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
@@ -8,12 +9,12 @@ import zipfile
 # Note that we have 400 MiB quota, please use it wisely.
 # See https://github.com/pypi/support/issues/3792 .
 # Please also sync the value with the one in Dockerfile.
-VLLM_MAX_SIZE_MB = int(os.environ.get('VLLM_MAX_SIZE_MB', 400))
+VLLM_MAX_SIZE_MB = int(os.environ.get("VLLM_MAX_SIZE_MB", 400))
 
 
 def print_top_10_largest_files(zip_file):
     """Print the top 10 largest files in the given zip file."""
-    with zipfile.ZipFile(zip_file, 'r') as z:
+    with zipfile.ZipFile(zip_file, "r") as z:
         file_sizes = [(f, z.getinfo(f).file_size) for f in z.namelist()]
         file_sizes.sort(key=lambda x: x[1], reverse=True)
         for f, size in file_sizes[:10]:
@@ -28,14 +29,18 @@ def check_wheel_size(directory):
                 wheel_path = os.path.join(root, file_name)
                 wheel_size_mb = os.path.getsize(wheel_path) / (1024 * 1024)
                 if wheel_size_mb > VLLM_MAX_SIZE_MB:
-                    print(f"Not allowed: Wheel {wheel_path} is larger "
-                          f"({wheel_size_mb:.2f} MB) than the limit "
-                          f"({VLLM_MAX_SIZE_MB} MB).")
+                    print(
+                        f"Not allowed: Wheel {wheel_path} is larger "
+                        f"({wheel_size_mb:.2f} MB) than the limit "
+                        f"({VLLM_MAX_SIZE_MB} MB)."
+                    )
                     print_top_10_largest_files(wheel_path)
                     return 1
                 else:
-                    print(f"Wheel {wheel_path} is within the allowed size "
-                          f"({wheel_size_mb:.2f} MB).")
+                    print(
+                        f"Wheel {wheel_path} is within the allowed size "
+                        f"({wheel_size_mb:.2f} MB)."
+                    )
     return 0
 
 
@@ -45,4 +50,4 @@ if __name__ == "__main__":
         sys.exit(1)
 
     directory = sys.argv[1]
-    sys.exit(check_wheel_size(directory))
+    sys.exit(check_wheel_size(directory))

.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
@@ -22,5 +23,5 @@ with open("index.html", "w") as f:
     print(f"Generated index.html for {args.wheel}")
     # cloudfront requires escaping the '+' character
     f.write(
-        template.format(wheel=filename,
-                        wheel_html_escaped=filename.replace("+", "%2B")))
+        template.format(wheel=filename, wheel_html_escaped=filename.replace("+", "%2B"))
+    )

.buildkite/lm-eval-harness/configs/Meta-Llama-3.2-1B-Instruct-FP8-compressed-tensors.yaml

@@ -0,0 +1,11 @@
+# bash .buildkite/lm-eval-harness/run-lm-eval-gsm-vllm-baseline.sh -m RedHatAI/Llama-3.2-1B-Instruct-FP8 -b "auto" -l 1319 -f 5 -t 1
+model_name: "RedHatAI/Llama-3.2-1B-Instruct-FP8"
+tasks:
+- name: "gsm8k"
+  metrics:
+  - name: "exact_match,strict-match"
+    value: 0.335
+  - name: "exact_match,flexible-extract"
+    value: 0.323
+limit: 1319
+num_fewshot: 5

.buildkite/lm-eval-harness/configs/Qwen2.5-1.5B-Instruct.yaml

@@ -0,0 +1,11 @@
+# bash .buildkite/lm-eval-harness/run-lm-eval-gsm-vllm-baseline.sh -m Qwen/Qwen2.5-1.5B-Instruct -b auto -l 1319 -f 5 -t 1
+model_name: "Qwen/Qwen2.5-1.5B-Instruct"
+tasks:
+- name: "gsm8k"
+  metrics:
+  - name: "exact_match,strict-match"
+    value: 0.54
+  - name: "exact_match,flexible-extract"
+    value: 0.59
+limit: 1319
+num_fewshot: 5

.buildkite/lm-eval-harness/configs/Qwen2.5-VL-3B-Instruct-FP8-dynamic.yaml

@@ -0,0 +1,11 @@
+# bash .buildkite/lm-eval-harness/run-lm-eval-gsm-vllm-baseline.sh -m RedHatAI/Qwen2.5-VL-3B-Instruct-FP8-Dynamic -b auto -l 1319 -f 5 -t 1
+model_name: "RedHatAI/Qwen2.5-VL-3B-Instruct-FP8-Dynamic"
+tasks:
+- name: "gsm8k"
+  metrics:
+  - name: "exact_match,strict-match"
+    value: 0.47
+  - name: "exact_match,flexible-extract"
+    value: 0.64
+limit: 1319
+num_fewshot: 5

.buildkite/lm-eval-harness/configs/models-large.txt

@@ -3,3 +3,4 @@ Meta-Llama-3-70B-Instruct.yaml
 Mixtral-8x7B-Instruct-v0.1.yaml
 Qwen2-57B-A14-Instruct.yaml
 DeepSeek-V2-Lite-Chat.yaml
+Meta-Llama-3-8B-QQQ.yaml

.buildkite/lm-eval-harness/configs/models-small.txt

@@ -1,10 +1,6 @@
-Meta-Llama-3-8B-Instruct.yaml
-Meta-Llama-3-8B-Instruct-FP8-compressed-tensors.yaml
+Qwen2.5-1.5B-Instruct.yaml
 Meta-Llama-3.2-1B-Instruct-INT8-compressed-tensors.yaml
 Meta-Llama-3-8B-Instruct-INT8-compressed-tensors-asym.yaml
 Meta-Llama-3-8B-Instruct-nonuniform-compressed-tensors.yaml
-Meta-Llama-3-8B-Instruct-Channelwise-compressed-tensors.yaml
+Qwen2.5-VL-3B-Instruct-FP8-dynamic.yaml
 Qwen1.5-MoE-W4A16-compressed-tensors.yaml
-Qwen2-1.5B-Instruct-INT8-compressed-tensors.yaml
-Qwen2-1.5B-Instruct-FP8W8.yaml
-Meta-Llama-3-8B-QQQ.yaml

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

@@ -0,0 +1,44 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+from pathlib import Path
+
+import pytest
+
+
+def pytest_addoption(parser):
+    parser.addoption(
+        "--config-list-file",
+        action="store",
+        help="Path to the file listing model config YAMLs (one per line)",
+    )
+    parser.addoption(
+        "--tp-size",
+        action="store",
+        default="1",
+        help="Tensor parallel size to use for evaluation",
+    )
+
+
+@pytest.fixture(scope="session")
+def config_list_file(pytestconfig, config_dir):
+    rel_path = pytestconfig.getoption("--config-list-file")
+    return config_dir / rel_path
+
+
+@pytest.fixture(scope="session")
+def tp_size(pytestconfig):
+    return pytestconfig.getoption("--tp-size")
+
+
+def pytest_generate_tests(metafunc):
+    if "config_filename" in metafunc.fixturenames:
+        rel_path = metafunc.config.getoption("--config-list-file")
+        config_list_file = Path(rel_path).resolve()
+        config_dir = config_list_file.parent
+        with open(config_list_file, encoding="utf-8") as f:
+            configs = [
+                config_dir / line.strip()
+                for line in f
+                if line.strip() and not line.startswith("#")
+            ]
+        metafunc.parametrize("config_filename", configs)

.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/run-tests.sh

@@ -1,59 +0,0 @@
-#!/bin/bash
-
-usage() {
-    echo``
-    echo "Runs lm eval harness on GSM8k using vllm and compares to "
-    echo "precomputed baseline (measured by HF transformers.)"
-    echo
-    echo "usage: ${0} <options>"
-    echo
-    echo "  -c    - path to the test data config (e.g. configs/small-models.txt)"
-    echo "  -t    - tensor parallel size"
-    echo
-}
-
-SUCCESS=0
-
-while getopts "c:t:" OPT; do
-  case ${OPT} in
-    c ) 
-        CONFIG="$OPTARG"
-        ;;
-    t )
-        TP_SIZE="$OPTARG"
-        ;;
-    \? )
-        usage
-        exit 1
-        ;;
-  esac
-done
-
-# Parse list of configs.
-IFS=$'\n' read -d '' -r -a MODEL_CONFIGS < "$CONFIG"
-
-for MODEL_CONFIG in "${MODEL_CONFIGS[@]}"
-do
-    LOCAL_SUCCESS=0
-    
-    echo "=== RUNNING MODEL: $MODEL_CONFIG WITH TP SIZE: $TP_SIZE==="
-
-    export LM_EVAL_TEST_DATA_FILE=$PWD/configs/${MODEL_CONFIG}
-    export LM_EVAL_TP_SIZE=$TP_SIZE
-    pytest -s test_lm_eval_correctness.py || LOCAL_SUCCESS=$?
-
-    if [[ $LOCAL_SUCCESS == 0 ]]; then
-        echo "=== PASSED MODEL: ${MODEL_CONFIG} ==="
-    else
-        echo "=== FAILED MODEL: ${MODEL_CONFIG} ==="
-    fi
-
-    SUCCESS=$((SUCCESS + LOCAL_SUCCESS))
-
-done
-
-if [ "${SUCCESS}" -eq "0" ]; then
-    exit 0
-else
-    exit 1
-fi

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

@@ -1,69 +1,57 @@
 # 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
 
-* export LM_EVAL_TEST_DATA_FILE=configs/Meta-Llama-3-70B-Instruct.yaml
-* export LM_EVAL_TP_SIZE=4 
-* pytest -s test_lm_eval_correctness.py
+pytest -s -v test_lm_eval_correctness.py \
+    --config-list-file=configs/models-small.txt \
+    --tp-size=1
 """
 
-import os
-from pathlib import Path
-
 import lm_eval
-import numpy
-import pytest
+import numpy as np
 import yaml
 
 RTOL = 0.08
-TEST_DATA_FILE = os.environ.get(
-    "LM_EVAL_TEST_DATA_FILE",
-    ".buildkite/lm-eval-harness/configs/Meta-Llama-3-8B-Instruct.yaml")
-
-TP_SIZE = os.environ.get("LM_EVAL_TP_SIZE", 1)
 
 
-def launch_lm_eval(eval_config):
-    trust_remote_code = eval_config.get('trust_remote_code', False)
-
-    model_args = f"pretrained={eval_config['model_name']}," \
-                 f"tensor_parallel_size={TP_SIZE}," \
-                 f"add_bos_token=true," \
-                 f"trust_remote_code={trust_remote_code}"
-
+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"max_model_len={max_model_len}"
+    )
     results = lm_eval.simple_evaluate(
         model="vllm",
         model_args=model_args,
         tasks=[task["name"] for task in eval_config["tasks"]],
         num_fewshot=eval_config["num_fewshot"],
         limit=eval_config["limit"],
-        batch_size="auto")
-
+        batch_size="auto",
+    )
     return results
 
 
-def test_lm_eval_correctness():
-    eval_config = yaml.safe_load(
-        Path(TEST_DATA_FILE).read_text(encoding="utf-8"))
+def test_lm_eval_correctness_param(config_filename, tp_size):
+    eval_config = yaml.safe_load(config_filename.read_text(encoding="utf-8"))
 
-    if eval_config[
-            "model_name"] == "nm-testing/Meta-Llama-3-70B-Instruct-FBGEMM-nonuniform":  #noqa: E501
-        pytest.skip("FBGEMM is currently failing on main.")
+    results = launch_lm_eval(eval_config, tp_size)
 
-    # Launch eval requests.
-    results = launch_lm_eval(eval_config)
-
-    # Confirm scores match ground truth.
     success = True
     for task in eval_config["tasks"]:
         for metric in task["metrics"]:
             ground_truth = metric["value"]
             measured_value = results["results"][task["name"]][metric["name"]]
-            print(f'{task["name"]} | {metric["name"]}: '
-                  f'ground_truth={ground_truth} | measured={measured_value}')
-            success = success and numpy.isclose(
-                ground_truth, measured_value, rtol=RTOL)
+            print(
+                f"{task['name']} | {metric['name']}: "
+                f"ground_truth={ground_truth} | measured={measured_value}"
+            )
+            success = success and np.isclose(ground_truth, measured_value, rtol=RTOL)
 
-    # Assert at the end, print all scores even on failure for debugging.
     assert success

.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`:
@@ -113,12 +127,36 @@ WARNING: The benchmarking script will save json results by itself, so please do
 
 ### Visualizing the results
 
-The `convert-results-json-to-markdown.py` helps you put the benchmarking results inside a markdown table, by formatting [descriptions.md](tests/descriptions.md) with real benchmarking results.
+The `convert-results-json-to-markdown.py` helps you put the benchmarking results inside a markdown table, by formatting [descriptions.md](performance-benchmarks-descriptions.md) with real benchmarking results.
 You can find the result presented as a table inside the `buildkite/performance-benchmark` job page.
 If you do not see the table, please wait till the benchmark finish running.
 The json version of the table (together with the json version of the benchmark) will be also attached to the markdown file.
 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)",
@@ -65,18 +70,32 @@ def read_markdown(file):
 
 
 def results_to_json(latency, throughput, serving):
-    return json.dumps({
-        'latency': latency.to_dict(),
-        'throughput': throughput.to_dict(),
-        'serving': serving.to_dict()
-    })
+    return json.dumps(
+        {
+            "latency": latency.to_dict(),
+            "throughput": throughput.to_dict(),
+            "serving": serving.to_dict(),
+        }
+    )
+
+
+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"):
-
         with open(test_file) as f:
             raw_result = json.loads(f.read())
 
@@ -120,7 +139,8 @@ if __name__ == "__main__":
             for perc in [10, 25, 50, 75, 90, 99]:
                 # Multiply 1000 to convert the time unit from s to ms
                 raw_result.update(
-                    {f"P{perc}": 1000 * raw_result["percentiles"][str(perc)]})
+                    {f"P{perc}": 1000 * raw_result["percentiles"][str(perc)]}
+                )
             raw_result["avg_latency"] = raw_result["avg_latency"] * 1000
 
             # add the result to raw_result
@@ -153,26 +173,48 @@ if __name__ == "__main__":
     serving_results = pd.DataFrame.from_dict(serving_results)
     throughput_results = pd.DataFrame.from_dict(throughput_results)
 
-    raw_results_json = results_to_json(latency_results, throughput_results,
-                                       serving_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
+    )
 
     # remapping the key, for visualization purpose
     if not latency_results.empty:
-        latency_results = latency_results[list(
-            latency_column_mapping.keys())].rename(
-                columns=latency_column_mapping)
+        latency_results = latency_results[list(latency_column_mapping.keys())].rename(
+            columns=latency_column_mapping
+        )
     if not serving_results.empty:
-        serving_results = serving_results[list(
-            serving_column_mapping.keys())].rename(
-                columns=serving_column_mapping)
+        serving_results = serving_results[list(serving_column_mapping.keys())].rename(
+            columns=serving_column_mapping
+        )
     if not throughput_results.empty:
-        throughput_results = throughput_results[list(
-            throughput_results_column_mapping.keys())].rename(
-                columns=throughput_results_column_mapping)
+        throughput_results = throughput_results[
+            list(throughput_results_column_mapping.keys())
+        ].rename(columns=throughput_results_column_mapping)
 
-    processed_results_json = results_to_json(latency_results,
-                                             throughput_results,
-                                             serving_results)
+    processed_results_json = results_to_json(
+        latency_results, throughput_results, serving_results
+    )
 
     for df in [latency_results, serving_results, throughput_results]:
         if df.empty:
@@ -184,38 +226,43 @@ if __name__ == "__main__":
         # The GPUs sometimes come in format of "GPUTYPE\nGPUTYPE\n...",
         # we want to turn it into "8xGPUTYPE"
         df["GPU"] = df["GPU"].apply(
-            lambda x: f"{len(x.split('\n'))}x{x.split('\n')[0]}")
+            lambda x: f"{len(x.split('\n'))}x{x.split('\n')[0]}"
+        )
 
     # get markdown tables
-    latency_md_table = tabulate(latency_results,
-                                headers='keys',
-                                tablefmt='pipe',
-                                showindex=False)
-    serving_md_table = tabulate(serving_results,
-                                headers='keys',
-                                tablefmt='pipe',
-                                showindex=False)
-    throughput_md_table = tabulate(throughput_results,
-                                   headers='keys',
-                                   tablefmt='pipe',
-                                   showindex=False)
+    latency_md_table = tabulate(
+        latency_results, headers="keys", tablefmt="pipe", showindex=False
+    )
+    serving_md_table = tabulate(
+        serving_results, headers="keys", tablefmt="pipe", showindex=False
+    )
+    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:
-
-        results = read_markdown("../.buildkite/nightly-benchmarks/" +
-                                "performance-benchmarks-descriptions.md")
+        results = read_markdown(
+            "../.buildkite/nightly-benchmarks/"
+            + "performance-benchmarks-descriptions.md"
+        )
         results = results.format(
             latency_tests_markdown_table=latency_md_table,
             throughput_tests_markdown_table=throughput_md_table,
             serving_tests_markdown_table=serving_md_table,
-            benchmarking_results_in_json_string=processed_results_json)
+            platform_markdown_table=platform_md_table,
+            benchmarking_results_in_json_string=processed_results_json,
+        )
         f.write(results)
 
     # document benchmarking results in json
     with open(results_folder / "benchmark_results.json", "w") as f:
-
-        results = latency_results.to_dict(
-            orient='records') + throughput_results.to_dict(
-                orient='records') + serving_results.to_dict(orient='records')
+        results = (
+            latency_results.to_dict(orient="records")
+            + throughput_results.to_dict(orient="records")
+            + serving_results.to_dict(orient="records")
+        )
         f.write(json.dumps(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
 
@@ -14,15 +15,12 @@ def main(model, cachedir):
 
 if __name__ == "__main__":
     parser = argparse.ArgumentParser(
-        description="Download and save Hugging Face tokenizer")
-    parser.add_argument("--model",
-                        type=str,
-                        required=True,
-                        help="Name of the model")
-    parser.add_argument("--cachedir",
-                        type=str,
-                        required=True,
-                        help="Directory to save the tokenizer")
+        description="Download and save Hugging Face tokenizer"
+    )
+    parser.add_argument("--model", type=str, required=True, help="Name of the model")
+    parser.add_argument(
+        "--cachedir", type=str, required=True, help="Directory to save the tokenizer"
+    )
 
     args = parser.parse_args()
     main(args.model, args.cachedir)

.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
@@ -11,33 +12,33 @@ from tabulate import tabulate
 
 def parse_arguments():
     parser = argparse.ArgumentParser(
-        description=
-        'Parse command line arguments for summary-nightly-results script.')
-    parser.add_argument('--results-folder',
-                        type=str,
-                        required=True,
-                        help='The folder where the results are stored.')
-    parser.add_argument('--description',
-                        type=str,
-                        required=True,
-                        help='Description of the results.')
+        description="Parse command line arguments for summary-nightly-results script."
+    )
+    parser.add_argument(
+        "--results-folder",
+        type=str,
+        required=True,
+        help="The folder where the results are stored.",
+    )
+    parser.add_argument(
+        "--description", type=str, required=True, help="Description of the results."
+    )
 
     args = parser.parse_args()
     return args
 
 
 def get_perf(df, method, model, metric):
-
     means = []
 
     for qps in [2, 4, 8, 16, "inf"]:
-        target = df['Test name'].str.contains(model)
-        target = target & df['Engine'].str.contains(method)
-        target = target & df['Test name'].str.contains("qps_" + str(qps))
+        target = df["Test name"].str.contains(model)
+        target = target & df["Engine"].str.contains(method)
+        target = target & df["Test name"].str.contains("qps_" + str(qps))
         filtered_df = df[target]
 
         if filtered_df.empty:
-            means.append(0.)
+            means.append(0.0)
         else:
             means.append(filtered_df[metric].values[0])
 
@@ -45,7 +46,6 @@ def get_perf(df, method, model, metric):
 
 
 def get_perf_w_std(df, method, model, metric):
-
     if metric in ["TTFT", "ITL"]:
         mean = get_perf(df, method, model, "Mean " + metric + " (ms)")
         mean = mean.tolist()
@@ -60,7 +60,8 @@ def get_perf_w_std(df, method, model, metric):
     else:
         assert metric == "Tput"
         mean = get_perf(df, method, model, "Input Tput (tok/s)") + get_perf(
-            df, method, model, "Output Tput (tok/s)")
+            df, method, model, "Output Tput (tok/s)"
+        )
         mean = mean.tolist()
         std = None
 
@@ -80,18 +81,17 @@ def main(args):
     # generate markdown table
     df = pd.DataFrame.from_dict(results)
 
-    md_table = tabulate(df, headers='keys', tablefmt='pipe', showindex=False)
+    md_table = tabulate(df, headers="keys", tablefmt="pipe", showindex=False)
 
     with open(args.description) as f:
         description = f.read()
 
-    description = description.format(
-        nightly_results_benchmarking_table=md_table)
+    description = description.format(nightly_results_benchmarking_table=md_table)
 
     with open("nightly_results.md", "w") as f:
         f.write(description)
 
 
-if __name__ == '__main__':
+if __name__ == "__main__":
     args = parse_arguments()
     main(args)

.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
@@ -34,10 +35,8 @@ serving_column_mapping = {
 }
 
 if __name__ == "__main__":
-
     # collect results
     for test_file in results_folder.glob("*.json"):
-
         with open(test_file) as f:
             raw_result = json.loads(f.read())
 
@@ -56,17 +55,16 @@ if __name__ == "__main__":
     serving_results = pd.DataFrame.from_dict(serving_results)
 
     if not serving_results.empty:
-        serving_results = serving_results[list(
-            serving_column_mapping.keys())].rename(
-                columns=serving_column_mapping)
+        serving_results = serving_results[list(serving_column_mapping.keys())].rename(
+            columns=serving_column_mapping
+        )
 
-    serving_md_table_with_headers = tabulate(serving_results,
-                                             headers='keys',
-                                             tablefmt='pipe',
-                                             showindex=False)
+    serving_md_table_with_headers = tabulate(
+        serving_results, headers="keys", tablefmt="pipe", showindex=False
+    )
     # remove the first line of header
-    serving_md_table_lines = serving_md_table_with_headers.split('\n')
-    serving_md_table_without_header = '\n'.join(serving_md_table_lines[2:])
+    serving_md_table_lines = serving_md_table_with_headers.split("\n")
+    serving_md_table_without_header = "\n".join(serving_md_table_lines[2:])
 
     prefix = datetime.datetime.now().strftime("%Y-%m-%d_%H-%M-%S")
     prefix = prefix + "_" + os.environ.get("CURRENT_LLM_SERVING_ENGINE")
@@ -76,10 +74,9 @@ if __name__ == "__main__":
         # document results with header.
         # for those who wants to reproduce our benchmark.
         f.write(serving_md_table_with_headers)
-        f.write('\n')
+        f.write("\n")
 
     # document benchmarking results in json
     with open(results_folder / f"{prefix}_nightly_results.json", "w") as f:
-
-        results = serving_results.to_dict(orient='records')
+        results = serving_results.to_dict(orient="records")
         f.write(json.dumps(results))

.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/pyproject.toml

@@ -0,0 +1,46 @@
+# This local pyproject file is part of the migration from yapf to ruff format.
+# It uses the same core rules as the main pyproject.toml file, but with the
+# following differences:
+# - ruff line length is overridden to 88
+# - deprecated typing ignores (UP006, UP035) have been removed
+
+[tool.ruff]
+line-length = 88
+
+[tool.ruff.lint.per-file-ignores]
+"vllm/third_party/**" = ["ALL"]
+"vllm/version.py" = ["F401"]
+"vllm/_version.py" = ["ALL"]
+
+[tool.ruff.lint]
+select = [
+    # pycodestyle
+    "E",
+    # Pyflakes
+    "F",
+    # pyupgrade
+    "UP",
+    # flake8-bugbear
+    "B",
+    # flake8-simplify
+    "SIM",
+    # isort
+    "I",
+    # flake8-logging-format
+    "G",
+]
+ignore = [
+    # star imports
+    "F405", "F403",
+    # lambda expression assignment
+    "E731",
+    # Loop control variable not used within loop body
+    "B007",
+    # f-string format
+    "UP032",
+    # Can remove once 3.10+ is the minimum Python version
+    "UP007",
+]
+
+[tool.ruff.format]
+docstring-code-format = true

.buildkite/release-pipeline.yaml

@@ -1,20 +1,22 @@
 steps:
-  - label: "Build wheel - CUDA 12.4"
+  - label: "Build wheel - CUDA 12.8"
+    id: build-wheel-cuda-12-8
     agents:
       queue: cpu_queue_postmerge
     commands:
-      - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=12.4.0 --tag vllm-ci:build-image --target build --progress plain -f docker/Dockerfile ."
+      - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=12.8.1 --tag vllm-ci:build-image --target build --progress plain -f docker/Dockerfile ."
       - "mkdir artifacts"
       - "docker run --rm -v $(pwd)/artifacts:/artifacts_host vllm-ci:build-image bash -c 'cp -r dist /artifacts_host && chmod -R a+rw /artifacts_host'"
       - "bash .buildkite/scripts/upload-wheels.sh"
     env:
       DOCKER_BUILDKIT: "1"
 
-  - label: "Build wheel - CUDA 12.1"
+  - label: "Build wheel - CUDA 12.6"
+    id: build-wheel-cuda-12-6
     agents:
       queue: cpu_queue_postmerge
     commands:
-      - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=12.1.0 --tag vllm-ci:build-image --target build --progress plain -f docker/Dockerfile ."
+      - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=12.6.3 --build-arg torch_cuda_arch_list='7.0 7.5 8.0 8.9 9.0+PTX' --tag vllm-ci:build-image --target build --progress plain -f docker/Dockerfile ."
       - "mkdir artifacts"
       - "docker run --rm -v $(pwd)/artifacts:/artifacts_host vllm-ci:build-image bash -c 'cp -r dist /artifacts_host && chmod -R a+rw /artifacts_host'"
       - "bash .buildkite/scripts/upload-wheels.sh"
@@ -28,10 +30,11 @@ 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:
-      - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=11.8.0 --tag vllm-ci:build-image --target build --progress plain -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=11.8.0 --build-arg torch_cuda_arch_list='7.0 7.5 8.0 8.9 9.0+PTX' --tag vllm-ci:build-image --target build --progress plain -f docker/Dockerfile ."
       - "mkdir artifacts"
       - "docker run --rm -v $(pwd)/artifacts:/artifacts_host vllm-ci:build-image bash -c 'cp -r dist /artifacts_host && chmod -R a+rw /artifacts_host'"
       - "bash .buildkite/scripts/upload-wheels.sh"
@@ -44,33 +47,49 @@ 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=12.4.0 --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"
     agents:
       queue: tpu_queue_postmerge
     commands:
+      - "yes | docker system prune -a"
+      - "git fetch --all"
       - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --tag vllm/vllm-tpu:nightly --tag vllm/vllm-tpu:$BUILDKITE_COMMIT --progress plain -f docker/Dockerfile.tpu ."
       - "docker push vllm/vllm-tpu:nightly"
       - "docker push vllm/vllm-tpu:$BUILDKITE_COMMIT"
     plugins:
       - docker-login#v3.0.0:
-          username: vllm
+          username: vllmbot
           password-env: DOCKERHUB_TOKEN
     env:
       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
@@ -82,7 +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:$(buildkite-agent meta-data get release-version)"
     env:
       DOCKER_BUILDKIT: "1"
@@ -98,6 +118,7 @@ steps:
     commands:
       - "aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin public.ecr.aws/q9t5s3a7"
       - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg GIT_REPO_CHECK=1 --tag public.ecr.aws/q9t5s3a7/vllm-neuron-release-repo:$(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

@@ -3,6 +3,9 @@
 # This script runs test inside the corresponding ROCm docker container.
 set -o pipefail
 
+# Export Python path
+export PYTHONPATH=".."
+
 # Print ROCm version
 echo "--- Confirming Clean Initial State"
 while true; do
@@ -74,38 +77,71 @@ HF_MOUNT="/root/.cache/huggingface"
 
 commands=$@
 echo "Commands:$commands"
+
+if [[ $commands == *"pytest -v -s basic_correctness/test_basic_correctness.py"* ]]; then
+  commands=${commands//"pytest -v -s basic_correctness/test_basic_correctness.py"/"VLLM_USE_TRITON_FLASH_ATTN=0 pytest -v -s basic_correctness/test_basic_correctness.py"}
+fi
+
+if [[ $commands == *"pytest -v -s models/test_registry.py"* ]]; then
+  commands=${commands//"pytest -v -s models/test_registry.py"/"pytest -v -s models/test_registry.py -k 'not BambaForCausalLM and not GritLM and not Mamba2ForCausalLM and not Zamba2ForCausalLM'"}
+fi
+
+if [[ $commands == *"VLLM_USE_V1=0 pytest -v -s models/test_initialization.py -k 'not llama4 and not plamo2'"* ]]; then
+  commands=${commands//"VLLM_USE_V1=0 pytest -v -s models/test_initialization.py -k 'not llama4 and not plamo2'"/"VLLM_USE_V1=0 pytest -v -s models/test_initialization.py -k 'not llama4 and not plamo2 and not BambaForCausalLM and not Gemma2ForCausalLM and not Grok1ModelForCausalLM and not Zamba2ForCausalLM and not Gemma2Model and not GritLM'"}
+fi
+
+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 "* ]]; then
+if [[ $commands == *" kernels/core"* ]]; then
   commands="${commands} \
-  --ignore=kernels/test_attention_selector.py \
-  --ignore=kernels/test_blocksparse_attention.py \
-  --ignore=kernels/test_causal_conv1d.py \
-  --ignore=kernels/test_cutlass.py \
-  --ignore=kernels/test_encoder_decoder_attn.py \
-  --ignore=kernels/test_flash_attn.py \
-  --ignore=kernels/test_flashinfer.py \
-  --ignore=kernels/test_int8_quant.py \
-  --ignore=kernels/test_machete_gemm.py \
-  --ignore=kernels/test_mamba_ssm.py \
-  --ignore=kernels/test_marlin_gemm.py \
-  --ignore=kernels/test_moe.py \
-  --ignore=kernels/test_prefix_prefill.py \
-  --ignore=kernels/test_rand.py \
-  --ignore=kernels/test_sampler.py \
-  --ignore=kernels/test_cascade_flash_attn.py \
-  --ignore=kernels/test_mamba_mixer2.py \
-  --ignore=kernels/test_aqlm.py \
-  --ignore=kernels/test_machete_mm.py \
-  --ignore=kernels/test_mha_attn.py \
-  --ignore=kernels/test_block_fp8.py \
-  --ignore=kernels/test_cutlass_moe.py \
-  --ignore=kernels/test_mamba_ssm_ssd.py \
-  --ignore=kernels/test_attention.py \
-  --ignore=kernels/test_block_int8.py \
-  --ignore=kernels/test_fused_quant_layernorm.py \
-  --ignore=kernels/test_int8_kernel.py \
-  --ignore=kernels/test_triton_moe_ptpc_fp8.py \
-  --ignore=kernels/test_permute_cols.py"
+  --ignore=kernels/core/test_fused_quant_layernorm.py \
+  --ignore=kernels/core/test_permute_cols.py"
+fi
+
+if [[ $commands == *" kernels/attention"* ]]; then
+  commands="${commands} \
+  --ignore=kernels/attention/test_attention_selector.py \
+  --ignore=kernels/attention/test_encoder_decoder_attn.py \
+  --ignore=kernels/attention/test_flash_attn.py \
+  --ignore=kernels/attention/test_flashinfer.py \
+  --ignore=kernels/attention/test_prefix_prefill.py \
+  --ignore=kernels/attention/test_cascade_flash_attn.py \
+  --ignore=kernels/attention/test_mha_attn.py \
+  --ignore=kernels/attention/test_lightning_attn.py \
+  --ignore=kernels/attention/test_attention.py"
+fi
+
+if [[ $commands == *" kernels/quantization"* ]]; then
+  commands="${commands} \
+  --ignore=kernels/quantization/test_int8_quant.py \
+  --ignore=kernels/quantization/test_aqlm.py \
+  --ignore=kernels/quantization/test_machete_mm.py \
+  --ignore=kernels/quantization/test_block_fp8.py \
+  --ignore=kernels/quantization/test_block_int8.py \
+  --ignore=kernels/quantization/test_marlin_gemm.py \
+  --ignore=kernels/quantization/test_cutlass_scaled_mm.py \
+  --ignore=kernels/quantization/test_int8_kernel.py"
+fi
+
+if [[ $commands == *" kernels/mamba"* ]]; then
+  commands="${commands} \
+  --ignore=kernels/mamba/test_mamba_mixer2.py \
+  --ignore=kernels/mamba/test_causal_conv1d.py \
+  --ignore=kernels/mamba/test_mamba_ssm_ssd.py"
+fi
+
+if [[ $commands == *" kernels/moe"* ]]; then
+  commands="${commands} \
+  --ignore=kernels/moe/test_moe.py \
+  --ignore=kernels/moe/test_cutlass_moe.py \
+  --ignore=kernels/moe/test_triton_moe_ptpc_fp8.py"
 fi
 
 #ignore certain Entrypoints/openai tests
@@ -147,6 +183,8 @@ fi
 
 
 PARALLEL_JOB_COUNT=8
+MYPYTHONPATH=".."
+
 # check if the command contains shard flag, we will run all shards in parallel because the host have 8 GPUs. 
 if [[ $commands == *"--shard-id="* ]]; then
   # assign job count as the number of shards used   
@@ -167,6 +205,7 @@ if [[ $commands == *"--shard-id="* ]]; then
         -e AWS_SECRET_ACCESS_KEY \
         -v "${HF_CACHE}:${HF_MOUNT}" \
         -e "HF_HOME=${HF_MOUNT}" \
+        -e "PYTHONPATH=${MYPYTHONPATH}" \
         --name "${container_name}_${GPU}" \
         "${image_name}" \
         /bin/bash -c "${commands_gpu}" \
@@ -197,6 +236,7 @@ else
           -e AWS_SECRET_ACCESS_KEY \
           -v "${HF_CACHE}:${HF_MOUNT}" \
           -e "HF_HOME=${HF_MOUNT}" \
+          -e "PYTHONPATH=${MYPYTHONPATH}" \
           --name "${container_name}" \
           "${image_name}" \
           /bin/bash -c "${commands}"

.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
@@ -32,9 +33,12 @@ function cpu_tests() {
     set -e
     pip install pytest pytest-asyncio einops peft Pillow soundfile transformers_stream_generator matplotlib
     pip install sentence-transformers datamodel_code_generator
-    pytest -v -s tests/models/embedding/language/test_cls_models.py::test_classification_models[float-jason9693/Qwen2.5-1.5B-apeach]
-    pytest -v -s tests/models/embedding/language/test_embedding.py::test_models[half-BAAI/bge-base-en-v1.5]
-    pytest -v -s tests/models/encoder_decoder/language -m cpu_model"
+    pytest -v -s tests/models/language/generation/test_bart.py -m cpu_model
+    pytest -v -s tests/models/language/generation/test_common.py::test_models[False-5-32-openai-community/gpt2]
+    pytest -v -s tests/models/language/generation/test_common.py::test_models[False-5-32-facebook/opt-125m]
+    pytest -v -s tests/models/language/generation/test_common.py::test_models[False-5-32-google/gemma-1.1-2b-it]
+    pytest -v -s tests/models/language/pooling/test_classification.py::test_models[float-jason9693/Qwen2.5-1.5B-apeach]
+    pytest -v -s tests/models/language/pooling/test_embedding.py -m cpu_model"
 }
 
 # 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

@@ -2,23 +2,55 @@
 
 # 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
 # override the exit code of the script, so we need to use
-# separate remove_docker_container and remove_docker_container_and_exit
+# separate remove_docker_containers and remove_docker_containers_and_exit
 # functions, while other platforms only need one remove_docker_container
 # function.
 EXITCODE=1
-remove_docker_container() { docker rm -f hpu-test || true; }
-remove_docker_container_and_exit() { remove_docker_container; exit $EXITCODE; }
-trap remove_docker_container_and_exit EXIT
-remove_docker_container
+remove_docker_containers() { docker rm -f hpu-plugin-v1-test || true; }
+trap 'remove_docker_containers; exit $EXITCODE;' EXIT
+remove_docker_containers
+
+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"
 
-# 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
 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

@@ -11,13 +11,14 @@ container_name="neuron_$(tr -dc A-Za-z0-9 < /dev/urandom | head -c 10; echo)"
 HF_CACHE="$(realpath ~)/huggingface"
 mkdir -p "${HF_CACHE}"
 HF_MOUNT="/root/.cache/huggingface"
+HF_TOKEN=$(aws secretsmanager get-secret-value  --secret-id "ci/vllm-neuron/hf-token" --region us-west-2 --query 'SecretString' --output text | jq -r .VLLM_NEURON_CI_HF_TOKEN)
 
 NEURON_COMPILE_CACHE_URL="$(realpath ~)/neuron_compile_cache"
 mkdir -p "${NEURON_COMPILE_CACHE_URL}"
 NEURON_COMPILE_CACHE_MOUNT="/root/.cache/neuron_compile_cache"
 
 # Try building the docker image
-aws ecr get-login-password --region us-west-2 | docker login --username AWS --password-stdin 763104351884.dkr.ecr.us-west-2.amazonaws.com
+aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin public.ecr.aws
 
 # prune old image and containers to save disk space, and only once a day
 # by using a timestamp file in tmp.
@@ -47,8 +48,17 @@ trap remove_docker_container EXIT
 docker run --rm -it --device=/dev/neuron0 --network bridge \
        -v "${HF_CACHE}:${HF_MOUNT}" \
        -e "HF_HOME=${HF_MOUNT}" \
+       -e "HF_TOKEN=${HF_TOKEN}" \
        -v "${NEURON_COMPILE_CACHE_URL}:${NEURON_COMPILE_CACHE_MOUNT}" \
        -e "NEURON_COMPILE_CACHE_URL=${NEURON_COMPILE_CACHE_MOUNT}" \
        --name "${container_name}" \
        ${image_name} \
-       /bin/bash -c "python3 /workspace/vllm/examples/offline_inference/neuron.py && python3 -m pytest /workspace/vllm/tests/neuron/1_core/ -v --capture=tee-sys && python3 -m pytest /workspace/vllm/tests/neuron/2_core/ -v --capture=tee-sys"
+       /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\";
+                python3 -m pytest \$f -v --capture=tee-sys;
+            done
+       "

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

@@ -1,54 +1,187 @@
 #!/bin/bash
 
-set -xue
+set -xu
+
+
+remove_docker_container() { 
+    docker rm -f tpu-test || true; 
+    docker rm -f vllm-tpu || true;
+}
+
+trap remove_docker_container EXIT
+
+# Remove the container that might not be cleaned up in the previous run.
+remove_docker_container
 
 # Build the docker image.
 docker build -f docker/Dockerfile.tpu -t vllm-tpu .
 
 # Set up cleanup.
-remove_docker_container() { docker rm -f tpu-test || true; }
-trap remove_docker_container EXIT
-# Remove the container that might not be cleaned up in the previous run.
-remove_docker_container
+cleanup_docker() {
+  # Get Docker's root directory
+  docker_root=$(docker info -f '{{.DockerRootDir}}')
+  if [ -z "$docker_root" ]; then
+    echo "Failed to determine Docker root directory."
+    exit 1
+  fi
+  echo "Docker root directory: $docker_root"
+  # Check disk usage of the filesystem where Docker's root directory is located
+  disk_usage=$(df "$docker_root" | tail -1 | awk '{print $5}' | sed 's/%//')
+  # Define the threshold
+  threshold=70
+  if [ "$disk_usage" -gt "$threshold" ]; then
+    echo "Disk usage is above $threshold%. Cleaning up Docker images and volumes..."
+    # Remove dangling images (those that are not tagged and not used by any container)
+    docker image prune -f
+    # Remove unused volumes / force the system prune for old images as well.
+    docker volume prune -f && docker system prune --force --filter "until=72h" --all
+    echo "Docker images and volumes cleanup completed."
+  else
+    echo "Disk usage is below $threshold%. No cleanup needed."
+  fi
+}
+cleanup_docker
 
 # For HF_TOKEN.
 source /etc/environment
-# Run a simple end-to-end example.
+
 docker run --privileged --net host --shm-size=16G -it \
     -e "HF_TOKEN=$HF_TOKEN" --name tpu-test \
-    vllm-tpu /bin/bash -c "python3 -m pip install git+https://github.com/thuml/depyf.git \
-    && python3 -m pip install pytest pytest-asyncio tpu-info \
-    && python3 -m pip install lm_eval[api]==0.4.4 \
-    && export VLLM_XLA_CACHE_PATH= \
-    && export VLLM_USE_V1=1 \
-    && export VLLM_XLA_CHECK_RECOMPILATION=1 \
-    && echo HARDWARE \
-    && tpu-info \
-    && echo TEST_0 \
-    && pytest -v -s /workspace/vllm/tests/v1/tpu/test_perf.py \
-    && echo TEST_1 \
-    && pytest -v -s /workspace/vllm/tests/tpu/test_compilation.py \
-    && echo TEST_2 \
-    && pytest -v -s /workspace/vllm/tests/v1/tpu/test_basic.py \
-    && echo TEST_3 \
-    && pytest -v -s /workspace/vllm/tests/entrypoints/llm/test_accuracy.py::test_lm_eval_accuracy_v1_engine \
-    && echo TEST_4 \
-    && pytest -s -v /workspace/vllm/tests/tpu/test_quantization_accuracy.py \
-    && echo TEST_5 \
-    && python3 /workspace/vllm/examples/offline_inference/tpu.py \
-    && echo TEST_6 \
-    && pytest -s -v /workspace/vllm/tests/v1/tpu/worker/test_tpu_model_runner.py \
-    && echo TEST_7 \
-    && pytest -s -v /workspace/vllm/tests/v1/tpu/test_sampler.py \
-    && echo TEST_8 \
-    && pytest -s -v /workspace/vllm/tests/v1/tpu/test_topk_topp_sampler.py \
-    && echo TEST_9 \
-    && pytest -s -v /workspace/vllm/tests/v1/tpu/test_multimodal.py \
-    && echo TEST_10 \
-    && pytest -s -v /workspace/vllm/tests/v1/tpu/test_pallas.py \
-    && echo TEST_11 \
-    && pytest -s -v /workspace/vllm/tests/v1/entrypoints/llm/test_struct_output_generate.py" \
+    vllm-tpu /bin/bash -c '
+set -e # Exit immediately if a command exits with a non-zero status.
+set -u # Treat unset variables as an error.
 
+echo "--- Starting script inside Docker container ---"
 
+# Create results directory
+RESULTS_DIR=$(mktemp -d)
+# If mktemp fails, set -e will cause the script to exit.
+echo "Results will be stored in: $RESULTS_DIR"
+
+# Install dependencies
+echo "--- Installing Python dependencies ---"
+python3 -m pip install --progress-bar off git+https://github.com/thuml/depyf.git \
+    && python3 -m pip install --progress-bar off pytest pytest-asyncio tpu-info \
+    && python3 -m pip install --progress-bar off lm_eval[api]==0.4.4
+echo "--- Python dependencies installed ---"
+export VLLM_USE_V1=1
+export VLLM_XLA_CHECK_RECOMPILATION=1
+export VLLM_XLA_CACHE_PATH=
+echo "Using VLLM V1"
+
+echo "--- Hardware Information ---"
+# tpu-info
+echo "--- Starting Tests ---"
+set +e
+overall_script_exit_code=0
+
+# --- Test Definitions ---
+# If a test fails, this function will print logs and will not cause the main script to exit.
+run_test() {
+    local test_num=$1
+    local test_name=$2
+    local test_command=$3
+    local log_file="$RESULTS_DIR/test_${test_num}.log"
+    local actual_exit_code
+
+    echo "--- TEST_$test_num: Running $test_name ---"
+    
+    # Execute the test command.
+    eval "$test_command" > >(tee -a "$log_file") 2> >(tee -a "$log_file" >&2)
+    actual_exit_code=$?
+
+    echo "TEST_${test_num}_COMMAND_EXIT_CODE: $actual_exit_code" # This goes to main log
+    echo "TEST_${test_num}_COMMAND_EXIT_CODE: $actual_exit_code" >> "$log_file" # Also to per-test log
+
+    if [ "$actual_exit_code" -ne 0 ]; then
+        echo "TEST_$test_num ($test_name) FAILED with exit code $actual_exit_code." >&2
+        echo "--- Log for failed TEST_$test_num ($test_name) ---" >&2
+        if [ -f "$log_file" ]; then
+            cat "$log_file" >&2
+        else
+            echo "Log file $log_file not found for TEST_$test_num ($test_name)." >&2
+        fi
+        echo "--- End of log for TEST_$test_num ($test_name) ---" >&2
+        return "$actual_exit_code" # Return the failure code
+    else
+        echo "TEST_$test_num ($test_name) PASSED."
+        return 0 # Return success
+    fi
+}
+
+# Helper function to call run_test and update the overall script exit code
+run_and_track_test() {
+    local test_num_arg="$1"
+    local test_name_arg="$2"
+    local test_command_arg="$3"
+
+    # Run the test
+    run_test "$test_num_arg" "$test_name_arg" "$test_command_arg"
+    local test_specific_exit_code=$?
+
+    # If the test failed, set the overall script exit code to 1
+    if [ "$test_specific_exit_code" -ne 0 ]; then
+        # No need for extra echo here, run_test already logged the failure.
+        overall_script_exit_code=1
+    fi
+}
+
+# --- Actual Test Execution ---
+run_and_track_test 0 "test_perf.py" \
+    "python3 -m pytest -s -v /workspace/vllm/tests/v1/tpu/test_perf.py"
+run_and_track_test 1 "test_compilation.py" \
+    "python3 -m pytest -s -v /workspace/vllm/tests/tpu/test_compilation.py"
+run_and_track_test 2 "test_basic.py" \
+    "python3 -m pytest -s -v /workspace/vllm/tests/v1/tpu/test_basic.py"
+run_and_track_test 3 "test_accuracy.py::test_lm_eval_accuracy_v1_engine" \
+    "python3 -m pytest -s -v /workspace/vllm/tests/entrypoints/llm/test_accuracy.py::test_lm_eval_accuracy_v1_engine"
+run_and_track_test 4 "test_quantization_accuracy.py" \
+    "python3 -m pytest -s -v /workspace/vllm/tests/tpu/test_quantization_accuracy.py"
+run_and_track_test 5 "examples/offline_inference/tpu.py" \
+    "python3 /workspace/vllm/examples/offline_inference/tpu.py"
+run_and_track_test 6 "test_tpu_model_runner.py" \
+    "python3 -m pytest -s -v /workspace/vllm/tests/v1/tpu/worker/test_tpu_model_runner.py"
+run_and_track_test 7 "test_sampler.py" \
+    "python3 -m pytest -s -v /workspace/vllm/tests/v1/tpu/test_sampler.py"
+run_and_track_test 8 "test_topk_topp_sampler.py" \
+    "python3 -m pytest -s -v /workspace/vllm/tests/v1/tpu/test_topk_topp_sampler.py"
+run_and_track_test 9 "test_multimodal.py" \
+    "python3 -m pytest -s -v /workspace/vllm/tests/v1/tpu/test_multimodal.py"
+run_and_track_test 10 "test_pallas.py" \
+    "python3 -m pytest -s -v /workspace/vllm/tests/v1/tpu/test_pallas.py"
+run_and_track_test 11 "test_struct_output_generate.py" \
+    "python3 -m pytest -s -v /workspace/vllm/tests/v1/entrypoints/llm/test_struct_output_generate.py -k \"not test_structured_output_with_reasoning_matrices\""
+run_and_track_test 12 "test_moe_pallas.py" \
+    "python3 -m pytest -s -v /workspace/vllm/tests/tpu/test_moe_pallas.py"
+run_and_track_test 13 "test_lora.py" \
+    "VLLM_XLA_CHECK_RECOMPILATION=0 python3 -m pytest -s -v /workspace/vllm/tests/tpu/lora/test_lora.py"
+run_and_track_test 14 "test_tpu_qkv_linear.py" \
+    "python3 -m pytest -s -v /workspace/vllm/tests/v1/tpu/test_tpu_qkv_linear.py"
+run_and_track_test 15 "test_spmd_model_weight_loading.py" \
+    "python3 -m pytest -s -v /workspace/vllm/tests/v1/tpu/test_spmd_model_weight_loading.py"
+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 [ "$overall_script_exit_code" -ne 0 ]; then
+    echo "--- One or more tests FAILED. Overall script exiting with failure code 1. ---"
+else
+    echo "--- All tests have completed and PASSED. Overall script exiting with success code 0. ---"
+fi
+exit "$overall_script_exit_code"
+' # IMPORTANT: This is the closing single quote for the bash -c "..." command. Ensure it is present and correct.
+
+# Capture the exit code of the docker run command
+DOCKER_RUN_EXIT_CODE=$?
+
+# The trap will run for cleanup.
+# Exit the main script with the Docker run command's exit code.
+if [ "$DOCKER_RUN_EXIT_CODE" -ne 0 ]; then
+    echo "Docker run command failed with exit code $DOCKER_RUN_EXIT_CODE."
+    exit "$DOCKER_RUN_EXIT_CODE"
+else
+    echo "Docker run command completed successfully."
+    exit 0
+fi
 # TODO: This test fails because it uses RANDOM_SEED sampling
-# && VLLM_USE_V1=1 pytest -v -s /workspace/vllm/tests/tpu/test_custom_dispatcher.py \
+# pytest -v -s /workspace/vllm/tests/tpu/test_custom_dispatcher.py \

.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/scripts/upload-wheels.sh

@@ -50,11 +50,11 @@ aws s3 cp "$normal_wheel" "s3://vllm-wheels/$BUILDKITE_COMMIT/"
 if [[ $normal_wheel == *"cu118"* ]]; then
     # if $normal_wheel matches cu118, do not upload the index.html
     echo "Skipping index files for cu118 wheels"
-elif [[ $normal_wheel == *"cu121"* ]]; then
-    # if $normal_wheel matches cu121, do not upload the index.html
-    echo "Skipping index files for cu121 wheels"
+elif [[ $normal_wheel == *"cu126"* ]]; then
+    # if $normal_wheel matches cu126, do not upload the index.html
+    echo "Skipping index files for cu126 wheels"
 else
-    # only upload index.html for cu124 wheels (default wheels)
+    # only upload index.html for cu128 wheels (default wheels)
     aws s3 cp index.html "s3://vllm-wheels/$BUILDKITE_COMMIT/vllm/index.html"
     aws s3 cp "s3://vllm-wheels/nightly/index.html" "s3://vllm-wheels/$BUILDKITE_COMMIT/index.html"
 fi
@@ -66,12 +66,13 @@ aws s3 cp "$normal_wheel" "s3://vllm-wheels/nightly/"
 if [[ $normal_wheel == *"cu118"* ]]; then
     # if $normal_wheel matches cu118, do not upload the index.html
     echo "Skipping index files for cu118 wheels"
-elif [[ $normal_wheel == *"cu121"* ]]; then
-    # if $normal_wheel matches cu121, do not upload the index.html
-    echo "Skipping index files for cu121 wheels"
+elif [[ $normal_wheel == *"cu126"* ]]; then
+    # if $normal_wheel matches cu126, do not upload the index.html
+    echo "Skipping index files for cu126 wheels"
 else
-    # only upload index.html for cu124 wheels (default wheels)
+    # only upload index.html for cu128 wheels (default wheels)
     aws s3 cp index.html "s3://vllm-wheels/nightly/vllm/index.html"
 fi
 
-aws s3 cp "$wheel" "s3://vllm-wheels/$version/"
+aws s3 cp "$wheel" "s3://vllm-wheels/$version/"
+aws s3 cp index.html "s3://vllm-wheels/$version/vllm/index.html"

.buildkite/test-pipeline.yaml

@@ -32,16 +32,27 @@ steps:
 ##### fast check tests  #####
 
 - label: Documentation Build # 2min
-  working_dir: "/vllm-workspace/test_docs/docs"
+  mirror_hardwares: [amdexperimental]
+  working_dir: "/vllm-workspace/test_docs"
   fast_check: true
   no_gpu: True
   commands:
-  - pip install -r ../../requirements/docs.txt
-  - SPHINXOPTS=\"-W\" make html
-  # Check API reference (if it fails, you may have missing mock imports)
-  - grep \"sig sig-object py\" build/html/api/inference_params.html
+  - pip install -r ../requirements/docs.txt
+  # 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:
   - vllm/
   - tests/mq_llm_engine
@@ -57,11 +68,13 @@ steps:
   - pytest -v -s async_engine # AsyncLLMEngine
   - NUM_SCHEDULER_STEPS=4 pytest -v -s async_engine/test_async_llm_engine.py
   - pytest -v -s test_inputs.py
+  - pytest -v -s test_outputs.py
   - pytest -v -s multimodal
   - pytest -v -s test_utils.py # Utils
   - pytest -v -s worker # Worker
 
 - label: Python-only Installation Test
+  mirror_hardwares: [amdexperimental]
   source_file_dependencies:
   - tests/standalone_tests/python_only_compile.sh
   - setup.py
@@ -69,7 +82,7 @@ steps:
   - bash standalone_tests/python_only_compile.sh
 
 - label: Basic Correctness Test # 30min
-  #mirror_hardwares: [amd]
+  mirror_hardwares: [amdexperimental, amdproduction]
   fast_check: true
   torch_nightly: true
   source_file_dependencies:
@@ -86,6 +99,7 @@ steps:
   - VLLM_TEST_ENABLE_ARTIFICIAL_PREEMPT=1 pytest -v -s basic_correctness/test_preemption.py
 
 - label: Chunked Prefill Test
+  mirror_hardwares: [amdexperimental, amdproduction]
   source_file_dependencies:
   - vllm/
   - tests/basic_correctness/test_chunked_prefill
@@ -94,7 +108,7 @@ steps:
   - VLLM_ATTENTION_BACKEND=FLASH_ATTN pytest -v -s basic_correctness/test_chunked_prefill.py
 
 - label: Core Test # 10min
-  mirror_hardwares: [amd]
+  mirror_hardwares: [amdexperimental, amdproduction]
   fast_check: true
   source_file_dependencies:
   - vllm/core
@@ -103,16 +117,14 @@ 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
   torch_nightly: true
-  #mirror_hardwares: [amd]
   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
@@ -121,11 +133,24 @@ 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
-  - pytest -v -s entrypoints/openai --ignore=entrypoints/openai/test_oot_registration.py  --ignore=entrypoints/openai/test_chat_with_tool_reasoning.py --ignore=entrypoints/openai/correctness/ --ignore=entrypoints/openai/test_openai_schema.py
-  - 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: 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
+
 - label: Distributed Tests (4 GPUs) # 10min
+  mirror_hardwares: [amdexperimental]
   working_dir: "/vllm-workspace/tests"
   num_gpus: 4
   source_file_dependencies:
@@ -133,32 +158,55 @@ steps:
   - vllm/core/
   - tests/distributed/test_utils
   - tests/distributed/test_pynccl
-  - tests/spec_decode/e2e/test_integration_dist_tp4
+  - tests/distributed/test_events
   - 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
   - torchrun --nproc-per-node=4 distributed/test_torchrun_example.py
+  # 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 spec_decode/e2e/test_integration_dist_tp4.py
+  - pytest -v -s distributed/test_events.py
   # TODO: create a dedicated test section for multi-GPU example tests
   # when we have multiple distributed example tests
   - pushd ../examples/offline_inference
-  - python3 rlhf.py
-  - RAY_DEDUP_LOGS=0 python3 rlhf_colocate.py
+  - VLLM_ALLOW_INSECURE_SERIALIZATION=1 python3 rlhf.py
+  - 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: [amd]
+  mirror_hardwares: [amdexperimental, amdproduction]
   num_gpus: 2
   source_file_dependencies:
   - vllm/
@@ -166,13 +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: [amd]
+  mirror_hardwares: [amdexperimental]
   source_file_dependencies:
   - vllm/
   - tests/test_regression
@@ -182,7 +235,7 @@ steps:
   working_dir: "/vllm-workspace/tests" # optional
 
 - label: Engine Test # 10min
-  mirror_hardwares: [amd]
+  mirror_hardwares: [amdexperimental]
   source_file_dependencies:
   - vllm/
   - tests/engine
@@ -190,13 +243,14 @@ steps:
   - tests/test_sequence
   - tests/test_config
   - tests/test_logger
+  - tests/test_vllm_port
   commands:
-  - pytest -v -s engine test_sequence.py test_config.py test_logger.py
+  - pytest -v -s engine test_sequence.py test_config.py test_logger.py test_vllm_port.py
   # OOM in the CI unless we run this separately
   - pytest -v -s tokenization
 
 - label: V1 Test
-  #mirror_hardwares: [amd]
+  mirror_hardwares: [amdexperimental]
   source_file_dependencies:
     - vllm/
     - tests/v1
@@ -209,10 +263,12 @@ steps:
     - pytest -v -s v1/worker
     - 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_stats.py
     - pytest -v -s v1/test_utils.py
     - pytest -v -s v1/test_oracle.py
+    - pytest -v -s v1/test_metrics_reader.py
     # TODO: accuracy does not match, whether setting
     # VLLM_USE_FLASHINFER_SAMPLER or not on H100.
     - pytest -v -s v1/e2e
@@ -221,8 +277,8 @@ steps:
     - pytest -v -s entrypoints/openai/correctness/test_lmeval.py::test_lm_eval_accuracy_v1_engine
 
 - label: Examples Test # 25min
+  mirror_hardwares: [amdexperimental]
   working_dir: "/vllm-workspace/examples"
-  #mirror_hardwares: [amd]
   source_file_dependencies:
   - vllm/entrypoints
   - examples/
@@ -235,9 +291,9 @@ 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 other/tensorize_vllm_model.py --model facebook/opt-125m serialize --serialized-directory /tmp/ --suffix v1 && python3 other/tensorize_vllm_model.py --model facebook/opt-125m deserialize --path-to-tensors /tmp/vllm/facebook/opt-125m/v1/model.tensors
+    - 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
     - python3 offline_inference/encoder_decoder_multimodal.py --model-type whisper --seed 0
     - python3 offline_inference/basic/classify.py
@@ -246,14 +302,24 @@ steps:
     - VLLM_USE_V1=0 python3 offline_inference/profiling.py --model facebook/opt-125m run_num_steps --num-steps 2
 
 - label: Prefix Caching Test # 9min
-  mirror_hardwares: [amd]
+  mirror_hardwares: [amdexperimental, amdproduction]
   source_file_dependencies:
   - vllm/
   - tests/prefix_caching
   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:
   - vllm/model_executor/layers
   - vllm/sampling_metadata.py
@@ -263,29 +329,8 @@ steps:
     - pytest -v -s samplers
     - VLLM_USE_FLASHINFER_SAMPLER=1 pytest -v -s samplers
 
-- label: LogitsProcessor Test # 5min
-  mirror_hardwares: [amd]
-  source_file_dependencies:
-  - vllm/model_executor/layers
-  - vllm/model_executor/guided_decoding
-  - tests/test_logits_processor
-  - tests/model_executor/test_guided_processors
-  commands:
-    - pytest -v -s test_logits_processor.py
-    - pytest -v -s model_executor/test_guided_processors.py
-
-- label: Speculative decoding tests # 40min
-  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: [amd]
+  mirror_hardwares: [amdexperimental, amdproduction]
   source_file_dependencies:
   - vllm/lora
   - tests/lora
@@ -293,15 +338,22 @@ steps:
   parallelism: 4
 
 - label: PyTorch Compilation Unit Tests
+  mirror_hardwares: [amdexperimental]
+  torch_nightly: true
   source_file_dependencies:
     - vllm/
     - tests/compile
   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
 
 - label: PyTorch Fullgraph Smoke Test # 9min
+  mirror_hardwares: [amdexperimental, amdproduction]
+  torch_nightly: true
   source_file_dependencies:
   - vllm/
   - tests/compile
@@ -310,8 +362,11 @@ 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]
+  torch_nightly: true
   source_file_dependencies:
   - vllm/
   - tests/compile
@@ -319,6 +374,7 @@ steps:
   - pytest -v -s compile/test_full_graph.py
 
 - label: Kernels Core Operation Test
+  mirror_hardwares: [amdexperimental, amdproduction]
   source_file_dependencies:
   - csrc/
   - tests/kernels/core
@@ -326,6 +382,7 @@ steps:
     - pytest -v -s kernels/core
 
 - label: Kernels Attention Test %N
+  mirror_hardwares: [amdexperimental, amdproduction]
   source_file_dependencies:
   - csrc/attention/
   - vllm/attention
@@ -336,6 +393,7 @@ steps:
   parallelism: 2
 
 - label: Kernels Quantization Test %N
+  mirror_hardwares: [amdexperimental, amdproduction]
   source_file_dependencies:
   - csrc/quantization/
   - vllm/model_executor/layers/quantization
@@ -345,6 +403,7 @@ steps:
   parallelism: 2
 
 - label: Kernels MoE Test
+  mirror_hardwares: [amdexperimental]
   source_file_dependencies:
   - csrc/moe/
   - tests/kernels/moe
@@ -353,6 +412,7 @@ steps:
     - pytest -v -s kernels/moe
 
 - label: Kernels Mamba Test
+  mirror_hardwares: [amdexperimental]
   source_file_dependencies:
   - csrc/mamba/
   - tests/kernels/mamba
@@ -360,48 +420,69 @@ steps:
     - pytest -v -s kernels/mamba
 
 - label: Tensorizer Test # 11min
-  # mirror_hardwares: [amd]
+  mirror_hardwares: [amdexperimental]
   soft_fail: true
   source_file_dependencies:
   - vllm/model_executor/model_loader
   - tests/tensorizer_loader
+  - tests/entrypoints/openai/test_tensorizer_entrypoint.py
   commands:
     - apt-get update && apt-get install -y curl libsodium23
     - export VLLM_WORKER_MULTIPROC_METHOD=spawn
     - pytest -v -s tensorizer_loader
+    - pytest -v -s entrypoints/openai/test_tensorizer_entrypoint.py
+
+- label: Model Executor Test
+  mirror_hardwares: [amdexperimental, amdproduction]
+  soft_fail: true
+  source_file_dependencies:
+  - vllm/model_executor
+  - tests/model_executor
+  commands:
+    - apt-get update && apt-get install -y curl libsodium23
+    - export VLLM_WORKER_MULTIPROC_METHOD=spawn
+    - pytest -v -s model_executor
 
 - label: Benchmarks # 9min
+  mirror_hardwares: [amdexperimental, amdproduction]
   working_dir: "/vllm-workspace/.buildkite"
-  mirror_hardwares: [amd]
   source_file_dependencies:
   - benchmarks/
   commands:
   - bash scripts/run-benchmarks.sh
 
 - label: Benchmarks CLI Test # 10min
+  mirror_hardwares: [amdexperimental, amdproduction]
   source_file_dependencies:
   - vllm/
   - tests/benchmarks/
   commands:
   - pytest -v -s benchmarks/
 
-- label: Quantization Test # 33min
+- label: Quantization Test
+  mirror_hardwares: [amdexperimental]
   source_file_dependencies:
   - csrc/
   - vllm/model_executor/layers/quantization
   - tests/quantization
-  command: VLLM_TEST_FORCE_LOAD_FORMAT=auto pytest -v -s 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
+  mirror_hardwares: [amdexperimental]
   working_dir: "/vllm-workspace/.buildkite/lm-eval-harness"
   source_file_dependencies:
   - csrc/
   - vllm/model_executor/layers/quantization
   commands:
   - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-  - bash ./run-tests.sh -c configs/models-small.txt -t 1
+  - pytest -s -v test_lm_eval_correctness.py --config-list-file=configs/models-small.txt --tp-size=1
 
 - label: OpenAI API correctness
+  mirror_hardwares: [amdexperimental]
   source_file_dependencies:
   - csrc/
   - vllm/entrypoints/openai/
@@ -410,6 +491,7 @@ steps:
   - pytest -s entrypoints/openai/correctness/
 
 - label: Encoder Decoder tests # 5min
+  mirror_hardwares: [amdexperimental]
   source_file_dependencies:
   - vllm/
   - tests/encoder_decoder
@@ -417,8 +499,8 @@ steps:
     - pytest -v -s encoder_decoder
 
 - label: OpenAI-Compatible Tool Use # 20 min
+  mirror_hardwares: [amdexperimental]
   fast_check: false
-  #mirror_hardwares: [ amd ]
   source_file_dependencies:
     - vllm/
     - tests/tool_use
@@ -430,92 +512,115 @@ steps:
 #####  models test  #####
 
 - label: Basic Models Test # 24min
+  mirror_hardwares: [amdexperimental]
+  torch_nightly: true
   source_file_dependencies:
   - vllm/
   - tests/models
   commands:
     - pytest -v -s models/test_transformers.py
     - pytest -v -s models/test_registry.py
-    # V1 Test: https://github.com/vllm-project/vllm/issues/14531
-    - VLLM_USE_V1=0 pytest -v -s models/test_initialization.py -k 'not llama4 and not plamo2'
-    - VLLM_USE_V1=0 pytest -v -s models/test_initialization.py -k 'llama4'
-    - VLLM_USE_V1=0 pytest -v -s models/test_initialization.py -k 'plamo2'
+    - pytest -v -s models/test_utils.py
+    - pytest -v -s models/test_vision.py
+    - pytest -v -s models/test_initialization.py
 
-- label: Language Models Test (Standard) # 32min
-  #mirror_hardwares: [amd]
+- label: Language Models Test (Standard)
+  mirror_hardwares: [amdexperimental]
+  torch_nightly: true
   source_file_dependencies:
   - vllm/
-  - tests/models/decoder_only/language
-  - tests/models/embedding/language
-  - tests/models/encoder_decoder/language
+  - tests/models/language
   commands:
     # Install causal-conv1d for plamo2 models here, as it is not compatible with pip-compile.
-    - pip install causal-conv1d
-    - pytest -v -s models/decoder_only/language -m 'core_model or quant_model'
-    - pytest -v -s models/embedding/language -m core_model
+    - pip install 'git+https://github.com/Dao-AILab/causal-conv1d@v1.5.0.post8'
+    - pip freeze | grep -E 'torch'
+    - pytest -v -s models/language -m core_model
 
-- label: Language Models Test (Extended) # 1h10min
-  optional: true
+- label: Language Models Test (Hybrid) # 35 min
+  mirror_hardwares: [amdexperimental]
+  torch_nightly: true
   source_file_dependencies:
   - vllm/
-  - tests/models/decoder_only/language
-  - tests/models/embedding/language
-  - tests/models/encoder_decoder/language
+  - tests/models/language/generation
   commands:
     # Install causal-conv1d for plamo2 models here, as it is not compatible with pip-compile.
-    - pip install causal-conv1d
-    - pytest -v -s models/decoder_only/language -m 'not core_model and not quant_model'
-    - pytest -v -s models/embedding/language -m 'not core_model'
+    - pip install 'git+https://github.com/Dao-AILab/causal-conv1d@v1.5.0.post8'
+    - pytest -v -s models/language/generation -m hybrid_model
 
-- label: Multi-Modal Models Test (Standard) # 40min
-  #mirror_hardwares: [amd]
-  source_file_dependencies:
-  - vllm/
-  - tests/models/decoder_only/audio_language
-  - tests/models/decoder_only/vision_language
-  - tests/models/embedding/vision_language
-  - tests/models/encoder_decoder/audio_language
-  - tests/models/encoder_decoder/vision_language
-  commands:
-    - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
-    - pytest -v -s models/multimodal
-    - pytest -v -s models/decoder_only/audio_language -m 'core_model or quant_model'
-    - pytest -v -s models/decoder_only/vision_language -m 'core_model or quant_model'
-    - pytest -v -s models/embedding/vision_language -m core_model
-    - pytest -v -s models/encoder_decoder/audio_language -m core_model
-    - pytest -v -s models/encoder_decoder/language -m core_model
-    - pytest -v -s models/encoder_decoder/vision_language -m core_model
-    - pytest -v -s models/decoder_only/vision_language/test_interleaved.py
-
-- label: Multi-Modal Models Test (Extended) 1 # 48m
+- label: Language Models Test (Extended Generation) # 1hr20min
+  mirror_hardwares: [amdexperimental]
   optional: true
   source_file_dependencies:
   - vllm/
-  - tests/models/decoder_only/audio_language
-  - tests/models/decoder_only/vision_language
-  - tests/models/embedding/vision_language
-  - tests/models/encoder_decoder/vision_language
+  - tests/models/language/generation
   commands:
-    - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
-    - pytest -v -s models/decoder_only/audio_language -m 'not core_model and not quant_model'
-    - pytest -v -s models/decoder_only/vision_language/test_models.py -m 'split(group=0) and not core_model and not quant_model'
-    - pytest -v -s --ignore models/decoder_only/vision_language/test_models.py models/decoder_only/vision_language -m 'not core_model and not quant_model'
-    - pytest -v -s models/embedding/vision_language -m 'not core_model'
-    - pytest -v -s models/encoder_decoder/language -m 'not core_model'
-    - pytest -v -s models/encoder_decoder/vision_language -m 'not core_model'
+    # 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) and (not hybrid_model)'
 
-- label: Multi-Modal Models Test (Extended) 2 # 38m
+- label: Language Models Test (Extended Pooling)  # 36min
+  mirror_hardwares: [amdexperimental]
   optional: true
   source_file_dependencies:
   - vllm/
-  - tests/models/decoder_only/vision_language
+  - tests/models/language/pooling
+  commands:
+    - pytest -v -s models/language/pooling -m 'not core_model'
+
+- label: Multi-Modal Models Test (Standard)
+  mirror_hardwares: [amdexperimental]
+  torch_nightly: true
+  source_file_dependencies:
+  - vllm/
+  - tests/models/multimodal
   commands:
     - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
-    - pytest -v -s models/decoder_only/vision_language/test_models.py -m 'split(group=1) and not core_model and not quant_model'
+    - pip freeze | grep -E 'torch'
+    - pytest -v -s models/multimodal/processing
+    - pytest -v -s --ignore models/multimodal/generation/test_whisper.py models/multimodal -m core_model
+    - cd .. && pytest -v -s tests/models/multimodal/generation/test_whisper.py -m core_model  # Otherwise, mp_method="spawn" doesn't work
+
+- label: Multi-Modal Models Test (Extended) 1
+  mirror_hardwares: [amdexperimental]
+  optional: true
+  source_file_dependencies:
+  - vllm/
+  - tests/models/multimodal
+  commands:
+    - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
+    - pytest -v -s --ignore models/multimodal/generation/test_common.py --ignore models/multimodal/processing models/multimodal -m 'not core_model'
+
+- label: Multi-Modal Models Test (Extended) 2
+  mirror_hardwares: [amdexperimental]
+  optional: true
+  source_file_dependencies:
+  - vllm/
+  - tests/models/multimodal
+  commands:
+    - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
+    - 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]
+  optional: true
+  source_file_dependencies:
+  - vllm/
+  - tests/models/multimodal
+  commands:
+    - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
+    - pytest -v -s models/multimodal/generation/test_common.py -m 'split(group=1) and not core_model'
+
+- label: Quantized Models Test
+  mirror_hardwares: [amdexperimental, amdproduction]
+  source_file_dependencies:
+  - vllm/model_executor/layers/quantization
+  - tests/models/quantization
+  commands:
+    - pytest -v -s models/quantization
 
 # This test is used only in PR development phase to test individual models and should never run on main
 - label: Custom Models Test
-  mirror_hardwares: [amd]
+  mirror_hardwares: [amdexperimental, amdproduction]
   optional: true
   commands:
     - echo 'Testing custom models...'
@@ -523,11 +628,23 @@ 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  #####
 
 - label: Distributed Comm Ops Test # 7min
-  mirror_hardwares: [amd]
+  mirror_hardwares: [amdexperimental, amdproduction]
   working_dir: "/vllm-workspace/tests"
   num_gpus: 2
   source_file_dependencies:
@@ -538,6 +655,7 @@ steps:
   - pytest -v -s distributed/test_shm_broadcast.py
 
 - label: 2 Node Tests (4 GPUs in total) # 16min
+  mirror_hardwares: [amdexperimental]
   working_dir: "/vllm-workspace/tests"
   num_gpus: 2
   num_nodes: 2
@@ -547,16 +665,21 @@ 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: [amd]
+  mirror_hardwares: [amdexperimental]
   working_dir: "/vllm-workspace/tests"
   num_gpus: 2
   source_file_dependencies:
@@ -571,9 +694,13 @@ 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
   - pytest -v -s ./compile/test_wrapper.py
@@ -581,26 +708,25 @@ steps:
   - TARGET_TEST_SUITE=L4 pytest basic_correctness/ -v -s -m 'distributed(num_gpus=2)'
   # Avoid importing model tests that cause CUDA reinitialization error
   - pytest models/test_transformers.py -v -s -m 'distributed(num_gpus=2)'
-  - pytest models/encoder_decoder/language/test_bart.py -v -s -m 'distributed(num_gpus=2)'
-  - pytest models/encoder_decoder/vision_language/test_broadcast.py -v -s -m 'distributed(num_gpus=2)'
-  - pytest models/decoder_only/vision_language/test_models.py -v -s -m 'distributed(num_gpus=2)'
+  - pytest models/language -v -s -m 'distributed(num_gpus=2)'
+  - pytest models/multimodal -v -s -m 'distributed(num_gpus=2)'
   # test sequence parallel
   - 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
 
 - label: Plugin Tests (2 GPUs) # 40min
+  mirror_hardwares: [amdexperimental]
   working_dir: "/vllm-workspace/tests"
   num_gpus: 2
   source_file_dependencies:
   - vllm/plugins/
   - tests/plugins/
   commands:
-  # begin platform plugin tests, all the code in-between runs on dummy platform
+  # begin platform plugin and general plugin tests, all the code in-between runs on dummy platform
   - pip install -e ./plugins/vllm_add_dummy_platform
   - pytest -v -s plugins_tests/test_platform_plugins.py
   - pip uninstall vllm_add_dummy_platform -y
@@ -611,8 +737,10 @@ steps:
   - pytest -v -s distributed/test_distributed_oot.py
   - pytest -v -s entrypoints/openai/test_oot_registration.py # it needs a clean process
   - pytest -v -s models/test_oot_registration.py # it needs a clean process
+  - pytest -v -s plugins/lora_resolvers # unit tests for in-tree lora resolver plugins
 
 - label: Multi-step Tests (4 GPUs) # 36min
+  mirror_hardwares: [amdexperimental, amdproduction]
   working_dir: "/vllm-workspace/tests"
   num_gpus: 4
   source_file_dependencies:
@@ -633,6 +761,7 @@ steps:
   - pytest -v -s multi_step/test_correctness_llm.py
 
 - label: Pipeline Parallelism Test # 45min
+  mirror_hardwares: [amdexperimental, amdproduction]
   working_dir: "/vllm-workspace/tests"
   num_gpus: 4
   source_file_dependencies:
@@ -646,6 +775,7 @@ steps:
   - pytest -v -s distributed/test_pipeline_parallel.py
 
 - label: LoRA TP Test (Distributed)
+  mirror_hardwares: [amdexperimental, amdproduction]
   num_gpus: 4
   source_file_dependencies:
   - vllm/lora
@@ -661,6 +791,7 @@ steps:
 
 
 - label: Weight Loading Multiple GPU Test  # 33min
+  mirror_hardwares: [amdexperimental]
   working_dir: "/vllm-workspace/tests"
   num_gpus: 2
   source_file_dependencies:
@@ -670,6 +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]
   working_dir: "/vllm-workspace/tests"
   num_gpus: 2
   gpu: a100
@@ -708,4 +840,4 @@ steps:
   - vllm/model_executor/layers/quantization
   commands:
   - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-  - bash ./run-tests.sh -c configs/models-large.txt -t 4
+  - pytest -s -v test_lm_eval_correctness.py --config-list-file=configs/models-large.txt --tp-size=4

.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

@@ -10,14 +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
-CMakeLists.txt @tlrmchlsmth
+/vllm/lora @jeejeelee
+/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
@@ -26,8 +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
@@ -35,8 +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

.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
@@ -21,12 +31,12 @@ body:
       It is suggested to download and execute the latest script, as vllm might frequently update the diagnosis information needed for accurately and quickly responding to issues.
     value: |
       <details>
-      <summary>The output of `python collect_env.py`</summary>
+      <summary>The output of <code>python collect_env.py</code></summary>
 
       ```text
       Your output of `python collect_env.py` here
       ```
-      
+
       </details>
   validations:
     required: true
@@ -75,20 +85,20 @@ body:
       ```
 
       ```
-      The error message you got, with the full traceback.
+      The error message you got, with the full traceback and the error logs with [dump_input.py:##] if present.
       ```
   validations:
     required: true
 - type: markdown
   attributes:
-    value: >
-      ⚠️ Please separate bugs of `transformers` implementation or usage from bugs of `vllm`. If you think anything is wrong with the models' output:
+    value: |
+      ⚠️ Please separate bugs of `transformers` implementation or usage from bugs of `vllm`. If you think anything is wrong with the model's output:
 
       - Try the counterpart of `transformers` first. If the error appears, please go to [their issues](https://github.com/huggingface/transformers/issues?q=is%3Aissue+is%3Aopen+sort%3Aupdated-desc).
 
       - If the error only appears in vllm, please provide the detailed script of how you run `transformers` and `vllm`, also highlight the difference and what you expect.
 
-      Thanks for contributing 🎉!
+      Thanks for reporting 🙏!
 - type: checkboxes
   id: askllm
   attributes:

.github/ISSUE_TEMPLATE/450-ci-failure.yml

@@ -0,0 +1,69 @@
+name: 🧪 CI failure report
+description: Report a failing test.
+title: "[CI Failure]: "
+labels: ["ci-failure"]
+
+body:
+- type: markdown
+  attributes:
+    value: >
+      #### Include the name of the failing Buildkite step and test file in the title.
+- type: input
+  attributes:
+    label: Name of failing test
+    description: |
+      Paste in the fully-qualified name of the failing test from the logs.
+    placeholder: |
+      `path/to/test_file.py::test_name[params]`
+  validations:
+    required: true
+- type: checkboxes
+  attributes:
+    label: Basic information
+    description: Select all items that apply to the failing test.
+    options:
+      - label: Flaky test
+      - label: Can reproduce locally
+      - label: Caused by external libraries (e.g. bug in `transformers`)
+- type: textarea
+  attributes:
+    label: 🧪 Describe the failing test
+    description: |
+      Please provide a clear and concise description of the failing test.
+    placeholder: |
+      A clear and concise description of the failing test.
+  
+      ```
+      The error message you got, with the full traceback and the error logs with [dump_input.py:##] if present.
+      ```
+  validations:
+    required: true
+- type: textarea
+  attributes:
+    label: 📝 History of failing test
+    description: |
+      Since when did the test start to fail?
+      You can look up its history via [Buildkite Test Suites](https://buildkite.com/organizations/vllm/analytics/suites/ci-1/tests?branch=main).
+
+      If you have time, identify the PR that caused the test to fail on main. You can do so via the following methods:
+
+      - Use Buildkite Test Suites to find the PR where the test failure first occurred, and reproduce the failure locally.
+
+      - Run [`git bisect`](https://git-scm.com/docs/git-bisect) locally.
+
+      - Manually unblock Buildkite steps for suspected PRs on main and check the results. (authorized users only)
+    placeholder: |
+      Approximate timeline and/or problematic PRs
+
+      A link to the Buildkite analytics of the failing test (if available)
+  validations:
+    required: true
+- type: textarea
+  attributes:
+    label: CC List.
+    description: >
+      The list of people you want to CC. Usually, this includes those who worked on the PR that failed the test.
+- type: markdown
+  attributes:
+    value: >
+      Thanks for reporting 🙏!

.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/overview.html>** (anything written below this line will be removed by GitHub Actions)
+**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:
@@ -58,7 +145,7 @@ pull_request_rules:
       - files~=^benchmarks/structured_schemas/
       - files=benchmarks/benchmark_serving_structured_output.py
       - files=benchmarks/run_structured_output_benchmark.sh
-      - files=docs/source/features/structured_outputs.md
+      - files=docs/features/structured_outputs.md
       - files=examples/offline_inference/structured_outputs.py
       - files=examples/online_serving/openai_chat_completion_structured_outputs.py
       - files=examples/online_serving/openai_chat_completion_structured_outputs_with_reasoning.py
@@ -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:
@@ -135,9 +225,7 @@ pull_request_rules:
       - files~=^tests/entrypoints/openai/tool_parsers/
       - files=tests/entrypoints/openai/test_chat_with_tool_reasoning.py
       - files~=^vllm/entrypoints/openai/tool_parsers/
-      - files=docs/source/features/tool_calling.md
-      - files=docs/source/getting_started/examples/openai_chat_completion_client_with_tools.md
-      - files=docs/source/getting_started/examples/chat_with_tools.md
+      - files=docs/features/tool_calling.md
       - files~=^examples/tool_chat_*
       - files=examples/offline_inference/chat_with_tools.py
       - files=examples/online_serving/openai_chat_completion_client_with_tools_required.py
@@ -163,6 +251,17 @@ pull_request_rules:
 
        https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork
 
+- name: assign reviewer for tensorizer changes
+  conditions:
+      - files~=^vllm/model_executor/model_loader/tensorizer.py
+      - files~=^vllm/model_executor/model_loader/tensorizer_loader.py
+      - files~=^tests/entrypoints/openai/test_tensorizer_entrypoint.py
+      - files~=^tests/tensorizer_loader/
+  actions:
+    assign:
+      users:
+        - "sangstar"
+
 - name: remove 'needs-rebase' label when conflict is resolved
   conditions:
       - -conflict

.github/scripts/cleanup_pr_body.sh

@@ -26,7 +26,7 @@ sed -i '/\*\*BEFORE SUBMITTING, PLEASE READ.*\*\*/,$d' "${NEW}"
 
 # Remove HTML <details> section that includes <summary> text of "PR Checklist (Click to Expand)"
 python3 - <<EOF
-import re
+import regex as re
 
 with open("${NEW}", "r") as file:
     content = file.read()

.github/workflows/add_label_automerge.yml

@@ -1,4 +1,6 @@
 name: Add label on auto-merge enabled
+permissions:
+    pull-requests: write
 on:
     pull_request_target:
         types:

.github/workflows/cleanup_pr_body.yml

@@ -20,7 +20,12 @@ jobs:
         with:
           python-version: '3.12'
 
+      - name: Install Python dependencies
+        run: |
+          python3 -m pip install --upgrade pip
+          python3 -m pip install regex
+
       - name: Update PR description
         env:
           GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-        run: .github/scripts/cleanup_pr_body.sh "${{ github.event.number }}"
+        run: bash .github/scripts/cleanup_pr_body.sh "${{ github.event.number }}"

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

@@ -2,6 +2,9 @@ name: Lint and Deploy Charts
 
 on: pull_request
 
+permissions:
+  contents: read
+
 jobs:
   lint-and-deploy:
     runs-on: ubuntu-latest
@@ -65,8 +68,8 @@ 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=$AWS_SECRET_ACCESS_KEY --set resources.requests.cpu=1 --set resources.requests.memory=4Gi --set resources.limits.cpu=2 --set resources.limits.memory=5Gi --set image.env[0].name=VLLM_CPU_KVCACHE_SPACE --set image.env[1].name=VLLM_LOGGING_LEVEL --set image.env[2].name=VLLM_CPU_CI_ENV --set-string image.env[0].value="1" --set-string image.env[1].value="DEBUG" --set-string image.env[2].value="1" --set-string extraInit.s3modelpath="opt-125m/" --set-string 'resources.limits.nvidia\.com/gpu=0' --set-string 'resources.requests.nvidia\.com/gpu=0' --set-string image.repository="vllm-cpu-env"
+
       - name: curl test
         run: |
           kubectl -n ns-vllm port-forward service/test-vllm-service 8001:80 &
@@ -79,4 +82,4 @@ jobs:
                           "max_tokens": 7,
                           "temperature": 0
                   }'):$CODE"
-          echo "$CODE"
+          echo "$CODE"

.github/workflows/pre-commit.yml

@@ -5,6 +5,9 @@ on:
   push:
     branches: [main]
 
+permissions:
+  contents: read
+
 jobs:
   pre-commit:
     runs-on: ubuntu-latest

.github/workflows/reminder_comment.yml

@@ -1,4 +1,6 @@
 name: PR Reminder Comment Bot
+permissions:
+  pull-requests: write
 on:
   pull_request_target:
     types: [opened]

.gitignore

@@ -77,10 +77,6 @@ instance/
 # Scrapy stuff:
 .scrapy
 
-# Sphinx documentation
-docs/_build/
-docs/source/getting_started/examples/
-
 # PyBuilder
 .pybuilder/
 target/
@@ -150,6 +146,8 @@ venv.bak/
 
 # mkdocs documentation
 /site
+docs/argparse
+docs/examples
 
 # mypy
 .mypy_cache/
@@ -203,5 +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

@@ -11,51 +11,59 @@ 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: v0.9.3
+  rev: v0.11.7
   hooks:
   - id: ruff
     args: [--output-format, github, --fix]
-- repo: https://github.com/codespell-project/codespell
-  rev: v2.4.0
+  - id: ruff-format
+    files: ^(.buildkite|benchmarks|examples)/.*
+- 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: 0a0b7a830386ba6a31c2ec8316849ae4d1b8240d # 6.0.0
+  rev: 6.0.1
   hooks:
   - id: isort
 - repo: https://github.com/pre-commit/mirrors-clang-format
-  rev: v19.1.7
+  rev: v20.1.3
   hooks:
   - id: clang-format
     exclude: 'csrc/(moe/topk_softmax_kernels.cu|quantization/gguf/(ggml-common.h|dequantize.cuh|vecdotq.cuh|mmq.cuh|mmvq.cuh))|vllm/third_party/.*'
     types_or: [c++, cuda]
     args: [--style=file, --verbose]
 - repo: https://github.com/jackdewinter/pymarkdown
-  rev: v0.9.27
+  rev: v0.9.29
   hooks:
   - id: pymarkdown
+    exclude: '.*\.inc\.md'
     args: [fix]
 - repo: https://github.com/rhysd/actionlint
   rev: v1.7.7
   hooks:
   - id: actionlint
 - repo: https://github.com/astral-sh/uv-pre-commit
-  rev: 0.6.2
+  rev: 0.6.17
   hooks:
     - id: pip-compile
-      args: [requirements/test.in, -o, requirements/test.txt]
+      args: [requirements/test.in, -o, requirements/test.txt, --index-strategy, unsafe-best-match, --torch-backend, cu128]
       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"
     language: python
     types: [python]
-    additional_dependencies: &mypy_deps [mypy==1.11.1, types-cachetools, types-setuptools, types-PyYAML, types-requests]
+    additional_dependencies: &mypy_deps [mypy==1.11.1, types-cachetools, types-setuptools, types-PyYAML, types-requests, pydantic]
     stages: [pre-commit] # Don't run in CI
   - id: mypy-3.9 # TODO: Use https://github.com/pre-commit/mirrors-mypy when mypy setup is less awkward
     name: Run mypy for Python 3.9
@@ -101,8 +109,8 @@ repos:
     args:
       - -c
       - |
-        if ! grep -q "^Signed-off-by: $(git config user.name) <$(git config user.email)>" .git/COMMIT_EDITMSG; then
-          printf "\nSigned-off-by: $(git config user.name) <$(git config user.email)>\n" >> .git/COMMIT_EDITMSG
+        if ! grep -q "^Signed-off-by: $(git config user.name) <$(git config user.email)>" "$(git rev-parse --git-path COMMIT_EDITMSG)"; then
+          printf "\nSigned-off-by: $(git config user.name) <$(git config user.email)>\n" >> "$(git rev-parse --git-path COMMIT_EDITMSG)"
         fi
     language: system
     verbose: true
@@ -112,6 +120,11 @@ repos:
     entry: python tools/check_spdx_header.py
     language: python
     types: [python]
+  - id: check-root-lazy-imports
+    name: Check root lazy imports
+    entry: python tools/check_init_lazy_imports.py
+    language: python
+    types: [python]
   - id: check-filenames
     name: Check for spaces in all filenames
     entry: bash
@@ -125,12 +138,39 @@ repos:
     name: Update Dockerfile dependency graph
     entry: tools/update-dockerfile-graph.sh
     language: script
-    files: ^docker/Dockerfile$
+  - id: enforce-import-regex-instead-of-re
+    name: Enforce import regex as re
+    entry: python tools/enforce_regex_import.py
+    language: python
+    types: [python]
     pass_filenames: false
+    additional_dependencies: [regex]
+  # forbid directly import triton
+  - id: forbid-direct-triton-import
+    name: "Forbid direct 'import triton'"
+    entry: python tools/check_triton_import.py
+    language: python
+    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

.readthedocs.yaml

@@ -8,12 +8,8 @@ build:
   tools:
     python: "3.12"
 
-sphinx:
-  configuration: docs/source/conf.py
-  fail_on_warning: true
-
-# If using Sphinx, optionally build your docs in additional formats such as PDF
-formats: []
+mkdocs:
+  configuration: mkdocs.yaml
 
 # Optionally declare the Python requirements required to build your docs
 python:

CMakeLists.txt

@@ -15,7 +15,6 @@ project(vllm_extensions LANGUAGES CXX)
 
 # CUDA by default, can be overridden by using -DVLLM_TARGET_DEVICE=... (used by setup.py)
 set(VLLM_TARGET_DEVICE "cuda" CACHE STRING "Target device backend for vLLM")
-
 message(STATUS "Build type: ${CMAKE_BUILD_TYPE}")
 message(STATUS "Target device: ${VLLM_TARGET_DEVICE}")
 
@@ -24,15 +23,15 @@ include(${CMAKE_CURRENT_LIST_DIR}/cmake/utils.cmake)
 # Suppress potential warnings about unused manually-specified variables
 set(ignoreMe "${VLLM_PYTHON_PATH}")
 
+# Prevent installation of dependencies (cutlass) by default.
+install(CODE "set(CMAKE_INSTALL_LOCAL_ONLY TRUE)" ALL_COMPONENTS)
+
 #
 # Supported python versions.  These versions will be searched in order, the
 # first match will be selected.  These should be kept in sync with setup.py.
 #
 set(PYTHON_SUPPORTED_VERSIONS "3.9" "3.10" "3.11" "3.12")
 
-# Supported NVIDIA architectures.
-set(CUDA_SUPPORTED_ARCHS "7.0;7.2;7.5;8.0;8.6;8.7;8.9;9.0;10.0;10.1;12.0")
-
 # Supported AMD GPU architectures.
 set(HIP_SUPPORTED_ARCHS "gfx906;gfx908;gfx90a;gfx942;gfx950;gfx1030;gfx1100;gfx1101;gfx1200;gfx1201")
 
@@ -46,8 +45,8 @@ 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 "2.6.0")
-set(TORCH_SUPPORTED_VERSION_ROCM "2.6.0")
+set(TORCH_SUPPORTED_VERSION_CUDA "2.7.1")
+set(TORCH_SUPPORTED_VERSION_ROCM "2.7.0")
 
 #
 # Try to find python package with an executable that exactly matches
@@ -80,6 +79,15 @@ endif()
 #
 find_package(Torch REQUIRED)
 
+# Supported NVIDIA architectures.
+# This check must happen after find_package(Torch) because that's when CMAKE_CUDA_COMPILER_VERSION gets defined
+if(DEFINED CMAKE_CUDA_COMPILER_VERSION AND
+   CMAKE_CUDA_COMPILER_VERSION VERSION_GREATER_EQUAL 12.8)
+  set(CUDA_SUPPORTED_ARCHS "7.0;7.2;7.5;8.0;8.6;8.7;8.9;9.0;10.0;10.1;12.0")
+else()
+  set(CUDA_SUPPORTED_ARCHS "7.0;7.2;7.5;8.0;8.6;8.7;8.9;9.0")
+endif()
+
 #
 # Forward the non-CUDA device extensions to external CMake scripts.
 #
@@ -163,7 +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.
@@ -174,9 +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
@@ -184,7 +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.
@@ -227,20 +230,24 @@ endif()
 #
 
 set(VLLM_EXT_SRC
+  "csrc/mamba/mamba_ssm/selective_scan_fwd.cu"
   "csrc/cache_kernels.cu"
   "csrc/attention/paged_attention_v1.cu"
   "csrc/attention/paged_attention_v2.cu"
   "csrc/attention/merge_attn_states.cu"
+  "csrc/attention/vertical_slash_index.cu"
   "csrc/pos_encoding_kernels.cu"
   "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"
   "csrc/quantization/fp8/common.cu"
   "csrc/quantization/fused_kernels/fused_layernorm_dynamic_per_token_quant.cu"
   "csrc/quantization/gguf/gguf_kernel.cu"
+  "csrc/quantization/activation_kernels.cu"
   "csrc/cuda_utils_kernels.cu"
   "csrc/prepare_inputs/advance_step.cu"
   "csrc/custom_all_reduce.cu"
@@ -249,9 +256,8 @@ set(VLLM_EXT_SRC
 if(VLLM_GPU_LANG STREQUAL "CUDA")
   SET(CUTLASS_ENABLE_HEADERS_ONLY ON CACHE BOOL "Enable only the header library")
 
-  # Set CUTLASS_REVISION manually -- its revision detection doesn't work in this case.
-  # Please keep this in sync with FetchContent_Declare line below.
-  set(CUTLASS_REVISION "v3.9.0" CACHE STRING "CUTLASS revision to use")
+  # Set CUTLASS_REVISION. Used for FetchContent. Also fixes some bogus messages when building.
+  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})
@@ -269,7 +275,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
         cutlass
         GIT_REPOSITORY https://github.com/nvidia/cutlass.git
         # Please keep this in sync with CUTLASS_REVISION line above.
-        GIT_TAG v3.9.0
+        GIT_TAG ${CUTLASS_REVISION}
         GIT_PROGRESS TRUE
 
         # Speed up CUTLASS download by retrieving only the specified GIT_TAG instead of the history.
@@ -281,14 +287,13 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
   FetchContent_MakeAvailable(cutlass)
 
   list(APPEND VLLM_EXT_SRC
-    "csrc/mamba/mamba_ssm/selective_scan_fwd.cu"
-    "csrc/mamba/causal_conv1d/causal_conv1d.cu"
     "csrc/quantization/aqlm/gemm_kernels.cu"
     "csrc/quantization/awq/gemm_kernels.cu"
     "csrc/permute_cols.cu"
     "csrc/quantization/cutlass_w8a8/scaled_mm_entry.cu"
     "csrc/quantization/fp4/nvfp4_quant_entry.cu"
     "csrc/quantization/fp4/nvfp4_scaled_mm_entry.cu"
+    "csrc/quantization/fp4/nvfp4_blockwise_moe_kernel.cu"
     "csrc/sparse/cutlass/sparse_scaled_mm_entry.cu"
     "csrc/cutlass_extensions/common.cpp"
     "csrc/attention/mla/cutlass_mla_entry.cu")
@@ -300,10 +305,55 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
   # Only build Marlin kernels if we are building for at least some compatible archs.
   # Keep building Marlin for 9.0 as there are some group sizes and shapes that
   # are not supported by Machete yet.
-  cuda_archs_loose_intersection(MARLIN_ARCHS "8.0;8.6;8.7;8.9;9.0;10.0;10.1;12.0" "${CUDA_ARCHS}")
+  # 9.0 for latest bf16 atomicAdd PTX
+  cuda_archs_loose_intersection(MARLIN_ARCHS "8.0;8.7;9.0+PTX" "${CUDA_ARCHS}")
   if (MARLIN_ARCHS)
+
+    #
+    # For the Marlin kernels we automatically generate sources for various
+    # preselected input type pairs and schedules.
+    # Generate sources:
+    set(MARLIN_GEN_SCRIPT
+      ${CMAKE_CURRENT_SOURCE_DIR}/csrc/quantization/gptq_marlin/generate_kernels.py)
+    file(MD5 ${MARLIN_GEN_SCRIPT} MARLIN_GEN_SCRIPT_HASH)
+
+    message(STATUS "Marlin generation script hash: ${MARLIN_GEN_SCRIPT_HASH}")
+    message(STATUS "Last run Marlin generate script hash: $CACHE{MARLIN_GEN_SCRIPT_HASH}")
+
+    if (NOT DEFINED CACHE{MARLIN_GEN_SCRIPT_HASH}
+        OR NOT $CACHE{MARLIN_GEN_SCRIPT_HASH} STREQUAL ${MARLIN_GEN_SCRIPT_HASH})
+      execute_process(
+        COMMAND ${CMAKE_COMMAND} -E env
+        PYTHONPATH=$PYTHONPATH
+          ${Python_EXECUTABLE} ${MARLIN_GEN_SCRIPT}
+        RESULT_VARIABLE marlin_generation_result
+        OUTPUT_VARIABLE marlin_generation_result
+        OUTPUT_FILE ${CMAKE_CURRENT_BINARY_DIR}/marlin_generation.log
+        ERROR_FILE ${CMAKE_CURRENT_BINARY_DIR}/marlin_generation.log
+      )
+
+      if (NOT marlin_generation_result EQUAL 0)
+        message(FATAL_ERROR "Marlin generation failed."
+                            " Result: \"${marlin_generation_result}\""
+                            "\nCheck the log for details: "
+                            "${CMAKE_CURRENT_BINARY_DIR}/marlin_generation.log")
+      else()
+        set(MARLIN_GEN_SCRIPT_HASH ${MARLIN_GEN_SCRIPT_HASH}
+            CACHE STRING "Last run Marlin generate script hash" FORCE)
+        message(STATUS "Marlin generation completed successfully.")
+      endif()
+    else()
+      message(STATUS "Marlin generation script has not changed, skipping generation.")
+    endif()
+
+    file(GLOB MARLIN_TEMPLATE_KERNEL_SRC "csrc/quantization/gptq_marlin/kernel_*.cu")
+    set_gencode_flags_for_srcs(
+      SRCS "${MARLIN_TEMPLATE_KERNEL_SRC}"
+      CUDA_ARCHS "${MARLIN_ARCHS}")
+
+    list(APPEND VLLM_EXT_SRC ${MARLIN_TEMPLATE_KERNEL_SRC})
+
     set(MARLIN_SRCS
-       "csrc/quantization/fp8/fp8_marlin.cu"
        "csrc/quantization/marlin/dense/marlin_cuda_kernel.cu"
        "csrc/quantization/marlin/sparse/marlin_24_cuda_kernel.cu"
        "csrc/quantization/marlin/qqq/marlin_qqq_gemm_kernel.cu"
@@ -341,7 +391,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
   # The cutlass_scaled_mm kernels for Hopper (c3x, i.e. CUTLASS 3.x) require
   # CUDA 12.0 or later
   cuda_archs_loose_intersection(SCALED_MM_ARCHS "9.0a;" "${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"
@@ -357,7 +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 "
@@ -368,13 +418,44 @@ 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(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND SCALED_MM_ARCHS)
     set(SRCS
       "csrc/quantization/cutlass_w8a8/scaled_mm_c3x_sm100.cu"
       "csrc/quantization/cutlass_w8a8/c3x/scaled_mm_sm100_fp8.cu"
+      "csrc/quantization/cutlass_w8a8/c3x/scaled_mm_blockwise_sm100_fp8.cu"
     )
     set_gencode_flags_for_srcs(
       SRCS "${SRCS}"
@@ -385,7 +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 ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND SCALED_MM_ARCHS)
       message(STATUS "Not building scaled_mm_c3x_sm100 as CUDA Compiler version is "
                      "not >= 12.8, we recommend upgrading to CUDA 12.8 or "
                      "later if you intend on running FP8 quantized models on "
@@ -399,8 +480,9 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
   #
   # For the cutlass_scaled_mm kernels we want to build the c2x (CUTLASS 2.x)
   # kernels for the remaining archs that are not already built for 3x.
+  # (Build 8.9 for FP8)
   cuda_archs_loose_intersection(SCALED_MM_2X_ARCHS
-    "7.5;8.0;8.6;8.7;8.9;9.0;10.0;10.1;12.0" "${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 ${SCALED_MM_3X_ARCHS})
   if (SCALED_MM_2X_ARCHS)
@@ -427,7 +509,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
   # The 2:4 sparse kernels cutlass_scaled_sparse_mm and cutlass_compressor
   # require CUDA 12.2 or later (and only work on Hopper).
   cuda_archs_loose_intersection(SCALED_MM_ARCHS "9.0a;" "${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}"
@@ -436,7 +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.")
@@ -448,15 +530,18 @@ 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(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND FP4_ARCHS)
     set(SRCS
       "csrc/quantization/fp4/nvfp4_quant_kernels.cu"
-      "csrc/quantization/fp4/nvfp4_scaled_mm_kernels.cu")
+      "csrc/quantization/fp4/nvfp4_experts_quant.cu"
+      "csrc/quantization/fp4/nvfp4_scaled_mm_kernels.cu"
+      "csrc/quantization/fp4/nvfp4_blockwise_moe_kernel.cu")
     set_gencode_flags_for_srcs(
       SRCS "${SRCS}"
       CUDA_ARCHS "${FP4_ARCHS}")
     list(APPEND VLLM_EXT_SRC "${SRCS}")
     list(APPEND VLLM_GPU_FLAGS "-DENABLE_NVFP4=1")
+    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.")
@@ -466,9 +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}")
@@ -486,13 +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;" "${CUDA_ARCHS}")
+  # The MoE kernel cutlass_moe_mm requires CUDA 12.3 or later (and ONLY works
+  # on Hopper). get_cutlass_(pplx_)moe_mm_data should only be compiled
+  # if it's possible to compile MoE kernels that use its output.
+  cuda_archs_loose_intersection(SCALED_MM_ARCHS "9.0a" "${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 "${SCALED_MM_ARCHS}")
@@ -506,6 +591,46 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
                      "if you intend on running FP8 quantized MoE models on Hopper.")
     else()
       message(STATUS "Not building grouped_mm_c3x as no compatible archs found "
+                     "in CUDA target architectures.")
+    endif()
+  endif()
+
+  # moe_data.cu is used by all CUTLASS MoE kernels.
+  cuda_archs_loose_intersection(CUTLASS_MOE_DATA_ARCHS "9.0a;10.0a" "${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 ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND SCALED_MM_ARCHS)
+      message(STATUS "Not building blockwise_scaled_group_mm_sm100 kernels as CUDA Compiler version is "
+                     "not >= 12.8, we recommend upgrading to CUDA 12.8 or later "
+                     "if you intend on running FP8 quantized MoE models on Blackwell.")
+    else()
+      message(STATUS "Not building blockwise_scaled_group_mm_sm100 as no compatible archs found "
                      "in CUDA target architectures")
     endif()
   endif()
@@ -516,7 +641,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
   # The machete kernels only work on hopper and require CUDA 12.0 or later.
   # Only build Machete kernels if we are building for something compatible with sm90a
   cuda_archs_loose_intersection(MACHETE_ARCHS "9.0a" "${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.
@@ -568,7 +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 "
@@ -582,6 +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
@@ -627,7 +760,8 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
     CUDA_ARCHS "${CUDA_ARCHS}")
 
   list(APPEND VLLM_MOE_EXT_SRC "${VLLM_MOE_WNA16_SRC}")
-  cuda_archs_loose_intersection(MARLIN_MOE_ARCHS "8.0;8.6;8.7;8.9;9.0;10.0;10.1;12.0" "${CUDA_ARCHS}")
+  # 9.0 for latest bf16 atomicAdd PTX
+  cuda_archs_loose_intersection(MARLIN_MOE_ARCHS "8.0;8.7;9.0+PTX" "${CUDA_ARCHS}")
   if (MARLIN_MOE_ARCHS)
 
     #
@@ -645,7 +779,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
         OR NOT $CACHE{MOE_MARLIN_GEN_SCRIPT_HASH} STREQUAL ${MOE_MARLIN_GEN_SCRIPT_HASH})
       execute_process(
         COMMAND ${CMAKE_COMMAND} -E env
-        PYTHONPATH=${CMAKE_CURRENT_SOURCE_DIR}/csrc/cutlass_extensions/:${CUTLASS_DIR}/python/:${VLLM_PYTHON_PATH}:$PYTHONPATH
+        PYTHONPATH=$PYTHONPATH
           ${Python_EXECUTABLE} ${MOE_MARLIN_GEN_SCRIPT}
         RESULT_VARIABLE moe_marlin_generation_result
         OUTPUT_VARIABLE moe_marlin_generation_output
@@ -681,6 +815,17 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
   endif()
 endif()
 
+if(VLLM_GPU_LANG STREQUAL "CUDA")
+  set(MOE_PERMUTE_SRC
+      "csrc/moe/permute_unpermute_kernels/moe_permute_unpermute_kernel.cu"
+      "csrc/moe/moe_permute_unpermute_op.cu")
+
+  set_gencode_flags_for_srcs(
+    SRCS "${MARLIN_PERMUTE_SRC}"
+    CUDA_ARCHS "${MOE_PERMUTE_ARCHS}")
+
+  list(APPEND VLLM_MOE_EXT_SRC "${MOE_PERMUTE_SRC}")
+endif()
 message(STATUS "Enabling moe extension.")
 define_gpu_extension_target(
   _moe_C
@@ -689,6 +834,8 @@ define_gpu_extension_target(
   SOURCES ${VLLM_MOE_EXT_SRC}
   COMPILE_FLAGS ${VLLM_GPU_FLAGS}
   ARCHITECTURES ${VLLM_GPU_ARCHES}
+  INCLUDE_DIRECTORIES ${CUTLASS_INCLUDE_DIR}
+  INCLUDE_DIRECTORIES ${CUTLASS_TOOLS_UTIL_INCLUDE_DIR}
   USE_SABI 3
   WITH_SOABI)
 
@@ -715,5 +862,7 @@ endif()
 # For CUDA we also build and ship some external projects.
 if (VLLM_GPU_LANG STREQUAL "CUDA")
     include(cmake/external_projects/flashmla.cmake)
+
+    # vllm-flash-attn should be last as it overwrites some CMake functions
     include(cmake/external_projects/vllm_flash_attn.cmake)
 endif ()

CONTRIBUTING.md

@@ -1,3 +1,3 @@
 # Contributing to vLLM
 
-You may find information about contributing to vLLM on [docs.vllm.ai](https://docs.vllm.ai/en/latest/contributing/overview.html).
+You may find information about contributing to vLLM on [docs.vllm.ai](https://docs.vllm.ai/en/latest/contributing).

README.md

@@ -1,7 +1,7 @@
 <p align="center">
   <picture>
-    <source media="(prefers-color-scheme: dark)" srcset="https://raw.githubusercontent.com/vllm-project/vllm/main/docs/source/assets/logos/vllm-logo-text-dark.png">
-    <img alt="vLLM" src="https://raw.githubusercontent.com/vllm-project/vllm/main/docs/source/assets/logos/vllm-logo-text-light.png" width=55%>
+    <source media="(prefers-color-scheme: dark)" srcset="https://raw.githubusercontent.com/vllm-project/vllm/main/docs/assets/logos/vllm-logo-text-dark.png">
+    <img alt="vLLM" src="https://raw.githubusercontent.com/vllm-project/vllm/main/docs/assets/logos/vllm-logo-text-light.png" width=55%>
   </picture>
 </p>
 
@@ -16,18 +16,20 @@ Easy, fast, and cheap LLM serving for everyone
 ---
 
 *Latest News* 🔥
+- [2025/05] We hosted [NYC vLLM Meetup](https://lu.ma/c1rqyf1f)! Please find the meetup slides [here](https://docs.google.com/presentation/d/1_q_aW_ioMJWUImf1s1YM-ZhjXz8cUeL0IJvaquOYBeA/edit?usp=sharing).
+- [2025/05] vLLM is now a hosted project under PyTorch Foundation! Please find the announcement [here](https://pytorch.org/blog/pytorch-foundation-welcomes-vllm/).
 - [2025/04] We hosted [Asia Developer Day](https://www.sginnovate.com/event/limited-availability-morning-evening-slots-remaining-inaugural-vllm-asia-developer-day)! Please find the meetup slides from the vLLM team [here](https://docs.google.com/presentation/d/19cp6Qu8u48ihB91A064XfaXruNYiBOUKrBxAmDOllOo/edit?usp=sharing).
-- [2025/03] We hosted [vLLM x Ollama Inference Night](https://lu.ma/vllm-ollama)! Please find the meetup slides from the vLLM team [here](https://docs.google.com/presentation/d/16T2PDD1YwRnZ4Tu8Q5r6n53c5Lr5c73UV9Vd2_eBo4U/edit?usp=sharing).
-- [2025/03] We hosted [the first vLLM China Meetup](https://mp.weixin.qq.com/s/n77GibL2corAtQHtVEAzfg)! Please find the meetup slides from vLLM team [here](https://docs.google.com/presentation/d/1REHvfQMKGnvz6p3Fd23HhSO4c8j5WPGZV0bKYLwnHyQ/edit?usp=sharing).
-- [2025/03] We hosted [the East Coast vLLM Meetup](https://lu.ma/7mu4k4xx)! Please find the meetup slides [here](https://docs.google.com/presentation/d/1NHiv8EUFF1NLd3fEYODm56nDmL26lEeXCaDgyDlTsRs/edit#slide=id.g31441846c39_0_0).
-- [2025/02] We hosted [the ninth vLLM meetup](https://lu.ma/h7g3kuj9) with Meta! Please find the meetup slides from vLLM team [here](https://docs.google.com/presentation/d/1jzC_PZVXrVNSFVCW-V4cFXb6pn7zZ2CyP_Flwo05aqg/edit?usp=sharing) and AMD [here](https://drive.google.com/file/d/1Zk5qEJIkTmlQ2eQcXQZlljAx3m9s7nwn/view?usp=sharing). The slides from Meta will not be posted.
 - [2025/01] We are excited to announce the alpha release of vLLM V1: A major architectural upgrade with 1.7x speedup! Clean code, optimized execution loop, zero-overhead prefix caching, enhanced multimodal support, and more. Please check out our blog post [here](https://blog.vllm.ai/2025/01/27/v1-alpha-release.html).
-- [2025/01] We hosted [the eighth vLLM meetup](https://lu.ma/zep56hui) with Google Cloud! Please find the meetup slides from vLLM team [here](https://docs.google.com/presentation/d/1epVkt4Zu8Jz_S5OhEHPc798emsYh2BwYfRuDDVEF7u4/edit?usp=sharing), and Google Cloud team [here](https://drive.google.com/file/d/1h24pHewANyRL11xy5dXUbvRC9F9Kkjix/view?usp=sharing).
-- [2024/12] vLLM joins [pytorch ecosystem](https://pytorch.org/blog/vllm-joins-pytorch)! Easy, Fast, and Cheap LLM Serving for Everyone!
 
 <details>
 <summary>Previous News</summary>
 
+- [2025/03] We hosted [vLLM x Ollama Inference Night](https://lu.ma/vllm-ollama)! Please find the meetup slides from the vLLM team [here](https://docs.google.com/presentation/d/16T2PDD1YwRnZ4Tu8Q5r6n53c5Lr5c73UV9Vd2_eBo4U/edit?usp=sharing).
+- [2025/03] We hosted [the first vLLM China Meetup](https://mp.weixin.qq.com/s/n77GibL2corAtQHtVEAzfg)! Please find the meetup slides from vLLM team [here](https://docs.google.com/presentation/d/1REHvfQMKGnvz6p3Fd23HhSO4c8j5WPGZV0bKYLwnHyQ/edit?usp=sharing).
+- [2025/03] We hosted [the East Coast vLLM Meetup](https://lu.ma/7mu4k4xx)! Please find the meetup slides [here](https://docs.google.com/presentation/d/1NHiv8EUFF1NLd3fEYODm56nDmL26lEeXCaDgyDlTsRs/edit#slide=id.g31441846c39_0_0).
+- [2025/02] We hosted [the ninth vLLM meetup](https://lu.ma/h7g3kuj9) with Meta! Please find the meetup slides from vLLM team [here](https://docs.google.com/presentation/d/1jzC_PZVXrVNSFVCW-V4cFXb6pn7zZ2CyP_Flwo05aqg/edit?usp=sharing) and AMD [here](https://drive.google.com/file/d/1Zk5qEJIkTmlQ2eQcXQZlljAx3m9s7nwn/view?usp=sharing). The slides from Meta will not be posted.
+- [2025/01] We hosted [the eighth vLLM meetup](https://lu.ma/zep56hui) with Google Cloud! Please find the meetup slides from vLLM team [here](https://docs.google.com/presentation/d/1epVkt4Zu8Jz_S5OhEHPc798emsYh2BwYfRuDDVEF7u4/edit?usp=sharing), and Google Cloud team [here](https://drive.google.com/file/d/1h24pHewANyRL11xy5dXUbvRC9F9Kkjix/view?usp=sharing).
+- [2024/12] vLLM joins [pytorch ecosystem](https://pytorch.org/blog/vllm-joins-pytorch)! Easy, Fast, and Cheap LLM Serving for Everyone!
 - [2024/11] We hosted [the seventh vLLM meetup](https://lu.ma/h0qvrajz) with Snowflake! Please find the meetup slides from vLLM team [here](https://docs.google.com/presentation/d/1e3CxQBV3JsfGp30SwyvS3eM_tW-ghOhJ9PAJGK6KR54/edit?usp=sharing), and Snowflake team [here](https://docs.google.com/presentation/d/1qF3RkDAbOULwz9WK5TOltt2fE9t6uIc_hVNLFAaQX6A/edit?usp=sharing).
 - [2024/10] We have just created a developer slack ([slack.vllm.ai](https://slack.vllm.ai)) focusing on coordinating contributions and discussing features. Please feel free to join us there!
 - [2024/10] Ray Summit 2024 held a special track for vLLM! Please find the opening talk slides from the vLLM team [here](https://docs.google.com/presentation/d/1B_KQxpHBTRa_mDF-tR6i8rWdOU5QoTZNcEg2MKZxEHM/edit?usp=sharing). Learn more from the [talks](https://www.youtube.com/playlist?list=PLzTswPQNepXl6AQwifuwUImLPFRVpksjR) from other vLLM contributors and users!
@@ -56,28 +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), 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
+- 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).
@@ -98,14 +98,14 @@ Visit our [documentation](https://docs.vllm.ai/en/latest/) to learn more.
 ## Contributing
 
 We welcome and value any contributions and collaborations.
-Please check out [Contributing to vLLM](https://docs.vllm.ai/en/stable/contributing/overview.html) for how to get involved.
+Please check out [Contributing to vLLM](https://docs.vllm.ai/en/latest/contributing/index.html) for how to get involved.
 
 ## Sponsors
 
 vLLM is a community project. Our compute resources for development and testing are supported by the following organizations. Thank you for your support!
 
 <!-- Note: Please sort them in alphabetical order. -->
-<!-- Note: Please keep these consistent with docs/source/community/sponsors.md -->
+<!-- Note: Please keep these consistent with docs/community/sponsors.md -->
 Cash Donations:
 - a16z
 - Dropbox
@@ -152,12 +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)

SECURITY.md

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

benchmarks/README.md

@@ -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>
@@ -64,6 +64,12 @@ become available.
       <td style="text-align: center;">✅</td>
       <td><code>lmms-lab/LLaVA-OneVision-Data</code>, <code>Aeala/ShareGPT_Vicuna_unfiltered</code></td>
     </tr>
+    <tr>
+      <td><strong>Custom</strong></td>
+      <td style="text-align: center;">✅</td>
+      <td style="text-align: center;">✅</td>
+      <td>Local file: <code>data.jsonl</code></td>
+    </tr>
   </tbody>
 </table>
 
@@ -76,7 +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
 
@@ -124,7 +133,40 @@ P99 ITL (ms):                            8.39
 ==================================================
 ```
 
-### VisionArena Benchmark for Vision Language Models
+**Custom Dataset**
+
+If the dataset you want to benchmark is not supported yet in vLLM, even then you can benchmark on it using `CustomDataset`. Your data needs to be in `.jsonl` format and needs to have "prompt" field per entry, e.g., data.jsonl
+
+```
+{"prompt": "What is the capital of India?"}
+{"prompt": "What is the capital of Iran?"}
+{"prompt": "What is the capital of China?"}
+``` 
+
+```bash
+# start server
+VLLM_USE_V1=1 vllm serve meta-llama/Llama-3.1-8B-Instruct --disable-log-requests
+```
+
+```bash
+# run benchmarking script
+python3 benchmarks/benchmark_serving.py --port 9001 --save-result --save-detailed \
+  --backend vllm \
+  --model meta-llama/Llama-3.1-8B-Instruct \
+  --endpoint /v1/completions \
+  --dataset-name custom \
+  --dataset-path <path-to-your-data-jsonl> \
+  --custom-skip-chat-template \
+  --num-prompts 80 \
+  --max-concurrency 1 \
+  --temperature=0.3 \
+  --top-p=0.75 \
+  --result-dir "./log/"
+```
+
+You can skip applying chat template if your data already has it by using `--custom-skip-chat-template`.
+
+**VisionArena Benchmark for Vision Language Models**
 
 ```bash
 # need a model with vision capability here
@@ -142,14 +184,13 @@ 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 \
-    --speculative-model "[ngram]" \
-    --ngram_prompt_lookup_min 2 \
-    --ngram-prompt-lookup-max 5 \
-    --num_speculative_tokens 5
+    --speculative-config $'{"method": "ngram",
+    "num_speculative_tokens": 5, "prompt_lookup_max": 5,
+    "prompt_lookup_min": 2}'
 ```
 
 ``` bash
@@ -160,7 +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
@@ -204,7 +245,17 @@ python3 vllm/benchmarks/benchmark_serving.py \
     --seed 42
 ```
 
-### Running With Sampling Parameters
+**`philschmid/mt-bench`**
+
+``` bash
+python3 vllm/benchmarks/benchmark_serving.py \
+    --model Qwen/QwQ-32B \
+    --dataset-name hf \
+    --dataset-path philschmid/mt-bench \
+    --num-prompts 80
+```
+
+**Running With Sampling Parameters**
 
 When using OpenAI-compatible backends such as `vllm`, optional sampling
 parameters can be specified. Example client command:
@@ -222,8 +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 \
@@ -241,7 +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 \
@@ -261,7 +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 \
@@ -274,10 +344,9 @@ python3 vllm/benchmarks/benchmark_throughput.py \
     --output-len=100 \
     --num-prompts=2048 \
     --async-engine \
-    --speculative-model="[ngram]" \
-    --ngram_prompt_lookup_min=2 \
-    --ngram-prompt-lookup-max=5 \
-    --num_speculative_tokens=5
+    --speculative-config $'{"method": "ngram",
+    "num_speculative_tokens": 5, "prompt_lookup_max": 5,
+    "prompt_lookup_min": 2}'
 ```
 
 ```
@@ -286,7 +355,7 @@ Total num prompt tokens:  261136
 Total num output tokens:  204800
 ```
 
-### Other HuggingFaceDataset Examples
+**Other HuggingFaceDataset Examples**
 
 **`lmms-lab/LLaVA-OneVision-Data`**
 
@@ -325,7 +394,7 @@ python3 benchmarks/benchmark_throughput.py \
   --num-prompts 10
 ```
 
-### Benchmark with LoRA Adapters
+**Benchmark with LoRA Adapters**
 
 ``` bash
 # download dataset
@@ -341,3 +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/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
@@ -12,8 +13,7 @@ from typing import Optional, Union
 import aiohttp
 import huggingface_hub.constants
 from tqdm.asyncio import tqdm
-from transformers import (AutoTokenizer, PreTrainedTokenizer,
-                          PreTrainedTokenizerFast)
+from transformers import AutoTokenizer, PreTrainedTokenizer, PreTrainedTokenizerFast
 
 # NOTE(simon): do not import vLLM here so the benchmark script
 # can run without vLLM installed.
@@ -43,8 +43,7 @@ class RequestFuncOutput:
     latency: float = 0.0
     output_tokens: int = 0
     ttft: float = 0.0  # Time to first token
-    itl: list[float] = field(
-        default_factory=list)  # list of inter-token latencies
+    itl: list[float] = field(default_factory=list)  # list of inter-token latencies
     tpot: float = 0.0  # avg next-token latencies
     prompt_len: int = 0
     error: str = ""
@@ -57,8 +56,9 @@ async def async_request_tgi(
     api_url = request_func_input.api_url
     assert api_url.endswith("generate_stream")
 
-    async with aiohttp.ClientSession(trust_env=True,
-                                     timeout=AIOHTTP_TIMEOUT) as session:
+    async with aiohttp.ClientSession(
+        trust_env=True, timeout=AIOHTTP_TIMEOUT
+    ) as session:
         params = {
             "max_new_tokens": request_func_input.output_len,
             "do_sample": True,
@@ -105,8 +105,7 @@ async def async_request_tgi(
 
                         # Decoding phase
                         else:
-                            output.itl.append(timestamp -
-                                              most_recent_timestamp)
+                            output.itl.append(timestamp - most_recent_timestamp)
 
                         most_recent_timestamp = timestamp
 
@@ -133,8 +132,9 @@ async def async_request_trt_llm(
     api_url = request_func_input.api_url
     assert api_url.endswith("generate_stream")
 
-    async with aiohttp.ClientSession(trust_env=True,
-                                     timeout=AIOHTTP_TIMEOUT) as session:
+    async with aiohttp.ClientSession(
+        trust_env=True, timeout=AIOHTTP_TIMEOUT
+    ) as session:
         payload = {
             "accumulate_tokens": True,
             "text_input": request_func_input.prompt,
@@ -159,8 +159,7 @@ async def async_request_trt_llm(
                         if not chunk_bytes:
                             continue
 
-                        chunk = chunk_bytes.decode("utf-8").removeprefix(
-                            "data:")
+                        chunk = chunk_bytes.decode("utf-8").removeprefix("data:")
 
                         data = json.loads(chunk)
                         output.generated_text += data["text_output"]
@@ -172,8 +171,7 @@ async def async_request_trt_llm(
 
                         # Decoding phase
                         else:
-                            output.itl.append(timestamp -
-                                              most_recent_timestamp)
+                            output.itl.append(timestamp - most_recent_timestamp)
 
                         most_recent_timestamp = timestamp
 
@@ -197,15 +195,23 @@ async def async_request_deepspeed_mii(
     request_func_input: RequestFuncInput,
     pbar: Optional[tqdm] = None,
 ) -> RequestFuncOutput:
-    async with aiohttp.ClientSession(trust_env=True,
-                                     timeout=AIOHTTP_TIMEOUT) as session:
+    api_url = request_func_input.api_url
+    assert api_url.endswith(("completions", "profile")), (
+        "OpenAI Completions API URL must end with 'completions' or 'profile'."
+    )
 
+    async with aiohttp.ClientSession(
+        trust_env=True, timeout=AIOHTTP_TIMEOUT
+    ) as session:
         payload = {
+            "model": request_func_input.model,
             "prompt": request_func_input.prompt,
             "max_tokens": request_func_input.output_len,
             "temperature": 0.01,  # deepspeed-mii does not accept 0.0 temp.
             "top_p": 1.0,
         }
+        headers = {"Authorization": f"Bearer {os.environ.get('OPENAI_API_KEY')}"}
+
         output = RequestFuncOutput()
         output.prompt_len = request_func_input.prompt_len
 
@@ -216,19 +222,21 @@ async def async_request_deepspeed_mii(
 
         st = time.perf_counter()
         try:
-            async with session.post(url=request_func_input.api_url,
-                                    json=payload) as response:
+            async with session.post(
+                url=api_url, json=payload, headers=headers
+            ) as response:
                 if response.status == 200:
                     parsed_resp = await response.json()
                     output.latency = time.perf_counter() - st
                     if "choices" in parsed_resp:
-                        output.generated_text = parsed_resp["choices"][0][
-                            "text"]
+                        output.generated_text = parsed_resp["choices"][0]["text"]
                     elif "text" in parsed_resp:
                         output.generated_text = parsed_resp["text"][0]
                     else:
-                        output.error = ("Unexpected response format: "
-                                        "neither 'choices' nor 'text' found")
+                        output.error = (
+                            "Unexpected response format: "
+                            "neither 'choices' nor 'text' found"
+                        )
                         output.success = False
                     output.success = True
                 else:
@@ -249,17 +257,20 @@ async def async_request_openai_completions(
     pbar: Optional[tqdm] = None,
 ) -> RequestFuncOutput:
     api_url = request_func_input.api_url
-    assert api_url.endswith(
-        ("completions", "profile")
-    ), "OpenAI Completions API URL must end with 'completions' or 'profile'."
+    assert api_url.endswith(("completions", "profile")), (
+        "OpenAI Completions API URL must end with 'completions' or 'profile'."
+    )
 
-    async with aiohttp.ClientSession(trust_env=True,
-                                     timeout=AIOHTTP_TIMEOUT) as session:
+    async with aiohttp.ClientSession(
+        trust_env=True, timeout=AIOHTTP_TIMEOUT
+    ) as session:
         payload = {
-            "model": request_func_input.model_name \
-                if request_func_input.model_name else request_func_input.model,
+            "model": request_func_input.model_name
+            if request_func_input.model_name
+            else request_func_input.model,
             "prompt": request_func_input.prompt,
             "temperature": 0.0,
+            "repetition_penalty": 1.0,
             "max_tokens": request_func_input.output_len,
             "logprobs": request_func_input.logprobs,
             "stream": True,
@@ -271,9 +282,7 @@ async def async_request_openai_completions(
             payload["ignore_eos"] = request_func_input.ignore_eos
         if request_func_input.extra_body:
             payload.update(request_func_input.extra_body)
-        headers = {
-            "Authorization": f"Bearer {os.environ.get('OPENAI_API_KEY')}"
-        }
+        headers = {"Authorization": f"Bearer {os.environ.get('OPENAI_API_KEY')}"}
 
         output = RequestFuncOutput()
         output.prompt_len = request_func_input.prompt_len
@@ -282,8 +291,9 @@ async def async_request_openai_completions(
         st = time.perf_counter()
         most_recent_timestamp = st
         try:
-            async with session.post(url=api_url, json=payload,
-                                    headers=headers) as response:
+            async with session.post(
+                url=api_url, json=payload, headers=headers
+            ) as response:
                 if response.status == 200:
                     first_chunk_received = False
                     async for chunk_bytes in response.content:
@@ -291,8 +301,7 @@ async def async_request_openai_completions(
                         if not chunk_bytes:
                             continue
 
-                        chunk = chunk_bytes.decode("utf-8").removeprefix(
-                            "data: ")
+                        chunk = chunk_bytes.decode("utf-8").removeprefix("data: ")
                         if chunk != "[DONE]":
                             data = json.loads(chunk)
 
@@ -312,21 +321,20 @@ async def async_request_openai_completions(
 
                                 # Decoding phase
                                 else:
-                                    output.itl.append(timestamp -
-                                                      most_recent_timestamp)
+                                    output.itl.append(timestamp - most_recent_timestamp)
 
                                 most_recent_timestamp = timestamp
                                 generated_text += text or ""
-                            elif usage := data.get("usage"):
-                                output.output_tokens = usage.get(
-                                    "completion_tokens")
+                            if usage := data.get("usage"):
+                                output.output_tokens = usage.get("completion_tokens")
                     if first_chunk_received:
                         output.success = True
                     else:
                         output.success = False
                         output.error = (
                             "Never received a valid chunk to calculate TTFT."
-                            "This response will be marked as failed!")
+                            "This response will be marked as failed!"
+                        )
                     output.generated_text = generated_text
                     output.latency = most_recent_timestamp - st
                 else:
@@ -347,23 +355,22 @@ async def async_request_openai_chat_completions(
     pbar: Optional[tqdm] = None,
 ) -> RequestFuncOutput:
     api_url = request_func_input.api_url
-    assert api_url.endswith(
-        ("chat/completions", "profile")
-    ), "OpenAI Chat Completions API URL must end with 'chat/completions'."
+    assert api_url.endswith(("chat/completions", "profile")), (
+        "OpenAI Chat Completions API URL must end with 'chat/completions'."
+    )
 
-    async with aiohttp.ClientSession(trust_env=True,
-                                     timeout=AIOHTTP_TIMEOUT) as session:
+    async with aiohttp.ClientSession(
+        trust_env=True, timeout=AIOHTTP_TIMEOUT
+    ) as session:
         content = [{"type": "text", "text": request_func_input.prompt}]
         if request_func_input.multi_modal_content:
             content.append(request_func_input.multi_modal_content)
         payload = {
-            "model": request_func_input.model_name \
-                if request_func_input.model_name else request_func_input.model,
+            "model": request_func_input.model_name
+            if request_func_input.model_name
+            else request_func_input.model,
             "messages": [
-                {
-                    "role": "user",
-                    "content": content
-                },
+                {"role": "user", "content": content},
             ],
             "temperature": 0.0,
             "max_completion_tokens": request_func_input.output_len,
@@ -389,16 +396,22 @@ async def async_request_openai_chat_completions(
         st = time.perf_counter()
         most_recent_timestamp = st
         try:
-            async with session.post(url=api_url, json=payload,
-                                    headers=headers) as response:
+            async with session.post(
+                url=api_url, json=payload, headers=headers
+            ) as response:
                 if response.status == 200:
                     async for chunk_bytes in response.content:
                         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)
@@ -412,13 +425,11 @@ async def async_request_openai_chat_completions(
 
                                 # Decoding phase
                                 else:
-                                    output.itl.append(timestamp -
-                                                      most_recent_timestamp)
+                                    output.itl.append(timestamp - most_recent_timestamp)
 
                                 generated_text += content or ""
                             elif usage := data.get("usage"):
-                                output.output_tokens = usage.get(
-                                    "completion_tokens")
+                                output.output_tokens = usage.get("completion_tokens")
 
                             most_recent_timestamp = timestamp
 
@@ -444,25 +455,28 @@ async def async_request_openai_audio(
 ) -> RequestFuncOutput:
     # Lazy import without PlaceholderModule to avoid vllm dep.
     import soundfile
+
     api_url = request_func_input.api_url
-    assert api_url.endswith(
-        ("transcriptions", "translations"
-         )), "OpenAI Chat Completions API URL must end with 'transcriptions' "
+    assert api_url.endswith(("transcriptions", "translations")), (
+        "OpenAI Chat Completions API URL must end with 'transcriptions' "
+    )
     "or `translations`."
 
-    async with aiohttp.ClientSession(trust_env=True,
-                                     timeout=AIOHTTP_TIMEOUT) as session:
+    async with aiohttp.ClientSession(
+        trust_env=True, timeout=AIOHTTP_TIMEOUT
+    ) as session:
         content = [{"type": "text", "text": request_func_input.prompt}]
         payload = {
-            "model": request_func_input.model_name \
-                if request_func_input.model_name else request_func_input.model,
+            "model": request_func_input.model_name
+            if request_func_input.model_name
+            else request_func_input.model,
             "temperature": 0.0,
             "max_completion_tokens": request_func_input.output_len,
             "stream": True,
             "language": "en",
             # Flattened due to multipart/form-data
             "stream_include_usage": True,
-            "stream_continuous_usage_stats": True
+            "stream_continuous_usage_stats": True,
         }
         if request_func_input.extra_body:
             payload.update(request_func_input.extra_body)
@@ -477,9 +491,9 @@ async def async_request_openai_audio(
             buffer.seek(0)
             return buffer
 
-        with to_bytes(*request_func_input.multi_modal_content['audio']) as f:
+        with to_bytes(*request_func_input.multi_modal_content["audio"]) as f:
             form = aiohttp.FormData()
-            form.add_field('file', f, content_type='audio/wav')
+            form.add_field("file", f, content_type="audio/wav")
             for key, value in payload.items():
                 form.add_field(key, str(value))
 
@@ -491,24 +505,22 @@ async def async_request_openai_audio(
             st = time.perf_counter()
             most_recent_timestamp = st
             try:
-                async with session.post(url=api_url,
-                                        data=form,
-                                        headers=headers) as response:
+                async with session.post(
+                    url=api_url, data=form, headers=headers
+                ) as response:
                     if response.status == 200:
                         async for chunk_bytes in response.content:
                             chunk_bytes = chunk_bytes.strip()
                             if not chunk_bytes:
                                 continue
 
-                            chunk = chunk_bytes.decode("utf-8").removeprefix(
-                                "data: ")
+                            chunk = chunk_bytes.decode("utf-8").removeprefix("data: ")
                             if chunk != "[DONE]":
                                 timestamp = time.perf_counter()
                                 data = json.loads(chunk)
 
                                 if choices := data.get("choices"):
-                                    content = choices[0]["delta"].get(
-                                        "content")
+                                    content = choices[0]["delta"].get("content")
                                     # First token
                                     if ttft == 0.0:
                                         ttft = timestamp - st
@@ -517,12 +529,14 @@ async def async_request_openai_audio(
                                     # Decoding phase
                                     else:
                                         output.itl.append(
-                                            timestamp - most_recent_timestamp)
+                                            timestamp - most_recent_timestamp
+                                        )
 
                                     generated_text += content or ""
                                 elif usage := data.get("usage"):
                                     output.output_tokens = usage.get(
-                                        "completion_tokens")
+                                        "completion_tokens"
+                                    )
 
                                 most_recent_timestamp = timestamp
 
@@ -543,7 +557,7 @@ async def async_request_openai_audio(
 
 
 def get_model(pretrained_model_name_or_path: str) -> str:
-    if os.getenv('VLLM_USE_MODELSCOPE', 'False').lower() == 'true':
+    if os.getenv("VLLM_USE_MODELSCOPE", "False").lower() == "true":
         from modelscope import snapshot_download
 
         from vllm.model_executor.model_loader.weight_utils import get_lock
@@ -554,7 +568,8 @@ def get_model(pretrained_model_name_or_path: str) -> str:
             model_path = snapshot_download(
                 model_id=pretrained_model_name_or_path,
                 local_files_only=huggingface_hub.constants.HF_HUB_OFFLINE,
-                ignore_file_pattern=[".*.pt", ".*.safetensors", ".*.bin"])
+                ignore_file_pattern=[".*.pt", ".*.safetensors", ".*.bin"],
+            )
 
             return model_path
     return pretrained_model_name_or_path
@@ -567,23 +582,23 @@ def get_tokenizer(
     **kwargs,
 ) -> Union[PreTrainedTokenizer, PreTrainedTokenizerFast]:
     if pretrained_model_name_or_path is not None and not os.path.exists(
-            pretrained_model_name_or_path):
-        pretrained_model_name_or_path = get_model(
-            pretrained_model_name_or_path)
+        pretrained_model_name_or_path
+    ):
+        pretrained_model_name_or_path = get_model(pretrained_model_name_or_path)
     if tokenizer_mode == "slow":
         if kwargs.get("use_fast", False):
-            raise ValueError(
-                "Cannot use the fast tokenizer in slow tokenizer mode.")
+            raise ValueError("Cannot use the fast tokenizer in slow tokenizer mode.")
         kwargs["use_fast"] = False
     if tokenizer_mode == "mistral":
         try:
             from vllm.transformers_utils.tokenizer import MistralTokenizer
         except ImportError as e:
-            raise ImportError("MistralTokenizer requires vllm package.\n"
-                              "Please install it with `pip install vllm` "
-                              "to use mistral tokenizer mode.") from e
-        return MistralTokenizer.from_pretrained(
-            str(pretrained_model_name_or_path))
+            raise ImportError(
+                "MistralTokenizer requires vllm package.\n"
+                "Please install it with `pip install vllm` "
+                "to use mistral tokenizer mode."
+            ) from e
+        return MistralTokenizer.from_pretrained(str(pretrained_model_name_or_path))
     else:
         return AutoTokenizer.from_pretrained(
             pretrained_model_name_or_path,
@@ -603,10 +618,11 @@ ASYNC_REQUEST_FUNCS = {
     "tensorrt-llm": async_request_trt_llm,
     "scalellm": async_request_openai_completions,
     "sglang": async_request_openai_completions,
+    "llama.cpp": async_request_openai_completions,
 }
 
 OPENAI_COMPATIBLE_BACKENDS = [
-    k for k, v in ASYNC_REQUEST_FUNCS.items()
-    if v in (async_request_openai_completions,
-             async_request_openai_chat_completions)
+    k
+    for k, v in ASYNC_REQUEST_FUNCS.items()
+    if v in (async_request_openai_completions, async_request_openai_chat_completions)
 ]

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
@@ -9,9 +10,6 @@ generation. Supported dataset types include:
   - BurstGPT
   - HuggingFace
   - VisionArena
-
-TODO: Implement CustomDataset to parse a JSON file and convert its contents into
-SampleRequest instances, similar to the approach used in ShareGPT.
 """
 
 import base64
@@ -35,6 +33,7 @@ from transformers import PreTrainedTokenizerBase
 from vllm.lora.request import LoRARequest
 from vllm.lora.utils import get_adapter_absolute_path
 from vllm.multimodal import MultiModalDataDict
+from vllm.multimodal.image import convert_image_mode
 from vllm.transformers_utils.tokenizer import AnyTokenizer, get_lora_tokenizer
 
 logger = logging.getLogger(__name__)
@@ -82,14 +81,12 @@ class BenchmarkDataset(ABC):
         self.dataset_path = dataset_path
         # Set the random seed, ensuring that a None value is replaced with the
         # default seed.
-        self.random_seed = (random_seed
-                            if random_seed is not None else self.DEFAULT_SEED)
+        self.random_seed = random_seed if random_seed is not None else self.DEFAULT_SEED
         self.data = None
 
     def apply_multimodal_chat_transformation(
-            self,
-            prompt: str,
-            mm_content: Optional[MultiModalDataDict] = None) -> list[dict]:
+        self, prompt: str, mm_content: Optional[MultiModalDataDict] = None
+    ) -> list[dict]:
         """
         Transform a prompt and optional multimodal content into a chat format.
         This method is used for chat models that expect a specific conversation
@@ -111,8 +108,7 @@ class BenchmarkDataset(ABC):
             NotImplementedError: If a subclass does not implement this method.
         """
         # TODO (jenniferzhao): add support for downloading data
-        raise NotImplementedError(
-            "load_data must be implemented in subclasses.")
+        raise NotImplementedError("load_data must be implemented in subclasses.")
 
     def get_random_lora_request(
         self,
@@ -158,8 +154,9 @@ class BenchmarkDataset(ABC):
         return lora_request, lora_tokenizer_cache[lora_id] or tokenizer
 
     @abstractmethod
-    def sample(self, tokenizer: PreTrainedTokenizerBase,
-               num_requests: int) -> list[SampleRequest]:
+    def sample(
+        self, tokenizer: PreTrainedTokenizerBase, num_requests: int
+    ) -> list[SampleRequest]:
         """
         Abstract method to generate sample requests from the dataset.
 
@@ -177,8 +174,9 @@ class BenchmarkDataset(ABC):
         """
         raise NotImplementedError("sample must be implemented in subclasses.")
 
-    def maybe_oversample_requests(self, requests: list[SampleRequest],
-                                  num_requests: int) -> None:
+    def maybe_oversample_requests(
+        self, requests: list[SampleRequest], num_requests: int
+    ) -> None:
         """
         Oversamples the list of requests if its size is less than the desired
         number.
@@ -189,11 +187,9 @@ class BenchmarkDataset(ABC):
         """
         if len(requests) < num_requests:
             random.seed(self.random_seed)
-            additional = random.choices(requests,
-                                        k=num_requests - len(requests))
+            additional = random.choices(requests, k=num_requests - len(requests))
             requests.extend(additional)
-            logger.info("Oversampled requests to reach %d total samples.",
-                        num_requests)
+            logger.info("Oversampled requests to reach %d total samples.", num_requests)
 
 
 # -----------------------------------------------------------------------------
@@ -218,14 +214,14 @@ def is_valid_sequence(
     """
     # Check for invalid conditions
     prompt_too_short = prompt_len < min_len
-    output_too_short = (not skip_min_output_len_check) and (output_len
-                                                            < min_len)
+    output_too_short = (not skip_min_output_len_check) and (output_len < min_len)
     prompt_too_long = prompt_len > max_prompt_len
     combined_too_long = (prompt_len + output_len) > max_total_len
 
     # Return True if none of the invalid conditions are met
-    return not (prompt_too_short or output_too_short or prompt_too_long
-                or combined_too_long)
+    return not (
+        prompt_too_short or output_too_short or prompt_too_long or combined_too_long
+    )
 
 
 @cache
@@ -257,28 +253,28 @@ def process_image(image: Any) -> Mapping[str, Any]:
     Raises:
         ValueError: If the input is not a supported type.
     """
-    if isinstance(image, dict) and 'bytes' in image:
-        image = Image.open(BytesIO(image['bytes']))
+    if isinstance(image, dict) and "bytes" in image:
+        image = Image.open(BytesIO(image["bytes"]))
     if isinstance(image, Image.Image):
-        image = image.convert("RGB")
+        image = convert_image_mode(image, "RGB")
         with io.BytesIO() as image_data:
             image.save(image_data, format="JPEG")
-            image_base64 = base64.b64encode(
-                image_data.getvalue()).decode("utf-8")
+            image_base64 = base64.b64encode(image_data.getvalue()).decode("utf-8")
         return {
             "type": "image_url",
-            "image_url": {
-                "url": f"data:image/jpeg;base64,{image_base64}"
-            },
+            "image_url": {"url": f"data:image/jpeg;base64,{image_base64}"},
         }
 
     if isinstance(image, str):
-        image_url = (image if image.startswith(
-            ("http://", "file://")) else f"file://{image}")
+        image_url = (
+            image if image.startswith(("http://", "file://")) else f"file://{image}"
+        )
         return {"type": "image_url", "image_url": {"url": image_url}}
 
-    raise ValueError(f"Invalid image input {image}. Must be a PIL.Image.Image"
-                     " or str or dictionary with raw image bytes.")
+    raise ValueError(
+        f"Invalid image input {image}. Must be a PIL.Image.Image"
+        " or str or dictionary with raw image bytes."
+    )
 
 
 # -----------------------------------------------------------------------------
@@ -315,42 +311,60 @@ class RandomDataset(BenchmarkDataset):
         )
 
         vocab_size = tokenizer.vocab_size
+        num_special_tokens = tokenizer.num_special_tokens_to_add()
+        real_input_len = input_len - num_special_tokens
 
-        prefix_token_ids = (np.random.randint(
-            0, vocab_size, size=prefix_len).tolist() if prefix_len > 0 else [])
+        prefix_token_ids = (
+            np.random.randint(0, vocab_size, size=prefix_len).tolist()
+            if prefix_len > 0
+            else []
+        )
 
         # New sampling logic: [X * (1 - b), X * (1 + b)]
-        input_low = int(input_len * (1 - range_ratio))
-        input_high = int(input_len * (1 + range_ratio))
+        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
         logger.info("Sampling input_len from [%s, %s]", input_low, input_high)
-        logger.info("Sampling output_len from [%s, %s]", output_low,
-                    output_high)
+        logger.info("Sampling output_len from [%s, %s]", output_low, output_high)
 
-        input_lens = np.random.randint(input_low,
-                                       input_high + 1,
-                                       size=num_requests)
-        output_lens = np.random.randint(output_low,
-                                        output_high + 1,
-                                        size=num_requests)
+        input_lens = np.random.randint(input_low, input_high + 1, size=num_requests)
+        output_lens = np.random.randint(output_low, output_high + 1, size=num_requests)
         offsets = np.random.randint(0, vocab_size, size=num_requests)
 
         requests = []
         for i in range(num_requests):
-            inner_seq = ((offsets[i] + i + np.arange(input_lens[i])) %
-                         vocab_size).tolist()
+            inner_seq = (
+                (offsets[i] + i + np.arange(input_lens[i])) % vocab_size
+            ).tolist()
             token_sequence = prefix_token_ids + inner_seq
             prompt = tokenizer.decode(token_sequence)
+            # After decoding the prompt we have to encode and decode it again.
+            # This is done because in some cases N consecutive tokens
+            # give a string tokenized into != N number of tokens.
+            # For example for GPT2Tokenizer:
+            # [6880, 6881] -> ['Ġcalls', 'here'] ->
+            # [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)[
+                :total_input_len
+            ]
+            prompt = tokenizer.decode(re_encoded_sequence)
+            total_input_len = len(re_encoded_sequence)
             requests.append(
                 SampleRequest(
                     prompt=prompt,
                     prompt_len=total_input_len,
                     expected_output_len=int(output_lens[i]),
-                ))
+                )
+            )
         return requests
 
 
@@ -377,7 +391,8 @@ class ShareGPTDataset(BenchmarkDataset):
             self.data = json.load(f)
         # Filter entries with at least two conversation turns.
         self.data = [
-            entry for entry in self.data
+            entry
+            for entry in self.data
             if "conversations" in entry and len(entry["conversations"]) >= 2
         ]
         random.seed(self.random_seed)
@@ -403,31 +418,123 @@ class ShareGPTDataset(BenchmarkDataset):
             )
 
             lora_request, tokenizer = self.get_random_lora_request(
-                tokenizer=tokenizer, max_loras=max_loras, lora_path=lora_path)
+                tokenizer=tokenizer, max_loras=max_loras, lora_path=lora_path
+            )
             prompt_ids = tokenizer(prompt).input_ids
             completion_ids = tokenizer(completion).input_ids
             prompt_len = len(prompt_ids)
-            new_output_len = (len(completion_ids)
-                              if output_len is None else output_len)
-            if not is_valid_sequence(prompt_len,
-                                     new_output_len,
-                                     skip_min_output_len_check=output_len
-                                     is not None):
+            new_output_len = len(completion_ids) if output_len is None else output_len
+            if not is_valid_sequence(
+                prompt_len,
+                new_output_len,
+                skip_min_output_len_check=output_len is not None,
+            ):
                 continue
             if enable_multimodal_chat:
-                prompt = self.apply_multimodal_chat_transformation(
-                    prompt, None)
+                prompt = self.apply_multimodal_chat_transformation(prompt, None)
             samples.append(
                 SampleRequest(
                     prompt=prompt,
                     prompt_len=prompt_len,
                     expected_output_len=new_output_len,
                     lora_request=lora_request,
-                ))
+                )
+            )
         self.maybe_oversample_requests(samples, num_requests)
         return samples
 
 
+# -----------------------------------------------------------------------------
+# Custom Dataset Implementation
+# -----------------------------------------------------------------------------
+
+
+class CustomDataset(BenchmarkDataset):
+    """
+    Implements the Custom dataset.  Loads data from a JSONL file and generates
+    sample requests based on conversation turns. E.g.,
+    ```
+    {"prompt": "What is the capital of India?"}
+    {"prompt": "What is the capital of Iran?"}
+    {"prompt": "What is the capital of China?"}
+    ```
+    """
+
+    def __init__(self, **kwargs) -> None:
+        super().__init__(**kwargs)
+        self.load_data()
+
+    def load_data(self) -> None:
+        if self.dataset_path is None:
+            raise ValueError("dataset_path must be provided for loading data.")
+
+        # self.data will be a list of dictionaries
+        # e.g., [{"prompt": "What is the capital of India?"}, ...]
+        # This will be the standardized format which load_data()
+        # has to convert into depending on the filetype of dataset_path.
+        # sample() will assume this standardized format of self.data
+        self.data = []
+
+        # Load the JSONL file
+        if self.dataset_path.endswith(".jsonl"):
+            jsonl_data = pd.read_json(path_or_buf=self.dataset_path, lines=True)
+
+            # check if the JSONL file has a 'prompt' column
+            if "prompt" not in jsonl_data.columns:
+                raise ValueError("JSONL file must contain a 'prompt' column.")
+
+            # Convert each row to a dictionary and append to self.data
+            # This will convert the DataFrame to a list of dictionaries
+            # where each dictionary corresponds to a row in the DataFrame.
+            # This is the standardized format we want for self.data
+            for _, row in jsonl_data.iterrows():
+                self.data.append(row.to_dict())
+        else:
+            raise NotImplementedError(
+                "Only JSONL format is supported for CustomDataset."
+            )
+
+        random.seed(self.random_seed)
+        random.shuffle(self.data)
+
+    def sample(
+        self,
+        tokenizer: PreTrainedTokenizerBase,
+        num_requests: int,
+        lora_path: Optional[str] = None,
+        max_loras: Optional[int] = None,
+        output_len: Optional[int] = None,
+        enable_multimodal_chat: bool = False,
+        skip_chat_template: bool = False,
+        **kwargs,
+    ) -> list:
+        sampled_requests = []
+        for item in self.data:
+            if len(sampled_requests) >= num_requests:
+                break
+            prompt = item["prompt"]
+
+            # apply template
+            if not skip_chat_template:
+                prompt = tokenizer.apply_chat_template(
+                    [{"role": "user", "content": prompt}],
+                    add_generation_prompt=True,
+                    tokenize=False,
+                )
+
+            prompt_len = len(tokenizer(prompt).input_ids)
+            sampled_requests.append(
+                SampleRequest(
+                    prompt=prompt,
+                    prompt_len=prompt_len,
+                    expected_output_len=output_len,
+                )
+            )
+        self.maybe_oversample_requests(sampled_requests, num_requests)
+
+        return sampled_requests
+
+
 # -----------------------------------------------------------------------------
 # Sonnet Dataset Implementation
 # -----------------------------------------------------------------------------
@@ -469,20 +576,20 @@ class SonnetDataset(BenchmarkDataset):
     ) -> list:
         # Calculate average token length for a poem line.
         tokenized_lines = [tokenizer(line).input_ids for line in self.data]
-        avg_len = sum(len(tokens)
-                      for tokens in tokenized_lines) / len(tokenized_lines)
+        avg_len = sum(len(tokens) for tokens in tokenized_lines) / len(tokenized_lines)
 
         # Build the base prompt.
         base_prompt = "Pick as many lines as you can from these poem lines:\n"
         base_msg = [{"role": "user", "content": base_prompt}]
-        base_fmt = tokenizer.apply_chat_template(base_msg,
-                                                 add_generation_prompt=True,
-                                                 tokenize=False)
+        base_fmt = tokenizer.apply_chat_template(
+            base_msg, add_generation_prompt=True, tokenize=False
+        )
         base_offset = len(tokenizer(base_fmt).input_ids)
         if input_len <= base_offset:
             raise ValueError(
                 f"'input_len' must be higher than the base prompt length "
-                f"({base_offset}).")
+                f"({base_offset})."
+            )
 
         # Determine how many poem lines to use.
         num_input_lines = round((input_len - base_offset) / avg_len)
@@ -491,21 +598,23 @@ class SonnetDataset(BenchmarkDataset):
 
         samples = []
         while len(samples) < num_requests:
-            extra_lines = random.choices(self.data,
-                                         k=num_input_lines - num_prefix_lines)
+            extra_lines = random.choices(
+                self.data, k=num_input_lines - num_prefix_lines
+            )
             prompt = f"{base_prompt}{''.join(prefix_lines + extra_lines)}"
             msg = [{"role": "user", "content": prompt}]
             prompt_formatted = tokenizer.apply_chat_template(
-                msg, add_generation_prompt=True, tokenize=False)
+                msg, add_generation_prompt=True, tokenize=False
+            )
             prompt_len = len(tokenizer(prompt_formatted).input_ids)
             if prompt_len <= input_len:
                 samples.append(
                     SampleRequest(
-                        prompt=prompt_formatted
-                        if return_prompt_formatted else prompt,
+                        prompt=prompt_formatted if return_prompt_formatted else prompt,
                         prompt_len=prompt_len,
                         expected_output_len=output_len,
-                    ))
+                    )
+                )
         return samples
 
 
@@ -525,7 +634,9 @@ class BurstGPTDataset(BenchmarkDataset):
         super().__init__(**kwargs)
         self.load_data()
 
-    def load_data(self, ):
+    def load_data(
+        self,
+    ):
         if self.dataset_path is None:
             raise ValueError("dataset_path must be provided for loading data.")
 
@@ -539,8 +650,7 @@ class BurstGPTDataset(BenchmarkDataset):
 
     def _sample_loaded_data(self, num_requests: int) -> list:
         if num_requests <= len(self.data):
-            data = self.data.sample(n=num_requests,
-                                    random_state=self.random_seed)
+            data = self.data.sample(n=num_requests, random_state=self.random_seed)
         else:
             data = self.data.sample(
                 n=num_requests,
@@ -564,7 +674,8 @@ class BurstGPTDataset(BenchmarkDataset):
             input_len = int(data[i][2])
             output_len = int(data[i][3])
             lora_req, tokenizer = self.get_random_lora_request(
-                tokenizer=tokenizer, max_loras=max_loras, lora_path=lora_path)
+                tokenizer=tokenizer, max_loras=max_loras, lora_path=lora_path
+            )
             vocab_size = tokenizer.vocab_size
             # Generate a synthetic prompt: a list of token IDs computed as (i +
             # j) modulo vocab_size.
@@ -576,7 +687,8 @@ class BurstGPTDataset(BenchmarkDataset):
                     prompt_len=input_len,
                     expected_output_len=output_len,
                     lora_request=lora_req,
-                ))
+                )
+            )
         return samples
 
 
@@ -592,6 +704,7 @@ class HuggingFaceDataset(BenchmarkDataset):
         self,
         dataset_path: str,
         dataset_split: str,
+        no_stream: bool = False,
         dataset_subset: Optional[str] = None,
         **kwargs,
     ) -> None:
@@ -599,6 +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:
@@ -607,7 +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)
 
@@ -619,20 +733,23 @@ class HuggingFaceDataset(BenchmarkDataset):
 
 class ConversationDataset(HuggingFaceDataset):
     """Dataset for conversation data with multimodal support."""
+
     SUPPORTED_DATASET_PATHS = {
-        'lmms-lab/LLaVA-OneVision-Data', 'Aeala/ShareGPT_Vicuna_unfiltered'
+        "lmms-lab/LLaVA-OneVision-Data",
+        "Aeala/ShareGPT_Vicuna_unfiltered",
     }
     IS_MULTIMODAL = True
 
-    def sample(self,
-               tokenizer: PreTrainedTokenizerBase,
-               num_requests: int,
-               output_len: Optional[int] = None,
-               enable_multimodal_chat: bool = False,
-               **kwargs) -> list:
+    def sample(
+        self,
+        tokenizer: PreTrainedTokenizerBase,
+        num_requests: int,
+        output_len: Optional[int] = None,
+        enable_multimodal_chat: bool = False,
+        **kwargs,
+    ) -> list:
         # Filter examples with at least 2 conversations
-        filtered_data = self.data.filter(
-            lambda x: len(x["conversations"]) >= 2)
+        filtered_data = self.data.filter(lambda x: len(x["conversations"]) >= 2)
         sampled_requests = []
         dynamic_output = output_len is None
 
@@ -648,24 +765,22 @@ class ConversationDataset(HuggingFaceDataset):
             completion_len = len(completion_ids)
             output_len = completion_len if dynamic_output else output_len
             assert isinstance(output_len, int) and output_len > 0
-            if dynamic_output and not is_valid_sequence(
-                    prompt_len, completion_len):
+            if dynamic_output and not is_valid_sequence(prompt_len, completion_len):
                 continue
-            mm_content = process_image(
-                item["image"]) if "image" in item else None
+            mm_content = process_image(item["image"]) if "image" in item else None
             if enable_multimodal_chat:
                 # Note: when chat is enabled the request prompt_len is no longer
                 # accurate and we will be using request output to count the
                 # actual prompt len and output len
-                prompt = self.apply_multimodal_chat_transformation(
-                    prompt, mm_content)
+                prompt = self.apply_multimodal_chat_transformation(prompt, mm_content)
             sampled_requests.append(
                 SampleRequest(
                     prompt=prompt,
                     prompt_len=prompt_len,
                     expected_output_len=output_len,
                     multi_modal_data=mm_content,
-                ))
+                )
+            )
         self.maybe_oversample_requests(sampled_requests, num_requests)
         return sampled_requests
 
@@ -682,10 +797,8 @@ class VisionArenaDataset(HuggingFaceDataset):
 
     DEFAULT_OUTPUT_LEN = 128
     SUPPORTED_DATASET_PATHS = {
-        "lmarena-ai/VisionArena-Chat":
-        lambda x: x["conversation"][0][0]["content"],
-        "lmarena-ai/vision-arena-bench-v0.1":
-        lambda x: x["turns"][0][0]["content"]
+        "lmarena-ai/VisionArena-Chat": lambda x: x["conversation"][0][0]["content"],
+        "lmarena-ai/vision-arena-bench-v0.1": lambda x: x["turns"][0][0]["content"],
     }
     IS_MULTIMODAL = True
 
@@ -697,16 +810,14 @@ class VisionArenaDataset(HuggingFaceDataset):
         enable_multimodal_chat: bool = False,
         **kwargs,
     ) -> list:
-        output_len = (output_len
-                      if output_len is not None else self.DEFAULT_OUTPUT_LEN)
+        output_len = output_len if output_len is not None else self.DEFAULT_OUTPUT_LEN
         sampled_requests = []
         for item in self.data:
             if len(sampled_requests) >= num_requests:
                 break
             parser_fn = self.SUPPORTED_DATASET_PATHS.get(self.dataset_path)
             if parser_fn is None:
-                raise ValueError(
-                    f"Unsupported dataset path: {self.dataset_path}")
+                raise ValueError(f"Unsupported dataset path: {self.dataset_path}")
             prompt = parser_fn(item)
             mm_content = process_image(item["images"][0])
             prompt_len = len(tokenizer(prompt).input_ids)
@@ -714,15 +825,15 @@ class VisionArenaDataset(HuggingFaceDataset):
                 # Note: when chat is enabled the request prompt_len is no longer
                 # accurate and we will be using request output to count the
                 # actual prompt len
-                prompt = self.apply_multimodal_chat_transformation(
-                    prompt, mm_content)
+                prompt = self.apply_multimodal_chat_transformation(prompt, mm_content)
             sampled_requests.append(
                 SampleRequest(
                     prompt=prompt,
                     prompt_len=prompt_len,
                     expected_output_len=output_len,
                     multi_modal_data=mm_content,
-                ))
+                )
+            )
         self.maybe_oversample_requests(sampled_requests, num_requests)
         return sampled_requests
 
@@ -747,26 +858,91 @@ class InstructCoderDataset(HuggingFaceDataset):
         "likaixin/InstructCoder",
     }
 
-    def sample(self,
-               tokenizer: PreTrainedTokenizerBase,
-               num_requests: int,
-               output_len: Optional[int] = None,
-               enable_multimodal_chat: bool = False,
-               **kwargs) -> list:
-        output_len = (output_len
-                      if output_len is not None else self.DEFAULT_OUTPUT_LEN)
+    def sample(
+        self,
+        tokenizer: PreTrainedTokenizerBase,
+        num_requests: int,
+        output_len: Optional[int] = None,
+        enable_multimodal_chat: bool = False,
+        **kwargs,
+    ) -> list:
+        output_len = output_len if output_len is not None else self.DEFAULT_OUTPUT_LEN
         sampled_requests = []
         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(
                     prompt=prompt,
                     prompt_len=prompt_len,
                     expected_output_len=output_len,
-                ))
+                )
+            )
+        self.maybe_oversample_requests(sampled_requests, num_requests)
+        return sampled_requests
+
+
+# -----------------------------------------------------------------------------
+# MT-Bench Dataset Implementation
+# -----------------------------------------------------------------------------
+
+
+class MTBenchDataset(HuggingFaceDataset):
+    """
+    MT-Bench Dataset.
+    https://huggingface.co/datasets/philschmid/mt-bench
+
+    We create a single turn dataset for MT-Bench.
+    This is similar to Spec decoding benchmark setup in vLLM
+    https://github.com/vllm-project/vllm/blob/9d98ab5ec/examples/offline_inference/eagle.py#L14-L18
+    """  # noqa: E501
+
+    DEFAULT_OUTPUT_LEN = 256  # avg len used in SD bench in vLLM
+    SUPPORTED_DATASET_PATHS = {
+        "philschmid/mt-bench",
+    }
+
+    def sample(
+        self,
+        tokenizer: PreTrainedTokenizerBase,
+        num_requests: int,
+        output_len: Optional[int] = None,
+        enable_multimodal_chat: bool = False,
+        **kwargs,
+    ) -> list:
+        output_len = output_len if output_len is not None else self.DEFAULT_OUTPUT_LEN
+        sampled_requests = []
+
+        for item in self.data:
+            if len(sampled_requests) >= num_requests:
+                break
+            prompt = item["turns"][0]
+
+            # 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(
+                    prompt=prompt,
+                    prompt_len=prompt_len,
+                    expected_output_len=output_len,
+                )
+            )
         self.maybe_oversample_requests(sampled_requests, num_requests)
         return sampled_requests
 
@@ -780,23 +956,27 @@ class AIMODataset(HuggingFaceDataset):
     """
     Dataset class for processing a AIMO dataset with reasoning questions.
     """
+
     SUPPORTED_DATASET_PATHS = {
-        "AI-MO/aimo-validation-aime", "AI-MO/NuminaMath-1.5",
-        "AI-MO/NuminaMath-CoT"
+        "AI-MO/aimo-validation-aime",
+        "AI-MO/NuminaMath-1.5",
+        "AI-MO/NuminaMath-CoT",
     }
 
-    def sample(self,
-               tokenizer: PreTrainedTokenizerBase,
-               num_requests: int,
-               output_len: Optional[int] = None,
-               **kwargs) -> list:
+    def sample(
+        self,
+        tokenizer: PreTrainedTokenizerBase,
+        num_requests: int,
+        output_len: Optional[int] = None,
+        **kwargs,
+    ) -> list:
         sampled_requests = []
         dynamic_output = output_len is None
 
         for item in self.data:
             if len(sampled_requests) >= num_requests:
                 break
-            prompt, completion = item['problem'], item["solution"]
+            prompt, completion = item["problem"], item["solution"]
 
             prompt_ids = tokenizer(prompt).input_ids
             completion_ids = tokenizer(completion).input_ids
@@ -804,10 +984,9 @@ class AIMODataset(HuggingFaceDataset):
             completion_len = len(completion_ids)
             output_len = completion_len if dynamic_output else output_len
             assert isinstance(output_len, int) and output_len > 0
-            if dynamic_output and not is_valid_sequence(prompt_len,
-                                                        completion_len,
-                                                        max_prompt_len=2048,
-                                                        max_total_len=32000):
+            if dynamic_output and not is_valid_sequence(
+                prompt_len, completion_len, max_prompt_len=2048, max_total_len=32000
+            ):
                 continue
             sampled_requests.append(
                 SampleRequest(
@@ -815,11 +994,100 @@ class AIMODataset(HuggingFaceDataset):
                     prompt_len=prompt_len,
                     expected_output_len=output_len,
                     multi_modal_data=None,
-                ))
+                )
+            )
         self.maybe_oversample_requests(sampled_requests, num_requests)
         return sampled_requests
 
 
+# -----------------------------------------------------------------------------
+# Next Edit Prediction Dataset Implementation
+# -----------------------------------------------------------------------------
+
+
+zeta_prompt = """### Instruction:
+You are a code completion assistant and your task is to analyze user edits and then rewrite an excerpt that the user provides, suggesting the appropriate edits within the excerpt, taking into account the cursor location.
+
+### User Edits:
+
+{}
+
+### User Excerpt:
+
+{}
+
+### Response:
+
+"""  # noqa: E501
+
+
+def _format_zeta_prompt(
+    sample: dict, original_start_marker: str = "<|editable_region_start|>"
+) -> dict:
+    """Format the zeta prompt for the Next Edit Prediction (NEP) dataset.
+
+    This function formats examples from the NEP dataset
+    into prompts and expected outputs. It could be
+    further extended to support more NEP datasets.
+
+    Args:
+        sample: The dataset sample containing events,
+            inputs, and outputs.
+        original_start_marker: The marker indicating the
+            start of the editable region. Defaults to
+            "<|editable_region_start|>".
+
+    Returns:
+        A dictionary with the formatted prompts and expected outputs.
+    """
+    events = sample["events"]
+    input = sample["input"]
+    output = sample["output"]
+    prompt = zeta_prompt.format(events, input)
+
+    # following the original implementation, extract the focused region
+    # from the raw output
+    output_start_index = output.find(original_start_marker)
+    output_focused_region = output[output_start_index:]
+    expected_output = output_focused_region
+
+    return {"prompt": prompt, "expected_output": expected_output}
+
+
+class NextEditPredictionDataset(HuggingFaceDataset):
+    """
+    Dataset class for processing a Next Edit Prediction dataset.
+    """
+
+    SUPPORTED_DATASET_PATHS = {
+        "zed-industries/zeta",
+    }
+    MAPPING_PROMPT_FUNCS = {
+        "zed-industries/zeta": _format_zeta_prompt,
+    }
+
+    def sample(self, tokenizer: PreTrainedTokenizerBase, num_requests: int, **kwargs):
+        formatting_prompt_func = self.MAPPING_PROMPT_FUNCS.get(self.dataset_path)
+        if formatting_prompt_func is None:
+            raise ValueError(f"Unsupported dataset path: {self.dataset_path}")
+        samples = []
+        for sample in self.data:
+            sample = formatting_prompt_func(sample)
+            samples.append(
+                SampleRequest(
+                    prompt=sample["prompt"],
+                    prompt_len=len(tokenizer(sample["prompt"]).input_ids),
+                    expected_output_len=len(
+                        tokenizer(sample["expected_output"]).input_ids
+                    ),
+                )
+            )
+            if len(samples) >= num_requests:
+                break
+        self.maybe_oversample_requests(samples, num_requests)
+        return samples
+
+
 # -----------------------------------------------------------------------------
 # ASR Dataset Implementation
 # -----------------------------------------------------------------------------
@@ -842,18 +1110,22 @@ class ASRDataset(HuggingFaceDataset):
     | AMI            | Meetings                               | Spontaneous              | ihm, sdm                    |
     +----------------+----------------------------------------+--------------------------+-----------------------------+
 
-    """ # noqa: E501
+    """  # noqa: E501
+
     SUPPORTED_DATASET_PATHS = {
-        "openslr/librispeech_asr", "facebook/voxpopuli", "LIUM/tedlium",
-        "edinburghcstr/ami", "speechcolab/gigaspeech", "kensho/spgispeech"
+        "openslr/librispeech_asr",
+        "facebook/voxpopuli",
+        "LIUM/tedlium",
+        "edinburghcstr/ami",
+        "speechcolab/gigaspeech",
+        "kensho/spgispeech",
     }
 
     DEFAULT_OUTPUT_LEN = 128
     IS_MULTIMODAL = True
 
     # TODO Whisper-specific. Abstract interface when more models are supported.
-    TRANSCRIPTION_PREAMBLE = "<|startoftranscript|><|en|><|transcribe|>"\
-                              "<|notimestamps|>"
+    TRANSCRIPTION_PREAMBLE = "<|startoftranscript|><|en|><|transcribe|><|notimestamps|>"
     skip_long_audios: bool = True
 
     def sample(
@@ -864,8 +1136,8 @@ class ASRDataset(HuggingFaceDataset):
         **kwargs,
     ) -> list:
         import librosa
-        output_len = (output_len
-                      if output_len is not None else self.DEFAULT_OUTPUT_LEN)
+
+        output_len = output_len if output_len is not None else self.DEFAULT_OUTPUT_LEN
         prompt = ASRDataset.TRANSCRIPTION_PREAMBLE
         prompt_len = len(tokenizer(prompt).input_ids)
         sampled_requests = []
@@ -888,10 +1160,14 @@ class ASRDataset(HuggingFaceDataset):
                     prompt_len=prompt_len,
                     expected_output_len=output_len,
                     multi_modal_data=mm_content,
-                ))
+                )
+            )
         if skipped:
-            logger.warning("%d samples discarded from dataset due to" \
-                           " their length being greater than" \
-                           " what Whisper supports.", skipped)
+            logger.warning(
+                "%d samples discarded from dataset due to"
+                " their length being greater than"
+                " what Whisper supports.",
+                skipped,
+            )
         self.maybe_oversample_requests(sampled_requests, num_requests)
         return sampled_requests

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
@@ -6,14 +7,13 @@ import dataclasses
 import json
 import os
 import time
-from pathlib import Path
 from typing import Any, Optional
 
 import numpy as np
-import torch
-from benchmark_utils import convert_to_pytorch_benchmark_format, write_to_json
 from tqdm import tqdm
 
+import vllm.envs as envs
+from benchmark_utils import convert_to_pytorch_benchmark_format, write_to_json
 from vllm import LLM, SamplingParams
 from vllm.engine.arg_utils import EngineArgs
 from vllm.inputs import PromptType
@@ -21,13 +21,14 @@ from vllm.sampling_params import BeamSearchParams
 from vllm.utils import FlexibleArgumentParser
 
 
-def save_to_pytorch_benchmark_format(args: argparse.Namespace,
-                                     results: dict[str, Any]) -> None:
+def save_to_pytorch_benchmark_format(
+    args: argparse.Namespace, results: dict[str, Any]
+) -> None:
     pt_records = convert_to_pytorch_benchmark_format(
         args=args,
         metrics={"latency": results["latencies"]},
-        extra_info={k: results[k]
-                    for k in ["avg_latency", "percentiles"]})
+        extra_info={k: results[k] for k in ["avg_latency", "percentiles"]},
+    )
     if pt_records:
         pt_file = f"{os.path.splitext(args.output_json)[0]}.pytorch.json"
         write_to_json(pt_file, pt_records)
@@ -42,9 +43,11 @@ def main(args: argparse.Namespace):
     # the engine will automatically process the request in multiple batches.
     llm = LLM(**dataclasses.asdict(engine_args))
     assert llm.llm_engine.model_config.max_model_len >= (
-        args.input_len +
-        args.output_len), ("Please ensure that max_model_len is greater than"
-                           " the sum of input_len and output_len.")
+        args.input_len + args.output_len
+    ), (
+        "Please ensure that max_model_len is greater than"
+        " the sum of input_len and output_len."
+    )
 
     sampling_params = SamplingParams(
         n=args.n,
@@ -55,18 +58,16 @@ def main(args: argparse.Namespace):
         detokenize=not args.disable_detokenize,
     )
     print(sampling_params)
-    dummy_prompt_token_ids = np.random.randint(10000,
-                                               size=(args.batch_size,
-                                                     args.input_len))
-    dummy_prompts: list[PromptType] = [{
-        "prompt_token_ids": batch
-    } for batch in dummy_prompt_token_ids.tolist()]
+    dummy_prompt_token_ids = np.random.randint(
+        10000, size=(args.batch_size, args.input_len)
+    )
+    dummy_prompts: list[PromptType] = [
+        {"prompt_token_ids": batch} for batch in dummy_prompt_token_ids.tolist()
+    ]
 
     def llm_generate():
         if not args.use_beam_search:
-            llm.generate(dummy_prompts,
-                         sampling_params=sampling_params,
-                         use_tqdm=False)
+            llm.generate(dummy_prompts, sampling_params=sampling_params, use_tqdm=False)
         else:
             llm.beam_search(
                 dummy_prompts,
@@ -79,16 +80,9 @@ def main(args: argparse.Namespace):
 
     def run_to_completion(profile_dir: Optional[str] = None):
         if profile_dir:
-            with torch.profiler.profile(
-                    activities=[
-                        torch.profiler.ProfilerActivity.CPU,
-                        torch.profiler.ProfilerActivity.CUDA,
-                    ],
-                    on_trace_ready=torch.profiler.tensorboard_trace_handler(
-                        str(profile_dir)),
-            ) as p:
-                llm_generate()
-            print(p.key_averages().table(sort_by="self_cuda_time_total"))
+            llm.start_profile()
+            llm_generate()
+            llm.stop_profile()
         else:
             start_time = time.perf_counter()
             llm_generate()
@@ -101,10 +95,7 @@ def main(args: argparse.Namespace):
         run_to_completion(profile_dir=None)
 
     if args.profile:
-        profile_dir = args.profile_result_dir
-        if not profile_dir:
-            profile_dir = (Path(".") / "vllm_benchmark_result" /
-                           f"latency_result_{time.time()}")
+        profile_dir = envs.VLLM_TORCH_PROFILER_DIR
         print(f"Profiling (results will be saved to '{profile_dir}')...")
         run_to_completion(profile_dir=profile_dir)
         return
@@ -132,10 +123,11 @@ 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.")
+        "requests till completion."
+    )
     parser.add_argument("--input-len", type=int, default=32)
     parser.add_argument("--output-len", type=int, default=128)
     parser.add_argument("--batch-size", type=int, default=8)
@@ -152,22 +144,14 @@ if __name__ == "__main__":
         default=10,
         help="Number of iterations to run for warmup.",
     )
-    parser.add_argument("--num-iters",
-                        type=int,
-                        default=30,
-                        help="Number of iterations to run.")
+    parser.add_argument(
+        "--num-iters", type=int, default=30, help="Number of iterations to run."
+    )
     parser.add_argument(
         "--profile",
         action="store_true",
         help="profile the generation process of a single batch",
     )
-    parser.add_argument(
-        "--profile-result-dir",
-        type=str,
-        default=None,
-        help=("path to save the pytorch profiler output. Can be visualized "
-              "with ui.perfetto.dev or Tensorboard."),
-    )
     parser.add_argument(
         "--output-json",
         type=str,
@@ -177,10 +161,26 @@ if __name__ == "__main__":
     parser.add_argument(
         "--disable-detokenize",
         action="store_true",
-        help=("Do not detokenize responses (i.e. do not include "
-              "detokenization time in the latency measurement)"),
+        help=(
+            "Do not detokenize responses (i.e. do not include "
+            "detokenization time in the latency measurement)"
+        ),
     )
 
     parser = EngineArgs.add_cli_args(parser)
+    # V1 enables prefix caching by default which skews the latency
+    # numbers. We need to disable prefix caching by default.
+    parser.set_defaults(enable_prefix_caching=False)
+
+    return parser
+
+
+if __name__ == "__main__":
+    parser = create_argument_parser()
     args = parser.parse_args()
+    if args.profile and not envs.VLLM_TORCH_PROFILER_DIR:
+        raise OSError(
+            "The environment variable 'VLLM_TORCH_PROFILER_DIR' is not set. "
+            "Please set it to a valid path to use torch profiler."
+        )
     main(args)

benchmarks/benchmark_long_document_qa_throughput.py

@@ -1,4 +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.
 
@@ -76,7 +77,7 @@ def repeat_prompts(prompts, repeat_count, mode: str):
             - 'random': Shuffle the prompts randomly after repetition.
             - 'tile': Repeat the entire prompt list in sequence.
               Example: [1, 2, 3] -> [1, 2, 3, 1, 2, 3].
-            - 'interleave': Repeat each prompt consecutively before moving to 
+            - 'interleave': Repeat each prompt consecutively before moving to
               the next. Example: [1, 2, 3] -> [1, 1, 2, 2, 3, 3].
 
     Returns:
@@ -86,20 +87,21 @@ def repeat_prompts(prompts, repeat_count, mode: str):
         ValueError: If an invalid mode is provided.
     """
     print("Repeat mode: ", mode)
-    if mode == 'random':
+    if mode == "random":
         repeated_prompts = prompts * repeat_count
         random.shuffle(repeated_prompts)
         return repeated_prompts
-    elif mode == 'tile':
+    elif mode == "tile":
         return prompts * repeat_count
-    elif mode == 'interleave':
+    elif mode == "interleave":
         repeated_prompts = []
         for prompt in prompts:
             repeated_prompts.extend([prompt] * repeat_count)
         return repeated_prompts
     else:
-        raise ValueError(f"Invalid mode: {mode}, only support "
-                         "'random', 'tile', 'interleave'")
+        raise ValueError(
+            f"Invalid mode: {mode}, only support 'random', 'tile', 'interleave'"
+        )
 
 
 def main(args):
@@ -109,16 +111,16 @@ def main(args):
     # we append the document id at the beginning to avoid any of the document
     # being the prefix of other documents
     prompts = [
-        str(i) + ' '.join(['hi'] * args.document_length)
+        str(i) + " ".join(["hi"] * args.document_length)
         for i in range(args.num_documents)
     ]
 
     prompts = repeat_prompts(prompts, args.repeat_count, mode=args.repeat_mode)
 
     warmup_prompts = [
-        "This is warm up request " + str(i) + \
-                ' '.join(['hi'] * args.document_length)
-        for i in range(args.num_documents)]
+        "This is warm up request " + str(i) + " ".join(["hi"] * args.document_length)
+        for i in range(args.num_documents)
+    ]
 
     # Create the LLM engine
     engine_args = EngineArgs.from_cli_args(args)
@@ -140,45 +142,61 @@ def main(args):
     )
 
 
-if __name__ == "__main__":
+def create_argument_parser():
     parser = FlexibleArgumentParser(
-        description=
-        'Benchmark the performance with or without automatic prefix caching.')
+        description="Benchmark the performance with or "
+        "without automatic prefix caching."
+    )
 
     parser.add_argument(
-        '--document-length',
+        "--document-length",
         type=int,
         # Roughly the number of tokens for a system paper,
         # excluding images
         default=20000,
-        help='Range of input lengths for sampling prompts,'
-        'specified as "min:max" (e.g., "128:256").')
+        help="Range of input lengths for sampling prompts, "
+        'specified as "min:max" (e.g., "128:256").',
+    )
 
-    parser.add_argument('--num-documents',
-                        type=int,
-                        default=8,
-                        help='Range of input lengths for sampling prompts,'
-                        'specified as "min:max" (e.g., "128:256").')
+    parser.add_argument(
+        "--num-documents",
+        type=int,
+        default=8,
+        help="Range of input lengths for sampling prompts, "
+        'specified as "min:max" (e.g., "128:256").',
+    )
 
-    parser.add_argument('--output-len', type=int, default=10)
+    parser.add_argument("--output-len", type=int, default=10)
 
-    parser.add_argument('--repeat-count',
-                        type=int,
-                        default=2,
-                        help='Number of times to repeat each prompt')
+    parser.add_argument(
+        "--repeat-count",
+        type=int,
+        default=2,
+        help="Number of times to repeat each prompt",
+    )
 
-    parser.add_argument("--repeat-mode",
-                        type=str,
-                        default='random',
-                        help='The mode to repeat prompts. The supported '
-                        'modes are "random", "tile", and "interleave". '
-                        'See repeat_prompts() in the source code for details.')
+    parser.add_argument(
+        "--repeat-mode",
+        type=str,
+        default="random",
+        help="The mode to repeat prompts. The supported "
+        'modes are "random", "tile", and "interleave". '
+        "See repeat_prompts() in the source code for details.",
+    )
 
-    parser.add_argument("--shuffle-seed",
-                        type=int,
-                        default=0,
-                        help='Random seed when the repeat mode is "random"')
+    parser.add_argument(
+        "--shuffle-seed",
+        type=int,
+        default=0,
+        help='Random seed when the repeat mode is "random"',
+    )
 
     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.
 
@@ -63,8 +64,7 @@ class Request:
     output_len: int
 
 
-def sample_tokens(tokenizer: PreTrainedTokenizerBase,
-                  length: int) -> list[int]:
+def sample_tokens(tokenizer: PreTrainedTokenizerBase, length: int) -> list[int]:
     vocab = tokenizer.get_vocab()
     all_special_ids = set(tokenizer.all_special_ids)
 
@@ -91,8 +91,10 @@ def sample_requests_from_dataset(
     # Filter out the conversations with less than 2 turns.
     dataset = [data for data in dataset if len(data["conversations"]) >= 2]
     # Only keep the first two turns of each conversation.
-    dataset = [(data["conversations"][0]["value"],
-                data["conversations"][1]["value"]) for data in dataset]
+    dataset = [
+        (data["conversations"][0]["value"], data["conversations"][1]["value"])
+        for data in dataset
+    ]
 
     # Shuffle the dataset.
     random.shuffle(dataset)
@@ -113,8 +115,9 @@ def sample_requests_from_dataset(
         completion = dataset[i][1]
         completion_token_ids = tokenizer(completion).input_ids
         prompt_len = len(prompt_token_ids)
-        output_len = (len(completion_token_ids)
-                      if fixed_output_len is None else fixed_output_len)
+        output_len = (
+            len(completion_token_ids) if fixed_output_len is None else fixed_output_len
+        )
         if min_len <= prompt_len <= max_len:
             filtered_requests.append(Request(prompt, prompt_len, output_len))
 
@@ -128,27 +131,27 @@ def sample_requests_from_random(
     fixed_output_len: Optional[int],
     prefix_len: int,
 ) -> list[Request]:
-
     requests = []
     prefix_token_ids = sample_tokens(tokenizer, prefix_len)
     min_len, max_len = input_length_range
 
     for i in range(num_requests):
         unique_part_token_ids = sample_tokens(
-            tokenizer,
-            random.randint(min_len - prefix_len, max_len - prefix_len))
+            tokenizer, random.randint(min_len - prefix_len, max_len - prefix_len)
+        )
         prompt_token_ids = prefix_token_ids + unique_part_token_ids
         prompt = tokenizer.decode(prompt_token_ids)
         prompt_len = len(prompt_token_ids)
-        assert (min_len <= prompt_len <= max_len
-                ), f"prompt_len {prompt_len} out of range {min_len}:{max_len}"
+        assert min_len <= prompt_len <= max_len, (
+            f"prompt_len {prompt_len} out of range {min_len}:{max_len}"
+        )
         requests.append(Request(prompt, prompt_len, fixed_output_len))
     return requests
 
 
-def repeat_and_sort_requests(requests: list[Request],
-                             repeat_count: int,
-                             sort: bool = False) -> list[str]:
+def repeat_and_sort_requests(
+    requests: list[Request], repeat_count: int, sort: bool = False
+) -> list[str]:
     repeated_requests = requests * repeat_count
     if sort:
         repeated_requests.sort(key=lambda x: x[1])
@@ -159,14 +162,14 @@ def repeat_and_sort_requests(requests: list[Request],
 
 def main(args):
     tokenizer = get_tokenizer(args.model, trust_remote_code=True)
-    input_length_range = tuple(map(int, args.input_length_range.split(':')))
+    input_length_range = tuple(map(int, args.input_length_range.split(":")))
     random.seed(args.seed)
     if args.dataset_path is not None:
         if args.prefix_len > 0:
-            raise ValueError("prefix-len is not supported when "
-                             "dataset-path is provided.")
-        print(f"Start to sample {args.num_prompts} prompts "
-              f"from {args.dataset_path}")
+            raise ValueError(
+                "prefix-len is not supported when dataset-path is provided."
+            )
+        print(f"Start to sample {args.num_prompts} prompts from {args.dataset_path}")
         filtered_requests = sample_requests_from_dataset(
             dataset_path=args.dataset_path,
             num_requests=args.num_prompts,
@@ -196,14 +199,16 @@ def main(args):
 
     llm = LLM(**dataclasses.asdict(engine_args))
 
-    sampling_params = SamplingParams(temperature=0,
-                                     max_tokens=args.output_len,
-                                     detokenize=not args.disable_detokenize)
+    sampling_params = SamplingParams(
+        temperature=0,
+        max_tokens=args.output_len,
+        detokenize=not args.disable_detokenize,
+    )
 
     print("Testing filtered requests")
-    prompts = repeat_and_sort_requests(filtered_requests,
-                                       repeat_count=args.repeat_count,
-                                       sort=args.sort)
+    prompts = repeat_and_sort_requests(
+        filtered_requests, repeat_count=args.repeat_count, sort=args.sort
+    )
 
     print("------start generating------")
     test_prefix(
@@ -213,31 +218,37 @@ def main(args):
     )
 
 
-if __name__ == "__main__":
+def create_argument_parser():
     parser = FlexibleArgumentParser(
-        description=
-        'Benchmark the performance with or without automatic prefix caching.')
-    parser.add_argument("--dataset-path",
-                        type=str,
-                        default=None,
-                        help="Path to the dataset.")
-    parser.add_argument('--output-len', type=int, default=10)
-    parser.add_argument('--num-prompts',
-                        type=int,
-                        required=True,
-                        help="Number of the prompts sampled from dataset")
-    parser.add_argument('--repeat-count',
-                        type=int,
-                        default=1,
-                        help='Number of times to repeat each prompt')
-    parser.add_argument('--sort',
-                        action='store_true',
-                        help='Sort prompts by input length')
-    parser.add_argument('--input-length-range',
-                        type=str,
-                        required=True,
-                        help='Range of input lengths for sampling prompts,'
-                        'specified as "min:max" (e.g., "128:256").')
+        description="Benchmark the performance with or without "
+        "automatic prefix caching."
+    )
+    parser.add_argument(
+        "--dataset-path", type=str, default=None, help="Path to the dataset."
+    )
+    parser.add_argument("--output-len", type=int, default=10)
+    parser.add_argument(
+        "--num-prompts",
+        type=int,
+        required=True,
+        help="Number of the prompts sampled from dataset",
+    )
+    parser.add_argument(
+        "--repeat-count",
+        type=int,
+        default=1,
+        help="Number of times to repeat each prompt",
+    )
+    parser.add_argument(
+        "--sort", action="store_true", help="Sort prompts by input length"
+    )
+    parser.add_argument(
+        "--input-length-range",
+        type=str,
+        required=True,
+        help="Range of input lengths for sampling prompts,"
+        'specified as "min:max" (e.g., "128:256").',
+    )
     parser.add_argument(
         "--prefix-len",
         type=int,
@@ -248,12 +259,20 @@ if __name__ == "__main__":
         "when dataset-path is not provided.",
     )
     parser.add_argument(
-        '--disable-detokenize',
-        action='store_true',
-        help=("Do not detokenize responses (i.e. do not include "
-              "detokenization time in the latency measurement)"),
+        "--disable-detokenize",
+        action="store_true",
+        help=(
+            "Do not detokenize responses (i.e. do not include "
+            "detokenization time in the latency measurement)"
+        ),
     )
 
     parser = EngineArgs.add_cli_args(parser)
+
+    return parser
+
+
+if __name__ == "__main__":
+    parser = create_argument_parser()
     args = parser.parse_args()
     main(args)

benchmarks/benchmark_prioritization.py

@@ -1,5 +1,7 @@
 # SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 """Benchmark offline prioritization."""
+
 import argparse
 import dataclasses
 import json
@@ -13,7 +15,7 @@ from vllm.engine.arg_utils import EngineArgs
 from vllm.utils import FlexibleArgumentParser
 
 
-#Select a equi-probable random priority
+# Select a equi-probable random priority
 def get_random_flag():
     return 0 if random.random() < 0.5 else 1
 
@@ -33,8 +35,10 @@ def sample_requests(
     # Filter out the conversations with less than 2 turns.
     dataset = [data for data in dataset if len(data["conversations"]) >= 2]
     # Only keep the first two turns of each conversation.
-    dataset = [(data["conversations"][0]["value"],
-                data["conversations"][1]["value"]) for data in dataset]
+    dataset = [
+        (data["conversations"][0]["value"], data["conversations"][1]["value"])
+        for data in dataset
+    ]
 
     # Shuffle the dataset.
     random.shuffle(dataset)
@@ -51,8 +55,9 @@ def sample_requests(
         completion = dataset[i][1]
         completion_token_ids = tokenizer(completion).input_ids
         prompt_len = len(prompt_token_ids)
-        output_len = len(completion_token_ids
-                         ) if fixed_output_len is None else fixed_output_len
+        output_len = (
+            len(completion_token_ids) if fixed_output_len is None else fixed_output_len
+        )
         if prompt_len < 4 or output_len < 4:
             # Prune too short sequences.
             continue
@@ -74,13 +79,16 @@ def run_vllm(
     disable_detokenize: bool = False,
 ) -> float:
     from vllm import LLM, SamplingParams
+
     llm = LLM(**dataclasses.asdict(engine_args))
 
     assert all(
         llm.llm_engine.model_config.max_model_len >= (request[1] + request[2])
-        for request in requests), (
-            "Please ensure that max_model_len is greater than the sum of"
-            " input_len and output_len for all requests.")
+        for request in requests
+    ), (
+        "Please ensure that max_model_len is greater than the sum of"
+        " input_len and output_len for all requests."
+    )
 
     # Add the requests to the engine.
     prompts = []
@@ -97,7 +105,8 @@ def run_vllm(
                 ignore_eos=True,
                 max_tokens=output_len,
                 detokenize=not disable_detokenize,
-            ))
+            )
+        )
 
     start = time.perf_counter()
     llm.generate(prompts, sampling_params, priority=priority, use_tqdm=True)
@@ -111,26 +120,33 @@ def main(args: argparse.Namespace):
 
     # Sample the requests.
     tokenizer = AutoTokenizer.from_pretrained(
-        args.tokenizer, trust_remote_code=args.trust_remote_code)
+        args.tokenizer, trust_remote_code=args.trust_remote_code
+    )
     if args.dataset is None:
         # Synthesize a prompt with the given input length.
         prompt = "hi" * (args.input_len - 1)
-        requests = [(prompt, args.input_len, args.output_len,
-                     get_random_flag()) for _ in range(args.num_prompts)]
+        requests = [
+            (prompt, args.input_len, args.output_len, get_random_flag())
+            for _ in range(args.num_prompts)
+        ]
     else:
-        requests = sample_requests(args.dataset, args.num_prompts, tokenizer,
-                                   args.output_len)
+        requests = sample_requests(
+            args.dataset, args.num_prompts, tokenizer, args.output_len
+        )
 
     if args.backend == "vllm":
-        elapsed_time = run_vllm(requests, args.n,
-                                EngineArgs.from_cli_args(args),
-                                args.disable_detokenize)
+        elapsed_time = run_vllm(
+            requests, args.n, EngineArgs.from_cli_args(args), args.disable_detokenize
+        )
     else:
         raise ValueError(f"Unknown backend: {args.backend}")
-    total_num_tokens = sum(prompt_len + output_len
-                           for _, prompt_len, output_len, priority in requests)
-    print(f"Throughput: {len(requests) / elapsed_time:.2f} requests/s, "
-          f"{total_num_tokens / elapsed_time:.2f} tokens/s")
+    total_num_tokens = sum(
+        prompt_len + output_len for _, prompt_len, output_len, priority in requests
+    )
+    print(
+        f"Throughput: {len(requests) / elapsed_time:.2f} requests/s, "
+        f"{total_num_tokens / elapsed_time:.2f} tokens/s"
+    )
 
     # Output JSON results if specified
     if args.output_json:
@@ -145,46 +161,55 @@ 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")
-    parser.add_argument("--dataset",
-                        type=str,
-                        default=None,
-                        help="Path to the dataset.")
-    parser.add_argument("--input-len",
-                        type=int,
-                        default=None,
-                        help="Input prompt length for each request")
-    parser.add_argument("--output-len",
-                        type=int,
-                        default=None,
-                        help="Output length for each request. Overrides the "
-                        "output length from the dataset.")
-    parser.add_argument("--n",
-                        type=int,
-                        default=1,
-                        help="Number of generated sequences per prompt.")
-    parser.add_argument("--num-prompts",
-                        type=int,
-                        default=200,
-                        help="Number of prompts to process.")
     parser.add_argument(
-        '--output-json',
+        "--backend", type=str, choices=["vllm", "hf", "mii"], default="vllm"
+    )
+    parser.add_argument(
+        "--dataset", type=str, default=None, help="Path to the dataset."
+    )
+    parser.add_argument(
+        "--input-len",
+        type=int,
+        default=None,
+        help="Input prompt length for each request",
+    )
+    parser.add_argument(
+        "--output-len",
+        type=int,
+        default=None,
+        help="Output length for each request. Overrides the "
+        "output length from the dataset.",
+    )
+    parser.add_argument(
+        "--n", type=int, default=1, help="Number of generated sequences per prompt."
+    )
+    parser.add_argument(
+        "--num-prompts", type=int, default=200, help="Number of prompts to process."
+    )
+    parser.add_argument(
+        "--output-json",
         type=str,
         default=None,
-        help='Path to save the throughput results in JSON format.')
+        help="Path to save the throughput results in JSON format.",
+    )
     parser.add_argument(
-        '--disable-detokenize',
-        action='store_true',
-        help=("Do not detokenize responses (i.e. do not include "
-              "detokenization time in the latency measurement)"),
+        "--disable-detokenize",
+        action="store_true",
+        help=(
+            "Do not detokenize responses (i.e. do not include "
+            "detokenization time in the latency measurement)"
+        ),
     )
 
     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_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
 
@@ -19,6 +19,7 @@ On the client side, run:
         --endpoint /generate_stream
     to the end of the command above.
 """
+
 import argparse
 import asyncio
 import copy
@@ -36,11 +37,15 @@ from typing import Optional
 import datasets
 import numpy as np
 import pandas as pd
-from backend_request_func import (ASYNC_REQUEST_FUNCS, RequestFuncInput,
-                                  RequestFuncOutput)
 from tqdm.asyncio import tqdm
 from transformers import PreTrainedTokenizerBase
 
+from backend_request_func import (
+    ASYNC_REQUEST_FUNCS,
+    RequestFuncInput,
+    RequestFuncOutput,
+)
+
 try:
     from vllm.transformers_utils.tokenizer import get_tokenizer
 except ImportError:
@@ -52,7 +57,8 @@ except ImportError:
     from argparse import ArgumentParser as FlexibleArgumentParser
 
 from vllm.v1.structured_output.backend_xgrammar import (
-    has_xgrammar_unsupported_json_features)
+    has_xgrammar_unsupported_json_features,
+)
 
 MILLISECONDS_TO_SECONDS_CONVERSION = 1000
 
@@ -98,6 +104,7 @@ class SampleRequest:
         prompt_len: The length of the prompt in tokens.
         expected_output_len: The expected length of the output in tokens.
     """
+
     prompt: str
     prompt_len: int
     expected_output_len: int
@@ -106,45 +113,45 @@ class SampleRequest:
     completion: str = None
 
 
-def sample_requests(tokenizer: PreTrainedTokenizerBase,
-                    args: argparse.Namespace) -> list[SampleRequest]:
-    if args.dataset == 'json' or args.dataset == 'json-unique':
+def sample_requests(
+    tokenizer: PreTrainedTokenizerBase, args: argparse.Namespace
+) -> list[SampleRequest]:
+    if args.dataset == "json" or args.dataset == "json-unique":
         if args.json_schema_path is None:
             dir_path = os.path.dirname(os.path.realpath(__file__))
-            args.json_schema_path = os.path.join(dir_path,
-                                                 "structured_schemas",
-                                                 "structured_schema_1.json")
+            args.json_schema_path = os.path.join(
+                dir_path, "structured_schemas", "structured_schema_1.json"
+            )
         json_schemas = []
         with open(args.json_schema_path) as f:
             schema = json.load(f)
 
-        if args.dataset == 'json-unique':
-            json_schemas = [
-                copy.deepcopy(schema) for _ in range(args.num_prompts)
-            ]
+        if args.dataset == "json-unique":
+            json_schemas = [copy.deepcopy(schema) for _ in range(args.num_prompts)]
             for i in range(len(json_schemas)):
-                json_schemas[i]["properties"][
-                    f"__optional_field_{uuid.uuid4()}"] = {
-                        "type":
-                        "string",
-                        "description":
-                        "An unique optional field to avoid cached schemas"
-                    }
+                if "properties" not in json_schemas[i]:
+                    json_schemas[i]["properties"] = {}
+                json_schemas[i]["properties"][f"__optional_field_{uuid.uuid4()}"] = {
+                    "type": "string",
+                    "description": "An unique optional field to avoid cached schemas",
+                }
         else:
             json_schemas = [schema] * args.num_prompts
 
         def gen_prompt(index: int):
-            return f"Generate an example of a user profile given the following schema: {json.dumps(get_schema(index))}"  # noqa: E501
+            return f"Generate an example of a brief user profile given the following schema: {json.dumps(get_schema(index))}"  # noqa: E501
 
         def get_schema(index: int):
             return json_schemas[index % len(json_schemas)]
 
         requests = [
-            SampleRequest(prompt=gen_prompt(i),
-                          prompt_len=len(tokenizer(gen_prompt(i)).input_ids),
-                          expected_output_len=args.output_len,
-                          schema=get_schema(i),
-                          structure_type=args.structure_type)
+            SampleRequest(
+                prompt=gen_prompt(i),
+                prompt_len=len(tokenizer(gen_prompt(i)).input_ids),
+                expected_output_len=args.output_len,
+                schema=get_schema(i),
+                structure_type=args.structure_type,
+            )
             for i in range(args.num_prompts)
         ]
 
@@ -168,11 +175,13 @@ def sample_requests(tokenizer: PreTrainedTokenizerBase,
         input_len = len(tokenizer(prompt).input_ids)
         print(f"Input length of the prompt: {input_len} tokens")
         requests = [
-            SampleRequest(prompt=prompt,
-                          prompt_len=input_len,
-                          expected_output_len=args.output_len,
-                          schema=schema,
-                          structure_type=args.structure_type)
+            SampleRequest(
+                prompt=prompt,
+                prompt_len=input_len,
+                expected_output_len=args.output_len,
+                schema=schema,
+                structure_type=args.structure_type,
+            )
             for _ in range(args.num_prompts)
         ]
 
@@ -186,11 +195,13 @@ def sample_requests(tokenizer: PreTrainedTokenizerBase,
         input_len = len(tokenizer(prompt).input_ids)
         print(f"Input length of the prompt: {input_len} tokens")
         requests = [
-            SampleRequest(prompt=prompt,
-                          prompt_len=input_len,
-                          expected_output_len=args.output_len,
-                          schema=regex,
-                          structure_type=args.structure_type)
+            SampleRequest(
+                prompt=prompt,
+                prompt_len=input_len,
+                expected_output_len=args.output_len,
+                schema=regex,
+                structure_type=args.structure_type,
+            )
             for _ in range(args.num_prompts)
         ]
 
@@ -201,47 +212,55 @@ def sample_requests(tokenizer: PreTrainedTokenizerBase,
         input_len = len(tokenizer(prompt).input_ids)
         print(f"Input length of the prompt: {input_len} tokens")
         requests = [
-            SampleRequest(prompt=prompt,
-                          prompt_len=input_len,
-                          expected_output_len=args.output_len,
-                          schema=choice,
-                          structure_type=args.structure_type)
+            SampleRequest(
+                prompt=prompt,
+                prompt_len=input_len,
+                expected_output_len=args.output_len,
+                schema=choice,
+                structure_type=args.structure_type,
+            )
             for _ in range(args.num_prompts)
         ]
 
     elif args.dataset == "xgrammar_bench":
         requests: list[SampleRequest] = []
-        dataset = datasets.load_dataset("NousResearch/json-mode-eval",
-                                        split="train")
+        dataset = datasets.load_dataset("NousResearch/json-mode-eval", split="train")
         full_dataset_len = len(dataset)
 
         def _filter_func(item):
             import json
+
             schema = json.loads(item["schema"])
             return not has_xgrammar_unsupported_json_features(schema)
 
         dataset = dataset.filter(_filter_func)
         num_filtered_out = full_dataset_len - len(dataset)
-        print(f"dataset has {len(dataset)} entries after filtering "
-              f"out {num_filtered_out} entries with unsupported features")
+        print(
+            f"dataset has {len(dataset)} entries after filtering "
+            f"out {num_filtered_out} entries with unsupported features"
+        )
         len_dataset = len(dataset)
         for data_point_idx in range(args.num_prompts):
             idx = data_point_idx
             while idx >= len_dataset:
                 idx -= len_dataset
             schema = dataset["schema"][idx]
-            prompt = tokenizer.apply_chat_template(dataset["prompt"][idx],
-                                                   tokenize=False)
+            prompt = tokenizer.apply_chat_template(
+                dataset["prompt"][idx], tokenize=False, add_generation_prompt=True
+            )
             input_len = len(tokenizer(prompt).input_ids)
             completion = dataset["completion"][idx]
 
             requests.append(
-                SampleRequest(prompt=prompt,
-                              prompt_len=input_len,
-                              expected_output_len=args.output_len,
-                              schema=schema,
-                              structure_type=args.structure_type,
-                              completion=completion))
+                SampleRequest(
+                    prompt=prompt,
+                    prompt_len=input_len,
+                    expected_output_len=args.output_len,
+                    schema=schema,
+                    structure_type=args.structure_type,
+                    completion=completion,
+                )
+            )
 
     return requests
 
@@ -273,7 +292,8 @@ async def get_request(
 
     # Calculate scale parameter theta to maintain the desired request_rate.
     assert burstiness > 0, (
-        f"A positive burstiness factor is expected, but given {burstiness}.")
+        f"A positive burstiness factor is expected, but given {burstiness}."
+    )
     theta = 1.0 / (request_rate * burstiness)
 
     for i, request in enumerate(input_requests):
@@ -315,8 +335,8 @@ def calculate_metrics(
             # multiple output tokens may be bundled together
             # Note : this may inflate the output token count slightly
             output_len = len(
-                tokenizer(outputs[i].generated_text,
-                          add_special_tokens=False).input_ids)
+                tokenizer(outputs[i].generated_text, add_special_tokens=False).input_ids
+            )
             actual_output_lens.append(output_len)
             total_input += input_requests[i].prompt_len
             tpot = 0
@@ -340,16 +360,19 @@ def calculate_metrics(
 
         if "ttft" in goodput_config_dict:
             valid_metrics.append(ttfts)
-            slo_values.append(goodput_config_dict["ttft"] /
-                              MILLISECONDS_TO_SECONDS_CONVERSION)
+            slo_values.append(
+                goodput_config_dict["ttft"] / MILLISECONDS_TO_SECONDS_CONVERSION
+            )
         if "tpot" in goodput_config_dict:
             valid_metrics.append(all_tpots)
-            slo_values.append(goodput_config_dict["tpot"] /
-                              MILLISECONDS_TO_SECONDS_CONVERSION)
+            slo_values.append(
+                goodput_config_dict["tpot"] / MILLISECONDS_TO_SECONDS_CONVERSION
+            )
         if "e2el" in goodput_config_dict:
             valid_metrics.append(e2els)
-            slo_values.append(goodput_config_dict["e2el"] /
-                              MILLISECONDS_TO_SECONDS_CONVERSION)
+            slo_values.append(
+                goodput_config_dict["e2el"] / MILLISECONDS_TO_SECONDS_CONVERSION
+            )
 
         for req_metric in zip(*valid_metrics):
             is_good_req = all([s >= r for s, r in zip(slo_values, req_metric)])
@@ -360,7 +383,8 @@ def calculate_metrics(
         warnings.warn(
             "All requests failed. This is likely due to a misconfiguration "
             "on the benchmark arguments.",
-            stacklevel=2)
+            stacklevel=2,
+        )
     metrics = BenchmarkMetrics(
         completed=completed,
         total_input=total_input,
@@ -369,27 +393,31 @@ def calculate_metrics(
         request_goodput=good_completed / dur_s,
         output_throughput=sum(actual_output_lens) / dur_s,
         total_token_throughput=(total_input + sum(actual_output_lens)) / dur_s,
-        mean_ttft_ms=np.mean(ttfts or 0) *
-        1000,  # ttfts is empty if streaming is not supported by backend
+        mean_ttft_ms=np.mean(ttfts or 0)
+        * 1000,  # ttfts is empty if streaming is not supported by backend
         std_ttft_ms=np.std(ttfts or 0) * 1000,
         median_ttft_ms=np.median(ttfts or 0) * 1000,
-        percentiles_ttft_ms=[(p, np.percentile(ttfts or 0, p) * 1000)
-                             for p in selected_percentiles],
+        percentiles_ttft_ms=[
+            (p, np.percentile(ttfts or 0, p) * 1000) for p in selected_percentiles
+        ],
         mean_tpot_ms=np.mean(tpots or 0) * 1000,
         std_tpot_ms=np.std(tpots or 0) * 1000,
         median_tpot_ms=np.median(tpots or 0) * 1000,
-        percentiles_tpot_ms=[(p, np.percentile(tpots or 0, p) * 1000)
-                             for p in selected_percentiles],
+        percentiles_tpot_ms=[
+            (p, np.percentile(tpots or 0, p) * 1000) for p in selected_percentiles
+        ],
         mean_itl_ms=np.mean(itls or 0) * 1000,
         std_itl_ms=np.std(itls or 0) * 1000,
         median_itl_ms=np.median(itls or 0) * 1000,
-        percentiles_itl_ms=[(p, np.percentile(itls or 0, p) * 1000)
-                            for p in selected_percentiles],
+        percentiles_itl_ms=[
+            (p, np.percentile(itls or 0, p) * 1000) for p in selected_percentiles
+        ],
         mean_e2el_ms=np.mean(e2els or 0) * 1000,
         std_e2el_ms=np.std(e2els or 0) * 1000,
         median_e2el_ms=np.median(e2els or 0) * 1000,
-        percentiles_e2el_ms=[(p, np.percentile(e2els or 0, p) * 1000)
-                             for p in selected_percentiles],
+        percentiles_e2el_ms=[
+            (p, np.percentile(e2els or 0, p) * 1000) for p in selected_percentiles
+        ],
     )
 
     return metrics, actual_output_lens
@@ -411,7 +439,6 @@ async def benchmark(
     ignore_eos: bool,
     max_concurrency: Optional[int],
     structured_output_ratio: float,
-    structured_output_backend: str,
     goodput_config_dict: Optional[dict[str, float]] = None,
 ):
     if backend in ASYNC_REQUEST_FUNCS:
@@ -423,18 +450,17 @@ async def benchmark(
         extra_body = {}
         # Add the schema to the extra_body
         extra_body[request.structure_type] = request.schema
-        # Add the specific structured_output_backend
-        extra_body["guided_decoding_backend"] = structured_output_backend
         return extra_body
 
     print("Starting initial single prompt test run...")
     structured_output_req_idx = random.sample(
-        range(len(input_requests)),
-        int(len(input_requests) * structured_output_ratio))
+        range(len(input_requests)), int(len(input_requests) * structured_output_ratio)
+    )
 
     test_request = input_requests[0]
-    test_req_extra_body = (prepare_extra_body(test_request)
-                           if 0 in structured_output_req_idx else None)
+    test_req_extra_body = (
+        prepare_extra_body(test_request) if 0 in structured_output_req_idx else None
+    )
     test_input = RequestFuncInput(
         model=model_id,
         prompt=test_request.prompt,
@@ -448,7 +474,8 @@ async def benchmark(
     if not test_output.success:
         raise ValueError(
             "Initial test run failed - Please make sure benchmark arguments "
-            f"are correctly specified. Error: {test_output.error}")
+            f"are correctly specified. Error: {test_output.error}"
+        )
     else:
         print("Initial test run completed. Starting main benchmark run...")
 
@@ -467,10 +494,7 @@ async def benchmark(
         if profile_output.success:
             print("Profiler started")
 
-    if burstiness == 1.0:
-        distribution = "Poisson process"
-    else:
-        distribution = "Gamma distribution"
+    distribution = "Poisson process" if burstiness == 1.0 else "Gamma distribution"
 
     print(f"Traffic request rate: {request_rate}")
     print(f"Burstiness factor: {burstiness} ({distribution})")
@@ -482,24 +506,21 @@ async def benchmark(
     # and it will simplify the code in limited_request_func.
     #    semaphore = (asyncio.Semaphore(max_concurrency)
     #                 if max_concurrency else contextlib.nullcontext())
-    semaphore = (asyncio.Semaphore(max_concurrency)
-                 if max_concurrency else None)
+    semaphore = asyncio.Semaphore(max_concurrency) if max_concurrency else None
 
     async def limited_request_func(request_func_input, pbar):
         if semaphore is None:
-            return await request_func(request_func_input=request_func_input,
-                                      pbar=pbar)
+            return await request_func(request_func_input=request_func_input, pbar=pbar)
         async with semaphore:
-            return await request_func(request_func_input=request_func_input,
-                                      pbar=pbar)
+            return await request_func(request_func_input=request_func_input, pbar=pbar)
 
     benchmark_start_time = time.perf_counter()
     tasks: list[asyncio.Task] = []
     expected: list[str] = []
-    async for i, request in get_request(input_requests, request_rate,
-                                        burstiness):
-        extra_body = prepare_extra_body(
-            request) if i in structured_output_req_idx else None
+    async for i, request in get_request(input_requests, request_rate, burstiness):
+        extra_body = (
+            prepare_extra_body(request) if i in structured_output_req_idx else None
+        )
         request_func_input = RequestFuncInput(
             model=model_id,
             prompt=request.prompt,
@@ -512,8 +533,9 @@ async def benchmark(
         expected.append(request.completion)
         tasks.append(
             asyncio.create_task(
-                limited_request_func(request_func_input=request_func_input,
-                                     pbar=pbar)))
+                limited_request_func(request_func_input=request_func_input, pbar=pbar)
+            )
+        )
     outputs: list[RequestFuncOutput] = await asyncio.gather(*tasks)
 
     if profile:
@@ -545,54 +567,58 @@ async def benchmark(
         goodput_config_dict=goodput_config_dict,
     )
 
-    print("{s:{c}^{n}}".format(s=' Serving Benchmark Result ', n=50, c='='))
+    print("{s:{c}^{n}}".format(s=" Serving Benchmark Result ", n=50, c="="))
     print("{:<40} {:<10}".format("Successful requests:", metrics.completed))
-    print("{:<40} {:<10.2f}".format("Benchmark duration (s):",
-                                    benchmark_duration))
+    print("{:<40} {:<10.2f}".format("Benchmark duration (s):", benchmark_duration))
     print("{:<40} {:<10}".format("Total input tokens:", metrics.total_input))
-    print("{:<40} {:<10}".format("Total generated tokens:",
-                                 metrics.total_output))
-    print("{:<40} {:<10.2f}".format("Request throughput (req/s):",
-                                    metrics.request_throughput))
+    print("{:<40} {:<10}".format("Total generated tokens:", metrics.total_output))
+    print(
+        "{:<40} {:<10.2f}".format(
+            "Request throughput (req/s):", metrics.request_throughput
+        )
+    )
     if goodput_config_dict:
-        print("{:<40} {:<10.2f}".format("Request goodput (req/s):",
-                                        metrics.request_goodput))
-    print("{:<40} {:<10.2f}".format("Output token throughput (tok/s):",
-                                    metrics.output_throughput))
-    print("{:<40} {:<10.2f}".format("Total Token throughput (tok/s):",
-                                    metrics.total_token_throughput))
+        print(
+            "{:<40} {:<10.2f}".format(
+                "Request goodput (req/s):", metrics.request_goodput
+            )
+        )
+    print(
+        "{:<40} {:<10.2f}".format(
+            "Output token throughput (tok/s):", metrics.output_throughput
+        )
+    )
+    print(
+        "{:<40} {:<10.2f}".format(
+            "Total Token throughput (tok/s):", metrics.total_token_throughput
+        )
+    )
 
     result = {
-        "duration":
-        benchmark_duration,
-        "completed":
-        metrics.completed,
-        "total_input_tokens":
-        metrics.total_input,
-        "total_output_tokens":
-        metrics.total_output,
-        "request_throughput":
-        metrics.request_throughput,
-        "output_throughput":
-        metrics.output_throughput,
-        "total_token_throughput":
-        metrics.total_token_throughput,
-        "ttft_description":
-        pd.Series([output.ttft for output in outputs]).describe().to_dict(),
-        "tpot_description":
-        pd.Series([output.tpot for output in outputs]).describe().to_dict(),
+        "duration": benchmark_duration,
+        "completed": metrics.completed,
+        "total_input_tokens": metrics.total_input,
+        "total_output_tokens": metrics.total_output,
+        "request_throughput": metrics.request_throughput,
+        "output_throughput": metrics.output_throughput,
+        "total_token_throughput": metrics.total_token_throughput,
+        "ttft_description": pd.Series([output.ttft for output in outputs])
+        .describe()
+        .to_dict(),
+        "tpot_description": pd.Series([output.tpot for output in outputs])
+        .describe()
+        .to_dict(),
         "input_lens": [output.prompt_len for output in outputs],
-        "output_lens":
-        actual_output_lens,
+        "output_lens": actual_output_lens,
         "ttfts": [output.ttft for output in outputs],
         "itls": [output.itl for output in outputs],
         "errors": [output.error for output in outputs],
     }
 
-    ret = [{
-        'generated': output.generated_text,
-        'expected': gt
-    } for output, gt in zip(outputs, expected)]
+    ret = [
+        {"generated": output.generated_text, "expected": gt}
+        for output, gt in zip(outputs, expected)
+    ]
 
     def process_one_metric(
         # E.g., "ttft"
@@ -606,29 +632,35 @@ async def benchmark(
         # metric.
         if metric_attribute_name not in selected_percentile_metrics:
             return
-        print("{s:{c}^{n}}".format(s=metric_header, n=50, c='-'))
-        print("{:<40} {:<10.2f}".format(
-            f"Mean {metric_name} (ms):",
-            getattr(metrics, f"mean_{metric_attribute_name}_ms")))
-        print("{:<40} {:<10.2f}".format(
-            f"Median {metric_name} (ms):",
-            getattr(metrics, f"median_{metric_attribute_name}_ms")))
+        print("{s:{c}^{n}}".format(s=metric_header, n=50, c="-"))
+        print(
+            "{:<40} {:<10.2f}".format(
+                f"Mean {metric_name} (ms):",
+                getattr(metrics, f"mean_{metric_attribute_name}_ms"),
+            )
+        )
+        print(
+            "{:<40} {:<10.2f}".format(
+                f"Median {metric_name} (ms):",
+                getattr(metrics, f"median_{metric_attribute_name}_ms"),
+            )
+        )
         result[f"mean_{metric_attribute_name}_ms"] = getattr(
-            metrics, f"mean_{metric_attribute_name}_ms")
+            metrics, f"mean_{metric_attribute_name}_ms"
+        )
         result[f"median_{metric_attribute_name}_ms"] = getattr(
-            metrics, f"median_{metric_attribute_name}_ms")
+            metrics, f"median_{metric_attribute_name}_ms"
+        )
         result[f"std_{metric_attribute_name}_ms"] = getattr(
-            metrics, f"std_{metric_attribute_name}_ms")
-        for p, value in getattr(metrics,
-                                f"percentiles_{metric_attribute_name}_ms"):
+            metrics, f"std_{metric_attribute_name}_ms"
+        )
+        for p, value in getattr(metrics, f"percentiles_{metric_attribute_name}_ms"):
             p_word = str(int(p)) if int(p) == p else str(p)
-            print("{:<40} {:<10.2f}".format(f"P{p_word} {metric_name} (ms):",
-                                            value))
+            print("{:<40} {:<10.2f}".format(f"P{p_word} {metric_name} (ms):", value))
             result[f"p{p_word}_{metric_attribute_name}_ms"] = value
 
     process_one_metric("ttft", "TTFT", "Time to First Token")
-    process_one_metric("tpot", "TPOT",
-                       "Time per Output Token (excl. 1st token)")
+    process_one_metric("tpot", "TPOT", "Time per Output Token (excl. 1st token)")
     process_one_metric("itl", "ITL", "Inter-token Latency")
     process_one_metric("e2el", "E2EL", "End-to-end Latency")
 
@@ -638,13 +670,13 @@ async def benchmark(
 
 
 def evaluate(ret, args):
-
     def _eval_correctness_json(expected, actual):
         # extract json string from string using regex
-        import re
-        actual = actual.replace('\n', '').replace(' ', '').strip()
+        import regex as re
+
+        actual = actual.replace("\n", "").replace(" ", "").strip()
         try:
-            actual = re.search(r'\{.*\}', actual).group()
+            actual = re.search(r"\{.*\}", actual).group()
             actual = json.loads(actual)
         except Exception:
             return False
@@ -655,29 +687,33 @@ def evaluate(ret, args):
         return actual in args.choice
 
     def _eval_correctness_regex(expected, actual):
-        import re
+        import regex as re
+
         return re.match(args.regex, actual) is not None
 
     def _eval_correctness(expected, actual):
-        if args.structure_type == 'guided_json':
+        if args.structure_type == "guided_json":
             return _eval_correctness_json(expected, actual)
-        elif args.structure_type == 'guided_regex':
+        elif args.structure_type == "guided_regex":
             return _eval_correctness_regex(expected, actual)
-        elif args.structure_type == 'guided_choice':
+        elif args.structure_type == "guided_choice":
             return _eval_correctness_choice(expected, actual)
         else:
             return None
 
     scores = []
     for res in ret:
-        score = _eval_correctness(res['expected'], res['generated'])
-        res['correctness'] = score
+        score = _eval_correctness(res["expected"], res["generated"])
+        res["correctness"] = score
         scores.append(score)
 
     not_none_scores = [score for score in scores if score is not None]
 
-    return (sum(not_none_scores) / len(not_none_scores) *
-            100) if len(not_none_scores) > 0 else None
+    return (
+        (sum(not_none_scores) / len(not_none_scores) * 100)
+        if len(not_none_scores) > 0
+        else None
+    )
 
 
 def parse_goodput(slo_pairs):
@@ -689,9 +725,10 @@ def parse_goodput(slo_pairs):
     except ValueError as err:
         raise argparse.ArgumentTypeError(
             "Invalid format found for service level objectives. "
-            "Specify service level objectives for goodput as \"KEY:VALUE\" "
+            'Specify service level objectives for goodput as "KEY:VALUE" '
             "pairs, where the key is a metric name, and the value is a "
-            "number in milliseconds.") from err
+            "number in milliseconds."
+        ) from err
     return goodput_config_dict
 
 
@@ -705,12 +742,14 @@ def check_goodput_args(args):
                 raise ValueError(
                     f"Invalid metric name found, {slo_name}: {slo_val}. "
                     "The service level objective name should be one of "
-                    f"{str(VALID_NAMES)}. ")
+                    f"{str(VALID_NAMES)}. "
+                )
             if slo_val < 0:
                 raise ValueError(
                     f"Invalid value found, {slo_name}: {slo_val}. "
                     "The service level objective value should be "
-                    "non-negative.")
+                    "non-negative."
+                )
     return goodput_config_dict
 
 
@@ -736,19 +775,19 @@ def main(args: argparse.Namespace):
         tokenizer_mode=args.tokenizer_mode,
     )
 
-    if args.dataset == 'grammar':
-        args.structure_type = 'guided_grammar'
-    elif args.dataset == 'regex':
-        args.structure_type = 'guided_regex'
-    elif args.dataset == 'choice':
-        args.structure_type = 'guided_choice'
+    if args.dataset == "grammar":
+        args.structure_type = "guided_grammar"
+    elif args.dataset == "regex":
+        args.structure_type = "guided_regex"
+    elif args.dataset == "choice":
+        args.structure_type = "guided_choice"
     else:
-        args.structure_type = 'guided_json'
+        args.structure_type = "guided_json"
 
     if args.no_structured_output:
         args.structured_output_ratio = 0
     if args.save_results:
-        result_file_name = f'{args.structured_output_ratio}guided'
+        result_file_name = f"{args.structured_output_ratio}guided"
         result_file_name += f"_{backend}"
         result_file_name += f"_{args.request_rate}qps"
         result_file_name += f"_{args.model.split('/')[-1]}"
@@ -776,37 +815,29 @@ def main(args: argparse.Namespace):
             disable_tqdm=args.disable_tqdm,
             profile=args.profile,
             selected_percentile_metrics=args.percentile_metrics.split(","),
-            selected_percentiles=[
-                float(p) for p in args.metric_percentiles.split(",")
-            ],
+            selected_percentiles=[float(p) for p in args.metric_percentiles.split(",")],
             ignore_eos=args.ignore_eos,
             max_concurrency=args.max_concurrency,
             structured_output_ratio=args.structured_output_ratio,
-            structured_output_backend=args.structured_output_backend,
             goodput_config_dict=goodput_config_dict,
-        ))
+        )
+    )
 
     # Save config and results to json
     score = evaluate(ret, args)
-    print("correct_rate(%)", score, '\n')
+    print("correct_rate(%)", score, "\n")
     if args.save_results:
         results = {
-            "backend":
-            backend,
-            "model_id":
-            model_id,
-            "tokenizer_id":
-            tokenizer_id,
-            "num_prompts":
-            args.num_prompts,
-            "request_rate":
-            args.request_rate if args.request_rate < float("inf") else "inf",
-            "burstiness":
-            args.burstiness,
-            "max_concurrency":
-            args.max_concurrency,
-            "correct_rate(%)":
-            score
+            "backend": backend,
+            "model_id": model_id,
+            "tokenizer_id": tokenizer_id,
+            "num_prompts": args.num_prompts,
+            "request_rate": args.request_rate
+            if args.request_rate < float("inf")
+            else "inf",
+            "burstiness": args.burstiness,
+            "max_concurrency": args.max_concurrency,
+            "correct_rate(%)": score,
         }
         results = {"outputs": ret, **results, **benchmark_result}
 
@@ -815,13 +846,14 @@ def main(args: argparse.Namespace):
             result_file_name = args.result_filename
         if args.result_dir:
             result_file_name = os.path.join(args.result_dir, result_file_name)
-        with open(result_file_name, "w", encoding='utf-8') as outfile:
+        with open(result_file_name, "w", encoding="utf-8") as outfile:
             json.dump(results, outfile, indent=4)
 
 
-if __name__ == "__main__":
+def create_argument_parser():
     parser = FlexibleArgumentParser(
-        description="Benchmark the online serving throughput.")
+        description="Benchmark the online serving throughput."
+    )
     parser.add_argument(
         "--backend",
         type=str,
@@ -843,16 +875,14 @@ if __name__ == "__main__":
         default="/v1/completions",
         help="API endpoint.",
     )
-    parser.add_argument("--dataset",
-                        default='json',
-                        choices=[
-                            'json', 'json-unique', 'grammar', 'regex',
-                            'choice', 'xgrammar_bench'
-                        ])
-    parser.add_argument("--json_schema_path",
-                        type=str,
-                        default=None,
-                        help="Path to json schema.")
+    parser.add_argument(
+        "--dataset",
+        default="json",
+        choices=["json", "json-unique", "grammar", "regex", "choice", "xgrammar_bench"],
+    )
+    parser.add_argument(
+        "--json-schema-path", type=str, default=None, help="Path to json schema."
+    )
     parser.add_argument(
         "--max-concurrency",
         type=int,
@@ -864,7 +894,8 @@ if __name__ == "__main__":
         "initiated, this argument will control how many are actually allowed "
         "to execute at a time. This means that when used in combination, the "
         "actual request rate may be lower than specified with --request-rate, "
-        "if the server is not processing requests fast enough to keep up.")
+        "if the server is not processing requests fast enough to keep up.",
+    )
     parser.add_argument(
         "--model",
         type=str,
@@ -874,15 +905,13 @@ if __name__ == "__main__":
     parser.add_argument(
         "--tokenizer",
         type=str,
-        help=
-        "Name or path of the tokenizer, if not using the default tokenizer.",  # noqa: E501
+        help="Name or path of the tokenizer, if not using the default tokenizer.",  # noqa: E501
     )
     parser.add_argument(
         "--tokenizer-mode",
         type=str,
         default="auto",
-        help=
-        "Name or path of the tokenizer, if not using the default tokenizer.",  # noqa: E501
+        help="Name or path of the tokenizer, if not using the default tokenizer.",  # noqa: E501
     )
     parser.add_argument(
         "--num-prompts",
@@ -959,52 +988,56 @@ if __name__ == "__main__":
         "--ignore-eos",
         action="store_true",
         help="Set ignore_eos flag when sending the benchmark request."
-        "Warning: ignore_eos is not supported in deepspeed_mii and tgi.")
+        "Warning: ignore_eos is not supported in deepspeed_mii and tgi.",
+    )
     parser.add_argument(
         "--percentile-metrics",
         type=str,
         default="ttft,tpot,itl",
         help="Comma-separated list of selected metrics to report percentils. "
         "This argument specifies the metrics to report percentiles. "
-        "Allowed metric names are \"ttft\", \"tpot\", \"itl\", \"e2el\". "
-        "Default value is \"ttft,tpot,itl\".")
+        'Allowed metric names are "ttft", "tpot", "itl", "e2el". '
+        'Default value is "ttft,tpot,itl".',
+    )
     parser.add_argument(
         "--metric-percentiles",
         type=str,
         default="99",
         help="Comma-separated list of percentiles for selected metrics. "
-        "To report 25-th, 50-th, and 75-th percentiles, use \"25,50,75\". "
-        "Default value is \"99\". "
-        "Use \"--percentile-metrics\" to select metrics.",
+        'To report 25-th, 50-th, and 75-th percentiles, use "25,50,75". '
+        'Default value is "99". '
+        'Use "--percentile-metrics" to select metrics.',
     )
     parser.add_argument(
         "--goodput",
         nargs="+",
         required=False,
-        help="Specify service level objectives for goodput as \"KEY:VALUE\" "
+        help='Specify service level objectives for goodput as "KEY:VALUE" '
         "pairs, where the key is a metric name, and the value is in "
-        "milliseconds. Multiple \"KEY:VALUE\" pairs can be provided, "
+        'milliseconds. Multiple "KEY:VALUE" pairs can be provided, '
         "separated by spaces. Allowed request level metric names are "
-        "\"ttft\", \"tpot\", \"e2el\". For more context on the definition of "
+        '"ttft", "tpot", "e2el". For more context on the definition of '
         "goodput, refer to DistServe paper: https://arxiv.org/pdf/2401.09670 "
-        "and the blog: https://hao-ai-lab.github.io/blogs/distserve")
+        "and the blog: https://hao-ai-lab.github.io/blogs/distserve",
+    )
 
-    parser.add_argument("--no-structured-output",
-                        action='store_true',
-                        default=False,
-                        help="Whether to disable JSON decoding or not.")
-    parser.add_argument("--structured-output-ratio",
-                        type=float,
-                        default=1.0,
-                        help="Ratio of Structured Outputs requests")
-    parser.add_argument("--structured-output-backend",
-                        type=str,
-                        choices=[
-                            "outlines", "lm-format-enforcer", "xgrammar",
-                            "guidance", "auto"
-                        ],
-                        default="auto",
-                        help="Backend to use for structured outputs")
+    parser.add_argument(
+        "--no-structured-output",
+        action="store_true",
+        default=False,
+        help="Whether to disable JSON decoding or not.",
+    )
+    parser.add_argument(
+        "--structured-output-ratio",
+        type=float,
+        default=1.0,
+        help="Ratio of Structured Outputs requests",
+    )
 
+    return parser
+
+
+if __name__ == "__main__":
+    parser = create_argument_parser()
     args = parser.parse_args()
     main(args)

benchmarks/benchmark_throughput.py

@@ -1,5 +1,7 @@
 # SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 """Benchmark offline inference throughput."""
+
 import argparse
 import dataclasses
 import json
@@ -11,18 +13,25 @@ from typing import Any, Optional, Union
 
 import torch
 import uvloop
-from benchmark_dataset import (AIMODataset, BurstGPTDataset,
-                               ConversationDataset, InstructCoderDataset,
-                               RandomDataset, SampleRequest, ShareGPTDataset,
-                               SonnetDataset, VisionArenaDataset)
-from benchmark_utils import convert_to_pytorch_benchmark_format, write_to_json
 from tqdm import tqdm
-from transformers import (AutoModelForCausalLM, AutoTokenizer,
-                          PreTrainedTokenizerBase)
+from transformers import AutoModelForCausalLM, AutoTokenizer, PreTrainedTokenizerBase
 
+from benchmark_dataset import (
+    AIMODataset,
+    BurstGPTDataset,
+    ConversationDataset,
+    InstructCoderDataset,
+    RandomDataset,
+    SampleRequest,
+    ShareGPTDataset,
+    SonnetDataset,
+    VisionArenaDataset,
+)
+from benchmark_utils import convert_to_pytorch_benchmark_format, write_to_json
 from vllm.engine.arg_utils import AsyncEngineArgs, EngineArgs
 from vllm.entrypoints.openai.api_server import (
-    build_async_engine_client_from_engine_args)
+    build_async_engine_client_from_engine_args,
+)
 from vllm.inputs import TextPrompt, TokensPrompt
 from vllm.lora.request import LoRARequest
 from vllm.outputs import RequestOutput
@@ -37,23 +46,30 @@ def run_vllm(
     disable_detokenize: bool = False,
 ) -> tuple[float, Optional[list[RequestOutput]]]:
     from vllm import LLM, SamplingParams
+
     llm = LLM(**dataclasses.asdict(engine_args))
     assert all(
-        llm.llm_engine.model_config.max_model_len >= (
-            request.prompt_len + request.expected_output_len)
-        for request in requests), (
-            "Please ensure that max_model_len is greater than the sum of"
-            " prompt_len and expected_output_len for all requests.")
+        llm.llm_engine.model_config.max_model_len
+        >= (request.prompt_len + request.expected_output_len)
+        for request in requests
+    ), (
+        "Please ensure that max_model_len is greater than the sum of"
+        " prompt_len and expected_output_len for all requests."
+    )
     # Add the requests to the engine.
     prompts: list[Union[TextPrompt, TokensPrompt]] = []
     sampling_params: list[SamplingParams] = []
     for request in requests:
         prompts.append(
-            TokensPrompt(prompt_token_ids=request.prompt["prompt_token_ids"],
-                       multi_modal_data=request.multi_modal_data)
-            if "prompt_token_ids" in request.prompt else \
-            TextPrompt(prompt=request.prompt,
-                       multi_modal_data=request.multi_modal_data))
+            TokensPrompt(
+                prompt_token_ids=request.prompt["prompt_token_ids"],
+                multi_modal_data=request.multi_modal_data,
+            )
+            if "prompt_token_ids" in request.prompt
+            else TextPrompt(
+                prompt=request.prompt, multi_modal_data=request.multi_modal_data
+            )
+        )
         sampling_params.append(
             SamplingParams(
                 n=n,
@@ -62,7 +78,8 @@ def run_vllm(
                 ignore_eos=True,
                 max_tokens=request.expected_output_len,
                 detokenize=not disable_detokenize,
-            ))
+            )
+        )
     lora_requests: Optional[list[LoRARequest]] = None
     if engine_args.enable_lora:
         lora_requests = [request.lora_request for request in requests]
@@ -72,16 +89,15 @@ def run_vllm(
     outputs = None
     if not use_beam_search:
         start = time.perf_counter()
-        outputs = llm.generate(prompts,
-                               sampling_params,
-                               lora_request=lora_requests,
-                               use_tqdm=True)
+        outputs = llm.generate(
+            prompts, sampling_params, lora_request=lora_requests, use_tqdm=True
+        )
         end = time.perf_counter()
     else:
         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()
@@ -91,30 +107,35 @@ def run_vllm(
                 beam_width=n,
                 max_tokens=output_len,
                 ignore_eos=True,
-            ))
+            ),
+        )
         end = time.perf_counter()
     return end - start, outputs
 
 
 def run_vllm_chat(
-        requests: list[SampleRequest],
-        n: int,
-        engine_args: EngineArgs,
-        disable_detokenize: bool = False) -> tuple[float, list[RequestOutput]]:
+    requests: list[SampleRequest],
+    n: int,
+    engine_args: EngineArgs,
+    disable_detokenize: bool = False,
+) -> tuple[float, list[RequestOutput]]:
     """
     Run vLLM chat benchmark. This function is recommended ONLY for benchmarking
     multimodal models as it properly handles multimodal inputs and chat
     formatting. For non-multimodal models, use run_vllm() instead.
     """
     from vllm import LLM, SamplingParams
+
     llm = LLM(**dataclasses.asdict(engine_args))
 
     assert all(
-        llm.llm_engine.model_config.max_model_len >= (
-            request.prompt_len + request.expected_output_len)
-        for request in requests), (
-            "Please ensure that max_model_len is greater than the sum of "
-            "prompt_len and expected_output_len for all requests.")
+        llm.llm_engine.model_config.max_model_len
+        >= (request.prompt_len + request.expected_output_len)
+        for request in requests
+    ), (
+        "Please ensure that max_model_len is greater than the sum of "
+        "prompt_len and expected_output_len for all requests."
+    )
 
     prompts = []
     sampling_params: list[SamplingParams] = []
@@ -128,7 +149,8 @@ def run_vllm_chat(
                 ignore_eos=True,
                 max_tokens=request.expected_output_len,
                 detokenize=not disable_detokenize,
-            ))
+            )
+        )
     start = time.perf_counter()
     outputs = llm.chat(prompts, sampling_params, use_tqdm=True)
     end = time.perf_counter()
@@ -145,13 +167,17 @@ async def run_vllm_async(
     from vllm import SamplingParams
 
     async with build_async_engine_client_from_engine_args(
-            engine_args, disable_frontend_multiprocessing) as llm:
+        engine_args, disable_frontend_multiprocessing
+    ) as llm:
+        model_config = await llm.get_model_config()
         assert all(
-            llm.model_config.max_model_len >= (request.prompt_len +
-                                               request.expected_output_len)
-            for request in requests), (
-                "Please ensure that max_model_len is greater than the sum of"
-                " prompt_len and expected_output_len for all requests.")
+            model_config.max_model_len
+            >= (request.prompt_len + request.expected_output_len)
+            for request in requests
+        ), (
+            "Please ensure that max_model_len is greater than the sum of"
+            " prompt_len and expected_output_len for all requests."
+        )
 
         # Add the requests to the engine.
         prompts: list[Union[TextPrompt, TokensPrompt]] = []
@@ -159,11 +185,15 @@ async def run_vllm_async(
         lora_requests: list[Optional[LoRARequest]] = []
         for request in requests:
             prompts.append(
-                TokensPrompt(prompt_token_ids=request.prompt["prompt_token_ids"],
-                        multi_modal_data=request.multi_modal_data)
-                if "prompt_token_ids" in request.prompt else \
-                TextPrompt(prompt=request.prompt,
-                           multi_modal_data=request.multi_modal_data))
+                TokensPrompt(
+                    prompt_token_ids=request.prompt["prompt_token_ids"],
+                    multi_modal_data=request.multi_modal_data,
+                )
+                if "prompt_token_ids" in request.prompt
+                else TextPrompt(
+                    prompt=request.prompt, multi_modal_data=request.multi_modal_data
+                )
+            )
             sampling_params.append(
                 SamplingParams(
                     n=n,
@@ -172,17 +202,16 @@ async def run_vllm_async(
                     ignore_eos=True,
                     max_tokens=request.expected_output_len,
                     detokenize=not disable_detokenize,
-                ))
+                )
+            )
             lora_requests.append(request.lora_request)
 
         generators = []
         start = time.perf_counter()
-        for i, (prompt, sp,
-                lr) in enumerate(zip(prompts, sampling_params, lora_requests)):
-            generator = llm.generate(prompt,
-                                     sp,
-                                     lora_request=lr,
-                                     request_id=f"test{i}")
+        for i, (prompt, sp, lr) in enumerate(
+            zip(prompts, sampling_params, lora_requests)
+        ):
+            generator = llm.generate(prompt, sp, lora_request=lr, request_id=f"test{i}")
             generators.append(generator)
         all_gens = merge_async_iterators(*generators)
         async for i, res in all_gens:
@@ -201,7 +230,8 @@ def run_hf(
     disable_detokenize: bool = False,
 ) -> float:
     llm = AutoModelForCausalLM.from_pretrained(
-        model, torch_dtype=torch.float16, trust_remote_code=trust_remote_code)
+        model, torch_dtype=torch.float16, trust_remote_code=trust_remote_code
+    )
     if llm.config.model_type == "llama":
         # To enable padding in the HF backend.
         tokenizer.pad_token = tokenizer.eos_token
@@ -224,14 +254,15 @@ def run_hf(
             # Check if we can add more requests to the batch.
             next_prompt_len = requests[i + 1].prompt_len
             next_output_len = requests[i + 1].expected_output_len
-            if (max(max_prompt_len, next_prompt_len) +
-                    max(max_output_len, next_output_len)) <= 2048:
+            if (
+                max(max_prompt_len, next_prompt_len)
+                + max(max_output_len, next_output_len)
+            ) <= 2048:
                 # We can add more requests to the batch.
                 continue
 
         # Generate the sequences.
-        input_ids = tokenizer(batch, return_tensors="pt",
-                              padding=True).input_ids
+        input_ids = tokenizer(batch, return_tensors="pt", padding=True).input_ids
         llm_outputs = llm.generate(
             input_ids=input_ids.cuda(),
             do_sample=True,
@@ -261,6 +292,7 @@ def run_mii(
     output_len: int,
 ) -> float:
     from mii import client, serve
+
     llm = serve(model, tensor_parallel=tensor_parallel_size)
     prompts = [request.prompt for request in requests]
 
@@ -272,8 +304,9 @@ def run_mii(
     return end - start
 
 
-def save_to_pytorch_benchmark_format(args: argparse.Namespace,
-                                     results: dict[str, Any]) -> None:
+def save_to_pytorch_benchmark_format(
+    args: argparse.Namespace, results: dict[str, Any]
+) -> None:
     pt_records = convert_to_pytorch_benchmark_format(
         args=args,
         metrics={
@@ -281,9 +314,9 @@ def save_to_pytorch_benchmark_format(args: argparse.Namespace,
             "tokens_per_second": [results["tokens_per_second"]],
         },
         extra_info={
-            k: results[k]
-            for k in ["elapsed_time", "num_requests", "total_num_tokens"]
-        })
+            k: results[k] for k in ["elapsed_time", "num_requests", "total_num_tokens"]
+        },
+    )
     if pt_records:
         # Don't use json suffix here as we don't want CI to pick it up
         pt_file = f"{os.path.splitext(args.output_json)[0]}.pytorch.json"
@@ -315,30 +348,32 @@ def get_requests(args, tokenizer):
             sample_kwargs["enable_multimodal_chat"] = True
     elif args.dataset_name == "sonnet":
         assert tokenizer.chat_template or tokenizer.default_chat_template, (
-            "Tokenizer/model must have chat template for sonnet dataset.")
+            "Tokenizer/model must have chat template for sonnet dataset."
+        )
         dataset_cls = SonnetDataset
         sample_kwargs["prefix_len"] = args.prefix_len
         sample_kwargs["return_prompt_formatted"] = True
     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
-            common_kwargs['dataset_split'] = "train"
+            common_kwargs["dataset_subset"] = None
+            common_kwargs["dataset_split"] = "train"
             sample_kwargs["enable_multimodal_chat"] = True
         elif args.dataset_path in InstructCoderDataset.SUPPORTED_DATASET_PATHS:
             dataset_cls = InstructCoderDataset
-            common_kwargs['dataset_split'] = "train"
+            common_kwargs["dataset_split"] = "train"
         elif args.dataset_path in ConversationDataset.SUPPORTED_DATASET_PATHS:
             dataset_cls = ConversationDataset
-            common_kwargs['dataset_subset'] = args.hf_subset
-            common_kwargs['dataset_split'] = args.hf_split
+            common_kwargs["dataset_subset"] = args.hf_subset
+            common_kwargs["dataset_split"] = args.hf_split
             sample_kwargs["enable_multimodal_chat"] = True
         elif args.dataset_path in AIMODataset.SUPPORTED_DATASET_PATHS:
             dataset_cls = AIMODataset
-            common_kwargs['dataset_subset'] = None
-            common_kwargs['dataset_split'] = "train"
+            common_kwargs["dataset_subset"] = None
+            common_kwargs["dataset_split"] = "train"
     else:
         raise ValueError(f"Unknown dataset name: {args.dataset_name}")
     # Remove None values
@@ -353,10 +388,10 @@ def main(args: argparse.Namespace):
     random.seed(args.seed)
     # Sample the requests.
     tokenizer = AutoTokenizer.from_pretrained(
-        args.tokenizer, trust_remote_code=args.trust_remote_code)
+        args.tokenizer, trust_remote_code=args.trust_remote_code
+    )
     requests = get_requests(args, tokenizer)
-    is_multi_modal = any(request.multi_modal_data is not None
-                         for request in requests)
+    is_multi_modal = any(request.multi_modal_data is not None for request in requests)
     request_outputs: Optional[list[RequestOutput]] = None
     if args.backend == "vllm":
         if args.async_engine:
@@ -367,23 +402,34 @@ def main(args: argparse.Namespace):
                     AsyncEngineArgs.from_cli_args(args),
                     args.disable_frontend_multiprocessing,
                     args.disable_detokenize,
-                ))
+                )
+            )
         else:
             elapsed_time, request_outputs = run_vllm(
-                requests, args.n, EngineArgs.from_cli_args(args),
-                args.disable_detokenize)
+                requests,
+                args.n,
+                EngineArgs.from_cli_args(args),
+                args.disable_detokenize,
+            )
     elif args.backend == "hf":
         assert args.tensor_parallel_size == 1
-        elapsed_time = run_hf(requests, args.model, tokenizer, args.n,
-                              args.hf_max_batch_size, args.trust_remote_code,
-                              args.disable_detokenize)
+        elapsed_time = run_hf(
+            requests,
+            args.model,
+            tokenizer,
+            args.n,
+            args.hf_max_batch_size,
+            args.trust_remote_code,
+            args.disable_detokenize,
+        )
     elif args.backend == "mii":
-        elapsed_time = run_mii(requests, args.model, args.tensor_parallel_size,
-                               args.output_len)
+        elapsed_time = run_mii(
+            requests, args.model, args.tensor_parallel_size, args.output_len
+        )
     elif args.backend == "vllm-chat":
         elapsed_time, request_outputs = run_vllm_chat(
-            requests, args.n, EngineArgs.from_cli_args(args),
-            args.disable_detokenize)
+            requests, args.n, EngineArgs.from_cli_args(args), args.disable_detokenize
+        )
     else:
         raise ValueError(f"Unknown backend: {args.backend}")
 
@@ -395,28 +441,31 @@ def main(args: argparse.Namespace):
         for ro in request_outputs:
             if not isinstance(ro, RequestOutput):
                 continue
-            total_prompt_tokens += len(
-                ro.prompt_token_ids) if ro.prompt_token_ids else 0
-            total_output_tokens += sum(
-                len(o.token_ids) for o in ro.outputs if o)
+            total_prompt_tokens += (
+                len(ro.prompt_token_ids) if ro.prompt_token_ids else 0
+            )
+            total_output_tokens += sum(len(o.token_ids) for o in ro.outputs if o)
         total_num_tokens = total_prompt_tokens + total_output_tokens
     else:
-        total_num_tokens = sum(r.prompt_len + r.expected_output_len
-                               for r in requests)
+        total_num_tokens = sum(r.prompt_len + r.expected_output_len for r in requests)
         total_output_tokens = sum(r.expected_output_len for r in requests)
         total_prompt_tokens = total_num_tokens - total_output_tokens
 
     if is_multi_modal and args.backend != "vllm-chat":
-        print("\033[91mWARNING\033[0m: Multi-modal request with "
-              f"{args.backend} backend detected. The "
-              "following metrics are not accurate because image tokens are not"
-              " counted. See vllm-project/vllm/issues/9778 for details.")
+        print(
+            "\033[91mWARNING\033[0m: Multi-modal request with "
+            f"{args.backend} backend detected. The "
+            "following metrics are not accurate because image tokens are not"
+            " counted. See vllm-project/vllm/issues/9778 for details."
+        )
         # TODO(vllm-project/vllm/issues/9778): Count multi-modal token length.
         # vllm-chat backend counts the image tokens now
 
-    print(f"Throughput: {len(requests) / elapsed_time:.2f} requests/s, "
-          f"{total_num_tokens / elapsed_time:.2f} total tokens/s, "
-          f"{total_output_tokens / elapsed_time:.2f} output tokens/s")
+    print(
+        f"Throughput: {len(requests) / elapsed_time:.2f} requests/s, "
+        f"{total_num_tokens / elapsed_time:.2f} total tokens/s, "
+        f"{total_output_tokens / elapsed_time:.2f} output tokens/s"
+    )
     print(f"Total num prompt tokens:  {total_prompt_tokens}")
     print(f"Total num output tokens:  {total_output_tokens}")
 
@@ -444,7 +493,8 @@ def validate_args(args):
         warnings.warn(
             "The '--dataset' argument will be deprecated in the next release. "
             "Please use '--dataset-name' and '--dataset-path' instead.",
-            stacklevel=2)
+            stacklevel=2,
+        )
         args.dataset_path = args.dataset
 
     if not getattr(args, "tokenizer", None):
@@ -457,9 +507,8 @@ def validate_args(args):
 
     # === Dataset Configuration ===
     if not args.dataset and not args.dataset_path:
-        print(
-            "When dataset path is not set, it will default to random dataset")
-        args.dataset_name = 'random'
+        print("When dataset path is not set, it will default to random dataset")
+        args.dataset_name = "random"
         if args.input_len is None:
             raise ValueError("input_len must be provided for a random dataset")
 
@@ -467,41 +516,55 @@ def validate_args(args):
     # --hf-subset and --hf-split: only used
     # when dataset_name is 'hf'
     if args.dataset_name != "hf" and (
-            getattr(args, "hf_subset", None) is not None
-            or getattr(args, "hf_split", None) is not None):
-        warnings.warn("--hf-subset and --hf-split will be ignored \
+        getattr(args, "hf_subset", None) is not None
+        or getattr(args, "hf_split", None) is not None
+    ):
+        warnings.warn(
+            "--hf-subset and --hf-split will be ignored \
                 since --dataset-name is not 'hf'.",
-                      stacklevel=2)
+            stacklevel=2,
+        )
     elif args.dataset_name == "hf":
         if args.dataset_path in (
-                VisionArenaDataset.SUPPORTED_DATASET_PATHS.keys()
-                | ConversationDataset.SUPPORTED_DATASET_PATHS):
-            assert args.backend == "vllm-chat", f"{args.dataset_path} needs to use vllm-chat as the backend."  #noqa: E501
-        elif args.dataset_path in (InstructCoderDataset.SUPPORTED_DATASET_PATHS
-                                   | AIMODataset.SUPPORTED_DATASET_PATHS):
-            assert args.backend == "vllm", f"{args.dataset_path} needs to use vllm as the backend."  #noqa: E501
+            VisionArenaDataset.SUPPORTED_DATASET_PATHS.keys()
+            | ConversationDataset.SUPPORTED_DATASET_PATHS
+        ):
+            assert args.backend == "vllm-chat", (
+                f"{args.dataset_path} needs to use vllm-chat as the backend."
+            )  # noqa: E501
+        elif args.dataset_path in (
+            InstructCoderDataset.SUPPORTED_DATASET_PATHS
+            | AIMODataset.SUPPORTED_DATASET_PATHS
+        ):
+            assert args.backend == "vllm", (
+                f"{args.dataset_path} needs to use vllm as the backend."
+            )  # noqa: E501
         else:
-            raise ValueError(
-                f"{args.dataset_path} is not supported by hf dataset.")
+            raise ValueError(f"{args.dataset_path} is not supported by hf dataset.")
 
     # --random-range-ratio: only used when dataset_name is 'random'
-    if args.dataset_name != 'random' and args.random_range_ratio is not None:
-        warnings.warn("--random-range-ratio will be ignored since \
+    if args.dataset_name != "random" and args.random_range_ratio is not None:
+        warnings.warn(
+            "--random-range-ratio will be ignored since \
                 --dataset-name is not 'random'.",
-                      stacklevel=2)
+            stacklevel=2,
+        )
 
     # --prefix-len: only used when dataset_name is 'random', 'sonnet', or not
     # set.
-    if args.dataset_name not in {"random", "sonnet", None
-                                 } and args.prefix_len is not None:
-        warnings.warn("--prefix-len will be ignored since --dataset-name\
+    if (
+        args.dataset_name not in {"random", "sonnet", None}
+        and args.prefix_len is not None
+    ):
+        warnings.warn(
+            "--prefix-len will be ignored since --dataset-name\
                  is not 'random', 'sonnet', or not set.",
-                      stacklevel=2)
+            stacklevel=2,
+        )
 
     # === LoRA Settings ===
     if getattr(args, "enable_lora", False) and args.backend != "vllm":
-        raise ValueError(
-            "LoRA benchmarking is only supported for vLLM backend")
+        raise ValueError("LoRA benchmarking is only supported for vLLM backend")
     if getattr(args, "enable_lora", False) and args.lora_path is None:
         raise ValueError("LoRA path must be provided when enable_lora is True")
 
@@ -511,8 +574,10 @@ def validate_args(args):
     if args.backend != "hf" and args.hf_max_batch_size is not None:
         raise ValueError("HF max batch size is only for HF backend.")
 
-    if args.backend in {"hf", "mii"} and getattr(args, "quantization",
-                                                 None) is not None:
+    if (
+        args.backend in {"hf", "mii"}
+        and getattr(args, "quantization", None) is not None
+    ):
         raise ValueError("Quantization is only for vLLM backend.")
 
     if args.backend == "mii" and args.dtype != "auto":
@@ -520,29 +585,37 @@ def validate_args(args):
     if args.backend == "mii" and args.n != 1:
         raise ValueError("n must be 1 for MII backend.")
     if args.backend == "mii" and args.tokenizer != args.model:
-        raise ValueError(
-            "Tokenizer must be the same as the model for MII backend.")
+        raise ValueError("Tokenizer must be the same as the model for MII backend.")
 
     # --data-parallel is not supported currently.
     # https://github.com/vllm-project/vllm/issues/16222
     if args.data_parallel_size > 1:
         raise ValueError(
             "Data parallel is not supported in offline benchmark, \
-            please use benchmark serving instead")
+            please use benchmark serving instead"
+        )
 
 
-if __name__ == "__main__":
+def create_argument_parser():
     parser = FlexibleArgumentParser(description="Benchmark the throughput.")
-    parser.add_argument("--backend",
-                        type=str,
-                        choices=["vllm", "hf", "mii", "vllm-chat"],
-                        default="vllm")
+    parser.add_argument(
+        "--backend",
+        type=str,
+        choices=["vllm", "hf", "mii", "vllm-chat"],
+        default="vllm",
+    )
     parser.add_argument(
         "--dataset-name",
         type=str,
         choices=["sharegpt", "random", "sonnet", "burstgpt", "hf"],
         help="Name of the dataset to benchmark on.",
-        default="sharegpt")
+        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,
@@ -550,57 +623,70 @@ if __name__ == "__main__":
         help="Path to the ShareGPT dataset, will be deprecated in\
             the next release. The dataset is expected to "
         "be a json in form of list[dict[..., conversations: "
-        "list[dict[..., value: <prompt_or_response>]]]]")
-    parser.add_argument("--dataset-path",
-                        type=str,
-                        default=None,
-                        help="Path to the dataset")
-    parser.add_argument("--input-len",
-                        type=int,
-                        default=None,
-                        help="Input prompt length for each request")
-    parser.add_argument("--output-len",
-                        type=int,
-                        default=None,
-                        help="Output length for each request. Overrides the "
-                        "output length from the dataset.")
-    parser.add_argument("--n",
-                        type=int,
-                        default=1,
-                        help="Number of generated sequences per prompt.")
-    parser.add_argument("--num-prompts",
-                        type=int,
-                        default=1000,
-                        help="Number of prompts to process.")
-    parser.add_argument("--hf-max-batch-size",
-                        type=int,
-                        default=None,
-                        help="Maximum batch size for HF backend.")
+        "list[dict[..., value: <prompt_or_response>]]]]",
+    )
     parser.add_argument(
-        '--output-json',
+        "--dataset-path", type=str, default=None, help="Path to the dataset"
+    )
+    parser.add_argument(
+        "--input-len",
+        type=int,
+        default=None,
+        help="Input prompt length for each request",
+    )
+    parser.add_argument(
+        "--output-len",
+        type=int,
+        default=None,
+        help="Output length for each request. Overrides the "
+        "output length from the dataset.",
+    )
+    parser.add_argument(
+        "--n", type=int, default=1, help="Number of generated sequences per prompt."
+    )
+    parser.add_argument(
+        "--num-prompts", type=int, default=1000, help="Number of prompts to process."
+    )
+    parser.add_argument(
+        "--hf-max-batch-size",
+        type=int,
+        default=None,
+        help="Maximum batch size for HF backend.",
+    )
+    parser.add_argument(
+        "--output-json",
         type=str,
         default=None,
-        help='Path to save the throughput results in JSON format.')
-    parser.add_argument("--async-engine",
-                        action='store_true',
-                        default=False,
-                        help="Use vLLM async engine rather than LLM class.")
-    parser.add_argument("--disable-frontend-multiprocessing",
-                        action='store_true',
-                        default=False,
-                        help="Disable decoupled async engine frontend.")
+        help="Path to save the throughput results in JSON format.",
+    )
+    parser.add_argument(
+        "--async-engine",
+        action="store_true",
+        default=False,
+        help="Use vLLM async engine rather than LLM class.",
+    )
+    parser.add_argument(
+        "--disable-frontend-multiprocessing",
+        action="store_true",
+        default=False,
+        help="Disable decoupled async engine frontend.",
+    )
     parser.add_argument(
         "--disable-detokenize",
         action="store_true",
-        help=("Do not detokenize the response (i.e. do not include "
-              "detokenization time in the measurement)"))
+        help=(
+            "Do not detokenize the response (i.e. do not include "
+            "detokenization time in the measurement)"
+        ),
+    )
     # LoRA
     parser.add_argument(
         "--lora-path",
         type=str,
         default=None,
-        help="Path to the lora adapters to use. This can be an absolute path, "
-        "a relative path, or a Hugging Face model identifier.")
+        help="Path to the LoRA adapters to use. This can be an absolute path, "
+        "a relative path, or a Hugging Face model identifier.",
+    )
     parser.add_argument(
         "--prefix-len",
         type=int,
@@ -614,7 +700,8 @@ if __name__ == "__main__":
         f"prefix_len (default: {SonnetDataset.DEFAULT_PREFIX_LEN}) "
         "controls how much of the input is fixed lines versus "
         "random lines, but the total input length remains approximately "
-        "input_len tokens.")
+        "input_len tokens.",
+    )
     # random dataset
     parser.add_argument(
         "--random-range-ratio",
@@ -628,16 +715,20 @@ if __name__ == "__main__":
     )
 
     # hf dtaset
-    parser.add_argument("--hf-subset",
-                        type=str,
-                        default=None,
-                        help="Subset of the HF dataset.")
-    parser.add_argument("--hf-split",
-                        type=str,
-                        default=None,
-                        help="Split of the HF dataset.")
+    parser.add_argument(
+        "--hf-subset", type=str, default=None, help="Subset of the HF dataset."
+    )
+    parser.add_argument(
+        "--hf-split", type=str, default=None, help="Split of the HF dataset."
+    )
 
     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
@@ -7,9 +8,9 @@ import os
 from typing import Any
 
 
-def convert_to_pytorch_benchmark_format(args: argparse.Namespace,
-                                        metrics: dict[str, list],
-                                        extra_info: dict[str, Any]) -> list:
+def convert_to_pytorch_benchmark_format(
+    args: argparse.Namespace, metrics: dict[str, list], extra_info: dict[str, Any]
+) -> list:
     """
     Save the benchmark results in the format used by PyTorch OSS benchmark with
     on metric per record
@@ -37,12 +38,12 @@ def convert_to_pytorch_benchmark_format(args: argparse.Namespace,
             },
         }
 
-        tp = record["benchmark"]["extra_info"]["args"].get(
-            "tensor_parallel_size")
+        tp = record["benchmark"]["extra_info"]["args"].get("tensor_parallel_size")
         # Save tensor_parallel_size parameter if it's part of the metadata
         if not tp and "tensor_parallel_size" in extra_info:
-            record["benchmark"]["extra_info"]["args"][
-                "tensor_parallel_size"] = extra_info["tensor_parallel_size"]
+            record["benchmark"]["extra_info"]["args"]["tensor_parallel_size"] = (
+                extra_info["tensor_parallel_size"]
+            )
 
         records.append(record)
 
@@ -50,7 +51,6 @@ def convert_to_pytorch_benchmark_format(args: argparse.Namespace,
 
 
 class InfEncoder(json.JSONEncoder):
-
     def clear_inf(self, o: Any):
         if isinstance(o, dict):
             return {k: self.clear_inf(v) for k, v in o.items()}
@@ -66,4 +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
@@ -23,8 +24,9 @@ DEFAULT_TP_SIZES = [1]
 
 
 # bench
-def bench_fn(label: str, sub_label: str, description: str, fn: Callable, *args,
-             **kwargs) -> TMeasurement:
+def bench_fn(
+    label: str, sub_label: str, description: str, fn: Callable, *args, **kwargs
+) -> TMeasurement:
     min_run_time = 1
 
     globals = {
@@ -41,16 +43,18 @@ def bench_fn(label: str, sub_label: str, description: str, fn: Callable, *args,
     ).blocked_autorange(min_run_time=min_run_time)
 
 
-def bench_int8(dtype: torch.dtype, m: int, k: int, n: int, label: str,
-               sub_label: str) -> Iterable[TMeasurement]:
+def bench_int8(
+    dtype: torch.dtype, m: int, k: int, n: int, label: str, sub_label: str
+) -> Iterable[TMeasurement]:
     assert dtype == torch.int8
     b_compressed, e, a, b = make_rand_sparse_tensors(torch.int8, m, n, k)
     scale_a = torch.tensor(1.0, device="cuda", dtype=torch.float32)
     scale_b = torch.tensor(1.0, device="cuda", dtype=torch.float32)
-    bias = torch.zeros((n, ), device="cuda", dtype=torch.bfloat16)
+    bias = torch.zeros((n,), device="cuda", dtype=torch.bfloat16)
 
-    out = ops.cutlass_scaled_sparse_mm(a, b_compressed, e, scale_a, scale_b,
-                                       torch.bfloat16)
+    out = ops.cutlass_scaled_sparse_mm(
+        a, b_compressed, e, scale_a, scale_b, torch.bfloat16
+    )
     out_ref = ops.cutlass_scaled_mm(a, b, scale_a, scale_b, torch.bfloat16)
 
     if not torch.allclose(out, out_ref):
@@ -63,54 +67,107 @@ def bench_int8(dtype: torch.dtype, m: int, k: int, n: int, label: str,
     timers = []
     # pytorch impl - bfloat16
     timers.append(
-        bench_fn(label, sub_label, "pytorch_bf16_bf16_bf16_matmul-no-scales",
-                 torch.mm, a.to(dtype=torch.bfloat16),
-                 b.to(dtype=torch.bfloat16)))
+        bench_fn(
+            label,
+            sub_label,
+            "pytorch_bf16_bf16_bf16_matmul-no-scales",
+            torch.mm,
+            a.to(dtype=torch.bfloat16),
+            b.to(dtype=torch.bfloat16),
+        )
+    )
 
     # pytorch impl - float16
     timers.append(
-        bench_fn(label, sub_label,
-                 "pytorch_fp16_fp16_fp16_matmul-no-scales", torch.mm,
-                 a.to(dtype=torch.float16), b.to(dtype=torch.float16)))
+        bench_fn(
+            label,
+            sub_label,
+            "pytorch_fp16_fp16_fp16_matmul-no-scales",
+            torch.mm,
+            a.to(dtype=torch.float16),
+            b.to(dtype=torch.float16),
+        )
+    )
 
     # cutlass impl
     timers.append(
-        bench_fn(label, sub_label, "cutlass_i8_i8_bf16_scaled_mm",
-                 ops.cutlass_scaled_mm, a, b, scale_a, scale_b,
-                 torch.bfloat16))
+        bench_fn(
+            label,
+            sub_label,
+            "cutlass_i8_i8_bf16_scaled_mm",
+            ops.cutlass_scaled_mm,
+            a,
+            b,
+            scale_a,
+            scale_b,
+            torch.bfloat16,
+        )
+    )
 
     # cutlass with bias
     timers.append(
-        bench_fn(label, sub_label, "cutlass_i8_i8_bf16_scaled_mm_bias",
-                 ops.cutlass_scaled_mm, a, b, scale_a, scale_b, torch.bfloat16,
-                 bias))
+        bench_fn(
+            label,
+            sub_label,
+            "cutlass_i8_i8_bf16_scaled_mm_bias",
+            ops.cutlass_scaled_mm,
+            a,
+            b,
+            scale_a,
+            scale_b,
+            torch.bfloat16,
+            bias,
+        )
+    )
 
     # cutlass sparse impl
     timers.append(
-        bench_fn(label, sub_label, "cutlass_i8_i8_bf16_scaled_sparse_mm",
-                 ops.cutlass_scaled_sparse_mm, a, b_compressed, e, scale_a,
-                 scale_b, torch.bfloat16))
+        bench_fn(
+            label,
+            sub_label,
+            "cutlass_i8_i8_bf16_scaled_sparse_mm",
+            ops.cutlass_scaled_sparse_mm,
+            a,
+            b_compressed,
+            e,
+            scale_a,
+            scale_b,
+            torch.bfloat16,
+        )
+    )
 
     # cutlass sparse with bias
     timers.append(
-        bench_fn(label, sub_label, "cutlass_i8_i8_bf16_scaled_sparse_mm_bias",
-                 ops.cutlass_scaled_sparse_mm, a, b_compressed, e, scale_a,
-                 scale_b, torch.bfloat16, bias))
+        bench_fn(
+            label,
+            sub_label,
+            "cutlass_i8_i8_bf16_scaled_sparse_mm_bias",
+            ops.cutlass_scaled_sparse_mm,
+            a,
+            b_compressed,
+            e,
+            scale_a,
+            scale_b,
+            torch.bfloat16,
+            bias,
+        )
+    )
 
     return timers
 
 
-def bench_fp8(dtype: torch.dtype, m: int, k: int, n: int, label: str,
-              sub_label: str) -> Iterable[TMeasurement]:
+def bench_fp8(
+    dtype: torch.dtype, m: int, k: int, n: int, label: str, sub_label: str
+) -> Iterable[TMeasurement]:
     assert dtype == torch.float8_e4m3fn
-    b_compressed, e, a, b = make_rand_sparse_tensors(torch.float8_e4m3fn, m, n,
-                                                     k)
+    b_compressed, e, a, b = make_rand_sparse_tensors(torch.float8_e4m3fn, m, n, k)
     scale_a = torch.tensor(1.0, device="cuda", dtype=torch.float32)
     scale_b = torch.tensor(1.0, device="cuda", dtype=torch.float32)
-    bias = torch.zeros((n, ), device="cuda", dtype=torch.bfloat16)
+    bias = torch.zeros((n,), device="cuda", dtype=torch.bfloat16)
 
-    out = ops.cutlass_scaled_sparse_mm(a, b_compressed, e, scale_a, scale_b,
-                                       torch.bfloat16)
+    out = ops.cutlass_scaled_sparse_mm(
+        a, b_compressed, e, scale_a, scale_b, torch.bfloat16
+    )
     out_ref = ops.cutlass_scaled_mm(a, b, scale_a, scale_b, torch.bfloat16)
 
     if not torch.allclose(out, out_ref):
@@ -124,97 +181,165 @@ def bench_fp8(dtype: torch.dtype, m: int, k: int, n: int, label: str,
 
     # pytorch impl w. bf16
     timers.append(
-        bench_fn(label, sub_label, "pytorch_bf16_bf16_bf16_matmul-no-scales",
-                 torch.mm, a.to(dtype=torch.bfloat16, device="cuda"),
-                 b.to(dtype=torch.bfloat16, device="cuda")))
+        bench_fn(
+            label,
+            sub_label,
+            "pytorch_bf16_bf16_bf16_matmul-no-scales",
+            torch.mm,
+            a.to(dtype=torch.bfloat16, device="cuda"),
+            b.to(dtype=torch.bfloat16, device="cuda"),
+        )
+    )
 
     # pytorch impl: bf16 output, without fp8 fast accum
     timers.append(
-        bench_fn(label,
-                 sub_label,
-                 "pytorch_fp8_fp8_bf16_scaled_mm",
-                 torch._scaled_mm,
-                 a,
-                 b,
-                 scale_a=scale_a,
-                 scale_b=scale_b,
-                 out_dtype=torch.bfloat16))
+        bench_fn(
+            label,
+            sub_label,
+            "pytorch_fp8_fp8_bf16_scaled_mm",
+            torch._scaled_mm,
+            a,
+            b,
+            scale_a=scale_a,
+            scale_b=scale_b,
+            out_dtype=torch.bfloat16,
+        )
+    )
 
     # pytorch impl: bf16 output, with fp8 fast accum
     timers.append(
-        bench_fn(label,
-                 sub_label,
-                 "pytorch_fp8_fp8_bf16_scaled_mm_fast_accum",
-                 torch._scaled_mm,
-                 a,
-                 b,
-                 scale_a=scale_a,
-                 scale_b=scale_b,
-                 out_dtype=torch.bfloat16,
-                 use_fast_accum=True))
+        bench_fn(
+            label,
+            sub_label,
+            "pytorch_fp8_fp8_bf16_scaled_mm_fast_accum",
+            torch._scaled_mm,
+            a,
+            b,
+            scale_a=scale_a,
+            scale_b=scale_b,
+            out_dtype=torch.bfloat16,
+            use_fast_accum=True,
+        )
+    )
 
     # pytorch impl: fp16 output, without fp8 fast accum
     timers.append(
-        bench_fn(label,
-                 sub_label,
-                 "pytorch_fp8_fp8_fp16_scaled_mm",
-                 torch._scaled_mm,
-                 a,
-                 b,
-                 scale_a=scale_a,
-                 scale_b=scale_b,
-                 out_dtype=torch.float16))
+        bench_fn(
+            label,
+            sub_label,
+            "pytorch_fp8_fp8_fp16_scaled_mm",
+            torch._scaled_mm,
+            a,
+            b,
+            scale_a=scale_a,
+            scale_b=scale_b,
+            out_dtype=torch.float16,
+        )
+    )
 
     # pytorch impl: fp16 output, with fp8 fast accum
     timers.append(
-        bench_fn(label,
-                 sub_label,
-                 "pytorch_fp8_fp8_fp16_scaled_mm_fast_accum",
-                 torch._scaled_mm,
-                 a,
-                 b,
-                 scale_a=scale_a,
-                 scale_b=scale_b,
-                 out_dtype=torch.float16,
-                 use_fast_accum=True))
+        bench_fn(
+            label,
+            sub_label,
+            "pytorch_fp8_fp8_fp16_scaled_mm_fast_accum",
+            torch._scaled_mm,
+            a,
+            b,
+            scale_a=scale_a,
+            scale_b=scale_b,
+            out_dtype=torch.float16,
+            use_fast_accum=True,
+        )
+    )
 
     # cutlass impl: bf16 output
     timers.append(
-        bench_fn(label, sub_label, "cutlass_fp8_fp8_bf16_scaled_mm",
-                 ops.cutlass_scaled_mm, a, b, scale_a, scale_b,
-                 torch.bfloat16))
+        bench_fn(
+            label,
+            sub_label,
+            "cutlass_fp8_fp8_bf16_scaled_mm",
+            ops.cutlass_scaled_mm,
+            a,
+            b,
+            scale_a,
+            scale_b,
+            torch.bfloat16,
+        )
+    )
 
     # cutlass impl: bf16 output
     timers.append(
-        bench_fn(label, sub_label, "cutlass_fp8_fp8_bf16_scaled_sparse_mm",
-                 ops.cutlass_scaled_sparse_mm, a, b_compressed, e, scale_a,
-                 scale_b, torch.bfloat16))
+        bench_fn(
+            label,
+            sub_label,
+            "cutlass_fp8_fp8_bf16_scaled_sparse_mm",
+            ops.cutlass_scaled_sparse_mm,
+            a,
+            b_compressed,
+            e,
+            scale_a,
+            scale_b,
+            torch.bfloat16,
+        )
+    )
 
     # cutlass impl: fp16 output
     timers.append(
-        bench_fn(label, sub_label, "cutlass_fp8_fp8_fp16_scaled_sparse_mm",
-                 ops.cutlass_scaled_sparse_mm, a, b_compressed, e, scale_a,
-                 scale_b, torch.float16))
+        bench_fn(
+            label,
+            sub_label,
+            "cutlass_fp8_fp8_fp16_scaled_sparse_mm",
+            ops.cutlass_scaled_sparse_mm,
+            a,
+            b_compressed,
+            e,
+            scale_a,
+            scale_b,
+            torch.float16,
+        )
+    )
 
     # cutlass impl: bf16 output, with bias
     timers.append(
-        bench_fn(label, sub_label,
-                 "cutlass_fp8_fp8_bf16_scaled_sparse_mm_bias",
-                 ops.cutlass_scaled_sparse_mm, a, b_compressed, e, scale_a,
-                 scale_b, torch.bfloat16, bias))
+        bench_fn(
+            label,
+            sub_label,
+            "cutlass_fp8_fp8_bf16_scaled_sparse_mm_bias",
+            ops.cutlass_scaled_sparse_mm,
+            a,
+            b_compressed,
+            e,
+            scale_a,
+            scale_b,
+            torch.bfloat16,
+            bias,
+        )
+    )
 
     # cutlass impl: fp16 output, with bias
     timers.append(
-        bench_fn(label, sub_label,
-                 "cutlass_fp8_fp8_fp16_scaled_sparse_mm_bias",
-                 ops.cutlass_scaled_sparse_mm, a, b_compressed, e, scale_a,
-                 scale_b, torch.float16, bias.to(dtype=torch.float16)))
+        bench_fn(
+            label,
+            sub_label,
+            "cutlass_fp8_fp8_fp16_scaled_sparse_mm_bias",
+            ops.cutlass_scaled_sparse_mm,
+            a,
+            b_compressed,
+            e,
+            scale_a,
+            scale_b,
+            torch.float16,
+            bias.to(dtype=torch.float16),
+        )
+    )
 
     return timers
 
 
-def bench(dtype: torch.dtype, m: int, k: int, n: int, label: str,
-          sub_label: str) -> Iterable[TMeasurement]:
+def bench(
+    dtype: torch.dtype, m: int, k: int, n: int, label: str, sub_label: str
+) -> Iterable[TMeasurement]:
     if dtype == torch.int8:
         return bench_int8(dtype, m, k, n, label, sub_label)
     if dtype == torch.float8_e4m3fn:
@@ -228,12 +353,12 @@ def print_timers(timers: Iterable[TMeasurement]):
     compare.print()
 
 
-def run(dtype: torch.dtype,
-        MKNs: Iterable[tuple[int, int, int]]) -> Iterable[TMeasurement]:
+def run(
+    dtype: torch.dtype, MKNs: Iterable[tuple[int, int, int]]
+) -> Iterable[TMeasurement]:
     results = []
     for m, k, n in MKNs:
-        timers = bench(dtype, m, k, n, f"scaled-{dtype}-gemm",
-                       f"MKN=({m}x{k}x{n})")
+        timers = bench(dtype, m, k, n, f"scaled-{dtype}-gemm", f"MKN=({m}x{k}x{n})")
         print_timers(timers)
         results.extend(timers)
 
@@ -241,10 +366,12 @@ def run(dtype: torch.dtype,
 
 
 # output makers
-def make_output(data: Iterable[TMeasurement],
-                MKNs: Iterable[tuple[int, int, int]],
-                base_description: str,
-                timestamp=None):
+def make_output(
+    data: Iterable[TMeasurement],
+    MKNs: Iterable[tuple[int, int, int]],
+    base_description: str,
+    timestamp=None,
+):
     print(f"== All Results {base_description} ====")
     print_timers(data)
 
@@ -258,8 +385,7 @@ def make_output(data: Iterable[TMeasurement],
 
 
 def run_square_bench(args):
-    dim_sizes = list(
-        range(args.dim_start, args.dim_end + 1, args.dim_increment))
+    dim_sizes = list(range(args.dim_start, args.dim_end + 1, args.dim_increment))
     MKNs = list(zip(dim_sizes, dim_sizes, dim_sizes))
     data = run(args.dtype, MKNs)
 
@@ -319,7 +445,7 @@ def run_model_bench(args):
         pkl.dump(all_data, f)
 
 
-if __name__ == '__main__':
+if __name__ == "__main__":
 
     def to_torch_dtype(dt):
         if dt == "int8":
@@ -344,12 +470,15 @@ Benchmark Cutlass GEMM.
     Output:
         - a .pkl file, that is a list of raw torch.benchmark.utils.Measurements for the pytorch and cutlass implementations for the various GEMMs.
             """,  # noqa: E501
-        formatter_class=argparse.RawTextHelpFormatter)
+        formatter_class=argparse.RawTextHelpFormatter,
+    )
 
-    parser.add_argument("--dtype",
-                        type=to_torch_dtype,
-                        required=True,
-                        help="Available options are ['int8', 'fp8']")
+    parser.add_argument(
+        "--dtype",
+        type=to_torch_dtype,
+        required=True,
+        help="Available options are ['int8', 'fp8']",
+    )
     subparsers = parser.add_subparsers(dest="cmd")
 
     square_parser = subparsers.add_parser("square_bench")
@@ -368,19 +497,19 @@ Benchmark Cutlass GEMM.
     range_parser.set_defaults(func=run_range_bench)
 
     model_parser = subparsers.add_parser("model_bench")
-    model_parser.add_argument("--models",
-                              nargs="+",
-                              type=str,
-                              default=DEFAULT_MODELS,
-                              choices=WEIGHT_SHAPES.keys())
-    model_parser.add_argument("--tp-sizes",
-                              nargs="+",
-                              type=int,
-                              default=DEFAULT_TP_SIZES)
-    model_parser.add_argument("--batch-sizes",
-                              nargs="+",
-                              type=int,
-                              default=DEFAULT_BATCH_SIZES)
+    model_parser.add_argument(
+        "--models",
+        nargs="+",
+        type=str,
+        default=DEFAULT_MODELS,
+        choices=WEIGHT_SHAPES.keys(),
+    )
+    model_parser.add_argument(
+        "--tp-sizes", nargs="+", type=int, default=DEFAULT_TP_SIZES
+    )
+    model_parser.add_argument(
+        "--batch-sizes", nargs="+", type=int, default=DEFAULT_BATCH_SIZES
+    )
     model_parser.set_defaults(func=run_model_bench)
 
     args = parser.parse_args()

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
@@ -10,8 +11,9 @@ import vllm._custom_ops as ops
 
 def to_fp8(tensor: torch.Tensor) -> torch.Tensor:
     finfo = torch.finfo(torch.float8_e4m3fn)
-    return torch.round(tensor.clamp(
-        min=finfo.min, max=finfo.max)).to(dtype=torch.float8_e4m3fn)
+    return torch.round(tensor.clamp(min=finfo.min, max=finfo.max)).to(
+        dtype=torch.float8_e4m3fn
+    )
 
 
 def to_int8(tensor: torch.Tensor) -> torch.Tensor:
@@ -26,10 +28,11 @@ def to_fp16(tensor: torch.Tensor) -> torch.Tensor:
     return tensor.to(dtype=torch.float16)
 
 
-def make_rand_tensors(dtype: torch.dtype, m: int, n: int,
-                      k: int) -> tuple[torch.Tensor, torch.Tensor]:
-    a = torch.randn((m, k), device='cuda') * 5
-    b = torch.randn((n, k), device='cuda').t() * 5
+def make_rand_tensors(
+    dtype: torch.dtype, m: int, n: int, k: int
+) -> tuple[torch.Tensor, torch.Tensor]:
+    a = torch.randn((m, k), device="cuda") * 5
+    b = torch.randn((n, k), device="cuda").t() * 5
 
     if dtype == torch.int8:
         return to_int8(a), to_int8(b)
@@ -49,9 +52,7 @@ def prune_to_2_4(tensor):
 
     # Create binary mask
     mask = torch.zeros_like(reshaped)
-    mask.scatter_(dim=1,
-                  index=indices,
-                  src=torch.ones_like(indices, dtype=mask.dtype))
+    mask.scatter_(dim=1, index=indices, src=torch.ones_like(indices, dtype=mask.dtype))
 
     # Apply mask and reshape back
     pruned = reshaped * mask
@@ -62,10 +63,11 @@ def prune_to_2_4(tensor):
     return pruned.reshape(original_shape)
 
 
-def make_rand_sparse_tensors(dtype: torch.dtype, m: int, n: int,
-                             k: int) -> tuple[torch.Tensor, torch.Tensor]:
-    a = torch.randn((m, k), device='cuda') * 5
-    b = torch.randn((n, k), device='cuda').t() * 5
+def make_rand_sparse_tensors(
+    dtype: torch.dtype, m: int, n: int, k: int
+) -> tuple[torch.Tensor, torch.Tensor]:
+    a = torch.randn((m, k), device="cuda") * 5
+    b = torch.randn((n, k), device="cuda").t() * 5
 
     b = prune_to_2_4(b.t()).t()
 
@@ -86,9 +88,9 @@ def make_rand_sparse_tensors(dtype: torch.dtype, m: int, n: int,
     return b_compressed, e, a, b
 
 
-def make_n_rand_sparse_tensors(num_tensors: int, dtype: torch.dtype,
-                        m: int, n: int, k: int) -> \
-                        tuple[Iterable[torch.Tensor], Iterable[torch.Tensor]]:
+def make_n_rand_sparse_tensors(
+    num_tensors: int, dtype: torch.dtype, m: int, n: int, k: int
+) -> tuple[Iterable[torch.Tensor], Iterable[torch.Tensor]]:
     ABs = []
     for _ in range(num_tensors):
         b_comp, e, a, b = make_rand_sparse_tensors(dtype, m, n, k)

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
@@ -16,8 +17,9 @@ from weight_shapes import WEIGHT_SHAPES
 
 from vllm import _custom_ops as ops
 from vllm.model_executor.layers.quantization.utils.fp8_utils import (
-    w8a8_block_fp8_matmul)
-from vllm.utils import FlexibleArgumentParser
+    w8a8_block_fp8_matmul,
+)
+from vllm.utils import FlexibleArgumentParser, cdiv
 
 DEFAULT_MODELS = list(WEIGHT_SHAPES.keys())
 DEFAULT_BATCH_SIZES = [1, 16, 32, 64, 128, 256, 512]
@@ -25,8 +27,9 @@ DEFAULT_TP_SIZES = [1]
 
 
 # bench
-def bench_fn(label: str, sub_label: str, description: str, fn: Callable, *args,
-             **kwargs) -> TMeasurement:
+def bench_fn(
+    label: str, sub_label: str, description: str, fn: Callable, *args, **kwargs
+) -> TMeasurement:
     min_run_time = 1
 
     globals = {
@@ -44,45 +47,48 @@ def bench_fn(label: str, sub_label: str, description: str, fn: Callable, *args,
 
 
 def bench_int8(
-        dtype: torch.dtype,
-        m: int,
-        k: int,
-        n: int,
-        label: str,
-        sub_label: str,
-        bench_kernels: Optional[list[str]] = None) -> Iterable[TMeasurement]:
+    dtype: torch.dtype,
+    m: int,
+    k: int,
+    n: int,
+    label: str,
+    sub_label: str,
+    bench_kernels: Optional[list[str]] = None,
+) -> Iterable[TMeasurement]:
     """Benchmark INT8-based kernels."""
     assert dtype == torch.int8
     a, b = make_rand_tensors(torch.int8, m, n, k)
     scale_a = torch.tensor(1.0, device="cuda", dtype=torch.float32)
     scale_b = torch.tensor(1.0, device="cuda", dtype=torch.float32)
-    bias = torch.zeros((n, ), device="cuda", dtype=torch.bfloat16)
-    azp = torch.zeros((m, ), device="cuda", dtype=torch.int32)
-    azp_adj = torch.zeros((n, ), device="cuda", dtype=torch.int32)
+    bias = torch.zeros((n,), device="cuda", dtype=torch.bfloat16)
+    azp = torch.zeros((m,), device="cuda", dtype=torch.int32)
+    azp_adj = torch.zeros((n,), device="cuda", dtype=torch.int32)
 
     bench_fns = {
-        "pytorch_bf16_bf16_bf16_matmul-no-scales":
-        lambda: torch.mm(a.to(dtype=torch.bfloat16), b.to(dtype=torch.bfloat16)
-                         ),
-        "pytorch_fp16_fp16_fp16_matmul-no-scales":
-        lambda: torch.mm(a.to(dtype=torch.float16), b.to(dtype=torch.float16)),
-        "cutlass_i8_i8_bf16_scaled_mm":
-        lambda: ops.cutlass_scaled_mm(a, b, scale_a, scale_b, torch.bfloat16),
-        "cutlass_i8_i8_bf16_scaled_mm_bias":
-        lambda: ops.cutlass_scaled_mm(a, b, scale_a, scale_b, torch.bfloat16,
-                                      bias),
-        "cutlass_i8_i8_bf16_scaled_mm_azp":
-        lambda: ops.cutlass_scaled_mm_azp(a, b, scale_a, scale_b, torch.
-                                          bfloat16, azp_adj),
-        "cutlass_i8_i8_bf16_scaled_mm_azp_bias":
-        lambda: ops.cutlass_scaled_mm_azp(a, b, scale_a, scale_b, torch.
-                                          bfloat16, azp_adj, None, bias),
-        "cutlass_i8_i8_bf16_scaled_mm_azp_pt":
-        lambda: ops.cutlass_scaled_mm_azp(a, b, scale_a, scale_b, torch.
-                                          bfloat16, azp_adj, azp),
-        "cutlass_i8_i8_bf16_scaled_mm_azp_pt_bias":
-        lambda: ops.cutlass_scaled_mm_azp(a, b, scale_a, scale_b, torch.
-                                          bfloat16, azp_adj, azp, bias),
+        "pytorch_bf16_bf16_bf16_matmul-no-scales": lambda: torch.mm(
+            a.to(dtype=torch.bfloat16), b.to(dtype=torch.bfloat16)
+        ),
+        "pytorch_fp16_fp16_fp16_matmul-no-scales": lambda: torch.mm(
+            a.to(dtype=torch.float16), b.to(dtype=torch.float16)
+        ),
+        "cutlass_i8_i8_bf16_scaled_mm": lambda: ops.cutlass_scaled_mm(
+            a, b, scale_a, scale_b, torch.bfloat16
+        ),
+        "cutlass_i8_i8_bf16_scaled_mm_bias": lambda: ops.cutlass_scaled_mm(
+            a, b, scale_a, scale_b, torch.bfloat16, bias
+        ),
+        "cutlass_i8_i8_bf16_scaled_mm_azp": lambda: ops.cutlass_scaled_mm_azp(
+            a, b, scale_a, scale_b, torch.bfloat16, azp_adj
+        ),
+        "cutlass_i8_i8_bf16_scaled_mm_azp_bias": lambda: ops.cutlass_scaled_mm_azp(
+            a, b, scale_a, scale_b, torch.bfloat16, azp_adj, None, bias
+        ),
+        "cutlass_i8_i8_bf16_scaled_mm_azp_pt": lambda: ops.cutlass_scaled_mm_azp(
+            a, b, scale_a, scale_b, torch.bfloat16, azp_adj, azp
+        ),
+        "cutlass_i8_i8_bf16_scaled_mm_azp_pt_bias": lambda: ops.cutlass_scaled_mm_azp(
+            a, b, scale_a, scale_b, torch.bfloat16, azp_adj, azp, bias
+        ),
     }
 
     timers = []
@@ -96,73 +102,68 @@ def bench_int8(
 
 
 def bench_fp8(
-        dtype: torch.dtype,
-        m: int,
-        k: int,
-        n: int,
-        label: str,
-        sub_label: str,
-        bench_kernels: Optional[list[str]] = None) -> Iterable[TMeasurement]:
+    dtype: torch.dtype,
+    m: int,
+    k: int,
+    n: int,
+    label: str,
+    sub_label: str,
+    bench_kernels: Optional[list[str]] = None,
+) -> Iterable[TMeasurement]:
     """Benchmark FP8-based kernels."""
     assert dtype == torch.float8_e4m3fn
     a, b = make_rand_tensors(torch.float8_e4m3fn, m, n, k)
     a_cont = a.contiguous()
     scale_a = torch.tensor(1.0, device="cuda", dtype=torch.float32)
     scale_b = torch.tensor(1.0, device="cuda", dtype=torch.float32)
-    block_scale_a = torch.rand((m, k // 128),
-                               device="cuda",
-                               dtype=torch.float32)
-    block_scale_b = torch.rand((k // 128, n // 128),
-                               device="cuda",
-                               dtype=torch.float32)
+
+    block_scale_a = torch.rand((m, cdiv(k, 128)), device="cuda", dtype=torch.float32)
+    block_scale_b = torch.rand(
+        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()
-    bias = torch.zeros((n, ), device="cuda", dtype=torch.bfloat16)
+    bias = torch.zeros((n,), device="cuda", dtype=torch.bfloat16)
 
     print(m, k, n)
 
     bench_fns = {
-        "pytorch_bf16_bf16_bf16_matmul-no-scales":
-        lambda: torch.mm(a.to(dtype=torch.bfloat16), b.to(dtype=torch.bfloat16)
-                         ),
-        "pytorch_fp16_fp16_fp16_matmul-no-scales":
-        lambda: torch.mm(a.to(dtype=torch.float16), b.to(dtype=torch.float16)),
-        "pytorch_fp8_fp8_fp16_scaled_mm":
-        lambda: torch._scaled_mm(
-            a, b, scale_a, scale_b, out_dtype=torch.float16),
-        "pytorch_fp8_fp8_fp16_scaled_mm_fast_accum":
-        lambda: torch._scaled_mm(a,
-                                 b,
-                                 scale_a,
-                                 scale_b,
-                                 out_dtype=torch.float16,
-                                 use_fast_accum=True),
-        "pytorch_fp8_fp8_bf16_scaled_mm":
-        lambda: torch._scaled_mm(
-            a, b, scale_a, scale_b, out_dtype=torch.bfloat16),
-        "pytorch_fp8_fp8_bf16_scaled_mm_fast_accum":
-        lambda: torch._scaled_mm(a,
-                                 b,
-                                 scale_a,
-                                 scale_b,
-                                 out_dtype=torch.bfloat16,
-                                 use_fast_accum=True),
-        "cutlass_fp8_fp8_bf16_scaled_mm":
-        lambda: ops.cutlass_scaled_mm(a, b, scale_a, scale_b, torch.bfloat16),
-        "cutlass_fp8_fp8_fp16_scaled_mm":
-        lambda: ops.cutlass_scaled_mm(a, b, scale_a, scale_b, torch.float16),
-        "cutlass_fp8_fp8_bf16_scaled_mm_bias":
-        lambda: ops.cutlass_scaled_mm(a, b, scale_a, scale_b, torch.bfloat16,
-                                      bias),
-        "cutlass_fp8_fp8_fp16_scaled_mm_bias":
-        lambda: ops.cutlass_scaled_mm(a, b, scale_a, scale_b, torch.float16,
-                                      bias.to(dtype=torch.float16)),
-        "triton_fp8_fp8_fp16_scaled_mm_blockwise":
-        lambda: w8a8_block_fp8_matmul(a_cont, b.t(), block_scale_a,
-                                      block_scale_b.t(), (128, 128)),
-        "cutlass_fp8_fp8_fp16_scaled_mm_blockwise":
-        lambda: ops.cutlass_scaled_mm(a, b, block_scale_a_M_major,
-                                      block_scale_b_K_major, torch.float16),
+        "pytorch_bf16_bf16_bf16_matmul-no-scales": lambda: torch.mm(
+            a.to(dtype=torch.bfloat16), b.to(dtype=torch.bfloat16)
+        ),
+        "pytorch_fp16_fp16_fp16_matmul-no-scales": lambda: torch.mm(
+            a.to(dtype=torch.float16), b.to(dtype=torch.float16)
+        ),
+        "pytorch_fp8_fp8_fp16_scaled_mm": lambda: torch._scaled_mm(
+            a, b, scale_a, scale_b, out_dtype=torch.float16
+        ),
+        "pytorch_fp8_fp8_fp16_scaled_mm_fast_accum": lambda: torch._scaled_mm(
+            a, b, scale_a, scale_b, out_dtype=torch.float16, use_fast_accum=True
+        ),
+        "pytorch_fp8_fp8_bf16_scaled_mm": lambda: torch._scaled_mm(
+            a, b, scale_a, scale_b, out_dtype=torch.bfloat16
+        ),
+        "pytorch_fp8_fp8_bf16_scaled_mm_fast_accum": lambda: torch._scaled_mm(
+            a, b, scale_a, scale_b, out_dtype=torch.bfloat16, use_fast_accum=True
+        ),
+        "cutlass_fp8_fp8_bf16_scaled_mm": lambda: ops.cutlass_scaled_mm(
+            a, b, scale_a, scale_b, torch.bfloat16
+        ),
+        "cutlass_fp8_fp8_fp16_scaled_mm": lambda: ops.cutlass_scaled_mm(
+            a, b, scale_a, scale_b, torch.float16
+        ),
+        "cutlass_fp8_fp8_bf16_scaled_mm_bias": lambda: ops.cutlass_scaled_mm(
+            a, b, scale_a, scale_b, torch.bfloat16, bias
+        ),
+        "cutlass_fp8_fp8_fp16_scaled_mm_bias": lambda: ops.cutlass_scaled_mm(
+            a, b, scale_a, scale_b, torch.float16, bias.to(dtype=torch.float16)
+        ),
+        "triton_fp8_fp8_fp16_scaled_mm_blockwise": lambda: w8a8_block_fp8_matmul(
+            a_cont, b.t(), block_scale_a, block_scale_b.t(), (128, 128)
+        ),
+        "cutlass_fp8_fp8_fp16_scaled_mm_blockwise": lambda: ops.cutlass_scaled_mm(
+            a, b, block_scale_a_M_major, block_scale_b_K_major, torch.float16
+        ),
     }
 
     timers = []
@@ -175,13 +176,15 @@ def bench_fp8(
     return timers
 
 
-def bench(dtype: torch.dtype,
-          m: int,
-          k: int,
-          n: int,
-          label: str,
-          sub_label: str,
-          bench_kernels: Optional[list[str]] = None) -> Iterable[TMeasurement]:
+def bench(
+    dtype: torch.dtype,
+    m: int,
+    k: int,
+    n: int,
+    label: str,
+    sub_label: str,
+    bench_kernels: Optional[list[str]] = None,
+) -> Iterable[TMeasurement]:
     if dtype == torch.int8:
         return bench_int8(dtype, m, k, n, label, sub_label, bench_kernels)
     if dtype == torch.float8_e4m3fn:
@@ -195,27 +198,33 @@ def print_timers(timers: Iterable[TMeasurement]):
     compare.print()
 
 
-def run(dtype: torch.dtype,
-        MKNs: Iterable[tuple[int, int, int]],
-        bench_kernels: Optional[list[str]] = None) -> Iterable[TMeasurement]:
+def run(
+    dtype: torch.dtype,
+    MKNs: Iterable[tuple[int, int, int]],
+    bench_kernels: Optional[list[str]] = None,
+) -> Iterable[TMeasurement]:
     results = []
     for m, k, n in MKNs:
-        timers = bench(dtype,
-                       m,
-                       k,
-                       n,
-                       f"scaled-{dtype}-gemm",
-                       f"MKN=({m}x{k}x{n})",
-                       bench_kernels=bench_kernels)
+        timers = bench(
+            dtype,
+            m,
+            k,
+            n,
+            f"scaled-{dtype}-gemm",
+            f"MKN=({m}x{k}x{n})",
+            bench_kernels=bench_kernels,
+        )
         print_timers(timers)
         results.extend(timers)
     return results
 
 
-def make_output(data: Iterable[TMeasurement],
-                MKNs: Iterable[tuple[int, int, int]],
-                base_description: str,
-                timestamp=None):
+def make_output(
+    data: Iterable[TMeasurement],
+    MKNs: Iterable[tuple[int, int, int]],
+    base_description: str,
+    timestamp=None,
+):
     print(f"== All Results {base_description} ====")
     print_timers(data)
 
@@ -226,8 +235,7 @@ def make_output(data: Iterable[TMeasurement],
 
 
 def run_square_bench(args):
-    dim_sizes = list(
-        range(args.dim_start, args.dim_end + 1, args.dim_increment))
+    dim_sizes = list(range(args.dim_start, args.dim_end + 1, args.dim_increment))
     MKNs = list(zip(dim_sizes, dim_sizes, dim_sizes))
     data = run(args.dtype, MKNs, bench_kernels=args.kernels)
     make_output(data, MKNs, f"square_bench-{args.dtype}")
@@ -285,7 +293,7 @@ def run_model_bench(args):
         pkl.dump(all_data, f)
 
 
-if __name__ == '__main__':
+if __name__ == "__main__":
 
     def to_torch_dtype(dt):
         if dt == "int8":
@@ -310,19 +318,21 @@ Benchmark Cutlass GEMM.
     Output:
         - a .pkl file, that is a list of raw torch.benchmark.utils.Measurements for the pytorch and cutlass implementations for the various GEMMs.
             """,  # noqa: E501
-        formatter_class=argparse.RawTextHelpFormatter)
+        formatter_class=argparse.RawTextHelpFormatter,
+    )
 
-    parser.add_argument("--dtype",
-                        type=to_torch_dtype,
-                        required=True,
-                        help="Available options are ['int8', 'fp8']")
+    parser.add_argument(
+        "--dtype",
+        type=to_torch_dtype,
+        required=True,
+        help="Available options are ['int8', 'fp8']",
+    )
     parser.add_argument(
         "--kernels",
         nargs="+",
         type=str,
         default=None,
-        help=
-        "Exact names of the kernels to benchmark. If not set, runs all kernels."
+        help="Exact names of the kernels to benchmark. If not set, runs all kernels.",
     )
 
     subparsers = parser.add_subparsers(dest="cmd")
@@ -343,19 +353,19 @@ Benchmark Cutlass GEMM.
     range_parser.set_defaults(func=run_range_bench)
 
     model_parser = subparsers.add_parser("model_bench")
-    model_parser.add_argument("--models",
-                              nargs="+",
-                              type=str,
-                              default=DEFAULT_MODELS,
-                              choices=WEIGHT_SHAPES.keys())
-    model_parser.add_argument("--tp-sizes",
-                              nargs="+",
-                              type=int,
-                              default=DEFAULT_TP_SIZES)
-    model_parser.add_argument("--batch-sizes",
-                              nargs="+",
-                              type=int,
-                              default=DEFAULT_BATCH_SIZES)
+    model_parser.add_argument(
+        "--models",
+        nargs="+",
+        type=str,
+        default=DEFAULT_MODELS,
+        choices=WEIGHT_SHAPES.keys(),
+    )
+    model_parser.add_argument(
+        "--tp-sizes", nargs="+", type=int, default=DEFAULT_TP_SIZES
+    )
+    model_parser.add_argument(
+        "--batch-sizes", nargs="+", type=int, default=DEFAULT_BATCH_SIZES
+    )
     model_parser.set_defaults(func=run_model_bench)
 
     args = parser.parse_args()

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)
@@ -42,4 +43,4 @@ WEIGHT_SHAPES = {
         ([8192, 57344], 1),
         ([28672, 8192], 0),
     ],
-}
+}

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
 
@@ -12,39 +13,37 @@ app = Quart(__name__)
 
 async def forward_request(url, data):
     async with aiohttp.ClientSession(timeout=AIOHTTP_TIMEOUT) as session:
-        headers = {
-            "Authorization": f"Bearer {os.environ.get('OPENAI_API_KEY')}"
-        }
-        async with session.post(url=url, json=data,
-                                headers=headers) as response:
+        headers = {"Authorization": f"Bearer {os.environ.get('OPENAI_API_KEY')}"}
+        async with session.post(url=url, json=data, headers=headers) as response:
             if response.status == 200:
                 # if response.headers.get('Transfer-Encoding') == 'chunked':
                 if True:
-                    async for chunk_bytes in response.content.iter_chunked(
-                            1024):
+                    async for chunk_bytes in response.content.iter_chunked(1024):
                         yield chunk_bytes
                 else:
                     content = await response.read()
                     yield content
 
 
-@app.route('/v1/completions', methods=['POST'])
+@app.route("/v1/completions", methods=["POST"])
 async def handle_request():
     try:
         original_request_data = await request.get_json()
 
         prefill_request = original_request_data.copy()
         # change max_tokens = 1 to let it only do prefill
-        prefill_request['max_tokens'] = 1
+        prefill_request["max_tokens"] = 1
 
         # finish prefill
-        async for _ in forward_request('http://localhost:8100/v1/completions',
-                                       prefill_request):
+        async for _ in forward_request(
+            "http://localhost:8100/v1/completions", prefill_request
+        ):
             continue
 
         # return decode
-        generator = forward_request('http://localhost:8200/v1/completions',
-                                    original_request_data)
+        generator = forward_request(
+            "http://localhost:8200/v1/completions", original_request_data
+        )
         response = await make_response(generator)
         response.timeout = None
 
@@ -53,11 +52,12 @@ async def handle_request():
     except Exception as e:
         import sys
         import traceback
+
         exc_info = sys.exc_info()
         print("Error occurred in disagg prefill proxy server")
         print(e)
         print("".join(traceback.format_exception(*exc_info)))
 
 
-if __name__ == '__main__':
+if __name__ == "__main__":
     app.run(port=8000)

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
@@ -8,7 +9,6 @@ from aiohttp import web
 
 
 class RoundRobinProxy:
-
     def __init__(self, target_ports):
         self.target_ports = target_ports
         self.port_cycle = itertools.cycle(self.target_ports)
@@ -21,14 +21,15 @@ class RoundRobinProxy:
             try:
                 # Forward the request
                 async with session.request(
-                        method=request.method,
-                        url=target_url,
-                        headers=request.headers,
-                        data=request.content,
+                    method=request.method,
+                    url=target_url,
+                    headers=request.headers,
+                    data=request.content,
                 ) as response:
                     # Start sending the response
-                    resp = web.StreamResponse(status=response.status,
-                                              headers=response.headers)
+                    resp = web.StreamResponse(
+                        status=response.status, headers=response.headers
+                    )
                     await resp.prepare(request)
 
                     # Stream the response content
@@ -45,11 +46,11 @@ class RoundRobinProxy:
 async def main():
     proxy = RoundRobinProxy([8100, 8200])
     app = web.Application()
-    app.router.add_route('*', '/{path:.*}', proxy.handle_request)
+    app.router.add_route("*", "/{path:.*}", proxy.handle_request)
 
     runner = web.AppRunner(app)
     await runner.setup()
-    site = web.TCPSite(runner, 'localhost', 8000)
+    site = web.TCPSite(runner, "localhost", 8000)
     await site.start()
 
     print("Proxy server started on http://localhost:8000")
@@ -58,5 +59,5 @@ async def main():
     await asyncio.Event().wait()
 
 
-if __name__ == '__main__':
+if __name__ == "__main__":
     asyncio.run(main())

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
 
@@ -6,43 +7,41 @@ import matplotlib.pyplot as plt
 import pandas as pd
 
 if __name__ == "__main__":
-
     data = []
-    for name in ['disagg_prefill', 'chunked_prefill']:
+    for name in ["disagg_prefill", "chunked_prefill"]:
         for qps in [2, 4, 6, 8]:
             with open(f"results/{name}-qps-{qps}.json") as f:
                 x = json.load(f)
-                x['name'] = name
-                x['qps'] = qps
+                x["name"] = name
+                x["qps"] = qps
                 data.append(x)
 
     df = pd.DataFrame.from_dict(data)
-    dis_df = df[df['name'] == 'disagg_prefill']
-    chu_df = df[df['name'] == 'chunked_prefill']
+    dis_df = df[df["name"] == "disagg_prefill"]
+    chu_df = df[df["name"] == "chunked_prefill"]
 
-    plt.style.use('bmh')
-    plt.rcParams['font.size'] = 20
+    plt.style.use("bmh")
+    plt.rcParams["font.size"] = 20
 
     for key in [
-            'mean_ttft_ms', 'median_ttft_ms', 'p99_ttft_ms', 'mean_itl_ms',
-            'median_itl_ms', 'p99_itl_ms'
+        "mean_ttft_ms",
+        "median_ttft_ms",
+        "p99_ttft_ms",
+        "mean_itl_ms",
+        "median_itl_ms",
+        "p99_itl_ms",
     ]:
-
         fig, ax = plt.subplots(figsize=(11, 7))
-        plt.plot(dis_df['qps'],
-                 dis_df[key],
-                 label='disagg_prefill',
-                 marker='o',
-                 linewidth=4)
-        plt.plot(chu_df['qps'],
-                 chu_df[key],
-                 label='chunked_prefill',
-                 marker='o',
-                 linewidth=4)
+        plt.plot(
+            dis_df["qps"], dis_df[key], label="disagg_prefill", marker="o", linewidth=4
+        )
+        plt.plot(
+            chu_df["qps"], chu_df[key], label="chunked_prefill", marker="o", linewidth=4
+        )
         ax.legend()
 
-        ax.set_xlabel('QPS')
+        ax.set_xlabel("QPS")
         ax.set_ylabel(key)
         ax.set_ylim(bottom=0)
-        fig.savefig(f'results/{key}.png')
+        fig.savefig(f"results/{key}.png")
         plt.close(fig)

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
@@ -24,10 +25,12 @@ class bench_params_t:
     dtype: torch.dtype
 
     def description(self):
-        return (f'N {self.num_tokens} '
-                f'x D {self.hidden_size} '
-                f'x R {self.add_residual} '
-                f'x DT {self.dtype}')
+        return (
+            f"N {self.num_tokens} "
+            f"x D {self.hidden_size} "
+            f"x R {self.add_residual} "
+            f"x DT {self.dtype}"
+        )
 
 
 def get_bench_params() -> list[bench_params_t]:
@@ -38,15 +41,19 @@ def get_bench_params() -> list[bench_params_t]:
     DTYPES = [torch.bfloat16, torch.float]
 
     combinations = product(NUM_TOKENS, HIDDEN_SIZES, ADD_RESIDUAL, DTYPES)
-    bench_params = list(map(lambda x: \
-        bench_params_t(x[0], x[1], x[2], x[3]), combinations))
+    bench_params = list(
+        map(lambda x: bench_params_t(x[0], x[1], x[2], x[3]), combinations)
+    )
     return bench_params
 
 
 # Reference impls
-def unfused_int8_impl(rms_norm_layer: RMSNorm, x: torch.Tensor,
-                      residual: Optional[torch.Tensor],
-                      quant_dtype: torch.dtype):
+def unfused_int8_impl(
+    rms_norm_layer: RMSNorm,
+    x: torch.Tensor,
+    residual: Optional[torch.Tensor],
+    quant_dtype: torch.dtype,
+):
     # Norm
     torch_out = None
     if residual is None:
@@ -58,9 +65,12 @@ def unfused_int8_impl(rms_norm_layer: RMSNorm, x: torch.Tensor,
     torch_out, _, _ = ops.scaled_int8_quant(torch_out)
 
 
-def unfused_fp8_impl(rms_norm_layer: RMSNorm, x: torch.Tensor,
-                     residual: Optional[torch.Tensor],
-                     quant_dtype: torch.dtype):
+def unfused_fp8_impl(
+    rms_norm_layer: RMSNorm,
+    x: torch.Tensor,
+    residual: Optional[torch.Tensor],
+    quant_dtype: torch.dtype,
+):
     # Norm
     torch_out = None
     if residual is None:
@@ -73,22 +83,27 @@ def unfused_fp8_impl(rms_norm_layer: RMSNorm, x: torch.Tensor,
 
 
 def fused_impl(
-        rms_norm_layer: RMSNorm,  # this stores the weights
-        x: torch.Tensor,
-        residual: Optional[torch.Tensor],
-        quant_dtype: torch.dtype):
-    out, _ = ops.rms_norm_dynamic_per_token_quant(x,
-                                                  rms_norm_layer.weight,
-                                                  1e-6,
-                                                  quant_dtype,
-                                                  residual=residual)
+    rms_norm_layer: RMSNorm,  # this stores the weights
+    x: torch.Tensor,
+    residual: Optional[torch.Tensor],
+    quant_dtype: torch.dtype,
+):
+    out, _ = ops.rms_norm_dynamic_per_token_quant(
+        x, rms_norm_layer.weight, 1e-6, quant_dtype, residual=residual
+    )
 
 
 # Bench functions
-def bench_fn(rms_norm_layer: RMSNorm, x: torch.Tensor, residual: torch.Tensor,
-             quant_dtype: torch.dtype, label: str, sub_label: str,
-             fn: Callable, description: str) -> TMeasurement:
-
+def bench_fn(
+    rms_norm_layer: RMSNorm,
+    x: torch.Tensor,
+    residual: torch.Tensor,
+    quant_dtype: torch.dtype,
+    label: str,
+    sub_label: str,
+    fn: Callable,
+    description: str,
+) -> TMeasurement:
     min_run_time = 1
 
     globals = {
@@ -106,43 +121,81 @@ def bench_fn(rms_norm_layer: RMSNorm, x: torch.Tensor, residual: torch.Tensor,
         description=description,
     ).blocked_autorange(min_run_time=min_run_time)
 
-def bench(params: bench_params_t, label: str, sub_label: str) \
-        -> Iterable[TMeasurement]:
 
+def bench(params: bench_params_t, label: str, sub_label: str) -> Iterable[TMeasurement]:
     # Make inputs
     layer = RMSNorm(params.hidden_size, 1e-6).to(dtype=params.dtype)
     # Make weights
     layer.weight.data.normal_(mean=1.0, std=0.1)
     # Make inputs
     scale = 1 / params.hidden_size
-    x = torch.randn(params.num_tokens,
-                    params.hidden_size,
-                    dtype=params.dtype,
-                    device='cuda') * scale
-    residual = (torch.randn_like(x) * scale).to(device='cuda') \
-            if params.add_residual else None
+    x = (
+        torch.randn(
+            params.num_tokens, params.hidden_size, dtype=params.dtype, device="cuda"
+        )
+        * scale
+    )
+    residual = (
+        (torch.randn_like(x) * scale).to(device="cuda") if params.add_residual else None
+    )
 
     timers = []
 
     # unfused int8 impl.
     timers.append(
-        bench_fn(layer, x, residual, torch.int8, label, sub_label,
-                 unfused_int8_impl, "unfused_int8_impl"))
+        bench_fn(
+            layer,
+            x,
+            residual,
+            torch.int8,
+            label,
+            sub_label,
+            unfused_int8_impl,
+            "unfused_int8_impl",
+        )
+    )
 
     # unfused fp8 impl.
     timers.append(
-        bench_fn(layer, x, residual, torch.float8_e4m3fn, label, sub_label,
-                 unfused_fp8_impl, "unfused_fp8_impl"))
+        bench_fn(
+            layer,
+            x,
+            residual,
+            torch.float8_e4m3fn,
+            label,
+            sub_label,
+            unfused_fp8_impl,
+            "unfused_fp8_impl",
+        )
+    )
 
     # fused int8 impl.
     timers.append(
-        bench_fn(layer, x, residual, torch.int8, label, sub_label, fused_impl,
-                 "fused_int8_impl"))
+        bench_fn(
+            layer,
+            x,
+            residual,
+            torch.int8,
+            label,
+            sub_label,
+            fused_impl,
+            "fused_int8_impl",
+        )
+    )
 
     # fused fp8 impl.
     timers.append(
-        bench_fn(layer, x, residual, torch.float8_e4m3fn, label, sub_label,
-                 fused_impl, "fused_fp8_impl"))
+        bench_fn(
+            layer,
+            x,
+            residual,
+            torch.float8_e4m3fn,
+            label,
+            sub_label,
+            fused_impl,
+            "fused_fp8_impl",
+        )
+    )
 
     print_timers(timers)
 
@@ -157,13 +210,12 @@ def print_timers(timers: Iterable[TMeasurement]):
 
 
 def main():
-    torch.set_default_device('cuda')
+    torch.set_default_device("cuda")
     bench_params = get_bench_params()
 
     timers = []
     for bp in tqdm(bench_params):
-        timers.extend(
-            bench(bp, "rms-norm-dynamic-per-token-quant", bp.description()))
+        timers.extend(bench(bp, "rms-norm-dynamic-per-token-quant", bp.description()))
     print_timers(timers)
 
     # pickle all the results
@@ -172,5 +224,5 @@ def main():
         pkl.dump(timers, f)
 
 
-if __name__ == '__main__':
+if __name__ == "__main__":
     main()

benchmarks/kernels/bench_fp8_gemm.py

@@ -0,0 +1,159 @@
+# 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_fp8_quant as vllm_scaled_fp8_quant
+from vllm.triton_utils import triton
+
+PROVIDER_CFGS = {
+    "torch-bf16": dict(enabled=True),
+    "fp8-tensor-w-token-a": dict(
+        w="tensor", a="token", no_a_quant=False, enabled=False
+    ),
+    "fp8-tensor-w-tensor-a": dict(
+        w="tensor", a="tensor", no_a_quant=False, enabled=True
+    ),
+    "fp8-channel-w-token-a": dict(
+        w="channel", a="token", no_a_quant=False, enabled=True
+    ),
+    "fp8-channel-w-tensor-a": dict(
+        w="channel", a="tensor", no_a_quant=False, enabled=False
+    ),
+    "fp8-tensor-w-token-a-noquant": dict(
+        w="tensor", a="token", no_a_quant=True, enabled=False
+    ),
+    "fp8-tensor-w-tensor-a-noquant": dict(
+        w="tensor", a="tensor", no_a_quant=True, enabled=True
+    ),
+    "fp8-channel-w-token-a-noquant": dict(
+        w="channel", a="token", no_a_quant=True, enabled=True
+    ),
+    "fp8-channel-w-tensor-a-noquant": dict(
+        w="channel", a="tensor", no_a_quant=True, enabled=False
+    ),
+}
+
+_enabled = [k for k, v in PROVIDER_CFGS.items() if v["enabled"]]
+
+
+def _quant_weight_fp8(b: torch.Tensor, w_type: str, device: str):
+    if w_type == "tensor":
+        scale_b = torch.ones(1, device=device, dtype=torch.float32)
+        b_fp8, scale_b_fp8 = vllm_scaled_fp8_quant(b, scale_b)
+    else:
+        b_fp8, scale_b_fp8 = vllm_scaled_fp8_quant(b, use_per_token_if_dynamic=True)
+    return b_fp8.t(), scale_b_fp8
+
+
+def build_fp8_runner(cfg, a, b, dtype, device):
+    b_fp8, scale_b_fp8 = _quant_weight_fp8(b, cfg["w"], device)
+
+    scale_a_const = (
+        torch.ones(1, device=device, dtype=torch.float32)
+        if cfg["a"] == "tensor"
+        else None
+    )
+
+    if cfg["no_a_quant"]:
+        if cfg["a"] == "tensor":
+            a_fp8, scale_a_fp8 = vllm_scaled_fp8_quant(a, scale_a_const)
+        else:
+            a_fp8, scale_a_fp8 = vllm_scaled_fp8_quant(a, use_per_token_if_dynamic=True)
+
+        def run():
+            return vllm_scaled_mm(a_fp8, b_fp8, scale_a_fp8, scale_b_fp8, dtype)
+
+        return run
+
+    if cfg["a"] == "tensor":
+
+        def run():
+            a_fp8, scale_a_fp8 = vllm_scaled_fp8_quant(a, scale_a_const)
+            return vllm_scaled_mm(a_fp8, b_fp8, scale_a_fp8, scale_b_fp8, dtype)
+
+    else:
+
+        def run():
+            a_fp8, scale_a_fp8 = vllm_scaled_fp8_quant(a, use_per_token_if_dynamic=True)
+            return vllm_scaled_mm(a_fp8, b_fp8, scale_a_fp8, scale_b_fp8, dtype)
+
+    return run
+
+
+@triton.testing.perf_report(
+    triton.testing.Benchmark(
+        x_names=["batch_size"],
+        x_vals=[1, 16, 64, 128, 256, 512, 1024, 2048, 4096, 8192, 16384],
+        x_log=False,
+        line_arg="provider",
+        line_vals=_enabled,
+        line_names=_enabled,
+        ylabel="TFLOP/s (larger is better)",
+        plot_name="BF16 vs FP8 GEMMs",
+        args={},
+    )
+)
+def benchmark(batch_size, provider, N, K):
+    M = batch_size
+    device = "cuda"
+    dtype = torch.bfloat16
+
+    a = torch.randn((M, K), device=device, dtype=dtype)
+    b = torch.randn((N, K), device=device, dtype=dtype)
+
+    quantiles = [0.5, 0.2, 0.8]
+
+    if provider == "torch-bf16":
+        ms, min_ms, max_ms = triton.testing.do_bench_cudagraph(
+            lambda: torch.nn.functional.linear(a, b), quantiles=quantiles
+        )
+    else:
+        cfg = PROVIDER_CFGS[provider]
+        run_quant = build_fp8_runner(cfg, a, b, dtype, device)
+        ms, min_ms, max_ms = triton.testing.do_bench_cudagraph(
+            lambda: run_quant(), quantiles=quantiles
+        )
+
+    to_tflops = lambda t_ms: (2 * M * N * K) * 1e-12 / (t_ms * 1e-3)
+    return to_tflops(ms), to_tflops(max_ms), to_tflops(min_ms)
+
+
+def prepare_shapes(args):
+    out = []
+    for model, tp_size in itertools.product(args.models, args.tp_sizes):
+        for KN, tp_dim in copy.deepcopy(WEIGHT_SHAPES[model]):
+            KN[tp_dim] //= tp_size
+            KN.append(model)
+            out.append(KN)
+    return out
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--models",
+        nargs="+",
+        type=str,
+        default=["meta-llama/Llama-3.1-8B-Instruct"],
+        choices=list(WEIGHT_SHAPES.keys()),
+    )
+    parser.add_argument("--tp-sizes", nargs="+", type=int, default=[1])
+    args = parser.parse_args()
+
+    for K, N, model in prepare_shapes(args):
+        print(f"{model}, N={N} K={K}, BF16 vs FP8 GEMMs TFLOP/s:")
+        benchmark.run(
+            print_data=True,
+            show_plots=True,
+            save_path=f"bench_fp8_res_n{N}_k{K}",
+            N=N,
+            K=K,
+        )
+
+    print("Benchmark finished!")

benchmarks/kernels/bench_int8_gemm.py

@@ -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
@@ -9,32 +10,39 @@ import torch.nn.functional as F
 
 from vllm import _custom_ops as ops
 from vllm.model_executor.layers.quantization.aqlm import (
-    dequantize_weight, generic_dequantize_gemm, get_int_dtype,
-    optimized_dequantize_gemm)
+    dequantize_weight,
+    generic_dequantize_gemm,
+    get_int_dtype,
+    optimized_dequantize_gemm,
+)
 from vllm.utils import FlexibleArgumentParser
 
-os.environ['CUDA_VISIBLE_DEVICES'] = '0'
+os.environ["CUDA_VISIBLE_DEVICES"] = "0"
 
 
 def torch_mult(
-        input: torch.Tensor,  #  [..., in_features]
-        weights: torch.Tensor,
-        scales: torch.Tensor,  #  [num_out_groups, 1, 1, 1]
+    # [..., in_features]
+    input: torch.Tensor,
+    weights: torch.Tensor,
+    # [num_out_groups, 1, 1, 1]
+    scales: torch.Tensor,
 ) -> torch.Tensor:
     output = F.linear(input, weights)
     return output
 
 
 def dequant_out_scale(
-    input: torch.Tensor,  #  [..., in_features]
-    codes: torch.IntTensor,  #  [num_out_groups, num_in_groups, num_codebooks]
-    codebooks: torch.
-    Tensor,  #  [num_codebooks, codebook_size, out_group_size, in_group_size]
-    scales: torch.Tensor,  #  [num_out_groups, 1, 1, 1]
+    # [..., in_features]
+    input: torch.Tensor,
+    # [num_out_groups, num_in_groups, num_codebooks]
+    codes: torch.IntTensor,
+    # [num_codebooks, codebook_size, out_group_size, in_group_size]
+    codebooks: torch.Tensor,
+    # [num_out_groups, 1, 1, 1]
+    scales: torch.Tensor,
     output_partition_sizes: torch.IntTensor,
     bias: Optional[torch.Tensor],
 ) -> torch.Tensor:
-
     weights = ops.aqlm_dequant(codes, codebooks, output_partition_sizes)
 
     if bias is None:
@@ -46,40 +54,42 @@ def dequant_out_scale(
         flattened_output *= b_scales
         return flattened_output.view(orig_shape)
     else:
-        b_scales = scales.view(scales.shape[:-3] + (-1, )).expand(
-            -1, weights.shape[1])
+        b_scales = scales.view(scales.shape[:-3] + (-1,)).expand(-1, weights.shape[1])
         weights *= b_scales
         return F.linear(input, weights, bias)
 
 
 def dequant_weight_scale(
-    input: torch.Tensor,  #  [..., in_features]
-    codes: torch.IntTensor,  #  [num_out_groups, num_in_groups, num_codebooks]
-    codebooks: torch.
-    Tensor,  #  [num_codebooks, codebook_size, out_group_size, in_group_size]
-    scales: torch.Tensor,  #  [num_out_groups, 1, 1, 1]
+    # [..., in_features]
+    input: torch.Tensor,
+    # [num_out_groups, num_in_groups, num_codebooks]
+    codes: torch.IntTensor,
+    # [num_codebooks, codebook_size, out_group_size, in_group_size]
+    codebooks: torch.Tensor,
+    # [num_out_groups, 1, 1, 1]
+    scales: torch.Tensor,
     output_partition_sizes: torch.IntTensor,
     bias: Optional[torch.Tensor],
 ) -> torch.Tensor:
-
     weights = ops.aqlm_dequant(codes, codebooks, output_partition_sizes)
 
-    b_scales = scales.view(scales.shape[:-3] + (-1, )).expand(
-        -1, weights.shape[1])
+    b_scales = scales.view(scales.shape[:-3] + (-1,)).expand(-1, weights.shape[1])
     weights *= b_scales
     return F.linear(input, weights, bias)
 
 
 def dequant_no_scale(
-    input: torch.Tensor,  #  [..., in_features]
-    codes: torch.IntTensor,  #  [num_out_groups, num_in_groups, num_codebooks]
-    codebooks: torch.
-    Tensor,  #  [num_codebooks, codebook_size, out_group_size, in_group_size]
-    scales: torch.Tensor,  #  [num_out_groups, 1, 1, 1]
+    # [..., in_features]
+    input: torch.Tensor,
+    # [num_out_groups, num_in_groups, num_codebooks]
+    codes: torch.IntTensor,
+    # [num_codebooks, codebook_size, out_group_size, in_group_size]
+    codebooks: torch.Tensor,
+    # [num_out_groups, 1, 1, 1]
+    scales: torch.Tensor,
     output_partition_sizes: torch.IntTensor,
     bias: Optional[torch.Tensor],
 ) -> torch.Tensor:
-
     weights = ops.aqlm_dequant(codes, codebooks, output_partition_sizes)
 
     return F.linear(input, weights, bias)
@@ -89,23 +99,26 @@ def dequant_no_scale(
 # the generic pytorch version.
 # Just visual comparison.
 def dequant_test(k: int, parts: torch.Tensor, nbooks: int, bits: int) -> None:
-
     n = int(parts.sum().item())
 
-    device = torch.device('cuda:0')
+    device = torch.device("cuda:0")
 
     code_range = (1 << bits) // 2
     ingroups = 8
 
-    codes = torch.randint(-code_range,
-                          code_range,
-                          size=(n, k // ingroups, nbooks),
-                          dtype=get_int_dtype(bits),
-                          device=device)
+    codes = torch.randint(
+        -code_range,
+        code_range,
+        size=(n, k // ingroups, nbooks),
+        dtype=get_int_dtype(bits),
+        device=device,
+    )
 
-    codebooks = torch.randn(size=(parts.shape[0] * nbooks, 1 << bits, 1, 8),
-                            dtype=torch.float16,
-                            device=device)
+    codebooks = torch.randn(
+        size=(parts.shape[0] * nbooks, 1 << bits, 1, 8),
+        dtype=torch.float16,
+        device=device,
+    )
 
     count = 0
     for index in range(16):
@@ -138,24 +151,25 @@ def dequant_test(k: int, parts: torch.Tensor, nbooks: int, bits: int) -> None:
 
 
 def main():
-
     parser = FlexibleArgumentParser(description="Benchmark aqlm performance.")
 
     # Add arguments
-    parser.add_argument("--nbooks",
-                        type=int,
-                        default=1,
-                        help="Number of codebooks (default: 1)")
-    parser.add_argument("--bits",
-                        type=int,
-                        default=16,
-                        help="Number of bits per code element (default: 16)")
+    parser.add_argument(
+        "--nbooks", type=int, default=1, help="Number of codebooks (default: 1)"
+    )
+    parser.add_argument(
+        "--bits",
+        type=int,
+        default=16,
+        help="Number of bits per code element (default: 16)",
+    )
     parser.add_argument(
         "--test",
         type=bool,
         default=False,
         help="Run the decompression/dequant tester rather than benchmarking "
-        "(default: False)")
+        "(default: False)",
+    )
 
     # Parse the arguments
     args = parser.parse_args()
@@ -165,7 +179,7 @@ def main():
     bits = args.bits
 
     if args.test:
-        dequant_test(4096, torch.tensor((4096, )), nbooks, bits)
+        dequant_test(4096, torch.tensor((4096,)), nbooks, bits)
         return
 
     # Otherwise, benchmark.
@@ -184,31 +198,54 @@ def main():
     with open(filename, "w") as f:
         sys.stdout = f
 
-        print('m | k | n | n parts', end='')
+        print("m | k | n | n parts", end="")
         for method in methods:
-            print(f" | {method.__name__.replace('_', ' ')} (µs)", end='')
-        print('')
+            print(f" | {method.__name__.replace('_', ' ')} (µs)", end="")
+        print("")
 
         # These are reasonable prefill sizes.
-        ksandpartions = ((4096, (4096, 4096, 4096)), (4096, (4096, )),
-                         (4096, (11008, 11008)), (11008, (4096, )))
+        ksandpartions = (
+            (4096, (4096, 4096, 4096)),
+            (4096, (4096,)),
+            (4096, (11008, 11008)),
+            (11008, (4096,)),
+        )
 
         # reasonable ranges for m.
         for m in [
-                1, 2, 4, 8, 10, 12, 14, 16, 24, 32, 48, 52, 56, 64, 96, 112,
-                128, 256, 512, 1024, 1536, 2048, 3072, 4096
+            1,
+            2,
+            4,
+            8,
+            10,
+            12,
+            14,
+            16,
+            24,
+            32,
+            48,
+            52,
+            56,
+            64,
+            96,
+            112,
+            128,
+            256,
+            512,
+            1024,
+            1536,
+            2048,
+            3072,
+            4096,
         ]:
-            print(f'{m}', file=sys.__stdout__)
+            print(f"{m}", file=sys.__stdout__)
             for ksp in ksandpartions:
-                run_grid(m, ksp[0], torch.tensor(ksp[1]), nbooks, bits,
-                         methods)
+                run_grid(m, ksp[0], torch.tensor(ksp[1]), nbooks, bits, methods)
 
         sys.stdout = sys.__stdout__
 
 
-def run_grid(m: int, k: int, parts: torch.Tensor, nbooks: int, bits: int,
-             methods):
-
+def run_grid(m: int, k: int, parts: torch.Tensor, nbooks: int, bits: int, methods):
     # I didn't see visible improvements from increasing these, but feel free :)
     num_warmup_trials = 1
     num_trials = 1
@@ -229,7 +266,7 @@ def run_grid(m: int, k: int, parts: torch.Tensor, nbooks: int, bits: int,
             )
 
     n = parts.sum().item()
-    print(f'{m} | {k} | {n} | {parts.tolist()}', end='')
+    print(f"{m} | {k} | {n} | {parts.tolist()}", end="")
 
     for method in methods:
         best_time_us = 1e20
@@ -249,32 +286,36 @@ def run_grid(m: int, k: int, parts: torch.Tensor, nbooks: int, bits: int,
             if kernel_dur_us < best_time_us:
                 best_time_us = kernel_dur_us
 
-        print(f' | {kernel_dur_us:.0f}', end='')
+        print(f" | {kernel_dur_us:.0f}", end="")
 
-    print('')
+    print("")
 
 
-def run_timing(num_calls: int, m: int, k: int, parts: torch.Tensor,
-               nbooks: int, bits: int, method) -> float:
-
+def run_timing(
+    num_calls: int, m: int, k: int, parts: torch.Tensor, nbooks: int, bits: int, method
+) -> float:
     n = int(parts.sum().item())
 
-    device = torch.device('cuda:0')
+    device = torch.device("cuda:0")
 
     input = torch.randn((1, m, k), dtype=torch.float16, device=device)
 
     code_range = (1 << bits) // 2
     ingroups = 8
 
-    codes = torch.randint(-code_range,
-                          code_range,
-                          size=(n, k // ingroups, nbooks),
-                          dtype=get_int_dtype(bits),
-                          device=device)
+    codes = torch.randint(
+        -code_range,
+        code_range,
+        size=(n, k // ingroups, nbooks),
+        dtype=get_int_dtype(bits),
+        device=device,
+    )
 
-    codebooks = torch.randn(size=(parts.shape[0] * nbooks, 1 << bits, 1, 8),
-                            dtype=torch.float16,
-                            device=device)
+    codebooks = torch.randn(
+        size=(parts.shape[0] * nbooks, 1 << bits, 1, 8),
+        dtype=torch.float16,
+        device=device,
+    )
 
     scales = torch.randn(size=(n, 1, 1, 1), dtype=torch.float16, device=device)
 

benchmarks/kernels/benchmark_bitblas.py

@@ -1,29 +1,36 @@
 # SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 # Copyright (c) Microsoft Corporation.
 # Licensed under the MIT License.
 
 from vllm.model_executor.layers.quantization.utils.bitblas_utils import (
-    MINIMUM_BITBLAS_VERSION)
+    MINIMUM_BITBLAS_VERSION,
+)
 
 try:
     import bitblas
+
     if bitblas.__version__ < MINIMUM_BITBLAS_VERSION:
-        raise ImportError("bitblas version is wrong. Please "
-                          f"install bitblas>={MINIMUM_BITBLAS_VERSION}")
+        raise ImportError(
+            "bitblas version is wrong. Please "
+            f"install bitblas>={MINIMUM_BITBLAS_VERSION}"
+        )
 except ImportError as e:
     bitblas_import_exception = e
-    raise ValueError("Trying to use the bitblas backend, but could not import"
-                     f"with the following error: {bitblas_import_exception}. "
-                     "Please install bitblas through the following command: "
-                     f"`pip install bitblas>={MINIMUM_BITBLAS_VERSION}`"
-                     ) from bitblas_import_exception
+    raise ValueError(
+        "Trying to use the bitblas backend, but could not import"
+        f"with the following error: {bitblas_import_exception}. "
+        "Please install bitblas through the following command: "
+        f"`pip install bitblas>={MINIMUM_BITBLAS_VERSION}`"
+    ) from bitblas_import_exception
 
 from bitblas import Matmul, MatmulConfig, auto_detect_nvidia_target
 
 from vllm.utils import FlexibleArgumentParser
 
 parser = FlexibleArgumentParser(
-    description="Benchmark BitBLAS int4 on a specific target.")
+    description="Benchmark BitBLAS int4 on a specific target."
+)
 
 # Add arguments to the parser
 parser.add_argument(
@@ -32,10 +39,9 @@ parser.add_argument(
     default=auto_detect_nvidia_target(),
     help="Specify the target device for benchmarking.",
 )
-parser.add_argument("--group_size",
-                    type=int,
-                    default=None,
-                    help="Group size for grouped quantization.")
+parser.add_argument(
+    "--group_size", type=int, default=None, help="Group size for grouped quantization."
+)
 parser.add_argument(
     "--A_dtype",
     type=str,
@@ -82,17 +88,17 @@ parser.add_argument(
     choices=["nt", "nn"],
     help="Matrix layout, 'nt' for non-transpose A and transpose W.",
 )
-parser.add_argument("--with_bias",
-                    action="store_true",
-                    help="Include bias in the benchmark.")
+parser.add_argument(
+    "--with_bias", action="store_true", help="Include bias in the benchmark."
+)
 parser.add_argument(
     "--with_scaling",
     action="store_true",
     help="Include scaling factor in the quantization.",
 )
-parser.add_argument("--with_zeros",
-                    action="store_true",
-                    help="Include zeros in the quantization.")
+parser.add_argument(
+    "--with_zeros", action="store_true", help="Include zeros in the quantization."
+)
 parser.add_argument(
     "--zeros_mode",
     type=str,
@@ -170,8 +176,7 @@ shapes = [
 ]
 
 # Build test shapes with all the shared arguments
-test_shapes = [(MatmulConfig, Matmul, (*shape, *shared_args))
-               for shape in shapes]
+test_shapes = [(MatmulConfig, Matmul, (*shape, *shared_args)) for shape in shapes]
 
 benchmark_sets = []
 benchmark_sets.extend(test_shapes)
@@ -206,12 +211,12 @@ for config_key, values in benchmark_results.items():
     func_name = args_split[0]
     input_args_str = "-".join(args_split[1:])
     col_widths[0] = max(col_widths[0], len(func_name) + 2, len(headers[0]) + 2)
-    col_widths[1] = max(col_widths[1],
-                        len(input_args_str) + 2,
-                        len(headers[1]) + 2)
-    col_widths[2] = max(col_widths[2],
-                        len(f"{values['BitBLAS_top20_latency']:.3f} ms") + 2,
-                        len(headers[2]) + 2)
+    col_widths[1] = max(col_widths[1], len(input_args_str) + 2, len(headers[1]) + 2)
+    col_widths[2] = max(
+        col_widths[2],
+        len(f"{values['BitBLAS_top20_latency']:.3f} ms") + 2,
+        len(headers[2]) + 2,
+    )
     # break only if you want to measure widths from a single example;
     # otherwise, let it loop over all items.
 
@@ -232,5 +237,6 @@ for config_key, values in benchmark_results.items():
         f"{values['BitBLAS_top20_latency']:.3f} ms",
     ]
     row_str = "".join(
-        [str(cell).ljust(col_widths[idx]) for idx, cell in enumerate(row)])
+        [str(cell).ljust(col_widths[idx]) for idx, cell in enumerate(row)]
+    )
     print(row_str)

benchmarks/kernels/benchmark_cutlass_fp4_moe.py

@@ -0,0 +1,490 @@
+# 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
+activations. The triton_moe kernel takes in fp8 weights(tensor scaled to fp8)
+and 16-bit activations.
+"""
+
+import nvtx
+import torch
+import torch.utils.benchmark as benchmark
+
+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_fp4
+from vllm.model_executor.layers.fused_moe.fused_moe import fused_experts, fused_topk
+from vllm.scalar_type import scalar_types
+from vllm.utils import FlexibleArgumentParser
+
+WEIGHT_SHAPES_MOE = {
+    "nvidia/DeepSeek-R1-FP4": [
+        [256, 8, 2048, 7168],
+    ],
+}
+
+DEFAULT_MODELS = [
+    "nvidia/DeepSeek-R1-FP4",
+]
+
+DEFAULT_BATCH_SIZES = [4, 8, 16, 32, 64, 128, 256, 512, 1024, 2048]
+DEFAULT_TP_SIZES = [1]
+
+PER_ACT_TOKEN_OPTS = [False]
+PER_OUT_CH_OPTS = [False]
+FLOAT4_E2M1_MAX = scalar_types.float4_e2m1f.max()
+FLOAT8_E4M3_MAX = torch.finfo(torch.float8_e4m3fn).max
+
+
+def to_fp8(tensor: torch.Tensor):
+    finfo = torch.finfo(torch.float8_e4m3fn)
+    return torch.round(tensor.clamp(min=finfo.min, max=finfo.max)).to(
+        dtype=torch.float8_e4m3fn
+    )
+
+
+def bench_run(
+    results: list[benchmark.Measurement],
+    model: str,
+    num_experts: int,
+    topk: int,
+    per_act_token: bool,
+    per_out_ch: bool,
+    mkn: tuple[int, int, int],
+):
+    label = "NVFP4 Blockscaled CUTLASS MOE vs FP8 Tensor Scaled Triton"
+
+    sub_label = (
+        "{}, num_experts={}, topk={}, per_act_token={} per_out_ch={}, MKN=({})".format(
+            model, num_experts, topk, per_act_token, per_out_ch, mkn
+        )
+    )
+
+    print(f"Testing: {sub_label}")
+
+    (m, k, n) = mkn
+
+    dtype = torch.half
+    device = "cuda"
+    a = torch.randn((m, k), device=device, dtype=dtype) / 10
+    w1 = torch.randn((num_experts, 2 * n, k), device=device, dtype=dtype) / 10
+    w2 = torch.randn((num_experts, k, n), device=device, dtype=dtype) / 10
+
+    _, a_fp8_scale = ops.scaled_fp8_quant(a)
+
+    w1_fp8q = torch.empty(
+        (num_experts, 2 * n, k), device=device, dtype=torch.float8_e4m3fn
+    )
+    w2_fp8q = torch.empty((num_experts, k, n), device=device, dtype=torch.float8_e4m3fn)
+    w1_fp8scale = torch.empty((num_experts, 1, 1), device=device, dtype=torch.float32)
+    w2_fp8scale = torch.empty((num_experts, 1, 1), device=device, dtype=torch.float32)
+
+    for expert in range(num_experts):
+        w1_fp8q[expert], w1_fp8scale[expert] = ops.scaled_fp8_quant(w1[expert])
+        w2_fp8q[expert], w2_fp8scale[expert] = ops.scaled_fp8_quant(w2[expert])
+
+    w1_fp8q_notransp = w1_fp8q.clone()
+    w2_fp8q_notransp = w2_fp8q.clone()
+    w1_fp8q = w1_fp8q.transpose(1, 2)
+    w2_fp8q = w2_fp8q.transpose(1, 2)
+
+    score = torch.randn((m, num_experts), device=device, dtype=dtype)
+
+    topk_weights, topk_ids, _ = fused_topk(a, score, topk, renormalize=False)
+
+    quant_blocksize = 16
+    w1_blockscale = torch.empty(
+        (num_experts, 2 * n, k // quant_blocksize),
+        device=device,
+        dtype=torch.float8_e4m3fn,
+    )
+    w2_blockscale = torch.empty(
+        (num_experts, k, n // quant_blocksize), device=device, dtype=torch.float8_e4m3fn
+    )
+
+    # n_b_scales = 2 * n if per_out_ch else 1
+    # k_b_scales = k if per_out_ch else 1
+    w1_fp4 = torch.empty((num_experts, 2 * n, k // 2), device=device, dtype=torch.uint8)
+    w2_fp4 = torch.empty((num_experts, k, n // 2), device=device, dtype=torch.uint8)
+
+    w1_gs = torch.empty((num_experts,), device=device, dtype=torch.float32)
+    w2_gs = torch.empty((num_experts,), device=device, dtype=torch.float32)
+    a1_gs = torch.ones((num_experts,), device=device, dtype=torch.float32)
+    a2_gs = torch.ones((num_experts,), device=device, dtype=torch.float32)
+
+    for expert in range(num_experts):
+        w1_e = w1[expert]
+        w2_e = w2[expert]
+        w1_amax = torch.abs(w1_e).max().to(torch.float32)
+        w2_amax = torch.abs(w2_e).max().to(torch.float32)
+        w1_gs[expert] = FLOAT8_E4M3_MAX * FLOAT4_E2M1_MAX / w1_amax
+        w2_gs[expert] = FLOAT8_E4M3_MAX * FLOAT4_E2M1_MAX / w2_amax
+
+        w1_fp4[expert], w1_blockscale[expert] = ops.scaled_fp4_quant(
+            w1_e, w1_gs[expert]
+        )
+
+        w2_fp4[expert], w2_blockscale[expert] = ops.scaled_fp4_quant(
+            w2_e, w2_gs[expert]
+        )
+
+    def run_triton_moe(
+        a: torch.Tensor,
+        w1: torch.Tensor,
+        w2: torch.Tensor,
+        topk_weights: torch.Tensor,
+        topk_ids: torch.Tensor,
+        w1_scale: torch.Tensor,
+        w2_scale: torch.Tensor,
+        a_fp8_scale: torch.Tensor,
+        num_repeats: int,
+    ):
+        for _ in range(num_repeats):
+            fused_experts(
+                a,
+                w1,
+                w2,
+                topk_weights,
+                topk_ids,
+                use_fp8_w8a8=True,
+                w1_scale=w1_scale,
+                w2_scale=w2_scale,
+                a1_scale=a_fp8_scale,
+            )
+
+    def run_cutlass_moe_fp4(
+        a: torch.Tensor,
+        w1_fp4: torch.Tensor,
+        w2_fp4: torch.Tensor,
+        w1_blockscale: torch.Tensor,
+        w2_blockscale: torch.Tensor,
+        w1_gs: torch.Tensor,
+        w2_gs: torch.Tensor,
+        a1_gs: torch.Tensor,
+        a2_gs: torch.Tensor,
+        topk_weights: torch.Tensor,
+        topk_ids: torch.Tensor,
+        m: int,
+        n: int,
+        k: int,
+        e: int,
+        device: torch.device,
+        num_repeats: int,
+    ):
+        for _ in range(num_repeats):
+            with nvtx.annotate("cutlass_moe_fp4", color="green"):
+                cutlass_moe_fp4(
+                    a=a,
+                    a1_gscale=a1_gs,
+                    a2_gscale=a2_gs,
+                    w1_fp4=w1_fp4,
+                    w1_blockscale=w1_blockscale,
+                    w1_alphas=w1_gs,
+                    w2_fp4=w2_fp4,
+                    w2_blockscale=w2_blockscale,
+                    w2_alphas=w2_gs,
+                    topk_weights=topk_weights,
+                    topk_ids=topk_ids,
+                    m=m,
+                    n=n,
+                    k=k,
+                    e=num_experts,
+                    device=device,
+                )
+
+    def run_cutlass_from_graph(
+        a: torch.Tensor,
+        a1_gscale: torch.Tensor,
+        w1_fp4: torch.Tensor,
+        w1_blockscale: torch.Tensor,
+        w1_alphas: torch.Tensor,
+        a2_gscale: torch.Tensor,
+        w2_fp4: torch.Tensor,
+        w2_blockscale: torch.Tensor,
+        w2_alphas: torch.Tensor,
+        topk_weights: torch.Tensor,
+        topk_ids: torch.Tensor,
+        m: int,
+        n: int,
+        k: int,
+        e: int,
+        device: torch.device,
+    ):
+        with set_current_vllm_config(
+            VllmConfig(parallel_config=ParallelConfig(pipeline_parallel_size=1))
+        ):
+            return cutlass_moe_fp4(
+                a=a,
+                a1_gscale=a1_gs,
+                w1_fp4=w1_fp4,
+                w1_blockscale=w1_blockscale,
+                w1_alphas=w1_alphas,
+                a2_gscale=a2_gs,
+                w2_fp4=w2_fp4,
+                w2_blockscale=w2_blockscale,
+                w2_alphas=w2_alphas,
+                topk_weights=topk_weights,
+                topk_ids=topk_ids,
+                m=m,
+                n=n,
+                k=k,
+                e=num_experts,
+                device=device,
+            )
+
+    def run_triton_from_graph(
+        a: torch.Tensor,
+        w1: torch.Tensor,
+        w2: torch.Tensor,
+        topk_weights: torch.Tensor,
+        topk_ids: torch.Tensor,
+        w1_scale: torch.Tensor,
+        w2_scale: torch.Tensor,
+        a_fp8_scale: torch.Tensor,
+    ):
+        with set_current_vllm_config(
+            VllmConfig(parallel_config=ParallelConfig(pipeline_parallel_size=1))
+        ):
+            return fused_experts(
+                a,
+                w1,
+                w2,
+                topk_weights,
+                topk_ids,
+                use_fp8_w8a8=True,
+                w1_scale=w1_scale,
+                w2_scale=w2_scale,
+                a1_scale=a_fp8_scale,
+            )
+
+    def replay_graph(graph, num_repeats):
+        for _ in range(num_repeats):
+            graph.replay()
+        torch.cuda.synchronize()
+
+    cutlass_stream = torch.cuda.Stream()
+    cutlass_graph = torch.cuda.CUDAGraph()
+    with torch.cuda.graph(cutlass_graph, stream=cutlass_stream):
+        run_cutlass_from_graph(
+            a=a,
+            a1_gscale=a1_gs,
+            w1_fp4=w1_fp4,
+            w1_blockscale=w1_blockscale,
+            w1_alphas=w1_gs,
+            a2_gscale=a2_gs,
+            w2_fp4=w2_fp4,
+            w2_blockscale=w2_blockscale,
+            w2_alphas=w2_gs,
+            topk_weights=topk_weights,
+            topk_ids=topk_ids,
+            m=m,
+            n=n,
+            k=k,
+            e=num_experts,
+            device=device,
+        )
+    torch.cuda.synchronize()
+
+    triton_stream = torch.cuda.Stream()
+    triton_graph = torch.cuda.CUDAGraph()
+    with torch.cuda.graph(triton_graph, stream=triton_stream):
+        run_triton_from_graph(
+            a,
+            w1_fp8q_notransp,
+            w2_fp8q_notransp,
+            topk_weights,
+            topk_ids,
+            w1_fp8scale,
+            w2_fp8scale,
+            a_fp8_scale,
+        )
+    torch.cuda.synchronize()
+
+    min_run_time = 5
+    num_warmup = 5
+    num_runs = 25
+
+    globals = {
+        # Baseline params
+        "w1": w1,
+        "w2": w2,
+        "score": score,
+        "topk": topk,
+        "w1_fp8q_notransp": w1_fp8q_notransp,
+        "w2_fp8q_notransp": w2_fp8q_notransp,
+        "w1_fp8scale": w1_fp8scale,
+        "w2_fp8scale": w2_fp8scale,
+        "a_fp8_scale": a_fp8_scale,
+        # Cutlass params
+        "a": a,
+        "a1_gscale": a1_gs,
+        "w1_fp4": w1_fp4,
+        "w1_blockscale": w1_blockscale,
+        "w1_alphas": w1_gs,
+        "a2_gscale": a2_gs,
+        "w2_fp4": w2_fp4,
+        "w2_blockscale": w2_blockscale,
+        "w2_alphas": w2_gs,
+        "topk_weights": topk_weights,
+        "topk_ids": topk_ids,
+        "m": m,
+        "n": n,
+        "k": k,
+        "e": num_experts,
+        "device": device,
+        # cuda graph params
+        "cutlass_graph": cutlass_graph,
+        "triton_graph": triton_graph,
+        # Gen params
+        "num_runs": num_runs,
+        # Kernels
+        "run_triton_moe": run_triton_moe,
+        "run_cutlass_moe_fp4": run_cutlass_moe_fp4,
+        "replay_graph": replay_graph,
+    }
+
+    # Warmup
+    run_triton_moe(
+        a,
+        w1_fp8q_notransp,
+        w2_fp8q_notransp,
+        topk_weights,
+        topk_ids,
+        w1_fp8scale,
+        w2_fp8scale,
+        a_fp8_scale,
+        num_warmup,
+    )
+
+    results.append(
+        benchmark.Timer(
+            stmt="run_triton_moe(a, w1_fp8q_notransp, w2_fp8q_notransp, topk_weights, topk_ids, w1_fp8scale, w2_fp8scale, a_fp8_scale, num_runs)",  # noqa: E501
+            globals=globals,
+            label=label,
+            sub_label=sub_label,
+            description="triton_moe",
+        ).blocked_autorange(min_run_time=min_run_time)
+    )
+
+    # Warmup
+    replay_graph(triton_graph, num_warmup)
+
+    results.append(
+        benchmark.Timer(
+            stmt="replay_graph(triton_graph, num_runs)",
+            globals=globals,
+            label=label,
+            sub_label=sub_label,
+            description="triton_moe_cuda_graphs",
+        ).blocked_autorange(min_run_time=min_run_time)
+    )
+
+    # Warmup
+
+    run_cutlass_moe_fp4(
+        a,
+        w1_fp4,
+        w2_fp4,
+        w1_blockscale,
+        w2_blockscale,
+        w1_gs,
+        w2_gs,
+        a1_gs,
+        a2_gs,
+        topk_weights,
+        topk_ids,
+        m,
+        n,
+        k,
+        num_experts,
+        device,
+        num_warmup,
+    )
+
+    results.append(
+        benchmark.Timer(
+            stmt="run_cutlass_moe_fp4(a, w1_fp4, w2_fp4, w1_blockscale, w2_blockscale, w1_alphas, w2_alphas, a1_gscale, a2_gscale, topk_weights, topk_ids, m, n, k, e, device, num_runs)",  # noqa: E501
+            globals=globals,
+            label=label,
+            sub_label=sub_label,
+            description="cutlass_moe_fp4",
+        ).blocked_autorange(min_run_time=min_run_time)
+    )
+
+    # Warmup
+    replay_graph(cutlass_graph, num_warmup)
+
+    results.append(
+        benchmark.Timer(
+            stmt="replay_graph(cutlass_graph, num_runs)",
+            globals=globals,
+            label=label,
+            sub_label=sub_label,
+            description="cutlass_moe_fp4_cuda_graphs",
+        ).blocked_autorange(min_run_time=min_run_time)
+    )
+
+
+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:
+        for tp in args.tp_sizes:
+            for layer in WEIGHT_SHAPES_MOE[model]:
+                num_experts = layer[0]
+                topk = layer[1]
+                size_k = layer[2]
+                size_n = layer[3] // tp
+
+                if len(args.limit_k) > 0 and size_k not in args.limit_k:
+                    continue
+
+                if len(args.limit_n) > 0 and size_n not in args.limit_n:
+                    continue
+
+                for per_act_token in PER_ACT_TOKEN_OPTS:
+                    for per_out_ch in PER_OUT_CH_OPTS:
+                        for size_m in args.batch_sizes:
+                            mkn = (size_m, size_k, size_n)
+                            bench_run(
+                                results,
+                                model,
+                                num_experts,
+                                topk,
+                                per_act_token,
+                                per_out_ch,
+                                mkn,
+                            )
+
+    compare = benchmark.Compare(results)
+    compare.print()
+
+
+if __name__ == "__main__":
+    parser = FlexibleArgumentParser(
+        description="Benchmark NVFP4 CUTLASS MOE across specified models/shapes/batches"
+    )
+    parser.add_argument(
+        "--models",
+        nargs="+",
+        type=str,
+        default=DEFAULT_MODELS,
+        choices=WEIGHT_SHAPES_MOE.keys(),
+    )
+    parser.add_argument("--tp-sizes", nargs="+", type=int, default=DEFAULT_TP_SIZES)
+    parser.add_argument(
+        "--batch-sizes", nargs="+", type=int, default=DEFAULT_BATCH_SIZES
+    )
+    parser.add_argument("--limit-k", nargs="+", type=int, default=[])
+    parser.add_argument("--limit-n", nargs="+", type=int, default=[])
+    parser.add_argument("--limit-num-groups", nargs="+", type=int, default=[])
+    parser.add_argument("--limit-per-act-token", nargs="+", type=int, default=[])
+    parser.add_argument("--limit-per-out-ch", nargs="+", type=int, default=[])
+
+    args = parser.parse_args()
+    main(args)

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
@@ -6,14 +7,18 @@ 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.fused_moe import (cutlass_moe_fp8,
-                                                            fused_experts,
-                                                            fused_topk)
+from vllm.model_executor.layers.fused_moe.cutlass_moe import cutlass_moe_fp8
+from vllm.model_executor.layers.fused_moe.fused_moe import (
+    fused_experts,
+    fused_topk,
+)
 from vllm.utils import FlexibleArgumentParser
 
 DEFAULT_MODELS = [
-    "nm-testing/Mixtral-8x7B-Instruct-v0.1", "nm-testing/deepseekv2-lite",
-    "ibm-granite/granite-3.0-1b-a400m", "ibm-granite/granite-3.0-3b-a800m"
+    "nm-testing/Mixtral-8x7B-Instruct-v0.1",
+    "nm-testing/deepseekv2-lite",
+    "ibm-granite/granite-3.0-1b-a400m",
+    "ibm-granite/granite-3.0-3b-a800m",
 ]
 DEFAULT_BATCH_SIZES = [1, 4, 8, 16, 32, 64, 128, 256, 512]
 DEFAULT_TP_SIZES = [1]
@@ -24,19 +29,27 @@ PER_OUT_CH_OPTS = [False]
 
 def to_fp8(tensor: torch.Tensor):
     finfo = torch.finfo(torch.float8_e4m3fn)
-    return torch.round(tensor.clamp(
-        min=finfo.min, max=finfo.max)).to(dtype=torch.float8_e4m3fn)
+    return torch.round(tensor.clamp(min=finfo.min, max=finfo.max)).to(
+        dtype=torch.float8_e4m3fn
+    )
 
 
-def bench_run(results: list[benchmark.Measurement], model: str,
-              num_experts: int, topk: int, per_act_token: bool,
-              per_out_ch: bool, mkn: tuple[int, int, int]):
+def bench_run(
+    results: list[benchmark.Measurement],
+    model: str,
+    num_experts: int,
+    topk: int,
+    per_act_token: bool,
+    per_out_ch: bool,
+    mkn: tuple[int, int, int],
+):
     label = "Quant Matmul"
 
     sub_label = (
-        "{}, num_experts={}, topk={}, per_act_token={} per_out_ch={}, "
-        "MKN=({})".format(model, num_experts, topk, per_act_token, per_out_ch,
-                          mkn))
+        "{}, num_experts={}, topk={}, per_act_token={} per_out_ch={}, MKN=({})".format(
+            model, num_experts, topk, per_act_token, per_out_ch, mkn
+        )
+    )
 
     print(f"Testing: {sub_label}")
 
@@ -50,122 +63,142 @@ def bench_run(results: list[benchmark.Measurement], model: str,
 
     _, a_scale = ops.scaled_fp8_quant(a)
 
-    w1_q = torch.empty((num_experts, 2 * n, k),
-                       device="cuda",
-                       dtype=torch.float8_e4m3fn)
-    w2_q = torch.empty((num_experts, k, n),
-                       device="cuda",
-                       dtype=torch.float8_e4m3fn)
-    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)
+    w1_q = torch.empty(
+        (num_experts, 2 * n, k), device="cuda", dtype=torch.float8_e4m3fn
+    )
+    w2_q = torch.empty((num_experts, k, n), device="cuda", dtype=torch.float8_e4m3fn)
+    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)
 
     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)
 
-    topk_weights, topk_ids = fused_topk(a, score, topk, renormalize=False)
+    topk_weights, topk_ids, token_expert_indices = fused_topk(
+        a, score, topk, renormalize=False
+    )
 
-    def run_triton_moe(a: torch.Tensor, w1: torch.Tensor, w2: torch.Tensor,
-                       topk_weights: torch.Tensor, topk_ids: torch.Tensor,
-                       w1_scale: torch.Tensor, w2_scale: torch.Tensor,
-                       a_scale: torch.Tensor, num_repeats: int):
-        for _ in range(num_repeats):
-            fused_experts(a,
-                          w1,
-                          w2,
-                          topk_weights,
-                          topk_ids,
-                          use_fp8_w8a8=True,
-                          w1_scale=w1_scale,
-                          w2_scale=w2_scale,
-                          a1_scale=a_scale)
+    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_cutlass_moe(a: torch.Tensor, a_scale: torch.Tensor,
-                        w1: torch.Tensor, w2: torch.Tensor,
-                        w1_scale: torch.Tensor, w2_scale: torch.Tensor,
-                        topk_weights: torch.Tensor, topk_ids: torch.Tensor,
-                        ab_strides1: torch.Tensor, c_strides1: torch.Tensor,
-                        ab_strides2: torch.Tensor, c_strides2: torch.Tensor,
-                        num_repeats: int):
+    def run_triton_moe(
+        a: torch.Tensor,
+        w1: torch.Tensor,
+        w2: torch.Tensor,
+        topk_weights: torch.Tensor,
+        topk_ids: torch.Tensor,
+        w1_scale: torch.Tensor,
+        w2_scale: torch.Tensor,
+        a_scale: torch.Tensor,
+        num_repeats: int,
+    ):
         for _ in range(num_repeats):
-            cutlass_moe_fp8(a,
-                            w1,
-                            w2,
-                            w1_scale,
-                            w2_scale,
-                            topk_weights,
-                            topk_ids,
-                            ab_strides1,
-                            c_strides1,
-                            ab_strides2,
-                            c_strides2,
-                            a1_scale=a_scale)
+            fused_experts(
+                a,
+                w1,
+                w2,
+                topk_weights,
+                topk_ids,
+                use_fp8_w8a8=True,
+                w1_scale=w1_scale,
+                w2_scale=w2_scale,
+                a1_scale=a_scale,
+            )
+
+    def run_cutlass_moe(
+        a: torch.Tensor,
+        a_scale: torch.Tensor,
+        w1: torch.Tensor,
+        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,
+        per_act_token: bool,
+        num_repeats: int,
+    ):
+        for _ in range(num_repeats):
+            cutlass_moe_fp8(
+                a,
+                w1,
+                w2,
+                topk_weights,
+                topk_ids,
+                w1_scale,
+                w2_scale,
+                ab_strides1,
+                ab_strides2,
+                c_strides1,
+                c_strides2,
+                per_act_token,
+                a1_scale=None,
+            )
 
     def run_cutlass_from_graph(
-            a: torch.Tensor, a_scale: torch.Tensor, w1_q: torch.Tensor,
-            w2_q: torch.Tensor, w1_scale: torch.Tensor, w2_scale: torch.Tensor,
-            topk_weights: torch.Tensor, topk_ids: torch.Tensor,
-            ab_strides1: torch.Tensor, c_strides1: torch.Tensor,
-            ab_strides2: torch.Tensor, c_strides2: torch.Tensor):
+        a: torch.Tensor,
+        a_scale: torch.Tensor,
+        w1_q: torch.Tensor,
+        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,
+    ):
         with set_current_vllm_config(
-                VllmConfig(parallel_config=ParallelConfig(
-                    pipeline_parallel_size=1))):
-            return cutlass_moe_fp8(a,
-                                   w1_q,
-                                   w2_q,
-                                   w1_scale,
-                                   w2_scale,
-                                   topk_weights,
-                                   topk_ids,
-                                   ab_strides1,
-                                   c_strides1,
-                                   ab_strides2,
-                                   c_strides2,
-                                   a1_scale=a_scale)
+            VllmConfig(parallel_config=ParallelConfig(pipeline_parallel_size=1))
+        ):
+            return cutlass_moe_fp8(
+                a,
+                w1_q,
+                w2_q,
+                topk_weights,
+                topk_ids,
+                w1_scale,
+                w2_scale,
+                ab_strides1,
+                ab_strides2,
+                c_strides1,
+                c_strides2,
+                per_act_token,
+                a1_scale=None,
+            )
 
-    def run_triton_from_graph(a: torch.Tensor, w1: torch.Tensor,
-                              w2: torch.Tensor, topk_weights: torch.Tensor,
-                              topk_ids: torch.Tensor, w1_scale: torch.Tensor,
-                              w2_scale: torch.Tensor, a_scale: torch.Tensor):
+    def run_triton_from_graph(
+        a: torch.Tensor,
+        w1: torch.Tensor,
+        w2: torch.Tensor,
+        topk_weights: torch.Tensor,
+        topk_ids: torch.Tensor,
+        w1_scale: torch.Tensor,
+        w2_scale: torch.Tensor,
+        a_scale: torch.Tensor,
+    ):
         with set_current_vllm_config(
-                VllmConfig(parallel_config=ParallelConfig(
-                    pipeline_parallel_size=1))):
-            return fused_experts(a,
-                                 w1,
-                                 w2,
-                                 topk_weights,
-                                 topk_ids,
-                                 use_fp8_w8a8=True,
-                                 w1_scale=w1_scale,
-                                 w2_scale=w2_scale,
-                                 a1_scale=a_scale)
+            VllmConfig(parallel_config=ParallelConfig(pipeline_parallel_size=1))
+        ):
+            return fused_experts(
+                a,
+                w1,
+                w2,
+                topk_weights,
+                topk_ids,
+                use_fp8_w8a8=True,
+                w1_scale=w1_scale,
+                w2_scale=w2_scale,
+                a1_scale=a_scale,
+            )
 
     def replay_graph(graph, num_repeats):
         for _ in range(num_repeats):
@@ -175,16 +208,35 @@ def bench_run(results: list[benchmark.Measurement], model: str,
     cutlass_stream = torch.cuda.Stream()
     cutlass_graph = torch.cuda.CUDAGraph()
     with torch.cuda.graph(cutlass_graph, stream=cutlass_stream):
-        run_cutlass_from_graph(a, a_scale, w1_q, w2_q, w1_scale, w2_scale,
-                               topk_weights, topk_ids, ab_strides1, c_strides1,
-                               ab_strides2, c_strides2)
+        run_cutlass_from_graph(
+            a,
+            a_scale,
+            w1_q,
+            w2_q,
+            w1_scale,
+            w2_scale,
+            ab_strides1,
+            ab_strides2,
+            c_strides1,
+            c_strides2,
+            topk_weights,
+            topk_ids,
+        )
     torch.cuda.synchronize()
 
     triton_stream = torch.cuda.Stream()
     triton_graph = torch.cuda.CUDAGraph()
     with torch.cuda.graph(triton_graph, stream=triton_stream):
-        run_triton_from_graph(a, w1_q_notransp, w2_q_notransp, topk_weights,
-                              topk_ids, w1_scale, w2_scale, a_scale)
+        run_triton_from_graph(
+            a,
+            w1_q,
+            w2_q,
+            topk_weights,
+            topk_ids,
+            w1_scale,
+            w2_scale,
+            a_scale,
+        )
     torch.cuda.synchronize()
 
     min_run_time = 5
@@ -197,17 +249,16 @@ def bench_run(results: list[benchmark.Measurement], model: str,
         "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,
@@ -224,18 +275,27 @@ def bench_run(results: list[benchmark.Measurement], model: str,
     }
 
     # Warmup
-    run_triton_moe(a, w1_q_notransp, w2_q_notransp, topk_weights, topk_ids,
-                   w1_scale, w2_scale, a_scale, num_warmup)
+    run_triton_moe(
+        a,
+        w1_q,
+        w2_q,
+        topk_weights,
+        topk_ids,
+        w1_scale,
+        w2_scale,
+        a_scale,
+        num_warmup,
+    )
 
     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,
             description="triton_moe",
-        ).blocked_autorange(min_run_time=min_run_time))
+        ).blocked_autorange(min_run_time=min_run_time)
+    )
 
     # Warmup
     replay_graph(triton_graph, num_warmup)
@@ -247,22 +307,36 @@ def bench_run(results: list[benchmark.Measurement], model: str,
             label=label,
             sub_label=sub_label,
             description="triton_moe_cuda_graphs",
-        ).blocked_autorange(min_run_time=min_run_time))
+        ).blocked_autorange(min_run_time=min_run_time)
+    )
 
     # Warmup
-    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_warmup)
+    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_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,
             description="grouped_gemm_moe",
-        ).blocked_autorange(min_run_time=min_run_time))
+        ).blocked_autorange(min_run_time=min_run_time)
+    )
 
     # Warmup
     replay_graph(cutlass_graph, num_warmup)
@@ -274,7 +348,8 @@ def bench_run(results: list[benchmark.Measurement], model: str,
             label=label,
             sub_label=sub_label,
             description="grouped_gemm_moe_cuda_graphs",
-        ).blocked_autorange(min_run_time=min_run_time))
+        ).blocked_autorange(min_run_time=min_run_time)
+    )
 
 
 def main(args):
@@ -302,8 +377,15 @@ def main(args):
                     for per_out_ch in PER_OUT_CH_OPTS:
                         for size_m in DEFAULT_BATCH_SIZES:
                             mkn = (size_m, size_k, size_n)
-                            bench_run(results, model, num_experts, topk,
-                                      per_act_token, per_out_ch, mkn)
+                            bench_run(
+                                results,
+                                model,
+                                num_experts,
+                                topk,
+                                per_act_token,
+                                per_out_ch,
+                                mkn,
+                            )
 
     compare = benchmark.Compare(results)
     compare.print()
@@ -311,7 +393,8 @@ def main(args):
 
 if __name__ == "__main__":
     parser = FlexibleArgumentParser(
-        description="Benchmark Marlin across specified models/shapes/batches")
+        description="Benchmark Marlin across specified models/shapes/batches"
+    )
     parser.add_argument(
         "--models",
         nargs="+",
@@ -319,21 +402,14 @@ if __name__ == "__main__":
         default=DEFAULT_MODELS,
         choices=WEIGHT_SHAPES_MOE.keys(),
     )
-    parser.add_argument("--tp-sizes",
-                        nargs="+",
-                        type=int,
-                        default=DEFAULT_TP_SIZES)
-    parser.add_argument("--batch-sizes",
-                        nargs="+",
-                        type=int,
-                        default=DEFAULT_BATCH_SIZES)
+    parser.add_argument("--tp-sizes", nargs="+", type=int, default=DEFAULT_TP_SIZES)
+    parser.add_argument(
+        "--batch-sizes", nargs="+", type=int, default=DEFAULT_BATCH_SIZES
+    )
     parser.add_argument("--limit-k", nargs="+", type=int, default=[])
     parser.add_argument("--limit-n", nargs="+", type=int, default=[])
     parser.add_argument("--limit-num-groups", nargs="+", type=int, default=[])
-    parser.add_argument("--limit-per-act-token",
-                        nargs="+",
-                        type=int,
-                        default=[])
+    parser.add_argument("--limit-per-act-token", nargs="+", type=int, default=[])
     parser.add_argument("--limit-per-out-ch", nargs="+", type=int, default=[])
 
     args = parser.parse_args()

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
 
@@ -10,14 +11,16 @@ from vllm.utils import STR_DTYPE_TO_TORCH_DTYPE, FlexibleArgumentParser
 
 
 @torch.inference_mode()
-def main(num_tokens: int,
-         hidden_size: int,
-         add_residual: bool,
-         dtype: torch.dtype,
-         seed: int = 0,
-         do_profile: bool = False,
-         num_warmup_iters: int = 5,
-         num_iters: int = 100) -> None:
+def main(
+    num_tokens: int,
+    hidden_size: int,
+    add_residual: bool,
+    dtype: torch.dtype,
+    seed: int = 0,
+    do_profile: bool = False,
+    num_warmup_iters: int = 5,
+    num_iters: int = 100,
+) -> None:
     current_platform.seed_everything(seed)
     torch.set_default_device("cuda")
 
@@ -56,33 +59,35 @@ def main(num_tokens: int,
     print(f"Kernel running time: {latency * 1000000:.3f} us")
 
 
-if __name__ == '__main__':
-    parser = FlexibleArgumentParser(
-        description="Benchmark the layernorm kernel.")
+if __name__ == "__main__":
+    parser = FlexibleArgumentParser(description="Benchmark the layernorm kernel.")
     parser.add_argument("--num-tokens", type=int, default=4096)
     parser.add_argument("--hidden-size", type=int, default=8192)
     parser.add_argument("--add-residual", action="store_true")
-    parser.add_argument("--dtype",
-                        type=str,
-                        choices=["half", "bfloat16", "float"],
-                        default="half")
+    parser.add_argument(
+        "--dtype", type=str, choices=["half", "bfloat16", "float"], default="half"
+    )
     parser.add_argument("--seed", type=int, default=0)
     parser.add_argument("--profile", action="store_true")
     parser.add_argument("--num-warmup-iters", type=int, default=5)
-    parser.add_argument("--num-iters",
-                        type=int,
-                        default=100,
-                        help="Number of benchmark iterations. "
-                        "If --profile is set, this number is ignored")
+    parser.add_argument(
+        "--num-iters",
+        type=int,
+        default=100,
+        help="Number of benchmark iterations. "
+        "If --profile is set, this number is ignored",
+    )
 
     args = parser.parse_args()
     print(args)
 
-    main(num_tokens=args.num_tokens,
-         hidden_size=args.hidden_size,
-         add_residual=args.add_residual,
-         dtype=STR_DTYPE_TO_TORCH_DTYPE[args.dtype],
-         seed=args.seed,
-         do_profile=args.profile,
-         num_warmup_iters=args.num_warmup_iters,
-         num_iters=args.num_iters)
+    main(
+        num_tokens=args.num_tokens,
+        hidden_size=args.hidden_size,
+        add_residual=args.add_residual,
+        dtype=STR_DTYPE_TO_TORCH_DTYPE[args.dtype],
+        seed=args.seed,
+        do_profile=args.profile,
+        num_warmup_iters=args.num_warmup_iters,
+        num_iters=args.num_iters,
+    )

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
@@ -20,12 +21,18 @@ from weight_shapes import WEIGHT_SHAPES
 
 from vllm import _custom_ops as ops
 from vllm.model_executor.layers.quantization.utils.marlin_utils import (
-    GPTQ_MARLIN_MAX_PARALLEL, GPTQ_MARLIN_MIN_THREAD_N, marlin_permute_scales,
-    marlin_zero_points)
+    GPTQ_MARLIN_MAX_PARALLEL,
+    GPTQ_MARLIN_MIN_THREAD_N,
+    marlin_permute_scales,
+    marlin_zero_points,
+)
 from vllm.model_executor.layers.quantization.utils.marlin_utils_test import (
-    MarlinWorkspace)
+    MarlinWorkspace,
+)
 from vllm.model_executor.layers.quantization.utils.quant_utils import (
-    pack_rows, quantize_weights)
+    pack_rows,
+    quantize_weights,
+)
 from vllm.scalar_type import ScalarType, scalar_types
 from vllm.utils import FlexibleArgumentParser
 
@@ -82,12 +89,14 @@ def rand_data(shape, dtype=torch.float16, scale=1):
         return torch.randint(-15, 15, shape, dtype=dtype, device="cuda")
 
 
-def quantize_and_pack(atype: torch.dtype,
-                      w: torch.Tensor,
-                      wtype: ScalarType,
-                      stype: Optional[torch.dtype],
-                      group_size: Optional[int],
-                      zero_points: bool = False):
+def quantize_and_pack(
+    atype: torch.dtype,
+    w: torch.Tensor,
+    wtype: ScalarType,
+    stype: Optional[torch.dtype],
+    group_size: Optional[int],
+    zero_points: bool = False,
+):
     assert wtype.is_integer(), "TODO: support floating point weights"
 
     w_ref, w_q, w_s, w_zp = quantize_weights(
@@ -96,21 +105,24 @@ def quantize_and_pack(atype: torch.dtype,
         group_size=group_size,
         zero_points=zero_points,
         # to match how the kernel applies zps
-        ref_zero_points_after_scales=True)
+        ref_zero_points_after_scales=True,
+    )
 
     w_q = pack_rows(w_q, wtype.size_bits, *w_q.shape)
     return w_ref, w_q, w_s, w_zp
 
 
-def create_bench_tensors(shape: tuple[int, int, int], types: TypeConfig,
-                         group_size: Optional[int]) -> list[BenchmarkTensors]:
+def create_bench_tensors(
+    shape: tuple[int, int, int], types: TypeConfig, group_size: Optional[int]
+) -> list[BenchmarkTensors]:
     m, n, k = shape
 
     # we want to make sure that weights don't fit into L2 cache between runs so
     #  we construct enough weights to exceed L2 cache, which is 50mb on a H100
     #  so we target total weight size > 2*50mb
-    num_weights = math.ceil(2 * 50 * 1024**2 * 8 /
-                            (k * n * types.weight_type.size_bits))
+    num_weights = math.ceil(
+        2 * 50 * 1024**2 * 8 / (k * n * types.weight_type.size_bits)
+    )
 
     a = rand_data((m, k), types.act_type, scale=5)
 
@@ -124,8 +136,13 @@ def create_bench_tensors(shape: tuple[int, int, int], types: TypeConfig,
             w = w.to(torch.float16)
 
         w_ref, w_q_packed, w_s, w_zp = quantize_and_pack(
-            a.dtype, w, types.weight_type, types.group_scale_type, group_size,
-            types.group_zero_type is not None)
+            a.dtype,
+            w,
+            types.weight_type,
+            types.group_scale_type,
+            group_size,
+            types.group_zero_type is not None,
+        )
 
         if not a.dtype.is_floating_point:
             aiinfo = torch.iinfo(a.dtype)
@@ -133,21 +150,30 @@ def create_bench_tensors(shape: tuple[int, int, int], types: TypeConfig,
 
         w_ref = w_ref.to(torch.float32)
 
-        w_ch_s = None if types.channel_scale_type is None else\
-            rand_data((n,), types.channel_scale_type)
-        w_tok_s = None if types.token_scale_type is None else\
-            rand_data((m,), types.token_scale_type)
+        w_ch_s = (
+            None
+            if types.channel_scale_type is None
+            else rand_data((n,), types.channel_scale_type)
+        )
+        w_tok_s = (
+            None
+            if types.token_scale_type is None
+            else rand_data((m,), types.token_scale_type)
+        )
 
         benchmark_tensors.append(
-            BenchmarkTensors(w_ref=w_ref,
-                             a=a,
-                             w_q=w_q_packed,
-                             wtype=types.weight_type,
-                             w_g_s=w_s,
-                             w_g_zp=w_zp,
-                             group_size=group_size,
-                             w_ch_s=w_ch_s,
-                             w_tok_s=w_tok_s))
+            BenchmarkTensors(
+                w_ref=w_ref,
+                a=a,
+                w_q=w_q_packed,
+                wtype=types.weight_type,
+                w_g_s=w_s,
+                w_g_zp=w_zp,
+                group_size=group_size,
+                w_ch_s=w_ch_s,
+                w_tok_s=w_tok_s,
+            )
+        )
 
     return benchmark_tensors
 
@@ -170,50 +196,59 @@ def cutlass_scaled_mm_create_bench_fn(bt: BenchmarkTensors) -> Callable:
         scale_b = torch.tensor(1.0, dtype=torch.float32, device=bt.a.device)
     w_col_major = bt.w_ref.to(bt.a.dtype).t().contiguous().t()
     return lambda: ops.cutlass_scaled_mm(
-        bt.a, w_col_major, scale_a, scale_b, out_dtype=torch.float16)
+        bt.a, w_col_major, scale_a, scale_b, out_dtype=torch.float16
+    )
 
 
 def marlin_create_bench_fn(bt: BenchmarkTensors) -> Callable:
     device = bt.a.device
 
-    workspace = MarlinWorkspace(bt.w_ref.shape[1], GPTQ_MARLIN_MIN_THREAD_N,
-                                GPTQ_MARLIN_MAX_PARALLEL)
+    workspace = MarlinWorkspace(
+        bt.w_ref.shape[1], GPTQ_MARLIN_MIN_THREAD_N, GPTQ_MARLIN_MAX_PARALLEL
+    )
 
     if bt.w_g_zp is None:
         w_zp = torch.empty(0, dtype=torch.int, device=device)
     else:
-        w_zp = marlin_zero_points(bt.w_g_zp, bt.w_ref.shape[0],
-                                  bt.w_ref.shape[1], bt.wtype.size_bits)
+        w_zp = marlin_zero_points(
+            bt.w_g_zp, bt.w_ref.shape[0], bt.w_ref.shape[1], bt.wtype.size_bits
+        )
 
     if bt.group_size is None:
         w_s = torch.tensor([], device="cuda", dtype=torch.half)
     else:
-        w_s = marlin_permute_scales(bt.w_g_s, bt.w_ref.shape[0],
-                                    bt.w_ref.shape[1], bt.group_size)
+        w_s = marlin_permute_scales(
+            bt.w_g_s, bt.w_ref.shape[0], bt.w_ref.shape[1], bt.group_size
+        )
 
     sort_indices = torch.empty(0, dtype=torch.int, device=device)
     g_idx = torch.empty(0, dtype=torch.int, device=device)
-    w_q = ops.gptq_marlin_repack(bt.w_q, sort_indices, bt.w_ref.shape[0],
-                                 bt.w_ref.shape[1], bt.wtype.size_bits)
+    w_q = ops.gptq_marlin_repack(
+        bt.w_q, sort_indices, bt.w_ref.shape[0], bt.w_ref.shape[1], bt.wtype.size_bits
+    )
 
     if bt.a.dtype.is_floating_point:
         assert bt.w_ch_s is None
         assert bt.w_tok_s is None
         assert bt.group_size is not None
 
-        fn = lambda: ops.gptq_marlin_gemm(a=bt.a,
-                                          b_q_weight=w_q,
-                                          b_scales=w_s,
-                                          b_zeros=w_zp,
-                                          g_idx=g_idx,
-                                          perm=sort_indices,
-                                          workspace=workspace.scratch,
-                                          b_q_type=bt.wtype,
-                                          size_m=bt.a.shape[0],
-                                          size_n=bt.w_ref.shape[1],
-                                          size_k=bt.w_ref.shape[0],
-                                          is_k_full=True,
-                                          is_zp_float=False)
+        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,
+            workspace=workspace.scratch,
+            b_q_type=bt.wtype,
+            size_m=bt.a.shape[0],
+            size_n=bt.w_ref.shape[1],
+            size_k=bt.w_ref.shape[0],
+            is_k_full=True,
+            is_zp_float=False,
+        )
     else:
         assert bt.a.dtype == torch.int8
         assert bt.wtype == scalar_types.uint4b8
@@ -221,36 +256,35 @@ def marlin_create_bench_fn(bt: BenchmarkTensors) -> Callable:
         if bt.w_ch_s is not None:
             s_ch = bt.w_ch_s.to(torch.float32)
         else:
-            s_ch = torch.ones(bt.w_ref.shape[1],
-                              dtype=torch.float32,
-                              device=device)
+            s_ch = torch.ones(bt.w_ref.shape[1], dtype=torch.float32, device=device)
 
         if bt.w_tok_s is not None:
             s_tok = bt.w_tok_s.to(torch.float32)
         else:
-            s_tok = torch.ones(bt.a.shape[0],
-                               dtype=torch.float32,
-                               device=device)
+            s_tok = torch.ones(bt.a.shape[0], dtype=torch.float32, device=device)
 
-        fn = lambda: ops.marlin_qqq_gemm(a=bt.a,
-                                         b_q_weight=w_q,
-                                         s_group=w_s,
-                                         s_tok=s_tok,
-                                         s_ch=s_ch,
-                                         workspace=workspace.scratch,
-                                         size_m=bt.a.shape[0],
-                                         size_n=bt.w_ref.shape[1],
-                                         size_k=bt.w_ref.shape[0])
+        fn = lambda: ops.marlin_qqq_gemm(
+            a=bt.a,
+            b_q_weight=w_q,
+            s_group=w_s,
+            s_tok=s_tok,
+            s_ch=s_ch,
+            workspace=workspace.scratch,
+            size_m=bt.a.shape[0],
+            size_n=bt.w_ref.shape[1],
+            size_k=bt.w_ref.shape[0],
+        )
 
     return fn
 
 
-def machete_create_bench_fn(bt: BenchmarkTensors,
-                            out_type=torch.dtype,
-                            schedule=None) -> Callable:
+def machete_create_bench_fn(
+    bt: BenchmarkTensors, out_type=torch.dtype, schedule=None
+) -> Callable:
     w_q = bt.w_q.t().contiguous().t()  # make col major
-    w_q = ops.machete_prepack_B(w_q, bt.a.dtype, bt.wtype,
-                                None if bt.w_g_s is None else bt.w_g_s.dtype)
+    w_q = ops.machete_prepack_B(
+        w_q, bt.a.dtype, bt.wtype, None if bt.w_g_s is None else bt.w_g_s.dtype
+    )
 
     w_g_zp = bt.w_g_zp
     if w_g_zp is not None:
@@ -275,26 +309,24 @@ def machete_create_bench_fn(bt: BenchmarkTensors,
 # bench
 
 
-def bench_fns(label: str, sub_label: str, description: str,
-              fns: list[Callable]):
-
+def bench_fns(label: str, sub_label: str, description: str, fns: list[Callable]):
     min_run_time = 1 if not NVTX_PROFILE else 0.1
     res = TBenchmark.Timer(
         stmt="""
         for fn in fns:
             fn()
         """,
-        globals={
-            "fns": fns
-        },
+        globals={"fns": fns},
         label=label,
         sub_label=sub_label,
         description=description,
     ).blocked_autorange(min_run_time=min_run_time)
 
     if NVTX_PROFILE:
-        with nvtx.annotate("mm-bench"), nvtx.annotate(
-                f"{label}|{sub_label}|{description}"):
+        with (
+            nvtx.annotate("mm-bench"),
+            nvtx.annotate(f"{label}|{sub_label}|{description}"),
+        ):
             fns[0]()
 
     return res
@@ -304,19 +336,20 @@ _SWEEP_SCHEDULES_RESULTS: Optional[pd.DataFrame] = None
 _SWEEP_SCHEDULES_RESULTS_CSV: Optional[str] = None
 
 
-def bench(types: TypeConfig,
-          group_size: int,
-          m: int,
-          k: int,
-          n: int,
-          label: str,
-          sub_label: str,
-          sweep_schedules: bool = True) -> list[TMeasurement]:
+def bench(
+    types: TypeConfig,
+    group_size: int,
+    m: int,
+    k: int,
+    n: int,
+    label: str,
+    sub_label: str,
+    sweep_schedules: bool = True,
+) -> list[TMeasurement]:
     benchmark_tensors = create_bench_tensors((m, n, k), types, group_size)
     sub_label += f", L={len(benchmark_tensors)}"
 
-    name_type_string = f"W{types.weight_type}"+\
-                       f"-A{terse_type_name(types.act_type)}"
+    name_type_string = f"W{types.weight_type}" + f"-A{terse_type_name(types.act_type)}"
     if types.group_scale_type is not None:
         name_type_string += f"-GS{terse_type_name(types.group_scale_type)}"
     if types.group_zero_type is not None:
@@ -332,31 +365,45 @@ def bench(types: TypeConfig,
     # pytorch impl
     timers.append(
         bench_fns(
-            label, sub_label, "torch.matmul (fp16)",
-            [torch_matmul_f16_create_bench_fn(bt)
-             for bt in benchmark_tensors]))
+            label,
+            sub_label,
+            "torch.matmul (fp16)",
+            [torch_matmul_f16_create_bench_fn(bt) for bt in benchmark_tensors],
+        )
+    )
 
     if types.act_type == torch.int8 or types.act_type == torch.float8_e4m3fn:
         timers.append(
             bench_fns(
-                label, sub_label,
-                f"cutlass_scaled_mm ({terse_type_name(types.act_type)})", [
-                    cutlass_scaled_mm_create_bench_fn(bt)
-                    for bt in benchmark_tensors
-                ]))
+                label,
+                sub_label,
+                f"cutlass_scaled_mm ({terse_type_name(types.act_type)})",
+                [cutlass_scaled_mm_create_bench_fn(bt) for bt in benchmark_tensors],
+            )
+        )
 
     if types.act_type != torch.float8_e4m3fn:
         timers.append(
-            bench_fns(label, sub_label, f"marlin ({name_type_string})",
-                      [marlin_create_bench_fn(bt)
-                       for bt in benchmark_tensors]))
+            bench_fns(
+                label,
+                sub_label,
+                f"marlin ({name_type_string})",
+                [marlin_create_bench_fn(bt) for bt in benchmark_tensors],
+            )
+        )
 
     # machete
     timers.append(
-        bench_fns(label, sub_label, f"machete ({name_type_string})", [
-            machete_create_bench_fn(bt, out_type=types.output_type)
-            for bt in benchmark_tensors
-        ]))
+        bench_fns(
+            label,
+            sub_label,
+            f"machete ({name_type_string})",
+            [
+                machete_create_bench_fn(bt, out_type=types.output_type)
+                for bt in benchmark_tensors
+            ],
+        )
+    )
 
     if sweep_schedules:
         global _SWEEP_SCHEDULES_RESULTS
@@ -371,7 +418,8 @@ def bench(types: TypeConfig,
             group_zeros_type=types.group_zero_type,
             token_scales_type=types.token_scale_type,
             channel_scales_type=types.channel_scale_type,
-            out_type=types.output_type)
+            out_type=types.output_type,
+        )
 
         if schedules is None or len(schedules) == 0:
             raise ValueError("No schedules found to sweep")
@@ -383,11 +431,17 @@ def bench(types: TypeConfig,
             if schedule_M >= 2 * max(m, 16) or schedule_M < m // 4:
                 continue
 
-            res = bench_fns(label, sub_label, "machete_best", [
-                machete_create_bench_fn(
-                    bt, out_type=types.output_type, schedule=schedule)
-                for bt in benchmark_tensors
-            ])
+            res = bench_fns(
+                label,
+                sub_label,
+                "machete_best",
+                [
+                    machete_create_bench_fn(
+                        bt, out_type=types.output_type, schedule=schedule
+                    )
+                    for bt in benchmark_tensors
+                ],
+            )
 
             results_row = {
                 "M": m,
@@ -398,10 +452,8 @@ def bench(types: TypeConfig,
                 "median": res.median,
             }
             if _SWEEP_SCHEDULES_RESULTS is None:
-                _SWEEP_SCHEDULES_RESULTS = pd.DataFrame(
-                    columns=results_row.keys())
-            _SWEEP_SCHEDULES_RESULTS.\
-                loc[len(_SWEEP_SCHEDULES_RESULTS)] = results_row
+                _SWEEP_SCHEDULES_RESULTS = pd.DataFrame(columns=results_row.keys())
+            _SWEEP_SCHEDULES_RESULTS.loc[len(_SWEEP_SCHEDULES_RESULTS)] = results_row
 
             print(f"  {res.median:5.5} ", schedule)
             if not best or res.median < best.median:
@@ -422,8 +474,9 @@ def print_timers(timers: list[TMeasurement]):
 def run(args, MKNs: Iterable[tuple[int, int, int]]) -> Iterable[TMeasurement]:
     types = TypeConfig(
         act_type=args.act_type,
-        weight_type=scalar_types.uint4b8 if args.group_zero_type is None \
-            else scalar_types.uint4,
+        weight_type=scalar_types.uint4b8
+        if args.group_zero_type is None
+        else scalar_types.uint4,
         output_type=args.out_type,
         group_scale_type=args.group_scale_type,
         group_zero_type=args.group_zero_type,
@@ -433,14 +486,16 @@ def run(args, MKNs: Iterable[tuple[int, int, int]]) -> Iterable[TMeasurement]:
 
     results: list[TMeasurement] = []
     for m, k, n in MKNs:
-        timers = bench(types,
-                       args.group_size,
-                       m,
-                       k,
-                       n,
-                       f"{args.act_type}-gemm",
-                       f"MKN=({m}x{k}x{n})",
-                       sweep_schedules=args.sweep_schedules)
+        timers = bench(
+            types,
+            args.group_size,
+            m,
+            k,
+            n,
+            f"{args.act_type}-gemm",
+            f"MKN=({m}x{k}x{n})",
+            sweep_schedules=args.sweep_schedules,
+        )
         print_timers(timers)
         results.extend(timers)
 
@@ -454,7 +509,6 @@ def make_output(
     base_description: str,
     timestamp=None,
 ):
-
     print(f"== All Results {base_description} ====")
     print_timers(data)
 
@@ -468,8 +522,7 @@ def make_output(
 
 
 def run_square_bench(args):
-    dim_sizes = list(
-        range(args.dim_start, args.dim_end + 1, args.dim_increment))
+    dim_sizes = list(range(args.dim_start, args.dim_end + 1, args.dim_increment))
     MKNs = list(zip(dim_sizes, dim_sizes, dim_sizes))
     data = run(args.dtype, args.sweep_schedules, MKNs)
 
@@ -479,8 +532,9 @@ def run_square_bench(args):
 def run_range_bench(args):
     m_start, k_start, n_start = (int(x) for x in args.dim_start.split(","))
     m_end, k_end, n_end = (int(x) for x in args.dim_end.split(","))
-    m_increment, k_increment, n_increment = \
-        (int(x) for x in args.dim_increment.split(","))
+    m_increment, k_increment, n_increment = (
+        int(x) for x in args.dim_increment.split(",")
+    )
     Ms = list(range(m_start, m_end + 1, m_increment))
     Ks = list(range(k_start, k_end + 1, k_increment))
     Ns = list(range(n_start, n_end + 1, n_increment))
@@ -492,7 +546,6 @@ def run_range_bench(args):
 
 
 def run_model_bench(args):
-
     print("Benchmarking models:")
     for i, model in enumerate(args.models):
         print(f"[{i}]  {model}")
@@ -535,10 +588,13 @@ def run_model_bench(args):
     with open(f"model_bench-{type_string}-{timestr}.pkl", "wb") as f:
         args_dict = vars(args)
         args_dict.pop("func")
-        pkl.dump({
-            "args": args_dict,
-            "results": all_results,
-        }, f)
+        pkl.dump(
+            {
+                "args": args_dict,
+                "results": all_results,
+            },
+            f,
+        )
 
 
 if __name__ == "__main__":
@@ -554,7 +610,6 @@ if __name__ == "__main__":
         }[dt]
 
     class ToTorchDtype(argparse.Action):
-
         def __call__(self, parser, namespace, values, option_string=None):
             setattr(namespace, self.dest, to_torch_dtype(values))
 
@@ -580,32 +635,32 @@ Benchmark Machete GEMM.
         "--act-type",
         action=ToTorchDtype,
         required=True,
-        choices=['bfloat16', 'float16', 'int8', 'float8_e4m3fn'],
+        choices=["bfloat16", "float16", "int8", "float8_e4m3fn"],
     )
     parser.add_argument(
         "--group-scale-type",
         action=ToTorchDtype,
-        choices=['bfloat16', 'float16'],
+        choices=["bfloat16", "float16"],
     )
     parser.add_argument(
         "--group-zero-type",
         type=to_torch_dtype,
-        choices=['bfloat16', 'float16'],
+        choices=["bfloat16", "float16"],
     )
     parser.add_argument(
         "--channel-scale-type",
         action=ToTorchDtype,
-        choices=['float'],
+        choices=["float"],
     )
     parser.add_argument(
         "--token-scale-type",
         action=ToTorchDtype,
-        choices=['float'],
+        choices=["float"],
     )
     parser.add_argument(
         "--out-type",
         action=ToTorchDtype,
-        choices=['bfloat16', 'float16'],
+        choices=["bfloat16", "float16"],
     )
     parser.add_argument(
         "--group-size",
@@ -618,9 +673,11 @@ Benchmark Machete GEMM.
         action="store_true",
         help="Run a sweep over all supported schedules",
     )
-    parser.add_argument("--sweep-csv-out",
-                        help="CSV to store sweep results",
-                        default="sch_sweep_results.csv")
+    parser.add_argument(
+        "--sweep-csv-out",
+        help="CSV to store sweep results",
+        default="sch_sweep_results.csv",
+    )
     subparsers = parser.add_subparsers(dest="cmd", required=True)
 
     square_parser = subparsers.add_parser("square_bench")
@@ -634,17 +691,20 @@ Benchmark Machete GEMM.
         "--dim-start",
         type=str,
         required=True,
-        help="Start value for M,K,N as common separated list")
+        help="Start value for M,K,N as common separated list",
+    )
     range_parser.add_argument(
         "--dim-end",
         type=str,
         required=True,
-        help="End value (inclusive) for M,K,N as common separated list")
+        help="End value (inclusive) for M,K,N as common separated list",
+    )
     range_parser.add_argument(
         "--dim-increment",
         type=str,
         required=True,
-        help="Increment value for M,K,N as common separated list")
+        help="Increment value for M,K,N as common separated list",
+    )
     range_parser.set_defaults(func=run_range_bench)
 
     model_parser = subparsers.add_parser("model_bench")
@@ -655,14 +715,12 @@ Benchmark Machete GEMM.
         default=DEFAULT_MODELS,
         choices=WEIGHT_SHAPES.keys(),
     )
-    model_parser.add_argument("--tp-sizes",
-                              nargs="+",
-                              type=int,
-                              default=DEFAULT_TP_SIZES)
-    model_parser.add_argument("--batch-sizes",
-                              nargs="+",
-                              type=int,
-                              default=DEFAULT_BATCH_SIZES)
+    model_parser.add_argument(
+        "--tp-sizes", nargs="+", type=int, default=DEFAULT_TP_SIZES
+    )
+    model_parser.add_argument(
+        "--batch-sizes", nargs="+", type=int, default=DEFAULT_BATCH_SIZES
+    )
     model_parser.set_defaults(func=run_model_bench)
 
     args = parser.parse_args()

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
@@ -6,20 +7,43 @@ from benchmark_shapes import WEIGHT_SHAPES
 
 from vllm import _custom_ops as ops
 from vllm.model_executor.layers.quantization.gptq_marlin_24 import (
-    GPTQ_MARLIN_24_MAX_PARALLEL, GPTQ_MARLIN_24_MIN_THREAD_N,
-    GPTQ_MARLIN_24_SUPPORTED_GROUP_SIZES, GPTQ_MARLIN_24_SUPPORTED_QUANT_TYPES)
+    GPTQ_MARLIN_24_MAX_PARALLEL,
+    GPTQ_MARLIN_24_MIN_THREAD_N,
+    GPTQ_MARLIN_24_SUPPORTED_GROUP_SIZES,
+    GPTQ_MARLIN_24_SUPPORTED_QUANT_TYPES,
+)
 from vllm.model_executor.layers.quantization.utils.allspark_utils import (
-    ALLSPARK_AMPERE_M_CUBLAS_THRESHOLD, ALLSPARK_SUPPORTED_QUANT_TYPES)
+    ALLSPARK_AMPERE_M_CUBLAS_THRESHOLD,
+    ALLSPARK_SUPPORTED_QUANT_TYPES,
+)
 from vllm.model_executor.layers.quantization.utils.marlin_utils import (
-    GPTQ_MARLIN_MAX_PARALLEL, GPTQ_MARLIN_MIN_THREAD_N,
-    MARLIN_SUPPORTED_GROUP_SIZES, query_marlin_supported_quant_types)
+    GPTQ_MARLIN_MAX_PARALLEL,
+    GPTQ_MARLIN_MIN_THREAD_N,
+    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, marlin_quantize)
+    MarlinWorkspace,
+    awq_marlin_quantize,
+    marlin_quantize,
+)
 from vllm.model_executor.layers.quantization.utils.marlin_utils_test_24 import (
-    marlin_24_quantize)
+    marlin_24_quantize,
+)
 from vllm.model_executor.layers.quantization.utils.quant_utils import (
-    gptq_pack, gptq_quantize_weights, quantize_weights, sort_weights)
-from vllm.scalar_type import ScalarType
+    gptq_pack,
+    gptq_quantize_weights,
+    quantize_weights,
+    sort_weights,
+)
+from vllm.scalar_type import ScalarType, scalar_types
 from vllm.utils import FlexibleArgumentParser
 
 DEFAULT_MODELS = ["meta-llama/Llama-2-7b-hf/TP1"]
@@ -29,78 +53,156 @@ ACT_ORDER_OPTS = [False, True]
 K_FULL_OPTS = [False, True]
 
 
-def bench_run(results: list[benchmark.Measurement], model: str,
-              act_order: bool, is_k_full: bool, quant_type: ScalarType,
-              group_size: int, size_m: int, size_k: int, size_n: int):
+def bench_run(
+    results: list[benchmark.Measurement],
+    model: str,
+    act_order: bool,
+    is_k_full: bool,
+    quant_type: ScalarType,
+    group_size: int,
+    size_m: int,
+    size_k: int,
+    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))
-
+    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_24_supported = (
+        quant_type in GPTQ_MARLIN_24_SUPPORTED_QUANT_TYPES
+        and group_size in GPTQ_MARLIN_24_SUPPORTED_GROUP_SIZES
+    )
+    repack_supported = (
+        quant_type in GPTQ_MARLIN_24_SUPPORTED_QUANT_TYPES
+        and group_size in MARLIN_SUPPORTED_GROUP_SIZES
+    )
+    allspark_supported = (
+        quant_type in ALLSPARK_SUPPORTED_QUANT_TYPES
+        and group_size == -1
+        and not act_order
+        and is_k_full
+    )
+
+    def gen_marlin_params():
+        # Marlin quant
+        marlin_g_idx = marlin_sort_indices = marlin_zp = marlin_s2 = None
+        if quant_type == scalar_types.float4_e2m1f:
+            if group_size != 16 or act_order:
+                return
+            marlin_w_ref, marlin_q_w, marlin_s, marlin_s2 = rand_marlin_weight_fp4_like(
+                b.T, group_size
+            )
+        elif quant_type == scalar_types.float8_e4m3fn:
+            if group_size not in [-1, 128] or act_order:
+                return
+            marlin_w_ref, marlin_q_w, marlin_s = marlin_quant_fp8_torch(b.T, group_size)
+        elif group_size == 16:
+            return
+        elif has_zp:
+            marlin_w_ref, marlin_q_w, marlin_s, marlin_zp = awq_marlin_quantize(
+                b, quant_type, group_size
+            )
+        else:
+            marlin_w_ref, marlin_q_w, marlin_s, marlin_g_idx, marlin_sort_indices, _ = (
+                marlin_quantize(b, quant_type, group_size, act_order)
+            )
+        return (
+            marlin_w_ref,
+            marlin_q_w,
+            marlin_s,
+            marlin_s2,
+            marlin_zp,
+            marlin_g_idx,
+            marlin_sort_indices,
+        )
+
+    def gen_marlin_24_params():
+        marlin_24_w_ref = marlin_24_q_w_comp = marlin_24_meta = marlin_24_s = None
+        if marlin_24_supported:
+            (marlin_24_w_ref, marlin_24_q_w_comp, marlin_24_meta, marlin_24_s) = (
+                marlin_24_quantize(b, quant_type, group_size)
+            )
+        return (marlin_24_w_ref, marlin_24_q_w_comp, marlin_24_meta, marlin_24_s)
+
+    def gen_repack_params():
+        q_w_gptq = None
+        repack_sort_indices = None
+        if repack_supported:
+            (w_ref, q_w, s, g_idx, rand_perm) = gptq_quantize_weights(
+                b, quant_type, group_size, act_order
+            )
+            q_w_gptq = gptq_pack(q_w, quant_type.size_bits, size_k, size_n)
+
+            # For act_order, sort the "weights" and "g_idx"
+            # so that group ids are increasing
+            repack_sort_indices = torch.empty(0, dtype=torch.int, device=b.device)
+            if act_order:
+                (q_w, g_idx, repack_sort_indices) = sort_weights(q_w, g_idx)
+        return q_w_gptq, repack_sort_indices
+
+    def gen_allspark_params():
+        qw_reorder = s_reorder = zp_reorder = sm_count = sm_version = (
+            CUBLAS_M_THRESHOLD
+        ) = None
+        nonlocal allspark_supported
+        if allspark_supported:
+            properties = torch.cuda.get_device_properties(b.device.index)
+            sm_count = properties.multi_processor_count
+            sm_version = properties.major * 10 + properties.minor
+
+            supported_arch = sm_version >= 80 and sm_version < 90
+            allspark_supported = allspark_supported and supported_arch
+            if supported_arch:
+                w_ref, qw, s, zp = quantize_weights(b, quant_type, group_size, has_zp)
+                qw = qw.to(torch.uint8)
+
+                qw_reorder, s_reorder, zp_reorder = ops.allspark_repack_weight(
+                    qw, s, zp, has_zp
+                )
+                CUBLAS_M_THRESHOLD = ALLSPARK_AMPERE_M_CUBLAS_THRESHOLD
+        return (
+            qw_reorder,
+            s_reorder,
+            zp_reorder,
+            sm_count,
+            sm_version,
+            CUBLAS_M_THRESHOLD,
+        )
 
-    # Marlin quant
     (
         marlin_w_ref,
         marlin_q_w,
         marlin_s,
+        marlin_s2,
+        marlin_zp,
         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_zp = torch.empty(0, dtype=torch.int, device=b.device)
-
-    # GPTQ quant
-    (w_ref, q_w, s, g_idx,
-     rand_perm) = gptq_quantize_weights(b, quant_type, group_size, act_order)
-    q_w_gptq = gptq_pack(q_w, quant_type.size_bits, size_k, size_n)
-
-    # For act_order, sort the "weights" and "g_idx"
-    # so that group ids are increasing
-    repack_sort_indices = torch.empty(0, dtype=torch.int, device=b.device)
-    if act_order:
-        (q_w, g_idx, repack_sort_indices) = sort_weights(q_w, g_idx)
+    ) = 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)
-    marlin_zp = torch.zeros_like(marlin_s, dtype=torch.int)
-
-    # AllSpark W8A16 quant
-    as_supported_case = (quant_type in ALLSPARK_SUPPORTED_QUANT_TYPES
-                         and group_size == -1 and not act_order and is_k_full)
-    if as_supported_case:
-        properties = torch.cuda.get_device_properties(b.device.index)
-        sm_count = properties.multi_processor_count
-        sm_version = properties.major * 10 + properties.minor
-
-        supported_arch = (sm_version >= 80 and sm_version < 90)
-        as_supported_case = as_supported_case and supported_arch
-        if supported_arch:
-            has_zp = False
-            w_ref, qw, s, zp = quantize_weights(b, quant_type, group_size,
-                                                has_zp)
-            qw = qw.to(torch.uint8)
-
-            qw_reorder, s_reorder, zp_reorder = \
-                ops.allspark_repack_weight(
-                qw, s, zp, has_zp)
-            CUBLAS_M_THRESHOLD = ALLSPARK_AMPERE_M_CUBLAS_THRESHOLD
+    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
@@ -110,15 +212,14 @@ def bench_run(results: list[benchmark.Measurement], model: str,
         "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
@@ -131,13 +232,12 @@ def bench_run(results: list[benchmark.Measurement], model: str,
         "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,
@@ -148,7 +248,7 @@ def bench_run(results: list[benchmark.Measurement], model: str,
     min_run_time = 1
 
     # Warmup pytorch
-    for i in range(5):
+    for _ in range(5):
         torch.matmul(a, marlin_w_ref)
 
     results.append(
@@ -158,67 +258,67 @@ def bench_run(results: list[benchmark.Measurement], model: str,
             label=label,
             sub_label=sub_label,
             description="pytorch_gemm",
-        ).blocked_autorange(min_run_time=min_run_time))
+        ).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, 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",
-        ).blocked_autorange(min_run_time=min_run_time))
+            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,
             description="gptq_marlin_gemm_fp32",
-        ).blocked_autorange(min_run_time=min_run_time))
+        ).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
+                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
                 globals=globals,
                 label=label,
                 sub_label=sub_label,
                 description="gptq_marlin_24_gemm",
-            ).blocked_autorange(min_run_time=min_run_time))
+            ).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 as_supported_case:
+    if repack_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
+                stmt="q_res = gptq_marlin_repack(q_w_gptq, repack_sort_indices, size_k, size_n, quant_type.size_bits)",  # noqa: E501
+                globals=globals,
+                label=label,
+                sub_label=sub_label,
+                description="gptq_marlin_repack",
+            ).blocked_autorange(min_run_time=min_run_time)
+        )
+
+    if allspark_supported:
+        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
                 globals=globals,
                 label=label,
                 sub_label=sub_label,
                 description="allspark_w8a16_gemm_fp32",
-            ).blocked_autorange(min_run_time=min_run_time))
+            ).blocked_autorange(min_run_time=min_run_time)
+        )
 
 
 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:
@@ -233,37 +333,53 @@ def main(args):
                 continue
 
             for act_order in ACT_ORDER_OPTS:
-                if len(args.limit_act_order
-                       ) > 0 and act_order not in args.limit_act_order:
+                if (
+                    len(args.limit_act_order) > 0
+                    and act_order not in args.limit_act_order
+                ):
                     continue
 
                 for is_k_full in K_FULL_OPTS:
-                    if len(args.limit_k_full
-                           ) > 0 and is_k_full not in args.limit_k_full:
+                    if (
+                        len(args.limit_k_full) > 0
+                        and is_k_full not in args.limit_k_full
+                    ):
                         continue
 
-                    for quant_type in query_marlin_supported_quant_types(
-                            False):
-                        if len(args.limit_num_bits) > 0 and \
-                            quant_type.size_bits not in args.limit_num_bits:
+                    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:
-                            if len(
-                                    args.limit_group_size
-                            ) > 0 and group_size not in args.limit_group_size:
+                        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
+                            ):
                                 continue
 
                             # For act_order, the group_size must be less than
                             # size_k
-                            if act_order and (group_size == size_k
-                                              or group_size == -1):
+                            if act_order and (group_size == size_k or group_size == -1):
                                 continue
 
                             for size_m in args.batch_sizes:
-                                bench_run(results, model, act_order, is_k_full,
-                                          quant_type, group_size, size_m,
-                                          size_k, size_n)
+                                bench_run(
+                                    results,
+                                    model,
+                                    act_order,
+                                    is_k_full,
+                                    quant_type,
+                                    group_size,
+                                    size_m,
+                                    size_k,
+                                    size_n,
+                                )
 
     compare = benchmark.Compare(results)
     compare.print()
@@ -274,7 +390,8 @@ def main(args):
 #
 if __name__ == "__main__":
     parser = FlexibleArgumentParser(
-        description="Benchmark Marlin across specified models/shapes/batches")
+        description="Benchmark Marlin across specified models/shapes/batches"
+    )
     parser.add_argument(
         "--models",
         nargs="+",
@@ -282,10 +399,9 @@ if __name__ == "__main__":
         default=DEFAULT_MODELS,
         choices=WEIGHT_SHAPES.keys(),
     )
-    parser.add_argument("--batch-sizes",
-                        nargs="+",
-                        type=int,
-                        default=DEFAULT_BATCH_SIZES)
+    parser.add_argument(
+        "--batch-sizes", nargs="+", type=int, default=DEFAULT_BATCH_SIZES
+    )
     parser.add_argument("--limit-k", nargs="+", type=int, default=[])
     parser.add_argument("--limit-n", nargs="+", type=int, default=[])
     parser.add_argument("--limit-group-size", nargs="+", type=int, default=[])

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
@@ -10,12 +11,12 @@ from typing import Any, TypedDict
 
 import ray
 import torch
-import triton
 from ray.experimental.tqdm_ray import tqdm
-from transformers import AutoConfig
 
 from vllm.model_executor.layers.fused_moe.fused_moe import *
 from vllm.platforms import current_platform
+from vllm.transformers_utils.config import get_config
+from vllm.triton_utils import triton
 from vllm.utils import FlexibleArgumentParser
 
 FP8_DTYPE = current_platform.fp8_dtype()
@@ -30,57 +31,64 @@ class BenchmarkConfig(TypedDict):
     num_stages: int
 
 
-def benchmark_config(config: BenchmarkConfig,
-                     num_tokens: int,
-                     num_experts: int,
-                     shard_intermediate_size: int,
-                     hidden_size: int,
-                     topk: int,
-                     dtype: torch.dtype,
-                     use_fp8_w8a8: bool,
-                     use_int8_w8a16: bool,
-                     num_iters: int = 100,
-                     block_quant_shape: List[int] = None,
-                     use_deep_gemm: bool = False) -> float:
+def benchmark_config(
+    config: BenchmarkConfig,
+    num_tokens: int,
+    num_experts: int,
+    shard_intermediate_size: int,
+    hidden_size: int,
+    topk: int,
+    dtype: torch.dtype,
+    use_fp8_w8a8: bool,
+    use_int8_w8a16: bool,
+    num_iters: int = 100,
+    block_quant_shape: list[int] = None,
+    use_deep_gemm: bool = False,
+) -> float:
     init_dtype = torch.float16 if use_fp8_w8a8 else dtype
     x = torch.randn(num_tokens, hidden_size, dtype=dtype)
     if use_int8_w8a16:
-        w1 = torch.randint(-127,
-                           127, (
-                               num_experts,
-                               shard_intermediate_size,
-                               hidden_size,
-                           ),
-                           dtype=torch.int8)
-        w2 = torch.randint(-127,
-                           127, (
-                               num_experts,
-                               hidden_size,
-                               shard_intermediate_size // 2,
-                           ),
-                           dtype=torch.int8)
+        w1 = torch.randint(
+            -127,
+            127,
+            (
+                num_experts,
+                shard_intermediate_size,
+                hidden_size,
+            ),
+            dtype=torch.int8,
+        )
+        w2 = torch.randint(
+            -127,
+            127,
+            (
+                num_experts,
+                hidden_size,
+                shard_intermediate_size // 2,
+            ),
+            dtype=torch.int8,
+        )
     else:
-        w1 = torch.randn(num_experts,
-                         shard_intermediate_size,
-                         hidden_size,
-                         dtype=init_dtype)
-        w2 = torch.randn(num_experts,
-                         hidden_size,
-                         shard_intermediate_size // 2,
-                         dtype=init_dtype)
-    gating_output = torch.randn(num_iters,
-                                num_tokens,
-                                num_experts,
-                                dtype=torch.float32)
+        w1 = torch.randn(
+            num_experts, shard_intermediate_size, hidden_size, dtype=init_dtype
+        )
+        w2 = torch.randn(
+            num_experts, hidden_size, shard_intermediate_size // 2, dtype=init_dtype
+        )
+    gating_output = torch.randn(num_iters, num_tokens, num_experts, dtype=torch.float32)
 
     w1_scale = None
     w2_scale = None
     a1_scale = None
     a2_scale = None
     if use_int8_w8a16:
-        w1_scale = torch.randn((num_experts, 2 * shard_intermediate_size),
-                               dtype=torch.float32)
+        w1_scale = torch.randn(
+            (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]
@@ -92,10 +100,14 @@ def benchmark_config(config: BenchmarkConfig,
             n_tiles_w2 = (K + block_n - 1) // block_n
             k_tiles_w1 = (K + block_k - 1) // block_k
             k_tiles_w2 = (N + block_k - 1) // block_k
-            w1_scale = torch.rand((E, n_tiles_w1, k_tiles_w1),
-                                  dtype=torch.float32) * factor_for_scale
-            w2_scale = torch.rand((E, n_tiles_w2, k_tiles_w2),
-                                  dtype=torch.float32) * factor_for_scale
+            w1_scale = (
+                torch.rand((E, n_tiles_w1, k_tiles_w1), dtype=torch.float32)
+                * factor_for_scale
+            )
+            w2_scale = (
+                torch.rand((E, n_tiles_w2, k_tiles_w2), dtype=torch.float32)
+                * factor_for_scale
+            )
         else:
             w1_scale = torch.randn(num_experts, dtype=torch.float32)
             w2_scale = torch.randn(num_experts, dtype=torch.float32)
@@ -113,10 +125,12 @@ def benchmark_config(config: BenchmarkConfig,
 
     def run():
         from vllm.model_executor.layers.fused_moe import override_config
+
         with override_config(config):
             if use_deep_gemm:
-                topk_weights, topk_ids = fused_topk(x, input_gating, topk,
-                                                    False)
+                topk_weights, topk_ids, token_expert_indices = fused_topk(
+                    x, input_gating, topk, False
+                )
                 return fused_experts(
                     x,
                     w1,
@@ -212,8 +226,7 @@ def get_rocm_tuning_space(use_fp16):
     return param_ranges
 
 
-def get_configs_compute_bound(use_fp16,
-                              block_quant_shape) -> list[dict[str, int]]:
+def get_configs_compute_bound(use_fp16, block_quant_shape) -> list[dict[str, int]]:
     configs: list[BenchmarkConfig] = []
 
     if current_platform.is_rocm():
@@ -249,20 +262,25 @@ def get_configs_compute_bound(use_fp16,
     if block_quant_shape is not None and not use_fp16:
         block_n, block_k = block_quant_shape[0], block_quant_shape[1]
         for config in configs[:]:
-            if config["BLOCK_SIZE_K"] % block_k != 0 or config[
-                    "BLOCK_SIZE_N"] % block_n != 0:
+            if (
+                config["BLOCK_SIZE_K"] % block_k != 0
+                or config["BLOCK_SIZE_N"] % block_n != 0
+            ):
                 configs.remove(config)
     return configs
 
 
-def prune_rocm_search_space(num_tokens, shard_intermediate_size, hidden_size,
-                            search_space, is_fp16, topk):
+def prune_rocm_search_space(
+    num_tokens, shard_intermediate_size, hidden_size, search_space, is_fp16, topk
+):
     N1, K1 = shard_intermediate_size, hidden_size
     N2, K2 = hidden_size, shard_intermediate_size // 2
-    pruned_space_1 = prune_rocm_configs(num_tokens * topk, N1, K1,
-                                        search_space, is_fp16)
-    pruned_space_2 = prune_rocm_configs(num_tokens * topk, N2, K2,
-                                        search_space, is_fp16)
+    pruned_space_1 = prune_rocm_configs(
+        num_tokens * topk, N1, K1, search_space, is_fp16
+    )
+    pruned_space_2 = prune_rocm_configs(
+        num_tokens * topk, N2, K2, search_space, is_fp16
+    )
     search_space = merge_unique_dicts(pruned_space_1, pruned_space_2)
     return search_space
 
@@ -300,14 +318,14 @@ def prune_rocm_configs(M, N, K, configs, is_fp16=True):
         SPLIT_K = config.get("SPLIT_K", 1)
         GROUP_M = config.get("GROUP_SIZE_M")
         if is_fp16:
-            if (matrix_instr_nonkdim > BLOCK_SIZE_M
-                    or matrix_instr_nonkdim > BLOCK_SIZE_N):
+            if (
+                matrix_instr_nonkdim > BLOCK_SIZE_M
+                or matrix_instr_nonkdim > BLOCK_SIZE_N
+            ):
                 continue
-            if (matrix_instr_nonkdim >= M
-                    and matrix_instr_nonkdim != BLOCK_SIZE_M):
+            if matrix_instr_nonkdim >= M and matrix_instr_nonkdim != BLOCK_SIZE_M:
                 continue
-            if (matrix_instr_nonkdim >= N
-                    and matrix_instr_nonkdim != BLOCK_SIZE_N):
+            if matrix_instr_nonkdim >= N and matrix_instr_nonkdim != BLOCK_SIZE_N:
                 continue
         # Skip BLOCK_SIZE that is too large compare to M/N
         # unless BLOCK_SIZE is already small enough
@@ -328,8 +346,10 @@ def prune_rocm_configs(M, N, K, configs, is_fp16=True):
             continue
         # out of shared memory resource
         # TODO (zhanglx): This does not consider the LDS usage in the epilogue
-        LDS = (BLOCK_SIZE_K * BLOCK_SIZE_M * elemBytes_a +
-               BLOCK_SIZE_K * BLOCK_SIZE_N * elemBytes_b)
+        LDS = (
+            BLOCK_SIZE_K * BLOCK_SIZE_M * elemBytes_a
+            + BLOCK_SIZE_K * BLOCK_SIZE_N * elemBytes_b
+        )
         if LDS > 65536:
             continue
         # Skip small block sizes and num_warps for large gemm
@@ -363,7 +383,6 @@ def merge_unique_dicts(list1, list2):
 
 @ray.remote(num_gpus=1)
 class BenchmarkWorker:
-
     def __init__(self, seed: int) -> None:
         torch.set_default_device("cuda")
         current_platform.seed_everything(seed)
@@ -383,40 +402,44 @@ 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)
-        dtype_str = get_config_dtype_str(dtype,
-                                         use_int8_w8a16=use_int8_w8a16,
-                                         use_fp8_w8a8=use_fp8_w8a8)
+        dtype_str = get_config_dtype_str(
+            dtype, use_int8_w8a16=use_int8_w8a16, use_fp8_w8a8=use_fp8_w8a8
+        )
         # NOTE(woosuk): The current naming convention uses w2.shape[2], which
         # is the intermediate size after silu_and_mul.
-        op_config = get_moe_configs(num_experts, shard_intermediate_size // 2,
-                                    dtype_str)
+        op_config = get_moe_configs(
+            num_experts, shard_intermediate_size // 2, dtype_str
+        )
         if op_config is None:
-            config = get_default_config(num_tokens,
-                                        num_experts,
-                                        shard_intermediate_size,
-                                        hidden_size,
-                                        topk,
-                                        dtype_str,
-                                        is_marlin=False)
+            config = get_default_config(
+                num_tokens,
+                num_experts,
+                shard_intermediate_size,
+                hidden_size,
+                topk,
+                dtype_str,
+                is_marlin=False,
+            )
         else:
-            config = op_config[min(op_config.keys(),
-                                   key=lambda x: abs(x - num_tokens))]
-        kernel_time = benchmark_config(config,
-                                       num_tokens,
-                                       num_experts,
-                                       shard_intermediate_size,
-                                       hidden_size,
-                                       topk,
-                                       dtype,
-                                       use_fp8_w8a8,
-                                       use_int8_w8a16,
-                                       num_iters=100,
-                                       block_quant_shape=block_quant_shape,
-                                       use_deep_gemm=use_deep_gemm)
+            config = op_config[min(op_config.keys(), key=lambda x: abs(x - num_tokens))]
+        kernel_time = benchmark_config(
+            config,
+            num_tokens,
+            num_experts,
+            shard_intermediate_size,
+            hidden_size,
+            topk,
+            dtype,
+            use_fp8_w8a8,
+            use_int8_w8a16,
+            num_iters=100,
+            block_quant_shape=block_quant_shape,
+            use_deep_gemm=use_deep_gemm,
+        )
         return config, kernel_time
 
     def tune(
@@ -437,13 +460,22 @@ class BenchmarkWorker:
         best_time = float("inf")
         if current_platform.is_rocm():
             is_fp16 = not (use_fp8_w8a8 or use_int8_w8a16)
-            search_space = prune_rocm_search_space(num_tokens,
-                                                   shard_intermediate_size,
-                                                   hidden_size, search_space,
-                                                   is_fp16, topk)
+            search_space = prune_rocm_search_space(
+                num_tokens,
+                shard_intermediate_size,
+                hidden_size,
+                search_space,
+                is_fp16,
+                topk,
+            )
 
-        with torch.cuda.device(self.device_id) if current_platform.is_rocm(
-        ) else nullcontext():
+        need_device_guard = False
+        if current_platform.is_rocm():
+            visible_device = os.environ.get("ROCR_VISIBLE_DEVICES", None)
+            if visible_device != f"{self.device_id}":
+                need_device_guard = True
+
+        with torch.cuda.device(self.device_id) if need_device_guard else nullcontext():
             for config in tqdm(search_space):
                 try:
                     kernel_time = benchmark_config(
@@ -458,7 +490,8 @@ class BenchmarkWorker:
                         use_int8_w8a16,
                         num_iters=20,
                         block_quant_shape=block_quant_shape,
-                        use_deep_gemm=use_deep_gemm)
+                        use_deep_gemm=use_deep_gemm,
+                    )
                 except triton.runtime.autotuner.OutOfResources:
                     # Some configurations may be invalid and fail to compile.
                     continue
@@ -474,42 +507,44 @@ class BenchmarkWorker:
 
 def sort_config(config: BenchmarkConfig) -> BenchmarkConfig:
     return {
-        "BLOCK_SIZE_M":
-        config["BLOCK_SIZE_M"],
-        "BLOCK_SIZE_N":
-        config["BLOCK_SIZE_N"],
-        "BLOCK_SIZE_K":
-        config["BLOCK_SIZE_K"],
-        "GROUP_SIZE_M":
-        config["GROUP_SIZE_M"],
-        "num_warps":
-        config["num_warps"],
-        "num_stages":
-        config["num_stages"],
-        **({
-            "waves_per_eu": config["waves_per_eu"]
-        } if "waves_per_eu" in config else {}),
-        **({
-            "matrix_instr_nonkdim": config["matrix_instr_nonkdim"]
-        } if "matrix_instr_nonkdim" in config else {}),
-        **({
-            "kpack": config["kpack"]
-        } if "kpack" in config else {}),
+        "BLOCK_SIZE_M": config["BLOCK_SIZE_M"],
+        "BLOCK_SIZE_N": config["BLOCK_SIZE_N"],
+        "BLOCK_SIZE_K": config["BLOCK_SIZE_K"],
+        "GROUP_SIZE_M": config["GROUP_SIZE_M"],
+        "num_warps": config["num_warps"],
+        "num_stages": config["num_stages"],
+        **(
+            {"waves_per_eu": config["waves_per_eu"]} if "waves_per_eu" in config else {}
+        ),
+        **(
+            {"matrix_instr_nonkdim": config["matrix_instr_nonkdim"]}
+            if "matrix_instr_nonkdim" in config
+            else {}
+        ),
+        **({"kpack": config["kpack"]} if "kpack" in config else {}),
     }
 
 
-def save_configs(configs: dict[int, BenchmarkConfig], num_experts: int,
-                 shard_intermediate_size: int, hidden_size: int, topk: int,
-                 dtype: torch.dtype, use_fp8_w8a8: bool, use_int8_w8a16: bool,
-                 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)
+def save_configs(
+    configs: dict[int, BenchmarkConfig],
+    num_experts: int,
+    shard_intermediate_size: int,
+    hidden_size: int,
+    topk: int,
+    dtype: torch.dtype,
+    use_fp8_w8a8: bool,
+    use_int8_w8a16: bool,
+    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
+    )
 
     # NOTE(woosuk): The current naming convention uses w2.shape[2], which
     # is the intermediate size after silu_and_mul.
-    filename = get_config_file_name(num_experts, shard_intermediate_size // 2,
-                                    dtype_str, block_quant_shape)
+    filename = get_config_file_name(
+        num_experts, shard_intermediate_size // 2, dtype_str, block_quant_shape
+    )
 
     print(f"Writing best config to {filename}...")
     with open(filename, "w") as f:
@@ -518,18 +553,19 @@ def save_configs(configs: dict[int, BenchmarkConfig], num_experts: int,
 
 
 def get_weight_block_size_safety(config, default_value=None):
-
-    quantization_config = getattr(config, 'quantization_config', {})
+    quantization_config = getattr(config, "quantization_config", {})
     if isinstance(quantization_config, dict):
-        return quantization_config.get('weight_block_size', default_value)
+        return quantization_config.get("weight_block_size", default_value)
     return default_value
 
 
 def main(args: argparse.Namespace):
     print(args)
-    block_quant_shape = None
-    config = AutoConfig.from_pretrained(
-        args.model, trust_remote_code=args.trust_remote_code)
+
+    config = get_config(model=args.model, trust_remote_code=args.trust_remote_code)
+    if args.model_prefix:
+        config = getattr(config, args.model_prefix)
+
     if config.architectures[0] == "DbrxForCausalLM":
         E = config.ffn_config.moe_num_experts
         topk = config.ffn_config.moe_top_k
@@ -540,18 +576,25 @@ def main(args: argparse.Namespace):
         topk = config.num_experts_per_tok
         intermediate_size = config.intermediate_size
         shard_intermediate_size = 2 * intermediate_size // args.tp_size
-    elif (config.architectures[0] == "DeepseekV3ForCausalLM"
-          or config.architectures[0] == "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
         shard_intermediate_size = 2 * intermediate_size // args.tp_size
-        block_quant_shape = get_weight_block_size_safety(config)
-    elif config.architectures[0] == "Qwen2MoeForCausalLM":
+    elif config.architectures[0] in ("Qwen2MoeForCausalLM", "Qwen3MoeForCausalLM"):
         E = config.num_experts
         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()
@@ -565,17 +608,44 @@ def main(args: argparse.Namespace):
     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)
 
     if args.batch_size is None:
         batch_sizes = [
-            1, 2, 4, 8, 16, 24, 32, 48, 64, 96, 128, 256, 512, 1024, 1536,
-            2048, 3072, 4096
+            1,
+            2,
+            4,
+            8,
+            16,
+            24,
+            32,
+            48,
+            64,
+            96,
+            128,
+            256,
+            512,
+            1024,
+            1536,
+            2048,
+            3072,
+            4096,
         ]
     else:
-        batch_sizes = [args.batch_size]
+        batch_sizes = args.batch_size
 
     use_deep_gemm = bool(args.use_deep_gemm)
 
+    if current_platform.is_rocm() and "HIP_VISIBLE_DEVICES" in os.environ:
+        # Ray will set ROCR_VISIBLE_DEVICES for device visibility
+        logger.warning(
+            "Ray uses ROCR_VISIBLE_DEVICES to control device accessibility."
+            "Replacing HIP_VISIBLE_DEVICES with ROCR_VISIBLE_DEVICES."
+        )
+        val = os.environ["HIP_VISIBLE_DEVICES"]
+        os.environ["ROCR_VISIBLE_DEVICES"] = val
+        del os.environ["HIP_VISIBLE_DEVICES"]
+
     ray.init()
     num_gpus = int(ray.available_resources()["GPU"])
     workers = [BenchmarkWorker.remote(args.seed) for _ in range(num_gpus)]
@@ -598,25 +668,59 @@ def main(args: argparse.Namespace):
 
         start = time.time()
         configs = _distribute(
-            "tune", [(batch_size, E, shard_intermediate_size, hidden_size,
-                      topk, dtype, use_fp8_w8a8, use_int8_w8a16, search_space,
-                      block_quant_shape, use_deep_gemm)
-                     for batch_size in batch_sizes])
+            "tune",
+            [
+                (
+                    batch_size,
+                    E,
+                    shard_intermediate_size,
+                    hidden_size,
+                    topk,
+                    dtype,
+                    use_fp8_w8a8,
+                    use_int8_w8a16,
+                    search_space,
+                    block_quant_shape,
+                    use_deep_gemm,
+                )
+                for batch_size in batch_sizes
+            ],
+        )
         best_configs = {
-            M: sort_config(config)
-            for M, config in zip(batch_sizes, configs)
+            M: sort_config(config) for M, config in zip(batch_sizes, configs)
         }
-        save_configs(best_configs, E, shard_intermediate_size, hidden_size,
-                     topk, dtype, use_fp8_w8a8, use_int8_w8a16,
-                     block_quant_shape)
+        save_configs(
+            best_configs,
+            E,
+            shard_intermediate_size,
+            hidden_size,
+            topk,
+            dtype,
+            use_fp8_w8a8,
+            use_int8_w8a16,
+            block_quant_shape,
+        )
         end = time.time()
         print(f"Tuning took {end - start:.2f} seconds")
     else:
         outputs = _distribute(
             "benchmark",
-            [(batch_size, E, shard_intermediate_size, hidden_size, topk, dtype,
-              use_fp8_w8a8, use_int8_w8a16, block_quant_shape, use_deep_gemm)
-             for batch_size in batch_sizes])
+            [
+                (
+                    batch_size,
+                    E,
+                    shard_intermediate_size,
+                    hidden_size,
+                    topk,
+                    dtype,
+                    use_fp8_w8a8,
+                    use_int8_w8a16,
+                    block_quant_shape,
+                    use_deep_gemm,
+                )
+                for batch_size in batch_sizes
+            ],
+        )
 
         for batch_size, (config, kernel_time) in zip(batch_sizes, outputs):
             print(f"Batch size: {batch_size}, config: {config}")
@@ -625,23 +729,21 @@ def main(args: argparse.Namespace):
 
 if __name__ == "__main__":
     parser = FlexibleArgumentParser()
-    parser.add_argument("--model",
-                        type=str,
-                        default="mistralai/Mixtral-8x7B-Instruct-v0.1")
-    parser.add_argument("--tp-size",
-                        "-tp",
-                        "--tensor-parallel-size",
-                        type=int,
-                        default=2)
-    parser.add_argument("--dtype",
-                        type=str,
-                        choices=["auto", "fp8_w8a8", "int8_w8a16"],
-                        default="auto")
+    parser.add_argument(
+        "--model", type=str, default="mistralai/Mixtral-8x7B-Instruct-v0.1"
+    )
+    parser.add_argument(
+        "--tp-size", "-tp", "--tensor-parallel-size", type=int, default=2
+    )
+    parser.add_argument(
+        "--dtype", type=str, choices=["auto", "fp8_w8a8", "int8_w8a16"], default="auto"
+    )
     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)
     args = parser.parse_args()
 
     main(args)

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

@@ -0,0 +1,418 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+import argparse
+from typing import Any, TypedDict
+
+import ray
+import torch
+from transformers import AutoConfig
+
+from vllm.model_executor.layers.fused_moe.deep_gemm_moe import (
+    _moe_permute,
+    _moe_unpermute_and_reduce,
+)
+from vllm.model_executor.layers.fused_moe.fused_moe import *
+from vllm.model_executor.layers.fused_moe.moe_permute_unpermute import *
+from vllm.model_executor.layers.fused_moe.utils import _fp8_quantize
+from vllm.platforms import current_platform
+from vllm.utils import FlexibleArgumentParser
+
+FP8_DTYPE = current_platform.fp8_dtype()
+
+
+class BenchmarkConfig(TypedDict):
+    BLOCK_SIZE_M: int
+    BLOCK_SIZE_N: int
+    BLOCK_SIZE_K: int
+    GROUP_SIZE_M: int
+    num_warps: int
+    num_stages: int
+
+
+def benchmark_permute(
+    num_tokens: int,
+    num_experts: int,
+    hidden_size: int,
+    topk: int,
+    dtype: torch.dtype,
+    use_fp8_w8a8: bool,
+    use_int8_w8a16: bool,
+    num_iters: int = 100,
+    use_customized_permute: bool = False,
+) -> float:
+    # init_dtype = torch.float16 if use_fp8_w8a8 else dtype
+    hidden_states = torch.randn(num_tokens, hidden_size, dtype=dtype)
+    # output_hidden_states = torch.empty_like(hidden_states)
+    if use_fp8_w8a8:
+        align_block_size = 128  # deepgemm needs 128 m aligned block
+        qhidden_states, scale = _fp8_quantize(hidden_states, None, None)
+    else:
+        align_block_size = None
+        qhidden_states = hidden_states
+
+    gating_output = torch.randn(num_iters, num_tokens, num_experts, dtype=torch.float32)
+
+    input_gating = torch.randn(num_tokens, num_experts, dtype=torch.float32)
+    topk_weights, topk_ids, token_expert_indices = fused_topk(
+        qhidden_states, input_gating, topk, False
+    )
+
+    def prepare(i: int):
+        input_gating.copy_(gating_output[i])
+
+    def run():
+        if use_customized_permute:
+            (permuted_hidden_states, first_token_off, inv_perm_idx, m_indices) = (
+                moe_permute(
+                    qhidden_states,
+                    topk_weights=topk_weights,
+                    topk_ids=topk_ids,
+                    token_expert_indices=token_expert_indices,
+                    topk=topk,
+                    n_expert=num_experts,
+                    n_local_expert=num_experts,
+                    expert_map=None,
+                    align_block_size=align_block_size,
+                )
+            )
+        else:
+            (
+                permuted_hidden_states,
+                a1q_scale,
+                sorted_token_ids,
+                expert_ids,
+                inv_perm,
+            ) = _moe_permute(
+                qhidden_states, None, topk_ids, num_experts, None, align_block_size
+            )
+
+    # JIT compilation & warmup
+    run()
+    torch.cuda.synchronize()
+
+    # Capture 10 invocations with CUDA graph
+    graph = torch.cuda.CUDAGraph()
+    with torch.cuda.graph(graph):
+        for _ in range(10):
+            run()
+    torch.cuda.synchronize()
+
+    # Warmup
+    for _ in range(5):
+        graph.replay()
+    torch.cuda.synchronize()
+
+    start_event = torch.cuda.Event(enable_timing=True)
+    end_event = torch.cuda.Event(enable_timing=True)
+
+    latencies: list[float] = []
+    for i in range(num_iters):
+        prepare(i)
+        torch.cuda.synchronize()
+
+        start_event.record()
+        graph.replay()
+        end_event.record()
+        end_event.synchronize()
+        latencies.append(start_event.elapsed_time(end_event))
+    avg = sum(latencies) / (num_iters * 10) * 1000  # us
+    graph.reset()
+    return avg
+
+
+def benchmark_unpermute(
+    num_tokens: int,
+    num_experts: int,
+    hidden_size: int,
+    topk: int,
+    dtype: torch.dtype,
+    use_fp8_w8a8: bool,
+    use_int8_w8a16: bool,
+    num_iters: int = 100,
+    use_customized_permute: bool = False,
+) -> float:
+    # init_dtype = torch.float16 if use_fp8_w8a8 else dtype
+    hidden_states = torch.randn(num_tokens, hidden_size, dtype=dtype)
+    output_hidden_states = torch.empty_like(hidden_states)
+    if use_fp8_w8a8:
+        align_block_size = 128  # deepgemm needs 128 m aligned block
+        qhidden_states, scale = _fp8_quantize(hidden_states, None, None)
+    else:
+        align_block_size = None
+        qhidden_states = hidden_states
+
+    input_gating = torch.randn(num_tokens, num_experts, dtype=torch.float32)
+
+    topk_weights, topk_ids, token_expert_indices = fused_topk(
+        qhidden_states, input_gating, topk, False
+    )
+
+    def prepare():
+        if use_customized_permute:
+            (permuted_hidden_states, first_token_off, inv_perm_idx, m_indices) = (
+                moe_permute(
+                    qhidden_states,
+                    topk_weights=topk_weights,
+                    topk_ids=topk_ids,
+                    token_expert_indices=token_expert_indices,
+                    topk=topk,
+                    n_expert=num_experts,
+                    n_local_expert=num_experts,
+                    expert_map=None,
+                    align_block_size=align_block_size,
+                )
+            )
+            # convert to fp16/bf16 as gemm output
+            return (
+                permuted_hidden_states.to(dtype),
+                first_token_off,
+                inv_perm_idx,
+                m_indices,
+            )
+        else:
+            (
+                permuted_qhidden_states,
+                a1q_scale,
+                sorted_token_ids,
+                expert_ids,
+                inv_perm,
+            ) = _moe_permute(
+                qhidden_states, None, topk_ids, num_experts, None, align_block_size
+            )
+            # convert to fp16/bf16 as gemm output
+            return (
+                permuted_qhidden_states.to(dtype),
+                a1q_scale,
+                sorted_token_ids,
+                expert_ids,
+                inv_perm,
+            )
+
+    def run(input: tuple):
+        if use_customized_permute:
+            (permuted_hidden_states, first_token_off, inv_perm_idx, m_indices) = input
+            moe_unpermute(
+                permuted_hidden_states,
+                topk_weights,
+                topk_ids,
+                inv_perm_idx,
+                first_token_off,
+                topk,
+                num_experts,
+                num_experts,
+            )
+        else:
+            (
+                permuted_hidden_states,
+                a1q_scale,
+                sorted_token_ids,
+                expert_ids,
+                inv_perm,
+            ) = input
+            _moe_unpermute_and_reduce(
+                output_hidden_states, permuted_hidden_states, inv_perm, topk_weights
+            )
+
+    # JIT compilation & warmup
+    input = prepare()
+    run(input)
+    torch.cuda.synchronize()
+
+    # Capture 10 invocations with CUDA graph
+    graph = torch.cuda.CUDAGraph()
+    with torch.cuda.graph(graph):
+        for _ in range(10):
+            run(input)
+    torch.cuda.synchronize()
+
+    # Warmup
+    for _ in range(5):
+        graph.replay()
+    torch.cuda.synchronize()
+
+    start_event = torch.cuda.Event(enable_timing=True)
+    end_event = torch.cuda.Event(enable_timing=True)
+
+    latencies: list[float] = []
+    for i in range(num_iters):
+        torch.cuda.synchronize()
+        start_event.record()
+        graph.replay()
+        end_event.record()
+        end_event.synchronize()
+        latencies.append(start_event.elapsed_time(end_event))
+    avg = sum(latencies) / (num_iters * 10) * 1000  # us
+    graph.reset()
+    return avg
+
+
+@ray.remote(num_gpus=1)
+class BenchmarkWorker:
+    def __init__(self, seed: int) -> None:
+        torch.set_default_device("cuda")
+        current_platform.seed_everything(seed)
+        self.seed = seed
+        # Get the device ID to allocate tensors and kernels
+        # on the respective GPU. This is required for Ray to work
+        # correctly with multi-GPU tuning on the ROCm platform.
+        self.device_id = int(ray.get_gpu_ids()[0])
+
+    def benchmark(
+        self,
+        num_tokens: int,
+        num_experts: int,
+        hidden_size: int,
+        topk: int,
+        dtype: torch.dtype,
+        use_fp8_w8a8: bool,
+        use_int8_w8a16: bool,
+        use_customized_permute: bool = False,
+    ) -> tuple[dict[str, int], float]:
+        current_platform.seed_everything(self.seed)
+
+        permute_time = benchmark_permute(
+            num_tokens,
+            num_experts,
+            hidden_size,
+            topk,
+            dtype,
+            use_fp8_w8a8,
+            use_int8_w8a16,
+            num_iters=100,
+            use_customized_permute=use_customized_permute,
+        )
+        unpermute_time = benchmark_unpermute(
+            num_tokens,
+            num_experts,
+            hidden_size,
+            topk,
+            dtype,
+            use_fp8_w8a8,
+            use_int8_w8a16,
+            num_iters=100,
+            use_customized_permute=use_customized_permute,
+        )
+        return permute_time, unpermute_time
+
+
+def get_weight_block_size_safety(config, default_value=None):
+    quantization_config = getattr(config, "quantization_config", {})
+    if isinstance(quantization_config, dict):
+        return quantization_config.get("weight_block_size", default_value)
+    return default_value
+
+
+def main(args: argparse.Namespace):
+    print(args)
+
+    config = AutoConfig.from_pretrained(
+        args.model, trust_remote_code=args.trust_remote_code
+    )
+    if config.architectures[0] == "DbrxForCausalLM":
+        E = config.ffn_config.moe_num_experts
+        topk = config.ffn_config.moe_top_k
+    elif config.architectures[0] == "JambaForCausalLM":
+        E = config.num_experts
+        topk = config.num_experts_per_tok
+    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
+    elif config.architectures[0] in ["Qwen2MoeForCausalLM", "Qwen3MoeForCausalLM"]:
+        E = config.num_experts
+        topk = config.num_experts_per_tok
+
+    else:
+        # Support for llama4
+        config = config.get_text_config()
+        # Default: Mixtral.
+        E = config.num_local_experts
+        topk = config.num_experts_per_tok
+
+    hidden_size = config.hidden_size
+    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"
+    use_customized_permute = args.use_customized_permute
+
+    if args.batch_size is None:
+        batch_sizes = [
+            1,
+            2,
+            4,
+            8,
+            16,
+            24,
+            32,
+            48,
+            64,
+            96,
+            128,
+            256,
+            512,
+            1024,
+            1536,
+            2048,
+            3072,
+            4096,
+        ]
+    else:
+        batch_sizes = [args.batch_size]
+
+    ray.init()
+    num_gpus = int(ray.available_resources()["GPU"])
+    workers = [BenchmarkWorker.remote(args.seed) for _ in range(num_gpus)]
+
+    def _distribute(method: str, inputs: list[Any]) -> list[Any]:
+        outputs = []
+        worker_idx = 0
+        for input_args in inputs:
+            worker = workers[worker_idx]
+            worker_method = getattr(worker, method)
+            output = worker_method.remote(*input_args)
+            outputs.append(output)
+            worker_idx = (worker_idx + 1) % num_gpus
+        return ray.get(outputs)
+
+    outputs = _distribute(
+        "benchmark",
+        [
+            (
+                batch_size,
+                E,
+                hidden_size,
+                topk,
+                dtype,
+                use_fp8_w8a8,
+                use_int8_w8a16,
+                use_customized_permute,
+            )
+            for batch_size in batch_sizes
+        ],
+    )
+
+    for batch_size, (permute, unpermute) in zip(batch_sizes, outputs):
+        print(f"Batch size: {batch_size}")
+        print(f"Permute time: {permute:.2f} us")
+        print(f"Unpermute time: {unpermute:.2f} us")
+
+
+if __name__ == "__main__":
+    parser = FlexibleArgumentParser()
+    parser.add_argument(
+        "--model", type=str, default="mistralai/Mixtral-8x7B-Instruct-v0.1"
+    )
+    parser.add_argument(
+        "--dtype", type=str, choices=["auto", "fp8_w8a8", "int8_w8a16"], default="auto"
+    )
+    parser.add_argument("--use-customized-permute", action="store_true")
+    parser.add_argument("--seed", type=int, default=0)
+    parser.add_argument("--batch-size", type=int, required=False)
+    parser.add_argument("--trust-remote-code", action="store_true")
+    args = parser.parse_args()
+
+    main(args)

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
@@ -9,8 +10,11 @@ import torch
 from vllm import _custom_ops as ops
 from vllm.logger import init_logger
 from vllm.platforms import current_platform
-from vllm.utils import (STR_DTYPE_TO_TORCH_DTYPE, FlexibleArgumentParser,
-                        create_kv_caches_with_random)
+from vllm.utils import (
+    STR_DTYPE_TO_TORCH_DTYPE,
+    FlexibleArgumentParser,
+    create_kv_caches_with_random,
+)
 
 logger = init_logger(__name__)
 
@@ -38,19 +42,15 @@ def main(
     current_platform.seed_everything(seed)
 
     scale = float(1.0 / (head_size**0.5))
-    query = torch.empty(num_seqs,
-                        num_query_heads,
-                        head_size,
-                        dtype=dtype,
-                        device=device)
+    query = torch.empty(
+        num_seqs, num_query_heads, head_size, dtype=dtype, device=device
+    )
     query.uniform_(-scale, scale)
 
     assert num_query_heads % num_kv_heads == 0
     alibi_slopes = None
     if use_alibi:
-        alibi_slopes = torch.randn(num_query_heads,
-                                   dtype=torch.float,
-                                   device=device)
+        alibi_slopes = torch.randn(num_query_heads, dtype=torch.float, device=device)
 
     seq_lens = [seq_len for _ in range(num_seqs)]
     max_seq_len = max(seq_lens)
@@ -61,24 +61,23 @@ def main(
     block_tables_lst: list[list[int]] = []
     for _ in range(num_seqs):
         block_table = [
-            random.randint(0, NUM_BLOCKS - 1)
-            for _ in range(max_num_blocks_per_seq)
+            random.randint(0, NUM_BLOCKS - 1) for _ in range(max_num_blocks_per_seq)
         ]
         block_tables_lst.append(block_table)
 
-    block_tables = torch.tensor(block_tables_lst,
-                                dtype=torch.int,
-                                device=device)
+    block_tables = torch.tensor(block_tables_lst, dtype=torch.int, device=device)
 
     # Create the KV cache.
-    key_caches, value_caches = create_kv_caches_with_random(NUM_BLOCKS,
-                                                            block_size,
-                                                            1,
-                                                            num_kv_heads,
-                                                            head_size,
-                                                            kv_cache_dtype,
-                                                            dtype,
-                                                            device=device)
+    key_caches, value_caches = create_kv_caches_with_random(
+        NUM_BLOCKS,
+        block_size,
+        1,
+        num_kv_heads,
+        head_size,
+        kv_cache_dtype,
+        dtype,
+        device=device,
+    )
     key_cache, value_cache = key_caches[0], value_caches[0]
 
     # Prepare for the paged attention kernel.
@@ -86,11 +85,11 @@ def main(
     if version == "v2":
         if current_platform.is_rocm():
             global PARTITION_SIZE
-            if not args.custom_paged_attn:
+            if not args.custom_paged_attn and not current_platform.is_navi():
                 PARTITION_SIZE = 1024
             else:
                 PARTITION_SIZE = PARTITION_SIZE_ROCM
-        num_partitions = ((max_seq_len + PARTITION_SIZE - 1) // PARTITION_SIZE)
+        num_partitions = (max_seq_len + PARTITION_SIZE - 1) // PARTITION_SIZE
         tmp_output = torch.empty(
             size=(num_seqs, num_query_heads, num_partitions, head_size),
             dtype=output.dtype,
@@ -110,9 +109,7 @@ def main(
         start_time = time.perf_counter()
 
         # Using default kv_scale
-        k_scale = v_scale = torch.tensor(1.0,
-                                         dtype=torch.float32,
-                                         device=device)
+        k_scale = v_scale = torch.tensor(1.0, dtype=torch.float32, device=device)
 
         for _ in range(num_iters):
             if version == "v1":
@@ -166,6 +163,7 @@ def main(
                         scale,
                         block_tables,
                         seq_lens,
+                        None,
                         block_size,
                         max_seq_len,
                         alibi_slopes,
@@ -195,30 +193,29 @@ def main(
     print(f"Kernel running time: {latency * 1000000:.3f} us")
 
 
-if __name__ == '__main__':
-    logger.warning("This script benchmarks the paged attention kernel. "
-                   "By default this is no longer used in vLLM inference.")
+if __name__ == "__main__":
+    logger.warning(
+        "This script benchmarks the paged attention kernel. "
+        "By default this is no longer used in vLLM inference."
+    )
 
-    parser = FlexibleArgumentParser(
-        description="Benchmark the paged attention kernel.")
-    parser.add_argument("--version",
-                        type=str,
-                        choices=["v1", "v2"],
-                        default="v2")
+    parser = FlexibleArgumentParser(description="Benchmark the paged attention kernel.")
+    parser.add_argument("--version", type=str, choices=["v1", "v2"], default="v2")
     parser.add_argument("--batch-size", type=int, default=8)
     parser.add_argument("--seq-len", type=int, default=4096)
     parser.add_argument("--num-query-heads", type=int, default=64)
     parser.add_argument("--num-kv-heads", type=int, default=8)
-    parser.add_argument("--head-size",
-                        type=int,
-                        choices=[64, 80, 96, 112, 120, 128, 192, 256],
-                        default=128)
+    parser.add_argument(
+        "--head-size",
+        type=int,
+        choices=[64, 80, 96, 112, 120, 128, 192, 256],
+        default=128,
+    )
     parser.add_argument("--block-size", type=int, choices=[16, 32], default=16)
     parser.add_argument("--use-alibi", action="store_true")
-    parser.add_argument("--dtype",
-                        type=str,
-                        choices=["half", "bfloat16", "float"],
-                        default="half")
+    parser.add_argument(
+        "--dtype", type=str, choices=["half", "bfloat16", "float"], default="half"
+    )
     parser.add_argument("--seed", type=int, default=0)
     parser.add_argument("--profile", action="store_true")
     parser.add_argument(
@@ -228,10 +225,11 @@ if __name__ == '__main__':
         default="auto",
         help="Data type for kv cache storage. If 'auto', will use model "
         "data type. CUDA 11.8+ supports fp8 (=fp8_e4m3) and fp8_e5m2. "
-        "ROCm (AMD GPU) supports fp8 (=fp8_e4m3)")
-    parser.add_argument("--custom-paged-attn",
-                        action="store_true",
-                        help="Use custom paged attention")
+        "ROCm (AMD GPU) supports fp8 (=fp8_e4m3)",
+    )
+    parser.add_argument(
+        "--custom-paged-attn", action="store_true", help="Use custom paged attention"
+    )
     args = parser.parse_args()
     print(args)
 

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
 
@@ -10,15 +11,17 @@ from vllm.utils import STR_DTYPE_TO_TORCH_DTYPE, FlexibleArgumentParser
 
 
 @torch.inference_mode()
-def main(num_tokens: int,
-         hidden_size: int,
-         static_scale: bool,
-         quant_dtype: torch.dtype,
-         dtype: torch.dtype,
-         seed: int = 0,
-         do_profile: bool = False,
-         num_warmup_iters: int = 5,
-         num_iters: int = 100) -> None:
+def main(
+    num_tokens: int,
+    hidden_size: int,
+    static_scale: bool,
+    quant_dtype: torch.dtype,
+    dtype: torch.dtype,
+    seed: int = 0,
+    do_profile: bool = False,
+    num_warmup_iters: int = 5,
+    num_iters: int = 100,
+) -> None:
     current_platform.seed_everything(seed)
     torch.set_default_device("cuda")
 
@@ -56,7 +59,7 @@ def main(num_tokens: int,
     print(f"Kernel running time: {latency * 1000000:.3f} us")
 
 
-if __name__ == '__main__':
+if __name__ == "__main__":
 
     def to_torch_dtype(dt):
         if dt == "int8":
@@ -66,37 +69,40 @@ if __name__ == '__main__':
         raise ValueError(f"Unsupported dtype: {dt}")
 
     parser = FlexibleArgumentParser(
-        description="Benchmark the quantization (fp8 or int8) kernel.")
+        description="Benchmark the quantization (fp8 or int8) kernel."
+    )
     parser.add_argument("--num-tokens", type=int, default=4096)
     parser.add_argument("--hidden-size", type=int, default=8192)
     parser.add_argument("--static-scale", action="store_true")
-    parser.add_argument("--quant-dtype",
-                        type=str,
-                        choices=["fp8", "int8"],
-                        default="int8")
-    parser.add_argument("--dtype",
-                        type=str,
-                        choices=["half", "bfloat16", "float"],
-                        default="half")
+    parser.add_argument(
+        "--quant-dtype", type=str, choices=["fp8", "int8"], default="int8"
+    )
+    parser.add_argument(
+        "--dtype", type=str, choices=["half", "bfloat16", "float"], default="half"
+    )
 
     parser.add_argument("--seed", type=int, default=0)
     parser.add_argument("--profile", action="store_true")
     parser.add_argument("--num-warmup-iters", type=int, default=5)
-    parser.add_argument("--num-iters",
-                        type=int,
-                        default=100,
-                        help="Number of benchmark iterations. "
-                        "If --profile is set, this number is ignored")
+    parser.add_argument(
+        "--num-iters",
+        type=int,
+        default=100,
+        help="Number of benchmark iterations. "
+        "If --profile is set, this number is ignored",
+    )
 
     args = parser.parse_args()
     print(args)
 
-    main(num_tokens=args.num_tokens,
-         hidden_size=args.hidden_size,
-         static_scale=args.static_scale,
-         quant_dtype=to_torch_dtype(args.quant_dtype),
-         dtype=STR_DTYPE_TO_TORCH_DTYPE[args.dtype],
-         seed=args.seed,
-         do_profile=args.profile,
-         num_warmup_iters=args.num_warmup_iters,
-         num_iters=args.num_iters)
+    main(
+        num_tokens=args.num_tokens,
+        hidden_size=args.hidden_size,
+        static_scale=args.static_scale,
+        quant_dtype=to_torch_dtype(args.quant_dtype),
+        dtype=STR_DTYPE_TO_TORCH_DTYPE[args.dtype],
+        seed=args.seed,
+        do_profile=args.profile,
+        num_warmup_iters=args.num_warmup_iters,
+        num_iters=args.num_iters,
+    )