Enable v1 metrics tests (#20953)

Signed-off-by: Seiji Eicher <seiji@anyscale.com>

.buildkite/check-wheel-size.py

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

.buildkite/generate_index.py

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

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

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

.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,add_bos_token=true,trust_remote_code=true,max_model_len=4096" \
   --tasks gsm8k --num_fewshot "$FEWSHOT" --limit "$LIMIT" \
   --batch_size "$BATCH_SIZE"

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

@@ -1,4 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 """
 LM eval harness on model to compare vs HF baseline computed offline.
 Configs are found in configs/$MODEL.yaml
@@ -17,12 +18,14 @@ RTOL = 0.08
 
 def launch_lm_eval(eval_config, tp_size):
     trust_remote_code = eval_config.get("trust_remote_code", False)
+    max_model_len = eval_config.get("max_model_len", 4096)
     model_args = (
         f"pretrained={eval_config['model_name']},"
         f"tensor_parallel_size={tp_size},"
         f"enforce_eager=true,"
         f"add_bos_token=true,"
-        f"trust_remote_code={trust_remote_code}"
+        f"trust_remote_code={trust_remote_code},"
+        f"max_model_len={max_model_len}"
     )
     results = lm_eval.simple_evaluate(
         model="vllm",

.buildkite/nightly-benchmarks/README.md

@@ -11,7 +11,7 @@ See [vLLM performance dashboard](https://perf.vllm.ai) for the latest performanc
 
 ## Performance benchmark quick overview
 
-**Benchmarking Coverage**: latency, throughput and fix-qps serving on A100 (the support for FP8 benchmark on H100 is coming!), with different models.
+**Benchmarking Coverage**: latency, throughput and fix-qps serving on A100 (the support for FP8 benchmark on H100 is coming!) and Intel® Xeon® Processors, with different models.
 
 **Benchmarking Duration**: about 1hr.
 
@@ -31,13 +31,27 @@ Performance benchmark will be triggered when:
 - A PR being merged into vllm.
 - Every commit for those PRs with `perf-benchmarks` label AND `ready` label.
 
+Manually Trigger the benchmark
+
+```bash
+bash .buildkite/nightly-benchmarks/scripts/run-performance-benchmarks.sh
+```
+
+Runtime environment variables:
+- `ON_CPU`: set the value to '1' on Intel® Xeon® Processors. Default value is 0.
+- `SERVING_JSON`: JSON file to use for the serving tests. Default value is empty string (use default file).
+- `LATENCY_JSON`: JSON file to use for the latency tests. Default value is empty string (use default file).
+- `THROUGHPUT_JSON`: JSON file to use for the throughout tests. Default value is empty string (use default file).
+- `REMOTE_HOST`: IP for the remote vLLM service to benchmark. Default value is empty string.
+- `REMOTE_PORT`: Port for the remote vLLM service to benchmark. Default value is empty string.
+
 Nightly benchmark will be triggered when:
 - Every commit for those PRs with `perf-benchmarks` label and `nightly-benchmarks` label.
 
 ## Performance benchmark details
 
 See [performance-benchmarks-descriptions.md](performance-benchmarks-descriptions.md) for detailed descriptions, and use `tests/latency-tests.json`, `tests/throughput-tests.json`, `tests/serving-tests.json` to configure the test cases.
-
+> NOTE: For Intel® Xeon® Processors, use `tests/latency-tests-cpu.json`, `tests/throughput-tests-cpu.json`, `tests/serving-tests-cpu.json` instead.
 ### Latency test
 
 Here is an example of one test inside `latency-tests.json`:
@@ -119,6 +133,30 @@ If you do not see the table, please wait till the benchmark finish running.
 The json version of the table (together with the json version of the benchmark) will be also attached to the markdown file.
 The raw benchmarking results (in the format of json files) are in the `Artifacts` tab of the benchmarking.
 
+The `compare-json-results.py` helps to compare benchmark results JSON files converted using `convert-results-json-to-markdown.py`.
+When run, benchmark script generates results under `benchmark/results` folder, along with the `benchmark_results.md` and `benchmark_results.json`.
+`compare-json-results.py` compares two `benchmark_results.json` files and provides performance ratio e.g. for Output Tput, Median TTFT and Median TPOT.
+
+Here is an example using the script to compare result_a and result_b without detail test name.
+`python3 compare-json-results.py -f results_a/benchmark_results.json -f results_b/benchmark_results.json --ignore_test_name`
+
+|    | results_a/benchmark_results.json | results_b/benchmark_results.json | perf_ratio        |
+|----|----------------------------------------|----------------------------------------|----------|
+| 0  | 142.633982                             | 156.526018                             | 1.097396 |
+| 1  | 241.620334                             | 294.018783                             | 1.216863 |
+| 2  | 218.298905                             | 262.664916                             | 1.203235 |
+| 3  | 242.743860                             | 299.816190                             | 1.235113 |
+
+Here is an example using the script to compare result_a and result_b with detail test name.
+`python3 compare-json-results.py -f results_a/benchmark_results.json -f results_b/benchmark_results.json`
+|   | results_a/benchmark_results.json_name | results_a/benchmark_results.json | results_b/benchmark_results.json_name | results_b/benchmark_results.json | perf_ratio        |
+|---|---------------------------------------------|----------------------------------------|---------------------------------------------|----------------------------------------|----------|
+| 0 | serving_llama8B_tp1_sharegpt_qps_1          | 142.633982                             | serving_llama8B_tp1_sharegpt_qps_1          | 156.526018                             | 1.097396 |
+| 1 | serving_llama8B_tp1_sharegpt_qps_16         | 241.620334                             | serving_llama8B_tp1_sharegpt_qps_16         | 294.018783                             | 1.216863 |
+| 2 | serving_llama8B_tp1_sharegpt_qps_4          | 218.298905                             | serving_llama8B_tp1_sharegpt_qps_4          | 262.664916                             | 1.203235 |
+| 3 | serving_llama8B_tp1_sharegpt_qps_inf        | 242.743860                             | serving_llama8B_tp1_sharegpt_qps_inf        | 299.816190                             | 1.235113 |
+| 4 | serving_llama8B_tp2_random_1024_128_qps_1   | 96.613390                              | serving_llama8B_tp4_random_1024_128_qps_1   | 108.404853                             | 1.122048 |
+
 ## Nightly test details
 
 See [nightly-descriptions.md](nightly-descriptions.md) for the detailed description on test workload, models and docker containers of benchmarking other llm engines.

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

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

.buildkite/nightly-benchmarks/performance-benchmarks-descriptions.md

@@ -4,7 +4,8 @@
 - Input length: 32 tokens.
 - Output length: 128 tokens.
 - Batch size: fixed (8).
-- Models: llama-3.1 8B, llama-3 70B, mixtral 8x7B.
+- GPU Models: llama-3.1 8B, llama-3 70B, mixtral 8x7B.
+- CPU Models: llama-3.1 8B.
 - Evaluation metrics: end-to-end latency (mean, median, p99).
 
 {latency_tests_markdown_table}
@@ -14,7 +15,8 @@
 - Input length: randomly sample 200 prompts from ShareGPT dataset (with fixed random seed).
 - Output length: the corresponding output length of these 200 prompts.
 - Batch size: dynamically determined by vllm to achieve maximum throughput.
-- Models: llama-3.1 8B, llama-3 70B, mixtral 8x7B.
+- GPU Models: llama-3.1 8B, llama-3 70B, mixtral 8x7B.
+- CPU Models: llama-3.1 8B.
 - Evaluation metrics: throughput.
 
 {throughput_tests_markdown_table}
@@ -25,12 +27,18 @@
 - Output length: the corresponding output length of these 200 prompts.
 - Batch size: dynamically determined by vllm and the arrival pattern of the requests.
 - **Average QPS (query per second)**: 1, 4, 16 and inf. QPS = inf means all requests come at once. For other QPS values, the arrival time of each query is determined using a random Poisson process (with fixed random seed).
-- Models: llama-3.1 8B, llama-3 70B, mixtral 8x7B.
-- We also added a speculative decoding test for llama-3 70B, under QPS 2
+- GPU Models: llama-3.1 8B, llama-3 70B, mixtral 8x7B.
+- We also added a speculative decoding test for llama-3 70B on GPU, under QPS 2
+- CPU Models: llama-3.1 8B.
 - Evaluation metrics: throughput, TTFT (time to the first token, with mean, median and p99), ITL (inter-token latency, with mean, median and p99).
+- For CPU, we added random dataset tests to benchmark fixed input/output length with 100 prompts.
 
 {serving_tests_markdown_table}
 
+## Platform Information
+
+{platform_markdown_table}
+
 ## json version of the benchmarking tables
 
 This section contains the data of the markdown tables above in JSON format.

.buildkite/nightly-benchmarks/scripts/compare-json-results.py

@@ -0,0 +1,66 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+import argparse
+
+import pandas as pd
+
+
+def compare_data_columns(
+    files, name_column, data_column, drop_column, ignore_test_name=False
+):
+    print("\ncompare_data_column: " + data_column)
+    frames = []
+    compare_frames = []
+    for file in files:
+        data_df = pd.read_json(file)
+        serving_df = data_df.dropna(subset=[drop_column], ignore_index=True)
+        if ignore_test_name is False:
+            serving_df = serving_df.rename(columns={name_column: file + "_name"})
+            frames.append(serving_df[file + "_name"])
+        serving_df = serving_df.rename(columns={data_column: file})
+        frames.append(serving_df[file])
+        compare_frames.append(serving_df[file])
+        if len(compare_frames) >= 2:
+            # Compare numbers among two files
+            ratio_df = compare_frames[1] / compare_frames[0]
+            frames.append(ratio_df)
+            compare_frames.pop(1)
+
+    concat_df = pd.concat(frames, axis=1)
+    return concat_df
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "-f", "--file", action="append", type=str, help="input file name"
+    )
+    parser.add_argument(
+        "--ignore_test_name", action="store_true", help="ignore_test_name or not"
+    )
+    args = parser.parse_args()
+    files = args.file
+    print("comparing : " + ", ".join(files))
+
+    drop_column = "P99"
+    name_column = "Test name"
+    data_cols_to_compare = ["Output Tput (tok/s)", "Median TTFT (ms)", "Median"]
+    html_msgs_for_data_cols = [
+        "Compare Output Tokens /n",
+        "Median TTFT /n",
+        "Median TPOT /n",
+    ]
+    ignore_test_name = args.ignore_test_name
+    with open("perf_comparison.html", "w") as text_file:
+        for i in range(len(data_cols_to_compare)):
+            output_df = compare_data_columns(
+                files,
+                name_column,
+                data_cols_to_compare[i],
+                drop_column,
+                ignore_test_name=ignore_test_name,
+            )
+            print(output_df)
+            html = output_df.to_html()
+            text_file.write(html_msgs_for_data_cols[i])
+            text_file.write(html)

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

@@ -1,10 +1,13 @@
 # SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 import json
 import os
+from importlib import util
 from pathlib import Path
 
 import pandas as pd
+import psutil
 from tabulate import tabulate
 
 results_folder = Path("results/")
@@ -28,11 +31,11 @@ throughput_results = []
 throughput_results_column_mapping = {
     "test_name": "Test name",
     "gpu_type": "GPU",
-    # "num_requests": "# of req.",
-    # "total_num_tokens": "Total # of tokens",
-    # "elapsed_time": "Elapsed time (s)",
+    "num_requests": "# of req.",
+    "total_num_tokens": "Total # of tokens",
+    "elapsed_time": "Elapsed time (s)",
     "requests_per_second": "Tput (req/s)",
-    # "tokens_per_second": "Tput (tok/s)",
+    "tokens_per_second": "Tput (tok/s)",
 }
 
 # serving results and the keys that will be printed into markdown
@@ -40,16 +43,18 @@ serving_results = []
 serving_column_mapping = {
     "test_name": "Test name",
     "gpu_type": "GPU",
-    # "completed": "# of req.",
+    "completed": "# of req.",
     "request_throughput": "Tput (req/s)",
-    # "input_throughput": "Input Tput (tok/s)",
-    # "output_throughput": "Output Tput (tok/s)",
+    "total_token_throughput": "Total Token Tput (tok/s)",
+    "output_throughput": "Output Tput (tok/s)",
+    "total_input_tokens": "Total input tokens",
+    "total_output_tokens": "Total output tokens",
     "mean_ttft_ms": "Mean TTFT (ms)",
     "median_ttft_ms": "Median TTFT (ms)",
     "p99_ttft_ms": "P99 TTFT (ms)",
-    # "mean_tpot_ms": "Mean TPOT (ms)",
-    # "median_tpot_ms": "Median",
-    # "p99_tpot_ms": "P99",
+    "mean_tpot_ms": "Mean TPOT (ms)",
+    "median_tpot_ms": "Median",
+    "p99_tpot_ms": "P99",
     "mean_itl_ms": "Mean ITL (ms)",
     "median_itl_ms": "Median ITL (ms)",
     "p99_itl_ms": "P99 ITL (ms)",
@@ -74,6 +79,20 @@ def results_to_json(latency, throughput, serving):
     )
 
 
+def get_size_with_unit(bytes, suffix="B"):
+    """
+    Scale bytes to its proper format
+    e.g:
+        1253656 => '1.20MB'
+        1253656678 => '1.17GB'
+    """
+    factor = 1024
+    for unit in ["", "K", "M", "G", "T", "P"]:
+        if bytes < factor:
+            return f"{bytes:.2f}{unit}{suffix}"
+        bytes /= factor
+
+
 if __name__ == "__main__":
     # collect results
     for test_file in results_folder.glob("*.json"):
@@ -154,6 +173,27 @@ if __name__ == "__main__":
     serving_results = pd.DataFrame.from_dict(serving_results)
     throughput_results = pd.DataFrame.from_dict(throughput_results)
 
+    svmem = psutil.virtual_memory()
+    platform_data = {
+        "Physical cores": [psutil.cpu_count(logical=False)],
+        "Total cores": [psutil.cpu_count(logical=True)],
+        "Total Memory": [get_size_with_unit(svmem.total)],
+    }
+
+    if util.find_spec("numa") is not None:
+        from numa import info
+
+        platform_data["Total NUMA nodes"] = [info.get_num_configured_nodes()]
+
+    if util.find_spec("cpuinfo") is not None:
+        from cpuinfo import get_cpu_info
+
+        platform_data["CPU Brand"] = [get_cpu_info()["brand_raw"]]
+
+    platform_results = pd.DataFrame.from_dict(
+        platform_data, orient="index", columns=["Platform Info"]
+    )
+
     raw_results_json = results_to_json(
         latency_results, throughput_results, serving_results
     )
@@ -199,6 +239,9 @@ if __name__ == "__main__":
     throughput_md_table = tabulate(
         throughput_results, headers="keys", tablefmt="pipe", showindex=False
     )
+    platform_md_table = tabulate(
+        platform_results, headers="keys", tablefmt="pipe", showindex=True
+    )
 
     # document the result
     with open(results_folder / "benchmark_results.md", "w") as f:
@@ -210,6 +253,7 @@ if __name__ == "__main__":
             latency_tests_markdown_table=latency_md_table,
             throughput_tests_markdown_table=throughput_md_table,
             serving_tests_markdown_table=serving_md_table,
+            platform_markdown_table=platform_md_table,
             benchmarking_results_in_json_string=processed_results_json,
         )
         f.write(results)

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

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

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

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

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

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

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

@@ -31,6 +31,20 @@ check_gpus() {
   echo "GPU type is $gpu_type"
 }
 
+check_cpus() {
+  # check the number of CPUs and NUMA Node and GPU type.
+  declare -g numa_count=$(python3 -c  "from numa import info;numa_size = info.get_num_configured_nodes(); print(numa_size)")
+  if [[ $numa_count -gt 0 ]]; then
+    echo "NUMA found."
+    echo $numa_count
+  else
+    echo "Need at least 1 NUMA to run benchmarking."
+    exit 1
+  fi
+  declare -g gpu_type="cpu"
+  echo "GPU type is $gpu_type"
+}
+
 check_hf_token() {
   # check if HF_TOKEN is available and valid
   if [[ -z "$HF_TOKEN" ]]; then
@@ -69,6 +83,22 @@ json2args() {
   echo "$args"
 }
 
+json2envs() {
+  # transforms the JSON string to environment variables.
+  # example:
+  # input: { "VLLM_CPU_KVCACHE_SPACE": 5 }
+  # output: VLLM_CPU_KVCACHE_SPACE=5
+  local json_string=$1
+  local args=$(
+    echo "$json_string" | jq -r '
+      to_entries |
+      map((.key ) + "=" + (.value | tostring)) |
+      join(" ")
+    '
+  )
+  echo "$args"
+}
+
 wait_for_server() {
   # wait for vllm server to start
   # return 1 if vllm server crashes
@@ -158,15 +188,24 @@ run_latency_tests() {
     # get arguments
     latency_params=$(echo "$params" | jq -r '.parameters')
     latency_args=$(json2args "$latency_params")
+    latency_environment_variables=$(echo "$params" | jq -r '.environment_variables')
+    latency_envs=$(json2envs "$latency_environment_variables")
 
     # 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 $test_name."
-      continue
+    if [ "$ON_CPU" == "1" ];then
+      if [[ $numa_count -lt $tp ]]; then
+        echo "Required tensor-parallel-size $tp but only $numa_count NUMA nodes found. Skip testcase $test_name."
+        continue
+      fi
+    else
+      if [[ $gpu_count -lt $tp ]]; then
+        echo "Required tensor-parallel-size $tp but only $gpu_count GPU found. Skip testcase $test_name."
+        continue
+      fi
     fi
 
-    latency_command="python3 benchmark_latency.py \
+    latency_command=" $latency_envs python3 benchmark_latency.py \
       --output-json $RESULTS_FOLDER/${test_name}.json \
       $latency_args"
 
@@ -216,15 +255,24 @@ run_throughput_tests() {
     # get arguments
     throughput_params=$(echo "$params" | jq -r '.parameters')
     throughput_args=$(json2args "$throughput_params")
+    throughput_environment_variables=$(echo "$params" | jq -r '.environment_variables')
+    throughput_envs=$(json2envs "$throughput_environment_variables")
 
     # check if there is enough GPU to run the test
     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 $test_name."
-      continue
+    if [ "$ON_CPU" == "1" ];then
+      if [[ $numa_count -lt $tp ]]; then
+        echo "Required tensor-parallel-size $tp but only $numa_count NUMA nodes found. Skip testcase $test_name."
+        continue
+      fi
+    else
+      if [[ $gpu_count -lt $tp ]]; then
+        echo "Required tensor-parallel-size $tp but only $gpu_count GPU found. Skip testcase $test_name."
+        continue
+      fi
     fi
 
-    throughput_command="python3 benchmark_throughput.py \
+    throughput_command=" $throughput_envs python3 benchmark_throughput.py \
       --output-json $RESULTS_FOLDER/${test_name}.json \
       $throughput_args"
 
@@ -272,18 +320,27 @@ run_serving_tests() {
 
     # get client and server arguments
     server_params=$(echo "$params" | jq -r '.server_parameters')
+    server_envs=$(echo "$params" | jq -r '.server_environment_variables')
     client_params=$(echo "$params" | jq -r '.client_parameters')
     server_args=$(json2args "$server_params")
+    server_envs=$(json2envs "$server_envs")
     client_args=$(json2args "$client_params")
     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
+    # check if there is enough resources 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 $test_name."
-      continue
+    if [ "$ON_CPU" == "1" ];then
+      if [[ $numa_count -lt $tp ]]; then
+        echo "Required tensor-parallel-size $tp but only $numa_count NUMA nodes found. Skip testcase $test_name."
+        continue
+      fi
+    else
+      if [[ $gpu_count -lt $tp ]]; then
+        echo "Required tensor-parallel-size $tp but only $gpu_count GPU found. Skip testcase $test_name."
+        continue
+      fi
     fi
 
     # check if server model and client model is aligned
@@ -294,23 +351,33 @@ run_serving_tests() {
       continue
     fi
 
-    server_command="python3 \
+    server_command="$server_envs python3 \
       -m vllm.entrypoints.openai.api_server \
       $server_args"
 
     # run the server
     echo "Running test case $test_name"
     echo "Server command: $server_command"
-    bash -c "$server_command" &
-    server_pid=$!
-
-    # wait until the server is alive
-    if wait_for_server; then
-      echo ""
-      echo "vllm server is up and running."
+    # support remote vllm server
+    client_remote_args=""
+    if [[ -z "${REMOTE_HOST}" ]]; then
+      bash -c "$server_command" &
+      server_pid=$!
+      # wait until the server is alive
+      if wait_for_server; then
+        echo ""
+        echo "vLLM server is up and running."
+      else
+        echo ""
+        echo "vLLM failed to start within the timeout period."
+      fi
     else
-      echo ""
-      echo "vllm failed to start within the timeout period."
+      server_command="Using Remote Server $REMOTE_HOST $REMOTE_PORT"
+      if [[ ${REMOTE_PORT} ]]; then
+        client_remote_args=" --host=$REMOTE_HOST --port=$REMOTE_PORT "
+      else
+        client_remote_args=" --host=$REMOTE_HOST "
+      fi
     fi
 
     # iterate over different QPS
@@ -332,7 +399,7 @@ run_serving_tests() {
         --result-filename ${new_test_name}.json \
         --request-rate $qps \
         --metadata "tensor_parallel_size=$tp" \
-        $client_args"
+        $client_args $client_remote_args "
 
       echo "Running test case $test_name with qps $qps"
       echo "Client command: $client_command"
@@ -360,7 +427,14 @@ run_serving_tests() {
 }
 
 main() {
-  check_gpus
+  local ARCH
+  ARCH=''
+  if [ "$ON_CPU" == "1" ];then
+     check_cpus
+     ARCH='-cpu'
+  else
+     check_gpus
+  fi
   check_hf_token
 
   # Set to v1 to run v1 benchmark
@@ -386,9 +460,9 @@ main() {
   QUICK_BENCHMARK_ROOT=../.buildkite/nightly-benchmarks/
 
   # benchmarking
-  run_serving_tests $QUICK_BENCHMARK_ROOT/tests/serving-tests.json
-  run_latency_tests $QUICK_BENCHMARK_ROOT/tests/latency-tests.json
-  run_throughput_tests $QUICK_BENCHMARK_ROOT/tests/throughput-tests.json
+  run_serving_tests $QUICK_BENCHMARK_ROOT/tests/"${SERVING_JSON:-serving-tests$ARCH.json}"
+  run_latency_tests $QUICK_BENCHMARK_ROOT/tests/"${LATENCY_JSON:-latency-tests$ARCH.json}"
+  run_throughput_tests $QUICK_BENCHMARK_ROOT/tests/"${THROUGHPUT_JSON:-throughput-tests$ARCH.json}"
 
   # postprocess benchmarking results
   pip install tabulate pandas

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

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

.buildkite/nightly-benchmarks/tests/latency-tests-cpu.json

@@ -0,0 +1,30 @@
+[
+    {
+        "test_name": "latency_llama8B_tp1",
+        "environment_variables": {
+	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
+	    "VLLM_CPU_KVCACHE_SPACE": 40
+        },
+        "parameters": {
+            "model": "meta-llama/Meta-Llama-3.1-8B-Instruct",
+            "tensor_parallel_size": 1,
+            "load_format": "dummy",
+            "num_iters_warmup": 5,
+            "num_iters": 15
+        }
+    },
+    {
+        "test_name": "latency_llama8B_tp4",
+        "environment_variables": {
+	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
+	    "VLLM_CPU_KVCACHE_SPACE": 40
+        },
+        "parameters": {
+            "model": "meta-llama/Meta-Llama-3.1-8B-Instruct",
+            "tensor_parallel_size": 4,
+            "load_format": "dummy",
+            "num_iters_warmup": 5,
+            "num_iters": 15
+        }
+    }
+]

.buildkite/nightly-benchmarks/tests/serving-tests-cpu.json

@@ -0,0 +1,158 @@
+[
+    {
+        "test_name": "serving_llama8B_tp1_sharegpt",
+        "qps_list": [1, 4, 16, "inf"],
+        "server_environment_variables": {
+            "VLLM_RPC_TIMEOUT": 100000,
+	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
+	    "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
+	    "VLLM_CPU_KVCACHE_SPACE": 40
+        },
+        "server_parameters": {
+            "model": "meta-llama/Meta-Llama-3.1-8B-Instruct",
+            "tensor_parallel_size": 1,
+	    "dtype": "bfloat16",
+	    "distributed_executor_backend": "mp",
+	    "block_size": 128,
+	    "trust_remote_code": "",
+            "disable_log_stats": "",
+            "disable_log_requests": "",
+	    "enforce_eager": "",
+            "load_format": "dummy"
+        },
+        "client_parameters": {
+            "model": "meta-llama/Meta-Llama-3.1-8B-Instruct",
+            "backend": "vllm",
+            "dataset_name": "sharegpt",
+            "dataset_path": "./ShareGPT_V3_unfiltered_cleaned_split.json",
+	    "max_concurrency": 60,
+            "num_prompts": 200
+        }
+    },
+    {
+        "test_name": "serving_llama8B_tp2_sharegpt",
+        "qps_list": [1, 4, 16, "inf"],
+        "server_environment_variables": {
+            "VLLM_RPC_TIMEOUT": 100000,
+	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
+	    "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
+	    "VLLM_CPU_KVCACHE_SPACE": 40
+        },
+        "server_parameters": {
+            "model": "meta-llama/Meta-Llama-3.1-8B-Instruct",
+            "tensor_parallel_size": 2,
+	    "dtype": "bfloat16",
+	    "distributed_executor_backend": "mp",
+	    "block_size": 128,
+	    "trust_remote_code": "",
+            "disable_log_stats": "",
+            "disable_log_requests": "",
+	    "enforce_eager": "",
+            "load_format": "dummy"
+        },
+        "client_parameters": {
+            "model": "meta-llama/Meta-Llama-3.1-8B-Instruct",
+            "backend": "vllm",
+            "dataset_name": "sharegpt",
+            "dataset_path": "./ShareGPT_V3_unfiltered_cleaned_split.json",
+	    "max_concurrency": 60,
+            "num_prompts": 200
+        }
+    },
+    {
+        "test_name": "serving_llama8B_tp4_sharegpt",
+        "qps_list": [1, 4, 16, "inf"],
+        "server_environment_variables": {
+            "VLLM_RPC_TIMEOUT": 100000,
+	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
+	    "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
+	    "VLLM_CPU_KVCACHE_SPACE": 40
+        },
+        "server_parameters": {
+            "model": "meta-llama/Meta-Llama-3.1-8B-Instruct",
+            "tensor_parallel_size": 4,
+	    "dtype": "bfloat16",
+	    "distributed_executor_backend": "mp",
+	    "block_size": 128,
+	    "trust_remote_code": "",
+            "disable_log_stats": "",
+            "disable_log_requests": "",
+	    "enforce_eager": "",
+            "load_format": "dummy"
+        },
+        "client_parameters": {
+            "model": "meta-llama/Meta-Llama-3.1-8B-Instruct",
+            "backend": "vllm",
+            "dataset_name": "sharegpt",
+            "dataset_path": "./ShareGPT_V3_unfiltered_cleaned_split.json",
+	    "max_concurrency": 60,
+            "num_prompts": 200
+        }
+    },
+    {
+        "test_name": "serving_llama8B_tp4_random_1024_128",
+        "qps_list": [1, 4, 16, "inf"],
+        "server_environment_variables": {
+            "VLLM_RPC_TIMEOUT": 100000,
+	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
+	    "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
+	    "VLLM_CPU_KVCACHE_SPACE": 40
+        },
+        "server_parameters": {
+            "model": "meta-llama/Meta-Llama-3.1-8B-Instruct",
+            "tensor_parallel_size": 4,
+	    "dtype": "bfloat16",
+	    "distributed_executor_backend": "mp",
+	    "block_size": 128,
+	    "trust_remote_code": "",
+	    "enable_chunked_prefill": "",
+            "disable_log_stats": "",
+            "disable_log_requests": "",
+	    "enforce_eager": "",
+            "load_format": "dummy"
+        },
+        "client_parameters": {
+            "model": "meta-llama/Meta-Llama-3.1-8B-Instruct",
+            "backend": "vllm",
+            "dataset_name": "random",
+	    "random-input-len": 1024,
+	    "random-output-len": 128,
+	    "ignore-eos": "",
+	    "max_concurrency": 100,
+            "num_prompts": 100
+        }
+    },
+    {
+        "test_name": "serving_llama8B_pp6_random_1024_128",
+        "qps_list": [1, 4, 16, "inf"],
+        "server_environment_variables": {
+            "VLLM_RPC_TIMEOUT": 100000,
+	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
+	    "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
+	    "VLLM_CPU_KVCACHE_SPACE": 40
+        },
+        "server_parameters": {
+            "model": "meta-llama/Meta-Llama-3.1-8B-Instruct",
+            "pipeline_parallel_size": 6,
+	    "dtype": "bfloat16",
+	    "distributed_executor_backend": "mp",
+	    "block_size": 128,
+	    "trust_remote_code": "",
+	    "enable_chunked_prefill": "",
+            "disable_log_stats": "",
+            "disable_log_requests": "",
+	    "enforce_eager": "",
+            "load_format": "dummy"
+        },
+        "client_parameters": {
+            "model": "meta-llama/Meta-Llama-3.1-8B-Instruct",
+            "backend": "vllm",
+            "dataset_name": "random",
+	    "random-input-len": 1024,
+	    "random-output-len": 128,
+	    "ignore-eos": "",
+	    "max_concurrency": 100,
+            "num_prompts": 100
+        }
+    }
+]

.buildkite/nightly-benchmarks/tests/throughput-tests-cpu.json

@@ -0,0 +1,32 @@
+[
+    {
+        "test_name": "throughput_llama8B_tp1",
+        "environment_variables": {
+	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
+	    "VLLM_CPU_KVCACHE_SPACE": 40
+        },
+        "parameters": {
+            "model": "meta-llama/Meta-Llama-3.1-8B-Instruct",
+            "tensor_parallel_size": 1,
+            "load_format": "dummy",
+            "dataset": "./ShareGPT_V3_unfiltered_cleaned_split.json",
+            "num_prompts": 200,
+            "backend": "vllm"
+        }
+    },
+    {
+        "test_name": "throughput_llama8B_tp4",
+        "environment_variables": {
+	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
+	    "VLLM_CPU_KVCACHE_SPACE": 40
+        },
+        "parameters": {
+            "model": "meta-llama/Meta-Llama-3.1-8B-Instruct",
+            "tensor_parallel_size": 4,
+            "load_format": "dummy",
+            "dataset": "./ShareGPT_V3_unfiltered_cleaned_split.json",
+            "num_prompts": 200,
+            "backend": "vllm"
+        }
+    }
+]

.buildkite/release-pipeline.yaml

@@ -1,5 +1,6 @@
 steps:
   - label: "Build wheel - CUDA 12.8"
+    id: build-wheel-cuda-12-8
     agents:
       queue: cpu_queue_postmerge
     commands:
@@ -11,6 +12,7 @@ steps:
       DOCKER_BUILDKIT: "1"
 
   - label: "Build wheel - CUDA 12.6"
+    id: build-wheel-cuda-12-6
     agents:
       queue: cpu_queue_postmerge
     commands:
@@ -28,6 +30,7 @@ steps:
 
   - label: "Build wheel - CUDA 11.8"
     # depends_on: block-build-cu118-wheel
+    id: build-wheel-cuda-11-8
     agents:
       queue: cpu_queue_postmerge
     commands:
@@ -44,13 +47,26 @@ steps:
 
   - label: "Build release image"
     depends_on: block-release-image-build
+    id: build-release-image
     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=12.8.1 --tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT --target vllm-openai --progress plain -f docker/Dockerfile ."
+      - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=12.8.1 --build-arg INSTALL_KV_CONNECTORS=true --tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT --target vllm-openai --progress plain -f docker/Dockerfile ."
       - "docker push public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT"
 
+  - label: "Annotate release workflow"
+    depends_on:
+      - build-release-image
+      - build-wheel-cuda-12-8
+      - build-wheel-cuda-12-6
+      - build-wheel-cuda-11-8
+    id: annotate-release-workflow
+    agents:
+      queue: cpu_queue_postmerge
+    commands:
+      - "bash .buildkite/scripts/annotate-release.sh"
+
   - label: "Build and publish TPU release image"
     depends_on: ~
     if: build.env("NIGHTLY") == "1"
@@ -70,9 +86,10 @@ steps:
       DOCKER_BUILDKIT: "1"
 
   - input: "Provide Release version here"
+    id: input-release-version
     fields:
       - text: "What is the release version?"
-        key: "release-version"
+        key: release-version
 
   - block: "Build CPU release image"
     key: block-cpu-release-image-build
@@ -84,7 +101,8 @@ steps:
       queue: cpu_queue_postmerge
     commands:
       - "aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin public.ecr.aws/q9t5s3a7"
-      - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg GIT_REPO_CHECK=1 --tag public.ecr.aws/q9t5s3a7/vllm-cpu-release-repo:$(buildkite-agent meta-data get release-version) --tag public.ecr.aws/q9t5s3a7/vllm-cpu-release-repo:latest --progress plain --target vllm-openai -f docker/Dockerfile.cpu ."
+      - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg GIT_REPO_CHECK=1 --build-arg VLLM_CPU_AVX512BF16=true --build-arg VLLM_CPU_AVX512VNNI=true --tag public.ecr.aws/q9t5s3a7/vllm-cpu-release-repo:$(buildkite-agent meta-data get release-version) --tag public.ecr.aws/q9t5s3a7/vllm-cpu-release-repo:latest --progress plain --target vllm-openai -f docker/Dockerfile.cpu ."
+      - "docker push public.ecr.aws/q9t5s3a7/vllm-cpu-release-repo:latest"
       - "docker push public.ecr.aws/q9t5s3a7/vllm-cpu-release-repo:$(buildkite-agent meta-data get release-version)"
     env:
       DOCKER_BUILDKIT: "1"
@@ -100,6 +118,7 @@ steps:
     commands:
       - "aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin public.ecr.aws/q9t5s3a7"
       - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg GIT_REPO_CHECK=1 --tag public.ecr.aws/q9t5s3a7/vllm-neuron-release-repo:$(buildkite-agent meta-data get release-version) --tag public.ecr.aws/q9t5s3a7/vllm-neuron-release-repo:latest --progress plain -f docker/Dockerfile.neuron ."
+      - "docker push public.ecr.aws/q9t5s3a7/vllm-neuron-release-repo:latest"
       - "docker push public.ecr.aws/q9t5s3a7/vllm-neuron-release-repo:$(buildkite-agent meta-data get release-version)"
     env:
       DOCKER_BUILDKIT: "1"

.buildkite/scripts/annotate-release.sh

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

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

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

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

@@ -94,6 +94,10 @@ if [[ $commands == *"pytest -v -s compile/test_basic_correctness.py"* ]]; then
   commands=${commands//"pytest -v -s compile/test_basic_correctness.py"/"VLLM_USE_TRITON_FLASH_ATTN=0 pytest -v -s compile/test_basic_correctness.py"}
 fi
 
+if [[ $commands == *"pytest -v -s lora"* ]]; then
+  commands=${commands//"pytest -v -s lora"/"VLLM_ROCM_CUSTOM_PAGED_ATTN=0 pytest -v -s lora"}
+fi
+
 #ignore certain kernels tests
 if [[ $commands == *" kernels/core"* ]]; then
   commands="${commands} \
@@ -103,10 +107,8 @@ fi
 
 if [[ $commands == *" kernels/attention"* ]]; then
   commands="${commands} \
-  --ignore=kernels/attention/stest_attention_selector.py \
-  --ignore=kernels/attention/test_blocksparse_attention.py \
-  --ignore=kernels/attention/test_encoder_decoder_attn.py \
   --ignore=kernels/attention/test_attention_selector.py \
+  --ignore=kernels/attention/test_encoder_decoder_attn.py \
   --ignore=kernels/attention/test_flash_attn.py \
   --ignore=kernels/attention/test_flashinfer.py \
   --ignore=kernels/attention/test_prefix_prefill.py \

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

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

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

@@ -6,75 +6,83 @@ set -ex
 
 # allow to bind to different cores
 CORE_RANGE=${CORE_RANGE:-48-95}
+OMP_CORE_RANGE=${OMP_CORE_RANGE:-48-95}
 NUMA_NODE=${NUMA_NODE:-1}
 
+export CMAKE_BUILD_PARALLEL_LEVEL=32
+
 # Setup cleanup
 remove_docker_container() { 
     set -e; 
-    docker rm -f cpu-test-"$BUILDKITE_BUILD_NUMBER"-"$NUMA_NODE" cpu-test-"$BUILDKITE_BUILD_NUMBER"-avx2-"$NUMA_NODE" || true; 
-    docker image rm cpu-test-"$BUILDKITE_BUILD_NUMBER" cpu-test-"$BUILDKITE_BUILD_NUMBER"-avx2 || true; 
+    docker rm -f cpu-test-"$NUMA_NODE" cpu-test-"$NUMA_NODE"-avx2 || true; 
 }
 trap remove_docker_container EXIT
 remove_docker_container
 
 # Try building the docker image
-numactl -C "$CORE_RANGE" -N "$NUMA_NODE" docker build --tag cpu-test-"$BUILDKITE_BUILD_NUMBER" --target vllm-test -f docker/Dockerfile.cpu .
-numactl -C "$CORE_RANGE" -N "$NUMA_NODE" docker build --build-arg VLLM_CPU_DISABLE_AVX512="true" --tag cpu-test-"$BUILDKITE_BUILD_NUMBER"-avx2 --target vllm-test -f docker/Dockerfile.cpu .
+numactl -C "$CORE_RANGE" -N "$NUMA_NODE" docker build --tag cpu-test-"$NUMA_NODE" --target vllm-test -f docker/Dockerfile.cpu .
+numactl -C "$CORE_RANGE" -N "$NUMA_NODE" docker build --build-arg VLLM_CPU_DISABLE_AVX512="true" --tag cpu-test-"$NUMA_NODE"-avx2 --target vllm-test -f docker/Dockerfile.cpu .
 
 # Run the image, setting --shm-size=4g for tensor parallel.
-docker run -itd --entrypoint /bin/bash -v ~/.cache/huggingface:/root/.cache/huggingface --cpuset-cpus="$CORE_RANGE"  \
- --cpuset-mems="$NUMA_NODE" --privileged=true -e HF_TOKEN --env VLLM_CPU_KVCACHE_SPACE=4 --shm-size=4g --name cpu-test-"$BUILDKITE_BUILD_NUMBER"-"$NUMA_NODE" cpu-test-"$BUILDKITE_BUILD_NUMBER"
-docker run -itd --entrypoint /bin/bash -v ~/.cache/huggingface:/root/.cache/huggingface --cpuset-cpus="$CORE_RANGE" \
- --cpuset-mems="$NUMA_NODE" --privileged=true -e HF_TOKEN --env VLLM_CPU_KVCACHE_SPACE=4 --shm-size=4g --name cpu-test-"$BUILDKITE_BUILD_NUMBER"-avx2-"$NUMA_NODE" cpu-test-"$BUILDKITE_BUILD_NUMBER"-avx2
+docker run -itd --cpuset-cpus="$CORE_RANGE" --cpuset-mems="$NUMA_NODE" --entrypoint /bin/bash -v ~/.cache/huggingface:/root/.cache/huggingface --privileged=true -e HF_TOKEN --env VLLM_CPU_KVCACHE_SPACE=4 --env VLLM_CPU_CI_ENV=1 --shm-size=4g --name cpu-test-"$NUMA_NODE" cpu-test-"$NUMA_NODE"
+docker run -itd --cpuset-cpus="$CORE_RANGE" --cpuset-mems="$NUMA_NODE" --entrypoint /bin/bash -v ~/.cache/huggingface:/root/.cache/huggingface --privileged=true -e HF_TOKEN --env VLLM_CPU_KVCACHE_SPACE=4 --env VLLM_CPU_CI_ENV=1 --shm-size=4g --name cpu-test-"$NUMA_NODE"-avx2 cpu-test-"$NUMA_NODE"-avx2
 
 function cpu_tests() {
   set -e
   export NUMA_NODE=$2
-  export BUILDKITE_BUILD_NUMBER=$3
+
+  # list packages
+  docker exec cpu-test-"$NUMA_NODE"-avx2 bash -c "
+    set -e
+    pip list"
+
+  docker exec cpu-test-"$NUMA_NODE" bash -c "
+    set -e
+    pip list"
 
   # offline inference
-  docker exec cpu-test-"$BUILDKITE_BUILD_NUMBER"-avx2-"$NUMA_NODE" bash -c "
+  docker exec cpu-test-"$NUMA_NODE"-avx2 bash -c "
     set -e
     python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m"
 
   # Run basic model test
-  docker exec cpu-test-"$BUILDKITE_BUILD_NUMBER"-"$NUMA_NODE" bash -c "
+  docker exec cpu-test-"$NUMA_NODE" bash -c "
     set -e
-    pytest -v -s tests/kernels/test_cache.py -m cpu_model
-    pytest -v -s tests/kernels/test_mla_decode_cpu.py -m cpu_model
-    pytest -v -s tests/models/decoder_only/language -m cpu_model
-    pytest -v -s tests/models/embedding/language -m cpu_model
-    pytest -v -s tests/models/encoder_decoder/language -m cpu_model
-    pytest -v -s tests/models/decoder_only/audio_language -m cpu_model
-    pytest -v -s tests/models/decoder_only/vision_language -m cpu_model"
+    # Note: disable until supports V1
+    # pytest -v -s tests/kernels/attention/test_cache.py -m cpu_model
+    # pytest -v -s tests/kernels/attention/test_mla_decode_cpu.py -m cpu_model
+
+    # Note: disable Bart until supports V1
+    pytest -v -s tests/models/language/generation -m cpu_model \
+                --ignore=tests/models/language/generation/test_bart.py
+    VLLM_CPU_SGL_KERNEL=1 pytest -v -s tests/models/language/generation -m cpu_model \
+                --ignore=tests/models/language/generation/test_bart.py
+
+    pytest -v -s tests/models/language/pooling -m cpu_model
+    pytest -v -s tests/models/multimodal/generation \
+                --ignore=tests/models/multimodal/generation/test_mllama.py \
+                --ignore=tests/models/multimodal/generation/test_pixtral.py \
+                -m cpu_model"
 
   # Run compressed-tensor test
-  docker exec cpu-test-"$BUILDKITE_BUILD_NUMBER"-"$NUMA_NODE" bash -c "
+  docker exec cpu-test-"$NUMA_NODE" bash -c "
     set -e
     pytest -s -v \
-    tests/quantization/test_compressed_tensors.py::test_compressed_tensors_w8a8_static_setup \
-    tests/quantization/test_compressed_tensors.py::test_compressed_tensors_w8a8_dynamic_per_token"
+    tests/quantization/test_compressed_tensors.py::test_compressed_tensors_w8a8_logprobs[False-10-32-neuralmagic/Llama-3.2-1B-quantized.w8a8]" 
 
+  # Note: disable it until supports V1
   # Run AWQ test
-  docker exec cpu-test-"$BUILDKITE_BUILD_NUMBER"-"$NUMA_NODE" bash -c "
-    set -e
-    pytest -s -v \
-    tests/quantization/test_ipex_quant.py"
-
-  # Run chunked-prefill and prefix-cache test
-  docker exec cpu-test-"$BUILDKITE_BUILD_NUMBER"-"$NUMA_NODE" bash -c "
-    set -e
-    pytest -s -v -k cpu_model \
-    tests/basic_correctness/test_chunked_prefill.py"  
+  # docker exec cpu-test-"$NUMA_NODE" bash -c "
+  #   set -e
+  #   VLLM_USE_V1=0 pytest -s -v \
+  #   tests/quantization/test_ipex_quant.py"
 
   # online serving
-  docker exec cpu-test-"$BUILDKITE_BUILD_NUMBER"-"$NUMA_NODE" bash -c "
+  docker exec cpu-test-"$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 \
+    VLLM_CPU_CI_ENV=0 python3 benchmarks/benchmark_serving.py \
       --backend vllm \
       --dataset-name random \
       --model facebook/opt-125m \
@@ -83,7 +91,7 @@ function cpu_tests() {
       --tokenizer facebook/opt-125m"
 
   # Run multi-lora tests
-  docker exec cpu-test-"$BUILDKITE_BUILD_NUMBER"-"$NUMA_NODE" bash -c "
+  docker exec cpu-test-"$NUMA_NODE" bash -c "
     set -e
     pytest -s -v \
     tests/lora/test_qwen2vl.py"
@@ -91,4 +99,4 @@ function cpu_tests() {
 
 # 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 $BUILDKITE_BUILD_NUMBER"
+timeout 1.5h bash -c "cpu_tests $CORE_RANGE $NUMA_NODE"

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

@@ -2,10 +2,32 @@
 
 # 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
+set -exuo pipefail
 
 # Try building the docker image
-docker build -t hpu-test-env -f docker/Dockerfile.hpu .
+cat <<EOF | docker build -t hpu-plugin-v1-test-env -f - .
+FROM gaudi-base-image:latest
+
+COPY ./ /workspace/vllm
+
+WORKDIR /workspace/vllm
+
+ENV no_proxy=localhost,127.0.0.1
+ENV PT_HPU_ENABLE_LAZY_COLLECTIVES=true
+
+RUN VLLM_TARGET_DEVICE=empty pip install .
+RUN pip install git+https://github.com/vllm-project/vllm-gaudi.git
+
+# install development dependencies (for testing)
+RUN python3 -m pip install -e tests/vllm_test_utils
+
+WORKDIR /workspace/
+
+RUN git clone https://github.com/vllm-project/vllm-gaudi.git
+
+RUN ln -s /workspace/vllm/tests && ln -s /workspace/vllm/examples && ln -s /workspace/vllm/benchmarks
+
+EOF
 
 # Setup cleanup
 # certain versions of HPU software stack have a bug that can
@@ -14,13 +36,21 @@ docker build -t hpu-test-env -f docker/Dockerfile.hpu .
 # functions, while other platforms only need one remove_docker_container
 # function.
 EXITCODE=1
-remove_docker_containers() { docker rm -f hpu-test || true; docker rm -f hpu-test-tp2 || true; }
-remove_docker_containers_and_exit() { remove_docker_containers; exit $EXITCODE; }
-trap remove_docker_containers_and_exit EXIT
+remove_docker_containers() { docker rm -f hpu-plugin-v1-test || true; }
+trap 'remove_docker_containers; exit $EXITCODE;' EXIT
 remove_docker_containers
 
-# 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/generate.py --model facebook/opt-125m
-docker run --runtime=habana --name=hpu-test-tp2 --network=host -e HABANA_VISIBLE_DEVICES=all -e VLLM_SKIP_WARMUP=true --entrypoint="" hpu-test-env python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m --tensor-parallel-size 2
+echo "Running HPU plugin v1 test"
+docker run --rm --runtime=habana --name=hpu-plugin-v1-test --network=host \
+  -e HABANA_VISIBLE_DEVICES=all \
+  hpu-plugin-v1-test-env \
+  /bin/bash "/workspace/vllm-gaudi/tests/upstream_tests/ci_tests.sh"
 
 EXITCODE=$?
+if [ $EXITCODE -eq 0 ]; then
+  echo "Test with basic model passed"
+else
+  echo "Test with basic model FAILED with exit code: $EXITCODE" >&2
+fi
+
+# The trap will handle the container removal and final exit.

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

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

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

@@ -70,7 +70,7 @@ export VLLM_XLA_CACHE_PATH=
 echo "Using VLLM V1"
 
 echo "--- Hardware Information ---"
-tpu-info
+# tpu-info
 echo "--- Starting Tests ---"
 set +e
 overall_script_exit_code=0
@@ -150,11 +150,17 @@ run_and_track_test 9 "test_multimodal.py" \
 run_and_track_test 10 "test_pallas.py" \
     "python3 -m pytest -s -v /workspace/vllm/tests/v1/tpu/test_pallas.py"
 run_and_track_test 11 "test_struct_output_generate.py" \
-    "python3 -m pytest -s -v /workspace/vllm/tests/v1/entrypoints/llm/test_struct_output_generate.py"
+    "python3 -m pytest -s -v /workspace/vllm/tests/v1/entrypoints/llm/test_struct_output_generate.py -k \"not test_structured_output_with_reasoning_matrices\""
 run_and_track_test 12 "test_moe_pallas.py" \
     "python3 -m pytest -s -v /workspace/vllm/tests/tpu/test_moe_pallas.py"
 run_and_track_test 13 "test_lora.py" \
     "VLLM_XLA_CHECK_RECOMPILATION=0 python3 -m pytest -s -v /workspace/vllm/tests/tpu/lora/test_lora.py"
+run_and_track_test 14 "test_tpu_qkv_linear.py" \
+    "python3 -m pytest -s -v /workspace/vllm/tests/v1/tpu/test_tpu_qkv_linear.py"
+run_and_track_test 15 "test_spmd_model_weight_loading.py" \
+    "python3 -m pytest -s -v /workspace/vllm/tests/v1/tpu/test_spmd_model_weight_loading.py"
+run_and_track_test 16 "test_kv_cache_update_kernel.py" \
+    "python3 -m pytest -s -v /workspace/vllm/tests/v1/tpu/test_kv_cache_update_kernel.py"
 
 # After all tests have been attempted, exit with the overall status.
 if [ "$overall_script_exit_code" -ne 0 ]; then

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

@@ -11,8 +11,8 @@ container_name="xpu_${BUILDKITE_COMMIT}_$(tr -dc A-Za-z0-9 < /dev/urandom | head
 docker build -t ${image_name} -f docker/Dockerfile.xpu .
 
 # Setup cleanup
-remove_docker_container() { 
-  docker rm -f "${container_name}" || true; 
+remove_docker_container() {
+  docker rm -f "${container_name}" || true;
   docker image rm -f "${image_name}" || true;
   docker system prune -f || true;
 }
@@ -26,6 +26,9 @@ docker run \
     --name "${container_name}" \
     "${image_name}" \
     sh -c '
-    VLLM_USE_V1=0 python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m
-    VLLM_USE_V1=0 python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m -tp 2
+    VLLM_USE_V1=1 python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m --block-size 64 --enforce-eager
+    VLLM_USE_V1=1 python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m --block-size 64 --enforce-eager -tp 2 --distributed-executor-backend ray
+    VLLM_USE_V1=1 python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m --block-size 64 --enforce-eager -tp 2 --distributed-executor-backend mp
+    cd tests
+    pytest -v -s v1/core
 '

.buildkite/scripts/rerun-test.sh

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

.buildkite/scripts/tpu/cleanup_docker.sh

@@ -0,0 +1,24 @@
+#!/bin/bash
+
+set -euo pipefail
+
+docker_root=$(docker info -f '{{.DockerRootDir}}')
+if [ -z "$docker_root" ]; then
+  echo "Failed to determine Docker root directory."
+  exit 1
+fi
+echo "Docker root directory: $docker_root"
+# Check disk usage of the filesystem where Docker's root directory is located
+disk_usage=$(df "$docker_root" | tail -1 | awk '{print $5}' | sed 's/%//')
+# Define the threshold
+threshold=70
+if [ "$disk_usage" -gt "$threshold" ]; then
+  echo "Disk usage is above $threshold%. Cleaning up Docker images and volumes..."
+  # Remove dangling images (those that are not tagged and not used by any container)
+  docker image prune -f
+  # Remove unused volumes / force the system prune for old images as well.
+  docker volume prune -f && docker system prune --force --filter "until=72h" --all
+  echo "Docker images and volumes cleanup completed."
+else
+  echo "Disk usage is below $threshold%. No cleanup needed."
+fi

.buildkite/scripts/tpu/config_v6e_1.env

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

.buildkite/scripts/tpu/docker_run_bm.sh

@@ -0,0 +1,92 @@
+#!/bin/bash
+
+if [ ! -f "$1" ]; then
+  echo "Error: The env file '$1' does not exist."
+  exit 1  # Exit the script with a non-zero status to indicate an error
+fi
+
+ENV_FILE=$1
+
+# For testing on local vm, use `set -a` to export all variables
+source /etc/environment
+source $ENV_FILE
+
+remove_docker_container() { 
+    docker rm -f tpu-test || true; 
+    docker rm -f vllm-tpu || true;
+    docker rm -f $CONTAINER_NAME || true;
+}
+
+trap remove_docker_container EXIT
+
+# Remove the container that might not be cleaned up in the previous run.
+remove_docker_container
+
+LOG_ROOT=$(mktemp -d)
+# If mktemp fails, set -e will cause the script to exit.
+echo "Results will be stored in: $LOG_ROOT"
+
+if [ -z "$HF_TOKEN" ]; then
+  echo "Error: HF_TOKEN is not set or is empty."  
+  exit 1
+fi
+
+# Make sure mounted disk or dir exists
+if [ ! -d "$DOWNLOAD_DIR" ]; then
+    echo "Error: Folder $DOWNLOAD_DIR does not exist. This is useually a mounted drive. If no mounted drive, just create a folder."
+    exit 1
+fi
+
+echo "Run model $MODEL"
+echo
+
+echo "starting docker...$CONTAINER_NAME"
+echo    
+docker run \
+ -v $DOWNLOAD_DIR:$DOWNLOAD_DIR \
+ --env-file $ENV_FILE \
+ -e HF_TOKEN="$HF_TOKEN" \
+ -e TARGET_COMMIT=$BUILDKITE_COMMIT \
+ -e MODEL=$MODEL \
+ -e WORKSPACE=/workspace \
+ --name $CONTAINER_NAME \
+ -d \
+ --privileged \
+ --network host \
+ -v /dev/shm:/dev/shm \
+ vllm/vllm-tpu-bm tail -f /dev/null
+
+echo "run script..."
+echo
+docker exec "$CONTAINER_NAME" /bin/bash -c ".buildkite/scripts/tpu/run_bm.sh"
+
+echo "copy result back..."
+VLLM_LOG="$LOG_ROOT/$TEST_NAME"_vllm_log.txt
+BM_LOG="$LOG_ROOT/$TEST_NAME"_bm_log.txt
+docker cp "$CONTAINER_NAME:/workspace/vllm_log.txt" "$VLLM_LOG" 
+docker cp "$CONTAINER_NAME:/workspace/bm_log.txt" "$BM_LOG"
+
+throughput=$(grep "Request throughput (req/s):" "$BM_LOG" | sed 's/[^0-9.]//g')
+echo "throughput for $TEST_NAME at $BUILDKITE_COMMIT: $throughput"
+
+if [ "$BUILDKITE" = "true" ]; then
+  echo "Running inside Buildkite"
+  buildkite-agent artifact upload "$VLLM_LOG" 
+  buildkite-agent artifact upload "$BM_LOG"
+else
+  echo "Not running inside Buildkite"
+fi
+
+#
+# compare the throughput with EXPECTED_THROUGHPUT 
+# and assert meeting the expectation
+# 
+if [[ -z "$throughput" || ! "$throughput" =~ ^[0-9]+([.][0-9]+)?$ ]]; then
+  echo "Failed to get the throughput"
+  exit 1
+fi
+
+if (( $(echo "$throughput < $EXPECTED_THROUGHPUT" | bc -l) )); then
+  echo "Error: throughput($throughput) is less than expected($EXPECTED_THROUGHPUT)"
+  exit 1
+fi

.buildkite/scripts/tpu/quantized_v6e_1.env

@@ -0,0 +1,14 @@
+# Environment config
+TEST_NAME=llama8bw8a8
+CONTAINER_NAME=vllm-tpu
+
+# vllm config
+MODEL=RedHatAI/Meta-Llama-3.1-8B-Instruct-quantized.w8a8
+MAX_NUM_SEQS=128
+MAX_NUM_BATCHED_TOKENS=1024
+TENSOR_PARALLEL_SIZE=1
+MAX_MODEL_LEN=2048
+DOWNLOAD_DIR=/mnt/disks/persist
+EXPECTED_THROUGHPUT=10.0
+INPUT_LEN=1800
+OUTPUT_LEN=128

.buildkite/scripts/tpu/run_bm.sh

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

.buildkite/test-pipeline.yaml

@@ -41,6 +41,16 @@ steps:
   # TODO: add `--strict` once warnings in docstrings are fixed
   - mkdocs build
 
+- label: Pytorch Nightly Dependency Override Check # 2min
+  # if this test fails, it means the nightly torch version is not compatible with some
+  # of the dependencies. Please check the error message and add the package to whitelist
+  # in /vllm/tools/generate_nightly_torch_test.py
+  soft_fail: true
+  source_file_dependencies:
+  - requirements/nightly_torch_test.txt
+  commands:
+  - bash standalone_tests/pytorch_nightly_dependency.sh
+
 - label: Async Engine, Inputs, Utils, Worker Test # 24min
   mirror_hardwares: [amdexperimental]
   source_file_dependencies:
@@ -89,7 +99,7 @@ steps:
   - VLLM_TEST_ENABLE_ARTIFICIAL_PREEMPT=1 pytest -v -s basic_correctness/test_preemption.py
 
 - label: Chunked Prefill Test
-  mirror_hardwares: [amdexperimental]
+  mirror_hardwares: [amdexperimental, amdproduction]
   source_file_dependencies:
   - vllm/
   - tests/basic_correctness/test_chunked_prefill
@@ -107,7 +117,7 @@ steps:
   commands:
   - pytest -v -s core
 
-- label: Entrypoints Test # 40min
+- label: Entrypoints Test (LLM) # 40min
   mirror_hardwares: [amdexperimental]
   working_dir: "/vllm-workspace/tests"
   fast_check: true
@@ -115,8 +125,6 @@ steps:
   source_file_dependencies:
   - vllm/
   - tests/entrypoints/llm
-  - tests/entrypoints/openai
-  - tests/entrypoints/test_chat_utils
   - tests/entrypoints/offline_mode
   commands:
   - export VLLM_WORKER_MULTIPROC_METHOD=spawn
@@ -125,9 +133,21 @@ steps:
   - 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
   - VLLM_USE_V1=0 pytest -v -s entrypoints/llm/test_guided_generate.py # it needs a clean process
+  - VLLM_USE_V1=0 pytest -v -s entrypoints/offline_mode # Needs to avoid interference with other tests
+
+- label: Entrypoints Test (API Server) # 40min
+  mirror_hardwares: [amdexperimental]
+  working_dir: "/vllm-workspace/tests"
+  fast_check: true
+  torch_nightly: true
+  source_file_dependencies:
+  - vllm/
+  - tests/entrypoints/openai
+  - tests/entrypoints/test_chat_utils
+  commands:
+  - export VLLM_WORKER_MULTIPROC_METHOD=spawn
   - pytest -v -s entrypoints/openai --ignore=entrypoints/openai/test_chat_with_tool_reasoning.py --ignore=entrypoints/openai/test_oot_registration.py --ignore=entrypoints/openai/test_tensorizer_entrypoint.py --ignore=entrypoints/openai/correctness/
   - pytest -v -s entrypoints/test_chat_utils.py
-  - VLLM_USE_V1=0 pytest -v -s entrypoints/offline_mode # Needs to avoid interference with other tests
 
 - label: Distributed Tests (4 GPUs) # 10min
   mirror_hardwares: [amdexperimental]
@@ -139,12 +159,13 @@ steps:
   - tests/distributed/test_utils
   - tests/distributed/test_pynccl
   - tests/distributed/test_events
-  - tests/spec_decode/e2e/test_integration_dist_tp4
   - tests/compile/test_basic_correctness
   - examples/offline_inference/rlhf.py
   - examples/offline_inference/rlhf_colocate.py
   - tests/examples/offline_inference/data_parallel.py
   - tests/v1/test_async_llm_dp.py
+  - tests/v1/test_external_lb_dp.py
+  - tests/v1/engine/test_engine_core_client.py
   commands:
   # test with tp=2 and external_dp=2
   - VLLM_USE_V1=0 torchrun --nproc-per-node=4 distributed/test_torchrun_example.py
@@ -152,13 +173,14 @@ steps:
   # test with tp=2 and pp=2
   - PP_SIZE=2 torchrun --nproc-per-node=4 distributed/test_torchrun_example.py
   # test with internal dp
-  - python3 ../examples/offline_inference/data_parallel.py
+  - python3 ../examples/offline_inference/data_parallel.py --enforce-eager
   - TP_SIZE=2 DP_SIZE=2 pytest -v -s v1/test_async_llm_dp.py
+  - TP_SIZE=2 DP_SIZE=2 pytest -v -s v1/test_external_lb_dp.py
+  - pytest -v -s v1/engine/test_engine_core_client.py::test_kv_cache_events_dp
   - pytest -v -s distributed/test_utils.py
   - pytest -v -s compile/test_basic_correctness.py
   - pytest -v -s distributed/test_pynccl.py
   - pytest -v -s distributed/test_events.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
   - pushd ../examples/offline_inference
@@ -166,6 +188,23 @@ steps:
   - VLLM_ALLOW_INSECURE_SERIALIZATION=1 RAY_DEDUP_LOGS=0 python3 rlhf_colocate.py
   - popd
 
+- label: EPLB Algorithm Test
+  working_dir: "/vllm-workspace/tests"
+  source_file_dependencies:
+  - vllm/distributed/eplb
+  - tests/distributed/test_eplb_algo.py
+  commands:
+  - pytest -v -s distributed/test_eplb_algo.py
+
+- label: EPLB Execution Test # 5min
+  working_dir: "/vllm-workspace/tests"
+  num_gpus: 4
+  source_file_dependencies:
+  - vllm/distributed/eplb
+  - tests/distributed/test_eplb_execute.py
+  commands:
+  - pytest -v -s distributed/test_eplb_execute.py
+
 - label: Metrics, Tracing Test # 10min
   mirror_hardwares: [amdexperimental, amdproduction]
   num_gpus: 2
@@ -175,13 +214,18 @@ steps:
   - tests/tracing
   commands:
   - pytest -v -s metrics
+  - "pip install \
+      'opentelemetry-sdk>=1.26.0' \
+      'opentelemetry-api>=1.26.0' \
+      'opentelemetry-exporter-otlp>=1.26.0' \
+      'opentelemetry-semantic-conventions-ai>=0.4.1'"
   - pytest -v -s tracing
 
 ##### fast check tests  #####
 #####  1 GPU test  #####
 
 - label: Regression Test # 5min
-  mirror_hardwares: [amdexperimental, amdproduction]
+  mirror_hardwares: [amdexperimental]
   source_file_dependencies:
   - vllm/
   - tests/test_regression
@@ -191,7 +235,7 @@ steps:
   working_dir: "/vllm-workspace/tests" # optional
 
 - label: Engine Test # 10min
-  mirror_hardwares: [amdexperimental, amdproduction]
+  mirror_hardwares: [amdexperimental]
   source_file_dependencies:
   - vllm/
   - tests/engine
@@ -220,6 +264,7 @@ steps:
     - pytest -v -s v1/structured_output
     - pytest -v -s v1/spec_decode
     - pytest -v -s v1/kv_connector/unit
+    - pytest -v -s v1/metrics
     - pytest -v -s v1/test_serial_utils.py
     - pytest -v -s v1/test_utils.py
     - pytest -v -s v1/test_oracle.py
@@ -246,7 +291,7 @@ steps:
     - python3 offline_inference/llm_engine_example.py
     - python3 offline_inference/audio_language.py --seed 0
     - python3 offline_inference/vision_language.py --seed 0
-    - python3 offline_inference/vision_language_embedding.py --seed 0
+    - python3 offline_inference/vision_language_pooling.py --seed 0
     - python3 offline_inference/vision_language_multi_image.py --seed 0
     - VLLM_USE_V1=0 python3 others/tensorize_vllm_model.py --model facebook/opt-125m serialize --serialized-directory /tmp/ --suffix v1 && python3 others/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
@@ -264,6 +309,15 @@ steps:
   commands:
     - pytest -v -s prefix_caching
 
+
+- label: Platform Tests (CUDA)
+  mirror_hardwares: [amdexperimental]
+  source_file_dependencies:
+  - vllm/
+  - tests/cuda
+  commands:
+    - pytest -v -s cuda/test_cuda_context.py
+
 - label: Samplers Test # 36min
   mirror_hardwares: [amdexperimental]
   source_file_dependencies:
@@ -275,19 +329,8 @@ steps:
     - pytest -v -s samplers
     - VLLM_USE_FLASHINFER_SAMPLER=1 pytest -v -s samplers
 
-- label: Speculative decoding tests # 40min
-  mirror_hardwares: [amdexperimental]
-  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 --ignore=spec_decode/e2e/test_mtp_correctness.py
-    - pytest -v -s spec_decode/e2e/test_eagle_correctness.py
-
 - label: LoRA Test %N # 15min each
-  mirror_hardwares: [amdexperimental]
+  mirror_hardwares: [amdexperimental, amdproduction]
   source_file_dependencies:
   - vllm/lora
   - tests/lora
@@ -295,7 +338,7 @@ steps:
   parallelism: 4
 
 - label: PyTorch Compilation Unit Tests
-  mirror_hardwares: [amdexperimental, amdproduction]
+  mirror_hardwares: [amdexperimental]
   torch_nightly: true
   source_file_dependencies:
     - vllm/
@@ -303,6 +346,7 @@ steps:
   commands:
     - pytest -v -s compile/test_pass_manager.py
     - pytest -v -s compile/test_fusion.py
+    - pytest -v -s compile/test_fusion_attn.py
     - pytest -v -s compile/test_silu_mul_quant_fusion.py
     - pytest -v -s compile/test_sequence_parallelism.py
     - pytest -v -s compile/test_async_tp.py
@@ -318,6 +362,7 @@ steps:
   # these tests need to be separated, cannot combine
   - pytest -v -s compile/piecewise/test_simple.py
   - pytest -v -s compile/piecewise/test_toy_llama.py
+  - pytest -v -s compile/piecewise/test_full_cudagraph.py
 
 - label: PyTorch Fullgraph Test # 18min
   mirror_hardwares: [amdexperimental, amdproduction]
@@ -375,7 +420,7 @@ steps:
     - pytest -v -s kernels/mamba
 
 - label: Tensorizer Test # 11min
-  mirror_hardwares: [amdexperimental, amdproduction]
+  mirror_hardwares: [amdexperimental]
   soft_fail: true
   source_file_dependencies:
   - vllm/model_executor/model_loader
@@ -421,6 +466,9 @@ steps:
   - vllm/model_executor/layers/quantization
   - tests/quantization
   commands:
+  # temporary install here since we need nightly, will move to requirements/test.in
+  # after torchao 0.12 release
+  - pip install --pre torchao --index-url https://download.pytorch.org/whl/nightly/cu126
   - VLLM_TEST_FORCE_LOAD_FORMAT=auto pytest -v -s quantization
 
 - label: LM Eval Small Models # 53min
@@ -464,7 +512,7 @@ steps:
 #####  models test  #####
 
 - label: Basic Models Test # 24min
-  mirror_hardwares: [amdexperimental, amdproduction]
+  mirror_hardwares: [amdexperimental]
   torch_nightly: true
   source_file_dependencies:
   - vllm/
@@ -488,6 +536,17 @@ steps:
     - pip freeze | grep -E 'torch'
     - pytest -v -s models/language -m core_model
 
+- label: Language Models Test (Hybrid) # 35 min
+  mirror_hardwares: [amdexperimental]
+  torch_nightly: true
+  source_file_dependencies:
+  - vllm/
+  - tests/models/language/generation
+  commands:
+    # Install causal-conv1d for plamo2 models here, as it is not compatible with pip-compile.
+    - pip install 'git+https://github.com/Dao-AILab/causal-conv1d@v1.5.0.post8'
+    - pytest -v -s models/language/generation -m hybrid_model
+
 - label: Language Models Test (Extended Generation) # 1hr20min
   mirror_hardwares: [amdexperimental]
   optional: true
@@ -497,7 +556,7 @@ steps:
   commands:
     # Install causal-conv1d for plamo2 models here, as it is not compatible with pip-compile.
     - pip install 'git+https://github.com/Dao-AILab/causal-conv1d@v1.5.0.post8'
-    - pytest -v -s models/language/generation -m 'not core_model'
+    - pytest -v -s models/language/generation -m '(not core_model) and (not hybrid_model)'
 
 - label: Language Models Test (Extended Pooling)  # 36min
   mirror_hardwares: [amdexperimental]
@@ -542,7 +601,7 @@ steps:
     - pytest -v -s models/multimodal/generation/test_common.py -m 'split(group=0) and not core_model'
 
 - label: Multi-Modal Models Test (Extended) 3
-  mirror_hardwares: [amdexperimental, amdproduction]
+  mirror_hardwares: [amdexperimental]
   optional: true
   source_file_dependencies:
   - vllm/
@@ -569,6 +628,18 @@ steps:
     # e.g. pytest -v -s models/encoder_decoder/vision_language/test_mllama.py
     # *To avoid merge conflicts, remember to REMOVE (not just comment out) them before merging the PR*
 
+- label: Transformers Nightly Models Test
+  working_dir: "/vllm-workspace/"
+  optional: true
+  commands:
+    - pip install --upgrade git+https://github.com/huggingface/transformers
+    - pytest -v -s tests/models/test_initialization.py
+    - pytest -v -s tests/models/multimodal/processing/
+    - pytest -v -s tests/models/multimodal/test_mapping.py
+    - python3 examples/offline_inference/basic/chat.py
+    - python3 examples/offline_inference/audio_language.py --model-type whisper
+    - python3 examples/offline_inference/vision_language.py --model-type qwen2_5_vl
+
 #####  1 GPU test  #####
 #####  multi gpus test  #####
 
@@ -594,13 +665,18 @@ steps:
   - vllm/executor/
   - vllm/model_executor/models/
   - tests/distributed/
+  - tests/examples/offline_inference/data_parallel.py
   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 'Same node test passed'
+    - NUM_NODES=2 torchrun --nnodes 2 --nproc-per-node=2 --rdzv_backend=c10d --rdzv_endpoint=192.168.10.10 distributed/test_node_count.py | grep 'Node count test passed'
+    - python3 ../examples/offline_inference/data_parallel.py --dp-size=2 --tp-size=1 --node-size=2 --node-rank=0 --master-addr=192.168.10.10 --master-port=12345 --enforce-eager --trust-remote-code
     - 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 'Same node test passed'
+    - NUM_NODES=2 torchrun --nnodes 2 --nproc-per-node=2 --rdzv_backend=c10d --rdzv_endpoint=192.168.10.10 distributed/test_node_count.py | grep 'Node count test passed'
+    - python3 ../examples/offline_inference/data_parallel.py --dp-size=2 --tp-size=1 --node-size=2 --node-rank=1 --master-addr=192.168.10.10 --master-port=12345 --enforce-eager --trust-remote-code
 
 - label: Distributed Tests (2 GPUs) # 40min
   mirror_hardwares: [amdexperimental]
@@ -618,10 +694,12 @@ steps:
   - vllm/worker/model_runner.py
   - entrypoints/llm/test_collective_rpc.py
   - tests/v1/test_async_llm_dp.py
+  - tests/v1/test_external_lb_dp.py
   - tests/v1/entrypoints/openai/test_multi_api_servers.py
   - vllm/v1/engine/
   commands:
   - TP_SIZE=1 DP_SIZE=2 pytest -v -s v1/test_async_llm_dp.py
+  - TP_SIZE=1 DP_SIZE=2 pytest -v -s v1/test_external_lb_dp.py
   - DP_SIZE=2 pytest -v -s v1/entrypoints/openai/test_multi_api_servers.py
   - pytest -v -s entrypoints/llm/test_collective_rpc.py
   - pytest -v -s ./compile/test_basic_correctness.py
@@ -636,7 +714,6 @@ steps:
   - pytest -v -s distributed/test_sequence_parallel.py
   # this test fails consistently.
   # TODO: investigate and fix
-  # - pytest -v -s spec_decode/e2e/test_integration_dist_tp2.py
   - VLLM_USE_V1=0 CUDA_VISIBLE_DEVICES=0,1 pytest -v -s test_sharded_state_loader.py
   - VLLM_USE_V1=0 CUDA_VISIBLE_DEVICES=0,1 pytest -v -s kv_transfer/test_disagg.py
   - CUDA_VISIBLE_DEVICES=0,1 pytest -v -s v1/shutdown
@@ -663,7 +740,7 @@ steps:
   - pytest -v -s plugins/lora_resolvers # unit tests for in-tree lora resolver plugins
 
 - label: Multi-step Tests (4 GPUs) # 36min
-  mirror_hardwares: [amdexperimental]
+  mirror_hardwares: [amdexperimental, amdproduction]
   working_dir: "/vllm-workspace/tests"
   num_gpus: 4
   source_file_dependencies:
@@ -724,7 +801,7 @@ steps:
     - bash weight_loading/run_model_weight_loading_test.sh -c weight_loading/models.txt
 
 - label: Weight Loading Multiple GPU Test - Large Models # optional
-  mirror_hardwares: [amdexperimental] 
+  mirror_hardwares: [amdexperimental]
   working_dir: "/vllm-workspace/tests"
   num_gpus: 2
   gpu: a100

.gemini/config.yaml

@@ -0,0 +1,6 @@
+# https://developers.google.com/gemini-code-assist/docs/customize-gemini-behavior-github
+have_fun: false  # Just review the code
+code_review:
+  comment_severity_threshold: HIGH  # Reduce quantity of comments
+  pull_request_opened:
+    summary: false  # Don't summarize the PR in a separate comment

.github/CODEOWNERS

@@ -10,15 +10,22 @@
 /vllm/worker/worker.py @zhuohan123 @youkaichao @alexm-redhat @comaniac @njhill
 /vllm/model_executor/layers/sampler.py @zhuohan123 @youkaichao @alexm-redhat @comaniac @njhill
 /vllm/model_executor/layers/quantization @mgoin @robertgshaw2-redhat @tlrmchlsmth
-/vllm/model_executor/guided_decoding @mgoin @russellb
+/vllm/model_executor/guided_decoding @mgoin @russellb @aarnphm
 /vllm/multimodal @DarkLight1337 @ywang96
 /vllm/vllm_flash_attn @LucasWilkinson
 /vllm/lora @jeejeelee
-CMakeLists.txt @tlrmchlsmth
+/vllm/reasoning @aarnphm
+/vllm/entrypoints @aarnphm
+/vllm/compilation @zou3519 @youkaichao @ProExpertProg
+CMakeLists.txt @tlrmchlsmth @LucasWilkinson
+
+# Any change to the VllmConfig changes can have a large user-facing impact,
+# so spam a lot of people
+/vllm/config.py @simon-mo @WoosukKwon @youkaichao @robertgshaw2-redhat @mgoin @tlrmchlsmth @houseroad @hmellor
 
 # vLLM V1
 /vllm/v1 @WoosukKwon @robertgshaw2-redhat @njhill @ywang96 @comaniac @alexm-redhat
-/vllm/v1/structured_output @mgoin @russellb
+/vllm/v1/structured_output @mgoin @russellb @aarnphm
 
 # Test ownership
 /.buildkite/lm-eval-harness @mgoin @simon-mo
@@ -27,8 +34,8 @@ CMakeLists.txt @tlrmchlsmth
 /tests/distributed/test_multi_node_assignment.py @youkaichao
 /tests/distributed/test_pipeline_parallel.py @youkaichao
 /tests/distributed/test_same_node.py @youkaichao
-/tests/entrypoints @DarkLight1337 @robertgshaw2-redhat @simon-mo
-/tests/entrypoints/llm/test_guided_generate.py @mgoin @russellb
+/tests/entrypoints @DarkLight1337 @robertgshaw2-redhat @simon-mo @aarnphm
+/tests/entrypoints/llm/test_guided_generate.py @mgoin @russellb @aarnphm
 /tests/kernels @tlrmchlsmth @WoosukKwon
 /tests/model_executor/test_guided_processors.py @mgoin @russellb
 /tests/models @DarkLight1337 @ywang96
@@ -36,13 +43,12 @@ CMakeLists.txt @tlrmchlsmth
 /tests/multimodal @DarkLight1337 @ywang96
 /tests/prefix_caching @comaniac @KuntaiDu
 /tests/quantization @mgoin @robertgshaw2-redhat
-/tests/spec_decode @njhill @LiuXiaoxuanPKU
 /tests/test_inputs.py @DarkLight1337 @ywang96
-/tests/v1/entrypoints/llm/test_struct_output_generate.py @mgoin @russellb
-/tests/v1/structured_output @mgoin @russellb
+/tests/v1/entrypoints/llm/test_struct_output_generate.py @mgoin @russellb @aarnphm
+/tests/v1/structured_output @mgoin @russellb @aarnphm
 /tests/weight_loading @mgoin @youkaichao
 /tests/lora @jeejeelee
 
 # Docs
 /docs @hmellor
-mkdocs.yaml @hmellor
+mkdocs.yaml @hmellor

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

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

.github/PULL_REQUEST_TEMPLATE.md

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

.github/mergify.yml

@@ -27,6 +27,22 @@ pull_request_rules:
       add:
         - ci/build
 
+- name: label-deepseek
+  description: Automatically apply deepseek label
+  conditions:
+    - or:
+      - files~=^examples/.*deepseek.*\.py
+      - files~=^tests/.*deepseek.*\.py
+      - files~=^vllm/entrypoints/openai/tool_parsers/.*deepseek.*\.py
+      - files~=^vllm/model_executor/models/.*deepseek.*\.py
+      - files~=^vllm/reasoning/.*deepseek.*\.py
+      - files~=^vllm/transformers_utils/.*deepseek.*\.py
+      - title~=(?i)DeepSeek
+  actions:
+    label:
+      add:
+        - deepseek
+
 - name: label-frontend
   description: Automatically apply frontend label
   conditions:
@@ -36,6 +52,21 @@ pull_request_rules:
       add:
         - frontend
 
+- name: label-llama
+  description: Automatically apply llama label
+  conditions:
+    - or:
+      - files~=^examples/.*llama.*\.py
+      - files~=^tests/.*llama.*\.py
+      - files~=^vllm/entrypoints/openai/tool_parsers/llama.*\.py
+      - files~=^vllm/model_executor/models/.*llama.*\.py
+      - files~=^vllm/transformers_utils/configs/.*llama.*\.py
+      - title~=(?i)llama
+  actions:
+    label:
+      add:
+        - llama
+
 - name: label-multi-modality
   description: Automatically apply multi-modality label
   conditions:
@@ -43,14 +74,70 @@ pull_request_rules:
       - files~=^vllm/multimodal/
       - files~=^tests/multimodal/
       - files~=^tests/models/multimodal/
-      - files~=^tests/models/*/audio_language/
-      - files~=^tests/models/*/vision_language/
       - files=tests/models/test_vision.py
   actions:
     label:
       add:
         - multi-modality
 
+- name: label-new-model
+  description: Automatically apply new-model label
+  conditions:
+    - and:
+      - files~=^vllm/model_executor/models/
+      - files=vllm/model_executor/models/registry.py
+  actions:
+    label:
+      add:
+        - new-model
+
+- name: label-performance
+  description: Automatically apply performance label
+  conditions:
+    - or:
+      - files~=^benchmarks/
+      - files~=^vllm/benchmarks/
+      - files~=^tests/benchmarks/
+      - files~=^\.buildkite/nightly-benchmarks/
+  actions:
+    label:
+      add:
+        - performance
+
+- name: label-qwen
+  description: Automatically apply qwen label
+  conditions:
+    - or:
+      - files~=^examples/.*qwen.*\.py
+      - files~=^tests/.*qwen.*\.py
+      - files~=^vllm/model_executor/models/.*qwen.*\.py
+      - files~=^vllm/reasoning/.*qwen.*\.py
+      - title~=(?i)Qwen
+  actions:
+    label:
+      add:
+        - qwen
+
+- name: label-rocm
+  description: Automatically apply rocm label
+  conditions:
+    - or:
+      - files~=^csrc/rocm/
+      - files~=^docker/Dockerfile.rocm
+      - files~=^requirements/rocm.*\.txt
+      - files~=^vllm/attention/backends/rocm.*\.py
+      - files~=^vllm/attention/ops/rocm.*\.py
+      - files~=^vllm/model_executor/layers/fused_moe/rocm.*\.py
+      - files~=^vllm/v1/attention/backends/mla/rocm.*\.py
+      - files~=^tests/kernels/.*_rocm.*\.py
+      - files=vllm/platforms/rocm.py
+      - title~=(?i)AMD
+      - title~=(?i)ROCm
+  actions:
+    label:
+      add:
+        - rocm
+
 - name: label-structured-output
   description: Automatically apply structured-output label
   conditions:
@@ -77,9 +164,12 @@ pull_request_rules:
   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/
+      - files~=^vllm/v1/spec_decode/
+      - files~=^tests/v1/spec_decode/
+      - files~=^examples/.*(spec_decode|mlpspeculator|eagle|speculation).*\.py
+      - files~=^vllm/model_executor/models/.*eagle.*\.py
+      - files=vllm/model_executor/models/mlp_speculator.py
+      - files~=^vllm/transformers_utils/configs/(eagle|medusa|mlp_speculator)\.py
   actions:
     label:
       add:

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

@@ -68,7 +68,7 @@ jobs:
           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=$AWS_SECRET_ACCESS_KEY --set resources.requests.cpu=1 --set resources.requests.memory=4Gi --set resources.limits.cpu=2 --set resources.limits.memory=5Gi --set image.env[0].name=VLLM_CPU_KVCACHE_SPACE --set image.env[1].name=VLLM_LOGGING_LEVEL --set-string image.env[0].value="1" --set-string image.env[1].value="DEBUG" --set-string extraInit.s3modelpath="opt-125m/" --set-string 'resources.limits.nvidia\.com/gpu=0' --set-string 'resources.requests.nvidia\.com/gpu=0' --set-string image.repository="vllm-cpu-env"
+          helm install --wait --wait-for-jobs --timeout 5m0s --debug --create-namespace --namespace=ns-vllm test-vllm examples/online_serving/chart-helm -f examples/online_serving/chart-helm/values.yaml --set secrets.s3endpoint=http://minio:9000 --set secrets.s3bucketname=testbucket --set secrets.s3accesskeyid=$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 image.env[2].name=VLLM_CPU_CI_ENV --set-string image.env[0].value="1" --set-string image.env[1].value="DEBUG" --set-string image.env[2].value="1" --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: |

.gitignore

@@ -146,6 +146,7 @@ venv.bak/
 
 # mkdocs documentation
 /site
+docs/argparse
 docs/examples
 
 # mypy
@@ -200,5 +201,5 @@ benchmarks/**/*.json
 actionlint
 shellcheck*/
 
-# Ingore moe/marlin_moe gen code
+# Ignore moe/marlin_moe gen code
 csrc/moe/marlin_moe_wna16/kernel_*

.pre-commit-config.yaml

@@ -11,6 +11,8 @@ repos:
   hooks:
   - id: yapf
     args: [--in-place, --verbose]
+    # Keep the same list from yapfignore here to avoid yapf failing without any inputs
+    exclude: '(.buildkite|benchmarks|build|examples)/.*'
 - repo: https://github.com/astral-sh/ruff-pre-commit
   rev: v0.11.7
   hooks:
@@ -18,12 +20,10 @@ repos:
     args: [--output-format, github, --fix]
   - id: ruff-format
     files: ^(.buildkite|benchmarks|examples)/.*
-- repo: https://github.com/codespell-project/codespell
-  rev: v2.4.1
+- repo: https://github.com/crate-ci/typos
+  rev: v1.34.0
   hooks:
-  - id: codespell
-    additional_dependencies: ['tomli']
-    args: ['--toml', 'pyproject.toml']
+  - id: typos
 - repo: https://github.com/PyCQA/isort
   rev: 6.0.1
   hooks:
@@ -53,6 +53,11 @@ repos:
       files: ^requirements/test\.(in|txt)$
 - repo: local
   hooks:
+  - id: format-torch-nightly-test
+    name: reformat nightly_torch_test.txt to be in sync with test.in
+    language: python
+    entry: python tools/generate_nightly_torch_test.py
+    files: ^requirements/test\.(in|txt)$
   - id: mypy-local
     name: Run mypy for local Python installation
     entry: tools/mypy.sh 0 "local"
@@ -115,6 +120,11 @@ repos:
     entry: python tools/check_spdx_header.py
     language: python
     types: [python]
+  - id: check-root-lazy-imports
+    name: Check root lazy imports
+    entry: python tools/check_init_lazy_imports.py
+    language: python
+    types: [python]
   - id: check-filenames
     name: Check for spaces in all filenames
     entry: bash
@@ -143,10 +153,24 @@ repos:
     types: [python]
     pass_filenames: false
     additional_dependencies: [regex]
+  - id: check-pickle-imports
+    name: Prevent new pickle/cloudpickle imports
+    entry: python tools/check_pickle_imports.py
+    language: python
+    types: [python]
+    pass_filenames: false
+    additional_dependencies: [pathspec, regex]
+  - id: validate-config
+    name: Validate configuration has default values and that each field has a docstring
+    entry: python tools/validate_config.py
+    language: python
+    types: [python]
+    pass_filenames: true
+    files: vllm/config.py|tests/test_config.py|vllm/entrypoints/openai/cli_args.py
   # Keep `suggestion` last
   - id: suggestion
     name: Suggestion
-    entry: bash -c 'echo "To bypass pre-commit hooks, add --no-verify to git commit."'
+    entry: bash -c 'echo "To bypass all the pre-commit hooks, add --no-verify to git commit. To skip a specific hook, prefix the commit command with SKIP=<hook-id>."'
     language: system
     verbose: true
     pass_filenames: false

CMakeLists.txt

@@ -45,7 +45,7 @@ set(HIP_SUPPORTED_ARCHS "gfx906;gfx908;gfx90a;gfx942;gfx950;gfx1030;gfx1100;gfx1
 # requirements.txt files and should be kept consistent.  The ROCm torch
 # versions are derived from docker/Dockerfile.rocm
 #
-set(TORCH_SUPPORTED_VERSION_CUDA "2.7.0")
+set(TORCH_SUPPORTED_VERSION_CUDA "2.7.1")
 set(TORCH_SUPPORTED_VERSION_ROCM "2.7.0")
 
 #
@@ -171,7 +171,6 @@ if(NVCC_THREADS AND VLLM_GPU_LANG STREQUAL "CUDA")
   list(APPEND VLLM_GPU_FLAGS "--threads=${NVCC_THREADS}")
 endif()
 
-
 #
 # Use FetchContent for C++ dependencies that are compiled as part of vLLM's build process.
 # setup.py will override FETCHCONTENT_BASE_DIR to play nicely with sccache.
@@ -182,9 +181,6 @@ include(FetchContent)
 file(MAKE_DIRECTORY ${FETCHCONTENT_BASE_DIR}) # Ensure the directory exists
 message(STATUS "FetchContent base directory: ${FETCHCONTENT_BASE_DIR}")
 
-#
-# Set rocm version dev int.
-#
 if(VLLM_GPU_LANG STREQUAL "HIP")
   #
   # Overriding the default -O set up by cmake, adding ggdb3 for the most verbose devug info
@@ -192,7 +188,6 @@ if(VLLM_GPU_LANG STREQUAL "HIP")
   set(CMAKE_${VLLM_GPU_LANG}_FLAGS_DEBUG "${CMAKE_${VLLM_GPU_LANG}_FLAGS_DEBUG} -O0 -ggdb3")
   set(CMAKE_CXX_FLAGS_DEBUG "${CMAKE_CXX_FLAGS_DEBUG} -O0 -ggdb3")
 
-
   #
   # Certain HIP functions are marked as [[nodiscard]], yet vllm ignores the result which generates
   # a lot of warnings that always mask real issues. Suppressing until this is properly addressed.
@@ -236,7 +231,6 @@ endif()
 
 set(VLLM_EXT_SRC
   "csrc/mamba/mamba_ssm/selective_scan_fwd.cu"
-  "csrc/mamba/causal_conv1d/causal_conv1d.cu"
   "csrc/cache_kernels.cu"
   "csrc/attention/paged_attention_v1.cu"
   "csrc/attention/paged_attention_v2.cu"
@@ -246,6 +240,7 @@ set(VLLM_EXT_SRC
   "csrc/activation_kernels.cu"
   "csrc/layernorm_kernels.cu"
   "csrc/layernorm_quant_kernels.cu"
+  "csrc/sampler.cu"
   "csrc/cuda_view.cu"
   "csrc/quantization/gptq/q_gemm.cu"
   "csrc/quantization/compressed_tensors/int8_quant_kernels.cu"
@@ -262,7 +257,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
   SET(CUTLASS_ENABLE_HEADERS_ONLY ON CACHE BOOL "Enable only the header library")
 
   # Set CUTLASS_REVISION. Used for FetchContent. Also fixes some bogus messages when building.
-  set(CUTLASS_REVISION "v3.9.2" CACHE STRING "CUTLASS revision to use")
+  set(CUTLASS_REVISION "v4.0.0" CACHE STRING "CUTLASS revision to use")
 
   # Use the specified CUTLASS source directory for compilation if VLLM_CUTLASS_SRC_DIR is provided
   if (DEFINED ENV{VLLM_CUTLASS_SRC_DIR})
@@ -311,7 +306,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
   # Keep building Marlin for 9.0 as there are some group sizes and shapes that
   # are not supported by Machete yet.
   # 9.0 for latest bf16 atomicAdd PTX
-  cuda_archs_loose_intersection(MARLIN_ARCHS "8.0;9.0+PTX" "${CUDA_ARCHS}")
+  cuda_archs_loose_intersection(MARLIN_ARCHS "8.0;8.7;9.0+PTX" "${CUDA_ARCHS}")
   if (MARLIN_ARCHS)
 
     #
@@ -396,7 +391,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
   # The cutlass_scaled_mm kernels for Hopper (c3x, i.e. CUTLASS 3.x) require
   # CUDA 12.0 or later
   cuda_archs_loose_intersection(SCALED_MM_ARCHS "9.0a;" "${CUDA_ARCHS}")
-  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.0 AND SCALED_MM_ARCHS)
+  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.0 AND SCALED_MM_ARCHS)
     set(SRCS
        "csrc/quantization/cutlass_w8a8/scaled_mm_c3x_sm90.cu"
        "csrc/quantization/cutlass_w8a8/c3x/scaled_mm_sm90_fp8.cu"
@@ -412,7 +407,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
     list(APPEND SCALED_MM_3X_ARCHS "${SCALED_MM_ARCHS}")
     message(STATUS "Building scaled_mm_c3x_sm90 for archs: ${SCALED_MM_ARCHS}")
   else()
-    if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.0 AND SCALED_MM_ARCHS)
+    if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.0 AND SCALED_MM_ARCHS)
       message(STATUS "Not building scaled_mm_c3x_sm90 as CUDA Compiler version is "
                      "not >= 12.0, we recommend upgrading to CUDA 12.0 or "
                      "later if you intend on running FP8 quantized models on "
@@ -423,10 +418,40 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
     endif()
   endif()
 
-  # The cutlass_scaled_mm kernels for Blackwell (c3x, i.e. CUTLASS 3.x) require
+
+  # The cutlass_scaled_mm kernels for Geforce Blackwell SM120 (c3x, i.e. CUTLASS 3.x) require
   # CUDA 12.8 or later
-  cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0a;10.1a;12.0a" "${CUDA_ARCHS}")
-  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.8 AND SCALED_MM_ARCHS)
+  cuda_archs_loose_intersection(SCALED_MM_ARCHS "12.0;12.0a" "${CUDA_ARCHS}")
+  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND SCALED_MM_ARCHS)
+    set(SRCS
+      "csrc/quantization/cutlass_w8a8/scaled_mm_c3x_sm120.cu"
+      "csrc/quantization/cutlass_w8a8/c3x/scaled_mm_sm120_fp8.cu"
+    )
+    set_gencode_flags_for_srcs(
+      SRCS "${SRCS}"
+      CUDA_ARCHS "${SCALED_MM_ARCHS}")
+    list(APPEND VLLM_EXT_SRC "${SRCS}")
+    list(APPEND VLLM_GPU_FLAGS "-DENABLE_SCALED_MM_SM120=1")
+    # Let scaled_mm_c2x know it doesn't need to build these arches
+    list(APPEND SCALED_MM_3X_ARCHS "${SCALED_MM_ARCHS}")
+    message(STATUS "Building scaled_mm_c3x_sm120 for archs: ${SCALED_MM_ARCHS}")
+  else()
+    if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND SCALED_MM_ARCHS)
+      message(STATUS "Not building scaled_mm_c3x_sm120 as CUDA Compiler version is "
+                     "not >= 12.8, we recommend upgrading to CUDA 12.8 or "
+                     "later if you intend on running FP8 quantized models on "
+                     "Blackwell.")
+    else()
+      message(STATUS "Not building scaled_mm_c3x_120 as no compatible archs found "
+                     "in CUDA target architectures")
+    endif()
+  endif()
+
+
+  # The cutlass_scaled_mm kernels for Blackwell SM100 (c3x, i.e. CUTLASS 3.x)
+  # require CUDA 12.8 or later
+  cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0a;10.1a" "${CUDA_ARCHS}")
+  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND SCALED_MM_ARCHS)
     set(SRCS
       "csrc/quantization/cutlass_w8a8/scaled_mm_c3x_sm100.cu"
       "csrc/quantization/cutlass_w8a8/c3x/scaled_mm_sm100_fp8.cu"
@@ -441,7 +466,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
     list(APPEND SCALED_MM_3X_ARCHS "${SCALED_MM_ARCHS}")
     message(STATUS "Building scaled_mm_c3x_sm100 for archs: ${SCALED_MM_ARCHS}")
   else()
-    if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.8 AND SCALED_MM_ARCHS)
+    if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND SCALED_MM_ARCHS)
       message(STATUS "Not building scaled_mm_c3x_sm100 as CUDA Compiler version is "
                      "not >= 12.8, we recommend upgrading to CUDA 12.8 or "
                      "later if you intend on running FP8 quantized models on "
@@ -457,7 +482,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
   # kernels for the remaining archs that are not already built for 3x.
   # (Build 8.9 for FP8)
   cuda_archs_loose_intersection(SCALED_MM_2X_ARCHS
-    "7.5;8.0;8.9+PTX" "${CUDA_ARCHS}")
+    "7.5;8.0;8.7;8.9+PTX" "${CUDA_ARCHS}")
   # subtract out the archs that are already built for 3x
   list(REMOVE_ITEM SCALED_MM_2X_ARCHS ${SCALED_MM_3X_ARCHS})
   if (SCALED_MM_2X_ARCHS)
@@ -484,7 +509,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
   # The 2:4 sparse kernels cutlass_scaled_sparse_mm and cutlass_compressor
   # require CUDA 12.2 or later (and only work on Hopper).
   cuda_archs_loose_intersection(SCALED_MM_ARCHS "9.0a;" "${CUDA_ARCHS}")
-  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.2 AND SCALED_MM_ARCHS)
+  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.2 AND SCALED_MM_ARCHS)
     set(SRCS "csrc/sparse/cutlass/sparse_scaled_mm_c3x.cu")
     set_gencode_flags_for_srcs(
       SRCS "${SRCS}"
@@ -493,7 +518,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
     list(APPEND VLLM_GPU_FLAGS "-DENABLE_SPARSE_SCALED_MM_C3X=1")
     message(STATUS "Building sparse_scaled_mm_c3x for archs: ${SCALED_MM_ARCHS}")
   else()
-    if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.2 AND SCALED_MM_ARCHS)
+    if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.2 AND SCALED_MM_ARCHS)
       message(STATUS "Not building sparse_scaled_mm_c3x kernels as CUDA Compiler version is "
                      "not >= 12.2, we recommend upgrading to CUDA 12.2 or later "
                      "if you intend on running FP8 sparse quantized models on Hopper.")
@@ -505,7 +530,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
 
   # FP4 Archs and flags
   cuda_archs_loose_intersection(FP4_ARCHS "10.0a" "${CUDA_ARCHS}")
-  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.8 AND FP4_ARCHS)
+  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND FP4_ARCHS)
     set(SRCS
       "csrc/quantization/fp4/nvfp4_quant_kernels.cu"
       "csrc/quantization/fp4/nvfp4_experts_quant.cu"
@@ -516,6 +541,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
       CUDA_ARCHS "${FP4_ARCHS}")
     list(APPEND VLLM_EXT_SRC "${SRCS}")
     list(APPEND VLLM_GPU_FLAGS "-DENABLE_NVFP4=1")
+    list(APPEND VLLM_GPU_FLAGS "-DENABLE_CUTLASS_MOE_SM100=1")
     message(STATUS "Building NVFP4 for archs: ${FP4_ARCHS}")
   else()
     message(STATUS "Not building NVFP4 as no compatible archs were found.")
@@ -525,9 +551,10 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
 
   # CUTLASS MLA Archs and flags
   cuda_archs_loose_intersection(MLA_ARCHS "10.0a" "${CUDA_ARCHS}")
-  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.8 AND MLA_ARCHS)
+  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND MLA_ARCHS)
     set(SRCS
-      "csrc/attention/mla/cutlass_mla_kernels.cu")
+      "csrc/attention/mla/cutlass_mla_kernels.cu"
+      "csrc/attention/mla/sm100_cutlass_mla_kernel.cu")
     set_gencode_flags_for_srcs(
       SRCS "${SRCS}"
       CUDA_ARCHS "${MLA_ARCHS}")
@@ -545,13 +572,12 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
 
   # CUTLASS MoE kernels
 
-  # The MoE kernel cutlass_moe_mm requires CUDA 12.3 or later (and only works
-  # on Hopper). get_cutlass_moe_mm_data should only be compiled if it's possible
-  # to compile MoE kernels that use its output.
-  cuda_archs_loose_intersection(SCALED_MM_ARCHS "9.0a;10.0a" "${CUDA_ARCHS}")
+  # The MoE kernel cutlass_moe_mm requires CUDA 12.3 or later (and ONLY works
+  # on Hopper). get_cutlass_(pplx_)moe_mm_data should only be compiled
+  # if it's possible to compile MoE kernels that use its output.
+  cuda_archs_loose_intersection(SCALED_MM_ARCHS "9.0a" "${CUDA_ARCHS}")
   if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.3 AND SCALED_MM_ARCHS)
-    set(SRCS "csrc/quantization/cutlass_w8a8/moe/grouped_mm_c3x.cu"
-             "csrc/quantization/cutlass_w8a8/moe/moe_data.cu")
+    set(SRCS "csrc/quantization/cutlass_w8a8/moe/grouped_mm_c3x.cu")
     set_gencode_flags_for_srcs(
       SRCS "${SRCS}"
       CUDA_ARCHS "${SCALED_MM_ARCHS}")
@@ -565,6 +591,46 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
                      "if you intend on running FP8 quantized MoE models on Hopper.")
     else()
       message(STATUS "Not building grouped_mm_c3x as no compatible archs found "
+                     "in CUDA target architectures.")
+    endif()
+  endif()
+
+  # moe_data.cu is used by all CUTLASS MoE kernels.
+  cuda_archs_loose_intersection(CUTLASS_MOE_DATA_ARCHS "9.0a;10.0a" "${CUDA_ARCHS}")
+  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.3 AND CUTLASS_MOE_DATA_ARCHS)
+    set(SRCS "csrc/quantization/cutlass_w8a8/moe/moe_data.cu")
+    set_gencode_flags_for_srcs(
+      SRCS "${SRCS}"
+      CUDA_ARCHS "${CUTLASS_MOE_DATA_ARCHS}")
+    list(APPEND VLLM_EXT_SRC "${SRCS}")
+    message(STATUS "Building moe_data for archs: ${CUTLASS_MOE_DATA_ARCHS}")
+  else()
+    if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.3 AND CUTLASS_MOE_DATA_ARCHS)
+      message(STATUS "Not building moe_data as CUDA Compiler version is "
+                     "not >= 12.3, we recommend upgrading to CUDA 12.3 or later "
+                     "if you intend on running FP8 quantized MoE models on Hopper or Blackwell.")
+    else()
+      message(STATUS "Not building moe_data as no compatible archs found "
+                     "in CUDA target architectures.")
+    endif()
+  endif()
+  
+  cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0a" "${CUDA_ARCHS}")
+  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND SCALED_MM_ARCHS)
+    set(SRCS "csrc/quantization/cutlass_w8a8/moe/blockwise_scaled_group_mm_sm100.cu")
+    set_gencode_flags_for_srcs(
+      SRCS "${SRCS}"
+      CUDA_ARCHS "${SCALED_MM_ARCHS}")
+    list(APPEND VLLM_EXT_SRC "${SRCS}")
+    list(APPEND VLLM_GPU_FLAGS "-DENABLE_CUTLASS_MOE_SM100=1")
+    message(STATUS "Building blockwise_scaled_group_mm_sm100 for archs: ${SCALED_MM_ARCHS}")
+  else()
+    if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND SCALED_MM_ARCHS)
+      message(STATUS "Not building blockwise_scaled_group_mm_sm100 kernels as CUDA Compiler version is "
+                     "not >= 12.8, we recommend upgrading to CUDA 12.8 or later "
+                     "if you intend on running FP8 quantized MoE models on Blackwell.")
+    else()
+      message(STATUS "Not building blockwise_scaled_group_mm_sm100 as no compatible archs found "
                      "in CUDA target architectures")
     endif()
   endif()
@@ -575,7 +641,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
   # The machete kernels only work on hopper and require CUDA 12.0 or later.
   # Only build Machete kernels if we are building for something compatible with sm90a
   cuda_archs_loose_intersection(MACHETE_ARCHS "9.0a" "${CUDA_ARCHS}")
-  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.0 AND MACHETE_ARCHS)
+  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.0 AND MACHETE_ARCHS)
     #
     # For the Machete kernels we automatically generate sources for various
     # preselected input type pairs and schedules.
@@ -627,7 +693,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
 
     message(STATUS "Building Machete kernels for archs: ${MACHETE_ARCHS}")
   else()
-    if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.0
+    if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.0
         AND MACHETE_ARCHS)
       message(STATUS "Not building Machete kernels as CUDA Compiler version is "
                      "not >= 12.0, we recommend upgrading to CUDA 12.0 or "
@@ -641,6 +707,14 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
 # if CUDA endif
 endif()
 
+if (VLLM_GPU_LANG STREQUAL "HIP")
+  # Add QuickReduce kernels
+  list(APPEND VLLM_EXT_SRC
+    "csrc/custom_quickreduce.cu"
+  )
+# if ROCM endif
+endif()
+
 message(STATUS "Enabling C extension.")
 define_gpu_extension_target(
   _C
@@ -687,7 +761,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
 
   list(APPEND VLLM_MOE_EXT_SRC "${VLLM_MOE_WNA16_SRC}")
   # 9.0 for latest bf16 atomicAdd PTX
-  cuda_archs_loose_intersection(MARLIN_MOE_ARCHS "8.0;9.0+PTX" "${CUDA_ARCHS}")
+  cuda_archs_loose_intersection(MARLIN_MOE_ARCHS "8.0;8.7;9.0+PTX" "${CUDA_ARCHS}")
   if (MARLIN_MOE_ARCHS)
 
     #

README.md

@@ -58,28 +58,26 @@ vLLM is fast with:
 - Efficient management of attention key and value memory with [**PagedAttention**](https://blog.vllm.ai/2023/06/20/vllm.html)
 - Continuous batching of incoming requests
 - Fast model execution with CUDA/HIP graph
-- Quantizations: [GPTQ](https://arxiv.org/abs/2210.17323), [AWQ](https://arxiv.org/abs/2306.00978), [AutoRound](https://arxiv.org/abs/2309.05516),INT4, INT8, and FP8.
-- Optimized CUDA kernels, including integration with FlashAttention and FlashInfer.
+- Quantizations: [GPTQ](https://arxiv.org/abs/2210.17323), [AWQ](https://arxiv.org/abs/2306.00978), [AutoRound](https://arxiv.org/abs/2309.05516), INT4, INT8, and FP8
+- Optimized CUDA kernels, including integration with FlashAttention and FlashInfer
 - Speculative decoding
 - Chunked prefill
 
-**Performance benchmark**: We include a performance benchmark at the end of [our blog post](https://blog.vllm.ai/2024/09/05/perf-update.html). It compares the performance of vLLM against other LLM serving engines ([TensorRT-LLM](https://github.com/NVIDIA/TensorRT-LLM), [SGLang](https://github.com/sgl-project/sglang) and [LMDeploy](https://github.com/InternLM/lmdeploy)). The implementation is under [nightly-benchmarks folder](.buildkite/nightly-benchmarks/) and you can [reproduce](https://github.com/vllm-project/vllm/issues/8176) this benchmark using our one-click runnable script.
-
 vLLM is flexible and easy to use with:
 
 - Seamless integration with popular Hugging Face models
 - High-throughput serving with various decoding algorithms, including *parallel sampling*, *beam search*, and more
-- Tensor parallelism and pipeline parallelism support for distributed inference
+- Tensor, pipeline, data and expert parallelism support for distributed inference
 - Streaming outputs
 - OpenAI-compatible API server
-- Support NVIDIA GPUs, AMD CPUs and GPUs, Intel CPUs and GPUs, PowerPC CPUs, TPU, and AWS Neuron.
+- Support NVIDIA GPUs, AMD CPUs and GPUs, Intel CPUs and GPUs, PowerPC CPUs, TPU, and AWS Neuron
 - Prefix caching support
 - Multi-LoRA support
 
 vLLM seamlessly supports most popular open-source models on HuggingFace, including:
 - Transformer-like LLMs (e.g., Llama)
 - Mixture-of-Expert LLMs (e.g., Mixtral, Deepseek-V2 and V3)
-- Embedding Models (e.g. E5-Mistral)
+- Embedding Models (e.g., E5-Mistral)
 - Multi-modal LLMs (e.g., LLaVA)
 
 Find the full list of supported models [here](https://docs.vllm.ai/en/latest/models/supported_models.html).
@@ -154,12 +152,14 @@ If you use vLLM for your research, please cite our [paper](https://arxiv.org/abs
 
 ## Contact Us
 
+<!-- --8<-- [start:contact-us] -->
 - For technical questions and feature requests, please use GitHub [Issues](https://github.com/vllm-project/vllm/issues) or [Discussions](https://github.com/vllm-project/vllm/discussions)
 - For discussing with fellow users, please use the [vLLM Forum](https://discuss.vllm.ai)
-- coordinating contributions and development, please use [Slack](https://slack.vllm.ai)
+- For coordinating contributions and development, please use [Slack](https://slack.vllm.ai)
 - For security disclosures, please use GitHub's [Security Advisories](https://github.com/vllm-project/vllm/security/advisories) feature
 - For collaborations and partnerships, please contact us at [vllm-questions@lists.berkeley.edu](mailto:vllm-questions@lists.berkeley.edu)
+<!-- --8<-- [end:contact-us] -->
 
 ## Media Kit
 
-- If you wish to use vLLM's logo, please refer to [our media kit repo](https://github.com/vllm-project/media-kit).
+- If you wish to use vLLM's logo, please refer to [our media kit repo](https://github.com/vllm-project/media-kit)

benchmarks/README.md

@@ -4,7 +4,7 @@ This README guides you through running benchmark tests with the extensive
 datasets supported on vLLM. It’s a living document, updated as new features and datasets
 become available.
 
-## Dataset Overview
+**Dataset Overview**
 
 <table style="width:100%; border-collapse: collapse;">
   <thead>
@@ -82,7 +82,10 @@ become available.
 **Note**: HuggingFace dataset's `dataset-name` should be set to `hf`
 
 ---
-## Example - Online Benchmark
+<details>
+<summary><b>🚀 Example - Online Benchmark</b></summary>
+
+<br/>
 
 First start serving your model
 
@@ -130,7 +133,8 @@ P99 ITL (ms):                            8.39
 ==================================================
 ```
 
-### Custom Dataset
+**Custom Dataset**
+
 If the dataset you want to benchmark is not supported yet in vLLM, even then you can benchmark on it using `CustomDataset`. Your data needs to be in `.jsonl` format and needs to have "prompt" field per entry, e.g., data.jsonl
 
 ```
@@ -162,7 +166,7 @@ python3 benchmarks/benchmark_serving.py --port 9001 --save-result --save-detaile
 
 You can skip applying chat template if your data already has it by using `--custom-skip-chat-template`.
 
-### VisionArena Benchmark for Vision Language Models
+**VisionArena Benchmark for Vision Language Models**
 
 ```bash
 # need a model with vision capability here
@@ -180,7 +184,7 @@ python3 vllm/benchmarks/benchmark_serving.py \
   --num-prompts 1000
 ```
 
-### InstructCoder Benchmark with Speculative Decoding
+**InstructCoder Benchmark with Speculative Decoding**
 
 ``` bash
 VLLM_USE_V1=1 vllm serve meta-llama/Meta-Llama-3-8B-Instruct \
@@ -197,7 +201,7 @@ python3 benchmarks/benchmark_serving.py \
     --num-prompts 2048
 ```
 
-### Other HuggingFaceDataset Examples
+**Other HuggingFaceDataset Examples**
 
 ```bash
 vllm serve Qwen/Qwen2-VL-7B-Instruct --disable-log-requests
@@ -251,7 +255,7 @@ python3 vllm/benchmarks/benchmark_serving.py \
     --num-prompts 80
 ```
 
-### Running With Sampling Parameters
+**Running With Sampling Parameters**
 
 When using OpenAI-compatible backends such as `vllm`, optional sampling
 parameters can be specified. Example client command:
@@ -269,8 +273,27 @@ python3 vllm/benchmarks/benchmark_serving.py \
   --num-prompts 10
 ```
 
----
-## Example - Offline Throughput Benchmark
+**Running With Ramp-Up Request Rate**
+
+The benchmark tool also supports ramping up the request rate over the
+duration of the benchmark run. This can be useful for stress testing the
+server or finding the maximum throughput that it can handle, given some latency budget.
+
+Two ramp-up strategies are supported:
+- `linear`: Increases the request rate linearly from a start value to an end value.
+- `exponential`: Increases the request rate exponentially.
+
+The following arguments can be used to control the ramp-up:
+- `--ramp-up-strategy`: The ramp-up strategy to use (`linear` or `exponential`).
+- `--ramp-up-start-rps`: The request rate at the beginning of the benchmark.
+- `--ramp-up-end-rps`: The request rate at the end of the benchmark.
+
+</details>
+
+<details>
+<summary><b>📈 Example - Offline Throughput Benchmark</b></summary>
+
+<br/>
 
 ```bash
 python3 vllm/benchmarks/benchmark_throughput.py \
@@ -288,7 +311,7 @@ Total num prompt tokens:  5014
 Total num output tokens:  1500
 ```
 
-### VisionArena Benchmark for Vision Language Models
+**VisionArena Benchmark for Vision Language Models**
 
 ``` bash
 python3 vllm/benchmarks/benchmark_throughput.py \
@@ -308,7 +331,7 @@ Total num prompt tokens:  14527
 Total num output tokens:  1280
 ```
 
-### InstructCoder Benchmark with Speculative Decoding
+**InstructCoder Benchmark with Speculative Decoding**
 
 ``` bash
 VLLM_WORKER_MULTIPROC_METHOD=spawn \
@@ -332,7 +355,7 @@ Total num prompt tokens:  261136
 Total num output tokens:  204800
 ```
 
-### Other HuggingFaceDataset Examples
+**Other HuggingFaceDataset Examples**
 
 **`lmms-lab/LLaVA-OneVision-Data`**
 
@@ -371,7 +394,7 @@ python3 benchmarks/benchmark_throughput.py \
   --num-prompts 10
 ```
 
-### Benchmark with LoRA Adapters
+**Benchmark with LoRA Adapters**
 
 ``` bash
 # download dataset
@@ -387,3 +410,196 @@ python3 vllm/benchmarks/benchmark_throughput.py \
   --enable-lora \
   --lora-path yard1/llama-2-7b-sql-lora-test
   ```
+
+</details>
+
+<details>
+<summary><b>🛠️ Example - Structured Output Benchmark</b></summary>
+
+<br/>
+
+Benchmark the performance of structured output generation (JSON, grammar, regex).
+
+**Server Setup**
+
+```bash
+vllm serve NousResearch/Hermes-3-Llama-3.1-8B --disable-log-requests
+```
+
+**JSON Schema Benchmark**
+
+```bash
+python3 benchmarks/benchmark_serving_structured_output.py \
+  --backend vllm \
+  --model NousResearch/Hermes-3-Llama-3.1-8B \
+  --dataset json \
+  --structured-output-ratio 1.0 \
+  --request-rate 10 \
+  --num-prompts 1000
+```
+
+**Grammar-based Generation Benchmark**
+
+```bash
+python3 benchmarks/benchmark_serving_structured_output.py \
+  --backend vllm \
+  --model NousResearch/Hermes-3-Llama-3.1-8B \
+  --dataset grammar \
+  --structure-type grammar \
+  --request-rate 10 \
+  --num-prompts 1000
+```
+
+**Regex-based Generation Benchmark**
+
+```bash
+python3 benchmarks/benchmark_serving_structured_output.py \
+  --backend vllm \
+  --model NousResearch/Hermes-3-Llama-3.1-8B \
+  --dataset regex \
+  --request-rate 10 \
+  --num-prompts 1000
+```
+
+**Choice-based Generation Benchmark**
+
+```bash
+python3 benchmarks/benchmark_serving_structured_output.py \
+  --backend vllm \
+  --model NousResearch/Hermes-3-Llama-3.1-8B \
+  --dataset choice \
+  --request-rate 10 \
+  --num-prompts 1000
+```
+
+**XGrammar Benchmark Dataset**
+
+```bash
+python3 benchmarks/benchmark_serving_structured_output.py \
+  --backend vllm \
+  --model NousResearch/Hermes-3-Llama-3.1-8B \
+  --dataset xgrammar_bench \
+  --request-rate 10 \
+  --num-prompts 1000
+```
+
+</details>
+
+<details>
+<summary><b>📚 Example - Long Document QA Benchmark</b></summary>
+
+<br/>
+
+Benchmark the performance of long document question-answering with prefix caching.
+
+**Basic Long Document QA Test**
+
+```bash
+python3 benchmarks/benchmark_long_document_qa_throughput.py \
+  --model meta-llama/Llama-2-7b-chat-hf \
+  --enable-prefix-caching \
+  --num-documents 16 \
+  --document-length 2000 \
+  --output-len 50 \
+  --repeat-count 5
+```
+
+**Different Repeat Modes**
+
+```bash
+# Random mode (default) - shuffle prompts randomly
+python3 benchmarks/benchmark_long_document_qa_throughput.py \
+  --model meta-llama/Llama-2-7b-chat-hf \
+  --enable-prefix-caching \
+  --num-documents 8 \
+  --document-length 3000 \
+  --repeat-count 3 \
+  --repeat-mode random
+
+# Tile mode - repeat entire prompt list in sequence
+python3 benchmarks/benchmark_long_document_qa_throughput.py \
+  --model meta-llama/Llama-2-7b-chat-hf \
+  --enable-prefix-caching \
+  --num-documents 8 \
+  --document-length 3000 \
+  --repeat-count 3 \
+  --repeat-mode tile
+
+# Interleave mode - repeat each prompt consecutively
+python3 benchmarks/benchmark_long_document_qa_throughput.py \
+  --model meta-llama/Llama-2-7b-chat-hf \
+  --enable-prefix-caching \
+  --num-documents 8 \
+  --document-length 3000 \
+  --repeat-count 3 \
+  --repeat-mode interleave
+```
+
+</details>
+
+<details>
+<summary><b>🗂️ Example - Prefix Caching Benchmark</b></summary>
+
+<br/>
+
+Benchmark the efficiency of automatic prefix caching.
+
+**Fixed Prompt with Prefix Caching**
+
+```bash
+python3 benchmarks/benchmark_prefix_caching.py \
+  --model meta-llama/Llama-2-7b-chat-hf \
+  --enable-prefix-caching \
+  --num-prompts 1 \
+  --repeat-count 100 \
+  --input-length-range 128:256
+```
+
+**ShareGPT Dataset with Prefix Caching**
+
+```bash
+# download dataset
+# wget https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/resolve/main/ShareGPT_V3_unfiltered_cleaned_split.json
+
+python3 benchmarks/benchmark_prefix_caching.py \
+  --model meta-llama/Llama-2-7b-chat-hf \
+  --dataset-path /path/ShareGPT_V3_unfiltered_cleaned_split.json \
+  --enable-prefix-caching \
+  --num-prompts 20 \
+  --repeat-count 5 \
+  --input-length-range 128:256
+```
+
+</details>
+
+<details>
+<summary><b>⚡ Example - Request Prioritization Benchmark</b></summary>
+
+<br/>
+
+Benchmark the performance of request prioritization in vLLM.
+
+**Basic Prioritization Test**
+
+```bash
+python3 benchmarks/benchmark_prioritization.py \
+  --model meta-llama/Llama-2-7b-chat-hf \
+  --input-len 128 \
+  --output-len 64 \
+  --num-prompts 100 \
+  --scheduling-policy priority
+```
+
+**Multiple Sequences per Prompt**
+
+```bash
+python3 benchmarks/benchmark_prioritization.py \
+  --model meta-llama/Llama-2-7b-chat-hf \
+  --input-len 128 \
+  --output-len 64 \
+  --num-prompts 100 \
+  --scheduling-policy priority \
+  --n 2
+```
+
+</details>

benchmarks/auto_tune.sh

@@ -1,212 +0,0 @@
-#!/bin/bash
-
-# This script aims to tune the best server parameter combinations to maximize throughput for given requirement. 
-# The current server parameter combination is  max_num_seqs and max_num_batched_tokens
-# It also supports additional requirement: e2e latency and prefix cache. 
-
-# Pre-requisite:
-# 1. Checkout to your branch, install/ update the correct running env. For TPU, activate conda env and install the corresponding torch, xla version. 
-# 2. If the model is customized, replace the MODEL's config with the customized config.
-# 3. Set variables (ALL REQUIRED)
-#   BASE: your directory for vllm repo
-#   MODEL: the model served by vllm
-#   DOWNLOAD_DIR: directory to download and load model weights.
-#   INPUT_LEN: request input len
-#   OUTPUT_LEN: request output len
-#   MIN_CACHE_HIT_PCT: prefix cache rate
-#   MAX_LATENCY_ALLOWED_MS: (e2e) latency requirement. If there's no latency requirement, set it to a large number like 1000000000
-# 4. Run the script, it might take a long time, you can use tmux to avoid the script stop if disconnection happens.
-# 5. The final result will be saved in RESULT file. 
-
-
-# Example use cases 
-# 1. Given input_len=1800, output_len=20, what's the best max_num_seqs and max_num_batched_tokens to get highest throughput?
-# Use INPUT_LEN=1800,  OUTPUT_LEN=20, MIN_CACHE_HIT_PCT=0, MAX_LATENCY_ALLOWED_MS=100000000000
-# 2. If we have latency requirement to be lower than 500ms, what's the best server parameter?
-# Use INPUT_LEN=1800,  OUTPUT_LEN=20, MIN_CACHE_HIT_PCT=0, MAX_LATENCY_ALLOWED_MS=500
-# 3. If we want to reach 60% prefix cache, what's the best server parameter? 
-# Use INPUT_LEN=1800,  OUTPUT_LEN=20, MIN_CACHE_HIT_PCT=60, MAX_LATENCY_ALLOWED_MS=500
-
-TAG=$(date +"%Y_%m_%d_%H_%M")
-BASE=""
-MODEL="meta-llama/Llama-3.1-8B-Instruct"
-DOWNLOAD_DIR=""
-INPUT_LEN=4000
-OUTPUT_LEN=16
-MIN_CACHE_HIT_PCT_PCT=0
-MAX_LATENCY_ALLOWED_MS=100000000000
-
-LOG_FOLDER="$BASE/auto-benchmark/$TAG"
-RESULT="$LOG_FOLDER/result.txt"
-
-echo "result file$ $RESULT"
-echo "model: $MODEL"
-echo
-
-rm -rf $LOG_FOLDER
-mkdir -p $LOG_FOLDER
-
-cd "$BASE/vllm"
-# create sonnet-4x.txt so that we can sample 2048 tokens for input
-echo "" > benchmarks/sonnet_4x.txt
-for _ in {1..4}
-do
-cat benchmarks/sonnet.txt >> benchmarks/sonnet_4x.txt
-done
-
-pip install datasets
-
-current_hash=$(git rev-parse HEAD)
-echo "hash:$current_hash" >> "$RESULT"
-echo "current_hash: $current_hash"
-
-best_throughput=0
-best_max_num_seqs=0
-best_num_batched_tokens=0
-best_goodput=0
-run_benchmark() {
-    local max_num_seqs=$1
-    local max_num_batched_tokens=$2
-    echo "max_num_seq: $max_num_seqs, max_num_batched_tokens: $max_num_batched_tokens"
-    local vllm_log="$LOG_FOLDER/vllm_log_${max_num_seqs}_${max_num_batched_tokens}.txt"
-    echo "vllm_log: $vllm_log"
-    echo
-    rm -f $vllm_log
-
-    # start the server
-    VLLM_USE_V1=1 VLLM_SERVER_DEV_MODE=1 vllm serve $MODEL \
-        --disable-log-requests \
-        --port 8004 \
-        --gpu-memory-utilization 0.98 \
-        --max-num-seqs $max_num_seqs \
-        --max-num-batched-tokens $max_num_batched_tokens \
-        --tensor-parallel-size 1 \
-        --enable-prefix-caching \
-        --load-format dummy \
-        --download-dir $DOWNLOAD_DIR \
-        --max-model-len $(( INPUT_LEN+OUTPUT_LEN )) > "$vllm_log" 2>&1 &
-    echo "wait for 10 minutes.."
-    echo
-    # wait for 10 minutes...
-    server_started=0
-    for i in {1..60}; do        
-        if grep -Fq "Application startup complete" "$vllm_log"; then
-            echo "Application started"
-            server_started=1
-            break
-        else
-            # echo "wait for 10 seconds..."
-            sleep 10
-        fi
-    done
- 
-    if (( ! server_started )); then
-        echo "server did not start within 10 minutes, terminate the benchmarking. Please check server log at $vllm_log"
-        echo "pkill -f vllm"
-        echo
-        pkill vllm
-        sleep 10
-        return 1
-    fi
-    
-    echo "run benchmark test..."
-    echo
-    meet_latency_requirement=0
-    # get a basic qps by using request-rate inf
-    bm_log="$LOG_FOLDER/bm_log_${max_num_seqs}_${max_num_batched_tokens}_requestrate_inf.txt"
-    prefix_len=$(( INPUT_LEN * MIN_CACHE_HIT_PCT / 100 ))
-    python benchmarks/benchmark_serving.py \
-        --backend vllm \
-        --model $MODEL  \
-        --dataset-name sonnet \
-        --dataset-path benchmarks/sonnet_4x.txt \
-        --sonnet-input-len $INPUT_LEN \
-        --sonnet-output-len $OUTPUT_LEN \
-        --ignore-eos \
-        --disable-tqdm \
-        --request-rate inf \
-        --percentile-metrics ttft,tpot,itl,e2el \
-        --goodput e2el:$MAX_LATENCY_ALLOWED_MS \
-        --num-prompts 100 \
-        --sonnet-prefix-len $prefix_len \
-        --port 8004 > "$bm_log"
-    through_put=$(grep "Request throughput (req/s):" "$bm_log" | sed 's/[^0-9.]//g')
-    e2el=$(grep "P99 E2EL (ms):" "$bm_log" | awk '{print $NF}')
-    goodput=$(grep "Request goodput (req/s):" "$bm_log" | sed 's/[^0-9.]//g')
-
-    if (( $(echo "$e2el <= $MAX_LATENCY_ALLOWED_MS" | bc -l) )); then
-        meet_latency_requirement=1
-    fi
-
-    if (( ! meet_latency_requirement )); then
-    # start from request-rate as int(through_put) + 1
-        request_rate=$((${through_put%.*} + 1))
-        while ((request_rate > 0)); do
-            # clear prefix cache
-            curl -X POST http://0.0.0.0:8004/reset_prefix_cache
-            sleep 5
-            bm_log="$LOG_FOLDER/bm_log_${max_num_seqs}_${max_num_batched_tokens}_requestrate_${request_rate}.txt"
-            python benchmarks/benchmark_serving.py \
-                --backend vllm \
-                --model $MODEL  \
-                --dataset-name sonnet \
-                --dataset-path benchmarks/sonnet_4x.txt \
-                --sonnet-input-len $INPUT_LEN \
-                --sonnet-output-len $OUTPUT_LEN \
-                --ignore_eos \
-                --disable-tqdm \
-                --request-rate $request_rate \
-                --percentile-metrics ttft,tpot,itl,e2el \
-                --goodput e2el:$MAX_LATENCY_ALLOWED_MS \
-                --num-prompts 100 \
-                --sonnet-prefix-len $prefix_len \
-                --port 8004 > "$bm_log"
-            through_put=$(grep "Request throughput (req/s):" "$bm_log" | sed 's/[^0-9.]//g')
-            e2el=$(grep "P99 E2EL (ms):" "$bm_log" | awk '{print $NF}')
-            goodput=$(grep "Request goodput (req/s):" "$bm_log" | sed 's/[^0-9.]//g')
-            if (( $(echo "$e2el <= $MAX_LATENCY_ALLOWED_MS" | bc -l) )); then
-                meet_latency_requirement=1
-                break
-            fi
-            request_rate=$((request_rate-1))
-        done
-    fi
-    # write the results and update the best result.
-    if ((meet_latency_requirement)); then
-        echo "max_num_seqs: $max_num_seqs, max_num_batched_tokens: $max_num_batched_tokens, request_rate: $request_rate, e2el: $e2el, through put: $through_put, goodput: $goodput"
-        echo "max_num_seqs: $max_num_seqs, max_num_batched_tokens: $max_num_batched_tokens, request_rate: $request_rate, e2el: $e2el, through put: $through_put, goodput: $goodput" >> "$RESULT"
-        if (( $(echo "$through_put > $best_throughput" | bc -l) )); then
-            best_throughput=$through_put
-            best_max_num_seqs=$max_num_seqs
-            best_num_batched_tokens=$max_num_batched_tokens
-            best_goodput=$goodput
-        fi
-    else
-        echo "max_num_seqs: $max_num_seqs, max_num_batched_tokens: $max_num_batched_tokens does not meet latency requirement ${MAX_LATENCY_ALLOWED_MS}"
-        echo "max_num_seqs: $max_num_seqs, max_num_batched_tokens: $max_num_batched_tokens does not meet latency requirement ${MAX_LATENCY_ALLOWED_MS}" >> "$RESULT"
-    fi
-
-    echo "best_max_num_seqs: $best_max_num_seqs, best_num_batched_tokens: $best_num_batched_tokens, best_throughput: $best_throughput"
-
-    echo "pkill -f vllm"
-    echo
-    pkill vllm
-    sleep 10
-    rm -f $vllm_log
-    printf '=%.0s' $(seq 1 20)
-    return 0
-}
-
-
-num_seqs_list="128 256"
-num_batched_tokens_list="512 1024 2048 4096"
-for num_seqs in $num_seqs_list; do
-    for num_batched_tokens in $num_batched_tokens_list; do
-        run_benchmark $num_seqs $num_batched_tokens
-        exit 0
-    done
-done
-echo "finish permutations"
-echo "best_max_num_seqs: $best_max_num_seqs, best_num_batched_tokens: $best_num_batched_tokens, best_throughput: $best_throughput"
-echo "best_max_num_seqs: $best_max_num_seqs, best_num_batched_tokens: $best_num_batched_tokens, best_throughput: $best_throughput" >> "$RESULT"
-

benchmarks/auto_tune/README.md

@@ -0,0 +1,137 @@
+# Automated vLLM Server Parameter Tuning
+
+This script automates the process of finding the optimal server parameter combination (`max-num-seqs` and `max-num-batched-tokens`) to maximize throughput for a vLLM server. It also supports additional constraints such as E2E latency and prefix cache hit rate.
+
+## Table of Contents
+- [Prerequisites](#prerequisites)
+- [Configuration](#configuration)
+- [How to Run](#how-to-run)
+- [Example Use Cases](#example-use-cases)
+- [Output](#output)
+- [How It Works](#how-it-works)
+
+## Prerequisites
+
+Before running the script, please ensure the following steps are completed:
+
+1. **Clone vLLM & Set Up Branch**: Clone the vLLM repository and check out to your desired branch.
+
+```bash
+git clone https://github.com/vllm-project/vllm.git
+cd vllm
+# git checkout <your-branch>
+```
+
+1. **Install Environment**: Install or update the correct running environment. For TPU usage, activate your `conda` environment and install the corresponding `torch` and `torch_xla` versions.
+
+2. **Model Configuration**: If you are using a customized model, ensure its configuration files are correctly placed and accessible.
+
+## Configuration
+
+You must set the following variables at the top of the script before execution.
+
+| Variable | Description | Example Value |
+| --- | --- | --- |
+| `BASE` | **Required.** The absolute path to the parent directory of your vLLM repository directory. | `"$HOME"` |
+| `MODEL` | **Required.** The Hugging Face model identifier to be served by vllm. | `"meta-llama/Llama-3.1-8B-Instruct"` |
+| `SYSTEM`| **Required.** The hardware you are running on. Choices: `TPU` or `GPU`. (For other systems, it might not support saving profiles) | `"TPU"` |
+| `TP` | **Required.** The tensor-parallelism size. | `1` |
+| `DOWNLOAD_DIR` | **Required.** Directory to download and load model weights from. | `""` (default download path) |
+| `INPUT_LEN` | **Required.** Request input length. | `4000` |
+| `OUTPUT_LEN` | **Required.** Request output length. | `16` |
+| `MIN_CACHE_HIT_PCT` | Prefix cache hit rate in percentage (0-100). Set to `0` to disable. | `60` |
+| `MAX_LATENCY_ALLOWED_MS` | The maximum allowed P99 end-to-end latency in milliseconds. Set to a very large number (e.g., `100000000000`) to effectively ignore the latency constraint. | `500` |
+| `NUM_SEQS_LIST` | A space-separated string of `max-num-seqs` values to test. | `"128 256"` |
+| `NUM_BATCHED_TOKENS_LIST` | A space-separated string of `max-num-batched-tokens` values to test. | `"1024 2048 4096"` |
+
+**Note**: The default `NUM_SEQS_LIST` and `NUM_BATCHED_TOKENS_LIST` are set for medium-sized inputs/outputs. For very short contexts (e.g., 20 input, 20 output tokens), you may need to test larger values for `max-num-seqs`.
+
+## How to Run
+
+1. **Configure**: Edit the script and set the variables in the [Configuration](#configuration) section.
+2. **Execute**: Run the script. Since the process can take a long time, it is highly recommended to use a terminal multiplexer like `tmux` or `screen` to prevent the script from stopping if your connection is lost.
+
+```
+cd <FOLDER_OF_THIS_SCRIPT>
+bash auto_tune.sh
+```
+
+    Please note that the `bash auto_tune.sh` command cannot contain full or partial path with keyword `vllm`, otherwise `pkill -f vllm` command will also kill this script itself.
+
+## Example Use Cases
+
+Here are a few examples of how to configure the script for different goals:
+
+### 1. Maximize Throughput (No Latency Constraint)
+- **Goal**: Find the best `max-num-seqs` and `max-num-batched-tokens` to get the highest possible throughput for 1800 input tokens and 20 output tokens.
+- **Configuration**:
+
+```bash
+INPUT_LEN=1800
+OUTPUT_LEN=20
+MIN_CACHE_HIT_PCT=0
+MAX_LATENCY_ALLOWED_MS=100000000000 # A very large number
+```
+
+#### 2. Maximize Throughput with a Latency Requirement
+- **Goal**: Find the best server parameters when P99 end-to-end latency must be below 500ms.
+- **Configuration**:
+
+```bash
+INPUT_LEN=1800
+OUTPUT_LEN=20
+MIN_CACHE_HIT_PCT=0
+MAX_LATENCY_ALLOWED_MS=500
+```
+
+#### 3. Maximize Throughput with Prefix Caching and Latency Requirements
+- **Goal**: Find the best server parameters assuming a 60% prefix cache hit rate and a latency requirement of 500ms.
+- **Configuration**:
+
+```bash
+INPUT_LEN=1800
+OUTPUT_LEN=20
+MIN_CACHE_HIT_PCT=60
+MAX_LATENCY_ALLOWED_MS=500
+```
+
+## Output
+
+After the script finishes, you will find the results in a new, timestamped directory created inside `$BASE/auto-benchmark/`.
+
+- **Log Files**: The directory (`$BASE/auto-benchmark/YYYY_MM_DD_HH_MM/`) contains detailed logs for each run:
+    - `vllm_log_...txt`: The log output from the vLLM server for each parameter combination.
+    - `bm_log_...txt`: The log output from the `benchmark_serving.py` script for each benchmark run.
+
+- **Final Result Summary**: A file named `result.txt` is created in the log directory. It contains a summary of each tested combination and concludes with the overall best parameters found.
+
+```
+# Example result.txt content
+hash:a1b2c3d4...
+max_num_seqs: 128, max_num_batched_tokens: 2048, request_rate: 10.0, e2el: 450.5, throughput: 9.8, goodput: 9.8
+max_num_seqs: 128, max_num_batched_tokens: 4096 does not meet latency requirement 500
+...
+best_max_num_seqs: 256, best_num_batched_tokens: 2048, best_throughput: 12.5, profile saved in: /home/user/vllm/auto-benchmark/2024_08_01_10_30/profile
+```
+
+  If it cannot find the best parameters, the final row will be `best_max_num_seqs: 0, best_num_batched_tokens: 0, best_throughput: 0`. This can be due to either the server not starting properly, or the latency requirement being too strict.
+
+- **Profiler Trace**: A directory named `profile` is created inside the log directory. It contains the profiler trace file (e.g., `.xplane.pb` for TPU or a `.json` trace for GPU) from the single best-performing run.
+
+## How It Works
+
+The script follows a systematic process to find the optimal parameters:
+
+1. **Find Max GPU Memory Utilization**: The script first determines the highest safe `gpu-memory-utilization` (starting from 0.98 and decreasing) that does not cause an Out-Of-Memory (OOM) error when launching the server. This ensures the benchmark runs use the maximum available memory without crashing.
+
+2. **Iterate and Benchmark**: It then enters a nested loop, iterating through every combination of `max-num-seqs` and `max-num-batched-tokens` provided in the configuration lists.
+
+3. **Latency-Aware Throughput Search**: For each parameter combination:
+    - The vLLM server is started.
+    - A benchmark is first run with an infinite request rate (`--request-rate inf`).
+    - If the resulting P99 E2E latency is within the `MAX_LATENCY_ALLOWED_MS` limit, this throughput is considered the maximum for this configuration.
+    - If the latency is too high, the script performs a search by iteratively decreasing the request rate until the latency constraint is met. This finds the highest sustainable throughput for the given parameters and latency requirement.
+
+4. **Track Best Result**: Throughout the process, the script tracks the parameter combination that has yielded the highest valid throughput so far.
+
+5. **Profile Collection**: For the best-performing run, the script saves the vLLM profiler output, which can be used for deep-dive performance analysis with tools like TensorBoard.

benchmarks/auto_tune/auto_tune.sh

@@ -0,0 +1,247 @@
+#!/bin/bash
+
+# This script aims to tune the best server parameter combinations to maximize throughput for given requirement. 
+# See details in README (benchmarks/auto_tune/README.md).
+
+TAG=$(date +"%Y_%m_%d_%H_%M")
+BASE=""
+MODEL="meta-llama/Llama-3.1-8B-Instruct"
+SYSTEM="TPU"
+TP=1
+DOWNLOAD_DIR=""
+INPUT_LEN=4000
+OUTPUT_LEN=16
+MIN_CACHE_HIT_PCT=0
+MAX_LATENCY_ALLOWED_MS=100000000000
+NUM_SEQS_LIST="128 256"
+NUM_BATCHED_TOKENS_LIST="512 1024 2048 4096"
+
+LOG_FOLDER="$BASE/auto-benchmark/$TAG"
+RESULT="$LOG_FOLDER/result.txt"
+PROFILE_PATH="$LOG_FOLDER/profile"
+
+echo "result file: $RESULT"
+echo "model: $MODEL"
+
+rm -rf $LOG_FOLDER
+rm -rf $PROFILE_PATH
+mkdir -p $LOG_FOLDER
+mkdir -p $PROFILE_PATH
+
+cd "$BASE/vllm"
+
+pip install -q datasets
+
+current_hash=$(git rev-parse HEAD)
+echo "hash:$current_hash" >> "$RESULT"
+echo "current_hash: $current_hash"
+
+best_throughput=0
+best_max_num_seqs=0
+best_num_batched_tokens=0
+best_goodput=0
+
+start_server() {
+    local gpu_memory_utilization=$1
+    local max_num_seqs=$2
+    local max_num_batched_tokens=$3
+    local vllm_log=$4
+    local profile_dir=$5
+    
+    pkill -f vllm
+
+    VLLM_USE_V1=1 VLLM_SERVER_DEV_MODE=1 VLLM_TORCH_PROFILER_DIR=$profile_dir vllm serve $MODEL \
+        --disable-log-requests \
+        --port 8004 \
+        --gpu-memory-utilization $gpu_memory_utilization \
+        --max-num-seqs $max_num_seqs \
+        --max-num-batched-tokens $max_num_batched_tokens \
+        --tensor-parallel-size $TP \
+        --enable-prefix-caching \
+        --load-format dummy \
+        --download-dir "$DOWNLOAD_DIR" \
+        --max-model-len $(( INPUT_LEN+OUTPUT_LEN )) > "$vllm_log" 2>&1 &
+
+    # wait for 10 minutes...
+    server_started=0
+    for i in {1..60}; do  
+        RESPONSE=$(curl -s -X GET "http://0.0.0.0:8004/health" -w "%{http_code}" -o /dev/stdout)
+        STATUS_CODE=$(echo "$RESPONSE" | tail -n 1) 
+        if [[ "$STATUS_CODE" -eq 200 ]]; then
+            server_started=1
+            break
+        else
+            sleep 10
+        fi
+    done
+    if (( ! server_started )); then
+        echo "server did not start within 10 minutes. Please check server log at $vllm_log".
+        return 1
+    else
+        return 0
+    fi
+}
+
+update_best_profile() {
+    local profile_dir=$1
+    local profile_index=$2
+    sorted_paths=($(find "$profile_dir" -maxdepth 1 -not -path "$profile_dir" | sort))
+    selected_profile_file=
+    if [[ "$SYSTEM" == "TPU" ]]; then
+        selected_profile_file="${sorted_paths[$profile_index]}/*.xplane.pb"
+    fi 
+    if [[ "$SYSTEM" == "GPU" ]]; then
+        selected_profile_file="${sorted_paths[$profile_index]}"
+    fi 
+    rm -f $PROFILE_PATH/*
+    cp $selected_profile_file $PROFILE_PATH
+}
+
+run_benchmark() {
+    local max_num_seqs=$1
+    local max_num_batched_tokens=$2
+    local gpu_memory_utilization=$3
+    echo "max_num_seq: $max_num_seqs, max_num_batched_tokens: $max_num_batched_tokens"
+    local vllm_log="$LOG_FOLDER/vllm_log_${max_num_seqs}_${max_num_batched_tokens}.txt"
+    local profile_dir="$LOG_FOLDER/profile_${max_num_seqs}_${max_num_batched_tokens}"
+    echo "vllm_log: $vllm_log"
+    echo
+    rm -f $vllm_log
+    mkdir -p $profile_dir
+    pkill -f vllm
+    local profile_index=0
+
+    echo "starting server..."
+    start_server $gpu_memory_utilization $max_num_seqs $max_num_batched_tokens $vllm_log $profile_dir
+    result=$?
+    if [[ "$result" -eq 1 ]]; then
+        echo "server failed to start. gpu_memory_utilization:$gpu_memory_utilization, max_num_seqs:$max_num_seqs, max_num_batched_tokens: $max_num_batched_tokens"
+    else
+        echo "server started."
+    fi
+    echo
+    
+    echo "run benchmark test..."
+    meet_latency_requirement=0
+    # get a basic qps by using request-rate inf
+    bm_log="$LOG_FOLDER/bm_log_${max_num_seqs}_${max_num_batched_tokens}_requestrate_inf.txt"
+    prefix_len=$(( INPUT_LEN * MIN_CACHE_HIT_PCT / 100 ))
+    python benchmarks/benchmark_serving.py \
+        --backend vllm \
+        --model $MODEL  \
+        --dataset-name random \
+        --random-input-len $INPUT_LEN \
+        --random-output-len $OUTPUT_LEN \
+        --ignore-eos \
+        --disable-tqdm \
+        --request-rate inf \
+        --percentile-metrics ttft,tpot,itl,e2el \
+        --goodput e2el:$MAX_LATENCY_ALLOWED_MS \
+        --num-prompts 1000 \
+        --random-prefix-len $prefix_len \
+        --port 8004 \
+        --profile &> "$bm_log"
+    throughput=$(grep "Request throughput (req/s):" "$bm_log" | sed 's/[^0-9.]//g')
+    e2el=$(grep "P99 E2EL (ms):" "$bm_log" | awk '{print $NF}')
+    goodput=$(grep "Request goodput (req/s):" "$bm_log" | sed 's/[^0-9.]//g')
+
+    if (( $(echo "$e2el <= $MAX_LATENCY_ALLOWED_MS" | bc -l) )); then
+        meet_latency_requirement=1
+        request_rate=inf
+    fi
+
+    if (( ! meet_latency_requirement )); then
+    # start from request-rate as int(throughput) + 1
+        request_rate=$((${throughput%.*} + 1))
+        while ((request_rate > 0)); do
+            profile_index=$((profile_index+1))
+            # clear prefix cache
+            curl -X POST http://0.0.0.0:8004/reset_prefix_cache
+            sleep 5
+            bm_log="$LOG_FOLDER/bm_log_${max_num_seqs}_${max_num_batched_tokens}_requestrate_${request_rate}.txt"
+            python benchmarks/benchmark_serving.py \
+                --backend vllm \
+                --model $MODEL  \
+                --dataset-name random \
+                --random-input-len $INPUT_LEN \
+                --random-output-len $OUTPUT_LEN \
+                --ignore-eos \
+                --disable-tqdm \
+                --request-rate $request_rate \
+                --percentile-metrics ttft,tpot,itl,e2el \
+                --goodput e2el:$MAX_LATENCY_ALLOWED_MS \
+                --num-prompts 100 \
+                --random-prefix-len $prefix_len \
+                --port 8004 &> "$bm_log"
+            throughput=$(grep "Request throughput (req/s):" "$bm_log" | sed 's/[^0-9.]//g')
+            e2el=$(grep "P99 E2EL (ms):" "$bm_log" | awk '{print $NF}')
+            goodput=$(grep "Request goodput (req/s):" "$bm_log" | sed 's/[^0-9.]//g')
+            if (( $(echo "$e2el <= $MAX_LATENCY_ALLOWED_MS" | bc -l) )); then
+                meet_latency_requirement=1
+                break
+            fi
+            request_rate=$((request_rate-1))
+        done
+    fi
+    # write the results and update the best result.
+    if ((meet_latency_requirement)); then
+        echo "max_num_seqs: $max_num_seqs, max_num_batched_tokens: $max_num_batched_tokens, request_rate: $request_rate, e2el: $e2el, throughput: $throughput, goodput: $goodput"
+        echo "max_num_seqs: $max_num_seqs, max_num_batched_tokens: $max_num_batched_tokens, request_rate: $request_rate, e2el: $e2el, throughput: $throughput, goodput: $goodput" >> "$RESULT"
+        if (( $(echo "$throughput > $best_throughput" | bc -l) )); then
+            best_throughput=$throughput
+            best_max_num_seqs=$max_num_seqs
+            best_num_batched_tokens=$max_num_batched_tokens
+            best_goodput=$goodput
+            if [[ "$SYSTEM" == "TPU" ]]; then
+                update_best_profile "$profile_dir/plugins/profile" $profile_index
+            fi
+            if [[ "$SYSTEM" == "GPU" ]]; then
+                update_best_profile "$profile_dir" $profile_index
+            fi
+        fi
+    else
+        echo "max_num_seqs: $max_num_seqs, max_num_batched_tokens: $max_num_batched_tokens does not meet latency requirement ${MAX_LATENCY_ALLOWED_MS}"
+        echo "max_num_seqs: $max_num_seqs, max_num_batched_tokens: $max_num_batched_tokens does not meet latency requirement ${MAX_LATENCY_ALLOWED_MS}" >> "$RESULT"
+    fi
+
+    echo "best_max_num_seqs: $best_max_num_seqs, best_num_batched_tokens: $best_num_batched_tokens, best_throughput: $best_throughput"
+
+    pkill vllm
+    sleep 10
+    printf '=%.0s' $(seq 1 20)
+    return 0
+}
+
+read -r -a num_seqs_list <<< "$NUM_SEQS_LIST"
+read -r -a num_batched_tokens_list <<< "$NUM_BATCHED_TOKENS_LIST"
+
+# first find out the max gpu-memory-utilization without HBM OOM.
+gpu_memory_utilization=0.98
+find_gpu_memory_utilization=0
+while (( $(echo "$gpu_memory_utilization >= 0.9" | bc -l) )); do
+    start_server $gpu_memory_utilization "${num_seqs_list[-1]}" "${num_batched_tokens_list[-1]}" "$LOG_FOLDER/vllm_log_gpu_memory_utilization_$gpu_memory_utilization.log"
+    result=$?
+    if [[ "$result" -eq 0 ]]; then
+        find_gpu_memory_utilization=1
+        break
+    else
+        gpu_memory_utilization=$(echo "$gpu_memory_utilization - 0.01" | bc)
+    fi
+done
+
+if [[ "$find_gpu_memory_utilization" -eq 1 ]]; then
+    echo "Using gpu_memory_utilization=$gpu_memory_utilization to serve model."
+else
+    echo "Cannot find a proper gpu_memory_utilization over 0.9 to serve the model, please check logs in $LOG_FOLDER."
+    exit 1
+fi
+
+for num_seqs in "${num_seqs_list[@]}"; do
+    for num_batched_tokens in "${num_batched_tokens_list[@]}"; do
+        run_benchmark $num_seqs $num_batched_tokens $gpu_memory_utilization
+    done
+done
+echo "finish permutations"
+echo "best_max_num_seqs: $best_max_num_seqs, best_num_batched_tokens: $best_num_batched_tokens, best_throughput: $best_throughput, profile saved in: $PROFILE_PATH"
+echo "best_max_num_seqs: $best_max_num_seqs, best_num_batched_tokens: $best_num_batched_tokens, best_throughput: $best_throughput, profile saved in: $PROFILE_PATH" >> "$RESULT"
+

benchmarks/backend_request_func.py

@@ -1,4 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 import io
 import json
@@ -403,8 +404,14 @@ async def async_request_openai_chat_completions(
                         chunk_bytes = chunk_bytes.strip()
                         if not chunk_bytes:
                             continue
+                        chunk_bytes = chunk_bytes.decode("utf-8")
+                        # NOTE: SSE comments (often used as pings) start with a colon.
+                        # These are not JSON data payload and should be skipped.
+                        if chunk_bytes.startswith(":"):
+                            continue
+
+                        chunk = chunk_bytes.removeprefix("data: ")
 
-                        chunk = chunk_bytes.decode("utf-8").removeprefix("data: ")
                         if chunk != "[DONE]":
                             timestamp = time.perf_counter()
                             data = json.loads(chunk)

benchmarks/benchmark_dataset.py

@@ -1,4 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 """
 This module defines a framework for sampling benchmark requests from various
 datasets. Each dataset subclass of BenchmarkDataset must implement sample
@@ -323,6 +324,9 @@ class RandomDataset(BenchmarkDataset):
         input_low = int(real_input_len * (1 - range_ratio))
         input_high = int(real_input_len * (1 + range_ratio))
         output_low = int(output_len * (1 - range_ratio))
+        # Ensure the lower bound for output length is at least 1 to prevent
+        # sampling 0 tokens, which can cause request failures.
+        output_low = max(output_low, 1)
         output_high = int(output_len * (1 + range_ratio))
 
         # Add logging for debugging
@@ -348,11 +352,12 @@ class RandomDataset(BenchmarkDataset):
             # [1650, 939, 486] -> ['Ġcall', 'sh', 'ere']
             # To avoid uncontrolled change of the prompt length,
             # the encoded sequence is truncated before being decode again.
+            total_input_len = prefix_len + int(input_lens[i])
             re_encoded_sequence = tokenizer.encode(prompt, add_special_tokens=False)[
-                : input_lens[i]
+                :total_input_len
             ]
             prompt = tokenizer.decode(re_encoded_sequence)
-            total_input_len = prefix_len + int(input_lens[i])
+            total_input_len = len(re_encoded_sequence)
             requests.append(
                 SampleRequest(
                     prompt=prompt,
@@ -699,6 +704,7 @@ class HuggingFaceDataset(BenchmarkDataset):
         self,
         dataset_path: str,
         dataset_split: str,
+        no_stream: bool = False,
         dataset_subset: Optional[str] = None,
         **kwargs,
     ) -> None:
@@ -706,6 +712,7 @@ class HuggingFaceDataset(BenchmarkDataset):
 
         self.dataset_split = dataset_split
         self.dataset_subset = dataset_subset
+        self.load_stream = not no_stream
         self.load_data()
 
     def load_data(self) -> None:
@@ -714,7 +721,7 @@ class HuggingFaceDataset(BenchmarkDataset):
             self.dataset_path,
             name=self.dataset_subset,
             split=self.dataset_split,
-            streaming=True,
+            streaming=self.load_stream,
         )
         self.data = self.data.shuffle(seed=self.random_seed)
 
@@ -864,7 +871,15 @@ class InstructCoderDataset(HuggingFaceDataset):
         for item in self.data:
             if len(sampled_requests) >= num_requests:
                 break
-            prompt = f"{item['instruction']}:\n{item['input']}"
+            prompt = f"{item['input']}\n\n{item['instruction']} Just output \
+            the code, do not include any explanation."
+
+            # apply template
+            prompt = tokenizer.apply_chat_template(
+                [{"role": "user", "content": prompt}],
+                add_generation_prompt=True,
+                tokenize=False,
+            )
             prompt_len = len(tokenizer(prompt).input_ids)
             sampled_requests.append(
                 SampleRequest(

benchmarks/benchmark_latency.py

@@ -1,4 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 """Benchmark the latency of processing a single batch of requests."""
 
 import argparse
@@ -122,7 +123,7 @@ def main(args: argparse.Namespace):
         save_to_pytorch_benchmark_format(args, results)
 
 
-if __name__ == "__main__":
+def create_argument_parser():
     parser = FlexibleArgumentParser(
         description="Benchmark the latency of processing a single batch of "
         "requests till completion."
@@ -170,6 +171,12 @@ if __name__ == "__main__":
     # V1 enables prefix caching by default which skews the latency
     # numbers. We need to disable prefix caching by default.
     parser.set_defaults(enable_prefix_caching=False)
+
+    return parser
+
+
+if __name__ == "__main__":
+    parser = create_argument_parser()
     args = parser.parse_args()
     if args.profile and not envs.VLLM_TORCH_PROFILER_DIR:
         raise OSError(

benchmarks/benchmark_long_document_qa_throughput.py

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

benchmarks/benchmark_prefix_caching.py

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

benchmarks/benchmark_prioritization.py

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

benchmarks/benchmark_serving.py

@@ -1,4 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 r"""Benchmark online serving throughput.
 
 On the server side, run one of the following commands:
@@ -32,7 +33,7 @@ import warnings
 from collections.abc import AsyncGenerator, Iterable
 from dataclasses import dataclass
 from datetime import datetime
-from typing import Any, Optional
+from typing import Any, Literal, Optional
 
 import numpy as np
 from tqdm.asyncio import tqdm
@@ -106,14 +107,42 @@ class BenchmarkMetrics:
     percentiles_e2el_ms: list[tuple[float, float]]
 
 
+def _get_current_request_rate(
+    ramp_up_strategy: Optional[Literal["linear", "exponential"]],
+    ramp_up_start_rps: Optional[int],
+    ramp_up_end_rps: Optional[int],
+    request_index: int,
+    total_requests: int,
+    request_rate: float,
+) -> float:
+    if (
+        ramp_up_strategy
+        and ramp_up_start_rps is not None
+        and ramp_up_end_rps is not None
+    ):
+        progress = request_index / max(total_requests - 1, 1)
+        if ramp_up_strategy == "linear":
+            increase = (ramp_up_end_rps - ramp_up_start_rps) * progress
+            return ramp_up_start_rps + increase
+        elif ramp_up_strategy == "exponential":
+            ratio = ramp_up_end_rps / ramp_up_start_rps
+            return ramp_up_start_rps * (ratio**progress)
+        else:
+            raise ValueError(f"Unknown ramp-up strategy: {ramp_up_strategy}")
+    return request_rate
+
+
 async def get_request(
     input_requests: list[SampleRequest],
     request_rate: float,
     burstiness: float = 1.0,
-) -> AsyncGenerator[SampleRequest, None]:
+    ramp_up_strategy: Optional[Literal["linear", "exponential"]] = None,
+    ramp_up_start_rps: Optional[int] = None,
+    ramp_up_end_rps: Optional[int] = None,
+) -> AsyncGenerator[tuple[SampleRequest, float], None]:
     """
     Asynchronously generates requests at a specified rate
-    with OPTIONAL burstiness.
+    with OPTIONAL burstiness and OPTIONAL ramp-up strategy.
 
     Args:
         input_requests:
@@ -128,22 +157,44 @@ async def get_request(
             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.
+         ramp_up_strategy (optional):
+            The ramp-up strategy. Can be "linear" or "exponential".
+            If None, uses constant request rate (specified by request_rate).
+        ramp_up_start_rps (optional):
+            The starting request rate for ramp-up.
+        ramp_up_end_rps (optional):
+            The ending request rate for ramp-up.
     """
-    input_requests: Iterable[SampleRequest] = 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)
+    # Convert to list to get length for ramp-up calculations
+    if isinstance(input_requests, Iterable) and not isinstance(input_requests, list):
+        input_requests = list(input_requests)
+
+    total_requests = len(input_requests)
+    request_index = 0
 
     for request in input_requests:
-        yield request
+        current_request_rate = _get_current_request_rate(
+            ramp_up_strategy,
+            ramp_up_start_rps,
+            ramp_up_end_rps,
+            request_index,
+            total_requests,
+            request_rate,
+        )
 
-        if request_rate == float("inf"):
+        yield request, current_request_rate
+
+        request_index += 1
+
+        if current_request_rate == float("inf"):
             # If the request rate is infinity, then we don't need to wait.
             continue
 
+        theta = 1.0 / (current_request_rate * burstiness)
+
         # Sample the request interval from the gamma distribution.
         # If burstiness is 1, it follows exponential distribution.
         interval = np.random.gamma(shape=burstiness, scale=theta)
@@ -289,6 +340,9 @@ async def benchmark(
     max_concurrency: Optional[int],
     lora_modules: Optional[Iterable[str]],
     extra_body: Optional[dict],
+    ramp_up_strategy: Optional[Literal["linear", "exponential"]] = None,
+    ramp_up_start_rps: Optional[int] = None,
+    ramp_up_end_rps: Optional[int] = None,
 ):
     if backend in ASYNC_REQUEST_FUNCS:
         request_func = ASYNC_REQUEST_FUNCS[backend]
@@ -352,7 +406,15 @@ async def benchmark(
 
     distribution = "Poisson process" if burstiness == 1.0 else "Gamma distribution"
 
-    print(f"Traffic request rate: {request_rate}")
+    if ramp_up_strategy is not None:
+        print(
+            f"Traffic ramp-up strategy: {ramp_up_strategy}. Will increase "
+            f"RPS from {ramp_up_start_rps} to {ramp_up_end_rps} RPS over "
+            "the duration of the benchmark."
+        )
+    else:
+        print(f"Traffic request rate: {request_rate} RPS.")
+
     print(f"Burstiness factor: {burstiness} ({distribution})")
     print(f"Maximum request concurrency: {max_concurrency}")
 
@@ -372,7 +434,34 @@ async def benchmark(
 
     benchmark_start_time = time.perf_counter()
     tasks: list[asyncio.Task] = []
-    async for request in get_request(input_requests, request_rate, burstiness):
+
+    rps_change_events = []
+    last_int_rps = -1
+    if ramp_up_strategy is not None and ramp_up_start_rps is not None:
+        last_int_rps = ramp_up_start_rps
+        rps_change_events.append(
+            {
+                "rps": last_int_rps,
+                "timestamp": datetime.now().isoformat(),
+            }
+        )
+
+    async for request, current_request_rate in get_request(
+        input_requests,
+        request_rate,
+        burstiness,
+        ramp_up_strategy,
+        ramp_up_start_rps,
+        ramp_up_end_rps,
+    ):
+        if ramp_up_strategy is not None:
+            current_int_rps = int(current_request_rate)
+            if current_int_rps > last_int_rps:
+                timestamp = datetime.now().isoformat()
+                for rps_val in range(last_int_rps + 1, current_int_rps + 1):
+                    rps_change_events.append({"rps": rps_val, "timestamp": timestamp})
+                last_int_rps = current_int_rps
+
         prompt, prompt_len, output_len, mm_content = (
             request.prompt,
             request.prompt_len,
@@ -396,11 +485,8 @@ async def benchmark(
             ignore_eos=ignore_eos,
             extra_body=extra_body,
         )
-        tasks.append(
-            asyncio.create_task(
-                limited_request_func(request_func_input=request_func_input, pbar=pbar)
-            )
-        )
+        task = limited_request_func(request_func_input=request_func_input, pbar=pbar)
+        tasks.append(asyncio.create_task(task))
     outputs: list[RequestFuncOutput] = await asyncio.gather(*tasks)
 
     if profile:
@@ -465,7 +551,7 @@ 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,
+        "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],
@@ -476,6 +562,9 @@ async def benchmark(
         "errors": [output.error for output in outputs],
     }
 
+    if rps_change_events:
+        result["rps_change_events"] = rps_change_events
+
     def process_one_metric(
         # E.g., "ttft"
         metric_attribute_name: str,
@@ -609,6 +698,26 @@ def main(args: argparse.Namespace):
     tokenizer_id = args.tokenizer if args.tokenizer is not None else args.model
     tokenizer_mode = args.tokenizer_mode
 
+    # Validate ramp-up arguments
+    if args.ramp_up_strategy is not None:
+        if args.request_rate != float("inf"):
+            raise ValueError(
+                "When using ramp-up, do not specify --request-rate. "
+                "The request rate will be controlled by ramp-up parameters. "
+                "Please remove the --request-rate argument."
+            )
+        if args.ramp_up_start_rps is None or args.ramp_up_end_rps is None:
+            raise ValueError(
+                "When using --ramp-up-strategy, both --ramp-up-start-rps and "
+                "--ramp-up-end-rps must be specified"
+            )
+        if args.ramp_up_start_rps < 0 or args.ramp_up_end_rps < 0:
+            raise ValueError("Ramp-up start and end RPS must be non-negative")
+        if args.ramp_up_start_rps > args.ramp_up_end_rps:
+            raise ValueError("Ramp-up start RPS must be less than end RPS")
+        if args.ramp_up_strategy == "exponential" and args.ramp_up_start_rps == 0:
+            raise ValueError("For exponential ramp-up, the start RPS cannot be 0.")
+
     if args.base_url is not None:
         api_url = f"{args.base_url}{args.endpoint}"
         base_url = f"{args.base_url}"
@@ -716,6 +825,7 @@ def main(args: argparse.Namespace):
             dataset_subset=args.hf_subset,
             dataset_split=args.hf_split,
             random_seed=args.seed,
+            no_stream=args.no_stream,
         ).sample(
             num_requests=args.num_prompts,
             tokenizer=tokenizer,
@@ -801,6 +911,9 @@ def main(args: argparse.Namespace):
             max_concurrency=args.max_concurrency,
             lora_modules=args.lora_modules,
             extra_body=sampling_params,
+            ramp_up_strategy=args.ramp_up_strategy,
+            ramp_up_start_rps=args.ramp_up_start_rps,
+            ramp_up_end_rps=args.ramp_up_end_rps,
         )
     )
 
@@ -833,6 +946,11 @@ def main(args: argparse.Namespace):
         result_json["burstiness"] = args.burstiness
         result_json["max_concurrency"] = args.max_concurrency
 
+        if args.ramp_up_strategy is not None:
+            result_json["ramp_up_strategy"] = args.ramp_up_strategy
+            result_json["ramp_up_start_rps"] = args.ramp_up_start_rps
+            result_json["ramp_up_end_rps"] = args.ramp_up_end_rps
+
         # Merge with benchmark result
         result_json = {**result_json, **benchmark_result}
 
@@ -858,7 +976,10 @@ def main(args: argparse.Namespace):
             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.ramp_up_strategy is not None:
+            file_name = f"{backend}-ramp-up-{args.ramp_up_strategy}-{args.ramp_up_start_rps}qps-{args.ramp_up_end_rps}qps{max_concurrency_str}-{base_model_id}-{current_dt}.json"  # noqa
+        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:
@@ -874,7 +995,7 @@ def main(args: argparse.Namespace):
         save_to_pytorch_benchmark_format(args, result_json, file_name)
 
 
-if __name__ == "__main__":
+def create_argument_parser():
     parser = FlexibleArgumentParser(
         description="Benchmark the online serving throughput."
     )
@@ -913,6 +1034,11 @@ if __name__ == "__main__":
         help="Path to the sharegpt/sonnet dataset. "
         "Or the huggingface dataset ID if using HF dataset.",
     )
+    parser.add_argument(
+        "--no-stream",
+        action="store_true",
+        help="Do not load the dataset in streaming mode.",
+    )
     parser.add_argument(
         "--max-concurrency",
         type=int,
@@ -1224,6 +1350,35 @@ if __name__ == "__main__":
         "script chooses a LoRA module at random.",
     )
 
-    args = parser.parse_args()
+    parser.add_argument(
+        "--ramp-up-strategy",
+        type=str,
+        default=None,
+        choices=["linear", "exponential"],
+        help="The ramp-up strategy. This would be used to "
+        "ramp up the request rate from initial RPS to final "
+        "RPS rate (specified by --ramp-up-start-rps and --ramp-up-end-rps). "
+        "over the duration of the benchmark.",
+    )
+    parser.add_argument(
+        "--ramp-up-start-rps",
+        type=int,
+        default=None,
+        help="The starting request rate for ramp-up (RPS). "
+        "Needs to be specified when --ramp-up-strategy is used.",
+    )
+    parser.add_argument(
+        "--ramp-up-end-rps",
+        type=int,
+        default=None,
+        help="The ending request rate for ramp-up (RPS). "
+        "Needs to be specified when --ramp-up-strategy is used.",
+    )
 
+    return parser
+
+
+if __name__ == "__main__":
+    parser = create_argument_parser()
+    args = parser.parse_args()
     main(args)

benchmarks/benchmark_serving_structured_output.py

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

benchmarks/benchmark_throughput.py

@@ -1,4 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 """Benchmark offline inference throughput."""
 
 import argparse
@@ -96,7 +97,7 @@ def run_vllm(
         assert lora_requests is None, "BeamSearch API does not support LoRA"
         prompts = [request.prompt for request in requests]
         # output_len should be the same for all requests.
-        output_len = requests[0][2]
+        output_len = requests[0].expected_output_len
         for request in requests:
             assert request.expected_output_len == output_len
         start = time.perf_counter()
@@ -355,6 +356,7 @@ def get_requests(args, tokenizer):
     elif args.dataset_name == "burstgpt":
         dataset_cls = BurstGPTDataset
     elif args.dataset_name == "hf":
+        common_kwargs["no_stream"] = args.no_stream
         if args.dataset_path in VisionArenaDataset.SUPPORTED_DATASET_PATHS:
             dataset_cls = VisionArenaDataset
             common_kwargs["dataset_subset"] = None
@@ -594,7 +596,7 @@ def validate_args(args):
         )
 
 
-if __name__ == "__main__":
+def create_argument_parser():
     parser = FlexibleArgumentParser(description="Benchmark the throughput.")
     parser.add_argument(
         "--backend",
@@ -609,6 +611,11 @@ if __name__ == "__main__":
         help="Name of the dataset to benchmark on.",
         default="sharegpt",
     )
+    parser.add_argument(
+        "--no-stream",
+        action="store_true",
+        help="Do not load the dataset in streaming mode.",
+    )
     parser.add_argument(
         "--dataset",
         type=str,
@@ -716,6 +723,12 @@ if __name__ == "__main__":
     )
 
     parser = AsyncEngineArgs.add_cli_args(parser)
+
+    return parser
+
+
+if __name__ == "__main__":
+    parser = create_argument_parser()
     args = parser.parse_args()
     if args.tokenizer is None:
         args.tokenizer = args.model

benchmarks/benchmark_utils.py

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

benchmarks/cutlass_benchmarks/sparse_benchmarks.py

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

benchmarks/cutlass_benchmarks/utils.py

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

benchmarks/cutlass_benchmarks/w8a8_benchmarks.py

@@ -1,4 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 import argparse
 import copy
@@ -18,7 +19,7 @@ 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
+from vllm.utils import FlexibleArgumentParser, cdiv
 
 DEFAULT_MODELS = list(WEIGHT_SHAPES.keys())
 DEFAULT_BATCH_SIZES = [1, 16, 32, 64, 128, 256, 512]
@@ -116,14 +117,9 @@ def bench_fp8(
     scale_a = torch.tensor(1.0, device="cuda", dtype=torch.float32)
     scale_b = torch.tensor(1.0, device="cuda", dtype=torch.float32)
 
-    def ceil_div(x: int, y: int) -> int:
-        return (x + y - 1) // y
-
-    block_scale_a = torch.rand(
-        (m, ceil_div(k, 128)), device="cuda", dtype=torch.float32
-    )
+    block_scale_a = torch.rand((m, cdiv(k, 128)), device="cuda", dtype=torch.float32)
     block_scale_b = torch.rand(
-        ceil_div(k, 128), ceil_div(n, 128), device="cuda", dtype=torch.float32
+        cdiv(k, 128), cdiv(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()

benchmarks/cutlass_benchmarks/weight_shapes.py

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

benchmarks/disagg_benchmarks/disagg_prefill_proxy_server.py

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

benchmarks/disagg_benchmarks/round_robin_proxy.py

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

benchmarks/disagg_benchmarks/visualize_benchmark_results.py

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

benchmarks/fused_kernels/layernorm_rms_benchmarks.py

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

benchmarks/kernels/bench_fp8_gemm.py

@@ -1,14 +1,89 @@
 # SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 import argparse
 import copy
 import itertools
 
 import torch
-import triton
 from weight_shapes import WEIGHT_SHAPES
 
 from vllm._custom_ops import cutlass_scaled_mm as vllm_scaled_mm
 from vllm._custom_ops import scaled_fp8_quant as vllm_scaled_fp8_quant
+from vllm.triton_utils import triton
+
+PROVIDER_CFGS = {
+    "torch-bf16": dict(enabled=True),
+    "fp8-tensor-w-token-a": dict(
+        w="tensor", a="token", no_a_quant=False, enabled=False
+    ),
+    "fp8-tensor-w-tensor-a": dict(
+        w="tensor", a="tensor", no_a_quant=False, enabled=True
+    ),
+    "fp8-channel-w-token-a": dict(
+        w="channel", a="token", no_a_quant=False, enabled=True
+    ),
+    "fp8-channel-w-tensor-a": dict(
+        w="channel", a="tensor", no_a_quant=False, enabled=False
+    ),
+    "fp8-tensor-w-token-a-noquant": dict(
+        w="tensor", a="token", no_a_quant=True, enabled=False
+    ),
+    "fp8-tensor-w-tensor-a-noquant": dict(
+        w="tensor", a="tensor", no_a_quant=True, enabled=True
+    ),
+    "fp8-channel-w-token-a-noquant": dict(
+        w="channel", a="token", no_a_quant=True, enabled=True
+    ),
+    "fp8-channel-w-tensor-a-noquant": dict(
+        w="channel", a="tensor", no_a_quant=True, enabled=False
+    ),
+}
+
+_enabled = [k for k, v in PROVIDER_CFGS.items() if v["enabled"]]
+
+
+def _quant_weight_fp8(b: torch.Tensor, w_type: str, device: str):
+    if w_type == "tensor":
+        scale_b = torch.ones(1, device=device, dtype=torch.float32)
+        b_fp8, scale_b_fp8 = vllm_scaled_fp8_quant(b, scale_b)
+    else:
+        b_fp8, scale_b_fp8 = vllm_scaled_fp8_quant(b, use_per_token_if_dynamic=True)
+    return b_fp8.t(), scale_b_fp8
+
+
+def build_fp8_runner(cfg, a, b, dtype, device):
+    b_fp8, scale_b_fp8 = _quant_weight_fp8(b, cfg["w"], device)
+
+    scale_a_const = (
+        torch.ones(1, device=device, dtype=torch.float32)
+        if cfg["a"] == "tensor"
+        else None
+    )
+
+    if cfg["no_a_quant"]:
+        if cfg["a"] == "tensor":
+            a_fp8, scale_a_fp8 = vllm_scaled_fp8_quant(a, scale_a_const)
+        else:
+            a_fp8, scale_a_fp8 = vllm_scaled_fp8_quant(a, use_per_token_if_dynamic=True)
+
+        def run():
+            return vllm_scaled_mm(a_fp8, b_fp8, scale_a_fp8, scale_b_fp8, dtype)
+
+        return run
+
+    if cfg["a"] == "tensor":
+
+        def run():
+            a_fp8, scale_a_fp8 = vllm_scaled_fp8_quant(a, scale_a_const)
+            return vllm_scaled_mm(a_fp8, b_fp8, scale_a_fp8, scale_b_fp8, dtype)
+
+    else:
+
+        def run():
+            a_fp8, scale_a_fp8 = vllm_scaled_fp8_quant(a, use_per_token_if_dynamic=True)
+            return vllm_scaled_mm(a_fp8, b_fp8, scale_a_fp8, scale_b_fp8, dtype)
+
+    return run
 
 
 @triton.testing.perf_report(
@@ -17,28 +92,8 @@ from vllm._custom_ops import scaled_fp8_quant as vllm_scaled_fp8_quant
         x_vals=[1, 16, 64, 128, 256, 512, 1024, 2048, 4096, 8192, 16384],
         x_log=False,
         line_arg="provider",
-        line_vals=[
-            "torch-bf16",
-            # "fp8-tensor-w-token-a",
-            "fp8-tensor-w-tensor-a",
-            "fp8-channel-w-token-a",
-            # "fp8-channel-w-tensor-a",
-            # "fp8-tensor-w-token-a-noquant",
-            "fp8-tensor-w-tensor-a-noquant",
-            "fp8-channel-w-token-a-noquant",
-            # "fp8-channel-w-tensor-a-noquant",
-        ],
-        line_names=[
-            "torch-bf16",
-            # "fp8-tensor-w-token-a",
-            "fp8-tensor-w-tensor-a",
-            "fp8-channel-w-token-a",
-            # "fp8-channel-w-tensor-a",
-            # "fp8-tensor-w-token-a-noquant",
-            "fp8-tensor-w-tensor-a-noquant",
-            "fp8-channel-w-token-a-noquant",
-            # "fp8-channel-w-tensor-a-noquant",
-        ],
+        line_vals=_enabled,
+        line_names=_enabled,
         ylabel="TFLOP/s (larger is better)",
         plot_name="BF16 vs FP8 GEMMs",
         args={},
@@ -49,144 +104,34 @@ def benchmark(batch_size, provider, N, K):
     device = "cuda"
     dtype = torch.bfloat16
 
-    # Create input tensors
     a = torch.randn((M, K), device=device, dtype=dtype)
     b = torch.randn((N, K), device=device, dtype=dtype)
 
     quantiles = [0.5, 0.2, 0.8]
 
-    if "torch-bf16" in provider:
+    if provider == "torch-bf16":
         ms, min_ms, max_ms = triton.testing.do_bench_cudagraph(
             lambda: torch.nn.functional.linear(a, b), quantiles=quantiles
         )
-
-    elif "fp8" in provider:
-        # Weights are always quantized ahead of time
-        if "noquant" in provider:
-            # For no quantization, we just measure the GEMM
-            if "tensor-w-token-a" in provider:
-                # Dynamic per-token quant for A, per-tensor quant for B
-                b_fp8, scale_b_fp8 = vllm_scaled_fp8_quant(b)
-                assert scale_b_fp8.numel() == 1
-                a_fp8, scale_a_fp8 = vllm_scaled_fp8_quant(
-                    a, use_per_token_if_dynamic=True
-                )
-
-                def run_quant():
-                    return vllm_scaled_mm(a_fp8, b_fp8, scale_a_fp8, scale_b_fp8, dtype)
-
-            elif "tensor-w-tensor-a" in provider:
-                # Static per-tensor quantization with fixed scales
-                # for both A and B
-                scale_a = torch.tensor([1.0], device=device, dtype=torch.float32)
-                scale_b = torch.tensor([1.0], device=device, dtype=torch.float32)
-                b_fp8, scale_b_fp8 = vllm_scaled_fp8_quant(b, scale_b)
-                assert scale_b_fp8.numel() == 1
-                a_fp8, scale_a_fp8 = vllm_scaled_fp8_quant(a, scale_a)
-
-                def run_quant():
-                    return vllm_scaled_mm(a_fp8, b_fp8, scale_a_fp8, scale_b_fp8, dtype)
-
-            elif "channel-w-token-a" in provider:
-                # Static per-channel quantization for weights, per-token
-                # quant for A
-                scale_b = torch.tensor((N,), device=device, dtype=torch.float32)
-                b_fp8, scale_b_fp8 = vllm_scaled_fp8_quant(b, scale_b)
-                scale_b_fp8 = scale_b_fp8.expand(N).contiguous()
-                assert scale_b_fp8.numel() == N
-                a_fp8, scale_a_fp8 = vllm_scaled_fp8_quant(
-                    a, use_per_token_if_dynamic=True
-                )
-
-                def run_quant():
-                    return vllm_scaled_mm(a_fp8, b_fp8, scale_a_fp8, scale_b_fp8, dtype)
-
-            elif "channel-w-tensor-a" in provider:
-                # Static per-channel quantization for weights, per-tensor
-                # quant for A
-                scale_a = torch.tensor([1.0], device=device, dtype=torch.float32)
-                scale_b = torch.tensor((N,), device=device, dtype=torch.float32)
-                b_fp8, scale_b_fp8 = vllm_scaled_fp8_quant(b, scale_b)
-                scale_b_fp8 = scale_b_fp8.expand(N).contiguous()
-                assert scale_b_fp8.numel() == N
-                a_fp8, scale_a_fp8 = vllm_scaled_fp8_quant(a, scale_a)
-
-                def run_quant():
-                    return vllm_scaled_mm(a_fp8, b_fp8, scale_a_fp8, scale_b_fp8, dtype)
-
-        else:
-            # In these cases, we quantize the activations during the GEMM call
-            if "tensor-w-token-a" in provider:
-                # Dynamic per-token quant for A, per-tensor quant for B
-                b_fp8, scale_b_fp8 = vllm_scaled_fp8_quant(b)
-                assert scale_b_fp8.numel() == 1
-
-                def run_quant():
-                    a_fp8, scale_a_fp8 = vllm_scaled_fp8_quant(
-                        a, use_per_token_if_dynamic=True
-                    )
-                    return vllm_scaled_mm(a_fp8, b_fp8, scale_a_fp8, scale_b_fp8, dtype)
-
-            elif "tensor-w-tensor-a" in provider:
-                # Static per-tensor quantization with fixed scales
-                # for both A and B
-                scale_a = torch.tensor([1.0], device=device, dtype=torch.float32)
-                scale_b = torch.tensor([1.0], device=device, dtype=torch.float32)
-                b_fp8, scale_b_fp8 = vllm_scaled_fp8_quant(b, scale_b)
-                assert scale_b_fp8.numel() == 1
-
-                def run_quant():
-                    a_fp8, scale_a_fp8 = vllm_scaled_fp8_quant(a, scale_a)
-                    return vllm_scaled_mm(a_fp8, b_fp8, scale_a_fp8, scale_b_fp8, dtype)
-
-            elif "channel-w-token-a" in provider:
-                # Static per-channel quantization for weights, per-token
-                # quant for A
-                scale_b = torch.tensor((N,), device=device, dtype=torch.float32)
-                b_fp8, scale_b_fp8 = vllm_scaled_fp8_quant(b, scale_b)
-                scale_b_fp8 = scale_b_fp8.expand(N).contiguous()
-                assert scale_b_fp8.numel() == N
-
-                def run_quant():
-                    a_fp8, scale_a_fp8 = vllm_scaled_fp8_quant(
-                        a, use_per_token_if_dynamic=True
-                    )
-                    return vllm_scaled_mm(a_fp8, b_fp8, scale_a_fp8, scale_b_fp8, dtype)
-
-            elif "channel-w-tensor-a" in provider:
-                # Static per-channel quantization for weights, per-tensor
-                # quant for A
-                scale_a = torch.tensor([1.0], device=device, dtype=torch.float32)
-                scale_b = torch.tensor((N,), device=device, dtype=torch.float32)
-                b_fp8, scale_b_fp8 = vllm_scaled_fp8_quant(b, scale_b)
-                scale_b_fp8 = scale_b_fp8.expand(N).contiguous()
-                assert scale_b_fp8.numel() == N
-
-                def run_quant():
-                    a_fp8, scale_a_fp8 = vllm_scaled_fp8_quant(a, scale_a)
-                    return vllm_scaled_mm(a_fp8, b_fp8, scale_a_fp8, scale_b_fp8, dtype)
-
-        b_fp8 = b_fp8.t()
-
+    else:
+        cfg = PROVIDER_CFGS[provider]
+        run_quant = build_fp8_runner(cfg, a, b, dtype, device)
         ms, min_ms, max_ms = triton.testing.do_bench_cudagraph(
             lambda: run_quant(), quantiles=quantiles
         )
 
-    # Calculate TFLOP/s, two flops per multiply-add
-    tflops = lambda ms: (2 * M * N * K) * 1e-12 / (ms * 1e-3)
-    return tflops(ms), tflops(max_ms), tflops(min_ms)
+    to_tflops = lambda t_ms: (2 * M * N * K) * 1e-12 / (t_ms * 1e-3)
+    return to_tflops(ms), to_tflops(max_ms), to_tflops(min_ms)
 
 
 def prepare_shapes(args):
-    KN_model_names = []
-    models_tps = list(itertools.product(args.models, args.tp_sizes))
-    for model, tp_size in models_tps:
-        assert model in WEIGHT_SHAPES
-        for KN, tp_split_dim in copy.deepcopy(WEIGHT_SHAPES[model]):
-            KN[tp_split_dim] = KN[tp_split_dim] // tp_size
+    out = []
+    for model, tp_size in itertools.product(args.models, args.tp_sizes):
+        for KN, tp_dim in copy.deepcopy(WEIGHT_SHAPES[model]):
+            KN[tp_dim] //= tp_size
             KN.append(model)
-            KN_model_names.append(KN)
-    return KN_model_names
+            out.append(KN)
+    return out
 
 
 if __name__ == "__main__":
@@ -196,21 +141,13 @@ if __name__ == "__main__":
         nargs="+",
         type=str,
         default=["meta-llama/Llama-3.1-8B-Instruct"],
-        choices=[*WEIGHT_SHAPES.keys()],
-        help="List of models to benchmark",
-    )
-    parser.add_argument(
-        "--tp-sizes",
-        nargs="+",
-        type=int,
-        default=[1],
-        help="List of tensor parallel sizes",
+        choices=list(WEIGHT_SHAPES.keys()),
     )
+    parser.add_argument("--tp-sizes", nargs="+", type=int, default=[1])
     args = parser.parse_args()
 
-    KN_model_names = prepare_shapes(args)
-    for K, N, model_name in KN_model_names:
-        print(f"{model_name}, N={N} K={K}, BF16 vs FP8 GEMMs TFLOP/s:")
+    for K, N, model in prepare_shapes(args):
+        print(f"{model}, N={N} K={K}, BF16 vs FP8 GEMMs TFLOP/s:")
         benchmark.run(
             print_data=True,
             show_plots=True,

benchmarks/kernels/bench_int8_gemm.py

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

benchmarks/kernels/bench_nvfp4_gemm.py

@@ -0,0 +1,141 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+import argparse
+import copy
+import itertools
+
+import torch
+from weight_shapes import WEIGHT_SHAPES
+
+from vllm import _custom_ops as ops
+from vllm.platforms import current_platform
+from vllm.scalar_type import scalar_types
+from vllm.triton_utils import triton
+
+if not current_platform.has_device_capability(100):
+    raise RuntimeError("NVFP4 requires compute capability of 10.0 (Blackwell)")
+
+
+FLOAT4_E2M1_MAX = scalar_types.float4_e2m1f.max()
+FLOAT8_E4M3_MAX = torch.finfo(torch.float8_e4m3fn).max
+
+PROVIDER_CFGS = {
+    "torch-bf16": dict(enabled=True),
+    "nvfp4": dict(no_a_quant=False, enabled=True),
+    "nvfp4-noquant": dict(no_a_quant=True, enabled=True),
+}
+
+_enabled = [k for k, v in PROVIDER_CFGS.items() if v["enabled"]]
+
+
+def _quant_weight_nvfp4(b: torch.Tensor, device: str):
+    # Compute global scale for weight
+    b_amax = torch.abs(b).max().to(torch.float32)
+    b_global_scale = FLOAT8_E4M3_MAX * FLOAT4_E2M1_MAX / b_amax
+    b_fp4, scale_b_fp4 = ops.scaled_fp4_quant(b, b_global_scale)
+    return b_fp4, scale_b_fp4, b_global_scale
+
+
+def build_nvfp4_runner(cfg, a, b, dtype, device):
+    b_fp4, scale_b_fp4, b_global_scale = _quant_weight_nvfp4(b, device)
+
+    # Compute global scale for activation
+    # NOTE: This is generally provided ahead-of-time by the model checkpoint.
+    a_amax = torch.abs(a).max().to(torch.float32)
+    a_global_scale = FLOAT8_E4M3_MAX * FLOAT4_E2M1_MAX / a_amax
+
+    # Alpha for the GEMM operation
+    alpha = 1.0 / (a_global_scale * b_global_scale)
+
+    if cfg["no_a_quant"]:
+        # Pre-quantize activation
+        a_fp4, scale_a_fp4 = ops.scaled_fp4_quant(a, a_global_scale)
+
+        def run():
+            return ops.cutlass_scaled_fp4_mm(
+                a_fp4, b_fp4, scale_a_fp4, scale_b_fp4, alpha, dtype
+            )
+
+        return run
+
+    # Quantize activation on-the-fly
+    def run():
+        a_fp4, scale_a_fp4 = ops.scaled_fp4_quant(a, a_global_scale)
+        return ops.cutlass_scaled_fp4_mm(
+            a_fp4, b_fp4, scale_a_fp4, scale_b_fp4, alpha, dtype
+        )
+
+    return run
+
+
+@triton.testing.perf_report(
+    triton.testing.Benchmark(
+        x_names=["batch_size"],
+        x_vals=[1, 16, 64, 128, 256, 512, 1024, 2048, 4096, 8192, 16384],
+        x_log=False,
+        line_arg="provider",
+        line_vals=_enabled,
+        line_names=_enabled,
+        ylabel="TFLOP/s (larger is better)",
+        plot_name="BF16 vs NVFP4 GEMMs",
+        args={},
+    )
+)
+def benchmark(batch_size, provider, N, K):
+    M = batch_size
+    device = "cuda"
+    dtype = torch.bfloat16
+
+    a = torch.randn((M, K), device=device, dtype=dtype)
+    b = torch.randn((N, K), device=device, dtype=dtype)
+
+    quantiles = [0.5, 0.2, 0.8]
+
+    if provider == "torch-bf16":
+        ms, min_ms, max_ms = triton.testing.do_bench_cudagraph(
+            lambda: torch.nn.functional.linear(a, b), quantiles=quantiles
+        )
+    else:
+        cfg = PROVIDER_CFGS[provider]
+        run_quant = build_nvfp4_runner(cfg, a, b, dtype, device)
+        ms, min_ms, max_ms = triton.testing.do_bench_cudagraph(
+            lambda: run_quant(), quantiles=quantiles
+        )
+
+    to_tflops = lambda t_ms: (2 * M * N * K) * 1e-12 / (t_ms * 1e-3)
+    return to_tflops(ms), to_tflops(max_ms), to_tflops(min_ms)
+
+
+def prepare_shapes(args):
+    out = []
+    for model, tp_size in itertools.product(args.models, args.tp_sizes):
+        for KN, tp_dim in copy.deepcopy(WEIGHT_SHAPES[model]):
+            KN[tp_dim] //= tp_size
+            KN.append(model)
+            out.append(KN)
+    return out
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--models",
+        nargs="+",
+        type=str,
+        default=["meta-llama/Llama-3.1-8B-Instruct"],
+        choices=list(WEIGHT_SHAPES.keys()),
+    )
+    parser.add_argument("--tp-sizes", nargs="+", type=int, default=[1])
+    args = parser.parse_args()
+
+    for K, N, model in prepare_shapes(args):
+        print(f"{model}, N={N} K={K}, BF16 vs NVFP4 GEMMs TFLOP/s:")
+        benchmark.run(
+            print_data=True,
+            show_plots=True,
+            save_path=f"bench_nvfp4_res_n{N}_k{K}",
+            N=N,
+            K=K,
+        )
+
+    print("Benchmark finished!")

benchmarks/kernels/bench_per_token_quant_fp8.py

@@ -0,0 +1,98 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+import itertools
+from typing import Callable
+
+import torch
+
+from vllm import _custom_ops as ops
+from vllm.config import CompilationConfig, VllmConfig, set_current_vllm_config
+from vllm.model_executor.layers.quantization.input_quant_fp8 import QuantFP8
+from vllm.model_executor.layers.quantization.utils.quant_utils import GroupShape
+from vllm.triton_utils import triton
+
+
+# TODO(luka): use standalone_compile utility
+def with_dyn_arg(fn: Callable, arg_index: int, dim_index: int):
+    def inner(*args):
+        torch._dynamo.mark_dynamic(args[arg_index], dim_index)
+        return fn(*args)
+
+    return inner
+
+
+torch._dynamo.config.recompile_limit = 8888
+compilation_config = CompilationConfig(custom_ops=["none"])
+with set_current_vllm_config(VllmConfig(compilation_config=compilation_config)):
+    torch_per_token_quant_fp8 = torch.compile(
+        QuantFP8(False, GroupShape.PER_TOKEN),
+        fullgraph=True,
+        dynamic=False,  # recompile for different shapes
+    )
+
+    # First dim is explicitly dynamic to simulate vLLM usage
+    torch_per_token_quant_fp8 = with_dyn_arg(torch_per_token_quant_fp8, 0, 0)
+
+
+def cuda_per_token_quant_fp8(
+    input: torch.Tensor,
+) -> tuple[torch.Tensor, torch.Tensor]:
+    return ops.scaled_fp8_quant(input)
+
+
+def calculate_diff(batch_size: int, seq_len: int):
+    """Calculate difference between Triton and CUDA implementations."""
+    device = torch.device("cuda")
+    x = torch.rand((batch_size * seq_len, 4096), dtype=torch.float16, device=device)
+
+    torch_out, torch_scale = torch_per_token_quant_fp8(x)
+    cuda_out, cuda_scale = cuda_per_token_quant_fp8(x)
+
+    if torch.allclose(
+        cuda_out.to(torch.float32), torch_out.to(torch.float32), rtol=1e-3, atol=1e-5
+    ) and torch.allclose(cuda_scale, torch_scale, rtol=1e-3, atol=1e-5):
+        print("✅ All implementations match")
+    else:
+        print("❌ Implementations differ")
+
+
+batch_size_range = [1, 16, 32, 64, 128]
+seq_len_range = [1, 16, 64, 128, 256, 512, 1024, 2048, 4096]
+
+configs = list(itertools.product(batch_size_range, seq_len_range))
+
+
+@triton.testing.perf_report(
+    triton.testing.Benchmark(
+        x_names=["batch_size", "seq_len"],
+        x_vals=configs,
+        line_arg="provider",
+        line_vals=["torch", "cuda"],
+        line_names=["Torch", "CUDA"],
+        styles=[("blue", "-"), ("green", "-")],
+        ylabel="us",
+        plot_name="per-token-dynamic-quant-fp8-performance",
+        args={},
+    )
+)
+def benchmark_quantization(batch_size, seq_len, provider):
+    dtype = torch.float16
+    device = torch.device("cuda")
+
+    x = torch.randn(batch_size * seq_len, 4096, device=device, dtype=dtype)
+
+    quantiles = [0.5, 0.2, 0.8]
+
+    if provider == "torch":
+        fn = lambda: torch_per_token_quant_fp8(x.clone())
+    elif provider == "cuda":
+        fn = lambda: cuda_per_token_quant_fp8(x.clone())
+
+    ms, min_ms, max_ms = triton.testing.do_bench_cudagraph(fn, quantiles=quantiles)
+
+    return 1000 * ms, 1000 * max_ms, 1000 * min_ms
+
+
+if __name__ == "__main__":
+    calculate_diff(batch_size=4, seq_len=4096)
+    benchmark_quantization.run(print_data=True)

benchmarks/kernels/benchmark_aqlm.py

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

benchmarks/kernels/benchmark_bitblas.py

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

benchmarks/kernels/benchmark_cutlass_fp4_moe.py

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

benchmarks/kernels/benchmark_grouped_gemm_cutlass.py

@@ -1,4 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 import torch
 import torch.utils.benchmark as benchmark
@@ -6,8 +7,8 @@ from benchmark_shapes import WEIGHT_SHAPES_MOE
 
 from vllm import _custom_ops as ops
 from vllm.config import ParallelConfig, VllmConfig, set_current_vllm_config
+from vllm.model_executor.layers.fused_moe.cutlass_moe import cutlass_moe_fp8
 from vllm.model_executor.layers.fused_moe.fused_moe import (
-    cutlass_moe_fp8,
     fused_experts,
     fused_topk,
 )
@@ -69,18 +70,9 @@ def bench_run(
     w1_scale = torch.empty((num_experts, 1, 1), device="cuda", dtype=torch.float32)
     w2_scale = torch.empty((num_experts, 1, 1), device="cuda", dtype=torch.float32)
 
-    ab_strides1 = torch.full((num_experts,), k, device="cuda", dtype=torch.int64)
-    c_strides1 = torch.full((num_experts,), 2 * n, device="cuda", dtype=torch.int64)
-    ab_strides2 = torch.full((num_experts,), n, device="cuda", dtype=torch.int64)
-    c_strides2 = torch.full((num_experts,), k, device="cuda", dtype=torch.int64)
-
     for expert in range(num_experts):
         w1_q[expert], w1_scale[expert] = ops.scaled_fp8_quant(w1[expert])
         w2_q[expert], w2_scale[expert] = ops.scaled_fp8_quant(w2[expert])
-    w1_q_notransp = w1_q.clone()
-    w2_q_notransp = w2_q.clone()
-    w1_q = w1_q.transpose(1, 2)
-    w2_q = w2_q.transpose(1, 2)
 
     score = torch.randn((m, num_experts), device="cuda", dtype=dtype)
 
@@ -88,6 +80,11 @@ def bench_run(
         a, score, topk, renormalize=False
     )
 
+    ab_strides1 = torch.full((num_experts,), k, device="cuda", dtype=torch.int64)
+    ab_strides2 = torch.full((num_experts,), n, device="cuda", dtype=torch.int64)
+    c_strides1 = torch.full((num_experts,), 2 * n, device="cuda", dtype=torch.int64)
+    c_strides2 = torch.full((num_experts,), k, device="cuda", dtype=torch.int64)
+
     def run_triton_moe(
         a: torch.Tensor,
         w1: torch.Tensor,
@@ -119,12 +116,13 @@ def bench_run(
         w2: torch.Tensor,
         w1_scale: torch.Tensor,
         w2_scale: torch.Tensor,
+        ab_strides1: torch.Tensor,
+        ab_strides2: torch.Tensor,
+        c_strides1: torch.Tensor,
+        c_strides2: torch.Tensor,
         topk_weights: torch.Tensor,
         topk_ids: torch.Tensor,
-        ab_strides1: torch.Tensor,
-        c_strides1: torch.Tensor,
-        ab_strides2: torch.Tensor,
-        c_strides2: torch.Tensor,
+        per_act_token: bool,
         num_repeats: int,
     ):
         for _ in range(num_repeats):
@@ -132,15 +130,16 @@ def bench_run(
                 a,
                 w1,
                 w2,
-                w1_scale,
-                w2_scale,
                 topk_weights,
                 topk_ids,
+                w1_scale,
+                w2_scale,
                 ab_strides1,
-                c_strides1,
                 ab_strides2,
+                c_strides1,
                 c_strides2,
-                a1_scale=a_scale,
+                per_act_token,
+                a1_scale=None,
             )
 
     def run_cutlass_from_graph(
@@ -150,12 +149,12 @@ def bench_run(
         w2_q: torch.Tensor,
         w1_scale: torch.Tensor,
         w2_scale: torch.Tensor,
+        ab_strides1: torch.Tensor,
+        ab_strides2: torch.Tensor,
+        c_strides1: torch.Tensor,
+        c_strides2: torch.Tensor,
         topk_weights: torch.Tensor,
         topk_ids: torch.Tensor,
-        ab_strides1: torch.Tensor,
-        c_strides1: torch.Tensor,
-        ab_strides2: torch.Tensor,
-        c_strides2: torch.Tensor,
     ):
         with set_current_vllm_config(
             VllmConfig(parallel_config=ParallelConfig(pipeline_parallel_size=1))
@@ -164,15 +163,16 @@ def bench_run(
                 a,
                 w1_q,
                 w2_q,
-                w1_scale,
-                w2_scale,
                 topk_weights,
                 topk_ids,
+                w1_scale,
+                w2_scale,
                 ab_strides1,
-                c_strides1,
                 ab_strides2,
+                c_strides1,
                 c_strides2,
-                a1_scale=a_scale,
+                per_act_token,
+                a1_scale=None,
             )
 
     def run_triton_from_graph(
@@ -215,12 +215,12 @@ def bench_run(
             w2_q,
             w1_scale,
             w2_scale,
+            ab_strides1,
+            ab_strides2,
+            c_strides1,
+            c_strides2,
             topk_weights,
             topk_ids,
-            ab_strides1,
-            c_strides1,
-            ab_strides2,
-            c_strides2,
         )
     torch.cuda.synchronize()
 
@@ -229,8 +229,8 @@ def bench_run(
     with torch.cuda.graph(triton_graph, stream=triton_stream):
         run_triton_from_graph(
             a,
-            w1_q_notransp,
-            w2_q_notransp,
+            w1_q,
+            w2_q,
             topk_weights,
             topk_ids,
             w1_scale,
@@ -249,17 +249,16 @@ def bench_run(
         "w2": w2,
         "score": score,
         "topk": topk,
-        "w1_q_notransp": w1_q_notransp,
-        "w2_q_notransp": w2_q_notransp,
         # Cutlass params
         "a_scale": a_scale,
         "w1_q": w1_q,
         "w2_q": w2_q,
         "w1_scale": w1_scale,
         "w2_scale": w2_scale,
+        "per_act_token": per_act_token,
         "ab_strides1": ab_strides1,
-        "c_strides1": c_strides1,
         "ab_strides2": ab_strides2,
+        "c_strides1": c_strides1,
         "c_strides2": c_strides2,
         # cuda graph params
         "cutlass_graph": cutlass_graph,
@@ -278,8 +277,8 @@ def bench_run(
     # Warmup
     run_triton_moe(
         a,
-        w1_q_notransp,
-        w2_q_notransp,
+        w1_q,
+        w2_q,
         topk_weights,
         topk_ids,
         w1_scale,
@@ -290,7 +289,7 @@ def bench_run(
 
     results.append(
         benchmark.Timer(
-            stmt="run_triton_moe(a, w1_q_notransp, w2_q_notransp, topk_weights, topk_ids, w1_scale, w2_scale, a_scale, num_runs)",  # noqa: E501
+            stmt="run_triton_moe(a, w1_q, w2_q, topk_weights, topk_ids, w1_scale, w2_scale, a_scale, num_runs)",  # noqa: E501
             globals=globals,
             label=label,
             sub_label=sub_label,
@@ -319,18 +318,19 @@ def bench_run(
         w2_q,
         w1_scale,
         w2_scale,
+        ab_strides1,
+        ab_strides2,
+        c_strides1,
+        c_strides2,
         topk_weights,
         topk_ids,
-        ab_strides1,
-        c_strides1,
-        ab_strides2,
-        c_strides2,
+        per_act_token,
         num_warmup,
     )
 
     results.append(
         benchmark.Timer(
-            stmt="run_cutlass_moe(a, a_scale, w1_q, w2_q, w1_scale, w2_scale, topk_weights, topk_ids, ab_strides1, c_strides1, ab_strides2, c_strides2, num_runs)",  # noqa: E501
+            stmt="run_cutlass_moe(a, a_scale, w1_q, w2_q, w1_scale, w2_scale, ab_strides1, ab_strides2, c_strides1, c_strides2, topk_weights, topk_ids, per_act_token, num_runs)",  # noqa: E501
             globals=globals,
             label=label,
             sub_label=sub_label,

benchmarks/kernels/benchmark_layernorm.py

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

benchmarks/kernels/benchmark_lora.py

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

benchmarks/kernels/benchmark_machete.py

@@ -1,4 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 import argparse
 import copy
@@ -233,8 +234,10 @@ def marlin_create_bench_fn(bt: BenchmarkTensors) -> Callable:
 
         fn = lambda: ops.gptq_marlin_gemm(
             a=bt.a,
+            c=None,
             b_q_weight=w_q,
             b_scales=w_s,
+            global_scale=None,
             b_zeros=w_zp,
             g_idx=g_idx,
             perm=sort_indices,

benchmarks/kernels/benchmark_marlin.py

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

benchmarks/kernels/benchmark_moe.py

@@ -1,4 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 import argparse
 import json
@@ -6,7 +7,6 @@ import time
 from contextlib import nullcontext
 from datetime import datetime
 from itertools import product
-from types import SimpleNamespace
 from typing import Any, TypedDict
 
 import ray
@@ -42,7 +42,7 @@ def benchmark_config(
     use_fp8_w8a8: bool,
     use_int8_w8a16: bool,
     num_iters: int = 100,
-    block_quant_shape: List[int] = None,
+    block_quant_shape: list[int] = None,
     use_deep_gemm: bool = False,
 ) -> float:
     init_dtype = torch.float16 if use_fp8_w8a8 else dtype
@@ -86,6 +86,9 @@ def benchmark_config(
             (num_experts, 2 * shard_intermediate_size), dtype=torch.float32
         )
         w2_scale = torch.randn((hidden_size, num_experts), dtype=torch.float32)
+    if use_deep_gemm:
+        # we use the default block shape for deepgemm
+        block_quant_shape = [128, 128]
     if use_fp8_w8a8:
         if block_quant_shape:
             block_n, block_k = block_quant_shape[0], block_quant_shape[1]
@@ -399,7 +402,7 @@ class BenchmarkWorker:
         dtype: torch.dtype,
         use_fp8_w8a8: bool,
         use_int8_w8a16: bool,
-        block_quant_shape: List[int] = None,
+        block_quant_shape: list[int] = None,
         use_deep_gemm: bool = False,
     ) -> tuple[dict[str, int], float]:
         current_platform.seed_everything(self.seed)
@@ -531,7 +534,7 @@ def save_configs(
     dtype: torch.dtype,
     use_fp8_w8a8: bool,
     use_int8_w8a16: bool,
-    block_quant_shape: List[int],
+    block_quant_shape: list[int],
 ) -> None:
     dtype_str = get_config_dtype_str(
         dtype, use_int8_w8a16=use_int8_w8a16, use_fp8_w8a8=use_fp8_w8a8
@@ -562,7 +565,6 @@ def main(args: argparse.Namespace):
     config = get_config(model=args.model, trust_remote_code=args.trust_remote_code)
     if args.model_prefix:
         config = getattr(config, args.model_prefix)
-    config = SimpleNamespace(**config)
 
     if config.architectures[0] == "DbrxForCausalLM":
         E = config.ffn_config.moe_num_experts
@@ -574,7 +576,11 @@ def main(args: argparse.Namespace):
         topk = config.num_experts_per_tok
         intermediate_size = config.intermediate_size
         shard_intermediate_size = 2 * intermediate_size // args.tp_size
-    elif config.architectures[0] in ("DeepseekV3ForCausalLM", "DeepseekV2ForCausalLM"):
+    elif config.architectures[0] in (
+        "DeepseekV3ForCausalLM",
+        "DeepseekV2ForCausalLM",
+        "Glm4MoeForCausalLM",
+    ):
         E = config.n_routed_experts
         topk = config.num_experts_per_tok
         intermediate_size = config.moe_intermediate_size
@@ -584,6 +590,11 @@ def main(args: argparse.Namespace):
         topk = config.num_experts_per_tok
         intermediate_size = config.moe_intermediate_size
         shard_intermediate_size = 2 * intermediate_size // args.tp_size
+    elif config.architectures[0] in ("HunYuanMoEV1ForCausalLM"):
+        E = config.num_experts
+        topk = config.moe_topk[0]
+        intermediate_size = config.moe_intermediate_size[0]
+        shard_intermediate_size = 2 * intermediate_size // args.tp_size
     else:
         # Support for llama4
         config = config.get_text_config()
@@ -594,11 +605,7 @@ def main(args: argparse.Namespace):
         shard_intermediate_size = 2 * intermediate_size // args.tp_size
 
     hidden_size = config.hidden_size
-    dtype = (
-        torch.float16
-        if current_platform.is_rocm()
-        else getattr(torch, config.torch_dtype)
-    )
+    dtype = torch.float16 if current_platform.is_rocm() else config.torch_dtype
     use_fp8_w8a8 = args.dtype == "fp8_w8a8"
     use_int8_w8a16 = args.dtype == "int8_w8a16"
     block_quant_shape = get_weight_block_size_safety(config)
@@ -625,7 +632,7 @@ def main(args: argparse.Namespace):
             4096,
         ]
     else:
-        batch_sizes = [args.batch_size]
+        batch_sizes = args.batch_size
 
     use_deep_gemm = bool(args.use_deep_gemm)
 
@@ -733,7 +740,7 @@ if __name__ == "__main__":
     )
     parser.add_argument("--use-deep-gemm", action="store_true")
     parser.add_argument("--seed", type=int, default=0)
-    parser.add_argument("--batch-size", type=int, required=False)
+    parser.add_argument("--batch-size", type=int, nargs="+", required=False)
     parser.add_argument("--tune", action="store_true")
     parser.add_argument("--trust-remote-code", action="store_true")
     parser.add_argument("--model-prefix", type=str, required=False)

benchmarks/kernels/benchmark_moe_align_block_size.py

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

benchmarks/kernels/benchmark_moe_permute_unpermute.py

@@ -1,4 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 import argparse
 from typing import Any, TypedDict
@@ -317,6 +318,7 @@ def main(args: argparse.Namespace):
     elif (
         config.architectures[0] == "DeepseekV3ForCausalLM"
         or config.architectures[0] == "DeepseekV2ForCausalLM"
+        or config.architectures[0] == "Glm4MoeForCausalLM"
     ):
         E = config.n_routed_experts
         topk = config.num_experts_per_tok

benchmarks/kernels/benchmark_paged_attention.py

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

benchmarks/kernels/benchmark_quant.py

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

benchmarks/kernels/benchmark_rmsnorm.py

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

benchmarks/kernels/benchmark_rope.py

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

benchmarks/kernels/benchmark_shapes.py

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

benchmarks/kernels/benchmark_trtllm_attention.py

@@ -0,0 +1,240 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+import csv
+import os
+import random
+from datetime import datetime
+
+import flashinfer
+import torch
+
+FLOAT32_BYTES = torch.finfo(torch.float).bits // 8
+
+# KV Cache Layout for TRT-LLM
+# kv_cache_shape = (num_blocks, 2, num_kv_heads, page_size, head_dim)
+
+
+def to_float8(x, dtype=torch.float8_e4m3fn):
+    finfo = torch.finfo(dtype)
+    min_val, max_val = x.aminmax()
+    amax = torch.maximum(min_val.abs(), max_val.abs()).clamp(min=1e-12)
+    scale = finfo.max / amax * 0.1
+    x_scl_sat = (x * scale).clamp(min=finfo.min, max=finfo.max)
+    return x_scl_sat.to(dtype), scale.float().reciprocal()
+
+
+@torch.no_grad()
+def benchmark_decode(
+    num_seqs,
+    max_seq_len,
+    page_size=16,
+    dtype=torch.bfloat16,
+    kv_layout="HND",
+    num_kv_heads=8,
+    kv_cache_dtype="auto",
+    head_dim=128,
+    warmup=10,
+    trials=20,
+):
+    torch.set_default_device("cuda")
+    device = "cuda"
+    torch.manual_seed(0)
+
+    # Currently only HEAD_GRP_SIZE == 8 is supported
+    HEAD_GRP_SIZE = 8
+    MAX_SEQ_LEN = max_seq_len
+
+    # large number to reduce kv_cache reuse
+    NUM_BLOCKS = int(256000 / page_size)
+
+    workspace_buffer = torch.empty(1024 * 1024 * 1024, dtype=torch.int8, device=device)
+
+    # For decode, batch_size is num_decode_token
+    num_qo_heads = num_kv_heads * HEAD_GRP_SIZE
+    sm_scale = float(1.0 / (head_dim**0.5))
+    q = torch.randn(num_seqs, num_qo_heads, head_dim, device=device, dtype=dtype)
+    kv_lens = [random.randint(1, MAX_SEQ_LEN) for _ in range(num_seqs)]
+
+    max_kv_len = max(kv_lens)
+    kv_lens_tensor = torch.tensor(kv_lens, dtype=torch.int, device=device)
+    max_num_blocks_per_seq = (max_kv_len + page_size - 1) // page_size
+
+    block_tables = torch.randint(
+        0, NUM_BLOCKS, (num_seqs, max_num_blocks_per_seq), dtype=torch.int32
+    )
+
+    kv_cache_shape = (NUM_BLOCKS, 2, num_kv_heads, page_size, head_dim)
+    kv_cache = torch.randn(size=kv_cache_shape, device=device, dtype=dtype)
+    k_scale = v_scale = 1.0
+
+    if kv_cache_dtype.startswith("fp8"):
+        kv_cache, _ = to_float8(kv_cache)
+
+    # Benchmark TRT decode
+    def trt_decode():
+        return flashinfer.decode.trtllm_batch_decode_with_kv_cache(
+            q,
+            kv_cache,
+            workspace_buffer,
+            num_qo_heads,
+            num_kv_heads,
+            sm_scale,
+            block_tables,
+            kv_lens_tensor,
+            page_size,
+            max_kv_len,
+            kv_cache_dtype,
+            k_scale,
+            v_scale,
+        )
+
+    def time_fn(fn, warmup=10, trials=20):
+        torch.cuda.synchronize()
+        start = torch.cuda.Event(enable_timing=True)
+        end = torch.cuda.Event(enable_timing=True)
+        times = []
+        for i in range(warmup):
+            fn()
+        for i in range(trials):
+            start.record()
+            fn()
+            end.record()
+            torch.cuda.synchronize()
+            times.append(start.elapsed_time(end))  # ms
+        return sum(times) / len(times), torch.std(torch.tensor(times))
+
+    # TRT Decode
+    trt_mean, trt_std = time_fn(trt_decode)
+
+    kv_indptr = [0]
+    kv_indices = []
+    kv_last_page_lens = []
+    for i in range(num_seqs):
+        seq_len = kv_lens[i]
+        assert seq_len > 0
+        num_blocks = (seq_len + page_size - 1) // page_size
+        kv_indices.extend(block_tables[i, :num_blocks])
+        kv_indptr.append(kv_indptr[-1] + num_blocks)
+        kv_last_page_len = seq_len % page_size
+        if kv_last_page_len == 0:
+            kv_last_page_len = page_size
+        kv_last_page_lens.append(kv_last_page_len)
+
+    kv_indptr = torch.tensor(kv_indptr, dtype=torch.int32)
+    kv_indices = torch.tensor(kv_indices, dtype=torch.int32)
+    kv_last_page_lens = torch.tensor(kv_last_page_lens, dtype=torch.int32)
+
+    wrapper = flashinfer.BatchDecodeWithPagedKVCacheWrapper(
+        workspace_buffer,
+        kv_layout,
+        use_tensor_cores=((num_qo_heads // num_kv_heads) > 4),
+    )
+
+    wrapper.plan(
+        kv_indptr,
+        kv_indices,
+        kv_last_page_lens,
+        num_qo_heads,
+        num_kv_heads,
+        head_dim,
+        page_size,
+        "NONE",
+        q_data_type=dtype,
+        kv_data_type=torch.float8_e4m3fn if kv_cache_dtype.startswith("fp8") else dtype,
+    )
+
+    def baseline_decode():
+        return wrapper.run(q, kv_cache, sm_scale, k_scale, v_scale)
+
+    baseline_mean, baseline_std = time_fn(baseline_decode)
+
+    # Calculate percentage speedup (positive means TRT is faster)
+    speedup_percent = (baseline_mean - trt_mean) / baseline_mean
+
+    print(
+        f"\t{num_seqs}\t{max_seq_len}\t{trt_mean:.3f}\t{trt_std.item():.3f}"
+        f"\t{baseline_mean:.3f}\t{baseline_std.item():.3f}\t{speedup_percent:.3f}"
+    )
+
+    # Return results for CSV writing
+    return {
+        "num_seqs": num_seqs,
+        "trt_mean": trt_mean,
+        "trt_std": trt_std.item(),
+        "baseline_mean": baseline_mean,
+        "baseline_std": baseline_std.item(),
+        "speedup_percent": speedup_percent,
+        "q_dtype": str(dtype),
+        "kv_cache_dtype": kv_cache_dtype,
+        "page_size": page_size,
+        "num_kv_heads": num_kv_heads,
+        "head_dim": head_dim,
+        "max_seq_len": max_seq_len,
+    }
+
+
+def write_results_to_csv(results, filename=None):
+    """Write benchmark results to CSV file."""
+    if filename is None:
+        timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")
+        filename = f"flashinfer_trtllm_benchmark_{timestamp}.csv"
+
+    fieldnames = [
+        "num_seqs",
+        "trt_mean",
+        "trt_std",
+        "baseline_mean",
+        "baseline_std",
+        "speedup_percent",
+        "q_dtype",
+        "kv_cache_dtype",
+        "page_size",
+        "num_kv_heads",
+        "head_dim",
+        "max_seq_len",
+    ]
+
+    file_exists = os.path.exists(filename)
+
+    with open(filename, "a", newline="") as csvfile:
+        writer = csv.DictWriter(csvfile, fieldnames=fieldnames)
+
+        if not file_exists:
+            writer.writeheader()
+
+        for result in results:
+            writer.writerow(result)
+
+    print(f"Results written to {filename}")
+
+
+if __name__ == "__main__":
+    num_seqs = [1, 4, 8, 16, 32, 64, 128, 256]
+    max_seq_lens = [1024, 2048, 4096, 8192, 16384, 32768, 65536, 131072]
+    all_results = []
+
+    print("Running benchmark for kv_cache_dtype: bfloat16")
+    print(
+        "\tnum_seqs\tmax_seq_len\ttrt_mean\ttrt_std\tbaseline_mean\tbaseline_std\tspeedup_percent"
+    )
+    for max_seq_len in max_seq_lens:
+        for bs in num_seqs:
+            result = benchmark_decode(
+                bs, max_seq_len, dtype=torch.bfloat16, kv_cache_dtype="auto"
+            )
+            all_results.append(result)
+
+    print("Running benchmark for q_dtype = bfloat16, kv_cache_dtype: fp8")
+    print(
+        "\tnum_seqs\tmax_seq_len\ttrt_mean\ttrt_std\tbaseline_mean\tbaseline_std\tspeedup_percent"
+    )
+    for max_seq_len in max_seq_lens:
+        for bs in num_seqs:
+            result = benchmark_decode(
+                bs, max_seq_len, dtype=torch.bfloat16, kv_cache_dtype="fp8"
+            )
+            all_results.append(result)
+
+    # Write all results to CSV
+    write_results_to_csv(all_results)

benchmarks/kernels/benchmark_w8a8_block_fp8.py

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

benchmarks/kernels/deepgemm/benchmark_fp8_block_dense_gemm.py

@@ -1,4 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 # fmt: off
 # ruff: noqa: E501
 import time
@@ -84,12 +85,6 @@ def benchmark_shape(m: int,
 
     # === DeepGEMM Implementation ===
     def deepgemm_gemm():
-        # A quantization is inside the loop as it depends on activations
-        # A_deepgemm, A_scale_deepgemm = per_token_cast_to_fp8(A)
-        # A_deepgemm, A_scale_deepgemm = per_token_group_quant_fp8(
-        #     A, block_size[1])
-        # A_scale_aligned = get_col_major_tma_aligned_tensor(A_scale_deepgemm)
-        # C_deepgemm = torch.empty((m, n), device='cuda', dtype=torch.bfloat16)
         deep_gemm.gemm_fp8_fp8_bf16_nt((A_deepgemm, A_scale_deepgemm),
                                        (B_deepgemm, B_scale_deepgemm),
                                        C_deepgemm)
@@ -97,8 +92,6 @@ def benchmark_shape(m: int,
 
     # === vLLM Triton Implementation ===
     def vllm_triton_gemm():
-        # A quantization is inside the loop as it depends on activations
-        # A_vllm, A_scale_vllm = per_token_group_quant_fp8(A, block_size[1])
         return w8a8_block_fp8_matmul(A_vllm,
                                      B_vllm,
                                      A_scale_vllm,
@@ -108,9 +101,6 @@ def benchmark_shape(m: int,
 
     # === vLLM CUTLASS Implementation ===
     def vllm_cutlass_gemm():
-        # A quantization is inside the loop as it depends on activations
-        # A_vllm_cutlass, A_scale_vllm_cutlass = per_token_group_quant_fp8(
-        #     A, block_size[1], column_major_scales=True)
         return ops.cutlass_scaled_mm(A_vllm_cutlass,
                                      B_vllm.T,
                                      scale_a=A_scale_vllm_cutlass,

benchmarks/kernels/graph_machete_bench.py

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

benchmarks/kernels/utils.py

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

benchmarks/kernels/weight_shapes.py

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

benchmarks/kv_cache/benchmark_block_pool.py

@@ -0,0 +1,108 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+import gc
+import time
+from typing import Optional
+
+from tabulate import tabulate
+
+from vllm.utils import FlexibleArgumentParser
+from vllm.v1.core.block_pool import BlockPool
+
+
+class Metric:
+    def __init__(self) -> None:
+        self.cnt: int = 0
+        self.sum_v: int = 0
+        self.max_v: Optional[int] = None
+
+    def update(self, v: int) -> None:
+        self.cnt += 1
+        self.sum_v += v
+        if self.max_v is None:
+            self.max_v = v
+        else:
+            self.max_v = max(self.max_v, v)
+
+    def avg_v(self) -> float:
+        return self.sum_v * 1.0 / self.cnt
+
+
+def main(args):
+    rows = []
+    for allocate_block in args.allocate_blocks:
+        # Enforce a GC collect ahead to minimize the impact among runs
+        gc.collect()
+        block_pool = BlockPool(num_gpu_blocks=args.num_gpu_blocks, enable_caching=True)
+
+        get_blocks_metric: Metric = Metric()
+        free_blocks_metric: Metric = Metric()
+        for _ in range(args.num_iteration):
+            t1 = time.monotonic_ns()
+            blocks = block_pool.get_new_blocks(allocate_block)
+            t2 = time.monotonic_ns()
+            block_pool.free_blocks(blocks)
+            t3 = time.monotonic_ns()
+            get_blocks_metric.update(t2 - t1)
+            free_blocks_metric.update(t3 - t2)
+
+        if get_blocks_metric.max_v is not None and free_blocks_metric.max_v is not None:
+            rows.append(
+                [
+                    get_blocks_metric.cnt,
+                    args.num_gpu_blocks,
+                    allocate_block,
+                    get_blocks_metric.avg_v() / 1000000,
+                    get_blocks_metric.max_v / 1000000.0,
+                    free_blocks_metric.avg_v() / 1000000,
+                    free_blocks_metric.max_v / 1000000.0,
+                ]
+            )
+        else:
+            print(
+                "No valid metrics found."
+                f" {get_blocks_metric.max_v=} {free_blocks_metric.max_v=}"
+            )
+
+    print(
+        tabulate(
+            rows,
+            headers=[
+                "Iterations",
+                "Total\nBlocks",
+                "Allocated\nBlocks",
+                "Get Blocks\nAvg (ms)",
+                "Get Blocks\nMax (ms)",
+                "Free Blocks\nAvg (ms)",
+                "Free Blocks\nMax (ms)",
+            ],
+            tablefmt="grid",
+            floatfmt=".6f",
+        )
+    )
+
+
+def invoke_main() -> None:
+    parser = FlexibleArgumentParser(
+        description="Benchmark the performance of BlockPool for KV Cache."
+    )
+    parser.add_argument("--num-gpu-blocks", type=int, default=100000)
+    parser.add_argument(
+        "--num-iteration",
+        type=int,
+        default=1000,
+        help="Number of iterations to run to stablize final data readings",
+    )
+    parser.add_argument(
+        "--allocate-blocks",
+        type=int,
+        nargs="*",
+        default=[10, 50, 100, 500, 1000],
+        help="Number of blocks to allocate",
+    )
+    args = parser.parse_args()
+    main(args)
+
+
+if __name__ == "__main__":
+    invoke_main()  # pragma: no cover

benchmarks/overheads/benchmark_hashing.py

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

cmake/cpu_extension.cmake

@@ -12,9 +12,8 @@ endif()
 #
 # Define environment variables for special configurations
 #
-if(DEFINED ENV{VLLM_CPU_AVX512BF16})
-    set(ENABLE_AVX512BF16 ON)
-endif()
+set(ENABLE_AVX512BF16 $ENV{VLLM_CPU_AVX512BF16})
+set(ENABLE_AVX512VNNI $ENV{VLLM_CPU_AVX512VNNI})
 
 include_directories("${CMAKE_SOURCE_DIR}/csrc")
 
@@ -75,6 +74,7 @@ if (MACOSX_FOUND AND CMAKE_SYSTEM_PROCESSOR STREQUAL "arm64")
 else()
     find_isa(${CPUINFO} "avx2" AVX2_FOUND)
     find_isa(${CPUINFO} "avx512f" AVX512_FOUND)
+    find_isa(${CPUINFO} "Power11" POWER11_FOUND)
     find_isa(${CPUINFO} "POWER10" POWER10_FOUND)
     find_isa(${CPUINFO} "POWER9" POWER9_FOUND)
     find_isa(${CPUINFO} "asimd" ASIMD_FOUND) # Check for ARM NEON support
@@ -95,24 +95,48 @@ if (AVX512_FOUND AND NOT AVX512_DISABLED)
         if (CMAKE_CXX_COMPILER_ID STREQUAL "GNU" AND
             CMAKE_CXX_COMPILER_VERSION VERSION_GREATER_EQUAL 12.3)
             list(APPEND CXX_COMPILE_FLAGS "-mavx512bf16")
+            set(ENABLE_AVX512BF16 ON)
         else()
+            set(ENABLE_AVX512BF16 OFF)
             message(WARNING "Disable AVX512-BF16 ISA support, requires gcc/g++ >= 12.3")
         endif()
     else()
+        set(ENABLE_AVX512BF16 OFF)
         message(WARNING "Disable AVX512-BF16 ISA support, no avx512_bf16 found in local CPU flags." " If cross-compilation is required, please set env VLLM_CPU_AVX512BF16=1.")
     endif()
+
+    find_isa(${CPUINFO} "avx512_vnni" AVX512VNNI_FOUND)
+    if (AVX512VNNI_FOUND OR ENABLE_AVX512VNNI)
+        if (CMAKE_CXX_COMPILER_ID STREQUAL "GNU" AND
+            CMAKE_CXX_COMPILER_VERSION VERSION_GREATER_EQUAL 12.3)
+            list(APPEND CXX_COMPILE_FLAGS "-mavx512vnni")
+            set(ENABLE_AVX512VNNI ON)
+        else()
+            set(ENABLE_AVX512VNNI OFF)
+            message(WARNING "Disable AVX512-VNNI ISA support, requires gcc/g++ >= 12.3")
+        endif()
+    else()
+        set(ENABLE_AVX512VNNI OFF)
+        message(WARNING "Disable AVX512-VNNI ISA support, no avx512_vnni found in local CPU flags." " If cross-compilation is required, please set env VLLM_CPU_AVX512VNNI=1.")
+    endif()
     
 elseif (AVX2_FOUND)
     list(APPEND CXX_COMPILE_FLAGS "-mavx2")
     message(WARNING "vLLM CPU backend using AVX2 ISA")
     
-elseif (POWER9_FOUND OR POWER10_FOUND)
+elseif (POWER9_FOUND OR POWER10_FOUND OR POWER11_FOUND)
     message(STATUS "PowerPC detected")
-    # Check for PowerPC VSX support
-    list(APPEND CXX_COMPILE_FLAGS
-        "-mvsx"
-        "-mcpu=native"
-        "-mtune=native")
+    if (POWER9_FOUND)
+        list(APPEND CXX_COMPILE_FLAGS
+            "-mvsx"
+            "-mcpu=power9"
+            "-mtune=power9")
+    elseif (POWER10_FOUND OR POWER11_FOUND)
+        list(APPEND CXX_COMPILE_FLAGS
+            "-mvsx"
+            "-mcpu=power10"
+            "-mtune=power10")
+    endif()
 
 elseif (ASIMD_FOUND)
     message(STATUS "ARMv8 or later architecture detected")
@@ -141,17 +165,32 @@ else()
 endif()
 
 #
-# Build oneDNN for W8A8 GEMM kernels (only for x86-AVX512 platforms)
-#
-if (AVX512_FOUND AND NOT AVX512_DISABLED)
+# Build oneDNN for W8A8 GEMM kernels (only for x86-AVX512 /ARM platforms)
+# Flag to enable ACL kernels for AARCH64 platforms
+if ( VLLM_BUILD_ACL STREQUAL "ON")
+    set(USE_ACL ON)
+else()
+    set(USE_ACL OFF)
+endif()
+
+if ((AVX512_FOUND AND NOT AVX512_DISABLED) OR ASIMD_FOUND)
     FetchContent_Declare(
         oneDNN
         GIT_REPOSITORY https://github.com/oneapi-src/oneDNN.git
-        GIT_TAG  v3.7.1
+        GIT_TAG  v3.8.1
         GIT_PROGRESS TRUE
         GIT_SHALLOW TRUE
     )
 
+    if(USE_ACL)
+        find_library(ARM_COMPUTE_LIBRARY NAMES arm_compute PATHS $ENV{ACL_ROOT_DIR}/build/)
+        if(NOT ARM_COMPUTE_LIBRARY)
+            message(FATAL_ERROR "Could not find ARM Compute Library: please set ACL_ROOT_DIR")
+        endif()
+        set(ONEDNN_AARCH64_USE_ACL "ON")
+        set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wl,-rpath,$ENV{ACL_ROOT_DIR}/build/")
+        endif()
+
     set(ONEDNN_LIBRARY_TYPE "STATIC")
     set(ONEDNN_BUILD_DOC "OFF")
     set(ONEDNN_BUILD_EXAMPLES "OFF")
@@ -224,11 +263,29 @@ if (AVX512_FOUND AND NOT AVX512_DISABLED)
         "csrc/cpu/quant.cpp"
         "csrc/cpu/shm.cpp"
         ${VLLM_EXT_SRC})
+    if (ENABLE_AVX512BF16 AND ENABLE_AVX512VNNI)
+        set(VLLM_EXT_SRC
+            "csrc/cpu/sgl-kernels/gemm.cpp"
+            "csrc/cpu/sgl-kernels/gemm_int8.cpp"
+            "csrc/cpu/sgl-kernels/gemm_fp8.cpp"
+            "csrc/cpu/sgl-kernels/moe.cpp"
+            "csrc/cpu/sgl-kernels/moe_int8.cpp"
+            "csrc/cpu/sgl-kernels/moe_fp8.cpp"
+            ${VLLM_EXT_SRC})
+        add_compile_definitions(-DCPU_CAPABILITY_AVX512)
+    endif()
 elseif(POWER10_FOUND)
     set(VLLM_EXT_SRC
         "csrc/cpu/quant.cpp"
         ${VLLM_EXT_SRC})
 endif()
+if (ASIMD_FOUND)
+    set(VLLM_EXT_SRC
+        "csrc/cpu/quant.cpp"
+        ${VLLM_EXT_SRC})
+endif()
+
+message(STATUS "CPU extension source files: ${VLLM_EXT_SRC}")
 
 #
 # Define extension targets

cmake/external_projects/vllm_flash_attn.cmake

@@ -38,7 +38,7 @@ else()
   FetchContent_Declare(
           vllm-flash-attn
           GIT_REPOSITORY https://github.com/vllm-project/flash-attention.git
-          GIT_TAG 8798f27777fb57f447070301bf33a9f9c607f491
+          GIT_TAG 1c2624e53c078854e0637ee566c72fe2107e75f4
           GIT_PROGRESS TRUE
           # Don't share the vllm-flash-attn build between build types
           BINARY_DIR ${CMAKE_BINARY_DIR}/vllm-flash-attn

cmake/hipify.py

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

cmake/utils.cmake

@@ -122,6 +122,7 @@ function (get_torch_gpu_compiler_flags OUT_GPU_FLAGS GPU_LANG)
       "-DENABLE_FP8"
       "-U__HIP_NO_HALF_CONVERSIONS__"
       "-U__HIP_NO_HALF_OPERATORS__"
+      "-Werror=unused-variable"
       "-fno-gpu-rdc")
 
   endif()
@@ -264,8 +265,8 @@ macro(set_gencode_flags_for_srcs)
 endmacro()
 
 #
-# For the given `SRC_CUDA_ARCHS` list of gencode versions in the form 
-#  `<major>.<minor>[letter]` compute the "loose intersection" with the 
+# For the given `SRC_CUDA_ARCHS` list of gencode versions in the form
+#  `<major>.<minor>[letter]` compute the "loose intersection" with the
 #  `TGT_CUDA_ARCHS` list of gencodes. We also support the `+PTX` suffix in
 #  `SRC_CUDA_ARCHS` which indicates that the PTX code should be built when there
 #  is a CUDA_ARCH in `TGT_CUDA_ARCHS` that is equal to or larger than the
@@ -277,7 +278,7 @@ endmacro()
 #  in `SRC_CUDA_ARCHS` that is less or equal to the version in `TGT_CUDA_ARCHS`.
 # We have special handling for x.0a, if x.0a is in `SRC_CUDA_ARCHS` and x.0 is
 #  in `TGT_CUDA_ARCHS` then we should remove x.0a from `SRC_CUDA_ARCHS` and add
-#  x.0a to the result (and remove x.0 from TGT_CUDA_ARCHS). 
+#  x.0a to the result (and remove x.0 from TGT_CUDA_ARCHS).
 # The result is stored in `OUT_CUDA_ARCHS`.
 #
 # Example:
@@ -312,21 +313,16 @@ function(cuda_archs_loose_intersection OUT_CUDA_ARCHS SRC_CUDA_ARCHS TGT_CUDA_AR
   # if x.0a is in SRC_CUDA_ARCHS and x.0 is in CUDA_ARCHS then we should
   # remove x.0a from SRC_CUDA_ARCHS and add x.0a to _CUDA_ARCHS
   set(_CUDA_ARCHS)
-  if ("9.0a" IN_LIST _SRC_CUDA_ARCHS)
-    list(REMOVE_ITEM _SRC_CUDA_ARCHS "9.0a")
-    if ("9.0" IN_LIST TGT_CUDA_ARCHS)
-      list(REMOVE_ITEM _TGT_CUDA_ARCHS "9.0")
-      set(_CUDA_ARCHS "9.0a")
+  foreach(_arch ${_SRC_CUDA_ARCHS})
+    if(_arch MATCHES "\\a$")
+      list(REMOVE_ITEM _SRC_CUDA_ARCHS "${_arch}")
+      string(REPLACE "a" "" _base "${_arch}")
+      if ("${_base}" IN_LIST TGT_CUDA_ARCHS)
+        list(REMOVE_ITEM _TGT_CUDA_ARCHS "${_base}")
+        list(APPEND _CUDA_ARCHS "${_arch}")
+      endif()
     endif()
-  endif()
-
-  if ("10.0a" IN_LIST _SRC_CUDA_ARCHS)
-    list(REMOVE_ITEM _SRC_CUDA_ARCHS "10.0a")
-    if ("10.0" IN_LIST TGT_CUDA_ARCHS)
-      list(REMOVE_ITEM _TGT_CUDA_ARCHS "10.0")
-      set(_CUDA_ARCHS "10.0a")
-    endif()
-  endif()
+  endforeach()
 
   list(SORT _SRC_CUDA_ARCHS COMPARE NATURAL ORDER ASCENDING)
 
@@ -358,7 +354,7 @@ function(cuda_archs_loose_intersection OUT_CUDA_ARCHS SRC_CUDA_ARCHS TGT_CUDA_AR
   endforeach()
 
   list(REMOVE_DUPLICATES _CUDA_ARCHS)
-  
+
   # reapply +PTX suffix to architectures that requested PTX
   set(_FINAL_ARCHS)
   foreach(_arch ${_CUDA_ARCHS})
@@ -369,7 +365,7 @@ function(cuda_archs_loose_intersection OUT_CUDA_ARCHS SRC_CUDA_ARCHS TGT_CUDA_AR
     endif()
   endforeach()
   set(_CUDA_ARCHS ${_FINAL_ARCHS})
-  
+
   set(${OUT_CUDA_ARCHS} ${_CUDA_ARCHS} PARENT_SCOPE)
 endfunction()
 

csrc/attention/attention_kernels.cuh

@@ -24,6 +24,7 @@
 
 #include "attention_dtypes.h"
 #include "attention_utils.cuh"
+#include "cuda_compat.h"
 
 #ifdef USE_ROCM
   #include <hip/hip_bf16.h>
@@ -33,12 +34,6 @@ typedef __hip_bfloat16 __nv_bfloat16;
   #include "../quantization/fp8/nvidia/quant_utils.cuh"
 #endif
 
-#ifndef USE_ROCM
-  #define WARP_SIZE 32
-#else
-  #define WARP_SIZE warpSize
-#endif
-
 #define MAX(a, b) ((a) > (b) ? (a) : (b))
 #define MIN(a, b) ((a) < (b) ? (a) : (b))
 #define DIVIDE_ROUND_UP(a, b) (((a) + (b) - 1) / (b))
@@ -670,7 +665,6 @@ __global__ void paged_attention_v2_reduce_kernel(
 
 }  // namespace vllm
 
-#undef WARP_SIZE
 #undef MAX
 #undef MIN
 #undef DIVIDE_ROUND_UP