgs.add_cli_args(parser)
-
-    return parser
-
-
-if __name__ == "__main__":
-    parser = create_argument_parser()
     args = parser.parse_args()
     main(args)

benchmarks/benchmark_prioritization.py

@@ -1,5 +1,4 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 """Benchmark offline prioritization."""
 
 import argparse
@@ -161,7 +160,7 @@ def main(args: argparse.Namespace):
             json.dump(results, f, indent=4)
 
 
-def create_argument_parser():
+if __name__ == "__main__":
     parser = FlexibleArgumentParser(description="Benchmark the throughput.")
     parser.add_argument(
         "--backend", type=str, choices=["vllm", "hf", "mii"], default="vllm"
@@ -204,12 +203,6 @@ def create_argument_parser():
     )
 
     parser = EngineArgs.add_cli_args(parser)
-
-    return parser
-
-
-if __name__ == "__main__":
-    parser = create_argument_parser()
     args = parser.parse_args()
     if args.tokenizer is None:
         args.tokenizer = args.model

benchmarks/benchmark_serving.py

@@ -1,5 +1,4 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 r"""Benchmark online serving throughput.
 
 On the server side, run one of the following commands:
@@ -61,7 +60,6 @@ from benchmark_dataset import (
     ASRDataset,
     BurstGPTDataset,
     ConversationDataset,
-    CustomDataset,
     HuggingFaceDataset,
     InstructCoderDataset,
     MTBenchDataset,
@@ -629,16 +627,7 @@ def main(args: argparse.Namespace):
             "'--dataset-path' if required."
         )
 
-    if args.dataset_name == "custom":
-        dataset = CustomDataset(dataset_path=args.dataset_path)
-        input_requests = dataset.sample(
-            num_requests=args.num_prompts,
-            tokenizer=tokenizer,
-            output_len=args.custom_output_len,
-            skip_chat_template=args.custom_skip_chat_template,
-        )
-
-    elif args.dataset_name == "sonnet":
+    if args.dataset_name == "sonnet":
         dataset = SonnetDataset(dataset_path=args.dataset_path)
         # For the "sonnet" dataset, formatting depends on the backend.
         if args.backend == "openai-chat":
@@ -773,10 +762,6 @@ def main(args: argparse.Namespace):
     if "temperature" not in sampling_params:
         sampling_params["temperature"] = 0.0  # Default to greedy decoding.
 
-    if args.backend == "llama.cpp":
-        # Disable prompt caching in llama.cpp backend
-        sampling_params["cache_prompt"] = False
-
     # Avoid GC processing "static" data - reduce pause times.
     gc.collect()
     gc.freeze()
@@ -849,8 +834,6 @@ def main(args: argparse.Namespace):
             ]:
                 if field in result_json:
                     del result_json[field]
-                if field in benchmark_result:
-                    del benchmark_result[field]
 
         # Save to file
         base_model_id = model_id.split("/")[-1]
@@ -863,7 +846,6 @@ def main(args: argparse.Namespace):
         if args.result_filename:
             file_name = args.result_filename
         if args.result_dir:
-            os.makedirs(args.result_dir, exist_ok=True)
             file_name = os.path.join(args.result_dir, file_name)
         with open(
             file_name, mode="a+" if args.append_result else "w", encoding="utf-8"
@@ -875,7 +857,7 @@ def main(args: argparse.Namespace):
         save_to_pytorch_benchmark_format(args, result_json, file_name)
 
 
-def create_argument_parser():
+if __name__ == "__main__":
     parser = FlexibleArgumentParser(
         description="Benchmark the online serving throughput."
     )
@@ -904,7 +886,7 @@ def create_argument_parser():
         "--dataset-name",
         type=str,
         default="sharegpt",
-        choices=["sharegpt", "burstgpt", "sonnet", "random", "hf", "custom"],
+        choices=["sharegpt", "burstgpt", "sonnet", "random", "hf"],
         help="Name of the dataset to benchmark on.",
     )
     parser.add_argument(
@@ -1074,19 +1056,6 @@ def create_argument_parser():
     )
 
     # group for dataset specific arguments
-    custom_group = parser.add_argument_group("custom dataset options")
-    custom_group.add_argument(
-        "--custom-output-len",
-        type=int,
-        default=256,
-        help="Number of output tokens per request, used only for custom dataset.",
-    )
-    custom_group.add_argument(
-        "--custom-skip-chat-template",
-        action="store_true",
-        help="Skip applying chat template to prompt, used only for custom dataset.",
-    )
-
     sonnet_group = parser.add_argument_group("sonnet dataset options")
     sonnet_group.add_argument(
         "--sonnet-input-len",
@@ -1225,10 +1194,6 @@ def create_argument_parser():
         "script chooses a LoRA module at random.",
     )
 
-    return parser
-
-
-if __name__ == "__main__":
-    parser = create_argument_parser()
     args = parser.parse_args()
+
     main(args)

benchmarks/benchmark_serving_structured_output.py

@@ -1,5 +1,4 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 r"""Benchmark online serving throughput with structured outputs.
 
 On the server side, run one of the following commands:
@@ -12,6 +11,7 @@ On the client side, run:
         --model <your_model> \
         --dataset json \
         --structured-output-ratio 1.0 \
+        --structured-output-backend auto \
         --request-rate 10 \
         --num-prompts 1000
 
@@ -850,7 +850,7 @@ def main(args: argparse.Namespace):
             json.dump(results, outfile, indent=4)
 
 
-def create_argument_parser():
+if __name__ == "__main__":
     parser = FlexibleArgumentParser(
         description="Benchmark the online serving throughput."
     )
@@ -1034,10 +1034,5 @@ def create_argument_parser():
         help="Ratio of Structured Outputs requests",
     )
 
-    return parser
-
-
-if __name__ == "__main__":
-    parser = create_argument_parser()
     args = parser.parse_args()
     main(args)

benchmarks/benchmark_throughput.py

@@ -1,5 +1,4 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 """Benchmark offline inference throughput."""
 
 import argparse
@@ -97,7 +96,7 @@ def run_vllm(
         assert lora_requests is None, "BeamSearch API does not support LoRA"
         prompts = [request.prompt for request in requests]
         # output_len should be the same for all requests.
-        output_len = requests[0].expected_output_len
+        output_len = requests[0][2]
         for request in requests:
             assert request.expected_output_len == output_len
         start = time.perf_counter()
@@ -595,7 +594,7 @@ def validate_args(args):
         )
 
 
-def create_argument_parser():
+if __name__ == "__main__":
     parser = FlexibleArgumentParser(description="Benchmark the throughput.")
     parser.add_argument(
         "--backend",
@@ -717,12 +716,6 @@ def create_argument_parser():
     )
 
     parser = AsyncEngineArgs.add_cli_args(parser)
-
-    return parser
-
-
-if __name__ == "__main__":
-    parser = create_argument_parser()
     args = parser.parse_args()
     if args.tokenizer is None:
         args.tokenizer = args.model

benchmarks/benchmark_utils.py

@@ -1,5 +1,4 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 import argparse
 import json
@@ -66,9 +65,4 @@ class InfEncoder(json.JSONEncoder):
 
 def write_to_json(filename: str, records: list) -> None:
     with open(filename, "w") as f:
-        json.dump(
-            records,
-            f,
-            cls=InfEncoder,
-            default=lambda o: f"<{type(o).__name__} object is not JSON serializable>",
-        )
+        json.dump(records, f, cls=InfEncoder)

benchmarks/cutlass_benchmarks/sparse_benchmarks.py

@@ -1,5 +1,4 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 import argparse
 import copy

benchmarks/cutlass_benchmarks/utils.py

@@ -1,5 +1,4 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 # Cutlass bench utils
 from collections.abc import Iterable

benchmarks/cutlass_benchmarks/w8a8_benchmarks.py

@@ -1,5 +1,4 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 import argparse
 import copy

benchmarks/cutlass_benchmarks/weight_shapes.py

@@ -1,5 +1,4 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 # Weight Shapes are in the format
 # ([K, N], TP_SPLIT_DIM)

benchmarks/disagg_benchmarks/disagg_prefill_proxy_server.py

@@ -1,5 +1,4 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 import os
 

benchmarks/disagg_benchmarks/round_robin_proxy.py

@@ -1,5 +1,4 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 import asyncio
 import itertools

benchmarks/disagg_benchmarks/visualize_benchmark_results.py

@@ -1,5 +1,4 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 import json
 

benchmarks/fused_kernels/layernorm_rms_benchmarks.py

@@ -1,5 +1,4 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 import pickle as pkl
 import time

benchmarks/kernels/bench_fp8_gemm.py

@@ -1,158 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-import argparse
-import copy
-import itertools
-
-import torch
-from weight_shapes import WEIGHT_SHAPES
-
-from vllm._custom_ops import cutlass_scaled_mm as vllm_scaled_mm
-from vllm._custom_ops import scaled_fp8_quant as vllm_scaled_fp8_quant
-from vllm.triton_utils import triton
-
-PROVIDER_CFGS = {
-    "torch-bf16": dict(enabled=True),
-    "fp8-tensor-w-token-a": dict(
-        w="tensor", a="token", no_a_quant=False, enabled=False
-    ),
-    "fp8-tensor-w-tensor-a": dict(
-        w="tensor", a="tensor", no_a_quant=False, enabled=True
-    ),
-    "fp8-channel-w-token-a": dict(
-        w="channel", a="token", no_a_quant=False, enabled=True
-    ),
-    "fp8-channel-w-tensor-a": dict(
-        w="channel", a="tensor", no_a_quant=False, enabled=False
-    ),
-    "fp8-tensor-w-token-a-noquant": dict(
-        w="tensor", a="token", no_a_quant=True, enabled=False
-    ),
-    "fp8-tensor-w-tensor-a-noquant": dict(
-        w="tensor", a="tensor", no_a_quant=True, enabled=True
-    ),
-    "fp8-channel-w-token-a-noquant": dict(
-        w="channel", a="token", no_a_quant=True, enabled=True
-    ),
-    "fp8-channel-w-tensor-a-noquant": dict(
-        w="channel", a="tensor", no_a_quant=True, enabled=False
-    ),
-}
-
-_enabled = [k for k, v in PROVIDER_CFGS.items() if v["enabled"]]
-
-
-def _quant_weight_fp8(b: torch.Tensor, w_type: str, device: str):
-    if w_type == "tensor":
-        scale_b = torch.ones(1, device=device, dtype=torch.float32)
-        b_fp8, scale_b_fp8 = vllm_scaled_fp8_quant(b, scale_b)
-    else:
-        b_fp8, scale_b_fp8 = vllm_scaled_fp8_quant(b, use_per_token_if_dynamic=True)
-    return b_fp8.t(), scale_b_fp8
-
-
-def build_fp8_runner(cfg, a, b, dtype, device):
-    b_fp8, scale_b_fp8 = _quant_weight_fp8(b, cfg["w"], device)
-
-    scale_a_const = (
-        torch.ones(1, device=device, dtype=torch.float32)
-        if cfg["a"] == "tensor"
-        else None
-    )
-
-    if cfg["no_a_quant"]:
-        if cfg["a"] == "tensor":
-            a_fp8, scale_a_fp8 = vllm_scaled_fp8_quant(a, scale_a_const)
-        else:
-            a_fp8, scale_a_fp8 = vllm_scaled_fp8_quant(a, use_per_token_if_dynamic=True)
-
-        def run():
-            return vllm_scaled_mm(a_fp8, b_fp8, scale_a_fp8, scale_b_fp8, dtype)
-
-        return run
-
-    if cfg["a"] == "tensor":
-
-        def run():
-            a_fp8, scale_a_fp8 = vllm_scaled_fp8_quant(a, scale_a_const)
-            return vllm_scaled_mm(a_fp8, b_fp8, scale_a_fp8, scale_b_fp8, dtype)
-
-    else:
-
-        def run():
-            a_fp8, scale_a_fp8 = vllm_scaled_fp8_quant(a, use_per_token_if_dynamic=True)
-            return vllm_scaled_mm(a_fp8, b_fp8, scale_a_fp8, scale_b_fp8, dtype)
-
-    return run
-
-
-@triton.testing.perf_report(
-    triton.testing.Benchmark(
-        x_names=["batch_size"],
-        x_vals=[1, 16, 64, 128, 256, 512, 1024, 2048, 4096, 8192, 16384],
-        x_log=False,
-        line_arg="provider",
-        line_vals=_enabled,
-        line_names=_enabled,
-        ylabel="TFLOP/s (larger is better)",
-        plot_name="BF16 vs FP8 GEMMs",
-        args={},
-    )
-)
-def benchmark(batch_size, provider, N, K):
-    M = batch_size
-    device = "cuda"
-    dtype = torch.bfloat16
-
-    a = torch.randn((M, K), device=device, dtype=dtype)
-    b = torch.randn((N, K), device=device, dtype=dtype)
-
-    quantiles = [0.5, 0.2, 0.8]
-
-    if provider == "torch-bf16":
-        ms, min_ms, max_ms = triton.testing.do_bench_cudagraph(
-            lambda: torch.nn.functional.linear(a, b), quantiles=quantiles
-        )
-    else:
-        cfg = PROVIDER_CFGS[provider]
-        run_quant = build_fp8_runner(cfg, a, b, dtype, device)
-        ms, min_ms, max_ms = triton.testing.do_bench_cudagraph(
-            lambda: run_quant(), quantiles=quantiles
-        )
-
-    to_tflops = lambda t_ms: (2 * M * N * K) * 1e-12 / (t_ms * 1e-3)
-    return to_tflops(ms), to_tflops(max_ms), to_tflops(min_ms)
-
-
-def prepare_shapes(args):
-    out = []
-    for model, tp_size in itertools.product(args.models, args.tp_sizes):
-        for KN, tp_dim in copy.deepcopy(WEIGHT_SHAPES[model]):
-            KN[tp_dim] //= tp_size
-            KN.append(model)
-            out.append(KN)
-    return out
-
-
-if __name__ == "__main__":
-    parser = argparse.ArgumentParser()
-    parser.add_argument(
-        "--models",
-        nargs="+",
-        type=str,
-        default=["meta-llama/Llama-3.1-8B-Instruct"],
-        choices=list(WEIGHT_SHAPES.keys()),
-    )
-    parser.add_argument("--tp-sizes", nargs="+", type=int, default=[1])
-    args = parser.parse_args()
-
-    for K, N, model in prepare_shapes(args):
-        print(f"{model}, N={N} K={K}, BF16 vs FP8 GEMMs TFLOP/s:")
-        benchmark.run(
-            print_data=True,
-            show_plots=True,
-            save_path=f"bench_fp8_res_n{N}_k{K}",
-            N=N,
-            K=K,
-        )
-
-    print("Benchmark finished!")

benchmarks/kernels/bench_int8_gemm.py

@@ -1,169 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-import argparse
-import copy
-import itertools
-
-import torch
-from weight_shapes import WEIGHT_SHAPES
-
-from vllm._custom_ops import cutlass_scaled_mm as vllm_scaled_mm
-from vllm._custom_ops import scaled_int8_quant as vllm_scaled_int8_quant
-from vllm.triton_utils import triton
-
-PROVIDER_CFGS = {
-    "torch-bf16": dict(enabled=True),
-    "int8-tensor-w-token-a": dict(
-        w="tensor", a="token", no_a_quant=False, enabled=False
-    ),
-    "int8-tensor-w-tensor-a": dict(
-        w="tensor", a="tensor", no_a_quant=False, enabled=True
-    ),
-    "int8-channel-w-token-a": dict(
-        w="channel", a="token", no_a_quant=False, enabled=True
-    ),
-    "int8-channel-w-tensor-a": dict(
-        w="channel", a="tensor", no_a_quant=False, enabled=False
-    ),
-    "int8-tensor-w-token-a-noquant": dict(
-        w="tensor", a="token", no_a_quant=True, enabled=False
-    ),
-    "int8-tensor-w-tensor-a-noquant": dict(
-        w="tensor", a="tensor", no_a_quant=True, enabled=True
-    ),
-    "int8-channel-w-token-a-noquant": dict(
-        w="channel", a="token", no_a_quant=True, enabled=True
-    ),
-    "int8-channel-w-tensor-a-noquant": dict(
-        w="channel", a="tensor", no_a_quant=True, enabled=False
-    ),
-}
-
-
-def _quant_weight(b, w_type, device):
-    if w_type == "tensor":
-        scale_b = torch.ones(1, device=device, dtype=torch.float32)
-        b_int8, scale_b_int8, _ = vllm_scaled_int8_quant(b, scale_b)
-        assert scale_b_int8.numel() == 1
-    else:  # channel
-        b_int8, scale_b_int8, _ = vllm_scaled_int8_quant(b)
-        assert scale_b_int8.numel() == b.shape[0]
-    return b_int8.t(), scale_b_int8
-
-
-def build_int8_runner(cfg, a, b, dtype, device):
-    # quant before running the kernel
-    b_int8, scale_b_int8 = _quant_weight(b, cfg["w"], device)
-
-    scale_a_const = None
-    if cfg["a"] == "tensor":
-        scale_a_const = torch.ones(1, device=device, dtype=torch.float32)
-
-    # no quant, create activation ahead
-    if cfg["no_a_quant"]:
-        if cfg["a"] == "tensor":
-            a_int8, scale_a_int8, _ = vllm_scaled_int8_quant(a, scale_a_const)
-        else:  # token
-            a_int8, scale_a_int8, _ = vllm_scaled_int8_quant(a)
-
-        def run_quant():
-            return vllm_scaled_mm(a_int8, b_int8, scale_a_int8, scale_b_int8, dtype)
-
-        return run_quant
-
-    # dynamic quant, create activation inside
-    if cfg["a"] == "tensor":
-
-        def run_quant():
-            a_int8, scale_a_int8, _ = vllm_scaled_int8_quant(a, scale_a_const)
-            return vllm_scaled_mm(a_int8, b_int8, scale_a_int8, scale_b_int8, dtype)
-
-    else:  # token
-
-        def run_quant():
-            a_int8, scale_a_int8, _ = vllm_scaled_int8_quant(a)
-            return vllm_scaled_mm(a_int8, b_int8, scale_a_int8, scale_b_int8, dtype)
-
-    return run_quant
-
-
-_enabled = [k for k, v in PROVIDER_CFGS.items() if v.get("enabled")]
-
-
-@triton.testing.perf_report(
-    triton.testing.Benchmark(
-        x_names=["batch_size"],
-        x_vals=[1, 16, 64, 128, 256, 512, 1024, 2048, 4096, 8192, 16384],
-        x_log=False,
-        line_arg="provider",
-        line_vals=_enabled,
-        line_names=[k for k in _enabled],
-        ylabel="TFLOP/s (larger is better)",
-        plot_name="BF16 vs INT8 GEMMs",
-        args={},
-    )
-)
-def benchmark(batch_size, provider, N, K):
-    M = batch_size
-    device = "cuda"
-    dtype = torch.bfloat16
-    a = torch.randn((M, K), device=device, dtype=dtype)
-    b = torch.randn((N, K), device=device, dtype=dtype)
-
-    quantiles = [0.5, 0.2, 0.8]
-
-    if provider == "torch-bf16":
-        ms, min_ms, max_ms = triton.testing.do_bench_cudagraph(
-            lambda: torch.nn.functional.linear(a, b), quantiles=quantiles
-        )
-    else:
-        cfg = PROVIDER_CFGS[provider]
-        run_quant = build_int8_runner(cfg, a, b, dtype, device)
-        ms, min_ms, max_ms = triton.testing.do_bench_cudagraph(
-            lambda: run_quant(), quantiles=quantiles
-        )
-
-    to_tflops = lambda t_ms: (2 * M * N * K) * 1e-12 / (t_ms * 1e-3)
-    return to_tflops(ms), to_tflops(max_ms), to_tflops(min_ms)
-
-
-def prepare_shapes(args):
-    KN_model_names = []
-    for model, tp_size in itertools.product(args.models, args.tp_sizes):
-        for KN, tp_dim in copy.deepcopy(WEIGHT_SHAPES[model]):
-            KN[tp_dim] //= tp_size
-            KN.append(model)
-            KN_model_names.append(KN)
-    return KN_model_names
-
-
-if __name__ == "__main__":
-    parser = argparse.ArgumentParser()
-    parser.add_argument(
-        "--models",
-        nargs="+",
-        type=str,
-        default=["meta-llama/Llama-3.1-8B-Instruct"],
-        choices=list(WEIGHT_SHAPES.keys()),
-        help="List of models to benchmark",
-    )
-    parser.add_argument(
-        "--tp-sizes",
-        nargs="+",
-        type=int,
-        default=[1],
-        help="List of tensor parallel sizes",
-    )
-    args = parser.parse_args()
-
-    for K, N, model in prepare_shapes(args):
-        print(f"{model}, N={N} K={K}, BF16 vs INT8 GEMMs TFLOP/s:")
-        benchmark.run(
-            print_data=True,
-            show_plots=True,
-            save_path=f"bench_int8_res_n{N}_k{K}",
-            N=N,
-            K=K,
-        )
-
-    print("Benchmark finished!")

benchmarks/kernels/benchmark_aqlm.py

@@ -1,5 +1,4 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 import os
 import sys

benchmarks/kernels/benchmark_bitblas.py

@@ -1,5 +1,4 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 # Copyright (c) Microsoft Corporation.
 # Licensed under the MIT License.
 

benchmarks/kernels/benchmark_cutlass_fp4_moe.py

@@ -1,5 +1,4 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 """
 Benchmark the performance of the cutlass_moe_fp4 kernel vs the triton_moe
 kernel. The cutlass_moe_fp4 kernel takes in fp4 quantized weights and 16-bit
@@ -91,7 +90,7 @@ def bench_run(
 
     score = torch.randn((m, num_experts), device=device, dtype=dtype)
 
-    topk_weights, topk_ids, _ = fused_topk(a, score, topk, renormalize=False)
+    topk_weights, topk_ids = fused_topk(a, score, topk, renormalize=False)
 
     quant_blocksize = 16
     w1_blockscale = torch.empty(

benchmarks/kernels/benchmark_grouped_gemm_cutlass.py

@@ -1,5 +1,4 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 import torch
 import torch.utils.benchmark as benchmark
@@ -7,8 +6,8 @@ from benchmark_shapes import WEIGHT_SHAPES_MOE
 
 from vllm import _custom_ops as ops
 from vllm.config import ParallelConfig, VllmConfig, set_current_vllm_config
-from vllm.model_executor.layers.fused_moe.cutlass_moe import cutlass_moe_fp8
 from vllm.model_executor.layers.fused_moe.fused_moe import (
+    cutlass_moe_fp8,
     fused_experts,
     fused_topk,
 )
@@ -70,9 +69,18 @@ def bench_run(
     w1_scale = torch.empty((num_experts, 1, 1), device="cuda", dtype=torch.float32)
     w2_scale = torch.empty((num_experts, 1, 1), device="cuda", dtype=torch.float32)
 
+    ab_strides1 = torch.full((num_experts,), k, device="cuda", dtype=torch.int64)
+    c_strides1 = torch.full((num_experts,), 2 * n, device="cuda", dtype=torch.int64)
+    ab_strides2 = torch.full((num_experts,), n, device="cuda", dtype=torch.int64)
+    c_strides2 = torch.full((num_experts,), k, device="cuda", dtype=torch.int64)
+
     for expert in range(num_experts):
         w1_q[expert], w1_scale[expert] = ops.scaled_fp8_quant(w1[expert])
         w2_q[expert], w2_scale[expert] = ops.scaled_fp8_quant(w2[expert])
+    w1_q_notransp = w1_q.clone()
+    w2_q_notransp = w2_q.clone()
+    w1_q = w1_q.transpose(1, 2)
+    w2_q = w2_q.transpose(1, 2)
 
     score = torch.randn((m, num_experts), device="cuda", dtype=dtype)
 
@@ -113,6 +121,10 @@ def bench_run(
         w2_scale: torch.Tensor,
         topk_weights: torch.Tensor,
         topk_ids: torch.Tensor,
+        ab_strides1: torch.Tensor,
+        c_strides1: torch.Tensor,
+        ab_strides2: torch.Tensor,
+        c_strides2: torch.Tensor,
         num_repeats: int,
     ):
         for _ in range(num_repeats):
@@ -120,10 +132,14 @@ def bench_run(
                 a,
                 w1,
                 w2,
-                topk_weights,
-                topk_ids,
                 w1_scale,
                 w2_scale,
+                topk_weights,
+                topk_ids,
+                ab_strides1,
+                c_strides1,
+                ab_strides2,
+                c_strides2,
                 a1_scale=a_scale,
             )
 
@@ -136,6 +152,10 @@ def bench_run(
         w2_scale: torch.Tensor,
         topk_weights: torch.Tensor,
         topk_ids: torch.Tensor,
+        ab_strides1: torch.Tensor,
+        c_strides1: torch.Tensor,
+        ab_strides2: torch.Tensor,
+        c_strides2: torch.Tensor,
     ):
         with set_current_vllm_config(
             VllmConfig(parallel_config=ParallelConfig(pipeline_parallel_size=1))
@@ -144,10 +164,14 @@ def bench_run(
                 a,
                 w1_q,
                 w2_q,
-                topk_weights,
-                topk_ids,
                 w1_scale,
                 w2_scale,
+                topk_weights,
+                topk_ids,
+                ab_strides1,
+                c_strides1,
+                ab_strides2,
+                c_strides2,
                 a1_scale=a_scale,
             )
 
@@ -193,6 +217,10 @@ def bench_run(
             w2_scale,
             topk_weights,
             topk_ids,
+            ab_strides1,
+            c_strides1,
+            ab_strides2,
+            c_strides2,
         )
     torch.cuda.synchronize()
 
@@ -201,8 +229,8 @@ def bench_run(
     with torch.cuda.graph(triton_graph, stream=triton_stream):
         run_triton_from_graph(
             a,
-            w1_q,
-            w2_q,
+            w1_q_notransp,
+            w2_q_notransp,
             topk_weights,
             topk_ids,
             w1_scale,
@@ -221,12 +249,18 @@ def bench_run(
         "w2": w2,
         "score": score,
         "topk": topk,
+        "w1_q_notransp": w1_q_notransp,
+        "w2_q_notransp": w2_q_notransp,
         # Cutlass params
         "a_scale": a_scale,
         "w1_q": w1_q,
         "w2_q": w2_q,
         "w1_scale": w1_scale,
         "w2_scale": w2_scale,
+        "ab_strides1": ab_strides1,
+        "c_strides1": c_strides1,
+        "ab_strides2": ab_strides2,
+        "c_strides2": c_strides2,
         # cuda graph params
         "cutlass_graph": cutlass_graph,
         "triton_graph": triton_graph,
@@ -244,8 +278,8 @@ def bench_run(
     # Warmup
     run_triton_moe(
         a,
-        w1_q,
-        w2_q,
+        w1_q_notransp,
+        w2_q_notransp,
         topk_weights,
         topk_ids,
         w1_scale,
@@ -256,7 +290,7 @@ def bench_run(
 
     results.append(
         benchmark.Timer(
-            stmt="run_triton_moe(a, w1_q, w2_q, topk_weights, topk_ids, w1_scale, w2_scale, a_scale, num_runs)",  # noqa: E501
+            stmt="run_triton_moe(a, w1_q_notransp, w2_q_notransp, topk_weights, topk_ids, w1_scale, w2_scale, a_scale, num_runs)",  # noqa: E501
             globals=globals,
             label=label,
             sub_label=sub_label,
@@ -287,12 +321,16 @@ def bench_run(
         w2_scale,
         topk_weights,
         topk_ids,
+        ab_strides1,
+        c_strides1,
+        ab_strides2,
+        c_strides2,
         num_warmup,
     )
 
     results.append(
         benchmark.Timer(
-            stmt="run_cutlass_moe(a, a_scale, w1_q, w2_q, w1_scale, w2_scale, topk_weights, topk_ids, num_runs)",  # noqa: E501
+            stmt="run_cutlass_moe(a, a_scale, w1_q, w2_q, w1_scale, w2_scale, topk_weights, topk_ids, ab_strides1, c_strides1, ab_strides2, c_strides2, num_runs)",  # noqa: E501
             globals=globals,
             label=label,
             sub_label=sub_label,

benchmarks/kernels/benchmark_layernorm.py

@@ -1,5 +1,4 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 import time
 

benchmarks/kernels/benchmark_lora.py

@@ -1,5 +1,4 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 import argparse
 import copy

benchmarks/kernels/benchmark_machete.py

@@ -1,5 +1,4 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 import argparse
 import copy

benchmarks/kernels/benchmark_marlin.py

@@ -1,5 +1,4 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 import torch
 import torch.utils.benchmark as benchmark
@@ -22,16 +21,8 @@ from vllm.model_executor.layers.quantization.utils.marlin_utils import (
     MARLIN_SUPPORTED_GROUP_SIZES,
     query_marlin_supported_quant_types,
 )
-from vllm.model_executor.layers.quantization.utils.marlin_utils_fp4 import (
-    FP4_MARLIN_SUPPORTED_GROUP_SIZES,
-    rand_marlin_weight_fp4_like,
-)
-from vllm.model_executor.layers.quantization.utils.marlin_utils_fp8 import (
-    marlin_quant_fp8_torch,
-)
 from vllm.model_executor.layers.quantization.utils.marlin_utils_test import (
     MarlinWorkspace,
-    awq_marlin_quantize,
     marlin_quantize,
 )
 from vllm.model_executor.layers.quantization.utils.marlin_utils_test_24 import (
@@ -43,7 +34,7 @@ from vllm.model_executor.layers.quantization.utils.quant_utils import (
     quantize_weights,
     sort_weights,
 )
-from vllm.scalar_type import ScalarType, scalar_types
+from vllm.scalar_type import ScalarType
 from vllm.utils import FlexibleArgumentParser
 
 DEFAULT_MODELS = ["meta-llama/Llama-2-7b-hf/TP1"]
@@ -65,144 +56,80 @@ def bench_run(
     size_n: int,
 ):
     label = "Quant Matmul"
+
     sub_label = "{}, act={} k_full={}, q={}, g={}, MKN=({}x{}x{})".format(
         model, act_order, is_k_full, str(quant_type), group_size, size_m, size_k, size_n
     )
+
     print(f"Testing: {sub_label}")
 
     a = torch.randn(size_m, size_k).to(torch.half).cuda()
     b = torch.rand(size_k, size_n).to(torch.half).cuda()
-    has_zp = quant_type in [scalar_types.uint4, scalar_types.uint8]
-    if act_order and (group_size == -1 or group_size == size_k or has_zp):
-        return
-    if size_k % group_size != 0:
-        return
 
-    marlin_24_supported = (
-        quant_type in GPTQ_MARLIN_24_SUPPORTED_QUANT_TYPES
-        and group_size in GPTQ_MARLIN_24_SUPPORTED_GROUP_SIZES
-    )
-    repack_supported = (
-        quant_type in GPTQ_MARLIN_24_SUPPORTED_QUANT_TYPES
-        and group_size in MARLIN_SUPPORTED_GROUP_SIZES
-    )
-    allspark_supported = (
-        quant_type in ALLSPARK_SUPPORTED_QUANT_TYPES
-        and group_size == -1
-        and not act_order
-        and is_k_full
-    )
-
-    def gen_marlin_params():
-        # Marlin quant
-        marlin_g_idx = marlin_sort_indices = marlin_zp = marlin_s2 = None
-        if quant_type == scalar_types.float4_e2m1f:
-            if group_size != 16 or act_order:
-                return
-            marlin_w_ref, marlin_q_w, marlin_s, marlin_s2 = rand_marlin_weight_fp4_like(
-                b.T, group_size
-            )
-        elif quant_type == scalar_types.float8_e4m3fn:
-            if group_size not in [-1, 128] or act_order:
-                return
-            marlin_w_ref, marlin_q_w, marlin_s = marlin_quant_fp8_torch(b.T, group_size)
-        elif group_size == 16:
-            return
-        elif has_zp:
-            marlin_w_ref, marlin_q_w, marlin_s, marlin_zp = awq_marlin_quantize(
-                b, quant_type, group_size
-            )
-        else:
-            marlin_w_ref, marlin_q_w, marlin_s, marlin_g_idx, marlin_sort_indices, _ = (
-                marlin_quantize(b, quant_type, group_size, act_order)
-            )
-        return (
-            marlin_w_ref,
-            marlin_q_w,
-            marlin_s,
-            marlin_s2,
-            marlin_zp,
-            marlin_g_idx,
-            marlin_sort_indices,
-        )
-
-    def gen_marlin_24_params():
-        marlin_24_w_ref = marlin_24_q_w_comp = marlin_24_meta = marlin_24_s = None
-        if marlin_24_supported:
-            (marlin_24_w_ref, marlin_24_q_w_comp, marlin_24_meta, marlin_24_s) = (
-                marlin_24_quantize(b, quant_type, group_size)
-            )
-        return (marlin_24_w_ref, marlin_24_q_w_comp, marlin_24_meta, marlin_24_s)
-
-    def gen_repack_params():
-        q_w_gptq = None
-        repack_sort_indices = None
-        if repack_supported:
-            (w_ref, q_w, s, g_idx, rand_perm) = gptq_quantize_weights(
-                b, quant_type, group_size, act_order
-            )
-            q_w_gptq = gptq_pack(q_w, quant_type.size_bits, size_k, size_n)
-
-            # For act_order, sort the "weights" and "g_idx"
-            # so that group ids are increasing
-            repack_sort_indices = torch.empty(0, dtype=torch.int, device=b.device)
-            if act_order:
-                (q_w, g_idx, repack_sort_indices) = sort_weights(q_w, g_idx)
-        return q_w_gptq, repack_sort_indices
-
-    def gen_allspark_params():
-        qw_reorder = s_reorder = zp_reorder = sm_count = sm_version = (
-            CUBLAS_M_THRESHOLD
-        ) = None
-        nonlocal allspark_supported
-        if allspark_supported:
-            properties = torch.cuda.get_device_properties(b.device.index)
-            sm_count = properties.multi_processor_count
-            sm_version = properties.major * 10 + properties.minor
-
-            supported_arch = sm_version >= 80 and sm_version < 90
-            allspark_supported = allspark_supported and supported_arch
-            if supported_arch:
-                w_ref, qw, s, zp = quantize_weights(b, quant_type, group_size, has_zp)
-                qw = qw.to(torch.uint8)
-
-                qw_reorder, s_reorder, zp_reorder = ops.allspark_repack_weight(
-                    qw, s, zp, has_zp
-                )
-                CUBLAS_M_THRESHOLD = ALLSPARK_AMPERE_M_CUBLAS_THRESHOLD
-        return (
-            qw_reorder,
-            s_reorder,
-            zp_reorder,
-            sm_count,
-            sm_version,
-            CUBLAS_M_THRESHOLD,
-        )
+    a_tmp = torch.zeros(size_m, size_k).to(torch.half).cuda()
 
+    # Marlin quant
     (
         marlin_w_ref,
         marlin_q_w,
         marlin_s,
-        marlin_s2,
-        marlin_zp,
         marlin_g_idx,
         marlin_sort_indices,
-    ) = gen_marlin_params()
-    marlin_24_w_ref, marlin_24_q_w_comp, marlin_24_meta, marlin_24_s = (
-        gen_marlin_24_params()
+        marlin_rand_perm,
+    ) = marlin_quantize(b, quant_type, group_size, act_order)
+
+    # Marlin_24 quant
+    (marlin_24_w_ref, marlin_24_q_w_comp, marlin_24_meta, marlin_24_s) = (
+        marlin_24_quantize(b, quant_type, group_size)
     )
-    q_w_gptq, repack_sort_indices = gen_repack_params()
-    qw_reorder, s_reorder, zp_reorder, sm_count, sm_version, CUBLAS_M_THRESHOLD = (
-        gen_allspark_params()
+
+    marlin_zp = torch.empty(0, dtype=torch.int, device=b.device)
+
+    # GPTQ quant
+    (w_ref, q_w, s, g_idx, rand_perm) = gptq_quantize_weights(
+        b, quant_type, group_size, act_order
     )
+    q_w_gptq = gptq_pack(q_w, quant_type.size_bits, size_k, size_n)
+
+    # For act_order, sort the "weights" and "g_idx"
+    # so that group ids are increasing
+    repack_sort_indices = torch.empty(0, dtype=torch.int, device=b.device)
+    if act_order:
+        (q_w, g_idx, repack_sort_indices) = sort_weights(q_w, g_idx)
 
     # Prepare
     marlin_workspace = MarlinWorkspace(
         size_n, GPTQ_MARLIN_MIN_THREAD_N, GPTQ_MARLIN_MAX_PARALLEL
     )
+
     marlin_24_workspace = MarlinWorkspace(
         size_n, GPTQ_MARLIN_24_MIN_THREAD_N, GPTQ_MARLIN_24_MAX_PARALLEL
     )
+    marlin_zp = torch.zeros_like(marlin_s, dtype=torch.int)
+
+    # AllSpark W8A16 quant
+    as_supported_case = (
+        quant_type in ALLSPARK_SUPPORTED_QUANT_TYPES
+        and group_size == -1
+        and not act_order
+        and is_k_full
+    )
+    if as_supported_case:
+        properties = torch.cuda.get_device_properties(b.device.index)
+        sm_count = properties.multi_processor_count
+        sm_version = properties.major * 10 + properties.minor
+
+        supported_arch = sm_version >= 80 and sm_version < 90
+        as_supported_case = as_supported_case and supported_arch
+        if supported_arch:
+            has_zp = False
+            w_ref, qw, s, zp = quantize_weights(b, quant_type, group_size, has_zp)
+            qw = qw.to(torch.uint8)
+
+            qw_reorder, s_reorder, zp_reorder = ops.allspark_repack_weight(
+                qw, s, zp, has_zp
+            )
+            CUBLAS_M_THRESHOLD = ALLSPARK_AMPERE_M_CUBLAS_THRESHOLD
 
     globals = {
         # Gen params
@@ -212,14 +139,15 @@ def bench_run(
         "size_n": size_n,
         "size_k": size_k,
         "a": a,
+        "a_tmp": a_tmp,
         # Marlin params
         "marlin_w_ref": marlin_w_ref,
         "marlin_q_w": marlin_q_w,
         "marlin_s": marlin_s,
-        "marlin_s2": marlin_s2,
         "marlin_zp": marlin_zp,
         "marlin_g_idx": marlin_g_idx,
         "marlin_sort_indices": marlin_sort_indices,
+        "marlin_rand_perm": marlin_rand_perm,
         "marlin_workspace": marlin_workspace,
         "is_k_full": is_k_full,
         # Marlin_24 params
@@ -232,12 +160,12 @@ def bench_run(
         "q_w_gptq": q_w_gptq,
         "repack_sort_indices": repack_sort_indices,
         # AllSpark W8A16 params
-        "qw_reorder": qw_reorder,
-        "s_reorder": s_reorder,
-        "zp_reorder": zp_reorder,
-        "sm_count": sm_count,
-        "sm_version": sm_version,
-        "CUBLAS_M_THRESHOLD": CUBLAS_M_THRESHOLD,
+        "qw_reorder": qw_reorder if as_supported_case else None,
+        "s_reorder": s_reorder if as_supported_case else None,
+        "zp_reorder": zp_reorder if as_supported_case else None,
+        "sm_count": sm_count if as_supported_case else None,
+        "sm_version": sm_version if as_supported_case else None,
+        "CUBLAS_M_THRESHOLD": CUBLAS_M_THRESHOLD if as_supported_case else None,
         # Kernels
         "gptq_marlin_gemm": ops.gptq_marlin_gemm,
         "gptq_marlin_24_gemm": ops.gptq_marlin_24_gemm,
@@ -248,7 +176,7 @@ def bench_run(
     min_run_time = 1
 
     # Warmup pytorch
-    for _ in range(5):
+    for i in range(5):
         torch.matmul(a, marlin_w_ref)
 
     results.append(
@@ -263,17 +191,17 @@ def bench_run(
 
     results.append(
         benchmark.Timer(
-            stmt="output = gptq_marlin_gemm(a, None, marlin_q_w, marlin_s, marlin_s2, marlin_zp, marlin_g_idx, marlin_sort_indices, marlin_workspace.scratch, quant_type, size_m, size_n, size_k, is_k_full, False, False, False)",  # noqa: E501
+            stmt="output = gptq_marlin_gemm(a, marlin_q_w, marlin_s, marlin_zp, marlin_g_idx, marlin_sort_indices, marlin_workspace.scratch, quant_type, size_m, size_n, size_k, is_k_full, False, False, False)",  # noqa: E501
             globals=globals,
             label=label,
             sub_label=sub_label,
-            description="gptq_marlin_gemm",
+            description="gptq_marlin_gemm_fp16",
         ).blocked_autorange(min_run_time=min_run_time)
     )
 
     results.append(
         benchmark.Timer(
-            stmt="output = gptq_marlin_gemm(a, None, marlin_q_w, marlin_s, marlin_s2, marlin_zp, marlin_g_idx, marlin_sort_indices, marlin_workspace.scratch, quant_type, size_m, size_n, size_k, is_k_full, False, True, False)",  # noqa: E501
+            stmt="output = gptq_marlin_gemm(a, marlin_q_w, marlin_s, marlin_zp, marlin_g_idx, marlin_sort_indices, marlin_workspace.scratch, quant_type, size_m, size_n, size_k, is_k_full, False, True, False)",  # noqa: E501
             globals=globals,
             label=label,
             sub_label=sub_label,
@@ -281,7 +209,10 @@ def bench_run(
         ).blocked_autorange(min_run_time=min_run_time)
     )
 
-    if marlin_24_supported:
+    if (
+        quant_type in GPTQ_MARLIN_24_SUPPORTED_QUANT_TYPES
+        and group_size in GPTQ_MARLIN_24_SUPPORTED_GROUP_SIZES
+    ):
         results.append(
             benchmark.Timer(
                 stmt="output = gptq_marlin_24_gemm(a, marlin_24_q_w_comp, marlin_24_meta, marlin_24_s, marlin_24_workspace.scratch, quant_type, size_m, size_n, size_k)",  # noqa: E501
@@ -292,18 +223,17 @@ def bench_run(
             ).blocked_autorange(min_run_time=min_run_time)
         )
 
-    if repack_supported:
-        results.append(
-            benchmark.Timer(
-                stmt="q_res = gptq_marlin_repack(q_w_gptq, repack_sort_indices, size_k, size_n, quant_type.size_bits)",  # noqa: E501
-                globals=globals,
-                label=label,
-                sub_label=sub_label,
-                description="gptq_marlin_repack",
-            ).blocked_autorange(min_run_time=min_run_time)
-        )
+    results.append(
+        benchmark.Timer(
+            stmt="q_res = gptq_marlin_repack(q_w_gptq, repack_sort_indices, size_k, size_n, quant_type.size_bits)",  # noqa: E501
+            globals=globals,
+            label=label,
+            sub_label=sub_label,
+            description="gptq_marlin_repack",
+        ).blocked_autorange(min_run_time=min_run_time)
+    )
 
-    if allspark_supported:
+    if as_supported_case:
         results.append(
             benchmark.Timer(
                 stmt="output = allspark_w8a16_gemm(a, qw_reorder, s_reorder, zp_reorder, size_n, group_size, sm_count, sm_version, CUBLAS_M_THRESHOLD, False, True)",  # noqa: E501
@@ -319,6 +249,7 @@ def main(args):
     print("Benchmarking models:")
     for i, model in enumerate(args.models):
         print(f"[{i}]  {model}")
+
     results: list[benchmark.Measurement] = []
 
     for model in args.models:
@@ -346,17 +277,14 @@ def main(args):
                     ):
                         continue
 
-                    for quant_type in query_marlin_supported_quant_types():
+                    for quant_type in query_marlin_supported_quant_types(False):
                         if (
                             len(args.limit_num_bits) > 0
                             and quant_type.size_bits not in args.limit_num_bits
                         ):
                             continue
 
-                        for group_size in (
-                            MARLIN_SUPPORTED_GROUP_SIZES
-                            + FP4_MARLIN_SUPPORTED_GROUP_SIZES
-                        ):
+                        for group_size in MARLIN_SUPPORTED_GROUP_SIZES:
                             if (
                                 len(args.limit_group_size) > 0
                                 and group_size not in args.limit_group_size

benchmarks/kernels/benchmark_moe.py

@@ -1,5 +1,4 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 import argparse
 import json
@@ -7,6 +6,7 @@ import time
 from contextlib import nullcontext
 from datetime import datetime
 from itertools import product
+from types import SimpleNamespace
 from typing import Any, TypedDict
 
 import ray
@@ -42,7 +42,7 @@ def benchmark_config(
     use_fp8_w8a8: bool,
     use_int8_w8a16: bool,
     num_iters: int = 100,
-    block_quant_shape: list[int] = None,
+    block_quant_shape: List[int] = None,
     use_deep_gemm: bool = False,
 ) -> float:
     init_dtype = torch.float16 if use_fp8_w8a8 else dtype
@@ -399,7 +399,7 @@ class BenchmarkWorker:
         dtype: torch.dtype,
         use_fp8_w8a8: bool,
         use_int8_w8a16: bool,
-        block_quant_shape: list[int] = None,
+        block_quant_shape: List[int] = None,
         use_deep_gemm: bool = False,
     ) -> tuple[dict[str, int], float]:
         current_platform.seed_everything(self.seed)
@@ -531,7 +531,7 @@ def save_configs(
     dtype: torch.dtype,
     use_fp8_w8a8: bool,
     use_int8_w8a16: bool,
-    block_quant_shape: list[int],
+    block_quant_shape: List[int],
 ) -> None:
     dtype_str = get_config_dtype_str(
         dtype, use_int8_w8a16=use_int8_w8a16, use_fp8_w8a8=use_fp8_w8a8
@@ -562,6 +562,7 @@ def main(args: argparse.Namespace):
     config = get_config(model=args.model, trust_remote_code=args.trust_remote_code)
     if args.model_prefix:
         config = getattr(config, args.model_prefix)
+    config = SimpleNamespace(**config)
 
     if config.architectures[0] == "DbrxForCausalLM":
         E = config.ffn_config.moe_num_experts
@@ -593,7 +594,11 @@ def main(args: argparse.Namespace):
         shard_intermediate_size = 2 * intermediate_size // args.tp_size
 
     hidden_size = config.hidden_size
-    dtype = torch.float16 if current_platform.is_rocm() else config.torch_dtype
+    dtype = (
+        torch.float16
+        if current_platform.is_rocm()
+        else getattr(torch, config.torch_dtype)
+    )
     use_fp8_w8a8 = args.dtype == "fp8_w8a8"
     use_int8_w8a16 = args.dtype == "int8_w8a16"
     block_quant_shape = get_weight_block_size_safety(config)

benchmarks/kernels/benchmark_moe_align_block_size.py

@@ -1,159 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-import argparse
-import itertools
-
-import torch
-
-from vllm import _custom_ops as ops
-from vllm.model_executor.layers.fused_moe.moe_align_block_size import (
-    moe_align_block_size_triton,
-)
-from vllm.triton_utils import triton
-
-
-def get_topk_ids(num_tokens: int, num_experts: int, topk: int) -> torch.Tensor:
-    return torch.stack(
-        [
-            torch.randperm(num_experts, dtype=torch.int32, device="cuda")[:topk]
-            for _ in range(num_tokens)
-        ]
-    )
-
-
-def check_correctness(num_tokens, num_experts=256, block_size=256, topk=8):
-    """
-    Verifies vllm vs. Triton
-    """
-    topk_ids = get_topk_ids(num_tokens, num_experts, topk)
-
-    # 1. malloc space for triton and vllm
-    # malloc enough space (max_num_tokens_padded) for the sorted ids
-    max_num_tokens_padded = topk_ids.numel() + num_experts * (block_size - 1)
-    sorted_ids_triton = torch.empty(
-        (max_num_tokens_padded,), dtype=torch.int32, device="cuda"
-    )
-    sorted_ids_triton.fill_(topk_ids.numel())  # fill with sentinel value
-    expert_ids_triton = torch.zeros(
-        (max_num_tokens_padded // block_size,), dtype=torch.int32, device="cuda"
-    )
-    num_tokens_post_pad_triton = torch.empty((1,), dtype=torch.int32, device="cuda")
-
-    sorted_ids_vllm = torch.empty_like(sorted_ids_triton)
-    sorted_ids_vllm.fill_(topk_ids.numel())
-    expert_ids_vllm = torch.zeros_like(expert_ids_triton)
-    num_tokens_post_pad_vllm = torch.empty_like(num_tokens_post_pad_triton)
-
-    # 2. run implementations
-    moe_align_block_size_triton(
-        topk_ids,
-        num_experts,
-        block_size,
-        sorted_ids_triton,
-        expert_ids_triton,
-        num_tokens_post_pad_triton,
-    )
-
-    ops.moe_align_block_size(
-        topk_ids,
-        num_experts,
-        block_size,
-        sorted_ids_vllm,
-        expert_ids_vllm,
-        num_tokens_post_pad_vllm,
-    )
-    print(f"✅ VLLM implementation works with {num_experts} experts!")
-
-    # 3. compare results
-    if torch.allclose(expert_ids_triton, expert_ids_vllm) and torch.allclose(
-        num_tokens_post_pad_triton, num_tokens_post_pad_vllm
-    ):
-        print("✅ Triton and VLLM implementations match.")
-    else:
-        print("❌ Triton and VLLM implementations DO NOT match.")
-        print("Triton expert_ids:", expert_ids_triton)
-        print("VLLM expert_ids:", expert_ids_vllm)
-        print("Triton num_tokens_post_pad:", num_tokens_post_pad_triton)
-        print("VLLM num_tokens_post_pad:", num_tokens_post_pad_vllm)
-
-
-# test configurations
-num_tokens_range = [1, 16, 256, 4096]
-num_experts_range = [16, 64, 224, 256, 280, 512]
-topk_range = [1, 2, 8]
-configs = list(itertools.product(num_tokens_range, num_experts_range, topk_range))
-
-
-@triton.testing.perf_report(
-    triton.testing.Benchmark(
-        x_names=["num_tokens", "num_experts", "topk"],
-        x_vals=configs,
-        line_arg="provider",
-        line_vals=["vllm", "triton"],  # "triton"
-        line_names=["VLLM", "Triton"],  # "Triton"
-        plot_name="moe-align-block-size-performance",
-        args={},
-    )
-)
-def benchmark(num_tokens, num_experts, topk, provider):
-    """Benchmark function for Triton."""
-    block_size = 256
-    topk_ids = get_topk_ids(num_tokens, num_experts, topk)
-
-    max_num_tokens_padded = topk_ids.numel() + num_experts * (block_size - 1)
-    sorted_ids = torch.empty((max_num_tokens_padded,), dtype=torch.int32, device="cuda")
-    sorted_ids.fill_(topk_ids.numel())
-    max_num_m_blocks = max_num_tokens_padded // block_size
-    expert_ids = torch.empty((max_num_m_blocks,), dtype=torch.int32, device="cuda")
-    num_tokens_post_pad = torch.empty((1,), dtype=torch.int32, device="cuda")
-
-    quantiles = [0.5, 0.2, 0.8]
-
-    if provider == "vllm":
-        ms, min_ms, max_ms = triton.testing.do_bench(
-            lambda: ops.moe_align_block_size(
-                topk_ids,
-                num_experts,
-                block_size,
-                sorted_ids.clone(),
-                expert_ids.clone(),
-                num_tokens_post_pad.clone(),
-            ),
-            quantiles=quantiles,
-        )
-    elif provider == "triton":
-        ms, min_ms, max_ms = triton.testing.do_bench(
-            lambda: moe_align_block_size_triton(
-                topk_ids,
-                num_experts,
-                block_size,
-                sorted_ids.clone(),
-                expert_ids.clone(),
-                num_tokens_post_pad.clone(),
-            ),
-            quantiles=quantiles,
-        )
-
-    return 1000 * ms, 1000 * max_ms, 1000 * min_ms
-
-
-if __name__ == "__main__":
-    parser = argparse.ArgumentParser()
-    parser.add_argument(
-        "--num_experts",
-        type=int,
-        default=64,
-        choices=[8, 16, 32, 64, 128, 256],
-    )
-    parser.add_argument(
-        "--topk",
-        type=int,
-        default=8,
-        choices=[2, 4, 8],
-        help="Top-k value for correctness check.",
-    )
-    args = parser.parse_args()
-
-    print("Running correctness check...")
-    check_correctness(num_tokens=1024, num_experts=args.num_experts, topk=args.topk)
-    benchmark.run(print_data=True, show_plots=True)

benchmarks/kernels/benchmark_moe_permute_unpermute.py

@@ -1,5 +1,4 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 import argparse
 from typing import Any, TypedDict

benchmarks/kernels/benchmark_paged_attention.py

@@ -1,5 +1,4 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 import random
 import time

benchmarks/kernels/benchmark_quant.py

@@ -1,5 +1,4 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 import time
 

benchmarks/kernels/benchmark_rmsnorm.py

@@ -1,5 +1,4 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 import itertools
 from typing import Optional, Union

benchmarks/kernels/benchmark_rope.py

@@ -1,5 +1,4 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 from itertools import accumulate
 from typing import Optional
@@ -23,7 +22,7 @@ def benchmark_rope_kernels_multi_lora(
     seed: int,
     device: str,
     max_position: int = 8192,
-    base: float = 10000,
+    base: int = 10000,
 ) -> None:
     current_platform.seed_everything(seed)
     torch.set_default_device(device)

benchmarks/kernels/benchmark_shapes.py

@@ -1,5 +1,4 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 WEIGHT_SHAPES = {
     "ideal": [[4 * 256 * 32, 256 * 32]],

benchmarks/kernels/benchmark_w8a8_block_fp8.py

@@ -1,5 +1,4 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 # Adapted from sglang quantization/tuning_block_wise_kernel.py
 
 import argparse

benchmarks/kernels/deepgemm/benchmark_fp8_block_dense_gemm.py

@@ -1,5 +1,4 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 # fmt: off
 # ruff: noqa: E501
 import time

benchmarks/kernels/graph_machete_bench.py

@@ -1,5 +1,4 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 import math
 import pickle

benchmarks/kernels/utils.py

@@ -1,5 +1,4 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 import dataclasses
 from collections.abc import Iterable

benchmarks/kernels/weight_shapes.py

@@ -1,5 +1,4 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 # Weight Shapes are in the format
 # ([K, N], TP_SPLIT_DIM)
@@ -49,50 +48,4 @@ WEIGHT_SHAPES = {
         ([16384, 106496], 1),
         ([53248, 16384], 0),
     ],
-    "meta-llama/Llama-3.1-8B-Instruct": [
-        ([4096, 6144], 1),
-        ([4096, 4096], 0),
-        ([4096, 28672], 1),
-        ([14336, 4096], 0),
-    ],
-    "meta-llama/Llama-3.3-70B-Instruct": [
-        ([8192, 10240], 1),
-        ([8192, 8192], 0),
-        ([8192, 57344], 1),
-        ([28672, 8192], 0),
-    ],
-    "mistralai/Mistral-Large-Instruct-2407": [
-        ([12288, 14336], 1),
-        ([12288, 12288], 0),
-        ([12288, 57344], 1),
-        ([28672, 12288], 0),
-    ],
-    "Qwen/Qwen2.5-7B-Instruct": [
-        ([3584, 4608], 1),
-        ([3584, 3584], 0),
-        ([3584, 37888], 1),
-        ([18944, 3584], 0),
-    ],
-    "Qwen/Qwen2.5-32B-Instruct": [
-        ([5120, 7168], 1),
-        ([5120, 5120], 0),
-        ([5120, 55296], 1),
-        ([27648, 5120], 0),
-    ],
-    "Qwen/Qwen2.5-72B-Instruct": [
-        ([8192, 10240], 1),
-        ([8192, 8192], 0),
-        ([8192, 59136], 1),
-        ([29568, 8192], 0),
-    ],
-    "deepseek-ai/DeepSeek-Coder-V2-Lite-Instruct": [
-        ([2048, 3072], 1),
-        ([2048, 4096], 1),
-        ([2048, 2048], 0),
-        ([2048, 576], 0),
-        ([2048, 21888], 1),
-        ([10944, 2048], 0),
-        ([2048, 2816], 1),
-        ([1408, 2048], 0),
-    ],
 }

benchmarks/overheads/benchmark_hashing.py

@@ -1,5 +1,4 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 import cProfile
 import pstats

cmake/cpu_extension.cmake

@@ -75,7 +75,6 @@ if (MACOSX_FOUND AND CMAKE_SYSTEM_PROCESSOR STREQUAL "arm64")
 else()
     find_isa(${CPUINFO} "avx2" AVX2_FOUND)
     find_isa(${CPUINFO} "avx512f" AVX512_FOUND)
-    find_isa(${CPUINFO} "Power11" POWER11_FOUND)
     find_isa(${CPUINFO} "POWER10" POWER10_FOUND)
     find_isa(${CPUINFO} "POWER9" POWER9_FOUND)
     find_isa(${CPUINFO} "asimd" ASIMD_FOUND) # Check for ARM NEON support
@@ -107,19 +106,13 @@ elseif (AVX2_FOUND)
     list(APPEND CXX_COMPILE_FLAGS "-mavx2")
     message(WARNING "vLLM CPU backend using AVX2 ISA")
     
-elseif (POWER9_FOUND OR POWER10_FOUND OR POWER11_FOUND)
+elseif (POWER9_FOUND OR POWER10_FOUND)
     message(STATUS "PowerPC detected")
-    if (POWER9_FOUND)
-        list(APPEND CXX_COMPILE_FLAGS
-            "-mvsx"
-            "-mcpu=power9"
-            "-mtune=power9")
-    elseif (POWER10_FOUND OR POWER11_FOUND)
-        list(APPEND CXX_COMPILE_FLAGS
-            "-mvsx"
-            "-mcpu=power10"
-            "-mtune=power10")
-    endif()
+    # Check for PowerPC VSX support
+    list(APPEND CXX_COMPILE_FLAGS
+        "-mvsx"
+        "-mcpu=native"
+        "-mtune=native")
 
 elseif (ASIMD_FOUND)
     message(STATUS "ARMv8 or later architecture detected")

cmake/external_projects/vllm_flash_attn.cmake

@@ -38,7 +38,7 @@ else()
   FetchContent_Declare(
           vllm-flash-attn
           GIT_REPOSITORY https://github.com/vllm-project/flash-attention.git
-          GIT_TAG 763ad155a1c826f71ff318f41edb1e4e5e376ddb
+          GIT_TAG 8798f27777fb57f447070301bf33a9f9c607f491
           GIT_PROGRESS TRUE
           # Don't share the vllm-flash-attn build between build types
           BINARY_DIR ${CMAKE_BINARY_DIR}/vllm-flash-attn
@@ -46,38 +46,22 @@ else()
 endif()
 
 
-# Ensure the vllm/vllm_flash_attn directory exists before installation
-install(CODE "file(MAKE_DIRECTORY \"\${CMAKE_INSTALL_PREFIX}/vllm/vllm_flash_attn\")" ALL_COMPONENTS)
-
-# Make sure vllm-flash-attn install rules are nested under vllm/
-# This is here to support installing all components under the same prefix with cmake --install.
-# setup.py installs every component separately but uses the same prefix for all.
-# ALL_COMPONENTS is used to avoid duplication for FA2 and FA3,
-# and these statements don't hurt when installing neither component.
-install(CODE "set(CMAKE_INSTALL_LOCAL_ONLY FALSE)" ALL_COMPONENTS)
-install(CODE "set(OLD_CMAKE_INSTALL_PREFIX \"\${CMAKE_INSTALL_PREFIX}\")" ALL_COMPONENTS)
-install(CODE "set(CMAKE_INSTALL_PREFIX \"\${CMAKE_INSTALL_PREFIX}/vllm/\")" ALL_COMPONENTS)
-
 # Fetch the vllm-flash-attn library
 FetchContent_MakeAvailable(vllm-flash-attn)
 message(STATUS "vllm-flash-attn is available at ${vllm-flash-attn_SOURCE_DIR}")
 
-# Restore the install prefix
-install(CODE "set(CMAKE_INSTALL_PREFIX \"\${OLD_CMAKE_INSTALL_PREFIX}\")" ALL_COMPONENTS)
-install(CODE "set(CMAKE_INSTALL_LOCAL_ONLY TRUE)" ALL_COMPONENTS)
-
 # Copy over the vllm-flash-attn python files (duplicated for fa2 and fa3, in
 # case only one is built, in the case both are built redundant work is done)
 install(
   DIRECTORY ${vllm-flash-attn_SOURCE_DIR}/vllm_flash_attn/
-  DESTINATION vllm/vllm_flash_attn
+  DESTINATION vllm_flash_attn
   COMPONENT _vllm_fa2_C
   FILES_MATCHING PATTERN "*.py"
 )
 
 install(
   DIRECTORY ${vllm-flash-attn_SOURCE_DIR}/vllm_flash_attn/
-  DESTINATION vllm/vllm_flash_attn
+  DESTINATION vllm_flash_attn
   COMPONENT _vllm_fa3_C
   FILES_MATCHING PATTERN "*.py"
 )

cmake/hipify.py

@@ -1,6 +1,5 @@
 #!/usr/bin/env python3
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 #
 # A command line tool for running pytorch's hipify preprocessor on CUDA

cmake/utils.cmake

@@ -39,6 +39,35 @@ function (run_python OUT EXPR ERR_MSG)
   set(${OUT} ${PYTHON_OUT} PARENT_SCOPE)
 endfunction()
 
+# Generate CUDA sources via a Python script with caching based on script hash
+function(generate_cuda_sources NAME SCRIPT PYTHONPATH_PREPEND GLOB OUT_SRCS)
+  string(TOUPPER "${NAME}" _UPPER_NAME)
+  set(_CACHE_VAR "${_UPPER_NAME}_GEN_SCRIPT_HASH")
+  file(MD5 "${SCRIPT}" _GEN_HASH)
+  message(STATUS "${NAME} generation script hash: ${_GEN_HASH}")
+  message(STATUS "Last run ${NAME} generation script hash: $CACHE{${_CACHE_VAR}}")
+  if(NOT DEFINED CACHE{${_CACHE_VAR}} OR NOT $CACHE{${_CACHE_VAR}} STREQUAL "${_GEN_HASH}")
+    message(STATUS "Running ${NAME} generation script: ${SCRIPT}, with PYTHONPATH prepend: ${PYTHONPATH_PREPEND}")
+    execute_process(
+      COMMAND ${CMAKE_COMMAND} -E env
+        PYTHONPATH=${PYTHONPATH_PREPEND}:$ENV{PYTHONPATH}
+        ${Python_EXECUTABLE} "${SCRIPT}"
+      RESULT_VARIABLE _GEN_RESULT
+      OUTPUT_FILE "${CMAKE_CURRENT_BINARY_DIR}/${NAME}_generation.log"
+      ERROR_FILE "${CMAKE_CURRENT_BINARY_DIR}/${NAME}_generation.log"
+    )
+    if(NOT _GEN_RESULT EQUAL 0)
+      message(FATAL_ERROR "${NAME} generation failed. Result: \"${_GEN_RESULT}\"\nCheck the log for details: ${CMAKE_CURRENT_BINARY_DIR}/${NAME}_generation.log")
+    else()
+      set(${_CACHE_VAR} "${_GEN_HASH}" CACHE STRING "Last run ${NAME} generation script hash" FORCE)
+      message(STATUS "${NAME} generation completed successfully.")
+    endif()
+  else()
+    message(STATUS "${NAME} generation script has not changed, skipping generation.")
+  endif()
+  file(GLOB ${OUT_SRCS} "${GLOB}")
+endfunction()
+
 # Run `EXPR` in python after importing `PKG`. Use the result of this to extend
 # `CMAKE_PREFIX_PATH` so the torch cmake configuration can be imported.
 macro (append_cmake_prefix_path PKG EXPR)
@@ -76,7 +105,7 @@ function (hipify_sources_target OUT_SRCS NAME ORIG_SRCS)
   set(CSRC_BUILD_DIR ${CMAKE_CURRENT_BINARY_DIR}/csrc)
   add_custom_target(
     hipify${NAME}
-    COMMAND ${Python_EXECUTABLE} ${CMAKE_SOURCE_DIR}/cmake/hipify.py -p ${CMAKE_SOURCE_DIR}/csrc -o ${CSRC_BUILD_DIR} ${SRCS}
+    COMMAND ${CMAKE_SOURCE_DIR}/cmake/hipify.py -p ${CMAKE_SOURCE_DIR}/csrc -o ${CSRC_BUILD_DIR} ${SRCS}
     DEPENDS ${CMAKE_SOURCE_DIR}/cmake/hipify.py ${SRCS}
     BYPRODUCTS ${HIP_SRCS}
     COMMENT "Running hipify on ${NAME} extension source files.")
@@ -86,6 +115,87 @@ function (hipify_sources_target OUT_SRCS NAME ORIG_SRCS)
   set(${OUT_SRCS} ${HIP_SRCS} PARENT_SCOPE)
 endfunction()
 
+# Macro to conditionally include CUDA sources based on architecture and CUDA
+# compiler version, optionally generating sources via a Python script.
+# Usage:
+#   optional_cuda_sources(
+#     NAME <name>
+#     [MIN_VERSION <version>]
+#     ARCHS <arch1;arch2;...>
+#     SRCS <source1> [<source2> ...]
+#     [FLAGS <flag1> ...]
+#     [VERSION_MSG <line1> [<line2> ...]]
+#     [NO_ARCH_MSG <line1> [<line2> ...]]
+#     [GEN_SCRIPT <path/to/generate_script.py>]
+#     [GEN_PYTHONPATH_PREPEND <str_to_prepend_to_pythonpath>]
+#     [GEN_GLOB <glob_pattern_for_generated_sources>]
+# This will check if `CMAKE_CUDA_COMPILER_VERSION` is greater than or equal
+# to `MIN_VERSION` and the `cuda_archs_loose_intersection` of `ARCHS` and 
+# `CUDA_ARCHS` (taken from global scope) is not empty. If both these conditions
+# are met, it will:
+#  1) set the gencode flags of the sources in `SRCS` to be the 
+#     `cuda_archs_loose_intersection` of `ARCHS` and `CUDA_ARCHS`.
+#  2) append the sources in `SRCS` to the global `VLLM_EXT_SRC` variable.
+#  3) append the flags in `FLAGS` to the global `VLLM_GPU_FLAGS` variable.
+#
+# This will also run GEN_SCRIPT (if supplied and the hash of the script does not
+# match the latest in the cmake cache), with GEN_PYTHONPATH_PREPEND prepended to
+# to the PYTHONPATH when calling  before globbing sources matching GEN_GLOB
+# and appending them alongside SRCS (with the gencodes set)
+macro(optional_cuda_sources)
+  set(oneValueArgs NAME MIN_VERSION GEN_SCRIPT GEN_PYTHONPATH_PREPEND GEN_GLOB OUT_SRCS_VAR)
+  set(multiValueArgs ARCHS SRCS FLAGS VERSION_MSG NO_ARCH_MSG)
+  cmake_parse_arguments(OCS "" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
+  if(NOT OCS_NAME)
+    message(FATAL_ERROR "optional_cuda_sources: NAME is required")
+  endif()
+  if(NOT OCS_ARCHS)
+    message(FATAL_ERROR "optional_cuda_sources ${OCS_NAME}: ARCHS is required")
+  endif()
+  if(NOT OCS_SRCS AND NOT OCS_GEN_SCRIPT)
+    message(FATAL_ERROR "optional_cuda_sources ${OCS_NAME}: either SRCS or GEN_SCRIPT must be provided")
+  endif()
+  if(NOT OCS_MIN_VERSION)
+    set(OCS_MIN_VERSION "0.0")
+  endif()
+  cuda_archs_loose_intersection(_OCS_ARCHS "${OCS_ARCHS}" "${CUDA_ARCHS}")
+  if(CMAKE_CUDA_COMPILER_VERSION VERSION_GREATER_EQUAL ${OCS_MIN_VERSION})
+    if(_OCS_ARCHS)
+      set(_OCS_SRCS ${OCS_SRCS})
+      # Generate sources if a script is provided
+      if(OCS_GEN_SCRIPT AND OCS_GEN_GLOB)
+        generate_cuda_sources(
+          ${OCS_NAME} "${OCS_GEN_SCRIPT}" "${OCS_GEN_PYTHONPATH_PREPEND}" "${OCS_GEN_GLOB}" _OCS_GEN_SRCS)
+        list(APPEND _OCS_SRCS ${_OCS_GEN_SRCS})
+      endif()
+      set_gencode_flags_for_srcs(SRCS "${_OCS_SRCS}" CUDA_ARCHS "${_OCS_ARCHS}")
+      if(OCS_OUT_SRCS_VAR)
+        list(APPEND ${OCS_OUT_SRCS_VAR} ${_OCS_SRCS})
+      else()
+        list(APPEND VLLM_EXT_SRC ${_OCS_SRCS})
+      endif()
+      if(OCS_FLAGS)
+        list(APPEND VLLM_GPU_FLAGS ${OCS_FLAGS})
+      endif()
+      message(STATUS "Building ${OCS_NAME} for archs: ${_OCS_ARCHS}")
+    else()
+      if(OCS_NO_ARCH_MSG)
+        list(JOIN OCS_NO_ARCH_MSG "\n   " _OCS_NO_ARCH_JOINED)
+        message(STATUS "${_OCS_NO_ARCH_JOINED}")
+      else()
+        message(STATUS "Not building ${OCS_NAME}: no compatible architectures found in CUDA target architectures")
+      endif()
+    endif()
+  else()
+    if(OCS_VERSION_MSG)
+      list(JOIN OCS_VERSION_MSG "\n   " _OCS_VERSION_JOINED)
+      message(STATUS "${_OCS_VERSION_JOINED}")
+    else()
+      message(STATUS "Not building ${OCS_NAME}: CUDA Compiler version is less than ${OCS_MIN_VERSION}")
+    endif()
+  endif()
+endmacro()
+
 #
 # Get additional GPU compiler flags from torch.
 #
@@ -122,7 +232,6 @@ function (get_torch_gpu_compiler_flags OUT_GPU_FLAGS GPU_LANG)
       "-DENABLE_FP8"
       "-U__HIP_NO_HALF_CONVERSIONS__"
       "-U__HIP_NO_HALF_OPERATORS__"
-      "-Werror=unused-variable"
       "-fno-gpu-rdc")
 
   endif()

csrc/attention/merge_attn_states.cu

@@ -143,14 +143,6 @@ void merge_attn_states_launcher(torch::Tensor& output,
   const uint pack_size = 16 / sizeof(scalar_t);
   TORCH_CHECK(head_size % pack_size == 0,
               "headsize must be multiple of pack_size:", pack_size);
-  TORCH_CHECK(output.stride(-2) == head_size && output.stride(-1) == 1,
-              "output heads must be contiguous in memory");
-  TORCH_CHECK(
-      prefix_output.stride(-2) == head_size && prefix_output.stride(-1) == 1,
-      "prefix_output heads must be contiguous in memory");
-  TORCH_CHECK(
-      suffix_output.stride(-2) == head_size && suffix_output.stride(-1) == 1,
-      "suffix_output heads must be contiguous in memory");
   float* output_lse_ptr = nullptr;
   if (output_lse.has_value()) {
     output_lse_ptr = output_lse.value().data_ptr<float>();

csrc/attention/mla/cutlass_mla_kernels.cu

@@ -119,7 +119,7 @@ typename T::Fmha::Arguments args_from_options(
       {static_cast<ElementOut*>(out.data_ptr()), stride_O,
        static_cast<ElementAcc*>(nullptr), stride_LSE},
       hw_info,
-      1,        // split_kv
+      -1,       // split_kv
       nullptr,  // is_var_split_kv
   };
   // TODO(kaixih@nvidia): When split_kv=-1 and is_var_split_kv=false, we compute

csrc/attention/paged_attention_v1.cu

@@ -65,6 +65,9 @@ void paged_attention_v1_launcher(
   int kv_block_stride = key_cache.stride(0);
   int kv_head_stride = key_cache.stride(1);
 
+  [[maybe_unused]] int thread_group_size = MAX(WARP_SIZE / BLOCK_SIZE, 1);
+  assert(head_size % thread_group_size == 0);
+
   // NOTE: alibi_slopes is optional.
   const float* alibi_slopes_ptr =
       alibi_slopes
@@ -190,4 +193,4 @@ void paged_attention_v1(
 #undef WARP_SIZE
 #undef MAX
 #undef MIN
-#undef DIVIDE_ROUND_UP
+#undef DIVIDE_ROUND_UP

csrc/attention/paged_attention_v2.cu

@@ -66,6 +66,9 @@ void paged_attention_v2_launcher(
   int kv_block_stride = key_cache.stride(0);
   int kv_head_stride = key_cache.stride(1);
 
+  [[maybe_unused]] int thread_group_size = MAX(WARP_SIZE / BLOCK_SIZE, 1);
+  assert(head_size % thread_group_size == 0);
+
   // NOTE: alibi_slopes is optional.
   const float* alibi_slopes_ptr =
       alibi_slopes
@@ -200,4 +203,4 @@ void paged_attention_v2(
 #undef WARP_SIZE
 #undef MAX
 #undef MIN
-#undef DIVIDE_ROUND_UP
+#undef DIVIDE_ROUND_UP

csrc/cpu/attention.cpp

@@ -137,8 +137,8 @@ FORCE_INLINE std::pair<T, T> reduceSoftmaxAlibi(T* data, const int size,
 }
 
 template <typename T>
-FORCE_INLINE void reducePartitionSoftmax(const T* max_data, T* sum_data,
-                                         const int size) {
+FORCE_INLINE void reducePartitonSoftmax(const T* max_data, T* sum_data,
+                                        const int size) {
   T max = max_data[0];
   for (int i = 1; i < size; ++i) {
     max = max >= max_data[i] ? max : max_data[i];
@@ -634,7 +634,7 @@ struct paged_attention_v2_impl {
 
         if (partition_num == 1) continue;
 
-        reducePartitionSoftmax(
+        reducePartitonSoftmax(
             max_logits + seq_idx * num_heads * max_num_partitions +
                 head_idx * max_num_partitions,
             exp_sums + seq_idx * num_heads * max_num_partitions +

csrc/cpu/cpu_types_x86.hpp

@@ -83,7 +83,7 @@ struct FP16Vec16 : public Vec<FP16Vec16> {
   explicit FP16Vec16(const void* ptr)
       : reg((__m256i)_mm256_loadu_si256((__m256i*)ptr)) {}
 
-  // non-temporal load
+  // non-temproal load
   explicit FP16Vec16(bool, void* ptr)
       : reg(_mm256_stream_load_si256((__m256i*)ptr)) {}
 
@@ -120,7 +120,7 @@ struct BF16Vec16 : public Vec<BF16Vec16> {
   explicit BF16Vec16(const void* ptr)
       : reg((__m256i)_mm256_loadu_si256((__m256i*)ptr)) {}
 
-  // non-temporal load
+  // non-temproal load
   explicit BF16Vec16(bool, void* ptr)
       : reg(_mm256_stream_load_si256((__m256i*)ptr)) {}
 
@@ -327,7 +327,7 @@ struct FP32Vec16 : public Vec<FP32Vec16> {
   // normal load
   explicit FP32Vec16(const float* ptr) : reg(_mm512_loadu_ps(ptr)) {}
 
-  // non-temporal load
+  // non-temproal load
   explicit FP32Vec16(bool, void* ptr)
       : reg((__m512)_mm512_stream_load_si512(ptr)) {}
 
@@ -576,7 +576,7 @@ struct INT8Vec64 : public Vec<INT8Vec64> {
   // normal load
   explicit INT8Vec64(void* ptr) : reg(_mm512_loadu_epi8(ptr)) {}
 
-  // non-temporal load
+  // non-temproal load
   explicit INT8Vec64(bool, void* ptr) : reg(_mm512_stream_load_si512(ptr)) {}
 
   void save(void* ptr) const { _mm512_storeu_epi8(ptr, reg); }
@@ -587,7 +587,7 @@ struct INT8Vec64 : public Vec<INT8Vec64> {
     _mm512_mask_storeu_epi8(ptr, mask, reg);
   }
 
-  // non-temporal save
+  // non-temproal save
   void nt_save(int8_t* ptr) { _mm512_stream_si512((__m512i*)ptr, reg); }
 };
 #endif

csrc/cpu/utils.cpp

@@ -54,7 +54,8 @@ std::string init_cpu_threads_env(const std::string& cpu_ids) {
     *(src_mask->maskp) = *(src_mask->maskp) ^ *(mask->maskp);
     int page_num = numa_migrate_pages(pid, src_mask, mask);
     if (page_num == -1) {
-      TORCH_WARN("numa_migrate_pages failed. errno: " + std::to_string(errno));
+      TORCH_CHECK(false,
+                  "numa_migrate_pages failed. errno: " + std::to_string(errno));
     }
 
     // restrict memory allocation node.
@@ -104,4 +105,4 @@ std::string init_cpu_threads_env(const std::string& cpu_ids) {
 
   return ss.str();
 }
-#endif
+#endif

csrc/cutlass_extensions/vllm_cutlass_library_extension.py

@@ -1,5 +1,4 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 import enum
 from typing import Union

csrc/moe/marlin_moe_wna16/generate_kernels.py

@@ -1,5 +1,4 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 import glob
 import itertools
 import os

csrc/moe/moe_align_sum_kernels.cu

@@ -13,45 +13,232 @@
 namespace vllm {
 namespace moe {
 
-template <typename scalar_t>
-__global__ void moe_align_block_size_kernel(
-    const scalar_t* __restrict__ topk_ids,
-    int32_t* __restrict__ sorted_token_ids, int32_t* __restrict__ expert_ids,
-    int32_t* __restrict__ total_tokens_post_pad, int32_t num_experts,
-    int32_t padded_num_experts, int32_t experts_per_warp, int32_t block_size,
-    size_t numel, int32_t* __restrict__ cumsum) {
-  extern __shared__ int32_t shared_counts[];
+namespace {
+__device__ __forceinline__ int32_t index(int32_t total_col, int32_t row,
+                                         int32_t col) {
+  // don't worry about overflow because num_experts is relatively small
+  return row * total_col + col;
+}
+}  // namespace
 
-  const int warp_id = threadIdx.x / WARP_SIZE;
-  const int my_expert_start = warp_id * experts_per_warp;
+template <typename scalar_t, typename token_cnts_t>
+__global__ void moe_align_block_size_kernel(scalar_t* __restrict__ topk_ids,
+                                            int32_t* sorted_token_ids,
+                                            int32_t* expert_ids,
+                                            int32_t* total_tokens_post_pad,
+                                            int32_t num_experts,
+                                            int32_t block_size, size_t numel) {
+  const size_t tokens_per_thread = CEILDIV(numel, blockDim.x);
+  const size_t start_idx = threadIdx.x * tokens_per_thread;
 
-  for (int i = 0; i < experts_per_warp; ++i) {
-    if (my_expert_start + i < padded_num_experts) {
-      shared_counts[warp_id * experts_per_warp + i] = 0;
+  extern __shared__ int32_t shared_mem[];
+  int32_t* cumsum = shared_mem;  // 1d tensor with shape (num_experts + 1)
+  token_cnts_t* tokens_cnts =
+      (token_cnts_t*)(shared_mem + num_experts +
+                      1);  // 2d tensor with shape (blockDim.x + 1, num_experts)
+
+  for (int i = 0; i < num_experts; ++i) {
+    tokens_cnts[index(num_experts, threadIdx.x + 1, i)] = 0;
+  }
+
+  /**
+   * In the first step we compute token_cnts[thread_index + 1][expert_index],
+   * which counts how many tokens in the token shard of thread_index are
+   * assigned to expert expert_index.
+   */
+  for (int i = start_idx; i < numel && i < start_idx + tokens_per_thread; ++i) {
+    ++tokens_cnts[index(num_experts, threadIdx.x + 1, topk_ids[i])];
+  }
+
+  __syncthreads();
+
+  // For each expert we accumulate the token counts from the different threads.
+  if (threadIdx.x < num_experts) {
+    tokens_cnts[index(num_experts, 0, threadIdx.x)] = 0;
+    for (int i = 1; i <= blockDim.x; ++i) {
+      tokens_cnts[index(num_experts, i, threadIdx.x)] +=
+          tokens_cnts[index(num_experts, i - 1, threadIdx.x)];
     }
   }
 
   __syncthreads();
 
-  const size_t tid = threadIdx.x;
-  const size_t stride = blockDim.x;
-
-  for (size_t i = tid; i < numel; i += stride) {
-    int expert_id = topk_ids[i];
-    int warp_idx = expert_id / experts_per_warp;
-    int expert_offset = expert_id % experts_per_warp;
-    atomicAdd(&shared_counts[warp_idx * experts_per_warp + expert_offset], 1);
+  // We accumulate the token counts of all experts in thread 0.
+  if (threadIdx.x == 0) {
+    cumsum[0] = 0;
+    for (int i = 1; i <= num_experts; ++i) {
+      cumsum[i] = cumsum[i - 1] +
+                  CEILDIV(tokens_cnts[index(num_experts, blockDim.x, i - 1)],
+                          block_size) *
+                      block_size;
+    }
+    *total_tokens_post_pad = static_cast<int32_t>(cumsum[num_experts]);
   }
 
   __syncthreads();
 
+  /**
+   * For each expert, each thread processes the tokens of the corresponding
+   * blocks and stores the corresponding expert_id for each block.
+   */
+  if (threadIdx.x < num_experts) {
+    for (int i = cumsum[threadIdx.x]; i < cumsum[threadIdx.x + 1];
+         i += block_size) {
+      expert_ids[i / block_size] = threadIdx.x;
+    }
+  }
+
+  /**
+   * Each thread processes a token shard, calculating the index of each token
+   * after sorting by expert number. Given the example topk_ids =
+   * [0,1,2,1,2,3,0,3,4] and block_size = 4, then the output would be [0, 6, *,
+   * *, 1, 3, *, *, 2, 4, *, *, 5, 7, *, *, 8, *, *, *], where * represents a
+   * padding value(preset in python).
+   */
+  for (int i = start_idx; i < numel && i < start_idx + tokens_per_thread; ++i) {
+    int32_t expert_id = topk_ids[i];
+    /** The cumsum[expert_id] stores the starting index of the tokens that the
+     * expert with expert_id needs to process, and
+     * tokens_cnts[threadIdx.x][expert_id] stores the indices of the tokens
+     * processed by the expert with expert_id within the current thread's token
+     * shard.
+     */
+    int32_t rank_post_pad =
+        tokens_cnts[index(num_experts, threadIdx.x, expert_id)] +
+        cumsum[expert_id];
+    sorted_token_ids[rank_post_pad] = i;
+    ++tokens_cnts[index(num_experts, threadIdx.x, expert_id)];
+  }
+}
+
+// TODO(simon): this is temporarily adapted from
+// https://github.com/sgl-project/sglang/commit/31548116a8dc8c6df7e146e0587335a59fc5b9d7
+// we did this to unblock Deepseek V3 but there should be a better
+// implementation to manage shared memory.
+template <typename scalar_t>
+__global__ void moe_align_block_size_global_mem_kernel(
+    scalar_t* __restrict__ topk_ids, int32_t* sorted_token_ids,
+    int32_t* expert_ids, int32_t* total_tokens_post_pad, int32_t num_experts,
+    int32_t block_size, size_t numel, int32_t* tokens_cnts, int32_t* cumsum) {
+  const size_t tokens_per_thread = CEILDIV(numel, blockDim.x);
+  const size_t start_idx = threadIdx.x * tokens_per_thread;
+
+  for (int i = 0; i < num_experts; ++i) {
+    tokens_cnts[index(num_experts, threadIdx.x + 1, i)] = 0;
+  }
+
+  /**
+   * In the first step we compute token_cnts[thread_index + 1][expert_index],
+   * which counts how many tokens in the token shard of thread_index are
+   * assigned to expert expert_index.
+   */
+  for (int i = start_idx; i < numel && i < start_idx + tokens_per_thread; ++i) {
+    ++tokens_cnts[index(num_experts, threadIdx.x + 1, topk_ids[i])];
+  }
+
+  __syncthreads();
+
+  // For each expert we accumulate the token counts from the different threads.
+  if (threadIdx.x < num_experts) {
+    tokens_cnts[index(num_experts, 0, threadIdx.x)] = 0;
+    for (int i = 1; i <= blockDim.x; ++i) {
+      tokens_cnts[index(num_experts, i, threadIdx.x)] +=
+          tokens_cnts[index(num_experts, i - 1, threadIdx.x)];
+    }
+  }
+
+  __syncthreads();
+
+  // We accumulate the token counts of all experts in thread 0.
+  if (threadIdx.x == 0) {
+    cumsum[0] = 0;
+    for (int i = 1; i <= num_experts; ++i) {
+      cumsum[i] = cumsum[i - 1] +
+                  CEILDIV(tokens_cnts[index(num_experts, blockDim.x, i - 1)],
+                          block_size) *
+                      block_size;
+    }
+    *total_tokens_post_pad = cumsum[num_experts];
+  }
+
+  __syncthreads();
+
+  /**
+   * For each expert, each thread processes the tokens of the corresponding
+   * blocks and stores the corresponding expert_id for each block.
+   */
+  if (threadIdx.x < num_experts) {
+    for (int i = cumsum[threadIdx.x]; i < cumsum[threadIdx.x + 1];
+         i += block_size) {
+      expert_ids[i / block_size] = threadIdx.x;
+    }
+  }
+
+  /**
+   * Each thread processes a token shard, calculating the index of each token
+   * after sorting by expert number. Given the example topk_ids =
+   * [0,1,2,1,2,3,0,3,4] and block_size = 4, then the output would be [0, 6, *,
+   * *, 1, 3, *, *, 2, 4, *, *, 5, 7, *, *, 8, *, *, *], where * represents a
+   * padding value(preset in python).
+   */
+  for (int i = start_idx; i < numel && i < start_idx + tokens_per_thread; ++i) {
+    int32_t expert_id = topk_ids[i];
+    /** The cumsum[expert_id] stores the starting index of the tokens that the
+     * expert with expert_id needs to process, and
+     * tokens_cnts[threadIdx.x][expert_id] stores the indices of the tokens
+     * processed by the expert with expert_id within the current thread's token
+     * shard.
+     */
+    int32_t rank_post_pad =
+        tokens_cnts[index(num_experts, threadIdx.x, expert_id)] +
+        cumsum[expert_id];
+    sorted_token_ids[rank_post_pad] = i;
+    ++tokens_cnts[index(num_experts, threadIdx.x, expert_id)];
+  }
+}
+
+// taken from
+// https://github.com/sgl-project/sglang/commit/cdae77b03dfc6fec3863630550b45bbfc789f957
+template <typename scalar_t>
+__global__ void sgl_moe_align_block_size_kernel(
+    scalar_t* __restrict__ topk_ids, int32_t* sorted_token_ids,
+    int32_t* expert_ids, int32_t* total_tokens_post_pad, int32_t num_experts,
+    int32_t block_size, size_t numel, int32_t* cumsum) {
+  __shared__ int32_t shared_counts[32][8];
+
+  const int warp_id = threadIdx.x / 32;
+  const int experts_per_warp = 8;
+  const int my_expert_start = warp_id * experts_per_warp;
+
+  // Initialize shared_counts for this warp's experts
+  for (int i = 0; i < experts_per_warp; ++i) {
+    if (my_expert_start + i < num_experts) {
+      shared_counts[warp_id][i] = 0;
+    }
+  }
+
+  __syncthreads();
+
+  const size_t tokens_per_thread = CEILDIV(numel, blockDim.x);
+  const size_t start_idx = threadIdx.x * tokens_per_thread;
+
+  for (int i = start_idx; i < numel && i < start_idx + tokens_per_thread; ++i) {
+    int expert_id = topk_ids[i];
+    int warp_idx = expert_id / experts_per_warp;
+    int expert_offset = expert_id % experts_per_warp;
+    atomicAdd(&shared_counts[warp_idx][expert_offset], 1);
+  }
+
+  __syncthreads();
+
+  // Single thread computes cumulative sum and total tokens
   if (threadIdx.x == 0) {
     cumsum[0] = 0;
     for (int i = 1; i <= num_experts; ++i) {
       int expert_count = 0;
       int warp_idx = (i - 1) / experts_per_warp;
       int expert_offset = (i - 1) % experts_per_warp;
-      expert_count = shared_counts[warp_idx * experts_per_warp + expert_offset];
+      expert_count = shared_counts[warp_idx][expert_offset];
 
       cumsum[i] =
           cumsum[i - 1] + CEILDIV(expert_count, block_size) * block_size;
@@ -61,6 +248,7 @@ __global__ void moe_align_block_size_kernel(
 
   __syncthreads();
 
+  // Assign expert IDs to blocks
   if (threadIdx.x < num_experts) {
     for (int i = cumsum[threadIdx.x]; i < cumsum[threadIdx.x + 1];
          i += block_size) {
@@ -69,11 +257,13 @@ __global__ void moe_align_block_size_kernel(
   }
 }
 
+// taken from
+// https://github.com/sgl-project/sglang/commit/cdae77b03dfc6fec3863630550b45bbfc789f957
 template <typename scalar_t>
-__global__ void count_and_sort_expert_tokens_kernel(
-    const scalar_t* __restrict__ topk_ids,
-    int32_t* __restrict__ sorted_token_ids, int32_t* __restrict__ cumsum_buffer,
-    size_t numel) {
+__global__ void sgl_moe_token_sort_kernel(scalar_t* __restrict__ topk_ids,
+                                          int32_t* sorted_token_ids,
+                                          int32_t* cumsum_buffer,
+                                          size_t numel) {
   const size_t tid = blockIdx.x * blockDim.x + threadIdx.x;
   const size_t stride = blockDim.x * gridDim.x;
 
@@ -100,138 +290,132 @@ __global__ void moe_sum_kernel(
   }
 }
 
-template <typename scalar_t>
-__global__ void moe_align_block_size_small_batch_expert_kernel(
-    const scalar_t* __restrict__ topk_ids,
-    int32_t* __restrict__ sorted_token_ids, int32_t* __restrict__ expert_ids,
-    int32_t* __restrict__ total_tokens_post_pad, int32_t num_experts,
-    int32_t block_size, size_t numel) {
-  const size_t tid = threadIdx.x;
-  const size_t stride = blockDim.x;
-
-  extern __shared__ int32_t shared_mem[];
-  int32_t* cumsum = shared_mem;
-  int32_t* tokens_cnts = (int32_t*)(shared_mem + num_experts + 1);
-
-  for (int i = 0; i < num_experts; ++i) {
-    tokens_cnts[(threadIdx.x + 1) * num_experts + i] = 0;
-  }
-
-  for (size_t i = tid; i < numel; i += stride) {
-    ++tokens_cnts[(threadIdx.x + 1) * num_experts + topk_ids[i]];
-  }
-
-  __syncthreads();
-
-  if (threadIdx.x < num_experts) {
-    tokens_cnts[threadIdx.x] = 0;
-    for (int i = 1; i <= blockDim.x; ++i) {
-      tokens_cnts[i * num_experts + threadIdx.x] +=
-          tokens_cnts[(i - 1) * num_experts + threadIdx.x];
-    }
-  }
-
-  __syncthreads();
-
-  if (threadIdx.x == 0) {
-    cumsum[0] = 0;
-    for (int i = 1; i <= num_experts; ++i) {
-      cumsum[i] =
-          cumsum[i - 1] +
-          CEILDIV(tokens_cnts[blockDim.x * num_experts + i - 1], block_size) *
-              block_size;
-    }
-    *total_tokens_post_pad = static_cast<int32_t>(cumsum[num_experts]);
-  }
-
-  __syncthreads();
-
-  if (threadIdx.x < num_experts) {
-    for (int i = cumsum[threadIdx.x]; i < cumsum[threadIdx.x + 1];
-         i += block_size) {
-      expert_ids[i / block_size] = threadIdx.x;
-    }
-  }
-
-  for (size_t i = tid; i < numel; i += stride) {
-    int32_t expert_id = topk_ids[i];
-    int32_t rank_post_pad =
-        tokens_cnts[threadIdx.x * num_experts + expert_id] + cumsum[expert_id];
-    sorted_token_ids[rank_post_pad] = i;
-    ++tokens_cnts[threadIdx.x * num_experts + expert_id];
-  }
-}
-
 }  // namespace moe
 }  // namespace vllm
 
-// taken from
-// https://github.com/sgl-project/sglang/blob/8b5f83ed3b7d2a49ad5c5cd5aa61c5d502f47dbc
 void moe_align_block_size(torch::Tensor topk_ids, int64_t num_experts,
                           int64_t block_size, torch::Tensor sorted_token_ids,
                           torch::Tensor experts_ids,
                           torch::Tensor num_tokens_post_pad) {
   const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
 
-  int64_t padded_num_experts =
-      ((num_experts + WARP_SIZE - 1) / WARP_SIZE) * WARP_SIZE;
-  int experts_per_warp = WARP_SIZE;
-  int threads = 1024;
-  threads = ((threads + WARP_SIZE - 1) / WARP_SIZE) * WARP_SIZE;
+  int device_max_shared_mem;
+  auto dev = topk_ids.get_device();
+  cudaDeviceGetAttribute(&device_max_shared_mem,
+                         cudaDevAttrMaxSharedMemoryPerBlockOptin, dev);
 
-  VLLM_DISPATCH_INTEGRAL_AND_UNSIGNED_TYPES(
-      topk_ids.scalar_type(), "moe_align_block_size_kernel", [&] {
-        // calc needed amount of shared mem for `cumsum` tensors
-        auto options_int =
-            torch::TensorOptions().dtype(torch::kInt).device(topk_ids.device());
-        torch::Tensor cumsum_buffer =
-            torch::zeros({num_experts + 1}, options_int);
-        bool small_batch_expert_mode =
-            (topk_ids.numel() < 1024) && (num_experts <= 64);
+  const int32_t num_thread = max((int32_t)num_experts, WARP_SIZE);
+  const int32_t shared_mem_i32 =
+      ((num_thread + 1) * num_experts + (num_experts + 1)) * sizeof(int32_t);
+  const int32_t shared_mem_i16 =
+      ((num_thread + 1) * num_experts) * sizeof(uint16_t) +
+      (num_experts + 1) * sizeof(int32_t);
 
-        if (small_batch_expert_mode) {
-          const int32_t threads = max((int32_t)num_experts, WARP_SIZE);
-          const int32_t shared_mem_size =
-              ((threads + 1) * num_experts + (num_experts + 1)) *
-              sizeof(int32_t);
+  bool use_global_memory = false;
+  bool use_i16 = false;  // Use uint16_t for shared memory token counts
+  if (shared_mem_i32 < device_max_shared_mem) {
+    // Do nothing in this case. We're all set to use int32_t token counts
+  } else if (shared_mem_i16 < device_max_shared_mem &&
+             topk_ids.numel() <= 65535) {
+    // when nelements of topk_ids is smaller than 65535 (max value of uint16),
+    // element value of token_cnts would also smaller than 65535,
+    // so we can use uint16 as dtype of token_cnts
+    use_i16 = true;
+  } else {
+    use_global_memory = true;
+  }
 
-          auto small_batch_expert_kernel =
-              vllm::moe::moe_align_block_size_small_batch_expert_kernel<
-                  scalar_t>;
-          small_batch_expert_kernel<<<1, threads, shared_mem_size, stream>>>(
+  if (use_global_memory) {
+    VLLM_DISPATCH_INTEGRAL_AND_UNSIGNED_TYPES(
+        topk_ids.scalar_type(), "moe_align_block_size_global_mem_kernel", [&] {
+          // calc needed amount of shared mem for `tokens_cnts` and `cumsum`
+          // tensors
+          const int32_t num_thread = max((int32_t)num_experts, WARP_SIZE);
+
+          auto options_int = torch::TensorOptions()
+                                 .dtype(torch::kInt)
+                                 .device(topk_ids.device());
+          torch::Tensor token_cnts_buffer =
+              torch::empty({(num_experts + 1) * num_experts}, options_int);
+          torch::Tensor cumsum_buffer =
+              torch::empty({num_experts + 1}, options_int);
+
+          auto kernel =
+              vllm::moe::moe_align_block_size_global_mem_kernel<scalar_t>;
+          kernel<<<1, num_thread, 0, stream>>>(
+              topk_ids.data_ptr<scalar_t>(),
+              sorted_token_ids.data_ptr<int32_t>(),
+              experts_ids.data_ptr<int32_t>(),
+              num_tokens_post_pad.data_ptr<int32_t>(), num_experts, block_size,
+              topk_ids.numel(), token_cnts_buffer.data_ptr<int32_t>(),
+              cumsum_buffer.data_ptr<int32_t>());
+        });
+  } else if (use_i16) {
+    VLLM_DISPATCH_INTEGRAL_AND_UNSIGNED_TYPES(
+        topk_ids.scalar_type(), "moe_align_block_size_kernel", [&] {
+          // set dynamic shared mem
+          auto kernel =
+              vllm::moe::moe_align_block_size_kernel<scalar_t, uint16_t>;
+          AT_CUDA_CHECK(VLLM_DevFuncAttribute_SET_MaxDynamicSharedMemorySize(
+              (void*)kernel, shared_mem_i16));
+          kernel<<<1, num_thread, shared_mem_i16, stream>>>(
               topk_ids.data_ptr<scalar_t>(),
               sorted_token_ids.data_ptr<int32_t>(),
               experts_ids.data_ptr<int32_t>(),
               num_tokens_post_pad.data_ptr<int32_t>(), num_experts, block_size,
               topk_ids.numel());
-        } else {
-          auto align_kernel = vllm::moe::moe_align_block_size_kernel<scalar_t>;
-
-          size_t num_warps = CEILDIV(padded_num_experts, experts_per_warp);
-          size_t shared_mem_size =
-              num_warps * experts_per_warp * sizeof(int32_t);
-
-          align_kernel<<<1, threads, shared_mem_size, stream>>>(
+        });
+  } else {
+    VLLM_DISPATCH_INTEGRAL_AND_UNSIGNED_TYPES(
+        topk_ids.scalar_type(), "moe_align_block_size_kernel", [&] {
+          auto kernel =
+              vllm::moe::moe_align_block_size_kernel<scalar_t, int32_t>;
+          AT_CUDA_CHECK(VLLM_DevFuncAttribute_SET_MaxDynamicSharedMemorySize(
+              (void*)kernel, shared_mem_i32));
+          kernel<<<1, num_thread, shared_mem_i32, stream>>>(
               topk_ids.data_ptr<scalar_t>(),
               sorted_token_ids.data_ptr<int32_t>(),
               experts_ids.data_ptr<int32_t>(),
-              num_tokens_post_pad.data_ptr<int32_t>(), num_experts,
-              padded_num_experts, experts_per_warp, block_size,
-              topk_ids.numel(), cumsum_buffer.data_ptr<int32_t>());
+              num_tokens_post_pad.data_ptr<int32_t>(), num_experts, block_size,
+              topk_ids.numel());
+        });
+  }
+}
 
-          const int block_threads = std::min(256, (int)threads);
-          const int num_blocks =
-              (topk_ids.numel() + block_threads - 1) / block_threads;
-          const int max_blocks = 65535;
-          const int actual_blocks = std::min(num_blocks, max_blocks);
+void sgl_moe_align_block_size(torch::Tensor topk_ids, int64_t num_experts,
+                              int64_t block_size,
+                              torch::Tensor sorted_token_ids,
+                              torch::Tensor experts_ids,
+                              torch::Tensor num_tokens_post_pad) {
+  const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
+  TORCH_CHECK(num_experts == 256,
+              "sgl_moe_align_block_size kernel only supports deepseek v3.");
 
-          auto sort_kernel =
-              vllm::moe::count_and_sort_expert_tokens_kernel<scalar_t>;
-          sort_kernel<<<actual_blocks, block_threads, 0, stream>>>(
-              topk_ids.data_ptr<scalar_t>(),
-              sorted_token_ids.data_ptr<int32_t>(),
-              cumsum_buffer.data_ptr<int32_t>(), topk_ids.numel());
-        }
+  VLLM_DISPATCH_INTEGRAL_AND_UNSIGNED_TYPES(
+      topk_ids.scalar_type(), "sgl_moe_align_block_size_kernel", [&] {
+        // calc needed amount of shared mem for `cumsum` tensors
+        auto options_int =
+            torch::TensorOptions().dtype(torch::kInt).device(topk_ids.device());
+        torch::Tensor cumsum_buffer =
+            torch::zeros({num_experts + 1}, options_int);
+
+        auto align_kernel =
+            vllm::moe::sgl_moe_align_block_size_kernel<scalar_t>;
+        align_kernel<<<1, 1024, 0, stream>>>(
+            topk_ids.data_ptr<scalar_t>(), sorted_token_ids.data_ptr<int32_t>(),
+            experts_ids.data_ptr<int32_t>(),
+            num_tokens_post_pad.data_ptr<int32_t>(), num_experts, block_size,
+            topk_ids.numel(), cumsum_buffer.data_ptr<int32_t>());
+
+        const int block_threads = 256;
+        const int num_blocks =
+            (topk_ids.numel() + block_threads - 1) / block_threads;
+        const int max_blocks = 65535;
+        const int actual_blocks = std::min(num_blocks, max_blocks);
+        auto sort_kernel = vllm::moe::sgl_moe_token_sort_kernel<scalar_t>;
+        sort_kernel<<<actual_blocks, block_threads, 0, stream>>>(
+            topk_ids.data_ptr<scalar_t>(), sorted_token_ids.data_ptr<int32_t>(),
+            cumsum_buffer.data_ptr<int32_t>(), topk_ids.numel());
       });
 }
 

csrc/moe/moe_ops.h

@@ -12,6 +12,12 @@ void moe_align_block_size(torch::Tensor topk_ids, int64_t num_experts,
                           int64_t block_size, torch::Tensor sorted_token_ids,
                           torch::Tensor experts_ids,
                           torch::Tensor num_tokens_post_pad);
+
+void sgl_moe_align_block_size(torch::Tensor topk_ids, int64_t num_experts,
+                              int64_t block_size,
+                              torch::Tensor sorted_token_ids,
+                              torch::Tensor experts_ids,
+                              torch::Tensor num_tokens_post_pad);
 #ifndef USE_ROCM
 torch::Tensor moe_wna16_gemm(torch::Tensor input, torch::Tensor output,
                              torch::Tensor b_qweight, torch::Tensor b_scales,
@@ -24,8 +30,4 @@ torch::Tensor moe_wna16_gemm(torch::Tensor input, torch::Tensor output,
                              int64_t BLOCK_SIZE_K, int64_t bit);
 #endif
 
-bool moe_permute_unpermute_supported();
-
-void shuffle_rows(const torch::Tensor& input_tensor,
-                  const torch::Tensor& dst2src_map,
-                  torch::Tensor& output_tensor);
+bool moe_permute_unpermute_supported();

csrc/moe/moe_permute_unpermute_op.cu

@@ -12,7 +12,7 @@ void moe_permute(
     const torch::Tensor& input,                      // [n_token, hidden]
     const torch::Tensor& topk_weights,               //[n_token, topk]
     torch::Tensor& topk_ids,                         // [n_token, topk]
-    const torch::Tensor& token_expert_indices,       // [n_token, topk]
+    const torch::Tensor& token_expert_indicies,      // [n_token, topk]
     const std::optional<torch::Tensor>& expert_map,  // [n_expert]
     int64_t n_expert, int64_t n_local_expert, int64_t topk,
     const std::optional<int64_t>& align_block_size,
@@ -27,15 +27,15 @@ void moe_permute(
               "expert_first_token_offset must be int64");
   TORCH_CHECK(topk_ids.scalar_type() == at::ScalarType::Int,
               "topk_ids must be int32");
-  TORCH_CHECK(token_expert_indices.scalar_type() == at::ScalarType::Int,
-              "token_expert_indices must be int32");
+  TORCH_CHECK(token_expert_indicies.scalar_type() == at::ScalarType::Int,
+              "token_expert_indicies must be int32");
   TORCH_CHECK(src_row_id2dst_row_id_map.scalar_type() == at::ScalarType::Int,
               "src_row_id2dst_row_id_map must be int32");
   TORCH_CHECK(expert_first_token_offset.size(0) == n_local_expert + 1,
               "expert_first_token_offset shape != n_local_expert+1")
   TORCH_CHECK(
-      src_row_id2dst_row_id_map.sizes() == token_expert_indices.sizes(),
-      "token_expert_indices shape must be same as src_row_id2dst_row_id_map");
+      src_row_id2dst_row_id_map.sizes() == token_expert_indicies.sizes(),
+      "token_expert_indicies shape must be same as src_row_id2dst_row_id_map");
   auto n_token = input.sizes()[0];
   auto n_hidden = input.sizes()[1];
   auto align_block_size_value =
@@ -71,7 +71,7 @@ void moe_permute(
                              expert_map_ptr, n_expert, stream);
   }
   // expert sort topk expert id and scan expert id get expert_first_token_offset
-  sortAndScanExpert(get_ptr<int>(topk_ids), get_ptr<int>(token_expert_indices),
+  sortAndScanExpert(get_ptr<int>(topk_ids), get_ptr<int>(token_expert_indicies),
                     get_ptr<int>(permuted_experts_id),
                     get_ptr<int>(dst_row_id2src_row_id_map),
                     get_ptr<int64_t>(expert_first_token_offset), n_token,
@@ -130,67 +130,11 @@ void moe_unpermute(
   });
 }
 
-template <typename T>
-__global__ void shuffleInputRowsKernel(const T* input,
-                                       const int32_t* dst2src_map, T* output,
-                                       int64_t num_src_rows,
-                                       int64_t num_dst_rows, int64_t num_cols) {
-  int64_t dest_row_idx = blockIdx.x;
-  int64_t const source_row_idx = dst2src_map[dest_row_idx];
-
-  if (blockIdx.x < num_dst_rows) {
-    // Load 128-bits per thread
-    constexpr int64_t ELEM_PER_THREAD = 128 / sizeof(T) / 8;
-    using DataElem = cutlass::Array<T, ELEM_PER_THREAD>;
-
-    // Duplicate and permute rows
-    auto const* source_row_ptr =
-        reinterpret_cast<DataElem const*>(input + source_row_idx * num_cols);
-    auto* dest_row_ptr =
-        reinterpret_cast<DataElem*>(output + dest_row_idx * num_cols);
-
-    int64_t const start_offset = threadIdx.x;
-    int64_t const stride = blockDim.x;
-    int64_t const num_elems_in_col = num_cols / ELEM_PER_THREAD;
-
-    for (int elem_index = start_offset; elem_index < num_elems_in_col;
-         elem_index += stride) {
-      dest_row_ptr[elem_index] = source_row_ptr[elem_index];
-    }
-  }
-}
-
-void shuffle_rows(const torch::Tensor& input_tensor,
-                  const torch::Tensor& dst2src_map,
-                  torch::Tensor& output_tensor) {
-  TORCH_CHECK(input_tensor.scalar_type() == output_tensor.scalar_type(),
-              "Input and output tensors must have the same data type");
-
-  auto stream = at::cuda::getCurrentCUDAStream().stream();
-  int64_t const blocks = output_tensor.size(0);
-  int64_t const threads = 256;
-  int64_t const num_dest_rows = output_tensor.size(0);
-  int64_t const num_src_rows = input_tensor.size(0);
-  int64_t const num_cols = input_tensor.size(1);
-
-  TORCH_CHECK(!(num_cols % (128 / sizeof(input_tensor.scalar_type()) / 8)),
-              "num_cols must be divisible by 128 / "
-              "sizeof(input_tensor.scalar_type()) / 8");
-
-  MOE_DISPATCH(input_tensor.scalar_type(), [&] {
-    shuffleInputRowsKernel<scalar_t><<<blocks, threads, 0, stream>>>(
-        reinterpret_cast<scalar_t*>(input_tensor.data_ptr()),
-        dst2src_map.data_ptr<int32_t>(),
-        reinterpret_cast<scalar_t*>(output_tensor.data_ptr()), num_src_rows,
-        num_dest_rows, num_cols);
-  });
-}
-
 #else
 
 void moe_permute(const torch::Tensor& input, const torch::Tensor& topk_weights,
                  torch::Tensor& topk_ids,
-                 const torch::Tensor& token_expert_indices,
+                 const torch::Tensor& token_expert_indicies,
                  const std::optional<torch::Tensor>& expert_map,
                  int64_t n_expert, int64_t n_local_expert, int64_t topk,
                  const std::optional<int64_t>& align_block_size,
@@ -203,7 +147,7 @@ void moe_permute(const torch::Tensor& input, const torch::Tensor& topk_weights,
 
 void moe_unpermute(const torch::Tensor& input,
                    const torch::Tensor& topk_weights, torch::Tensor& topk_ids,
-                   const torch::Tensor& token_expert_indices,
+                   const torch::Tensor& token_expert_indicies,
                    const std::optional<torch::Tensor>& expert_map,
                    int64_t n_expert, int64_t n_local_expert, int64_t topk,
                    const std::optional<int64_t>& align_block_size,

csrc/moe/permute_unpermute_kernels/dispatch.h

@@ -14,13 +14,12 @@
     __VA_ARGS__();                                         \
     break;                                                 \
   }
-#define MOE_DISPATCH_FLOAT_CASE(...)                            \
-  MOE_DISPATCH_CASE(at::ScalarType::Float, __VA_ARGS__)         \
-  MOE_DISPATCH_CASE(at::ScalarType::Half, __VA_ARGS__)          \
-  MOE_DISPATCH_CASE(at::ScalarType::BFloat16, __VA_ARGS__)      \
-  MOE_DISPATCH_CASE(at::ScalarType::Float8_e5m2, __VA_ARGS__)   \
-  MOE_DISPATCH_CASE(at::ScalarType::Float8_e4m3fn, __VA_ARGS__) \
-  MOE_DISPATCH_CASE(at::ScalarType::Byte, __VA_ARGS__)
+#define MOE_DISPATCH_FLOAT_CASE(...)                          \
+  MOE_DISPATCH_CASE(at::ScalarType::Float, __VA_ARGS__)       \
+  MOE_DISPATCH_CASE(at::ScalarType::Half, __VA_ARGS__)        \
+  MOE_DISPATCH_CASE(at::ScalarType::BFloat16, __VA_ARGS__)    \
+  MOE_DISPATCH_CASE(at::ScalarType::Float8_e5m2, __VA_ARGS__) \
+  MOE_DISPATCH_CASE(at::ScalarType::Float8_e4m3fn, __VA_ARGS__)
 
 #define MOE_DISPATCH(TYPE, ...) \
   MOE_SWITCH(TYPE, MOE_DISPATCH_FLOAT_CASE(__VA_ARGS__))
@@ -40,11 +39,6 @@ template <>
 struct ScalarType2CudaType<at::ScalarType::BFloat16> {
   using type = __nv_bfloat16;
 };
-// uint8 for packed fp4
-template <>
-struct ScalarType2CudaType<at::ScalarType::Byte> {
-  using type = uint8_t;
-};
 
 // #if __CUDA_ARCH__ >= 890
 // fp8

csrc/moe/permute_unpermute_kernels/moe_permute_unpermute_kernel.inl

@@ -20,6 +20,7 @@ __global__ void expandInputRowsKernel(
   int expert_id = sorted_experts[expanded_dest_row];
 
   extern __shared__ int64_t smem_expert_first_token_offset[];
+  int64_t align_expanded_row_accumulate = 0;
   if constexpr (ALIGN_BLOCK_SIZE) {
     // load g2s
     for (int idx = threadIdx.x; idx < num_local_experts + 1;
@@ -62,6 +63,7 @@ __global__ void expandInputRowsKernel(
     using DataElem = cutlass::Array<T, ELEM_PER_THREAD>;
 
     // Duplicate and permute rows
+    int64_t const source_k_rank = expanded_source_row / num_rows;
     int64_t const source_row = expanded_source_row % num_rows;
 
     auto const* source_row_ptr =
@@ -158,6 +160,7 @@ __global__ void finalizeMoeRoutingKernel(
        elem_index += stride) {
     ComputeElem thread_output;
     thread_output.fill(0);
+    float row_rescale{0.f};
     for (int k_idx = 0; k_idx < k; ++k_idx) {
       int64_t const expanded_original_row = original_row + k_idx * num_rows;
       int64_t const expanded_permuted_row =
@@ -174,6 +177,8 @@ __global__ void finalizeMoeRoutingKernel(
       auto const* expanded_permuted_rows_row_ptr =
           expanded_permuted_rows_v + expanded_permuted_row * num_elems_in_col;
 
+      int64_t const expert_idx = expert_for_source_row[k_offset];
+
       ComputeElem expert_result = arrayConvert<InputElem, ComputeElem>(
           expanded_permuted_rows_row_ptr[elem_index]);
       thread_output = thread_output + row_scale * (expert_result);

csrc/moe/topk_softmax_kernels.cu

@@ -425,7 +425,7 @@ void topkGatingSoftmaxLauncherHelper(const float* input, const bool* finished, f
 
 #define LAUNCH_SOFTMAX(NUM_EXPERTS, WARPS_PER_TB)                       \
     topkGatingSoftmaxLauncherHelper<NUM_EXPERTS, WARPS_PER_TB>(         \
-        gating_output, nullptr, topk_weights, topk_indices,            \
+        gating_output, nullptr, topk_weights, topk_indicies,            \
         token_expert_indices, num_tokens, topk, 0, num_experts,         \
         stream);
 
@@ -433,7 +433,7 @@ template <typename IndType>
 void topkGatingSoftmaxKernelLauncher(
     const float* gating_output,
     float* topk_weights,
-    IndType* topk_indices,
+    IndType* topk_indicies,
     int* token_expert_indices,
     float* softmax_workspace,
     const int num_tokens,
@@ -476,7 +476,7 @@ void topkGatingSoftmaxKernelLauncher(
             moeSoftmax<TPB><<<num_tokens, TPB, 0, stream>>>(
                 gating_output, nullptr, softmax_workspace, num_experts);
             moeTopK<TPB><<<num_tokens, TPB, 0, stream>>>(
-                softmax_workspace, nullptr, topk_weights, topk_indices, token_expert_indices,
+                softmax_workspace, nullptr, topk_weights, topk_indicies, token_expert_indices,
                 num_experts, topk, 0, num_experts);
         }
     }
@@ -516,8 +516,9 @@ void topk_softmax(
             topk,
             stream);
     }
-    else if (topk_indices.scalar_type() == at::ScalarType::UInt32)
+    else
     {
+        assert(topk_indices.scalar_type() == at::ScalarType::UInt32);
         vllm::moe::topkGatingSoftmaxKernelLauncher(
             gating_output.data_ptr<float>(),
             topk_weights.data_ptr<float>(),
@@ -529,17 +530,4 @@ void topk_softmax(
             topk,
             stream);
     }
-    else {
-        assert(topk_indices.scalar_type() == at::ScalarType::Int64);
-        vllm::moe::topkGatingSoftmaxKernelLauncher(
-            gating_output.data_ptr<float>(),
-            topk_weights.data_ptr<float>(),
-            topk_indices.data_ptr<int64_t>(),
-            token_expert_indices.data_ptr<int>(),
-            softmax_workspace.data_ptr<float>(),
-            num_tokens,
-            num_experts,
-            topk,
-            stream);
-    }
 }

csrc/moe/torch_bindings.cpp

@@ -22,6 +22,15 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, m) {
       "                     Tensor! num_tokens_post_pad) -> ()");
   m.impl("moe_align_block_size", torch::kCUDA, &moe_align_block_size);
 
+  // temporarily adapted from
+  // https://github.com/sgl-project/sglang/commit/ded9fcd09a43d5e7d5bb31a2bc3e9fc21bf65d2a
+  m.def(
+      "sgl_moe_align_block_size(Tensor topk_ids, int num_experts,"
+      "                         int block_size, Tensor! sorted_token_ids,"
+      "                         Tensor! experts_ids,"
+      "                         Tensor! num_tokens_post_pad) -> ()");
+  m.impl("sgl_moe_align_block_size", torch::kCUDA, &sgl_moe_align_block_size);
+
 #ifndef USE_ROCM
   m.def(
       "moe_wna16_gemm(Tensor input, Tensor! output, Tensor b_qweight, "
@@ -57,7 +66,7 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, m) {
 
   m.def(
       "moe_permute(Tensor input, Tensor topk_weight, Tensor! topk_ids,"
-      "Tensor token_expert_indices, Tensor? expert_map, int n_expert,"
+      "Tensor token_expert_indicies, Tensor? expert_map, int n_expert,"
       "int n_local_expert,"
       "int topk, int? align_block_size,Tensor! permuted_input, Tensor! "
       "expert_first_token_offset, Tensor! src_row_id2dst_row_id_map, Tensor! "
@@ -72,12 +81,6 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, m) {
   m.def("moe_permute_unpermute_supported() -> bool");
   m.impl("moe_permute_unpermute_supported", &moe_permute_unpermute_supported);
 
-  // Row shuffle for MoE
-  m.def(
-      "shuffle_rows(Tensor input_tensor, Tensor dst2src_map, Tensor! "
-      "output_tensor) -> ()");
-  m.impl("shuffle_rows", torch::kCUDA, &shuffle_rows);
-
 #endif
 }
 

csrc/ops.h

@@ -92,11 +92,6 @@ void rms_norm(torch::Tensor& out, torch::Tensor& input, torch::Tensor& weight,
 void fused_add_rms_norm(torch::Tensor& input, torch::Tensor& residual,
                         torch::Tensor& weight, double epsilon);
 
-void apply_repetition_penalties_(torch::Tensor& logits,
-                                 const torch::Tensor& prompt_mask,
-                                 const torch::Tensor& output_mask,
-                                 const torch::Tensor& repetition_penalties);
-
 void rms_norm_static_fp8_quant(torch::Tensor& out, torch::Tensor& input,
                                torch::Tensor& weight, torch::Tensor& scale,
                                double epsilon);
@@ -236,8 +231,7 @@ void cutlass_moe_mm(
     torch::Tensor const& b_tensors, torch::Tensor const& a_scales,
     torch::Tensor const& b_scales, torch::Tensor const& expert_offsets,
     torch::Tensor const& problem_sizes, torch::Tensor const& a_strides,
-    torch::Tensor const& b_strides, torch::Tensor const& c_strides,
-    bool per_act_token, bool per_out_ch);
+    torch::Tensor const& b_strides, torch::Tensor const& c_strides);
 
 void cutlass_fp4_group_mm(
     torch::Tensor& output, const torch::Tensor& a, const torch::Tensor& b,
@@ -249,16 +243,7 @@ void get_cutlass_moe_mm_data(
     const torch::Tensor& topk_ids, torch::Tensor& expert_offsets,
     torch::Tensor& problem_sizes1, torch::Tensor& problem_sizes2,
     torch::Tensor& input_permutation, torch::Tensor& output_permutation,
-    const int64_t num_experts, const int64_t n, const int64_t k,
-    const std::optional<torch::Tensor>& blockscale_offsets);
-
-void get_cutlass_pplx_moe_mm_data(torch::Tensor& expert_offsets,
-                                  torch::Tensor& problem_sizes1,
-                                  torch::Tensor& problem_sizes2,
-                                  const torch::Tensor& expert_num_tokens,
-                                  const int64_t num_local_experts,
-                                  const int64_t padded_m, const int64_t n,
-                                  const int64_t k);
+    const int64_t num_experts, const int64_t n, const int64_t k);
 
 void cutlass_scaled_mm_azp(torch::Tensor& out, torch::Tensor const& a,
                            torch::Tensor const& b,

csrc/prepare_inputs/advance_step.cu

@@ -274,6 +274,7 @@ void advance_step_flashinfer(
   cudaDeviceGetAttribute(&blocks, cudaDevAttrMultiProcessorCount, dev);
   cudaDeviceGetAttribute(&threads, cudaDevAttrMaxThreadsPerBlock, dev);
 
+  [[maybe_unused]] int block_tables_stride = block_tables.stride(0);
   TORCH_CHECK((blocks * threads > num_queries),
               "multi-step: not enough threads to map to num_queries = ",
               num_queries, " block_tables.stride(0) = ", block_tables.stride(0),