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

Signed-off-by: youkaichao <youkaichao@gmail.com>
[Misc] Remove duplicated DeepSeek V2/V3 model definition (#12793 )
2025-02-06 15:23:50 +08:00 · 2025-02-05 23:16:20 -08:00 · 2025-02-05 22:24:57 -08:00 · 2025-02-06 13:25:54 +08:00 · 2025-02-05 19:54:13 -08:00 · 2025-02-05 19:23:35 -08:00
1606 changed files with 152833 additions and 45210 deletions
--- a/.buildkite/check-wheel-size.py
+++ b/.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):
--- a/.buildkite/generate_index.py
+++ b/.buildkite/generate_index.py
@@ -0,0 +1,26 @@
 # SPDX-License-Identifier: Apache-2.0
 import argparse
 import os
 template = """<!DOCTYPE html>
 <html>
    <body>
    <h1>Links for vLLM</h1/>
        <a href="../{wheel_html_escaped}">{wheel}</a><br/>
    </body>
 </html>
 """
 parser = argparse.ArgumentParser()
 parser.add_argument("--wheel", help="The wheel path.", required=True)
 args = parser.parse_args()
 filename = os.path.basename(args.wheel)
 with open("index.html", "w") as f:
    print(f"Generated index.html for {args.wheel}")
    # cloudfront requires escaping the '+' character
    f.write(
        template.format(wheel=filename,
                        wheel_html_escaped=filename.replace("+", "%2B")))
--- a/.buildkite/lm-eval-harness/configs/Meta-Llama-3.2-1B-Instruct-INT8-compressed-tensors.yaml
+++ b/.buildkite/lm-eval-harness/configs/Meta-Llama-3.2-1B-Instruct-INT8-compressed-tensors.yaml
@@ -0,0 +1,11 @@
 # bash .buildkite/lm-eval-harness/run-lm-eval-gsm-vllm-baseline.sh -m neuralmagic/Llama-3.2-1B-Instruct-quantized.w8a8 -b "auto" -l 1000 -f 5 -t 1
 model_name: "neuralmagic/Llama-3.2-1B-Instruct-quantized.w8a8"
 tasks:
 - name: "gsm8k"
  metrics:
  - name: "exact_match,strict-match"
    value: 0.356
  - name: "exact_match,flexible-extract"
    value: 0.358
 limit: 1000
 num_fewshot: 5
--- a/.buildkite/lm-eval-harness/configs/SparseLlama3.1_2of4_fp8_compressed.yaml
+++ b/.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 
--- a/.buildkite/lm-eval-harness/configs/models-small.txt
+++ b/.buildkite/lm-eval-harness/configs/models-small.txt
@@ -1,6 +1,6 @@
 Meta-Llama-3-8B-Instruct.yaml
 Meta-Llama-3-8B-Instruct-FP8-compressed-tensors.yaml
-Meta-Llama-3-8B-Instruct-INT8-compressed-tensors.yaml
+Meta-Llama-3.2-1B-Instruct-INT8-compressed-tensors.yaml
 Meta-Llama-3-8B-Instruct-INT8-compressed-tensors-asym.yaml
 Meta-Llama-3-8B-Instruct-nonuniform-compressed-tensors.yaml
 Meta-Llama-3-8B-Instruct-Channelwise-compressed-tensors.yaml
--- a/.buildkite/lm-eval-harness/run-lm-eval-gsm-hf-baseline.sh
+++ b/.buildkite/lm-eval-harness/run-lm-eval-gsm-hf-baseline.sh
@@ -41,6 +41,6 @@ while getopts "m:b:l:f:" OPT; do
 done
 lm_eval --model hf \
-  --model_args pretrained=$MODEL,parallelize=True \
+  --model_args "pretrained=$MODEL,parallelize=True" \
-  --tasks gsm8k --num_fewshot $FEWSHOT --limit $LIMIT \
+  --tasks gsm8k --num_fewshot "$FEWSHOT" --limit "$LIMIT" \
-  --batch_size $BATCH_SIZE
+  --batch_size "$BATCH_SIZE"
--- a/.buildkite/lm-eval-harness/run-lm-eval-gsm-vllm-baseline.sh
+++ b/.buildkite/lm-eval-harness/run-lm-eval-gsm-vllm-baseline.sh
@@ -46,6 +46,6 @@ while getopts "m:b:l:f:t:" OPT; do
 done
 lm_eval --model vllm \
-  --model_args pretrained=$MODEL,tensor_parallel_size=$TP_SIZE,distributed_executor_backend="ray",trust_remote_code=true,max_model_len=4096 \
+  --model_args "pretrained=$MODEL,tensor_parallel_size=$TP_SIZE,distributed_executor_backend=ray,trust_remote_code=true,max_model_len=4096" \
-  --tasks gsm8k --num_fewshot $FEWSHOT --limit $LIMIT \
+  --tasks gsm8k --num_fewshot "$FEWSHOT" --limit "$LIMIT" \
-  --batch_size $BATCH_SIZE
+  --batch_size "$BATCH_SIZE"
--- a/.buildkite/lm-eval-harness/run-tests.sh
+++ b/.buildkite/lm-eval-harness/run-tests.sh
@@ -30,7 +30,7 @@ while getopts "c:t:" OPT; do
 done
 # Parse list of configs.
-IFS=$'\n' read -d '' -r -a MODEL_CONFIGS < $CONFIG
+IFS=$'\n' read -d '' -r -a MODEL_CONFIGS < "$CONFIG"
 for MODEL_CONFIG in "${MODEL_CONFIGS[@]}"
 do
--- a/.buildkite/lm-eval-harness/test_lm_eval_correctness.py
+++ b/.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/$MODEL.yaml
--- a/.buildkite/nightly-benchmarks/benchmark-pipeline.yaml
+++ b/.buildkite/nightly-benchmarks/benchmark-pipeline.yaml
@@ -1,5 +1,6 @@
 steps:
  - label: "Wait for container to be ready"
    key: wait-for-container-image
    agents:
      queue: A100
    plugins:
@@ -9,16 +10,18 @@ steps:
          - image: badouralix/curl-jq
            command:
            - sh .buildkite/nightly-benchmarks/scripts/wait-for-image.sh
-  - wait
+
  - label: "A100"
    # skip: "use this flag to conditionally skip the benchmark step, useful for PR testing"
    agents:
      queue: A100
    depends_on: wait-for-container-image
    plugins:
    - kubernetes:
        podSpec:
          priorityClassName: perf-benchmark
          containers:
-          - image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
+          - image: public.ecr.aws/q9t5s3a7/vllm-ci-postmerge-repo:$BUILDKITE_COMMIT
            command:
            - bash .buildkite/nightly-benchmarks/scripts/run-performance-benchmarks.sh
            resources:
@@ -41,20 +44,49 @@ steps:
          - name: devshm
            emptyDir:
              medium: Memory
  # - label: "H100"
  #   agents:
  #     queue: H100
  #   plugins:
  #   - docker#v5.11.0:
  #       image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
  #       command:
  #       - bash
  #       - .buildkite/nightly-benchmarks/run-benchmarks-suite.sh
  #       mount-buildkite-agent: true
  #       propagate-environment: true
  #       ipc: host
  #       gpus: all
  #       environment:
  #       - VLLM_USAGE_SOURCE
  #       - HF_TOKEN
  - label: "H200"
    # 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:$BUILDKITE_COMMIT
        command:
        - bash
        - .buildkite/nightly-benchmarks/scripts/run-performance-benchmarks.sh
        mount-buildkite-agent: true
        propagate-environment: true
        ipc: host
        gpus: 4,5,6,7
        volumes:
          - /data/benchmark-hf-cache:/root/.cache/huggingface
        environment:
        - VLLM_USAGE_SOURCE
        - HF_TOKEN
  #- block: "Run H100 Benchmark"
    #key: block-h100
    #depends_on: ~
  - label: "H100"
    # skip: "use this flag to conditionally skip the benchmark step, useful for PR testing"
    agents:
      queue: H100
    depends_on: wait-for-container-image
    plugins:
    - docker#v5.12.0:
        image: public.ecr.aws/q9t5s3a7/vllm-ci-postmerge-repo:$BUILDKITE_COMMIT
        command:
        - bash
        - .buildkite/nightly-benchmarks/scripts/run-performance-benchmarks.sh
        mount-buildkite-agent: true
        propagate-environment: true
        ipc: host
        gpus: all # see CUDA_VISIBLE_DEVICES for actual GPUs used
        volumes:
          - /data/benchmark-hf-cache:/root/.cache/huggingface
        environment:
        - VLLM_USAGE_SOURCE
        - HF_TOKEN
--- a/.buildkite/nightly-benchmarks/scripts/convert-results-json-to-markdown.py
+++ b/.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
@@ -56,7 +58,7 @@ serving_column_mapping = {
 def read_markdown(file):
    if os.path.exists(file):
-        with open(file, "r") as f:
+        with open(file) as f:
            return f.read() + "\n"
    else:
        return f"{file} not found.\n"
@@ -75,14 +77,14 @@ if __name__ == "__main__":
    # collect results
    for test_file in results_folder.glob("*.json"):
-        with open(test_file, "r") as f:
+        with open(test_file) as f:
            raw_result = json.loads(f.read())
        if "serving" in str(test_file):
            # this result is generated via `benchmark_serving.py`
            # attach the benchmarking command to raw_result
-            with open(test_file.with_suffix(".commands"), "r") as f:
+            with open(test_file.with_suffix(".commands")) as f:
                command = json.loads(f.read())
            raw_result.update(command)
@@ -97,7 +99,7 @@ if __name__ == "__main__":
            # this result is generated via `benchmark_latency.py`
            # attach the benchmarking command to raw_result
-            with open(test_file.with_suffix(".commands"), "r") as f:
+            with open(test_file.with_suffix(".commands")) as f:
                command = json.loads(f.read())
            raw_result.update(command)
@@ -119,7 +121,7 @@ if __name__ == "__main__":
            # this result is generated via `benchmark_throughput.py`
            # attach the benchmarking command to raw_result
-            with open(test_file.with_suffix(".commands"), "r") as f:
+            with open(test_file.with_suffix(".commands")) as f:
                command = json.loads(f.read())
            raw_result.update(command)
@@ -157,6 +159,18 @@ if __name__ == "__main__":
                                             throughput_results,
                                             serving_results)
    for df in [latency_results, serving_results, throughput_results]:
        if df.empty:
            continue
        # Sort all dataframes by their respective "Test name" columns
        df.sort_values(by="Test name", inplace=True)
        # The GPUs sometimes come in format of "GPUTYPE\nGPUTYPE\n...",
        # we want to turn it into "8xGPUTYPE"
        df["GPU"] = df["GPU"].apply(
            lambda x: f"{len(x.split('\n'))}x{x.split('\n')[0]}")
    # get markdown tables
    latency_md_table = tabulate(latency_results,
                                headers='keys',
--- a/.buildkite/nightly-benchmarks/scripts/download-tokenizer.py
+++ b/.buildkite/nightly-benchmarks/scripts/download-tokenizer.py
@@ -1,3 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
 import argparse
 from transformers import AutoTokenizer
--- a/.buildkite/nightly-benchmarks/scripts/generate-nightly-markdown.py
+++ b/.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
@@ -72,7 +74,7 @@ def main(args):
    # collect results
    for test_file in results_folder.glob("*_nightly_results.json"):
-        with open(test_file, "r") as f:
+        with open(test_file) as f:
            results = results + json.loads(f.read())
    # generate markdown table
@@ -80,7 +82,7 @@ def main(args):
    md_table = tabulate(df, headers='keys', tablefmt='pipe', showindex=False)
-    with open(args.description, "r") as f:
+    with open(args.description) as f:
        description = f.read()
    description = description.format(
--- a/.buildkite/nightly-benchmarks/scripts/get-lmdeploy-modelname.py
+++ b/.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")
--- a/.buildkite/nightly-benchmarks/scripts/launch-server.sh
+++ b/.buildkite/nightly-benchmarks/scripts/launch-server.sh
@@ -50,31 +50,30 @@ launch_trt_server() {
  git clone https://github.com/triton-inference-server/tensorrtllm_backend.git
  git lfs install
  cd tensorrtllm_backend
-  git checkout $trt_llm_version
+  git checkout "$trt_llm_version"
  tensorrtllm_backend_dir=$(pwd)
  git submodule update --init --recursive
  # build trtllm engine
  cd /tensorrtllm_backend
-  cd ./tensorrt_llm/examples/${model_type}
+  cd "./tensorrt_llm/examples/${model_type}"
  python3 convert_checkpoint.py \
-    --model_dir ${model_path} \
+    --model_dir "${model_path}" \
-    --dtype ${model_dtype} \
+    --dtype "${model_dtype}" \
-    --tp_size ${model_tp_size} \
+    --tp_size "${model_tp_size}" \
-    --output_dir ${trt_model_path}
+    --output_dir "${trt_model_path}"
  trtllm-build \
-    --checkpoint_dir ${trt_model_path} \
+    --checkpoint_dir "${trt_model_path}" \
    --use_fused_mlp \
    --reduce_fusion disable \
    --workers 8 \
-    --gpt_attention_plugin ${model_dtype} \
+    --gpt_attention_plugin "${model_dtype}" \
-    --gemm_plugin ${model_dtype} \
+    --gemm_plugin "${model_dtype}" \
-    --tp_size ${model_tp_size} \
+    --tp_size "${model_tp_size}" \
-    --max_batch_size ${max_batch_size} \
+    --max_batch_size "${max_batch_size}" \
-    --max_input_len ${max_input_len} \
+    --max_input_len "${max_input_len}" \
-    --max_seq_len ${max_seq_len} \
+    --max_seq_len "${max_seq_len}" \
-    --max_num_tokens ${max_num_tokens} \
+    --max_num_tokens "${max_num_tokens}" \
-    --output_dir ${trt_engine_path}
+    --output_dir "${trt_engine_path}"
  # handle triton protobuf files and launch triton server
  cd /tensorrtllm_backend
@@ -82,15 +81,15 @@ launch_trt_server() {
  cp -r all_models/inflight_batcher_llm/* triton_model_repo/
  cd triton_model_repo
  rm -rf ./tensorrt_llm/1/*
-  cp -r ${trt_engine_path}/* ./tensorrt_llm/1
+  cp -r "${trt_engine_path}"/* ./tensorrt_llm/1
  python3 ../tools/fill_template.py -i tensorrt_llm/config.pbtxt triton_backend:tensorrtllm,engine_dir:/tensorrtllm_backend/triton_model_repo/tensorrt_llm/1,decoupled_mode:true,batching_strategy:inflight_fused_batching,batch_scheduler_policy:guaranteed_no_evict,exclude_input_in_output:true,triton_max_batch_size:2048,max_queue_delay_microseconds:0,max_beam_width:1,max_queue_size:2048,enable_kv_cache_reuse:false
-  python3 ../tools/fill_template.py -i preprocessing/config.pbtxt triton_max_batch_size:2048,tokenizer_dir:$model_path,preprocessing_instance_count:5
+  python3 ../tools/fill_template.py -i preprocessing/config.pbtxt "triton_max_batch_size:2048,tokenizer_dir:$model_path,preprocessing_instance_count:5"
-  python3 ../tools/fill_template.py -i postprocessing/config.pbtxt triton_max_batch_size:2048,tokenizer_dir:$model_path,postprocessing_instance_count:5,skip_special_tokens:false
+  python3 ../tools/fill_template.py -i postprocessing/config.pbtxt "triton_max_batch_size:2048,tokenizer_dir:$model_path,postprocessing_instance_count:5,skip_special_tokens:false"
-  python3 ../tools/fill_template.py -i ensemble/config.pbtxt triton_max_batch_size:$max_batch_size
+  python3 ../tools/fill_template.py -i ensemble/config.pbtxt triton_max_batch_size:"$max_batch_size"
-  python3 ../tools/fill_template.py -i tensorrt_llm_bls/config.pbtxt triton_max_batch_size:$max_batch_size,decoupled_mode:true,accumulate_tokens:"False",bls_instance_count:1
+  python3 ../tools/fill_template.py -i tensorrt_llm_bls/config.pbtxt "triton_max_batch_size:$max_batch_size,decoupled_mode:true,accumulate_tokens:False,bls_instance_count:1"
  cd /tensorrtllm_backend
  python3 scripts/launch_triton_server.py \
-    --world_size=${model_tp_size} \
+    --world_size="${model_tp_size}" \
    --model_repo=/tensorrtllm_backend/triton_model_repo &
 }
@@ -98,10 +97,7 @@ launch_trt_server() {
 launch_tgi_server() {
  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')
  server_args=$(json2args "$server_params")
  if echo "$common_params" | jq -e 'has("fp8")' >/dev/null; then
@@ -129,10 +125,7 @@ launch_tgi_server() {
 launch_lmdeploy_server() {
  model=$(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 "$common_params" | jq -r '.num_prompts')
  server_args=$(json2args "$server_params")
  server_command="lmdeploy serve api_server $model \
@@ -149,10 +142,7 @@ launch_sglang_server() {
  model=$(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 "$common_params" | jq -r '.num_prompts')
  server_args=$(json2args "$server_params")
  if echo "$common_params" | jq -e 'has("fp8")' >/dev/null; then
@@ -185,10 +175,7 @@ launch_vllm_server() {
  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')
  server_args=$(json2args "$server_params")
  if echo "$common_params" | jq -e 'has("fp8")' >/dev/null; then
@@ -217,19 +204,19 @@ launch_vllm_server() {
 main() {
-  if [[ $CURRENT_LLM_SERVING_ENGINE == "trt" ]]; then
+  if [[ "$CURRENT_LLM_SERVING_ENGINE" == "trt" ]]; then
    launch_trt_server
  fi
-  if [[ $CURRENT_LLM_SERVING_ENGINE == "tgi" ]]; then
+  if [[ "$CURRENT_LLM_SERVING_ENGINE" == "tgi" ]]; then
    launch_tgi_server
  fi
-  if [[ $CURRENT_LLM_SERVING_ENGINE == "lmdeploy" ]]; then
+  if [[ "$CURRENT_LLM_SERVING_ENGINE" == "lmdeploy" ]]; then
    launch_lmdeploy_server
  fi
-  if [[ $CURRENT_LLM_SERVING_ENGINE == "sglang" ]]; then
+  if [[ "$CURRENT_LLM_SERVING_ENGINE" == "sglang" ]]; then
    launch_sglang_server
  fi
--- a/.buildkite/nightly-benchmarks/scripts/nightly-annotate.sh
+++ b/.buildkite/nightly-benchmarks/scripts/nightly-annotate.sh
@@ -16,10 +16,10 @@ main() {
    fi
    # initial annotation
-    description="$VLLM_SOURCE_CODE_LOC/.buildkite/nightly-benchmarks/nightly-descriptions.md"
+    #description="$VLLM_SOURCE_CODE_LOC/.buildkite/nightly-benchmarks/nightly-descriptions.md"
    # download results
-    cd $VLLM_SOURCE_CODE_LOC/benchmarks
+    cd "$VLLM_SOURCE_CODE_LOC/benchmarks"
    mkdir -p results/
    /workspace/buildkite-agent artifact download 'results/*nightly_results.json' results/
    ls
@@ -30,20 +30,20 @@ main() {
    /workspace/buildkite-agent artifact upload "results.zip"
    # upload benchmarking scripts
-    cd $VLLM_SOURCE_CODE_LOC/
+    cd "$VLLM_SOURCE_CODE_LOC/"
    zip -r nightly-benchmarks.zip .buildkite/ benchmarks/
    /workspace/buildkite-agent artifact upload "nightly-benchmarks.zip"
-    cd $VLLM_SOURCE_CODE_LOC/.buildkite/nightly-benchmarks/
+    cd "$VLLM_SOURCE_CODE_LOC/.buildkite/nightly-benchmarks/"
    # upload benchmarking pipeline
    /workspace/buildkite-agent artifact upload "nightly-pipeline.yaml"
-    cd $VLLM_SOURCE_CODE_LOC/.buildkite/nightly-benchmarks/
+    cd "$VLLM_SOURCE_CODE_LOC/.buildkite/nightly-benchmarks/"
    /workspace/buildkite-agent annotate --style "success" --context "nightly-benchmarks-results" --append < nightly-annotation.md
-    # 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
--- a/.buildkite/nightly-benchmarks/scripts/run-nightly-benchmarks.sh
+++ b/.buildkite/nightly-benchmarks/scripts/run-nightly-benchmarks.sh
@@ -12,7 +12,7 @@ check_gpus() {
    echo "Need at least 1 GPU to run benchmarking."
    exit 1
  fi
-  declare -g gpu_type=$(echo $(nvidia-smi --query-gpu=name --format=csv,noheader) | awk '{print $2}')
+  declare -g gpu_type="$(nvidia-smi --query-gpu=name --format=csv,noheader | awk '{print $2}')"
  echo "GPU type is $gpu_type"
 }
@@ -102,7 +102,7 @@ kill_gpu_processes() {
  pkill -f text-generation
  pkill -f lmdeploy
-  while [ $(nvidia-smi --query-gpu=memory.used --format=csv,noheader,nounits | head -n 1) -ge 1000 ]; do
+  while [ "$(nvidia-smi --query-gpu=memory.used --format=csv,noheader,nounits | head -n 1)" -ge 1000 ]; do
    sleep 1
  done
 }
@@ -119,8 +119,8 @@ wait_for_server() {
 ensure_installed() {
  # Ensure that the given command is installed by apt-get
  local cmd=$1
-  if ! which $cmd >/dev/null; then
+  if ! which "$cmd" >/dev/null; then
-    apt-get update && apt-get install -y $cmd
+    apt-get update && apt-get install -y "$cmd"
  fi
 }
@@ -173,13 +173,11 @@ run_serving_tests() {
      echo "Reuse previous server for test case $test_name"
    else
      kill_gpu_processes
-      bash $VLLM_SOURCE_CODE_LOC/.buildkite/nightly-benchmarks/scripts/launch-server.sh \
+      bash "$VLLM_SOURCE_CODE_LOC/.buildkite/nightly-benchmarks/scripts/launch-server.sh" \
        "$server_params" "$common_params"
    fi
-    wait_for_server
+    if wait_for_server; then
    if [ $? -eq 0 ]; then
      echo ""
      echo "$CURRENT_LLM_SERVING_ENGINE server is up and running."
    else
@@ -190,13 +188,13 @@ run_serving_tests() {
    # prepare tokenizer
    # this is required for lmdeploy.
-    cd $VLLM_SOURCE_CODE_LOC/benchmarks
+    cd "$VLLM_SOURCE_CODE_LOC/benchmarks"
    rm -rf /tokenizer_cache
    mkdir /tokenizer_cache
    python3 ../.buildkite/nightly-benchmarks/scripts/download-tokenizer.py \
      --model "$model" \
      --cachedir /tokenizer_cache
-    cd $VLLM_SOURCE_CODE_LOC/benchmarks
+    cd "$VLLM_SOURCE_CODE_LOC/benchmarks"
    # change model name for lmdeploy (it will not follow standard hf name)
@@ -303,15 +301,113 @@ 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() {
  # download sharegpt dataset
-  cd $VLLM_SOURCE_CODE_LOC/benchmarks
+  cd "$VLLM_SOURCE_CODE_LOC/benchmarks"
  wget https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/resolve/main/ShareGPT_V3_unfiltered_cleaned_split.json
  # duplicate sonnet by 4x, to allow benchmarking with input length 2048
-  cd $VLLM_SOURCE_CODE_LOC/benchmarks
+  cd "$VLLM_SOURCE_CODE_LOC/benchmarks"
  echo "" > sonnet_4x.txt
  for _ in {1..4}
  do
@@ -330,26 +426,35 @@ 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
-  cd $VLLM_SOURCE_CODE_LOC/benchmarks
+  cd "$VLLM_SOURCE_CODE_LOC/benchmarks"
  declare -g RESULTS_FOLDER=results/
  mkdir -p $RESULTS_FOLDER
-  BENCHMARK_ROOT=$VLLM_SOURCE_CODE_LOC/.buildkite/nightly-benchmarks/
+  BENCHMARK_ROOT="$VLLM_SOURCE_CODE_LOC/.buildkite/nightly-benchmarks/"
  # run the test
-  run_serving_tests $BENCHMARK_ROOT/tests/nightly-tests.json
+  run_serving_tests "$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 $BENCHMARK_ROOT/scripts/summary-nightly-results.py
+  python3 "$BENCHMARK_ROOT/scripts/summary-nightly-results.py"
  upload_to_buildkite
 }
--- a/.buildkite/nightly-benchmarks/scripts/run-performance-benchmarks.sh
+++ b/.buildkite/nightly-benchmarks/scripts/run-performance-benchmarks.sh
@@ -6,6 +6,7 @@
 # Do not set -e, as the mixtral 8x22B model tends to crash occasionally
 # and we still want to see other benchmarking results even when mixtral crashes.
 set -x
 set -o pipefail
 check_gpus() {
@@ -17,7 +18,7 @@ check_gpus() {
    echo "Need at least 1 GPU to run benchmarking."
    exit 1
  fi
-  declare -g gpu_type=$(echo $(nvidia-smi --query-gpu=name --format=csv,noheader) | awk '{print $2}')
+  declare -g gpu_type=$(nvidia-smi --query-gpu=name --format=csv,noheader | awk '{print $2}')
  echo "GPU type is $gpu_type"
 }
@@ -85,15 +86,11 @@ kill_gpu_processes() {
  ps -aux
  lsof -t -i:8000 | xargs -r kill -9
-  pkill -f pt_main_thread
+  pgrep python3 | xargs -r kill -9
  # this line doesn't work now
  # ps aux | grep python | grep openai | awk '{print $2}' | xargs -r kill -9
  pkill -f python3
  pkill -f /usr/bin/python3
  # wait until GPU memory usage smaller than 1GB
-  while [ $(nvidia-smi --query-gpu=memory.used --format=csv,noheader,nounits | head -n 1) -ge 1000 ]; do
+  while [ "$(nvidia-smi --query-gpu=memory.used --format=csv,noheader,nounits | head -n 1)" -ge 1000 ]; do
    sleep 1
  done
@@ -117,7 +114,7 @@ upload_to_buildkite() {
  fi
  # Use the determined command to annotate and upload artifacts
-  $BUILDKITE_AGENT_COMMAND annotate --style "info" --context "$BUILDKITE_LABEL-benchmark-results" <$RESULTS_FOLDER/benchmark_results.md
+  $BUILDKITE_AGENT_COMMAND annotate --style "info" --context "$BUILDKITE_LABEL-benchmark-results" < "$RESULTS_FOLDER/benchmark_results.md"
  $BUILDKITE_AGENT_COMMAND artifact upload "$RESULTS_FOLDER/*"
 }
@@ -150,7 +147,7 @@ run_latency_tests() {
    # check if there is enough GPU to run the test
    tp=$(echo "$latency_params" | jq -r '.tensor_parallel_size')
    if [[ $gpu_count -lt $tp ]]; then
-      echo "Required tensor-parallel-size $tp but only $gpu_count GPU found. Skip testcase $testname."
+      echo "Required tensor-parallel-size $tp but only $gpu_count GPU found. Skip testcase $test_name."
      continue
    fi
@@ -206,9 +203,9 @@ run_throughput_tests() {
    throughput_args=$(json2args "$throughput_params")
    # check if there is enough GPU to run the test
-    tp=$(echo $throughput_params | jq -r '.tensor_parallel_size')
+    tp=$(echo "$throughput_params" | jq -r '.tensor_parallel_size')
    if [[ $gpu_count -lt $tp ]]; then
-      echo "Required tensor-parallel-size $tp but only $gpu_count GPU found. Skip testcase $testname."
+      echo "Required tensor-parallel-size $tp but only $gpu_count GPU found. Skip testcase $test_name."
      continue
    fi
@@ -270,7 +267,7 @@ run_serving_tests() {
    # check if there is enough GPU to run the test
    tp=$(echo "$server_params" | jq -r '.tensor_parallel_size')
    if [[ $gpu_count -lt $tp ]]; then
-      echo "Required tensor-parallel-size $tp but only $gpu_count GPU found. Skip testcase $testname."
+      echo "Required tensor-parallel-size $tp but only $gpu_count GPU found. Skip testcase $test_name."
      continue
    fi
@@ -278,7 +275,7 @@ run_serving_tests() {
    server_model=$(echo "$server_params" | jq -r '.model')
    client_model=$(echo "$client_params" | jq -r '.model')
    if [[ $server_model != "$client_model" ]]; then
-      echo "Server model and client model must be the same. Skip testcase $testname."
+      echo "Server model and client model must be the same. Skip testcase $test_name."
      continue
    fi
@@ -289,12 +286,11 @@ run_serving_tests() {
    # run the server
    echo "Running test case $test_name"
    echo "Server command: $server_command"
-    eval "$server_command" &
+    bash -c "$server_command" &
    server_pid=$!
    # wait until the server is alive
-    wait_for_server
+    if wait_for_server; then
    if [ $? -eq 0 ]; then
      echo ""
      echo "vllm server is up and running."
    else
@@ -323,7 +319,7 @@ run_serving_tests() {
      echo "Running test case $test_name with qps $qps"
      echo "Client command: $client_command"
-      eval "$client_command"
+      bash -c "$client_command"
      # record the benchmarking commands
      jq_output=$(jq -n \
--- a/.buildkite/nightly-benchmarks/scripts/summary-nightly-results.py
+++ b/.buildkite/nightly-benchmarks/scripts/summary-nightly-results.py
@@ -1,3 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
 import datetime
 import json
 import os
@@ -36,11 +38,11 @@ if __name__ == "__main__":
    # collect results
    for test_file in results_folder.glob("*.json"):
-        with open(test_file, "r") as f:
+        with open(test_file) as f:
            raw_result = json.loads(f.read())
        # attach the benchmarking command to raw_result
-        with open(test_file.with_suffix(".commands"), "r") as f:
+        with open(test_file.with_suffix(".commands")) as f:
            command = json.loads(f.read())
        raw_result.update(command)
--- a/.buildkite/nightly-benchmarks/scripts/wait-for-image.sh
+++ b/.buildkite/nightly-benchmarks/scripts/wait-for-image.sh
@@ -1,12 +1,12 @@
 #!/bin/sh
-TOKEN=$(curl -s -L "https://public.ecr.aws/token?service=public.ecr.aws&scope=repository:q9t5s3a7/vllm-ci-test-repo:pull" | jq -r .token)
+TOKEN=$(curl -s -L "https://public.ecr.aws/token?service=public.ecr.aws&scope=repository:q9t5s3a7/vllm-ci-postmerge-repo:pull" | jq -r .token)
-URL="https://public.ecr.aws/v2/q9t5s3a7/vllm-ci-test-repo/manifests/$BUILDKITE_COMMIT"
+URL="https://public.ecr.aws/v2/q9t5s3a7/vllm-ci-postmerge-repo/manifests/$BUILDKITE_COMMIT"
 TIMEOUT_SECONDS=10
 retries=0
 while [ $retries -lt 1000 ]; do
-    if [ $(curl -s --max-time $TIMEOUT_SECONDS -L -H "Authorization: Bearer $TOKEN" -o /dev/null -w "%{http_code}" $URL) -eq 200 ]; then
+    if [ "$(curl -s --max-time "$TIMEOUT_SECONDS" -L -H "Authorization: Bearer $TOKEN" -o /dev/null -w "%{http_code}" "$URL")" -eq 200 ]; then
        exit 0
    fi
--- a/.buildkite/nightly-benchmarks/tests/genai-perf-tests.json
+++ b/.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": {
        }
    }
 ]
--- a/.buildkite/release-pipeline.yaml
+++ b/.buildkite/release-pipeline.yaml
@@ -1,33 +1,77 @@
 steps:
  - label: "Build wheel - CUDA 12.1"
    agents:
-      queue: cpu_queue
+      queue: cpu_queue_postmerge
    commands:
      - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=12.1.0 --tag vllm-ci:build-image --target build --progress plain ."
      - "mkdir artifacts"
      - "docker run --rm -v $(pwd)/artifacts:/artifacts_host vllm-ci:build-image bash -c 'cp -r dist /artifacts_host && chmod -R a+rw /artifacts_host'"
-      # rename the files to change linux -> manylinux1
+      - "bash .buildkite/upload-wheels.sh"
      - "for f in artifacts/dist/*.whl; do mv -- \"$$f\" \"$${f/linux/manylinux1}\"; done"
      - "mv artifacts/dist/$(ls artifacts/dist) artifacts/dist/vllm-1.0.0.dev-cp38-abi3-manylinux1_x86_64.whl"
      - "aws s3 cp artifacts/dist/vllm-1.0.0.dev-cp38-abi3-manylinux1_x86_64.whl s3://vllm-wheels/$BUILDKITE_COMMIT/vllm-1.0.0.dev-cp38-abi3-manylinux1_x86_64.whl"
      - "aws s3 cp artifacts/dist/vllm-1.0.0.dev-cp38-abi3-manylinux1_x86_64.whl s3://vllm-wheels/nightly/vllm-1.0.0.dev-cp38-abi3-manylinux1_x86_64.whl"
    env:
      DOCKER_BUILDKIT: "1"
-  - block: "Build CUDA 11.8 wheel"
+  # Note(simon): We can always build CUDA 11.8 wheel to ensure the build is working.
-    key: block-build-cu118-wheel
+  # However, this block can be uncommented to save some compute hours.
  # - block: "Build CUDA 11.8 wheel"
  #   key: block-build-cu118-wheel
  - label: "Build wheel - CUDA 11.8"
-    depends_on: block-build-cu118-wheel
+    # depends_on: block-build-cu118-wheel
    agents:
-      queue: cpu_queue
+      queue: cpu_queue_postmerge
    commands:
      - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=11.8.0 --tag vllm-ci:build-image --target build --progress plain ."
      - "mkdir artifacts"
      - "docker run --rm -v $(pwd)/artifacts:/artifacts_host vllm-ci:build-image bash -c 'cp -r dist /artifacts_host && chmod -R a+rw /artifacts_host'"
-      # rename the files to change linux -> manylinux1
+      - "bash .buildkite/upload-wheels.sh"
-      - "for f in artifacts/dist/*.whl; do mv -- \"$$f\" \"$${f/linux/manylinux1}\"; done"
+    env:
-      - "aws s3 cp --recursive artifacts/dist s3://vllm-wheels/$BUILDKITE_COMMIT/"
+      DOCKER_BUILDKIT: "1"
-      - "aws s3 cp --recursive artifacts/dist s3://vllm-wheels/nightly/"
+
  - block: "Build release image"
    depends_on: ~
    key: block-release-image-build
  - label: "Build release image"
    depends_on: block-release-image-build
    agents:
      queue: cpu_queue_postmerge
    commands:
      - "aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin public.ecr.aws/q9t5s3a7"
      - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=12.1.0 --tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT --target vllm-openai --progress plain ."
      - "docker push public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT"
  - label: "Build and publish TPU release image"
    depends_on: ~
    if: build.env("NIGHTLY") == "1"
    agents:
      queue: tpu_queue_postmerge
    commands:
      - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --tag vllm/vllm-tpu:nightly --tag vllm/vllm-tpu:$BUILDKITE_COMMIT --progress plain -f Dockerfile.tpu ."
      - "docker push vllm/vllm-tpu:nightly"
      - "docker push vllm/vllm-tpu:$BUILDKITE_COMMIT"
    plugins:
      - docker-login#v3.0.0:
          username: vllm
          password-env: DOCKERHUB_TOKEN
    env:
      DOCKER_BUILDKIT: "1"
  - input: "Provide Release version here"
    fields:
      - text: "What is the release version?"
        key: "release-version"
  - block: "Build CPU release image"
    key: block-cpu-release-image-build
    depends_on: ~
  - label: "Build and publish CPU release image"
    depends_on: block-cpu-release-image-build
    agents:
      queue: cpu_queue_postmerge
    commands:
      - "aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin public.ecr.aws/q9t5s3a7"
      - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg GIT_REPO_CHECK=1 --tag public.ecr.aws/q9t5s3a7/vllm-cpu-release-repo:$(buildkite-agent meta-data get release-version) --progress plain -f Dockerfile.cpu ."
      - "docker push public.ecr.aws/q9t5s3a7/vllm-cpu-release-repo:$(buildkite-agent meta-data get release-version)"
    env:
      DOCKER_BUILDKIT: "1"
--- a/.buildkite/run-amd-test.sh
+++ b/.buildkite/run-amd-test.sh
@@ -1,3 +1,5 @@
 #!/bin/bash
 # This script runs test inside the corresponding ROCm docker container.
 set -o pipefail
@@ -31,8 +33,8 @@ cleanup_docker() {
    echo "Disk usage is above $threshold%. Cleaning up Docker images and volumes..."
    # Remove dangling images (those that are not tagged and not used by any container)
    docker image prune -f
-    # Remove unused volumes
+    # Remove unused volumes / force the system prune for old images as well.
-    docker volume prune -f
+    docker volume prune -f && docker system prune --force --filter "until=72h" --all
    echo "Docker images and volumes cleanup completed."
  else
    echo "Disk usage is below $threshold%. No cleanup needed."
@@ -57,17 +59,17 @@ done
 echo "--- Pulling container" 
 image_name="rocm/vllm-ci:${BUILDKITE_COMMIT}"
 container_name="rocm_${BUILDKITE_COMMIT}_$(tr -dc A-Za-z0-9 < /dev/urandom | head -c 10; echo)"
-docker pull ${image_name}
+docker pull "${image_name}"
 remove_docker_container() {
-   docker rm -f ${container_name} || docker image rm -f ${image_name} || true
+   docker rm -f "${container_name}" || docker image rm -f "${image_name}" || true
 }
 trap remove_docker_container EXIT
 echo "--- Running container"
 HF_CACHE="$(realpath ~)/huggingface"
-mkdir -p ${HF_CACHE}
+mkdir -p "${HF_CACHE}"
 HF_MOUNT="/root/.cache/huggingface"
 commands=$@
@@ -83,7 +85,6 @@ if [[ $commands == *" kernels "* ]]; then
  --ignore=kernels/test_encoder_decoder_attn.py \
  --ignore=kernels/test_flash_attn.py \
  --ignore=kernels/test_flashinfer.py \
  --ignore=kernels/test_gguf.py \
  --ignore=kernels/test_int8_quant.py \
  --ignore=kernels/test_machete_gemm.py \
  --ignore=kernels/test_mamba_ssm.py \
@@ -107,35 +108,36 @@ fi
 PARALLEL_JOB_COUNT=8
 # check if the command contains shard flag, we will run all shards in parallel because the host have 8 GPUs. 
 if [[ $commands == *"--shard-id="* ]]; then
  # assign job count as the number of shards used   
  commands=${commands//"--num-shards= "/"--num-shards=${PARALLEL_JOB_COUNT} "}
  for GPU in $(seq 0 $(($PARALLEL_JOB_COUNT-1))); do
-    #replace shard arguments
+    # assign shard-id for each shard
-    commands=${commands//"--shard-id= "/"--shard-id=${GPU} "}
+    commands_gpu=${commands//"--shard-id= "/"--shard-id=${GPU} "}
-    commands=${commands//"--num-shards= "/"--num-shards=${PARALLEL_JOB_COUNT} "}
+    echo "Shard ${GPU} commands:$commands_gpu"
    echo "Shard ${GPU} commands:$commands"
    docker run \
        --device /dev/kfd --device /dev/dri \
        --network host \
        --shm-size=16gb \
        --rm \
-        -e HIP_VISIBLE_DEVICES=${GPU} \
+        -e HIP_VISIBLE_DEVICES="${GPU}" \
        -e HF_TOKEN \
-        -v ${HF_CACHE}:${HF_MOUNT} \
+        -v "${HF_CACHE}:${HF_MOUNT}" \
-        -e HF_HOME=${HF_MOUNT} \
+        -e "HF_HOME=${HF_MOUNT}" \
-        --name ${container_name}_${GPU}  \
+        --name "${container_name}_${GPU}" \
-        ${image_name} \
+        "${image_name}" \
-        /bin/bash -c "${commands}" \
+        /bin/bash -c "${commands_gpu}" \
        |& while read -r line; do echo ">>Shard $GPU: $line"; done &
    PIDS+=($!)
  done
  #wait for all processes to finish and collect exit codes
-  for pid in ${PIDS[@]}; do
+  for pid in "${PIDS[@]}"; do
-    wait ${pid}
+    wait "${pid}"
    STATUS+=($?)
  done
-  for st in ${STATUS[@]}; do
+  for st in "${STATUS[@]}"; do
    if [[ ${st} -ne 0 ]]; then
      echo "One of the processes failed with $st"
-      exit ${st}
+      exit "${st}"
    fi
  done
 else
@@ -146,9 +148,9 @@ else
          --rm \
          -e HIP_VISIBLE_DEVICES=0 \
          -e HF_TOKEN \
-          -v ${HF_CACHE}:${HF_MOUNT} \
+          -v "${HF_CACHE}:${HF_MOUNT}" \
-          -e HF_HOME=${HF_MOUNT} \
+          -e "HF_HOME=${HF_MOUNT}" \
-          --name ${container_name} \
+          --name "${container_name}" \
-          ${image_name} \
+          "${image_name}" \
          /bin/bash -c "${commands}"
 fi
--- a/.buildkite/run-benchmarks.sh
+++ b/.buildkite/run-benchmarks.sh
@@ -1,3 +1,5 @@
 #!/bin/bash
 # This script is run by buildkite to run the benchmarks and upload the results to buildkite
 set -ex
--- a/.buildkite/run-cpu-test-ppc64le.sh
+++ b/.buildkite/run-cpu-test-ppc64le.sh
@@ -1,39 +1,14 @@
 #!/bin/bash
 # This script build the CPU docker image and run the offline inference inside the container.
 # It serves a sanity check for compilation and basic model usage.
 set -ex
 # Try building the docker image
 docker build -t cpu-test -f Dockerfile.ppc64le .
 # Setup cleanup
-remove_docker_container() { docker rm -f cpu-test || true; }
+remove_docker_container() { docker rm -f cpu-test || true; docker system prune -f; }
 trap remove_docker_container EXIT
 remove_docker_container
-# Run the image, setting --shm-size=4g for tensor parallel.
+# Try building the docker image
-source /etc/environment
+docker build -t cpu-test -f Dockerfile.ppc64le .
 #docker run -itd --entrypoint /bin/bash -v ~/.cache/huggingface:/root/.cache/huggingface --privileged=true --network host -e HF_TOKEN --env VLLM_CPU_KVCACHE_SPACE=4 --shm-size=4g --name cpu-test cpu-test
 docker run -itd --entrypoint /bin/bash -v ~/.cache/huggingface:/root/.cache/huggingface --privileged=true --network host -e HF_TOKEN=$HF_TOKEN --name cpu-test cpu-test
 # Run basic model test
 docker exec cpu-test bash -c "
  pip install pytest matplotlib einops transformers_stream_generator
  pytest -v -s tests/models -m \"not vlm\" \
    --ignore=tests/models/test_embedding.py \
    --ignore=tests/models/test_oot_registration.py \
    --ignore=tests/models/test_registry.py \
    --ignore=tests/models/test_jamba.py \
    --ignore=tests/models/test_mamba.py \
    --ignore=tests/models/test_danube3_4b.py" # Mamba kernels and Danube3-4B on CPU is not supported
 # online inference
 docker exec cpu-test bash -c "
  python3 -m vllm.entrypoints.openai.api_server --model facebook/opt-125m & 
  timeout 600 bash -c 'until curl localhost:8000/v1/models; do sleep 1; done' || exit 1
  python3 benchmarks/benchmark_serving.py \
    --backend vllm \
    --dataset-name random \
    --model facebook/opt-125m \
    --num-prompts 20 \
    --endpoint /v1/completions \
    --tokenizer facebook/opt-125m"
--- a/.buildkite/run-cpu-test.sh
+++ b/.buildkite/run-cpu-test.sh
@@ -1,57 +1,88 @@
 #!/bin/bash
 # This script build the CPU docker image and run the offline inference inside the container.
 # It serves a sanity check for compilation and basic model usage.
 set -ex
 # allow to bind to different cores
 CORE_RANGE=${CORE_RANGE:-48-95}
 NUMA_NODE=${NUMA_NODE:-1}
 # Try building the docker image
-numactl -C 48-95 -N 1 docker build -t cpu-test -f Dockerfile.cpu .
+numactl -C "$CORE_RANGE" -N "$NUMA_NODE" docker build -t cpu-test-"$BUILDKITE_BUILD_NUMBER" -f Dockerfile.cpu .
-numactl -C 48-95 -N 1 docker build --build-arg VLLM_CPU_DISABLE_AVX512="true" -t cpu-test-avx2 -f Dockerfile.cpu .
+numactl -C "$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 cpu-test-avx2 || 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=48-95 \
+docker run -itd --entrypoint /bin/bash -v ~/.cache/huggingface:/root/.cache/huggingface --cpuset-cpus="$CORE_RANGE"  \
- --cpuset-mems=1 --privileged=true --network host -e HF_TOKEN --env VLLM_CPU_KVCACHE_SPACE=4 --shm-size=4g --name cpu-test cpu-test
+ --cpuset-mems="$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-"$BUILDKITE_BUILD_NUMBER"
-docker run -itd --entrypoint /bin/bash -v ~/.cache/huggingface:/root/.cache/huggingface --cpuset-cpus=48-95 \
+docker run -itd --entrypoint /bin/bash -v ~/.cache/huggingface:/root/.cache/huggingface --cpuset-cpus="$CORE_RANGE" \
- --cpuset-mems=1 --privileged=true --network host -e HF_TOKEN --env VLLM_CPU_KVCACHE_SPACE=4 --shm-size=4g --name cpu-test-avx2 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
-# offline inference
+function cpu_tests() {
-docker exec cpu-test-avx2 bash -c "python3 examples/offline_inference.py"
+  set -e
  export NUMA_NODE=$2
-# Run basic model test
+  # offline inference
-docker exec cpu-test bash -c "
+  docker exec cpu-test-"$BUILDKITE_BUILD_NUMBER"-avx2-"$NUMA_NODE" bash -c "
-  pip install pytest matplotlib einops transformers_stream_generator datamodel_code_generator
+    set -e
-  pytest -v -s tests/models/encoder_decoder/language
+    python3 examples/offline_inference/basic.py"
  pytest -v -s tests/models/decoder_only/language \
    --ignore=tests/models/test_fp8.py \
    --ignore=tests/models/decoder_only/language/test_jamba.py \
    --ignore=tests/models/decoder_only/language/test_mamba.py \
    --ignore=tests/models/decoder_only/language/test_granitemoe.py \
    --ignore=tests/models/decoder_only/language/test_danube3_4b.py" # Mamba and Danube3-4B on CPU is not supported
-# Run compressed-tensor test
+  # Run basic model test
-docker exec cpu-test bash -c "
+  docker exec cpu-test-"$BUILDKITE_BUILD_NUMBER"-"$NUMA_NODE" bash -c "
-  pytest -s -v \
+    set -e
-  tests/quantization/test_compressed_tensors.py::test_compressed_tensors_w8a8_static_setup \
+    pip install -r vllm/requirements-test.txt
-  tests/quantization/test_compressed_tensors.py::test_compressed_tensors_w8a8_dynamic_per_token"
+    pytest -v -s tests/models/decoder_only/language -m cpu_model
    pytest -v -s tests/models/embedding/language -m cpu_model
    pytest -v -s tests/models/encoder_decoder/language -m cpu_model
    pytest -v -s tests/models/decoder_only/audio_language -m cpu_model
    pytest -v -s tests/models/decoder_only/vision_language -m cpu_model"
-# Run AWQ test
+  # Run compressed-tensor test
-docker exec cpu-test bash -c "
+  docker exec cpu-test-"$BUILDKITE_BUILD_NUMBER"-"$NUMA_NODE" bash -c "
-  pytest -s -v \
+    set -e
-  tests/quantization/test_ipex_quant.py"
+    pytest -s -v \
    tests/quantization/test_compressed_tensors.py::test_compressed_tensors_w8a8_static_setup \
    tests/quantization/test_compressed_tensors.py::test_compressed_tensors_w8a8_dynamic_per_token"
-# online inference
+  # Run AWQ test
-docker exec cpu-test bash -c "
+  docker exec cpu-test-"$BUILDKITE_BUILD_NUMBER"-"$NUMA_NODE" bash -c "
-  export VLLM_CPU_KVCACHE_SPACE=10 
+    set -e
-  export VLLM_CPU_OMP_THREADS_BIND=48-92 
+    pytest -s -v \
-  python3 -m vllm.entrypoints.openai.api_server --model facebook/opt-125m & 
+    tests/quantization/test_ipex_quant.py"
-  timeout 600 bash -c 'until curl localhost:8000/v1/models; do sleep 1; done' || exit 1
+
-  python3 benchmarks/benchmark_serving.py \
+  # Run chunked-prefill and prefix-cache test
-    --backend vllm \
+  docker exec cpu-test-"$BUILDKITE_BUILD_NUMBER"-"$NUMA_NODE" bash -c "
-    --dataset-name random \
+    set -e
-    --model facebook/opt-125m \
+    pytest -s -v -k cpu_model \
-    --num-prompts 20 \
+    tests/basic_correctness/test_chunked_prefill.py"  
-    --endpoint /v1/completions \
+
-    --tokenizer facebook/opt-125m"
+  # online serving
  docker exec cpu-test-"$BUILDKITE_BUILD_NUMBER"-"$NUMA_NODE" bash -c "
    set -e
    export VLLM_CPU_KVCACHE_SPACE=10 
    export VLLM_CPU_OMP_THREADS_BIND=$1
    python3 -m vllm.entrypoints.openai.api_server --model facebook/opt-125m --dtype half & 
    timeout 600 bash -c 'until curl localhost:8000/v1/models; do sleep 1; done' || exit 1
    python3 benchmarks/benchmark_serving.py \
      --backend vllm \
      --dataset-name random \
      --model facebook/opt-125m \
      --num-prompts 20 \
      --endpoint /v1/completions \
      --tokenizer facebook/opt-125m"
  # Run multi-lora tests
  docker exec cpu-test-"$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 40 mins.
 export -f cpu_tests
 timeout 40m bash -c "cpu_tests $CORE_RANGE $NUMA_NODE"
--- a/.buildkite/run-gh200-test.sh
+++ b/.buildkite/run-gh200-test.sh
@@ -0,0 +1,28 @@
 #!/bin/bash
 # This script build the GH200 docker image and run the offline inference inside the container.
 # It serves a sanity check for compilation and basic model usage.
 set -ex
 # Skip the new torch installation during build since we are using the specified version for arm64 in the Dockerfile
 python3 use_existing_torch.py
 # Try building the docker image
 DOCKER_BUILDKIT=1 docker build . \
  --target vllm-openai \
  --platform "linux/arm64" \
  -t gh200-test \
  --build-arg max_jobs=66 \
  --build-arg nvcc_threads=2 \
  --build-arg torch_cuda_arch_list="9.0+PTX" \
  --build-arg vllm_fa_cmake_gpu_arches="90-real"
 # Setup cleanup
 remove_docker_container() { docker rm -f gh200-test || true; }
 trap remove_docker_container EXIT
 remove_docker_container
 # Run the image and test offline inference
 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/cli.py --model meta-llama/Llama-3.2-1B
 '
--- a/.buildkite/run-hpu-test.sh
+++ b/.buildkite/run-hpu-test.sh
@@ -0,0 +1,24 @@
 #!/bin/bash
 # This script build the CPU docker image and run the offline inference inside the container.
 # It serves a sanity check for compilation and basic model usage.
 set -ex
 # Try building the docker image
 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; }
 remove_docker_container_and_exit() { remove_docker_container; exit $EXITCODE; }
 trap remove_docker_container_and_exit EXIT
 remove_docker_container
 # Run the image and launch offline inference
 docker run --runtime=habana --name=hpu-test --network=host -e HABANA_VISIBLE_DEVICES=all -e VLLM_SKIP_WARMUP=true --entrypoint="" hpu-test-env python3 examples/offline_inference/basic.py
 EXITCODE=$?
--- a/.buildkite/run-multi-node-test.sh
+++ b/.buildkite/run-multi-node-test.sh
@@ -14,7 +14,7 @@ DOCKER_IMAGE=$4
 shift 4
 COMMANDS=("$@")
-if [ ${#COMMANDS[@]} -ne $NUM_NODES ]; then
+if [ ${#COMMANDS[@]} -ne "$NUM_NODES" ]; then
    echo "The number of commands must be equal to the number of nodes."
    echo "Number of nodes: $NUM_NODES"
    echo "Number of commands: ${#COMMANDS[@]}"
@@ -23,7 +23,7 @@ fi
 echo "List of commands"
 for command in "${COMMANDS[@]}"; do
-    echo $command
+    echo "$command"
 done
 start_network() {
@@ -36,7 +36,7 @@ start_nodes() {
        for node_gpu in $(seq 0 $(($NUM_GPUS - 1))); do
            DEVICE_NUM=$(($node * $NUM_GPUS + $node_gpu))
            GPU_DEVICES+=$(($DEVICE_NUM))
-            if [ $node_gpu -lt $(($NUM_GPUS - 1)) ]; then
+            if [ "$node_gpu" -lt $(($NUM_GPUS - 1)) ]; then
                GPU_DEVICES+=','
            fi
        done
@@ -49,17 +49,20 @@ start_nodes() {
        # 3. map the huggingface cache directory to the container
        # 3. assign ip addresses to the containers (head node: 192.168.10.10, worker nodes:
        #    starting from 192.168.10.11)
-        docker run -d --gpus "$GPU_DEVICES" --shm-size=10.24gb -e HF_TOKEN -v ~/.cache/huggingface:/root/.cache/huggingface --name node$node --network docker-net --ip 192.168.10.$((10 + $node)) --rm $DOCKER_IMAGE /bin/bash -c "tail -f /dev/null"
+        docker run -d --gpus "$GPU_DEVICES" --shm-size=10.24gb -e HF_TOKEN \
            -v ~/.cache/huggingface:/root/.cache/huggingface --name "node$node" \
            --network docker-net --ip 192.168.10.$((10 + $node)) --rm "$DOCKER_IMAGE" \
            /bin/bash -c "tail -f /dev/null"
        # organize containers into a ray cluster
-        if [ $node -eq 0 ]; then
+        if [ "$node" -eq 0 ]; then
            # start the ray head node
-            docker exec -d node$node /bin/bash -c "ray start --head --port=6379 --block"
+            docker exec -d "node$node" /bin/bash -c "ray start --head --port=6379 --block"
            # wait for the head node to be ready
            sleep 10
        else
            # start the ray worker nodes, and connect them to the head node
-            docker exec -d node$node /bin/bash -c "ray start --address=192.168.10.10:6379 --block"
+            docker exec -d "node$node" /bin/bash -c "ray start --address=192.168.10.10:6379 --block"
        fi
    done
@@ -79,22 +82,22 @@ run_nodes() {
        for node_gpu in $(seq 0 $(($NUM_GPUS - 1))); do
            DEVICE_NUM=$(($node * $NUM_GPUS + $node_gpu))
            GPU_DEVICES+=$(($DEVICE_NUM))
-            if [ $node_gpu -lt $(($NUM_GPUS - 1)) ]; then
+            if [ "$node_gpu" -lt $(($NUM_GPUS - 1)) ]; then
                GPU_DEVICES+=','
            fi
        done
        GPU_DEVICES+='"'
        echo "Running node$node with GPU devices: $GPU_DEVICES"
-        if [ $node -ne 0 ]; then
+        if [ "$node" -ne 0 ]; then
-            docker exec -d node$node /bin/bash -c "cd $WORKING_DIR ; ${COMMANDS[$node]}"
+            docker exec -d "node$node" /bin/bash -c "cd $WORKING_DIR ; ${COMMANDS[$node]}"
        else
-            docker exec node$node /bin/bash -c "cd $WORKING_DIR ; ${COMMANDS[$node]}"
+            docker exec "node$node" /bin/bash -c "cd $WORKING_DIR ; ${COMMANDS[$node]}"
        fi
    done
 }
 cleanup() {
    for node in $(seq 0 $(($NUM_NODES-1))); do
-        docker stop node$node
+        docker stop "node$node"
    done
    docker network rm docker-net
 }
--- a/.buildkite/run-neuron-test.sh
+++ b/.buildkite/run-neuron-test.sh
@@ -1,6 +1,20 @@
 #!/bin/bash
 # 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 "${NEURON_COMPILE_CACHE_URL}"
 NEURON_COMPILE_CACHE_MOUNT="/root/.cache/neuron_compile_cache"
 # Try building the docker image
 aws ecr get-login-password --region us-west-2 | docker login --username AWS --password-stdin 763104351884.dkr.ecr.us-west-2.amazonaws.com
@@ -11,41 +25,33 @@ if [ -f /tmp/neuron-docker-build-timestamp ]; then
    last_build=$(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)
-        echo $current_time > /tmp/neuron-docker-build-timestamp
+        docker image prune -f
        # Remove unused volumes / force the system prune for old images as well.
        docker volume prune -f && docker system prune -f
        # Remove huggingface model artifacts and compiler cache
        rm -rf "${HF_MOUNT:?}/*"
        rm -rf "${NEURON_COMPILE_CACHE_MOUNT:?}/*"
        echo "$current_time" > /tmp/neuron-docker-build-timestamp
    fi
 else
-    echo $(date +%s) > /tmp/neuron-docker-build-timestamp
+    date "+%s" > /tmp/neuron-docker-build-timestamp
 fi
-docker build -t neuron -f Dockerfile.neuron .
+docker build -t "${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 $timeout seconds"
            break
        fi
    done
 }
 wait_for_server_to_start
 # Test a simple prompt
 curl -X POST -H "Content-Type: application/json" \
    localhost:8000/generate \
    -d '{"prompt": "San Francisco is a"}'
--- a/.buildkite/run-openvino-test.sh
+++ b/.buildkite/run-openvino-test.sh
@@ -1,3 +1,5 @@
 #!/bin/bash
 # This script build the OpenVINO docker image and run the offline inference inside the container.
 # It serves a sanity check for compilation and basic model usage.
 set -ex
@@ -11,4 +13,4 @@ trap remove_docker_container EXIT
 remove_docker_container
 # Run the image and launch offline inference
-docker run --network host --env VLLM_OPENVINO_KVCACHE_SPACE=1 --name openvino-test openvino-test python3 /workspace/vllm/examples/offline_inference.py
+docker run --network host --env VLLM_OPENVINO_KVCACHE_SPACE=1 --name openvino-test openvino-test python3 /workspace/examples/offline_inference/basic.py
--- a/.buildkite/run-tpu-test.sh
+++ b/.buildkite/run-tpu-test.sh
@@ -1,3 +1,5 @@
 #!/bin/bash
 set -e
 # Build the docker image.
@@ -12,4 +14,13 @@ remove_docker_container
 # For HF_TOKEN.
 source /etc/environment
 # Run a simple end-to-end example.
-docker run --privileged --net host --shm-size=16G -it -e HF_TOKEN=$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  && 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"
--- a/.buildkite/run-xpu-test.sh
+++ b/.buildkite/run-xpu-test.sh
@@ -1,3 +1,5 @@
 #!/bin/bash
 # This script build the CPU docker image and run the offline inference inside the container.
 # It serves a sanity check for compilation and basic model usage.
 set -ex
@@ -10,5 +12,8 @@ remove_docker_container() { docker rm -f xpu-test || true; }
 trap remove_docker_container EXIT
 remove_docker_container
-# Run the image and launch offline inference
+# Run the image and test offline inference/tensor parallel
-docker run --network host --name xpu-test --device /dev/dri -v /dev/dri/by-path:/dev/dri/by-path --entrypoint="" xpu-test python3 examples/offline_inference.py
+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/basic.py
    python3 examples/offline_inference/cli.py -tp 2
 '
--- a/.buildkite/test-pipeline.yaml
+++ b/.buildkite/test-pipeline.yaml
@@ -9,7 +9,7 @@
 # label(str): the name of the test. emoji allowed.
 # fast_check(bool): whether to run this on each commit on fastcheck pipeline.
 # fast_check_only(bool): run this test on fastcheck pipeline only
-# optional(bool): never run this test by default (i.e. need to unblock manually)
+# optional(bool): never run this test by default (i.e. need to unblock manually) unless it's scheduled nightly run.
 # command(str): the single command to run for tests. incompatible with commands.
 # commands(list): the list of commands to run for test. incompatbile with command.
 # mirror_hardwares(list): the list of hardwares to run the test on as well. currently only supports [amd]
@@ -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
@@ -50,7 +50,9 @@ steps:
  - tests/multimodal
  - tests/test_utils
  - tests/worker
  - tests/standalone_tests/lazy_imports.py
  commands:
  - 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
@@ -59,6 +61,13 @@ steps:
  - pytest -v -s test_utils.py # Utils
  - pytest -v -s worker # Worker
 - label: Python-only Installation Test
  source_file_dependencies:
  - tests/standalone_tests/python_only_compile.sh
  - setup.py
  commands:
  - bash standalone_tests/python_only_compile.sh
 - label: Basic Correctness Test # 30min
  #mirror_hardwares: [amd]
  fast_check: true
@@ -67,7 +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
@@ -97,14 +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
@@ -118,10 +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 
@@ -163,27 +179,47 @@ steps:
  # OOM in the CI unless we run this separately
  - pytest -v -s tokenization
- label: Examples Test # 15min
+- label: V1 Test
  #mirror_hardwares: [amd]
  source_file_dependencies:
    - vllm/
    - tests/v1
  commands:
    # 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"
  #mirror_hardwares: [amd]
  source_file_dependencies:
  - vllm/entrypoints
  - examples/
  commands:
-    - pip install awscli tensorizer # for llava example and tensorizer test
+    - 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_profile.py --model facebook/opt-125m
+    - python3 offline_inference/classification.py
    - python3 offline_inference/embedding.py
    - python3 offline_inference/scoring.py
    - python3 offline_inference/profiling.py --model facebook/opt-125m run_num_steps --num-steps 2
 - label: Prefix Caching Test # 9min
-  #mirror_hardwares: [amd]
+  mirror_hardwares: [amd]
  source_file_dependencies:
  - vllm/
  - tests/prefix_caching
@@ -195,6 +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
@@ -203,23 +240,29 @@ steps:
  mirror_hardwares: [amd]
  source_file_dependencies:
  - vllm/model_executor/layers
  - vllm/model_executor/guided_decoding
  - tests/test_logits_processor
-  command: pytest -v -s test_logits_processor.py
+  - tests/model_executor/test_guided_processors
  commands: 
    - pytest -v -s test_logits_processor.py
    - pytest -v -s model_executor/test_guided_processors.py
- label: Speculative decoding tests # 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
+  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
@@ -229,15 +272,16 @@ steps:
  - tests/compile
  commands:
  - pytest -v -s compile/test_basic_correctness.py
  # these tests need to be separated, cannot combine
  - pytest -v -s compile/piecewise/test_simple.py
  - pytest -v -s compile/piecewise/test_toy_llama.py
-# TODO: re-write in comparison tests, and fix symbolic shape
+- label: "PyTorch Fullgraph Test" # 18min
-# for quantization ops.
+  source_file_dependencies:
-# - label: "PyTorch Fullgraph Test" # 18min
+  - vllm/
-#   source_file_dependencies:
+  - tests/compile
-#   - vllm/
+  commands:
-#   - tests/compile
+  - pytest -v -s compile/test_full_graph.py
 #   commands:
 #   - pytest -v -s compile/test_full_graph.py
 - label: Kernels Test %N # 1h each
  mirror_hardwares: [amd]
@@ -266,7 +310,6 @@ steps:
  source_file_dependencies:
  - benchmarks/
  commands:
  - pip install aiohttp
  - bash run-benchmarks.sh
 - label: Quantization Test # 33min
@@ -303,46 +346,84 @@ steps:
 #####  models test  #####
- label: Basic Models Test # 3min
+- label: Basic Models Test # 24min
  source_file_dependencies:
  - 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/*.py --ignore=models/test_oot_registration.py
+    - pytest -v -s models/test_initialization.py
- label: Decoder-only Language Models Test # 1h36min
+- label: Language Models Test (Standard) # 32min
  #mirror_hardwares: [amd]
  source_file_dependencies:
  - vllm/
  - tests/models/decoder_only/language
  - tests/models/embedding/language
  - tests/models/encoder_decoder/language
  commands:
-    - pytest -v -s models/decoder_only/language
+    - pytest -v -s models/decoder_only/language -m 'core_model or quant_model'
    - pytest -v -s models/embedding/language -m core_model
- label: Decoder-only Multi-Modal Models Test # 1h31min
+- label: Language Models Test (Extended) # 1h10min
  optional: true
  source_file_dependencies:
  - vllm/
  - tests/models/decoder_only/language
  - tests/models/embedding/language
  - tests/models/encoder_decoder/language
  commands:
    - 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) # 40min
  #mirror_hardwares: [amd]
  source_file_dependencies:
  - vllm/
  - tests/models/decoder_only/audio_language
  - tests/models/decoder_only/vision_language
  commands:
    - pytest -v -s models/decoder_only/audio_language
    - pytest -v -s models/decoder_only/vision_language
 - label: Other Models Test # 6min
  #mirror_hardwares: [amd]
  source_file_dependencies:
  - vllm/
  - tests/models/embedding/language
  - tests/models/embedding/vision_language
-  - tests/models/encoder_decoder/language
+  - tests/models/encoder_decoder/audio_language
  - tests/models/encoder_decoder/vision_language
  commands:
-    - pytest -v -s models/embedding/language
+    - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
-    - pytest -v -s models/embedding/vision_language
+    - pytest -v -s models/multimodal
-    - pytest -v -s models/encoder_decoder/language
+    - pytest -v -s models/decoder_only/audio_language -m 'core_model or quant_model'
-    - pytest -v -s models/encoder_decoder/vision_language
+    - 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 # 48m
  optional: true
  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/vision_language
  commands:
    - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
    - pytest -v -s models/decoder_only/audio_language -m 'not core_model and not quant_model'
    - pytest -v -s models/decoder_only/vision_language/test_models.py -m 'split(group=0) and not core_model and not quant_model'
    # HACK - run phi3v tests separately to sidestep this transformers bug
    # https://github.com/huggingface/transformers/issues/34307
    - pytest -v -s models/decoder_only/vision_language/test_phi3v.py
    - pytest -v -s --ignore models/decoder_only/vision_language/test_models.py --ignore models/decoder_only/vision_language/test_phi3v.py models/decoder_only/vision_language -m 'not core_model and not quant_model'
    - pytest -v -s models/embedding/vision_language -m 'not core_model'
    - pytest -v -s models/encoder_decoder/language -m 'not core_model'
    - pytest -v -s models/encoder_decoder/vision_language -m 'not core_model'
 - label: Multi-Modal Models Test (Extended) 2 # 38m
  optional: true
  source_file_dependencies:
  - vllm/
  - tests/models/decoder_only/vision_language
  commands:
    - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
    - pytest -v -s models/decoder_only/vision_language/test_models.py -m 'split(group=1) and not core_model and not quant_model'
 # This test is used only in PR development phase to test individual models and should never run on main
 - label: Custom Models Test
@@ -378,11 +459,11 @@ steps:
  - tests/distributed/
  commands:
  - # the following commands are for the first node, with ip 192.168.10.10 (ray environment already set up)
-    - VLLM_TEST_SAME_HOST=0 torchrun --nnodes 2 --nproc-per-node=2 --rdzv_backend=c10d --rdzv_endpoint=192.168.10.10 distributed/test_same_node.py | grep -q 'Same node test passed'
+    - VLLM_TEST_SAME_HOST=0 torchrun --nnodes 2 --nproc-per-node=2 --rdzv_backend=c10d --rdzv_endpoint=192.168.10.10 distributed/test_same_node.py | grep 'Same node test passed'
    - VLLM_MULTI_NODE=1 pytest -v -s distributed/test_multi_node_assignment.py
    - VLLM_MULTI_NODE=1 pytest -v -s distributed/test_pipeline_parallel.py
  - # the following commands are for the second node, with ip 192.168.10.11 (ray environment already set up)
-    - VLLM_TEST_SAME_HOST=0 torchrun --nnodes 2 --nproc-per-node=2 --rdzv_backend=c10d --rdzv_endpoint=192.168.10.10 distributed/test_same_node.py | grep -q 'Same node test passed'
+    - VLLM_TEST_SAME_HOST=0 torchrun --nnodes 2 --nproc-per-node=2 --rdzv_backend=c10d --rdzv_endpoint=192.168.10.10 distributed/test_same_node.py | grep 'Same node test passed'
 - label: Distributed Tests (2 GPUs) # 40min
  #mirror_hardwares: [amd]
@@ -395,20 +476,46 @@ steps:
  - vllm/model_executor/models/
  - tests/distributed/
  - vllm/compilation
  - 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 -q 'Same node test passed'
+  - 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_2_gpus
+  - 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/encoder_decoder/language/test_bart.py -v -s -m distributed_2_gpus
+  - pytest models/test_transformers.py -v -s -m 'distributed(num_gpus=2)'
-  - pytest models/encoder_decoder/vision_language/test_broadcast.py -v -s -m distributed_2_gpus
+  - pytest models/encoder_decoder/language/test_bart.py -v -s -m 'distributed(num_gpus=2)'
-  - pytest models/decoder_only/vision_language/test_broadcast.py -v -s -m distributed_2_gpus
+  - pytest models/encoder_decoder/vision_language/test_broadcast.py -v -s -m 'distributed(num_gpus=2)'
-  - pytest -v -s spec_decode/e2e/test_integration_dist_tp2.py
+  - pytest models/decoder_only/vision_language/test_models.py -v -s -m 'distributed(num_gpus=2)'
  # this test fails consistently.
  # TODO: investigate and fix
  # - 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
-  - CUDA_VISIBLE_DEVICES=0,1 pytest -v -s test_sharded_state_loader.py
+  - pytest -v -s entrypoints/openai/test_oot_registration.py # it needs a clean process
-  - CUDA_VISIBLE_DEVICES=0,1 pytest -v -s distributed/test_utils.py
+  - 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"
@@ -425,7 +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
@@ -441,18 +550,23 @@ steps:
  - pytest -v -s distributed/test_pp_cudagraph.py
  - pytest -v -s distributed/test_pipeline_parallel.py
- label: LoRA Long Context (Distributed) # 11min
+- label: LoRA TP Test (Distributed)
  # This test runs llama 13B, so it is required to run on 4 GPUs.
  num_gpus: 4
  soft_fail: true
  source_file_dependencies:
  - vllm/lora
-  - tests/lora/test_long_context
+  - tests/lora
  commands:
    # FIXIT: find out which code initialize cuda before running the test
    # before the fix, we need to use spawn to test it
    - export VLLM_WORKER_MULTIPROC_METHOD=spawn
    # This test runs llama 13B, so it is required to run on 4 GPUs.
    - pytest -v -s -x lora/test_long_context.py
    # There is some Tensor Parallelism related processing logic in LoRA that 
    # 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
  working_dir: "/vllm-workspace/tests"
@@ -480,6 +594,7 @@ steps:
 - label: Distributed Tests (A100) # optional
  gpu: a100
  optional: true
  num_gpus: 4
  source_file_dependencies:
  - vllm/
@@ -487,11 +602,13 @@ steps:
  # NOTE: don't test llama model here, it seems hf implementation is buggy
  # see https://github.com/vllm-project/vllm/pull/5689 for details
  - pytest -v -s distributed/test_custom_all_reduce.py
-  - TARGET_TEST_SUITE=A100 pytest basic_correctness/ -v -s -m distributed_2_gpus
+  - torchrun --nproc_per_node=2 distributed/test_ca_buffer_sharing.py
  - TARGET_TEST_SUITE=A100 pytest basic_correctness/ -v -s -m 'distributed(num_gpus=2)'
  - pytest -v -s -x lora/test_mixtral.py
 - label: LM Eval Large Models # optional
  gpu: a100
  optional: true
  num_gpus: 4
  working_dir: "/vllm-workspace/.buildkite/lm-eval-harness"
  source_file_dependencies:
--- a/.buildkite/upload-wheels.sh
+++ b/.buildkite/upload-wheels.sh
@@ -0,0 +1,71 @@
 #!/usr/bin/env bash
 set -ex
 # Assume wheels are in artifacts/dist/*.whl
 wheel_files=(artifacts/dist/*.whl)
 # Check that exactly one wheel is found
 if [[ ${#wheel_files[@]} -ne 1 ]]; then
  echo "Error: Expected exactly one wheel file in artifacts/dist/, but found ${#wheel_files[@]}"
  exit 1
 fi
 # Get the single wheel file
 wheel="${wheel_files[0]}"
 # Rename 'linux' to 'manylinux1' in the wheel filename
 new_wheel="${wheel/linux/manylinux1}"
 mv -- "$wheel" "$new_wheel"
 wheel="$new_wheel"
 # Extract the version from the wheel
 version=$(unzip -p "$wheel" '**/METADATA' | grep '^Version: ' | cut -d' ' -f2)
 echo "Version: $version"
 normal_wheel="$wheel" # Save the original wheel filename
 # If the version contains "dev", rename it to v1.0.0.dev for consistency
 if [[ $version == *dev* ]]; then
    suffix="${version##*.}"
    if [[ $suffix == cu* ]]; then
        new_version="1.0.0.dev+${suffix}"
    else
        new_version="1.0.0.dev"
    fi
    new_wheel="${wheel/$version/$new_version}"
    # use cp to keep both files in the artifacts directory
    cp -- "$wheel" "$new_wheel"
    wheel="$new_wheel"
    version="$new_version"
 fi
 # Upload the wheel to S3
 python3 .buildkite/generate_index.py --wheel "$normal_wheel"
 # generate index for this commit
 aws s3 cp "$wheel" "s3://vllm-wheels/$BUILDKITE_COMMIT/"
 aws s3 cp "$normal_wheel" "s3://vllm-wheels/$BUILDKITE_COMMIT/"
 if [[ $normal_wheel == *"cu118"* ]]; then
    # if $normal_wheel matches cu118, do not upload the index.html
    echo "Skipping index files for cu118 wheels"
 else
    # only upload index.html for cu12 wheels (default wheels)
    aws s3 cp index.html "s3://vllm-wheels/$BUILDKITE_COMMIT/vllm/index.html"
    aws s3 cp "s3://vllm-wheels/nightly/index.html" "s3://vllm-wheels/$BUILDKITE_COMMIT/index.html"
 fi
 # generate index for nightly
 aws s3 cp "$wheel" "s3://vllm-wheels/nightly/"
 aws s3 cp "$normal_wheel" "s3://vllm-wheels/nightly/"
 if [[ $normal_wheel == *"cu118"* ]]; then
    # if $normal_wheel matches cu118, do not upload the index.html
    echo "Skipping index files for cu118 wheels"
 else
    # only upload index.html for cu12 wheels (default wheels)
    aws s3 cp index.html "s3://vllm-wheels/nightly/vllm/index.html"
 fi
 aws s3 cp "$wheel" "s3://vllm-wheels/$version/"
--- a/.github/CODEOWNERS
+++ b/.github/CODEOWNERS
@@ -2,29 +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 @WoosukKwon @zhuohan123 @youkaichao @alexm-neuralmagic @comaniac @njhill
+/vllm/core @zhuohan123 @youkaichao @alexm-redhat @comaniac @njhill
-/vllm/engine/llm_engine.py @WoosukKwon @zhuohan123 @youkaichao @alexm-neuralmagic @comaniac @njhill
+/vllm/engine/llm_engine.py @zhuohan123 @youkaichao @alexm-redhat @comaniac @njhill
-/vllm/executor/executor_base.py @WoosukKwon @zhuohan123 @youkaichao @alexm-neuralmagic @comaniac @njhill
+/vllm/executor/executor_base.py @zhuohan123 @youkaichao @alexm-redhat @comaniac @njhill
-/vllm/worker/worker_base.py @WoosukKwon @zhuohan123 @youkaichao @alexm-neuralmagic @comaniac @njhill
+/vllm/worker/worker_base.py @zhuohan123 @youkaichao @alexm-redhat @comaniac @njhill
-/vllm/worker/worker.py @WoosukKwon @zhuohan123 @youkaichao @alexm-neuralmagic @comaniac @njhill
+/vllm/worker/worker.py @zhuohan123 @youkaichao @alexm-redhat @comaniac @njhill
-/vllm/model_executor/layers/sampler.py @WoosukKwon @zhuohan123 @youkaichao @alexm-neuralmagic @comaniac @njhill
+/vllm/model_executor/layers/sampler.py @zhuohan123 @youkaichao @alexm-redhat @comaniac @njhill
-CMakeLists.txt @tlrmchlsmth @WoosukKwon
+/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-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
--- a/.github/FUNDING.yml
+++ b/.github/FUNDING.yml
@@ -1,2 +1,2 @@
 github: [vllm-project]
-open_collective: [vllm]
+open_collective: vllm
--- a/.github/ISSUE_TEMPLATE/400-bug-report.yml
+++ b/.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
--- a/.github/ISSUE_TEMPLATE/500-feature-request.yml
+++ b/.github/ISSUE_TEMPLATE/500-feature-request.yml
--- a/.github/ISSUE_TEMPLATE/600-new-model.yml
+++ b/.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.
--- a/.github/ISSUE_TEMPLATE/700-performance-discussion.yml
+++ b/.github/ISSUE_TEMPLATE/700-performance-discussion.yml
--- a/.github/ISSUE_TEMPLATE/800-misc-discussion.yml
+++ b/.github/ISSUE_TEMPLATE/800-misc-discussion.yml
--- a/.github/PULL_REQUEST_TEMPLATE.md
+++ b/.github/PULL_REQUEST_TEMPLATE.md
@@ -2,73 +2,4 @@ FILL IN THE PR DESCRIPTION HERE
 FIX #xxxx (*link existing issues this PR will resolve*)
-**BEFORE SUBMITTING, PLEASE READ THE CHECKLIST BELOW AND FILL IN THE DESCRIPTION ABOVE**
+**BEFORE SUBMITTING, PLEASE READ https://docs.vllm.ai/en/latest/contributing/overview.html **
 ---
 <details>
 <!-- inside this <details> section, markdown rendering does not work, so we use raw html here. -->
 <summary><b> PR Checklist (Click to Expand) </b></summary>
 <p>Thank you for your contribution to vLLM! Before submitting the pull request, please ensure the PR meets the following criteria. This helps vLLM maintain the code quality and improve the efficiency of the review process.</p>
 <h3>PR Title and Classification</h3>
 <p>Only specific types of PRs will be reviewed. The PR title is prefixed appropriately to indicate the type of change. Please use one of the following:</p>
 <ul>
    <li><code>[Bugfix]</code> for bug fixes.</li>
    <li><code>[CI/Build]</code> for build or continuous integration improvements.</li>
    <li><code>[Doc]</code> for documentation fixes and improvements.</li>
    <li><code>[Model]</code> for adding a new model or improving an existing model. Model name should appear in the title.</li>
    <li><code>[Frontend]</code> For changes on the vLLM frontend (e.g., OpenAI API server, <code>LLM</code> class, etc.) </li>
    <li><code>[Kernel]</code> for changes affecting CUDA kernels or other compute kernels.</li>
    <li><code>[Core]</code> for changes in the core vLLM logic (e.g., <code>LLMEngine</code>, <code>AsyncLLMEngine</code>, <code>Scheduler</code>, etc.)</li>
    <li><code>[Hardware][Vendor]</code> for hardware-specific changes. Vendor name should appear in the prefix (e.g., <code>[Hardware][AMD]</code>).</li>
    <li><code>[Misc]</code> for PRs that do not fit the above categories. Please use this sparingly.</li>
 </ul>
 <p><strong>Note:</strong> If the PR spans more than one category, please include all relevant prefixes.</p>
 <h3>Code Quality</h3>
 <p>The PR need to meet the following code quality standards:</p>
 <ul>
    <li>We adhere to <a href="https://google.github.io/styleguide/pyguide.html">Google Python style guide</a> and <a href="https://google.github.io/styleguide/cppguide.html">Google C++ style guide</a>.</li>
    <li>Pass all linter checks. Please use <a href="https://github.com/vllm-project/vllm/blob/main/format.sh"><code>format.sh</code></a> to format your code.</li>
    <li>The code need to be well-documented to ensure future contributors can easily understand the code.</li>
    <li>Include sufficient tests to ensure the project to stay correct and robust. This includes both unit tests and integration tests.</li>
    <li>Please add documentation to <code>docs/source/</code> if the PR modifies the user-facing behaviors of vLLM. It helps vLLM user understand and utilize the new features or changes.</li>
 </ul>
 <h3>Adding or changing kernels</h3>
 <p>Each custom kernel needs a schema and one or more implementations to be registered with PyTorch.</p>
 <ul>
    <li>Make sure custom ops are registered following PyTorch guidelines: <a href="https://pytorch.org/tutorials/advanced/cpp_custom_ops.html#cpp-custom-ops-tutorial">Custom C++ and CUDA Operators</a> and <a href="https://docs.google.com/document/d/1_W62p8WJOQQUzPsJYa7s701JXt0qf2OfLub2sbkHOaU">The Custom Operators Manual</a></li>
    <li>Custom operations that return <code>Tensors</code> require meta-functions. Meta-functions should be implemented and registered in python so that dynamic dims can be handled automatically. See above documents for a description of meta-functions.</li>
    <li>Use <a href="https://pytorch.org/docs/stable/library.html#torch.library.opcheck"><code>torch.libary.opcheck()</code></a> to test the function registration and meta-function for any registered ops.  See <code>tests/kernels</code> for examples.</li>
    <li>When changing the C++ signature of an existing op, the schema must be updated to reflect the changes.</li>
    <li>If a new custom type is needed, see the following document: <a href="https://docs.google.com/document/d/18fBMPuOJ0fY5ZQ6YyrHUppw9FA332CpNtgB6SOIgyuA">Custom Class Support in PT2</a>.
 </ul>
 <h3>Notes for Large Changes</h3>
 <p>Please keep the changes as concise as possible. For major architectural changes (>500 LOC excluding kernel/data/config/test), we would expect a GitHub issue (RFC) discussing the technical design and justification. Otherwise, we will tag it with <code>rfc-required</code> and might not go through the PR.</p>
 <h3>What to Expect for the Reviews</h3>
 <p>The goal of the vLLM team is to be a <i>transparent reviewing machine</i>. We would like to make the review process transparent and efficient and make sure no contributor feel confused or frustrated. However, the vLLM team is small, so we need to prioritize some PRs over others. Here is what you can expect from the review process: </p>
 <ul>
    <li> After the PR is submitted, the PR will be assigned to a reviewer. Every reviewer will pick up the PRs based on their expertise and availability.</li>
    <li> After the PR is assigned, the reviewer will provide status update every 2-3 days. If the PR is not reviewed within 7 days, please feel free to ping the reviewer or the vLLM team.</li>
    <li> After the review, the reviewer will put an <code> action-required</code> label on the PR if there are changes required. The contributor should address the comments and ping the reviewer to re-review the PR.</li>
    <li> Please respond to all comments within a reasonable time frame. If a comment isn't clear or you disagree with a suggestion, feel free to ask for clarification or discuss the suggestion.
 </li>
 </ul>
 <h3>Thank You</h3>
 <p> Finally, thank you for taking the time to read these guidelines and for your interest in contributing to vLLM. Your contributions make vLLM a great tool for everyone! </p>
 </details>
--- a/.github/dependabot.yml
+++ b/.github/dependabot.yml
@@ -5,3 +5,27 @@ updates:
    directory: "/"
    schedule:
      interval: "weekly"
  - package-ecosystem: "pip"
    directory: "/"
    schedule:
      interval: "weekly"
    labels: ["dependencies"]
    open-pull-requests-limit: 5
    reviewers: ["khluu", "simon-mo"]
    allow:
      - dependency-type: "all"
    ignore:
      - dependency-name: "*"
        update-types: ["version-update:semver-patch"]
      - dependency-name: "torch"
      - dependency-name: "torchvision"
      - dependency-name: "xformers"
      - dependency-name: "lm-format-enforcer"
      - dependency-name: "gguf"
      - dependency-name: "compressed-tensors"
      - dependency-name: "ray[adag]"
      - dependency-name: "lm-eval"
    groups:
      minor-update:
        applies-to: version-updates
        update-types: ["minor"]
--- a/.github/mergify.yml
+++ b/.github/mergify.yml
@@ -0,0 +1,97 @@
 pull_request_rules:
 - name: label-documentation
  description: Automatically apply documentation label
  conditions:
    - or:
      - files~=^[^/]+\.md$
      - files~=^docs/
  actions:
    label:
      add:
        - documentation
 - name: label-ci-build
  description: Automatically apply ci/build label
  conditions:
    - or:
      - files~=^\.github/
      - files~=\.buildkite/
      - files~=^cmake/
      - files=CMakeLists.txt
      - files~=^Dockerfile
      - files~=^requirements.*\.txt
      - files=setup.py
  actions:
    label:
      add:
        - ci/build
 - name: label-frontend
  description: Automatically apply frontend label
  conditions:
    - files~=^vllm/entrypoints/
  actions:
    label:
      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
      - -closed
  actions:
    label:
      add:
        - needs-rebase
    comment:
      message: |
       This pull request has merge conflicts that must be resolved before it can be
       merged. Please rebase the PR, @{{author}}.
       https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork
 - name: remove 'needs-rebase' label when conflict is resolved
  conditions:
      - -conflict
      - -closed
  actions:
    label:
      remove:
        - needs-rebase
--- a/.github/scripts/cleanup_pr_body.sh
+++ b/.github/scripts/cleanup_pr_body.sh
@@ -0,0 +1,50 @@
 #!/bin/bash
 set -eu
 # ensure 1 argument is passed
 if [ "$#" -ne 1 ]; then
    echo "Usage: $0 <pr_number>"
    exit 1
 fi
 PR_NUMBER=$1
 OLD=/tmp/orig_pr_body.txt
 NEW=/tmp/new_pr_body.txt
 gh pr view --json body --template "{{.body}}" "${PR_NUMBER}" > "${OLD}"
 cp "${OLD}" "${NEW}"
 # Remove "FIX #xxxx (*link existing issues this PR will resolve*)"
 sed -i '/FIX #xxxx.*$/d' "${NEW}"
 # Remove "FILL IN THE PR DESCRIPTION HERE"
 sed -i '/FILL IN THE PR DESCRIPTION HERE/d' "${NEW}"
 # Remove all lines after and including "**BEFORE SUBMITTING, PLEASE READ THE CHECKLIST BELOW AND FILL IN THE DESCRIPTION ABOVE**"
 sed -i '/\*\*BEFORE SUBMITTING, PLEASE READ.*\*\*/,$d' "${NEW}"
 # Remove HTML <details> section that includes <summary> text of "PR Checklist (Click to Expand)"
 python3 - <<EOF
 import re
 with open("${NEW}", "r") as file:
    content = file.read()
 pattern = re.compile(r'(---\n\n)?<details>.*?<summary>.*?PR Checklist \(Click to Expand\).*?</summary>.*?</details>', re.DOTALL)
 content = re.sub(pattern, '', content)
 with open("${NEW}", "w") as file:
    file.write(content)
 EOF
 # Run this only if ${NEW} is different than ${OLD}
 if ! cmp -s "${OLD}" "${NEW}"; then
    gh pr edit --body-file "${NEW}" "${PR_NUMBER}"
    echo
    echo "Updated PR body:"
    echo
    cat "${NEW}"
 else
    echo "No changes needed"
 fi
--- a/.github/workflows/actionlint.yml
+++ b/.github/workflows/actionlint.yml
@@ -1,37 +0,0 @@
 name: Lint GitHub Actions workflows
 on:
  push:
    branches:
      - "main"
    paths:
      - '.github/workflows/*.ya?ml'
      - '.github/workflows/actionlint.*'
  pull_request:
    branches:
      - "main"
    paths:
      - '.github/workflows/*.ya?ml'
      - '.github/workflows/actionlint.*'
 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@eef61447b9ff4aafe5dcd4e0bbf5d482be7e7871 # v4.2.1
        with:
          fetch-depth: 0
      - name: "Run actionlint"
        run: |
          tools/actionlint.sh -color
--- a/.github/workflows/add_label_automerge.yml
+++ b/.github/workflows/add_label_automerge.yml
@@ -8,7 +8,7 @@ jobs:
        runs-on: ubuntu-latest
        steps:
            -   name: Add label
-                uses: actions/github-script@v7
+                uses: actions/github-script@60a0d83039c74a4aee543508d2ffcb1c3799cdea # v7.0.1
                with:
                    script: |
                        github.rest.issues.addLabels({
--- a/.github/workflows/clang-format.yml
+++ b/.github/workflows/clang-format.yml
@@ -1,41 +0,0 @@
 name: clang-format
 on:
  # Trigger the workflow on push or pull request,
  # but only for the main branch
  push:
    branches:
      - main
  pull_request:
    branches:
      - main
 jobs:
  clang-format:
    runs-on: ubuntu-latest
    strategy:
      matrix:
        python-version: ["3.11"]
    steps:
    - uses: actions/checkout@v4
    - name: Set up Python ${{ matrix.python-version }}
      uses: actions/setup-python@v5
      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
--- a/.github/workflows/cleanup_pr_body.yml
+++ b/.github/workflows/cleanup_pr_body.yml
@@ -0,0 +1,26 @@
 name: Cleanup PR Body
 on:
  pull_request_target:
    types: [opened, reopened, edited]
 permissions:
  pull-requests: write
 jobs:
  update-description:
    runs-on: ubuntu-latest
    steps:
      - name: Checkout repository
        uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
      - name: Set up Python
        uses: actions/setup-python@0b93645e9fea7318ecaed2b359559ac225c90a2b # v5.3.0
        with:
          python-version: '3.12'
      - name: Update PR description
        env:
          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
        run: .github/scripts/cleanup_pr_body.sh "${{ github.event.number }}"
--- a/.github/workflows/lint-and-deploy.yaml
+++ b/.github/workflows/lint-and-deploy.yaml
@@ -0,0 +1,82 @@
 name: Lint and Deploy Charts
 on: pull_request
 jobs:
  lint-and-deploy:
    runs-on: ubuntu-latest
    steps:
      - name: Checkout
        uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
        with:
          fetch-depth: 0
      - name: Set up Helm
        uses: azure/setup-helm@fe7b79cd5ee1e45176fcad797de68ecaf3ca4814 # v4.2.0
        with:
          version: v3.14.4
       #Python is required because ct lint runs Yamale and yamllint which require Python.
      - uses: actions/setup-python@0b93645e9fea7318ecaed2b359559ac225c90a2b # v5.3.0
        with:
          python-version: '3.13'
      - name: Set up chart-testing
        uses: helm/chart-testing-action@e6669bcd63d7cb57cb4380c33043eebe5d111992 # v2.6.1
        with:
          version: v3.10.1
      - name: Run chart-testing (lint)
        run: ct lint --target-branch ${{ github.event.repository.default_branch }} --chart-dirs examples/online_serving/chart-helm --charts examples/online_serving/chart-helm
      - name: Setup minio
        run: |
          docker network create vllm-net
          docker run -d -p 9000:9000 --name minio --net vllm-net \
                     -e "MINIO_ACCESS_KEY=minioadmin" \
                     -e "MINIO_SECRET_KEY=minioadmin" \
                     -v /tmp/data:/data \
                     -v /tmp/config:/root/.minio \
                     minio/minio server /data
          export AWS_ACCESS_KEY_ID=minioadmin
          export AWS_SECRET_ACCESS_KEY=minioadmin
          export AWS_EC2_METADATA_DISABLED=true
          mkdir opt-125m
          cd opt-125m && curl -O -Ls "https://huggingface.co/facebook/opt-125m/resolve/main/{pytorch_model.bin,config.json,generation_config.json,merges.txt,special_tokens_map.json,tokenizer_config.json,vocab.json}" && cd ..
          aws --endpoint-url http://127.0.0.1:9000/ s3 mb s3://testbucket
          aws --endpoint-url http://127.0.0.1:9000/ s3 cp opt-125m/ s3://testbucket/opt-125m --recursive
      - name: Create kind cluster
        uses: helm/kind-action@0025e74a8c7512023d06dc019c617aa3cf561fde # v1.10.0
      - name: Build the Docker image vllm cpu
        run: docker buildx build -f Dockerfile.cpu -t vllm-cpu-env .
      - name: Configuration of docker images, network and namespace for the kind cluster
        run: |
          docker pull amazon/aws-cli:2.6.4
          kind load docker-image  amazon/aws-cli:2.6.4 --name chart-testing
          kind load docker-image vllm-cpu-env:latest --name chart-testing
          docker network connect vllm-net "$(docker ps -aqf "name=chart-testing-control-plane")"
          kubectl create ns ns-vllm
      - name: Run chart-testing (install)
        run: |
          export AWS_ACCESS_KEY_ID=minioadmin
          export AWS_SECRET_ACCESS_KEY=minioadmin
          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=$AWS_SECRET_ACCESS_KEY --set resources.requests.cpu=1 --set resources.requests.memory=4Gi --set resources.limits.cpu=2 --set resources.limits.memory=5Gi --set image.env[0].name=VLLM_CPU_KVCACHE_SPACE --set image.env[1].name=VLLM_LOGGING_LEVEL --set-string image.env[0].value="1" --set-string image.env[1].value="DEBUG" --set-string extraInit.s3modelpath="opt-125m/" --set-string 'resources.limits.nvidia\.com/gpu=0' --set-string 'resources.requests.nvidia\.com/gpu=0' --set-string image.repository="vllm-cpu-env"
      - name: curl test
        run: |
          kubectl -n ns-vllm port-forward service/test-vllm-service 8001:80 &
          sleep 10
          CODE="$(curl -v -f --location http://localhost:8001/v1/completions \
                  --header "Content-Type: application/json" \
                  --data '{
                          "model": "opt-125m",
                          "prompt": "San Francisco is a",
                          "max_tokens": 7,
                          "temperature": 0
                  }'):$CODE"
          echo "$CODE"
--- a/.github/workflows/matchers/mypy.json
+++ b/.github/workflows/matchers/mypy.json
@@ -0,0 +1,16 @@
 {
  "problemMatcher": [
    {
      "owner": "mypy",
      "pattern": [
        {
          "regexp": "^(.+):(\\d+):\\s(error|warning):\\s(.+)$",
          "file": 1,
          "line": 2,
          "severity": 3,
          "message": 4
        }
      ]
    }
  ]
 }
--- a/.github/workflows/mypy.yaml
+++ b/.github/workflows/mypy.yaml
@@ -1,35 +0,0 @@
 name: mypy
 on:
  # Trigger the workflow on push or pull request,
  # but only for the main branch
  push:
    branches:
      - main
  pull_request:
    branches:
      - main
 jobs:
  mypy:
    runs-on: ubuntu-latest
    strategy:
      matrix:
        python-version: ["3.8", "3.9", "3.10", "3.11", "3.12"]
    steps:
    - uses: actions/checkout@v4
    - name: Set up Python ${{ matrix.python-version }}
      uses: actions/setup-python@v5
      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: |
        tools/mypy.sh
--- a/.github/workflows/pre-commit.yml
+++ b/.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
--- a/.github/workflows/publish.yml
+++ b/.github/workflows/publish.yml
@@ -21,7 +21,7 @@ jobs:
      upload_url: ${{ steps.create_release.outputs.upload_url }}
    steps:
      - name: Checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
      - name: Extract branch info
        shell: bash
@@ -30,7 +30,7 @@ jobs:
      - name: Create Release
        id: create_release
-        uses: "actions/github-script@v7"
+        uses: actions/github-script@60a0d83039c74a4aee543508d2ffcb1c3799cdea # v7.0.1
        env:
          RELEASE_TAG: ${{ env.release_tag }}
        with:
@@ -39,67 +39,68 @@ jobs:
            const script = require('.github/workflows/scripts/create_release.js')
            await script(github, context, core)
-  wheel:
+  # NOTE(simon): No longer build wheel using Github Actions. See buildkite's release workflow. 
-    name: Build Wheel
+  # wheel:
-    runs-on: ${{ matrix.os }}
+  #   name: Build Wheel
-    needs: release
+  #   runs-on: ${{ matrix.os }}
  #   needs: release
-    strategy:
+  #   strategy:
-      fail-fast: false
+  #     fail-fast: false
-      matrix:
+  #     matrix:
-          os: ['ubuntu-20.04']
+  #         os: ['ubuntu-20.04']
-          python-version: ['3.8', '3.9', '3.10', '3.11', '3.12']
+  #         python-version: ['3.9', '3.10', '3.11', '3.12']
-          pytorch-version: ['2.4.0']  # Must be the most recent version that meets requirements-cuda.txt.
+  #         pytorch-version: ['2.4.0']  # Must be the most recent version that meets requirements-cuda.txt.
-          cuda-version: ['11.8', '12.1']
+  #         cuda-version: ['11.8', '12.1']
-    steps:
+  #   steps:
-      - name: Checkout
+  #     - name: Checkout
-        uses: actions/checkout@v4
+  #       uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
-      - name: Setup ccache
+  #     - name: Setup ccache
-        uses: hendrikmuhs/ccache-action@v1.2
+  #       uses: hendrikmuhs/ccache-action@ed74d11c0b343532753ecead8a951bb09bb34bc9 # v1.2.14
-        with:
+  #       with:
-          create-symlink: true
+  #         create-symlink: true
-          key: ${{ github.job }}-${{ matrix.python-version }}-${{ matrix.cuda-version }}
+  #         key: ${{ github.job }}-${{ matrix.python-version }}-${{ matrix.cuda-version }}
-      - name: Set up Linux Env
+  #     - name: Set up Linux Env
-        if: ${{ runner.os == 'Linux' }}
+  #       if: ${{ runner.os == 'Linux' }}
-        run: |
+  #       run: |
-          bash -x .github/workflows/scripts/env.sh
+  #         bash -x .github/workflows/scripts/env.sh
-      - name: Set up Python
+  #     - name: Set up Python
-        uses: actions/setup-python@v5
+  #       uses: actions/setup-python@0b93645e9fea7318ecaed2b359559ac225c90a2b # v5.3.0
-        with:
+  #       with:
-            python-version: ${{ matrix.python-version }}
+  #           python-version: ${{ matrix.python-version }}
-      - name: Install CUDA ${{ matrix.cuda-version }}
+  #     - name: Install CUDA ${{ matrix.cuda-version }}
-        run: |
+  #       run: |
-          bash -x .github/workflows/scripts/cuda-install.sh ${{ matrix.cuda-version }} ${{ matrix.os }}
+  #         bash -x .github/workflows/scripts/cuda-install.sh ${{ matrix.cuda-version }} ${{ matrix.os }}
-      - name: Install PyTorch ${{ matrix.pytorch-version }} with CUDA ${{ matrix.cuda-version }}
+  #     - name: Install PyTorch ${{ matrix.pytorch-version }} with CUDA ${{ matrix.cuda-version }}
-        run: |
+  #       run: |
-          bash -x .github/workflows/scripts/pytorch-install.sh ${{ matrix.python-version }} ${{ matrix.pytorch-version }} ${{ matrix.cuda-version }}
+  #         bash -x .github/workflows/scripts/pytorch-install.sh ${{ matrix.python-version }} ${{ matrix.pytorch-version }} ${{ matrix.cuda-version }}
-      - name: Build wheel
+  #     - name: Build wheel
-        shell: bash
+  #       shell: bash
-        env:
+  #       env:
-          CMAKE_BUILD_TYPE: Release # do not compile with debug symbol to reduce wheel size
+  #         CMAKE_BUILD_TYPE: Release # do not compile with debug symbol to reduce wheel size
-        run: |
+  #       run: |
-          bash -x .github/workflows/scripts/build.sh ${{ matrix.python-version }} ${{ matrix.cuda-version }}
+  #         bash -x .github/workflows/scripts/build.sh ${{ matrix.python-version }} ${{ matrix.cuda-version }}
-          wheel_name=$(find dist -name "*whl" -print0 | xargs -0 -n 1 basename)
+  #         wheel_name=$(find dist -name "*whl" -print0 | xargs -0 -n 1 basename)
-          asset_name=${wheel_name//"linux"/"manylinux1"}
+  #         asset_name=${wheel_name//"linux"/"manylinux1"}
-          echo "wheel_name=${wheel_name}" >> "$GITHUB_ENV"
+  #         echo "wheel_name=${wheel_name}" >> "$GITHUB_ENV"
-          echo "asset_name=${asset_name}" >> "$GITHUB_ENV"
+  #         echo "asset_name=${asset_name}" >> "$GITHUB_ENV"
-      - name: Upload Release Asset
+  #     - name: Upload Release Asset
-        uses: actions/upload-release-asset@v1
+  #       uses: actions/upload-release-asset@e8f9f06c4b078e705bd2ea027f0926603fc9b4d5 # v1.0.2
-        env:
+  #       env:
-          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+  #         GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-        with:
+  #       with:
-          upload_url: ${{ needs.release.outputs.upload_url }}
+  #         upload_url: ${{ needs.release.outputs.upload_url }}
-          asset_path: ./dist/${{ env.wheel_name }}
+  #         asset_path: ./dist/${{ env.wheel_name }}
-          asset_name: ${{ env.asset_name }}
+  #         asset_name: ${{ env.asset_name }}
-          asset_content_type: application/*
+  #         asset_content_type: application/*
      # (Danielkinz): This last step will publish the .whl to pypi. Warning: untested
      # - name: Publish package
--- a/.github/workflows/reminder_comment.yml
+++ b/.github/workflows/reminder_comment.yml
@@ -2,20 +2,24 @@ name: PR Reminder Comment Bot
 on:
  pull_request_target:
    types: [opened]
 jobs:
  pr_reminder:
    runs-on: ubuntu-latest
    steps:
      - name: Remind to run full CI on PR
-        uses: actions/github-script@v7
+        uses: actions/github-script@60a0d83039c74a4aee543508d2ffcb1c3799cdea # v7.0.1
        with:
          script: |
            github.rest.issues.createComment({
              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 }}
--- a/.github/workflows/ruff.yml
+++ b/.github/workflows/ruff.yml
@@ -1,37 +0,0 @@
 name: ruff
 on:
  # Trigger the workflow on push or pull request,
  # but only for the main branch
  push:
    branches:
      - main
  pull_request:
    branches:
      - main
 jobs:
  ruff:
    runs-on: ubuntu-latest
    strategy:
      matrix:
        python-version: ["3.8", "3.9", "3.10", "3.11", "3.12"]
    steps:
    - uses: actions/checkout@v4
    - name: Set up Python ${{ matrix.python-version }}
      uses: actions/setup-python@v5
      with:
        python-version: ${{ matrix.python-version }}
    - name: Install dependencies
      run: |
        python -m pip install --upgrade pip
        pip install -r requirements-lint.txt
    - name: Analysing the code with ruff
      run: |
        ruff check .
    - name: Spelling check with codespell
      run: |
        codespell --toml pyproject.toml
    - name: Run isort
      run: |
        isort . --check-only
--- a/.github/workflows/scripts/cuda-install.sh
+++ b/.github/workflows/scripts/cuda-install.sh
@@ -1,16 +1,16 @@
 #!/bin/bash
 # Replace '.' with '-' ex: 11.8 -> 11-8
-cuda_version=$(echo $1 | tr "." "-")
+cuda_version=$(echo "$1" | tr "." "-")
 # Removes '-' and '.' ex: ubuntu-20.04 -> ubuntu2004
-OS=$(echo $2 | tr -d ".\-")
+OS=$(echo "$2" | tr -d ".\-")
 # Installs CUDA
-wget -nv https://developer.download.nvidia.com/compute/cuda/repos/${OS}/x86_64/cuda-keyring_1.1-1_all.deb
+wget -nv "https://developer.download.nvidia.com/compute/cuda/repos/${OS}/x86_64/cuda-keyring_1.1-1_all.deb"
 sudo dpkg -i cuda-keyring_1.1-1_all.deb
 rm cuda-keyring_1.1-1_all.deb
 sudo apt -qq update
-sudo apt -y install cuda-${cuda_version} cuda-nvcc-${cuda_version} cuda-libraries-dev-${cuda_version}
+sudo apt -y install "cuda-${cuda_version}" "cuda-nvcc-${cuda_version}" "cuda-libraries-dev-${cuda_version}"
 sudo apt clean
 # Test nvcc
--- a/.github/workflows/scripts/pytorch-install.sh
+++ b/.github/workflows/scripts/pytorch-install.sh
@@ -6,7 +6,7 @@ cuda_version=$3
 # Install torch
 $python_executable -m pip install numpy pyyaml scipy ipython mkl mkl-include ninja cython typing pandas typing-extensions dataclasses setuptools && conda clean -ya
-$python_executable -m pip install torch==${pytorch_version}+cu${cuda_version//./} --extra-index-url https://download.pytorch.org/whl/cu${cuda_version//./}
+$python_executable -m pip install torch=="${pytorch_version}+cu${cuda_version//./}" --extra-index-url "https://download.pytorch.org/whl/cu${cuda_version//./}"
 # Print version information
 $python_executable --version
--- a/.github/workflows/stale.yml
+++ b/.github/workflows/stale.yml
@@ -0,0 +1,52 @@
 name: 'Close inactive issues and PRs'
 on:
  schedule:
    # Daily at 1:30 AM UTC
    - cron: '30 1 * * *'
 jobs:
  close-issues-and-pull-requests:
    permissions:
      issues: write
      pull-requests: write
      actions: write
    runs-on: ubuntu-latest
    steps:
      - uses: actions/stale@28ca1036281a5e5922ead5184a1bbf96e5fc984e # v9.0.0
        with:
          # Increasing this value ensures that changes to this workflow
          # propagate to all issues and PRs in days rather than months
          operations-per-run: 1000
          exempt-draft-pr: true
          exempt-issue-labels: 'keep-open'
          exempt-pr-labels: 'keep-open'
          labels-to-add-when-unstale: 'unstale'
          labels-to-remove-when-stale: 'unstale'
          days-before-issue-stale: 90
          days-before-issue-close: 30
          stale-issue-label: 'stale'
          stale-issue-message: >
            This issue has been automatically marked as stale because it has not
            had any activity within 90 days. It will be automatically closed if no
            further activity occurs within 30 days. Leave a comment if
            you feel this issue should remain open. Thank you!
          close-issue-message: >
            This issue has been automatically closed due to inactivity. Please
            feel free to reopen if you feel it is still relevant. Thank you!
          days-before-pr-stale: 90
          days-before-pr-close: 30
          stale-pr-label: 'stale'
          stale-pr-message: >
            This pull request has been automatically marked as stale because it
            has not had any activity within 90 days. It will be automatically
            closed if no further activity occurs within 30 days. Leave a comment
            if you feel this pull request should remain open. Thank you!
          close-pr-message: >
            This pull request has been automatically closed due to inactivity.
            Please feel free to reopen if you intend to continue working on it.
            Thank you!
--- a/.github/workflows/yapf.yml
+++ b/.github/workflows/yapf.yml
@@ -1,31 +0,0 @@
 name: yapf
 on:
  # Trigger the workflow on push or pull request,
  # but only for the main branch
  push:
    branches:
      - main
  pull_request:
    branches:
      - main
 jobs:
  yapf:
    runs-on: ubuntu-latest
    strategy:
      matrix:
        python-version: ["3.8", "3.9", "3.10", "3.11", "3.12"]
    steps:
    - uses: actions/checkout@v4
    - name: Set up Python ${{ matrix.python-version }}
      uses: actions/setup-python@v5
      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 .
--- a/.gitignore
+++ b/.gitignore
@@ -79,8 +79,7 @@ instance/
 # Sphinx documentation
 docs/_build/
-docs/source/getting_started/examples/*.rst
+docs/source/getting_started/examples/
 !**/*.template.rst
 # PyBuilder
 .pybuilder/
@@ -202,3 +201,4 @@ benchmarks/*.json
 # Linting
 actionlint
 shellcheck*/
--- a/.pre-commit-config.yaml
+++ b/.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
--- a/.readthedocs.yaml
+++ b/.readthedocs.yaml
@@ -6,17 +6,16 @@ version: 2
 build:
  os: ubuntu-22.04
  tools:
-    python: "3.8"
+    python: "3.12"
 sphinx:
-   configuration: docs/source/conf.py
+  configuration: docs/source/conf.py
-   fail_on_warning: true
+  fail_on_warning: true
 # If using Sphinx, optionally build your docs in additional formats such as PDF
 formats: []
 # Optionally declare the Python requirements required to build your docs
 python:
-   install:
+  install:
-   - requirements: docs/requirements-docs.txt
+    - requirements: docs/requirements-docs.txt
--- a/.shellcheckrc
+++ b/.shellcheckrc
@@ -0,0 +1,9 @@
 # rules currently disabled:
 #
 #   SC1091 (info): Not following: <sourced file> was not specified as input (see shellcheck -x)
 #   SC2004 (style): $/${} is unnecessary on arithmetic variables.
 #   SC2129 (style): Consider using { cmd1; cmd2; } >> file instead of individual redirects.
 #   SC2155 (warning): Declare and assign separately to avoid masking return values.
 #   SC2164 (warning): Use 'cd ... || exit' or 'cd ... || return' in case cd fails.
 #
 disable=SC1091,SC2004,SC2129,SC2155,SC2164
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -24,20 +24,17 @@ 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.
 #
-set(PYTHON_SUPPORTED_VERSIONS "3.8" "3.9" "3.10" "3.11" "3.12")
+set(PYTHON_SUPPORTED_VERSIONS "3.9" "3.10" "3.11" "3.12")
 # Supported NVIDIA architectures.
-set(CUDA_SUPPORTED_ARCHS "7.0;7.5;8.0;8.6;8.9;9.0")
+set(CUDA_SUPPORTED_ARCHS "7.0;7.2;7.5;8.0;8.6;8.7;8.9;9.0")
 # Supported AMD GPU architectures.
-set(HIP_SUPPORTED_ARCHS "gfx906;gfx908;gfx90a;gfx940;gfx941;gfx942;gfx1030;gfx1100")
+set(HIP_SUPPORTED_ARCHS "gfx906;gfx908;gfx90a;gfx940;gfx941;gfx942;gfx1030;gfx1100;gfx1101")
 #
 # Supported/expected torch versions for CUDA/ROCm.
@@ -49,8 +46,8 @@ set(HIP_SUPPORTED_ARCHS "gfx906;gfx908;gfx90a;gfx940;gfx941;gfx942;gfx1030;gfx11
 # requirements.txt files and should be kept consistent.  The ROCm torch
 # versions are derived from Dockerfile.rocm
 #
-set(TORCH_SUPPORTED_VERSION_CUDA "2.4.0")
+set(TORCH_SUPPORTED_VERSION_CUDA "2.5.1")
-set(TORCH_SUPPORTED_VERSION_ROCM "2.5.0")
+set(TORCH_SUPPORTED_VERSION_ROCM "2.5.1")
 #
 # Try to find python package with an executable that exactly matches
@@ -83,24 +80,6 @@ endif()
 #
 find_package(Torch REQUIRED)
 #
 message(STATUS "Enabling core extension.")
 # Define _core_C extension
 #  built for (almost) every target platform, (excludes TPU and Neuron)
 set(VLLM_EXT_SRC
  "csrc/core/torch_bindings.cpp")
 define_gpu_extension_target(
  _core_C
  DESTINATION vllm
  LANGUAGE CXX
  SOURCES ${VLLM_EXT_SRC}
  COMPILE_FLAGS ${CXX_COMPILE_FLAGS}
  USE_SABI 3
  WITH_SOABI)
 #
 # Forward the non-CUDA device extensions to external CMake scripts.
 #
@@ -187,33 +166,61 @@ endif()
 #
 # Use FetchContent for C++ dependencies that are compiled as part of vLLM's build process.
-# Configure it to place files in vllm/.deps, in order to play nicely with sccache.
+# setup.py will override FETCHCONTENT_BASE_DIR to play nicely with sccache.
 # Each dependency that produces build artifacts should override its BINARY_DIR to avoid
 # conflicts between build types. It should instead be set to ${CMAKE_BINARY_DIR}/<dependency>.
 #
 include(FetchContent)
-get_filename_component(PROJECT_ROOT_DIR "${CMAKE_CURRENT_SOURCE_DIR}" ABSOLUTE)
+file(MAKE_DIRECTORY ${FETCHCONTENT_BASE_DIR}) # Ensure the directory exists
 file(MAKE_DIRECTORY "${FETCHCONTENT_BASE_DIR}")
 set(FETCHCONTENT_BASE_DIR "${PROJECT_ROOT_DIR}/.deps")
 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
 #
 set(VLLM_EXT_SRC
  "csrc/cache_kernels.cu"
-  "csrc/attention/attention_kernels.cu"
+  "csrc/attention/paged_attention_v1.cu"
  "csrc/attention/paged_attention_v2.cu"
  "csrc/pos_encoding_kernels.cu"
  "csrc/activation_kernels.cu"
  "csrc/layernorm_kernels.cu"
  "csrc/layernorm_quant_kernels.cu"
  "csrc/quantization/gptq/q_gemm.cu"
  "csrc/quantization/compressed_tensors/int8_quant_kernels.cu"
  "csrc/quantization/fp8/common.cu"
  "csrc/quantization/fused_kernels/fused_layernorm_dynamic_per_token_quant.cu"
  "csrc/quantization/gguf/gguf_kernel.cu"
  "csrc/cuda_utils_kernels.cu"
  "csrc/moe_align_block_size_kernels.cu"
  "csrc/prepare_inputs/advance_step.cu"
  "csrc/torch_bindings.cpp")
@@ -221,19 +228,32 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
  SET(CUTLASS_ENABLE_HEADERS_ONLY ON CACHE BOOL "Enable only the header library")
  # Set CUTLASS_REVISION manually -- its revision detection doesn't work in this case.
-  set(CUTLASS_REVISION "v3.5.1" CACHE STRING "CUTLASS revision to use")
+  set(CUTLASS_REVISION "v3.6.0" CACHE STRING "CUTLASS revision to use")
-  FetchContent_Declare(
+  # Use the specified CUTLASS source directory for compilation if VLLM_CUTLASS_SRC_DIR is provided
  if (DEFINED ENV{VLLM_CUTLASS_SRC_DIR})
    set(VLLM_CUTLASS_SRC_DIR $ENV{VLLM_CUTLASS_SRC_DIR})
  endif()
  if(VLLM_CUTLASS_SRC_DIR)
    if(NOT IS_ABSOLUTE VLLM_CUTLASS_SRC_DIR)
      get_filename_component(VLLM_CUTLASS_SRC_DIR "${VLLM_CUTLASS_SRC_DIR}" ABSOLUTE)
    endif()
    message(STATUS "The VLLM_CUTLASS_SRC_DIR is set, using ${VLLM_CUTLASS_SRC_DIR} for compilation")
    FetchContent_Declare(cutlass SOURCE_DIR ${VLLM_CUTLASS_SRC_DIR})
  else()
    FetchContent_Declare(
        cutlass
        GIT_REPOSITORY https://github.com/nvidia/cutlass.git
-        GIT_TAG v3.5.1
+        GIT_TAG v3.7.0
        GIT_PROGRESS TRUE
        # Speed up CUTLASS download by retrieving only the specified GIT_TAG instead of the history.
        # Important: If GIT_SHALLOW is enabled then GIT_TAG works only with branch names and tags.
        # So if the GIT_TAG above is updated to a commit hash, GIT_SHALLOW must be set to FALSE
        GIT_SHALLOW TRUE
-  )
+    )
  endif()
  FetchContent_MakeAvailable(cutlass)
  list(APPEND VLLM_EXT_SRC
@@ -241,10 +261,12 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
    "csrc/mamba/causal_conv1d/causal_conv1d.cu"
    "csrc/quantization/aqlm/gemm_kernels.cu"
    "csrc/quantization/awq/gemm_kernels.cu"
    "csrc/quantization/gguf/gguf_kernel.cu"
    "csrc/custom_all_reduce.cu"
    "csrc/permute_cols.cu"
-    "csrc/quantization/cutlass_w8a8/scaled_mm_entry.cu")
+    "csrc/quantization/cutlass_w8a8/scaled_mm_entry.cu"
    "csrc/sparse/cutlass/sparse_scaled_mm_entry.cu"
    "csrc/sparse/cutlass/sparse_compressor_entry.cu"
    "csrc/cutlass_extensions/common.cpp")
  set_gencode_flags_for_srcs(
    SRCS "${VLLM_EXT_SRC}"
@@ -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;8.9;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"
@@ -270,15 +292,19 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
    message(STATUS "Building Marlin kernels for archs: ${MARLIN_ARCHS}")
  else()
    message(STATUS "Not building Marlin kernels as no compatible archs found"
-                   "in CUDA target architectures")
+                   " in CUDA target architectures")
  endif()
  #
  # The cutlass_scaled_mm kernels for Hopper (c3x, i.e. CUTLASS 3.x) require
-  # CUDA 12.0 or later (and only work on Hopper, 9.0/9.0a for now).
+  # CUDA 12.0 or later (and only work on Hopper, 9.0a for now).
-  cuda_archs_loose_intersection(SCALED_MM_3X_ARCHS "9.0;9.0a" "${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}")
@@ -305,7 +331,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
  # For the cutlass_scaled_mm kernels we want to build the c2x (CUTLASS 2.x)
  # kernels for the remaining archs that are not already built for 3x.
  cuda_archs_loose_intersection(SCALED_MM_2X_ARCHS
-    "7.5;8.0;8.6;8.9;9.0" "${CUDA_ARCHS}")
+    "7.5;8.0;8.6;8.7;8.9;9.0" "${CUDA_ARCHS}")
  # subtract out the archs that are already built for 3x
  list(REMOVE_ITEM SCALED_MM_2X_ARCHS ${SCALED_MM_3X_ARCHS})
  if (SCALED_MM_2X_ARCHS)
@@ -326,6 +352,31 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
    endif()
  endif()
  #
  # 2:4 Sparse Kernels
  # The 2:4 sparse kernels cutlass_scaled_sparse_mm and cutlass_compressor
  # require CUDA 12.2 or later (and only work on Hopper, 9.0a for now).
  if(${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")
    set_gencode_flags_for_srcs(
      SRCS "${SRCS}"
      CUDA_ARCHS "${SCALED_MM_3X_ARCHS}")
    list(APPEND VLLM_EXT_SRC "${SRCS}")
    list(APPEND VLLM_GPU_FLAGS "-DENABLE_SPARSE_SCALED_MM_C3X=1")
    message(STATUS "Building sparse_scaled_mm_c3x for archs: ${SCALED_MM_3X_ARCHS}")
  else()
    if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.2 AND SCALED_MM_3X_ARCHS)
      message(STATUS "Not building sparse_scaled_mm_c3x kernels as CUDA Compiler version is "
                     "not >= 12.2, we recommend upgrading to CUDA 12.2 or later "
                     "if you intend on running FP8 sparse quantized models on Hopper.")
    else()
      message(STATUS "Not building sparse_scaled_mm_c3x as no compatible archs found "
                     "in CUDA target architectures")
    endif()
  endif()
  #
  # Machete kernels
@@ -407,7 +458,7 @@ define_gpu_extension_target(
  SOURCES ${VLLM_EXT_SRC}
  COMPILE_FLAGS ${VLLM_GPU_FLAGS}
  ARCHITECTURES ${VLLM_GPU_ARCHES}
-  INCLUDE_DIRECTORIES ${CUTLASS_INCLUDE_DIR}
+  INCLUDE_DIRECTORIES ${CUTLASS_INCLUDE_DIR};${CUTLASS_TOOLS_UTIL_INCLUDE_DIR}
  USE_SABI 3
  WITH_SOABI)
@@ -423,6 +474,7 @@ target_compile_definitions(_C PRIVATE CUTLASS_ENABLE_DIRECT_CUDA_DRIVER_CALL=1)
 set(VLLM_MOE_EXT_SRC
  "csrc/moe/torch_bindings.cpp"
  "csrc/moe/moe_align_sum_kernels.cu"
  "csrc/moe/topk_softmax_kernels.cu")
 set_gencode_flags_for_srcs(
@@ -430,7 +482,7 @@ set_gencode_flags_for_srcs(
  CUDA_ARCHS "${CUDA_ARCHS}")
 if(VLLM_GPU_LANG STREQUAL "CUDA")
-  cuda_archs_loose_intersection(MARLIN_MOE_ARCHS "8.0;8.6;8.9;9.0" "${CUDA_ARCHS}")
+  cuda_archs_loose_intersection(MARLIN_MOE_ARCHS "8.0;8.6;8.7;8.9;9.0" "${CUDA_ARCHS}")
  if (MARLIN_MOE_ARCHS)
    set(MARLIN_MOE_SRC
        "csrc/moe/marlin_kernels/marlin_moe_kernel.h"
@@ -450,7 +502,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
    message(STATUS "Building Marlin MOE kernels for archs: ${MARLIN_MOE_ARCHS}")
  else()
    message(STATUS "Not building Marlin MOE kernels as no compatible archs found"
-                   "in CUDA target architectures")
+                   " in CUDA target architectures")
  endif()
 endif()
@@ -485,7 +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 ()
@@ -508,7 +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.
@@ -520,41 +572,41 @@ 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 013f0c4fc47e6574060879d9734c1df8c5c273bd
+          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/
+  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"
+  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"
 )
 # Nothing after vllm-flash-attn, see comment about macros above
--- a/CODE_OF_CONDUCT.md
+++ b/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
--- a/CONTRIBUTING.md
+++ b/CONTRIBUTING.md
@@ -1,50 +1,3 @@
 # Contributing to vLLM
-Thank you for your interest in contributing to vLLM! Our community is open to everyone and welcomes all kinds of contributions, no matter how small or large. There are several ways you can contribute to the project:
+You may find information about contributing to vLLM on [docs.vllm.ai](https://docs.vllm.ai/en/latest/contributing/overview.html).
 - Identify and report any issues or bugs.
 - Request or add support for a new model.
 - Suggest or implement new features.
 - Improve documentation or contribute a how-to guide. 
 We also believe in the power of community support; thus, answering queries, offering PR reviews, and assisting others are also highly regarded and beneficial contributions.
 Finally, one of the most impactful ways to support us is by raising awareness about vLLM. Talk about it in your blog posts and highlight how it's driving your incredible projects. Express your support on social media if you're using vLLM, or simply offer your appreciation by starring our repository!
 ## Developing
 Depending on the kind of development you'd like to do (e.g. Python, CUDA), you can choose to build vLLM with or without compilation. Check out the [building from source](https://docs.vllm.ai/en/latest/getting_started/installation.html#build-from-source) documentation for details.
 ## Testing
 ```bash
 pip install -r requirements-dev.txt
 # linting and formatting
 bash format.sh
 # Static type checking
 mypy
 # Unit tests
 pytest tests/
 ```
 **Note:** Currently, the repository does not pass the ``mypy`` tests.
 ## Contribution Guidelines
 ### Issues
 If you encounter a bug or have a feature request, please [search existing issues](https://github.com/vllm-project/vllm/issues?q=is%3Aissue) first to see if it has already been reported. If not, please [file a new issue](https://github.com/vllm-project/vllm/issues/new/choose), providing as much relevant information as possible.
 > [!IMPORTANT]
 > If you discover a security vulnerability, please follow the instructions [here](/SECURITY.md#reporting-a-vulnerability).
 ### Pull Requests & Code Reviews
 Please check the PR checklist in the [PR template](.github/PULL_REQUEST_TEMPLATE.md) for detailed guide for contribution.
 ### Thank You
 Finally, thank you for taking the time to read these guidelines and for your interest in contributing to vLLM.
 All of your contributions help make vLLM a great tool and community for everyone!
--- a/34
+++ b/34
@@ -0,0 +1,34 @@
 Developer Certificate of Origin
 Version 1.1
 Copyright (C) 2004, 2006 The Linux Foundation and its contributors.
 Everyone is permitted to copy and distribute verbatim copies of this
 license document, but changing it is not allowed.
 Developer's Certificate of Origin 1.1
 By making a contribution to this project, I certify that:
 (a) The contribution was created in whole or in part by me and I
    have the right to submit it under the open source license
    indicated in the file; or
 (b) The contribution is based upon previous work that, to the best
    of my knowledge, is covered under an appropriate open source
    license and I have the right under that license to submit that
    work with modifications, whether created in whole or in part
    by me, under the same open source license (unless I am
    permitted to submit under a different license), as indicated
    in the file; or
 (c) The contribution was provided directly to me by some other
    person who certified (a), (b) or (c) and I have not modified
    it.
 (d) I understand and agree that this project and the contribution
    are public and that a record of the contribution (including all
    personal information I submit with it, including my sign-off) is
    maintained indefinitely and may be redistributed consistent with
    this project or the open source license(s) involved.
--- a/101
+++ b/101
@@ -2,8 +2,8 @@
 # to run the OpenAI compatible server.
 # Please update any changes made here to
-# docs/source/dev/dockerfile/dockerfile.rst 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 ####################
@@ -11,6 +11,7 @@ ARG CUDA_VERSION=12.4.1
 FROM nvidia/cuda:${CUDA_VERSION}-devel-ubuntu20.04 AS base
 ARG CUDA_VERSION=12.4.1
 ARG PYTHON_VERSION=3.12
 ARG TARGETPLATFORM
 ENV DEBIAN_FRONTEND=noninteractive
 # Install Python and other dependencies
@@ -44,12 +45,21 @@ RUN ldconfig /usr/local/cuda-$(echo $CUDA_VERSION | cut -d. -f1,2)/compat/
 WORKDIR /workspace
 # install build and runtime dependencies
 # arm64 (GH200) build follows the practice of "use existing pytorch" build,
 # we need to install torch and torchvision from the nightly builds first,
 # pytorch will not appear as a vLLM dependency in all of the following steps
 # after this step
 RUN --mount=type=cache,target=/root/.cache/pip \
    if [ "$TARGETPLATFORM" = "linux/arm64" ]; then \
        python3 -m pip install --index-url https://download.pytorch.org/whl/nightly/cu126 "torch==2.7.0.dev20250121+cu126" "torchvision==0.22.0.dev20250121";  \
    fi
 COPY requirements-common.txt requirements-common.txt
 COPY requirements-cuda.txt requirements-cuda.txt
 RUN --mount=type=cache,target=/root/.cache/pip \
    python3 -m pip install -r requirements-cuda.txt
 # cuda arch list used by torch
 # can be useful for both `dev` and `test`
 # explicitly set the list to avoid issues with torch 2.2
@@ -63,6 +73,7 @@ ENV VLLM_FA_CMAKE_GPU_ARCHES=${vllm_fa_cmake_gpu_arches}
 #################### WHEEL BUILD IMAGE ####################
 FROM base AS build
 ARG TARGETPLATFORM
 # install build dependencies
 COPY requirements-build.txt requirements-build.txt
@@ -115,8 +126,8 @@ RUN --mount=type=cache,target=/root/.cache/ccache \
 # Check the size of the wheel if RUN_WHEEL_CHECK is true
 COPY .buildkite/check-wheel-size.py check-wheel-size.py
-# Default max size of the wheel is 250MB
+# sync the default value with .buildkite/check-wheel-size.py
-ARG VLLM_MAX_SIZE_MB=250
+ARG VLLM_MAX_SIZE_MB=400
 ENV VLLM_MAX_SIZE_MB=$VLLM_MAX_SIZE_MB
 ARG RUN_WHEEL_CHECK=true
 RUN if [ "$RUN_WHEEL_CHECK" = "true" ]; then \
@@ -134,15 +145,17 @@ COPY requirements-test.txt requirements-test.txt
 COPY requirements-dev.txt requirements-dev.txt
 RUN --mount=type=cache,target=/root/.cache/pip \
    python3 -m pip install -r requirements-dev.txt
 #################### DEV IMAGE ####################
 #################### vLLM installation IMAGE ####################
 # image with vLLM installed
-FROM nvidia/cuda:${CUDA_VERSION}-base-ubuntu22.04 AS vllm-base
+# TODO: Restore to base image after FlashInfer AOT wheel fixed
 FROM nvidia/cuda:${CUDA_VERSION}-devel-ubuntu22.04 AS vllm-base
 ARG CUDA_VERSION=12.4.1
 ARG PYTHON_VERSION=3.12
 WORKDIR /vllm-workspace
 ENV DEBIAN_FRONTEND=noninteractive
 ARG TARGETPLATFORM
 RUN PYTHON_VERSION_STR=$(echo ${PYTHON_VERSION} | sed 's/\.//g') && \
    echo "export PYTHON_VERSION_STR=${PYTHON_VERSION_STR}" >> /etc/environment
@@ -151,7 +164,7 @@ RUN PYTHON_VERSION_STR=$(echo ${PYTHON_VERSION} | sed 's/\.//g') && \
 RUN echo 'tzdata tzdata/Areas select America' | debconf-set-selections \
    && echo 'tzdata tzdata/Zones/America select Los_Angeles' | debconf-set-selections \
    && apt-get update -y \
-    && apt-get install -y ccache software-properties-common git curl sudo vim python3-pip \
+    && apt-get install -y ccache software-properties-common git curl wget sudo vim python3-pip \
    && apt-get install -y ffmpeg libsm6 libxext6 libgl1 \
    && add-apt-repository ppa:deadsnakes/ppa \
    && apt-get update -y \
@@ -168,17 +181,45 @@ RUN echo 'tzdata tzdata/Areas select America' | debconf-set-selections \
 # or future versions of triton.
 RUN ldconfig /usr/local/cuda-$(echo $CUDA_VERSION | cut -d. -f1,2)/compat/
-# install vllm wheel first, so that torch etc will be installed
+# arm64 (GH200) build follows the practice of "use existing pytorch" build,
 # we need to install torch and torchvision from the nightly builds first,
 # pytorch will not appear as a vLLM dependency in all of the following steps
 # after this step
 RUN --mount=type=cache,target=/root/.cache/pip \
    if [ "$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";  \
    fi
 # Install vllm wheel first, so that torch etc will be installed.
 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
-RUN --mount=type=cache,target=/root/.cache/pip \
+# How to build this FlashInfer wheel:
-    . /etc/environment && \
+# $ export FLASHINFER_ENABLE_AOT=1
-    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
+# $ # Note we remove 7.0 from the arch list compared to the list below, since FlashInfer only supports sm75+
-COPY examples examples
+# $ export TORCH_CUDA_ARCH_LIST='7.5 8.0 8.6 8.9 9.0+PTX'
-#################### vLLM installation IMAGE ####################
+# $ 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 [ "$TARGETPLATFORM" != "linux/arm64" ]; then \
    python3 -m pip install https://wheels.vllm.ai/flashinfer/524304395bd1d8cd7d07db083859523fcaa246a4/flashinfer_python-0.2.0.post1-cp${PYTHON_VERSION_STR}-cp${PYTHON_VERSION_STR}-linux_x86_64.whl; \
 fi
 COPY examples examples
 # Although we build Flashinfer with AOT mode, there's still
 # some issues w.r.t. JIT compilation. Therefore we need to
 # install build dependencies for JIT compilation.
 # TODO: Remove this once FlashInfer AOT wheel is fixed
 COPY requirements-build.txt requirements-build.txt
 RUN --mount=type=cache,target=/root/.cache/pip \
    python3 -m pip install -r requirements-build.txt
 #################### vLLM installation IMAGE ####################
 #################### TEST IMAGE ####################
 # image to run unit testing suite
@@ -191,24 +232,48 @@ ADD . /vllm-workspace/
 RUN --mount=type=cache,target=/root/.cache/pip \
    python3 -m pip install -r requirements-dev.txt
 # install development dependencies (for testing)
 RUN --mount=type=cache,target=/root/.cache/pip \
    python3 -m pip install -e tests/vllm_test_utils
 # enable fast downloads from hf (for testing)
 RUN --mount=type=cache,target=/root/.cache/pip \
    python3 -m pip install hf_transfer
 ENV HF_HUB_ENABLE_HF_TRANSFER 1
 # Copy in the v1 package for testing (it isn't distributed yet)
 COPY vllm/v1 /usr/local/lib/python3.12/dist-packages/vllm/v1
 # doc requires source code
 # we hide them inside `test_docs/` , so that this source code
 # will not be imported by other tests
 RUN mkdir test_docs
 RUN mv docs test_docs/
 RUN mv vllm test_docs/
 #################### TEST IMAGE ####################
 #################### OPENAI API SERVER ####################
-# openai api server alternative
+# base openai image with additional requirements, for any subsequent openai-style images
-FROM vllm-base AS vllm-openai
+FROM vllm-base AS vllm-openai-base
 # install additional dependencies for openai api server
 RUN --mount=type=cache,target=/root/.cache/pip \
-    pip install accelerate hf_transfer 'modelscope!=1.15.0' bitsandbytes>=0.44.0 timm==0.9.10
+    if [ "$TARGETPLATFORM" = "linux/arm64" ]; then \
        pip install accelerate hf_transfer 'modelscope!=1.15.0' 'bitsandbytes>=0.42.0' 'timm==0.9.10' boto3 runai-model-streamer runai-model-streamer[s3]; \
    else \
        pip install accelerate hf_transfer 'modelscope!=1.15.0' 'bitsandbytes>=0.45.0' 'timm==0.9.10' boto3 runai-model-streamer runai-model-streamer[s3]; \
    fi
 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 ####################
--- a/Dockerfile.arm
+++ b/Dockerfile.arm
@@ -0,0 +1,62 @@
 # This vLLM Dockerfile is used to construct an image that can build and run vLLM on ARM CPU platform.
 FROM ubuntu:22.04 AS cpu-test-arm
 ENV CCACHE_DIR=/root/.cache/ccache
 ENV CMAKE_CXX_COMPILER_LAUNCHER=ccache
 RUN --mount=type=cache,target=/var/cache/apt \
    apt-get update -y \
    && apt-get install -y curl ccache git wget vim numactl gcc-12 g++-12 python3 python3-pip libtcmalloc-minimal4 libnuma-dev \
    && apt-get install -y ffmpeg libsm6 libxext6 libgl1 \
    && update-alternatives --install /usr/bin/gcc gcc /usr/bin/gcc-12 10 --slave /usr/bin/g++ g++ /usr/bin/g++-12
 # tcmalloc provides better memory allocation efficiency, e.g., holding memory in caches to speed up access of commonly-used objects.
 RUN --mount=type=cache,target=/root/.cache/pip \
    pip install py-cpuinfo  # Use this to gather CPU info and optimize based on ARM Neoverse cores
 # Set LD_PRELOAD for tcmalloc on ARM
 ENV LD_PRELOAD="/usr/lib/aarch64-linux-gnu/libtcmalloc_minimal.so.4"
 RUN echo 'ulimit -c 0' >> ~/.bashrc
 WORKDIR /workspace
 ARG PIP_EXTRA_INDEX_URL="https://download.pytorch.org/whl/cpu"
 ENV PIP_EXTRA_INDEX_URL=${PIP_EXTRA_INDEX_URL}
 RUN --mount=type=cache,target=/root/.cache/pip \
    --mount=type=bind,src=requirements-build.txt,target=requirements-build.txt \
    pip install --upgrade pip && \
    pip install -r requirements-build.txt
 FROM cpu-test-arm AS build
 WORKDIR /workspace/vllm
 RUN --mount=type=cache,target=/root/.cache/pip \
    --mount=type=bind,src=requirements-common.txt,target=requirements-common.txt \
    --mount=type=bind,src=requirements-cpu.txt,target=requirements-cpu.txt \
    pip install -v -r requirements-cpu.txt
 COPY . .
 ARG GIT_REPO_CHECK=0
 RUN --mount=type=bind,source=.git,target=.git \
    if [ "$GIT_REPO_CHECK" != 0 ]; then bash tools/check_repo.sh ; fi
 # Disabling AVX512 specific optimizations for ARM
 ARG VLLM_CPU_DISABLE_AVX512="true"
 ENV VLLM_CPU_DISABLE_AVX512=${VLLM_CPU_DISABLE_AVX512}
 RUN --mount=type=cache,target=/root/.cache/pip \
    --mount=type=cache,target=/root/.cache/ccache \
    --mount=type=bind,source=.git,target=.git \
    VLLM_TARGET_DEVICE=cpu python3 setup.py bdist_wheel && \
    pip install dist/*.whl && \
    rm -rf dist
 WORKDIR /workspace/
 RUN ln -s /workspace/vllm/tests && ln -s /workspace/vllm/examples && ln -s /workspace/vllm/benchmarks
 ENTRYPOINT ["python3", "-m", "vllm.entrypoints.openai.api_server"]
--- a/Dockerfile.cpu
+++ b/Dockerfile.cpu
@@ -16,13 +16,13 @@ RUN --mount=type=cache,target=/var/cache/apt \
 # intel-openmp provides additional performance improvement vs. openmp
 # tcmalloc provides better memory allocation efficiency, e.g, holding memory in caches to speed up access of commonly-used objects.
 RUN --mount=type=cache,target=/root/.cache/pip \
-    pip install intel-openmp
+    pip install intel-openmp==2025.0.1
 ENV LD_PRELOAD="/usr/lib/x86_64-linux-gnu/libtcmalloc_minimal.so.4:/usr/local/lib/libiomp5.so"
 RUN echo 'ulimit -c 0' >> ~/.bashrc
-RUN pip install intel_extension_for_pytorch==2.4.0
+RUN pip install intel_extension_for_pytorch==2.5.0
 WORKDIR /workspace
@@ -62,4 +62,8 @@ WORKDIR /workspace/
 RUN ln -s /workspace/vllm/tests && ln -s /workspace/vllm/examples && ln -s /workspace/vllm/benchmarks
 # install development dependencies (for testing)
 RUN --mount=type=cache,target=/root/.cache/pip \
    pip install -e tests/vllm_test_utils
 ENTRYPOINT ["python3", "-m", "vllm.entrypoints.openai.api_server"]
--- a/Dockerfile.hpu
+++ b/Dockerfile.hpu
@@ -0,0 +1,21 @@
 FROM vault.habana.ai/gaudi-docker/1.19.1/ubuntu22.04/habanalabs/pytorch-installer-2.5.1:latest
 COPY ./ /workspace/vllm
 WORKDIR /workspace/vllm
 RUN pip install -v -r requirements-hpu.txt
 ENV no_proxy=localhost,127.0.0.1
 ENV PT_HPU_ENABLE_LAZY_COLLECTIVES=true
 RUN VLLM_TARGET_DEVICE=hpu python3 setup.py install
 # install development dependencies (for testing)
 RUN python3 -m pip install -e tests/vllm_test_utils
 WORKDIR /workspace/
 RUN ln -s /workspace/vllm/tests && ln -s /workspace/vllm/examples && ln -s /workspace/vllm/benchmarks
 ENTRYPOINT ["python3", "-m", "vllm.entrypoints.openai.api_server"]
--- a/Dockerfile.neuron
+++ b/Dockerfile.neuron
@@ -1,5 +1,6 @@
 # default base image
-ARG BASE_IMAGE="public.ecr.aws/neuron/pytorch-inference-neuronx:2.1.2-neuronx-py310-sdk2.20.0-ubuntu20.04"
+# https://gallery.ecr.aws/neuron/pytorch-inference-neuronx
 ARG BASE_IMAGE="public.ecr.aws/neuron/pytorch-inference-neuronx:2.5.1-neuronx-py310-sdk2.21.0-ubuntu22.04"
 FROM $BASE_IMAGE
@@ -14,16 +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 ${APP_MOUNT}/vllm
 RUN python3 -m pip install --upgrade pip
 RUN python3 -m pip install --no-cache-dir fastapi ninja tokenizers pandas
-RUN python3 -m pip install sentencepiece transformers==4.36.2 -U
+RUN python3 -m pip install sentencepiece transformers==4.45.2 -U
 RUN python3 -m pip install transformers-neuronx --extra-index-url=https://pip.repos.neuron.amazonaws.com -U
-RUN python3 -m pip install --pre neuronx-cc==2.15.* --extra-index-url=https://pip.repos.neuron.amazonaws.com -U
+RUN python3 -m pip install neuronx-cc==2.16.345.0 --extra-index-url=https://pip.repos.neuron.amazonaws.com -U
 RUN python3 -m pip install pytest
 COPY . .
 ARG GIT_REPO_CHECK=0
@@ -31,11 +33,17 @@ RUN --mount=type=bind,source=.git,target=.git \
    if [ "$GIT_REPO_CHECK" != 0 ]; then bash tools/check_repo.sh ; fi
 RUN python3 -m pip install -U \
-        cmake>=3.26 ninja packaging setuptools-scm>=8 wheel jinja2 \
+        'cmake>=3.26' ninja packaging 'setuptools-scm>=8' wheel jinja2 \
        -r requirements-neuron.txt
 ENV VLLM_TARGET_DEVICE neuron
 RUN --mount=type=bind,source=.git,target=.git \
-    pip install --no-build-isolation -v -e . \
+    pip install --no-build-isolation -v -e .
 # 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"]
--- a/Dockerfile.openvino
+++ b/Dockerfile.openvino
@@ -14,12 +14,16 @@ ARG GIT_REPO_CHECK=0
 RUN --mount=type=bind,source=.git,target=.git \
    if [ "$GIT_REPO_CHECK" != 0 ]; then bash tools/check_repo.sh ; fi
 RUN python3 -m pip install -U pip
 # install build requirements
-RUN PIP_EXTRA_INDEX_URL="https://download.pytorch.org/whl/cpu" python3 -m pip install -r /workspace/vllm/requirements-build.txt
+RUN PIP_EXTRA_INDEX_URL="https://download.pytorch.org/whl/cpu" python3 -m pip install -r /workspace/requirements-build.txt
 # build vLLM with OpenVINO backend
-RUN PIP_EXTRA_INDEX_URL="https://download.pytorch.org/whl/cpu" VLLM_TARGET_DEVICE="openvino" python3 -m pip install /workspace/vllm/
+RUN PIP_EXTRA_INDEX_URL="https://download.pytorch.org/whl/cpu" VLLM_TARGET_DEVICE="openvino" python3 -m pip install /workspace
-COPY examples/ /workspace/vllm/examples
+COPY examples/ /workspace/examples
-COPY benchmarks/ /workspace/vllm/benchmarks
+COPY benchmarks/ /workspace/benchmarks
 # install development dependencies (for testing)
 RUN python3 -m pip install -e tests/vllm_test_utils
 CMD ["/bin/bash"]
--- a/Dockerfile.ppc64le
+++ b/Dockerfile.ppc64le
@@ -4,12 +4,12 @@ USER root
 ENV PATH="/usr/local/cargo/bin:$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
@@ -18,19 +18,20 @@ 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 \
+        'cmake>=3.26' ninja packaging 'setuptools-scm>=8' wheel jinja2 \
        torch==2.3.1 \
        -r requirements-cpu.txt \
        xformers uvloop==0.20.0
 RUN --mount=type=bind,source=.git,target=.git \
    VLLM_TARGET_DEVICE=cpu python3 setup.py install
 # install development dependencies (for testing)
 RUN python3 -m pip install -e tests/vllm_test_utils
 WORKDIR /workspace/
 RUN ln -s /workspace/vllm/tests && ln -s /workspace/vllm/examples && ln -s /workspace/vllm/benchmarks
-ENTRYPOINT ["python3", "-m", "vllm.entrypoints.openai.api_server"]
+ENTRYPOINT ["/opt/conda/bin/python3", "-m", "vllm.entrypoints.openai.api_server"]
--- a/Dockerfile.rocm
+++ b/Dockerfile.rocm
@@ -1,169 +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
 WORKDIR ${COMMON_WORKDIR}
-# Install torch == 2.6.0 on ROCm
+
-RUN --mount=type=cache,target=/root/.cache/pip \
+# -----------------------
-    case "$(ls /opt | grep -Po 'rocm-[0-9]\.[0-9]')" in \
+# vLLM fetch stages
-        *"rocm-6.2"*) \
+FROM base AS fetch_vllm_0
-            python3 -m pip uninstall -y torch torchvision \
+ONBUILD COPY ./ vllm/
-            && python3 -m pip install --pre \
+FROM base AS fetch_vllm_1
-                torch==2.6.0.dev20240918 \
+ARG VLLM_REPO="https://github.com/vllm-project/vllm.git"
-                setuptools-scm>=8 \
+ARG VLLM_BRANCH="main"
-                torchvision==0.20.0.dev20240918 \
+ONBUILD RUN git clone ${VLLM_REPO} \
-                --extra-index-url https://download.pytorch.org/whl/nightly/rocm6.2;; \
+	    && cd vllm \
 	    && git checkout ${VLLM_BRANCH}
 FROM fetch_vllm_${REMOTE_VLLM} AS fetch_vllm
 # -----------------------
 # vLLM build stages
 FROM fetch_vllm AS build_vllm
 ARG USE_CYTHON
 # Build vLLM
 RUN cd vllm \
    && python3 -m pip install -r requirements-rocm.txt \
    && python3 setup.py clean --all  \
    && 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
 # -----------------------
 # Test vLLM image
 FROM base AS test
 RUN python3 -m pip install --upgrade pip && rm -rf /var/lib/apt/lists/*
 # 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
 WORKDIR /vllm-workspace
 ARG COMMON_WORKDIR
 COPY --from=build_vllm ${COMMON_WORKDIR}/vllm /vllm-workspace
 # install development dependencies (for testing)
 RUN cd /vllm-workspace \
    && rm -rf vllm \
    && python3 -m pip install -e tests/vllm_test_utils \
    && python3 -m pip install lm-eval[api]==0.4.4 \
    && python3 -m pip install pytest-shard
 # -----------------------
 # Final vLLM image
 FROM base AS final
 RUN python3 -m pip install --upgrade pip && rm -rf /var/lib/apt/lists/*
 # Error related to odd state for numpy 1.20.3 where there is no METADATA etc, but an extra LICENSES_bundled.txt.
 # Manually remove it so that later steps of numpy upgrade can continue
 RUN case "$(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
-ENV LLVM_SYMBOLIZER_PATH=/opt/rocm/llvm/bin/llvm-symbolizer
+RUN python3 -m pip install --upgrade huggingface-hub[cli]
-ENV PATH=$PATH:/opt/rocm/bin:/libtorch/bin:
+ARG BUILD_RPD
-ENV LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/opt/rocm/lib/:/libtorch/lib:
+RUN if [ ${BUILD_RPD} -eq "1" ]; then \
-ENV CPLUS_INCLUDE_PATH=$CPLUS_INCLUDE_PATH:/libtorch/include:/libtorch/include/torch/csrc/api/include/:/opt/rocm/include/:
+    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
-ENV PYTORCH_ROCM_ARCH=${PYTORCH_ROCM_ARCH}
+# Install vLLM
-ENV CCACHE_DIR=/root/.cache/ccache
+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
-### AMD-SMI build stage
+# Copy over the benchmark scripts as well
-FROM base AS build_amdsmi
+COPY --from=export_vllm /benchmarks ${COMMON_WORKDIR}/vllm/benchmarks
-# Build amdsmi wheel always
+COPY --from=export_vllm /examples ${COMMON_WORKDIR}/vllm/examples
 RUN cd /opt/rocm/share/amd_smi \
    && python3 -m pip wheel . --wheel-dir=/install
 ### Flash-Attention wheel build stage
 FROM base AS build_fa
 ARG BUILD_FA
 ARG FA_GFX_ARCHS
 ARG FA_BRANCH
 # Build ROCm flash-attention wheel if `BUILD_FA = 1`
 RUN --mount=type=cache,target=${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 [ "$GIT_REPO_CHECK" != 0 ]; then bash tools/check_repo.sh ; fi
 # Package upgrades for useful functionality or to avoid dependency issues
 RUN --mount=type=cache,target=/root/.cache/pip \
    python3 -m pip install --upgrade numba scipy huggingface-hub[cli] pytest-shard
 # Workaround for ray >= 2.10.0
 ENV RAY_EXPERIMENTAL_NOSET_ROCR_VISIBLE_DEVICES=1
 # Silences the HF Tokenizers warning
 ENV TOKENIZERS_PARALLELISM=false
-RUN --mount=type=cache,target=${CCACHE_DIR} \
+# Performance environment variable.
-    --mount=type=bind,source=.git,target=.git \
+ENV HIP_FORCE_DEV_KERNARG=1
    --mount=type=cache,target=/root/.cache/pip \
    python3 -m pip install -Ur requirements-rocm.txt \
    && python3 setup.py clean --all \
    && python3 setup.py develop
 # Copy amdsmi wheel into final image
 RUN --mount=type=bind,from=build_amdsmi,src=/install,target=/install \
    mkdir -p libs \
    && 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 --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
 RUN --mount=type=bind,from=build_fa,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 flash-attn; fi
 # Install wheels that were built to the final image
 RUN --mount=type=cache,target=/root/.cache/pip \
    if ls libs/*.whl; then \
    python3 -m pip install libs/*.whl; fi
 CMD ["/bin/bash"]
--- a/Dockerfile.rocm_base
+++ b/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: ${FA_REPO}" >> /app/versions.txt
--- a/Dockerfile.tpu
+++ b/Dockerfile.tpu
@@ -1,4 +1,4 @@
-ARG NIGHTLY_DATE="20240828"
+ARG NIGHTLY_DATE="20250124"
 ARG BASE_IMAGE="us-central1-docker.pkg.dev/tpu-pytorch-releases/docker/xla:nightly_3.10_tpuvm_$NIGHTLY_DATE"
 FROM $BASE_IMAGE
@@ -9,12 +9,6 @@ RUN apt-get update && apt-get install -y \
    git \
    ffmpeg libsm6 libxext6 libgl1
 # Install the TPU and Pallas dependencies.
 RUN --mount=type=cache,target=/root/.cache/pip \
    python3 -m pip install torch_xla[tpu] -f https://storage.googleapis.com/libtpu-releases/index.html
 RUN --mount=type=cache,target=/root/.cache/pip \
    python3 -m pip install torch_xla[pallas] -f https://storage.googleapis.com/jax-releases/jax_nightly_releases.html -f https://storage.googleapis.com/jax-releases/jaxlib_nightly_releases.html
 # Build vLLM.
 COPY . .
 ARG GIT_REPO_CHECK=0
@@ -25,8 +19,10 @@ ENV VLLM_TARGET_DEVICE="tpu"
 RUN --mount=type=cache,target=/root/.cache/pip \
    --mount=type=bind,source=.git,target=.git \
    python3 -m pip install \
        cmake>=3.26 ninja packaging setuptools-scm>=8 wheel jinja2 \
        -r requirements-tpu.txt
 RUN python3 setup.py develop
 # install development dependencies (for testing)
 RUN python3 -m pip install -e tests/vllm_test_utils
 CMD ["/bin/bash"]
--- a/Dockerfile.xpu
+++ b/Dockerfile.xpu
@@ -30,9 +30,19 @@ COPY requirements-common.txt /workspace/vllm/requirements-common.txt
 RUN --mount=type=cache,target=/root/.cache/pip \
    pip install --no-cache-dir \
    --extra-index-url https://pytorch-extension.intel.com/release-whl/stable/xpu/us/ \
    -r requirements-xpu.txt
 RUN git clone https://github.com/intel/pti-gpu && \
    cd pti-gpu/sdk && \
    git checkout 6c491f07a777ed872c2654ca9942f1d0dde0a082 && \
    mkdir build && \
    cd build && \
    cmake -DCMAKE_BUILD_TYPE=Release -DCMAKE_TOOLCHAIN_FILE=../cmake/toolchains/icpx_toolchain.cmake -DBUILD_TESTING=OFF .. && \
    make -j && \
    cmake --install . --config Release --prefix "/usr/local"
 ENV LD_LIBRARY_PATH="$LD_LIBRARY_PATH:/usr/local/lib/"
 COPY . .
 ARG GIT_REPO_CHECK
 RUN --mount=type=bind,source=.git,target=.git \
@@ -54,5 +64,6 @@ RUN --mount=type=cache,target=/root/.cache/pip \
 ENV VLLM_USAGE_SOURCE production-docker-image \
    TRITON_XPU_PROFILE 1
-
+# install development dependencies (for testing)
 RUN python3 -m pip install -e tests/vllm_test_utils
 ENTRYPOINT ["python3", "-m", "vllm.entrypoints.openai.api_server"]
--- a/README.md
+++ b/README.md
@@ -10,13 +10,18 @@ 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!
- [2024/10] Ray Summit 2024 held a special track for vLLM! Please find the opening talk slides from the vLLM team [here](https://docs.google.com/presentation/d/1B_KQxpHBTRa_mDF-tR6i8rWdOU5QoTZNcEg2MKZxEHM/edit?usp=sharing). Learn more from the [talks](https://raysummit.anyscale.com/flow/anyscale/raysummit2024/landing/page/sessioncatalog?tab.day=20241001&search.sessiontracks=1719251906298001uzJ2) from other vLLM contributors and users!
+- [2024/10] Ray Summit 2024 held a special track for vLLM! Please find the opening talk slides from the vLLM team [here](https://docs.google.com/presentation/d/1B_KQxpHBTRa_mDF-tR6i8rWdOU5QoTZNcEg2MKZxEHM/edit?usp=sharing). Learn more from the [talks](https://www.youtube.com/playlist?list=PLzTswPQNepXl6AQwifuwUImLPFRVpksjR) from other vLLM contributors and users!
 - [2024/09] We hosted [the sixth vLLM meetup](https://lu.ma/87q3nvnh) with NVIDIA! Please find the meetup slides [here](https://docs.google.com/presentation/d/1wrLGwytQfaOTd5wCGSPNhoaW3nq0E-9wqyP7ny93xRs/edit?usp=sharing).
 - [2024/07] We hosted [the fifth vLLM meetup](https://lu.ma/lp0gyjqr) with AWS! Please find the meetup slides [here](https://docs.google.com/presentation/d/1RgUD8aCfcHocghoP3zmXzck9vX3RCI9yfUAB2Bbcl4Y/edit?usp=sharing).
 - [2024/07] In partnership with Meta, vLLM officially supports Llama 3.1 with FP8 quantization and pipeline parallelism! Please check out our blog post [here](https://blog.vllm.ai/2024/07/23/llama31.html).
@@ -31,10 +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.
@@ -57,7 +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)
@@ -65,16 +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
@@ -87,27 +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.
@@ -126,7 +139,10 @@ 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.
 ## Media Kit
 * If you wish to use vLLM's logo, please refer to [our media kit repo](https://github.com/vllm-project/media-kit).
--- a/SECURITY.md
+++ b/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).
 ---
--- a/benchmarks/README.md
+++ b/benchmarks/README.md
@@ -6,3 +6,14 @@ You can download the dataset by running:
 ```bash
 wget https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/resolve/main/ShareGPT_V3_unfiltered_cleaned_split.json
 ```
 ## Downloading the ShareGPT4V dataset
 The json file refers to several image datasets (coco, llava, etc.). The benchmark scripts
 will ignore a datapoint if the referred image is missing.
 ```bash
 wget https://huggingface.co/datasets/Lin-Chen/ShareGPT4V/resolve/main/sharegpt4v_instruct_gpt4-vision_cap100k.json
 mkdir coco -p
 wget http://images.cocodataset.org/zips/train2017.zip -O coco/train2017.zip
 unzip coco/train2017.zip -d coco/
 ```
--- a/benchmarks/backend_request_func.py
+++ b/benchmarks/backend_request_func.py
@@ -1,3 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
 import json
 import os
 import sys
@@ -22,8 +24,10 @@ 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
    multi_modal_content: Optional[dict] = None
    ignore_eos: bool = False
@@ -33,9 +37,11 @@ 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
    tpot: float = 0.0  # avg next-token latencies
    prompt_len: int = 0
    error: str = ""
@@ -47,13 +53,15 @@ 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,
            "do_sample": True,
            "temperature": 0.01,  # TGI does not accept 0.0 temperature.
            "top_p": 0.99,  # TGI does not accept 1.0 top_p.
            "truncate": request_func_input.prompt_len,
            # TGI does not accept ignore_eos flag.
        }
        payload = {
@@ -75,11 +83,11 @@ async def async_request_tgi(
                            continue
                        chunk_bytes = chunk_bytes.decode("utf-8")
-                        #NOTE: Sometimes TGI returns a ping response without
+                        # NOTE: Sometimes TGI returns a ping response without
                        # any data, we should skip it.
                        if chunk_bytes.startswith(":"):
                            continue
-                        chunk = remove_prefix(chunk_bytes, "data:")
+                        chunk = chunk_bytes.removeprefix("data:")
                        data = json.loads(chunk)
                        timestamp = time.perf_counter()
@@ -118,7 +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,
@@ -144,15 +153,15 @@ async def async_request_trt_llm(
                        if not chunk_bytes:
                            continue
-                        chunk = remove_prefix(chunk_bytes.decode("utf-8"),
+                        chunk = chunk_bytes.decode("utf-8").removeprefix(
-                                              "data:")
+                            "data:")
                        data = json.loads(chunk)
                        output.generated_text += data["text_output"]
                        timestamp = time.perf_counter()
                        # First token
                        if ttft == 0.0:
-                            ttft = time.perf_counter() - st
+                            ttft = timestamp - st
                            output.ttft = ttft
                        # Decoding phase
@@ -182,7 +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 = {
@@ -230,17 +240,25 @@ 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 = {
            "Authorization": f"Bearer {os.environ.get('OPENAI_API_KEY')}"
        }
@@ -249,32 +267,34 @@ 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:
            async with session.post(url=api_url, json=payload,
                                    headers=headers) as response:
                if response.status == 200:
                    first_chunk_received = False
                    async for chunk_bytes in response.content:
                        chunk_bytes = chunk_bytes.strip()
                        if not chunk_bytes:
                            continue
-                        chunk = remove_prefix(chunk_bytes.decode("utf-8"),
+                        chunk = chunk_bytes.decode("utf-8").removeprefix(
-                                              "data: ")
+                            "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 ttft == 0.0:
+                                if not first_chunk_received:
                                    first_chunk_received = True
                                    ttft = time.perf_counter() - st
                                    output.ttft = ttft
@@ -284,11 +304,19 @@ 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:
                        output.success = False
                        output.error = (
                            "Never received a valid chunk to calculate TTFT."
                            "This response will be marked as failed!")
                    output.generated_text = generated_text
-                    output.success = True
+                    output.latency = most_recent_timestamp - st
                    output.latency = latency
                else:
                    output.error = response.reason or ""
                    output.success = False
@@ -311,12 +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",
@@ -324,10 +354,16 @@ async def async_request_openai_chat_completions(
                },
            ],
            "temperature": 0.0,
-            "max_tokens": request_func_input.output_len,
+            "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 = {
            "Content-Type": "application/json",
            "Authorization": f"Bearer {os.environ.get('OPENAI_API_KEY')}",
@@ -349,19 +385,17 @@ async def async_request_openai_chat_completions(
                        if not chunk_bytes:
                            continue
-                        chunk = remove_prefix(chunk_bytes.decode("utf-8"),
+                        chunk = chunk_bytes.decode("utf-8").removeprefix(
-                                              "data: ")
+                            "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
@@ -369,13 +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
@@ -389,14 +426,6 @@ async def async_request_openai_chat_completions(
    return output
 # Since vllm must support Python 3.8, we can't use str.removeprefix(prefix)
 # introduced in Python 3.9
 def remove_prefix(text: str, prefix: str) -> str:
    if text.startswith(prefix):
        return text[len(prefix):]
    return text
 def get_model(pretrained_model_name_or_path: str) -> str:
    if os.getenv('VLLM_USE_MODELSCOPE', 'False').lower() == 'true':
        from modelscope import snapshot_download
@@ -411,14 +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 = {
--- a/benchmarks/benchmark_guided.py
+++ b/benchmarks/benchmark_guided.py
@@ -0,0 +1,495 @@
 # SPDX-License-Identifier: Apache-2.0
 """Benchmark guided decoding throughput."""
 import argparse
 import dataclasses
 import json
 import os
 import random
 import time
 from typing import List
 import datasets
 import pandas as pd
 import uvloop
 from transformers import AutoTokenizer, PreTrainedTokenizerBase
 from vllm.engine.arg_utils import AsyncEngineArgs, EngineArgs
 from vllm.entrypoints.openai.api_server import (
    build_async_engine_client_from_engine_args)
 from vllm.sampling_params import GuidedDecodingParams
 from vllm.utils import FlexibleArgumentParser, merge_async_iterators
@dataclasses.dataclass
 class SampleRequest:
    """A class representing a single inference request for benchmarking.
    Attributes:
        prompt: The input text prompt for the model.
        multi_modal_data: Optional dictionary containing multi-modal data (e.g.
            images).
        prompt_len: The length of the prompt in tokens.
        expected_output_len: The expected length of the output in tokens.
    """
    prompt: str
    prompt_len: int
    expected_output_len: int
    schema: dict
    structure_type: str = 'json'
    completion: str = None
 def run_vllm(requests: List[SampleRequest],
             engine_args: EngineArgs,
             n: int,
             guided_decoding_rate: float = 1.0,
             warmup: bool = False) -> float:
    from vllm import LLM, SamplingParams
    llm = LLM(**vars(engine_args))
    # Add the requests to the engine.
    prompts: List[str] = []
    sampling_params: List[SamplingParams] = []
    # create a list containing random selected true or false
    guided_decoding_req_idx = random.sample(
        range(len(requests)), int(len(requests) * guided_decoding_rate))
    if warmup:
        print(">>>>> Running warmup prompt, for the first 5")
        # We setup the first 5 requests to warmup FSM
        # if using xgrammar dataset, we will skip warmup
        warmup_requests = requests[:5]
        for i, request in enumerate(warmup_requests):
            prompts.append(request.prompt)
            sampling_params.append(
                SamplingParams(
                    n=n,
                    temperature=1.0,
                    top_p=1.0,
                    ignore_eos=True,
                    max_tokens=request.expected_output_len,
                    guided_decoding=GuidedDecodingParams(json=request.schema)
                    if guided_decoding_rate > 0 else None,
                ))
        llm.generate(prompts, sampling_params, use_tqdm=False)
    print(">>>>> Benchmark started...")
    prompts = []
    sampling_params = []
    for i, request in enumerate(requests):
        prompts.append(request.prompt)
        sampling_params.append(
            SamplingParams(
                n=n,
                temperature=1.0,
                top_p=1.0,
                ignore_eos=True,
                max_tokens=request.expected_output_len,
                guided_decoding=GuidedDecodingParams(
                    **{request.structure_type: request.schema})
                if i in guided_decoding_req_idx else None,
            ))
    start = time.perf_counter()
    outputs = llm.generate(prompts, sampling_params, use_tqdm=False)
    ret = []
    for output, request in zip(outputs, requests):
        generated_text = output.outputs[0].text
        ret.append({
            "generated": generated_text,
            "expected": request.completion
        })
    end = time.perf_counter()
    return end - start, ret
 async def run_vllm_async(
        requests: List[SampleRequest],
        engine_args: AsyncEngineArgs,
        n: int,
        guided_decoding_rate: float = 1.0,
        warmup: bool = False,
        disable_frontend_multiprocessing: bool = False) -> float:
    from vllm import SamplingParams
    async with build_async_engine_client_from_engine_args(
            engine_args, disable_frontend_multiprocessing) as llm:
        # Add the requests to the engine.
        prompts: List[str] = []
        sampling_params: List[SamplingParams] = []
        guided_decoding_req_idx = random.sample(
            range(len(requests)), int(len(requests) * guided_decoding_rate))
        if warmup:
            print(">>>>>> Running warmup prompt, for the first 5")
            # We setup the first 5 requests to warmup FSM
            # if using xgrammar dataset, we will skip warmup
            warmup_requests = requests[:5]
            for i, request in enumerate(warmup_requests):
                prompts.append(request.prompt)
                sampling_params.append(
                    SamplingParams(
                        n=n,
                        temperature=1.0,
                        top_p=1.0,
                        ignore_eos=True,
                        max_tokens=request.expected_output_len,
                        guided_decoding=GuidedDecodingParams(
                            json=request.schema)
                        if guided_decoding_rate > 0 else None,
                    ))
            generators = []
            for i, (prompt, sp) in enumerate(zip(prompts, sampling_params)):
                generator = llm.generate(prompt, sp, request_id=f"test{i}")
                generators.append(generator)
            all_gens = merge_async_iterators(*generators)
            async for i, res in all_gens:
                pass
        print(">>>>> Benchmark started...")
        prompts = []
        sampling_params = []
        for i, request in enumerate(requests):
            prompts.append(request.prompt)
            sampling_params.append(
                SamplingParams(
                    n=n,
                    temperature=1.0,
                    top_p=1.0,
                    ignore_eos=True,
                    max_tokens=request.expected_output_len,
                    guided_decoding=GuidedDecodingParams(json=request.schema)
                    if i in guided_decoding_req_idx else None,
                ))
        generators = []
        start_time = []
        latencies = []
        start = time.perf_counter()
        for i, (prompt, sp) in enumerate(zip(prompts, sampling_params)):
            generator = llm.generate(prompt, sp, request_id=f"test{i}")
            generators.append(generator)
            start_time.append(time.perf_counter())
            latencies.append([])
        all_gens = merge_async_iterators(*generators)
        generated_texts = [''] * len(requests)
        async for i, res in all_gens:
            generated_texts[i] = res.outputs[0].text
            lat = time.perf_counter() - start_time[i]
            latencies[i].append(lat)
        ret = [{
            'generated': gt,
            'expected': req.completion
        } for gt, req in zip(generated_texts, requests)]
        end = time.perf_counter()
        first_latency = pd.Series([lat[0] * 1000 for lat in latencies])
        next_latency = pd.Series([(lat[-1] - lat[0]) / len(lat[1:]) * 1000
                                  for lat in latencies])
        return end - start, ret, (first_latency, next_latency)
 def sample_requests(tokenizer: PreTrainedTokenizerBase,
                    args: argparse.Namespace) -> List[SampleRequest]:
    if args.dataset == 'json':
        if args.json_schema_path is None:
            dir_path = os.path.dirname(os.path.realpath(__file__))
            args.json_schema_path = os.path.join(dir_path,
                                                 "structured_schemas",
                                                 "structured_schema_1.json")
        with open(args.json_schema_path) as f:
            schema = json.load(f)
        prompt = f"Generate an example of a user profile given the following schema: {json.dumps(schema)}"  # noqa: E501
        input_len = len(tokenizer(prompt).input_ids)
        print(f"Input length of the prompt: {input_len} tokens")
        requests = [
            SampleRequest(prompt=prompt,
                          prompt_len=input_len,
                          expected_output_len=args.output_len,
                          schema=schema,
                          structure_type=args.structure_type)
            for _ in range(args.num_prompts)
        ]
    elif args.dataset == "grammar":
        schema = """
            ?start: select_statement
            ?select_statement: "SELECT " column_list " FROM " table_name
            ?column_list: column_name ("," column_name)*
            ?table_name: identifier
            ?column_name: identifier
            ?identifier: /[a-zA-Z_][a-zA-Z0-9_]*/
        """
        prompt = "Generate an SQL query to show the 'username' \
            and 'email' from the 'users' table."
        input_len = len(tokenizer(prompt).input_ids)
        print(f"Input length of the prompt: {input_len} tokens")
        requests = [
            SampleRequest(prompt=prompt,
                          prompt_len=input_len,
                          expected_output_len=args.output_len,
                          schema=schema,
                          structure_type=args.structure_type)
            for _ in range(args.num_prompts)
        ]
    elif args.dataset == "regex":
        regex = r"\w+@\w+\.com\n"
        args.regex = regex
        prompt = "Generate an email address for Alan Turing, \
            who works in Enigma. End in .com and new line. \
                Example result: alan.turing@enigma.com\n"
        input_len = len(tokenizer(prompt).input_ids)
        print(f"Input length of the prompt: {input_len} tokens")
        requests = [
            SampleRequest(prompt=prompt,
                          prompt_len=input_len,
                          expected_output_len=args.output_len,
                          schema=regex,
                          structure_type=args.structure_type)
            for _ in range(args.num_prompts)
        ]
    elif args.dataset == "choice":
        choice = ["Positive", "Negative"]
        args.choice = choice
        prompt = "Classify this sentiment: vLLM is wonderful!"
        input_len = len(tokenizer(prompt).input_ids)
        print(f"Input length of the prompt: {input_len} tokens")
        requests = [
            SampleRequest(prompt=prompt,
                          prompt_len=input_len,
                          expected_output_len=args.output_len,
                          schema=choice,
                          structure_type=args.structure_type)
            for _ in range(args.num_prompts)
        ]
    elif args.dataset == "xgrammar_bench":
        args.warmup = False
        requests: List[SampleRequest] = []
        dataset = datasets.load_dataset("NousResearch/json-mode-eval",
                                        split="train")
        print(f"dataset has {len(dataset)} entries")
        len_dataset = len(dataset)
        for data_point_idx in range(args.num_prompts):
            idx = data_point_idx
            while idx >= len_dataset:
                idx -= len_dataset
            schema = dataset["schema"][idx]
            prompt = tokenizer.apply_chat_template(dataset["prompt"][idx],
                                                   tokenize=False)
            input_len = len(tokenizer(prompt).input_ids)
            completion = dataset["completion"][idx]
            requests.append(
                SampleRequest(prompt=prompt,
                              prompt_len=input_len,
                              expected_output_len=args.output_len,
                              schema=schema,
                              completion=completion))
    return requests
 def evaluate(ret, args):
    def _eval_correctness_json(expected, actual):
        # extract json string from string using regex
        import re
        actual = actual.replace('\n', '').replace(' ', '').strip()
        try:
            actual = re.search(r'\{.*\}', actual).group()
            actual = json.loads(actual)
        except Exception:
            return False
        return True
    def _eval_correctness_choice(expected, actual):
        return actual in args.choice
    def _eval_correctness_regex(expected, actual):
        import re
        return re.match(args.regex, actual) is not None
    def _eval_correctness(expected, actual):
        if args.structure_type == 'json':
            return _eval_correctness_json(expected, actual)
        elif args.structure_type == 'regex':
            return _eval_correctness_regex(expected, actual)
        elif args.structure_type == 'choice':
            return _eval_correctness_choice(expected, actual)
        else:
            return None
    scores = []
    for res in ret:
        score = _eval_correctness(res['expected'], res['generated'])
        res['correctness'] = score
        scores.append(score)
    not_none_scores = [score for score in scores if score is not None]
    return (sum(not_none_scores) / len(not_none_scores) *
            100) if len(not_none_scores) > 0 else None
 def main(args: argparse.Namespace):
    print(args)
    random.seed(args.seed)
    # async engine is working for 'regex', 'choice' and 'grammar'
    if args.dataset == 'grammar':
        args.structure_type = 'grammar'
        args.async_engine = False
    elif args.dataset == 'regex':
        args.structure_type = 'regex'
        args.async_engine = False
    elif args.dataset == 'choice':
        args.structure_type = 'choice'
        args.async_engine = False
    else:
        args.structure_type = 'json'
    if args.no_guided_decoding:
        args.guided_decoding_ratio = 0
    if args.save_results:
        result_file_name = f'{args.guided_decoding_ratio}guided'
        result_file_name += f"_{args.model.split('/')[-1]}"
        result_file_name += f"_{args.dataset}"
        result_file_name += f"_{args.num_prompts}"
        result_file_name += f"_out{args.output_len}"
        result_file_name += f"_async{args.async_engine}"
        result_file_name += f"_warmup{args.warmup}"
        result_file_name += f"_chunkedprefill{args.enable_chunked_prefill}"
        result_file_name += ".txt"
    else:
        result_file_name = None
    # Synthesize a prompt with the given input length.
    tokenizer = AutoTokenizer.from_pretrained(
        args.tokenizer, trust_remote_code=args.trust_remote_code)
    requests = sample_requests(tokenizer, args)
    if args.async_engine:
        engine_args = AsyncEngineArgs.from_cli_args(args)
        elapsed_time, ret, (first_latency, next_latency) = uvloop.run(
            run_vllm_async(requests, engine_args, args.n,
                           args.guided_decoding_ratio, args.warmup,
                           args.disable_frontend_multiprocessing))
    else:
        engine_args = EngineArgs.from_cli_args(args)
        elapsed_time, ret = run_vllm(requests, engine_args, args.n,
                                     args.guided_decoding_ratio, args.warmup)
        first_latency, next_latency = None, None
    score = evaluate(ret, args)
    total_num_tokens = sum(request.prompt_len + request.expected_output_len
                           for request in requests)
    total_output_tokens = sum(request.expected_output_len
                              for request in requests)
    if first_latency is not None:
        latency_breakdown = "\nFirst token latency(msecs):\n"
        latency_breakdown += f"{first_latency.describe()}"
        latency_breakdown += "\nNext token latency(msecs):\n"
        latency_breakdown += f"{next_latency.describe()}"
    print(
        f"Throughput: {len(requests) / elapsed_time:.2f} requests/s, "
        f"{total_num_tokens / elapsed_time:.2f} total tokens/s, "
        f"{total_output_tokens / elapsed_time:.2f} output tokens/s",
        f"Correct rate is {score} %",
        f"{latency_breakdown if first_latency is not None else ''}")
    # Output JSON results if specified
    if args.output_json or result_file_name:
        results = {
            "elapsed_time": elapsed_time,
            "num_requests": len(requests),
            "total_num_tokens": total_num_tokens,
            "total_output_tokens": total_output_tokens,
            "requests_per_second": len(requests) / elapsed_time,
            "tokens_per_second": f"{total_num_tokens / elapsed_time:.2f}",
            "output_tokens_per_second":
            f"{total_output_tokens / elapsed_time:.2f}",
            "correct_rate(%)": score
        }
        results = {"outputs": ret, **results}
        if first_latency is not None:
            results["first_token_latency(msecs)"] = first_latency.describe(
            ).to_dict()
            results["next_token_latency(msecs)"] = next_latency.describe(
            ).to_dict()
        if args.output_json:
            with open(args.output_json, "w") as f:
                json.dump(results, f, indent=4)
        elif result_file_name:
            with open(result_file_name, "w") as f:
                json.dump(results, f, indent=4)
 if __name__ == "__main__":
    parser = FlexibleArgumentParser(description="Benchmark guided decoding.")
    parser = AsyncEngineArgs.add_cli_args(parser)
    parser.add_argument("--output-len",
                        type=int,
                        default=512,
                        help="Output length for each request. Overrides the "
                        "output length from the dataset.")
    parser.add_argument(
        "--dataset",
        default='json',
        choices=['json', 'grammar', 'regex', 'choice', 'xgrammar_bench'])
    parser.add_argument("--json_schema_path",
                        type=str,
                        default=None,
                        help="Path to json schema.")
    parser.add_argument("--n",
                        type=int,
                        default=1,
                        help="Number of generated sequences per prompt.")
    parser.add_argument("--num-prompts",
                        type=int,
                        default=10,
                        help="Number of prompts to process.")
    parser.add_argument(
        '--output-json',
        type=str,
        default=None,
        help='Path to save the throughput results in JSON format.')
    parser.add_argument("--async-engine",
                        action='store_true',
                        default=False,
                        help="Use vLLM async engine rather than LLM class.")
    parser.add_argument("--no-guided-decoding",
                        action='store_true',
                        default=False,
                        help="Whether to disable JSON decoding or not.")
    parser.add_argument("--guided-decoding-ratio",
                        type=float,
                        default=1.0,
                        help="Ratio of Guided Decoding requests")
    parser.add_argument("--disable-frontend-multiprocessing",
                        action='store_true',
                        default=False,
                        help="Disable decoupled async engine frontend.")
    parser.add_argument("--warmup",
                        action="store_true",
                        default=False,
                        help="Run warmup prompts before benchmark.")
    parser.add_argument("--save-results",
                        action="store_true",
                        default=False,
                        help="save output results.")
    args = parser.parse_args()
    if args.tokenizer is None:
        args.tokenizer = args.model
    main(args)
--- a/benchmarks/benchmark_latency.py
+++ b/benchmarks/benchmark_latency.py
@@ -1,5 +1,7 @@
 # SPDX-License-Identifier: Apache-2.0
 """Benchmark the latency of processing a single batch of requests."""
 import argparse
 import dataclasses
 import json
 import time
 from pathlib import Path
@@ -10,43 +12,20 @@ import torch
 from tqdm import tqdm
 from vllm import LLM, SamplingParams
-from vllm.engine.arg_utils import DEVICE_OPTIONS, EngineArgs
+from vllm.engine.arg_utils import EngineArgs
 from vllm.inputs import PromptType
-from vllm.model_executor.layers.quantization import QUANTIZATION_METHODS
+from vllm.sampling_params import BeamSearchParams
 from vllm.utils import FlexibleArgumentParser
 def main(args: argparse.Namespace):
    print(args)
    engine_args = EngineArgs.from_cli_args(args)
    # NOTE(woosuk): If the request cannot be processed in a single batch,
    # the engine will automatically process the request in multiple batches.
-    llm = LLM(
+    llm = LLM(**dataclasses.asdict(engine_args))
        model=args.model,
        speculative_model=args.speculative_model,
        num_speculative_tokens=args.num_speculative_tokens,
        speculative_draft_tensor_parallel_size=\
            args.speculative_draft_tensor_parallel_size,
        tokenizer=args.tokenizer,
        quantization=args.quantization,
        tensor_parallel_size=args.tensor_parallel_size,
        trust_remote_code=args.trust_remote_code,
        dtype=args.dtype,
        max_model_len=args.max_model_len,
        enforce_eager=args.enforce_eager,
        kv_cache_dtype=args.kv_cache_dtype,
        quantization_param_path=args.quantization_param_path,
        device=args.device,
        ray_workers_use_nsight=args.ray_workers_use_nsight,
        enable_chunked_prefill=args.enable_chunked_prefill,
        download_dir=args.download_dir,
        block_size=args.block_size,
        gpu_memory_utilization=args.gpu_memory_utilization,
        load_format=args.load_format,
        distributed_executor_backend=args.distributed_executor_backend,
        otlp_traces_endpoint=args.otlp_traces_endpoint,
        enable_prefix_caching=args.enable_prefix_caching,
    )
    sampling_params = SamplingParams(
        n=args.n,
@@ -63,6 +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(
@@ -72,15 +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
@@ -125,19 +114,6 @@ if __name__ == '__main__':
    parser = FlexibleArgumentParser(
        description='Benchmark the latency of processing a single batch of '
        'requests till completion.')
    parser.add_argument('--model', type=str, default='facebook/opt-125m')
    parser.add_argument('--speculative-model', type=str, default=None)
    parser.add_argument('--num-speculative-tokens', type=int, default=None)
    parser.add_argument('--speculative-draft-tensor-parallel-size',
                        '-spec-draft-tp',
                        type=int,
                        default=None)
    parser.add_argument('--tokenizer', type=str, default=None)
    parser.add_argument('--quantization',
                        '-q',
                        choices=[*QUANTIZATION_METHODS, None],
                        default=None)
    parser.add_argument('--tensor-parallel-size', '-tp', type=int, default=1)
    parser.add_argument('--input-len', type=int, default=32)
    parser.add_argument('--output-len', type=int, default=128)
    parser.add_argument('--batch-size', type=int, default=8)
@@ -154,45 +130,6 @@ if __name__ == '__main__':
                        type=int,
                        default=30,
                        help='Number of iterations to run.')
    parser.add_argument('--trust-remote-code',
                        action='store_true',
                        help='trust remote code from huggingface')
    parser.add_argument(
        '--max-model-len',
        type=int,
        default=None,
        help='Maximum length of a sequence (including prompt and output). '
        'If None, will be derived from the model.')
    parser.add_argument(
        '--dtype',
        type=str,
        default='auto',
        choices=['auto', 'half', 'float16', 'bfloat16', 'float', 'float32'],
        help='data type for model weights and activations. '
        'The "auto" option will use FP16 precision '
        'for FP32 and FP16 models, and BF16 precision '
        'for BF16 models.')
    parser.add_argument('--enforce-eager',
                        action='store_true',
                        help='enforce eager mode and disable CUDA graph')
    parser.add_argument(
        '--kv-cache-dtype',
        type=str,
        choices=['auto', 'fp8', 'fp8_e5m2', 'fp8_e4m3'],
        default="auto",
        help='Data type for kv cache storage. If "auto", will use model '
        'data type. CUDA 11.8+ supports fp8 (=fp8_e4m3) and fp8_e5m2. '
        'ROCm (AMD GPU) supports fp8 (=fp8_e4m3)')
    parser.add_argument(
        '--quantization-param-path',
        type=str,
        default=None,
        help='Path to the JSON file containing the KV cache scaling factors. '
        'This should generally be supplied, when KV cache dtype is FP8. '
        'Otherwise, KV cache scaling factors default to 1.0, which may cause '
        'accuracy issues. FP8_E5M2 (without scaling) is only supported on '
        'cuda version greater than 11.8. On ROCm (AMD GPU), FP8_E4M3 is '
        'instead supported for common inference criteria.')
    parser.add_argument(
        '--profile',
        action='store_true',
@@ -203,78 +140,12 @@ if __name__ == '__main__':
        default=None,
        help=('path to save the pytorch profiler output. Can be visualized '
              'with ui.perfetto.dev or Tensorboard.'))
    parser.add_argument("--device",
                        type=str,
                        default="auto",
                        choices=DEVICE_OPTIONS,
                        help='device type for vLLM execution')
    parser.add_argument('--block-size',
                        type=int,
                        default=16,
                        help='block size of key/value cache')
    parser.add_argument(
        '--enable-chunked-prefill',
        action='store_true',
        help='If True, the prefill requests can be chunked based on the '
        'max_num_batched_tokens')
    parser.add_argument("--enable-prefix-caching",
                        action='store_true',
                        help="Enable automatic prefix caching")
    parser.add_argument(
        "--ray-workers-use-nsight",
        action='store_true',
        help="If specified, use nsight to profile ray workers",
    )
    parser.add_argument('--download-dir',
                        type=str,
                        default=None,
                        help='directory to download and load the weights, '
                        'default to the default cache dir of huggingface')
    parser.add_argument(
        '--output-json',
        type=str,
        default=None,
        help='Path to save the latency results in JSON format.')
-    parser.add_argument('--gpu-memory-utilization',
+
-                        type=float,
+    parser = EngineArgs.add_cli_args(parser)
                        default=0.9,
                        help='the fraction of GPU memory to be used for '
                        'the model executor, which can range from 0 to 1.'
                        'If unspecified, will use the default value of 0.9.')
    parser.add_argument(
        '--load-format',
        type=str,
        default=EngineArgs.load_format,
        choices=[
            'auto', 'pt', 'safetensors', 'npcache', 'dummy', 'tensorizer',
            'bitsandbytes'
        ],
        help='The format of the model weights to load.\n\n'
        '* "auto" will try to load the weights in the safetensors format '
        'and fall back to the pytorch bin format if safetensors format '
        'is not available.\n'
        '* "pt" will load the weights in the pytorch bin format.\n'
        '* "safetensors" will load the weights in the safetensors format.\n'
        '* "npcache" will load the weights in pytorch format and store '
        'a numpy cache to speed up the loading.\n'
        '* "dummy" will initialize the weights with random values, '
        'which is mainly for profiling.\n'
        '* "tensorizer" will load the weights using tensorizer from '
        'CoreWeave. See the Tensorize vLLM Model script in the Examples'
        'section for more information.\n'
        '* "bitsandbytes" will load the weights using bitsandbytes '
        'quantization.\n')
    parser.add_argument(
        '--distributed-executor-backend',
        choices=['ray', 'mp'],
        default=None,
        help='Backend to use for distributed serving. When more than 1 GPU '
        'is used, will be automatically set to "ray" if installed '
        'or "mp" (multiprocessing) otherwise.')
    parser.add_argument(
        '--otlp-traces-endpoint',
        type=str,
        default=None,
        help='Target URL to which OpenTelemetry traces will be sent.')
    args = parser.parse_args()
    main(args)
--- a/benchmarks/benchmark_long_document_qa_throughput.py
+++ b/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)
--- a/benchmarks/benchmark_prefix_caching.py
+++ b/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
@@ -25,6 +27,7 @@ ShareGPT example usage:
        --input-length-range 128:256
 """
 import dataclasses
 import json
 import random
 import time
@@ -33,6 +36,7 @@ from typing import List, Optional, Tuple
 from transformers import PreTrainedTokenizerBase
 from vllm import LLM, SamplingParams
 from vllm.engine.arg_utils import EngineArgs
 from vllm.utils import FlexibleArgumentParser
 try:
@@ -52,13 +56,30 @@ def test_prefix(llm=None, sampling_params=None, prompts=None):
    print(f"cost time {end_time - start_time}")
-def sample_requests(
+@dataclasses.dataclass
 class Request:
    prompt: str
    prompt_len: int
    output_len: int
 def sample_tokens(tokenizer: PreTrainedTokenizerBase, length: int) -> str:
    vocab = tokenizer.get_vocab()
    # Remove the special tokens.
    vocab = {
        k: v
        for k, v in vocab.items() if k not in tokenizer.all_special_ids
    }
    return random.choices(list(vocab.values()), k=length)
 def sample_requests_from_dataset(
    dataset_path: str,
    num_requests: int,
    tokenizer: PreTrainedTokenizerBase,
    input_length_range: Tuple[int, int],
    fixed_output_len: Optional[int],
-) -> List[Tuple[str, int, int]]:
+) -> List[Request]:
    if fixed_output_len is not None and fixed_output_len < 4:
        raise ValueError("output_len too small")
@@ -75,31 +96,55 @@ def sample_requests(
    random.shuffle(dataset)
    min_len, max_len = input_length_range
    assert min_len >= 0 and max_len >= min_len, "input_length_range too small"
    # Filter out sequences that are too long or too short
-    filtered_dataset: List[Tuple[str, int, int]] = []
+    filtered_requests: List[Request] = []
    for i in range(len(dataset)):
-        if len(filtered_dataset) == num_requests:
+        if len(filtered_requests) == num_requests:
            break
        # Tokenize the prompts and completions.
-        prompt = dataset[i][0]
+        prompt_token_ids = tokenizer(dataset[i][0]).input_ids
-        prompt_token_ids = tokenizer(prompt).input_ids
+        prompt = tokenizer.decode(prompt_token_ids)
        completion = dataset[i][1]
        completion_token_ids = tokenizer(completion).input_ids
        prompt_len = len(prompt_token_ids)
-        output_len = len(completion_token_ids
+        output_len = (len(completion_token_ids)
-                         ) if fixed_output_len is None else fixed_output_len
+                      if fixed_output_len is None else fixed_output_len)
        if prompt_len < 4 or output_len < 4:
            # Prune too short sequences.
            continue
        if min_len <= prompt_len <= max_len:
-            filtered_dataset.append((prompt, prompt_len, output_len))
+            filtered_requests.append(Request(prompt, prompt_len, output_len))
-    return filtered_dataset
+    return filtered_requests
-def repeat_and_sort_requests(requests: List[Tuple[str, int, int]],
+def sample_requests_from_random(
    num_requests: int,
    tokenizer: PreTrainedTokenizerBase,
    input_length_range: Tuple[int, int],
    fixed_output_len: Optional[int],
    prefix_len: int,
 ) -> List[Request]:
    requests = []
    prefix_token_ids = sample_tokens(tokenizer, prefix_len)
    min_len, max_len = input_length_range
    for i in range(num_requests):
        unique_part_token_ids = sample_tokens(
            tokenizer,
            random.randint(min_len - prefix_len, max_len - prefix_len))
        prompt_token_ids = prefix_token_ids + unique_part_token_ids
        prompt = tokenizer.decode(prompt_token_ids)
        prompt_len = len(prompt_token_ids)
        assert (min_len <= prompt_len <= max_len
                ), f"prompt_len {prompt_len} out of range {min_len}:{max_len}"
        requests.append(Request(prompt, prompt_len, fixed_output_len))
    return requests
 def repeat_and_sort_requests(requests: List[Request],
                             repeat_count: int,
                             sort: bool = False) -> List[str]:
    repeated_requests = requests * repeat_count
@@ -107,7 +152,7 @@ def repeat_and_sort_requests(requests: List[Tuple[str, int, int]],
        repeated_requests.sort(key=lambda x: x[1])
    else:
        random.shuffle(repeated_requests)
-    return [req[0] for req in repeated_requests]
+    return [req.prompt for req in repeated_requests]
 def main(args):
@@ -115,9 +160,12 @@ def main(args):
    input_length_range = tuple(map(int, args.input_length_range.split(':')))
    random.seed(args.seed)
    if args.dataset_path is not None:
-        print(f"Start to sample {args.num_prompts} prompts"
+        if args.prefix_len > 0:
-              "from {args.dataset_path}")
+            raise ValueError("prefix-len is not supported when "
-        filtered_datasets = sample_requests(
+                             "dataset-path is provided.")
        print(f"Start to sample {args.num_prompts} prompts "
              f"from {args.dataset_path}")
        filtered_requests = sample_requests_from_dataset(
            dataset_path=args.dataset_path,
            num_requests=args.num_prompts,
            tokenizer=tokenizer,
@@ -125,31 +173,34 @@ def main(args):
            fixed_output_len=args.output_len,
        )
    else:
-        prompt_len = len(tokenizer(PROMPT).input_ids)
+        print(f"Start to sample {args.num_prompts} prompts from random")
-        filtered_datasets = [(PROMPT, prompt_len, args.output_len)
+        filtered_requests = sample_requests_from_random(
-                             ] * args.num_prompts
+            num_requests=args.num_prompts,
            tokenizer=tokenizer,
            input_length_range=input_length_range,
            fixed_output_len=args.output_len,
            prefix_len=args.prefix_len,
        )
-    llm = LLM(model=args.model,
+    # Print some helpful stats of the requests.
-              tokenizer_mode='auto',
+    print(f"Sampled {len(filtered_requests)} requests.")
-              trust_remote_code=True,
+    prompt_lens = [req.prompt_len for req in filtered_requests]
-              enforce_eager=True,
+    print(f"Average input length: {sum(prompt_lens) / len(prompt_lens)}")
-              tensor_parallel_size=args.tensor_parallel_size,
+    print(f"P50 input length: {sorted(prompt_lens)[len(prompt_lens) // 2]}")
-              enable_prefix_caching=args.enable_prefix_caching)
+    print(f"Min Prompt Length: {min(prompt_lens)}")
    print(f"Max Prompt Length: {max(prompt_lens)}")
    engine_args = EngineArgs.from_cli_args(args)
    llm = LLM(**dataclasses.asdict(engine_args))
    sampling_params = SamplingParams(temperature=0, max_tokens=args.output_len)
-    print("Testing filtered datasets")
+    print("Testing filtered requests")
-    prompts = repeat_and_sort_requests(filtered_datasets,
+    prompts = repeat_and_sort_requests(filtered_requests,
                                       repeat_count=args.repeat_count,
                                       sort=args.sort)
    print("------warm up------")
    test_prefix(
        llm=llm,
        prompts=prompts,
        sampling_params=sampling_params,
    )
    print("------start generating------")
    test_prefix(
        llm=llm,
@@ -162,37 +213,37 @@ if __name__ == "__main__":
    parser = FlexibleArgumentParser(
        description=
        'Benchmark the performance with or without automatic prefix caching.')
    parser.add_argument('--model',
                        type=str,
                        default='baichuan-inc/Baichuan2-13B-Chat')
    parser.add_argument("--dataset-path",
                        type=str,
                        default=None,
                        help="Path to the dataset.")
    parser.add_argument('--tensor-parallel-size', '-tp', type=int, default=1)
    parser.add_argument('--output-len', type=int, default=10)
    parser.add_argument('--enable-prefix-caching',
                        action='store_true',
                        help='enable prefix caching')
    parser.add_argument('--num-prompts',
                        type=int,
-                        default=1,
+                        required=True,
                        help="Number of the prompts sampled from dataset")
    parser.add_argument('--repeat-count',
                        type=int,
-                        default=100,
+                        default=1,
                        help='Number of times to repeat each prompt')
    parser.add_argument('--sort',
                        action='store_true',
                        help='Sort prompts by input length')
    parser.add_argument('--input-length-range',
                        type=str,
-                        default='128:256',
+                        required=True,
                        help='Range of input lengths for sampling prompts,'
                        'specified as "min:max" (e.g., "128:256").')
-    parser.add_argument("--seed",
+    parser.add_argument(
-                        type=int,
+        "--prefix-len",
-                        default=0,
+        type=int,
-                        help='Random seed for reproducibility')
+        default=0,
        help="Specifies the length of a common prefix to be "
        "added to the input prompt. The input-length-range will "
        "subtract this length when filtering prompts. Only used "
        "when dataset-path is not provided.",
    )
    parser = EngineArgs.add_cli_args(parser)
    args = parser.parse_args()
    main(args)
--- a/benchmarks/benchmark_prioritization.py
+++ b/benchmarks/benchmark_prioritization.py
@@ -1,5 +1,7 @@
 # SPDX-License-Identifier: Apache-2.0
 """Benchmark offline prioritization."""
 import argparse
 import dataclasses
 import json
 import random
 import time
@@ -7,7 +9,8 @@ from typing import List, Optional, Tuple
 from transformers import AutoTokenizer, PreTrainedTokenizerBase
-from vllm.model_executor.layers.quantization import QUANTIZATION_METHODS
+from vllm.engine.arg_utils import EngineArgs
 from vllm.utils import FlexibleArgumentParser
 def sample_requests(
@@ -62,46 +65,11 @@ def sample_requests(
 def run_vllm(
    requests: List[Tuple[str, int, int]],
    model: str,
    tokenizer: str,
    quantization: Optional[str],
    tensor_parallel_size: int,
    seed: int,
    n: int,
-    trust_remote_code: bool,
+    engine_args: EngineArgs,
    dtype: str,
    max_model_len: Optional[int],
    enforce_eager: bool,
    kv_cache_dtype: str,
    quantization_param_path: Optional[str],
    device: str,
    enable_prefix_caching: bool,
    enable_chunked_prefill: bool,
    max_num_batched_tokens: int,
    gpu_memory_utilization: float = 0.9,
    download_dir: Optional[str] = None,
 ) -> float:
    from vllm import LLM, SamplingParams
-    llm = LLM(
+    llm = LLM(**dataclasses.asdict(engine_args))
        model=model,
        tokenizer=tokenizer,
        quantization=quantization,
        tensor_parallel_size=tensor_parallel_size,
        seed=seed,
        trust_remote_code=trust_remote_code,
        dtype=dtype,
        max_model_len=max_model_len,
        gpu_memory_utilization=gpu_memory_utilization,
        enforce_eager=enforce_eager,
        kv_cache_dtype=kv_cache_dtype,
        quantization_param_path=quantization_param_path,
        device=device,
        enable_prefix_caching=enable_prefix_caching,
        download_dir=download_dir,
        enable_chunked_prefill=enable_chunked_prefill,
        max_num_batched_tokens=max_num_batched_tokens,
        disable_log_stats=False,
    )
    # Add the requests to the engine.
    prompts = []
@@ -142,16 +110,8 @@ def main(args: argparse.Namespace):
                                   args.output_len)
    if args.backend == "vllm":
-        elapsed_time = run_vllm(requests, args.model, args.tokenizer,
+        elapsed_time = run_vllm(requests, args.n,
-                                args.quantization, args.tensor_parallel_size,
+                                EngineArgs.from_cli_args(args))
                                args.seed, args.n, args.trust_remote_code,
                                args.dtype, args.max_model_len,
                                args.enforce_eager, args.kv_cache_dtype,
                                args.quantization_param_path, args.device,
                                args.enable_prefix_caching,
                                args.enable_chunked_prefill,
                                args.max_num_batched_tokens,
                                args.gpu_memory_utilization, args.download_dir)
    else:
        raise ValueError(f"Unknown backend: {args.backend}")
    total_num_tokens = sum(prompt_len + output_len
@@ -173,7 +133,7 @@ def main(args: argparse.Namespace):
 if __name__ == "__main__":
-    parser = argparse.ArgumentParser(description="Benchmark the throughput.")
+    parser = FlexibleArgumentParser(description="Benchmark the throughput.")
    parser.add_argument("--backend",
                        type=str,
                        choices=["vllm", "hf", "mii"],
@@ -191,13 +151,6 @@ if __name__ == "__main__":
                        default=None,
                        help="Output length for each request. Overrides the "
                        "output length from the dataset.")
    parser.add_argument("--model", type=str, default="facebook/opt-125m")
    parser.add_argument("--tokenizer", type=str, default=None)
    parser.add_argument('--quantization',
                        '-q',
                        choices=[*QUANTIZATION_METHODS, None],
                        default=None)
    parser.add_argument("--tensor-parallel-size", "-tp", type=int, default=1)
    parser.add_argument("--n",
                        type=int,
                        default=1,
@@ -206,81 +159,13 @@ if __name__ == "__main__":
                        type=int,
                        default=200,
                        help="Number of prompts to process.")
    parser.add_argument("--seed", type=int, default=0)
    parser.add_argument('--trust-remote-code',
                        action='store_true',
                        help='trust remote code from huggingface')
    parser.add_argument(
        '--max-model-len',
        type=int,
        default=None,
        help='Maximum length of a sequence (including prompt and output). '
        'If None, will be derived from the model.')
    parser.add_argument(
        '--dtype',
        type=str,
        default='auto',
        choices=['auto', 'half', 'float16', 'bfloat16', 'float', 'float32'],
        help='data type for model weights and activations. '
        'The "auto" option will use FP16 precision '
        'for FP32 and FP16 models, and BF16 precision '
        'for BF16 models.')
    parser.add_argument('--gpu-memory-utilization',
                        type=float,
                        default=0.9,
                        help='the fraction of GPU memory to be used for '
                        'the model executor, which can range from 0 to 1.'
                        'If unspecified, will use the default value of 0.9.')
    parser.add_argument("--enforce-eager",
                        action="store_true",
                        help="enforce eager execution")
    parser.add_argument(
        '--kv-cache-dtype',
        type=str,
        choices=['auto', 'fp8', 'fp8_e5m2', 'fp8_e4m3'],
        default="auto",
        help='Data type for kv cache storage. If "auto", will use model '
        'data type. CUDA 11.8+ supports fp8 (=fp8_e4m3) and fp8_e5m2. '
        'ROCm (AMD GPU) supports fp8 (=fp8_e4m3)')
    parser.add_argument(
        '--quantization-param-path',
        type=str,
        default=None,
        help='Path to the JSON file containing the KV cache scaling factors. '
        'This should generally be supplied, when KV cache dtype is FP8. '
        'Otherwise, KV cache scaling factors default to 1.0, which may cause '
        'accuracy issues. FP8_E5M2 (without scaling) is only supported on '
        'cuda version greater than 11.8. On ROCm (AMD GPU), FP8_E4M3 is '
        'instead supported for common inference criteria.')
    parser.add_argument(
        "--device",
        type=str,
        default="cuda",
        choices=["cuda", "cpu"],
        help='device type for vLLM execution, supporting CUDA and CPU.')
    parser.add_argument(
        "--enable-prefix-caching",
        action='store_true',
        help="enable automatic prefix caching for vLLM backend.")
    parser.add_argument("--enable-chunked-prefill",
                        action='store_true',
                        help="enable chunked prefill for vLLM backend.")
    parser.add_argument('--max-num-batched-tokens',
                        type=int,
                        default=None,
                        help='maximum number of batched tokens per '
                        'iteration')
    parser.add_argument('--download-dir',
                        type=str,
                        default=None,
                        help='directory to download and load the weights, '
                        'default to the default cache dir of huggingface')
    parser.add_argument(
        '--output-json',
        type=str,
        default=None,
        help='Path to save the throughput results in JSON format.')
    parser = EngineArgs.add_cli_args(parser)
    args = parser.parse_args()
    if args.tokenizer is None:
        args.tokenizer = args.model
--- a/benchmarks/benchmark_serving.py
+++ b/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
@@ -53,6 +55,8 @@ try:
 except ImportError:
    from argparse import ArgumentParser as FlexibleArgumentParser
 MILLISECONDS_TO_SECONDS_CONVERSION = 1000
@dataclass
 class BenchmarkMetrics:
@@ -60,6 +64,7 @@ class BenchmarkMetrics:
    total_input: int
    total_output: int
    request_throughput: float
    request_goodput: float
    output_throughput: float
    total_token_throughput: float
    mean_ttft_ms: float
@@ -196,22 +201,80 @@ def sample_sonnet_requests(
    return sampled_requests
-def sample_hf_requests(
+def sample_vision_arena_requests(
-    dataset_path: str,
+    dataset,
    dataset_subset: str,
    dataset_split: str,
    num_requests: int,
    tokenizer: PreTrainedTokenizerBase,
    fixed_output_len: Optional[int] = None,
 ) -> List[Tuple[str, str, int, Optional[Dict[str, Collection[str]]]]]:
    sampled_requests: List[Tuple[str, int, int, Dict[str,
                                                     Collection[str]]]] = []
    for data in dataset:
        if len(sampled_requests) == num_requests:
            break
        prompt = data["turns"][0][0]['content']
        prompt_token_ids = tokenizer(prompt).input_ids
        if fixed_output_len is None:
            # Default max output len is set to 128
            print("--hf-output-len is not provided. Using default value 128.")
            fixed_output_len = 128
        prompt_len = len(prompt_token_ids)
        output_len = fixed_output_len
        assert isinstance(
            data["images"][0],
            Image), ("Input image format must be `PIL.Image.Image`, "
                     f"given {type(data['image'])}.")
        image: Image = data["images"][0]
        image = image.convert("RGB")
        image_data = io.BytesIO()
        image.save(image_data, format='JPEG')
        image_base64 = base64.b64encode(image_data.getvalue()).decode("utf-8")
        mm_content = {
            "type": "image_url",
            "image_url": {
                "url": f"data:image/jpeg;base64,{image_base64}"
            },
        }
        sampled_requests.append((prompt, prompt_len, output_len, mm_content))
    return sampled_requests
 def sample_hf_requests(
    dataset_path: str,
    dataset_subset: Optional[str],
    dataset_split: str,
    num_requests: int,
    tokenizer: PreTrainedTokenizerBase,
    random_seed: int,
    fixed_output_len: Optional[int] = None,
 ) -> List[Tuple[str, str, int, Optional[Dict[str, Collection[str]]]]]:
    # Special case for vision_arena dataset
    if dataset_path == 'lmarena-ai/vision-arena-bench-v0.1' \
        and dataset_subset is None:
        assert dataset_split == "train"
        dataset = load_dataset(dataset_path,
                               name=dataset_subset,
                               split=dataset_split,
                               streaming=True)
        dataset = dataset.shuffle(seed=random_seed)
        return sample_vision_arena_requests(dataset, num_requests, tokenizer,
                                            fixed_output_len)
    dataset = load_dataset(dataset_path,
                           name=dataset_subset,
                           split=dataset_split,
                           streaming=True)
    assert "conversations" in dataset.features, (
        "HF Dataset must have 'conversations' column.")
-    filtered_dataset = dataset.shuffle().filter(
+    filter_func = lambda x: len(x["conversations"]) >= 2
-        lambda x: len(x["conversations"]) >= 2)
+    filtered_dataset = dataset.shuffle(seed=random_seed).filter(filter_func)
    sampled_requests: List[Tuple[str, int, int, Dict[str,
                                                     Collection[str]]]] = []
    for data in filtered_dataset:
@@ -247,6 +310,19 @@ def sample_hf_requests(
                    "url": f"data:image/jpeg;base64,{image_base64}"
                },
            }
        elif "image" in data and isinstance(data["image"], str):
            if (data["image"].startswith("http://") or \
                data["image"].startswith("file://")):
                image_url = data["image"]
            else:
                image_url = f"file://{data['image']}"
            mm_content = {
                "type": "image_url",
                "image_url": {
                    "url": image_url
                },
            }
        else:
            mm_content = None
@@ -293,8 +369,33 @@ def sample_random_requests(
 async def get_request(
    input_requests: List[Tuple[str, int, int]],
    request_rate: float,
    burstiness: float = 1.0,
 ) -> AsyncGenerator[Tuple[str, int, int], None]:
    """
    Asynchronously generates requests at a specified rate 
    with OPTIONAL burstiness.
    Args:
        input_requests: 
            A list of input requests, each represented as a tuple.
        request_rate: 
            The rate at which requests are generated (requests/s).
        burstiness (optional): 
            The burstiness factor of the request generation. 
            Only takes effect when request_rate is not inf.
            Default value is 1, which follows a Poisson process.
            Otherwise, the request intervals follow a gamma distribution.
            A lower burstiness value (0 < burstiness < 1) results 
            in more bursty requests, while a higher burstiness value 
            (burstiness > 1) results in a more uniform arrival of requests.
    """
    input_requests = iter(input_requests)
    # Calculate scale parameter theta to maintain the desired request_rate.
    assert burstiness > 0, (
        f"A positive burstiness factor is expected, but given {burstiness}.")
    theta = 1.0 / (request_rate * burstiness)
    for request in input_requests:
        yield request
@@ -302,8 +403,9 @@ async def get_request(
            # If the request rate is infinity, then we don't need to wait.
            continue
-        # Sample the request interval from the exponential distribution.
+        # Sample the request interval from the gamma distribution.
-        interval = np.random.exponential(1.0 / request_rate)
+        # If burstiness is 1, it follows exponential distribution.
        interval = np.random.gamma(shape=burstiness, scale=theta)
        # The next request will be sent after the interval.
        await asyncio.sleep(interval)
@@ -315,28 +417,39 @@ def calculate_metrics(
    tokenizer: PreTrainedTokenizerBase,
    selected_percentile_metrics: List[str],
    selected_percentiles: List[float],
    goodput_config_dict: Dict[str, float],
 ) -> Tuple[BenchmarkMetrics, List[int]]:
    actual_output_lens: List[int] = []
    total_input = 0
    completed = 0
    good_completed = 0
    itls: List[float] = []
    tpots: List[float] = []
    all_tpots: List[float] = []
    ttfts: List[float] = []
    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:
-            # Note : this may inflate the output token count slightly
+                # We use the tokenizer to count the number of output tokens
-            output_len = len(
+                # for some serving backends instead of looking at
-                tokenizer(outputs[i].generated_text,
+                # len(outputs[i].itl) since multiple output tokens may be
-                          add_special_tokens=False).input_ids)
+                # bundled together
                # Note : this may inflate the output token count slightly
                output_len = len(
                    tokenizer(outputs[i].generated_text,
                              add_special_tokens=False).input_ids)
            actual_output_lens.append(output_len)
            total_input += input_requests[i][1]
            tpot = 0
            if output_len > 1:
-                tpots.append(
+                latency_minus_ttft = outputs[i].latency - outputs[i].ttft
-                    (outputs[i].latency - outputs[i].ttft) / (output_len - 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)
            itls += outputs[i].itl
            ttfts.append(outputs[i].ttft)
            e2els.append(outputs[i].latency)
@@ -344,6 +457,28 @@ def calculate_metrics(
        else:
            actual_output_lens.append(0)
    if goodput_config_dict:
        valid_metrics = []
        slo_values = []
        if "ttft" in goodput_config_dict:
            valid_metrics.append(ttfts)
            slo_values.append(goodput_config_dict["ttft"] /
                              MILLISECONDS_TO_SECONDS_CONVERSION)
        if "tpot" in goodput_config_dict:
            valid_metrics.append(all_tpots)
            slo_values.append(goodput_config_dict["tpot"] /
                              MILLISECONDS_TO_SECONDS_CONVERSION)
        if "e2el" in goodput_config_dict:
            valid_metrics.append(e2els)
            slo_values.append(goodput_config_dict["e2el"] /
                              MILLISECONDS_TO_SECONDS_CONVERSION)
        for req_metric in zip(*valid_metrics):
            is_good_req = all([s >= r for s, r in zip(slo_values, req_metric)])
            if is_good_req:
                good_completed += 1
    if completed == 0:
        warnings.warn(
            "All requests failed. This is likely due to a misconfiguration "
@@ -354,6 +489,7 @@ def calculate_metrics(
        total_input=total_input,
        total_output=sum(actual_output_lens),
        request_throughput=completed / dur_s,
        request_goodput=good_completed / dur_s,
        output_throughput=sum(actual_output_lens) / dur_s,
        total_token_throughput=(total_input + sum(actual_output_lens)) / dur_s,
        mean_ttft_ms=np.mean(ttfts or 0) *
@@ -372,9 +508,9 @@ def calculate_metrics(
        median_itl_ms=np.median(itls or 0) * 1000,
        percentiles_itl_ms=[(p, np.percentile(itls or 0, p) * 1000)
                            for p in selected_percentiles],
-        mean_e2el_ms=np.median(e2els or 0) * 1000,
+        mean_e2el_ms=np.mean(e2els or 0) * 1000,
        std_e2el_ms=np.std(e2els or 0) * 1000,
-        median_e2el_ms=np.mean(e2els or 0) * 1000,
+        median_e2el_ms=np.median(e2els or 0) * 1000,
        percentiles_e2el_ms=[(p, np.percentile(e2els or 0, p) * 1000)
                             for p in selected_percentiles],
    )
@@ -387,16 +523,20 @@ 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],
    best_of: int,
    request_rate: float,
    burstiness: float,
    disable_tqdm: bool,
    profile: bool,
    selected_percentile_metrics: List[str],
    selected_percentiles: List[str],
    ignore_eos: bool,
    goodput_config_dict: Dict[str, float],
    max_concurrency: Optional[int],
 ):
    if backend in ASYNC_REQUEST_FUNCS:
        request_func = ASYNC_REQUEST_FUNCS[backend]
@@ -412,6 +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,
@@ -432,6 +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,
@@ -444,15 +586,38 @@ async def benchmark(
        if profile_output.success:
            print("Profiler started")
    if burstiness == 1.0:
        distribution = "Poisson process"
    else:
        distribution = "Gamma distribution"
    print(f"Traffic request rate: {request_rate}")
    print(f"Burstiness factor: {burstiness} ({distribution})")
    print(f"Maximum request concurrency: {max_concurrency}")
    pbar = None if disable_tqdm else tqdm(total=len(input_requests))
    # This can be used once the minimum Python version is 3.10 or higher,
    # and it will simplify the code in limited_request_func.
    #    semaphore = (asyncio.Semaphore(max_concurrency)
    #                 if max_concurrency else contextlib.nullcontext())
    semaphore = (asyncio.Semaphore(max_concurrency)
                 if max_concurrency else None)
    async def limited_request_func(request_func_input, pbar):
        if semaphore is None:
            return await request_func(request_func_input=request_func_input,
                                      pbar=pbar)
        async with semaphore:
            return await request_func(request_func_input=request_func_input,
                                      pbar=pbar)
    benchmark_start_time = time.perf_counter()
    tasks: List[asyncio.Task] = []
-    async for request in get_request(input_requests, request_rate):
+    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,
@@ -463,8 +628,8 @@ async def benchmark(
                                              ignore_eos=ignore_eos)
        tasks.append(
            asyncio.create_task(
-                request_func(request_func_input=request_func_input,
+                limited_request_func(request_func_input=request_func_input,
-                             pbar=pbar)))
+                                     pbar=pbar)))
    outputs: List[RequestFuncOutput] = await asyncio.gather(*tasks)
    if profile:
@@ -494,6 +659,7 @@ async def benchmark(
        tokenizer=tokenizer,
        selected_percentile_metrics=selected_percentile_metrics,
        selected_percentiles=selected_percentiles,
        goodput_config_dict=goodput_config_dict,
    )
    print("{s:{c}^{n}}".format(s=' Serving Benchmark Result ', n=50, c='='))
@@ -505,6 +671,9 @@ async def benchmark(
                                 metrics.total_output))
    print("{:<40} {:<10.2f}".format("Request throughput (req/s):",
                                    metrics.request_throughput))
    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):",
                                    metrics.output_throughput))
    print("{:<40} {:<10.2f}".format("Total Token throughput (tok/s):",
@@ -516,6 +685,8 @@ async def benchmark(
        "total_input_tokens": metrics.total_input,
        "total_output_tokens": metrics.total_output,
        "request_throughput": metrics.request_throughput,
        "request_goodput:":
        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],
@@ -569,6 +740,41 @@ async def benchmark(
    return result
 def check_goodput_args(args):
    # Check and parse goodput arguments
    goodput_config_dict = {}
    VALID_NAMES = ["ttft", "tpot", "e2el"]
    if args.goodput:
        goodput_config_dict = parse_goodput(args.goodput)
        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}. "
                    "The service level objective name should be one of "
                    f"{str(VALID_NAMES)}. ")
            if slo_val < 0:
                raise ValueError(
                    f"Invalid value found, {slo_name}: {slo_val}. "
                    "The service level objective value should be "
                    "non-negative.")
    return goodput_config_dict
 def parse_goodput(slo_pairs):
    goodput_config_dict = {}
    try:
        for slo_pair in slo_pairs:
            slo_name, slo_val = slo_pair.split(":")
            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 goodput_config_dict
 def main(args: argparse.Namespace):
    print(args)
    random.seed(args.seed)
@@ -576,7 +782,9 @@ 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
    if args.base_url is not None:
        api_url = f"{args.base_url}{args.endpoint}"
@@ -586,6 +794,7 @@ def main(args: argparse.Namespace):
        base_url = f"http://{args.host}:{args.port}"
    tokenizer = get_tokenizer(tokenizer_id,
                              tokenizer_mode=tokenizer_mode,
                              trust_remote_code=args.trust_remote_code)
    if args.dataset is not None:
@@ -646,6 +855,7 @@ def main(args: argparse.Namespace):
            dataset_split=args.hf_split,
            num_requests=args.num_prompts,
            tokenizer=tokenizer,
            random_seed=args.seed,
            fixed_output_len=args.hf_output_len,
        )
@@ -662,17 +872,25 @@ def main(args: argparse.Namespace):
    else:
        raise ValueError(f"Unknown dataset: {args.dataset_name}")
    goodput_config_dict = check_goodput_args(args)
    # Avoid GC processing "static" data - reduce pause times.
    gc.collect()
    gc.freeze()
    benchmark_result = asyncio.run(
        benchmark(
            backend=backend,
            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,
            best_of=args.best_of,
            request_rate=args.request_rate,
            burstiness=args.burstiness,
            disable_tqdm=args.disable_tqdm,
            profile=args.profile,
            selected_percentile_metrics=args.percentile_metrics.split(","),
@@ -680,6 +898,8 @@ def main(args: argparse.Namespace):
                float(p) for p in args.metric_percentiles.split(",")
            ],
            ignore_eos=args.ignore_eos,
            goodput_config_dict=goodput_config_dict,
            max_concurrency=args.max_concurrency,
        ))
    # Save config and results to json
@@ -707,15 +927,19 @@ 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
        # Merge with benchmark result
        result_json = {**result_json, **benchmark_result}
        # Save to file
        base_model_id = model_id.split("/")[-1]
-        file_name = f"{backend}-{args.request_rate}qps-{base_model_id}-{current_dt}.json"  #noqa
+        max_concurrency_str = (f"-concurrency{args.max_concurrency}"
                               if args.max_concurrency is not None else "")
        file_name = f"{backend}-{args.request_rate}qps{max_concurrency_str}-{base_model_id}-{current_dt}.json"  #noqa
        if args.result_filename:
            file_name = args.result_filename
        if args.result_dir:
@@ -766,6 +990,19 @@ if __name__ == "__main__":
                        default=None,
                        help="Path to the sharegpt/sonnet dataset. "
                        "Or the huggingface dataset ID if using HF dataset.")
    parser.add_argument(
        "--max-concurrency",
        type=int,
        default=None,
        help="Maximum number of concurrent requests. This can be used "
        "to help simulate an environment where a higher level component "
        "is enforcing a maximum number of concurrent requests. While the "
        "--request-rate argument controls the rate at which requests are "
        "initiated, this argument will control how many are actually allowed "
        "to execute at a time. This means that when used in combination, the "
        "actual request rate may be lower than specified with --request-rate, "
        "if the server is not processing requests fast enough to keep up.")
    parser.add_argument(
        "--model",
        type=str,
@@ -808,8 +1045,20 @@ if __name__ == "__main__":
        default=float("inf"),
        help="Number of requests per second. If this is inf, "
        "then all the requests are sent at time 0. "
-        "Otherwise, we use Poisson process to synthesize "
+        "Otherwise, we use Poisson process or gamma distribution "
-        "the request arrival times.",
+        "to synthesize the request arrival times.",
    )
    parser.add_argument(
        "--burstiness",
        type=float,
        default=1.0,
        help="Burstiness factor of the request generation. "
        "Only take effect when request_rate is not inf. "
        "Default value is 1, which follows Poisson process. "
        "Otherwise, the request intervals follow a gamma distribution. "
        "A lower burstiness value (0 < burstiness < 1) results in more "
        "bursty requests. A higher burstiness value (burstiness > 1) "
        "results in a more uniform arrival of requests.",
    )
    parser.add_argument("--seed", type=int, default=0)
    parser.add_argument(
@@ -879,6 +1128,17 @@ if __name__ == "__main__":
        "Default value is \"99\". "
        "Use \"--percentile-metrics\" to select metrics.",
    )
    parser.add_argument(
        "--goodput",
        nargs="+",
        required=False,
        help="Specify service level objectives for goodput as \"KEY:VALUE\" "
        "pairs, where the key is a metric name, and the value is in "
        "milliseconds. Multiple \"KEY:VALUE\" pairs can be provided, "
        "separated by spaces. Allowed request level metric names are "
        "\"ttft\", \"tpot\", \"e2el\". For more context on the definition of "
        "goodput, refer to DistServe paper: https://arxiv.org/pdf/2401.09670 "
        "and the blog: https://hao-ai-lab.github.io/blogs/distserve")
    # group for dataset specific arguments
    sonnet_group = parser.add_argument_group("sonnet dataset options")
@@ -960,5 +1220,22 @@ if __name__ == "__main__":
        "from the sampled HF dataset.",
    )
    parser.add_argument(
        '--tokenizer-mode',
        type=str,
        default="auto",
        choices=['auto', 'slow', 'mistral'],
        help='The tokenizer mode.\n\n* "auto" will use the '
        'fast tokenizer if available.\n* "slow" will '
        '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)
--- a/benchmarks/benchmark_serving_guided.py
+++ b/benchmarks/benchmark_serving_guided.py
@@ -0,0 +1,882 @@
 # SPDX-License-Identifier: Apache-2.0
 r"""Benchmark online serving throughput with guided decoding.
 On the server side, run one of the following commands:
    (vLLM OpenAI API server)
    vllm serve <your_model> --disable-log-requests
    (TGI backend)
    ./launch_tgi_server.sh <your_model> <max_batch_total_tokens>
 On the client side, run:
    python benchmarks/benchmark_serving.py \
        --backend <backend> \
        --model <your_model> \
        --dataset json \
        --guided-decoding-ratio 1.0 \
        --guided-decoding-backend xgrammar \
        --request-rate 10 \
        --num-prompts 1000
    when using tgi backend, add
        --endpoint /generate_stream
    to the end of the command above.
 """
 import argparse
 import asyncio
 import dataclasses
 import json
 import os
 import random
 import time
 import warnings
 from dataclasses import dataclass
 from typing import AsyncGenerator, List, Optional, Tuple
 import datasets
 import numpy as np
 import pandas as pd
 from backend_request_func import (ASYNC_REQUEST_FUNCS, RequestFuncInput,
                                  RequestFuncOutput)
 from tqdm.asyncio import tqdm
 from transformers import PreTrainedTokenizerBase
 try:
    from vllm.transformers_utils.tokenizer import get_tokenizer
 except ImportError:
    from backend_request_func import get_tokenizer
 try:
    from vllm.utils import FlexibleArgumentParser
 except ImportError:
    from argparse import ArgumentParser as FlexibleArgumentParser
 MILLISECONDS_TO_SECONDS_CONVERSION = 1000
@dataclass
 class BenchmarkMetrics:
    completed: int
    total_input: int
    total_output: int
    request_throughput: float
    request_goodput: float
    output_throughput: float
    total_token_throughput: float
    mean_ttft_ms: float
    median_ttft_ms: float
    std_ttft_ms: float
    percentiles_ttft_ms: List[Tuple[float, float]]
    mean_tpot_ms: float
    median_tpot_ms: float
    std_tpot_ms: float
    percentiles_tpot_ms: List[Tuple[float, float]]
    mean_itl_ms: float
    median_itl_ms: float
    std_itl_ms: float
    percentiles_itl_ms: List[Tuple[float, float]]
    # E2EL stands for end-to-end latency per request.
    # It is the time taken on the client side from sending
    # a request to receiving a complete response.
    mean_e2el_ms: float
    median_e2el_ms: float
    std_e2el_ms: float
    percentiles_e2el_ms: List[Tuple[float, float]]
@dataclasses.dataclass
 class SampleRequest:
    """A class representing a single inference request for benchmarking.
    Attributes:
        prompt: The input text prompt for the model.
        multi_modal_data: Optional dictionary containing multi-modal data (e.g.
            images).
        prompt_len: The length of the prompt in tokens.
        expected_output_len: The expected length of the output in tokens.
    """
    prompt: str
    prompt_len: int
    expected_output_len: int
    schema: dict
    structure_type: str
    completion: str = None
 def sample_requests(tokenizer: PreTrainedTokenizerBase,
                    args: argparse.Namespace) -> List[SampleRequest]:
    if args.dataset == 'json':
        if args.json_schema_path is None:
            dir_path = os.path.dirname(os.path.realpath(__file__))
            args.json_schema_path = os.path.join(dir_path,
                                                 "structured_schemas",
                                                 "structured_schema_1.json")
        with open(args.json_schema_path) as f:
            schema = json.load(f)
        prompt = f"Generate an example of a user profile given the following schema: {json.dumps(schema)}"  # noqa: E501
        input_len = len(tokenizer(prompt).input_ids)
        print(f"Input length of the prompt: {input_len} tokens")
        requests = [
            SampleRequest(prompt=prompt,
                          prompt_len=input_len,
                          expected_output_len=args.output_len,
                          schema=schema,
                          structure_type=args.structure_type)
            for _ in range(args.num_prompts)
        ]
    elif args.dataset == "grammar":
        schema = """
            ?start: select_statement
            ?select_statement: "SELECT " column_list " FROM " table_name
            ?column_list: column_name ("," column_name)*
            ?table_name: identifier
            ?column_name: identifier
            ?identifier: /[a-zA-Z_][a-zA-Z0-9_]*/
        """
        prompt = "Generate an SQL query to show the 'username' \
            and 'email' from the 'users' table."
        input_len = len(tokenizer(prompt).input_ids)
        print(f"Input length of the prompt: {input_len} tokens")
        requests = [
            SampleRequest(prompt=prompt,
                          prompt_len=input_len,
                          expected_output_len=args.output_len,
                          schema=schema,
                          structure_type=args.structure_type)
            for _ in range(args.num_prompts)
        ]
    elif args.dataset == "regex":
        regex = r"\w+@\w+\.com\n"
        args.regex = regex
        prompt = "Generate an email address for Alan Turing, \
            who works in Enigma. End in .com and new line. \
                Example result: alan.turing@enigma.com\n"
        input_len = len(tokenizer(prompt).input_ids)
        print(f"Input length of the prompt: {input_len} tokens")
        requests = [
            SampleRequest(prompt=prompt,
                          prompt_len=input_len,
                          expected_output_len=args.output_len,
                          schema=regex,
                          structure_type=args.structure_type)
            for _ in range(args.num_prompts)
        ]
    elif args.dataset == "choice":
        choice = ["Positive", "Negative"]
        args.choice = choice
        prompt = "Classify this sentiment: vLLM is wonderful!"
        input_len = len(tokenizer(prompt).input_ids)
        print(f"Input length of the prompt: {input_len} tokens")
        requests = [
            SampleRequest(prompt=prompt,
                          prompt_len=input_len,
                          expected_output_len=args.output_len,
                          schema=choice,
                          structure_type=args.structure_type)
            for _ in range(args.num_prompts)
        ]
    elif args.dataset == "xgrammar_bench":
        requests: List[SampleRequest] = []
        dataset = datasets.load_dataset("NousResearch/json-mode-eval",
                                        split="train")
        print(f"dataset has {len(dataset)} entries")
        len_dataset = len(dataset)
        for data_point_idx in range(args.num_prompts):
            idx = data_point_idx
            while idx >= len_dataset:
                idx -= len_dataset
            schema = dataset["schema"][idx]
            prompt = tokenizer.apply_chat_template(dataset["prompt"][idx],
                                                   tokenize=False)
            input_len = len(tokenizer(prompt).input_ids)
            completion = dataset["completion"][idx]
            requests.append(
                SampleRequest(prompt=prompt,
                              prompt_len=input_len,
                              expected_output_len=args.output_len,
                              schema=schema,
                              structure_type=args.structure_type,
                              completion=completion))
    return requests
 async def get_request(
    input_requests: List[SampleRequest],
    request_rate: float,
    burstiness: float = 1.0,
 ) -> AsyncGenerator[Tuple[int, SampleRequest], None]:
    """
    Asynchronously generates requests at a specified rate 
    with OPTIONAL burstiness.
    Args:
        input_requests: 
            A list of input requests, each represented as a tuple.
        request_rate: 
            The rate at which requests are generated (requests/s).
        burstiness (optional): 
            The burstiness factor of the request generation. 
            Only takes effect when request_rate is not inf.
            Default value is 1, which follows a Poisson process.
            Otherwise, the request intervals follow a gamma distribution.
            A lower burstiness value (0 < burstiness < 1) results 
            in more bursty requests, while a higher burstiness value 
            (burstiness > 1) results in a more uniform arrival of requests.
    """
    input_requests = iter(input_requests)
    # Calculate scale parameter theta to maintain the desired request_rate.
    assert burstiness > 0, (
        f"A positive burstiness factor is expected, but given {burstiness}.")
    theta = 1.0 / (request_rate * burstiness)
    for i, request in enumerate(input_requests):
        yield i, request
        if request_rate == float("inf"):
            # If the request rate is infinity, then we don't need to wait.
            continue
        # Sample the request interval from the gamma distribution.
        # If burstiness is 1, it follows exponential distribution.
        interval = np.random.gamma(shape=burstiness, scale=theta)
        # The next request will be sent after the interval.
        await asyncio.sleep(interval)
 def calculate_metrics(
    input_requests: List[Tuple[str, int, int]],
    outputs: List[RequestFuncOutput],
    dur_s: float,
    tokenizer: PreTrainedTokenizerBase,
    selected_percentile_metrics: List[str],
    selected_percentiles: List[float],
 ) -> Tuple[BenchmarkMetrics, List[int]]:
    actual_output_lens: List[int] = []
    total_input = 0
    completed = 0
    good_completed = 0
    itls: List[float] = []
    tpots: List[float] = []
    all_tpots: List[float] = []
    ttfts: List[float] = []
    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
            # 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,
                          add_special_tokens=False).input_ids)
            actual_output_lens.append(output_len)
            total_input += input_requests[i].prompt_len
            tpot = 0
            if output_len > 1:
                tpot = (outputs[i].latency - outputs[i].ttft) / (output_len -
                                                                 1)
                tpots.append(tpot)
            outputs[i].tpot = sum(tpots) / len(tpots) if len(tpots) else 0
            # Note: if output_len <= 1, we regard tpot as 0 for goodput
            all_tpots.append(tpot)
            itls += outputs[i].itl
            ttfts.append(outputs[i].ttft)
            e2els.append(outputs[i].latency)
            completed += 1
        else:
            actual_output_lens.append(0)
    if completed == 0:
        warnings.warn(
            "All requests failed. This is likely due to a misconfiguration "
            "on the benchmark arguments.",
            stacklevel=2)
    metrics = BenchmarkMetrics(
        completed=completed,
        total_input=total_input,
        total_output=sum(actual_output_lens),
        request_throughput=completed / dur_s,
        request_goodput=good_completed / dur_s,
        output_throughput=sum(actual_output_lens) / dur_s,
        total_token_throughput=(total_input + sum(actual_output_lens)) / dur_s,
        mean_ttft_ms=np.mean(ttfts or 0) *
        1000,  # ttfts is empty if streaming is not supported by backend
        std_ttft_ms=np.std(ttfts or 0) * 1000,
        median_ttft_ms=np.median(ttfts or 0) * 1000,
        percentiles_ttft_ms=[(p, np.percentile(ttfts or 0, p) * 1000)
                             for p in selected_percentiles],
        mean_tpot_ms=np.mean(tpots or 0) * 1000,
        std_tpot_ms=np.std(tpots or 0) * 1000,
        median_tpot_ms=np.median(tpots or 0) * 1000,
        percentiles_tpot_ms=[(p, np.percentile(tpots or 0, p) * 1000)
                             for p in selected_percentiles],
        mean_itl_ms=np.mean(itls or 0) * 1000,
        std_itl_ms=np.std(itls or 0) * 1000,
        median_itl_ms=np.median(itls or 0) * 1000,
        percentiles_itl_ms=[(p, np.percentile(itls or 0, p) * 1000)
                            for p in selected_percentiles],
        mean_e2el_ms=np.mean(e2els or 0) * 1000,
        std_e2el_ms=np.std(e2els or 0) * 1000,
        median_e2el_ms=np.median(e2els or 0) * 1000,
        percentiles_e2el_ms=[(p, np.percentile(e2els or 0, p) * 1000)
                             for p in selected_percentiles],
    )
    return metrics, actual_output_lens
 async def benchmark(
    backend: str,
    api_url: str,
    base_url: str,
    model_id: str,
    tokenizer: PreTrainedTokenizerBase,
    input_requests: List[SampleRequest],
    request_rate: float,
    burstiness: float,
    disable_tqdm: bool,
    profile: bool,
    selected_percentile_metrics: List[str],
    selected_percentiles: List[str],
    ignore_eos: bool,
    max_concurrency: Optional[int],
    guided_decoding_ratio: float,
    guided_decoding_backend: str,
 ):
    if backend in ASYNC_REQUEST_FUNCS:
        request_func = ASYNC_REQUEST_FUNCS[backend]
    else:
        raise ValueError(f"Unknown backend: {backend}")
    def prepare_extra_body(request) -> dict:
        extra_body = {}
        # Add the schema to the extra_body
        extra_body[request.structure_type] = request.schema
        # Add the specific guided_decoding_backend
        extra_body["guided_decoding_backend"] = guided_decoding_backend
        return extra_body
    print("Starting initial single prompt test run...")
    guided_decoding_req_idx = random.sample(
        range(len(input_requests)),
        int(len(input_requests) * guided_decoding_ratio))
    test_request = input_requests[0]
    test_input = RequestFuncInput(
        model=model_id,
        prompt=test_request.prompt,
        api_url=api_url,
        prompt_len=test_request.prompt_len,
        output_len=test_request.expected_output_len,
        ignore_eos=ignore_eos,
        extra_body=prepare_extra_body(test_request),
    )
    test_output = await request_func(request_func_input=test_input)
    if not test_output.success:
        raise ValueError(
            "Initial test run failed - Please make sure benchmark arguments "
            f"are correctly specified. Error: {test_output.error}")
    else:
        print("Initial test run completed. Starting main benchmark run...")
    if profile:
        print("Starting profiler...")
        profile_input = RequestFuncInput(
            model=model_id,
            prompt=test_request.prompt,
            api_url=base_url + "/start_profile",
            prompt_len=test_request.prompt_len,
            output_len=test_request.expected_output_len,
            ignore_eos=ignore_eos,
            extra_body=prepare_extra_body(test_request),
        )
        profile_output = await request_func(request_func_input=profile_input)
        if profile_output.success:
            print("Profiler started")
    if burstiness == 1.0:
        distribution = "Poisson process"
    else:
        distribution = "Gamma distribution"
    print(f"Traffic request rate: {request_rate}")
    print(f"Burstiness factor: {burstiness} ({distribution})")
    print(f"Maximum request concurrency: {max_concurrency}")
    pbar = None if disable_tqdm else tqdm(total=len(input_requests))
    # This can be used once the minimum Python version is 3.10 or higher,
    # and it will simplify the code in limited_request_func.
    #    semaphore = (asyncio.Semaphore(max_concurrency)
    #                 if max_concurrency else contextlib.nullcontext())
    semaphore = (asyncio.Semaphore(max_concurrency)
                 if max_concurrency else None)
    async def limited_request_func(request_func_input, pbar):
        if semaphore is None:
            return await request_func(request_func_input=request_func_input,
                                      pbar=pbar)
        async with semaphore:
            return await request_func(request_func_input=request_func_input,
                                      pbar=pbar)
    benchmark_start_time = time.perf_counter()
    tasks: List[asyncio.Task] = []
    expected: List[str] = []
    async for i, request in get_request(input_requests, request_rate,
                                        burstiness):
        extra_body = prepare_extra_body(
            request) if i in guided_decoding_req_idx else None
        request_func_input = RequestFuncInput(
            model=model_id,
            prompt=request.prompt,
            api_url=api_url,
            prompt_len=request.prompt_len,
            output_len=request.expected_output_len,
            ignore_eos=ignore_eos,
            extra_body=extra_body,
        )
        expected.append(request.completion)
        tasks.append(
            asyncio.create_task(
                limited_request_func(request_func_input=request_func_input,
                                     pbar=pbar)))
    outputs: List[RequestFuncOutput] = await asyncio.gather(*tasks)
    if profile:
        print("Stopping profiler...")
        profile_input = RequestFuncInput(
            model=model_id,
            prompt=test_request.prompt,
            api_url=base_url + "/stop_profile",
            prompt_len=test_request.prompt_len,
            output_len=test_request.expected_output_len,
            extra_body={test_request.structure_type: test_request.schema},
        )
        profile_output = await request_func(request_func_input=profile_input)
        if profile_output.success:
            print("Profiler stopped")
    if pbar is not None:
        pbar.close()
    benchmark_duration = time.perf_counter() - benchmark_start_time
    metrics, actual_output_lens = calculate_metrics(
        input_requests=input_requests,
        outputs=outputs,
        dur_s=benchmark_duration,
        tokenizer=tokenizer,
        selected_percentile_metrics=selected_percentile_metrics,
        selected_percentiles=selected_percentiles,
    )
    print("{s:{c}^{n}}".format(s=' Serving Benchmark Result ', n=50, c='='))
    print("{:<40} {:<10}".format("Successful requests:", metrics.completed))
    print("{:<40} {:<10.2f}".format("Benchmark duration (s):",
                                    benchmark_duration))
    print("{:<40} {:<10}".format("Total input tokens:", metrics.total_input))
    print("{:<40} {:<10}".format("Total generated tokens:",
                                 metrics.total_output))
    print("{:<40} {:<10.2f}".format("Request throughput (req/s):",
                                    metrics.request_throughput))
    print("{:<40} {:<10.2f}".format("Output token throughput (tok/s):",
                                    metrics.output_throughput))
    print("{:<40} {:<10.2f}".format("Total Token throughput (tok/s):",
                                    metrics.total_token_throughput))
    result = {
        "duration":
        benchmark_duration,
        "completed":
        metrics.completed,
        "total_input_tokens":
        metrics.total_input,
        "total_output_tokens":
        metrics.total_output,
        "request_throughput":
        metrics.request_throughput,
        "output_throughput":
        metrics.output_throughput,
        "total_token_throughput":
        metrics.total_token_throughput,
        "ttft_description":
        pd.Series([output.ttft for output in outputs]).describe().to_dict(),
        "tpot_description":
        pd.Series([output.tpot for output in outputs]).describe().to_dict(),
        "input_lens": [output.prompt_len for output in outputs],
        "output_lens":
        actual_output_lens,
        "ttfts": [output.ttft for output in outputs],
        "itls": [output.itl for output in outputs],
        "errors": [output.error for output in outputs],
    }
    ret = [{
        'generated': output.generated_text,
        'expected': gt
    } for output, gt in zip(outputs, expected)]
    def process_one_metric(
        # E.g., "ttft"
        metric_attribute_name: str,
        # E.g., "TTFT"
        metric_name: str,
        # E.g., "Time to First Token"
        metric_header: str,
    ):
        # This function prints and adds statistics of the specified
        # metric.
        if metric_attribute_name not in selected_percentile_metrics:
            return
        print("{s:{c}^{n}}".format(s=metric_header, n=50, c='-'))
        print("{:<40} {:<10.2f}".format(
            f"Mean {metric_name} (ms):",
            getattr(metrics, f"mean_{metric_attribute_name}_ms")))
        print("{:<40} {:<10.2f}".format(
            f"Median {metric_name} (ms):",
            getattr(metrics, f"median_{metric_attribute_name}_ms")))
        result[f"mean_{metric_attribute_name}_ms"] = getattr(
            metrics, f"mean_{metric_attribute_name}_ms")
        result[f"median_{metric_attribute_name}_ms"] = getattr(
            metrics, f"median_{metric_attribute_name}_ms")
        result[f"std_{metric_attribute_name}_ms"] = getattr(
            metrics, f"std_{metric_attribute_name}_ms")
        for p, value in getattr(metrics,
                                f"percentiles_{metric_attribute_name}_ms"):
            p_word = str(int(p)) if int(p) == p else str(p)
            print("{:<40} {:<10.2f}".format(f"P{p_word} {metric_name} (ms):",
                                            value))
            result[f"p{p_word}_{metric_attribute_name}_ms"] = value
    process_one_metric("ttft", "TTFT", "Time to First Token")
    process_one_metric("tpot", "TPOT",
                       "Time per Output Token (excl. 1st token)")
    process_one_metric("itl", "ITL", "Inter-token Latency")
    process_one_metric("e2el", "E2EL", "End-to-end Latency")
    print("=" * 50)
    return result, ret
 def evaluate(ret, args):
    def _eval_correctness_json(expected, actual):
        # extract json string from string using regex
        import re
        actual = actual.replace('\n', '').replace(' ', '').strip()
        try:
            actual = re.search(r'\{.*\}', actual).group()
            actual = json.loads(actual)
        except Exception:
            return False
        return True
    def _eval_correctness_choice(expected, actual):
        return actual in args.choice
    def _eval_correctness_regex(expected, actual):
        import re
        return re.match(args.regex, actual) is not None
    def _eval_correctness(expected, actual):
        if args.structure_type == 'guided_json':
            return _eval_correctness_json(expected, actual)
        elif args.structure_type == 'guided_regex':
            return _eval_correctness_regex(expected, actual)
        elif args.structure_type == 'guided_choice':
            return _eval_correctness_choice(expected, actual)
        else:
            return None
    scores = []
    for res in ret:
        score = _eval_correctness(res['expected'], res['generated'])
        res['correctness'] = score
        scores.append(score)
    not_none_scores = [score for score in scores if score is not None]
    return (sum(not_none_scores) / len(not_none_scores) *
            100) if len(not_none_scores) > 0 else None
 def main(args: argparse.Namespace):
    print(args)
    random.seed(args.seed)
    np.random.seed(args.seed)
    backend = args.backend
    model_id = args.model
    tokenizer_id = args.tokenizer if args.tokenizer is not None else args.model
    if args.base_url is not None:
        api_url = f"{args.base_url}{args.endpoint}"
        base_url = f"{args.base_url}"
    else:
        api_url = f"http://{args.host}:{args.port}{args.endpoint}"
        base_url = f"http://{args.host}:{args.port}"
    tokenizer = get_tokenizer(tokenizer_id,
                              trust_remote_code=args.trust_remote_code)
    if args.dataset == 'grammar':
        args.structure_type = 'guided_grammar'
    elif args.dataset == 'regex':
        args.structure_type = 'guided_regex'
    elif args.dataset == 'choice':
        args.structure_type = 'guided_choice'
    else:
        args.structure_type = 'guided_json'
    if args.no_guided_decoding:
        args.guided_decoding_ratio = 0
    if args.save_results:
        result_file_name = f'{args.guided_decoding_ratio}guided'
        result_file_name += f"_{backend}"
        result_file_name += f"_{args.request_rate}qps"
        result_file_name += f"_{args.model.split('/')[-1]}"
        result_file_name += f"_{args.dataset}"
        result_file_name += f"_{args.num_prompts}"
        result_file_name += f"_out{args.output_len}"
        result_file_name += ".txt"
    else:
        result_file_name = None
    input_requests = sample_requests(tokenizer, args)
    benchmark_result, ret = asyncio.run(
        benchmark(
            backend=backend,
            api_url=api_url,
            base_url=base_url,
            model_id=model_id,
            tokenizer=tokenizer,
            input_requests=input_requests,
            request_rate=args.request_rate,
            burstiness=args.burstiness,
            disable_tqdm=args.disable_tqdm,
            profile=args.profile,
            selected_percentile_metrics=args.percentile_metrics.split(","),
            selected_percentiles=[
                float(p) for p in args.metric_percentiles.split(",")
            ],
            ignore_eos=args.ignore_eos,
            max_concurrency=args.max_concurrency,
            guided_decoding_ratio=args.guided_decoding_ratio,
            guided_decoding_backend=args.guided_decoding_backend,
        ))
    # Save config and results to json
    score = evaluate(ret, args)
    print("correct_rate(%)", score, '\n')
    if args.save_results:
        results = {
            "backend":
            backend,
            "model_id":
            model_id,
            "tokenizer_id":
            tokenizer_id,
            "num_prompts":
            args.num_prompts,
            "request_rate":
            args.request_rate if args.request_rate < float("inf") else "inf",
            "burstiness":
            args.burstiness,
            "max_concurrency":
            args.max_concurrency,
            "correct_rate(%)":
            score
        }
        results = {"outputs": ret, **results, **benchmark_result}
        # Save to file
        if args.result_filename:
            result_file_name = args.result_filename
        if args.result_dir:
            result_file_name = os.path.join(args.result_dir, result_file_name)
        with open(result_file_name, "w", encoding='utf-8') as outfile:
            json.dump(results, outfile, indent=4)
 if __name__ == "__main__":
    parser = FlexibleArgumentParser(
        description="Benchmark the online serving throughput.")
    parser.add_argument(
        "--backend",
        type=str,
        default="vllm",
        choices=list(ASYNC_REQUEST_FUNCS.keys()),
    )
    parser.add_argument(
        "--base-url",
        type=str,
        default=None,
        help="Server or API base url if not using http host and port.",
    )
    parser.add_argument("--host", type=str, default="localhost")
    parser.add_argument("--port", type=int, default=8000)
    parser.add_argument(
        "--endpoint",
        type=str,
        default="/v1/completions",
        help="API endpoint.",
    )
    parser.add_argument(
        "--dataset",
        default='json',
        choices=['json', 'grammar', 'regex', 'choice', 'xgrammar_bench'])
    parser.add_argument("--json_schema_path",
                        type=str,
                        default=None,
                        help="Path to json schema.")
    parser.add_argument(
        "--max-concurrency",
        type=int,
        default=None,
        help="Maximum number of concurrent requests. This can be used "
        "to help simulate an environment where a higher level component "
        "is enforcing a maximum number of concurrent requests. While the "
        "--request-rate argument controls the rate at which requests are "
        "initiated, this argument will control how many are actually allowed "
        "to execute at a time. This means that when used in combination, the "
        "actual request rate may be lower than specified with --request-rate, "
        "if the server is not processing requests fast enough to keep up.")
    parser.add_argument(
        "--model",
        type=str,
        required=True,
        help="Name of the model.",
    )
    parser.add_argument(
        "--tokenizer",
        type=str,
        help=
        "Name or path of the tokenizer, if not using the default tokenizer.",  # noqa: E501
    )
    parser.add_argument(
        "--num-prompts",
        type=int,
        default=1000,
        help="Number of prompts to process.",
    )
    parser.add_argument(
        "--output-len",
        type=int,
        default=128,
        help="Number of output tokens.",
    )
    parser.add_argument(
        "--request-rate",
        type=float,
        default=float("inf"),
        help="Number of requests per second. If this is inf, "
        "then all the requests are sent at time 0. "
        "Otherwise, we use Poisson process or gamma distribution "
        "to synthesize the request arrival times.",
    )
    parser.add_argument(
        "--burstiness",
        type=float,
        default=1.0,
        help="Burstiness factor of the request generation. "
        "Only take effect when request_rate is not inf. "
        "Default value is 1, which follows Poisson process. "
        "Otherwise, the request intervals follow a gamma distribution. "
        "A lower burstiness value (0 < burstiness < 1) results in more "
        "bursty requests. A higher burstiness value (burstiness > 1) "
        "results in a more uniform arrival of requests.",
    )
    parser.add_argument("--seed", type=int, default=0)
    parser.add_argument(
        "--trust-remote-code",
        action="store_true",
        help="Trust remote code from huggingface",
    )
    parser.add_argument(
        "--disable-tqdm",
        action="store_true",
        help="Specify to disable tqdm progress bar.",
    )
    parser.add_argument(
        "--save-results",
        action="store_true",
        help="Specify to save benchmark results to a json file",
    )
    parser.add_argument(
        "--profile",
        action="store_true",
        help="Use Torch Profiler. The endpoint must be launched with "
        "VLLM_TORCH_PROFILER_DIR to enable profiler.",
    )
    parser.add_argument(
        "--result-dir",
        type=str,
        default=None,
        help="Specify directory to save benchmark json results."
        "If not specified, results are saved in the current directory.",
    )
    parser.add_argument(
        "--result-filename",
        type=str,
        default=None,
        help="Specify the filename to save benchmark json results."
        "If not specified, results will be saved in "
        "{backend}-{args.request_rate}qps-{base_model_id}-{current_dt}.json"
        " format.",
    )
    parser.add_argument(
        "--ignore-eos",
        action="store_true",
        help="Set ignore_eos flag when sending the benchmark request."
        "Warning: ignore_eos is not supported in deepspeed_mii and tgi.")
    parser.add_argument(
        "--percentile-metrics",
        type=str,
        default="ttft,tpot,itl",
        help="Comma-seperated list of selected metrics to report percentils. "
        "This argument specifies the metrics to report percentiles. "
        "Allowed metric names are \"ttft\", \"tpot\", \"itl\", \"e2el\". "
        "Default value is \"ttft,tpot,itl\".")
    parser.add_argument(
        "--metric-percentiles",
        type=str,
        default="99",
        help="Comma-seperated list of percentiles for selected metrics. "
        "To report 25-th, 50-th, and 75-th percentiles, use \"25,50,75\". "
        "Default value is \"99\". "
        "Use \"--percentile-metrics\" to select metrics.",
    )
    parser.add_argument("--no-guided-decoding",
                        action='store_true',
                        default=False,
                        help="Whether to disable JSON decoding or not.")
    parser.add_argument("--guided-decoding-ratio",
                        type=float,
                        default=1.0,
                        help="Ratio of Guided Decoding requests")
    parser.add_argument("--guided-decoding-backend",
                        type=str,
                        choices=["outlines", "lm-format-enforcer", "xgrammar"],
                        default="xgrammar",
                        help="Backend to use for guided decoding")
    args = parser.parse_args()
    main(args)
--- a/benchmarks/benchmark_throughput.py
+++ b/benchmarks/benchmark_throughput.py
@@ -1,30 +1,100 @@
 # SPDX-License-Identifier: Apache-2.0
 """Benchmark offline inference throughput."""
 import argparse
 import dataclasses
 import json
 import random
 import time
-from typing import List, Optional, Tuple
+from functools import cache
 from typing import Dict, List, Optional, Tuple
 import torch
 import uvloop
 from PIL import Image
 from tqdm import tqdm
 from transformers import (AutoModelForCausalLM, AutoTokenizer,
                          PreTrainedTokenizerBase)
-from vllm.engine.arg_utils import DEVICE_OPTIONS, AsyncEngineArgs, EngineArgs
+from vllm.engine.arg_utils import AsyncEngineArgs, EngineArgs
 from vllm.entrypoints.openai.api_server import (
    build_async_engine_client_from_engine_args)
-from vllm.model_executor.layers.quantization import QUANTIZATION_METHODS
+from vllm.inputs import TextPrompt
 from vllm.lora.request import LoRARequest
 from vllm.lora.utils import get_adapter_absolute_path
 from vllm.multimodal import MultiModalDataDict
 from vllm.sampling_params import BeamSearchParams
 from vllm.transformers_utils.tokenizer import AnyTokenizer, get_lora_tokenizer
 from vllm.utils import FlexibleArgumentParser, merge_async_iterators
-def sample_requests(
+@dataclasses.dataclass
-    dataset_path: str,
+class SampleRequest:
-    num_requests: int,
+    """A class representing a single inference request for benchmarking.
-    tokenizer: PreTrainedTokenizerBase,
+
-    fixed_output_len: Optional[int],
+    Attributes:
-) -> List[Tuple[str, int, int]]:
+        prompt: The input text prompt for the model.
        prompt_len: The length of the prompt in tokens.
        expected_output_len: The expected length of the output in tokens.
        multi_modal_data: Optional dictionary containing multi-modal data (e.g.
            images).
        lora_request: Optional LoRARequest specifying the LoRA to use. 
    """
    prompt: str
    prompt_len: int
    expected_output_len: int
    multi_modal_data: Optional[MultiModalDataDict] = None
    lora_request: Optional[LoRARequest] = None
 def _get_prompt_for_image_model(question: str, *, model: str) -> str:
    """Prepend and append special tokens around the question to form a prompt.
    Args:
        question: The input question text to wrap with special tokens
        model: The name of the model being used, to determine which special
            tokens to add
    Returns:
        The formatted prompt string with appropriate special tokens for the
            model
    Raises:
        ValueError: If an unsupported model name is provided
    """
    model = model.lower()
    if "pixtral" in model:
        return f"<s>[INST]{question}\n[IMG][/INST]"
    raise ValueError(f"Unsupported model {model}")
@cache
 def lora_path_on_disk(lora_path: str) -> str:
    return get_adapter_absolute_path(lora_path)
 lora_tokenizer_cache: Dict[int, AnyTokenizer] = {}
 def get_random_lora_request(
        args: argparse.Namespace
 ) -> Tuple[LoRARequest, Optional[AnyTokenizer]]:
    global lora_tokenizer_cache
    lora_id = random.randint(1, args.max_loras)
    lora_request = LoRARequest(lora_name=str(lora_id),
                               lora_int_id=lora_id,
                               lora_path=lora_path_on_disk(args.lora_path))
    if lora_id not in lora_tokenizer_cache:
        lora_tokenizer_cache[lora_id] = get_lora_tokenizer(lora_request)
    return lora_request, lora_tokenizer_cache[lora_id]
 def sample_requests(tokenizer: PreTrainedTokenizerBase,
                    args: argparse.Namespace) -> List[SampleRequest]:
    dataset_path: str = args.dataset
    num_requests: int = args.num_prompts
    fixed_output_len: Optional[int] = args.output_len
    model: str = args.model
    if fixed_output_len is not None and fixed_output_len < 4:
        raise ValueError("output_len too small")
@@ -33,24 +103,46 @@ def sample_requests(
        dataset = json.load(f)
    # Filter out the conversations with less than 2 turns.
    dataset = [data for data in dataset if len(data["conversations"]) >= 2]
    # Only keep the first two turns of each conversation.
    dataset = [(data["conversations"][0]["value"],
                data["conversations"][1]["value"]) for data in dataset]
    # Shuffle the dataset.
    random.shuffle(dataset)
    # Filter out sequences that are too long or too short
-    filtered_dataset: List[Tuple[str, int, int]] = []
+    filtered_dataset: List[SampleRequest] = []
-    for i in range(len(dataset)):
+    for data in tqdm(dataset,
                     total=len(filtered_dataset),
                     desc="sampling requests"):
        if len(filtered_dataset) == num_requests:
            break
        # Only keep the first two turns of each conversation.
        prompt = data["conversations"][0]["value"]
        completion = data["conversations"][1]["value"]
        multi_modal_data: Optional[MultiModalDataDict] = None
        if "image" in data:
            multi_modal_data = multi_modal_data or {}
            image_path = data["image"]
            # TODO(vllm-project/vllm/issues/9778): Support multiple images.
            assert isinstance(image_path,
                              str), "Only support single image input"
            try:
                multi_modal_data["image"] = Image.open(image_path).convert(
                    "RGB")
            except FileNotFoundError:
                # Ignore datapoint where asset is missing
                continue
            prompt = _get_prompt_for_image_model(question=prompt, model=model)
        request_tokenizer = tokenizer
        lora_request: Optional[LoRARequest] = None
        if args.enable_lora:
            lora_request, lora_tokenizer = get_random_lora_request(args)
            if lora_tokenizer:
                request_tokenizer = lora_tokenizer
        # Tokenize the prompts and completions.
-        prompt = dataset[i][0]
+        prompt_token_ids = request_tokenizer(prompt).input_ids
-        prompt_token_ids = tokenizer(prompt).input_ids
+        completion_token_ids = request_tokenizer(completion).input_ids
        completion = dataset[i][1]
        completion_token_ids = tokenizer(completion).input_ids
        prompt_len = len(prompt_token_ids)
        output_len = len(completion_token_ids
                         ) if fixed_output_len is None else fixed_output_len
@@ -60,87 +152,59 @@ def sample_requests(
        if prompt_len > 1024 or prompt_len + output_len > 2048:
            # Prune too long sequences.
            continue
-        filtered_dataset.append((prompt, prompt_len, output_len))
+        filtered_dataset.append(
            SampleRequest(prompt=prompt,
                          prompt_len=prompt_len,
                          expected_output_len=output_len,
                          multi_modal_data=multi_modal_data,
                          lora_request=lora_request))
    return filtered_dataset
 def run_vllm(
-    requests: List[Tuple[str, int, int]],
+    requests: List[SampleRequest],
    model: str,
    tokenizer: str,
    quantization: Optional[str],
    tensor_parallel_size: int,
    seed: int,
    n: int,
-    trust_remote_code: bool,
+    engine_args: EngineArgs,
    dtype: str,
    max_model_len: Optional[int],
    enforce_eager: bool,
    kv_cache_dtype: str,
    quantization_param_path: Optional[str],
    device: str,
    enable_prefix_caching: bool,
    enable_chunked_prefill: bool,
    max_num_batched_tokens: int,
    distributed_executor_backend: Optional[str],
    gpu_memory_utilization: float = 0.9,
    num_scheduler_steps: int = 1,
    download_dir: Optional[str] = None,
    load_format: str = EngineArgs.load_format,
    disable_async_output_proc: bool = False,
 ) -> float:
    from vllm import LLM, SamplingParams
-    llm = LLM(
+    llm = LLM(**dataclasses.asdict(engine_args))
        model=model,
        tokenizer=tokenizer,
        quantization=quantization,
        tensor_parallel_size=tensor_parallel_size,
        seed=seed,
        trust_remote_code=trust_remote_code,
        dtype=dtype,
        max_model_len=max_model_len,
        gpu_memory_utilization=gpu_memory_utilization,
        enforce_eager=enforce_eager,
        kv_cache_dtype=kv_cache_dtype,
        quantization_param_path=quantization_param_path,
        device=device,
        enable_prefix_caching=enable_prefix_caching,
        download_dir=download_dir,
        enable_chunked_prefill=enable_chunked_prefill,
        max_num_batched_tokens=max_num_batched_tokens,
        distributed_executor_backend=distributed_executor_backend,
        load_format=load_format,
        num_scheduler_steps=num_scheduler_steps,
        disable_async_output_proc=disable_async_output_proc,
    )
    # Add the requests to the engine.
-    prompts: List[str] = []
+    prompts: List[TextPrompt] = []
    sampling_params: List[SamplingParams] = []
-    for prompt, _, output_len in requests:
+    for request in requests:
-        prompts.append(prompt)
+        prompts.append(
            TextPrompt(prompt=request.prompt,
                       multi_modal_data=request.multi_modal_data))
        sampling_params.append(
            SamplingParams(
                n=n,
                temperature=1.0,
                top_p=1.0,
                ignore_eos=True,
-                max_tokens=output_len,
+                max_tokens=request.expected_output_len,
            ))
    lora_requests: Optional[List[LoRARequest]] = None
    if engine_args.enable_lora:
        lora_requests = [request.lora_request for request in requests]
    use_beam_search = False
    if not use_beam_search:
        start = time.perf_counter()
-        llm.generate(prompts, sampling_params, use_tqdm=True)
+        llm.generate(prompts,
                     sampling_params,
                     lora_request=lora_requests,
                     use_tqdm=True)
        end = time.perf_counter()
    else:
-        prompts = [prompt for prompt, _, _ in requests]
+        assert lora_requests is None, "BeamSearch API does not support LoRA"
        prompts = [request.prompt for request in requests]
        # output_len should be the same for all requests.
        output_len = requests[0][2]
-        for prompt, input_len, _output_len in requests:
+        for request in requests:
-            assert _output_len == output_len
+            assert request.expected_output_len == output_len
        start = time.perf_counter()
        llm.beam_search(
            prompts,
@@ -154,79 +218,42 @@ def run_vllm(
 async def run_vllm_async(
-    requests: List[Tuple[str, int, int]],
+    requests: List[SampleRequest],
    model: str,
    tokenizer: str,
    quantization: Optional[str],
    tensor_parallel_size: int,
    seed: int,
    n: int,
-    trust_remote_code: bool,
+    engine_args: AsyncEngineArgs,
    dtype: str,
    max_model_len: Optional[int],
    enforce_eager: bool,
    kv_cache_dtype: str,
    quantization_param_path: Optional[str],
    device: str,
    enable_prefix_caching: bool,
    enable_chunked_prefill: bool,
    max_num_batched_tokens: int,
    distributed_executor_backend: Optional[str],
    gpu_memory_utilization: float = 0.9,
    num_scheduler_steps: int = 1,
    download_dir: Optional[str] = None,
    load_format: str = EngineArgs.load_format,
    disable_async_output_proc: bool = False,
    disable_frontend_multiprocessing: bool = False,
 ) -> float:
    from vllm import SamplingParams
    engine_args = AsyncEngineArgs(
        model=model,
        tokenizer=tokenizer,
        quantization=quantization,
        tensor_parallel_size=tensor_parallel_size,
        seed=seed,
        trust_remote_code=trust_remote_code,
        dtype=dtype,
        max_model_len=max_model_len,
        gpu_memory_utilization=gpu_memory_utilization,
        enforce_eager=enforce_eager,
        kv_cache_dtype=kv_cache_dtype,
        quantization_param_path=quantization_param_path,
        device=device,
        enable_prefix_caching=enable_prefix_caching,
        download_dir=download_dir,
        enable_chunked_prefill=enable_chunked_prefill,
        max_num_batched_tokens=max_num_batched_tokens,
        distributed_executor_backend=distributed_executor_backend,
        load_format=load_format,
        num_scheduler_steps=num_scheduler_steps,
        disable_async_output_proc=disable_async_output_proc,
        worker_use_ray=False,
        disable_log_requests=True,
    )
    async with build_async_engine_client_from_engine_args(
            engine_args, disable_frontend_multiprocessing) as llm:
        # Add the requests to the engine.
-        prompts: List[str] = []
+        prompts: List[TextPrompt] = []
        sampling_params: List[SamplingParams] = []
-        for prompt, _, output_len in requests:
+        lora_requests: List[Optional[LoRARequest]] = []
-            prompts.append(prompt)
+        for request in requests:
            prompts.append(
                TextPrompt(prompt=request.prompt,
                           multi_modal_data=request.multi_modal_data))
            sampling_params.append(
                SamplingParams(
                    n=n,
                    temperature=1.0,
                    top_p=1.0,
                    ignore_eos=True,
-                    max_tokens=output_len,
+                    max_tokens=request.expected_output_len,
                ))
            lora_requests.append(request.lora_request)
        generators = []
        start = time.perf_counter()
-        for i, (prompt, sp) in enumerate(zip(prompts, sampling_params)):
+        for i, (prompt, sp,
-            generator = llm.generate(prompt, sp, request_id=f"test{i}")
+                lr) in enumerate(zip(prompts, sampling_params, lora_requests)):
            generator = llm.generate(prompt,
                                     sp,
                                     lora_request=lr,
                                     request_id=f"test{i}")
            generators.append(generator)
        all_gens = merge_async_iterators(*generators)
        async for i, res in all_gens:
@@ -236,7 +263,7 @@ async def run_vllm_async(
 def run_hf(
-    requests: List[Tuple[str, int, int]],
+    requests: List[SampleRequest],
    model: str,
    tokenizer: PreTrainedTokenizerBase,
    n: int,
@@ -294,14 +321,14 @@ def run_hf(
 def run_mii(
-    requests: List[Tuple[str, int, int]],
+    requests: List[SampleRequest],
    model: str,
    tensor_parallel_size: int,
    output_len: int,
 ) -> float:
    from mii import client, serve
    llm = serve(model, tensor_parallel=tensor_parallel_size)
-    prompts = [prompt for prompt, _, _ in requests]
+    prompts = [request.prompt for request in requests]
    start = time.perf_counter()
    llm.generate(prompts, max_new_tokens=output_len)
@@ -319,32 +346,62 @@ def main(args: argparse.Namespace):
    tokenizer = AutoTokenizer.from_pretrained(
        args.tokenizer, trust_remote_code=args.trust_remote_code)
    if args.dataset is None:
-        # Synthesize a prompt with the given input length.
+        vocab_size = tokenizer.vocab_size
-        prompt = "hi" * (args.input_len - 1)
+        requests = []
-        requests = [(prompt, args.input_len, args.output_len)
+        for _ in range(args.num_prompts):
-                    for _ in range(args.num_prompts)]
+
            request_tokenizer = tokenizer
            lora_request: Optional[LoRARequest] = None
            if args.enable_lora:
                lora_request, lora_tokenizer = get_random_lora_request(args)
                if lora_tokenizer:
                    request_tokenizer = lora_tokenizer
            # Synthesize a prompt with the given input length.
            candidate_ids = [
                random.randint(0, vocab_size - 1)
                for _ in range(args.input_len)
            ]
            # As tokenizer may add additional tokens like BOS, we need to try
            # different lengths to get the desired input length.
            for _ in range(5):  # Max attempts to correct
                candidate_prompt = request_tokenizer.decode(candidate_ids)
                tokenized_len = len(request_tokenizer.encode(candidate_prompt))
                if tokenized_len == args.input_len:
                    break
                # Adjust length based on difference
                diff = args.input_len - tokenized_len
                if diff > 0:
                    candidate_ids.extend([
                        random.randint(100, vocab_size - 100)
                        for _ in range(diff)
                    ])
                else:
                    candidate_ids = candidate_ids[:diff]
            requests.append(
                SampleRequest(prompt=candidate_prompt,
                              prompt_len=args.input_len,
                              expected_output_len=args.output_len,
                              lora_request=lora_request))
    else:
-        requests = sample_requests(args.dataset, args.num_prompts, tokenizer,
+        requests = sample_requests(tokenizer, args)
                                   args.output_len)
    is_multi_modal = any(request.multi_modal_data is not None
                         for request in requests)
    if args.backend == "vllm":
        run_args = [
            requests, args.model, args.tokenizer, args.quantization,
            args.tensor_parallel_size, args.seed, args.n,
            args.trust_remote_code, args.dtype, args.max_model_len,
            args.enforce_eager, args.kv_cache_dtype,
            args.quantization_param_path, args.device,
            args.enable_prefix_caching, args.enable_chunked_prefill,
            args.max_num_batched_tokens, args.distributed_executor_backend,
            args.gpu_memory_utilization, args.num_scheduler_steps,
            args.download_dir, args.load_format, args.disable_async_output_proc
        ]
        if args.async_engine:
-            run_args.append(args.disable_frontend_multiprocessing)
+            elapsed_time = uvloop.run(
-            elapsed_time = uvloop.run(run_vllm_async(*run_args))
+                run_vllm_async(
                    requests,
                    args.n,
                    AsyncEngineArgs.from_cli_args(args),
                    args.disable_frontend_multiprocessing,
                ))
        else:
-            elapsed_time = run_vllm(*run_args)
+            elapsed_time = run_vllm(requests, args.n,
                                    EngineArgs.from_cli_args(args))
    elif args.backend == "hf":
        assert args.tensor_parallel_size == 1
        elapsed_time = run_hf(requests, args.model, tokenizer, args.n,
@@ -354,10 +411,18 @@ def main(args: argparse.Namespace):
                               args.output_len)
    else:
        raise ValueError(f"Unknown backend: {args.backend}")
-    total_num_tokens = sum(prompt_len + output_len
+    total_num_tokens = sum(request.prompt_len + request.expected_output_len
-                           for _, prompt_len, output_len in requests)
+                           for request in requests)
    total_output_tokens = sum(request.expected_output_len
                              for request in requests)
    if is_multi_modal:
        print("\033[91mWARNING\033[0m: Multi-modal request detected. The "
              "following metrics are not accurate because image tokens are not"
              " counted. See vllm-project/vllm/issues/9778 for details.")
        # TODO(vllm-project/vllm/issues/9778): Count molti-modal token length.
    print(f"Throughput: {len(requests) / elapsed_time:.2f} requests/s, "
-          f"{total_num_tokens / elapsed_time:.2f} tokens/s")
+          f"{total_num_tokens / elapsed_time:.2f} total tokens/s, "
          f"{total_output_tokens / elapsed_time:.2f} output tokens/s")
    # Output JSON results if specified
    if args.output_json:
@@ -381,7 +446,9 @@ if __name__ == "__main__":
    parser.add_argument("--dataset",
                        type=str,
                        default=None,
-                        help="Path to the dataset.")
+                        help="Path to the dataset. The dataset is expected to "
                        "be a json in form of List[Dict[..., conversations: "
                        "List[Dict[..., value: <prompt_or_response>]]]]")
    parser.add_argument("--input-len",
                        type=int,
                        default=None,
@@ -391,13 +458,6 @@ if __name__ == "__main__":
                        default=None,
                        help="Output length for each request. Overrides the "
                        "output length from the dataset.")
    parser.add_argument("--model", type=str, default="facebook/opt-125m")
    parser.add_argument("--tokenizer", type=str, default=None)
    parser.add_argument('--quantization',
                        '-q',
                        choices=[*QUANTIZATION_METHODS, None],
                        default=None)
    parser.add_argument("--tensor-parallel-size", "-tp", type=int, default=1)
    parser.add_argument("--n",
                        type=int,
                        default=1,
@@ -406,123 +466,15 @@ if __name__ == "__main__":
                        type=int,
                        default=1000,
                        help="Number of prompts to process.")
    parser.add_argument("--seed", type=int, default=0)
    parser.add_argument("--hf-max-batch-size",
                        type=int,
                        default=None,
                        help="Maximum batch size for HF backend.")
    parser.add_argument('--trust-remote-code',
                        action='store_true',
                        help='trust remote code from huggingface')
    parser.add_argument(
        '--max-model-len',
        type=int,
        default=None,
        help='Maximum length of a sequence (including prompt and output). '
        'If None, will be derived from the model.')
    parser.add_argument(
        '--dtype',
        type=str,
        default='auto',
        choices=['auto', 'half', 'float16', 'bfloat16', 'float', 'float32'],
        help='data type for model weights and activations. '
        'The "auto" option will use FP16 precision '
        'for FP32 and FP16 models, and BF16 precision '
        'for BF16 models.')
    parser.add_argument('--gpu-memory-utilization',
                        type=float,
                        default=0.9,
                        help='the fraction of GPU memory to be used for '
                        'the model executor, which can range from 0 to 1.'
                        'If unspecified, will use the default value of 0.9.')
    parser.add_argument("--enforce-eager",
                        action="store_true",
                        help="enforce eager execution")
    parser.add_argument(
        '--kv-cache-dtype',
        type=str,
        choices=['auto', 'fp8', 'fp8_e5m2', 'fp8_e4m3'],
        default="auto",
        help='Data type for kv cache storage. If "auto", will use model '
        'data type. CUDA 11.8+ supports fp8 (=fp8_e4m3) and fp8_e5m2. '
        'ROCm (AMD GPU) supports fp8 (=fp8_e4m3)')
    parser.add_argument(
        '--quantization-param-path',
        type=str,
        default=None,
        help='Path to the JSON file containing the KV cache scaling factors. '
        'This should generally be supplied, when KV cache dtype is FP8. '
        'Otherwise, KV cache scaling factors default to 1.0, which may cause '
        'accuracy issues. FP8_E5M2 (without scaling) is only supported on '
        'cuda version greater than 11.8. On ROCm (AMD GPU), FP8_E4M3 is '
        'instead supported for common inference criteria.')
    parser.add_argument("--device",
                        type=str,
                        default="auto",
                        choices=DEVICE_OPTIONS,
                        help='device type for vLLM execution')
    parser.add_argument(
        "--num-scheduler-steps",
        type=int,
        default=1,
        help="Maximum number of forward steps per scheduler call.")
    parser.add_argument(
        "--enable-prefix-caching",
        action='store_true',
        help="Enable automatic prefix caching for vLLM backend.")
    parser.add_argument("--enable-chunked-prefill",
                        action='store_true',
                        help="enable chunked prefill for vLLM backend.")
    parser.add_argument('--max-num-batched-tokens',
                        type=int,
                        default=None,
                        help='maximum number of batched tokens per '
                        'iteration')
    parser.add_argument('--download-dir',
                        type=str,
                        default=None,
                        help='directory to download and load the weights, '
                        'default to the default cache dir of huggingface')
    parser.add_argument(
        '--output-json',
        type=str,
        default=None,
        help='Path to save the throughput results in JSON format.')
    parser.add_argument(
        '--distributed-executor-backend',
        choices=['ray', 'mp'],
        default=None,
        help='Backend to use for distributed serving. When more than 1 GPU '
        'is used, will be automatically set to "ray" if installed '
        'or "mp" (multiprocessing) otherwise.')
    parser.add_argument(
        '--load-format',
        type=str,
        default=EngineArgs.load_format,
        choices=[
            'auto', 'pt', 'safetensors', 'npcache', 'dummy', 'tensorizer',
            'bitsandbytes'
        ],
        help='The format of the model weights to load.\n\n'
        '* "auto" will try to load the weights in the safetensors format '
        'and fall back to the pytorch bin format if safetensors format '
        'is not available.\n'
        '* "pt" will load the weights in the pytorch bin format.\n'
        '* "safetensors" will load the weights in the safetensors format.\n'
        '* "npcache" will load the weights in pytorch format and store '
        'a numpy cache to speed up the loading.\n'
        '* "dummy" will initialize the weights with random values, '
        'which is mainly for profiling.\n'
        '* "tensorizer" will load the weights using tensorizer from '
        'CoreWeave. See the Tensorize vLLM Model script in the Examples'
        'section for more information.\n'
        '* "bitsandbytes" will load the weights using bitsandbytes '
        'quantization.\n')
    parser.add_argument(
        "--disable-async-output-proc",
        action='store_true',
        default=False,
        help="Disable async output processor for vLLM backend.")
    parser.add_argument("--async-engine",
                        action='store_true',
                        default=False,
@@ -531,6 +483,15 @@ if __name__ == "__main__":
                        action='store_true',
                        default=False,
                        help="Disable decoupled async engine frontend.")
    # LoRA
    parser.add_argument(
        "--lora-path",
        type=str,
        default=None,
        help="Path to the lora adapters to use. This can be an absolute path, "
        "a relative path, or a Hugging Face model identifier.")
    parser = AsyncEngineArgs.add_cli_args(parser)
    args = parser.parse_args()
    if args.tokenizer is None:
        args.tokenizer = args.model
@@ -539,6 +500,8 @@ if __name__ == "__main__":
        assert args.output_len is not None
    else:
        assert args.input_len is None
    if args.enable_lora:
        assert args.lora_path is not None
    if args.backend == "vllm":
        if args.hf_max_batch_size is not None:
@@ -548,6 +511,9 @@ if __name__ == "__main__":
            raise ValueError("HF max batch size is required for HF backend.")
        if args.quantization is not None:
            raise ValueError("Quantization is only for vLLM backend.")
        if args.enable_lora is not None:
            raise ValueError("LoRA benchmarking is only supported for vLLM"
                             " backend")
    elif args.backend == "mii":
        if args.dtype != "auto":
            raise ValueError("dtype must be auto for MII backend.")
@@ -560,4 +526,7 @@ if __name__ == "__main__":
        if args.tokenizer != args.model:
            raise ValueError("Tokenizer must be the same as the model for MII "
                             "backend.")
        if args.enable_lora is not None:
            raise ValueError("LoRA benchmarking is only supported for vLLM"
                             " backend")
    main(args)
--- a/benchmarks/cutlass_benchmarks/sparse_benchmarks.py
+++ b/benchmarks/cutlass_benchmarks/sparse_benchmarks.py
@@ -0,0 +1,386 @@
 # 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
 import torch
 import torch.utils.benchmark as TBenchmark
 from torch.utils.benchmark import Measurement as TMeasurement
 from utils import make_rand_sparse_tensors
 from weight_shapes import WEIGHT_SHAPES
 from vllm import _custom_ops as ops
 from vllm.utils import FlexibleArgumentParser
 DEFAULT_MODELS = list(WEIGHT_SHAPES.keys())
 DEFAULT_BATCH_SIZES = [1, 16, 32, 64, 128, 256, 512]
 DEFAULT_TP_SIZES = [1]
 # bench
 def bench_fn(label: str, sub_label: str, description: str, fn: Callable, *args,
             **kwargs) -> TMeasurement:
    min_run_time = 1
    globals = {
        "args": args,
        "kwargs": kwargs,
        "fn": fn,
    }
    return TBenchmark.Timer(
        stmt="fn(*args, **kwargs)",
        globals=globals,
        label=label,
        sub_label=sub_label,
        description=description,
    ).blocked_autorange(min_run_time=min_run_time)
 def bench_int8(dtype: torch.dtype, m: int, k: int, n: int, label: str,
               sub_label: str) -> Iterable[TMeasurement]:
    assert dtype == torch.int8
    b_compressed, e, a, b = make_rand_sparse_tensors(torch.int8, m, n, k)
    scale_a = torch.tensor(1.0, device="cuda", dtype=torch.float32)
    scale_b = torch.tensor(1.0, device="cuda", dtype=torch.float32)
    bias = torch.zeros((n, ), device="cuda", dtype=torch.bfloat16)
    out = ops.cutlass_scaled_sparse_mm(a, b_compressed, e, scale_a, scale_b,
                                       torch.bfloat16)
    out_ref = ops.cutlass_scaled_mm(a, b, scale_a, scale_b, torch.bfloat16)
    if not torch.allclose(out, out_ref):
        print("Incorrect results")
        print(out)
        print(out_ref)
    else:
        print("Correct results")
    timers = []
    # pytorch impl - bfloat16
    timers.append(
        bench_fn(label, sub_label, "pytorch_bf16_bf16_bf16_matmul-no-scales",
                 torch.mm, a.to(dtype=torch.bfloat16),
                 b.to(dtype=torch.bfloat16)))
    # 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 sparse impl
    timers.append(
        bench_fn(label, sub_label, "cutlass_i8_i8_bf16_scaled_sparse_mm",
                 ops.cutlass_scaled_sparse_mm, a, b_compressed, e, scale_a,
                 scale_b, torch.bfloat16))
    # cutlass sparse with bias
    timers.append(
        bench_fn(label, sub_label, "cutlass_i8_i8_bf16_scaled_sparse_mm_bias",
                 ops.cutlass_scaled_sparse_mm, a, b_compressed, e, scale_a,
                 scale_b, torch.bfloat16, bias))
    return timers
 def bench_fp8(dtype: torch.dtype, m: int, k: int, n: int, label: str,
              sub_label: str) -> Iterable[TMeasurement]:
    assert dtype == torch.float8_e4m3fn
    b_compressed, e, a, b = make_rand_sparse_tensors(torch.float8_e4m3fn, m, n,
                                                     k)
    scale_a = torch.tensor(1.0, device="cuda", dtype=torch.float32)
    scale_b = torch.tensor(1.0, device="cuda", dtype=torch.float32)
    bias = torch.zeros((n, ), device="cuda", dtype=torch.bfloat16)
    out = ops.cutlass_scaled_sparse_mm(a, b_compressed, e, scale_a, scale_b,
                                       torch.bfloat16)
    out_ref = ops.cutlass_scaled_mm(a, b, scale_a, scale_b, torch.bfloat16)
    if not torch.allclose(out, out_ref):
        print("Incorrect results")
        print(out)
        print(out_ref)
    else:
        print("Correct results")
    timers = []
    # pytorch impl w. bf16
    timers.append(
        bench_fn(label, sub_label, "pytorch_bf16_bf16_bf16_matmul-no-scales",
                 torch.mm, a.to(dtype=torch.bfloat16, device="cuda"),
                 b.to(dtype=torch.bfloat16, device="cuda")))
    # pytorch impl: bf16 output, without fp8 fast accum
    timers.append(
        bench_fn(label,
                 sub_label,
                 "pytorch_fp8_fp8_bf16_scaled_mm",
                 torch._scaled_mm,
                 a,
                 b,
                 scale_a=scale_a,
                 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))
    # cutlass impl: bf16 output
    timers.append(
        bench_fn(label, sub_label, "cutlass_fp8_fp8_bf16_scaled_mm",
                 ops.cutlass_scaled_mm, a, b, scale_a, scale_b,
                 torch.bfloat16))
    # cutlass impl: bf16 output
    timers.append(
        bench_fn(label, sub_label, "cutlass_fp8_fp8_bf16_scaled_sparse_mm",
                 ops.cutlass_scaled_sparse_mm, a, b_compressed, e, scale_a,
                 scale_b, torch.bfloat16))
    # cutlass impl: fp16 output
    timers.append(
        bench_fn(label, sub_label, "cutlass_fp8_fp8_fp16_scaled_sparse_mm",
                 ops.cutlass_scaled_sparse_mm, a, b_compressed, e, scale_a,
                 scale_b, torch.float16))
    # cutlass impl: bf16 output, with bias
    timers.append(
        bench_fn(label, sub_label,
                 "cutlass_fp8_fp8_bf16_scaled_sparse_mm_bias",
                 ops.cutlass_scaled_sparse_mm, a, b_compressed, e, 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_sparse_mm_bias",
                 ops.cutlass_scaled_sparse_mm, a, b_compressed, e, 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,
          sub_label: str) -> Iterable[TMeasurement]:
    if dtype == torch.int8:
        return bench_int8(dtype, m, k, n, label, sub_label)
    if dtype == torch.float8_e4m3fn:
        return bench_fp8(dtype, m, k, n, label, sub_label)
    raise ValueError("unsupported type")
 # runner
 def print_timers(timers: Iterable[TMeasurement]):
    compare = TBenchmark.Compare(timers)
    compare.print()
 def run(dtype: torch.dtype,
        MKNs: Iterable[Tuple[int, int, int]]) -> Iterable[TMeasurement]:
    results = []
    for m, k, n in MKNs:
        timers = bench(dtype, m, k, n, f"scaled-{dtype}-gemm",
                       f"MKN=({m}x{k}x{n})")
        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,
                timestamp=None):
    print(f"== All Results {base_description} ====")
    print_timers(data)
    # pickle all the results
    timestamp = int(time.time()) if timestamp is None else timestamp
    with open(f"{base_description}-{timestamp}.pkl", "wb") as f:
        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)
    make_output(data, MKNs, f"square_bench-{args.dtype}")
 def run_range_bench(args):
    dim_sizes = list(range(args.dim_start, args.dim_end, args.dim_increment))
    n = len(dim_sizes)
    Ms = [args.m_constant] * n if args.m_constant is not None else dim_sizes
    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)
    make_output(data, MKNs, f"range_bench-{args.dtype}")
 def run_model_bench(args):
    print("Benchmarking models:")
    for i, model in enumerate(args.models):
        print(f"[{i}]  {model}")
    def model_shapes(model_name: str, tp_size: int) -> List[Tuple[int, int]]:
        KNs = []
        for KN, tp_split_dim in copy.deepcopy(WEIGHT_SHAPES[model_name]):
            KN[tp_split_dim] = KN[tp_split_dim] // tp_size
            KNs.append(KN)
        return KNs
    model_bench_data = []
    models_tps = list(itertools.product(args.models, args.tp_sizes))
    for model, tp_size in models_tps:
        Ms = args.batch_sizes
        KNs = model_shapes(model, tp_size)
        MKNs = []
        for m in Ms:
            for k, n in KNs:
                MKNs.append((m, k, n))
        data = run(args.dtype, MKNs)
        model_bench_data.append(data)
    # Print all results
    for data, model_tp in zip(model_bench_data, models_tps):
        model, tp_size = model_tp
        print(f"== Results {args.dtype} {model}-TP{tp_size} ====")
        print_timers(data)
    timestamp = int(time.time())
    all_data = []
    for d in model_bench_data:
        all_data.extend(d)
    # pickle all data
    with open(f"model_bench-{args.dtype}-{timestamp}.pkl", "wb") as f:
        pkl.dump(all_data, f)
 if __name__ == '__main__':
    def to_torch_dtype(dt):
        if dt == "int8":
            return torch.int8
        if dt == "fp8":
            return torch.float8_e4m3fn
        raise ValueError("unsupported dtype")
    parser = FlexibleArgumentParser(
        description="""
 Benchmark Cutlass GEMM.
    To run square GEMMs:
        python3 ./benchmarks/cutlass_benchmarks/sparse_benchmarks.py --dtype fp8 square_bench --dim-start 128 --dim-end 512 --dim-increment 64
    To run constant N and K and sweep M:
        python3 ./benchmarks/cutlass_benchmarks/sparse_benchmarks.py --dtype fp8 range_bench --dim-start 128 --dim-end 512 --dim-increment 64 --n-constant 16384 --k-constant 16384
    To run dimensions from a model:
        python3 ./benchmarks/cutlass_benchmarks/sparse_benchmarks.py --dtype fp8 model_bench --models meta-llama/Llama-2-7b-hf --batch-sizes 16 --tp-sizes 1
    Output:
        - a .pkl file, that is a list of raw torch.benchmark.utils.Measurements for the pytorch and cutlass implementations for the various GEMMs.
            """,  # noqa: E501
        formatter_class=argparse.RawTextHelpFormatter)
    parser.add_argument("--dtype",
                        type=to_torch_dtype,
                        required=True,
                        help="Available options are ['int8', 'fp8']")
    subparsers = parser.add_subparsers(dest="cmd")
    square_parser = subparsers.add_parser("square_bench")
    square_parser.add_argument("--dim-start", type=int, required=True)
    square_parser.add_argument("--dim-end", type=int, required=True)
    square_parser.add_argument("--dim-increment", type=int, required=True)
    square_parser.set_defaults(func=run_square_bench)
    range_parser = subparsers.add_parser("range_bench")
    range_parser.add_argument("--dim-start", type=int, required=True)
    range_parser.add_argument("--dim-end", type=int, required=True)
    range_parser.add_argument("--dim-increment", type=int, required=True)
    range_parser.add_argument("--m-constant", type=int, default=None)
    range_parser.add_argument("--n-constant", type=int, default=None)
    range_parser.add_argument("--k-constant", type=int, default=None)
    range_parser.set_defaults(func=run_range_bench)
    model_parser = subparsers.add_parser("model_bench")
    model_parser.add_argument("--models",
                              nargs="+",
                              type=str,
                              default=DEFAULT_MODELS,
                              choices=WEIGHT_SHAPES.keys())
    model_parser.add_argument("--tp-sizes",
                              nargs="+",
                              type=int,
                              default=DEFAULT_TP_SIZES)
    model_parser.add_argument("--batch-sizes",
                              nargs="+",
                              type=int,
                              default=DEFAULT_BATCH_SIZES)
    model_parser.set_defaults(func=run_model_bench)
    args = parser.parse_args()
    args.func(args)
--- a/benchmarks/cutlass_benchmarks/utils.py
+++ b/benchmarks/cutlass_benchmarks/utils.py
@@ -0,0 +1,98 @@
 # SPDX-License-Identifier: Apache-2.0
 # Cutlass bench utils
 from typing import Iterable, Tuple
 import torch
 import vllm._custom_ops as ops
 def to_fp8(tensor: torch.Tensor) -> torch.Tensor:
    finfo = torch.finfo(torch.float8_e4m3fn)
    return torch.round(tensor.clamp(
        min=finfo.min, max=finfo.max)).to(dtype=torch.float8_e4m3fn)
 def to_int8(tensor: torch.Tensor) -> torch.Tensor:
    return torch.round(tensor.clamp(min=-128, max=127)).to(dtype=torch.int8)
 def to_bf16(tensor: torch.Tensor) -> torch.Tensor:
    return tensor.to(dtype=torch.bfloat16)
 def to_fp16(tensor: torch.Tensor) -> torch.Tensor:
    return tensor.to(dtype=torch.float16)
 def make_rand_tensors(dtype: torch.dtype, m: int, n: int,
                      k: int) -> Tuple[torch.Tensor, torch.Tensor]:
    a = torch.randn((m, k), device='cuda') * 5
    b = torch.randn((n, k), device='cuda').t() * 5
    if dtype == torch.int8:
        return to_int8(a), to_int8(b)
    if dtype == torch.float8_e4m3fn:
        return to_fp8(a), to_fp8(b)
    raise ValueError("unsupported dtype")
 def prune_to_2_4(tensor):
    # Reshape tensor to [N, 4] where N is number of groups of 4
    original_shape = tensor.shape
    reshaped = tensor.reshape(-1, 4)
    # Get indices of top 2 absolute values in each group of 4
    _, indices = torch.topk(torch.abs(reshaped), k=2, dim=1)
    # Create binary mask
    mask = torch.zeros_like(reshaped)
    mask.scatter_(dim=1,
                  index=indices,
                  src=torch.ones_like(indices, dtype=mask.dtype))
    # Apply mask and reshape back
    pruned = reshaped * mask
    # Turn all -0.0 to 0.0
    pruned[pruned == -0.0] = 0.0
    return pruned.reshape(original_shape)
 def make_rand_sparse_tensors(dtype: torch.dtype, m: int, n: int,
                             k: int) -> Tuple[torch.Tensor, torch.Tensor]:
    a = torch.randn((m, k), device='cuda') * 5
    b = torch.randn((n, k), device='cuda').t() * 5
    b = prune_to_2_4(b.t()).t()
    if dtype == torch.int8:
        a, b = to_int8(a), to_int8(b)
    elif dtype == torch.float8_e4m3fn:
        a, b = to_fp8(a), to_fp8(b)
    elif dtype == torch.float16:
        a, b = to_fp16(a), to_fp16(b)
    elif dtype == torch.bfloat16:
        a, b = to_bf16(a), to_bf16(b)
    else:
        raise ValueError("unsupported dtype")
    b_compressed, e = ops.cutlass_sparse_compress(b.t())
    # Compressed B, Metadata, Original A, B
    return b_compressed, e, a, b
 def make_n_rand_sparse_tensors(num_tensors: int, dtype: torch.dtype,
                        m: int, n: int, k: int) -> \
                        Tuple[Iterable[torch.Tensor], Iterable[torch.Tensor]]:
    ABs = []
    for _ in range(num_tensors):
        b_comp, e, a, b = make_rand_sparse_tensors(dtype, m, n, k)
        if b_comp is not None:
            ABs.append(make_rand_sparse_tensors(dtype, m, n, k))
    BComps, Es, As, Bs = zip(*ABs)
    return list(BComps), list(Es), list(As), list(Bs)
--- a/benchmarks/cutlass_benchmarks/w8a8_benchmarks.py
+++ b/benchmarks/cutlass_benchmarks/w8a8_benchmarks.py
@@ -1,47 +1,27 @@
 # 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
 from torch.utils.benchmark import Measurement as TMeasurement
 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())
 DEFAULT_BATCH_SIZES = [1, 16, 32, 64, 128, 256, 512]
 DEFAULT_TP_SIZES = [1]
 # helpers
 def to_fp8(tensor: torch.Tensor) -> torch.Tensor:
    finfo = torch.finfo(torch.float8_e4m3fn)
    return torch.round(tensor.clamp(
        min=finfo.min, max=finfo.max)).to(dtype=torch.float8_e4m3fn)
 def to_int8(tensor: torch.Tensor) -> torch.Tensor:
    return torch.round(tensor.clamp(min=-128, max=127)).to(dtype=torch.int8)
 def make_rand_tensors(dtype: torch.dtype, m: int, n: int,
                      k: int) -> Tuple[torch.Tensor, torch.Tensor]:
    a = torch.randn((m, k), device='cuda') * 5
    b = torch.randn((n, k), device='cuda').t() * 5
    if dtype == torch.int8:
        return to_int8(a), to_int8(b)
    if dtype == torch.float8_e4m3fn:
        return to_fp8(a), to_fp8(b)
    raise ValueError("unsupported dtype")
 # bench
 def bench_fn(label: str, sub_label: str, description: str, fn: Callable, *args,
@@ -62,8 +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)
@@ -72,155 +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)
    print(m, k, n)
    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)),
        "pytorch_fp8_fp8_fp16_scaled_mm":
        lambda: torch._scaled_mm(
            a, b, scale_a, scale_b, out_dtype=torch.float16),
        "pytorch_fp8_fp8_fp16_scaled_mm_fast_accum":
        lambda: torch._scaled_mm(a,
                                 b,
                                 scale_a,
                                 scale_b,
                                 out_dtype=torch.float16,
                                 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),
    }
    timers = []
-
+    for name, fn in bench_fns.items():
-    # pytorch impl w. bf16
+        # If bench_kernels is None, run all. Otherwise, run only exact matches.
-    timers.append(
+        if bench_kernels is None or name in bench_kernels:
-        bench_fn(label, sub_label, "pytorch_bf16_bf16_bf16_matmul-no-scales",
+            print(f"Running {name}")
-                 torch.mm, a.to(dtype=torch.bfloat16, device="cuda"),
+            timers.append(bench_fn(label, sub_label, name, fn))
                 b.to(dtype=torch.bfloat16, device="cuda")))
    # pytorch impl: bf16 output, without fp8 fast accum
    timers.append(
        bench_fn(label,
                 sub_label,
                 "pytorch_fp8_fp8_bf16_scaled_mm",
                 torch._scaled_mm,
                 a,
                 b,
                 scale_a=scale_a,
                 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))
    # cutlass impl: bf16 output
    timers.append(
        bench_fn(label, sub_label, "cutlass_fp8_fp8_bf16_scaled_mm",
                 ops.cutlass_scaled_mm, a, b, scale_a, scale_b,
                 torch.bfloat16))
    # cutlass impl: fp16 output
    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")
@@ -231,18 +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,
@@ -256,15 +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}")
@@ -275,8 +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}")
@@ -302,7 +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
@@ -352,6 +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")
--- a/benchmarks/cutlass_benchmarks/weight_shapes.py
+++ b/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:
--- a/benchmarks/disagg_benchmarks/disagg_overhead_benchmark.sh
+++ b/benchmarks/disagg_benchmarks/disagg_overhead_benchmark.sh
@@ -0,0 +1,145 @@
 #!/bin/bash
 # benchmark the overhead of disaggregated prefill.
 # methodology:
 # - send all request to prefill vLLM instance. It will buffer KV cache.
 # - then send all request to decode instance. 
 # - The TTFT of decode instance is the overhead.
 set -ex
 kill_gpu_processes() {
  # kill all processes on GPU.
  pgrep pt_main_thread | xargs -r kill -9
  pgrep python3 | xargs -r kill -9
  sleep 10
  # remove vllm config file
  rm -rf ~/.config/vllm
  # Print the GPU memory usage
  # so that we know if all GPU processes are killed.
  gpu_memory_usage=$(nvidia-smi --query-gpu=memory.used --format=csv,noheader,nounits -i 0)
  # The memory usage should be 0 MB.
  echo "GPU 0 Memory Usage: $gpu_memory_usage MB"
 }
 wait_for_server() {
  # wait for vllm server to start
  # return 1 if vllm server crashes
  local port=$1
  timeout 1200 bash -c "
    until curl -s localhost:${port}/v1/completions > /dev/null; do
      sleep 1
    done" && return 0 || return 1
 }
 benchmark() {
  export VLLM_LOGGING_LEVEL=DEBUG
  export VLLM_HOST_IP=$(hostname -I | awk '{print $1}')
  # compare chunked prefill with disaggregated prefill
  results_folder="./results"
  model="meta-llama/Meta-Llama-3.1-8B-Instruct"
  dataset_name="sonnet"
  dataset_path="../sonnet_4x.txt"
  num_prompts=10
  qps=$1
  prefix_len=50
  input_len=2048
  output_len=$2
  CUDA_VISIBLE_DEVICES=0 python3 \
    -m vllm.entrypoints.openai.api_server \
    --model $model \
    --port 8100 \
    --max-model-len 10000 \
    --gpu-memory-utilization 0.6 \
    --kv-transfer-config \
    '{"kv_connector":"PyNcclConnector","kv_role":"kv_producer","kv_rank":0,"kv_parallel_size":2,"kv_buffer_size":5e9}' &
  CUDA_VISIBLE_DEVICES=1 python3 \
    -m vllm.entrypoints.openai.api_server \
    --model $model \
    --port 8200 \
    --max-model-len 10000 \
    --gpu-memory-utilization 0.6 \
    --kv-transfer-config \
    '{"kv_connector":"PyNcclConnector","kv_role":"kv_consumer","kv_rank":1,"kv_parallel_size":2,"kv_buffer_size":5e9}' &
  wait_for_server 8100
  wait_for_server 8200
  # let the prefill instance finish prefill
  python3 ../benchmark_serving.py \
          --backend vllm \
          --model $model \
          --dataset-name $dataset_name \
          --dataset-path $dataset_path \
          --sonnet-input-len $input_len \
          --sonnet-output-len "$output_len" \
          --sonnet-prefix-len $prefix_len \
          --num-prompts $num_prompts \
          --port 8100 \
          --save-result \
          --result-dir $results_folder \
          --result-filename disagg_prefill_tp1.json \
          --request-rate "inf"
  # send the request to decode.
  # The TTFT of this command will be the overhead of disagg prefill impl.
  python3 ../benchmark_serving.py \
          --backend vllm \
          --model $model \
          --dataset-name $dataset_name \
          --dataset-path $dataset_path \
          --sonnet-input-len $input_len \
          --sonnet-output-len "$output_len" \
          --sonnet-prefix-len $prefix_len \
          --num-prompts $num_prompts \
          --port 8200 \
          --save-result \
          --result-dir $results_folder \
          --result-filename disagg_prefill_tp1_overhead.json \
          --request-rate "$qps"
  kill_gpu_processes
 }
 main() {
  (which wget && which curl) || (apt-get update && apt-get install -y wget curl)
  (which jq) || (apt-get -y install jq)
  (which socat) || (apt-get -y install socat)
  pip install quart httpx datasets
  cd "$(dirname "$0")"
  cd ..
  # create sonnet-4x.txt
  echo "" > sonnet_4x.txt
  for _ in {1..4}
  do
    cat sonnet.txt >> sonnet_4x.txt
  done
  cd disagg_benchmarks
  rm -rf results
  mkdir results
  default_qps=1
  default_output_len=1
  benchmark $default_qps $default_output_len
 }
 main "$@"
--- a/benchmarks/disagg_benchmarks/disagg_performance_benchmark.sh
+++ b/benchmarks/disagg_benchmarks/disagg_performance_benchmark.sh
@@ -0,0 +1,163 @@
 #!/bin/bash
 # Requirement: 2x GPUs.
 # Model: meta-llama/Meta-Llama-3.1-8B-Instruct
 # Query: 1024 input tokens, 6 output tokens, QPS 2/4/6/8, 100 requests
 # Resource: 2x GPU
 # Approaches:
 # 2. Chunked prefill: 2 vllm instance with tp=4, equivalent to 1 tp=4 instance with QPS 4
 # 3. Disaggregated prefill: 1 prefilling instance and 1 decoding instance
 # Prefilling instance: max_output_token=1
 # Decoding instance: force the input tokens be the same across requests to bypass prefilling
 set -ex
 kill_gpu_processes() {
  # kill all processes on GPU.
  pgrep pt_main_thread | xargs -r kill -9
  pgrep python3 | xargs -r kill -9
  for port in 8000 8100 8200; do lsof -t -i:$port | xargs -r kill -9; done
  sleep 1
 }
 wait_for_server() {
  # wait for vllm server to start
  # return 1 if vllm server crashes
  local port=$1
  timeout 1200 bash -c "
    until curl -s localhost:${port}/v1/completions > /dev/null; do
      sleep 1
    done" && return 0 || return 1
 }
 launch_chunked_prefill() {
  model="meta-llama/Meta-Llama-3.1-8B-Instruct"
  # disagg prefill
  CUDA_VISIBLE_DEVICES=0 python3 \
    -m vllm.entrypoints.openai.api_server \
    --model $model \
    --port 8100 \
    --max-model-len 10000 \
    --enable-chunked-prefill \
    --gpu-memory-utilization 0.6 &
  CUDA_VISIBLE_DEVICES=1 python3 \
    -m vllm.entrypoints.openai.api_server \
    --model $model \
    --port 8200 \
    --max-model-len 10000 \
    --enable-chunked-prefill \
    --gpu-memory-utilization 0.6 &
  wait_for_server 8100
  wait_for_server 8200
  python3 round_robin_proxy.py &
  sleep 1
 }
 launch_disagg_prefill() {
  model="meta-llama/Meta-Llama-3.1-8B-Instruct" 
  # disagg prefill
  CUDA_VISIBLE_DEVICES=0 python3 \
    -m vllm.entrypoints.openai.api_server \
    --model $model \
    --port 8100 \
    --max-model-len 10000 \
    --gpu-memory-utilization 0.6 \
    --kv-transfer-config \
    '{"kv_connector":"PyNcclConnector","kv_role":"kv_producer","kv_rank":0,"kv_parallel_size":2,"kv_buffer_size":5e9}' &
  CUDA_VISIBLE_DEVICES=1 python3 \
    -m vllm.entrypoints.openai.api_server \
    --model $model \
    --port 8200 \
    --max-model-len 10000 \
    --gpu-memory-utilization 0.6 \
    --kv-transfer-config \
    '{"kv_connector":"PyNcclConnector","kv_role":"kv_consumer","kv_rank":1,"kv_parallel_size":2,"kv_buffer_size":5e9}' &
  wait_for_server 8100
  wait_for_server 8200
  python3 disagg_prefill_proxy_server.py &
  sleep 1
 }
 benchmark() {
  results_folder="./results"
  model="meta-llama/Meta-Llama-3.1-8B-Instruct"
  dataset_name="sonnet"
  dataset_path="../sonnet_4x.txt"
  num_prompts=100
  qps=$1
  prefix_len=50
  input_len=1024
  output_len=$2
  tag=$3
  python3 ../benchmark_serving.py \
          --backend vllm \
          --model $model \
          --dataset-name $dataset_name \
          --dataset-path $dataset_path \
          --sonnet-input-len $input_len \
          --sonnet-output-len "$output_len" \
          --sonnet-prefix-len $prefix_len \
          --num-prompts $num_prompts \
          --port 8000 \
          --save-result \
          --result-dir $results_folder \
          --result-filename "$tag"-qps-"$qps".json \
          --request-rate "$qps"
  sleep 2
 }
 main() {
  (which wget && which curl) || (apt-get update && apt-get install -y wget curl)
  (which jq) || (apt-get -y install jq)
  (which socat) || (apt-get -y install socat)
  (which lsof) || (apt-get -y install lsof)
  pip install quart httpx matplotlib aiohttp datasets
  cd "$(dirname "$0")"
  cd ..
  # create sonnet-4x.txt so that we can sample 2048 tokens for input
  echo "" > sonnet_4x.txt
  for _ in {1..4}
  do
    cat sonnet.txt >> sonnet_4x.txt
  done
  cd disagg_benchmarks
  rm -rf results
  mkdir results
  default_output_len=6
  export VLLM_HOST_IP=$(hostname -I | awk '{print $1}')
  launch_chunked_prefill
  for qps in 2 4 6 8; do
  benchmark $qps $default_output_len chunked_prefill
  done
  kill_gpu_processes
  launch_disagg_prefill
  for qps in 2 4 6 8; do
  benchmark $qps $default_output_len disagg_prefill
  done
  kill_gpu_processes
  python3 visualize_benchmark_results.py
 }
 main "$@"
--- a/benchmarks/disagg_benchmarks/disagg_prefill_proxy_server.py
+++ b/benchmarks/disagg_benchmarks/disagg_prefill_proxy_server.py
@@ -0,0 +1,63 @@
 # SPDX-License-Identifier: Apache-2.0
 import os
 import aiohttp
 from quart import Quart, make_response, request
 AIOHTTP_TIMEOUT = aiohttp.ClientTimeout(total=6 * 60 * 60)
 app = Quart(__name__)
 async def forward_request(url, data):
    async with aiohttp.ClientSession(timeout=AIOHTTP_TIMEOUT) as session:
        headers = {
            "Authorization": f"Bearer {os.environ.get('OPENAI_API_KEY')}"
        }
        async with session.post(url=url, json=data,
                                headers=headers) as response:
            if response.status == 200:
                # if response.headers.get('Transfer-Encoding') == 'chunked':
                if True:
                    async for chunk_bytes in response.content.iter_chunked(
                            1024):
                        yield chunk_bytes
                else:
                    content = await response.read()
                    yield content
@app.route('/v1/completions', methods=['POST'])
 async def handle_request():
    try:
        original_request_data = await request.get_json()
        prefill_request = original_request_data.copy()
        # change max_tokens = 1 to let it only do prefill
        prefill_request['max_tokens'] = 1
        # finish prefill
        async for _ in forward_request('http://localhost:8100/v1/completions',
                                       prefill_request):
            continue
        # return decode
        generator = forward_request('http://localhost:8200/v1/completions',
                                    original_request_data)
        response = await make_response(generator)
        response.timeout = None
        return response
    except Exception as e:
        import sys
        import traceback
        exc_info = sys.exc_info()
        print("Error occurred in disagg prefill proxy server")
        print(e)
        print("".join(traceback.format_exception(*exc_info)))
 if __name__ == '__main__':
    app.run(port=8000)
--- a/benchmarks/disagg_benchmarks/round_robin_proxy.py
+++ b/benchmarks/disagg_benchmarks/round_robin_proxy.py
@@ -0,0 +1,62 @@
 # SPDX-License-Identifier: Apache-2.0
 import asyncio
 import itertools
 import aiohttp
 from aiohttp import web
 class RoundRobinProxy:
    def __init__(self, target_ports):
        self.target_ports = target_ports
        self.port_cycle = itertools.cycle(self.target_ports)
    async def handle_request(self, request):
        target_port = next(self.port_cycle)
        target_url = f"http://localhost:{target_port}{request.path_qs}"
        async with aiohttp.ClientSession() as session:
            try:
                # Forward the request
                async with session.request(
                        method=request.method,
                        url=target_url,
                        headers=request.headers,
                        data=request.content,
                ) as response:
                    # Start sending the response
                    resp = web.StreamResponse(status=response.status,
                                              headers=response.headers)
                    await resp.prepare(request)
                    # Stream the response content
                    async for chunk in response.content.iter_any():
                        await resp.write(chunk)
                    await resp.write_eof()
                    return resp
            except Exception as e:
                return web.Response(text=f"Error: {str(e)}", status=500)
 async def main():
    proxy = RoundRobinProxy([8100, 8200])
    app = web.Application()
    app.router.add_route('*', '/{path:.*}', proxy.handle_request)
    runner = web.AppRunner(app)
    await runner.setup()
    site = web.TCPSite(runner, 'localhost', 8000)
    await site.start()
    print("Proxy server started on http://localhost:8000")
    # Keep the server running
    await asyncio.Event().wait()
 if __name__ == '__main__':
    asyncio.run(main())
--- a/Show More
+++ b/Show More
`@@ -1,2 +1,2 @@`
	`github: [vllm-project]`	`github: [vllm-project]`
	`open_collective: [vllm]`	`open_collective: vllm`