Merge branch 'main' into wye-refactor-quant-folder

Remove redundant assignment in xfer_buffers, This is a little fix (#24732 )
Signed-off-by: ChenTaoyu-SJTU <ctynb@qq.com>
2025-10-20 14:53:52 +08:00 · 2025-09-13 07:38:47 -07:00 · 2025-09-13 08:11:59 +00:00 · 2025-09-13 15:45:11 +08:00 · 2025-09-13 00:43:33 -07:00 · 2025-09-13 00:34:53 -07:00
1339 changed files with 89234 additions and 40394 deletions
--- a/.buildkite/check-wheel-size.py
+++ b/.buildkite/check-wheel-size.py
@ -5,11 +5,11 @@ import os
 import sys
 import zipfile

-# Read the VLLM_MAX_SIZE_MB environment variable, defaulting to 400 MiB
-# Note that we have 400 MiB quota, please use it wisely.
-# See https://github.com/pypi/support/issues/3792 .
+# Read the VLLM_MAX_SIZE_MB environment variable, defaulting to 450 MiB
+# Note that we have 800 MiB quota, please use it wisely.
+# See https://github.com/pypi/support/issues/6326 .
 # Please also sync the value with the one in Dockerfile.
-VLLM_MAX_SIZE_MB = int(os.environ.get("VLLM_MAX_SIZE_MB", 400))
+VLLM_MAX_SIZE_MB = int(os.environ.get("VLLM_MAX_SIZE_MB", 450))


 def print_top_10_largest_files(zip_file):
--- 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
@ -141,7 +141,7 @@ When run, benchmark script generates results under `benchmark/results` folder, a
 `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 with Model, Dataset name, input/output lenght, max concurrency and qps.
+Here is an example using the script to compare result_a and result_b with Model, Dataset name, input/output length, max concurrency and qps.
 `python3 compare-json-results.py -f results_a/benchmark_results.json -f results_b/benchmark_results.json`

 |   | Model | Dataset Name | Input Len | Output Len | # of max concurrency | qps  | results_a/benchmark_results.json | results_b/benchmark_results.json | perf_ratio        |
--- 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
@ -3,44 +3,129 @@
 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, info_cols, drop_column, debug=False
 ):
-    print("\ncompare_data_column: " + data_column)
+    """
+    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:
-        data_df = pd.read_json(file)
-        serving_df = data_df.dropna(subset=[drop_column], ignore_index=True)
-        # Show all info columns in the first couple columns
-        if not frames:
-            for col in info_cols:
-                if col not in serving_df.columns:
-                    print(f"Skipping missing column: {col}")
-                    continue
-                frames.append(serving_df[col])
-        # only show test name under debug mode
-        if debug is True:
-            serving_df = serving_df.rename(columns={name_column: file + "_name"})
-            frames.append(serving_df[file + "_name"])
+        df = pd.read_json(file, orient="records")

-        file = "/".join(file.split("/")[:-1])
-        serving_df = serving_df.rename(columns={data_column: file})
-        frames.append(serving_df[file])
-        raw_data_cols.append(file)
-        compare_frames.append(serving_df[file])
+        # 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:
-            # Compare numbers among two files
-            ratio_df = compare_frames[1] / compare_frames[0]
-            frames.append(ratio_df)
-            compare_frames.pop(1)
+            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)
+    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

@ -67,6 +152,15 @@ def split_json_by_tp_pp(

    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",
@ -124,7 +218,7 @@ if __name__ == "__main__":
        "--xaxis",
        type=str,
        default="# of max concurrency.",
-        help="column name to use as X Axis in comparision graph",
+        help="column name to use as X Axis in comparison graph",
    )
    args = parser.parse_args()

@ -181,7 +275,6 @@ if __name__ == "__main__":
                    f"Expected subset: {filtered_info_cols}, "
                    f"but DataFrame has: {list(output_df.columns)}"
                )
-
            output_df_sorted = output_df.sort_values(by=existing_group_cols)
            output_groups = output_df_sorted.groupby(existing_group_cols, dropna=False)
            for name, group in output_groups:
@ -189,8 +282,7 @@ if __name__ == "__main__":
                text_file.write(html_msgs_for_data_cols[i])
                text_file.write(html)

-                if plot is True:
-                    import pandas as pd
+                if plot and plotly_found:
                    import plotly.express as px

                    df = group[raw_data_cols]
--- 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/tests/serving-tests-cpu-snc2.json
+++ b/.buildkite/nightly-benchmarks/tests/serving-tests-cpu-snc2.json
@ -1,6 +1,6 @@
 [
    {
-        "test_name": "serving_llama8B_tp1_sharegpt",
+        "test_name": "serving_llama8B_bf16_tp1_sharegpt",
        "qps_list": ["inf"],
        "max_concurrency_list": [12, 16, 24, 32, 64, 128, 200],
        "server_environment_variables": {
@ -32,7 +32,7 @@
        }
    },
    {
-        "test_name": "serving_llama8B_tp2_sharegpt",
+        "test_name": "serving_llama8B_bf16_tp2_sharegpt",
        "qps_list": ["inf"],
        "max_concurrency_list": [12, 16, 24, 32, 64, 128, 200],
        "server_environment_variables": {
@ -64,7 +64,7 @@
        }
    },
    {
-        "test_name": "serving_llama8B_tp4_sharegpt",
+        "test_name": "serving_llama8B_bf16_tp4_sharegpt",
        "qps_list": ["inf"],
        "max_concurrency_list": [12, 16, 24, 32, 64, 128, 200],
        "server_environment_variables": {
@ -96,7 +96,7 @@
        }
    },
    {
-        "test_name": "serving_llama8B_tp1_random_128_128",
+        "test_name": "serving_llama8B_bf16_tp1_random_128_128",
        "qps_list": ["inf"],
        "max_concurrency_list": [12, 16, 24, 32, 64, 128, 200, 1000],
        "server_environment_variables": {
@ -131,7 +131,7 @@
        }
    },
    {
-        "test_name": "serving_llama8B_tp2_random_128_128",
+        "test_name": "serving_llama8B_bf16_tp2_random_128_128",
        "qps_list": ["inf"],
        "max_concurrency_list": [12, 16, 24, 32, 64, 128, 200, 1000],
        "server_environment_variables": {
@ -166,7 +166,7 @@
        }
    },
    {
-        "test_name": "serving_llama8B_tp4_random_128_128",
+        "test_name": "serving_llama8B_bf16_tp4_random_128_128",
        "qps_list": ["inf"],
        "max_concurrency_list": [12, 16, 24, 32, 64, 128, 200, 1000],
        "server_environment_variables": {
@ -198,5 +198,413 @@
 	    "random-output-len": 128,
            "num_prompts": 1000
        }
+    },
+    {
+        "test_name": "serving_llama8B_int8_tp1_sharegpt",
+        "qps_list": ["inf"],
+        "max_concurrency_list": [12, 16, 24, 32, 64, 128, 200],
+        "server_environment_variables": {
+            "VLLM_RPC_TIMEOUT": 100000,
+	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
+	    "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
+	    "VLLM_CPU_SGL_KERNEL": 1,
+	    "VLLM_CPU_KVCACHE_SPACE": 40
+        },
+        "server_parameters": {
+            "model": "RedHatAI/Meta-Llama-3.1-8B-Instruct-quantized.w8a8",
+            "tensor_parallel_size": 1,
+	    "dtype": "bfloat16",
+	    "distributed_executor_backend": "mp",
+	    "block_size": 128,
+	    "trust_remote_code": "",
+            "disable_log_stats": "",
+	    "enforce_eager": "",
+	    "max_num_batched_tokens": 2048,
+	    "max_num_seqs": 256,
+            "load_format": "dummy"
+        },
+        "client_parameters": {
+            "model": "RedHatAI/Meta-Llama-3.1-8B-Instruct-quantized.w8a8",
+            "backend": "vllm",
+            "dataset_name": "sharegpt",
+            "dataset_path": "./ShareGPT_V3_unfiltered_cleaned_split.json",
+            "num_prompts": 200
+        }
+    },
+    {
+        "test_name": "serving_llama8B_int8_tp2_sharegpt",
+        "qps_list": ["inf"],
+        "max_concurrency_list": [12, 16, 24, 32, 64, 128, 200],
+        "server_environment_variables": {
+            "VLLM_RPC_TIMEOUT": 100000,
+	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
+	    "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
+	    "VLLM_CPU_SGL_KERNEL": 1,
+	    "VLLM_CPU_KVCACHE_SPACE": 40
+        },
+        "server_parameters": {
+            "model": "RedHatAI/Meta-Llama-3.1-8B-Instruct-quantized.w8a8",
+            "tensor_parallel_size": 2,
+	    "dtype": "bfloat16",
+	    "distributed_executor_backend": "mp",
+	    "block_size": 128,
+	    "trust_remote_code": "",
+            "disable_log_stats": "",
+	    "enforce_eager": "",
+	    "max_num_batched_tokens": 2048,
+	    "max_num_seqs": 256,
+            "load_format": "dummy"
+        },
+        "client_parameters": {
+            "model": "RedHatAI/Meta-Llama-3.1-8B-Instruct-quantized.w8a8",
+            "backend": "vllm",
+            "dataset_name": "sharegpt",
+            "dataset_path": "./ShareGPT_V3_unfiltered_cleaned_split.json",
+            "num_prompts": 200
+        }
+    },
+    {
+        "test_name": "serving_llama8B_int8_tp4_sharegpt",
+        "qps_list": ["inf"],
+        "max_concurrency_list": [12, 16, 24, 32, 64, 128, 200],
+        "server_environment_variables": {
+            "VLLM_RPC_TIMEOUT": 100000,
+	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
+	    "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
+	    "VLLM_CPU_SGL_KERNEL": 1,
+	    "VLLM_CPU_KVCACHE_SPACE": 40
+        },
+        "server_parameters": {
+            "model": "RedHatAI/Meta-Llama-3.1-8B-Instruct-quantized.w8a8",
+            "tensor_parallel_size": 4,
+	    "dtype": "bfloat16",
+	    "distributed_executor_backend": "mp",
+	    "block_size": 128,
+	    "trust_remote_code": "",
+            "disable_log_stats": "",
+	    "enforce_eager": "",
+	    "max_num_batched_tokens": 2048,
+	    "max_num_seqs": 256,
+            "load_format": "dummy"
+        },
+        "client_parameters": {
+            "model": "RedHatAI/Meta-Llama-3.1-8B-Instruct-quantized.w8a8",
+            "backend": "vllm",
+            "dataset_name": "sharegpt",
+            "dataset_path": "./ShareGPT_V3_unfiltered_cleaned_split.json",
+            "num_prompts": 200
+        }
+    },
+    {
+        "test_name": "serving_llama8B_int8_tp1_random_128_128",
+        "qps_list": ["inf"],
+        "max_concurrency_list": [12, 16, 24, 32, 64, 128, 200, 1000],
+        "server_environment_variables": {
+            "VLLM_RPC_TIMEOUT": 100000,
+	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
+	    "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
+	    "VLLM_CPU_SGL_KERNEL": 1,
+	    "VLLM_CPU_KVCACHE_SPACE": 40
+        },
+        "server_parameters": {
+            "model": "RedHatAI/Meta-Llama-3.1-8B-Instruct-quantized.w8a8",
+            "tensor_parallel_size": 1,
+	    "dtype": "bfloat16",
+	    "distributed_executor_backend": "mp",
+	    "block_size": 128,
+	    "trust_remote_code": "",
+	    "enable_chunked_prefill": "",
+            "disable_log_stats": "",
+	    "enforce_eager": "",
+	    "max_num_batched_tokens": 2048,
+	    "max_num_seqs": 256,
+            "load_format": "dummy"
+        },
+        "client_parameters": {
+            "model": "RedHatAI/Meta-Llama-3.1-8B-Instruct-quantized.w8a8",
+            "backend": "vllm",
+            "dataset_name": "random",
+	    "random-input-len": 128,
+	    "random-output-len": 128,
+	    "ignore-eos": "",
+            "num_prompts": 1000
+        }
+    },
+    {
+        "test_name": "serving_llama8B_int8_tp2_random_128_128",
+        "qps_list": ["inf"],
+        "max_concurrency_list": [12, 16, 24, 32, 64, 128, 200, 1000],
+        "server_environment_variables": {
+            "VLLM_RPC_TIMEOUT": 100000,
+	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
+	    "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
+	    "VLLM_CPU_SGL_KERNEL": 1,
+	    "VLLM_CPU_KVCACHE_SPACE": 40
+        },
+        "server_parameters": {
+            "model": "RedHatAI/Meta-Llama-3.1-8B-Instruct-quantized.w8a8",
+            "tensor_parallel_size": 2,
+	    "dtype": "bfloat16",
+	    "distributed_executor_backend": "mp",
+	    "block_size": 128,
+	    "trust_remote_code": "",
+	    "enable_chunked_prefill": "",
+            "disable_log_stats": "",
+	    "enforce_eager": "",
+	    "max_num_batched_tokens": 2048,
+	    "max_num_seqs": 256,
+            "load_format": "dummy"
+        },
+        "client_parameters": {
+            "model": "RedHatAI/Meta-Llama-3.1-8B-Instruct-quantized.w8a8",
+            "backend": "vllm",
+            "dataset_name": "random",
+	    "random-input-len": 128,
+	    "random-output-len": 128,
+	    "ignore-eos": "",
+            "num_prompts": 1000
+        }
+    },
+    {
+        "test_name": "serving_llama8B_int8_tp4_random_128_128",
+        "qps_list": ["inf"],
+        "max_concurrency_list": [12, 16, 24, 32, 64, 128, 200, 1000],
+        "server_environment_variables": {
+            "VLLM_RPC_TIMEOUT": 100000,
+	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
+	    "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
+	    "VLLM_CPU_SGL_KERNEL": 1,
+	    "VLLM_CPU_KVCACHE_SPACE": 40
+        },
+        "server_parameters": {
+            "model": "RedHatAI/Meta-Llama-3.1-8B-Instruct-quantized.w8a8",
+            "tensor_parallel_size": 4,
+	    "dtype": "bfloat16",
+	    "distributed_executor_backend": "mp",
+	    "block_size": 128,
+	    "trust_remote_code": "",
+	    "enable_chunked_prefill": "",
+            "disable_log_stats": "",
+	    "enforce_eager": "",
+	    "max_num_batched_tokens": 2048,
+	    "max_num_seqs": 256,
+            "load_format": "dummy"
+        },
+        "client_parameters": {
+            "model": "RedHatAI/Meta-Llama-3.1-8B-Instruct-quantized.w8a8",
+            "backend": "vllm",
+            "dataset_name": "random",
+	    "random-input-len": 128,
+	    "random-output-len": 128,
+	    "ignore-eos": "",
+            "num_prompts": 1000
+        }
+    },
+    {
+        "test_name": "serving_llama8B_int4_tp1_sharegpt",
+        "qps_list": ["inf"],
+        "max_concurrency_list": [12, 16, 24, 32, 64, 128, 200],
+        "server_environment_variables": {
+            "VLLM_RPC_TIMEOUT": 100000,
+	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
+	    "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
+	    "VLLM_CPU_SGL_KERNEL": 1,
+	    "VLLM_CPU_KVCACHE_SPACE": 40
+        },
+        "server_parameters": {
+            "model": "hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4",
+	    "quantization": "awq",
+            "tensor_parallel_size": 1,
+	    "dtype": "bfloat16",
+	    "distributed_executor_backend": "mp",
+	    "block_size": 128,
+	    "trust_remote_code": "",
+            "disable_log_stats": "",
+	    "enforce_eager": "",
+	    "max_num_batched_tokens": 2048,
+	    "max_num_seqs": 256,
+            "load_format": "dummy"
+        },
+        "client_parameters": {
+            "model": "hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4",
+            "backend": "vllm",
+            "dataset_name": "sharegpt",
+            "dataset_path": "./ShareGPT_V3_unfiltered_cleaned_split.json",
+            "num_prompts": 200
+        }
+    },
+    {
+        "test_name": "serving_llama8B_int4_tp2_sharegpt",
+        "qps_list": ["inf"],
+        "max_concurrency_list": [12, 16, 24, 32, 64, 128, 200],
+        "server_environment_variables": {
+            "VLLM_RPC_TIMEOUT": 100000,
+	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
+	    "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
+	    "VLLM_CPU_SGL_KERNEL": 1,
+	    "VLLM_CPU_KVCACHE_SPACE": 40
+        },
+        "server_parameters": {
+            "model": "hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4",
+	    "quantization": "awq",
+            "tensor_parallel_size": 2,
+	    "dtype": "bfloat16",
+	    "distributed_executor_backend": "mp",
+	    "block_size": 128,
+	    "trust_remote_code": "",
+            "disable_log_stats": "",
+	    "enforce_eager": "",
+	    "max_num_batched_tokens": 2048,
+	    "max_num_seqs": 256,
+            "load_format": "dummy"
+        },
+        "client_parameters": {
+            "model": "hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4",
+            "backend": "vllm",
+            "dataset_name": "sharegpt",
+            "dataset_path": "./ShareGPT_V3_unfiltered_cleaned_split.json",
+            "num_prompts": 200
+        }
+    },
+    {
+        "test_name": "serving_llama8B_int4_tp4_sharegpt",
+        "qps_list": ["inf"],
+        "max_concurrency_list": [12, 16, 24, 32, 64, 128, 200],
+        "server_environment_variables": {
+            "VLLM_RPC_TIMEOUT": 100000,
+	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
+	    "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
+	    "VLLM_CPU_SGL_KERNEL": 1,
+	    "VLLM_CPU_KVCACHE_SPACE": 40
+        },
+        "server_parameters": {
+            "model": "hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4",
+	    "quantization": "awq",
+            "tensor_parallel_size": 4,
+	    "dtype": "bfloat16",
+	    "distributed_executor_backend": "mp",
+	    "block_size": 128,
+	    "trust_remote_code": "",
+            "disable_log_stats": "",
+	    "enforce_eager": "",
+	    "max_num_batched_tokens": 2048,
+	    "max_num_seqs": 256,
+            "load_format": "dummy"
+        },
+        "client_parameters": {
+            "model": "hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4",
+            "backend": "vllm",
+            "dataset_name": "sharegpt",
+            "dataset_path": "./ShareGPT_V3_unfiltered_cleaned_split.json",
+            "num_prompts": 200
+        }
+    },
+    {
+        "test_name": "serving_llama8B_int4_tp1_random_128_128",
+        "qps_list": ["inf"],
+        "max_concurrency_list": [12, 16, 24, 32, 64, 128, 200, 1000],
+        "server_environment_variables": {
+            "VLLM_RPC_TIMEOUT": 100000,
+	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
+	    "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
+	    "VLLM_CPU_SGL_KERNEL": 1,
+	    "VLLM_CPU_KVCACHE_SPACE": 40
+        },
+        "server_parameters": {
+            "model": "hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4",
+	    "quantization": "awq",
+            "tensor_parallel_size": 1,
+	    "dtype": "bfloat16",
+	    "distributed_executor_backend": "mp",
+	    "block_size": 128,
+	    "trust_remote_code": "",
+	    "enable_chunked_prefill": "",
+            "disable_log_stats": "",
+	    "enforce_eager": "",
+	    "max_num_batched_tokens": 2048,
+	    "max_num_seqs": 256,
+            "load_format": "dummy"
+        },
+        "client_parameters": {
+            "model": "hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4",
+            "backend": "vllm",
+            "dataset_name": "random",
+	    "random-input-len": 128,
+	    "random-output-len": 128,
+	    "ignore-eos": "",
+            "num_prompts": 1000
+        }
+    },
+    {
+        "test_name": "serving_llama8B_int4_tp2_random_128_128",
+        "qps_list": ["inf"],
+        "max_concurrency_list": [12, 16, 24, 32, 64, 128, 200, 1000],
+        "server_environment_variables": {
+            "VLLM_RPC_TIMEOUT": 100000,
+	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
+	    "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
+	    "VLLM_CPU_SGL_KERNEL": 1,
+	    "VLLM_CPU_KVCACHE_SPACE": 40
+        },
+        "server_parameters": {
+            "model": "hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4",
+	    "quantization": "awq",
+            "tensor_parallel_size": 2,
+	    "dtype": "bfloat16",
+	    "distributed_executor_backend": "mp",
+	    "block_size": 128,
+	    "trust_remote_code": "",
+	    "enable_chunked_prefill": "",
+            "disable_log_stats": "",
+	    "enforce_eager": "",
+	    "max_num_batched_tokens": 2048,
+	    "max_num_seqs": 256,
+            "load_format": "dummy"
+        },
+        "client_parameters": {
+            "model": "hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4",
+            "backend": "vllm",
+            "dataset_name": "random",
+	    "random-input-len": 128,
+	    "random-output-len": 128,
+	    "ignore-eos": "",
+            "num_prompts": 1000
+        }
+    },
+    {
+        "test_name": "serving_llama8B_int4_tp4_random_128_128",
+        "qps_list": ["inf"],
+        "max_concurrency_list": [12, 16, 24, 32, 64, 128, 200, 1000],
+        "server_environment_variables": {
+            "VLLM_RPC_TIMEOUT": 100000,
+	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
+	    "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
+	    "VLLM_CPU_SGL_KERNEL": 1,
+	    "VLLM_CPU_KVCACHE_SPACE": 40
+        },
+        "server_parameters": {
+            "model": "hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4",
+	    "quantization": "awq",
+            "tensor_parallel_size": 4,
+	    "dtype": "bfloat16",
+	    "distributed_executor_backend": "mp",
+	    "block_size": 128,
+	    "trust_remote_code": "",
+	    "enable_chunked_prefill": "",
+            "disable_log_stats": "",
+	    "enforce_eager": "",
+	    "max_num_batched_tokens": 2048,
+	    "max_num_seqs": 256,
+            "load_format": "dummy"
+        },
+        "client_parameters": {
+            "model": "hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4",
+            "backend": "vllm",
+            "dataset_name": "random",
+	    "random-input-len": 128,
+	    "random-output-len": 128,
+	    "ignore-eos": "",
+            "num_prompts": 1000
+        }
    }
 ]
--- a/.buildkite/nightly-benchmarks/tests/serving-tests-cpu-snc3.json
+++ b/.buildkite/nightly-benchmarks/tests/serving-tests-cpu-snc3.json
@ -1,6 +1,6 @@
 [
    {
-        "test_name": "serving_llama8B_pp1_sharegpt",
+        "test_name": "serving_llama8B_bf16_pp1_sharegpt",
        "qps_list": ["inf"],
        "max_concurrency_list": [12, 16, 24, 32, 64, 128, 200],
        "server_environment_variables": {
@ -32,7 +32,39 @@
        }
    },
    {
-        "test_name": "serving_llama8B_pp3_sharegpt",
+        "test_name": "serving_llama8B_bf16_tp2_sharegpt",
+        "qps_list": ["inf"],
+        "max_concurrency_list": [12, 16, 24, 32, 64, 128, 200],
+        "server_environment_variables": {
+            "VLLM_RPC_TIMEOUT": 100000,
+	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
+	    "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
+	    "VLLM_CPU_SGL_KERNEL": 1,
+	    "VLLM_CPU_KVCACHE_SPACE": 40
+        },
+        "server_parameters": {
+            "model": "meta-llama/Llama-3.1-8B-Instruct",
+            "tensor_parallel_size": 2,
+	    "dtype": "bfloat16",
+	    "distributed_executor_backend": "mp",
+	    "block_size": 128,
+	    "trust_remote_code": "",
+            "disable_log_stats": "",
+	    "enforce_eager": "",
+	    "max_num_batched_tokens": 2048,
+	    "max_num_seqs": 256,
+            "load_format": "dummy"
+        },
+        "client_parameters": {
+            "model": "meta-llama/Llama-3.1-8B-Instruct",
+            "backend": "vllm",
+            "dataset_name": "sharegpt",
+            "dataset_path": "./ShareGPT_V3_unfiltered_cleaned_split.json",
+            "num_prompts": 200
+        }
+    },
+    {
+        "test_name": "serving_llama8B_bf16_pp3_sharegpt",
        "qps_list": ["inf"],
        "max_concurrency_list": [12, 16, 24, 32, 64, 128, 200],
        "server_environment_variables": {
@ -64,7 +96,7 @@
        }
    },
    {
-        "test_name": "serving_llama8B_tp2pp3_sharegpt",
+        "test_name": "serving_llama8B_bf16_tp2pp3_sharegpt",
        "qps_list": ["inf"],
        "max_concurrency_list": [12, 16, 24, 32, 64, 128, 200],
        "server_environment_variables": {
@ -97,7 +129,7 @@
        }
    },
    {
-        "test_name": "serving_llama8B_pp1_random_128_128",
+        "test_name": "serving_llama8B_bf16_pp1_random_128_128",
        "qps_list": ["inf"],
        "max_concurrency_list": [12, 16, 24, 32, 64, 128, 200, 1000],
        "server_environment_variables": {
@ -132,7 +164,42 @@
        }
    },
    {
-        "test_name": "serving_llama8B_pp3_random_128_128",
+        "test_name": "serving_llama8B_bf16_tp2_random_128_128",
+        "qps_list": ["inf"],
+        "max_concurrency_list": [12, 16, 24, 32, 64, 128, 200, 1000],
+        "server_environment_variables": {
+            "VLLM_RPC_TIMEOUT": 100000,
+	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
+	    "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
+	    "VLLM_CPU_SGL_KERNEL": 1,
+	    "VLLM_CPU_KVCACHE_SPACE": 40
+        },
+        "server_parameters": {
+            "model": "meta-llama/Llama-3.1-8B-Instruct",
+            "tensor_parallel_size": 2,
+	    "dtype": "bfloat16",
+	    "distributed_executor_backend": "mp",
+	    "block_size": 128,
+	    "trust_remote_code": "",
+	    "enable_chunked_prefill": "",
+            "disable_log_stats": "",
+	    "enforce_eager": "",
+	    "max_num_batched_tokens": 2048,
+	    "max_num_seqs": 256,
+            "load_format": "dummy"
+        },
+        "client_parameters": {
+            "model": "meta-llama/Llama-3.1-8B-Instruct",
+            "backend": "vllm",
+            "dataset_name": "random",
+	    "random-input-len": 128,
+	    "random-output-len": 128,
+	    "ignore-eos": "",
+            "num_prompts": 1000
+        }
+    },
+    {
+        "test_name": "serving_llama8B_bf16_pp3_random_128_128",
        "qps_list": ["inf"],
        "max_concurrency_list": [12, 16, 24, 32, 64, 128, 200, 1000],
        "server_environment_variables": {
@ -167,7 +234,7 @@
        }
    },
    {
-        "test_name": "serving_llama8B_tp2pp3_random_128_128",
+        "test_name": "serving_llama8B_bf16_tp2pp3_random_128_128",
        "qps_list": ["inf"],
        "max_concurrency_list": [12, 16, 24, 32, 64, 128, 200, 1000],
        "server_environment_variables": {
@ -201,5 +268,553 @@
 	    "ignore-eos": "",
            "num_prompts": 1000
        }
+    },
+    {
+        "test_name": "serving_llama8B_int8_pp1_sharegpt",
+        "qps_list": ["inf"],
+        "max_concurrency_list": [12, 16, 24, 32, 64, 128, 200],
+        "server_environment_variables": {
+            "VLLM_RPC_TIMEOUT": 100000,
+	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
+	    "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
+	    "VLLM_CPU_SGL_KERNEL": 1,
+	    "VLLM_CPU_KVCACHE_SPACE": 40
+        },
+        "server_parameters": {
+            "model": "RedHatAI/Meta-Llama-3.1-8B-Instruct-quantized.w8a8",
+            "pipeline_parallel_size": 1,
+	    "dtype": "bfloat16",
+	    "distributed_executor_backend": "mp",
+	    "block_size": 128,
+	    "trust_remote_code": "",
+            "disable_log_stats": "",
+	    "enforce_eager": "",
+	    "max_num_batched_tokens": 2048,
+	    "max_num_seqs": 256,
+            "load_format": "dummy"
+        },
+        "client_parameters": {
+            "model": "RedHatAI/Meta-Llama-3.1-8B-Instruct-quantized.w8a8",
+            "backend": "vllm",
+            "dataset_name": "sharegpt",
+            "dataset_path": "./ShareGPT_V3_unfiltered_cleaned_split.json",
+            "num_prompts": 200
+        }
+    },
+    {
+        "test_name": "serving_llama8B_int8_tp2_sharegpt",
+        "qps_list": ["inf"],
+        "max_concurrency_list": [12, 16, 24, 32, 64, 128, 200],
+        "server_environment_variables": {
+            "VLLM_RPC_TIMEOUT": 100000,
+	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
+	    "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
+	    "VLLM_CPU_SGL_KERNEL": 1,
+	    "VLLM_CPU_KVCACHE_SPACE": 40
+        },
+        "server_parameters": {
+            "model": "RedHatAI/Meta-Llama-3.1-8B-Instruct-quantized.w8a8",
+            "tensor_parallel_size": 2,
+	    "dtype": "bfloat16",
+	    "distributed_executor_backend": "mp",
+	    "block_size": 128,
+	    "trust_remote_code": "",
+            "disable_log_stats": "",
+	    "enforce_eager": "",
+	    "max_num_batched_tokens": 2048,
+	    "max_num_seqs": 256,
+            "load_format": "dummy"
+        },
+        "client_parameters": {
+            "model": "RedHatAI/Meta-Llama-3.1-8B-Instruct-quantized.w8a8",
+            "backend": "vllm",
+            "dataset_name": "sharegpt",
+            "dataset_path": "./ShareGPT_V3_unfiltered_cleaned_split.json",
+            "num_prompts": 200
+        }
+    },
+    {
+        "test_name": "serving_llama8B_int8_pp3_sharegpt",
+        "qps_list": ["inf"],
+        "max_concurrency_list": [12, 16, 24, 32, 64, 128, 200],
+        "server_environment_variables": {
+            "VLLM_RPC_TIMEOUT": 100000,
+	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
+	    "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
+	    "VLLM_CPU_SGL_KERNEL": 1,
+	    "VLLM_CPU_KVCACHE_SPACE": 40
+        },
+        "server_parameters": {
+            "model": "RedHatAI/Meta-Llama-3.1-8B-Instruct-quantized.w8a8",
+            "pipeline_parallel_size": 3,
+	    "dtype": "bfloat16",
+	    "distributed_executor_backend": "mp",
+	    "block_size": 128,
+	    "trust_remote_code": "",
+            "disable_log_stats": "",
+	    "enforce_eager": "",
+	    "max_num_batched_tokens": 2048,
+	    "max_num_seqs": 256,
+            "load_format": "dummy"
+        },
+        "client_parameters": {
+            "model": "RedHatAI/Meta-Llama-3.1-8B-Instruct-quantized.w8a8",
+            "backend": "vllm",
+            "dataset_name": "sharegpt",
+            "dataset_path": "./ShareGPT_V3_unfiltered_cleaned_split.json",
+            "num_prompts": 200
+        }
+    },
+    {
+        "test_name": "serving_llama8B_int8_tp2pp3_sharegpt",
+        "qps_list": ["inf"],
+        "max_concurrency_list": [12, 16, 24, 32, 64, 128, 200],
+        "server_environment_variables": {
+            "VLLM_RPC_TIMEOUT": 100000,
+	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
+	    "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
+	    "VLLM_CPU_SGL_KERNEL": 1,
+	    "VLLM_CPU_KVCACHE_SPACE": 40
+        },
+        "server_parameters": {
+            "model": "RedHatAI/Meta-Llama-3.1-8B-Instruct-quantized.w8a8",
+            "tensor_parallel_size": 2,
+            "pipeline_parallel_size": 3,
+	    "dtype": "bfloat16",
+	    "distributed_executor_backend": "mp",
+	    "block_size": 128,
+	    "trust_remote_code": "",
+            "disable_log_stats": "",
+	    "enforce_eager": "",
+	    "max_num_batched_tokens": 2048,
+	    "max_num_seqs": 256,
+            "load_format": "dummy"
+        },
+        "client_parameters": {
+            "model": "RedHatAI/Meta-Llama-3.1-8B-Instruct-quantized.w8a8",
+            "backend": "vllm",
+            "dataset_name": "sharegpt",
+            "dataset_path": "./ShareGPT_V3_unfiltered_cleaned_split.json",
+            "num_prompts": 200
+        }
+    },
+    {
+        "test_name": "serving_llama8B_int8_pp1_random_128_128",
+        "qps_list": ["inf"],
+        "max_concurrency_list": [12, 16, 24, 32, 64, 128, 200, 1000],
+        "server_environment_variables": {
+            "VLLM_RPC_TIMEOUT": 100000,
+	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
+	    "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
+	    "VLLM_CPU_SGL_KERNEL": 1,
+	    "VLLM_CPU_KVCACHE_SPACE": 40
+        },
+        "server_parameters": {
+            "model": "RedHatAI/Meta-Llama-3.1-8B-Instruct-quantized.w8a8",
+            "pipeline_parallel_size": 1,
+	    "dtype": "bfloat16",
+	    "distributed_executor_backend": "mp",
+	    "block_size": 128,
+	    "trust_remote_code": "",
+	    "enable_chunked_prefill": "",
+            "disable_log_stats": "",
+	    "enforce_eager": "",
+	    "max_num_batched_tokens": 2048,
+	    "max_num_seqs": 256,
+            "load_format": "dummy"
+        },
+        "client_parameters": {
+            "model": "RedHatAI/Meta-Llama-3.1-8B-Instruct-quantized.w8a8",
+            "backend": "vllm",
+            "dataset_name": "random",
+	    "random-input-len": 128,
+	    "random-output-len": 128,
+	    "ignore-eos": "",
+            "num_prompts": 1000
+        }
+    },
+    {
+        "test_name": "serving_llama8B_int8_tp2_random_128_128",
+        "qps_list": ["inf"],
+        "max_concurrency_list": [12, 16, 24, 32, 64, 128, 200, 1000],
+        "server_environment_variables": {
+            "VLLM_RPC_TIMEOUT": 100000,
+	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
+	    "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
+	    "VLLM_CPU_SGL_KERNEL": 1,
+	    "VLLM_CPU_KVCACHE_SPACE": 40
+        },
+        "server_parameters": {
+            "model": "RedHatAI/Meta-Llama-3.1-8B-Instruct-quantized.w8a8",
+            "tensor_parallel_size": 2,
+	    "dtype": "bfloat16",
+	    "distributed_executor_backend": "mp",
+	    "block_size": 128,
+	    "trust_remote_code": "",
+	    "enable_chunked_prefill": "",
+            "disable_log_stats": "",
+	    "enforce_eager": "",
+	    "max_num_batched_tokens": 2048,
+	    "max_num_seqs": 256,
+            "load_format": "dummy"
+        },
+        "client_parameters": {
+            "model": "RedHatAI/Meta-Llama-3.1-8B-Instruct-quantized.w8a8",
+            "backend": "vllm",
+            "dataset_name": "random",
+	    "random-input-len": 128,
+	    "random-output-len": 128,
+	    "ignore-eos": "",
+            "num_prompts": 1000
+        }
+    },
+    {
+        "test_name": "serving_llama8B_int8_pp3_random_128_128",
+        "qps_list": ["inf"],
+        "max_concurrency_list": [12, 16, 24, 32, 64, 128, 200, 1000],
+        "server_environment_variables": {
+            "VLLM_RPC_TIMEOUT": 100000,
+	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
+	    "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
+	    "VLLM_CPU_SGL_KERNEL": 1,
+	    "VLLM_CPU_KVCACHE_SPACE": 40
+        },
+        "server_parameters": {
+            "model": "RedHatAI/Meta-Llama-3.1-8B-Instruct-quantized.w8a8",
+            "pipeline_parallel_size": 3,
+	    "dtype": "bfloat16",
+	    "distributed_executor_backend": "mp",
+	    "block_size": 128,
+	    "trust_remote_code": "",
+	    "enable_chunked_prefill": "",
+            "disable_log_stats": "",
+	    "enforce_eager": "",
+	    "max_num_batched_tokens": 2048,
+	    "max_num_seqs": 256,
+            "load_format": "dummy"
+        },
+        "client_parameters": {
+            "model": "RedHatAI/Meta-Llama-3.1-8B-Instruct-quantized.w8a8",
+            "backend": "vllm",
+            "dataset_name": "random",
+	    "random-input-len": 128,
+	    "random-output-len": 128,
+	    "ignore-eos": "",
+            "num_prompts": 1000
+        }
+    },
+    {
+        "test_name": "serving_llama8B_int8_tp2pp3_random_128_128",
+        "qps_list": ["inf"],
+        "max_concurrency_list": [12, 16, 24, 32, 64, 128, 200, 1000],
+        "server_environment_variables": {
+            "VLLM_RPC_TIMEOUT": 100000,
+	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
+	    "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
+	    "VLLM_CPU_SGL_KERNEL": 1,
+	    "VLLM_CPU_KVCACHE_SPACE": 40
+        },
+        "server_parameters": {
+            "model": "RedHatAI/Meta-Llama-3.1-8B-Instruct-quantized.w8a8",
+            "tensor_parallel_size": 2,
+            "pipeline_parallel_size": 3,
+	    "dtype": "bfloat16",
+	    "distributed_executor_backend": "mp",
+	    "block_size": 128,
+	    "trust_remote_code": "",
+	    "enable_chunked_prefill": "",
+            "disable_log_stats": "",
+	    "enforce_eager": "",
+	    "max_num_batched_tokens": 2048,
+	    "max_num_seqs": 256,
+            "load_format": "dummy"
+        },
+        "client_parameters": {
+            "model": "RedHatAI/Meta-Llama-3.1-8B-Instruct-quantized.w8a8",
+            "backend": "vllm",
+            "dataset_name": "random",
+	    "random-input-len": 128,
+	    "random-output-len": 128,
+	    "ignore-eos": "",
+            "num_prompts": 1000
+        }
+    },
+    {
+        "test_name": "serving_llama8B_int4_pp1_sharegpt",
+        "qps_list": ["inf"],
+        "max_concurrency_list": [12, 16, 24, 32, 64, 128, 200],
+        "server_environment_variables": {
+            "VLLM_RPC_TIMEOUT": 100000,
+	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
+	    "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
+	    "VLLM_CPU_SGL_KERNEL": 1,
+	    "VLLM_CPU_KVCACHE_SPACE": 40
+        },
+        "server_parameters": {
+            "model": "hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4",
+	    "quantization": "awq",
+            "pipeline_parallel_size": 1,
+	    "dtype": "bfloat16",
+	    "distributed_executor_backend": "mp",
+	    "block_size": 128,
+	    "trust_remote_code": "",
+            "disable_log_stats": "",
+	    "enforce_eager": "",
+	    "max_num_batched_tokens": 2048,
+	    "max_num_seqs": 256,
+            "load_format": "dummy"
+        },
+        "client_parameters": {
+            "model": "hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4",
+            "backend": "vllm",
+            "dataset_name": "sharegpt",
+            "dataset_path": "./ShareGPT_V3_unfiltered_cleaned_split.json",
+            "num_prompts": 200
+        }
+    },
+    {
+        "test_name": "serving_llama8B_int4_tp2_sharegpt",
+        "qps_list": ["inf"],
+        "max_concurrency_list": [12, 16, 24, 32, 64, 128, 200],
+        "server_environment_variables": {
+            "VLLM_RPC_TIMEOUT": 100000,
+	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
+	    "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
+	    "VLLM_CPU_SGL_KERNEL": 1,
+	    "VLLM_CPU_KVCACHE_SPACE": 40
+        },
+        "server_parameters": {
+            "model": "hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4",
+	    "quantization": "awq",
+            "tensor_parallel_size": 2,
+	    "dtype": "bfloat16",
+	    "distributed_executor_backend": "mp",
+	    "block_size": 128,
+	    "trust_remote_code": "",
+            "disable_log_stats": "",
+	    "enforce_eager": "",
+	    "max_num_batched_tokens": 2048,
+	    "max_num_seqs": 256,
+            "load_format": "dummy"
+        },
+        "client_parameters": {
+            "model": "hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4",
+            "backend": "vllm",
+            "dataset_name": "sharegpt",
+            "dataset_path": "./ShareGPT_V3_unfiltered_cleaned_split.json",
+            "num_prompts": 200
+        }
+    },
+    {
+        "test_name": "serving_llama8B_int4_pp3_sharegpt",
+        "qps_list": ["inf"],
+        "max_concurrency_list": [12, 16, 24, 32, 64, 128, 200],
+        "server_environment_variables": {
+            "VLLM_RPC_TIMEOUT": 100000,
+	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
+	    "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
+	    "VLLM_CPU_SGL_KERNEL": 1,
+	    "VLLM_CPU_KVCACHE_SPACE": 40
+        },
+        "server_parameters": {
+            "model": "hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4",
+	    "quantization": "awq",
+            "pipeline_parallel_size": 3,
+	    "dtype": "bfloat16",
+	    "distributed_executor_backend": "mp",
+	    "block_size": 128,
+	    "trust_remote_code": "",
+            "disable_log_stats": "",
+	    "enforce_eager": "",
+	    "max_num_batched_tokens": 2048,
+	    "max_num_seqs": 256,
+            "load_format": "dummy"
+        },
+        "client_parameters": {
+            "model": "hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4",
+            "backend": "vllm",
+            "dataset_name": "sharegpt",
+            "dataset_path": "./ShareGPT_V3_unfiltered_cleaned_split.json",
+            "num_prompts": 200
+        }
+    },
+    {
+        "test_name": "serving_llama8B_int4_tp2pp3_sharegpt",
+        "qps_list": ["inf"],
+        "max_concurrency_list": [12, 16, 24, 32, 64, 128, 200],
+        "server_environment_variables": {
+            "VLLM_RPC_TIMEOUT": 100000,
+	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
+	    "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
+	    "VLLM_CPU_SGL_KERNEL": 1,
+	    "VLLM_CPU_KVCACHE_SPACE": 40
+        },
+        "server_parameters": {
+            "model": "hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4",
+	    "quantization": "awq",
+            "tensor_parallel_size": 2,
+            "pipeline_parallel_size": 3,
+	    "dtype": "bfloat16",
+	    "distributed_executor_backend": "mp",
+	    "block_size": 128,
+	    "trust_remote_code": "",
+            "disable_log_stats": "",
+	    "enforce_eager": "",
+	    "max_num_batched_tokens": 2048,
+	    "max_num_seqs": 256,
+            "load_format": "dummy"
+        },
+        "client_parameters": {
+            "model": "hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4",
+            "backend": "vllm",
+            "dataset_name": "sharegpt",
+            "dataset_path": "./ShareGPT_V3_unfiltered_cleaned_split.json",
+            "num_prompts": 200
+        }
+    },
+    {
+        "test_name": "serving_llama8B_int4_pp1_random_128_128",
+        "qps_list": ["inf"],
+        "max_concurrency_list": [12, 16, 24, 32, 64, 128, 200, 1000],
+        "server_environment_variables": {
+            "VLLM_RPC_TIMEOUT": 100000,
+	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
+	    "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
+	    "VLLM_CPU_SGL_KERNEL": 1,
+	    "VLLM_CPU_KVCACHE_SPACE": 40
+        },
+        "server_parameters": {
+            "model": "hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4",
+	    "quantization": "awq",
+            "pipeline_parallel_size": 1,
+	    "dtype": "bfloat16",
+	    "distributed_executor_backend": "mp",
+	    "block_size": 128,
+	    "trust_remote_code": "",
+	    "enable_chunked_prefill": "",
+            "disable_log_stats": "",
+	    "enforce_eager": "",
+	    "max_num_batched_tokens": 2048,
+	    "max_num_seqs": 256,
+            "load_format": "dummy"
+        },
+        "client_parameters": {
+            "model": "hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4",
+            "backend": "vllm",
+            "dataset_name": "random",
+	    "random-input-len": 128,
+	    "random-output-len": 128,
+	    "ignore-eos": "",
+            "num_prompts": 1000
+        }
+    },
+    {
+        "test_name": "serving_llama8B_int4_tp2_random_128_128",
+        "qps_list": ["inf"],
+        "max_concurrency_list": [12, 16, 24, 32, 64, 128, 200, 1000],
+        "server_environment_variables": {
+            "VLLM_RPC_TIMEOUT": 100000,
+	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
+	    "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
+	    "VLLM_CPU_SGL_KERNEL": 1,
+	    "VLLM_CPU_KVCACHE_SPACE": 40
+        },
+        "server_parameters": {
+            "model": "hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4",
+	    "quantization": "awq",
+            "tensor_parallel_size": 2,
+	    "dtype": "bfloat16",
+	    "distributed_executor_backend": "mp",
+	    "block_size": 128,
+	    "trust_remote_code": "",
+	    "enable_chunked_prefill": "",
+            "disable_log_stats": "",
+	    "enforce_eager": "",
+	    "max_num_batched_tokens": 2048,
+	    "max_num_seqs": 256,
+            "load_format": "dummy"
+        },
+        "client_parameters": {
+            "model": "hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4",
+            "backend": "vllm",
+            "dataset_name": "random",
+	    "random-input-len": 128,
+	    "random-output-len": 128,
+	    "ignore-eos": "",
+            "num_prompts": 1000
+        }
+    },
+    {
+        "test_name": "serving_llama8B_int4_pp3_random_128_128",
+        "qps_list": ["inf"],
+        "max_concurrency_list": [12, 16, 24, 32, 64, 128, 200, 1000],
+        "server_environment_variables": {
+            "VLLM_RPC_TIMEOUT": 100000,
+	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
+	    "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
+	    "VLLM_CPU_SGL_KERNEL": 1,
+	    "VLLM_CPU_KVCACHE_SPACE": 40
+        },
+        "server_parameters": {
+            "model": "hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4",
+	    "quantization": "awq",
+            "pipeline_parallel_size": 3,
+	    "dtype": "bfloat16",
+	    "distributed_executor_backend": "mp",
+	    "block_size": 128,
+	    "trust_remote_code": "",
+	    "enable_chunked_prefill": "",
+            "disable_log_stats": "",
+	    "enforce_eager": "",
+	    "max_num_batched_tokens": 2048,
+	    "max_num_seqs": 256,
+            "load_format": "dummy"
+        },
+        "client_parameters": {
+            "model": "hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4",
+            "backend": "vllm",
+            "dataset_name": "random",
+	    "random-input-len": 128,
+	    "random-output-len": 128,
+	    "ignore-eos": "",
+            "num_prompts": 1000
+        }
+    },
+    {
+        "test_name": "serving_llama8B_int4_tp2pp3_random_128_128",
+        "qps_list": ["inf"],
+        "max_concurrency_list": [12, 16, 24, 32, 64, 128, 200, 1000],
+        "server_environment_variables": {
+            "VLLM_RPC_TIMEOUT": 100000,
+	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
+	    "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
+	    "VLLM_CPU_SGL_KERNEL": 1,
+	    "VLLM_CPU_KVCACHE_SPACE": 40
+        },
+        "server_parameters": {
+            "model": "hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4",
+	    "quantization": "awq",
+            "tensor_parallel_size": 2,
+            "pipeline_parallel_size": 3,
+	    "dtype": "bfloat16",
+	    "distributed_executor_backend": "mp",
+	    "block_size": 128,
+	    "trust_remote_code": "",
+	    "enable_chunked_prefill": "",
+            "disable_log_stats": "",
+	    "enforce_eager": "",
+	    "max_num_batched_tokens": 2048,
+	    "max_num_seqs": 256,
+            "load_format": "dummy"
+        },
+        "client_parameters": {
+            "model": "hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4",
+            "backend": "vllm",
+            "dataset_name": "random",
+	    "random-input-len": 128,
+	    "random-output-len": 128,
+	    "ignore-eos": "",
+            "num_prompts": 1000
+        }
    }
 ]
--- a/.buildkite/release-pipeline.yaml
+++ b/.buildkite/release-pipeline.yaml
@ -1,21 +1,24 @@
 steps:
-  # aarch64 + CUDA builds
-  - label: "Build arm64 wheel - CUDA 12.8"
-    id: build-wheel-arm64-cuda-12-8
+  # aarch64 + CUDA builds. PyTorch 2.8 aarch64 + CUDA wheel is only available on CUDA 12.9
+  - label: "Build arm64 wheel - CUDA 12.9"
+    id: build-wheel-arm64-cuda-12-9
    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 ."
+      - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=12.9.1 --build-arg torch_cuda_arch_list='8.7 9.0 10.0+PTX 12.0' --tag vllm-ci:build-image --target build --progress plain -f docker/Dockerfile ."
      - "mkdir artifacts"
      - "docker run --rm -v $(pwd)/artifacts:/artifacts_host vllm-ci:build-image bash -c 'cp -r dist /artifacts_host && chmod -R a+rw /artifacts_host'"
      - "bash .buildkite/scripts/upload-wheels.sh"
    env:
      DOCKER_BUILDKIT: "1"

-  # x86 + CUDA builds
+  - block: "Build CUDA 12.8 wheel"
+    key: block-build-cu128-wheel
+
  - label: "Build wheel - CUDA 12.8"
+    depends_on: block-build-cu128-wheel
    id: build-wheel-cuda-12-8
    agents:
      queue: cpu_queue_postmerge
@ -27,7 +30,12 @@ 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
@ -39,44 +47,63 @@ steps:
    env:
      DOCKER_BUILDKIT: "1"

-  # Note(simon): We can always build CUDA 11.8 wheel to ensure the build is working.
-  # However, this block can be uncommented to save some compute hours.
-  # - block: "Build CUDA 11.8 wheel"
-  #   key: block-build-cu118-wheel
-
-  - label: "Build wheel - CUDA 11.8"
-    # depends_on: block-build-cu118-wheel
-    id: build-wheel-cuda-11-8
+  # x86 + CUDA builds
+  - label: "Build wheel - CUDA 12.9"
+    depends_on: ~
+    id: build-wheel-cuda-12-9
    agents:
      queue: cpu_queue_postmerge
    commands:
-      - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=11.8.0 --build-arg torch_cuda_arch_list='7.0 7.5 8.0 8.9 9.0+PTX' --tag vllm-ci:build-image --target build --progress plain -f docker/Dockerfile ."
+      - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=12.9.1 --build-arg torch_cuda_arch_list='7.0 7.5 8.0 8.9 9.0+PTX' --tag vllm-ci:build-image --target build --progress plain -f docker/Dockerfile ."
      - "mkdir artifacts"
      - "docker run --rm -v $(pwd)/artifacts:/artifacts_host vllm-ci:build-image bash -c 'cp -r dist /artifacts_host && chmod -R a+rw /artifacts_host'"
      - "bash .buildkite/scripts/upload-wheels.sh"
    env:
      DOCKER_BUILDKIT: "1"

-  - block: "Build release image"
+  - label: "Build release image (x86)"
    depends_on: ~
-    key: block-release-image-build
-
-  - label: "Build release image"
-    depends_on: block-release-image-build
-    id: build-release-image
+    id: build-release-image-x86
    agents:
      queue: cpu_queue_postmerge
    commands:
      - "aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin public.ecr.aws/q9t5s3a7"
-      - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=12.8.1 --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-$(uname -m) --target vllm-openai --progress plain -f docker/Dockerfile ."
+      - "docker push public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)"
+      # re-tag to default image tag and push, just in case arm64 build fails
+      - "docker tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m) public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT"
      - "docker push public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT"

+  # PyTorch 2.8 aarch64 + CUDA wheel is only available on CUDA 12.9
+  - label: "Build release image (arm64)"
+    depends_on: ~
+    id: build-release-image-arm64
+    agents:
+      queue: arm64_cpu_queue_postmerge
+    commands:
+      - "aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin public.ecr.aws/q9t5s3a7"
+      - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=12.9.1 --build-arg torch_cuda_arch_list='8.7 9.0 10.0+PTX 12.0' --build-arg INSTALL_KV_CONNECTORS=true --tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m) --target vllm-openai --progress plain -f docker/Dockerfile ."
+      - "docker push public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)"
+
+  # Add job to create multi-arch manifest
+  - label: "Create multi-arch manifest"
+    depends_on:
+      - build-release-image-x86
+      - build-release-image-arm64
+    id: create-multi-arch-manifest
+    agents:
+      queue: cpu_queue_postmerge
+    commands:
+      - "aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin public.ecr.aws/q9t5s3a7"
+      - "docker manifest create public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-x86_64 public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-aarch64 --amend"
+      - "docker manifest push public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT"
+
  - label: "Annotate release workflow"
    depends_on:
-      - build-release-image
+      - create-multi-arch-manifest
      - build-wheel-cuda-12-8
      - build-wheel-cuda-12-6
-      - build-wheel-cuda-11-8
+      - build-wheel-cuda-12-9
    id: annotate-release-workflow
    agents:
      queue: cpu_queue_postmerge
@ -123,18 +150,24 @@ steps:
    env:
      DOCKER_BUILDKIT: "1"

-  - block: "Build Neuron release image"
-    key: block-neuron-release-image-build
-    depends_on: ~
-
-  - label: "Build and publish Neuron release image"
-    depends_on: block-neuron-release-image-build
+  - label: "Build and publish nightly multi-arch image to DockerHub"
+    depends_on:
+      - create-multi-arch-manifest
+    if: build.env("NIGHTLY") == "1"
    agents:
-      queue: neuron-postmerge
+      queue: cpu_queue_postmerge
    commands:
      - "aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin public.ecr.aws/q9t5s3a7"
-      - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg GIT_REPO_CHECK=1 --tag public.ecr.aws/q9t5s3a7/vllm-neuron-release-repo:$(buildkite-agent meta-data get release-version) --tag public.ecr.aws/q9t5s3a7/vllm-neuron-release-repo:latest --progress plain -f docker/Dockerfile.neuron ."
-      - "docker push public.ecr.aws/q9t5s3a7/vllm-neuron-release-repo:latest"
-      - "docker push public.ecr.aws/q9t5s3a7/vllm-neuron-release-repo:$(buildkite-agent meta-data get release-version)"
+      - "docker pull public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT"
+      - "docker tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT vllm/vllm-openai:nightly"
+      - "docker tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT vllm/vllm-openai:nightly-$BUILDKITE_COMMIT"
+      - "docker push vllm/vllm-openai:nightly"
+      - "docker push vllm/vllm-openai:nightly-$BUILDKITE_COMMIT"
+      # Clean up old nightly builds (keep only last 14)
+      - "bash .buildkite/scripts/cleanup-nightly-builds.sh"
+    plugins:
+      - docker-login#v3.0.0:
+          username: vllmbot
+          password-env: DOCKERHUB_TOKEN
    env:
      DOCKER_BUILDKIT: "1"
--- a/.buildkite/scripts/cleanup-nightly-builds.sh
+++ b/.buildkite/scripts/cleanup-nightly-builds.sh
@ -0,0 +1,97 @@
+#!/bin/bash
+
+set -ex
+
+# Clean up old nightly builds from DockerHub, keeping only the last 14 builds
+# This script uses DockerHub API to list and delete old tags with "nightly-" prefix
+
+# DockerHub API endpoint for vllm/vllm-openai repository
+REPO_API_URL="https://hub.docker.com/v2/repositories/vllm/vllm-openai/tags"
+
+# Get DockerHub token from environment
+if [ -z "$DOCKERHUB_TOKEN" ]; then
+    echo "Error: DOCKERHUB_TOKEN environment variable is not set"
+    exit 1
+fi
+
+# Function to get all tags from DockerHub
+get_all_tags() {
+    local page=1
+    local all_tags=""
+    
+    while true; do
+        local response=$(curl -s -H "Authorization: Bearer $DOCKERHUB_TOKEN" \
+            "$REPO_API_URL?page=$page&page_size=100")
+        
+        # Get both last_updated timestamp and tag name, separated by |
+        local tags=$(echo "$response" | jq -r '.results[] | select(.name | startswith("nightly-")) | "\(.last_updated)|\(.name)"')
+        
+        if [ -z "$tags" ]; then
+            break
+        fi
+        
+        all_tags="$all_tags$tags"$'\n'
+        page=$((page + 1))
+    done
+    
+    # Sort by timestamp (newest first) and extract just the tag names
+    echo "$all_tags" | sort -r | cut -d'|' -f2
+}
+
+delete_tag() {
+    local tag_name="$1"
+    echo "Deleting tag: $tag_name"
+    
+    local delete_url="https://hub.docker.com/v2/repositories/vllm/vllm-openai/tags/$tag_name"
+    local response=$(curl -s -X DELETE -H "Authorization: Bearer $DOCKERHUB_TOKEN" "$delete_url")
+    
+    if echo "$response" | jq -e '.detail' > /dev/null 2>&1; then
+        echo "Warning: Failed to delete tag $tag_name: $(echo "$response" | jq -r '.detail')"
+    else
+        echo "Successfully deleted tag: $tag_name"
+    fi
+}
+
+# Get all nightly- prefixed tags, sorted by last_updated timestamp (newest first)
+echo "Fetching all tags from DockerHub..."
+all_tags=$(get_all_tags)
+
+if [ -z "$all_tags" ]; then
+    echo "No tags found to clean up"
+    exit 0
+fi
+
+# Count total tags
+total_tags=$(echo "$all_tags" | wc -l)
+echo "Found $total_tags tags"
+
+# Keep only the last 14 builds (including the current one)
+tags_to_keep=14
+tags_to_delete=$((total_tags - tags_to_keep))
+
+if [ $tags_to_delete -le 0 ]; then
+    echo "No tags need to be deleted (only $total_tags tags found, keeping $tags_to_keep)"
+    exit 0
+fi
+
+echo "Will delete $tags_to_delete old tags, keeping the newest $tags_to_keep"
+
+# Get tags to delete (skip the first $tags_to_keep tags)
+tags_to_delete_list=$(echo "$all_tags" | tail -n +$((tags_to_keep + 1)))
+
+if [ -z "$tags_to_delete_list" ]; then
+    echo "No tags to delete"
+    exit 0
+fi
+
+# Delete old tags
+echo "Deleting old tags..."
+while IFS= read -r tag; do
+    if [ -n "$tag" ]; then
+        delete_tag "$tag"
+        # Add a small delay to avoid rate limiting
+        sleep 1
+    fi
+done <<< "$tags_to_delete_list"
+
+echo "Cleanup completed successfully"
--- a/.buildkite/scripts/hardware_ci/run-amd-test.sh
+++ b/.buildkite/scripts/hardware_ci/run-amd-test.sh
@ -164,7 +164,6 @@ if [[ $commands == *" entrypoints/llm "* ]]; then
  --ignore=entrypoints/llm/test_chat.py \
  --ignore=entrypoints/llm/test_accuracy.py \
  --ignore=entrypoints/llm/test_init.py \
-  --ignore=entrypoints/llm/test_generate_multiple_loras.py \
  --ignore=entrypoints/llm/test_prompt_validation.py "}
 fi

--- a/.buildkite/scripts/hardware_ci/run-cpu-test.sh
+++ b/.buildkite/scripts/hardware_ci/run-cpu-test.sh
@ -25,8 +25,8 @@ numactl -C "$CORE_RANGE" -N "$NUMA_NODE" docker build --tag cpu-test-"$NUMA_NODE
 numactl -C "$CORE_RANGE" -N "$NUMA_NODE" docker build --build-arg VLLM_CPU_DISABLE_AVX512="true" --tag cpu-test-"$NUMA_NODE"-avx2 --target vllm-test -f docker/Dockerfile.cpu .

 # Run the image, setting --shm-size=4g for tensor parallel.
-docker run -itd --cpuset-cpus="$CORE_RANGE" --cpuset-mems="$NUMA_NODE" --entrypoint /bin/bash -v ~/.cache/huggingface:/root/.cache/huggingface --privileged=true -e HF_TOKEN --env VLLM_CPU_KVCACHE_SPACE=4 --env VLLM_CPU_CI_ENV=1 -e E2E_OMP_THREADS="$OMP_CORE_RANGE" --shm-size=4g --name cpu-test-"$NUMA_NODE" cpu-test-"$NUMA_NODE"
-docker run -itd --cpuset-cpus="$CORE_RANGE" --cpuset-mems="$NUMA_NODE" --entrypoint /bin/bash -v ~/.cache/huggingface:/root/.cache/huggingface --privileged=true -e HF_TOKEN --env VLLM_CPU_KVCACHE_SPACE=4 --env VLLM_CPU_CI_ENV=1 -e E2E_OMP_THREADS="$OMP_CORE_RANGE" --shm-size=4g --name cpu-test-"$NUMA_NODE"-avx2 cpu-test-"$NUMA_NODE"-avx2
+docker run -itd --cpuset-cpus="$CORE_RANGE" --cpuset-mems="$NUMA_NODE" --entrypoint /bin/bash -v ~/.cache/huggingface:/root/.cache/huggingface --privileged=true -e HF_TOKEN --env VLLM_CPU_KVCACHE_SPACE=16 --env VLLM_CPU_CI_ENV=1 -e E2E_OMP_THREADS="$OMP_CORE_RANGE" --shm-size=4g --name cpu-test-"$NUMA_NODE" cpu-test-"$NUMA_NODE"
+docker run -itd --cpuset-cpus="$CORE_RANGE" --cpuset-mems="$NUMA_NODE" --entrypoint /bin/bash -v ~/.cache/huggingface:/root/.cache/huggingface --privileged=true -e HF_TOKEN --env VLLM_CPU_KVCACHE_SPACE=16 --env VLLM_CPU_CI_ENV=1 -e E2E_OMP_THREADS="$OMP_CORE_RANGE" --shm-size=4g --name cpu-test-"$NUMA_NODE"-avx2 cpu-test-"$NUMA_NODE"-avx2

 function cpu_tests() {
  set -e
@ -46,21 +46,26 @@ function cpu_tests() {
    set -e
    python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m"

+  # Run kernel tests
+  docker exec cpu-test-"$NUMA_NODE" bash -c "
+    set -e
+    pytest -x -v -s tests/kernels/test_onednn.py"
+
  # Run basic model test
  docker exec cpu-test-"$NUMA_NODE" bash -c "
    set -e
    # Note: disable until supports V1
-    # pytest -v -s tests/kernels/attention/test_cache.py -m cpu_model
-    # pytest -v -s tests/kernels/attention/test_mla_decode_cpu.py -m cpu_model
+    # pytest -x -v -s tests/kernels/attention/test_cache.py -m cpu_model
+    # pytest -x -v -s tests/kernels/attention/test_mla_decode_cpu.py -m cpu_model

    # Note: disable Bart until supports V1
-    pytest -v -s tests/models/language/generation -m cpu_model \
+    pytest -x -v -s tests/models/language/generation -m cpu_model \
                --ignore=tests/models/language/generation/test_bart.py
-    VLLM_CPU_SGL_KERNEL=1 pytest -v -s tests/models/language/generation -m cpu_model \
+    VLLM_CPU_SGL_KERNEL=1 pytest -x -v -s tests/models/language/generation -m cpu_model \
                --ignore=tests/models/language/generation/test_bart.py

-    pytest -v -s tests/models/language/pooling -m cpu_model
-    pytest -v -s tests/models/multimodal/generation \
+    pytest -x -v -s tests/models/language/pooling -m cpu_model
+    pytest -x -v -s tests/models/multimodal/generation \
                --ignore=tests/models/multimodal/generation/test_mllama.py \
                --ignore=tests/models/multimodal/generation/test_pixtral.py \
                -m cpu_model"
@ -68,35 +73,51 @@ function cpu_tests() {
  # Run compressed-tensor test
  docker exec cpu-test-"$NUMA_NODE" bash -c "
    set -e
-    pytest -s -v \
+    pytest -x -s -v \
    tests/quantization/test_compressed_tensors.py::test_compressed_tensors_w8a8_logprobs[False-10-32-neuralmagic/Llama-3.2-1B-quantized.w8a8]"

  # Note: disable it until supports V1
  # Run AWQ test
  # docker exec cpu-test-"$NUMA_NODE" bash -c "
  #   set -e
-  #   VLLM_USE_V1=0 pytest -s -v \
+  #   VLLM_USE_V1=0 pytest -x -s -v \
  #   tests/quantization/test_ipex_quant.py"

  # Run multi-lora tests
  docker exec cpu-test-"$NUMA_NODE" bash -c "
    set -e
-    pytest -s -v \
+    pytest -x -s -v \
    tests/lora/test_qwen2vl.py"

-  # online serving
+  # online serving: tp+pp
  docker exec cpu-test-"$NUMA_NODE" bash -c '
    set -e
    VLLM_CPU_OMP_THREADS_BIND=$E2E_OMP_THREADS VLLM_CPU_SGL_KERNEL=1 vllm serve meta-llama/Llama-3.2-3B-Instruct -tp=2 -pp=2 &
+    server_pid=$!
    timeout 600 bash -c "until curl localhost:8000/v1/models; do sleep 1; done" || exit 1
    vllm bench serve \
      --backend vllm \
      --dataset-name random \
      --model meta-llama/Llama-3.2-3B-Instruct \
      --num-prompts 20 \
-      --endpoint /v1/completions'
+      --endpoint /v1/completions
+    kill -s SIGTERM $server_pid &'
+
+  # online serving: tp+dp
+  docker exec cpu-test-"$NUMA_NODE" bash -c '
+    set -e
+    VLLM_CPU_OMP_THREADS_BIND=$E2E_OMP_THREADS VLLM_CPU_SGL_KERNEL=1 vllm serve meta-llama/Llama-3.2-3B-Instruct -tp=2 -dp=2 &
+    server_pid=$!
+    timeout 600 bash -c "until curl localhost:8000/v1/models; do sleep 1; done" || exit 1
+    vllm bench serve \
+      --backend vllm \
+      --dataset-name random \
+      --model meta-llama/Llama-3.2-3B-Instruct \
+      --num-prompts 20 \
+      --endpoint /v1/completions
+    kill -s SIGTERM $server_pid &'
 }

 # 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 $NUMA_NODE"
--- a/.buildkite/scripts/hardware_ci/run-neuron-test.sh
+++ b/.buildkite/scripts/hardware_ci/run-neuron-test.sh
@ -1,64 +0,0 @@
-#!/bin/bash
-
-# This script build the Neuron docker image and run the API server inside the container.
-# It serves a sanity check for compilation and basic model usage.
-set -e
-set -v
-
-image_name="neuron/vllm-ci"
-container_name="neuron_$(tr -dc A-Za-z0-9 < /dev/urandom | head -c 10; echo)"
-
-HF_CACHE="$(realpath ~)/huggingface"
-mkdir -p "${HF_CACHE}"
-HF_MOUNT="/root/.cache/huggingface"
-HF_TOKEN=$(aws secretsmanager get-secret-value  --secret-id "ci/vllm-neuron/hf-token" --region us-west-2 --query 'SecretString' --output text | jq -r .VLLM_NEURON_CI_HF_TOKEN)
-
-NEURON_COMPILE_CACHE_URL="$(realpath ~)/neuron_compile_cache"
-mkdir -p "${NEURON_COMPILE_CACHE_URL}"
-NEURON_COMPILE_CACHE_MOUNT="/root/.cache/neuron_compile_cache"
-
-# Try building the docker image
-aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin public.ecr.aws
-
-# prune old image and containers to save disk space, and only once a day
-# by using a timestamp file in tmp.
-if [ -f /tmp/neuron-docker-build-timestamp ]; then
-    last_build=$(cat /tmp/neuron-docker-build-timestamp)
-    current_time=$(date +%s)
-    if [ $((current_time - last_build)) -gt 86400 ]; then
-        # Remove dangling images (those that are not tagged and not used by any container)
-        docker image prune -f
-        # Remove unused volumes / force the system prune for old images as well.
-        docker volume prune -f && docker system prune -f
-        echo "$current_time" > /tmp/neuron-docker-build-timestamp
-    fi
-else
-    date "+%s" > /tmp/neuron-docker-build-timestamp
-fi
-
-docker build -t "${image_name}" -f docker/Dockerfile.neuron .
-
-# Setup cleanup
-remove_docker_container() {
-    docker image rm -f "${image_name}" || true;
-}
-trap remove_docker_container EXIT
-
-# Run the image
-docker run --rm -it --device=/dev/neuron0 --network bridge \
-       -v "${HF_CACHE}:${HF_MOUNT}" \
-       -e "HF_HOME=${HF_MOUNT}" \
-       -e "HF_TOKEN=${HF_TOKEN}" \
-       -v "${NEURON_COMPILE_CACHE_URL}:${NEURON_COMPILE_CACHE_MOUNT}" \
-       -e "NEURON_COMPILE_CACHE_URL=${NEURON_COMPILE_CACHE_MOUNT}" \
-       --name "${container_name}" \
-       ${image_name} \
-       /bin/bash -c "
-            set -e; # Exit on first error
-            python3 /workspace/vllm/examples/offline_inference/neuron.py;
-            python3 -m pytest /workspace/vllm/tests/neuron/1_core/ -v --capture=tee-sys;
-            for f in /workspace/vllm/tests/neuron/2_core/*.py; do
-                echo \"Running test file: \$f\";
-                python3 -m pytest \$f -v --capture=tee-sys;
-            done
-       "
--- 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: $RESULTS_DIR"
 echo "--- Installing Python dependencies ---"
 python3 -m pip install --progress-bar off git+https://github.com/thuml/depyf.git \
    && python3 -m pip install --progress-bar off pytest pytest-asyncio tpu-info \
-    && python3 -m pip install --progress-bar off lm_eval[api]==0.4.4 \
+    && python3 -m pip install --progress-bar off "lm-eval @ git+https://github.com/EleutherAI/lm-evaluation-harness.git@206b7722158f58c35b7ffcd53b035fdbdda5126d" \
    && python3 -m pip install --progress-bar off hf-transfer
 echo "--- Python dependencies installed ---"
 export VLLM_USE_V1=1
--- a/.buildkite/scripts/hardware_ci/run-xpu-test.sh
+++ b/.buildkite/scripts/hardware_ci/run-xpu-test.sh
@ -23,20 +23,27 @@ docker run \
    --device /dev/dri \
    -v /dev/dri/by-path:/dev/dri/by-path \
    --entrypoint="" \
+    -e "HF_TOKEN=${HF_TOKEN}" \
+    -e "ZE_AFFINITY_MASK=${ZE_AFFINITY_MASK}" \
    --name "${container_name}" \
    "${image_name}" \
-    sh -c '
-    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
+    bash -c '
+    set -e
+    echo $ZE_AFFINITY_MASK
+    pip install tblib==3.1.0
+    python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m --block-size 64 --enforce-eager
+    python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m --block-size 64 -O3 -O.cudagraph_mode=NONE
+    python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m --block-size 64 --enforce-eager -tp 2 --distributed-executor-backend ray
+    python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m --block-size 64 --enforce-eager -tp 2 --distributed-executor-backend mp
+    VLLM_ATTENTION_BACKEND=TRITON_ATTN_VLLM_V1 python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m --block-size 64 --enforce-eager
    cd tests
    pytest -v -s v1/core
    pytest -v -s v1/engine
    pytest -v -s v1/sample --ignore=v1/sample/test_logprobs.py --ignore=v1/sample/test_logprobs_e2e.py
    pytest -v -s v1/worker --ignore=v1/worker/test_gpu_model_runner.py
    pytest -v -s v1/structured_output
-    pytest -v -s v1/spec_decode --ignore=v1/spec_decode/test_max_len.py --ignore=v1/spec_decode/test_eagle.py
-    pytest -v -s v1/kv_connector/unit --ignore=v1/kv_connector/unit/test_multi_connector.py --ignore=v1/kv_connector/unit/test_nixl_connector.py
+    pytest -v -s v1/spec_decode --ignore=v1/spec_decode/test_max_len.py --ignore=v1/spec_decode/test_eagle.py --ignore=v1/spec_decode/test_tree_attention.py
+    pytest -v -s v1/kv_connector/unit --ignore=v1/kv_connector/unit/test_multi_connector.py --ignore=v1/kv_connector/unit/test_nixl_connector.py --ignore=v1/kv_connector/unit/test_shared_storage_connector.py
    pytest -v -s v1/test_serial_utils.py
    pytest -v -s v1/test_utils.py
    pytest -v -s v1/test_metrics_reader.py
--- a/.buildkite/scripts/tpu/cleanup_docker.sh
+++ b/.buildkite/scripts/tpu/cleanup_docker.sh
@ -17,7 +17,7 @@ if [ "$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"

@ -47,14 +58,15 @@ python3 .buildkite/generate_index.py --wheel "$normal_wheel"
 aws s3 cp "$wheel" "s3://vllm-wheels/$BUILDKITE_COMMIT/"
 aws s3 cp "$normal_wheel" "s3://vllm-wheels/$BUILDKITE_COMMIT/"

-if [[ $normal_wheel == *"cu118"* ]]; then
-    # if $normal_wheel matches cu118, do not upload the index.html
-    echo "Skipping index files for cu118 wheels"
-elif [[ $normal_wheel == *"cu126"* ]]; then
+if [[ $normal_wheel == *"cu126"* ]]; then
    # if $normal_wheel matches cu126, do not upload the index.html
    echo "Skipping index files for cu126 wheels"
+elif [[ $normal_wheel == *"cu128"* ]]; then
+    # if $normal_wheel matches cu128, do not upload the index.html
+    echo "Skipping index files for cu128 wheels"
 else
-    # only upload index.html for cu128 wheels (default wheels)
+    # only upload index.html for cu129 wheels (default wheels) as it
+    # is available on both x86 and arm64
    aws s3 cp index.html "s3://vllm-wheels/$BUILDKITE_COMMIT/vllm/index.html"
    aws s3 cp "s3://vllm-wheels/nightly/index.html" "s3://vllm-wheels/$BUILDKITE_COMMIT/index.html"
 fi
@ -63,14 +75,15 @@ fi
 aws s3 cp "$wheel" "s3://vllm-wheels/nightly/"
 aws s3 cp "$normal_wheel" "s3://vllm-wheels/nightly/"

-if [[ $normal_wheel == *"cu118"* ]]; then
-    # if $normal_wheel matches cu118, do not upload the index.html
-    echo "Skipping index files for cu118 wheels"
-elif [[ $normal_wheel == *"cu126"* ]]; then
+if [[ $normal_wheel == *"cu126"* ]]; then
    # if $normal_wheel matches cu126, do not upload the index.html
    echo "Skipping index files for cu126 wheels"
+elif [[ $normal_wheel == *"cu128"* ]]; then
+    # if $normal_wheel matches cu128, do not upload the index.html
+    echo "Skipping index files for cu128 wheels"
 else
-    # only upload index.html for cu128 wheels (default wheels)
+    # only upload index.html for cu129 wheels (default wheels) as it
+    # is available on both x86 and arm64
    aws s3 cp index.html "s3://vllm-wheels/nightly/vllm/index.html"
 fi

--- a/.buildkite/test-pipeline.yaml
+++ b/.buildkite/test-pipeline.yaml
@ -41,7 +41,8 @@ steps:
  commands:
  - bash standalone_tests/pytorch_nightly_dependency.sh

- label: Async Engine, Inputs, Utils, Worker Test # 24min
+- label: Async Engine, Inputs, Utils, Worker Test # 36min
+  timeout_in_minutes: 50
  mirror_hardwares: [amdexperimental]
  source_file_dependencies:
  - vllm/
@ -53,6 +54,7 @@ steps:
  - tests/utils_
  - tests/worker
  - tests/standalone_tests/lazy_imports.py
+  - tests/transformers_utils
  commands:
  - python3 standalone_tests/lazy_imports.py
  - pytest -v -s mq_llm_engine # MQLLMEngine
@ -62,8 +64,10 @@ steps:
  - pytest -v -s multimodal
  - pytest -v -s utils_ # Utils
  - pytest -v -s worker # Worker
+  - pytest -v -s transformers_utils # transformers_utils

- label: Python-only Installation Test
+- label: Python-only Installation Test # 10min
+  timeout_in_minutes: 20
  mirror_hardwares: [amdexperimental]
  source_file_dependencies:
  - tests/standalone_tests/python_only_compile.sh
@ -71,7 +75,8 @@ steps:
  commands:
  - bash standalone_tests/python_only_compile.sh

- label: Basic Correctness Test # 30min
+- label: Basic Correctness Test # 20min
+  timeout_in_minutes: 30
  mirror_hardwares: [amdexperimental]
  fast_check: true
  torch_nightly: true
@ -88,16 +93,8 @@ 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
+- label: Core Test # 22min
+  timeout_in_minutes: 35
  mirror_hardwares: [amdexperimental]
  fast_check: true
  source_file_dependencies:
@ -107,7 +104,19 @@ steps:
  commands:
  - pytest -v -s core

- label: Entrypoints Test (LLM) # 40min
+- label: Entrypoints Unit Tests # 5min
+  timeout_in_minutes: 10
+  working_dir: "/vllm-workspace/tests"
+  fast_check: true
+  source_file_dependencies:
+  - vllm/entrypoints
+  - tests/entrypoints/
+  commands:
+  - pytest -v -s entrypoints/openai/tool_parsers
+  - pytest -v -s entrypoints/ --ignore=entrypoints/llm --ignore=entrypoints/openai --ignore=entrypoints/offline_mode --ignore=entrypoints/test_chat_utils.py  --ignore=entrypoints/pooling
+
+- label: Entrypoints Integration Test (LLM) # 30min
+  timeout_in_minutes: 40
  mirror_hardwares: [amdexperimental]
  working_dir: "/vllm-workspace/tests"
  fast_check: true
@ -118,13 +127,13 @@ steps:
  - tests/entrypoints/offline_mode
  commands:
  - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-  - pytest -v -s entrypoints/llm --ignore=entrypoints/llm/test_lazy_outlines.py --ignore=entrypoints/llm/test_generate.py --ignore=entrypoints/llm/test_generate_multiple_loras.py --ignore=entrypoints/llm/test_collective_rpc.py
+  - pytest -v -s entrypoints/llm --ignore=entrypoints/llm/test_lazy_outlines.py --ignore=entrypoints/llm/test_generate.py --ignore=entrypoints/llm/test_collective_rpc.py
  - pytest -v -s entrypoints/llm/test_lazy_outlines.py # it needs a clean process
  - pytest -v -s entrypoints/llm/test_generate.py # it needs a clean process
-  - pytest -v -s entrypoints/llm/test_generate_multiple_loras.py # it needs a clean process
  - VLLM_USE_V1=0 pytest -v -s entrypoints/offline_mode # Needs to avoid interference with other tests

- label: Entrypoints Test (API Server) # 40min
+- label: Entrypoints Integration Test (API Server) # 100min
+  timeout_in_minutes: 130
  mirror_hardwares: [amdexperimental]
  working_dir: "/vllm-workspace/tests"
  fast_check: true
@ -135,10 +144,25 @@ 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 --ignore=entrypoints/openai/tool_parsers/
  - pytest -v -s entrypoints/test_chat_utils.py

- label: Distributed Tests (4 GPUs) # 10min
+- label: Entrypoints Integration Test (Pooling)
+  timeout_in_minutes: 50
+  mirror_hardwares: [amdexperimental]
+  working_dir: "/vllm-workspace/tests"
+  fast_check: true
+  torch_nightly: true
+  source_file_dependencies:
+  - vllm/
+  - tests/entrypoints/pooling
+  commands:
+  - export VLLM_WORKER_MULTIPROC_METHOD=spawn
+  - pytest -v -s entrypoints/pooling
+
+- label: Distributed Tests (4 GPUs) # 35min
+  timeout_in_minutes: 50
  mirror_hardwares: [amdexperimental]
  working_dir: "/vllm-workspace/tests"
  num_gpus: 4
@ -181,7 +205,8 @@ steps:
  - VLLM_ALLOW_INSECURE_SERIALIZATION=1 RAY_DEDUP_LOGS=0 python3 rlhf_colocate.py
  - popd

- label: EPLB Algorithm Test
+- label: EPLB Algorithm Test # 5min
+  timeout_in_minutes: 15
  working_dir: "/vllm-workspace/tests"
  source_file_dependencies:
  - vllm/distributed/eplb
@ -190,6 +215,7 @@ steps:
  - pytest -v -s distributed/test_eplb_algo.py

 - label: EPLB Execution Test # 5min
+  timeout_in_minutes: 15
  working_dir: "/vllm-workspace/tests"
  num_gpus: 4
  source_file_dependencies:
@ -198,13 +224,14 @@ steps:
  commands:
  - pytest -v -s distributed/test_eplb_execute.py

- label: Metrics, Tracing Test # 10min
+- label: Metrics, Tracing Test # 12min
+  timeout_in_minutes: 20
  mirror_hardwares: [amdexperimental]
  num_gpus: 2
  source_file_dependencies:
  - vllm/
  - tests/metrics
-  - tests/tracing
+  - tests/v1/tracing
  commands:
  - pytest -v -s metrics
  - "pip install \
@ -217,7 +244,8 @@ steps:
 ##### fast check tests  #####
 #####  1 GPU test  #####

- label: Regression Test # 5min
+- label: Regression Test # 7min
+  timeout_in_minutes: 20
  mirror_hardwares: [amdexperimental]
  source_file_dependencies:
  - vllm/
@ -227,7 +255,8 @@ steps:
  - pytest -v -s test_regression.py
  working_dir: "/vllm-workspace/tests" # optional

- label: Engine Test # 10min
+- label: Engine Test # 25min
+  timeout_in_minutes: 40
  mirror_hardwares: [amdexperimental]
  source_file_dependencies:
  - vllm/
@ -242,7 +271,29 @@ steps:
  # OOM in the CI unless we run this separately
  - pytest -v -s tokenization

- label: V1 Test
+- label: V1 Test e2e + engine # 30min
+  timeout_in_minutes: 45
+  mirror_hardwares: [amdexperimental]
+  source_file_dependencies:
+    - vllm/
+    - tests/v1
+  commands:
+    # TODO: accuracy does not match, whether setting
+    # VLLM_USE_FLASHINFER_SAMPLER or not on H100.
+    - pytest -v -s v1/e2e
+    - pytest -v -s v1/engine
+
+- label: V1 Test entrypoints # 35min
+  timeout_in_minutes: 50
+  mirror_hardwares: [amdexperimental]
+  source_file_dependencies:
+    - vllm/
+    - tests/v1
+  commands:
+    - pytest -v -s v1/entrypoints
+
+- label: V1 Test others # 42min
+  timeout_in_minutes: 60
  mirror_hardwares: [amdexperimental]
  source_file_dependencies:
    - vllm/
@ -250,8 +301,7 @@ steps:
  commands:
    # split the test to avoid interference
    - 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
@ -263,14 +313,12 @@ steps:
    - pytest -v -s v1/test_utils.py
    - pytest -v -s v1/test_oracle.py
    - pytest -v -s v1/test_metrics_reader.py
-    # TODO: accuracy does not match, whether setting
-    # VLLM_USE_FLASHINFER_SAMPLER or not on H100.
-    - pytest -v -s v1/e2e
    # Integration test for streaming correctness (requires special branch).
    - pip install -U git+https://github.com/robertgshaw2-redhat/lm-evaluation-harness.git@streaming-api
    - pytest -v -s entrypoints/openai/correctness/test_lmeval.py::test_lm_eval_accuracy_v1_engine

- label: Examples Test # 25min
+- label: Examples Test # 30min
+  timeout_in_minutes: 45
  mirror_hardwares: [amdexperimental]
  working_dir: "/vllm-workspace/examples"
  source_file_dependencies:
@ -288,23 +336,14 @@ steps:
    - python3 offline_inference/vision_language_pooling.py --seed 0
    - python3 offline_inference/vision_language_multi_image.py --seed 0
    - VLLM_USE_V1=0 python3 others/tensorize_vllm_model.py --model facebook/opt-125m serialize --serialized-directory /tmp/ --suffix v1 && python3 others/tensorize_vllm_model.py --model facebook/opt-125m deserialize --path-to-tensors /tmp/vllm/facebook/opt-125m/v1/model.tensors
-    - python3 offline_inference/encoder_decoder.py
    - python3 offline_inference/encoder_decoder_multimodal.py --model-type whisper --seed 0
    - python3 offline_inference/basic/classify.py
    - python3 offline_inference/basic/embed.py
    - 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)
+- label: Platform Tests (CUDA) # 4min
+  timeout_in_minutes: 15
  mirror_hardwares: [amdexperimental]
  source_file_dependencies:
  - vllm/
@ -312,7 +351,8 @@ steps:
  commands:
    - pytest -v -s cuda/test_cuda_context.py

- label: Samplers Test # 36min
+- label: Samplers Test # 56min
+  timeout_in_minutes: 75
  mirror_hardwares: [amdexperimental]
  source_file_dependencies:
  - vllm/model_executor/layers
@ -323,15 +363,23 @@ steps:
    - pytest -v -s samplers
    - VLLM_USE_FLASHINFER_SAMPLER=1 pytest -v -s samplers

- label: LoRA Test %N # 15min each
+- label: LoRA Test %N # 20min each
+  timeout_in_minutes: 30
  mirror_hardwares: [amdexperimental]
  source_file_dependencies:
  - vllm/lora
  - tests/lora
-  command: pytest -v -s lora --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --ignore=lora/test_chatglm3_tp.py --ignore=lora/test_llama_tp.py
+  commands:
+    - pytest -v -s lora \
+      --shard-id=$$BUILDKITE_PARALLEL_JOB \
+      --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT \
+      --ignore=lora/test_chatglm3_tp.py \
+      --ignore=lora/test_llama_tp.py \
+      --ignore=lora/test_llm_with_multi_loras.py
  parallelism: 4

- label: PyTorch Compilation Unit Tests
+- label: PyTorch Compilation Unit Tests # 15min
+  timeout_in_minutes: 30
  mirror_hardwares: [amdexperimental]
  torch_nightly: true
  source_file_dependencies:
@ -345,8 +393,10 @@ 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
+- label: PyTorch Fullgraph Smoke Test # 15min
+  timeout_in_minutes: 30
  mirror_hardwares: [amdexperimental]
  torch_nightly: true
  source_file_dependencies:
@ -354,12 +404,10 @@ steps:
  - tests/compile
  commands:
  - pytest -v -s compile/test_basic_correctness.py
-  # these tests need to be separated, cannot combine
-  - pytest -v -s compile/piecewise/test_simple.py
-  - pytest -v -s compile/piecewise/test_toy_llama.py
-  - pytest -v -s compile/piecewise/test_full_cudagraph.py
+  - pytest -v -s compile/piecewise/

- label: PyTorch Fullgraph Test # 18min
+- label: PyTorch Fullgraph Test # 20min
+  timeout_in_minutes: 30
  mirror_hardwares: [amdexperimental]
  torch_nightly: true
  source_file_dependencies:
@ -368,7 +416,8 @@ steps:
  commands:
  - pytest -v -s compile/test_full_graph.py

- label: Kernels Core Operation Test
+- label: Kernels Core Operation Test # 48min
+  timeout_in_minutes: 75
  mirror_hardwares: [amdexperimental]
  source_file_dependencies:
  - csrc/
@ -376,7 +425,8 @@ steps:
  commands:
    - pytest -v -s kernels/core

- label: Kernels Attention Test %N
+- label: Kernels Attention Test %N # 23min
+  timeout_in_minutes: 35
  mirror_hardwares: [amdexperimental]
  source_file_dependencies:
  - csrc/attention/
@ -387,7 +437,8 @@ steps:
    - pytest -v -s kernels/attention --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT
  parallelism: 2

- label: Kernels Quantization Test %N
+- label: Kernels Quantization Test %N # 64min
+  timeout_in_minutes: 90
  mirror_hardwares: [amdexperimental]
  source_file_dependencies:
  - csrc/quantization/
@ -397,18 +448,21 @@ steps:
    - pytest -v -s kernels/quantization --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT
  parallelism: 2

- label: Kernels MoE Test %N
+- label: Kernels MoE Test %N # 40min
+  timeout_in_minutes: 60
  mirror_hardwares: [amdexperimental]
  source_file_dependencies:
  - csrc/quantization/cutlass_w8a8/moe/
  - csrc/moe/
  - tests/kernels/moe
  - vllm/model_executor/layers/fused_moe/
+  - vllm/distributed/device_communicators/
  commands:
    - pytest -v -s kernels/moe --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT
  parallelism: 2

- label: Kernels Mamba Test
+- label: Kernels Mamba Test # 31min
+  timeout_in_minutes: 45
  mirror_hardwares: [amdexperimental]
  source_file_dependencies:
  - csrc/mamba/
@ -416,7 +470,8 @@ steps:
  commands:
    - pytest -v -s kernels/mamba

- label: Tensorizer Test # 11min
+- label: Tensorizer Test # 14min
+  timeout_in_minutes: 25
  mirror_hardwares: [amdexperimental]
  source_file_dependencies:
  - vllm/model_executor/model_loader
@ -428,7 +483,8 @@ steps:
    - pytest -v -s tensorizer_loader
    - pytest -v -s entrypoints/openai/test_tensorizer_entrypoint.py

- label: Model Executor Test
+- label: Model Executor Test # 7min
+  timeout_in_minutes: 20
  mirror_hardwares: [amdexperimental]
  source_file_dependencies:
  - vllm/model_executor
@ -438,7 +494,8 @@ steps:
    - export VLLM_WORKER_MULTIPROC_METHOD=spawn
    - pytest -v -s model_executor

- label: Benchmarks # 9min
+- label: Benchmarks # 11min
+  timeout_in_minutes: 20
  mirror_hardwares: [amdexperimental]
  working_dir: "/vllm-workspace/.buildkite"
  source_file_dependencies:
@ -446,7 +503,8 @@ steps:
  commands:
  - bash scripts/run-benchmarks.sh

- label: Benchmarks CLI Test # 10min
+- label: Benchmarks CLI Test # 7min
+  timeout_in_minutes: 20
  mirror_hardwares: [amdexperimental]
  source_file_dependencies:
  - vllm/
@ -454,7 +512,8 @@ steps:
  commands:
  - pytest -v -s benchmarks/

- label: Quantization Test
+- label: Quantization Test # 70min
+  timeout_in_minutes: 90
  mirror_hardwares: [amdexperimental]
  source_file_dependencies:
  - csrc/
@ -462,21 +521,25 @@ steps:
  - tests/quantization
  commands:
  # temporary install here since we need nightly, will move to requirements/test.in
-  # after torchao 0.12 release
-  - pip install --pre torchao --index-url https://download.pytorch.org/whl/nightly/cu126
+  # after torchao 0.12 release, and pin a working version of torchao nightly here
+
+  # since torchao nightly is only compatible with torch nightly currently
+  # https://github.com/pytorch/ao/issues/2919, we'll have to skip new torchao tests for now
+  # we can only upgrade after this is resolved
+  - pip install --pre torchao==0.13.0.dev20250814 --index-url https://download.pytorch.org/whl/nightly/cu128
  - VLLM_TEST_FORCE_LOAD_FORMAT=auto pytest -v -s quantization

 - label: LM Eval Small Models # 53min
+  timeout_in_minutes: 75
  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
+- label: OpenAI API correctness # 22min
+  timeout_in_minutes: 30
  mirror_hardwares: [amdexperimental]
  source_file_dependencies:
  - csrc/
@ -485,7 +548,8 @@ steps:
  commands: # LMEval+Transcription WER check
  - pytest -s entrypoints/openai/correctness/

- label: Encoder Decoder tests # 5min
+- label: Encoder Decoder tests # 12min
+  timeout_in_minutes: 20
  mirror_hardwares: [amdexperimental]
  source_file_dependencies:
  - vllm/
@ -493,7 +557,8 @@ steps:
  commands:
    - pytest -v -s encoder_decoder

- label: OpenAI-Compatible Tool Use # 20 min
+- label: OpenAI-Compatible Tool Use # 23 min
+  timeout_in_minutes: 35
  mirror_hardwares: [amdexperimental]
  fast_check: false
  source_file_dependencies:
@ -506,30 +571,82 @@ steps:

 #####  models test  #####

- label: Basic Models Test # 24min
+- label: Basic Models Tests (Initialization)
+  timeout_in_minutes: 45
  mirror_hardwares: [amdexperimental]
  torch_nightly: true
  source_file_dependencies:
  - vllm/
-  - tests/models
+  - tests/models/test_initialization.py
  commands:
-    - pytest -v -s models/test_transformers.py
-    - pytest -v -s models/test_registry.py
-    - pytest -v -s models/test_utils.py
-    - pytest -v -s models/test_vision.py
-    - pytest -v -s models/test_initialization.py
+    # Run a subset of model initialization tests
+    - pytest -v -s models/test_initialization.py::test_can_initialize_small_subset

- label: Language Models Test (Standard)
+- label: Basic Models Tests (Extra Initialization) %N
+  timeout_in_minutes: 45
+  mirror_hardwares: [amdexperimental]
+  torch_nightly: true
+  source_file_dependencies:
+  - vllm/model_executor/models/
+  - tests/models/test_initialization.py
+  commands:
+    # Only when vLLM model source is modified - test initialization of a large
+    # subset of supported models (the complement of the small subset in the above
+    # test.) Also run if model initialization test file is modified
+    - pytest -v -s models/test_initialization.py \
+             -k 'not test_can_initialize_small_subset' \
+             --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT \
+             --shard-id=$$BUILDKITE_PARALLEL_JOB
+  parallelism: 2
+
+- label: Basic Models Tests (Other)
+  timeout_in_minutes: 45
+  mirror_hardwares: [amdexperimental]
+  torch_nightly: true
+  source_file_dependencies:
+  - vllm/
+  - tests/models/test_transformers.py
+  - tests/models/test_registry.py
+  - tests/models/test_utils.py
+  - tests/models/test_vision.py
+  commands:
+    - pytest -v -s models/test_transformers.py \
+                   models/test_registry.py \
+                   models/test_utils.py \
+                   models/test_vision.py
+
+- label: Language Models Tests (Standard)
+  timeout_in_minutes: 25
  mirror_hardwares: [amdexperimental]
  torch_nightly: true
  source_file_dependencies:
  - vllm/
  - tests/models/language
  commands:
+    # Test standard language models, excluding a subset of slow tests
    - pip freeze | grep -E 'torch'
-    - pytest -v -s models/language -m core_model
+    - pytest -v -s models/language -m 'core_model and (not slow_test)'

- label: Language Models Test (Hybrid) # 35 min
+- label: Language Models Tests (Extra Standard) %N
+  timeout_in_minutes: 45
+  mirror_hardwares: [amdexperimental]
+  torch_nightly: true
+  source_file_dependencies:
+  - vllm/model_executor/models/
+  - tests/models/language/pooling/test_embedding.py
+  - tests/models/language/generation/test_common.py
+  - tests/models/language/pooling/test_classification.py
+  commands:
+    # Shard slow subset of standard language models tests. Only run when model
+    # source is modified, or when specified test files are modified
+    - pip freeze | grep -E 'torch'
+    - pytest -v -s models/language -m 'core_model and slow_test' \
+             --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT \
+             --shard-id=$$BUILDKITE_PARALLEL_JOB
+  parallelism: 2
+
+- label: Language Models Tests (Hybrid) %N
+  timeout_in_minutes: 75
  mirror_hardwares: [amdexperimental]
  torch_nightly: true
  source_file_dependencies:
@ -540,9 +657,15 @@ steps:
    # Note: also needed to run plamo2 model in vLLM
    - uv pip install --system --no-build-isolation 'git+https://github.com/state-spaces/mamba@v2.2.5'
    - uv pip install --system --no-build-isolation 'git+https://github.com/Dao-AILab/causal-conv1d@v1.5.2'
-    - pytest -v -s models/language/generation -m hybrid_model
+    # Shard hybrid language model tests
+    - pytest -v -s models/language/generation \
+                   -m hybrid_model \
+                   --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT \
+                   --shard-id=$$BUILDKITE_PARALLEL_JOB
+  parallelism: 2

- label: Language Models Test (Extended Generation) # 1hr20min
+- label: Language Models Test (Extended Generation) # 80min
+  timeout_in_minutes: 110
  mirror_hardwares: [amdexperimental]
  optional: true
  source_file_dependencies:
@ -553,7 +676,18 @@ steps:
    - pip install 'git+https://github.com/Dao-AILab/causal-conv1d@v1.5.0.post8'
    - pytest -v -s models/language/generation -m '(not core_model) and (not hybrid_model)'

+- label: Language Models Test (PPL)
+  timeout_in_minutes: 110
+  mirror_hardwares: [amdexperimental]
+  optional: true
+  source_file_dependencies:
+  - vllm/
+  - tests/models/language/generation_ppl_test
+  commands:
+    - pytest -v -s models/language/generation_ppl_test
+
 - label: Language Models Test (Extended Pooling)  # 36min
+  timeout_in_minutes: 50
  mirror_hardwares: [amdexperimental]
  optional: true
  source_file_dependencies:
@ -562,7 +696,27 @@ steps:
  commands:
    - pytest -v -s models/language/pooling -m 'not core_model'

- label: Multi-Modal Models Test (Standard)
+- label: Language Models Test (MTEB)
+  timeout_in_minutes: 110
+  mirror_hardwares: [amdexperimental]
+  optional: true
+  source_file_dependencies:
+  - vllm/
+  - tests/models/language/pooling_mteb_test
+  commands:
+    - pytest -v -s models/language/pooling_mteb_test
+
+- label: Multi-Modal Processor Test # 44min
+  timeout_in_minutes: 60
+  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
+
+- label: Multi-Modal Models Test (Standard) # 60min
+  timeout_in_minutes: 80
  mirror_hardwares: [amdexperimental]
  torch_nightly: true
  source_file_dependencies:
@ -571,10 +725,8 @@ 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"
-    - cd .. && pytest -v -s tests/models/multimodal/generation/test_whisper.py -m core_model  # Otherwise, mp_method="spawn" doesn't work
+    - pytest -v -s models/multimodal -m core_model --ignore models/multimodal/generation/test_whisper.py --ignore models/multimodal/processing
+    - cd .. && VLLM_WORKER_MULTIPROC_METHOD=spawn pytest -v -s tests/models/multimodal/generation/test_whisper.py -m core_model  # Otherwise, mp_method="spawn" doesn't work

 - label: Multi-Modal Models Test (Extended) 1
  mirror_hardwares: [amdexperimental]
@ -584,7 +736,7 @@ 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]
@ -606,7 +758,8 @@ steps:
    - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
    - pytest -v -s models/multimodal/generation/test_common.py -m 'split(group=1) and not core_model'

- label: Quantized Models Test
+- label: Quantized Models Test # 45 min
+  timeout_in_minutes: 60
  mirror_hardwares: [amdexperimental]
  source_file_dependencies:
  - vllm/model_executor/layers/quantization
@ -636,7 +789,8 @@ steps:
    - python3 examples/offline_inference/audio_language.py --model-type whisper
    - python3 examples/offline_inference/vision_language.py --model-type qwen2_5_vl

- label: Blackwell Test
+- label: Blackwell Test # 38 min
+  timeout_in_minutes: 60
  working_dir: "/vllm-workspace/"
  gpu: b200
  # optional: true
@ -647,8 +801,10 @@ 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
@ -656,20 +812,28 @@ steps:
    # num_heads2 broken by https://github.com/flashinfer-ai/flashinfer/issues/1353
    - pytest -v -s tests/kernels/attention/test_flashinfer.py -k 'not num_heads2'
    - pytest -v -s tests/kernels/attention/test_flashinfer_trtllm_attention.py
-    - pytest -v -s tests/kernels/test_cutlass_mla_decode.py
+    - pytest -v -s tests/kernels/attention/test_cutlass_mla_decode.py
+    - pytest -v -s tests/kernels/attention/test_flashinfer_mla_decode.py
    # Quantization
    - pytest -v -s tests/kernels/quantization/test_cutlass_scaled_mm.py -k 'fp8'
    - pytest -v -s tests/kernels/quantization/test_nvfp4_quant.py
+    - pytest -v -s tests/kernels/quantization/test_silu_nvfp4_quant_fusion.py
    - pytest -v -s tests/kernels/quantization/test_nvfp4_scaled_mm.py
+    - pytest -v -s tests/kernels/quantization/test_flashinfer_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
+    - pytest -v -s tests/compile/test_silu_mul_quant_fusion.py

 #####  1 GPU test  #####
 #####  multi gpus test  #####

 - label: Distributed Comm Ops Test # 7min
+  timeout_in_minutes: 20
  mirror_hardwares: [amdexperimental]
  working_dir: "/vllm-workspace/tests"
  num_gpus: 2
@ -679,8 +843,11 @@ steps:
  commands:
  - pytest -v -s distributed/test_comm_ops.py
  - pytest -v -s distributed/test_shm_broadcast.py
+  - pytest -v -s distributed/test_shm_buffer.py
+  - pytest -v -s distributed/test_shm_storage.py

 - label: 2 Node Tests (4 GPUs in total) # 16min
+  timeout_in_minutes: 30
  mirror_hardwares: [amdexperimental]
  working_dir: "/vllm-workspace/tests"
  num_gpus: 2
@ -704,7 +871,8 @@ steps:
    - NUM_NODES=2 torchrun --nnodes 2 --nproc-per-node=2 --rdzv_backend=c10d --rdzv_endpoint=192.168.10.10 distributed/test_node_count.py | grep 'Node count test passed'
    - python3 ../examples/offline_inference/data_parallel.py --dp-size=2 --tp-size=1 --node-size=2 --node-rank=1 --master-addr=192.168.10.10 --master-port=12345 --enforce-eager --trust-remote-code

- label: Distributed Tests (2 GPUs) # 40min
+- label: Distributed Tests (2 GPUs) # 110min
+  timeout_in_minutes: 150
  mirror_hardwares: [amdexperimental]
  working_dir: "/vllm-workspace/tests"
  num_gpus: 2
@ -735,7 +903,8 @@ steps:
  # Avoid importing model tests that cause CUDA reinitialization error
  - pytest models/test_transformers.py -v -s -m 'distributed(num_gpus=2)'
  - pytest models/language -v -s -m 'distributed(num_gpus=2)'
-  - pytest models/multimodal -v -s -m 'distributed(num_gpus=2)'
+  - pytest models/multimodal -v -s -m 'distributed(num_gpus=2)' --ignore models/multimodal/generation/test_whisper.py
+  - VLLM_WORKER_MULTIPROC_METHOD=spawn pytest models/multimodal/generation/test_whisper.py -v -s -m 'distributed(num_gpus=2)'
  # test sequence parallel
  - pytest -v -s distributed/test_sequence_parallel.py
  # this test fails consistently.
@ -745,6 +914,7 @@ steps:
  - pytest -v -s models/multimodal/generation/test_maverick.py

 - label: Plugin Tests (2 GPUs) # 40min
+  timeout_in_minutes: 60
  mirror_hardwares: [amdexperimental]
  working_dir: "/vllm-workspace/tests"
  num_gpus: 2
@ -757,6 +927,11 @@ steps:
  - pytest -v -s plugins_tests/test_platform_plugins.py
  - pip uninstall vllm_add_dummy_platform -y
  # end platform plugin tests
+  # begin io_processor plugins test, all the code in between uses the prithvi_io_processor plugin
+  - pip install -e ./plugins/prithvi_io_processor_plugin
+  - pytest -v -s plugins_tests/test_io_processor_plugins.py
+  - pip uninstall prithvi_io_processor_plugin -y
+  # end io_processor plugins test
  # other tests continue here:
  - pytest -v -s plugins_tests/test_scheduler_plugins.py
  - pip install -e ./plugins/vllm_add_dummy_model
@ -765,7 +940,8 @@ steps:
  - pytest -v -s models/test_oot_registration.py # it needs a clean process
  - pytest -v -s plugins/lora_resolvers # unit tests for in-tree lora resolver plugins

- label: Pipeline Parallelism Test # 45min
+- label: Pipeline + Context Parallelism Test # 45min
+  timeout_in_minutes: 60
  mirror_hardwares: [amdexperimental]
  working_dir: "/vllm-workspace/tests"
  num_gpus: 4
@ -778,8 +954,10 @@ steps:
  commands:
  - pytest -v -s distributed/test_pp_cudagraph.py
  - pytest -v -s distributed/test_pipeline_parallel.py
+  # - pytest -v -s distributed/test_context_parallel.py # TODO: enable it on Hopper runners or add triton MLA support

- label: LoRA TP Test (Distributed)
+- label: LoRA TP Test (Distributed) # 17 min
+  timeout_in_minutes: 30
  mirror_hardwares: [amdexperimental]
  num_gpus: 4
  source_file_dependencies:
@ -793,13 +971,15 @@ steps:
    # requires multi-GPU testing for validation.
    - pytest -v -s -x lora/test_chatglm3_tp.py
    - pytest -v -s -x lora/test_llama_tp.py
-    - pytest -v -s -x lora/test_multi_loras_with_tp.py
+    - pytest -v -s -x lora/test_llm_with_multi_loras.py


 - label: Weight Loading Multiple GPU Test  # 33min
+  timeout_in_minutes: 45
  mirror_hardwares: [amdexperimental]
  working_dir: "/vllm-workspace/tests"
  num_gpus: 2
+  optional: true
  source_file_dependencies:
  - vllm/
  - tests/weight_loading
@ -847,3 +1027,10 @@ 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/.bc-linter.yml
+++ b/.github/.bc-linter.yml
@ -0,0 +1,24 @@
+# doc: https://github.com/pytorch/test-infra/blob/main/tools/stronghold/docs/bc_linter_config.md
+version: 1
+paths:
+# We temporarily disable globally, and will only enable with `annotations.include`
+# include:
+#   - "vllm/v1/attetion/*.py"
+#   - "vllm/v1/core/*.py"
+exclude:
+  - "**/*.py"
+
+scan:
+  functions: true        # check free functions and methods
+  classes: true          # check classes/dataclasses
+  public_only: true      # ignore names starting with "_" at any level
+
+annotations:
+  include:               # decorators that force‑include a symbol
+    - name: "bc_linter_include"  # matched by simple name or dotted suffix
+      propagate_to_members: false # for classes, include methods/inner classes
+  exclude:               # decorators that force‑exclude a symbol
+    - name: "bc_linter_skip"     # matched by simple name or dotted suffix
+      propagate_to_members: true  # for classes, exclude methods/inner classes
+
+excluded_violations: []  # e.g. ["ParameterRenamed", "FieldTypeChanged"]
--- a/.github/CODEOWNERS
+++ b/.github/CODEOWNERS
@ -5,17 +5,21 @@
 /vllm/attention/backends/abstract.py @WoosukKwon @zhuohan123 @youkaichao @alexm-redhat @comaniac @njhill
 /vllm/core @zhuohan123 @youkaichao @alexm-redhat @comaniac @njhill
 /vllm/engine/llm_engine.py @zhuohan123 @youkaichao @alexm-redhat @comaniac @njhill
-/vllm/executor/executor_base.py @zhuohan123 @youkaichao @alexm-redhat @comaniac @njhill
-/vllm/worker/worker_base.py @zhuohan123 @youkaichao @alexm-redhat @comaniac @njhill
+/vllm/executor/executor_base.py @zhuohan123 @youkaichao @alexm-redhat @comaniac @njhill @22quinn
+/vllm/worker/worker_base.py @zhuohan123 @youkaichao @alexm-redhat @comaniac @njhill @22quinn
 /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/sampler.py @zhuohan123 @youkaichao @alexm-redhat @comaniac @njhill @NickLucche
 /vllm/model_executor/layers/quantization @mgoin @robertgshaw2-redhat @tlrmchlsmth @yewentao256
-/vllm/multimodal @DarkLight1337 @ywang96
+/vllm/model_executor/layers/mamba @tdoublep
+/vllm/model_executor/model_loader @22quinn
+/vllm/multimodal @DarkLight1337 @ywang96 @NickLucche
+/vllm/v1/sample @22quinn @houseroad
 /vllm/vllm_flash_attn @LucasWilkinson
 /vllm/lora @jeejeelee
-/vllm/reasoning @aarnphm
-/vllm/entrypoints @aarnphm
+/vllm/reasoning @aarnphm @chaunceyjiang
+/vllm/entrypoints @aarnphm @chaunceyjiang
 /vllm/compilation @zou3519 @youkaichao @ProExpertProg
+/vllm/distributed/kv_transfer @NickLucche
 CMakeLists.txt @tlrmchlsmth @LucasWilkinson

 # Any change to the VllmConfig changes can have a large user-facing impact,
@ -24,26 +28,32 @@ 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/structured_output @mgoin @russellb @aarnphm @benchislett
+/vllm/v1/spec_decode @benchislett @luccafong
+/vllm/v1/attention/backends/triton_attn.py @tdoublep
+/vllm/v1/core @heheda12345
+/vllm/v1/kv_cache_interface.py @heheda12345

 # 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
-/tests/entrypoints @DarkLight1337 @robertgshaw2-redhat @simon-mo @aarnphm
+/tests/entrypoints @DarkLight1337 @robertgshaw2-redhat @simon-mo @aarnphm @NickLucche
 /tests/kernels @tlrmchlsmth @WoosukKwon @yewentao256
 /tests/models @DarkLight1337 @ywang96
-/tests/multimodal @DarkLight1337 @ywang96
+/tests/multimodal @DarkLight1337 @ywang96 @NickLucche
 /tests/prefix_caching @comaniac @KuntaiDu
 /tests/quantization @mgoin @robertgshaw2-redhat @yewentao256
 /tests/test_inputs.py @DarkLight1337 @ywang96
 /tests/v1/entrypoints/llm/test_struct_output_generate.py @mgoin @russellb @aarnphm
 /tests/v1/structured_output @mgoin @russellb @aarnphm
+/tests/v1/core @heheda12345
 /tests/weight_loading @mgoin @youkaichao @yewentao256
 /tests/lora @jeejeelee
+/tests/models/language/generation/test_hybrid.py @tdoublep
+/tests/v1/kv_connector/nixl_integration @NickLucche

 # Docs
 /docs @hmellor
@ -65,6 +75,9 @@ mkdocs.yaml @hmellor
 /vllm/attention/backends/dual_chunk_flash_attn.py @sighingnow
 /vllm/model_executor/models/qwen* @sighingnow

+# MTP-specific files
+/vllm/model_executor/models/deepseek_mtp.py @luccafong
+
 # Mistral-specific files
 /vllm/model_executor/models/mistral*.py @patrickvonplaten
 /vllm/model_executor/models/mixtral*.py @patrickvonplaten
@ -72,3 +85,20 @@ 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
+
+# ROCm related: specify owner with write access to notify AMD folks for careful code review
+/docker/Dockerfile.rocm* @gshtras
+/vllm/v1/attention/backends/rocm*.py @gshtras
+/vllm/v1/attention/backends/mla/rocm*.py @gshtras
+/vllm/attention/ops/rocm*.py @gshtras
+/vllm/model_executor/layers/fused_moe/rocm*.py @gshtras
+
+# TPU
+/vllm/v1/worker/tpu* @NickLucche
+/vllm/platforms/tpu.py @NickLucche
+/vllm/v1/sample/tpu @NickLucche
+/vllm/tests/v1/tpu @NickLucche
--- 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/mergify.yml
+++ b/.github/mergify.yml
@ -124,9 +124,16 @@ pull_request_rules:
    - or:
      - files~=^examples/.*gpt[-_]?oss.*\.py
      - files~=^tests/.*gpt[-_]?oss.*\.py
+      - files~=^tests/entrypoints/openai/test_response_api_with_harmony.py
+      - files~=^tests/entrypoints/test_context.py
      - files~=^vllm/model_executor/models/.*gpt[-_]?oss.*\.py
      - files~=^vllm/model_executor/layers/.*gpt[-_]?oss.*\.py
+      - files~=^vllm/entrypoints/harmony_utils.py
+      - files~=^vllm/entrypoints/tool_server.py
+      - files~=^vllm/entrypoints/tool.py
+      - files~=^vllm/entrypoints/context.py
      - title~=(?i)gpt[-_]?oss
+      - title~=(?i)harmony
  actions:
    label:
      add:
@ -273,6 +280,20 @@ pull_request_rules:
      users:
        - "sangstar"

+- name: assign reviewer for modelopt changes
+  conditions:
+    - or:
+        - files~=^vllm/model_executor/layers/quantization/modelopt\.py$
+        - files~=^vllm/model_executor/layers/quantization/__init__\.py$
+        - files~=^tests/models/quantization/test_modelopt\.py$
+        - files~=^tests/quantization/test_modelopt\.py$
+        - files~=^tests/models/quantization/test_nvfp4\.py$
+        - files~=^docs/features/quantization/modelopt\.md$
+  actions:
+    assign:
+      users:
+        - "Edwardf0t1"
+
 - name: remove 'needs-rebase' label when conflict is resolved
  conditions:
      - -conflict
--- a/.github/scale-config.yml
+++ b/.github/scale-config.yml
@ -0,0 +1,21 @@
+# scale-config.yml:
+#   Powers what instance types are available for GHA auto-scaled
+#   runners. Runners listed here will be available as self hosted
+#   runners, configuration is directly pulled from the main branch.
+# runner_types:
+#   runner_label:
+#     instance_type: m4.large
+#     os: linux
+#     # min_available defaults to the global cfg in the ALI Terraform
+#     min_available: undefined
+#     # when max_available value is not defined, no max runners is enforced
+#     max_available: undefined
+#     disk_size: 50
+#     is_ephemeral: true
+
+runner_types:
+  linux.2xlarge:
+    disk_size: 150
+    instance_type: c5.2xlarge
+    is_ephemeral: true
+    os: linux
--- a/.github/workflows/add_label_automerge.yml
+++ b/.github/workflows/add_label_automerge.yml
@ -10,7 +10,7 @@ jobs:
        runs-on: ubuntu-latest
        steps:
            -   name: Add label
-                uses: actions/github-script@60a0d83039c74a4aee543508d2ffcb1c3799cdea # v7.0.1
+                uses: actions/github-script@ed597411d8f924073f98dfc5c65a23a2325f34cd # v8.0.0
                with:
                    script: |
                        github.rest.issues.addLabels({
--- a/.github/workflows/bc-lint.yml
+++ b/.github/workflows/bc-lint.yml
@ -0,0 +1,29 @@
+name: BC Lint
+
+on:
+  pull_request:
+    types:
+      - opened
+      - synchronize
+      - reopened
+      - labeled
+      - unlabeled
+
+jobs:
+  bc_lint:
+    if: github.repository_owner == 'vllm-project'
+    runs-on: ubuntu-latest
+    steps:
+      - name: Run BC Lint Action
+        uses: pytorch/test-infra/.github/actions/bc-lint@main
+        with:
+          repo: ${{ github.event.pull_request.head.repo.full_name }}
+          base_sha: ${{ github.event.pull_request.base.sha }}
+          head_sha: ${{ github.event.pull_request.head.sha }}
+          suppression: ${{ contains(github.event.pull_request.labels.*.name, 'suppress-bc-linter') }}
+          docs_link: 'https://github.com/pytorch/test-infra/wiki/BC-Linter'
+          config_dir: .github
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.event.pull_request.number || github.sha }}
+  cancel-in-progress: true
--- a/.github/workflows/cleanup_pr_body.yml
+++ b/.github/workflows/cleanup_pr_body.yml
@ -16,7 +16,7 @@ jobs:
        uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2

      - name: Set up Python
-        uses: actions/setup-python@42375524e23c412d93fb67b49958b491fce71c38 # v5.4.0
+        uses: actions/setup-python@e797f83bcb11b83ae66e0230d6156d7c80228e7c # v6.0.0
        with:
          python-version: '3.12'

--- a/.github/workflows/issue_autolabel.yml
+++ b/.github/workflows/issue_autolabel.yml
@ -0,0 +1,309 @@
+name: Label issues based on keywords
+on:
+  issues:
+    types: [opened, edited, reopened]
+permissions:
+  issues: write          # needed so the workflow can add labels
+  contents: read
+concurrency:
+  group: issue-labeler-${{ github.event.issue.number }}
+  cancel-in-progress: true
+jobs:
+  add-labels:
+    runs-on: ubuntu-latest
+    steps:
+      - name: Label issues based on keywords
+        uses: actions/github-script@ed597411d8f924073f98dfc5c65a23a2325f34cd  # v8.0.0
+        with:
+          script: |
+            // Configuration: Add new labels and keywords here
+            const labelConfig = {
+              rocm: {
+                // Keyword search - matches whole words only (with word boundaries)
+                keywords: [
+                  {
+                    term: "composable kernel",
+                    searchIn: "both"
+                  },
+                  {
+                    term: "rccl",
+                    searchIn: "body"  // only search in body
+                  },
+                  {
+                    term: "migraphx",
+                    searchIn: "title"  // only search in title
+                  },
+                  {
+                    term: "hipgraph",
+                    searchIn: "both"
+                  },
+                  {
+                    term: "ROCm System Management Interface",
+                    searchIn: "body"
+                  },
+                ],
+                
+                // Substring search - matches anywhere in text (partial matches)
+                substrings: [
+                  {
+                    term: "VLLM_ROCM_",
+                    searchIn: "both"
+                  },
+                  {
+                    term: "aiter",
+                    searchIn: "title"
+                  },
+                  {
+                    term: "rocm",
+                    searchIn: "title"
+                  },
+                  {
+                    term: "amd",
+                    searchIn: "title"
+                  },
+                  {
+                    term: "hip-",
+                    searchIn: "both"
+                  },
+                  {
+                    term: "gfx",
+                    searchIn: "both"
+                  },
+                  {
+                    term: "cdna",
+                    searchIn: "both"
+                  },
+                  {
+                    term: "rdna",
+                    searchIn: "both"
+                  },
+                  {
+                    term: "torch_hip",
+                    searchIn: "body"  // only in body
+                  },
+                  {
+                    term: "_hip",
+                    searchIn: "both"
+                  },
+                  {
+                    term: "hip_",
+                    searchIn: "both"
+                  },
+                  
+                  // ROCm tools and libraries
+                  {
+                    term: "hipify",
+                    searchIn: "both"
+                  },
+                ],
+                
+                // Regex patterns - for complex pattern matching
+                regexPatterns: [
+                  {
+                    pattern: "\\bmi\\d{3}[a-z]*\\b",
+                    description: "AMD GPU names (mi + 3 digits + optional letters)",
+                    flags: "gi",
+                    searchIn: "both"  // "title", "body", or "both"
+                  }
+                ],
+              },
+            };
+            
+            // Helper function to create regex based on search type
+            function createSearchRegex(term, type) {
+              // Escape special regex characters in the term
+              const escapedTerm = term.replace(/[.*+?^${}()|[\]\\]/g, '\\$&');
+              
+              switch (type) {
+                case 'keyword':
+                  // Word boundary search - matches whole words only
+                  return new RegExp(`\\b${escapedTerm}\\b`, "gi");
+                case 'substring':
+                  // Substring search - matches anywhere in the text
+                  return new RegExp(escapedTerm, "gi");
+                default:
+                  throw new Error(`Unknown search type: ${type}`);
+              }
+            }
+            
+            // Helper function to find matching terms in text with line information
+            function findMatchingTermsWithLines(text, searchTerms = [], searchType = 'keyword', searchLocation = '') {
+              const matches = [];
+              const lines = text.split('\n');
+              
+              for (const termConfig of searchTerms) {
+                let regex;
+                let term, searchIn, pattern, description, flags;
+                
+                // Handle different input formats (string or object)
+                if (typeof termConfig === 'string') {
+                  term = termConfig;
+                  searchIn = 'both'; // default
+                } else {
+                  term = termConfig.term;
+                  searchIn = termConfig.searchIn || 'both';
+                  pattern = termConfig.pattern;
+                  description = termConfig.description;
+                  flags = termConfig.flags;
+                }
+                
+                // Skip if this term shouldn't be searched in the current location
+                if (searchIn !== 'both' && searchIn !== searchLocation) {
+                  continue;
+                }
+                
+                // Create appropriate regex
+                if (searchType === 'regex') {
+                  regex = new RegExp(pattern, flags || "gi");
+                } else {
+                  regex = createSearchRegex(term, searchType);
+                }
+                
+                const termMatches = [];
+                
+                // Check each line for matches
+                lines.forEach((line, lineIndex) => {
+                  const lineMatches = line.match(regex);
+                  if (lineMatches) {
+                    lineMatches.forEach(match => {
+                      termMatches.push({
+                        match: match,
+                        lineNumber: lineIndex + 1,
+                        lineContent: line.trim(),
+                        searchType: searchType,
+                        searchLocation: searchLocation,
+                        originalTerm: term || pattern,
+                        description: description,
+                        // Show context around the match in the line
+                        context: line.length > 100 ? 
+                          line.substring(Math.max(0, line.toLowerCase().indexOf(match.toLowerCase()) - 30), 
+                                       line.toLowerCase().indexOf(match.toLowerCase()) + match.length + 30) + '...' 
+                          : line.trim()
+                      });
+                    });
+                  }
+                });
+                
+                if (termMatches.length > 0) {
+                  matches.push({
+                    term: term || (description || pattern),
+                    searchType: searchType,
+                    searchLocation: searchLocation,
+                    searchIn: searchIn,
+                    pattern: pattern,
+                    matches: termMatches,
+                    count: termMatches.length
+                  });
+                }
+              }
+              
+              return matches;
+            }
+            
+            // Helper function to check if label should be added
+            async function processLabel(labelName, config) {
+              const body = context.payload.issue.body || "";
+              const title = context.payload.issue.title || "";
+              
+              core.notice(`Processing label: ${labelName}`);
+              core.notice(`Issue Title: "${title}"`);
+              core.notice(`Issue Body length: ${body.length} characters`);
+              
+              let shouldAddLabel = false;
+              let allMatches = [];
+              let reason = '';
+              
+              const keywords = config.keywords || [];
+              const substrings = config.substrings || [];
+              const regexPatterns = config.regexPatterns || [];
+              
+              core.notice(`Searching with ${keywords.length} keywords, ${substrings.length} substrings, and ${regexPatterns.length} regex patterns`);
+              
+              // Search in title
+              if (title.trim()) {
+                core.notice(`Searching in title: "${title}"`);
+                
+                const titleKeywordMatches = findMatchingTermsWithLines(title, keywords, 'keyword', 'title');
+                const titleSubstringMatches = findMatchingTermsWithLines(title, substrings, 'substring', 'title');
+                const titleRegexMatches = findMatchingTermsWithLines(title, regexPatterns, 'regex', 'title');
+                
+                allMatches.push(...titleKeywordMatches, ...titleSubstringMatches, ...titleRegexMatches);
+              }
+              
+              // Search in body
+              if (body.trim()) {
+                core.notice(`Searching in body (${body.length} characters)`);
+                
+                const bodyKeywordMatches = findMatchingTermsWithLines(body, keywords, 'keyword', 'body');
+                const bodySubstringMatches = findMatchingTermsWithLines(body, substrings, 'substring', 'body');
+                const bodyRegexMatches = findMatchingTermsWithLines(body, regexPatterns, 'regex', 'body');
+                
+                allMatches.push(...bodyKeywordMatches, ...bodySubstringMatches, ...bodyRegexMatches);
+              }
+              
+              if (allMatches.length > 0) {
+                core.notice(`Found ${allMatches.length} matching term(s):`);
+                
+                for (const termMatch of allMatches) {
+                  const locationText = termMatch.searchLocation === 'title' ? 'title' : 'body';
+                  const searchInText = termMatch.searchIn === 'both' ? 'both' : termMatch.searchIn;
+                  
+                  if (termMatch.searchType === 'regex') {
+                    core.notice(`  📍 Regex: "${termMatch.term}" (pattern: ${termMatch.pattern}) found ${termMatch.count} time(s) in ${locationText} (configured to search in: ${searchInText}):`);
+                  } else {
+                    core.notice(`  📍 Term: "${termMatch.term}" (${termMatch.searchType} search) found ${termMatch.count} time(s) in ${locationText} (configured to search in: ${searchInText}):`);
+                  }
+                  
+                  // Show details for each match
+                  termMatch.matches.forEach((match, index) => {
+                    core.notice(`    ${index + 1}. Line ${match.lineNumber} in ${match.searchLocation}: "${match.match}" [${match.searchType}]`);
+                    if (match.description) {
+                      core.notice(`       Description: ${match.description}`);
+                    }
+                    core.notice(`       Context: ${match.context}`);
+                    if (match.lineContent !== match.context) {
+                      core.notice(`       Full line: ${match.lineContent}`);
+                    }
+                  });
+                }
+                
+                shouldAddLabel = true;
+                const totalMatches = allMatches.reduce((sum, t) => sum + t.count, 0);
+                const titleMatches = allMatches.filter(t => t.searchLocation === 'title').reduce((sum, t) => sum + t.count, 0);
+                const bodyMatches = allMatches.filter(t => t.searchLocation === 'body').reduce((sum, t) => sum + t.count, 0);
+                const keywordMatches = allMatches.filter(t => t.searchType === 'keyword').reduce((sum, t) => sum + t.count, 0);
+                const substringMatches = allMatches.filter(t => t.searchType === 'substring').reduce((sum, t) => sum + t.count, 0);
+                const regexMatches = allMatches.filter(t => t.searchType === 'regex').reduce((sum, t) => sum + t.count, 0);
+                
+                reason = `Found ${totalMatches} total matches (${titleMatches} in title, ${bodyMatches} in body) - ${keywordMatches} keyword matches, ${substringMatches} substring matches, ${regexMatches} regex matches`;
+              }
+              
+              core.notice(`Final decision: ${shouldAddLabel ? 'ADD LABEL' : 'DO NOT ADD LABEL'}`);
+              core.notice(`Reason: ${reason || 'No matching terms found'}`);
+              
+              if (shouldAddLabel) {
+                const existingLabels = context.payload.issue.labels.map(l => l.name);
+                if (!existingLabels.includes(labelName)) {
+                  await github.rest.issues.addLabels({
+                    owner: context.repo.owner,
+                    repo: context.repo.repo,
+                    issue_number: context.issue.number,
+                    labels: [labelName],
+                  });
+                  core.notice(`Label "${labelName}" added. ${reason}`);
+                  return true;
+                }
+                core.notice(`Label "${labelName}" already present.`);
+                return false;
+              }
+              
+              core.notice(`No matching terms found for label "${labelName}".`);
+              return false;
+            }
+            
+            // Process all configured labels
+            const processLabels = Object.entries(labelConfig)
+              .map(([labelName, config]) => processLabel(labelName, config));
+            const labelsAdded = await Promise.all(processLabels);
+            const numLabelsAdded = labelsAdded.reduce((x, y) => x + y, 0);
+            core.notice(`Processing complete. ${numLabelsAdded} label(s) added.`);
--- 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/pre-commit.yml
+++ b/.github/workflows/pre-commit.yml
@ -17,7 +17,7 @@ jobs:
    runs-on: ubuntu-latest
    steps:
    - uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
-    - uses: actions/setup-python@42375524e23c412d93fb67b49958b491fce71c38 # v5.4.0
+    - uses: actions/setup-python@e797f83bcb11b83ae66e0230d6156d7c80228e7c # v6.0.0
      with:
        python-version: "3.12"
    - run: echo "::add-matcher::.github/workflows/matchers/actionlint.json"
--- 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
@ -9,19 +9,46 @@ jobs:
    runs-on: ubuntu-latest
    steps:
      - name: Remind to run full CI on PR
-        uses: actions/github-script@60a0d83039c74a4aee543508d2ffcb1c3799cdea # v7.0.1
+        uses: actions/github-script@ed597411d8f924073f98dfc5c65a23a2325f34cd # v8.0.0
        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/.github/workflows/stale.yml
+++ b/.github/workflows/stale.yml
@ -13,7 +13,7 @@ jobs:
      actions: write
    runs-on: ubuntu-latest
    steps:
-      - uses: actions/stale@5bef64f19d7facfb25b37b414482c7164d639639 # v9.1.0
+      - uses: actions/stale@3a9db7e6a41a89f618792c92c0e97cc736e1b13f # v10.0.0
        with:
          # Increasing this value ensures that changes to this workflow
          # propagate to all issues and PRs in days rather than months
--- a/.gitignore
+++ b/.gitignore
@ -4,7 +4,7 @@
 # vllm-flash-attn built from source
 vllm/vllm_flash_attn/*

-# triton jit 
+# triton jit
 .triton

 # Byte-compiled / optimized / DLL files
@ -177,6 +177,14 @@ cython_debug/
 # VSCode
 .vscode/

+# Claude
+CLAUDE.md
+.claude/
+
+# Codex
+AGENTS.md
+.codex/
+
 # DS Store
 .DS_Store

@ -209,4 +217,4 @@ shellcheck*/
 csrc/moe/marlin_moe_wna16/kernel_*

 # Ignore ep_kernels_workspace folder
-ep_kernels_workspace/
+ep_kernels_workspace/
--- a/.pre-commit-config.yaml
+++ b/.pre-commit-config.yaml
@ -21,7 +21,7 @@ repos:
  - id: ruff-format
    files: ^(.buildkite|benchmarks|examples)/.*
 - repo: https://github.com/crate-ci/typos
-  rev: v1.34.0
+  rev: v1.35.5
  hooks:
  - id: typos
 - repo: https://github.com/PyCQA/isort
--- a/.yapfignore
+++ b/.yapfignore
@ -1 +1,2 @@
 collect_env.py
+vllm/model_executor/layers/fla/ops/*.py
--- 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")
@ -45,8 +45,8 @@ set(HIP_SUPPORTED_ARCHS "gfx906;gfx908;gfx90a;gfx942;gfx950;gfx1030;gfx1100;gfx1
 # requirements.txt files and should be kept consistent.  The ROCm torch
 # versions are derived from docker/Dockerfile.rocm
 #
-set(TORCH_SUPPORTED_VERSION_CUDA "2.7.1")
-set(TORCH_SUPPORTED_VERSION_ROCM "2.7.0")
+set(TORCH_SUPPORTED_VERSION_CUDA "2.8.0")
+set(TORCH_SUPPORTED_VERSION_ROCM "2.8.0")

 #
 # Try to find python package with an executable that exactly matches
@ -243,8 +243,8 @@ set(VLLM_EXT_SRC
  "csrc/sampler.cu"
  "csrc/cuda_view.cu"
  "csrc/quantization/gptq/q_gemm.cu"
-  "csrc/quantization/compressed_tensors/int8_quant_kernels.cu"
-  "csrc/quantization/fp8/common.cu"
+  "csrc/quantization/w8a8/int8/scaled_quant.cu"
+  "csrc/quantization/w8a8/fp8/common.cu"
  "csrc/quantization/fused_kernels/fused_layernorm_dynamic_per_token_quant.cu"
  "csrc/quantization/gguf/gguf_kernel.cu"
  "csrc/quantization/activation_kernels.cu"
@ -288,14 +288,15 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
  list(APPEND VLLM_EXT_SRC
    "csrc/quantization/awq/gemm_kernels.cu"
    "csrc/permute_cols.cu"
-    "csrc/quantization/cutlass_w8a8/scaled_mm_entry.cu"
+    "csrc/quantization/w8a8/cutlass/scaled_mm_entry.cu"
    "csrc/quantization/fp4/nvfp4_quant_entry.cu"
    "csrc/quantization/fp4/nvfp4_scaled_mm_entry.cu"
    "csrc/quantization/fp4/nvfp4_blockwise_moe_kernel.cu"
    "csrc/sparse/cutlass/sparse_scaled_mm_entry.cu"
    "csrc/cutlass_extensions/common.cpp"
    "csrc/attention/mla/cutlass_mla_entry.cu"
-    "csrc/quantization/fp8/per_token_group_quant.cu")
+    "csrc/quantization/w8a8/fp8/per_token_group_quant.cu"
+    "csrc/quantization/w8a8/int8/per_token_group_quant.cu")

  set_gencode_flags_for_srcs(
    SRCS "${VLLM_EXT_SRC}"
@ -357,9 +358,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
    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")
@ -401,11 +400,11 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
  cuda_archs_loose_intersection(SCALED_MM_ARCHS "9.0a;" "${CUDA_ARCHS}")
  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.0 AND SCALED_MM_ARCHS)
    set(SRCS
-       "csrc/quantization/cutlass_w8a8/scaled_mm_c3x_sm90.cu"
-       "csrc/quantization/cutlass_w8a8/c3x/scaled_mm_sm90_fp8.cu"
-       "csrc/quantization/cutlass_w8a8/c3x/scaled_mm_sm90_int8.cu"
-       "csrc/quantization/cutlass_w8a8/c3x/scaled_mm_azp_sm90_int8.cu"
-       "csrc/quantization/cutlass_w8a8/c3x/scaled_mm_blockwise_sm90_fp8.cu")
+       "csrc/quantization/w8a8/cutlass/scaled_mm_c3x_sm90.cu"
+       "csrc/quantization/w8a8/cutlass/c3x/scaled_mm_sm90_fp8.cu"
+       "csrc/quantization/w8a8/cutlass/c3x/scaled_mm_sm90_int8.cu"
+       "csrc/quantization/w8a8/cutlass/c3x/scaled_mm_azp_sm90_int8.cu"
+       "csrc/quantization/w8a8/cutlass/c3x/scaled_mm_blockwise_sm90_fp8.cu")
    set_gencode_flags_for_srcs(
      SRCS "${SRCS}"
      CUDA_ARCHS "${SCALED_MM_ARCHS}")
@ -432,9 +431,9 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
  cuda_archs_loose_intersection(SCALED_MM_ARCHS "12.0;12.0a" "${CUDA_ARCHS}")
  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND SCALED_MM_ARCHS)
    set(SRCS
-      "csrc/quantization/cutlass_w8a8/scaled_mm_c3x_sm120.cu"
-      "csrc/quantization/cutlass_w8a8/c3x/scaled_mm_sm120_fp8.cu"
-      "csrc/quantization/cutlass_w8a8/c3x/scaled_mm_blockwise_sm120_fp8.cu"
+      "csrc/quantization/w8a8/cutlass/scaled_mm_c3x_sm120.cu"
+      "csrc/quantization/w8a8/cutlass/c3x/scaled_mm_sm120_fp8.cu"
+      "csrc/quantization/w8a8/cutlass/c3x/scaled_mm_blockwise_sm120_fp8.cu"
    )
    set_gencode_flags_for_srcs(
      SRCS "${SRCS}"
@ -462,9 +461,9 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
  cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0a;10.1a" "${CUDA_ARCHS}")
  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND SCALED_MM_ARCHS)
    set(SRCS
-      "csrc/quantization/cutlass_w8a8/scaled_mm_c3x_sm100.cu"
-      "csrc/quantization/cutlass_w8a8/c3x/scaled_mm_sm100_fp8.cu"
-      "csrc/quantization/cutlass_w8a8/c3x/scaled_mm_blockwise_sm100_fp8.cu"
+      "csrc/quantization/w8a8/cutlass/scaled_mm_c3x_sm100.cu"
+      "csrc/quantization/w8a8/cutlass/c3x/scaled_mm_sm100_fp8.cu"
+      "csrc/quantization/w8a8/cutlass/c3x/scaled_mm_blockwise_sm100_fp8.cu"
    )
    set_gencode_flags_for_srcs(
      SRCS "${SRCS}"
@ -495,7 +494,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
  # subtract out the archs that are already built for 3x
  list(REMOVE_ITEM SCALED_MM_2X_ARCHS ${SCALED_MM_3X_ARCHS})
  if (SCALED_MM_2X_ARCHS)
-    set(SRCS "csrc/quantization/cutlass_w8a8/scaled_mm_c2x.cu")
+    set(SRCS "csrc/quantization/w8a8/cutlass/scaled_mm_c2x.cu")
    set_gencode_flags_for_srcs(
      SRCS "${SRCS}"
      CUDA_ARCHS "${SCALED_MM_2X_ARCHS}")
@ -543,6 +542,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND FP4_ARCHS)
    set(SRCS
      "csrc/quantization/fp4/nvfp4_quant_kernels.cu"
+      "csrc/quantization/fp4/activation_nvfp4_quant_fusion_kernels.cu"
      "csrc/quantization/fp4/nvfp4_scaled_mm_sm120_kernels.cu")
    set_gencode_flags_for_srcs(
      SRCS "${SRCS}"
@ -561,6 +561,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND FP4_ARCHS)
    set(SRCS
      "csrc/quantization/fp4/nvfp4_quant_kernels.cu"
+      "csrc/quantization/fp4/activation_nvfp4_quant_fusion_kernels.cu"
      "csrc/quantization/fp4/nvfp4_experts_quant.cu"
      "csrc/quantization/fp4/nvfp4_scaled_mm_kernels.cu"
      "csrc/quantization/fp4/nvfp4_blockwise_moe_kernel.cu")
@ -605,7 +606,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
  # if it's possible to compile MoE kernels that use its output.
  cuda_archs_loose_intersection(SCALED_MM_ARCHS "9.0a" "${CUDA_ARCHS}")
  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.3 AND SCALED_MM_ARCHS)
-    set(SRCS "csrc/quantization/cutlass_w8a8/moe/grouped_mm_c3x_sm90.cu")
+    set(SRCS "csrc/quantization/w8a8/cutlass/moe/grouped_mm_c3x_sm90.cu")
    set_gencode_flags_for_srcs(
      SRCS "${SRCS}"
      CUDA_ARCHS "${SCALED_MM_ARCHS}")
@ -625,7 +626,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")

  cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0a" "${CUDA_ARCHS}")
  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND SCALED_MM_ARCHS)
-    set(SRCS "csrc/quantization/cutlass_w8a8/moe/grouped_mm_c3x_sm100.cu")
+    set(SRCS "csrc/quantization/w8a8/cutlass/moe/grouped_mm_c3x_sm100.cu")
    set_gencode_flags_for_srcs(
      SRCS "${SRCS}"
      CUDA_ARCHS "${SCALED_MM_ARCHS}")
@ -646,7 +647,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
  # moe_data.cu is used by all CUTLASS MoE kernels.
  cuda_archs_loose_intersection(CUTLASS_MOE_DATA_ARCHS "9.0a;10.0a" "${CUDA_ARCHS}")
  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.3 AND CUTLASS_MOE_DATA_ARCHS)
-    set(SRCS "csrc/quantization/cutlass_w8a8/moe/moe_data.cu")
+    set(SRCS "csrc/quantization/w8a8/cutlass/moe/moe_data.cu")
    set_gencode_flags_for_srcs(
      SRCS "${SRCS}"
      CUDA_ARCHS "${CUTLASS_MOE_DATA_ARCHS}")
@ -665,7 +666,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")

  cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0a" "${CUDA_ARCHS}")
  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND SCALED_MM_ARCHS)
-    set(SRCS "csrc/quantization/cutlass_w8a8/moe/blockwise_scaled_group_mm_sm100.cu")
+    set(SRCS "csrc/quantization/w8a8/cutlass/moe/blockwise_scaled_group_mm_sm100.cu")
    set_gencode_flags_for_srcs(
      SRCS "${SRCS}"
      CUDA_ARCHS "${SCALED_MM_ARCHS}")
@ -752,6 +753,33 @@ 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()

@ -792,7 +820,9 @@ 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")
--- a/MANIFEST.in
+++ b/MANIFEST.in
@ -2,7 +2,6 @@ include LICENSE
 include requirements/common.txt
 include requirements/cuda.txt
 include requirements/rocm.txt
-include requirements/neuron.txt
 include requirements/cpu.txt
 include CMakeLists.txt

--- a/README.md
+++ b/README.md
@ -14,18 +14,25 @@ Easy, fast, and cheap LLM serving for everyone
 | <a href="https://docs.vllm.ai"><b>Documentation</b></a> | <a href="https://blog.vllm.ai/"><b>Blog</b></a> | <a href="https://arxiv.org/abs/2309.06180"><b>Paper</b></a> | <a href="https://x.com/vllm_project"><b>Twitter/X</b></a> | <a href="https://discuss.vllm.ai"><b>User Forum</b></a> | <a href="https://slack.vllm.ai"><b>Developer Slack</b></a> |
 </p>

+---
+Join us at the [PyTorch Conference, October 22-23](https://events.linuxfoundation.org/pytorch-conference/) and [Ray Summit, November 3-5](https://www.anyscale.com/ray-summit/2025) in San Francisco for our latest updates on vLLM and to meet the vLLM team! Register now for the largest vLLM community events of the year!
+
 ---

 *Latest News* 🔥

- [2025/08] 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).
+- [2025/08] We hosted [vLLM Shenzhen Meetup](https://mp.weixin.qq.com/s/k8ZBO1u2_2odgiKWH_GVTQ) focusing on the ecosystem around vLLM! Please find the meetup slides [here](https://drive.google.com/drive/folders/1Ua2SVKVSu-wp5vou_6ElraDt2bnKhiEA).
+- [2025/08] We hosted [vLLM Singapore Meetup](https://www.sginnovate.com/event/vllm-sg-meet). We shared V1 updates, disaggregated serving and MLLM speedups with speakers from Embedded LLM, AMD, WekaIO, and A*STAR. Please find the meetup slides [here](https://drive.google.com/drive/folders/1ncf3GyqLdqFaB6IeB834E5TZJPLAOiXZ?usp=sharing).
+- [2025/08] We hosted [vLLM Shanghai Meetup](https://mp.weixin.qq.com/s/pDmAXHcN7Iqc8sUKgJgGtg) focusing on building, developing, and integrating with vLLM! Please find the meetup slides [here](https://drive.google.com/drive/folders/1OvLx39wnCGy_WKq8SiVKf7YcxxYI3WCH).
 - [2025/05] vLLM is now a hosted project under PyTorch Foundation! Please find the announcement [here](https://pytorch.org/blog/pytorch-foundation-welcomes-vllm/).
 - [2025/01] We are excited to announce the alpha release of vLLM V1: A major architectural upgrade with 1.7x speedup! Clean code, optimized execution loop, zero-overhead prefix caching, enhanced multimodal support, and more. Please check out our blog post [here](https://blog.vllm.ai/2025/01/27/v1-alpha-release.html).

 <details>
 <summary>Previous News</summary>

+- [2025/08] We hosted [vLLM Korea Meetup](https://luma.com/cgcgprmh) with Red Hat and Rebellions! We shared the latest advancements in vLLM along with project spotlights from the vLLM Korea community. Please find the meetup slides [here](https://drive.google.com/file/d/1bcrrAE1rxUgx0mjIeOWT6hNe2RefC5Hm/view).
+- [2025/08] We hosted [vLLM Beijing Meetup](https://mp.weixin.qq.com/s/dgkWg1WFpWGO2jCdTqQHxA) focusing on large-scale LLM deployment! Please find the meetup slides [here](https://drive.google.com/drive/folders/1Pid6NSFLU43DZRi0EaTcPgXsAzDvbBqF) and the recording [here](https://www.chaspark.com/#/live/1166916873711665152).
+- [2025/05] We hosted [NYC vLLM Meetup](https://lu.ma/c1rqyf1f)! Please find the meetup slides [here](https://docs.google.com/presentation/d/1_q_aW_ioMJWUImf1s1YM-ZhjXz8cUeL0IJvaquOYBeA/edit?usp=sharing).
 - [2025/04] We hosted [Asia Developer Day](https://www.sginnovate.com/event/limited-availability-morning-evening-slots-remaining-inaugural-vllm-asia-developer-day)! Please find the meetup slides from the vLLM team [here](https://docs.google.com/presentation/d/19cp6Qu8u48ihB91A064XfaXruNYiBOUKrBxAmDOllOo/edit?usp=sharing).
 - [2025/03] We hosted [vLLM x Ollama Inference Night](https://lu.ma/vllm-ollama)! Please find the meetup slides from the vLLM team [here](https://docs.google.com/presentation/d/16T2PDD1YwRnZ4Tu8Q5r6n53c5Lr5c73UV9Vd2_eBo4U/edit?usp=sharing).
 - [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).
@ -74,7 +81,7 @@ vLLM is flexible and easy to use with:
 - Tensor, pipeline, data and expert parallelism support for distributed inference
 - Streaming outputs
 - OpenAI-compatible API server
- Support NVIDIA GPUs, AMD CPUs and GPUs, Intel CPUs and GPUs, PowerPC CPUs, TPU, and AWS Neuron
+- Support for NVIDIA GPUs, AMD CPUs and GPUs, Intel CPUs and GPUs, PowerPC CPUs, and TPU. Additionally, support for diverse hardware plugins such as Intel Gaudi, IBM Spyre and Huawei Ascend.
 - Prefix caching support
 - Multi-LoRA support

--- a/SECURITY.md
+++ b/SECURITY.md
@ -42,4 +42,9 @@ For certain security issues of CRITICAL, HIGH, or MODERATE severity level, we ma

 * If you wish to be added to the prenotification group, please send an email copying all the members of the [vulnerability management team](https://docs.vllm.ai/en/latest/contributing/vulnerability_management.html). Each vendor contact will be analyzed on a case-by-case basis.

+* Organizations and vendors who either ship or use vLLM, are eligible to join the prenotification group if they meet at least one of the following qualifications
+    * Substantial internal deployment leveraging the upstream vLLM project.
+    * Established internal security teams and comprehensive compliance measures.
+    * Active and consistent contributions to the upstream vLLM project.
+
 * We may withdraw organizations from receiving future prenotifications if they release fixes or any other information about issues before they are public. Group membership may also change based on policy refinements for who may be included.
--- a/benchmarks/README.md
+++ b/benchmarks/README.md
@ -32,6 +32,14 @@ become available.
        <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>
@ -51,6 +59,12 @@ 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>
@ -81,6 +95,24 @@ become available.
      <td style="text-align: center;">✅</td>
      <td><code>lmms-lab/LLaVA-OneVision-Data</code>, <code>Aeala/ShareGPT_Vicuna_unfiltered</code></td>
    </tr>
+    <tr>
+      <td><strong>HuggingFace-MTBench</strong></td>
+      <td style="text-align: center;">✅</td>
+      <td style="text-align: center;">✅</td>
+      <td><code>philschmid/mt-bench</code></td>
+    </tr>
+    <tr>
+      <td><strong>HuggingFace-Blazedit</strong></td>
+      <td style="text-align: center;">✅</td>
+      <td style="text-align: center;">✅</td>
+      <td><code>vdaita/edit_5k_char</code>, <code>vdaita/edit_10k_char</code></td>
+    </tr>
+    <tr>
+      <td><strong>Spec Bench</strong></td>
+      <td style="text-align: center;">✅</td>
+      <td style="text-align: center;">✅</td>
+      <td><code>wget https://raw.githubusercontent.com/hemingkx/Spec-Bench/refs/heads/main/data/spec_bench/question.jsonl</code></td>
+    </tr>
    <tr>
      <td><strong>Custom</strong></td>
      <td style="text-align: center;">✅</td>
@ -96,7 +128,12 @@ become available.

 🚧: to be supported

-**Note**: HuggingFace dataset's `dataset-name` should be set to `hf`
+**Note**: HuggingFace dataset's `dataset-name` should be set to `hf`.
+For local `dataset-path`, please set `hf-name` to its Hugging Face ID like
+
+```bash
+--dataset-path /datasets/VisionArena-Chat/ --hf-name lmarena-ai/VisionArena-Chat
+```

 ## 🚀 Example - Online Benchmark

@ -194,6 +231,7 @@ 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 \
@ -219,6 +257,43 @@ vllm bench serve \
    --num-prompts 2048
 ```

+### Spec Bench Benchmark with Speculative Decoding
+
+``` bash
+VLLM_USE_V1=1 vllm serve meta-llama/Meta-Llama-3-8B-Instruct \
+    --speculative-config $'{"method": "ngram",
+    "num_speculative_tokens": 5, "prompt_lookup_max": 5,
+    "prompt_lookup_min": 2}'
+```
+
+[SpecBench dataset](https://github.com/hemingkx/Spec-Bench)
+
+Run all categories:
+
+``` bash
+# Download the dataset using:
+# wget https://raw.githubusercontent.com/hemingkx/Spec-Bench/refs/heads/main/data/spec_bench/question.jsonl
+
+vllm bench serve \
+    --model meta-llama/Meta-Llama-3-8B-Instruct \
+    --dataset-name spec_bench \ 
+    --dataset-path "<YOUR_DOWNLOADED_PATH>/data/spec_bench/question.jsonl" \
+    --num-prompts -1
+```
+
+Available categories include `[writing, roleplay, reasoning, math, coding, extraction, stem, humanities, translation, summarization, qa, math_reasoning, rag]`.
+
+Run only a specific category like "summarization":
+
+``` bash
+vllm bench serve \
+    --model meta-llama/Meta-Llama-3-8B-Instruct \
+    --dataset-name spec_bench \ 
+    --dataset-path "<YOUR_DOWNLOADED_PATH>/data/spec_bench/question.jsonl" \
+    --num-prompts -1
+    --spec-bench-category "summarization"
+```
+
 ### Other HuggingFaceDataset Examples

 ```bash
@ -230,6 +305,7 @@ 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 \
@ -244,6 +320,7 @@ 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 \
@ -273,6 +350,18 @@ vllm bench serve \
    --num-prompts 80
 ```

+`vdaita/edit_5k_char` or `vdaita/edit_10k_char`:
+
+``` bash
+vllm bench serve \
+    --model Qwen/QwQ-32B \
+    --dataset-name hf \
+    --dataset-path vdaita/edit_5k_char \
+    --num-prompts 90 \
+    --blazedit-min-distance 0.01 \
+    --blazedit-max-distance 0.99
+```
+
 ### Running With Sampling Parameters

 When using OpenAI-compatible backends such as `vllm`, optional sampling
@ -609,7 +698,7 @@ vllm bench serve \
  --prefix-repetition-prefix-len 512 \
  --prefix-repetition-suffix-len 128 \
  --prefix-repetition-num-prefixes 5 \
-  --prefix-repetition-output-len 128 
+  --prefix-repetition-output-len 128
 ```

 </details>
@ -672,7 +761,7 @@ python -m vllm.entrypoints.openai.api_server \
 Send requests with images:

 ```bash
-python benchmarks/benchmark_serving.py \
+vllm bench serve \
  --backend openai-chat \
  --model Qwen/Qwen2.5-VL-7B-Instruct \
  --dataset-name sharegpt \
@ -684,4 +773,102 @@ python benchmarks/benchmark_serving.py \
  --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
+vllm bench serve \
+  --backend openai-chat \
+  --model Qwen/Qwen2.5-VL-7B-Instruct \
+  --dataset-name sharegpt \
+  --dataset-path /path/to/ShareGPT4Video/llava_v1_5_mix665k_with_video_chatgpt72k_share4video28k.json \
+  --num-prompts 100 \
+  --save-result \
+  --result-dir ~/vllm_benchmark_results \
+  --save-detailed \
+  --endpoint /v1/chat/completion
+```
+
+### Synthetic Random Images (random-mm)
+
+Generate synthetic image inputs alongside random text prompts to stress-test vision models without external datasets.
+
+Notes:
+
+- Works only with online benchmark via the OpenAI  backend (`--backend openai-chat`) and endpoint `/v1/chat/completions`.
+- Video sampling is not yet implemented.
+
+Start the server (example):
+
+```bash
+vllm serve Qwen/Qwen2.5-VL-3B-Instruct \
+  --dtype bfloat16 \
+  --max-model-len 16384 \
+  --limit-mm-per-prompt '{"image": 3, "video": 0}' \
+  --mm-processor-kwargs max_pixels=1003520
+```
+
+Benchmark. It is recommended to use the flag `--ignore-eos` to simulate real responses. You can set the size of the output via the arg `random-output-len`.
+
+Ex.1: Fixed number of items and a single image resolution, enforcing generation of approx 40 tokens:
+
+```bash
+vllm bench serve \
+  --backend openai-chat \
+  --model Qwen/Qwen2.5-VL-3B-Instruct \
+  --endpoint /v1/chat/completions \
+  --dataset-name random-mm \
+  --num-prompts 100 \
+  --max-concurrency 10 \
+  --random-prefix-len 25 \
+  --random-input-len 300 \
+  --random-output-len 40 \
+  --random-range-ratio 0.2 \
+  --random-mm-base-items-per-request 2 \
+  --random-mm-limit-mm-per-prompt '{"image": 3, "video": 0}' \
+  --random-mm-bucket-config '{(224, 224, 1): 1.0}' \
+  --request-rate inf \
+  --ignore-eos \
+  --seed 42
+```
+
+The number of items per request can be controlled by passing multiple image buckets:
+
+```bash
+  --random-mm-base-items-per-request 2 \
+  --random-mm-num-mm-items-range-ratio 0.5 \
+  --random-mm-limit-mm-per-prompt '{"image": 4, "video": 0}' \
+  --random-mm-bucket-config '{(256, 256, 1): 0.7, (720, 1280, 1): 0.3}' \
+```
+
+Flags specific to `random-mm`:
+
+- `--random-mm-base-items-per-request`: base number of multimodal items per request.
+- `--random-mm-num-mm-items-range-ratio`: vary item count uniformly in the closed integer range [floor(n·(1−r)), ceil(n·(1+r))]. Set r=0 to keep it fixed; r=1 allows 0 items.
+- `--random-mm-limit-mm-per-prompt`: per-modality hard caps, e.g. '{"image": 3, "video": 0}'.
+- `--random-mm-bucket-config`: dict mapping (H, W, T) → probability. Entries with probability 0 are removed; remaining probabilities are renormalized to sum to 1. Use T=1 for images. Set any T>1 for videos (video sampling not yet supported).
+
+Behavioral notes:
+
+- If the requested base item count cannot be satisfied under the provided per-prompt limits, the tool raises an error rather than silently clamping.
+
+How sampling works:
+
+- Determine per-request item count k by sampling uniformly from the integer range defined by `--random-mm-base-items-per-request` and `--random-mm-num-mm-items-range-ratio`, then clamp k to at most the sum of per-modality limits.
+- For each of the k items, sample a bucket (H, W, T) according to the normalized probabilities in `--random-mm-bucket-config`, while tracking how many items of each modality have been added.
+- If a modality (e.g., image) reaches its limit from `--random-mm-limit-mm-per-prompt`, all buckets of that modality are excluded and the remaining bucket probabilities are renormalized before continuing.
+This should be seen as an edge case, and if this behavior can be avoided by setting `--random-mm-limit-mm-per-prompt` to a large number. Note that this might result in errors due to engine config `--limit-mm-per-prompt`.
+- The resulting request contains synthetic image data in `multi_modal_data` (OpenAI Chat format). When `random-mm` is used with the OpenAI Chat backend, prompts remain text and MM content is attached via `multi_modal_data`.
+
 </details>
--- a/benchmarks/auto_tune/README.md
+++ b/benchmarks/auto_tune/README.md
@ -31,6 +31,12 @@ cd vllm

 You must set the following variables at the top of the script before execution.

+   Note: You can also override the default values below via environment variables when running the script.
+
+```bash
+MODEL=meta-llama/Llama-3.3-70B-Instruct SYSTEM=TPU TP=8 DOWNLOAD_DIR='' INPUT_LEN=128 OUTPUT_LEN=2048 MAX_MODEL_LEN=2300 MIN_CACHE_HIT_PCT=0 MAX_LATENCY_ALLOWED_MS=100000000000 NUM_SEQS_LIST="128 256" NUM_BATCHED_TOKENS_LIST="1024 2048 4096" VLLM_LOGGING_LEVEL=DEBUG bash auto_tune.sh
+```
+
 | Variable | Description | Example Value |
 | --- | --- | --- |
 | `BASE` | **Required.** The absolute path to the parent directory of your vLLM repository directory. | `"$HOME"` |
--- a/benchmarks/auto_tune/auto_tune.sh
+++ b/benchmarks/auto_tune/auto_tune.sh
@ -5,25 +5,41 @@

 TAG=$(date +"%Y_%m_%d_%H_%M")
 SCRIPT_DIR=$( cd -- "$( dirname -- "${BASH_SOURCE[0]}" )" &> /dev/null && pwd )
-BASE="$SCRIPT_DIR/../../.."
-MODEL="meta-llama/Llama-3.1-8B-Instruct"
-SYSTEM="TPU"
-TP=1
-DOWNLOAD_DIR=""
-INPUT_LEN=4000
-OUTPUT_LEN=16
-MAX_MODEL_LEN=4096
-MIN_CACHE_HIT_PCT=0
-MAX_LATENCY_ALLOWED_MS=100000000000
-NUM_SEQS_LIST="128 256"
-NUM_BATCHED_TOKENS_LIST="512 1024 2048 4096"
+VLLM_LOGGING_LEVEL=${VLLM_LOGGING_LEVEL:-INFO}
+BASE=${BASE:-"$SCRIPT_DIR/../../.."}
+MODEL=${MODEL:-"meta-llama/Llama-3.1-8B-Instruct"}
+SYSTEM=${SYSTEM:-"TPU"}
+TP=${TP:-1}
+DOWNLOAD_DIR=${DOWNLOAD_DIR:-""}
+INPUT_LEN=${INPUT_LEN:-4000}
+OUTPUT_LEN=${OUTPUT_LEN:-16}
+MAX_MODEL_LEN=${MAX_MODEL_LEN:-4096}
+MIN_CACHE_HIT_PCT=${MIN_CACHE_HIT_PCT:-0}
+MAX_LATENCY_ALLOWED_MS=${MAX_LATENCY_ALLOWED_MS:-100000000000}
+NUM_SEQS_LIST=${NUM_SEQS_LIST:-"128 256"}
+NUM_BATCHED_TOKENS_LIST=${NUM_BATCHED_TOKENS_LIST:-"512 1024 2048 4096"}

 LOG_FOLDER="$BASE/auto-benchmark/$TAG"
 RESULT="$LOG_FOLDER/result.txt"
 PROFILE_PATH="$LOG_FOLDER/profile"

-echo "result file: $RESULT"
-echo "model: $MODEL"
+echo "====================== AUTO TUNE PARAMETERS ===================="
+echo "SCRIPT_DIR=$SCRIPT_DIR"
+echo "BASE=$BASE"
+echo "MODEL=$MODEL"
+echo "SYSTEM=$SYSTEM"
+echo "TP=$TP"
+echo "DOWNLOAD_DIR=$DOWNLOAD_DIR"
+echo "INPUT_LEN=$INPUT_LEN"
+echo "OUTPUT_LEN=$OUTPUT_LEN"
+echo "MAX_MODEL_LEN=$MAX_MODEL_LEN"
+echo "MIN_CACHE_HIT_PCT=$MIN_CACHE_HIT_PCT"
+echo "MAX_LATENCY_ALLOWED_MS=$MAX_LATENCY_ALLOWED_MS"
+echo "NUM_SEQS_LIST=$NUM_SEQS_LIST"
+echo "NUM_BATCHED_TOKENS_LIST=$NUM_BATCHED_TOKENS_LIST"
+echo "VLLM_LOGGING_LEVEL=$VLLM_LOGGING_LEVEL"
+echo "RESULT_FILE=$RESULT"
+echo "====================== AUTO TUNEPARAMETERS ===================="

 rm -rf $LOG_FOLDER
 rm -rf $PROFILE_PATH
@ -213,7 +229,7 @@ run_benchmark() {

    pkill -if vllm
    sleep 10
-    printf '=%.0s' $(seq 1 20)
+    echo "===================="
    return 0
 }

--- 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_block_pool.py
+++ b/benchmarks/benchmark_block_pool.py
@ -57,7 +57,7 @@ def invoke_main() -> None:
        "--num-iteration",
        type=int,
        default=1000,
-        help="Number of iterations to run to stablize final data readings",
+        help="Number of iterations to run to stabilize final data readings",
    )
    parser.add_argument(
        "--allocate-blocks",
--- 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
@ -351,7 +403,7 @@ class RandomDataset(BenchmarkDataset):
            # [6880, 6881] -> ['Ġcalls', 'here'] ->
            # [1650, 939, 486] -> ['Ġcall', 'sh', 'ere']
            # To avoid uncontrolled change of the prompt length,
-            # the encoded sequence is truncated before being decode again.
+            # the encoded sequence is truncated before being decoded again.
            total_input_len = prefix_len + int(input_lens[i])
            re_encoded_sequence = tokenizer.encode(prompt, add_special_tokens=False)[
                :total_input_len
@ -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
@ -430,9 +486,10 @@ class ShareGPTDataset(BenchmarkDataset):
                skip_min_output_len_check=output_len is not None,
            ):
                continue
-            # TODO: Also support ShareGPT4Video.
            if image_path := entry.get("image"):
                mm_content = process_image(image_path)
+            elif video_path := entry.get("video"):
+                mm_content = process_video(video_path)
            else:
                mm_content = None
            if enable_multimodal_chat:
@ -444,9 +501,11 @@ class ShareGPTDataset(BenchmarkDataset):
                    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


@ -512,10 +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"]
@ -534,9 +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

@ -578,6 +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.
@ -603,6 +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
@ -613,14 +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


@ -672,6 +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 = []
@ -693,6 +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
@ -752,12 +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:
@ -785,9 +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


@ -814,11 +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)
@ -838,9 +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


@ -870,15 +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(
@ -892,9 +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


@ -924,12 +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]
@ -947,9 +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


@ -974,10 +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:
@ -1000,9 +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


@ -1072,12 +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(
@ -1086,11 +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


@ -1139,6 +1242,7 @@ class ASRDataset(HuggingFaceDataset):
        tokenizer: PreTrainedTokenizerBase,
        num_requests: int,
        output_len: Optional[int] = None,
+        request_id_prefix: str = "",
        **kwargs,
    ) -> list:
        import librosa
@ -1148,6 +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
@ -1166,8 +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"
@ -1175,5 +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_latency.py
+++ b/benchmarks/benchmark_latency.py
@ -1,191 +1,17 @@
 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-"""Benchmark the latency of processing a single batch of requests."""
-
-import argparse
-import dataclasses
-import json
-import os
-import time
-from typing import Any, Optional
-
-import numpy as np
-from tqdm import tqdm
-from typing_extensions import deprecated
-
-import vllm.envs as envs
-from benchmark_utils import convert_to_pytorch_benchmark_format, write_to_json
-from vllm import LLM, SamplingParams
-from vllm.engine.arg_utils import EngineArgs
-from vllm.inputs import PromptType
-from vllm.sampling_params import BeamSearchParams
-from vllm.utils import FlexibleArgumentParser
-
-
-def save_to_pytorch_benchmark_format(
-    args: argparse.Namespace, results: dict[str, Any]
-) -> None:
-    pt_records = convert_to_pytorch_benchmark_format(
-        args=args,
-        metrics={"latency": results["latencies"]},
-        extra_info={k: results[k] for k in ["avg_latency", "percentiles"]},
-    )
-    if pt_records:
-        pt_file = f"{os.path.splitext(args.output_json)[0]}.pytorch.json"
-        write_to_json(pt_file, pt_records)
-
-
-@deprecated(
-    "benchmark_latency.py is deprecated and will be removed in a "
-    "future version. Please use 'vllm bench latency' instead.",
-)
-def main(args: argparse.Namespace):
-    print(args)
-
-    engine_args = EngineArgs.from_cli_args(args)
-
-    # NOTE(woosuk): If the request cannot be processed in a single batch,
-    # the engine will automatically process the request in multiple batches.
-    llm = LLM(**dataclasses.asdict(engine_args))
-    assert llm.llm_engine.model_config.max_model_len >= (
-        args.input_len + args.output_len
-    ), (
-        "Please ensure that max_model_len is greater than"
-        " the sum of input_len and output_len."
-    )
-
-    sampling_params = SamplingParams(
-        n=args.n,
-        temperature=1.0,
-        top_p=1.0,
-        ignore_eos=True,
-        max_tokens=args.output_len,
-        detokenize=not args.disable_detokenize,
-    )
-    print(sampling_params)
-    dummy_prompt_token_ids = np.random.randint(
-        10000, size=(args.batch_size, args.input_len)
-    )
-    dummy_prompts: list[PromptType] = [
-        {"prompt_token_ids": batch} for batch in dummy_prompt_token_ids.tolist()
-    ]
-
-    def llm_generate():
-        if not args.use_beam_search:
-            llm.generate(dummy_prompts, sampling_params=sampling_params, use_tqdm=False)
-        else:
-            llm.beam_search(
-                dummy_prompts,
-                BeamSearchParams(
-                    beam_width=args.n,
-                    max_tokens=args.output_len,
-                    ignore_eos=True,
-                ),
-            )
-
-    def run_to_completion(profile_dir: Optional[str] = None):
-        if profile_dir:
-            llm.start_profile()
-            llm_generate()
-            llm.stop_profile()
-        else:
-            start_time = time.perf_counter()
-            llm_generate()
-            end_time = time.perf_counter()
-            latency = end_time - start_time
-            return latency
-
-    print("Warming up...")
-    for _ in tqdm(range(args.num_iters_warmup), desc="Warmup iterations"):
-        run_to_completion(profile_dir=None)
-
-    if args.profile:
-        profile_dir = envs.VLLM_TORCH_PROFILER_DIR
-        print(f"Profiling (results will be saved to '{profile_dir}')...")
-        run_to_completion(profile_dir=profile_dir)
-        return
-
-    # Benchmark.
-    latencies = []
-    for _ in tqdm(range(args.num_iters), desc="Profiling iterations"):
-        latencies.append(run_to_completion(profile_dir=None))
-    latencies = np.array(latencies)
-    percentages = [10, 25, 50, 75, 90, 99]
-    percentiles = np.percentile(latencies, percentages)
-    print(f"Avg latency: {np.mean(latencies)} seconds")
-    for percentage, percentile in zip(percentages, percentiles):
-        print(f"{percentage}% percentile latency: {percentile} seconds")
-
-    # Output JSON results if specified
-    if args.output_json:
-        results = {
-            "avg_latency": np.mean(latencies),
-            "latencies": latencies.tolist(),
-            "percentiles": dict(zip(percentages, percentiles.tolist())),
-        }
-        with open(args.output_json, "w") as f:
-            json.dump(results, f, indent=4)
-        save_to_pytorch_benchmark_format(args, results)
-
-
-def create_argument_parser():
-    parser = FlexibleArgumentParser(
-        description="Benchmark the latency of processing a single batch of "
-        "requests till completion."
-    )
-    parser.add_argument("--input-len", type=int, default=32)
-    parser.add_argument("--output-len", type=int, default=128)
-    parser.add_argument("--batch-size", type=int, default=8)
-    parser.add_argument(
-        "--n",
-        type=int,
-        default=1,
-        help="Number of generated sequences per prompt.",
-    )
-    parser.add_argument("--use-beam-search", action="store_true")
-    parser.add_argument(
-        "--num-iters-warmup",
-        type=int,
-        default=10,
-        help="Number of iterations to run for warmup.",
-    )
-    parser.add_argument(
-        "--num-iters", type=int, default=30, help="Number of iterations to run."
-    )
-    parser.add_argument(
-        "--profile",
-        action="store_true",
-        help="profile the generation process of a single batch",
-    )
-    parser.add_argument(
-        "--output-json",
-        type=str,
-        default=None,
-        help="Path to save the latency results in JSON format.",
-    )
-    parser.add_argument(
-        "--disable-detokenize",
-        action="store_true",
-        help=(
-            "Do not detokenize responses (i.e. do not include "
-            "detokenization time in the latency measurement)"
-        ),
-    )
-
-    parser = EngineArgs.add_cli_args(parser)
-    # V1 enables prefix caching by default which skews the latency
-    # numbers. We need to disable prefix caching by default.
-    parser.set_defaults(enable_prefix_caching=False)
-
-    return parser
-
+import sys

 if __name__ == "__main__":
-    parser = create_argument_parser()
-    args = parser.parse_args()
-    if args.profile and not envs.VLLM_TORCH_PROFILER_DIR:
-        raise OSError(
-            "The environment variable 'VLLM_TORCH_PROFILER_DIR' is not set. "
-            "Please set it to a valid path to use torch profiler."
-        )
-    main(args)
+    print("""DEPRECATED: This script has been moved to the vLLM CLI.
+
+Please use the following command instead:
+    vllm bench latency
+
+For help with the new command, run:
+    vllm bench latency --help
+
+Alternatively, you can run the new command directly with:
+    python -m vllm.entrypoints.cli.main bench latency --help
+""")
+    sys.exit(1)
--- a/benchmarks/benchmark_ngram_proposer.py
+++ b/benchmarks/benchmark_ngram_proposer.py
@ -77,7 +77,7 @@ def invoke_main() -> None:
        "--num-iteration",
        type=int,
        default=100,
-        help="Number of iterations to run to stablize final data readings",
+        help="Number of iterations to run to stabilize final data readings",
    )
    parser.add_argument(
        "--num-req", type=int, default=128, help="Number of requests in the batch"
--- a/benchmarks/benchmark_serving.py
+++ b/benchmarks/benchmark_serving.py
--- a/benchmarks/benchmark_serving_structured_output.py
+++ b/benchmarks/benchmark_serving_structured_output.py
@ -998,7 +998,7 @@ def create_argument_parser():
        "--percentile-metrics",
        type=str,
        default="ttft,tpot,itl",
-        help="Comma-separated list of selected metrics to report percentils. "
+        help="Comma-separated list of selected metrics to report percentiles. "
        "This argument specifies the metrics to report percentiles. "
        'Allowed metric names are "ttft", "tpot", "itl", "e2el". '
        'Default value is "ttft,tpot,itl".',
--- a/benchmarks/benchmark_throughput.py
+++ b/benchmarks/benchmark_throughput.py
@ -1,742 +1,17 @@
 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-"""Benchmark offline inference throughput."""
-
-import argparse
-import dataclasses
-import json
-import os
-import random
-import time
-import warnings
-from typing import Any, Optional, Union
-
-import torch
-import uvloop
-from tqdm import tqdm
-from transformers import AutoModelForCausalLM, AutoTokenizer, PreTrainedTokenizerBase
-from typing_extensions import deprecated
-
-from benchmark_dataset import (
-    AIMODataset,
-    BurstGPTDataset,
-    ConversationDataset,
-    InstructCoderDataset,
-    RandomDataset,
-    SampleRequest,
-    ShareGPTDataset,
-    SonnetDataset,
-    VisionArenaDataset,
-)
-from benchmark_utils import convert_to_pytorch_benchmark_format, write_to_json
-from vllm.engine.arg_utils import AsyncEngineArgs, EngineArgs
-from vllm.entrypoints.openai.api_server import (
-    build_async_engine_client_from_engine_args,
-)
-from vllm.inputs import TextPrompt, TokensPrompt
-from vllm.lora.request import LoRARequest
-from vllm.outputs import RequestOutput
-from vllm.sampling_params import BeamSearchParams
-from vllm.utils import FlexibleArgumentParser, merge_async_iterators
-
-
-def run_vllm(
-    requests: list[SampleRequest],
-    n: int,
-    engine_args: EngineArgs,
-    disable_detokenize: bool = False,
-) -> tuple[float, Optional[list[RequestOutput]]]:
-    from vllm import LLM, SamplingParams
-
-    llm = LLM(**dataclasses.asdict(engine_args))
-    assert all(
-        llm.llm_engine.model_config.max_model_len
-        >= (request.prompt_len + request.expected_output_len)
-        for request in requests
-    ), (
-        "Please ensure that max_model_len is greater than the sum of"
-        " prompt_len and expected_output_len for all requests."
-    )
-    # Add the requests to the engine.
-    prompts: list[Union[TextPrompt, TokensPrompt]] = []
-    sampling_params: list[SamplingParams] = []
-    for request in requests:
-        prompts.append(
-            TokensPrompt(
-                prompt_token_ids=request.prompt["prompt_token_ids"],
-                multi_modal_data=request.multi_modal_data,
-            )
-            if "prompt_token_ids" in request.prompt
-            else TextPrompt(
-                prompt=request.prompt, multi_modal_data=request.multi_modal_data
-            )
-        )
-        sampling_params.append(
-            SamplingParams(
-                n=n,
-                temperature=1.0,
-                top_p=1.0,
-                ignore_eos=True,
-                max_tokens=request.expected_output_len,
-                detokenize=not disable_detokenize,
-            )
-        )
-    lora_requests: Optional[list[LoRARequest]] = None
-    if engine_args.enable_lora:
-        lora_requests = [request.lora_request for request in requests]
-
-    use_beam_search = False
-
-    outputs = None
-    if not use_beam_search:
-        start = time.perf_counter()
-        outputs = llm.generate(
-            prompts, sampling_params, lora_request=lora_requests, use_tqdm=True
-        )
-        end = time.perf_counter()
-    else:
-        assert lora_requests is None, "BeamSearch API does not support LoRA"
-        prompts = [request.prompt for request in requests]
-        # output_len should be the same for all requests.
-        output_len = requests[0].expected_output_len
-        for request in requests:
-            assert request.expected_output_len == output_len
-        start = time.perf_counter()
-        llm.beam_search(
-            prompts,
-            BeamSearchParams(
-                beam_width=n,
-                max_tokens=output_len,
-                ignore_eos=True,
-            ),
-        )
-        end = time.perf_counter()
-    return end - start, outputs
-
-
-def run_vllm_chat(
-    requests: list[SampleRequest],
-    n: int,
-    engine_args: EngineArgs,
-    disable_detokenize: bool = False,
-) -> tuple[float, list[RequestOutput]]:
-    """
-    Run vLLM chat benchmark. This function is recommended ONLY for benchmarking
-    multimodal models as it properly handles multimodal inputs and chat
-    formatting. For non-multimodal models, use run_vllm() instead.
-    """
-    from vllm import LLM, SamplingParams
-
-    llm = LLM(**dataclasses.asdict(engine_args))
-
-    assert all(
-        llm.llm_engine.model_config.max_model_len
-        >= (request.prompt_len + request.expected_output_len)
-        for request in requests
-    ), (
-        "Please ensure that max_model_len is greater than the sum of "
-        "prompt_len and expected_output_len for all requests."
-    )
-
-    prompts = []
-    sampling_params: list[SamplingParams] = []
-    for request in requests:
-        prompts.append(request.prompt)
-        sampling_params.append(
-            SamplingParams(
-                n=n,
-                temperature=1.0,
-                top_p=1.0,
-                ignore_eos=True,
-                max_tokens=request.expected_output_len,
-                detokenize=not disable_detokenize,
-            )
-        )
-    start = time.perf_counter()
-    outputs = llm.chat(prompts, sampling_params, use_tqdm=True)
-    end = time.perf_counter()
-    return end - start, outputs
-
-
-async def run_vllm_async(
-    requests: list[SampleRequest],
-    n: int,
-    engine_args: AsyncEngineArgs,
-    disable_frontend_multiprocessing: bool = False,
-    disable_detokenize: bool = False,
-) -> float:
-    from vllm import SamplingParams
-
-    async with build_async_engine_client_from_engine_args(
-        engine_args,
-        disable_frontend_multiprocessing=disable_frontend_multiprocessing,
-    ) as llm:
-        model_config = await llm.get_model_config()
-        assert all(
-            model_config.max_model_len
-            >= (request.prompt_len + request.expected_output_len)
-            for request in requests
-        ), (
-            "Please ensure that max_model_len is greater than the sum of"
-            " prompt_len and expected_output_len for all requests."
-        )
-
-        # Add the requests to the engine.
-        prompts: list[Union[TextPrompt, TokensPrompt]] = []
-        sampling_params: list[SamplingParams] = []
-        lora_requests: list[Optional[LoRARequest]] = []
-        for request in requests:
-            prompts.append(
-                TokensPrompt(
-                    prompt_token_ids=request.prompt["prompt_token_ids"],
-                    multi_modal_data=request.multi_modal_data,
-                )
-                if "prompt_token_ids" in request.prompt
-                else TextPrompt(
-                    prompt=request.prompt, multi_modal_data=request.multi_modal_data
-                )
-            )
-            sampling_params.append(
-                SamplingParams(
-                    n=n,
-                    temperature=1.0,
-                    top_p=1.0,
-                    ignore_eos=True,
-                    max_tokens=request.expected_output_len,
-                    detokenize=not disable_detokenize,
-                )
-            )
-            lora_requests.append(request.lora_request)
-
-        generators = []
-        start = time.perf_counter()
-        for i, (prompt, sp, lr) in enumerate(
-            zip(prompts, sampling_params, lora_requests)
-        ):
-            generator = llm.generate(prompt, sp, lora_request=lr, request_id=f"test{i}")
-            generators.append(generator)
-        all_gens = merge_async_iterators(*generators)
-        async for i, res in all_gens:
-            pass
-        end = time.perf_counter()
-        return end - start
-
-
-def run_hf(
-    requests: list[SampleRequest],
-    model: str,
-    tokenizer: PreTrainedTokenizerBase,
-    n: int,
-    max_batch_size: int,
-    trust_remote_code: bool,
-    disable_detokenize: bool = False,
-) -> float:
-    llm = AutoModelForCausalLM.from_pretrained(
-        model, torch_dtype=torch.float16, trust_remote_code=trust_remote_code
-    )
-    if llm.config.model_type == "llama":
-        # To enable padding in the HF backend.
-        tokenizer.pad_token = tokenizer.eos_token
-    llm = llm.cuda()
-
-    pbar = tqdm(total=len(requests))
-    start = time.perf_counter()
-    batch: list[str] = []
-    max_prompt_len = 0
-    max_output_len = 0
-    for i in range(len(requests)):
-        prompt = requests[i].prompt
-        prompt_len = requests[i].prompt_len
-        output_len = requests[i].expected_output_len
-        # Add the prompt to the batch.
-        batch.append(prompt)
-        max_prompt_len = max(max_prompt_len, prompt_len)
-        max_output_len = max(max_output_len, output_len)
-        if len(batch) < max_batch_size and i != len(requests) - 1:
-            # Check if we can add more requests to the batch.
-            next_prompt_len = requests[i + 1].prompt_len
-            next_output_len = requests[i + 1].expected_output_len
-            if (
-                max(max_prompt_len, next_prompt_len)
-                + max(max_output_len, next_output_len)
-            ) <= 2048:
-                # We can add more requests to the batch.
-                continue
-
-        # Generate the sequences.
-        input_ids = tokenizer(batch, return_tensors="pt", padding=True).input_ids
-        llm_outputs = llm.generate(
-            input_ids=input_ids.cuda(),
-            do_sample=True,
-            num_return_sequences=n,
-            temperature=1.0,
-            top_p=1.0,
-            use_cache=True,
-            max_new_tokens=max_output_len,
-        )
-        if not disable_detokenize:
-            # Include the decoding time.
-            tokenizer.batch_decode(llm_outputs, skip_special_tokens=True)
-        pbar.update(len(batch))
-
-        # Clear the batch.
-        batch = []
-        max_prompt_len = 0
-        max_output_len = 0
-    end = time.perf_counter()
-    return end - start
-
-
-def run_mii(
-    requests: list[SampleRequest],
-    model: str,
-    tensor_parallel_size: int,
-    output_len: int,
-) -> float:
-    from mii import client, serve
-
-    llm = serve(model, tensor_parallel=tensor_parallel_size)
-    prompts = [request.prompt for request in requests]
-
-    start = time.perf_counter()
-    llm.generate(prompts, max_new_tokens=output_len)
-    end = time.perf_counter()
-    client = client(model)
-    client.terminate_server()
-    return end - start
-
-
-def save_to_pytorch_benchmark_format(
-    args: argparse.Namespace, results: dict[str, Any]
-) -> None:
-    pt_records = convert_to_pytorch_benchmark_format(
-        args=args,
-        metrics={
-            "requests_per_second": [results["requests_per_second"]],
-            "tokens_per_second": [results["tokens_per_second"]],
-        },
-        extra_info={
-            k: results[k] for k in ["elapsed_time", "num_requests", "total_num_tokens"]
-        },
-    )
-    if pt_records:
-        # Don't use json suffix here as we don't want CI to pick it up
-        pt_file = f"{os.path.splitext(args.output_json)[0]}.pytorch.json"
-        write_to_json(pt_file, pt_records)
-
-
-def get_requests(args, tokenizer):
-    # Common parameters for all dataset types.
-    common_kwargs = {
-        "dataset_path": args.dataset_path,
-        "random_seed": args.seed,
-    }
-    sample_kwargs = {
-        "tokenizer": tokenizer,
-        "lora_path": args.lora_path,
-        "max_loras": args.max_loras,
-        "num_requests": args.num_prompts,
-        "input_len": args.input_len,
-        "output_len": args.output_len,
-    }
-
-    if args.dataset_path is None or args.dataset_name == "random":
-        sample_kwargs["range_ratio"] = args.random_range_ratio
-        sample_kwargs["prefix_len"] = args.prefix_len
-        dataset_cls = RandomDataset
-    elif args.dataset_name == "sharegpt":
-        dataset_cls = ShareGPTDataset
-        if args.backend == "vllm-chat":
-            sample_kwargs["enable_multimodal_chat"] = True
-    elif args.dataset_name == "sonnet":
-        assert tokenizer.chat_template or tokenizer.default_chat_template, (
-            "Tokenizer/model must have chat template for sonnet dataset."
-        )
-        dataset_cls = SonnetDataset
-        sample_kwargs["prefix_len"] = args.prefix_len
-        sample_kwargs["return_prompt_formatted"] = True
-    elif args.dataset_name == "burstgpt":
-        dataset_cls = BurstGPTDataset
-    elif args.dataset_name == "hf":
-        common_kwargs["no_stream"] = args.no_stream
-        if args.dataset_path in VisionArenaDataset.SUPPORTED_DATASET_PATHS:
-            dataset_cls = VisionArenaDataset
-            common_kwargs["dataset_subset"] = None
-            common_kwargs["dataset_split"] = "train"
-            sample_kwargs["enable_multimodal_chat"] = True
-        elif args.dataset_path in InstructCoderDataset.SUPPORTED_DATASET_PATHS:
-            dataset_cls = InstructCoderDataset
-            common_kwargs["dataset_split"] = "train"
-        elif args.dataset_path in ConversationDataset.SUPPORTED_DATASET_PATHS:
-            dataset_cls = ConversationDataset
-            common_kwargs["dataset_subset"] = args.hf_subset
-            common_kwargs["dataset_split"] = args.hf_split
-            sample_kwargs["enable_multimodal_chat"] = True
-        elif args.dataset_path in AIMODataset.SUPPORTED_DATASET_PATHS:
-            dataset_cls = AIMODataset
-            common_kwargs["dataset_subset"] = None
-            common_kwargs["dataset_split"] = "train"
-    else:
-        raise ValueError(f"Unknown dataset name: {args.dataset_name}")
-    # Remove None values
-    sample_kwargs = {k: v for k, v in sample_kwargs.items() if v is not None}
-    return dataset_cls(**common_kwargs).sample(**sample_kwargs)
-
-
-@deprecated(
-    "benchmark_throughput.py is deprecated and will be removed in a "
-    "future version. Please use 'vllm bench throughput' instead.",
-)
-def main(args: argparse.Namespace):
-    if args.seed is None:
-        args.seed = 0
-    print(args)
-    random.seed(args.seed)
-    # Sample the requests.
-    tokenizer = AutoTokenizer.from_pretrained(
-        args.tokenizer, trust_remote_code=args.trust_remote_code
-    )
-    requests = get_requests(args, tokenizer)
-    is_multi_modal = any(request.multi_modal_data is not None for request in requests)
-    request_outputs: Optional[list[RequestOutput]] = None
-    if args.backend == "vllm":
-        if args.async_engine:
-            elapsed_time = uvloop.run(
-                run_vllm_async(
-                    requests,
-                    args.n,
-                    AsyncEngineArgs.from_cli_args(args),
-                    args.disable_frontend_multiprocessing,
-                    args.disable_detokenize,
-                )
-            )
-        else:
-            elapsed_time, request_outputs = run_vllm(
-                requests,
-                args.n,
-                EngineArgs.from_cli_args(args),
-                args.disable_detokenize,
-            )
-    elif args.backend == "hf":
-        assert args.tensor_parallel_size == 1
-        elapsed_time = run_hf(
-            requests,
-            args.model,
-            tokenizer,
-            args.n,
-            args.hf_max_batch_size,
-            args.trust_remote_code,
-            args.disable_detokenize,
-        )
-    elif args.backend == "mii":
-        elapsed_time = run_mii(
-            requests, args.model, args.tensor_parallel_size, args.output_len
-        )
-    elif args.backend == "vllm-chat":
-        elapsed_time, request_outputs = run_vllm_chat(
-            requests, args.n, EngineArgs.from_cli_args(args), args.disable_detokenize
-        )
-    else:
-        raise ValueError(f"Unknown backend: {args.backend}")
-
-    if request_outputs:
-        # Note: with the vllm and vllm-chat backends,
-        # we have request_outputs, which we use to count tokens.
-        total_prompt_tokens = 0
-        total_output_tokens = 0
-        for ro in request_outputs:
-            if not isinstance(ro, RequestOutput):
-                continue
-            total_prompt_tokens += (
-                len(ro.prompt_token_ids) if ro.prompt_token_ids else 0
-            )
-            total_output_tokens += sum(len(o.token_ids) for o in ro.outputs if o)
-        total_num_tokens = total_prompt_tokens + total_output_tokens
-    else:
-        total_num_tokens = sum(r.prompt_len + r.expected_output_len for r in requests)
-        total_output_tokens = sum(r.expected_output_len for r in requests)
-        total_prompt_tokens = total_num_tokens - total_output_tokens
-
-    if is_multi_modal and args.backend != "vllm-chat":
-        print(
-            "\033[91mWARNING\033[0m: Multi-modal request with "
-            f"{args.backend} backend detected. The "
-            "following metrics are not accurate because image tokens are not"
-            " counted. See vllm-project/vllm/issues/9778 for details."
-        )
-        # TODO(vllm-project/vllm/issues/9778): Count multi-modal token length.
-        # vllm-chat backend counts the image tokens now
-
-    print(
-        f"Throughput: {len(requests) / elapsed_time:.2f} requests/s, "
-        f"{total_num_tokens / elapsed_time:.2f} total tokens/s, "
-        f"{total_output_tokens / elapsed_time:.2f} output tokens/s"
-    )
-    print(f"Total num prompt tokens:  {total_prompt_tokens}")
-    print(f"Total num output tokens:  {total_output_tokens}")
-
-    # Output JSON results if specified
-    if args.output_json:
-        results = {
-            "elapsed_time": elapsed_time,
-            "num_requests": len(requests),
-            "total_num_tokens": total_num_tokens,
-            "requests_per_second": len(requests) / elapsed_time,
-            "tokens_per_second": total_num_tokens / elapsed_time,
-        }
-        with open(args.output_json, "w") as f:
-            json.dump(results, f, indent=4)
-        save_to_pytorch_benchmark_format(args, results)
-
-
-def validate_args(args):
-    """
-    Validate command-line arguments.
-    """
-
-    # === Deprecation and Defaulting ===
-    if args.dataset is not None:
-        warnings.warn(
-            "The '--dataset' argument will be deprecated in the next release. "
-            "Please use '--dataset-name' and '--dataset-path' instead.",
-            stacklevel=2,
-        )
-        args.dataset_path = args.dataset
-
-    if not getattr(args, "tokenizer", None):
-        args.tokenizer = args.model
-
-    # === Backend Validation ===
-    valid_backends = {"vllm", "hf", "mii", "vllm-chat"}
-    if args.backend not in valid_backends:
-        raise ValueError(f"Unsupported backend: {args.backend}")
-
-    # === Dataset Configuration ===
-    if not args.dataset and not args.dataset_path:
-        print("When dataset path is not set, it will default to random dataset")
-        args.dataset_name = "random"
-        if args.input_len is None:
-            raise ValueError("input_len must be provided for a random dataset")
-
-    # === Dataset Name Specific Checks ===
-    # --hf-subset and --hf-split: only used
-    # when dataset_name is 'hf'
-    if args.dataset_name != "hf" and (
-        getattr(args, "hf_subset", None) is not None
-        or getattr(args, "hf_split", None) is not None
-    ):
-        warnings.warn(
-            "--hf-subset and --hf-split will be ignored \
-                since --dataset-name is not 'hf'.",
-            stacklevel=2,
-        )
-    elif args.dataset_name == "hf":
-        if args.dataset_path in (
-            VisionArenaDataset.SUPPORTED_DATASET_PATHS.keys()
-            | ConversationDataset.SUPPORTED_DATASET_PATHS
-        ):
-            assert args.backend == "vllm-chat", (
-                f"{args.dataset_path} needs to use vllm-chat as the backend."
-            )  # noqa: E501
-        elif args.dataset_path in (
-            InstructCoderDataset.SUPPORTED_DATASET_PATHS
-            | AIMODataset.SUPPORTED_DATASET_PATHS
-        ):
-            assert args.backend == "vllm", (
-                f"{args.dataset_path} needs to use vllm as the backend."
-            )  # noqa: E501
-        else:
-            raise ValueError(f"{args.dataset_path} is not supported by hf dataset.")
-
-    # --random-range-ratio: only used when dataset_name is 'random'
-    if args.dataset_name != "random" and args.random_range_ratio is not None:
-        warnings.warn(
-            "--random-range-ratio will be ignored since \
-                --dataset-name is not 'random'.",
-            stacklevel=2,
-        )
-
-    # --prefix-len: only used when dataset_name is 'random', 'sonnet', or not
-    # set.
-    if (
-        args.dataset_name not in {"random", "sonnet", None}
-        and args.prefix_len is not None
-    ):
-        warnings.warn(
-            "--prefix-len will be ignored since --dataset-name\
-                 is not 'random', 'sonnet', or not set.",
-            stacklevel=2,
-        )
-
-    # === LoRA Settings ===
-    if getattr(args, "enable_lora", False) and args.backend != "vllm":
-        raise ValueError("LoRA benchmarking is only supported for vLLM backend")
-    if getattr(args, "enable_lora", False) and args.lora_path is None:
-        raise ValueError("LoRA path must be provided when enable_lora is True")
-
-    # === Backend-specific Validations ===
-    if args.backend == "hf" and args.hf_max_batch_size is None:
-        raise ValueError("HF max batch size is required for HF backend")
-    if args.backend != "hf" and args.hf_max_batch_size is not None:
-        raise ValueError("HF max batch size is only for HF backend.")
-
-    if (
-        args.backend in {"hf", "mii"}
-        and getattr(args, "quantization", None) is not None
-    ):
-        raise ValueError("Quantization is only for vLLM backend.")
-
-    if args.backend == "mii" and args.dtype != "auto":
-        raise ValueError("dtype must be auto for MII backend.")
-    if args.backend == "mii" and args.n != 1:
-        raise ValueError("n must be 1 for MII backend.")
-    if args.backend == "mii" and args.tokenizer != args.model:
-        raise ValueError("Tokenizer must be the same as the model for MII backend.")
-
-    # --data-parallel is not supported currently.
-    # https://github.com/vllm-project/vllm/issues/16222
-    if args.data_parallel_size > 1:
-        raise ValueError(
-            "Data parallel is not supported in offline benchmark, \
-            please use benchmark serving instead"
-        )
-
-
-def create_argument_parser():
-    parser = FlexibleArgumentParser(description="Benchmark the throughput.")
-    parser.add_argument(
-        "--backend",
-        type=str,
-        choices=["vllm", "hf", "mii", "vllm-chat"],
-        default="vllm",
-    )
-    parser.add_argument(
-        "--dataset-name",
-        type=str,
-        choices=["sharegpt", "random", "sonnet", "burstgpt", "hf"],
-        help="Name of the dataset to benchmark on.",
-        default="sharegpt",
-    )
-    parser.add_argument(
-        "--no-stream",
-        action="store_true",
-        help="Do not load the dataset in streaming mode.",
-    )
-    parser.add_argument(
-        "--dataset",
-        type=str,
-        default=None,
-        help="Path to the ShareGPT dataset, will be deprecated in\
-            the next release. The dataset is expected to "
-        "be a json in form of list[dict[..., conversations: "
-        "list[dict[..., value: <prompt_or_response>]]]]",
-    )
-    parser.add_argument(
-        "--dataset-path", type=str, default=None, help="Path to the dataset"
-    )
-    parser.add_argument(
-        "--input-len",
-        type=int,
-        default=None,
-        help="Input prompt length for each request",
-    )
-    parser.add_argument(
-        "--output-len",
-        type=int,
-        default=None,
-        help="Output length for each request. Overrides the "
-        "output length from the dataset.",
-    )
-    parser.add_argument(
-        "--n", type=int, default=1, help="Number of generated sequences per prompt."
-    )
-    parser.add_argument(
-        "--num-prompts", type=int, default=1000, help="Number of prompts to process."
-    )
-    parser.add_argument(
-        "--hf-max-batch-size",
-        type=int,
-        default=None,
-        help="Maximum batch size for HF backend.",
-    )
-    parser.add_argument(
-        "--output-json",
-        type=str,
-        default=None,
-        help="Path to save the throughput results in JSON format.",
-    )
-    parser.add_argument(
-        "--async-engine",
-        action="store_true",
-        default=False,
-        help="Use vLLM async engine rather than LLM class.",
-    )
-    parser.add_argument(
-        "--disable-frontend-multiprocessing",
-        action="store_true",
-        default=False,
-        help="Disable decoupled async engine frontend.",
-    )
-    parser.add_argument(
-        "--disable-detokenize",
-        action="store_true",
-        help=(
-            "Do not detokenize the response (i.e. do not include "
-            "detokenization time in the measurement)"
-        ),
-    )
-    # LoRA
-    parser.add_argument(
-        "--lora-path",
-        type=str,
-        default=None,
-        help="Path to the LoRA adapters to use. This can be an absolute path, "
-        "a relative path, or a Hugging Face model identifier.",
-    )
-    parser.add_argument(
-        "--prefix-len",
-        type=int,
-        default=None,
-        help=f"Number of prefix tokens to be used in RandomDataset "
-        "and SonnetDataset. For RandomDataset, the total input "
-        "length is the sum of prefix-len (default: "
-        f"{RandomDataset.DEFAULT_PREFIX_LEN}) and a random context length "
-        "sampled from [input_len * (1 - range_ratio), "
-        "input_len * (1 + range_ratio)]. For SonnetDataset, "
-        f"prefix_len (default: {SonnetDataset.DEFAULT_PREFIX_LEN}) "
-        "controls how much of the input is fixed lines versus "
-        "random lines, but the total input length remains approximately "
-        "input_len tokens.",
-    )
-    # random dataset
-    parser.add_argument(
-        "--random-range-ratio",
-        type=float,
-        default=None,
-        help=f"Range ratio (default : {RandomDataset.DEFAULT_RANGE_RATIO}) "
-        "for sampling input/output length, "
-        "used only for RandomDataset. Must be in the range [0, 1) to "
-        "define a symmetric sampling range "
-        "[length * (1 - range_ratio), length * (1 + range_ratio)].",
-    )
-
-    # hf dtaset
-    parser.add_argument(
-        "--hf-subset", type=str, default=None, help="Subset of the HF dataset."
-    )
-    parser.add_argument(
-        "--hf-split", type=str, default=None, help="Split of the HF dataset."
-    )
-
-    parser = AsyncEngineArgs.add_cli_args(parser)
-
-    return parser
-
+import sys

 if __name__ == "__main__":
-    parser = create_argument_parser()
-    args = parser.parse_args()
-    if args.tokenizer is None:
-        args.tokenizer = args.model
-    validate_args(args)
-    main(args)
+    print("""DEPRECATED: This script has been moved to the vLLM CLI.
+
+Please use the following command instead:
+    vllm bench throughput
+
+For help with the new command, run:
+    vllm bench throughput --help
+
+Alternatively, you can run the new command directly with:
+    python -m vllm.entrypoints.cli.main bench throughput --help
+""")
+    sys.exit(1)
--- a/benchmarks/disagg_benchmarks/disagg_overhead_benchmark.sh
+++ b/benchmarks/disagg_benchmarks/disagg_overhead_benchmark.sh
@ -62,7 +62,7 @@ benchmark() {
    --max-model-len 10000 \
    --gpu-memory-utilization 0.6 \
    --kv-transfer-config \
-    '{"kv_connector":"PyNcclConnector","kv_role":"kv_producer","kv_rank":0,"kv_parallel_size":2,"kv_buffer_size":5e9}' &
+    '{"kv_connector":"P2pNcclConnector","kv_role":"kv_producer","kv_rank":0,"kv_parallel_size":2,"kv_buffer_size":5e9}' &


  CUDA_VISIBLE_DEVICES=1 python3 \
@ -72,7 +72,7 @@ benchmark() {
    --max-model-len 10000 \
    --gpu-memory-utilization 0.6 \
    --kv-transfer-config \
-    '{"kv_connector":"PyNcclConnector","kv_role":"kv_consumer","kv_rank":1,"kv_parallel_size":2,"kv_buffer_size":5e9}' &
+    '{"kv_connector":"P2pNcclConnector","kv_role":"kv_consumer","kv_rank":1,"kv_parallel_size":2,"kv_buffer_size":5e9}' &

  wait_for_server 8100
  wait_for_server 8200
--- a/benchmarks/disagg_benchmarks/disagg_performance_benchmark.sh
+++ b/benchmarks/disagg_benchmarks/disagg_performance_benchmark.sh
@ -69,7 +69,7 @@ launch_disagg_prefill() {
    --max-model-len 10000 \
    --gpu-memory-utilization 0.6 \
    --kv-transfer-config \
-    '{"kv_connector":"PyNcclConnector","kv_role":"kv_producer","kv_rank":0,"kv_parallel_size":2,"kv_buffer_size":5e9}' &
+    '{"kv_connector":"P2pNcclConnector","kv_role":"kv_producer","kv_rank":0,"kv_parallel_size":2,"kv_buffer_size":5e9}' &

  CUDA_VISIBLE_DEVICES=1 python3 \
    -m vllm.entrypoints.openai.api_server \
@ -78,7 +78,7 @@ launch_disagg_prefill() {
    --max-model-len 10000 \
    --gpu-memory-utilization 0.6 \
    --kv-transfer-config \
-    '{"kv_connector":"PyNcclConnector","kv_role":"kv_consumer","kv_rank":1,"kv_parallel_size":2,"kv_buffer_size":5e9}' &
+    '{"kv_connector":"P2pNcclConnector","kv_role":"kv_consumer","kv_rank":1,"kv_parallel_size":2,"kv_buffer_size":5e9}' &

  wait_for_server 8100
  wait_for_server 8200
--- a/benchmarks/kernels/bench_block_fp8_gemm.py
+++ b/benchmarks/kernels/bench_block_fp8_gemm.py
@ -0,0 +1,145 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+import torch
+
+from vllm.model_executor.layers.quantization.utils.fp8_utils import (
+    apply_w8a8_block_fp8_linear,
+)
+from vllm.model_executor.layers.quantization.utils.w8a8_utils import (
+    CUTLASS_BLOCK_FP8_SUPPORTED,
+)
+from vllm.platforms import current_platform
+from vllm.triton_utils import triton as vllm_triton
+
+assert current_platform.is_cuda(), (
+    "Only support benchmarking w8a8 block fp8 kernel on CUDA device."
+)
+
+# DeepSeek-V3 weight shapes
+DEEPSEEK_V3_SHAPES = [
+    (512 + 64, 7168),
+    (2112, 7168),
+    ((128 + 64) * 128, 7168),
+    (128 * (128 + 128), 512),
+    (7168, 16384),
+    (7168, 18432),
+    (18432 * 2, 7168),
+    (24576, 1536),
+    (12288, 7168),
+    (4096, 7168),
+    (7168, 2048),
+]
+
+
+def build_w8a8_block_fp8_runner(M, N, K, block_size, device, use_cutlass):
+    """Build runner function for w8a8 block fp8 matmul."""
+    factor_for_scale = 1e-2
+
+    fp8_info = torch.finfo(torch.float8_e4m3fn)
+    fp8_max, fp8_min = fp8_info.max, fp8_info.min
+
+    # Create random FP8 tensors
+    A_ref = (torch.rand(M, K, dtype=torch.bfloat16, device=device) - 0.5) * 2 * fp8_max
+
+    B_ref = (torch.rand(N, K, dtype=torch.bfloat16, device=device) - 0.5) * 2 * fp8_max
+    B = B_ref.clamp(min=fp8_min, max=fp8_max).to(torch.float8_e4m3fn)
+
+    # Create scales
+    block_n, block_k = block_size[0], block_size[1]
+    n_tiles = (N + block_n - 1) // block_n
+    k_tiles = (K + block_k - 1) // block_k
+
+    Bs = (
+        torch.rand(n_tiles, k_tiles, dtype=torch.float32, device=device)
+        * factor_for_scale
+    )
+
+    # SM90 CUTLASS requires row-major format for scales
+    if use_cutlass and current_platform.is_device_capability(90):
+        Bs = Bs.T.contiguous()
+
+    def run():
+        if use_cutlass:
+            return apply_w8a8_block_fp8_linear(
+                A_ref, B, block_size, Bs, cutlass_block_fp8_supported=True
+            )
+        else:
+            return apply_w8a8_block_fp8_linear(
+                A_ref, B, block_size, Bs, cutlass_block_fp8_supported=False
+            )
+
+    return run
+
+
+# Determine available providers
+available_providers = ["torch-bf16", "w8a8-block-fp8-triton"]
+plot_title = "BF16 vs W8A8 Block FP8 GEMMs"
+
+if CUTLASS_BLOCK_FP8_SUPPORTED:
+    available_providers.append("w8a8-block-fp8-cutlass")
+
+
+@vllm_triton.testing.perf_report(
+    vllm_triton.testing.Benchmark(
+        x_names=["batch_size"],
+        x_vals=[1, 16, 64, 128, 256, 512, 1024, 2048, 4096, 8192, 16384],
+        x_log=False,
+        line_arg="provider",
+        line_vals=available_providers,
+        line_names=available_providers,
+        ylabel="TFLOP/s (larger is better)",
+        plot_name="BF16 vs W8A8 Block FP8 GEMMs",
+        args={},
+    )
+)
+def benchmark_tflops(batch_size, provider, N, K, block_size=(128, 128)):
+    M = batch_size
+    device = "cuda"
+
+    quantiles = [0.5, 0.2, 0.8]
+
+    if provider == "torch-bf16":
+        a = torch.randn((M, K), device=device, dtype=torch.bfloat16)
+        b = torch.randn((N, K), device=device, dtype=torch.bfloat16)
+        ms, min_ms, max_ms = vllm_triton.testing.do_bench_cudagraph(
+            lambda: torch.nn.functional.linear(a, b), quantiles=quantiles
+        )
+    elif provider == "w8a8-block-fp8-triton":
+        run_w8a8_triton = build_w8a8_block_fp8_runner(
+            M, N, K, block_size, device, use_cutlass=False
+        )
+        ms, min_ms, max_ms = vllm_triton.testing.do_bench_cudagraph(
+            lambda: run_w8a8_triton(), quantiles=quantiles
+        )
+    elif provider == "w8a8-block-fp8-cutlass":
+        run_w8a8_cutlass = build_w8a8_block_fp8_runner(
+            M, N, K, block_size, device, use_cutlass=True
+        )
+        ms, min_ms, max_ms = vllm_triton.testing.do_bench_cudagraph(
+            lambda: run_w8a8_cutlass(), quantiles=quantiles
+        )
+    else:
+        raise ValueError(f"Unknown provider: {provider}")
+
+    to_tflops = lambda t_ms: (2 * M * N * K) * 1e-12 / (t_ms * 1e-3)
+    return to_tflops(ms), to_tflops(max_ms), to_tflops(min_ms)
+
+
+if __name__ == "__main__":
+    block_size = (128, 128)
+
+    for N, K in DEEPSEEK_V3_SHAPES:
+        print(f"\nBenchmarking DeepSeek-V3, N={N} K={K}")
+
+        print(f"TFLOP/s comparison (block_size={block_size}):")
+        benchmark_tflops.run(
+            print_data=True,
+            # show_plots=False,
+            # save_path=f"bench_w8a8_block_fp8_tflops_n{N}_k{K}",
+            N=N,
+            K=K,
+            block_size=block_size,
+        )
+
+    print("\nBenchmark finished!")
--- a/benchmarks/kernels/benchmark_activation.py
+++ b/benchmarks/kernels/benchmark_activation.py
@ -0,0 +1,104 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+# benchmark custom activation op performance
+import itertools
+
+import torch
+
+import vllm.model_executor.layers.activation  # noqa F401
+from vllm.model_executor.custom_op import CustomOp
+from vllm.platforms import current_platform
+from vllm.triton_utils import triton
+from vllm.utils import STR_DTYPE_TO_TORCH_DTYPE, FlexibleArgumentParser
+
+batch_size_range = [1, 16, 32, 64, 128]
+seq_len_range = [1, 16, 64, 128, 256, 512, 1024, 2048, 4096]
+intermediate_size = [3072, 9728, 12288]
+configs = list(itertools.product(batch_size_range, seq_len_range, intermediate_size))
+
+
+def benchmark_activation(
+    batch_size: int,
+    seq_len: int,
+    intermediate_size: int,
+    provider: str,
+    func_name: str,
+    dtype: torch.dtype,
+):
+    device = "cuda"
+    num_tokens = batch_size * seq_len
+    dim = intermediate_size
+    current_platform.seed_everything(42)
+    torch.set_default_device(device)
+
+    if func_name == "gelu_and_mul":
+        layer = CustomOp.op_registry[func_name](approximate="none")
+    elif func_name == "gelu_and_mul_tanh":
+        layer = CustomOp.op_registry["gelu_and_mul"](approximate="tanh")
+    elif func_name == "fatrelu_and_mul":
+        threshold = 0.5
+        layer = CustomOp.op_registry[func_name](threshold)
+    else:
+        layer = CustomOp.op_registry[func_name]()
+
+    x = torch.randn(num_tokens, dim, dtype=dtype, device=device)
+    compiled_layer = torch.compile(layer.forward_native)
+
+    if provider == "custom":
+        fn = lambda: layer(x)
+    elif provider == "compiled":
+        fn = lambda: compiled_layer(x)
+
+    ms, min_ms, max_ms = triton.testing.do_bench_cudagraph(
+        fn, quantiles=[0.5, 0.2, 0.8]
+    )
+    return ms, max_ms, min_ms
+
+
+if __name__ == "__main__":
+    parser = FlexibleArgumentParser(description="Benchmark the custom activation op.")
+    parser.add_argument(
+        "--func-name",
+        type=str,
+        choices=[
+            "mul_and_silu",
+            "silu_and_mul",
+            "gelu_and_mul",
+            "gelu_and_mul_tanh",
+            "fatrelu_and_mul",
+            "swigluoai_and_mul",
+            "gelu_new",
+            "gelu_fast",
+            "quick_gelu",
+        ],
+        default="silu_and_mul",
+    )
+    parser.add_argument(
+        "--dtype", type=str, choices=["half", "bfloat16", "float"], default="bfloat16"
+    )
+    args = parser.parse_args()
+    assert args
+
+    func_name = args.func_name
+    dtype = STR_DTYPE_TO_TORCH_DTYPE[args.dtype]
+
+    perf_report = triton.testing.perf_report(
+        triton.testing.Benchmark(
+            x_names=["batch_size", "seq_len", "intermediate_size"],
+            x_vals=configs,
+            line_arg="provider",
+            line_vals=["custom", "compiled"],
+            line_names=["Custom OP", "Compiled"],
+            styles=[("blue", "-"), ("green", "-")],
+            ylabel="ms",
+            plot_name=f"{func_name}-op-performance",
+            args={},
+        )
+    )
+
+    perf_report(
+        lambda batch_size, seq_len, intermediate_size, provider: benchmark_activation(
+            batch_size, seq_len, intermediate_size, provider, func_name, dtype
+        )
+    ).run(print_data=True)
--- a/benchmarks/kernels/benchmark_device_communicators.py
+++ b/benchmarks/kernels/benchmark_device_communicators.py
@ -0,0 +1,486 @@
+#!/usr/bin/env python3
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+"""
+Benchmark script for device communicators:
+CustomAllreduce (oneshot, twoshot), PyNcclCommunicator,
+and SymmMemCommunicator (multimem, two-shot).
+
+Usage:
+    torchrun --nproc_per_node=<N> benchmark_device_communicators.py [options]
+
+Example:
+    torchrun --nproc_per_node=2 benchmark_device_communicators.py
+    --sequence-lengths 512 1024 2048 --num-warmup 10 --num-trials 100
+"""
+
+import json
+import os
+import time
+from contextlib import nullcontext
+from typing import Callable, Optional
+
+import torch
+import torch.distributed as dist
+from torch.distributed import ProcessGroup
+
+from vllm.distributed.device_communicators.custom_all_reduce import CustomAllreduce
+from vllm.distributed.device_communicators.pynccl import PyNcclCommunicator
+from vllm.distributed.device_communicators.symm_mem import SymmMemCommunicator
+from vllm.logger import init_logger
+from vllm.utils import FlexibleArgumentParser
+
+logger = init_logger(__name__)
+
+# Default sequence lengths to benchmark
+DEFAULT_SEQUENCE_LENGTHS = [128, 512, 1024, 2048, 4096, 8192]
+
+# Fixed hidden size and dtype for all benchmarks
+HIDDEN_SIZE = 8192
+BENCHMARK_DTYPE = torch.bfloat16
+
+# CUDA graph settings
+CUDA_GRAPH_CAPTURE_CYCLES = 10
+
+
+class CommunicatorBenchmark:
+    """Benchmark class for testing device communicators."""
+
+    def __init__(
+        self,
+        rank: int,
+        world_size: int,
+        device: torch.device,
+        cpu_group: ProcessGroup,
+        sequence_lengths: list[int],
+    ):
+        self.rank = rank
+        self.world_size = world_size
+        self.device = device
+        self.cpu_group = cpu_group
+
+        # Calculate max_size_override based on largest sequence length
+        max_seq_len = max(sequence_lengths)
+        max_tensor_elements = max_seq_len * HIDDEN_SIZE
+        self.max_size_override = max_tensor_elements * BENCHMARK_DTYPE.itemsize + 1
+
+        # Initialize communicators
+        self.custom_allreduce = None
+        self.pynccl_comm = None
+        self.symm_mem_comm = None
+        self.symm_mem_comm_multimem = None
+        self.symm_mem_comm_two_shot = None
+
+        self._init_communicators()
+
+    def _init_communicators(self):
+        """Initialize all available communicators."""
+        try:
+            self.custom_allreduce = CustomAllreduce(
+                group=self.cpu_group,
+                device=self.device,
+                max_size=self.max_size_override,
+            )
+            if not self.custom_allreduce.disabled:
+                logger.info("Rank %s: CustomAllreduce initialized", self.rank)
+            else:
+                logger.info("Rank %s: CustomAllreduce disabled", self.rank)
+        except Exception as e:
+            logger.warning(
+                "Rank %s: Failed to initialize CustomAllreduce: %s", self.rank, e
+            )
+            self.custom_allreduce = None
+
+        try:
+            self.pynccl_comm = PyNcclCommunicator(
+                group=self.cpu_group, device=self.device
+            )
+            if not self.pynccl_comm.disabled:
+                logger.info("Rank %s: PyNcclCommunicator initialized", self.rank)
+            else:
+                logger.info("Rank %s: PyNcclCommunicator disabled", self.rank)
+                self.pynccl_comm = None
+        except Exception as e:
+            logger.warning(
+                "Rank %s: Failed to initialize PyNcclCommunicator: %s", self.rank, e
+            )
+            self.pynccl_comm = None
+
+        # Initialize variants for SymmMemCommunicator
+        try:
+            self.symm_mem_comm_multimem = SymmMemCommunicator(
+                group=self.cpu_group,
+                device=self.device,
+                force_multimem=True,
+                max_size_override=self.max_size_override,
+            )
+            if not self.symm_mem_comm_multimem.disabled:
+                logger.info(
+                    "Rank %s: SymmMemCommunicator (multimem) initialized", self.rank
+                )
+            else:
+                self.symm_mem_comm_multimem = None
+        except Exception as e:
+            logger.warning(
+                "Rank %s: Failed to initialize SymmMemCommunicator (multimem): %s",
+                self.rank,
+                e,
+            )
+            self.symm_mem_comm_multimem = None
+
+        try:
+            self.symm_mem_comm_two_shot = SymmMemCommunicator(
+                group=self.cpu_group,
+                device=self.device,
+                force_multimem=False,
+                max_size_override=self.max_size_override,
+            )
+            if not self.symm_mem_comm_two_shot.disabled:
+                logger.info(
+                    "Rank %s: SymmMemCommunicator (two_shot) initialized", self.rank
+                )
+            else:
+                self.symm_mem_comm_two_shot = None
+        except Exception as e:
+            logger.warning(
+                "Rank %s: Failed to initialize SymmMemCommunicator (two_shot): %s",
+                self.rank,
+                e,
+            )
+            self.symm_mem_comm_two_shot = None
+
+    def benchmark_allreduce(
+        self, sequence_length: int, num_warmup: int, num_trials: int
+    ) -> dict[str, float]:
+        """Benchmark allreduce operations for all available communicators."""
+
+        results = {}
+
+        # Define communicators with their benchmark functions
+        communicators = []
+
+        if self.custom_allreduce is not None:
+            comm = self.custom_allreduce
+            # CustomAllreduce one-shot
+            communicators.append(
+                (
+                    "ca_1stage",
+                    lambda t, c=comm: c.custom_all_reduce(t),
+                    lambda t, c=comm: c.should_custom_ar(t),
+                    comm.capture(),
+                    "1stage",  # env variable value
+                )
+            )
+            # CustomAllreduce two-shot
+            communicators.append(
+                (
+                    "ca_2stage",
+                    lambda t, c=comm: c.custom_all_reduce(t),
+                    lambda t, c=comm: c.should_custom_ar(t),
+                    comm.capture(),
+                    "2stage",  # env variable value
+                )
+            )
+
+        if self.pynccl_comm is not None:
+            comm = self.pynccl_comm
+            communicators.append(
+                (
+                    "pynccl",
+                    lambda t, c=comm: c.all_reduce(t),
+                    lambda t: True,  # Always available if initialized
+                    nullcontext(),
+                    None,  # no env variable needed
+                )
+            )
+
+        if self.symm_mem_comm_multimem is not None:
+            comm = self.symm_mem_comm_multimem
+            communicators.append(
+                (
+                    "symm_mem_multimem",
+                    lambda t, c=comm: c.all_reduce(t),
+                    lambda t, c=comm: c.should_use_symm_mem(t),
+                    nullcontext(),
+                    None,  # no env variable needed
+                )
+            )
+
+        if self.symm_mem_comm_two_shot is not None:
+            comm = self.symm_mem_comm_two_shot
+            communicators.append(
+                (
+                    "symm_mem_two_shot",
+                    lambda t, c=comm: c.all_reduce(t),
+                    lambda t, c=comm: c.should_use_symm_mem(t),
+                    nullcontext(),
+                    None,  # no env variable needed
+                )
+            )
+
+        # Benchmark each communicator
+        for name, allreduce_fn, should_use_fn, context, env_var in communicators:
+            # Set environment variable if needed
+            if env_var is not None:
+                os.environ["VLLM_CUSTOM_ALLREDUCE_ALGO"] = env_var
+            else:
+                # Clear the environment variable to avoid interference
+                os.environ.pop("VLLM_CUSTOM_ALLREDUCE_ALGO", None)
+
+            latency = self.benchmark_allreduce_single(
+                sequence_length,
+                allreduce_fn,
+                should_use_fn,
+                context,
+                num_warmup,
+                num_trials,
+            )
+            if latency is not None:
+                results[name] = latency
+
+        return results
+
+    def benchmark_allreduce_single(
+        self,
+        sequence_length: int,
+        allreduce_fn: Callable[[torch.Tensor], Optional[torch.Tensor]],
+        should_use_fn: Callable[[torch.Tensor], bool],
+        context,
+        num_warmup: int,
+        num_trials: int,
+    ) -> Optional[float]:
+        """Benchmark method with CUDA graph optimization."""
+        try:
+            # Create test tensor (2D: sequence_length x hidden_size)
+            tensor = torch.randn(
+                sequence_length, HIDDEN_SIZE, dtype=BENCHMARK_DTYPE, device=self.device
+            )
+            if not should_use_fn(tensor):
+                return None
+
+            torch.cuda.synchronize()
+            stream = torch.cuda.Stream()
+            with torch.cuda.stream(stream):
+                graph_input = tensor.clone()
+
+                # Warmup before capture
+                for _ in range(3):
+                    allreduce_fn(graph_input)
+
+                # Capture the graph using context manager
+                with context:
+                    graph = torch.cuda.CUDAGraph()
+                    with torch.cuda.graph(graph):
+                        for _ in range(CUDA_GRAPH_CAPTURE_CYCLES):
+                            allreduce_fn(graph_input)
+
+            torch.cuda.synchronize()
+            for _ in range(num_warmup):
+                graph.replay()
+            torch.cuda.synchronize()
+
+            torch.cuda.synchronize()
+            start_time = time.perf_counter()
+
+            for _ in range(num_trials):
+                graph.replay()
+            torch.cuda.synchronize()
+
+            end_time = time.perf_counter()
+
+            # Convert to ms and divide by CUDA_GRAPH_CAPTURE_CYCLES
+            return (
+                (end_time - start_time) / num_trials / CUDA_GRAPH_CAPTURE_CYCLES * 1000
+            )
+
+        except Exception as e:
+            logger.error("CUDA graph benchmark failed: %s", e)
+            raise RuntimeError(
+                f"CUDA graph benchmark failed for communicator: {e}"
+            ) from e
+
+
+def _calculate_speedup_info(comm_results: dict[str, float]) -> str:
+    """Calculate speedup information for a single tensor size."""
+    if not comm_results:
+        return "N/A"
+
+    # Find the fastest communicator
+    fastest_comm = min(comm_results.keys(), key=lambda k: comm_results[k])
+    fastest_time = comm_results[fastest_comm]
+
+    # Calculate speedup vs PyNccl if available
+    if "pynccl" in comm_results:
+        pynccl_time = comm_results["pynccl"]
+        speedup = pynccl_time / fastest_time
+        return f"{fastest_comm} ({speedup:.2f}x)"
+    else:
+        return f"{fastest_comm} (N/A)"
+
+
+def print_results(
+    results: dict[str, dict[str, float]], sequence_lengths: list[int], world_size: int
+):
+    """Print benchmark results in a formatted table."""
+
+    print(f"\n{'=' * 130}")
+    print("Device Communicator Benchmark Results")
+    print(
+        f"World Size: {world_size}, Data Type: {BENCHMARK_DTYPE}, "
+        f"Hidden Size: {HIDDEN_SIZE}"
+    )
+    print(f"{'=' * 130}")
+
+    # Get all communicator names
+    all_comms = set()
+    for size_results in results.values():
+        all_comms.update(size_results.keys())
+
+    all_comms = sorted(list(all_comms))
+
+    # Print header
+    header = f"{'Tensor Shape':<20}{'Tensor Size':<15}"
+    for comm in all_comms:
+        header += f"{comm:<20}"
+    header += f"{'Best (Speedup vs PyNccl)':<30}"
+    print(header)
+    print("-" * len(header))
+
+    # Print results for each sequence length
+    for seq_len in sequence_lengths:
+        if seq_len in results:
+            # Calculate tensor size in elements and bytes
+            tensor_elements = seq_len * HIDDEN_SIZE
+            tensor_bytes = tensor_elements * BENCHMARK_DTYPE.itemsize
+
+            # Format tensor size (MB)
+            tensor_size_mb = tensor_bytes / (1024 * 1024)
+            tensor_size_str = f"{tensor_size_mb:.2f} MB"
+
+            # Format tensor shape
+            tensor_shape = f"({seq_len}, {HIDDEN_SIZE})"
+
+            row = f"{tensor_shape:<20}{tensor_size_str:<15}"
+            for comm in all_comms:
+                if comm in results[seq_len]:
+                    row += f"{results[seq_len][comm]:<20.3f}"
+                else:
+                    row += f"{'N/A':<20}"
+
+            # Calculate speedup information
+            speedup_info = _calculate_speedup_info(results[seq_len])
+            row += f"{speedup_info:<30}"
+
+            print(row)
+
+    print(f"{'=' * 130}")
+    print("All times are in milliseconds (ms) per allreduce operation")
+    print("Speedup column shows: fastest_algorithm (speedup_vs_pynccl)")
+
+
+def main():
+    parser = FlexibleArgumentParser(description="Benchmark device communicators")
+
+    parser.add_argument(
+        "--sequence-lengths",
+        type=int,
+        nargs="+",
+        default=DEFAULT_SEQUENCE_LENGTHS,
+        help="Sequence lengths to benchmark (tensor shape: seq_len x hidden_size)",
+    )
+
+    parser.add_argument(
+        "--num-warmup", type=int, default=5, help="Number of warmup iterations"
+    )
+
+    parser.add_argument(
+        "--num-trials", type=int, default=50, help="Number of benchmark trials"
+    )
+
+    parser.add_argument("--output-json", type=str, help="Output results to JSON file")
+
+    args = parser.parse_args()
+
+    # Initialize distributed
+    if not dist.is_initialized():
+        dist.init_process_group(backend="gloo")
+    rank = dist.get_rank()
+    world_size = dist.get_world_size()
+
+    # Set device
+    device = torch.device(f"cuda:{rank}")
+    torch.cuda.set_device(device)
+
+    # Get CPU process group
+    cpu_group = dist.new_group(backend="gloo")
+
+    # Disable USE_SYMM_MEM to avoid affecting the max_sizes
+    # in symm_mem and custom_all_reduce for benchmark
+    os.environ["VLLM_ALLREDUCE_USE_SYMM_MEM"] = "0"
+
+    # Initialize benchmark
+    benchmark = CommunicatorBenchmark(
+        rank, world_size, device, cpu_group, args.sequence_lengths
+    )
+
+    # Run benchmarks
+    all_results = {}
+
+    for seq_len in args.sequence_lengths:
+        if rank == 0:
+            logger.info(
+                "Benchmarking sequence length: %s (tensor shape: %s x %s)",
+                seq_len,
+                seq_len,
+                HIDDEN_SIZE,
+            )
+
+        results = benchmark.benchmark_allreduce(
+            sequence_length=seq_len,
+            num_warmup=args.num_warmup,
+            num_trials=args.num_trials,
+        )
+
+        all_results[seq_len] = results
+
+        # Synchronize between ranks
+        dist.barrier()
+
+    # Print results (only rank 0)
+    if rank == 0:
+        print_results(all_results, args.sequence_lengths, world_size)
+
+        # Save to JSON if requested
+        if args.output_json:
+            # Add speedup information to results
+            enhanced_results = {}
+            for seq_len, comm_results in all_results.items():
+                enhanced_results[seq_len] = {
+                    "timings": comm_results,
+                    "speedup_info": _calculate_speedup_info(comm_results),
+                }
+
+            output_data = {
+                "world_size": world_size,
+                "dtype": str(BENCHMARK_DTYPE),
+                "hidden_size": HIDDEN_SIZE,
+                "sequence_lengths": args.sequence_lengths,
+                "num_warmup": args.num_warmup,
+                "num_trials": args.num_trials,
+                "cuda_graph_capture_cycles": CUDA_GRAPH_CAPTURE_CYCLES,
+                "results": enhanced_results,
+            }
+
+            with open(args.output_json, "w") as f:
+                json.dump(output_data, f, indent=2)
+
+            logger.info("Results saved to %s", args.output_json)
+
+    # Cleanup
+    if cpu_group != dist.group.WORLD:
+        dist.destroy_process_group(cpu_group)
+
+
+if __name__ == "__main__":
+    main()
--- 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_lora.py
+++ b/benchmarks/kernels/benchmark_lora.py
@ -637,7 +637,7 @@ def bench_optype(
    # Clear LoRA optimization hash-maps.
    _LORA_A_PTR_DICT.clear()
    _LORA_B_PTR_DICT.clear()
-    # Run bench function so that _LORA_A_PTR_DICT and _LORA_B_PTR_DICT are setup
+    # Run bench function so that _LORA_A_PTR_DICT and _LORA_B_PTR_DICT are set up
    for kwargs in kwargs_list:
        op_type.bench_fn()(**kwargs)
    torch.cuda.synchronize()
--- a/benchmarks/kernels/benchmark_machete.py
+++ b/benchmarks/kernels/benchmark_machete.py
@ -253,28 +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

@ -305,6 +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
@ -406,6 +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
@ -419,8 +419,10 @@ class BenchmarkWorker:
        )
        # NOTE(woosuk): The current naming convention uses w2.shape[2], which
        # is the intermediate size after silu_and_mul.
+        block_n = block_quant_shape[0] if block_quant_shape else None
+        block_k = block_quant_shape[1] if block_quant_shape else None
        op_config = get_moe_configs(
-            num_experts, shard_intermediate_size // 2, dtype_str
+            num_experts, shard_intermediate_size // 2, dtype_str, block_n, block_k
        )
        if op_config is None:
            config = get_default_config(
@ -430,7 +432,7 @@ class BenchmarkWorker:
                hidden_size,
                topk,
                dtype_str,
-                is_marlin=False,
+                block_quant_shape,
            )
        else:
            config = op_config[min(op_config.keys(), key=lambda x: abs(x - num_tokens))]
@ -592,7 +594,11 @@ def main(args: argparse.Namespace):
        E = config.n_routed_experts
        topk = config.num_experts_per_tok
        intermediate_size = config.moe_intermediate_size
-    elif config.architectures[0] in ("Qwen2MoeForCausalLM", "Qwen3MoeForCausalLM"):
+    elif config.architectures[0] in (
+        "Qwen2MoeForCausalLM",
+        "Qwen3MoeForCausalLM",
+        "Qwen3NextForCausalLM",
+    ):
        E = config.num_experts
        topk = config.num_experts_per_tok
        intermediate_size = config.moe_intermediate_size
@ -676,7 +682,11 @@ def main(args: argparse.Namespace):
        is_fp16 = not (use_fp8_w8a8 or use_int8_w8a16)
        search_space = get_configs_compute_bound(is_fp16, block_quant_shape)
        print(f"Start tuning over {len(search_space)} configurations...")
-
+        if use_deep_gemm:
+            raise ValueError(
+                "Tuning with --use-deep-gemm is not supported as it only tunes Triton "
+                "kernels. Please remove the flag."
+            )
        start = time.time()
        configs = _distribute(
            "tune",
--- a/benchmarks/kernels/benchmark_polynorm.py
+++ b/benchmarks/kernels/benchmark_polynorm.py
@ -0,0 +1,155 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+import itertools
+
+import torch
+
+from vllm import _custom_ops as vllm_ops
+from vllm.triton_utils import triton
+
+
+def polynorm_naive(
+    x: torch.Tensor,
+    weight: torch.Tensor,
+    bias: torch.Tensor,
+    eps: float = 1e-6,
+):
+    orig_shape = x.shape
+    x = x.view(-1, x.shape[-1])
+
+    def norm(x, eps: float):
+        return x / torch.sqrt(x.pow(2).mean(-1, keepdim=True) + eps)
+
+    x = x.float()
+    return (
+        (
+            weight[0] * norm(x**3, eps)
+            + weight[1] * norm(x**2, eps)
+            + weight[2] * norm(x, eps)
+            + bias
+        )
+        .to(weight.dtype)
+        .view(orig_shape)
+    )
+
+
+def polynorm_vllm(
+    x: torch.Tensor,
+    weight: torch.Tensor,
+    bias: torch.Tensor,
+    eps: float = 1e-6,
+):
+    orig_shape = x.shape
+    x = x.view(-1, x.shape[-1])
+
+    out = torch.empty_like(x)
+    vllm_ops.poly_norm(out, x, weight, bias, eps)
+    output = out
+
+    output = output.view(orig_shape)
+    return output
+
+
+def calculate_diff(batch_size, seq_len, hidden_dim):
+    dtype = torch.bfloat16
+    x = torch.randn(batch_size, seq_len, hidden_dim, dtype=dtype, device="cuda")
+    weight = torch.ones(3, dtype=dtype, device="cuda")
+    bias = torch.ones(1, dtype=dtype, device="cuda")
+
+    output_naive = polynorm_naive(x, weight, bias)
+    output_vllm = polynorm_vllm(x, weight, bias)
+
+    if torch.allclose(output_naive, output_vllm, atol=1e-2, rtol=1e-2):
+        print("✅ All implementations match")
+    else:
+        print("❌ Implementations differ")
+
+
+batch_size_range = [2**i for i in range(0, 7, 2)]
+seq_length_range = [2**i for i in range(6, 11, 1)]
+dim_range = [2048, 4096]
+configs = list(itertools.product(dim_range, batch_size_range, seq_length_range))
+
+
+def get_benchmark():
+    @triton.testing.perf_report(
+        triton.testing.Benchmark(
+            x_names=["dim", "batch_size", "seq_len"],
+            x_vals=[list(_) for _ in configs],
+            line_arg="provider",
+            line_vals=["naive", "vllm"],
+            line_names=["Naive", "vLLM"],
+            styles=[("blue", "-"), ("red", "-")],
+            ylabel="us",
+            plot_name="polynorm-perf",
+            args={},
+        )
+    )
+    def benchmark(dim, batch_size, seq_len, provider):
+        dtype = torch.bfloat16
+        hidden_dim = dim * 4
+
+        x = torch.randn(batch_size, seq_len, hidden_dim, dtype=dtype, device="cuda")
+        weight = torch.ones(3, dtype=dtype, device="cuda")
+        bias = torch.ones(1, dtype=dtype, device="cuda")
+
+        quantiles = [0.5, 0.2, 0.8]
+
+        if provider == "naive":
+            ms, min_ms, max_ms = triton.testing.do_bench(
+                lambda: polynorm_naive(x, weight, bias),
+                quantiles=quantiles,
+            )
+        else:
+            ms, min_ms, max_ms = triton.testing.do_bench(
+                lambda: polynorm_vllm(x, weight, bias),
+                quantiles=quantiles,
+            )
+
+        return 1000 * ms, 1000 * max_ms, 1000 * min_ms
+
+    return benchmark
+
+
+if __name__ == "__main__":
+    import argparse
+
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--batch-size",
+        type=int,
+        default=4,
+        help="Batch size",
+    )
+    parser.add_argument(
+        "--seq-len",
+        type=int,
+        default=128,
+        help="Sequence length",
+    )
+    parser.add_argument(
+        "--hidden-dim",
+        type=int,
+        default=8192,
+        help="Intermediate size of MLP",
+    )
+    parser.add_argument(
+        "--save-path",
+        type=str,
+        default="./configs/polnorm/",
+        help="Path to save polnorm benchmark results",
+    )
+
+    args = parser.parse_args()
+
+    # Run correctness test
+    calculate_diff(
+        batch_size=args.batch_size,
+        seq_len=args.seq_len,
+        hidden_dim=args.hidden_dim,
+    )
+
+    benchmark = get_benchmark()
+    # Run performance benchmark
+    benchmark.run(print_data=True, save_path=args.save_path)
--- a/benchmarks/kernels/benchmark_silu_mul_fp8_quant.py
+++ b/benchmarks/kernels/benchmark_silu_mul_fp8_quant.py
@ -0,0 +1,675 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+from collections.abc import Callable
+
+import matplotlib.pyplot as plt
+import numpy as np
+import torch
+
+from vllm.model_executor.layers.fused_moe.batched_deep_gemm_moe import (
+    silu_mul_fp8_quant_deep_gemm_cuda,
+)
+from vllm.platforms import current_platform
+from vllm.triton_utils import tl, triton
+from vllm.utils.deep_gemm import is_deep_gemm_e8m0_used
+
+
+@triton.jit
+def _silu_mul_fp8_quant_deep_gemm(
+    # Pointers ------------------------------------------------------------
+    input_ptr,  # 16-bit activations (E, T, 2*H)
+    y_q_ptr,  # fp8 quantized activations (E, T, H)
+    y_s_ptr,  # 16-bit scales (E, T, G)
+    counts_ptr,  # int32 num tokens per expert (E)
+    # Sizes ---------------------------------------------------------------
+    H: tl.constexpr,  # hidden dimension (per output)
+    GROUP_SIZE: tl.constexpr,  # elements per group (usually 128)
+    # Strides for input (elements) ---------------------------------------
+    stride_i_e,
+    stride_i_t,
+    stride_i_h,
+    # Strides for y_q (elements) -----------------------------------------
+    stride_yq_e,
+    stride_yq_t,
+    stride_yq_h,
+    # Strides for y_s (elements) -----------------------------------------
+    stride_ys_e,
+    stride_ys_t,
+    stride_ys_g,
+    # Stride for counts (elements)
+    stride_counts_e,
+    # Numeric params ------------------------------------------------------
+    eps: tl.constexpr,
+    fp8_min: tl.constexpr,
+    fp8_max: tl.constexpr,
+    use_ue8m0: tl.constexpr,
+    # Meta ---------------------------------------------------------------
+    BLOCK: tl.constexpr,
+    NUM_STAGES: tl.constexpr,
+):
+    G = H // GROUP_SIZE
+
+    # map program id -> (e, g)
+    pid = tl.program_id(0)
+    e = pid // G
+    g = pid % G
+
+    e = e.to(tl.int64)
+    g = g.to(tl.int64)
+
+    # number of valid tokens for this expert
+    n_tokens = tl.load(counts_ptr + e * stride_counts_e).to(tl.int64)
+
+    cols = tl.arange(0, BLOCK).to(tl.int64)
+    mask = cols < BLOCK
+
+    base_input_offset = e * stride_i_e + g * GROUP_SIZE * stride_i_h
+    base_gate_offset = base_input_offset + cols * stride_i_h
+    base_up_offset = base_input_offset + H * stride_i_h + cols * stride_i_h
+    base_yq_offset = e * stride_yq_e + g * GROUP_SIZE * stride_yq_h + cols * stride_yq_h
+    base_ys_offset = e * stride_ys_e + g * stride_ys_g
+
+    for t in tl.range(0, n_tokens, num_stages=NUM_STAGES):
+        gate = tl.load(
+            input_ptr + base_gate_offset + t * stride_i_t, mask=mask, other=0.0
+        ).to(tl.float32)
+        up = tl.load(input_ptr + base_up_offset + t * stride_i_t, mask=mask, other=0.0)
+
+        gate = gate * (1.0 / (1.0 + tl.exp(-gate)))
+        y = gate * up
+
+        y_s = tl.maximum(tl.max(tl.abs(y)), eps) / fp8_max
+        if use_ue8m0:
+            y_s = tl.exp2(tl.ceil(tl.log2(y_s)))
+
+        y_q = tl.clamp(y / y_s, fp8_min, fp8_max).to(y_q_ptr.dtype.element_ty)
+
+        tl.store(y_q_ptr + base_yq_offset + t * stride_yq_t, y_q, mask=mask)
+        tl.store(y_s_ptr + base_ys_offset + t * stride_ys_t, y_s)
+
+
+def silu_mul_fp8_quant_deep_gemm_triton(
+    y: torch.Tensor,  # (E, T, 2*H)
+    tokens_per_expert: torch.Tensor,  # (E,) number of valid tokens per expert
+    num_parallel_tokens,
+    group_size: int = 128,
+    eps: float = 1e-10,
+) -> tuple[torch.Tensor, torch.Tensor]:
+    """Quantize silu(y[..., :H]) * y[..., H:] to FP8 with group per-token scales
+
+    y has shape (E, T, 2*H). The first half of the last dimension is
+    silu-activated, multiplied by the second half, then quantized into FP8.
+
+    Returns `(y_q, y_s)` where
+    * `y_q`: FP8 tensor, shape (E, T, H), same layout as y[..., :H]
+    * `y_s`: FP32 tensor, shape (E, T, H // group_size), strides (T*G, 1, T)
+    """
+    assert y.ndim == 3, "y must be (E, T, 2*H)"
+    E, T, H2 = y.shape
+    assert H2 % 2 == 0, "last dim of y must be even (2*H)"
+    H = H2 // 2
+    G = (H + group_size - 1) // group_size
+    assert H % group_size == 0, "H must be divisible by group_size"
+    assert tokens_per_expert.ndim == 1 and tokens_per_expert.shape[0] == E, (
+        "tokens_per_expert must be shape (E,)"
+    )
+    tokens_per_expert = tokens_per_expert.to(device=y.device, dtype=torch.int32)
+
+    # allocate outputs
+    fp8_dtype = torch.float8_e4m3fn
+    y_q = torch.empty((E, T, H), dtype=fp8_dtype, device=y.device)
+
+    # strides (elements)
+    stride_i_e, stride_i_t, stride_i_h = y.stride()
+    stride_yq_e, stride_yq_t, stride_yq_h = y_q.stride()
+
+    # desired scale strides (elements): (T*G, 1, T)
+    stride_ys_e = T * G
+    stride_ys_t = 1
+    stride_ys_g = T
+    y_s = torch.empty_strided(
+        (E, T, G),
+        (stride_ys_e, stride_ys_t, stride_ys_g),
+        dtype=torch.float32,
+        device=y.device,
+    )
+
+    stride_cnt_e = tokens_per_expert.stride()[0]
+
+    # Static grid over experts and H-groups.
+    # A loop inside the kernel handles the token dim
+    grid = (E * G,)
+
+    f_info = torch.finfo(fp8_dtype)
+    fp8_max = f_info.max
+    fp8_min = f_info.min
+
+    _silu_mul_fp8_quant_deep_gemm[grid](
+        y,
+        y_q,
+        y_s,
+        tokens_per_expert,
+        H,
+        group_size,
+        stride_i_e,
+        stride_i_t,
+        stride_i_h,
+        stride_yq_e,
+        stride_yq_t,
+        stride_yq_h,
+        stride_ys_e,
+        stride_ys_t,
+        stride_ys_g,
+        stride_cnt_e,
+        eps,
+        fp8_min,
+        fp8_max,
+        is_deep_gemm_e8m0_used(),
+        BLOCK=group_size,
+        NUM_STAGES=4,
+        num_warps=1,
+    )
+
+    return y_q, y_s
+
+
+# Parse generation strategies
+strategies = ["uniform", "max_t", "first_t"]
+
+
+def benchmark(
+    kernel: Callable,
+    E: int,
+    T: int,
+    H: int,
+    total_tokens: int,
+    num_parallel_tokens: int = 64,
+    G: int = 128,
+    runs: int = 200,
+    num_warmups: int = 20,
+    gen_strategy: str = "default",
+    iterations_per_run: int = 20,
+):
+    def generate_data(seed_offset=0):
+        """Generate input data with given seed offset"""
+        current_platform.seed_everything(42 + seed_offset)
+        y = torch.rand((E, T, 2 * H), dtype=torch.bfloat16, device="cuda").contiguous()
+
+        if gen_strategy == "uniform":
+            r = torch.rand(size=(E,), device="cuda")
+            r /= r.sum()
+            r *= total_tokens
+            tokens_per_expert = r.int()
+            tokens_per_expert = torch.minimum(
+                tokens_per_expert,
+                torch.ones((E,), device=r.device, dtype=torch.int) * T,
+            )
+        elif gen_strategy == "max_t":
+            tokens_per_expert = torch.empty(size=(E,), dtype=torch.int32, device="cuda")
+            tokens_per_expert.fill_(total_tokens / E)
+        elif gen_strategy == "first_t":
+            tokens_per_expert = torch.zeros(size=(E,), dtype=torch.int32, device="cuda")
+            tokens_per_expert[0] = min(T, total_tokens)
+        else:
+            raise ValueError(f"Unknown generation strategy: {gen_strategy}")
+        return y, tokens_per_expert
+
+    dataset_count = 4
+    # Pre-generate different input matrices for each iteration to avoid cache effects
+    data_sets = [generate_data(i) for i in range(dataset_count)]
+
+    # Warmup
+    y, tokens_per_expert = data_sets[0]
+    for _ in range(num_warmups):
+        kernel(
+            y, tokens_per_expert, num_parallel_tokens=num_parallel_tokens, group_size=G
+        )
+    torch.cuda.synchronize()
+
+    start_event = torch.cuda.Event(enable_timing=True)
+    end_event = torch.cuda.Event(enable_timing=True)
+
+    # Benchmark
+    latencies: list[float] = []
+    for _ in range(runs):
+        torch.cuda.synchronize()
+
+        start_event.record()
+        for i in range(iterations_per_run):
+            y, tokens_per_expert = data_sets[i % dataset_count]
+            kernel(
+                y,
+                tokens_per_expert,
+                num_parallel_tokens=num_parallel_tokens,
+                group_size=G,
+            )
+        end_event.record()
+        end_event.synchronize()
+
+        total_time_ms = start_event.elapsed_time(end_event)
+        per_iter_time_ms = total_time_ms / iterations_per_run
+        latencies.append(per_iter_time_ms)
+
+    # Use median instead of average for better outlier handling
+    median_time_ms = np.median(latencies)
+    median_time_s = median_time_ms / 1000
+
+    # Calculate actual work done (using first dataset for consistency)
+    _, tokens_per_expert = data_sets[0]
+    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 / median_time_s / 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 / median_time_s / 1e9
+
+    HOPPER_BANDWIDTH_TBPS = 3.35
+    return (
+        median_time_ms,
+        gflops,
+        memory_bw,
+        (memory_bw / (HOPPER_BANDWIDTH_TBPS * 1024)) * 100,
+    )
+
+
+def create_comparison_plot(
+    ratio, cuda_times, baseline_times, config_labels, strategy_name, id
+):
+    """Create a comparison plot for a specific generation strategy"""
+    fig, ax = plt.subplots(1, 1, figsize=(16, 6))
+
+    # Configure x-axis positions
+    x = np.arange(len(config_labels))
+    width = 0.35
+
+    # Execution Time plot (lower is better)
+    ax.bar(
+        x - width / 2, cuda_times, width, label="CUDA Kernel", alpha=0.8, color="blue"
+    )
+    ax.bar(
+        x + width / 2,
+        baseline_times,
+        width,
+        label="Baseline",
+        alpha=0.8,
+        color="orange",
+    )
+
+    # Add speedup labels over each bar pair
+    for i in range(len(x)):
+        speedup = ratio[i]
+        max_height = max(cuda_times[i], baseline_times[i])
+        ax.text(
+            x[i],
+            max_height + max_height * 0.02,
+            f"{speedup:.2f}x",
+            ha="center",
+            va="bottom",
+            fontweight="bold",
+            fontsize=9,
+        )
+
+    ax.set_xlabel("Configuration")
+    ax.set_ylabel("% Utilization")
+    ax.set_title(
+        f"Memory Bandwidth Utilization (%) - {strategy_name}\n(Higher is Better)"
+    )
+    ax.set_xticks(x)
+    ax.set_xticklabels(config_labels, rotation=45, ha="right")
+    ax.legend()
+    ax.grid(True, alpha=0.3)
+
+    plt.tight_layout()
+    return fig, ax
+
+
+def create_combined_plot(all_results):
+    """Create a combined plot with all strategies in one PNG"""
+    num_strategies = len(all_results)
+    fig, axes = plt.subplots(num_strategies, 1, figsize=(20, 6 * num_strategies))
+
+    if num_strategies == 1:
+        axes = [axes]
+
+    for idx, (
+        strategy_name,
+        ratio,
+        cuda_times,
+        baseline_times,
+        config_labels,
+    ) in enumerate(all_results):
+        ax = axes[idx]
+
+        # Configure x-axis positions
+        x = np.arange(len(config_labels))
+        width = 0.35
+
+        # Execution Time plot (lower is better)
+        ax.bar(
+            x - width / 2,
+            cuda_times,
+            width,
+            label="CUDA Kernel",
+            alpha=0.8,
+            color="blue",
+        )
+        ax.bar(
+            x + width / 2,
+            baseline_times,
+            width,
+            label="Baseline",
+            alpha=0.8,
+            color="orange",
+        )
+
+        # Add speedup labels over each bar pair
+        for i in range(len(x)):
+            speedup = ratio[i]
+            max_height = max(cuda_times[i], baseline_times[i])
+            ax.text(
+                x[i],
+                max_height + max_height * 0.02,
+                f"{speedup:.2f}x",
+                ha="center",
+                va="bottom",
+                fontweight="bold",
+                fontsize=9,
+            )
+
+        ax.set_xlabel("Configuration")
+        ax.set_ylabel("% Utilization")
+        ax.set_title(
+            f"Memory Bandwidth Utilization (%) - {strategy_name}\n(Higher is Better)"
+        )
+        ax.set_xticks(x)
+        ax.set_xticklabels(config_labels, rotation=45, ha="right")
+        ax.legend()
+        ax.grid(True, alpha=0.3)
+
+    plt.tight_layout()
+    filename = "../../silu_bench/silu_benchmark_combined.png"
+    plt.savefig(filename, dpi=300, bbox_inches="tight")
+    plt.show()
+
+    return filename
+
+
+outer_dim = 7168
+configs = [
+    # DeepSeekV3 Configs
+    (8, 1024, 7168),
+    # DeepSeekV3 Configs
+    (32, 1024, 7168),
+    # DeepSeekV3 Configs
+    (256, 1024, 7168),
+]
+
+runs = 100
+num_warmups = 20
+
+strategy_descriptions = {
+    "uniform": "Uniform Random",
+    "max_t": "Even Assignment",
+    "first_t": "experts[0] = T, experts[1:] = 0",
+}
+
+print(f"GPU: {torch.cuda.get_device_name()}")
+print(f"Testing strategies: {', '.join(strategies)}")
+print(f"Configurations: {len(configs)} configs")
+
+all_results = []
+
+# Run benchmarks for each strategy
+for id, strategy in enumerate(strategies):
+    print(f"\n{'=' * 60}")
+    print(f"Testing strategy: {strategy_descriptions[strategy]}")
+    print(f"{'=' * 60}")
+
+    # Collect benchmark data for both algorithms
+    config_labels = []
+    config_x_axis = []
+    all_cuda_results = []
+    all_baseline_results = []
+    all_ratios = []
+
+    for E, T, H in configs:
+        total_tokens_config = [8 * E, 16 * E, 32 * E, 64 * E, 128 * E, 256 * E]
+        config_x_axis.append(total_tokens_config)
+
+        cuda_results = []
+        baseline_results = []
+        ratios = []
+
+        for total_tokens in total_tokens_config:
+            config_label = f"E={E},T={T},H={H},TT={total_tokens}"
+            config_labels.append(config_label)
+
+            # CUDA kernel results
+            time_ms_cuda, gflops, gbps, perc = benchmark(
+                silu_mul_fp8_quant_deep_gemm_cuda,
+                E,
+                T,
+                H,
+                total_tokens,
+                runs=runs,
+                num_warmups=num_warmups,
+                gen_strategy=strategy,
+            )
+            cuda_results.append((time_ms_cuda, gflops, gbps, perc))
+
+            # Baseline results
+            time_ms_triton, gflops, gbps, perc = benchmark(
+                silu_mul_fp8_quant_deep_gemm_triton,
+                E,
+                T,
+                H,
+                total_tokens,
+                runs=runs,
+                num_warmups=num_warmups,
+                gen_strategy=strategy,
+            )
+            baseline_results.append((time_ms_triton, gflops, gbps, perc))
+            ratios.append(time_ms_triton / time_ms_cuda)
+
+            print(f"Completed: {config_label}")
+        all_cuda_results.append(cuda_results)
+        all_baseline_results.append(baseline_results)
+        all_ratios.append(ratios)
+
+    # Store results for combined plotting
+    all_results.append(
+        (
+            strategy_descriptions[strategy],
+            all_ratios,
+            all_cuda_results,
+            all_baseline_results,
+            config_labels,
+            config_x_axis,
+        )
+    )
+
+    # Print summary table for this strategy
+    print(f"\nSummary Table - {strategy_descriptions[strategy]}:")
+    print(f"{'Config':<20} {'CUDA Time(ms)':<12} {'Base Time(ms)':<12} {'Speedup':<8}")
+    print("-" * 60)
+
+    for i, (E, T, H) in enumerate(configs):
+        speedup = baseline_results[i][0] / cuda_results[i][0]
+        config_label = f"E={E:3d},T={T:4d},H={H:4d}"
+        print(
+            f"{config_label:<20} {cuda_results[i][0]:8.5f} "
+            f"{baseline_results[i][0]:8.5f} {speedup:6.2f}x"
+        )
+
+
+def create_total_tokens_plot(all_results):
+    num_strategies = len(all_results)
+    num_configs = len(configs)
+
+    # Create side-by-side subplots: 2 columns for speedup and bandwidth percentage
+    fig, axs = plt.subplots(
+        num_strategies, num_configs * 2, figsize=(28, 6 * num_strategies)
+    )
+
+    # Add main title to the entire figure
+    fig.suptitle(
+        "Performance Analysis: Speedup vs Bandwidth Utilization (Triton & CUDA)",
+        fontsize=16,
+        fontweight="bold",
+        y=0.98,
+    )
+
+    # Handle single strategy case
+    if num_strategies == 1:
+        axs = axs.reshape(1, -1)
+
+    # Handle single config case
+    if num_configs == 1:
+        axs = axs.reshape(-1, 2)
+
+    for strategy_idx, result in enumerate(all_results):
+        (
+            strategy_name,
+            all_ratios,
+            all_cuda_results,
+            all_baseline_results,
+            config_labels,
+            config_x_axis,
+        ) = result
+
+        for config_idx in range(num_configs):
+            # Speedup plot (left column)
+            ax_speedup = axs[strategy_idx, config_idx * 2]
+            # Bandwidth plot (right column)
+            ax_bandwidth = axs[strategy_idx, config_idx * 2 + 1]
+
+            E, T, H = configs[config_idx]
+            ratios = all_ratios[config_idx]
+            total_tokens_values = config_x_axis[config_idx]
+
+            # Extract CUDA and Triton bandwidth percentages
+            cuda_bandwidth_percentages = [
+                result[3] for result in all_cuda_results[config_idx]
+            ]
+            triton_bandwidth_percentages = [
+                result[3] for result in all_baseline_results[config_idx]
+            ]
+
+            # Plot speedup ratios vs total tokens (left plot)
+            ax_speedup.plot(
+                total_tokens_values, ratios, "bo-", linewidth=3, markersize=8
+            )
+            ax_speedup.set_title(
+                f"{strategy_name}\nSpeedup (CUDA/Triton)\nE={E}, T={T}, H={H}",
+                fontsize=12,
+                fontweight="bold",
+            )
+            ax_speedup.set_xlabel("Total Tokens", fontweight="bold", fontsize=11)
+            ax_speedup.set_ylabel("Speedup Ratio", fontweight="bold", fontsize=11)
+            ax_speedup.grid(True, alpha=0.3)
+
+            ax_bandwidth.plot(
+                total_tokens_values,
+                cuda_bandwidth_percentages,
+                "ro-",
+                linewidth=3,
+                markersize=8,
+                label="CUDA",
+            )
+            ax_bandwidth.plot(
+                total_tokens_values,
+                triton_bandwidth_percentages,
+                "go-",
+                linewidth=3,
+                markersize=8,
+                label="Triton",
+            )
+            ax_bandwidth.set_title(
+                f"{strategy_name}\nBandwidth Utilization (Hopper)\nE={E}, T={T}, H={H}",
+                fontsize=12,
+                fontweight="bold",
+            )
+            ax_bandwidth.set_xlabel("Total Tokens", fontweight="bold", fontsize=11)
+            ax_bandwidth.set_ylabel(
+                "% of Peak Bandwidth", fontweight="bold", fontsize=11
+            )
+            ax_bandwidth.legend(prop={"weight": "bold"})
+            ax_bandwidth.grid(True, alpha=0.3)
+
+            # Format x-axis labels for both plots
+            for ax in [ax_speedup, ax_bandwidth]:
+                ax.set_xticks(total_tokens_values)
+                ax.set_xticklabels(
+                    [
+                        f"{tt // 1000}K" if tt >= 1000 else str(tt)
+                        for tt in total_tokens_values
+                    ],
+                    fontweight="bold",
+                )
+                # Make tick labels bold
+                for label in ax.get_xticklabels() + ax.get_yticklabels():
+                    label.set_fontweight("bold")
+
+            # Add value labels on speedup points
+            for x, y in zip(total_tokens_values, ratios):
+                ax_speedup.annotate(
+                    f"{y:.2f}x",
+                    (x, y),
+                    textcoords="offset points",
+                    xytext=(0, 12),
+                    ha="center",
+                    fontsize=10,
+                    fontweight="bold",
+                    bbox=dict(boxstyle="round,pad=0.3", facecolor="white", alpha=0.7),
+                )
+
+            # Add value labels on CUDA bandwidth points
+            for x, y in zip(total_tokens_values, cuda_bandwidth_percentages):
+                ax_bandwidth.annotate(
+                    f"{y:.1f}%",
+                    (x, y),
+                    textcoords="offset points",
+                    xytext=(0, 12),
+                    ha="center",
+                    fontsize=9,
+                    fontweight="bold",
+                    bbox=dict(boxstyle="round,pad=0.2", facecolor="red", alpha=0.3),
+                )
+
+            # Add value labels on Triton bandwidth points
+            for x, y in zip(total_tokens_values, triton_bandwidth_percentages):
+                ax_bandwidth.annotate(
+                    f"{y:.1f}%",
+                    (x, y),
+                    textcoords="offset points",
+                    xytext=(0, -15),
+                    ha="center",
+                    fontsize=9,
+                    fontweight="bold",
+                    bbox=dict(boxstyle="round,pad=0.2", facecolor="green", alpha=0.3),
+                )
+
+    plt.tight_layout()
+    plt.subplots_adjust(top=0.93)  # Make room for main title
+    filename = "silu_benchmark_total_tokens.png"
+    plt.savefig(filename, dpi=300, bbox_inches="tight")
+    plt.show()
+
+    return filename
+
+
+# Create combined plot with all strategies
+combined_plot_filename = create_total_tokens_plot(all_results)
+
+print(f"\n{'=' * 60}")
+print("Benchmark Complete!")
+print(f"Generated combined plot: {combined_plot_filename}")
+print(f"{'=' * 60}")
--- 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 +250,44 @@ 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, 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 +266,43 @@ 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_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, 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,
@ -56,7 +56,7 @@ def w8a8_block_matmul(
        Bs: The per-block quantization scale for `B`.
        block_size: The block size for per-block quantization.
                    It should be 2-dim, e.g., [128, 128].
-        output_dytpe: The dtype of the returned tensor.
+        output_dtype: The dtype of the returned tensor.

    Returns:
        torch.Tensor: The result of matmul.
@ -141,6 +141,7 @@ def get_weight_shapes(tp_size):
    # cannot TP
    total = [
        (512 + 64, 7168),
+        (2112, 7168),
        ((128 + 64) * 128, 7168),
        (128 * (128 + 128), 512),
        (7168, 16384),
--- 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,
@ -960,7 +962,7 @@ async def main_mp(

    # At this point all the clients finished,
    # collect results (TTFT, TPOT, etc.) from all the clients.
-    # This needs to happens before calling join on the clients
+    # This needs to happen before calling join on the clients
    # (result_queue should be emptied).
    while not result_queue.empty():
        client_metrics.append(result_queue.get())
@ -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
@ -1,6 +1,7 @@
 include(FetchContent)

 set(CMAKE_CXX_STANDARD_REQUIRED ON)
+set(CMAKE_CXX_STANDARD 17)
 set(CMAKE_CXX_EXTENSIONS ON)
 set(CMAKE_EXPORT_COMPILE_COMMANDS ON)

@ -87,6 +88,7 @@ is_avx512_disabled(AVX512_DISABLED)

 if (MACOSX_FOUND AND CMAKE_SYSTEM_PROCESSOR STREQUAL "arm64")
    message(STATUS "Apple Silicon Detected")
+    set(APPLE_SILICON_FOUND TRUE)
    set(ENABLE_NUMA OFF)
    check_sysctl(hw.optional.neon ASIMD_FOUND)
    check_sysctl(hw.optional.arm.FEAT_BF16 ARM_BF16_FOUND)
@ -182,17 +184,17 @@ 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 AND NOT APPLE_SILICON_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 +206,7 @@ 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 +219,23 @@ 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 +262,6 @@ 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 +275,11 @@ 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 57b4e68b9f9d94750b46de8f8dbd2bfcc86edd4f
+          GIT_TAG ee4d25bd84e0cbc7e0b9b9685085fd5db2dcb62a
          GIT_PROGRESS TRUE
          # Don't share the vllm-flash-attn build between build types
          BINARY_DIR ${CMAKE_BINARY_DIR}/vllm-flash-attn
--- a/csrc/attention/attention_kernels.cuh
+++ b/csrc/attention/attention_kernels.cuh
@ -28,10 +28,10 @@

 #ifdef USE_ROCM
  #include <hip/hip_bf16.h>
-  #include "../quantization/fp8/amd/quant_utils.cuh"
+  #include "../quantization/w8a8/fp8/amd/quant_utils.cuh"
 typedef __hip_bfloat16 __nv_bfloat16;
 #else
-  #include "../quantization/fp8/nvidia/quant_utils.cuh"
+  #include "../quantization/w8a8/fp8/nvidia/quant_utils.cuh"
 #endif

 #define MAX(a, b) ((a) > (b) ? (a) : (b))
--- a/csrc/attention/mla/sm100_cutlass_mla_kernel.cu
+++ b/csrc/attention/mla/sm100_cutlass_mla_kernel.cu
@ -36,12 +36,14 @@ limitations under the License.
 #if !defined(CUDA_VERSION) || CUDA_VERSION < 12040
 void sm100_cutlass_mla_decode(
    torch::Tensor const& out,
+    torch::Tensor const& lse,
    torch::Tensor const& q_nope,
    torch::Tensor const& q_pe,
    torch::Tensor const& kv_c_and_k_pe_cache,
    torch::Tensor const& seq_lens,
    torch::Tensor const& page_table,
    torch::Tensor const& workspace,
+    double sm_scale,
    int64_t num_kv_splits) {
  TORCH_CHECK(false, "CUDA version must be >= 12.4 for cutlass_mla_decode");
 }
@ -64,11 +66,11 @@ struct IsPersistent {
  static const bool value = v;
 };

-template <typename T, bool IsPaged128, typename PersistenceOption = IsPersistent<true>>
+template <typename T, typename TOut, bool IsPaged128, typename PersistenceOption = IsPersistent<true>>
 struct MlaSm100 {
  using Element = T;
  using ElementAcc = float;
-  using ElementOut = T;
+  using ElementOut = TOut;

  using TileShape = Shape<_128, _128, Shape<_512, _64>>;
  using TileShapeH = cute::tuple_element_t<0, TileShape>;
@ -99,6 +101,7 @@ struct MlaSm100 {
 template <typename T>
 typename T::Fmha::Arguments args_from_options(
    at::Tensor const& out,
+    at::Tensor const& lse,
    at::Tensor const& q_nope,
    at::Tensor const& q_pe,
    at::Tensor const& kv_c_and_k_pe_cache,
@ -162,12 +165,15 @@ typename T::Fmha::Arguments args_from_options(
       stride_PT,
       page_count_total,
       page_size},
-      {static_cast<ElementOut*>(out.data_ptr()), stride_O, static_cast<ElementAcc*>(nullptr), stride_LSE},
+      {static_cast<ElementOut*>(out.data_ptr()),
+       stride_O,
+       static_cast<ElementAcc*>(lse.defined() ? lse.data_ptr() : nullptr),
+       stride_LSE},
      hw_info,
      // 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
@ -178,9 +184,10 @@ typename T::Fmha::Arguments args_from_options(
  return arguments;
 }

-template <typename Element, bool IsPaged128, typename PersistenceOption>
+template <typename Element, typename ElementOut, bool IsPaged128, typename PersistenceOption>
 void runMla(
    at::Tensor const& out,
+    at::Tensor const& lse,
    at::Tensor const& q_nope,
    at::Tensor const& q_pe,
    at::Tensor const& kv_c_and_k_pe_cache,
@ -190,9 +197,9 @@ void runMla(
    double sm_scale,
    int64_t num_kv_splits,
    cudaStream_t stream) {
-  using MlaSm100Type = MlaSm100<Element, IsPaged128, PersistenceOption>;
+  using MlaSm100Type = MlaSm100<Element, ElementOut, IsPaged128, PersistenceOption>;
  typename MlaSm100Type::Fmha fmha;
-  auto arguments = args_from_options<MlaSm100Type>(out, q_nope, q_pe, kv_c_and_k_pe_cache, seq_lens, page_table, sm_scale, num_kv_splits);
+  auto arguments = args_from_options<MlaSm100Type>(out, lse, q_nope, q_pe, kv_c_and_k_pe_cache, seq_lens, page_table, sm_scale, num_kv_splits);

  CUTLASS_CHECK(fmha.can_implement(arguments));

@ -214,6 +221,7 @@ void runMla(

 void sm100_cutlass_mla_decode(
    torch::Tensor const& out,
+    torch::Tensor const& lse,
    torch::Tensor const& q_nope,
    torch::Tensor const& q_pe,
    torch::Tensor const& kv_c_and_k_pe_cache,
@ -233,14 +241,14 @@ void sm100_cutlass_mla_decode(
  DISPATCH_BOOL(page_size == 128, IsPaged128, [&] {
    DISPATCH_BOOL(num_kv_splits <= 1, NotManualSplitKV, [&] {
      if (in_dtype == at::ScalarType::Half) {
-        runMla<cutlass::half_t, IsPaged128, IsPersistent<NotManualSplitKV>>(
-          out, q_nope, q_pe, kv_c_and_k_pe_cache, seq_lens, page_table, workspace, sm_scale, num_kv_splits, stream);
+        runMla<cutlass::half_t, cutlass::half_t, IsPaged128, IsPersistent<NotManualSplitKV>>(
+          out, lse, q_nope, q_pe, kv_c_and_k_pe_cache, seq_lens, page_table, workspace, sm_scale, num_kv_splits, stream);
      } else if (in_dtype == at::ScalarType::BFloat16) {
-        runMla<cutlass::bfloat16_t, IsPaged128, IsPersistent<NotManualSplitKV>>(
-          out, q_nope, q_pe, kv_c_and_k_pe_cache, seq_lens, page_table, workspace, sm_scale, num_kv_splits, stream);
+        runMla<cutlass::bfloat16_t, cutlass::bfloat16_t, IsPaged128, IsPersistent<NotManualSplitKV>>(
+          out, lse, q_nope, q_pe, kv_c_and_k_pe_cache, seq_lens, page_table, workspace, sm_scale, num_kv_splits, stream);
      } else if (in_dtype == at::ScalarType::Float8_e4m3fn) {
-        runMla<cutlass::float_e4m3_t, IsPaged128, IsPersistent<NotManualSplitKV>>(
-          out, q_nope, q_pe, kv_c_and_k_pe_cache, seq_lens, page_table, workspace, sm_scale, num_kv_splits, stream);
+        runMla<cutlass::float_e4m3_t, cutlass::bfloat16_t, IsPaged128, IsPersistent<NotManualSplitKV>>(
+          out, lse, q_nope, q_pe, kv_c_and_k_pe_cache, seq_lens, page_table, workspace, sm_scale, num_kv_splits, stream);
      } else {
        TORCH_CHECK(false, "Unsupported input data type of MLA");
      }
@ -253,7 +261,7 @@ void sm100_cutlass_mla_decode(
 int64_t sm100_cutlass_mla_get_workspace_size(int64_t max_seq_len, int64_t num_batches, int64_t sm_count, int64_t num_kv_splits) {
  // Workspace size depends on ElementAcc and ElementLSE (same as ElementAcc)
  // which are float, so Element type here doesn't matter.
-  using MlaSm100Type = MlaSm100<cutlass::half_t, true>;
+  using MlaSm100Type = MlaSm100<cutlass::half_t, cutlass::half_t, true>;

  // Get split kv. Requires problem shape and sm_count only.
  typename MlaSm100Type::Fmha::Arguments arguments;
@ -264,7 +272,7 @@ 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 +40,19 @@ 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);
+
+// TODO(hc): cp_gather_cache need support scaled kvcahe in the future.
+void cp_gather_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);
--- a/csrc/cache_kernels.cu
+++ b/csrc/cache_kernels.cu
@ -1,6 +1,7 @@
 #include <torch/all.h>
 #include <ATen/cuda/CUDAContext.h>
 #include <c10/cuda/CUDAGuard.h>
+#include <c10/cuda/CUDAException.h>

 #include "cuda_utils.h"
 #include "cuda_compat.h"
@ -8,9 +9,9 @@
 #include "quantization/vectorization_utils.cuh"

 #ifdef USE_ROCM
-  #include "quantization/fp8/amd/quant_utils.cuh"
+  #include "quantization/w8a8/fp8/amd/quant_utils.cuh"
 #else
-  #include "quantization/fp8/nvidia/quant_utils.cuh"
+  #include "quantization/w8a8/fp8/nvidia/quant_utils.cuh"
 #endif

 #include <algorithm>
@ -624,9 +625,9 @@ 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 +635,7 @@ __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 +677,16 @@ __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) {
@ -705,8 +713,144 @@ __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>>>(            \
+// 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_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, 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();
+
+  int32_t block_size = src_cache.size(1);
+  int32_t entry_size = src_cache.flatten(2, -1).size(2);
+
+  TORCH_CHECK(block_table.dtype() == torch::kInt32,
+              "block_table must be int32");
+  TORCH_CHECK(cu_seq_lens.dtype() == torch::kInt32,
+              "cu_seq_lens must be int32");
+  if (seq_starts.has_value()) {
+    TORCH_CHECK(seq_starts.value().dtype() == torch::kInt32,
+                "seq_starts must be int32");
+  }
+
+  TORCH_CHECK(src_cache.device() == dst.device(),
+              "src_cache and dst must be on the same device");
+  TORCH_CHECK(src_cache.device() == block_table.device(),
+              "src_cache and block_table must be on the same device");
+  TORCH_CHECK(src_cache.device() == cu_seq_lens.device(),
+              "src_cache and cu_seq_lens must be on the same device");
+  if (seq_starts.has_value()) {
+    TORCH_CHECK(src_cache.device() == seq_starts.value().device(),
+                "src_cache and seq_starts must be on the same device");
+  }
+
+  int64_t block_table_stride = block_table.stride(0);
+  int64_t cache_block_stride = src_cache.stride(0);
+  int64_t cache_entry_stride = src_cache.stride(1);
+  int64_t dst_entry_stride = dst.stride(0);
+
+  // Decide on the number of splits based on the batch size.
+  int num_splits = batch_size > 128 ? 2 : batch_size > 64 ? 4 : 16;
+  dim3 grid(batch_size, num_splits);
+  dim3 block(1024);
+
+  const int32_t* seq_starts_ptr =
+      seq_starts.has_value() ? seq_starts.value().data_ptr<int32_t>() : nullptr;
+
+  DISPATCH_BY_KV_CACHE_DTYPE(dst.dtype(), kv_cache_dtype, CALL_GATHER_CACHE);
+}
+
+namespace vllm {
+template <typename scalar_t>
+// Note(hc): The cp_gather_cache allows seq_starts to no longer be divisible by
+// block_size.
+__global__ void cp_gather_cache(
+    const scalar_t* __restrict__ src_cache,   // [NUM_BLOCKS, BLOCK_SIZE,
+                                              // ENTRY_SIZE]
+    scalar_t* __restrict__ dst,               // [TOT_TOKENS, ENTRY_SIZE]
+    const int32_t* __restrict__ block_table,  // [BATCH, BLOCK_INDICES]
+    const int32_t* __restrict__ cu_seq_lens,  // [BATCH+1]
+    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 int32_t* __restrict__ seq_starts  // Optional: starting offsets per
+                                            // batch
+) {
+  const int64_t bid = blockIdx.x;  // Batch ID
+  const int32_t num_splits = gridDim.y;
+  const int32_t split = blockIdx.y;
+  const int32_t seq_start = cu_seq_lens[bid];
+  const int32_t seq_end = cu_seq_lens[bid + 1];
+  const int32_t seq_len = seq_end - seq_start;
+  const int32_t tot_slots = seq_len;
+  const int32_t split_slots = cuda_utils::ceil_div(tot_slots, num_splits);
+
+  const int32_t split_start = split * split_slots;
+  const int32_t split_end = min((split + 1) * split_slots, tot_slots);
+
+  const bool is_active_split = (split_start < tot_slots);
+
+  if (!is_active_split) return;
+
+  // Adjust the pointer for the block_table for this batch.
+  // If seq_starts is provided, compute an offset based on it
+  const int32_t batch_offset = bid * block_table_stride;
+  int32_t offset = split_start;
+  if (seq_starts != nullptr) {
+    offset += seq_starts[bid];
+  }
+  int32_t offset_div = offset / block_size;
+  offset = offset % block_size;
+  const int32_t* batch_block_table = block_table + batch_offset;
+
+  // Adjust dst pointer based on the cumulative sequence lengths.
+  dst += seq_start * dst_entry_stride;
+
+  auto copy_entry = [&](const scalar_t* __restrict__ _src,
+                        scalar_t* __restrict__ _dst) {
+    for (int i = threadIdx.x; i < entry_size; i += blockDim.x)
+      _dst[i] = _src[i];
+  };
+
+  for (int pid = split_start; pid < split_end; ++pid) {
+    auto block_id = batch_block_table[offset_div];
+    auto block_start_ptr = src_cache + block_id * cache_block_stride;
+    auto block_dst_ptr = dst + pid * dst_entry_stride;
+    copy_entry(block_start_ptr + offset * cache_entry_stride, block_dst_ptr);
+    offset += 1;
+    // bump to next block
+    if (offset == block_size) {
+      offset_div += 1;
+      offset = 0;
+    }
+  }
+}
+}  // namespace vllm
+
+// Macro to dispatch the kernel based on the data type.
+#define CALL_CP_GATHER_CACHE(CPY_DTYPE)                                 \
+  vllm::cp_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>(), \
@ -716,9 +860,9 @@ __global__ void gather_cache(
 // 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(
+//  - Optionally, seq_starts (if provided) offsets the starting slot index by
+//  seq_starts[bid]
+void cp_gather_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]
@ -769,11 +913,11 @@ void gather_cache(
      seq_starts.has_value() ? seq_starts.value().data_ptr<int32_t>() : nullptr;

  if (dtype_bits == 32) {
-    CALL_GATHER_CACHE(uint32_t);
+    CALL_CP_GATHER_CACHE(uint32_t);
  } else if (dtype_bits == 16) {
-    CALL_GATHER_CACHE(uint16_t);
+    CALL_CP_GATHER_CACHE(uint16_t);
  } else if (dtype_bits == 8) {
-    CALL_GATHER_CACHE(uint8_t);
+    CALL_CP_GATHER_CACHE(uint8_t);
  } else {
    TORCH_CHECK(false, "Unsupported data type width: ", dtype_bits);
  }
--- a/csrc/cpu/cpu_types_vxe.hpp
+++ b/csrc/cpu/cpu_types_vxe.hpp
@ -12,7 +12,7 @@ namespace vec_op {
 #define vec_sub(a, b) ((a) - (b))
 #define vec_mul(a, b) ((a) * (b))
 #define vec_div(a, b) ((a) / (b))
-#define vec_sr(a, b) ((a) >> (b))  // Vector Shift Right Algebaic
+#define vec_sr(a, b) ((a) >> (b))  // Vector Shift Right Algebraic
 #define vec_sl(a, b) ((a) << (b))  // Vector Shift Left

 // FIXME: FP16 is not fully supported in Torch-CPU
--- 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 +1,523 @@
+#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;
+}
+
+DNNLScratchPadManager::DNNLScratchPadManager() : size_(0), ptr_(nullptr) {
+  this->realloc(allocation_unit * 128);
+}
+
+void DNNLScratchPadManager::realloc(size_t new_size) {
+  new_size = round(new_size);
+  if (new_size > size_) {
+    ptr_ = std::aligned_alloc(64, new_size);
+    size_ = new_size;
+  }
+}
+
+DNNLScratchPadManager* DNNLScratchPadManager::get_dnnl_scratchpad_manager() {
+  static DNNLScratchPadManager manager;
+  return &manager;
+}
+
+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));
+  }
+};
+
+template <>
+struct hash<MatMulPrimitiveHandler::ClassMatmulCacheKey> {
+  size_t operator()(
+      const MatMulPrimitiveHandler::ClassMatmulCacheKey& val) const {
+    return hash<dnnl_dim_t>()(val.b_n_size) ^ hash<dnnl_dim_t>()(val.b_k_size);
+  }
+};
+
+template <>
+struct hash<MatMulPrimitiveHandler::MSizeCacheKey> {
+  size_t operator()(const MatMulPrimitiveHandler::MSizeCacheKey& val) const {
+    return hash<dnnl_dim_t>()(val.a_m_size) ^
+           hash<dnnl_dim_t>()(val.a_m_stride) ^ 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;
+}
+
+bool operator==(const MatMulPrimitiveHandler::ClassMatmulCacheKey& l,
+                const MatMulPrimitiveHandler::ClassMatmulCacheKey& r) {
+  return l.b_n_size == r.b_n_size && l.b_k_size == r.b_k_size;
+}
+
+bool operator==(const MatMulPrimitiveHandler::MSizeCacheKey& l,
+                const MatMulPrimitiveHandler::MSizeCacheKey& r) {
+  return l.a_m_size == r.a_m_size && l.a_m_stride == r.a_m_stride &&
+         l.use_bias == r.use_bias && 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);
+
+  auto&& [scratchpad_storage, scratchpad_mem_desc] = get_runtime_memory_ptr(5);
+  scratchpad_storage->set_data_handle(
+      DNNLScratchPadManager::get_dnnl_scratchpad_manager()->get_data<void>());
+
+  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);
+    auto manager = DNNLScratchPadManager::get_dnnl_scratchpad_manager();
+    manager->realloc(desc.scratchpad_desc().get_size());
+    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());
+
+  memory_cache_[DNNL_ARG_SCRATCHPAD] =
+      dnnl::memory({{b_n_size_}, dnnl::memory::data_type::f32, {1}},
+                   default_engine(), nullptr);
+  set_runtime_memory_ptr(5, memory_cache_[DNNL_ARG_SCRATCHPAD].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;
+
+  attr.set_scratchpad_mode(dnnl::scratchpad_mode::user);
+
+  // 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);
+  }
+}
+
+MatMulPrimitiveHandler::MatMulPrimitiveHandler(const Args& args)
+    : DNNLMatMulPrimitiveHandler(
+          static_cast<DNNLMatMulPrimitiveHandler::Args>(args), args.ab_type),
+      m_size_cache_(nullptr) {
+  assert(ab_type_ == dnnl::memory::data_type::f32 ||
+         ab_type_ == dnnl::memory::data_type::bf16 ||
+         ab_type_ == dnnl::memory::data_type::f16);
+  prepack_weight(args.b_ptr,
+                 create_primitive_desc(
+                     MSizeCacheKey{.a_m_size = DNNL_RUNTIME_DIM_VAL,
+                                   .a_m_stride = DNNL_RUNTIME_DIM_VAL,
+                                   .use_bias = false,
+                                   .bias_type = dnnl::memory::data_type::undef},
+                     true)
+                     .weights_desc());
+  init_runtime_memory_cache(args);
+}
+
+static std::shared_ptr<MatMulPrimitiveHandler::MSizeCache>
+get_matul_class_primitive_cache(
+    const MatMulPrimitiveHandler::ClassMatmulCacheKey& key,
+    int64_t cache_size) {
+  static MatMulPrimitiveHandler::ClassMatmulCache cache(128);
+  assert(cache_size > 0);
+  return cache.get_or_create(key, [&]() {
+    return std::make_shared<MatMulPrimitiveHandler::MSizeCache>(cache_size);
+  });
+}
+
+void MatMulPrimitiveHandler::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;
+  a_mem_desc->format_desc.blocking.strides[0] = args.a_m_stride;
+  c_storage->set_data_handle((void*)args.c_ptr);
+  c_mem_desc->dims[0] = args.a_m_size;
+
+  if (args.use_bias) {
+    auto&& [bias_storage, bias_mem_desc] = get_runtime_memory_ptr(2);
+    bias_storage->set_data_handle((void*)args.bias_ptr);
+  }
+
+  dnnl::matmul matmul = get_matmul_cache(args);
+
+  auto&& [scratchpad_storage, scratchpad_mem_desc] = get_runtime_memory_ptr(3);
+  scratchpad_storage->set_data_handle(
+      DNNLScratchPadManager::get_dnnl_scratchpad_manager()->get_data<void>());
+
+  matmul.execute(default_stream(), memory_cache_);
+  default_stream().wait();
+}
+
+dnnl::matmul MatMulPrimitiveHandler::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_};
+    m_size_cache_ = get_matul_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);
+    auto manager = DNNLScratchPadManager::get_dnnl_scratchpad_manager();
+    manager->realloc(desc.scratchpad_desc().get_size());
+    return dnnl::matmul(desc);
+  });
+}
+
+dnnl::matmul::primitive_desc MatMulPrimitiveHandler::create_primitive_desc(
+    const MSizeCacheKey& key, bool first_time) {
+  dnnl::memory::desc a_md;
+  dnnl::memory::desc b_md;
+  if (first_time) {
+    a_md = dnnl::memory::desc({key.a_m_size, b_k_size_}, b_type_,
+                              dnnl::memory::format_tag::ab);
+    b_md = dnnl::memory::desc({b_k_size_, b_n_size_}, b_type_,
+                              dnnl::memory::format_tag::any);
+  } else {
+    a_md = dnnl::memory::desc({key.a_m_size, b_k_size_}, b_type_,
+                              {key.a_m_stride, 1});
+    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;
+  attr.set_scratchpad_mode(dnnl::scratchpad_mode::user);
+
+  if (key.use_bias) {
+    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);
+  }
+}
+
+void MatMulPrimitiveHandler::init_runtime_memory_cache(const Args& args) {
+  memory_cache_[DNNL_ARG_SRC] = dnnl::memory(
+      {{1, b_k_size_}, b_type_, {b_k_size_, 1}}, 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());
+
+  memory_cache_[DNNL_ARG_BIAS] =
+      dnnl::memory({{b_n_size_}, dnnl::memory::data_type::f32, {1}},
+                   default_engine(), nullptr);
+  set_runtime_memory_ptr(2, memory_cache_[DNNL_ARG_BIAS].get());
+
+  memory_cache_[DNNL_ARG_SCRATCHPAD] =
+      dnnl::memory({{b_n_size_}, dnnl::memory::data_type::f32, {1}},
+                   default_engine(), nullptr);
+  set_runtime_memory_ptr(3, memory_cache_[DNNL_ARG_SCRATCHPAD].get());
+}
--- a/csrc/cpu/dnnl_helper.h
+++ b/csrc/cpu/dnnl_helper.h
@ -0,0 +1,243 @@
+#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 DNNLScratchPadManager {
+ public:
+  static constexpr size_t allocation_unit = 4 * 1024 * 1024;  // 4KB
+
+  static DNNLScratchPadManager* get_dnnl_scratchpad_manager();
+
+  DNNLScratchPadManager();
+
+  template <typename T>
+  T* get_data() {
+    return reinterpret_cast<T*>(ptr_);
+  }
+
+  static size_t round(size_t size) {
+    return ((size + allocation_unit - 1) / allocation_unit) * allocation_unit;
+  }
+
+  void realloc(size_t new_size);
+
+ private:
+  size_t size_;
+  void* ptr_;
+};
+
+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_;
+};
+
+class MatMulPrimitiveHandler : public DNNLMatMulPrimitiveHandler {
+ public:
+  struct Args : public DNNLMatMulPrimitiveHandler::Args {
+    dnnl::memory::data_type ab_type;
+  };
+
+  struct ClassMatmulCacheKey {
+    dnnl_dim_t b_n_size;
+    dnnl_dim_t b_k_size;
+
+    friend bool operator==(const ClassMatmulCacheKey& l,
+                           const ClassMatmulCacheKey& r);
+  };
+
+  struct MSizeCacheKey {
+    dnnl_dim_t a_m_size;
+    dnnl_dim_t a_m_stride;
+    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 void* a_ptr;
+    const void* bias_ptr;
+    void* c_ptr;
+  };
+
+ public:
+  MatMulPrimitiveHandler(const Args& args);
+
+  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:
+  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 +1,549 @@
+#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;
+    float azp_val = 0.0f;
+    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.bias_type = get_dnnl_type<void>();
+  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);
+        }
+      });
+}
+
+int64_t create_onednn_mm_handler(const torch::Tensor& b,
+                                 int64_t primitive_cache_size) {
+  TORCH_CHECK(b.dim() == 2);
+
+  MatMulPrimitiveHandler::Args args;
+  args.primitive_cache_size = primitive_cache_size;
+
+  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();
+
+  VLLM_DISPATCH_FLOATING_TYPES(b.scalar_type(), "create_onednn_mm_handler",
+                               [&] {
+                                 args.c_type = get_dnnl_type<scalar_t>();
+                                 args.ab_type = get_dnnl_type<scalar_t>();
+                               });
+
+  return reinterpret_cast<int64_t>(new MatMulPrimitiveHandler(args));
+}
+
+void onednn_mm(torch::Tensor& c,        // [M, OC], row-major
+               const torch::Tensor& a,  // [M, IC], row-major
+               const std::optional<torch::Tensor>& bias, int64_t handler) {
+  CPU_KERNEL_GUARD_IN(onednn_mm)
+  TORCH_CHECK(a.dim() == 2);
+  TORCH_CHECK(a.stride(-1) == 1);
+  TORCH_CHECK(c.is_contiguous());
+  MatMulPrimitiveHandler* ptr =
+      reinterpret_cast<MatMulPrimitiveHandler*>(handler);
+
+  MatMulPrimitiveHandler::ExecArgs exec_args;
+  exec_args.a_m_size = a.size(0);
+  exec_args.a_m_stride = a.stride(0);
+
+  VLLM_DISPATCH_FLOATING_TYPES(a.scalar_type(), "onednn_mm", [&] {
+    if (bias.has_value()) {
+      exec_args.use_bias = true;
+      exec_args.bias_type = get_dnnl_type<scalar_t>();
+      exec_args.bias_ptr = bias->data_ptr<scalar_t>();
+    } else {
+      exec_args.use_bias = false;
+      exec_args.bias_type = get_dnnl_type<void>();
+      exec_args.bias_ptr = nullptr;
+    }
+    exec_args.a_ptr = a.data_ptr<scalar_t>();
+    exec_args.c_ptr = c.data_ptr<scalar_t>();
+
+    ptr->execute(exec_args);
+  });
+}
--- 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/sgl-kernels/moe.cpp
+++ b/csrc/cpu/sgl-kernels/moe.cpp
@ -215,7 +215,7 @@ int moe_align_block_size(
      offsets[mb + 1] = sorted_id_size(sorted_ids + mb * BLOCK_M);
    }
  });
-  // TODO: do we need to vecterize this ?
+  // TODO: do we need to vectorize this ?
  for (int mb = 0; mb < num_token_blocks; ++mb) {
    offsets[mb + 1] += offsets[mb];
  }
--- a/csrc/cpu/torch_bindings.cpp
+++ b/csrc/cpu/torch_bindings.cpp
@ -6,25 +6,26 @@

 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);
+
+int64_t create_onednn_mm_handler(const torch::Tensor& b,
+                                 int64_t primitive_cache_size);
+
+void onednn_mm(torch::Tensor& c, const torch::Tensor& a,
+               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 +152,37 @@ 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 GEMM handler
+  ops.def(
+      "create_onednn_mm_handler(Tensor b, int "
+      "primitive_cache_size) -> int",
+      &create_onednn_mm_handler);
+
+  // oneDNN GEMM
+  ops.def(
+      "onednn_mm(Tensor! c, Tensor a, Tensor? bias, "
+      "int handler) -> ()");
+  ops.impl("onednn_mm", torch::kCPU, &onednn_mm);
+
+  // 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 +198,6 @@ 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/custom_all_reduce.cuh
+++ b/csrc/custom_all_reduce.cuh
@ -15,6 +15,8 @@ typedef __hip_bfloat16 nv_bfloat16;
 #include <map>
 #include <unordered_map>
 #include <vector>
+#include <cstdlib>
+#include <cstring>

 namespace vllm {
 #define CUDACHECK(cmd)                                              \
@ -555,22 +557,47 @@ class CustomAllreduce {
    size /= d;
    auto bytes = size * sizeof(typename packed_t<T>::P);
    int blocks = std::min(block_limit, (size + threads - 1) / threads);
+
+    // Check environment variable once
+    const char* env_algo = std::getenv("VLLM_CUSTOM_ALLREDUCE_ALGO");
+    bool force_1stage = false;
+    bool force_2stage = false;
+    if (env_algo != nullptr) {
+      if (std::strcmp(env_algo, "1stage") == 0 ||
+          std::strcmp(env_algo, "oneshot") == 0) {
+        force_1stage = true;
+      } else if (std::strcmp(env_algo, "2stage") == 0 ||
+                 std::strcmp(env_algo, "twoshot") == 0) {
+        force_2stage = true;
+      } else {
+        throw std::runtime_error(
+            "Invalid VLLM_CUSTOM_ALLREDUCE_ALGO: " + std::string(env_algo) +
+            ". Valid values: 1stage, oneshot, 2stage, twoshot");
+      }
+    }
+
 #define KL(ngpus, name)                                                       \
  name<T, ngpus><<<blocks, threads, 0, stream>>>(ptrs, sg_, self_sg_, output, \
                                                 rank_, size);
-#define REDUCE_CASE(ngpus)                            \
-  case ngpus: {                                       \
-    if (world_size_ == 2) {                           \
-      KL(ngpus, cross_device_reduce_1stage);          \
-    } else if (fully_connected_) {                    \
-      if ((world_size_ <= 4 && bytes < 512 * 1024) || \
-          (world_size_ <= 8 && bytes < 256 * 1024)) { \
-        KL(ngpus, cross_device_reduce_1stage);        \
-      } else {                                        \
-        KL(ngpus, cross_device_reduce_2stage);        \
-      }                                               \
-    }                                                 \
-    break;                                            \
+#define REDUCE_CASE(ngpus)                              \
+  case ngpus: {                                         \
+    if (force_1stage) {                                 \
+      KL(ngpus, cross_device_reduce_1stage);            \
+    } else if (force_2stage) {                          \
+      KL(ngpus, cross_device_reduce_2stage);            \
+    } else {                                            \
+      if (world_size_ == 2) {                           \
+        KL(ngpus, cross_device_reduce_1stage);          \
+      } else if (fully_connected_) {                    \
+        if ((world_size_ <= 4 && bytes < 512 * 1024) || \
+            (world_size_ <= 8 && bytes < 256 * 1024)) { \
+          KL(ngpus, cross_device_reduce_1stage);        \
+        } else {                                        \
+          KL(ngpus, cross_device_reduce_2stage);        \
+        }                                               \
+      }                                                 \
+    }                                                   \
+    break;                                              \
  }

    switch (world_size_) {
--- a/csrc/cutlass_extensions/gemm/collective/collective_builder.hpp
+++ b/csrc/cutlass_extensions/gemm/collective/collective_builder.hpp
@ -1,123 +0,0 @@
-// Modified from: cutlass/gemm/collective/builders/sm90_gmma_builder.inl
-// clang-format off
-#pragma once
-
-#include "cutlass/gemm/collective/builders/sm90_gmma_builder.inl"
-
-#include "cutlass_extensions/gemm/collective/sm90_mma_tma_gmma_ss_warpspecialized_fp8_blockwise_scaling.hpp"
-
-
-/////////////////////////////////////////////////////////////////////////////////////////////////
-
-namespace cutlass::gemm::collective {
-
-/////////////////////////////////////////////////////////////////////////////////////////////////
-
-// GMMA_TMA_WS_SS (BlockScaled Builders)
-template <
-  class ElementA,
-  class GmemLayoutATag,
-  int AlignmentA,
-  class ElementB,
-  class GmemLayoutBTag,
-  int AlignmentB,
-  class ElementAccumulator,
-  class TileShape_MNK,
-  class ClusterShape_MNK,
-  class StageCountType,
-  int ScaleGranularityM
->
-struct CollectiveBuilder<
-    arch::Sm90,
-    arch::OpClassTensorOp,
-    ElementA,
-    GmemLayoutATag,
-    AlignmentA,
-    ElementB,
-    GmemLayoutBTag,
-    AlignmentB,
-    ElementAccumulator,
-    TileShape_MNK,
-    ClusterShape_MNK,
-    StageCountType,
-    KernelTmaWarpSpecializedCooperativeFP8BlockScaledSubGroupMAccum<ScaleGranularityM>,
-    cute::enable_if_t<
-      not detail::is_use_rmem_A<ElementA, GmemLayoutATag, ElementB, GmemLayoutBTag>()>
-> {
-  using KernelScheduleType = KernelTmaWarpSpecializedCooperativeFP8BlockScaledSubGroupMAccum<ScaleGranularityM>;
-
-  static_assert(is_static<TileShape_MNK>::value);
-  static_assert(is_static<ClusterShape_MNK>::value);
-#ifndef CUTLASS_SM90_COLLECTIVE_BUILDER_SUPPORTED
-  static_assert(cutlass::detail::dependent_false<ElementA>, "Unsupported Toolkit for SM90 Collective Builder\n");
-#endif
-  static_assert(detail::is_aligned<ElementA, AlignmentA, ElementB, AlignmentB, detail::tma_alignment_bytes>(),
-                "Should meet TMA alignment requirement\n");
-
-  static constexpr bool IsArrayOfPointersGemm = (cute::is_any_of_v<KernelScheduleType,
-                                                                   KernelPtrArrayTmaWarpSpecializedCooperative,
-                                                                   KernelPtrArrayTmaWarpSpecializedPingpong>);
-  static constexpr bool IsFP8Input = detail::is_input_fp8<ElementA, ElementB>();
-  static_assert((!IsFP8Input || !IsArrayOfPointersGemm),
-                "KernelTmaWarpSpecializedCooperativeFP8BlockScaledAccum is only compatible with FP8 Blocked Scaled version right now.");
-
-  // For fp32 types, map to tf32 MMA value type
-  using ElementAMma = cute::conditional_t<cute::is_same_v<ElementA, float>, tfloat32_t, ElementA>;
-  using ElementBMma = cute::conditional_t<cute::is_same_v<ElementB, float>, tfloat32_t, ElementB>;
-
-  static constexpr cute::GMMA::Major GmmaMajorA = detail::gmma_ss_tag_to_major_A<ElementAMma, GmemLayoutATag>();
-  static constexpr cute::GMMA::Major GmmaMajorB = detail::gmma_ss_tag_to_major_B<ElementBMma, GmemLayoutBTag>();
-
-  static constexpr bool IsCooperative = cute::is_any_of_v<KernelScheduleType,
-                                                          KernelTmaWarpSpecializedCooperative,
-                                                          KernelPtrArrayTmaWarpSpecializedCooperative,
-                                                          KernelTmaWarpSpecializedCooperativeFP8BlockScaledSubGroupMAccum<ScaleGranularityM>>;
-  using AtomLayoutMNK = cute::conditional_t<IsCooperative,
-      Layout<Shape<_2,_1,_1>>, Layout<Shape<_1,_1,_1>>>;
-
-  using TiledMma = decltype(cute::make_tiled_mma(cute::GMMA::ss_op_selector<
-      ElementAMma, ElementBMma, ElementAccumulator, TileShape_MNK, GmmaMajorA, GmmaMajorB>(), AtomLayoutMNK{}));
-
-  using GmemTiledCopyA = decltype(detail::sm90_cluster_shape_to_tma_atom(shape<1>(ClusterShape_MNK{})));
-  using GmemTiledCopyB = decltype(detail::sm90_cluster_shape_to_tma_atom(shape<0>(ClusterShape_MNK{})));
-
-  using SmemLayoutAtomA = decltype(detail::ss_smem_selector<
-      GmmaMajorA, ElementAMma, decltype(cute::get<0>(TileShape_MNK{})), decltype(cute::get<2>(TileShape_MNK{}))>());
-  using SmemLayoutAtomB = decltype(detail::ss_smem_selector<
-      GmmaMajorB, ElementBMma, decltype(cute::get<1>(TileShape_MNK{})), decltype(cute::get<2>(TileShape_MNK{}))>());
-
-  static constexpr size_t TensorMapStorage = IsArrayOfPointersGemm ? sizeof(cute::TmaDescriptor) * 2 /* for A and B */ : 0;
-  static constexpr int KernelSmemCarveout = static_cast<int>(TensorMapStorage);
-
-  static constexpr int PipelineStages = detail::compute_stage_count_or_override<detail::sm90_smem_capacity_bytes - KernelSmemCarveout,
-      ElementAMma, ElementBMma, TileShape_MNK>(StageCountType{});
-  using DispatchPolicy = MainloopSm90TmaGmmaWarpSpecializedBlockScalingSubGroupMFP8<PipelineStages, ClusterShape_MNK, KernelScheduleType, ScaleGranularityM>;
-
-  using SmemCopyAtomA = void;
-  using SmemCopyAtomB = void;
-
-  using CollectiveOp = CollectiveMma<
-      DispatchPolicy,
-      TileShape_MNK,
-      ElementA,
-      TagToStrideA_t<GmemLayoutATag>,
-      ElementB,
-      TagToStrideB_t<GmemLayoutBTag>,
-      TiledMma,
-      GmemTiledCopyA,
-      SmemLayoutAtomA,
-      SmemCopyAtomA,
-      cute::identity,
-      GmemTiledCopyB,
-      SmemLayoutAtomB,
-      SmemCopyAtomB,
-      cute::identity
-    >;
-};
-
-
-/////////////////////////////////////////////////////////////////////////////////////////////////
-
-}  // namespace cutlass::gemm::collective
-
-/////////////////////////////////////////////////////////////////////////////////////////////////
--- a/csrc/cutlass_extensions/gemm/collective/fp8_accumulation.hpp
+++ b/csrc/cutlass_extensions/gemm/collective/fp8_accumulation.hpp
@ -1,183 +0,0 @@
-// clang-format off
-// adapted from: https://github.com/soundOfDestiny/cutlass/blob/a4208aa6958864923505cade9c63eb2a6daf16e5/include/cutlass/gemm/collective/fp8_accumulation.hpp
-
-/***************************************************************************************************
- * Copyright (c) 2023 - 2024 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
- * SPDX-License-Identifier: BSD-3-Clause
- *
- * Redistribution and use in source and binary forms, with or without
- * modification, are permitted provided that the following conditions are met:
- *
- * 1. Redistributions of source code must retain the above copyright notice, this
- * list of conditions and the following disclaimer.
- *
- * 2. Redistributions in binary form must reproduce the above copyright notice,
- * this list of conditions and the following disclaimer in the documentation
- * and/or other materials provided with the distribution.
- *
- * 3. Neither the name of the copyright holder nor the names of its
- * contributors may be used to endorse or promote products derived from
- * this software without specific prior written permission.
- *
- * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
- * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
- * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
- * DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
- * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
- * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
- * SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
- * CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
- * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
- * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
- *
- **************************************************************************************************/
-
-#pragma once
-
-#include "cute/algorithm/clear.hpp"
-#include "cute/tensor.hpp"
-
-//////////////////////////////////////////////////////////////////////////////
-///////////////////////////////////FP8 Accumulation///////////////////////////
-//////////////////////////////////////////////////////////////////////////////
-/// This class provides API to promote (add) or scale (multiply_add) the results
-/// from the tensor core accumulators to the main accumulators when the number 
-/// of MMAs reaches the max number of MMA interval specified by user, after that
-/// the tensor core accumulators are zeroed.
-//////////////////////////////////////////////////////////////////////////////
-
-namespace cutlass::gemm::collective {
-
-template <
-    class EngineAccum,
-    class LayoutAccum>
-struct GmmaFP8AccumulationWithScale {  
-  using TensorAccum = cute::Tensor<EngineAccum, LayoutAccum>;
-  using ElementAccumulator = typename EngineAccum::value_type;
-
-  static_assert(is_static<LayoutAccum>::value, "Accumulator Layout should be static");
-  static_assert(is_rmem<TensorAccum>::value , "Accumulator tensor must be rmem resident.");
-
-private:
-  TensorAccum& accum_;
-  TensorAccum accum_temp_;
-
-  uint32_t accum_promotion_interval_;         // defines the max num of executed MMAs after which accum should be promoted.
-  uint32_t mma_count_per_mainloop_iteration_; // num of MMAs per k_tile of mainloop
-  uint32_t mma_count_;                        // current executed MMAs
-  uint32_t reset_accum_flag_;                 // accum needs to be zeroed or not. 
-
-  // promote or `add` the partial accumulators to main accumulator (FADD).
-  CUTLASS_DEVICE
-  void promote_core() {
-    warpgroup_wait<0>();
-    CUTLASS_PRAGMA_UNROLL
-    for (int i = 0; i < size(accum_); ++i) {
-      accum_(i) += accum_temp_(i);
-    }
-  }
-
-  // `multiply` scale the partial accumulators and `add` to main accumulator (FFMA).
-  template <
-    class EngineScale,
-    class LayoutScale>
-  CUTLASS_DEVICE
-  void scale_core(const cute::Tensor<EngineScale, LayoutScale> &scale) {
-    using TensorScale = cute::Tensor<EngineScale, LayoutScale>;
-
-    static_assert(is_static<LayoutScale>::value, "Scale Layout should be static");
-    static_assert(is_rmem<TensorScale>::value , "Scale tensor must be rmem resident.");
-
-    static_assert(LayoutAccum{}.shape() == LayoutScale{}.shape(), "Accumulator and scale must have same shape.");
-
-    warpgroup_wait<0>();
-    CUTLASS_PRAGMA_UNROLL
-    for (int i = 0; i < size(accum_); ++i) {
-      accum_(i) += accum_temp_(i) * scale(i);
-    }
-  }
-
-public:
-  CUTLASS_DEVICE
-  GmmaFP8AccumulationWithScale(
-      TensorAccum &accum,
-      uint32_t accum_promotion_interval,
-      uint32_t mma_count_per_mainloop_iteration)
-      : accum_(accum), 
-        accum_promotion_interval_(accum_promotion_interval),
-        mma_count_per_mainloop_iteration_(mma_count_per_mainloop_iteration),
-        mma_count_(0), 
-        reset_accum_flag_(0) 
-  {
-    accum_temp_ = cute::make_fragment_like(accum);
-  }
-
-  //
-  // Methods (Common)
-  //
-
-  CUTLASS_DEVICE 
-  TensorAccum& operator()() {
-    return accum_temp_;
-  }
-
-  /// prepare the MMA accumulators when initialization or zeroing is required.
-  CUTLASS_DEVICE
-  bool prepare_if_needed() { 
-    return reset_accum_flag_;
-  }
-
-  //
-  // Methods (for FADD version)
-  //
-
-  /// promote (add) the results from the MMA accumulators to main accumulator if needed.
-  CUTLASS_DEVICE
-  void promote_if_needed() {
-    mma_count_ += mma_count_per_mainloop_iteration_;
-    reset_accum_flag_ = __shfl_sync(0xffffffff, mma_count_ == accum_promotion_interval_, 0);
-    if (reset_accum_flag_) {
-      promote_core();
-      mma_count_ = 0;
-    }
-  }
-
-  /// promote (add) the residue results from the MMA accumulators to main accumulator if needed.
-  CUTLASS_DEVICE
-  void promote_residue_if_needed() {
-    if (__shfl_sync(0xffffffff, mma_count_ > 0, 0)) {
-      promote_core();
-    }
-  }
-
-  //
-  // Methods (for FFMA version)
-  //
-
-  /// scale (multiply_add) the results from the MMA accumulators to main accumulator if needed.
-  template <
-    class EngineScale,
-    class LayoutScale>
-  CUTLASS_DEVICE
-  void scale_if_needed(const cute::Tensor<EngineScale, LayoutScale> &scale) {
-    mma_count_ += mma_count_per_mainloop_iteration_;
-    reset_accum_flag_ = __shfl_sync(0xffffffff, mma_count_ == accum_promotion_interval_, 0);
-    if (reset_accum_flag_) {
-      scale_core(scale);
-      mma_count_ = 0;
-    }
-  }
-
-  /// scale (multiply_add) the residue results from the MMA accumulators to main accumulator if needed.
-  template <
-    class EngineScale,
-    class LayoutScale>
-  CUTLASS_DEVICE
-  void scale_residue_if_needed(const cute::Tensor<EngineScale, LayoutScale> &scale) {
-    if (__shfl_sync(0xffffffff, mma_count_ > 0, 0)) {
-      scale_core(scale);
-    }
-  }
-};
-
-} // namespace cutlass::gemm::collective
--- a/csrc/cutlass_extensions/gemm/collective/sm90_mma_tma_gmma_ss_warpspecialized_fp8_blockwise_scaling.hpp
+++ b/csrc/cutlass_extensions/gemm/collective/sm90_mma_tma_gmma_ss_warpspecialized_fp8_blockwise_scaling.hpp
@ -1,729 +0,0 @@
-// clang-format off
-// Adapted (Heavily) from: https://github.com/soundOfDestiny/cutlass/blob/9d997ce0dea4c5fa1a617db6b7ff29aa9235822c/include/cutlass/gemm/collective/sm90_mma_tma_gmma_ss_warpspecialized_fp8_blockwise_scaling.hpp
-
-/***************************************************************************************************
- * Copyright (c) 2023 - 2024 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
- * SPDX-License-Identifier: BSD-3-Clause
- *
- * Redistribution and use in source and binary forms, with or without
- * modification, are permitted provided that the following conditions are met:
- *
- * 1. Redistributions of source code must retain the above copyright notice, this
- * list of conditions and the following disclaimer.
- *
- * 2. Redistributions in binary form must reproduce the above copyright notice,
- * this list of conditions and the following disclaimer in the documentation
- * and/or other materials provided with the distribution.
- *
- * 3. Neither the name of the copyright holder nor the names of its
- * contributors may be used to endorse or promote products derived from
- * this software without specific prior written permission.
- *
- * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
- * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
- * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
- * DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
- * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
- * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
- * SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
- * CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
- * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
- * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
- *
- **************************************************************************************************/
-
-#pragma once
-
-#include "cutlass/cutlass.h"
-#include "cutlass/gemm/dispatch_policy.hpp"
-#include "cutlass/trace.h"
-#include "cutlass/numeric_types.h"
-
-#include "cute/arch/cluster_sm90.hpp"
-#include "cute/arch/copy_sm80.hpp"
-#include "cute/arch/copy_sm90.hpp"
-#include "cute/algorithm/functional.hpp"
-#include "cute/atom/mma_atom.hpp"
-#include "cute/algorithm/gemm.hpp"
-#include "cute/numeric/arithmetic_tuple.hpp"
-
-#include "cutlass_extensions/gemm/dispatch_policy.hpp"
-#include "cutlass_extensions/gemm/collective/fp8_accumulation.hpp"
-
-/////////////////////////////////////////////////////////////////////////////////////////////////
-
-namespace cutlass::gemm::collective {
-using namespace cute;
-
-/////////////////////////////////////////////////////////////////////////////////////////////////
-
-// WarpSpecialized Mainloop
-template <
-  int Stages,
-  class ClusterShape,
-  class KernelSchedule,
-  int ScaleGranularityM_,
-  class TileShape_,
-  class ElementA_,
-  class StrideA_,
-  class ElementB_,
-  class StrideB_,
-  class TiledMma_,
-  class GmemTiledCopyA_,
-  class SmemLayoutAtomA_,
-  class SmemCopyAtomA_,
-  class TransformA_,
-  class GmemTiledCopyB_,
-  class SmemLayoutAtomB_,
-  class SmemCopyAtomB_,
-  class TransformB_>
-struct CollectiveMma<
-    MainloopSm90TmaGmmaWarpSpecializedBlockScalingSubGroupMFP8<Stages, ClusterShape, KernelSchedule, ScaleGranularityM_>,
-    TileShape_,
-    ElementA_,
-    StrideA_,
-    ElementB_,
-    StrideB_,
-    TiledMma_,
-    GmemTiledCopyA_,
-    SmemLayoutAtomA_,
-    SmemCopyAtomA_,
-    TransformA_,
-    GmemTiledCopyB_,
-    SmemLayoutAtomB_,
-    SmemCopyAtomB_,
-    TransformB_>
-{
-  //
-  // Type Aliases
-  //
-  using DispatchPolicy = MainloopSm90TmaGmmaWarpSpecializedBlockScalingSubGroupMFP8<Stages, ClusterShape, KernelSchedule, ScaleGranularityM_>;
-  using TileShape = TileShape_;
-  using ElementA = ElementA_;
-  using StrideA = StrideA_;
-  using ElementB = ElementB_;
-  using StrideB = StrideB_;
-  using TiledMma = TiledMma_;
-  using ElementAccumulator = typename TiledMma::ValTypeC;
-  using ElementBlockScale = ElementAccumulator;
-  using GmemTiledCopyA = GmemTiledCopyA_;
-  using GmemTiledCopyB = GmemTiledCopyB_;
-  using SmemLayoutAtomA = SmemLayoutAtomA_;
-  using SmemLayoutAtomB = SmemLayoutAtomB_;
-  using SmemCopyAtomA = SmemCopyAtomA_;
-  using SmemCopyAtomB = SmemCopyAtomB_;
-  using TransformA = TransformA_;
-  using TransformB = TransformB_;
-  using ArchTag = typename DispatchPolicy::ArchTag;
-
-  using CtaShape_MNK = decltype(shape_div(TileShape{}, ClusterShape{}));
-  using MainloopPipeline = cutlass::PipelineTmaAsync<DispatchPolicy::Stages>;
-  using PipelineState = cutlass::PipelineState<DispatchPolicy::Stages>;
-  using PipelineParams = typename MainloopPipeline::Params;
-
-  // Two threads per CTA are producers (1 for operand tile and 32 for scales)
-  static constexpr int NumProducerThreadEvents = 33; 
-
-  static constexpr int ScaleGranularityM = ScaleGranularityM_ == 0 ? size<0>(TileShape{}) : ScaleGranularityM_;
-  static constexpr int ScaleMsPerTile = size<0>(TileShape{}) / ScaleGranularityM;
-
-  static_assert(cute::rank(SmemLayoutAtomA{}) == 2, "SmemLayoutAtom must be rank 2 (M/N, K)");
-  static_assert((size<0>(TileShape{}) % size<0>(SmemLayoutAtomA{})) == 0, "SmemLayoutAtom must evenly divide tile shape.");
-  static_assert((size<2>(TileShape{}) % size<1>(SmemLayoutAtomA{})) == 0, "SmemLayoutAtom must evenly divide tile shape.");
-
-  static_assert(cute::rank(SmemLayoutAtomB{}) == 2, "SmemLayoutAtom must be rank 2 (M/N, K)");
-  static_assert((size<1>(TileShape{}) % size<0>(SmemLayoutAtomB{})) == 0, "SmemLayoutAtom must evenly divide tile shape.");
-  static_assert((size<2>(TileShape{}) % size<1>(SmemLayoutAtomB{})) == 0, "SmemLayoutAtom must evenly divide tile shape.");
-
-  static_assert((size<0>(TileShape{}) % ScaleGranularityM) == 0, "FP8 scaling granularity must evenly divide tile shape along M.");
-
-  // Tile along modes in a way that maximizes the TMA box size.
-  using SmemLayoutA = decltype(tile_to_shape(
-      SmemLayoutAtomA{},
-      make_shape(shape<0>(TileShape{}), shape<2>(TileShape{}), Int<DispatchPolicy::Stages>{}),
-      cute::conditional_t< ::cutlass::gemm::detail::is_major<0,StrideA>(), Step<_2,_1,_3>, Step<_1,_2,_3>>{}));
-  using SmemLayoutB = decltype(tile_to_shape(
-      SmemLayoutAtomB{},
-      make_shape(shape<1>(TileShape{}), shape<2>(TileShape{}), Int<DispatchPolicy::Stages>{}),
-      cute::conditional_t< ::cutlass::gemm::detail::is_major<0,StrideB>(), Step<_2,_1,_3>, Step<_1,_2,_3>>{}));
-  
-  // Block scaling gmem-to-smem copy atom 
-  using SmemBlockScalingCopyAtomA = Copy_Atom<SM80_CP_ASYNC_CACHEALWAYS<ElementBlockScale>, ElementBlockScale>;
-  using SmemBlockScalingCopyAtomB = Copy_Atom<SM80_CP_ASYNC_CACHEALWAYS<ElementBlockScale>, ElementBlockScale>;
-  
-  // Block scaling smem layout
-  using SmemLayoutScaleA = Layout<Shape<Int<ScaleMsPerTile>, Int<DispatchPolicy::Stages>>>;
-  using SmemLayoutScaleB = Layout<Shape<Int<DispatchPolicy::Stages>>, Stride<_1>>; // `ScaleNsPerTile` is always 1.
-
-  static_assert(DispatchPolicy::Stages >= 2, "Specialization requires Stages set to value 1 or more.");
-  static_assert(cute::is_base_of<cute::GMMA::DescriptorIterator, typename TiledMma::FrgTypeA>::value &&
-                cute::is_base_of<cute::GMMA::DescriptorIterator, typename TiledMma::FrgTypeB>::value,
-                "MMA atom must source both A and B operand from smem_desc for this mainloop.");
-  static_assert(cute::is_same_v<GmemTiledCopyA, SM90_TMA_LOAD> || cute::is_same_v<GmemTiledCopyA, SM90_TMA_LOAD_MULTICAST>,
-      "GmemTiledCopy - invalid SM90 TMA copy atom specified.");
-  static_assert(cute::is_same_v<GmemTiledCopyB, SM90_TMA_LOAD> || cute::is_same_v<GmemTiledCopyB, SM90_TMA_LOAD_MULTICAST>,
-      "GmemTiledCopy - invalid SM90 TMA copy atom specified.");
-  static_assert(cute::is_same_v<ElementAccumulator, ElementBlockScale>,
-             "ElementAccumulator and ElementBlockScale should be same datatype");
-
-  struct SharedStorage
-  {
-    struct TensorStorage : cute::aligned_struct<128> {
-      cute::array_aligned<typename TiledMma::ValTypeA, cute::cosize_v<SmemLayoutA>> smem_A;  // mxk
-      cute::array_aligned<typename TiledMma::ValTypeB, cute::cosize_v<SmemLayoutB>> smem_B;  // nxk
-      cute::array_aligned<ElementBlockScale, cute::cosize_v<SmemLayoutScaleA>> smem_scale_A; // ScaleMsPerTile x k
-      cute::array_aligned<ElementBlockScale, cute::cosize_v<SmemLayoutScaleB>> smem_scale_B; // 1xk
-    } tensors;
-
-    using PipelineStorage = typename MainloopPipeline::SharedStorage;
-    PipelineStorage pipeline;
-  };
-  using TensorStorage = typename SharedStorage::TensorStorage;
-  using PipelineStorage = typename SharedStorage::PipelineStorage;
-
-  // Host side kernel arguments
-  struct Arguments {
-    ElementA const* ptr_A;
-    StrideA dA;
-    ElementB const* ptr_B;
-    StrideB dB;
-    ElementBlockScale const* ptr_scale_A; 
-    ElementBlockScale const* ptr_scale_B;
-  };
-
-  // Device side kernel params
-  struct Params {
-    // Assumption: StrideA is congruent with Problem_MK
-    using TMA_A = decltype(make_tma_copy_A_sm90(
-        GmemTiledCopyA{},
-        make_tensor(static_cast<ElementA const*>(nullptr), repeat_like(StrideA{}, int32_t(0)), StrideA{}),
-        SmemLayoutA{}(_,_,0),
-        TileShape{},
-        ClusterShape{}));
-    // Assumption: StrideB is congruent with Problem_NK
-    using TMA_B = decltype(make_tma_copy_B_sm90(
-        GmemTiledCopyB{},
-        make_tensor(static_cast<ElementB const*>(nullptr), repeat_like(StrideB{}, int32_t(0)), StrideB{}),
-        SmemLayoutB{}(_,_,0),
-        TileShape{},
-        ClusterShape{}));
-    TMA_A tma_load_a;
-    TMA_B tma_load_b;
-    uint32_t tma_transaction_bytes = TmaTransactionBytes;
-    uint32_t tma_transaction_bytes_mk = TmaTransactionBytesMK;
-    uint32_t tma_transaction_bytes_nk = TmaTransactionBytesNK;
-    // Block scaling factors for A and B
-    ElementBlockScale const* ptr_scale_A; 
-    ElementBlockScale const* ptr_scale_B;
-  };
-
-  //
-  // Methods
-  //
-
-  template <class ProblemShape>
-  static constexpr Params
-  to_underlying_arguments(ProblemShape const& problem_shape, Arguments const& args, void* workspace) {
-    (void) workspace;
-
-    // Optionally append 1s until problem shape is rank-4 (MNKL), in case it is only rank-3 (MNK)
-    auto problem_shape_MNKL = append<4>(problem_shape, 1);
-    auto [M,N,K,L] = problem_shape_MNKL;
-
-    auto ptr_A = reinterpret_cast<ElementA const*>(args.ptr_A);
-    auto ptr_B = reinterpret_cast<ElementB const*>(args.ptr_B);
-
-    Tensor tensor_a = make_tensor(ptr_A, make_layout(make_shape(M,K,L), args.dA));
-    Tensor tensor_b = make_tensor(ptr_B, make_layout(make_shape(N,K,L), args.dB));
-    typename Params::TMA_A tma_load_a = make_tma_copy_A_sm90(
-        GmemTiledCopyA{},
-        tensor_a,
-        SmemLayoutA{}(_,_,cute::Int<0>{}),
-        TileShape{},
-        ClusterShape{});
-    typename Params::TMA_B tma_load_b = make_tma_copy_B_sm90(
-        GmemTiledCopyB{},
-        tensor_b,
-        SmemLayoutB{}(_,_,cute::Int<0>{}),
-        TileShape{},
-        ClusterShape{});
-    uint32_t transaction_bytes_mk = TmaTransactionBytesMK;
-    uint32_t transaction_bytes_nk = TmaTransactionBytesNK;
-    uint32_t transaction_bytes = transaction_bytes_mk + transaction_bytes_nk;
-
-    return {
-      tma_load_a,
-      tma_load_b,
-      transaction_bytes,
-      transaction_bytes_mk,
-      transaction_bytes_nk,
-      args.ptr_scale_A,
-      args.ptr_scale_B
-    };
-  }
-
-  template<class ProblemShape>
-  static bool
-  can_implement(
-      ProblemShape const& problem_shape,
-      [[maybe_unused]] Arguments const& args) {
-    constexpr int tma_alignment_bits = 128;
-    auto problem_shape_MNKL = append<4>(problem_shape, 1);
-    auto [M,N,K,L] = problem_shape_MNKL;
-    
-    bool implementable = true;
-    constexpr int min_tma_aligned_elements_A = tma_alignment_bits / cutlass::sizeof_bits<ElementA>::value;
-    implementable = implementable && cutlass::detail::check_alignment<min_tma_aligned_elements_A>(cute::make_shape(M,K,L), StrideA{});
-    constexpr int min_tma_aligned_elements_B = tma_alignment_bits / cutlass::sizeof_bits<ElementB>::value;
-    implementable = implementable && cutlass::detail::check_alignment<min_tma_aligned_elements_B>(cute::make_shape(N,K,L), StrideB{});
-
-    if (!implementable) {
-      CUTLASS_TRACE_HOST("  CAN IMPLEMENT: Problem Size doesn't meet the minimum alignment requirements for TMA.\n");
-    }
-    return implementable;
-  }
-
-  static constexpr int K_PIPE_MAX = DispatchPolicy::Stages;
-  static constexpr int K_PIPE_MMAS = 1;
-  static constexpr uint32_t TmaTransactionBytesMK =
-        cutlass::bits_to_bytes(size<0>(SmemLayoutA{}) * size<1>(SmemLayoutA{}) * static_cast<uint32_t>(sizeof_bits<ElementA>::value));
-  static constexpr uint32_t TmaTransactionBytesNK =
-        cutlass::bits_to_bytes(size<0>(SmemLayoutB{}) * size<1>(SmemLayoutB{}) * static_cast<uint32_t>(sizeof_bits<ElementB>::value));
-  static constexpr uint32_t TmaTransactionBytes = TmaTransactionBytesMK + TmaTransactionBytesNK;
-
-  /// Issue Tma Descriptor Prefetch -- ideally from a single thread for best performance
-  CUTLASS_DEVICE
-  static void prefetch_tma_descriptors(Params const& mainloop_params)
-  {
-    cute::prefetch_tma_descriptor(mainloop_params.tma_load_a.get_tma_descriptor());
-    cute::prefetch_tma_descriptor(mainloop_params.tma_load_b.get_tma_descriptor());
-  }
-
-  /// Set up the data needed by this collective for load and mma.
-  /// Returns a tuple of tensors. The collective and the kernel layer have the contract
-  /// Returned tuple must contain at least two elements, with the first two elements being:
-  /// gA_mkl - The tma tensor, A after a local tile so it has shape  (BLK_M,BLK_K,m,k,l)
-  /// gB_nkl - The tma tensor, B after a local tile so it has shape  (BLK_N,BLK_K,n,k,l)
-  template <class ProblemShape_MNKL>
-  CUTLASS_DEVICE auto
-  load_init(ProblemShape_MNKL const& problem_shape_MNKL, Params const& mainloop_params) const {
-    using X = Underscore;
-    // Separate out problem shape for convenience
-    auto [M,N,K,L] = problem_shape_MNKL;
-
-    // TMA requires special handling of strides to deal with coord codomain mapping
-    // Represent the full tensors -- get these from TMA
-    Tensor mA_mkl = mainloop_params.tma_load_a.get_tma_tensor(make_shape(M,K,L));                            // (m,k,l)
-    Tensor mB_nkl = mainloop_params.tma_load_b.get_tma_tensor(make_shape(N,K,L));                            // (n,k,l)
-
-    // Make tiled views, defer the slice
-    Tensor gA_mkl = local_tile(mA_mkl, TileShape{}, make_coord(_,_,_), Step<_1, X,_1>{});        // (BLK_M,BLK_K,m,k,l)
-    Tensor gB_nkl = local_tile(mB_nkl, TileShape{}, make_coord(_,_,_), Step< X,_1,_1>{});        // (BLK_N,BLK_K,n,k,l)
-
-    constexpr auto scales_m = Int<ScaleMsPerTile>{};
-    auto tM = get<2>(gA_mkl.shape());
-    auto tN = get<2>(gB_nkl.shape());
-    auto tK = get<3>(gA_mkl.shape());
-
-    // Make the tiled views of scale tensors
-    auto scaleA_shape = make_shape(M / ScaleGranularityM, tK, L); // (scale_m,k,l)
-    auto scaleA_layout = make_ordered_layout(scaleA_shape,  Step<_0, _1, _2>{});
-    auto scaleB_shape = make_shape(tN, tK, L); // (n,k,l)
-    auto scaleB_layout = make_ordered_layout(scaleB_shape, Step<_1, _0, _2>{});
-
-    // Note that mScaleA_mkl and mScaleB_nkl are already blocked tiled in the `m` host and 
-    // gScaleA_mkl and gScaleB_nkl in `g` global memory are same as mScaleA_mkl and mScaleB_nkl.
-    Tensor mScaleA_mkl = make_tensor(make_gmem_ptr(mainloop_params.ptr_scale_A), scaleA_layout); // (scale_m,k,l)
-    Tensor mScaleB_nkl = make_tensor(make_gmem_ptr(mainloop_params.ptr_scale_B), scaleB_layout); // (n,k,l)
-
-    return cute::make_tuple(gA_mkl, gB_nkl, mScaleA_mkl, mScaleB_nkl);
-  }
-
-  /// Perform a collective-scoped matrix multiply-accumulate
-  /// Producer Perspective
-  template <
-    class TensorA, class TensorB,
-    class TensorScaleA, class TensorScaleB,
-    class KTileIterator, class BlockCoord
-  >
-  CUTLASS_DEVICE void
-  load(
-      Params const& mainloop_params,
-      MainloopPipeline pipeline,
-      PipelineState smem_pipe_write,
-      cute::tuple<TensorA, TensorB, TensorScaleA, TensorScaleB> const& load_inputs,
-      BlockCoord const& blk_coord,
-      KTileIterator k_tile_iter, int k_tile_count,
-      int thread_idx,
-      uint32_t block_rank_in_cluster,
-      TensorStorage& shared_tensors) {
-    int lane_predicate = cute::elect_one_sync();
-
-    // Blockscaling: Tma loads for load_input and CpAsync for load_scale
-    Tensor sA = make_tensor(make_smem_ptr(shared_tensors.smem_A.data()), SmemLayoutA{});        // (BLK_M,BLK_K,PIPE)
-    Tensor sB = make_tensor(make_smem_ptr(shared_tensors.smem_B.data()), SmemLayoutB{});        // (BLK_N,BLK_K,PIPE)
-    Tensor sScaleA = make_tensor(cute::make_smem_ptr(shared_tensors.smem_scale_A.data()), SmemLayoutScaleA{}); // (ScaleMsPerTile,k)
-    Tensor sScaleB = make_tensor(cute::make_smem_ptr(shared_tensors.smem_scale_B.data()), SmemLayoutScaleB{}); // (k)
-
-    //
-    // Prepare the TMA loads for A and B
-    //
-
-    constexpr uint32_t cluster_shape_x = get<0>(ClusterShape());
-    uint2 cluster_local_block_id = {block_rank_in_cluster % cluster_shape_x, block_rank_in_cluster / cluster_shape_x};
-
-    Tensor gA_mkl = get<0>(load_inputs);
-    Tensor gB_nkl = get<1>(load_inputs);
-
-    auto block_tma_a = mainloop_params.tma_load_a.get_slice(cluster_local_block_id.y);
-    auto block_tma_b = mainloop_params.tma_load_b.get_slice(cluster_local_block_id.x);
-
-    // Partition the inputs based on the current block coordinates.
-    auto [m_coord, n_coord, k_coord, l_coord] = blk_coord;
-    Tensor gA = gA_mkl(_,_,m_coord,_,l_coord);                                                     // (BLK_M,BLK_K,k)
-    Tensor gB = gB_nkl(_,_,n_coord,_,l_coord);                                                     // (BLK_N,BLK_K,k)
-
-
-    // Block scaling: load_scale has scaling tensors in global memory which are not tiled
-    Tensor mScaleA_mkl = get<2>(load_inputs);
-    Tensor mScaleB_nkl = get<3>(load_inputs);
-    auto scales_m = get<0>(mScaleA_mkl.shape());
-
-    Tensor cScaleA_mkl = make_identity_tensor(mScaleA_mkl.shape());
-
-    Tensor gScaleA = local_tile( 
-      mScaleA_mkl, make_tile(Int<ScaleMsPerTile>{}), 
-      make_coord(m_coord,_,l_coord));                   // (ScaleMsPerTile,k,1)
-    Tensor cScaleA = local_tile( 
-      cScaleA_mkl, make_tile(Int<ScaleMsPerTile>{}), 
-      make_coord(m_coord,_,l_coord));
-    Tensor gScaleB = mScaleB_nkl(n_coord,_,l_coord);                                           // (1,k,1)
-
-    // TODO: test `scale_copy_a` with `ScaleMsPerTile` < 128
-    TiledCopy scale_copy_a = make_tiled_copy(SmemBlockScalingCopyAtomA{}, 
-      Layout<Shape<_32>>{}, Layout<Shape<_1>>{}); // (1,1,1)
-    TiledCopy scale_copy_b = make_tiled_copy(SmemBlockScalingCopyAtomB{}, 
-      Layout<Shape<_1>>{}, Layout<Shape<_1>>{}); // (1,1,1)
-    ThrCopy thr_scale_copy_a = scale_copy_a.get_slice(threadIdx.x);
-    ThrCopy thr_scale_copy_b = scale_copy_b.get_slice(threadIdx.x);
-    
-    Tensor tAgA_ScaleA = thr_scale_copy_a.partition_S(gScaleA);
-    Tensor tAcA_ScaleA = thr_scale_copy_a.partition_S(cScaleA);
-    Tensor tAsA_ScaleA = thr_scale_copy_a.partition_D(sScaleA);
-    
-    Tensor tBgB_ScaleB = thr_scale_copy_b.partition_S(gScaleB);
-    Tensor tBsB_ScaleB = thr_scale_copy_b.partition_D(sScaleB);
-
-    // Applies the mapping from block_tma_a
-    Tensor tAgA = block_tma_a.partition_S(gA);                                              // (TMA,TMA_M,TMA_K,k)
-    Tensor tAsA = block_tma_a.partition_D(sA);                                              // (TMA,TMA_M,TMA_K,PIPE)
-
-    Tensor tBgB = block_tma_b.partition_S(gB);                                              // (TMA,TMA_N,TMA_K,k)
-    Tensor tBsB = block_tma_b.partition_D(sB);                                              // (TMA,TMA_N,TMA_K,PIPE)
-
-    uint16_t mcast_mask_a = 0;
-    uint16_t mcast_mask_b = 0;
-
-    // Issue TmaLoads for GEMM operands A/B and CpAsync for scale tensors
-    // Maps the tile -> block, value
-    if constexpr (cute::is_same_v<GmemTiledCopyA, SM90_TMA_LOAD_MULTICAST>) {
-      auto block_layout = Layout<typename DispatchPolicy::ClusterShape>{};                       // (m,n) -> block_id
-      for (int n = 0; n < size<1>(block_layout); ++n) {
-        mcast_mask_a |= (uint16_t(1) << block_layout(cluster_local_block_id.x,n,Int<0>{}));
-      }
-    }
-
-    if constexpr (cute::is_same_v<GmemTiledCopyB, SM90_TMA_LOAD_MULTICAST>) {
-      auto block_layout = Layout<typename DispatchPolicy::ClusterShape>{};                       // (m,n) -> block_id
-      for (int m = 0; m < size<0>(block_layout); ++m) {
-        mcast_mask_b |= (uint16_t(1) << block_layout(m,cluster_local_block_id.y,Int<0>{}));
-      }
-    }
-
-    // Allocate predicate tensors for a_scales (since we can't guarantee that 
-    // all scales are valid, since we could have a partial tiles along M)
-    Tensor tApA_ScaleA = make_tensor<bool>(shape(tAsA_ScaleA(_,_,0)));
-    #pragma unroll
-    for (int i = 0; i < size(tApA_ScaleA); ++i) {
-      tApA_ScaleA(i) = get<0>(tAcA_ScaleA(i)) < scales_m;
-    }
-
-    // Mainloop
-    CUTLASS_PRAGMA_NO_UNROLL
-    for ( ; k_tile_count > 0; --k_tile_count) {
-      // LOCK smem_pipe_write for _writing_
-      pipeline.producer_acquire(smem_pipe_write);
-
-      //
-      // Copy gmem to smem for *k_tile_iter
-      //
-      int write_stage = smem_pipe_write.index();
-      using BarrierType = typename MainloopPipeline::ProducerBarrierType;
-      BarrierType* tma_barrier = pipeline.producer_get_barrier(smem_pipe_write);
-
-      // Copy operands A and B from global memory to shared memory
-      if (lane_predicate) copy(mainloop_params.tma_load_a.with(*tma_barrier, mcast_mask_a), tAgA(_,_,_,*k_tile_iter), tAsA(_,_,_,write_stage));
-      if (lane_predicate) copy(mainloop_params.tma_load_b.with(*tma_barrier, mcast_mask_b), tBgB(_,_,_,*k_tile_iter), tBsB(_,_,_,write_stage));
-
-      // Copy scale tensors from global memory to shared memory
-      copy_if(scale_copy_a, tApA_ScaleA, tAgA_ScaleA(_,_,*k_tile_iter), tAsA_ScaleA(_,_,write_stage));
-      copy(scale_copy_b, tBgB_ScaleB(_,*k_tile_iter), tBsB_ScaleB(_,write_stage));
-      pipeline.producer_commit(smem_pipe_write, cutlass::arch::cpasync_barrier_arrive_noinc);
-
-      ++k_tile_iter;
-
-      // Advance smem_pipe_write
-      ++smem_pipe_write;
-    }
-  }
-
-  /// Perform a Producer Epilogue to prevent early exit of blocks in a Cluster
-  CUTLASS_DEVICE void
-  load_tail(
-      MainloopPipeline pipeline,
-      PipelineState smem_pipe_write) {
-    int lane_predicate = cute::elect_one_sync();
-
-    // Issue the epilogue waits
-    if (lane_predicate) {
-      /* This helps avoid early exit of blocks in Cluster
-       * Waits for all stages to either be released (all
-       * Consumer UNLOCKs), or if the stage was never used
-       * then would just be acquired since the phase was
-       * still inverted from make_producer_start_state
-       */
-      pipeline.producer_tail(smem_pipe_write);
-    }
-  }
-
-  /// Perform a collective-scoped matrix multiply-accumulate
-  /// Consumer Perspective
-  template <
-    class FrgTensorC
-  >
-  CUTLASS_DEVICE void
-  mma(MainloopPipeline pipeline,
-      PipelineState smem_pipe_read,
-      FrgTensorC& accum,
-      int k_tile_count,
-      int thread_idx,
-      TensorStorage& shared_tensors,
-      Params const& mainloop_params) {
-
-
-    static_assert(is_rmem<FrgTensorC>::value, "C tensor must be rmem resident.");
-    static_assert(cute::rank(SmemLayoutA{}) == 3, "Smem layout must be rank 3.");
-    static_assert(cute::rank(SmemLayoutB{}) == 3, "Smem layout must be rank 3.");
-    static_assert(cute::is_void_v<SmemCopyAtomA>,
-      "SM90 GMMA mainloops cannot have a non-void copy atom for smem sourced instructions.");
-    static_assert(cute::is_void_v<SmemCopyAtomB>,
-      "SM90 GMMA mainloops cannot have a non-void copy atom for smem sourced instructions.");
-
-    Tensor sA = make_tensor(make_smem_ptr(shared_tensors.smem_A.data()), SmemLayoutA{});          // (BLK_M,BLK_K,PIPE)
-    Tensor sB = make_tensor(make_smem_ptr(shared_tensors.smem_B.data()), SmemLayoutB{});          // (BLK_N,BLK_K,PIPE)
-    
-    // Block scaling
-    Tensor sScaleAViewAsC = make_tensor(cute::make_smem_ptr(shared_tensors.smem_scale_A.data()),
-      Layout<
-        Shape<Shape<Int<ScaleGranularityM>, Int<ScaleMsPerTile>>, cute::tuple_element_t<1, TileShape>, Int<DispatchPolicy::Stages>>,
-        Stride<Stride<_0, _1>, _0, Int<ScaleMsPerTile>>
-      >{}); // ((ScaleGranularityM,ScaleMsPerTile),n,k)
-    Tensor sScaleB = make_tensor(cute::make_smem_ptr(shared_tensors.smem_scale_B.data()), SmemLayoutScaleB{}); // (k)
-
-    //
-    // Define C accumulators and A/B partitioning
-    //
-    
-    // Layout of warp group to thread mapping
-
-    static_assert(stride<0>(typename TiledMma::ALayout{}) == 0 and 
-                  stride<0>(typename TiledMma::BLayout{}) == 0 and
-                  size<0>(typename TiledMma::ALayout{}) == NumThreadsPerWarpGroup and
-                  size<0>(typename TiledMma::BLayout{}) == NumThreadsPerWarpGroup, 
-                  "Stride of the first mode must be 0 and the size of the mode must be NumThreadsPerWarpGroup");
-
-    constexpr int MmaWarpGroups = size(TiledMma{}) / NumThreadsPerWarpGroup;
-    Layout warp_group_thread_layout = make_layout(Int<MmaWarpGroups>{}, 
-                                                  Int<NumThreadsPerWarpGroup>{});
-
-    int warp_group_idx = __shfl_sync(0xFFFFFFFF, thread_idx / NumThreadsPerWarpGroup, 0);
-
-    TiledMma tiled_mma;
-    auto thread_mma = tiled_mma.get_slice(warp_group_thread_layout(warp_group_idx));
-
-    Tensor tCsScaleAViewAsC = tiled_mma.get_slice(thread_idx).partition_C(sScaleAViewAsC);    // (MMA,MMA_M,MMA_N,PIPE), `thread_mma` above is correct when partitioning A and B, but it is not correct when partitioning C.
-
-    Tensor tCsA = thread_mma.partition_A(sA);                                                 // (MMA,MMA_M,MMA_K,PIPE)
-    Tensor tCsB = thread_mma.partition_B(sB);                                                 // (MMA,MMA_N,MMA_K,PIPE)
-
-    // Allocate "fragments/descriptors"
-    Tensor tCrA = thread_mma.make_fragment_A(tCsA);                                           // (MMA,MMA_M,MMA_K,PIPE)
-    Tensor tCrB = thread_mma.make_fragment_B(tCsB);                                           // (MMA,MMA_N,MMA_K,PIPE)
-
-    CUTE_STATIC_ASSERT_V(size<1>(tCsA) == size<1>(accum));                                                         // M
-    CUTE_STATIC_ASSERT_V(size<1>(tCsB) == size<2>(accum));                                                         // N
-    CUTE_STATIC_ASSERT_V(size<2>(tCsA) == size<2>(tCsB));                                                          // K
-    CUTE_STATIC_ASSERT_V(size<3>(tCsA) == size<3>(tCsB));                                                       // PIPE
-    CUTE_STATIC_ASSERT_V(Int<DispatchPolicy::Stages>{} == size<2>(sA));                                         // PIPE
-    CUTE_STATIC_ASSERT_V(Int<DispatchPolicy::Stages>{} == size<2>(sB));                                         // PIPE
-
-    //
-    // PIPELINED MAIN LOOP
-    //
-    static_assert((0 <= K_PIPE_MMAS) && (K_PIPE_MMAS <  K_PIPE_MAX),
-        "ERROR : Incorrect number of MMAs in flight");
-
-    // We release buffers to producer warps(dma load) with some mmas in flight
-    PipelineState smem_pipe_release = smem_pipe_read;
-    
-    // Per block scale values for operand A and B
-
-    using RegLayoutScaleAViewAsC = decltype(make_layout_like(tCsScaleAViewAsC(_, _, _, 0).layout())); // `make_layout_like` makes a compact layout.
-    using RegLayoutScaleAEssential = decltype(filter_zeros(RegLayoutScaleAViewAsC{}.stride(), RegLayoutScaleAViewAsC{}.shape())); // an interface to traverse the underlying storage for the compact layout mentioned above
-
-    Tensor tCrScaleAViewAsC = make_tensor<ElementBlockScale>(RegLayoutScaleAViewAsC{});              // (MMA,MMA_M,MMA_N)
-    ElementBlockScale scale_b;
-
-    // Prologue GMMAs
-    int prologue_mma_count = min(K_PIPE_MMAS, k_tile_count);
-
-    tiled_mma.accumulate_ = GMMA::ScaleOut::Zero;
-
-    GmmaFP8AccumulationWithScale accumulation(accum, size<2>(TileShape{}) / size<2>(typename TiledMma::AtomShape_MNK{}), size<2>(tCrA));
-    warpgroup_fence_operand(accumulation());
-    CUTLASS_PRAGMA_UNROLL
-    for (int k_tile_prologue = prologue_mma_count; k_tile_prologue > 0; --k_tile_prologue)
-    {
-      // WAIT on smem_pipe_read until its data are available (phase bit flips from rdPhaseBit value)
-      auto barrier_token = pipeline.consumer_try_wait(smem_pipe_read);
-      pipeline.consumer_wait(smem_pipe_read, barrier_token);
-
-      if (accumulation.prepare_if_needed()) {
-        tiled_mma.accumulate_ = GMMA::ScaleOut::Zero;
-      }
-
-      int read_stage = smem_pipe_read.index();
-      
-      // Load per block scale values from shared memory to registers.
-      scale_b = sScaleB[read_stage];
-      CUTLASS_PRAGMA_UNROLL
-      for (int i = 0; i < size(RegLayoutScaleAEssential{}); i++) {
-        tCrScaleAViewAsC.data()[i] = tCsScaleAViewAsC(_, _, _, read_stage)(idx2crd(i, RegLayoutScaleAEssential{}));
-      }
-      if constexpr (ScaleMsPerTile == 1) {
-        static_assert(size(RegLayoutScaleAEssential{}) == 1);
-        tCrScaleAViewAsC.data()[0] = __shfl_sync(0xffffffff, tCrScaleAViewAsC.data()[0] * scale_b, 0); // `tCrScaleAViewAsC.data()[0]` are all same in a warp group when `ScaleMsPerTile == 1`.
-      } else {
-        CUTLASS_PRAGMA_UNROLL
-        for (int i = 0; i < size(RegLayoutScaleAEssential{}); i++) {
-          tCrScaleAViewAsC.data()[i] = tCrScaleAViewAsC.data()[i] * scale_b;
-        }
-      }
-
-      warpgroup_arrive();
-      // Unroll the K mode manually to set scale D to 1
-      CUTLASS_PRAGMA_UNROLL
-      for (int k_block = 0; k_block < size<2>(tCrA); ++k_block) {
-        // (V,M,K) x (V,N,K) => (V,M,N)
-        cute::gemm(tiled_mma, tCrA(_,_,k_block,read_stage), tCrB(_,_,k_block,read_stage), accumulation());
-        tiled_mma.accumulate_ = GMMA::ScaleOut::One;
-      }
-      warpgroup_commit_batch();
-
-      // Block scale the accumulators with reg tensor `tCrScaleAViewAsC`
-      accumulation.scale_if_needed(tCrScaleAViewAsC);
-
-      ++smem_pipe_read;
-    }
-
-    warpgroup_fence_operand(accumulation());
-    // Mainloop GMMAs
-    k_tile_count -= prologue_mma_count;
-
-    CUTLASS_PRAGMA_NO_UNROLL
-    for ( ; k_tile_count > 0; --k_tile_count)
-    {
-      // WAIT on smem_pipe_read until its data are available (phase bit flips from rdPhaseBit value)
-      auto barrier_token = pipeline.consumer_try_wait(smem_pipe_read);
-      pipeline.consumer_wait(smem_pipe_read, barrier_token);
-
-      //
-      // Compute on k_tile
-      //
-
-      int read_stage = smem_pipe_read.index();
-
-      // Load per block scale values from shared memory to registers (at most twice per block along M and exactly once per block along N) 
-      scale_b = sScaleB[read_stage];
-      CUTLASS_PRAGMA_UNROLL
-      for (int i = 0; i < size(RegLayoutScaleAEssential{}); i++) {
-        tCrScaleAViewAsC.data()[i] = tCsScaleAViewAsC(_, _, _, read_stage)(idx2crd(i, RegLayoutScaleAEssential{}));
-      }
-      if constexpr (ScaleMsPerTile == 1) {
-        static_assert(size(RegLayoutScaleAEssential{}) == 1);
-        tCrScaleAViewAsC.data()[0] = __shfl_sync(0xffffffff, tCrScaleAViewAsC.data()[0] * scale_b, 0); // `tCrScaleAViewAsC.data()[0]` are all same in a warp group when `ScaleMsPerTile == 1`.
-      } else {
-        CUTLASS_PRAGMA_UNROLL
-        for (int i = 0; i < size(RegLayoutScaleAEssential{}); i++) {
-          tCrScaleAViewAsC.data()[i] = tCrScaleAViewAsC.data()[i] * scale_b;
-        }
-      }
-
-      if (accumulation.prepare_if_needed()) {
-        tiled_mma.accumulate_ = GMMA::ScaleOut::Zero;
-      }
-
-      warpgroup_fence_operand(accumulation());
-      warpgroup_arrive();
-      // Unroll the K mode manually to set scale D to 1
-      CUTLASS_PRAGMA_UNROLL
-      for (int k_block = 0; k_block < size<2>(tCrA); ++k_block) {
-        // (V,M,K) x (V,N,K) => (V,M,N)
-        cute::gemm(tiled_mma, tCrA(_,_,k_block,read_stage), tCrB(_,_,k_block,read_stage), accumulation());
-        tiled_mma.accumulate_ = GMMA::ScaleOut::One;
-      }
-      warpgroup_commit_batch();
-
-      /// Wait on the GMMA barrier for K_PIPE_MMAS (or fewer) outstanding to ensure smem_pipe_write is consumed
-      warpgroup_wait<K_PIPE_MMAS>();
-      warpgroup_fence_operand(accumulation());
-
-      // Block scale the accumulators with reg tensor `tCrScaleAViewAsC`
-      accumulation.scale_if_needed(tCrScaleAViewAsC);
-
-      pipeline.consumer_release(smem_pipe_release);                 // UNLOCK smem_pipe_release, done _computing_ on it
-
-      // Advance smem_pipe_read and smem_pipe_release
-      ++smem_pipe_read;
-      ++smem_pipe_release;
-    }
-    
-    accumulation.scale_residue_if_needed(tCrScaleAViewAsC);
-
-    warpgroup_fence_operand(accumulation());
-  }
-
-  /// Perform a Consumer Epilogue to release all buffers
-  CUTLASS_DEVICE void
-  mma_tail(MainloopPipeline pipeline, PipelineState smem_pipe_release, int k_tile_count) {
-    // Prologue GMMAs
-    int prologue_mma_count = min(K_PIPE_MMAS, k_tile_count);
-    k_tile_count -= prologue_mma_count;
-
-    smem_pipe_release.advance(k_tile_count);
-
-    // Wait on all GMMAs to complete
-    warpgroup_wait<0>();
-
-    for (int count = 0; count < prologue_mma_count; ++count) {
-      pipeline.consumer_release(smem_pipe_release);                 // UNLOCK smem_pipe_release, done _computing_ on it
-      ++smem_pipe_release;
-    }
-  }
-};
-
-/////////////////////////////////////////////////////////////////////////////////////////////////
-
-} // namespace cutlass::gemm::collective
-
-/////////////////////////////////////////////////////////////////////////////////////////////////
--- a/csrc/cutlass_extensions/gemm/dispatch_policy.hpp
+++ b/csrc/cutlass_extensions/gemm/dispatch_policy.hpp
@ -1,39 +0,0 @@
-#pragma once
-
-#include "cutlass/gemm/dispatch_policy.hpp"
-
-namespace cutlass::gemm {
-
-//////////////////////////////////////////////////////////////////////////////
-
-// FP8 related policies (including Blocked Scaled Accumulation)
-//  `ScaleGranularityM` specifies scaling granularity along M, while zero-value
-//  `ScaleGranularityM` indicates that scaling granularity is
-//  `size<0>(TileShape_MNK{})` along M.
-template <int ScaleGranularityM = 0>
-struct KernelTmaWarpSpecializedCooperativeFP8BlockScaledSubGroupMAccum
-    : KernelTmaWarpSpecializedCooperative {};
-
-// n-buffer in smem (Hopper TMA), pipelined with Hopper GMMA and TMA, Warp
-// specialized dynamic schedule For FP8 kernels with Block Scaling
-template <int Stages_, class ClusterShape_ = Shape<_1, _1, _1>,
-          class KernelSchedule = KernelTmaWarpSpecialized,
-          int ScaleGranularityM =
-              0  // `ScaleGranularityM` specifies scaling granularity along M,
-                 // while zero-value `ScaleGranularityM` indicates that scaling
-                 // granularity is `size<0>(TileShape_MNK{})` along M.
-          >
-struct MainloopSm90TmaGmmaWarpSpecializedBlockScalingSubGroupMFP8
-    : MainloopSm90TmaGmmaWarpSpecialized<Stages_, ClusterShape_,
-                                         KernelSchedule> {
-  static_assert(
-      cute::is_same_v<
-          KernelSchedule,
-          KernelTmaWarpSpecializedCooperativeFP8BlockScaledSubGroupMAccum<
-              ScaleGranularityM>>,
-      "KernelSchedule must be one of the warp specialized policies");
-};
-
-//////////////////////////////////////////////////////////////////////////////
-
-}  // namespace cutlass::gemm
--- a/csrc/cutlass_extensions/vllm_collective_builder.cuh
+++ b/csrc/cutlass_extensions/vllm_collective_builder.cuh
@ -1,6 +1,6 @@
 #pragma once

-#include "cutlass_extensions/gemm/collective/collective_builder.hpp"
+#include "cutlass/gemm/collective/collective_builder.hpp"

 namespace cutlass::gemm::collective {
 using namespace cute;
--- a/csrc/dispatch_utils.h
+++ b/csrc/dispatch_utils.h
@ -19,6 +19,13 @@
 #define VLLM_DISPATCH_FLOATING_TYPES(TYPE, NAME, ...) \
  AT_DISPATCH_SWITCH(TYPE, NAME, VLLM_DISPATCH_CASE_FLOATING_TYPES(__VA_ARGS__))

+#define VLLM_DISPATCH_CASE_HALF_TYPES(...)            \
+  AT_DISPATCH_CASE(at::ScalarType::Half, __VA_ARGS__) \
+  AT_DISPATCH_CASE(at::ScalarType::BFloat16, __VA_ARGS__)
+
+#define VLLM_DISPATCH_HALF_TYPES(TYPE, NAME, ...) \
+  AT_DISPATCH_SWITCH(TYPE, NAME, VLLM_DISPATCH_CASE_HALF_TYPES(__VA_ARGS__))
+
 // ROCm devices might use either fn or fnuz, so set up dispatch table for both.
 // A host-based check at runtime will create a preferred FP8 type for ROCm
 // such that the correct kernel is dispatched.
--- a/csrc/layernorm_kernels.cu
+++ b/csrc/layernorm_kernels.cu
@ -140,6 +140,211 @@ fused_add_rms_norm_kernel(
  }
 }

+/* Function specialization in the case of FP16/BF16 tensors.
+   Additional optimizations we can make in this case are
+   packed and vectorized operations, which help with the
+   memory latency bottleneck.
+
+   _f16VecPN struct extends _f16Vec to add operations specifically required for
+   polynomial normalization (poly norm).
+   The original _f16Vec does not include the sum-of-powers computation or
+   in-place polynomial normalization logic. */
+template <typename scalar_t, int width>
+struct alignas(16) _f16VecPN : _f16Vec<scalar_t, width> {
+  using Base = _f16Vec<scalar_t, width>;
+  using Converter = typename Base::Converter;
+  using T1 = typename Base::T1;
+  using T2 = typename Base::T2;
+  using Base::data;
+
+  __device__ auto sum_pows() const {
+    float s2 = 0.0f, s4 = 0.0f, s6 = 0.0f;
+
+#pragma unroll
+    for (int i = 0; i < width; i += 2) {
+      float2 z = Converter::convert(T2{data[i], data[i + 1]});
+      float x2 = z.x * z.x;
+      float x4 = x2 * x2;
+      float x6 = x4 * x2;
+
+      float y2 = z.y * z.y;
+      float y4 = y2 * y2;
+      float y6 = y4 * y2;
+
+      s2 += x2 + y2;
+      s4 += x4 + y4;
+      s6 += x6 + y6;
+    }
+    return std::make_tuple(s2, s4, s6);
+  }
+
+  __device__ void poly_norm_inplace(const float w2_inv_std,
+                                    const float w1_inv_std2,
+                                    const float w0_inv_std3, const float bias) {
+#pragma unroll
+    for (int i = 0; i < width; i += 2) {
+      float2 z = Converter::convert(T2{data[i], data[i + 1]});
+
+      float x2 = z.x * z.x;
+      float x3 = x2 * z.x;
+      z.x = w2_inv_std * z.x + w1_inv_std2 * x2 + w0_inv_std3 * x3 + bias;
+
+      float y2 = z.y * z.y;
+      float y3 = y2 * z.y;
+      z.y = w2_inv_std * z.y + w1_inv_std2 * y2 + w0_inv_std3 * y3 + bias;
+
+      auto out = Converter::convert(z);
+      data[i] = out.x;
+      data[i + 1] = out.y;
+    }
+  }
+};
+
+template <typename scalar_t, int width>
+__global__ std::enable_if_t<(width > 0) && _typeConvert<scalar_t>::exists>
+poly_norm_kernel(scalar_t* __restrict__ out,           // [..., hidden_size]
+                 const scalar_t* __restrict__ input,   // [..., hidden_size]
+                 const scalar_t* __restrict__ weight,  // [3]
+                 const scalar_t* __restrict__ bias,    // [1]
+                 const float epsilon, const int hidden_size) {
+  // Sanity checks on our vector struct and type-punned pointer arithmetic
+  static_assert(std::is_pod_v<_f16VecPN<scalar_t, width>>);
+  static_assert(sizeof(_f16VecPN<scalar_t, width>) == sizeof(scalar_t) * width);
+
+  /* These and the argument pointers are all declared `restrict` as they are
+     not aliased in practice. Argument pointers should not be dereferenced
+     in this kernel as that would be undefined behavior */
+  auto* __restrict__ input_v =
+      reinterpret_cast<const _f16VecPN<scalar_t, width>*>(input);
+  const int vec_hidden_size = hidden_size / width;
+  float variance = 0.0f;
+  float variance2 = 0.0f;
+  float variance3 = 0.0f;
+
+  for (int idx = threadIdx.x; idx < vec_hidden_size; idx += blockDim.x) {
+    int id = blockIdx.x * vec_hidden_size + idx;
+    _f16VecPN<scalar_t, width> temp = input_v[id];
+    auto [x2, x4, x6] = temp.sum_pows();
+
+    variance += x2;
+    variance2 += x4;
+    variance3 += x6;
+  }
+
+  float3 thread_variances = make_float3(variance, variance2, variance3);
+
+  struct SumOp {
+    __device__ float3 operator()(const float3& a, const float3& b) const {
+      return make_float3(a.x + b.x, a.y + b.y, a.z + b.z);
+    }
+  };
+
+  using BlockReduce = cub::BlockReduce<float3, 1024>;
+  __shared__ typename BlockReduce::TempStorage reduceStore;
+  float3 block_variances =
+      BlockReduce(reduceStore).Reduce(thread_variances, SumOp{}, blockDim.x);
+
+  variance = block_variances.x;
+  variance2 = block_variances.y;
+  variance3 = block_variances.z;
+
+  __shared__ float s_w2_inv_std;
+  __shared__ float s_w1_inv_std2;
+  __shared__ float s_w0_inv_std3;
+  __shared__ float s_bias;
+
+  if (threadIdx.x == 0) {
+    float w0 = (float)weight[0];
+    float w1 = (float)weight[1];
+    float w2 = (float)weight[2];
+    s_bias = (float)bias[0];
+
+    s_w2_inv_std = w2 * rsqrtf(variance / hidden_size + epsilon);
+    s_w1_inv_std2 = w1 * rsqrtf(variance2 / hidden_size + epsilon);
+    s_w0_inv_std3 = w0 * rsqrtf(variance3 / hidden_size + epsilon);
+  }
+  __syncthreads();
+
+  auto* __restrict__ out_v = reinterpret_cast<_f16VecPN<scalar_t, width>*>(out);
+
+  for (int idx = threadIdx.x; idx < vec_hidden_size; idx += blockDim.x) {
+    int id = blockIdx.x * vec_hidden_size + idx;
+    _f16VecPN<scalar_t, width> temp = input_v[id];
+    temp.poly_norm_inplace(s_w2_inv_std, s_w1_inv_std2, s_w0_inv_std3, s_bias);
+    out_v[id] = temp;
+  }
+}
+
+/* Generic poly_norm_kernel
+   The width field is not used here but necessary for other specializations.
+ */
+template <typename scalar_t, int width>
+__global__ std::enable_if_t<(width == 0) || !_typeConvert<scalar_t>::exists>
+poly_norm_kernel(scalar_t* __restrict__ out,           // [..., hidden_size]
+                 const scalar_t* __restrict__ input,   // [..., hidden_size]
+                 const scalar_t* __restrict__ weight,  // [3]
+                 const scalar_t* __restrict__ bias,    // [1]
+                 const float epsilon, const int hidden_size) {
+  float variance = 0.0f;
+  float variance2 = 0.0f;
+  float variance3 = 0.0f;
+
+  for (int idx = threadIdx.x; idx < hidden_size; idx += blockDim.x) {
+    float x = (float)input[blockIdx.x * hidden_size + idx];
+    float x2 = x * x;
+    float x4 = x2 * x2;
+    float x6 = x4 * x2;
+
+    variance += x2;
+    variance2 += x4;
+    variance3 += x6;
+  }
+
+  float3 thread_variances = make_float3(variance, variance2, variance3);
+
+  struct SumOp {
+    __device__ float3 operator()(const float3& a, const float3& b) const {
+      return make_float3(a.x + b.x, a.y + b.y, a.z + b.z);
+    }
+  };
+
+  using BlockReduce = cub::BlockReduce<float3, 1024>;
+  __shared__ typename BlockReduce::TempStorage reduceStore;
+  float3 block_variances =
+      BlockReduce(reduceStore).Reduce(thread_variances, SumOp{}, blockDim.x);
+
+  variance = block_variances.x;
+  variance2 = block_variances.y;
+  variance3 = block_variances.z;
+
+  __shared__ float s_w2_inv_std;
+  __shared__ float s_w1_inv_std2;
+  __shared__ float s_w0_inv_std3;
+  __shared__ float s_bias;
+
+  if (threadIdx.x == 0) {
+    float w0 = (float)weight[0];
+    float w1 = (float)weight[1];
+    float w2 = (float)weight[2];
+    s_bias = (float)bias[0];
+
+    s_w2_inv_std = w2 * rsqrtf(variance / hidden_size + epsilon);
+    s_w1_inv_std2 = w1 * rsqrtf(variance2 / hidden_size + epsilon);
+    s_w0_inv_std3 = w0 * rsqrtf(variance3 / hidden_size + epsilon);
+  }
+  __syncthreads();
+
+  for (int idx = threadIdx.x; idx < hidden_size; idx += blockDim.x) {
+    float x = (float)input[blockIdx.x * hidden_size + idx];
+    float x2 = x * x;
+    float x3 = x2 * x;
+
+    out[blockIdx.x * hidden_size + idx] =
+        (scalar_t)(x * s_w2_inv_std + x2 * s_w1_inv_std2 + x3 * s_w0_inv_std3 +
+                   s_bias);
+  }
+}
+
 }  // namespace vllm

 void rms_norm(torch::Tensor& out,     // [..., hidden_size]
@ -219,3 +424,49 @@ void fused_add_rms_norm(torch::Tensor& input,     // [..., hidden_size]
    LAUNCH_FUSED_ADD_RMS_NORM(0);
  }
 }
+
+#define LAUNCH_FUSED_POLY_NORM(width)                                         \
+  VLLM_DISPATCH_FLOATING_TYPES(input.scalar_type(), "poly_norm_kernel", [&] { \
+    vllm::poly_norm_kernel<scalar_t, width><<<grid, block, 0, stream>>>(      \
+        out.data_ptr<scalar_t>(), input.data_ptr<scalar_t>(),                 \
+        weight.data_ptr<scalar_t>(), bias.data_ptr<scalar_t>(), epsilon,      \
+        hidden_size);                                                         \
+  });
+
+void poly_norm(torch::Tensor& out,     // [..., hidden_size]
+               torch::Tensor& input,   // [..., hidden_size]
+               torch::Tensor& weight,  // [3]
+               torch::Tensor& bias,    // [1]
+               double epsilon) {
+  TORCH_CHECK(out.is_contiguous());
+  TORCH_CHECK(input.is_contiguous());
+  TORCH_CHECK(out.data_ptr() != input.data_ptr());
+
+  int hidden_size = input.size(-1);
+  int num_tokens = input.numel() / hidden_size;
+
+  dim3 grid(num_tokens);
+  /* This kernel is memory-latency bound in many scenarios.
+     When num_tokens is large, a smaller block size allows
+     for increased block occupancy on CUs and better latency
+     hiding on global mem ops. */
+  const int max_block_size = (num_tokens < 256) ? 1024 : 256;
+  dim3 block(std::min(hidden_size, max_block_size));
+  const at::cuda::OptionalCUDAGuard device_guard(device_of(input));
+  const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
+  /*If the tensor types are FP16/BF16, try to use the optimized kernel
+    with packed + vectorized ops.
+    Max optimization is achieved with a width-8 vector of FP16/BF16s
+    since we can load at most 128 bits at once in a global memory op.
+    However, this requires each tensor's data to be aligned to 16
+    bytes.
+   */
+  auto inp_ptr = reinterpret_cast<std::uintptr_t>(input.data_ptr());
+  auto out_ptr = reinterpret_cast<std::uintptr_t>(out.data_ptr());
+  bool ptrs_are_aligned = inp_ptr % 16 == 0 && out_ptr % 16 == 0;
+  if (ptrs_are_aligned && hidden_size % 8 == 0) {
+    LAUNCH_FUSED_POLY_NORM(8);
+  } else {
+    LAUNCH_FUSED_POLY_NORM(0);
+  }
+}
--- a/csrc/layernorm_quant_kernels.cu
+++ b/csrc/layernorm_quant_kernels.cu
@ -6,7 +6,7 @@
 */

 #include "type_convert.cuh"
-#include "quantization/fp8/common.cuh"
+#include "quantization/w8a8/fp8/common.cuh"
 #include "dispatch_utils.h"

 #include <torch/cuda.h>
--- a/csrc/mamba/mamba_ssm/selective_scan_fwd.cu
+++ b/csrc/mamba/mamba_ssm/selective_scan_fwd.cu
@ -27,11 +27,12 @@

 template<int kNThreads_, int kNItems_, int kNRows_, bool kIsEvenLen_,
         bool kIsVariableB_, bool kIsVariableC_,
-         bool kHasZ_, bool kVarlen_, typename input_t_, typename weight_t_>
+         bool kHasZ_, bool kVarlen_, typename input_t_, typename weight_t_, typename state_t_>
 struct Selective_Scan_fwd_kernel_traits {
    static_assert(kNItems_ % 4 == 0);
    using input_t = input_t_;
    using weight_t = weight_t_;
+    using state_t = state_t_;
    static constexpr int kNThreads = kNThreads_;
    // Setting MinBlocksPerMP to be 3 (instead of 2) for 128 threads improves occupancy.
    static constexpr int kMinBlocks = kNThreads < 128 ? 5 : 3;
@ -132,7 +133,7 @@ void selective_scan_fwd_kernel(SSMParamsBase params) {
    input_t *Bvar = reinterpret_cast<input_t *>(params.B_ptr) + sequence_start_index * params.B_batch_stride + group_id * params.B_group_stride;
    weight_t *C = reinterpret_cast<weight_t *>(params.C_ptr) + dim_id * kNRows * params.C_d_stride;
    input_t *Cvar = reinterpret_cast<input_t *>(params.C_ptr) + sequence_start_index * params.C_batch_stride + group_id * params.C_group_stride;
-    input_t *ssm_states = reinterpret_cast<input_t *>(params.ssm_states_ptr) + 
+    typename Ktraits::state_t *ssm_states = reinterpret_cast<typename Ktraits::state_t *>(params.ssm_states_ptr) + 
    cache_index * params.ssm_states_batch_stride + 
    dim_id * kNRows * params.ssm_states_dim_stride;
    
@ -261,7 +262,7 @@ void selective_scan_fwd_kernel(SSMParamsBase params) {
                if (threadIdx.x == 0) {
                    smem_running_prefix[state_idx] = prefix_op.running_prefix;
                    if (chunk == n_chunks - 1) {
-                        ssm_states[state_idx * params.ssm_states_dstate_stride] = input_t(prefix_op.running_prefix.y);
+                        ssm_states[state_idx * params.ssm_states_dstate_stride] = typename Ktraits::state_t(prefix_op.running_prefix.y);
                    }
                }
                #pragma unroll
@ -310,7 +311,7 @@ void selective_scan_fwd_kernel(SSMParamsBase params) {
    }
 }

-template<int kNThreads, int kNItems, typename input_t, typename weight_t>
+template<int kNThreads, int kNItems, typename input_t, typename weight_t, typename state_t>
 void selective_scan_fwd_launch(SSMParamsBase &params, cudaStream_t stream) {
    // Only kNRows == 1 is tested for now, which ofc doesn't differ from previously when we had each block
    // processing 1 row.
@ -321,7 +322,7 @@ void selective_scan_fwd_launch(SSMParamsBase &params, cudaStream_t stream) {
    BOOL_SWITCH(params.seqlen % (kNThreads * kNItems) == 0, kIsEvenLen, [&] {
        BOOL_SWITCH(params.z_ptr != nullptr , kHasZ, [&] {
            BOOL_SWITCH(params.query_start_loc_ptr != nullptr , kVarlen, [&] {
-                using Ktraits = Selective_Scan_fwd_kernel_traits<kNThreads, kNItems, kNRows, kIsEvenLen, kIsVariableB, kIsVariableC, kHasZ,  kVarlen, input_t, weight_t>;
+                using Ktraits = Selective_Scan_fwd_kernel_traits<kNThreads, kNItems, kNRows, kIsEvenLen, kIsVariableB, kIsVariableC, kHasZ,  kVarlen, input_t, weight_t, state_t>;
                constexpr int kSmemSize = Ktraits::kSmemSize + kNRows * MAX_DSTATE * sizeof(typename Ktraits::scan_t);
                dim3 grid(params.batch, params.dim / kNRows);
                auto kernel = &selective_scan_fwd_kernel<Ktraits>;
@ -341,59 +342,78 @@ void selective_scan_fwd_launch(SSMParamsBase &params, cudaStream_t stream) {
    });
 }

-template<typename input_t, typename weight_t>
+template<typename input_t, typename weight_t, typename state_t>
 void selective_scan_fwd_cuda(SSMParamsBase &params, cudaStream_t stream) {

    #ifndef USE_ROCM
        if (params.seqlen <= 128) {           
-            selective_scan_fwd_launch<32, 4, input_t, weight_t>(params, stream);
+            selective_scan_fwd_launch<32, 4, input_t, weight_t, state_t>(params, stream);
        } else if (params.seqlen <= 256) {
-            selective_scan_fwd_launch<32, 8, input_t, weight_t>(params, stream);
+            selective_scan_fwd_launch<32, 8, input_t, weight_t, state_t>(params, stream);
        } else if (params.seqlen <= 512) {
-            selective_scan_fwd_launch<32, 16, input_t, weight_t>(params, stream);
+            selective_scan_fwd_launch<32, 16, input_t, weight_t, state_t>(params, stream);
        } else if (params.seqlen <= 1024) {
-            selective_scan_fwd_launch<64, 16, input_t, weight_t>(params, stream);
+            selective_scan_fwd_launch<64, 16, input_t, weight_t, state_t>(params, stream);
        } else {
-            selective_scan_fwd_launch<128, 16, input_t, weight_t>(params, stream);
+            selective_scan_fwd_launch<128, 16, input_t, weight_t, state_t>(params, stream);
        }
    #else
        if (params.seqlen <= 256) {
-            selective_scan_fwd_launch<64, 4, input_t, weight_t>(params, stream);
+            selective_scan_fwd_launch<64, 4, input_t, weight_t, state_t>(params, stream);
        } else if (params.seqlen <= 512) {
-            selective_scan_fwd_launch<64, 8, input_t, weight_t>(params, stream);
+            selective_scan_fwd_launch<64, 8, input_t, weight_t, state_t>(params, stream);
        } else if (params.seqlen <= 1024) {
-            selective_scan_fwd_launch<64, 16, input_t, weight_t>(params, stream);
+            selective_scan_fwd_launch<64, 16, input_t, weight_t, state_t>(params, stream);
        } else {
-            selective_scan_fwd_launch<128, 16, input_t, weight_t>(params, stream);
+            selective_scan_fwd_launch<128, 16, input_t, weight_t, state_t>(params, stream);
        }
    #endif
 }

-template void selective_scan_fwd_cuda<at::BFloat16, float>(SSMParamsBase &params, cudaStream_t stream);
-template void selective_scan_fwd_cuda<at::Half, float>(SSMParamsBase &params, cudaStream_t stream);
-template void selective_scan_fwd_cuda<float, float>(SSMParamsBase &params, cudaStream_t stream);
+template void selective_scan_fwd_cuda<at::BFloat16, float, at::BFloat16>(SSMParamsBase &params, cudaStream_t stream);
+template void selective_scan_fwd_cuda<at::BFloat16, float, float>(SSMParamsBase &params, cudaStream_t stream);
+template void selective_scan_fwd_cuda<at::Half, float, at::Half>(SSMParamsBase &params, cudaStream_t stream);
+template void selective_scan_fwd_cuda<at::Half, float, float>(SSMParamsBase &params, cudaStream_t stream);
+template void selective_scan_fwd_cuda<float, float, float>(SSMParamsBase &params, cudaStream_t stream);

 #define CHECK_SHAPE(x, ...) TORCH_CHECK(x.sizes() == torch::IntArrayRef({__VA_ARGS__}), #x " must have shape (" #__VA_ARGS__ ")")

-#define DISPATCH_WTYPE_ITYPE_FLOAT_AND_HALF_AND_BF16(ITYPE, NAME, ...)              \
+#define DISPATCH_WTYPE_ITYPE_FLOAT_AND_HALF_AND_BF16(ITYPE, STYPE, NAME, ...)       \
    if (ITYPE == at::ScalarType::Half) {                                            \
        using input_t = at::Half;                                                   \
        using weight_t = float;                                                     \
-        __VA_ARGS__();                                                              \
+        if (STYPE == at::ScalarType::Half) {                                        \
+            using state_t = at::Half;                                               \
+            __VA_ARGS__();                                                          \
+        } else if (STYPE == at::ScalarType::Float) {                                \
+            using state_t = float;                                                  \
+            __VA_ARGS__();                                                          \
+        } else {                                                                    \
+            AT_ERROR(#NAME, " not implemented for state type '", toString(STYPE), "'"); \
+        }                                                                           \
    } else if (ITYPE == at::ScalarType::BFloat16) {                                 \
        using input_t = at::BFloat16;                                               \
        using weight_t = float;                                                     \
-        __VA_ARGS__();                                                              \
+        if (STYPE == at::ScalarType::BFloat16) {                                    \
+            using state_t = at::BFloat16;                                           \
+            __VA_ARGS__();                                                          \
+        } else if (STYPE == at::ScalarType::Float) {                                \
+            using state_t = float;                                                  \
+            __VA_ARGS__();                                                          \
+        } else {                                                                    \
+            AT_ERROR(#NAME, " not implemented for state type '", toString(STYPE), "'"); \
+        }                                                                           \
    } else if (ITYPE == at::ScalarType::Float)  {                                   \
        using input_t = float;                                                      \
        using weight_t = float;                                                     \
+        using state_t = float;                                                      \
        __VA_ARGS__();                                                              \
    } else {                                                                        \
        AT_ERROR(#NAME, " not implemented for input type '", toString(ITYPE), "'"); \
    }


-template<typename input_t, typename weight_t>
+template<typename input_t, typename weight_t, typename state_t>
 void selective_scan_fwd_cuda(SSMParamsBase &params, cudaStream_t stream);

 void set_ssm_params_fwd(SSMParamsBase &params,
@ -648,7 +668,9 @@ void selective_scan_fwd(const torch::Tensor &u, const torch::Tensor &delta,

    // Right now u has BHL layout and delta has HBL layout, and we want out to have HBL layout
    at::Tensor out = delta;
-    TORCH_CHECK(ssm_states.scalar_type() == input_type);
+    // ssm_states can now be either the same as input_type or float32
+    auto state_type = ssm_states.scalar_type();
+    TORCH_CHECK(state_type == input_type || state_type == at::ScalarType::Float);
    TORCH_CHECK(ssm_states.is_cuda());
    TORCH_CHECK(ssm_states.stride(-1) == 1);

@ -670,7 +692,7 @@ void selective_scan_fwd(const torch::Tensor &u, const torch::Tensor &delta,
    
    const at::cuda::OptionalCUDAGuard device_guard(device_of(u));
    auto stream = at::cuda::getCurrentCUDAStream().stream();
-    DISPATCH_WTYPE_ITYPE_FLOAT_AND_HALF_AND_BF16(u.scalar_type(), "selective_scan_fwd", [&] {
-        selective_scan_fwd_cuda<input_t, weight_t>(params, stream);
+    DISPATCH_WTYPE_ITYPE_FLOAT_AND_HALF_AND_BF16(u.scalar_type(), ssm_states.scalar_type(), "selective_scan_fwd", [&] {
+        selective_scan_fwd_cuda<input_t, weight_t, state_t>(params, stream);
    });
 }
--- a/csrc/moe/grouped_topk_kernels.cu
+++ b/csrc/moe/grouped_topk_kernels.cu
@ -0,0 +1,758 @@
+/*
+ * 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 float kNegInfinity = INFINITY * -1;
+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 = kNegInfinity;
+  T topk_group_value = kNegInfinity;
+  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 = kNegInfinity;
+      }
+      pre_count_equal_to_top_value = count_equal_to_top_value;
+      count_equal_to_top_value = __popc(__ballot_sync(
+          FULL_WARP_MASK, (value == cuda_cast<T, float>(kNegInfinity))));
+    }
+    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_cast<T, float>(kNegInfinity));
+  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_cast<T, float>(kNegInfinity);
+          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/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/Show More
+++ b/Show More