fix: gptq marlin weight loading failure (#23066)

Signed-off-by: calvin chen <wen.chen@dynamia.ai>

.buildkite/nightly-benchmarks/README.md

@@ -7,7 +7,7 @@ This directory contains two sets of benchmark for vllm.
 - Performance benchmark: benchmark vllm's performance under various workload, for **developers** to gain clarity on whether their PR improves/degrades vllm's performance
 - Nightly benchmark: compare vllm's performance against alternatives (tgi, trt-llm and lmdeploy), for **the public** to know when to choose vllm.
 
-See [vLLM performance dashboard](https://perf.vllm.ai) for the latest performance benchmark results and [vLLM GitHub README](https://github.com/vllm-project/vllm/blob/main/README.md) for latest nightly benchmark results.
+See [vLLM performance dashboard](https://hud.pytorch.org/benchmark/llms?repoName=vllm-project%2Fvllm) for the latest performance benchmark results and [vLLM GitHub README](https://github.com/vllm-project/vllm/blob/main/README.md) for latest nightly benchmark results.
 
 ## Performance benchmark quick overview
 
@@ -28,6 +28,7 @@ See [vLLM performance dashboard](https://perf.vllm.ai) for the latest performanc
 ## Trigger the benchmark
 
 Performance benchmark will be triggered when:
+
 - A PR being merged into vllm.
 - Every commit for those PRs with `perf-benchmarks` label AND `ready` label.
 
@@ -38,6 +39,7 @@ bash .buildkite/nightly-benchmarks/scripts/run-performance-benchmarks.sh
 ```
 
 Runtime environment variables:
+
 - `ON_CPU`: set the value to '1' on Intel® Xeon® Processors. Default value is 0.
 - `SERVING_JSON`: JSON file to use for the serving tests. Default value is empty string (use default file).
 - `LATENCY_JSON`: JSON file to use for the latency tests. Default value is empty string (use default file).
@@ -46,12 +48,14 @@ Runtime environment variables:
 - `REMOTE_PORT`: Port for the remote vLLM service to benchmark. Default value is empty string.
 
 Nightly benchmark will be triggered when:
+
 - Every commit for those PRs with `perf-benchmarks` label and `nightly-benchmarks` label.
 
 ## Performance benchmark details
 
 See [performance-benchmarks-descriptions.md](performance-benchmarks-descriptions.md) for detailed descriptions, and use `tests/latency-tests.json`, `tests/throughput-tests.json`, `tests/serving-tests.json` to configure the test cases.
 > NOTE: For Intel® Xeon® Processors, use `tests/latency-tests-cpu.json`, `tests/throughput-tests-cpu.json`, `tests/serving-tests-cpu.json` instead.
+>
 ### Latency test
 
 Here is an example of one test inside `latency-tests.json`:
@@ -100,7 +104,6 @@ We test the throughput by using `vllm bench serve` with request rate = inf to co
             "tensor_parallel_size": 1,
             "swap_space": 16,
             "disable_log_stats": "",
-            "disable_log_requests": "",
             "load_format": "dummy"
         },
         "client_parameters": {
@@ -135,27 +138,20 @@ The raw benchmarking results (in the format of json files) are in the `Artifacts
 
 The `compare-json-results.py` helps to compare benchmark results JSON files converted using `convert-results-json-to-markdown.py`.
 When run, benchmark script generates results under `benchmark/results` folder, along with the `benchmark_results.md` and `benchmark_results.json`.
-`compare-json-results.py` compares two `benchmark_results.json` files and provides performance ratio e.g. for Output Tput, Median TTFT and Median TPOT.
+`compare-json-results.py` compares two `benchmark_results.json` files and provides performance ratio e.g. for Output Tput, Median TTFT and Median TPOT.  
+If only one benchmark_results.json is passed, `compare-json-results.py` compares different TP and PP configurations in the benchmark_results.json instead.
 
-Here is an example using the script to compare result_a and result_b without detail test name.
-`python3 compare-json-results.py -f results_a/benchmark_results.json -f results_b/benchmark_results.json --ignore_test_name`
-
-|    | results_a/benchmark_results.json | results_b/benchmark_results.json | perf_ratio        |
-|----|----------------------------------------|----------------------------------------|----------|
-| 0  | 142.633982                             | 156.526018                             | 1.097396 |
-| 1  | 241.620334                             | 294.018783                             | 1.216863 |
-| 2  | 218.298905                             | 262.664916                             | 1.203235 |
-| 3  | 242.743860                             | 299.816190                             | 1.235113 |
-
-Here is an example using the script to compare result_a and result_b with detail test name.
+Here is an example using the script to compare result_a and result_b with Model, Dataset name, input/output lenght, max concurrency and qps.
 `python3 compare-json-results.py -f results_a/benchmark_results.json -f results_b/benchmark_results.json`
-|   | results_a/benchmark_results.json_name | results_a/benchmark_results.json | results_b/benchmark_results.json_name | results_b/benchmark_results.json | perf_ratio        |
-|---|---------------------------------------------|----------------------------------------|---------------------------------------------|----------------------------------------|----------|
-| 0 | serving_llama8B_tp1_sharegpt_qps_1          | 142.633982                             | serving_llama8B_tp1_sharegpt_qps_1          | 156.526018                             | 1.097396 |
-| 1 | serving_llama8B_tp1_sharegpt_qps_16         | 241.620334                             | serving_llama8B_tp1_sharegpt_qps_16         | 294.018783                             | 1.216863 |
-| 2 | serving_llama8B_tp1_sharegpt_qps_4          | 218.298905                             | serving_llama8B_tp1_sharegpt_qps_4          | 262.664916                             | 1.203235 |
-| 3 | serving_llama8B_tp1_sharegpt_qps_inf        | 242.743860                             | serving_llama8B_tp1_sharegpt_qps_inf        | 299.816190                             | 1.235113 |
-| 4 | serving_llama8B_tp2_random_1024_128_qps_1   | 96.613390                              | serving_llama8B_tp4_random_1024_128_qps_1   | 108.404853                             | 1.122048 |
+
+|   | Model | Dataset Name | Input Len | Output Len | # of max concurrency | qps  | results_a/benchmark_results.json | results_b/benchmark_results.json | perf_ratio        |
+|----|---------------------------------------|--------|-----|-----|------|-----|-----------|----------|----------|
+| 0  | meta-llama/Meta-Llama-3.1-8B-Instruct | random | 128 | 128 | 1000 | 1 | 142.633982                             | 156.526018                             | 1.097396 |
+| 1  | meta-llama/Meta-Llama-3.1-8B-Instruct | random | 128 | 128 | 1000 | inf| 241.620334                             | 294.018783                             | 1.216863 |
+
+A comparison diagram will be generated below the table.
+Here is an example to compare between 96c/results_gnr_96c_091_tp2pp3 and 128c/results_gnr_128c_091_tp2pp3
+<img width="1886" height="828" alt="image" src="https://github.com/user-attachments/assets/c02a43ef-25d0-4fd6-90e5-2169a28682dd" />
 
 ## Nightly test details
 
@@ -164,9 +160,9 @@ See [nightly-descriptions.md](nightly-descriptions.md) for the detailed descript
 ### Workflow
 
 - The [nightly-pipeline.yaml](nightly-pipeline.yaml) specifies the docker containers for different LLM serving engines.
-- Inside each container, we run [run-nightly-suite.sh](run-nightly-suite.sh), which will probe the serving engine of the current container.
-- The `run-nightly-suite.sh` will redirect the request to `tests/run-[llm serving engine name]-nightly.sh`, which parses the workload described in [nightly-tests.json](tests/nightly-tests.json) and performs the benchmark.
-- At last, we run [scripts/plot-nightly-results.py](scripts/plot-nightly-results.py) to collect and plot the final benchmarking results, and update the results to buildkite.
+- Inside each container, we run [scripts/run-nightly-benchmarks.sh](scripts/run-nightly-benchmarks.sh), which will probe the serving engine of the current container.
+- The `scripts/run-nightly-benchmarks.sh` will parse the workload described in [nightly-tests.json](tests/nightly-tests.json) and launch the right benchmark for the specified serving engine via `scripts/launch-server.sh`.
+- At last, we run [scripts/summary-nightly-results.py](scripts/summary-nightly-results.py) to collect and plot the final benchmarking results, and update the results to buildkite.
 
 ### Nightly tests
 
@@ -176,6 +172,6 @@ In [nightly-tests.json](tests/nightly-tests.json), we include the command line a
 
 The docker containers for benchmarking are specified in `nightly-pipeline.yaml`.
 
-WARNING: the docker versions are HARD-CODED and SHOULD BE ALIGNED WITH `nightly-descriptions.md`. The docker versions need to be hard-coded as there are several version-specific bug fixes inside `tests/run-[llm serving engine name]-nightly.sh`.
+WARNING: the docker versions are HARD-CODED and SHOULD BE ALIGNED WITH `nightly-descriptions.md`. The docker versions need to be hard-coded as there are several version-specific bug fixes inside `scripts/run-nightly-benchmarks.sh` and `scripts/launch-server.sh`.
 
 WARNING: populating `trt-llm` to latest version is not easy, as it requires updating several protobuf files in [tensorrt-demo](https://github.com/neuralmagic/tensorrt-demo.git).

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

@@ -1,3 +1,4 @@
+# Nightly benchmark annotation
 
 ## Description
 
@@ -13,15 +14,15 @@ Please download the visualization scripts in the post
 
 - Find the docker we use in `benchmarking pipeline`
 - Deploy the docker, and inside the docker:
-  - Download `nightly-benchmarks.zip`.
-  - In the same folder, run the following code:
+    - Download `nightly-benchmarks.zip`.
+    - In the same folder, run the following code:
 
-  ```bash
-  export HF_TOKEN=<your HF token>
-  apt update
-  apt install -y git
-  unzip nightly-benchmarks.zip
-  VLLM_SOURCE_CODE_LOC=./ bash .buildkite/nightly-benchmarks/scripts/run-nightly-benchmarks.sh
-  ```
+    ```bash
+    export HF_TOKEN=<your HF token>
+    apt update
+    apt install -y git
+    unzip nightly-benchmarks.zip
+    VLLM_SOURCE_CODE_LOC=./ bash .buildkite/nightly-benchmarks/scripts/run-nightly-benchmarks.sh
+    ```
 
 And the results will be inside `./benchmarks/results`.

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

@@ -13,25 +13,25 @@ Latest reproduction guilde: [github issue link](https://github.com/vllm-project/
 ## Setup
 
 - Docker images:
-  - vLLM: `vllm/vllm-openai:v0.6.2`
-  - SGLang: `lmsysorg/sglang:v0.3.2-cu121`
-  - LMDeploy: `openmmlab/lmdeploy:v0.6.1-cu12`
-  - TensorRT-LLM: `nvcr.io/nvidia/tritonserver:24.07-trtllm-python-py3`
-    - *NOTE: we uses r24.07 as the current implementation only works for this version. We are going to bump this up.*
-  - Check [nightly-pipeline.yaml](nightly-pipeline.yaml) for the concrete docker images, specs and commands we use for the benchmark.
+    - vLLM: `vllm/vllm-openai:v0.6.2`
+    - SGLang: `lmsysorg/sglang:v0.3.2-cu121`
+    - LMDeploy: `openmmlab/lmdeploy:v0.6.1-cu12`
+    - TensorRT-LLM: `nvcr.io/nvidia/tritonserver:24.07-trtllm-python-py3`
+        - *NOTE: we uses r24.07 as the current implementation only works for this version. We are going to bump this up.*
+    - Check [nightly-pipeline.yaml](nightly-pipeline.yaml) for the concrete docker images, specs and commands we use for the benchmark.
 - Hardware
-  - 8x Nvidia A100 GPUs
+    - 8x Nvidia A100 GPUs
 - Workload:
-  - Dataset
-    - ShareGPT dataset
-    - Prefill-heavy dataset (in average 462 input tokens, 16 tokens as output)
-    - Decode-heavy dataset (in average 462 input tokens, 256 output tokens)
-    - Check [nightly-tests.json](tests/nightly-tests.json) for the concrete configuration of datasets we use.
-  - Models: llama-3 8B, llama-3 70B.
-    - We do not use llama 3.1 as it is incompatible with trt-llm r24.07. ([issue](https://github.com/NVIDIA/TensorRT-LLM/issues/2105)).
-  - Average QPS (query per second): 2, 4, 8, 16, 32 and inf.
-    - Queries are randomly sampled, and arrival patterns are determined via Poisson process, but all with fixed random seed.
-  - Evaluation metrics: Throughput (higher the better), TTFT (time to the first token, lower the better), ITL (inter-token latency, lower the better).
+    - Dataset
+        - ShareGPT dataset
+        - Prefill-heavy dataset (in average 462 input tokens, 16 tokens as output)
+        - Decode-heavy dataset (in average 462 input tokens, 256 output tokens)
+        - Check [nightly-tests.json](tests/nightly-tests.json) for the concrete configuration of datasets we use.
+    - Models: llama-3 8B, llama-3 70B.
+        - We do not use llama 3.1 as it is incompatible with trt-llm r24.07. ([issue](https://github.com/NVIDIA/TensorRT-LLM/issues/2105)).
+    - Average QPS (query per second): 2, 4, 8, 16, 32 and inf.
+        - Queries are randomly sampled, and arrival patterns are determined via Poisson process, but all with fixed random seed.
+    - Evaluation metrics: Throughput (higher the better), TTFT (time to the first token, lower the better), ITL (inter-token latency, lower the better).
 
 ## Known issues
 

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

@@ -1,3 +1,4 @@
+# Performance benchmarks descriptions
 
 ## Latency tests
 

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

@@ -1,24 +1,38 @@
 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 import argparse
+import json
+import os
 
 import pandas as pd
 
 
 def compare_data_columns(
-    files, name_column, data_column, drop_column, ignore_test_name=False
+    files, name_column, data_column, info_cols, drop_column, debug=False
 ):
     print("\ncompare_data_column: " + data_column)
     frames = []
+    raw_data_cols = []
     compare_frames = []
     for file in files:
         data_df = pd.read_json(file)
         serving_df = data_df.dropna(subset=[drop_column], ignore_index=True)
-        if ignore_test_name is False:
+        # Show all info columns in the first couple columns
+        if not frames:
+            for col in info_cols:
+                if col not in serving_df.columns:
+                    print(f"Skipping missing column: {col}")
+                    continue
+                frames.append(serving_df[col])
+        # only show test name under debug mode
+        if debug is True:
             serving_df = serving_df.rename(columns={name_column: file + "_name"})
             frames.append(serving_df[file + "_name"])
+
+        file = "/".join(file.split("/")[:-1])
         serving_df = serving_df.rename(columns={data_column: file})
         frames.append(serving_df[file])
+        raw_data_cols.append(file)
         compare_frames.append(serving_df[file])
         if len(compare_frames) >= 2:
             # Compare numbers among two files
@@ -27,7 +41,68 @@ def compare_data_columns(
             compare_frames.pop(1)
 
     concat_df = pd.concat(frames, axis=1)
-    return concat_df
+    print(raw_data_cols)
+    return concat_df, raw_data_cols
+
+
+def split_json_by_tp_pp(
+    input_file: str = "benchmark_results.json", output_root: str = "."
+) -> list[str]:
+    """
+    Split a benchmark JSON into separate folders by (TP Size, PP Size).
+
+    Creates: <output_root>/tp{TP}_pp{PP}/benchmark_results.json
+    Returns: list of file paths written.
+    """
+    # Load JSON data into DataFrame
+    with open(input_file, encoding="utf-8") as f:
+        data = json.load(f)
+
+    # If the JSON is a dict with a list under common keys, use that list
+    if isinstance(data, dict):
+        for key in ("results", "serving_results", "benchmarks", "data"):
+            if isinstance(data.get(key), list):
+                data = data[key]
+                break
+
+    df = pd.DataFrame(data)
+
+    # Handle alias column names
+    rename_map = {
+        "tp_size": "TP Size",
+        "tensor_parallel_size": "TP Size",
+        "pp_size": "PP Size",
+        "pipeline_parallel_size": "PP Size",
+    }
+    df.rename(
+        columns={k: v for k, v in rename_map.items() if k in df.columns}, inplace=True
+    )
+
+    # Ensure TP/PP columns exist (default to 1 if missing)
+    if "TP Size" not in df.columns:
+        df["TP Size"] = 1
+    if "PP Size" not in df.columns:
+        df["PP Size"] = 1
+
+    # make sure TP/PP are numeric ints with no NaN
+    df["TP Size"] = (
+        pd.to_numeric(df.get("TP Size", 1), errors="coerce").fillna(1).astype(int)
+    )
+    df["PP Size"] = (
+        pd.to_numeric(df.get("PP Size", 1), errors="coerce").fillna(1).astype(int)
+    )
+
+    # Split into separate folders
+    saved_paths: list[str] = []
+    for (tp, pp), group_df in df.groupby(["TP Size", "PP Size"], dropna=False):
+        folder_name = os.path.join(output_root, f"tp{int(tp)}_pp{int(pp)}")
+        os.makedirs(folder_name, exist_ok=True)
+        filepath = os.path.join(folder_name, "benchmark_results.json")
+        group_df.to_json(filepath, orient="records", indent=2, force_ascii=False)
+        print(f"Saved: {filepath}")
+        saved_paths.append(filepath)
+
+    return saved_paths
 
 
 if __name__ == "__main__":
@@ -36,31 +111,105 @@ if __name__ == "__main__":
         "-f", "--file", action="append", type=str, help="input file name"
     )
     parser.add_argument(
-        "--ignore_test_name", action="store_true", help="ignore_test_name or not"
+        "--debug", action="store_true", help="show all information for debugging"
+    )
+    parser.add_argument(
+        "--plot",
+        action=argparse.BooleanOptionalAction,
+        default=True,
+        help="plot perf diagrams or not --no-plot --plot",
+    )
+    parser.add_argument(
+        "-x",
+        "--xaxis",
+        type=str,
+        default="# of max concurrency.",
+        help="column name to use as X Axis in comparision graph",
     )
     args = parser.parse_args()
-    files = args.file
-    print("comparing : " + ", ".join(files))
 
     drop_column = "P99"
     name_column = "Test name"
+    info_cols = [
+        "Model",
+        "Dataset Name",
+        "Input Len",
+        "Output Len",
+        "TP Size",
+        "PP Size",
+        "# of max concurrency.",
+        "qps",
+    ]
     data_cols_to_compare = ["Output Tput (tok/s)", "Median TTFT (ms)", "Median"]
     html_msgs_for_data_cols = [
         "Compare Output Tokens /n",
         "Median TTFT /n",
         "Median TPOT /n",
     ]
-    ignore_test_name = args.ignore_test_name
+
+    if len(args.file) == 1:
+        files = split_json_by_tp_pp(args.file[0], output_root="splits")
+        info_cols = [c for c in info_cols if c not in ("TP Size", "PP Size")]
+    else:
+        files = args.file
+    print("comparing : " + ", ".join(files))
+    debug = args.debug
+    plot = args.plot
+    # For Plot feature, assign y axis from one of info_cols
+    y_axis_index = info_cols.index(args.xaxis) if args.xaxis in info_cols else 6
     with open("perf_comparison.html", "w") as text_file:
         for i in range(len(data_cols_to_compare)):
-            output_df = compare_data_columns(
+            output_df, raw_data_cols = compare_data_columns(
                 files,
                 name_column,
                 data_cols_to_compare[i],
+                info_cols,
                 drop_column,
-                ignore_test_name=ignore_test_name,
+                debug=debug,
             )
-            print(output_df)
-            html = output_df.to_html()
-            text_file.write(html_msgs_for_data_cols[i])
-            text_file.write(html)
+
+            # For Plot feature, insert y axis from one of info_cols
+            raw_data_cols.insert(0, info_cols[y_axis_index])
+
+            filtered_info_cols = info_cols[:-2]
+            existing_group_cols = [
+                c for c in filtered_info_cols if c in output_df.columns
+            ]
+            if not existing_group_cols:
+                raise ValueError(
+                    f"No valid group-by columns  "
+                    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_groups = output_df_sorted.groupby(existing_group_cols, dropna=False)
+            for name, group in output_groups:
+                html = group.to_html()
+                text_file.write(html_msgs_for_data_cols[i])
+                text_file.write(html)
+
+                if plot is True:
+                    import pandas as pd
+                    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,
+                    )
+                    # 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

@@ -1,17 +1,19 @@
 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
+import argparse
 import json
 import os
+import shlex
 from importlib import util
 from pathlib import Path
+from typing import Any
 
 import pandas as pd
 import psutil
+import regex as re
 from tabulate import tabulate
 
-results_folder = Path("results/")
-
 # latency results and the keys that will be printed into markdown
 latency_results = []
 latency_column_mapping = {
@@ -42,13 +44,22 @@ throughput_results_column_mapping = {
 serving_results = []
 serving_column_mapping = {
     "test_name": "Test name",
+    "model_id": "Model",
+    "dataset_name": "Dataset Name",
+    "input_len": "Input Len",
+    "output_len": "Output Len",
+    "tp_size": "TP Size",
+    "pp_size": "PP Size",
+    "dtype": "dtype",
     "gpu_type": "GPU",
     "completed": "# of req.",
+    "qps": "qps",
+    "max_concurrency": "# of max concurrency.",
     "request_throughput": "Tput (req/s)",
     "total_token_throughput": "Total Token Tput (tok/s)",
     "output_throughput": "Output Tput (tok/s)",
-    "total_input_tokens": "Total input tokens",
-    "total_output_tokens": "Total output tokens",
+    # "total_input_tokens": "Total input tokens",
+    # "total_output_tokens": "Total output tokens",
     "mean_ttft_ms": "Mean TTFT (ms)",
     "median_ttft_ms": "Median TTFT (ms)",
     "p99_ttft_ms": "P99 TTFT (ms)",
@@ -93,7 +104,104 @@ def get_size_with_unit(bytes, suffix="B"):
         bytes /= factor
 
 
+def _coerce(val: str) -> Any:
+    """Best-effort type coercion from string to Python types."""
+    low = val.lower()
+    if low == "null":
+        return None
+    if low == "true":
+        return True
+    if low == "false":
+        return False
+    # integers
+    if re.fullmatch(r"[+-]?\d+", val):
+        try:
+            return int(val)
+        except ValueError:
+            pass
+    # floats (keep 'inf'/'-inf'/'nan' as strings)
+    if re.fullmatch(r"[+-]?\d*\.\d+", val):
+        try:
+            return float(val)
+        except ValueError:
+            pass
+    return val
+
+
+def parse_client_command(cmd: str) -> dict[str, Any]:
+    """Parse the client_command shell string into {executable, script, args}."""
+    toks = shlex.split(cmd)
+    if len(toks) < 2:
+        raise ValueError("client_command must include an executable and a script")
+    executable, script = toks[0], toks[1]
+    args: dict[str, Any] = {}
+
+    i = 2
+    while i < len(toks):
+        t = toks[i]
+        if t.startswith("--"):
+            # --key=value or --key (value) or boolean flag
+            if "=" in t:
+                key, val = t.split("=", 1)
+                if key == "--metadata":
+                    md = {}
+                    if val:
+                        if "=" in val:
+                            k, v = val.split("=", 1)
+                            md[k] = _coerce(v)
+                        else:
+                            md[val] = True
+                    args[key] = md
+                else:
+                    args[key] = _coerce(val)
+                i += 1
+                continue
+
+            key = t
+
+            # Special: consume metadata k=v pairs until next --flag
+            if key == "--metadata":
+                i += 1
+                md = {}
+                while i < len(toks) and not toks[i].startswith("--"):
+                    pair = toks[i]
+                    if "=" in pair:
+                        k, v = pair.split("=", 1)
+                        md[k] = _coerce(v)
+                    else:
+                        md[pair] = True
+                    i += 1
+                args[key] = md
+                continue
+
+            # Standard: check if next token is a value (not a flag)
+            if i + 1 < len(toks) and not toks[i + 1].startswith("--"):
+                args[key] = _coerce(toks[i + 1])
+                i += 2
+            else:
+                # lone flag -> True
+                args[key] = True
+                i += 1
+        else:
+            # unexpected positional; skip
+            i += 1
+
+    return {"executable": executable, "script": script, "args": args}
+
+
 if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "-r",
+        "--result",
+        type=str,
+        default="results",
+        help="Folder name for benchmark output results.",
+    )
+    args = parser.parse_args()
+    results_folder = Path(args.result)
+    if not results_folder.exists():
+        raise FileNotFoundError(f"results folder does not exist: {results_folder}")
     # collect results
     for test_file in results_folder.glob("*.json"):
         with open(test_file) as f:
@@ -101,7 +209,6 @@ if __name__ == "__main__":
 
         if "serving" in str(test_file):
             # this result is generated via `vllm bench serve` command
-
             # attach the benchmarking command to raw_result
             try:
                 with open(test_file.with_suffix(".commands")) as f:
@@ -109,12 +216,44 @@ if __name__ == "__main__":
             except OSError as e:
                 print(e)
                 continue
+            # Parse Server Command Arg
+            out: dict[str, Any] = {
+                "server_command": parse_client_command(command["server_command"])
+            }
+            parse_args = [
+                "--tensor-parallel-size",
+                "--pipeline-parallel-size",
+                "--dtype",
+            ]
+            col_mapping = ["tp_size", "pp_size", "dtype"]
+            for index, arg in enumerate(parse_args):
+                if arg in out["server_command"]["args"]:
+                    raw_result.update(
+                        {col_mapping[index]: out["server_command"]["args"][arg]}
+                    )
 
+            # Parse Client Command Arg
+            out: dict[str, Any] = {
+                "client_command": parse_client_command(command["client_command"])
+            }
+            parse_args = [
+                "--dataset-name",
+                "--random-input-len",
+                "--random-output-len",
+                "--request-rate",
+            ]
+            col_mapping = ["dataset_name", "input_len", "output_len", "qps"]
+
+            for index, arg in enumerate(parse_args):
+                if arg in out["client_command"]["args"]:
+                    raw_result.update(
+                        {col_mapping[index]: out["client_command"]["args"][arg]}
+                    )
+            # Add Server, Client command
             raw_result.update(command)
 
             # update the test name of this result
             raw_result.update({"test_name": test_file.stem})
-
             # add the result to raw_result
             serving_results.append(raw_result)
             continue
@@ -204,7 +343,10 @@ if __name__ == "__main__":
             columns=latency_column_mapping
         )
     if not serving_results.empty:
-        serving_results = serving_results[list(serving_column_mapping.keys())].rename(
+        valid_columns = [
+            col for col in serving_column_mapping if col in serving_results.columns
+        ]
+        serving_results = serving_results[valid_columns].rename(
             columns=serving_column_mapping
         )
     if not throughput_results.empty:
@@ -244,7 +386,9 @@ if __name__ == "__main__":
     )
 
     # document the result
-    with open(results_folder / "benchmark_results.md", "w") as f:
+    md_file = "benchmark_results.md"
+    json_file = "benchmark_results.json"
+    with open(results_folder / md_file, "w") as f:
         results = read_markdown(
             "../.buildkite/nightly-benchmarks/"
             + "performance-benchmarks-descriptions.md"
@@ -259,7 +403,7 @@ if __name__ == "__main__":
         f.write(results)
 
     # document benchmarking results in json
-    with open(results_folder / "benchmark_results.json", "w") as f:
+    with open(results_folder / json_file, "w") as f:
         results = (
             latency_results.to_dict(orient="records")
             + throughput_results.to_dict(orient="records")

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

@@ -33,7 +33,7 @@ check_gpus() {
 
 check_cpus() {
   # check the number of CPUs and NUMA Node and GPU type.
-  declare -g numa_count=$(python3 -c  "from numa import info;numa_size = info.get_num_configured_nodes(); print(numa_size)")
+  declare -g numa_count=$(lscpu | grep "NUMA node(s):" | awk '{print $3}')
   if [[ $numa_count -gt 0 ]]; then
     echo "NUMA found."
     echo $numa_count
@@ -194,9 +194,11 @@ run_latency_tests() {
 
     # check if there is enough GPU to run the test
     tp=$(echo "$latency_params" | jq -r '.tensor_parallel_size')
-    if [ "$ON_CPU" == "1" ];then
-      if [[ $numa_count -lt $tp ]]; then
-        echo "Required tensor-parallel-size $tp but only $numa_count NUMA nodes found. Skip testcase $test_name."
+    if [ "$ON_CPU" == "1" ]; then
+      pp=$(echo "$latency_params" | jq -r '.pipeline_parallel_size')
+      world_size=$(($tp*$pp))
+      if [[ $numa_count -lt $world_size  && -z "${REMOTE_HOST}" ]]; then
+        echo "Required world-size $world_size but only $numa_count NUMA nodes found. Skip testcase $test_name."
         continue
       fi
     else
@@ -261,9 +263,11 @@ run_throughput_tests() {
 
     # check if there is enough GPU to run the test
     tp=$(echo "$throughput_params" | jq -r '.tensor_parallel_size')
-    if [ "$ON_CPU" == "1" ];then
-      if [[ $numa_count -lt $tp ]]; then
-        echo "Required tensor-parallel-size $tp but only $numa_count NUMA nodes found. Skip testcase $test_name."
+    if [ "$ON_CPU" == "1" ]; then
+      pp=$(echo "$throughput_params" | jq -r '.pipeline_parallel_size')
+      world_size=$(($tp*$pp))
+      if [[ $numa_count -lt $world_size  && -z "${REMOTE_HOST}" ]]; then
+        echo "Required world-size $world_size but only $numa_count NUMA nodes found. Skip testcase $test_name."
         continue
       fi
     else
@@ -329,12 +333,21 @@ run_serving_tests() {
     qps_list=$(echo "$params" | jq -r '.qps_list')
     qps_list=$(echo "$qps_list" | jq -r '.[] | @sh')
     echo "Running over qps list $qps_list"
+    max_concurrency_list=$(echo "$params" | jq -r '.max_concurrency_list')
+    if [[ -z "$max_concurrency_list" || "$max_concurrency_list" == "null" ]]; then
+        num_prompts=$(echo "$client_params" | jq -r '.num_prompts')
+        max_concurrency_list="[$num_prompts]"
+    fi
+    max_concurrency_list=$(echo "$max_concurrency_list" | jq -r '.[] | @sh')
+    echo "Running over max concurrency list $max_concurrency_list"
 
     # check if there is enough resources to run the test
     tp=$(echo "$server_params" | jq -r '.tensor_parallel_size')
-    if [ "$ON_CPU" == "1" ];then
-      if [[ $numa_count -lt $tp ]]; then
-        echo "Required tensor-parallel-size $tp but only $numa_count NUMA nodes found. Skip testcase $test_name."
+    if [ "$ON_CPU" == "1" ]; then
+      pp=$(echo "$server_params" | jq -r '.pipeline_parallel_size')
+      world_size=$(($tp*$pp))
+      if [[ $numa_count -lt $world_size  && -z "${REMOTE_HOST}" ]]; then
+        echo "Required world-size $world_size but only $numa_count NUMA nodes found. Skip testcase $test_name."
         continue
       fi
     else
@@ -390,35 +403,39 @@ run_serving_tests() {
         echo "now qps is $qps"
       fi
 
-      new_test_name=$test_name"_qps_"$qps
+      # iterate over different max_concurrency
+      for max_concurrency in $max_concurrency_list; do
+        new_test_name=$test_name"_qps_"$qps"_concurrency_"$max_concurrency
+        echo " new test name $new_test_name"
+        # pass the tensor parallel size to the client so that it can be displayed
+        # on the benchmark dashboard
+        client_command="vllm bench serve \
+          --save-result \
+          --result-dir $RESULTS_FOLDER \
+          --result-filename ${new_test_name}.json \
+          --request-rate $qps \
+          --max-concurrency $max_concurrency \
+          --metadata "tensor_parallel_size=$tp" \
+          $client_args $client_remote_args "
 
-      # pass the tensor parallel size to the client so that it can be displayed
-      # on the benchmark dashboard
-      client_command="vllm bench serve \
-        --save-result \
-        --result-dir $RESULTS_FOLDER \
-        --result-filename ${new_test_name}.json \
-        --request-rate $qps \
-        --metadata "tensor_parallel_size=$tp" \
-        $client_args $client_remote_args "
+        echo "Running test case $test_name with qps $qps"
+        echo "Client command: $client_command"
 
-      echo "Running test case $test_name with qps $qps"
-      echo "Client command: $client_command"
+        bash -c "$client_command"
 
-      bash -c "$client_command"
-
-      # record the benchmarking commands
-      jq_output=$(jq -n \
-        --arg server "$server_command" \
-        --arg client "$client_command" \
-        --arg gpu "$gpu_type" \
-        '{
-          server_command: $server,
-          client_command: $client,
-          gpu_type: $gpu
-        }')
-      echo "$jq_output" >"$RESULTS_FOLDER/${new_test_name}.commands"
+        # record the benchmarking commands
+        jq_output=$(jq -n \
+          --arg server "$server_command" \
+          --arg client "$client_command" \
+          --arg gpu "$gpu_type" \
+          '{
+            server_command: $server,
+            client_command: $client,
+            gpu_type: $gpu
+          }')
+        echo "$jq_output" >"$RESULTS_FOLDER/${new_test_name}.commands"
 
+      done
     done
 
     # clean up

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

@@ -11,9 +11,7 @@
         },
         "vllm_server_parameters": {
             "disable_log_stats": "",
-            "disable_log_requests": "",
             "gpu_memory_utilization": 0.9,
-            "num_scheduler_steps": 10,
             "max_num_seqs": 512,
             "dtype": "bfloat16"
         },

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

@@ -6,7 +6,7 @@
 	    "VLLM_CPU_KVCACHE_SPACE": 40
         },
         "parameters": {
-            "model": "meta-llama/Meta-Llama-3.1-8B-Instruct",
+            "model": "meta-llama/Llama-3.1-8B-Instruct",
             "tensor_parallel_size": 1,
             "load_format": "dummy",
             "num_iters_warmup": 5,
@@ -20,7 +20,7 @@
 	    "VLLM_CPU_KVCACHE_SPACE": 40
         },
         "parameters": {
-            "model": "meta-llama/Meta-Llama-3.1-8B-Instruct",
+            "model": "meta-llama/Llama-3.1-8B-Instruct",
             "tensor_parallel_size": 4,
             "load_format": "dummy",
             "num_iters_warmup": 5,

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

@@ -35,9 +35,7 @@
         }, 
         "vllm_server_parameters": {
             "disable_log_stats": "",
-            "disable_log_requests": "",
             "gpu_memory_utilization": 0.9,
-            "num_scheduler_steps": 10,
             "max_num_seqs": 512,
             "dtype": "bfloat16"
         },
@@ -90,9 +88,7 @@
         }, 
         "vllm_server_parameters": {
             "disable_log_stats": "",
-            "disable_log_requests": "",
             "gpu_memory_utilization": 0.9,
-            "num_scheduler_steps": 10,
             "max_num_seqs": 512,
             "dtype": "bfloat16"
         },
@@ -145,9 +141,7 @@
         }, 
         "vllm_server_parameters": {
             "disable_log_stats": "",
-            "disable_log_requests": "",
             "gpu_memory_utilization": 0.9,
-            "num_scheduler_steps": 10,
             "max_num_seqs": 512,
             "dtype": "bfloat16"
         },
@@ -197,9 +191,7 @@
         }, 
         "vllm_server_parameters": {
             "disable_log_stats": "",
-            "disable_log_requests": "",
             "gpu_memory_utilization": 0.9,
-            "num_scheduler_steps": 10,
             "max_num_seqs": 512,
             "dtype": "bfloat16"
         },
@@ -251,9 +243,7 @@
         }, 
         "vllm_server_parameters": {
             "disable_log_stats": "",
-            "disable_log_requests": "",
             "gpu_memory_utilization": 0.9,
-            "num_scheduler_steps": 10,
             "max_num_seqs": 512,
             "dtype": "bfloat16"
         },
@@ -305,9 +295,7 @@
         }, 
         "vllm_server_parameters": {
             "disable_log_stats": "",
-            "disable_log_requests": "",
             "gpu_memory_utilization": 0.9,
-            "num_scheduler_steps": 10,
             "max_num_seqs": 512,
             "dtype": "bfloat16"
         },

.buildkite/nightly-benchmarks/tests/serving-tests-cpu-snc2.json

@@ -0,0 +1,202 @@
+[
+    {
+        "test_name": "serving_llama8B_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": "meta-llama/Llama-3.1-8B-Instruct",
+            "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": "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_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_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_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": "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": 128,
+	    "ignore-eos": "",
+            "num_prompts": 1000
+        }
+    },
+    {
+        "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": {
+            "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_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,
+            "num_prompts": 1000
+        }
+    }
+]

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

@@ -0,0 +1,205 @@
+[
+    {
+        "test_name": "serving_llama8B_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": "meta-llama/Llama-3.1-8B-Instruct",
+            "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": "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_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": "meta-llama/Llama-3.1-8B-Instruct",
+            "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": "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_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": "meta-llama/Llama-3.1-8B-Instruct",
+            "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": "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_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": "meta-llama/Llama-3.1-8B-Instruct",
+            "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": "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_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": "meta-llama/Llama-3.1-8B-Instruct",
+            "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": "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_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": "meta-llama/Llama-3.1-8B-Instruct",
+            "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": "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
+        }
+    }
+]

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

@@ -2,104 +2,112 @@
     {
         "test_name": "serving_llama8B_tp1_sharegpt",
         "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/Meta-Llama-3.1-8B-Instruct",
+            "model": "meta-llama/Llama-3.1-8B-Instruct",
             "tensor_parallel_size": 1,
 	    "dtype": "bfloat16",
 	    "distributed_executor_backend": "mp",
 	    "block_size": 128,
 	    "trust_remote_code": "",
             "disable_log_stats": "",
-            "disable_log_requests": "",
 	    "enforce_eager": "",
+	    "max_num_batched_tokens": 2048,
+	    "max_num_seqs": 256,
             "load_format": "dummy"
         },
         "client_parameters": {
-            "model": "meta-llama/Meta-Llama-3.1-8B-Instruct",
+            "model": "meta-llama/Llama-3.1-8B-Instruct",
             "backend": "vllm",
             "dataset_name": "sharegpt",
             "dataset_path": "./ShareGPT_V3_unfiltered_cleaned_split.json",
-	    "max_concurrency": 60,
             "num_prompts": 200
         }
     },
     {
         "test_name": "serving_llama8B_tp2_sharegpt",
         "qps_list": [1, 4, 16, "inf"],
+        "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/Meta-Llama-3.1-8B-Instruct",
+            "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": "",
-            "disable_log_requests": "",
 	    "enforce_eager": "",
+	    "max_num_batched_tokens": 2048,
+	    "max_num_seqs": 256,
             "load_format": "dummy"
         },
         "client_parameters": {
-            "model": "meta-llama/Meta-Llama-3.1-8B-Instruct",
+            "model": "meta-llama/Llama-3.1-8B-Instruct",
             "backend": "vllm",
             "dataset_name": "sharegpt",
             "dataset_path": "./ShareGPT_V3_unfiltered_cleaned_split.json",
-	    "max_concurrency": 60,
             "num_prompts": 200
         }
     },
     {
         "test_name": "serving_llama8B_tp4_sharegpt",
         "qps_list": [1, 4, 16, "inf"],
+        "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/Meta-Llama-3.1-8B-Instruct",
+            "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": "",
-            "disable_log_requests": "",
 	    "enforce_eager": "",
+	    "max_num_batched_tokens": 2048,
+	    "max_num_seqs": 256,
             "load_format": "dummy"
         },
         "client_parameters": {
-            "model": "meta-llama/Meta-Llama-3.1-8B-Instruct",
+            "model": "meta-llama/Llama-3.1-8B-Instruct",
             "backend": "vllm",
             "dataset_name": "sharegpt",
             "dataset_path": "./ShareGPT_V3_unfiltered_cleaned_split.json",
-	    "max_concurrency": 60,
             "num_prompts": 200
         }
     },
     {
         "test_name": "serving_llama8B_tp4_random_1024_128",
         "qps_list": [1, 4, 16, "inf"],
+        "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/Meta-Llama-3.1-8B-Instruct",
+            "model": "meta-llama/Llama-3.1-8B-Instruct",
             "tensor_parallel_size": 4,
 	    "dtype": "bfloat16",
 	    "distributed_executor_backend": "mp",
@@ -107,32 +115,34 @@
 	    "trust_remote_code": "",
 	    "enable_chunked_prefill": "",
             "disable_log_stats": "",
-            "disable_log_requests": "",
 	    "enforce_eager": "",
+	    "max_num_batched_tokens": 2048,
+	    "max_num_seqs": 256,
             "load_format": "dummy"
         },
         "client_parameters": {
-            "model": "meta-llama/Meta-Llama-3.1-8B-Instruct",
+            "model": "meta-llama/Llama-3.1-8B-Instruct",
             "backend": "vllm",
             "dataset_name": "random",
 	    "random-input-len": 1024,
 	    "random-output-len": 128,
 	    "ignore-eos": "",
-	    "max_concurrency": 100,
             "num_prompts": 100
         }
     },
     {
         "test_name": "serving_llama8B_pp6_random_1024_128",
         "qps_list": [1, 4, 16, "inf"],
+        "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/Meta-Llama-3.1-8B-Instruct",
+            "model": "meta-llama/Llama-3.1-8B-Instruct",
             "pipeline_parallel_size": 6,
 	    "dtype": "bfloat16",
 	    "distributed_executor_backend": "mp",
@@ -140,18 +150,18 @@
 	    "trust_remote_code": "",
 	    "enable_chunked_prefill": "",
             "disable_log_stats": "",
-            "disable_log_requests": "",
 	    "enforce_eager": "",
+	    "max_num_batched_tokens": 2048,
+	    "max_num_seqs": 256,
             "load_format": "dummy"
         },
         "client_parameters": {
-            "model": "meta-llama/Meta-Llama-3.1-8B-Instruct",
+            "model": "meta-llama/Llama-3.1-8B-Instruct",
             "backend": "vllm",
             "dataset_name": "random",
 	    "random-input-len": 1024,
 	    "random-output-len": 128,
 	    "ignore-eos": "",
-	    "max_concurrency": 100,
             "num_prompts": 100
         }
     }

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

@@ -7,7 +7,6 @@
             "tensor_parallel_size": 1,
             "swap_space": 16,
             "disable_log_stats": "",
-            "disable_log_requests": "",
             "load_format": "dummy"
         },
         "client_parameters": {
@@ -26,7 +25,6 @@
             "tensor_parallel_size": 4,
             "swap_space": 16,
             "disable_log_stats": "",
-            "disable_log_requests": "",
             "load_format": "dummy"
         },
         "client_parameters": {
@@ -45,7 +43,6 @@
             "tensor_parallel_size": 2,
             "swap_space": 16,
             "disable_log_stats": "",
-            "disable_log_requests": "",
             "load_format": "dummy"
         },
         "client_parameters": {
@@ -60,8 +57,7 @@
         "test_name": "serving_llama70B_tp4_sharegpt_specdecode",
         "qps_list": [2],
         "server_parameters": {
-            "model": "meta-llama/Meta-Llama-3.1-70B-Instruct",
-            "disable_log_requests": "", 
+            "model": "meta-llama/Meta-Llama-3.1-70B-Instruct", 
             "tensor_parallel_size": 4,
             "swap_space": 16,
             "speculative_config": {

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

@@ -6,7 +6,7 @@
 	    "VLLM_CPU_KVCACHE_SPACE": 40
         },
         "parameters": {
-            "model": "meta-llama/Meta-Llama-3.1-8B-Instruct",
+            "model": "meta-llama/Llama-3.1-8B-Instruct",
             "tensor_parallel_size": 1,
             "load_format": "dummy",
             "dataset": "./ShareGPT_V3_unfiltered_cleaned_split.json",
@@ -21,7 +21,7 @@
 	    "VLLM_CPU_KVCACHE_SPACE": 40
         },
         "parameters": {
-            "model": "meta-llama/Meta-Llama-3.1-8B-Instruct",
+            "model": "meta-llama/Llama-3.1-8B-Instruct",
             "tensor_parallel_size": 4,
             "load_format": "dummy",
             "dataset": "./ShareGPT_V3_unfiltered_cleaned_split.json",

.buildkite/release-pipeline.yaml

@@ -1,4 +1,20 @@
 steps:
+  # aarch64 + CUDA builds
+  - 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=12.8.1 --build-arg torch_cuda_arch_list='8.7 9.0 10.0+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"
+
+  # x86 + CUDA builds
   - label: "Build wheel - CUDA 12.8"
     id: build-wheel-cuda-12-8
     agents:

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

@@ -121,7 +121,6 @@ fi
 if [[ $commands == *" kernels/quantization"* ]]; then
   commands="${commands} \
   --ignore=kernels/quantization/test_int8_quant.py \
-  --ignore=kernels/quantization/test_aqlm.py \
   --ignore=kernels/quantization/test_machete_mm.py \
   --ignore=kernels/quantization/test_block_fp8.py \
   --ignore=kernels/quantization/test_block_int8.py \

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

@@ -78,6 +78,12 @@ function cpu_tests() {
   #   VLLM_USE_V1=0 pytest -s -v \
   #   tests/quantization/test_ipex_quant.py"
 
+  # Run multi-lora tests
+  docker exec cpu-test-"$NUMA_NODE" bash -c "
+    set -e
+    pytest -s -v \
+    tests/lora/test_qwen2vl.py"
+
   # online serving
   docker exec cpu-test-"$NUMA_NODE" bash -c '
     set -e
@@ -89,12 +95,6 @@ function cpu_tests() {
       --model meta-llama/Llama-3.2-3B-Instruct \
       --num-prompts 20 \
       --endpoint /v1/completions'
-
-  # Run multi-lora tests
-  docker exec cpu-test-"$NUMA_NODE" bash -c "
-    set -e
-    pytest -s -v \
-    tests/lora/test_qwen2vl.py"
 }
 
 # All of CPU tests are expected to be finished less than 40 mins.

.buildkite/scripts/hardware_ci/run-gh200-test.sh

@@ -16,8 +16,7 @@ DOCKER_BUILDKIT=1 docker build . \
   --build-arg max_jobs=66 \
   --build-arg nvcc_threads=2 \
   --build-arg RUN_WHEEL_CHECK=false \
-  --build-arg torch_cuda_arch_list="9.0+PTX" \
-  --build-arg vllm_fa_cmake_gpu_arches="90-real"
+  --build-arg torch_cuda_arch_list="9.0+PTX"
 
 # Setup cleanup
 remove_docker_container() { docker rm -f gh200-test || true; }

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

@@ -4,8 +4,7 @@ set -xu
 
 
 remove_docker_container() { 
-    docker rm -f tpu-test || true; 
-    docker rm -f vllm-tpu || true;
+    docker rm -f tpu-test || true;
 }
 
 trap remove_docker_container EXIT
@@ -129,7 +128,7 @@ run_and_track_test() {
 
 # --- Actual Test Execution ---
 run_and_track_test 1 "test_struct_output_generate.py" \
-    "HF_HUB_DISABLE_XET=1 python3 -m pytest -s -v /workspace/vllm/tests/v1/entrypoints/llm/test_struct_output_generate.py -k \"not test_structured_output_with_reasoning_matrices\""
+    "python3 -m pytest -s -v /workspace/vllm/tests/v1/entrypoints/llm/test_struct_output_generate.py -k \"not test_structured_output_with_reasoning_matrices\""
 run_and_track_test 2 "test_moe_pallas.py" \
     "python3 -m pytest -s -v /workspace/vllm/tests/tpu/test_moe_pallas.py"
 run_and_track_test 3 "test_lora.py" \
@@ -140,6 +139,8 @@ run_and_track_test 5 "test_spmd_model_weight_loading.py" \
     "python3 -m pytest -s -v /workspace/vllm/tests/v1/tpu/test_spmd_model_weight_loading.py"
 run_and_track_test 6 "test_kv_cache_update_kernel.py" \
     "python3 -m pytest -s -v /workspace/vllm/tests/v1/tpu/test_kv_cache_update_kernel.py"
+run_and_track_test 7 "test_tpu_int8.py" \
+    "python3 -m pytest -s -v /workspace/vllm/tests/v1/tpu/test_tpu_int8.py"
 
 # After all tests have been attempted, exit with the overall status.
 if [ "$overall_script_exit_code" -ne 0 ]; then

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

@@ -5,7 +5,6 @@ set -xu
 
 remove_docker_container() { 
     docker rm -f tpu-test || true; 
-    docker rm -f vllm-tpu || true;
 }
 
 trap remove_docker_container EXIT
@@ -135,7 +134,7 @@ run_and_track_test 1 "test_compilation.py" \
 run_and_track_test 2 "test_basic.py" \
     "python3 -m pytest -s -v /workspace/vllm/tests/v1/tpu/test_basic.py"
 run_and_track_test 3 "test_accuracy.py::test_lm_eval_accuracy_v1_engine" \
-    "HF_HUB_DISABLE_XET=1 python3 -m pytest -s -v /workspace/vllm/tests/entrypoints/llm/test_accuracy.py::test_lm_eval_accuracy_v1_engine"
+    "python3 -m pytest -s -v /workspace/vllm/tests/entrypoints/llm/test_accuracy.py::test_lm_eval_accuracy_v1_engine"
 run_and_track_test 4 "test_quantization_accuracy.py" \
     "python3 -m pytest -s -v /workspace/vllm/tests/tpu/test_quantization_accuracy.py"
 run_and_track_test 5 "examples/offline_inference/tpu.py" \

.buildkite/scripts/tpu/config_v6e_1.env

@@ -1,6 +1,6 @@
 # Environment config
 TEST_NAME=llama8b
-CONTAINER_NAME=vllm-tpu
+CONTAINER_NAME=tpu-test
 
 # vllm config
 MODEL=meta-llama/Llama-3.1-8B-Instruct

.buildkite/scripts/tpu/docker_run_bm.sh

@@ -12,8 +12,6 @@ source /etc/environment
 source $ENV_FILE
 
 remove_docker_container() { 
-    docker rm -f tpu-test || true; 
-    docker rm -f vllm-tpu || true;
     docker rm -f $CONTAINER_NAME || true;
 }
 

.buildkite/scripts/tpu/quantized_v6e_1.env

@@ -1,6 +1,6 @@
 # Environment config
 TEST_NAME=llama8bw8a8
-CONTAINER_NAME=vllm-tpu
+CONTAINER_NAME=tpu-test
 
 # vllm config
 MODEL=RedHatAI/Meta-Llama-3.1-8B-Instruct-quantized.w8a8

.buildkite/scripts/tpu/run_bm.sh

@@ -44,7 +44,6 @@ echo
 
 VLLM_USE_V1=1 vllm serve $MODEL \
  --seed 42 \
- --disable-log-requests \
  --max-num-seqs $MAX_NUM_SEQS \
  --max-num-batched-tokens $MAX_NUM_BATCHED_TOKENS \
  --tensor-parallel-size $TENSOR_PARALLEL_SIZE \

.buildkite/test-pipeline.yaml

@@ -31,16 +31,6 @@
 steps:
 ##### fast check tests  #####
 
-- label: Documentation Build # 2min
-  mirror_hardwares: [amdexperimental]
-  working_dir: "/vllm-workspace/test_docs"
-  fast_check: true
-  no_gpu: True
-  commands:
-  - pip install -r ../requirements/docs.txt
-  # TODO: add `--strict` once warnings in docstrings are fixed
-  - mkdocs build
-
 - label: Pytorch Nightly Dependency Override Check # 2min
   # if this test fails, it means the nightly torch version is not compatible with some
   # of the dependencies. Please check the error message and add the package to whitelist
@@ -57,20 +47,20 @@ steps:
   - vllm/
   - tests/mq_llm_engine
   - tests/async_engine
-  - tests/test_inputs
+  - tests/test_inputs.py
+  - tests/test_outputs.py
   - tests/multimodal
-  - tests/test_utils
+  - tests/utils_
   - tests/worker
   - tests/standalone_tests/lazy_imports.py
   commands:
   - python3 standalone_tests/lazy_imports.py
   - pytest -v -s mq_llm_engine # MQLLMEngine
   - pytest -v -s async_engine # AsyncLLMEngine
-  - NUM_SCHEDULER_STEPS=4 pytest -v -s async_engine/test_async_llm_engine.py
   - pytest -v -s test_inputs.py
   - pytest -v -s test_outputs.py
   - pytest -v -s multimodal
-  - pytest -v -s test_utils.py # Utils
+  - pytest -v -s utils_ # Utils
   - pytest -v -s worker # Worker
 
 - label: Python-only Installation Test
@@ -82,7 +72,7 @@ steps:
   - bash standalone_tests/python_only_compile.sh
 
 - label: Basic Correctness Test # 30min
-  mirror_hardwares: [amdexperimental, amdproduction]
+  mirror_hardwares: [amdexperimental]
   fast_check: true
   torch_nightly: true
   source_file_dependencies:
@@ -99,7 +89,7 @@ steps:
   - VLLM_TEST_ENABLE_ARTIFICIAL_PREEMPT=1 pytest -v -s basic_correctness/test_preemption.py
 
 - label: Chunked Prefill Test
-  mirror_hardwares: [amdexperimental, amdproduction]
+  mirror_hardwares: [amdexperimental]
   source_file_dependencies:
   - vllm/
   - tests/basic_correctness/test_chunked_prefill
@@ -108,7 +98,7 @@ steps:
   - VLLM_ATTENTION_BACKEND=FLASH_ATTN pytest -v -s basic_correctness/test_chunked_prefill.py
 
 - label: Core Test # 10min
-  mirror_hardwares: [amdexperimental, amdproduction]
+  mirror_hardwares: [amdexperimental]
   fast_check: true
   source_file_dependencies:
   - vllm/core
@@ -128,11 +118,10 @@ steps:
   - tests/entrypoints/offline_mode
   commands:
   - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-  - pytest -v -s entrypoints/llm --ignore=entrypoints/llm/test_lazy_outlines.py --ignore=entrypoints/llm/test_generate.py --ignore=entrypoints/llm/test_generate_multiple_loras.py --ignore=entrypoints/llm/test_guided_generate.py --ignore=entrypoints/llm/test_collective_rpc.py
+  - pytest -v -s entrypoints/llm --ignore=entrypoints/llm/test_lazy_outlines.py --ignore=entrypoints/llm/test_generate.py --ignore=entrypoints/llm/test_generate_multiple_loras.py --ignore=entrypoints/llm/test_collective_rpc.py
   - pytest -v -s entrypoints/llm/test_lazy_outlines.py # it needs a clean process
   - pytest -v -s entrypoints/llm/test_generate.py # it needs a clean process
   - pytest -v -s entrypoints/llm/test_generate_multiple_loras.py # it needs a clean process
-  - VLLM_USE_V1=0 pytest -v -s entrypoints/llm/test_guided_generate.py # it needs a clean process
   - VLLM_USE_V1=0 pytest -v -s entrypoints/offline_mode # Needs to avoid interference with other tests
 
 - label: Entrypoints Test (API Server) # 40min
@@ -210,7 +199,7 @@ steps:
   - pytest -v -s distributed/test_eplb_execute.py
 
 - label: Metrics, Tracing Test # 10min
-  mirror_hardwares: [amdexperimental, amdproduction]
+  mirror_hardwares: [amdexperimental]
   num_gpus: 2
   source_file_dependencies:
   - vllm/
@@ -229,7 +218,7 @@ steps:
 #####  1 GPU test  #####
 
 - label: Regression Test # 5min
-  mirror_hardwares: [amdexperimental, amdproduction]
+  mirror_hardwares: [amdexperimental]
   source_file_dependencies:
   - vllm/
   - tests/test_regression
@@ -264,6 +253,7 @@ steps:
     - pytest -v -s v1/engine
     - pytest -v -s v1/entrypoints
     - pytest -v -s v1/sample
+    - pytest -v -s v1/logits_processors
     - pytest -v -s v1/worker
     - pytest -v -s v1/structured_output
     - pytest -v -s v1/spec_decode
@@ -281,7 +271,7 @@ steps:
     - pytest -v -s entrypoints/openai/correctness/test_lmeval.py::test_lm_eval_accuracy_v1_engine
 
 - label: Examples Test # 25min
-  mirror_hardwares: [amdexperimental, amdproduction]
+  mirror_hardwares: [amdexperimental]
   working_dir: "/vllm-workspace/examples"
   source_file_dependencies:
   - vllm/entrypoints
@@ -306,7 +296,7 @@ steps:
     - VLLM_USE_V1=0 python3 offline_inference/profiling.py --model facebook/opt-125m run_num_steps --num-steps 2
 
 - label: Prefix Caching Test # 9min
-  mirror_hardwares: [amdexperimental, amdproduction]
+  mirror_hardwares: [amdexperimental]
   source_file_dependencies:
   - vllm/
   - tests/prefix_caching
@@ -315,7 +305,7 @@ steps:
 
 
 - label: Platform Tests (CUDA)
-  mirror_hardwares: [amdexperimental, amdproduction]
+  mirror_hardwares: [amdexperimental]
   source_file_dependencies:
   - vllm/
   - tests/cuda
@@ -354,9 +344,10 @@ steps:
     - pytest -v -s compile/test_silu_mul_quant_fusion.py
     - pytest -v -s compile/test_sequence_parallelism.py
     - pytest -v -s compile/test_async_tp.py
+    - pytest -v -s compile/test_fusion_all_reduce.py
 
 - label: PyTorch Fullgraph Smoke Test # 9min
-  mirror_hardwares: [amdexperimental, amdproduction]
+  mirror_hardwares: [amdexperimental]
   torch_nightly: true
   source_file_dependencies:
   - vllm/
@@ -369,7 +360,7 @@ steps:
   - pytest -v -s compile/piecewise/test_full_cudagraph.py
 
 - label: PyTorch Fullgraph Test # 18min
-  mirror_hardwares: [amdexperimental, amdproduction]
+  mirror_hardwares: [amdexperimental]
   torch_nightly: true
   source_file_dependencies:
   - vllm/
@@ -378,7 +369,7 @@ steps:
   - pytest -v -s compile/test_full_graph.py
 
 - label: Kernels Core Operation Test
-  mirror_hardwares: [amdexperimental, amdproduction]
+  mirror_hardwares: [amdexperimental]
   source_file_dependencies:
   - csrc/
   - tests/kernels/core
@@ -403,20 +394,22 @@ steps:
   - vllm/model_executor/layers/quantization
   - tests/kernels/quantization
   commands:
-    - pytest -v -s kernels/quantization  --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT
+    - pytest -v -s kernels/quantization --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT
   parallelism: 2
 
-- label: Kernels MoE Test
+- label: Kernels MoE Test %N
   mirror_hardwares: [amdexperimental]
   source_file_dependencies:
+  - csrc/quantization/cutlass_w8a8/moe/
   - csrc/moe/
   - tests/kernels/moe
   - vllm/model_executor/layers/fused_moe/
   commands:
-    - pytest -v -s kernels/moe
+    - pytest -v -s kernels/moe --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT
+  parallelism: 2
 
 - label: Kernels Mamba Test
-  mirror_hardwares: [amdexperimental, amdproduction]
+  mirror_hardwares: [amdexperimental]
   source_file_dependencies:
   - csrc/mamba/
   - tests/kernels/mamba
@@ -424,8 +417,7 @@ steps:
     - pytest -v -s kernels/mamba
 
 - label: Tensorizer Test # 11min
-  mirror_hardwares: [amdexperimental, amdproduction]
-  soft_fail: true
+  mirror_hardwares: [amdexperimental]
   source_file_dependencies:
   - vllm/model_executor/model_loader
   - tests/tensorizer_loader
@@ -437,7 +429,7 @@ steps:
     - pytest -v -s entrypoints/openai/test_tensorizer_entrypoint.py
 
 - label: Model Executor Test
-  mirror_hardwares: [amdexperimental, amdproduction]
+  mirror_hardwares: [amdexperimental]
   source_file_dependencies:
   - vllm/model_executor
   - tests/model_executor
@@ -447,7 +439,7 @@ steps:
     - pytest -v -s model_executor
 
 - label: Benchmarks # 9min
-  mirror_hardwares: [amdexperimental, amdproduction]
+  mirror_hardwares: [amdexperimental]
   working_dir: "/vllm-workspace/.buildkite"
   source_file_dependencies:
   - benchmarks/
@@ -455,7 +447,7 @@ steps:
   - bash scripts/run-benchmarks.sh
 
 - label: Benchmarks CLI Test # 10min
-  mirror_hardwares: [amdexperimental, amdproduction]
+  mirror_hardwares: [amdexperimental]
   source_file_dependencies:
   - vllm/
   - tests/benchmarks/
@@ -494,7 +486,7 @@ steps:
   - pytest -s entrypoints/openai/correctness/
 
 - label: Encoder Decoder tests # 5min
-  mirror_hardwares: [amdexperimental, amdproduction]
+  mirror_hardwares: [amdexperimental]
   source_file_dependencies:
   - vllm/
   - tests/encoder_decoder
@@ -502,7 +494,7 @@ steps:
     - pytest -v -s encoder_decoder
 
 - label: OpenAI-Compatible Tool Use # 20 min
-  mirror_hardwares: [amdexperimental, amdproduction]
+  mirror_hardwares: [amdexperimental]
   fast_check: false
   source_file_dependencies:
     - vllm/
@@ -534,8 +526,6 @@ steps:
   - vllm/
   - tests/models/language
   commands:
-    # Install causal-conv1d for plamo2 models here, as it is not compatible with pip-compile.
-    - pip install 'git+https://github.com/Dao-AILab/causal-conv1d@v1.5.0.post8'
     - pip freeze | grep -E 'torch'
     - pytest -v -s models/language -m core_model
 
@@ -546,8 +536,10 @@ steps:
   - vllm/
   - tests/models/language/generation
   commands:
-    # Install causal-conv1d for plamo2 models here, as it is not compatible with pip-compile.
-    - pip install 'git+https://github.com/Dao-AILab/causal-conv1d@v1.5.0.post8'
+    # Install fast path packages for testing against transformers
+    # Note: also needed to run plamo2 model in vLLM
+    - uv pip install --system --no-build-isolation 'git+https://github.com/state-spaces/mamba@v2.2.5'
+    - uv pip install --system --no-build-isolation 'git+https://github.com/Dao-AILab/causal-conv1d@v1.5.2'
     - pytest -v -s models/language/generation -m hybrid_model
 
 - label: Language Models Test (Extended Generation) # 1hr20min
@@ -580,7 +572,8 @@ steps:
     - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
     - pip freeze | grep -E 'torch'
     - pytest -v -s models/multimodal/processing
-    - pytest -v -s --ignore models/multimodal/generation/test_whisper.py models/multimodal -m core_model
+    - pytest -v -s --ignore models/multimodal/generation/test_whisper.py --ignore models/multimodal/test_tensor_schema.py models/multimodal -m core_model
+    - pytest -v -s models/multimodal/test_tensor_schema.py -m core_model  # Needs mp_method="spawn"
     - cd .. && pytest -v -s tests/models/multimodal/generation/test_whisper.py -m core_model  # Otherwise, mp_method="spawn" doesn't work
 
 - label: Multi-Modal Models Test (Extended) 1
@@ -623,7 +616,7 @@ steps:
 
 # This test is used only in PR development phase to test individual models and should never run on main
 - label: Custom Models Test
-  mirror_hardwares: [amdexperimental, amdproduction]
+  mirror_hardwares: [amdexperimental]
   optional: true
   commands:
     - echo 'Testing custom models...'
@@ -643,11 +636,41 @@ steps:
     - python3 examples/offline_inference/audio_language.py --model-type whisper
     - python3 examples/offline_inference/vision_language.py --model-type qwen2_5_vl
 
+- label: Blackwell Test
+  working_dir: "/vllm-workspace/"
+  gpu: b200
+  # optional: true
+  source_file_dependencies:
+  - csrc/quantization/fp4/
+  - csrc/attention/mla/
+  - csrc/quantization/cutlass_w8a8/moe/
+  - vllm/model_executor/layers/fused_moe/cutlass_moe.py
+  - vllm/model_executor/layers/fused_moe/flashinfer_cutlass_moe.py
+  - vllm/model_executor/layers/fused_moe/flashinfer_cutlass_prepare_finalize.py
+  - vllm/v1/attention/backends/flashinfer.py
+  - vllm/compilation/fusion.py
+  commands:
+    - nvidia-smi
+    - python3 examples/offline_inference/basic/chat.py
+    # Attention
+    # num_heads2 broken by https://github.com/flashinfer-ai/flashinfer/issues/1353
+    - pytest -v -s tests/kernels/attention/test_flashinfer.py -k 'not num_heads2'
+    - pytest -v -s tests/kernels/attention/test_flashinfer_trtllm_attention.py
+    - pytest -v -s tests/kernels/test_cutlass_mla_decode.py
+    # Quantization
+    - pytest -v -s tests/kernels/quantization/test_cutlass_scaled_mm.py -k 'fp8'
+    - pytest -v -s tests/kernels/quantization/test_nvfp4_quant.py
+    - pytest -v -s tests/kernels/quantization/test_nvfp4_scaled_mm.py
+    - pytest -v -s tests/kernels/quantization/test_flashinfer_nvfp4_scaled_mm.py
+    - pytest -v -s tests/kernels/moe/test_nvfp4_moe.py
+    # Fusion
+    - pytest -v -s tests/compile/test_fusion_all_reduce.py
+
 #####  1 GPU test  #####
 #####  multi gpus test  #####
 
 - label: Distributed Comm Ops Test # 7min
-  mirror_hardwares: [amdexperimental, amdproduction]
+  mirror_hardwares: [amdexperimental]
   working_dir: "/vllm-workspace/tests"
   num_gpus: 2
   source_file_dependencies:
@@ -718,7 +741,6 @@ steps:
   # this test fails consistently.
   # TODO: investigate and fix
   - VLLM_USE_V1=0 CUDA_VISIBLE_DEVICES=0,1 pytest -v -s test_sharded_state_loader.py
-  - VLLM_USE_V1=0 CUDA_VISIBLE_DEVICES=0,1 pytest -v -s kv_transfer/test_disagg.py
   - CUDA_VISIBLE_DEVICES=0,1 pytest -v -s v1/shutdown
   - pytest -v -s models/multimodal/generation/test_maverick.py
 
@@ -743,29 +765,8 @@ steps:
   - pytest -v -s models/test_oot_registration.py # it needs a clean process
   - pytest -v -s plugins/lora_resolvers # unit tests for in-tree lora resolver plugins
 
-- label: Multi-step Tests (4 GPUs) # 36min
-  mirror_hardwares: [amdexperimental, amdproduction]
-  working_dir: "/vllm-workspace/tests"
-  num_gpus: 4
-  source_file_dependencies:
-  - vllm/model_executor/layers/sampler.py
-  - vllm/sequence.py
-  - vllm/worker/worker_base.py
-  - vllm/worker/worker.py
-  - vllm/worker/multi_step_worker.py
-  - vllm/worker/model_runner_base.py
-  - vllm/worker/model_runner.py
-  - vllm/worker/multi_step_model_runner.py
-  - vllm/engine
-  - tests/multi_step
-  commands:
-  # this test is quite flaky
-  # TODO: investigate and fix.
-  # - pytest -v -s multi_step/test_correctness_async_llm.py
-  - pytest -v -s multi_step/test_correctness_llm.py
-
 - label: Pipeline Parallelism Test # 45min
-  mirror_hardwares: [amdexperimental, amdproduction]
+  mirror_hardwares: [amdexperimental]
   working_dir: "/vllm-workspace/tests"
   num_gpus: 4
   source_file_dependencies:
@@ -779,7 +780,7 @@ steps:
   - pytest -v -s distributed/test_pipeline_parallel.py
 
 - label: LoRA TP Test (Distributed)
-  mirror_hardwares: [amdexperimental, amdproduction]
+  mirror_hardwares: [amdexperimental]
   num_gpus: 4
   source_file_dependencies:
   - vllm/lora
@@ -792,6 +793,7 @@ steps:
     # requires multi-GPU testing for validation.
     - pytest -v -s -x lora/test_chatglm3_tp.py
     - pytest -v -s -x lora/test_llama_tp.py
+    - pytest -v -s -x lora/test_multi_loras_with_tp.py
 
 
 - label: Weight Loading Multiple GPU Test  # 33min

.github/CODEOWNERS

@@ -9,8 +9,7 @@
 /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
-/vllm/model_executor/guided_decoding @mgoin @russellb @aarnphm
+/vllm/model_executor/layers/quantization @mgoin @robertgshaw2-redhat @tlrmchlsmth @yewentao256
 /vllm/multimodal @DarkLight1337 @ywang96
 /vllm/vllm_flash_attn @LucasWilkinson
 /vllm/lora @jeejeelee
@@ -21,7 +20,7 @@ CMakeLists.txt @tlrmchlsmth @LucasWilkinson
 
 # Any change to the VllmConfig changes can have a large user-facing impact,
 # so spam a lot of people
-/vllm/config.py @simon-mo @WoosukKwon @youkaichao @robertgshaw2-redhat @mgoin @tlrmchlsmth @houseroad @hmellor
+/vllm/config @simon-mo @WoosukKwon @youkaichao @robertgshaw2-redhat @mgoin @tlrmchlsmth @houseroad @hmellor @yewentao256 @ProExpertProg
 
 # vLLM V1
 /vllm/v1 @WoosukKwon @robertgshaw2-redhat @njhill @ywang96 @comaniac @alexm-redhat
@@ -35,18 +34,15 @@ CMakeLists.txt @tlrmchlsmth @LucasWilkinson
 /tests/distributed/test_pipeline_parallel.py @youkaichao
 /tests/distributed/test_same_node.py @youkaichao
 /tests/entrypoints @DarkLight1337 @robertgshaw2-redhat @simon-mo @aarnphm
-/tests/entrypoints/llm/test_guided_generate.py @mgoin @russellb @aarnphm
-/tests/kernels @tlrmchlsmth @WoosukKwon
-/tests/model_executor/test_guided_processors.py @mgoin @russellb
+/tests/kernels @tlrmchlsmth @WoosukKwon @yewentao256
 /tests/models @DarkLight1337 @ywang96
-/tests/multi_step @alexm-redhat @comaniac
 /tests/multimodal @DarkLight1337 @ywang96
 /tests/prefix_caching @comaniac @KuntaiDu
-/tests/quantization @mgoin @robertgshaw2-redhat
+/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/weight_loading @mgoin @youkaichao
+/tests/weight_loading @mgoin @youkaichao @yewentao256
 /tests/lora @jeejeelee
 
 # Docs
@@ -64,3 +60,15 @@ mkdocs.yaml @hmellor
 /vllm/v1/worker/^xpu @jikunshang
 /vllm/platforms/xpu.py @jikunshang
 /docker/Dockerfile.xpu @jikunshang
+
+# Qwen-specific files
+/vllm/attention/backends/dual_chunk_flash_attn.py @sighingnow
+/vllm/model_executor/models/qwen* @sighingnow
+
+# Mistral-specific files
+/vllm/model_executor/models/mistral*.py @patrickvonplaten
+/vllm/model_executor/models/mixtral*.py @patrickvonplaten
+/vllm/model_executor/models/voxtral*.py @patrickvonplaten
+/vllm/model_executor/models/pixtral*.py @patrickvonplaten
+/vllm/transformers_utils/configs/mistral.py @patrickvonplaten
+/vllm/transformers_utils/tokenizers/mistral.py @patrickvonplaten

.github/PULL_REQUEST_TEMPLATE.md

@@ -1,10 +1,5 @@
-## Essential Elements of an Effective PR Description Checklist
-- [ ] The purpose of the PR, such as "Fix some issue (link existing issues this PR will resolve)".
-- [ ] The test plan, such as providing test command.
-- [ ] The test results, such as pasting the results comparison before and after, or e2e results
-- [ ] (Optional) The necessary documentation update, such as updating `supported_models.md` and `examples` for a new model.
-
-PLEASE FILL IN THE PR DESCRIPTION HERE ENSURING ALL CHECKLIST ITEMS ABOVE HAVE BEEN CONSIDERED.
+<!-- markdownlint-disable -->
+PLEASE FILL IN THE PR DESCRIPTION HERE ENSURING ALL CHECKLIST ITEMS (AT THE BOTTOM) HAVE BEEN CONSIDERED.
 
 ## Purpose
 
@@ -14,5 +9,14 @@ PLEASE FILL IN THE PR DESCRIPTION HERE ENSURING ALL CHECKLIST ITEMS ABOVE HAVE B
 
 ## (Optional) Documentation Update
 
-<!--- pyml disable-next-line no-emphasis-as-heading -->
+---
+<details>
+<summary> Essential Elements of an Effective PR Description Checklist </summary>
+
+- [ ] The purpose of the PR, such as "Fix some issue (link existing issues this PR will resolve)".
+- [ ] The test plan, such as providing test command.
+- [ ] The test results, such as pasting the results comparison before and after, or e2e results
+- [ ] (Optional) The necessary documentation update, such as updating `supported_models.md` and `examples` for a new model.
+</details>
+
 **BEFORE SUBMITTING, PLEASE READ <https://docs.vllm.ai/en/latest/contributing>** (anything written below this line will be removed by GitHub Actions)

.github/mergify.yml

@@ -118,6 +118,20 @@ pull_request_rules:
       add:
         - qwen
 
+- name: label-gpt-oss
+  description: Automatically apply gpt-oss label
+  conditions:
+    - or:
+      - files~=^examples/.*gpt[-_]?oss.*\.py
+      - files~=^tests/.*gpt[-_]?oss.*\.py
+      - files~=^vllm/model_executor/models/.*gpt[-_]?oss.*\.py
+      - files~=^vllm/model_executor/layers/.*gpt[-_]?oss.*\.py
+      - title~=(?i)gpt[-_]?oss
+  actions:
+    label:
+      add:
+        - gpt-oss
+
 - name: label-rocm
   description: Automatically apply rocm label
   conditions:
@@ -149,9 +163,6 @@ pull_request_rules:
       - files=examples/offline_inference/structured_outputs.py
       - files=examples/online_serving/openai_chat_completion_structured_outputs.py
       - files=examples/online_serving/openai_chat_completion_structured_outputs_with_reasoning.py
-      - files~=^vllm/model_executor/guided_decoding/
-      - files=tests/model_executor/test_guided_processors.py
-      - files=tests/entrypoints/llm/test_guided_generate.py
       - files~=^tests/v1/structured_output/
       - files=tests/v1/entrypoints/llm/test_guided_generate.py
       - files~=^vllm/v1/structured_output/

.github/scripts/cleanup_pr_body.sh

@@ -15,11 +15,11 @@ NEW=/tmp/new_pr_body.txt
 gh pr view --json body --template "{{.body}}" "${PR_NUMBER}" > "${OLD}"
 cp "${OLD}" "${NEW}"
 
-# Remove "FIX #xxxx (*link existing issues this PR will resolve*)"
-sed -i '/FIX #xxxx.*$/d' "${NEW}"
+# Remove markdown comments (like the <!-- markdownlint-disable --> at the start)
+sed -i '/<!--.*-->$/d' "${NEW}"
 
-# Remove "FILL IN THE PR DESCRIPTION HERE"
-sed -i '/FILL IN THE PR DESCRIPTION HERE/d' "${NEW}"
+# Remove "PLEASE FILL IN THE PR DESCRIPTION HERE ENSURING ALL CHECKLIST ITEMS (AT THE BOTTOM) HAVE BEEN CONSIDERED."
+sed -i '/PLEASE FILL IN THE PR DESCRIPTION HERE.*$/d' "${NEW}"
 
 # Remove all lines after and including "**BEFORE SUBMITTING, PLEASE READ THE CHECKLIST BELOW AND FILL IN THE DESCRIPTION ABOVE**"
 sed -i '/\*\*BEFORE SUBMITTING, PLEASE READ.*\*\*/,$d' "${NEW}"

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

@@ -2,12 +2,16 @@ name: Lint and Deploy Charts
 
 on: pull_request
 
+concurrency:
+  group: ${{ github.workflow }}-${{ github.ref }}
+  cancel-in-progress: true
+
 permissions:
   contents: read
 
 jobs:
   lint-and-deploy:
-    runs-on: ubuntu-24.04-arm
+    runs-on: ubuntu-latest
     steps:
       - name: Checkout
         uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2

.github/workflows/matchers/markdownlint.json

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

.github/workflows/pre-commit.yml

@@ -5,6 +5,10 @@ on:
   push:
     branches: [main]
 
+concurrency:
+  group: ${{ github.workflow }}-${{ github.ref }}
+  cancel-in-progress: ${{ github.event_name == 'pull_request' }}
+
 permissions:
   contents: read
 
@@ -17,6 +21,7 @@ jobs:
       with:
         python-version: "3.12"
     - run: echo "::add-matcher::.github/workflows/matchers/actionlint.json"
+    - run: echo "::add-matcher::.github/workflows/matchers/markdownlint.json"
     - run: echo "::add-matcher::.github/workflows/matchers/mypy.json"
     - uses: pre-commit/action@2c7b3805fd2a0fd8c1884dcaebf91fc102a13ecd # v3.0.1
       with:

.github/workflows/scripts/build.sh

@@ -15,7 +15,6 @@ $python_executable -m pip install -r requirements/build.txt -r requirements/cuda
 export MAX_JOBS=1
 # Make sure release wheels are built for the following architectures
 export TORCH_CUDA_ARCH_LIST="7.0 7.5 8.0 8.6 8.9 9.0+PTX"
-export VLLM_FA_CMAKE_GPU_ARCHES="80-real;90-real"
 
 bash tools/check_repo.sh
 

.gitignore

@@ -4,6 +4,9 @@
 # vllm-flash-attn built from source
 vllm/vllm_flash_attn/*
 
+# triton jit 
+.triton
+
 # Byte-compiled / optimized / DLL files
 __pycache__/
 *.py[cod]
@@ -147,7 +150,8 @@ venv.bak/
 # mkdocs documentation
 /site
 docs/argparse
-docs/examples
+docs/examples/*
+!docs/examples/README.md
 
 # mypy
 .mypy_cache/
@@ -203,3 +207,6 @@ shellcheck*/
 
 # Ignore moe/marlin_moe gen code
 csrc/moe/marlin_moe_wna16/kernel_*
+
+# Ignore ep_kernels_workspace folder
+ep_kernels_workspace/

.markdownlint.yaml

@@ -0,0 +1,13 @@
+MD007:
+  indent: 4
+MD013: false
+MD024:
+  siblings_only: true
+MD033: false
+MD042: false
+MD045: false
+MD046: false
+MD051: false
+MD052: false
+MD053: false
+MD059: false

.pre-commit-config.yaml

@@ -35,12 +35,12 @@ repos:
     exclude: 'csrc/(moe/topk_softmax_kernels.cu|quantization/gguf/(ggml-common.h|dequantize.cuh|vecdotq.cuh|mmq.cuh|mmvq.cuh))|vllm/third_party/.*'
     types_or: [c++, cuda]
     args: [--style=file, --verbose]
-- repo: https://github.com/jackdewinter/pymarkdown
-  rev: v0.9.29
+- repo: https://github.com/igorshubovych/markdownlint-cli
+  rev: v0.45.0
   hooks:
-  - id: pymarkdown
+  - id: markdownlint
     exclude: '.*\.inc\.md'
-    args: [fix]
+    stages: [manual] # Only run in CI
 - repo: https://github.com/rhysd/actionlint
   rev: v1.7.7
   hooks:

.readthedocs.yaml

@@ -7,6 +7,9 @@ build:
   os: ubuntu-22.04
   tools:
     python: "3.12"
+  jobs:
+    post_checkout:
+      - git fetch --unshallow || true
 
 mkdocs:
   configuration: mkdocs.yaml

CMakeLists.txt

@@ -249,7 +249,6 @@ set(VLLM_EXT_SRC
   "csrc/quantization/gguf/gguf_kernel.cu"
   "csrc/quantization/activation_kernels.cu"
   "csrc/cuda_utils_kernels.cu"
-  "csrc/prepare_inputs/advance_step.cu"
   "csrc/custom_all_reduce.cu"
   "csrc/torch_bindings.cpp")
 
@@ -287,7 +286,6 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
   FetchContent_MakeAvailable(cutlass)
 
   list(APPEND VLLM_EXT_SRC
-    "csrc/quantization/aqlm/gemm_kernels.cu"
     "csrc/quantization/awq/gemm_kernels.cu"
     "csrc/permute_cols.cu"
     "csrc/quantization/cutlass_w8a8/scaled_mm_entry.cu"
@@ -351,6 +349,10 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
     set_gencode_flags_for_srcs(
       SRCS "${MARLIN_TEMPLATE_KERNEL_SRC}"
       CUDA_ARCHS "${MARLIN_ARCHS}")
+    if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8)
+      set_source_files_properties(${MARLIN_TEMPLATE_KERNEL_SRC}
+        PROPERTIES COMPILE_FLAGS "-static-global-template-stub=false")
+    endif()
 
     list(APPEND VLLM_EXT_SRC ${MARLIN_TEMPLATE_KERNEL_SRC})
 
@@ -364,7 +366,12 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
     set_gencode_flags_for_srcs(
       SRCS "${MARLIN_SRCS}"
       CUDA_ARCHS "${MARLIN_ARCHS}")
+    if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8)
+      set_source_files_properties("csrc/quantization/gptq_marlin/gptq_marlin.cu"
+        PROPERTIES COMPILE_FLAGS "-static-global-template-stub=false")
+    endif()
     list(APPEND VLLM_EXT_SRC "${MARLIN_SRCS}")
+
     message(STATUS "Building Marlin kernels for archs: ${MARLIN_ARCHS}")
   else()
     message(STATUS "Not building Marlin kernels as no compatible archs found"
@@ -427,6 +434,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
     set(SRCS
       "csrc/quantization/cutlass_w8a8/scaled_mm_c3x_sm120.cu"
       "csrc/quantization/cutlass_w8a8/c3x/scaled_mm_sm120_fp8.cu"
+      "csrc/quantization/cutlass_w8a8/c3x/scaled_mm_blockwise_sm120_fp8.cu"
     )
     set_gencode_flags_for_srcs(
       SRCS "${SRCS}"
@@ -529,6 +537,25 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
     endif()
   endif()
 
+  # The nvfp4_scaled_mm_sm120 kernels for Geforce Blackwell SM120 require
+  # CUDA 12.8 or later
+  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"
+      "csrc/quantization/fp4/nvfp4_scaled_mm_sm120_kernels.cu")
+    set_gencode_flags_for_srcs(
+      SRCS "${SRCS}"
+      CUDA_ARCHS "${FP4_ARCHS}")
+    list(APPEND VLLM_EXT_SRC "${SRCS}")
+    list(APPEND VLLM_GPU_FLAGS "-DENABLE_NVFP4_SM120=1")
+    message(STATUS "Building NVFP4 for archs: ${FP4_ARCHS}")
+  else()
+    message(STATUS "Not building NVFP4 as no compatible archs were found.")
+    # clear FP4_ARCHS
+    set(FP4_ARCHS)
+  endif()
+
   # FP4 Archs and flags
   cuda_archs_loose_intersection(FP4_ARCHS "10.0a" "${CUDA_ARCHS}")
   if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND FP4_ARCHS)
@@ -541,7 +568,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
       SRCS "${SRCS}"
       CUDA_ARCHS "${FP4_ARCHS}")
     list(APPEND VLLM_EXT_SRC "${SRCS}")
-    list(APPEND VLLM_GPU_FLAGS "-DENABLE_NVFP4=1")
+    list(APPEND VLLM_GPU_FLAGS "-DENABLE_NVFP4_SM100=1")
     list(APPEND VLLM_GPU_FLAGS "-DENABLE_CUTLASS_MOE_SM100=1")
     message(STATUS "Building NVFP4 for archs: ${FP4_ARCHS}")
   else()
@@ -834,6 +861,10 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
     set_gencode_flags_for_srcs(
       SRCS "${MOE_WNAA16_MARLIN_SRC}"
       CUDA_ARCHS "${MARLIN_MOE_ARCHS}")
+    if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8)
+      set_source_files_properties(${MOE_WNAA16_MARLIN_SRC}
+        PROPERTIES COMPILE_FLAGS "-static-global-template-stub=false")
+    endif()
 
     list(APPEND VLLM_MOE_EXT_SRC ${MOE_WNAA16_MARLIN_SRC})
 

README.md

@@ -1,3 +1,4 @@
+<!-- markdownlint-disable MD001 MD041 -->
 <p align="center">
   <picture>
     <source media="(prefers-color-scheme: dark)" srcset="https://raw.githubusercontent.com/vllm-project/vllm/main/docs/assets/logos/vllm-logo-text-dark.png">
@@ -16,14 +17,16 @@ Easy, fast, and cheap LLM serving for everyone
 ---
 
 *Latest News* 🔥
+
+- [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).
 - [2025/05] vLLM is now a hosted project under PyTorch Foundation! Please find the announcement [here](https://pytorch.org/blog/pytorch-foundation-welcomes-vllm/).
-- [2025/04] We hosted [Asia Developer Day](https://www.sginnovate.com/event/limited-availability-morning-evening-slots-remaining-inaugural-vllm-asia-developer-day)! Please find the meetup slides from the vLLM team [here](https://docs.google.com/presentation/d/19cp6Qu8u48ihB91A064XfaXruNYiBOUKrBxAmDOllOo/edit?usp=sharing).
 - [2025/01] 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>
 
+- [2025/04] We hosted [Asia Developer Day](https://www.sginnovate.com/event/limited-availability-morning-evening-slots-remaining-inaugural-vllm-asia-developer-day)! Please find the meetup slides from the vLLM team [here](https://docs.google.com/presentation/d/19cp6Qu8u48ihB91A064XfaXruNYiBOUKrBxAmDOllOo/edit?usp=sharing).
 - [2025/03] We hosted [vLLM x Ollama Inference Night](https://lu.ma/vllm-ollama)! Please find the meetup slides from the vLLM team [here](https://docs.google.com/presentation/d/16T2PDD1YwRnZ4Tu8Q5r6n53c5Lr5c73UV9Vd2_eBo4U/edit?usp=sharing).
 - [2025/03] We hosted [the first vLLM China Meetup](https://mp.weixin.qq.com/s/n77GibL2corAtQHtVEAzfg)! Please find the meetup slides from vLLM team [here](https://docs.google.com/presentation/d/1REHvfQMKGnvz6p3Fd23HhSO4c8j5WPGZV0bKYLwnHyQ/edit?usp=sharing).
 - [2025/03] We hosted [the East Coast vLLM Meetup](https://lu.ma/7mu4k4xx)! Please find the meetup slides [here](https://docs.google.com/presentation/d/1NHiv8EUFF1NLd3fEYODm56nDmL26lEeXCaDgyDlTsRs/edit#slide=id.g31441846c39_0_0).
@@ -46,6 +49,7 @@ Easy, fast, and cheap LLM serving for everyone
 </details>
 
 ---
+
 ## About
 
 vLLM is a fast and easy-to-use library for LLM inference and serving.
@@ -75,6 +79,7 @@ vLLM is flexible and easy to use with:
 - Multi-LoRA support
 
 vLLM seamlessly supports most popular open-source models on HuggingFace, including:
+
 - Transformer-like LLMs (e.g., Llama)
 - Mixture-of-Expert LLMs (e.g., Mixtral, Deepseek-V2 and V3)
 - Embedding Models (e.g., E5-Mistral)
@@ -91,6 +96,7 @@ pip install vllm
 ```
 
 Visit our [documentation](https://docs.vllm.ai/en/latest/) to learn more.
+
 - [Installation](https://docs.vllm.ai/en/latest/getting_started/installation.html)
 - [Quickstart](https://docs.vllm.ai/en/latest/getting_started/quickstart.html)
 - [List of Supported Models](https://docs.vllm.ai/en/latest/models/supported_models.html)
@@ -107,6 +113,7 @@ vLLM is a community project. Our compute resources for development and testing a
 <!-- Note: Please sort them in alphabetical order. -->
 <!-- Note: Please keep these consistent with docs/community/sponsors.md -->
 Cash Donations:
+
 - a16z
 - Dropbox
 - Sequoia Capital
@@ -114,6 +121,8 @@ Cash Donations:
 - ZhenFund
 
 Compute Resources:
+
+- Alibaba Cloud
 - AMD
 - Anyscale
 - AWS
@@ -153,7 +162,7 @@ If you use vLLM for your research, please cite our [paper](https://arxiv.org/abs
 ## Contact Us
 
 <!-- --8<-- [start:contact-us] -->
-- For technical questions and feature requests, please use GitHub [Issues](https://github.com/vllm-project/vllm/issues) or [Discussions](https://github.com/vllm-project/vllm/discussions)
+- For technical questions and feature requests, please use GitHub [Issues](https://github.com/vllm-project/vllm/issues)
 - For discussing with fellow users, please use the [vLLM Forum](https://discuss.vllm.ai)
 - For coordinating contributions and development, please use [Slack](https://slack.vllm.ai)
 - For security disclosures, please use GitHub's [Security Advisories](https://github.com/vllm-project/vllm/security/advisories) feature

RELEASE.md

@@ -60,9 +60,10 @@ Please note: **No feature work allowed for cherry picks**. All PRs that are cons
 Before each release, we perform end-to-end performance validation to ensure no regressions are introduced. This validation uses the [vllm-benchmark workflow](https://github.com/pytorch/pytorch-integration-testing/actions/workflows/vllm-benchmark.yml) on PyTorch CI.
 
 **Current Coverage:**
+
 * Models: Llama3, Llama4, and Mixtral
 * Hardware: NVIDIA H100 and AMD MI300x
-* *Note: Coverage may change based on new model releases and hardware availability*
+* _Note: Coverage may change based on new model releases and hardware availability_
 
 **Performance Validation Process:**
 
@@ -71,11 +72,13 @@ Request write access to the [pytorch/pytorch-integration-testing](https://github
 
 **Step 2: Review Benchmark Setup**
 Familiarize yourself with the benchmark configurations:
+
 * [CUDA setup](https://github.com/pytorch/pytorch-integration-testing/tree/main/vllm-benchmarks/benchmarks/cuda)
 * [ROCm setup](https://github.com/pytorch/pytorch-integration-testing/tree/main/vllm-benchmarks/benchmarks/rocm)
 
 **Step 3: Run the Benchmark**
 Navigate to the [vllm-benchmark workflow](https://github.com/pytorch/pytorch-integration-testing/actions/workflows/vllm-benchmark.yml) and configure:
+
 * **vLLM branch**: Set to the release branch (e.g., `releases/v0.9.2`)
 * **vLLM commit**: Set to the RC commit hash
 

SECURITY.md

@@ -1,13 +1,45 @@
 # Security Policy
 
-## Reporting a Vulnerability
+## Reporting security issues
 
-If you believe you have found a security vulnerability in vLLM, we encourage you to let us know right away. We will investigate all legitimate reports and do our best to quickly fix the problem.
+Please report security issues privately using [the vulnerability submission form](https://github.com/vllm-project/vllm/security/advisories/new).
 
-Please report security issues privately using [the vulnerability submission form](https://github.com/vllm-project/vllm/security/advisories/new). Reports will then be triaged by the [vulnerability management team](https://docs.vllm.ai/en/latest/contributing/vulnerability_management.html).
+## Issue triage
 
----
+Reports will then be triaged by the [vulnerability management team](https://docs.vllm.ai/en/latest/contributing/vulnerability_management.html).
+
+## Threat model
 
 Please see the [Security Guide in the vLLM documentation](https://docs.vllm.ai/en/latest/usage/security.html) for more information on vLLM's security assumptions and recommendations.
 
 Please see [PyTorch's Security Policy](https://github.com/pytorch/pytorch/blob/main/SECURITY.md) for more information and recommendations on how to securely interact with models.
+
+## Issue severity
+
+We will determine the risk of each issue, taking into account our experience dealing with past issues, versions affected, common defaults, and use cases. We use the following severity categories:
+
+### CRITICAL Severity
+
+Vulnerabilities that allow remote attackers to execute arbitrary code, take full control of the system, or significantly compromise confidentiality, integrity, or availability without any interaction or privileges needed, examples include remote code execution via network, deserialization issues that allow exploit chains. Generally those issues which are rated as CVSS  ≥ 9.0.
+
+### HIGH Severity
+
+Serious security flaws that allow elevated impact—like RCE in specific, limited contexts or significant data loss—but require advanced conditions or some trust, examples include RCE in advanced deployment modes (e.g. multi-node), or high impact issues where some sort of privileged network access is required. These issues typically have CVSS scores between 7.0 and 8.9
+
+### MODERATE Severity
+
+Vulnerabilities that cause denial of service or partial disruption, but do not allow arbitrary code execution or data breach and have limited impact. These issues have a CVSS rating between 4.0 and 6.9
+
+### LOW Severity
+
+Minor issues such as informational disclosures, logging errors, non-exploitable flaws, or weaknesses that require local or high-privilege access and offer negligible impact. Examples include side channel attacks or hash collisions. These issues often have CVSS scores less than 4.0
+
+## Prenotification policy
+
+For certain security issues of CRITICAL, HIGH, or MODERATE severity level, we may prenotify certain organizations or vendors that ship vLLM. The purpose of this prenotification is to allow for a coordinated release of fixes for severe issues.
+
+* This prenotification will be in the form of a private email notification. It may also include adding security contacts to the GitHub security advisory, typically a few days before release.
+
+* If you wish to be added to the prenotification group, please send an email copying all the members of the [vulnerability management team](https://docs.vllm.ai/en/latest/contributing/vulnerability_management.html). Each vendor contact will be analyzed on a case-by-case basis.
+
+* We may withdraw organizations from receiving future prenotifications if they release fixes or any other information about issues before they are public. Group membership may also change based on policy refinements for who may be included.

benchmarks/README.md

@@ -4,7 +4,7 @@ This README guides you through running benchmark tests with the extensive
 datasets supported on vLLM. It’s a living document, updated as new features and datasets
 become available.
 
-**Dataset Overview**
+## Dataset Overview
 
 <table style="width:100%; border-collapse: collapse;">
   <thead>
@@ -22,6 +22,17 @@ become available.
       <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>BurstGPT</strong></td>
       <td style="text-align: center;">✅</td>
@@ -29,7 +40,7 @@ become available.
       <td><code>wget https://github.com/HPMLL/BurstGPT/releases/download/v1.1/BurstGPT_without_fails_2.csv</code></td>
     </tr>
     <tr>
-      <td><strong>Sonnet</strong></td>
+      <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>
@@ -40,6 +51,12 @@ become available.
       <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>
@@ -81,16 +98,17 @@ become available.
 
 **Note**: HuggingFace dataset's `dataset-name` should be set to `hf`
 
----
+## 🚀 Example - Online Benchmark
+
 <details>
-<summary><b>🚀 Example - Online Benchmark</b></summary>
+<summary>Show more</summary>
 
 <br/>
 
 First start serving your model
 
 ```bash
-vllm serve NousResearch/Hermes-3-Llama-3.1-8B --disable-log-requests
+vllm serve NousResearch/Hermes-3-Llama-3.1-8B
 ```
 
 Then run the benchmarking script
@@ -109,7 +127,7 @@ vllm bench serve \
 
 If successful, you will see the following output
 
-```
+```text
 ============ Serving Benchmark Result ============
 Successful requests:                     10
 Benchmark duration (s):                  5.78
@@ -133,11 +151,11 @@ P99 ITL (ms):                            8.39
 ==================================================
 ```
 
-**Custom Dataset**
+### Custom Dataset
 
 If the dataset you want to benchmark is not supported yet in vLLM, even then you can benchmark on it using `CustomDataset`. Your data needs to be in `.jsonl` format and needs to have "prompt" field per entry, e.g., data.jsonl
 
-```
+```json
 {"prompt": "What is the capital of India?"}
 {"prompt": "What is the capital of Iran?"}
 {"prompt": "What is the capital of China?"}
@@ -145,7 +163,7 @@ If the dataset you want to benchmark is not supported yet in vLLM, even then you
 
 ```bash
 # start server
-VLLM_USE_V1=1 vllm serve meta-llama/Llama-3.1-8B-Instruct --disable-log-requests
+VLLM_USE_V1=1 vllm serve meta-llama/Llama-3.1-8B-Instruct
 ```
 
 ```bash
@@ -166,11 +184,11 @@ vllm bench serve --port 9001 --save-result --save-detailed \
 
 You can skip applying chat template if your data already has it by using `--custom-skip-chat-template`.
 
-**VisionArena Benchmark for Vision Language Models**
+### VisionArena Benchmark for Vision Language Models
 
 ```bash
 # need a model with vision capability here
-vllm serve Qwen/Qwen2-VL-7B-Instruct --disable-log-requests
+vllm serve Qwen/Qwen2-VL-7B-Instruct
 ```
 
 ```bash
@@ -184,7 +202,7 @@ vllm bench serve \
   --num-prompts 1000
 ```
 
-**InstructCoder Benchmark with Speculative Decoding**
+### InstructCoder Benchmark with Speculative Decoding
 
 ``` bash
 VLLM_USE_V1=1 vllm serve meta-llama/Meta-Llama-3-8B-Instruct \
@@ -201,13 +219,13 @@ vllm bench serve \
     --num-prompts 2048
 ```
 
-**Other HuggingFaceDataset Examples**
+### Other HuggingFaceDataset Examples
 
 ```bash
-vllm serve Qwen/Qwen2-VL-7B-Instruct --disable-log-requests
+vllm serve Qwen/Qwen2-VL-7B-Instruct
 ```
 
-**`lmms-lab/LLaVA-OneVision-Data`**
+`lmms-lab/LLaVA-OneVision-Data`:
 
 ```bash
 vllm bench serve \
@@ -221,7 +239,7 @@ vllm bench serve \
   --num-prompts 10
 ```
 
-**`Aeala/ShareGPT_Vicuna_unfiltered`**
+`Aeala/ShareGPT_Vicuna_unfiltered`:
 
 ```bash
 vllm bench serve \
@@ -234,7 +252,7 @@ vllm bench serve \
   --num-prompts 10
 ```
 
-**`AI-MO/aimo-validation-aime`**
+`AI-MO/aimo-validation-aime`:
 
 ``` bash
 vllm bench serve \
@@ -245,7 +263,7 @@ vllm bench serve \
     --seed 42
 ```
 
-**`philschmid/mt-bench`**
+`philschmid/mt-bench`:
 
 ``` bash
 vllm bench serve \
@@ -255,7 +273,7 @@ vllm bench serve \
     --num-prompts 80
 ```
 
-**Running With Sampling Parameters**
+### Running With Sampling Parameters
 
 When using OpenAI-compatible backends such as `vllm`, optional sampling
 parameters can be specified. Example client command:
@@ -273,25 +291,29 @@ vllm bench serve \
   --num-prompts 10
 ```
 
-**Running With Ramp-Up Request Rate**
+### 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><b>📈 Example - Offline Throughput Benchmark</b></summary>
+<summary>Show more</summary>
 
 <br/>
 
@@ -305,15 +327,15 @@ vllm bench throughput \
 
 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**
+### VisionArena Benchmark for Vision Language Models
 
-``` bash
+```bash
 vllm bench throughput \
   --model Qwen/Qwen2-VL-7B-Instruct \
   --backend vllm-chat \
@@ -325,13 +347,13 @@ vllm bench throughput \
 
 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**
+### InstructCoder Benchmark with Speculative Decoding
 
 ``` bash
 VLLM_WORKER_MULTIPROC_METHOD=spawn \
@@ -349,15 +371,15 @@ vllm bench throughput \
     "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**
+### Other HuggingFaceDataset Examples
 
-**`lmms-lab/LLaVA-OneVision-Data`**
+`lmms-lab/LLaVA-OneVision-Data`:
 
 ```bash
 vllm bench throughput \
@@ -370,7 +392,7 @@ vllm bench throughput \
   --num-prompts 10
 ```
 
-**`Aeala/ShareGPT_Vicuna_unfiltered`**
+`Aeala/ShareGPT_Vicuna_unfiltered`:
 
 ```bash
 vllm bench throughput \
@@ -382,7 +404,7 @@ vllm bench throughput \
   --num-prompts 10
 ```
 
-**`AI-MO/aimo-validation-aime`**
+`AI-MO/aimo-validation-aime`:
 
 ```bash
 vllm bench throughput \
@@ -394,7 +416,7 @@ vllm bench throughput \
   --num-prompts 10
 ```
 
-**Benchmark with LoRA Adapters**
+Benchmark with LoRA adapters:
 
 ``` bash
 # download dataset
@@ -413,20 +435,22 @@ vllm bench throughput \
 
 </details>
 
+## 🛠️ Example - Structured Output Benchmark
+
 <details>
-<summary><b>🛠️ Example - Structured Output Benchmark</b></summary>
+<summary>Show more</summary>
 
 <br/>
 
 Benchmark the performance of structured output generation (JSON, grammar, regex).
 
-**Server Setup**
+### Server Setup
 
 ```bash
-vllm serve NousResearch/Hermes-3-Llama-3.1-8B --disable-log-requests
+vllm serve NousResearch/Hermes-3-Llama-3.1-8B
 ```
 
-**JSON Schema Benchmark**
+### JSON Schema Benchmark
 
 ```bash
 python3 benchmarks/benchmark_serving_structured_output.py \
@@ -438,7 +462,7 @@ python3 benchmarks/benchmark_serving_structured_output.py \
   --num-prompts 1000
 ```
 
-**Grammar-based Generation Benchmark**
+### Grammar-based Generation Benchmark
 
 ```bash
 python3 benchmarks/benchmark_serving_structured_output.py \
@@ -450,7 +474,7 @@ python3 benchmarks/benchmark_serving_structured_output.py \
   --num-prompts 1000
 ```
 
-**Regex-based Generation Benchmark**
+### Regex-based Generation Benchmark
 
 ```bash
 python3 benchmarks/benchmark_serving_structured_output.py \
@@ -461,7 +485,7 @@ python3 benchmarks/benchmark_serving_structured_output.py \
   --num-prompts 1000
 ```
 
-**Choice-based Generation Benchmark**
+### Choice-based Generation Benchmark
 
 ```bash
 python3 benchmarks/benchmark_serving_structured_output.py \
@@ -472,7 +496,7 @@ python3 benchmarks/benchmark_serving_structured_output.py \
   --num-prompts 1000
 ```
 
-**XGrammar Benchmark Dataset**
+### XGrammar Benchmark Dataset
 
 ```bash
 python3 benchmarks/benchmark_serving_structured_output.py \
@@ -485,14 +509,16 @@ python3 benchmarks/benchmark_serving_structured_output.py \
 
 </details>
 
+## 📚 Example - Long Document QA Benchmark
+
 <details>
-<summary><b>📚 Example - Long Document QA Benchmark</b></summary>
+<summary>Show more</summary>
 
 <br/>
 
 Benchmark the performance of long document question-answering with prefix caching.
 
-**Basic Long Document QA Test**
+### Basic Long Document QA Test
 
 ```bash
 python3 benchmarks/benchmark_long_document_qa_throughput.py \
@@ -504,7 +530,7 @@ python3 benchmarks/benchmark_long_document_qa_throughput.py \
   --repeat-count 5
 ```
 
-**Different Repeat Modes**
+### Different Repeat Modes
 
 ```bash
 # Random mode (default) - shuffle prompts randomly
@@ -537,14 +563,16 @@ python3 benchmarks/benchmark_long_document_qa_throughput.py \
 
 </details>
 
+## 🗂️ Example - Prefix Caching Benchmark
+
 <details>
-<summary><b>🗂️ Example - Prefix Caching Benchmark</b></summary>
+<summary>Show more</summary>
 
 <br/>
 
 Benchmark the efficiency of automatic prefix caching.
 
-**Fixed Prompt with Prefix Caching**
+### Fixed Prompt with Prefix Caching
 
 ```bash
 python3 benchmarks/benchmark_prefix_caching.py \
@@ -555,7 +583,7 @@ python3 benchmarks/benchmark_prefix_caching.py \
   --input-length-range 128:256
 ```
 
-**ShareGPT Dataset with Prefix Caching**
+### ShareGPT Dataset with Prefix Caching
 
 ```bash
 # download dataset
@@ -570,16 +598,32 @@ python3 benchmarks/benchmark_prefix_caching.py \
   --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><b>⚡ Example - Request Prioritization Benchmark</b></summary>
+<summary>Show more</summary>
 
 <br/>
 
 Benchmark the performance of request prioritization in vLLM.
 
-**Basic Prioritization Test**
+### Basic Prioritization Test
 
 ```bash
 python3 benchmarks/benchmark_prioritization.py \
@@ -590,7 +634,7 @@ python3 benchmarks/benchmark_prioritization.py \
   --scheduling-policy priority
 ```
 
-**Multiple Sequences per Prompt**
+### Multiple Sequences per Prompt
 
 ```bash
 python3 benchmarks/benchmark_prioritization.py \
@@ -603,3 +647,41 @@ python3 benchmarks/benchmark_prioritization.py \
 ```
 
 </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
+```
+
+</details>

benchmarks/auto_tune/README.md

@@ -3,6 +3,7 @@
 This script automates the process of finding the optimal server parameter combination (`max-num-seqs` and `max-num-batched-tokens`) to maximize throughput for a vLLM server. It also supports additional constraints such as E2E latency and prefix cache hit rate.
 
 ## Table of Contents
+
 - [Prerequisites](#prerequisites)
 - [Configuration](#configuration)
 - [How to Run](#how-to-run)
@@ -52,7 +53,7 @@ You must set the following variables at the top of the script before execution.
 1. **Configure**: Edit the script and set the variables in the [Configuration](#configuration) section.
 2. **Execute**: Run the script. Since the process can take a long time, it is highly recommended to use a terminal multiplexer like `tmux` or `screen` to prevent the script from stopping if your connection is lost.
 
-```
+```bash
 cd <FOLDER_OF_THIS_SCRIPT>
 bash auto_tune.sh
 ```
@@ -64,6 +65,7 @@ bash auto_tune.sh
 Here are a few examples of how to configure the script for different goals:
 
 ### 1. Maximize Throughput (No Latency Constraint)
+
 - **Goal**: Find the best `max-num-seqs` and `max-num-batched-tokens` to get the highest possible throughput for 1800 input tokens and 20 output tokens.
 - **Configuration**:
 
@@ -76,6 +78,7 @@ MAX_LATENCY_ALLOWED_MS=100000000000 # A very large number
 ```
 
 #### 2. Maximize Throughput with a Latency Requirement
+
 - **Goal**: Find the best server parameters when P99 end-to-end latency must be below 500ms.
 - **Configuration**:
 
@@ -88,6 +91,7 @@ MAX_LATENCY_ALLOWED_MS=500
 ```
 
 #### 3. Maximize Throughput with Prefix Caching and Latency Requirements
+
 - **Goal**: Find the best server parameters assuming a 60% prefix cache hit rate and a latency requirement of 500ms.
 - **Configuration**:
 
@@ -109,7 +113,7 @@ After the script finishes, you will find the results in a new, timestamped direc
 
 - **Final Result Summary**: A file named `result.txt` is created in the log directory. It contains a summary of each tested combination and concludes with the overall best parameters found.
 
-```
+```text
 # Example result.txt content
 hash:a1b2c3d4...
 max_num_seqs: 128, max_num_batched_tokens: 2048, request_rate: 10.0, e2el: 450.5, throughput: 9.8, goodput: 9.8

benchmarks/auto_tune/auto_tune.sh

@@ -49,6 +49,7 @@ best_throughput=0
 best_max_num_seqs=0
 best_num_batched_tokens=0
 best_goodput=0
+best_request_rate=0
 
 start_server() {
     local gpu_memory_utilization=$1
@@ -57,19 +58,35 @@ start_server() {
     local vllm_log=$4
     local profile_dir=$5
 
-    pkill -f vllm
+    pkill -if vllm
 
-    VLLM_USE_V1=1 VLLM_SERVER_DEV_MODE=1 VLLM_TORCH_PROFILER_DIR=$profile_dir vllm serve $MODEL \
-        --disable-log-requests \
-        --port 8004 \
-        --gpu-memory-utilization $gpu_memory_utilization \
-        --max-num-seqs $max_num_seqs \
-        --max-num-batched-tokens $max_num_batched_tokens \
-        --tensor-parallel-size $TP \
-        --enable-prefix-caching \
-        --load-format dummy \
-        --download-dir "$DOWNLOAD_DIR" \
-        --max-model-len $MAX_MODEL_LEN > "$vllm_log" 2>&1 &
+    # Define the common arguments as a bash array.
+    # Each argument and its value are separate elements.
+    local common_args_array=(
+        "$MODEL"
+        "--disable-log-requests"
+        "--port" "8004"
+        "--gpu-memory-utilization" "$gpu_memory_utilization"
+        "--max-num-seqs" "$max_num_seqs"
+        "--max-num-batched-tokens" "$max_num_batched_tokens"
+        "--tensor-parallel-size" "$TP"
+        "--enable-prefix-caching"
+        "--load-format" "dummy"
+        "--download-dir" "$DOWNLOAD_DIR"
+        "--max-model-len" "$MAX_MODEL_LEN"
+    )
+
+    # Use the array expansion "${common_args_array[@]}"
+    # This correctly passes each element as a separate argument.
+    if [[ -n "$profile_dir" ]]; then
+        # Start server with profiling enabled
+        VLLM_USE_V1=1 VLLM_SERVER_DEV_MODE=1 VLLM_TORCH_PROFILER_DIR=$profile_dir \
+            vllm serve "${common_args_array[@]}" > "$vllm_log" 2>&1 &
+    else
+        # Start server without profiling
+        VLLM_USE_V1=1 VLLM_SERVER_DEV_MODE=1 \
+            vllm serve "${common_args_array[@]}" > "$vllm_log" 2>&1 &
+    fi
 
     # wait for 10 minutes...
     server_started=0
@@ -83,6 +100,7 @@ start_server() {
             sleep 10
         fi
     done
+
     if (( ! server_started )); then
         echo "server did not start within 10 minutes. Please check server log at $vllm_log".
         return 1
@@ -91,37 +109,20 @@ start_server() {
     fi
 }
 
-update_best_profile() {
-    local profile_dir=$1
-    local profile_index=$2
-    sorted_paths=($(find "$profile_dir" -maxdepth 1 -not -path "$profile_dir" | sort))
-    selected_profile_file=
-    if [[ "$SYSTEM" == "TPU" ]]; then
-        selected_profile_file="${sorted_paths[$profile_index]}/*.xplane.pb"
-    fi
-    if [[ "$SYSTEM" == "GPU" ]]; then
-        selected_profile_file="${sorted_paths[$profile_index]}"
-    fi
-    rm -f $PROFILE_PATH/*
-    cp $selected_profile_file $PROFILE_PATH
-}
-
 run_benchmark() {
     local max_num_seqs=$1
     local max_num_batched_tokens=$2
     local gpu_memory_utilization=$3
     echo "max_num_seq: $max_num_seqs, max_num_batched_tokens: $max_num_batched_tokens"
     local vllm_log="$LOG_FOLDER/vllm_log_${max_num_seqs}_${max_num_batched_tokens}.txt"
-    local profile_dir="$LOG_FOLDER/profile_${max_num_seqs}_${max_num_batched_tokens}"
     echo "vllm_log: $vllm_log"
     echo
     rm -f $vllm_log
-    mkdir -p $profile_dir
-    pkill -f vllm
-    local profile_index=0
+    pkill -if vllm
 
     echo "starting server..."
-    start_server $gpu_memory_utilization $max_num_seqs $max_num_batched_tokens $vllm_log $profile_dir
+    # Call start_server without a profile_dir to avoid profiling overhead
+    start_server $gpu_memory_utilization $max_num_seqs $max_num_batched_tokens $vllm_log ""
     result=$?
     if [[ "$result" -eq 1 ]]; then
         echo "server failed to start. gpu_memory_utilization:$gpu_memory_utilization, max_num_seqs:$max_num_seqs, max_num_batched_tokens: $max_num_batched_tokens"
@@ -135,7 +136,8 @@ run_benchmark() {
     # get a basic qps by using request-rate inf
     bm_log="$LOG_FOLDER/bm_log_${max_num_seqs}_${max_num_batched_tokens}_requestrate_inf.txt"
     prefix_len=$(( INPUT_LEN * MIN_CACHE_HIT_PCT / 100 ))
-adjusted_input_len=$(( INPUT_LEN - prefix_len ))
+    adjusted_input_len=$(( INPUT_LEN - prefix_len ))
+    # --profile flag is removed from this call
     vllm bench serve \
         --backend vllm \
         --model $MODEL  \
@@ -149,8 +151,7 @@ adjusted_input_len=$(( INPUT_LEN - prefix_len ))
         --goodput e2el:$MAX_LATENCY_ALLOWED_MS \
         --num-prompts 1000 \
         --random-prefix-len $prefix_len \
-        --port 8004 \
-        --profile &> "$bm_log"
+        --port 8004 &> "$bm_log"
     throughput=$(grep "Request throughput (req/s):" "$bm_log" | sed 's/[^0-9.]//g')
     e2el=$(grep "P99 E2EL (ms):" "$bm_log" | awk '{print $NF}')
     goodput=$(grep "Request goodput (req/s):" "$bm_log" | sed 's/[^0-9.]//g')
@@ -164,7 +165,6 @@ adjusted_input_len=$(( INPUT_LEN - prefix_len ))
     # start from request-rate as int(throughput) + 1
         request_rate=$((${throughput%.*} + 1))
         while ((request_rate > 0)); do
-            profile_index=$((profile_index+1))
             # clear prefix cache
             curl -X POST http://0.0.0.0:8004/reset_prefix_cache
             sleep 5
@@ -202,12 +202,7 @@ adjusted_input_len=$(( INPUT_LEN - prefix_len ))
             best_max_num_seqs=$max_num_seqs
             best_num_batched_tokens=$max_num_batched_tokens
             best_goodput=$goodput
-            if [[ "$SYSTEM" == "TPU" ]]; then
-                update_best_profile "$profile_dir/plugins/profile" $profile_index
-            fi
-            if [[ "$SYSTEM" == "GPU" ]]; then
-                update_best_profile "$profile_dir" $profile_index
-            fi
+            best_request_rate=$request_rate
         fi
     else
         echo "max_num_seqs: $max_num_seqs, max_num_batched_tokens: $max_num_batched_tokens does not meet latency requirement ${MAX_LATENCY_ALLOWED_MS}"
@@ -216,7 +211,7 @@ adjusted_input_len=$(( INPUT_LEN - prefix_len ))
 
     echo "best_max_num_seqs: $best_max_num_seqs, best_num_batched_tokens: $best_num_batched_tokens, best_throughput: $best_throughput"
 
-    pkill vllm
+    pkill -if vllm
     sleep 10
     printf '=%.0s' $(seq 1 20)
     return 0
@@ -229,7 +224,8 @@ read -r -a num_batched_tokens_list <<< "$NUM_BATCHED_TOKENS_LIST"
 gpu_memory_utilization=0.98
 find_gpu_memory_utilization=0
 while (( $(echo "$gpu_memory_utilization >= 0.9" | bc -l) )); do
-    start_server $gpu_memory_utilization "${num_seqs_list[-1]}" "${num_batched_tokens_list[-1]}" "$LOG_FOLDER/vllm_log_gpu_memory_utilization_$gpu_memory_utilization.log"
+    # Pass empty string for profile_dir argument
+    start_server $gpu_memory_utilization "${num_seqs_list[-1]}" "${num_batched_tokens_list[-1]}" "$LOG_FOLDER/vllm_log_gpu_memory_utilization_$gpu_memory_utilization.log" ""
     result=$?
     if [[ "$result" -eq 0 ]]; then
         find_gpu_memory_utilization=1
@@ -252,5 +248,45 @@ for num_seqs in "${num_seqs_list[@]}"; do
     done
 done
 echo "finish permutations"
+
+# =================================================================================
+# FINAL PROFILING RUN FOR THE BEST CONFIGURATION
+# =================================================================================
+if (( $(echo "$best_throughput > 0" | bc -l) )); then
+    echo
+    echo "Benchmark tuning finished. Now running profiling on the best configuration found..."
+    echo "Best config: max_num_seqs: $best_max_num_seqs, max_num_batched_tokens: $best_num_batched_tokens, throughput: $best_throughput"
+    echo
+
+    vllm_log="$LOG_FOLDER/vllm_log_BEST_PROFILE.txt"
+    bm_log="$LOG_FOLDER/bm_log_BEST_PROFILE.txt"
+
+    # Start server with the best params and profiling ENABLED
+    echo "Starting server for profiling..."
+    start_server $gpu_memory_utilization $best_max_num_seqs $best_num_batched_tokens "$vllm_log" "$PROFILE_PATH"
+
+    # Run benchmark with the best params and the --profile flag
+    echo "Running benchmark with profiling..."
+    prefix_len=$(( INPUT_LEN * MIN_CACHE_HIT_PCT / 100 ))
+    adjusted_input_len=$(( INPUT_LEN - prefix_len ))
+    vllm bench serve \
+        --backend vllm \
+        --model $MODEL \
+        --dataset-name random \
+        --random-input-len $adjusted_input_len \
+        --random-output-len $OUTPUT_LEN \
+        --ignore-eos \
+        --disable-tqdm \
+        --request-rate $best_request_rate \
+        --percentile-metrics ttft,tpot,itl,e2el \
+        --goodput e2el:$MAX_LATENCY_ALLOWED_MS \
+        --num-prompts 100 \
+        --random-prefix-len $prefix_len \
+        --port 8004 \
+        --profile &> "$bm_log"
+else
+    echo "No configuration met the latency requirements. Skipping final profiling run."
+fi
+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/backend_request_func.py

@@ -31,7 +31,7 @@ class RequestFuncInput:
     model_name: Optional[str] = None
     logprobs: Optional[int] = None
     extra_body: Optional[dict] = None
-    multi_modal_content: Optional[dict] = None
+    multi_modal_content: Optional[dict | list[dict]] = None
     ignore_eos: bool = False
     language: Optional[str] = None
 
@@ -364,7 +364,15 @@ async def async_request_openai_chat_completions(
     ) as session:
         content = [{"type": "text", "text": request_func_input.prompt}]
         if request_func_input.multi_modal_content:
-            content.append(request_func_input.multi_modal_content)
+            mm_content = request_func_input.multi_modal_content
+            if isinstance(mm_content, list):
+                content.extend(mm_content)
+            elif isinstance(mm_content, dict):
+                content.append(mm_content)
+            else:
+                raise TypeError(
+                    "multi_modal_content must be a dict or list[dict] for openai-chat"
+                )
         payload = {
             "model": request_func_input.model_name
             if request_func_input.model_name
@@ -491,7 +499,10 @@ async def async_request_openai_audio(
             buffer.seek(0)
             return buffer
 
-        with to_bytes(*request_func_input.multi_modal_content["audio"]) as f:
+        mm_audio = request_func_input.multi_modal_content
+        if not isinstance(mm_audio, dict) or "audio" not in mm_audio:
+            raise TypeError("multi_modal_content must be a dict containing 'audio'")
+        with to_bytes(*mm_audio["audio"]) as f:
             form = aiohttp.FormData()
             form.add_field("file", f, content_type="audio/wav")
             for key, value in payload.items():

benchmarks/benchmark_block_pool.py

@@ -0,0 +1,74 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+import gc
+
+from tabulate import tabulate
+
+from benchmark_utils import TimeCollector
+from vllm.utils import FlexibleArgumentParser
+from vllm.v1.core.block_pool import BlockPool
+
+
+def main(args):
+    rows = []
+    for allocate_block in args.allocate_blocks:
+        # Enforce a GC collect ahead to minimize the impact among runs
+        gc.collect()
+        block_pool = BlockPool(num_gpu_blocks=args.num_gpu_blocks, enable_caching=True)
+
+        get_blocks_times = TimeCollector(TimeCollector.US)
+        free_blocks_times = TimeCollector(TimeCollector.US)
+        for _ in range(args.num_iteration):
+            with get_blocks_times:
+                blocks = block_pool.get_new_blocks(allocate_block)
+            with free_blocks_times:
+                block_pool.free_blocks(blocks)
+
+        rows.append(
+            [get_blocks_times.cnt, args.num_gpu_blocks, allocate_block]
+            + get_blocks_times.dump_avg_max()
+            + free_blocks_times.dump_avg_max()
+        )
+
+    print(
+        tabulate(
+            rows,
+            headers=[
+                "Iterations",
+                "Total\nBlocks",
+                "Allocated\nBlocks",
+                "Get Blocks\nAvg (us)",
+                "Get Blocks\nMax (us)",
+                "Free Blocks\nAvg (us)",
+                "Free Blocks\nMax (us)",
+            ],
+            tablefmt="grid",
+            floatfmt=".3f",
+        )
+    )
+
+
+def invoke_main() -> None:
+    parser = FlexibleArgumentParser(
+        description="Benchmark the performance of BlockPool for KV Cache."
+    )
+    parser.add_argument("--num-gpu-blocks", type=int, default=100000)
+    parser.add_argument(
+        "--num-iteration",
+        type=int,
+        default=1000,
+        help="Number of iterations to run to stablize final data readings",
+    )
+    parser.add_argument(
+        "--allocate-blocks",
+        type=int,
+        nargs="*",
+        default=[10, 50, 100, 500, 1000],
+        help="Number of blocks to allocate",
+    )
+    args = parser.parse_args()
+    main(args)
+
+
+if __name__ == "__main__":
+    invoke_main()  # pragma: no cover

benchmarks/benchmark_dataset.py

@@ -52,7 +52,7 @@ class SampleRequest:
     prompt: Union[str, Any]
     prompt_len: int
     expected_output_len: int
-    multi_modal_data: Optional[Union[MultiModalDataDict, dict]] = None
+    multi_modal_data: Optional[Union[MultiModalDataDict, dict, list[dict]]] = None
     lora_request: Optional[LoRARequest] = None
 
 
@@ -430,14 +430,20 @@ class ShareGPTDataset(BenchmarkDataset):
                 skip_min_output_len_check=output_len is not None,
             ):
                 continue
+            # TODO: Also support ShareGPT4Video.
+            if image_path := entry.get("image"):
+                mm_content = process_image(image_path)
+            else:
+                mm_content = None
             if enable_multimodal_chat:
-                prompt = self.apply_multimodal_chat_transformation(prompt, None)
+                prompt = self.apply_multimodal_chat_transformation(prompt, mm_content)
             samples.append(
                 SampleRequest(
                     prompt=prompt,
                     prompt_len=prompt_len,
                     expected_output_len=new_output_len,
                     lora_request=lora_request,
+                    multi_modal_data=mm_content,
                 )
             )
         self.maybe_oversample_requests(samples, num_requests)

benchmarks/benchmark_ngram_proposer.py

@@ -0,0 +1,112 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+import gc
+
+import numpy as np
+from tabulate import tabulate
+
+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
+
+
+def main(args):
+    rows = []
+    for max_ngram in args.max_ngram:
+        collector = TimeCollector(TimeCollector.US)
+
+        model_config = ModelConfig(
+            model="facebook/opt-125m",
+            task="generate",
+            max_model_len=args.num_token + args.num_spec_token,
+            tokenizer="facebook/opt-125m",
+            tokenizer_mode="auto",
+            dtype="auto",
+            seed=None,
+            trust_remote_code=False,
+        )
+        proposer = NgramProposer(
+            vllm_config=VllmConfig(
+                model_config=model_config,
+                speculative_config=SpeculativeConfig(
+                    prompt_lookup_min=args.min_ngram,
+                    prompt_lookup_max=max_ngram,
+                    num_speculative_tokens=args.num_spec_token,
+                    method="ngram",
+                ),
+            )
+        )
+
+        # Warm up
+        proposer.propose(np.random.randint(0, 20, (args.num_token,)))
+
+        gc.collect()
+        for _ in range(args.num_iteration):
+            tokens = np.random.randint(0, 20, (args.num_req, args.num_token))
+            with collector:
+                for i in range(args.num_req):
+                    proposer.propose(tokens[i, :])
+        rows.append(
+            [args.num_req, args.num_token, args.min_ngram, max_ngram]
+            + collector.dump_avg_max()
+        )
+
+    print(
+        tabulate(
+            rows,
+            headers=[
+                "# Request",
+                "# Token",
+                "Min Ngram",
+                "Max Ngram",
+                "Avg (us)",
+                "Max (us)",
+            ],
+            tablefmt="grid",
+            floatfmt=".3f",
+        )
+    )
+
+
+def invoke_main() -> None:
+    parser = FlexibleArgumentParser(
+        description="Benchmark the performance of N-gram speculative decode drafting"
+    )
+    parser.add_argument(
+        "--num-iteration",
+        type=int,
+        default=100,
+        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"
+    )
+    parser.add_argument(
+        "--num-token", type=int, default=1500, help="Number of tokens for each request"
+    )
+    parser.add_argument(
+        "--min-ngram",
+        type=int,
+        default=3,
+        help="Minimum n-gram to match",
+    )
+    parser.add_argument(
+        "--max-ngram",
+        type=int,
+        nargs="*",
+        default=[5, 7, 10, 15, 20],
+        help="Maximum n-gram to match",
+    )
+    parser.add_argument(
+        "--num-spec-token",
+        type=int,
+        default=3,
+        help="Number of speculative tokens to generate",
+    )
+    args = parser.parse_args()
+    main(args)
+
+
+if __name__ == "__main__":
+    invoke_main()  # pragma: no cover

benchmarks/benchmark_serving.py

@@ -5,8 +5,7 @@ r"""Benchmark online serving throughput.
 On the server side, run one of the following commands:
     vLLM OpenAI API server
     vllm serve <your_model> \
-        --swap-space 16 \
-        --disable-log-requests
+        --swap-space 16
 
 On the client side, run:
     python benchmarks/benchmark_serving.py \
@@ -264,7 +263,14 @@ async def benchmark(
         input_requests[0].multi_modal_data,
     )
 
-    assert test_mm_content is None or isinstance(test_mm_content, dict)
+    assert (
+        test_mm_content is None
+        or isinstance(test_mm_content, dict)
+        or (
+            isinstance(test_mm_content, list)
+            and all(isinstance(item, dict) for item in test_mm_content)
+        )
+    ), "multi_modal_data must be a dict or list[dict]"
     test_input = RequestFuncInput(
         model=model_id,
         model_name=model_name,
@@ -396,20 +402,6 @@ async def benchmark(
         tasks.append(asyncio.create_task(task))
     outputs: list[RequestFuncOutput] = await asyncio.gather(*tasks)
 
-    if profile:
-        print("Stopping profiler...")
-        profile_input = RequestFuncInput(
-            model=model_id,
-            prompt=test_prompt,
-            api_url=base_url + "/stop_profile",
-            prompt_len=test_prompt_len,
-            output_len=test_output_len,
-            logprobs=logprobs,
-        )
-        profile_output = await request_func(request_func_input=profile_input)
-        if profile_output.success:
-            print("Profiler stopped")
-
     if pbar is not None:
         pbar.close()
 
@@ -427,6 +419,10 @@ async def benchmark(
 
     print("{s:{c}^{n}}".format(s=" Serving Benchmark Result ", n=50, c="="))
     print("{:<40} {:<10}".format("Successful requests:", metrics.completed))
+    if max_concurrency is not None:
+        print("{:<40} {:<10}".format("Maximum request concurrency:", max_concurrency))
+    if request_rate != float("inf"):
+        print("{:<40} {:<10.2f}".format("Request rate configured (RPS):", request_rate))
     print("{:<40} {:<10.2f}".format("Benchmark duration (s):", benchmark_duration))
     print("{:<40} {:<10}".format("Total input tokens:", metrics.total_input))
     print("{:<40} {:<10}".format("Total generated tokens:", metrics.total_output))
@@ -518,6 +514,20 @@ async def benchmark(
 
     print("=" * 50)
 
+    if profile:
+        print("Stopping profiler...")
+        profile_input = RequestFuncInput(
+            model=model_id,
+            prompt=test_prompt,
+            api_url=base_url + "/stop_profile",
+            prompt_len=test_prompt_len,
+            output_len=test_output_len,
+            logprobs=logprobs,
+        )
+        profile_output = await request_func(request_func_input=profile_input)
+        if profile_output.success:
+            print("Profiler stopped")
+
     return result
 
 

benchmarks/benchmark_serving_structured_output.py

@@ -4,7 +4,7 @@ r"""Benchmark online serving throughput with structured outputs.
 
 On the server side, run one of the following commands:
     (vLLM OpenAI API server)
-    vllm serve <your_model> --disable-log-requests
+    vllm serve <your_model>
 
 On the client side, run:
     python benchmarks/benchmark_serving_structured_output.py \
@@ -538,20 +538,6 @@ async def benchmark(
         )
     outputs: list[RequestFuncOutput] = await asyncio.gather(*tasks)
 
-    if profile:
-        print("Stopping profiler...")
-        profile_input = RequestFuncInput(
-            model=model_id,
-            prompt=test_request.prompt,
-            api_url=base_url + "/stop_profile",
-            prompt_len=test_request.prompt_len,
-            output_len=test_request.expected_output_len,
-            extra_body={test_request.structure_type: test_request.schema},
-        )
-        profile_output = await request_func(request_func_input=profile_input)
-        if profile_output.success:
-            print("Profiler stopped")
-
     if pbar is not None:
         pbar.close()
 
@@ -569,6 +555,10 @@ async def benchmark(
 
     print("{s:{c}^{n}}".format(s=" Serving Benchmark Result ", n=50, c="="))
     print("{:<40} {:<10}".format("Successful requests:", metrics.completed))
+    if max_concurrency is not None:
+        print("{:<40} {:<10}".format("Maximum request concurrency:", max_concurrency))
+    if request_rate != float("inf"):
+        print("{:<40} {:<10.2f}".format("Request rate configured (RPS):", request_rate))
     print("{:<40} {:<10.2f}".format("Benchmark duration (s):", benchmark_duration))
     print("{:<40} {:<10}".format("Total input tokens:", metrics.total_input))
     print("{:<40} {:<10}".format("Total generated tokens:", metrics.total_output))
@@ -666,6 +656,20 @@ async def benchmark(
 
     print("=" * 50)
 
+    if profile:
+        print("Stopping profiler...")
+        profile_input = RequestFuncInput(
+            model=model_id,
+            prompt=test_request.prompt,
+            api_url=base_url + "/stop_profile",
+            prompt_len=test_request.prompt_len,
+            output_len=test_request.expected_output_len,
+            extra_body={test_request.structure_type: test_request.schema},
+        )
+        profile_output = await request_func(request_func_input=profile_input)
+        if profile_output.success:
+            print("Profiler stopped")
+
     return result, ret
 
 

benchmarks/benchmark_utils.py

@@ -1,11 +1,12 @@
 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-
 import argparse
 import json
 import math
 import os
-from typing import Any
+import time
+from types import TracebackType
+from typing import Any, Optional, Union
 
 
 def convert_to_pytorch_benchmark_format(
@@ -72,3 +73,53 @@ def write_to_json(filename: str, records: list) -> None:
             cls=InfEncoder,
             default=lambda o: f"<{type(o).__name__} object is not JSON serializable>",
         )
+
+
+# Collect time and generate time metrics
+#
+# Example Usage:
+#   collector = TimeCollector(TimeCollector.US)
+#   for _ in range(total_iteration):
+#      with collector:
+#          ...
+#   collector.dump_avg_max()
+class TimeCollector:
+    NS: int = 1
+    US: int = NS * 1000
+    MS: int = US * 1000
+    S: int = MS * 1000
+
+    def __init__(self, scale: int) -> None:
+        self.cnt: int = 0
+        self._sum: int = 0
+        self._max: Optional[int] = None
+        self.scale = scale
+        self.start_time: int = time.monotonic_ns()
+
+    def collect(self, v: int) -> None:
+        self.cnt += 1
+        self._sum += v
+        if self._max is None:
+            self._max = v
+        else:
+            self._max = max(self._max, v)
+
+    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) -> Union[float, str]:
+        return self._max / self.scale if self._max else "N/A"
+
+    def dump_avg_max(self) -> list[Union[float, str]]:
+        return [self.avg(), self.max()]
+
+    def __enter__(self) -> None:
+        self.start_time = time.monotonic_ns()
+
+    def __exit__(
+        self,
+        exc_type: Optional[type[BaseException]],
+        exc_value: Optional[BaseException],
+        exc_traceback: Optional[TracebackType],
+    ) -> None:
+        self.collect(time.monotonic_ns() - self.start_time)

benchmarks/disagg_benchmarks/disagg_prefill_proxy_server.py

@@ -1,63 +1,199 @@
 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
+import argparse
+import asyncio
+import logging
 import os
 
 import aiohttp
-from quart import Quart, make_response, request
+from quart import Quart, Response, make_response, request
+from rate_limiter import RateLimiter
+from request_queue import RequestQueue
 
-AIOHTTP_TIMEOUT = aiohttp.ClientTimeout(total=6 * 60 * 60)
-
-app = Quart(__name__)
+# Configure logging
+logging.basicConfig(level=logging.INFO)
+logger = logging.getLogger(__name__)
 
 
-async def forward_request(url, data):
-    async with aiohttp.ClientSession(timeout=AIOHTTP_TIMEOUT) as session:
+def parse_args():
+    """parse command line arguments"""
+    parser = argparse.ArgumentParser(description="vLLM P/D disaggregation proxy server")
+
+    # Add args
+    parser.add_argument(
+        "--timeout",
+        type=float,
+        default=300,
+        help="Timeout for backend service requests in seconds (default: 300)",
+    )
+    parser.add_argument(
+        "--max-concurrent",
+        type=int,
+        default=100,
+        help="Maximum concurrent requests to backend services (default: 100)",
+    )
+    parser.add_argument(
+        "--queue-size",
+        type=int,
+        default=500,
+        help="Maximum number of requests in the queue (default: 500)",
+    )
+    parser.add_argument(
+        "--rate-limit",
+        type=int,
+        default=40,
+        help="Maximum requests per second (default: 40)",
+    )
+    parser.add_argument(
+        "--port",
+        type=int,
+        default=8000,
+        help="Port to run the server on (default: 8000)",
+    )
+    parser.add_argument(
+        "--prefill-url",
+        type=str,
+        default="http://localhost:8100/v1/completions",
+        help="Prefill service endpoint URL",
+    )
+    parser.add_argument(
+        "--decode-url",
+        type=str,
+        default="http://localhost:8200/v1/completions",
+        help="Decode service endpoint URL",
+    )
+
+    return parser.parse_args()
+
+
+def main():
+    """parse command line arguments"""
+    args = parse_args()
+
+    # Initialize configuration using command line parameters
+    AIOHTTP_TIMEOUT = aiohttp.ClientTimeout(total=args.timeout)
+    MAX_CONCURRENT_REQUESTS = args.max_concurrent
+    REQUEST_QUEUE_SIZE = args.queue_size
+    RATE_LIMIT = args.rate_limit
+    PREFILL_SERVICE_URL = args.prefill_url
+    DECODE_SERVICE_URL = args.decode_url
+    PORT = args.port
+
+    app = Quart(__name__)
+
+    # Initialize the rate limiter and request queue
+    rate_limiter = RateLimiter(RATE_LIMIT)
+    request_queue = RequestQueue(MAX_CONCURRENT_REQUESTS, REQUEST_QUEUE_SIZE)
+
+    # Attach the configuration object to the application instance
+    app.config.update(
+        {
+            "AIOHTTP_TIMEOUT": AIOHTTP_TIMEOUT,
+            "rate_limiter": rate_limiter,
+            "request_queue": request_queue,
+            "PREFILL_SERVICE_URL": PREFILL_SERVICE_URL,
+            "DECODE_SERVICE_URL": DECODE_SERVICE_URL,
+        }
+    )
+
+    # Start queue processing on app startup
+    @app.before_serving
+    async def startup():
+        """Start request processing task when app starts serving"""
+        asyncio.create_task(request_queue.process())
+
+    async def forward_request(url, data):
+        """Forward request to backend service with rate limiting and error handling"""
         headers = {"Authorization": f"Bearer {os.environ.get('OPENAI_API_KEY')}"}
-        async with session.post(url=url, json=data, headers=headers) as response:
-            if response.status == 200:
-                # if response.headers.get('Transfer-Encoding') == 'chunked':
-                if True:
-                    async for chunk_bytes in response.content.iter_chunked(1024):
-                        yield chunk_bytes
-                else:
-                    content = await response.read()
-                    yield content
 
-
-@app.route("/v1/completions", methods=["POST"])
-async def handle_request():
-    try:
-        original_request_data = await request.get_json()
-
-        prefill_request = original_request_data.copy()
-        # change max_tokens = 1 to let it only do prefill
-        prefill_request["max_tokens"] = 1
-
-        # finish prefill
-        async for _ in forward_request(
-            "http://localhost:8100/v1/completions", prefill_request
+        # Use rate limiter as context manager
+        async with (
+            rate_limiter,
+            aiohttp.ClientSession(timeout=AIOHTTP_TIMEOUT) as session,
         ):
-            continue
+            try:
+                async with session.post(
+                    url=url, json=data, headers=headers
+                ) as response:
+                    if response.status == 200:
+                        # Stream response chunks
+                        async for chunk_bytes in response.content.iter_chunked(1024):
+                            yield chunk_bytes
+                    else:
+                        # Handle backend service errors
+                        error_text = await response.text()
+                        logger.error(
+                            "Backend service error: %s - %s",
+                            response.status,
+                            error_text,
+                        )
+                        yield b'{"error": "Backend service error"}'
+            except aiohttp.ClientError as e:
+                # Handle connection errors
+                logger.error("Connection error to %s: %s", url, str(e))
+                yield b'{"error": "Service unavailable"}'
+            except asyncio.TimeoutError:
+                # Handle timeout errors
+                logger.error("Timeout connecting to %s", url)
+                yield b'{"error": "Service timeout"}'
 
-        # return decode
-        generator = forward_request(
-            "http://localhost:8200/v1/completions", original_request_data
-        )
-        response = await make_response(generator)
-        response.timeout = None
+    async def process_request():
+        """Process a single request through prefill and decode stages"""
+        try:
+            original_request_data = await request.get_json()
 
-        return response
+            # Create prefill request (max_tokens=1)
+            prefill_request = original_request_data.copy()
+            prefill_request["max_tokens"] = 1
 
-    except Exception as e:
-        import sys
-        import traceback
+            # Execute prefill stage
+            async for _ in forward_request(PREFILL_SERVICE_URL, prefill_request):
+                continue
 
-        exc_info = sys.exc_info()
-        print("Error occurred in disagg prefill proxy server")
-        print(e)
-        print("".join(traceback.format_exception(*exc_info)))
+            # Execute decode stage and stream response
+            generator = forward_request(DECODE_SERVICE_URL, original_request_data)
+            response = await make_response(generator)
+            response.timeout = None  # Disable timeout for streaming response
+            return response
+
+        except Exception:
+            logger.exception("Error processing request")
+            return Response(
+                response=b'{"error": "Internal server error"}',
+                status=500,
+                content_type="application/json",
+            )
+
+    @app.route("/v1/completions", methods=["POST"])
+    async def handle_request():
+        """Handle incoming API requests with concurrency and rate limiting"""
+        # Create task for request processing
+        task = asyncio.create_task(process_request())
+
+        # Enqueue request or reject if queue is full
+        if not await request_queue.enqueue(task):
+            return Response(
+                response=b'{"error": "Server busy, try again later"}',
+                status=503,
+                content_type="application/json",
+            )
+
+        try:
+            # Return the response from the processing task
+            return await task
+        except asyncio.CancelledError:
+            # Handle task cancellation (timeout or queue full)
+            logger.warning("Request cancelled due to timeout or queue full")
+            return Response(
+                response=b'{"error": "Request cancelled"}',
+                status=503,
+                content_type="application/json",
+            )
+
+    # Start the Quart server with host can be set to 0.0.0.0
+    app.run(port=PORT)
 
 
 if __name__ == "__main__":
-    app.run(port=8000)
+    main()

benchmarks/disagg_benchmarks/rate_limiter.py

@@ -0,0 +1,45 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+import asyncio
+import time
+
+
+class RateLimiter:
+    """Token bucket rate limiter implementation"""
+
+    def __init__(self, rate_limit):
+        self.rate_limit = rate_limit  # Requests per second
+        self.num_available_tokens = rate_limit  # Available tokens
+        self.last_refill = time.monotonic()  # Last token refill time
+        self.lock = asyncio.Lock()  # Synchronization lock
+
+    async def acquire(self):
+        """Acquire a token from the rate limiter"""
+        while True:
+            async with self.lock:
+                current_time = time.monotonic()
+                elapsed = current_time - self.last_refill
+
+                # Refill num_available_tokens if more than 1 second has passed
+                if elapsed > 1.0:
+                    self.num_available_tokens = self.rate_limit
+                    self.last_refill = current_time
+
+                # Check if num_available_tokens are available
+                if self.num_available_tokens > 0:
+                    self.num_available_tokens -= 1
+                    return True
+
+                # Calculate wait time if no num_available_tokens available
+                wait_time = 1.0 - elapsed
+            await asyncio.sleep(wait_time)
+
+    async def __aenter__(self):
+        """Enter async context manager - acquire token"""
+        await self.acquire()
+        return self
+
+    async def __aexit__(self, exc_type, exc_value, traceback):
+        """Exit async context manager - no cleanup needed"""
+        pass

benchmarks/disagg_benchmarks/request_queue.py

@@ -0,0 +1,39 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+import asyncio
+from collections import deque
+
+
+class RequestQueue:
+    """Request queue manager with concurrency control"""
+
+    def __init__(self, max_concurrent, max_queue_size):
+        # Maximum concurrent requests
+        self.max_concurrent = max_concurrent
+        self.max_queue_size = max_queue_size  # Maximum queue size
+        # Concurrency control
+        self.semaphore = asyncio.Semaphore(max_concurrent)
+        self.queue = deque()  # Request queue
+        self.queue_size = 0  # Current queue size
+        self.lock = asyncio.Lock()  # Sync queue Lock
+
+    async def enqueue(self, task):
+        """Add a request task to the queue"""
+        async with self.lock:
+            if self.queue_size >= self.max_queue_size:
+                return False
+
+            self.queue.append(task)
+            self.queue_size += 1
+            return True
+
+    async def process(self):
+        """Process queued requests using semaphore for concurrency control"""
+        while True:
+            if self.queue:
+                async with self.semaphore, self.lock:
+                    task = self.queue.popleft()
+                    self.queue_size -= 1
+                    await task
+            await asyncio.sleep(0.01)  # Yield control to event loop

benchmarks/kernels/benchmark_aqlm.py

@@ -1,345 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-
-import os
-import sys
-from typing import Optional
-
-import torch
-import torch.nn.functional as F
-
-from vllm import _custom_ops as ops
-from vllm.model_executor.layers.quantization.aqlm import (
-    dequantize_weight,
-    generic_dequantize_gemm,
-    get_int_dtype,
-    optimized_dequantize_gemm,
-)
-from vllm.utils import FlexibleArgumentParser
-
-os.environ["CUDA_VISIBLE_DEVICES"] = "0"
-
-
-def torch_mult(
-    # [..., in_features]
-    input: torch.Tensor,
-    weights: torch.Tensor,
-    # [num_out_groups, 1, 1, 1]
-    scales: torch.Tensor,
-) -> torch.Tensor:
-    output = F.linear(input, weights)
-    return output
-
-
-def dequant_out_scale(
-    # [..., in_features]
-    input: torch.Tensor,
-    # [num_out_groups, num_in_groups, num_codebooks]
-    codes: torch.IntTensor,
-    # [num_codebooks, codebook_size, out_group_size, in_group_size]
-    codebooks: torch.Tensor,
-    # [num_out_groups, 1, 1, 1]
-    scales: torch.Tensor,
-    output_partition_sizes: torch.IntTensor,
-    bias: Optional[torch.Tensor],
-) -> torch.Tensor:
-    weights = ops.aqlm_dequant(codes, codebooks, output_partition_sizes)
-
-    if bias is None:
-        output = F.linear(input, weights, bias)
-        orig_shape = output.shape
-        flattened_output = output.view(-1, output.size(-1))
-        f_scales = scales.view(-1, scales.shape[0])
-        b_scales = f_scales.expand(flattened_output.shape[0], -1)
-        flattened_output *= b_scales
-        return flattened_output.view(orig_shape)
-    else:
-        b_scales = scales.view(scales.shape[:-3] + (-1,)).expand(-1, weights.shape[1])
-        weights *= b_scales
-        return F.linear(input, weights, bias)
-
-
-def dequant_weight_scale(
-    # [..., in_features]
-    input: torch.Tensor,
-    # [num_out_groups, num_in_groups, num_codebooks]
-    codes: torch.IntTensor,
-    # [num_codebooks, codebook_size, out_group_size, in_group_size]
-    codebooks: torch.Tensor,
-    # [num_out_groups, 1, 1, 1]
-    scales: torch.Tensor,
-    output_partition_sizes: torch.IntTensor,
-    bias: Optional[torch.Tensor],
-) -> torch.Tensor:
-    weights = ops.aqlm_dequant(codes, codebooks, output_partition_sizes)
-
-    b_scales = scales.view(scales.shape[:-3] + (-1,)).expand(-1, weights.shape[1])
-    weights *= b_scales
-    return F.linear(input, weights, bias)
-
-
-def dequant_no_scale(
-    # [..., in_features]
-    input: torch.Tensor,
-    # [num_out_groups, num_in_groups, num_codebooks]
-    codes: torch.IntTensor,
-    # [num_codebooks, codebook_size, out_group_size, in_group_size]
-    codebooks: torch.Tensor,
-    # [num_out_groups, 1, 1, 1]
-    scales: torch.Tensor,
-    output_partition_sizes: torch.IntTensor,
-    bias: Optional[torch.Tensor],
-) -> torch.Tensor:
-    weights = ops.aqlm_dequant(codes, codebooks, output_partition_sizes)
-
-    return F.linear(input, weights, bias)
-
-
-# Compare the optimized 1x16 and 2x8 cuda decompression/dequant kernels against
-# the generic pytorch version.
-# Just visual comparison.
-def dequant_test(k: int, parts: torch.Tensor, nbooks: int, bits: int) -> None:
-    n = int(parts.sum().item())
-
-    device = torch.device("cuda:0")
-
-    code_range = (1 << bits) // 2
-    ingroups = 8
-
-    codes = torch.randint(
-        -code_range,
-        code_range,
-        size=(n, k // ingroups, nbooks),
-        dtype=get_int_dtype(bits),
-        device=device,
-    )
-
-    codebooks = torch.randn(
-        size=(parts.shape[0] * nbooks, 1 << bits, 1, 8),
-        dtype=torch.float16,
-        device=device,
-    )
-
-    count = 0
-    for index in range(16):
-        for i in range(8):
-            for book in range(nbooks):
-                codebooks[book, index, 0, i] = count * (10**book)
-            count += 1
-
-    print("codes shape", codes.shape)
-
-    for i in range(16):
-        for book in range(nbooks):
-            codes[0, i, book] = i
-            codes[0, -i, book] = i
-
-    weights = dequantize_weight(codes, codebooks, None)
-    weights2 = ops.aqlm_dequant(codes, codebooks, parts)
-
-    print("weights shape:", weights.shape)
-    print("weights2 shape:", weights2.shape)
-
-    print("weights are:", weights)
-    print("weights2 are:", weights2)
-
-    print("first 128 weights are", weights[0, 0:128].to(torch.int32))
-    print("first 128 weights2 are:", weights2[0, 0:128].to(torch.int32))
-
-    print("last 128 weights are", weights[0, -128:])
-    print("last 128 weights2 are:", weights2[0, -128:])
-
-
-def main():
-    parser = FlexibleArgumentParser(description="Benchmark aqlm performance.")
-
-    # Add arguments
-    parser.add_argument(
-        "--nbooks", type=int, default=1, help="Number of codebooks (default: 1)"
-    )
-    parser.add_argument(
-        "--bits",
-        type=int,
-        default=16,
-        help="Number of bits per code element (default: 16)",
-    )
-    parser.add_argument(
-        "--test",
-        type=bool,
-        default=False,
-        help="Run the decompression/dequant tester rather than benchmarking "
-        "(default: False)",
-    )
-
-    # Parse the arguments
-    args = parser.parse_args()
-
-    # Extract values
-    nbooks = args.nbooks
-    bits = args.bits
-
-    if args.test:
-        dequant_test(4096, torch.tensor((4096,)), nbooks, bits)
-        return
-
-    # Otherwise, benchmark.
-    methods = [
-        ops.aqlm_gemm,
-        dequant_out_scale,
-        generic_dequantize_gemm,
-        optimized_dequantize_gemm,
-        dequant_weight_scale,
-        torch_mult,
-        dequant_no_scale,
-    ]
-
-    filename = f"./aqlm_benchmark_{nbooks}x{bits}.csv"
-    print(f"writing benchmarks to file {filename}")
-    with open(filename, "w") as f:
-        sys.stdout = f
-
-        print("m | k | n | n parts", end="")
-        for method in methods:
-            print(f" | {method.__name__.replace('_', ' ')} (µs)", end="")
-        print("")
-
-        # These are reasonable prefill sizes.
-        ksandpartions = (
-            (4096, (4096, 4096, 4096)),
-            (4096, (4096,)),
-            (4096, (11008, 11008)),
-            (11008, (4096,)),
-        )
-
-        # reasonable ranges for m.
-        for m in [
-            1,
-            2,
-            4,
-            8,
-            10,
-            12,
-            14,
-            16,
-            24,
-            32,
-            48,
-            52,
-            56,
-            64,
-            96,
-            112,
-            128,
-            256,
-            512,
-            1024,
-            1536,
-            2048,
-            3072,
-            4096,
-        ]:
-            print(f"{m}", file=sys.__stdout__)
-            for ksp in ksandpartions:
-                run_grid(m, ksp[0], torch.tensor(ksp[1]), nbooks, bits, methods)
-
-        sys.stdout = sys.__stdout__
-
-
-def run_grid(m: int, k: int, parts: torch.Tensor, nbooks: int, bits: int, methods):
-    # I didn't see visible improvements from increasing these, but feel free :)
-    num_warmup_trials = 1
-    num_trials = 1
-
-    num_calls = 100
-
-    # warmup.
-    for method in methods:
-        for _ in range(num_warmup_trials):
-            run_timing(
-                num_calls=num_calls,
-                m=m,
-                k=k,
-                parts=parts,
-                nbooks=nbooks,
-                bits=bits,
-                method=method,
-            )
-
-    n = parts.sum().item()
-    print(f"{m} | {k} | {n} | {parts.tolist()}", end="")
-
-    for method in methods:
-        best_time_us = 1e20
-        for _ in range(num_trials):
-            kernel_dur_ms = run_timing(
-                num_calls=num_calls,
-                m=m,
-                k=k,
-                parts=parts,
-                nbooks=nbooks,
-                bits=bits,
-                method=method,
-            )
-
-            kernel_dur_us = 1000 * kernel_dur_ms
-
-            if kernel_dur_us < best_time_us:
-                best_time_us = kernel_dur_us
-
-        print(f" | {kernel_dur_us:.0f}", end="")
-
-    print("")
-
-
-def run_timing(
-    num_calls: int, m: int, k: int, parts: torch.Tensor, nbooks: int, bits: int, method
-) -> float:
-    n = int(parts.sum().item())
-
-    device = torch.device("cuda:0")
-
-    input = torch.randn((1, m, k), dtype=torch.float16, device=device)
-
-    code_range = (1 << bits) // 2
-    ingroups = 8
-
-    codes = torch.randint(
-        -code_range,
-        code_range,
-        size=(n, k // ingroups, nbooks),
-        dtype=get_int_dtype(bits),
-        device=device,
-    )
-
-    codebooks = torch.randn(
-        size=(parts.shape[0] * nbooks, 1 << bits, 1, 8),
-        dtype=torch.float16,
-        device=device,
-    )
-
-    scales = torch.randn(size=(n, 1, 1, 1), dtype=torch.float16, device=device)
-
-    # for comparison to just a pytorch mult.
-    weights = torch.randn((n, k), dtype=torch.float16, device=device)
-
-    start_event = torch.cuda.Event(enable_timing=True)
-    end_event = torch.cuda.Event(enable_timing=True)
-
-    start_event.record()
-
-    if method is torch_mult:
-        for i in range(num_calls):
-            torch_mult(input, weights, scales)
-    else:
-        for i in range(num_calls):
-            method(input, codes, codebooks, scales, parts, None)
-
-    end_event.record()
-    end_event.synchronize()
-
-    dur_ms = start_event.elapsed_time(end_event) / num_calls
-    return dur_ms
-
-
-if __name__ == "__main__":
-    sys.exit(main())

benchmarks/kernels/benchmark_bitblas.py

@@ -3,6 +3,8 @@
 # Copyright (c) Microsoft Corporation.
 # Licensed under the MIT License.
 
+from packaging import version
+
 from vllm.model_executor.layers.quantization.utils.bitblas_utils import (
     MINIMUM_BITBLAS_VERSION,
 )
@@ -10,7 +12,7 @@ from vllm.model_executor.layers.quantization.utils.bitblas_utils import (
 try:
     import bitblas
 
-    if bitblas.__version__ < MINIMUM_BITBLAS_VERSION:
+    if version.parse(bitblas.__version__) < version.parse(MINIMUM_BITBLAS_VERSION):
         raise ImportError(
             "bitblas version is wrong. Please "
             f"install bitblas>={MINIMUM_BITBLAS_VERSION}"

benchmarks/kernels/benchmark_machete.py

@@ -236,6 +236,7 @@ def marlin_create_bench_fn(bt: BenchmarkTensors) -> Callable:
             a=bt.a,
             c=None,
             b_q_weight=w_q,
+            b_bias=None,
             b_scales=w_s,
             global_scale=None,
             b_zeros=w_zp,

benchmarks/kernels/benchmark_moe.py

@@ -3,6 +3,7 @@
 
 import argparse
 import json
+import os
 import time
 from contextlib import nullcontext
 from datetime import datetime
@@ -22,6 +23,13 @@ from vllm.utils import FlexibleArgumentParser
 FP8_DTYPE = current_platform.fp8_dtype()
 
 
+def ensure_divisibility(numerator, denominator, text):
+    """Ensure that numerator is divisible by the denominator."""
+    assert numerator % denominator == 0, "{} {} is not divisible by tp {}.".format(
+        text, numerator, denominator
+    )
+
+
 class BenchmarkConfig(TypedDict):
     BLOCK_SIZE_M: int
     BLOCK_SIZE_N: int
@@ -535,6 +543,7 @@ def save_configs(
     use_fp8_w8a8: bool,
     use_int8_w8a16: bool,
     block_quant_shape: list[int],
+    save_dir: str,
 ) -> None:
     dtype_str = get_config_dtype_str(
         dtype, use_int8_w8a16=use_int8_w8a16, use_fp8_w8a8=use_fp8_w8a8
@@ -545,7 +554,8 @@ def save_configs(
     filename = get_config_file_name(
         num_experts, shard_intermediate_size // 2, dtype_str, block_quant_shape
     )
-
+    os.makedirs(save_dir, exist_ok=True)
+    filename = os.path.join(save_dir, filename)
     print(f"Writing best config to {filename}...")
     with open(filename, "w") as f:
         json.dump(configs, f, indent=4)
@@ -570,12 +580,10 @@ 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
-        shard_intermediate_size = 2 * intermediate_size // args.tp_size
     elif config.architectures[0] == "JambaForCausalLM":
         E = config.num_experts
         topk = config.num_experts_per_tok
         intermediate_size = config.intermediate_size
-        shard_intermediate_size = 2 * intermediate_size // args.tp_size
     elif config.architectures[0] in (
         "DeepseekV3ForCausalLM",
         "DeepseekV2ForCausalLM",
@@ -584,17 +592,14 @@ def main(args: argparse.Namespace):
         E = config.n_routed_experts
         topk = config.num_experts_per_tok
         intermediate_size = config.moe_intermediate_size
-        shard_intermediate_size = 2 * intermediate_size // args.tp_size
     elif config.architectures[0] in ("Qwen2MoeForCausalLM", "Qwen3MoeForCausalLM"):
         E = config.num_experts
         topk = config.num_experts_per_tok
         intermediate_size = config.moe_intermediate_size
-        shard_intermediate_size = 2 * intermediate_size // args.tp_size
     elif config.architectures[0] in ("HunYuanMoEV1ForCausalLM"):
         E = config.num_experts
         topk = config.moe_topk[0]
         intermediate_size = config.moe_intermediate_size[0]
-        shard_intermediate_size = 2 * intermediate_size // args.tp_size
     else:
         # Support for llama4
         config = config.get_text_config()
@@ -602,8 +607,14 @@ def main(args: argparse.Namespace):
         E = config.num_local_experts
         topk = config.num_experts_per_tok
         intermediate_size = config.intermediate_size
+    enable_ep = bool(args.enable_expert_parallel)
+    if enable_ep:
+        ensure_divisibility(E, args.tp_size, "Number of experts")
+        E = E // args.tp_size
+        shard_intermediate_size = 2 * intermediate_size
+    else:
+        ensure_divisibility(intermediate_size, args.tp_size, "intermediate_size")
         shard_intermediate_size = 2 * intermediate_size // args.tp_size
-
     hidden_size = config.hidden_size
     dtype = torch.float16 if current_platform.is_rocm() else config.torch_dtype
     use_fp8_w8a8 = args.dtype == "fp8_w8a8"
@@ -699,6 +710,7 @@ def main(args: argparse.Namespace):
             use_fp8_w8a8,
             use_int8_w8a16,
             block_quant_shape,
+            args.save_dir,
         )
         end = time.time()
         print(f"Tuning took {end - start:.2f} seconds")
@@ -735,10 +747,14 @@ if __name__ == "__main__":
     parser.add_argument(
         "--tp-size", "-tp", "--tensor-parallel-size", type=int, default=2
     )
+    parser.add_argument("--enable-expert-parallel", "-enable-ep", action="store_true")
     parser.add_argument(
         "--dtype", type=str, choices=["auto", "fp8_w8a8", "int8_w8a16"], default="auto"
     )
     parser.add_argument("--use-deep-gemm", action="store_true")
+    parser.add_argument(
+        "--save-dir", type=str, default="./", help="Directory to save tuned results"
+    )
     parser.add_argument("--seed", type=int, default=0)
     parser.add_argument("--batch-size", type=int, nargs="+", required=False)
     parser.add_argument("--tune", action="store_true")

benchmarks/kernels/benchmark_mrope.py

@@ -0,0 +1,328 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+# This script benchmarks the mrope kernel (mainly for Qwen2VL and Qwen2.5VL models).
+# It generates test data, runs benchmarks, and saves results to a CSV file.
+#
+# The CSV file (named with current date/time) contains these columns:
+# model_name, tp_size, num_tokens, num_heads, num_kv_heads, head_dim, max_position,
+# rope_theta, is_neox_style, rope_scaling, dtype, torch_mean, torch_median, torch_p99,
+# torch_min, torch_max, triton_mean, triton_median, triton_p99, triton_min, triton_max,
+# speedup
+#
+# == Usage Examples ==
+#
+# Single model benchmark:
+# python3 benchmark_mrope.py --model-name Qwen/Qwen2-VL-7B-Instruct --tp-size 1 \
+#   --warmup-iter 10 --benchmark-iter 100 --dtype bfloat16 --seed 0 --num-tokens 1024
+#
+# All models benchmark:
+# python3 benchmark_mrope.py --model-name "" --tp-size 1 --warmup-iter 10 \
+#   --benchmark-iter 100 --dtype bfloat16 --seed 0 --num-tokens 1024
+#
+# All models with different TP sizes:
+# python3 benchmark_mrope.py --model-name "" --tp-size 1 2 4 8 --warmup-iter 10 \
+#   --benchmark-iter 100 --dtype bfloat16 --seed 0 --num-tokens 1024
+#
+# All models with different token counts:
+# python3 benchmark_mrope.py --model-name "" --tp-size 1 --warmup-iter 10 \
+#   --benchmark-iter 100 --dtype bfloat16 --seed 0 --num-tokens 1024 4096 16384
+import csv
+import os
+import time
+from datetime import datetime
+from typing import Any
+
+import numpy as np
+import torch
+
+from vllm.model_executor.layers.rotary_embedding import get_rope
+from vllm.platforms import current_platform
+from vllm.transformers_utils.config import get_config
+from vllm.utils import FlexibleArgumentParser
+
+device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+
+
+def generate_test_data(
+    num_tokens: int,
+    num_q_heads: int,
+    num_kv_heads: int,
+    head_size: int,
+    max_position_embeddings: int,
+    dtype: torch.dtype,
+    device: torch.device,
+):
+    """Generate test data for given configuration."""
+    # Create 2D positions (3, num_tokens) for multimodal case
+    positions = torch.randint(
+        0, max_position_embeddings // 4, (3, num_tokens), device=device
+    )
+
+    # Create query and key tensors
+    query = torch.randn(num_tokens, num_q_heads * head_size, dtype=dtype, device=device)
+    key = torch.randn(num_tokens, num_kv_heads * head_size, dtype=dtype, device=device)
+
+    return positions, query, key
+
+
+def calculate_stats(times: list[float]) -> dict[str, float]:
+    """Calculate statistics from a list of times."""
+    times_array = np.array(times)
+    return {
+        "mean": np.mean(times_array),
+        "median": np.median(times_array),
+        "p99": np.percentile(times_array, 99),
+        "min": np.min(times_array),
+        "max": np.max(times_array),
+    }
+
+
+def benchmark_mrope(
+    model_name: str,
+    num_tokens: int,
+    head_dim: int,
+    tp_size: int,
+    num_heads: int,
+    num_kv_heads: int,
+    max_position: int = 8192,
+    rope_theta: float = 10000,
+    is_neox_style: bool = True,
+    rope_scaling: dict[str, Any] = None,
+    dtype: torch.dtype = torch.bfloat16,
+    seed: int = 0,
+    warmup_iter: int = 10,
+    benchmark_iter: int = 100,
+    csv_writer=None,
+):
+    current_platform.seed_everything(seed)
+    torch.set_default_device(device)
+    # the parameters to compute the q k v size based on tp_size
+    mrope_helper_class = get_rope(
+        head_size=head_dim,
+        rotary_dim=head_dim,
+        max_position=max_position,
+        base=rope_theta,
+        is_neox_style=is_neox_style,
+        rope_scaling=rope_scaling,
+        dtype=dtype,
+    ).to(device=device)
+
+    print(80 * "=")
+    print(
+        f"Evaluating model: {model_name} "
+        f"with tp_size: {tp_size} "
+        f"and num_tokens: {num_tokens}, "
+        f"dtype: {dtype}"
+    )
+
+    # create q k v input tensors
+    # create rotary pos emb input tensors
+    positions, query, key = generate_test_data(
+        num_tokens, num_heads, num_kv_heads, head_dim, max_position, dtype, device
+    )
+
+    # Warm up
+    for _ in range(warmup_iter):
+        mrope_helper_class.forward_native(
+            positions,
+            query.clone(),
+            key.clone(),
+        )
+
+        mrope_helper_class.forward_cuda(
+            positions,
+            query.clone(),
+            key.clone(),
+        )
+
+    torch.cuda.synchronize()
+
+    # Time reference implementation
+    torch_times = []
+    for _ in range(benchmark_iter):
+        query_clone = query.clone()
+        key_clone = key.clone()
+        torch.cuda.synchronize()
+        start_time = time.time()
+
+        mrope_helper_class.forward_native(
+            positions,
+            query_clone,
+            key_clone,
+        )
+
+        torch.cuda.synchronize()
+        torch_times.append(time.time() - start_time)
+
+    # Time triton kernel implementation
+    triton_times = []
+    for _ in range(benchmark_iter):
+        query_clone = query.clone()
+        key_clone = key.clone()
+        torch.cuda.synchronize()
+        start_time = time.time()
+        mrope_helper_class.forward_cuda(
+            positions,
+            query_clone,
+            key_clone,
+        )
+        torch.cuda.synchronize()
+        triton_times.append(time.time() - start_time)
+
+    # Calculate statistics
+    torch_stats = calculate_stats(torch_times)
+    triton_stats = calculate_stats(triton_times)
+    print(f"\nPerformance for config ({num_tokens}, {num_heads}, {num_kv_heads}):")
+
+    print(
+        f"Torch implementation: "
+        f"mean={torch_stats['mean']:.8f}s, "
+        f"median={torch_stats['median']:.8f}s, "
+        f"p99={torch_stats['p99']:.8f}s"
+    )
+
+    print(
+        f"Triton implementation: "
+        f"mean={triton_stats['mean']:.8f}s, "
+        f"median={triton_stats['median']:.8f}s, "
+        f"p99={triton_stats['p99']:.8f}s"
+    )
+
+    print(
+        f"Triton Speedup over Torch: {torch_stats['mean'] / triton_stats['mean']:.8f}x"
+    )
+
+    # Write to CSV
+    if csv_writer:
+        row = [
+            model_name,
+            tp_size,
+            num_tokens,
+            num_heads,
+            num_kv_heads,
+            head_dim,
+            max_position,
+            rope_theta,
+            is_neox_style,
+            str(rope_scaling),
+            str(dtype).split(".")[-1],
+            torch_stats["mean"],
+            torch_stats["median"],
+            torch_stats["p99"],
+            torch_stats["min"],
+            torch_stats["max"],
+            triton_stats["mean"],
+            triton_stats["median"],
+            triton_stats["p99"],
+            triton_stats["min"],
+            triton_stats["max"],
+            torch_stats["mean"] / triton_stats["mean"],  # speedup
+        ]
+        csv_writer.writerow(row)
+
+    return torch_stats, triton_stats
+
+
+if __name__ == "__main__":
+    parser = FlexibleArgumentParser(
+        description="Benchmark the rotary embedding kernels."
+    )
+    parser.add_argument("--model-name", type=str, default="")
+    parser.add_argument("--tp-size", type=int, default=1)
+    parser.add_argument("--warmup-iter", type=int, default=10)
+    parser.add_argument("--benchmark-iter", type=int, default=100)
+    parser.add_argument("--dtype", type=str, choices=["bfloat16"], default="bfloat16")
+    parser.add_argument("--seed", type=int, default=0)
+    parser.add_argument("--num-tokens", type=int, nargs="+", required=False)
+    parser.add_argument("--trust-remote-code", action="store_true")
+    parser.add_argument("--output-csv", type=str, default="mrope_benchmark_results.csv")
+    args = parser.parse_args()
+    print(args)
+
+    # Create CSV file for results
+    timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")
+    csv_filename = f"{os.path.splitext(args.output_csv)[0]}_{timestamp}.csv"
+
+    with open(csv_filename, "w", newline="") as csvfile:
+        csv_writer = csv.writer(csvfile)
+        # Write header
+        header = [
+            "model_name",
+            "tp_size",
+            "num_tokens",
+            "num_heads",
+            "num_kv_heads",
+            "head_dim",
+            "max_position",
+            "rope_theta",
+            "is_neox_style",
+            "rope_scaling",
+            "dtype",
+            "torch_mean",
+            "torch_median",
+            "torch_p99",
+            "torch_min",
+            "torch_max",
+            "triton_mean",
+            "triton_median",
+            "triton_p99",
+            "triton_min",
+            "triton_max",
+            "speedup",
+        ]
+        csv_writer.writerow(header)
+
+        model_tp_dict = {}
+        if args.model_name == "":
+            model_tp_dict = {
+                "Qwen/Qwen2-VL-2B-Instruct": [1],
+                "Qwen/Qwen2-VL-7B-Instruct": [1],
+                "Qwen/Qwen2-VL-72B-Instruct": [2, 4, 8],
+                "Qwen/Qwen2.5-VL-3B-Instruct": [1, 2, 4, 8],
+                "Qwen/Qwen2.5-VL-7B-Instruct": [1, 2, 4, 8],
+                "Qwen/Qwen2.5-VL-72B-Instruct": [2, 4, 8],
+            }
+        else:
+            model_tp_dict[args.model_name] = [args.tp_size]
+
+        if args.num_tokens is None:
+            num_tokens_list = [2**i for i in range(0, 18)]
+        else:
+            num_tokens_list = args.num_tokens
+
+        for model_name, tp_list in model_tp_dict.items():
+            config = get_config(model_name, trust_remote_code=args.trust_remote_code)
+            for tp_size in tp_list:
+                # get the model config
+                total_num_kv_heads = config.num_key_value_heads
+                total_num_heads = config.num_attention_heads
+                num_heads = total_num_heads // tp_size
+                num_kv_heads = max(1, total_num_kv_heads // tp_size)
+                head_dim = config.hidden_size // total_num_heads
+                q_size = num_heads * head_dim
+                kv_size = num_kv_heads * head_dim
+                is_neox_style = True
+                rope_theta = config.rope_theta
+                max_position = config.max_position_embeddings
+
+                for num_tokens in num_tokens_list:
+                    benchmark_mrope(
+                        model_name=model_name,
+                        num_tokens=num_tokens,
+                        head_dim=head_dim,
+                        tp_size=tp_size,
+                        num_heads=num_heads,
+                        num_kv_heads=num_kv_heads,
+                        max_position=max_position,
+                        rope_theta=rope_theta,
+                        is_neox_style=is_neox_style,
+                        rope_scaling=config.rope_scaling,
+                        dtype=getattr(torch, args.dtype),
+                        seed=args.seed,
+                        warmup_iter=args.warmup_iter,
+                        benchmark_iter=args.benchmark_iter,
+                        csv_writer=csv_writer,
+                    )
+
+    print(f"Benchmark results saved to {csv_filename}")

benchmarks/kernels/benchmark_per_token_group_quant.py

@@ -0,0 +1,159 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+import argparse
+import math
+from contextlib import contextmanager
+from typing import Callable
+from unittest.mock import patch
+
+import torch
+
+from vllm.model_executor.layers.quantization.utils import fp8_utils, int8_utils
+from vllm.platforms import current_platform
+
+
+@contextmanager
+def _triton_mode():
+    """Temporarily force the Triton fallback path"""
+    with patch("vllm.platforms.current_platform.is_cuda", return_value=False):
+        yield
+
+
+def _time_cuda(
+    fn: Callable[[], tuple[torch.Tensor, torch.Tensor]],
+    warmup_iters: int,
+    bench_iters: int,
+) -> float:
+    # warmup
+    for _ in range(warmup_iters):
+        fn()
+    torch.cuda.synchronize()
+
+    start = torch.cuda.Event(enable_timing=True)
+    end = torch.cuda.Event(enable_timing=True)
+
+    start.record()
+    for _ in range(bench_iters):
+        fn()
+    end.record()
+    torch.cuda.synchronize()
+
+    return start.elapsed_time(end) / bench_iters  # ms/iter
+
+
+def _run_single(
+    shape: tuple[int, int],
+    group_size: int,
+    dtype: str,
+    *,
+    column_major: bool = False,
+    scale_ue8m0: bool = False,
+    warmup_iters: int,
+    bench_iters: int,
+) -> None:
+    num_tokens, hidden_dim = shape
+
+    device = torch.device("cuda")
+    torch.manual_seed(42)
+    x = torch.randn(num_tokens, hidden_dim, device=device, dtype=torch.bfloat16) * 8
+
+    if dtype == "fp8":
+
+        def cuda_impl():
+            return fp8_utils.per_token_group_quant_fp8(
+                x,
+                group_size,
+                column_major_scales=column_major,
+                use_ue8m0=scale_ue8m0,
+            )
+
+        def triton_impl():
+            with _triton_mode():
+                return fp8_utils.per_token_group_quant_fp8(
+                    x,
+                    group_size,
+                    column_major_scales=column_major,
+                    use_ue8m0=scale_ue8m0,
+                )
+    elif dtype == "int8":
+
+        def cuda_impl():
+            return int8_utils.per_token_group_quant_int8(x, group_size)
+
+        def triton_impl():
+            with _triton_mode():
+                return int8_utils.per_token_group_quant_int8(x, group_size)
+    else:
+        raise ValueError("dtype must be 'fp8' or 'int8'")
+
+    cuda_ms = _time_cuda(cuda_impl, warmup_iters, bench_iters)
+    triton_ms = _time_cuda(triton_impl, warmup_iters, bench_iters)
+
+    speedup = triton_ms / cuda_ms if cuda_ms else math.inf
+
+    cfg_desc = (
+        f"shape={shape}  gs={group_size:<3}  col_major={column_major:<5}  "
+        f"ue8m0={scale_ue8m0:<5}  dtype={dtype}"
+    )
+    print(
+        f"{cfg_desc:55} | CUDA {cuda_ms:7.3f} ms  | Triton {triton_ms:7.3f} ms  | "
+        f"speed-up ×{speedup:5.2f}"
+    )
+
+
+def parse_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--warmup-iters", type=int, default=10)
+    parser.add_argument("--bench-iters", type=int, default=100)
+    parser.add_argument("--dtype", choices=["fp8", "int8", "both"], default="both")
+    return parser.parse_args()
+
+
+if __name__ == "__main__":
+    if not current_platform.is_cuda():
+        raise RuntimeError("CUDA device is required to run this benchmark.")
+
+    args = parse_args()
+    warmup_iters, bench_iters = args.warmup_iters, args.bench_iters
+
+    shapes = [(32, 128), (64, 256), (16, 512)]
+    group_sizes = [64, 128]
+
+    dtypes = ["fp8", "int8"] if args.dtype == "both" else [args.dtype]
+
+    header = (
+        "Configuration".ljust(55)
+        + " | "
+        + "CUDA (ms)".center(12)
+        + " | "
+        + "Triton (ms)".center(13)
+        + " | "
+        + "Speed-up"
+    )
+    print(header)
+    print("-" * len(header))
+
+    for dtype in dtypes:
+        for shape in shapes:
+            for gs in group_sizes:
+                if dtype == "fp8":
+                    for col_major in (False, True):
+                        for ue8m0 in (False, True):
+                            _run_single(
+                                shape,
+                                gs,
+                                dtype,
+                                column_major=col_major,
+                                scale_ue8m0=ue8m0,
+                                warmup_iters=warmup_iters,
+                                bench_iters=bench_iters,
+                            )
+                else:  # INT8 has no col-major / ue8m0 switches
+                    _run_single(
+                        shape,
+                        gs,
+                        dtype,
+                        warmup_iters=warmup_iters,
+                        bench_iters=bench_iters,
+                    )

benchmarks/kernels/benchmark_reshape_and_cache_flash.py

@@ -0,0 +1,156 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+from __future__ import annotations
+
+import random
+import time
+
+import torch
+from tabulate import tabulate
+
+from vllm import _custom_ops as ops
+from vllm.logger import init_logger
+from vllm.platforms import current_platform
+from vllm.utils import (
+    STR_DTYPE_TO_TORCH_DTYPE,
+    FlexibleArgumentParser,
+    create_kv_caches_with_random_flash,
+)
+
+logger = init_logger(__name__)
+
+
+@torch.inference_mode()
+def run_benchmark(
+    num_tokens: int,
+    num_heads: int,
+    head_size: int,
+    block_size: int,
+    num_blocks: int,
+    dtype: torch.dtype,
+    kv_cache_dtype: str,
+    kv_cache_layout: str,
+    num_iters: int,
+    device: str = "cuda",
+) -> float:
+    """Return latency (seconds) for given num_tokens."""
+
+    if kv_cache_dtype == "fp8" and head_size % 16:
+        raise ValueError("fp8 kv-cache requires head_size to be a multiple of 16.")
+
+    current_platform.seed_everything(42)
+    torch.set_default_device(device)
+
+    # create random key / value tensors [T, H, D].
+    key = torch.randn(num_tokens, num_heads, head_size, dtype=dtype, device=device)
+    value = torch.randn_like(key)
+
+    # prepare the slot mapping.
+    # each token is assigned a unique slot in the KV-cache.
+    num_slots = block_size * num_blocks
+    if num_tokens > num_slots:
+        raise ValueError("num_tokens cannot exceed the total number of cache slots")
+    slot_mapping_lst = random.sample(range(num_slots), num_tokens)
+    slot_mapping = torch.tensor(slot_mapping_lst, dtype=torch.long, device=device)
+
+    key_caches, value_caches = create_kv_caches_with_random_flash(
+        num_blocks,
+        block_size,
+        1,  # num_layers
+        num_heads,
+        head_size,
+        kv_cache_dtype,
+        dtype,
+        device=device,
+        cache_layout=kv_cache_layout,
+    )
+    key_cache, value_cache = key_caches[0], value_caches[0]
+
+    # compute per-kernel scaling factors for fp8 conversion (if used).
+    k_scale = (key.amax() / 64.0).to(torch.float32)
+    v_scale = (value.amax() / 64.0).to(torch.float32)
+
+    def run_cuda_benchmark(n_iters: int) -> float:
+        nonlocal key, value, key_cache, value_cache, slot_mapping
+        torch.cuda.synchronize()
+        start = time.perf_counter()
+        for _ in range(n_iters):
+            ops.reshape_and_cache_flash(
+                key,
+                value,
+                key_cache,
+                value_cache,
+                slot_mapping,
+                kv_cache_dtype,
+                k_scale,
+                v_scale,
+            )
+        torch.cuda.synchronize()
+        end = time.perf_counter()
+        return (end - start) / n_iters
+
+    # warm-up
+    run_cuda_benchmark(3)
+
+    lat = run_cuda_benchmark(num_iters)
+
+    # free tensors to mitigate OOM when sweeping
+    del key, value, key_cache, value_cache, slot_mapping
+    torch.cuda.empty_cache()
+
+    return lat
+
+
+def main(args):
+    rows = []
+    for layout in ["NHD", "HND"]:
+        for exp in range(1, 17):
+            n_tok = 2**exp
+            lat = run_benchmark(
+                num_tokens=n_tok,
+                num_heads=args.num_heads,
+                head_size=args.head_size,
+                block_size=args.block_size,
+                num_blocks=args.num_blocks,
+                dtype=STR_DTYPE_TO_TORCH_DTYPE[args.dtype],
+                kv_cache_dtype=args.kv_cache_dtype,
+                kv_cache_layout=layout,
+                num_iters=args.iters,
+                device="cuda",
+            )
+            rows.append([n_tok, layout, f"{lat * 1e6:.3f}"])
+
+    print(tabulate(rows, headers=["num_tokens", "layout", "latency (µs)"]))
+
+
+if __name__ == "__main__":
+    parser = FlexibleArgumentParser()
+
+    parser.add_argument("--num-heads", type=int, default=128)
+    parser.add_argument(
+        "--head-size",
+        type=int,
+        choices=[64, 80, 96, 112, 120, 128, 192, 256],
+        default=128,
+    )
+    parser.add_argument("--block-size", type=int, choices=[16, 32], default=16)
+    parser.add_argument("--num-blocks", type=int, default=128 * 512)
+
+    parser.add_argument(
+        "--dtype",
+        type=str,
+        choices=["half", "bfloat16", "float"],
+        default="bfloat16",
+    )
+
+    parser.add_argument(
+        "--kv-cache-dtype",
+        type=str,
+        choices=["auto", "fp8"],
+        default="auto",
+    )
+
+    parser.add_argument("--iters", type=int, default=100)
+    args = parser.parse_args()
+
+    main(args)

benchmarks/kernels/benchmark_trtllm_decode_attention.py

@@ -41,7 +41,6 @@ def benchmark_decode(
     device = "cuda"
     torch.manual_seed(0)
 
-    # Currently only HEAD_GRP_SIZE == 8 is supported
     HEAD_GRP_SIZE = 8
     MAX_SEQ_LEN = max_seq_len
 
@@ -71,22 +70,20 @@ def benchmark_decode(
     if kv_cache_dtype.startswith("fp8"):
         kv_cache, _ = to_float8(kv_cache)
 
+    output_trtllm = torch.empty(q.shape, dtype=dtype)
+
     # Benchmark TRT decode
     def trt_decode():
         return flashinfer.decode.trtllm_batch_decode_with_kv_cache(
             q,
             kv_cache,
             workspace_buffer,
-            num_qo_heads,
-            num_kv_heads,
-            sm_scale,
             block_tables,
             kv_lens_tensor,
-            page_size,
             max_kv_len,
-            kv_cache_dtype,
-            k_scale,
-            v_scale,
+            bmm1_scale=k_scale * sm_scale,
+            bmm2_scale=v_scale,
+            out=output_trtllm,
         )
 
     def time_fn(fn, warmup=10, trials=20):
@@ -125,6 +122,8 @@ def benchmark_decode(
     kv_indices = torch.tensor(kv_indices, dtype=torch.int32)
     kv_last_page_lens = torch.tensor(kv_last_page_lens, dtype=torch.int32)
 
+    output_baseline = torch.empty(q.shape, dtype=dtype)
+
     wrapper = flashinfer.BatchDecodeWithPagedKVCacheWrapper(
         workspace_buffer,
         kv_layout,
@@ -145,7 +144,7 @@ def benchmark_decode(
     )
 
     def baseline_decode():
-        return wrapper.run(q, kv_cache, sm_scale, k_scale, v_scale)
+        return wrapper.run(q, kv_cache, sm_scale, k_scale, v_scale, output_baseline)
 
     baseline_mean, baseline_std = time_fn(baseline_decode)
 
@@ -214,25 +213,39 @@ if __name__ == "__main__":
     max_seq_lens = [1024, 2048, 4096, 8192, 16384, 32768, 65536, 131072]
     all_results = []
 
-    print("Running benchmark for kv_cache_dtype: bfloat16")
     print(
-        "\tnum_seqs\tmax_seq_len\ttrt_mean\ttrt_std\tbaseline_mean\tbaseline_std\tspeedup_percent"
+        "Running benchmark for q_dtype = bfloat16, kv_cache_dtype: bfloat16, "
+        "output_dtype: bfloat16"
+    )
+    print(
+        "\tnum_seqs\tmax_seq_len\ttrt_mean\ttrt_std\tbaseline_mean\t"
+        "baseline_std\tspeedup_percent"
     )
     for max_seq_len in max_seq_lens:
         for bs in num_seqs:
             result = benchmark_decode(
-                bs, max_seq_len, dtype=torch.bfloat16, kv_cache_dtype="auto"
+                bs,
+                max_seq_len,
+                dtype=torch.bfloat16,
+                kv_cache_dtype="auto",
             )
             all_results.append(result)
 
-    print("Running benchmark for q_dtype = bfloat16, kv_cache_dtype: fp8")
     print(
-        "\tnum_seqs\tmax_seq_len\ttrt_mean\ttrt_std\tbaseline_mean\tbaseline_std\tspeedup_percent"
+        "Running benchmark for q_dtype = bfloat16, kv_cache_dtype: fp8, "
+        "output_dtype: bfloat16"
+    )
+    print(
+        "\tnum_seqs\tmax_seq_len\ttrt_mean\ttrt_std\tbaseline_mean\t"
+        "baseline_std\tspeedup_percent"
     )
     for max_seq_len in max_seq_lens:
         for bs in num_seqs:
             result = benchmark_decode(
-                bs, max_seq_len, dtype=torch.bfloat16, kv_cache_dtype="fp8"
+                bs,
+                max_seq_len,
+                dtype=torch.bfloat16,
+                kv_cache_dtype="fp8",
             )
             all_results.append(result)
 

benchmarks/kernels/benchmark_trtllm_prefill_attention.py

@@ -0,0 +1,250 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+import csv
+import os
+import random
+from datetime import datetime
+
+import flashinfer
+import torch
+
+FLOAT32_BYTES = torch.finfo(torch.float).bits // 8
+
+# KV Cache Layout for TRT-LLM
+# kv_cache_shape = (num_blocks, 2, num_kv_heads, page_size, head_dim)
+
+
+def to_float8(x, dtype=torch.float8_e4m3fn):
+    finfo = torch.finfo(dtype)
+    min_val, max_val = x.aminmax()
+    amax = torch.maximum(min_val.abs(), max_val.abs()).clamp(min=1e-12)
+    scale = finfo.max / amax * 0.1
+    x_scl_sat = (x * scale).clamp(min=finfo.min, max=finfo.max)
+    return x_scl_sat.to(dtype), scale.float().reciprocal()
+
+
+@torch.no_grad()
+def benchmark_prefill(
+    num_seqs,
+    max_seq_len,
+    page_size=16,
+    dtype=torch.bfloat16,
+    kv_layout="HND",
+    num_kv_heads=8,
+    kv_cache_dtype="auto",
+    head_dim=128,
+    warmup=10,
+    trials=20,
+):
+    torch.set_default_device("cuda")
+    torch.manual_seed(0)
+
+    HEAD_GRP_SIZE = 8
+    MAX_SEQ_LEN = max_seq_len
+
+    # large number to reduce kv_cache reuse
+    NUM_BLOCKS = int(256000 / page_size)
+
+    workspace_buffer = torch.empty(1024 * 1024 * 1024, dtype=torch.int8)
+
+    num_qo_heads = num_kv_heads * HEAD_GRP_SIZE
+    sm_scale = float(1.0 / (head_dim**0.5))
+
+    q_lens = [random.randint(1, MAX_SEQ_LEN) for _ in range(num_seqs)]
+    q_lens[-1] = MAX_SEQ_LEN
+    max_q_len = max(q_lens)
+    q_indptr = torch.cat(
+        [
+            torch.tensor([0], dtype=torch.int32),
+            torch.cumsum(
+                torch.tensor(q_lens, dtype=torch.int32), dim=0, dtype=torch.int32
+            ),
+        ]
+    )
+    q = torch.randn(sum(q_lens), num_qo_heads, head_dim, dtype=dtype)
+
+    kv_lens = [random.randint(0, MAX_SEQ_LEN) for _ in range(num_seqs)]
+    kv_lens[-1] = MAX_SEQ_LEN
+
+    seq_lens = [q_len + kv_len for q_len, kv_len in zip(q_lens, kv_lens)]
+    max_seq_len = max(seq_lens)
+    seq_lens_tensor = torch.tensor(seq_lens, dtype=torch.int32)
+
+    max_num_blocks_per_seq = (max_seq_len + page_size - 1) // page_size
+    block_tables = torch.randint(
+        0, NUM_BLOCKS, (num_seqs, max_num_blocks_per_seq), dtype=torch.int32
+    )
+
+    kv_cache_shape = (NUM_BLOCKS, 2, num_kv_heads, page_size, head_dim)
+    kv_cache = torch.randn(size=kv_cache_shape, dtype=dtype)
+    k_scale = v_scale = 1.0
+
+    if kv_cache_dtype.startswith("fp8"):
+        kv_cache, _ = to_float8(kv_cache)
+
+    output_trtllm = torch.empty(q.shape, dtype=dtype)
+
+    kv_indptr = [0]
+    kv_indices = []
+    kv_last_page_lens = []
+    for i in range(num_seqs):
+        seq_len = seq_lens[i]
+        assert seq_len > 0
+        num_blocks = (seq_len + page_size - 1) // page_size
+        kv_indices.extend(block_tables[i, :num_blocks])
+        kv_indptr.append(kv_indptr[-1] + num_blocks)
+        kv_last_page_len = seq_len % page_size
+        if kv_last_page_len == 0:
+            kv_last_page_len = page_size
+        kv_last_page_lens.append(kv_last_page_len)
+
+    kv_indptr = torch.tensor(kv_indptr, dtype=torch.int32)
+    kv_indices = torch.tensor(kv_indices, dtype=torch.int32)
+    kv_last_page_lens = torch.tensor(kv_last_page_lens, dtype=torch.int32)
+
+    output_baseline = torch.empty(q.shape, dtype=dtype)
+
+    wrapper = flashinfer.BatchPrefillWithPagedKVCacheWrapper(
+        workspace_buffer, kv_layout
+    )
+    wrapper.plan(
+        q_indptr,
+        kv_indptr,
+        kv_indices,
+        kv_last_page_lens,
+        num_qo_heads,
+        num_kv_heads,
+        head_dim,
+        page_size,
+        causal=True,
+        sm_scale=sm_scale,
+        q_data_type=dtype,
+        kv_data_type=kv_cache.dtype,
+    )
+
+    def time_fn(fn, warmup=10, trials=20):
+        torch.cuda.synchronize()
+        start = torch.cuda.Event(enable_timing=True)
+        end = torch.cuda.Event(enable_timing=True)
+        times = []
+        for i in range(warmup):
+            fn()
+        for i in range(trials):
+            start.record()
+            fn()
+            end.record()
+            torch.cuda.synchronize()
+            times.append(start.elapsed_time(end))  # ms
+        return sum(times) / len(times), torch.std(torch.tensor(times))
+
+    def baseline_prefill():
+        return wrapper.run(
+            q, kv_cache, k_scale=k_scale, v_scale=v_scale, out=output_baseline
+        )
+
+    def trt_prefill():
+        return flashinfer.prefill.trtllm_batch_context_with_kv_cache(
+            query=q,
+            kv_cache=kv_cache,
+            workspace_buffer=workspace_buffer,
+            block_tables=block_tables,
+            seq_lens=seq_lens_tensor,
+            max_q_len=max_q_len,
+            max_kv_len=max_seq_len,
+            bmm1_scale=k_scale * sm_scale,
+            bmm2_scale=v_scale,
+            batch_size=num_seqs,
+            cum_seq_lens_q=q_indptr,
+            cum_seq_lens_kv=kv_indptr,
+            out=output_trtllm,
+        )
+
+    trt_mean, trt_std = time_fn(trt_prefill)
+    baseline_mean, baseline_std = time_fn(baseline_prefill)
+
+    # Calculate percentage speedup (positive means TRT is faster)
+    speedup_percent = (baseline_mean - trt_mean) / baseline_mean
+
+    print(
+        f"\t{num_seqs}\t{max_seq_len}\t{trt_mean:.5f}\t{trt_std.item():.5f}"
+        f"\t{baseline_mean:.5f}\t{baseline_std.item():.5f}\t{speedup_percent:.5f}"
+    )
+
+    # Return results for CSV writing
+    return {
+        "num_seqs": num_seqs,
+        "trt_mean": trt_mean,
+        "trt_std": trt_std.item(),
+        "baseline_mean": baseline_mean,
+        "baseline_std": baseline_std.item(),
+        "speedup_percent": speedup_percent,
+        "q_dtype": str(dtype),
+        "kv_cache_dtype": kv_cache_dtype,
+        "page_size": page_size,
+        "num_kv_heads": num_kv_heads,
+        "head_dim": head_dim,
+        "max_seq_len": max_seq_len,
+    }
+
+
+def write_results_to_csv(results, filename=None):
+    """Write benchmark results to CSV file."""
+    if filename is None:
+        timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")
+        filename = f"flashinfer_trtllm_benchmark_{timestamp}.csv"
+
+    fieldnames = [
+        "num_seqs",
+        "trt_mean",
+        "trt_std",
+        "baseline_mean",
+        "baseline_std",
+        "speedup_percent",
+        "q_dtype",
+        "kv_cache_dtype",
+        "page_size",
+        "num_kv_heads",
+        "head_dim",
+        "max_seq_len",
+    ]
+
+    file_exists = os.path.exists(filename)
+
+    with open(filename, "a", newline="") as csvfile:
+        writer = csv.DictWriter(csvfile, fieldnames=fieldnames)
+
+        if not file_exists:
+            writer.writeheader()
+
+        for result in results:
+            writer.writerow(result)
+
+    print(f"Results written to {filename}")
+
+
+if __name__ == "__main__":
+    num_seqs = [1, 4, 8, 16, 32, 64, 128, 256]
+    max_seq_lens = [1024, 2048, 4096, 8192, 16384, 32768, 65536, 131072]
+    all_results = []
+
+    print(
+        "Running benchmark for q_dtype = bfloat16, kv_cache_dtype: bfloat16, "
+        "output_dtype: bfloat16"
+    )
+    print(
+        "\tnum_seqs\tmax_seq_len\ttrt_mean\ttrt_std\tbaseline_mean\t"
+        "baseline_std\tspeedup_percent"
+    )
+    for max_seq_len in max_seq_lens:
+        for bs in num_seqs:
+            result = benchmark_prefill(
+                bs,
+                max_seq_len,
+                dtype=torch.bfloat16,
+                kv_cache_dtype="auto",
+            )
+            all_results.append(result)
+
+    # Write all results to CSV
+    write_results_to_csv(all_results)

benchmarks/kernels/deepgemm/README.md

@@ -8,7 +8,7 @@ Currently this just includes dense GEMMs and only works on Hopper GPUs.
 
 You need to install vLLM in your usual fashion, then install DeepGEMM from source in its own directory:
 
-```
+```bash
 git clone --recursive https://github.com/deepseek-ai/DeepGEMM
 cd DeepGEMM
 python setup.py install
@@ -17,7 +17,7 @@ uv pip install -e .
 
 ## Usage
 
-```
+```console
 python benchmark_fp8_block_dense_gemm.py
 INFO 02-26 21:55:13 [__init__.py:207] Automatically detected platform cuda.
 ===== STARTING FP8 GEMM BENCHMARK =====

benchmarks/kernels/deepgemm/benchmark_fp8_block_dense_gemm.py

@@ -4,49 +4,16 @@
 # ruff: noqa: E501
 import time
 
-# Import DeepGEMM functions
-import deep_gemm
 import torch
-from deep_gemm import calc_diff, ceil_div, get_col_major_tma_aligned_tensor
 
-# Import vLLM functions
 from vllm import _custom_ops as ops
 from vllm.model_executor.layers.quantization.utils.fp8_utils import (
+    get_col_major_tma_aligned_tensor,
     per_token_group_quant_fp8,
     w8a8_block_fp8_matmul,
 )
 from vllm.triton_utils import triton
-
-
-# Copied from
-# https://github.com/deepseek-ai/DeepGEMM/blob/78cacf70d41d15d688bd493ebc85845f7f2a3d5d/tests/test_core.py#L9
-def per_token_cast_to_fp8(
-        x: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
-    """Convert tensor to FP8 format with per-token scaling."""
-    assert x.dim() == 2 and x.size(1) % 128 == 0
-    m, n = x.shape
-    x_view = x.view(m, -1, 128)
-    x_amax = x_view.abs().float().amax(dim=2).view(m, -1).clamp(1e-4)
-    return (x_view * (448.0 / x_amax.unsqueeze(2))).to(
-        torch.float8_e4m3fn).view(m, n), (x_amax / 448.0).view(m, -1)
-
-
-# Copied from
-# https://github.com/deepseek-ai/DeepGEMM/blob/78cacf70d41d15d688bd493ebc85845f7f2a3d5d/tests/test_core.py#L17
-def per_block_cast_to_fp8(
-        x: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
-    """Convert tensor to FP8 format with per-block scaling."""
-    assert x.dim() == 2
-    m, n = x.shape
-    x_padded = torch.zeros((ceil_div(m, 128) * 128, ceil_div(n, 128) * 128),
-                           dtype=x.dtype,
-                           device=x.device)
-    x_padded[:m, :n] = x
-    x_view = x_padded.view(-1, 128, x_padded.size(1) // 128, 128)
-    x_amax = x_view.abs().float().amax(dim=(1, 3), keepdim=True).clamp(1e-4)
-    x_scaled = (x_view * (448.0 / x_amax)).to(torch.float8_e4m3fn)
-    return x_scaled.view_as(x_padded)[:m, :n].contiguous(), (
-        x_amax / 448.0).view(x_view.size(0), x_view.size(2))
+from vllm.utils.deep_gemm import calc_diff, fp8_gemm_nt, per_block_cast_to_fp8
 
 
 def benchmark_shape(m: int,
@@ -69,14 +36,14 @@ def benchmark_shape(m: int,
 
     # Pre-quantize B for all implementations
     # (weights can be pre-quantized offline)
-    B_deepgemm, B_scale_deepgemm = per_block_cast_to_fp8(B)
-    B_vllm, B_scale_vllm = per_block_cast_to_fp8(B)
+    B_deepgemm, B_scale_deepgemm = per_block_cast_to_fp8(B, [128, 128], use_ue8m0=True)
+    B_vllm, B_scale_vllm = per_block_cast_to_fp8(B, [128, 128], use_ue8m0=True)
 
     # Block size configuration
     block_size = [128, 128]
 
     # Pre-quantize A for all implementations
-    A_deepgemm, A_scale_deepgemm = per_token_cast_to_fp8(A)
+    A_deepgemm, A_scale_deepgemm = per_token_group_quant_fp8(A, block_size[1])
     A_scale_deepgemm = get_col_major_tma_aligned_tensor(A_scale_deepgemm)
     C_deepgemm = torch.empty((m, n), device='cuda', dtype=torch.bfloat16)
     A_vllm, A_scale_vllm = per_token_group_quant_fp8(A, block_size[1])
@@ -85,7 +52,7 @@ def benchmark_shape(m: int,
 
     # === DeepGEMM Implementation ===
     def deepgemm_gemm():
-        deep_gemm.gemm_fp8_fp8_bf16_nt((A_deepgemm, A_scale_deepgemm),
+        fp8_gemm_nt((A_deepgemm, A_scale_deepgemm),
                                        (B_deepgemm, B_scale_deepgemm),
                                        C_deepgemm)
         return C_deepgemm

benchmarks/kv_cache/benchmark_block_pool.py

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

benchmarks/multi_turn/README.md

@@ -0,0 +1,71 @@
+# Benchmark KV Cache Offloading with Multi-Turn Conversations
+
+The requirements (pip) for `benchmark_serving_multi_turn.py` can be found in `requirements.txt`
+
+First start serving your model
+
+```bash
+export MODEL_NAME=/models/meta-llama/Meta-Llama-3.1-8B-Instruct/
+
+vllm serve $MODEL_NAME --disable-log-requests
+```
+
+## Synthetic Multi-Turn Conversations
+
+Download the following text file (used for generation of synthetic conversations)
+
+```bash
+wget https://www.gutenberg.org/ebooks/1184.txt.utf-8
+mv 1184.txt.utf-8 pg1184.txt
+```
+
+The filename `pg1184.txt` is used in `generate_multi_turn.json` (see `"text_files"`).
+
+But you may use other text files if you prefer (using this specific file is not required).
+
+Then run the benchmarking script
+
+```bash
+export MODEL_NAME=/models/meta-llama/Meta-Llama-3.1-8B-Instruct/
+
+python benchmark_serving_multi_turn.py --model $MODEL_NAME --input-file generate_multi_turn.json \
+--num-clients 2 --max-active-conversations 6
+```
+
+You can edit the file `generate_multi_turn.json` to change the conversation parameters (number of turns, etc.).
+
+If successful, you will see the following output
+
+```bash
+----------------------------------------------------------------------------------------------------
+Statistics summary:
+runtime_sec = 215.810
+requests_per_sec = 0.769
+----------------------------------------------------------------------------------------------------
+                   count     mean     std      min      25%      50%      75%      90%      99%      max
+ttft_ms            166.0    78.22   67.63    45.91    59.94    62.26    64.43    69.66   353.18   567.54
+tpot_ms            166.0    25.37    0.57    24.40    25.07    25.31    25.50    25.84    27.50    28.05
+latency_ms         166.0  2591.07  326.90  1998.53  2341.62  2573.01  2860.10  3003.50  3268.46  3862.94
+input_num_turns    166.0     7.43    4.57     1.00     3.00     7.00    11.00    13.00    17.00    17.00
+input_num_tokens   166.0  2006.20  893.56   522.00  1247.75  2019.00  2718.00  3233.00  3736.45  3899.00
+output_num_tokens  166.0   100.01   11.80    80.00    91.00    99.00   109.75   116.00   120.00   120.00
+output_num_chunks  166.0    99.01   11.80    79.00    90.00    98.00   108.75   115.00   119.00   119.00
+----------------------------------------------------------------------------------------------------
+```
+
+## ShareGPT Conversations
+
+To run with the ShareGPT data, download the following ShareGPT dataset:
+`https://huggingface.co/datasets/philschmid/sharegpt-raw/blob/main/sharegpt_20230401_clean_lang_split.json`
+
+Use the `convert_sharegpt_to_openai.py` script to convert the dataset to a format supported by `benchmark_serving_multi_turn.py`
+
+```bash
+python convert_sharegpt_to_openai.py sharegpt_20230401_clean_lang_split.json sharegpt_conv_128.json --seed=99 --max-items=128
+```
+
+The script will convert the ShareGPT dataset to a dataset with the standard user/assistant roles.
+
+The flag `--max-items=128` is used to sample 128 conversations from the original dataset (change as needed).
+
+Use the output JSON file `sharegpt_conv_128.json` as the `--input-file` for `benchmark_serving_multi_turn.py`.

benchmarks/multi_turn/bench_dataset.py

@@ -0,0 +1,493 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+from abc import ABC, abstractmethod
+from statistics import mean
+from typing import Any, NamedTuple, Optional, Union
+
+import numpy as np  # type: ignore
+import pandas as pd  # type: ignore
+from bench_utils import (
+    TEXT_SEPARATOR,
+    Color,
+    logger,
+)
+from transformers import AutoTokenizer  # type: ignore
+
+# Conversation ID is a string (e.g: "UzTK34D")
+ConvId = str
+
+# A list of dicts (dicts with keys "id" and "messages")
+ShareGptConversations = list[dict[str, Any]]
+
+# A list of dicts (dicts with keys "role" and "content")
+MessagesList = list[dict[str, str]]
+
+# Map conversation ID to conversation messages
+ConversationsMap = list[ConvId, MessagesList]
+
+
+class Distribution(ABC):
+    @abstractmethod
+    def sample(self, size: int = 1) -> np.ndarray:
+        pass
+
+
+class UniformDistribution(Distribution):
+    def __init__(
+        self,
+        min_val: Union[int, float],
+        max_val: Union[int, float],
+        is_integer: bool = True,
+    ) -> None:
+        self.min_val = min_val
+        self.max_val = max_val
+        self.is_integer = is_integer
+
+    def sample(self, size: int = 1) -> np.ndarray:
+        if self.is_integer:
+            return np.random.randint(
+                int(self.min_val), int(self.max_val + 1), size=size
+            )
+        else:
+            return np.random.uniform(self.min_val, self.max_val, size=size)
+
+    def __repr__(self) -> str:
+        return f"UniformDistribution[{self.min_val}, {self.max_val}]"
+
+
+class ConstantDistribution(Distribution):
+    def __init__(self, value: Union[int, float]) -> None:
+        self.value = value
+        self.max_val = value
+
+    def sample(self, size: int = 1) -> np.ndarray:
+        return np.full(shape=size, fill_value=self.value)
+
+    def __repr__(self) -> str:
+        return f"Constant[{self.value}]"
+
+
+class ZipfDistribution(Distribution):
+    def __init__(self, alpha: float, max_val: Optional[int] = None) -> None:
+        self.alpha = alpha
+        self.max_val = max_val
+
+    def sample(self, size: int = 1) -> np.ndarray:
+        samples = np.random.zipf(self.alpha, size=size)
+        if self.max_val:
+            samples = np.minimum(samples, self.max_val)
+        return samples
+
+    def __repr__(self) -> str:
+        return f"ZipfDistribution[{self.alpha}]"
+
+
+class PoissonDistribution(Distribution):
+    def __init__(self, alpha: float, max_val: Optional[int] = None) -> None:
+        self.alpha = alpha
+        self.max_val = max_val
+
+    def sample(self, size: int = 1) -> np.ndarray:
+        samples = np.random.poisson(self.alpha, size=size)
+        if self.max_val:
+            samples = np.minimum(samples, self.max_val)
+        return samples
+
+    def __repr__(self) -> str:
+        return f"PoissonDistribution[{self.alpha}]"
+
+
+class LognormalDistribution(Distribution):
+    def __init__(
+        self, mean: float, sigma: float, max_val: Optional[int] = None
+    ) -> None:
+        self.mean = mean
+        self.sigma = sigma
+        self.max_val = max_val
+
+    def sample(self, size: int = 1) -> np.ndarray:
+        samples = np.random.lognormal(mean=self.mean, sigma=self.sigma, size=size)
+        if self.max_val:
+            samples = np.minimum(samples, self.max_val)
+
+        return np.round(samples).astype(int)
+
+    def __repr__(self) -> str:
+        return f"LognormalDistribution[{self.mean}, {self.sigma}]"
+
+
+class GenConvArgs(NamedTuple):
+    num_conversations: int
+    text_files: list[str]
+    input_num_turns: Distribution
+    input_common_prefix_num_tokens: Distribution
+    input_prefix_num_tokens: Distribution
+    input_num_tokens: Distribution
+    output_num_tokens: Distribution
+    print_stats: bool
+
+
+def verify_field_exists(
+    conf: dict, field_name: str, section: str, subsection: str
+) -> None:
+    if field_name not in conf:
+        raise ValueError(
+            f"Missing field '{field_name}' in {section=} and {subsection=}"
+        )
+
+
+def get_random_distribution(
+    conf: dict, section: str, subsection: str, optional: bool = False
+) -> Distribution:
+    # section can be "prompt_input" or "prompt_output" (both required)
+    conf = conf[section]
+
+    if optional and subsection not in conf:
+        # Optional subsection, if not found assume the value is always 0
+        return ConstantDistribution(0)
+
+    # subsection can be "num_turns", "num_tokens" or "prefix_num_tokens"
+    if subsection not in conf:
+        raise ValueError(f"Missing subsection {subsection} in section {section}")
+
+    conf = conf[subsection]
+
+    distribution = conf.get("distribution")
+    if distribution is None:
+        raise ValueError(
+            f"Missing field 'distribution' in {section=} and {subsection=}"
+        )
+
+    if distribution == "constant":
+        verify_field_exists(conf, "value", section, subsection)
+        return ConstantDistribution(conf["value"])
+
+    elif distribution == "zipf":
+        verify_field_exists(conf, "alpha", section, subsection)
+        max_val = conf.get("max", None)
+        return ZipfDistribution(conf["alpha"], max_val=max_val)
+
+    elif distribution == "poisson":
+        verify_field_exists(conf, "alpha", section, subsection)
+        max_val = conf.get("max", None)
+        return PoissonDistribution(conf["alpha"], max_val=max_val)
+
+    elif distribution == "lognormal":
+        verify_field_exists(conf, "mean", section, subsection)
+        verify_field_exists(conf, "sigma", section, subsection)
+        max_val = conf.get("max", None)
+        return LognormalDistribution(conf["mean"], conf["sigma"], max_val=max_val)
+
+    elif distribution == "uniform":
+        verify_field_exists(conf, "min", section, subsection)
+        verify_field_exists(conf, "max", section, subsection)
+
+        min_value = conf["min"]
+        max_value = conf["max"]
+
+        assert min_value > 0
+        assert min_value <= max_value
+
+        is_integer = isinstance(min_value, int) and isinstance(max_value, int)
+        return UniformDistribution(min_value, max_value, is_integer)
+    else:
+        raise ValueError(f"Unknown distribution: {distribution}")
+
+
+def parse_input_json_file(conf: dict) -> GenConvArgs:
+    # Validate the input file
+    assert isinstance(conf, dict)
+    required_fields = [
+        "filetype",
+        "num_conversations",
+        "text_files",
+        "prompt_input",
+        "prompt_output",
+    ]
+    for field in required_fields:
+        assert field in conf, f"Missing field {field} in input {conf}"
+
+    assert conf["filetype"] == "generate_conversations"
+
+    assert conf["num_conversations"] > 0, "num_conversations should be larger than zero"
+
+    text_files = conf["text_files"]
+
+    assert isinstance(text_files, list), "Field 'text_files' should be a list"
+    assert len(text_files) > 0, (
+        "Field 'text_files' should be a list with at least one file"
+    )
+
+    # Parse the parameters for the prompt input/output workload
+    input_num_turns = get_random_distribution(conf, "prompt_input", "num_turns")
+    input_num_tokens = get_random_distribution(conf, "prompt_input", "num_tokens")
+    input_common_prefix_num_tokens = get_random_distribution(
+        conf, "prompt_input", "common_prefix_num_tokens", optional=True
+    )
+    input_prefix_num_tokens = get_random_distribution(
+        conf, "prompt_input", "prefix_num_tokens"
+    )
+    output_num_tokens = get_random_distribution(conf, "prompt_output", "num_tokens")
+
+    print_stats: bool = conf.get("print_stats", False)
+    assert isinstance(print_stats, bool), (
+        "Field 'print_stats' should be either 'true' or 'false'"
+    )
+
+    args = GenConvArgs(
+        num_conversations=conf["num_conversations"],
+        text_files=text_files,
+        input_num_turns=input_num_turns,
+        input_common_prefix_num_tokens=input_common_prefix_num_tokens,
+        input_prefix_num_tokens=input_prefix_num_tokens,
+        input_num_tokens=input_num_tokens,
+        output_num_tokens=output_num_tokens,
+        print_stats=print_stats,
+    )
+    return args
+
+
+def print_conv_stats(conversations: ConversationsMap, tokenizer: AutoTokenizer) -> None:
+    # Collect statistics
+    conv_stats: list[dict[Any, Any]] = []
+    req_stats: list[int] = []
+
+    print("\nCollecting statistics...")
+    for messages in conversations.values():
+        # messages is a list of dicts
+        user_tokens: list[int] = []
+        assistant_tokens: list[int] = []
+        request_tokens: list[int] = []
+
+        req_tokens = 0
+        for m in messages:
+            content = m["content"]
+            num_tokens = len(tokenizer(content).input_ids)
+
+            if m["role"] == "user":
+                user_tokens.append(num_tokens)
+                # New user prompt including all chat history
+                req_tokens += num_tokens
+                request_tokens.append(req_tokens)
+
+            elif m["role"] == "assistant":
+                assistant_tokens.append(num_tokens)
+                # Update assistant answer
+                # (will be part of chat history for the next user prompt)
+                req_tokens += num_tokens
+
+        item_stats = {
+            "conversation_turns": len(messages),
+            "user_tokens": mean(user_tokens),
+            "assistant_tokens": mean(assistant_tokens),
+        }
+
+        conv_stats.append(item_stats)
+        req_stats.extend(request_tokens)
+
+    # Print statistics
+    percentiles = [0.25, 0.5, 0.75, 0.9, 0.99]
+
+    print(TEXT_SEPARATOR)
+    print(f"{Color.YELLOW}Conversations statistics:{Color.RESET}")
+    print(TEXT_SEPARATOR)
+    df = pd.DataFrame(conv_stats)
+    print(df.describe(percentiles=percentiles).transpose())
+    print(TEXT_SEPARATOR)
+    print(f"{Color.YELLOW}Request statistics:{Color.RESET}")
+    print(TEXT_SEPARATOR)
+    df = pd.DataFrame(req_stats, columns=["request_tokens"])
+    print(df.describe(percentiles=percentiles).transpose())
+    print(TEXT_SEPARATOR)
+
+
+def generate_conversations(
+    args: GenConvArgs, tokenizer: AutoTokenizer
+) -> ConversationsMap:
+    # Text for all user prompts
+    # (text from the input text files will be appended to this line)
+    base_prompt_text = "Please rewrite the following text and add more content: "
+    base_prompt_token_count = len(
+        tokenizer.encode(base_prompt_text, add_special_tokens=False)
+    )
+
+    logger.info(f"{Color.PURPLE}Generating conversations...{Color.RESET}")
+    logger.info(args)
+
+    list_of_tokens = []
+
+    for filename in args.text_files:
+        # Load text file that will be used to generate prompts
+        with open(filename) as file:
+            data = file.read()
+            tokens_in_file = tokenizer.encode(data, add_special_tokens=False)
+            list_of_tokens.extend(tokens_in_file)
+
+    conversations: ConversationsMap = {}
+    conv_id = 0
+
+    # Generate number of turns for every conversation
+    turn_count: np.ndarray = args.input_num_turns.sample(args.num_conversations)
+
+    # Turn count should be at least 2 (one user prompt and one assistant answer)
+    turn_count = np.maximum(turn_count, 2)
+
+    # Round up to an even number (every user prompt should have an answer)
+    turn_count = turn_count + (turn_count % 2)
+
+    # Generate number of prefix tokens for every conversation
+    conv_prefix_tokens: np.ndarray = args.input_prefix_num_tokens.sample(
+        args.num_conversations
+    )
+
+    # Used to reduce shared text between conversations
+    # (jump/skip over text sections between conversations)
+    base_offset = 0
+
+    # Common prefix size for all conversations (only 1 sample required)
+    common_prefix_text = ""
+    common_prefix_tokens: int = args.input_common_prefix_num_tokens.sample(1)[0]
+    if common_prefix_tokens > 0:
+        # Using "." at the end to separate sentences
+        common_prefix_text = (
+            tokenizer.decode(list_of_tokens[: common_prefix_tokens - 2]) + "."
+        )
+        base_offset += common_prefix_tokens
+
+    for conv_id in range(args.num_conversations):
+        # Generate a single conversation
+        messages: MessagesList = []
+
+        nturns = turn_count[conv_id]
+
+        # User prompt token count per turn (with lower limit)
+        input_token_count: np.ndarray = args.input_num_tokens.sample(nturns)
+        input_token_count = np.maximum(input_token_count, base_prompt_token_count)
+
+        # Assistant answer token count per turn (with lower limit)
+        output_token_count: np.ndarray = args.output_num_tokens.sample(nturns)
+        output_token_count = np.maximum(output_token_count, 1)
+
+        user_turn = True
+        for turn_id in range(nturns):
+            if user_turn:
+                role = "user"
+                num_tokens = input_token_count[turn_id]
+
+                # Generate the user prompt,
+                # use a unique prefix (the conv_id) for each conversation
+                # (to avoid shared prefix between conversations)
+                content = f"{conv_id} is a nice number... "
+
+                if len(common_prefix_text) > 0 and turn_id == 0:
+                    content = common_prefix_text + content
+
+                # Update the number of tokens left for the content
+                num_tokens -= len(tokenizer.encode(content, add_special_tokens=False))
+
+                if turn_id == 0:
+                    prefix_num_tokens = conv_prefix_tokens[conv_id]
+                    if prefix_num_tokens > 0:
+                        # Add prefix text (context) to the first turn
+                        start_offset = base_offset
+                        end_offset = start_offset + prefix_num_tokens
+                        assert len(list_of_tokens) > end_offset, (
+                            "Not enough input text to generate "
+                            f"{prefix_num_tokens} tokens for the "
+                            f"prefix text ({start_offset=}, {end_offset=})"
+                        )
+
+                        content += f"{conv_id}, " + tokenizer.decode(
+                            list_of_tokens[start_offset:end_offset]
+                        )
+                        base_offset += prefix_num_tokens
+
+                # Add the actual user prompt/question after the prefix text
+                content += base_prompt_text
+                num_tokens -= base_prompt_token_count
+
+                if num_tokens > 0:
+                    # Add text from the input file (to reach the desired token count)
+                    start_offset = base_offset + turn_id * input_token_count.max()
+                    end_offset = start_offset + num_tokens
+                    assert len(list_of_tokens) > end_offset, (
+                        f"Not enough input text to generate {num_tokens} tokens "
+                        f"for the prompt ({start_offset=}, {end_offset=})"
+                    )
+
+                    # Convert tokens back to text
+                    content += tokenizer.decode(list_of_tokens[start_offset:end_offset])
+            else:
+                role = "assistant"
+                # This content will not be used as input to the LLM server
+                # (actual answers will be used instead).
+                # Content is only required to determine the min_tokens/max_tokens
+                # (inputs to the LLM server).
+                num_tokens = output_token_count[turn_id]
+                assert len(list_of_tokens) > num_tokens, (
+                    f"Not enough input text to generate {num_tokens} "
+                    "tokens for assistant content"
+                )
+                content = tokenizer.decode(list_of_tokens[:num_tokens])
+
+            # Append the user/assistant message to the list of messages
+            messages.append({"role": role, "content": content})
+            user_turn = not user_turn
+
+        # Add the new conversation
+        conversations[f"CONV_ID_{conv_id}"] = messages
+
+        # Increase base offset for the next conversation
+        base_offset += nturns
+
+    if args.print_stats:
+        print_conv_stats(conversations, tokenizer)
+
+    return conversations
+
+
+def conversations_list_to_dict(input_list: ShareGptConversations) -> ConversationsMap:
+    conversations: ConversationsMap = {}
+
+    for item in input_list:
+        conv_id: str = item["id"]
+        assert isinstance(conv_id, str)
+
+        assert conv_id not in conversations, (
+            f"Conversation ID {conv_id} found more than once in the input"
+        )
+
+        messages: MessagesList = item["messages"]
+        assert isinstance(messages, list), (
+            f"Conversation messages should be a list (ID: {conv_id})"
+        )
+        assert len(messages) > 0, f"Conversation with no messages (ID: {conv_id})"
+
+        conversations[conv_id] = messages
+
+    logger.info(f"Using {len(conversations)} unique conversations (IDs)")
+    assert len(conversations) == len(input_list)
+
+    # Print statistics about the selected conversations
+    stats: list[dict[str, Any]] = []
+    for conv_data in conversations.values():
+        stats.append({"num_turns": len(conv_data)})
+
+    print(TEXT_SEPARATOR)
+    print(f"{Color.YELLOW}Conversations statistics:{Color.RESET}")
+    print(TEXT_SEPARATOR)
+    percentiles = [0.25, 0.5, 0.75, 0.9, 0.99, 0.999, 0.9999]
+    conv_stats = pd.DataFrame(stats).describe(percentiles=percentiles)
+    print(conv_stats.transpose())
+    print(TEXT_SEPARATOR)
+
+    return conversations
+
+
+def conversations_dict_to_list(input_dict: ConversationsMap) -> ShareGptConversations:
+    output: ShareGptConversations = []
+    for conv_id, conv_data in input_dict.items():
+        new_item = {"id": conv_id, "messages": conv_data}
+        output.append(new_item)
+
+    return output

benchmarks/multi_turn/bench_utils.py

@@ -0,0 +1,28 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+import logging
+from enum import Enum
+
+
+class Color(Enum):
+    RED = "\033[91m"
+    GREEN = "\033[92m"
+    BLUE = "\033[94m"
+    PURPLE = "\033[95m"
+    CYAN = "\033[96m"
+    YELLOW = "\033[93m"
+    RESET = "\033[0m"
+
+    def __str__(self):
+        return self.value
+
+
+TEXT_SEPARATOR = "-" * 100
+
+# Configure the logger
+logging.basicConfig(
+    level=logging.INFO,
+    format="%(asctime)s [%(levelname)s] - %(message)s",
+    datefmt="%d-%m-%Y %H:%M:%S",
+)
+logger = logging.getLogger(__name__)

benchmarks/multi_turn/convert_sharegpt_to_openai.py

@@ -0,0 +1,354 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""
+Download dataset from:
+https://huggingface.co/datasets/philschmid/sharegpt-raw/blob/main/sharegpt_20230401_clean_lang_split.json
+
+Convert to OpenAI API:
+export INPUT_FILE=sharegpt_20230401_clean_lang_split.json
+python convert_sharegpt_to_openai.py $INPUT_FILE sharegpt_conv_128.json --max-items=128
+"""
+
+import argparse
+import json
+import random
+from statistics import mean
+from typing import Any, Optional
+
+import pandas as pd  # type: ignore
+import tqdm  # type: ignore
+from transformers import AutoTokenizer  # type: ignore
+
+
+def has_non_english_chars(text: str) -> bool:
+    return not text.isascii()
+
+
+def content_is_valid(
+    content: str, min_content_len: Optional[int], max_content_len: Optional[int]
+) -> bool:
+    if min_content_len and len(content) < min_content_len:
+        return False
+
+    if max_content_len and len(content) > max_content_len:
+        return False
+
+    return has_non_english_chars(content)
+
+
+def print_stats(
+    conversations: "list[dict[Any, Any]]", tokenizer: Optional[AutoTokenizer] = None
+) -> None:
+    # Collect statistics
+    stats = []
+
+    print("\nCollecting statistics...")
+    for item in tqdm.tqdm(conversations):
+        # item has "id" and "messages"
+        messages = item["messages"]
+
+        user_turns = 0
+        assistant_turns = 0
+        user_words = 0
+        assistant_words = 0
+        conv_chars = 0
+
+        user_tokens: list[int] = []
+        assistant_tokens: list[int] = []
+
+        for m in messages:
+            content = m["content"]
+            conv_chars += len(content)
+            content_num_words = content.count(" ") + 1
+
+            num_tokens = 0
+            if tokenizer:
+                num_tokens = len(tokenizer(m["content"]).input_ids)
+
+            if m["role"] == "user":
+                user_turns += 1
+                user_words += content_num_words
+                if tokenizer:
+                    user_tokens.append(num_tokens)
+
+            elif m["role"] == "assistant":
+                assistant_turns += 1
+                assistant_words += content_num_words
+                if tokenizer:
+                    assistant_tokens.append(num_tokens)
+
+        # assert user_turns == assistant_turns, \
+        # f"Invalid conversation ID {item['id']}"
+
+        conv_words = user_words + assistant_words
+        item_stats = {
+            "user_turns": user_turns,
+            "assistant_turns": assistant_turns,
+            "user_words": user_words,
+            "assistant_words": assistant_words,
+            "conv_turns": len(messages),
+            "conv_words": conv_words,
+            "conv_characters": conv_chars,
+        }
+
+        if len(user_tokens) > 0:
+            item_stats["user_tokens"] = int(mean(user_tokens))
+
+        if len(assistant_tokens) > 0:
+            item_stats["assistant_tokens"] = int(mean(assistant_tokens))
+
+        stats.append(item_stats)
+
+    print("\nStatistics:")
+    percentiles = [0.25, 0.5, 0.75, 0.9, 0.99, 0.999, 0.9999]
+    df = pd.DataFrame(stats)
+    print(df.describe(percentiles=percentiles).transpose())
+
+
+def convert_sharegpt_to_openai(
+    seed: int,
+    input_file: str,
+    output_file: str,
+    max_items: Optional[int],
+    min_content_len: Optional[int] = None,
+    max_content_len: Optional[int] = None,
+    min_turns: Optional[int] = None,
+    max_turns: Optional[int] = None,
+    model: Optional[str] = None,
+) -> None:
+    if min_turns and max_turns:
+        assert min_turns <= max_turns
+
+    if min_content_len and max_content_len:
+        # Verify that min is not larger than max if both were given
+        assert min_content_len <= max_content_len
+
+    print(
+        f"Input parameters:\n{seed=}, {max_items=}, {min_content_len=},"
+        f" {max_content_len=}, {min_turns=}, {max_turns=}\n"
+    )
+
+    random.seed(seed)
+
+    tokenizer = None
+    if model is not None:
+        print(f"Loading tokenizer from: {model}")
+        tokenizer = AutoTokenizer.from_pretrained(model)
+
+    # Read the ShareGPT JSON file
+    print(f"Reading file: {input_file}")
+    with open(input_file, encoding="utf-8") as f:
+        # Should be a list of dicts
+        # Each dict should have "id" (string) and "conversations" (list of dicts)
+        sharegpt_data = json.load(f)
+
+    assert isinstance(sharegpt_data, list), "Input file should contain a list of dicts"
+
+    print(f"Total items in input file: {len(sharegpt_data):,}")
+
+    print(f"Shuffling dataset with seed {seed}")
+    random.shuffle(sharegpt_data)
+
+    # Map conversation ID to the all the messages
+    conversation_parts: dict[str, list[Any]] = {}
+
+    for item in tqdm.tqdm(sharegpt_data):
+        assert "id" in item, "Missing key 'id'"
+        assert "conversations" in item, "Missing key 'conversations'"
+
+        # Conversation ID (e.g: "hiWPlMD") and part/session (0, 1, 2, etc.)
+        conv_id, _ = item["id"].split("_")
+        new_turns = item["conversations"]
+
+        if conv_id not in conversation_parts:
+            # Start new conversation
+            conversation_parts[conv_id] = []
+        elif len(conversation_parts[conv_id]) > 0 and len(new_turns) > 0:
+            prev_turns = conversation_parts[conv_id][-1]
+            if prev_turns[-1]["from"] == new_turns[0]["from"]:
+                new_turns = new_turns[1:]
+
+        if len(new_turns) > 0:
+            # We assume that parts are in order in the ShareGPT dataset
+            conversation_parts[conv_id].append(new_turns)
+
+    dataset: list[dict[str, Any]] = []
+    for conv_id, conv_parts in conversation_parts.items():
+        new_item = {"id": conv_id}
+
+        conversations: list[dict[str, str]] = []
+
+        # Merge all parts
+        for conv_part in conv_parts:
+            conversations.extend(conv_part)
+
+        if len(conversations) > 0:
+            new_item["conversations"] = conversations
+            dataset.append(new_item)
+
+    print(f"Total unique conversations (IDs) in input file: {len(dataset):,}")
+
+    # Final output data
+    final_openai_dataset: list[dict] = []
+
+    # Filter conversations from the ShareGPT dataset and convert to OpenAI format
+    for item in tqdm.tqdm(dataset):
+        messages: list[dict] = []
+
+        assert "id" in item, "Missing key 'id'"
+        assert "conversations" in item, "Missing key 'conversations'"
+
+        conv_id = item["id"]
+        conversations = item["conversations"]
+
+        if min_turns is not None and len(conversations) < min_turns:
+            # Skip short conversations
+            continue
+
+        # Convert each message in the conversation, up to max_turns if specified
+        for i, turn in enumerate(conversations):
+            assert "from" in turn and "value" in turn, (
+                f"Invalid conversation ID {conv_id} - missing 'from' or 'value'"
+            )
+
+            role = None
+            turn_from = turn["from"]
+
+            if turn_from in {"human", "user"}:
+                role = "user"
+            elif turn_from in {"gpt", "bing", "chatgpt", "bard"}:
+                role = "assistant"
+            elif turn_from == "system":
+                role = "system"
+
+            assert role is not None, (
+                f"Invalid conversation ID {conv_id} - 'from'='{turn_from}' is invalid"
+            )
+
+            if i == 0 and role != "user":
+                # If the first message is from assistant (gpt), skip it.
+                # this happens when the conversation is a follow-up
+                # to a previous conversation (from the same user).
+                continue
+
+            if max_turns is not None and i >= max_turns:
+                break
+
+            # Convert message to OpenAI format (with "role" and "content")
+            content = turn["value"]
+            messages.append({"role": role, "content": content})
+
+        # Add the converted conversation to the OpenAI format
+        if len(messages) > 0:
+            valid_messages = True
+
+            # First turn should always be from the user
+            user_turn = True
+
+            for m in messages:
+                # Make sure that turns alternate between user and assistant
+                if (user_turn and m["role"] != "user") or (
+                    not user_turn and m["role"] != "assistant"
+                ):
+                    valid_messages = False
+                    break
+
+                user_turn = not user_turn
+
+                content = m["content"]
+                valid_messages = content_is_valid(
+                    content, min_content_len, max_content_len
+                )
+                if not valid_messages:
+                    break
+
+            if valid_messages is True:
+                final_openai_dataset.append({"id": conv_id, "messages": messages})
+
+    assert len(final_openai_dataset) > 0, "Final number of conversations is zero"
+
+    print_stats(final_openai_dataset)
+
+    print_stats_again = False
+    if max_items is not None and len(final_openai_dataset) > max_items:
+        print(f"\n\nSampling {max_items} items from the dataset...")
+        print_stats_again = True
+        final_openai_dataset = random.sample(final_openai_dataset, max_items)
+
+    if print_stats_again:
+        # Print stats after the dataset changed
+        print_stats(final_openai_dataset, tokenizer)
+
+    # Write the converted data to a new JSON file
+    final_size = len(final_openai_dataset)
+    print(f"\nTotal conversations converted (after filtering): {final_size:,}")
+    print(f"\nWriting file: {output_file}")
+    with open(output_file, "w", encoding="utf-8") as f:
+        json.dump(final_openai_dataset, f, ensure_ascii=False, indent=2)
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser(
+        description="Convert ShareGPT dataset to OpenAI API format"
+    )
+    parser.add_argument("input_file", help="Path to the input ShareGPT JSON file")
+    parser.add_argument(
+        "output_file", help="Path to the output OpenAI format JSON file"
+    )
+    parser.add_argument(
+        "--seed", type=int, default=0, help="Seed for random number generators"
+    )
+    parser.add_argument(
+        "--max-items",
+        type=int,
+        default=None,
+        help="Maximum number of items in the output file",
+    )
+    parser.add_argument(
+        "--min-turns",
+        type=int,
+        default=None,
+        help="Minimum number of turns per conversation",
+    )
+    parser.add_argument(
+        "--max-turns",
+        type=int,
+        default=None,
+        help="Maximum number of turns per conversation",
+    )
+    parser.add_argument(
+        "--min-content-len",
+        type=int,
+        default=None,
+        help="Min number of characters in the messages' content",
+    )
+    parser.add_argument(
+        "--max-content-len",
+        type=int,
+        default=None,
+        help="Max number of characters in the messages' content",
+    )
+    parser.add_argument(
+        "--model",
+        type=str,
+        default=None,
+        help="LLM model, only the tokenizer will be used",
+    )
+
+    args = parser.parse_args()
+
+    convert_sharegpt_to_openai(
+        args.seed,
+        args.input_file,
+        args.output_file,
+        args.max_items,
+        args.min_content_len,
+        args.max_content_len,
+        args.min_turns,
+        args.max_turns,
+        args.model,
+    )
+
+
+if __name__ == "__main__":
+    main()

benchmarks/multi_turn/generate_multi_turn.json

@@ -0,0 +1,35 @@
+{
+    "filetype": "generate_conversations",
+    "num_conversations": 24,
+    "text_files": ["pg1184.txt"],
+    "print_stats": false,
+    "prompt_input": {
+        "num_turns": {
+            "distribution": "uniform",
+            "min": 12,
+            "max": 18
+        },
+        "common_prefix_num_tokens": {
+            "distribution": "constant",
+            "value": 500
+        },
+        "prefix_num_tokens": {
+            "distribution": "lognormal",
+            "mean": 6,
+            "sigma": 4,
+            "max": 1500
+        },
+        "num_tokens": {
+            "distribution": "uniform",
+            "min": 120,
+            "max": 160
+        }
+    },
+    "prompt_output": {
+        "num_tokens": {
+            "distribution": "uniform",
+            "min": 80,
+            "max": 120
+        }
+    }
+}

benchmarks/multi_turn/requirements.txt

@@ -0,0 +1,5 @@
+numpy>=1.24
+pandas>=2.0.0
+aiohttp>=3.10
+transformers>=4.46
+xlsxwriter>=3.2.1

cmake/external_projects/flashmla.cmake

@@ -19,7 +19,7 @@ else()
   FetchContent_Declare(
         flashmla
         GIT_REPOSITORY https://github.com/vllm-project/FlashMLA.git
-        GIT_TAG 575f7724b9762f265bbee5889df9c7d630801845
+        GIT_TAG 0e43e774597682284358ff2c54530757b654b8d1
         GIT_PROGRESS TRUE
         CONFIGURE_COMMAND ""
         BUILD_COMMAND ""
@@ -37,9 +37,9 @@ cuda_archs_loose_intersection(FLASH_MLA_ARCHS "9.0a" "${CUDA_ARCHS}")
 if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.3 AND FLASH_MLA_ARCHS)
     set(FlashMLA_SOURCES
         ${flashmla_SOURCE_DIR}/csrc/flash_api.cpp
-        ${flashmla_SOURCE_DIR}/csrc/flash_fwd_mla_bf16_sm90.cu
-        ${flashmla_SOURCE_DIR}/csrc/flash_fwd_mla_fp16_sm90.cu
-        ${flashmla_SOURCE_DIR}/csrc/flash_fwd_mla_metadata.cu)
+        ${flashmla_SOURCE_DIR}/csrc/kernels/splitkv_mla.cu
+        ${flashmla_SOURCE_DIR}/csrc/kernels/mla_combine.cu
+        ${flashmla_SOURCE_DIR}/csrc/kernels/get_mla_metadata.cu)
 
     set(FlashMLA_INCLUDES
         ${flashmla_SOURCE_DIR}/csrc/cutlass/include

cmake/external_projects/vllm_flash_attn.cmake

@@ -38,7 +38,7 @@ else()
   FetchContent_Declare(
           vllm-flash-attn
           GIT_REPOSITORY https://github.com/vllm-project/flash-attention.git
-          GIT_TAG 1c2624e53c078854e0637ee566c72fe2107e75f4
+          GIT_TAG 57b4e68b9f9d94750b46de8f8dbd2bfcc86edd4f
           GIT_PROGRESS TRUE
           # Don't share the vllm-flash-attn build between build types
           BINARY_DIR ${CMAKE_BINARY_DIR}/vllm-flash-attn

cmake/utils.cmake

@@ -467,6 +467,12 @@ function (define_gpu_extension_target GPU_MOD_NAME)
   if (GPU_LANGUAGE STREQUAL "HIP")
     # Make this target dependent on the hipify preprocessor step.
     add_dependencies(${GPU_MOD_NAME} hipify${GPU_MOD_NAME})
+    # Make sure we include the hipified versions of the headers, and avoid conflicts with the ones in the original source folder
+    target_include_directories(${GPU_MOD_NAME} PRIVATE ${CMAKE_CURRENT_BINARY_DIR}/csrc
+      ${GPU_INCLUDE_DIRECTORIES})
+  else()
+    target_include_directories(${GPU_MOD_NAME} PRIVATE csrc
+      ${GPU_INCLUDE_DIRECTORIES})
   endif()
 
   if (GPU_ARCHITECTURES)
@@ -482,8 +488,6 @@ function (define_gpu_extension_target GPU_MOD_NAME)
   target_compile_definitions(${GPU_MOD_NAME} PRIVATE
     "-DTORCH_EXTENSION_NAME=${GPU_MOD_NAME}")
 
-  target_include_directories(${GPU_MOD_NAME} PRIVATE csrc
-    ${GPU_INCLUDE_DIRECTORIES})
 
   target_link_libraries(${GPU_MOD_NAME} PRIVATE torch ${GPU_LIBRARIES})
 

csrc/activation_kernels.cu

@@ -128,6 +128,45 @@ __global__ void act_and_mul_kernel_with_param(
   }
 }
 
+template <typename T>
+__device__ __forceinline__ T swigluoai_and_mul(const T& gate, const T& up,
+                                               float alpha, float limit) {
+  // clamp gate: min=None, max=limit
+  const float gate_f = (float)gate;
+  const float clamped_gate = gate_f > limit ? limit : gate_f;
+
+  // clamp up: min=-limit, max=limit
+  const float up_f = (float)up;
+  const float clamped_up =
+      up_f > limit ? limit : (up_f < -limit ? -limit : up_f);
+
+  // glu = gate * sigmoid(gate * alpha)
+  const float sigmoid_val = 1.0f / (1.0f + expf(-clamped_gate * alpha));
+  const float glu = clamped_gate * sigmoid_val;
+
+  // (up + 1) * glu
+  return (T)((clamped_up + 1.0f) * glu);
+}
+
+template <typename scalar_t,
+          scalar_t (*ACT_FN)(const scalar_t&, const scalar_t&, const float,
+                             const float)>
+__global__ void swigluoai_and_mul_kernel(
+    scalar_t* __restrict__ out,          // [..., d]
+    const scalar_t* __restrict__ input,  // [..., 2, d]
+    const int d, const float alpha, const float limit) {
+  const int64_t token_idx = blockIdx.x;
+  // TODO: Vectorize loads and stores.
+  for (int64_t idx = threadIdx.x; idx < d; idx += blockDim.x) {
+    // gate = x[..., ::2]  (even indices)
+    const scalar_t gate = VLLM_LDG(&input[token_idx * 2 * d + 2 * idx]);
+    // up = x[..., 1::2]   (odd indices)
+    const scalar_t up = VLLM_LDG(&input[token_idx * 2 * d + 2 * idx + 1]);
+
+    out[token_idx * d + idx] = ACT_FN(gate, up, alpha, limit);
+  }
+}
+
 }  // namespace vllm
 
 #define LAUNCH_ACTIVATION_GATE_KERNEL_WITH_PARAM(KERNEL, PARAM)         \
@@ -145,11 +184,31 @@ __global__ void act_and_mul_kernel_with_param(
                                          PARAM);                        \
       });
 
+#define LAUNCH_SIGLUOAI_AND_MUL(KERNEL, ALPHA, LIMIT)                          \
+  int d = input.size(-1) / 2;                                                  \
+  int64_t num_tokens = input.numel() / input.size(-1);                         \
+  dim3 grid(num_tokens);                                                       \
+  dim3 block(std::min(d, 1024));                                               \
+  const at::cuda::OptionalCUDAGuard device_guard(device_of(input));            \
+  const cudaStream_t stream = at::cuda::getCurrentCUDAStream();                \
+  VLLM_DISPATCH_FLOATING_TYPES(                                                \
+      input.scalar_type(), "clamp_swiglu_kernel_with_params", [&] {            \
+        vllm::swigluoai_and_mul_kernel<scalar_t, KERNEL<scalar_t>>             \
+            <<<grid, block, 0, stream>>>(out.data_ptr<scalar_t>(),             \
+                                         input.data_ptr<scalar_t>(), d, ALPHA, \
+                                         LIMIT);                               \
+      });
+
 void fatrelu_and_mul(torch::Tensor& out,    // [..., d],
                      torch::Tensor& input,  // [..., 2 * d]
                      double threshold) {
   LAUNCH_ACTIVATION_GATE_KERNEL_WITH_PARAM(vllm::fatrelu_kernel, threshold);
 }
+void swigluoai_and_mul(torch::Tensor& out,    // [..., d]
+                       torch::Tensor& input,  // [..., 2 * d]
+                       double alpha, double limit) {
+  LAUNCH_SIGLUOAI_AND_MUL(vllm::swigluoai_and_mul, alpha, limit);
+}
 namespace vllm {
 
 // Element-wise activation kernel template.

csrc/cache_kernels.cu

@@ -5,6 +5,7 @@
 #include "cuda_utils.h"
 #include "cuda_compat.h"
 #include "dispatch_utils.h"
+#include "quantization/vectorization_utils.cuh"
 
 #ifdef USE_ROCM
   #include "quantization/fp8/amd/quant_utils.cuh"
@@ -261,14 +262,26 @@ __global__ void reshape_and_cache_kernel(
   }
 }
 
+// Used by vectorization_utils to copy/convert one element
+template <typename OutT, typename InT, Fp8KVCacheDataType kv_dt>
+struct CopyWithScaleOp {
+  float scale;
+
+  __device__ __forceinline__ void operator()(OutT& dst, const InT src) const {
+    if constexpr (kv_dt == Fp8KVCacheDataType::kAuto) {
+      dst = static_cast<OutT>(src);
+    } else {
+      dst = fp8::scaled_convert<OutT, InT, kv_dt>(src, scale);
+    }
+  }
+};
+
 template <typename scalar_t, typename cache_t, Fp8KVCacheDataType kv_dt>
 __global__ void reshape_and_cache_flash_kernel(
     const scalar_t* __restrict__ key,    // [num_tokens, num_heads, head_size]
     const scalar_t* __restrict__ value,  // [num_tokens, num_heads, head_size]
-    cache_t* __restrict__ key_cache,     // [num_blocks, block_size, num_heads,
-                                         // head_size]
-    cache_t* __restrict__ value_cache,   // [num_blocks, block_size, num_heads,
-                                         // head_size]
+    cache_t* __restrict__ key_cache,     // NHD or HND, shape see comments below
+    cache_t* __restrict__ value_cache,   // same above
     const int64_t* __restrict__ slot_mapping,  // [num_tokens]
     const int64_t block_stride, const int64_t page_stride,
     const int64_t head_stride, const int64_t key_stride,
@@ -282,25 +295,58 @@ __global__ void reshape_and_cache_flash_kernel(
   }
   const int64_t block_idx = slot_idx / block_size;
   const int64_t block_offset = slot_idx % block_size;
-  const int n = num_heads * head_size;
-  for (int i = threadIdx.x; i < n; i += blockDim.x) {
-    const int64_t src_key_idx = token_idx * key_stride + i;
-    const int64_t src_value_idx = token_idx * value_stride + i;
-    const int head_idx = i / head_size;
-    const int head_offset = i % head_size;
-    const int64_t tgt_key_value_idx = block_idx * block_stride +
-                                      block_offset * page_stride +
-                                      head_idx * head_stride + head_offset;
-    scalar_t tgt_key = key[src_key_idx];
-    scalar_t tgt_value = value[src_value_idx];
-    if constexpr (kv_dt == Fp8KVCacheDataType::kAuto) {
-      key_cache[tgt_key_value_idx] = tgt_key;
-      value_cache[tgt_key_value_idx] = tgt_value;
-    } else {
-      key_cache[tgt_key_value_idx] =
-          fp8::scaled_convert<cache_t, scalar_t, kv_dt>(tgt_key, *k_scale);
-      value_cache[tgt_key_value_idx] =
-          fp8::scaled_convert<cache_t, scalar_t, kv_dt>(tgt_value, *v_scale);
+  const int n_elems = num_heads * head_size;
+
+  // pointers to the beginning of the source row for this token.
+  const scalar_t* __restrict__ key_src = key + token_idx * key_stride;
+  const scalar_t* __restrict__ value_src = value + token_idx * value_stride;
+
+  // find the start position inside the kv-cache for this token.
+  cache_t* __restrict__ key_dst =
+      key_cache + block_idx * block_stride + block_offset * page_stride;
+  cache_t* __restrict__ value_dst =
+      value_cache + block_idx * block_stride + block_offset * page_stride;
+
+  // this is true for the NHD layout where `head_stride == head_size`
+  const bool is_contiguous_heads = (head_stride == head_size);
+
+  float k_scale_val = (kv_dt == Fp8KVCacheDataType::kAuto) ? 0.f : *k_scale;
+  float v_scale_val = (kv_dt == Fp8KVCacheDataType::kAuto) ? 0.f : *v_scale;
+  constexpr int VEC_SIZE = (sizeof(scalar_t) == 2) ? 8 : 4;
+  CopyWithScaleOp<cache_t, scalar_t, kv_dt> k_op{k_scale_val};
+  CopyWithScaleOp<cache_t, scalar_t, kv_dt> v_op{v_scale_val};
+  if (is_contiguous_heads) {
+    // NHD layout
+    // kv cache: [num_blocks, block_size, num_heads, head_size]
+    vectorize_with_alignment<VEC_SIZE>(key_src, key_dst, n_elems, threadIdx.x,
+                                       blockDim.x, k_op);
+
+    vectorize_with_alignment<VEC_SIZE>(value_src, value_dst, n_elems,
+                                       threadIdx.x, blockDim.x, v_op);
+
+  } else {
+    // HND layout: heads are strided, but each head_size segment is contiguous
+    // kv cache: [num_blocks, num_heads, block_size, head_size]
+    const int lane = threadIdx.x & 31;     // 0..31 within warp
+    const int warp_id = threadIdx.x >> 5;  // warp index within block
+    const int warps_per_block = blockDim.x >> 5;
+
+    for (int head = warp_id; head < num_heads; head += warps_per_block) {
+      const scalar_t* __restrict__ k_src_h = key_src + head * head_size;
+      const scalar_t* __restrict__ v_src_h = value_src + head * head_size;
+
+      cache_t* __restrict__ k_dst_h =
+          key_dst + static_cast<int64_t>(head) * head_stride;
+      cache_t* __restrict__ v_dst_h =
+          value_dst + static_cast<int64_t>(head) * head_stride;
+
+      // within each head, let the 32 threads of the warp perform the vector
+      // copy
+      vectorize_with_alignment<VEC_SIZE>(k_src_h, k_dst_h, head_size, lane, 32,
+                                         k_op);
+
+      vectorize_with_alignment<VEC_SIZE>(v_src_h, v_dst_h, head_size, lane, 32,
+                                         v_op);
     }
   }
 }

csrc/core/scalar_type.hpp

@@ -321,6 +321,8 @@ static inline constexpr auto kFE3M2f =
     ScalarType::float_(3, 2, true, ScalarType::NAN_NONE);
 static inline constexpr auto kFE4M3fn =
     ScalarType::float_(4, 3, true, ScalarType::NAN_EXTD_RANGE_MAX_MIN);
+static inline constexpr auto kFE8M0fnu =
+    ScalarType(8, 0, false, 0, true, ScalarType::NAN_EXTD_RANGE_MAX_MIN);
 static inline constexpr auto kFE5M2 = ScalarType::float_IEEE754(5, 2);
 static inline constexpr auto kFE8M7 = ScalarType::float_IEEE754(8, 7);
 static inline constexpr auto kFE5M10 = ScalarType::float_IEEE754(5, 10);

csrc/cpu/quant.cpp

@@ -16,12 +16,14 @@ struct KernelVecType<float> {
   using cvt_vec_type = vec_op::FP32Vec16;
 };
 
+#if !defined(__aarch64__) || defined(ARM_BF16_SUPPORT)
 template <>
 struct KernelVecType<c10::BFloat16> {
   using load_vec_type = vec_op::BF16Vec16;
   using azp_adj_load_vec_type = vec_op::INT32Vec16;
   using cvt_vec_type = vec_op::FP32Vec16;
 };
+#endif
 
 template <>
 struct KernelVecType<c10::Half> {

csrc/cutlass_extensions/common.hpp

@@ -60,3 +60,13 @@ struct enable_sm100_only : Kernel {
 #endif
   }
 };
+
+template <typename Kernel>
+struct enable_sm120_only : Kernel {
+  template <typename... Args>
+  CUTLASS_DEVICE void operator()(Args&&... args) {
+#if defined __CUDA_ARCH__ && __CUDA_ARCH__ == 1200
+    Kernel::operator()(std::forward<Args>(args)...);
+#endif
+  }
+};

csrc/mamba/mamba_ssm/selective_scan.h

@@ -45,6 +45,9 @@ struct SSMParamsBase {
     index_t out_d_stride;
     index_t out_z_batch_stride;
     index_t out_z_d_stride;
+    index_t ssm_states_batch_stride;
+    index_t ssm_states_dim_stride;  
+    index_t ssm_states_dstate_stride;
 
     // Common data pointers.
     void *__restrict__ A_ptr;

csrc/mamba/mamba_ssm/selective_scan_fwd.cu

@@ -132,8 +132,10 @@ void selective_scan_fwd_kernel(SSMParamsBase params) {
     input_t *Bvar = reinterpret_cast<input_t *>(params.B_ptr) + sequence_start_index * params.B_batch_stride + group_id * params.B_group_stride;
     weight_t *C = reinterpret_cast<weight_t *>(params.C_ptr) + dim_id * kNRows * params.C_d_stride;
     input_t *Cvar = reinterpret_cast<input_t *>(params.C_ptr) + sequence_start_index * params.C_batch_stride + group_id * params.C_group_stride;
-    input_t *ssm_states = reinterpret_cast<input_t *>(params.ssm_states_ptr) + (cache_index * params.dim + dim_id * kNRows) * params.dstate;
-
+    input_t *ssm_states = reinterpret_cast<input_t *>(params.ssm_states_ptr) + 
+    cache_index * params.ssm_states_batch_stride + 
+    dim_id * kNRows * params.ssm_states_dim_stride;
+    
     float D_val[kNRows] = {0};
     if (params.D_ptr != nullptr) {
         #pragma unroll
@@ -248,7 +250,7 @@ void selective_scan_fwd_kernel(SSMParamsBase params) {
                 }
                 // Initialize running total
 
-                scan_t running_prefix = chunk > 0 ? smem_running_prefix[state_idx + r * MAX_DSTATE] : make_float2(1.0, has_initial_state ? float(ssm_states[state_idx]): 0.0);
+                scan_t running_prefix = chunk > 0 ? smem_running_prefix[state_idx + r * MAX_DSTATE] : make_float2(1.0, has_initial_state ? float(ssm_states[state_idx * params.ssm_states_dstate_stride]): 0.0);
 
                 SSMScanPrefixCallbackOp<weight_t> prefix_op(running_prefix);
                 typename Ktraits::BlockScanT(smem_scan).InclusiveScan(
@@ -259,7 +261,7 @@ void selective_scan_fwd_kernel(SSMParamsBase params) {
                 if (threadIdx.x == 0) {
                     smem_running_prefix[state_idx] = prefix_op.running_prefix;
                     if (chunk == n_chunks - 1) {
-                        ssm_states[state_idx] = input_t(prefix_op.running_prefix.y);
+                        ssm_states[state_idx * params.ssm_states_dstate_stride] = input_t(prefix_op.running_prefix.y);
                     }
                 }
                 #pragma unroll
@@ -481,6 +483,10 @@ void set_ssm_params_fwd(SSMParamsBase &params,
         params.out_batch_stride = out.stride(1);
         params.out_d_stride = out.stride(0);
 
+        params.ssm_states_batch_stride = ssm_states.stride(0);
+        params.ssm_states_dim_stride = ssm_states.stride(1);  
+        params.ssm_states_dstate_stride = ssm_states.stride(2);
+
     }
     else{
         if (!is_variable_B) {
@@ -509,6 +515,10 @@ void set_ssm_params_fwd(SSMParamsBase &params,
         }
         params.out_batch_stride = out.stride(0);
         params.out_d_stride = out.stride(1);
+        
+        params.ssm_states_batch_stride = ssm_states.stride(0);
+        params.ssm_states_dim_stride = ssm_states.stride(1);  
+        params.ssm_states_dstate_stride = ssm_states.stride(2);
     }
 }
 

csrc/moe/marlin_moe_wna16/generate_kernels.py

@@ -20,6 +20,7 @@ namespace MARLIN_NAMESPACE_NAME {
 TEMPLATE = ("template __global__ void Marlin<"
             "{{scalar_t}}, "
             "{{w_type_id}}, "
+            "{{s_type_id}}, "
             "{{threads}}, "
             "{{thread_m_blocks}}, "
             "{{thread_n_blocks}}, "
@@ -77,6 +78,7 @@ def generate_new_kernels():
             if scalar_type == "vllm::kFE4M3fn" and group_blocks not in [-1, 8]:
                 continue
             # nvfp4 only supports group_size == 16
+            # mxfp4 only supports group_size == 32
             if scalar_type == "vllm::kFE2M1f" and group_blocks not in [1, 2]:
                 continue
             # other quantization methods don't support group_size = 16
@@ -89,9 +91,22 @@ def generate_new_kernels():
 
             c_dtype = "half" if dtype == "fp16" else "nv_bfloat16"
 
+            if scalar_type == "vllm::kFE2M1f" and group_blocks == 1:
+                s_type = "vllm::kFE4M3fn"
+            elif scalar_type == "vllm::kFE2M1f" and group_blocks == 2:
+                s_type = "vllm::kFE8M0fnu"
+                if dtype == "fp16":
+                    # we cannot safely dequantize e8m0 to fp16, so skip this
+                    continue
+            elif dtype == "fp16":
+                s_type = "vllm::kFloat16"
+            elif dtype == "bf16":
+                s_type = "vllm::kBFloat16"
+
             template_str = jinja2.Template(TEMPLATE).render(
                 scalar_t=c_dtype,
                 w_type_id=scalar_type + ".id()",
+                s_type_id=s_type + ".id()",
                 threads=threads,
                 thread_m_blocks=max(m_blocks, 1),
                 thread_n_blocks=n_blocks,

csrc/moe/marlin_moe_wna16/kernel.h

@@ -7,23 +7,25 @@
 #include "quantization/gptq_marlin/marlin_dtypes.cuh"
 #include "core/scalar_type.hpp"
 
-#define MARLIN_KERNEL_PARAMS                                          \
-  const int4 *__restrict__ A, const int4 *__restrict__ B,             \
-      int4 *__restrict__ C, int4 *__restrict__ C_tmp,                 \
-      const int4 *__restrict__ scales_ptr,                            \
-      const uint16_t *__restrict__ scale2_ptr,                        \
-      const int4 *__restrict__ zp_ptr, const int *__restrict__ g_idx, \
-      const int32_t *__restrict__ sorted_token_ids_ptr,               \
-      const int32_t *__restrict__ expert_ids_ptr,                     \
-      const int32_t *__restrict__ num_tokens_past_padded_ptr,         \
-      const float *__restrict__ topk_weights_ptr, int top_k,          \
-      bool mul_topk_weights, bool is_ep, int num_groups, int prob_m,  \
-      int prob_n, int prob_k, int *locks, bool use_atomic_add,        \
+#define MARLIN_KERNEL_PARAMS                                                  \
+  const int4 *__restrict__ A, const int4 *__restrict__ B,                     \
+      int4 *__restrict__ C, int4 *__restrict__ C_tmp,                         \
+      const int4 *__restrict__ b_bias_ptr,                                    \
+      const int4 *__restrict__ scales_ptr,                                    \
+      const uint16_t *__restrict__ scale2_ptr,                                \
+      const int4 *__restrict__ zp_ptr, const int *__restrict__ g_idx,         \
+      const int32_t *__restrict__ sorted_token_ids_ptr,                       \
+      const int32_t *__restrict__ expert_ids_ptr,                             \
+      const int32_t *__restrict__ num_tokens_past_padded_ptr,                 \
+      const float *__restrict__ topk_weights_ptr, int top_k,                  \
+      bool mul_topk_weights, bool is_ep, int num_groups, int prob_m,          \
+      int prob_n, int prob_k, int *locks, bool has_bias, bool use_atomic_add, \
       bool use_fp32_reduce, int max_shared_mem
 
 namespace MARLIN_NAMESPACE_NAME {
 template <typename scalar_t,  // compute dtype, half or nv_float16
           const vllm::ScalarTypeId w_type_id,  // weight ScalarType id
+          const vllm::ScalarTypeId s_type_id,  // weight scale ScalarType id
           const int threads,          // number of threads in a threadblock
           const int thread_m_blocks,  // number of 16x16 blocks in the m
                                       // dimension (batchsize) of the

csrc/moe/marlin_moe_wna16/marlin_template.h

@@ -280,6 +280,7 @@ __device__ inline void wait_negative_and_add(int* lock) {
 
 template <typename scalar_t,  // compute dtype, half or nv_float16
           const vllm::ScalarTypeId w_type_id,  // weight ScalarType id
+          const vllm::ScalarTypeId s_type_id,  // weight scale ScalarType id
           const int threads,          // number of threads in a threadblock
           const int thread_m_blocks,  // number of 16x16 blocks in the m
                                       // dimension (batchsize) of the
@@ -299,6 +300,7 @@ __global__ void Marlin(
     const int4* __restrict__ B,  // 4bit quantized weight matrix of shape kxn
     int4* __restrict__ C,        // fp16 output buffer of shape mxn
     int4* __restrict__ C_tmp,    // fp32 tmp output buffer (for reduce)
+    const int4* __restrict__ b_bias_ptr,
     const int4* __restrict__ scales_ptr,  // fp16 quantization scales of shape
                                           // (k/groupsize)xn
     const uint16_t* __restrict__ scale2_ptr,  // fp16 global scale (for nvfp4
@@ -318,8 +320,9 @@ __global__ void Marlin(
     int prob_n,             // output dimension n
     int prob_k,             // reduction dimension k
     int* locks,             // extra global storage for barrier synchronization
-    bool use_atomic_add,    // whether to use atomic add to reduce
-    bool use_fp32_reduce,   // whether to use fp32 global reduce
+    bool has_bias,
+    bool use_atomic_add,   // whether to use atomic add to reduce
+    bool use_fp32_reduce,  // whether to use fp32 global reduce
     int max_shared_mem) {
   // Each threadblock processes one "stripe" of the B matrix with (roughly) the
   // same size, which might involve multiple column "slices" (of width 16 *
@@ -342,12 +345,23 @@ __global__ void Marlin(
 
   extern __shared__ int4 sh[];
   static constexpr auto w_type = vllm::ScalarType::from_id(w_type_id);
+  static constexpr auto s_type = vllm::ScalarType::from_id(s_type_id);
+  if constexpr (w_type == vllm::kFE2M1f) {
+    static_assert(s_type == vllm::kFE4M3fn && group_blocks == 1 ||
+                  s_type == vllm::kFE8M0fnu && group_blocks == 2);
+  } else if constexpr (std::is_same<scalar_t, nv_bfloat16>::value) {
+    static_assert(s_type == vllm::kBFloat16);
+  } else if constexpr (std::is_same<scalar_t, half>::value) {
+    static_assert(s_type == vllm::kFloat16);
+  }
+
   constexpr bool has_zp = w_type == vllm::kU4 || w_type == vllm::kU8;
   constexpr bool is_int_type = w_type == vllm::kU4 || w_type == vllm::kU8 ||
                                w_type == vllm::kU4B8 || w_type == vllm::kU8B128;
   // see comments of dequant.h for more details
   constexpr bool dequant_skip_flop =
-      !is_int_type ||
+      w_type == vllm::kFE4M3fn ||
+      w_type == vllm::kFE2M1f && s_type == vllm::kFE4M3fn ||
       has_zp && !is_zp_float && !std::is_same<scalar_t, nv_bfloat16>::value ||
       has_zp && !is_zp_float && !(w_type == vllm::kU8);
 
@@ -365,6 +379,7 @@ __global__ void Marlin(
   const int zp_expert_stride =
       is_zp_float ? prob_n * prob_k / group_size / 8
                   : prob_n * prob_k / group_size / (pack_factor * 4);
+  const int b_bias_expert_stride = prob_n / 8;
 
   // parallel: num valid moe blocks
   int num_tokens_past_padded = num_tokens_past_padded_ptr[0];
@@ -475,7 +490,7 @@ __global__ void Marlin(
         for (int i = 0; i < 4; i++) {
           int idx = tid4 * 4 + i;
           idx = idx < block_num_valid_tokens ? idx : 0;
-          if constexpr (w_type == vllm::kFE2M1f) {
+          if constexpr (w_type == vllm::kFE2M1f && s_type == vllm::kFE4M3fn) {
             sh_block_topk_weights[idx] = __hmul2(
                 global_scale, Dtype::num2num2(Dtype::float2num(
                                   topk_weights_ptr[sh_block_sorted_ids[idx]])));
@@ -513,7 +528,7 @@ __global__ void Marlin(
       expert_id = expert_ids_ptr[block_id];
     }
 
-    if constexpr (w_type == vllm::kFE2M1f) {
+    if constexpr (w_type == vllm::kFE2M1f && s_type == vllm::kFE4M3fn) {
       uint16_t val = scale2_ptr[expert_id];
       global_scale = Dtype::num2num2(*reinterpret_cast<scalar_t*>(&val));
     }
@@ -526,6 +541,9 @@ __global__ void Marlin(
     if constexpr (has_act_order) {
       g_idx += (expert_id - old_expert_id) * prob_k;
     }
+    if (has_bias) {
+      b_bias_ptr += (expert_id - old_expert_id) * b_bias_expert_stride;
+    }
 
     read_moe_block_data(block_id);
   };
@@ -721,7 +739,7 @@ __global__ void Marlin(
 
     s_sh_rd = 8 * ((threadIdx.x / 32) % (thread_n_blocks / 4)) +
               (threadIdx.x % 32) / 4;
-    s_sh_rd = s_sh_rd * 2 + warp_row % 2;
+    s_sh_rd = s_sh_rd * 2 + (warp_row / group_blocks) % 2;
 
   } else if constexpr (group_blocks != -1)
     s_sh_rd = 8 * ((threadIdx.x / 32) % (thread_n_blocks / 4)) +
@@ -734,6 +752,18 @@ __global__ void Marlin(
     s_sh_rd = 8 * ((threadIdx.x / 32) % (thread_n_blocks / 4)) +
               (threadIdx.x % 32) % 4;
 
+  int bias_sh_rd;
+  if constexpr (m_block_size_8) {
+    bias_sh_rd = 8 * ((threadIdx.x / 32) % (thread_n_blocks / 4)) +
+                 (threadIdx.x % 32) / 8;
+  } else {
+    bias_sh_rd = 8 * ((threadIdx.x / 32) % (thread_n_blocks / 4)) +
+                 (threadIdx.x % 32) % 4;
+  }
+
+  int bias_sh_wr = threadIdx.x;
+  int bias_gl_rd = (thread_n_blocks * 16 / 8) * slice_col + threadIdx.x;
+
   // Zero-points have the same read layout as the scales
   // (without column-wise case)
   constexpr int num_col_threads = 8;
@@ -793,7 +823,19 @@ __global__ void Marlin(
   constexpr int sh_b_size = stages * b_sh_stage;
   int4* sh_b = sh_new;
   int4* sh_red = sh_new;
-  int4* sh_g_idx = sh_b + (sh_red_size > sh_b_size ? sh_red_size : sh_b_size);
+
+  constexpr int sh_size_b_red_min =
+      (sh_red_size < sh_b_size ? sh_red_size : sh_b_size);
+  constexpr int sh_size_b_red_max =
+      (sh_red_size > sh_b_size ? sh_red_size : sh_b_size);
+  constexpr int sh_bias_size = (thread_n_blocks * 16 / 8);
+  constexpr int sh_b_red_bias_size =
+      sh_size_b_red_max > (sh_size_b_red_min + sh_bias_size)
+          ? sh_size_b_red_max
+          : (sh_size_b_red_min + sh_bias_size);
+
+  int4* sh_bias = sh_new + sh_size_b_red_min;
+  int4* sh_g_idx = sh_new + sh_b_red_bias_size;
   int4* sh_zp = sh_g_idx + (stages * g_idx_stage);
   constexpr int sh_s_size = has_act_order ? (act_s_max_num_groups * s_sh_stride)
                                           : (stages * s_sh_stage);
@@ -803,9 +845,9 @@ __global__ void Marlin(
   static_assert(thread_m_blocks * 16 * thread_n_blocks * 16 / 8 <=
                 stages * b_sh_stage);
   int4* sh_a = sh_s + sh_s_size;
-  constexpr int shm_size_used =
-      moe_block_size + stages * (g_idx_stage + zp_sh_stage) + sh_s_size +
-      (sh_red_size > sh_b_size ? sh_red_size : sh_b_size);
+  constexpr int shm_size_used = moe_block_size +
+                                stages * (g_idx_stage + zp_sh_stage) +
+                                sh_s_size + sh_b_red_bias_size;
 
   // all remaining shared memory is used to cache A (input)
   // sh_a_max_row is at least ` stages * 16 * thread_m_blocks `
@@ -816,7 +858,8 @@ __global__ void Marlin(
   FragA frag_a[2][thread_m_blocks];
   I4 frag_b_quant[2][b_thread_vecs];
   FragC frag_c[thread_m_blocks][4][2];
-  FragS frag_s[2][4];                    // No act-order
+  FragS frag_s[2][4];  // No act-order
+  FragS frag_bias[2][4];
   FragS act_frag_s[2][4][4];             // For act-order
   int frag_qzp[2][num_ints_per_thread];  // Zero-points
   FragZP frag_zp;                        // Zero-points in fp16
@@ -1065,10 +1108,15 @@ __global__ void Marlin(
           if constexpr (w_type_id != vllm::kFE2M1f.id()) {
             reinterpret_cast<int4*>(&frag_s[k % 2])[0] =
                 sh_s_stage[s_sh_rd + cur_group_id * s_sh_stride];
-          } else {
+          } else if constexpr (group_blocks == 1 || thread_k_blocks > 4) {
             reinterpret_cast<int2*>(&frag_s[k % 2])[0] =
                 reinterpret_cast<int2*>(
                     sh_s_stage)[s_sh_rd + cur_group_id * (2 * s_sh_stride)];
+          } else {
+            reinterpret_cast<int2*>(&frag_s[k % 2])[0] =
+                reinterpret_cast<int2*>(
+                    sh_s_stage)[s_sh_rd + cur_group_id * (2 * s_sh_stride) +
+                                k % 2];
           }
         }
       }
@@ -1281,9 +1329,9 @@ __global__ void Marlin(
       int s_quant_0 = reinterpret_cast<int*>(frag_s[k2])[0];
       int s_quant_1 = reinterpret_cast<int*>(frag_s[k2])[1];
 
-      dequant_fp8_scales<scalar_t2>(s_quant_0,
-                                    reinterpret_cast<scalar_t2*>(&frag_s[k2]));
-      dequant_fp8_scales<scalar_t2>(
+      dequant_fp8_scales<scalar_t2, s_type_id>(
+          s_quant_0, reinterpret_cast<scalar_t2*>(&frag_s[k2]));
+      dequant_fp8_scales<scalar_t2, s_type_id>(
           s_quant_1, reinterpret_cast<scalar_t2*>(&frag_s[k2]) + 2);
     }
 
@@ -1566,7 +1614,7 @@ __global__ void Marlin(
   // Write out the reduce final result in the correct layout. We only actually
   // reshuffle matrix fragments in this step, the reduction above is performed
   // in fragment layout.
-  auto write_result = [&]() {
+  auto write_result = [&](bool last) {
     int c_gl_stride = prob_n / 8;
     constexpr int c_sh_stride = 2 * thread_n_blocks + 1;
     int c_gl_wr_delta = c_gl_stride * (threads / (2 * thread_n_blocks));
@@ -1592,7 +1640,7 @@ __global__ void Marlin(
 
     // We first reorder in shared memory to guarantee the most efficient final
     // global write patterns
-    auto write = [&](int idx, float c0, float c1, FragS& s) {
+    auto write = [&](int idx, float c0, float c1, FragS& s, FragS& b_bias) {
       scalar_t2 res =
           Dtype::nums2num2(Dtype::float2num(c0), Dtype::float2num(c1));
 
@@ -1601,14 +1649,27 @@ __global__ void Marlin(
       if constexpr (!has_act_order && group_blocks == -1 &&
                     w_type.size_bits() == 4 &&
                     (has_zp && dequant_skip_flop || !has_zp)) {
-        res = __hmul2(res, s[0]);
+        scalar_t2 tmp_scale = s[0];
+        if constexpr (m_block_size_8) {
+          tmp_scale = Dtype::num2num2(
+              reinterpret_cast<scalar_t*>(&s[0])[(threadIdx.x % 8) / 4]);
+        }
+        res = __hmul2(res, tmp_scale);
       }
 
-      if constexpr (w_type == vllm::kFE2M1f) {
+      if constexpr (w_type == vllm::kFE2M1f && s_type == vllm::kFE4M3fn) {
         if (!mul_topk_weights) {
           res = __hmul2(res, global_scale);
         }
       }
+      if (has_bias && last) {
+        scalar_t2 tmp_bias = b_bias[0];
+        if constexpr (m_block_size_8) {
+          tmp_bias = Dtype::num2num2(
+              reinterpret_cast<scalar_t*>(&b_bias[0])[(threadIdx.x % 8) / 4]);
+        }
+        res = __hadd2(res, tmp_bias);
+      }
 
       if constexpr (m_block_size_8) {
         ((scalar_t*)sh_red)[idx] = res.x;
@@ -1626,19 +1687,25 @@ __global__ void Marlin(
           if constexpr (m_block_size_8) {
             int wr = c_sh_wr + 16 * j;
             write(wr, frag_c[i][j][0][0], frag_c[i][j][0][1],
-                  frag_s[j / 2][2 * (j % 2) + 0]);
+                  frag_s[j / 2][2 * (j % 2) + 0],
+                  frag_bias[j / 2][2 * (j % 2) + 0]);
             write(wr + 8, frag_c[i][j][0][2], frag_c[i][j][0][3],
-                  frag_s[j / 2][2 * (j % 2) + 1]);
+                  frag_s[j / 2][2 * (j % 2) + 1],
+                  frag_bias[j / 2][2 * (j % 2) + 1]);
           } else {
             int wr = c_sh_wr + 8 * j;
             write(wr + (4 * c_sh_stride) * 0 + 0, frag_c[i][j][0][0],
-                  frag_c[i][j][0][1], frag_s[j / 2][2 * (j % 2) + 0]);
+                  frag_c[i][j][0][1], frag_s[j / 2][2 * (j % 2) + 0],
+                  frag_bias[j / 2][2 * (j % 2) + 0]);
             write(wr + (4 * c_sh_stride) * 8 + 0, frag_c[i][j][0][2],
-                  frag_c[i][j][0][3], frag_s[j / 2][2 * (j % 2) + 0]);
+                  frag_c[i][j][0][3], frag_s[j / 2][2 * (j % 2) + 0],
+                  frag_bias[j / 2][2 * (j % 2) + 0]);
             write(wr + (4 * c_sh_stride) * 0 + 4, frag_c[i][j][1][0],
-                  frag_c[i][j][1][1], frag_s[j / 2][2 * (j % 2) + 1]);
+                  frag_c[i][j][1][1], frag_s[j / 2][2 * (j % 2) + 1],
+                  frag_bias[j / 2][2 * (j % 2) + 1]);
             write(wr + (4 * c_sh_stride) * 8 + 4, frag_c[i][j][1][2],
-                  frag_c[i][j][1][3], frag_s[j / 2][2 * (j % 2) + 1]);
+                  frag_c[i][j][1][3], frag_s[j / 2][2 * (j % 2) + 1],
+                  frag_bias[j / 2][2 * (j % 2) + 1]);
           }
         }
         c_sh_wr += 16 * (4 * c_sh_stride);
@@ -1805,6 +1872,14 @@ __global__ void Marlin(
       }
 
       thread_block_reduce();
+
+      if (has_bias && last) {
+        __syncthreads();
+        cp_async4_pred(&sh_bias[bias_sh_wr], &b_bias_ptr[bias_gl_rd],
+                       threadIdx.x < 16 * thread_n_blocks / 8);
+        cp_async_fence();
+      }
+
       if constexpr (!has_act_order && group_blocks == -1 &&
                     (has_zp && dequant_skip_flop || !has_zp)) {
         if (w_type.size_bits() == 8 || (last || use_atomic_add)) {
@@ -1867,11 +1942,20 @@ __global__ void Marlin(
         }
         barrier_release(&locks[locks_off], last);
       }
+
+      if (has_bias && last) {
+        cp_async_wait<0>();
+        __syncthreads();
+        reinterpret_cast<int4*>(&frag_bias)[0] = sh_bias[bias_sh_rd];
+        reinterpret_cast<int4*>(&frag_bias)[1] = sh_bias[bias_sh_rd + 4];
+        __syncthreads();
+      }
+
       if (use_atomic_add && slice_count > 1 && slice_idx != 0)
         wait_negative_and_add(&locks[locks_off]);
       if (last || use_atomic_add)
         // only the last block in a slice actually writes the result
-        write_result();
+        write_result(last);
       int old_slice_row = slice_row;
       slice_row = 0;
       slice_col_par++;
@@ -1904,6 +1988,7 @@ __global__ void Marlin(
           for (int i = 0; i < b_sh_wr_iters; i++) B_ptr[i] -= b_gl_stride;
         }
 
+        bias_gl_rd = (thread_n_blocks * 16 / 8) * slice_col + threadIdx.x;
         // Update slice k/n for scales loading
         if constexpr (has_act_order) {
           slice_k_start = tb_k * slice_row;

csrc/moe/marlin_moe_wna16/ops.cu

@@ -51,8 +51,9 @@ __global__ void permute_cols_kernel(
 }  // namespace marlin
 
 torch::Tensor moe_wna16_marlin_gemm(
-    torch::Tensor& a, std::optional<torch::Tensor> const& c_or_none,
-    torch::Tensor& b_q_weight, torch::Tensor& b_scales,
+    torch::Tensor& a, std::optional<torch::Tensor> c_or_none,
+    torch::Tensor& b_q_weight,
+    std::optional<torch::Tensor> const& b_bias_or_none, torch::Tensor& b_scales,
     std::optional<torch::Tensor> const& b_zeros_or_none,
     std::optional<torch::Tensor> const& g_idx_or_none,
     std::optional<torch::Tensor> const& perm_or_none, torch::Tensor& workspace,
@@ -212,7 +213,7 @@ int get_kernel_cache_size(thread_config_t const& th_config, bool m_block_size_8,
   // Get B size
   int tb_k = th_config.thread_k;
   int tb_n = th_config.thread_n;
-  int tb_m = thread_m_blocks * (m_block_size_8 ? 8 : 16);
+  int tb_m = thread_m_blocks * 16;
 
   // shm size for block_sorted_ids/rd_block_sorted_ids/block_topk_weights
   // both of them requires tb_m * 4 bytes (tb_m * int32 or tb_m * float32)
@@ -220,6 +221,11 @@ int get_kernel_cache_size(thread_config_t const& th_config, bool m_block_size_8,
   int sh_a_size = pipe_stages * (tb_m * tb_k) * 2;
   int sh_b_size = pipe_stages * (tb_k * tb_n / pack_factor) * 4;
   int sh_red_size = tb_m * (tb_n + 8) * 2;
+  int sh_bias_size = tb_n * 2;
+  int tmp_size =
+      (sh_b_size > sh_red_size ? sh_red_size : sh_b_size) + sh_bias_size;
+  tmp_size = max(max(sh_b_size, sh_red_size), tmp_size);
+
   int sh_s_size =
       get_scales_cache_size(th_config, prob_m, prob_n, prob_k, num_bits,
                             group_size, has_act_order, is_k_full);
@@ -234,8 +240,8 @@ int get_kernel_cache_size(thread_config_t const& th_config, bool m_block_size_8,
       sh_zp_size = sh_s_size / 2;
   }
 
-  int total_size = max(sh_b_size, sh_red_size) + sh_a_size + sh_s_size +
-                   sh_zp_size + sh_g_idx_size + sh_block_meta_size;
+  int total_size = tmp_size + sh_a_size + sh_s_size + sh_zp_size +
+                   sh_g_idx_size + sh_block_meta_size;
 
   return total_size;
 }
@@ -270,20 +276,25 @@ bool is_valid_config(thread_config_t const& th_config, bool m_block_size_8,
   int cache_size = get_kernel_cache_size(
       th_config, m_block_size_8, thread_m_blocks, prob_m, prob_n, prob_k,
       num_bits, group_size, has_act_order, is_k_full, has_zp, is_zp_float);
-  return cache_size <= max_shared_mem;
+  return cache_size + 512 <= max_shared_mem;
 }
 
-  #define _GET_IF(W_TYPE, THREAD_M_BLOCKS, THREAD_N_BLOCKS, THREAD_K_BLOCKS, \
-                  M_BLOCK_SIZE_8, GROUP_BLOCKS, NUM_THREADS, IS_ZP_FLOAT)    \
-    else if (q_type == W_TYPE && thread_m_blocks == THREAD_M_BLOCKS &&       \
-             thread_n_blocks == THREAD_N_BLOCKS &&                           \
-             thread_k_blocks == THREAD_K_BLOCKS &&                           \
-             m_block_size_8 == M_BLOCK_SIZE_8 &&                             \
-             group_blocks == GROUP_BLOCKS && num_threads == NUM_THREADS &&   \
-             is_zp_float == IS_ZP_FLOAT) {                                   \
-      kernel = Marlin<scalar_t, W_TYPE.id(), NUM_THREADS, THREAD_M_BLOCKS,   \
-                      THREAD_N_BLOCKS, THREAD_K_BLOCKS, M_BLOCK_SIZE_8,      \
-                      pipe_stages, GROUP_BLOCKS, IS_ZP_FLOAT>;               \
+  #define _GET_IF(W_TYPE, THREAD_M_BLOCKS, THREAD_N_BLOCKS, THREAD_K_BLOCKS,   \
+                  M_BLOCK_SIZE_8, GROUP_BLOCKS, NUM_THREADS, IS_ZP_FLOAT)      \
+    else if (q_type == W_TYPE && thread_m_blocks == THREAD_M_BLOCKS &&         \
+             thread_n_blocks == THREAD_N_BLOCKS &&                             \
+             thread_k_blocks == THREAD_K_BLOCKS &&                             \
+             m_block_size_8 == M_BLOCK_SIZE_8 &&                               \
+             group_blocks == GROUP_BLOCKS && num_threads == NUM_THREADS &&     \
+             is_zp_float == IS_ZP_FLOAT) {                                     \
+      constexpr auto S_TYPE =                                                  \
+          W_TYPE == vllm::kFE2M1f                                              \
+              ? (GROUP_BLOCKS == 1 ? vllm::kFE4M3fn : vllm::kFE8M0fnu)         \
+              : (std::is_same<scalar_t, half>::value ? vllm::kFloat16          \
+                                                     : vllm::kBFloat16);       \
+      kernel = Marlin<scalar_t, W_TYPE.id(), S_TYPE.id(), NUM_THREADS,         \
+                      THREAD_M_BLOCKS, THREAD_N_BLOCKS, THREAD_K_BLOCKS,       \
+                      M_BLOCK_SIZE_8, pipe_stages, GROUP_BLOCKS, IS_ZP_FLOAT>; \
     }
 
   // COMMON: cases for (group_blocks in [-1, 2, 4, 8] and is_zp_float == false)
@@ -335,31 +346,45 @@ bool is_valid_config(thread_config_t const& th_config, bool m_block_size_8,
     _GET_IF(W_TYPE, 2, N_BLOCKS, K_BLOCKS, false, 8, NUM_THREADS, false)  \
     _GET_IF(W_TYPE, 3, N_BLOCKS, K_BLOCKS, false, -1, NUM_THREADS, false) \
     _GET_IF(W_TYPE, 3, N_BLOCKS, K_BLOCKS, false, 8, NUM_THREADS, false)  \
-                                                                          \
     _GET_IF(W_TYPE, 4, N_BLOCKS, K_BLOCKS, false, -1, NUM_THREADS, false) \
     _GET_IF(W_TYPE, 4, N_BLOCKS, K_BLOCKS, false, 8, NUM_THREADS, false)
 
-  #define FP4_GET_IF_M1(W_TYPE, N_BLOCKS, K_BLOCKS, NUM_THREADS)        \
-    _GET_IF(W_TYPE, 1, N_BLOCKS, K_BLOCKS, true, 1, NUM_THREADS, false) \
-    _GET_IF(W_TYPE, 1, N_BLOCKS, K_BLOCKS, false, 1, NUM_THREADS, false)
-
-  #define FP4_GET_IF_M234(W_TYPE, N_BLOCKS, K_BLOCKS, NUM_THREADS)       \
-    _GET_IF(W_TYPE, 2, N_BLOCKS, K_BLOCKS, false, 1, NUM_THREADS, false) \
-    _GET_IF(W_TYPE, 3, N_BLOCKS, K_BLOCKS, false, 1, NUM_THREADS, false) \
-    _GET_IF(W_TYPE, 4, N_BLOCKS, K_BLOCKS, false, 1, NUM_THREADS, false)
-
-  #define FP4_GET_IF(W_TYPE)            \
-    FP4_GET_IF_M1(W_TYPE, 8, 8, 256)    \
-    FP4_GET_IF_M1(W_TYPE, 8, 4, 128)    \
-    FP4_GET_IF_M234(W_TYPE, 16, 4, 256) \
-    FP4_GET_IF_M234(W_TYPE, 8, 4, 128)
-
   #define BIGGROUP_GET_IF(W_TYPE)            \
     BIGGROUP_GET_IF_M1(W_TYPE, 8, 8, 256)    \
     BIGGROUP_GET_IF_M1(W_TYPE, 8, 4, 128)    \
     BIGGROUP_GET_IF_M234(W_TYPE, 16, 4, 256) \
     BIGGROUP_GET_IF_M234(W_TYPE, 8, 4, 128)
 
+  #define NVFP4_GET_IF_M1(W_TYPE, N_BLOCKS, K_BLOCKS, NUM_THREADS)      \
+    _GET_IF(W_TYPE, 1, N_BLOCKS, K_BLOCKS, true, 1, NUM_THREADS, false) \
+    _GET_IF(W_TYPE, 1, N_BLOCKS, K_BLOCKS, false, 1, NUM_THREADS, false)
+
+  #define NVFP4_GET_IF_M234(W_TYPE, N_BLOCKS, K_BLOCKS, NUM_THREADS)     \
+    _GET_IF(W_TYPE, 2, N_BLOCKS, K_BLOCKS, false, 1, NUM_THREADS, false) \
+    _GET_IF(W_TYPE, 3, N_BLOCKS, K_BLOCKS, false, 1, NUM_THREADS, false) \
+    _GET_IF(W_TYPE, 4, N_BLOCKS, K_BLOCKS, false, 1, NUM_THREADS, false)
+
+  #define NVFP4_GET_IF(W_TYPE)            \
+    NVFP4_GET_IF_M1(W_TYPE, 8, 8, 256)    \
+    NVFP4_GET_IF_M1(W_TYPE, 8, 4, 128)    \
+    NVFP4_GET_IF_M234(W_TYPE, 16, 4, 256) \
+    NVFP4_GET_IF_M234(W_TYPE, 8, 4, 128)
+
+  #define MXFP4_GET_IF_M1(W_TYPE, N_BLOCKS, K_BLOCKS, NUM_THREADS)      \
+    _GET_IF(W_TYPE, 1, N_BLOCKS, K_BLOCKS, true, 2, NUM_THREADS, false) \
+    _GET_IF(W_TYPE, 1, N_BLOCKS, K_BLOCKS, false, 2, NUM_THREADS, false)
+
+  #define MXFP4_GET_IF_M234(W_TYPE, N_BLOCKS, K_BLOCKS, NUM_THREADS)     \
+    _GET_IF(W_TYPE, 2, N_BLOCKS, K_BLOCKS, false, 2, NUM_THREADS, false) \
+    _GET_IF(W_TYPE, 3, N_BLOCKS, K_BLOCKS, false, 2, NUM_THREADS, false) \
+    _GET_IF(W_TYPE, 4, N_BLOCKS, K_BLOCKS, false, 2, NUM_THREADS, false)
+
+  #define MXFP4_GET_IF(W_TYPE)            \
+    MXFP4_GET_IF_M1(W_TYPE, 8, 8, 256)    \
+    MXFP4_GET_IF_M1(W_TYPE, 8, 4, 128)    \
+    MXFP4_GET_IF_M234(W_TYPE, 16, 4, 256) \
+    MXFP4_GET_IF_M234(W_TYPE, 8, 4, 128)
+
   // We currently have 4-bit models only with group_blocks == 4
   #define FZP_GET_IF_M1(W_TYPE, N_BLOCKS, K_BLOCKS, NUM_THREADS)       \
     _GET_IF(W_TYPE, 1, N_BLOCKS, K_BLOCKS, true, 4, NUM_THREADS, true) \
@@ -408,12 +433,17 @@ MarlinFuncPtr get_marlin_kernel(const vllm::ScalarType q_type,
   COMMON_GET_IF(vllm::kU4B8)
   COMMON_GET_IF(vllm::kU8B128)
 
-  BIGGROUP_GET_IF(vllm::kFE4M3fn)
+  NVFP4_GET_IF(vllm::kFE2M1f)
 
-  FP4_GET_IF(vllm::kFE2M1f)
+  BIGGROUP_GET_IF(vllm::kFE4M3fn)
 
   ACT_GET_IF(vllm::kU4B8)
   ACT_GET_IF(vllm::kU8B128)
+  if (std::is_same<scalar_t, nv_bfloat16>::value) {
+    if (false) {
+    }
+    MXFP4_GET_IF(vllm::kFE2M1f)
+  }
 
   return kernel;
 }
@@ -482,16 +512,16 @@ exec_config_t determine_exec_config(const vllm::ScalarType& q_type, int prob_m,
 }
 
 template <typename scalar_t>
-void marlin_mm(const void* A, const void* B, void* C, void* C_tmp, void* s,
-               void* s2, void* zp, void* g_idx, void* perm, void* a_tmp,
-               void* sorted_token_ids, void* expert_ids,
+void marlin_mm(const void* A, const void* B, void* C, void* C_tmp, void* b_bias,
+               void* s, void* s2, void* zp, void* g_idx, void* perm,
+               void* a_tmp, void* sorted_token_ids, void* expert_ids,
                void* num_tokens_past_padded, void* topk_weights,
                int moe_block_size, int top_k, bool mul_topk_weights, bool is_ep,
                int prob_m, int prob_n, int prob_k, void* workspace,
-               vllm::ScalarType const& q_type, bool has_act_order,
-               bool is_k_full, bool has_zp, int num_groups, int group_size,
-               int dev, cudaStream_t stream, int thread_k, int thread_n,
-               int sms, bool use_atomic_add, bool use_fp32_reduce,
+               vllm::ScalarType const& q_type, bool has_bias,
+               bool has_act_order, bool is_k_full, bool has_zp, int num_groups,
+               int group_size, int dev, cudaStream_t stream, int thread_k,
+               int thread_n, int sms, bool use_atomic_add, bool use_fp32_reduce,
                bool is_zp_float) {
   int thread_m_blocks = div_ceil(moe_block_size, 16);
   bool m_block_size_8 = moe_block_size == 8;
@@ -538,6 +568,7 @@ void marlin_mm(const void* A, const void* B, void* C, void* C_tmp, void* s,
   const int4* B_ptr = (const int4*)B;
   int4* C_ptr = (int4*)C;
   int4* C_tmp_ptr = (int4*)C_tmp;
+  const int4* bias_ptr = (const int4*)b_bias;
   const int4* s_ptr = (const int4*)s;
   const uint16_t* s2_ptr = (const uint16_t*)s2;
   const int4* zp_ptr = (const int4*)zp;
@@ -648,10 +679,10 @@ void marlin_mm(const void* A, const void* B, void* C, void* C_tmp, void* s,
   // avoid ">>>" being formatted to "> > >"
   // clang-format off
   kernel<<<blocks, num_threads, max_shared_mem, stream>>>(
-      A_ptr, B_ptr, C_ptr, C_tmp_ptr, s_ptr, s2_ptr, zp_ptr, g_idx_ptr,
+      A_ptr, B_ptr, C_ptr, C_tmp_ptr, bias_ptr, s_ptr, s2_ptr, zp_ptr, g_idx_ptr,
       sorted_token_ids_ptr, expert_ids_ptr, num_tokens_past_padded_ptr,
       topk_weights_ptr, top_k, mul_topk_weights, is_ep, num_groups, prob_m,
-      prob_n, prob_k, locks, use_atomic_add, use_fp32_reduce, max_shared_mem);
+      prob_n, prob_k, locks, has_bias, use_atomic_add, use_fp32_reduce, max_shared_mem);
   // clang-format on
 }
 
@@ -659,7 +690,8 @@ void marlin_mm(const void* A, const void* B, void* C, void* C_tmp, void* s,
 
 torch::Tensor moe_wna16_marlin_gemm(
     torch::Tensor& a, std::optional<torch::Tensor> const& c_or_none,
-    torch::Tensor& b_q_weight, torch::Tensor& b_scales,
+    torch::Tensor& b_q_weight,
+    std::optional<torch::Tensor> const& b_bias_or_none, torch::Tensor& b_scales,
     std::optional<torch::Tensor> const& global_scale_or_none,
     std::optional<torch::Tensor> const& b_zeros_or_none,
     std::optional<torch::Tensor> const& g_idx_or_none,
@@ -766,7 +798,6 @@ torch::Tensor moe_wna16_marlin_gemm(
   num_groups = b_scales.size(1);
 
   torch::Tensor g_idx, perm, a_tmp;
-  ;
   if (g_idx_or_none.has_value() && perm_or_none.has_value()) {
     g_idx = g_idx_or_none.value();
     perm = perm_or_none.value();
@@ -815,12 +846,24 @@ torch::Tensor moe_wna16_marlin_gemm(
   torch::Tensor global_scale;
   if (global_scale_or_none.has_value()) {
     global_scale = global_scale_or_none.value();
-    TORCH_CHECK(b_q_type == vllm::kFE2M1f,
-                "global_scale can only be used for float4_e2m1f.");
+    TORCH_CHECK(b_q_type == vllm::kFE2M1f && group_size == 16,
+                "global_scale can only be used for nvfp4 format.");
   } else {
     global_scale = torch::empty({0}, options);
-    TORCH_CHECK(!(b_q_type == vllm::kFE2M1f),
-                "the global_scale parameter must be passed for float4_e2m1f.");
+    TORCH_CHECK(!(b_q_type == vllm::kFE2M1f && group_size == 16),
+                "the global_scale parameter must be passed for nvfp4 format.");
+  }
+
+  bool has_bias = b_bias_or_none.has_value();
+  torch::Tensor b_bias;
+  if (has_bias) {
+    b_bias = b_bias_or_none.value();
+    TORCH_CHECK(b_bias.device().is_cuda(), "b_bias is not on GPU");
+    TORCH_CHECK(b_bias.is_contiguous(), "b_bias is not contiguous");
+    TORCH_CHECK(b_bias.size(1) == size_n, "b_bias.size(0) != size_n");
+    TORCH_CHECK(b_bias.stride(1) == 1, "b_bias.stride(1) != 1");
+  } else {
+    b_bias = torch::empty({0}, options);
   }
 
   torch::Tensor b_zeros;
@@ -832,7 +875,6 @@ torch::Tensor moe_wna16_marlin_gemm(
     b_zeros = torch::empty({0}, options);
   }
   bool has_zp = b_zeros.size(-1) > 0;
-
   if (has_zp) {
     TORCH_CHECK(
         b_q_type == vllm::kU4 || b_q_type == vllm::kU8,
@@ -890,41 +932,58 @@ torch::Tensor moe_wna16_marlin_gemm(
   if (a.scalar_type() == at::ScalarType::Half) {
     void* scales_ptr;
     if (b_q_type == vllm::kFE2M1f) {
-      scales_ptr = b_scales.data_ptr<at::Float8_e4m3fn>();
+      if (group_size == 16)
+        scales_ptr = b_scales.data_ptr<at::Float8_e4m3fn>();
+      else if (group_size == 32)
+        scales_ptr = b_scales.data_ptr<at::Float8_e8m0fnu>();
+      else
+        TORCH_CHECK(false,
+                    "float4_e2m1f only supports group_size == 16 (NVFP4) ",
+                    "and group_size == 32 (MXFP4)");
     } else {
       scales_ptr = b_scales.data_ptr<at::Half>();
     }
 
     MARLIN_NAMESPACE_NAME::marlin_mm<half>(
         a.data_ptr<at::Half>(), b_q_weight.data_ptr(), c.data_ptr<at::Half>(),
-        c_tmp.data_ptr<float>(), scales_ptr, global_scale.data_ptr<at::Half>(),
-        b_zeros.data_ptr(), g_idx.data_ptr(), perm.data_ptr(),
-        a_tmp.data_ptr<at::Half>(), sorted_token_ids.data_ptr(),
-        expert_ids.data_ptr(), num_tokens_past_padded.data_ptr(),
-        topk_weights.data_ptr(), moe_block_size, top_k, mul_topk_weights, is_ep,
-        size_m, size_n, size_k, workspace.data_ptr(), b_q_type, has_act_order,
-        is_k_full, has_zp, num_groups, group_size, dev,
+        c_tmp.data_ptr<float>(), b_bias.data_ptr<at::Half>(), scales_ptr,
+        global_scale.data_ptr<at::Half>(), b_zeros.data_ptr(), g_idx.data_ptr(),
+        perm.data_ptr(), a_tmp.data_ptr<at::Half>(),
+        sorted_token_ids.data_ptr(), expert_ids.data_ptr(),
+        num_tokens_past_padded.data_ptr(), topk_weights.data_ptr(),
+        moe_block_size, top_k, mul_topk_weights, is_ep, size_m, size_n, size_k,
+        workspace.data_ptr(), b_q_type, has_bias, has_act_order, is_k_full,
+        has_zp, num_groups, group_size, dev,
         at::cuda::getCurrentCUDAStream(dev), thread_k, thread_n, sms,
         use_atomic_add, use_fp32_reduce, is_zp_float);
   } else if (a.scalar_type() == at::ScalarType::BFloat16) {
     void* scales_ptr;
     if (b_q_type == vllm::kFE2M1f) {
-      scales_ptr = b_scales.data_ptr<at::Float8_e4m3fn>();
+      if (group_size == 16)
+        scales_ptr = b_scales.data_ptr<at::Float8_e4m3fn>();
+      else if (group_size == 32)
+        scales_ptr = b_scales.data_ptr<at::Float8_e8m0fnu>();
+      else
+        TORCH_CHECK(false,
+                    "float4_e2m1f only supports group_size == 16 (NVFP4) ",
+                    "and group_size == 32 (MXFP4)");
     } else {
       scales_ptr = b_scales.data_ptr<at::BFloat16>();
     }
 
     MARLIN_NAMESPACE_NAME::marlin_mm<nv_bfloat16>(
         a.data_ptr<at::BFloat16>(), b_q_weight.data_ptr(),
-        c.data_ptr<at::BFloat16>(), c_tmp.data_ptr<float>(), scales_ptr,
+        c.data_ptr<at::BFloat16>(), c_tmp.data_ptr<float>(),
+        b_bias.data_ptr<at::BFloat16>(), scales_ptr,
         global_scale.data_ptr<at::BFloat16>(), b_zeros.data_ptr(),
         g_idx.data_ptr(), perm.data_ptr(), a_tmp.data_ptr<at::BFloat16>(),
         sorted_token_ids.data_ptr(), expert_ids.data_ptr(),
         num_tokens_past_padded.data_ptr(), topk_weights.data_ptr(),
         moe_block_size, top_k, mul_topk_weights, is_ep, size_m, size_n, size_k,
-        workspace.data_ptr(), b_q_type, has_act_order, is_k_full, has_zp,
-        num_groups, group_size, dev, at::cuda::getCurrentCUDAStream(dev),
-        thread_k, thread_n, sms, use_atomic_add, use_fp32_reduce, is_zp_float);
+        workspace.data_ptr(), b_q_type, has_bias, has_act_order, is_k_full,
+        has_zp, num_groups, group_size, dev,
+        at::cuda::getCurrentCUDAStream(dev), thread_k, thread_n, sms,
+        use_atomic_add, use_fp32_reduce, is_zp_float);
   } else {
     TORCH_CHECK(false,
                 "moe_wna16_marlin_gemm only supports bfloat16 and float16");

csrc/moe/topk_softmax_kernels.cu

@@ -24,9 +24,12 @@
 #ifndef USE_ROCM
     #include <cub/util_type.cuh>
     #include <cub/cub.cuh>
+    #include <cuda/std/functional>
+    using AddOp = cuda::std::plus<float>;
 #else
     #include <hipcub/util_type.hpp>
     #include <hipcub/hipcub.hpp>
+    using AddOp = cub::Sum; 
 #endif
 
 #define MAX(a, b) ((a) > (b) ? (a) : (b))
@@ -62,7 +65,6 @@ __launch_bounds__(TPB) __global__
 
     const int thread_row_offset = blockIdx.x * num_cols;
 
-    cub::Sum sum;
     float threadData(-FLT_MAX);
 
     // Don't touch finished rows.
@@ -92,7 +94,7 @@ __launch_bounds__(TPB) __global__
         threadData += exp((static_cast<float>(input[idx]) - float_max));
     }
 
-    const auto Z = BlockReduce(tmpStorage).Reduce(threadData, sum);
+    const auto Z = BlockReduce(tmpStorage).Reduce(threadData, AddOp());
 
     if (threadIdx.x == 0)
     {
@@ -186,7 +188,9 @@ __launch_bounds__(TPB) __global__ void moeTopK(
   It fuses the softmax, max and argmax into a single kernel.
 
   Limitations:
-  1) This implementation is intended for when the number of experts is a small power of 2.
+  1) This implementation is optimized for when the number of experts is a small power of 2.
+     Additionally it also supports when number of experts is multiple of 64 which is still
+     faster than the computing softmax and topK separately (only tested on CUDA yet).
   2) This implementation assumes k is small, but will work for any k.
 */
 
@@ -196,8 +200,6 @@ __launch_bounds__(WARPS_PER_CTA* WARP_SIZE_PARAM) __global__
         int* source_rows, const int k, const int start_expert, const int end_expert)
 {
     // We begin by enforcing compile time assertions and setting up compile time constants.
-    static_assert(VPT == (VPT & -VPT), "VPT must be power of 2");
-    static_assert(NUM_EXPERTS == (NUM_EXPERTS & -NUM_EXPERTS), "NUM_EXPERTS must be power of 2");
     static_assert(BYTES_PER_LDG == (BYTES_PER_LDG & -BYTES_PER_LDG), "BYTES_PER_LDG must be power of 2");
     static_assert(BYTES_PER_LDG <= 16, "BYTES_PER_LDG must be leq 16");
 
@@ -405,12 +407,10 @@ struct TopkConstants
 };
 } // namespace detail
 
-template <int EXPERTS, int WARPS_PER_TB, int WARP_SIZE_PARAM, typename IndType>
+template <int EXPERTS, int WARPS_PER_TB, int WARP_SIZE_PARAM, int MAX_BYTES_PER_LDG, typename IndType>
 void topkGatingSoftmaxLauncherHelper(const float* input, const bool* finished, float* output, IndType* indices,
     int* source_row, const int num_rows, const int k, const int start_expert, const int end_expert, cudaStream_t stream)
 {
-    static constexpr std::size_t MAX_BYTES_PER_LDG = 16;
-
     static constexpr int BYTES_PER_LDG = MIN(MAX_BYTES_PER_LDG, sizeof(float) * EXPERTS);
     using Constants = detail::TopkConstants<EXPERTS, BYTES_PER_LDG, WARP_SIZE_PARAM>;
     static constexpr int VPT = Constants::VPT;
@@ -423,21 +423,27 @@ void topkGatingSoftmaxLauncherHelper(const float* input, const bool* finished, f
         input, finished, output, num_rows, indices, source_row, k, start_expert, end_expert);
 }
 
-#define LAUNCH_SOFTMAX(NUM_EXPERTS, WARPS_PER_TB)                                \
-    switch (warpSize) {                                                          \
-        case 32:                                                                 \
-            topkGatingSoftmaxLauncherHelper<NUM_EXPERTS, WARPS_PER_TB, 32>(      \
-                gating_output, nullptr, topk_weights, topk_indices,              \
-                token_expert_indices, num_tokens, topk, 0, num_experts, stream); \
-            break;                                                               \
-        case 64:                                                                 \
-            topkGatingSoftmaxLauncherHelper<NUM_EXPERTS, WARPS_PER_TB, 64>(      \
-                gating_output, nullptr, topk_weights, topk_indices,              \
-                token_expert_indices, num_tokens, topk, 0, num_experts, stream); \
-            break;                                                               \
-        default:                                                                 \
-            TORCH_CHECK(false, "Unsupported warp size: ", warpSize);             \
+#ifndef USE_ROCM
+#define LAUNCH_SOFTMAX(NUM_EXPERTS, WARPS_PER_TB, MAX_BYTES)                          \
+    static_assert(WARP_SIZE == 32,                                                    \
+                  "Unsupported warp size. Only 32 is supported for CUDA");            \
+    topkGatingSoftmaxLauncherHelper<NUM_EXPERTS, WARPS_PER_TB, WARP_SIZE, MAX_BYTES>( \
+        gating_output, nullptr, topk_weights, topk_indices,                           \
+        token_expert_indices, num_tokens, topk, 0, num_experts, stream);
+#else
+#define LAUNCH_SOFTMAX(NUM_EXPERTS, WARPS_PER_TB, MAX_BYTES)                             \
+    if (WARP_SIZE == 64) {                                                               \
+        topkGatingSoftmaxLauncherHelper<NUM_EXPERTS, WARPS_PER_TB, 64, MAX_BYTES>(       \
+            gating_output, nullptr, topk_weights, topk_indices,                          \
+            token_expert_indices, num_tokens, topk, 0, num_experts, stream);             \
+    } else if (WARP_SIZE == 32) {                                                        \
+        topkGatingSoftmaxLauncherHelper<NUM_EXPERTS, WARPS_PER_TB, 32, MAX_BYTES>(       \
+            gating_output, nullptr, topk_weights, topk_indices,                          \
+            token_expert_indices, num_tokens, topk, 0, num_experts, stream);             \
+    } else {                                                                             \
+        assert(false && "Unsupported warp size. Only 32 and 64 are supported for ROCm"); \
     }
+#endif
 
 template <typename IndType>
 void topkGatingSoftmaxKernelLauncher(
@@ -451,38 +457,64 @@ void topkGatingSoftmaxKernelLauncher(
     const int topk,
     cudaStream_t stream) {
     static constexpr int WARPS_PER_TB = 4;
-    auto warpSize = WARP_SIZE;
+    static constexpr int BYTES_PER_LDG_POWER_OF_2 = 16;
+#ifndef USE_ROCM
+    static constexpr int BYTES_PER_LDG_MULTIPLE_64 = 8;
+#endif
     switch (num_experts) {
         case 1:
-            LAUNCH_SOFTMAX(1, WARPS_PER_TB);
+            LAUNCH_SOFTMAX(1, WARPS_PER_TB, BYTES_PER_LDG_POWER_OF_2);
             break;
         case 2:
-            LAUNCH_SOFTMAX(2, WARPS_PER_TB);
+            LAUNCH_SOFTMAX(2, WARPS_PER_TB, BYTES_PER_LDG_POWER_OF_2);
             break;
         case 4:
-            LAUNCH_SOFTMAX(4, WARPS_PER_TB);
+            LAUNCH_SOFTMAX(4, WARPS_PER_TB, BYTES_PER_LDG_POWER_OF_2);
             break;
         case 8:
-            LAUNCH_SOFTMAX(8, WARPS_PER_TB);
+            LAUNCH_SOFTMAX(8, WARPS_PER_TB, BYTES_PER_LDG_POWER_OF_2);
             break;
         case 16:
-            LAUNCH_SOFTMAX(16, WARPS_PER_TB);
+            LAUNCH_SOFTMAX(16, WARPS_PER_TB, BYTES_PER_LDG_POWER_OF_2);
             break;
         case 32:
-            LAUNCH_SOFTMAX(32, WARPS_PER_TB);
+            LAUNCH_SOFTMAX(32, WARPS_PER_TB, BYTES_PER_LDG_POWER_OF_2);
             break;
         case 64:
-            LAUNCH_SOFTMAX(64, WARPS_PER_TB);
+            LAUNCH_SOFTMAX(64, WARPS_PER_TB, BYTES_PER_LDG_POWER_OF_2);
             break;
         case 128:
-            LAUNCH_SOFTMAX(128, WARPS_PER_TB);
+            LAUNCH_SOFTMAX(128, WARPS_PER_TB, BYTES_PER_LDG_POWER_OF_2);
             break;
         case 256:
-            LAUNCH_SOFTMAX(256, WARPS_PER_TB);
+            LAUNCH_SOFTMAX(256, WARPS_PER_TB, BYTES_PER_LDG_POWER_OF_2);
             break;
+        case 512:
+            LAUNCH_SOFTMAX(512, WARPS_PER_TB, BYTES_PER_LDG_POWER_OF_2);
+            break;
+        // (CUDA only) support multiples of 64 when num_experts is not power of 2.
+        // ROCm uses WARP_SIZE 64 so 8 bytes loading won't fit for some of num_experts,
+        // alternatively we can test 4 bytes loading and enable it in future.
+#ifndef USE_ROCM
+        case 192:
+            LAUNCH_SOFTMAX(192, WARPS_PER_TB, BYTES_PER_LDG_MULTIPLE_64);
+            break;
+        case 320:
+            LAUNCH_SOFTMAX(320, WARPS_PER_TB, BYTES_PER_LDG_MULTIPLE_64);
+            break;
+        case 384:
+            LAUNCH_SOFTMAX(384, WARPS_PER_TB, BYTES_PER_LDG_MULTIPLE_64);
+            break;
+        case 448:
+            LAUNCH_SOFTMAX(448, WARPS_PER_TB, BYTES_PER_LDG_MULTIPLE_64);
+            break;
+        case 576:
+            LAUNCH_SOFTMAX(576, WARPS_PER_TB, BYTES_PER_LDG_MULTIPLE_64);
+            break;
+#endif
         default: {
             TORCH_CHECK(softmax_workspace != nullptr,
-                "softmax_workspace must be provided for num_experts that are not a power of 2.");
+                "softmax_workspace must be provided for num_experts that are not a power of 2 or multiple of 64.");
             static constexpr int TPB = 256;
             moeSoftmax<TPB><<<num_tokens, TPB, 0, stream>>>(
                 gating_output, nullptr, softmax_workspace, num_experts);

csrc/moe/torch_bindings.cpp

@@ -35,7 +35,8 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, m) {
 
   m.def(
       "moe_wna16_marlin_gemm(Tensor! a, Tensor? c_or_none,"
-      "Tensor! b_q_weight, Tensor! b_scales, Tensor? global_scale, Tensor? "
+      "Tensor! b_q_weight, Tensor? b_bias_or_none,"
+      "Tensor! b_scales, Tensor? global_scale, Tensor? "
       "b_zeros_or_none,"
       "Tensor? g_idx_or_none, Tensor? perm_or_none, Tensor! workspace,"
       "Tensor sorted_token_ids,"

csrc/ops.h

@@ -138,6 +138,8 @@ void gelu_tanh_and_mul(torch::Tensor& out, torch::Tensor& input);
 
 void fatrelu_and_mul(torch::Tensor& out, torch::Tensor& input,
                      double threshold);
+void swigluoai_and_mul(torch::Tensor& out, torch::Tensor& input,
+                       double alpha = 1.702, double limit = 7.0);
 
 void gelu_new(torch::Tensor& out, torch::Tensor& input);
 
@@ -145,22 +147,6 @@ void gelu_fast(torch::Tensor& out, torch::Tensor& input);
 
 void gelu_quick(torch::Tensor& out, torch::Tensor& input);
 
-void advance_step_flashattn(int64_t num_seqs, int64_t num_queries,
-                            int64_t block_size, torch::Tensor& input_tokens,
-                            torch::Tensor& sampled_token_ids,
-                            torch::Tensor& input_positions,
-                            torch::Tensor& seq_lens,
-                            torch::Tensor& slot_mapping,
-                            torch::Tensor& block_tables);
-
-void advance_step_flashinfer(
-    int64_t num_seqs, int64_t num_queries, int64_t block_size,
-    torch::Tensor& input_tokens, torch::Tensor& sampled_token_ids,
-    torch::Tensor& input_positions, torch::Tensor& seq_lens,
-    torch::Tensor& slot_mapping, torch::Tensor& block_tables,
-    torch::Tensor& paged_kv_indices, torch::Tensor& paged_kv_indptr,
-    torch::Tensor& paged_kv_last_page_len, torch::Tensor& block_table_bounds);
-
 void cutlass_mla_decode(torch::Tensor const& out, torch::Tensor const& q_nope,
                         torch::Tensor const& q_pe,
                         torch::Tensor const& kv_c_and_k_pe_cache,
@@ -170,15 +156,6 @@ void cutlass_mla_decode(torch::Tensor const& out, torch::Tensor const& q_nope,
 torch::Tensor get_cuda_view_from_cpu_tensor(torch::Tensor& cpu_tensor);
 
 #ifndef USE_ROCM
-torch::Tensor aqlm_gemm(const torch::Tensor& input, const torch::Tensor& codes,
-                        const torch::Tensor& codebooks,
-                        const torch::Tensor& scales,
-                        const std::vector<int64_t>& codebook_partition_sizes,
-                        const std::optional<torch::Tensor>& bias);
-
-torch::Tensor aqlm_dequant(
-    const torch::Tensor& codes, const torch::Tensor& codebooks,
-    const std::vector<int64_t>& codebook_partition_sizes);
 
 torch::Tensor awq_gemm(torch::Tensor _in_feats, torch::Tensor _kernel,
                        torch::Tensor _scaling_factors, torch::Tensor _zeros,

csrc/prepare_inputs/advance_step.cu

@@ -1,336 +0,0 @@
-/*
- * The goal of this GPU kernel is to advance input tensors on the GPU directly
- * PR: https://github.com/vllm-project/vllm/pull/6338
- * Current restrictions:
- *     1. Specialized for DraftModelRunner
- *     2. Supports flash_attn only
- */
-
-#include "advance_step.cuh"
-
-namespace prepare_inputs {
-
-//
-template <int const num_threads>
-__global__ void advance_step_flashattn_kernel(
-    int num_seqs, int num_queries, int block_size, long* input_tokens_ptr,
-    long const* sampled_token_ids_ptr, long* input_positions_ptr,
-    int* seq_lens_ptr, long* slot_mapping_ptr, int const* block_tables_ptr,
-    int64_t const block_tables_stride) {
-  int const n_pad = num_seqs - num_queries;
-  if (n_pad && blockIdx.x == 0) {
-    // Handle cuda graph padding
-    int const offset = num_queries;
-    for (int i = threadIdx.x; i < n_pad; i += blockDim.x) {
-      input_tokens_ptr[offset + i] = 0;
-      input_positions_ptr[offset + i] = 0;
-      slot_mapping_ptr[offset + i] = -1;
-    }
-  }
-
-  int num_query_blocks = div_ceil(num_queries, num_threads);
-
-  if (blockIdx.x >= num_query_blocks) {
-    return;
-  }
-
-  int cur_query_id = blockIdx.x * num_threads + threadIdx.x;
-
-  if (cur_query_id >= num_queries) {
-    return;
-  }
-
-  // Update input_tokens
-  input_tokens_ptr[cur_query_id] = sampled_token_ids_ptr[cur_query_id];
-
-  int seq_len = seq_lens_ptr[cur_query_id];
-  int next_seq_len = seq_len + 1;
-  int next_input_pos = next_seq_len - 1;
-
-  // Update seq_lens
-  seq_lens_ptr[cur_query_id] = next_seq_len;
-  // Update input_positions
-  input_positions_ptr[cur_query_id] = next_input_pos;
-
-  int const* seq_block_tables_ptr =
-      block_tables_ptr + block_tables_stride * cur_query_id;
-
-  int block_index = next_input_pos / block_size;
-  int block_offset = next_input_pos % block_size;
-
-  int slot_num = seq_block_tables_ptr[block_index] * block_size + block_offset;
-  // Update slot_mapping
-  slot_mapping_ptr[cur_query_id] = slot_num;
-}
-
-inline void verify_tensor(std::string const& name, torch::Tensor const& t,
-                          int64_t const size_0, int64_t const size_1,
-                          c10::ScalarType const type) {
-  bool size_0_cond = true;
-  if (size_0 != -1) {
-    size_0_cond = t.size(0) == size_0;
-  }
-
-  bool size_1_cond = true;
-  if (size_1 != -1) {
-    size_1_cond = t.size(1) == size_1;
-  }
-
-  bool is_contiguous = t.is_contiguous();
-  bool same_type = t.dtype() == type;
-
-  bool pass = size_0_cond && size_1_cond && is_contiguous && same_type;
-  if (!pass) {
-    TORCH_CHECK(false, "tensor: name = ", name, ", shape = ", t.sizes(),
-                " is_cont = ", t.is_contiguous(), ", type = ", t.dtype(),
-                " is not as expected: shape = [", size_0, ", ", size_1,
-                "], type = ", type);
-  }
-}
-
-/// each thread processes a block per query
-__global__ void advance_step_flashinfer_kernel(
-    int num_threads, int num_seqs, int num_queries, int block_size,
-    long* input_tokens_ptr, long const* sampled_token_ids_ptr,
-    long* input_positions_ptr, int* seq_lens_ptr, long* slot_mapping_ptr,
-    int const* block_tables_ptr, int64_t const block_tables_stride,
-    int* paged_kv_last_page_len_ptr, int* block_table_bound_ptr) {
-  int const n_pad = num_seqs - num_queries;
-  if (n_pad && blockIdx.x == 0) {
-    // Handle cuda graph padding
-    int const offset = num_queries;
-    for (int i = threadIdx.x; i < n_pad; i += blockDim.x) {
-      input_tokens_ptr[offset + i] = 0;
-      input_positions_ptr[offset + i] = 0;
-      slot_mapping_ptr[offset + i] = -1;
-    }
-  }
-  int num_query_blocks = div_ceil(num_queries, num_threads);
-
-  if (blockIdx.x < num_query_blocks) {
-    int cur_query_id = blockIdx.x * num_threads + threadIdx.x;
-
-    if (cur_query_id < num_queries) {
-      // Update input_tokens
-      input_tokens_ptr[cur_query_id] = sampled_token_ids_ptr[cur_query_id];
-
-      int seq_len = seq_lens_ptr[cur_query_id];
-      int next_seq_len = seq_len + 1;
-      int next_input_pos = next_seq_len - 1;
-
-      // Update seq_lens
-      seq_lens_ptr[cur_query_id] = next_seq_len;
-      // Update input_positions
-      input_positions_ptr[cur_query_id] = next_input_pos;
-
-      int const* seq_block_tables_ptr =
-          block_tables_ptr + block_tables_stride * cur_query_id;
-
-      int block_index = next_input_pos / block_size;
-      int block_offset = next_input_pos % block_size;
-
-      // Update paged_kv_last_page_len
-      paged_kv_last_page_len_ptr[cur_query_id] = block_offset + 1;
-
-      int slot_num =
-          seq_block_tables_ptr[block_index] * block_size + block_offset;
-      // Update slot_mapping
-      slot_mapping_ptr[cur_query_id] = slot_num;
-      block_table_bound_ptr[cur_query_id] = div_ceil(next_seq_len, block_size);
-    }
-  }
-}
-
-__global__ void advance_step_flashinfer_indptr_kernel(
-    int num_threads, int num_seqs, int num_queries, int* paged_kv_indptr_ptr,
-    int* block_table_bound_ptr) {
-  int idx = blockIdx.x * num_threads + threadIdx.x;
-  // Update paged_kv_indptr
-  if (idx == 0) {
-    paged_kv_indptr_ptr[idx] = 0;
-  }
-  if (idx < num_queries) {
-    int sum = 0;
-    for (int i = 0; i <= idx; ++i) {
-      sum += block_table_bound_ptr[i];
-    }
-    paged_kv_indptr_ptr[idx + 1] = sum;
-  }
-}
-
-__global__ void advance_step_flashinfer_indices_kernel(
-    int num_seqs, int num_queries, int const* block_tables_ptr,
-    int64_t const max_num_blocks_per_seq, int* paged_kv_indices_ptr,
-    int* paged_kv_indptr_ptr, int* block_table_bound_ptr) {
-  // note: max_num_blocks_per_seq = block_tables.stride(0)
-  int tid = blockIdx.x * blockDim.x + threadIdx.x;
-
-  // when cuda graphs are enabled, paged_kv_indptr tensor
-  // has to be updated for the padded queries
-  // tid represents a query# for paged_kv_indptr tensor
-  if (num_queries < tid && tid <= num_seqs) {
-    paged_kv_indptr_ptr[tid] = paged_kv_indptr_ptr[num_queries];
-  }
-
-  // each thread processes a block_ptr in block_tables
-  // block_tables shape: [num_queries, max_num_blocks_per_seq]
-  // paged_kv_indices is flattened block_tables.
-  for (int idx = tid; idx < (num_seqs * max_num_blocks_per_seq);
-       idx += (gridDim.x * blockDim.x)) {
-    // block_tables-row = paged_kv_indptr[queryNum]
-    int queryNum = idx / max_num_blocks_per_seq;
-    int col = idx % max_num_blocks_per_seq;
-    if (queryNum < num_queries && col < block_table_bound_ptr[queryNum]) {
-      int indices_arr_idx = paged_kv_indptr_ptr[queryNum] + col;
-      int block_tables_idx = queryNum * max_num_blocks_per_seq + col;
-      paged_kv_indices_ptr[indices_arr_idx] =
-          block_tables_ptr[block_tables_idx];
-    }
-  }
-}
-
-void advance_step_flashattn(int num_seqs, int num_queries, int block_size,
-                            torch::Tensor& input_tokens,       // type: long
-                            torch::Tensor& sampled_token_ids,  // type: long
-                            torch::Tensor& input_positions,    // type: long
-                            torch::Tensor& seq_lens,           // type: int
-                            torch::Tensor& slot_mapping,       // type: long
-                            torch::Tensor& block_tables) {     // type: int
-
-  if (logging) {
-    printf("advance_step_flashattn:\n");
-    printf("  num_seqs = %d\n", num_seqs);
-    printf("  num_queries = %d\n", num_queries);
-    printf("  block_size = %d\n", block_size);
-  }
-  // Verify all tensors
-  verify_tensor("input_tokens", input_tokens, num_seqs, -1, at::kLong);
-  verify_tensor("sampled_token_ids", sampled_token_ids, num_queries, 1,
-                at::kLong);
-  verify_tensor("input_positions", input_positions, num_seqs, -1, at::kLong);
-  verify_tensor("seq_lens", seq_lens, num_seqs, -1, at::kInt);
-  verify_tensor("slot_mapping", slot_mapping, num_seqs, -1, at::kLong);
-  verify_tensor("block_tables", block_tables, num_seqs, -1, at::kInt);
-
-  int dev = sampled_token_ids.get_device();
-  cudaStream_t stream = at::cuda::getCurrentCUDAStream(dev);
-
-  int blocks;
-  cudaDeviceGetAttribute(&blocks, cudaDevAttrMultiProcessorCount, dev);
-
-  advance_step_flashattn_kernel<max_threads>
-      <<<blocks, max_threads, 0, stream>>>(
-          num_seqs, num_queries, block_size,
-          reinterpret_cast<long*>(input_tokens.data_ptr()),
-          reinterpret_cast<long const*>(sampled_token_ids.data_ptr()),
-          reinterpret_cast<long*>(input_positions.data_ptr()),
-          reinterpret_cast<int*>(seq_lens.data_ptr()),
-          reinterpret_cast<long*>(slot_mapping.data_ptr()),
-          reinterpret_cast<int const*>(block_tables.data_ptr()),
-          block_tables.stride(0));
-}
-
-void advance_step_flashinfer(
-    int num_seqs, int num_queries, int block_size,
-    torch::Tensor& input_tokens,            // type: long
-    torch::Tensor& sampled_token_ids,       // type: long
-    torch::Tensor& input_positions,         // type: long
-    torch::Tensor& seq_lens,                // type: int
-    torch::Tensor& slot_mapping,            // type: long
-    torch::Tensor& block_tables,            // type: int
-    torch::Tensor& paged_kv_indices,        // type: int
-    torch::Tensor& paged_kv_indptr,         // type: int
-    torch::Tensor& paged_kv_last_page_len,  // type: int
-    torch::Tensor& block_table_bound) {     // type: int
-
-  if (logging) {
-    printf("advance_step_flashinfer:\n");
-    printf("  num_seqs = %d\n", num_seqs);
-    printf("  num_queries = %d\n", num_queries);
-    printf("  block_size = %d\n", block_size);
-    printf("  block_tables.stride(0) = %zu\n", block_tables.stride(0));
-  }
-  // Verify all tensors
-  verify_tensor("input_tokens", input_tokens, num_seqs, -1, at::kLong);
-  // verify_tensor("sampled_token_ids", sampled_token_ids, num_queries, 1,
-  //               at::kLong);
-  verify_tensor("input_positions", input_positions, num_seqs, -1, at::kLong);
-  verify_tensor("seq_lens", seq_lens, num_seqs, -1, at::kInt);
-  verify_tensor("slot_mapping", slot_mapping, num_seqs, -1, at::kLong);
-  verify_tensor("block_tables", block_tables, num_seqs, -1, at::kInt);
-
-  verify_tensor("paged_kv_indices", paged_kv_indices, -1, -1, at::kInt);
-  verify_tensor("paged_kv_indptr", paged_kv_indptr, num_seqs + 1, -1, at::kInt);
-  verify_tensor("paged_kv_last_page_len", paged_kv_last_page_len, num_seqs, -1,
-                at::kInt);
-
-  verify_tensor("block_table_bound", block_table_bound, num_seqs, -1, at::kInt);
-
-  int dev = sampled_token_ids.get_device();
-  cudaStream_t stream = at::cuda::getCurrentCUDAStream(dev);
-
-  int blocks;
-  int threads;
-  cudaDeviceGetAttribute(&blocks, cudaDevAttrMultiProcessorCount, dev);
-  cudaDeviceGetAttribute(&threads, cudaDevAttrMaxThreadsPerBlock, dev);
-
-  TORCH_CHECK((blocks * threads > num_queries),
-              "multi-step: not enough threads to map to num_queries = ",
-              num_queries, " block_tables.stride(0) = ", block_tables.stride(0),
-              " blocks = ", blocks, " max_threads = ", threads);
-  if (logging) {
-    printf("launching kernels with %d blocks and %d threads\n", blocks,
-           threads);
-  }
-  advance_step_flashinfer_kernel<<<blocks, threads, 0, stream>>>(
-      threads, num_seqs, num_queries, block_size,
-      reinterpret_cast<long*>(input_tokens.data_ptr()),
-      reinterpret_cast<long const*>(sampled_token_ids.data_ptr()),
-      reinterpret_cast<long*>(input_positions.data_ptr()),
-      reinterpret_cast<int*>(seq_lens.data_ptr()),
-      reinterpret_cast<long*>(slot_mapping.data_ptr()),
-      reinterpret_cast<int const*>(block_tables.data_ptr()),
-      block_tables.stride(0),
-      reinterpret_cast<int*>(paged_kv_last_page_len.data_ptr()),
-      reinterpret_cast<int*>(block_table_bound.data_ptr()));
-
-  advance_step_flashinfer_indptr_kernel<<<blocks, threads, 0, stream>>>(
-      threads, num_seqs, num_queries,
-      reinterpret_cast<int*>(paged_kv_indptr.data_ptr()),
-      reinterpret_cast<int*>(block_table_bound.data_ptr()));
-
-  advance_step_flashinfer_indices_kernel<<<blocks, threads, 0, stream>>>(
-      num_seqs, num_queries,
-      reinterpret_cast<int const*>(block_tables.data_ptr()),
-      block_tables.stride(0),
-      reinterpret_cast<int*>(paged_kv_indices.data_ptr()),
-      reinterpret_cast<int*>(paged_kv_indptr.data_ptr()),
-      reinterpret_cast<int*>(block_table_bound.data_ptr()));
-}
-
-}  // namespace prepare_inputs
-
-void advance_step_flashattn(int64_t num_seqs, int64_t num_queries,
-                            int64_t block_size, torch::Tensor& input_tokens,
-                            torch::Tensor& sampled_token_ids,
-                            torch::Tensor& input_positions,
-                            torch::Tensor& seq_lens,
-                            torch::Tensor& slot_mapping,
-                            torch::Tensor& block_tables) {
-  prepare_inputs::advance_step_flashattn(
-      num_seqs, num_queries, block_size, input_tokens, sampled_token_ids,
-      input_positions, seq_lens, slot_mapping, block_tables);
-}
-
-void advance_step_flashinfer(
-    int64_t num_seqs, int64_t num_queries, int64_t block_size,
-    torch::Tensor& input_tokens, torch::Tensor& sampled_token_ids,
-    torch::Tensor& input_positions, torch::Tensor& seq_lens,
-    torch::Tensor& slot_mapping, torch::Tensor& block_tables,
-    torch::Tensor& paged_kv_indices, torch::Tensor& paged_kv_indptr,
-    torch::Tensor& paged_kv_last_page_len, torch::Tensor& block_table_bound) {
-  prepare_inputs::advance_step_flashinfer(
-      num_seqs, num_queries, block_size, input_tokens, sampled_token_ids,
-      input_positions, seq_lens, slot_mapping, block_tables, paged_kv_indices,
-      paged_kv_indptr, paged_kv_last_page_len, block_table_bound);
-}

csrc/prepare_inputs/advance_step.cuh

@@ -1,19 +0,0 @@
-#pragma once
-
-#include <torch/all.h>
-
-#include <ATen/cuda/CUDAContext.h>
-#include <c10/cuda/CUDAGuard.h>
-#include <cuda.h>
-#include <cuda_fp16.h>
-#include <cuda_runtime.h>
-#include <iostream>
-
-namespace prepare_inputs {
-
-static constexpr int max_threads = 256;
-static constexpr bool logging = false;
-
-constexpr int div_ceil(int a, int b) { return (a + b - 1) / b; }
-
-}  // namespace prepare_inputs

csrc/quantization/aqlm/gemm_kernels.cu

@@ -1,597 +0,0 @@
-/*
- * Modified by Neural Magic
- * Adapted from https://github.com/Vahe1994/AQLM
- *
- * 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.
- */
-
-#include <cuda.h>
-#include <cuda_fp16.h>
-#include <cuda_runtime.h>
-#include <torch/all.h>
-#include <c10/cuda/CUDAStream.h>
-#include <c10/cuda/CUDAGuard.h>
-
-#include <iostream>
-#include <cstdlib>
-
-namespace vllm {
-namespace aqlm {
-
-__global__ void Code1x16MatVec(
-    const int4* __restrict__ A, const int4* __restrict__ B,
-    int4* __restrict__ C, const int4* __restrict__ codebook, const int prob_m,
-    const int prob_k,
-    const int4 codebook_a_sizes,  // cumulative sizes of A spanning each
-                                  // codebook, at most 3 long.
-    const int codebook_stride     // as int4.
-) {
-  int a_gl_stride = prob_k / 8 / 8;
-  int a_gl_rd = (blockDim.x / 32) * blockIdx.x + (threadIdx.x / 32);
-  bool pred = a_gl_rd < prob_m;
-
-  if (pred) {
-    // advance to the correct codebook, this easy because we only multiply one
-    // column of the codebook.
-    auto codebook_size = &codebook_a_sizes.x;
-    while (a_gl_rd >= *codebook_size) {
-      codebook += codebook_stride;
-      ++codebook_size;
-    }
-  }
-
-  int b_gl_rd = 0;
-  int c_gl_wr = a_gl_rd;
-  a_gl_rd = a_gl_stride * a_gl_rd + threadIdx.x % 32;
-  int a_gl_end = a_gl_rd + a_gl_stride - threadIdx.x % 32;
-
-  __shared__ int4 sh_b[32 * 9];
-  float res = 0;
-
-  int iters = (prob_k / 8 + 8 * 32 - 1) / (8 * 32);
-  while (iters--) {
-    // We pad shared memory to avoid bank conflicts during reads
-    __syncthreads();
-    for (int i = threadIdx.x; i < 32 * 8; i += blockDim.x) {
-      if (b_gl_rd + i < prob_k / 8) sh_b[9 * (i / 8) + i % 8] = B[b_gl_rd + i];
-    }
-    __syncthreads();
-    b_gl_rd += 32 * 8;
-
-    int b_sh_rd = 9 * (threadIdx.x % 32);
-    if (pred && a_gl_rd < a_gl_end) {
-      const uint16_t* enc = reinterpret_cast<const uint16_t*>(&A[a_gl_rd]);
-#pragma unroll
-      for (int i = 0; i < 8; i++) {
-        uint32_t dec[4];
-        // We bypass the L1 cache to avoid massive amounts of memory streaming
-        // that doesn't actually help us; this brings > 2x speedup.
-        asm volatile("ld.cg.global.v4.u32 {%0, %1, %2, %3}, [%4];"
-                     : "=r"(dec[0]), "=r"(dec[1]), "=r"(dec[2]), "=r"(dec[3])
-                     : "l"((void*)&codebook[enc[i]]));
-        half2* a = reinterpret_cast<half2*>(&dec);
-        half2* b = reinterpret_cast<half2*>(&sh_b[b_sh_rd]);
-        half2 res2 = {};
-#pragma unroll
-        for (int j = 0; j < 4; j++) res2 = __hfma2(a[j], b[j], res2);
-        res += __half2float(res2.x) + __half2float(res2.y);
-        b_sh_rd++;
-      }
-      a_gl_rd += 32;
-    }
-  }
-
-  if (pred) {
-#pragma unroll
-    for (int i = 16; i > 0; i /= 2) res += __shfl_down_sync(0xffffffff, res, i);
-    if (threadIdx.x % 32 == 0)
-      reinterpret_cast<__half*>(C)[c_gl_wr] = __float2half(res);
-  }
-}
-
-__global__ void Code2x8MatVec(
-    const int4* __restrict__ A, const int4* __restrict__ B,
-    int4* __restrict__ C, const int4* __restrict__ codebook, int prob_m,
-    int prob_k,
-    const int4 codebook_a_sizes,  // cumulative sizes of A spanning each
-                                  // codebook, at most 3 long.
-    const int codebook_stride     // as int4.
-
-) {
-  int a_gl_stride = prob_k / 8 / 8;
-  int a_gl_rd = (blockDim.x / 32) * blockIdx.x + (threadIdx.x / 32);
-  bool pred = a_gl_rd < prob_m;
-
-  if (pred) {
-    // advance to the correct codebook, this easy because we only multiply one
-    // column of the codebook.
-    auto codebook_size = &codebook_a_sizes.x;
-    while (a_gl_rd >= *codebook_size) {
-      codebook += codebook_stride;
-      ++codebook_size;
-    }
-  }
-
-  int b_gl_rd = 0;
-  int c_gl_wr = a_gl_rd;
-  a_gl_rd = a_gl_stride * a_gl_rd + threadIdx.x % 32;
-  int a_gl_end = a_gl_rd + a_gl_stride - threadIdx.x % 32;
-  int lane = threadIdx.x % 8;
-
-  extern __shared__ int4 sh[];
-  int4* sh_b = sh;
-  int4* sh_code = sh_b + 32 * 9;
-  int4* sh_code0 = sh_code;
-  int4* sh_code1 = sh_code + 256 * 8;
-
-  for (int i = threadIdx.x; i < 2 * 256; i += blockDim.x) {
-    int4 dec = codebook[i];
-#pragma unroll
-    for (int j = 0; j < 8; j++) sh_code[8 * i + (j + lane) % 8] = dec;
-  }
-  __syncthreads();
-
-  float res = 0;
-
-  int iters = (prob_k / 8 + 8 * 32 - 1) / (8 * 32);
-  while (iters--) {
-    // We pad shared memory to avoid bank conflicts during reads
-    __syncthreads();
-    for (int i = threadIdx.x; i < 32 * 8; i += blockDim.x) {
-      if (b_gl_rd + i < prob_k / 8) sh_b[9 * (i / 8) + i % 8] = B[b_gl_rd + i];
-    }
-    __syncthreads();
-    b_gl_rd += 32 * 8;
-
-    int b_sh_rd = 9 * (threadIdx.x % 32);
-    if (pred && a_gl_rd < a_gl_end) {
-      const uint8_t* enc = reinterpret_cast<const uint8_t*>(&A[a_gl_rd]);
-#pragma unroll
-      for (int i = 0; i < 8; i++) {
-        half2* a0 =
-            reinterpret_cast<half2*>(&sh_code0[8 * enc[2 * i + 0] + lane]);
-        half2* a1 =
-            reinterpret_cast<half2*>(&sh_code1[8 * enc[2 * i + 1] + lane]);
-        half2* b = reinterpret_cast<half2*>(&sh_b[b_sh_rd]);
-        half2 res2 = {};
-#pragma unroll
-        for (int j = 0; j < 4; j++)
-          res2 = __hfma2(__hadd2(a0[j], a1[j]), b[j], res2);
-        res += __half2float(res2.x) + __half2float(res2.y);
-        b_sh_rd++;
-      }
-      a_gl_rd += 32;
-    }
-  }
-
-  if (pred) {
-#pragma unroll
-    for (int i = 16; i > 0; i /= 2) res += __shfl_down_sync(0xffffffff, res, i);
-    if (threadIdx.x % 32 == 0)
-      reinterpret_cast<__half*>(C)[c_gl_wr] = __float2half(res);
-  }
-}
-
-__global__ void Code1x16Dequant(
-    const int4* __restrict__ A, int4* __restrict__ C,
-    const int4* __restrict__ codebook, int prob_m, int prob_k,
-    const int4 codebook_a_sizes,  // cumulative sizes of A spanning each
-                                  // codebook, at most 3 long, sums to m.
-    const int codebook_stride     // as int4
-) {
-  int a_gl_stride = prob_k / 8 / 8;
-  int a_gl_rd = (blockDim.x / 32) * blockIdx.x + (threadIdx.x / 32);
-  bool pred = a_gl_rd < prob_m;
-
-  if (pred) {
-    // advance to the correct codebook, this easy because we only multiply one
-    // column of the codebook.
-    auto codebook_size = &codebook_a_sizes.x;
-    while (a_gl_rd >= *codebook_size) {
-      codebook += codebook_stride;
-      ++codebook_size;
-    }
-  }
-
-  a_gl_rd = a_gl_stride * a_gl_rd + threadIdx.x % 32;
-  int a_gl_end = a_gl_rd + a_gl_stride - threadIdx.x % 32;
-
-  int c_gl_stride = prob_k / 8;
-  int c_gl_wr = (blockDim.x / 32) * blockIdx.x + (threadIdx.x / 32);
-  c_gl_wr = c_gl_stride * c_gl_wr + (threadIdx.x % 32) * 8;
-
-  int iters = (prob_k / 8 - 1) / (8 * 32) + 1;
-  while (iters--) {
-    if (pred && a_gl_rd < a_gl_end) {
-      const uint16_t* enc = reinterpret_cast<const uint16_t*>(&A[a_gl_rd]);
-#pragma unroll
-      for (int i = 0; i < 8; i++) {
-        int4 chunk;
-        auto dec = reinterpret_cast<uint32_t*>(&chunk);
-        // We bypass the L1 cache to avoid massive amounts of memory streaming
-        // that doesn't actually help us; this brings > 2x speedup.
-        asm volatile("ld.cg.global.v4.u32 {%0, %1, %2, %3}, [%4];"
-                     : "=r"(dec[0]), "=r"(dec[1]), "=r"(dec[2]), "=r"(dec[3])
-                     : "l"((void*)&codebook[enc[i]]));
-
-        C[a_gl_rd * 8 + i] = chunk;
-      }
-    }
-    a_gl_rd += 32;
-  }
-}
-
-__global__ void Code2x8Dequant(
-    const int4* __restrict__ A, int4* __restrict__ C,
-    const int4* __restrict__ codebook, int prob_m, int prob_k,
-    const int4
-        codebook_a_sizes,  // cumulative sizes of A spanning each codebook, at
-                           // most 3 long, corresponds to cols.
-    const int codebook_stride  // as int4
-) {
-  int a_gl_stride = prob_k / 8 / 8;
-  int a_gl_rd = (blockDim.x / 32) * blockIdx.x + (threadIdx.x / 32);
-  bool pred = a_gl_rd < prob_m;
-
-  if (pred) {
-    // advance to the correct codebook, this easy because we only multiply one
-    // column of the codebook.
-    auto codebook_size = &codebook_a_sizes.x;
-    while (a_gl_rd >= *codebook_size) {
-      codebook += codebook_stride;
-      ++codebook_size;
-    }
-  }
-
-  a_gl_rd = a_gl_stride * a_gl_rd + threadIdx.x % 32;
-  int a_gl_end = a_gl_rd + a_gl_stride - threadIdx.x % 32;
-  int lane = threadIdx.x % 8;
-
-  int c_gl_stride = prob_k / 8;
-  int c_gl_wr = (blockDim.x / 32) * blockIdx.x + (threadIdx.x / 32);
-  c_gl_wr = c_gl_stride * c_gl_wr + (threadIdx.x % 32) * 8;
-
-  extern __shared__ int4 sh[];
-  int4* sh_code = sh;
-  int4* sh_code0 = sh_code;
-  int4* sh_code1 = sh_code + 256 * 8;
-
-  for (int i = threadIdx.x; i < 2 * 256; i += blockDim.x) {
-    int4 dec = codebook[i];
-#pragma unroll
-    for (int j = 0; j < 8; j++) sh_code[8 * i + (j + lane) % 8] = dec;
-  }
-  __syncthreads();
-
-  int iters = (prob_k / 8 - 1) / (8 * 32) + 1;
-  while (iters--) {
-    if (pred && a_gl_rd < a_gl_end) {
-      const uint8_t* enc = reinterpret_cast<const uint8_t*>(&A[a_gl_rd]);
-#pragma unroll
-      for (int i = 0; i < 8; i++) {
-        int4 chunk;
-        half2* a0 =
-            reinterpret_cast<half2*>(&sh_code0[8 * enc[2 * i + 0] + lane]);
-        half2* a1 =
-            reinterpret_cast<half2*>(&sh_code1[8 * enc[2 * i + 1] + lane]);
-#pragma unroll
-        for (int j = 0; j < 4; j++)
-          reinterpret_cast<half2*>(&chunk)[j] = __hadd2(a0[j], a1[j]);
-        C[a_gl_rd * 8 + i] = chunk;
-      }
-    }
-    a_gl_rd += 32;
-  }
-}
-
-inline int ceildiv(int a, int b) { return (a + b - 1) / b; }
-
-const int THREAD_M = 16;
-
-void code1x16_matvec_cuda(const void* __restrict__ A,
-                          const void* __restrict__ B, void* __restrict__ C,
-                          const void* __restrict__ codebook, int prob_m,
-                          int prob_k, const int4 codebook_a_sizes,
-                          const int codebook_stride) {
-  int sms;
-  cudaDeviceGetAttribute(&sms, cudaDevAttrMultiProcessorCount, 0);
-  int waves = 0;
-  int thread_m;
-  do {
-    waves++;
-    thread_m = ceildiv(prob_m, waves * sms);
-  } while (thread_m > THREAD_M);
-
-  int blocks = ceildiv(prob_m, thread_m);
-  int threads = 32 * thread_m;
-  cudaStream_t stream = at::cuda::getCurrentCUDAStream().stream();
-  Code1x16MatVec<<<blocks, threads, 16 * 32 * 9, stream>>>(
-      (const int4*)A, (const int4*)B, (int4*)C, (const int4*)codebook, prob_m,
-      prob_k, codebook_a_sizes, codebook_stride);
-}
-
-void code2x8_matvec_cuda(const void* __restrict__ A, const void* __restrict__ B,
-                         void* __restrict__ C,
-                         const void* __restrict__ codebook, int prob_m,
-                         int prob_k, const int4 codebook_a_sizes,
-                         const int codebook_stride) {
-  int sms;
-  cudaDeviceGetAttribute(&sms, cudaDevAttrMultiProcessorCount, 0);
-  int waves = 0;
-  int thread_m;
-  do {
-    waves++;
-    thread_m = ceildiv(prob_m, waves * sms);
-  } while (thread_m > THREAD_M);
-
-  int blocks = ceildiv(prob_m, thread_m);
-  int threads = 32 * thread_m;
-  int shared = 16 * (2 * 256 * 8 + 32 * 9);
-  cudaFuncSetAttribute(Code2x8MatVec,
-                       cudaFuncAttributeMaxDynamicSharedMemorySize, shared);
-  cudaStream_t stream = at::cuda::getCurrentCUDAStream().stream();
-  Code2x8MatVec<<<blocks, threads, shared, stream>>>(
-      (const int4*)A, (const int4*)B, (int4*)C, (const int4*)codebook, prob_m,
-      prob_k, codebook_a_sizes, codebook_stride);
-}
-
-void code1x16_dequant_cuda(
-    const void* __restrict__ A, void* __restrict__ C,
-    const void* __restrict__ codebook, int prob_m, int prob_k,
-    const int4 codebook_a_sizes,  // cumulative sizes of A spanning each
-                                  // codebook, at most 3 long.
-    const int codebook_stride     // as int4.
-) {
-  int sms;
-  cudaDeviceGetAttribute(&sms, cudaDevAttrMultiProcessorCount, 0);
-  int waves = 0;
-  int thread_m;
-  do {
-    waves++;
-    thread_m = ceildiv(prob_m, waves * sms);
-  } while (thread_m > THREAD_M);
-
-  int blocks = ceildiv(prob_m, thread_m);
-  int threads = 32 * thread_m;
-  cudaStream_t stream = at::cuda::getCurrentCUDAStream().stream();
-  Code1x16Dequant<<<blocks, threads, 0, stream>>>(
-      (const int4*)A, (int4*)C, (const int4*)codebook, prob_m, prob_k,
-      codebook_a_sizes,  // cumulative sizes of A spanning each codebook, at
-                         // most 3 long.
-      codebook_stride    // as int4.
-  );
-}
-
-// Dequantizes the code and codebook into weights.
-void code2x8_dequant_cuda(
-    const void* __restrict__ A, void* __restrict__ C,
-    const void* __restrict__ codebook, int prob_m, int prob_k,
-    const int4
-        codebook_a_sizes,  // cumulative sizes of A spanning each codebook, at
-                           // most 3 long, corresponds to cols.
-    const int codebook_stride  // as int4
-) {
-  int sms;
-  cudaDeviceGetAttribute(&sms, cudaDevAttrMultiProcessorCount, 0);
-  int waves = 0;
-  int thread_m;
-  do {
-    waves++;
-    thread_m = ceildiv(prob_m, waves * sms);
-  } while (thread_m > THREAD_M);
-
-  int blocks = ceildiv(prob_m, thread_m);
-  int threads = 32 * thread_m;
-  int shared = 16 * (2 * 256 * 8 + 32 * 9);
-  cudaStream_t stream = at::cuda::getCurrentCUDAStream().stream();
-
-  cudaFuncSetAttribute(Code2x8Dequant,
-                       cudaFuncAttributeMaxDynamicSharedMemorySize, shared);
-  Code2x8Dequant<<<blocks, threads, shared, stream>>>(
-      (const int4*)A, (int4*)C, (const int4*)codebook, prob_m, prob_k,
-      codebook_a_sizes, codebook_stride);
-}
-
-int codebook_stride(const torch::Tensor& codebooks) {
-  return codebooks.stride(0) * codebooks.element_size() / sizeof(int4);
-}
-
-void code1x16_matvec(
-    const torch::Tensor& A, const torch::Tensor& B, torch::Tensor& C,
-    const torch::Tensor& codebook,
-    const int4 codebook_a_sizes  // cumulative sizes of A spanning each
-                                 // codebook, at most 3 long.
-) {
-  const at::cuda::OptionalCUDAGuard device_guard(device_of(A));
-  int prob_m = C.size(0);
-  int prob_k = B.size(0);
-
-  code1x16_matvec_cuda(A.data_ptr(), B.data_ptr(), C.data_ptr(),
-                       codebook.data_ptr(), prob_m, prob_k, codebook_a_sizes,
-                       codebook_stride(codebook));
-}
-
-torch::Tensor code1x16_matmat(const torch::Tensor& input,
-                              const torch::Tensor& codes,
-                              const torch::Tensor& codebooks,
-                              const torch::Tensor& scales,
-                              const int4 codebook_a_sizes,
-                              const std::optional<torch::Tensor>& bias) {
-  auto input_sizes = input.sizes();
-  auto out_features = codes.size(0) * codebooks.size(2);
-  auto flat_input = input.reshape({-1, input.size(-1)});
-  auto flat_output = torch::empty(
-      {flat_input.size(0), out_features},
-      torch::TensorOptions().dtype(input.dtype()).device(input.device()));
-
-  for (int i = 0; i < flat_input.size(0); ++i) {
-    auto input_vec = flat_input.index({i});
-    auto output_vec = flat_output.index({i});
-    code1x16_matvec(codes.squeeze(2), input_vec, output_vec, codebooks,
-                    codebook_a_sizes);
-  }
-  flat_output *= scales.flatten().unsqueeze(0);
-
-  if (bias.has_value()) {
-    flat_output += bias->unsqueeze(0);
-  }
-
-  auto output_sizes = input_sizes.vec();
-  output_sizes.pop_back();
-  output_sizes.push_back(-1);
-  auto output = flat_output.reshape(output_sizes);
-  return output;
-}
-
-void code2x8_matvec(const torch::Tensor& A, const torch::Tensor& B,
-                    torch::Tensor& C, const torch::Tensor& codebook,
-                    const int4 codebook_a_sizes) {
-  const at::cuda::OptionalCUDAGuard device_guard(device_of(A));
-  int prob_m = C.size(0);
-  int prob_k = B.size(0);
-  code2x8_matvec_cuda(A.data_ptr(), B.data_ptr(), C.data_ptr(),
-                      codebook.data_ptr(), prob_m, prob_k, codebook_a_sizes,
-                      2 * codebook_stride(codebook));
-}
-
-torch::Tensor code2x8_matmat(const torch::Tensor& input,
-                             const torch::Tensor& codes,
-                             const torch::Tensor& codebooks,
-                             const torch::Tensor& scales,
-                             const int4 codebook_a_sizes,
-                             const std::optional<torch::Tensor>& bias) {
-  auto input_sizes = input.sizes();
-  auto out_features = codes.size(0) * codebooks.size(2);
-  auto flat_input = input.reshape({-1, input.size(-1)});
-  auto flat_output = torch::empty(
-      {flat_input.size(0), out_features},
-      torch::TensorOptions().dtype(input.dtype()).device(input.device()));
-
-  for (int i = 0; i < flat_input.size(0); ++i) {
-    auto input_vec = flat_input.index({i});
-    auto output_vec = flat_output.index({i});
-    code2x8_matvec(codes.squeeze(2), input_vec, output_vec, codebooks,
-                   codebook_a_sizes);
-  }
-  flat_output *= scales.flatten().unsqueeze(0);
-  if (bias.has_value()) {
-    flat_output += bias->unsqueeze(0);
-  }
-
-  auto output_sizes = input_sizes.vec();
-  output_sizes.pop_back();
-  output_sizes.push_back(-1);
-  auto output = flat_output.reshape(output_sizes);
-  return output;
-}
-
-// Accumulate the partition sizes.
-int4 accumulate_sizes(const std::vector<int64_t>& codebook_partition_sizes) {
-  int4 cumulative_sizes;
-  auto cumulative_size = &cumulative_sizes.x;
-  size_t i = 0;
-  int last = 0;
-  assert(codebook_partition_sizes.size() <= 4);
-  for (; i < codebook_partition_sizes.size(); ++i, ++cumulative_size) {
-    *cumulative_size = codebook_partition_sizes[i] + last;
-    last = *cumulative_size;
-  }
-  // fill in the rest with unreachable.
-  for (; i < 4; ++i, ++cumulative_size) {
-    *cumulative_size = last * 10;
-  }
-  return cumulative_sizes;
-}
-
-}  // namespace aqlm
-}  // namespace vllm
-
-torch::Tensor aqlm_gemm(const torch::Tensor& input, const torch::Tensor& codes,
-                        const torch::Tensor& codebooks,
-                        const torch::Tensor& scales,
-                        const std::vector<int64_t>& codebook_partition_sizes,
-                        const std::optional<torch::Tensor>& bias) {
-  int4 cumulative_sizes =
-      vllm::aqlm::accumulate_sizes(codebook_partition_sizes);
-
-  int const nbooks = codebooks.size(0) / codebook_partition_sizes.size();
-  int const entries = codebooks.size(1);
-
-  if (nbooks == 1 && entries == (1 << 16)) {
-    return vllm::aqlm::code1x16_matmat(input, codes, codebooks, scales,
-                                       cumulative_sizes, bias);
-  }
-  if (nbooks == 2 && entries == (1 << 8)) {
-    return vllm::aqlm::code2x8_matmat(input, codes, codebooks, scales,
-                                      cumulative_sizes, bias);
-  }
-
-  TORCH_CHECK(false, "AQLM with ", nbooks, " codebooks and ", entries,
-              " entries is not currently supported.")
-  return {};
-}
-
-torch::Tensor aqlm_dequant(
-    const torch::Tensor& codes, const torch::Tensor& codebooks,
-    const std::vector<int64_t>& codebook_partition_sizes) {
-  int4 cumulative_sizes =
-      vllm::aqlm::accumulate_sizes(codebook_partition_sizes);
-
-  int const nbooks = codebooks.size(0) / codebook_partition_sizes.size();
-  int const entries = codebooks.size(1);
-
-  const at::cuda::OptionalCUDAGuard device_guard(device_of(codes));
-  int rows = codes.size(1);
-  int cols = codes.size(0);
-
-  auto in_features = codes.size(1) * 8;
-  auto out_features = codes.size(0);
-
-  assert(out_features == std::accumulate(codebook_partition_sizes.begin(),
-                                         codebook_partition_sizes.end(), 0));
-
-  auto weights = torch::empty({out_features, in_features},
-                              torch::TensorOptions()
-                                  .dtype(codebooks.dtype())
-                                  .device(codebooks.device()));
-
-  if (nbooks == 1 && entries == (1 << 16)) {
-    vllm::aqlm::code1x16_dequant_cuda(codes.data_ptr(), weights.data_ptr(),
-                                      codebooks.data_ptr(), out_features,
-                                      in_features, cumulative_sizes,
-                                      vllm::aqlm::codebook_stride(codebooks));
-
-    // if you wanted to flip to scaling the weights, (though it's 30%-ish slower
-    // and not consistent with gemv implementation.) weights *=
-    // scales.index({"...", 0, 0});
-
-    return weights;
-  }
-
-  if (nbooks == 2 && entries == (1 << 8)) {
-    vllm::aqlm::code2x8_dequant_cuda(codes.data_ptr(), weights.data_ptr(),
-                                     codebooks.data_ptr(), out_features,
-                                     in_features, cumulative_sizes,
-                                     vllm::aqlm::codebook_stride(codebooks));
-
-    // if you wanted to flip to scaling the weights, (though it's 30%-ish slower
-    // and not consistent with gemv implementation) weights *=
-    // scales.index({"...", 0, 0});
-
-    return weights;
-  }
-
-  TORCH_CHECK(false, "AQLM with ", nbooks, " codebooks and ", entries,
-              " entries is not currently supported.")
-  return {};
-}

csrc/quantization/compressed_tensors/int8_quant_kernels.cu

@@ -1,7 +1,9 @@
 #include <ATen/cuda/CUDAContext.h>
 #include <torch/all.h>
 
-#include "../per_token_group_quant_8bit.h"
+#ifndef USE_ROCM
+  #include "../per_token_group_quant_8bit.h"
+#endif
 
 #include <cmath>
 
@@ -339,10 +341,12 @@ void dynamic_scaled_int8_quant(
       });
 }
 
+#ifndef USE_ROCM
 void per_token_group_quant_int8(const torch::Tensor& input,
                                 torch::Tensor& output_q,
                                 torch::Tensor& output_s, int64_t group_size,
                                 double eps, double int8_min, double int8_max) {
   per_token_group_quant_8bit(input, output_q, output_s, group_size, eps,
                              int8_min, int8_max);
-}
+}
+#endif

csrc/quantization/cutlass_w8a8/Epilogues.md

@@ -86,6 +86,7 @@ D = s_a s_b \widehat A \widehat B
 ```
 
 Epilogue parameters:
+
 - `scale_a` is the scale for activations, can be per-tensor (scalar) or per-token (column-vector).
 - `scale_b` is the scale for weights, can be per-tensor (scalar) or per-channel (row-vector).
 
@@ -135,7 +136,7 @@ That is precomputed and stored in `azp_with_adj` as a row-vector.
 Epilogue parameters:
 
 - `scale_a` is the scale for activations, can be per-tensor (scalar) or per-token (column-vector).
-  - Generally this will be per-tensor as the zero-points are per-tensor.
+    - Generally this will be per-tensor as the zero-points are per-tensor.
 - `scale_b` is the scale for weights, can be per-tensor (scalar) or per-channel (row-vector).
 - `azp_with_adj` is the precomputed zero-point term ($` z_a J_a \widehat B `$), is per-channel (row-vector).
 - `bias` is the bias, is always per-channel (row-vector).
@@ -152,7 +153,7 @@ That means the zero-point term $` z_a J_a \widehat B `$ becomes an outer product
 Epilogue parameters:
 
 - `scale_a` is the scale for activations, can be per-tensor (scalar) or per-token (column-vector).
-  - Generally this will be per-token as the zero-points are per-token.
+    - Generally this will be per-token as the zero-points are per-token.
 - `scale_b` is the scale for weights, can be per-tensor (scalar) or per-channel (row-vector).
 - `azp_adj` is the precomputed zero-point adjustment term ($` \mathbf 1 \widehat B `$), is per-channel (row-vector).
 - `azp` is the zero-point (`z_a`), is per-token (column-vector).

csrc/quantization/cutlass_w8a8/c3x/scaled_mm_blockwise_sm120_fp8.cu

@@ -0,0 +1,23 @@
+#include "scaled_mm_kernels.hpp"
+#include "scaled_mm_blockwise_sm120_fp8_dispatch.cuh"
+#include "cutlass_extensions/epilogue/scaled_mm_epilogues_c3x.hpp"
+
+namespace vllm {
+
+void cutlass_scaled_mm_blockwise_sm120_fp8(torch::Tensor& out,
+                                           torch::Tensor const& a,
+                                           torch::Tensor const& b,
+                                           torch::Tensor const& a_scales,
+                                           torch::Tensor const& b_scales) {
+  if (out.dtype() == torch::kBFloat16) {
+    cutlass_gemm_blockwise_sm120_fp8_dispatch<cutlass::bfloat16_t>(
+        out, a, b, a_scales, b_scales);
+
+  } else {
+    TORCH_CHECK(out.dtype() == torch::kFloat16);
+    cutlass_gemm_blockwise_sm120_fp8_dispatch<cutlass::half_t>(
+        out, a, b, a_scales, b_scales);
+  }
+}
+
+}  // namespace vllm

csrc/quantization/cutlass_w8a8/c3x/scaled_mm_blockwise_sm120_fp8_dispatch.cuh

@@ -0,0 +1,183 @@
+#pragma once
+
+#include "cuda_utils.h"
+#include "cutlass/cutlass.h"
+#include "cutlass/numeric_types.h"
+
+#include "cute/tensor.hpp"
+#include "cutlass/tensor_ref.h"
+#include "cutlass/gemm/dispatch_policy.hpp"
+#include "cutlass/gemm/collective/collective_builder.hpp"
+#include "cutlass/gemm/device/gemm_universal_adapter.h"
+#include "cutlass/gemm/kernel/gemm_universal.hpp"
+#include "cutlass/gemm/kernel/tile_scheduler_params.h"
+#include "cutlass/epilogue/dispatch_policy.hpp"
+#include "cutlass/epilogue/collective/collective_builder.hpp"
+
+#include "cutlass_extensions/gemm/dispatch_policy.hpp"
+#include "cutlass_extensions/gemm/collective/collective_builder.hpp"
+
+#include "cutlass_gemm_caller.cuh"
+
+namespace vllm {
+
+using namespace cute;
+
+// clang-format off
+template <class OutType, int ScaleGranularityM,
+          int ScaleGranularityN, int ScaleGranularityK,
+          class MmaTileShape, class ClusterShape,
+          class EpilogueScheduler, class MainloopScheduler>
+struct cutlass_3x_gemm_fp8_blockwise {
+  using ElementAB = cutlass::float_e4m3_t;
+
+  using ElementA = ElementAB;
+  using LayoutA = cutlass::layout::RowMajor;
+  using LayoutA_Transpose = typename cutlass::layout::LayoutTranspose<LayoutA>::type;
+  static constexpr int AlignmentA = 128 / cutlass::sizeof_bits<ElementA>::value;
+
+  using ElementB = ElementAB;
+  // ColumnMajor is used for B to match the CUTLASS convention.
+  using LayoutB = cutlass::layout::ColumnMajor;
+  using LayoutB_Transpose = typename cutlass::layout::LayoutTranspose<LayoutB>::type;
+  static constexpr int AlignmentB = 128 / cutlass::sizeof_bits<ElementB>::value;
+
+  using ElementD = OutType;
+  using LayoutD = cutlass::layout::RowMajor;
+  using LayoutD_Transpose = typename cutlass::layout::LayoutTranspose<LayoutD>::type;
+  static constexpr int AlignmentD = 128 / cutlass::sizeof_bits<ElementD>::value;
+
+  using ElementC = void; // TODO: support bias
+  using LayoutC = LayoutD;
+  using LayoutC_Transpose = LayoutD_Transpose;
+  static constexpr int AlignmentC = AlignmentD;
+
+  using ElementAccumulator = float;
+  using ElementCompute = float;
+  using ElementBlockScale = float; 
+
+  using ScaleConfig = cutlass::detail::Sm120BlockwiseScaleConfig<
+        ScaleGranularityM, ScaleGranularityN, ScaleGranularityK,
+        cute::UMMA::Major::MN, cute::UMMA::Major::K>;
+
+  // layout_SFA and layout_SFB cannot be swapped since they are deduced.
+  using LayoutSFA = decltype(ScaleConfig::deduce_layoutSFA());
+  using LayoutSFB = decltype(ScaleConfig::deduce_layoutSFB());
+
+  using ArchTag = cutlass::arch::Sm120;
+  using OperatorClass = cutlass::arch::OpClassTensorOp;
+
+  static constexpr auto RoundStyle = cutlass::FloatRoundStyle::round_to_nearest;
+  using ElementScalar = float;
+  using DefaultOperation = cutlass::epilogue::fusion::LinearCombination<ElementD, ElementCompute, ElementC, ElementScalar, RoundStyle>;
+  using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
+      ArchTag,
+      OperatorClass,
+      MmaTileShape,
+      ClusterShape,
+      cutlass::epilogue::collective::EpilogueTileAuto,
+      ElementAccumulator,
+      ElementCompute,
+      ElementC,
+      LayoutC,
+      AlignmentC,
+      ElementD,
+      LayoutD,
+      AlignmentD,
+      EpilogueScheduler,
+      DefaultOperation
+  >::CollectiveOp;
+ 
+  using StageCountType = cutlass::gemm::collective::StageCountAuto; 
+  using CollectiveMainloop = 
+      typename cutlass::gemm::collective::CollectiveBuilder<
+          ArchTag,
+          OperatorClass,
+          ElementA,
+          cute::tuple<LayoutA, LayoutSFA>,
+          AlignmentA,
+          ElementB,
+          cute::tuple<LayoutB, LayoutSFB>,
+          AlignmentB,
+          ElementAccumulator,
+          MmaTileShape,
+          ClusterShape,
+          cutlass::gemm::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
+          MainloopScheduler
+      >::CollectiveOp;
+
+  using KernelType = enable_sm120_only<cutlass::gemm::kernel::GemmUniversal<
+      Shape<int, int, int, int>, CollectiveMainloop, CollectiveEpilogue>>;
+
+  struct GemmKernel : public KernelType {};
+};
+
+template <typename Gemm>
+void cutlass_gemm_caller_blockwise(torch::Tensor& out, torch::Tensor const& a,
+                                   torch::Tensor const& b,
+                                   torch::Tensor const& a_scales,
+                                   torch::Tensor const& b_scales) {
+  using GemmKernel = typename Gemm::GemmKernel;
+  using StrideA = typename Gemm::GemmKernel::StrideA;
+  using StrideB = typename Gemm::GemmKernel::StrideB;
+  using StrideD = typename Gemm::GemmKernel::StrideD;
+  using StrideC = typename Gemm::GemmKernel::StrideC;
+  using LayoutSFA = typename Gemm::LayoutSFA;
+  using LayoutSFB = typename Gemm::LayoutSFB;
+  using ScaleConfig = typename Gemm::ScaleConfig;
+
+  using ElementAB = typename Gemm::ElementAB;
+  using ElementD = typename Gemm::ElementD;
+
+  int32_t m = a.size(0), n = b.size(1), k = a.size(1);
+
+  StrideA a_stride;
+  StrideB b_stride;
+  StrideC c_stride;
+  a_stride =
+      cutlass::make_cute_packed_stride(StrideA{}, cute::make_shape(m, k, 1));
+  b_stride =
+      cutlass::make_cute_packed_stride(StrideB{}, cute::make_shape(n, k, 1));
+  c_stride =
+      cutlass::make_cute_packed_stride(StrideC{}, cute::make_shape(m, n, 1));
+
+  LayoutSFA layout_SFA = 
+      ScaleConfig::tile_atom_to_shape_SFA(make_shape(m, n, k, 1));
+  LayoutSFB layout_SFB = 
+      ScaleConfig::tile_atom_to_shape_SFB(make_shape(m, n, k, 1));
+
+  auto a_ptr = static_cast<ElementAB*>(a.data_ptr());
+  auto b_ptr = static_cast<ElementAB*>(b.data_ptr());
+  auto a_scales_ptr = static_cast<float*>(a_scales.data_ptr());
+  auto b_scales_ptr = static_cast<float*>(b_scales.data_ptr());
+
+  auto mainloop_args = [&](){
+    return typename GemmKernel::MainloopArguments{
+        a_ptr,        a_stride,   b_ptr,        b_stride,
+        a_scales_ptr, layout_SFA, b_scales_ptr, layout_SFB
+    };
+  }();
+  auto prob_shape = cute::make_shape(m, n, k, 1);
+
+  auto c_ptr = static_cast<ElementD*>(out.data_ptr());
+  typename GemmKernel::EpilogueArguments epilogue_args{
+      {}, c_ptr, c_stride, c_ptr, c_stride};
+  c3x::cutlass_gemm_caller<GemmKernel>(a.device(), prob_shape, mainloop_args,
+                                       epilogue_args);
+}
+
+template <typename OutType>
+void cutlass_gemm_blockwise_sm120_fp8_dispatch(torch::Tensor& out,
+                                               torch::Tensor const& a,
+                                               torch::Tensor const& b,
+                                               torch::Tensor const& a_scales,
+                                               torch::Tensor const& b_scales) {
+  // TODO: better heuristics
+  cutlass_gemm_caller_blockwise<cutlass_3x_gemm_fp8_blockwise<
+      OutType, 1, 128, 128, Shape<_128, _128, _128>,
+      Shape<_1, _1, _1>, cutlass::epilogue::collective::EpilogueScheduleAuto,
+      cutlass::gemm::collective::KernelScheduleAuto>>(
+      out, a, b, a_scales, b_scales);
+}
+
+}  // namespace vllm

csrc/quantization/cutlass_w8a8/c3x/scaled_mm_kernels.hpp

@@ -47,4 +47,10 @@ void cutlass_scaled_mm_blockwise_sm100_fp8(torch::Tensor& out,
                                            torch::Tensor const& b,
                                            torch::Tensor const& a_scales,
                                            torch::Tensor const& b_scales);
+
+void cutlass_scaled_mm_blockwise_sm120_fp8(torch::Tensor& out,
+                                           torch::Tensor const& a,
+                                           torch::Tensor const& b,
+                                           torch::Tensor const& a_scales,
+                                           torch::Tensor const& b_scales);
 }  // namespace vllm

csrc/quantization/cutlass_w8a8/c3x/scaled_mm_sm90_fp8.cu

@@ -1,6 +1,5 @@
 #include "scaled_mm_kernels.hpp"
 #include "scaled_mm_sm90_fp8_dispatch.cuh"
-#include "cutlass_extensions/epilogue/scaled_mm_epilogues_c3x.hpp"
 
 namespace vllm {
 
@@ -13,11 +12,11 @@ void cutlass_scaled_mm_sm90_fp8(torch::Tensor& out, torch::Tensor const& a,
   if (bias) {
     TORCH_CHECK(bias->dtype() == out.dtype(),
                 "currently bias dtype must match output dtype ", out.dtype());
-    return cutlass_scaled_mm_sm90_fp8_epilogue<c3x::ScaledEpilogueBias>(
-        out, a, b, a_scales, b_scales, *bias);
+    return cutlass_scaled_mm_sm90_fp8_epilogue<true>(out, a, b, a_scales,
+                                                     b_scales, *bias);
   } else {
-    return cutlass_scaled_mm_sm90_fp8_epilogue<c3x::ScaledEpilogue>(
-        out, a, b, a_scales, b_scales);
+    return cutlass_scaled_mm_sm90_fp8_epilogue<false>(out, a, b, a_scales,
+                                                      b_scales);
   }
 }