[Misc] Improve error message for incorrect pynvml (#12809)

Signed-off-by: youkaichao <youkaichao@gmail.com>

.buildkite/check-wheel-size.py

@@ -1,9 +1,14 @@
+# SPDX-License-Identifier: Apache-2.0
+
 import os
 import sys
 import zipfile
 
-# Read the VLLM_MAX_SIZE_MB environment variable, defaulting to 250 MB
+# Read the VLLM_MAX_SIZE_MB environment variable, defaulting to 400 MiB
-VLLM_MAX_SIZE_MB = int(os.environ.get('VLLM_MAX_SIZE_MB', 250))
+# Note that we have 400 MiB quota, please use it wisely.
+# See https://github.com/pypi/support/issues/3792 .
+# Please also sync the value with the one in Dockerfile.
+VLLM_MAX_SIZE_MB = int(os.environ.get('VLLM_MAX_SIZE_MB', 400))
 
 
 def print_top_10_largest_files(zip_file):

.buildkite/generate_index.py

@@ -1,3 +1,5 @@
+# SPDX-License-Identifier: Apache-2.0
+
 import argparse
 import os
 

.buildkite/lm-eval-harness/configs/SparseLlama3.1_2of4_fp8_compressed.yaml

@@ -0,0 +1,11 @@
+# bash ./run-lm-eval-gsm-vllm-baseline.sh -m nm-testing/SparseLlama-3.1-8B-gsm8k-pruned.2of4-chnl_wts_per_tok_dyn_act_fp8-BitM -b "auto" -t 2
+model_name: "nm-testing/SparseLlama-3.1-8B-gsm8k-pruned.2of4-chnl_wts_per_tok_dyn_act_fp8-BitM"
+tasks:
+- name: "gsm8k"
+  metrics:
+  - name: "exact_match,strict-match"
+    value: 0.6353
+  - name: "exact_match,flexible-extract"
+    value: 0.637
+limit: null
+num_fewshot: null 

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

@@ -1,3 +1,4 @@
+# SPDX-License-Identifier: Apache-2.0
 """
 LM eval harness on model to compare vs HF baseline computed offline.
 Configs are found in configs/$MODEL.yaml

.buildkite/nightly-benchmarks/benchmark-pipeline.yaml

@@ -1,5 +1,6 @@
 steps:
   - label: "Wait for container to be ready"
+    key: wait-for-container-image
     agents:
       queue: A100
     plugins:
@@ -10,12 +11,11 @@ steps:
             command:
             - sh .buildkite/nightly-benchmarks/scripts/wait-for-image.sh
 
-  - wait
-
   - label: "A100"
     # skip: "use this flag to conditionally skip the benchmark step, useful for PR testing"
     agents:
       queue: A100
+    depends_on: wait-for-container-image
     plugins:
     - kubernetes:
         podSpec:
@@ -49,6 +49,7 @@ steps:
     # skip: "use this flag to conditionally skip the benchmark step, useful for PR testing"
     agents:
       queue: H200
+    depends_on: wait-for-container-image
     plugins:
     - docker#v5.12.0:
         image: public.ecr.aws/q9t5s3a7/vllm-ci-postmerge-repo:$BUILDKITE_COMMIT
@@ -73,7 +74,7 @@ steps:
     # skip: "use this flag to conditionally skip the benchmark step, useful for PR testing"
     agents:
       queue: H100
-    depends_on: block-h100
+    depends_on: wait-for-container-image
     plugins:
     - docker#v5.12.0:
         image: public.ecr.aws/q9t5s3a7/vllm-ci-postmerge-repo:$BUILDKITE_COMMIT

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

@@ -1,3 +1,5 @@
+# SPDX-License-Identifier: Apache-2.0
+
 import json
 import os
 from pathlib import Path

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

@@ -1,3 +1,5 @@
+# SPDX-License-Identifier: Apache-2.0
+
 import argparse
 
 from transformers import AutoTokenizer

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

@@ -1,3 +1,5 @@
+# SPDX-License-Identifier: Apache-2.0
+
 import argparse
 import json
 from pathlib import Path

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

@@ -1,3 +1,5 @@
+# SPDX-License-Identifier: Apache-2.0
+
 from lmdeploy.serve.openai.api_client import APIClient
 
 api_client = APIClient("http://localhost:8000")

.buildkite/nightly-benchmarks/scripts/nightly-annotate.sh

@@ -43,7 +43,7 @@ main() {
     
 
 
-    # The figures should be genereated by a separate process outside the CI/CD pipeline
+    # The figures should be generated by a separate process outside the CI/CD pipeline
 
     # # generate figures
     # python3 -m pip install tabulate pandas matplotlib

.buildkite/nightly-benchmarks/scripts/run-nightly-benchmarks.sh

@@ -301,6 +301,104 @@ run_serving_tests() {
   kill_gpu_processes
 }
 
+run_genai_perf_tests() {
+  # run genai-perf tests 
+
+  # $1: a json file specifying genai-perf test cases
+  local genai_perf_test_file
+  genai_perf_test_file=$1
+
+  # Iterate over genai-perf tests
+  jq -c '.[]' "$genai_perf_test_file" | while read -r params; do
+    # get the test name, and append the GPU type back to it.
+    test_name=$(echo "$params" | jq -r '.test_name')    
+    
+    # if TEST_SELECTOR is set, only run the test cases that match the selector
+    if [[ -n "$TEST_SELECTOR" ]] && [[ ! "$test_name" =~ $TEST_SELECTOR ]]; then
+      echo "Skip test case $test_name."
+      continue
+    fi
+    
+    # prepend the current serving engine to the test name
+    test_name=${CURRENT_LLM_SERVING_ENGINE}_${test_name}
+
+    # get common parameters
+    common_params=$(echo "$params" | jq -r '.common_parameters')
+    model=$(echo "$common_params" | jq -r '.model')
+    tp=$(echo "$common_params" | jq -r '.tp')
+    dataset_name=$(echo "$common_params" | jq -r '.dataset_name')
+    dataset_path=$(echo "$common_params" | jq -r '.dataset_path')
+    port=$(echo "$common_params" | jq -r '.port')
+    num_prompts=$(echo "$common_params" | jq -r '.num_prompts')
+    reuse_server=$(echo "$common_params" | jq -r '.reuse_server')
+
+    # get client and server arguments
+    server_params=$(echo "$params" | jq -r ".${CURRENT_LLM_SERVING_ENGINE}_server_parameters")
+    qps_list=$(echo "$params" | jq -r '.qps_list')
+    qps_list=$(echo "$qps_list" | jq -r '.[] | @sh')
+    echo "Running over qps list $qps_list"
+
+    # check if there is enough GPU to run the test
+    if [[ $gpu_count -lt $tp ]]; then
+      echo "Required num-shard $tp but only $gpu_count GPU found. Skip testcase $test_name."
+      continue
+    fi
+
+    if [[ $reuse_server == "true" ]]; then
+      echo "Reuse previous server for test case $test_name"
+    else
+      kill_gpu_processes
+      bash "$VLLM_SOURCE_CODE_LOC/.buildkite/nightly-benchmarks/scripts/launch-server.sh" \
+        "$server_params" "$common_params"
+    fi
+
+    if wait_for_server; then
+      echo ""
+      echo "$CURRENT_LLM_SERVING_ENGINE server is up and running."
+    else
+      echo ""
+      echo "$CURRENT_LLM_SERVING_ENGINE failed to start within the timeout period."
+      break
+    fi
+
+    # iterate over different QPS
+    for qps in $qps_list; do
+      # remove the surrounding single quote from qps
+      if [[ "$qps" == *"inf"* ]]; then
+        echo "qps was $qps"
+        qps=$num_prompts
+        echo "now qps is $qps"
+      fi
+    
+      new_test_name=$test_name"_qps_"$qps
+      backend=$CURRENT_LLM_SERVING_ENGINE
+      
+      if [[ "$backend" == *"vllm"* ]]; then
+        backend="vllm"
+      fi
+      #TODO: add output dir.
+      client_command="genai-perf profile \
+        -m $model \
+        --service-kind openai \
+        --backend vllm \
+        --endpoint-type chat \
+        --streaming \
+        --url localhost:$port \
+        --request-rate $qps \
+        --num-prompts $num_prompts \
+      "
+
+    echo "Client command: $client_command"
+
+    eval "$client_command"
+
+    #TODO: process/record outputs
+    done
+  done
+
+  kill_gpu_processes
+
+}
 
 prepare_dataset() {
 
@@ -328,12 +426,17 @@ main() {
 
   pip install -U transformers
 
+  pip install -r requirements-dev.txt
+  which genai-perf
+
   # check storage
   df -h
 
   ensure_installed wget
   ensure_installed curl
   ensure_installed jq
+  # genai-perf dependency
+  ensure_installed libb64-0d
 
   prepare_dataset
 
@@ -345,6 +448,10 @@ main() {
   # run the test
   run_serving_tests "$BENCHMARK_ROOT/tests/nightly-tests.json"
 
+  # run genai-perf tests
+  run_genai_perf_tests "$BENCHMARK_ROOT/tests/genai-perf-tests.json"
+  mv artifacts/ $RESULTS_FOLDER/
+
   # upload benchmark results to buildkite
   python3 -m pip install tabulate pandas
   python3 "$BENCHMARK_ROOT/scripts/summary-nightly-results.py"

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

@@ -1,3 +1,5 @@
+# SPDX-License-Identifier: Apache-2.0
+
 import datetime
 import json
 import os

.buildkite/nightly-benchmarks/tests/genai-perf-tests.json

@@ -0,0 +1,23 @@
+[
+    {
+        "test_name": "llama8B_tp1_genai_perf",
+        "qps_list": [4,8,16,32],
+        "common_parameters": {
+            "model": "meta-llama/Meta-Llama-3-8B-Instruct",
+            "tp": 1,
+            "port": 8000,
+            "num_prompts": 500,
+            "reuse_server": false
+        },
+        "vllm_server_parameters": {
+            "disable_log_stats": "",
+            "disable_log_requests": "",
+            "gpu_memory_utilization": 0.9,
+            "num_scheduler_steps": 10,
+            "max_num_seqs": 512,
+            "dtype": "bfloat16"
+        },
+        "genai_perf_input_parameters": {
+        }
+    }
+]

.buildkite/release-pipeline.yaml

@@ -56,6 +56,11 @@ steps:
     env:
       DOCKER_BUILDKIT: "1"
 
+  - input: "Provide Release version here"
+    fields:
+      - text: "What is the release version?"
+        key: "release-version"
+
   - block: "Build CPU release image"
     key: block-cpu-release-image-build
     depends_on: ~
@@ -66,7 +71,7 @@ steps:
       queue: cpu_queue_postmerge
     commands:
       - "aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin public.ecr.aws/q9t5s3a7"
-      - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg GIT_REPO_CHECK=1 --tag public.ecr.aws/q9t5s3a7/vllm-cpu-release-repo:$RELEASE_VERSION --progress plain -f Dockerfile.cpu ."
+      - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg GIT_REPO_CHECK=1 --tag public.ecr.aws/q9t5s3a7/vllm-cpu-release-repo:$(buildkite-agent meta-data get release-version) --progress plain -f Dockerfile.cpu ."
-      - "docker push public.ecr.aws/q9t5s3a7/vllm-cpu-release-repo:$RELEASE_VERSION"
+      - "docker push public.ecr.aws/q9t5s3a7/vllm-cpu-release-repo:$(buildkite-agent meta-data get release-version)"
     env:
       DOCKER_BUILDKIT: "1"

.buildkite/run-cpu-test.sh

@@ -9,36 +9,33 @@ CORE_RANGE=${CORE_RANGE:-48-95}
 NUMA_NODE=${NUMA_NODE:-1}
 
 # Try building the docker image
-numactl -C "$CORE_RANGE" -N "$NUMA_NODE" docker build -t cpu-test -f Dockerfile.cpu .
+numactl -C "$CORE_RANGE" -N "$NUMA_NODE" docker build -t cpu-test-"$BUILDKITE_BUILD_NUMBER" -f Dockerfile.cpu .
-numactl -C "$CORE_RANGE" -N "$NUMA_NODE" docker build --build-arg VLLM_CPU_DISABLE_AVX512="true" -t cpu-test-avx2 -f Dockerfile.cpu .
+numactl -C "$CORE_RANGE" -N "$NUMA_NODE" docker build --build-arg VLLM_CPU_DISABLE_AVX512="true" -t cpu-test-"$BUILDKITE_BUILD_NUMBER"-avx2 -f Dockerfile.cpu .
 
 # Setup cleanup
-remove_docker_container() { docker rm -f cpu-test-"$NUMA_NODE" cpu-test-avx2-"$NUMA_NODE" || true; }
+remove_docker_container() { set -e; docker rm -f cpu-test-"$BUILDKITE_BUILD_NUMBER"-"$NUMA_NODE" cpu-test-"$BUILDKITE_BUILD_NUMBER"-avx2-"$NUMA_NODE" || true; }
 trap remove_docker_container EXIT
 remove_docker_container
 
 # Run the image, setting --shm-size=4g for tensor parallel.
 docker run -itd --entrypoint /bin/bash -v ~/.cache/huggingface:/root/.cache/huggingface --cpuset-cpus="$CORE_RANGE"  \
- --cpuset-mems="$NUMA_NODE" --privileged=true --network host -e HF_TOKEN --env VLLM_CPU_KVCACHE_SPACE=4 --shm-size=4g --name cpu-test-"$NUMA_NODE" cpu-test
+ --cpuset-mems="$NUMA_NODE" --privileged=true --network host -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="$NUMA_NODE" --privileged=true --network host -e HF_TOKEN --env VLLM_CPU_KVCACHE_SPACE=4 --shm-size=4g --name cpu-test-avx2-"$NUMA_NODE" cpu-test-avx2
+ --cpuset-mems="$NUMA_NODE" --privileged=true --network host -e HF_TOKEN --env VLLM_CPU_KVCACHE_SPACE=4 --shm-size=4g --name cpu-test-"$BUILDKITE_BUILD_NUMBER"-avx2-"$NUMA_NODE" cpu-test-"$BUILDKITE_BUILD_NUMBER"-avx2
 
 function cpu_tests() {
   set -e
   export NUMA_NODE=$2
 
   # offline inference
-  docker exec cpu-test-avx2-"$NUMA_NODE" bash -c "
+  docker exec cpu-test-"$BUILDKITE_BUILD_NUMBER"-avx2-"$NUMA_NODE" bash -c "
     set -e
-    python3 examples/offline_inference.py"
+    python3 examples/offline_inference/basic.py"
 
   # 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
-    pip install pytest pytest-asyncio \
+    pip install -r vllm/requirements-test.txt
-      decord einops librosa peft Pillow sentence-transformers soundfile \
-      transformers_stream_generator matplotlib datamodel_code_generator
-    pip install torchvision --index-url https://download.pytorch.org/whl/cpu
     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
@@ -46,26 +43,26 @@ function cpu_tests() {
     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
     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 inference
+  # 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=$1
@@ -78,8 +75,14 @@ function cpu_tests() {
       --num-prompts 20 \
       --endpoint /v1/completions \
       --tokenizer facebook/opt-125m"
+
+  # Run multi-lora tests
+  docker exec cpu-test-"$BUILDKITE_BUILD_NUMBER"-"$NUMA_NODE" bash -c "
+    set -e
+    pytest -s -v \
+    tests/lora/test_qwen2vl.py"
 }
 
-# All of CPU tests are expected to be finished less than 25 mins.
+# All of CPU tests are expected to be finished less than 40 mins.
 export -f cpu_tests
-timeout 30m bash -c "cpu_tests $CORE_RANGE $NUMA_NODE"
+timeout 40m bash -c "cpu_tests $CORE_RANGE $NUMA_NODE"

.buildkite/run-gh200-test.sh

@@ -23,6 +23,6 @@ trap remove_docker_container EXIT
 remove_docker_container
 
 # Run the image and test offline inference
-docker run --name gh200-test --gpus=all --entrypoint="" gh200-test bash -c '
+docker run -e HF_TOKEN -v /root/.cache/huggingface:/root/.cache/huggingface --name gh200-test --gpus=all --entrypoint="" gh200-test bash -c '
-    python3 examples/offline_inference.py
+    python3 examples/offline_inference/cli.py --model meta-llama/Llama-3.2-1B
 '

.buildkite/run-hpu-test.sh

@@ -8,9 +8,17 @@ set -ex
 docker build -t hpu-test-env -f Dockerfile.hpu .
 
 # Setup cleanup
+# certain versions of HPU software stack have a bug that can
+# override the exit code of the script, so we need to use
+# separate remove_docker_container and remove_docker_container_and_exit
+# functions, while other platforms only need one remove_docker_container
+# function.
+EXITCODE=1
 remove_docker_container() { docker rm -f hpu-test || true; }
-trap remove_docker_container EXIT
+remove_docker_container_and_exit() { remove_docker_container; exit $EXITCODE; }
+trap remove_docker_container_and_exit EXIT
 remove_docker_container
 
 # Run the image and launch offline inference
-docker run --runtime=habana --name=hpu-test --network=host -e HABANA_VISIBLE_DEVICES=all -e VLLM_SKIP_WARMUP=true --entrypoint="" hpu-test-env python3 examples/offline_inference.py
+docker run --runtime=habana --name=hpu-test --network=host -e HABANA_VISIBLE_DEVICES=all -e VLLM_SKIP_WARMUP=true --entrypoint="" hpu-test-env python3 examples/offline_inference/basic.py
+EXITCODE=$?

.buildkite/run-neuron-test.sh

@@ -3,6 +3,18 @@
 # This script build the Neuron docker image and run the API server inside the container.
 # It serves a sanity check for compilation and basic model usage.
 set -e
+set -v
+
+image_name="neuron/vllm-ci"
+container_name="neuron_$(tr -dc A-Za-z0-9 < /dev/urandom | head -c 10; echo)"
+
+HF_CACHE="$(realpath ~)/huggingface"
+mkdir -p "${HF_CACHE}"
+HF_MOUNT="/root/.cache/huggingface"
+
+NEURON_COMPILE_CACHE_URL="$(realpath ~)/neuron_compile_cache"
+mkdir -p "${NEURON_COMPILE_CACHE_URL}"
+NEURON_COMPILE_CACHE_MOUNT="/root/.cache/neuron_compile_cache"
 
 # Try building the docker image
 aws ecr get-login-password --region us-west-2 | docker login --username AWS --password-stdin 763104351884.dkr.ecr.us-west-2.amazonaws.com
@@ -13,41 +25,33 @@ if [ -f /tmp/neuron-docker-build-timestamp ]; then
     last_build=$(cat /tmp/neuron-docker-build-timestamp)
     current_time=$(date +%s)
     if [ $((current_time - last_build)) -gt 86400 ]; then
-        docker system prune -f
+        # Remove dangling images (those that are not tagged and not used by any container)
+        docker image prune -f
+        # Remove unused volumes / force the system prune for old images as well.
+        docker volume prune -f && docker system prune -f
+        # Remove huggingface model artifacts and compiler cache
+        rm -rf "${HF_MOUNT:?}/*"
+        rm -rf "${NEURON_COMPILE_CACHE_MOUNT:?}/*"
         echo "$current_time" > /tmp/neuron-docker-build-timestamp
     fi
 else
     date "+%s" > /tmp/neuron-docker-build-timestamp
 fi
 
-docker build -t neuron -f Dockerfile.neuron .
+docker build -t "${image_name}" -f Dockerfile.neuron .
 
 # Setup cleanup
-remove_docker_container() { docker rm -f neuron || true; }
+remove_docker_container() {
+    docker image rm -f "${image_name}" || true;
+}
 trap remove_docker_container EXIT
-remove_docker_container
 
 # Run the image
-docker run --device=/dev/neuron0 --device=/dev/neuron1 --network host --name neuron neuron python3 -m vllm.entrypoints.api_server \
+docker run --rm -it --device=/dev/neuron0 --device=/dev/neuron1 --network host \
-       --model TinyLlama/TinyLlama-1.1B-Chat-v1.0 --max-num-seqs 8 --max-model-len 128 --block-size 128 --device neuron --tensor-parallel-size 2 &
+       -v "${HF_CACHE}:${HF_MOUNT}" \
-
+       -e "HF_HOME=${HF_MOUNT}" \
-# Wait for the server to start
+       -v "${NEURON_COMPILE_CACHE_URL}:${NEURON_COMPILE_CACHE_MOUNT}" \
-wait_for_server_to_start() {
+       -e "NEURON_COMPILE_CACHE_URL=${NEURON_COMPILE_CACHE_MOUNT}" \
-    timeout=300
+       --name "${container_name}" \
-    counter=0
+       ${image_name} \
-
+       /bin/bash -c "python3 /workspace/vllm/examples/offline_inference/neuron.py && python3 -m pytest /workspace/vllm/tests/neuron/ -v --capture=tee-sys"
-    while [ "$(curl -s -o /dev/null -w '%{http_code}' localhost:8000/health)" != "200" ]; do
-        sleep 1
-        counter=$((counter + 1))
-        if [ $counter -ge $timeout ]; then
-            echo "Timeout after $timeout seconds"
-            break
-        fi
-    done
-}
-wait_for_server_to_start
-
-# Test a simple prompt
-curl -X POST -H "Content-Type: application/json" \
-    localhost:8000/generate \
-    -d '{"prompt": "San Francisco is a"}'

.buildkite/run-openvino-test.sh

@@ -13,4 +13,4 @@ trap remove_docker_container EXIT
 remove_docker_container
 
 # Run the image and launch offline inference
-docker run --network host --env VLLM_OPENVINO_KVCACHE_SPACE=1 --name openvino-test openvino-test python3 /workspace/examples/offline_inference.py
+docker run --network host --env VLLM_OPENVINO_KVCACHE_SPACE=1 --name openvino-test openvino-test python3 /workspace/examples/offline_inference/basic.py

.buildkite/run-tpu-test.sh

@@ -14,4 +14,13 @@ 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=$HF_TOKEN" --name tpu-test vllm-tpu /bin/bash -c "python3 -m pip install git+https://github.com/thuml/depyf.git && python3 -m pip install pytest && python3 -m pip install lm_eval[api]==0.4.4 && pytest -v -s /workspace/vllm/tests/entrypoints/openai/test_accuracy.py && pytest -v -s /workspace/vllm/tests/tpu/test_custom_dispatcher.py && python3 /workspace/vllm/tests/tpu/test_compilation.py && python3 /workspace/vllm/examples/offline_inference_tpu.py"
+docker run --privileged --net host --shm-size=16G -it \
+    -e "HF_TOKEN=$HF_TOKEN" --name tpu-test \
+    vllm-tpu /bin/bash -c "python3 -m pip install git+https://github.com/thuml/depyf.git \
+    && python3 -m pip install pytest \
+    && python3 -m pip install lm_eval[api]==0.4.4 \
+    && pytest -v -s /workspace/vllm/tests/entrypoints/openai/test_accuracy.py \
+    && pytest -v -s /workspace/vllm/tests/tpu/test_custom_dispatcher.py \
+    && python3 /workspace/vllm/tests/tpu/test_compilation.py \
+    && python3 /workspace/vllm/tests/tpu/test_quantization_accuracy.py \
+    && python3 /workspace/vllm/examples/offline_inference/tpu.py"

.buildkite/run-xpu-test.sh

@@ -14,6 +14,6 @@ remove_docker_container
 
 # Run the image and test offline inference/tensor parallel
 docker run --name xpu-test --device /dev/dri -v /dev/dri/by-path:/dev/dri/by-path --entrypoint="" xpu-test sh -c '
-    python3 examples/offline_inference.py
+    python3 examples/offline_inference/basic.py
-    python3 examples/offline_inference_cli.py -tp 2
+    python3 examples/offline_inference/cli.py -tp 2
 '

.buildkite/test-pipeline.yaml

@@ -38,7 +38,7 @@ steps:
   - pip install -r requirements-docs.txt
   - SPHINXOPTS=\"-W\" make html
   # Check API reference (if it fails, you may have missing mock imports)
-  - grep \"sig sig-object py\" build/html/dev/sampling_params.html
+  - grep \"sig sig-object py\" build/html/api/inference_params.html
 
 - label: Async Engine, Inputs, Utils, Worker Test # 24min
   fast_check: true
@@ -50,9 +50,9 @@ steps:
   - tests/multimodal
   - tests/test_utils
   - tests/worker
-  - tests/standalone_tests/lazy_torch_compile.py
+  - tests/standalone_tests/lazy_imports.py
   commands:
-  - python3 standalone_tests/lazy_torch_compile.py
+  - python3 standalone_tests/lazy_imports.py
   - pytest -v -s mq_llm_engine # MQLLMEngine
   - pytest -v -s async_engine # AsyncLLMEngine
   - NUM_SCHEDULER_STEPS=4 pytest -v -s async_engine/test_async_llm_engine.py
@@ -76,7 +76,9 @@ steps:
   - tests/basic_correctness/test_basic_correctness
   - tests/basic_correctness/test_cpu_offload
   - tests/basic_correctness/test_preemption
+  - tests/basic_correctness/test_cumem.py
   commands:
+  - pytest -v -s basic_correctness/test_cumem.py
   - pytest -v -s basic_correctness/test_basic_correctness.py
   - pytest -v -s basic_correctness/test_cpu_offload.py
   - VLLM_TEST_ENABLE_ARTIFICIAL_PREEMPT=1 pytest -v -s basic_correctness/test_preemption.py
@@ -106,14 +108,12 @@ steps:
   source_file_dependencies:
   - vllm/
   commands:
-  - pip install -e ./plugins/vllm_add_dummy_model
+  - pytest -v -s entrypoints/llm --ignore=entrypoints/llm/test_lazy_outlines.py --ignore=entrypoints/llm/test_generate.py --ignore=entrypoints/llm/test_generate_multiple_loras.py --ignore=entrypoints/llm/test_guided_generate.py --ignore=entrypoints/llm/test_collective_rpc.py
-  - pytest -v -s entrypoints/llm --ignore=entrypoints/llm/test_lazy_outlines.py --ignore=entrypoints/llm/test_generate.py --ignore=entrypoints/llm/test_generate_multiple_loras.py --ignore=entrypoints/llm/test_guided_generate.py
   - pytest -v -s entrypoints/llm/test_lazy_outlines.py # it needs a clean process
   - pytest -v -s entrypoints/llm/test_generate.py # it needs a clean process
   - pytest -v -s entrypoints/llm/test_generate_multiple_loras.py # it needs a clean process
   - pytest -v -s entrypoints/llm/test_guided_generate.py # it needs a clean process
   - pytest -v -s entrypoints/openai --ignore=entrypoints/openai/test_oot_registration.py
-  - pytest -v -s entrypoints/openai/test_oot_registration.py # it needs a clean process
   - pytest -v -s entrypoints/test_chat_utils.py
   - pytest -v -s entrypoints/offline_mode # Needs to avoid interference with other tests
 
@@ -127,11 +127,17 @@ steps:
   - tests/distributed
   - tests/spec_decode/e2e/test_integration_dist_tp4
   - tests/compile
+  - examples/offline_inference/rlhf.py
+  - examples/offline_inference/ray_placement.py
   commands:
   - pytest -v -s distributed/test_utils.py
   - pytest -v -s compile/test_basic_correctness.py
   - pytest -v -s distributed/test_pynccl.py
   - pytest -v -s spec_decode/e2e/test_integration_dist_tp4.py
+  # TODO: create a dedicated test section for multi-GPU example tests
+  # when we have multiple distributed example tests
+  - python3 ../examples/offline_inference/rlhf.py
+  - RAY_DEDUP_LOGS=0 python3 ../examples/offline_inference/ray_placement.py
 
 - label: Metrics, Tracing Test # 10min
   num_gpus: 2 
@@ -179,7 +185,16 @@ steps:
     - vllm/
     - tests/v1
   commands:
-    - VLLM_USE_V1=1 pytest -v -s v1
+    # split the test to avoid interference
+    - VLLM_USE_V1=1 pytest -v -s v1/core
+    - VLLM_USE_V1=1 pytest -v -s v1/engine
+    - VLLM_USE_V1=1 pytest -v -s v1/sample
+    - VLLM_USE_V1=1 pytest -v -s v1/worker
+    - VLLM_USE_V1=1 pytest -v -s v1/test_stats.py
+    - VLLM_USE_V1=1 pytest -v -s v1/test_utils.py
+    # TODO: accuracy does not match, whether setting
+    # VLLM_USE_FLASHINFER_SAMPLER or not on H100.
+    - VLLM_USE_V1=1 pytest -v -s v1/e2e
 
 - label: Examples Test # 25min
   working_dir: "/vllm-workspace/examples"
@@ -189,19 +204,19 @@ steps:
   - examples/
   commands:
     - pip install tensorizer # for tensorizer test
-    - python3 offline_inference.py
+    - python3 offline_inference/basic.py
-    - python3 cpu_offload.py
+    - python3 offline_inference/cpu_offload.py
-    - python3 offline_inference_chat.py
+    - python3 offline_inference/chat.py
-    - python3 offline_inference_with_prefix.py
+    - python3 offline_inference/prefix_caching.py
-    - python3 llm_engine_example.py
+    - python3 offline_inference/llm_engine_example.py
-    - python3 offline_inference_vision_language.py
+    - python3 offline_inference/vision_language.py
-    - python3 offline_inference_vision_language_multi_image.py
+    - python3 offline_inference/vision_language_multi_image.py
-    - python3 tensorize_vllm_model.py --model facebook/opt-125m serialize --serialized-directory /tmp/ --suffix v1 && python3 tensorize_vllm_model.py --model facebook/opt-125m deserialize --path-to-tensors /tmp/vllm/facebook/opt-125m/v1/model.tensors
+    - python3 other/tensorize_vllm_model.py --model facebook/opt-125m serialize --serialized-directory /tmp/ --suffix v1 && python3 other/tensorize_vllm_model.py --model facebook/opt-125m deserialize --path-to-tensors /tmp/vllm/facebook/opt-125m/v1/model.tensors
-    - python3 offline_inference_encoder_decoder.py
+    - python3 offline_inference/encoder_decoder.py
-    - python3 offline_inference_classification.py
+    - python3 offline_inference/classification.py
-    - python3 offline_inference_embedding.py
+    - python3 offline_inference/embedding.py
-    - python3 offline_inference_scoring.py
+    - python3 offline_inference/scoring.py
-    - python3 offline_profile.py --model facebook/opt-125m run_num_steps --num-steps 2
+    - python3 offline_inference/profiling.py --model facebook/opt-125m run_num_steps --num-steps 2
 
 - label: Prefix Caching Test # 9min
   mirror_hardwares: [amd]
@@ -216,6 +231,7 @@ steps:
   - vllm/model_executor/layers
   - vllm/sampling_metadata.py
   - tests/samplers
+  - tests/conftest.py
   commands:
     - pytest -v -s samplers
     - VLLM_USE_FLASHINFER_SAMPLER=1 pytest -v -s samplers
@@ -231,20 +247,22 @@ steps:
     - pytest -v -s test_logits_processor.py
     - pytest -v -s model_executor/test_guided_processors.py
 
-- label: Speculative decoding tests # 30min
+- label: Speculative decoding tests # 40min
   source_file_dependencies:
   - vllm/spec_decode
   - tests/spec_decode
+  - vllm/model_executor/models/eagle.py
   commands:
     - pytest -v -s spec_decode/e2e/test_multistep_correctness.py
     - VLLM_ATTENTION_BACKEND=FLASH_ATTN pytest -v -s spec_decode --ignore=spec_decode/e2e/test_multistep_correctness.py
+    - pytest -v -s spec_decode/e2e/test_eagle_correctness.py
 
 - label: LoRA Test %N # 15min each
   mirror_hardwares: [amd]
   source_file_dependencies:
   - vllm/lora
   - tests/lora
-  command: pytest -v -s lora --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --ignore=lora/test_long_context.py --ignore=lora/test_chatglm3_tp.py --ignore=lora/test_llama_tp.py
+  command: pytest -v -s lora --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --ignore=lora/test_long_context.py --ignore=lora/test_chatglm3_tp.py --ignore=lora/test_llama_tp.py --ignore=lora/test_minicpmv_tp.py
   parallelism: 4
 
 - label: "PyTorch Fullgraph Smoke Test" # 9min
@@ -333,8 +351,7 @@ steps:
   - vllm/
   - tests/models
   commands:
-    - pip install -e ./plugins/vllm_add_dummy_model
+    - pytest -v -s models/test_transformers.py
-    - pytest -v -s models/test_oot_registration.py # it needs a clean process
     - pytest -v -s models/test_registry.py
     - pytest -v -s models/test_initialization.py
 
@@ -360,23 +377,26 @@ steps:
     - pytest -v -s models/decoder_only/language -m 'not core_model and not quant_model'
     - pytest -v -s models/embedding/language -m 'not core_model'
 
-- label: Multi-Modal Models Test (Standard) # 28min
+- label: Multi-Modal Models Test (Standard) # 40min
   #mirror_hardwares: [amd]
   source_file_dependencies:
   - vllm/
   - tests/models/decoder_only/audio_language
   - tests/models/decoder_only/vision_language
   - tests/models/embedding/vision_language
+  - tests/models/encoder_decoder/audio_language
   - tests/models/encoder_decoder/vision_language
   commands:
     - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
+    - pytest -v -s models/multimodal
     - pytest -v -s models/decoder_only/audio_language -m 'core_model or quant_model'
     - pytest -v -s --ignore models/decoder_only/vision_language/test_phi3v.py models/decoder_only/vision_language -m 'core_model or quant_model'
     - pytest -v -s models/embedding/vision_language -m core_model
+    - pytest -v -s models/encoder_decoder/audio_language -m core_model
     - pytest -v -s models/encoder_decoder/language -m core_model
     - pytest -v -s models/encoder_decoder/vision_language -m core_model
 
-- label: Multi-Modal Models Test (Extended) 1 # 1h16m
+- label: Multi-Modal Models Test (Extended) 1 # 48m
   optional: true
   source_file_dependencies:
   - vllm/
@@ -459,21 +479,44 @@ steps:
   - vllm/worker/worker_base.py
   - vllm/worker/worker.py
   - vllm/worker/model_runner.py
+  - entrypoints/llm/test_collective_rpc.py
   commands:
+  - pytest -v -s entrypoints/llm/test_collective_rpc.py
+  - torchrun --nproc-per-node=2 distributed/test_torchrun_example.py
   - pytest -v -s ./compile/test_basic_correctness.py
   - pytest -v -s ./compile/test_wrapper.py
   - VLLM_TEST_SAME_HOST=1 torchrun --nproc-per-node=4 distributed/test_same_node.py | grep 'Same node test passed'
   - TARGET_TEST_SUITE=L4 pytest basic_correctness/ -v -s -m 'distributed(num_gpus=2)'
   # Avoid importing model tests that cause CUDA reinitialization error
+  - pytest models/test_transformers.py -v -s -m 'distributed(num_gpus=2)'
   - pytest models/encoder_decoder/language/test_bart.py -v -s -m 'distributed(num_gpus=2)'
   - pytest models/encoder_decoder/vision_language/test_broadcast.py -v -s -m 'distributed(num_gpus=2)'
   - pytest models/decoder_only/vision_language/test_models.py -v -s -m 'distributed(num_gpus=2)'
-  - pytest -v -s spec_decode/e2e/test_integration_dist_tp2.py
+  # this test fails consistently.
-  - pip install -e ./plugins/vllm_add_dummy_model
+  # TODO: investigate and fix
-  - pytest -v -s distributed/test_distributed_oot.py
+  # - pytest -v -s spec_decode/e2e/test_integration_dist_tp2.py
   - CUDA_VISIBLE_DEVICES=0,1 pytest -v -s test_sharded_state_loader.py
   - CUDA_VISIBLE_DEVICES=0,1 pytest -v -s kv_transfer/disagg_test.py
 
+- label: Plugin Tests (2 GPUs) # 40min
+  working_dir: "/vllm-workspace/tests"
+  num_gpus: 2
+  fast_check: true
+  source_file_dependencies:
+  - vllm/plugins/
+  - tests/plugins/
+  commands:
+  # begin platform plugin tests, all the code in-between runs on dummy platform
+  - pip install -e ./plugins/vllm_add_dummy_platform
+  - pytest -v -s plugins_tests/test_platform_plugins.py
+  - pip uninstall vllm_add_dummy_platform -y
+  # end platform plugin tests
+  # other tests continue here:
+  - pip install -e ./plugins/vllm_add_dummy_model
+  - pytest -v -s distributed/test_distributed_oot.py
+  - pytest -v -s entrypoints/openai/test_oot_registration.py # it needs a clean process
+  - pytest -v -s models/test_oot_registration.py # it needs a clean process
+
 - label: Multi-step Tests (4 GPUs) # 36min
   working_dir: "/vllm-workspace/tests"
   num_gpus: 4
@@ -489,7 +532,9 @@ steps:
   - vllm/engine
   - tests/multi_step
   commands:
-  - pytest -v -s multi_step/test_correctness_async_llm.py
+  # this test is quite flaky
+  # TODO: investigate and fix.
+  # - pytest -v -s multi_step/test_correctness_async_llm.py
   - pytest -v -s multi_step/test_correctness_llm.py
 
 - label: Pipeline Parallelism Test # 45min
@@ -520,6 +565,7 @@ steps:
     # requires multi-GPU testing for validation.
     - pytest -v -s -x lora/test_chatglm3_tp.py
     - pytest -v -s -x lora/test_llama_tp.py
+    - pytest -v -s -x lora/test_minicpmv_tp.py
 
 
 - label: Weight Loading Multiple GPU Test  # 33min

.github/CODEOWNERS

@@ -2,32 +2,35 @@
 # for more info about CODEOWNERS file
 
 # This lists cover the "core" components of vLLM that require careful review
-/vllm/attention/backends/abstract.py @WoosukKwon @zhuohan123 @youkaichao @alexm-neuralmagic @comaniac @njhill
+/vllm/attention/backends/abstract.py @WoosukKwon @zhuohan123 @youkaichao @alexm-redhat @comaniac @njhill
-/vllm/core @zhuohan123 @youkaichao @alexm-neuralmagic @comaniac @njhill
+/vllm/core @zhuohan123 @youkaichao @alexm-redhat @comaniac @njhill
-/vllm/engine/llm_engine.py @zhuohan123 @youkaichao @alexm-neuralmagic @comaniac @njhill
+/vllm/engine/llm_engine.py @zhuohan123 @youkaichao @alexm-redhat @comaniac @njhill
-/vllm/executor/executor_base.py @zhuohan123 @youkaichao @alexm-neuralmagic @comaniac @njhill
+/vllm/executor/executor_base.py @zhuohan123 @youkaichao @alexm-redhat @comaniac @njhill
-/vllm/worker/worker_base.py @zhuohan123 @youkaichao @alexm-neuralmagic @comaniac @njhill
+/vllm/worker/worker_base.py @zhuohan123 @youkaichao @alexm-redhat @comaniac @njhill
-/vllm/worker/worker.py @zhuohan123 @youkaichao @alexm-neuralmagic @comaniac @njhill
+/vllm/worker/worker.py @zhuohan123 @youkaichao @alexm-redhat @comaniac @njhill
-/vllm/model_executor/layers/sampler.py @zhuohan123 @youkaichao @alexm-neuralmagic @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
+/vllm/multimodal @DarkLight1337 @ywang96
 CMakeLists.txt @tlrmchlsmth
 
 # vLLM V1
-/vllm/v1 @WoosukKwon @robertgshaw2-neuralmagic @njhill @ywang96 @comaniac @alexm-neuralmagic
+/vllm/v1 @WoosukKwon @robertgshaw2-redhat @njhill @ywang96 @comaniac @alexm-redhat
 
 # Test ownership
-/tests/async_engine @njhill @robertgshaw2-neuralmagic @simon-mo
+/tests/async_engine @njhill @robertgshaw2-redhat @simon-mo
 /tests/test_inputs.py @DarkLight1337 @ywang96
-/tests/entrypoints @DarkLight1337 @robertgshaw2-neuralmagic @simon-mo
+/tests/entrypoints @DarkLight1337 @robertgshaw2-redhat @simon-mo
 /tests/models @DarkLight1337 @ywang96
 /tests/multimodal @DarkLight1337 @ywang96
 /tests/prefix_caching @comaniac @KuntaiDu
 /tests/spec_decode @njhill @LiuXiaoxuanPKU
 /tests/kernels @tlrmchlsmth @WoosukKwon
-/tests/quantization @mgoin @robertgshaw2-neuralmagic
+/tests/quantization @mgoin @robertgshaw2-redhat
 /.buildkite/lm-eval-harness @mgoin @simon-mo
 /tests/distributed/test_multi_node_assignment.py @youkaichao
 /tests/distributed/test_pipeline_parallel.py @youkaichao
 /tests/distributed/test_same_node.py @youkaichao
-/tests/multi_step @alexm-neuralmagic @comaniac
+/tests/multi_step @alexm-redhat @comaniac
 /tests/weight_loading @mgoin @youkaichao
 /tests/basic_correctness/test_chunked_prefill @rkooo567 @comaniac

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

@@ -30,15 +30,6 @@ body:
       </details>
   validations:
     required: true
-- type: textarea
-  attributes:
-    label: Model Input Dumps
-    description: |
-      If you are facing crashing due to illegal memory access or other issues with model execution, vLLM may dump the problematic input of the model. In this case, you will see the message `Error in model execution (input dumped to /tmp/err_xxx.pkl)`. If you see this message, please zip the file (because GitHub doesn't support .pkl file format) and upload it here. This will help us to reproduce the issue and facilitate the debugging process.
-    placeholder: |
-      Upload the dumped input file.
-  validations:
-    required: false
 - type: textarea
   attributes:
     label: 🐛 Describe the bug

.github/ISSUE_TEMPLATE/600-new-model.yml

@@ -9,7 +9,7 @@ body:
     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+).
 
-      #### We also highly recommend you read https://docs.vllm.ai/en/latest/models/adding_model.html first to understand how to add a new model.
+      #### We also highly recommend you read https://docs.vllm.ai/en/latest/contributing/model/adding_model.html first to understand how to add a new model.
 - type: textarea
   attributes:
     label: The model to consider.

.github/mergify.yml

@@ -35,6 +35,43 @@ pull_request_rules:
       add:
         - frontend
 
+- name: label-structured-output
+  description: Automatically apply structured-output label
+  conditions:
+    - or:
+      - files~=^vllm/model_executor/guided_decoding/
+      - files=tests/model_executor/test_guided_processors.py
+      - files=tests/entrypoints/llm/test_guided_generate.py
+      - files=benchmarks/benchmark_serving_guided.py
+      - files=benchmarks/benchmark_guided.py
+  actions:
+    label:
+      add:
+        - structured-output
+
+- name: label-speculative-decoding
+  description: Automatically apply speculative-decoding label
+  conditions:
+    - or:
+      - files~=^vllm/spec_decode/
+      - files=vllm/model_executor/layers/spec_decode_base_sampler.py
+      - files~=^tests/spec_decode/
+  actions:
+    label:
+      add:
+        - speculative-decoding
+
+- name: label-v1
+  description: Automatically apply v1 label
+  conditions:
+    - or:
+      - files~=^vllm/v1/
+      - files~=^tests/v1/
+  actions:
+    label:
+      add:
+        - v1
+
 - name: ping author on conflicts and add 'needs-rebase' label
   conditions:
       - conflict

.github/workflows/actionlint.yml

@@ -1,40 +0,0 @@
-name: Lint GitHub Actions workflows
-on:
-  push:
-    branches:
-      - "main"
-    paths:
-      - '.github/workflows/*.ya?ml'
-      - '.github/workflows/actionlint.*'
-      - '.github/workflows/matchers/actionlint.json'
-  pull_request:
-    branches:
-      - "main"
-    paths:
-      - '.github/workflows/*.ya?ml'
-      - '.github/workflows/actionlint.*'
-      - '.github/workflows/matchers/actionlint.json'
-
-env:
-  LC_ALL: en_US.UTF-8
-
-defaults:
-  run:
-    shell: bash
-
-permissions:
-  contents: read
-
-jobs:
-  actionlint:
-    runs-on: ubuntu-latest
-    steps:
-      - name: "Checkout"
-        uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
-        with:
-          fetch-depth: 0
-
-      - name: "Run actionlint"
-        run: |
-          echo "::add-matcher::.github/workflows/matchers/actionlint.json"
-          tools/actionlint.sh -color

.github/workflows/clang-format.yml

@@ -1,53 +0,0 @@
-name: clang-format
-
-on:
-  # Trigger the workflow on push or pull request,
-  # but only for the main branch
-  push:
-    branches:
-      - main
-    paths:
-      - '**/*.h'
-      - '**/*.cpp'
-      - '**/*.cu'
-      - '**/*.cuh'
-      - '.github/workflows/clang-format.yml'
-  pull_request:
-    branches:
-      - main
-    paths:
-      - '**/*.h'
-      - '**/*.cpp'
-      - '**/*.cu'
-      - '**/*.cuh'
-      - '.github/workflows/clang-format.yml'
-
-jobs:
-  clang-format:
-    runs-on: ubuntu-latest
-    strategy:
-      matrix:
-        python-version: ["3.11"]
-    steps:
-    - uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
-    - name: Set up Python ${{ matrix.python-version }}
-      uses: actions/setup-python@0b93645e9fea7318ecaed2b359559ac225c90a2b # v5.3.0
-      with:
-        python-version: ${{ matrix.python-version }}
-    - name: Install dependencies
-      run: |
-        python -m pip install --upgrade pip
-        pip install clang-format==18.1.5
-    - name: Running clang-format
-      run: |
-        EXCLUDES=(
-            'csrc/moe/topk_softmax_kernels.cu'
-            'csrc/quantization/gguf/ggml-common.h'
-            'csrc/quantization/gguf/dequantize.cuh'
-            'csrc/quantization/gguf/vecdotq.cuh'
-            'csrc/quantization/gguf/mmq.cuh'
-            'csrc/quantization/gguf/mmvq.cuh'
-        )
-        find csrc/ \( -name '*.h' -o -name '*.cpp' -o -name '*.cu' -o -name '*.cuh' \) -print \
-            | grep -vFf <(printf "%s\n" "${EXCLUDES[@]}") \
-            | xargs clang-format --dry-run --Werror

.github/workflows/codespell.yml

@@ -1,45 +0,0 @@
-name: codespell
-
-on:
-  # Trigger the workflow on push or pull request,
-  # but only for the main branch
-  push:
-    branches:
-      - main
-    paths:
-      - "**/*.py"
-      - "**/*.md"
-      - "**/*.rst"
-      - pyproject.toml
-      - requirements-lint.txt
-      - .github/workflows/codespell.yml
-  pull_request:
-    branches:
-      - main
-    paths:
-      - "**/*.py"
-      - "**/*.md"
-      - "**/*.rst"
-      - pyproject.toml
-      - requirements-lint.txt
-      - .github/workflows/codespell.yml
-
-jobs:
-  codespell:
-    runs-on: ubuntu-latest
-    strategy:
-      matrix:
-        python-version: ["3.12"]
-    steps:
-    - uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
-    - name: Set up Python ${{ matrix.python-version }}
-      uses: actions/setup-python@0b93645e9fea7318ecaed2b359559ac225c90a2b # v5.3.0
-      with:
-        python-version: ${{ matrix.python-version }}
-    - name: Install dependencies
-      run: |
-        python -m pip install --upgrade pip
-        pip install -r requirements-lint.txt
-    - name: Spelling check with codespell
-      run: |
-        codespell --toml pyproject.toml

.github/workflows/lint-and-deploy.yaml

@@ -27,7 +27,7 @@ jobs:
           version: v3.10.1
 
       - name: Run chart-testing (lint)
-        run: ct lint --target-branch ${{ github.event.repository.default_branch }} --chart-dirs examples/chart-helm --charts examples/chart-helm
+        run: ct lint --target-branch ${{ github.event.repository.default_branch }} --chart-dirs examples/online_serving/chart-helm --charts examples/online_serving/chart-helm
 
       - name: Setup minio
         run: |
@@ -64,7 +64,8 @@ jobs:
         run: |
           export AWS_ACCESS_KEY_ID=minioadmin
           export AWS_SECRET_ACCESS_KEY=minioadmin
-          helm install --wait --wait-for-jobs --timeout 5m0s --debug --create-namespace --namespace=ns-vllm test-vllm examples/chart-helm -f examples/chart-helm/values.yaml --set secrets.s3endpoint=http://minio:9000 --set secrets.s3bucketname=testbucket --set secrets.s3accesskeyid=$AWS_ACCESS_KEY_ID --set secrets.s3accesskey=$AWS_SECRET_ACCESS_KEY --set resources.requests.cpu=1 --set resources.requests.memory=4Gi --set resources.limits.cpu=2 --set resources.limits.memory=5Gi --set image.env[0].name=VLLM_CPU_KVCACHE_SPACE --set image.env[1].name=VLLM_LOGGING_LEVEL --set-string image.env[0].value="1" --set-string image.env[1].value="DEBUG" --set-string extraInit.s3modelpath="opt-125m/" --set-string 'resources.limits.nvidia\.com/gpu=0' --set-string 'resources.requests.nvidia\.com/gpu=0' --set-string image.repository="vllm-cpu-env"
+          sleep 30 && kubectl -n ns-vllm logs -f "$(kubectl -n ns-vllm get pods | awk '/deployment/ {print $1;exit}')" &
+          helm install --wait --wait-for-jobs --timeout 5m0s --debug --create-namespace --namespace=ns-vllm test-vllm examples/online_serving/chart-helm -f examples/online_serving/chart-helm/values.yaml --set secrets.s3endpoint=http://minio:9000 --set secrets.s3bucketname=testbucket --set secrets.s3accesskeyid=$AWS_ACCESS_KEY_ID --set secrets.s3accesskey=$AWS_SECRET_ACCESS_KEY --set resources.requests.cpu=1 --set resources.requests.memory=4Gi --set resources.limits.cpu=2 --set resources.limits.memory=5Gi --set image.env[0].name=VLLM_CPU_KVCACHE_SPACE --set image.env[1].name=VLLM_LOGGING_LEVEL --set-string image.env[0].value="1" --set-string image.env[1].value="DEBUG" --set-string extraInit.s3modelpath="opt-125m/" --set-string 'resources.limits.nvidia\.com/gpu=0' --set-string 'resources.requests.nvidia\.com/gpu=0' --set-string image.repository="vllm-cpu-env"
     
       - name: curl test
         run: |

.github/workflows/matchers/ruff.json

@@ -1,17 +0,0 @@
-{
-    "problemMatcher": [
-      {
-        "owner": "ruff",
-        "pattern": [
-          {
-            "regexp": "^(.+?):(\\d+):(\\d+): (\\w+): (.+)$",
-            "file": 1,
-            "line": 2,
-            "column": 3,
-            "code": 4,
-            "message": 5
-          }
-        ]
-      }
-    ]
-  }

.github/workflows/mypy.yaml

@@ -1,51 +0,0 @@
-name: mypy
-
-on:
-  # Trigger the workflow on push or pull request,
-  # but only for the main branch
-  push:
-    branches:
-      - main
-    paths:
-      - '**/*.py'
-      - '.github/workflows/mypy.yaml'
-      - 'tools/mypy.sh'
-      - 'pyproject.toml'
-  pull_request:
-    branches:
-      - main
-    # This workflow is only relevant when one of the following files changes.
-    # However, we have github configured to expect and require this workflow
-    # to run and pass before github with auto-merge a pull request. Until github
-    # allows more flexible auto-merge policy, we can just run this on every PR.
-    # It doesn't take that long to run, anyway.
-    #paths:
-    #  - '**/*.py'
-    #  - '.github/workflows/mypy.yaml'
-    #  - 'tools/mypy.sh'
-    #  - 'pyproject.toml'
-
-jobs:
-  mypy:
-    runs-on: ubuntu-latest
-    strategy:
-      matrix:
-        python-version: ["3.9", "3.10", "3.11", "3.12"]
-    steps:
-    - uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
-    - name: Set up Python ${{ matrix.python-version }}
-      uses: actions/setup-python@0b93645e9fea7318ecaed2b359559ac225c90a2b # v5.3.0
-      with:
-        python-version: ${{ matrix.python-version }}
-    - name: Install dependencies
-      run: |
-        python -m pip install --upgrade pip
-        pip install mypy==1.11.1
-        pip install types-setuptools
-        pip install types-PyYAML
-        pip install types-requests
-        pip install types-setuptools
-    - name: Mypy
-      run: |
-        echo "::add-matcher::.github/workflows/matchers/mypy.json"
-        tools/mypy.sh 1 ${{ matrix.python-version }}

.github/workflows/png-lint.yml

@@ -1,37 +0,0 @@
-name: Lint PNG exports from excalidraw
-on:
-  push:
-    branches:
-      - "main"
-    paths:
-      - '*.excalidraw.png'
-      - '.github/workflows/png-lint.yml'
-  pull_request:
-    branches:
-      - "main"
-    paths:
-      - '*.excalidraw.png'
-      - '.github/workflows/png-lint.yml'
-
-env:
-  LC_ALL: en_US.UTF-8
-
-defaults:
-  run:
-    shell: bash
-
-permissions:
-  contents: read
-
-jobs:
-  actionlint:
-    runs-on: ubuntu-latest
-    steps:
-      - name: "Checkout"
-        uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
-        with:
-          fetch-depth: 0
-
-      - name: "Run png-lint.sh to check excalidraw exported images"
-        run: |
-          tools/png-lint.sh

.github/workflows/pre-commit.yml

@@ -0,0 +1,19 @@
+name: pre-commit
+
+on:
+  pull_request:
+  push:
+    branches: [main]
+
+jobs:
+  pre-commit:
+    runs-on: ubuntu-latest
+    steps:
+    - uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
+    - uses: actions/setup-python@0b93645e9fea7318ecaed2b359559ac225c90a2b # v5.3.0
+      with:
+        python-version: "3.12"
+    - run: echo "::add-matcher::.github/workflows/matchers/actionlint.json"
+    - uses: pre-commit/action@2c7b3805fd2a0fd8c1884dcaebf91fc102a13ecd # v3.0.1
+      with:
+        extra_args: --all-files --hook-stage manual

.github/workflows/reminder_comment.yml

@@ -2,7 +2,6 @@ name: PR Reminder Comment Bot
 on:
   pull_request_target:
     types: [opened]
-
 jobs:
   pr_reminder:
     runs-on: ubuntu-latest
@@ -15,7 +14,12 @@ jobs:
               owner: context.repo.owner,
               repo: context.repo.repo,
               issue_number: context.issue.number,
-              body: '👋 Hi! Thank you for contributing to the vLLM project.\n Just a reminder: PRs would not trigger full CI run by default. Instead, it would only run `fastcheck` CI which starts running only a small and essential subset of CI tests to quickly catch errors. You can run other CI tests on top of those by going to your `fastcheck` build on Buildkite UI (linked in the PR checks section) and unblock them. If you do not have permission to unblock, ping `simon-mo` or `khluu` to add you in our Buildkite org. \n\nOnce the PR is approved and ready to go, your PR reviewer(s) can run CI to test the changes comprehensively before merging.\n\n To run CI, PR reviewers can do one of these:\n- Add `ready` label to the PR\n- Enable auto-merge.\n\n🚀'
+              body: '👋 Hi! Thank you for contributing to the vLLM project.\n\n' +
+                '💬 Join our developer Slack at https://slack.vllm.ai to discuss your PR in #pr-reviews, coordinate on features in #feat- channels, or join special interest groups in #sig- channels.\n\n' +
+                'Just a reminder: PRs would not trigger full CI run by default. Instead, it would only run `fastcheck` CI which starts running only a small and essential subset of CI tests to quickly catch errors. You can run other CI tests on top of those by going to your `fastcheck` build on Buildkite UI (linked in the PR checks section) and unblock them. If you do not have permission to unblock, ping `simon-mo` or `khluu` to add you in our Buildkite org.\n\n' +
+                'Once the PR is approved and ready to go, your PR reviewer(s) can run CI to test the changes comprehensively before merging.\n\n' +
+                'To run CI, PR reviewers can either: Add `ready` label to the PR or enable auto-merge.\n\n' +
+                '🚀'
             })
         env:
           GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}

.github/workflows/ruff.yml

@@ -1,52 +0,0 @@
-name: ruff
-
-on:
-  # Trigger the workflow on push or pull request,
-  # but only for the main branch
-  push:
-    branches:
-      - main
-    paths:
-      - "**/*.py"
-      - pyproject.toml
-      - requirements-lint.txt
-      - .github/workflows/matchers/ruff.json
-      - .github/workflows/ruff.yml
-  pull_request:
-    branches:
-      - main
-    # This workflow is only relevant when one of the following files changes.
-    # However, we have github configured to expect and require this workflow
-    # to run and pass before github with auto-merge a pull request. Until github
-    # allows more flexible auto-merge policy, we can just run this on every PR.
-    # It doesn't take that long to run, anyway.
-    #paths:
-    #  - "**/*.py"
-    #  - pyproject.toml
-    #  - requirements-lint.txt
-    #  - .github/workflows/matchers/ruff.json
-    #  - .github/workflows/ruff.yml
-
-jobs:
-  ruff:
-    runs-on: ubuntu-latest
-    strategy:
-      matrix:
-        python-version: ["3.12"]
-    steps:
-      - uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
-      - name: Set up Python ${{ matrix.python-version }}
-        uses: actions/setup-python@0b93645e9fea7318ecaed2b359559ac225c90a2b # v5.3.0
-        with:
-          python-version: ${{ matrix.python-version }}
-      - name: Install dependencies
-        run: |
-          python -m pip install --upgrade pip
-          pip install -r requirements-lint.txt
-      - name: Analysing the code with ruff
-        run: |
-          echo "::add-matcher::.github/workflows/matchers/ruff.json"
-          ruff check --output-format github .
-      - name: Run isort
-        run: |
-          isort . --check-only

.github/workflows/shellcheck.yml

@@ -1,37 +0,0 @@
-name: Lint shell scripts
-on:
-  push:
-    branches:
-      - "main"
-    paths:
-      - '**/*.sh'
-      - '.github/workflows/shellcheck.yml'
-  pull_request:
-    branches:
-      - "main"
-    paths:
-      - '**/*.sh'
-      - '.github/workflows/shellcheck.yml'
-
-env:
-  LC_ALL: en_US.UTF-8
-
-defaults:
-  run:
-    shell: bash
-
-permissions:
-  contents: read
-
-jobs:
-  shellcheck:
-    runs-on: ubuntu-latest
-    steps:
-      - name: "Checkout"
-        uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
-        with:
-          fetch-depth: 0
-
-      - name: "Check shell scripts"
-        run: |
-          tools/shellcheck.sh

.github/workflows/sphinx-lint.yml

@@ -1,32 +0,0 @@
-name: Lint documentation
-
-on:
-  push:
-    branches:
-      - main
-    paths:
-      - "docs/**"
-  pull_request:
-    branches:
-      - main
-    paths:
-      - "docs/**"
-
-jobs:
-  sphinx-lint:
-    runs-on: ubuntu-latest
-    strategy:
-      matrix:
-        python-version: ["3.12"]
-    steps:
-      - uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
-      - name: Set up Python ${{ matrix.python-version }}
-        uses: actions/setup-python@0b93645e9fea7318ecaed2b359559ac225c90a2b # v5.3.0
-        with:
-          python-version: ${{ matrix.python-version }}
-      - name: Install dependencies
-        run: |
-          python -m pip install --upgrade pip
-          pip install -r requirements-lint.txt
-      - name: Linting docs
-        run: tools/sphinx-lint.sh

.github/workflows/yapf.yml

@@ -1,38 +0,0 @@
-name: yapf
-
-on:
-  # Trigger the workflow on push or pull request,
-  # but only for the main branch
-  push:
-    branches:
-      - main
-    paths:
-      - "**/*.py"
-      - .github/workflows/yapf.yml
-  pull_request:
-    branches:
-      - main
-    paths:
-      - "**/*.py"
-      - .github/workflows/yapf.yml
-
-jobs:
-  yapf:
-    runs-on: ubuntu-latest
-    strategy:
-      matrix:
-        python-version: ["3.12"]
-    steps:
-      - uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
-      - name: Set up Python ${{ matrix.python-version }}
-        uses: actions/setup-python@0b93645e9fea7318ecaed2b359559ac225c90a2b # v5.3.0
-        with:
-          python-version: ${{ matrix.python-version }}
-      - name: Install dependencies
-        run: |
-          python -m pip install --upgrade pip
-          pip install yapf==0.32.0
-          pip install toml==0.10.2
-      - name: Running yapf
-        run: |
-          yapf --diff --recursive .

.gitignore

@@ -79,10 +79,7 @@ instance/
 
 # Sphinx documentation
 docs/_build/
-docs/source/getting_started/examples/*.rst
+docs/source/getting_started/examples/
-!**/*.template.rst
-docs/source/getting_started/examples/*.md
-!**/*.template.md
 
 # PyBuilder
 .pybuilder/

.pre-commit-config.yaml

@@ -0,0 +1,110 @@
+default_stages:
+  - pre-commit # Run locally
+  - manual # Run in CI
+repos:
+- repo: https://github.com/google/yapf
+  rev: v0.43.0
+  hooks:
+  - id: yapf
+    args: [--in-place, --verbose]
+    additional_dependencies: [toml] # TODO: Remove when yapf is upgraded
+- repo: https://github.com/astral-sh/ruff-pre-commit
+  rev: v0.9.3
+  hooks:
+  - id: ruff
+    args: [--output-format, github]
+- repo: https://github.com/codespell-project/codespell
+  rev: v2.4.0
+  hooks:
+  - id: codespell
+    exclude: 'benchmarks/sonnet.txt|(build|tests/(lora/data|models/fixtures|prompts))/.*'
+- repo: https://github.com/PyCQA/isort
+  rev: 5.13.2
+  hooks:
+  - id: isort
+- repo: https://github.com/pre-commit/mirrors-clang-format
+  rev: v19.1.7
+  hooks:
+  - id: clang-format
+    exclude: 'csrc/(moe/topk_softmax_kernels.cu|quantization/gguf/(ggml-common.h|dequantize.cuh|vecdotq.cuh|mmq.cuh|mmvq.cuh))'
+    types_or: [c++, cuda]
+    args: [--style=file, --verbose]
+- repo: https://github.com/jackdewinter/pymarkdown
+  rev: v0.9.27
+  hooks:
+  - id: pymarkdown
+    files: docs/.*
+- repo: https://github.com/rhysd/actionlint
+  rev: v1.7.7
+  hooks:
+  - id: actionlint
+- repo: local
+  hooks:
+  - id: mypy-local
+    name: Run mypy for local Python installation
+    entry: tools/mypy.sh 0 "local"
+    language: python
+    types: [python]
+    additional_dependencies: &mypy_deps [mypy==1.11.1, 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
+    entry: tools/mypy.sh 1 "3.9"
+    language: python
+    types: [python]
+    additional_dependencies: *mypy_deps
+    stages: [manual] # Only run in CI
+  - id: mypy-3.10 # TODO: Use https://github.com/pre-commit/mirrors-mypy when mypy setup is less awkward
+    name: Run mypy for Python 3.10
+    entry: tools/mypy.sh 1 "3.10"
+    language: python
+    types: [python]
+    additional_dependencies: *mypy_deps
+    stages: [manual] # Only run in CI
+  - id: mypy-3.11 # TODO: Use https://github.com/pre-commit/mirrors-mypy when mypy setup is less awkward
+    name: Run mypy for Python 3.11
+    entry: tools/mypy.sh 1 "3.11"
+    language: python
+    types: [python]
+    additional_dependencies: *mypy_deps
+    stages: [manual] # Only run in CI
+  - id: mypy-3.12 # TODO: Use https://github.com/pre-commit/mirrors-mypy when mypy setup is less awkward
+    name: Run mypy for Python 3.12
+    entry: tools/mypy.sh 1 "3.12"
+    language: python
+    types: [python]
+    additional_dependencies: *mypy_deps
+    stages: [manual] # Only run in CI
+  - id: shellcheck
+    name: Lint shell scripts
+    entry: tools/shellcheck.sh
+    language: script
+    types: [shell]
+  - id: png-lint
+    name: Lint PNG exports from excalidraw
+    entry: tools/png-lint.sh
+    language: script
+    types: [png]
+  - id: signoff-commit
+    name: Sign-off Commit
+    entry: bash
+    args:
+      - -c
+      - |
+        if ! grep -q "^Signed-off-by: $(git config user.name) <$(git config user.email)>" .git/COMMIT_EDITMSG; then
+          printf "\nSigned-off-by: $(git config user.name) <$(git config user.email)>\n" >> .git/COMMIT_EDITMSG
+        fi
+    language: system
+    verbose: true
+    stages: [commit-msg]
+  - id: check-spdx-header
+    name: Check SPDX headers
+    entry: python tools/check_spdx_header.py
+    language: python
+    types: [python]
+  - id: suggestion
+    name: Suggestion
+    entry: bash -c 'echo "To bypass pre-commit hooks, add --no-verify to git commit."'
+    language: system
+    verbose: true
+    pass_filenames: false

CMakeLists.txt

@@ -24,9 +24,6 @@ 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.
@@ -181,6 +178,31 @@ message(STATUS "FetchContent base directory: ${FETCHCONTENT_BASE_DIR}")
 # Define other extension targets
 #
 
+#
+# cumem_allocator extension
+#
+
+set(VLLM_CUMEM_EXT_SRC
+  "csrc/cumem_allocator.cpp")
+
+set_gencode_flags_for_srcs(
+  SRCS "${VLLM_CUMEM_EXT_SRC}"
+  CUDA_ARCHS "${CUDA_ARCHS}")
+
+if(VLLM_GPU_LANG STREQUAL "CUDA")
+  message(STATUS "Enabling cumem allocator extension.")
+  # link against cuda driver library
+  list(APPEND CUMEM_LIBS cuda)
+  define_gpu_extension_target(
+    cumem_allocator
+    DESTINATION vllm
+    LANGUAGE CXX
+    SOURCES ${VLLM_CUMEM_EXT_SRC}
+    LIBRARIES ${CUMEM_LIBS}
+    USE_SABI 3.8
+    WITH_SOABI)
+endif()
+
 #
 # _C extension
 #
@@ -223,13 +245,13 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
     FetchContent_Declare(
         cutlass
         GIT_REPOSITORY https://github.com/nvidia/cutlass.git
-        GIT_TAG 8aa95dbb888be6d81c6fbf7169718c5244b53227
+        GIT_TAG v3.7.0
         GIT_PROGRESS TRUE
 
         # Speed up CUTLASS download by retrieving only the specified GIT_TAG instead of the history.
         # Important: If GIT_SHALLOW is enabled then GIT_TAG works only with branch names and tags.
         # So if the GIT_TAG above is updated to a commit hash, GIT_SHALLOW must be set to FALSE
-        GIT_SHALLOW FALSE
+        GIT_SHALLOW TRUE
     )
   endif()
   FetchContent_MakeAvailable(cutlass)
@@ -253,7 +275,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
   # Only build Marlin kernels if we are building for at least some compatible archs.
   # Keep building Marlin for 9.0 as there are some group sizes and shapes that
   # are not supported by Machete yet.
-  cuda_archs_loose_intersection(MARLIN_ARCHS "8.0;8.6;8.7;8.9;9.0" ${CUDA_ARCHS})
+  cuda_archs_loose_intersection(MARLIN_ARCHS "8.0;8.6;8.7;8.9;9.0" "${CUDA_ARCHS}")
   if (MARLIN_ARCHS)
     set(MARLIN_SRCS
        "csrc/quantization/fp8/fp8_marlin.cu"
@@ -274,10 +296,15 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
   endif()
 
   # The cutlass_scaled_mm kernels for Hopper (c3x, i.e. CUTLASS 3.x) require
-  # CUDA 12.0 or later (and only work on Hopper, 9.0/9.0a for now).
+  # CUDA 12.0 or later (and only work on Hopper, 9.0a for now).
-  cuda_archs_loose_intersection(SCALED_MM_3X_ARCHS "9.0;9.0a" "${CUDA_ARCHS}")
+  cuda_archs_loose_intersection(SCALED_MM_3X_ARCHS "9.0a" "${CUDA_ARCHS}")
   if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.0 AND SCALED_MM_3X_ARCHS)
-    set(SRCS "csrc/quantization/cutlass_w8a8/scaled_mm_c3x.cu")
+    set(SRCS 
+       "csrc/quantization/cutlass_w8a8/scaled_mm_c3x.cu"
+       "csrc/quantization/cutlass_w8a8/c3x/scaled_mm_sm90_fp8.cu"
+       "csrc/quantization/cutlass_w8a8/c3x/scaled_mm_sm90_int8.cu"
+       "csrc/quantization/cutlass_w8a8/c3x/scaled_mm_azp_sm90_int8.cu"
+       "csrc/quantization/cutlass_w8a8/c3x/scaled_mm_blockwise_sm90_fp8.cu")
     set_gencode_flags_for_srcs(
       SRCS "${SRCS}"
       CUDA_ARCHS "${SCALED_MM_3X_ARCHS}")
@@ -329,7 +356,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
   # 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, 9.0/9.0a for now).
+  # require CUDA 12.2 or later (and only work on Hopper, 9.0a for now).
   if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.2 AND SCALED_MM_3X_ARCHS)
     set(SRCS "csrc/sparse/cutlass/sparse_compressor_c3x.cu"
              "csrc/sparse/cutlass/sparse_scaled_mm_c3x.cu")
@@ -510,7 +537,7 @@ if(VLLM_GPU_LANG STREQUAL "HIP")
 endif()
 
 # vllm-flash-attn currently only supported on CUDA
-if (NOT VLLM_TARGET_DEVICE STREQUAL "cuda")
+if (NOT VLLM_GPU_LANG STREQUAL "CUDA")
   return()
 endif ()
 
@@ -533,7 +560,7 @@ endif()
 # They should be identical but if they aren't, this is a massive footgun.
 #
 # The vllm-flash-attn install rules are nested under vllm to make sure the library gets installed in the correct place.
-# To only install vllm-flash-attn, use --component vllm_flash_attn_c.
+# To only install vllm-flash-attn, use --component _vllm_fa2_C (for FA2) or --component _vllm_fa3_C (for FA3).
 # If no component is specified, vllm-flash-attn is still installed.
 
 # If VLLM_FLASH_ATTN_SRC_DIR is set, vllm-flash-attn is installed from that directory instead of downloading.
@@ -545,42 +572,40 @@ if (DEFINED ENV{VLLM_FLASH_ATTN_SRC_DIR})
 endif()
 
 if(VLLM_FLASH_ATTN_SRC_DIR)
-  FetchContent_Declare(vllm-flash-attn SOURCE_DIR ${VLLM_FLASH_ATTN_SRC_DIR})
+  FetchContent_Declare(
+          vllm-flash-attn SOURCE_DIR 
+          ${VLLM_FLASH_ATTN_SRC_DIR}
+          BINARY_DIR ${CMAKE_BINARY_DIR}/vllm-flash-attn
+  )
 else()
   FetchContent_Declare(
           vllm-flash-attn
           GIT_REPOSITORY https://github.com/vllm-project/flash-attention.git
-          GIT_TAG 04325b6798bcc326c86fb35af62d05a9c8c8eceb
+          GIT_TAG d4e09037abf588af1ec47d0e966b237ee376876c
           GIT_PROGRESS TRUE
           # Don't share the vllm-flash-attn build between build types
           BINARY_DIR ${CMAKE_BINARY_DIR}/vllm-flash-attn
   )
 endif()
 
-# Set the parent build flag so that the vllm-flash-attn library does not redo compile flag and arch initialization.
-set(VLLM_PARENT_BUILD ON)
-
-# Ensure the vllm/vllm_flash_attn directory exists before installation
-install(CODE "file(MAKE_DIRECTORY \"\${CMAKE_INSTALL_PREFIX}/vllm/vllm_flash_attn\")" COMPONENT vllm_flash_attn_c)
-
-# Make sure vllm-flash-attn install rules are nested under vllm/
-install(CODE "set(CMAKE_INSTALL_LOCAL_ONLY FALSE)" COMPONENT vllm_flash_attn_c)
-install(CODE "set(OLD_CMAKE_INSTALL_PREFIX \"\${CMAKE_INSTALL_PREFIX}\")" COMPONENT vllm_flash_attn_c)
-install(CODE "set(CMAKE_INSTALL_PREFIX \"\${CMAKE_INSTALL_PREFIX}/vllm/\")" COMPONENT vllm_flash_attn_c)
 
 # 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
+# Copy over the vllm-flash-attn python files (duplicated for fa2 and fa3, in
-install(CODE "set(CMAKE_INSTALL_PREFIX \"\${OLD_CMAKE_INSTALL_PREFIX}\")" COMPONENT vllm_flash_attn_c)
+# case only one is built, in the case both are built redundant work is done)
-install(CODE "set(CMAKE_INSTALL_LOCAL_ONLY TRUE)" COMPONENT vllm_flash_attn_c)
-
-# Copy over the vllm-flash-attn python files
 install(
   DIRECTORY ${vllm-flash-attn_SOURCE_DIR}/vllm_flash_attn/
-        DESTINATION vllm/vllm_flash_attn
+  DESTINATION vllm_flash_attn
-        COMPONENT vllm_flash_attn_c
+  COMPONENT _vllm_fa2_C
+  FILES_MATCHING PATTERN "*.py"
+)
+
+install(
+  DIRECTORY ${vllm-flash-attn_SOURCE_DIR}/vllm_flash_attn/
+  DESTINATION vllm_flash_attn
+  COMPONENT _vllm_fa3_C
   FILES_MATCHING PATTERN "*.py"
 )
 

CODE_OF_CONDUCT.md

@@ -61,7 +61,7 @@ representative at an online or offline/IRL event.
 
 Instances of abusive, harassing, or otherwise unacceptable behavior may be
 reported to the community leaders responsible for enforcement in the #code-of-conduct
-channel in the [vLLM Discord](https://discord.com/invite/jz7wjKhh6g).
+channel in the [vLLM Slack](https://slack.vllm.ai).
 All complaints will be reviewed and investigated promptly and fairly.
 
 All community leaders are obligated to respect the privacy and security of the

Dockerfile

@@ -2,8 +2,8 @@
 # to run the OpenAI compatible server.
 
 # Please update any changes made here to
-# docs/source/dev/dockerfile/dockerfile.md and
+# docs/source/contributing/dockerfile/dockerfile.md and
-# docs/source/assets/dev/dockerfile-stages-dependency.png
+# docs/source/assets/contributing/dockerfile-stages-dependency.png
 
 ARG CUDA_VERSION=12.4.1
 #################### BASE BUILD IMAGE ####################
@@ -52,7 +52,7 @@ WORKDIR /workspace
 # after this step
 RUN --mount=type=cache,target=/root/.cache/pip \
     if [ "$TARGETPLATFORM" = "linux/arm64" ]; then \
-        python3 -m pip install --index-url https://download.pytorch.org/whl/nightly/cu124 "torch==2.6.0.dev20241210+cu124" "torchvision==0.22.0.dev20241215";  \
+        python3 -m pip install --index-url https://download.pytorch.org/whl/nightly/cu126 "torch==2.7.0.dev20250121+cu126" "torchvision==0.22.0.dev20250121";  \
     fi
 
 COPY requirements-common.txt requirements-common.txt
@@ -126,8 +126,8 @@ RUN --mount=type=cache,target=/root/.cache/ccache \
 
 # Check the size of the wheel if RUN_WHEEL_CHECK is true
 COPY .buildkite/check-wheel-size.py check-wheel-size.py
-# Default max size of the wheel is 250MB
+# sync the default value with .buildkite/check-wheel-size.py
-ARG VLLM_MAX_SIZE_MB=250
+ARG VLLM_MAX_SIZE_MB=400
 ENV VLLM_MAX_SIZE_MB=$VLLM_MAX_SIZE_MB
 ARG RUN_WHEEL_CHECK=true
 RUN if [ "$RUN_WHEEL_CHECK" = "true" ]; then \
@@ -149,7 +149,8 @@ RUN --mount=type=cache,target=/root/.cache/pip \
 
 #################### vLLM installation IMAGE ####################
 # image with vLLM installed
-FROM nvidia/cuda:${CUDA_VERSION}-base-ubuntu22.04 AS vllm-base
+# TODO: Restore to base image after FlashInfer AOT wheel fixed
+FROM nvidia/cuda:${CUDA_VERSION}-devel-ubuntu22.04 AS vllm-base
 ARG CUDA_VERSION=12.4.1
 ARG PYTHON_VERSION=3.12
 WORKDIR /vllm-workspace
@@ -194,12 +195,30 @@ RUN --mount=type=bind,from=build,src=/workspace/dist,target=/vllm-workspace/dist
     --mount=type=cache,target=/root/.cache/pip \
     python3 -m pip install dist/*.whl --verbose
 
+# How to build this FlashInfer wheel:
+# $ export FLASHINFER_ENABLE_AOT=1
+# $ # Note we remove 7.0 from the arch list compared to the list below, since FlashInfer only supports sm75+
+# $ export TORCH_CUDA_ARCH_LIST='7.5 8.0 8.6 8.9 9.0+PTX'
+# $ git clone https://github.com/flashinfer-ai/flashinfer.git --recursive
+# $ cd flashinfer
+# $ git checkout 524304395bd1d8cd7d07db083859523fcaa246a4
+# $ python3 setup.py bdist_wheel --dist-dir=dist --verbose
+
 RUN --mount=type=cache,target=/root/.cache/pip \
 . /etc/environment && \
 if [ "$TARGETPLATFORM" != "linux/arm64" ]; then \
-    python3 -m pip install https://github.com/flashinfer-ai/flashinfer/releases/download/v0.1.6/flashinfer-0.1.6+cu121torch2.4-cp${PYTHON_VERSION_STR}-cp${PYTHON_VERSION_STR}-linux_x86_64.whl; \
+    python3 -m pip install https://wheels.vllm.ai/flashinfer/524304395bd1d8cd7d07db083859523fcaa246a4/flashinfer_python-0.2.0.post1-cp${PYTHON_VERSION_STR}-cp${PYTHON_VERSION_STR}-linux_x86_64.whl; \
 fi
 COPY examples examples
+
+# Although we build Flashinfer with AOT mode, there's still
+# some issues w.r.t. JIT compilation. Therefore we need to
+# install build dependencies for JIT compilation.
+# TODO: Remove this once FlashInfer AOT wheel is fixed
+COPY requirements-build.txt requirements-build.txt
+RUN --mount=type=cache,target=/root/.cache/pip \
+    python3 -m pip install -r requirements-build.txt
+
 #################### vLLM installation IMAGE ####################
 
 #################### TEST IMAGE ####################
@@ -234,8 +253,8 @@ RUN mv vllm test_docs/
 #################### TEST IMAGE ####################
 
 #################### OPENAI API SERVER ####################
-# openai api server alternative
+# base openai image with additional requirements, for any subsequent openai-style images
-FROM vllm-base AS vllm-openai
+FROM vllm-base AS vllm-openai-base
 
 # install additional dependencies for openai api server
 RUN --mount=type=cache,target=/root/.cache/pip \
@@ -247,5 +266,14 @@ RUN --mount=type=cache,target=/root/.cache/pip \
 
 ENV VLLM_USAGE_SOURCE production-docker-image
 
+# define sagemaker first, so it is not default from `docker build`
+FROM vllm-openai-base AS vllm-sagemaker
+
+COPY examples/online_serving/sagemaker-entrypoint.sh .
+RUN chmod +x sagemaker-entrypoint.sh
+ENTRYPOINT ["./sagemaker-entrypoint.sh"]
+
+FROM vllm-openai-base AS vllm-openai
+
 ENTRYPOINT ["python3", "-m", "vllm.entrypoints.openai.api_server"]
 #################### OPENAI API SERVER ####################

Dockerfile.cpu

@@ -26,10 +26,10 @@ RUN pip install intel_extension_for_pytorch==2.5.0
 
 WORKDIR /workspace
 
-COPY requirements-build.txt requirements-build.txt
 ARG PIP_EXTRA_INDEX_URL="https://download.pytorch.org/whl/cpu"
 ENV PIP_EXTRA_INDEX_URL=${PIP_EXTRA_INDEX_URL}
 RUN --mount=type=cache,target=/root/.cache/pip \
+    --mount=type=bind,src=requirements-build.txt,target=requirements-build.txt \
     pip install --upgrade pip && \
     pip install -r requirements-build.txt
 
@@ -37,9 +37,9 @@ FROM cpu-test-1 AS build
 
 WORKDIR /workspace/vllm
 
-COPY requirements-common.txt requirements-common.txt
-COPY requirements-cpu.txt requirements-cpu.txt
 RUN --mount=type=cache,target=/root/.cache/pip \
+    --mount=type=bind,src=requirements-common.txt,target=requirements-common.txt \
+    --mount=type=bind,src=requirements-cpu.txt,target=requirements-cpu.txt \
     pip install -v -r requirements-cpu.txt
 
 COPY . .

Dockerfile.hpu

@@ -1,4 +1,4 @@
-FROM vault.habana.ai/gaudi-docker/1.18.0/ubuntu22.04/habanalabs/pytorch-installer-2.4.0:latest
+FROM vault.habana.ai/gaudi-docker/1.19.1/ubuntu22.04/habanalabs/pytorch-installer-2.5.1:latest
 
 COPY ./ /workspace/vllm
 

Dockerfile.neuron

@@ -1,6 +1,6 @@
 # default base image
 # https://gallery.ecr.aws/neuron/pytorch-inference-neuronx
-ARG BASE_IMAGE="public.ecr.aws/neuron/pytorch-inference-neuronx:2.1.2-neuronx-py310-sdk2.20.2-ubuntu20.04"
+ARG BASE_IMAGE="public.ecr.aws/neuron/pytorch-inference-neuronx:2.5.1-neuronx-py310-sdk2.21.0-ubuntu22.04"
 
 FROM $BASE_IMAGE
 
@@ -15,16 +15,17 @@ RUN apt-get update && \
         ffmpeg libsm6 libxext6 libgl1
 
 ### Mount Point ###
-# When launching the container, mount the code directory to /app
+# When launching the container, mount the code directory to /workspace
-ARG APP_MOUNT=/app
+ARG APP_MOUNT=/workspace
 VOLUME [ ${APP_MOUNT} ]
 WORKDIR ${APP_MOUNT}/vllm
 
 RUN python3 -m pip install --upgrade pip
 RUN python3 -m pip install --no-cache-dir fastapi ninja tokenizers pandas
-RUN python3 -m pip install sentencepiece transformers==4.36.2 -U
+RUN python3 -m pip install sentencepiece transformers==4.45.2 -U
 RUN python3 -m pip install transformers-neuronx --extra-index-url=https://pip.repos.neuron.amazonaws.com -U
-RUN python3 -m pip install --pre neuronx-cc==2.15.* --extra-index-url=https://pip.repos.neuron.amazonaws.com -U
+RUN python3 -m pip install neuronx-cc==2.16.345.0 --extra-index-url=https://pip.repos.neuron.amazonaws.com -U
+RUN python3 -m pip install pytest
 
 COPY . .
 ARG GIT_REPO_CHECK=0
@@ -42,4 +43,7 @@ RUN --mount=type=bind,source=.git,target=.git \
 # install development dependencies (for testing)
 RUN python3 -m pip install -e tests/vllm_test_utils
 
+# overwrite entrypoint to run bash script
+RUN echo "import subprocess; import sys; subprocess.check_call(sys.argv[1:])" > /usr/local/bin/dockerd-entrypoint.py
+
 CMD ["/bin/bash"]

Dockerfile.openvino

@@ -14,6 +14,7 @@ ARG GIT_REPO_CHECK=0
 RUN --mount=type=bind,source=.git,target=.git \
     if [ "$GIT_REPO_CHECK" != 0 ]; then bash tools/check_repo.sh ; fi
 
+RUN python3 -m pip install -U pip
 # install build requirements
 RUN PIP_EXTRA_INDEX_URL="https://download.pytorch.org/whl/cpu" python3 -m pip install -r /workspace/requirements-build.txt
 # build vLLM with OpenVINO backend

Dockerfile.ppc64le

@@ -4,12 +4,12 @@ USER root
 
 ENV PATH="/usr/local/cargo/bin:$PATH:/opt/conda/bin/"
 
-RUN apt-get update -y && apt-get install -y git wget curl vim libnuma-dev libsndfile-dev libprotobuf-dev build-essential ffmpeg libsm6 libxext6 libgl1 
+RUN apt-get update -y && apt-get install -y git wget kmod curl vim libnuma-dev libsndfile-dev libprotobuf-dev build-essential ffmpeg libsm6 libxext6 libgl1 libssl-dev 
 
 # Some packages in requirements-cpu are installed here
 # IBM provides optimized packages for ppc64le processors in the open-ce project for mamba
 # Currently these may not be available for venv or pip directly
-RUN micromamba install -y -n base -c https://ftp.osuosl.org/pub/open-ce/1.11.0-p10/ -c defaults python=3.10 torchvision-cpu=0.16.2 rust && micromamba clean --all --yes
+RUN micromamba install -y -n base -c https://ftp.osuosl.org/pub/open-ce/1.11.0-p10/ -c defaults python=3.10 rust && micromamba clean --all --yes
 
 COPY ./ /workspace/vllm
 
@@ -18,11 +18,9 @@ ARG GIT_REPO_CHECK=0
 RUN --mount=type=bind,source=.git,target=.git \
     if [ "$GIT_REPO_CHECK" != 0 ]; then bash tools/check_repo.sh; fi
 
-# These packages will be in rocketce eventually
 RUN --mount=type=cache,target=/root/.cache/pip  \
-    pip install -v --prefer-binary --extra-index-url https://repo.fury.io/mgiessing \
+    RUSTFLAGS='-L /opt/conda/lib' pip install -v --prefer-binary --extra-index-url https://repo.fury.io/mgiessing \
         'cmake>=3.26' ninja packaging 'setuptools-scm>=8' wheel jinja2 \
-        torch==2.3.1 \
         -r requirements-cpu.txt \
         xformers uvloop==0.20.0
 

Dockerfile.rocm

@@ -1,174 +1,119 @@
-# Default ROCm 6.2 base image
+# default base image
-ARG BASE_IMAGE="rocm/pytorch:rocm6.2_ubuntu20.04_py3.9_pytorch_release_2.3.0"
+ARG REMOTE_VLLM="0"
+ARG USE_CYTHON="0"
+ARG BUILD_RPD="1"
+ARG COMMON_WORKDIR=/app
+ARG BASE_IMAGE=rocm/vllm-dev:base
 
-# Default ROCm ARCHes to build vLLM for.
+FROM ${BASE_IMAGE} AS base
-ARG PYTORCH_ROCM_ARCH="gfx908;gfx90a;gfx942;gfx1100"
 
-# Whether to install CK-based flash-attention
+ARG ARG_PYTORCH_ROCM_ARCH
-# If 0, will not install flash-attention
+ENV PYTORCH_ROCM_ARCH=${ARG_PYTORCH_ROCM_ARCH:-${PYTORCH_ROCM_ARCH}}
-ARG BUILD_FA="1"
-ARG FA_GFX_ARCHS="gfx90a;gfx942"
-ARG FA_BRANCH="3cea2fb"
-
-# Whether to build triton on rocm
-ARG BUILD_TRITON="1"
-ARG TRITON_BRANCH="e192dba"
-
-### Base image build stage
-FROM $BASE_IMAGE AS base
-
-# Import arg(s) defined before this build stage
-ARG PYTORCH_ROCM_ARCH
 
 # Install some basic utilities
-RUN apt-get update && apt-get install python3 python3-pip -y
+RUN apt-get update -q -y && apt-get install -q -y \
-RUN apt-get update && apt-get install -y \
+    sqlite3 libsqlite3-dev libfmt-dev libmsgpack-dev libsuitesparse-dev
-    curl \
+# Remove sccache    
-    ca-certificates \
+RUN python3 -m pip install --upgrade pip && pip install setuptools_scm
-    sudo \
-    git \
-    bzip2 \
-    libx11-6 \
-    build-essential \
-    wget \
-    unzip \
-    tmux \
-    ccache \
- && rm -rf /var/lib/apt/lists/*
-
-# When launching the container, mount the code directory to /vllm-workspace
-ARG APP_MOUNT=/vllm-workspace
-WORKDIR ${APP_MOUNT}
-
-RUN python3 -m pip install --upgrade pip
-# Remove sccache so it doesn't interfere with ccache
-# TODO: implement sccache support across components
 RUN apt-get purge -y sccache; python3 -m pip uninstall -y sccache; rm -f "$(which sccache)"
-
+ARG COMMON_WORKDIR
-# Install torch == 2.6.0 on ROCm
+WORKDIR ${COMMON_WORKDIR}
-RUN --mount=type=cache,target=/root/.cache/pip \
-    case "$(ls /opt | grep -Po 'rocm-[0-9]\.[0-9]')" in \
-        *"rocm-6.2"*) \
-            python3 -m pip uninstall -y torch torchvision \
-            && python3 -m pip install --pre \
-                torch==2.6.0.dev20241113+rocm6.2 \
-                'setuptools-scm>=8' \
-                torchvision==0.20.0.dev20241113+rocm6.2 \
-                --extra-index-url https://download.pytorch.org/whl/nightly/rocm6.2;; \
-        *) ;; esac
-
-ENV LLVM_SYMBOLIZER_PATH=/opt/rocm/llvm/bin/llvm-symbolizer
-ENV PATH=$PATH:/opt/rocm/bin:/libtorch/bin:
-ENV LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/opt/rocm/lib/:/libtorch/lib:
-ENV CPLUS_INCLUDE_PATH=$CPLUS_INCLUDE_PATH:/libtorch/include:/libtorch/include/torch/csrc/api/include/:/opt/rocm/include/:
-
-ENV PYTORCH_ROCM_ARCH=${PYTORCH_ROCM_ARCH}
-ENV CCACHE_DIR=/root/.cache/ccache
 
 
-### AMD-SMI build stage
+# -----------------------
-FROM base AS build_amdsmi
+# vLLM fetch stages
-# Build amdsmi wheel always
+FROM base AS fetch_vllm_0
-RUN cd /opt/rocm/share/amd_smi \
+ONBUILD COPY ./ vllm/
-    && python3 -m pip wheel . --wheel-dir=/install
+FROM base AS fetch_vllm_1
+ARG VLLM_REPO="https://github.com/vllm-project/vllm.git"
+ARG VLLM_BRANCH="main"
+ONBUILD RUN git clone ${VLLM_REPO} \
+	    && cd vllm \
+	    && git checkout ${VLLM_BRANCH}
+FROM fetch_vllm_${REMOTE_VLLM} AS fetch_vllm
 
-
+# -----------------------
-### Flash-Attention wheel build stage
+# vLLM build stages
-FROM base AS build_fa
+FROM fetch_vllm AS build_vllm
-ARG BUILD_FA
+ARG USE_CYTHON
-ARG FA_GFX_ARCHS
+# Build vLLM
-ARG FA_BRANCH
+RUN cd vllm \
-# Build ROCm flash-attention wheel if `BUILD_FA = 1`
+    && python3 -m pip install -r requirements-rocm.txt \
-RUN --mount=type=cache,target=${CCACHE_DIR} \
-    if [ "$BUILD_FA" = "1" ]; then \
-        mkdir -p libs \
-        && cd libs \
-        && git clone https://github.com/ROCm/flash-attention.git \
-        && cd flash-attention \
-        && git checkout "${FA_BRANCH}" \
-        && git submodule update --init \
-        && GPU_ARCHS="${FA_GFX_ARCHS}" python3 setup.py bdist_wheel --dist-dir=/install; \
-    # Create an empty directory otherwise as later build stages expect one
-    else mkdir -p /install; \
-    fi
-
-
-### Triton wheel build stage
-FROM base AS build_triton
-ARG BUILD_TRITON
-ARG TRITON_BRANCH
-# Build triton wheel if `BUILD_TRITON = 1`
-RUN --mount=type=cache,target=${CCACHE_DIR} \
-    if [ "$BUILD_TRITON" = "1" ]; then \
-    mkdir -p libs \
-    && cd libs \
-    && python3 -m pip install ninja cmake wheel pybind11 \
-    && git clone https://github.com/OpenAI/triton.git \
-    && cd triton \
-    && git checkout "${TRITON_BRANCH}" \
-    && cd python \
-    && python3 setup.py bdist_wheel --dist-dir=/install; \
-    # Create an empty directory otherwise as later build stages expect one
-    else mkdir -p /install; \
-    fi
-
-
-### Final vLLM build stage
-FROM base AS final
-# Import the vLLM development directory from the build context
-COPY . .
-ARG GIT_REPO_CHECK=0
-RUN --mount=type=bind,source=.git,target=.git \
-    if [ "$GIT_REPO_CHECK" != 0 ]; then bash tools/check_repo.sh ; fi
-
-RUN python3 -m pip install --upgrade pip
-
-# Package upgrades for useful functionality or to avoid dependency issues
-RUN --mount=type=cache,target=/root/.cache/pip \
-    python3 -m pip install --upgrade numba scipy huggingface-hub[cli] pytest-shard
-
-
-# Workaround for ray >= 2.10.0
-ENV RAY_EXPERIMENTAL_NOSET_ROCR_VISIBLE_DEVICES=1
-# Silences the HF Tokenizers warning
-ENV TOKENIZERS_PARALLELISM=false
-
-RUN --mount=type=cache,target=${CCACHE_DIR} \
-    --mount=type=bind,source=.git,target=.git \
-    --mount=type=cache,target=/root/.cache/pip \
-    python3 -m pip install -Ur requirements-rocm.txt \
     && python3 setup.py clean --all  \
-    && python3 setup.py develop
+    && if [ ${USE_CYTHON} -eq "1" ]; then python3 setup_cython.py build_ext --inplace; fi \
+    && python3 setup.py bdist_wheel --dist-dir=dist
+FROM scratch AS export_vllm
+ARG COMMON_WORKDIR
+COPY --from=build_vllm ${COMMON_WORKDIR}/vllm/dist/*.whl /
+COPY --from=build_vllm ${COMMON_WORKDIR}/vllm/requirements*.txt /
+COPY --from=build_vllm ${COMMON_WORKDIR}/vllm/benchmarks /benchmarks
+COPY --from=build_vllm ${COMMON_WORKDIR}/vllm/tests /tests
+COPY --from=build_vllm ${COMMON_WORKDIR}/vllm/examples /examples
+COPY --from=build_vllm ${COMMON_WORKDIR}/vllm/.buildkite /.buildkite
 
-# Copy amdsmi wheel into final image
+# -----------------------
-RUN --mount=type=bind,from=build_amdsmi,src=/install,target=/install \
+# Test vLLM image
-    mkdir -p libs \
+FROM base AS test
-    && cp /install/*.whl libs \
-    # Preemptively uninstall to avoid same-version no-installs
-    && python3 -m pip uninstall -y amdsmi;
 
-# Copy triton wheel(s) into final image if they were built
+RUN python3 -m pip install --upgrade pip && rm -rf /var/lib/apt/lists/*
-RUN --mount=type=bind,from=build_triton,src=/install,target=/install \
-    mkdir -p libs \
-    && if ls /install/*.whl; then \
-        cp /install/*.whl libs \
-        # Preemptively uninstall to avoid same-version no-installs
-        && python3 -m pip uninstall -y triton; fi
 
-# Copy flash-attn wheel(s) into final image if they were built
+# Install vLLM
-RUN --mount=type=bind,from=build_fa,src=/install,target=/install \
+RUN --mount=type=bind,from=export_vllm,src=/,target=/install \
-    mkdir -p libs \
+    cd /install \
-    && if ls /install/*.whl; then \
+    && pip install -U -r requirements-rocm.txt \
-        cp /install/*.whl libs \
+    && pip uninstall -y vllm \
-        # Preemptively uninstall to avoid same-version no-installs
+    && pip install *.whl
-        && python3 -m pip uninstall -y flash-attn; fi
 
-# Install wheels that were built to the final image
+WORKDIR /vllm-workspace
-RUN --mount=type=cache,target=/root/.cache/pip \
+ARG COMMON_WORKDIR
-    if ls libs/*.whl; then \
+COPY --from=build_vllm ${COMMON_WORKDIR}/vllm /vllm-workspace
-    python3 -m pip install libs/*.whl; fi
 
 # install development dependencies (for testing)
-RUN python3 -m pip install -e tests/vllm_test_utils
+RUN cd /vllm-workspace \
+    && rm -rf vllm \
+    && python3 -m pip install -e tests/vllm_test_utils \
+    && python3 -m pip install lm-eval[api]==0.4.4 \
+    && python3 -m pip install pytest-shard
+
+# -----------------------
+# Final vLLM image
+FROM base AS final
+
+RUN python3 -m pip install --upgrade pip && rm -rf /var/lib/apt/lists/*
+# Error related to odd state for numpy 1.20.3 where there is no METADATA etc, but an extra LICENSES_bundled.txt.
+# Manually remove it so that later steps of numpy upgrade can continue
+RUN case "$(which python3)" in \
+        *"/opt/conda/envs/py_3.9"*) \
+            rm -rf /opt/conda/envs/py_3.9/lib/python3.9/site-packages/numpy-1.20.3.dist-info/;; \
+        *) ;; esac
+
+RUN python3 -m pip install --upgrade huggingface-hub[cli]
+ARG BUILD_RPD
+RUN if [ ${BUILD_RPD} -eq "1" ]; then \
+    git clone -b nvtx_enabled https://github.com/ROCm/rocmProfileData.git \
+    && cd rocmProfileData/rpd_tracer \
+    && pip install -r requirements.txt && cd ../ \
+    && make && make install \
+    && cd hipMarker && python3 setup.py install ; fi
+
+# Install vLLM
+RUN --mount=type=bind,from=export_vllm,src=/,target=/install \
+    cd /install \
+    && pip install -U -r requirements-rocm.txt \
+    && pip uninstall -y vllm \
+    && pip install *.whl
+
+ARG COMMON_WORKDIR
+
+# Copy over the benchmark scripts as well
+COPY --from=export_vllm /benchmarks ${COMMON_WORKDIR}/vllm/benchmarks
+COPY --from=export_vllm /examples ${COMMON_WORKDIR}/vllm/examples
+
+ENV RAY_EXPERIMENTAL_NOSET_ROCR_VISIBLE_DEVICES=1
+ENV TOKENIZERS_PARALLELISM=false
+
+# Performance environment variable.
+ENV HIP_FORCE_DEV_KERNARG=1
 
 CMD ["/bin/bash"]
+

Dockerfile.rocm_base

@@ -0,0 +1,158 @@
+ARG BASE_IMAGE=rocm/dev-ubuntu-22.04:6.3.1-complete
+ARG HIPBLASLT_BRANCH="4d40e36"
+ARG HIPBLAS_COMMON_BRANCH="7c1566b"
+ARG LEGACY_HIPBLASLT_OPTION=
+ARG RCCL_BRANCH="648a58d"
+ARG RCCL_REPO="https://github.com/ROCm/rccl"
+ARG TRITON_BRANCH="e5be006"
+ARG TRITON_REPO="https://github.com/triton-lang/triton.git"
+ARG PYTORCH_BRANCH="8d4926e"
+ARG PYTORCH_VISION_BRANCH="v0.19.1"
+ARG PYTORCH_REPO="https://github.com/pytorch/pytorch.git"
+ARG PYTORCH_VISION_REPO="https://github.com/pytorch/vision.git"
+ARG FA_BRANCH="b7d29fb"
+ARG FA_REPO="https://github.com/ROCm/flash-attention.git"
+
+FROM ${BASE_IMAGE} AS base
+
+ENV PATH=/opt/rocm/llvm/bin:$PATH
+ENV ROCM_PATH=/opt/rocm
+ENV LD_LIBRARY_PATH=/opt/rocm/lib:/usr/local/lib:
+ARG PYTORCH_ROCM_ARCH=gfx90a;gfx942
+ENV PYTORCH_ROCM_ARCH=${PYTORCH_ROCM_ARCH}
+
+ARG PYTHON_VERSION=3.12
+
+RUN mkdir -p /app
+WORKDIR /app
+ENV DEBIAN_FRONTEND=noninteractive
+
+# Install Python and other dependencies
+RUN apt-get update -y \
+    && apt-get install -y software-properties-common git curl sudo vim less \
+    && add-apt-repository ppa:deadsnakes/ppa \
+    && apt-get update -y \
+    && apt-get install -y python${PYTHON_VERSION} python${PYTHON_VERSION}-dev python${PYTHON_VERSION}-venv \
+       python${PYTHON_VERSION}-lib2to3 python-is-python3  \
+    && update-alternatives --install /usr/bin/python3 python3 /usr/bin/python${PYTHON_VERSION} 1 \
+    && update-alternatives --set python3 /usr/bin/python${PYTHON_VERSION} \
+    && ln -sf /usr/bin/python${PYTHON_VERSION}-config /usr/bin/python3-config \
+    && curl -sS https://bootstrap.pypa.io/get-pip.py | python${PYTHON_VERSION} \
+    && python3 --version && python3 -m pip --version
+
+RUN pip install -U packaging cmake ninja wheel setuptools pybind11 Cython
+
+FROM base AS build_hipblaslt
+ARG HIPBLASLT_BRANCH
+ARG HIPBLAS_COMMON_BRANCH
+# Set to "--legacy_hipblas_direct" for ROCm<=6.2
+ARG LEGACY_HIPBLASLT_OPTION
+RUN git clone https://github.com/ROCm/hipBLAS-common.git
+RUN cd hipBLAS-common \
+    && git checkout ${HIPBLAS_COMMON_BRANCH} \
+    && mkdir build \
+    && cd build \
+    && cmake .. \
+    && make package \
+    && dpkg -i ./*.deb
+RUN git clone https://github.com/ROCm/hipBLASLt
+RUN cd hipBLASLt \
+    && git checkout ${HIPBLASLT_BRANCH} \
+    && ./install.sh -d --architecture ${PYTORCH_ROCM_ARCH} ${LEGACY_HIPBLASLT_OPTION} \
+    && cd build/release \
+    && make package
+RUN mkdir -p /app/install && cp /app/hipBLASLt/build/release/*.deb /app/hipBLAS-common/build/*.deb /app/install
+
+FROM base AS build_rccl
+ARG RCCL_BRANCH
+ARG RCCL_REPO
+RUN git clone ${RCCL_REPO}
+RUN cd rccl \
+    && git checkout ${RCCL_BRANCH} \
+    && ./install.sh -p --amdgpu_targets ${PYTORCH_ROCM_ARCH}
+RUN mkdir -p /app/install && cp /app/rccl/build/release/*.deb /app/install
+
+FROM base AS build_triton
+ARG TRITON_BRANCH
+ARG TRITON_REPO
+RUN git clone ${TRITON_REPO}
+RUN cd triton \
+    && git checkout ${TRITON_BRANCH} \
+    && cd python \
+    && python3 setup.py bdist_wheel --dist-dir=dist
+RUN mkdir -p /app/install && cp /app/triton/python/dist/*.whl /app/install
+
+FROM base AS build_amdsmi
+RUN cd /opt/rocm/share/amd_smi \
+    && pip wheel . --wheel-dir=dist
+RUN mkdir -p /app/install && cp /opt/rocm/share/amd_smi/dist/*.whl /app/install
+
+FROM base AS build_pytorch
+ARG PYTORCH_BRANCH
+ARG PYTORCH_VISION_BRANCH
+ARG PYTORCH_REPO
+ARG PYTORCH_VISION_REPO
+ARG FA_BRANCH
+ARG FA_REPO
+RUN git clone ${PYTORCH_REPO} pytorch
+RUN cd pytorch && git checkout ${PYTORCH_BRANCH} && \
+    pip install -r requirements.txt && git submodule update --init --recursive \
+    && python3 tools/amd_build/build_amd.py \
+    && CMAKE_PREFIX_PATH=$(python3 -c 'import sys; print(sys.prefix)') python3 setup.py bdist_wheel --dist-dir=dist \
+    && pip install dist/*.whl
+RUN git clone ${PYTORCH_VISION_REPO} vision
+RUN cd vision && git checkout ${PYTORCH_VISION_BRANCH} \
+    && python3 setup.py bdist_wheel --dist-dir=dist \
+    && pip install dist/*.whl
+RUN git clone ${FA_REPO}
+RUN cd flash-attention \
+    && git checkout ${FA_BRANCH} \
+    && git submodule update --init \
+    && MAX_JOBS=64 GPU_ARCHS=${PYTORCH_ROCM_ARCH} python3 setup.py bdist_wheel --dist-dir=dist
+RUN mkdir -p /app/install && cp /app/pytorch/dist/*.whl /app/install \
+    && cp /app/vision/dist/*.whl /app/install \
+    && cp /app/flash-attention/dist/*.whl /app/install
+
+FROM base AS final
+RUN --mount=type=bind,from=build_hipblaslt,src=/app/install/,target=/install \
+    dpkg -i /install/*deb \
+    && sed -i 's/, hipblaslt-dev \(.*\), hipcub-dev/, hipcub-dev/g' /var/lib/dpkg/status \
+    && sed -i 's/, hipblaslt \(.*\), hipfft/, hipfft/g' /var/lib/dpkg/status
+RUN --mount=type=bind,from=build_rccl,src=/app/install/,target=/install \
+    dpkg -i /install/*deb \
+    && sed -i 's/, rccl-dev \(.*\), rocalution/, rocalution/g' /var/lib/dpkg/status \
+    && sed -i 's/, rccl \(.*\), rocalution/, rocalution/g' /var/lib/dpkg/status
+RUN --mount=type=bind,from=build_triton,src=/app/install/,target=/install \
+    pip install /install/*.whl
+RUN --mount=type=bind,from=build_amdsmi,src=/app/install/,target=/install \
+    pip install /install/*.whl
+RUN --mount=type=bind,from=build_pytorch,src=/app/install/,target=/install \
+    pip install /install/*.whl
+
+ARG BASE_IMAGE
+ARG HIPBLASLT_BRANCH
+ARG LEGACY_HIPBLASLT_OPTION
+ARG RCCL_BRANCH
+ARG RCCL_REPO
+ARG TRITON_BRANCH
+ARG TRITON_REPO
+ARG PYTORCH_BRANCH
+ARG PYTORCH_VISION_BRANCH
+ARG PYTORCH_REPO
+ARG PYTORCH_VISION_REPO
+ARG FA_BRANCH
+ARG FA_REPO
+RUN echo "BASE_IMAGE: ${BASE_IMAGE}" > /app/versions.txt \
+    && echo "HIPBLAS_COMMON_BRANCH: ${HIPBLAS_COMMON_BRANCH}" >> /app/versions.txt \
+    && echo "HIPBLASLT_BRANCH: ${HIPBLASLT_BRANCH}" >> /app/versions.txt \
+    && echo "LEGACY_HIPBLASLT_OPTION: ${LEGACY_HIPBLASLT_OPTION}" >> /app/versions.txt \
+    && echo "RCCL_BRANCH: ${RCCL_BRANCH}" >> /app/versions.txt \
+    && echo "RCCL_REPO: ${RCCL_REPO}" >> /app/versions.txt \
+    && echo "TRITON_BRANCH: ${TRITON_BRANCH}" >> /app/versions.txt \
+    && echo "TRITON_REPO: ${TRITON_REPO}" >> /app/versions.txt \
+    && echo "PYTORCH_BRANCH: ${PYTORCH_BRANCH}" >> /app/versions.txt \
+    && echo "PYTORCH_VISION_BRANCH: ${PYTORCH_VISION_BRANCH}" >> /app/versions.txt \
+    && echo "PYTORCH_REPO: ${PYTORCH_REPO}" >> /app/versions.txt \
+    && echo "PYTORCH_VISION_REPO: ${PYTORCH_VISION_REPO}" >> /app/versions.txt \
+    && echo "FA_BRANCH: ${FA_BRANCH}" >> /app/versions.txt \
+    && echo "FA_REPO: ${FA_REPO}" >> /app/versions.txt

Dockerfile.tpu

@@ -1,4 +1,4 @@
-ARG NIGHTLY_DATE="20241017"
+ARG NIGHTLY_DATE="20250124"
 ARG BASE_IMAGE="us-central1-docker.pkg.dev/tpu-pytorch-releases/docker/xla:nightly_3.10_tpuvm_$NIGHTLY_DATE"
 
 FROM $BASE_IMAGE

README.md

@@ -10,12 +10,14 @@ Easy, fast, and cheap LLM serving for everyone
 </h3>
 
 <p align="center">
-| <a href="https://docs.vllm.ai"><b>Documentation</b></a> | <a href="https://vllm.ai"><b>Blog</b></a> | <a href="https://arxiv.org/abs/2309.06180"><b>Paper</b></a> | <a href="https://discord.gg/jz7wjKhh6g"><b>Discord</b></a> | <a href="https://x.com/vllm_project"><b>Twitter/X</b></a> | <a href="https://slack.vllm.ai"><b>Developer Slack</b></a> |
+| <a href="https://docs.vllm.ai"><b>Documentation</b></a> | <a href="https://vllm.ai"><b>Blog</b></a> | <a href="https://arxiv.org/abs/2309.06180"><b>Paper</b></a> | <a href="https://x.com/vllm_project"><b>Twitter/X</b></a> | <a href="https://slack.vllm.ai"><b>Developer Slack</b></a> |
 </p>
 
 ---
 
 *Latest News* 🔥
+- [2025/01] We are excited to announce the alpha release of vLLM V1: A major architectural upgrade with 1.7x speedup! Clean code, optimized execution loop, zero-overhead prefix caching, enhanced multimodal support, and more. Please check out our blog post [here](https://blog.vllm.ai/2025/01/27/v1-alpha-release.html).
+- [2025/01] We hosted [the eighth vLLM meetup](https://lu.ma/zep56hui) with Google Cloud! Please find the meetup slides from vLLM team [here](https://docs.google.com/presentation/d/1epVkt4Zu8Jz_S5OhEHPc798emsYh2BwYfRuDDVEF7u4/edit?usp=sharing).
 - [2024/12] vLLM joins [pytorch ecosystem](https://pytorch.org/blog/vllm-joins-pytorch)! Easy, Fast, and Cheap LLM Serving for Everyone!
 - [2024/11] We hosted [the seventh vLLM meetup](https://lu.ma/h0qvrajz) with Snowflake! Please find the meetup slides from vLLM team [here](https://docs.google.com/presentation/d/1e3CxQBV3JsfGp30SwyvS3eM_tW-ghOhJ9PAJGK6KR54/edit?usp=sharing), and Snowflake team [here](https://docs.google.com/presentation/d/1qF3RkDAbOULwz9WK5TOltt2fE9t6uIc_hVNLFAaQX6A/edit?usp=sharing).
 - [2024/10] We have just created a developer slack ([slack.vllm.ai](https://slack.vllm.ai)) focusing on coordinating contributions and discussing features. Please feel free to join us there!
@@ -34,10 +36,12 @@ Easy, fast, and cheap LLM serving for everyone
 ## About
 vLLM is a fast and easy-to-use library for LLM inference and serving.
 
+Originally developed in the [Sky Computing Lab](https://sky.cs.berkeley.edu) at UC Berkeley, vLLM has evolved into a community-driven project with contributions from both academia and industry.
+
 vLLM is fast with:
 
 - State-of-the-art serving throughput
-- Efficient management of attention key and value memory with **PagedAttention**
+- 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), INT4, INT8, and FP8.
@@ -60,7 +64,7 @@ vLLM is flexible and easy to use with:
 
 vLLM seamlessly supports most popular open-source models on HuggingFace, including:
 - Transformer-like LLMs (e.g., Llama)
-- Mixture-of-Expert LLMs (e.g., Mixtral)
+- Mixture-of-Expert LLMs (e.g., Mixtral, Deepseek-V2 and V3)
 - Embedding Models (e.g. E5-Mistral)
 - Multi-modal LLMs (e.g., LLaVA)
 
@@ -68,16 +72,16 @@ Find the full list of supported models [here](https://docs.vllm.ai/en/latest/mod
 
 ## Getting Started
 
-Install vLLM with `pip` or [from source](https://vllm.readthedocs.io/en/latest/getting_started/installation.html#build-from-source):
+Install vLLM with `pip` or [from source](https://docs.vllm.ai/en/latest/getting_started/installation/gpu/index.html#build-wheel-from-source):
 
 ```bash
 pip install vllm
 ```
 
-Visit our [documentation](https://vllm.readthedocs.io/en/latest/) to learn more.
+Visit our [documentation](https://docs.vllm.ai/en/latest/) to learn more.
-- [Installation](https://vllm.readthedocs.io/en/latest/getting_started/installation.html)
+- [Installation](https://docs.vllm.ai/en/latest/getting_started/installation/index.html)
-- [Quickstart](https://vllm.readthedocs.io/en/latest/getting_started/quickstart.html)
+- [Quickstart](https://docs.vllm.ai/en/latest/getting_started/quickstart.html)
-- [Supported Models](https://vllm.readthedocs.io/en/latest/models/supported_models.html)
+- [List of Supported Models](https://docs.vllm.ai/en/latest/models/supported_models.html)
 
 ## Contributing
 
@@ -90,28 +94,33 @@ vLLM is a community project. Our compute resources for development and testing a
 
 <!-- Note: Please sort them in alphabetical order. -->
 <!-- Note: Please keep these consistent with docs/source/community/sponsors.md -->
-
+Cash Donations:
 - a16z
+- Dropbox
+- Sequoia Capital
+- Skywork AI
+- ZhenFund
+
+Compute Resources:
 - AMD
 - Anyscale
 - AWS
 - Crusoe Cloud
 - Databricks
 - DeepInfra
-- Dropbox
 - Google Cloud
 - Lambda Lab
 - Nebius
+- Novita AI
 - NVIDIA
 - Replicate
 - Roblox
 - RunPod
-- Sequoia Capital
-- Skywork AI
 - Trainy
 - UC Berkeley
 - UC San Diego
-- ZhenFund
+
+Slack Sponsor: Anyscale
 
 We also have an official fundraising venue through [OpenCollective](https://opencollective.com/vllm). We plan to use the fund to support the development, maintenance, and adoption of vLLM.
 
@@ -130,8 +139,7 @@ If you use vLLM for your research, please cite our [paper](https://arxiv.org/abs
 ## Contact Us
 
 * For technical questions and feature requests, please use Github issues or discussions.
-* For discussing with fellow users, please use Discord.
+* For discussing with fellow users and coordinating contributions and development, please use Slack.
-* For coordinating contributions and development, please use Slack.
 * For security disclosures, please use Github's security advisory feature.
 * For collaborations and partnerships, please contact us at vllm-questions AT lists.berkeley.edu.
 

SECURITY.md

@@ -4,7 +4,7 @@
 
 If you believe you have found a security vulnerability in vLLM, we encourage you to let us know right away. We will investigate all legitimate reports and do our best to quickly fix the problem.
 
-Please report security issues privately using [the vulnerability submission form](https://github.com/vllm-project/vllm/security/advisories/new).
+Please report security issues privately using [the vulnerability submission form](https://github.com/vllm-project/vllm/security/advisories/new). Reports will then be triaged by the [vulnerability management team](https://docs.vllm.ai/en/latest/contributing/vulnerability_management.html).
 
 ---
 

benchmarks/backend_request_func.py

@@ -1,3 +1,5 @@
+# SPDX-License-Identifier: Apache-2.0
+
 import json
 import os
 import sys
@@ -22,6 +24,7 @@ class RequestFuncInput:
     prompt_len: int
     output_len: int
     model: str
+    model_name: Optional[str] = None
     best_of: int = 1
     logprobs: Optional[int] = None
     extra_body: Optional[dict] = None
@@ -34,6 +37,7 @@ class RequestFuncOutput:
     generated_text: str = ""
     success: bool = False
     latency: float = 0.0
+    output_tokens: int = 0
     ttft: float = 0.0  # Time to first token
     itl: List[float] = field(
         default_factory=list)  # List of inter-token latencies
@@ -49,7 +53,8 @@ async def async_request_tgi(
     api_url = request_func_input.api_url
     assert api_url.endswith("generate_stream")
 
-    async with aiohttp.ClientSession(timeout=AIOHTTP_TIMEOUT) as session:
+    async with aiohttp.ClientSession(trust_env=True,
+                                     timeout=AIOHTTP_TIMEOUT) as session:
         params = {
             "best_of": request_func_input.best_of,
             "max_new_tokens": request_func_input.output_len,
@@ -121,7 +126,8 @@ async def async_request_trt_llm(
     api_url = request_func_input.api_url
     assert api_url.endswith("generate_stream")
 
-    async with aiohttp.ClientSession(timeout=AIOHTTP_TIMEOUT) as session:
+    async with aiohttp.ClientSession(trust_env=True,
+                                     timeout=AIOHTTP_TIMEOUT) as session:
         assert request_func_input.best_of == 1
         payload = {
             "accumulate_tokens": True,
@@ -155,7 +161,7 @@ async def async_request_trt_llm(
                         timestamp = time.perf_counter()
                         # First token
                         if ttft == 0.0:
-                            ttft = time.perf_counter() - st
+                            ttft = timestamp - st
                             output.ttft = ttft
 
                         # Decoding phase
@@ -185,7 +191,8 @@ async def async_request_deepspeed_mii(
     request_func_input: RequestFuncInput,
     pbar: Optional[tqdm] = None,
 ) -> RequestFuncOutput:
-    async with aiohttp.ClientSession(timeout=AIOHTTP_TIMEOUT) as session:
+    async with aiohttp.ClientSession(trust_env=True,
+                                     timeout=AIOHTTP_TIMEOUT) as session:
         assert request_func_input.best_of == 1
 
         payload = {
@@ -233,17 +240,23 @@ async def async_request_openai_completions(
         ("completions", "profile")
     ), "OpenAI Completions API URL must end with 'completions' or 'profile'."
 
-    async with aiohttp.ClientSession(timeout=AIOHTTP_TIMEOUT) as session:
+    async with aiohttp.ClientSession(trust_env=True,
+                                     timeout=AIOHTTP_TIMEOUT) as session:
         payload = {
-            "model": request_func_input.model,
+            "model": request_func_input.model_name \
+                if request_func_input.model_name else request_func_input.model,
             "prompt": request_func_input.prompt,
             "temperature": 0.0,
             "best_of": request_func_input.best_of,
             "max_tokens": request_func_input.output_len,
             "logprobs": request_func_input.logprobs,
             "stream": True,
-            "ignore_eos": request_func_input.ignore_eos,
+            "stream_options": {
+                "include_usage": True,
+            },
         }
+        if request_func_input.ignore_eos:
+            payload["ignore_eos"] = request_func_input.ignore_eos
         if request_func_input.extra_body:
             payload.update(request_func_input.extra_body)
         headers = {
@@ -254,7 +267,6 @@ async def async_request_openai_completions(
         output.prompt_len = request_func_input.prompt_len
 
         generated_text = ""
-        ttft = 0.0
         st = time.perf_counter()
         most_recent_timestamp = st
         try:
@@ -269,15 +281,16 @@ async def async_request_openai_completions(
 
                         chunk = chunk_bytes.decode("utf-8").removeprefix(
                             "data: ")
-                        if chunk == "[DONE]":
+                        if chunk != "[DONE]":
-                            latency = time.perf_counter() - st
-                        else:
                             data = json.loads(chunk)
 
                             # NOTE: Some completion API might have a last
                             # usage summary response without a token so we
                             # want to check a token was generated
-                            if data["choices"][0]["text"]:
+                            if choices := data.get("choices"):
+                                # Note that text could be empty here
+                                # e.g. for special tokens
+                                text = choices[0].get("text")
                                 timestamp = time.perf_counter()
                                 # First token
                                 if not first_chunk_received:
@@ -291,7 +304,10 @@ async def async_request_openai_completions(
                                                       most_recent_timestamp)
 
                                 most_recent_timestamp = timestamp
-                                generated_text += data["choices"][0]["text"]
+                                generated_text += text or ""
+                            elif usage := data.get("usage"):
+                                output.output_tokens = usage.get(
+                                    "completion_tokens")
                     if first_chunk_received:
                         output.success = True
                     else:
@@ -300,7 +316,7 @@ async def async_request_openai_completions(
                             "Never received a valid chunk to calculate TTFT."
                             "This response will be marked as failed!")
                     output.generated_text = generated_text
-                    output.latency = latency
+                    output.latency = most_recent_timestamp - st
                 else:
                     output.error = response.reason or ""
                     output.success = False
@@ -323,12 +339,14 @@ async def async_request_openai_chat_completions(
         "chat/completions"
     ), "OpenAI Chat Completions API URL must end with 'chat/completions'."
 
-    async with aiohttp.ClientSession(timeout=AIOHTTP_TIMEOUT) as session:
+    async with aiohttp.ClientSession(trust_env=True,
+                                     timeout=AIOHTTP_TIMEOUT) as session:
         content = [{"type": "text", "text": request_func_input.prompt}]
         if request_func_input.multi_modal_content:
             content.append(request_func_input.multi_modal_content)
         payload = {
-            "model": request_func_input.model,
+            "model": request_func_input.model_name \
+                if request_func_input.model_name else request_func_input.model,
             "messages": [
                 {
                     "role": "user",
@@ -338,8 +356,12 @@ async def async_request_openai_chat_completions(
             "temperature": 0.0,
             "max_completion_tokens": request_func_input.output_len,
             "stream": True,
-            "ignore_eos": request_func_input.ignore_eos,
+            "stream_options": {
+                "include_usage": True,
+            },
         }
+        if request_func_input.ignore_eos:
+            payload["ignore_eos"] = request_func_input.ignore_eos
         if request_func_input.extra_body:
             payload.update(request_func_input.extra_body)
         headers = {
@@ -365,17 +387,15 @@ async def async_request_openai_chat_completions(
 
                         chunk = chunk_bytes.decode("utf-8").removeprefix(
                             "data: ")
-                        if chunk == "[DONE]":
+                        if chunk != "[DONE]":
-                            latency = time.perf_counter() - st
-                        else:
                             timestamp = time.perf_counter()
                             data = json.loads(chunk)
 
-                            delta = data["choices"][0]["delta"]
+                            if choices := data.get("choices"):
-                            if delta.get("content", None):
+                                content = choices[0]["delta"].get("content")
                                 # First token
                                 if ttft == 0.0:
-                                    ttft = time.perf_counter() - st
+                                    ttft = timestamp - st
                                     output.ttft = ttft
 
                                 # Decoding phase
@@ -383,13 +403,16 @@ async def async_request_openai_chat_completions(
                                     output.itl.append(timestamp -
                                                       most_recent_timestamp)
 
-                                generated_text += delta["content"]
+                                generated_text += content or ""
+                            elif usage := data.get("usage"):
+                                output.output_tokens = usage.get(
+                                    "completion_tokens")
 
                             most_recent_timestamp = timestamp
 
                     output.generated_text = generated_text
                     output.success = True
-                    output.latency = latency
+                    output.latency = most_recent_timestamp - st
                 else:
                     output.error = response.reason or ""
                     output.success = False
@@ -417,14 +440,35 @@ def get_model(pretrained_model_name_or_path: str) -> str:
 
 
 def get_tokenizer(
-    pretrained_model_name_or_path: str, trust_remote_code: bool
+    pretrained_model_name_or_path: str,
+    tokenizer_mode: str = "auto",
+    trust_remote_code: bool = False,
+    **kwargs,
 ) -> Union[PreTrainedTokenizer, PreTrainedTokenizerFast]:
     if pretrained_model_name_or_path is not None and not os.path.exists(
             pretrained_model_name_or_path):
         pretrained_model_name_or_path = get_model(
             pretrained_model_name_or_path)
-    return AutoTokenizer.from_pretrained(pretrained_model_name_or_path,
+    if tokenizer_mode == "slow":
-                                         trust_remote_code=trust_remote_code)
+        if kwargs.get("use_fast", False):
+            raise ValueError(
+                "Cannot use the fast tokenizer in slow tokenizer mode.")
+        kwargs["use_fast"] = False
+    if tokenizer_mode == "mistral":
+        try:
+            from vllm.transformers_utils.tokenizer import MistralTokenizer
+        except ImportError as e:
+            raise ImportError("MistralTokenizer requires vllm package.\n"
+                              "Please install it with `pip install vllm` "
+                              "to use mistral tokenizer mode.") from e
+        return MistralTokenizer.from_pretrained(
+            str(pretrained_model_name_or_path))
+    else:
+        return AutoTokenizer.from_pretrained(
+            pretrained_model_name_or_path,
+            trust_remote_code=trust_remote_code,
+            **kwargs,
+        )
 
 
 ASYNC_REQUEST_FUNCS = {

benchmarks/benchmark_guided.py

@@ -1,3 +1,4 @@
+# SPDX-License-Identifier: Apache-2.0
 """Benchmark guided decoding throughput."""
 import argparse
 import dataclasses

benchmarks/benchmark_latency.py

@@ -1,3 +1,4 @@
+# SPDX-License-Identifier: Apache-2.0
 """Benchmark the latency of processing a single batch of requests."""
 import argparse
 import dataclasses
@@ -13,6 +14,7 @@ from tqdm import tqdm
 from vllm import LLM, SamplingParams
 from vllm.engine.arg_utils import EngineArgs
 from vllm.inputs import PromptType
+from vllm.sampling_params import BeamSearchParams
 from vllm.utils import FlexibleArgumentParser
 
 
@@ -40,6 +42,20 @@ def main(args: argparse.Namespace):
         "prompt_token_ids": batch
     } for batch in dummy_prompt_token_ids.tolist()]
 
+    def llm_generate():
+        if not args.use_beam_search:
+            llm.generate(dummy_prompts,
+                         sampling_params=sampling_params,
+                         use_tqdm=False)
+        else:
+            llm.beam_search(
+                dummy_prompts,
+                BeamSearchParams(
+                    beam_width=args.n,
+                    max_tokens=args.output_len,
+                    ignore_eos=True,
+                ))
+
     def run_to_completion(profile_dir: Optional[str] = None):
         if profile_dir:
             with torch.profiler.profile(
@@ -49,15 +65,11 @@ def main(args: argparse.Namespace):
                     ],
                     on_trace_ready=torch.profiler.tensorboard_trace_handler(
                         str(profile_dir))) as p:
-                llm.generate(dummy_prompts,
+                llm_generate()
-                             sampling_params=sampling_params,
+            print(p.key_averages().table(sort_by="self_cuda_time_total"))
-                             use_tqdm=False)
-            print(p.key_averages())
         else:
             start_time = time.perf_counter()
-            llm.generate(dummy_prompts,
+            llm_generate()
-                         sampling_params=sampling_params,
-                         use_tqdm=False)
             end_time = time.perf_counter()
             latency = end_time - start_time
             return latency

benchmarks/benchmark_long_document_qa_throughput.py

@@ -0,0 +1,184 @@
+# SPDX-License-Identifier: Apache-2.0
+"""
+Offline benchmark to test the long document QA throughput.
+
+Example usage:
+    # This workload samples 8 different prompts with a default input
+    # length of 20000 tokens, then replicates each prompt 2 times 
+    # in random order.
+    python benchmark_long_document_qa_throughput.py \
+        --model meta-llama/Llama-2-7b-chat-hf \
+        --enable-prefix-caching \
+        --num-documents 8 \
+        --repeat-count 2 
+
+Commandline arguments:
+    --num-documents: The number of documents to sample prompts from.
+
+    --document-length: The length of each document in tokens. 
+                       (Optional, default: 20000)
+
+    --output-len: The number of tokens to generate for each prompt.
+                  (Optional, default: 10)
+
+    --repeat-count: The number of times to repeat each prompt.
+                    (Optional, default: 2)
+
+    --repeat-mode: The mode to repeat prompts. The supported modes are:
+        - 'random': shuffle the prompts randomly. (Default)
+        - 'tile': the entire prompt list is repeated in sequence. (Potentially
+                  lowest cache hit)
+        - 'interleave': each prompt is repeated consecutively before 
+                        moving to the next element. (Highest cache hit)
+    
+    --shuffle-seed: Random seed when the repeat mode is "random".
+                    (Optional, default: 0)
+
+In the meantime, it also supports all the vLLM engine args to initialize the 
+LLM engine. You can refer to the `vllm.engine.arg_utils.EngineArgs` for more
+details.
+"""
+
+import dataclasses
+import random
+import time
+
+from vllm import LLM, SamplingParams
+from vllm.engine.arg_utils import EngineArgs
+from vllm.utils import FlexibleArgumentParser
+
+
+def test_long_document_qa(llm=None, sampling_params=None, prompts=None):
+    """
+    Test long document QA with the given prompts and sampling parameters.
+    Print the time spent in processing all the prompts.
+
+    Args:
+        llm: The language model used for generating responses.
+        sampling_params: Sampling parameter used to generate the response.
+        prompts: A list of prompt strings to be processed by the LLM.
+    """
+    start_time = time.time()
+    llm.generate(prompts, sampling_params=sampling_params)
+    end_time = time.time()
+    print(f"Time to execute all requests: {end_time - start_time:.4f} secs")
+
+
+def repeat_prompts(prompts, repeat_count, mode: str):
+    """
+    Repeat each prompt in the list for a specified number of times.
+    The order of prompts in the output list depends on the mode.
+
+    Args:
+        prompts: A list of prompts to be repeated.
+        repeat_count: The number of times each prompt is repeated.
+        mode: The mode of repetition. Supported modes are:
+            - 'random': Shuffle the prompts randomly after repetition.
+            - 'tile': Repeat the entire prompt list in sequence.
+              Example: [1, 2, 3] -> [1, 2, 3, 1, 2, 3].
+            - 'interleave': Repeat each prompt consecutively before moving to 
+              the next. Example: [1, 2, 3] -> [1, 1, 2, 2, 3, 3].
+
+    Returns:
+        A list of repeated prompts in the specified order.
+
+    Raises:
+        ValueError: If an invalid mode is provided.
+    """
+    print("Repeat mode: ", mode)
+    if mode == 'random':
+        repeated_prompts = prompts * repeat_count
+        random.shuffle(repeated_prompts)
+        return repeated_prompts
+    elif mode == 'tile':
+        return prompts * repeat_count
+    elif mode == 'interleave':
+        repeated_prompts = []
+        for prompt in prompts:
+            repeated_prompts.extend([prompt] * repeat_count)
+        return repeated_prompts
+    else:
+        raise ValueError(f"Invalid mode: {mode}, only support "
+                         "'random', 'tile', 'interleave'")
+
+
+def main(args):
+    random.seed(args.shuffle_seed)
+
+    # Prepare the prompts:
+    # we append the document id at the beginning to avoid any of the document
+    # being the prefix of other documents
+    prompts = [
+        str(i) + ' '.join(['hi'] * args.document_length)
+        for i in range(args.num_documents)
+    ]
+
+    prompts = repeat_prompts(prompts, args.repeat_count, mode=args.repeat_mode)
+
+    warmup_prompts = [
+        "This is warm up request " + str(i) + \
+                ' '.join(['hi'] * args.document_length)
+        for i in range(args.num_documents)]
+
+    # Create the LLM engine
+    engine_args = EngineArgs.from_cli_args(args)
+    llm = LLM(**dataclasses.asdict(engine_args))
+    sampling_params = SamplingParams(temperature=0, max_tokens=args.output_len)
+
+    print("------warm up------")
+    test_long_document_qa(
+        llm=llm,
+        prompts=warmup_prompts,
+        sampling_params=sampling_params,
+    )
+
+    print("------start generating------")
+    test_long_document_qa(
+        llm=llm,
+        prompts=prompts,
+        sampling_params=sampling_params,
+    )
+
+
+if __name__ == "__main__":
+    parser = FlexibleArgumentParser(
+        description=
+        'Benchmark the performance with or without automatic prefix caching.')
+
+    parser.add_argument(
+        '--document-length',
+        type=int,
+        # Roughly the number of tokens for a system paper,
+        # excluding images
+        default=20000,
+        help='Range of input lengths for sampling prompts,'
+        'specified as "min:max" (e.g., "128:256").')
+
+    parser.add_argument('--num-documents',
+                        type=int,
+                        default=8,
+                        help='Range of input lengths for sampling prompts,'
+                        'specified as "min:max" (e.g., "128:256").')
+
+    parser.add_argument('--output-len', type=int, default=10)
+
+    parser.add_argument('--repeat-count',
+                        type=int,
+                        default=2,
+                        help='Number of times to repeat each prompt')
+
+    parser.add_argument("--repeat-mode",
+                        type=str,
+                        default='random',
+                        help='The mode to repeat prompts. The supported '
+                        'modes are "random", "tile", and "interleave". '
+                        'See repeat_prompts() in the source code for details.')
+
+    parser.add_argument("--shuffle-seed",
+                        type=int,
+                        default=0,
+                        help='Random seed when the repeat mode is "random"')
+
+    parser = EngineArgs.add_cli_args(parser)
+    args = parser.parse_args()
+    main(args)

benchmarks/benchmark_prefix_caching.py

@@ -1,3 +1,4 @@
+# SPDX-License-Identifier: Apache-2.0
 """
 Benchmark the efficiency of prefix caching.
 
@@ -10,7 +11,8 @@ Fixed example usage:
         --model meta-llama/Llama-2-7b-chat-hf \
         --enable-prefix-caching \
         --num-prompts 1 \
-        --repeat-count 100
+        --repeat-count 100 \
+        --input-length-range 128:256
 
 ShareGPT example usage:
     # This command samples 20 prompts with input lengths

benchmarks/benchmark_prioritization.py

@@ -1,3 +1,4 @@
+# SPDX-License-Identifier: Apache-2.0
 """Benchmark offline prioritization."""
 import argparse
 import dataclasses

benchmarks/benchmark_serving.py

@@ -1,3 +1,4 @@
+# SPDX-License-Identifier: Apache-2.0
 r"""Benchmark online serving throughput.
 
 On the server side, run one of the following commands:
@@ -25,6 +26,7 @@ On the client side, run:
 import argparse
 import asyncio
 import base64
+import gc
 import io
 import json
 import os
@@ -199,7 +201,7 @@ def sample_sonnet_requests(
     return sampled_requests
 
 
-def sample_mmmu_pro_vision_requests(
+def sample_vision_arena_requests(
     dataset,
     num_requests: int,
     tokenizer: PreTrainedTokenizerBase,
@@ -211,13 +213,7 @@ def sample_mmmu_pro_vision_requests(
         if len(sampled_requests) == num_requests:
             break
 
-        # MMMU-Pro vision direct prompt
+        prompt = data["turns"][0][0]['content']
-        # Ref: https://github.com/MMMU-Benchmark/MMMU/blob/6ce42f4d8f70c1841c67867152648974415b5cac/mmmu-pro/prompts.yaml#L5
-        prompt = (
-            "Answer with the option letter from the given choices directly. "
-            "The last line of your response should be of the following "
-            "format: 'Answer: $LETTER' (without quotes) where LETTER is one of "
-            "options.")
 
         prompt_token_ids = tokenizer(prompt).input_ids
         if fixed_output_len is None:
@@ -229,10 +225,10 @@ def sample_mmmu_pro_vision_requests(
         output_len = fixed_output_len
 
         assert isinstance(
-            data["image"],
+            data["images"][0],
             Image), ("Input image format must be `PIL.Image.Image`, "
                      f"given {type(data['image'])}.")
-        image: Image = data["image"]
+        image: Image = data["images"][0]
         image = image.convert("RGB")
         image_data = io.BytesIO()
         image.save(image_data, format='JPEG')
@@ -251,7 +247,7 @@ def sample_mmmu_pro_vision_requests(
 
 def sample_hf_requests(
     dataset_path: str,
-    dataset_subset: str,
+    dataset_subset: Optional[str],
     dataset_split: str,
     num_requests: int,
     tokenizer: PreTrainedTokenizerBase,
@@ -259,19 +255,17 @@ def sample_hf_requests(
     fixed_output_len: Optional[int] = None,
 ) -> List[Tuple[str, str, int, Optional[Dict[str, Collection[str]]]]]:
 
-    # Special case for MMMU-Pro vision dataset
+    # Special case for vision_arena dataset
-    if dataset_path == 'MMMU/MMMU_Pro' and dataset_subset == 'vision':
+    if dataset_path == 'lmarena-ai/vision-arena-bench-v0.1' \
-        assert dataset_split == "test"
+        and dataset_subset is None:
+        assert dataset_split == "train"
         dataset = load_dataset(dataset_path,
                                name=dataset_subset,
                                split=dataset_split,
                                streaming=True)
-        assert "image" in dataset.features, (
+        dataset = dataset.shuffle(seed=random_seed)
-            "MMMU/MMMU_Pro vision dataset must have 'image' column.")
+        return sample_vision_arena_requests(dataset, num_requests, tokenizer,
-        filter_func = lambda x: isinstance(x["image"], Image)
+                                            fixed_output_len)
-        dataset = dataset.shuffle(seed=random_seed).filter(filter_func)
-        return sample_mmmu_pro_vision_requests(dataset, num_requests,
-                                               tokenizer, fixed_output_len)
 
     dataset = load_dataset(dataset_path,
                            name=dataset_subset,
@@ -423,7 +417,7 @@ def calculate_metrics(
     tokenizer: PreTrainedTokenizerBase,
     selected_percentile_metrics: List[str],
     selected_percentiles: List[float],
-    gootput_config_dict: Dict[str, float],
+    goodput_config_dict: Dict[str, float],
 ) -> Tuple[BenchmarkMetrics, List[int]]:
     actual_output_lens: List[int] = []
     total_input = 0
@@ -436,9 +430,13 @@ def calculate_metrics(
     e2els: List[float] = []
     for i in range(len(outputs)):
         if outputs[i].success:
-            # We use the tokenizer to count the number of output tokens for all
+            output_len = outputs[i].output_tokens
-            # serving backends instead of looking at len(outputs[i].itl) since
+
-            # multiple output tokens may be bundled together
+            if output_len is None:
+                # We use the tokenizer to count the number of output tokens
+                # for some serving backends instead of looking at
+                # len(outputs[i].itl) since multiple output tokens may be
+                # bundled together
                 # Note : this may inflate the output token count slightly
                 output_len = len(
                     tokenizer(outputs[i].generated_text,
@@ -447,8 +445,8 @@ def calculate_metrics(
             total_input += input_requests[i][1]
             tpot = 0
             if output_len > 1:
-                tpot = (outputs[i].latency - outputs[i].ttft) / (output_len -
+                latency_minus_ttft = outputs[i].latency - outputs[i].ttft
-                                                                 1)
+                tpot = latency_minus_ttft / (output_len - 1)
                 tpots.append(tpot)
             # Note: if output_len <= 1, we regard tpot as 0 for goodput
             all_tpots.append(tpot)
@@ -459,21 +457,21 @@ def calculate_metrics(
         else:
             actual_output_lens.append(0)
 
-    if gootput_config_dict:
+    if goodput_config_dict:
         valid_metrics = []
         slo_values = []
 
-        if "ttft" in gootput_config_dict:
+        if "ttft" in goodput_config_dict:
             valid_metrics.append(ttfts)
-            slo_values.append(gootput_config_dict["ttft"] /
+            slo_values.append(goodput_config_dict["ttft"] /
                               MILLISECONDS_TO_SECONDS_CONVERSION)
-        if "tpot" in gootput_config_dict:
+        if "tpot" in goodput_config_dict:
             valid_metrics.append(all_tpots)
-            slo_values.append(gootput_config_dict["tpot"] /
+            slo_values.append(goodput_config_dict["tpot"] /
                               MILLISECONDS_TO_SECONDS_CONVERSION)
-        if "e2el" in gootput_config_dict:
+        if "e2el" in goodput_config_dict:
             valid_metrics.append(e2els)
-            slo_values.append(gootput_config_dict["e2el"] /
+            slo_values.append(goodput_config_dict["e2el"] /
                               MILLISECONDS_TO_SECONDS_CONVERSION)
 
         for req_metric in zip(*valid_metrics):
@@ -525,6 +523,7 @@ async def benchmark(
     api_url: str,
     base_url: str,
     model_id: str,
+    model_name: str,
     tokenizer: PreTrainedTokenizerBase,
     input_requests: List[Tuple[str, int, int]],
     logprobs: Optional[int],
@@ -536,7 +535,7 @@ async def benchmark(
     selected_percentile_metrics: List[str],
     selected_percentiles: List[str],
     ignore_eos: bool,
-    gootput_config_dict: Dict[str, float],
+    goodput_config_dict: Dict[str, float],
     max_concurrency: Optional[int],
 ):
     if backend in ASYNC_REQUEST_FUNCS:
@@ -553,6 +552,7 @@ async def benchmark(
             "Multi-modal content is only supported on 'openai-chat' backend.")
     test_input = RequestFuncInput(
         model=model_id,
+        model_name=model_name,
         prompt=test_prompt,
         api_url=api_url,
         prompt_len=test_prompt_len,
@@ -573,6 +573,7 @@ async def benchmark(
     if profile:
         print("Starting profiler...")
         profile_input = RequestFuncInput(model=model_id,
+                                         model_name=model_name,
                                          prompt=test_prompt,
                                          api_url=base_url + "/start_profile",
                                          prompt_len=test_prompt_len,
@@ -616,6 +617,7 @@ async def benchmark(
     async for request in get_request(input_requests, request_rate, burstiness):
         prompt, prompt_len, output_len, mm_content = request
         request_func_input = RequestFuncInput(model=model_id,
+                                              model_name=model_name,
                                               prompt=prompt,
                                               api_url=api_url,
                                               prompt_len=prompt_len,
@@ -657,7 +659,7 @@ async def benchmark(
         tokenizer=tokenizer,
         selected_percentile_metrics=selected_percentile_metrics,
         selected_percentiles=selected_percentiles,
-        gootput_config_dict=gootput_config_dict,
+        goodput_config_dict=goodput_config_dict,
     )
 
     print("{s:{c}^{n}}".format(s=' Serving Benchmark Result ', n=50, c='='))
@@ -669,7 +671,7 @@ async def benchmark(
                                  metrics.total_output))
     print("{:<40} {:<10.2f}".format("Request throughput (req/s):",
                                     metrics.request_throughput))
-    if gootput_config_dict:
+    if goodput_config_dict:
         print("{:<40} {:<10.2f}".format("Request goodput (req/s):",
                                         metrics.request_goodput))
     print("{:<40} {:<10.2f}".format("Output token throughput (tok/s):",
@@ -684,7 +686,7 @@ async def benchmark(
         "total_output_tokens": metrics.total_output,
         "request_throughput": metrics.request_throughput,
         "request_goodput:":
-        metrics.request_goodput if gootput_config_dict else None,
+        metrics.request_goodput if goodput_config_dict else None,
         "output_throughput": metrics.output_throughput,
         "total_token_throughput": metrics.total_token_throughput,
         "input_lens": [output.prompt_len for output in outputs],
@@ -740,11 +742,11 @@ async def benchmark(
 
 def check_goodput_args(args):
     # Check and parse goodput arguments
-    gootput_config_dict = {}
+    goodput_config_dict = {}
     VALID_NAMES = ["ttft", "tpot", "e2el"]
     if args.goodput:
-        gootput_config_dict = parse_goodput(args.goodput)
+        goodput_config_dict = parse_goodput(args.goodput)
-        for slo_name, slo_val in gootput_config_dict.items():
+        for slo_name, slo_val in goodput_config_dict.items():
             if slo_name not in VALID_NAMES:
                 raise ValueError(
                     f"Invalid metric name found, {slo_name}: {slo_val}. "
@@ -755,22 +757,22 @@ def check_goodput_args(args):
                     f"Invalid value found, {slo_name}: {slo_val}. "
                     "The service level objective value should be "
                     "non-negative.")
-    return gootput_config_dict
+    return goodput_config_dict
 
 
 def parse_goodput(slo_pairs):
-    gootput_config_dict = {}
+    goodput_config_dict = {}
     try:
         for slo_pair in slo_pairs:
             slo_name, slo_val = slo_pair.split(":")
-            gootput_config_dict[slo_name] = float(slo_val)
+            goodput_config_dict[slo_name] = float(slo_val)
     except ValueError as err:
         raise argparse.ArgumentTypeError(
             "Invalid format found for service level objectives. "
             "Specify service level objectives for goodput as \"KEY:VALUE\" "
             "pairs, where the key is a metric name, and the value is a "
             "number in milliseconds.") from err
-    return gootput_config_dict
+    return goodput_config_dict
 
 
 def main(args: argparse.Namespace):
@@ -780,6 +782,7 @@ def main(args: argparse.Namespace):
 
     backend = args.backend
     model_id = args.model
+    model_name = args.served_model_name
     tokenizer_id = args.tokenizer if args.tokenizer is not None else args.model
     tokenizer_mode = args.tokenizer_mode
 
@@ -869,7 +872,11 @@ def main(args: argparse.Namespace):
     else:
         raise ValueError(f"Unknown dataset: {args.dataset_name}")
 
-    gootput_config_dict = check_goodput_args(args)
+    goodput_config_dict = check_goodput_args(args)
+
+    # Avoid GC processing "static" data - reduce pause times.
+    gc.collect()
+    gc.freeze()
 
     benchmark_result = asyncio.run(
         benchmark(
@@ -877,6 +884,7 @@ def main(args: argparse.Namespace):
             api_url=api_url,
             base_url=base_url,
             model_id=model_id,
+            model_name=model_name,
             tokenizer=tokenizer,
             input_requests=input_requests,
             logprobs=args.logprobs,
@@ -890,7 +898,7 @@ def main(args: argparse.Namespace):
                 float(p) for p in args.metric_percentiles.split(",")
             ],
             ignore_eos=args.ignore_eos,
-            gootput_config_dict=gootput_config_dict,
+            goodput_config_dict=goodput_config_dict,
             max_concurrency=args.max_concurrency,
         ))
 
@@ -919,8 +927,8 @@ def main(args: argparse.Namespace):
                     )
 
         # Traffic
-        result_json["request_rate"] = (
+        result_json["request_rate"] = (args.request_rate if args.request_rate
-            args.request_rate if args.request_rate < float("inf") else "inf")
+                                       < float("inf") else "inf")
         result_json["burstiness"] = args.burstiness
         result_json["max_concurrency"] = args.max_concurrency
 
@@ -1222,5 +1230,12 @@ if __name__ == "__main__":
         'always use the slow tokenizer. \n* '
         '"mistral" will always use the `mistral_common` tokenizer.')
 
+    parser.add_argument("--served-model-name",
+                        type=str,
+                        default=None,
+                        help="The model name used in the API. "
+                        "If not specified, the model name will be the "
+                        "same as the ``--model`` argument. ")
+
     args = parser.parse_args()
     main(args)

benchmarks/benchmark_serving_guided.py

@@ -1,3 +1,4 @@
+# SPDX-License-Identifier: Apache-2.0
 r"""Benchmark online serving throughput with guided decoding.
 
 On the server side, run one of the following commands:

benchmarks/benchmark_throughput.py

@@ -1,3 +1,4 @@
+# SPDX-License-Identifier: Apache-2.0
 """Benchmark offline inference throughput."""
 import argparse
 import dataclasses

benchmarks/cutlass_benchmarks/sparse_benchmarks.py

@@ -1,3 +1,5 @@
+# SPDX-License-Identifier: Apache-2.0
+
 import argparse
 import copy
 import itertools

benchmarks/cutlass_benchmarks/utils.py

@@ -1,3 +1,5 @@
+# SPDX-License-Identifier: Apache-2.0
+
 # Cutlass bench utils
 from typing import Iterable, Tuple
 

benchmarks/cutlass_benchmarks/w8a8_benchmarks.py

@@ -1,9 +1,11 @@
+# SPDX-License-Identifier: Apache-2.0
+
 import argparse
 import copy
 import itertools
 import pickle as pkl
 import time
-from typing import Callable, Iterable, List, Tuple
+from typing import Callable, Iterable, List, Optional, Tuple
 
 import torch
 import torch.utils.benchmark as TBenchmark
@@ -12,6 +14,8 @@ from utils import make_rand_tensors
 from weight_shapes import WEIGHT_SHAPES
 
 from vllm import _custom_ops as ops
+from vllm.model_executor.layers.quantization.utils.fp8_utils import (
+    w8a8_block_fp8_matmul)
 from vllm.utils import FlexibleArgumentParser
 
 DEFAULT_MODELS = list(WEIGHT_SHAPES.keys())
@@ -38,8 +42,15 @@ def bench_fn(label: str, sub_label: str, description: str, fn: Callable, *args,
     ).blocked_autorange(min_run_time=min_run_time)
 
 
-def bench_int8(dtype: torch.dtype, m: int, k: int, n: int, label: str,
+def bench_int8(
-               sub_label: str) -> Iterable[TMeasurement]:
+        dtype: torch.dtype,
+        m: int,
+        k: int,
+        n: int,
+        label: str,
+        sub_label: str,
+        bench_kernels: Optional[List[str]] = None) -> Iterable[TMeasurement]:
+    """Benchmark INT8-based kernels."""
     assert dtype == torch.int8
     a, b = make_rand_tensors(torch.int8, m, n, k)
     scale_a = torch.tensor(1.0, device="cuda", dtype=torch.float32)
@@ -48,155 +59,132 @@ def bench_int8(dtype: torch.dtype, m: int, k: int, n: int, label: str,
     azp = torch.zeros((m, ), device="cuda", dtype=torch.int32)
     azp_adj = torch.zeros((n, ), device="cuda", dtype=torch.int32)
 
+    bench_fns = {
+        "pytorch_bf16_bf16_bf16_matmul-no-scales":
+        lambda: torch.mm(a.to(dtype=torch.bfloat16), b.to(dtype=torch.bfloat16)
+                         ),
+        "pytorch_fp16_fp16_fp16_matmul-no-scales":
+        lambda: torch.mm(a.to(dtype=torch.float16), b.to(dtype=torch.float16)),
+        "cutlass_i8_i8_bf16_scaled_mm":
+        lambda: ops.cutlass_scaled_mm(a, b, scale_a, scale_b, torch.bfloat16),
+        "cutlass_i8_i8_bf16_scaled_mm_bias":
+        lambda: ops.cutlass_scaled_mm(a, b, scale_a, scale_b, torch.bfloat16,
+                                      bias),
+        "cutlass_i8_i8_bf16_scaled_mm_azp":
+        lambda: ops.cutlass_scaled_mm_azp(a, b, scale_a, scale_b, torch.
+                                          bfloat16, azp_adj),
+        "cutlass_i8_i8_bf16_scaled_mm_azp_bias":
+        lambda: ops.cutlass_scaled_mm_azp(a, b, scale_a, scale_b, torch.
+                                          bfloat16, azp_adj, None, bias),
+        "cutlass_i8_i8_bf16_scaled_mm_azp_pt":
+        lambda: ops.cutlass_scaled_mm_azp(a, b, scale_a, scale_b, torch.
+                                          bfloat16, azp_adj, azp),
+        "cutlass_i8_i8_bf16_scaled_mm_azp_pt_bias":
+        lambda: ops.cutlass_scaled_mm_azp(a, b, scale_a, scale_b, torch.
+                                          bfloat16, azp_adj, azp, bias),
+    }
+
     timers = []
-    # pytorch impl - bfloat16
+    for name, fn in bench_fns.items():
-    timers.append(
+        # If bench_kernels is None, run all. Otherwise, run only exact matches.
-        bench_fn(label, sub_label, "pytorch_bf16_bf16_bf16_matmul-no-scales",
+        if bench_kernels is None or name in bench_kernels:
-                 torch.mm, a.to(dtype=torch.bfloat16),
+            print(f"Running {name}")
-                 b.to(dtype=torch.bfloat16)))
+            timers.append(bench_fn(label, sub_label, name, fn))
-
-    # pytorch impl - float16
-    timers.append(
-        bench_fn(label, sub_label,
-                 "pytorch_fp16_fp16_fp16_matmul-no-scales", torch.mm,
-                 a.to(dtype=torch.float16), b.to(dtype=torch.float16)))
-
-    # cutlass impl
-    timers.append(
-        bench_fn(label, sub_label, "cutlass_i8_i8_bf16_scaled_mm",
-                 ops.cutlass_scaled_mm, a, b, scale_a, scale_b,
-                 torch.bfloat16))
-
-    # cutlass with bias
-    timers.append(
-        bench_fn(label, sub_label, "cutlass_i8_i8_bf16_scaled_mm_bias",
-                 ops.cutlass_scaled_mm, a, b, scale_a, scale_b, torch.bfloat16,
-                 bias))
-
-    # cutlass with azp per-tensor
-    timers.append(
-        bench_fn(label, sub_label, "cutlass_i8_i8_bf16_scaled_mm_azp",
-                 ops.cutlass_scaled_mm_azp, a, b, scale_a, scale_b,
-                 torch.bfloat16, azp_adj))
-
-    # cutlass with azp per-tensor + bias
-    timers.append(
-        bench_fn(label, sub_label, "cutlass_i8_i8_bf16_scaled_mm_azp_bias",
-                 ops.cutlass_scaled_mm_azp, a, b, scale_a, scale_b,
-                 torch.bfloat16, azp_adj, None, bias))
-
-    # cutlass with azp per-token
-    timers.append(
-        bench_fn(label, sub_label, "cutlass_i8_i8_bf16_scaled_mm_azp_pt",
-                 ops.cutlass_scaled_mm_azp, a, b, scale_a, scale_b,
-                 torch.bfloat16, azp_adj, azp))
-
-    # cutlass with azp per-token + bias
-    timers.append(
-        bench_fn(label, sub_label, "cutlass_i8_i8_bf16_scaled_mm_azp_pt_bias",
-                 ops.cutlass_scaled_mm_azp, a, b, scale_a, scale_b,
-                 torch.bfloat16, azp_adj, azp, bias))
 
     return timers
 
 
-def bench_fp8(dtype: torch.dtype, m: int, k: int, n: int, label: str,
+def bench_fp8(
-              sub_label: str) -> Iterable[TMeasurement]:
+        dtype: torch.dtype,
+        m: int,
+        k: int,
+        n: int,
+        label: str,
+        sub_label: str,
+        bench_kernels: Optional[List[str]] = None) -> Iterable[TMeasurement]:
+    """Benchmark FP8-based kernels."""
     assert dtype == torch.float8_e4m3fn
     a, b = make_rand_tensors(torch.float8_e4m3fn, m, n, k)
+    a_cont = a.contiguous()
     scale_a = torch.tensor(1.0, device="cuda", dtype=torch.float32)
     scale_b = torch.tensor(1.0, device="cuda", dtype=torch.float32)
+    block_scale_a = torch.rand((m, k // 128),
+                               device="cuda",
+                               dtype=torch.float32)
+    block_scale_b = torch.rand((k // 128, n // 128),
+                               device="cuda",
+                               dtype=torch.float32)
+    block_scale_a_M_major = block_scale_a.t().contiguous().t()
+    block_scale_b_K_major = block_scale_b.t().contiguous().t()
     bias = torch.zeros((n, ), device="cuda", dtype=torch.bfloat16)
 
-    timers = []
+    print(m, k, n)
 
-    # pytorch impl w. bf16
+    bench_fns = {
-    timers.append(
+        "pytorch_bf16_bf16_bf16_matmul-no-scales":
-        bench_fn(label, sub_label, "pytorch_bf16_bf16_bf16_matmul-no-scales",
+        lambda: torch.mm(a.to(dtype=torch.bfloat16), b.to(dtype=torch.bfloat16)
-                 torch.mm, a.to(dtype=torch.bfloat16, device="cuda"),
+                         ),
-                 b.to(dtype=torch.bfloat16, device="cuda")))
+        "pytorch_fp16_fp16_fp16_matmul-no-scales":
-
+        lambda: torch.mm(a.to(dtype=torch.float16), b.to(dtype=torch.float16)),
-    # pytorch impl: bf16 output, without fp8 fast accum
+        "pytorch_fp8_fp8_fp16_scaled_mm":
-    timers.append(
+        lambda: torch._scaled_mm(
-        bench_fn(label,
+            a, b, scale_a, scale_b, out_dtype=torch.float16),
-                 sub_label,
+        "pytorch_fp8_fp8_fp16_scaled_mm_fast_accum":
-                 "pytorch_fp8_fp8_bf16_scaled_mm",
+        lambda: torch._scaled_mm(a,
-                 torch._scaled_mm,
-                 a,
                                  b,
-                 scale_a=scale_a,
+                                 scale_a,
-                 scale_b=scale_b,
+                                 scale_b,
-                 out_dtype=torch.bfloat16))
-
-    # pytorch impl: bf16 output, with fp8 fast accum
-    timers.append(
-        bench_fn(label,
-                 sub_label,
-                 "pytorch_fp8_fp8_bf16_scaled_mm_fast_accum",
-                 torch._scaled_mm,
-                 a,
-                 b,
-                 scale_a=scale_a,
-                 scale_b=scale_b,
-                 out_dtype=torch.bfloat16,
-                 use_fast_accum=True))
-
-    # pytorch impl: fp16 output, without fp8 fast accum
-    timers.append(
-        bench_fn(label,
-                 sub_label,
-                 "pytorch_fp8_fp8_fp16_scaled_mm",
-                 torch._scaled_mm,
-                 a,
-                 b,
-                 scale_a=scale_a,
-                 scale_b=scale_b,
-                 out_dtype=torch.float16))
-
-    # pytorch impl: fp16 output, with fp8 fast accum
-    timers.append(
-        bench_fn(label,
-                 sub_label,
-                 "pytorch_fp8_fp8_fp16_scaled_mm_fast_accum",
-                 torch._scaled_mm,
-                 a,
-                 b,
-                 scale_a=scale_a,
-                 scale_b=scale_b,
                                  out_dtype=torch.float16,
-                 use_fast_accum=True))
+                                 use_fast_accum=True),
+        "pytorch_fp8_fp8_bf16_scaled_mm":
+        lambda: torch._scaled_mm(
+            a, b, scale_a, scale_b, out_dtype=torch.bfloat16),
+        "pytorch_fp8_fp8_bf16_scaled_mm_fast_accum":
+        lambda: torch._scaled_mm(a,
+                                 b,
+                                 scale_a,
+                                 scale_b,
+                                 out_dtype=torch.bfloat16,
+                                 use_fast_accum=True),
+        "cutlass_fp8_fp8_bf16_scaled_mm":
+        lambda: ops.cutlass_scaled_mm(a, b, scale_a, scale_b, torch.bfloat16),
+        "cutlass_fp8_fp8_fp16_scaled_mm":
+        lambda: ops.cutlass_scaled_mm(a, b, scale_a, scale_b, torch.float16),
+        "cutlass_fp8_fp8_bf16_scaled_mm_bias":
+        lambda: ops.cutlass_scaled_mm(a, b, scale_a, scale_b, torch.bfloat16,
+                                      bias),
+        "cutlass_fp8_fp8_fp16_scaled_mm_bias":
+        lambda: ops.cutlass_scaled_mm(a, b, scale_a, scale_b, torch.float16,
+                                      bias.to(dtype=torch.float16)),
+        "triton_fp8_fp8_fp16_scaled_mm_blockwise":
+        lambda: w8a8_block_fp8_matmul(a_cont, b.t(), block_scale_a,
+                                      block_scale_b.t(), (128, 128)),
+        "cutlass_fp8_fp8_fp16_scaled_mm_blockwise":
+        lambda: ops.cutlass_scaled_mm(a, b, block_scale_a_M_major,
+                                      block_scale_b_K_major, torch.float16),
+    }
 
-    # cutlass impl: bf16 output
+    timers = []
-    timers.append(
+    for name, fn in bench_fns.items():
-        bench_fn(label, sub_label, "cutlass_fp8_fp8_bf16_scaled_mm",
+        # If bench_kernels is None, run all. Otherwise, run only exact matches.
-                 ops.cutlass_scaled_mm, a, b, scale_a, scale_b,
+        if bench_kernels is None or name in bench_kernels:
-                 torch.bfloat16))
+            print(f"Running {name}")
-    # cutlass impl: fp16 output
+            timers.append(bench_fn(label, sub_label, name, fn))
-    timers.append(
-        bench_fn(label, sub_label, "cutlass_fp8_fp8_fp16_scaled_mm",
-                 ops.cutlass_scaled_mm, a, b, scale_a, scale_b, torch.float16))
-
-    # cutlass impl: bf16 output, with bias
-    timers.append(
-        bench_fn(label, sub_label, "cutlass_fp8_fp8_bf16_scaled_mm_bias",
-                 ops.cutlass_scaled_mm, a, b, scale_a, scale_b, torch.bfloat16,
-                 bias))
-
-    # cutlass impl: fp16 output, with bias
-    timers.append(
-        bench_fn(label, sub_label, "cutlass_fp8_fp8_fp16_scaled_mm_bias",
-                 ops.cutlass_scaled_mm, a, b, scale_a, scale_b, torch.float16,
-                 bias.to(dtype=torch.float16)))
 
     return timers
 
 
-def bench(dtype: torch.dtype, m: int, k: int, n: int, label: str,
+def bench(dtype: torch.dtype,
-          sub_label: str) -> Iterable[TMeasurement]:
+          m: int,
+          k: int,
+          n: int,
+          label: str,
+          sub_label: str,
+          bench_kernels: Optional[List[str]] = None) -> Iterable[TMeasurement]:
     if dtype == torch.int8:
-        return bench_int8(dtype, m, k, n, label, sub_label)
+        return bench_int8(dtype, m, k, n, label, sub_label, bench_kernels)
     if dtype == torch.float8_e4m3fn:
-        return bench_fp8(dtype, m, k, n, label, sub_label)
+        return bench_fp8(dtype, m, k, n, label, sub_label, bench_kernels)
     raise ValueError("unsupported type")
 
 
@@ -207,18 +195,22 @@ def print_timers(timers: Iterable[TMeasurement]):
 
 
 def run(dtype: torch.dtype,
-        MKNs: Iterable[Tuple[int, int, int]]) -> Iterable[TMeasurement]:
+        MKNs: Iterable[Tuple[int, int, int]],
+        bench_kernels: Optional[List[str]] = None) -> Iterable[TMeasurement]:
     results = []
     for m, k, n in MKNs:
-        timers = bench(dtype, m, k, n, f"scaled-{dtype}-gemm",
+        timers = bench(dtype,
-                       f"MKN=({m}x{k}x{n})")
+                       m,
+                       k,
+                       n,
+                       f"scaled-{dtype}-gemm",
+                       f"MKN=({m}x{k}x{n})",
+                       bench_kernels=bench_kernels)
         print_timers(timers)
         results.extend(timers)
-
     return results
 
 
-# output makers
 def make_output(data: Iterable[TMeasurement],
                 MKNs: Iterable[Tuple[int, int, int]],
                 base_description: str,
@@ -232,15 +224,11 @@ def make_output(data: Iterable[TMeasurement],
         pkl.dump(data, f)
 
 
-# argparse runners
-
-
 def run_square_bench(args):
     dim_sizes = list(
         range(args.dim_start, args.dim_end + 1, args.dim_increment))
     MKNs = list(zip(dim_sizes, dim_sizes, dim_sizes))
-    data = run(args.dtype, MKNs)
+    data = run(args.dtype, MKNs, bench_kernels=args.kernels)
-
     make_output(data, MKNs, f"square_bench-{args.dtype}")
 
 
@@ -251,8 +239,7 @@ def run_range_bench(args):
     Ks = [args.k_constant] * n if args.k_constant is not None else dim_sizes
     Ns = [args.n_constant] * n if args.n_constant is not None else dim_sizes
     MKNs = list(zip(Ms, Ks, Ns))
-    data = run(args.dtype, MKNs)
+    data = run(args.dtype, MKNs, bench_kernels=args.kernels)
-
     make_output(data, MKNs, f"range_bench-{args.dtype}")
 
 
@@ -278,7 +265,7 @@ def run_model_bench(args):
             for k, n in KNs:
                 MKNs.append((m, k, n))
 
-        data = run(args.dtype, MKNs)
+        data = run(args.dtype, MKNs, bench_kernels=args.kernels)
         model_bench_data.append(data)
 
     # Print all results
@@ -328,6 +315,15 @@ Benchmark Cutlass GEMM.
                         type=to_torch_dtype,
                         required=True,
                         help="Available options are ['int8', 'fp8']")
+    parser.add_argument(
+        "--kernels",
+        nargs="+",
+        type=str,
+        default=None,
+        help=
+        "Exact names of the kernels to benchmark. If not set, runs all kernels."
+    )
+
     subparsers = parser.add_subparsers(dest="cmd")
 
     square_parser = subparsers.add_parser("square_bench")

benchmarks/cutlass_benchmarks/weight_shapes.py

@@ -1,3 +1,5 @@
+# SPDX-License-Identifier: Apache-2.0
+
 # Weight Shapes are in the format
 # ([K, N], TP_SPLIT_DIM)
 # Example:

benchmarks/disagg_benchmarks/disagg_prefill_proxy_server.py

@@ -1,3 +1,5 @@
+# SPDX-License-Identifier: Apache-2.0
+
 import os
 
 import aiohttp

benchmarks/disagg_benchmarks/round_robin_proxy.py

@@ -1,3 +1,5 @@
+# SPDX-License-Identifier: Apache-2.0
+
 import asyncio
 import itertools
 

benchmarks/disagg_benchmarks/visualize_benchmark_results.py

@@ -1,3 +1,5 @@
+# SPDX-License-Identifier: Apache-2.0
+
 import json
 
 import matplotlib.pyplot as plt

benchmarks/fused_kernels/layernorm_rms_benchmarks.py

@@ -1,3 +1,5 @@
+# SPDX-License-Identifier: Apache-2.0
+
 import pickle as pkl
 import time
 from dataclasses import dataclass

benchmarks/kernels/benchmark_aqlm.py

@@ -1,3 +1,5 @@
+# SPDX-License-Identifier: Apache-2.0
+
 import os
 import sys
 from typing import Optional

benchmarks/kernels/benchmark_layernorm.py

@@ -1,3 +1,5 @@
+# SPDX-License-Identifier: Apache-2.0
+
 import time
 
 import torch

benchmarks/kernels/benchmark_machete.py

@@ -1,3 +1,5 @@
+# SPDX-License-Identifier: Apache-2.0
+
 import argparse
 import copy
 import itertools

benchmarks/kernels/benchmark_marlin.py

@@ -1,3 +1,5 @@
+# SPDX-License-Identifier: Apache-2.0
+
 from typing import List
 
 import torch

benchmarks/kernels/benchmark_moe.py

@@ -1,6 +1,9 @@
+# SPDX-License-Identifier: Apache-2.0
+
 import argparse
 import time
 from datetime import datetime
+from itertools import product
 from typing import Any, Dict, List, Tuple, TypedDict
 
 import ray
@@ -13,6 +16,9 @@ from vllm.model_executor.layers.fused_moe.fused_moe import *
 from vllm.platforms import current_platform
 from vllm.utils import FlexibleArgumentParser
 
+FP8_DTYPE = torch.float8_e4m3fnuz if current_platform.is_rocm(
+) else torch.float8_e4m3fn
+
 
 class BenchmarkConfig(TypedDict):
     BLOCK_SIZE_M: int
@@ -80,8 +86,8 @@ def benchmark_config(
         a1_scale = torch.randn(1, dtype=torch.float32)
         a2_scale = torch.randn(1, dtype=torch.float32)
 
-        w1 = w1.to(torch.float8_e4m3fn)
+        w1 = w1.to(FP8_DTYPE)
-        w2 = w2.to(torch.float8_e4m3fn)
+        w2 = w2.to(FP8_DTYPE)
 
     input_gating = torch.empty(num_tokens, num_experts, dtype=torch.float32)
 
@@ -141,28 +147,172 @@ def benchmark_config(
     return avg
 
 
-def get_configs_compute_bound() -> List[Dict[str, int]]:
+def get_rocm_tuning_space(use_fp16):
+    block_mn_range = [16, 32, 64, 128, 256]
+    block_k_range = [16, 32, 64, 128, 256]
+    if not use_fp16:
+        block_k_range.remove(16)  # BLOCK_K=16 not supported for fp8
+    num_warps_range = [1, 2, 4, 8]
+    group_m_range = [1, 4, 8, 16, 32]
+    num_stage_range = [2]
+    waves_per_eu_range = [0]
+    matrix_instr_nonkdim_range = [16, 32] if use_fp16 else []
+    kpack_range = [1, 2] if use_fp16 else []
+
+    param_ranges = {
+        "BLOCK_SIZE_M": block_mn_range,
+        "BLOCK_SIZE_N": block_mn_range,
+        "BLOCK_SIZE_K": block_k_range,
+        "GROUP_SIZE_M": group_m_range,
+        "num_warps": num_warps_range,
+        "num_stages": num_stage_range,
+        "waves_per_eu": waves_per_eu_range,
+    }
+    if use_fp16:
+        param_ranges["matrix_instr_nonkdim"] = matrix_instr_nonkdim_range
+        param_ranges["kpack"] = kpack_range
+
+    return param_ranges
+
+
+def get_configs_compute_bound(use_fp16) -> List[Dict[str, int]]:
+    configs: List[BenchmarkConfig] = []
+
+    if current_platform.is_rocm():
+        param_ranges = get_rocm_tuning_space(use_fp16)
+    else:
         # Reduced search space for faster tuning.
         # TODO(woosuk): Increase the search space and use a performance model to
         # prune the search space.
-    configs: List[BenchmarkConfig] = []
+        block_m_range = [16, 32, 64, 128, 256]
-    for num_stages in [2, 3, 4, 5]:
+        block_n_range = [32, 64, 128, 256]
-        for block_m in [16, 32, 64, 128, 256]:
+        block_k_range = [64, 128, 256]
-            for block_k in [64, 128, 256]:
+        num_warps_range = [4, 8]
-                for block_n in [32, 64, 128, 256]:
+        group_m_range = [1, 16, 32, 64]
-                    for num_warps in [4, 8]:
+        num_stage_range = [2, 3, 4, 5]
-                        for group_size in [1, 16, 32, 64]:
+
-                            configs.append({
+        param_ranges = {
-                                "BLOCK_SIZE_M": block_m,
+            "BLOCK_SIZE_M": block_m_range,
-                                "BLOCK_SIZE_N": block_n,
+            "BLOCK_SIZE_N": block_n_range,
-                                "BLOCK_SIZE_K": block_k,
+            "BLOCK_SIZE_K": block_k_range,
-                                "GROUP_SIZE_M": group_size,
+            "GROUP_SIZE_M": group_m_range,
-                                "num_warps": num_warps,
+            "num_warps": num_warps_range,
-                                "num_stages": num_stages,
+            "num_stages": num_stage_range,
-                            })
+        }
+
+    keys, values = zip(*param_ranges.items())
+    for config_values in product(*values):
+        config = dict(zip(keys, config_values))
+        configs.append(config)
     return configs
 
 
+def prune_rocm_search_space(num_tokens, shard_intermediate_size, hidden_size,
+                            search_space, is_fp16):
+    N1, K1 = shard_intermediate_size, hidden_size
+    N2, K2 = hidden_size, shard_intermediate_size // 2
+    pruned_space_1 = prune_rocm_configs(num_tokens * 2, N1, K1, search_space,
+                                        is_fp16)
+    pruned_space_2 = prune_rocm_configs(num_tokens * 2, N2, K2, search_space,
+                                        is_fp16)
+    search_space = merge_unique_dicts(pruned_space_1, pruned_space_2)
+    return search_space
+
+
+# The following code is inspired by ROCm/Triton GEMM tuning script:
+# https://github.com/ROCm/triton/blob/triton-mlir/scripts/amd/gemm/tune_gemm.py#L89
+def prune_rocm_configs(M, N, K, configs, is_fp16=True):
+    pruned_configs = []
+    elemBytes_a = 2 if is_fp16 else 1
+    elemBytes_b = 2 if is_fp16 else 1
+
+    mfma = 16 if M < 32 or N < 32 else 32
+
+    # TODO (zhanglx): figure out the boundary between large and small gemms
+    large_gemm = False
+    if M >= 2048 and N >= 2048:
+        large_gemm = True
+
+    for config in configs:
+        BLOCK_SIZE_M = config.get("BLOCK_SIZE_M")
+        BLOCK_SIZE_N = config.get("BLOCK_SIZE_N")
+        BLOCK_SIZE_K = config.get("BLOCK_SIZE_K")
+        num_warps = config.get("num_warps")
+
+        if is_fp16:
+            matrix_instr_nonkdim = config.get("matrix_instr_nonkdim")
+            if matrix_instr_nonkdim > mfma:
+                continue
+        if mfma == 4 and BLOCK_SIZE_K < 64:
+            continue
+        # some layouts could not work properly in case
+        # number elements per thread is less 1
+        if BLOCK_SIZE_M * BLOCK_SIZE_N < 64:
+            continue
+        SPLIT_K = config.get("SPLIT_K", 1)
+        GROUP_M = config.get("GROUP_SIZE_M")
+        if is_fp16:
+            if (matrix_instr_nonkdim > BLOCK_SIZE_M
+                    or matrix_instr_nonkdim > BLOCK_SIZE_N):
+                continue
+            if (matrix_instr_nonkdim >= M
+                    and matrix_instr_nonkdim != BLOCK_SIZE_M):
+                continue
+            if (matrix_instr_nonkdim >= N
+                    and matrix_instr_nonkdim != BLOCK_SIZE_N):
+                continue
+        # Skip BLOCK_SIZE that is too large compare to M/N
+        # unless BLOCK_SIZE is already small enough
+        if M * 2 < BLOCK_SIZE_M and BLOCK_SIZE_M != 16:
+            continue
+        if N * 2 < BLOCK_SIZE_N and BLOCK_SIZE_N != 16:
+            continue
+        # skip large split_k when not necessary
+        if SPLIT_K != 1 and not need_split_k(M, N, K):
+            continue
+        # skip split_k that leads to EVEN_K = false
+        leap = SPLIT_K * BLOCK_SIZE_K
+        modv = K % leap
+        if modv != 0:
+            continue
+        # skip large GROUP_M
+        if GROUP_M * BLOCK_SIZE_M > M and GROUP_M != 1:
+            continue
+        # out of shared memory resource
+        # TODO (zhanglx): This does not consider the LDS usage in the epilogue
+        LDS = (BLOCK_SIZE_K * BLOCK_SIZE_M * elemBytes_a +
+               BLOCK_SIZE_K * BLOCK_SIZE_N * elemBytes_b)
+        if LDS > 65536:
+            continue
+        # Skip small block sizes and num_warps for large gemm
+        # For fp16 and f8, we want to only use BLOCK_SIZE >= 64
+        if large_gemm:
+            if BLOCK_SIZE_M < 64 or BLOCK_SIZE_N < 64:
+                continue
+            if BLOCK_SIZE_K < 64:
+                continue
+            if num_warps < 4:
+                continue
+
+        pruned_configs.append(config)
+
+    return pruned_configs
+
+
+def need_split_k(SIZE_M, SIZE_N, SIZE_K):
+    return (SIZE_M < 64 or SIZE_N < 64) and SIZE_K > 1024
+
+
+def merge_unique_dicts(list1, list2):
+    result = []
+    combined_list = list1.copy()
+    combined_list.extend(list2)
+    for dictionary in combined_list:
+        if dictionary not in result:
+            result.append(dictionary)
+    return result
+
+
 @ray.remote(num_gpus=1)
 class BenchmarkWorker:
 
@@ -170,6 +320,10 @@ class BenchmarkWorker:
         torch.set_default_device("cuda")
         current_platform.seed_everything(seed)
         self.seed = seed
+        # Get the device ID to allocate tensors and kernels
+        # on the respective GPU. This is required for Ray to work
+        # correctly with multi-GPU tuning on the ROCm platform.
+        self.device_id = int(ray.get_gpu_ids()[0])
 
     def benchmark(
         self,
@@ -191,9 +345,13 @@ class BenchmarkWorker:
         op_config = get_moe_configs(num_experts, shard_intermediate_size // 2,
                                     dtype_str)
         if op_config is None:
-            config = get_default_config(num_tokens, num_experts,
+            config = get_default_config(num_tokens,
-                                        shard_intermediate_size, hidden_size,
+                                        num_experts,
-                                        topk, dtype_str)
+                                        shard_intermediate_size,
+                                        hidden_size,
+                                        topk,
+                                        dtype_str,
+                                        is_marlin=False)
         else:
             config = op_config[min(op_config.keys(),
                                    key=lambda x: abs(x - num_tokens))]
@@ -217,6 +375,14 @@ class BenchmarkWorker:
     ) -> Dict[str, int]:
         best_config = None
         best_time = float("inf")
+        if current_platform.is_rocm():
+            is_fp16 = not (use_fp8_w8a8 or use_int8_w8a16)
+            search_space = prune_rocm_search_space(num_tokens,
+                                                   shard_intermediate_size,
+                                                   hidden_size, search_space,
+                                                   is_fp16)
+
+        with torch.cuda.device(self.device_id):
             for config in tqdm(search_space):
                 try:
                     kernel_time = benchmark_config(config,
@@ -228,7 +394,7 @@ class BenchmarkWorker:
                                                    dtype,
                                                    use_fp8_w8a8,
                                                    use_int8_w8a16,
-                                               num_iters=10)
+                                                   num_iters=20)
                 except triton.runtime.autotuner.OutOfResources:
                     # Some configurations may be invalid and fail to compile.
                     continue
@@ -244,12 +410,27 @@ class BenchmarkWorker:
 
 def sort_config(config: BenchmarkConfig) -> BenchmarkConfig:
     return {
-        "BLOCK_SIZE_M": config["BLOCK_SIZE_M"],
+        "BLOCK_SIZE_M":
-        "BLOCK_SIZE_N": config["BLOCK_SIZE_N"],
+        config["BLOCK_SIZE_M"],
-        "BLOCK_SIZE_K": config["BLOCK_SIZE_K"],
+        "BLOCK_SIZE_N":
-        "GROUP_SIZE_M": config["GROUP_SIZE_M"],
+        config["BLOCK_SIZE_N"],
-        "num_warps": config["num_warps"],
+        "BLOCK_SIZE_K":
-        "num_stages": config["num_stages"],
+        config["BLOCK_SIZE_K"],
+        "GROUP_SIZE_M":
+        config["GROUP_SIZE_M"],
+        "num_warps":
+        config["num_warps"],
+        "num_stages":
+        config["num_stages"],
+        **({
+            "waves_per_eu": config["waves_per_eu"]
+        } if "waves_per_eu" in config else {}),
+        **({
+            "matrix_instr_nonkdim": config["matrix_instr_nonkdim"]
+        } if "matrix_instr_nonkdim" in config else {}),
+        **({
+            "kpack": config["kpack"]
+        } if "kpack" in config else {}),
     }
 
 
@@ -275,7 +456,8 @@ def save_configs(configs: Dict[int, BenchmarkConfig], num_experts: int,
 def main(args: argparse.Namespace):
     print(args)
 
-    config = AutoConfig.from_pretrained(args.model)
+    config = AutoConfig.from_pretrained(
+        args.model, trust_remote_code=args.trust_remote_code)
     if config.architectures[0] == "DbrxForCausalLM":
         E = config.ffn_config.moe_num_experts
         topk = config.ffn_config.moe_top_k
@@ -286,6 +468,11 @@ def main(args: argparse.Namespace):
         topk = config.num_experts_per_tok
         intermediate_size = config.intermediate_size
         shard_intermediate_size = 2 * intermediate_size // args.tp_size
+    elif config.architectures[0] == "DeepseekV3ForCausalLM":
+        E = config.n_routed_experts
+        topk = config.num_experts_per_tok
+        intermediate_size = config.moe_intermediate_size
+        shard_intermediate_size = 2 * intermediate_size // args.tp_size
     else:
         # Default: Mixtral.
         E = config.num_local_experts
@@ -294,7 +481,7 @@ def main(args: argparse.Namespace):
         shard_intermediate_size = 2 * intermediate_size // args.tp_size
 
     hidden_size = config.hidden_size
-    dtype = config.torch_dtype
+    dtype = torch.float16 if current_platform.is_rocm() else config.torch_dtype
     use_fp8_w8a8 = args.dtype == "fp8_w8a8"
     use_int8_w8a16 = args.dtype == "int8_w8a16"
 
@@ -322,7 +509,8 @@ def main(args: argparse.Namespace):
         return ray.get(outputs)
 
     if args.tune:
-        search_space = get_configs_compute_bound()
+        is_fp16 = not (use_fp8_w8a8 or use_int8_w8a16)
+        search_space = get_configs_compute_bound(is_fp16)
         print(f"Start tuning over {len(search_space)} configurations...")
 
         start = time.time()
@@ -354,7 +542,11 @@ if __name__ == "__main__":
     parser.add_argument("--model",
                         type=str,
                         default="mistralai/Mixtral-8x7B-Instruct-v0.1")
-    parser.add_argument("--tp-size", "-tp", type=int, default=2)
+    parser.add_argument("--tp-size",
+                        "-tp",
+                        "--tensor-parallel-size",
+                        type=int,
+                        default=2)
     parser.add_argument("--dtype",
                         type=str,
                         choices=["auto", "fp8_w8a8", "int8_w8a16"],
@@ -362,6 +554,7 @@ if __name__ == "__main__":
     parser.add_argument("--seed", type=int, default=0)
     parser.add_argument("--batch-size", type=int, required=False)
     parser.add_argument("--tune", action="store_true")
+    parser.add_argument("--trust-remote-code", action="store_true")
     args = parser.parse_args()
 
     main(args)

benchmarks/kernels/benchmark_paged_attention.py

@@ -1,3 +1,5 @@
+# SPDX-License-Identifier: Apache-2.0
+
 import random
 import time
 from typing import List, Optional
@@ -98,7 +100,9 @@ def main(
         start_time = time.perf_counter()
 
         # Using default kv_scale
-        k_scale = v_scale = 1.0
+        k_scale = v_scale = torch.tensor(1.0,
+                                         dtype=torch.float32,
+                                         device=device)
 
         for _ in range(num_iters):
             if version == "v1":

benchmarks/kernels/benchmark_quant.py

@@ -1,3 +1,5 @@
+# SPDX-License-Identifier: Apache-2.0
+
 import time
 
 import torch

benchmarks/kernels/benchmark_rmsnorm.py

@@ -1,3 +1,5 @@
+# SPDX-License-Identifier: Apache-2.0
+
 import itertools
 from typing import Optional, Tuple, Union
 

benchmarks/kernels/benchmark_rope.py

@@ -1,3 +1,5 @@
+# SPDX-License-Identifier: Apache-2.0
+
 from itertools import accumulate
 from typing import List, Optional
 

benchmarks/kernels/benchmark_shapes.py

@@ -1,3 +1,5 @@
+# SPDX-License-Identifier: Apache-2.0
+
 WEIGHT_SHAPES = {
     "ideal": [[4 * 256 * 32, 256 * 32]],
     "mistralai/Mistral-7B-v0.1/TP1": [

benchmarks/kernels/graph_machete_bench.py

@@ -1,3 +1,5 @@
+# SPDX-License-Identifier: Apache-2.0
+
 import math
 import pickle
 import re

benchmarks/kernels/utils.py

@@ -0,0 +1,212 @@
+# SPDX-License-Identifier: Apache-2.0
+
+import dataclasses
+from typing import Any, Callable, Iterable, Optional
+
+import torch
+import torch.utils.benchmark as TBenchmark
+from torch.utils.benchmark import Measurement as TMeasurement
+
+
+@dataclasses.dataclass
+class CudaGraphBenchParams:
+    num_ops_in_cuda_graph: int
+
+
+@dataclasses.dataclass
+class ArgPool:
+    """
+    When some argument of the benchmarking function is annotated with this type,
+    the benchmarking class (BenchMM) will collapse the argument to a pick a
+    single value from the given list of values, during function invocation.
+    For every invocation during a benchmarking run, it will choose a
+    different value from the list.
+    """
+    values: Iterable[Any]
+
+    def __getitem__(self, index):
+        return self.values[index]
+
+
+class Bench:
+
+    class ArgsIterator:
+
+        def __init__(self, args_list, kwargs_list):
+            assert len(args_list) == len(kwargs_list)
+            self.args_list = args_list
+            self.kwargs_list = kwargs_list
+            self.n = len(self.args_list)
+            self.idx = 0
+
+        def __next__(self):
+            while True:
+                yield (self.args_list[self.idx], self.kwargs_list[self.idx])
+                self.idx += 1
+                self.idx = self.idx % self.n
+
+        def reset(self):
+            self.idx = 0
+
+        @property
+        def n_args(self):
+            return self.n
+
+    def __init__(self, cuda_graph_params: Optional[CudaGraphBenchParams],
+                 label: str, sub_label: str, description: str, fn: Callable,
+                 *args, **kwargs):
+
+        self.cuda_graph_params = cuda_graph_params
+        self.use_cuda_graph = self.cuda_graph_params is not None
+        self.label = label
+        self.sub_label = sub_label
+        self.description = description
+        self.fn = fn
+
+        # Process args
+        self._args = args
+        self._kwargs = kwargs
+        self.args_list, self.kwargs_list = self.collapse_argpool(
+            *args, **kwargs)
+        self.args_iterator = self.ArgsIterator(self.args_list,
+                                               self.kwargs_list)
+
+        # Cudagraph runner
+        self.g = None
+        if self.use_cuda_graph:
+            self.g = self.get_cuda_graph_runner()
+
+        # benchmark run params
+        self.min_run_time = 1
+
+    def collapse_argpool(self, *args, **kwargs):
+        argpool_args = [arg for arg in args if isinstance(arg, ArgPool)] + [
+            arg for arg in kwargs.values() if isinstance(arg, ArgPool)
+        ]
+        if len(argpool_args) == 0:
+            return [args], [kwargs]
+
+        # Make sure all argpools are of the same size
+        argpool_size = len(argpool_args[0].values)
+        assert all([argpool_size == len(arg.values) for arg in argpool_args])
+
+        # create copies of the args
+        args_list = []
+        kwargs_list = []
+        for _ in range(argpool_size):
+            args_list.append(args)
+            kwargs_list.append(kwargs.copy())
+
+        for i in range(argpool_size):
+            # collapse args; Just pick the ith value
+            args_list[i] = tuple([
+                arg[i] if isinstance(arg, ArgPool) else arg
+                for arg in args_list[i]
+            ])
+
+            # collapse kwargs
+            kwargs_i = kwargs_list[i]
+            arg_pool_keys = [
+                k for k, v in kwargs_i.items() if isinstance(v, ArgPool)
+            ]
+            for k in arg_pool_keys:
+                # again just pick the ith value
+                kwargs_i[k] = kwargs_i[k][i]
+            kwargs_list[i] = kwargs_i
+
+        return args_list, kwargs_list
+
+    def get_cuda_graph_runner(self):
+        assert self.use_cuda_graph
+        assert self.args_iterator is not None
+
+        num_graph_ops = self.cuda_graph_params.num_ops_in_cuda_graph
+
+        # warmup
+        args_it = self.args_iterator.__next__()
+        for _ in range(2):
+            args, kwargs = next(args_it)
+            self.fn(*args, **kwargs)
+
+        self.args_iterator.reset()
+        args_it = self.args_iterator.__next__()
+        stream = torch.cuda.Stream()
+        with torch.cuda.stream(stream):
+            g = torch.cuda.CUDAGraph()
+            with torch.cuda.graph(g):
+                for _ in range(num_graph_ops):
+                    args, kwargs = next(args_it)
+                    self.fn(*args, **kwargs)
+        return g
+
+    def run_cudagrah(self) -> TMeasurement:
+        assert self.use_cuda_graph
+        globals = {'g': self.g}
+
+        return TBenchmark.Timer(
+            stmt="g.replay()",
+            globals=globals,
+            label=(
+                f"{self.label}"
+                f" | cugraph {self.cuda_graph_params.num_ops_in_cuda_graph} ops"
+            ),
+            sub_label=self.sub_label,
+            description=self.description,
+        ).blocked_autorange(min_run_time=self.min_run_time)
+
+    def run_eager(self) -> TMeasurement:
+        setup = None
+        stmt = None
+        globals = None
+
+        has_arg_pool = self.args_iterator.n_args > 1
+        if has_arg_pool:
+            setup = '''
+                    args_iterator.reset()
+                    args_it = args_iterator.__next__()
+                    '''
+            stmt = '''
+                    args, kwargs = next(args_it)
+                    fn(*args, **kwargs)
+                    '''
+            globals = {'fn': self.fn, 'args_iterator': self.args_iterator}
+        else:
+            # no arg pool. Just use the args and kwargs directly
+            self.args_iterator.reset()
+            args_it = self.args_iterator.__next__()
+            args, kwargs = next(args_it)
+
+            setup = ""
+            stmt = '''
+                    fn(*args, **kwargs)
+                   '''
+            globals = {'fn': self.fn, 'args': args, 'kwargs': kwargs}
+
+        return TBenchmark.Timer(
+            stmt=stmt,
+            setup=setup,
+            globals=globals,
+            label=self.label,
+            sub_label=self.sub_label,
+            description=self.description,
+        ).blocked_autorange(min_run_time=self.min_run_time)
+
+    def run(self) -> TMeasurement:
+        timer = None
+        if self.use_cuda_graph:  # noqa SIM108
+            timer = self.run_cudagrah()
+        else:
+            timer = self.run_eager()
+        if not timer.meets_confidence() or timer.has_warnings:
+            print("Doesn't meet confidence - re-running bench ...")
+            return self.run()
+        return timer
+
+    def __enter__(self):
+        return self
+
+    def __exit__(self, exc_type, exc_value, traceback):
+        if exc_type:
+            print(f"exc type {exc_type}")
+            print(f"exc value {exc_value}")
+            print(f"exc traceback {traceback}")

benchmarks/kernels/weight_shapes.py

@@ -1,3 +1,5 @@
+# SPDX-License-Identifier: Apache-2.0
+
 # Weight Shapes are in the format
 # ([K, N], TP_SPLIT_DIM)
 # Example:

benchmarks/overheads/benchmark_hashing.py

@@ -1,3 +1,5 @@
+# SPDX-License-Identifier: Apache-2.0
+
 import cProfile
 import pstats
 

cmake/cpu_extension.cmake

@@ -4,6 +4,11 @@ set(CMAKE_CXX_STANDARD_REQUIRED ON)
 set(CMAKE_CXX_EXTENSIONS ON)
 set(CMAKE_EXPORT_COMPILE_COMMANDS ON)
 
+if (${CMAKE_SYSTEM_NAME} MATCHES "Darwin")
+    set(MACOSX_FOUND TRUE)
+endif()
+
+
 #
 # Define environment variables for special configurations
 #
@@ -13,6 +18,9 @@ endif()
 
 include_directories("${CMAKE_SOURCE_DIR}/csrc")
 
+
+set (ENABLE_NUMA TRUE)
+
 #
 # Check the compile flags
 #
@@ -22,17 +30,27 @@ if (CMAKE_SYSTEM_PROCESSOR MATCHES "x86_64")
         "-mf16c"
     )
 endif()
+
+if(MACOSX_FOUND)
+    list(APPEND CXX_COMPILE_FLAGS
+        "-Xpreprocessor"
+        "-fopenmp"
+        "-DVLLM_CPU_EXTENSION")
+else()
     list(APPEND CXX_COMPILE_FLAGS
         "-fopenmp"
         "-DVLLM_CPU_EXTENSION")
+endif()
 
+if (NOT MACOSX_FOUND)
     execute_process(COMMAND cat /proc/cpuinfo
                     RESULT_VARIABLE CPUINFO_RET
                     OUTPUT_VARIABLE CPUINFO)
-
     if (NOT CPUINFO_RET EQUAL 0)
         message(FATAL_ERROR "Failed to check CPU features via /proc/cpuinfo")
     endif()
+endif()
+
 
 function (find_isa CPUINFO TARGET OUT)
     string(FIND ${CPUINFO} ${TARGET} ISA_FOUND)
@@ -54,12 +72,17 @@ endfunction()
 
 is_avx512_disabled(AVX512_DISABLED)
 
+if (MACOSX_FOUND AND CMAKE_SYSTEM_PROCESSOR STREQUAL "arm64")
+    set(APPLE_SILICON_FOUND TRUE)
+else()
     find_isa(${CPUINFO} "avx2" AVX2_FOUND)
     find_isa(${CPUINFO} "avx512f" AVX512_FOUND)
     find_isa(${CPUINFO} "POWER10" POWER10_FOUND)
     find_isa(${CPUINFO} "POWER9" POWER9_FOUND)
     find_isa(${CPUINFO} "asimd" ASIMD_FOUND) # Check for ARM NEON support
     find_isa(${CPUINFO} "bf16" ARM_BF16_FOUND) # Check for ARM BF16 support
+endif()
+
 
 if (AVX512_FOUND AND NOT AVX512_DISABLED)
     list(APPEND CXX_COMPILE_FLAGS
@@ -103,6 +126,9 @@ elseif (ASIMD_FOUND)
         set(MARCH_FLAGS "-march=armv8.2-a+dotprod+fp16")  
     endif()
     list(APPEND CXX_COMPILE_FLAGS ${MARCH_FLAGS})     
+elseif(APPLE_SILICON_FOUND)
+    message(STATUS "Apple Silicon Detected")
+    set(ENABLE_NUMA OFF)
 else()
     message(FATAL_ERROR "vLLM CPU backend requires AVX512, AVX2, Power9+ ISA or ARMv8 support.")
 endif()
@@ -139,7 +165,12 @@ endif()
 
 message(STATUS "CPU extension compile flags: ${CXX_COMPILE_FLAGS}")
 
+if(ENABLE_NUMA)
     list(APPEND LIBS numa)
+else()
+    message(STATUS "NUMA is disabled")
+    add_compile_definitions(-DVLLM_NUMA_DISABLED)
+endif()
 
 #
 # _C extension

cmake/hipify.py

@@ -1,4 +1,5 @@
 #!/usr/bin/env python3
+# SPDX-License-Identifier: Apache-2.0
 
 #
 # A command line tool for running pytorch's hipify preprocessor on CUDA

cmake/utils.cmake

@@ -58,8 +58,8 @@ function (hipify_sources_target OUT_SRCS NAME ORIG_SRCS)
   #
   set(SRCS ${ORIG_SRCS})
   set(CXX_SRCS ${ORIG_SRCS})
-  list(FILTER SRCS EXCLUDE REGEX "\.(cc)|(cpp)$")
+  list(FILTER SRCS EXCLUDE REGEX "\.(cc)|(cpp)|(hip)$")
-  list(FILTER CXX_SRCS INCLUDE REGEX "\.(cc)|(cpp)$")
+  list(FILTER CXX_SRCS INCLUDE REGEX "\.(cc)|(cpp)|(hip)$")
 
   #
   # Generate ROCm/HIP source file names from CUDA file names.
@@ -259,7 +259,7 @@ endmacro()
 #  in `SRC_CUDA_ARCHS` that is less or equal to the version in `TGT_CUDA_ARCHS`.
 # We have special handling for 9.0a, if 9.0a is in `SRC_CUDA_ARCHS` and 9.0 is
 #  in `TGT_CUDA_ARCHS` then we should remove 9.0a from `SRC_CUDA_ARCHS` and add
-#  9.0a to the result. 
+#  9.0a to the result (and remove 9.0 from TGT_CUDA_ARCHS). 
 # The result is stored in `OUT_CUDA_ARCHS`.
 #
 # Example:
@@ -270,32 +270,45 @@ endmacro()
 #
 function(cuda_archs_loose_intersection OUT_CUDA_ARCHS SRC_CUDA_ARCHS TGT_CUDA_ARCHS)
   list(REMOVE_DUPLICATES SRC_CUDA_ARCHS)
+  set(TGT_CUDA_ARCHS_ ${TGT_CUDA_ARCHS})
 
   # if 9.0a is in SRC_CUDA_ARCHS and 9.0 is in CUDA_ARCHS then we should
   # remove 9.0a from SRC_CUDA_ARCHS and add 9.0a to _CUDA_ARCHS
   set(_CUDA_ARCHS)
   if ("9.0a" IN_LIST SRC_CUDA_ARCHS)
     list(REMOVE_ITEM SRC_CUDA_ARCHS "9.0a")
-    if ("9.0" IN_LIST TGT_CUDA_ARCHS)
+    if ("9.0" IN_LIST TGT_CUDA_ARCHS_)
+      list(REMOVE_ITEM TGT_CUDA_ARCHS_ "9.0")
       set(_CUDA_ARCHS "9.0a")
     endif()
   endif()
 
   list(SORT SRC_CUDA_ARCHS COMPARE NATURAL ORDER ASCENDING)
 
-  # for each ARCH in CUDA_ARCHS find the highest arch in SRC_CUDA_ARCHS that is 
+  # for each ARCH in TGT_CUDA_ARCHS find the highest arch in SRC_CUDA_ARCHS that
-  # less or eqault to ARCH
+  # is less or equal to ARCH (but has the same major version since SASS binary
-  foreach(_ARCH ${CUDA_ARCHS})
+  # compatibility is only forward compatible within the same major version).
+  foreach(_ARCH ${TGT_CUDA_ARCHS_})
     set(_TMP_ARCH)
+    # Extract the major version of the target arch
+    string(REGEX REPLACE "^([0-9]+)\\..*$" "\\1" TGT_ARCH_MAJOR "${_ARCH}")
     foreach(_SRC_ARCH ${SRC_CUDA_ARCHS})
+      # Extract the major version of the source arch
+      string(REGEX REPLACE "^([0-9]+)\\..*$" "\\1" SRC_ARCH_MAJOR "${_SRC_ARCH}")
+      # Check major-version match AND version-less-or-equal
       if (_SRC_ARCH VERSION_LESS_EQUAL _ARCH)
-      set(_TMP_ARCH ${_SRC_ARCH})
+        if (SRC_ARCH_MAJOR STREQUAL TGT_ARCH_MAJOR)
+          set(_TMP_ARCH "${_SRC_ARCH}")
+        endif()
       else()
+        # If we hit a version greater than the target, we can break
         break()
       endif()
     endforeach()
+
+    # If we found a matching _TMP_ARCH, append it to _CUDA_ARCHS
     if (_TMP_ARCH)
-    list(APPEND _CUDA_ARCHS ${_TMP_ARCH})
+      list(APPEND _CUDA_ARCHS "${_TMP_ARCH}")
     endif()
   endforeach()
 

collect_env.py

@@ -1,3 +1,5 @@
+# SPDX-License-Identifier: Apache-2.0
+
 # ruff: noqa
 # code borrowed from https://github.com/pytorch/pytorch/blob/main/torch/utils/collect_env.py
 

csrc/activation_kernels.cu

@@ -9,8 +9,16 @@
 
 namespace vllm {
 
+template <typename scalar_t, scalar_t (*ACT_FN)(const scalar_t&),
+          bool act_first>
+__device__ __forceinline__ scalar_t compute(const scalar_t& x,
+                                            const scalar_t& y) {
+  return act_first ? ACT_FN(x) * y : x * ACT_FN(y);
+}
 // Activation and gating kernel template.
-template <typename scalar_t, scalar_t (*ACT_FN)(const scalar_t&)>
+
+template <typename scalar_t, scalar_t (*ACT_FN)(const scalar_t&),
+          bool act_first>
 __global__ void act_and_mul_kernel(
     scalar_t* __restrict__ out,          // [..., d]
     const scalar_t* __restrict__ input,  // [..., 2, d]
@@ -19,7 +27,7 @@ __global__ void act_and_mul_kernel(
   for (int64_t idx = threadIdx.x; idx < d; idx += blockDim.x) {
     const scalar_t x = VLLM_LDG(&input[token_idx * 2 * d + idx]);
     const scalar_t y = VLLM_LDG(&input[token_idx * 2 * d + d + idx]);
-    out[token_idx * d + idx] = ACT_FN(x) * y;
+    out[token_idx * d + idx] = compute<scalar_t, ACT_FN, act_first>(x, y);
   }
 }
 
@@ -55,7 +63,9 @@ __device__ __forceinline__ T gelu_tanh_kernel(const T& x) {
 }  // namespace vllm
 
 // Launch activation and gating kernel.
-#define LAUNCH_ACTIVATION_GATE_KERNEL(KERNEL)                            \
+// Use ACT_FIRST (bool) indicating whether to apply the activation function
+// first.
+#define LAUNCH_ACTIVATION_GATE_KERNEL(KERNEL, ACT_FIRST)                 \
   int d = input.size(-1) / 2;                                            \
   int64_t num_tokens = input.numel() / input.size(-1);                   \
   dim3 grid(num_tokens);                                                 \
@@ -64,7 +74,7 @@ __device__ __forceinline__ T gelu_tanh_kernel(const T& x) {
   const cudaStream_t stream = at::cuda::getCurrentCUDAStream();          \
   VLLM_DISPATCH_FLOATING_TYPES(                                          \
       input.scalar_type(), "act_and_mul_kernel", [&] {                   \
-        vllm::act_and_mul_kernel<scalar_t, KERNEL<scalar_t>>             \
+        vllm::act_and_mul_kernel<scalar_t, KERNEL<scalar_t>, ACT_FIRST>  \
             <<<grid, block, 0, stream>>>(out.data_ptr<scalar_t>(),       \
                                          input.data_ptr<scalar_t>(), d); \
       });
@@ -72,19 +82,27 @@ __device__ __forceinline__ T gelu_tanh_kernel(const T& x) {
 void silu_and_mul(torch::Tensor& out,    // [..., d]
                   torch::Tensor& input)  // [..., 2 * d]
 {
-  LAUNCH_ACTIVATION_GATE_KERNEL(vllm::silu_kernel);
+  LAUNCH_ACTIVATION_GATE_KERNEL(vllm::silu_kernel, true);
+}
+
+void mul_and_silu(torch::Tensor& out,    // [..., d]
+                  torch::Tensor& input)  // [..., 2 * d]
+{
+  // The difference between mul_and_silu and silu_and_mul is that mul_and_silu
+  // applies the silu to the latter half of the input.
+  LAUNCH_ACTIVATION_GATE_KERNEL(vllm::silu_kernel, false);
 }
 
 void gelu_and_mul(torch::Tensor& out,    // [..., d]
                   torch::Tensor& input)  // [..., 2 * d]
 {
-  LAUNCH_ACTIVATION_GATE_KERNEL(vllm::gelu_kernel);
+  LAUNCH_ACTIVATION_GATE_KERNEL(vllm::gelu_kernel, true);
 }
 
 void gelu_tanh_and_mul(torch::Tensor& out,    // [..., d]
                        torch::Tensor& input)  // [..., 2 * d]
 {
-  LAUNCH_ACTIVATION_GATE_KERNEL(vllm::gelu_tanh_kernel);
+  LAUNCH_ACTIVATION_GATE_KERNEL(vllm::gelu_tanh_kernel, true);
 }
 
 namespace vllm {

csrc/attention/attention_kernels.cuh

@@ -105,7 +105,7 @@ __device__ void paged_attention_kernel(
     const int max_num_blocks_per_seq,
     const float* __restrict__ alibi_slopes,  // [num_heads]
     const int q_stride, const int kv_block_stride, const int kv_head_stride,
-    const float k_scale, const float v_scale, const int tp_rank,
+    const float* k_scale, const float* v_scale, const int tp_rank,
     const int blocksparse_local_blocks, const int blocksparse_vert_stride,
     const int blocksparse_block_size, const int blocksparse_head_sliding_step) {
   const int seq_idx = blockIdx.y;
@@ -285,7 +285,7 @@ __device__ void paged_attention_kernel(
           Quant_vec k_vec_quant = *reinterpret_cast<const Quant_vec*>(
               k_ptr + offset1 * BLOCK_SIZE * x + offset2);
           k_vecs[j] = fp8::scaled_convert<K_vec, Quant_vec, KV_DTYPE>(
-              k_vec_quant, k_scale);
+              k_vec_quant, *k_scale);
         }
       }
 
@@ -415,7 +415,7 @@ __device__ void paged_attention_kernel(
               *reinterpret_cast<const V_quant_vec*>(v_ptr + offset);
           // Vector conversion from V_quant_vec to V_vec.
           v_vec = fp8::scaled_convert<V_vec, V_quant_vec, KV_DTYPE>(v_quant_vec,
-                                                                    v_scale);
+                                                                    *v_scale);
         }
         if (block_idx == num_seq_blocks - 1) {
           // NOTE(woosuk): When v_vec contains the tokens that are out of the
@@ -513,7 +513,7 @@ __global__ void paged_attention_v1_kernel(
     const int max_num_blocks_per_seq,
     const float* __restrict__ alibi_slopes,  // [num_heads]
     const int q_stride, const int kv_block_stride, const int kv_head_stride,
-    const float k_scale, const float v_scale, const int tp_rank,
+    const float* k_scale, const float* v_scale, const int tp_rank,
     const int blocksparse_local_blocks, const int blocksparse_vert_stride,
     const int blocksparse_block_size, const int blocksparse_head_sliding_step) {
   paged_attention_kernel<scalar_t, cache_t, HEAD_SIZE, BLOCK_SIZE, NUM_THREADS,
@@ -549,7 +549,7 @@ __global__ void paged_attention_v2_kernel(
     const int max_num_blocks_per_seq,
     const float* __restrict__ alibi_slopes,  // [num_heads]
     const int q_stride, const int kv_block_stride, const int kv_head_stride,
-    const float k_scale, const float v_scale, const int tp_rank,
+    const float* k_scale, const float* v_scale, const int tp_rank,
     const int blocksparse_local_blocks, const int blocksparse_vert_stride,
     const int blocksparse_block_size, const int blocksparse_head_sliding_step) {
   paged_attention_kernel<scalar_t, cache_t, HEAD_SIZE, BLOCK_SIZE, NUM_THREADS,

csrc/attention/paged_attention_v1.cu

@@ -41,7 +41,7 @@
           out_ptr, query_ptr, key_cache_ptr, value_cache_ptr, num_kv_heads, \
           scale, block_tables_ptr, seq_lens_ptr, max_num_blocks_per_seq,    \
           alibi_slopes_ptr, q_stride, kv_block_stride, kv_head_stride,      \
-          k_scale, v_scale, tp_rank, blocksparse_local_blocks,              \
+          k_scale_ptr, v_scale_ptr, tp_rank, blocksparse_local_blocks,      \
           blocksparse_vert_stride, blocksparse_block_size,                  \
           blocksparse_head_sliding_step);
 
@@ -53,10 +53,10 @@ void paged_attention_v1_launcher(
     torch::Tensor& out, torch::Tensor& query, torch::Tensor& key_cache,
     torch::Tensor& value_cache, int num_kv_heads, float scale,
     torch::Tensor& block_tables, torch::Tensor& seq_lens, int max_seq_len,
-    const c10::optional<torch::Tensor>& alibi_slopes, float k_scale,
+    const std::optional<torch::Tensor>& alibi_slopes, torch::Tensor& k_scale,
-    float v_scale, const int tp_rank, const int blocksparse_local_blocks,
+    torch::Tensor& v_scale, const int tp_rank,
-    const int blocksparse_vert_stride, const int blocksparse_block_size,
+    const int blocksparse_local_blocks, const int blocksparse_vert_stride,
-    const int blocksparse_head_sliding_step) {
+    const int blocksparse_block_size, const int blocksparse_head_sliding_step) {
   int num_seqs = query.size(0);
   int num_heads = query.size(1);
   int head_size = query.size(2);
@@ -80,6 +80,8 @@ void paged_attention_v1_launcher(
   CACHE_T* value_cache_ptr = reinterpret_cast<CACHE_T*>(value_cache.data_ptr());
   int* block_tables_ptr = block_tables.data_ptr<int>();
   int* seq_lens_ptr = seq_lens.data_ptr<int>();
+  const float* k_scale_ptr = reinterpret_cast<const float*>(k_scale.data_ptr());
+  const float* v_scale_ptr = reinterpret_cast<const float*>(v_scale.data_ptr());
 
   constexpr int NUM_WARPS = NUM_THREADS / WARP_SIZE;
   int padded_max_seq_len =
@@ -176,9 +178,10 @@ void paged_attention_v1(
     torch::Tensor& block_tables,  // [num_seqs, max_num_blocks_per_seq]
     torch::Tensor& seq_lens,      // [num_seqs]
     int64_t block_size, int64_t max_seq_len,
-    const c10::optional<torch::Tensor>& alibi_slopes,
+    const std::optional<torch::Tensor>& alibi_slopes,
-    const std::string& kv_cache_dtype, double k_scale, double v_scale,
+    const std::string& kv_cache_dtype, torch::Tensor& k_scale,
-    const int64_t tp_rank, const int64_t blocksparse_local_blocks,
+    torch::Tensor& v_scale, const int64_t tp_rank,
+    const int64_t blocksparse_local_blocks,
     const int64_t blocksparse_vert_stride, const int64_t blocksparse_block_size,
     const int64_t blocksparse_head_sliding_step) {
   const bool is_block_sparse = (blocksparse_vert_stride > 1);

csrc/attention/paged_attention_v2.cu

@@ -37,7 +37,7 @@
           exp_sums_ptr, max_logits_ptr, tmp_out_ptr, query_ptr, key_cache_ptr, \
           value_cache_ptr, num_kv_heads, scale, block_tables_ptr,              \
           seq_lens_ptr, max_num_blocks_per_seq, alibi_slopes_ptr, q_stride,    \
-          kv_block_stride, kv_head_stride, k_scale, v_scale, tp_rank,          \
+          kv_block_stride, kv_head_stride, k_scale_ptr, v_scale_ptr, tp_rank,  \
           blocksparse_local_blocks, blocksparse_vert_stride,                   \
           blocksparse_block_size, blocksparse_head_sliding_step);              \
   vllm::paged_attention_v2_reduce_kernel<T, HEAD_SIZE, NUM_THREADS,            \
@@ -54,10 +54,10 @@ void paged_attention_v2_launcher(
     torch::Tensor& tmp_out, torch::Tensor& query, torch::Tensor& key_cache,
     torch::Tensor& value_cache, int num_kv_heads, float scale,
     torch::Tensor& block_tables, torch::Tensor& seq_lens, int max_seq_len,
-    const c10::optional<torch::Tensor>& alibi_slopes, float k_scale,
+    const std::optional<torch::Tensor>& alibi_slopes, torch::Tensor& k_scale,
-    float v_scale, const int tp_rank, const int blocksparse_local_blocks,
+    torch::Tensor& v_scale, const int tp_rank,
-    const int blocksparse_vert_stride, const int blocksparse_block_size,
+    const int blocksparse_local_blocks, const int blocksparse_vert_stride,
-    const int blocksparse_head_sliding_step) {
+    const int blocksparse_block_size, const int blocksparse_head_sliding_step) {
   int num_seqs = query.size(0);
   int num_heads = query.size(1);
   int head_size = query.size(2);
@@ -84,6 +84,8 @@ void paged_attention_v2_launcher(
   CACHE_T* value_cache_ptr = reinterpret_cast<CACHE_T*>(value_cache.data_ptr());
   int* block_tables_ptr = block_tables.data_ptr<int>();
   int* seq_lens_ptr = seq_lens.data_ptr<int>();
+  const float* k_scale_ptr = reinterpret_cast<const float*>(k_scale.data_ptr());
+  const float* v_scale_ptr = reinterpret_cast<const float*>(v_scale.data_ptr());
 
   constexpr int NUM_WARPS = NUM_THREADS / WARP_SIZE;
   int max_num_partitions = DIVIDE_ROUND_UP(max_seq_len, PARTITION_SIZE);
@@ -187,9 +189,10 @@ void paged_attention_v2(
     torch::Tensor& block_tables,  // [num_seqs, max_num_blocks_per_seq]
     torch::Tensor& seq_lens,      // [num_seqs]
     int64_t block_size, int64_t max_seq_len,
-    const c10::optional<torch::Tensor>& alibi_slopes,
+    const std::optional<torch::Tensor>& alibi_slopes,
-    const std::string& kv_cache_dtype, double k_scale, double v_scale,
+    const std::string& kv_cache_dtype, torch::Tensor& k_scale,
-    const int64_t tp_rank, const int64_t blocksparse_local_blocks,
+    torch::Tensor& v_scale, const int64_t tp_rank,
+    const int64_t blocksparse_local_blocks,
     const int64_t blocksparse_vert_stride, const int64_t blocksparse_block_size,
     const int64_t blocksparse_head_sliding_step) {
   const bool is_block_sparse = (blocksparse_vert_stride > 1);

csrc/cache.h

@@ -15,18 +15,26 @@ void copy_blocks(std::vector<torch::Tensor> const& key_caches,
                  std::vector<torch::Tensor> const& value_caches,
                  const torch::Tensor& block_mapping);
 
+void copy_blocks_mla(std::vector<torch::Tensor> const& kv_caches,
+                     const torch::Tensor& block_mapping);
+
 void reshape_and_cache(torch::Tensor& key, torch::Tensor& value,
                        torch::Tensor& key_cache, torch::Tensor& value_cache,
                        torch::Tensor& slot_mapping,
-                       const std::string& kv_cache_dtype, const double k_scale,
+                       const std::string& kv_cache_dtype,
-                       const double v_scale);
+                       torch::Tensor& k_scale, torch::Tensor& v_scale);
 
 void reshape_and_cache_flash(torch::Tensor& key, torch::Tensor& value,
                              torch::Tensor& key_cache,
                              torch::Tensor& value_cache,
                              torch::Tensor& slot_mapping,
                              const std::string& kv_cache_dtype,
-                             const double k_scale, const double v_scale);
+                             torch::Tensor& k_scale, torch::Tensor& v_scale);
+
+void concat_and_cache_mla(torch::Tensor& kv_c, torch::Tensor& k_pe,
+                          torch::Tensor& kv_cache, torch::Tensor& slot_mapping,
+                          const std::string& kv_cache_dtype,
+                          torch::Tensor& scale);
 
 // Just for unittest
 void convert_fp8(torch::Tensor& dst_cache, torch::Tensor& src_cache,

csrc/cache_kernels.cu

@@ -46,7 +46,10 @@ void swap_blocks(torch::Tensor& src, torch::Tensor& dst,
   char* src_ptr = static_cast<char*>(src.data_ptr());
   char* dst_ptr = static_cast<char*>(dst.data_ptr());
 
-  const int64_t block_size_in_bytes = src.element_size() * src[0].numel();
+  // We use the stride instead of numel in case the cache is padded for memory
+  // alignment reasons, we assume the blocks data (inclusive of any padding)
+  // is contiguous in memory
+  const int64_t block_size_in_bytes = src.element_size() * src.stride(0);
   const at::cuda::OptionalCUDAGuard device_guard(
       src_device.is_cuda() ? src_device : dst_device);
   const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
@@ -93,6 +96,24 @@ __global__ void copy_blocks_kernel(int64_t* key_cache_ptrs,
   }
 }
 
+// Kernel for MLA, which works on a single joint kv_cache
+// Grid: (num_layers, num_pairs)
+template <typename scalar_t>
+__global__ void copy_blocks_mla_kernel(
+    int64_t* cache_ptrs, const int64_t* __restrict__ block_mapping,
+    const int mem_footprint_per_block) {
+  const int layer_idx = blockIdx.x;
+  const int pair_idx = blockIdx.y;
+  scalar_t* cache = reinterpret_cast<scalar_t*>(cache_ptrs[layer_idx]);
+  int64_t src_block = block_mapping[2 * pair_idx];
+  int64_t dst_block = block_mapping[2 * pair_idx + 1];
+  int64_t src_offset = src_block * mem_footprint_per_block;
+  int64_t dst_offset = dst_block * mem_footprint_per_block;
+  for (int i = threadIdx.x; i < mem_footprint_per_block; i += blockDim.x) {
+    cache[dst_offset + i] = cache[src_offset + i];
+  }
+}
+
 }  // namespace vllm
 
 // Note: the key_caches and value_caches vectors are constant but
@@ -147,6 +168,42 @@ void copy_blocks(std::vector<torch::Tensor> const& key_caches,
       }));
 }
 
+// copy blocks kernel for MLA (assumes a joint KV-cache)
+void copy_blocks_mla(std::vector<torch::Tensor> const& kv_caches,
+                     const torch::Tensor& block_mapping) {
+  int num_layers = kv_caches.size();
+  if (num_layers == 0) {
+    return;
+  }
+  torch::Device cache_device = kv_caches[0].device();
+  TORCH_CHECK(cache_device.is_cuda(), "kv_cache must be on CUDA");
+
+  std::vector<int64_t> cache_ptrs(num_layers);
+  for (int layer_idx = 0; layer_idx < num_layers; ++layer_idx) {
+    cache_ptrs[layer_idx] =
+        reinterpret_cast<int64_t>(kv_caches[layer_idx].data_ptr());
+  }
+  torch::Tensor cache_ptrs_tensor =
+      torch::from_blob(cache_ptrs.data(), {num_layers}, torch::kInt64)
+          .to(cache_device);
+
+  int num_pairs = block_mapping.size(0);
+  // We use the stride instead of numel in case the cache is padded for memory
+  // alignment reasons, we assume the blocks data (inclusive of any padding)
+  // is contiguous in memory
+  int mem_footprint_per_block = kv_caches[0].stride(0);
+  dim3 grid(num_layers, num_pairs);
+  dim3 block(std::min(1024, mem_footprint_per_block));
+  const at::cuda::OptionalCUDAGuard device_guard(cache_device);
+  const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
+  VLLM_DISPATCH_FLOATING_AND_BYTE_TYPES(
+      kv_caches[0].scalar_type(), "copy_blocks_mla_kernel", ([&] {
+        vllm::copy_blocks_mla_kernel<scalar_t><<<grid, block, 0, stream>>>(
+            cache_ptrs_tensor.data_ptr<int64_t>(),
+            block_mapping.data_ptr<int64_t>(), mem_footprint_per_block);
+      }));
+}
+
 namespace vllm {
 
 template <typename scalar_t, typename cache_t, Fp8KVCacheDataType kv_dt>
@@ -159,8 +216,8 @@ __global__ void reshape_and_cache_kernel(
                                          // block_size]
     const int64_t* __restrict__ slot_mapping,  // [num_tokens]
     const int key_stride, const int value_stride, const int num_heads,
-    const int head_size, const int block_size, const int x, const float k_scale,
+    const int head_size, const int block_size, const int x,
-    const float v_scale) {
+    const float* k_scale, const float* v_scale) {
   const int64_t token_idx = blockIdx.x;
   const int64_t slot_idx = slot_mapping[token_idx];
   if (slot_idx < 0) {
@@ -196,9 +253,9 @@ __global__ void reshape_and_cache_kernel(
       value_cache[tgt_value_idx] = tgt_value;
     } else {
       key_cache[tgt_key_idx] =
-          fp8::scaled_convert<cache_t, scalar_t, kv_dt>(tgt_key, k_scale);
+          fp8::scaled_convert<cache_t, scalar_t, kv_dt>(tgt_key, *k_scale);
       value_cache[tgt_value_idx] =
-          fp8::scaled_convert<cache_t, scalar_t, kv_dt>(tgt_value, v_scale);
+          fp8::scaled_convert<cache_t, scalar_t, kv_dt>(tgt_value, *v_scale);
     }
   }
 }
@@ -214,7 +271,7 @@ __global__ void reshape_and_cache_flash_kernel(
     const int64_t* __restrict__ slot_mapping,  // [num_tokens]
     const int block_stride, const int key_stride, const int value_stride,
     const int num_heads, const int head_size, const int block_size,
-    const float k_scale, const float v_scale) {
+    const float* k_scale, const float* v_scale) {
   const int64_t token_idx = blockIdx.x;
   const int64_t slot_idx = slot_mapping[token_idx];
   // NOTE: slot_idx can be -1 if the token is padded
@@ -239,12 +296,57 @@ __global__ void reshape_and_cache_flash_kernel(
       value_cache[tgt_key_value_idx] = tgt_value;
     } else {
       key_cache[tgt_key_value_idx] =
-          fp8::scaled_convert<cache_t, scalar_t, kv_dt>(tgt_key, k_scale);
+          fp8::scaled_convert<cache_t, scalar_t, kv_dt>(tgt_key, *k_scale);
       value_cache[tgt_key_value_idx] =
-          fp8::scaled_convert<cache_t, scalar_t, kv_dt>(tgt_value, v_scale);
+          fp8::scaled_convert<cache_t, scalar_t, kv_dt>(tgt_value, *v_scale);
     }
   }
 }
+
+template <typename scalar_t, typename cache_t, Fp8KVCacheDataType kv_dt>
+__global__ void concat_and_cache_mla_kernel(
+    const scalar_t* __restrict__ kv_c,  // [num_tokens, kv_lora_rank]
+    const scalar_t* __restrict__ k_pe,  // [num_tokens, pe_dim]
+    cache_t* __restrict__ kv_cache,  // [num_blocks, block_size, (kv_lora_rank
+                                     // + pe_dim)]
+    const int64_t* __restrict__ slot_mapping,  // [num_tokens]
+    const int block_stride,                    //
+    const int entry_stride,                    //
+    const int kv_c_stride,                     //
+    const int k_pe_stride,                     //
+    const int kv_lora_rank,                    //
+    const int pe_dim,                          //
+    const int block_size,                      //
+    const float* scale                         //
+) {
+  const int64_t token_idx = blockIdx.x;
+  const int64_t slot_idx = slot_mapping[token_idx];
+  // NOTE: slot_idx can be -1 if the token is padded
+  if (slot_idx < 0) {
+    return;
+  }
+  const int64_t block_idx = slot_idx / block_size;
+  const int64_t block_offset = slot_idx % block_size;
+
+  auto copy = [&](const scalar_t* __restrict__ src, cache_t* __restrict__ dst,
+                  int src_stride, int dst_stride, int size, int offset) {
+    for (int i = threadIdx.x; i < size; i += blockDim.x) {
+      const int64_t src_idx = token_idx * src_stride + i;
+      const int64_t dst_idx =
+          block_idx * block_stride + block_offset * entry_stride + i + offset;
+      if constexpr (kv_dt == Fp8KVCacheDataType::kAuto) {
+        dst[dst_idx] = src[src_idx];
+      } else {
+        dst[dst_idx] =
+            fp8::scaled_convert<cache_t, scalar_t, kv_dt>(src[src_idx], *scale);
+      }
+    }
+  };
+
+  copy(kv_c, kv_cache, kv_c_stride, block_stride, kv_lora_rank, 0);
+  copy(k_pe, kv_cache, k_pe_stride, block_stride, pe_dim, kv_lora_rank);
+}
+
 }  // namespace vllm
 
 // KV_T is the stored data type of kv-cache.
@@ -258,7 +360,9 @@ __global__ void reshape_and_cache_flash_kernel(
           reinterpret_cast<CACHE_T*>(key_cache.data_ptr()),           \
           reinterpret_cast<CACHE_T*>(value_cache.data_ptr()),         \
           slot_mapping.data_ptr<int64_t>(), key_stride, value_stride, \
-          num_heads, head_size, block_size, x, k_scale, v_scale);
+          num_heads, head_size, block_size, x,                        \
+          reinterpret_cast<const float*>(k_scale.data_ptr()),         \
+          reinterpret_cast<const float*>(v_scale.data_ptr()));
 
 void reshape_and_cache(
     torch::Tensor& key,    // [num_tokens, num_heads, head_size]
@@ -268,8 +372,8 @@ void reshape_and_cache(
     torch::Tensor&
         value_cache,  // [num_blocks, num_heads, head_size, block_size]
     torch::Tensor& slot_mapping,  // [num_tokens]
-    const std::string& kv_cache_dtype, const double k_scale,
+    const std::string& kv_cache_dtype, torch::Tensor& k_scale,
-    const double v_scale) {
+    torch::Tensor& v_scale) {
   int num_tokens = key.size(0);
   int num_heads = key.size(1);
   int head_size = key.size(2);
@@ -299,7 +403,9 @@ void reshape_and_cache(
           reinterpret_cast<CACHE_T*>(key_cache.data_ptr()),           \
           reinterpret_cast<CACHE_T*>(value_cache.data_ptr()),         \
           slot_mapping.data_ptr<int64_t>(), block_stride, key_stride, \
-          value_stride, num_heads, head_size, block_size, k_scale, v_scale);
+          value_stride, num_heads, head_size, block_size,             \
+          reinterpret_cast<const float*>(k_scale.data_ptr()),         \
+          reinterpret_cast<const float*>(v_scale.data_ptr()));
 
 void reshape_and_cache_flash(
     torch::Tensor& key,        // [num_tokens, num_heads, head_size]
@@ -308,8 +414,8 @@ void reshape_and_cache_flash(
     torch::Tensor&
         value_cache,  // [num_blocks, block_size, num_heads, head_size]
     torch::Tensor& slot_mapping,  // [num_tokens] or [num_actual_tokens]
-    const std::string& kv_cache_dtype, const double k_scale,
+    const std::string& kv_cache_dtype, torch::Tensor& k_scale,
-    const double v_scale) {
+    torch::Tensor& v_scale) {
   // NOTE(woosuk): In vLLM V1, key.size(0) can be different from
   // slot_mapping.size(0) because of padding for CUDA graphs.
   // In vLLM V0, key.size(0) is always equal to slot_mapping.size(0) because
@@ -339,6 +445,57 @@ void reshape_and_cache_flash(
                              CALL_RESHAPE_AND_CACHE_FLASH);
 }
 
+// KV_T is the stored data type of kv-cache.
+// CACHE_T is the data type of key and value tensors.
+// KV_DTYPE is the real data type of kv-cache.
+#define CALL_CONCAT_AND_CACHE_MLA(KV_T, CACHE_T, KV_DTYPE)              \
+  vllm::concat_and_cache_mla_kernel<KV_T, CACHE_T, KV_DTYPE>            \
+      <<<grid, block, 0, stream>>>(                                     \
+          reinterpret_cast<KV_T*>(kv_c.data_ptr()),                     \
+          reinterpret_cast<KV_T*>(k_pe.data_ptr()),                     \
+          reinterpret_cast<CACHE_T*>(kv_cache.data_ptr()),              \
+          slot_mapping.data_ptr<int64_t>(), block_stride, entry_stride, \
+          kv_c_stride, k_pe_stride, kv_lora_rank, pe_dim, block_size,   \
+          reinterpret_cast<const float*>(scale.data_ptr()));
+
+void concat_and_cache_mla(
+    torch::Tensor& kv_c,          // [num_tokens, kv_lora_rank]
+    torch::Tensor& k_pe,          // [num_tokens, pe_dim]
+    torch::Tensor& kv_cache,      // [num_blocks, block_size, (kv_lora_rank +
+                                  // pe_dim)]
+    torch::Tensor& slot_mapping,  // [num_tokens] or [num_actual_tokens]
+    const std::string& kv_cache_dtype, torch::Tensor& scale) {
+  // NOTE(woosuk): In vLLM V1, key.size(0) can be different from
+  // slot_mapping.size(0) because of padding for CUDA graphs.
+  // In vLLM V0, key.size(0) is always equal to slot_mapping.size(0) because
+  // both include padding.
+  // In vLLM V1, however, key.size(0) can be larger than slot_mapping.size(0)
+  // since key includes padding for CUDA graphs, while slot_mapping does not.
+  // In this case, slot_mapping.size(0) represents the actual number of tokens
+  // before padding.
+  // For compatibility with both cases, we use slot_mapping.size(0) as the
+  // number of tokens.
+  int num_tokens = slot_mapping.size(0);
+  int kv_lora_rank = kv_c.size(1);
+  int pe_dim = k_pe.size(1);
+  int block_size = kv_cache.size(1);
+
+  TORCH_CHECK(kv_cache.size(2) == kv_lora_rank + pe_dim);
+
+  int kv_c_stride = kv_c.stride(0);
+  int k_pe_stride = k_pe.stride(0);
+  int block_stride = kv_cache.stride(0);
+  int entry_stride = kv_cache.stride(1);
+
+  dim3 grid(num_tokens);
+  dim3 block(std::min(kv_lora_rank, 512));
+  const at::cuda::OptionalCUDAGuard device_guard(device_of(kv_c));
+  const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
+
+  DISPATCH_BY_KV_CACHE_DTYPE(kv_c.dtype(), kv_cache_dtype,
+                             CALL_CONCAT_AND_CACHE_MLA);
+}
+
 namespace vllm {
 
 template <typename Tout, typename Tin, Fp8KVCacheDataType kv_dt>

csrc/core/math.hpp

@@ -1,7 +1,14 @@
+#pragma once
+
 #include <climits>
 #include <iostream>
 
-inline uint32_t next_pow_2(uint32_t const num) {
+inline constexpr uint32_t next_pow_2(uint32_t const num) {
   if (num <= 1) return num;
   return 1 << (CHAR_BIT * sizeof(num) - __builtin_clz(num - 1));
 }
+
+template <typename T>
+inline constexpr std::enable_if_t<std::is_integral_v<T>, T> ceil_div(T a, T b) {
+  return (a + b - 1) / b;
+}