Add pathlib.Path support and docstrings for environment variables

- Convert path-related variables to use pathlib.Path types with standardized parsing - Add comprehensive docstrings based on original comments for better documentation - Implement standardized path parsing with expanduser and absolute path conversion - Support both Path and Optional[Path] types in parsing logic - Maintain backwards compatibility while providing type-safe path handling Co-authored-by: ProExpertProg <11367180+ProExpertProg@users.noreply.github.com>
Implement standardized environment variable parsing as requested
2025-08-26 02:49:34 +00:00 · 2025-08-26 01:22:13 +00:00 · 2025-08-26 00:45:58 +00:00 · 2025-08-26 00:42:28 +00:00 · 2025-08-26 00:30:21 +00:00 · 2025-08-25 16:44:51 -07:00
730 changed files with 40799 additions and 21896 deletions
--- a/.buildkite/generate_index.py
+++ b/.buildkite/generate_index.py
@ -8,7 +8,8 @@ template = """<!DOCTYPE html>
 <html>
    <body>
    <h1>Links for vLLM</h1/>
-        <a href="../{wheel_html_escaped}">{wheel}</a><br/>
+        <a href="../{x86_wheel_html_escaped}">{x86_wheel}</a><br/>
+        <a href="../{arm_wheel_html_escaped}">{arm_wheel}</a><br/>
    </body>
 </html>
 """
@ -21,7 +22,25 @@ filename = os.path.basename(args.wheel)

 with open("index.html", "w") as f:
    print(f"Generated index.html for {args.wheel}")
+    # sync the abi tag with .buildkite/scripts/upload-wheels.sh
+    if "x86_64" in filename:
+        x86_wheel = filename
+        arm_wheel = filename.replace("x86_64", "aarch64").replace(
+            "manylinux1", "manylinux2014"
+        )
+    elif "aarch64" in filename:
+        x86_wheel = filename.replace("aarch64", "x86_64").replace(
+            "manylinux2014", "manylinux1"
+        )
+        arm_wheel = filename
+    else:
+        raise ValueError(f"Unsupported wheel: {filename}")
    # cloudfront requires escaping the '+' character
    f.write(
-        template.format(wheel=filename, wheel_html_escaped=filename.replace("+", "%2B"))
+        template.format(
+            x86_wheel=x86_wheel,
+            x86_wheel_html_escaped=x86_wheel.replace("+", "%2B"),
+            arm_wheel=arm_wheel,
+            arm_wheel_html_escaped=arm_wheel.replace("+", "%2B"),
+        )
    )
--- a/.buildkite/lm-eval-harness/configs/Meta-Llama-3-8B-QQQ.yaml
+++ b/.buildkite/lm-eval-harness/configs/Meta-Llama-3-8B-QQQ.yaml
@ -1,12 +0,0 @@
-# For vllm script, with -t option (tensor parallel size).
-# bash .buildkite/lm-eval-harness/run-lm-eval-gsm-vllm-baseline.sh -m HandH1998/QQQ-Llama-3-8b-g128 -b 32 -l 1000 -f 5 -t 1
-model_name: "HandH1998/QQQ-Llama-3-8b-g128"
-tasks:
- name: "gsm8k"
-  metrics:
-  - name: "exact_match,strict-match"
-    value: 0.419
-  - name: "exact_match,flexible-extract"
-    value: 0.416
-limit: 1000
-num_fewshot: 5
--- a/.buildkite/lm-eval-harness/configs/models-large.txt
+++ b/.buildkite/lm-eval-harness/configs/models-large.txt
@ -3,4 +3,3 @@ Meta-Llama-3-70B-Instruct.yaml
 Mixtral-8x7B-Instruct-v0.1.yaml
 Qwen2-57B-A14-Instruct.yaml
 DeepSeek-V2-Lite-Chat.yaml
-Meta-Llama-3-8B-QQQ.yaml
--- a/.buildkite/lm-eval-harness/run-lm-eval-gsm-hf-baseline.sh
+++ b/.buildkite/lm-eval-harness/run-lm-eval-gsm-hf-baseline.sh
@ -2,7 +2,7 @@
 # We can use this script to compute baseline accuracy on GSM for transformers.
 #
 # Make sure you have lm-eval-harness installed:
-#   pip install lm-eval==0.4.4
+#   pip install git+https://github.com/EleutherAI/lm-evaluation-harness.git@206b7722158f58c35b7ffcd53b035fdbdda5126d#egg=lm-eval[api]

 usage() {
    echo``
--- a/.buildkite/lm-eval-harness/run-lm-eval-gsm-vllm-baseline.sh
+++ b/.buildkite/lm-eval-harness/run-lm-eval-gsm-vllm-baseline.sh
@ -3,7 +3,7 @@
 # We use this for fp8, which HF does not support.
 #
 # Make sure you have lm-eval-harness installed:
-#   pip install lm-eval==0.4.4
+#   pip install git+https://github.com/EleutherAI/lm-evaluation-harness.git@206b7722158f58c35b7ffcd53b035fdbdda5126d#egg=lm-eval[api]

 usage() {
    echo``
--- a/.buildkite/nightly-benchmarks/README.md
+++ b/.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

@ -138,28 +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

--- a/.buildkite/nightly-benchmarks/nightly-descriptions.md
+++ b/.buildkite/nightly-benchmarks/nightly-descriptions.md
@ -17,7 +17,7 @@ Latest reproduction guilde: [github issue link](https://github.com/vllm-project/
    - 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.*
+        - *NOTE: we use 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
--- a/.buildkite/nightly-benchmarks/scripts/compare-json-results.py
+++ b/.buildkite/nightly-benchmarks/scripts/compare-json-results.py
@ -1,33 +1,202 @@
 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 import argparse
+import json
+import os
+from importlib import util

 import pandas as pd

+plotly_found = util.find_spec("plotly.express") is not None
+

 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 = []
-    compare_frames = []
-    for file in files:
-        data_df = pd.read_json(file)
-        serving_df = data_df.dropna(subset=[drop_column], ignore_index=True)
-        if ignore_test_name is False:
-            serving_df = serving_df.rename(columns={name_column: file + "_name"})
-            frames.append(serving_df[file + "_name"])
-        serving_df = serving_df.rename(columns={data_column: file})
-        frames.append(serving_df[file])
-        compare_frames.append(serving_df[file])
-        if len(compare_frames) >= 2:
-            # Compare numbers among two files
-            ratio_df = compare_frames[1] / compare_frames[0]
-            frames.append(ratio_df)
-            compare_frames.pop(1)
+    """
+    Align concatenation by keys derived from info_cols instead of row order.
+    - Pick one canonical key list: subset of info_cols present in ALL files.
+    - For each file: set index to those keys, aggregate duplicates
+    - (mean for metric, first for names).
+    - Concat along axis=1 (indexes align), then reset_index so callers can
+    - group by columns.
+    - If --debug, add a <file_label>_name column per file.
+    """
+    print("\ncompare_data_column:", data_column)

+    frames = []
+    raw_data_cols = []
+    compare_frames = []
+
+    # 1) choose a canonical key list from info_cols that exists in ALL files
+    cols_per_file = []
+    for f in files:
+        try:
+            df_tmp = pd.read_json(f, orient="records")
+        except Exception as err:
+            raise ValueError(f"Failed to read {f}") from err
+        cols_per_file.append(set(df_tmp.columns))
+
+    key_cols = [c for c in info_cols if all(c in cset for cset in cols_per_file)]
+    if not key_cols:
+        # soft fallback: use any info_cols present in the first file
+        key_cols = [c for c in info_cols if c in list(cols_per_file[0])]
+    if not key_cols:
+        raise ValueError(
+            "No common key columns found from info_cols across the input files."
+        )
+
+    # 2) build a single "meta" block (keys as columns) once, aligned by the key index
+    meta_added = False
+
+    for file in files:
+        df = pd.read_json(file, orient="records")
+
+        # Keep rows that actually have the compared metric (same as original behavior)
+        if drop_column in df.columns:
+            df = df.dropna(subset=[drop_column], ignore_index=True)
+
+        # Stabilize numeric key columns (harmless if missing)
+        for c in (
+            "Input Len",
+            "Output Len",
+            "TP Size",
+            "PP Size",
+            "# of max concurrency.",
+            "qps",
+        ):
+            if c in df.columns:
+                df[c] = pd.to_numeric(df[c], errors="coerce")
+
+        # Ensure all key columns exist
+        for c in key_cols:
+            if c not in df.columns:
+                df[c] = pd.NA
+
+        # Set index = key_cols and aggregate duplicates → unique MultiIndex
+        df_idx = df.set_index(key_cols, drop=False)
+
+        # meta (key columns), unique per key
+        meta = df_idx[key_cols]
+        if not meta.index.is_unique:
+            meta = meta.groupby(level=key_cols, dropna=False).first()
+
+        # metric series for this file, aggregated to one row per key
+        file_label = "/".join(file.split("/")[:-1]) or os.path.basename(file)
+        s = df_idx[data_column]
+        if not s.index.is_unique:
+            s = s.groupby(level=key_cols, dropna=False).mean()
+        s.name = file_label  # column label like original
+
+        # add meta once (from first file) so keys are the leftmost columns
+        if not meta_added:
+            frames.append(meta)
+            meta_added = True
+
+        # (NEW) debug: aligned test-name column per file
+        if debug and name_column in df_idx.columns:
+            name_s = df_idx[name_column]
+            if not name_s.index.is_unique:
+                name_s = name_s.groupby(level=key_cols, dropna=False).first()
+            name_s.name = f"{file_label}_name"
+            frames.append(name_s)
+
+        frames.append(s)
+        raw_data_cols.append(file_label)
+        compare_frames.append(s)
+
+        # Generalize ratio: for any file N>=2, add ratio (fileN / file1)
+        if len(compare_frames) >= 2:
+            base = compare_frames[0]
+            current = compare_frames[-1]
+            ratio = current / base
+            ratio = ratio.mask(base == 0)  # avoid inf when baseline is 0
+            ratio.name = f"Ratio 1 vs {len(compare_frames)}"
+            frames.append(ratio)
+
+    # 4) concat on columns with aligned MultiIndex;
+    # then reset_index to return keys as columns
    concat_df = pd.concat(frames, axis=1)
-    return concat_df
+    concat_df = concat_df.reset_index(drop=True).reset_index()
+    if "index" in concat_df.columns:
+        concat_df = concat_df.drop(columns=["index"])
+
+    # Ensure key/info columns appear first (in your info_cols order)
+    front = [c for c in info_cols if c in concat_df.columns]
+    rest = [c for c in concat_df.columns if c not in front]
+    concat_df = concat_df[front + rest]
+
+    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)
+
+    # Keep only "serving" tests
+    name_col = next(
+        (c for c in ["Test name", "test_name", "Test Name"] if c in df.columns), None
+    )
+    if name_col:
+        df = df[
+            df[name_col].astype(str).str.contains(r"serving", case=False, na=False)
+        ].copy()
+
+    # 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 +205,103 @@ 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 and plotly_found:
+                    import plotly.express as px
+
+                    df = group[raw_data_cols]
+                    df_sorted = df.sort_values(by=info_cols[y_axis_index])
+                    # Melt DataFrame for plotting
+                    df_melted = df_sorted.melt(
+                        id_vars=info_cols[y_axis_index],
+                        var_name="Configuration",
+                        value_name=data_cols_to_compare[i],
+                    )
+                    title = data_cols_to_compare[i] + " vs " + info_cols[y_axis_index]
+                    # Create Plotly line chart
+                    fig = px.line(
+                        df_melted,
+                        x=info_cols[y_axis_index],
+                        y=data_cols_to_compare[i],
+                        color="Configuration",
+                        title=title,
+                        markers=True,
+                    )
+                    # Export to HTML
+                    text_file.write(fig.to_html(full_html=True, include_plotlyjs="cdn"))
--- a/.buildkite/nightly-benchmarks/scripts/convert-results-json-to-markdown.py
+++ b/.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 = {
@ -42,14 +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)",
@ -94,7 +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:
@ -102,7 +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:
@ -110,12 +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
@ -205,7 +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:
@ -245,7 +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"
@ -260,7 +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")
--- a/.buildkite/nightly-benchmarks/scripts/run-nightly-benchmarks.sh
+++ b/.buildkite/nightly-benchmarks/scripts/run-nightly-benchmarks.sh
@ -382,7 +382,7 @@ run_genai_perf_tests() {
      client_command="genai-perf profile \
        -m $model \
        --service-kind openai \
-        --backend vllm \
+        --backend "$backend" \
        --endpoint-type chat \
        --streaming \
        --url localhost:$port \
--- a/.buildkite/nightly-benchmarks/scripts/run-performance-benchmarks.sh
+++ b/.buildkite/nightly-benchmarks/scripts/run-performance-benchmarks.sh
@ -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
--- a/.buildkite/nightly-benchmarks/tests/latency-tests-cpu.json
+++ b/.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,
--- a/.buildkite/nightly-benchmarks/tests/serving-tests-cpu-snc2.json
+++ b/.buildkite/nightly-benchmarks/tests/serving-tests-cpu-snc2.json
@ -1,7 +1,8 @@
 [
    {
        "test_name": "serving_llama8B_tp1_sharegpt",
-        "qps_list": [1, 4, 16, "inf"],
+        "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,
@ -10,7 +11,7 @@
 	    "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",
@ -23,17 +24,17 @@
            "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"],
+        "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,
@ -42,7 +43,7 @@
 	    "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",
@ -55,17 +56,17 @@
            "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"],
+        "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,
@ -74,7 +75,7 @@
 	    "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",
@ -87,17 +88,17 @@
            "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_tp1_random_128_128",
-        "qps_list": [1, 4, 16, "inf"],
+        "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,
@ -106,7 +107,7 @@
 	    "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",
@ -120,19 +121,19 @@
            "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": 128,
 	    "random-output-len": 128,
 	    "ignore-eos": "",
-	    "max_concurrency": 1000,
            "num_prompts": 1000
        }
    },
    {
        "test_name": "serving_llama8B_tp2_random_128_128",
-        "qps_list": [1, 4, 16, "inf"],
+        "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,
@ -141,7 +142,7 @@
 	    "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",
@ -155,19 +156,19 @@
            "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": 128,
 	    "random-output-len": 128,
 	    "ignore-eos": "",
-	    "max_concurrency": 1000,
            "num_prompts": 1000
        }
    },
    {
        "test_name": "serving_llama8B_tp4_random_128_128",
-        "qps_list": [1, 4, 16, "inf"],
+        "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,
@ -176,7 +177,7 @@
 	    "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",
@ -190,13 +191,11 @@
            "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": 128,
 	    "random-output-len": 128,
-	    "ignore-eos": "",
-	    "max_concurrency": 1000,
            "num_prompts": 1000
        }
    }
--- a/.buildkite/nightly-benchmarks/tests/serving-tests-cpu-snc3.json
+++ b/.buildkite/nightly-benchmarks/tests/serving-tests-cpu-snc3.json
@ -1,7 +1,8 @@
 [
    {
        "test_name": "serving_llama8B_pp1_sharegpt",
-        "qps_list": [1, 4, 16, "inf"],
+        "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,
@ -10,7 +11,7 @@
 	    "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": 1,
 	    "dtype": "bfloat16",
 	    "distributed_executor_backend": "mp",
@ -23,17 +24,17 @@
            "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_pp3_sharegpt",
-        "qps_list": [1, 4, 16, "inf"],
+        "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,
@ -42,7 +43,7 @@
 	    "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": 3,
 	    "dtype": "bfloat16",
 	    "distributed_executor_backend": "mp",
@ -55,17 +56,17 @@
            "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_tp2pp6_sharegpt",
-        "qps_list": [1, 4, 16, "inf"],
+        "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,
@ -74,7 +75,7 @@
 	    "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,
            "pipeline_parallel_size": 3,
 	    "dtype": "bfloat16",
@ -88,17 +89,17 @@
            "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_pp1_random_128_128",
-        "qps_list": [1, 4, 16, "inf"],
+        "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,
@ -107,7 +108,7 @@
 	    "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": 1,
 	    "dtype": "bfloat16",
 	    "distributed_executor_backend": "mp",
@ -121,28 +122,28 @@
            "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": 128,
 	    "random-output-len": 128,
 	    "ignore-eos": "",
-	    "max_concurrency": 1000,
            "num_prompts": 1000
        }
    },
    {
        "test_name": "serving_llama8B_pp3_random_128_128",
-        "qps_list": [1, 4, 16, "inf"],
+        "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_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": 3,
 	    "dtype": "bfloat16",
 	    "distributed_executor_backend": "mp",
@ -156,19 +157,19 @@
            "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": 128,
 	    "random-output-len": 128,
 	    "ignore-eos": "",
-	    "max_concurrency": 1000,
            "num_prompts": 1000
        }
    },
    {
        "test_name": "serving_llama8B_tp2pp3_random_128_128",
-        "qps_list": [1, 4, 16, "inf"],
+        "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,
@ -177,7 +178,7 @@
 	    "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,
            "pipeline_parallel_size": 3,
 	    "dtype": "bfloat16",
@ -192,13 +193,12 @@
            "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": 128,
 	    "random-output-len": 128,
 	    "ignore-eos": "",
-	    "max_concurrency": 1000,
            "num_prompts": 1000
        }
    }
--- a/.buildkite/nightly-benchmarks/tests/serving-tests-cpu.json
+++ b/.buildkite/nightly-benchmarks/tests/serving-tests-cpu.json
@ -2,6 +2,7 @@
    {
        "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,
@ -10,7 +11,7 @@
 	    "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",
@ -23,17 +24,17 @@
            "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,
@ -42,7 +43,7 @@
 	    "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",
@ -55,17 +56,17 @@
            "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,
@ -74,7 +75,7 @@
 	    "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",
@ -87,17 +88,17 @@
            "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,
@ -106,7 +107,7 @@
 	    "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",
@ -120,19 +121,19 @@
            "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,
@ -141,7 +142,7 @@
 	    "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",
@ -155,13 +156,12 @@
            "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
        }
    }
--- a/.buildkite/nightly-benchmarks/tests/throughput-tests-cpu.json
+++ b/.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",
--- a/.buildkite/release-pipeline.yaml
+++ b/.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:
@ -11,7 +27,12 @@ steps:
    env:
      DOCKER_BUILDKIT: "1"

+  - block: "Build CUDA 12.6 wheel"
+    key: block-build-cu126-wheel
+    depends_on: ~
+
  - label: "Build wheel - CUDA 12.6"
+    depends_on: block-build-cu126-wheel
    id: build-wheel-cuda-12-6
    agents:
      queue: cpu_queue_postmerge
@ -52,7 +73,7 @@ steps:
      queue: cpu_queue_postmerge
    commands:
      - "aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin public.ecr.aws/q9t5s3a7"
-      - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=12.8.1 --build-arg INSTALL_KV_CONNECTORS=true --tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT --target vllm-openai --progress plain -f docker/Dockerfile ."
+      - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=12.8.1 --build-arg FLASHINFER_AOT_COMPILE=true --build-arg INSTALL_KV_CONNECTORS=true --tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT --target vllm-openai --progress plain -f docker/Dockerfile ."
      - "docker push public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT"

  - label: "Annotate release workflow"
--- a/.buildkite/scripts/hardware_ci/run-amd-test.sh
+++ b/.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 \
--- a/.buildkite/scripts/hardware_ci/run-cpu-test.sh
+++ b/.buildkite/scripts/hardware_ci/run-cpu-test.sh
@ -46,6 +46,11 @@ function cpu_tests() {
    set -e
    python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m"

+  # Run kernel tests
+  docker exec cpu-test-"$NUMA_NODE" bash -c "
+    set -e
+    pytest -v -s tests/kernels/test_onednn.py"
+
  # Run basic model test
  docker exec cpu-test-"$NUMA_NODE" bash -c "
    set -e
@ -99,4 +104,4 @@ function cpu_tests() {

 # All of CPU tests are expected to be finished less than 40 mins.
 export -f cpu_tests
-timeout 1.5h bash -c "cpu_tests $CORE_RANGE $NUMA_NODE"
+timeout 2h bash -c "cpu_tests $CORE_RANGE $NUMA_NODE"
--- a/.buildkite/scripts/hardware_ci/run-tpu-v1-test-part2.sh
+++ b/.buildkite/scripts/hardware_ci/run-tpu-v1-test-part2.sh
@ -61,7 +61,7 @@ echo "Results will be stored in: $RESULTS_DIR"
 echo "--- Installing Python dependencies ---"
 python3 -m pip install --progress-bar off git+https://github.com/thuml/depyf.git \
    && python3 -m pip install --progress-bar off pytest pytest-asyncio tpu-info \
-    && python3 -m pip install --progress-bar off lm_eval[api]==0.4.4 \
+    && python3 -m pip install --progress-bar off "lm-eval @ git+https://github.com/EleutherAI/lm-evaluation-harness.git@206b7722158f58c35b7ffcd53b035fdbdda5126d" \
    && python3 -m pip install --progress-bar off hf-transfer
 echo "--- Python dependencies installed ---"
 export VLLM_USE_V1=1
--- a/.buildkite/scripts/hardware_ci/run-tpu-v1-test.sh
+++ b/.buildkite/scripts/hardware_ci/run-tpu-v1-test.sh
@ -61,7 +61,7 @@ echo "Results will be stored in: $RESULTS_DIR"
 echo "--- Installing Python dependencies ---"
 python3 -m pip install --progress-bar off git+https://github.com/thuml/depyf.git \
    && python3 -m pip install --progress-bar off pytest pytest-asyncio tpu-info \
-    && python3 -m pip install --progress-bar off lm_eval[api]==0.4.4 \
+    && python3 -m pip install --progress-bar off "lm-eval @ git+https://github.com/EleutherAI/lm-evaluation-harness.git@206b7722158f58c35b7ffcd53b035fdbdda5126d" \
    && python3 -m pip install --progress-bar off hf-transfer
 echo "--- Python dependencies installed ---"
 export VLLM_USE_V1=1
--- a/.buildkite/scripts/hardware_ci/run-xpu-test.sh
+++ b/.buildkite/scripts/hardware_ci/run-xpu-test.sh
@ -23,9 +23,13 @@ docker run \
    --device /dev/dri \
    -v /dev/dri/by-path:/dev/dri/by-path \
    --entrypoint="" \
+    -e "HF_TOKEN=${HF_TOKEN}" \
+    -e "ZE_AFFINITY_MASK=${ZE_AFFINITY_MASK}" \
    --name "${container_name}" \
    "${image_name}" \
-    sh -c '
+    bash -c '
+    set -e
+    echo $ZE_AFFINITY_MASK
    VLLM_USE_V1=1 python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m --block-size 64 --enforce-eager
    VLLM_USE_V1=1 python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m --block-size 64 --enforce-eager -tp 2 --distributed-executor-backend ray
    VLLM_USE_V1=1 python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m --block-size 64 --enforce-eager -tp 2 --distributed-executor-backend mp
@ -35,8 +39,8 @@ docker run \
    pytest -v -s v1/sample --ignore=v1/sample/test_logprobs.py --ignore=v1/sample/test_logprobs_e2e.py
    pytest -v -s v1/worker --ignore=v1/worker/test_gpu_model_runner.py
    pytest -v -s v1/structured_output
-    pytest -v -s v1/spec_decode --ignore=v1/spec_decode/test_max_len.py --ignore=v1/spec_decode/test_eagle.py
-    pytest -v -s v1/kv_connector/unit --ignore=v1/kv_connector/unit/test_multi_connector.py --ignore=v1/kv_connector/unit/test_nixl_connector.py
+    pytest -v -s v1/spec_decode --ignore=v1/spec_decode/test_max_len.py --ignore=v1/spec_decode/test_eagle.py --ignore=v1/spec_decode/test_tree_attention.py
+    pytest -v -s v1/kv_connector/unit --ignore=v1/kv_connector/unit/test_multi_connector.py --ignore=v1/kv_connector/unit/test_nixl_connector.py --ignore=v1/kv_connector/unit/test_shared_storage_connector.py
    pytest -v -s v1/test_serial_utils.py
    pytest -v -s v1/test_utils.py
    pytest -v -s v1/test_metrics_reader.py
--- a/.buildkite/scripts/tpu/cleanup_docker.sh
+++ b/.buildkite/scripts/tpu/cleanup_docker.sh
@ -17,7 +17,7 @@ if [ "$disk_usage" -gt "$threshold" ]; then
  # Remove dangling images (those that are not tagged and not used by any container)
  docker image prune -f
  # Remove unused volumes / force the system prune for old images as well.
-  docker volume prune -f && docker system prune --force --filter "until=72h" --all
+  docker volume prune -f && docker system prune --force --filter "until=24h" --all
  echo "Docker images and volumes cleanup completed."
 else
  echo "Disk usage is below $threshold%. No cleanup needed."
--- a/.buildkite/scripts/upload-wheels.sh
+++ b/.buildkite/scripts/upload-wheels.sh
@ -14,8 +14,19 @@ fi
 # Get the single wheel file
 wheel="${wheel_files[0]}"

-# Rename 'linux' to 'manylinux1' in the wheel filename
-new_wheel="${wheel/linux/manylinux1}"
+# Detect architecture and rename 'linux' to appropriate manylinux version
+arch=$(uname -m)
+if [[ $arch == "x86_64" ]]; then
+    manylinux_version="manylinux1"
+elif [[ $arch == "aarch64" ]]; then
+    manylinux_version="manylinux2014"
+else
+    echo "Warning: Unknown architecture $arch, using manylinux1 as default"
+    manylinux_version="manylinux1"
+fi
+
+# Rename 'linux' to the appropriate manylinux version in the wheel filename
+new_wheel="${wheel/linux/$manylinux_version}"
 mv -- "$wheel" "$new_wheel"
 wheel="$new_wheel"

--- a/.buildkite/test-pipeline.yaml
+++ b/.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
@ -98,15 +88,6 @@ steps:
  - pytest -v -s basic_correctness/test_cpu_offload.py
  - VLLM_TEST_ENABLE_ARTIFICIAL_PREEMPT=1 pytest -v -s basic_correctness/test_preemption.py

- label: Chunked Prefill Test
-  mirror_hardwares: [amdexperimental]
-  source_file_dependencies:
-  - vllm/
-  - tests/basic_correctness/test_chunked_prefill
-  commands:
-  - VLLM_ATTENTION_BACKEND=XFORMERS pytest -v -s basic_correctness/test_chunked_prefill.py
-  - VLLM_ATTENTION_BACKEND=FLASH_ATTN pytest -v -s basic_correctness/test_chunked_prefill.py
-
 - label: Core Test # 10min
  mirror_hardwares: [amdexperimental]
  fast_check: true
@ -145,7 +126,8 @@ steps:
  - tests/entrypoints/test_chat_utils
  commands:
  - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-  - pytest -v -s entrypoints/openai --ignore=entrypoints/openai/test_chat_with_tool_reasoning.py --ignore=entrypoints/openai/test_oot_registration.py --ignore=entrypoints/openai/test_tensorizer_entrypoint.py --ignore=entrypoints/openai/correctness/
+  - PYTHONPATH=/vllm-workspace pytest -v -s entrypoints/openai/test_collective_rpc.py # PYTHONPATH is needed to import custom Worker extension
+  - pytest -v -s entrypoints/openai --ignore=entrypoints/openai/test_chat_with_tool_reasoning.py --ignore=entrypoints/openai/test_oot_registration.py --ignore=entrypoints/openai/test_tensorizer_entrypoint.py --ignore=entrypoints/openai/correctness/ --ignore=entrypoints/openai/test_collective_rpc.py
  - pytest -v -s entrypoints/test_chat_utils.py

 - label: Distributed Tests (4 GPUs) # 10min
@ -262,7 +244,9 @@ steps:
    - pytest -v -s v1/core
    - pytest -v -s v1/engine
    - pytest -v -s v1/entrypoints
+    - pytest -v -s v1/executor
    - 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
@ -304,15 +288,6 @@ steps:
    - python3 offline_inference/basic/score.py
    - 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]
-  source_file_dependencies:
-  - vllm/
-  - tests/prefix_caching
-  commands:
-    - pytest -v -s prefix_caching
-
-
 - label: Platform Tests (CUDA)
  mirror_hardwares: [amdexperimental]
  source_file_dependencies:
@ -354,6 +329,7 @@ steps:
    - 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
+    - pytest -v -s compile/test_decorator.py

 - label: PyTorch Fullgraph Smoke Test # 9min
  mirror_hardwares: [amdexperimental]
@ -367,6 +343,7 @@ steps:
  - pytest -v -s compile/piecewise/test_simple.py
  - pytest -v -s compile/piecewise/test_toy_llama.py
  - pytest -v -s compile/piecewise/test_full_cudagraph.py
+  - pytest -v -s compile/piecewise/test_multiple_graphs.py

 - label: PyTorch Fullgraph Test # 18min
  mirror_hardwares: [amdexperimental]
@ -409,6 +386,7 @@ steps:
 - 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/
@ -476,13 +454,11 @@ steps:

 - label: LM Eval Small Models # 53min
  mirror_hardwares: [amdexperimental]
-  working_dir: "/vllm-workspace/.buildkite/lm-eval-harness"
  source_file_dependencies:
  - csrc/
  - vllm/model_executor/layers/quantization
  commands:
-  - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-  - pytest -s -v test_lm_eval_correctness.py --config-list-file=configs/models-small.txt --tp-size=1
+  - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-small.txt --tp-size=1

 - label: OpenAI API correctness
  mirror_hardwares: [amdexperimental]
@ -570,6 +546,15 @@ steps:
  commands:
    - pytest -v -s models/language/pooling -m 'not core_model'

+- label: Multi-Modal Processor Test
+  source_file_dependencies:
+  - vllm/
+  - tests/models/multimodal
+  commands:
+    - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
+    - pytest -v -s models/multimodal/processing --ignore models/multimodal/processing/test_tensor_schema.py
+    - pytest -v -s models/multimodal/processing/test_tensor_schema.py
+
 - label: Multi-Modal Models Test (Standard)
  mirror_hardwares: [amdexperimental]
  torch_nightly: true
@ -579,9 +564,7 @@ steps:
  commands:
    - 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 --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"
+    - pytest -v -s models/multimodal -m core_model --ignore models/multimodal/generation/test_whisper.py --ignore models/multimodal/processing
    - 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
@ -592,7 +575,7 @@ steps:
  - tests/models/multimodal
  commands:
    - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
-    - pytest -v -s --ignore models/multimodal/generation/test_common.py --ignore models/multimodal/processing models/multimodal -m 'not core_model'
+    - pytest -v -s models/multimodal -m 'not core_model' --ignore models/multimodal/generation/test_common.py --ignore models/multimodal/processing

 - label: Multi-Modal Models Test (Extended) 2
  mirror_hardwares: [amdexperimental]
@ -655,8 +638,10 @@ steps:
  - 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/model_executor/layers/quantization/utils/flashinfer_utils.py
  - vllm/v1/attention/backends/flashinfer.py
  - vllm/compilation/fusion.py
+  - vllm/compilation/fusion_attn.py
  commands:
    - nvidia-smi
    - python3 examples/offline_inference/basic/chat.py
@ -669,9 +654,13 @@ steps:
    - 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
+    - pytest -v -s tests/kernels/moe/test_mxfp4_moe.py
    # Fusion
    - pytest -v -s tests/compile/test_fusion_all_reduce.py
+    - pytest -v -s tests/compile/test_fusion_attn.py::test_attention_quant_pattern
+    - pytest -v -s tests/kernels/moe/test_flashinfer.py

 #####  1 GPU test  #####
 #####  multi gpus test  #####
@ -854,3 +843,10 @@ steps:
  commands:
  - export VLLM_WORKER_MULTIPROC_METHOD=spawn
  - pytest -s -v test_lm_eval_correctness.py --config-list-file=configs/models-large.txt --tp-size=4
+
+- label: Qwen MoE EP Test # optional
+  gpu: h200
+  optional: true
+  num_gpus: 2
+  commands:
+    - CUDA_VISIBLE_DEVICES=1,2 VLLM_ALL2ALL_BACKEND=deepep_high_throughput VLLM_USE_DEEP_GEMM=1 VLLM_LOGGING_LEVEL=DEBUG python3 /vllm-workspace/examples/offline_inference/data_parallel.py --model Qwen/Qwen1.5-MoE-A2.7B --tp-size=1  --dp-size=2 --max-model-len 2048
--- a/.github/CODEOWNERS
+++ b/.github/CODEOWNERS
@ -10,6 +10,7 @@
 /vllm/worker/worker.py @zhuohan123 @youkaichao @alexm-redhat @comaniac @njhill
 /vllm/model_executor/layers/sampler.py @zhuohan123 @youkaichao @alexm-redhat @comaniac @njhill
 /vllm/model_executor/layers/quantization @mgoin @robertgshaw2-redhat @tlrmchlsmth @yewentao256
+/vllm/model_executor/layers/mamba @tdoublep
 /vllm/multimodal @DarkLight1337 @ywang96
 /vllm/vllm_flash_attn @LucasWilkinson
 /vllm/lora @jeejeelee
@ -25,11 +26,11 @@ CMakeLists.txt @tlrmchlsmth @LucasWilkinson
 # vLLM V1
 /vllm/v1 @WoosukKwon @robertgshaw2-redhat @njhill @ywang96 @comaniac @alexm-redhat
 /vllm/v1/structured_output @mgoin @russellb @aarnphm
+/vllm/v1/attention/backends/triton_attn.py @tdoublep

 # Test ownership
 /.buildkite/lm-eval-harness @mgoin @simon-mo
 /tests/async_engine @njhill @robertgshaw2-redhat @simon-mo
-/tests/basic_correctness/test_chunked_prefill @rkooo567 @comaniac
 /tests/distributed/test_multi_node_assignment.py @youkaichao
 /tests/distributed/test_pipeline_parallel.py @youkaichao
 /tests/distributed/test_same_node.py @youkaichao
@ -44,6 +45,7 @@ CMakeLists.txt @tlrmchlsmth @LucasWilkinson
 /tests/v1/structured_output @mgoin @russellb @aarnphm
 /tests/weight_loading @mgoin @youkaichao @yewentao256
 /tests/lora @jeejeelee
+/tests/models/language/generation/test_hybrid.py @tdoublep

 # Docs
 /docs @hmellor
@ -72,3 +74,9 @@ mkdocs.yaml @hmellor
 /vllm/model_executor/models/pixtral*.py @patrickvonplaten
 /vllm/transformers_utils/configs/mistral.py @patrickvonplaten
 /vllm/transformers_utils/tokenizers/mistral.py @patrickvonplaten
+
+# Kernels
+/vllm/attention/ops/chunked_prefill_paged_decode.py @tdoublep
+/vllm/attention/ops/triton_unified_attention.py @tdoublep
+
+
--- a/.github/PULL_REQUEST_TEMPLATE.md
+++ b/.github/PULL_REQUEST_TEMPLATE.md
@ -7,8 +7,6 @@ PLEASE FILL IN THE PR DESCRIPTION HERE ENSURING ALL CHECKLIST ITEMS (AT THE BOTT

 ## Test Result

-## (Optional) Documentation Update
-
 ---
 <details>
 <summary> Essential Elements of an Effective PR Description Checklist </summary>
@ -17,6 +15,7 @@ PLEASE FILL IN THE PR DESCRIPTION HERE ENSURING ALL CHECKLIST ITEMS (AT THE BOTT
 - [ ] 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.
+- [ ] (Optional) Release notes update. If your change is user facing, please update the release notes draft in the [Google Doc](https://docs.google.com/document/d/1YyVqrgX4gHTtrstbq8oWUImOyPCKSGnJ7xtTpmXzlRs/edit?tab=t.0).
 </details>

 **BEFORE SUBMITTING, PLEASE READ <https://docs.vllm.ai/en/latest/contributing>** (anything written below this line will be removed by GitHub Actions)
--- a/.github/workflows/lint-and-deploy.yaml
+++ b/.github/workflows/lint-and-deploy.yaml
@ -1,89 +0,0 @@
-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-latest
-    steps:
-      - name: Checkout
-        uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
-        with:
-          fetch-depth: 0
-
-      - name: Set up Helm
-        uses: azure/setup-helm@b9e51907a09c216f16ebe8536097933489208112 # v4.3.0
-        with:
-          version: v3.14.4
-
-       #Python is required because ct lint runs Yamale and yamllint which require Python.
-      - uses: actions/setup-python@42375524e23c412d93fb67b49958b491fce71c38 # v5.4.0
-        with:
-          python-version: '3.13'
-
-      - name: Set up chart-testing
-        uses: helm/chart-testing-action@0d28d3144d3a25ea2cc349d6e59901c4ff469b3b # v2.7.0
-        with:
-          version: v3.10.1
-
-      - name: Run chart-testing (lint)
-        run: ct lint --target-branch ${{ github.event.repository.default_branch }} --chart-dirs examples/online_serving/chart-helm --charts examples/online_serving/chart-helm
-
-      - name: Setup minio
-        run: |
-          docker network create vllm-net
-          docker run -d -p 9000:9000 --name minio --net vllm-net \
-                     -e "MINIO_ACCESS_KEY=minioadmin" \
-                     -e "MINIO_SECRET_KEY=minioadmin" \
-                     -v /tmp/data:/data \
-                     -v /tmp/config:/root/.minio \
-                     minio/minio server /data
-          export AWS_ACCESS_KEY_ID=minioadmin
-          export AWS_SECRET_ACCESS_KEY=minioadmin
-          export AWS_EC2_METADATA_DISABLED=true
-          mkdir opt-125m
-          cd opt-125m && curl -O -Ls "https://huggingface.co/facebook/opt-125m/resolve/main/{pytorch_model.bin,config.json,generation_config.json,merges.txt,special_tokens_map.json,tokenizer_config.json,vocab.json}" && cd ..
-          aws --endpoint-url http://127.0.0.1:9000/ s3 mb s3://testbucket
-          aws --endpoint-url http://127.0.0.1:9000/ s3 cp opt-125m/ s3://testbucket/opt-125m --recursive
-
-      - name: Create kind cluster
-        uses: helm/kind-action@a1b0e391336a6ee6713a0583f8c6240d70863de3 # v1.12.0
-
-      - name: Build the Docker image vllm cpu
-        run: docker buildx build -f docker/Dockerfile.cpu -t vllm-cpu-env .
-
-      - name: Configuration of docker images, network and namespace for the kind cluster
-        run: |
-          docker pull amazon/aws-cli:2.6.4
-          kind load docker-image  amazon/aws-cli:2.6.4 --name chart-testing
-          kind load docker-image vllm-cpu-env:latest --name chart-testing
-          docker network connect vllm-net "$(docker ps -aqf "name=chart-testing-control-plane")"
-          kubectl create ns ns-vllm
-
-      - name: Run chart-testing (install)
-        run: |
-          export AWS_ACCESS_KEY_ID=minioadmin
-          export AWS_SECRET_ACCESS_KEY=minioadmin
-          sleep 30 && kubectl -n ns-vllm logs -f "$(kubectl -n ns-vllm get pods | awk '/deployment/ {print $1;exit}')" &
-          helm install --wait --wait-for-jobs --timeout 5m0s --debug --create-namespace --namespace=ns-vllm test-vllm examples/online_serving/chart-helm -f examples/online_serving/chart-helm/values.yaml --set secrets.s3endpoint=http://minio:9000 --set secrets.s3bucketname=testbucket --set secrets.s3accesskeyid=$AWS_ACCESS_KEY_ID --set secrets.s3accesskey=$AWS_SECRET_ACCESS_KEY --set resources.requests.cpu=1 --set resources.requests.memory=4Gi --set resources.limits.cpu=2 --set resources.limits.memory=5Gi --set image.env[0].name=VLLM_CPU_KVCACHE_SPACE --set image.env[1].name=VLLM_LOGGING_LEVEL --set image.env[2].name=VLLM_CPU_CI_ENV --set-string image.env[0].value="1" --set-string image.env[1].value="DEBUG" --set-string image.env[2].value="1" --set-string extraInit.s3modelpath="opt-125m/" --set-string 'resources.limits.nvidia\.com/gpu=0' --set-string 'resources.requests.nvidia\.com/gpu=0' --set-string image.repository="vllm-cpu-env"
-
-      - name: curl test
-        run: |
-          kubectl -n ns-vllm port-forward service/test-vllm-service 8001:80 &
-          sleep 10
-          CODE="$(curl -v -f --location http://localhost:8001/v1/completions \
-                  --header "Content-Type: application/json" \
-                  --data '{
-                          "model": "opt-125m",
-                          "prompt": "San Francisco is a",
-                          "max_tokens": 7,
-                          "temperature": 0
-                  }'):$CODE"
-          echo "$CODE"
--- a/.github/workflows/publish.yml
+++ b/.github/workflows/publish.yml
@ -1,111 +0,0 @@
-# This workflow will upload a Python Package to Release asset
-# For more information see: https://help.github.com/en/actions/language-and-framework-guides/using-python-with-github-actions
-
-name: Create Release
-
-on:
-  push:
-    tags:
-      - v*
-
-# Needed to create release and upload assets
-permissions:
-  contents: write
-
-jobs:
-  release:
-    # Retrieve tag and create release
-    name: Create Release
-    runs-on: ubuntu-latest
-    outputs:
-      upload_url: ${{ steps.create_release.outputs.upload_url }}
-    steps:
-      - name: Checkout
-        uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
-
-      - name: Extract branch info
-        shell: bash
-        run: |
-          echo "release_tag=${GITHUB_REF#refs/*/}" >> "$GITHUB_ENV"
-
-      - name: Create Release
-        id: create_release
-        uses: actions/github-script@60a0d83039c74a4aee543508d2ffcb1c3799cdea # v7.0.1
-        env:
-          RELEASE_TAG: ${{ env.release_tag }}
-        with:
-          github-token: "${{ secrets.GITHUB_TOKEN }}"
-          script: |
-            const script = require('.github/workflows/scripts/create_release.js')
-            await script(github, context, core)
-
-  # NOTE(simon): No longer build wheel using GitHub Actions. See buildkite's release workflow. 
-  # wheel:
-  #   name: Build Wheel
-  #   runs-on: ${{ matrix.os }}
-  #   needs: release
-
-  #   strategy:
-  #     fail-fast: false
-  #     matrix:
-  #         os: ['ubuntu-20.04']
-  #         python-version: ['3.9', '3.10', '3.11', '3.12']
-  #         pytorch-version: ['2.4.0']  # Must be the most recent version that meets requirements/cuda.txt.
-  #         cuda-version: ['11.8', '12.1']
-
-  #   steps:
-  #     - name: Checkout
-  #       uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
-
-  #     - name: Setup ccache
-  #       uses: hendrikmuhs/ccache-action@ed74d11c0b343532753ecead8a951bb09bb34bc9 # v1.2.14
-  #       with:
-  #         create-symlink: true
-  #         key: ${{ github.job }}-${{ matrix.python-version }}-${{ matrix.cuda-version }}
-
-  #     - name: Set up Linux Env
-  #       if: ${{ runner.os == 'Linux' }}
-  #       run: |
-  #         bash -x .github/workflows/scripts/env.sh
-
-  #     - name: Set up Python
-  #       uses: actions/setup-python@0b93645e9fea7318ecaed2b359559ac225c90a2b # v5.3.0
-  #       with:
-  #           python-version: ${{ matrix.python-version }}
-
-  #     - name: Install CUDA ${{ matrix.cuda-version }}
-  #       run: |
-  #         bash -x .github/workflows/scripts/cuda-install.sh ${{ matrix.cuda-version }} ${{ matrix.os }}
-
-  #     - name: Install PyTorch ${{ matrix.pytorch-version }} with CUDA ${{ matrix.cuda-version }}
-  #       run: |
-  #         bash -x .github/workflows/scripts/pytorch-install.sh ${{ matrix.python-version }} ${{ matrix.pytorch-version }} ${{ matrix.cuda-version }}
-
-  #     - name: Build wheel
-  #       shell: bash
-  #       env:
-  #         CMAKE_BUILD_TYPE: Release # do not compile with debug symbol to reduce wheel size
-  #       run: |
-  #         bash -x .github/workflows/scripts/build.sh ${{ matrix.python-version }} ${{ matrix.cuda-version }}
-  #         wheel_name=$(find dist -name "*whl" -print0 | xargs -0 -n 1 basename)
-  #         asset_name=${wheel_name//"linux"/"manylinux1"}
-  #         echo "wheel_name=${wheel_name}" >> "$GITHUB_ENV"
-  #         echo "asset_name=${asset_name}" >> "$GITHUB_ENV"
-
-  #     - name: Upload Release Asset
-  #       uses: actions/upload-release-asset@e8f9f06c4b078e705bd2ea027f0926603fc9b4d5 # v1.0.2
-  #       env:
-  #         GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-  #       with:
-  #         upload_url: ${{ needs.release.outputs.upload_url }}
-  #         asset_path: ./dist/${{ env.wheel_name }}
-  #         asset_name: ${{ env.asset_name }}
-  #         asset_content_type: application/*
-
-      # (Danielkinz): This last step will publish the .whl to pypi. Warning: untested
-      # - name: Publish package
-      #   uses: pypa/gh-action-pypi-publish@release/v1.8
-      #   with:
-      #     repository-url: https://test.pypi.org/legacy/
-      #     password: ${{ secrets.PYPI_API_TOKEN }}
-      #     skip-existing: true
--- a/.github/workflows/reminder_comment.yml
+++ b/.github/workflows/reminder_comment.yml
@ -12,16 +12,43 @@ jobs:
        uses: actions/github-script@60a0d83039c74a4aee543508d2ffcb1c3799cdea # v7.0.1
        with:
          script: |
-            github.rest.issues.createComment({
-              owner: context.repo.owner,
-              repo: context.repo.repo,
-              issue_number: context.issue.number,
-              body: '👋 Hi! Thank you for contributing to the vLLM project.\n\n' +
-                '💬 Join our developer Slack at https://slack.vllm.ai to discuss your PR in #pr-reviews, coordinate on features in #feat- channels, or join special interest groups in #sig- channels.\n\n' +
-                'Just a reminder: PRs would not trigger full CI run by default. Instead, it would only run `fastcheck` CI which starts running only a small and essential subset of CI tests to quickly catch errors. You can run other CI tests on top of those by going to your `fastcheck` build on Buildkite UI (linked in the PR checks section) and unblock them. If you do not have permission to unblock, ping `simon-mo` or `khluu` to add you in our Buildkite org.\n\n' +
-                'Once the PR is approved and ready to go, your PR reviewer(s) can run CI to test the changes comprehensively before merging.\n\n' +
-                'To run CI, PR reviewers can either: Add `ready` label to the PR or enable auto-merge.\n\n' +
-                '🚀'
-            })
+            try {
+              // Get the PR author
+              const prAuthor = context.payload.pull_request.user.login;
+              
+              // Check if this is the author's first PR in this repository
+              // Use GitHub's search API to find all PRs by this author
+              const { data: searchResults } = await github.rest.search.issuesAndPullRequests({
+                q: `repo:${context.repo.owner}/${context.repo.repo} type:pr author:${prAuthor}`,
+                per_page: 100  
+              });
+              
+              const authorPRCount = searchResults.total_count;
+              
+              console.log(`Found ${authorPRCount} PRs by ${prAuthor}`);
+              
+              // Only post comment if this is the first PR (only one PR by this author)
+              if (authorPRCount === 1) {
+                console.log(`Posting welcome comment for first-time contributor: ${prAuthor}`);
+                await github.rest.issues.createComment({
+                owner: context.repo.owner,
+                repo: context.repo.repo,
+                issue_number: context.issue.number,
+                body: '👋 Hi! Thank you for contributing to the vLLM project.\n\n' +
+                  '💬 Join our developer Slack at https://slack.vllm.ai to discuss your PR in #pr-reviews, coordinate on features in #feat- channels, or join special interest groups in #sig- channels.\n\n' +
+                  'Just a reminder: PRs would not trigger full CI run by default. Instead, it would only run `fastcheck` CI which starts running only a small and essential subset of CI tests to quickly catch errors. \n\n' +
+                  'You ask your reviewers to trigger select CI tests on top of `fastcheck` CI. \n\n' +
+                  'Once the PR is approved and ready to go, your PR reviewer(s) can run CI to test the changes comprehensively before merging.\n\n' +
+                  'To run CI, PR reviewers can either: Add `ready` label to the PR or enable auto-merge.\n\n' +
+                  'If you have any questions, please reach out to us on Slack at https://slack.vllm.ai.\n\n' +
+                  '🚀'
+                });
+              } else {
+                console.log(`Skipping comment for ${prAuthor} - not their first PR (${authorPRCount} PRs found)`);
+              }
+            } catch (error) {
+              console.error('Error checking PR history or posting comment:', error);
+              // Don't fail the workflow, just log the error
+            }
        env:
          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
--- a/.gitignore
+++ b/.gitignore
@ -207,3 +207,6 @@ shellcheck*/

 # Ignore moe/marlin_moe gen code
 csrc/moe/marlin_moe_wna16/kernel_*
+
+# Ignore ep_kernels_workspace folder
+ep_kernels_workspace/
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@ -30,7 +30,7 @@ install(CODE "set(CMAKE_INSTALL_LOCAL_ONLY TRUE)" ALL_COMPONENTS)
 # Supported python versions.  These versions will be searched in order, the
 # first match will be selected.  These should be kept in sync with setup.py.
 #
-set(PYTHON_SUPPORTED_VERSIONS "3.9" "3.10" "3.11" "3.12")
+set(PYTHON_SUPPORTED_VERSIONS "3.9" "3.10" "3.11" "3.12", "3.13")

 # Supported AMD GPU architectures.
 set(HIP_SUPPORTED_ARCHS "gfx906;gfx908;gfx90a;gfx942;gfx950;gfx1030;gfx1100;gfx1101;gfx1200;gfx1201")
@ -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"
@ -351,20 +349,27 @@ 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})

    set(MARLIN_SRCS
-       "csrc/quantization/marlin/dense/marlin_cuda_kernel.cu"
       "csrc/quantization/marlin/sparse/marlin_24_cuda_kernel.cu"
-       "csrc/quantization/marlin/qqq/marlin_qqq_gemm_kernel.cu"
       "csrc/quantization/gptq_marlin/gptq_marlin.cu"
       "csrc/quantization/gptq_marlin/gptq_marlin_repack.cu"
       "csrc/quantization/gptq_marlin/awq_marlin_repack.cu")
    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: ${MARLIN_ARCHS}")
  else()
    message(STATUS "Not building Marlin kernels as no compatible archs found"
@ -745,6 +750,33 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
                     "found in CUDA target architectures")
    endif()
  endif()
+
+  # Only build W4A8 kernels if we are building for something compatible with sm90a
+  cuda_archs_loose_intersection(W4A8_ARCHS "9.0a" "${CUDA_ARCHS}")
+  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.0 AND W4A8_ARCHS)
+    set(SRCS
+       "csrc/quantization/cutlass_w4a8/w4a8_mm_entry.cu")
+
+    set_gencode_flags_for_srcs(
+      SRCS "${SRCS}"
+      CUDA_ARCHS "${W4A8_ARCHS}")
+
+    list(APPEND VLLM_EXT_SRC "${SRCS}")
+
+    message(STATUS "Building W4A8 kernels for archs: ${W4A8_ARCHS}")
+  else()
+    if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.0
+        AND W4A8_ARCHS)
+      message(STATUS "Not building W4A8 kernels as CUDA Compiler version is "
+                     "not >= 12.0, we recommend upgrading to CUDA 12.0 or "
+                     "later if you intend on running w4a16 quantized models on "
+                     "Hopper.")
+    else()
+      message(STATUS "Not building W4A8 kernels as no compatible archs "
+                     "found in CUDA target architectures")
+    endif()
+  endif()
+
 # if CUDA endif
 endif()

@ -785,7 +817,9 @@ set(VLLM_MOE_EXT_SRC
  "csrc/moe/topk_softmax_kernels.cu")

 if(VLLM_GPU_LANG STREQUAL "CUDA")
-  list(APPEND VLLM_MOE_EXT_SRC "csrc/moe/moe_wna16.cu")
+  list(APPEND VLLM_MOE_EXT_SRC
+    "csrc/moe/moe_wna16.cu"
+    "csrc/moe/grouped_topk_kernels.cu")
 endif()

 if(VLLM_GPU_LANG STREQUAL "CUDA")
@ -854,6 +888,10 @@ 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})

--- a/README.md
+++ b/README.md
@ -18,14 +18,15 @@ Easy, fast, and cheap LLM serving for everyone

 *Latest News* 🔥

+- [2025/08] We hosted [vLLM Shanghai Meetup](https://mp.weixin.qq.com/s/pDmAXHcN7Iqc8sUKgJgGtg) focusing on building, developing, and integrating with vLLM! Please find the meetup slides [here](https://drive.google.com/drive/folders/1OvLx39wnCGy_WKq8SiVKf7YcxxYI3WCH).
 - [2025/08] We hosted [vLLM 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/).
 - [2025/01] 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>

+- [2025/05] We hosted [NYC vLLM Meetup](https://lu.ma/c1rqyf1f)! Please find the meetup slides [here](https://docs.google.com/presentation/d/1_q_aW_ioMJWUImf1s1YM-ZhjXz8cUeL0IJvaquOYBeA/edit?usp=sharing).
 - [2025/04] We hosted [Asia Developer Day](https://www.sginnovate.com/event/limited-availability-morning-evening-slots-remaining-inaugural-vllm-asia-developer-day)! Please find the meetup slides from the vLLM team [here](https://docs.google.com/presentation/d/19cp6Qu8u48ihB91A064XfaXruNYiBOUKrBxAmDOllOo/edit?usp=sharing).
 - [2025/03] We hosted [vLLM x Ollama Inference Night](https://lu.ma/vllm-ollama)! Please find the meetup slides from the vLLM team [here](https://docs.google.com/presentation/d/16T2PDD1YwRnZ4Tu8Q5r6n53c5Lr5c73UV9Vd2_eBo4U/edit?usp=sharing).
 - [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).
--- a/benchmarks/README.md
+++ b/benchmarks/README.md
@ -22,6 +22,25 @@ 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>ShareGPT4Video (Video)</strong></td>
+      <td style="text-align: center;">✅</td>
+      <td style="text-align: center;">✅</td>
+      <td>
+        <code>git clone https://huggingface.co/datasets/ShareGPT4Video/ShareGPT4Video</code>
+      </td>
+    </tr>
    <tr>
      <td><strong>BurstGPT</strong></td>
      <td style="text-align: center;">✅</td>
@ -29,7 +48,7 @@ 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 +59,18 @@ become available.
      <td style="text-align: center;">✅</td>
      <td><code>synthetic</code></td>
    </tr>
+    <tr>
+      <td><strong>RandomMultiModal (Image/Video)</strong></td>
+      <td style="text-align: center;">🟡</td>
+      <td style="text-align: center;">🚧</td>
+      <td><code>synthetic</code> </td>
+    </tr>
+    <tr>
+      <td><strong>Prefix Repetition</strong></td>
+      <td style="text-align: center;">✅</td>
+      <td style="text-align: center;">✅</td>
+      <td><code>synthetic</code></td>
+    </tr>
    <tr>
      <td><strong>HuggingFace-VisionArena</strong></td>
      <td style="text-align: center;">✅</td>
@ -177,6 +208,7 @@ vllm serve Qwen/Qwen2-VL-7B-Instruct
 ```bash
 vllm bench serve \
  --backend openai-chat \
+  --endpoint-type openai-chat \
  --model Qwen/Qwen2-VL-7B-Instruct \
  --endpoint /v1/chat/completions \
  --dataset-name hf \
@ -213,6 +245,7 @@ vllm serve Qwen/Qwen2-VL-7B-Instruct
 ```bash
 vllm bench serve \
  --backend openai-chat \
+  --endpoint-type openai-chat \
  --model Qwen/Qwen2-VL-7B-Instruct \
  --endpoint /v1/chat/completions \
  --dataset-name hf \
@ -227,6 +260,7 @@ vllm bench serve \
 ```bash
 vllm bench serve \
  --backend openai-chat \
+  --endpoint-type openai-chat \
  --model Qwen/Qwen2-VL-7B-Instruct \
  --endpoint /v1/chat/completions \
  --dataset-name hf \
@ -581,6 +615,20 @@ 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
@ -616,3 +664,139 @@ 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
+```
+
+### Videos (ShareGPT4Video)
+
+Start vLLM:
+
+```bash
+python -m vllm.entrypoints.openai.api_server \
+  --model Qwen/Qwen2.5-VL-7B-Instruct \
+  --dtype bfloat16 \
+  --limit-mm-per-prompt '{"video": 1}' \
+  --allowed-local-media-path /path/to/sharegpt4video/videos
+```
+
+Send requests with videos:
+
+```bash
+python benchmarks/benchmark_serving.py \
+  --backend openai-chat \
+  --model Qwen/Qwen2.5-VL-7B-Instruct \
+  --dataset-name sharegpt \
+  --dataset-path /path/to/ShareGPT4Video/llava_v1_5_mix665k_with_video_chatgpt72k_share4video28k.json \
+  --num-prompts 100 \
+  --save-result \
+  --result-dir ~/vllm_benchmark_results \
+  --save-detailed \
+  --endpoint /v1/chat/completion
+```
+
+### Synthetic Random Images (random-mm)
+
+Generate synthetic image inputs alongside random text prompts to stress-test vision models without external datasets.
+
+Notes:
+
+- Works only with online benchmark via the OpenAI  backend (`--backend openai-chat`) and endpoint `/v1/chat/completions`.
+- Video sampling is not yet implemented.
+
+Start the server (example):
+
+```bash
+vllm serve Qwen/Qwen2.5-VL-3B-Instruct \
+  --dtype bfloat16 \
+  --max-model-len 16384 \
+  --limit-mm-per-prompt '{"image": 3, "video": 0}' \
+  --mm-processor-kwargs max_pixels=1003520
+```
+
+Benchmark. It is recommended to use the flag `--ignore-eos` to simulate real responses. You can set the size of the output via the arg `random-output-len`.
+
+Ex.1: Fixed number of items and a single image resolutionm, enforcing generation of approx 40 tokens:
+
+```bash
+vllm bench serve \
+  --backend openai-chat \
+  --model Qwen/Qwen2.5-VL-3B-Instruct \
+  --endpoint /v1/chat/completions \
+  --dataset-name random-mm \
+  --num-prompts 100 \
+  --max-concurrency 10 \
+  --random-prefix-len 25 \
+  --random-input-len 300 \
+  --random-output-len 40 \
+  --random-range-ratio 0.2 \
+  --random-mm-base-items-per-request 2 \
+  --random-mm-limit-mm-per-prompt '{"image": 3, "video": 0}' \
+  --random-mm-bucket-config '{(224, 224, 1): 1.0}' \
+  --request-rate inf \
+  --ignore-eos \
+  --seed 42
+```
+
+The number of items per request can be controlled by passing multiple image buckets:
+
+```bash
+  --random-mm-base-items-per-request 2 \
+  --random-mm-num-mm-items-range-ratio 0.5 \
+  --random-mm-limit-mm-per-prompt '{"image": 4, "video": 0}' \
+  --random-mm-bucket-config '{(256, 256, 1): 0.7, (720, 1280, 1): 0.3}' \
+```
+
+Flags specific to `random-mm`:
+
+- `--random-mm-base-items-per-request`: base number of multimodal items per request.
+- `--random-mm-num-mm-items-range-ratio`: vary item count uniformly in the closed integer range [floor(n·(1−r)), ceil(n·(1+r))]. Set r=0 to keep it fixed; r=1 allows 0 items.
+- `--random-mm-limit-mm-per-prompt`: per-modality hard caps, e.g. '{"image": 3, "video": 0}'.
+- `--random-mm-bucket-config`: dict mapping (H, W, T) → probability. Entries with probability 0 are removed; remaining probabilities are renormalized to sum to 1. Use T=1 for images. Set any T>1 for videos (video sampling not yet supported).
+
+Behavioral notes:
+
+- If the requested base item count cannot be satisfied under the provided per-prompt limits, the tool raises an error rather than silently clamping.
+
+How sampling works:
+
+- Determine per-request item count k by sampling uniformly from the integer range defined by `--random-mm-base-items-per-request` and `--random-mm-num-mm-items-range-ratio`, then clamp k to at most the sum of per-modality limits.
+- For each of the k items, sample a bucket (H, W, T) according to the normalized probabilities in `--random-mm-bucket-config`, while tracking how many items of each modality have been added.
+- If a modality (e.g., image) reaches its limit from `--random-mm-limit-mm-per-prompt`, all buckets of that modality are excluded and the remaining bucket probabilities are renormalized before continuing.
+This should be seen as an edge case, and if this behavior can be avoided by setting `--random-mm-limit-mm-per-prompt` to a large number. Note that this might result in errors due to engine config `--limit-mm-per-prompt`.
+- The resulting request contains synthetic image data in `multi_modal_data` (OpenAI Chat format). When `random-mm` is used with the OpenAI Chat backend, prompts remain text and MM content is attached via `multi_modal_data`.
+
+</details>
--- a/benchmarks/backend_request_func.py
+++ b/benchmarks/backend_request_func.py
@ -34,6 +34,7 @@ class RequestFuncInput:
    multi_modal_content: Optional[dict | list[dict]] = None
    ignore_eos: bool = False
    language: Optional[str] = None
+    request_id: Optional[str] = None


@dataclass
@ -71,6 +72,9 @@ async def async_request_tgi(
            "inputs": request_func_input.prompt,
            "parameters": params,
        }
+        headers = None
+        if request_func_input.request_id:
+            headers = {"x-request-id": request_func_input.request_id}
        output = RequestFuncOutput()
        output.prompt_len = request_func_input.prompt_len
        if request_func_input.ignore_eos:
@ -82,7 +86,9 @@ async def async_request_tgi(
        st = time.perf_counter()
        most_recent_timestamp = st
        try:
-            async with session.post(url=api_url, json=payload) as response:
+            async with session.post(
+                url=api_url, json=payload, headers=headers
+            ) as response:
                if response.status == 200:
                    async for chunk_bytes in response.content:
                        chunk_bytes = chunk_bytes.strip()
@ -145,6 +151,9 @@ async def async_request_trt_llm(
        }
        if request_func_input.ignore_eos:
            payload["min_length"] = request_func_input.output_len
+        headers = None
+        if request_func_input.request_id:
+            headers = {"x-request-id": request_func_input.request_id}
        output = RequestFuncOutput()
        output.prompt_len = request_func_input.prompt_len

@ -152,7 +161,9 @@ async def async_request_trt_llm(
        st = time.perf_counter()
        most_recent_timestamp = st
        try:
-            async with session.post(url=api_url, json=payload) as response:
+            async with session.post(
+                url=api_url, json=payload, headers=headers
+            ) as response:
                if response.status == 200:
                    async for chunk_bytes in response.content:
                        chunk_bytes = chunk_bytes.strip()
@ -211,6 +222,8 @@ async def async_request_deepspeed_mii(
            "top_p": 1.0,
        }
        headers = {"Authorization": f"Bearer {os.environ.get('OPENAI_API_KEY')}"}
+        if request_func_input.request_id:
+            headers["x-request-id"] = request_func_input.request_id

        output = RequestFuncOutput()
        output.prompt_len = request_func_input.prompt_len
@ -283,6 +296,8 @@ async def async_request_openai_completions(
        if request_func_input.extra_body:
            payload.update(request_func_input.extra_body)
        headers = {"Authorization": f"Bearer {os.environ.get('OPENAI_API_KEY')}"}
+        if request_func_input.request_id:
+            headers["x-request-id"] = request_func_input.request_id

        output = RequestFuncOutput()
        output.prompt_len = request_func_input.prompt_len
@ -395,6 +410,8 @@ async def async_request_openai_chat_completions(
            "Content-Type": "application/json",
            "Authorization": f"Bearer {os.environ.get('OPENAI_API_KEY')}",
        }
+        if request_func_input.request_id:
+            headers["x-request-id"] = request_func_input.request_id

        output = RequestFuncOutput()
        output.prompt_len = request_func_input.prompt_len
@ -491,6 +508,8 @@ async def async_request_openai_audio(
        headers = {
            "Authorization": f"Bearer {os.environ.get('OPENAI_API_KEY')}",
        }
+        if request_func_input.request_id:
+            headers["x-request-id"] = request_func_input.request_id

        # Send audio file
        def to_bytes(y, sr):
--- a/benchmarks/benchmark_dataset.py
+++ b/benchmarks/benchmark_dataset.py
@ -19,6 +19,7 @@ import logging
 import random
 from abc import ABC, abstractmethod
 from collections.abc import Mapping
+from copy import deepcopy
 from dataclasses import dataclass
 from functools import cache
 from io import BytesIO
@ -54,6 +55,7 @@ class SampleRequest:
    expected_output_len: int
    multi_modal_data: Optional[Union[MultiModalDataDict, dict, list[dict]]] = None
    lora_request: Optional[LoRARequest] = None
+    request_id: Optional[str] = None


 # -----------------------------------------------------------------------------
@ -155,7 +157,10 @@ class BenchmarkDataset(ABC):

    @abstractmethod
    def sample(
-        self, tokenizer: PreTrainedTokenizerBase, num_requests: int
+        self,
+        tokenizer: PreTrainedTokenizerBase,
+        num_requests: int,
+        request_id_prefix: str = "",
    ) -> list[SampleRequest]:
        """
        Abstract method to generate sample requests from the dataset.
@ -167,6 +172,7 @@ class BenchmarkDataset(ABC):
            tokenizer (PreTrainedTokenizerBase): The tokenizer to be used
             for processing the dataset's text.
            num_requests (int): The number of sample requests to generate.
+            request_id_prefix (str) The prefix of request_id.

        Returns:
            list[SampleRequest]: A list of sample requests generated from the
@ -175,7 +181,10 @@ class BenchmarkDataset(ABC):
        raise NotImplementedError("sample must be implemented in subclasses.")

    def maybe_oversample_requests(
-        self, requests: list[SampleRequest], num_requests: int
+        self,
+        requests: list[SampleRequest],
+        num_requests: int,
+        request_id_prefix: str = "",
    ) -> None:
        """
        Oversamples the list of requests if its size is less than the desired
@ -183,11 +192,18 @@ class BenchmarkDataset(ABC):

        Args:
            requests (List[SampleRequest]): The current list of sampled
-            requests.  num_requests (int): The target number of requests.
+            requests.
+            num_requests (int): The target number of requests.
+            request_id_prefix (str) The prefix of the request ids.
        """
        if len(requests) < num_requests:
            random.seed(self.random_seed)
-            additional = random.choices(requests, k=num_requests - len(requests))
+            additional = deepcopy(
+                random.choices(requests, k=num_requests - len(requests))
+            )
+            for i in range(len(additional)):
+                req = additional[i]
+                req.request_id = request_id_prefix + str(len(requests) + i)
            requests.extend(additional)
            logger.info("Oversampled requests to reach %d total samples.", num_requests)

@ -277,6 +293,41 @@ def process_image(image: Any) -> Mapping[str, Any]:
    )


+def process_video(video: Any) -> Mapping[str, Any]:
+    """
+    Process a single video input and return a multimedia content dictionary.
+
+    Supports the following input types:
+
+    1. Dictionary with raw video bytes: - Expects a dict with a 'bytes' key
+       containing raw video data.
+
+    2. String input: - Treats the string as a URL or local file path.  -
+       Prepends "file://" if the string doesn't start with "http://" or
+       "file://".  - Returns a dictionary with the image URL.
+
+    Raises:
+        ValueError: If the input is not a supported type.
+    """
+    if isinstance(video, dict) and "bytes" in video:
+        video_bytes = video["bytes"]
+        video_base64 = base64.b64encode(video_bytes).decode("utf-8")
+        return {
+            "type": "video_url",
+            "video_url": {"url": f"data:video/mp4;base64,{video_base64}"},
+        }
+
+    if isinstance(video, str):
+        video_url = (
+            video if video.startswith(("http://", "file://")) else f"file://{video}"
+        )
+        return {"type": "video_url", "video_url": {"url": video_url}}
+
+    raise ValueError(
+        f"Invalid video input {video}. Must be a string of local path/remote url, or a dictionary with raw video bytes in the form of `{{'bytes': raw_video_bytes}}`."  # noqa: E501
+    )
+
+
 # -----------------------------------------------------------------------------
 # Random Dataset Implementation (Synthetic Data)
 # -----------------------------------------------------------------------------
@ -303,6 +354,7 @@ class RandomDataset(BenchmarkDataset):
        range_ratio: float = DEFAULT_RANGE_RATIO,
        input_len: int = DEFAULT_INPUT_LEN,
        output_len: int = DEFAULT_OUTPUT_LEN,
+        request_id_prefix: str = "",
        **kwargs,
    ) -> list[SampleRequest]:
        # Enforce range_ratio < 1
@ -363,8 +415,10 @@ class RandomDataset(BenchmarkDataset):
                    prompt=prompt,
                    prompt_len=total_input_len,
                    expected_output_len=int(output_lens[i]),
+                    request_id=request_id_prefix + str(i),
                )
            )
+
        return requests


@ -406,9 +460,11 @@ class ShareGPTDataset(BenchmarkDataset):
        max_loras: Optional[int] = None,
        output_len: Optional[int] = None,
        enable_multimodal_chat: bool = False,
+        request_id_prefix: str = "",
        **kwargs,
    ) -> list:
        samples: list = []
+        ind = 0
        for entry in self.data:
            if len(samples) >= num_requests:
                break
@ -430,17 +486,26 @@ class ShareGPTDataset(BenchmarkDataset):
                skip_min_output_len_check=output_len is not None,
            ):
                continue
+            if image_path := entry.get("image"):
+                mm_content = process_image(image_path)
+            elif video_path := entry.get("video"):
+                mm_content = process_video(video_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,
+                    request_id=request_id_prefix + str(ind),
                )
            )
-        self.maybe_oversample_requests(samples, num_requests)
+            ind += 1
+        self.maybe_oversample_requests(samples, num_requests, request_id_prefix)
        return samples


@ -506,10 +571,11 @@ class CustomDataset(BenchmarkDataset):
        output_len: Optional[int] = None,
        enable_multimodal_chat: bool = False,
        skip_chat_template: bool = False,
+        request_id_prefix: str = "",
        **kwargs,
    ) -> list:
        sampled_requests = []
-        for item in self.data:
+        for i, item in enumerate(self.data):
            if len(sampled_requests) >= num_requests:
                break
            prompt = item["prompt"]
@ -528,9 +594,12 @@ class CustomDataset(BenchmarkDataset):
                    prompt=prompt,
                    prompt_len=prompt_len,
                    expected_output_len=output_len,
+                    request_id=request_id_prefix + str(i),
                )
            )
-        self.maybe_oversample_requests(sampled_requests, num_requests)
+        self.maybe_oversample_requests(
+            sampled_requests, num_requests, request_id_prefix
+        )

        return sampled_requests

@ -572,6 +641,7 @@ class SonnetDataset(BenchmarkDataset):
        input_len: int = DEFAULT_INPUT_LEN,
        output_len: int = DEFAULT_OUTPUT_LEN,
        return_prompt_formatted: bool = False,
+        request_id_prefix: str = "",
        **kwargs,
    ) -> list:
        # Calculate average token length for a poem line.
@ -597,6 +667,7 @@ class SonnetDataset(BenchmarkDataset):
        prefix_lines = self.data[:num_prefix_lines]

        samples = []
+        ind = 0
        while len(samples) < num_requests:
            extra_lines = random.choices(
                self.data, k=num_input_lines - num_prefix_lines
@ -607,14 +678,17 @@ class SonnetDataset(BenchmarkDataset):
                msg, add_generation_prompt=True, tokenize=False
            )
            prompt_len = len(tokenizer(prompt_formatted).input_ids)
+
            if prompt_len <= input_len:
                samples.append(
                    SampleRequest(
                        prompt=prompt_formatted if return_prompt_formatted else prompt,
                        prompt_len=prompt_len,
                        expected_output_len=output_len,
+                        request_id=request_id_prefix + str(ind),
                    )
                )
+                ind += 1
        return samples


@ -666,6 +740,7 @@ class BurstGPTDataset(BenchmarkDataset):
        num_requests: int,
        max_loras: Optional[int] = None,
        lora_path: Optional[str] = None,
+        request_id_prefix: str = "",
        **kwargs,
    ) -> list[SampleRequest]:
        samples = []
@ -687,6 +762,7 @@ class BurstGPTDataset(BenchmarkDataset):
                    prompt_len=input_len,
                    expected_output_len=output_len,
                    lora_request=lora_req,
+                    request_id=request_id_prefix + str(i),
                )
            )
        return samples
@ -746,12 +822,14 @@ class ConversationDataset(HuggingFaceDataset):
        num_requests: int,
        output_len: Optional[int] = None,
        enable_multimodal_chat: bool = False,
+        request_id_prefix: str = "",
        **kwargs,
    ) -> list:
        # Filter examples with at least 2 conversations
        filtered_data = self.data.filter(lambda x: len(x["conversations"]) >= 2)
        sampled_requests = []
        dynamic_output = output_len is None
+        ind = 0

        for item in filtered_data:
            if len(sampled_requests) >= num_requests:
@ -779,9 +857,13 @@ class ConversationDataset(HuggingFaceDataset):
                    prompt_len=prompt_len,
                    expected_output_len=output_len,
                    multi_modal_data=mm_content,
+                    request_id=request_id_prefix + str(ind),
                )
            )
-        self.maybe_oversample_requests(sampled_requests, num_requests)
+            ind += 1
+        self.maybe_oversample_requests(
+            sampled_requests, num_requests, request_id_prefix
+        )
        return sampled_requests


@ -808,11 +890,12 @@ class VisionArenaDataset(HuggingFaceDataset):
        num_requests: int,
        output_len: Optional[int] = None,
        enable_multimodal_chat: bool = False,
+        request_id_prefix: str = "",
        **kwargs,
    ) -> list:
        output_len = output_len if output_len is not None else self.DEFAULT_OUTPUT_LEN
        sampled_requests = []
-        for item in self.data:
+        for i, item in enumerate(self.data):
            if len(sampled_requests) >= num_requests:
                break
            parser_fn = self.SUPPORTED_DATASET_PATHS.get(self.dataset_path)
@ -832,9 +915,12 @@ class VisionArenaDataset(HuggingFaceDataset):
                    prompt_len=prompt_len,
                    expected_output_len=output_len,
                    multi_modal_data=mm_content,
+                    request_id=request_id_prefix + str(i),
                )
            )
-        self.maybe_oversample_requests(sampled_requests, num_requests)
+        self.maybe_oversample_requests(
+            sampled_requests, num_requests, request_id_prefix
+        )
        return sampled_requests


@ -864,15 +950,18 @@ class InstructCoderDataset(HuggingFaceDataset):
        num_requests: int,
        output_len: Optional[int] = None,
        enable_multimodal_chat: bool = False,
+        request_id_prefix: str = "",
        **kwargs,
    ) -> list:
        output_len = output_len if output_len is not None else self.DEFAULT_OUTPUT_LEN
        sampled_requests = []
-        for item in self.data:
+        for i, item in enumerate(self.data):
            if len(sampled_requests) >= num_requests:
                break
-            prompt = f"{item['input']}\n\n{item['instruction']} Just output \
-            the code, do not include any explanation."
+            prompt = (
+                f"{item['input']}\n\n{item['instruction']} Just output "
+                "the code, do not include any explanation."
+            )

            # apply template
            prompt = tokenizer.apply_chat_template(
@ -886,9 +975,12 @@ class InstructCoderDataset(HuggingFaceDataset):
                    prompt=prompt,
                    prompt_len=prompt_len,
                    expected_output_len=output_len,
+                    request_id=request_id_prefix + str(i),
                )
            )
-        self.maybe_oversample_requests(sampled_requests, num_requests)
+        self.maybe_oversample_requests(
+            sampled_requests, num_requests, request_id_prefix
+        )
        return sampled_requests


@ -918,12 +1010,13 @@ class MTBenchDataset(HuggingFaceDataset):
        num_requests: int,
        output_len: Optional[int] = None,
        enable_multimodal_chat: bool = False,
+        request_id_prefix: str = "",
        **kwargs,
    ) -> list:
        output_len = output_len if output_len is not None else self.DEFAULT_OUTPUT_LEN
        sampled_requests = []

-        for item in self.data:
+        for i, item in enumerate(self.data):
            if len(sampled_requests) >= num_requests:
                break
            prompt = item["turns"][0]
@ -941,9 +1034,12 @@ class MTBenchDataset(HuggingFaceDataset):
                    prompt=prompt,
                    prompt_len=prompt_len,
                    expected_output_len=output_len,
+                    request_id=request_id_prefix + str(i),
                )
            )
-        self.maybe_oversample_requests(sampled_requests, num_requests)
+        self.maybe_oversample_requests(
+            sampled_requests, num_requests, request_id_prefix
+        )
        return sampled_requests


@ -968,10 +1064,12 @@ class AIMODataset(HuggingFaceDataset):
        tokenizer: PreTrainedTokenizerBase,
        num_requests: int,
        output_len: Optional[int] = None,
+        request_id_prefix: str = "",
        **kwargs,
    ) -> list:
        sampled_requests = []
        dynamic_output = output_len is None
+        ind = 0

        for item in self.data:
            if len(sampled_requests) >= num_requests:
@ -994,9 +1092,13 @@ class AIMODataset(HuggingFaceDataset):
                    prompt_len=prompt_len,
                    expected_output_len=output_len,
                    multi_modal_data=None,
+                    request_id=request_id_prefix + str(ind),
                )
            )
-        self.maybe_oversample_requests(sampled_requests, num_requests)
+            ind += 1
+        self.maybe_oversample_requests(
+            sampled_requests, num_requests, request_id_prefix
+        )
        return sampled_requests


@ -1066,12 +1168,18 @@ class NextEditPredictionDataset(HuggingFaceDataset):
        "zed-industries/zeta": _format_zeta_prompt,
    }

-    def sample(self, tokenizer: PreTrainedTokenizerBase, num_requests: int, **kwargs):
+    def sample(
+        self,
+        tokenizer: PreTrainedTokenizerBase,
+        num_requests: int,
+        request_id_prefix: str = "",
+        **kwargs,
+    ):
        formatting_prompt_func = self.MAPPING_PROMPT_FUNCS.get(self.dataset_path)
        if formatting_prompt_func is None:
            raise ValueError(f"Unsupported dataset path: {self.dataset_path}")
        samples = []
-        for sample in self.data:
+        for i, sample in enumerate(self.data):
            sample = formatting_prompt_func(sample)
            samples.append(
                SampleRequest(
@ -1080,11 +1188,12 @@ class NextEditPredictionDataset(HuggingFaceDataset):
                    expected_output_len=len(
                        tokenizer(sample["expected_output"]).input_ids
                    ),
+                    request_id=request_id_prefix + str(i),
                )
            )
            if len(samples) >= num_requests:
                break
-        self.maybe_oversample_requests(samples, num_requests)
+        self.maybe_oversample_requests(samples, num_requests, request_id_prefix)
        return samples


@ -1133,6 +1242,7 @@ class ASRDataset(HuggingFaceDataset):
        tokenizer: PreTrainedTokenizerBase,
        num_requests: int,
        output_len: Optional[int] = None,
+        request_id_prefix: str = "",
        **kwargs,
    ) -> list:
        import librosa
@ -1142,6 +1252,7 @@ class ASRDataset(HuggingFaceDataset):
        prompt_len = len(tokenizer(prompt).input_ids)
        sampled_requests = []
        skipped = 0
+        ind = 0
        for item in self.data:
            if len(sampled_requests) >= num_requests:
                break
@ -1160,8 +1271,10 @@ class ASRDataset(HuggingFaceDataset):
                    prompt_len=prompt_len,
                    expected_output_len=output_len,
                    multi_modal_data=mm_content,
+                    request_id=request_id_prefix + str(ind),
                )
            )
+            ind += 1
        if skipped:
            logger.warning(
                "%d samples discarded from dataset due to"
@ -1169,5 +1282,7 @@ class ASRDataset(HuggingFaceDataset):
                " what Whisper supports.",
                skipped,
            )
-        self.maybe_oversample_requests(sampled_requests, num_requests)
+        self.maybe_oversample_requests(
+            sampled_requests, num_requests, request_id_prefix
+        )
        return sampled_requests
--- a/benchmarks/benchmark_serving.py
+++ b/benchmarks/benchmark_serving.py
@ -375,11 +375,12 @@ async def benchmark(
                    rps_change_events.append({"rps": rps_val, "timestamp": timestamp})
                last_int_rps = current_int_rps

-        prompt, prompt_len, output_len, mm_content = (
+        prompt, prompt_len, output_len, mm_content, request_id = (
            request.prompt,
            request.prompt_len,
            request.expected_output_len,
            request.multi_modal_data,
+            request.request_id,
        )
        req_model_id, req_model_name = model_id, model_name
        if lora_modules:
@ -397,6 +398,7 @@ async def benchmark(
            multi_modal_content=mm_content,
            ignore_eos=ignore_eos,
            extra_body=extra_body,
+            request_id=request_id,
        )
        task = limited_request_func(request_func_input=request_func_input, pbar=pbar)
        tasks.append(asyncio.create_task(task))
@ -665,6 +667,7 @@ def main(args: argparse.Namespace):
            tokenizer=tokenizer,
            output_len=args.custom_output_len,
            skip_chat_template=args.custom_skip_chat_template,
+            request_id_prefix=args.request_id_prefix,
        )

    elif args.dataset_name == "sonnet":
@ -678,6 +681,7 @@ def main(args: argparse.Namespace):
                prefix_len=args.sonnet_prefix_len,
                tokenizer=tokenizer,
                return_prompt_formatted=False,
+                request_id_prefix=args.request_id_prefix,
            )
        else:
            assert tokenizer.chat_template or tokenizer.default_chat_template, (
@ -690,6 +694,7 @@ def main(args: argparse.Namespace):
                prefix_len=args.sonnet_prefix_len,
                tokenizer=tokenizer,
                return_prompt_formatted=True,
+                request_id_prefix=args.request_id_prefix,
            )

    elif args.dataset_name == "hf":
@ -751,6 +756,7 @@ def main(args: argparse.Namespace):
            num_requests=args.num_prompts,
            tokenizer=tokenizer,
            output_len=args.hf_output_len,
+            request_id_prefix=args.request_id_prefix,
        )

    else:
@ -762,10 +768,15 @@ def main(args: argparse.Namespace):
                tokenizer=tokenizer,
                num_requests=args.num_prompts,
                output_len=args.sharegpt_output_len,
+                request_id_prefix=args.request_id_prefix,
            ),
            "burstgpt": lambda: BurstGPTDataset(
                random_seed=args.seed, dataset_path=args.dataset_path
-            ).sample(tokenizer=tokenizer, num_requests=args.num_prompts),
+            ).sample(
+                tokenizer=tokenizer,
+                num_requests=args.num_prompts,
+                request_id_prefix=args.request_id_prefix,
+            ),
            "random": lambda: RandomDataset(dataset_path=args.dataset_path).sample(
                tokenizer=tokenizer,
                num_requests=args.num_prompts,
@ -773,6 +784,7 @@ def main(args: argparse.Namespace):
                input_len=args.random_input_len,
                output_len=args.random_output_len,
                range_ratio=args.random_range_ratio,
+                request_id_prefix=args.request_id_prefix,
            ),
        }

@ -1118,6 +1130,13 @@ def create_argument_parser():
        "goodput, refer to DistServe paper: https://arxiv.org/pdf/2401.09670 "
        "and the blog: https://hao-ai-lab.github.io/blogs/distserve",
    )
+    parser.add_argument(
+        "--request-id-prefix",
+        type=str,
+        required=False,
+        default="benchmark-serving",
+        help="Specify the prefix of request id.",
+    )

    # group for dataset specific arguments
    custom_group = parser.add_argument_group("custom dataset options")
--- a/benchmarks/benchmark_throughput.py
+++ b/benchmarks/benchmark_throughput.py
@ -597,8 +597,8 @@ def validate_args(args):
    # https://github.com/vllm-project/vllm/issues/16222
    if args.data_parallel_size > 1:
        raise ValueError(
-            "Data parallel is not supported in offline benchmark, \
-            please use benchmark serving instead"
+            "Data parallel is not supported in offline benchmark, "
+            "please use benchmark serving instead"
        )


--- a/benchmarks/disagg_benchmarks/disagg_prefill_proxy_server.py
+++ b/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()
--- a/benchmarks/disagg_benchmarks/rate_limiter.py
+++ b/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
--- a/benchmarks/disagg_benchmarks/request_queue.py
+++ b/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
--- a/benchmarks/kernels/benchmark_aqlm.py
+++ b/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())
--- a/benchmarks/kernels/benchmark_grouped_gemm_cutlass.py
+++ b/benchmarks/kernels/benchmark_grouped_gemm_cutlass.py
@ -80,6 +80,11 @@ def bench_run(
        a, score, topk, renormalize=False
    )

+    ab_strides1 = torch.full((num_experts,), k, device="cuda", dtype=torch.int64)
+    ab_strides2 = torch.full((num_experts,), n, device="cuda", dtype=torch.int64)
+    c_strides1 = torch.full((num_experts,), 2 * n, device="cuda", dtype=torch.int64)
+    c_strides2 = torch.full((num_experts,), k, device="cuda", dtype=torch.int64)
+
    def run_triton_moe(
        a: torch.Tensor,
        w1: torch.Tensor,
@ -111,6 +116,10 @@ def bench_run(
        w2: torch.Tensor,
        w1_scale: torch.Tensor,
        w2_scale: torch.Tensor,
+        ab_strides1: torch.Tensor,
+        ab_strides2: torch.Tensor,
+        c_strides1: torch.Tensor,
+        c_strides2: torch.Tensor,
        topk_weights: torch.Tensor,
        topk_ids: torch.Tensor,
        per_act_token: bool,
@ -125,6 +134,10 @@ def bench_run(
                topk_ids,
                w1_scale,
                w2_scale,
+                ab_strides1,
+                ab_strides2,
+                c_strides1,
+                c_strides2,
                per_act_token,
                a1_scale=None,
            )
@ -136,6 +149,10 @@ def bench_run(
        w2_q: torch.Tensor,
        w1_scale: torch.Tensor,
        w2_scale: torch.Tensor,
+        ab_strides1: torch.Tensor,
+        ab_strides2: torch.Tensor,
+        c_strides1: torch.Tensor,
+        c_strides2: torch.Tensor,
        topk_weights: torch.Tensor,
        topk_ids: torch.Tensor,
    ):
@ -150,6 +167,10 @@ def bench_run(
                topk_ids,
                w1_scale,
                w2_scale,
+                ab_strides1,
+                ab_strides2,
+                c_strides1,
+                c_strides2,
                per_act_token,
                a1_scale=None,
            )
@ -194,6 +215,10 @@ def bench_run(
            w2_q,
            w1_scale,
            w2_scale,
+            ab_strides1,
+            ab_strides2,
+            c_strides1,
+            c_strides2,
            topk_weights,
            topk_ids,
        )
@ -231,6 +256,10 @@ def bench_run(
        "w1_scale": w1_scale,
        "w2_scale": w2_scale,
        "per_act_token": per_act_token,
+        "ab_strides1": ab_strides1,
+        "ab_strides2": ab_strides2,
+        "c_strides1": c_strides1,
+        "c_strides2": c_strides2,
        # cuda graph params
        "cutlass_graph": cutlass_graph,
        "triton_graph": triton_graph,
@ -289,6 +318,10 @@ def bench_run(
        w2_q,
        w1_scale,
        w2_scale,
+        ab_strides1,
+        ab_strides2,
+        c_strides1,
+        c_strides2,
        topk_weights,
        topk_ids,
        per_act_token,
@ -297,7 +330,7 @@ def bench_run(

    results.append(
        benchmark.Timer(
-            stmt="run_cutlass_moe(a, a_scale, w1_q, w2_q, w1_scale, w2_scale, topk_weights, topk_ids, per_act_token, num_runs)",  # noqa: E501
+            stmt="run_cutlass_moe(a, a_scale, w1_q, w2_q, w1_scale, w2_scale, ab_strides1, ab_strides2, c_strides1, c_strides2, topk_weights, topk_ids, per_act_token, num_runs)",  # noqa: E501
            globals=globals,
            label=label,
            sub_label=sub_label,
--- a/benchmarks/kernels/benchmark_machete.py
+++ b/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,
@ -252,28 +253,7 @@ def marlin_create_bench_fn(bt: BenchmarkTensors) -> Callable:
    else:
        assert bt.a.dtype == torch.int8
        assert bt.wtype == scalar_types.uint4b8
-
-        if bt.w_ch_s is not None:
-            s_ch = bt.w_ch_s.to(torch.float32)
-        else:
-            s_ch = torch.ones(bt.w_ref.shape[1], dtype=torch.float32, device=device)
-
-        if bt.w_tok_s is not None:
-            s_tok = bt.w_tok_s.to(torch.float32)
-        else:
-            s_tok = torch.ones(bt.a.shape[0], dtype=torch.float32, device=device)
-
-        fn = lambda: ops.marlin_qqq_gemm(
-            a=bt.a,
-            b_q_weight=w_q,
-            s_group=w_s,
-            s_tok=s_tok,
-            s_ch=s_ch,
-            workspace=workspace.scratch,
-            size_m=bt.a.shape[0],
-            size_n=bt.w_ref.shape[1],
-            size_k=bt.w_ref.shape[0],
-        )
+        raise NotImplementedError("QQQ is not supported anymore")

    return fn

@ -304,6 +284,25 @@ def machete_create_bench_fn(
    )


+def cutlass_w4a8_create_bench_fn(
+    bt: BenchmarkTensors, out_type=torch.dtype, schedule=None
+) -> Callable:
+    w_q = bt.w_q.t().contiguous().t()  # make col major
+    w_q = ops.cutlass_encode_and_reorder_int4b(w_q)
+    # expects fp8 scales
+    w_s = ops.cutlass_pack_scale_fp8(bt.w_g_s.to(torch.float8_e4m3fn))
+
+    return lambda: ops.cutlass_w4a8_mm(
+        a=bt.a,
+        b_q=w_q,
+        b_group_scales=w_s,
+        b_group_size=bt.group_size,
+        b_channel_scales=bt.w_ch_s,
+        a_token_scales=bt.w_tok_s,
+        maybe_schedule=schedule,
+    )
+
+
 # impl

 # bench
@ -405,6 +404,20 @@ def bench(
        )
    )

+    # cutlass w4a8
+    if types.act_type == torch.float8_e4m3fn and group_size == 128:
+        timers.append(
+            bench_fns(
+                label,
+                sub_label,
+                f"cutlass w4a8 ({name_type_string})",
+                [
+                    cutlass_w4a8_create_bench_fn(bt, out_type=types.output_type)
+                    for bt in benchmark_tensors
+                ],
+            )
+        )
+
    if sweep_schedules:
        global _SWEEP_SCHEDULES_RESULTS

--- a/benchmarks/kernels/benchmark_moe.py
+++ b/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
@ -429,7 +430,6 @@ class BenchmarkWorker:
                hidden_size,
                topk,
                dtype_str,
-                is_marlin=False,
            )
        else:
            config = op_config[min(op_config.keys(), key=lambda x: abs(x - num_tokens))]
@ -542,6 +542,7 @@ 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
@ -552,7 +553,8 @@ 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)
@ -707,6 +709,7 @@ 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")
@ -748,6 +751,9 @@ if __name__ == "__main__":
        "--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")
--- a/benchmarks/kernels/benchmark_silu_mul_fp8_quant.py
+++ b/benchmarks/kernels/benchmark_silu_mul_fp8_quant.py
@ -0,0 +1,77 @@
+#!/usr/bin/env python3
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+import time
+
+import torch
+
+from vllm.model_executor.layers.fused_moe.batched_deep_gemm_moe import (
+    silu_mul_fp8_quant_deep_gemm,
+)
+from vllm.platforms import current_platform
+
+
+def benchmark(E, T, H, G=128, runs=50):
+    current_platform.seed_everything(42)
+    y = torch.randn((E, T, 2 * H), dtype=torch.bfloat16, device="cuda")
+    tokens_per_expert = torch.randint(
+        T // 2, T, size=(E,), dtype=torch.int32, device="cuda"
+    )
+
+    # Warmup
+    for _ in range(10):
+        silu_mul_fp8_quant_deep_gemm(y, tokens_per_expert, group_size=G)
+        torch.cuda.synchronize()
+
+    # Benchmark
+    torch.cuda.synchronize()
+    start = time.perf_counter()
+    for _ in range(runs):
+        silu_mul_fp8_quant_deep_gemm(y, tokens_per_expert, group_size=G)
+    torch.cuda.synchronize()
+
+    avg_time = (time.perf_counter() - start) / runs * 1000
+
+    # Calculate actual work done (only count valid tokens)
+    actual_tokens = tokens_per_expert.sum().item()
+    actual_elements = actual_tokens * H
+
+    # GFLOPS: operations per element = exp + 3 muls + 1 div + quantization ops ≈ 8 ops
+    ops_per_element = 8
+    total_ops = actual_elements * ops_per_element
+    gflops = total_ops / (avg_time / 1000) / 1e9
+
+    # Memory bandwidth: bfloat16 inputs (2 bytes), fp8 output (1 byte), scales (4 bytes)
+    input_bytes = actual_tokens * 2 * H * 2  # 2*H bfloat16 inputs
+    output_bytes = actual_tokens * H * 1  # H fp8 outputs
+    scale_bytes = actual_tokens * (H // G) * 4  # scales in float32
+    total_bytes = input_bytes + output_bytes + scale_bytes
+    memory_bw = total_bytes / (avg_time / 1000) / 1e9
+
+    return avg_time, gflops, memory_bw
+
+
+configs = [
+    (8, 32, 1024),
+    (16, 64, 2048),
+    (32, 128, 4096),
+    # DeepSeekV3 Configs
+    (256, 16, 7168),
+    (256, 32, 7168),
+    (256, 64, 7168),
+    (256, 128, 7168),
+    (256, 256, 7168),
+    (256, 512, 7168),
+    (256, 1024, 7168),
+]
+
+print(f"GPU: {torch.cuda.get_device_name()}")
+print(f"{'Config':<20} {'Time(ms)':<10} {'GFLOPS':<10} {'GB/s':<10}")
+print("-" * 50)
+
+for E, T, H in configs:
+    try:
+        time_ms, gflops, gbps = benchmark(E, T, H)
+        print(f"E={E:3d},T={T:4d},H={H:4d} {time_ms:8.3f} {gflops:8.1f} {gbps:8.1f}")
+    except Exception:
+        print(f"E={E:3d},T={T:4d},H={H:4d} FAILED")
--- a/benchmarks/kernels/benchmark_trtllm_decode_attention.py
+++ b/benchmarks/kernels/benchmark_trtllm_decode_attention.py
@ -3,16 +3,17 @@

 import csv
 import os
-import random
 from datetime import datetime
+from typing import Optional

 import flashinfer
 import torch

-FLOAT32_BYTES = torch.finfo(torch.float).bits // 8
+from vllm.utils import round_up

-# KV Cache Layout for TRT-LLM
-# kv_cache_shape = (num_blocks, 2, num_kv_heads, page_size, head_dim)
+FLOAT32_BYTES = torch.finfo(torch.float).bits // 8
+FP8_DTYPE = torch.float8_e4m3fn
+FP4_DTYPE = torch.uint8


 def to_float8(x, dtype=torch.float8_e4m3fn):
@ -26,65 +27,106 @@ def to_float8(x, dtype=torch.float8_e4m3fn):

@torch.no_grad()
 def benchmark_decode(
-    num_seqs,
-    max_seq_len,
-    page_size=16,
-    dtype=torch.bfloat16,
-    kv_layout="HND",
-    num_kv_heads=8,
-    kv_cache_dtype="auto",
-    head_dim=128,
-    warmup=10,
-    trials=20,
+    dtype: torch.dtype,
+    quant_dtypes: tuple[
+        Optional[torch.dtype], Optional[torch.dtype], Optional[torch.dtype]
+    ],
+    batch_size: int,
+    max_seq_len: int,
+    num_heads: tuple[int, int] = (64, 8),
+    head_size: int = 128,
+    kv_layout: str = "HND",
+    block_size: int = 16,
+    warmup: int = 10,
+    trials: int = 20,
 ):
    torch.set_default_device("cuda")
-    device = "cuda"
    torch.manual_seed(0)

-    HEAD_GRP_SIZE = 8
-    MAX_SEQ_LEN = max_seq_len
+    q_quant_dtype, kv_quant_dtype, o_quant_dtype = quant_dtypes
+    q_quant_dtype = q_quant_dtype or dtype
+    kv_quant_dtype = kv_quant_dtype or dtype
+    o_quant_dtype = o_quant_dtype or dtype
+
+    num_qo_heads, num_kv_heads = num_heads
+    assert num_qo_heads % num_kv_heads == 0
+
+    sm_scale = float(1.0 / (head_size**0.5))

    # large number to reduce kv_cache reuse
-    NUM_BLOCKS = int(256000 / page_size)
+    NUM_BLOCKS = int(256000 / block_size)

-    workspace_buffer = torch.empty(1024 * 1024 * 1024, dtype=torch.int8, device=device)
+    kv_cache_shape = None
+    if kv_layout == "NHD":
+        kv_cache_shape = (NUM_BLOCKS, 2, block_size, num_kv_heads, head_size)
+    elif kv_layout == "HND":
+        kv_cache_shape = (NUM_BLOCKS, 2, num_kv_heads, block_size, head_size)
+    else:
+        raise ValueError(f"Invalid kv_layout: {kv_layout}")

-    # For decode, batch_size is num_decode_token
-    num_qo_heads = num_kv_heads * HEAD_GRP_SIZE
-    sm_scale = float(1.0 / (head_dim**0.5))
-    q = torch.randn(num_seqs, num_qo_heads, head_dim, device=device, dtype=dtype)
-    kv_lens = [random.randint(1, MAX_SEQ_LEN) for _ in range(num_seqs)]
+    # Always using 1.0 scale to reflect the real perf in benchmarking
+    q_scale = 1.0
+    ref_query = torch.randn(batch_size, num_qo_heads, head_size, dtype=dtype)
+    if q_quant_dtype == FP8_DTYPE:
+        query, _ = to_float8(ref_query)
+    else:
+        query = ref_query

-    max_kv_len = max(kv_lens)
-    kv_lens_tensor = torch.tensor(kv_lens, dtype=torch.int, device=device)
-    max_num_blocks_per_seq = (max_kv_len + page_size - 1) // page_size
+    kv_lens = torch.randint(1, max_seq_len, (batch_size,), dtype=torch.int32)
+    kv_lens[-1] = max_seq_len

-    block_tables = torch.randint(
-        0, NUM_BLOCKS, (num_seqs, max_num_blocks_per_seq), dtype=torch.int32
-    )
+    seq_lens = kv_lens
+    max_seq_len = torch.max(seq_lens).item()

-    kv_cache_shape = (NUM_BLOCKS, 2, num_kv_heads, page_size, head_dim)
-    kv_cache = torch.randn(size=kv_cache_shape, device=device, dtype=dtype)
+    # Always using 1.0 scale to reflect the real perf in benchmarking
    k_scale = v_scale = 1.0
+    ref_kv_cache = torch.randn(kv_cache_shape, dtype=dtype)
+    if kv_quant_dtype == FP8_DTYPE:
+        kv_cache, _ = to_float8(ref_kv_cache)
+    else:
+        kv_cache = ref_kv_cache

-    if kv_cache_dtype.startswith("fp8"):
-        kv_cache, _ = to_float8(kv_cache)
+    max_num_blocks_per_seq = (max_seq_len + block_size - 1) // block_size
+    block_tables = torch.randint(
+        0, NUM_BLOCKS, (batch_size, max_num_blocks_per_seq), dtype=torch.int32
+    )
+    kv_indptr = [0]
+    kv_indices = []
+    kv_last_page_lens = []
+    for i in range(batch_size):
+        seq_len = seq_lens[i]
+        assert seq_len > 0
+        num_blocks = (seq_len + block_size - 1) // block_size
+        kv_indices.extend(block_tables[i, :num_blocks])
+        kv_indptr.append(kv_indptr[-1] + num_blocks)
+        kv_last_page_len = seq_len % block_size
+        if kv_last_page_len == 0:
+            kv_last_page_len = block_size
+        kv_last_page_lens.append(kv_last_page_len)

-    output_trtllm = torch.empty(q.shape, dtype=dtype)
+    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)
+    workspace_buffer = torch.zeros(1024 * 1024 * 1024, dtype=torch.int8)

-    # Benchmark TRT decode
-    def trt_decode():
-        return flashinfer.decode.trtllm_batch_decode_with_kv_cache(
-            q,
-            kv_cache,
-            workspace_buffer,
-            block_tables,
-            kv_lens_tensor,
-            max_kv_len,
-            bmm1_scale=k_scale * sm_scale,
-            bmm2_scale=v_scale,
-            out=output_trtllm,
-        )
+    wrapper = flashinfer.BatchDecodeWithPagedKVCacheWrapper(
+        workspace_buffer,
+        kv_layout,
+        use_tensor_cores=True,
+    )
+    wrapper.plan(
+        kv_indptr,
+        kv_indices,
+        kv_last_page_lens,
+        num_qo_heads,
+        num_kv_heads,
+        head_size,
+        block_size,
+        "NONE",
+        sm_scale=sm_scale,
+        q_data_type=dtype,
+        kv_data_type=dtype,
+    )

    def time_fn(fn, warmup=10, trials=20):
        torch.cuda.synchronize()
@ -101,74 +143,72 @@ def benchmark_decode(
            times.append(start.elapsed_time(end))  # ms
        return sum(times) / len(times), torch.std(torch.tensor(times))

-    # TRT Decode
-    trt_mean, trt_std = time_fn(trt_decode)
-
-    kv_indptr = [0]
-    kv_indices = []
-    kv_last_page_lens = []
-    for i in range(num_seqs):
-        seq_len = kv_lens[i]
-        assert seq_len > 0
-        num_blocks = (seq_len + page_size - 1) // page_size
-        kv_indices.extend(block_tables[i, :num_blocks])
-        kv_indptr.append(kv_indptr[-1] + num_blocks)
-        kv_last_page_len = seq_len % page_size
-        if kv_last_page_len == 0:
-            kv_last_page_len = page_size
-        kv_last_page_lens.append(kv_last_page_len)
-
-    kv_indptr = torch.tensor(kv_indptr, dtype=torch.int32)
-    kv_indices = torch.tensor(kv_indices, dtype=torch.int32)
-    kv_last_page_lens = torch.tensor(kv_last_page_lens, dtype=torch.int32)
-
-    output_baseline = torch.empty(q.shape, dtype=dtype)
-
-    wrapper = flashinfer.BatchDecodeWithPagedKVCacheWrapper(
-        workspace_buffer,
-        kv_layout,
-        use_tensor_cores=((num_qo_heads // num_kv_heads) > 4),
-    )
-
-    wrapper.plan(
-        kv_indptr,
-        kv_indices,
-        kv_last_page_lens,
-        num_qo_heads,
-        num_kv_heads,
-        head_dim,
-        page_size,
-        "NONE",
-        q_data_type=dtype,
-        kv_data_type=torch.float8_e4m3fn if kv_cache_dtype.startswith("fp8") else dtype,
-    )
+    o_scale = 1.0
+    o_sf_scale = None
+    output_baseline = torch.empty(ref_query.shape, dtype=dtype)
+    if o_quant_dtype == FP4_DTYPE:
+        o_sf_scale = 500.0
+        output_trtllm = flashinfer.utils.FP4Tensor(
+            torch.empty(query.shape[:-1] + (query.shape[-1] // 2,), dtype=torch.uint8),
+            torch.empty(
+                (
+                    round_up(query.shape[0], 128),
+                    round_up(query.shape[1] * query.shape[2] // 16, 4),
+                ),
+                dtype=torch.float8_e4m3fn,
+            ),
+        )
+    else:
+        output_trtllm = torch.empty(query.shape, dtype=o_quant_dtype)

    def baseline_decode():
-        return wrapper.run(q, kv_cache, sm_scale, k_scale, v_scale, output_baseline)
+        return wrapper.run(
+            ref_query,
+            ref_kv_cache,
+            k_scale=k_scale,
+            v_scale=v_scale,
+            out=output_baseline,
+        )
+
+    def trtllm_decode():
+        return flashinfer.decode.trtllm_batch_decode_with_kv_cache(
+            query=query,
+            kv_cache=kv_cache,
+            workspace_buffer=workspace_buffer,
+            block_tables=block_tables,
+            seq_lens=seq_lens,
+            max_seq_len=max_seq_len,
+            bmm1_scale=q_scale * k_scale * sm_scale,
+            bmm2_scale=v_scale / o_scale,
+            o_sf_scale=o_sf_scale,
+            out=output_trtllm,
+        )

    baseline_mean, baseline_std = time_fn(baseline_decode)
+    trtllm_mean, trtllm_std = time_fn(trtllm_decode)

    # Calculate percentage speedup (positive means TRT is faster)
-    speedup_percent = (baseline_mean - trt_mean) / baseline_mean
+    speedup_percent = (baseline_mean - trtllm_mean) / baseline_mean

    print(
-        f"\t{num_seqs}\t{max_seq_len}\t{trt_mean:.3f}\t{trt_std.item():.3f}"
+        f"\t{batch_size}\t{max_seq_len}\t{trtllm_mean:.3f}\t{trtllm_std.item():.3f}"
        f"\t{baseline_mean:.3f}\t{baseline_std.item():.3f}\t{speedup_percent:.3f}"
    )

    # Return results for CSV writing
    return {
-        "num_seqs": num_seqs,
-        "trt_mean": trt_mean,
-        "trt_std": trt_std.item(),
+        "batch_size": batch_size,
+        "trtllm_mean": trtllm_mean,
+        "trtllm_std": trtllm_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,
+        "q_dtype": str(q_quant_dtype),
+        "kv_cache_dtype": str(kv_quant_dtype),
+        "output_dtype": str(o_quant_dtype),
+        "block_size": block_size,
        "num_kv_heads": num_kv_heads,
-        "head_dim": head_dim,
+        "head_size": head_size,
        "max_seq_len": max_seq_len,
    }

@ -180,17 +220,18 @@ def write_results_to_csv(results, filename=None):
        filename = f"flashinfer_trtllm_benchmark_{timestamp}.csv"

    fieldnames = [
-        "num_seqs",
-        "trt_mean",
-        "trt_std",
+        "batch_size",
+        "trtllm_mean",
+        "trtllm_std",
        "baseline_mean",
        "baseline_std",
        "speedup_percent",
        "q_dtype",
        "kv_cache_dtype",
-        "page_size",
+        "output_dtype",
+        "block_size",
        "num_kv_heads",
-        "head_dim",
+        "head_size",
        "max_seq_len",
    ]

@ -209,45 +250,43 @@ def write_results_to_csv(results, filename=None):


 if __name__ == "__main__":
-    num_seqs = [1, 4, 8, 16, 32, 64, 128, 256]
+    batch_sizes = [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_decode(
-                bs,
-                max_seq_len,
-                dtype=torch.bfloat16,
-                kv_cache_dtype="auto",
-            )
-            all_results.append(result)
+    dtype = torch.bfloat16
+    quant_dtypes = [
+        # (q_quant_dtype, kv_quant_dtype, o_quant_dtype)
+        (None, None, None),
+        (None, FP8_DTYPE, None),
+        (FP8_DTYPE, FP8_DTYPE, FP8_DTYPE),
+        (FP8_DTYPE, FP8_DTYPE, FP4_DTYPE),
+    ]

-    print(
-        "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",
-            )
-            all_results.append(result)
+    for quant_dtype in quant_dtypes:
+        q_quant_dtype, kv_quant_dtype, o_quant_dtype = quant_dtype
+        q_quant_dtype = q_quant_dtype or dtype
+        kv_quant_dtype = kv_quant_dtype or dtype
+        o_quant_dtype = o_quant_dtype or dtype
+
+        print(
+            f"Running benchmark for q_dtype = {q_quant_dtype}, "
+            f"kv_cache_dtype: {kv_quant_dtype}, "
+            f"output_dtype: {o_quant_dtype}"
+        )
+        print(
+            "\tbatch_size\tmax_seq_len\ttrtllm_mean\ttrtllm_std\tbaseline_mean\t"
+            "baseline_std\tspeedup_percent"
+        )
+        for max_seq_len in max_seq_lens:
+            for bs in batch_sizes:
+                result = benchmark_decode(
+                    dtype=dtype,
+                    quant_dtypes=quant_dtype,
+                    batch_size=bs,
+                    max_seq_len=max_seq_len,
+                )
+                all_results.append(result)

    # Write all results to CSV
    write_results_to_csv(all_results)
--- a/benchmarks/kernels/benchmark_trtllm_prefill_attention.py
+++ b/benchmarks/kernels/benchmark_trtllm_prefill_attention.py
@ -3,16 +3,17 @@

 import csv
 import os
-import random
 from datetime import datetime
+from typing import Optional

 import flashinfer
 import torch

-FLOAT32_BYTES = torch.finfo(torch.float).bits // 8
+from vllm.utils import round_up

-# KV Cache Layout for TRT-LLM
-# kv_cache_shape = (num_blocks, 2, num_kv_heads, page_size, head_dim)
+FLOAT32_BYTES = torch.finfo(torch.float).bits // 8
+FP8_DTYPE = torch.float8_e4m3fn
+FP4_DTYPE = torch.uint8


 def to_float8(x, dtype=torch.float8_e4m3fn):
@ -26,84 +27,100 @@ def to_float8(x, dtype=torch.float8_e4m3fn):

@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,
+    dtype: torch.dtype,
+    quant_dtypes: tuple[
+        Optional[torch.dtype], Optional[torch.dtype], Optional[torch.dtype]
+    ],
+    batch_size: int,
+    max_seq_len: int,
+    num_heads: tuple[int, int] = (64, 8),
+    head_size: int = 128,
+    kv_layout: str = "HND",
+    block_size: int = 16,
+    warmup: int = 10,
+    trials: int = 20,
 ):
    torch.set_default_device("cuda")
    torch.manual_seed(0)

-    HEAD_GRP_SIZE = 8
-    MAX_SEQ_LEN = max_seq_len
+    q_quant_dtype, kv_quant_dtype, o_quant_dtype = quant_dtypes
+    q_quant_dtype = q_quant_dtype or dtype
+    kv_quant_dtype = kv_quant_dtype or dtype
+    o_quant_dtype = o_quant_dtype or dtype
+
+    max_q_len = max_kv_len = max_seq_len
+
+    num_qo_heads, num_kv_heads = num_heads
+    assert num_qo_heads % num_kv_heads == 0
+
+    sm_scale = float(1.0 / (head_size**0.5))

    # large number to reduce kv_cache reuse
-    NUM_BLOCKS = int(256000 / page_size)
+    NUM_BLOCKS = int(256000 / block_size)

-    workspace_buffer = torch.empty(1024 * 1024 * 1024, dtype=torch.int8)
+    kv_cache_shape = None
+    if kv_layout == "NHD":
+        kv_cache_shape = (NUM_BLOCKS, 2, block_size, num_kv_heads, head_size)
+    elif kv_layout == "HND":
+        kv_cache_shape = (NUM_BLOCKS, 2, num_kv_heads, block_size, head_size)
+    else:
+        raise ValueError(f"Invalid kv_layout: {kv_layout}")

-    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_lens = torch.randint(1, max_q_len, (batch_size,), dtype=torch.int32)
+    q_lens[-1] = max_q_len
    q_indptr = torch.cat(
        [
            torch.tensor([0], dtype=torch.int32),
-            torch.cumsum(
-                torch.tensor(q_lens, dtype=torch.int32), dim=0, dtype=torch.int32
-            ),
+            torch.cumsum(q_lens, 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
+    # Always using 1.0 scale to reflect the real perf in benchmarking
+    q_scale = 1.0
+    ref_query = torch.randn(
+        torch.sum(q_lens).item(), num_qo_heads, head_size, dtype=dtype
    )
+    if q_quant_dtype == FP8_DTYPE:
+        query, _ = to_float8(ref_query)
+    else:
+        query = ref_query

-    kv_cache_shape = (NUM_BLOCKS, 2, num_kv_heads, page_size, head_dim)
-    kv_cache = torch.randn(size=kv_cache_shape, dtype=dtype)
+    kv_lens = torch.randint(0, max_kv_len, (batch_size,), dtype=torch.int32)
+    kv_lens[-1] = max_kv_len
+
+    seq_lens = kv_lens + q_lens
+    max_seq_len = torch.max(seq_lens).item()
+
+    # Always using 1.0 scale to reflect the real perf in benchmarking
    k_scale = v_scale = 1.0
+    ref_kv_cache = torch.randn(kv_cache_shape, dtype=dtype)
+    if kv_quant_dtype == FP8_DTYPE:
+        kv_cache, _ = to_float8(ref_kv_cache)
+    else:
+        kv_cache = ref_kv_cache

-    if kv_cache_dtype.startswith("fp8"):
-        kv_cache, _ = to_float8(kv_cache)
-
-    output_trtllm = torch.empty(q.shape, dtype=dtype)
-
+    max_num_blocks_per_seq = (max_seq_len + block_size - 1) // block_size
+    block_tables = torch.randint(
+        0, NUM_BLOCKS, (batch_size, max_num_blocks_per_seq), dtype=torch.int32
+    )
    kv_indptr = [0]
    kv_indices = []
    kv_last_page_lens = []
-    for i in range(num_seqs):
+    for i in range(batch_size):
        seq_len = seq_lens[i]
        assert seq_len > 0
-        num_blocks = (seq_len + page_size - 1) // page_size
+        num_blocks = (seq_len + block_size - 1) // block_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
+        kv_last_page_len = seq_len % block_size
        if kv_last_page_len == 0:
-            kv_last_page_len = page_size
+            kv_last_page_len = block_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)
+    workspace_buffer = torch.zeros(1024 * 1024 * 1024, dtype=torch.int8)

    wrapper = flashinfer.BatchPrefillWithPagedKVCacheWrapper(
        workspace_buffer, kv_layout
@ -115,12 +132,12 @@ def benchmark_prefill(
        kv_last_page_lens,
        num_qo_heads,
        num_kv_heads,
-        head_dim,
-        page_size,
+        head_size,
+        block_size,
        causal=True,
        sm_scale=sm_scale,
        q_data_type=dtype,
-        kv_data_type=kv_cache.dtype,
+        kv_data_type=dtype,
    )

    def time_fn(fn, warmup=10, trials=20):
@ -138,52 +155,76 @@ def benchmark_prefill(
            times.append(start.elapsed_time(end))  # ms
        return sum(times) / len(times), torch.std(torch.tensor(times))

+    o_scale = 1.0
+    o_sf_scale = None
+    output_baseline = torch.empty(ref_query.shape, dtype=dtype)
+    if o_quant_dtype == FP4_DTYPE:
+        o_sf_scale = 500.0
+        output_trtllm = flashinfer.utils.FP4Tensor(
+            torch.empty(query.shape[:-1] + (query.shape[-1] // 2,), dtype=torch.uint8),
+            torch.empty(
+                (
+                    round_up(query.shape[0], 128),
+                    round_up(query.shape[1] * query.shape[2] // 16, 4),
+                ),
+                dtype=torch.float8_e4m3fn,
+            ),
+        )
+    else:
+        output_trtllm = torch.empty(query.shape, dtype=o_quant_dtype)
+
    def baseline_prefill():
        return wrapper.run(
-            q, kv_cache, k_scale=k_scale, v_scale=v_scale, out=output_baseline
+            ref_query,
+            ref_kv_cache,
+            k_scale=k_scale,
+            v_scale=v_scale,
+            out=output_baseline,
        )

-    def trt_prefill():
+    def trtllm_prefill():
        return flashinfer.prefill.trtllm_batch_context_with_kv_cache(
-            query=q,
+            query=query,
            kv_cache=kv_cache,
            workspace_buffer=workspace_buffer,
            block_tables=block_tables,
-            seq_lens=seq_lens_tensor,
+            seq_lens=seq_lens,
            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,
+            bmm1_scale=q_scale * k_scale * sm_scale,
+            bmm2_scale=v_scale / o_scale,
+            batch_size=batch_size,
            cum_seq_lens_q=q_indptr,
            cum_seq_lens_kv=kv_indptr,
+            o_sf_scale=o_sf_scale,
            out=output_trtllm,
        )

-    trt_mean, trt_std = time_fn(trt_prefill)
    baseline_mean, baseline_std = time_fn(baseline_prefill)
+    trtllm_mean, trtllm_std = time_fn(trtllm_prefill)

    # Calculate percentage speedup (positive means TRT is faster)
-    speedup_percent = (baseline_mean - trt_mean) / baseline_mean
+    speedup_percent = (baseline_mean - trtllm_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}"
+        f"\t{batch_size}\t{max_seq_len}\t{trtllm_mean:8.3f}\t{trtllm_std.item():8.3f}"
+        f"\t{baseline_mean:8.3f}\t{baseline_std.item():8.3f}\t{speedup_percent:8.3f}"
    )

    # Return results for CSV writing
    return {
-        "num_seqs": num_seqs,
-        "trt_mean": trt_mean,
-        "trt_std": trt_std.item(),
+        "batch_size": batch_size,
+        "trtllm_mean": trtllm_mean,
+        "trtllm_std": trtllm_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,
+        "q_dtype": str(q_quant_dtype),
+        "kv_cache_dtype": str(kv_quant_dtype),
+        "output_dtype": str(o_quant_dtype),
+        "block_size": block_size,
        "num_kv_heads": num_kv_heads,
-        "head_dim": head_dim,
+        "head_size": head_size,
        "max_seq_len": max_seq_len,
    }

@ -195,17 +236,18 @@ def write_results_to_csv(results, filename=None):
        filename = f"flashinfer_trtllm_benchmark_{timestamp}.csv"

    fieldnames = [
-        "num_seqs",
-        "trt_mean",
-        "trt_std",
+        "batch_size",
+        "trtllm_mean",
+        "trtllm_std",
        "baseline_mean",
        "baseline_std",
        "speedup_percent",
        "q_dtype",
        "kv_cache_dtype",
-        "page_size",
+        "output_dtype",
+        "block_size",
        "num_kv_heads",
-        "head_dim",
+        "head_size",
        "max_seq_len",
    ]

@ -224,27 +266,42 @@ def write_results_to_csv(results, filename=None):


 if __name__ == "__main__":
-    num_seqs = [1, 4, 8, 16, 32, 64, 128, 256]
+    batch_sizes = [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)
+    dtype = torch.bfloat16
+    quant_dtypes = [
+        # (q_quant_dtype, kv_quant_dtype, o_quant_dtype)
+        (None, None, None),
+        (FP8_DTYPE, FP8_DTYPE, FP8_DTYPE),
+        (FP8_DTYPE, FP8_DTYPE, FP4_DTYPE),
+    ]
+
+    for quant_dtype in quant_dtypes:
+        q_quant_dtype, kv_quant_dtype, o_quant_dtype = quant_dtype
+        q_quant_dtype = q_quant_dtype or dtype
+        kv_quant_dtype = kv_quant_dtype or dtype
+        o_quant_dtype = o_quant_dtype or dtype
+
+        print(
+            f"Running benchmark for q_dtype = {q_quant_dtype}, "
+            f"kv_cache_dtype: {kv_quant_dtype}, "
+            f"output_dtype: {o_quant_dtype}"
+        )
+        print(
+            "\tbatch_size\tmax_seq_len\ttrtllm_mean\ttrtllm_std\tbaseline_mean\t"
+            "baseline_std\tspeedup_percent"
+        )
+        for max_seq_len in max_seq_lens:
+            for bs in batch_sizes:
+                result = benchmark_prefill(
+                    dtype=dtype,
+                    quant_dtypes=quant_dtype,
+                    batch_size=bs,
+                    max_seq_len=max_seq_len,
+                )
+                all_results.append(result)

    # Write all results to CSV
    write_results_to_csv(all_results)
--- a/benchmarks/kernels/benchmark_w8a8_block_fp8.py
+++ b/benchmarks/kernels/benchmark_w8a8_block_fp8.py
@ -11,8 +11,8 @@ from datetime import datetime
 from typing import Any

 import torch
-import tqdm
 import triton
+from tqdm import tqdm

 from vllm.model_executor.layers.quantization.utils.fp8_utils import (
    _w8a8_block_fp8_matmul,
--- a/benchmarks/kernels/weight_shapes.py
+++ b/benchmarks/kernels/weight_shapes.py
@ -95,4 +95,10 @@ WEIGHT_SHAPES = {
        ([2048, 2816], 1),
        ([1408, 2048], 0),
    ],
+    "CohereLabs/c4ai-command-a-03-2025": [
+        ([12288, 14336], 1),
+        ([12288, 12288], 0),
+        ([12288, 73728], 1),
+        ([36864, 12288], 0),
+    ],
 }
--- a/benchmarks/multi_turn/README.md
+++ b/benchmarks/multi_turn/README.md
@ -5,11 +5,13 @@ The requirements (pip) for `benchmark_serving_multi_turn.py` can be found in `re
 First start serving your model

 ```bash
-export MODEL_NAME=/models/meta-llama/Meta-Llama-3.1-8B-Instruct/
+export MODEL_PATH=/models/meta-llama/Meta-Llama-3.1-8B-Instruct/

-vllm serve $MODEL_NAME --disable-log-requests
+vllm serve $MODEL_PATH --served-model-name Llama --disable-log-requests
 ```

+The variable `MODEL_PATH` should be a path to the model files (e.g. downloaded from huggingface).
+
 ## Synthetic Multi-Turn Conversations

 Download the following text file (used for generation of synthetic conversations)
@ -26,10 +28,10 @@ But you may use other text files if you prefer (using this specific file is not
 Then run the benchmarking script

 ```bash
-export MODEL_NAME=/models/meta-llama/Meta-Llama-3.1-8B-Instruct/
+export MODEL_PATH=/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
+python benchmark_serving_multi_turn.py --model $MODEL_PATH --served-model-name Llama \
+--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.).
--- a/benchmarks/multi_turn/benchmark_serving_multi_turn.py
+++ b/benchmarks/multi_turn/benchmark_serving_multi_turn.py
@ -825,9 +825,11 @@ def get_client_config(

    # Arguments for API requests
    chat_url = f"{args.url}/v1/chat/completions"
+    model_name = args.served_model_name if args.served_model_name else args.model
+
    req_args = RequestArgs(
        chat_url=chat_url,
-        model=args.model,
+        model=model_name,
        stream=not args.no_stream,
        limit_min_tokens=args.limit_min_tokens,
        limit_max_tokens=args.limit_max_tokens,
@ -1247,9 +1249,19 @@ async def main() -> None:
        default=0,
        help="Seed for random number generators (default: 0)",
    )
+
    parser.add_argument(
        "-m", "--model", type=str, required=True, help="Path of the LLM model"
    )
+    parser.add_argument(
+        "--served-model-name",
+        type=str,
+        default=None,
+        help="The model name used in the API. "
+        "If not specified, the model name will be the "
+        "same as the ``--model`` argument. ",
+    )
+
    parser.add_argument(
        "-u",
        "--url",
--- a/cmake/cpu_extension.cmake
+++ b/cmake/cpu_extension.cmake
@ -182,17 +182,17 @@ endif()
 #
 # Build oneDNN for W8A8 GEMM kernels (only for x86-AVX512 /ARM platforms)
 # Flag to enable ACL kernels for AARCH64 platforms
-if ( VLLM_BUILD_ACL STREQUAL "ON")
+if (VLLM_BUILD_ACL STREQUAL "ON")
    set(USE_ACL ON)
 else()
    set(USE_ACL OFF)
 endif()

-if ((AVX512_FOUND AND NOT AVX512_DISABLED) OR ASIMD_FOUND)
+if ((AVX512_FOUND AND NOT AVX512_DISABLED) OR ASIMD_FOUND OR POWER9_FOUND OR POWER10_FOUND OR POWER11_FOUND)
    FetchContent_Declare(
        oneDNN
        GIT_REPOSITORY https://github.com/oneapi-src/oneDNN.git
-        GIT_TAG  v3.8.1
+        GIT_TAG v3.9
        GIT_PROGRESS TRUE
        GIT_SHALLOW TRUE
    )
@ -204,7 +204,7 @@ if ((AVX512_FOUND AND NOT AVX512_DISABLED) OR ASIMD_FOUND)
        endif()
        set(ONEDNN_AARCH64_USE_ACL "ON")
        set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wl,-rpath,$ENV{ACL_ROOT_DIR}/build/")
-        endif()
+    endif()

    set(ONEDNN_LIBRARY_TYPE "STATIC")
    set(ONEDNN_BUILD_DOC "OFF")
@ -217,38 +217,23 @@ if ((AVX512_FOUND AND NOT AVX512_DISABLED) OR ASIMD_FOUND)
    set(ONEDNN_ENABLE_ITT_TASKS "OFF")
    set(ONEDNN_ENABLE_MAX_CPU_ISA "OFF")
    set(ONEDNN_ENABLE_CPU_ISA_HINTS "OFF")
+    set(ONEDNN_VERBOSE "OFF")
    set(CMAKE_POLICY_DEFAULT_CMP0077 NEW)

    FetchContent_MakeAvailable(oneDNN)
-    
-    list(APPEND LIBS dnnl)
-elseif(POWER10_FOUND)
-    FetchContent_Declare(
-        oneDNN
-        GIT_REPOSITORY https://github.com/oneapi-src/oneDNN.git
-        GIT_TAG v3.7.2
-        GIT_PROGRESS TRUE
-        GIT_SHALLOW TRUE
+    add_library(dnnl_ext OBJECT "csrc/cpu/dnnl_helper.cpp")
+    target_include_directories(
+        dnnl_ext
+        PUBLIC ${oneDNN_SOURCE_DIR}/include
+        PUBLIC ${oneDNN_BINARY_DIR}/include
+        PRIVATE ${oneDNN_SOURCE_DIR}/src
    )
-
-    set(ONEDNN_LIBRARY_TYPE "STATIC")
-    set(ONEDNN_BUILD_DOC "OFF")
-    set(ONEDNN_BUILD_EXAMPLES "OFF")
-    set(ONEDNN_BUILD_TESTS "OFF")
-    set(ONEDNN_ENABLE_WORKLOAD "INFERENCE")
-    set(ONEDNN_ENABLE_PRIMITIVE "MATMUL;REORDER")
-    set(ONEDNN_BUILD_GRAPH "OFF")
-    set(ONEDNN_ENABLE_JIT_PROFILING "OFF")
-    set(ONEDNN_ENABLE_ITT_TASKS "OFF")
-    set(ONEDNN_ENABLE_MAX_CPU_ISA "OFF")
-    set(ONEDNN_ENABLE_CPU_ISA_HINTS "OFF")
-    set(CMAKE_POLICY_DEFAULT_CMP0077 NEW)
-
-    set(DNNL_CPU_RUNTIME "OMP")
-
-    FetchContent_MakeAvailable(oneDNN)
-
-    list(APPEND LIBS dnnl)
+    target_link_libraries(dnnl_ext dnnl)
+    target_compile_options(dnnl_ext PRIVATE ${CXX_COMPILE_FLAGS} -fPIC)
+    list(APPEND LIBS dnnl_ext)
+    set(USE_ONEDNN ON)
+else()
+    set(USE_ONEDNN OFF)
 endif()

 message(STATUS "CPU extension compile flags: ${CXX_COMPILE_FLAGS}")
@ -275,7 +260,6 @@ set(VLLM_EXT_SRC

 if (AVX512_FOUND AND NOT AVX512_DISABLED)
    set(VLLM_EXT_SRC
-        "csrc/cpu/quant.cpp"
        "csrc/cpu/shm.cpp"
        ${VLLM_EXT_SRC})
    if (ENABLE_AVX512BF16 AND ENABLE_AVX512VNNI)
@ -289,14 +273,11 @@ if (AVX512_FOUND AND NOT AVX512_DISABLED)
            ${VLLM_EXT_SRC})
        add_compile_definitions(-DCPU_CAPABILITY_AVX512)
    endif()
-elseif(POWER10_FOUND)
-    set(VLLM_EXT_SRC
-        "csrc/cpu/quant.cpp"
-        ${VLLM_EXT_SRC})
 endif()
-if (ASIMD_FOUND)
+
+if(USE_ONEDNN)
    set(VLLM_EXT_SRC
-        "csrc/cpu/quant.cpp"
+        "csrc/cpu/dnnl_kernels.cpp"
        ${VLLM_EXT_SRC})
 endif()

--- a/cmake/external_projects/flashmla.cmake
+++ b/cmake/external_projects/flashmla.cmake
@ -19,7 +19,7 @@ else()
  FetchContent_Declare(
        flashmla
        GIT_REPOSITORY https://github.com/vllm-project/FlashMLA.git
-        GIT_TAG 0e43e774597682284358ff2c54530757b654b8d1
+        GIT_TAG a757314c04eedd166e329e846c820eb1bdd702de
        GIT_PROGRESS TRUE
        CONFIGURE_COMMAND ""
        BUILD_COMMAND ""
@ -37,13 +37,14 @@ 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/kernels/splitkv_mla.cu
+        ${flashmla_SOURCE_DIR}/csrc/kernels/get_mla_metadata.cu
        ${flashmla_SOURCE_DIR}/csrc/kernels/mla_combine.cu
-        ${flashmla_SOURCE_DIR}/csrc/kernels/get_mla_metadata.cu)
+        ${flashmla_SOURCE_DIR}/csrc/kernels/splitkv_mla.cu
+        ${flashmla_SOURCE_DIR}/csrc/kernels_fp8/flash_fwd_mla_fp8_sm90.cu)

    set(FlashMLA_INCLUDES
        ${flashmla_SOURCE_DIR}/csrc/cutlass/include
-        ${flashmla_SOURCE_DIR}/csrc/include)
+        ${flashmla_SOURCE_DIR}/csrc)

    set_gencode_flags_for_srcs(
        SRCS "${FlashMLA_SOURCES}"
--- a/cmake/external_projects/vllm_flash_attn.cmake
+++ b/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 93cf5a08f421a3efd0c4a7e005ef8f742b578ce0
+          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
--- a/csrc/activation_kernels.cu
+++ b/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.
--- a/csrc/attention/mla/sm100_cutlass_mla_kernel.cu
+++ b/csrc/attention/mla/sm100_cutlass_mla_kernel.cu
@ -167,7 +167,7 @@ typename T::Fmha::Arguments args_from_options(
      // TODO(trevor-m): Change split_kv back to -1 when
      // https://github.com/NVIDIA/cutlass/issues/2274 is fixed. Split_kv=1 will
      // perform worse with larger context length and smaller batch sizes.
-      num_kv_splits, // split_kv
+      static_cast<int>(num_kv_splits), // split_kv
      nullptr,       // is_var_split_kv
  };
  // TODO(kaixih@nvidia): When split_kv=-1 and is_var_split_kv=false, we compute
@ -264,7 +264,7 @@ int64_t sm100_cutlass_mla_get_workspace_size(int64_t max_seq_len, int64_t num_ba
  // Assumes device 0 when getting sm_count.
  arguments.hw_info.sm_count =
      sm_count <= 0 ? cutlass::KernelHardwareInfo::query_device_multiprocessor_count(/*device_id=*/0) : sm_count;
-  arguments.split_kv = num_kv_splits;
+  arguments.split_kv = static_cast<int>(num_kv_splits);
  MlaSm100Type::Fmha::set_split_kv(arguments);

  return MlaSm100Type::Fmha::get_workspace_size(arguments);
--- a/csrc/cache.h
+++ b/csrc/cache.h
@ -40,9 +40,11 @@ void concat_and_cache_mla(torch::Tensor& kv_c, torch::Tensor& k_pe,
 void convert_fp8(torch::Tensor& dst_cache, torch::Tensor& src_cache,
                 const double scale, const std::string& kv_cache_dtype);

-void gather_cache(
+void gather_and_maybe_dequant_cache(
    torch::Tensor const& src_cache,    // [NUM_BLOCKS, BLOCK_SIZE, ENTRIES...]
    torch::Tensor const& dst,          // [TOT_TOKENS, ENTRIES...]
    torch::Tensor const& block_table,  // [BATCH, BLOCK_INDICES]
    torch::Tensor const& cu_seq_lens,  // [BATCH+1]
-    int64_t batch_size, std::optional<torch::Tensor> seq_starts = std::nullopt);
+    int64_t batch_size, const std::string& kv_cache_dtype,
+    torch::Tensor const& scale,
+    std::optional<torch::Tensor> seq_starts = std::nullopt);
--- a/csrc/cache_kernels.cu
+++ b/csrc/cache_kernels.cu
@ -624,9 +624,9 @@ void convert_fp8(torch::Tensor& dst_cache, torch::Tensor& src_cache,
 namespace vllm {

 // grid is launched with dimensions (batch, num_splits)
-template <typename scalar_t>
-__global__ void gather_cache(
-    const scalar_t* __restrict__ src_cache,   // [NUM_BLOCKS, BLOCK_SIZE,
+template <typename scalar_t, typename cache_t, Fp8KVCacheDataType kv_dt>
+__global__ void gather_and_maybe_dequant_cache(
+    const cache_t* __restrict__ src_cache,    // [NUM_BLOCKS, BLOCK_SIZE,
                                              // ENTRIES...]
    scalar_t* __restrict__ dst,               // [TOT_TOKENS, ENTRIES...]
    const int32_t* __restrict__ block_table,  // [BATCH, BLOCK_INDICES]
@ -634,6 +634,7 @@ __global__ void gather_cache(
    const int32_t block_size, const int32_t entry_size,
    const int64_t block_table_stride, const int64_t cache_block_stride,
    const int64_t cache_entry_stride, const int64_t dst_entry_stride,
+    const float* __restrict__ scale,
    const int32_t* __restrict__ seq_starts) {  // Optional: starting offsets per
                                               // batch

@ -675,10 +676,16 @@ __global__ void gather_cache(
    if (partial_block_size) full_blocks_end -= 1;
  }

-  auto copy_entry = [&](const scalar_t* __restrict__ _src,
+  auto copy_entry = [&](const cache_t* __restrict__ _src,
                        scalar_t* __restrict__ _dst) {
-    for (int i = threadIdx.x; i < entry_size; i += blockDim.x)
-      _dst[i] = _src[i];
+    for (int i = threadIdx.x; i < entry_size; i += blockDim.x) {
+      if constexpr (kv_dt == Fp8KVCacheDataType::kAuto) {
+        _dst[i] = static_cast<scalar_t>(_src[i]);
+      } else {
+        _dst[i] =
+            fp8::scaled_convert<scalar_t, cache_t, kv_dt>(_src[i], *scale);
+      }
+    }
  };

  for (int pid = split_start; pid < full_blocks_end; ++pid) {
@ -705,25 +712,31 @@ __global__ void gather_cache(
 }  // namespace vllm

 // Macro to dispatch the kernel based on the data type.
-#define CALL_GATHER_CACHE(CPY_DTYPE)                                    \
-  vllm::gather_cache<CPY_DTYPE><<<grid, block, 0, stream>>>(            \
-      reinterpret_cast<CPY_DTYPE*>(src_cache.data_ptr()),               \
-      reinterpret_cast<CPY_DTYPE*>(dst.data_ptr()),                     \
-      block_table.data_ptr<int32_t>(), cu_seq_lens.data_ptr<int32_t>(), \
-      block_size, entry_size, block_table_stride, cache_block_stride,   \
-      cache_entry_stride, dst_entry_stride, seq_starts_ptr);
+// SCALAR_T is the data type of the destination tensor.
+// CACHE_T is the stored data type of kv-cache.
+// KV_DTYPE is the real data type of kv-cache.
+#define CALL_GATHER_CACHE(SCALAR_T, CACHE_T, KV_DTYPE)                      \
+  vllm::gather_and_maybe_dequant_cache<SCALAR_T, CACHE_T, KV_DTYPE>         \
+      <<<grid, block, 0, stream>>>(                                         \
+          reinterpret_cast<CACHE_T*>(src_cache.data_ptr()),                 \
+          reinterpret_cast<SCALAR_T*>(dst.data_ptr()),                      \
+          block_table.data_ptr<int32_t>(), cu_seq_lens.data_ptr<int32_t>(), \
+          block_size, entry_size, block_table_stride, cache_block_stride,   \
+          cache_entry_stride, dst_entry_stride,                             \
+          reinterpret_cast<const float*>(scale.data_ptr()), seq_starts_ptr);

 // Gather sequences from the cache into the destination tensor.
 //  - cu_seq_lens contains the cumulative sequence lengths for each batch
 //  - block_table contains the cache block indices for each sequence
 //  - Optionally, seq_starts (if provided) offsets the starting block index by
 //  (seq_starts[bid] / page_size)
-void gather_cache(
+void gather_and_maybe_dequant_cache(
    torch::Tensor const& src_cache,    // [NUM_BLOCKS, BLOCK_SIZE, ENTRIES...]
    torch::Tensor const& dst,          // [TOT_TOKENS, ENTRIES...]
    torch::Tensor const& block_table,  // [BATCH, BLOCK_INDICES]
    torch::Tensor const& cu_seq_lens,  // [BATCH+1]
-    int64_t batch_size,
+    int64_t batch_size, const std::string& kv_cache_dtype,
+    torch::Tensor const& scale,
    std::optional<torch::Tensor> seq_starts = std::nullopt) {
  at::cuda::OptionalCUDAGuard device_guard(src_cache.device());
  const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
@ -761,20 +774,8 @@ void gather_cache(
  dim3 grid(batch_size, num_splits);
  dim3 block(1024);

-  TORCH_CHECK(src_cache.dtype() == dst.dtype(),
-              "src_cache and dst must have the same dtype");
-
-  const int dtype_bits = src_cache.element_size() * 8;
  const int32_t* seq_starts_ptr =
      seq_starts.has_value() ? seq_starts.value().data_ptr<int32_t>() : nullptr;

-  if (dtype_bits == 32) {
-    CALL_GATHER_CACHE(uint32_t);
-  } else if (dtype_bits == 16) {
-    CALL_GATHER_CACHE(uint16_t);
-  } else if (dtype_bits == 8) {
-    CALL_GATHER_CACHE(uint8_t);
-  } else {
-    TORCH_CHECK(false, "Unsupported data type width: ", dtype_bits);
-  }
+  DISPATCH_BY_KV_CACHE_DTYPE(dst.dtype(), kv_cache_dtype, CALL_GATHER_CACHE);
 }
--- a/csrc/core/scalar_type.hpp
+++ b/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);
--- a/csrc/cpu/cpu_types_x86.hpp
+++ b/csrc/cpu/cpu_types_x86.hpp
@ -89,7 +89,7 @@ struct FP16Vec16 : public Vec<FP16Vec16> {

  explicit FP16Vec16(const FP32Vec16&);

-  void save(void* ptr) const { *reinterpret_cast<__m256i*>(ptr) = reg; }
+  void save(void* ptr) const { _mm256_storeu_si256((__m256i*)ptr, reg); }

  void save(void* ptr, const int elem_num) const {
    constexpr uint32_t M = 0xFFFFFFFF;
@ -126,7 +126,7 @@ struct BF16Vec16 : public Vec<BF16Vec16> {

  explicit BF16Vec16(const FP32Vec16&);

-  void save(void* ptr) const { *reinterpret_cast<__m256i*>(ptr) = reg; }
+  void save(void* ptr) const { _mm256_storeu_si256((__m256i*)ptr, reg); }

  void save(void* ptr, const int elem_num) const {
    constexpr uint32_t M = 0xFFFFFFFF;
@ -180,8 +180,8 @@ struct BF16Vec32 : public Vec<BF16Vec32> {
            (__m128i)vec8_data.reg, 1)) {}

  void save(void* ptr) const {
-    *reinterpret_cast<__m256i*>(ptr) = reg_low;
-    *reinterpret_cast<__m256i*>((__m256i*)ptr + 1) = reg_high;
+    _mm256_storeu_si256((__m256i*)ptr, reg_low);
+    _mm256_storeu_si256((__m256i*)ptr + 1, reg_high);
  }
 };
 #endif
--- a/csrc/cpu/dnnl_helper.cpp
+++ b/csrc/cpu/dnnl_helper.cpp
@ -0,0 +1,346 @@
+#include <list>
+#include <optional>
+
+#include "common/memory_desc.hpp"
+#include "common/memory.hpp"
+
+#include "dnnl_helper.h"
+
+static dnnl::engine& default_engine() {
+  static dnnl::engine engine(dnnl::engine::kind::cpu, 0);
+  return engine;
+}
+
+static dnnl::stream& default_stream() {
+  static dnnl::stream stream(default_engine());
+  return stream;
+}
+
+void release_dnnl_matmul_handler(int64_t handler) {
+  DNNLMatMulPrimitiveHandler* ptr =
+      reinterpret_cast<DNNLMatMulPrimitiveHandler*>(handler);
+  delete ptr;
+}
+
+template <typename KT, typename VT>
+class DNNLPrimitiveCache {
+ public:
+  using cache_value_t = std::pair<KT, VT>;
+  using result_value_t = VT;
+  using container_t = std::list<cache_value_t>;
+  using value_iterator_t = typename container_t::iterator;
+  using map_t = std::unordered_map<KT, value_iterator_t>;
+  using creator_t = VT (*)();
+
+ public:
+  DNNLPrimitiveCache(size_t capacity)
+      : capacity_(capacity),
+        values_(),
+        key_to_value_(std::min(256lu, capacity)) {
+    assert(capacity > 0);
+  }
+
+  template <typename F>
+  result_value_t get_or_create(const KT& key, F&& creator) {
+    std::optional<value_iterator_t> value = get_value(key);
+    if (value.has_value()) {
+      return value.value()->second;
+    } else {
+      return add_value({key, creator()})->second;
+    }
+  }
+
+  size_t size() const { return values_.size(); }
+
+ private:
+  void dump_data() {
+    std::stringstream ss;
+    ss << "table_id: " << std::hex << reinterpret_cast<size_t>(this) << std::dec
+       << "\n";
+    ss << "container: [";
+    for (auto&& iter : values_) {
+      ss << "(" << iter.first << ", " << std::hex
+         << reinterpret_cast<size_t>(iter.second.get()) << "), " << std::dec;
+    }
+    ss << "]\n";
+
+    ss << "map: [";
+    for (auto&& iter : key_to_value_) {
+      ss << "(" << iter.first << ", " << iter.second->first << ", " << std::hex
+         << reinterpret_cast<size_t>(iter.second->second.get()) << std::dec
+         << "), ";
+    }
+    ss << "]\n";
+    std::printf("%s\n", ss.str().c_str());
+  }
+
+  value_iterator_t add_value(cache_value_t&& new_value) {
+    if (size() == capacity_) {
+      cache_value_t& last_item = values_.back();
+      key_to_value_.erase(last_item.first);
+      values_.pop_back();
+    }
+
+    auto& added_value_ = values_.emplace_front(std::move(new_value));
+    key_to_value_.emplace(added_value_.first, values_.begin());
+    return values_.begin();
+  }
+
+  std::optional<value_iterator_t> get_value(const KT& key) {
+    if (key_to_value_.size() > 0 && key == values_.begin()->first) {
+      return values_.begin();
+    }
+
+    auto value_map_iterator = key_to_value_.find(key);
+    if (value_map_iterator != key_to_value_.end()) {
+      values_.splice(values_.begin(), values_, value_map_iterator->second);
+      return value_map_iterator->second;
+    } else {
+      return {};
+    }
+  }
+
+ private:
+  const size_t capacity_;
+  container_t values_;
+  map_t key_to_value_;
+};
+
+DNNLMatMulPrimitiveHandler::DNNLMatMulPrimitiveHandler(
+    const Args& args, dnnl::memory::data_type b_type)
+    : b_n_size_(args.b_n_size),
+      b_n_stride_(args.b_n_stride),
+      b_k_size_(args.b_k_size),
+      b_k_stride_(args.b_k_stride),
+      b_type_(b_type),
+      c_type_(args.c_type),
+      runtime_memory_ptrs_(8),
+      primitive_cache_size_(args.primitive_cache_size) {
+  assert(primitive_cache_size_ > 0);
+}
+
+void DNNLMatMulPrimitiveHandler::prepack_weight(
+    void* original_b_ptr, dnnl::memory::desc b_target_mem_desc) {
+  dnnl::memory::desc original_b_md({b_k_size_, b_n_size_}, b_type_,
+                                   {b_k_stride_, b_n_stride_});
+  dnnl::memory original_weight(original_b_md, default_engine(), original_b_ptr);
+  dnnl::memory packed_weight(b_target_mem_desc, default_engine());
+  {
+    dnnl::reorder(original_weight, packed_weight)
+        .execute(default_stream(), original_weight, packed_weight);
+    default_stream().wait();
+  }
+  memory_cache_[DNNL_ARG_WEIGHTS] = packed_weight;
+  b_target_mem_desc_ = b_target_mem_desc;
+}
+
+void DNNLMatMulPrimitiveHandler::set_runtime_memory_ptr(
+    size_t index, dnnl_memory* memory_ptr) {
+  dnnl::impl::memory_storage_t* mem_storage_ptr = memory_ptr->memory_storage();
+  dnnl_memory_desc* mem_desc = const_cast<dnnl_memory_desc*>(memory_ptr->md());
+  runtime_memory_ptrs_[index] = {mem_storage_ptr, mem_desc};
+}
+
+std::pair<dnnl::impl::memory_storage_t*, dnnl_memory_desc*>
+DNNLMatMulPrimitiveHandler::get_runtime_memory_ptr(size_t index) {
+  return runtime_memory_ptrs_[index];
+}
+
+namespace std {
+template <>
+struct hash<W8A8MatMulPrimitiveHandler::ClassMatmulCacheKey> {
+  size_t operator()(
+      const W8A8MatMulPrimitiveHandler::ClassMatmulCacheKey& val) const {
+    return hash<dnnl_dim_t>()(val.b_n_size) ^ hash<dnnl_dim_t>()(val.b_k_size) ^
+           hash<int>()(static_cast<int>(val.a_qs)) ^
+           hash<int>()(static_cast<int>(val.b_qs)) ^ hash<bool>()(val.use_azp) ^
+           hash<int>()(static_cast<int>(val.c_type));
+  }
+};
+
+template <>
+struct hash<W8A8MatMulPrimitiveHandler::MSizeCacheKey> {
+  size_t operator()(
+      const W8A8MatMulPrimitiveHandler::MSizeCacheKey& val) const {
+    return hash<dnnl_dim_t>()(val.a_m_size) ^ hash<bool>()(val.use_bias) ^
+           hash<int>()(static_cast<int>(val.bias_type));
+  }
+};
+}  // namespace std
+
+bool operator==(const W8A8MatMulPrimitiveHandler::ClassMatmulCacheKey& l,
+                const W8A8MatMulPrimitiveHandler::ClassMatmulCacheKey& r) {
+  return l.b_n_size == r.b_n_size && l.b_k_size == r.b_k_size &&
+         l.a_qs == r.a_qs && l.b_qs == r.b_qs && l.use_azp == r.use_azp &&
+         l.c_type == r.c_type;
+}
+
+bool operator==(const W8A8MatMulPrimitiveHandler::MSizeCacheKey& l,
+                const W8A8MatMulPrimitiveHandler::MSizeCacheKey& r) {
+  return l.use_bias == r.use_bias && l.a_m_size == r.a_m_size &&
+         l.bias_type == r.bias_type;
+}
+
+static std::shared_ptr<W8A8MatMulPrimitiveHandler::MSizeCache>
+get_w8a8_class_primitive_cache(
+    const W8A8MatMulPrimitiveHandler::ClassMatmulCacheKey& key,
+    int64_t cache_size) {
+  static W8A8MatMulPrimitiveHandler::ClassMatmulCache cache(128);
+  assert(cache_size > 0);
+  return cache.get_or_create(key, [&]() {
+    return std::make_shared<W8A8MatMulPrimitiveHandler::MSizeCache>(cache_size);
+  });
+}
+
+W8A8MatMulPrimitiveHandler::W8A8MatMulPrimitiveHandler(const Args& args)
+    : DNNLMatMulPrimitiveHandler(
+          static_cast<const DNNLMatMulPrimitiveHandler::Args&>(args),
+          dnnl::memory::data_type::s8),
+      use_azp_(args.use_a_zero_point),
+      a_qs_(args.a_quantization_strategy),
+      b_qs_(args.b_quantization_strategy),
+      m_size_cache_(nullptr) {
+  assert(a_qs_ != QuantizationStrategy::PER_OUTPUT_CHANNEL);
+  assert(b_qs_ != QuantizationStrategy::PER_TOKEN);
+  if (a_qs_ == QuantizationStrategy::PER_TOKEN) {
+    assert(!use_azp_);
+  };
+  prepack_weight(args.b_ptr,
+                 create_primitive_desc(
+                     MSizeCacheKey{.a_m_size = DNNL_RUNTIME_DIM_VAL,
+                                   .use_bias = false,
+                                   .bias_type = dnnl::memory::data_type::undef},
+                     true)
+                     .weights_desc());
+  init_runtime_memory_cache(args);
+}
+
+void W8A8MatMulPrimitiveHandler::execute(ExecArgs& args) {
+  auto&& [a_storage, a_mem_desc] = get_runtime_memory_ptr(0);
+  auto&& [c_storage, c_mem_desc] = get_runtime_memory_ptr(1);
+  a_storage->set_data_handle((void*)args.a_ptr);
+  a_mem_desc->dims[0] = args.a_m_size;
+  c_storage->set_data_handle((void*)args.c_ptr);
+  c_mem_desc->dims[0] = args.a_m_size;
+
+  if (a_qs_ == QuantizationStrategy::PER_TENSOR) {
+    auto&& [a_scale_storage, a_scale_mem_desc] = get_runtime_memory_ptr(2);
+    a_scale_storage->set_data_handle((void*)args.a_scales_ptr);
+  }
+  if (use_azp_) {
+    auto&& [a_zero_point_storage, a_zero_point_mem_desc] =
+        get_runtime_memory_ptr(3);
+    a_zero_point_storage->set_data_handle((void*)args.a_zero_points_ptr);
+  }
+
+  if (args.use_bias) {
+    auto&& [bias_storage, bias_mem_desc] = get_runtime_memory_ptr(4);
+    bias_storage->set_data_handle((void*)args.bias_ptr);
+  }
+
+  dnnl::matmul matmul = get_matmul_cache(args);
+  matmul.execute(default_stream(), memory_cache_);
+  default_stream().wait();
+}
+
+dnnl::matmul W8A8MatMulPrimitiveHandler::get_matmul_cache(
+    const MSizeCacheKey& key) {
+  if (m_size_cache_.get() == nullptr) {
+    ClassMatmulCacheKey key = {.b_n_size = b_n_size_,
+                               .b_k_size = b_k_size_,
+                               .a_qs = a_qs_,
+                               .b_qs = b_qs_,
+                               .use_azp = use_azp_,
+                               .c_type = c_type_};
+    m_size_cache_ = get_w8a8_class_primitive_cache(key, primitive_cache_size_);
+  }
+
+  return m_size_cache_->get_or_create(key, [&]() {
+    dnnl::matmul::primitive_desc desc = this->create_primitive_desc(key, false);
+    return dnnl::matmul(desc);
+  });
+}
+
+void W8A8MatMulPrimitiveHandler::init_runtime_memory_cache(const Args& args) {
+  memory_cache_[DNNL_ARG_SRC] = dnnl::memory({{1, b_k_size_},
+                                              dnnl::memory::data_type::s8,
+                                              dnnl::memory::format_tag::ab},
+                                             default_engine(), nullptr);
+  set_runtime_memory_ptr(0, memory_cache_[DNNL_ARG_SRC].get());
+  memory_cache_[DNNL_ARG_DST] =
+      dnnl::memory({{1, b_n_size_}, c_type_, dnnl::memory::format_tag::ab},
+                   default_engine(), nullptr);
+  set_runtime_memory_ptr(1, memory_cache_[DNNL_ARG_DST].get());
+
+  // For PER_TOKEN, scales will be applied in outside epilogue
+  if (a_qs_ == QuantizationStrategy::PER_TENSOR) {
+    memory_cache_[DNNL_ARG_ATTR_SCALES | DNNL_ARG_SRC] = dnnl::memory(
+        {{1}, dnnl::memory::data_type::f32, {1}}, default_engine(), nullptr);
+    set_runtime_memory_ptr(
+        2, memory_cache_[DNNL_ARG_ATTR_SCALES | DNNL_ARG_SRC].get());
+    if (use_azp_) {
+      memory_cache_[DNNL_ARG_ATTR_ZERO_POINTS | DNNL_ARG_SRC] = dnnl::memory(
+          {{1}, dnnl::memory::data_type::s32, {1}}, default_engine(), nullptr);
+      set_runtime_memory_ptr(
+          3, memory_cache_[DNNL_ARG_ATTR_ZERO_POINTS | DNNL_ARG_SRC].get());
+    }
+  }
+
+  if (b_qs_ == QuantizationStrategy::PER_TENSOR) {
+    memory_cache_[DNNL_ARG_ATTR_SCALES | DNNL_ARG_WEIGHTS] =
+        dnnl::memory({{1}, dnnl::memory::data_type::f32, {1}}, default_engine(),
+                     (void*)args.b_scales_ptr);
+  } else if (b_qs_ == QuantizationStrategy::PER_OUTPUT_CHANNEL) {
+    memory_cache_[DNNL_ARG_ATTR_SCALES | DNNL_ARG_WEIGHTS] =
+        dnnl::memory({{b_n_size_}, dnnl::memory::data_type::f32, {1}},
+                     default_engine(), (void*)args.b_scales_ptr);
+  }
+
+  memory_cache_[DNNL_ARG_BIAS] =
+      dnnl::memory({{b_n_size_}, dnnl::memory::data_type::f32, {1}},
+                   default_engine(), nullptr);
+  set_runtime_memory_ptr(4, memory_cache_[DNNL_ARG_BIAS].get());
+}
+
+dnnl::matmul::primitive_desc W8A8MatMulPrimitiveHandler::create_primitive_desc(
+    const MSizeCacheKey& key, bool first_time) {
+  dnnl::memory::desc a_md({key.a_m_size, b_k_size_},
+                          dnnl::memory::data_type::s8,
+                          dnnl::memory::format_tag::ab);
+  dnnl::memory::desc b_md;
+  if (first_time) {
+    b_md =
+        dnnl::memory::desc({b_k_size_, b_n_size_}, dnnl::memory::data_type::s8,
+                           dnnl::memory::format_tag::any);
+  } else {
+    b_md = b_target_mem_desc_;
+  }
+  dnnl::memory::desc c_md({key.a_m_size, b_n_size_}, c_type_,
+                          dnnl::memory::format_tag::ab);
+
+  dnnl::primitive_attr attr;
+  // For PER_TOKEN, scales will be applied in outside epilogue
+  if (a_qs_ == QuantizationStrategy::PER_TENSOR) {
+    attr.set_scales_mask(DNNL_ARG_SRC, 0);
+    if (use_azp_) {
+      attr.set_zero_points_mask(DNNL_ARG_SRC, 0);
+    }
+  }
+
+  if (b_qs_ == QuantizationStrategy::PER_TENSOR) {
+    attr.set_scales_mask(DNNL_ARG_WEIGHTS, 0);
+  } else if (b_qs_ == QuantizationStrategy::PER_OUTPUT_CHANNEL) {
+    attr.set_scales_mask(DNNL_ARG_WEIGHTS, 2);
+  }
+
+  if (key.use_bias) {
+    // For PER_TOKEN, bias will be applied in epilogue
+    assert(a_qs_ == QuantizationStrategy::PER_TENSOR);
+    dnnl::memory::desc bias_md({1, b_n_size_}, key.bias_type, {b_n_size_, 1});
+    return dnnl::matmul::primitive_desc(default_engine(), a_md, b_md, bias_md,
+                                        c_md, attr);
+  } else {
+    return dnnl::matmul::primitive_desc(default_engine(), a_md, b_md, c_md,
+                                        attr);
+  }
+}
--- a/csrc/cpu/dnnl_helper.h
+++ b/csrc/cpu/dnnl_helper.h
@ -0,0 +1,169 @@
+#ifndef DNNL_HELPER_H
+#define DNNL_HELPER_H
+
+#include <optional>
+#include <cassert>
+
+#include "oneapi/dnnl/dnnl.hpp"
+
+namespace c10 {
+struct BFloat16;
+struct Half;
+}  // namespace c10
+
+namespace dnnl {
+namespace impl {
+struct memory_storage_t;
+struct matmul_pd_t;
+struct matmul_desc_t;
+}  // namespace impl
+}  // namespace dnnl
+struct dnnl_memory_desc;
+
+template <typename KT, typename VT>
+class DNNLPrimitiveCache;
+
+template <typename T>
+struct DNNLType {
+  static constexpr dnnl::memory::data_type type =
+      dnnl::memory::data_type::undef;
+};
+
+template <>
+struct DNNLType<int8_t> {
+  static constexpr dnnl::memory::data_type type = dnnl::memory::data_type::s8;
+};
+
+template <>
+struct DNNLType<int32_t> {
+  static constexpr dnnl::memory::data_type type = dnnl::memory::data_type::s32;
+};
+
+template <>
+struct DNNLType<float> {
+  static constexpr dnnl::memory::data_type type = dnnl::memory::data_type::f32;
+};
+
+template <>
+struct DNNLType<c10::BFloat16> {
+  static constexpr dnnl::memory::data_type type = dnnl::memory::data_type::bf16;
+};
+
+template <>
+struct DNNLType<c10::Half> {
+  static constexpr dnnl::memory::data_type type = dnnl::memory::data_type::f16;
+};
+
+template <typename T>
+constexpr inline dnnl::memory::data_type get_dnnl_type() {
+  return DNNLType<std::decay_t<T>>::type;
+}
+
+class DNNLMatMulPrimitiveHandler {
+ public:
+  virtual ~DNNLMatMulPrimitiveHandler() = default;
+
+ protected:
+  struct Args {
+    dnnl_dim_t b_n_size;
+    dnnl_dim_t b_n_stride;
+    dnnl_dim_t b_k_size;
+    dnnl_dim_t b_k_stride;
+    void* b_ptr;
+    dnnl::memory::data_type c_type;
+    size_t primitive_cache_size;
+  };
+
+ protected:
+  DNNLMatMulPrimitiveHandler(const Args& args, dnnl::memory::data_type b_type);
+
+  void prepack_weight(void* original_b_ptr,
+                      dnnl::memory::desc b_target_mem_desc);
+
+  void set_runtime_memory_ptr(size_t index, dnnl_memory* memory_ptr);
+
+  std::pair<dnnl::impl::memory_storage_t*, dnnl_memory_desc*>
+  get_runtime_memory_ptr(size_t index);
+
+ protected:
+  const dnnl_dim_t b_n_size_;
+  const dnnl_dim_t b_n_stride_;
+  const dnnl_dim_t b_k_size_;
+  const dnnl_dim_t b_k_stride_;
+  dnnl::memory::data_type b_type_;
+  dnnl::memory::data_type c_type_;
+  std::unordered_map<int, dnnl::memory> memory_cache_;
+  std::vector<std::pair<dnnl::impl::memory_storage_t*, dnnl_memory_desc*>>
+      runtime_memory_ptrs_;
+  dnnl::memory::desc b_target_mem_desc_;
+  int64_t primitive_cache_size_;
+};
+
+class W8A8MatMulPrimitiveHandler : public DNNLMatMulPrimitiveHandler {
+ public:
+  enum class QuantizationStrategy { PER_TOKEN, PER_TENSOR, PER_OUTPUT_CHANNEL };
+
+  struct Args : public DNNLMatMulPrimitiveHandler::Args {
+    bool use_a_zero_point;
+    QuantizationStrategy a_quantization_strategy;
+    QuantizationStrategy b_quantization_strategy;
+    float* b_scales_ptr;
+  };
+
+  struct ClassMatmulCacheKey {
+    dnnl_dim_t b_n_size;
+    dnnl_dim_t b_k_size;
+    QuantizationStrategy a_qs;
+    QuantizationStrategy b_qs;
+    bool use_azp;
+    dnnl::memory::data_type c_type;
+
+    friend bool operator==(const ClassMatmulCacheKey& l,
+                           const ClassMatmulCacheKey& r);
+  };
+
+  struct MSizeCacheKey {
+    dnnl_dim_t a_m_size;
+    bool use_bias;
+    dnnl::memory::data_type bias_type;
+
+    friend bool operator==(const MSizeCacheKey& l, const MSizeCacheKey& r);
+  };
+
+  using MSizeCache = DNNLPrimitiveCache<MSizeCacheKey, dnnl::matmul>;
+  using ClassMatmulCache =
+      DNNLPrimitiveCache<ClassMatmulCacheKey, std::shared_ptr<MSizeCache>>;
+
+  struct ExecArgs : public MSizeCacheKey {
+    const int8_t* a_ptr;
+    const float* a_scales_ptr;
+    const int32_t* a_zero_points_ptr;
+    const void* bias_ptr;
+    void* c_ptr;
+  };
+
+ public:
+  W8A8MatMulPrimitiveHandler(const Args& args);
+
+  QuantizationStrategy get_input_scale_strategy() const { return a_qs_; }
+
+  bool get_input_use_zero_point() const { return use_azp_; }
+
+  void execute(ExecArgs& args);
+
+ private:
+  dnnl::matmul::primitive_desc create_primitive_desc(const MSizeCacheKey& key,
+                                                     bool first_time);
+
+  void init_runtime_memory_cache(const Args& args);
+
+  dnnl::matmul get_matmul_cache(const MSizeCacheKey& key);
+
+ private:
+  const bool use_azp_;
+  const QuantizationStrategy a_qs_;
+  const QuantizationStrategy b_qs_;
+  std::shared_ptr<MSizeCache> m_size_cache_;
+};
+
+#endif
--- a/csrc/cpu/dnnl_helper.hpp
+++ b/csrc/cpu/dnnl_helper.hpp
@ -1,206 +0,0 @@
-#ifndef DNNL_HELPER_HPP
-#define DNNL_HELPER_HPP
-
-#include <c10/util/BFloat16.h>
-#include <c10/util/Half.h>
-
-#include "oneapi/dnnl/dnnl.hpp"
-
-namespace {
-template <typename T>
-struct DNNLType {
-  static constexpr dnnl::memory::data_type type =
-      dnnl::memory::data_type::undef;
-};
-
-template <>
-struct DNNLType<int8_t> {
-  static constexpr dnnl::memory::data_type type = dnnl::memory::data_type::s8;
-};
-
-template <>
-struct DNNLType<int32_t> {
-  static constexpr dnnl::memory::data_type type = dnnl::memory::data_type::s32;
-};
-
-template <>
-struct DNNLType<float> {
-  static constexpr dnnl::memory::data_type type = dnnl::memory::data_type::f32;
-};
-
-template <>
-struct DNNLType<c10::BFloat16> {
-  static constexpr dnnl::memory::data_type type = dnnl::memory::data_type::bf16;
-};
-
-template <>
-struct DNNLType<c10::Half> {
-  static constexpr dnnl::memory::data_type type = dnnl::memory::data_type::f16;
-};
-
-template <typename T>
-constexpr inline dnnl::memory::data_type get_dnnl_type() {
-  return DNNLType<std::decay_t<T>>::type;
-}
-};  // namespace
-
-template <bool InputNoScale>
-class DNNLPrimitiveHelper {
- public:
-  // I8 input GEMM kernel (C = a_scales * A @ (b_scales * B^T) + bias)
-  // A: [M, K], row-major
-  // B: [K, N], column-major
-  // C: [M, N], row-major
-  // bias: [N], row-major, optional
-  // a_scales: [MS]
-  // b_scales: [NS]
-  // Note: Due to the limitation of oneDNN
-  // (https://github.com/oneapi-src/oneDNN/issues/1636), the quantized bias is
-  // not supported.
-
-  template <typename OutputT, typename BiasT>
-  static void gemm_s8s8_jit(const int8_t* a, const int8_t* b, OutputT* c,
-                            const BiasT* bias, dnnl_dim_t M, dnnl_dim_t N,
-                            dnnl_dim_t K, const float* a_scales,
-                            const float* b_scales, dnnl_dim_t MS,
-                            dnnl_dim_t NS) {
-    auto&& OutputType = get_dnnl_type<OutputT>();
-    auto&& BiasType = get_dnnl_type<BiasT>();
-
-    dnnl::memory::desc a_md({M, K}, dnnl::memory::data_type::s8, {K, 1});
-    dnnl::memory::desc b_md({K, N}, dnnl::memory::data_type::s8, {1, K});
-    dnnl::memory::desc c_md({M, N}, OutputType, {N, 1});
-
-    dnnl::primitive_attr attr;
-    if constexpr (!InputNoScale) {
-      if (MS == 1) {
-        // per-tensor
-        attr.set_scales_mask(DNNL_ARG_SRC, 0);
-      } else {
-        // per-token
-        TORCH_CHECK(false, "per-token quantization is unsupported.");
-      }
-    }
-
-    if (NS == 1) {
-      // per-tensor
-      attr.set_scales_mask(DNNL_ARG_WEIGHTS, 0);
-    } else {
-      // per-channel
-      attr.set_scales_mask(DNNL_ARG_WEIGHTS, 2);
-    }
-
-    dnnl::matmul::primitive_desc matmul_pd;
-// Create memory descriptors with format_tag::any for the primitive. This
-// enables the matmul primitive to choose memory layouts for an
-// optimized primitive implementation, and these layouts may differ from the
-// ones provided by the user.
-#ifdef __aarch64__
-    auto mat_src_md = dnnl::memory::desc({M, K}, dnnl::memory::data_type::s8,
-                                         dnnl::memory::format_tag::any);
-    auto mat_weights_md = dnnl::memory::desc(
-        {K, N}, dnnl::memory::data_type::s8, dnnl::memory::format_tag::any);
-    auto mat_dst_md =
-        dnnl::memory::desc({M, N}, OutputType, dnnl::memory::format_tag::any);
-    if (bias) {
-      dnnl::memory::desc bias_md({1, N}, BiasType, {N, 1});
-      matmul_pd = dnnl::matmul::primitive_desc(default_engine(), mat_src_md,
-                                               mat_weights_md, bias_md,
-                                               mat_dst_md, attr);
-    } else {
-      matmul_pd = dnnl::matmul::primitive_desc(
-          default_engine(), mat_src_md, mat_weights_md, mat_dst_md, attr);
-    }
-#else
-    if (bias) {
-      dnnl::memory::desc bias_md({1, N}, BiasType, {N, 1});
-      matmul_pd = dnnl::matmul::primitive_desc(default_engine(), a_md, b_md,
-                                               bias_md, c_md, attr);
-    } else {
-      matmul_pd = dnnl::matmul::primitive_desc(default_engine(), a_md, b_md,
-                                               c_md, attr);
-    }
-#endif
-    dnnl::matmul matmul(matmul_pd);
-
-    auto& engine = default_engine();
-
-    dnnl::memory a_m(a_md, engine, (void*)a);
-    dnnl::memory b_m(b_md, engine, (void*)b);
-    dnnl::memory c_m(c_md, engine, (void*)c);
-    dnnl::memory a_scales_m({{MS}, dnnl::memory::data_type::f32, {1}}, engine,
-                            (void*)a_scales);
-    dnnl::memory b_scales_m({{NS}, dnnl::memory::data_type::f32, {1}}, engine,
-                            (void*)b_scales);
-
-    auto& stream = default_stream();
-
-    auto mat_src_mem = a_m;
-    auto mat_weights_mem = b_m;
-    auto mat_dst_mem = c_m;
-#ifdef __aarch64__
-    if (matmul_pd.weights_desc() != b_m.get_desc()) {
-      mat_weights_mem = dnnl::memory(matmul_pd.weights_desc(), engine);
-      dnnl::reorder(b_m, mat_weights_mem).execute(stream, b_m, mat_weights_mem);
-    }
-#endif
-    if constexpr (InputNoScale) {
-      if (bias) {
-        dnnl::memory::desc bias_md({N}, BiasType, {1});
-        dnnl::memory bias_m(bias_md, engine, (void*)bias);
-        matmul.execute(
-            stream, {
-                        {DNNL_ARG_SRC, mat_src_mem},
-                        {DNNL_ARG_WEIGHTS, mat_weights_mem},
-                        {DNNL_ARG_BIAS, bias_m},
-                        {DNNL_ARG_DST, mat_dst_mem},
-                        {DNNL_ARG_ATTR_SCALES | DNNL_ARG_WEIGHTS, b_scales_m},
-                    });
-      } else {
-        matmul.execute(
-            stream, {
-                        {DNNL_ARG_SRC, mat_src_mem},
-                        {DNNL_ARG_WEIGHTS, mat_weights_mem},
-                        {DNNL_ARG_DST, mat_dst_mem},
-                        {DNNL_ARG_ATTR_SCALES | DNNL_ARG_WEIGHTS, b_scales_m},
-                    });
-      }
-    } else {
-      if (bias) {
-        dnnl::memory::desc bias_md({N}, BiasType, {1});
-        dnnl::memory bias_m(bias_md, engine, (void*)bias);
-        matmul.execute(
-            stream, {
-                        {DNNL_ARG_SRC, mat_src_mem},
-                        {DNNL_ARG_WEIGHTS, mat_weights_mem},
-                        {DNNL_ARG_BIAS, bias_m},
-                        {DNNL_ARG_DST, mat_dst_mem},
-                        {DNNL_ARG_ATTR_SCALES | DNNL_ARG_SRC, a_scales_m},
-                        {DNNL_ARG_ATTR_SCALES | DNNL_ARG_WEIGHTS, b_scales_m},
-                    });
-      } else {
-        matmul.execute(
-            stream, {
-                        {DNNL_ARG_SRC, mat_src_mem},
-                        {DNNL_ARG_WEIGHTS, mat_weights_mem},
-                        {DNNL_ARG_DST, mat_dst_mem},
-                        {DNNL_ARG_ATTR_SCALES | DNNL_ARG_SRC, a_scales_m},
-                        {DNNL_ARG_ATTR_SCALES | DNNL_ARG_WEIGHTS, b_scales_m},
-                    });
-      }
-    }
-    stream.wait();
-  }
-
- private:
-  static dnnl::engine& default_engine() {
-    static dnnl::engine engine(dnnl::engine::kind::cpu, 0);
-    return engine;
-  }
-
-  static dnnl::stream& default_stream() {
-    static dnnl::stream stream(default_engine());
-    return stream;
-  }
-};
-#endif
--- a/csrc/cpu/dnnl_kernels.cpp
+++ b/csrc/cpu/dnnl_kernels.cpp
@ -0,0 +1,494 @@
+#include "cpu_types.hpp"
+#include "dnnl_helper.h"
+
+namespace {
+template <typename scalar_t>
+struct KernelVecType {
+  using load_vec_type = void;
+  using cvt_vec_type = void;
+};
+
+template <>
+struct KernelVecType<float> {
+  using load_vec_type = vec_op::FP32Vec16;
+  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 cvt_vec_type = vec_op::FP32Vec16;
+};
+#endif
+
+template <>
+struct KernelVecType<c10::Half> {
+#if defined(__powerpc64__) || defined(__s390x__)
+  // Power architecture-specific vector type
+  using load_vec_type = vec_op::FP32Vec16;
+#else
+  // Fallback for other architectures
+  using load_vec_type = vec_op::FP16Vec16;
+#endif
+  using cvt_vec_type = vec_op::FP32Vec16;
+};
+
+template <bool AZP, typename scalar_t>
+void static_scaled_int8_quant_impl(const scalar_t* input, int8_t* output,
+                                   const float* scale, const int32_t* azp,
+                                   const int64_t num_tokens,
+                                   const int64_t input_stride,
+                                   const int64_t hidden_size) {
+  using load_vec_t = typename KernelVecType<scalar_t>::load_vec_type;
+  using cvt_vec_t = typename KernelVecType<scalar_t>::cvt_vec_type;
+  constexpr int64_t vec_elem_num = load_vec_t::VEC_ELEM_NUM;
+
+  constexpr float i8_min =
+      static_cast<float>(std::numeric_limits<int8_t>::min());
+  constexpr float i8_max =
+      static_cast<float>(std::numeric_limits<int8_t>::max());
+  const cvt_vec_t inv_scale(1.0 / *scale);
+  const cvt_vec_t i8_min_vec(i8_min);
+  const cvt_vec_t i8_max_vec(i8_max);
+
+  cvt_vec_t zp_vec;
+  if constexpr (AZP) {
+    zp_vec = cvt_vec_t(static_cast<float>(*azp));
+  }
+
+#pragma omp parallel for
+  for (int64_t i = 0; i < num_tokens; ++i) {
+    int64_t j = 0;
+    const scalar_t* input_ptr = input + i * input_stride;
+    int8_t* output_ptr = output + i * hidden_size;
+    for (; j < hidden_size - vec_elem_num; j += vec_elem_num) {
+      load_vec_t elems(input_ptr + j);
+      cvt_vec_t elems_fp32(elems);
+      elems_fp32 = elems_fp32 * inv_scale;
+
+      if constexpr (AZP) {
+        elems_fp32 = elems_fp32 + zp_vec;
+      }
+
+      elems_fp32 = elems_fp32.clamp(i8_min_vec, i8_max_vec);
+      vec_op::INT8Vec16 elems_int8(elems_fp32);
+      elems_int8.save(output_ptr + j);
+    }
+
+    load_vec_t elems(input_ptr + j);
+    cvt_vec_t elems_fp32(elems);
+    elems_fp32 = elems_fp32 * inv_scale;
+
+    if constexpr (AZP) {
+      elems_fp32 = elems_fp32 + zp_vec;
+    }
+
+    elems_fp32 = elems_fp32.clamp(i8_min_vec, i8_max_vec);
+    vec_op::INT8Vec16 elems_int8(elems_fp32);
+    elems_int8.save(output_ptr + j, hidden_size - j);
+  }
+}
+
+template <bool AZP, typename scalar_t>
+void dynamic_scaled_int8_quant_impl(const scalar_t* input, int8_t* output,
+                                    float* scale, int32_t* azp,
+                                    const int64_t num_tokens,
+                                    const int64_t input_stride,
+                                    const int64_t hidden_size) {
+  using load_vec_t = typename KernelVecType<scalar_t>::load_vec_type;
+  using cvt_vec_t = typename KernelVecType<scalar_t>::cvt_vec_type;
+  constexpr int vec_elem_num = load_vec_t::VEC_ELEM_NUM;
+
+  constexpr float i8_min =
+      static_cast<float>(std::numeric_limits<int8_t>::min());
+  constexpr float i8_max =
+      static_cast<float>(std::numeric_limits<int8_t>::max());
+  const cvt_vec_t i8_min_vec(i8_min);
+  const cvt_vec_t i8_max_vec(i8_max);
+
+#pragma omp parallel for
+  for (int64_t i = 0; i < num_tokens; ++i) {
+    cvt_vec_t max_value(std::numeric_limits<float>::lowest());
+    cvt_vec_t min_value(std::numeric_limits<float>::max());
+    {
+      int64_t j = 0;
+      const scalar_t* input_ptr = input + i * input_stride;
+      for (; j < hidden_size - vec_elem_num; j += vec_elem_num) {
+        load_vec_t elems(input_ptr + j);
+        cvt_vec_t elems_fp32(elems);
+        if constexpr (AZP) {
+          max_value = max_value.max(elems_fp32);
+          min_value = min_value.min(elems_fp32);
+        } else {
+          max_value = max_value.max(elems_fp32.abs());
+        }
+      }
+
+      load_vec_t elems(input_ptr + j);
+      cvt_vec_t elems_fp32(elems);
+
+      if (j + vec_elem_num == hidden_size) {
+        if constexpr (AZP) {
+          max_value = max_value.max(elems_fp32);
+          min_value = min_value.min(elems_fp32);
+        } else {
+          max_value = max_value.max(elems_fp32.abs());
+        }
+      } else {
+        if constexpr (AZP) {
+          max_value = max_value.max(elems_fp32, hidden_size - j);
+          min_value = min_value.min(elems_fp32, hidden_size - j);
+        } else {
+          max_value = max_value.max(elems_fp32.abs(), hidden_size - j);
+        }
+      }
+    }
+
+    float scale_val, azp_val;
+    if constexpr (AZP) {
+      float max_scalar = max_value.reduce_max();
+      float min_scalar = min_value.reduce_min();
+      scale_val = (max_scalar - min_scalar) / 255.0f;
+      azp_val = std::nearbyint(-128.0f - min_scalar / scale_val);
+      azp[i] = azp_val;
+      scale[i] = scale_val;
+    } else {
+      scale_val = max_value.reduce_max() / 127.0f;
+      scale[i] = scale_val;
+    }
+
+    const cvt_vec_t inv_scale(1.0 / scale_val);
+    const cvt_vec_t azp_vec(azp_val);
+
+    {
+      int64_t j = 0;
+      const scalar_t* input_ptr = input + i * input_stride;
+      int8_t* output_ptr = output + i * hidden_size;
+      for (; j < hidden_size - vec_elem_num; j += vec_elem_num) {
+        load_vec_t elems(input_ptr + j);
+        cvt_vec_t elems_fp32(elems);
+        elems_fp32 = (elems_fp32 * inv_scale);
+
+        if constexpr (AZP) {
+          elems_fp32 = elems_fp32 + azp_vec;
+        }
+        elems_fp32 = elems_fp32.clamp(i8_min_vec, i8_max_vec);
+        vec_op::INT8Vec16 elems_int8(elems_fp32);
+        elems_int8.save(output_ptr + j);
+      }
+
+      load_vec_t elems(input_ptr + j);
+      cvt_vec_t elems_fp32(elems);
+      elems_fp32 = (elems_fp32 * inv_scale);
+
+      if constexpr (AZP) {
+        elems_fp32 = elems_fp32 + azp_vec;
+      }
+      elems_fp32 = elems_fp32.clamp(i8_min_vec, i8_max_vec);
+      vec_op::INT8Vec16 elems_int8(elems_fp32);
+      elems_int8.save(output_ptr + j, hidden_size - j);
+    }
+  }
+}
+
+template <bool AZP, bool Bias, typename scalar_t>
+void dynamic_quant_epilogue(const float* input, scalar_t* output,
+                            const float* a_scale, const int32_t* azp,
+                            const float* azp_adj, const scalar_t* bias,
+                            const int64_t num_tokens,
+                            const int64_t hidden_size) {
+  CPU_KERNEL_GUARD_IN(dynamic_quant_epilogue)
+  using load_vec_t = typename KernelVecType<scalar_t>::load_vec_type;
+  using cvt_vec_t = typename KernelVecType<scalar_t>::cvt_vec_type;
+  constexpr int vec_elem_num = load_vec_t::VEC_ELEM_NUM;
+
+  const int64_t thread_num = omp_get_max_threads();
+  if (num_tokens > thread_num) {
+#pragma omp parallel for
+    for (int64_t i = 0; i < num_tokens; ++i) {
+      const float* input_ptr = input + i * hidden_size;
+      scalar_t* output_ptr = output + i * hidden_size;
+      int64_t j = 0;
+      cvt_vec_t token_scale_vec(a_scale[i]);
+      cvt_vec_t token_zp_scale_vec;
+      if constexpr (AZP) {
+        float zp_scale_val = a_scale[i] * static_cast<float>(azp[i]);
+        token_zp_scale_vec = cvt_vec_t(zp_scale_val);
+      }
+      for (; j < hidden_size - vec_elem_num; ++j) {
+        cvt_vec_t elems_fp32(input_ptr + j);
+        elems_fp32 = elems_fp32 * token_scale_vec;
+        if constexpr (AZP) {
+          cvt_vec_t azp_adj_fp32(azp_adj + j);
+          elems_fp32 = elems_fp32 - azp_adj_fp32 * token_zp_scale_vec;
+        }
+        if constexpr (Bias) {
+          load_vec_t bias_vec(bias + j);
+          cvt_vec_t bias_vec_fp32(bias_vec);
+          elems_fp32 = elems_fp32 + bias_vec_fp32;
+        }
+        load_vec_t elems_out(elems_fp32);
+        elems_out.save(output_ptr + j);
+      }
+      cvt_vec_t elems_fp32(input_ptr + j);
+      elems_fp32 = elems_fp32 * token_scale_vec;
+      if constexpr (AZP) {
+        cvt_vec_t azp_adj_fp32(azp_adj + j);
+        elems_fp32 = elems_fp32 - azp_adj_fp32 * token_zp_scale_vec;
+      }
+      if constexpr (Bias) {
+        load_vec_t bias_vec(bias + j);
+        cvt_vec_t bias_vec_fp32(bias_vec);
+        elems_fp32 = elems_fp32 + bias_vec_fp32;
+      }
+      load_vec_t elems_out(elems_fp32);
+      elems_out.save(output_ptr + j, hidden_size - j);
+    }
+  } else {
+    const int64_t vec_iteration =
+        (hidden_size + vec_elem_num - 1) / vec_elem_num;
+    const int64_t vec_iteration_per_thread =
+        (vec_iteration + thread_num - 1) / thread_num;
+    const int64_t elem_num_per_thread = vec_iteration_per_thread * vec_elem_num;
+#pragma omp parallel for schedule(static, 1)
+    for (int64_t i = 0; i < thread_num; ++i) {
+      const int64_t start = elem_num_per_thread * i;
+      const int64_t end = std::min(hidden_size, elem_num_per_thread + start);
+      for (int64_t j = 0; j < num_tokens; ++j) {
+        cvt_vec_t token_scale_vec(a_scale[j]);
+        cvt_vec_t token_zp_scale_vec;
+        if constexpr (AZP) {
+          float zp_scale_val = a_scale[j] * static_cast<float>(azp[j]);
+          token_zp_scale_vec = cvt_vec_t(zp_scale_val);
+        }
+        int64_t k = start;
+        const float* input_ptr = input + j * hidden_size;
+        scalar_t* output_ptr = output + j * hidden_size;
+        for (; k < end - vec_elem_num; k += vec_elem_num) {
+          cvt_vec_t elems_fp32(input_ptr + k);
+          elems_fp32 = elems_fp32 * token_scale_vec;
+          if constexpr (AZP) {
+            cvt_vec_t azp_adj_fp32(azp_adj + k);
+            elems_fp32 = elems_fp32 - azp_adj_fp32 * token_zp_scale_vec;
+          }
+          if constexpr (Bias) {
+            load_vec_t bias_vec(bias + k);
+            cvt_vec_t bias_vec_fp32(bias_vec);
+            elems_fp32 = elems_fp32 + bias_vec_fp32;
+          }
+          load_vec_t elems_out(elems_fp32);
+          elems_out.save(output_ptr + k);
+        }
+        if (k < end) {
+          cvt_vec_t elems_fp32(input_ptr + k);
+          elems_fp32 = elems_fp32 * token_scale_vec;
+          if constexpr (AZP) {
+            cvt_vec_t azp_adj_fp32(azp_adj + k);
+            elems_fp32 = elems_fp32 - azp_adj_fp32 * token_zp_scale_vec;
+          }
+          if constexpr (Bias) {
+            load_vec_t bias_vec(bias + k);
+            cvt_vec_t bias_vec_fp32(bias_vec);
+            elems_fp32 = elems_fp32 + bias_vec_fp32;
+          }
+          load_vec_t elems_out(elems_fp32);
+          elems_out.save(output_ptr + k, end - k);
+        }
+      }
+    }
+  }
+}
+}  // namespace
+
+int64_t create_onednn_scaled_mm_handler(
+    const torch::Tensor& b,         // [IC, OC], column-major
+    const torch::Tensor& b_scales,  // [1] or [OC]
+    at::ScalarType output_type, bool dynamic_act_quant, bool use_azp,
+    int64_t primitive_cache_size) {
+  TORCH_CHECK(b.dim() == 2);
+  TORCH_CHECK(b.stride(0) == 1);  // Column-major
+  TORCH_CHECK(b_scales.is_contiguous());
+
+  W8A8MatMulPrimitiveHandler::Args args;
+  args.primitive_cache_size = primitive_cache_size;
+
+  if (b_scales.numel() == 1) {
+    args.b_quantization_strategy =
+        W8A8MatMulPrimitiveHandler::QuantizationStrategy::PER_TENSOR;
+  } else {
+    TORCH_CHECK_EQ(b_scales.numel(), b.size(1));
+    args.b_quantization_strategy =
+        W8A8MatMulPrimitiveHandler::QuantizationStrategy::PER_OUTPUT_CHANNEL;
+  }
+  args.b_scales_ptr = b_scales.data_ptr<float>();
+  args.b_k_size = b.size(0);
+  args.b_k_stride = b.stride(0);
+  args.b_n_size = b.size(1);
+  args.b_n_stride = b.stride(1);
+  args.b_ptr = b.data_ptr<int8_t>();
+
+  if (dynamic_act_quant) {
+    // dynamic per-token, bias, A scales and A zps will be applied in outside.
+    args.a_quantization_strategy =
+        W8A8MatMulPrimitiveHandler::QuantizationStrategy::PER_TOKEN;
+    args.use_a_zero_point = false;
+  } else {
+    // static per-tensor
+    args.a_quantization_strategy =
+        W8A8MatMulPrimitiveHandler::QuantizationStrategy::PER_TENSOR;
+    args.use_a_zero_point = use_azp;
+  }
+
+  VLLM_DISPATCH_FLOATING_TYPES(output_type, "create_onednn_scaled_mm_handler",
+                               [&] {
+                                 if (dynamic_act_quant) {
+                                   args.c_type = get_dnnl_type<float>();
+                                 } else {
+                                   args.c_type = get_dnnl_type<scalar_t>();
+                                 }
+                               });
+
+  return reinterpret_cast<int64_t>(new W8A8MatMulPrimitiveHandler(args));
+}
+
+void onednn_scaled_mm(
+    torch::Tensor& c,                             // [M, OC], row-major
+    const torch::Tensor& a,                       // [M, IC], row-major
+    const torch::Tensor& a_scales,                // [M] or [1]
+    const std::optional<torch::Tensor>& azp,      // [M] or [1]
+    const std::optional<torch::Tensor>& azp_adj,  // [M] or [1]
+    const std::optional<torch::Tensor>& bias,     // [N]
+    int64_t handler) {
+  CPU_KERNEL_GUARD_IN(onednn_scaled_mm)
+  TORCH_CHECK(a.dim() == 2);
+  TORCH_CHECK(a.is_contiguous());
+  TORCH_CHECK(c.is_contiguous());
+  W8A8MatMulPrimitiveHandler* ptr =
+      reinterpret_cast<W8A8MatMulPrimitiveHandler*>(handler);
+  const int32_t* azp_ptr = nullptr;
+  if (azp.has_value()) {
+    azp_ptr = azp->data_ptr<int32_t>();
+  }
+  if (ptr->get_input_scale_strategy() ==
+      W8A8MatMulPrimitiveHandler::QuantizationStrategy::PER_TENSOR) {
+    TORCH_CHECK_EQ(a_scales.numel(), 1);
+  }
+
+  W8A8MatMulPrimitiveHandler::ExecArgs exec_args;
+  exec_args.a_ptr = a.data_ptr<int8_t>();
+  exec_args.a_m_size = a.size(0);
+  exec_args.bias_ptr = nullptr;
+  exec_args.use_bias = false;
+  exec_args.a_scales_ptr = nullptr;
+  exec_args.a_zero_points_ptr = nullptr;
+
+  VLLM_DISPATCH_FLOATING_TYPES(c.scalar_type(), "onednn_scaled_mm", [&] {
+    if (ptr->get_input_scale_strategy() ==
+        W8A8MatMulPrimitiveHandler::QuantizationStrategy::PER_TENSOR) {
+      if (bias.has_value()) {
+        exec_args.bias_ptr = bias->data_ptr<scalar_t>();
+        exec_args.bias_type = get_dnnl_type<scalar_t>();
+        exec_args.use_bias = true;
+      }
+      exec_args.a_scales_ptr = a_scales.data_ptr<float>();
+      exec_args.a_zero_points_ptr = azp_ptr;
+      exec_args.c_ptr = c.data_ptr<scalar_t>();
+      ptr->execute(exec_args);
+    } else if (ptr->get_input_scale_strategy() ==
+               W8A8MatMulPrimitiveHandler::QuantizationStrategy::PER_TOKEN) {
+      torch::Tensor tmp_fp32_out =
+          torch::empty_like(c, ::at::ScalarType::Float);
+      exec_args.c_ptr = tmp_fp32_out.data_ptr<float>();
+      ptr->execute(exec_args);
+      if (bias.has_value()) {
+        if (azp.has_value()) {
+          dynamic_quant_epilogue<true, true>(
+              tmp_fp32_out.data_ptr<float>(), c.data_ptr<scalar_t>(),
+              a_scales.data_ptr<float>(), azp_ptr, azp_adj->data_ptr<float>(),
+              bias->data_ptr<scalar_t>(), c.size(0), c.size(1));
+        } else {
+          dynamic_quant_epilogue<false, true>(
+              tmp_fp32_out.data_ptr<float>(), c.data_ptr<scalar_t>(),
+              a_scales.data_ptr<float>(), azp_ptr, nullptr,
+              bias->data_ptr<scalar_t>(), c.size(0), c.size(1));
+        }
+      } else {
+        if (azp.has_value()) {
+          dynamic_quant_epilogue<true, false>(
+              tmp_fp32_out.data_ptr<float>(), c.data_ptr<scalar_t>(),
+              a_scales.data_ptr<float>(), azp_ptr, azp_adj->data_ptr<float>(),
+              (scalar_t*)nullptr, c.size(0), c.size(1));
+        } else {
+          dynamic_quant_epilogue<false, false>(
+              tmp_fp32_out.data_ptr<float>(), c.data_ptr<scalar_t>(),
+              a_scales.data_ptr<float>(), azp_ptr, nullptr, (scalar_t*)nullptr,
+              c.size(0), c.size(1));
+        }
+      }
+    } else {
+      TORCH_CHECK(false, "invalid act quant type.");
+    }
+  });
+}
+
+// static-per-tensor quantization.
+void static_scaled_int8_quant(
+    torch::Tensor& out,          // [batch, hidden_size]
+    const torch::Tensor& input,  // [batch, hidden_size]
+    const torch::Tensor& scale, std::optional<torch::Tensor> const& azp) {
+  CPU_KERNEL_GUARD_IN(static_scaled_int8_quant)
+  TORCH_CHECK(out.is_contiguous());
+  TORCH_CHECK_EQ(input.dim(), 2);
+  TORCH_CHECK_EQ(input.stride(1), 1);
+  TORCH_CHECK(scale.numel() == 1);
+  TORCH_CHECK(!azp.has_value() || azp->numel() == 1);
+
+  const int64_t stride = input.stride(0);
+  const int64_t hidden_size = input.size(1);
+  const int64_t num_tokens = input.size(0);
+  VLLM_DISPATCH_FLOATING_TYPES(
+      input.scalar_type(), "static_scaled_int8_quant_impl", [&] {
+        if (azp.has_value()) {
+          static_scaled_int8_quant_impl<true>(
+              input.data_ptr<scalar_t>(), out.data_ptr<int8_t>(),
+              scale.data_ptr<float>(), azp->data_ptr<int32_t>(), num_tokens,
+              stride, hidden_size);
+        } else {
+          static_scaled_int8_quant_impl<false>(input.data_ptr<scalar_t>(),
+                                               out.data_ptr<int8_t>(),
+                                               scale.data_ptr<float>(), nullptr,
+                                               num_tokens, stride, hidden_size);
+        }
+      });
+}
+
+// dynamic-per-token quantization.
+void dynamic_scaled_int8_quant(
+    torch::Tensor& out,          // [batch, hidden_size]
+    const torch::Tensor& input,  // [batch, hidden_size]
+    torch::Tensor& scale,        // [batch, 1]
+    std::optional<torch::Tensor> const& azp) {
+  CPU_KERNEL_GUARD_IN(dynamic_scaled_int8_quant)
+  TORCH_CHECK(out.is_contiguous());
+  TORCH_CHECK_EQ(input.dim(), 2);
+  TORCH_CHECK_EQ(input.stride(1), 1);
+
+  const int64_t hidden_size = input.size(1);
+  const int64_t num_tokens = input.size(0);
+  const int64_t stride = input.stride(0);
+  VLLM_DISPATCH_FLOATING_TYPES(
+      input.scalar_type(), "dynamic_scaled_int8_quant_impl", [&] {
+        if (azp.has_value()) {
+          dynamic_scaled_int8_quant_impl<true>(
+              input.data_ptr<scalar_t>(), out.data_ptr<int8_t>(),
+              scale.data_ptr<float>(), azp->data_ptr<int32_t>(), num_tokens,
+              stride, hidden_size);
+        } else {
+          dynamic_scaled_int8_quant_impl<false>(
+              input.data_ptr<scalar_t>(), out.data_ptr<int8_t>(),
+              scale.data_ptr<float>(), nullptr, num_tokens, stride,
+              hidden_size);
+        }
+      });
+}
--- a/csrc/cpu/quant.cpp
+++ b/csrc/cpu/quant.cpp
@ -1,951 +0,0 @@
-#include "cpu_types.hpp"
-#include "dnnl_helper.hpp"
-
-namespace {
-template <typename scalar_t>
-struct KernelVecType {
-  using load_vec_type = void;
-  using azp_adj_load_vec_type = void;
-  using cvt_vec_type = void;
-};
-
-template <>
-struct KernelVecType<float> {
-  using load_vec_type = vec_op::FP32Vec16;
-  using azp_adj_load_vec_type = vec_op::INT32Vec16;
-  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> {
-#if defined(__powerpc64__) || defined(__s390x__)
-  // Power architecture-specific vector type
-  using load_vec_type = vec_op::FP32Vec16;
-#else
-  // Fallback for other architectures
-  using load_vec_type = vec_op::FP16Vec16;
-#endif
-  using azp_adj_load_vec_type = vec_op::INT32Vec16;
-  using cvt_vec_type = vec_op::FP32Vec16;
-};
-
-#if defined(__AVX512F__) || defined(__aarch64__)
-template <bool AZP, typename scalar_t>
-void static_scaled_int8_quant_impl(const scalar_t* input, int8_t* output,
-                                   const float* scale, const int32_t* azp,
-                                   const int num_tokens,
-                                   const int hidden_size) {
-  using load_vec_t = typename KernelVecType<scalar_t>::load_vec_type;
-  using cvt_vec_t = typename KernelVecType<scalar_t>::cvt_vec_type;
-  constexpr int vec_elem_num = load_vec_t::VEC_ELEM_NUM;
-
-  constexpr float i8_min =
-      static_cast<float>(std::numeric_limits<int8_t>::min());
-  constexpr float i8_max =
-      static_cast<float>(std::numeric_limits<int8_t>::max());
-  const cvt_vec_t inv_scale(1.0 / *scale);
-  const cvt_vec_t i8_min_vec(i8_min);
-  const cvt_vec_t i8_max_vec(i8_max);
-
-  cvt_vec_t zp_vec;
-  if constexpr (AZP) {
-    zp_vec = cvt_vec_t(static_cast<float>(*azp));
-  }
-
-  #pragma omp parallel for
-  for (int i = 0; i < num_tokens; ++i) {
-    int j = 0;
-    for (; j < hidden_size - vec_elem_num; j += vec_elem_num) {
-      load_vec_t elems(input + i * hidden_size + j);
-      cvt_vec_t elems_fp32(elems);
-      elems_fp32 = elems_fp32 * inv_scale;
-
-      if constexpr (AZP) {
-        elems_fp32 = elems_fp32 + zp_vec;
-      }
-
-      elems_fp32 = elems_fp32.clamp(i8_min_vec, i8_max_vec);
-      vec_op::INT8Vec16 elems_int8(elems_fp32);
-      elems_int8.save(output + i * hidden_size + j);
-    }
-
-    load_vec_t elems(input + i * hidden_size + j);
-    cvt_vec_t elems_fp32(elems);
-    elems_fp32 = elems_fp32 * inv_scale;
-
-    if constexpr (AZP) {
-      elems_fp32 = elems_fp32 + zp_vec;
-    }
-
-    elems_fp32 = elems_fp32.clamp(i8_min_vec, i8_max_vec);
-    vec_op::INT8Vec16 elems_int8(elems_fp32);
-    elems_int8.save(output + i * hidden_size + j, hidden_size - j);
-  }
-}
-
-template <bool AZP, typename scalar_t>
-void dynamic_scaled_int8_quant_impl(const scalar_t* input, int8_t* output,
-                                    float* scale, int32_t* azp,
-                                    const int num_tokens,
-                                    const int hidden_size) {
-  using load_vec_t = typename KernelVecType<scalar_t>::load_vec_type;
-  using cvt_vec_t = typename KernelVecType<scalar_t>::cvt_vec_type;
-  constexpr int vec_elem_num = load_vec_t::VEC_ELEM_NUM;
-
-  constexpr float i8_min =
-      static_cast<float>(std::numeric_limits<int8_t>::min());
-  constexpr float i8_max =
-      static_cast<float>(std::numeric_limits<int8_t>::max());
-  const cvt_vec_t i8_min_vec(i8_min);
-  const cvt_vec_t i8_max_vec(i8_max);
-
-  #pragma omp parallel for
-  for (int i = 0; i < num_tokens; ++i) {
-    cvt_vec_t max_value(std::numeric_limits<float>::lowest());
-    cvt_vec_t min_value(std::numeric_limits<float>::max());
-    {
-      int j = 0;
-      for (; j < hidden_size - vec_elem_num; j += vec_elem_num) {
-        load_vec_t elems(input + i * hidden_size + j);
-        cvt_vec_t elems_fp32(elems);
-        if constexpr (AZP) {
-          max_value = max_value.max(elems_fp32);
-          min_value = min_value.min(elems_fp32);
-        } else {
-          max_value = max_value.max(elems_fp32.abs());
-        }
-      }
-
-      load_vec_t elems(input + i * hidden_size + j);
-      cvt_vec_t elems_fp32(elems);
-
-      if (j + vec_elem_num == hidden_size) {
-        if constexpr (AZP) {
-          max_value = max_value.max(elems_fp32);
-          min_value = min_value.min(elems_fp32);
-        } else {
-          max_value = max_value.max(elems_fp32.abs());
-        }
-      } else {
-        if constexpr (AZP) {
-          max_value = max_value.max(elems_fp32, hidden_size - j);
-          min_value = min_value.min(elems_fp32, hidden_size - j);
-        } else {
-          max_value = max_value.max(elems_fp32.abs(), hidden_size - j);
-        }
-      }
-    }
-
-    float scale_val, azp_val;
-    if constexpr (AZP) {
-      float max_scalar = max_value.reduce_max();
-      float min_scalar = min_value.reduce_min();
-      scale_val = (max_scalar - min_scalar) / 255.0f;
-      azp_val = std::nearbyint(-128.0f - min_scalar / scale_val);
-      azp[i] = static_cast<int32_t>(azp_val);
-      scale[i] = scale_val;
-    } else {
-      scale_val = max_value.reduce_max() / 127.0f;
-      scale[i] = scale_val;
-    }
-
-    const cvt_vec_t inv_scale(1.0 / scale_val);
-    const cvt_vec_t azp_vec(azp_val);
-
-    {
-      int j = 0;
-      for (; j < hidden_size - vec_elem_num; j += vec_elem_num) {
-        load_vec_t elems(input + i * hidden_size + j);
-        cvt_vec_t elems_fp32(elems);
-        elems_fp32 = (elems_fp32 * inv_scale);
-
-        if constexpr (AZP) {
-          elems_fp32 = elems_fp32 + azp_vec;
-        }
-        elems_fp32 = elems_fp32.clamp(i8_min_vec, i8_max_vec);
-        vec_op::INT8Vec16 elems_int8(elems_fp32);
-        elems_int8.save(output + i * hidden_size + j);
-      }
-
-      load_vec_t elems(input + i * hidden_size + j);
-      cvt_vec_t elems_fp32(elems);
-      elems_fp32 = (elems_fp32 * inv_scale);
-
-      if constexpr (AZP) {
-        elems_fp32 = elems_fp32 + azp_vec;
-      }
-      elems_fp32 = elems_fp32.clamp(i8_min_vec, i8_max_vec);
-      vec_op::INT8Vec16 elems_int8(elems_fp32);
-      elems_int8.save(output + i * hidden_size + j, hidden_size - j);
-    }
-  }
-}
-
-template <bool PerChannel, typename scalar_t>
-void static_quant_epilogue(const float* input, scalar_t* output,
-                           const float a_scale, const float* b_scale,
-                           const int32_t* azp_with_adj, const int num_tokens,
-                           const int hidden_size) {
-  CPU_KERNEL_GUARD_IN(dynamic_output_scale_impl)
-  using load_vec_t = typename KernelVecType<scalar_t>::load_vec_type;
-  using azp_adj_load_vec_t =
-      typename KernelVecType<scalar_t>::azp_adj_load_vec_type;
-  using cvt_vec_t = typename KernelVecType<scalar_t>::cvt_vec_type;
-  constexpr int vec_elem_num = load_vec_t::VEC_ELEM_NUM;
-
-  #pragma omp parallel for
-  for (int i = 0; i < num_tokens; ++i) {
-    cvt_vec_t a_scale_vec(a_scale);
-    cvt_vec_t b_scale_vec(*b_scale);
-    cvt_vec_t scale_vec = a_scale_vec * b_scale_vec;
-
-    int j = 0;
-    for (; j < hidden_size - vec_elem_num; j += vec_elem_num) {
-      cvt_vec_t elems_fp32(input + i * hidden_size + j);
-      azp_adj_load_vec_t azp_adj_vec(azp_with_adj + j);
-      cvt_vec_t azp_adj_fp32(azp_adj_vec);
-
-      if constexpr (PerChannel) {
-        b_scale_vec = cvt_vec_t(b_scale + j);
-        scale_vec = b_scale_vec * a_scale_vec;
-      }
-
-      elems_fp32 = elems_fp32 - scale_vec * azp_adj_fp32;
-
-      load_vec_t elems_out(elems_fp32);
-      elems_out.save(output + i * hidden_size + j);
-    }
-
-    cvt_vec_t elems_fp32(input + i * hidden_size + j);
-    azp_adj_load_vec_t azp_adj_vec(azp_with_adj + j);
-    cvt_vec_t azp_adj_fp32(azp_adj_vec);
-
-    if constexpr (PerChannel) {
-      b_scale_vec = cvt_vec_t(b_scale + j);
-      scale_vec = b_scale_vec * a_scale_vec;
-    }
-
-    elems_fp32 = elems_fp32 - scale_vec * azp_adj_fp32;
-
-    load_vec_t elems_out(elems_fp32);
-    elems_out.save(output + i * hidden_size + j, hidden_size - j);
-  }
-}
-
-template <bool AZP, bool PerChannel, bool Bias, typename scalar_t>
-void dynamic_quant_epilogue(const float* input, scalar_t* output,
-                            const float* a_scale, const float* b_scale,
-                            const int32_t* azp, const int32_t* azp_adj,
-                            const scalar_t* bias, const int num_tokens,
-                            const int hidden_size) {
-  CPU_KERNEL_GUARD_IN(dynamic_quant_epilogue)
-  using load_vec_t = typename KernelVecType<scalar_t>::load_vec_type;
-  using azp_adj_load_vec_t =
-      typename KernelVecType<scalar_t>::azp_adj_load_vec_type;
-  using cvt_vec_t = typename KernelVecType<scalar_t>::cvt_vec_type;
-  constexpr int vec_elem_num = load_vec_t::VEC_ELEM_NUM;
-
-  #pragma omp parallel for
-  for (int i = 0; i < num_tokens; ++i) {
-    int j = 0;
-    cvt_vec_t token_scale_vec(a_scale[i]);
-    cvt_vec_t token_zp_scale_vec;
-    if constexpr (AZP) {
-      float zp_scale_val = a_scale[i] * static_cast<float>(azp[i]);
-      if constexpr (!PerChannel) {
-        zp_scale_val *= *b_scale;
-      }
-      token_zp_scale_vec = cvt_vec_t(zp_scale_val);
-    }
-
-    for (; j < hidden_size - vec_elem_num; j += vec_elem_num) {
-      cvt_vec_t elems_fp32(input + i * hidden_size + j);
-      elems_fp32 = elems_fp32 * token_scale_vec;
-
-      if constexpr (AZP) {
-        azp_adj_load_vec_t azp_adj_vec(azp_adj + j);
-        cvt_vec_t azp_adj_fp32(azp_adj_vec);
-        azp_adj_fp32 = azp_adj_fp32 * token_zp_scale_vec;
-
-        if constexpr (PerChannel) {
-          cvt_vec_t b_scale_vec(b_scale + j);
-          azp_adj_fp32 = azp_adj_fp32 * b_scale_vec;
-        }
-
-        elems_fp32 = elems_fp32 - azp_adj_fp32;
-      }
-
-      if constexpr (Bias) {
-        load_vec_t bias_vec(bias + j);
-        cvt_vec_t bias_vec_fp32(bias_vec);
-        elems_fp32 = elems_fp32 + bias_vec_fp32;
-      }
-
-      load_vec_t elems_out(elems_fp32);
-      elems_out.save(output + i * hidden_size + j);
-    }
-
-    cvt_vec_t elems_fp32(input + i * hidden_size + j);
-    elems_fp32 = elems_fp32 * token_scale_vec;
-
-    if constexpr (AZP) {
-      azp_adj_load_vec_t azp_adj_vec(azp_adj + j);
-      cvt_vec_t azp_adj_fp32(azp_adj_vec);
-      azp_adj_fp32 = azp_adj_fp32 * token_zp_scale_vec;
-
-      if constexpr (PerChannel) {
-        cvt_vec_t b_scale_vec(b_scale + j);
-        azp_adj_fp32 = azp_adj_fp32 * b_scale_vec;
-      }
-
-      elems_fp32 = elems_fp32 - azp_adj_fp32;
-    }
-
-    if constexpr (Bias) {
-      load_vec_t bias_vec(bias + j);
-      cvt_vec_t bias_vec_fp32(bias_vec);
-      elems_fp32 = elems_fp32 + bias_vec_fp32;
-    }
-
-    load_vec_t elems_out(elems_fp32);
-    elems_out.save(output + i * hidden_size + j, hidden_size - j);
-  }
-}
-#elif defined(__powerpc64__)
-template <bool AZP, typename scalar_t>
-void static_scaled_int8_quant_impl(const scalar_t* input, int8_t* output,
-                                   const float* scale, const int32_t* azp,
-                                   const int num_tokens,
-                                   const int hidden_size) {
-  using load_vec_t = typename KernelVecType<scalar_t>::load_vec_type;
-  using cvt_vec_t = typename KernelVecType<scalar_t>::cvt_vec_type;
-  constexpr int vec_elem_num = load_vec_t::VEC_ELEM_NUM;
-
-  constexpr float i8_min =
-      static_cast<float>(std::numeric_limits<int8_t>::min());
-  constexpr float i8_max =
-      static_cast<float>(std::numeric_limits<int8_t>::max());
-
-  const cvt_vec_t inv_scale(1.0 / *scale);
-  const cvt_vec_t i8_min_vec(i8_min);
-  const cvt_vec_t i8_max_vec(i8_max);
-
-  cvt_vec_t zp_vec;
-  if constexpr (AZP) {
-    zp_vec = cvt_vec_t(static_cast<float>(*azp));
-  }
-  #pragma omp parallel for
-  for (int i = 0; i < num_tokens; ++i) {
-    int j = 0;
-    for (; j < hidden_size - vec_elem_num; j += vec_elem_num) {
-      load_vec_t elems(input + i * hidden_size + j);
-      cvt_vec_t elems_fp32(elems);
-      elems_fp32 = elems_fp32 * inv_scale;
-      if constexpr (AZP) {
-        elems_fp32 = elems_fp32 + zp_vec;
-      }
-      elems_fp32 = elems_fp32.clamp(i8_min_vec, i8_max_vec);
-      vec_op::INT8Vec16 elems_int8(elems_fp32);
-      elems_int8.save(output + i * hidden_size + j);
-    }
-    load_vec_t elems(input + i * hidden_size + j);
-    cvt_vec_t elems_fp32(elems);
-    elems_fp32 = elems_fp32 * inv_scale;
-
-    if constexpr (AZP) {
-      elems_fp32 = elems_fp32 + zp_vec;
-    }
-
-    elems_fp32 = elems_fp32.clamp(i8_min_vec, i8_max_vec);
-    vec_op::INT8Vec16 elems_int8(elems_fp32);
-    elems_int8.save(output + i * hidden_size + j, hidden_size - j);
-  }
-}
-template <bool AZP, typename scalar_t>
-void dynamic_scaled_int8_quant_impl(const scalar_t* input, int8_t* output,
-                                    float* scale, int32_t* azp,
-                                    const int num_tokens,
-                                    const int hidden_size) {
-  using load_vec_t = typename KernelVecType<scalar_t>::load_vec_type;
-  using cvt_vec_t = typename KernelVecType<scalar_t>::cvt_vec_type;
-  constexpr int vec_elem_num = load_vec_t::VEC_ELEM_NUM;
-
-  constexpr float i8_min =
-      static_cast<float>(std::numeric_limits<int8_t>::min());
-  constexpr float i8_max =
-      static_cast<float>(std::numeric_limits<int8_t>::max());
-  const cvt_vec_t i8_min_vec(i8_min);
-  const cvt_vec_t i8_max_vec(i8_max);
-
-  #pragma omp parallel for
-  for (int i = 0; i < num_tokens; ++i) {
-    cvt_vec_t max_value(std::numeric_limits<float>::lowest());
-    cvt_vec_t min_value(std::numeric_limits<float>::max());
-    {
-      int j = 0;
-      for (; j < hidden_size - vec_elem_num; j += vec_elem_num) {
-        load_vec_t elems(input + i * hidden_size + j);
-        cvt_vec_t elems_fp32(elems);
-        if constexpr (AZP) {
-          max_value = max_value.max(elems_fp32);
-          min_value = min_value.min(elems_fp32);
-        } else {
-          max_value = max_value.max(elems_fp32.abs());
-        }
-      }
-
-      load_vec_t elems(input + i * hidden_size + j);
-      cvt_vec_t elems_fp32(elems);
-
-      if (j + vec_elem_num == hidden_size) {
-        if constexpr (AZP) {
-          max_value = max_value.max(elems_fp32);
-          min_value = min_value.min(elems_fp32);
-        } else {
-          max_value = max_value.max(elems_fp32.abs());
-        }
-      } else {
-        if constexpr (AZP) {
-          max_value = max_value.max(elems_fp32, hidden_size - j);
-          min_value = min_value.min(elems_fp32, hidden_size - j);
-        } else {
-          max_value = max_value.max(elems_fp32.abs(), hidden_size - j);
-        }
-      }
-    }
-
-    float scale_val, azp_val;
-    if constexpr (AZP) {
-      float max_scalar = max_value.reduce_max();
-      float min_scalar = min_value.reduce_min();
-      scale_val = (max_scalar - min_scalar) / 255.0f;
-      azp_val = std::nearbyint(-128.0f - min_scalar / scale_val);
-      azp[i] = static_cast<int32_t>(azp_val);
-      scale[i] = scale_val;
-    } else {
-      scale_val = max_value.reduce_max() / 127.0f;
-      scale[i] = scale_val;
-    }
-
-    const cvt_vec_t inv_scale(1.0 / scale_val);
-    const cvt_vec_t azp_vec(azp_val);
-
-    {
-      int j = 0;
-      for (; j < hidden_size - vec_elem_num; j += vec_elem_num) {
-        load_vec_t elems(input + i * hidden_size + j);
-        cvt_vec_t elems_fp32(elems);
-        elems_fp32 = (elems_fp32 * inv_scale);
-
-        if constexpr (AZP) {
-          elems_fp32 = elems_fp32 + azp_vec;
-        }
-        elems_fp32 = elems_fp32.clamp(i8_min_vec, i8_max_vec);
-        vec_op::INT8Vec16 elems_int8(elems_fp32);
-        elems_int8.save(output + i * hidden_size + j);
-      }
-
-      load_vec_t elems(input + i * hidden_size + j);
-      cvt_vec_t elems_fp32(elems);
-      elems_fp32 = (elems_fp32 * inv_scale);
-
-      if constexpr (AZP) {
-        elems_fp32 = elems_fp32 + azp_vec;
-      }
-      elems_fp32 = elems_fp32.clamp(i8_min_vec, i8_max_vec);
-      vec_op::INT8Vec16 elems_int8(elems_fp32);
-      elems_int8.save(output + i * hidden_size + j, hidden_size - j);
-    }
-  }
-}
-template <bool PerChannel, typename scalar_t>
-void static_quant_epilogue(const float* input, scalar_t* output,
-                           const float a_scale, const float* b_scale,
-                           const int32_t* azp_with_adj, const int num_tokens,
-                           const int hidden_size) {
-  CPU_KERNEL_GUARD_IN(dynamic_output_scale_impl)
-  using load_vec_t = typename KernelVecType<scalar_t>::load_vec_type;
-  using azp_adj_load_vec_t =
-      typename KernelVecType<scalar_t>::azp_adj_load_vec_type;
-  using cvt_vec_t = typename KernelVecType<scalar_t>::cvt_vec_type;
-  constexpr int vec_elem_num = load_vec_t::VEC_ELEM_NUM;
-
-  #pragma omp parallel for
-  for (int i = 0; i < num_tokens; ++i) {
-    cvt_vec_t a_scale_vec(a_scale);
-    cvt_vec_t b_scale_vec(*b_scale);
-    cvt_vec_t scale_vec = a_scale_vec * b_scale_vec;
-
-    int j = 0;
-    for (; j < hidden_size - vec_elem_num; j += vec_elem_num) {
-      cvt_vec_t elems_fp32(input + i * hidden_size + j);
-      azp_adj_load_vec_t azp_adj_vec(azp_with_adj + j);
-      cvt_vec_t azp_adj_fp32(azp_adj_vec);
-
-      if constexpr (PerChannel) {
-        b_scale_vec = cvt_vec_t(b_scale + j);
-        scale_vec = b_scale_vec * a_scale_vec;
-      }
-      elems_fp32 = elems_fp32 - scale_vec * azp_adj_fp32;
-      load_vec_t elems_out(elems_fp32);
-      elems_out.save(output + i * hidden_size + j);
-    }
-
-    cvt_vec_t elems_fp32(input + i * hidden_size + j);
-    azp_adj_load_vec_t azp_adj_vec(azp_with_adj + j);
-    cvt_vec_t azp_adj_fp32(azp_adj_vec);
-
-    if constexpr (PerChannel) {
-      b_scale_vec = cvt_vec_t(b_scale + j);
-      scale_vec = b_scale_vec * a_scale_vec;
-    }
-
-    elems_fp32 = elems_fp32 - scale_vec * azp_adj_fp32;
-
-    load_vec_t elems_out(elems_fp32);
-    elems_out.save(output + i * hidden_size + j, hidden_size - j);
-  }
-}
-template <bool AZP, bool PerChannel, bool Bias, typename scalar_t>
-void dynamic_quant_epilogue(const float* input, scalar_t* output,
-                            const float* a_scale, const float* b_scale,
-                            const int32_t* azp, const int32_t* azp_adj,
-                            const scalar_t* bias, const int num_tokens,
-                            const int hidden_size) {
-  CPU_KERNEL_GUARD_IN(dynamic_quant_epilogue)
-  using load_vec_t = typename KernelVecType<scalar_t>::load_vec_type;
-  using azp_adj_load_vec_t =
-      typename KernelVecType<scalar_t>::azp_adj_load_vec_type;
-  using cvt_vec_t = typename KernelVecType<scalar_t>::cvt_vec_type;
-  constexpr int vec_elem_num = load_vec_t::VEC_ELEM_NUM;
-
-  #pragma omp parallel for
-  for (int i = 0; i < num_tokens; ++i) {
-    int j = 0;
-    cvt_vec_t token_scale_vec(a_scale[i]);
-    cvt_vec_t token_zp_scale_vec;
-    if constexpr (AZP) {
-      float zp_scale_val = a_scale[i] * static_cast<float>(azp[i]);
-      if constexpr (!PerChannel) {
-        zp_scale_val *= *b_scale;
-      }
-      token_zp_scale_vec = cvt_vec_t(zp_scale_val);
-    }
-
-    for (; j < hidden_size - vec_elem_num; j += vec_elem_num) {
-      cvt_vec_t elems_fp32(input + i * hidden_size + j);
-      elems_fp32 = elems_fp32 * token_scale_vec;
-
-      if constexpr (AZP) {
-        azp_adj_load_vec_t azp_adj_vec(azp_adj + j);
-        cvt_vec_t azp_adj_fp32(azp_adj_vec);
-        azp_adj_fp32 = azp_adj_fp32 * token_zp_scale_vec;
-
-        if constexpr (PerChannel) {
-          cvt_vec_t b_scale_vec(b_scale + j);
-          azp_adj_fp32 = azp_adj_fp32 * b_scale_vec;
-        }
-
-        elems_fp32 = elems_fp32 - azp_adj_fp32;
-      }
-
-      if constexpr (Bias) {
-        load_vec_t bias_vec(bias + j);
-        cvt_vec_t bias_vec_fp32(bias_vec);
-        elems_fp32 = elems_fp32 + bias_vec_fp32;
-      }
-
-      load_vec_t elems_out(elems_fp32);
-      elems_out.save(output + i * hidden_size + j);
-    }
-
-    cvt_vec_t elems_fp32(input + i * hidden_size + j);
-    elems_fp32 = elems_fp32 * token_scale_vec;
-
-    if constexpr (AZP) {
-      azp_adj_load_vec_t azp_adj_vec(azp_adj + j);
-      cvt_vec_t azp_adj_fp32(azp_adj_vec);
-      azp_adj_fp32 = azp_adj_fp32 * token_zp_scale_vec;
-
-      if constexpr (PerChannel) {
-        cvt_vec_t b_scale_vec(b_scale + j);
-        azp_adj_fp32 = azp_adj_fp32 * b_scale_vec;
-      }
-
-      elems_fp32 = elems_fp32 - azp_adj_fp32;
-    }
-
-    if constexpr (Bias) {
-      load_vec_t bias_vec(bias + j);
-      cvt_vec_t bias_vec_fp32(bias_vec);
-      elems_fp32 = elems_fp32 + bias_vec_fp32;
-    }
-
-    load_vec_t elems_out(elems_fp32);
-    elems_out.save(output + i * hidden_size + j, hidden_size - j);
-  }
-}
-#else
-template <typename scalar_t>
-void static_scaled_int8_quant_impl(const scalar_t* input, int8_t* output,
-                                   const float* scale, const int32_t* azp,
-                                   const int num_tokens,
-                                   const int hidden_size) {
-  TORCH_CHECK(false,
-              "static_scaled_int8_quant_impl requires AVX512/powerpc64/AArch64 "
-              "support.")
-}
-
-template <typename scalar_t>
-void dynamic_scaled_int8_quant_impl(const scalar_t* input, int8_t* output,
-                                    float* scale, int32_t* azp,
-                                    const int num_tokens,
-                                    const int hidden_size) {
-  TORCH_CHECK(false,
-              "dynamic_scaled_int8_quant_impl requires "
-              "AVX512/powerpc64/AArch64 support.")
-}
-
-template <bool PerChannel, typename scalar_t>
-void static_quant_epilogue(const float* input, scalar_t* output,
-                           const float a_scale, const float* b_scale,
-                           const int32_t* azp_with_adj, const int num_tokens,
-                           const int hidden_size) {
-  TORCH_CHECK(
-      false, "static_quant_epilogue requires AVX512/powerpc64/AArch64 support.")
-}
-
-template <typename scalar_t>
-void dynamic_quant_epilogue(const float* input, scalar_t* output,
-                            const float* a_scale, const float* b_scale,
-                            const int32_t* azp, const int32_t* azp_with_adj,
-                            const scalar_t* bias, const int num_tokens,
-                            const int hidden_size) {
-  TORCH_CHECK(
-      false,
-      "dynamic_quant_epilogue requires AVX512/powerpc64/AArch64 support.")
-}
-#endif
-}  // namespace
-
-void int8_scaled_mm(torch::Tensor& c,               // [M, OC], row-major
-                    const torch::Tensor& a,         // [M, IC], row-major
-                    const torch::Tensor& b,         // [IC, OC], column-major
-                    const torch::Tensor& a_scales,  // [1] or [M]
-                    const torch::Tensor& b_scales,  // [1] or [OC]
-                    const std::optional<torch::Tensor>& bias  // [OC]
-) {
-  CPU_KERNEL_GUARD_IN(cutlass_scaled_mm)
-  // Checks for conformality
-  TORCH_CHECK(a.dtype() == torch::kInt8 && b.dtype() == torch::kInt8,
-              "int8_scaled_mm only supports INT8 inputs.")
-  TORCH_CHECK(a.dim() == 2 && b.dim() == 2 && c.dim() == 2);
-  TORCH_CHECK(c.size(0) == a.size(0) && a.size(1) == b.size(0) &&
-              b.size(1) == c.size(1));
-  TORCH_CHECK(a_scales.numel() == 1 || a_scales.numel() == a.size(0));
-  TORCH_CHECK(b_scales.numel() == 1 || b_scales.numel() == b.size(1));
-
-  // Check for strides and alignment
-  TORCH_CHECK(a.stride(1) == 1 && c.stride(1) == 1);  // Row-major
-  TORCH_CHECK(b.stride(0) == 1);                      // Column-major
-  TORCH_CHECK(c.stride(0) % 16 == 0 &&
-              b.stride(1) % 16 == 0);  // 16 Byte Alignment
-  TORCH_CHECK(a_scales.is_contiguous() && b_scales.is_contiguous());
-
-  if (bias) {
-    TORCH_CHECK(bias->numel() == b.size(1) && bias->is_contiguous() &&
-                bias->dim() == 1);
-  }
-
-  VLLM_DISPATCH_FLOATING_TYPES(c.scalar_type(), "int8_scaled_mm", [&] {
-    if (a_scales.numel() != 1) {
-      // per-token
-      // Note: oneDNN doesn't support per-token activation quantization
-      // Ideally we want to fuse the GEMM and the scale procedure with oneDNN
-      // JIT, the intermediate data is cached in registers or L1. But for now
-      // the oneDNN GEMM code generation only supports two quantization
-      // patterns: per-tensor or per-output-channel of weight.
-      // So we have to apply the per-token scale with a 'epilogue'. In C=s_a *
-      // s_b * (A@B) + bias, the C_inter = s_b * (A@B) is computed by oneDNN
-      // GEMM, then the per-token scale (and bias) is applied with the epilogue
-      // C=s_a * C_inter + bias.
-      torch::Tensor tmp_fp32_out =
-          torch::empty_like(c, ::at::ScalarType::Float);
-      // Compute C_inter=s_b * (A@B)
-      DNNLPrimitiveHelper<true>::gemm_s8s8_jit<float, void>(
-          a.data_ptr<int8_t>(), b.data_ptr<int8_t>(),
-          tmp_fp32_out.data_ptr<float>(), nullptr, a.size(0), b.size(1),
-          a.size(1), nullptr, b_scales.data_ptr<float>(), 0, b_scales.numel());
-      if (bias.has_value()) {
-        // Compute C=s_a * C_inter + bias
-        dynamic_quant_epilogue<false, true, true>(
-            tmp_fp32_out.data_ptr<float>(), c.data_ptr<scalar_t>(),
-            a_scales.data_ptr<float>(), nullptr, nullptr, nullptr,
-            bias->data_ptr<scalar_t>(), c.size(0), c.size(1));
-      } else {
-        // Compute C=s_a * C_inter
-        dynamic_quant_epilogue<false, true, false, scalar_t>(
-            tmp_fp32_out.data_ptr<float>(), c.data_ptr<scalar_t>(),
-            a_scales.data_ptr<float>(), nullptr, nullptr, nullptr, nullptr,
-            c.size(0), c.size(1));
-      }
-    } else {
-      // per-tensor
-      if (bias.has_value()) {
-        // Compute C=s_a * s_b * (A@B) + bias
-        DNNLPrimitiveHelper<false>::gemm_s8s8_jit(
-            a.data_ptr<int8_t>(), b.data_ptr<int8_t>(), c.data_ptr<scalar_t>(),
-            bias->data_ptr<scalar_t>(), a.size(0), b.size(1), a.size(1),
-            a_scales.data_ptr<float>(), b_scales.data_ptr<float>(),
-            a_scales.numel(), b_scales.numel());
-      } else {
-        // Compute C=s_a * s_b * (A@B)
-        DNNLPrimitiveHelper<false>::gemm_s8s8_jit<scalar_t, void>(
-            a.data_ptr<int8_t>(), b.data_ptr<int8_t>(), c.data_ptr<scalar_t>(),
-            nullptr, a.size(0), b.size(1), a.size(1),
-            a_scales.data_ptr<float>(), b_scales.data_ptr<float>(),
-            a_scales.numel(), b_scales.numel());
-      }
-    }
-  });
-}
-
-void int8_scaled_mm_azp(torch::Tensor& c,        // [M, OC], row-major
-                        const torch::Tensor& a,  // [M, IC], row-major
-                        const torch::Tensor& b,  // [IC, OC], column-major
-                        const torch::Tensor& a_scales,            // [1] or [M]
-                        const torch::Tensor& b_scales,            // [1] or [OC]
-                        const torch::Tensor& azp_adj,             // [OC]
-                        const std::optional<torch::Tensor>& azp,  // [1] or [M]
-                        const std::optional<torch::Tensor>& bias  // [OC]
-) {
-  CPU_KERNEL_GUARD_IN(cutlass_scaled_mm_azp)
-  // Checks for conformality
-  TORCH_CHECK(a.dtype() == torch::kInt8 && b.dtype() == torch::kInt8,
-              "int8_scaled_mm_azp only supports INT8 inputs.")
-  TORCH_CHECK(a.dim() == 2 && b.dim() == 2 && c.dim() == 2);
-  TORCH_CHECK(c.size(0) == a.size(0) && a.size(1) == b.size(0) &&
-              b.size(1) == c.size(1));
-  TORCH_CHECK(a_scales.numel() == 1 || a_scales.numel() == a.size(0));
-  TORCH_CHECK(b_scales.numel() == 1 || b_scales.numel() == b.size(1));
-
-  // Check for strides and alignment
-  TORCH_CHECK(a.stride(1) == 1 && c.stride(1) == 1);  // Row-major
-  TORCH_CHECK(b.stride(0) == 1);                      // Column-major
-  TORCH_CHECK(c.stride(0) % 16 == 0 &&
-              b.stride(1) % 16 == 0);  // 16 Byte Alignment
-  TORCH_CHECK(a_scales.is_contiguous() && b_scales.is_contiguous());
-
-  if (bias) {
-    TORCH_CHECK(bias->numel() == b.size(1) && bias->is_contiguous());
-  }
-  if (azp) {
-    TORCH_CHECK(azp->numel() == a.size(0) && azp->is_contiguous());
-  }
-  TORCH_CHECK(azp_adj.numel() == b.size(1) && azp_adj.is_contiguous());
-
-  // azp & bias types
-  TORCH_CHECK(azp_adj.dtype() == torch::kInt32);
-  TORCH_CHECK(!azp || azp->dtype() == torch::kInt32);
-  TORCH_CHECK(!bias || bias->dtype() == c.dtype(),
-              "currently bias dtype must match output dtype ", c.dtype());
-
-  VLLM_DISPATCH_FLOATING_TYPES(c.scalar_type(), "int8_scaled_mm_azp", [&] {
-    torch::Tensor tmp_fp32_out = torch::empty_like(c, ::at::ScalarType::Float);
-    if (a_scales.numel() != 1) {
-      // per-token
-      // Note: oneDNN doesn't support per-token activation quantization
-      // Compute C_inter=s_b * (A@B)
-      DNNLPrimitiveHelper<true>::gemm_s8s8_jit<float, void>(
-          a.data_ptr<int8_t>(), b.data_ptr<int8_t>(),
-          tmp_fp32_out.data_ptr<float>(), nullptr, a.size(0), b.size(1),
-          a.size(1), nullptr, b_scales.data_ptr<float>(), 0, b_scales.numel());
-      if (bias.has_value()) {
-        // Compute C=s_a * C_inter - s_a * s_b * azp * azp_adj + bias
-        if (b_scales.numel() != 1) {
-          // Per-Channel
-          dynamic_quant_epilogue<true, true, true>(
-              tmp_fp32_out.data_ptr<float>(), c.data_ptr<scalar_t>(),
-              a_scales.data_ptr<float>(), b_scales.data_ptr<float>(),
-              azp->data_ptr<int32_t>(), azp_adj.data_ptr<int32_t>(),
-              bias->data_ptr<scalar_t>(), c.size(0), c.size(1));
-        } else {
-          // Per-Tensor
-          dynamic_quant_epilogue<true, false, true>(
-              tmp_fp32_out.data_ptr<float>(), c.data_ptr<scalar_t>(),
-              a_scales.data_ptr<float>(), b_scales.data_ptr<float>(),
-              azp->data_ptr<int32_t>(), azp_adj.data_ptr<int32_t>(),
-              bias->data_ptr<scalar_t>(), c.size(0), c.size(1));
-        }
-      } else {
-        // Compute C=s_a * C_inter - s_a * s_b * azp * azp_adj
-        if (b_scales.numel() != 1) {
-          // Per-Channel
-          dynamic_quant_epilogue<true, true, false, scalar_t>(
-              tmp_fp32_out.data_ptr<float>(), c.data_ptr<scalar_t>(),
-              a_scales.data_ptr<float>(), b_scales.data_ptr<float>(),
-              azp->data_ptr<int32_t>(), azp_adj.data_ptr<int32_t>(), nullptr,
-              c.size(0), c.size(1));
-        } else {
-          // Per-Tensor
-          dynamic_quant_epilogue<true, false, false, scalar_t>(
-              tmp_fp32_out.data_ptr<float>(), c.data_ptr<scalar_t>(),
-              a_scales.data_ptr<float>(), b_scales.data_ptr<float>(),
-              azp->data_ptr<int32_t>(), azp_adj.data_ptr<int32_t>(), nullptr,
-              c.size(0), c.size(1));
-        }
-      }
-    } else {
-      // per-tensor
-      if (bias.has_value()) {
-        // Compute C_inter=s_a * s_b * (A@B) + bias
-        DNNLPrimitiveHelper<false>::gemm_s8s8_jit(
-            a.data_ptr<int8_t>(), b.data_ptr<int8_t>(),
-            tmp_fp32_out.data_ptr<float>(), bias->data_ptr<scalar_t>(),
-            a.size(0), b.size(1), a.size(1), a_scales.data_ptr<float>(),
-            b_scales.data_ptr<float>(), a_scales.numel(), b_scales.numel());
-      } else {
-        // Compute C_inter=s_a * s_b * (A@B)
-        DNNLPrimitiveHelper<false>::gemm_s8s8_jit<float, void>(
-            a.data_ptr<int8_t>(), b.data_ptr<int8_t>(),
-            tmp_fp32_out.data_ptr<float>(), nullptr, a.size(0), b.size(1),
-            a.size(1), a_scales.data_ptr<float>(), b_scales.data_ptr<float>(),
-            a_scales.numel(), b_scales.numel());
-      }
-
-      // Compute C=C_inter - s_a * s_b * azp_adj
-      if (b_scales.numel() != 1) {
-        // Per-Channel
-        static_quant_epilogue<true>(
-            tmp_fp32_out.data_ptr<float>(), c.data_ptr<scalar_t>(),
-            *a_scales.data_ptr<float>(), b_scales.data_ptr<float>(),
-            azp_adj.data_ptr<int32_t>(), a.size(0), b.size(1));
-      } else {
-        // Per-Tensor
-        static_quant_epilogue<false>(
-            tmp_fp32_out.data_ptr<float>(), c.data_ptr<scalar_t>(),
-            *a_scales.data_ptr<float>(), b_scales.data_ptr<float>(),
-            azp_adj.data_ptr<int32_t>(), a.size(0), b.size(1));
-      }
-    }
-  });
-}
-
-// static-per-tensor quantization.
-void static_scaled_int8_quant(torch::Tensor& out,          // [..., hidden_size]
-                              const torch::Tensor& input,  // [..., hidden_size]
-                              const torch::Tensor& scale,
-                              std::optional<torch::Tensor> const& azp) {
-  CPU_KERNEL_GUARD_IN(static_scaled_int8_quant)
-  TORCH_CHECK(input.is_contiguous());
-  TORCH_CHECK(out.is_contiguous());
-  TORCH_CHECK(scale.numel() == 1);
-  TORCH_CHECK(!azp.has_value() || azp->numel() == 1);
-
-  const int hidden_size = input.size(-1);
-  const int num_tokens = input.numel() / hidden_size;
-  VLLM_DISPATCH_FLOATING_TYPES(
-      input.scalar_type(), "static_scaled_int8_quant_impl", [&] {
-        if (azp.has_value()) {
-          static_scaled_int8_quant_impl<true>(
-              input.data_ptr<scalar_t>(), out.data_ptr<int8_t>(),
-              scale.data_ptr<float>(), azp->data_ptr<int32_t>(), num_tokens,
-              hidden_size);
-        } else {
-          static_scaled_int8_quant_impl<false>(
-              input.data_ptr<scalar_t>(), out.data_ptr<int8_t>(),
-              scale.data_ptr<float>(), nullptr, num_tokens, hidden_size);
-        }
-      });
-}
-
-// dynamic-per-token quantization.
-void dynamic_scaled_int8_quant(
-    torch::Tensor& out,          // [..., hidden_size]
-    const torch::Tensor& input,  // [..., hidden_size]
-    torch::Tensor& scale,        // [..., 1]
-    std::optional<torch::Tensor> const& azp) {
-  CPU_KERNEL_GUARD_IN(dynamic_scaled_int8_quant)
-  TORCH_CHECK(input.is_contiguous());
-  TORCH_CHECK(out.is_contiguous());
-
-  int const hidden_size = input.size(-1);
-  int const num_tokens = input.numel() / hidden_size;
-  VLLM_DISPATCH_FLOATING_TYPES(
-      input.scalar_type(), "dynamic_scaled_int8_quant_impl", [&] {
-        if (azp.has_value()) {
-          dynamic_scaled_int8_quant_impl<true>(
-              input.data_ptr<scalar_t>(), out.data_ptr<int8_t>(),
-              scale.data_ptr<float>(), azp->data_ptr<int32_t>(), num_tokens,
-              hidden_size);
-        } else {
-          dynamic_scaled_int8_quant_impl<false>(
-              input.data_ptr<scalar_t>(), out.data_ptr<int8_t>(),
-              scale.data_ptr<float>(), nullptr, num_tokens, hidden_size);
-        }
-      });
-}
-
-#if defined(__powerpc64__)
-void int8_scaled_mm_ppc64le(torch::Tensor& c,        // [M, OC], row-major
-                            const torch::Tensor& a,  // [M, IC], row-major
-                            const torch::Tensor& b,  // [IC, OC], column-major
-                            const torch::Tensor& a_scales,
-                            const torch::Tensor& b_scales,
-                            const std::optional<torch::Tensor>& bias  // [OC]
-) {
-  CPU_KERNEL_GUARD_IN(cutlass_scaled_mm)
-  // Checks for conformality
-  TORCH_CHECK(a.dtype() == torch::kInt8 && b.dtype() == torch::kInt8,
-              "int8_scaled_mm_ppc64le only supports INT8 inputs.");
-  TORCH_CHECK(a.dim() == 2 && b.dim() == 2 && c.dim() == 2);
-  TORCH_CHECK(c.size(0) == a.size(0) && a.size(1) == b.size(0) &&
-              b.size(1) == c.size(1));
-  // We dont need this
-  TORCH_CHECK(a_scales.numel() == 1 || a_scales.numel() == a.size(0));
-  TORCH_CHECK(b_scales.numel() == 1 || b_scales.numel() == b.size(1));
-
-  // Check for strides and alignment
-  TORCH_CHECK(a.stride(1) == 1 && c.stride(1) == 1);  // Row-major
-  TORCH_CHECK(b.stride(0) == 1);                      // Column-major
-  TORCH_CHECK(c.stride(0) % 16 == 0 &&
-              b.stride(1) % 16 == 0);  // 16 Byte Alignment
-  TORCH_CHECK(a_scales.is_contiguous() && b_scales.is_contiguous());
-
-  if (bias) {
-    TORCH_CHECK(bias->numel() == b.size(1) && bias->is_contiguous() &&
-                bias->dim() == 1);
-  }
-  VLLM_DISPATCH_FLOATING_TYPES(c.scalar_type(), "int8_scaled_mm_ppc64le", [&] {
-    torch::Tensor tmp_fp32_out = torch::empty_like(c, ::at::ScalarType::Float);
-    // Compute C_inter=s_b * (A@B)
-    DNNLPrimitiveHelper<true>::gemm_s8s8_jit<float, void>(
-        a.data_ptr<int8_t>(), b.data_ptr<int8_t>(),
-        tmp_fp32_out.data_ptr<float>(), nullptr, a.size(0), b.size(1),
-        a.size(1), nullptr, b_scales.data_ptr<float>(), 0, b_scales.numel());
-    if (bias.has_value()) {
-      // Compute C=s_a * C_inter + bias
-      dynamic_quant_epilogue<false, true, true>(
-          tmp_fp32_out.data_ptr<float>(), c.data_ptr<scalar_t>(),
-          a_scales.data_ptr<float>(), nullptr, nullptr, nullptr,
-          bias->data_ptr<scalar_t>(), c.size(0), c.size(1));
-    } else {
-      // Compute C=s_a * C_inter
-      dynamic_quant_epilogue<false, true, false, scalar_t>(
-          tmp_fp32_out.data_ptr<float>(), c.data_ptr<scalar_t>(),
-          a_scales.data_ptr<float>(), nullptr, nullptr, nullptr, nullptr,
-          c.size(0), c.size(1));
-    }
-  });
-}
-
-#endif
--- a/csrc/cpu/torch_bindings.cpp
+++ b/csrc/cpu/torch_bindings.cpp
@ -6,25 +6,20 @@

 std::string init_cpu_threads_env(const std::string& cpu_ids);

-void int8_scaled_mm(torch::Tensor& c, const torch::Tensor& a,
-                    const torch::Tensor& b, const torch::Tensor& a_scales,
-                    const torch::Tensor& b_scales,
-                    const std::optional<torch::Tensor>& bias);
+void release_dnnl_matmul_handler(int64_t handler);

-void int8_scaled_mm_azp(torch::Tensor& c, const torch::Tensor& a,
-                        const torch::Tensor& b, const torch::Tensor& a_scales,
-                        const torch::Tensor& b_scales,
-                        const torch::Tensor& azp_adj,
-                        const std::optional<torch::Tensor>& azp,
-                        const std::optional<torch::Tensor>& bias);
+int64_t create_onednn_scaled_mm_handler(const torch::Tensor& b,
+                                        const torch::Tensor& b_scales,
+                                        at::ScalarType output_type,
+                                        bool dynamic_act_quant, bool use_azp,
+                                        int64_t primitive_cache_size);

-#if defined(__powerpc64__)
-void int8_scaled_mm_ppc64le(torch::Tensor& c, const torch::Tensor& a,
-                            const torch::Tensor& b,
-                            const torch::Tensor& a_scales,
-                            const torch::Tensor& b_scales,
-                            const std::optional<torch::Tensor>& bias);
-#endif
+void onednn_scaled_mm(torch::Tensor& c, const torch::Tensor& a,
+                      const torch::Tensor& a_scales,
+                      const std::optional<torch::Tensor>& azp,
+                      const std::optional<torch::Tensor>& azp_adj,
+                      const std::optional<torch::Tensor>& bias,
+                      int64_t handler);

 void mla_decode_kvcache(torch::Tensor& out, torch::Tensor& query,
                        torch::Tensor& kv_cache, double scale,
@ -151,8 +146,25 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
  ops.impl("rotary_embedding", torch::kCPU, &rotary_embedding);

  // Quantization
-#if defined(__AVX512F__) || (defined(__aarch64__) && !defined(__APPLE__))
+#if defined(__AVX512F__) || (defined(__aarch64__) && !defined(__APPLE__)) || \
+    defined(__powerpc64__)
  at::Tag stride_tag = at::Tag::needs_fixed_stride_order;
+  // Helper function to release oneDNN handlers
+  ops.def("release_dnnl_matmul_handler(int handler) -> ()",
+          &release_dnnl_matmul_handler);
+
+  // Create oneDNN W8A8 handler
+  ops.def(
+      "create_onednn_scaled_mm_handler(Tensor b, Tensor b_scales, ScalarType "
+      "output_type, bool dynamic_act_quant, bool use_azp, int "
+      "primitive_cache_size) -> int",
+      &create_onednn_scaled_mm_handler);
+
+  // oneDNN scaled_mm for W8A8 with static per-tensor activation quantization
+  ops.def(
+      "onednn_scaled_mm(Tensor! c, Tensor a, Tensor a_scales, Tensor? azp, "
+      "Tensor? azp_adj, Tensor? bias, int handler) -> ()");
+  ops.impl("onednn_scaled_mm", torch::kCPU, &onednn_scaled_mm);

  // Compute int8 quantized tensor for given scaling factor.
  ops.def(
@ -168,50 +180,6 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
      {stride_tag});
  ops.impl("dynamic_scaled_int8_quant", torch::kCPU,
           &dynamic_scaled_int8_quant);
-  // W8A8 GEMM, supporting symmetric per-tensor or per-row/column
-  // quantization.
-  ops.def(
-      "cutlass_scaled_mm(Tensor! out, Tensor a,"
-      "                  Tensor b, Tensor a_scales,"
-      "                  Tensor b_scales, Tensor? bias) -> ()",
-      {stride_tag});
-  ops.impl("cutlass_scaled_mm", torch::kCPU, &int8_scaled_mm);
-  // w8a8 GEMM, supporting asymmetric per-tensor or per-row/column
-  // quantization.
-  ops.def(
-      "cutlass_scaled_mm_azp(Tensor! out, Tensor a,"
-      "                  Tensor b, Tensor a_scales,"
-      "                  Tensor b_scales, Tensor azp_adj,"
-      "                  Tensor? azp, Tensor? bias) -> ()",
-      {stride_tag});
-  ops.impl("cutlass_scaled_mm_azp", torch::kCPU, &int8_scaled_mm_azp);
-#elif defined(__powerpc64__)
-  // Compute int8 quantized tensor for given scaling factor.
-  ops.def(
-      "static_scaled_int8_quant(Tensor! out, Tensor input, Tensor scale,"
-      "Tensor? azp) -> ()");
-  ops.impl("static_scaled_int8_quant", torch::kCPU, &static_scaled_int8_quant);
-
-  // Compute int8 quantized tensor and scaling factor
-  ops.def(
-      "dynamic_scaled_int8_quant(Tensor! out, Tensor input, Tensor! scale, "
-      "Tensor!? azp) -> ()");
-  ops.impl("dynamic_scaled_int8_quant", torch::kCPU,
-           &dynamic_scaled_int8_quant);
-  // W8A8 GEMM, supporting symmetric quantization.
-  ops.def(
-      "cutlass_scaled_mm(Tensor! out, Tensor a,"
-      "                  Tensor b, Tensor a_scales,"
-      "                  Tensor b_scales, Tensor? bias) -> ()");
-  ops.impl("cutlass_scaled_mm", torch::kCPU, &int8_scaled_mm_ppc64le);
-  // w8a8 GEMM, supporting asymmetric per-tensor or per-row/column
-  // quantization.
-  ops.def(
-      "cutlass_scaled_mm_azp(Tensor! out, Tensor a,"
-      "                  Tensor b, Tensor a_scales,"
-      "                  Tensor b_scales, Tensor azp_adj,"
-      "                  Tensor? azp, Tensor? bias) -> ()");
-  ops.impl("cutlass_scaled_mm_azp", torch::kCPU, &int8_scaled_mm_azp);
 #endif

 // SHM CCL
--- a/csrc/moe/grouped_topk_kernels.cu
+++ b/csrc/moe/grouped_topk_kernels.cu
@ -0,0 +1,757 @@
+/*
+ * Adapted from
+ * https://github.com/NVIDIA/TensorRT-LLM/blob/v0.21.0/cpp/tensorrt_llm/kernels/noAuxTcKernels.cu
+ * Copyright (c) 2025, The vLLM team.
+ * SPDX-FileCopyrightText: Copyright (c) 1993-2024 NVIDIA CORPORATION &
+ * AFFILIATES. All rights reserved. SPDX-License-Identifier: Apache-2.0
+ *
+ * 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 <c10/cuda/CUDAStream.h>
+#include <torch/all.h>
+#include <cuda_fp16.h>
+#include <cuda_bf16.h>
+#include <cooperative_groups.h>
+#include <cooperative_groups/reduce.h>
+namespace cg = cooperative_groups;
+
+namespace vllm {
+namespace moe {
+
+constexpr unsigned FULL_WARP_MASK = 0xffffffff;
+constexpr int32_t WARP_SIZE = 32;
+constexpr int32_t BLOCK_SIZE = 512;
+constexpr int32_t NUM_WARPS_PER_BLOCK = BLOCK_SIZE / WARP_SIZE;
+
+namespace warp_topk {
+
+template <int size, typename T>
+__host__ __device__ constexpr T round_up_to_multiple_of(T len) {
+  if (len == 0) {
+    return 0;
+  }
+  return ((len - 1) / size + 1) * size;
+}
+
+template <typename T>
+constexpr __host__ __device__ bool isPowerOf2(T v) {
+  return (v && !(v & (v - 1)));
+}
+
+template <bool greater, typename T>
+__forceinline__ __device__ bool is_better_than(T val, T baseline) {
+  return (val > baseline && greater) || (val < baseline && !greater);
+}
+
+template <bool greater, typename T, typename idxT>
+__forceinline__ __device__ bool is_better_than(T val, T baseline, idxT index,
+                                               idxT baseline_index) {
+  bool res = (val > baseline && greater) || (val < baseline && !greater);
+  if (val == baseline) {
+    res = (index < baseline_index && greater) ||
+          (index < baseline_index && !greater);
+  }
+  return res;
+}
+
+template <typename T, typename idxT>
+int calc_smem_size_for_block_wide(int num_of_warp, int64_t k) {
+  int64_t cache_topk = (sizeof(T) + sizeof(idxT)) * num_of_warp * k;
+  int64_t n = std::max<int>(num_of_warp / 2 * k, num_of_warp * WARP_SIZE);
+  return max(cache_topk,
+             round_up_to_multiple_of<256>(n * sizeof(T)) + n * sizeof(idxT));
+}
+
+template <int size, bool ascending, bool reverse, typename T, typename idxT,
+          bool is_stable>
+struct BitonicMerge {
+  // input should be a bitonic sequence, and sort it to be a monotonic sequence
+  __device__ static void merge(T* __restrict__ val_arr,
+                               idxT* __restrict__ idx_arr) {
+    static_assert(isPowerOf2(size));
+    static_assert(size >= 2 * WARP_SIZE);
+    constexpr int arr_len = size / WARP_SIZE;
+
+    constexpr int stride = arr_len / 2;
+    for (int i = 0; i < stride; ++i) {
+      int const other_i = i + stride;
+      T& val = val_arr[i];
+      T& other_val = val_arr[other_i];
+      bool is_better;
+      if constexpr (is_stable) {
+        is_better = is_better_than<ascending>(val, other_val, idx_arr[i],
+                                              idx_arr[other_i]);
+      } else {
+        is_better = is_better_than<ascending>(val, other_val);
+      }
+
+      if (is_better) {
+        T tmp = val;
+        val = other_val;
+        other_val = tmp;
+
+        idxT tmp2 = idx_arr[i];
+        idx_arr[i] = idx_arr[other_i];
+        idx_arr[other_i] = tmp2;
+      }
+    }
+
+    BitonicMerge<size / 2, ascending, reverse, T, idxT, is_stable>::merge(
+        val_arr, idx_arr);
+    BitonicMerge<size / 2, ascending, reverse, T, idxT, is_stable>::merge(
+        val_arr + arr_len / 2, idx_arr + arr_len / 2);
+  }
+};
+
+template <int size, bool ascending, typename T, typename idxT, bool is_stable>
+struct BitonicSort {
+  __device__ static void sort(T* __restrict__ val_arr,
+                              idxT* __restrict__ idx_arr) {
+    static_assert(isPowerOf2(size));
+    static_assert(size >= 2 * WARP_SIZE);
+    constexpr int arr_len = size / WARP_SIZE;
+
+    BitonicSort<size / 2, true, T, idxT, is_stable>::sort(val_arr, idx_arr);
+    BitonicSort<size / 2, false, T, idxT, is_stable>::sort(
+        val_arr + arr_len / 2, idx_arr + arr_len / 2);
+    BitonicMerge<size, ascending, ascending, T, idxT, is_stable>::merge(
+        val_arr, idx_arr);
+  }
+};
+
+template <bool ascending, typename T, typename idxT, bool is_stable>
+struct BitonicSort<32, ascending, T, idxT, is_stable> {
+  __device__ static void sort(T* __restrict__ val_arr,
+                              idxT* __restrict__ idx_arr) {
+    int const lane = threadIdx.x % WARP_SIZE;
+
+    // ascending doesn't matter before merging since all we need is a bitonic
+    // sequence
+    for (int stage = 0; stage < 4; ++stage) {
+      for (int stride = (1 << stage); stride > 0; stride /= 2) {
+        bool reverse = (lane >> stage) & 2;
+        bool is_second = lane & stride;
+
+        T other = __shfl_xor_sync(FULL_WARP_MASK, *val_arr, stride);
+        idxT other_idx = __shfl_xor_sync(FULL_WARP_MASK, *idx_arr, stride);
+
+        bool is_better;
+        if constexpr (is_stable) {
+          if constexpr (ascending) {
+            is_better = ((*val_arr > other) ||
+                         ((*val_arr == other) && (*idx_arr < other_idx))) !=
+                        (reverse != is_second);
+          } else {
+            is_better = ((*val_arr > other) ||
+                         ((*val_arr == other) && (*idx_arr > other_idx))) !=
+                        (reverse != is_second);
+          }
+        } else {
+          is_better = (*val_arr != other &&
+                       (*val_arr > other) != (reverse != is_second));
+        }
+        if (is_better) {
+          *val_arr = other;
+          *idx_arr = other_idx;
+        }
+      }
+    }
+
+    BitonicMerge<32, ascending, ascending, T, idxT, is_stable>::merge(val_arr,
+                                                                      idx_arr);
+  }
+};
+
+template <bool ascending, bool reverse, typename T, typename idxT,
+          bool is_stable>
+struct BitonicMerge<32, ascending, reverse, T, idxT, is_stable> {
+  __device__ static void merge(T* __restrict__ val_arr,
+                               idxT* __restrict__ idx_arr) {
+    int const lane = threadIdx.x % WARP_SIZE;
+    for (int stride = WARP_SIZE / 2; stride > 0; stride /= 2) {
+      bool is_second = lane & stride;
+      T& val = *val_arr;
+      T other = __shfl_xor_sync(FULL_WARP_MASK, val, stride);
+      idxT& idx = *idx_arr;
+      idxT other_idx = __shfl_xor_sync(FULL_WARP_MASK, idx, stride);
+
+      bool is_better;
+      if constexpr (is_stable) {
+        if constexpr (ascending) {
+          is_better = ((*val_arr > other) ||
+                       ((*val_arr == other) && (*idx_arr < other_idx))) ==
+                      (reverse != is_second);  // for min
+        } else {
+          is_better = ((*val_arr > other) ||
+                       ((*val_arr == other) && (*idx_arr > other_idx))) ==
+                      (reverse != is_second);  // for max
+        }
+      } else {
+        is_better =
+            (val != other && ((val > other) == (ascending != is_second)));
+      }
+
+      if (is_better) {
+        val = other;
+        idx = other_idx;
+      }
+    }
+  }
+};
+
+template <int capacity, bool greater, typename T, typename idxT, bool is_stable>
+class WarpSort {
+ public:
+  __device__ WarpSort(idxT k, T dummy)
+      : lane_(threadIdx.x % WARP_SIZE), k_(k), dummy_(dummy) {
+    static_assert(capacity >= WARP_SIZE && isPowerOf2(capacity));
+
+    for (int i = 0; i < max_arr_len_; ++i) {
+      val_arr_[i] = dummy_;
+      idx_arr_[i] = 0;
+    }
+  }
+
+  // load and merge k sorted values
+  __device__ void load_sorted(T const* __restrict__ in,
+                              idxT const* __restrict__ in_idx, idxT start) {
+    idxT idx = start + WARP_SIZE - 1 - lane_;
+    for (int i = max_arr_len_ - 1; i >= 0; --i, idx += WARP_SIZE) {
+      if (idx < start + k_) {
+        T t = in[idx];
+        bool is_better;
+        if constexpr (is_stable) {
+          is_better =
+              is_better_than<greater>(t, val_arr_[i], in_idx[idx], idx_arr_[i]);
+        } else {
+          is_better = is_better_than<greater>(t, val_arr_[i]);
+        }
+        if (is_better) {
+          val_arr_[i] = t;
+          idx_arr_[i] = in_idx[idx];
+        }
+      }
+    }
+
+    BitonicMerge<capacity, greater, !greater, T, idxT, is_stable>::merge(
+        val_arr_, idx_arr_);
+  }
+
+  __device__ void dump(T* __restrict__ out, idxT* __restrict__ out_idx) const {
+    for (int i = 0; i < max_arr_len_; ++i) {
+      idxT out_i = i * WARP_SIZE + lane_;
+      if (out_i < k_) {
+        out[out_i] = val_arr_[i];
+        out_idx[out_i] = idx_arr_[i];
+      }
+    }
+  }
+
+  __device__ void dumpIdx(idxT* __restrict__ out_idx) const {
+    for (int i = 0; i < max_arr_len_; ++i) {
+      idxT out_i = i * WARP_SIZE + lane_;
+      if (out_i < k_) {
+        out_idx[out_i] = idx_arr_[i];
+      }
+    }
+  }
+
+ protected:
+  static constexpr int max_arr_len_ = capacity / WARP_SIZE;
+
+  T val_arr_[max_arr_len_];
+  idxT idx_arr_[max_arr_len_];
+
+  int const lane_;
+  idxT const k_;
+  T const dummy_;
+
+};  // end class WarpSort
+
+template <int capacity, bool greater, typename T, typename idxT, bool is_stable>
+class WarpSelect : public WarpSort<capacity, greater, T, idxT, is_stable> {
+ public:
+  __device__ WarpSelect(idxT k, T dummy)
+      : WarpSort<capacity, greater, T, idxT, is_stable>(k, dummy),
+        k_th_(dummy),
+        k_th_lane_((k - 1) % WARP_SIZE) {
+    extern __shared__ char smem_buf[];  // extern __shared__ T smem_buf[];
+
+    int const num_of_warp = blockDim.x / WARP_SIZE;
+    int const warp_id = threadIdx.x / WARP_SIZE;
+    val_smem_ = reinterpret_cast<T*>(smem_buf);
+    val_smem_ += warp_id * WARP_SIZE;
+    idx_smem_ = reinterpret_cast<idxT*>(
+        smem_buf +
+        round_up_to_multiple_of<256>(num_of_warp * sizeof(T) * WARP_SIZE));
+    idx_smem_ += warp_id * WARP_SIZE;
+  }
+
+  __device__ void add(T const* in, idxT start, idxT end) {
+    idxT const end_for_fullwarp =
+        round_up_to_multiple_of<WARP_SIZE>(end - start) + start;
+    for (idxT i = start + lane_; i < end_for_fullwarp; i += WARP_SIZE) {
+      T val = (i < end) ? in[i] : dummy_;
+      add(val, i);
+    }
+  }
+
+  __device__ void add(T val, idxT idx) {
+    bool do_add;
+    if constexpr (is_stable) {
+      do_add = is_better_than<greater>(val, k_th_, idx, k_th_idx_);
+    } else {
+      do_add = is_better_than<greater>(val, k_th_);
+    }
+
+    uint32_t mask = __ballot_sync(FULL_WARP_MASK, do_add);
+    if (mask == 0) {
+      return;
+    }
+
+    int pos = smem_buf_len_ + __popc(mask & ((0x1u << lane_) - 1));
+    if (do_add && pos < WARP_SIZE) {
+      val_smem_[pos] = val;
+      idx_smem_[pos] = idx;
+      do_add = false;
+    }
+    smem_buf_len_ += __popc(mask);
+    if (smem_buf_len_ >= WARP_SIZE) {
+      __syncwarp();
+      merge_buf_(val_smem_[lane_], idx_smem_[lane_]);
+      smem_buf_len_ -= WARP_SIZE;
+    }
+    if (do_add) {
+      pos -= WARP_SIZE;
+      val_smem_[pos] = val;
+      idx_smem_[pos] = idx;
+    }
+    __syncwarp();
+  }
+
+  __device__ void done() {
+    if (smem_buf_len_) {
+      T val = (lane_ < smem_buf_len_) ? val_smem_[lane_] : dummy_;
+      idxT idx = (lane_ < smem_buf_len_) ? idx_smem_[lane_] : 0;
+      merge_buf_(val, idx);
+    }
+
+    // after done(), smem is used for merging results among warps
+    __syncthreads();
+  }
+
+ private:
+  __device__ void set_k_th_() {
+    k_th_ = __shfl_sync(FULL_WARP_MASK, val_arr_[max_arr_len_ - 1], k_th_lane_);
+    if constexpr (is_stable) {
+      k_th_idx_ =
+          __shfl_sync(FULL_WARP_MASK, idx_arr_[max_arr_len_ - 1], k_th_lane_);
+    }
+  }
+
+  __device__ void merge_buf_(T val, idxT idx) {
+    BitonicSort<WARP_SIZE, greater, T, idxT, is_stable>::sort(&val, &idx);
+
+    T& old = val_arr_[max_arr_len_ - 1];
+
+    bool is_better;
+    if constexpr (is_stable) {
+      is_better =
+          is_better_than<greater>(val, old, idx, idx_arr_[max_arr_len_ - 1]);
+    } else {
+      is_better = is_better_than<greater>(val, old);
+    }
+
+    if (is_better) {
+      old = val;
+      idx_arr_[max_arr_len_ - 1] = idx;
+    }
+
+    BitonicMerge<capacity, greater, !greater, T, idxT, is_stable>::merge(
+        val_arr_, idx_arr_);
+
+    set_k_th_();
+  }
+
+  using WarpSort<capacity, greater, T, idxT, is_stable>::max_arr_len_;
+  using WarpSort<capacity, greater, T, idxT, is_stable>::val_arr_;
+  using WarpSort<capacity, greater, T, idxT, is_stable>::idx_arr_;
+  using WarpSort<capacity, greater, T, idxT, is_stable>::lane_;
+  using WarpSort<capacity, greater, T, idxT, is_stable>::k_;
+  using WarpSort<capacity, greater, T, idxT, is_stable>::dummy_;
+
+  T* val_smem_;
+  idxT* idx_smem_;
+  int smem_buf_len_ = 0;
+
+  T k_th_;
+  idxT k_th_idx_;
+  int const k_th_lane_;
+};  // end class WarpSelect
+}  // namespace warp_topk
+
+template <typename T_OUT, typename T_IN>
+__device__ inline T_OUT cuda_cast(T_IN val) {
+  return val;
+}
+
+template <>
+__device__ inline float cuda_cast<float, __nv_bfloat16>(__nv_bfloat16 val) {
+  return __bfloat162float(val);
+}
+
+template <typename T>
+__device__ void topk_with_k2(T* output, T const* input,
+                             cg::thread_block_tile<32> const& tile,
+                             int32_t const lane_id,
+                             int const num_experts_per_group) {
+  // Get the top2 per thread
+  T largest = -INFINITY;
+  T second_largest = -INFINITY;
+
+  if (num_experts_per_group > WARP_SIZE) {
+    for (int i = lane_id; i < num_experts_per_group; i += WARP_SIZE) {
+      T value = input[i];
+      if (value > largest) {
+        second_largest = largest;
+        largest = value;
+      } else if (value > second_largest) {
+        second_largest = value;
+      }
+    }
+  } else {
+    for (int i = lane_id; i < num_experts_per_group; i += WARP_SIZE) {
+      largest = input[i];
+    }
+  }
+
+  __syncwarp();  // Ensure all threads have valid data before reduction
+  // Get the top2 warpwise
+  T max1 = cg::reduce(tile, largest, cg::greater<T>());
+
+  T max2 = max1;
+  bool equal_to_max1 = (max1 == largest);
+
+  int count_max1 = __popc(__ballot_sync(FULL_WARP_MASK, equal_to_max1));
+
+  if (count_max1 == 1) {
+    largest = (largest == max1) ? second_largest : largest;
+    max2 = cg::reduce(tile, largest, cg::greater<T>());
+  }
+
+  if (lane_id == 0) {
+    *output = max1 + max2;
+  }
+}
+
+template <typename T>
+__global__ void topk_with_k2_kernel(T* output, T* input,
+                                    int64_t const num_tokens,
+                                    int64_t const num_cases,
+                                    int64_t const n_group,
+                                    int64_t const num_experts_per_group) {
+  int32_t warp_id = threadIdx.x / WARP_SIZE;
+  int32_t lane_id = threadIdx.x % WARP_SIZE;
+
+  int32_t case_id = blockIdx.x * NUM_WARPS_PER_BLOCK + warp_id;
+  if (case_id < num_cases) {
+    input += case_id * num_experts_per_group;
+    output += case_id;
+
+    cg::thread_block block = cg::this_thread_block();
+    cg::thread_block_tile<32> tile = cg::tiled_partition<32>(block);
+
+#if (defined(__CUDA_ARCH__) && (__CUDA_ARCH__ >= 900))
+    asm volatile("griddepcontrol.wait;");
+#endif
+    topk_with_k2(output, input, tile, lane_id, num_experts_per_group);
+  }
+#if (defined(__CUDA_ARCH__) && (__CUDA_ARCH__ >= 900))
+  asm volatile("griddepcontrol.launch_dependents;");
+#endif
+}
+
+template <typename T, typename IdxT>
+__global__ void group_idx_and_topk_idx_kernel(
+    T* scores, T const* group_scores, T* topk_values, IdxT* topk_indices,
+    T* scores_with_bias, int64_t const num_tokens, int64_t const n_group,
+    int64_t const topk_group, int64_t const topk, int64_t const num_experts,
+    int64_t const num_experts_per_group, bool renormalize,
+    double routed_scaling_factor) {
+  int32_t warp_id = threadIdx.x / WARP_SIZE;
+  int32_t lane_id = threadIdx.x % WARP_SIZE;
+  int32_t case_id =
+      blockIdx.x * NUM_WARPS_PER_BLOCK + warp_id;  // one per token
+  scores_with_bias += case_id * num_experts;
+  scores += case_id * num_experts;
+  group_scores += case_id * n_group;
+  topk_values += case_id * topk;
+  topk_indices += case_id * topk;
+
+  int32_t align_num_experts_per_group =
+      warp_topk::round_up_to_multiple_of<WARP_SIZE>(num_experts_per_group);
+
+  cg::thread_block block = cg::this_thread_block();
+  cg::thread_block_tile<32> tile = cg::tiled_partition<32>(block);
+
+  extern __shared__ char smem_buf[];  // NOTE: reuse the shared memory here to
+                                      // store the target topk idx
+  int32_t* s_topk_idx = reinterpret_cast<int32_t*>(smem_buf);
+  T* s_topk_value =
+      reinterpret_cast<T*>(s_topk_idx + NUM_WARPS_PER_BLOCK * topk) +
+      warp_id * topk;
+  s_topk_idx += warp_id * topk;
+
+  T value = cuda::std::numeric_limits<T>::min();
+  T topk_group_value = cuda::std::numeric_limits<T>::min();
+  int32_t num_equalto_topkth_group;
+
+#if (defined(__CUDA_ARCH__) && (__CUDA_ARCH__ >= 900))
+  asm volatile("griddepcontrol.wait;");  // I think all prolog can be put before
+                                         // acqbulk because it's ptr arithmetic
+#endif
+
+  if (case_id < num_tokens) {
+    // calculate group_idx
+    int32_t target_num_min = WARP_SIZE - n_group + topk_group;
+    if (lane_id < n_group &&
+        (isfinite(cuda_cast<float, T>(
+            group_scores[lane_id]))))  // The check is necessary to avoid
+                                       // abnormal input
+    {
+      value = group_scores[lane_id];
+    }
+
+    int count_equal_to_top_value = WARP_SIZE - n_group;
+    int pre_count_equal_to_top_value = 0;
+    // Use loop to find the largset top_group
+    while (count_equal_to_top_value < target_num_min) {
+      __syncwarp();  // Ensure all threads have valid data before reduction
+      topk_group_value = cg::reduce(tile, value, cg::greater<T>());
+      if (value == topk_group_value) {
+        value = cuda::std::numeric_limits<T>::min();
+      }
+      pre_count_equal_to_top_value = count_equal_to_top_value;
+      count_equal_to_top_value = __popc(__ballot_sync(
+          FULL_WARP_MASK, (value == cuda::std::numeric_limits<T>::min())));
+    }
+    num_equalto_topkth_group = target_num_min - pre_count_equal_to_top_value;
+  }
+  __syncthreads();
+
+  warp_topk::WarpSelect</*capability*/ WARP_SIZE, /*greater*/ true, T, int32_t,
+                        /* is_stable */ true>
+      queue((int32_t)topk, -INFINITY);
+
+  int count_equalto_topkth_group = 0;
+  bool if_proceed_next_topk =
+      (topk_group_value != cuda::std::numeric_limits<T>::min());
+  if (case_id < num_tokens && if_proceed_next_topk) {
+    for (int i_group = 0; i_group < n_group; i_group++) {
+      if ((group_scores[i_group] > topk_group_value) ||
+          ((group_scores[i_group] == topk_group_value) &&
+           (count_equalto_topkth_group < num_equalto_topkth_group))) {
+        int32_t offset = i_group * num_experts_per_group;
+        for (int32_t i = lane_id; i < align_num_experts_per_group;
+             i += WARP_SIZE) {
+          T candidates =
+              (i < num_experts_per_group) && isfinite(cuda_cast<float, T>(
+                                                 scores_with_bias[offset + i]))
+                  ? scores_with_bias[offset + i]
+                  : cuda::std::numeric_limits<T>::min();
+          queue.add(candidates, offset + i);
+        }
+        if (group_scores[i_group] == topk_group_value) {
+          count_equalto_topkth_group++;
+        }
+      }
+    }
+    queue.done();
+    __syncwarp();
+    // Get the topk_idx
+    queue.dumpIdx(s_topk_idx);
+    __syncwarp();
+  }
+
+  // Load the valid score value
+  // Calculate the summation
+  float topk_sum = 1e-20;
+  if (case_id < num_tokens && if_proceed_next_topk) {
+    for (int i = lane_id;
+         i < warp_topk::round_up_to_multiple_of<WARP_SIZE>(topk);
+         i += WARP_SIZE) {
+      T value =
+          i < topk
+              ? scores[s_topk_idx[i]]
+              : cuda_cast<T, float>(0.0f);  // Load the valid value of expert
+      if (i < topk) {
+        s_topk_value[i] = value;
+      }
+      topk_sum += reduce(tile, cuda_cast<float, T>(value), cg::plus<float>());
+    }
+  }
+
+  __syncthreads();
+
+  if (case_id < num_tokens) {
+    if (if_proceed_next_topk) {
+      for (int i = lane_id; i < topk; i += WARP_SIZE) {
+        float value;
+        if (renormalize) {
+          value = cuda_cast<float, T>(s_topk_value[i]) / topk_sum *
+                  routed_scaling_factor;
+        } else {
+          value = cuda_cast<float, T>(s_topk_value[i]) * routed_scaling_factor;
+        }
+        topk_indices[i] = s_topk_idx[i];
+        topk_values[i] = cuda_cast<T, float>(value);
+      }
+    } else {
+      for (int i = lane_id; i < topk; i += WARP_SIZE) {
+        topk_indices[i] = i;
+        topk_values[i] = cuda_cast<T, float>(1.0f / topk);
+      }
+    }
+    // Note: when if_proceed_next_topk==false, choose the first 8 experts as the
+    // default result.
+  }
+#if (defined(__CUDA_ARCH__) && (__CUDA_ARCH__ >= 900))
+  asm volatile("griddepcontrol.launch_dependents;");
+#endif
+}
+
+template <typename T, typename IdxT>
+void invokeNoAuxTc(T* scores, T* group_scores, T* topk_values,
+                   IdxT* topk_indices, T* scores_with_bias,
+                   int64_t const num_tokens, int64_t const num_experts,
+                   int64_t const n_group, int64_t const topk_group,
+                   int64_t const topk, bool const renormalize,
+                   double const routed_scaling_factor, bool enable_pdl = false,
+                   cudaStream_t const stream = 0) {
+  int64_t num_cases = num_tokens * n_group;
+  int64_t topk_with_k2_num_blocks = (num_cases - 1) / NUM_WARPS_PER_BLOCK + 1;
+  auto* kernel_instance1 = &topk_with_k2_kernel<T>;
+  cudaLaunchConfig_t config;
+  config.gridDim = topk_with_k2_num_blocks;
+  config.blockDim = BLOCK_SIZE;
+  config.dynamicSmemBytes = 0;
+  config.stream = stream;
+  cudaLaunchAttribute attrs[1];
+  attrs[0].id = cudaLaunchAttributeProgrammaticStreamSerialization;
+  attrs[0].val.programmaticStreamSerializationAllowed = enable_pdl;
+  config.numAttrs = 1;
+  config.attrs = attrs;
+  cudaLaunchKernelEx(&config, kernel_instance1, group_scores, scores_with_bias,
+                     num_tokens, num_cases, n_group, num_experts / n_group);
+
+  int64_t topk_with_k_group_num_blocks =
+      (num_tokens - 1) / NUM_WARPS_PER_BLOCK + 1;
+  size_t dynamic_smem_in_bytes =
+      warp_topk::calc_smem_size_for_block_wide<T, int32_t>(NUM_WARPS_PER_BLOCK,
+                                                           topk);
+  auto* kernel_instance2 = &group_idx_and_topk_idx_kernel<T, IdxT>;
+  config.gridDim = topk_with_k_group_num_blocks;
+  config.blockDim = BLOCK_SIZE;
+  config.dynamicSmemBytes = dynamic_smem_in_bytes;
+  config.stream = stream;
+  attrs[0].id = cudaLaunchAttributeProgrammaticStreamSerialization;
+  attrs[0].val.programmaticStreamSerializationAllowed = enable_pdl;
+  config.numAttrs = 1;
+  config.attrs = attrs;
+  cudaLaunchKernelEx(&config, kernel_instance2, scores, group_scores,
+                     topk_values, topk_indices, scores_with_bias, num_tokens,
+                     n_group, topk_group, topk, num_experts,
+                     num_experts / n_group, renormalize, routed_scaling_factor);
+}
+
+#define INSTANTIATE_NOAUX_TC(T, IdxT)                                       \
+  template void invokeNoAuxTc<T, IdxT>(                                     \
+      T * scores, T * group_scores, T * topk_values, IdxT * topk_indices,   \
+      T * scores_with_bias, int64_t const num_tokens,                       \
+      int64_t const num_experts, int64_t const n_group,                     \
+      int64_t const topk_group, int64_t const topk, bool const renormalize, \
+      double const routed_scaling_factor, bool enable_pdl,                  \
+      cudaStream_t const stream);
+
+INSTANTIATE_NOAUX_TC(float, int32_t);
+INSTANTIATE_NOAUX_TC(half, int32_t);
+INSTANTIATE_NOAUX_TC(__nv_bfloat16, int32_t);
+}  // end namespace moe
+}  // namespace vllm
+
+std::tuple<torch::Tensor, torch::Tensor> grouped_topk(
+    torch::Tensor const& scores, torch::Tensor const& scores_with_bias,
+    int64_t n_group, int64_t topk_group, int64_t topk, bool renormalize,
+    double routed_scaling_factor) {
+  auto data_type = scores_with_bias.scalar_type();
+  auto input_size = scores_with_bias.sizes();
+  int64_t num_tokens = input_size[0];
+  int64_t num_experts = input_size[1];
+  TORCH_CHECK(input_size.size() == 2, "scores_with_bias must be a 2D Tensor");
+  TORCH_CHECK(num_experts % n_group == 0,
+              "num_experts should be divisible by n_group");
+  TORCH_CHECK(n_group <= 32,
+              "n_group should be smaller than or equal to 32 for now");
+  TORCH_CHECK(topk <= 32, "topk should be smaller than or equal to 32 for now");
+
+  torch::Tensor group_scores = torch::empty(
+      {num_tokens, n_group}, torch::dtype(data_type).device(torch::kCUDA));
+  torch::Tensor topk_values = torch::empty(
+      {num_tokens, topk}, torch::dtype(data_type).device(torch::kCUDA));
+  torch::Tensor topk_indices = torch::empty(
+      {num_tokens, topk}, torch::dtype(torch::kInt32).device(torch::kCUDA));
+
+  auto stream = c10::cuda::getCurrentCUDAStream(scores_with_bias.get_device());
+
+  switch (data_type) {
+    case torch::kFloat16:
+      // Handle Float16
+      vllm::moe::invokeNoAuxTc<half, int32_t>(
+          reinterpret_cast<half*>(scores.mutable_data_ptr()),
+          reinterpret_cast<half*>(group_scores.mutable_data_ptr()),
+          reinterpret_cast<half*>(topk_values.mutable_data_ptr()),
+          reinterpret_cast<int32_t*>(topk_indices.mutable_data_ptr()),
+          reinterpret_cast<half*>(scores_with_bias.data_ptr()), num_tokens,
+          num_experts, n_group, topk_group, topk, renormalize,
+          routed_scaling_factor, false, stream);
+      break;
+    case torch::kFloat32:
+      // Handle Float32
+      vllm::moe::invokeNoAuxTc<float, int32_t>(
+          reinterpret_cast<float*>(scores.mutable_data_ptr()),
+          reinterpret_cast<float*>(group_scores.mutable_data_ptr()),
+          reinterpret_cast<float*>(topk_values.mutable_data_ptr()),
+          reinterpret_cast<int32_t*>(topk_indices.mutable_data_ptr()),
+          reinterpret_cast<float*>(scores_with_bias.data_ptr()), num_tokens,
+          num_experts, n_group, topk_group, topk, renormalize,
+          routed_scaling_factor, false, stream);
+      break;
+    case torch::kBFloat16:
+      // Handle BFloat16
+      vllm::moe::invokeNoAuxTc<__nv_bfloat16, int32_t>(
+          reinterpret_cast<__nv_bfloat16*>(scores.mutable_data_ptr()),
+          reinterpret_cast<__nv_bfloat16*>(group_scores.mutable_data_ptr()),
+          reinterpret_cast<__nv_bfloat16*>(topk_values.mutable_data_ptr()),
+          reinterpret_cast<int32_t*>(topk_indices.mutable_data_ptr()),
+          reinterpret_cast<__nv_bfloat16*>(scores_with_bias.data_ptr()),
+          num_tokens, num_experts, n_group, topk_group, topk, renormalize,
+          routed_scaling_factor, false, stream);
+      break;
+    default:
+      // Handle other data types
+      throw std::invalid_argument(
+          "Invalid dtype, only supports float16, float32, and bfloat16");
+      break;
+  }
+  return {topk_values, topk_indices};
+}
--- a/csrc/moe/marlin_moe_wna16/generate_kernels.py
+++ b/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,
--- a/csrc/moe/marlin_moe_wna16/kernel.h
+++ b/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
--- a/csrc/moe/marlin_moe_wna16/marlin_template.h
+++ b/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;
--- a/csrc/moe/marlin_moe_wna16/ops.cu
+++ b/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");
--- a/csrc/moe/moe_ops.h
+++ b/csrc/moe/moe_ops.h
@ -22,6 +22,11 @@ torch::Tensor moe_wna16_gemm(torch::Tensor input, torch::Tensor output,
                             torch::Tensor num_tokens_post_pad, int64_t top_k,
                             int64_t BLOCK_SIZE_M, int64_t BLOCK_SIZE_N,
                             int64_t BLOCK_SIZE_K, int64_t bit);
+
+std::tuple<torch::Tensor, torch::Tensor> grouped_topk(
+    torch::Tensor const& scores, torch::Tensor const& scores_with_bias,
+    int64_t n_group, int64_t topk_group, int64_t topk, bool renormalize,
+    double routed_scaling_factor);
 #endif

 bool moe_permute_unpermute_supported();
--- a/csrc/moe/moe_permute_unpermute_op.cu
+++ b/csrc/moe/moe_permute_unpermute_op.cu
@ -45,8 +45,6 @@ void moe_permute(
  auto copy_topk_ids = topk_ids.clone();  // copy topk_ids for preprocess
  auto permuted_experts_id = torch::empty_like(topk_ids);
  auto sorted_row_idx = torch::empty_like(inv_permuted_idx);
-  auto align_expert_first_token_offset =
-      torch::zeros_like(expert_first_token_offset);

  CubKeyValueSorter sorter{};
  int64_t* valid_num_ptr = nullptr;
@ -85,12 +83,14 @@ void moe_permute(
  });

  // get m_indices and update expert_first_token_offset with align block
-  getMIndices(get_ptr<int64_t>(expert_first_token_offset),
-              get_ptr<int64_t>(align_expert_first_token_offset),
-              get_ptr<int>(m_indices), n_local_expert, align_block_size_value,
-              stream);
+  // this is only required for DeepGemm and not required for CUTLASS group gemm
  if (align_block_size.has_value()) {
-    // update align_expert_first_token_offset
+    auto align_expert_first_token_offset =
+        torch::zeros_like(expert_first_token_offset);
+    getMIndices(get_ptr<int64_t>(expert_first_token_offset),
+                get_ptr<int64_t>(align_expert_first_token_offset),
+                get_ptr<int>(m_indices), n_local_expert, align_block_size_value,
+                stream);
    expert_first_token_offset.copy_(align_expert_first_token_offset);
  }
 }
@ -195,19 +195,14 @@ void moe_permute(const torch::Tensor& input, const torch::Tensor& topk_weights,
                 torch::Tensor& expert_first_token_offset,
                 torch::Tensor& src_row_id2dst_row_id_map,
                 torch::Tensor& m_indices) {
-  TORCH_CHECK(false, "moe_unpermute is not supported on CUDA < 12.0");
+  TORCH_CHECK(false, "moe_permute is not supported on CUDA < 12.0");
 }

-void moe_unpermute(const torch::Tensor& input,
-                   const torch::Tensor& topk_weights, torch::Tensor& topk_ids,
-                   const torch::Tensor& token_expert_indices,
-                   const std::optional<torch::Tensor>& expert_map,
-                   int64_t n_expert, int64_t n_local_expert, int64_t topk,
-                   const std::optional<int64_t>& align_block_size,
-                   torch::Tensor& permuted_input,
-                   torch::Tensor& expert_first_token_offset,
-                   torch::Tensor& src_row_id2dst_row_id_map,
-                   torch::Tensor& m_indices) {
+void moe_unpermute(
+    const torch::Tensor& permuted_hidden_states,
+    const torch::Tensor& topk_weights, const torch::Tensor& inv_permuted_idx,
+    const std::optional<torch::Tensor>& expert_first_token_offset, int64_t topk,
+    torch::Tensor& hidden_states) {
  TORCH_CHECK(false, "moe_unpermute is not supported on CUDA < 12.0");
 }

@ -224,4 +219,4 @@ bool moe_permute_unpermute_supported() {
 TORCH_LIBRARY_IMPL_EXPAND(TORCH_EXTENSION_NAME, CUDA, m) {
  m.impl("moe_permute", &moe_permute);
  m.impl("moe_unpermute", &moe_unpermute);
-}
+}
--- a/csrc/moe/torch_bindings.cpp
+++ b/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,"
@ -77,6 +78,12 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, m) {
      "output_tensor) -> ()");
  m.impl("shuffle_rows", torch::kCUDA, &shuffle_rows);

+  // Apply grouped topk routing to select experts.
+  m.def(
+      "grouped_topk(Tensor scores, Tensor scores_with_bias, int n_group, int "
+      "topk_group, int topk, bool renormalize, float "
+      "routed_scaling_factor) -> (Tensor, Tensor)");
+  m.impl("grouped_topk", torch::kCUDA, &grouped_topk);
 #endif
 }

--- a/csrc/ops.h
+++ b/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,
@ -252,6 +229,11 @@ void get_cutlass_moe_mm_data(
    const int64_t num_experts, const int64_t n, const int64_t k,
    const std::optional<torch::Tensor>& blockscale_offsets);

+void get_cutlass_moe_mm_problem_sizes(
+    const torch::Tensor& topk_ids, torch::Tensor& problem_sizes1,
+    torch::Tensor& problem_sizes2, const int64_t num_experts, const int64_t n,
+    const int64_t k, const std::optional<torch::Tensor>& blockscale_offsets);
+
 void get_cutlass_pplx_moe_mm_data(torch::Tensor& expert_offsets,
                                  torch::Tensor& problem_sizes1,
                                  torch::Tensor& problem_sizes2,
--- a/csrc/prepare_inputs/advance_step.cu
+++ b/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);
-}
--- a/csrc/prepare_inputs/advance_step.cuh
+++ b/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
--- a/csrc/quantization/aqlm/gemm_kernels.cu
+++ b/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 {};
-}
--- a/csrc/quantization/cutlass_w4a8/w4a8_mm_entry.cu
+++ b/csrc/quantization/cutlass_w4a8/w4a8_mm_entry.cu
@ -0,0 +1,418 @@
+//
+// Based off of:
+//   https://github.com/NVIDIA/cutlass/blob/main/examples/55_hopper_mixed_dtype_gemm/55_hopper_int4_fp8_gemm.cu
+//
+
+#include <ATen/cuda/CUDAContext.h>
+#include <c10/cuda/CUDAGuard.h>
+#include <torch/all.h>
+#include "cutlass_extensions/torch_utils.hpp"
+
+#include "core/registration.h"
+
+#include "cutlass/cutlass.h"
+
+#include "cute/tensor.hpp"
+#include "cutlass/gemm/collective/collective_builder.hpp"
+#include "cutlass/epilogue/collective/collective_builder.hpp"
+#include "cutlass/gemm/device/gemm_universal_adapter.h"
+#include "cutlass/gemm/kernel/gemm_universal.hpp"
+
+#include "cutlass/util/packed_stride.hpp"
+#include "cutlass/util/mixed_dtype_utils.hpp"
+
+#include "cutlass_extensions/common.hpp"
+#include "cutlass_extensions/epilogue/scaled_mm_epilogues_c3x.hpp"
+
+namespace vllm::cutlass_w4a8 {
+
+using namespace cute;
+
+// -------------------------------------------------------------------------------------
+// Static configuration shared across all instantiations
+// -------------------------------------------------------------------------------------
+using MmaType = cutlass::float_e4m3_t;  // A/scale element type
+using QuantType = cutlass::int4b_t;     // B element type (packed int4)
+
+static int constexpr TileShapeK = 128 * 8 / sizeof_bits<MmaType>::value;
+static int constexpr ScalePackSize = 8;  // pack 8 scale elements together
+static int constexpr PackFactor = 8;     // 8 4-bit packed into int32
+
+// A matrix configuration
+using ElementA = MmaType;                   // Element type for A matrix operand
+using LayoutA = cutlass::layout::RowMajor;  // Layout type for A matrix operand
+using LayoutA_Transpose =
+    typename cutlass::layout::LayoutTranspose<LayoutA>::type;
+constexpr int AlignmentA =
+    128 / cutlass::sizeof_bits<
+              ElementA>::value;  // Memory access granularity/alignment of A
+                                 // matrix in units of elements (up to 16 bytes)
+using StrideA = cutlass::detail::TagToStrideA_t<LayoutA>;
+
+// B matrix configuration
+using ElementB = QuantType;  // Element type for B matrix operand
+using LayoutB =
+    cutlass::layout::ColumnMajor;  // Layout type for B matrix operand
+using LayoutB_Transpose =
+    typename cutlass::layout::LayoutTranspose<LayoutB>::type;
+constexpr int AlignmentB =
+    128 / cutlass::sizeof_bits<
+              ElementB>::value;  // Memory access granularity/alignment of B
+                                 // matrix in units of elements (up to 16 bytes)
+using StrideB = cutlass::detail::TagToStrideB_t<LayoutB>;
+
+// Define the CuTe layout for reordered quantized tensor B
+// LayoutAtomQuant places values that will be read by the same thread in
+// contiguous locations in global memory. It specifies the reordering within a
+// single warp's fragment
+using LayoutAtomQuant =
+    decltype(cutlass::compute_memory_reordering_atom<MmaType>());
+using LayoutB_Reordered = decltype(cute::tile_to_shape(
+    LayoutAtomQuant{}, Layout<Shape<int, int, int>, StrideB>{}));
+
+// Group-wise scales
+using ElementScale = MmaType;
+using LayoutScale = cutlass::layout::RowMajor;
+
+// Per-tok, per-chan scales
+using ElementSChannel = float;
+
+// C/D matrix configuration
+using ElementC =
+    cutlass::bfloat16_t;  // Element type for C and D matrix operands
+using LayoutC =
+    cutlass::layout::RowMajor;  // Layout type for C and D matrix operands
+constexpr int AlignmentC =
+    128 / cutlass::sizeof_bits<
+              ElementC>::value;  // Memory access granularity/alignment of C
+                                 // matrix in units of elements (up to 16 bytes)
+
+using ElementD = ElementC;
+using LayoutD = LayoutC;
+constexpr int AlignmentD = 128 / cutlass::sizeof_bits<ElementD>::value;
+
+// Core kernel configurations
+using ElementAccumulator = float;     // Element type for internal accumulation
+using ElementCompute = float;         // Element type for epilogue computation
+using ArchTag = cutlass::arch::Sm90;  // Tag indicating the minimum SM that
+                                      // supports the intended feature
+using OperatorClass = cutlass::arch::OpClassTensorOp;  // Operator class tag
+using KernelSchedule =
+    cutlass::gemm::KernelTmaWarpSpecializedCooperative;  // Kernel to launch
+                                                         // based on the default
+                                                         // setting in the
+                                                         // Collective Builder
+using EpilogueSchedule = cutlass::epilogue::TmaWarpSpecializedCooperative;
+using EpilogueTileType = cutlass::epilogue::collective::EpilogueTileAuto;
+
+// ----------------------------------------------------------------------------
+// Kernel template — Tile/Cluster shapes
+// ----------------------------------------------------------------------------
+template <class TileShape_MN, class ClusterShape_MNK>
+struct W4A8GemmKernel {
+  using TileShape =
+      decltype(cute::append(TileShape_MN{}, cute::Int<TileShapeK>{}));
+  using ClusterShape = ClusterShape_MNK;
+
+  // Epilogue per-tok, per-chan scales
+  using ChTokScalesEpilogue =
+      typename vllm::c3x::ScaledEpilogue<ElementAccumulator, ElementD,
+                                         TileShape>;
+  using EVTCompute = typename ChTokScalesEpilogue::EVTCompute;
+  using CollectiveEpilogue =
+      typename cutlass::epilogue::collective::CollectiveBuilder<
+          ArchTag, OperatorClass, TileShape, ClusterShape, EpilogueTileType,
+          ElementAccumulator, ElementSChannel,
+          // Transpose layout of D here since we use explicit swap + transpose
+          // the void type for C tells the builder to allocate 0 smem for the C
+          // matrix. We can enable this if beta == 0 by changing ElementC to
+          // void below.
+          ElementC, typename cutlass::layout::LayoutTranspose<LayoutC>::type,
+          AlignmentC, ElementD,
+          typename cutlass::layout::LayoutTranspose<LayoutD>::type, AlignmentD,
+          EpilogueSchedule,  // This is the only epi supporting the required
+                             // swap + transpose.
+          EVTCompute>::CollectiveOp;
+
+  // The Scale information must get paired with the operand that will be scaled.
+  // In this example, B is scaled so we make a tuple of B's information and the
+  // scale information.
+  using CollectiveMainloopShuffled =
+      typename cutlass::gemm::collective::CollectiveBuilder<
+          ArchTag, OperatorClass,
+          cute::tuple<ElementB, cutlass::Array<ElementScale, ScalePackSize>>,
+          LayoutB_Reordered, AlignmentB, ElementA, LayoutA_Transpose,
+          AlignmentA, ElementAccumulator, TileShape, ClusterShape,
+          cutlass::gemm::collective::StageCountAutoCarveout<static_cast<int>(
+              sizeof(typename CollectiveEpilogue::SharedStorage))>,
+          KernelSchedule>::CollectiveOp;
+
+  using GemmKernelShuffled = cutlass::gemm::kernel::GemmUniversal<
+      Shape<int, int, int, int>,  // Indicates ProblemShape
+      CollectiveMainloopShuffled, CollectiveEpilogue>;
+  using GemmShuffled =
+      cutlass::gemm::device::GemmUniversalAdapter<GemmKernelShuffled>;
+
+  using StrideC = typename GemmKernelShuffled::StrideC;
+  using StrideD = typename GemmKernelShuffled::StrideD;
+  using StrideS = typename CollectiveMainloopShuffled::StrideScale;
+
+  static torch::Tensor mm(torch::Tensor const& A,
+                          torch::Tensor const& B,             // already packed
+                          torch::Tensor const& group_scales,  // already packed
+                          int64_t group_size,
+                          torch::Tensor const& channel_scales,
+                          torch::Tensor const& token_scales,
+                          std::optional<at::ScalarType> const& maybe_out_type) {
+    // TODO: param validation
+    int m = A.size(0);
+    int k = A.size(1);
+    int n = B.size(1);
+
+    // Allocate output
+    const at::cuda::OptionalCUDAGuard device_guard(device_of(A));
+    auto device = A.device();
+    auto stream = at::cuda::getCurrentCUDAStream(device.index());
+    torch::Tensor D =
+        torch::empty({m, n}, torch::TensorOptions()
+                                 .dtype(equivalent_scalar_type_v<ElementD>)
+                                 .device(device));
+    // prepare arg pointers
+    auto A_ptr = static_cast<MmaType const*>(A.const_data_ptr());
+    auto B_ptr = static_cast<QuantType const*>(B.const_data_ptr());
+    auto D_ptr = static_cast<ElementD*>(D.data_ptr());
+    // can we avoid harcode the 8 here
+    auto S_ptr =
+        static_cast<cutlass::Array<ElementScale, ScalePackSize> const*>(
+            group_scales.const_data_ptr());
+
+    // runtime layout for B
+    auto shape_B = cute::make_shape(n, k, 1);
+    LayoutB_Reordered layout_B_reordered =
+        cute::tile_to_shape(LayoutAtomQuant{}, shape_B);
+
+    // strides
+    int const scale_k = cutlass::ceil_div(k, group_size);
+    StrideA stride_A =
+        cutlass::make_cute_packed_stride(StrideA{}, cute::make_shape(m, k, 1));
+    // Reverse stride here due to swap and transpose
+    StrideD stride_D =
+        cutlass::make_cute_packed_stride(StrideD{}, cute::make_shape(n, m, 1));
+    StrideS stride_S = cutlass::make_cute_packed_stride(
+        StrideS{}, cute::make_shape(n, scale_k, 1));
+
+    // Create a structure of gemm kernel arguments suitable for invoking an
+    // instance of Gemm auto arguments =
+    // args_from_options<GemmShuffled>(options);
+    /// Populates a Gemm::Arguments structure from the given arguments
+    /// Swap the A and B tensors, as well as problem shapes here.
+    using Args = typename GemmShuffled::Arguments;
+    using MainloopArguments = typename GemmKernelShuffled::MainloopArguments;
+    using EpilogueArguments = typename GemmKernelShuffled::EpilogueArguments;
+
+    MainloopArguments mainloop_arguments{
+        B_ptr, layout_B_reordered, A_ptr,     stride_A,
+        S_ptr, stride_S,           group_size};
+
+    EpilogueArguments epilogue_arguments{
+        ChTokScalesEpilogue::prepare_args(channel_scales, token_scales),
+        nullptr,
+        {},  // no C
+        D_ptr,
+        stride_D};
+
+    Args arguments{cutlass::gemm::GemmUniversalMode::kGemm,
+                   {n, m, k, 1},  // shape
+                   mainloop_arguments,
+                   epilogue_arguments};
+
+    // Workspace
+    size_t workspace_size = GemmShuffled::get_workspace_size(arguments);
+    torch::Tensor workspace =
+        torch::empty(workspace_size,
+                     torch::TensorOptions().dtype(torch::kU8).device(device));
+
+    // Run GEMM
+    GemmShuffled gemm;
+    CUTLASS_CHECK(gemm.can_implement(arguments));
+    CUTLASS_CHECK(gemm.initialize(arguments, workspace.data_ptr(), stream));
+    CUTLASS_CHECK(gemm.run(stream));
+
+    return D;
+  }
+};
+
+// ----------------------------------------------------------------------------
+// Kernel instantiations and dispatch logic
+// ----------------------------------------------------------------------------
+using Kernel_256x128_1x1x1 =
+    W4A8GemmKernel<Shape<_256, _128>, Shape<_1, _1, _1>>;
+using Kernel_256x64_1x1x1 = W4A8GemmKernel<Shape<_256, _64>, Shape<_1, _1, _1>>;
+using Kernel_256x32_1x1x1 = W4A8GemmKernel<Shape<_256, _32>, Shape<_1, _1, _1>>;
+using Kernel_256x16_1x1x1 = W4A8GemmKernel<Shape<_256, _16>, Shape<_1, _1, _1>>;
+using Kernel_128x256_2x1x1 =
+    W4A8GemmKernel<Shape<_128, _256>, Shape<_2, _1, _1>>;
+using Kernel_128x256_1x1x1 =
+    W4A8GemmKernel<Shape<_128, _256>, Shape<_1, _1, _1>>;
+using Kernel_128x128_1x1x1 =
+    W4A8GemmKernel<Shape<_128, _128>, Shape<_1, _1, _1>>;
+using Kernel_128x64_1x1x1 = W4A8GemmKernel<Shape<_128, _64>, Shape<_1, _1, _1>>;
+using Kernel_128x32_1x1x1 = W4A8GemmKernel<Shape<_128, _32>, Shape<_1, _1, _1>>;
+using Kernel_128x16_1x1x1 = W4A8GemmKernel<Shape<_128, _16>, Shape<_1, _1, _1>>;
+
+torch::Tensor mm_dispatch(torch::Tensor const& A,
+                          torch::Tensor const& B,             // already packed
+                          torch::Tensor const& group_scales,  // already packed
+                          int64_t group_size,
+                          torch::Tensor const& channel_scales,
+                          torch::Tensor const& token_scales,
+                          std::optional<at::ScalarType> const& maybe_out_type,
+                          const std::string& schedule) {
+  if (schedule == "256x128_1x1x1") {
+    return Kernel_256x128_1x1x1::mm(A, B, group_scales, group_size,
+                                    channel_scales, token_scales,
+                                    maybe_out_type);
+  } else if (schedule == "256x64_1x1x1") {
+    return Kernel_256x64_1x1x1::mm(A, B, group_scales, group_size,
+                                   channel_scales, token_scales,
+                                   maybe_out_type);
+  } else if (schedule == "256x32_1x1x1") {
+    return Kernel_256x32_1x1x1::mm(A, B, group_scales, group_size,
+                                   channel_scales, token_scales,
+                                   maybe_out_type);
+  } else if (schedule == "256x16_1x1x1") {
+    return Kernel_256x16_1x1x1::mm(A, B, group_scales, group_size,
+                                   channel_scales, token_scales,
+                                   maybe_out_type);
+  } else if (schedule == "128x256_2x1x1") {
+    return Kernel_128x256_2x1x1::mm(A, B, group_scales, group_size,
+                                    channel_scales, token_scales,
+                                    maybe_out_type);
+  } else if (schedule == "128x256_1x1x1") {
+    return Kernel_128x256_1x1x1::mm(A, B, group_scales, group_size,
+                                    channel_scales, token_scales,
+                                    maybe_out_type);
+  } else if (schedule == "128x128_1x1x1") {
+    return Kernel_128x128_1x1x1::mm(A, B, group_scales, group_size,
+                                    channel_scales, token_scales,
+                                    maybe_out_type);
+  } else if (schedule == "128x64_1x1x1") {
+    return Kernel_128x64_1x1x1::mm(A, B, group_scales, group_size,
+                                   channel_scales, token_scales,
+                                   maybe_out_type);
+  } else if (schedule == "128x32_1x1x1") {
+    return Kernel_128x32_1x1x1::mm(A, B, group_scales, group_size,
+                                   channel_scales, token_scales,
+                                   maybe_out_type);
+  } else if (schedule == "128x16_1x1x1") {
+    return Kernel_128x16_1x1x1::mm(A, B, group_scales, group_size,
+                                   channel_scales, token_scales,
+                                   maybe_out_type);
+  }
+  TORCH_CHECK(false, "Unknown W4A8 schedule: ", schedule);
+  return {};
+}
+
+torch::Tensor mm(torch::Tensor const& A,
+                 torch::Tensor const& B,             // already packed
+                 torch::Tensor const& group_scales,  // already packed
+                 int64_t group_size, torch::Tensor const& channel_scales,
+                 torch::Tensor const& token_scales,
+                 std::optional<at::ScalarType> const& maybe_out_type,
+                 std::optional<std::string> maybe_schedule) {
+  // requested a specific schedule
+  if (maybe_schedule) {
+    return mm_dispatch(A, B, group_scales, group_size, channel_scales,
+                       token_scales, maybe_out_type, *maybe_schedule);
+  }
+  std::string schedule;
+  int M = A.size(0);
+  int K = A.size(1);
+  int N = B.size(1);
+  // heuristic
+  if (M <= 16) {
+    schedule = (K == 16384 && N == 18432) ? "256x16_1x1x1" : "128x16_1x1x1";
+  } else if (M <= 32) {
+    schedule = (K == 16384 && N == 18432) ? "256x32_1x1x1" : "128x32_1x1x1";
+  } else if (M <= 64) {
+    if (K == 16384 && N == 18432)
+      schedule = "256x64_1x1x1";
+    else if (N <= 8192 && K <= 8192)
+      schedule = "128x32_1x1x1";
+    else
+      schedule = "128x64_1x1x1";
+  } else if (M <= 128) {
+    if (K == 16384 && N == 18432)
+      schedule = "256x128_1x1x1";
+    else if (N <= 8192)
+      schedule = "128x64_1x1x1";
+    else
+      schedule = "128x128_1x1x1";
+  } else if (M <= 256) {
+    if (N <= 4096)
+      schedule = "128x64_1x1x1";
+    else if (N <= 8192)
+      schedule = "128x128_1x1x1";
+    else
+      schedule = "128x256_1x1x1";
+  } else if (M <= 512 && N <= 4096) {
+    schedule = "128x128_1x1x1";
+  } else if (M <= 1024) {
+    schedule = "128x256_1x1x1";
+  } else {
+    schedule = "128x256_2x1x1";
+  }
+  return mm_dispatch(A, B, group_scales, group_size, channel_scales,
+                     token_scales, maybe_out_type, schedule);
+}
+
+// ----------------------------------------------------------------------------
+// Pre-processing utils
+// ----------------------------------------------------------------------------
+torch::Tensor pack_scale_fp8(torch::Tensor const& scales) {
+  TORCH_CHECK(scales.dtype() == torch::kFloat8_e4m3fn);
+  TORCH_CHECK(scales.is_contiguous());
+  TORCH_CHECK(scales.is_cuda());
+
+  auto packed_scales = torch::empty(
+      {scales.numel() * ScalePackSize},
+      torch::TensorOptions().dtype(scales.dtype()).device(scales.device()));
+  auto scales_ptr = static_cast<MmaType const*>(scales.const_data_ptr());
+  auto packed_scales_ptr =
+      static_cast<cutlass::Array<ElementScale, ScalePackSize>*>(
+          packed_scales.data_ptr());
+
+  cutlass::pack_scale_fp8(scales_ptr, packed_scales_ptr, scales.numel());
+
+  return packed_scales;
+}
+
+torch::Tensor encode_and_reorder_int4b(torch::Tensor const& B) {
+  TORCH_CHECK(B.dtype() == torch::kInt32);
+  TORCH_CHECK(B.dim() == 2);
+
+  torch::Tensor B_packed = torch::empty_like(B);
+
+  int k = B.size(0) * PackFactor;  // logical k
+  int n = B.size(1);
+
+  auto B_ptr = static_cast<QuantType const*>(B.const_data_ptr());
+  auto B_packed_ptr = static_cast<QuantType*>(B_packed.data_ptr());
+  auto shape_B = cute::make_shape(n, k, 1);
+  auto layout_B = make_layout(shape_B, LayoutRight{});  // row major
+  LayoutB_Reordered layout_B_reordered =
+      cute::tile_to_shape(LayoutAtomQuant{}, shape_B);
+
+  cutlass::unified_encode_int4b(B_ptr, B_packed_ptr, n * k);
+  cutlass::reorder_tensor(B_packed_ptr, layout_B, layout_B_reordered);
+
+  return B_packed;
+}
+
+TORCH_LIBRARY_IMPL_EXPAND(TORCH_EXTENSION_NAME, CUDA, m) {
+  m.impl("cutlass_w4a8_mm", &mm);
+  m.impl("cutlass_pack_scale_fp8", &pack_scale_fp8);
+  m.impl("cutlass_encode_and_reorder_int4b", &encode_and_reorder_int4b);
+}
+
+}  // namespace vllm::cutlass_w4a8
--- a/csrc/quantization/cutlass_w8a8/moe/get_group_starts.cuh
+++ b/csrc/quantization/cutlass_w8a8/moe/get_group_starts.cuh
@ -10,7 +10,7 @@

 template <typename ElementAB, typename ElementC, typename ElementAccumulator>
 __global__ void get_group_gemm_starts(
-    int32_t* expert_offsets, ElementAB** a_offsets, ElementAB** b_offsets,
+    int64_t* expert_offsets, ElementAB** a_offsets, ElementAB** b_offsets,
    ElementC** out_offsets, ElementAccumulator** a_scales_offsets,
    ElementAccumulator** b_scales_offsets, ElementAB* a_base_as_int,
    ElementAB* b_base_as_int, ElementC* out_base_as_int,
@ -34,7 +34,7 @@ __global__ void get_group_gemm_starts(
  else if (out_tensors.dtype() == TENSOR_C_TYPE) {                         \
    get_group_gemm_starts<cutlass::float_e4m3_t, C_TYPE, float>            \
        <<<1, num_experts, 0, stream>>>(                                   \
-            static_cast<int32_t*>(expert_offsets.data_ptr()),              \
+            static_cast<int64_t*>(expert_offsets.data_ptr()),              \
            static_cast<cutlass::float_e4m3_t**>(a_ptrs.data_ptr()),       \
            static_cast<cutlass::float_e4m3_t**>(b_ptrs.data_ptr()),       \
            static_cast<C_TYPE**>(out_ptrs.data_ptr()),                    \
@ -61,6 +61,8 @@ void run_get_group_gemm_starts(
  TORCH_CHECK(b_tensors.dtype() == torch::kFloat8_e4m3fn);
  TORCH_CHECK(a_scales.dtype() == torch::kFloat32);
  TORCH_CHECK(b_scales.dtype() == torch::kFloat32);
+  // expect int64_t to avoid overflow during offset calculations
+  TORCH_CHECK(expert_offsets.dtype() == torch::kInt64);

  int num_experts = static_cast<int>(expert_offsets.size(0));
  bool per_act_token = a_scales.numel() != 1;
--- a/csrc/quantization/cutlass_w8a8/moe/moe_data.cu
+++ b/csrc/quantization/cutlass_w8a8/moe/moe_data.cu
@ -104,6 +104,53 @@ __global__ void compute_arg_sorts(const int32_t* __restrict__ topk_ids,
  }
 }

+namespace {
+inline void launch_compute_problem_sizes(const torch::Tensor& topk_ids,
+                                         torch::Tensor& problem_sizes1,
+                                         torch::Tensor& problem_sizes2,
+                                         torch::Tensor& atomic_buffer,
+                                         int64_t num_experts, int64_t n,
+                                         int64_t k, cudaStream_t stream,
+                                         const bool swap_ab) {
+  int num_threads = min(THREADS_PER_EXPERT, topk_ids.numel());
+
+  const int32_t* topk_ptr = static_cast<const int32_t*>(topk_ids.data_ptr());
+  int32_t* ps1_ptr = static_cast<int32_t*>(problem_sizes1.data_ptr());
+  int32_t* ps2_ptr = static_cast<int32_t*>(problem_sizes2.data_ptr());
+  int32_t* atomic_ptr = static_cast<int32_t*>(atomic_buffer.data_ptr());
+
+  if (swap_ab) {
+    compute_problem_sizes<true><<<num_experts, num_threads, 0, stream>>>(
+        topk_ptr, ps1_ptr, ps2_ptr, atomic_ptr,
+        static_cast<int>(topk_ids.numel()), static_cast<int>(n),
+        static_cast<int>(k));
+  } else {
+    compute_problem_sizes<false><<<num_experts, num_threads, 0, stream>>>(
+        topk_ptr, ps1_ptr, ps2_ptr, atomic_ptr,
+        static_cast<int>(topk_ids.numel()), static_cast<int>(n),
+        static_cast<int>(k));
+  }
+}
+}  // namespace
+
+void get_cutlass_moe_mm_problem_sizes_caller(
+    const torch::Tensor& topk_ids, torch::Tensor& problem_sizes1,
+    torch::Tensor& problem_sizes2, const int64_t num_experts, const int64_t n,
+    const int64_t k, const std::optional<torch::Tensor>& blockscale_offsets) {
+  auto stream = at::cuda::getCurrentCUDAStream(topk_ids.device().index());
+  auto options_int32 =
+      torch::TensorOptions().dtype(torch::kInt32).device(topk_ids.device());
+  torch::Tensor atomic_buffer = torch::zeros(num_experts, options_int32);
+
+  // Swap-AB should be disabled for FP4 path
+  bool may_swap_ab = (!blockscale_offsets.has_value()) &&
+                     (topk_ids.numel() <= SWAP_AB_THRESHOLD);
+
+  launch_compute_problem_sizes(topk_ids, problem_sizes1, problem_sizes2,
+                               atomic_buffer, num_experts, n, k, stream,
+                               may_swap_ab);
+}
+
 void get_cutlass_moe_mm_data_caller(
    const torch::Tensor& topk_ids, torch::Tensor& expert_offsets,
    torch::Tensor& problem_sizes1, torch::Tensor& problem_sizes2,
@ -121,21 +168,9 @@ void get_cutlass_moe_mm_data_caller(
  bool may_swap_ab = (!blockscale_offsets.has_value()) &&
                     (topk_ids.numel() <= SWAP_AB_THRESHOLD);

-  if (may_swap_ab) {
-    compute_problem_sizes<true><<<num_experts, num_threads, 0, stream>>>(
-        static_cast<const int32_t*>(topk_ids.data_ptr()),
-        static_cast<int32_t*>(problem_sizes1.data_ptr()),
-        static_cast<int32_t*>(problem_sizes2.data_ptr()),
-        static_cast<int32_t*>(atomic_buffer.data_ptr()), topk_ids.numel(), n,
-        k);
-  } else {
-    compute_problem_sizes<false><<<num_experts, num_threads, 0, stream>>>(
-        static_cast<const int32_t*>(topk_ids.data_ptr()),
-        static_cast<int32_t*>(problem_sizes1.data_ptr()),
-        static_cast<int32_t*>(problem_sizes2.data_ptr()),
-        static_cast<int32_t*>(atomic_buffer.data_ptr()), topk_ids.numel(), n,
-        k);
-  }
+  launch_compute_problem_sizes(topk_ids, problem_sizes1, problem_sizes2,
+                               atomic_buffer, num_experts, n, k, stream,
+                               may_swap_ab);

  if (blockscale_offsets.has_value()) {
    // fp4 path
@ -161,6 +196,7 @@ void get_cutlass_moe_mm_data_caller(
      topk_ids.size(1));
 }

+template <bool SWAP_AB>
 __global__ void compute_pplx_data(int32_t* expert_offsets,
                                  int32_t* problem_sizes1,
                                  int32_t* problem_sizes2,
@ -168,14 +204,23 @@ __global__ void compute_pplx_data(int32_t* expert_offsets,
                                  const int padded_m, const int n,
                                  const int k) {
  int expert_idx = threadIdx.x;
-
  expert_offsets[expert_idx] = expert_idx * padded_m;
-  problem_sizes1[expert_idx * 3] = expert_num_tokens[expert_idx];
-  problem_sizes1[expert_idx * 3 + 1] = 2 * n;
-  problem_sizes1[expert_idx * 3 + 2] = k;
-  problem_sizes2[expert_idx * 3] = expert_num_tokens[expert_idx];
-  problem_sizes2[expert_idx * 3 + 1] = k;
-  problem_sizes2[expert_idx * 3 + 2] = n;
+
+  if constexpr (!SWAP_AB) {
+    problem_sizes1[expert_idx * 3] = expert_num_tokens[expert_idx];
+    problem_sizes1[expert_idx * 3 + 1] = 2 * n;
+    problem_sizes1[expert_idx * 3 + 2] = k;
+    problem_sizes2[expert_idx * 3] = expert_num_tokens[expert_idx];
+    problem_sizes2[expert_idx * 3 + 1] = k;
+    problem_sizes2[expert_idx * 3 + 2] = n;
+  } else {
+    problem_sizes1[expert_idx * 3] = 2 * n;
+    problem_sizes1[expert_idx * 3 + 1] = expert_num_tokens[expert_idx];
+    problem_sizes1[expert_idx * 3 + 2] = k;
+    problem_sizes2[expert_idx * 3] = k;
+    problem_sizes2[expert_idx * 3 + 1] = expert_num_tokens[expert_idx];
+    problem_sizes2[expert_idx * 3 + 2] = n;
+  }
 }

 void get_cutlass_pplx_moe_mm_data_caller(torch::Tensor& expert_offsets,
@ -187,10 +232,19 @@ void get_cutlass_pplx_moe_mm_data_caller(torch::Tensor& expert_offsets,
                                         const int64_t n, const int64_t k) {
  auto stream = at::cuda::getCurrentCUDAStream(expert_offsets.device().index());

-  compute_pplx_data<<<1, num_local_experts, 0, stream>>>(
-      static_cast<int32_t*>(expert_offsets.data_ptr()),
-      static_cast<int32_t*>(problem_sizes1.data_ptr()),
-      static_cast<int32_t*>(problem_sizes2.data_ptr()),
-      static_cast<const int32_t*>(expert_num_tokens.data_ptr()), padded_m, n,
-      k);
+  if (num_local_experts * padded_m > SWAP_AB_THRESHOLD) {
+    compute_pplx_data<false><<<1, num_local_experts, 0, stream>>>(
+        static_cast<int32_t*>(expert_offsets.data_ptr()),
+        static_cast<int32_t*>(problem_sizes1.data_ptr()),
+        static_cast<int32_t*>(problem_sizes2.data_ptr()),
+        static_cast<const int32_t*>(expert_num_tokens.data_ptr()), padded_m, n,
+        k);
+  } else {
+    compute_pplx_data<true><<<1, num_local_experts, 0, stream>>>(
+        static_cast<int32_t*>(expert_offsets.data_ptr()),
+        static_cast<int32_t*>(problem_sizes1.data_ptr()),
+        static_cast<int32_t*>(problem_sizes2.data_ptr()),
+        static_cast<const int32_t*>(expert_num_tokens.data_ptr()), padded_m, n,
+        k);
+  }
 }
--- a/csrc/quantization/cutlass_w8a8/scaled_mm_entry.cu
+++ b/csrc/quantization/cutlass_w8a8/scaled_mm_entry.cu
@ -76,6 +76,11 @@ void get_cutlass_moe_mm_data_caller(
    const int64_t num_experts, const int64_t n, const int64_t k,
    const std::optional<torch::Tensor>& blockscale_offsets);

+void get_cutlass_moe_mm_problem_sizes_caller(
+    const torch::Tensor& topk_ids, torch::Tensor& problem_sizes1,
+    torch::Tensor& problem_sizes2, const int64_t num_experts, const int64_t n,
+    const int64_t k, const std::optional<torch::Tensor>& blockscale_offsets);
+
 void get_cutlass_pplx_moe_mm_data_caller(torch::Tensor& expert_offsets,
                                         torch::Tensor& problem_sizes1,
                                         torch::Tensor& problem_sizes2,
@ -293,6 +298,25 @@ void get_cutlass_moe_mm_data(
      version_num, ". Required capability: 90 or 100");
 }

+void get_cutlass_moe_mm_problem_sizes(
+    const torch::Tensor& topk_ids, torch::Tensor& problem_sizes1,
+    torch::Tensor& problem_sizes2, const int64_t num_experts, const int64_t n,
+    const int64_t k, const std::optional<torch::Tensor>& blockscale_offsets) {
+  int32_t version_num = get_sm_version_num();
+#if (defined ENABLE_CUTLASS_MOE_SM90 && ENABLE_CUTLASS_MOE_SM90) || \
+    (defined ENABLE_CUTLASS_MOE_SM100 && ENABLE_CUTLASS_MOE_SM100)
+  get_cutlass_moe_mm_problem_sizes_caller(topk_ids, problem_sizes1,
+                                          problem_sizes2, num_experts, n, k,
+                                          blockscale_offsets);
+  return;
+#endif
+  TORCH_CHECK_NOT_IMPLEMENTED(
+      false,
+      "No compiled get_cutlass_moe_mm_problem_sizes: no cutlass_scaled_mm "
+      "kernel for CUDA device capability: ",
+      version_num, ". Required capability: 90 or 100");
+}
+
 void get_cutlass_pplx_moe_mm_data(torch::Tensor& expert_offsets,
                                  torch::Tensor& problem_sizes1,
                                  torch::Tensor& problem_sizes2,
--- a/csrc/quantization/gptq_marlin/dequant.h
+++ b/csrc/quantization/gptq_marlin/dequant.h
@ -470,11 +470,12 @@ __device__ inline void dequant<nv_bfloat162, vllm::kFE2M1f.id(), false>(
  frag_b[0] = __hmul2(frag_b[0], bias_reg);
 }

-template <typename scalar_t2>
+template <typename scalar_t2, vllm::ScalarTypeId s_type_id>
 __device__ inline void dequant_fp8_scales(int q, scalar_t2* frag_b);

 template <>
-__device__ inline void dequant_fp8_scales<half2>(int q, half2* frag_b) {
+__device__ inline void dequant_fp8_scales<half2, vllm::kFE4M3fn.id()>(
+    int q, half2* frag_b) {
  int Out1 = (q & 0xFF00FF00) >> 1;
  ;
  q <<= 8;
@ -486,8 +487,8 @@ __device__ inline void dequant_fp8_scales<half2>(int q, half2* frag_b) {
 };

 template <>
-__device__ inline void dequant_fp8_scales<nv_bfloat162>(int q,
-                                                        nv_bfloat162* frag_b) {
+__device__ inline void dequant_fp8_scales<nv_bfloat162, vllm::kFE4M3fn.id()>(
+    int q, nv_bfloat162* frag_b) {
  constexpr int FP8_EXPONENT = 4, BF16_EXPONENT = 8;
  constexpr int RIGHT_SHIFT = BF16_EXPONENT - FP8_EXPONENT;
  constexpr int MASK = 0x7F007F00;
@ -502,6 +503,20 @@ __device__ inline void dequant_fp8_scales<nv_bfloat162>(int q,
  frag_b[0] = *reinterpret_cast<const nv_bfloat162*>(&Out2);
 }

+template <>
+__device__ inline void dequant_fp8_scales<nv_bfloat162, vllm::kFE8M0fnu.id()>(
+    int q, nv_bfloat162* frag_b) {
+  // In this conversion, 2 ** -127 in FP8E8M0 would become 0 in BF16,
+  // but we assume that such a extreme value would not occur in real models.
+  int Out1 = (q & 0xFF00FF00) >> 1;
+  q <<= 7;
+  int Out2 = q & 0x7F807F80;
+
+  // Note: reverse indexing is intentional because weights are permuted
+  frag_b[1] = *reinterpret_cast<const nv_bfloat162*>(&Out1);
+  frag_b[0] = *reinterpret_cast<const nv_bfloat162*>(&Out2);
+}
+
 #endif

 }  // namespace MARLIN_NAMESPACE_NAME
--- a/csrc/quantization/gptq_marlin/generate_kernels.py
+++ b/csrc/quantization/gptq_marlin/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}}, "
@ -78,7 +79,8 @@ def generate_new_kernels():
            if scalar_type == "vllm::kFE4M3fn" and group_blocks not in [-1, 8]:
                continue
            # nvfp4 only supports group_size == 16
-            if scalar_type == "vllm::kFE2M1f" and group_blocks != 1:
+            # 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
            if scalar_type != "vllm::kFE2M1f" and group_blocks == 1:
@ -97,10 +99,23 @@ def generate_new_kernels():
                # 4bit quantization and fp16
                is_zp_float_list.append(True)

+            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"
+
            for is_zp_float in is_zp_float_list:
                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,
--- a/csrc/quantization/gptq_marlin/gptq_marlin.cu
+++ b/csrc/quantization/gptq_marlin/gptq_marlin.cu
@ -48,7 +48,8 @@ __global__ void permute_cols_kernel(int4 const* __restrict__ a_int4_ptr,

 torch::Tensor gptq_marlin_gemm(
    torch::Tensor& a, std::optional<torch::Tensor> 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& b_zeros_or_none,
    std::optional<torch::Tensor> const& g_idx_or_none,
    std::optional<torch::Tensor> const& perm_or_none, torch::Tensor& workspace,
@ -187,7 +188,12 @@ int get_kernel_cache_size(thread_config_t const& th_config, int thread_m_blocks,
  int tb_m = thread_m_blocks * 16;
  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);
+  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);
@ -202,8 +208,8 @@ int get_kernel_cache_size(thread_config_t const& th_config, int thread_m_blocks,
      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;
+  int total_size =
+      tmp_size + sh_a_size + sh_s_size + sh_zp_size + sh_g_idx_size;

  return total_size;
 }
@ -237,20 +243,25 @@ bool is_valid_config(thread_config_t const& th_config, int thread_m_blocks,
  int cache_size = get_kernel_cache_size(
      th_config, 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)
@ -315,22 +326,39 @@ bool is_valid_config(thread_config_t const& th_config, int thread_m_blocks,
    BIGGROUP_GET_IF_M234(W_TYPE, 8, 4, 128)  \
    BIGGROUP_GET_IF_M234(W_TYPE, 4, 8, 128)

-  #define FP4_GET_IF_M1(W_TYPE, N_BLOCKS, K_BLOCKS, NUM_THREADS)        \
+  #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 FP4_GET_IF_M234(W_TYPE, N_BLOCKS, K_BLOCKS, NUM_THREADS)       \
+  #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 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_M1(W_TYPE, 4, 8, 128)    \
-    FP4_GET_IF_M234(W_TYPE, 16, 4, 256) \
-    FP4_GET_IF_M234(W_TYPE, 8, 4, 128)  \
-    FP4_GET_IF_M234(W_TYPE, 4, 8, 128)
+  #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_M1(W_TYPE, 4, 8, 128)    \
+    NVFP4_GET_IF_M234(W_TYPE, 16, 4, 256) \
+    NVFP4_GET_IF_M234(W_TYPE, 8, 4, 128)  \
+    NVFP4_GET_IF_M234(W_TYPE, 4, 8, 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_M1(W_TYPE, 4, 8, 128)    \
+    MXFP4_GET_IF_M234(W_TYPE, 16, 4, 256) \
+    MXFP4_GET_IF_M234(W_TYPE, 8, 4, 128)  \
+    MXFP4_GET_IF_M234(W_TYPE, 4, 8, 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)       \
@ -384,7 +412,7 @@ MarlinFuncPtr get_marlin_kernel(const vllm::ScalarType q_type,
  COMMON_GET_IF(vllm::kU4B8)
  COMMON_GET_IF(vllm::kU8B128)

-  FP4_GET_IF(vllm::kFE2M1f)
+  NVFP4_GET_IF(vllm::kFE2M1f)

  BIGGROUP_GET_IF(vllm::kFE4M3fn)

@ -396,6 +424,11 @@ MarlinFuncPtr get_marlin_kernel(const vllm::ScalarType q_type,
    }
    FZP_GET_IF(vllm::kU4)
  }
+  if (std::is_same<scalar_t, nv_bfloat16>::value) {
+    if (false) {
+    }
+    MXFP4_GET_IF(vllm::kFE2M1f)
+  }

  return kernel;
 }
@ -453,12 +486,12 @@ 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,
-               int prob_m, int prob_n, int prob_k, int lda, 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_init,
+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, int prob_m, int prob_n, int prob_k, int lda,
+               void* workspace, 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_init,
               int thread_n_init, int sms, bool use_atomic_add,
               bool use_fp32_reduce, bool is_zp_float) {
  if (has_zp) {
@ -503,6 +536,7 @@ 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;
@ -623,8 +657,9 @@ 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_new, stream>>>(
-        A_ptr, B_ptr, C_ptr, C_tmp_ptr, s_ptr, s2_ptr, zp_ptr, g_idx_ptr, num_groups,
-        prob_m_split, prob_n, prob_k, lda, locks, part_use_atomic_add,
+        A_ptr, B_ptr, C_ptr, C_tmp_ptr, bias_ptr, s_ptr, s2_ptr, zp_ptr,
+        g_idx_ptr, num_groups,
+        prob_m_split, prob_n, prob_k, lda, locks, has_bias, part_use_atomic_add,
        use_fp32_reduce, max_shared_mem_new);
    // clang-format on

@ -638,7 +673,8 @@ void marlin_mm(const void* A, const void* B, void* C, void* C_tmp, void* s,

 torch::Tensor gptq_marlin_gemm(
    torch::Tensor& a, std::optional<torch::Tensor> 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,
@ -785,12 +821,24 @@ torch::Tensor gptq_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(0) == size_n, "b_bias.size(0) != size_n");
+    TORCH_CHECK(b_bias.stride(0) == 1, "b_bias.stride(0) != 1");
+  } else {
+    b_bias = torch::empty({0}, options);
  }

  torch::Tensor b_zeros;
@ -857,34 +905,50 @@ torch::Tensor gptq_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::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>(), size_m, size_n, size_k, a.stride(0),
-        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);
+        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>(), size_m, size_n, size_k,
+        a.stride(0), 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::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>(),
        size_m, size_n, size_k, a.stride(0), workspace.data_ptr(), b_q_type,
-        has_act_order, is_k_full, has_zp, num_groups, group_size, dev,
+        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 {
--- a/csrc/quantization/gptq_marlin/kernel.h
+++ b/csrc/quantization/gptq_marlin/kernel.h
@ -10,15 +10,18 @@
 #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,          \
      int num_groups, int prob_m, int prob_n, int prob_k, int lda, int *locks, \
-      bool use_atomic_add, bool use_fp32_reduce, int max_shared_mem
+      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 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
--- a/csrc/quantization/gptq_marlin/marlin_template.h
+++ b/csrc/quantization/gptq_marlin/marlin_template.h
@ -39,6 +39,7 @@ 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
@ -271,6 +272,7 @@ __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
@ -290,6 +292,7 @@ __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
@ -297,12 +300,13 @@ __global__ void Marlin(
    const int4* __restrict__ zp_ptr,  // 4bit packed zero-points of shape
                                      // (k/groupsize)x(n/pack_factor)
    const int* __restrict__ g_idx,    // int32 group indices of shape k
-    int num_groups,        // number of scale groups per output channel
-    int prob_m,            // batch dimension m
-    int prob_n,            // output dimension n
-    int prob_k,            // reduction dimension k
-    int lda,               // A.stride(0), equal to prob_k is A is contiguous
-    int* locks,            // extra global storage for barrier synchronization
+    int num_groups,  // number of scale groups per output channel
+    int prob_m,      // batch dimension m
+    int prob_n,      // output dimension n
+    int prob_k,      // reduction dimension k
+    int lda,         // A.stride(0), equal to prob_k is A is contiguous
+    int* locks,      // extra global storage for barrier synchronization
+    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) {
@ -326,18 +330,29 @@ __global__ void Marlin(
  using FragZP = typename ScalarType<scalar_t>::FragZP;

  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);

  scalar_t2 global_scale;
-
-  if constexpr (w_type == vllm::kFE2M1f) {
+  if constexpr (w_type == vllm::kFE2M1f && s_type == vllm::kFE4M3fn) {
+    // NVFP4 format requires global scale
    uint16_t val = scale2_ptr[0];
    global_scale = Dtype::num2num2(*reinterpret_cast<scalar_t*>(&val));
  }
@ -589,7 +604,7 @@ __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)) +
@ -602,6 +617,18 @@ __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;
@ -670,7 +697,19 @@ __global__ void Marlin(
  constexpr int sh_b_size = stages * b_sh_stage;
  int4* sh_b = sh;
  int4* sh_red = sh;
-  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 + sh_size_b_red_min;
+  int4* sh_g_idx = sh + 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);
@ -680,15 +719,13 @@ __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 =
-  //     stages * (g_idx_stage + zp_sh_stage) + sh_s_size +
-  //     (sh_red_size > sh_b_size ? sh_red_size : sh_b_size);

  // Register storage for double buffer of shared memory reads.
  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
@ -923,10 +960,15 @@ __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];
          }
        }
      }
@ -1139,9 +1181,9 @@ __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);
    }

@ -1411,7 +1453,7 @@ __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));
@ -1438,7 +1480,7 @@ __global__ void Marlin(
    int c_gl_wr_end = c_gl_stride * prob_m;
    // 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));

@ -1447,12 +1489,25 @@ __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) {
        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;
@ -1470,19 +1525,25 @@ __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);
@ -1622,6 +1683,14 @@ __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)) {
@ -1684,11 +1753,20 @@ __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);
      slice_row = 0;
      slice_col_par++;
      slice_col++;
@ -1706,6 +1784,7 @@ __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;
--- a/csrc/quantization/machete/generate.py
+++ b/csrc/quantization/machete/generate.py
@ -349,9 +349,12 @@ def to_cute_constant(value: list[int]):


 def unique_schedules(impl_configs: list[ImplConfig]):
-    return list(
-        set(sch for impl_config in impl_configs
-            for sch in impl_config.schedules))
+    # Use dict over set for deterministic ordering
+    return list({
+        sch: None
+        for impl_config in impl_configs
+        for sch in impl_config.schedules
+    }.keys())


 def unsigned_type_with_bitwidth(num_bits):
@ -568,78 +571,79 @@ def generate():
                     itertools.repeat(default_heuristic))
    ]

-    # Stored as "condition": ((tile_shape_mn), (cluster_shape_mnk))
-    # TODO (LucasWilkinson): Further tuning required
-    qqq_tile_heuristic_config = {
-        #### M = 257+
-        # ((128, 256), (2, 1, 1)) Broken for QQQ types
-        # TODO (LucasWilkinson): Investigate further
-        # "M > 256 && K <= 16384 && N <= 4096": ((128, 128), (2, 1, 1)),
-        # "M > 256": ((128, 256), (2, 1, 1)),
-        "M > 256": ((128, 128), (2, 1, 1)),
-        #### M = 129-256
-        "M > 128 && K <= 4096 && N <= 4096": ((128, 64), (2, 1, 1)),
-        "M > 128 && K <= 8192 && N <= 8192": ((128, 128), (2, 1, 1)),
-        # ((128, 256), (2, 1, 1)) Broken for QQQ types
-        # TODO (LucasWilkinson): Investigate further
-        # "M > 128": ((128, 256), (2, 1, 1)),
-        "M > 128": ((128, 128), (2, 1, 1)),
-        #### M = 65-128
-        "M > 64 && K <= 4069 && N <= 4069": ((128, 32), (2, 1, 1)),
-        "M > 64 && K <= 4069 && N <= 8192": ((128, 64), (2, 1, 1)),
-        "M > 64 && K >= 8192 && N >= 12288": ((256, 128), (2, 1, 1)),
-        "M > 64": ((128, 128), (2, 1, 1)),
-        #### M = 33-64
-        "M > 32 && K <= 6144 && N <= 6144": ((128, 16), (1, 1, 1)),
-        # Broken for QQQ types
-        # TODO (LucasWilkinson): Investigate further
-        #"M > 32 && K >= 16384 && N >= 12288": ((256, 64), (2, 1, 1)),
-        "M > 32": ((128, 64), (2, 1, 1)),
-        #### M = 17-32
-        "M > 16 && K <= 12288 && N <= 8192": ((128, 32), (2, 1, 1)),
-        "M > 16": ((256, 32), (2, 1, 1)),
-        #### M = 1-16
-        "N >= 26624": ((256, 16), (1, 1, 1)),
-        None: ((128, 16), (1, 1, 1)),
-    }
+    # TODO: Support W4A8 when ready
+    # # Stored as "condition": ((tile_shape_mn), (cluster_shape_mnk))
+    # # TODO (LucasWilkinson): Further tuning required
+    # qqq_tile_heuristic_config = {
+    #     #### M = 257+
+    #     # ((128, 256), (2, 1, 1)) Broken for QQQ types
+    #     # TODO (LucasWilkinson): Investigate further
+    #     # "M > 256 && K <= 16384 && N <= 4096": ((128, 128), (2, 1, 1)),
+    #     # "M > 256": ((128, 256), (2, 1, 1)),
+    #     "M > 256": ((128, 128), (2, 1, 1)),
+    #     #### M = 129-256
+    #     "M > 128 && K <= 4096 && N <= 4096": ((128, 64), (2, 1, 1)),
+    #     "M > 128 && K <= 8192 && N <= 8192": ((128, 128), (2, 1, 1)),
+    #     # ((128, 256), (2, 1, 1)) Broken for QQQ types
+    #     # TODO (LucasWilkinson): Investigate further
+    #     # "M > 128": ((128, 256), (2, 1, 1)),
+    #     "M > 128": ((128, 128), (2, 1, 1)),
+    #     #### M = 65-128
+    #     "M > 64 && K <= 4069 && N <= 4069": ((128, 32), (2, 1, 1)),
+    #     "M > 64 && K <= 4069 && N <= 8192": ((128, 64), (2, 1, 1)),
+    #     "M > 64 && K >= 8192 && N >= 12288": ((256, 128), (2, 1, 1)),
+    #     "M > 64": ((128, 128), (2, 1, 1)),
+    #     #### M = 33-64
+    #     "M > 32 && K <= 6144 && N <= 6144": ((128, 16), (1, 1, 1)),
+    #     # Broken for QQQ types
+    #     # TODO (LucasWilkinson): Investigate further
+    #     #"M > 32 && K >= 16384 && N >= 12288": ((256, 64), (2, 1, 1)),
+    #     "M > 32": ((128, 64), (2, 1, 1)),
+    #     #### M = 17-32
+    #     "M > 16 && K <= 12288 && N <= 8192": ((128, 32), (2, 1, 1)),
+    #     "M > 16": ((256, 32), (2, 1, 1)),
+    #     #### M = 1-16
+    #     "N >= 26624": ((256, 16), (1, 1, 1)),
+    #     None: ((128, 16), (1, 1, 1)),
+    # }

-    # For now we use the same heuristic for all types
-    # Heuristic is currently tuned for H100s
-    qqq_heuristic = [
-        (cond, ScheduleConfig(*tile_config,
-                              **sch_common_params))  # type: ignore
-        for cond, tile_config in qqq_tile_heuristic_config.items()
-    ]
+    # # For now we use the same heuristic for all types
+    # # Heuristic is currently tuned for H100s
+    # qqq_heuristic = [
+    #     (cond, ScheduleConfig(*tile_config,
+    #                           **sch_common_params))  # type: ignore
+    #     for cond, tile_config in qqq_tile_heuristic_config.items()
+    # ]

-    QQQ_kernel_types = [
-        *(TypeConfig(
-            a=DataType.s8,
-            b=VLLMDataType.u4b8,
-            b_group_scale=b_group_scale,
-            b_group_zeropoint=DataType.void,
-            b_channel_scale=DataType.f32,
-            a_token_scale=DataType.f32,
-            out=DataType.f16,
-            accumulator=DataType.s32,
-        ) for b_group_scale in (DataType.f16, DataType.void)),
-        *(TypeConfig(
-            a=DataType.e4m3,
-            b=VLLMDataType.u4b8,
-            b_group_scale=b_group_scale,
-            b_group_zeropoint=DataType.void,
-            b_channel_scale=DataType.f32,
-            a_token_scale=DataType.f32,
-            out=DataType.f16,
-            accumulator=DataType.f32,
-        ) for b_group_scale in (DataType.f16, DataType.void)),
-    ]
+    # QQQ_kernel_types = [
+    #     *(TypeConfig(
+    #         a=DataType.s8,
+    #         b=VLLMDataType.u4b8,
+    #         b_group_scale=b_group_scale,
+    #         b_group_zeropoint=DataType.void,
+    #         b_channel_scale=DataType.f32,
+    #         a_token_scale=DataType.f32,
+    #         out=DataType.f16,
+    #         accumulator=DataType.s32,
+    #     ) for b_group_scale in (DataType.f16, DataType.void)),
+    #     *(TypeConfig(
+    #         a=DataType.e4m3,
+    #         b=VLLMDataType.u4b8,
+    #         b_group_scale=b_group_scale,
+    #         b_group_zeropoint=DataType.void,
+    #         b_channel_scale=DataType.f32,
+    #         a_token_scale=DataType.f32,
+    #         out=DataType.f16,
+    #         accumulator=DataType.f32,
+    #     ) for b_group_scale in (DataType.f16, DataType.void)),
+    # ]

-    impl_configs += [
-        ImplConfig(x[0], x[1], x[2])
-        for x in zip(QQQ_kernel_types,
-                     itertools.repeat(get_unique_schedules(qqq_heuristic)),
-                     itertools.repeat(qqq_heuristic))
-    ]
+    # impl_configs += [
+    #     ImplConfig(x[0], x[1], x[2])
+    #     for x in zip(QQQ_kernel_types,
+    #                  itertools.repeat(get_unique_schedules(qqq_heuristic)),
+    #                  itertools.repeat(qqq_heuristic))
+    # ]

    output_dir = os.path.join(SCRIPT_DIR, "generated")

--- a/csrc/quantization/marlin/dense/LICENSE
+++ b/csrc/quantization/marlin/dense/LICENSE
@ -1,209 +0,0 @@
-Contains code from https://github.com/IST-DASLab/marlin
-
-                                 Apache License
-                           Version 2.0, January 2004
-                        http://www.apache.org/licenses/
-
-   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
-
-   1. Definitions.
-
-      "License" shall mean the terms and conditions for use, reproduction,
-      and distribution as defined by Sections 1 through 9 of this document.
-
-      "Licensor" shall mean the copyright owner or entity authorized by
-      the copyright owner that is granting the License.
-
-      "Legal Entity" shall mean the union of the acting entity and all
-      other entities that control, are controlled by, or are under common
-      control with that entity. For the purposes of this definition,
-      "control" means (i) the power, direct or indirect, to cause the
-      direction or management of such entity, whether by contract or
-      otherwise, or (ii) ownership of fifty percent (50%) or more of the
-      outstanding shares, or (iii) beneficial ownership of such entity.
-
-      "You" (or "Your") shall mean an individual or Legal Entity
-      exercising permissions granted by this License.
-
-      "Source" form shall mean the preferred form for making modifications,
-      including but not limited to software source code, documentation
-      source, and configuration files.
-
-      "Object" form shall mean any form resulting from mechanical
-      transformation or translation of a Source form, including but
-      not limited to compiled object code, generated documentation,
-      and conversions to other media types.
-
-      "Work" shall mean the work of authorship, whether in Source or
-      Object form, made available under the License, as indicated by a
-      copyright notice that is included in or attached to the work
-      (an example is provided in the Appendix below).
-
-      "Derivative Works" shall mean any work, whether in Source or Object
-      form, that is based on (or derived from) the Work and for which the
-      editorial revisions, annotations, elaborations, or other modifications
-      represent, as a whole, an original work of authorship. For the purposes
-      of this License, Derivative Works shall not include works that remain
-      separable from, or merely link (or bind by name) to the interfaces of,
-      the Work and Derivative Works thereof.
-
-      "Contribution" shall mean any work of authorship, including
-      the original version of the Work and any modifications or additions
-      to that Work or Derivative Works thereof, that is intentionally
-      submitted to Licensor for inclusion in the Work by the copyright owner
-      or by an individual or Legal Entity authorized to submit on behalf of
-      the copyright owner. For the purposes of this definition, "submitted"
-      means any form of electronic, verbal, or written communication sent
-      to the Licensor or its representatives, including but not limited to
-      communication on electronic mailing lists, source code control systems,
-      and issue tracking systems that are managed by, or on behalf of, the
-      Licensor for the purpose of discussing and improving the Work, but
-      excluding communication that is conspicuously marked or otherwise
-      designated in writing by the copyright owner as "Not a Contribution."
-
-      "Contributor" shall mean Licensor and any individual or Legal Entity
-      on behalf of whom a Contribution has been received by Licensor and
-      subsequently incorporated within the Work.
-
-   2. Grant of Copyright License. Subject to the terms and conditions of
-      this License, each Contributor hereby grants to You a perpetual,
-      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
-      copyright license to reproduce, prepare Derivative Works of,
-      publicly display, publicly perform, sublicense, and distribute the
-      Work and such Derivative Works in Source or Object form.
-
-   3. Grant of Patent License. Subject to the terms and conditions of
-      this License, each Contributor hereby grants to You a perpetual,
-      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
-      (except as stated in this section) patent license to make, have made,
-      use, offer to sell, sell, import, and otherwise transfer the Work,
-      where such license applies only to those patent claims licensable
-      by such Contributor that are necessarily infringed by their
-      Contribution(s) alone or by combination of their Contribution(s)
-      with the Work to which such Contribution(s) was submitted. If You
-      institute patent litigation against any entity (including a
-      cross-claim or counterclaim in a lawsuit) alleging that the Work
-      or a Contribution incorporated within the Work constitutes direct
-      or contributory patent infringement, then any patent licenses
-      granted to You under this License for that Work shall terminate
-      as of the date such litigation is filed.
-
-   4. Redistribution. You may reproduce and distribute copies of the
-      Work or Derivative Works thereof in any medium, with or without
-      modifications, and in Source or Object form, provided that You
-      meet the following conditions:
-
-      (a) You must give any other recipients of the Work or
-          Derivative Works a copy of this License; and
-
-      (b) You must cause any modified files to carry prominent notices
-          stating that You changed the files; and
-
-      (c) You must retain, in the Source form of any Derivative Works
-          that You distribute, all copyright, patent, trademark, and
-          attribution notices from the Source form of the Work,
-          excluding those notices that do not pertain to any part of
-          the Derivative Works; and
-
-      (d) If the Work includes a "NOTICE" text file as part of its
-          distribution, then any Derivative Works that You distribute must
-          include a readable copy of the attribution notices contained
-          within such NOTICE file, excluding those notices that do not
-          pertain to any part of the Derivative Works, in at least one
-          of the following places: within a NOTICE text file distributed
-          as part of the Derivative Works; within the Source form or
-          documentation, if provided along with the Derivative Works; or,
-          within a display generated by the Derivative Works, if and
-          wherever such third-party notices normally appear. The contents
-          of the NOTICE file are for informational purposes only and
-          do not modify the License. You may add Your own attribution
-          notices within Derivative Works that You distribute, alongside
-          or as an addendum to the NOTICE text from the Work, provided
-          that such additional attribution notices cannot be construed
-          as modifying the License.
-
-      You may add Your own copyright statement to Your modifications and
-      may provide additional or different license terms and conditions
-      for use, reproduction, or distribution of Your modifications, or
-      for any such Derivative Works as a whole, provided Your use,
-      reproduction, and distribution of the Work otherwise complies with
-      the conditions stated in this License.
-
-   5. Submission of Contributions. Unless You explicitly state otherwise,
-      any Contribution intentionally submitted for inclusion in the Work
-      by You to the Licensor shall be under the terms and conditions of
-      this License, without any additional terms or conditions.
-      Notwithstanding the above, nothing herein shall supersede or modify
-      the terms of any separate license agreement you may have executed
-      with Licensor regarding such Contributions.
-
-   6. Trademarks. This License does not grant permission to use the trade
-      names, trademarks, service marks, or product names of the Licensor,
-      except as required for reasonable and customary use in describing the
-      origin of the Work and reproducing the content of the NOTICE file.
-
-   7. Disclaimer of Warranty. Unless required by applicable law or
-      agreed to in writing, Licensor provides the Work (and each
-      Contributor provides its Contributions) on an "AS IS" BASIS,
-      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
-      implied, including, without limitation, any warranties or conditions
-      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
-      PARTICULAR PURPOSE. You are solely responsible for determining the
-      appropriateness of using or redistributing the Work and assume any
-      risks associated with Your exercise of permissions under this License.
-
-   8. Limitation of Liability. In no event and under no legal theory,
-      whether in tort (including negligence), contract, or otherwise,
-      unless required by applicable law (such as deliberate and grossly
-      negligent acts) or agreed to in writing, shall any Contributor be
-      liable to You for damages, including any direct, indirect, special,
-      incidental, or consequential damages of any character arising as a
-      result of this License or out of the use or inability to use the
-      Work (including but not limited to damages for loss of goodwill,
-      work stoppage, computer failure or malfunction, or any and all
-      other commercial damages or losses), even if such Contributor
-      has been advised of the possibility of such damages.
-
-   9. Accepting Warranty or Additional Liability. While redistributing
-      the Work or Derivative Works thereof, You may choose to offer,
-      and charge a fee for, acceptance of support, warranty, indemnity,
-      or other liability obligations and/or rights consistent with this
-      License. However, in accepting such obligations, You may act only
-      on Your own behalf and on Your sole responsibility, not on behalf
-      of any other Contributor, and only if You agree to indemnify,
-      defend, and hold each Contributor harmless for any liability
-      incurred by, or claims asserted against, such Contributor by reason
-      of your accepting any such warranty or additional liability.
-
-   END OF TERMS AND CONDITIONS
-
-   APPENDIX: How to apply the Apache License to your work.
-
-      To apply the Apache License to your work, attach the following
-      boilerplate notice, with the fields enclosed by brackets "{}"
-      replaced with your own identifying information. (Don't include
-      the brackets!)  The text should be enclosed in the appropriate
-      comment syntax for the file format. We also recommend that a
-      file or class name and description of purpose be included on the
-      same "printed page" as the copyright notice for easier
-      identification within third-party archives.
-
-   Copyright {yyyy} {name of copyright owner}
-
-   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.
-
------------------------------------------------------------------------------------
-
-This product bundles various third-party components under other open source licenses.
-This section summarizes those components and their licenses. See licenses/
-for text of these licenses.
--- a/csrc/quantization/marlin/dense/common/base.h
+++ b/csrc/quantization/marlin/dense/common/base.h
@ -1,32 +0,0 @@
-/*
- * Modified by HandH1998
- * Modified by Neural Magic
- * Copyright (C) Marlin.2024 Elias Frantar
- *
- * 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.
- */
-
-#pragma once
-
-constexpr int ceildiv(int a, int b) { return (a + b - 1) / b; }
-
-// Instances of `Vec` are used to organize groups of >>registers<<, as needed
-// for instance as inputs to tensor core operations. Consequently, all
-// corresponding index accesses must be compile-time constants, which is why we
-// extensively use `#pragma unroll` throughout the kernel code to guarantee
-// this.
-template <typename T, int n>
-struct Vec {
-  T elems[n];
-  __device__ T& operator[](int i) { return elems[i]; }
-};
--- a/csrc/quantization/marlin/dense/common/mem.h
+++ b/csrc/quantization/marlin/dense/common/mem.h
@ -1,89 +0,0 @@
-/*
- * Modified by HandH1998
- * Modified by Neural Magic
- * Copyright (C) Marlin.2024 Elias Frantar
- *
- * 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.
- */
-
-#pragma once
-
-// Predicated asynchronous global->shared copy; used for inputs A where we apply
-// predication to handle batchsizes that are not multiples of 16.
-__device__ inline void cp_async4_pred(void* smem_ptr, const void* glob_ptr,
-                                      bool pred = true) {
-  const int BYTES = 16;
-  uint32_t smem = static_cast<uint32_t>(__cvta_generic_to_shared(smem_ptr));
-  asm volatile(
-      "{\n"
-      "   .reg .pred p;\n"
-      "   setp.ne.b32 p, %0, 0;\n"
-      "   @p cp.async.cg.shared.global [%1], [%2], %3;\n"
-      "}\n" ::"r"((int)pred),
-      "r"(smem), "l"(glob_ptr), "n"(BYTES));
-}
-
-// Asynchronous global->shared copy
-__device__ inline void cp_async4(void* smem_ptr, const void* glob_ptr) {
-  const int BYTES = 16;
-  uint32_t smem = static_cast<uint32_t>(__cvta_generic_to_shared(smem_ptr));
-  asm volatile(
-      "{\n"
-      "   cp.async.cg.shared.global [%0], [%1], %2;\n"
-      "}\n" ::"r"(smem),
-      "l"(glob_ptr), "n"(BYTES));
-}
-
-// Async copy fence.
-__device__ inline void cp_async_fence() {
-  asm volatile("cp.async.commit_group;\n" ::);
-}
-
-// Wait until at most `n` async copy stages are still pending.
-template <int n>
-__device__ inline void cp_async_wait() {
-  asm volatile("cp.async.wait_group %0;\n" ::"n"(n));
-}
-
-// Wait until barrier reaches `count`, then lock for current threadblock.
-__device__ inline void barrier_acquire(int* lock, int count) {
-  if (threadIdx.x == 0) {
-    int state = -1;
-    do
-      // Guarantee that subsequent writes by this threadblock will be visible
-      // globally.
-      asm volatile("ld.global.acquire.gpu.b32 %0, [%1];\n"
-                   : "=r"(state)
-                   : "l"(lock));
-    while (state != count);
-  }
-  __syncthreads();
-}
-
-// Release barrier and increment visitation count.
-__device__ inline void barrier_release(int* lock, bool reset = false) {
-  __syncthreads();
-  if (threadIdx.x == 0) {
-    if (reset) {
-      lock[0] = 0;
-      return;
-    }
-    int val = 1;
-    // Make sure that all writes since acquiring this barrier are visible
-    // globally, while releasing the barrier.
-    asm volatile("fence.acq_rel.gpu;\n");
-    asm volatile("red.relaxed.gpu.global.add.s32 [%0], %1;\n"
-                 :
-                 : "l"(lock), "r"(val));
-  }
-}
--- a/csrc/quantization/marlin/dense/marlin_cuda_kernel.cu
+++ b/csrc/quantization/marlin/dense/marlin_cuda_kernel.cu
--- a/csrc/quantization/marlin/qqq/marlin_qqq_gemm_kernel.cu
+++ b/csrc/quantization/marlin/qqq/marlin_qqq_gemm_kernel.cu
--- a/csrc/quantization/vectorization_utils.cuh
+++ b/csrc/quantization/vectorization_utils.cuh
@ -41,8 +41,10 @@ __device__ inline void vectorize_with_alignment(

    for (int i = tid; i < num_vec; i += stride) {
      vout_t tmp;
-      vec_op(tmp, v_in[i]);
-      v_out[i] = tmp;
+      // Make a local copy of the entire pack
+      vin_t src = v_in[i];  // <- encourages a single vector ld
+      vec_op(tmp, src);
+      v_out[i] = tmp;  // <- encourages a single vector st
    }
    return;
  }
@ -71,8 +73,10 @@ __device__ inline void vectorize_with_alignment(
  // 2. vectorize the main part
  for (int i = tid; i < num_vec; i += stride) {
    vout_t tmp;
-    vec_op(tmp, v_in[i]);
-    v_out[i] = tmp;
+    // Make a local copy of the entire pack
+    vin_t src = v_in[i];  // <- encourages a single vector ld
+    vec_op(tmp, src);
+    v_out[i] = tmp;  // <- encourages a single vector st
  }

  // 3. handle the tail
@ -125,7 +129,8 @@ __device__ inline void vectorize_read_with_alignment(const InT* in, int len,
    auto* v_in = reinterpret_cast<const vin_t*>(in);

    for (int i = tid; i < num_vec; i += stride) {
-      vec_op(v_in[i]);
+      vin_t tmp = v_in[i];
+      vec_op(tmp);
    }
    return;
  }
--- a/csrc/torch_bindings.cpp
+++ b/csrc/torch_bindings.cpp
@ -130,6 +130,12 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
  ops.def("fatrelu_and_mul(Tensor! out, Tensor input, float threshold) -> ()");
  ops.impl("fatrelu_and_mul", torch::kCUDA, &fatrelu_and_mul);

+  ops.def(
+      "swigluoai_and_mul(Tensor! out, Tensor input, float alpha=1.702, float "
+      "limit=7.0) "
+      "-> ()");
+  ops.impl("swigluoai_and_mul", torch::kCUDA, &swigluoai_and_mul);
+
  // GELU implementation used in GPT-2.
  ops.def("gelu_new(Tensor! out, Tensor input) -> ()");
  ops.impl("gelu_new", torch::kCUDA, &gelu_new);
@ -142,25 +148,6 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
  ops.def("gelu_quick(Tensor! out, Tensor input) -> ()");
  ops.impl("gelu_quick", torch::kCUDA, &gelu_quick);

-  // prepare_inputs advance_step
-  ops.def(
-      "advance_step_flashattn(int num_seqs, int num_queries, int block_size, "
-      "Tensor! input_tokens, Tensor sampled_token_ids, "
-      "Tensor! input_positions, Tensor! seq_lens, Tensor! slot_mapping, "
-      "Tensor block_tables) -> ()");
-  ops.impl("advance_step_flashattn", torch::kCUDA, &advance_step_flashattn);
-
-  ops.def(
-      "advance_step_flashinfer("
-      "    int num_seqs, int num_queries, int block_size,"
-      "    Tensor! input_tokens, Tensor sampled_token_ids,"
-      "    Tensor! input_positions, Tensor! seq_lens, Tensor! slot_mapping,"
-      "    Tensor block_tables, Tensor! paged_kv_indices,"
-      "    Tensor! paged_kv_indptr, Tensor! paged_kv_last_page_len,"
-      "    Tensor! block_table_bounds"
-      ") -> ()");
-  ops.impl("advance_step_flashinfer", torch::kCUDA, &advance_step_flashinfer);
-
  // Layernorm
  // Apply Root Mean Square (RMS) Normalization to the input tensor.
  ops.def(
@ -226,21 +213,6 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {

  // Quantization ops
 #ifndef USE_ROCM
-  // Quantized GEMM for AQLM.
-  ops.def(
-      "aqlm_gemm(Tensor input, Tensor codes, Tensor codebooks, "
-      "Tensor scales, int[] codebook_partition_sizes, Tensor? bias) "
-      "-> Tensor",
-      {stride_tag});
-  ops.impl("aqlm_gemm", torch::kCUDA, &aqlm_gemm);
-
-  // Decompression method for AQLM.
-  ops.def(
-      "aqlm_dequant(Tensor codes, Tensor codebooks, "
-      "int[] codebook_partition_sizes) -> Tensor",
-      {stride_tag});
-  ops.impl("aqlm_dequant", torch::kCUDA, &aqlm_dequant);
-
  // Quantized GEMM for AWQ.
  ops.def(
      "awq_gemm(Tensor _in_feats, Tensor _kernel, Tensor _scaling_factors, "
@ -269,14 +241,6 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
  // custom types:
  // https://docs.google.com/document/d/18fBMPuOJ0fY5ZQ6YyrHUppw9FA332CpNtgB6SOIgyuA

-  // Marlin (Dense) Optimized Quantized GEMM for GPTQ.
-  ops.def(
-      "marlin_gemm(Tensor a, Tensor b_q_weight, Tensor b_scales, "
-      "Tensor! workspace, SymInt size_m, SymInt size_n, SymInt size_k) -> "
-      "Tensor",
-      {stride_tag});
-  // conditionally compiled so impl in source file
-
  // Marlin_24 (Sparse) Optimized Quantized GEMM for GPTQ.
  ops.def(
      "gptq_marlin_24_gemm(Tensor a, Tensor b_q_weight, Tensor b_meta, "
@ -326,6 +290,7 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
  // gptq_marlin Optimized Quantized GEMM for GPTQ.
  ops.def(
      "gptq_marlin_gemm(Tensor a, Tensor? c_or_none, 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, int b_q_type, "
      "SymInt size_m, SymInt size_n, SymInt size_k, bool is_k_full, "
@ -344,6 +309,26 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
      "awq_marlin_repack(Tensor b_q_weight, SymInt size_k, "
      "SymInt size_n, int num_bits) -> Tensor");
  // conditionally compiled so impl registrations are in source file
+
+  // CUTLASS w4a8 GEMM
+  ops.def(
+      "cutlass_w4a8_mm("
+      "   Tensor A,"
+      "   Tensor B,"
+      "   Tensor group_scales,"
+      "   int    group_size,"
+      "   Tensor channel_scales,"
+      "   Tensor token_scales,"
+      "   ScalarType? out_type,"
+      "   str?   maybe_schedule"
+      ") -> Tensor",
+      {stride_tag});
+  // pack scales
+  ops.def("cutlass_pack_scale_fp8(Tensor scales) -> Tensor");
+  // encode and reorder weight matrix
+  ops.def("cutlass_encode_and_reorder_int4b(Tensor B) -> Tensor");
+  // conditionally compiled so impl registration is in source file
+
 #endif

  // Dequantization for GGML.
@ -380,15 +365,6 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
  ops.def("ggml_moe_get_block_size", &ggml_moe_get_block_size);

 #ifndef USE_ROCM
-  // marlin_qqq_gemm for QQQ.
-  ops.def(
-      "marlin_qqq_gemm(Tensor a, Tensor b_q_weight, "
-      "Tensor s_tok, Tensor s_ch, Tensor s_group, "
-      "Tensor! workspace, SymInt size_m, SymInt size_n, "
-      "SymInt size_k) -> Tensor",
-      {stride_tag});
-  // conditionally compiled so impl registration is in source file
-
  // CUTLASS nvfp4 block scaled GEMM
  ops.def(
      "cutlass_scaled_fp4_mm(Tensor! out, Tensor a, Tensor b,"
@ -467,6 +443,19 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
      {stride_tag});
  ops.impl("get_cutlass_moe_mm_data", torch::kCUDA, &get_cutlass_moe_mm_data);

+  // A function that computes problem sizes for each expert's multiplication
+  // used by the two mms called from fused MoE operation. It takes topk_ids as
+  // an input, and computes problem_sizes1 and problem_sizes2 only.
+  ops.def(
+      "get_cutlass_moe_mm_problem_sizes(Tensor topk_ids, "
+      "                                 Tensor! problem_sizes1, "
+      "                                 Tensor! problem_sizes2, "
+      "                                 int num_experts, int n, int k, "
+      "                                 Tensor? blockscale_offsets) -> ()",
+      {stride_tag});
+  ops.impl("get_cutlass_moe_mm_problem_sizes", torch::kCUDA,
+           &get_cutlass_moe_mm_problem_sizes);
+
  // A function that computes data required to run fused MoE with w8a8 grouped
  // GEMM and PPLX. It takes expert_num_tokens and non_zero_expert_idxs
  // as an input, and computes expert_offsets (token start indices of each
@ -703,11 +692,16 @@ TORCH_LIBRARY_EXPAND(CONCAT(TORCH_EXTENSION_NAME, _cache_ops), cache_ops) {
      "str kv_cache_dtype) -> ()");
  cache_ops.impl("convert_fp8", torch::kCUDA, &convert_fp8);

-  // Gather cache blocks from src_cache to dst.
+  // Gather cache blocks from src_cache to dst, dequantizing from
+  // src_cache's dtype to dst's dtype if necessary.
  cache_ops.def(
-      "gather_cache(Tensor src_cache, Tensor! dst, Tensor block_table, "
-      "Tensor cu_seq_lens, int batch_size, Tensor? seq_starts) -> ()");
-  cache_ops.impl("gather_cache", torch::kCUDA, &gather_cache);
+      "gather_and_maybe_dequant_cache(Tensor src_cache, Tensor! dst, "
+      "                               Tensor block_table, Tensor cu_seq_lens, "
+      "                               int batch_size, "
+      "                               str kv_cache_dtype, "
+      "                               Tensor scale, Tensor? seq_starts) -> ()");
+  cache_ops.impl("gather_and_maybe_dequant_cache", torch::kCUDA,
+                 &gather_and_maybe_dequant_cache);
 }

 TORCH_LIBRARY_EXPAND(CONCAT(TORCH_EXTENSION_NAME, _cuda_utils), cuda_utils) {
--- a/docker/Dockerfile
+++ b/docker/Dockerfile
@ -139,21 +139,6 @@ RUN ldconfig /usr/local/cuda-$(echo $CUDA_VERSION | cut -d. -f1,2)/compat/
 WORKDIR /workspace

 # install build and runtime dependencies
-
-# arm64 (GH200) build follows the practice of "use existing pytorch" build,
-# we need to install torch and torchvision from the nightly builds first,
-# pytorch will not appear as a vLLM dependency in all of the following steps
-# after this step
-RUN --mount=type=cache,target=/root/.cache/uv \
-    if [ "$TARGETPLATFORM" = "linux/arm64" ]; then \
-        uv pip install --system \
-            --index-url ${PYTORCH_CUDA_NIGHTLY_INDEX_BASE_URL}/cu$(echo $CUDA_VERSION | cut -d. -f1,2 | tr -d '.') \
-            "torch==2.8.0.dev20250318+cu128" "torchvision==0.22.0.dev20250319";  \
-        uv pip install --system \
-            --index-url ${PYTORCH_CUDA_NIGHTLY_INDEX_BASE_URL}/cu$(echo $CUDA_VERSION | cut -d. -f1,2 | tr -d '.') \
-            --pre pytorch_triton==3.3.0+gitab727c40; \
-    fi
-
 COPY requirements/common.txt requirements/common.txt
 COPY requirements/cuda.txt requirements/cuda.txt
 RUN --mount=type=cache,target=/root/.cache/uv \
@ -234,6 +219,8 @@ RUN --mount=type=cache,target=/root/.cache/uv \
        && sccache --show-stats; \
    fi

+ARG vllm_target_device="cuda"
+ENV VLLM_TARGET_DEVICE=${vllm_target_device}
 ENV CCACHE_DIR=/root/.cache/ccache
 RUN --mount=type=cache,target=/root/.cache/ccache \
    --mount=type=cache,target=/root/.cache/uv \
@ -385,31 +372,45 @@ RUN --mount=type=bind,from=build,src=/workspace/dist,target=/vllm-workspace/dist

 # Install FlashInfer from source
 ARG FLASHINFER_GIT_REPO="https://github.com/flashinfer-ai/flashinfer.git"
-# Keep this in sync with https://github.com/vllm-project/vllm/blob/main/requirements/cuda.txt
-# We use `--force-reinstall --no-deps` to avoid issues with the existing FlashInfer wheel.
-ARG FLASHINFER_GIT_REF="v0.2.11"
+# Keep this in sync with "flashinfer" extra in setup.py
+ARG FLASHINFER_GIT_REF="v0.2.12"
+# Flag to control whether to compile FlashInfer AOT kernels
+# Set to "true" to enable AOT compilation:
+# docker build --build-arg FLASHINFER_AOT_COMPILE=true ...
+ARG FLASHINFER_AOT_COMPILE=false
 RUN --mount=type=cache,target=/root/.cache/uv bash - <<'BASH'
  . /etc/environment
    git clone --depth 1 --recursive --shallow-submodules \
        --branch ${FLASHINFER_GIT_REF} \
        ${FLASHINFER_GIT_REPO} flashinfer
-    # Exclude CUDA arches for older versions (11.x and 12.0-12.7)
-    # TODO: Update this to allow setting TORCH_CUDA_ARCH_LIST as a build arg.
-    if [[ "${CUDA_VERSION}" == 11.* ]]; then
-        FI_TORCH_CUDA_ARCH_LIST="7.5 8.0 8.9"
-    elif [[ "${CUDA_VERSION}" == 12.[0-7]* ]]; then
-        FI_TORCH_CUDA_ARCH_LIST="7.5 8.0 8.9 9.0a"
-    else
-        # CUDA 12.8+ supports 10.0a and 12.0
-        FI_TORCH_CUDA_ARCH_LIST="7.5 8.0 8.9 9.0a 10.0a 12.0"
-    fi
-    echo "🏗️  Building FlashInfer for arches: ${FI_TORCH_CUDA_ARCH_LIST}"
-    # Needed to build AOT kernels
    pushd flashinfer
-        TORCH_CUDA_ARCH_LIST="${FI_TORCH_CUDA_ARCH_LIST}" \
-            python3 -m flashinfer.aot
-        TORCH_CUDA_ARCH_LIST="${FI_TORCH_CUDA_ARCH_LIST}" \
-            uv pip install --system --no-build-isolation --force-reinstall --no-deps .
+        if [ "${FLASHINFER_AOT_COMPILE}" = "true" ]; then
+            # Exclude CUDA arches for older versions (11.x and 12.0-12.7)
+            # TODO: Update this to allow setting TORCH_CUDA_ARCH_LIST as a build arg.
+            if [[ "${CUDA_VERSION}" == 11.* ]]; then
+                FI_TORCH_CUDA_ARCH_LIST="7.5 8.0 8.9"
+            elif [[ "${CUDA_VERSION}" == 12.[0-7]* ]]; then
+                FI_TORCH_CUDA_ARCH_LIST="7.5 8.0 8.9 9.0a"
+            else
+                # CUDA 12.8+ supports 10.0a and 12.0
+                FI_TORCH_CUDA_ARCH_LIST="7.5 8.0 8.9 9.0a 10.0a 12.0"
+            fi
+            echo "🏗️  Installing FlashInfer with AOT compilation for arches: ${FI_TORCH_CUDA_ARCH_LIST}"
+            # Build AOT kernels
+            TORCH_CUDA_ARCH_LIST="${FI_TORCH_CUDA_ARCH_LIST}" \
+                python3 -m flashinfer.aot
+            # Install with no-build-isolation since we already built AOT kernels
+            TORCH_CUDA_ARCH_LIST="${FI_TORCH_CUDA_ARCH_LIST}" \
+                uv pip install --system --no-build-isolation . \
+                --extra-index-url ${PYTORCH_CUDA_INDEX_BASE_URL}/cu$(echo $CUDA_VERSION | cut -d. -f1,2 | tr -d '.')
+            # Download pre-compiled cubins
+            TORCH_CUDA_ARCH_LIST="${FI_TORCH_CUDA_ARCH_LIST}" \
+                python3 -m flashinfer --download-cubin || echo "WARNING: Failed to download flashinfer cubins."
+        else
+            echo "🏗️  Installing FlashInfer without AOT compilation in JIT mode"
+            uv pip install --system . \
+                --extra-index-url ${PYTORCH_CUDA_INDEX_BASE_URL}/cu$(echo $CUDA_VERSION | cut -d. -f1,2 | tr -d '.')
+        fi
    popd
    rm -rf flashinfer
 BASH
@ -431,31 +432,19 @@ RUN --mount=type=cache,target=/root/.cache/uv \
        --extra-index-url ${PYTORCH_CUDA_INDEX_BASE_URL}/cu$(echo $CUDA_VERSION | cut -d. -f1,2 | tr -d '.')

 # Install DeepGEMM from source
-ARG DEEPGEMM_GIT_REPO="https://github.com/deepseek-ai/DeepGEMM.git"
 ARG DEEPGEMM_GIT_REF="7b6b5563b9d4c1ae07ffbce7f78ad3ac9204827c"
-RUN --mount=type=cache,target=/root/.cache/uv bash - <<'BASH'
-  . /etc/environment
-    CUDA_MAJOR="${CUDA_VERSION%%.*}"
-    CUDA_MINOR="${CUDA_VERSION#${CUDA_MAJOR}.}"
-    CUDA_MINOR="${CUDA_MINOR%%.*}"
-    if [ "$CUDA_MAJOR" -ge 12 ] && [ "$CUDA_MINOR" -ge 8 ]; then
-        git clone --recursive --shallow-submodules \
-            ${DEEPGEMM_GIT_REPO} deepgemm
-        echo "🏗️  Building DeepGEMM"
-        pushd deepgemm
-            git checkout ${DEEPGEMM_GIT_REF}
-            # Build DeepGEMM
-            # (Based on https://github.com/deepseek-ai/DeepGEMM/blob/main/install.sh)
-            rm -rf build dist
-            rm -rf *.egg-info
-            python3 setup.py bdist_wheel
-            uv pip install --system dist/*.whl
-        popd
-        rm -rf deepgemm
-    else
-        echo "Skipping DeepGEMM installation (requires CUDA 12.8+ but got ${CUDA_VERSION})"
-    fi
-BASH
+COPY tools/install_deepgemm.sh /tmp/install_deepgemm.sh
+RUN --mount=type=cache,target=/root/.cache/uv \
+    VLLM_DOCKER_BUILD_CONTEXT=1 /tmp/install_deepgemm.sh --cuda-version "${CUDA_VERSION}" --ref "${DEEPGEMM_GIT_REF}" \
+    && rm /tmp/install_deepgemm.sh
+
+# Install EP kernels(pplx-kernels and DeepEP), NixL
+COPY tools/ep_kernels/install_python_libraries.sh install_python_libraries.sh
+COPY tools/install_nixl.sh install_nixl.sh
+ENV CUDA_HOME=/usr/local/cuda
+RUN export TORCH_CUDA_ARCH_LIST="${TORCH_CUDA_ARCH_LIST:-9.0a+PTX}" \
+    && bash install_python_libraries.sh \
+    && bash install_nixl.sh --force

 #################### vLLM installation IMAGE ####################

@ -497,14 +486,11 @@ ENV HF_HUB_ENABLE_HF_TRANSFER 1
 # Copy in the v1 package for testing (it isn't distributed yet)
 COPY vllm/v1 /usr/local/lib/python${PYTHON_VERSION}/dist-packages/vllm/v1

-# doc requires source code
-# we hide them inside `test_docs/` , so that this source code
+# Source code is used in the `python_only_compile.sh` test
+# We hide it inside `src/` so that this source code
 # will not be imported by other tests
-RUN mkdir test_docs
-RUN mv docs test_docs/
-RUN cp -r examples test_docs/
-RUN mv vllm test_docs/
-RUN mv mkdocs.yaml test_docs/
+RUN mkdir src
+RUN mv vllm src/vllm
 #################### TEST IMAGE ####################

 #################### OPENAI API SERVER ####################
--- a/docker/Dockerfile.rocm
+++ b/docker/Dockerfile.rocm
@ -71,7 +71,7 @@ COPY --from=build_vllm ${COMMON_WORKDIR}/vllm /vllm-workspace
 RUN cd /vllm-workspace \
    && rm -rf vllm \
    && python3 -m pip install -e tests/vllm_test_utils \
-    && python3 -m pip install lm-eval[api]==0.4.4 \
+    && python3 -m pip install git+https://github.com/EleutherAI/lm-evaluation-harness.git@206b7722158f58c35b7ffcd53b035fdbdda5126d#egg=lm-eval[api] \
    && python3 -m pip install pytest-shard

 # -----------------------
--- a/docker/Dockerfile.s390x
+++ b/docker/Dockerfile.s390x
@ -16,7 +16,7 @@ ENV LANG=C.UTF-8 \
 RUN microdnf install -y \
    which procps findutils tar vim git gcc gcc-gfortran g++ make patch zlib-devel \
    libjpeg-turbo-devel libtiff-devel libpng-devel libwebp-devel freetype-devel harfbuzz-devel \
-    openssl-devel openblas openblas-devel autoconf automake libtool cmake numpy && \
+    openssl-devel openblas openblas-devel autoconf automake libtool cmake numpy libsndfile && \
    microdnf clean all

 # Python Installation
@ -136,6 +136,71 @@ RUN --mount=type=cache,target=/root/.cache/uv \
    mkdir -p /tmp/hf-xet/dist && \
    cp dist/*.whl /tmp/hf-xet/dist/

+# Build numba
+FROM python-install AS numba-builder
+
+ARG MAX_JOBS
+ARG NUMBA_VERSION=0.61.2
+
+WORKDIR /tmp
+
+# Clone all required dependencies
+RUN --mount=type=cache,target=/root/.cache/uv \
+    microdnf install ninja-build gcc gcc-c++ -y && \
+    git clone --recursive https://github.com/llvm/llvm-project.git -b llvmorg-15.0.7  && \
+    git clone --recursive https://github.com/numba/llvmlite.git -b v0.44.0 && \
+    git clone --recursive https://github.com/numba/numba.git -b ${NUMBA_VERSION} && \
+    cd llvm-project && mkdir build && cd  build && \
+    uv pip install 'cmake<4' setuptools numpy && \
+    export PREFIX=/usr/local && CMAKE_ARGS="${CMAKE_ARGS} -DLLVM_ENABLE_PROJECTS=lld;libunwind;compiler-rt" \
+    CFLAGS="$(echo $CFLAGS | sed 's/-fno-plt //g')" \
+    CXXFLAGS="$(echo $CXXFLAGS | sed 's/-fno-plt //g')" \
+    CMAKE_ARGS="${CMAKE_ARGS} -DFFI_INCLUDE_DIR=$PREFIX/include" \
+    CMAKE_ARGS="${CMAKE_ARGS} -DFFI_LIBRARY_DIR=$PREFIX/lib" \
+    cmake -DCMAKE_INSTALL_PREFIX="${PREFIX}"               \
+        -DCMAKE_BUILD_TYPE=Release                       \
+        -DCMAKE_LIBRARY_PATH="${PREFIX}"                 \
+        -DLLVM_ENABLE_LIBEDIT=OFF                        \
+        -DLLVM_ENABLE_LIBXML2=OFF                        \
+        -DLLVM_ENABLE_RTTI=ON                            \
+        -DLLVM_ENABLE_TERMINFO=OFF                       \
+        -DLLVM_INCLUDE_BENCHMARKS=OFF                    \
+        -DLLVM_INCLUDE_DOCS=OFF                          \
+        -DLLVM_INCLUDE_EXAMPLES=OFF                      \
+        -DLLVM_INCLUDE_GO_TESTS=OFF                      \
+        -DLLVM_INCLUDE_TESTS=OFF                         \
+        -DLLVM_INCLUDE_UTILS=ON                          \
+        -DLLVM_INSTALL_UTILS=ON                          \
+        -DLLVM_UTILS_INSTALL_DIR=libexec/llvm            \
+        -DLLVM_BUILD_LLVM_DYLIB=OFF                      \
+        -DLLVM_LINK_LLVM_DYLIB=OFF                       \
+        -DLLVM_EXPERIMENTAL_TARGETS_TO_BUILD=WebAssembly \
+        -DLLVM_ENABLE_FFI=ON                             \
+        -DLLVM_ENABLE_Z3_SOLVER=OFF                      \
+        -DLLVM_OPTIMIZED_TABLEGEN=ON                     \
+        -DCMAKE_POLICY_DEFAULT_CMP0111=NEW               \
+        -DCOMPILER_RT_BUILD_BUILTINS=ON                  \
+        -DCOMPILER_RT_BUILTINS_HIDE_SYMBOLS=OFF          \
+        -DCOMPILER_RT_BUILD_LIBFUZZER=OFF                \
+        -DCOMPILER_RT_BUILD_CRT=OFF                      \
+        -DCOMPILER_RT_BUILD_MEMPROF=OFF                  \
+        -DCOMPILER_RT_BUILD_PROFILE=OFF                  \
+        -DCOMPILER_RT_BUILD_SANITIZERS=OFF               \
+        -DCOMPILER_RT_BUILD_XRAY=OFF                     \
+        -DCOMPILER_RT_BUILD_GWP_ASAN=OFF                 \
+        -DCOMPILER_RT_BUILD_ORC=OFF                      \
+        -DCOMPILER_RT_INCLUDE_TESTS=OFF                  \
+        ${CMAKE_ARGS} -GNinja ../llvm                    \
+
+    && ninja install  . && \
+    #  build llvmlite
+    cd ../../llvmlite && python setup.py bdist_wheel && \
+    cd ../numba && \
+    if ! grep '#include "dynamic_annotations.h"' numba/_dispatcher.cpp; then \
+       sed -i '/#include "internal\/pycore_atomic.h"/i\#include "dynamic_annotations.h"' numba/_dispatcher.cpp; \
+    fi && python setup.py bdist_wheel
+
+
 # Final build stage
 FROM python-install AS vllm-cpu
 ARG PYTHON_VERSION
@ -163,23 +228,30 @@ RUN --mount=type=cache,target=/root/.cache/uv \
    --mount=type=bind,from=torch-vision,source=/tmp/vision/dist,target=/tmp/vision-wheels/ \
    --mount=type=bind,from=hf-xet-builder,source=/tmp/hf-xet/dist,target=/tmp/hf-xet-wheels/ \
    --mount=type=bind,from=torch,source=/tmp/pytorch/dist,target=/tmp/torch-wheels/ \
+    --mount=type=bind,from=numba-builder,source=/tmp/llvmlite/dist,target=/tmp/llvmlite-wheels/ \
+    --mount=type=bind,from=numba-builder,source=/tmp/numba/dist,target=/tmp/numba-wheels/ \
     sed -i '/^torch/d' requirements/build.txt && \
-     ARROW_WHL_FILE=$(ls /tmp/arrow-wheels/pyarrow-*.whl | head -n 1) && \
-     VISION_WHL_FILE=$(ls /tmp/vision-wheels/*.whl | head -n 1) && \
-     HF_XET_WHL_FILE=$(ls /tmp/hf-xet-wheels/*.whl | head -n 1) && \
-     TORCH_WHL_FILE=$(ls /tmp/torch-wheels/*.whl | head -n 1) && \
+     ARROW_WHL_FILE=$(ls /tmp/arrow-wheels/pyarrow-*.whl) && \
+     VISION_WHL_FILE=$(ls /tmp/vision-wheels/*.whl) && \
+     HF_XET_WHL_FILE=$(ls /tmp/hf-xet-wheels/*.whl) && \
+     TORCH_WHL_FILE=$(ls /tmp/torch-wheels/*.whl) && \
+     LLVM_WHL_FILE=$(ls /tmp/llvmlite-wheels/*.whl) && \
+     NUMBA_WHL_FILE=$(ls /tmp/numba-wheels/*.whl) && \
    uv pip install -v \    
        $ARROW_WHL_FILE  \
        $VISION_WHL_FILE \
        $HF_XET_WHL_FILE \
        $TORCH_WHL_FILE \
+        $LLVM_WHL_FILE \
+        $NUMBA_WHL_FILE \
        --index-strategy unsafe-best-match \
        -r requirements/build.txt \
-        -r requirements/cpu.txt 
+        -r requirements/cpu.txt
+

 # Build and install vllm
 RUN --mount=type=cache,target=/root/.cache/uv \
-    VLLM_TARGET_DEVICE=cpu python setup.py bdist_wheel && \
+    VLLM_TARGET_DEVICE=cpu VLLM_CPU_MOE_PREPACK=0 python setup.py bdist_wheel && \
    uv pip install "$(echo dist/*.whl)[tensorizer]"

 # setup non-root user for vllm
@ -196,4 +268,3 @@ WORKDIR /home/vllm

 # Set the default entrypoint
 ENTRYPOINT ["python", "-m", "vllm.entrypoints.openai.api_server"]
-
--- a/docker/Dockerfile.tpu
+++ b/docker/Dockerfile.tpu
@ -7,7 +7,8 @@ WORKDIR /workspace/vllm
 # Install some basic utilities
 RUN apt-get update && apt-get install -y \
    git \
-    ffmpeg libsm6 libxext6 libgl1
+    ffmpeg libsm6 libxext6 libgl1 && \
+    rm -rf /var/lib/apt/lists/*

 # Build vLLM.
 COPY . .
@ -16,6 +17,9 @@ RUN --mount=type=bind,source=.git,target=.git \
    if [ "$GIT_REPO_CHECK" != 0 ]; then bash tools/check_repo.sh; fi

 # Remove existing versions of dependencies
+# TODO: These packages will remain as dead weight in the Docker image layers.
+# We should find a way to build the image without uninstalling these.
+# Consider using a different base image.
 RUN pip uninstall -y torch torch_xla torchvision

 ENV VLLM_TARGET_DEVICE="tpu"
@ -23,9 +27,10 @@ RUN --mount=type=cache,target=/root/.cache/pip \
    --mount=type=bind,source=.git,target=.git \
    python3 -m pip install \
        -r requirements/tpu.txt
-RUN python3 -m pip install -e .
+
+RUN --mount=type=cache,target=/root/.cache/pip python3 -m pip install -e .

 # install development dependencies (for testing)
-RUN python3 -m pip install -e tests/vllm_test_utils
+RUN --mount=type=cache,target=/root/.cache/pip python3 -m pip install -e tests/vllm_test_utils

 CMD ["/bin/bash"]
--- a/Show More
+++ b/Show More