Simplify max model length auto selection

.buildkite/check-wheel-size.py

@@ -5,11 +5,11 @@ import os
 import sys
 import zipfile
 
-# Read the VLLM_MAX_SIZE_MB environment variable, defaulting to 500 MiB
-# Note that we have 800 MiB quota, please use it wisely.
-# See https://github.com/pypi/support/issues/6326 .
+# Read the VLLM_MAX_SIZE_MB environment variable, defaulting to 400 MiB
+# 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", 500))
+VLLM_MAX_SIZE_MB = int(os.environ.get("VLLM_MAX_SIZE_MB", 400))
 
 
 def print_top_10_largest_files(zip_file):

.buildkite/lm-eval-harness/configs/Meta-Llama-3-8B-QQQ.yaml

@@ -1,12 +0,0 @@
-# For vllm script, with -t option (tensor parallel size).
-# bash .buildkite/lm-eval-harness/run-lm-eval-gsm-vllm-baseline.sh -m HandH1998/QQQ-Llama-3-8b-g128 -b 32 -l 1000 -f 5 -t 1
-model_name: "HandH1998/QQQ-Llama-3-8b-g128"
-tasks:
-- name: "gsm8k"
-  metrics:
-  - name: "exact_match,strict-match"
-    value: 0.419
-  - name: "exact_match,flexible-extract"
-    value: 0.416
-limit: 1000
-num_fewshot: 5

.buildkite/lm-eval-harness/configs/Meta-Llama-4-Maverick-17B-128E-Instruct-FP8-MM.yaml

@@ -1,12 +0,0 @@
-# For hf script, without -t option (tensor parallel size).
-# bash .buildkite/lm-eval-harness/run-lm-eval-chartqa-vllm-vlm-baseline.sh -m meta-llama/Llama-4-Maverick-17B-128E-Instruct-FP8 -l 100 -t 8
-model_name: "meta-llama/Llama-4-Maverick-17B-128E-Instruct-FP8"
-backend: "vllm-vlm"
-tasks:
-- name: "chartqa"
-  metrics:
-  - name: "relaxed_accuracy,none"
-    # TODO(zhewenl): model card is 0.90, but the actual score is 0.80.
-    value: 0.80
-limit: 100
-num_fewshot: 0

.buildkite/lm-eval-harness/configs/Meta-Llama-4-Maverick-17B-128E-Instruct-FP8.yaml

@@ -1,10 +0,0 @@
-# For hf script, without -t option (tensor parallel size).
-# bash .buildkite/lm-eval-harness/run-lm-eval-mmlupro-vllm-baseline.sh -m meta-llama/Llama-4-Maverick-17B-128E-Instruct-FP8 -l 250 -t 8 -f 5
-model_name: "meta-llama/Llama-4-Maverick-17B-128E-Instruct-FP8"
-tasks:
-- name: "mmlu_pro"
-  metrics:
-  - name: "exact_match,custom-extract"
-    value: 0.80
-limit: 250 # will run on 250 * 14 subjects = 3500 samples
-num_fewshot: 5

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

@@ -1,5 +1,4 @@
-# For vllm script, with -t option (tensor parallel size)
-# bash .buildkite/lm-eval-harness/run-lm-eval-gsm-vllm-baseline.sh -m RedHatAI/Qwen2.5-VL-3B-Instruct-FP8-Dynamic -l 1319 -t 1
+# 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"

.buildkite/lm-eval-harness/configs/Qwen2.5-VL-7B-Instruct.yaml

@@ -1,12 +0,0 @@
-# For vllm script, with -t option (tensor parallel size).
-# bash .buildkite/lm-eval-harness/run-lm-eval-chartqa-vllm-vlm-baseline.sh -m Qwen/Qwen2.5-VL-7B-Instruct -l 2500 -t 1
-
-model_name: "Qwen/Qwen2.5-VL-7B-Instruct"
-backend: "vllm-vlm"
-tasks:
-- name: "chartqa"
-  metrics:
-  - name: "relaxed_accuracy,none"
-    value: 0.855
-limit: 2500
-num_fewshot: 0

.buildkite/lm-eval-harness/configs/models-large-h100.txt

@@ -1 +0,0 @@
-Meta-Llama-4-Maverick-17B-128E-Instruct-FP8.yaml

.buildkite/lm-eval-harness/configs/models-mm-large-h100.txt

@@ -1 +0,0 @@
-Meta-Llama-4-Maverick-17B-128E-Instruct-FP8-MM.yaml

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

@@ -1 +0,0 @@
-Qwen2.5-VL-7B-Instruct.yaml

.buildkite/lm-eval-harness/run-lm-eval-chartqa-vllm-vlm-baseline.sh

@@ -1,44 +0,0 @@
-#!/bin/bash
-# We can use this script to compute baseline accuracy on chartqa for vllm.
-#
-# Make sure you have lm-eval-harness installed:
-#   pip install lm-eval==0.4.9
-
-usage() {
-    echo``
-    echo "Runs lm eval harness on ChartQA using multimodal vllm."
-    echo "This pathway is intended to be used to create baselines for "
-    echo "our correctness tests in vllm's CI."
-    echo
-    echo "usage: ${0} <options>"
-    echo
-    echo "  -m    - huggingface stub or local directory of the model"
-    echo "  -l    - limit number of samples to run"
-    echo "  -t    - tensor parallel size to run at"
-    echo
-}
-
-while getopts "m:l:t:" OPT; do
-  case ${OPT} in
-    m ) 
-        MODEL="$OPTARG"
-        ;;
-    l ) 
-        LIMIT="$OPTARG"
-        ;;
-    t ) 
-        TP_SIZE="$OPTARG"
-        ;;
-    \? ) 
-        usage
-        exit 1
-        ;;
-  esac
-done
-
-lm_eval --model vllm-vlm \
-  --model_args "pretrained=$MODEL,tensor_parallel_size=$TP_SIZE" \
-  --tasks chartqa \
-  --batch_size auto \
-  --apply_chat_template \
-  --limit $LIMIT

.buildkite/lm-eval-harness/run-lm-eval-mmlupro-vllm-baseline.sh

@@ -1,50 +0,0 @@
-#!/bin/bash
-# We can use this script to compute baseline accuracy on MMLUPRO for vllm.
-# We use this for fp8, which HF does not support.
-#
-# Make sure you have lm-eval-harness installed:
-#   pip install git+https://github.com/EleutherAI/lm-evaluation-harness.git@206b7722158f58c35b7ffcd53b035fdbdda5126d#egg=lm-eval[api]
-
-usage() {
-    echo``
-    echo "Runs lm eval harness on MMLU Pro using huggingface transformers."
-    echo "This pathway is intended to be used to create baselines for "
-    echo "our automated nm-test-accuracy workflow"
-    echo
-    echo "usage: ${0} <options>"
-    echo
-    echo "  -m    - huggingface stub or local directory of the model"
-    echo "  -l    - limit number of samples to run"
-    echo "  -f    - number of fewshot samples to use"
-    echo "  -t    - tensor parallel size to run at"
-    echo
-}
-
-while getopts "m:b:l:f:t:" OPT; do
-  case ${OPT} in
-    m )
-        MODEL="$OPTARG"
-        ;;
-    b )
-        BATCH_SIZE="$OPTARG"
-        ;;
-    l )
-        LIMIT="$OPTARG"
-        ;;
-    f )
-        FEWSHOT="$OPTARG"
-        ;;
-    t )
-        TP_SIZE="$OPTARG"
-        ;;
-    \? )
-        usage
-        exit 1
-        ;;
-  esac
-done
-
-lm_eval --model vllm \
-  --model_args "pretrained=$MODEL,tensor_parallel_size=$TP_SIZE,add_bos_token=true,trust_remote_code=true,max_model_len=4096" \
-  --tasks mmlu_pro --num_fewshot "$FEWSHOT" --limit "$LIMIT" \
-  --batch_size auto

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

@@ -19,27 +19,21 @@ RTOL = 0.08
 def launch_lm_eval(eval_config, tp_size):
     trust_remote_code = eval_config.get("trust_remote_code", False)
     max_model_len = eval_config.get("max_model_len", 4096)
-    batch_size = eval_config.get("batch_size", "auto")
-    backend = eval_config.get("backend", "vllm")
     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},"
+        f"max_model_len={max_model_len}"
     )
     results = lm_eval.simple_evaluate(
-        model=backend,
+        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"],
-        # TODO(yeq): using chat template w/ fewshot_as_multiturn is supposed help
-        # text models. however, this is regressing measured strict-match for
-        # existing text models in CI, so only apply it for mm.
-        apply_chat_template=backend == "vllm-vlm",
-        batch_size=batch_size,
+        batch_size="auto",
     )
     return results
 

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

@@ -8,7 +8,7 @@ This benchmark aims to:
 
 Latest results: [results link](https://blog.vllm.ai/2024/09/05/perf-update.html), scroll to the end.
 
-Latest reproduction guide: [github issue link](https://github.com/vllm-project/vllm/issues/8176)
+Latest reproduction guilde: [github issue link](https://github.com/vllm-project/vllm/issues/8176)
 
 ## Setup
 

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

@@ -7,7 +7,6 @@ from importlib import util
 
 import pandas as pd
 
-pd.options.display.float_format = "{:.2f}".format
 plotly_found = util.find_spec("plotly.express") is not None
 
 
@@ -110,10 +109,7 @@ def compare_data_columns(
         if len(compare_frames) >= 2:
             base = compare_frames[0]
             current = compare_frames[-1]
-            if "P99" in data_column or "Median" in data_column:
-                ratio = base / current  # for latency
-            else:
-                ratio = current / base
+            ratio = current / base
             ratio = ratio.mask(base == 0)  # avoid inf when baseline is 0
             ratio.name = f"Ratio 1 vs {len(compare_frames)}"
             frames.append(ratio)
@@ -203,71 +199,6 @@ def split_json_by_tp_pp(
     return saved_paths
 
 
-def _add_limit_line(fig, y_value, label):
-    # Visible dashed line + annotation
-    fig.add_hline(
-        y=y_value,
-        line_dash="dash",
-        line_color="red" if "ttft" in label.lower() else "blue",
-        annotation_text=f"{label}: {y_value} ms",
-        annotation_position="top left",
-    )
-    # Optional: add a legend item (as a transparent helper trace)
-    if plot and plotly_found:
-        import plotly.graph_objects as go
-
-        fig.add_trace(
-            go.Scatter(
-                x=[None],
-                y=[None],
-                mode="lines",
-                line=dict(
-                    dash="dash", color="red" if "ttft" in label.lower() else "blue"
-                ),
-                name=f"{label}",
-            )
-        )
-
-
-def _find_concurrency_col(df: pd.DataFrame) -> str:
-    for c in [
-        "# of max concurrency.",
-        "# of max concurrency",
-        "Max Concurrency",
-        "max_concurrency",
-        "Concurrency",
-    ]:
-        if c in df.columns:
-            return c
-    # Fallback: guess an integer-like column (harmless if unused)
-    for c in df.columns:
-        if df[c].dtype.kind in "iu" and df[c].nunique() > 1 and df[c].min() >= 1:
-            return c
-    return "# of max concurrency."
-
-
-def _highlight_threshold(
-    df: pd.DataFrame, threshold: float
-) -> "pd.io.formats.style.Styler":
-    """Highlight numeric per-configuration columns with value <= threshold."""
-    conc_col = _find_concurrency_col(df)
-    key_cols = [
-        c
-        for c in ["Model", "Dataset Name", "Input Len", "Output Len", conc_col]
-        if c in df.columns
-    ]
-    conf_cols = [
-        c for c in df.columns if c not in key_cols and not str(c).startswith("Ratio")
-    ]
-    conf_cols = [c for c in conf_cols if pd.api.types.is_numeric_dtype(df[c])]
-    return df.style.map(
-        lambda v: "background-color:#e6ffe6;font-weight:bold;"
-        if pd.notna(v) and v <= threshold
-        else "",
-        subset=conf_cols,
-    )
-
-
 if __name__ == "__main__":
     parser = argparse.ArgumentParser()
     parser.add_argument(
@@ -287,28 +218,8 @@ if __name__ == "__main__":
         "--xaxis",
         type=str,
         default="# of max concurrency.",
-        help="column name to use as X Axis in comparison graph",
+        help="column name to use as X Axis in comparision graph",
     )
-    parser.add_argument(
-        "-l",
-        "--latency",
-        type=str,
-        default="p99",
-        help="take median|p99 for latency like TTFT/TPOT",
-    )
-    parser.add_argument(
-        "--ttft-max-ms",
-        type=float,
-        default=3000.0,
-        help="Reference limit for TTFT plots (ms)",
-    )
-    parser.add_argument(
-        "--tpot-max-ms",
-        type=float,
-        default=100.0,
-        help="Reference limit for TPOT plots (ms)",
-    )
-
     args = parser.parse_args()
 
     drop_column = "P99"
@@ -323,22 +234,12 @@ if __name__ == "__main__":
         "# of max concurrency.",
         "qps",
     ]
-
-    if "median" in args.latency:
-        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",
-        ]
-        drop_column = "P99"
-    elif "p99" in args.latency:
-        data_cols_to_compare = ["Output Tput (tok/s)", "P99 TTFT (ms)", "P99"]
-        html_msgs_for_data_cols = [
-            "Compare Output Tokens /n",
-            "P99 TTFT /n",
-            "P99 TPOT /n",
-        ]
+    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",
+    ]
 
     if len(args.file) == 1:
         files = split_json_by_tp_pp(args.file[0], output_root="splits")
@@ -374,83 +275,33 @@ if __name__ == "__main__":
                     f"Expected subset: {filtered_info_cols}, "
                     f"but DataFrame has: {list(output_df.columns)}"
                 )
-            # output_df_sorted = output_df.sort_values(by=existing_group_cols)
-            output_df_sorted = output_df.sort_values(by=args.xaxis)
+            output_df_sorted = output_df.sort_values(by=existing_group_cols)
             output_groups = output_df_sorted.groupby(existing_group_cols, dropna=False)
             for name, group in output_groups:
-                group_name = (
-                    ",".join(map(str, name)).replace(",", "_").replace("/", "-")
-                )
-                group_html_name = "perf_comparison_" + group_name + ".html"
-
-                metric_name = str(data_cols_to_compare[i]).lower()
-                if "tok/s" in metric_name:
-                    html = group.to_html()
-                elif "ttft" in metric_name:
-                    styler = _highlight_threshold(group, args.ttft_max_ms).format(
-                        {c: "{:.2f}" for c in group.select_dtypes("number").columns},
-                        na_rep="—",
-                    )
-                    html = styler.to_html(
-                        table_attributes='border="1" class="dataframe"'
-                    )
-                elif (
-                    "tpot" in metric_name
-                    or "median" in metric_name
-                    or "p99" in metric_name
-                ):
-                    styler = _highlight_threshold(group, args.tpot_max_ms).format(
-                        {c: "{:.2f}" for c in group.select_dtypes("number").columns},
-                        na_rep="—",
-                    )
-                    html = styler.to_html(
-                        table_attributes='border="1" class="dataframe"'
-                    )
-
+                html = group.to_html()
                 text_file.write(html_msgs_for_data_cols[i])
                 text_file.write(html)
-                with open(group_html_name, "a+") as sub_text_file:
-                    sub_text_file.write(html_msgs_for_data_cols[i])
-                    sub_text_file.write(html)
 
-                    if plot and plotly_found:
-                        import plotly.express as px
+                if plot and plotly_found:
+                    import plotly.express as px
 
-                        df = group[raw_data_cols]
-                        df_sorted = df.sort_values(by=info_cols[y_axis_index])
-                        # Melt DataFrame for plotting
-                        df_melted = df_sorted.melt(
-                            id_vars=info_cols[y_axis_index],
-                            var_name="Configuration",
-                            value_name=data_cols_to_compare[i],
-                        )
-                        title = (
-                            data_cols_to_compare[i] + " vs " + info_cols[y_axis_index]
-                        )
-                        # Create Plotly line chart
-                        fig = px.line(
-                            df_melted,
-                            x=info_cols[y_axis_index],
-                            y=data_cols_to_compare[i],
-                            color="Configuration",
-                            title=title,
-                            markers=True,
-                        )
-
-                        # ---- Add threshold lines based on metric name ----
-                        if "ttft" in metric_name:
-                            _add_limit_line(fig, args.ttft_max_ms, "TTFT limit")
-                        elif (
-                            "tpot" in metric_name
-                            or "median" in metric_name
-                            or "p99" in metric_name
-                        ):
-                            _add_limit_line(fig, args.tpot_max_ms, "TPOT limit")
-
-                        # Export to HTML
-                        text_file.write(
-                            fig.to_html(full_html=True, include_plotlyjs="cdn")
-                        )
-                        sub_text_file.write(
-                            fig.to_html(full_html=True, include_plotlyjs="cdn")
-                        )
+                    df = group[raw_data_cols]
+                    df_sorted = df.sort_values(by=info_cols[y_axis_index])
+                    # Melt DataFrame for plotting
+                    df_melted = df_sorted.melt(
+                        id_vars=info_cols[y_axis_index],
+                        var_name="Configuration",
+                        value_name=data_cols_to_compare[i],
+                    )
+                    title = data_cols_to_compare[i] + " vs " + info_cols[y_axis_index]
+                    # Create Plotly line chart
+                    fig = px.line(
+                        df_melted,
+                        x=info_cols[y_axis_index],
+                        y=data_cols_to_compare[i],
+                        color="Configuration",
+                        title=title,
+                        markers=True,
+                    )
+                    # Export to HTML
+                    text_file.write(fig.to_html(full_html=True, include_plotlyjs="cdn"))

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

@@ -63,11 +63,9 @@ serving_column_mapping = {
     "mean_ttft_ms": "Mean TTFT (ms)",
     "median_ttft_ms": "Median TTFT (ms)",
     "p99_ttft_ms": "P99 TTFT (ms)",
-    "std_ttft_ms": "STD TTFT (ms)",
     "mean_tpot_ms": "Mean TPOT (ms)",
     "median_tpot_ms": "Median",
     "p99_tpot_ms": "P99",
-    "std_tpot_ms": "STD TPOT (ms)",
     "mean_itl_ms": "Mean ITL (ms)",
     "median_itl_ms": "Median ITL (ms)",
     "p99_itl_ms": "P99 ITL (ms)",
@@ -370,7 +368,7 @@ 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: "{}x{}".format(len(x.split("\n")), x.split("\n")[0])
+            lambda x: f"{len(x.split('\n'))}x{x.split('\n')[0]}"
         )
 
     # get markdown tables

.buildkite/nightly-benchmarks/scripts/launch-server.sh

@@ -181,14 +181,18 @@ launch_vllm_server() {
   if echo "$common_params" | jq -e 'has("fp8")' >/dev/null; then
     echo "Key 'fp8' exists in common params. Use neuralmagic fp8 model for convenience."
     model=$(echo "$common_params" | jq -r '.neuralmagic_quantized_model')
-    server_command="vllm serve $model \
+    server_command="python3 \
+        -m vllm.entrypoints.openai.api_server \
         -tp $tp \
+        --model $model \
         --port $port \
         $server_args"
   else
     echo "Key 'fp8' does not exist in common params."
-    server_command="vllm serve $model \
+    server_command="python3 \
+        -m vllm.entrypoints.openai.api_server \
         -tp $tp \
+        --model $model \
         --port $port \
         $server_args"
   fi

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

@@ -365,7 +365,8 @@ run_serving_tests() {
       continue
     fi
 
-    server_command="$server_envs vllm serve \
+    server_command="$server_envs python3 \
+      -m vllm.entrypoints.openai.api_server \
       $server_args"
 
     # run the server
@@ -454,6 +455,11 @@ main() {
   fi
   check_hf_token
 
+  # Set to v1 to run v1 benchmark
+  if [[ "${ENGINE_VERSION:-v0}" == "v1" ]]; then
+    export VLLM_USE_V1=1
+  fi
+
   # dependencies
   (which wget && which curl) || (apt-get update && apt-get install -y wget curl)
   (which jq) || (apt-get update && apt-get -y install jq)
@@ -471,11 +477,6 @@ main() {
   mkdir -p $RESULTS_FOLDER
   QUICK_BENCHMARK_ROOT=../.buildkite/nightly-benchmarks/
 
-  # dump vllm info via vllm collect-env
-  env_output=$(vllm collect-env)
-
-  echo "$env_output" >"$RESULTS_FOLDER/vllm_env.txt"
-
   # benchmarking
   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}"

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

@@ -1,24 +1,28 @@
 [
     {
-        "test_name": "latency_llama8B_tp2",
+        "test_name": "latency_llama8B_tp1",
         "environment_variables": {
-            "VLLM_RPC_TIMEOUT": 100000,
 	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
-	    "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
-	    "VLLM_CPU_SGL_KERNEL": 1,
 	    "VLLM_CPU_KVCACHE_SPACE": 40
         },
         "parameters": {
             "model": "meta-llama/Llama-3.1-8B-Instruct",
-            "tensor_parallel_size": 2,
-	    "dtype": "bfloat16",
-	    "distributed_executor_backend": "mp",
-	    "block_size": 128,
-	    "trust_remote_code": "",
-            "disable_log_stats": "",
-	    "enforce_eager": "",
-	    "max_num_batched_tokens": 2048,
-	    "max_num_seqs": 256,
+            "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/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-snc2.json

@@ -1,6 +1,6 @@
 [
     {
-        "test_name": "serving_llama8B_bf16_tp1_sharegpt",
+        "test_name": "serving_llama8B_tp1_sharegpt",
         "qps_list": ["inf"],
         "max_concurrency_list": [12, 16, 24, 32, 64, 128, 200],
         "server_environment_variables": {
@@ -32,7 +32,7 @@
         }
     },
     {
-        "test_name": "serving_llama8B_bf16_tp2_sharegpt",
+        "test_name": "serving_llama8B_tp2_sharegpt",
         "qps_list": ["inf"],
         "max_concurrency_list": [12, 16, 24, 32, 64, 128, 200],
         "server_environment_variables": {
@@ -64,7 +64,7 @@
         }
     },
     {
-        "test_name": "serving_llama8B_bf16_tp4_sharegpt",
+        "test_name": "serving_llama8B_tp4_sharegpt",
         "qps_list": ["inf"],
         "max_concurrency_list": [12, 16, 24, 32, 64, 128, 200],
         "server_environment_variables": {
@@ -96,7 +96,7 @@
         }
     },
     {
-        "test_name": "serving_llama8B_bf16_tp1_random_128_128",
+        "test_name": "serving_llama8B_tp1_random_128_128",
         "qps_list": ["inf"],
         "max_concurrency_list": [12, 16, 24, 32, 64, 128, 200, 1000],
         "server_environment_variables": {
@@ -131,7 +131,7 @@
         }
     },
     {
-        "test_name": "serving_llama8B_bf16_tp2_random_128_128",
+        "test_name": "serving_llama8B_tp2_random_128_128",
         "qps_list": ["inf"],
         "max_concurrency_list": [12, 16, 24, 32, 64, 128, 200, 1000],
         "server_environment_variables": {
@@ -166,7 +166,7 @@
         }
     },
     {
-        "test_name": "serving_llama8B_bf16_tp4_random_128_128",
+        "test_name": "serving_llama8B_tp4_random_128_128",
         "qps_list": ["inf"],
         "max_concurrency_list": [12, 16, 24, 32, 64, 128, 200, 1000],
         "server_environment_variables": {
@@ -198,413 +198,5 @@
 	    "random-output-len": 128,
             "num_prompts": 1000
         }
-    },
-    {
-        "test_name": "serving_llama8B_int8_tp1_sharegpt",
-        "qps_list": ["inf"],
-        "max_concurrency_list": [12, 16, 24, 32, 64, 128, 200],
-        "server_environment_variables": {
-            "VLLM_RPC_TIMEOUT": 100000,
-	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
-	    "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
-	    "VLLM_CPU_SGL_KERNEL": 1,
-	    "VLLM_CPU_KVCACHE_SPACE": 40
-        },
-        "server_parameters": {
-            "model": "RedHatAI/Meta-Llama-3.1-8B-Instruct-quantized.w8a8",
-            "tensor_parallel_size": 1,
-	    "dtype": "bfloat16",
-	    "distributed_executor_backend": "mp",
-	    "block_size": 128,
-	    "trust_remote_code": "",
-            "disable_log_stats": "",
-	    "enforce_eager": "",
-	    "max_num_batched_tokens": 2048,
-	    "max_num_seqs": 256,
-            "load_format": "dummy"
-        },
-        "client_parameters": {
-            "model": "RedHatAI/Meta-Llama-3.1-8B-Instruct-quantized.w8a8",
-            "backend": "vllm",
-            "dataset_name": "sharegpt",
-            "dataset_path": "./ShareGPT_V3_unfiltered_cleaned_split.json",
-            "num_prompts": 200
-        }
-    },
-    {
-        "test_name": "serving_llama8B_int8_tp2_sharegpt",
-        "qps_list": ["inf"],
-        "max_concurrency_list": [12, 16, 24, 32, 64, 128, 200],
-        "server_environment_variables": {
-            "VLLM_RPC_TIMEOUT": 100000,
-	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
-	    "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
-	    "VLLM_CPU_SGL_KERNEL": 1,
-	    "VLLM_CPU_KVCACHE_SPACE": 40
-        },
-        "server_parameters": {
-            "model": "RedHatAI/Meta-Llama-3.1-8B-Instruct-quantized.w8a8",
-            "tensor_parallel_size": 2,
-	    "dtype": "bfloat16",
-	    "distributed_executor_backend": "mp",
-	    "block_size": 128,
-	    "trust_remote_code": "",
-            "disable_log_stats": "",
-	    "enforce_eager": "",
-	    "max_num_batched_tokens": 2048,
-	    "max_num_seqs": 256,
-            "load_format": "dummy"
-        },
-        "client_parameters": {
-            "model": "RedHatAI/Meta-Llama-3.1-8B-Instruct-quantized.w8a8",
-            "backend": "vllm",
-            "dataset_name": "sharegpt",
-            "dataset_path": "./ShareGPT_V3_unfiltered_cleaned_split.json",
-            "num_prompts": 200
-        }
-    },
-    {
-        "test_name": "serving_llama8B_int8_tp4_sharegpt",
-        "qps_list": ["inf"],
-        "max_concurrency_list": [12, 16, 24, 32, 64, 128, 200],
-        "server_environment_variables": {
-            "VLLM_RPC_TIMEOUT": 100000,
-	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
-	    "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
-	    "VLLM_CPU_SGL_KERNEL": 1,
-	    "VLLM_CPU_KVCACHE_SPACE": 40
-        },
-        "server_parameters": {
-            "model": "RedHatAI/Meta-Llama-3.1-8B-Instruct-quantized.w8a8",
-            "tensor_parallel_size": 4,
-	    "dtype": "bfloat16",
-	    "distributed_executor_backend": "mp",
-	    "block_size": 128,
-	    "trust_remote_code": "",
-            "disable_log_stats": "",
-	    "enforce_eager": "",
-	    "max_num_batched_tokens": 2048,
-	    "max_num_seqs": 256,
-            "load_format": "dummy"
-        },
-        "client_parameters": {
-            "model": "RedHatAI/Meta-Llama-3.1-8B-Instruct-quantized.w8a8",
-            "backend": "vllm",
-            "dataset_name": "sharegpt",
-            "dataset_path": "./ShareGPT_V3_unfiltered_cleaned_split.json",
-            "num_prompts": 200
-        }
-    },
-    {
-        "test_name": "serving_llama8B_int8_tp1_random_128_128",
-        "qps_list": ["inf"],
-        "max_concurrency_list": [12, 16, 24, 32, 64, 128, 200, 1000],
-        "server_environment_variables": {
-            "VLLM_RPC_TIMEOUT": 100000,
-	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
-	    "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
-	    "VLLM_CPU_SGL_KERNEL": 1,
-	    "VLLM_CPU_KVCACHE_SPACE": 40
-        },
-        "server_parameters": {
-            "model": "RedHatAI/Meta-Llama-3.1-8B-Instruct-quantized.w8a8",
-            "tensor_parallel_size": 1,
-	    "dtype": "bfloat16",
-	    "distributed_executor_backend": "mp",
-	    "block_size": 128,
-	    "trust_remote_code": "",
-	    "enable_chunked_prefill": "",
-            "disable_log_stats": "",
-	    "enforce_eager": "",
-	    "max_num_batched_tokens": 2048,
-	    "max_num_seqs": 256,
-            "load_format": "dummy"
-        },
-        "client_parameters": {
-            "model": "RedHatAI/Meta-Llama-3.1-8B-Instruct-quantized.w8a8",
-            "backend": "vllm",
-            "dataset_name": "random",
-	    "random-input-len": 128,
-	    "random-output-len": 128,
-	    "ignore-eos": "",
-            "num_prompts": 1000
-        }
-    },
-    {
-        "test_name": "serving_llama8B_int8_tp2_random_128_128",
-        "qps_list": ["inf"],
-        "max_concurrency_list": [12, 16, 24, 32, 64, 128, 200, 1000],
-        "server_environment_variables": {
-            "VLLM_RPC_TIMEOUT": 100000,
-	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
-	    "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
-	    "VLLM_CPU_SGL_KERNEL": 1,
-	    "VLLM_CPU_KVCACHE_SPACE": 40
-        },
-        "server_parameters": {
-            "model": "RedHatAI/Meta-Llama-3.1-8B-Instruct-quantized.w8a8",
-            "tensor_parallel_size": 2,
-	    "dtype": "bfloat16",
-	    "distributed_executor_backend": "mp",
-	    "block_size": 128,
-	    "trust_remote_code": "",
-	    "enable_chunked_prefill": "",
-            "disable_log_stats": "",
-	    "enforce_eager": "",
-	    "max_num_batched_tokens": 2048,
-	    "max_num_seqs": 256,
-            "load_format": "dummy"
-        },
-        "client_parameters": {
-            "model": "RedHatAI/Meta-Llama-3.1-8B-Instruct-quantized.w8a8",
-            "backend": "vllm",
-            "dataset_name": "random",
-	    "random-input-len": 128,
-	    "random-output-len": 128,
-	    "ignore-eos": "",
-            "num_prompts": 1000
-        }
-    },
-    {
-        "test_name": "serving_llama8B_int8_tp4_random_128_128",
-        "qps_list": ["inf"],
-        "max_concurrency_list": [12, 16, 24, 32, 64, 128, 200, 1000],
-        "server_environment_variables": {
-            "VLLM_RPC_TIMEOUT": 100000,
-	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
-	    "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
-	    "VLLM_CPU_SGL_KERNEL": 1,
-	    "VLLM_CPU_KVCACHE_SPACE": 40
-        },
-        "server_parameters": {
-            "model": "RedHatAI/Meta-Llama-3.1-8B-Instruct-quantized.w8a8",
-            "tensor_parallel_size": 4,
-	    "dtype": "bfloat16",
-	    "distributed_executor_backend": "mp",
-	    "block_size": 128,
-	    "trust_remote_code": "",
-	    "enable_chunked_prefill": "",
-            "disable_log_stats": "",
-	    "enforce_eager": "",
-	    "max_num_batched_tokens": 2048,
-	    "max_num_seqs": 256,
-            "load_format": "dummy"
-        },
-        "client_parameters": {
-            "model": "RedHatAI/Meta-Llama-3.1-8B-Instruct-quantized.w8a8",
-            "backend": "vllm",
-            "dataset_name": "random",
-	    "random-input-len": 128,
-	    "random-output-len": 128,
-	    "ignore-eos": "",
-            "num_prompts": 1000
-        }
-    },
-    {
-        "test_name": "serving_llama8B_int4_tp1_sharegpt",
-        "qps_list": ["inf"],
-        "max_concurrency_list": [12, 16, 24, 32, 64, 128, 200],
-        "server_environment_variables": {
-            "VLLM_RPC_TIMEOUT": 100000,
-	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
-	    "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
-	    "VLLM_CPU_SGL_KERNEL": 1,
-	    "VLLM_CPU_KVCACHE_SPACE": 40
-        },
-        "server_parameters": {
-            "model": "hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4",
-	    "quantization": "awq",
-            "tensor_parallel_size": 1,
-	    "dtype": "bfloat16",
-	    "distributed_executor_backend": "mp",
-	    "block_size": 128,
-	    "trust_remote_code": "",
-            "disable_log_stats": "",
-	    "enforce_eager": "",
-	    "max_num_batched_tokens": 2048,
-	    "max_num_seqs": 256,
-            "load_format": "dummy"
-        },
-        "client_parameters": {
-            "model": "hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4",
-            "backend": "vllm",
-            "dataset_name": "sharegpt",
-            "dataset_path": "./ShareGPT_V3_unfiltered_cleaned_split.json",
-            "num_prompts": 200
-        }
-    },
-    {
-        "test_name": "serving_llama8B_int4_tp2_sharegpt",
-        "qps_list": ["inf"],
-        "max_concurrency_list": [12, 16, 24, 32, 64, 128, 200],
-        "server_environment_variables": {
-            "VLLM_RPC_TIMEOUT": 100000,
-	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
-	    "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
-	    "VLLM_CPU_SGL_KERNEL": 1,
-	    "VLLM_CPU_KVCACHE_SPACE": 40
-        },
-        "server_parameters": {
-            "model": "hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4",
-	    "quantization": "awq",
-            "tensor_parallel_size": 2,
-	    "dtype": "bfloat16",
-	    "distributed_executor_backend": "mp",
-	    "block_size": 128,
-	    "trust_remote_code": "",
-            "disable_log_stats": "",
-	    "enforce_eager": "",
-	    "max_num_batched_tokens": 2048,
-	    "max_num_seqs": 256,
-            "load_format": "dummy"
-        },
-        "client_parameters": {
-            "model": "hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4",
-            "backend": "vllm",
-            "dataset_name": "sharegpt",
-            "dataset_path": "./ShareGPT_V3_unfiltered_cleaned_split.json",
-            "num_prompts": 200
-        }
-    },
-    {
-        "test_name": "serving_llama8B_int4_tp4_sharegpt",
-        "qps_list": ["inf"],
-        "max_concurrency_list": [12, 16, 24, 32, 64, 128, 200],
-        "server_environment_variables": {
-            "VLLM_RPC_TIMEOUT": 100000,
-	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
-	    "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
-	    "VLLM_CPU_SGL_KERNEL": 1,
-	    "VLLM_CPU_KVCACHE_SPACE": 40
-        },
-        "server_parameters": {
-            "model": "hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4",
-	    "quantization": "awq",
-            "tensor_parallel_size": 4,
-	    "dtype": "bfloat16",
-	    "distributed_executor_backend": "mp",
-	    "block_size": 128,
-	    "trust_remote_code": "",
-            "disable_log_stats": "",
-	    "enforce_eager": "",
-	    "max_num_batched_tokens": 2048,
-	    "max_num_seqs": 256,
-            "load_format": "dummy"
-        },
-        "client_parameters": {
-            "model": "hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4",
-            "backend": "vllm",
-            "dataset_name": "sharegpt",
-            "dataset_path": "./ShareGPT_V3_unfiltered_cleaned_split.json",
-            "num_prompts": 200
-        }
-    },
-    {
-        "test_name": "serving_llama8B_int4_tp1_random_128_128",
-        "qps_list": ["inf"],
-        "max_concurrency_list": [12, 16, 24, 32, 64, 128, 200, 1000],
-        "server_environment_variables": {
-            "VLLM_RPC_TIMEOUT": 100000,
-	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
-	    "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
-	    "VLLM_CPU_SGL_KERNEL": 1,
-	    "VLLM_CPU_KVCACHE_SPACE": 40
-        },
-        "server_parameters": {
-            "model": "hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4",
-	    "quantization": "awq",
-            "tensor_parallel_size": 1,
-	    "dtype": "bfloat16",
-	    "distributed_executor_backend": "mp",
-	    "block_size": 128,
-	    "trust_remote_code": "",
-	    "enable_chunked_prefill": "",
-            "disable_log_stats": "",
-	    "enforce_eager": "",
-	    "max_num_batched_tokens": 2048,
-	    "max_num_seqs": 256,
-            "load_format": "dummy"
-        },
-        "client_parameters": {
-            "model": "hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4",
-            "backend": "vllm",
-            "dataset_name": "random",
-	    "random-input-len": 128,
-	    "random-output-len": 128,
-	    "ignore-eos": "",
-            "num_prompts": 1000
-        }
-    },
-    {
-        "test_name": "serving_llama8B_int4_tp2_random_128_128",
-        "qps_list": ["inf"],
-        "max_concurrency_list": [12, 16, 24, 32, 64, 128, 200, 1000],
-        "server_environment_variables": {
-            "VLLM_RPC_TIMEOUT": 100000,
-	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
-	    "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
-	    "VLLM_CPU_SGL_KERNEL": 1,
-	    "VLLM_CPU_KVCACHE_SPACE": 40
-        },
-        "server_parameters": {
-            "model": "hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4",
-	    "quantization": "awq",
-            "tensor_parallel_size": 2,
-	    "dtype": "bfloat16",
-	    "distributed_executor_backend": "mp",
-	    "block_size": 128,
-	    "trust_remote_code": "",
-	    "enable_chunked_prefill": "",
-            "disable_log_stats": "",
-	    "enforce_eager": "",
-	    "max_num_batched_tokens": 2048,
-	    "max_num_seqs": 256,
-            "load_format": "dummy"
-        },
-        "client_parameters": {
-            "model": "hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4",
-            "backend": "vllm",
-            "dataset_name": "random",
-	    "random-input-len": 128,
-	    "random-output-len": 128,
-	    "ignore-eos": "",
-            "num_prompts": 1000
-        }
-    },
-    {
-        "test_name": "serving_llama8B_int4_tp4_random_128_128",
-        "qps_list": ["inf"],
-        "max_concurrency_list": [12, 16, 24, 32, 64, 128, 200, 1000],
-        "server_environment_variables": {
-            "VLLM_RPC_TIMEOUT": 100000,
-	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
-	    "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
-	    "VLLM_CPU_SGL_KERNEL": 1,
-	    "VLLM_CPU_KVCACHE_SPACE": 40
-        },
-        "server_parameters": {
-            "model": "hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4",
-	    "quantization": "awq",
-            "tensor_parallel_size": 4,
-	    "dtype": "bfloat16",
-	    "distributed_executor_backend": "mp",
-	    "block_size": 128,
-	    "trust_remote_code": "",
-	    "enable_chunked_prefill": "",
-            "disable_log_stats": "",
-	    "enforce_eager": "",
-	    "max_num_batched_tokens": 2048,
-	    "max_num_seqs": 256,
-            "load_format": "dummy"
-        },
-        "client_parameters": {
-            "model": "hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4",
-            "backend": "vllm",
-            "dataset_name": "random",
-	    "random-input-len": 128,
-	    "random-output-len": 128,
-	    "ignore-eos": "",
-            "num_prompts": 1000
-        }
     }
 ]

.buildkite/nightly-benchmarks/tests/serving-tests-cpu-snc3.json

@@ -1,6 +1,6 @@
 [
     {
-        "test_name": "serving_llama8B_bf16_pp1_sharegpt",
+        "test_name": "serving_llama8B_pp1_sharegpt",
         "qps_list": ["inf"],
         "max_concurrency_list": [12, 16, 24, 32, 64, 128, 200],
         "server_environment_variables": {
@@ -32,39 +32,7 @@
         }
     },
     {
-        "test_name": "serving_llama8B_bf16_tp2_sharegpt",
-        "qps_list": ["inf"],
-        "max_concurrency_list": [12, 16, 24, 32, 64, 128, 200],
-        "server_environment_variables": {
-            "VLLM_RPC_TIMEOUT": 100000,
-	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
-	    "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
-	    "VLLM_CPU_SGL_KERNEL": 1,
-	    "VLLM_CPU_KVCACHE_SPACE": 40
-        },
-        "server_parameters": {
-            "model": "meta-llama/Llama-3.1-8B-Instruct",
-            "tensor_parallel_size": 2,
-	    "dtype": "bfloat16",
-	    "distributed_executor_backend": "mp",
-	    "block_size": 128,
-	    "trust_remote_code": "",
-            "disable_log_stats": "",
-	    "enforce_eager": "",
-	    "max_num_batched_tokens": 2048,
-	    "max_num_seqs": 256,
-            "load_format": "dummy"
-        },
-        "client_parameters": {
-            "model": "meta-llama/Llama-3.1-8B-Instruct",
-            "backend": "vllm",
-            "dataset_name": "sharegpt",
-            "dataset_path": "./ShareGPT_V3_unfiltered_cleaned_split.json",
-            "num_prompts": 200
-        }
-    },
-    {
-        "test_name": "serving_llama8B_bf16_pp3_sharegpt",
+        "test_name": "serving_llama8B_pp3_sharegpt",
         "qps_list": ["inf"],
         "max_concurrency_list": [12, 16, 24, 32, 64, 128, 200],
         "server_environment_variables": {
@@ -96,39 +64,7 @@
         }
     },
     {
-        "test_name": "serving_llama8B_bf16_tp4_sharegpt",
-        "qps_list": ["inf"],
-        "max_concurrency_list": [12, 16, 24, 32, 64, 128, 200],
-        "server_environment_variables": {
-            "VLLM_RPC_TIMEOUT": 100000,
-	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
-	    "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
-	    "VLLM_CPU_SGL_KERNEL": 1,
-	    "VLLM_CPU_KVCACHE_SPACE": 40
-        },
-        "server_parameters": {
-            "model": "meta-llama/Llama-3.1-8B-Instruct",
-            "tensor_parallel_size": 4,
-	    "dtype": "bfloat16",
-	    "distributed_executor_backend": "mp",
-	    "block_size": 128,
-	    "trust_remote_code": "",
-            "disable_log_stats": "",
-	    "enforce_eager": "",
-	    "max_num_batched_tokens": 2048,
-	    "max_num_seqs": 256,
-            "load_format": "dummy"
-        },
-        "client_parameters": {
-            "model": "meta-llama/Llama-3.1-8B-Instruct",
-            "backend": "vllm",
-            "dataset_name": "sharegpt",
-            "dataset_path": "./ShareGPT_V3_unfiltered_cleaned_split.json",
-            "num_prompts": 200
-        }
-    },
-    {
-        "test_name": "serving_llama8B_bf16_tp2pp3_sharegpt",
+        "test_name": "serving_llama8B_tp2pp3_sharegpt",
         "qps_list": ["inf"],
         "max_concurrency_list": [12, 16, 24, 32, 64, 128, 200],
         "server_environment_variables": {
@@ -161,7 +97,7 @@
         }
     },
     {
-        "test_name": "serving_llama8B_bf16_pp1_random_128_128",
+        "test_name": "serving_llama8B_pp1_random_128_128",
         "qps_list": ["inf"],
         "max_concurrency_list": [12, 16, 24, 32, 64, 128, 200, 1000],
         "server_environment_variables": {
@@ -196,42 +132,7 @@
         }
     },
     {
-        "test_name": "serving_llama8B_bf16_tp2_random_128_128",
-        "qps_list": ["inf"],
-        "max_concurrency_list": [12, 16, 24, 32, 64, 128, 200, 1000],
-        "server_environment_variables": {
-            "VLLM_RPC_TIMEOUT": 100000,
-	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
-	    "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
-	    "VLLM_CPU_SGL_KERNEL": 1,
-	    "VLLM_CPU_KVCACHE_SPACE": 40
-        },
-        "server_parameters": {
-            "model": "meta-llama/Llama-3.1-8B-Instruct",
-            "tensor_parallel_size": 2,
-	    "dtype": "bfloat16",
-	    "distributed_executor_backend": "mp",
-	    "block_size": 128,
-	    "trust_remote_code": "",
-	    "enable_chunked_prefill": "",
-            "disable_log_stats": "",
-	    "enforce_eager": "",
-	    "max_num_batched_tokens": 2048,
-	    "max_num_seqs": 256,
-            "load_format": "dummy"
-        },
-        "client_parameters": {
-            "model": "meta-llama/Llama-3.1-8B-Instruct",
-            "backend": "vllm",
-            "dataset_name": "random",
-	    "random-input-len": 128,
-	    "random-output-len": 128,
-	    "ignore-eos": "",
-            "num_prompts": 1000
-        }
-    },
-    {
-        "test_name": "serving_llama8B_bf16_pp3_random_128_128",
+        "test_name": "serving_llama8B_pp3_random_128_128",
         "qps_list": ["inf"],
         "max_concurrency_list": [12, 16, 24, 32, 64, 128, 200, 1000],
         "server_environment_variables": {
@@ -266,42 +167,7 @@
         }
     },
     {
-        "test_name": "serving_llama8B_bf16_tp4_random_128_128",
-        "qps_list": ["inf"],
-        "max_concurrency_list": [12, 16, 24, 32, 64, 128, 200, 1000],
-        "server_environment_variables": {
-            "VLLM_RPC_TIMEOUT": 100000,
-	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
-	    "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
-	    "VLLM_CPU_SGL_KERNEL": 1,
-	    "VLLM_CPU_KVCACHE_SPACE": 40
-        },
-        "server_parameters": {
-            "model": "meta-llama/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": "",
-	    "enforce_eager": "",
-	    "max_num_batched_tokens": 2048,
-	    "max_num_seqs": 256,
-            "load_format": "dummy"
-        },
-        "client_parameters": {
-            "model": "meta-llama/Llama-3.1-8B-Instruct",
-            "backend": "vllm",
-            "dataset_name": "random",
-	    "random-input-len": 128,
-	    "random-output-len": 128,
-	    "ignore-eos": "",
-            "num_prompts": 1000
-        }
-    },
-    {
-        "test_name": "serving_llama8B_bf16_tp2pp3_random_128_128",
+        "test_name": "serving_llama8B_tp2pp3_random_128_128",
         "qps_list": ["inf"],
         "max_concurrency_list": [12, 16, 24, 32, 64, 128, 200, 1000],
         "server_environment_variables": {
@@ -335,689 +201,5 @@
 	    "ignore-eos": "",
             "num_prompts": 1000
         }
-    },
-    {
-        "test_name": "serving_llama8B_int8_pp1_sharegpt",
-        "qps_list": ["inf"],
-        "max_concurrency_list": [12, 16, 24, 32, 64, 128, 200],
-        "server_environment_variables": {
-            "VLLM_RPC_TIMEOUT": 100000,
-	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
-	    "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
-	    "VLLM_CPU_SGL_KERNEL": 1,
-	    "VLLM_CPU_KVCACHE_SPACE": 40
-        },
-        "server_parameters": {
-            "model": "RedHatAI/Meta-Llama-3.1-8B-Instruct-quantized.w8a8",
-            "pipeline_parallel_size": 1,
-	    "dtype": "bfloat16",
-	    "distributed_executor_backend": "mp",
-	    "block_size": 128,
-	    "trust_remote_code": "",
-            "disable_log_stats": "",
-	    "enforce_eager": "",
-	    "max_num_batched_tokens": 2048,
-	    "max_num_seqs": 256,
-            "load_format": "dummy"
-        },
-        "client_parameters": {
-            "model": "RedHatAI/Meta-Llama-3.1-8B-Instruct-quantized.w8a8",
-            "backend": "vllm",
-            "dataset_name": "sharegpt",
-            "dataset_path": "./ShareGPT_V3_unfiltered_cleaned_split.json",
-            "num_prompts": 200
-        }
-    },
-    {
-        "test_name": "serving_llama8B_int8_tp2_sharegpt",
-        "qps_list": ["inf"],
-        "max_concurrency_list": [12, 16, 24, 32, 64, 128, 200],
-        "server_environment_variables": {
-            "VLLM_RPC_TIMEOUT": 100000,
-	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
-	    "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
-	    "VLLM_CPU_SGL_KERNEL": 1,
-	    "VLLM_CPU_KVCACHE_SPACE": 40
-        },
-        "server_parameters": {
-            "model": "RedHatAI/Meta-Llama-3.1-8B-Instruct-quantized.w8a8",
-            "tensor_parallel_size": 2,
-	    "dtype": "bfloat16",
-	    "distributed_executor_backend": "mp",
-	    "block_size": 128,
-	    "trust_remote_code": "",
-            "disable_log_stats": "",
-	    "enforce_eager": "",
-	    "max_num_batched_tokens": 2048,
-	    "max_num_seqs": 256,
-            "load_format": "dummy"
-        },
-        "client_parameters": {
-            "model": "RedHatAI/Meta-Llama-3.1-8B-Instruct-quantized.w8a8",
-            "backend": "vllm",
-            "dataset_name": "sharegpt",
-            "dataset_path": "./ShareGPT_V3_unfiltered_cleaned_split.json",
-            "num_prompts": 200
-        }
-    },
-    {
-        "test_name": "serving_llama8B_int8_pp3_sharegpt",
-        "qps_list": ["inf"],
-        "max_concurrency_list": [12, 16, 24, 32, 64, 128, 200],
-        "server_environment_variables": {
-            "VLLM_RPC_TIMEOUT": 100000,
-	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
-	    "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
-	    "VLLM_CPU_SGL_KERNEL": 1,
-	    "VLLM_CPU_KVCACHE_SPACE": 40
-        },
-        "server_parameters": {
-            "model": "RedHatAI/Meta-Llama-3.1-8B-Instruct-quantized.w8a8",
-            "pipeline_parallel_size": 3,
-	    "dtype": "bfloat16",
-	    "distributed_executor_backend": "mp",
-	    "block_size": 128,
-	    "trust_remote_code": "",
-            "disable_log_stats": "",
-	    "enforce_eager": "",
-	    "max_num_batched_tokens": 2048,
-	    "max_num_seqs": 256,
-            "load_format": "dummy"
-        },
-        "client_parameters": {
-            "model": "RedHatAI/Meta-Llama-3.1-8B-Instruct-quantized.w8a8",
-            "backend": "vllm",
-            "dataset_name": "sharegpt",
-            "dataset_path": "./ShareGPT_V3_unfiltered_cleaned_split.json",
-            "num_prompts": 200
-        }
-    },
-    {
-        "test_name": "serving_llama8B_int8_tp4_sharegpt",
-        "qps_list": ["inf"],
-        "max_concurrency_list": [12, 16, 24, 32, 64, 128, 200],
-        "server_environment_variables": {
-            "VLLM_RPC_TIMEOUT": 100000,
-	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
-	    "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
-	    "VLLM_CPU_SGL_KERNEL": 1,
-	    "VLLM_CPU_KVCACHE_SPACE": 40
-        },
-        "server_parameters": {
-            "model": "RedHatAI/Meta-Llama-3.1-8B-Instruct-quantized.w8a8",
-            "tensor_parallel_size": 4,
-	    "dtype": "bfloat16",
-	    "distributed_executor_backend": "mp",
-	    "block_size": 128,
-	    "trust_remote_code": "",
-            "disable_log_stats": "",
-	    "enforce_eager": "",
-	    "max_num_batched_tokens": 2048,
-	    "max_num_seqs": 256,
-            "load_format": "dummy"
-        },
-        "client_parameters": {
-            "model": "RedHatAI/Meta-Llama-3.1-8B-Instruct-quantized.w8a8",
-            "backend": "vllm",
-            "dataset_name": "sharegpt",
-            "dataset_path": "./ShareGPT_V3_unfiltered_cleaned_split.json",
-            "num_prompts": 200
-        }
-    },
-    {
-        "test_name": "serving_llama8B_int8_tp2pp3_sharegpt",
-        "qps_list": ["inf"],
-        "max_concurrency_list": [12, 16, 24, 32, 64, 128, 200],
-        "server_environment_variables": {
-            "VLLM_RPC_TIMEOUT": 100000,
-	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
-	    "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
-	    "VLLM_CPU_SGL_KERNEL": 1,
-	    "VLLM_CPU_KVCACHE_SPACE": 40
-        },
-        "server_parameters": {
-            "model": "RedHatAI/Meta-Llama-3.1-8B-Instruct-quantized.w8a8",
-            "tensor_parallel_size": 2,
-            "pipeline_parallel_size": 3,
-	    "dtype": "bfloat16",
-	    "distributed_executor_backend": "mp",
-	    "block_size": 128,
-	    "trust_remote_code": "",
-            "disable_log_stats": "",
-	    "enforce_eager": "",
-	    "max_num_batched_tokens": 2048,
-	    "max_num_seqs": 256,
-            "load_format": "dummy"
-        },
-        "client_parameters": {
-            "model": "RedHatAI/Meta-Llama-3.1-8B-Instruct-quantized.w8a8",
-            "backend": "vllm",
-            "dataset_name": "sharegpt",
-            "dataset_path": "./ShareGPT_V3_unfiltered_cleaned_split.json",
-            "num_prompts": 200
-        }
-    },
-    {
-        "test_name": "serving_llama8B_int8_pp1_random_128_128",
-        "qps_list": ["inf"],
-        "max_concurrency_list": [12, 16, 24, 32, 64, 128, 200, 1000],
-        "server_environment_variables": {
-            "VLLM_RPC_TIMEOUT": 100000,
-	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
-	    "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
-	    "VLLM_CPU_SGL_KERNEL": 1,
-	    "VLLM_CPU_KVCACHE_SPACE": 40
-        },
-        "server_parameters": {
-            "model": "RedHatAI/Meta-Llama-3.1-8B-Instruct-quantized.w8a8",
-            "pipeline_parallel_size": 1,
-	    "dtype": "bfloat16",
-	    "distributed_executor_backend": "mp",
-	    "block_size": 128,
-	    "trust_remote_code": "",
-	    "enable_chunked_prefill": "",
-            "disable_log_stats": "",
-	    "enforce_eager": "",
-	    "max_num_batched_tokens": 2048,
-	    "max_num_seqs": 256,
-            "load_format": "dummy"
-        },
-        "client_parameters": {
-            "model": "RedHatAI/Meta-Llama-3.1-8B-Instruct-quantized.w8a8",
-            "backend": "vllm",
-            "dataset_name": "random",
-	    "random-input-len": 128,
-	    "random-output-len": 128,
-	    "ignore-eos": "",
-            "num_prompts": 1000
-        }
-    },
-    {
-        "test_name": "serving_llama8B_int8_tp2_random_128_128",
-        "qps_list": ["inf"],
-        "max_concurrency_list": [12, 16, 24, 32, 64, 128, 200, 1000],
-        "server_environment_variables": {
-            "VLLM_RPC_TIMEOUT": 100000,
-	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
-	    "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
-	    "VLLM_CPU_SGL_KERNEL": 1,
-	    "VLLM_CPU_KVCACHE_SPACE": 40
-        },
-        "server_parameters": {
-            "model": "RedHatAI/Meta-Llama-3.1-8B-Instruct-quantized.w8a8",
-            "tensor_parallel_size": 2,
-	    "dtype": "bfloat16",
-	    "distributed_executor_backend": "mp",
-	    "block_size": 128,
-	    "trust_remote_code": "",
-	    "enable_chunked_prefill": "",
-            "disable_log_stats": "",
-	    "enforce_eager": "",
-	    "max_num_batched_tokens": 2048,
-	    "max_num_seqs": 256,
-            "load_format": "dummy"
-        },
-        "client_parameters": {
-            "model": "RedHatAI/Meta-Llama-3.1-8B-Instruct-quantized.w8a8",
-            "backend": "vllm",
-            "dataset_name": "random",
-	    "random-input-len": 128,
-	    "random-output-len": 128,
-	    "ignore-eos": "",
-            "num_prompts": 1000
-        }
-    },
-    {
-        "test_name": "serving_llama8B_int8_pp3_random_128_128",
-        "qps_list": ["inf"],
-        "max_concurrency_list": [12, 16, 24, 32, 64, 128, 200, 1000],
-        "server_environment_variables": {
-            "VLLM_RPC_TIMEOUT": 100000,
-	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
-	    "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
-	    "VLLM_CPU_SGL_KERNEL": 1,
-	    "VLLM_CPU_KVCACHE_SPACE": 40
-        },
-        "server_parameters": {
-            "model": "RedHatAI/Meta-Llama-3.1-8B-Instruct-quantized.w8a8",
-            "pipeline_parallel_size": 3,
-	    "dtype": "bfloat16",
-	    "distributed_executor_backend": "mp",
-	    "block_size": 128,
-	    "trust_remote_code": "",
-	    "enable_chunked_prefill": "",
-            "disable_log_stats": "",
-	    "enforce_eager": "",
-	    "max_num_batched_tokens": 2048,
-	    "max_num_seqs": 256,
-            "load_format": "dummy"
-        },
-        "client_parameters": {
-            "model": "RedHatAI/Meta-Llama-3.1-8B-Instruct-quantized.w8a8",
-            "backend": "vllm",
-            "dataset_name": "random",
-	    "random-input-len": 128,
-	    "random-output-len": 128,
-	    "ignore-eos": "",
-            "num_prompts": 1000
-        }
-    },
-    {
-        "test_name": "serving_llama8B_int8_tp4_random_128_128",
-        "qps_list": ["inf"],
-        "max_concurrency_list": [12, 16, 24, 32, 64, 128, 200, 1000],
-        "server_environment_variables": {
-            "VLLM_RPC_TIMEOUT": 100000,
-	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
-	    "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
-	    "VLLM_CPU_SGL_KERNEL": 1,
-	    "VLLM_CPU_KVCACHE_SPACE": 40
-        },
-        "server_parameters": {
-            "model": "RedHatAI/Meta-Llama-3.1-8B-Instruct-quantized.w8a8",
-            "tensor_parallel_size": 4,
-	    "dtype": "bfloat16",
-	    "distributed_executor_backend": "mp",
-	    "block_size": 128,
-	    "trust_remote_code": "",
-	    "enable_chunked_prefill": "",
-            "disable_log_stats": "",
-	    "enforce_eager": "",
-	    "max_num_batched_tokens": 2048,
-	    "max_num_seqs": 256,
-            "load_format": "dummy"
-        },
-        "client_parameters": {
-            "model": "RedHatAI/Meta-Llama-3.1-8B-Instruct-quantized.w8a8",
-            "backend": "vllm",
-            "dataset_name": "random",
-	    "random-input-len": 128,
-	    "random-output-len": 128,
-	    "ignore-eos": "",
-            "num_prompts": 1000
-        }
-    },
-    {
-        "test_name": "serving_llama8B_int8_tp2pp3_random_128_128",
-        "qps_list": ["inf"],
-        "max_concurrency_list": [12, 16, 24, 32, 64, 128, 200, 1000],
-        "server_environment_variables": {
-            "VLLM_RPC_TIMEOUT": 100000,
-	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
-	    "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
-	    "VLLM_CPU_SGL_KERNEL": 1,
-	    "VLLM_CPU_KVCACHE_SPACE": 40
-        },
-        "server_parameters": {
-            "model": "RedHatAI/Meta-Llama-3.1-8B-Instruct-quantized.w8a8",
-            "tensor_parallel_size": 2,
-            "pipeline_parallel_size": 3,
-	    "dtype": "bfloat16",
-	    "distributed_executor_backend": "mp",
-	    "block_size": 128,
-	    "trust_remote_code": "",
-	    "enable_chunked_prefill": "",
-            "disable_log_stats": "",
-	    "enforce_eager": "",
-	    "max_num_batched_tokens": 2048,
-	    "max_num_seqs": 256,
-            "load_format": "dummy"
-        },
-        "client_parameters": {
-            "model": "RedHatAI/Meta-Llama-3.1-8B-Instruct-quantized.w8a8",
-            "backend": "vllm",
-            "dataset_name": "random",
-	    "random-input-len": 128,
-	    "random-output-len": 128,
-	    "ignore-eos": "",
-            "num_prompts": 1000
-        }
-    },
-    {
-        "test_name": "serving_llama8B_int4_pp1_sharegpt",
-        "qps_list": ["inf"],
-        "max_concurrency_list": [12, 16, 24, 32, 64, 128, 200],
-        "server_environment_variables": {
-            "VLLM_RPC_TIMEOUT": 100000,
-	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
-	    "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
-	    "VLLM_CPU_SGL_KERNEL": 1,
-	    "VLLM_CPU_KVCACHE_SPACE": 40
-        },
-        "server_parameters": {
-            "model": "hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4",
-	    "quantization": "awq",
-            "pipeline_parallel_size": 1,
-	    "dtype": "bfloat16",
-	    "distributed_executor_backend": "mp",
-	    "block_size": 128,
-	    "trust_remote_code": "",
-            "disable_log_stats": "",
-	    "enforce_eager": "",
-	    "max_num_batched_tokens": 2048,
-	    "max_num_seqs": 256,
-            "load_format": "dummy"
-        },
-        "client_parameters": {
-            "model": "hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4",
-            "backend": "vllm",
-            "dataset_name": "sharegpt",
-            "dataset_path": "./ShareGPT_V3_unfiltered_cleaned_split.json",
-            "num_prompts": 200
-        }
-    },
-    {
-        "test_name": "serving_llama8B_int4_tp2_sharegpt",
-        "qps_list": ["inf"],
-        "max_concurrency_list": [12, 16, 24, 32, 64, 128, 200],
-        "server_environment_variables": {
-            "VLLM_RPC_TIMEOUT": 100000,
-	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
-	    "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
-	    "VLLM_CPU_SGL_KERNEL": 1,
-	    "VLLM_CPU_KVCACHE_SPACE": 40
-        },
-        "server_parameters": {
-            "model": "hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4",
-	    "quantization": "awq",
-            "tensor_parallel_size": 2,
-	    "dtype": "bfloat16",
-	    "distributed_executor_backend": "mp",
-	    "block_size": 128,
-	    "trust_remote_code": "",
-            "disable_log_stats": "",
-	    "enforce_eager": "",
-	    "max_num_batched_tokens": 2048,
-	    "max_num_seqs": 256,
-            "load_format": "dummy"
-        },
-        "client_parameters": {
-            "model": "hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4",
-            "backend": "vllm",
-            "dataset_name": "sharegpt",
-            "dataset_path": "./ShareGPT_V3_unfiltered_cleaned_split.json",
-            "num_prompts": 200
-        }
-    },
-    {
-        "test_name": "serving_llama8B_int4_pp3_sharegpt",
-        "qps_list": ["inf"],
-        "max_concurrency_list": [12, 16, 24, 32, 64, 128, 200],
-        "server_environment_variables": {
-            "VLLM_RPC_TIMEOUT": 100000,
-	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
-	    "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
-	    "VLLM_CPU_SGL_KERNEL": 1,
-	    "VLLM_CPU_KVCACHE_SPACE": 40
-        },
-        "server_parameters": {
-            "model": "hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4",
-	    "quantization": "awq",
-            "pipeline_parallel_size": 3,
-	    "dtype": "bfloat16",
-	    "distributed_executor_backend": "mp",
-	    "block_size": 128,
-	    "trust_remote_code": "",
-            "disable_log_stats": "",
-	    "enforce_eager": "",
-	    "max_num_batched_tokens": 2048,
-	    "max_num_seqs": 256,
-            "load_format": "dummy"
-        },
-        "client_parameters": {
-            "model": "hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4",
-            "backend": "vllm",
-            "dataset_name": "sharegpt",
-            "dataset_path": "./ShareGPT_V3_unfiltered_cleaned_split.json",
-            "num_prompts": 200
-        }
-    },
-    {
-        "test_name": "serving_llama8B_int4_tp4_sharegpt",
-        "qps_list": ["inf"],
-        "max_concurrency_list": [12, 16, 24, 32, 64, 128, 200],
-        "server_environment_variables": {
-            "VLLM_RPC_TIMEOUT": 100000,
-	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
-	    "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
-	    "VLLM_CPU_SGL_KERNEL": 1,
-	    "VLLM_CPU_KVCACHE_SPACE": 40
-        },
-        "server_parameters": {
-            "model": "hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4",
-	    "quantization": "awq",
-            "tensor_parallel_size": 4,
-	    "dtype": "bfloat16",
-	    "distributed_executor_backend": "mp",
-	    "block_size": 128,
-	    "trust_remote_code": "",
-            "disable_log_stats": "",
-	    "enforce_eager": "",
-	    "max_num_batched_tokens": 2048,
-	    "max_num_seqs": 256,
-            "load_format": "dummy"
-        },
-        "client_parameters": {
-            "model": "hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4",
-            "backend": "vllm",
-            "dataset_name": "sharegpt",
-            "dataset_path": "./ShareGPT_V3_unfiltered_cleaned_split.json",
-            "num_prompts": 200
-        }
-    },
-    {
-        "test_name": "serving_llama8B_int4_tp2pp3_sharegpt",
-        "qps_list": ["inf"],
-        "max_concurrency_list": [12, 16, 24, 32, 64, 128, 200],
-        "server_environment_variables": {
-            "VLLM_RPC_TIMEOUT": 100000,
-	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
-	    "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
-	    "VLLM_CPU_SGL_KERNEL": 1,
-	    "VLLM_CPU_KVCACHE_SPACE": 40
-        },
-        "server_parameters": {
-            "model": "hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4",
-	    "quantization": "awq",
-            "tensor_parallel_size": 2,
-            "pipeline_parallel_size": 3,
-	    "dtype": "bfloat16",
-	    "distributed_executor_backend": "mp",
-	    "block_size": 128,
-	    "trust_remote_code": "",
-            "disable_log_stats": "",
-	    "enforce_eager": "",
-	    "max_num_batched_tokens": 2048,
-	    "max_num_seqs": 256,
-            "load_format": "dummy"
-        },
-        "client_parameters": {
-            "model": "hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4",
-            "backend": "vllm",
-            "dataset_name": "sharegpt",
-            "dataset_path": "./ShareGPT_V3_unfiltered_cleaned_split.json",
-            "num_prompts": 200
-        }
-    },
-    {
-        "test_name": "serving_llama8B_int4_pp1_random_128_128",
-        "qps_list": ["inf"],
-        "max_concurrency_list": [12, 16, 24, 32, 64, 128, 200, 1000],
-        "server_environment_variables": {
-            "VLLM_RPC_TIMEOUT": 100000,
-	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
-	    "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
-	    "VLLM_CPU_SGL_KERNEL": 1,
-	    "VLLM_CPU_KVCACHE_SPACE": 40
-        },
-        "server_parameters": {
-            "model": "hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4",
-	    "quantization": "awq",
-            "pipeline_parallel_size": 1,
-	    "dtype": "bfloat16",
-	    "distributed_executor_backend": "mp",
-	    "block_size": 128,
-	    "trust_remote_code": "",
-	    "enable_chunked_prefill": "",
-            "disable_log_stats": "",
-	    "enforce_eager": "",
-	    "max_num_batched_tokens": 2048,
-	    "max_num_seqs": 256,
-            "load_format": "dummy"
-        },
-        "client_parameters": {
-            "model": "hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4",
-            "backend": "vllm",
-            "dataset_name": "random",
-	    "random-input-len": 128,
-	    "random-output-len": 128,
-	    "ignore-eos": "",
-            "num_prompts": 1000
-        }
-    },
-    {
-        "test_name": "serving_llama8B_int4_tp2_random_128_128",
-        "qps_list": ["inf"],
-        "max_concurrency_list": [12, 16, 24, 32, 64, 128, 200, 1000],
-        "server_environment_variables": {
-            "VLLM_RPC_TIMEOUT": 100000,
-	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
-	    "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
-	    "VLLM_CPU_SGL_KERNEL": 1,
-	    "VLLM_CPU_KVCACHE_SPACE": 40
-        },
-        "server_parameters": {
-            "model": "hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4",
-	    "quantization": "awq",
-            "tensor_parallel_size": 2,
-	    "dtype": "bfloat16",
-	    "distributed_executor_backend": "mp",
-	    "block_size": 128,
-	    "trust_remote_code": "",
-	    "enable_chunked_prefill": "",
-            "disable_log_stats": "",
-	    "enforce_eager": "",
-	    "max_num_batched_tokens": 2048,
-	    "max_num_seqs": 256,
-            "load_format": "dummy"
-        },
-        "client_parameters": {
-            "model": "hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4",
-            "backend": "vllm",
-            "dataset_name": "random",
-	    "random-input-len": 128,
-	    "random-output-len": 128,
-	    "ignore-eos": "",
-            "num_prompts": 1000
-        }
-    },
-    {
-        "test_name": "serving_llama8B_int4_pp3_random_128_128",
-        "qps_list": ["inf"],
-        "max_concurrency_list": [12, 16, 24, 32, 64, 128, 200, 1000],
-        "server_environment_variables": {
-            "VLLM_RPC_TIMEOUT": 100000,
-	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
-	    "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
-	    "VLLM_CPU_SGL_KERNEL": 1,
-	    "VLLM_CPU_KVCACHE_SPACE": 40
-        },
-        "server_parameters": {
-            "model": "hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4",
-	    "quantization": "awq",
-            "pipeline_parallel_size": 3,
-	    "dtype": "bfloat16",
-	    "distributed_executor_backend": "mp",
-	    "block_size": 128,
-	    "trust_remote_code": "",
-	    "enable_chunked_prefill": "",
-            "disable_log_stats": "",
-	    "enforce_eager": "",
-	    "max_num_batched_tokens": 2048,
-	    "max_num_seqs": 256,
-            "load_format": "dummy"
-        },
-        "client_parameters": {
-            "model": "hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4",
-            "backend": "vllm",
-            "dataset_name": "random",
-	    "random-input-len": 128,
-	    "random-output-len": 128,
-	    "ignore-eos": "",
-            "num_prompts": 1000
-        }
-    },
-    {
-        "test_name": "serving_llama8B_int4_tp4_random_128_128",
-        "qps_list": ["inf"],
-        "max_concurrency_list": [12, 16, 24, 32, 64, 128, 200, 1000],
-        "server_environment_variables": {
-            "VLLM_RPC_TIMEOUT": 100000,
-	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
-	    "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
-	    "VLLM_CPU_SGL_KERNEL": 1,
-	    "VLLM_CPU_KVCACHE_SPACE": 40
-        },
-        "server_parameters": {
-            "model": "hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4",
-	    "quantization": "awq",
-            "tensor_parallel_size": 4,
-	    "dtype": "bfloat16",
-	    "distributed_executor_backend": "mp",
-	    "block_size": 128,
-	    "trust_remote_code": "",
-	    "enable_chunked_prefill": "",
-            "disable_log_stats": "",
-	    "enforce_eager": "",
-	    "max_num_batched_tokens": 2048,
-	    "max_num_seqs": 256,
-            "load_format": "dummy"
-        },
-        "client_parameters": {
-            "model": "hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4",
-            "backend": "vllm",
-            "dataset_name": "random",
-	    "random-input-len": 128,
-	    "random-output-len": 128,
-	    "ignore-eos": "",
-            "num_prompts": 1000
-        }
-    },
-    {
-        "test_name": "serving_llama8B_int4_tp2pp3_random_128_128",
-        "qps_list": ["inf"],
-        "max_concurrency_list": [12, 16, 24, 32, 64, 128, 200, 1000],
-        "server_environment_variables": {
-            "VLLM_RPC_TIMEOUT": 100000,
-	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
-	    "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
-	    "VLLM_CPU_SGL_KERNEL": 1,
-	    "VLLM_CPU_KVCACHE_SPACE": 40
-        },
-        "server_parameters": {
-            "model": "hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4",
-	    "quantization": "awq",
-            "tensor_parallel_size": 2,
-            "pipeline_parallel_size": 3,
-	    "dtype": "bfloat16",
-	    "distributed_executor_backend": "mp",
-	    "block_size": 128,
-	    "trust_remote_code": "",
-	    "enable_chunked_prefill": "",
-            "disable_log_stats": "",
-	    "enforce_eager": "",
-	    "max_num_batched_tokens": 2048,
-	    "max_num_seqs": 256,
-            "load_format": "dummy"
-        },
-        "client_parameters": {
-            "model": "hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4",
-            "backend": "vllm",
-            "dataset_name": "random",
-	    "random-input-len": 128,
-	    "random-output-len": 128,
-	    "ignore-eos": "",
-            "num_prompts": 1000
-        }
     }
 ]

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

@@ -2,7 +2,7 @@
     {
         "test_name": "serving_llama8B_tp1_sharegpt",
         "qps_list": [1, 4, 16, "inf"],
-        "max_concurrency_list": [32],
+        "max_concurrency_list": [12, 16, 24, 32, 64, 128, 200],
         "server_environment_variables": {
             "VLLM_RPC_TIMEOUT": 100000,
 	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
@@ -28,13 +28,13 @@
             "backend": "vllm",
             "dataset_name": "sharegpt",
             "dataset_path": "./ShareGPT_V3_unfiltered_cleaned_split.json",
-            "num_prompts": 32
+            "num_prompts": 200
         }
     },
     {
         "test_name": "serving_llama8B_tp2_sharegpt",
         "qps_list": [1, 4, 16, "inf"],
-        "max_concurrency_list": [32],
+        "max_concurrency_list": [12, 16, 24, 32, 64, 128, 200],
         "server_environment_variables": {
             "VLLM_RPC_TIMEOUT": 100000,
 	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
@@ -60,13 +60,13 @@
             "backend": "vllm",
             "dataset_name": "sharegpt",
             "dataset_path": "./ShareGPT_V3_unfiltered_cleaned_split.json",
-            "num_prompts": 32
+            "num_prompts": 200
         }
     },
     {
-        "test_name": "serving_llama8B_tp1_random_128_128",
+        "test_name": "serving_llama8B_tp4_sharegpt",
         "qps_list": [1, 4, 16, "inf"],
-        "max_concurrency_list": [32],
+        "max_concurrency_list": [12, 16, 24, 32, 64, 128, 200],
         "server_environment_variables": {
             "VLLM_RPC_TIMEOUT": 100000,
 	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
@@ -76,7 +76,39 @@
         },
         "server_parameters": {
             "model": "meta-llama/Llama-3.1-8B-Instruct",
-            "tensor_parallel_size": 1,
+            "tensor_parallel_size": 4,
+	    "dtype": "bfloat16",
+	    "distributed_executor_backend": "mp",
+	    "block_size": 128,
+	    "trust_remote_code": "",
+            "disable_log_stats": "",
+	    "enforce_eager": "",
+	    "max_num_batched_tokens": 2048,
+	    "max_num_seqs": 256,
+            "load_format": "dummy"
+        },
+        "client_parameters": {
+            "model": "meta-llama/Llama-3.1-8B-Instruct",
+            "backend": "vllm",
+            "dataset_name": "sharegpt",
+            "dataset_path": "./ShareGPT_V3_unfiltered_cleaned_split.json",
+            "num_prompts": 200
+        }
+    },
+    {
+        "test_name": "serving_llama8B_tp4_random_1024_128",
+        "qps_list": [1, 4, 16, "inf"],
+        "max_concurrency_list": [12, 16, 24, 32, 64, 128, 200],
+        "server_environment_variables": {
+            "VLLM_RPC_TIMEOUT": 100000,
+	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
+	    "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
+	    "VLLM_CPU_SGL_KERNEL": 1,
+	    "VLLM_CPU_KVCACHE_SPACE": 40
+        },
+        "server_parameters": {
+            "model": "meta-llama/Llama-3.1-8B-Instruct",
+            "tensor_parallel_size": 4,
 	    "dtype": "bfloat16",
 	    "distributed_executor_backend": "mp",
 	    "block_size": 128,
@@ -92,16 +124,16 @@
             "model": "meta-llama/Llama-3.1-8B-Instruct",
             "backend": "vllm",
             "dataset_name": "random",
-	    "random-input-len": 128,
+	    "random-input-len": 1024,
 	    "random-output-len": 128,
 	    "ignore-eos": "",
-            "num_prompts": 32
+            "num_prompts": 100
         }
     },
     {
-        "test_name": "serving_llama8B_tp2_random_128_128",
+        "test_name": "serving_llama8B_pp6_random_1024_128",
         "qps_list": [1, 4, 16, "inf"],
-        "max_concurrency_list": [32],
+        "max_concurrency_list": [12, 16, 24, 32, 64, 128, 200],
         "server_environment_variables": {
             "VLLM_RPC_TIMEOUT": 100000,
 	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
@@ -111,7 +143,7 @@
         },
         "server_parameters": {
             "model": "meta-llama/Llama-3.1-8B-Instruct",
-            "tensor_parallel_size": 2,
+            "pipeline_parallel_size": 6,
 	    "dtype": "bfloat16",
 	    "distributed_executor_backend": "mp",
 	    "block_size": 128,
@@ -127,150 +159,10 @@
             "model": "meta-llama/Llama-3.1-8B-Instruct",
             "backend": "vllm",
             "dataset_name": "random",
-	    "random-input-len": 128,
+	    "random-input-len": 1024,
 	    "random-output-len": 128,
 	    "ignore-eos": "",
-            "num_prompts": 32
-        }
-    },
-    {
-        "test_name": "serving_llama8B_tp1_random_128_2048",
-        "qps_list": [1, 4, 16, "inf"],
-        "max_concurrency_list": [32],
-        "server_environment_variables": {
-            "VLLM_RPC_TIMEOUT": 100000,
-	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
-	    "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
-	    "VLLM_CPU_SGL_KERNEL": 1,
-	    "VLLM_CPU_KVCACHE_SPACE": 40
-        },
-        "server_parameters": {
-            "model": "meta-llama/Llama-3.1-8B-Instruct",
-            "tensor_parallel_size": 1,
-	    "dtype": "bfloat16",
-	    "distributed_executor_backend": "mp",
-	    "block_size": 128,
-	    "trust_remote_code": "",
-	    "enable_chunked_prefill": "",
-            "disable_log_stats": "",
-	    "enforce_eager": "",
-	    "max_num_batched_tokens": 2048,
-	    "max_num_seqs": 256,
-            "load_format": "dummy"
-        },
-        "client_parameters": {
-            "model": "meta-llama/Llama-3.1-8B-Instruct",
-            "backend": "vllm",
-            "dataset_name": "random",
-	    "random-input-len": 128,
-	    "random-output-len": 2048,
-	    "ignore-eos": "",
-            "num_prompts": 32
-        }
-    },
-    {
-        "test_name": "serving_llama8B_tp2_random_128_2048",
-        "qps_list": [1, 4, 16, "inf"],
-        "max_concurrency_list": [32],
-        "server_environment_variables": {
-            "VLLM_RPC_TIMEOUT": 100000,
-	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
-	    "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
-	    "VLLM_CPU_SGL_KERNEL": 1,
-	    "VLLM_CPU_KVCACHE_SPACE": 40
-        },
-        "server_parameters": {
-            "model": "meta-llama/Llama-3.1-8B-Instruct",
-            "tensor_parallel_size": 2,
-	    "dtype": "bfloat16",
-	    "distributed_executor_backend": "mp",
-	    "block_size": 128,
-	    "trust_remote_code": "",
-	    "enable_chunked_prefill": "",
-            "disable_log_stats": "",
-	    "enforce_eager": "",
-	    "max_num_batched_tokens": 2048,
-	    "max_num_seqs": 256,
-            "load_format": "dummy"
-        },
-        "client_parameters": {
-            "model": "meta-llama/Llama-3.1-8B-Instruct",
-            "backend": "vllm",
-            "dataset_name": "random",
-	    "random-input-len": 128,
-	    "random-output-len": 2048,
-	    "ignore-eos": "",
-            "num_prompts": 32
-        }
-    },
-    {
-        "test_name": "serving_llama8B_tp1_random_2048_128",
-        "qps_list": [1, 4, 16, "inf"],
-        "max_concurrency_list": [32],
-        "server_environment_variables": {
-            "VLLM_RPC_TIMEOUT": 100000,
-	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
-	    "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
-	    "VLLM_CPU_SGL_KERNEL": 1,
-	    "VLLM_CPU_KVCACHE_SPACE": 40
-        },
-        "server_parameters": {
-            "model": "meta-llama/Llama-3.1-8B-Instruct",
-            "tensor_parallel_size": 1,
-	    "dtype": "bfloat16",
-	    "distributed_executor_backend": "mp",
-	    "block_size": 128,
-	    "trust_remote_code": "",
-	    "enable_chunked_prefill": "",
-            "disable_log_stats": "",
-	    "enforce_eager": "",
-	    "max_num_batched_tokens": 2048,
-	    "max_num_seqs": 256,
-            "load_format": "dummy"
-        },
-        "client_parameters": {
-            "model": "meta-llama/Llama-3.1-8B-Instruct",
-            "backend": "vllm",
-            "dataset_name": "random",
-	    "random-input-len": 2048,
-	    "random-output-len": 128,
-	    "ignore-eos": "",
-            "num_prompts": 32
-        }
-    },
-    {
-        "test_name": "serving_llama8B_tp2_random_2048_128",
-        "qps_list": [1, 4, 16, "inf"],
-        "max_concurrency_list": [32],
-        "server_environment_variables": {
-            "VLLM_RPC_TIMEOUT": 100000,
-	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
-	    "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
-	    "VLLM_CPU_SGL_KERNEL": 1,
-	    "VLLM_CPU_KVCACHE_SPACE": 40
-        },
-        "server_parameters": {
-            "model": "meta-llama/Llama-3.1-8B-Instruct",
-            "tensor_parallel_size": 2,
-	    "dtype": "bfloat16",
-	    "distributed_executor_backend": "mp",
-	    "block_size": 128,
-	    "trust_remote_code": "",
-	    "enable_chunked_prefill": "",
-            "disable_log_stats": "",
-	    "enforce_eager": "",
-	    "max_num_batched_tokens": 2048,
-	    "max_num_seqs": 256,
-            "load_format": "dummy"
-        },
-        "client_parameters": {
-            "model": "meta-llama/Llama-3.1-8B-Instruct",
-            "backend": "vllm",
-            "dataset_name": "random",
-	    "random-input-len": 2048,
-	    "random-output-len": 128,
-	    "ignore-eos": "",
-            "num_prompts": 32
+            "num_prompts": 100
         }
     }
 ]

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

@@ -1,24 +1,29 @@
 [
     {
-        "test_name": "throughput_llama8B_tp2",
+        "test_name": "throughput_llama8B_tp1",
         "environment_variables": {
-            "VLLM_RPC_TIMEOUT": 100000,
 	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
-	    "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
-	    "VLLM_CPU_SGL_KERNEL": 1,
 	    "VLLM_CPU_KVCACHE_SPACE": 40
         },
         "parameters": {
             "model": "meta-llama/Llama-3.1-8B-Instruct",
-            "tensor_parallel_size": 2,
-	    "dtype": "bfloat16",
-	    "distributed_executor_backend": "mp",
-	    "block_size": 128,
-	    "trust_remote_code": "",
-            "disable_log_stats": "",
-	    "enforce_eager": "",
-	    "max_num_batched_tokens": 2048,
-	    "max_num_seqs": 256,
+            "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/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

@@ -1,28 +1,13 @@
 steps:
   # aarch64 + CUDA builds
-  - label: "Build arm64 wheel - CUDA 12.9"
-    depends_on: ~
-    id: build-wheel-arm64-cuda-12-9
+  - label: "Build arm64 wheel - CUDA 12.8"
+    id: build-wheel-arm64-cuda-12-8
     agents:
       queue: arm64_cpu_queue_postmerge
     commands:
       # #NOTE: torch_cuda_arch_list is derived from upstream PyTorch build files here:
       # https://github.com/pytorch/pytorch/blob/main/.ci/aarch64_linux/aarch64_ci_build.sh#L7
-      - "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=12.9.1 --build-arg VLLM_MAIN_CUDA_VERSION=12.9 --build-arg torch_cuda_arch_list='8.7 8.9 9.0 10.0+PTX 12.0' --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"
-
-  # aarch64 build
-  - label: "Build arm64 CPU wheel"
-    depends_on: ~
-    id: build-wheel-arm64-cpu
-    agents:
-      queue: arm64_cpu_queue_postmerge
-    commands:
-      - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg GIT_REPO_CHECK=1 --build-arg VLLM_BUILD_ACL=ON --tag vllm-ci:build-image --target vllm-build --progress plain -f docker/Dockerfile.cpu ."
+      - "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 torch_cuda_arch_list='8.7 9.0 10.0+PTX 12.0' --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"
@@ -31,7 +16,6 @@ steps:
 
   # x86 + CUDA builds
   - label: "Build wheel - CUDA 12.8"
-    depends_on: ~
     id: build-wheel-cuda-12-8
     agents:
       queue: cpu_queue_postmerge
@@ -43,33 +27,41 @@ steps:
     env:
       DOCKER_BUILDKIT: "1"
 
-  - label: "Build wheel - CUDA 12.9"
+  - block: "Build CUDA 12.6 wheel"
+    key: block-build-cu126-wheel
     depends_on: ~
-    id: build-wheel-cuda-12-9
+
+  - label: "Build wheel - CUDA 12.6"
+    depends_on: block-build-cu126-wheel
+    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=12.9.1 --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"
     env:
       DOCKER_BUILDKIT: "1"
 
-  - label: "Build wheel - CUDA 13.0"
-    depends_on: ~
-    id: build-wheel-cuda-13-0
+  # Note(simon): We can always build CUDA 11.8 wheel to ensure the build is working.
+  # However, this block can be uncommented to save some compute hours.
+  # - block: "Build CUDA 11.8 wheel"
+  #   key: block-build-cu118-wheel
+
+  - label: "Build wheel - CUDA 11.8"
+    # depends_on: block-build-cu118-wheel
+    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=13.0.1 --build-arg BUILD_BASE_IMAGE=nvidia/cuda:13.0.1-devel-ubuntu22.04 --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"
     env:
       DOCKER_BUILDKIT: "1"
 
-  # Build release images (12.9)
   - label: "Build release image (x86)"
     depends_on: ~
     id: build-release-image-x86
@@ -77,7 +69,7 @@ steps:
       queue: cpu_queue_postmerge
     commands:
       - "aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin public.ecr.aws/q9t5s3a7"
-      - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=12.9.1 --build-arg FLASHINFER_AOT_COMPILE=true --build-arg INSTALL_KV_CONNECTORS=true --tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m) --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 FLASHINFER_AOT_COMPILE=true --build-arg INSTALL_KV_CONNECTORS=true --tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m) --target vllm-openai --progress plain -f docker/Dockerfile ."
       - "docker push public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)"
       # re-tag to default image tag and push, just in case arm64 build fails
       - "docker tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m) public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT"
@@ -90,7 +82,7 @@ steps:
       queue: arm64_cpu_queue_postmerge
     commands:
       - "aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin public.ecr.aws/q9t5s3a7"
-      - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=12.9.1 --build-arg FLASHINFER_AOT_COMPILE=true --build-arg torch_cuda_arch_list='8.7 8.9 9.0 10.0+PTX 12.0' --build-arg INSTALL_KV_CONNECTORS=true --tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m) --target vllm-openai --progress plain -f docker/Dockerfile ."
+      - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=12.8.1 --build-arg torch_cuda_arch_list='8.7 9.0 10.0+PTX 12.0' --build-arg INSTALL_KV_CONNECTORS=true --tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m) --target vllm-openai --progress plain -f docker/Dockerfile ."
       - "docker push public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)"
 
   # Add job to create multi-arch manifest
@@ -110,6 +102,8 @@ steps:
     depends_on:
       - create-multi-arch-manifest
       - build-wheel-cuda-12-8
+      - build-wheel-cuda-12-6
+      - build-wheel-cuda-11-8
     id: annotate-release-workflow
     agents:
       queue: cpu_queue_postmerge
@@ -156,46 +150,18 @@ steps:
     env:
       DOCKER_BUILDKIT: "1"
 
-  - block: "Build arm64 CPU release image"
-    key: block-arm64-cpu-release-image-build
+  - block: "Build Neuron release image"
+    key: block-neuron-release-image-build
     depends_on: ~
 
-  - label: "Build and publish arm64 CPU release image"
-    depends_on: block-arm64-cpu-release-image-build
+  - label: "Build and publish Neuron release image"
+    depends_on: block-neuron-release-image-build
     agents:
-      queue: arm64_cpu_queue_postmerge
+      queue: neuron-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-arm64-cpu-release-repo:$(buildkite-agent meta-data get release-version) --tag public.ecr.aws/q9t5s3a7/vllm-arm64-cpu-release-repo:latest --progress plain --target vllm-openai -f docker/Dockerfile.cpu ."
-      - "docker push public.ecr.aws/q9t5s3a7/vllm-arm64-cpu-release-repo:latest"
-      - "docker push public.ecr.aws/q9t5s3a7/vllm-arm64-cpu-release-repo:$(buildkite-agent meta-data get release-version)"
+      - "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"
-
-  - label: "Build and publish nightly multi-arch image to DockerHub"
-    depends_on:
-      - create-multi-arch-manifest
-    if: build.env("NIGHTLY") == "1"
-    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 pull public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-x86_64"
-      - "docker pull public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-aarch64"
-      - "docker tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-x86_64 vllm/vllm-openai:nightly-x86_64"
-      - "docker tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-aarch64 vllm/vllm-openai:nightly-aarch64"
-      - "docker push vllm/vllm-openai:nightly-x86_64"
-      - "docker push vllm/vllm-openai:nightly-aarch64"
-      - "docker manifest create vllm/vllm-openai:nightly vllm/vllm-openai:nightly-x86_64 vllm/vllm-openai:nightly-aarch64 --amend"
-      - "docker manifest create vllm/vllm-openai:nightly-$BUILDKITE_COMMIT vllm/vllm-openai:nightly-x86_64 vllm/vllm-openai:nightly-aarch64 --amend"
-      - "docker manifest push vllm/vllm-openai:nightly"
-      - "docker manifest push vllm/vllm-openai:nightly-$BUILDKITE_COMMIT"
-      # Clean up old nightly builds (keep only last 14)
-      - "bash .buildkite/scripts/cleanup-nightly-builds.sh"
-    plugins:
-      - docker-login#v3.0.0:
-          username: vllmbot
-          password-env: DOCKERHUB_TOKEN
-    env:
-      DOCKER_BUILDKIT: "1"
-      DOCKERHUB_USERNAME: "vllmbot"

.buildkite/scripts/annotate-release.sh

@@ -14,33 +14,18 @@ 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/${RELEASE_VERSION}/vllm-${RELEASE_VERSION}-cp38-abi3-manylinux2014_aarch64.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/${RELEASE_VERSION}+cu129/vllm-${RELEASE_VERSION}+cu129-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:${BUILDKITE_COMMIT}-x86_64
-docker pull public.ecr.aws/q9t5s3a7/vllm-release-repo:${BUILDKITE_COMMIT}-aarch64
-
-docker tag public.ecr.aws/q9t5s3a7/vllm-release-repo:${BUILDKITE_COMMIT}-x86_64 vllm/vllm-openai:x86_64
-docker tag vllm/vllm-openai:x86_64 vllm/vllm-openai:latest-x86_64
-docker tag vllm/vllm-openai:x86_64 vllm/vllm-openai:v${RELEASE_VERSION}-x86_64
-docker push vllm/vllm-openai:latest-x86_64
-docker push vllm/vllm-openai:v${RELEASE_VERSION}-x86_64
-
-docker tag public.ecr.aws/q9t5s3a7/vllm-release-repo:${BUILDKITE_COMMIT}-aarch64 vllm/vllm-openai:aarch64
-docker tag vllm/vllm-openai:aarch64 vllm/vllm-openai:latest-aarch64
-docker tag vllm/vllm-openai:aarch64 vllm/vllm-openai:v${RELEASE_VERSION}-aarch64
-docker push vllm/vllm-openai:latest-aarch64
-docker push vllm/vllm-openai:v${RELEASE_VERSION}-aarch64
-
-docker manifest create vllm/vllm-openai:latest vllm/vllm-openai:latest-x86_64 vllm/vllm-openai:latest-aarch64 --amend
-docker manifest create vllm/vllm-openai:v${RELEASE_VERSION} vllm/vllm-openai:v${RELEASE_VERSION}-x86_64 vllm/vllm-openai:v${RELEASE_VERSION}-aarch64 --amend
-docker manifest push vllm/vllm-openai:latest
-docker manifest push vllm/vllm-openai:v${RELEASE_VERSION}
+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/cleanup-nightly-builds.sh

@@ -1,120 +0,0 @@
-#!/bin/bash
-
-set -ex
-
-# Clean up old nightly builds from DockerHub, keeping only the last 14 builds
-# This script uses DockerHub API to list and delete old tags with "nightly-" prefix
-
-# DockerHub API endpoint for vllm/vllm-openai repository
-REPO_API_URL="https://hub.docker.com/v2/repositories/vllm/vllm-openai/tags"
-
-# Get DockerHub credentials from environment
-if [ -z "$DOCKERHUB_TOKEN" ]; then
-    echo "Error: DOCKERHUB_TOKEN environment variable is not set"
-    exit 1
-fi
-
-if [ -z "$DOCKERHUB_USERNAME" ]; then
-    echo "Error: DOCKERHUB_USERNAME environment variable is not set"
-    exit 1
-fi
-
-# Get DockerHub bearer token
-echo "Getting DockerHub bearer token..."
-set +x
-BEARER_TOKEN=$(curl -s -X POST \
-    -H "Content-Type: application/json" \
-    -d "{\"username\": \"$DOCKERHUB_USERNAME\", \"password\": \"$DOCKERHUB_TOKEN\"}" \
-    "https://hub.docker.com/v2/users/login" | jq -r '.token')
-set -x
-
-if [ -z "$BEARER_TOKEN" ] || [ "$BEARER_TOKEN" = "null" ]; then
-    echo "Error: Failed to get DockerHub bearer token"
-    exit 1
-fi
-
-# Function to get all tags from DockerHub
-get_all_tags() {
-    local page=1
-    local all_tags=""
-    
-    while true; do
-        set +x
-        local response=$(curl -s -H "Authorization: Bearer $BEARER_TOKEN" \
-            "$REPO_API_URL?page=$page&page_size=100")
-        set -x
-        
-        # Get both last_updated timestamp and tag name, separated by |
-        local tags=$(echo "$response" | jq -r '.results[] | select(.name | startswith("nightly-")) | "\(.last_updated)|\(.name)"')
-        
-        if [ -z "$tags" ]; then
-            break
-        fi
-        
-        all_tags="$all_tags$tags"$'\n'
-        page=$((page + 1))
-    done
-    
-    # Sort by timestamp (newest first) and extract just the tag names
-    echo "$all_tags" | sort -r | cut -d'|' -f2
-}
-
-delete_tag() {
-    local tag_name="$1"
-    echo "Deleting tag: $tag_name"
-    
-    local delete_url="https://hub.docker.com/v2/repositories/vllm/vllm-openai/tags/$tag_name"
-    set +x
-    local response=$(curl -s -X DELETE -H "Authorization: Bearer $BEARER_TOKEN" "$delete_url")
-    set -x
-    
-    if echo "$response" | jq -e '.detail' > /dev/null 2>&1; then
-        echo "Warning: Failed to delete tag $tag_name: $(echo "$response" | jq -r '.detail')"
-    else
-        echo "Successfully deleted tag: $tag_name"
-    fi
-}
-
-# Get all nightly- prefixed tags, sorted by last_updated timestamp (newest first)
-echo "Fetching all tags from DockerHub..."
-all_tags=$(get_all_tags)
-
-if [ -z "$all_tags" ]; then
-    echo "No tags found to clean up"
-    exit 0
-fi
-
-# Count total tags
-total_tags=$(echo "$all_tags" | wc -l)
-echo "Found $total_tags tags"
-
-# Keep only the last 14 builds (including the current one)
-tags_to_keep=14
-tags_to_delete=$((total_tags - tags_to_keep))
-
-if [ $tags_to_delete -le 0 ]; then
-    echo "No tags need to be deleted (only $total_tags tags found, keeping $tags_to_keep)"
-    exit 0
-fi
-
-echo "Will delete $tags_to_delete old tags, keeping the newest $tags_to_keep"
-
-# Get tags to delete (skip the first $tags_to_keep tags)
-tags_to_delete_list=$(echo "$all_tags" | tail -n +$((tags_to_keep + 1)))
-
-if [ -z "$tags_to_delete_list" ]; then
-    echo "No tags to delete"
-    exit 0
-fi
-
-# Delete old tags
-echo "Deleting old tags..."
-while IFS= read -r tag; do
-    if [ -n "$tag" ]; then
-        delete_tag "$tag"
-        # Add a small delay to avoid rate limiting
-        sleep 1
-    fi
-done <<< "$tags_to_delete_list"
-
-echo "Cleanup completed successfully"

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

@@ -86,6 +86,10 @@ 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
@@ -163,6 +167,12 @@ if [[ $commands == *" entrypoints/llm "* ]]; then
   --ignore=entrypoints/llm/test_prompt_validation.py "}
 fi
 
+#Obsolete currently
+##ignore certain Entrypoints/llm tests
+#if [[ $commands == *" && pytest -v -s entrypoints/llm/test_guided_generate.py"* ]]; then
+#  commands=${commands//" && pytest -v -s entrypoints/llm/test_guided_generate.py"/" "}
+#fi
+
 # --ignore=entrypoints/openai/test_encoder_decoder.py \
 # --ignore=entrypoints/openai/test_embedding.py \
 # --ignore=entrypoints/openai/test_oot_registration.py

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

@@ -25,28 +25,25 @@ function cpu_tests() {
 
   # offline inference
   podman exec -it "$container_id" bash -c "
-    set -xve
-    python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m" >> $HOME/test_basic.log
+    set -e
+    python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m"
 
   # Run basic model test
   podman exec -it "$container_id" bash -c "
-    set -evx
+    set -e
     pip install pytest pytest-asyncio einops peft Pillow soundfile transformers_stream_generator matplotlib
     pip install sentence-transformers datamodel_code_generator
-
-    # Note: disable Bart until supports V1
-    # pytest -v -s tests/models/language/generation/test_bart.py -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]
-    # TODO: Below test case tests/models/language/pooling/test_embedding.py::test_models[True-ssmits/Qwen2-7B-Instruct-embed-base] fails on ppc64le. Disabling it for time being.
-    # pytest -v -s tests/models/language/pooling/test_embedding.py -m cpu_model" >> $HOME/test_rest.log
+    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.
 
 export container_id
 export -f cpu_tests
-timeout 120m bash -c cpu_tests
+timeout 40m bash -c cpu_tests
 

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

@@ -25,8 +25,8 @@ numactl -C "$CORE_RANGE" -N "$NUMA_NODE" docker build --tag cpu-test-"$NUMA_NODE
 numactl -C "$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 --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=16 --env VLLM_CPU_CI_ENV=1 -e E2E_OMP_THREADS="$OMP_CORE_RANGE" --shm-size=4g --name cpu-test-"$NUMA_NODE" cpu-test-"$NUMA_NODE"
-docker run -itd --cpuset-cpus="$CORE_RANGE" --cpuset-mems="$NUMA_NODE" --entrypoint /bin/bash -v ~/.cache/huggingface:/root/.cache/huggingface --privileged=true -e HF_TOKEN --env VLLM_CPU_KVCACHE_SPACE=16 --env VLLM_CPU_CI_ENV=1 -e E2E_OMP_THREADS="$OMP_CORE_RANGE" --shm-size=4g --name cpu-test-"$NUMA_NODE"-avx2 cpu-test-"$NUMA_NODE"-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 -e E2E_OMP_THREADS="$OMP_CORE_RANGE" --shm-size=4g --name cpu-test-"$NUMA_NODE" cpu-test-"$NUMA_NODE"
+docker run -itd --cpuset-cpus="$CORE_RANGE" --cpuset-mems="$NUMA_NODE" --entrypoint /bin/bash -v ~/.cache/huggingface:/root/.cache/huggingface --privileged=true -e HF_TOKEN --env VLLM_CPU_KVCACHE_SPACE=4 --env VLLM_CPU_CI_ENV=1 -e E2E_OMP_THREADS="$OMP_CORE_RANGE" --shm-size=4g --name cpu-test-"$NUMA_NODE"-avx2 cpu-test-"$NUMA_NODE"-avx2
 
 function cpu_tests() {
   set -e
@@ -58,11 +58,15 @@ function cpu_tests() {
     # pytest -x -v -s tests/kernels/attention/test_cache.py -m cpu_model
     # pytest -x -v -s tests/kernels/attention/test_mla_decode_cpu.py -m cpu_model
 
-    pytest -x -v -s tests/models/language/generation -m cpu_model
-    VLLM_CPU_SGL_KERNEL=1 pytest -x -v -s tests/models/language/generation -m cpu_model
+    # Note: disable Bart until supports V1
+    pytest -x -v -s tests/models/language/generation -m cpu_model \
+                --ignore=tests/models/language/generation/test_bart.py
+    VLLM_CPU_SGL_KERNEL=1 pytest -x -v -s tests/models/language/generation -m cpu_model \
+                --ignore=tests/models/language/generation/test_bart.py
 
     pytest -x -v -s tests/models/language/pooling -m cpu_model
     pytest -x -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"
 
@@ -70,7 +74,7 @@ function cpu_tests() {
   docker exec cpu-test-"$NUMA_NODE" bash -c "
     set -e
     pytest -x -s -v \
-    tests/quantization/test_compressed_tensors.py::test_compressed_tensors_w8a8_logprobs"
+    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
@@ -85,33 +89,17 @@ function cpu_tests() {
     pytest -x -s -v \
     tests/lora/test_qwen2vl.py"
 
-  # online serving: tp+pp
+  # online serving
   docker exec cpu-test-"$NUMA_NODE" bash -c '
     set -e
     VLLM_CPU_OMP_THREADS_BIND=$E2E_OMP_THREADS VLLM_CPU_SGL_KERNEL=1 vllm serve meta-llama/Llama-3.2-3B-Instruct -tp=2 -pp=2 &
-    server_pid=$!
     timeout 600 bash -c "until curl localhost:8000/v1/models; do sleep 1; done" || exit 1
     vllm bench serve \
       --backend vllm \
       --dataset-name random \
       --model meta-llama/Llama-3.2-3B-Instruct \
       --num-prompts 20 \
-      --endpoint /v1/completions
-    kill -s SIGTERM $server_pid &'
-
-  # online serving: tp+dp
-  docker exec cpu-test-"$NUMA_NODE" bash -c '
-    set -e
-    VLLM_CPU_OMP_THREADS_BIND=$E2E_OMP_THREADS VLLM_CPU_SGL_KERNEL=1 vllm serve meta-llama/Llama-3.2-3B-Instruct -tp=2 -dp=2 &
-    server_pid=$!
-    timeout 600 bash -c "until curl localhost:8000/v1/models; do sleep 1; done" || exit 1
-    vllm bench serve \
-      --backend vllm \
-      --dataset-name random \
-      --model meta-llama/Llama-3.2-3B-Instruct \
-      --num-prompts 20 \
-      --endpoint /v1/completions
-    kill -s SIGTERM $server_pid &'
+      --endpoint /v1/completions'
 }
 
 # All of CPU tests are expected to be finished less than 40 mins.

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

@@ -0,0 +1,64 @@
+#!/bin/bash
+
+# This script build the Neuron docker image and run the API server inside the container.
+# It serves a sanity check for compilation and basic model usage.
+set -e
+set -v
+
+image_name="neuron/vllm-ci"
+container_name="neuron_$(tr -dc A-Za-z0-9 < /dev/urandom | head -c 10; echo)"
+
+HF_CACHE="$(realpath ~)/huggingface"
+mkdir -p "${HF_CACHE}"
+HF_MOUNT="/root/.cache/huggingface"
+HF_TOKEN=$(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-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.
+if [ -f /tmp/neuron-docker-build-timestamp ]; then
+    last_build=$(cat /tmp/neuron-docker-build-timestamp)
+    current_time=$(date +%s)
+    if [ $((current_time - last_build)) -gt 86400 ]; then
+        # Remove dangling images (those that are not tagged and not used by any container)
+        docker image prune -f
+        # Remove unused volumes / force the system prune for old images as well.
+        docker volume prune -f && docker system prune -f
+        echo "$current_time" > /tmp/neuron-docker-build-timestamp
+    fi
+else
+    date "+%s" > /tmp/neuron-docker-build-timestamp
+fi
+
+docker build -t "${image_name}" -f docker/Dockerfile.neuron .
+
+# Setup cleanup
+remove_docker_container() {
+    docker image rm -f "${image_name}" || true;
+}
+trap remove_docker_container EXIT
+
+# Run the image
+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 "
+            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-npu-test.sh

@@ -1,191 +0,0 @@
-#!/bin/bash
-
-# This script build the Ascend NPU docker image and run the offline inference inside the container.
-# It serves a sanity check for compilation and basic model usage.
-set -ex
-
-# Base ubuntu image with basic ascend development libraries and python installed
-VLLM_ASCEND_REPO="https://github.com/vllm-project/vllm-ascend.git"
-CONFIG_FILE_REMOTE_PATH="tests/e2e/vllm_interface/vllm_test.cfg"
-TEST_RUN_CONFIG_FILE="vllm_test.cfg"
-VLLM_ASCEND_TMP_DIR=
-# Get the test run configuration file from the vllm-ascend repository
-fetch_vllm_test_cfg() {
-    VLLM_ASCEND_TMP_DIR=$(mktemp -d)
-    # Ensure that the temporary directory is cleaned up when an exception occurs during configuration file retrieval
-    cleanup() {
-        rm -rf "${VLLM_ASCEND_TMP_DIR}"
-    }
-    trap cleanup EXIT
-
-    GIT_TRACE=1 git clone -v --depth 1 "${VLLM_ASCEND_REPO}" "${VLLM_ASCEND_TMP_DIR}"
-    if [ ! -f "${VLLM_ASCEND_TMP_DIR}/${CONFIG_FILE_REMOTE_PATH}" ]; then
-        echo "Error: file '${CONFIG_FILE_REMOTE_PATH}' does not exist in the warehouse" >&2
-        exit 1
-    fi
-
-    # If the file already exists locally, just overwrite it
-    cp "${VLLM_ASCEND_TMP_DIR}/${CONFIG_FILE_REMOTE_PATH}" "${TEST_RUN_CONFIG_FILE}"
-    echo "Copied ${CONFIG_FILE_REMOTE_PATH} to ${TEST_RUN_CONFIG_FILE}"
-
-    # Since the trap will be overwritten later, and when it is executed here, the task of cleaning up resources
-    # when the trap is abnormal has been completed, so the temporary resources are manually deleted here.
-    rm -rf "${VLLM_ASCEND_TMP_DIR}"
-    trap - EXIT
-}
-
-# Downloads test run configuration file from a remote URL.
-# Loads the configuration into the current script environment.
-get_config() {
-    if [ ! -f "${TEST_RUN_CONFIG_FILE}" ]; then
-        echo "Error: file '${TEST_RUN_CONFIG_FILE}' does not exist in the warehouse" >&2
-        exit 1
-    fi
-    source "${TEST_RUN_CONFIG_FILE}"
-    echo "Base docker image name that get from configuration: ${BASE_IMAGE_NAME}"
-    return 0
-}
-
-# get test running configuration.
-fetch_vllm_test_cfg
-get_config
-# Check if the function call was successful. If not, exit the script.
-if [ $? -ne 0 ]; then
-  exit 1
-fi
-
-image_name="npu/vllm-ci:${BUILDKITE_COMMIT}_${EPOCHSECONDS}"
-container_name="npu_${BUILDKITE_COMMIT}_$(tr -dc A-Za-z0-9 < /dev/urandom | head -c 10; echo)"
-
-# BUILDKITE_AGENT_NAME format is {hostname}-{agent_idx}-{npu_card_num}cards
-agent_idx=$(echo "${BUILDKITE_AGENT_NAME}" | awk -F'-' '{print $(NF-1)}')
-echo "agent_idx: ${agent_idx}"
-builder_name="cachebuilder${agent_idx}"
-builder_cache_dir="/mnt/docker-cache${agent_idx}"
-mkdir -p ${builder_cache_dir}
-
-# Try building the docker image
-cat <<EOF | DOCKER_BUILDKIT=1 docker build \
-    --add-host cache-service-vllm.nginx-pypi-cache.svc.cluster.local:${PYPI_CACHE_HOST} \
-    --builder ${builder_name} --cache-from type=local,src=${builder_cache_dir} \
-                           --cache-to type=local,dest=${builder_cache_dir},mode=max \
-    --progress=plain --load -t ${image_name} -f - .
-FROM ${BASE_IMAGE_NAME}
-
-# Define environments
-ENV DEBIAN_FRONTEND=noninteractive
-
-RUN pip config set global.index-url http://cache-service-vllm.nginx-pypi-cache.svc.cluster.local:${PYPI_CACHE_PORT}/pypi/simple && \
-    pip config set global.trusted-host cache-service-vllm.nginx-pypi-cache.svc.cluster.local && \
-    apt-get update -y && \
-    apt-get install -y python3-pip git vim wget net-tools gcc g++ cmake libnuma-dev && \
-    rm -rf /var/cache/apt/* && \
-    rm -rf /var/lib/apt/lists/*
-
-# Install for pytest to make the docker build cache layer always valid
-RUN --mount=type=cache,target=/root/.cache/pip \
-    pip install pytest>=6.0  modelscope
-
-WORKDIR /workspace/vllm
-
-# Install vLLM dependencies in advance. Effect: As long as common.txt remains unchanged, the docker cache layer will be valid.
-COPY requirements/common.txt /workspace/vllm/requirements/common.txt
-RUN --mount=type=cache,target=/root/.cache/pip \
-    pip install -r requirements/common.txt
-
-COPY . .
-
-# Install vLLM
-RUN --mount=type=cache,target=/root/.cache/pip \
-    VLLM_TARGET_DEVICE="empty" python3 -m pip install -v -e /workspace/vllm/ --extra-index https://download.pytorch.org/whl/cpu/ && \
-    python3 -m pip uninstall -y triton
-
-# Install vllm-ascend
-WORKDIR /workspace
-ARG VLLM_ASCEND_REPO=https://github.com/vllm-project/vllm-ascend.git
-ARG VLLM_ASCEND_TAG=main
-RUN git config --global url."https://gh-proxy.test.osinfra.cn/https://github.com/".insteadOf "https://github.com/" && \
-    git clone --depth 1 \$VLLM_ASCEND_REPO --branch \$VLLM_ASCEND_TAG /workspace/vllm-ascend
-
-# Install vllm dependencies in advance. Effect: As long as common.txt remains unchanged, the docker cache layer will be valid.
-RUN --mount=type=cache,target=/root/.cache/pip \
-    pip install -r /workspace/vllm-ascend/requirements.txt
-
-RUN --mount=type=cache,target=/root/.cache/pip \
-    export PIP_EXTRA_INDEX_URL=https://mirrors.huaweicloud.com/ascend/repos/pypi && \
-    source /usr/local/Ascend/ascend-toolkit/set_env.sh && \
-    source /usr/local/Ascend/nnal/atb/set_env.sh && \
-    export LD_LIBRARY_PATH=\$LD_LIBRARY_PATH:/usr/local/Ascend/ascend-toolkit/latest/`uname -i`-linux/devlib && \
-    python3 -m pip install -v -e /workspace/vllm-ascend/ --extra-index https://download.pytorch.org/whl/cpu/
-
-ENV VLLM_WORKER_MULTIPROC_METHOD=spawn
-ENV VLLM_USE_MODELSCOPE=True
-
-WORKDIR /workspace/vllm-ascend
-
-CMD ["/bin/bash"]
-
-EOF
-
-# Setup cleanup
-remove_docker_container() {
-  docker rm -f "${container_name}" || true;
-  docker image rm -f "${image_name}" || true;
-  docker system prune -f || true;
-}
-trap remove_docker_container EXIT
-
-# Generate corresponding --device args based on BUILDKITE_AGENT_NAME
-# Ascend NPU BUILDKITE_AGENT_NAME format is {hostname}-{agent_idx}-{npu_card_num}cards, and agent_idx starts from 1.
-#   e.g. atlas-a2-001-1-2cards means this is the 1-th agent on atlas-a2-001 host, and it has 2 NPU cards.
-#   returns --device /dev/davinci0 --device /dev/davinci1
-parse_and_gen_devices() {
-    local input="$1"
-    local index cards_num
-    if [[ "$input" =~ ([0-9]+)-([0-9]+)cards$ ]]; then
-        index="${BASH_REMATCH[1]}"
-        cards_num="${BASH_REMATCH[2]}"
-    else
-        echo "parse error" >&2
-        return 1
-    fi
-
-    local devices=""
-    local i=0
-    while (( i < cards_num )); do
-        local dev_idx=$(((index - 1)*cards_num + i ))
-        devices="$devices --device /dev/davinci${dev_idx}"
-        ((i++))
-    done
-
-    # trim leading space
-    devices="${devices#"${devices%%[![:space:]]*}"}"
-    # Output devices: assigned to the caller variable
-    printf '%s' "$devices"
-}
-
-devices=$(parse_and_gen_devices "${BUILDKITE_AGENT_NAME}") || exit 1
-
-# Run the image and execute the Out-Of-Tree (OOT) platform interface test case on Ascend NPU hardware.
-# This test checks whether the OOT platform interface is functioning properly in conjunction with
-# the hardware plugin vllm-ascend.
-model_cache_dir=/mnt/modelscope${agent_idx}
-mkdir -p ${model_cache_dir}
-docker run \
-    ${devices} \
-    --device /dev/davinci_manager \
-    --device /dev/devmm_svm \
-    --device /dev/hisi_hdc \
-    -v /usr/local/dcmi:/usr/local/dcmi \
-    -v /usr/local/bin/npu-smi:/usr/local/bin/npu-smi \
-    -v /usr/local/Ascend/driver/lib64/:/usr/local/Ascend/driver/lib64/ \
-    -v /usr/local/Ascend/driver/version.info:/usr/local/Ascend/driver/version.info \
-    -v /etc/ascend_install.info:/etc/ascend_install.info \
-    -v ${model_cache_dir}:/root/.cache/modelscope \
-    --entrypoint="" \
-    --name "${container_name}" \
-    "${image_name}" \
-    bash -c '
-    set -e
-    pytest -v -s tests/e2e/vllm_interface/
-'

.buildkite/scripts/hardware_ci/run-tpu-v1-test-part2.sh

@@ -62,11 +62,12 @@ 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 @ git+https://github.com/EleutherAI/lm-evaluation-harness.git@206b7722158f58c35b7ffcd53b035fdbdda5126d" \
-    && python3 -m pip install --progress-bar off hf-transfer tblib==3.1.0
+    && python3 -m pip install --progress-bar off hf-transfer
 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

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

@@ -62,11 +62,12 @@ 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 @ git+https://github.com/EleutherAI/lm-evaluation-harness.git@206b7722158f58c35b7ffcd53b035fdbdda5126d" \
-    && python3 -m pip install --progress-bar off hf-transfer tblib==3.1.0
+    && python3 -m pip install --progress-bar off hf-transfer
 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

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

@@ -30,19 +30,19 @@ docker run \
     bash -c '
     set -e
     echo $ZE_AFFINITY_MASK
-    pip install tblib==3.1.0
-    python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m --block-size 64 --enforce-eager
-    python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m --block-size 64 -O3 -O.cudagraph_mode=NONE
-    python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m --block-size 64 --enforce-eager -tp 2 --distributed-executor-backend ray
-    python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m --block-size 64 --enforce-eager -tp 2 --distributed-executor-backend mp
-    VLLM_ATTENTION_BACKEND=TRITON_ATTN 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
+    VLLM_USE_V1=1 python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m --block-size 64 -O3 -O.cudagraph_mode=NONE
+    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
     pytest -v -s v1/engine
     pytest -v -s v1/sample --ignore=v1/sample/test_logprobs.py --ignore=v1/sample/test_logprobs_e2e.py
     pytest -v -s v1/worker --ignore=v1/worker/test_gpu_model_runner.py
     pytest -v -s v1/structured_output
-    pytest -v -s v1/spec_decode --ignore=v1/spec_decode/test_max_len.py --ignore=v1/spec_decode/test_tree_attention.py
+    pytest -v -s v1/spec_decode --ignore=v1/spec_decode/test_max_len.py --ignore=v1/spec_decode/test_eagle.py --ignore=v1/spec_decode/test_tree_attention.py
     pytest -v -s v1/kv_connector/unit --ignore=v1/kv_connector/unit/test_multi_connector.py --ignore=v1/kv_connector/unit/test_nixl_connector.py --ignore=v1/kv_connector/unit/test_shared_storage_connector.py
     pytest -v -s v1/test_serial_utils.py
+    pytest -v -s v1/test_utils.py
+    pytest -v -s v1/test_metrics_reader.py
 '

.buildkite/scripts/run-benchmarks.sh

@@ -18,7 +18,7 @@ vllm bench throughput --input-len 256 --output-len 256 --output-json throughput_
 bench_throughput_exit_code=$?
 
 # run server-based benchmarks and upload the result to buildkite
-vllm serve meta-llama/Llama-2-7b-chat-hf &
+python3 -m vllm.entrypoints.openai.api_server --model meta-llama/Llama-2-7b-chat-hf &
 server_pid=$!
 wget https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/resolve/main/ShareGPT_V3_unfiltered_cleaned_split.json
 

.buildkite/scripts/run-prime-rl-test.sh

@@ -1,59 +0,0 @@
-#!/bin/bash
-# SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-
-# Setup script for Prime-RL integration tests
-# This script prepares the environment for running Prime-RL tests with nightly vLLM
-
-set -euo pipefail
-
-SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
-REPO_ROOT="$(cd "${SCRIPT_DIR}/../.." && pwd)"
-PRIME_RL_REPO="https://github.com/PrimeIntellect-ai/prime-rl.git"
-PRIME_RL_DIR="${REPO_ROOT}/prime-rl"
-
-echo "Setting up Prime-RL integration test environment..."
-
-# Clean up any existing Prime-RL directory
-if [ -d "${PRIME_RL_DIR}" ]; then
-    echo "Removing existing Prime-RL directory..."
-    rm -rf "${PRIME_RL_DIR}"
-fi
-
-# Install UV if not available
-if ! command -v uv &> /dev/null; then
-    echo "Installing UV package manager..."
-    curl -LsSf https://astral.sh/uv/install.sh | sh
-    source $HOME/.local/bin/env
-fi
-
-# Clone Prime-RL repository at specific branch for reproducible tests
-PRIME_RL_BRANCH="integ-vllm-main"
-echo "Cloning Prime-RL repository at branch: ${PRIME_RL_BRANCH}..."
-git clone --branch "${PRIME_RL_BRANCH}" --single-branch "${PRIME_RL_REPO}" "${PRIME_RL_DIR}"
-cd "${PRIME_RL_DIR}"
-
-echo "Setting up UV project environment..."
-export UV_PROJECT_ENVIRONMENT=/usr/local
-ln -s /usr/bin/python3 /usr/local/bin/python
-
-# Remove vllm pin from pyproject.toml
-echo "Removing vllm pin from pyproject.toml..."
-sed -i '/vllm==/d' pyproject.toml
-
-# Sync Prime-RL dependencies
-echo "Installing Prime-RL dependencies..."
-uv sync --inexact && uv sync --inexact --all-extras
-
-# Verify installation
-echo "Verifying installations..."
-uv run python -c "import vllm; print(f'vLLM version: {vllm.__version__}')"
-uv run python -c "import prime_rl; print('Prime-RL imported successfully')"
-
-echo "Prime-RL integration test environment setup complete!"
-
-echo "Running Prime-RL integration tests..."
-export WANDB_MODE=offline # this makes this test not require a WANDB_API_KEY
-uv run pytest -vs tests/integration/test_rl.py -m gpu
-
-echo "Prime-RL integration tests completed!"

.buildkite/scripts/tpu/quantized_v6e_1.env

@@ -9,6 +9,6 @@ MAX_NUM_BATCHED_TOKENS=1024
 TENSOR_PARALLEL_SIZE=1
 MAX_MODEL_LEN=2048
 DOWNLOAD_DIR=/mnt/disks/persist
-EXPECTED_THROUGHPUT=8.7
+EXPECTED_THROUGHPUT=10.0
 INPUT_LEN=1800
 OUTPUT_LEN=128

.buildkite/scripts/tpu/run_bm.sh

@@ -42,7 +42,7 @@ echo "lanching vllm..."
 echo "logging to $VLLM_LOG"
 echo
 
-vllm serve $MODEL \
+VLLM_USE_V1=1 vllm serve $MODEL \
  --seed 42 \
  --max-num-seqs $MAX_NUM_SEQS \
  --max-num-batched-tokens $MAX_NUM_BATCHED_TOKENS \

.buildkite/scripts/upload-wheels.sh

@@ -58,25 +58,31 @@ python3 .buildkite/generate_index.py --wheel "$normal_wheel"
 aws s3 cp "$wheel" "s3://vllm-wheels/$BUILDKITE_COMMIT/"
 aws s3 cp "$normal_wheel" "s3://vllm-wheels/$BUILDKITE_COMMIT/"
 
-if [[ $normal_wheel == *"cu129"* ]]; then
-    # only upload index.html for cu129 wheels (default wheels) as it
-    # is available on both x86 and arm64
+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 == *"cu126"* ]]; then
+    # if $normal_wheel matches cu126, do not upload the index.html
+    echo "Skipping index files for cu126 wheels"
+else
+    # only upload index.html for cu128 wheels (default wheels)
     aws s3 cp index.html "s3://vllm-wheels/$BUILDKITE_COMMIT/vllm/index.html"
     aws s3 cp "s3://vllm-wheels/nightly/index.html" "s3://vllm-wheels/$BUILDKITE_COMMIT/index.html"
-else
-    echo "Skipping index files for non-cu129 wheels"
 fi
 
 # generate index for nightly
 aws s3 cp "$wheel" "s3://vllm-wheels/nightly/"
 aws s3 cp "$normal_wheel" "s3://vllm-wheels/nightly/"
 
-if [[ $normal_wheel == *"cu129"* ]]; then
-    # only upload index.html for cu129 wheels (default wheels) as it
-    # is available on both x86 and arm64
-    aws s3 cp index.html "s3://vllm-wheels/nightly/vllm/index.html"
+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 == *"cu126"* ]]; then
+    # if $normal_wheel matches cu126, do not upload the index.html
+    echo "Skipping index files for cu126 wheels"
 else
-    echo "Skipping index files for non-cu129 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/"

.coveragerc

@@ -1,47 +0,0 @@
-[run]
-# Track the installed vllm package (this is what actually gets imported during tests)
-# Use wildcard pattern to match the installed location
-source =
-    vllm
-    */dist-packages/vllm
-    */site-packages/vllm
-omit =
-    */tests/*
-    */test_*
-    */__pycache__/*
-    */build/*
-    */dist/*
-    */vllm.egg-info/*
-    */third_party/*
-    */examples/*
-    */benchmarks/*
-    */docs/*
-
-[paths]
-# Map all possible vllm locations to a canonical "vllm" path
-# This ensures coverage.combine properly merges data from different test runs
-source =
-    vllm
-    /vllm-workspace/src/vllm
-    /vllm-workspace/vllm
-    */site-packages/vllm
-    */dist-packages/vllm
-
-[report]
-exclude_lines =
-    pragma: no cover
-    def __repr__
-    if self.debug:
-    if settings.DEBUG
-    raise AssertionError
-    raise NotImplementedError
-    if 0:
-    if __name__ == .__main__.:
-    class .*\bProtocol\):
-    @(abc\.)?abstractmethod
-
-[html]
-directory = htmlcov
-
-[xml]
-output = coverage.xml

.git-blame-ignore-revs

@@ -1,4 +0,0 @@
-# Migrate from `yapf` & `isort` to `ruff`
-d6953beb91da4e9c99be4c0a1304a2d24189535c
-# Convert `Optional[x]` to `x | None` and `Union[x, y]` to `x | y`
-8fcaaf6a165e661f63fc51be906bc05b0767332f

.github/.bc-linter.yml

@@ -1,24 +0,0 @@
-# doc: https://github.com/pytorch/test-infra/blob/main/tools/stronghold/docs/bc_linter_config.md
-version: 1
-paths:
-# We temporarily disable globally, and will only enable with `annotations.include`
-# include:
-#   - "vllm/v1/attetion/*.py"
-#   - "vllm/v1/core/*.py"
-exclude:
-  - "**/*.py"
-
-scan:
-  functions: true        # check free functions and methods
-  classes: true          # check classes/dataclasses
-  public_only: true      # ignore names starting with "_" at any level
-
-annotations:
-  include:               # decorators that force‑include a symbol
-    - name: "bc_linter_include"  # matched by simple name or dotted suffix
-      propagate_to_members: false # for classes, include methods/inner classes
-  exclude:               # decorators that force‑exclude a symbol
-    - name: "bc_linter_skip"     # matched by simple name or dotted suffix
-      propagate_to_members: true  # for classes, exclude methods/inner classes
-
-excluded_violations: []  # e.g. ["ParameterRenamed", "FieldTypeChanged"]

.github/CODEOWNERS

@@ -2,86 +2,64 @@
 # for more info about CODEOWNERS file
 
 # This lists cover the "core" components of vLLM that require careful review
-/vllm/attention @LucasWilkinson
 /vllm/attention/backends/abstract.py @WoosukKwon @zhuohan123 @youkaichao @alexm-redhat @comaniac @njhill
-/vllm/executor/executor_base.py @zhuohan123 @youkaichao @alexm-redhat @comaniac @njhill @22quinn
-/vllm/model_executor/layers/fused_moe @mgoin @pavanimajety
-/vllm/model_executor/layers/quantization @mgoin @robertgshaw2-redhat @tlrmchlsmth @yewentao256 @pavanimajety
+/vllm/core @zhuohan123 @youkaichao @alexm-redhat @comaniac @njhill
+/vllm/engine/llm_engine.py @zhuohan123 @youkaichao @alexm-redhat @comaniac @njhill
+/vllm/executor/executor_base.py @zhuohan123 @youkaichao @alexm-redhat @comaniac @njhill
+/vllm/worker/worker_base.py @zhuohan123 @youkaichao @alexm-redhat @comaniac @njhill
+/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 @yewentao256
 /vllm/model_executor/layers/mamba @tdoublep
-/vllm/model_executor/model_loader @22quinn
-/vllm/multimodal @DarkLight1337 @ywang96 @NickLucche
+/vllm/multimodal @DarkLight1337 @ywang96
 /vllm/vllm_flash_attn @LucasWilkinson
 /vllm/lora @jeejeelee
-/vllm/reasoning @aarnphm @chaunceyjiang
-/vllm/entrypoints @aarnphm @chaunceyjiang
+/vllm/reasoning @aarnphm
+/vllm/entrypoints @aarnphm
 /vllm/compilation @zou3519 @youkaichao @ProExpertProg
-/vllm/distributed/kv_transfer @NickLucche @ApostaC
 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 @simon-mo @WoosukKwon @youkaichao @robertgshaw2-redhat @mgoin @tlrmchlsmth @houseroad @hmellor @yewentao256 @ProExpertProg
-/vllm/config/cache.py @simon-mo @WoosukKwon @youkaichao @robertgshaw2-redhat @mgoin @tlrmchlsmth @houseroad @hmellor @yewentao256 @ProExpertProg @heheda12345
 
 # vLLM V1
-/vllm/v1/attention @LucasWilkinson
-/vllm/v1/attention/backends/mla @pavanimajety
-/vllm/v1/attention/backends/flashinfer.py @mgoin @pavanimajety
+/vllm/v1 @WoosukKwon @robertgshaw2-redhat @njhill @ywang96 @comaniac @alexm-redhat
+/vllm/v1/structured_output @mgoin @russellb @aarnphm
 /vllm/v1/attention/backends/triton_attn.py @tdoublep
-/vllm/v1/core @WoosukKwon @robertgshaw2-redhat @njhill @ywang96 @comaniac @alexm-redhat @heheda12345 @ApostaC
-/vllm/v1/sample @22quinn @houseroad @njhill
-/vllm/v1/spec_decode @benchislett @luccafong
-/vllm/v1/structured_output @mgoin @russellb @aarnphm @benchislett
-/vllm/v1/kv_cache_interface.py @heheda12345
-/vllm/v1/offloading @ApostaC
 
 # Test ownership
 /.buildkite/lm-eval-harness @mgoin @simon-mo
+/tests/async_engine @njhill @robertgshaw2-redhat @simon-mo
 /tests/distributed/test_multi_node_assignment.py @youkaichao
 /tests/distributed/test_pipeline_parallel.py @youkaichao
 /tests/distributed/test_same_node.py @youkaichao
-/tests/entrypoints @DarkLight1337 @robertgshaw2-redhat @simon-mo @aarnphm @NickLucche
-/tests/evals @mgoin
-/tests/kernels @mgoin @tlrmchlsmth @WoosukKwon @yewentao256
+/tests/entrypoints @DarkLight1337 @robertgshaw2-redhat @simon-mo @aarnphm
+/tests/kernels @tlrmchlsmth @WoosukKwon @yewentao256
 /tests/models @DarkLight1337 @ywang96
-/tests/multimodal @DarkLight1337 @ywang96 @NickLucche
-/tests/quantization @mgoin @robertgshaw2-redhat @yewentao256 @pavanimajety
+/tests/multimodal @DarkLight1337 @ywang96
+/tests/prefix_caching @comaniac @KuntaiDu
+/tests/quantization @mgoin @robertgshaw2-redhat @yewentao256
 /tests/test_inputs.py @DarkLight1337 @ywang96
 /tests/v1/entrypoints/llm/test_struct_output_generate.py @mgoin @russellb @aarnphm
 /tests/v1/structured_output @mgoin @russellb @aarnphm
-/tests/v1/core @WoosukKwon @robertgshaw2-redhat @njhill @ywang96 @comaniac @alexm-redhat @heheda12345 @ApostaC
 /tests/weight_loading @mgoin @youkaichao @yewentao256
 /tests/lora @jeejeelee
 /tests/models/language/generation/test_hybrid.py @tdoublep
-/tests/v1/kv_connector/nixl_integration @NickLucche
-/tests/v1/kv_connector @ApostaC
-/tests/v1/offloading @ApostaC
-
-# Transformers backend
-/vllm/model_executor/models/transformers @hmellor
-/tests/models/test_transformers.py @hmellor
 
 # Docs
-/docs/mkdocs @hmellor
-/docs/**/*.yml @hmellor
-/requirements/docs.txt @hmellor
-.readthedocs.yaml @hmellor
+/docs @hmellor
 mkdocs.yaml @hmellor
 
-# Linting
-.markdownlint.yaml @hmellor
-.pre-commit-config.yaml @hmellor
-/tools/pre_commit @hmellor
-
 # CPU
-/vllm/v1/worker/cpu* @bigPYJ1151
+/vllm/v1/worker/^cpu @bigPYJ1151
 /csrc/cpu @bigPYJ1151
 /vllm/platforms/cpu.py @bigPYJ1151
 /cmake/cpu_extension.cmake @bigPYJ1151
 /docker/Dockerfile.cpu @bigPYJ1151
 
 # Intel GPU
-/vllm/v1/worker/xpu* @jikunshang
+/vllm/v1/worker/^xpu @jikunshang
 /vllm/platforms/xpu.py @jikunshang
 /docker/Dockerfile.xpu @jikunshang
 
@@ -89,9 +67,6 @@ mkdocs.yaml @hmellor
 /vllm/attention/backends/dual_chunk_flash_attn.py @sighingnow
 /vllm/model_executor/models/qwen* @sighingnow
 
-# MTP-specific files
-/vllm/model_executor/models/deepseek_mtp.py @luccafong
-
 # Mistral-specific files
 /vllm/model_executor/models/mistral*.py @patrickvonplaten
 /vllm/model_executor/models/mixtral*.py @patrickvonplaten
@@ -111,19 +86,3 @@ mkdocs.yaml @hmellor
 /vllm/attention/ops/rocm*.py @gshtras
 /vllm/model_executor/layers/fused_moe/rocm*.py @gshtras
 
-# TPU
-/vllm/v1/worker/tpu* @NickLucche
-/vllm/platforms/tpu.py @NickLucche
-/vllm/v1/sample/tpu @NickLucche
-/vllm/tests/v1/tpu @NickLucche
-
-# KVConnector installation files
-/requirements/kv_connectors.txt @NickLucche
-
-# Pooling models
-/examples/*/pooling/ @noooop
-/tests/models/*/pooling* @noooop
-/tests/entrypoints/pooling @noooop
-/vllm/config/pooler.py @noooop
-/vllm/pooling_params.py @noooop
-/vllm/model_executor/layers/pooler.py @noooop

.github/ISSUE_TEMPLATE/750-RFC.yml

@@ -43,6 +43,10 @@ body:
       Any other things you would like to mention.
   validations:
     required: false
+- type: markdown
+  attributes:
+    value: >
+      Thanks for contributing 🎉! The vLLM core team hosts a biweekly RFC review session at 9:30AM Pacific Time, while most RFCs can be discussed online, you can optionally sign up for a slot to discuss your RFC online [here](https://docs.google.com/document/d/1CiLVBZeIVfR7_PNAKVSusxpceywkoOOB78qoWqHvSZc/edit).
 - type: checkboxes
   id: askllm
   attributes:

.github/mergify.yml

@@ -2,7 +2,6 @@ pull_request_rules:
 - name: label-documentation
   description: Automatically apply documentation label
   conditions:
-    - label != stale
     - or:
       - files~=^[^/]+\.md$
       - files~=^docs/
@@ -11,13 +10,10 @@ pull_request_rules:
     label:
       add:
         - documentation
-    comment:
-      message: "Documentation preview: https://vllm--{{number}}.org.readthedocs.build/en/{{number}}/"
 
 - name: label-ci-build
   description: Automatically apply ci/build label
   conditions:
-    - label != stale
     - or:
       - files~=^\.github/
       - files~=\.buildkite/
@@ -34,7 +30,6 @@ pull_request_rules:
 - name: label-deepseek
   description: Automatically apply deepseek label
   conditions:
-    - label != stale
     - or:
       - files~=^examples/.*deepseek.*\.py
       - files~=^tests/.*deepseek.*\.py
@@ -51,7 +46,6 @@ pull_request_rules:
 - name: label-frontend
   description: Automatically apply frontend label
   conditions:
-    - label != stale
     - files~=^vllm/entrypoints/
   actions:
     label:
@@ -61,7 +55,6 @@ pull_request_rules:
 - name: label-llama
   description: Automatically apply llama label
   conditions:
-    - label != stale
     - or:
       - files~=^examples/.*llama.*\.py
       - files~=^tests/.*llama.*\.py
@@ -77,7 +70,6 @@ pull_request_rules:
 - name: label-multi-modality
   description: Automatically apply multi-modality label
   conditions:
-    - label != stale
     - or:
       - files~=^vllm/multimodal/
       - files~=^tests/multimodal/
@@ -91,7 +83,6 @@ pull_request_rules:
 - name: label-new-model
   description: Automatically apply new-model label
   conditions:
-    - label != stale
     - and:
       - files~=^vllm/model_executor/models/
       - files=vllm/model_executor/models/registry.py
@@ -103,7 +94,6 @@ pull_request_rules:
 - name: label-performance
   description: Automatically apply performance label
   conditions:
-    - label != stale
     - or:
       - files~=^benchmarks/
       - files~=^vllm/benchmarks/
@@ -117,7 +107,6 @@ pull_request_rules:
 - name: label-qwen
   description: Automatically apply qwen label
   conditions:
-    - label != stale
     - or:
       - files~=^examples/.*qwen.*\.py
       - files~=^tests/.*qwen.*\.py
@@ -132,20 +121,12 @@ pull_request_rules:
 - name: label-gpt-oss
   description: Automatically apply gpt-oss label
   conditions:
-    - label != stale
     - or:
       - files~=^examples/.*gpt[-_]?oss.*\.py
       - files~=^tests/.*gpt[-_]?oss.*\.py
-      - files~=^tests/entrypoints/openai/test_response_api_with_harmony.py
-      - files~=^tests/entrypoints/test_context.py
       - files~=^vllm/model_executor/models/.*gpt[-_]?oss.*\.py
       - files~=^vllm/model_executor/layers/.*gpt[-_]?oss.*\.py
-      - files~=^vllm/entrypoints/harmony_utils.py
-      - files~=^vllm/entrypoints/tool_server.py
-      - files~=^vllm/entrypoints/tool.py
-      - files~=^vllm/entrypoints/context.py
       - title~=(?i)gpt[-_]?oss
-      - title~=(?i)harmony
   actions:
     label:
       add:
@@ -154,7 +135,6 @@ pull_request_rules:
 - name: label-rocm
   description: Automatically apply rocm label
   conditions:
-    - label != stale
     - or:
       - files~=^csrc/rocm/
       - files~=^docker/Dockerfile.rocm
@@ -175,7 +155,6 @@ pull_request_rules:
 - name: label-structured-output
   description: Automatically apply structured-output label
   conditions:
-    - label != stale
     - or:
       - files~=^benchmarks/structured_schemas/
       - files=benchmarks/benchmark_serving_structured_output.py
@@ -185,7 +164,7 @@ pull_request_rules:
       - files=examples/online_serving/openai_chat_completion_structured_outputs.py
       - files=examples/online_serving/openai_chat_completion_structured_outputs_with_reasoning.py
       - files~=^tests/v1/structured_output/
-      - files=tests/v1/entrypoints/llm/test_struct_output_generate.py
+      - files=tests/v1/entrypoints/llm/test_guided_generate.py
       - files~=^vllm/v1/structured_output/
   actions:
     label:
@@ -195,7 +174,6 @@ pull_request_rules:
 - name: label-speculative-decoding
   description: Automatically apply speculative-decoding label
   conditions:
-    - label != stale
     - or:
       - files~=^vllm/v1/spec_decode/
       - files~=^tests/v1/spec_decode/
@@ -211,7 +189,6 @@ pull_request_rules:
 - name: label-v1
   description: Automatically apply v1 label
   conditions:
-    - label != stale
     - or:
       - files~=^vllm/v1/
       - files~=^tests/v1/
@@ -224,7 +201,6 @@ pull_request_rules:
   description: Automatically apply tpu label
   # Keep this list in sync with `label-tpu-remove` conditions
   conditions:
-    - label != stale
     - or:
       - files~=tpu.py
       - files~=_tpu
@@ -240,7 +216,6 @@ pull_request_rules:
   description: Automatically remove tpu label
   # Keep this list in sync with `label-tpu` conditions
   conditions:
-    - label != stale
     - and:
       - -files~=tpu.py
       - -files~=_tpu
@@ -255,9 +230,9 @@ pull_request_rules:
 - name: label-tool-calling
   description: Automatically add tool-calling label
   conditions:
-    - label != stale
     - or:
       - files~=^tests/tool_use/
+      - files~=^tests/mistral_tool_use/
       - files~=^tests/entrypoints/openai/tool_parsers/
       - files=tests/entrypoints/openai/test_chat_with_tool_reasoning.py
       - files~=^vllm/entrypoints/openai/tool_parsers/
@@ -274,9 +249,8 @@ pull_request_rules:
 
 - name: ping author on conflicts and add 'needs-rebase' label
   conditions:
-    - label != stale
-    - conflict
-    - -closed
+      - conflict
+      - -closed
   actions:
     label:
       add:
@@ -290,55 +264,20 @@ pull_request_rules:
 
 - name: assign reviewer for tensorizer changes
   conditions:
-    - label != stale
-    - or:
       - 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/model_executor/model_loader/tensorizer_loader/
+      - files~=^tests/tensorizer_loader/
   actions:
     assign:
       users:
         - "sangstar"
 
-- name: assign reviewer for modelopt changes
-  conditions:
-    - label != stale
-    - or:
-        - files~=^vllm/model_executor/layers/quantization/modelopt\.py$
-        - files~=^vllm/model_executor/layers/quantization/__init__\.py$
-        - files~=^tests/models/quantization/test_modelopt\.py$
-        - files~=^tests/quantization/test_modelopt\.py$
-        - files~=^tests/models/quantization/test_nvfp4\.py$
-        - files~=^docs/features/quantization/modelopt\.md$
-  actions:
-    assign:
-      users:
-        - "Edwardf0t1"
-
 - name: remove 'needs-rebase' label when conflict is resolved
   conditions:
-    - -conflict
-    - -closed
+      - -conflict
+      - -closed
   actions:
     label:
       remove:
         - needs-rebase
-
-- name: label-kv-connector
-  description: Automatically apply kv-connector label
-  conditions:
-    - label != stale
-    - or:
-      - files~=^examples/online_serving/disaggregated[^/]*/.*
-      - files~=^examples/offline_inference/disaggregated[^/]*/.*
-      - files~=^examples/others/lmcache/
-      - files~=^tests/v1/kv_connector/
-      - files~=^vllm/distributed/kv_transfer/
-      - title~=(?i)\bP/?D\b
-      - title~=(?i)NIXL
-      - title~=(?i)LMCache
-  actions:
-    label:
-      add:
-        - kv-connector

.github/workflows/add_label_automerge.yml

@@ -10,7 +10,7 @@ jobs:
         runs-on: ubuntu-latest
         steps:
             -   name: Add label
-                uses: actions/github-script@ed597411d8f924073f98dfc5c65a23a2325f34cd # v8.0.0
+                uses: actions/github-script@60a0d83039c74a4aee543508d2ffcb1c3799cdea # v7.0.1
                 with:
                     script: |
                         github.rest.issues.addLabels({

.github/workflows/bc-lint.yml

@@ -1,29 +0,0 @@
-name: BC Lint
-
-on:
-  pull_request:
-    types:
-      - opened
-      - synchronize
-      - reopened
-      - labeled
-      - unlabeled
-
-jobs:
-  bc_lint:
-    if: github.repository_owner == 'vllm-project'
-    runs-on: ubuntu-latest
-    steps:
-      - name: Run BC Lint Action
-        uses: pytorch/test-infra/.github/actions/bc-lint@main
-        with:
-          repo: ${{ github.event.pull_request.head.repo.full_name }}
-          base_sha: ${{ github.event.pull_request.base.sha }}
-          head_sha: ${{ github.event.pull_request.head.sha }}
-          suppression: ${{ contains(github.event.pull_request.labels.*.name, 'suppress-bc-linter') }}
-          docs_link: 'https://github.com/pytorch/test-infra/wiki/BC-Linter'
-          config_dir: .github
-
-concurrency:
-  group: ${{ github.workflow }}-${{ github.event.pull_request.number || github.sha }}
-  cancel-in-progress: true

.github/workflows/cleanup_pr_body.yml

@@ -16,7 +16,7 @@ jobs:
         uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
 
       - name: Set up Python
-        uses: actions/setup-python@e797f83bcb11b83ae66e0230d6156d7c80228e7c # v6.0.0
+        uses: actions/setup-python@42375524e23c412d93fb67b49958b491fce71c38 # v5.4.0
         with:
           python-version: '3.12'
 

.github/workflows/issue_autolabel.yml

@@ -13,8 +13,7 @@ jobs:
     runs-on: ubuntu-latest
     steps:
       - name: Label issues based on keywords
-        id: label-step
-        uses: actions/github-script@ed597411d8f924073f98dfc5c65a23a2325f34cd  # v8.0.0
+        uses: actions/github-script@60a0d83039c74a4aee543508d2ffcb1c3799cdea  # v7.0.1
         with:
           script: |
             // Configuration: Add new labels and keywords here
@@ -43,16 +42,13 @@ jobs:
                     searchIn: "body"
                   },
                 ],
+                
                 // Substring search - matches anywhere in text (partial matches)
                 substrings: [
                   {
                     term: "VLLM_ROCM_",
                     searchIn: "both"
                   },
-                  {
-                    term: "aiter",
-                    searchIn: "title"
-                  },
                   {
                     term: "rocm",
                     searchIn: "title"
@@ -89,12 +85,14 @@ jobs:
                     term: "hip_",
                     searchIn: "both"
                   },
+                  
                   // ROCm tools and libraries
                   {
                     term: "hipify",
                     searchIn: "both"
                   },
                 ],
+                
                 // Regex patterns - for complex pattern matching
                 regexPatterns: [
                   {
@@ -105,17 +103,13 @@ jobs:
                   }
                 ],
               },
-              // Add more label configurations here as needed
-              // example: {
-              //   keywords: [...],
-              //   substrings: [...],
-              //   regexPatterns: [...]
-              // },
             };
+            
             // Helper function to create regex based on search type
             function createSearchRegex(term, type) {
               // Escape special regex characters in the term
               const escapedTerm = term.replace(/[.*+?^${}()|[\]\\]/g, '\\$&');
+              
               switch (type) {
                 case 'keyword':
                   // Word boundary search - matches whole words only
@@ -127,13 +121,16 @@ jobs:
                   throw new Error(`Unknown search type: ${type}`);
               }
             }
+            
             // Helper function to find matching terms in text with line information
             function findMatchingTermsWithLines(text, searchTerms = [], searchType = 'keyword', searchLocation = '') {
               const matches = [];
               const lines = text.split('\n');
+              
               for (const termConfig of searchTerms) {
                 let regex;
                 let term, searchIn, pattern, description, flags;
+                
                 // Handle different input formats (string or object)
                 if (typeof termConfig === 'string') {
                   term = termConfig;
@@ -145,17 +142,21 @@ jobs:
                   description = termConfig.description;
                   flags = termConfig.flags;
                 }
+                
                 // Skip if this term shouldn't be searched in the current location
                 if (searchIn !== 'both' && searchIn !== searchLocation) {
                   continue;
                 }
+                
                 // Create appropriate regex
                 if (searchType === 'regex') {
                   regex = new RegExp(pattern, flags || "gi");
                 } else {
                   regex = createSearchRegex(term, searchType);
                 }
+                
                 const termMatches = [];
+                
                 // Check each line for matches
                 lines.forEach((line, lineIndex) => {
                   const lineMatches = line.match(regex);
@@ -170,14 +171,15 @@ jobs:
                         originalTerm: term || pattern,
                         description: description,
                         // Show context around the match in the line
-                        context: line.length > 100 ?
-                          line.substring(Math.max(0, line.toLowerCase().indexOf(match.toLowerCase()) - 30),
-                                       line.toLowerCase().indexOf(match.toLowerCase()) + match.length + 30) + '...'
+                        context: line.length > 100 ? 
+                          line.substring(Math.max(0, line.toLowerCase().indexOf(match.toLowerCase()) - 30), 
+                                       line.toLowerCase().indexOf(match.toLowerCase()) + match.length + 30) + '...' 
                           : line.trim()
                       });
                     });
                   }
                 });
+                
                 if (termMatches.length > 0) {
                   matches.push({
                     term: term || (description || pattern),
@@ -190,48 +192,64 @@ jobs:
                   });
                 }
               }
+              
               return matches;
             }
+            
             // Helper function to check if label should be added
             async function processLabel(labelName, config) {
               const body = context.payload.issue.body || "";
               const title = context.payload.issue.title || "";
+              
               core.notice(`Processing label: ${labelName}`);
               core.notice(`Issue Title: "${title}"`);
               core.notice(`Issue Body length: ${body.length} characters`);
+              
               let shouldAddLabel = false;
               let allMatches = [];
               let reason = '';
+              
               const keywords = config.keywords || [];
               const substrings = config.substrings || [];
               const regexPatterns = config.regexPatterns || [];
+              
               core.notice(`Searching with ${keywords.length} keywords, ${substrings.length} substrings, and ${regexPatterns.length} regex patterns`);
+              
               // Search in title
               if (title.trim()) {
                 core.notice(`Searching in title: "${title}"`);
+                
                 const titleKeywordMatches = findMatchingTermsWithLines(title, keywords, 'keyword', 'title');
                 const titleSubstringMatches = findMatchingTermsWithLines(title, substrings, 'substring', 'title');
                 const titleRegexMatches = findMatchingTermsWithLines(title, regexPatterns, 'regex', 'title');
+                
                 allMatches.push(...titleKeywordMatches, ...titleSubstringMatches, ...titleRegexMatches);
               }
+              
               // Search in body
               if (body.trim()) {
                 core.notice(`Searching in body (${body.length} characters)`);
+                
                 const bodyKeywordMatches = findMatchingTermsWithLines(body, keywords, 'keyword', 'body');
                 const bodySubstringMatches = findMatchingTermsWithLines(body, substrings, 'substring', 'body');
                 const bodyRegexMatches = findMatchingTermsWithLines(body, regexPatterns, 'regex', 'body');
+                
                 allMatches.push(...bodyKeywordMatches, ...bodySubstringMatches, ...bodyRegexMatches);
               }
+              
               if (allMatches.length > 0) {
                 core.notice(`Found ${allMatches.length} matching term(s):`);
+                
                 for (const termMatch of allMatches) {
                   const locationText = termMatch.searchLocation === 'title' ? 'title' : 'body';
                   const searchInText = termMatch.searchIn === 'both' ? 'both' : termMatch.searchIn;
+                  
                   if (termMatch.searchType === 'regex') {
                     core.notice(`  📍 Regex: "${termMatch.term}" (pattern: ${termMatch.pattern}) found ${termMatch.count} time(s) in ${locationText} (configured to search in: ${searchInText}):`);
                   } else {
                     core.notice(`  📍 Term: "${termMatch.term}" (${termMatch.searchType} search) found ${termMatch.count} time(s) in ${locationText} (configured to search in: ${searchInText}):`);
                   }
+                  
                   // Show details for each match
                   termMatch.matches.forEach((match, index) => {
                     core.notice(`    ${index + 1}. Line ${match.lineNumber} in ${match.searchLocation}: "${match.match}" [${match.searchType}]`);
@@ -244,6 +262,7 @@ jobs:
                     }
                   });
                 }
+                
                 shouldAddLabel = true;
                 const totalMatches = allMatches.reduce((sum, t) => sum + t.count, 0);
                 const titleMatches = allMatches.filter(t => t.searchLocation === 'title').reduce((sum, t) => sum + t.count, 0);
@@ -251,10 +270,13 @@ jobs:
                 const keywordMatches = allMatches.filter(t => t.searchType === 'keyword').reduce((sum, t) => sum + t.count, 0);
                 const substringMatches = allMatches.filter(t => t.searchType === 'substring').reduce((sum, t) => sum + t.count, 0);
                 const regexMatches = allMatches.filter(t => t.searchType === 'regex').reduce((sum, t) => sum + t.count, 0);
+                
                 reason = `Found ${totalMatches} total matches (${titleMatches} in title, ${bodyMatches} in body) - ${keywordMatches} keyword matches, ${substringMatches} substring matches, ${regexMatches} regex matches`;
               }
+              
               core.notice(`Final decision: ${shouldAddLabel ? 'ADD LABEL' : 'DO NOT ADD LABEL'}`);
               core.notice(`Reason: ${reason || 'No matching terms found'}`);
+              
               if (shouldAddLabel) {
                 const existingLabels = context.payload.issue.labels.map(l => l.name);
                 if (!existingLabels.includes(labelName)) {
@@ -270,92 +292,14 @@ jobs:
                 core.notice(`Label "${labelName}" already present.`);
                 return false;
               }
+              
               core.notice(`No matching terms found for label "${labelName}".`);
               return false;
             }
+            
             // Process all configured labels
-            const labelsAddedResults = await Promise.all(
-              Object.entries(labelConfig).map(([labelName, config]) => 
-                processLabel(labelName, config).then(added => ({ labelName, added }))
-              )
-            );
-            
-            const numLabelsAdded = labelsAddedResults.filter(r => r.added).length;
-            core.notice(`Processing complete. ${numLabelsAdded} label(s) added.`);
-            
-            // Return which labels were added for the next step
-            const addedLabels = labelsAddedResults.filter(r => r.added).map(r => r.labelName);
-            core.setOutput('labels_added', JSON.stringify(addedLabels));
-            return addedLabels;
-
-      - name: CC users for labeled issues
-        if: steps.label-step.outputs.labels_added != '[]'
-        uses: actions/github-script@ed597411d8f924073f98dfc5c65a23a2325f34cd  # v8.0.0
-        with:
-          script: |
-            // Configuration: Map labels to GitHub users to CC
-            // You can add multiple users per label, and multiple label configurations
-            const ccConfig = {
-              rocm: {
-                users: ['hongxiayang', 'tjtanaa', 'vllmellm'],  // Add more users as needed: ['user1', 'user2', 'user3']
-                message: 'CC {users} for ROCm-related issue'  // {users} will be replaced with @mentions
-              },
-              // Add more label -> user mappings here
-              // Example:
-              // cuda: {
-              //   users: ['user1', 'user2'],
-              //   message: 'CC {users} for CUDA-related issue'
-              // },
-              // performance: {
-              //   users: ['perfexpert'],
-              //   message: 'CC {users} for performance issue'
-              // },
-            };
-            
-            const labelsAdded = JSON.parse('${{ steps.label-step.outputs.labels_added }}');
-            core.notice(`Labels added: ${labelsAdded.join(', ')}`);
-            
-            // Get existing comments to check for already mentioned users
-            const comments = await github.rest.issues.listComments({
-              owner: context.repo.owner,
-              repo: context.repo.repo,
-              issue_number: context.issue.number,
-            });
-            
-            const issueBody = context.payload.issue.body || '';
-            const allExistingText = issueBody + '\n' + comments.data.map(c => c.body).join('\n');
-            
-            // Process each label that was added
-            for (const label of labelsAdded) {
-              if (ccConfig[label]) {
-                const config = ccConfig[label];
-                const usersToMention = [];
-                
-                // Check which users haven't been mentioned yet
-                for (const user of config.users) {
-                  const mentionPattern = new RegExp(`@${user}\\b`, 'i');
-                  if (!mentionPattern.test(allExistingText)) {
-                    usersToMention.push(user);
-                  } else {
-                    core.notice(`@${user} already mentioned for label "${label}", skipping`);
-                  }
-                }
-                
-                // Post comment if there are users to mention
-                if (usersToMention.length > 0) {
-                  const mentions = usersToMention.map(u => `@${u}`).join(' ');
-                  const message = config.message.replace('{users}', mentions);
-                  
-                  await github.rest.issues.createComment({
-                    owner: context.repo.owner,
-                    repo: context.repo.repo,
-                    issue_number: context.issue.number,
-                    body: message
-                  });
-                  
-                  core.notice(`CC comment added for label "${label}": ${mentions}`);
-                } else {
-                  core.notice(`All users for label "${label}" already mentioned, skipping comment`);
-                }
-              }
-            }
+            const processLabels = Object.entries(labelConfig)
+              .map(([labelName, config]) => processLabel(labelName, config));
+            const labelsAdded = await Promise.all(processLabels);
+            const numLabelsAdded = labelsAdded.reduce((x, y) => x + y, 0);
+            core.notice(`Processing complete. ${numLabelsAdded} label(s) added.`);

.github/workflows/pre-commit.yml

@@ -17,7 +17,7 @@ jobs:
     runs-on: ubuntu-latest
     steps:
     - uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
-    - uses: actions/setup-python@e797f83bcb11b83ae66e0230d6156d7c80228e7c # v6.0.0
+    - uses: actions/setup-python@42375524e23c412d93fb67b49958b491fce71c38 # v5.4.0
       with:
         python-version: "3.12"
     - run: echo "::add-matcher::.github/workflows/matchers/actionlint.json"

.github/workflows/reminder_comment.yml

@@ -9,7 +9,7 @@ jobs:
     runs-on: ubuntu-latest
     steps:
       - name: Remind to run full CI on PR
-        uses: actions/github-script@ed597411d8f924073f98dfc5c65a23a2325f34cd # v8.0.0
+        uses: actions/github-script@60a0d83039c74a4aee543508d2ffcb1c3799cdea # v7.0.1
         with:
           script: |
             try {

.github/workflows/stale.yml

@@ -13,7 +13,7 @@ jobs:
       actions: write
     runs-on: ubuntu-latest
     steps:
-      - uses: actions/stale@5f858e3efba33a5ca4407a664cc011ad407f2008 # v10.1.0
+      - uses: actions/stale@5bef64f19d7facfb25b37b414482c7164d639639 # v9.1.0
         with:
           # Increasing this value ensures that changes to this workflow
           # propagate to all issues and PRs in days rather than months

.gitignore

@@ -4,7 +4,7 @@
 # vllm-flash-attn built from source
 vllm/vllm_flash_attn/*
 
-# triton jit
+# triton jit 
 .triton
 
 # Byte-compiled / optimized / DLL files
@@ -94,9 +94,6 @@ ipython_config.py
 # generated files
 **/generated/**
 
-# uv
-uv.lock
-
 # pyenv
 #   For a library or package, you might want to ignore these files since the code is
 #   intended to run in multiple environments; otherwise, check them in:
@@ -180,14 +177,6 @@ cython_debug/
 # VSCode
 .vscode/
 
-# Claude
-CLAUDE.md
-.claude/
-
-# Codex
-AGENTS.md
-.codex/
-
 # DS Store
 .DS_Store
 
@@ -220,4 +209,4 @@ shellcheck*/
 csrc/moe/marlin_moe_wna16/kernel_*
 
 # Ignore ep_kernels_workspace folder
-ep_kernels_workspace/
+ep_kernels_workspace/

.markdownlint.yaml

@@ -4,6 +4,7 @@ MD013: false
 MD024:
   siblings_only: true
 MD033: false
+MD042: false
 MD045: false
 MD046: false
 MD051: false

.pre-commit-config.yaml

@@ -6,19 +6,30 @@ default_stages:
   - manual # Run in CI
 exclude: 'vllm/third_party/.*'
 repos:
-- repo: https://github.com/astral-sh/ruff-pre-commit
-  rev: v0.14.0
+- repo: https://github.com/google/yapf
+  rev: v0.43.0
   hooks:
-  - id: ruff-check
+  - id: yapf
+    args: [--in-place, --verbose]
+    # Keep the same list from yapfignore here to avoid yapf failing without any inputs
+    exclude: '(.buildkite|benchmarks|build|examples)/.*'
+- repo: https://github.com/astral-sh/ruff-pre-commit
+  rev: v0.11.7
+  hooks:
+  - id: ruff
     args: [--output-format, github, --fix]
   - id: ruff-format
+    files: ^(.buildkite|benchmarks|examples)/.*
 - repo: https://github.com/crate-ci/typos
-  rev: v1.38.1
+  rev: v1.35.5
   hooks:
   - id: typos
-    args: [--force-exclude]
+- repo: https://github.com/PyCQA/isort
+  rev: 6.0.1
+  hooks:
+  - id: isort
 - repo: https://github.com/pre-commit/mirrors-clang-format
-  rev: v21.1.2
+  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/.*'
@@ -35,10 +46,10 @@ repos:
   hooks:
   - id: actionlint
 - repo: https://github.com/astral-sh/uv-pre-commit
-  rev: 0.9.1
+  rev: 0.6.17
   hooks:
     - id: pip-compile
-      args: [requirements/test.in, -o, requirements/test.txt, --index-strategy, unsafe-best-match, --torch-backend, cu129, --python-platform, x86_64-manylinux_2_28]
+      args: [requirements/test.in, -o, requirements/test.txt, --index-strategy, unsafe-best-match, --torch-backend, cu128]
       files: ^requirements/test\.(in|txt)$
 - repo: local
   hooks:
@@ -48,33 +59,39 @@ repos:
     entry: python tools/generate_nightly_torch_test.py
     files: ^requirements/test\.(in|txt)$
   - id: mypy-local
-    name: Run mypy locally for lowest supported Python version
-    entry: python tools/pre_commit/mypy.py 0 "3.10"
+    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, pydantic]
     stages: [pre-commit] # Don't run in CI
-    <<: &mypy_common
-      language: python
-      types_or: [python, pyi]
-      require_serial: true
-      additional_dependencies: [mypy==1.11.1, regex, types-cachetools, types-setuptools, types-PyYAML, types-requests, types-torch, pydantic]
+  - id: mypy-3.9 # TODO: Use https://github.com/pre-commit/mirrors-mypy when mypy setup is less awkward
+    name: Run mypy for Python 3.9
+    entry: tools/mypy.sh 1 "3.9"
+    language: python
+    types: [python]
+    additional_dependencies: *mypy_deps
+    stages: [manual] # Only run in CI
   - id: mypy-3.10 # TODO: Use https://github.com/pre-commit/mirrors-mypy when mypy setup is less awkward
     name: Run mypy for Python 3.10
-    entry: python tools/pre_commit/mypy.py 1 "3.10"
-    <<: *mypy_common
+    entry: tools/mypy.sh 1 "3.10"
+    language: python
+    types: [python]
+    additional_dependencies: *mypy_deps
     stages: [manual] # Only run in CI
   - id: mypy-3.11 # TODO: Use https://github.com/pre-commit/mirrors-mypy when mypy setup is less awkward
     name: Run mypy for Python 3.11
-    entry: python tools/pre_commit/mypy.py 1 "3.11"
-    <<: *mypy_common
+    entry: tools/mypy.sh 1 "3.11"
+    language: python
+    types: [python]
+    additional_dependencies: *mypy_deps
     stages: [manual] # Only run in CI
   - id: mypy-3.12 # TODO: Use https://github.com/pre-commit/mirrors-mypy when mypy setup is less awkward
     name: Run mypy for Python 3.12
-    entry: python tools/pre_commit/mypy.py 1 "3.12"
-    <<: *mypy_common
-    stages: [manual] # Only run in CI
-  - id: mypy-3.13 # TODO: Use https://github.com/pre-commit/mirrors-mypy when mypy setup is less awkward
-    name: Run mypy for Python 3.13
-    entry: python tools/pre_commit/mypy.py 1 "3.13"
-    <<: *mypy_common
+    entry: tools/mypy.sh 1 "3.12"
+    language: python
+    types: [python]
+    additional_dependencies: *mypy_deps
     stages: [manual] # Only run in CI
   - id: shellcheck
     name: Lint shell scripts
@@ -138,15 +155,18 @@ repos:
     additional_dependencies: [regex]
   - id: check-pickle-imports
     name: Prevent new pickle/cloudpickle imports
-    entry: python tools/pre_commit/check_pickle_imports.py
+    entry: python tools/check_pickle_imports.py
     language: python
     types: [python]
-    additional_dependencies: [regex]
+    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
-    additional_dependencies: [regex]
+    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

.readthedocs.yaml

@@ -13,7 +13,6 @@ build:
 
 mkdocs:
   configuration: mkdocs.yaml
-  fail_on_warning: true
 
 # Optionally declare the Python requirements required to build your docs
 python:

.yapfignore

@@ -1,2 +1 @@
 collect_env.py
-vllm/model_executor/layers/fla/ops/*.py

CMakeLists.txt

@@ -13,10 +13,6 @@ cmake_minimum_required(VERSION 3.26)
 # cmake --install . --component _C
 project(vllm_extensions LANGUAGES CXX)
 
-set(CMAKE_CXX_STANDARD 17)
-set(CMAKE_CXX_STANDARD_REQUIRED ON)
-
-
 # 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}")
@@ -34,10 +30,10 @@ 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 "3.10" "3.11" "3.12" "3.13")
+set(PYTHON_SUPPORTED_VERSIONS "3.9" "3.10" "3.11" "3.12" "3.13")
 
 # Supported AMD GPU architectures.
-set(HIP_SUPPORTED_ARCHS "gfx906;gfx908;gfx90a;gfx942;gfx950;gfx1030;gfx1100;gfx1101;gfx1200;gfx1201;gfx1150;gfx1151")
+set(HIP_SUPPORTED_ARCHS "gfx906;gfx908;gfx90a;gfx942;gfx950;gfx1030;gfx1100;gfx1101;gfx1200;gfx1201")
 
 #
 # Supported/expected torch versions for CUDA/ROCm.
@@ -49,8 +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 "2.9.0")
-set(TORCH_SUPPORTED_VERSION_ROCM "2.9.0")
+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
@@ -86,9 +82,6 @@ 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 13.0)
-  set(CUDA_SUPPORTED_ARCHS "7.5;8.0;8.6;8.7;8.9;9.0;10.0;11.0;12.0")
-elseif(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()
@@ -178,25 +171,6 @@ if(NVCC_THREADS AND VLLM_GPU_LANG STREQUAL "CUDA")
   list(APPEND VLLM_GPU_FLAGS "--threads=${NVCC_THREADS}")
 endif()
 
-#
-# Set compression mode for CUDA >=13.x.
-#
-if(VLLM_GPU_LANG STREQUAL "CUDA" AND
-   DEFINED CMAKE_CUDA_COMPILER_VERSION AND
-   CMAKE_CUDA_COMPILER_VERSION VERSION_GREATER_EQUAL 13.0)
-  list(APPEND VLLM_GPU_FLAGS "--compress-mode=size")
-endif()
-
-#
-# Set CUDA include flags for CXX compiler.
-#
-if(VLLM_GPU_LANG STREQUAL "CUDA")
-  set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -I${CUDA_TOOLKIT_ROOT_DIR}/include")
-  if(CUDA_VERSION VERSION_GREATER_EQUAL 13.0)
-    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -I${CUDA_TOOLKIT_ROOT_DIR}/include/cccl")
-  endif()
-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.
@@ -269,8 +243,8 @@ set(VLLM_EXT_SRC
   "csrc/sampler.cu"
   "csrc/cuda_view.cu"
   "csrc/quantization/gptq/q_gemm.cu"
-  "csrc/quantization/w8a8/int8/scaled_quant.cu"
-  "csrc/quantization/w8a8/fp8/common.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"
@@ -282,7 +256,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
   SET(CUTLASS_ENABLE_HEADERS_ONLY ON CACHE BOOL "Enable only the header library")
 
   # Set CUTLASS_REVISION. Used for FetchContent. Also fixes some bogus messages when building.
-  set(CUTLASS_REVISION "v4.2.1" CACHE STRING "CUTLASS revision to use")
+  set(CUTLASS_REVISION "v4.0.0" CACHE STRING "CUTLASS revision to use")
 
   # Use the specified CUTLASS source directory for compilation if VLLM_CUTLASS_SRC_DIR is provided
   if (DEFINED ENV{VLLM_CUTLASS_SRC_DIR})
@@ -314,13 +288,14 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
   list(APPEND VLLM_EXT_SRC
     "csrc/quantization/awq/gemm_kernels.cu"
     "csrc/permute_cols.cu"
-    "csrc/quantization/w8a8/cutlass/scaled_mm_entry.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/quantization/w8a8/fp8/per_token_group_quant.cu"
-    "csrc/quantization/w8a8/int8/per_token_group_quant.cu")
+    "csrc/attention/mla/cutlass_mla_entry.cu"
+    "csrc/quantization/fp8/per_token_group_quant.cu")
 
   set_gencode_flags_for_srcs(
     SRCS "${VLLM_EXT_SRC}"
@@ -424,11 +399,11 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
   cuda_archs_loose_intersection(SCALED_MM_ARCHS "9.0a;" "${CUDA_ARCHS}")
   if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.0 AND SCALED_MM_ARCHS)
     set(SRCS
-       "csrc/quantization/w8a8/cutlass/scaled_mm_c3x_sm90.cu"
-       "csrc/quantization/w8a8/cutlass/c3x/scaled_mm_sm90_fp8.cu"
-       "csrc/quantization/w8a8/cutlass/c3x/scaled_mm_sm90_int8.cu"
-       "csrc/quantization/w8a8/cutlass/c3x/scaled_mm_azp_sm90_int8.cu"
-       "csrc/quantization/w8a8/cutlass/c3x/scaled_mm_blockwise_sm90_fp8.cu")
+       "csrc/quantization/cutlass_w8a8/scaled_mm_c3x_sm90.cu"
+       "csrc/quantization/cutlass_w8a8/c3x/scaled_mm_sm90_fp8.cu"
+       "csrc/quantization/cutlass_w8a8/c3x/scaled_mm_sm90_int8.cu"
+       "csrc/quantization/cutlass_w8a8/c3x/scaled_mm_azp_sm90_int8.cu"
+       "csrc/quantization/cutlass_w8a8/c3x/scaled_mm_blockwise_sm90_fp8.cu")
     set_gencode_flags_for_srcs(
       SRCS "${SRCS}"
       CUDA_ARCHS "${SCALED_MM_ARCHS}")
@@ -452,16 +427,12 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
 
   # The cutlass_scaled_mm kernels for Geforce Blackwell SM120 (c3x, i.e. CUTLASS 3.x) require
   # CUDA 12.8 or later
-  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
-    cuda_archs_loose_intersection(SCALED_MM_ARCHS "12.0f" "${CUDA_ARCHS}")
-  else()
-    cuda_archs_loose_intersection(SCALED_MM_ARCHS "12.0a" "${CUDA_ARCHS}")
-  endif()
+  cuda_archs_loose_intersection(SCALED_MM_ARCHS "12.0;12.0a" "${CUDA_ARCHS}")
   if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND SCALED_MM_ARCHS)
     set(SRCS
-      "csrc/quantization/w8a8/cutlass/scaled_mm_c3x_sm120.cu"
-      "csrc/quantization/w8a8/cutlass/c3x/scaled_mm_sm120_fp8.cu"
-      "csrc/quantization/w8a8/cutlass/c3x/scaled_mm_blockwise_sm120_fp8.cu"
+      "csrc/quantization/cutlass_w8a8/scaled_mm_c3x_sm120.cu"
+      "csrc/quantization/cutlass_w8a8/c3x/scaled_mm_sm120_fp8.cu"
+      "csrc/quantization/cutlass_w8a8/c3x/scaled_mm_blockwise_sm120_fp8.cu"
     )
     set_gencode_flags_for_srcs(
       SRCS "${SRCS}"
@@ -486,16 +457,12 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
 
   # The cutlass_scaled_mm kernels for Blackwell SM100 (c3x, i.e. CUTLASS 3.x)
   # require CUDA 12.8 or later
-  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
-    cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0f;11.0f;12.0f" "${CUDA_ARCHS}")
-  else()
-    cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0a;10.1a;10.3a;12.0a;12.1a" "${CUDA_ARCHS}")
-  endif()
+  cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0a;10.1a" "${CUDA_ARCHS}")
   if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND SCALED_MM_ARCHS)
     set(SRCS
-      "csrc/quantization/w8a8/cutlass/scaled_mm_c3x_sm100.cu"
-      "csrc/quantization/w8a8/cutlass/c3x/scaled_mm_sm100_fp8.cu"
-      "csrc/quantization/w8a8/cutlass/c3x/scaled_mm_blockwise_sm100_fp8.cu"
+      "csrc/quantization/cutlass_w8a8/scaled_mm_c3x_sm100.cu"
+      "csrc/quantization/cutlass_w8a8/c3x/scaled_mm_sm100_fp8.cu"
+      "csrc/quantization/cutlass_w8a8/c3x/scaled_mm_blockwise_sm100_fp8.cu"
     )
     set_gencode_flags_for_srcs(
       SRCS "${SRCS}"
@@ -526,7 +493,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
   # subtract out the archs that are already built for 3x
   list(REMOVE_ITEM SCALED_MM_2X_ARCHS ${SCALED_MM_3X_ARCHS})
   if (SCALED_MM_2X_ARCHS)
-    set(SRCS "csrc/quantization/w8a8/cutlass/scaled_mm_c2x.cu")
+    set(SRCS "csrc/quantization/cutlass_w8a8/scaled_mm_c2x.cu")
     set_gencode_flags_for_srcs(
       SRCS "${SRCS}"
       CUDA_ARCHS "${SCALED_MM_2X_ARCHS}")
@@ -570,11 +537,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
 
   # The nvfp4_scaled_mm_sm120 kernels for Geforce Blackwell SM120 require
   # CUDA 12.8 or later
-  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
-    cuda_archs_loose_intersection(FP4_ARCHS "12.0f" "${CUDA_ARCHS}")
-  else()
-    cuda_archs_loose_intersection(FP4_ARCHS "12.0a" "${CUDA_ARCHS}")
-  endif()
+  cuda_archs_loose_intersection(FP4_ARCHS "12.0;12.0a" "${CUDA_ARCHS}")
   if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND FP4_ARCHS)
     set(SRCS
       "csrc/quantization/fp4/nvfp4_quant_kernels.cu"
@@ -593,11 +556,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
   endif()
 
   # FP4 Archs and flags
-  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
-    cuda_archs_loose_intersection(FP4_ARCHS "10.0f;11.0f;12.0f" "${CUDA_ARCHS}")
-  else()
-    cuda_archs_loose_intersection(FP4_ARCHS "10.0a;10.1a;12.0a;12.1a" "${CUDA_ARCHS}")
-  endif()
+  cuda_archs_loose_intersection(FP4_ARCHS "10.0a" "${CUDA_ARCHS}")
   if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND FP4_ARCHS)
     set(SRCS
       "csrc/quantization/fp4/nvfp4_quant_kernels.cu"
@@ -619,13 +578,10 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
   endif()
 
   # CUTLASS MLA Archs and flags
-  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
-    cuda_archs_loose_intersection(MLA_ARCHS "10.0f;11.0f;12.0f" "${CUDA_ARCHS}")
-  else()
-    cuda_archs_loose_intersection(MLA_ARCHS "10.0a;10.1a;10.3a;12.0a;12.1a" "${CUDA_ARCHS}")
-  endif()
+  cuda_archs_loose_intersection(MLA_ARCHS "10.0a" "${CUDA_ARCHS}")
   if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND MLA_ARCHS)
     set(SRCS
+      "csrc/attention/mla/cutlass_mla_kernels.cu"
       "csrc/attention/mla/sm100_cutlass_mla_kernel.cu")
     set_gencode_flags_for_srcs(
       SRCS "${SRCS}"
@@ -649,7 +605,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
   # 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/w8a8/cutlass/moe/grouped_mm_c3x_sm90.cu")
+    set(SRCS "csrc/quantization/cutlass_w8a8/moe/grouped_mm_c3x_sm90.cu")
     set_gencode_flags_for_srcs(
       SRCS "${SRCS}"
       CUDA_ARCHS "${SCALED_MM_ARCHS}")
@@ -667,13 +623,9 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
     endif()
   endif()
 
-  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
-    cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0f;11.0f" "${CUDA_ARCHS}")
-  else()
-    cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0a" "${CUDA_ARCHS}")
-  endif()
+  cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0a" "${CUDA_ARCHS}")
   if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND SCALED_MM_ARCHS)
-    set(SRCS "csrc/quantization/w8a8/cutlass/moe/grouped_mm_c3x_sm100.cu")
+    set(SRCS "csrc/quantization/cutlass_w8a8/moe/grouped_mm_c3x_sm100.cu")
     set_gencode_flags_for_srcs(
       SRCS "${SRCS}"
       CUDA_ARCHS "${SCALED_MM_ARCHS}")
@@ -692,13 +644,9 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
   endif()
 
   # moe_data.cu is used by all CUTLASS MoE kernels.
-  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
-    cuda_archs_loose_intersection(CUTLASS_MOE_DATA_ARCHS "9.0a;10.0f;11.0f;12.0f" "${CUDA_ARCHS}")
-  else()
-    cuda_archs_loose_intersection(CUTLASS_MOE_DATA_ARCHS "9.0a;10.0a;10.1a;10.3a;12.0a;12.1a" "${CUDA_ARCHS}")
-  endif()
+  cuda_archs_loose_intersection(CUTLASS_MOE_DATA_ARCHS "9.0a;10.0a" "${CUDA_ARCHS}")
   if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.3 AND CUTLASS_MOE_DATA_ARCHS)
-    set(SRCS "csrc/quantization/w8a8/cutlass/moe/moe_data.cu")
+    set(SRCS "csrc/quantization/cutlass_w8a8/moe/moe_data.cu")
     set_gencode_flags_for_srcs(
       SRCS "${SRCS}"
       CUDA_ARCHS "${CUTLASS_MOE_DATA_ARCHS}")
@@ -715,13 +663,9 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
     endif()
   endif()
 
-  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
-    cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0f;11.0f;12.0f" "${CUDA_ARCHS}")
-  else()
-    cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0a;10.1a;10.3a;12.0a;12.1a" "${CUDA_ARCHS}")
-  endif()
+  cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0a" "${CUDA_ARCHS}")
   if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND SCALED_MM_ARCHS)
-    set(SRCS "csrc/quantization/w8a8/cutlass/moe/blockwise_scaled_group_mm_sm100.cu")
+    set(SRCS "csrc/quantization/cutlass_w8a8/moe/blockwise_scaled_group_mm_sm100.cu")
     set_gencode_flags_for_srcs(
       SRCS "${SRCS}"
       CUDA_ARCHS "${SCALED_MM_ARCHS}")
@@ -835,17 +779,6 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
     endif()
   endif()
 
-  # Hadacore kernels
-  cuda_archs_loose_intersection(HADACORE_ARCHS "8.0;8.9;9.0" "${CUDA_ARCHS}")
-  if(HADACORE_ARCHS)
-    set(SRCS "csrc/quantization/hadamard/hadacore/hadamard_transform_cuda.cu")
-    set_gencode_flags_for_srcs(
-      SRCS "${SRCS}"
-      CUDA_ARCHS "${HADACORE_ARCHS}")
-    list(APPEND VLLM_EXT_SRC "${SRCS}")
-    message(STATUS "Building hadacore")
-  endif()
-
 # if CUDA endif
 endif()
 
@@ -883,7 +816,6 @@ target_compile_definitions(_C PRIVATE CUTLASS_ENABLE_DIRECT_CUDA_DRIVER_CALL=1)
 set(VLLM_MOE_EXT_SRC
   "csrc/moe/torch_bindings.cpp"
   "csrc/moe/moe_align_sum_kernels.cu"
-  "csrc/moe/moe_lora_align_sum_kernels.cu"
   "csrc/moe/topk_softmax_kernels.cu")
 
 if(VLLM_GPU_LANG STREQUAL "CUDA")
@@ -1008,7 +940,6 @@ endif()
 # For CUDA we also build and ship some external projects.
 if (VLLM_GPU_LANG STREQUAL "CUDA")
     include(cmake/external_projects/flashmla.cmake)
-    include(cmake/external_projects/qutlass.cmake)
 
     # vllm-flash-attn should be last as it overwrites some CMake functions
     include(cmake/external_projects/vllm_flash_attn.cmake)

MANIFEST.in

@@ -2,6 +2,7 @@ include LICENSE
 include requirements/common.txt
 include requirements/cuda.txt
 include requirements/rocm.txt
+include requirements/neuron.txt
 include requirements/cpu.txt
 include CMakeLists.txt
 

README.md

@@ -14,25 +14,19 @@ Easy, fast, and cheap LLM serving for everyone
 | <a href="https://docs.vllm.ai"><b>Documentation</b></a> | <a href="https://blog.vllm.ai/"><b>Blog</b></a> | <a href="https://arxiv.org/abs/2309.06180"><b>Paper</b></a> | <a href="https://x.com/vllm_project"><b>Twitter/X</b></a> | <a href="https://discuss.vllm.ai"><b>User Forum</b></a> | <a href="https://slack.vllm.ai"><b>Developer Slack</b></a> |
 </p>
 
----
-Join us at the [PyTorch Conference, October 22-23](https://events.linuxfoundation.org/pytorch-conference/) and [Ray Summit, November 3-5](https://www.anyscale.com/ray-summit/2025) in San Francisco for our latest updates on vLLM and to meet the vLLM team! Register now for the largest vLLM community events of the year!
-
 ---
 
 *Latest News* 🔥
 
-- [2025/09] We hosted [vLLM Toronto Meetup](https://luma.com/e80e0ymm) focused on tackling inference at scale and speculative decoding with speakers from NVIDIA and Red Hat! Please find the meetup slides [here](https://docs.google.com/presentation/d/1IYJYmJcu9fLpID5N5RbW_vO0XLo0CGOR14IXOjB61V8/edit?usp=sharing).
-- [2025/08] We hosted [vLLM Shenzhen Meetup](https://mp.weixin.qq.com/s/k8ZBO1u2_2odgiKWH_GVTQ) focusing on the ecosystem around vLLM! Please find the meetup slides [here](https://drive.google.com/drive/folders/1Ua2SVKVSu-wp5vou_6ElraDt2bnKhiEA).
-- [2025/08] We hosted [vLLM Singapore Meetup](https://www.sginnovate.com/event/vllm-sg-meet). We shared V1 updates, disaggregated serving and MLLM speedups with speakers from Embedded LLM, AMD, WekaIO, and A*STAR. Please find the meetup slides [here](https://drive.google.com/drive/folders/1ncf3GyqLdqFaB6IeB834E5TZJPLAOiXZ?usp=sharing).
 - [2025/08] We hosted [vLLM Shanghai Meetup](https://mp.weixin.qq.com/s/pDmAXHcN7Iqc8sUKgJgGtg) focusing on building, developing, and integrating with vLLM! Please find the meetup slides [here](https://drive.google.com/drive/folders/1OvLx39wnCGy_WKq8SiVKf7YcxxYI3WCH).
+- [2025/08] We hosted [vLLM Korea Meetup](https://luma.com/cgcgprmh) with Red Hat and Rebellions! We shared the latest advancements in vLLM along with project spotlights from the vLLM Korea community. Please find the meetup slides [here](https://drive.google.com/file/d/1bcrrAE1rxUgx0mjIeOWT6hNe2RefC5Hm/view).
+- [2025/08] We hosted [vLLM Beijing Meetup](https://mp.weixin.qq.com/s/dgkWg1WFpWGO2jCdTqQHxA) focusing on large-scale LLM deployment! Please find the meetup slides [here](https://drive.google.com/drive/folders/1Pid6NSFLU43DZRi0EaTcPgXsAzDvbBqF) and the recording [here](https://www.chaspark.com/#/live/1166916873711665152).
 - [2025/05] vLLM is now a hosted project under PyTorch Foundation! Please find the announcement [here](https://pytorch.org/blog/pytorch-foundation-welcomes-vllm/).
 - [2025/01] 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).
 
 <details>
 <summary>Previous News</summary>
 
-- [2025/08] We hosted [vLLM Korea Meetup](https://luma.com/cgcgprmh) with Red Hat and Rebellions! We shared the latest advancements in vLLM along with project spotlights from the vLLM Korea community. Please find the meetup slides [here](https://drive.google.com/file/d/1bcrrAE1rxUgx0mjIeOWT6hNe2RefC5Hm/view).
-- [2025/08] We hosted [vLLM Beijing Meetup](https://mp.weixin.qq.com/s/dgkWg1WFpWGO2jCdTqQHxA) focusing on large-scale LLM deployment! Please find the meetup slides [here](https://drive.google.com/drive/folders/1Pid6NSFLU43DZRi0EaTcPgXsAzDvbBqF) and the recording [here](https://www.chaspark.com/#/live/1166916873711665152).
 - [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).
 - [2025/04] 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).
@@ -82,7 +76,7 @@ vLLM is flexible and easy to use with:
 - Tensor, pipeline, data and expert parallelism support for distributed inference
 - Streaming outputs
 - OpenAI-compatible API server
-- Support for NVIDIA GPUs, AMD CPUs and GPUs, Intel CPUs and GPUs, PowerPC CPUs, and TPU. Additionally, support for diverse hardware plugins such as Intel Gaudi, IBM Spyre and Huawei Ascend.
+- Support NVIDIA GPUs, AMD CPUs and GPUs, Intel CPUs and GPUs, PowerPC CPUs, TPU, and AWS Neuron
 - Prefix caching support
 - Multi-LoRA support
 
@@ -149,7 +143,6 @@ Compute Resources:
 - Trainy
 - UC Berkeley
 - UC San Diego
-- Volcengine
 
 Slack Sponsor: Anyscale
 

benchmarks/README.md

@@ -1,20 +1,802 @@
-# Benchmarks
+# Benchmarking vLLM
 
-This directory used to contain vLLM's benchmark scripts and utilities for performance testing and evaluation.
+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.
 
-## Contents
+## Dataset Overview
 
-- **Serving benchmarks**: Scripts for testing online inference performance (latency, throughput)
-- **Throughput benchmarks**: Scripts for testing offline batch inference performance
-- **Specialized benchmarks**: Tools for testing specific features like structured output, prefix caching, long document QA, request prioritization, and multi-modal inference
-- **Dataset utilities**: Framework for loading and sampling from various benchmark datasets (ShareGPT, HuggingFace datasets, synthetic data, etc.)
+<table style="width:100%; border-collapse: collapse;">
+  <thead>
+    <tr>
+      <th style="width:15%; text-align: left;">Dataset</th>
+      <th style="width:10%; text-align: center;">Online</th>
+      <th style="width:10%; text-align: center;">Offline</th>
+      <th style="width:65%; text-align: left;">Data Path</th>
+    </tr>
+  </thead>
+  <tbody>
+    <tr>
+      <td><strong>ShareGPT</strong></td>
+      <td style="text-align: center;">✅</td>
+      <td style="text-align: center;">✅</td>
+      <td><code>wget https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/resolve/main/ShareGPT_V3_unfiltered_cleaned_split.json</code></td>
+    </tr>
+    <tr>
+      <td><strong>ShareGPT4V (Image)</strong></td>
+      <td style="text-align: center;">✅</td>
+      <td style="text-align: center;">✅</td>
+      <td>
+        <code>wget https://huggingface.co/datasets/Lin-Chen/ShareGPT4V/blob/main/sharegpt4v_instruct_gpt4-vision_cap100k.json</code>
+        <br>
+        <div>Note that the images need to be downloaded separately. For example, to download COCO's 2017 Train images:</div>
+        <code>wget http://images.cocodataset.org/zips/train2017.zip</code>
+      </td>
+    </tr>
+        <tr>
+      <td><strong>ShareGPT4Video (Video)</strong></td>
+      <td style="text-align: center;">✅</td>
+      <td style="text-align: center;">✅</td>
+      <td>
+        <code>git clone https://huggingface.co/datasets/ShareGPT4Video/ShareGPT4Video</code>
+      </td>
+    </tr>
+    <tr>
+      <td><strong>BurstGPT</strong></td>
+      <td style="text-align: center;">✅</td>
+      <td style="text-align: center;">✅</td>
+      <td><code>wget https://github.com/HPMLL/BurstGPT/releases/download/v1.1/BurstGPT_without_fails_2.csv</code></td>
+    </tr>
+    <tr>
+      <td><strong>Sonnet (deprecated)</strong></td>
+      <td style="text-align: center;">✅</td>
+      <td style="text-align: center;">✅</td>
+      <td>Local file: <code>benchmarks/sonnet.txt</code></td>
+    </tr>
+    <tr>
+      <td><strong>Random</strong></td>
+      <td style="text-align: center;">✅</td>
+      <td style="text-align: center;">✅</td>
+      <td><code>synthetic</code></td>
+    </tr>
+    <tr>
+      <td><strong>RandomMultiModal (Image/Video)</strong></td>
+      <td style="text-align: center;">🟡</td>
+      <td style="text-align: center;">🚧</td>
+      <td><code>synthetic</code> </td>
+    </tr>
+    <tr>
+      <td><strong>Prefix Repetition</strong></td>
+      <td style="text-align: center;">✅</td>
+      <td style="text-align: center;">✅</td>
+      <td><code>synthetic</code></td>
+    </tr>
+    <tr>
+      <td><strong>HuggingFace-VisionArena</strong></td>
+      <td style="text-align: center;">✅</td>
+      <td style="text-align: center;">✅</td>
+      <td><code>lmarena-ai/VisionArena-Chat</code></td>
+    </tr>
+    <tr>
+      <td><strong>HuggingFace-InstructCoder</strong></td>
+      <td style="text-align: center;">✅</td>
+      <td style="text-align: center;">✅</td>
+      <td><code>likaixin/InstructCoder</code></td>
+    </tr>
+      <tr>
+      <td><strong>HuggingFace-AIMO</strong></td>
+      <td style="text-align: center;">✅</td>
+      <td style="text-align: center;">✅</td>
+      <td><code>AI-MO/aimo-validation-aime</code> , <code>AI-MO/NuminaMath-1.5</code>, <code>AI-MO/NuminaMath-CoT</code></td>
+    </tr>
+    <tr>
+      <td><strong>HuggingFace-Other</strong></td>
+      <td style="text-align: center;">✅</td>
+      <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>
 
-## Usage
+✅: supported
 
-For detailed usage instructions, examples, and dataset information, see the [Benchmark CLI documentation](https://docs.vllm.ai/en/latest/contributing/benchmarks.html#benchmark-cli).
+🟡: Partial support
 
-For full CLI reference see:
+🚧: to be supported
 
-- <https://docs.vllm.ai/en/latest/cli/bench/latency.html>
-- <https://docs.vllm.ai/en/latest/cli/bench/serve.html>
-- <https://docs.vllm.ai/en/latest/cli/bench/throughput.html>
+**Note**: HuggingFace dataset's `dataset-name` should be set to `hf`
+
+## 🚀 Example - Online Benchmark
+
+<details>
+<summary>Show more</summary>
+
+<br/>
+
+First start serving your model
+
+```bash
+vllm serve NousResearch/Hermes-3-Llama-3.1-8B
+```
+
+Then run the benchmarking script
+
+```bash
+# download dataset
+# wget https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/resolve/main/ShareGPT_V3_unfiltered_cleaned_split.json
+vllm bench serve \
+  --backend vllm \
+  --model NousResearch/Hermes-3-Llama-3.1-8B \
+  --endpoint /v1/completions \
+  --dataset-name sharegpt \
+  --dataset-path <your data path>/ShareGPT_V3_unfiltered_cleaned_split.json \
+  --num-prompts 10
+```
+
+If successful, you will see the following output
+
+```text
+============ Serving Benchmark Result ============
+Successful requests:                     10
+Benchmark duration (s):                  5.78
+Total input tokens:                      1369
+Total generated tokens:                  2212
+Request throughput (req/s):              1.73
+Output token throughput (tok/s):         382.89
+Total Token throughput (tok/s):          619.85
+---------------Time to First Token----------------
+Mean TTFT (ms):                          71.54
+Median TTFT (ms):                        73.88
+P99 TTFT (ms):                           79.49
+-----Time per Output Token (excl. 1st token)------
+Mean TPOT (ms):                          7.91
+Median TPOT (ms):                        7.96
+P99 TPOT (ms):                           8.03
+---------------Inter-token Latency----------------
+Mean ITL (ms):                           7.74
+Median ITL (ms):                         7.70
+P99 ITL (ms):                            8.39
+==================================================
+```
+
+### Custom Dataset
+
+If the dataset you want to benchmark is not supported yet in vLLM, even then you can benchmark on it using `CustomDataset`. Your data needs to be in `.jsonl` format and needs to have "prompt" field per entry, e.g., data.jsonl
+
+```json
+{"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
+```
+
+```bash
+# run benchmarking script
+vllm bench serve --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
+vllm serve Qwen/Qwen2-VL-7B-Instruct
+```
+
+```bash
+vllm bench serve \
+  --backend openai-chat \
+  --endpoint-type openai-chat \
+  --model Qwen/Qwen2-VL-7B-Instruct \
+  --endpoint /v1/chat/completions \
+  --dataset-name hf \
+  --dataset-path lmarena-ai/VisionArena-Chat \
+  --hf-split train \
+  --num-prompts 1000
+```
+
+### InstructCoder Benchmark with Speculative Decoding
+
+``` bash
+VLLM_USE_V1=1 vllm serve meta-llama/Meta-Llama-3-8B-Instruct \
+    --speculative-config $'{"method": "ngram",
+    "num_speculative_tokens": 5, "prompt_lookup_max": 5,
+    "prompt_lookup_min": 2}'
+```
+
+``` bash
+vllm bench serve \
+    --model meta-llama/Meta-Llama-3-8B-Instruct \
+    --dataset-name hf \
+    --dataset-path likaixin/InstructCoder \
+    --num-prompts 2048
+```
+
+### Other HuggingFaceDataset Examples
+
+```bash
+vllm serve Qwen/Qwen2-VL-7B-Instruct
+```
+
+`lmms-lab/LLaVA-OneVision-Data`:
+
+```bash
+vllm bench serve \
+  --backend openai-chat \
+  --endpoint-type openai-chat \
+  --model Qwen/Qwen2-VL-7B-Instruct \
+  --endpoint /v1/chat/completions \
+  --dataset-name hf \
+  --dataset-path lmms-lab/LLaVA-OneVision-Data \
+  --hf-split train \
+  --hf-subset "chart2text(cauldron)" \
+  --num-prompts 10
+```
+
+`Aeala/ShareGPT_Vicuna_unfiltered`:
+
+```bash
+vllm bench serve \
+  --backend openai-chat \
+  --endpoint-type openai-chat \
+  --model Qwen/Qwen2-VL-7B-Instruct \
+  --endpoint /v1/chat/completions \
+  --dataset-name hf \
+  --dataset-path Aeala/ShareGPT_Vicuna_unfiltered \
+  --hf-split train \
+  --num-prompts 10
+```
+
+`AI-MO/aimo-validation-aime`:
+
+``` bash
+vllm bench serve \
+    --model Qwen/QwQ-32B \
+    --dataset-name hf \
+    --dataset-path AI-MO/aimo-validation-aime \
+    --num-prompts 10 \
+    --seed 42
+```
+
+`philschmid/mt-bench`:
+
+``` bash
+vllm bench serve \
+    --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:
+
+```bash
+vllm bench serve \
+  --backend vllm \
+  --model NousResearch/Hermes-3-Llama-3.1-8B \
+  --endpoint /v1/completions \
+  --dataset-name sharegpt \
+  --dataset-path <your data path>/ShareGPT_V3_unfiltered_cleaned_split.json \
+  --top-k 10 \
+  --top-p 0.9 \
+  --temperature 0.5 \
+  --num-prompts 10
+```
+
+### 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>
+
+## 📈 Example - Offline Throughput Benchmark
+
+<details>
+<summary>Show more</summary>
+
+<br/>
+
+```bash
+vllm bench throughput \
+  --model NousResearch/Hermes-3-Llama-3.1-8B \
+  --dataset-name sonnet \
+  --dataset-path vllm/benchmarks/sonnet.txt \
+  --num-prompts 10
+```
+
+If successful, you will see the following output
+
+```text
+Throughput: 7.15 requests/s, 4656.00 total tokens/s, 1072.15 output tokens/s
+Total num prompt tokens:  5014
+Total num output tokens:  1500
+```
+
+### VisionArena Benchmark for Vision Language Models
+
+```bash
+vllm bench throughput \
+  --model Qwen/Qwen2-VL-7B-Instruct \
+  --backend vllm-chat \
+  --dataset-name hf \
+  --dataset-path lmarena-ai/VisionArena-Chat \
+  --num-prompts 1000 \
+  --hf-split train
+```
+
+The `num prompt tokens` now includes image token counts
+
+```text
+Throughput: 2.55 requests/s, 4036.92 total tokens/s, 326.90 output tokens/s
+Total num prompt tokens:  14527
+Total num output tokens:  1280
+```
+
+### InstructCoder Benchmark with Speculative Decoding
+
+``` bash
+VLLM_WORKER_MULTIPROC_METHOD=spawn \
+VLLM_USE_V1=1 \
+vllm bench throughput \
+    --dataset-name=hf \
+    --dataset-path=likaixin/InstructCoder \
+    --model=meta-llama/Meta-Llama-3-8B-Instruct \
+    --input-len=1000 \
+    --output-len=100 \
+    --num-prompts=2048 \
+    --async-engine \
+    --speculative-config $'{"method": "ngram",
+    "num_speculative_tokens": 5, "prompt_lookup_max": 5,
+    "prompt_lookup_min": 2}'
+```
+
+```text
+Throughput: 104.77 requests/s, 23836.22 total tokens/s, 10477.10 output tokens/s
+Total num prompt tokens:  261136
+Total num output tokens:  204800
+```
+
+### Other HuggingFaceDataset Examples
+
+`lmms-lab/LLaVA-OneVision-Data`:
+
+```bash
+vllm bench throughput \
+  --model Qwen/Qwen2-VL-7B-Instruct \
+  --backend vllm-chat \
+  --dataset-name hf \
+  --dataset-path lmms-lab/LLaVA-OneVision-Data \
+  --hf-split train \
+  --hf-subset "chart2text(cauldron)" \
+  --num-prompts 10
+```
+
+`Aeala/ShareGPT_Vicuna_unfiltered`:
+
+```bash
+vllm bench throughput \
+  --model Qwen/Qwen2-VL-7B-Instruct \
+  --backend vllm-chat \
+  --dataset-name hf \
+  --dataset-path Aeala/ShareGPT_Vicuna_unfiltered \
+  --hf-split train \
+  --num-prompts 10
+```
+
+`AI-MO/aimo-validation-aime`:
+
+```bash
+vllm bench throughput \
+  --model Qwen/QwQ-32B \
+  --backend vllm \
+  --dataset-name hf \
+  --dataset-path AI-MO/aimo-validation-aime \
+  --hf-split train \
+  --num-prompts 10
+```
+
+Benchmark with LoRA adapters:
+
+``` bash
+# download dataset
+# wget https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/resolve/main/ShareGPT_V3_unfiltered_cleaned_split.json
+vllm bench throughput \
+  --model meta-llama/Llama-2-7b-hf \
+  --backend vllm \
+  --dataset_path <your data path>/ShareGPT_V3_unfiltered_cleaned_split.json \
+  --dataset_name sharegpt \
+  --num-prompts 10 \
+  --max-loras 2 \
+  --max-lora-rank 8 \
+  --enable-lora \
+  --lora-path yard1/llama-2-7b-sql-lora-test
+  ```
+
+</details>
+
+## 🛠️ Example - Structured Output Benchmark
+
+<details>
+<summary>Show more</summary>
+
+<br/>
+
+Benchmark the performance of structured output generation (JSON, grammar, regex).
+
+### Server Setup
+
+```bash
+vllm serve NousResearch/Hermes-3-Llama-3.1-8B
+```
+
+### 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>
+
+## 📚 Example - Long Document QA Benchmark
+
+<details>
+<summary>Show more</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>
+
+## 🗂️ Example - Prefix Caching Benchmark
+
+<details>
+<summary>Show more</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
+```
+
+### Prefix Repetition Dataset
+
+```bash
+vllm bench serve \
+  --backend openai \
+  --model meta-llama/Llama-2-7b-chat-hf \
+  --dataset-name prefix_repetition \
+  --num-prompts 100 \
+  --prefix-repetition-prefix-len 512 \
+  --prefix-repetition-suffix-len 128 \
+  --prefix-repetition-num-prefixes 5 \
+  --prefix-repetition-output-len 128
+```
+
+</details>
+
+## ⚡ Example - Request Prioritization Benchmark
+
+<details>
+<summary>Show more</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>
+
+## 👁️ Example - Multi-Modal Benchmark
+
+<details>
+<summary>Show more</summary>
+
+<br/>
+
+Benchmark the performance of multi-modal requests in vLLM.
+
+### Images (ShareGPT4V)
+
+Start vLLM:
+
+```bash
+python -m vllm.entrypoints.openai.api_server \
+  --model Qwen/Qwen2.5-VL-7B-Instruct \
+  --dtype bfloat16 \
+  --limit-mm-per-prompt '{"image": 1}' \
+  --allowed-local-media-path /path/to/sharegpt4v/images
+```
+
+Send requests with images:
+
+```bash
+python benchmarks/benchmark_serving.py \
+  --backend openai-chat \
+  --model Qwen/Qwen2.5-VL-7B-Instruct \
+  --dataset-name sharegpt \
+  --dataset-path /path/to/ShareGPT4V/sharegpt4v_instruct_gpt4-vision_cap100k.json \
+  --num-prompts 100 \
+  --save-result \
+  --result-dir ~/vllm_benchmark_results \
+  --save-detailed \
+  --endpoint /v1/chat/completion
+```
+
+### Videos (ShareGPT4Video)
+
+Start vLLM:
+
+```bash
+python -m vllm.entrypoints.openai.api_server \
+  --model Qwen/Qwen2.5-VL-7B-Instruct \
+  --dtype bfloat16 \
+  --limit-mm-per-prompt '{"video": 1}' \
+  --allowed-local-media-path /path/to/sharegpt4video/videos
+```
+
+Send requests with videos:
+
+```bash
+python benchmarks/benchmark_serving.py \
+  --backend openai-chat \
+  --model Qwen/Qwen2.5-VL-7B-Instruct \
+  --dataset-name sharegpt \
+  --dataset-path /path/to/ShareGPT4Video/llava_v1_5_mix665k_with_video_chatgpt72k_share4video28k.json \
+  --num-prompts 100 \
+  --save-result \
+  --result-dir ~/vllm_benchmark_results \
+  --save-detailed \
+  --endpoint /v1/chat/completion
+```
+
+### Synthetic Random Images (random-mm)
+
+Generate synthetic image inputs alongside random text prompts to stress-test vision models without external datasets.
+
+Notes:
+
+- Works only with online benchmark via the OpenAI  backend (`--backend openai-chat`) and endpoint `/v1/chat/completions`.
+- Video sampling is not yet implemented.
+
+Start the server (example):
+
+```bash
+vllm serve Qwen/Qwen2.5-VL-3B-Instruct \
+  --dtype bfloat16 \
+  --max-model-len 16384 \
+  --limit-mm-per-prompt '{"image": 3, "video": 0}' \
+  --mm-processor-kwargs max_pixels=1003520
+```
+
+Benchmark. It is recommended to use the flag `--ignore-eos` to simulate real responses. You can set the size of the output via the arg `random-output-len`.
+
+Ex.1: Fixed number of items and a single image resolution, enforcing generation of approx 40 tokens:
+
+```bash
+vllm bench serve \
+  --backend openai-chat \
+  --model Qwen/Qwen2.5-VL-3B-Instruct \
+  --endpoint /v1/chat/completions \
+  --dataset-name random-mm \
+  --num-prompts 100 \
+  --max-concurrency 10 \
+  --random-prefix-len 25 \
+  --random-input-len 300 \
+  --random-output-len 40 \
+  --random-range-ratio 0.2 \
+  --random-mm-base-items-per-request 2 \
+  --random-mm-limit-mm-per-prompt '{"image": 3, "video": 0}' \
+  --random-mm-bucket-config '{(224, 224, 1): 1.0}' \
+  --request-rate inf \
+  --ignore-eos \
+  --seed 42
+```
+
+The number of items per request can be controlled by passing multiple image buckets:
+
+```bash
+  --random-mm-base-items-per-request 2 \
+  --random-mm-num-mm-items-range-ratio 0.5 \
+  --random-mm-limit-mm-per-prompt '{"image": 4, "video": 0}' \
+  --random-mm-bucket-config '{(256, 256, 1): 0.7, (720, 1280, 1): 0.3}' \
+```
+
+Flags specific to `random-mm`:
+
+- `--random-mm-base-items-per-request`: base number of multimodal items per request.
+- `--random-mm-num-mm-items-range-ratio`: vary item count uniformly in the closed integer range [floor(n·(1−r)), ceil(n·(1+r))]. Set r=0 to keep it fixed; r=1 allows 0 items.
+- `--random-mm-limit-mm-per-prompt`: per-modality hard caps, e.g. '{"image": 3, "video": 0}'.
+- `--random-mm-bucket-config`: dict mapping (H, W, T) → probability. Entries with probability 0 are removed; remaining probabilities are renormalized to sum to 1. Use T=1 for images. Set any T>1 for videos (video sampling not yet supported).
+
+Behavioral notes:
+
+- If the requested base item count cannot be satisfied under the provided per-prompt limits, the tool raises an error rather than silently clamping.
+
+How sampling works:
+
+- Determine per-request item count k by sampling uniformly from the integer range defined by `--random-mm-base-items-per-request` and `--random-mm-num-mm-items-range-ratio`, then clamp k to at most the sum of per-modality limits.
+- For each of the k items, sample a bucket (H, W, T) according to the normalized probabilities in `--random-mm-bucket-config`, while tracking how many items of each modality have been added.
+- If a modality (e.g., image) reaches its limit from `--random-mm-limit-mm-per-prompt`, all buckets of that modality are excluded and the remaining bucket probabilities are renormalized before continuing.
+This should be seen as an edge case, and if this behavior can be avoided by setting `--random-mm-limit-mm-per-prompt` to a large number. Note that this might result in errors due to engine config `--limit-mm-per-prompt`.
+- The resulting request contains synthetic image data in `multi_modal_data` (OpenAI Chat format). When `random-mm` is used with the OpenAI Chat backend, prompts remain text and MM content is attached via `multi_modal_data`.
+
+</details>

benchmarks/auto_tune/README.md

@@ -31,12 +31,6 @@ cd vllm
 
 You must set the following variables at the top of the script before execution.
 
-   Note: You can also override the default values below via environment variables when running the script.
-
-```bash
-MODEL=meta-llama/Llama-3.3-70B-Instruct SYSTEM=TPU TP=8 DOWNLOAD_DIR='' INPUT_LEN=128 OUTPUT_LEN=2048 MAX_MODEL_LEN=2300 MIN_CACHE_HIT_PCT=0 MAX_LATENCY_ALLOWED_MS=100000000000 NUM_SEQS_LIST="128 256" NUM_BATCHED_TOKENS_LIST="1024 2048 4096" VLLM_LOGGING_LEVEL=DEBUG bash auto_tune.sh
-```
-
 | Variable | Description | Example Value |
 | --- | --- | --- |
 | `BASE` | **Required.** The absolute path to the parent directory of your vLLM repository directory. | `"$HOME"` |
@@ -149,70 +143,3 @@ The script follows a systematic process to find the optimal parameters:
 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.
-
-## Batched `auto_tune`
-
-The `batch_auto_tune.sh` script allows you to run multiple `auto_tune.sh` experiments sequentially from a single configuration file. It iterates through a list of parameter sets, executes `auto_tune.sh` for each, and records the results back into the input file.
-
-### Prerequisites
-
-- **jq**: This script requires `jq` to parse the JSON configuration file.
-- **gcloud**: If you plan to upload results to Google Cloud Storage, the `gcloud` CLI must be installed and authenticated.
-
-### How to Run
-
-1. **Create a JSON configuration file**: Create a file (e.g., `runs_config.json`) containing an array of JSON objects. Each object defines the parameters for a single `auto_tune.sh` run.
-
-2. **Execute the script**:
-
-    ```bash
-    bash batch_auto_tune.sh <path_to_json_file> [gcs_upload_path]
-    ```
-
-    - `<path_to_json_file>`: **Required.** Path to your JSON configuration file.
-    - `[gcs_upload_path]`: **Optional.** A GCS path (e.g., `gs://my-bucket/benchmark-results`) where the detailed results and profiles for each run will be uploaded. If this is empty, the results will be available on the local filesystem (see the log for `RESULT_FILE=/path/to/results/file.txt`).
-
-### Configuration File
-
-The JSON configuration file should contain an array of objects. Each object's keys correspond to the configuration variables for `auto_tune.sh` (see the [Configuration table above](#configuration)). These keys will be converted to uppercase environment variables for each run.
-
-Here is an example `runs_config.json` with two benchmark configurations:
-
-```json
-[
-  {
-    "base": "/home/user",
-    "model": "meta-llama/Llama-3.1-8B-Instruct",
-    "system": "TPU", # OR GPU
-    "tp": 8,
-    "input_len": 128,
-    "output_len": 2048,
-    "max_model_len": 2300,
-    "num_seqs_list": "128 256",
-    "num_batched_tokens_list": "8192 16384"
-  },
-  {
-    "base": "/home/user",
-    "model": "meta-llama/Llama-3.1-70B-Instruct",
-    "system": "TPU", # OR GPU
-    "tp": 8,
-    "input_len": 4000,
-    "output_len": 16,
-    "max_model_len": 4096,
-    "num_seqs_list": "64 128",
-    "num_batched_tokens_list": "4096 8192",
-    "max_latency_allowed_ms": 500
-  }
-]
-```
-
-### Output
-
-The script modifies the input JSON file in place, adding the results of each run to the corresponding object. The following fields are added:
-
-- `run_id`: A unique identifier for the run, derived from the timestamp.
-- `status`: The outcome of the run (`SUCCESS`, `FAILURE`, or `WARNING_NO_RESULT_FILE`).
-- `results`: The content of the `result.txt` file from the `auto_tune.sh` run.
-- `gcs_results`: The GCS URL where the run's artifacts are stored (if a GCS path was provided).
-
-A summary of successful and failed runs is also printed to the console upon completion.

benchmarks/auto_tune/auto_tune.sh

@@ -5,41 +5,25 @@
 
 TAG=$(date +"%Y_%m_%d_%H_%M")
 SCRIPT_DIR=$( cd -- "$( dirname -- "${BASH_SOURCE[0]}" )" &> /dev/null && pwd )
-VLLM_LOGGING_LEVEL=${VLLM_LOGGING_LEVEL:-INFO}
-BASE=${BASE:-"$SCRIPT_DIR/../../.."}
-MODEL=${MODEL:-"meta-llama/Llama-3.1-8B-Instruct"}
-SYSTEM=${SYSTEM:-"TPU"}
-TP=${TP:-1}
-DOWNLOAD_DIR=${DOWNLOAD_DIR:-""}
-INPUT_LEN=${INPUT_LEN:-4000}
-OUTPUT_LEN=${OUTPUT_LEN:-16}
-MAX_MODEL_LEN=${MAX_MODEL_LEN:-4096}
-MIN_CACHE_HIT_PCT=${MIN_CACHE_HIT_PCT:-0}
-MAX_LATENCY_ALLOWED_MS=${MAX_LATENCY_ALLOWED_MS:-100000000000}
-NUM_SEQS_LIST=${NUM_SEQS_LIST:-"128 256"}
-NUM_BATCHED_TOKENS_LIST=${NUM_BATCHED_TOKENS_LIST:-"512 1024 2048 4096"}
+BASE="$SCRIPT_DIR/../../.."
+MODEL="meta-llama/Llama-3.1-8B-Instruct"
+SYSTEM="TPU"
+TP=1
+DOWNLOAD_DIR=""
+INPUT_LEN=4000
+OUTPUT_LEN=16
+MAX_MODEL_LEN=4096
+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 "====================== AUTO TUNE PARAMETERS ===================="
-echo "SCRIPT_DIR=$SCRIPT_DIR"
-echo "BASE=$BASE"
-echo "MODEL=$MODEL"
-echo "SYSTEM=$SYSTEM"
-echo "TP=$TP"
-echo "DOWNLOAD_DIR=$DOWNLOAD_DIR"
-echo "INPUT_LEN=$INPUT_LEN"
-echo "OUTPUT_LEN=$OUTPUT_LEN"
-echo "MAX_MODEL_LEN=$MAX_MODEL_LEN"
-echo "MIN_CACHE_HIT_PCT=$MIN_CACHE_HIT_PCT"
-echo "MAX_LATENCY_ALLOWED_MS=$MAX_LATENCY_ALLOWED_MS"
-echo "NUM_SEQS_LIST=$NUM_SEQS_LIST"
-echo "NUM_BATCHED_TOKENS_LIST=$NUM_BATCHED_TOKENS_LIST"
-echo "VLLM_LOGGING_LEVEL=$VLLM_LOGGING_LEVEL"
-echo "RESULT_FILE=$RESULT"
-echo "====================== AUTO TUNEPARAMETERS ===================="
+echo "result file: $RESULT"
+echo "model: $MODEL"
 
 rm -rf $LOG_FOLDER
 rm -rf $PROFILE_PATH
@@ -74,7 +58,7 @@ start_server() {
     local vllm_log=$4
     local profile_dir=$5
 
-    pkill -if "vllm serve" || true
+    pkill -if vllm
 
     # Define the common arguments as a bash array.
     # Each argument and its value are separate elements.
@@ -96,22 +80,17 @@ start_server() {
     # This correctly passes each element as a separate argument.
     if [[ -n "$profile_dir" ]]; then
         # Start server with profiling enabled
-        VLLM_SERVER_DEV_MODE=1 VLLM_TORCH_PROFILER_DIR=$profile_dir \
+        VLLM_USE_V1=1 VLLM_SERVER_DEV_MODE=1 VLLM_TORCH_PROFILER_DIR=$profile_dir \
             vllm serve "${common_args_array[@]}" > "$vllm_log" 2>&1 &
     else
         # Start server without profiling
-        VLLM_SERVER_DEV_MODE=1 \
+        VLLM_USE_V1=1 VLLM_SERVER_DEV_MODE=1 \
             vllm serve "${common_args_array[@]}" > "$vllm_log" 2>&1 &
     fi
-    local server_pid=$!
 
     # wait for 10 minutes...
     server_started=0
     for i in {1..60}; do
-        # This line checks whether the server is still alive or not,
-        # since that we should always have permission to send signal to the server process.
-        kill -0 $server_pid 2> /dev/null || break
-
         RESPONSE=$(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
@@ -123,7 +102,7 @@ start_server() {
     done
 
     if (( ! server_started )); then
-        echo "server did not start within 10 minutes or crashed. Please check server log at $vllm_log".
+        echo "server did not start within 10 minutes. Please check server log at $vllm_log".
         return 1
     else
         return 0
@@ -139,7 +118,7 @@ run_benchmark() {
     echo "vllm_log: $vllm_log"
     echo
     rm -f $vllm_log
-    pkill -if "vllm serve" || true
+    pkill -if vllm
 
     echo "starting server..."
     # Call start_server without a profile_dir to avoid profiling overhead
@@ -232,9 +211,9 @@ run_benchmark() {
 
     echo "best_max_num_seqs: $best_max_num_seqs, best_num_batched_tokens: $best_num_batched_tokens, best_throughput: $best_throughput"
 
-    pkill -if "vllm serve" || true
+    pkill -if vllm
     sleep 10
-    echo "===================="
+    printf '=%.0s' $(seq 1 20)
     return 0
 }
 
@@ -308,6 +287,6 @@ if (( $(echo "$best_throughput > 0" | bc -l) )); then
 else
     echo "No configuration met the latency requirements. Skipping final profiling run."
 fi
-pkill -if "vllm serve" || true
+pkill -if vllm
 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/auto_tune/batch_auto_tune.sh

@@ -1,128 +0,0 @@
-#!/bin/bash
-
-INPUT_JSON="$1"
-GCS_PATH="$2" # Optional GCS path for uploading results for each run
-
-SCRIPT_DIR=$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" &>/dev/null && pwd)
-AUTOTUNE_SCRIPT="$SCRIPT_DIR/auto_tune.sh"
-
-if [[ -z "$INPUT_JSON" ]]; then
-  echo "Error: Input JSON file not provided."
-  echo "Usage: $0 <path_to_json_file> [gcs_upload_path]"
-  exit 1
-fi
-
-if [[ ! -f "$INPUT_JSON" ]]; then
-  echo "Error: File not found at '$INPUT_JSON'"
-  exit 1
-fi
-
-if ! command -v jq &> /dev/null; then
-    echo "Error: 'jq' command not found. Please install jq to process the JSON input."
-    exit 1
-fi
-
-if [[ -n "$GCS_PATH" ]] && ! command -v gcloud &> /dev/null; then
-    echo "Error: 'gcloud' command not found, but a GCS_PATH was provided."
-    exit 1
-fi
-
-SUCCESS_COUNT=0
-FAILURE_COUNT=0
-FAILED_RUNS=()
-SCRIPT_START_TIME=$(date +%s)
-
-json_content=$(cat "$INPUT_JSON")
-if ! num_runs=$(echo "$json_content" | jq 'length'); then
-  echo "Error: Invalid JSON in $INPUT_JSON. 'jq' failed to get array length." >&2
-  exit 1
-fi
-
-echo "Found $num_runs benchmark configurations in $INPUT_JSON."
-echo "Starting benchmark runs..."
-echo "--------------------------------------------------"
-
-for i in $(seq 0 $(($num_runs - 1))); do
-  run_object=$(echo "$json_content" | jq ".[$i]")
-
-  RUN_START_TIME=$(date +%s)
-  ENV_VARS_ARRAY=()
-  # Dynamically create env vars from the JSON object's keys
-  for key in $(echo "$run_object" | jq -r 'keys_unsorted[]'); do
-    value=$(echo "$run_object" | jq -r ".$key")
-    var_name=$(echo "$key" | tr '[:lower:]' '[:upper:]' | tr -cd 'A-Z0-9_')
-    ENV_VARS_ARRAY+=("${var_name}=${value}")
-  done
-
-  echo "Executing run #$((i+1))/$num_runs with parameters: ${ENV_VARS_ARRAY[*]}"
-
-  # Execute auto_tune.sh and capture output
-  RUN_OUTPUT_FILE=$(mktemp)
-  if env "${ENV_VARS_ARRAY[@]}" bash "$AUTOTUNE_SCRIPT" > >(tee -a "$RUN_OUTPUT_FILE") 2>&1; then
-    STATUS="SUCCESS"
-    ((SUCCESS_COUNT++))
-  else
-    STATUS="FAILURE"
-    ((FAILURE_COUNT++))
-    FAILED_RUNS+=("Run #$((i+1)): $(echo $run_object | jq -c .)")
-  fi
-
-  RUN_OUTPUT=$(<"$RUN_OUTPUT_FILE")
-  rm "$RUN_OUTPUT_FILE"
-
-  # Parse results and optionally upload them to GCS
-  RUN_ID=""
-  RESULTS=""
-  GCS_RESULTS_URL=""
-  if [[ "$STATUS" == "SUCCESS" ]]; then
-    RESULT_FILE_PATH=$(echo "$RUN_OUTPUT" | grep 'RESULT_FILE=' | tail -n 1 | cut -d'=' -f2 | tr -s '/' || true)
-
-    if [[ -n "$RESULT_FILE_PATH" && -f "$RESULT_FILE_PATH" ]]; then
-      RUN_ID=$(basename "$(dirname "$RESULT_FILE_PATH")")
-      RESULT_DIR=$(dirname "$RESULT_FILE_PATH")
-      RESULTS=$(cat "$RESULT_FILE_PATH")
-
-      if [[ -n "$GCS_PATH" ]]; then
-        GCS_RESULTS_URL="${GCS_PATH}/${RUN_ID}"
-        echo "Uploading results to GCS..."
-        if gcloud storage rsync --recursive "$RESULT_DIR/" "$GCS_RESULTS_URL"; then
-          echo "GCS upload successful."
-        else
-          echo "Warning: GCS upload failed for RUN_ID $RUN_ID."
-        fi
-      fi
-    else
-      echo "Warning: Could not find result file for a successful run."
-      STATUS="WARNING_NO_RESULT_FILE"
-    fi
-  fi
-
-  # Add the results back into the JSON object for this run
-  json_content=$(echo "$json_content" | jq --argjson i "$i" --arg run_id "$RUN_ID" --arg status "$STATUS" --arg results "$RESULTS" --arg gcs_results "$GCS_RESULTS_URL" \
-    '.[$i] += {run_id: $run_id, status: $status, results: $results, gcs_results: $gcs_results}')
-
-  RUN_END_TIME=$(date +%s)
-  echo "Run finished in $((RUN_END_TIME - RUN_START_TIME)) seconds. Status: $STATUS"
-  echo "--------------------------------------------------"
-
-  # Save intermediate progress back to the file
-  echo "$json_content" > "$INPUT_JSON.tmp" && mv "$INPUT_JSON.tmp" "$INPUT_JSON"
-
-done
-
-SCRIPT_END_TIME=$(date +%s)
-echo "All benchmark runs completed in $((SCRIPT_END_TIME - SCRIPT_START_TIME)) seconds."
-echo
-echo "====================== SUMMARY ======================"
-echo "Successful runs: $SUCCESS_COUNT"
-echo "Failed runs:     $FAILURE_COUNT"
-echo "==================================================="
-
-if [[ $FAILURE_COUNT -gt 0 ]]; then
-  echo "Details of failed runs (see JSON file for full parameters):"
-  for failed in "${FAILED_RUNS[@]}"; do
-    echo "  - $failed"
-  done
-fi
-
-echo "Updated results have been saved to '$INPUT_JSON'."

benchmarks/backend_request_func.py

@@ -8,6 +8,7 @@ import sys
 import time
 import traceback
 from dataclasses import dataclass, field
+from typing import Optional, Union
 
 import aiohttp
 import huggingface_hub.constants
@@ -27,13 +28,13 @@ class RequestFuncInput:
     prompt_len: int
     output_len: int
     model: str
-    model_name: str | None = None
-    logprobs: int | None = None
-    extra_body: dict | None = None
-    multi_modal_content: dict | list[dict] | None = None
+    model_name: Optional[str] = None
+    logprobs: Optional[int] = None
+    extra_body: Optional[dict] = None
+    multi_modal_content: Optional[dict | list[dict]] = None
     ignore_eos: bool = False
-    language: str | None = None
-    request_id: str | None = None
+    language: Optional[str] = None
+    request_id: Optional[str] = None
 
 
 @dataclass
@@ -51,7 +52,7 @@ class RequestFuncOutput:
 
 async def async_request_tgi(
     request_func_input: RequestFuncInput,
-    pbar: tqdm | None = None,
+    pbar: Optional[tqdm] = None,
 ) -> RequestFuncOutput:
     api_url = request_func_input.api_url
     assert api_url.endswith("generate_stream")
@@ -132,7 +133,7 @@ async def async_request_tgi(
 
 async def async_request_trt_llm(
     request_func_input: RequestFuncInput,
-    pbar: tqdm | None = None,
+    pbar: Optional[tqdm] = None,
 ) -> RequestFuncOutput:
     api_url = request_func_input.api_url
     assert api_url.endswith("generate_stream")
@@ -203,7 +204,7 @@ async def async_request_trt_llm(
 
 async def async_request_deepspeed_mii(
     request_func_input: RequestFuncInput,
-    pbar: tqdm | None = None,
+    pbar: Optional[tqdm] = None,
 ) -> RequestFuncOutput:
     api_url = request_func_input.api_url
     assert api_url.endswith(("completions", "profile")), (
@@ -266,7 +267,7 @@ async def async_request_deepspeed_mii(
 
 async def async_request_openai_completions(
     request_func_input: RequestFuncInput,
-    pbar: tqdm | None = None,
+    pbar: Optional[tqdm] = None,
 ) -> RequestFuncOutput:
     api_url = request_func_input.api_url
     assert api_url.endswith(("completions", "profile")), (
@@ -366,7 +367,7 @@ async def async_request_openai_completions(
 
 async def async_request_openai_chat_completions(
     request_func_input: RequestFuncInput,
-    pbar: tqdm | None = None,
+    pbar: Optional[tqdm] = None,
 ) -> RequestFuncOutput:
     api_url = request_func_input.api_url
     assert api_url.endswith(("chat/completions", "profile")), (
@@ -475,7 +476,7 @@ async def async_request_openai_chat_completions(
 
 async def async_request_openai_audio(
     request_func_input: RequestFuncInput,
-    pbar: tqdm | None = None,
+    pbar: Optional[tqdm] = None,
 ) -> RequestFuncOutput:
     # Lazy import without PlaceholderModule to avoid vllm dep.
     import soundfile
@@ -609,7 +610,7 @@ def get_tokenizer(
     tokenizer_mode: str = "auto",
     trust_remote_code: bool = False,
     **kwargs,
-) -> PreTrainedTokenizer | PreTrainedTokenizerFast:
+) -> Union[PreTrainedTokenizer, PreTrainedTokenizerFast]:
     if pretrained_model_name_or_path is not None and not os.path.exists(
         pretrained_model_name_or_path
     ):

benchmarks/benchmark_block_pool.py

@@ -2,10 +2,10 @@
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 import gc
 
-from benchmark_utils import TimeCollector
 from tabulate import tabulate
 
-from vllm.utils.argparse_utils import FlexibleArgumentParser
+from benchmark_utils import TimeCollector
+from vllm.utils import FlexibleArgumentParser
 from vllm.v1.core.block_pool import BlockPool
 
 
@@ -57,7 +57,7 @@ def invoke_main() -> None:
         "--num-iteration",
         type=int,
         default=1000,
-        help="Number of iterations to run to stabilize final data readings",
+        help="Number of iterations to run to stablize final data readings",
     )
     parser.add_argument(
         "--allocate-blocks",

benchmarks/benchmark_latency.py

@@ -1,17 +1,191 @@
 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-import sys
+"""Benchmark the latency of processing a single batch of requests."""
+
+import argparse
+import dataclasses
+import json
+import os
+import time
+from typing import Any, Optional
+
+import numpy as np
+from tqdm import tqdm
+from typing_extensions import deprecated
+
+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
+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:
+    pt_records = convert_to_pytorch_benchmark_format(
+        args=args,
+        metrics={"latency": results["latencies"]},
+        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)
+
+
+@deprecated(
+    "benchmark_latency.py is deprecated and will be removed in a "
+    "future version. Please use 'vllm bench latency' instead.",
+)
+def main(args: argparse.Namespace):
+    print(args)
+
+    engine_args = EngineArgs.from_cli_args(args)
+
+    # NOTE(woosuk): If the request cannot be processed in a single batch,
+    # the engine will automatically process the request in multiple batches.
+    llm = LLM(**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."
+    )
+
+    sampling_params = SamplingParams(
+        n=args.n,
+        temperature=1.0,
+        top_p=1.0,
+        ignore_eos=True,
+        max_tokens=args.output_len,
+        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()
+    ]
+
+    def llm_generate():
+        if not args.use_beam_search:
+            llm.generate(dummy_prompts, sampling_params=sampling_params, use_tqdm=False)
+        else:
+            llm.beam_search(
+                dummy_prompts,
+                BeamSearchParams(
+                    beam_width=args.n,
+                    max_tokens=args.output_len,
+                    ignore_eos=True,
+                ),
+            )
+
+    def run_to_completion(profile_dir: Optional[str] = None):
+        if profile_dir:
+            llm.start_profile()
+            llm_generate()
+            llm.stop_profile()
+        else:
+            start_time = time.perf_counter()
+            llm_generate()
+            end_time = time.perf_counter()
+            latency = end_time - start_time
+            return latency
+
+    print("Warming up...")
+    for _ in tqdm(range(args.num_iters_warmup), desc="Warmup iterations"):
+        run_to_completion(profile_dir=None)
+
+    if args.profile:
+        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
+
+    # Benchmark.
+    latencies = []
+    for _ in tqdm(range(args.num_iters), desc="Profiling iterations"):
+        latencies.append(run_to_completion(profile_dir=None))
+    latencies = np.array(latencies)
+    percentages = [10, 25, 50, 75, 90, 99]
+    percentiles = np.percentile(latencies, percentages)
+    print(f"Avg latency: {np.mean(latencies)} seconds")
+    for percentage, percentile in zip(percentages, percentiles):
+        print(f"{percentage}% percentile latency: {percentile} seconds")
+
+    # Output JSON results if specified
+    if args.output_json:
+        results = {
+            "avg_latency": np.mean(latencies),
+            "latencies": latencies.tolist(),
+            "percentiles": dict(zip(percentages, percentiles.tolist())),
+        }
+        with open(args.output_json, "w") as f:
+            json.dump(results, f, indent=4)
+        save_to_pytorch_benchmark_format(args, results)
+
+
+def create_argument_parser():
+    parser = FlexibleArgumentParser(
+        description="Benchmark the latency of processing a single batch of "
+        "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)
+    parser.add_argument(
+        "--n",
+        type=int,
+        default=1,
+        help="Number of generated sequences per prompt.",
+    )
+    parser.add_argument("--use-beam-search", action="store_true")
+    parser.add_argument(
+        "--num-iters-warmup",
+        type=int,
+        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(
+        "--profile",
+        action="store_true",
+        help="profile the generation process of a single batch",
+    )
+    parser.add_argument(
+        "--output-json",
+        type=str,
+        default=None,
+        help="Path to save the latency 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)"
+        ),
+    )
+
+    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__":
-    print("""DEPRECATED: This script has been moved to the vLLM CLI.
-
-Please use the following command instead:
-    vllm bench latency
-
-For help with the new command, run:
-    vllm bench latency --help
-
-Alternatively, you can run the new command directly with:
-    python -m vllm.entrypoints.cli.main bench latency --help
-""")
-    sys.exit(1)
+    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

@@ -46,7 +46,7 @@ import time
 
 from vllm import LLM, SamplingParams
 from vllm.engine.arg_utils import EngineArgs
-from vllm.utils.argparse_utils import FlexibleArgumentParser
+from vllm.utils import FlexibleArgumentParser
 
 
 def test_long_document_qa(llm=None, sampling_params=None, prompts=None):

benchmarks/benchmark_ngram_proposer.py

@@ -1,31 +1,17 @@
 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 import gc
-import time
-from unittest import mock
 
 import numpy as np
-from benchmark_utils import TimeCollector
 from tabulate import tabulate
 
-from vllm.config import (
-    CacheConfig,
-    DeviceConfig,
-    LoadConfig,
-    ModelConfig,
-    ParallelConfig,
-    SchedulerConfig,
-    SpeculativeConfig,
-    VllmConfig,
-)
-from vllm.platforms import current_platform
-from vllm.utils.argparse_utils import FlexibleArgumentParser
+from benchmark_utils import TimeCollector
+from vllm.config import ModelConfig, SpeculativeConfig, VllmConfig
+from vllm.utils import FlexibleArgumentParser
 from vllm.v1.spec_decode.ngram_proposer import NgramProposer
-from vllm.v1.worker.gpu_input_batch import InputBatch
-from vllm.v1.worker.gpu_model_runner import GPUModelRunner
 
 
-def benchmark_propose(args):
+def main(args):
     rows = []
     for max_ngram in args.max_ngram:
         collector = TimeCollector(TimeCollector.US)
@@ -83,93 +69,15 @@ def benchmark_propose(args):
     )
 
 
-def benchmark_batched_propose(args):
-    NUM_SPECULATIVE_TOKENS_NGRAM = 10
-    PROMPT_LOOKUP_MIN = 5
-    PROMPT_LOOKUP_MAX = 15
-    MAX_MODEL_LEN = int(1e7)
-    DEVICE = current_platform.device_type
-
-    model_config = ModelConfig(model="facebook/opt-125m", runner="generate")
-
-    speculative_config = SpeculativeConfig(
-        target_model_config=model_config,
-        target_parallel_config=ParallelConfig(),
-        method="ngram",
-        num_speculative_tokens=NUM_SPECULATIVE_TOKENS_NGRAM,
-        prompt_lookup_max=PROMPT_LOOKUP_MAX,
-        prompt_lookup_min=PROMPT_LOOKUP_MIN,
-    )
-
-    vllm_config = VllmConfig(
-        model_config=model_config,
-        cache_config=CacheConfig(),
-        speculative_config=speculative_config,
-        device_config=DeviceConfig(device=current_platform.device_type),
-        parallel_config=ParallelConfig(),
-        load_config=LoadConfig(),
-        scheduler_config=SchedulerConfig(),
-    )
-
-    # monkey patch vllm.v1.worker.gpu_model_runner.get_pp_group
-    mock_pp_group = mock.MagicMock()
-    mock_pp_group.world_size = 1
-    with mock.patch(
-        "vllm.v1.worker.gpu_model_runner.get_pp_group", return_value=mock_pp_group
-    ):
-        runner = GPUModelRunner(vllm_config, DEVICE)
-
-        # hack max model len
-        runner.max_model_len = MAX_MODEL_LEN
-        runner.drafter.max_model_len = MAX_MODEL_LEN
-
-        dummy_input_batch = InputBatch(
-            max_num_reqs=args.num_req,
-            max_model_len=MAX_MODEL_LEN,
-            max_num_batched_tokens=args.num_req * args.num_token,
-            device=DEVICE,
-            pin_memory=False,
-            vocab_size=256000,
-            block_sizes=[16],
-        )
-        dummy_input_batch._req_ids = list(str(id) for id in range(args.num_req))
-        dummy_input_batch.spec_decode_unsupported_reqs = ()
-        dummy_input_batch.num_tokens_no_spec = [args.num_token] * args.num_req
-        dummy_input_batch.token_ids_cpu = np.random.randint(
-            0, 20, (args.num_req, args.num_token)
-        )
-
-        runner.input_batch = dummy_input_batch
-
-        sampled_token_ids = [[0]] * args.num_req
-
-        print("Starting benchmark")
-        # first run is warmup so ignore it
-        for _ in range(args.num_iteration):
-            start = time.time()
-            runner.drafter.propose(
-                sampled_token_ids,
-                dummy_input_batch.req_ids,
-                dummy_input_batch.num_tokens_no_spec,
-                dummy_input_batch.token_ids_cpu,
-                dummy_input_batch.spec_decode_unsupported_reqs,
-            )
-            end = time.time()
-            print(f"Iteration time (s): {end - start}")
-
-
 def invoke_main() -> None:
     parser = FlexibleArgumentParser(
         description="Benchmark the performance of N-gram speculative decode drafting"
     )
-    parser.add_argument(
-        "--batched", action="store_true", help="consider time to prepare batch"
-    )
     parser.add_argument(
         "--num-iteration",
         type=int,
         default=100,
-        help="Number of iterations to run to stabilize final data readings",
+        help="Number of iterations to run to stablize final data readings",
     )
     parser.add_argument(
         "--num-req", type=int, default=128, help="Number of requests in the batch"
@@ -197,17 +105,8 @@ def invoke_main() -> None:
         help="Number of speculative tokens to generate",
     )
     args = parser.parse_args()
-
-    if not args.batched:
-        benchmark_propose(args)
-    else:
-        benchmark_batched_propose(args)
+    main(args)
 
 
-"""
-# Example command lines:
-# time python3 benchmarks/benchmark_ngram_proposer.py
-# time python3 benchmarks/benchmark_ngram_proposer.py --batched --num-iteration 4 --num-token 1000000 --num-req 128
-"""  # noqa: E501
 if __name__ == "__main__":
     invoke_main()  # pragma: no cover

benchmarks/benchmark_prefix_caching.py

@@ -32,12 +32,13 @@ import dataclasses
 import json
 import random
 import time
+from typing import Optional
 
 from transformers import PreTrainedTokenizerBase
 
 from vllm import LLM, SamplingParams
 from vllm.engine.arg_utils import EngineArgs
-from vllm.utils.argparse_utils import FlexibleArgumentParser
+from vllm.utils import FlexibleArgumentParser
 
 try:
     from vllm.transformers_utils.tokenizer import get_tokenizer
@@ -79,7 +80,7 @@ def sample_requests_from_dataset(
     num_requests: int,
     tokenizer: PreTrainedTokenizerBase,
     input_length_range: tuple[int, int],
-    fixed_output_len: int | None,
+    fixed_output_len: Optional[int],
 ) -> list[Request]:
     if fixed_output_len is not None and fixed_output_len < 4:
         raise ValueError("output_len too small")
@@ -127,7 +128,7 @@ def sample_requests_from_random(
     num_requests: int,
     tokenizer: PreTrainedTokenizerBase,
     input_length_range: tuple[int, int],
-    fixed_output_len: int | None,
+    fixed_output_len: Optional[int],
     prefix_len: int,
 ) -> list[Request]:
     requests = []

benchmarks/benchmark_prioritization.py

@@ -7,11 +7,12 @@ import dataclasses
 import json
 import random
 import time
+from typing import Optional
 
 from transformers import AutoTokenizer, PreTrainedTokenizerBase
 
 from vllm.engine.arg_utils import EngineArgs
-from vllm.utils.argparse_utils import FlexibleArgumentParser
+from vllm.utils import FlexibleArgumentParser
 
 
 # Select a equi-probable random priority
@@ -23,7 +24,7 @@ def sample_requests(
     dataset_path: str,
     num_requests: int,
     tokenizer: PreTrainedTokenizerBase,
-    fixed_output_len: int | None,
+    fixed_output_len: Optional[int],
 ) -> list[tuple[str, int, int, int]]:
     if fixed_output_len is not None and fixed_output_len < 4:
         raise ValueError("output_len too small")

benchmarks/benchmark_serving_structured_output.py

@@ -31,19 +31,20 @@ import time
 import uuid
 import warnings
 from collections.abc import AsyncGenerator
-from contextlib import nullcontext
 from dataclasses import dataclass
+from typing import Optional
 
 import datasets
 import numpy as np
 import pandas as pd
+from tqdm.asyncio import tqdm
+from transformers import PreTrainedTokenizerBase
+
 from backend_request_func import (
     ASYNC_REQUEST_FUNCS,
     RequestFuncInput,
     RequestFuncOutput,
 )
-from tqdm.asyncio import tqdm
-from transformers import PreTrainedTokenizerBase
 
 try:
     from vllm.transformers_utils.tokenizer import get_tokenizer
@@ -51,7 +52,7 @@ except ImportError:
     from backend_request_func import get_tokenizer
 
 try:
-    from vllm.utils.argparse_utils import FlexibleArgumentParser
+    from vllm.utils import FlexibleArgumentParser
 except ImportError:
     from argparse import ArgumentParser as FlexibleArgumentParser
 
@@ -316,7 +317,7 @@ def calculate_metrics(
     tokenizer: PreTrainedTokenizerBase,
     selected_percentile_metrics: list[str],
     selected_percentiles: list[float],
-    goodput_config_dict: dict[str, float] | None = None,
+    goodput_config_dict: Optional[dict[str, float]] = None,
 ) -> tuple[BenchmarkMetrics, list[int]]:
     actual_output_lens: list[int] = []
     total_input = 0
@@ -436,9 +437,9 @@ async def benchmark(
     selected_percentile_metrics: list[str],
     selected_percentiles: list[str],
     ignore_eos: bool,
-    max_concurrency: int | None,
+    max_concurrency: Optional[int],
     structured_output_ratio: float,
-    goodput_config_dict: dict[str, float] | None = None,
+    goodput_config_dict: Optional[dict[str, float]] = None,
 ):
     if backend in ASYNC_REQUEST_FUNCS:
         request_func = ASYNC_REQUEST_FUNCS[backend]
@@ -448,8 +449,7 @@ async def benchmark(
     def prepare_extra_body(request) -> dict:
         extra_body = {}
         # Add the schema to the extra_body
-        extra_body["structured_outputs"] = {}
-        extra_body["structured_outputs"][request.structure_type] = request.schema
+        extra_body[request.structure_type] = request.schema
         return extra_body
 
     print("Starting initial single prompt test run...")
@@ -502,9 +502,15 @@ async def benchmark(
 
     pbar = None if disable_tqdm else tqdm(total=len(input_requests))
 
-    semaphore = asyncio.Semaphore(max_concurrency) if max_concurrency else nullcontext()
+    # This can be used once the minimum Python version is 3.10 or higher,
+    # and it will simplify the code in limited_request_func.
+    #    semaphore = (asyncio.Semaphore(max_concurrency)
+    #                 if max_concurrency else contextlib.nullcontext())
+    semaphore = asyncio.Semaphore(max_concurrency) if max_concurrency else None
 
     async def limited_request_func(request_func_input, pbar):
+        if semaphore is None:
+            return await request_func(request_func_input=request_func_input, pbar=pbar)
         async with semaphore:
             return await request_func(request_func_input=request_func_input, pbar=pbar)
 
@@ -690,11 +696,11 @@ def evaluate(ret, args):
         return re.match(args.regex, actual) is not None
 
     def _eval_correctness(expected, actual):
-        if args.structure_type == "json":
+        if args.structure_type == "guided_json":
             return _eval_correctness_json(expected, actual)
-        elif args.structure_type == "regex":
+        elif args.structure_type == "guided_regex":
             return _eval_correctness_regex(expected, actual)
-        elif args.structure_type == "choice":
+        elif args.structure_type == "guided_choice":
             return _eval_correctness_choice(expected, actual)
         else:
             return None
@@ -774,18 +780,18 @@ def main(args: argparse.Namespace):
     )
 
     if args.dataset == "grammar":
-        args.structure_type = "grammar"
+        args.structure_type = "guided_grammar"
     elif args.dataset == "regex":
-        args.structure_type = "regex"
+        args.structure_type = "guided_regex"
     elif args.dataset == "choice":
-        args.structure_type = "choice"
+        args.structure_type = "guided_choice"
     else:
-        args.structure_type = "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}so"
+        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]}"
@@ -903,13 +909,13 @@ def create_argument_parser():
     parser.add_argument(
         "--tokenizer",
         type=str,
-        help="Name or path of the tokenizer, if not using the default tokenizer.",
+        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.",
+        help="Name or path of the tokenizer, if not using the default tokenizer.",  # noqa: E501
     )
     parser.add_argument(
         "--num-prompts",
@@ -992,7 +998,7 @@ def create_argument_parser():
         "--percentile-metrics",
         type=str,
         default="ttft,tpot,itl",
-        help="Comma-separated list of selected metrics to report percentiles. "
+        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".',

benchmarks/benchmark_throughput.py

@@ -1,17 +1,741 @@
 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-import sys
+"""Benchmark offline inference throughput."""
+
+import argparse
+import dataclasses
+import json
+import os
+import random
+import time
+import warnings
+from typing import Any, Optional, Union
+
+import torch
+import uvloop
+from tqdm import tqdm
+from transformers import AutoModelForCausalLM, AutoTokenizer, PreTrainedTokenizerBase
+from typing_extensions import deprecated
+
+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,
+)
+from vllm.inputs import TextPrompt, TokensPrompt
+from vllm.lora.request import LoRARequest
+from vllm.outputs import RequestOutput
+from vllm.sampling_params import BeamSearchParams
+from vllm.utils import FlexibleArgumentParser, merge_async_iterators
+
+
+def run_vllm(
+    requests: list[SampleRequest],
+    n: int,
+    engine_args: EngineArgs,
+    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."
+    )
+    # 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
+            )
+        )
+        sampling_params.append(
+            SamplingParams(
+                n=n,
+                temperature=1.0,
+                top_p=1.0,
+                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]
+
+    use_beam_search = False
+
+    outputs = None
+    if not use_beam_search:
+        start = time.perf_counter()
+        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"
+        # output_len should be the same for all requests.
+        output_len = requests[0].expected_output_len
+        for request in requests:
+            assert request.expected_output_len == output_len
+        start = time.perf_counter()
+        llm.beam_search(
+            prompts,
+            BeamSearchParams(
+                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]]:
+    """
+    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."
+    )
+
+    prompts = []
+    sampling_params: list[SamplingParams] = []
+    for request in requests:
+        prompts.append(request.prompt)
+        sampling_params.append(
+            SamplingParams(
+                n=n,
+                temperature=1.0,
+                top_p=1.0,
+                ignore_eos=True,
+                max_tokens=request.expected_output_len,
+                detokenize=not disable_detokenize,
+            )
+        )
+    start = time.perf_counter()
+    outputs = llm.chat(prompts, sampling_params, use_tqdm=True)
+    end = time.perf_counter()
+    return end - start, outputs
+
+
+async def run_vllm_async(
+    requests: list[SampleRequest],
+    n: int,
+    engine_args: AsyncEngineArgs,
+    disable_frontend_multiprocessing: bool = False,
+    disable_detokenize: bool = False,
+) -> float:
+    from vllm import SamplingParams
+
+    async with build_async_engine_client_from_engine_args(
+        engine_args,
+        disable_frontend_multiprocessing=disable_frontend_multiprocessing,
+    ) as llm:
+        model_config = await llm.get_model_config()
+        assert all(
+            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] = []
+        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
+                )
+            )
+            sampling_params.append(
+                SamplingParams(
+                    n=n,
+                    temperature=1.0,
+                    top_p=1.0,
+                    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}")
+            generators.append(generator)
+        all_gens = merge_async_iterators(*generators)
+        async for i, res in all_gens:
+            pass
+        end = time.perf_counter()
+        return end - start
+
+
+def run_hf(
+    requests: list[SampleRequest],
+    model: str,
+    tokenizer: PreTrainedTokenizerBase,
+    n: int,
+    max_batch_size: int,
+    trust_remote_code: bool,
+    disable_detokenize: bool = False,
+) -> float:
+    llm = AutoModelForCausalLM.from_pretrained(
+        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
+    llm = llm.cuda()
+
+    pbar = tqdm(total=len(requests))
+    start = time.perf_counter()
+    batch: list[str] = []
+    max_prompt_len = 0
+    max_output_len = 0
+    for i in range(len(requests)):
+        prompt = requests[i].prompt
+        prompt_len = requests[i].prompt_len
+        output_len = requests[i].expected_output_len
+        # Add the prompt to the batch.
+        batch.append(prompt)
+        max_prompt_len = max(max_prompt_len, prompt_len)
+        max_output_len = max(max_output_len, output_len)
+        if len(batch) < max_batch_size and i != len(requests) - 1:
+            # 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:
+                # We can add more requests to the batch.
+                continue
+
+        # Generate the sequences.
+        input_ids = tokenizer(batch, return_tensors="pt", padding=True).input_ids
+        llm_outputs = llm.generate(
+            input_ids=input_ids.cuda(),
+            do_sample=True,
+            num_return_sequences=n,
+            temperature=1.0,
+            top_p=1.0,
+            use_cache=True,
+            max_new_tokens=max_output_len,
+        )
+        if not disable_detokenize:
+            # Include the decoding time.
+            tokenizer.batch_decode(llm_outputs, skip_special_tokens=True)
+        pbar.update(len(batch))
+
+        # Clear the batch.
+        batch = []
+        max_prompt_len = 0
+        max_output_len = 0
+    end = time.perf_counter()
+    return end - start
+
+
+def run_mii(
+    requests: list[SampleRequest],
+    model: str,
+    tensor_parallel_size: int,
+    output_len: int,
+) -> float:
+    from mii import client, serve
+
+    llm = serve(model, tensor_parallel=tensor_parallel_size)
+    prompts = [request.prompt for request in requests]
+
+    start = time.perf_counter()
+    llm.generate(prompts, max_new_tokens=output_len)
+    end = time.perf_counter()
+    client = client(model)
+    client.terminate_server()
+    return end - start
+
+
+def save_to_pytorch_benchmark_format(
+    args: argparse.Namespace, results: dict[str, Any]
+) -> None:
+    pt_records = convert_to_pytorch_benchmark_format(
+        args=args,
+        metrics={
+            "requests_per_second": [results["requests_per_second"]],
+            "tokens_per_second": [results["tokens_per_second"]],
+        },
+        extra_info={
+            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"
+        write_to_json(pt_file, pt_records)
+
+
+def get_requests(args, tokenizer):
+    # Common parameters for all dataset types.
+    common_kwargs = {
+        "dataset_path": args.dataset_path,
+        "random_seed": args.seed,
+    }
+    sample_kwargs = {
+        "tokenizer": tokenizer,
+        "lora_path": args.lora_path,
+        "max_loras": args.max_loras,
+        "num_requests": args.num_prompts,
+        "input_len": args.input_len,
+        "output_len": args.output_len,
+    }
+
+    if args.dataset_path is None or args.dataset_name == "random":
+        sample_kwargs["range_ratio"] = args.random_range_ratio
+        sample_kwargs["prefix_len"] = args.prefix_len
+        dataset_cls = RandomDataset
+    elif args.dataset_name == "sharegpt":
+        dataset_cls = ShareGPTDataset
+        if args.backend == "vllm-chat":
+            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."
+        )
+        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"
+            sample_kwargs["enable_multimodal_chat"] = True
+        elif args.dataset_path in InstructCoderDataset.SUPPORTED_DATASET_PATHS:
+            dataset_cls = InstructCoderDataset
+            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
+            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"
+    else:
+        raise ValueError(f"Unknown dataset name: {args.dataset_name}")
+    # Remove None values
+    sample_kwargs = {k: v for k, v in sample_kwargs.items() if v is not None}
+    return dataset_cls(**common_kwargs).sample(**sample_kwargs)
+
+
+@deprecated(
+    "benchmark_throughput.py is deprecated and will be removed in a "
+    "future version. Please use 'vllm bench throughput' instead.",
+)
+def main(args: argparse.Namespace):
+    if args.seed is None:
+        args.seed = 0
+    print(args)
+    random.seed(args.seed)
+    # Sample the requests.
+    tokenizer = AutoTokenizer.from_pretrained(
+        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)
+    request_outputs: Optional[list[RequestOutput]] = None
+    if args.backend == "vllm":
+        if args.async_engine:
+            elapsed_time = uvloop.run(
+                run_vllm_async(
+                    requests,
+                    args.n,
+                    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,
+            )
+    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,
+        )
+    elif args.backend == "mii":
+        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
+        )
+    else:
+        raise ValueError(f"Unknown backend: {args.backend}")
+
+    if request_outputs:
+        # Note: with the vllm and vllm-chat backends,
+        # we have request_outputs, which we use to count tokens.
+        total_prompt_tokens = 0
+        total_output_tokens = 0
+        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_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_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."
+        )
+        # 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"Total num prompt tokens:  {total_prompt_tokens}")
+    print(f"Total num output tokens:  {total_output_tokens}")
+
+    # Output JSON results if specified
+    if args.output_json:
+        results = {
+            "elapsed_time": elapsed_time,
+            "num_requests": len(requests),
+            "total_num_tokens": total_num_tokens,
+            "requests_per_second": len(requests) / elapsed_time,
+            "tokens_per_second": total_num_tokens / elapsed_time,
+        }
+        with open(args.output_json, "w") as f:
+            json.dump(results, f, indent=4)
+        save_to_pytorch_benchmark_format(args, results)
+
+
+def validate_args(args):
+    """
+    Validate command-line arguments.
+    """
+
+    # === Deprecation and Defaulting ===
+    if args.dataset is not None:
+        warnings.warn(
+            "The '--dataset' argument will be deprecated in the next release. "
+            "Please use '--dataset-name' and '--dataset-path' instead.",
+            stacklevel=2,
+        )
+        args.dataset_path = args.dataset
+
+    if not getattr(args, "tokenizer", None):
+        args.tokenizer = args.model
+
+    # === Backend Validation ===
+    valid_backends = {"vllm", "hf", "mii", "vllm-chat"}
+    if args.backend not in valid_backends:
+        raise ValueError(f"Unsupported backend: {args.backend}")
+
+    # === 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"
+        if args.input_len is None:
+            raise ValueError("input_len must be provided for a random dataset")
+
+    # === Dataset Name Specific Checks ===
+    # --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 \
+                since --dataset-name is not 'hf'.",
+            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
+        else:
+            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 \
+                --dataset-name is not 'random'.",
+            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\
+                 is not 'random', 'sonnet', or not set.",
+            stacklevel=2,
+        )
+
+    # === LoRA Settings ===
+    if getattr(args, "enable_lora", False) and args.backend != "vllm":
+        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")
+
+    # === Backend-specific Validations ===
+    if args.backend == "hf" and args.hf_max_batch_size is None:
+        raise ValueError("HF max batch size is required for HF backend")
+    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
+    ):
+        raise ValueError("Quantization is only for vLLM backend.")
+
+    if args.backend == "mii" and args.dtype != "auto":
+        raise ValueError("dtype must be auto for MII backend.")
+    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.")
+
+    # --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"
+        )
+
+
+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(
+        "--dataset-name",
+        type=str,
+        choices=["sharegpt", "random", "sonnet", "burstgpt", "hf"],
+        help="Name of the dataset to benchmark on.",
+        default="sharegpt",
+    )
+    parser.add_argument(
+        "--no-stream",
+        action="store_true",
+        help="Do not load the dataset in streaming mode.",
+    )
+    parser.add_argument(
+        "--dataset",
+        type=str,
+        default=None,
+        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.",
+    )
+    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.",
+    )
+    parser.add_argument(
+        "--disable-detokenize",
+        action="store_true",
+        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.",
+    )
+    parser.add_argument(
+        "--prefix-len",
+        type=int,
+        default=None,
+        help=f"Number of prefix tokens to be used in RandomDataset "
+        "and SonnetDataset. For RandomDataset, the total input "
+        "length is the sum of prefix-len (default: "
+        f"{RandomDataset.DEFAULT_PREFIX_LEN}) and a random context length "
+        "sampled from [input_len * (1 - range_ratio), "
+        "input_len * (1 + range_ratio)]. For SonnetDataset, "
+        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.",
+    )
+    # random dataset
+    parser.add_argument(
+        "--random-range-ratio",
+        type=float,
+        default=None,
+        help=f"Range ratio (default : {RandomDataset.DEFAULT_RANGE_RATIO}) "
+        "for sampling input/output length, "
+        "used only for RandomDataset. Must be in the range [0, 1) to "
+        "define a symmetric sampling range "
+        "[length * (1 - range_ratio), length * (1 + range_ratio)].",
+    )
+
+    # 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 = AsyncEngineArgs.add_cli_args(parser)
+
+    return parser
+
 
 if __name__ == "__main__":
-    print("""DEPRECATED: This script has been moved to the vLLM CLI.
-
-Please use the following command instead:
-    vllm bench throughput
-
-For help with the new command, run:
-    vllm bench throughput --help
-
-Alternatively, you can run the new command directly with:
-    python -m vllm.entrypoints.cli.main bench throughput --help
-""")
-    sys.exit(1)
+    parser = create_argument_parser()
+    args = parser.parse_args()
+    if args.tokenizer is None:
+        args.tokenizer = args.model
+    validate_args(args)
+    main(args)

benchmarks/benchmark_utils.py

@@ -6,7 +6,7 @@ import math
 import os
 import time
 from types import TracebackType
-from typing import Any
+from typing import Any, Optional, Union
 
 
 def convert_to_pytorch_benchmark_format(
@@ -92,7 +92,7 @@ class TimeCollector:
     def __init__(self, scale: int) -> None:
         self.cnt: int = 0
         self._sum: int = 0
-        self._max: int | None = None
+        self._max: Optional[int] = None
         self.scale = scale
         self.start_time: int = time.monotonic_ns()
 
@@ -104,13 +104,13 @@ class TimeCollector:
         else:
             self._max = max(self._max, v)
 
-    def avg(self) -> float | str:
+    def avg(self) -> Union[float, str]:
         return self._sum * 1.0 / self.cnt / self.scale if self.cnt > 0 else "N/A"
 
-    def max(self) -> float | str:
+    def max(self) -> Union[float, str]:
         return self._max / self.scale if self._max else "N/A"
 
-    def dump_avg_max(self) -> list[float | str]:
+    def dump_avg_max(self) -> list[Union[float, str]]:
         return [self.avg(), self.max()]
 
     def __enter__(self) -> None:
@@ -118,8 +118,8 @@ class TimeCollector:
 
     def __exit__(
         self,
-        exc_type: type[BaseException] | None,
-        exc_value: BaseException | None,
-        exc_traceback: TracebackType | None,
+        exc_type: Optional[type[BaseException]],
+        exc_value: Optional[BaseException],
+        exc_traceback: Optional[TracebackType],
     ) -> None:
         self.collect(time.monotonic_ns() - self.start_time)

benchmarks/cutlass_benchmarks/sparse_benchmarks.py

@@ -6,7 +6,8 @@ import copy
 import itertools
 import pickle as pkl
 import time
-from collections.abc import Callable, Iterable
+from collections.abc import Iterable
+from typing import Callable
 
 import torch
 import torch.utils.benchmark as TBenchmark
@@ -15,7 +16,7 @@ from utils import make_rand_sparse_tensors
 from weight_shapes import WEIGHT_SHAPES
 
 from vllm import _custom_ops as ops
-from vllm.utils.argparse_utils import FlexibleArgumentParser
+from vllm.utils import FlexibleArgumentParser
 
 DEFAULT_MODELS = list(WEIGHT_SHAPES.keys())
 DEFAULT_BATCH_SIZES = [1, 16, 32, 64, 128, 256, 512]

benchmarks/cutlass_benchmarks/w8a8_benchmarks.py

@@ -6,7 +6,8 @@ import copy
 import itertools
 import pickle as pkl
 import time
-from collections.abc import Callable, Iterable
+from collections.abc import Iterable
+from typing import Callable, Optional
 
 import torch
 import torch.utils.benchmark as TBenchmark
@@ -16,10 +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_triton_block_scaled_mm,
+    w8a8_block_fp8_matmul,
 )
-from vllm.utils.argparse_utils import FlexibleArgumentParser
-from vllm.utils.math_utils import cdiv
+from vllm.utils import FlexibleArgumentParser, cdiv
 
 DEFAULT_MODELS = list(WEIGHT_SHAPES.keys())
 DEFAULT_BATCH_SIZES = [1, 16, 32, 64, 128, 256, 512]
@@ -53,7 +53,7 @@ def bench_int8(
     n: int,
     label: str,
     sub_label: str,
-    bench_kernels: list[str] | None = None,
+    bench_kernels: Optional[list[str]] = None,
 ) -> Iterable[TMeasurement]:
     """Benchmark INT8-based kernels."""
     assert dtype == torch.int8
@@ -108,7 +108,7 @@ def bench_fp8(
     n: int,
     label: str,
     sub_label: str,
-    bench_kernels: list[str] | None = None,
+    bench_kernels: Optional[list[str]] = None,
 ) -> Iterable[TMeasurement]:
     """Benchmark FP8-based kernels."""
     assert dtype == torch.float8_e4m3fn
@@ -158,7 +158,7 @@ def bench_fp8(
         "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_triton_block_scaled_mm(
+        "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(
@@ -183,7 +183,7 @@ def bench(
     n: int,
     label: str,
     sub_label: str,
-    bench_kernels: list[str] | None = None,
+    bench_kernels: Optional[list[str]] = None,
 ) -> Iterable[TMeasurement]:
     if dtype == torch.int8:
         return bench_int8(dtype, m, k, n, label, sub_label, bench_kernels)
@@ -201,7 +201,7 @@ def print_timers(timers: Iterable[TMeasurement]):
 def run(
     dtype: torch.dtype,
     MKNs: Iterable[tuple[int, int, int]],
-    bench_kernels: list[str] | None = None,
+    bench_kernels: Optional[list[str]] = None,
 ) -> Iterable[TMeasurement]:
     results = []
     for m, k, n in MKNs:

benchmarks/disagg_benchmarks/disagg_overhead_benchmark.sh

@@ -55,20 +55,24 @@ benchmark() {
   output_len=$2
 
 
-  CUDA_VISIBLE_DEVICES=0 vllm serve $model \
+  CUDA_VISIBLE_DEVICES=0 python3 \
+    -m vllm.entrypoints.openai.api_server \
+    --model $model \
     --port 8100 \
     --max-model-len 10000 \
     --gpu-memory-utilization 0.6 \
     --kv-transfer-config \
-    '{"kv_connector":"P2pNcclConnector","kv_role":"kv_producer","kv_rank":0,"kv_parallel_size":2,"kv_buffer_size":5e9}' &
+    '{"kv_connector":"PyNcclConnector","kv_role":"kv_producer","kv_rank":0,"kv_parallel_size":2,"kv_buffer_size":5e9}' &
 
 
-  CUDA_VISIBLE_DEVICES=1 vllm serve $model \
+  CUDA_VISIBLE_DEVICES=1 python3 \
+    -m vllm.entrypoints.openai.api_server \
+    --model $model \
     --port 8200 \
     --max-model-len 10000 \
     --gpu-memory-utilization 0.6 \
     --kv-transfer-config \
-    '{"kv_connector":"P2pNcclConnector","kv_role":"kv_consumer","kv_rank":1,"kv_parallel_size":2,"kv_buffer_size":5e9}' &
+    '{"kv_connector":"PyNcclConnector","kv_role":"kv_consumer","kv_rank":1,"kv_parallel_size":2,"kv_buffer_size":5e9}' &
 
   wait_for_server 8100
   wait_for_server 8200

benchmarks/disagg_benchmarks/disagg_performance_benchmark.sh

@@ -38,12 +38,16 @@ wait_for_server() {
 launch_chunked_prefill() {
   model="meta-llama/Meta-Llama-3.1-8B-Instruct"
   # disagg prefill
-  CUDA_VISIBLE_DEVICES=0 vllm serve $model \
+  CUDA_VISIBLE_DEVICES=0 python3 \
+    -m vllm.entrypoints.openai.api_server \
+    --model $model \
     --port 8100 \
     --max-model-len 10000 \
     --enable-chunked-prefill \
     --gpu-memory-utilization 0.6 &
-  CUDA_VISIBLE_DEVICES=1 vllm serve $model \
+  CUDA_VISIBLE_DEVICES=1 python3 \
+    -m vllm.entrypoints.openai.api_server \
+    --model $model \
     --port 8200 \
     --max-model-len 10000 \
     --enable-chunked-prefill \
@@ -58,19 +62,23 @@ launch_chunked_prefill() {
 launch_disagg_prefill() {
   model="meta-llama/Meta-Llama-3.1-8B-Instruct"
   # disagg prefill
-  CUDA_VISIBLE_DEVICES=0 vllm serve $model \
+  CUDA_VISIBLE_DEVICES=0 python3 \
+    -m vllm.entrypoints.openai.api_server \
+    --model $model \
     --port 8100 \
     --max-model-len 10000 \
     --gpu-memory-utilization 0.6 \
     --kv-transfer-config \
-    '{"kv_connector":"P2pNcclConnector","kv_role":"kv_producer","kv_rank":0,"kv_parallel_size":2,"kv_buffer_size":5e9}' &
+    '{"kv_connector":"PyNcclConnector","kv_role":"kv_producer","kv_rank":0,"kv_parallel_size":2,"kv_buffer_size":5e9}' &
 
-  CUDA_VISIBLE_DEVICES=1 vllm serve $model \
+  CUDA_VISIBLE_DEVICES=1 python3 \
+    -m vllm.entrypoints.openai.api_server \
+    --model $model \
     --port 8200 \
     --max-model-len 10000 \
     --gpu-memory-utilization 0.6 \
     --kv-transfer-config \
-    '{"kv_connector":"P2pNcclConnector","kv_role":"kv_consumer","kv_rank":1,"kv_parallel_size":2,"kv_buffer_size":5e9}' &
+    '{"kv_connector":"PyNcclConnector","kv_role":"kv_consumer","kv_rank":1,"kv_parallel_size":2,"kv_buffer_size":5e9}' &
 
   wait_for_server 8100
   wait_for_server 8200

benchmarks/fused_kernels/layernorm_rms_benchmarks.py

@@ -3,9 +3,10 @@
 
 import pickle as pkl
 import time
-from collections.abc import Callable, Iterable
+from collections.abc import Iterable
 from dataclasses import dataclass
 from itertools import product
+from typing import Callable, Optional
 
 import torch
 import torch.utils.benchmark as TBenchmark
@@ -50,7 +51,7 @@ def get_bench_params() -> list[bench_params_t]:
 def unfused_int8_impl(
     rms_norm_layer: RMSNorm,
     x: torch.Tensor,
-    residual: torch.Tensor | None,
+    residual: Optional[torch.Tensor],
     quant_dtype: torch.dtype,
 ):
     # Norm
@@ -67,7 +68,7 @@ def unfused_int8_impl(
 def unfused_fp8_impl(
     rms_norm_layer: RMSNorm,
     x: torch.Tensor,
-    residual: torch.Tensor | None,
+    residual: Optional[torch.Tensor],
     quant_dtype: torch.dtype,
 ):
     # Norm
@@ -84,7 +85,7 @@ def unfused_fp8_impl(
 def fused_impl(
     rms_norm_layer: RMSNorm,  # this stores the weights
     x: torch.Tensor,
-    residual: torch.Tensor | None,
+    residual: Optional[torch.Tensor],
     quant_dtype: torch.dtype,
 ):
     out, _ = ops.rms_norm_dynamic_per_token_quant(

benchmarks/kernels/bench_block_fp8_gemm.py

@@ -4,10 +4,7 @@
 import torch
 
 from vllm.model_executor.layers.quantization.utils.fp8_utils import (
-    apply_w8a8_block_fp8_linear,
-)
-from vllm.model_executor.layers.quantization.utils.w8a8_utils import (
-    CUTLASS_BLOCK_FP8_SUPPORTED,
+    w8a8_block_fp8_matmul,
 )
 from vllm.platforms import current_platform
 from vllm.triton_utils import triton as vllm_triton
@@ -19,7 +16,6 @@ assert current_platform.is_cuda(), (
 # DeepSeek-V3 weight shapes
 DEEPSEEK_V3_SHAPES = [
     (512 + 64, 7168),
-    (2112, 7168),
     ((128 + 64) * 128, 7168),
     (128 * (128 + 128), 512),
     (7168, 16384),
@@ -32,7 +28,7 @@ DEEPSEEK_V3_SHAPES = [
 ]
 
 
-def build_w8a8_block_fp8_runner(M, N, K, block_size, device, use_cutlass):
+def build_w8a8_block_fp8_runner(M, N, K, block_size, device):
     """Build runner function for w8a8 block fp8 matmul."""
     factor_for_scale = 1e-2
 
@@ -40,54 +36,37 @@ def build_w8a8_block_fp8_runner(M, N, K, block_size, device, use_cutlass):
     fp8_max, fp8_min = fp8_info.max, fp8_info.min
 
     # Create random FP8 tensors
-    A_ref = (torch.rand(M, K, dtype=torch.bfloat16, device=device) - 0.5) * 2 * fp8_max
+    A_fp32 = (torch.rand(M, K, dtype=torch.float32, device=device) - 0.5) * 2 * fp8_max
+    A = A_fp32.clamp(min=fp8_min, max=fp8_max).to(torch.float8_e4m3fn)
 
-    B_ref = (torch.rand(N, K, dtype=torch.bfloat16, device=device) - 0.5) * 2 * fp8_max
-    B = B_ref.clamp(min=fp8_min, max=fp8_max).to(torch.float8_e4m3fn)
+    B_fp32 = (torch.rand(N, K, dtype=torch.float32, device=device) - 0.5) * 2 * fp8_max
+    B = B_fp32.clamp(min=fp8_min, max=fp8_max).to(torch.float8_e4m3fn)
 
     # Create scales
     block_n, block_k = block_size[0], block_size[1]
     n_tiles = (N + block_n - 1) // block_n
     k_tiles = (K + block_k - 1) // block_k
 
+    As = torch.rand(M, k_tiles, dtype=torch.float32, device=device) * factor_for_scale
     Bs = (
         torch.rand(n_tiles, k_tiles, dtype=torch.float32, device=device)
         * factor_for_scale
     )
 
-    # SM90 CUTLASS requires row-major format for scales
-    if use_cutlass and current_platform.is_device_capability(90):
-        Bs = Bs.T.contiguous()
-
     def run():
-        if use_cutlass:
-            return apply_w8a8_block_fp8_linear(
-                A_ref, B, block_size, Bs, cutlass_block_fp8_supported=True
-            )
-        else:
-            return apply_w8a8_block_fp8_linear(
-                A_ref, B, block_size, Bs, cutlass_block_fp8_supported=False
-            )
+        return w8a8_block_fp8_matmul(A, B, As, Bs, block_size, torch.bfloat16)
 
     return run
 
 
-# Determine available providers
-available_providers = ["torch-bf16", "w8a8-block-fp8-triton"]
-plot_title = "BF16 vs W8A8 Block FP8 GEMMs"
-
-if CUTLASS_BLOCK_FP8_SUPPORTED:
-    available_providers.append("w8a8-block-fp8-cutlass")
-
-
 @vllm_triton.testing.perf_report(
     vllm_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=available_providers,
-        line_names=available_providers,
+        line_vals=["torch-bf16", "w8a8-block-fp8"],
+        line_names=["torch-bf16", "w8a8-block-fp8"],
         ylabel="TFLOP/s (larger is better)",
         plot_name="BF16 vs W8A8 Block FP8 GEMMs",
         args={},
@@ -105,22 +84,11 @@ def benchmark_tflops(batch_size, provider, N, K, block_size=(128, 128)):
         ms, min_ms, max_ms = vllm_triton.testing.do_bench_cudagraph(
             lambda: torch.nn.functional.linear(a, b), quantiles=quantiles
         )
-    elif provider == "w8a8-block-fp8-triton":
-        run_w8a8_triton = build_w8a8_block_fp8_runner(
-            M, N, K, block_size, device, use_cutlass=False
-        )
+    else:  # w8a8-block-fp8
+        run_w8a8 = build_w8a8_block_fp8_runner(M, N, K, block_size, device)
         ms, min_ms, max_ms = vllm_triton.testing.do_bench_cudagraph(
-            lambda: run_w8a8_triton(), quantiles=quantiles
+            lambda: run_w8a8(), quantiles=quantiles
         )
-    elif provider == "w8a8-block-fp8-cutlass":
-        run_w8a8_cutlass = build_w8a8_block_fp8_runner(
-            M, N, K, block_size, device, use_cutlass=True
-        )
-        ms, min_ms, max_ms = vllm_triton.testing.do_bench_cudagraph(
-            lambda: run_w8a8_cutlass(), quantiles=quantiles
-        )
-    else:
-        raise ValueError(f"Unknown provider: {provider}")
 
     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)

benchmarks/kernels/bench_mxfp4_qutlass.py

@@ -1,191 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-#
-# Copyright (C) 2025 Roberto L. Castro (Roberto.LopezCastro@ist.ac.at).
-# All Rights Reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#       http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-#
-
-import argparse
-import copy
-import itertools
-
-import torch
-from compressed_tensors.transform.utils.hadamard import deterministic_hadamard_matrix
-from weight_shapes import WEIGHT_SHAPES
-
-from vllm._custom_ops import fusedQuantizeMx, matmul_mxf4_bf16_tn
-from vllm.model_executor.layers.quantization.qutlass_utils import to_blocked
-from vllm.triton_utils import triton
-
-PROVIDER_CFGS = {
-    "torch-bf16": dict(enabled=True),
-    "mxfp4": dict(no_a_quant=False, enabled=True),
-    "mxfp4-noquant": dict(no_a_quant=True, enabled=True),
-}
-
-_enabled = [k for k, v in PROVIDER_CFGS.items() if v["enabled"]]
-
-
-def get_hadamard_matrix(group_size: int, dtype: torch.dtype, device: torch.device):
-    return (
-        deterministic_hadamard_matrix(group_size, dtype=dtype, device=device)
-        * group_size**-0.5
-    )
-
-
-def _quant_weight_mxfp4(
-    b: torch.Tensor, forward_hadamard_matrix: torch.Tensor, device: str
-):
-    weight_hf_e2m1, weight_hf_e8m0 = fusedQuantizeMx(
-        b, forward_hadamard_matrix, method="abs_max"
-    )
-    weight_hf_scale_block = to_blocked(weight_hf_e8m0, backend="triton")
-    return weight_hf_e2m1, weight_hf_scale_block
-
-
-def build_mxfp4_runner(cfg, a, b, forward_hadamard_matrix, dtype, device):
-    weight_hf_e2m1, weight_hf_scale_block = _quant_weight_mxfp4(
-        b, forward_hadamard_matrix, device
-    )
-    alpha = torch.tensor([1.0], device="cuda")
-
-    if cfg["no_a_quant"]:
-        # Pre-quantize activation
-        input_hf_e2m1, input_hf_e8m0 = fusedQuantizeMx(
-            a, forward_hadamard_matrix, method="abs_max"
-        )
-        input_hf_scale_block = to_blocked(input_hf_e8m0, backend="triton")
-
-        def run():
-            return matmul_mxf4_bf16_tn(
-                input_hf_e2m1,
-                weight_hf_e2m1,
-                input_hf_scale_block,
-                weight_hf_scale_block,
-                alpha,
-            )
-
-        return run
-
-    # Quantize activation on-the-fly
-    def run():
-        input_hf_e2m1, input_hf_e8m0 = fusedQuantizeMx(
-            a, forward_hadamard_matrix, method="abs_max"
-        )
-        input_hf_scale_block = to_blocked(input_hf_e8m0, backend="triton")
-        return matmul_mxf4_bf16_tn(
-            input_hf_e2m1,
-            weight_hf_e2m1,
-            input_hf_scale_block,
-            weight_hf_scale_block,
-            alpha,
-        )
-
-    return run
-
-
-@triton.testing.perf_report(
-    triton.testing.Benchmark(
-        x_names=["batch_size"],
-        x_vals=[
-            1,
-            4,
-            8,
-            16,
-            32,
-            64,
-            128,
-            256,
-            512,
-            1024,
-            2048,
-            4096,
-            8192,
-            16384,
-            24576,
-            32768,
-        ],
-        x_log=False,
-        line_arg="provider",
-        line_vals=_enabled,
-        line_names=_enabled,
-        ylabel="TFLOP/s (larger is better)",
-        plot_name="BF16 vs MXFP4 GEMMs",
-        args={},
-    )
-)
-def benchmark(batch_size, provider, N, K, had_size):
-    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)
-    forward_hadamard_matrix = get_hadamard_matrix(had_size, dtype, device)
-
-    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), rep=200, quantiles=quantiles
-        )
-    else:
-        cfg = PROVIDER_CFGS[provider]
-        run_quant = build_mxfp4_runner(
-            cfg, a, b, forward_hadamard_matrix, dtype, device
-        )
-        ms, min_ms, max_ms = triton.testing.do_bench_cudagraph(
-            lambda: run_quant(), rep=200, 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-3.3-70B-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):
-        for had_size in [32, 64, 128]:
-            print(f"{model}, N={N} K={K}, HAD={had_size}, BF16 vs MXFP4 GEMMs TFLOP/s:")
-            benchmark.run(
-                print_data=True,
-                show_plots=True,
-                save_path=f"bench_mxfp4_res_n{N}_k{K}",
-                N=N,
-                K=K,
-                had_size=had_size,
-            )
-
-    print("Benchmark finished!")

benchmarks/kernels/bench_nvfp4_gemm.py

@@ -3,7 +3,6 @@
 import argparse
 import copy
 import itertools
-import os
 
 import torch
 from weight_shapes import WEIGHT_SHAPES
@@ -24,45 +23,21 @@ PROVIDER_CFGS = {
     "torch-bf16": dict(enabled=True),
     "nvfp4": dict(no_a_quant=False, enabled=True),
     "nvfp4-noquant": dict(no_a_quant=True, enabled=True),
-    "fbgemm-nvfp4": dict(fbgemm=True, no_a_quant=False, enabled=True),
-    "fbgemm-nvfp4-noquant": dict(fbgemm=True, no_a_quant=True, enabled=True),
 }
 
-_needs_fbgemm = any(
-    v.get("fbgemm", False) for v in PROVIDER_CFGS.values() if v.get("enabled", False)
-)
-if _needs_fbgemm:
-    try:
-        from fbgemm_gpu.experimental.gemm.triton_gemm.fp4_quantize import (
-            triton_scale_nvfp4_quant,
-        )
-    except ImportError:
-        print(
-            "WARNING: FBGEMM providers are enabled but fbgemm_gpu is not installed. "
-            "These providers will be skipped. Please install fbgemm_gpu with: "
-            "'pip install fbgemm-gpu-genai' to run them."
-        )
-        # Disable FBGEMM providers so the benchmark can run.
-        for cfg in PROVIDER_CFGS.values():
-            if cfg.get("fbgemm"):
-                cfg["enabled"] = False
-
 _enabled = [k for k, v in PROVIDER_CFGS.items() if v["enabled"]]
 
 
-def _quant_weight_nvfp4(b: torch.Tensor, device: str, cfg):
+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
-    if "fbgemm" in cfg and cfg["fbgemm"]:
-        b_fp4, scale_b_fp4 = triton_scale_nvfp4_quant(b, b_global_scale)
-    else:
-        b_fp4, scale_b_fp4 = ops.scaled_fp4_quant(b, b_global_scale)
+    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, cfg)
+    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.
@@ -71,35 +46,6 @@ def build_nvfp4_runner(cfg, a, b, dtype, device):
 
     # Alpha for the GEMM operation
     alpha = 1.0 / (a_global_scale * b_global_scale)
-    if "fbgemm" in cfg and cfg["fbgemm"]:
-        if cfg["no_a_quant"]:
-            a_fp4, scale_a_fp4 = triton_scale_nvfp4_quant(a, a_global_scale)
-
-            def run():
-                return torch.ops.fbgemm.f4f4bf16(
-                    a_fp4,
-                    b_fp4,
-                    scale_a_fp4,
-                    scale_b_fp4,
-                    global_scale=alpha,
-                    use_mx=False,
-                )
-
-            return run
-        else:
-
-            def run():
-                a_fp4, scale_a_fp4 = triton_scale_nvfp4_quant(a, a_global_scale)
-                return torch.ops.fbgemm.f4f4bf16(
-                    a_fp4,
-                    b_fp4,
-                    scale_a_fp4,
-                    scale_b_fp4,
-                    global_scale=alpha,
-                    use_mx=False,
-                )
-
-            return run
 
     if cfg["no_a_quant"]:
         # Pre-quantize activation
@@ -184,13 +130,10 @@ if __name__ == "__main__":
 
     for K, N, model in prepare_shapes(args):
         print(f"{model}, N={N} K={K}, BF16 vs NVFP4 GEMMs TFLOP/s:")
-        save_dir = f"bench_nvfp4_res_n{N}_k{K}"
-        os.makedirs(save_dir, exist_ok=True)
-
         benchmark.run(
             print_data=True,
             show_plots=True,
-            save_path=save_dir,
+            save_path=f"bench_nvfp4_res_n{N}_k{K}",
             N=N,
             K=K,
         )

benchmarks/kernels/bench_nvfp4_qutlass.py

@@ -1,207 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-#
-# Copyright (C) 2025 Roberto L. Castro (Roberto.LopezCastro@ist.ac.at).
-# All Rights Reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#       http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-#
-
-import argparse
-import copy
-import itertools
-
-import torch
-from compressed_tensors.transform.utils.hadamard import deterministic_hadamard_matrix
-from weight_shapes import WEIGHT_SHAPES
-
-from vllm import _custom_ops as ops  # use existing nvfp4 gemm in vllm
-from vllm._custom_ops import fusedQuantizeNv
-from vllm.model_executor.layers.quantization.qutlass_utils import to_blocked
-from vllm.triton_utils import triton
-
-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 get_hadamard_matrix(group_size: int, dtype: torch.dtype, device: torch.device):
-    return (
-        deterministic_hadamard_matrix(group_size, dtype=dtype, device=device)
-        * group_size**-0.5
-    )
-
-
-def _quant_weight_nvfp4(
-    b: torch.Tensor,
-    forward_hadamard_matrix: torch.Tensor,
-    global_scale: torch.Tensor,
-    device: str,
-    M: int,
-    N: int,
-    K: int,
-):
-    weight_hf_e2m1, weight_hf_e8m0 = fusedQuantizeNv(
-        b, forward_hadamard_matrix, global_scale
-    )
-    weight_hf_scale_block = to_blocked(weight_hf_e8m0, backend="triton").view(
-        -1, K // 16
-    )
-    return weight_hf_e2m1, weight_hf_scale_block
-
-
-def build_nvfp4_runner(cfg, a, b, forward_hadamard_matrix, dtype, device, M, N, K):
-    alpha = torch.tensor([1.0], device="cuda")
-    global_scale = torch.tensor([1.0], device="cuda")
-    weight_hf_e2m1, weight_hf_scale_block = _quant_weight_nvfp4(
-        b, forward_hadamard_matrix, global_scale, device, M, N, K
-    )
-
-    if cfg["no_a_quant"]:
-        # Pre-quantize activation
-        input_hf_e2m1, input_hf_e8m0 = fusedQuantizeNv(
-            a, forward_hadamard_matrix, global_scale
-        )
-        input_hf_scale_block = to_blocked(input_hf_e8m0, backend="triton").view(
-            -1, K // 16
-        )
-
-        def run():
-            return ops.cutlass_scaled_fp4_mm(
-                input_hf_e2m1,
-                weight_hf_e2m1,
-                input_hf_scale_block,
-                weight_hf_scale_block,
-                alpha,
-                torch.bfloat16,
-            )
-
-        return run
-
-    # Quantize activation on-the-fly
-    def run():
-        input_hf_e2m1, input_hf_e8m0 = fusedQuantizeNv(
-            a, forward_hadamard_matrix, global_scale
-        )
-        input_hf_scale_block = to_blocked(input_hf_e8m0, backend="triton").view(
-            -1, K // 16
-        )
-        return ops.cutlass_scaled_fp4_mm(
-            input_hf_e2m1,
-            weight_hf_e2m1,
-            input_hf_scale_block,
-            weight_hf_scale_block,
-            alpha,
-            torch.bfloat16,
-        )
-
-    return run
-
-
-@triton.testing.perf_report(
-    triton.testing.Benchmark(
-        x_names=["batch_size"],
-        x_vals=[
-            1,
-            4,
-            8,
-            16,
-            32,
-            64,
-            128,
-            256,
-            512,
-            1024,
-            2048,
-            4096,
-            8192,
-            16384,
-            24576,
-            32768,
-        ],
-        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, had_size):
-    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)
-    forward_hadamard_matrix = get_hadamard_matrix(had_size, dtype, device)
-
-    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), rep=200, quantiles=quantiles
-        )
-    else:
-        cfg = PROVIDER_CFGS[provider]
-        run_quant = build_nvfp4_runner(
-            cfg, a, b, forward_hadamard_matrix, dtype, device, M, N, K
-        )
-        ms, min_ms, max_ms = triton.testing.do_bench_cudagraph(
-            lambda: run_quant(), rep=200, 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-3.3-70B-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):
-        for had_size in [16, 32, 64, 128]:
-            print(f"{model}, N={N} K={K}, HAD={had_size}, 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,
-                had_size=had_size,
-            )
-
-    print("Benchmark finished!")

benchmarks/kernels/bench_per_token_quant_fp8.py

@@ -1,27 +1,15 @@
 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 import itertools
-from collections.abc import Callable
-from unittest.mock import patch
+from typing import Callable
 
-import pandas as pd
 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
-from vllm.utils.argparse_utils import FlexibleArgumentParser
-from vllm.utils.torch_utils import STR_DTYPE_TO_TORCH_DTYPE
-
-
-def with_triton_mode(fn):
-    """Temporarily force the Triton fallback path"""
-
-    def wrapped(*args, **kwargs):
-        with patch("vllm.platforms.current_platform.is_cuda", return_value=False):
-            return fn(*args, **kwargs)
-
-    return wrapped
 
 
 # TODO(luka): use standalone_compile utility
@@ -33,238 +21,78 @@ def with_dyn_arg(fn: Callable, arg_index: int, dim_index: int):
     return inner
 
 
-def bench_compile(fn: Callable):
-    # recompile for different shapes
-    fwd = torch.compile(fn, fullgraph=True, dynamic=False)
+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
-    return with_dyn_arg(fwd, 0, 0)
+    torch_per_token_quant_fp8 = with_dyn_arg(torch_per_token_quant_fp8, 0, 0)
 
 
-torch._dynamo.config.recompile_limit = 8888
+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,
-    hidden_size: int,
-    group_shape: GroupShape,
-    dtype: torch.dtype,
-):
-    """Calculate the difference between Inductor and CUDA implementations."""
+def calculate_diff(batch_size: int, seq_len: int):
+    """Calculate difference between Triton and CUDA implementations."""
     device = torch.device("cuda")
-    x = torch.randn((batch_size, hidden_size), dtype=dtype, device=device)
+    x = torch.rand((batch_size * seq_len, 4096), dtype=torch.float16, device=device)
 
-    quant_fp8 = QuantFP8(False, group_shape, column_major_scales=False)
+    torch_out, torch_scale = torch_per_token_quant_fp8(x)
+    cuda_out, cuda_scale = cuda_per_token_quant_fp8(x)
 
-    torch_out, torch_scale = bench_compile(quant_fp8.forward_native)(x)
-    torch_eager_out, torch_eager_scale = quant_fp8.forward_native(x)
-    cuda_out, cuda_scale = quant_fp8.forward_cuda(x)
-
-    try:
-        torch.testing.assert_close(
-            cuda_out.to(torch.float32),
-            torch_out.to(torch.float32),
-            rtol=1e-3,
-            atol=1e-5,
-        )
-        torch.testing.assert_close(cuda_scale, torch_scale, rtol=1e-3, atol=1e-5)
-        torch.testing.assert_close(
-            cuda_out.to(torch.float32),
-            torch_eager_out.to(torch.float32),
-            rtol=1e-3,
-            atol=1e-5,
-        )
-        torch.testing.assert_close(cuda_scale, torch_eager_scale, rtol=1e-3, atol=1e-5)
+    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")
-    except AssertionError as e:
+    else:
         print("❌ Implementations differ")
-        print(e)
 
 
-configs = []
+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))
 
 
-def benchmark_quantization(
-    batch_size,
-    hidden_size,
-    provider,
-    group_shape: GroupShape,
-    col_major: bool,
-    dtype: torch.dtype,
-):
+@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, hidden_size, device=device, dtype=dtype)
+    x = torch.randn(batch_size * seq_len, 4096, device=device, dtype=dtype)
 
     quantiles = [0.5, 0.2, 0.8]
-    quant_fp8 = QuantFP8(False, group_shape, column_major_scales=col_major)
 
     if provider == "torch":
-        fn = lambda: bench_compile(quant_fp8.forward_native)(x.clone())
+        fn = lambda: torch_per_token_quant_fp8(x.clone())
     elif provider == "cuda":
-        fn = lambda: quant_fp8.forward_cuda(x.clone())
-    elif provider == "triton":
-        if not group_shape.is_per_group():
-            # Triton only supported for per-group
-            return 0, 0, 0
-
-        fn = lambda: with_triton_mode(quant_fp8.forward_cuda)(x.clone())
+        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
 
 
-# TODO(luka) extract to utils
-def compute_geomean_speedups(
-    df: pd.DataFrame,
-    baseline_col: str,
-    speedup_cols: list[str],
-    groupby_cols: list[str] | None = None,
-) -> pd.DataFrame:
-    """
-    Compute geometric mean speedups over a baseline column.
-
-    Args:
-        df: Input dataframe
-        baseline_col: Column to use as baseline
-        speedup_cols: Columns to compute speedups for
-        groupby_cols: Columns to group by. If None, compute over entire df.
-
-    Returns:
-        pd.DataFrame with geometric mean speedups
-    """
-    from scipy.stats import gmean
-
-    def geo_speedup(group: pd.DataFrame) -> pd.Series:
-        ratios = {
-            col: (group[baseline_col] / group[col]).values for col in speedup_cols
-        }
-        return pd.Series({col: gmean(vals) for col, vals in ratios.items()})
-
-    if groupby_cols is None:
-        result = geo_speedup(df).to_frame().T
-    else:
-        result = (
-            df.groupby(groupby_cols)
-            .apply(geo_speedup, include_groups=False)
-            .reset_index()
-        )
-
-    return result
-
-
 if __name__ == "__main__":
-    parser = FlexibleArgumentParser(
-        description="Benchmark the various implementations of QuantFP8 (dynamic-only)"
-    )
-    parser.add_argument("-c", "--check", action="store_true")
-    parser.add_argument(
-        "--dtype", type=str, choices=["half", "bfloat16", "float"], default="bfloat16"
-    )
-    parser.add_argument(
-        "--hidden-sizes",
-        type=int,
-        nargs="+",
-        default=[896, 1024, 2048, 4096, 7168],
-        help="Hidden sizes to benchmark",
-    )
-    parser.add_argument(
-        "--batch-sizes",
-        type=int,
-        nargs="+",
-        default=[1, 16, 128, 512, 1024],
-        help="Batch sizes to benchmark",
-    )
-    parser.add_argument(
-        "--group-sizes",
-        type=int,
-        nargs="+",
-        default=None,
-        help="Group sizes for GroupShape(1,N) to benchmark. "
-        "Use 0 for PER_TENSOR, -1 for PER_TOKEN (default: 0,-1,64,128)",
-    )
-    parser.add_argument(
-        "--no-column-major",
-        action="store_true",
-        help="Disable column-major scales testing",
-    )
-
-    args = parser.parse_args()
-    assert args
-
-    dtype = STR_DTYPE_TO_TORCH_DTYPE[args.dtype]
-
-    hidden_sizes = args.hidden_sizes
-    batch_sizes = args.batch_sizes
-
-    if args.group_sizes is not None:
-        group_shapes = []
-        for size in args.group_sizes:
-            if size == 0:
-                group_shapes.append(GroupShape.PER_TENSOR)
-            elif size == -1:
-                group_shapes.append(GroupShape.PER_TOKEN)
-            else:
-                group_shapes.append(GroupShape(1, size))
-    else:
-        group_shapes = [
-            GroupShape.PER_TENSOR,
-            GroupShape.PER_TOKEN,
-            GroupShape(1, 64),
-            GroupShape(1, 128),
-        ]
-
-    column_major_scales = [False] if args.no_column_major else [True, False]
-
-    config_gen = itertools.product(
-        group_shapes,
-        column_major_scales,
-        batch_sizes,
-        hidden_sizes,
-    )
-
-    # filter out column-major scales for non-group, reverse order
-    configs.extend(c[::-1] for c in config_gen if (c[0].is_per_group() or not c[1]))
-
-    print(f"Running {len(configs)} configurations:")
-    print(f"  Hidden sizes: {hidden_sizes}")
-    print(f"  Batch sizes: {batch_sizes}")
-    print(f"  Group shapes: {[str(g) for g in group_shapes]}")
-    print(f"  Column major scales: {column_major_scales}")
-    print()
-
-    if args.check:
-        for group_shape in group_shapes:
-            group_size = group_shape[1]
-            print(f"{group_size=}")
-            calculate_diff(
-                batch_size=4, hidden_size=4096, group_shape=group_shape, dtype=dtype
-            )
-
-    benchmark = triton.testing.perf_report(
-        triton.testing.Benchmark(
-            x_names=["hidden_size", "batch_size", "col_major", "group_shape"],
-            x_vals=configs,
-            line_arg="provider",
-            line_vals=["torch", "cuda", "triton"],
-            line_names=["Torch (Compiled)", "CUDA", "Triton"],
-            styles=[("blue", "-"), ("green", "-"), ("black", "-")],
-            ylabel="us",
-            plot_name="QuantFP8 performance",
-            args={},
-        )
-    )(benchmark_quantization)
-
-    df = benchmark.run(print_data=True, dtype=dtype, return_df=True)
-
-    # Print geomean speedups
-    geo_table_grouped = compute_geomean_speedups(
-        df,
-        baseline_col="Torch (Compiled)",
-        speedup_cols=["CUDA", "Triton"],
-        groupby_cols=["col_major", "group_shape"],
-    )
-
-    print("Speedup over Torch (Compiled)")
-    print(geo_table_grouped.to_string(index=False))
+    calculate_diff(batch_size=4, seq_len=4096)
+    benchmark_quantization.run(print_data=True)

benchmarks/kernels/benchmark_activation.py

@@ -1,105 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-
-# benchmark custom activation op performance
-import itertools
-
-import torch
-
-import vllm.model_executor.layers.activation  # noqa F401
-from vllm.model_executor.custom_op import CustomOp
-from vllm.platforms import current_platform
-from vllm.triton_utils import triton
-from vllm.utils.argparse_utils import FlexibleArgumentParser
-from vllm.utils.torch_utils import STR_DTYPE_TO_TORCH_DTYPE
-
-batch_size_range = [1, 16, 32, 64, 128]
-seq_len_range = [1, 16, 64, 128, 256, 512, 1024, 2048, 4096]
-intermediate_size = [3072, 9728, 12288]
-configs = list(itertools.product(batch_size_range, seq_len_range, intermediate_size))
-
-
-def benchmark_activation(
-    batch_size: int,
-    seq_len: int,
-    intermediate_size: int,
-    provider: str,
-    func_name: str,
-    dtype: torch.dtype,
-):
-    device = "cuda"
-    num_tokens = batch_size * seq_len
-    dim = intermediate_size
-    current_platform.seed_everything(42)
-    torch.set_default_device(device)
-
-    if func_name == "gelu_and_mul":
-        layer = CustomOp.op_registry[func_name](approximate="none")
-    elif func_name == "gelu_and_mul_tanh":
-        layer = CustomOp.op_registry["gelu_and_mul"](approximate="tanh")
-    elif func_name == "fatrelu_and_mul":
-        threshold = 0.5
-        layer = CustomOp.op_registry[func_name](threshold)
-    else:
-        layer = CustomOp.op_registry[func_name]()
-
-    x = torch.randn(num_tokens, dim, dtype=dtype, device=device)
-    compiled_layer = torch.compile(layer.forward_native)
-
-    if provider == "custom":
-        fn = lambda: layer(x)
-    elif provider == "compiled":
-        fn = lambda: compiled_layer(x)
-
-    ms, min_ms, max_ms = triton.testing.do_bench_cudagraph(
-        fn, quantiles=[0.5, 0.2, 0.8]
-    )
-    return ms, max_ms, min_ms
-
-
-if __name__ == "__main__":
-    parser = FlexibleArgumentParser(description="Benchmark the custom activation op.")
-    parser.add_argument(
-        "--func-name",
-        type=str,
-        choices=[
-            "mul_and_silu",
-            "silu_and_mul",
-            "gelu_and_mul",
-            "gelu_and_mul_tanh",
-            "fatrelu_and_mul",
-            "swigluoai_and_mul",
-            "gelu_new",
-            "gelu_fast",
-            "quick_gelu",
-        ],
-        default="silu_and_mul",
-    )
-    parser.add_argument(
-        "--dtype", type=str, choices=["half", "bfloat16", "float"], default="bfloat16"
-    )
-    args = parser.parse_args()
-    assert args
-
-    func_name = args.func_name
-    dtype = STR_DTYPE_TO_TORCH_DTYPE[args.dtype]
-
-    perf_report = triton.testing.perf_report(
-        triton.testing.Benchmark(
-            x_names=["batch_size", "seq_len", "intermediate_size"],
-            x_vals=configs,
-            line_arg="provider",
-            line_vals=["custom", "compiled"],
-            line_names=["Custom OP", "Compiled"],
-            styles=[("blue", "-"), ("green", "-")],
-            ylabel="ms",
-            plot_name=f"{func_name}-op-performance",
-            args={},
-        )
-    )
-
-    perf_report(
-        lambda batch_size, seq_len, intermediate_size, provider: benchmark_activation(
-            batch_size, seq_len, intermediate_size, provider, func_name, dtype
-        )
-    ).run(print_data=True)

benchmarks/kernels/benchmark_bitblas.py

@@ -28,7 +28,7 @@ except ImportError as e:
 
 from bitblas import Matmul, MatmulConfig, auto_detect_nvidia_target
 
-from vllm.utils.argparse_utils import FlexibleArgumentParser
+from vllm.utils import FlexibleArgumentParser
 
 parser = FlexibleArgumentParser(
     description="Benchmark BitBLAS int4 on a specific target."

benchmarks/kernels/benchmark_cutlass_fp4_moe.py

@@ -13,14 +13,10 @@ 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.config import (
-    fp8_w8a8_moe_quant_config,
-    nvfp4_moe_quant_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.argparse_utils import FlexibleArgumentParser
+from vllm.utils import FlexibleArgumentParser
 
 WEIGHT_SHAPES_MOE = {
     "nvidia/DeepSeek-R1-FP4": [
@@ -144,12 +140,6 @@ def bench_run(
         a_fp8_scale: torch.Tensor,
         num_repeats: int,
     ):
-        quant_config = fp8_w8a8_moe_quant_config(
-            w1_scale=w1_scale,
-            w2_scale=w2_scale,
-            a1_scale=a_fp8_scale,
-        )
-
         for _ in range(num_repeats):
             fused_experts(
                 a,
@@ -157,7 +147,10 @@ def bench_run(
                 w2,
                 topk_weights,
                 topk_ids,
-                quant_config=quant_config,
+                use_fp8_w8a8=True,
+                w1_scale=w1_scale,
+                w2_scale=w2_scale,
+                a1_scale=a_fp8_scale,
             )
 
     def run_cutlass_moe_fp4(
@@ -179,27 +172,25 @@ def bench_run(
         device: torch.device,
         num_repeats: int,
     ):
-        quant_config = nvfp4_moe_quant_config(
-            a1_gscale=a1_gs,
-            a2_gscale=a2_gs,
-            w1_scale=w1_blockscale,
-            w2_scale=w2_blockscale,
-            g1_alphas=w1_gs,
-            g2_alphas=w2_gs,
-        )
         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,
-                    quant_config=quant_config,
+                    device=device,
                 )
 
     def run_cutlass_from_graph(
@@ -220,29 +211,26 @@ def bench_run(
         e: int,
         device: torch.device,
     ):
-        quant_config = nvfp4_moe_quant_config(
-            a1_gscale=a1_gs,
-            a2_gscale=a2_gs,
-            w1_scale=w1_blockscale,
-            w2_scale=w2_blockscale,
-            g1_alphas=w1_gs,
-            g2_alphas=w2_gs,
-        )
-
         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,
-                quant_config=quant_config,
+                device=device,
             )
 
     def run_triton_from_graph(
@@ -258,18 +246,16 @@ def bench_run(
         with set_current_vllm_config(
             VllmConfig(parallel_config=ParallelConfig(pipeline_parallel_size=1))
         ):
-            quant_config = fp8_w8a8_moe_quant_config(
-                w1_scale=w1_scale,
-                w2_scale=w2_scale,
-                a1_scale=a_fp8_scale,
-            )
             return fused_experts(
                 a,
                 w1,
                 w2,
                 topk_weights,
                 topk_ids,
-                quant_config=quant_config,
+                use_fp8_w8a8=True,
+                w1_scale=w1_scale,
+                w2_scale=w2_scale,
+                a1_scale=a_fp8_scale,
             )
 
     def replay_graph(graph, num_repeats):

benchmarks/kernels/benchmark_cutlass_moe_fp8.py

@@ -1,406 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-"""
-Benchmark the performance of the cutlass_moe_fp8 kernel vs the triton_moe
-kernel. Both kernels take in fp8 quantized weights and 16-bit activations,
-but use different quantization strategies and backends.
-"""
-
-import nvtx
-import torch
-
-from vllm import _custom_ops as ops
-from vllm.model_executor.layers.fused_moe.config import fp8_w8a8_moe_quant_config
-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.platforms import current_platform
-from vllm.utils.argparse_utils import FlexibleArgumentParser
-
-# Weight shapes for different models: [num_experts, topk, hidden_size,
-# intermediate_size]
-WEIGHT_SHAPES_MOE = {
-    "mixtral-8x7b": [
-        [8, 2, 4096, 14336],
-    ],
-    "deepseek-v2": [
-        [160, 6, 5120, 12288],
-    ],
-    "custom-small": [
-        [8, 2, 2048, 7168],
-    ],
-    "glm45-fp8": [
-        [128, 8, 4096, 1408],
-    ],
-    "Llama-4-Maverick-17B-128E-Instruct-FP8": [
-        [128, 1, 5120, 8192],
-    ],
-}
-
-DEFAULT_MODELS = [
-    "mixtral-8x7b",
-]
-
-DEFAULT_BATCH_SIZES = [4, 8, 16, 32, 64, 128, 256, 512, 1024, 2048]
-DEFAULT_TP_SIZES = [1]
-
-PER_ACT_TOKEN_OPTS = [False, True]
-PER_OUT_CH_OPTS = [False, True]
-
-FP8_DTYPE = current_platform.fp8_dtype()
-
-
-def bench_run(
-    results: list,
-    model: str,
-    num_experts: int,
-    topk: int,
-    per_act_token: bool,
-    per_out_ch: bool,
-    mkn: tuple[int, int, int],
-):
-    (m, k, n) = mkn
-
-    dtype = torch.half
-    device = "cuda"
-
-    # Create input activations
-    a = torch.randn((m, k), device=device, dtype=dtype) / 10
-
-    # Create weights
-    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
-
-    # Create FP8 quantized weights and scales for both kernels
-    w1_fp8q = torch.empty((num_experts, 2 * n, k), device=device, dtype=FP8_DTYPE)
-    w2_fp8q = torch.empty((num_experts, k, n), device=device, dtype=FP8_DTYPE)
-
-    # Create scales based on quantization strategy
-    if per_out_ch:
-        # Per-channel quantization
-        w1_scale = torch.empty(
-            (num_experts, 2 * n, 1), device=device, dtype=torch.float32
-        )
-        w2_scale = torch.empty((num_experts, k, 1), device=device, dtype=torch.float32)
-    else:
-        # Per-tensor quantization
-        w1_scale = torch.empty((num_experts, 1, 1), device=device, dtype=torch.float32)
-        w2_scale = torch.empty((num_experts, 1, 1), device=device, dtype=torch.float32)
-
-    # Quantize weights
-    for expert in range(num_experts):
-        if per_out_ch:
-            # Per-channel quantization - not yet implemented properly
-            # For now, fall back to per-tensor quantization
-            w1_fp8q[expert], w1_scale_temp = ops.scaled_fp8_quant(w1[expert])
-            w2_fp8q[expert], w2_scale_temp = ops.scaled_fp8_quant(w2[expert])
-            # Expand scalar scales to the expected per-channel shape
-            w1_scale[expert] = w1_scale_temp.expand(2 * n, 1)
-            w2_scale[expert] = w2_scale_temp.expand(k, 1)
-        else:
-            # Per-tensor quantization
-            w1_fp8q[expert], w1_scale_temp = ops.scaled_fp8_quant(w1[expert])
-            w2_fp8q[expert], w2_scale_temp = ops.scaled_fp8_quant(w2[expert])
-            # Store scalar scales in [1, 1] tensors
-            w1_scale[expert, 0, 0] = w1_scale_temp
-            w2_scale[expert, 0, 0] = w2_scale_temp
-
-    # Prepare weights for CUTLASS (no transpose needed)
-    w1_fp8q_cutlass = w1_fp8q  # Keep original [E, 2N, K]
-    w2_fp8q_cutlass = w2_fp8q  # Keep original [E, K, N]
-
-    # Create router scores and get topk
-    score = torch.randn((m, num_experts), device=device, dtype=dtype)
-    topk_weights, topk_ids, _ = fused_topk(a, score, topk, renormalize=False)
-
-    # WORKAROUND: CUTLASS MoE FP8 has issues with per-token quantization
-    # Force per-tensor quantization for all cases to match working e2e setup
-    a1_scale = torch.full((), 1e-2, device=device, dtype=torch.float32)
-    a2_scale = torch.full((), 1e-2, device=device, dtype=torch.float32)
-
-    # Force per-tensor quantization for all cases
-    per_act_token = False
-
-    # Create stride tensors for CUTLASS
-    ab_strides1 = torch.full((num_experts,), k, dtype=torch.int64, device=device)
-    ab_strides2 = torch.full((num_experts,), n, dtype=torch.int64, device=device)
-    c_strides1 = torch.full((num_experts,), 2 * n, dtype=torch.int64, device=device)
-    c_strides2 = torch.full((num_experts,), k, dtype=torch.int64, device=device)
-
-    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,
-        a1_scale: torch.Tensor,
-        a2_scale: torch.Tensor,
-        num_repeats: int,
-    ):
-        quant_config = fp8_w8a8_moe_quant_config(
-            w1_scale=w1_scale,
-            w2_scale=w2_scale,
-            a1_scale=a1_scale,
-            a2_scale=a2_scale,
-            per_act_token_quant=per_act_token,
-            per_out_ch_quant=per_out_ch,
-        )
-
-        for _ in range(num_repeats):
-            fused_experts(
-                a,
-                w1,
-                w2,
-                topk_weights,
-                topk_ids,
-                quant_config=quant_config,
-            )
-
-    def run_cutlass_moe_fp8(
-        a: torch.Tensor,
-        w1: torch.Tensor,
-        w2: torch.Tensor,
-        topk_weights: torch.Tensor,
-        topk_ids: torch.Tensor,
-        ab_strides1: torch.Tensor,
-        ab_strides2: torch.Tensor,
-        c_strides1: torch.Tensor,
-        c_strides2: torch.Tensor,
-        w1_scale: torch.Tensor,
-        w2_scale: torch.Tensor,
-        a1_scale: torch.Tensor,
-        a2_scale: torch.Tensor,
-        num_repeats: int,
-    ):
-        quant_config = fp8_w8a8_moe_quant_config(
-            w1_scale=w1_scale,
-            w2_scale=w2_scale,
-            a1_scale=a1_scale,
-            a2_scale=a2_scale,
-            per_act_token_quant=per_act_token,
-            per_out_ch_quant=per_out_ch,
-        )
-
-        for _ in range(num_repeats):
-            with nvtx.annotate("cutlass_moe_fp8", color="blue"):
-                cutlass_moe_fp8(
-                    a=a,
-                    w1_q=w1,
-                    w2_q=w2,
-                    topk_weights=topk_weights,
-                    topk_ids=topk_ids,
-                    ab_strides1=ab_strides1,
-                    ab_strides2=ab_strides2,
-                    c_strides1=c_strides1,
-                    c_strides2=c_strides2,
-                    quant_config=quant_config,
-                    activation="silu",
-                    global_num_experts=num_experts,
-                )
-
-    # Pre-create quantization config to avoid creating it inside CUDA graph
-    quant_config = fp8_w8a8_moe_quant_config(
-        w1_scale=w1_scale,
-        w2_scale=w2_scale,
-        a1_scale=a1_scale,
-        a2_scale=a2_scale,
-        per_act_token_quant=per_act_token,
-        per_out_ch_quant=per_out_ch,
-    )
-
-    # Create CUDA graphs for CUTLASS (match benchmark_moe.py pattern exactly)
-    cutlass_stream = torch.cuda.Stream()
-    cutlass_graph = torch.cuda.CUDAGraph()
-    with torch.cuda.graph(cutlass_graph, stream=cutlass_stream):
-        # Capture 10 invocations like benchmark_moe.py
-        for _ in range(10):
-            cutlass_moe_fp8(
-                a=a,
-                w1_q=w1_fp8q_cutlass,
-                w2_q=w2_fp8q_cutlass,
-                topk_weights=topk_weights,
-                topk_ids=topk_ids,
-                ab_strides1=ab_strides1,
-                ab_strides2=ab_strides2,
-                c_strides1=c_strides1,
-                c_strides2=c_strides2,
-                quant_config=quant_config,
-                activation="silu",
-                global_num_experts=num_experts,
-            )
-    torch.cuda.synchronize()
-
-    # Create CUDA graphs for Triton (match benchmark_moe.py pattern exactly)
-    triton_stream = torch.cuda.Stream()
-    triton_graph = torch.cuda.CUDAGraph()
-    with torch.cuda.graph(triton_graph, stream=triton_stream):
-        # Capture 10 invocations like benchmark_moe.py
-        for _ in range(10):
-            fused_experts(
-                a,
-                w1_fp8q,
-                w2_fp8q,
-                topk_weights,
-                topk_ids,
-                quant_config=quant_config,
-            )
-    torch.cuda.synchronize()
-
-    def bench_cuda_graph(graph, num_warmup=5, num_iters=100):
-        """Benchmark CUDA graph using events like benchmark_moe.py"""
-        # Warmup
-        for _ in range(num_warmup):
-            graph.replay()
-        torch.cuda.synchronize()
-
-        # Timing
-        start_event = torch.cuda.Event(enable_timing=True)
-        end_event = torch.cuda.Event(enable_timing=True)
-
-        latencies = []
-        for _ 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))
-
-        # Divide by 10 since graph contains 10 calls
-        return sum(latencies) / (num_iters * 10)
-
-    # Benchmark parameters
-    num_warmup = 5
-    num_iters = 100
-
-    # Benchmark only CUDA graphs (more reliable and faster)
-    # Benchmark Triton MoE with CUDA graphs
-    triton_graph_time = bench_cuda_graph(
-        triton_graph, num_warmup=num_warmup, num_iters=num_iters
-    )
-
-    # Benchmark CUTLASS MoE with CUDA graphs
-    cutlass_graph_time = bench_cuda_graph(
-        cutlass_graph, num_warmup=num_warmup, num_iters=num_iters
-    )
-
-    # Convert ms to us and return results
-    triton_time_us = triton_graph_time * 1000
-    cutlass_time_us = cutlass_graph_time * 1000
-
-    return {
-        "batch_size": m,
-        "triton_time_us": triton_time_us,
-        "cutlass_time_us": cutlass_time_us,
-    }
-
-
-def main(args):
-    print("Benchmarking models:")
-    for i, model in enumerate(args.models):
-        print(f"[{i}]  {model}")
-
-    all_results = []
-
-    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 args.per_act_token_opts:
-                    for per_out_ch in args.per_out_ch_opts:
-                        print(
-                            f"\n=== {model}, experts={num_experts}, topk={topk},"
-                            f"per_act={per_act_token}, per_out_ch={per_out_ch} ==="
-                        )
-
-                        config_results = []
-                        for size_m in args.batch_sizes:
-                            mkn = (size_m, size_k, size_n)
-                            result = bench_run(
-                                [],  # Not used anymore
-                                model,
-                                num_experts,
-                                topk,
-                                per_act_token,
-                                per_out_ch,
-                                mkn,
-                            )
-                            if result:
-                                config_results.append(result)
-
-                        # Print results table for this configuration
-                        if config_results:
-                            print(
-                                f"\n{'Batch Size':<12}"
-                                f"{'Triton (us)':<15}"
-                                f"{'CUTLASS (us)':<15}"
-                            )
-                            print("-" * 45)
-                            for result in config_results:
-                                print(
-                                    f"{result['batch_size']:<12}"
-                                    f"{result['triton_time_us']:<15.2f}"
-                                    f"{result['cutlass_time_us']:<15.2f}"
-                                )
-
-                            all_results.extend(config_results)
-
-    print(f"\nTotal benchmarks completed: {len(all_results)}")
-
-
-if __name__ == "__main__":
-    parser = FlexibleArgumentParser(
-        description="""Benchmark CUTLASS FP8 MOE vs Triton FP8 FUSED MOE
-         across specified models/shapes/batches
-
-        Example usage:
-        python benchmark_cutlass_moe_fp8.py  \
-            --model "Llama-4-Maverick-17B-128E-Instruct-FP8"  \
-            --tp-sizes 8 \
-            --batch-size 2 4 8  \
-            --per-act-token-opts false \
-            --per-out-ch-opts false
-
-        """
-    )
-    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(
-        "--per-act-token-opts",
-        nargs="+",
-        type=lambda x: x.lower() == "true",
-        default=[False, True],
-        help="Per-activation token quantization options (true/false)",
-    )
-    parser.add_argument(
-        "--per-out-ch-opts",
-        nargs="+",
-        type=lambda x: x.lower() == "true",
-        default=[False, True],
-        help="Per-output channel quantization options (true/false)",
-    )
-
-    args = parser.parse_args()
-    main(args)

benchmarks/kernels/benchmark_device_communicators.py

@@ -1,508 +0,0 @@
-#!/usr/bin/env python3
-# SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-
-"""
-Benchmark script for device communicators:
-CustomAllreduce (oneshot, twoshot), PyNcclCommunicator,
-and SymmMemCommunicator (multimem, two-shot).
-
-for NCCL symmetric memory you need to set the environment variables
-NCCL_NVLS_ENABLE=1 NCCL_CUMEM_ENABLE=1 VLLM_USE_NCCL_SYMM_MEM=1, otherwise NCCL does
-not use fast NVLS implementation for all reduce.
-
-Usage:
-    torchrun --nproc_per_node=<N> benchmark_device_communicators.py [options]
-
-Example:
-    torchrun --nproc_per_node=2 benchmark_device_communicators.py
-    --sequence-lengths 512 1024 2048 --num-warmup 10 --num-trials 100
-"""
-
-import json
-import os
-import time
-from collections.abc import Callable
-from contextlib import nullcontext
-
-import torch
-import torch.distributed as dist
-from torch.distributed import ProcessGroup
-
-from vllm.distributed.device_communicators.custom_all_reduce import CustomAllreduce
-from vllm.distributed.device_communicators.pynccl import (
-    PyNcclCommunicator,
-    register_nccl_symmetric_ops,
-)
-from vllm.distributed.device_communicators.pynccl_allocator import (
-    set_graph_pool_id,
-)
-from vllm.distributed.device_communicators.symm_mem import SymmMemCommunicator
-from vllm.logger import init_logger
-from vllm.utils.argparse_utils import FlexibleArgumentParser
-
-logger = init_logger(__name__)
-
-# Default sequence lengths to benchmark
-DEFAULT_SEQUENCE_LENGTHS = [128, 512, 1024, 2048, 4096, 8192]
-
-# Fixed hidden size and dtype for all benchmarks
-HIDDEN_SIZE = 8192
-BENCHMARK_DTYPE = torch.bfloat16
-
-# CUDA graph settings
-CUDA_GRAPH_CAPTURE_CYCLES = 10
-
-
-class CommunicatorBenchmark:
-    """Benchmark class for testing device communicators."""
-
-    def __init__(
-        self,
-        rank: int,
-        world_size: int,
-        device: torch.device,
-        cpu_group: ProcessGroup,
-        sequence_lengths: list[int],
-    ):
-        self.rank = rank
-        self.world_size = world_size
-        self.device = device
-        self.cpu_group = cpu_group
-
-        # Calculate max_size_override based on largest sequence length
-        max_seq_len = max(sequence_lengths)
-        max_tensor_elements = max_seq_len * HIDDEN_SIZE
-        self.max_size_override = max_tensor_elements * BENCHMARK_DTYPE.itemsize + 1
-
-        # Initialize communicators
-        self.custom_allreduce = None
-        self.pynccl_comm = None
-        self.symm_mem_comm = None
-        self.symm_mem_comm_multimem = None
-        self.symm_mem_comm_two_shot = None
-
-        self._init_communicators()
-
-    def _init_communicators(self):
-        """Initialize all available communicators."""
-        try:
-            self.custom_allreduce = CustomAllreduce(
-                group=self.cpu_group,
-                device=self.device,
-                max_size=self.max_size_override,
-            )
-            if not self.custom_allreduce.disabled:
-                logger.info("Rank %s: CustomAllreduce initialized", self.rank)
-            else:
-                logger.info("Rank %s: CustomAllreduce disabled", self.rank)
-        except Exception as e:
-            logger.warning(
-                "Rank %s: Failed to initialize CustomAllreduce: %s", self.rank, e
-            )
-            self.custom_allreduce = None
-
-        try:
-            self.pynccl_comm = PyNcclCommunicator(
-                group=self.cpu_group, device=self.device
-            )
-            if not self.pynccl_comm.disabled:
-                logger.info("Rank %s: PyNcclCommunicator initialized", self.rank)
-                register_nccl_symmetric_ops(self.pynccl_comm)
-            else:
-                logger.info("Rank %s: PyNcclCommunicator disabled", self.rank)
-                self.pynccl_comm = None
-        except Exception as e:
-            logger.warning(
-                "Rank %s: Failed to initialize PyNcclCommunicator: %s", self.rank, e
-            )
-            self.pynccl_comm = None
-
-        # Initialize variants for SymmMemCommunicator
-        try:
-            self.symm_mem_comm_multimem = SymmMemCommunicator(
-                group=self.cpu_group,
-                device=self.device,
-                force_multimem=True,
-                max_size_override=self.max_size_override,
-            )
-            if not self.symm_mem_comm_multimem.disabled:
-                logger.info(
-                    "Rank %s: SymmMemCommunicator (multimem) initialized", self.rank
-                )
-            else:
-                self.symm_mem_comm_multimem = None
-        except Exception as e:
-            logger.warning(
-                "Rank %s: Failed to initialize SymmMemCommunicator (multimem): %s",
-                self.rank,
-                e,
-            )
-            self.symm_mem_comm_multimem = None
-
-        try:
-            self.symm_mem_comm_two_shot = SymmMemCommunicator(
-                group=self.cpu_group,
-                device=self.device,
-                force_multimem=False,
-                max_size_override=self.max_size_override,
-            )
-            if not self.symm_mem_comm_two_shot.disabled:
-                logger.info(
-                    "Rank %s: SymmMemCommunicator (two_shot) initialized", self.rank
-                )
-            else:
-                self.symm_mem_comm_two_shot = None
-        except Exception as e:
-            logger.warning(
-                "Rank %s: Failed to initialize SymmMemCommunicator (two_shot): %s",
-                self.rank,
-                e,
-            )
-            self.symm_mem_comm_two_shot = None
-
-    def benchmark_allreduce(
-        self, sequence_length: int, num_warmup: int, num_trials: int
-    ) -> dict[str, float]:
-        """Benchmark allreduce operations for all available communicators."""
-
-        results = {}
-
-        # Define communicators with their benchmark functions
-        communicators = []
-
-        if self.custom_allreduce is not None:
-            comm = self.custom_allreduce
-            # CustomAllreduce one-shot
-            communicators.append(
-                (
-                    "ca_1stage",
-                    lambda t, c=comm: c.custom_all_reduce(t),
-                    lambda t, c=comm: c.should_custom_ar(t),
-                    comm.capture(),
-                    "1stage",  # env variable value
-                )
-            )
-            # CustomAllreduce two-shot
-            communicators.append(
-                (
-                    "ca_2stage",
-                    lambda t, c=comm: c.custom_all_reduce(t),
-                    lambda t, c=comm: c.should_custom_ar(t),
-                    comm.capture(),
-                    "2stage",  # env variable value
-                )
-            )
-
-        if self.pynccl_comm is not None:
-            comm = self.pynccl_comm
-            communicators.append(
-                (
-                    "pynccl",
-                    lambda t, c=comm: c.all_reduce(t),
-                    lambda t: True,  # Always available if initialized
-                    nullcontext(),
-                    None,  # no env variable needed
-                )
-            )
-            communicators.append(
-                (
-                    "pynccl-symm",
-                    lambda t: torch.ops.vllm.all_reduce_symmetric_with_copy(t),
-                    lambda t: True,  # Always available if initialized
-                    nullcontext(),
-                    None,  # no env variable needed
-                )
-            )
-
-        if self.symm_mem_comm_multimem is not None:
-            comm = self.symm_mem_comm_multimem
-            communicators.append(
-                (
-                    "symm_mem_multimem",
-                    lambda t, c=comm: c.all_reduce(t),
-                    lambda t, c=comm: c.should_use_symm_mem(t),
-                    nullcontext(),
-                    None,  # no env variable needed
-                )
-            )
-
-        if self.symm_mem_comm_two_shot is not None:
-            comm = self.symm_mem_comm_two_shot
-            communicators.append(
-                (
-                    "symm_mem_two_shot",
-                    lambda t, c=comm: c.all_reduce(t),
-                    lambda t, c=comm: c.should_use_symm_mem(t),
-                    nullcontext(),
-                    None,  # no env variable needed
-                )
-            )
-
-        # Benchmark each communicator
-        for name, allreduce_fn, should_use_fn, context, env_var in communicators:
-            # Set environment variable if needed
-            if env_var is not None:
-                os.environ["VLLM_CUSTOM_ALLREDUCE_ALGO"] = env_var
-            else:
-                # Clear the environment variable to avoid interference
-                os.environ.pop("VLLM_CUSTOM_ALLREDUCE_ALGO", None)
-
-            latency = self.benchmark_allreduce_single(
-                sequence_length,
-                allreduce_fn,
-                should_use_fn,
-                context,
-                num_warmup,
-                num_trials,
-            )
-            if latency is not None:
-                results[name] = latency
-
-        return results
-
-    def benchmark_allreduce_single(
-        self,
-        sequence_length: int,
-        allreduce_fn: Callable[[torch.Tensor], torch.Tensor | None],
-        should_use_fn: Callable[[torch.Tensor], bool],
-        context,
-        num_warmup: int,
-        num_trials: int,
-    ) -> float | None:
-        """Benchmark method with CUDA graph optimization."""
-        try:
-            # Create test tensor (2D: sequence_length x hidden_size)
-            tensor = torch.randn(
-                sequence_length, HIDDEN_SIZE, dtype=BENCHMARK_DTYPE, device=self.device
-            )
-            if not should_use_fn(tensor):
-                return None
-
-            torch.cuda.synchronize()
-            stream = torch.cuda.Stream()
-            with torch.cuda.stream(stream):
-                graph_input = tensor.clone()
-
-                # Warmup before capture
-                for _ in range(3):
-                    allreduce_fn(graph_input)
-
-                # Capture the graph using context manager
-                with context:
-                    graph = torch.cuda.CUDAGraph()
-                    graph_pool = torch.cuda.graph_pool_handle()
-                    set_graph_pool_id(graph_pool)
-                    with torch.cuda.graph(graph, pool=graph_pool):
-                        for _ in range(CUDA_GRAPH_CAPTURE_CYCLES):
-                            allreduce_fn(graph_input)
-
-            torch.cuda.synchronize()
-            for _ in range(num_warmup):
-                graph.replay()
-            torch.cuda.synchronize()
-
-            torch.cuda.synchronize()
-            start_time = time.perf_counter()
-
-            for _ in range(num_trials):
-                graph.replay()
-            torch.cuda.synchronize()
-
-            end_time = time.perf_counter()
-
-            # Convert to ms and divide by CUDA_GRAPH_CAPTURE_CYCLES
-            return (
-                (end_time - start_time) / num_trials / CUDA_GRAPH_CAPTURE_CYCLES * 1000
-            )
-
-        except Exception as e:
-            logger.error("CUDA graph benchmark failed: %s", e)
-            raise RuntimeError(
-                f"CUDA graph benchmark failed for communicator: {e}"
-            ) from e
-
-
-def _calculate_speedup_info(comm_results: dict[str, float]) -> str:
-    """Calculate speedup information for a single tensor size."""
-    if not comm_results:
-        return "N/A"
-
-    # Find the fastest communicator
-    fastest_comm = min(comm_results.keys(), key=lambda k: comm_results[k])
-    fastest_time = comm_results[fastest_comm]
-
-    # Calculate speedup vs PyNccl if available
-    if "pynccl" in comm_results:
-        pynccl_time = comm_results["pynccl"]
-        speedup = pynccl_time / fastest_time
-        return f"{fastest_comm} ({speedup:.2f}x)"
-    else:
-        return f"{fastest_comm} (N/A)"
-
-
-def print_results(
-    results: dict[str, dict[str, float]], sequence_lengths: list[int], world_size: int
-):
-    """Print benchmark results in a formatted table."""
-
-    print(f"\n{'=' * 130}")
-    print("Device Communicator Benchmark Results")
-    print(
-        f"World Size: {world_size}, Data Type: {BENCHMARK_DTYPE}, "
-        f"Hidden Size: {HIDDEN_SIZE}"
-    )
-    print(f"{'=' * 130}")
-
-    # Get all communicator names
-    all_comms = set()
-    for size_results in results.values():
-        all_comms.update(size_results.keys())
-
-    all_comms = sorted(list(all_comms))
-
-    # Print header
-    header = f"{'Tensor Shape':<20}{'Tensor Size':<15}"
-    for comm in all_comms:
-        header += f"{comm:<20}"
-    header += f"{'Best (Speedup vs PyNccl)':<30}"
-    print(header)
-    print("-" * len(header))
-
-    # Print results for each sequence length
-    for seq_len in sequence_lengths:
-        if seq_len in results:
-            # Calculate tensor size in elements and bytes
-            tensor_elements = seq_len * HIDDEN_SIZE
-            tensor_bytes = tensor_elements * BENCHMARK_DTYPE.itemsize
-
-            # Format tensor size (MB)
-            tensor_size_mb = tensor_bytes / (1024 * 1024)
-            tensor_size_str = f"{tensor_size_mb:.2f} MB"
-
-            # Format tensor shape
-            tensor_shape = f"({seq_len}, {HIDDEN_SIZE})"
-
-            row = f"{tensor_shape:<20}{tensor_size_str:<15}"
-            for comm in all_comms:
-                if comm in results[seq_len]:
-                    row += f"{results[seq_len][comm]:<20.3f}"
-                else:
-                    row += f"{'N/A':<20}"
-
-            # Calculate speedup information
-            speedup_info = _calculate_speedup_info(results[seq_len])
-            row += f"{speedup_info:<30}"
-
-            print(row)
-
-    print(f"{'=' * 130}")
-    print("All times are in milliseconds (ms) per allreduce operation")
-    print("Speedup column shows: fastest_algorithm (speedup_vs_pynccl)")
-
-
-def main():
-    parser = FlexibleArgumentParser(description="Benchmark device communicators")
-
-    parser.add_argument(
-        "--sequence-lengths",
-        type=int,
-        nargs="+",
-        default=DEFAULT_SEQUENCE_LENGTHS,
-        help="Sequence lengths to benchmark (tensor shape: seq_len x hidden_size)",
-    )
-
-    parser.add_argument(
-        "--num-warmup", type=int, default=5, help="Number of warmup iterations"
-    )
-
-    parser.add_argument(
-        "--num-trials", type=int, default=50, help="Number of benchmark trials"
-    )
-
-    parser.add_argument("--output-json", type=str, help="Output results to JSON file")
-
-    args = parser.parse_args()
-
-    # Initialize distributed
-    if not dist.is_initialized():
-        dist.init_process_group(backend="gloo")
-    rank = dist.get_rank()
-    world_size = dist.get_world_size()
-
-    # Set device
-    device = torch.device(f"cuda:{rank}")
-    torch.cuda.set_device(device)
-
-    # Get CPU process group
-    cpu_group = dist.new_group(backend="gloo")
-
-    # Disable USE_SYMM_MEM to avoid affecting the max_sizes
-    # in symm_mem and custom_all_reduce for benchmark
-    os.environ["VLLM_ALLREDUCE_USE_SYMM_MEM"] = "0"
-
-    # Initialize benchmark
-    benchmark = CommunicatorBenchmark(
-        rank, world_size, device, cpu_group, args.sequence_lengths
-    )
-
-    # Run benchmarks
-    all_results = {}
-
-    for seq_len in args.sequence_lengths:
-        if rank == 0:
-            logger.info(
-                "Benchmarking sequence length: %s (tensor shape: %s x %s)",
-                seq_len,
-                seq_len,
-                HIDDEN_SIZE,
-            )
-
-        results = benchmark.benchmark_allreduce(
-            sequence_length=seq_len,
-            num_warmup=args.num_warmup,
-            num_trials=args.num_trials,
-        )
-
-        all_results[seq_len] = results
-
-        # Synchronize between ranks
-        dist.barrier()
-
-    # Print results (only rank 0)
-    if rank == 0:
-        print_results(all_results, args.sequence_lengths, world_size)
-
-        # Save to JSON if requested
-        if args.output_json:
-            # Add speedup information to results
-            enhanced_results = {}
-            for seq_len, comm_results in all_results.items():
-                enhanced_results[seq_len] = {
-                    "timings": comm_results,
-                    "speedup_info": _calculate_speedup_info(comm_results),
-                }
-
-            output_data = {
-                "world_size": world_size,
-                "dtype": str(BENCHMARK_DTYPE),
-                "hidden_size": HIDDEN_SIZE,
-                "sequence_lengths": args.sequence_lengths,
-                "num_warmup": args.num_warmup,
-                "num_trials": args.num_trials,
-                "cuda_graph_capture_cycles": CUDA_GRAPH_CAPTURE_CYCLES,
-                "results": enhanced_results,
-            }
-
-            with open(args.output_json, "w") as f:
-                json.dump(output_data, f, indent=2)
-
-            logger.info("Results saved to %s", args.output_json)
-
-    # Cleanup
-    if cpu_group != dist.group.WORLD:
-        dist.destroy_process_group(cpu_group)
-
-
-if __name__ == "__main__":
-    main()

benchmarks/kernels/benchmark_grouped_gemm_cutlass.py

@@ -7,13 +7,12 @@ 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.config import fp8_w8a8_moe_quant_config
 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.argparse_utils import FlexibleArgumentParser
+from vllm.utils import FlexibleArgumentParser
 
 DEFAULT_MODELS = [
     "nm-testing/Mixtral-8x7B-Instruct-v0.1",
@@ -97,11 +96,6 @@ def bench_run(
         a_scale: torch.Tensor,
         num_repeats: int,
     ):
-        quant_config = fp8_w8a8_moe_quant_config(
-            w1_scale=w1_scale,
-            w2_scale=w2_scale,
-            a1_scale=a_scale,
-        )
         for _ in range(num_repeats):
             fused_experts(
                 a,
@@ -109,7 +103,10 @@ def bench_run(
                 w2,
                 topk_weights,
                 topk_ids,
-                quant_config=quant_config,
+                use_fp8_w8a8=True,
+                w1_scale=w1_scale,
+                w2_scale=w2_scale,
+                a1_scale=a_scale,
             )
 
     def run_cutlass_moe(
@@ -128,12 +125,6 @@ def bench_run(
         per_act_token: bool,
         num_repeats: int,
     ):
-        quant_config = fp8_w8a8_moe_quant_config(
-            w1_scale=w1_scale,
-            w2_scale=w2_scale,
-            per_act_token_quant=per_act_token,
-        )
-
         for _ in range(num_repeats):
             cutlass_moe_fp8(
                 a,
@@ -141,11 +132,14 @@ def bench_run(
                 w2,
                 topk_weights,
                 topk_ids,
+                w1_scale,
+                w2_scale,
                 ab_strides1,
                 ab_strides2,
                 c_strides1,
                 c_strides2,
-                quant_config=quant_config,
+                per_act_token,
+                a1_scale=None,
             )
 
     def run_cutlass_from_graph(
@@ -162,12 +156,6 @@ def bench_run(
         topk_weights: torch.Tensor,
         topk_ids: torch.Tensor,
     ):
-        quant_config = fp8_w8a8_moe_quant_config(
-            w1_scale=w1_scale,
-            w2_scale=w2_scale,
-            per_act_token_quant=per_act_token,
-        )
-
         with set_current_vllm_config(
             VllmConfig(parallel_config=ParallelConfig(pipeline_parallel_size=1))
         ):
@@ -177,11 +165,14 @@ def bench_run(
                 w2_q,
                 topk_weights,
                 topk_ids,
+                w1_scale,
+                w2_scale,
                 ab_strides1,
                 ab_strides2,
                 c_strides1,
                 c_strides2,
-                quant_config=quant_config,
+                per_act_token,
+                a1_scale=None,
             )
 
     def run_triton_from_graph(
@@ -194,11 +185,6 @@ def bench_run(
         w2_scale: torch.Tensor,
         a_scale: torch.Tensor,
     ):
-        quant_config = fp8_w8a8_moe_quant_config(
-            w1_scale=w1_scale,
-            w2_scale=w2_scale,
-            a1_scale=a_scale,
-        )
         with set_current_vllm_config(
             VllmConfig(parallel_config=ParallelConfig(pipeline_parallel_size=1))
         ):
@@ -208,7 +194,10 @@ def bench_run(
                 w2,
                 topk_weights,
                 topk_ids,
-                quant_config=quant_config,
+                use_fp8_w8a8=True,
+                w1_scale=w1_scale,
+                w2_scale=w2_scale,
+                a1_scale=a_scale,
             )
 
     def replay_graph(graph, num_repeats):

benchmarks/kernels/benchmark_layernorm.py

@@ -7,8 +7,7 @@ import torch
 
 from vllm.model_executor.layers.layernorm import RMSNorm
 from vllm.platforms import current_platform
-from vllm.utils.argparse_utils import FlexibleArgumentParser
-from vllm.utils.torch_utils import STR_DTYPE_TO_TORCH_DTYPE
+from vllm.utils import STR_DTYPE_TO_TORCH_DTYPE, FlexibleArgumentParser
 
 
 @torch.inference_mode()

benchmarks/kernels/benchmark_lora.py

@@ -6,12 +6,11 @@ import copy
 import json
 import pickle
 import time
-from collections.abc import Callable
 from dataclasses import dataclass
 from enum import Enum, auto
 from itertools import product
 from pathlib import Path
-from typing import Any
+from typing import Any, Callable, Optional
 
 import torch
 import torch.utils.benchmark as TBenchmark
@@ -25,7 +24,7 @@ if HAS_TRITON:
     from vllm.lora.ops.triton_ops import LoRAKernelMeta, lora_expand, lora_shrink
     from vllm.lora.ops.triton_ops.utils import _LORA_A_PTR_DICT, _LORA_B_PTR_DICT
 
-from vllm.utils.argparse_utils import FlexibleArgumentParser
+from vllm.utils import FlexibleArgumentParser
 
 DEFAULT_MODELS = list(WEIGHT_SHAPES.keys())
 DEFAULT_TP_SIZES = [1]
@@ -80,9 +79,9 @@ def make_rand_lora_weight_tensor(
 
 
 def make_rand_tensors(
-    a_shape: tuple[int, ...],
-    b_shape: tuple[int, ...],
-    c_shape: tuple[int, ...],
+    a_shape: tuple[int],
+    b_shape: tuple[int],
+    c_shape: tuple[int],
     a_dtype: torch.dtype,
     b_dtype: torch.dtype,
     c_dtype: torch.dtype,
@@ -159,7 +158,7 @@ def ref_group_gemm(
     seq_lens_cpu: torch.Tensor,
     prompt_lora_mapping_cpu: torch.Tensor,
     scaling: float,
-    add_inputs: bool | None,
+    add_inputs: Optional[bool],
 ):
     """
     Torch group gemm reference implementation to test correctness of
@@ -244,7 +243,7 @@ class OpType(Enum):
         lora_rank: int,
         num_loras: int,
         num_slices: int,
-    ) -> tuple[tuple[int, ...], tuple[int, ...], tuple[int, ...]]:
+    ) -> tuple[tuple[int], tuple[int], tuple[int]]:
         """
         Given num_slices, return the shapes of the A, B, and C matrices
         in A x B = C, for the op_type
@@ -317,8 +316,8 @@ class BenchmarkContext:
     lora_rank: int
     sort_by_lora_id: bool
     dtype: torch.dtype
-    seq_length: int | None = None
-    num_slices: int | None = None  # num_slices for slice based ops
+    seq_length: Optional[int] = None
+    num_slices: Optional[int] = None  # num_slices for slice based ops
 
     def with_seq_length(self, seq_length: int) -> "BenchmarkContext":
         ctx = copy.copy(self)
@@ -465,11 +464,7 @@ class BenchmarkTensors:
         for field_name in LoRAKernelMeta.__dataclass_fields__:
             field = getattr(self.lora_kernel_meta, field_name)
             assert isinstance(field, torch.Tensor)
-            setattr(
-                self.lora_kernel_meta,
-                field_name,
-                to_device(field) if field_name != "no_lora_flag_cpu" else field,
-            )
+            setattr(self.lora_kernel_meta, field_name, to_device(field))
 
     def metadata(self) -> tuple[int, int, int]:
         """
@@ -517,7 +512,6 @@ class BenchmarkTensors:
             "lora_token_start_loc": self.lora_kernel_meta.lora_token_start_loc,
             "lora_ids": self.lora_kernel_meta.active_lora_ids,
             "scaling": 1.0,
-            "no_lora_flag_cpu": self.lora_kernel_meta.no_lora_flag_cpu,
         }
 
     def as_lora_expand_kwargs(self, add_inputs: bool) -> dict[str, Any]:
@@ -558,11 +552,10 @@ class BenchmarkTensors:
             "lora_ids": self.lora_kernel_meta.active_lora_ids,
             "offset_start": 0,
             "add_inputs": add_inputs,
-            "no_lora_flag_cpu": self.lora_kernel_meta.no_lora_flag_cpu,
         }
 
     def bench_fn_kwargs(
-        self, op_type: OpType, add_inputs: bool | None = None
+        self, op_type: OpType, add_inputs: Optional[bool] = None
     ) -> dict[str, Any]:
         if op_type.is_shrink_fn():
             assert add_inputs is None
@@ -576,7 +569,7 @@ class BenchmarkTensors:
         raise ValueError(f"Unrecognized optype {self}")
 
     def test_correctness(
-        self, op_type: OpType, expand_fn_add_inputs: bool | None
+        self, op_type: OpType, expand_fn_add_inputs: Optional[bool]
     ) -> bool:
         """
         Test correctness of op_type implementation against a grouped gemm
@@ -612,8 +605,8 @@ def bench_optype(
     ctx: BenchmarkContext,
     arg_pool_size: int,
     op_type: OpType,
-    cuda_graph_nops: int | None = None,
-    expand_fn_add_inputs: bool | None = None,
+    cuda_graph_nops: Optional[int] = None,
+    expand_fn_add_inputs: Optional[bool] = None,
     test_correctness: bool = False,
 ) -> TMeasurement:
     assert arg_pool_size >= 1
@@ -644,7 +637,7 @@ def bench_optype(
     # Clear LoRA optimization hash-maps.
     _LORA_A_PTR_DICT.clear()
     _LORA_B_PTR_DICT.clear()
-    # Run bench function so that _LORA_A_PTR_DICT and _LORA_B_PTR_DICT are set up
+    # Run bench function so that _LORA_A_PTR_DICT and _LORA_B_PTR_DICT are setup
     for kwargs in kwargs_list:
         op_type.bench_fn()(**kwargs)
     torch.cuda.synchronize()
@@ -680,7 +673,7 @@ def bench_torch_mm(
     ctx: BenchmarkContext,
     arg_pool_size: int,
     op_type: OpType,
-    cuda_graph_nops: int | None = None,
+    cuda_graph_nops: Optional[int] = None,
 ) -> TMeasurement:
     """
     Benchmark basic torch.mm as a roofline.
@@ -745,7 +738,7 @@ def use_cuda_graph_recommendation() -> str:
             """
 
 
-def print_timers(timers: list[TMeasurement], args: argparse.Namespace | None = None):
+def print_timers(timers: list[TMeasurement], args: Optional[argparse.Namespace] = None):
     compare = TBenchmark.Compare(timers)
     compare.print()
 

benchmarks/kernels/benchmark_machete.py

@@ -8,9 +8,10 @@ import math
 import os
 import pickle as pkl
 import time
-from collections.abc import Callable, Iterable
+from collections.abc import Iterable
 from dataclasses import dataclass
 from itertools import product
+from typing import Callable, Optional
 
 import pandas as pd
 import torch
@@ -33,7 +34,7 @@ from vllm.model_executor.layers.quantization.utils.quant_utils import (
     quantize_weights,
 )
 from vllm.scalar_type import ScalarType, scalar_types
-from vllm.utils.argparse_utils import FlexibleArgumentParser
+from vllm.utils import FlexibleArgumentParser
 
 DEFAULT_MODELS = ["meta-llama/Llama-3-8b", "meta-llama/Llama-2-70b-hf"]
 DEFAULT_BATCH_SIZES = [1, 16, 32, 64, 128, 256, 512, 1024]
@@ -62,23 +63,23 @@ class BenchmarkTensors:
     a: torch.Tensor
 
     w_q: torch.Tensor
-    group_size: int | None
+    group_size: Optional[int]
     wtype: ScalarType
     w_g_s: torch.Tensor
-    w_g_zp: torch.Tensor | None
-    w_ch_s: torch.Tensor | None
-    w_tok_s: torch.Tensor | None
+    w_g_zp: Optional[torch.Tensor]
+    w_ch_s: Optional[torch.Tensor]
+    w_tok_s: Optional[torch.Tensor]
 
 
 @dataclass
 class TypeConfig:
     act_type: torch.dtype
     weight_type: ScalarType
-    output_type: torch.dtype | None
-    group_scale_type: torch.dtype | None
-    group_zero_type: torch.dtype | None
-    channel_scale_type: torch.dtype | None
-    token_scale_type: torch.dtype | None
+    output_type: Optional[torch.dtype]
+    group_scale_type: Optional[torch.dtype]
+    group_zero_type: Optional[torch.dtype]
+    channel_scale_type: Optional[torch.dtype]
+    token_scale_type: Optional[torch.dtype]
 
 
 def rand_data(shape, dtype=torch.float16, scale=1):
@@ -92,8 +93,8 @@ def quantize_and_pack(
     atype: torch.dtype,
     w: torch.Tensor,
     wtype: ScalarType,
-    stype: torch.dtype | None,
-    group_size: int | None,
+    stype: Optional[torch.dtype],
+    group_size: Optional[int],
     zero_points: bool = False,
 ):
     assert wtype.is_integer(), "TODO: support floating point weights"
@@ -112,7 +113,7 @@ def quantize_and_pack(
 
 
 def create_bench_tensors(
-    shape: tuple[int, int, int], types: TypeConfig, group_size: int | None
+    shape: tuple[int, int, int], types: TypeConfig, group_size: Optional[int]
 ) -> list[BenchmarkTensors]:
     m, n, k = shape
 
@@ -330,8 +331,8 @@ def bench_fns(label: str, sub_label: str, description: str, fns: list[Callable])
     return res
 
 
-_SWEEP_SCHEDULES_RESULTS: pd.DataFrame | None = None
-_SWEEP_SCHEDULES_RESULTS_CSV: str | None = None
+_SWEEP_SCHEDULES_RESULTS: Optional[pd.DataFrame] = None
+_SWEEP_SCHEDULES_RESULTS_CSV: Optional[str] = None
 
 
 def bench(

benchmarks/kernels/benchmark_marlin.py

@@ -44,7 +44,7 @@ from vllm.model_executor.layers.quantization.utils.quant_utils import (
     sort_weights,
 )
 from vllm.scalar_type import ScalarType, scalar_types
-from vllm.utils.argparse_utils import FlexibleArgumentParser
+from vllm.utils import FlexibleArgumentParser
 
 DEFAULT_MODELS = ["meta-llama/Llama-2-7b-hf/TP1"]
 DEFAULT_BATCH_SIZES = [1, 16, 32, 64, 128, 256, 512, 1024, 2048, 4096, 8192]

benchmarks/kernels/benchmark_moe.py

@@ -14,15 +14,11 @@ import ray
 import torch
 from ray.experimental.tqdm_ray import tqdm
 
-from vllm.model_executor.layers.fused_moe.config import (
-    FusedMoEQuantConfig,
-    _get_config_dtype_str,
-)
 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.argparse_utils import FlexibleArgumentParser
+from vllm.utils import FlexibleArgumentParser
 
 FP8_DTYPE = current_platform.fp8_dtype()
 
@@ -138,36 +134,43 @@ def benchmark_config(
     def run():
         from vllm.model_executor.layers.fused_moe import override_config
 
-        if use_fp8_w8a8:
-            quant_dtype = torch.float8_e4m3fn
-        elif use_int8_w8a16:
-            quant_dtype = torch.int8
-        else:
-            quant_dtype = None
-
-        quant_config = FusedMoEQuantConfig.make(
-            quant_dtype=quant_dtype,
-            w1_scale=w1_scale,
-            w2_scale=w2_scale,
-            a1_scale=a1_scale,
-            a2_scale=a2_scale,
-            block_shape=block_quant_shape,
-        )
-
         with override_config(config):
-            topk_weights, topk_ids, token_expert_indices = fused_topk(
-                x, input_gating, topk, renormalize=not use_deep_gemm
-            )
-            return fused_experts(
-                x,
-                w1,
-                w2,
-                topk_weights,
-                topk_ids,
-                inplace=True,
-                quant_config=quant_config,
-                allow_deep_gemm=use_deep_gemm,
-            )
+            if use_deep_gemm:
+                topk_weights, topk_ids, token_expert_indices = fused_topk(
+                    x, input_gating, topk, False
+                )
+                return fused_experts(
+                    x,
+                    w1,
+                    w2,
+                    topk_weights,
+                    topk_ids,
+                    inplace=True,
+                    use_fp8_w8a8=use_fp8_w8a8,
+                    w1_scale=w1_scale,
+                    w2_scale=w2_scale,
+                    a1_scale=a1_scale,
+                    a2_scale=a2_scale,
+                    block_shape=block_quant_shape,
+                    allow_deep_gemm=True,
+                )
+            else:
+                fused_moe(
+                    x,
+                    w1,
+                    w2,
+                    input_gating,
+                    topk,
+                    renormalize=True,
+                    inplace=True,
+                    use_fp8_w8a8=use_fp8_w8a8,
+                    use_int8_w8a16=use_int8_w8a16,
+                    w1_scale=w1_scale,
+                    w2_scale=w2_scale,
+                    a1_scale=a1_scale,
+                    a2_scale=a2_scale,
+                    block_shape=block_quant_shape,
+                )
 
     # JIT compilation & warmup
     run()
@@ -411,7 +414,7 @@ class BenchmarkWorker:
         use_deep_gemm: bool = False,
     ) -> tuple[dict[str, int], float]:
         current_platform.seed_everything(self.seed)
-        dtype_str = _get_config_dtype_str(
+        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
@@ -544,7 +547,7 @@ def save_configs(
     block_quant_shape: list[int],
     save_dir: str,
 ) -> None:
-    dtype_str = _get_config_dtype_str(
+    dtype_str = get_config_dtype_str(
         dtype, use_int8_w8a16=use_int8_w8a16, use_fp8_w8a8=use_fp8_w8a8
     )
 
@@ -557,7 +560,7 @@ def save_configs(
     filename = os.path.join(save_dir, filename)
     print(f"Writing best config to {filename}...")
     with open(filename, "w") as f:
-        json.dump({"triton_version": triton.__version__, **configs}, f, indent=4)
+        json.dump(configs, f, indent=4)
         f.write("\n")
 
 
@@ -579,42 +582,26 @@ def main(args: argparse.Namespace):
         E = config.ffn_config.moe_num_experts
         topk = config.ffn_config.moe_top_k
         intermediate_size = config.ffn_config.ffn_hidden_size
-        hidden_size = config.hidden_size
     elif config.architectures[0] == "JambaForCausalLM":
         E = config.num_experts
         topk = config.num_experts_per_tok
         intermediate_size = config.intermediate_size
-        hidden_size = config.hidden_size
     elif config.architectures[0] in (
-        "DeepseekV2ForCausalLM",
         "DeepseekV3ForCausalLM",
-        "DeepseekV32ForCausalLM",
+        "DeepseekV2ForCausalLM",
         "Glm4MoeForCausalLM",
     ):
         E = config.n_routed_experts
         topk = config.num_experts_per_tok
         intermediate_size = config.moe_intermediate_size
-        hidden_size = config.hidden_size
-    elif config.architectures[0] in (
-        "Qwen2MoeForCausalLM",
-        "Qwen3MoeForCausalLM",
-        "Qwen3NextForCausalLM",
-    ):
+    elif config.architectures[0] in ("Qwen2MoeForCausalLM", "Qwen3MoeForCausalLM"):
         E = config.num_experts
         topk = config.num_experts_per_tok
         intermediate_size = config.moe_intermediate_size
-        hidden_size = config.hidden_size
-    elif config.architectures[0] == "Qwen3VLMoeForConditionalGeneration":
-        text_config = config.get_text_config()
-        E = text_config.num_experts
-        topk = text_config.num_experts_per_tok
-        intermediate_size = text_config.moe_intermediate_size
-        hidden_size = text_config.hidden_size
     elif config.architectures[0] in ("HunYuanMoEV1ForCausalLM"):
         E = config.num_experts
         topk = config.moe_topk[0]
         intermediate_size = config.moe_intermediate_size[0]
-        hidden_size = config.hidden_size
     else:
         # Support for llama4
         config = config.get_text_config()
@@ -622,7 +609,6 @@ def main(args: argparse.Namespace):
         E = config.num_local_experts
         topk = config.num_experts_per_tok
         intermediate_size = config.intermediate_size
-        hidden_size = config.hidden_size
     enable_ep = bool(args.enable_expert_parallel)
     if enable_ep:
         ensure_divisibility(E, args.tp_size, "Number of experts")
@@ -631,7 +617,8 @@ def main(args: argparse.Namespace):
     else:
         ensure_divisibility(intermediate_size, args.tp_size, "intermediate_size")
         shard_intermediate_size = 2 * intermediate_size // args.tp_size
-    dtype = torch.float16 if current_platform.is_rocm() else config.dtype
+    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"
     block_quant_shape = get_weight_block_size_safety(config)
@@ -691,11 +678,7 @@ def main(args: argparse.Namespace):
         is_fp16 = not (use_fp8_w8a8 or use_int8_w8a16)
         search_space = get_configs_compute_bound(is_fp16, block_quant_shape)
         print(f"Start tuning over {len(search_space)} configurations...")
-        if use_deep_gemm:
-            raise ValueError(
-                "Tuning with --use-deep-gemm is not supported as it only tunes Triton "
-                "kernels. Please remove the flag."
-            )
+
         start = time.time()
         configs = _distribute(
             "tune",