Merge remote-tracking branch 'nm/lwilkinson/fix-flashmla-full-cudagraph' into wide_ep_working_branch

Signed-off-by: Lucas Wilkinson <lwilkins@redhat.com>

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

@@ -46,6 +46,6 @@ while getopts "m:b:l:f:t:" OPT; do
 done
 
 lm_eval --model vllm \
-  --model_args "pretrained=$MODEL,tensor_parallel_size=$TP_SIZE,distributed_executor_backend=ray,trust_remote_code=true,max_model_len=4096" \
+  --model_args "pretrained=$MODEL,tensor_parallel_size=$TP_SIZE,add_bos_token=true,trust_remote_code=true,max_model_len=4096" \
   --tasks gsm8k --num_fewshot "$FEWSHOT" --limit "$LIMIT" \
   --batch_size "$BATCH_SIZE"

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

@@ -18,12 +18,14 @@ RTOL = 0.08
 
 def launch_lm_eval(eval_config, tp_size):
     trust_remote_code = eval_config.get("trust_remote_code", False)
+    max_model_len = eval_config.get("max_model_len", 4096)
     model_args = (
         f"pretrained={eval_config['model_name']},"
         f"tensor_parallel_size={tp_size},"
         f"enforce_eager=true,"
         f"add_bos_token=true,"
-        f"trust_remote_code={trust_remote_code}"
+        f"trust_remote_code={trust_remote_code},"
+        f"max_model_len={max_model_len}"
     )
     results = lm_eval.simple_evaluate(
         model="vllm",

.buildkite/nightly-benchmarks/README.md

@@ -11,7 +11,7 @@ See [vLLM performance dashboard](https://perf.vllm.ai) for the latest performanc
 
 ## Performance benchmark quick overview
 
-**Benchmarking Coverage**: latency, throughput and fix-qps serving on A100 (the support for FP8 benchmark on H100 is coming!), with different models.
+**Benchmarking Coverage**: latency, throughput and fix-qps serving on A100 (the support for FP8 benchmark on H100 is coming!) and Intel® Xeon® Processors, with different models.
 
 **Benchmarking Duration**: about 1hr.
 
@@ -31,13 +31,27 @@ Performance benchmark will be triggered when:
 - A PR being merged into vllm.
 - Every commit for those PRs with `perf-benchmarks` label AND `ready` label.
 
+Manually Trigger the benchmark
+
+```bash
+bash .buildkite/nightly-benchmarks/scripts/run-performance-benchmarks.sh
+```
+
+Runtime environment variables:
+- `ON_CPU`: set the value to '1' on Intel® Xeon® Processors. Default value is 0.
+- `SERVING_JSON`: JSON file to use for the serving tests. Default value is empty string (use default file).
+- `LATENCY_JSON`: JSON file to use for the latency tests. Default value is empty string (use default file).
+- `THROUGHPUT_JSON`: JSON file to use for the throughout tests. Default value is empty string (use default file).
+- `REMOTE_HOST`: IP for the remote vLLM service to benchmark. Default value is empty string.
+- `REMOTE_PORT`: Port for the remote vLLM service to benchmark. Default value is empty string.
+
 Nightly benchmark will be triggered when:
 - Every commit for those PRs with `perf-benchmarks` label and `nightly-benchmarks` label.
 
 ## Performance benchmark details
 
 See [performance-benchmarks-descriptions.md](performance-benchmarks-descriptions.md) for detailed descriptions, and use `tests/latency-tests.json`, `tests/throughput-tests.json`, `tests/serving-tests.json` to configure the test cases.
-
+> NOTE: For Intel® Xeon® Processors, use `tests/latency-tests-cpu.json`, `tests/throughput-tests-cpu.json`, `tests/serving-tests-cpu.json` instead.
 ### Latency test
 
 Here is an example of one test inside `latency-tests.json`:
@@ -60,7 +74,7 @@ Here is an example of one test inside `latency-tests.json`:
 In this example:
 
 - The `test_name` attributes is a unique identifier for the test. In `latency-tests.json`, it must start with `latency_`.
-- The `parameters` attribute control the command line arguments to be used for `benchmark_latency.py`. Note that please use underline `_` instead of the dash `-` when specifying the command line arguments, and `run-performance-benchmarks.sh` will convert the underline to dash when feeding the arguments to `benchmark_latency.py`. For example, the corresponding command line arguments for `benchmark_latency.py` will be `--model meta-llama/Meta-Llama-3-8B --tensor-parallel-size 1 --load-format dummy --num-iters-warmup 5 --num-iters 15`
+- The `parameters` attribute control the command line arguments to be used for `vllm bench latency`. Note that please use underline `_` instead of the dash `-` when specifying the command line arguments, and `run-performance-benchmarks.sh` will convert the underline to dash when feeding the arguments to `vllm bench latency`. For example, the corresponding command line arguments for `vllm bench latency` will be `--model meta-llama/Meta-Llama-3-8B --tensor-parallel-size 1 --load-format dummy --num-iters-warmup 5 --num-iters 15`
 
 Note that the performance numbers are highly sensitive to the value of the parameters. Please make sure the parameters are set correctly.
 
@@ -68,13 +82,13 @@ WARNING: The benchmarking script will save json results by itself, so please do
 
 ### Throughput test
 
-The tests are specified in `throughput-tests.json`. The syntax is similar to `latency-tests.json`, except for that the parameters will be fed forward to `benchmark_throughput.py`.
+The tests are specified in `throughput-tests.json`. The syntax is similar to `latency-tests.json`, except for that the parameters will be fed forward to `vllm bench throughput`.
 
 The number of this test is also stable -- a slight change on the value of this number might vary the performance numbers by a lot.
 
 ### Serving test
 
-We test the throughput by using `benchmark_serving.py` with request rate = inf to cover the online serving overhead. The corresponding parameters are in `serving-tests.json`, and here is an example:
+We test the throughput by using `vllm bench serve` with request rate = inf to cover the online serving overhead. The corresponding parameters are in `serving-tests.json`, and here is an example:
 
 ```json
 [
@@ -104,8 +118,8 @@ Inside this example:
 
 - The `test_name` attribute is also a unique identifier for the test. It must start with `serving_`.
 - The `server-parameters` includes the command line arguments for vLLM server.
-- The `client-parameters` includes the command line arguments for `benchmark_serving.py`.
-- The `qps_list` controls the list of qps for test. It will be used to configure the `--request-rate` parameter in `benchmark_serving.py`
+- The `client-parameters` includes the command line arguments for `vllm bench serve`.
+- The `qps_list` controls the list of qps for test. It will be used to configure the `--request-rate` parameter in `vllm bench serve`
 
 The number of this test is less stable compared to the delay and latency benchmarks (due to randomized sharegpt dataset sampling inside `benchmark_serving.py`), but a large change on this number (e.g. 5% change) still vary the output greatly.
 
@@ -119,6 +133,30 @@ If you do not see the table, please wait till the benchmark finish running.
 The json version of the table (together with the json version of the benchmark) will be also attached to the markdown file.
 The raw benchmarking results (in the format of json files) are in the `Artifacts` tab of the benchmarking.
 
+The `compare-json-results.py` helps to compare benchmark results JSON files converted using `convert-results-json-to-markdown.py`.
+When run, benchmark script generates results under `benchmark/results` folder, along with the `benchmark_results.md` and `benchmark_results.json`.
+`compare-json-results.py` compares two `benchmark_results.json` files and provides performance ratio e.g. for Output Tput, Median TTFT and Median TPOT.
+
+Here is an example using the script to compare result_a and result_b without detail test name.
+`python3 compare-json-results.py -f results_a/benchmark_results.json -f results_b/benchmark_results.json --ignore_test_name`
+
+|    | results_a/benchmark_results.json | results_b/benchmark_results.json | perf_ratio        |
+|----|----------------------------------------|----------------------------------------|----------|
+| 0  | 142.633982                             | 156.526018                             | 1.097396 |
+| 1  | 241.620334                             | 294.018783                             | 1.216863 |
+| 2  | 218.298905                             | 262.664916                             | 1.203235 |
+| 3  | 242.743860                             | 299.816190                             | 1.235113 |
+
+Here is an example using the script to compare result_a and result_b with detail test name.
+`python3 compare-json-results.py -f results_a/benchmark_results.json -f results_b/benchmark_results.json`
+|   | results_a/benchmark_results.json_name | results_a/benchmark_results.json | results_b/benchmark_results.json_name | results_b/benchmark_results.json | perf_ratio        |
+|---|---------------------------------------------|----------------------------------------|---------------------------------------------|----------------------------------------|----------|
+| 0 | serving_llama8B_tp1_sharegpt_qps_1          | 142.633982                             | serving_llama8B_tp1_sharegpt_qps_1          | 156.526018                             | 1.097396 |
+| 1 | serving_llama8B_tp1_sharegpt_qps_16         | 241.620334                             | serving_llama8B_tp1_sharegpt_qps_16         | 294.018783                             | 1.216863 |
+| 2 | serving_llama8B_tp1_sharegpt_qps_4          | 218.298905                             | serving_llama8B_tp1_sharegpt_qps_4          | 262.664916                             | 1.203235 |
+| 3 | serving_llama8B_tp1_sharegpt_qps_inf        | 242.743860                             | serving_llama8B_tp1_sharegpt_qps_inf        | 299.816190                             | 1.235113 |
+| 4 | serving_llama8B_tp2_random_1024_128_qps_1   | 96.613390                              | serving_llama8B_tp4_random_1024_128_qps_1   | 108.404853                             | 1.122048 |
+
 ## Nightly test details
 
 See [nightly-descriptions.md](nightly-descriptions.md) for the detailed description on test workload, models and docker containers of benchmarking other llm engines.

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

@@ -4,7 +4,8 @@
 - Input length: 32 tokens.
 - Output length: 128 tokens.
 - Batch size: fixed (8).
-- Models: llama-3.1 8B, llama-3 70B, mixtral 8x7B.
+- GPU Models: llama-3.1 8B, llama-3 70B, mixtral 8x7B.
+- CPU Models: llama-3.1 8B.
 - Evaluation metrics: end-to-end latency (mean, median, p99).
 
 {latency_tests_markdown_table}
@@ -14,7 +15,8 @@
 - Input length: randomly sample 200 prompts from ShareGPT dataset (with fixed random seed).
 - Output length: the corresponding output length of these 200 prompts.
 - Batch size: dynamically determined by vllm to achieve maximum throughput.
-- Models: llama-3.1 8B, llama-3 70B, mixtral 8x7B.
+- GPU Models: llama-3.1 8B, llama-3 70B, mixtral 8x7B.
+- CPU Models: llama-3.1 8B.
 - Evaluation metrics: throughput.
 
 {throughput_tests_markdown_table}
@@ -25,12 +27,18 @@
 - Output length: the corresponding output length of these 200 prompts.
 - Batch size: dynamically determined by vllm and the arrival pattern of the requests.
 - **Average QPS (query per second)**: 1, 4, 16 and inf. QPS = inf means all requests come at once. For other QPS values, the arrival time of each query is determined using a random Poisson process (with fixed random seed).
-- Models: llama-3.1 8B, llama-3 70B, mixtral 8x7B.
-- We also added a speculative decoding test for llama-3 70B, under QPS 2
+- GPU Models: llama-3.1 8B, llama-3 70B, mixtral 8x7B.
+- We also added a speculative decoding test for llama-3 70B on GPU, under QPS 2
+- CPU Models: llama-3.1 8B.
 - Evaluation metrics: throughput, TTFT (time to the first token, with mean, median and p99), ITL (inter-token latency, with mean, median and p99).
+- For CPU, we added random dataset tests to benchmark fixed input/output length with 100 prompts.
 
 {serving_tests_markdown_table}
 
+## Platform Information
+
+{platform_markdown_table}
+
 ## json version of the benchmarking tables
 
 This section contains the data of the markdown tables above in JSON format.

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

@@ -0,0 +1,66 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+import argparse
+
+import pandas as pd
+
+
+def compare_data_columns(
+    files, name_column, data_column, drop_column, ignore_test_name=False
+):
+    print("\ncompare_data_column: " + data_column)
+    frames = []
+    compare_frames = []
+    for file in files:
+        data_df = pd.read_json(file)
+        serving_df = data_df.dropna(subset=[drop_column], ignore_index=True)
+        if ignore_test_name is False:
+            serving_df = serving_df.rename(columns={name_column: file + "_name"})
+            frames.append(serving_df[file + "_name"])
+        serving_df = serving_df.rename(columns={data_column: file})
+        frames.append(serving_df[file])
+        compare_frames.append(serving_df[file])
+        if len(compare_frames) >= 2:
+            # Compare numbers among two files
+            ratio_df = compare_frames[1] / compare_frames[0]
+            frames.append(ratio_df)
+            compare_frames.pop(1)
+
+    concat_df = pd.concat(frames, axis=1)
+    return concat_df
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "-f", "--file", action="append", type=str, help="input file name"
+    )
+    parser.add_argument(
+        "--ignore_test_name", action="store_true", help="ignore_test_name or not"
+    )
+    args = parser.parse_args()
+    files = args.file
+    print("comparing : " + ", ".join(files))
+
+    drop_column = "P99"
+    name_column = "Test name"
+    data_cols_to_compare = ["Output Tput (tok/s)", "Median TTFT (ms)", "Median"]
+    html_msgs_for_data_cols = [
+        "Compare Output Tokens /n",
+        "Median TTFT /n",
+        "Median TPOT /n",
+    ]
+    ignore_test_name = args.ignore_test_name
+    with open("perf_comparison.html", "w") as text_file:
+        for i in range(len(data_cols_to_compare)):
+            output_df = compare_data_columns(
+                files,
+                name_column,
+                data_cols_to_compare[i],
+                drop_column,
+                ignore_test_name=ignore_test_name,
+            )
+            print(output_df)
+            html = output_df.to_html()
+            text_file.write(html_msgs_for_data_cols[i])
+            text_file.write(html)

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

@@ -3,9 +3,11 @@
 
 import json
 import os
+from importlib import util
 from pathlib import Path
 
 import pandas as pd
+import psutil
 from tabulate import tabulate
 
 results_folder = Path("results/")
@@ -29,11 +31,11 @@ throughput_results = []
 throughput_results_column_mapping = {
     "test_name": "Test name",
     "gpu_type": "GPU",
-    # "num_requests": "# of req.",
-    # "total_num_tokens": "Total # of tokens",
-    # "elapsed_time": "Elapsed time (s)",
+    "num_requests": "# of req.",
+    "total_num_tokens": "Total # of tokens",
+    "elapsed_time": "Elapsed time (s)",
     "requests_per_second": "Tput (req/s)",
-    # "tokens_per_second": "Tput (tok/s)",
+    "tokens_per_second": "Tput (tok/s)",
 }
 
 # serving results and the keys that will be printed into markdown
@@ -41,16 +43,18 @@ serving_results = []
 serving_column_mapping = {
     "test_name": "Test name",
     "gpu_type": "GPU",
-    # "completed": "# of req.",
+    "completed": "# of req.",
     "request_throughput": "Tput (req/s)",
-    # "input_throughput": "Input Tput (tok/s)",
-    # "output_throughput": "Output Tput (tok/s)",
+    "total_token_throughput": "Total Token Tput (tok/s)",
+    "output_throughput": "Output Tput (tok/s)",
+    "total_input_tokens": "Total input tokens",
+    "total_output_tokens": "Total output tokens",
     "mean_ttft_ms": "Mean TTFT (ms)",
     "median_ttft_ms": "Median TTFT (ms)",
     "p99_ttft_ms": "P99 TTFT (ms)",
-    # "mean_tpot_ms": "Mean TPOT (ms)",
-    # "median_tpot_ms": "Median",
-    # "p99_tpot_ms": "P99",
+    "mean_tpot_ms": "Mean TPOT (ms)",
+    "median_tpot_ms": "Median",
+    "p99_tpot_ms": "P99",
     "mean_itl_ms": "Mean ITL (ms)",
     "median_itl_ms": "Median ITL (ms)",
     "p99_itl_ms": "P99 ITL (ms)",
@@ -75,6 +79,20 @@ def results_to_json(latency, throughput, serving):
     )
 
 
+def get_size_with_unit(bytes, suffix="B"):
+    """
+    Scale bytes to its proper format
+    e.g:
+        1253656 => '1.20MB'
+        1253656678 => '1.17GB'
+    """
+    factor = 1024
+    for unit in ["", "K", "M", "G", "T", "P"]:
+        if bytes < factor:
+            return f"{bytes:.2f}{unit}{suffix}"
+        bytes /= factor
+
+
 if __name__ == "__main__":
     # collect results
     for test_file in results_folder.glob("*.json"):
@@ -82,7 +100,7 @@ if __name__ == "__main__":
             raw_result = json.loads(f.read())
 
         if "serving" in str(test_file):
-            # this result is generated via `benchmark_serving.py`
+            # this result is generated via `vllm bench serve` command
 
             # attach the benchmarking command to raw_result
             try:
@@ -102,7 +120,7 @@ if __name__ == "__main__":
             continue
 
         elif "latency" in f.name:
-            # this result is generated via `benchmark_latency.py`
+            # this result is generated via `vllm bench latency` command
 
             # attach the benchmarking command to raw_result
             try:
@@ -130,7 +148,7 @@ if __name__ == "__main__":
             continue
 
         elif "throughput" in f.name:
-            # this result is generated via `benchmark_throughput.py`
+            # this result is generated via `vllm bench throughput` command
 
             # attach the benchmarking command to raw_result
             try:
@@ -155,6 +173,27 @@ if __name__ == "__main__":
     serving_results = pd.DataFrame.from_dict(serving_results)
     throughput_results = pd.DataFrame.from_dict(throughput_results)
 
+    svmem = psutil.virtual_memory()
+    platform_data = {
+        "Physical cores": [psutil.cpu_count(logical=False)],
+        "Total cores": [psutil.cpu_count(logical=True)],
+        "Total Memory": [get_size_with_unit(svmem.total)],
+    }
+
+    if util.find_spec("numa") is not None:
+        from numa import info
+
+        platform_data["Total NUMA nodes"] = [info.get_num_configured_nodes()]
+
+    if util.find_spec("cpuinfo") is not None:
+        from cpuinfo import get_cpu_info
+
+        platform_data["CPU Brand"] = [get_cpu_info()["brand_raw"]]
+
+    platform_results = pd.DataFrame.from_dict(
+        platform_data, orient="index", columns=["Platform Info"]
+    )
+
     raw_results_json = results_to_json(
         latency_results, throughput_results, serving_results
     )
@@ -200,6 +239,9 @@ if __name__ == "__main__":
     throughput_md_table = tabulate(
         throughput_results, headers="keys", tablefmt="pipe", showindex=False
     )
+    platform_md_table = tabulate(
+        platform_results, headers="keys", tablefmt="pipe", showindex=True
+    )
 
     # document the result
     with open(results_folder / "benchmark_results.md", "w") as f:
@@ -211,6 +253,7 @@ if __name__ == "__main__":
             latency_tests_markdown_table=latency_md_table,
             throughput_tests_markdown_table=throughput_md_table,
             serving_tests_markdown_table=serving_md_table,
+            platform_markdown_table=platform_md_table,
             benchmarking_results_in_json_string=processed_results_json,
         )
         f.write(results)

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

@@ -73,7 +73,7 @@ get_current_llm_serving_engine() {
     echo "Container: vllm"
     # move to a completely irrelevant directory, to avoid import vllm from current folder
     export CURRENT_LLM_SERVING_ENGINE=vllm
-    
+
     return
   fi
 }
@@ -95,12 +95,14 @@ json2args() {
 }
 
 kill_gpu_processes() {
-  pkill -f python
-  pkill -f python3
-  pkill -f tritonserver
-  pkill -f pt_main_thread
-  pkill -f text-generation
-  pkill -f lmdeploy
+  pkill -f '[p]ython'
+  pkill -f '[p]ython3'
+  pkill -f '[t]ritonserver'
+  pkill -f '[p]t_main_thread'
+  pkill -f '[t]ext-generation'
+  pkill -f '[l]mdeploy'
+  # vLLM now names the process with VLLM prefix after https://github.com/vllm-project/vllm/pull/21445
+  pkill -f '[V]LLM'
 
   while [ "$(nvidia-smi --query-gpu=memory.used --format=csv,noheader,nounits | head -n 1)" -ge 1000 ]; do
     sleep 1
@@ -125,7 +127,7 @@ ensure_installed() {
 }
 
 run_serving_tests() {
-  # run serving tests using `benchmark_serving.py`
+  # run serving tests using `vllm bench serve` command
   # $1: a json file specifying serving test cases
 
   local serving_test_file
@@ -225,7 +227,7 @@ run_serving_tests() {
 
       if [[ "$dataset_name" = "sharegpt" ]]; then
 
-        client_command="python3 benchmark_serving.py \
+        client_command="vllm bench serve \
           --backend $backend \
           --tokenizer /tokenizer_cache \
           --model $model \
@@ -246,7 +248,7 @@ run_serving_tests() {
         sonnet_output_len=$(echo "$common_params" | jq -r '.sonnet_output_len')
         sonnet_prefix_len=$(echo "$common_params" | jq -r '.sonnet_prefix_len')
 
-        client_command="python3 benchmark_serving.py \
+        client_command="vllm bench serve \
           --backend $backend \
           --tokenizer /tokenizer_cache \
           --model $model \
@@ -265,13 +267,13 @@ run_serving_tests() {
           $client_args"
 
       else
-  
+
         echo "The dataset name must be either 'sharegpt' or 'sonnet'. Got $dataset_name."
         exit 1
 
       fi
 
-        
+
 
       echo "Running test case $test_name with qps $qps"
       echo "Client command: $client_command"
@@ -302,7 +304,7 @@ run_serving_tests() {
 }
 
 run_genai_perf_tests() {
-  # run genai-perf tests 
+  # run genai-perf tests
 
   # $1: a json file specifying genai-perf test cases
   local genai_perf_test_file
@@ -311,14 +313,14 @@ run_genai_perf_tests() {
   # Iterate over genai-perf tests
   jq -c '.[]' "$genai_perf_test_file" | while read -r params; do
     # get the test name, and append the GPU type back to it.
-    test_name=$(echo "$params" | jq -r '.test_name')    
-    
+    test_name=$(echo "$params" | jq -r '.test_name')
+
     # if TEST_SELECTOR is set, only run the test cases that match the selector
     if [[ -n "$TEST_SELECTOR" ]] && [[ ! "$test_name" =~ $TEST_SELECTOR ]]; then
       echo "Skip test case $test_name."
       continue
     fi
-    
+
     # prepend the current serving engine to the test name
     test_name=${CURRENT_LLM_SERVING_ENGINE}_${test_name}
 
@@ -369,10 +371,10 @@ run_genai_perf_tests() {
         qps=$num_prompts
         echo "now qps is $qps"
       fi
-    
+
       new_test_name=$test_name"_qps_"$qps
       backend=$CURRENT_LLM_SERVING_ENGINE
-      
+
       if [[ "$backend" == *"vllm"* ]]; then
         backend="vllm"
       fi
@@ -413,7 +415,7 @@ prepare_dataset() {
   do
     cat sonnet.txt >> sonnet_4x.txt
   done
-  
+
 }
 
 main() {

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

@@ -31,6 +31,20 @@ check_gpus() {
   echo "GPU type is $gpu_type"
 }
 
+check_cpus() {
+  # check the number of CPUs and NUMA Node and GPU type.
+  declare -g numa_count=$(python3 -c  "from numa import info;numa_size = info.get_num_configured_nodes(); print(numa_size)")
+  if [[ $numa_count -gt 0 ]]; then
+    echo "NUMA found."
+    echo $numa_count
+  else
+    echo "Need at least 1 NUMA to run benchmarking."
+    exit 1
+  fi
+  declare -g gpu_type="cpu"
+  echo "GPU type is $gpu_type"
+}
+
 check_hf_token() {
   # check if HF_TOKEN is available and valid
   if [[ -z "$HF_TOKEN" ]]; then
@@ -69,6 +83,22 @@ json2args() {
   echo "$args"
 }
 
+json2envs() {
+  # transforms the JSON string to environment variables.
+  # example:
+  # input: { "VLLM_CPU_KVCACHE_SPACE": 5 }
+  # output: VLLM_CPU_KVCACHE_SPACE=5
+  local json_string=$1
+  local args=$(
+    echo "$json_string" | jq -r '
+      to_entries |
+      map((.key ) + "=" + (.value | tostring)) |
+      join(" ")
+    '
+  )
+  echo "$args"
+}
+
 wait_for_server() {
   # wait for vllm server to start
   # return 1 if vllm server crashes
@@ -96,7 +126,8 @@ kill_gpu_processes() {
   ps -aux
   lsof -t -i:8000 | xargs -r kill -9
   pgrep python3 | xargs -r kill -9
-
+  # vLLM now names the process with VLLM prefix after https://github.com/vllm-project/vllm/pull/21445
+  pgrep VLLM | xargs -r kill -9
 
   # wait until GPU memory usage smaller than 1GB
   if command -v nvidia-smi; then
@@ -134,7 +165,7 @@ upload_to_buildkite() {
 }
 
 run_latency_tests() {
-  # run latency tests using `benchmark_latency.py`
+  # run latency tests using `vllm bench latency` command
   # $1: a json file specifying latency test cases
 
   local latency_test_file
@@ -158,15 +189,24 @@ run_latency_tests() {
     # get arguments
     latency_params=$(echo "$params" | jq -r '.parameters')
     latency_args=$(json2args "$latency_params")
+    latency_environment_variables=$(echo "$params" | jq -r '.environment_variables')
+    latency_envs=$(json2envs "$latency_environment_variables")
 
     # check if there is enough GPU to run the test
     tp=$(echo "$latency_params" | jq -r '.tensor_parallel_size')
-    if [[ $gpu_count -lt $tp ]]; then
-      echo "Required tensor-parallel-size $tp but only $gpu_count GPU found. Skip testcase $test_name."
-      continue
+    if [ "$ON_CPU" == "1" ];then
+      if [[ $numa_count -lt $tp ]]; then
+        echo "Required tensor-parallel-size $tp but only $numa_count NUMA nodes found. Skip testcase $test_name."
+        continue
+      fi
+    else
+      if [[ $gpu_count -lt $tp ]]; then
+        echo "Required tensor-parallel-size $tp but only $gpu_count GPU found. Skip testcase $test_name."
+        continue
+      fi
     fi
 
-    latency_command="python3 benchmark_latency.py \
+    latency_command=" $latency_envs vllm bench latency \
       --output-json $RESULTS_FOLDER/${test_name}.json \
       $latency_args"
 
@@ -192,7 +232,7 @@ run_latency_tests() {
 }
 
 run_throughput_tests() {
-  # run throughput tests using `benchmark_throughput.py`
+  # run throughput tests using `vllm bench throughput`
   # $1: a json file specifying throughput test cases
 
   local throughput_test_file
@@ -216,15 +256,24 @@ run_throughput_tests() {
     # get arguments
     throughput_params=$(echo "$params" | jq -r '.parameters')
     throughput_args=$(json2args "$throughput_params")
+    throughput_environment_variables=$(echo "$params" | jq -r '.environment_variables')
+    throughput_envs=$(json2envs "$throughput_environment_variables")
 
     # check if there is enough GPU to run the test
     tp=$(echo "$throughput_params" | jq -r '.tensor_parallel_size')
-    if [[ $gpu_count -lt $tp ]]; then
-      echo "Required tensor-parallel-size $tp but only $gpu_count GPU found. Skip testcase $test_name."
-      continue
+    if [ "$ON_CPU" == "1" ];then
+      if [[ $numa_count -lt $tp ]]; then
+        echo "Required tensor-parallel-size $tp but only $numa_count NUMA nodes found. Skip testcase $test_name."
+        continue
+      fi
+    else
+      if [[ $gpu_count -lt $tp ]]; then
+        echo "Required tensor-parallel-size $tp but only $gpu_count GPU found. Skip testcase $test_name."
+        continue
+      fi
     fi
 
-    throughput_command="python3 benchmark_throughput.py \
+    throughput_command=" $throughput_envs vllm bench throughput \
       --output-json $RESULTS_FOLDER/${test_name}.json \
       $throughput_args"
 
@@ -249,7 +298,7 @@ run_throughput_tests() {
 }
 
 run_serving_tests() {
-  # run serving tests using `benchmark_serving.py`
+  # run serving tests using `vllm bench serve` command
   # $1: a json file specifying serving test cases
 
   local serving_test_file
@@ -272,18 +321,27 @@ run_serving_tests() {
 
     # get client and server arguments
     server_params=$(echo "$params" | jq -r '.server_parameters')
+    server_envs=$(echo "$params" | jq -r '.server_environment_variables')
     client_params=$(echo "$params" | jq -r '.client_parameters')
     server_args=$(json2args "$server_params")
+    server_envs=$(json2envs "$server_envs")
     client_args=$(json2args "$client_params")
     qps_list=$(echo "$params" | jq -r '.qps_list')
     qps_list=$(echo "$qps_list" | jq -r '.[] | @sh')
     echo "Running over qps list $qps_list"
 
-    # check if there is enough GPU to run the test
+    # check if there is enough resources to run the test
     tp=$(echo "$server_params" | jq -r '.tensor_parallel_size')
-    if [[ $gpu_count -lt $tp ]]; then
-      echo "Required tensor-parallel-size $tp but only $gpu_count GPU found. Skip testcase $test_name."
-      continue
+    if [ "$ON_CPU" == "1" ];then
+      if [[ $numa_count -lt $tp ]]; then
+        echo "Required tensor-parallel-size $tp but only $numa_count NUMA nodes found. Skip testcase $test_name."
+        continue
+      fi
+    else
+      if [[ $gpu_count -lt $tp ]]; then
+        echo "Required tensor-parallel-size $tp but only $gpu_count GPU found. Skip testcase $test_name."
+        continue
+      fi
     fi
 
     # check if server model and client model is aligned
@@ -294,23 +352,33 @@ run_serving_tests() {
       continue
     fi
 
-    server_command="python3 \
+    server_command="$server_envs python3 \
       -m vllm.entrypoints.openai.api_server \
       $server_args"
 
     # run the server
     echo "Running test case $test_name"
     echo "Server command: $server_command"
-    bash -c "$server_command" &
-    server_pid=$!
-
-    # wait until the server is alive
-    if wait_for_server; then
-      echo ""
-      echo "vllm server is up and running."
+    # support remote vllm server
+    client_remote_args=""
+    if [[ -z "${REMOTE_HOST}" ]]; then
+      bash -c "$server_command" &
+      server_pid=$!
+      # wait until the server is alive
+      if wait_for_server; then
+        echo ""
+        echo "vLLM server is up and running."
+      else
+        echo ""
+        echo "vLLM failed to start within the timeout period."
+      fi
     else
-      echo ""
-      echo "vllm failed to start within the timeout period."
+      server_command="Using Remote Server $REMOTE_HOST $REMOTE_PORT"
+      if [[ ${REMOTE_PORT} ]]; then
+        client_remote_args=" --host=$REMOTE_HOST --port=$REMOTE_PORT "
+      else
+        client_remote_args=" --host=$REMOTE_HOST "
+      fi
     fi
 
     # iterate over different QPS
@@ -326,13 +394,13 @@ run_serving_tests() {
 
       # pass the tensor parallel size to the client so that it can be displayed
       # on the benchmark dashboard
-      client_command="python3 benchmark_serving.py \
+      client_command="vllm bench serve \
         --save-result \
         --result-dir $RESULTS_FOLDER \
         --result-filename ${new_test_name}.json \
         --request-rate $qps \
         --metadata "tensor_parallel_size=$tp" \
-        $client_args"
+        $client_args $client_remote_args "
 
       echo "Running test case $test_name with qps $qps"
       echo "Client command: $client_command"
@@ -360,7 +428,14 @@ run_serving_tests() {
 }
 
 main() {
-  check_gpus
+  local ARCH
+  ARCH=''
+  if [ "$ON_CPU" == "1" ];then
+     check_cpus
+     ARCH='-cpu'
+  else
+     check_gpus
+  fi
   check_hf_token
 
   # Set to v1 to run v1 benchmark
@@ -373,7 +448,7 @@ main() {
   (which jq) || (apt-get update && apt-get -y install jq)
   (which lsof) || (apt-get update && apt-get install -y lsof)
 
-  # get the current IP address, required by benchmark_serving.py
+  # get the current IP address, required by `vllm bench serve` command
   export VLLM_HOST_IP=$(hostname -I | awk '{print $1}')
   # turn of the reporting of the status of each request, to clean up the terminal output
   export VLLM_LOGGING_LEVEL="WARNING"
@@ -386,9 +461,9 @@ main() {
   QUICK_BENCHMARK_ROOT=../.buildkite/nightly-benchmarks/
 
   # benchmarking
-  run_serving_tests $QUICK_BENCHMARK_ROOT/tests/serving-tests.json
-  run_latency_tests $QUICK_BENCHMARK_ROOT/tests/latency-tests.json
-  run_throughput_tests $QUICK_BENCHMARK_ROOT/tests/throughput-tests.json
+  run_serving_tests $QUICK_BENCHMARK_ROOT/tests/"${SERVING_JSON:-serving-tests$ARCH.json}"
+  run_latency_tests $QUICK_BENCHMARK_ROOT/tests/"${LATENCY_JSON:-latency-tests$ARCH.json}"
+  run_throughput_tests $QUICK_BENCHMARK_ROOT/tests/"${THROUGHPUT_JSON:-throughput-tests$ARCH.json}"
 
   # postprocess benchmarking results
   pip install tabulate pandas

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

@@ -0,0 +1,30 @@
+[
+    {
+        "test_name": "latency_llama8B_tp1",
+        "environment_variables": {
+	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
+	    "VLLM_CPU_KVCACHE_SPACE": 40
+        },
+        "parameters": {
+            "model": "meta-llama/Meta-Llama-3.1-8B-Instruct",
+            "tensor_parallel_size": 1,
+            "load_format": "dummy",
+            "num_iters_warmup": 5,
+            "num_iters": 15
+        }
+    },
+    {
+        "test_name": "latency_llama8B_tp4",
+        "environment_variables": {
+	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
+	    "VLLM_CPU_KVCACHE_SPACE": 40
+        },
+        "parameters": {
+            "model": "meta-llama/Meta-Llama-3.1-8B-Instruct",
+            "tensor_parallel_size": 4,
+            "load_format": "dummy",
+            "num_iters_warmup": 5,
+            "num_iters": 15
+        }
+    }
+]

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

@@ -0,0 +1,158 @@
+[
+    {
+        "test_name": "serving_llama8B_tp1_sharegpt",
+        "qps_list": [1, 4, 16, "inf"],
+        "server_environment_variables": {
+            "VLLM_RPC_TIMEOUT": 100000,
+	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
+	    "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
+	    "VLLM_CPU_KVCACHE_SPACE": 40
+        },
+        "server_parameters": {
+            "model": "meta-llama/Meta-Llama-3.1-8B-Instruct",
+            "tensor_parallel_size": 1,
+	    "dtype": "bfloat16",
+	    "distributed_executor_backend": "mp",
+	    "block_size": 128,
+	    "trust_remote_code": "",
+            "disable_log_stats": "",
+            "disable_log_requests": "",
+	    "enforce_eager": "",
+            "load_format": "dummy"
+        },
+        "client_parameters": {
+            "model": "meta-llama/Meta-Llama-3.1-8B-Instruct",
+            "backend": "vllm",
+            "dataset_name": "sharegpt",
+            "dataset_path": "./ShareGPT_V3_unfiltered_cleaned_split.json",
+	    "max_concurrency": 60,
+            "num_prompts": 200
+        }
+    },
+    {
+        "test_name": "serving_llama8B_tp2_sharegpt",
+        "qps_list": [1, 4, 16, "inf"],
+        "server_environment_variables": {
+            "VLLM_RPC_TIMEOUT": 100000,
+	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
+	    "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
+	    "VLLM_CPU_KVCACHE_SPACE": 40
+        },
+        "server_parameters": {
+            "model": "meta-llama/Meta-Llama-3.1-8B-Instruct",
+            "tensor_parallel_size": 2,
+	    "dtype": "bfloat16",
+	    "distributed_executor_backend": "mp",
+	    "block_size": 128,
+	    "trust_remote_code": "",
+            "disable_log_stats": "",
+            "disable_log_requests": "",
+	    "enforce_eager": "",
+            "load_format": "dummy"
+        },
+        "client_parameters": {
+            "model": "meta-llama/Meta-Llama-3.1-8B-Instruct",
+            "backend": "vllm",
+            "dataset_name": "sharegpt",
+            "dataset_path": "./ShareGPT_V3_unfiltered_cleaned_split.json",
+	    "max_concurrency": 60,
+            "num_prompts": 200
+        }
+    },
+    {
+        "test_name": "serving_llama8B_tp4_sharegpt",
+        "qps_list": [1, 4, 16, "inf"],
+        "server_environment_variables": {
+            "VLLM_RPC_TIMEOUT": 100000,
+	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
+	    "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
+	    "VLLM_CPU_KVCACHE_SPACE": 40
+        },
+        "server_parameters": {
+            "model": "meta-llama/Meta-Llama-3.1-8B-Instruct",
+            "tensor_parallel_size": 4,
+	    "dtype": "bfloat16",
+	    "distributed_executor_backend": "mp",
+	    "block_size": 128,
+	    "trust_remote_code": "",
+            "disable_log_stats": "",
+            "disable_log_requests": "",
+	    "enforce_eager": "",
+            "load_format": "dummy"
+        },
+        "client_parameters": {
+            "model": "meta-llama/Meta-Llama-3.1-8B-Instruct",
+            "backend": "vllm",
+            "dataset_name": "sharegpt",
+            "dataset_path": "./ShareGPT_V3_unfiltered_cleaned_split.json",
+	    "max_concurrency": 60,
+            "num_prompts": 200
+        }
+    },
+    {
+        "test_name": "serving_llama8B_tp4_random_1024_128",
+        "qps_list": [1, 4, 16, "inf"],
+        "server_environment_variables": {
+            "VLLM_RPC_TIMEOUT": 100000,
+	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
+	    "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
+	    "VLLM_CPU_KVCACHE_SPACE": 40
+        },
+        "server_parameters": {
+            "model": "meta-llama/Meta-Llama-3.1-8B-Instruct",
+            "tensor_parallel_size": 4,
+	    "dtype": "bfloat16",
+	    "distributed_executor_backend": "mp",
+	    "block_size": 128,
+	    "trust_remote_code": "",
+	    "enable_chunked_prefill": "",
+            "disable_log_stats": "",
+            "disable_log_requests": "",
+	    "enforce_eager": "",
+            "load_format": "dummy"
+        },
+        "client_parameters": {
+            "model": "meta-llama/Meta-Llama-3.1-8B-Instruct",
+            "backend": "vllm",
+            "dataset_name": "random",
+	    "random-input-len": 1024,
+	    "random-output-len": 128,
+	    "ignore-eos": "",
+	    "max_concurrency": 100,
+            "num_prompts": 100
+        }
+    },
+    {
+        "test_name": "serving_llama8B_pp6_random_1024_128",
+        "qps_list": [1, 4, 16, "inf"],
+        "server_environment_variables": {
+            "VLLM_RPC_TIMEOUT": 100000,
+	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
+	    "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
+	    "VLLM_CPU_KVCACHE_SPACE": 40
+        },
+        "server_parameters": {
+            "model": "meta-llama/Meta-Llama-3.1-8B-Instruct",
+            "pipeline_parallel_size": 6,
+	    "dtype": "bfloat16",
+	    "distributed_executor_backend": "mp",
+	    "block_size": 128,
+	    "trust_remote_code": "",
+	    "enable_chunked_prefill": "",
+            "disable_log_stats": "",
+            "disable_log_requests": "",
+	    "enforce_eager": "",
+            "load_format": "dummy"
+        },
+        "client_parameters": {
+            "model": "meta-llama/Meta-Llama-3.1-8B-Instruct",
+            "backend": "vllm",
+            "dataset_name": "random",
+	    "random-input-len": 1024,
+	    "random-output-len": 128,
+	    "ignore-eos": "",
+	    "max_concurrency": 100,
+            "num_prompts": 100
+        }
+    }
+]

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

@@ -0,0 +1,32 @@
+[
+    {
+        "test_name": "throughput_llama8B_tp1",
+        "environment_variables": {
+	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
+	    "VLLM_CPU_KVCACHE_SPACE": 40
+        },
+        "parameters": {
+            "model": "meta-llama/Meta-Llama-3.1-8B-Instruct",
+            "tensor_parallel_size": 1,
+            "load_format": "dummy",
+            "dataset": "./ShareGPT_V3_unfiltered_cleaned_split.json",
+            "num_prompts": 200,
+            "backend": "vllm"
+        }
+    },
+    {
+        "test_name": "throughput_llama8B_tp4",
+        "environment_variables": {
+	    "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
+	    "VLLM_CPU_KVCACHE_SPACE": 40
+        },
+        "parameters": {
+            "model": "meta-llama/Meta-Llama-3.1-8B-Instruct",
+            "tensor_parallel_size": 4,
+            "load_format": "dummy",
+            "dataset": "./ShareGPT_V3_unfiltered_cleaned_split.json",
+            "num_prompts": 200,
+            "backend": "vllm"
+        }
+    }
+]

.buildkite/release-pipeline.yaml

@@ -52,7 +52,7 @@ steps:
       queue: cpu_queue_postmerge
     commands:
       - "aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin public.ecr.aws/q9t5s3a7"
-      - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=12.8.1 --tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT --target vllm-openai --progress plain -f docker/Dockerfile ."
+      - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=12.8.1 --build-arg INSTALL_KV_CONNECTORS=true --tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT --target vllm-openai --progress plain -f docker/Dockerfile ."
       - "docker push public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT"
 
   - label: "Annotate release workflow"
@@ -101,7 +101,7 @@ steps:
       queue: cpu_queue_postmerge
     commands:
       - "aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin public.ecr.aws/q9t5s3a7"
-      - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg GIT_REPO_CHECK=1 --tag public.ecr.aws/q9t5s3a7/vllm-cpu-release-repo:$(buildkite-agent meta-data get release-version) --tag public.ecr.aws/q9t5s3a7/vllm-cpu-release-repo:latest --progress plain --target vllm-openai -f docker/Dockerfile.cpu ."
+      - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg GIT_REPO_CHECK=1 --build-arg VLLM_CPU_AVX512BF16=true --build-arg VLLM_CPU_AVX512VNNI=true --tag public.ecr.aws/q9t5s3a7/vllm-cpu-release-repo:$(buildkite-agent meta-data get release-version) --tag public.ecr.aws/q9t5s3a7/vllm-cpu-release-repo:latest --progress plain --target vllm-openai -f docker/Dockerfile.cpu ."
       - "docker push public.ecr.aws/q9t5s3a7/vllm-cpu-release-repo:latest"
       - "docker push public.ecr.aws/q9t5s3a7/vllm-cpu-release-repo:$(buildkite-agent meta-data get release-version)"
     env:

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

@@ -107,10 +107,8 @@ fi
 
 if [[ $commands == *" kernels/attention"* ]]; then
   commands="${commands} \
-  --ignore=kernels/attention/stest_attention_selector.py \
-  --ignore=kernels/attention/test_blocksparse_attention.py \
-  --ignore=kernels/attention/test_encoder_decoder_attn.py \
   --ignore=kernels/attention/test_attention_selector.py \
+  --ignore=kernels/attention/test_encoder_decoder_attn.py \
   --ignore=kernels/attention/test_flash_attn.py \
   --ignore=kernels/attention/test_flashinfer.py \
   --ignore=kernels/attention/test_prefix_prefill.py \

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

@@ -6,15 +6,16 @@ set -ex
 
 # allow to bind to different cores
 CORE_RANGE=${CORE_RANGE:-48-95}
+# used for TP/PP E2E test
 OMP_CORE_RANGE=${OMP_CORE_RANGE:-48-95}
 NUMA_NODE=${NUMA_NODE:-1}
 
 export CMAKE_BUILD_PARALLEL_LEVEL=32
 
 # Setup cleanup
-remove_docker_container() { 
-    set -e; 
-    docker rm -f cpu-test-"$NUMA_NODE" cpu-test-"$NUMA_NODE"-avx2 || true; 
+remove_docker_container() {
+    set -e;
+    docker rm -f cpu-test-"$NUMA_NODE" cpu-test-"$NUMA_NODE"-avx2 || true;
 }
 trap remove_docker_container EXIT
 remove_docker_container
@@ -24,8 +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_OMP_THREADS_BIND="$OMP_CORE_RANGE" --env VLLM_CPU_CI_ENV=1 --shm-size=4g --name cpu-test-"$NUMA_NODE" cpu-test-"$NUMA_NODE"
-docker run -itd --cpuset-cpus="$CORE_RANGE" --cpuset-mems="$NUMA_NODE" --entrypoint /bin/bash -v ~/.cache/huggingface:/root/.cache/huggingface --privileged=true -e HF_TOKEN --env VLLM_CPU_KVCACHE_SPACE=4 --env VLLM_CPU_OMP_THREADS_BIND="$OMP_CORE_RANGE" --env VLLM_CPU_CI_ENV=1 --shm-size=4g --name cpu-test-"$NUMA_NODE"-avx2 cpu-test-"$NUMA_NODE"-avx2
+docker run -itd --cpuset-cpus="$CORE_RANGE" --cpuset-mems="$NUMA_NODE" --entrypoint /bin/bash -v ~/.cache/huggingface:/root/.cache/huggingface --privileged=true -e HF_TOKEN --env VLLM_CPU_KVCACHE_SPACE=4 --env VLLM_CPU_CI_ENV=1 -e E2E_OMP_THREADS="$OMP_CORE_RANGE" --shm-size=4g --name cpu-test-"$NUMA_NODE" cpu-test-"$NUMA_NODE"
+docker run -itd --cpuset-cpus="$CORE_RANGE" --cpuset-mems="$NUMA_NODE" --entrypoint /bin/bash -v ~/.cache/huggingface:/root/.cache/huggingface --privileged=true -e HF_TOKEN --env VLLM_CPU_KVCACHE_SPACE=4 --env VLLM_CPU_CI_ENV=1 -e E2E_OMP_THREADS="$OMP_CORE_RANGE" --shm-size=4g --name cpu-test-"$NUMA_NODE"-avx2 cpu-test-"$NUMA_NODE"-avx2
 
 function cpu_tests() {
   set -e
@@ -48,9 +49,16 @@ function cpu_tests() {
   # Run basic model test
   docker exec cpu-test-"$NUMA_NODE" bash -c "
     set -e
-    pytest -v -s tests/kernels/attention/test_cache.py -m cpu_model
-    pytest -v -s tests/kernels/attention/test_mla_decode_cpu.py -m cpu_model
-    pytest -v -s tests/models/language/generation -m cpu_model
+    # Note: disable until supports V1
+    # pytest -v -s tests/kernels/attention/test_cache.py -m cpu_model
+    # pytest -v -s tests/kernels/attention/test_mla_decode_cpu.py -m cpu_model
+
+    # Note: disable Bart until supports V1
+    pytest -v -s tests/models/language/generation -m cpu_model \
+                --ignore=tests/models/language/generation/test_bart.py
+    VLLM_CPU_SGL_KERNEL=1 pytest -v -s tests/models/language/generation -m cpu_model \
+                --ignore=tests/models/language/generation/test_bart.py
+
     pytest -v -s tests/models/language/pooling -m cpu_model
     pytest -v -s tests/models/multimodal/generation \
                 --ignore=tests/models/multimodal/generation/test_mllama.py \
@@ -61,33 +69,26 @@ function cpu_tests() {
   docker exec cpu-test-"$NUMA_NODE" bash -c "
     set -e
     pytest -s -v \
-    tests/quantization/test_compressed_tensors.py::test_compressed_tensors_w8a8_static_setup \
-    tests/quantization/test_compressed_tensors.py::test_compressed_tensors_w8a8_dynamic_per_token"
+    tests/quantization/test_compressed_tensors.py::test_compressed_tensors_w8a8_logprobs[False-10-32-neuralmagic/Llama-3.2-1B-quantized.w8a8]"
 
+  # Note: disable it until supports V1
   # Run AWQ test
-  docker exec cpu-test-"$NUMA_NODE" bash -c "
-    set -e
-    VLLM_USE_V1=0 pytest -s -v \
-    tests/quantization/test_ipex_quant.py"
-
-  # Run chunked-prefill and prefix-cache test
-  docker exec cpu-test-"$NUMA_NODE" bash -c "
-    set -e
-    pytest -s -v -k cpu_model \
-    tests/basic_correctness/test_chunked_prefill.py"  
+  # docker exec cpu-test-"$NUMA_NODE" bash -c "
+  #   set -e
+  #   VLLM_USE_V1=0 pytest -s -v \
+  #   tests/quantization/test_ipex_quant.py"
 
   # online serving
-  docker exec cpu-test-"$NUMA_NODE" bash -c "
+  docker exec cpu-test-"$NUMA_NODE" bash -c '
     set -e
-    python3 -m vllm.entrypoints.openai.api_server --model facebook/opt-125m --dtype half & 
-    timeout 600 bash -c 'until curl localhost:8000/v1/models; do sleep 1; done' || exit 1
-    VLLM_CPU_CI_ENV=0 python3 benchmarks/benchmark_serving.py \
+    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 &
+    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 facebook/opt-125m \
+      --model meta-llama/Llama-3.2-3B-Instruct \
       --num-prompts 20 \
-      --endpoint /v1/completions \
-      --tokenizer facebook/opt-125m"
+      --endpoint /v1/completions'
 
   # Run multi-lora tests
   docker exec cpu-test-"$NUMA_NODE" bash -c "
@@ -98,4 +99,4 @@ function cpu_tests() {
 
 # All of CPU tests are expected to be finished less than 40 mins.
 export -f cpu_tests
-timeout 1h bash -c "cpu_tests $CORE_RANGE $NUMA_NODE"
+timeout 1.5h bash -c "cpu_tests $CORE_RANGE $NUMA_NODE"

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

@@ -2,10 +2,32 @@
 
 # This script build the CPU docker image and run the offline inference inside the container.
 # It serves a sanity check for compilation and basic model usage.
-set -ex
+set -exuo pipefail
 
 # Try building the docker image
-docker build -t hpu-test-env -f docker/Dockerfile.hpu .
+cat <<EOF | docker build -t hpu-plugin-v1-test-env -f - .
+FROM gaudi-base-image:latest
+
+COPY ./ /workspace/vllm
+
+WORKDIR /workspace/vllm
+
+ENV no_proxy=localhost,127.0.0.1
+ENV PT_HPU_ENABLE_LAZY_COLLECTIVES=true
+
+RUN VLLM_TARGET_DEVICE=empty pip install .
+RUN pip install git+https://github.com/vllm-project/vllm-gaudi.git
+
+# install development dependencies (for testing)
+RUN python3 -m pip install -e tests/vllm_test_utils
+
+WORKDIR /workspace/
+
+RUN git clone https://github.com/vllm-project/vllm-gaudi.git
+
+RUN ln -s /workspace/vllm/tests && ln -s /workspace/vllm/examples && ln -s /workspace/vllm/benchmarks
+
+EOF
 
 # Setup cleanup
 # certain versions of HPU software stack have a bug that can
@@ -14,13 +36,21 @@ docker build -t hpu-test-env -f docker/Dockerfile.hpu .
 # functions, while other platforms only need one remove_docker_container
 # function.
 EXITCODE=1
-remove_docker_containers() { docker rm -f hpu-test || true; docker rm -f hpu-test-tp2 || true; }
-remove_docker_containers_and_exit() { remove_docker_containers; exit $EXITCODE; }
-trap remove_docker_containers_and_exit EXIT
+remove_docker_containers() { docker rm -f hpu-plugin-v1-test || true; }
+trap 'remove_docker_containers; exit $EXITCODE;' EXIT
 remove_docker_containers
 
-# Run the image and launch offline inference
-docker run --runtime=habana --name=hpu-test --network=host -e HABANA_VISIBLE_DEVICES=all -e VLLM_SKIP_WARMUP=true --entrypoint="" hpu-test-env python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m
-docker run --runtime=habana --name=hpu-test-tp2 --network=host -e HABANA_VISIBLE_DEVICES=all -e VLLM_SKIP_WARMUP=true --entrypoint="" hpu-test-env python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m --tensor-parallel-size 2
+echo "Running HPU plugin v1 test"
+docker run --rm --runtime=habana --name=hpu-plugin-v1-test --network=host \
+  -e HABANA_VISIBLE_DEVICES=all \
+  hpu-plugin-v1-test-env \
+  /bin/bash "/workspace/vllm-gaudi/tests/upstream_tests/ci_tests.sh"
 
 EXITCODE=$?
+if [ $EXITCODE -eq 0 ]; then
+  echo "Test with basic model passed"
+else
+  echo "Test with basic model FAILED with exit code: $EXITCODE" >&2
+fi
+
+# The trap will handle the container removal and final exit.

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

@@ -0,0 +1,166 @@
+#!/bin/bash
+
+set -xu
+
+
+remove_docker_container() { 
+    docker rm -f tpu-test || true; 
+    docker rm -f vllm-tpu || true;
+}
+
+trap remove_docker_container EXIT
+
+# Remove the container that might not be cleaned up in the previous run.
+remove_docker_container
+
+# Build the docker image.
+docker build -f docker/Dockerfile.tpu -t vllm-tpu .
+
+# Set up cleanup.
+cleanup_docker() {
+  # Get Docker's root directory
+  docker_root=$(docker info -f '{{.DockerRootDir}}')
+  if [ -z "$docker_root" ]; then
+    echo "Failed to determine Docker root directory."
+    exit 1
+  fi
+  echo "Docker root directory: $docker_root"
+  # Check disk usage of the filesystem where Docker's root directory is located
+  disk_usage=$(df "$docker_root" | tail -1 | awk '{print $5}' | sed 's/%//')
+  # Define the threshold
+  threshold=70
+  if [ "$disk_usage" -gt "$threshold" ]; then
+    echo "Disk usage is above $threshold%. Cleaning up Docker images and volumes..."
+    # Remove dangling images (those that are not tagged and not used by any container)
+    docker image prune -f
+    # Remove unused volumes / force the system prune for old images as well.
+    docker volume prune -f && docker system prune --force --filter "until=72h" --all
+    echo "Docker images and volumes cleanup completed."
+  else
+    echo "Disk usage is below $threshold%. No cleanup needed."
+  fi
+}
+cleanup_docker
+
+# For HF_TOKEN.
+source /etc/environment
+
+docker run --privileged --net host --shm-size=16G -it \
+    -e "HF_TOKEN=$HF_TOKEN" --name tpu-test \
+    vllm-tpu /bin/bash -c '
+set -e # Exit immediately if a command exits with a non-zero status.
+set -u # Treat unset variables as an error.
+
+echo "--- Starting script inside Docker container ---"
+
+# Create results directory
+RESULTS_DIR=$(mktemp -d)
+# If mktemp fails, set -e will cause the script to exit.
+echo "Results will be stored in: $RESULTS_DIR"
+
+# Install dependencies
+echo "--- Installing Python dependencies ---"
+python3 -m pip install --progress-bar off git+https://github.com/thuml/depyf.git \
+    && python3 -m pip install --progress-bar off pytest pytest-asyncio tpu-info \
+    && python3 -m pip install --progress-bar off lm_eval[api]==0.4.4 \
+    && python3 -m pip install --progress-bar off hf-transfer
+echo "--- Python dependencies installed ---"
+export VLLM_USE_V1=1
+export VLLM_XLA_CHECK_RECOMPILATION=1
+export VLLM_XLA_CACHE_PATH=
+echo "Using VLLM V1"
+
+echo "--- Hardware Information ---"
+# tpu-info
+echo "--- Starting Tests ---"
+set +e
+overall_script_exit_code=0
+
+# --- Test Definitions ---
+# If a test fails, this function will print logs and will not cause the main script to exit.
+run_test() {
+    local test_num=$1
+    local test_name=$2
+    local test_command=$3
+    local log_file="$RESULTS_DIR/test_${test_num}.log"
+    local actual_exit_code
+
+    echo "--- TEST_$test_num: Running $test_name ---"
+    
+    # Execute the test command.
+    eval "$test_command" > >(tee -a "$log_file") 2> >(tee -a "$log_file" >&2)
+    actual_exit_code=$?
+
+    echo "TEST_${test_num}_COMMAND_EXIT_CODE: $actual_exit_code" # This goes to main log
+    echo "TEST_${test_num}_COMMAND_EXIT_CODE: $actual_exit_code" >> "$log_file" # Also to per-test log
+
+    if [ "$actual_exit_code" -ne 0 ]; then
+        echo "TEST_$test_num ($test_name) FAILED with exit code $actual_exit_code." >&2
+        echo "--- Log for failed TEST_$test_num ($test_name) ---" >&2
+        if [ -f "$log_file" ]; then
+            cat "$log_file" >&2
+        else
+            echo "Log file $log_file not found for TEST_$test_num ($test_name)." >&2
+        fi
+        echo "--- End of log for TEST_$test_num ($test_name) ---" >&2
+        return "$actual_exit_code" # Return the failure code
+    else
+        echo "TEST_$test_num ($test_name) PASSED."
+        return 0 # Return success
+    fi
+}
+
+# Helper function to call run_test and update the overall script exit code
+run_and_track_test() {
+    local test_num_arg="$1"
+    local test_name_arg="$2"
+    local test_command_arg="$3"
+
+    # Run the test
+    run_test "$test_num_arg" "$test_name_arg" "$test_command_arg"
+    local test_specific_exit_code=$?
+
+    # If the test failed, set the overall script exit code to 1
+    if [ "$test_specific_exit_code" -ne 0 ]; then
+        # No need for extra echo here, run_test already logged the failure.
+        overall_script_exit_code=1
+    fi
+}
+
+# --- Actual Test Execution ---
+run_and_track_test 1 "test_struct_output_generate.py" \
+    "HF_HUB_DISABLE_XET=1 python3 -m pytest -s -v /workspace/vllm/tests/v1/entrypoints/llm/test_struct_output_generate.py -k \"not test_structured_output_with_reasoning_matrices\""
+run_and_track_test 2 "test_moe_pallas.py" \
+    "python3 -m pytest -s -v /workspace/vllm/tests/tpu/test_moe_pallas.py"
+run_and_track_test 3 "test_lora.py" \
+    "VLLM_XLA_CHECK_RECOMPILATION=0 python3 -m pytest -s -v /workspace/vllm/tests/tpu/lora/test_lora.py"
+run_and_track_test 4 "test_tpu_qkv_linear.py" \
+    "python3 -m pytest -s -v /workspace/vllm/tests/v1/tpu/test_tpu_qkv_linear.py"
+run_and_track_test 5 "test_spmd_model_weight_loading.py" \
+    "python3 -m pytest -s -v /workspace/vllm/tests/v1/tpu/test_spmd_model_weight_loading.py"
+run_and_track_test 6 "test_kv_cache_update_kernel.py" \
+    "python3 -m pytest -s -v /workspace/vllm/tests/v1/tpu/test_kv_cache_update_kernel.py"
+
+# After all tests have been attempted, exit with the overall status.
+if [ "$overall_script_exit_code" -ne 0 ]; then
+    echo "--- One or more tests FAILED. Overall script exiting with failure code 1. ---"
+else
+    echo "--- All tests have completed and PASSED. Overall script exiting with success code 0. ---"
+fi
+exit "$overall_script_exit_code"
+' # IMPORTANT: This is the closing single quote for the bash -c "..." command. Ensure it is present and correct.
+
+# Capture the exit code of the docker run command
+DOCKER_RUN_EXIT_CODE=$?
+
+# The trap will run for cleanup.
+# Exit the main script with the Docker run command's exit code.
+if [ "$DOCKER_RUN_EXIT_CODE" -ne 0 ]; then
+    echo "Docker run command failed with exit code $DOCKER_RUN_EXIT_CODE."
+    exit "$DOCKER_RUN_EXIT_CODE"
+else
+    echo "Docker run command completed successfully."
+    exit 0
+fi
+# TODO: This test fails because it uses RANDOM_SEED sampling
+# pytest -v -s /workspace/vllm/tests/tpu/test_custom_dispatcher.py \

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

@@ -62,7 +62,8 @@ 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[api]==0.4.4 \
+    && python3 -m pip install --progress-bar off hf-transfer
 echo "--- Python dependencies installed ---"
 export VLLM_USE_V1=1
 export VLLM_XLA_CHECK_RECOMPILATION=1
@@ -70,7 +71,7 @@ export VLLM_XLA_CACHE_PATH=
 echo "Using VLLM V1"
 
 echo "--- Hardware Information ---"
-tpu-info
+# tpu-info
 echo "--- Starting Tests ---"
 set +e
 overall_script_exit_code=0
@@ -134,7 +135,7 @@ run_and_track_test 1 "test_compilation.py" \
 run_and_track_test 2 "test_basic.py" \
     "python3 -m pytest -s -v /workspace/vllm/tests/v1/tpu/test_basic.py"
 run_and_track_test 3 "test_accuracy.py::test_lm_eval_accuracy_v1_engine" \
-    "python3 -m pytest -s -v /workspace/vllm/tests/entrypoints/llm/test_accuracy.py::test_lm_eval_accuracy_v1_engine"
+    "HF_HUB_DISABLE_XET=1 python3 -m pytest -s -v /workspace/vllm/tests/entrypoints/llm/test_accuracy.py::test_lm_eval_accuracy_v1_engine"
 run_and_track_test 4 "test_quantization_accuracy.py" \
     "python3 -m pytest -s -v /workspace/vllm/tests/tpu/test_quantization_accuracy.py"
 run_and_track_test 5 "examples/offline_inference/tpu.py" \
@@ -149,18 +150,6 @@ run_and_track_test 9 "test_multimodal.py" \
     "python3 -m pytest -s -v /workspace/vllm/tests/v1/tpu/test_multimodal.py"
 run_and_track_test 10 "test_pallas.py" \
     "python3 -m pytest -s -v /workspace/vllm/tests/v1/tpu/test_pallas.py"
-run_and_track_test 11 "test_struct_output_generate.py" \
-    "python3 -m pytest -s -v /workspace/vllm/tests/v1/entrypoints/llm/test_struct_output_generate.py -k \"not test_structured_output_with_reasoning_matrices\""
-run_and_track_test 12 "test_moe_pallas.py" \
-    "python3 -m pytest -s -v /workspace/vllm/tests/tpu/test_moe_pallas.py"
-run_and_track_test 13 "test_lora.py" \
-    "VLLM_XLA_CHECK_RECOMPILATION=0 python3 -m pytest -s -v /workspace/vllm/tests/tpu/lora/test_lora.py"
-run_and_track_test 14 "test_tpu_qkv_linear.py" \
-    "python3 -m pytest -s -v /workspace/vllm/tests/v1/tpu/test_tpu_qkv_linear.py"
-run_and_track_test 15 "test_spmd_model_weight_loading.py" \
-    "python3 -m pytest -s -v /workspace/vllm/tests/v1/tpu/test_spmd_model_weight_loading.py"
-run_and_track_test 16 "test_kv_cache_update_kernel.py" \
-    "python3 -m pytest -s -v /workspace/vllm/tests/v1/tpu/test_kv_cache_update_kernel.py"
 
 # After all tests have been attempted, exit with the overall status.
 if [ "$overall_script_exit_code" -ne 0 ]; then

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

@@ -11,8 +11,8 @@ container_name="xpu_${BUILDKITE_COMMIT}_$(tr -dc A-Za-z0-9 < /dev/urandom | head
 docker build -t ${image_name} -f docker/Dockerfile.xpu .
 
 # Setup cleanup
-remove_docker_container() { 
-  docker rm -f "${container_name}" || true; 
+remove_docker_container() {
+  docker rm -f "${container_name}" || true;
   docker image rm -f "${image_name}" || true;
   docker system prune -f || true;
 }
@@ -26,7 +26,18 @@ docker run \
     --name "${container_name}" \
     "${image_name}" \
     sh -c '
-    VLLM_USE_V1=0 python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m
-    VLLM_USE_V1=0 python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m -tp 2
     VLLM_USE_V1=1 python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m --block-size 64 --enforce-eager
+    VLLM_USE_V1=1 python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m --block-size 64 --enforce-eager -tp 2 --distributed-executor-backend ray
+    VLLM_USE_V1=1 python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m --block-size 64 --enforce-eager -tp 2 --distributed-executor-backend mp
+    cd tests
+    pytest -v -s v1/core
+    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/test_serial_utils.py
+    pytest -v -s v1/test_utils.py
+    pytest -v -s v1/test_metrics_reader.py
 '

.buildkite/scripts/run-benchmarks.sh

@@ -11,10 +11,10 @@ cd "$(dirname "${BASH_SOURCE[0]}")/../.."
 (which wget && which curl) || (apt-get update && apt-get install -y wget curl)
 
 # run python-based benchmarks and upload the result to buildkite
-python3 benchmarks/benchmark_latency.py --output-json latency_results.json 2>&1 | tee benchmark_latency.txt
+vllm bench latency --output-json latency_results.json 2>&1 | tee benchmark_latency.txt
 bench_latency_exit_code=$?
 
-python3 benchmarks/benchmark_throughput.py --input-len 256 --output-len 256 --output-json throughput_results.json 2>&1 | tee benchmark_throughput.txt
+vllm bench throughput --input-len 256 --output-len 256 --output-json throughput_results.json 2>&1 | tee benchmark_throughput.txt
 bench_throughput_exit_code=$?
 
 # run server-based benchmarks and upload the result to buildkite
@@ -24,7 +24,7 @@ wget https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/r
 
 # wait for server to start, timeout after 600 seconds
 timeout 600 bash -c 'until curl localhost:8000/v1/models; do sleep 1; done' || exit 1
-python3 benchmarks/benchmark_serving.py \
+vllm bench serve \
     --backend vllm \
     --dataset-name sharegpt \
     --dataset-path ./ShareGPT_V3_unfiltered_cleaned_split.json \

.buildkite/scripts/tpu/docker_run_bm.sh

@@ -22,16 +22,6 @@ trap remove_docker_container EXIT
 # Remove the container that might not be cleaned up in the previous run.
 remove_docker_container
 
-# Build docker image.
-# TODO: build the image outside the script and share the image with other
-# tpu test if building time is too long.
-DOCKER_BUILDKIT=1 docker build \
-  --build-arg max_jobs=16 \
-  --build-arg USE_SCCACHE=1 \
-  --build-arg GIT_REPO_CHECK=0 \
-  --tag vllm/vllm-tpu-bm \
-  --progress plain -f docker/Dockerfile.tpu .
-
 LOG_ROOT=$(mktemp -d)
 # If mktemp fails, set -e will cause the script to exit.
 echo "Results will be stored in: $LOG_ROOT"

.buildkite/scripts/tpu/quantized_v6e_1.env

@@ -0,0 +1,14 @@
+# Environment config
+TEST_NAME=llama8bw8a8
+CONTAINER_NAME=vllm-tpu
+
+# vllm config
+MODEL=RedHatAI/Meta-Llama-3.1-8B-Instruct-quantized.w8a8
+MAX_NUM_SEQS=128
+MAX_NUM_BATCHED_TOKENS=1024
+TENSOR_PARALLEL_SIZE=1
+MAX_MODEL_LEN=2048
+DOWNLOAD_DIR=/mnt/disks/persist
+EXPECTED_THROUGHPUT=10.0
+INPUT_LEN=1800
+OUTPUT_LEN=128

.buildkite/scripts/tpu/run_bm.sh

@@ -77,7 +77,7 @@ done
 echo "run benchmark test..."
 echo "logging to $BM_LOG"
 echo
-python benchmarks/benchmark_serving.py \
+vllm bench serve \
     --backend vllm \
     --model $MODEL  \
     --dataset-name sonnet \

.buildkite/test-pipeline.yaml

@@ -117,7 +117,7 @@ steps:
   commands:
   - pytest -v -s core
 
-- label: Entrypoints Test # 40min
+- label: Entrypoints Test (LLM) # 40min
   mirror_hardwares: [amdexperimental]
   working_dir: "/vllm-workspace/tests"
   fast_check: true
@@ -125,8 +125,6 @@ steps:
   source_file_dependencies:
   - vllm/
   - tests/entrypoints/llm
-  - tests/entrypoints/openai
-  - tests/entrypoints/test_chat_utils
   - tests/entrypoints/offline_mode
   commands:
   - export VLLM_WORKER_MULTIPROC_METHOD=spawn
@@ -135,9 +133,21 @@ steps:
   - pytest -v -s entrypoints/llm/test_generate.py # it needs a clean process
   - pytest -v -s entrypoints/llm/test_generate_multiple_loras.py # it needs a clean process
   - VLLM_USE_V1=0 pytest -v -s entrypoints/llm/test_guided_generate.py # it needs a clean process
+  - VLLM_USE_V1=0 pytest -v -s entrypoints/offline_mode # Needs to avoid interference with other tests
+
+- label: Entrypoints Test (API Server) # 40min
+  mirror_hardwares: [amdexperimental]
+  working_dir: "/vllm-workspace/tests"
+  fast_check: true
+  torch_nightly: true
+  source_file_dependencies:
+  - vllm/
+  - tests/entrypoints/openai
+  - tests/entrypoints/test_chat_utils
+  commands:
+  - export VLLM_WORKER_MULTIPROC_METHOD=spawn
   - pytest -v -s entrypoints/openai --ignore=entrypoints/openai/test_chat_with_tool_reasoning.py --ignore=entrypoints/openai/test_oot_registration.py --ignore=entrypoints/openai/test_tensorizer_entrypoint.py --ignore=entrypoints/openai/correctness/
   - pytest -v -s entrypoints/test_chat_utils.py
-  - VLLM_USE_V1=0 pytest -v -s entrypoints/offline_mode # Needs to avoid interference with other tests
 
 - label: Distributed Tests (4 GPUs) # 10min
   mirror_hardwares: [amdexperimental]
@@ -149,12 +159,14 @@ steps:
   - tests/distributed/test_utils
   - tests/distributed/test_pynccl
   - tests/distributed/test_events
-  - tests/spec_decode/e2e/test_integration_dist_tp4
   - tests/compile/test_basic_correctness
   - examples/offline_inference/rlhf.py
   - examples/offline_inference/rlhf_colocate.py
   - tests/examples/offline_inference/data_parallel.py
   - tests/v1/test_async_llm_dp.py
+  - tests/v1/test_external_lb_dp.py
+  - tests/v1/test_internal_lb_dp.py
+  - tests/v1/test_hybrid_lb_dp.py
   - tests/v1/engine/test_engine_core_client.py
   commands:
   # test with tp=2 and external_dp=2
@@ -163,14 +175,16 @@ steps:
   # test with tp=2 and pp=2
   - PP_SIZE=2 torchrun --nproc-per-node=4 distributed/test_torchrun_example.py
   # test with internal dp
-  - python3 ../examples/offline_inference/data_parallel.py
+  - python3 ../examples/offline_inference/data_parallel.py --enforce-eager
   - TP_SIZE=2 DP_SIZE=2 pytest -v -s v1/test_async_llm_dp.py
+  - TP_SIZE=2 DP_SIZE=2 pytest -v -s v1/test_external_lb_dp.py
+  - TP_SIZE=1 DP_SIZE=4 pytest -v -s v1/test_internal_lb_dp.py
+  - TP_SIZE=1 DP_SIZE=4 pytest -v -s v1/test_hybrid_lb_dp.py
   - pytest -v -s v1/engine/test_engine_core_client.py::test_kv_cache_events_dp
   - pytest -v -s distributed/test_utils.py
   - pytest -v -s compile/test_basic_correctness.py
   - pytest -v -s distributed/test_pynccl.py
   - pytest -v -s distributed/test_events.py
-  - pytest -v -s spec_decode/e2e/test_integration_dist_tp4.py
   # TODO: create a dedicated test section for multi-GPU example tests
   # when we have multiple distributed example tests
   - pushd ../examples/offline_inference
@@ -225,7 +239,7 @@ steps:
   working_dir: "/vllm-workspace/tests" # optional
 
 - label: Engine Test # 10min
-  mirror_hardwares: [amdexperimental, amdproduction]
+  mirror_hardwares: [amdexperimental]
   source_file_dependencies:
   - vllm/
   - tests/engine
@@ -254,6 +268,7 @@ steps:
     - pytest -v -s v1/structured_output
     - pytest -v -s v1/spec_decode
     - pytest -v -s v1/kv_connector/unit
+    - pytest -v -s v1/metrics
     - pytest -v -s v1/test_serial_utils.py
     - pytest -v -s v1/test_utils.py
     - pytest -v -s v1/test_oracle.py
@@ -262,11 +277,11 @@ steps:
     # 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-neuralmagic/lm-evaluation-harness.git@streaming-api
+    - 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
-  mirror_hardwares: [amdexperimental]
+  mirror_hardwares: [amdexperimental, amdproduction]
   working_dir: "/vllm-workspace/examples"
   source_file_dependencies:
   - vllm/entrypoints
@@ -280,7 +295,7 @@ steps:
     - python3 offline_inference/llm_engine_example.py
     - python3 offline_inference/audio_language.py --seed 0
     - python3 offline_inference/vision_language.py --seed 0
-    - python3 offline_inference/vision_language_embedding.py --seed 0
+    - python3 offline_inference/vision_language_pooling.py --seed 0
     - python3 offline_inference/vision_language_multi_image.py --seed 0
     - VLLM_USE_V1=0 python3 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
@@ -300,7 +315,7 @@ steps:
 
 
 - label: Platform Tests (CUDA)
-  mirror_hardwares: [amdexperimental]
+  mirror_hardwares: [amdexperimental, amdproduction]
   source_file_dependencies:
   - vllm/
   - tests/cuda
@@ -318,19 +333,8 @@ steps:
     - pytest -v -s samplers
     - VLLM_USE_FLASHINFER_SAMPLER=1 pytest -v -s samplers
 
-- label: Speculative decoding tests # 40min
-  mirror_hardwares: [amdexperimental]
-  source_file_dependencies:
-  - vllm/spec_decode
-  - tests/spec_decode
-  - vllm/model_executor/models/eagle.py
-  commands:
-    - pytest -v -s spec_decode/e2e/test_multistep_correctness.py
-    - VLLM_ATTENTION_BACKEND=FLASH_ATTN pytest -v -s spec_decode --ignore=spec_decode/e2e/test_multistep_correctness.py --ignore=spec_decode/e2e/test_mtp_correctness.py
-    - pytest -v -s spec_decode/e2e/test_eagle_correctness.py
-
 - label: LoRA Test %N # 15min each
-  mirror_hardwares: [amdexperimental, amdproduction]
+  mirror_hardwares: [amdexperimental]
   source_file_dependencies:
   - vllm/lora
   - tests/lora
@@ -338,7 +342,7 @@ steps:
   parallelism: 4
 
 - label: PyTorch Compilation Unit Tests
-  mirror_hardwares: [amdexperimental, amdproduction]
+  mirror_hardwares: [amdexperimental]
   torch_nightly: true
   source_file_dependencies:
     - vllm/
@@ -382,7 +386,7 @@ steps:
     - pytest -v -s kernels/core
 
 - label: Kernels Attention Test %N
-  mirror_hardwares: [amdexperimental, amdproduction]
+  mirror_hardwares: [amdexperimental]
   source_file_dependencies:
   - csrc/attention/
   - vllm/attention
@@ -393,7 +397,7 @@ steps:
   parallelism: 2
 
 - label: Kernels Quantization Test %N
-  mirror_hardwares: [amdexperimental, amdproduction]
+  mirror_hardwares: [amdexperimental]
   source_file_dependencies:
   - csrc/quantization/
   - vllm/model_executor/layers/quantization
@@ -412,7 +416,7 @@ steps:
     - pytest -v -s kernels/moe
 
 - label: Kernels Mamba Test
-  mirror_hardwares: [amdexperimental]
+  mirror_hardwares: [amdexperimental, amdproduction]
   source_file_dependencies:
   - csrc/mamba/
   - tests/kernels/mamba
@@ -434,7 +438,6 @@ steps:
 
 - label: Model Executor Test
   mirror_hardwares: [amdexperimental, amdproduction]
-  soft_fail: true
   source_file_dependencies:
   - vllm/model_executor
   - tests/model_executor
@@ -491,7 +494,7 @@ steps:
   - pytest -s entrypoints/openai/correctness/
 
 - label: Encoder Decoder tests # 5min
-  mirror_hardwares: [amdexperimental]
+  mirror_hardwares: [amdexperimental, amdproduction]
   source_file_dependencies:
   - vllm/
   - tests/encoder_decoder
@@ -499,7 +502,7 @@ steps:
     - pytest -v -s encoder_decoder
 
 - label: OpenAI-Compatible Tool Use # 20 min
-  mirror_hardwares: [amdexperimental]
+  mirror_hardwares: [amdexperimental, amdproduction]
   fast_check: false
   source_file_dependencies:
     - vllm/
@@ -512,7 +515,7 @@ steps:
 #####  models test  #####
 
 - label: Basic Models Test # 24min
-  mirror_hardwares: [amdexperimental, amdproduction]
+  mirror_hardwares: [amdexperimental]
   torch_nightly: true
   source_file_dependencies:
   - vllm/
@@ -536,6 +539,17 @@ steps:
     - pip freeze | grep -E 'torch'
     - pytest -v -s models/language -m core_model
 
+- label: Language Models Test (Hybrid) # 35 min
+  mirror_hardwares: [amdexperimental]
+  torch_nightly: true
+  source_file_dependencies:
+  - vllm/
+  - tests/models/language/generation
+  commands:
+    # Install causal-conv1d for plamo2 models here, as it is not compatible with pip-compile.
+    - pip install 'git+https://github.com/Dao-AILab/causal-conv1d@v1.5.0.post8'
+    - pytest -v -s models/language/generation -m hybrid_model
+
 - label: Language Models Test (Extended Generation) # 1hr20min
   mirror_hardwares: [amdexperimental]
   optional: true
@@ -545,7 +559,7 @@ steps:
   commands:
     # Install causal-conv1d for plamo2 models here, as it is not compatible with pip-compile.
     - pip install 'git+https://github.com/Dao-AILab/causal-conv1d@v1.5.0.post8'
-    - pytest -v -s models/language/generation -m 'not core_model'
+    - pytest -v -s models/language/generation -m '(not core_model) and (not hybrid_model)'
 
 - label: Language Models Test (Extended Pooling)  # 36min
   mirror_hardwares: [amdexperimental]
@@ -590,7 +604,7 @@ steps:
     - pytest -v -s models/multimodal/generation/test_common.py -m 'split(group=0) and not core_model'
 
 - label: Multi-Modal Models Test (Extended) 3
-  mirror_hardwares: [amdexperimental, amdproduction]
+  mirror_hardwares: [amdexperimental]
   optional: true
   source_file_dependencies:
   - vllm/
@@ -600,7 +614,7 @@ steps:
     - pytest -v -s models/multimodal/generation/test_common.py -m 'split(group=1) and not core_model'
 
 - label: Quantized Models Test
-  mirror_hardwares: [amdexperimental, amdproduction]
+  mirror_hardwares: [amdexperimental]
   source_file_dependencies:
   - vllm/model_executor/layers/quantization
   - tests/models/quantization
@@ -617,6 +631,18 @@ steps:
     # e.g. pytest -v -s models/encoder_decoder/vision_language/test_mllama.py
     # *To avoid merge conflicts, remember to REMOVE (not just comment out) them before merging the PR*
 
+- label: Transformers Nightly Models Test
+  working_dir: "/vllm-workspace/"
+  optional: true
+  commands:
+    - pip install --upgrade git+https://github.com/huggingface/transformers
+    - pytest -v -s tests/models/test_initialization.py
+    - pytest -v -s tests/models/multimodal/processing/
+    - pytest -v -s tests/models/multimodal/test_mapping.py
+    - python3 examples/offline_inference/basic/chat.py
+    - python3 examples/offline_inference/audio_language.py --model-type whisper
+    - python3 examples/offline_inference/vision_language.py --model-type qwen2_5_vl
+
 #####  1 GPU test  #####
 #####  multi gpus test  #####
 
@@ -671,10 +697,12 @@ steps:
   - vllm/worker/model_runner.py
   - entrypoints/llm/test_collective_rpc.py
   - tests/v1/test_async_llm_dp.py
+  - tests/v1/test_external_lb_dp.py
   - tests/v1/entrypoints/openai/test_multi_api_servers.py
   - vllm/v1/engine/
   commands:
   - TP_SIZE=1 DP_SIZE=2 pytest -v -s v1/test_async_llm_dp.py
+  - TP_SIZE=1 DP_SIZE=2 pytest -v -s v1/test_external_lb_dp.py
   - DP_SIZE=2 pytest -v -s v1/entrypoints/openai/test_multi_api_servers.py
   - pytest -v -s entrypoints/llm/test_collective_rpc.py
   - pytest -v -s ./compile/test_basic_correctness.py
@@ -689,10 +717,10 @@ steps:
   - pytest -v -s distributed/test_sequence_parallel.py
   # this test fails consistently.
   # TODO: investigate and fix
-  # - pytest -v -s spec_decode/e2e/test_integration_dist_tp2.py
   - VLLM_USE_V1=0 CUDA_VISIBLE_DEVICES=0,1 pytest -v -s test_sharded_state_loader.py
   - VLLM_USE_V1=0 CUDA_VISIBLE_DEVICES=0,1 pytest -v -s kv_transfer/test_disagg.py
   - CUDA_VISIBLE_DEVICES=0,1 pytest -v -s v1/shutdown
+  - pytest -v -s models/multimodal/generation/test_maverick.py
 
 - label: Plugin Tests (2 GPUs) # 40min
   mirror_hardwares: [amdexperimental]

.gemini/config.yaml

@@ -0,0 +1,6 @@
+# https://developers.google.com/gemini-code-assist/docs/customize-gemini-behavior-github
+have_fun: false  # Just review the code
+code_review:
+  comment_severity_threshold: HIGH  # Reduce quantity of comments
+  pull_request_opened:
+    summary: false  # Don't summarize the PR in a separate comment

.github/CODEOWNERS

@@ -16,7 +16,8 @@
 /vllm/lora @jeejeelee
 /vllm/reasoning @aarnphm
 /vllm/entrypoints @aarnphm
-CMakeLists.txt @tlrmchlsmth
+/vllm/compilation @zou3519 @youkaichao @ProExpertProg
+CMakeLists.txt @tlrmchlsmth @LucasWilkinson
 
 # Any change to the VllmConfig changes can have a large user-facing impact,
 # so spam a lot of people
@@ -42,7 +43,6 @@ CMakeLists.txt @tlrmchlsmth
 /tests/multimodal @DarkLight1337 @ywang96
 /tests/prefix_caching @comaniac @KuntaiDu
 /tests/quantization @mgoin @robertgshaw2-redhat
-/tests/spec_decode @njhill @LiuXiaoxuanPKU
 /tests/test_inputs.py @DarkLight1337 @ywang96
 /tests/v1/entrypoints/llm/test_struct_output_generate.py @mgoin @russellb @aarnphm
 /tests/v1/structured_output @mgoin @russellb @aarnphm
@@ -52,3 +52,15 @@ CMakeLists.txt @tlrmchlsmth
 # Docs
 /docs @hmellor
 mkdocs.yaml @hmellor
+
+# CPU
+/vllm/v1/worker/^cpu @bigPYJ1151
+/csrc/cpu @bigPYJ1151
+/vllm/platforms/cpu.py @bigPYJ1151
+/cmake/cpu_extension.cmake @bigPYJ1151
+/docker/Dockerfile.cpu @bigPYJ1151
+
+# Intel GPU
+/vllm/v1/worker/^xpu @jikunshang
+/vllm/platforms/xpu.py @jikunshang
+/docker/Dockerfile.xpu @jikunshang

.github/ISSUE_TEMPLATE/750-RFC.yml

@@ -46,7 +46,7 @@ body:
 - type: markdown
   attributes:
     value: >
-      Thanks for contributing 🎉!
+      Thanks for contributing 🎉! The vLLM core team hosts a biweekly RFC review session at 9:30AM Pacific Time, while most RFCs can be discussed online, you can optionally sign up for a slot to discuss your RFC online [here](https://docs.google.com/document/d/1CiLVBZeIVfR7_PNAKVSusxpceywkoOOB78qoWqHvSZc/edit).
 - type: checkboxes
   id: askllm
   attributes:

.github/mergify.yml

@@ -27,6 +27,22 @@ pull_request_rules:
       add:
         - ci/build
 
+- name: label-deepseek
+  description: Automatically apply deepseek label
+  conditions:
+    - or:
+      - files~=^examples/.*deepseek.*\.py
+      - files~=^tests/.*deepseek.*\.py
+      - files~=^vllm/entrypoints/openai/tool_parsers/.*deepseek.*\.py
+      - files~=^vllm/model_executor/models/.*deepseek.*\.py
+      - files~=^vllm/reasoning/.*deepseek.*\.py
+      - files~=^vllm/transformers_utils/.*deepseek.*\.py
+      - title~=(?i)DeepSeek
+  actions:
+    label:
+      add:
+        - deepseek
+
 - name: label-frontend
   description: Automatically apply frontend label
   conditions:
@@ -58,14 +74,23 @@ pull_request_rules:
       - files~=^vllm/multimodal/
       - files~=^tests/multimodal/
       - files~=^tests/models/multimodal/
-      - files~=^tests/models/*/audio_language/
-      - files~=^tests/models/*/vision_language/
       - files=tests/models/test_vision.py
   actions:
     label:
       add:
         - multi-modality
 
+- name: label-new-model
+  description: Automatically apply new-model label
+  conditions:
+    - and:
+      - files~=^vllm/model_executor/models/
+      - files=vllm/model_executor/models/registry.py
+  actions:
+    label:
+      add:
+        - new-model
+
 - name: label-performance
   description: Automatically apply performance label
   conditions:
@@ -139,9 +164,12 @@ pull_request_rules:
   description: Automatically apply speculative-decoding label
   conditions:
     - or:
-      - files~=^vllm/spec_decode/
-      - files=vllm/model_executor/layers/spec_decode_base_sampler.py
-      - files~=^tests/spec_decode/
+      - files~=^vllm/v1/spec_decode/
+      - files~=^tests/v1/spec_decode/
+      - files~=^examples/.*(spec_decode|mlpspeculator|eagle|speculation).*\.py
+      - files~=^vllm/model_executor/models/.*eagle.*\.py
+      - files=vllm/model_executor/models/mlp_speculator.py
+      - files~=^vllm/transformers_utils/configs/(eagle|medusa|mlp_speculator)\.py
   actions:
     label:
       add:

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

@@ -7,7 +7,7 @@ permissions:
 
 jobs:
   lint-and-deploy:
-    runs-on: ubuntu-latest
+    runs-on: ubuntu-24.04-arm
     steps:
       - name: Checkout
         uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
@@ -68,7 +68,7 @@ jobs:
           export AWS_ACCESS_KEY_ID=minioadmin
           export AWS_SECRET_ACCESS_KEY=minioadmin
           sleep 30 && kubectl -n ns-vllm logs -f "$(kubectl -n ns-vllm get pods | awk '/deployment/ {print $1;exit}')" &
-          helm install --wait --wait-for-jobs --timeout 5m0s --debug --create-namespace --namespace=ns-vllm test-vllm examples/online_serving/chart-helm -f examples/online_serving/chart-helm/values.yaml --set secrets.s3endpoint=http://minio:9000 --set secrets.s3bucketname=testbucket --set secrets.s3accesskeyid=$AWS_ACCESS_KEY_ID --set secrets.s3accesskey=$AWS_SECRET_ACCESS_KEY --set resources.requests.cpu=1 --set resources.requests.memory=4Gi --set resources.limits.cpu=2 --set resources.limits.memory=5Gi --set image.env[0].name=VLLM_CPU_KVCACHE_SPACE --set image.env[1].name=VLLM_LOGGING_LEVEL --set-string image.env[0].value="1" --set-string image.env[1].value="DEBUG" --set-string extraInit.s3modelpath="opt-125m/" --set-string 'resources.limits.nvidia\.com/gpu=0' --set-string 'resources.requests.nvidia\.com/gpu=0' --set-string image.repository="vllm-cpu-env"
+          helm install --wait --wait-for-jobs --timeout 5m0s --debug --create-namespace --namespace=ns-vllm test-vllm examples/online_serving/chart-helm -f examples/online_serving/chart-helm/values.yaml --set secrets.s3endpoint=http://minio:9000 --set secrets.s3bucketname=testbucket --set secrets.s3accesskeyid=$AWS_ACCESS_KEY_ID --set secrets.s3accesskey=$AWS_SECRET_ACCESS_KEY --set resources.requests.cpu=1 --set resources.requests.memory=4Gi --set resources.limits.cpu=2 --set resources.limits.memory=5Gi --set image.env[0].name=VLLM_CPU_KVCACHE_SPACE --set image.env[1].name=VLLM_LOGGING_LEVEL --set image.env[2].name=VLLM_CPU_CI_ENV --set-string image.env[0].value="1" --set-string image.env[1].value="DEBUG" --set-string image.env[2].value="1" --set-string extraInit.s3modelpath="opt-125m/" --set-string 'resources.limits.nvidia\.com/gpu=0' --set-string 'resources.requests.nvidia\.com/gpu=0' --set-string image.repository="vllm-cpu-env"
 
       - name: curl test
         run: |

.gitignore

@@ -146,6 +146,7 @@ venv.bak/
 
 # mkdocs documentation
 /site
+docs/argparse
 docs/examples
 
 # mypy

.pre-commit-config.yaml

@@ -21,7 +21,7 @@ repos:
   - id: ruff-format
     files: ^(.buildkite|benchmarks|examples)/.*
 - repo: https://github.com/crate-ci/typos
-  rev: v1.32.0
+  rev: v1.34.0
   hooks:
   - id: typos
 - repo: https://github.com/PyCQA/isort
@@ -160,10 +160,17 @@ repos:
     types: [python]
     pass_filenames: false
     additional_dependencies: [pathspec, regex]
+  - id: validate-config
+    name: Validate configuration has default values and that each field has a docstring
+    entry: python tools/validate_config.py
+    language: python
+    types: [python]
+    pass_filenames: true
+    files: vllm/config.py|tests/test_config.py|vllm/entrypoints/openai/cli_args.py
   # Keep `suggestion` last
   - id: suggestion
     name: Suggestion
-    entry: bash -c 'echo "To bypass pre-commit hooks, add --no-verify to git commit."'
+    entry: bash -c 'echo "To bypass all the pre-commit hooks, add --no-verify to git commit. To skip a specific hook, prefix the commit command with SKIP=<hook-id>."'
     language: system
     verbose: true
     pass_filenames: false

CMakeLists.txt

@@ -45,7 +45,7 @@ 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 "2.7.0")
+set(TORCH_SUPPORTED_VERSION_CUDA "2.7.1")
 set(TORCH_SUPPORTED_VERSION_ROCM "2.7.0")
 
 #
@@ -171,7 +171,6 @@ if(NVCC_THREADS AND VLLM_GPU_LANG STREQUAL "CUDA")
   list(APPEND VLLM_GPU_FLAGS "--threads=${NVCC_THREADS}")
 endif()
 
-
 #
 # Use FetchContent for C++ dependencies that are compiled as part of vLLM's build process.
 # setup.py will override FETCHCONTENT_BASE_DIR to play nicely with sccache.
@@ -232,7 +231,6 @@ endif()
 
 set(VLLM_EXT_SRC
   "csrc/mamba/mamba_ssm/selective_scan_fwd.cu"
-  "csrc/mamba/causal_conv1d/causal_conv1d.cu"
   "csrc/cache_kernels.cu"
   "csrc/attention/paged_attention_v1.cu"
   "csrc/attention/paged_attention_v2.cu"
@@ -259,7 +257,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
   SET(CUTLASS_ENABLE_HEADERS_ONLY ON CACHE BOOL "Enable only the header library")
 
   # Set CUTLASS_REVISION. Used for FetchContent. Also fixes some bogus messages when building.
-  set(CUTLASS_REVISION "v3.9.2" CACHE STRING "CUTLASS revision to use")
+  set(CUTLASS_REVISION "v4.0.0" CACHE STRING "CUTLASS revision to use")
 
   # Use the specified CUTLASS source directory for compilation if VLLM_CUTLASS_SRC_DIR is provided
   if (DEFINED ENV{VLLM_CUTLASS_SRC_DIR})
@@ -298,7 +296,8 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
     "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/attention/mla/cutlass_mla_entry.cu"
+    "csrc/quantization/fp8/per_token_group_quant.cu")
 
   set_gencode_flags_for_srcs(
     SRCS "${VLLM_EXT_SRC}"
@@ -393,7 +392,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
   # The cutlass_scaled_mm kernels for Hopper (c3x, i.e. CUTLASS 3.x) require
   # CUDA 12.0 or later
   cuda_archs_loose_intersection(SCALED_MM_ARCHS "9.0a;" "${CUDA_ARCHS}")
-  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.0 AND SCALED_MM_ARCHS)
+  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.0 AND SCALED_MM_ARCHS)
     set(SRCS
        "csrc/quantization/cutlass_w8a8/scaled_mm_c3x_sm90.cu"
        "csrc/quantization/cutlass_w8a8/c3x/scaled_mm_sm90_fp8.cu"
@@ -409,7 +408,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
     list(APPEND SCALED_MM_3X_ARCHS "${SCALED_MM_ARCHS}")
     message(STATUS "Building scaled_mm_c3x_sm90 for archs: ${SCALED_MM_ARCHS}")
   else()
-    if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.0 AND SCALED_MM_ARCHS)
+    if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.0 AND SCALED_MM_ARCHS)
       message(STATUS "Not building scaled_mm_c3x_sm90 as CUDA Compiler version is "
                      "not >= 12.0, we recommend upgrading to CUDA 12.0 or "
                      "later if you intend on running FP8 quantized models on "
@@ -420,10 +419,40 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
     endif()
   endif()
 
+
+  # The cutlass_scaled_mm kernels for Geforce Blackwell SM120 (c3x, i.e. CUTLASS 3.x) require
+  # CUDA 12.8 or later
+  cuda_archs_loose_intersection(SCALED_MM_ARCHS "12.0;12.0a" "${CUDA_ARCHS}")
+  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND SCALED_MM_ARCHS)
+    set(SRCS
+      "csrc/quantization/cutlass_w8a8/scaled_mm_c3x_sm120.cu"
+      "csrc/quantization/cutlass_w8a8/c3x/scaled_mm_sm120_fp8.cu"
+    )
+    set_gencode_flags_for_srcs(
+      SRCS "${SRCS}"
+      CUDA_ARCHS "${SCALED_MM_ARCHS}")
+    list(APPEND VLLM_EXT_SRC "${SRCS}")
+    list(APPEND VLLM_GPU_FLAGS "-DENABLE_SCALED_MM_SM120=1")
+    # Let scaled_mm_c2x know it doesn't need to build these arches
+    list(APPEND SCALED_MM_3X_ARCHS "${SCALED_MM_ARCHS}")
+    message(STATUS "Building scaled_mm_c3x_sm120 for archs: ${SCALED_MM_ARCHS}")
+  else()
+    if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND SCALED_MM_ARCHS)
+      message(STATUS "Not building scaled_mm_c3x_sm120 as CUDA Compiler version is "
+                     "not >= 12.8, we recommend upgrading to CUDA 12.8 or "
+                     "later if you intend on running FP8 quantized models on "
+                     "Blackwell.")
+    else()
+      message(STATUS "Not building scaled_mm_c3x_120 as no compatible archs found "
+                     "in CUDA target architectures")
+    endif()
+  endif()
+
+
   # The cutlass_scaled_mm kernels for Blackwell SM100 (c3x, i.e. CUTLASS 3.x)
   # require CUDA 12.8 or later
   cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0a;10.1a" "${CUDA_ARCHS}")
-  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.8 AND SCALED_MM_ARCHS)
+  if(${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"
@@ -438,7 +467,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
     list(APPEND SCALED_MM_3X_ARCHS "${SCALED_MM_ARCHS}")
     message(STATUS "Building scaled_mm_c3x_sm100 for archs: ${SCALED_MM_ARCHS}")
   else()
-    if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.8 AND SCALED_MM_ARCHS)
+    if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND SCALED_MM_ARCHS)
       message(STATUS "Not building scaled_mm_c3x_sm100 as CUDA Compiler version is "
                      "not >= 12.8, we recommend upgrading to CUDA 12.8 or "
                      "later if you intend on running FP8 quantized models on "
@@ -481,7 +510,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
   # The 2:4 sparse kernels cutlass_scaled_sparse_mm and cutlass_compressor
   # require CUDA 12.2 or later (and only work on Hopper).
   cuda_archs_loose_intersection(SCALED_MM_ARCHS "9.0a;" "${CUDA_ARCHS}")
-  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.2 AND SCALED_MM_ARCHS)
+  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.2 AND SCALED_MM_ARCHS)
     set(SRCS "csrc/sparse/cutlass/sparse_scaled_mm_c3x.cu")
     set_gencode_flags_for_srcs(
       SRCS "${SRCS}"
@@ -490,7 +519,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
     list(APPEND VLLM_GPU_FLAGS "-DENABLE_SPARSE_SCALED_MM_C3X=1")
     message(STATUS "Building sparse_scaled_mm_c3x for archs: ${SCALED_MM_ARCHS}")
   else()
-    if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.2 AND SCALED_MM_ARCHS)
+    if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.2 AND SCALED_MM_ARCHS)
       message(STATUS "Not building sparse_scaled_mm_c3x kernels as CUDA Compiler version is "
                      "not >= 12.2, we recommend upgrading to CUDA 12.2 or later "
                      "if you intend on running FP8 sparse quantized models on Hopper.")
@@ -502,7 +531,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
 
   # FP4 Archs and flags
   cuda_archs_loose_intersection(FP4_ARCHS "10.0a" "${CUDA_ARCHS}")
-  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.8 AND FP4_ARCHS)
+  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND FP4_ARCHS)
     set(SRCS
       "csrc/quantization/fp4/nvfp4_quant_kernels.cu"
       "csrc/quantization/fp4/nvfp4_experts_quant.cu"
@@ -523,9 +552,10 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
 
   # CUTLASS MLA Archs and flags
   cuda_archs_loose_intersection(MLA_ARCHS "10.0a" "${CUDA_ARCHS}")
-  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.8 AND MLA_ARCHS)
+  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND MLA_ARCHS)
     set(SRCS
-      "csrc/attention/mla/cutlass_mla_kernels.cu")
+      "csrc/attention/mla/cutlass_mla_kernels.cu"
+      "csrc/attention/mla/sm100_cutlass_mla_kernel.cu")
     set_gencode_flags_for_srcs(
       SRCS "${SRCS}"
       CUDA_ARCHS "${MLA_ARCHS}")
@@ -548,7 +578,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
   # if it's possible to compile MoE kernels that use its output.
   cuda_archs_loose_intersection(SCALED_MM_ARCHS "9.0a" "${CUDA_ARCHS}")
   if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.3 AND SCALED_MM_ARCHS)
-    set(SRCS "csrc/quantization/cutlass_w8a8/moe/grouped_mm_c3x.cu")
+    set(SRCS "csrc/quantization/cutlass_w8a8/moe/grouped_mm_c3x_sm90.cu")
     set_gencode_flags_for_srcs(
       SRCS "${SRCS}"
       CUDA_ARCHS "${SCALED_MM_ARCHS}")
@@ -562,7 +592,27 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
                      "if you intend on running FP8 quantized MoE models on Hopper.")
     else()
       message(STATUS "Not building grouped_mm_c3x as no compatible archs found "
-                     "in CUDA target architectures")
+                     "in CUDA target architectures.")
+    endif()
+  endif()
+
+  cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0a" "${CUDA_ARCHS}")
+  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND SCALED_MM_ARCHS)
+    set(SRCS "csrc/quantization/cutlass_w8a8/moe/grouped_mm_c3x_sm100.cu")
+    set_gencode_flags_for_srcs(
+      SRCS "${SRCS}"
+      CUDA_ARCHS "${SCALED_MM_ARCHS}")
+    list(APPEND VLLM_EXT_SRC "${SRCS}")
+    list(APPEND VLLM_GPU_FLAGS "-DENABLE_CUTLASS_MOE_SM100=1")
+    message(STATUS "Building grouped_mm_c3x for archs: ${SCALED_MM_ARCHS}")
+  else()
+    if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND SCALED_MM_ARCHS)
+      message(STATUS "Not building grouped_mm_c3x kernels as CUDA Compiler version is "
+                     "not >= 12.8, we recommend upgrading to CUDA 12.8 or later "
+                     "if you intend on running FP8 quantized MoE models on Blackwell.")
+    else()
+      message(STATUS "Not building grouped_mm_c3x as no compatible archs found "
+                     "in CUDA target architectures.")
     endif()
   endif()
 
@@ -574,7 +624,37 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
       SRCS "${SRCS}"
       CUDA_ARCHS "${CUTLASS_MOE_DATA_ARCHS}")
     list(APPEND VLLM_EXT_SRC "${SRCS}")
-  endif() 
+    message(STATUS "Building moe_data for archs: ${CUTLASS_MOE_DATA_ARCHS}")
+  else()
+    if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.3 AND CUTLASS_MOE_DATA_ARCHS)
+      message(STATUS "Not building moe_data as CUDA Compiler version is "
+                     "not >= 12.3, we recommend upgrading to CUDA 12.3 or later "
+                     "if you intend on running FP8 quantized MoE models on Hopper or Blackwell.")
+    else()
+      message(STATUS "Not building moe_data as no compatible archs found "
+                     "in CUDA target architectures.")
+    endif()
+  endif()
+
+  cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0a" "${CUDA_ARCHS}")
+  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND SCALED_MM_ARCHS)
+    set(SRCS "csrc/quantization/cutlass_w8a8/moe/blockwise_scaled_group_mm_sm100.cu")
+    set_gencode_flags_for_srcs(
+      SRCS "${SRCS}"
+      CUDA_ARCHS "${SCALED_MM_ARCHS}")
+    list(APPEND VLLM_EXT_SRC "${SRCS}")
+    list(APPEND VLLM_GPU_FLAGS "-DENABLE_CUTLASS_MOE_SM100=1")
+    message(STATUS "Building blockwise_scaled_group_mm_sm100 for archs: ${SCALED_MM_ARCHS}")
+  else()
+    if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND SCALED_MM_ARCHS)
+      message(STATUS "Not building blockwise_scaled_group_mm_sm100 kernels as CUDA Compiler version is "
+                     "not >= 12.8, we recommend upgrading to CUDA 12.8 or later "
+                     "if you intend on running FP8 quantized MoE models on Blackwell.")
+    else()
+      message(STATUS "Not building blockwise_scaled_group_mm_sm100 as no compatible archs found "
+                     "in CUDA target architectures")
+    endif()
+  endif()
 
   #
   # Machete kernels
@@ -582,7 +662,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
   # The machete kernels only work on hopper and require CUDA 12.0 or later.
   # Only build Machete kernels if we are building for something compatible with sm90a
   cuda_archs_loose_intersection(MACHETE_ARCHS "9.0a" "${CUDA_ARCHS}")
-  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.0 AND MACHETE_ARCHS)
+  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.0 AND MACHETE_ARCHS)
     #
     # For the Machete kernels we automatically generate sources for various
     # preselected input type pairs and schedules.
@@ -634,7 +714,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
 
     message(STATUS "Building Machete kernels for archs: ${MACHETE_ARCHS}")
   else()
-    if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.0
+    if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.0
         AND MACHETE_ARCHS)
       message(STATUS "Not building Machete kernels as CUDA Compiler version is "
                      "not >= 12.0, we recommend upgrading to CUDA 12.0 or "
@@ -688,6 +768,14 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
   list(APPEND VLLM_MOE_EXT_SRC "csrc/moe/moe_wna16.cu")
 endif()
 
+if(VLLM_GPU_LANG STREQUAL "CUDA")
+  set(MOE_PERMUTE_SRC
+      "csrc/moe/permute_unpermute_kernels/moe_permute_unpermute_kernel.cu"
+      "csrc/moe/moe_permute_unpermute_op.cu")
+
+  list(APPEND VLLM_MOE_EXT_SRC "${MOE_PERMUTE_SRC}")
+endif()
+
 set_gencode_flags_for_srcs(
   SRCS "${VLLM_MOE_EXT_SRC}"
   CUDA_ARCHS "${CUDA_ARCHS}")
@@ -756,17 +844,6 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
   endif()
 endif()
 
-if(VLLM_GPU_LANG STREQUAL "CUDA")
-  set(MOE_PERMUTE_SRC
-      "csrc/moe/permute_unpermute_kernels/moe_permute_unpermute_kernel.cu"
-      "csrc/moe/moe_permute_unpermute_op.cu")
-
-  set_gencode_flags_for_srcs(
-    SRCS "${MARLIN_PERMUTE_SRC}"
-    CUDA_ARCHS "${MOE_PERMUTE_ARCHS}")
-
-  list(APPEND VLLM_MOE_EXT_SRC "${MOE_PERMUTE_SRC}")
-endif()
 message(STATUS "Enabling moe extension.")
 define_gpu_extension_target(
   _moe_C

README.md

@@ -63,13 +63,11 @@ vLLM is fast with:
 - Speculative decoding
 - Chunked prefill
 
-**Performance benchmark**: We include a performance benchmark at the end of [our blog post](https://blog.vllm.ai/2024/09/05/perf-update.html). It compares the performance of vLLM against other LLM serving engines ([TensorRT-LLM](https://github.com/NVIDIA/TensorRT-LLM), [SGLang](https://github.com/sgl-project/sglang) and [LMDeploy](https://github.com/InternLM/lmdeploy)). The implementation is under [nightly-benchmarks folder](.buildkite/nightly-benchmarks/) and you can [reproduce](https://github.com/vllm-project/vllm/issues/8176) this benchmark using our one-click runnable script.
-
 vLLM is flexible and easy to use with:
 
 - Seamless integration with popular Hugging Face models
 - High-throughput serving with various decoding algorithms, including *parallel sampling*, *beam search*, and more
-- Tensor parallelism and pipeline parallelism support for distributed inference
+- Tensor, pipeline, data and expert parallelism support for distributed inference
 - Streaming outputs
 - OpenAI-compatible API server
 - Support NVIDIA GPUs, AMD CPUs and GPUs, Intel CPUs and GPUs, PowerPC CPUs, TPU, and AWS Neuron

RELEASE.md

@@ -52,3 +52,36 @@ After branch cut, we approach finalizing the release branch with clear criteria
 * Release branch specific changes (e.g. change version identifiers or CI fixes)
 
 Please note: **No feature work allowed for cherry picks**. All PRs that are considered for cherry-picks need to be merged on trunk, the only exception are Release branch specific changes.
+
+## Manual validations
+
+### E2E Performance Validation
+
+Before each release, we perform end-to-end performance validation to ensure no regressions are introduced. This validation uses the [vllm-benchmark workflow](https://github.com/pytorch/pytorch-integration-testing/actions/workflows/vllm-benchmark.yml) on PyTorch CI.
+
+**Current Coverage:**
+* Models: Llama3, Llama4, and Mixtral
+* Hardware: NVIDIA H100 and AMD MI300x
+* *Note: Coverage may change based on new model releases and hardware availability*
+
+**Performance Validation Process:**
+
+**Step 1: Get Access**
+Request write access to the [pytorch/pytorch-integration-testing](https://github.com/pytorch/pytorch-integration-testing) repository to run the benchmark workflow.
+
+**Step 2: Review Benchmark Setup**
+Familiarize yourself with the benchmark configurations:
+* [CUDA setup](https://github.com/pytorch/pytorch-integration-testing/tree/main/vllm-benchmarks/benchmarks/cuda)
+* [ROCm setup](https://github.com/pytorch/pytorch-integration-testing/tree/main/vllm-benchmarks/benchmarks/rocm)
+
+**Step 3: Run the Benchmark**
+Navigate to the [vllm-benchmark workflow](https://github.com/pytorch/pytorch-integration-testing/actions/workflows/vllm-benchmark.yml) and configure:
+* **vLLM branch**: Set to the release branch (e.g., `releases/v0.9.2`)
+* **vLLM commit**: Set to the RC commit hash
+
+**Step 4: Review Results**
+Once the workflow completes, benchmark results will be available on the [vLLM benchmark dashboard](https://hud.pytorch.org/benchmark/llms?repoName=vllm-project%2Fvllm) under the corresponding branch and commit.
+
+**Step 5: Performance Comparison**
+Compare the current results against the previous release to verify no performance regressions have occurred. Here is an
+example of [v0.9.1 vs v0.9.2](https://hud.pytorch.org/benchmark/llms?startTime=Thu%2C%2017%20Apr%202025%2021%3A43%3A50%20GMT&stopTime=Wed%2C%2016%20Jul%202025%2021%3A43%3A50%20GMT&granularity=week&lBranch=releases/v0.9.1&lCommit=b6553be1bc75f046b00046a4ad7576364d03c835&rBranch=releases/v0.9.2&rCommit=a5dd03c1ebc5e4f56f3c9d3dc0436e9c582c978f&repoName=vllm-project%2Fvllm&benchmarkName=&modelName=All%20Models&backendName=All%20Backends&modeName=All%20Modes&dtypeName=All%20DType&deviceName=All%20Devices&archName=All%20Platforms).

benchmarks/README.md

@@ -98,7 +98,7 @@ Then run the benchmarking script
 ```bash
 # download dataset
 # wget https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/resolve/main/ShareGPT_V3_unfiltered_cleaned_split.json
-python3 vllm/benchmarks/benchmark_serving.py \
+vllm bench serve \
   --backend vllm \
   --model NousResearch/Hermes-3-Llama-3.1-8B \
   --endpoint /v1/completions \
@@ -111,25 +111,25 @@ If successful, you will see the following output
 
 ```
 ============ Serving Benchmark Result ============
-Successful requests:                     10        
-Benchmark duration (s):                  5.78      
-Total input tokens:                      1369      
-Total generated tokens:                  2212      
-Request throughput (req/s):              1.73      
-Output token throughput (tok/s):         382.89    
-Total Token throughput (tok/s):          619.85    
+Successful requests:                     10
+Benchmark duration (s):                  5.78
+Total input tokens:                      1369
+Total generated tokens:                  2212
+Request throughput (req/s):              1.73
+Output token throughput (tok/s):         382.89
+Total Token throughput (tok/s):          619.85
 ---------------Time to First Token----------------
-Mean TTFT (ms):                          71.54     
-Median TTFT (ms):                        73.88     
-P99 TTFT (ms):                           79.49     
+Mean TTFT (ms):                          71.54
+Median TTFT (ms):                        73.88
+P99 TTFT (ms):                           79.49
 -----Time per Output Token (excl. 1st token)------
-Mean TPOT (ms):                          7.91      
-Median TPOT (ms):                        7.96      
-P99 TPOT (ms):                           8.03      
+Mean TPOT (ms):                          7.91
+Median TPOT (ms):                        7.96
+P99 TPOT (ms):                           8.03
 ---------------Inter-token Latency----------------
-Mean ITL (ms):                           7.74      
-Median ITL (ms):                         7.70      
-P99 ITL (ms):                            8.39      
+Mean ITL (ms):                           7.74
+Median ITL (ms):                         7.70
+P99 ITL (ms):                            8.39
 ==================================================
 ```
 
@@ -141,7 +141,7 @@ If the dataset you want to benchmark is not supported yet in vLLM, even then you
 {"prompt": "What is the capital of India?"}
 {"prompt": "What is the capital of Iran?"}
 {"prompt": "What is the capital of China?"}
-``` 
+```
 
 ```bash
 # start server
@@ -150,7 +150,7 @@ VLLM_USE_V1=1 vllm serve meta-llama/Llama-3.1-8B-Instruct --disable-log-requests
 
 ```bash
 # run benchmarking script
-python3 benchmarks/benchmark_serving.py --port 9001 --save-result --save-detailed \
+vllm bench serve --port 9001 --save-result --save-detailed \
   --backend vllm \
   --model meta-llama/Llama-3.1-8B-Instruct \
   --endpoint /v1/completions \
@@ -174,7 +174,7 @@ vllm serve Qwen/Qwen2-VL-7B-Instruct --disable-log-requests
 ```
 
 ```bash
-python3 vllm/benchmarks/benchmark_serving.py \
+vllm bench serve \
   --backend openai-chat \
   --model Qwen/Qwen2-VL-7B-Instruct \
   --endpoint /v1/chat/completions \
@@ -194,7 +194,7 @@ VLLM_USE_V1=1 vllm serve meta-llama/Meta-Llama-3-8B-Instruct \
 ```
 
 ``` bash
-python3 benchmarks/benchmark_serving.py \
+vllm bench serve \
     --model meta-llama/Meta-Llama-3-8B-Instruct \
     --dataset-name hf \
     --dataset-path likaixin/InstructCoder \
@@ -210,7 +210,7 @@ vllm serve Qwen/Qwen2-VL-7B-Instruct --disable-log-requests
 **`lmms-lab/LLaVA-OneVision-Data`**
 
 ```bash
-python3 vllm/benchmarks/benchmark_serving.py \
+vllm bench serve \
   --backend openai-chat \
   --model Qwen/Qwen2-VL-7B-Instruct \
   --endpoint /v1/chat/completions \
@@ -224,7 +224,7 @@ python3 vllm/benchmarks/benchmark_serving.py \
 **`Aeala/ShareGPT_Vicuna_unfiltered`**
 
 ```bash
-python3 vllm/benchmarks/benchmark_serving.py \
+vllm bench serve \
   --backend openai-chat \
   --model Qwen/Qwen2-VL-7B-Instruct \
   --endpoint /v1/chat/completions \
@@ -237,7 +237,7 @@ python3 vllm/benchmarks/benchmark_serving.py \
 **`AI-MO/aimo-validation-aime`**
 
 ``` bash
-python3 vllm/benchmarks/benchmark_serving.py \
+vllm bench serve \
     --model Qwen/QwQ-32B \
     --dataset-name hf \
     --dataset-path AI-MO/aimo-validation-aime \
@@ -248,7 +248,7 @@ python3 vllm/benchmarks/benchmark_serving.py \
 **`philschmid/mt-bench`**
 
 ``` bash
-python3 vllm/benchmarks/benchmark_serving.py \
+vllm bench serve \
     --model Qwen/QwQ-32B \
     --dataset-name hf \
     --dataset-path philschmid/mt-bench \
@@ -261,7 +261,7 @@ When using OpenAI-compatible backends such as `vllm`, optional sampling
 parameters can be specified. Example client command:
 
 ```bash
-python3 vllm/benchmarks/benchmark_serving.py \
+vllm bench serve \
   --backend vllm \
   --model NousResearch/Hermes-3-Llama-3.1-8B \
   --endpoint /v1/completions \
@@ -296,7 +296,7 @@ The following arguments can be used to control the ramp-up:
 <br/>
 
 ```bash
-python3 vllm/benchmarks/benchmark_throughput.py \
+vllm bench throughput \
   --model NousResearch/Hermes-3-Llama-3.1-8B \
   --dataset-name sonnet \
   --dataset-path vllm/benchmarks/sonnet.txt \
@@ -314,7 +314,7 @@ Total num output tokens:  1500
 **VisionArena Benchmark for Vision Language Models**
 
 ``` bash
-python3 vllm/benchmarks/benchmark_throughput.py \
+vllm bench throughput \
   --model Qwen/Qwen2-VL-7B-Instruct \
   --backend vllm-chat \
   --dataset-name hf \
@@ -336,7 +336,7 @@ Total num output tokens:  1280
 ``` bash
 VLLM_WORKER_MULTIPROC_METHOD=spawn \
 VLLM_USE_V1=1 \
-python3 vllm/benchmarks/benchmark_throughput.py \
+vllm bench throughput \
     --dataset-name=hf \
     --dataset-path=likaixin/InstructCoder \
     --model=meta-llama/Meta-Llama-3-8B-Instruct \
@@ -360,7 +360,7 @@ Total num output tokens:  204800
 **`lmms-lab/LLaVA-OneVision-Data`**
 
 ```bash
-python3 vllm/benchmarks/benchmark_throughput.py \
+vllm bench throughput \
   --model Qwen/Qwen2-VL-7B-Instruct \
   --backend vllm-chat \
   --dataset-name hf \
@@ -373,7 +373,7 @@ python3 vllm/benchmarks/benchmark_throughput.py \
 **`Aeala/ShareGPT_Vicuna_unfiltered`**
 
 ```bash
-python3 vllm/benchmarks/benchmark_throughput.py \
+vllm bench throughput \
   --model Qwen/Qwen2-VL-7B-Instruct \
   --backend vllm-chat \
   --dataset-name hf \
@@ -385,7 +385,7 @@ python3 vllm/benchmarks/benchmark_throughput.py \
 **`AI-MO/aimo-validation-aime`**
 
 ```bash
-python3 benchmarks/benchmark_throughput.py \
+vllm bench throughput \
   --model Qwen/QwQ-32B \
   --backend vllm \
   --dataset-name hf \
@@ -399,7 +399,7 @@ python3 benchmarks/benchmark_throughput.py \
 ``` bash
 # download dataset
 # wget https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/resolve/main/ShareGPT_V3_unfiltered_cleaned_split.json
-python3 vllm/benchmarks/benchmark_throughput.py \
+vllm bench throughput \
   --model meta-llama/Llama-2-7b-hf \
   --backend vllm \
   --dataset_path <your data path>/ShareGPT_V3_unfiltered_cleaned_split.json \

benchmarks/auto_tune/README.md

@@ -0,0 +1,141 @@
+# Automated vLLM Server Parameter Tuning
+
+This script automates the process of finding the optimal server parameter combination (`max-num-seqs` and `max-num-batched-tokens`) to maximize throughput for a vLLM server. It also supports additional constraints such as E2E latency and prefix cache hit rate.
+
+## Table of Contents
+- [Prerequisites](#prerequisites)
+- [Configuration](#configuration)
+- [How to Run](#how-to-run)
+- [Example Use Cases](#example-use-cases)
+- [Output](#output)
+- [How It Works](#how-it-works)
+
+## Prerequisites
+
+Before running the script, please ensure the following steps are completed:
+
+1. **Clone vLLM & Set Up Branch**: Clone the vLLM repository and check out to your desired branch.
+
+```bash
+git clone https://github.com/vllm-project/vllm.git
+cd vllm
+# git checkout <your-branch>
+```
+
+1. **Install Environment**: Install or update the correct running environment. For TPU usage, activate your `conda` environment and install the corresponding `torch` and `torch_xla` versions.
+
+2. **Model Configuration**: If you are using a customized model, ensure its configuration files are correctly placed and accessible.
+
+## Configuration
+
+You must set the following variables at the top of the script before execution.
+
+| Variable | Description | Example Value |
+| --- | --- | --- |
+| `BASE` | **Required.** The absolute path to the parent directory of your vLLM repository directory. | `"$HOME"` |
+| `MODEL` | **Required.** The Hugging Face model identifier to be served by vllm. | `"meta-llama/Llama-3.1-8B-Instruct"` |
+| `SYSTEM`| **Required.** The hardware you are running on. Choices: `TPU` or `GPU`. (For other systems, it might not support saving profiles) | `"TPU"` |
+| `TP` | **Required.** The tensor-parallelism size. | `1` |
+| `DOWNLOAD_DIR` | **Required.** Directory to download and load model weights from. | `""` (default download path) |
+| `INPUT_LEN` | **Required.** Request input length. | `4000` |
+| `OUTPUT_LEN` | **Required.** Request output length. | `16` |
+| `MAX_MODEL_LEN` | **Required.** Max model length. | `4096` |
+| `MIN_CACHE_HIT_PCT` | Prefix cache hit rate in percentage (0-100). Set to `0` to disable. | `60` |
+| `MAX_LATENCY_ALLOWED_MS` | The maximum allowed P99 end-to-end latency in milliseconds. Set to a very large number (e.g., `100000000000`) to effectively ignore the latency constraint. | `500` |
+| `NUM_SEQS_LIST` | A space-separated string of `max-num-seqs` values to test. | `"128 256"` |
+| `NUM_BATCHED_TOKENS_LIST` | A space-separated string of `max-num-batched-tokens` values to test. | `"1024 2048 4096"` |
+
+**Note**: The default `NUM_SEQS_LIST` and `NUM_BATCHED_TOKENS_LIST` are set for medium-sized inputs/outputs. For very short contexts (e.g., 20 input, 20 output tokens), you may need to test larger values for `max-num-seqs`.
+
+## How to Run
+
+1. **Configure**: Edit the script and set the variables in the [Configuration](#configuration) section.
+2. **Execute**: Run the script. Since the process can take a long time, it is highly recommended to use a terminal multiplexer like `tmux` or `screen` to prevent the script from stopping if your connection is lost.
+
+```
+cd <FOLDER_OF_THIS_SCRIPT>
+bash auto_tune.sh
+```
+
+    Please note that the `bash auto_tune.sh` command cannot contain full or partial path with keyword `vllm`, otherwise `pkill -f vllm` command will also kill this script itself.
+
+## Example Use Cases
+
+Here are a few examples of how to configure the script for different goals:
+
+### 1. Maximize Throughput (No Latency Constraint)
+- **Goal**: Find the best `max-num-seqs` and `max-num-batched-tokens` to get the highest possible throughput for 1800 input tokens and 20 output tokens.
+- **Configuration**:
+
+```bash
+INPUT_LEN=1800
+OUTPUT_LEN=20
+MAX_MODEL_LEN=2048
+MIN_CACHE_HIT_PCT=0
+MAX_LATENCY_ALLOWED_MS=100000000000 # A very large number
+```
+
+#### 2. Maximize Throughput with a Latency Requirement
+- **Goal**: Find the best server parameters when P99 end-to-end latency must be below 500ms.
+- **Configuration**:
+
+```bash
+INPUT_LEN=1800
+OUTPUT_LEN=20
+MAX_MODEL_LEN=2048
+MIN_CACHE_HIT_PCT=0
+MAX_LATENCY_ALLOWED_MS=500
+```
+
+#### 3. Maximize Throughput with Prefix Caching and Latency Requirements
+- **Goal**: Find the best server parameters assuming a 60% prefix cache hit rate and a latency requirement of 500ms.
+- **Configuration**:
+
+```bash
+INPUT_LEN=1800
+OUTPUT_LEN=20
+MAX_MODEL_LEN=2048
+MIN_CACHE_HIT_PCT=60
+MAX_LATENCY_ALLOWED_MS=500
+```
+
+## Output
+
+After the script finishes, you will find the results in a new, timestamped directory created inside `$BASE/auto-benchmark/`.
+
+- **Log Files**: The directory (`$BASE/auto-benchmark/YYYY_MM_DD_HH_MM/`) contains detailed logs for each run:
+    - `vllm_log_...txt`: The log output from the vLLM server for each parameter combination.
+    - `bm_log_...txt`: The log output from the `vllm bench serve` command for each benchmark run.
+
+- **Final Result Summary**: A file named `result.txt` is created in the log directory. It contains a summary of each tested combination and concludes with the overall best parameters found.
+
+```
+# Example result.txt content
+hash:a1b2c3d4...
+max_num_seqs: 128, max_num_batched_tokens: 2048, request_rate: 10.0, e2el: 450.5, throughput: 9.8, goodput: 9.8
+max_num_seqs: 128, max_num_batched_tokens: 4096 does not meet latency requirement 500
+...
+best_max_num_seqs: 256, best_num_batched_tokens: 2048, best_throughput: 12.5, profile saved in: /home/user/vllm/auto-benchmark/2024_08_01_10_30/profile
+```
+
+  If it cannot find the best parameters, the final row will be `best_max_num_seqs: 0, best_num_batched_tokens: 0, best_throughput: 0`. This can be due to either the server not starting properly, or the latency requirement being too strict.
+
+- **Profiler Trace**: A directory named `profile` is created inside the log directory. It contains the profiler trace file (e.g., `.xplane.pb` for TPU or a `.json` trace for GPU) from the single best-performing run.
+
+## How It Works
+
+The script follows a systematic process to find the optimal parameters:
+
+1. **Find Max GPU Memory Utilization**: The script first determines the highest safe `gpu-memory-utilization` (starting from 0.98 and decreasing) that does not cause an Out-Of-Memory (OOM) error when launching the server. This ensures the benchmark runs use the maximum available memory without crashing.
+
+2. **Iterate and Benchmark**: It then enters a nested loop, iterating through every combination of `max-num-seqs` and `max-num-batched-tokens` provided in the configuration lists.
+
+3. **Latency-Aware Throughput Search**: For each parameter combination:
+    - The vLLM server is started.
+    - A benchmark is first run with an infinite request rate (`--request-rate inf`).
+    - If the resulting P99 E2E latency is within the `MAX_LATENCY_ALLOWED_MS` limit, this throughput is considered the maximum for this configuration.
+    - If the latency is too high, the script performs a search by iteratively decreasing the request rate until the latency constraint is met. This finds the highest sustainable throughput for the given parameters and latency requirement.
+
+4. **Track Best Result**: Throughout the process, the script tracks the parameter combination that has yielded the highest valid throughput so far.
+
+5. **Profile Collection**: For the best-performing run, the script saves the vLLM profiler output, which can be used for deep-dive performance analysis with tools like TensorBoard.

benchmarks/auto_tune/auto_tune.sh

@@ -1,45 +1,18 @@
 #!/bin/bash
 
-# This script aims to tune the best server parameter combinations to maximize throughput for given requirement. 
-# The current server parameter combination is  max_num_seqs and max_num_batched_tokens
-# It also supports additional requirement: e2e latency and prefix cache. 
-
-# Pre-requisite:
-# 1. Checkout to your branch, install/ update the correct running env. For TPU, activate conda env and install the corresponding torch, xla version. 
-# 2. If the model is customized, replace the MODEL's config with the customized config.
-# 3. Set variables (ALL REQUIRED)
-#   BASE: your directory for vllm repo
-#   MODEL: the model served by vllm
-#   SYSTEM: the hardware, choice TPU or GPU, for other systems, "get best profile" might not support.
-#   TP: ways of tensor parallelism
-#   DOWNLOAD_DIR: directory to download and load model weights.
-#   INPUT_LEN: request input len
-#   OUTPUT_LEN: request output len
-#   MIN_CACHE_HIT_PCT: prefix cache rate
-#   MAX_LATENCY_ALLOWED_MS: (e2e) latency requirement. If there's no latency requirement, set it to a large number like 1000000000
-#   NUM_SEQS_LIST: a list of `max-num-seqs` you want to loop with.
-#   NUM_BATCHED_TOKENS_LIST: a list of `max-num-batched-tokens` you want to loop with.
-#   Note that the default NUM_SEQS_LIST and NUM_BATCHED_TOKENS_LIST are set for medium size input/output len, for extra short context (such as 20:20), you might need to include larger numbers in NUM_SEQS_LIST.
-# 4. Run the script, it might take a long time, you can use tmux to avoid the script stop if disconnection happens.
-# 5. The final result will be saved in RESULT file. 
-
-
-# Example use cases 
-# 1. Given input_len=1800, output_len=20, what's the best max_num_seqs and max_num_batched_tokens to get highest throughput?
-# Use INPUT_LEN=1800,  OUTPUT_LEN=20, MIN_CACHE_HIT_PCT=0, MAX_LATENCY_ALLOWED_MS=100000000000
-# 2. If we have latency requirement to be lower than 500ms, what's the best server parameter?
-# Use INPUT_LEN=1800,  OUTPUT_LEN=20, MIN_CACHE_HIT_PCT=0, MAX_LATENCY_ALLOWED_MS=500
-# 3. If we want to reach 60% prefix cache, what's the best server parameter? 
-# Use INPUT_LEN=1800,  OUTPUT_LEN=20, MIN_CACHE_HIT_PCT=60, MAX_LATENCY_ALLOWED_MS=500
+# This script aims to tune the best server parameter combinations to maximize throughput for given requirement.
+# See details in README (benchmarks/auto_tune/README.md).
 
 TAG=$(date +"%Y_%m_%d_%H_%M")
-BASE=""
+SCRIPT_DIR=$( cd -- "$( dirname -- "${BASH_SOURCE[0]}" )" &> /dev/null && pwd )
+BASE="$SCRIPT_DIR/../../.."
 MODEL="meta-llama/Llama-3.1-8B-Instruct"
 SYSTEM="TPU"
 TP=1
 DOWNLOAD_DIR=""
 INPUT_LEN=4000
 OUTPUT_LEN=16
+MAX_MODEL_LEN=4096
 MIN_CACHE_HIT_PCT=0
 MAX_LATENCY_ALLOWED_MS=100000000000
 NUM_SEQS_LIST="128 256"
@@ -65,6 +38,13 @@ current_hash=$(git rev-parse HEAD)
 echo "hash:$current_hash" >> "$RESULT"
 echo "current_hash: $current_hash"
 
+TOTAL_LEN=$((INPUT_LEN + OUTPUT_LEN))
+RED='\033[0;31m'
+if (( TOTAL_LEN > MAX_MODEL_LEN )); then
+    echo -e "${RED}FAILED: INPUT_LEN($INPUT_LEN) + OUTPUT_LEN($OUTPUT_LEN) = $TOTAL_LEN, which is > MAX_MODEL_LEN = $MAX_MODEL_LEN.\033[0m" >&2
+    exit 1
+fi
+
 best_throughput=0
 best_max_num_seqs=0
 best_num_batched_tokens=0
@@ -76,7 +56,7 @@ start_server() {
     local max_num_batched_tokens=$3
     local vllm_log=$4
     local profile_dir=$5
-    
+
     pkill -f vllm
 
     VLLM_USE_V1=1 VLLM_SERVER_DEV_MODE=1 VLLM_TORCH_PROFILER_DIR=$profile_dir vllm serve $MODEL \
@@ -89,13 +69,13 @@ start_server() {
         --enable-prefix-caching \
         --load-format dummy \
         --download-dir "$DOWNLOAD_DIR" \
-        --max-model-len $(( INPUT_LEN+OUTPUT_LEN )) > "$vllm_log" 2>&1 &
+        --max-model-len $MAX_MODEL_LEN > "$vllm_log" 2>&1 &
 
     # wait for 10 minutes...
     server_started=0
-    for i in {1..60}; do  
+    for i in {1..60}; do
         RESPONSE=$(curl -s -X GET "http://0.0.0.0:8004/health" -w "%{http_code}" -o /dev/stdout)
-        STATUS_CODE=$(echo "$RESPONSE" | tail -n 1) 
+        STATUS_CODE=$(echo "$RESPONSE" | tail -n 1)
         if [[ "$STATUS_CODE" -eq 200 ]]; then
             server_started=1
             break
@@ -118,10 +98,10 @@ update_best_profile() {
     selected_profile_file=
     if [[ "$SYSTEM" == "TPU" ]]; then
         selected_profile_file="${sorted_paths[$profile_index]}/*.xplane.pb"
-    fi 
+    fi
     if [[ "$SYSTEM" == "GPU" ]]; then
         selected_profile_file="${sorted_paths[$profile_index]}"
-    fi 
+    fi
     rm -f $PROFILE_PATH/*
     cp $selected_profile_file $PROFILE_PATH
 }
@@ -149,17 +129,18 @@ run_benchmark() {
         echo "server started."
     fi
     echo
-    
+
     echo "run benchmark test..."
     meet_latency_requirement=0
     # get a basic qps by using request-rate inf
     bm_log="$LOG_FOLDER/bm_log_${max_num_seqs}_${max_num_batched_tokens}_requestrate_inf.txt"
     prefix_len=$(( INPUT_LEN * MIN_CACHE_HIT_PCT / 100 ))
-    python benchmarks/benchmark_serving.py \
+adjusted_input_len=$(( INPUT_LEN - prefix_len ))
+    vllm bench serve \
         --backend vllm \
         --model $MODEL  \
         --dataset-name random \
-        --random-input-len $INPUT_LEN \
+        --random-input-len $adjusted_input_len \
         --random-output-len $OUTPUT_LEN \
         --ignore-eos \
         --disable-tqdm \
@@ -188,11 +169,11 @@ run_benchmark() {
             curl -X POST http://0.0.0.0:8004/reset_prefix_cache
             sleep 5
             bm_log="$LOG_FOLDER/bm_log_${max_num_seqs}_${max_num_batched_tokens}_requestrate_${request_rate}.txt"
-            python benchmarks/benchmark_serving.py \
+            vllm bench serve \
                 --backend vllm \
                 --model $MODEL  \
                 --dataset-name random \
-                --random-input-len $INPUT_LEN \
+                --random-input-len $adjusted_input_len \
                 --random-output-len $OUTPUT_LEN \
                 --ignore-eos \
                 --disable-tqdm \
@@ -273,4 +254,3 @@ done
 echo "finish permutations"
 echo "best_max_num_seqs: $best_max_num_seqs, best_num_batched_tokens: $best_num_batched_tokens, best_throughput: $best_throughput, profile saved in: $PROFILE_PATH"
 echo "best_max_num_seqs: $best_max_num_seqs, best_num_batched_tokens: $best_num_batched_tokens, best_throughput: $best_throughput, profile saved in: $PROFILE_PATH" >> "$RESULT"
-

benchmarks/benchmark_dataset.py

@@ -324,6 +324,9 @@ class RandomDataset(BenchmarkDataset):
         input_low = int(real_input_len * (1 - range_ratio))
         input_high = int(real_input_len * (1 + range_ratio))
         output_low = int(output_len * (1 - range_ratio))
+        # Ensure the lower bound for output length is at least 1 to prevent
+        # sampling 0 tokens, which can cause request failures.
+        output_low = max(output_low, 1)
         output_high = int(output_len * (1 + range_ratio))
 
         # Add logging for debugging
@@ -701,6 +704,7 @@ class HuggingFaceDataset(BenchmarkDataset):
         self,
         dataset_path: str,
         dataset_split: str,
+        no_stream: bool = False,
         dataset_subset: Optional[str] = None,
         **kwargs,
     ) -> None:
@@ -708,6 +712,7 @@ class HuggingFaceDataset(BenchmarkDataset):
 
         self.dataset_split = dataset_split
         self.dataset_subset = dataset_subset
+        self.load_stream = not no_stream
         self.load_data()
 
     def load_data(self) -> None:
@@ -716,7 +721,7 @@ class HuggingFaceDataset(BenchmarkDataset):
             self.dataset_path,
             name=self.dataset_subset,
             split=self.dataset_split,
-            streaming=True,
+            streaming=self.load_stream,
         )
         self.data = self.data.shuffle(seed=self.random_seed)
 

benchmarks/benchmark_latency.py

@@ -11,6 +11,7 @@ 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
@@ -34,6 +35,10 @@ def save_to_pytorch_benchmark_format(
         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)
 

benchmarks/benchmark_serving.py

@@ -30,7 +30,7 @@ import os
 import random
 import time
 import warnings
-from collections.abc import AsyncGenerator, Iterable
+from collections.abc import Iterable
 from dataclasses import dataclass
 from datetime import datetime
 from typing import Any, Literal, Optional
@@ -38,6 +38,7 @@ from typing import Any, Literal, Optional
 import numpy as np
 from tqdm.asyncio import tqdm
 from transformers import PreTrainedTokenizerBase
+from typing_extensions import deprecated
 
 from backend_request_func import (
     ASYNC_REQUEST_FUNCS,
@@ -73,6 +74,7 @@ from benchmark_dataset import (
     VisionArenaDataset,
 )
 from benchmark_utils import convert_to_pytorch_benchmark_format, write_to_json
+from vllm.benchmarks.serve import get_request
 
 MILLISECONDS_TO_SECONDS_CONVERSION = 1000
 
@@ -107,101 +109,6 @@ class BenchmarkMetrics:
     percentiles_e2el_ms: list[tuple[float, float]]
 
 
-def _get_current_request_rate(
-    ramp_up_strategy: Optional[Literal["linear", "exponential"]],
-    ramp_up_start_rps: Optional[int],
-    ramp_up_end_rps: Optional[int],
-    request_index: int,
-    total_requests: int,
-    request_rate: float,
-) -> float:
-    if (
-        ramp_up_strategy
-        and ramp_up_start_rps is not None
-        and ramp_up_end_rps is not None
-    ):
-        progress = request_index / max(total_requests - 1, 1)
-        if ramp_up_strategy == "linear":
-            increase = (ramp_up_end_rps - ramp_up_start_rps) * progress
-            return ramp_up_start_rps + increase
-        elif ramp_up_strategy == "exponential":
-            ratio = ramp_up_end_rps / ramp_up_start_rps
-            return ramp_up_start_rps * (ratio**progress)
-        else:
-            raise ValueError(f"Unknown ramp-up strategy: {ramp_up_strategy}")
-    return request_rate
-
-
-async def get_request(
-    input_requests: list[SampleRequest],
-    request_rate: float,
-    burstiness: float = 1.0,
-    ramp_up_strategy: Optional[Literal["linear", "exponential"]] = None,
-    ramp_up_start_rps: Optional[int] = None,
-    ramp_up_end_rps: Optional[int] = None,
-) -> AsyncGenerator[tuple[SampleRequest, float], None]:
-    """
-    Asynchronously generates requests at a specified rate
-    with OPTIONAL burstiness and OPTIONAL ramp-up strategy.
-
-    Args:
-        input_requests:
-            A list of input requests, each represented as a SampleRequest.
-        request_rate:
-            The rate at which requests are generated (requests/s).
-        burstiness (optional):
-            The burstiness factor of the request generation.
-            Only takes effect when request_rate is not inf.
-            Default value is 1, which follows a Poisson process.
-            Otherwise, the request intervals follow a gamma distribution.
-            A lower burstiness value (0 < burstiness < 1) results
-            in more bursty requests, while a higher burstiness value
-            (burstiness > 1) results in a more uniform arrival of requests.
-         ramp_up_strategy (optional):
-            The ramp-up strategy. Can be "linear" or "exponential".
-            If None, uses constant request rate (specified by request_rate).
-        ramp_up_start_rps (optional):
-            The starting request rate for ramp-up.
-        ramp_up_end_rps (optional):
-            The ending request rate for ramp-up.
-    """
-    assert burstiness > 0, (
-        f"A positive burstiness factor is expected, but given {burstiness}."
-    )
-    # Convert to list to get length for ramp-up calculations
-    if isinstance(input_requests, Iterable) and not isinstance(input_requests, list):
-        input_requests = list(input_requests)
-
-    total_requests = len(input_requests)
-    request_index = 0
-
-    for request in input_requests:
-        current_request_rate = _get_current_request_rate(
-            ramp_up_strategy,
-            ramp_up_start_rps,
-            ramp_up_end_rps,
-            request_index,
-            total_requests,
-            request_rate,
-        )
-
-        yield request, current_request_rate
-
-        request_index += 1
-
-        if current_request_rate == float("inf"):
-            # If the request rate is infinity, then we don't need to wait.
-            continue
-
-        theta = 1.0 / (current_request_rate * burstiness)
-
-        # Sample the request interval from the gamma distribution.
-        # If burstiness is 1, it follows exponential distribution.
-        interval = np.random.gamma(shape=burstiness, scale=theta)
-        # The next request will be sent after the interval.
-        await asyncio.sleep(interval)
-
-
 def calculate_metrics(
     input_requests: list[SampleRequest],
     outputs: list[RequestFuncOutput],
@@ -551,7 +458,7 @@ async def benchmark(
         "total_input_tokens": metrics.total_input,
         "total_output_tokens": metrics.total_output,
         "request_throughput": metrics.request_throughput,
-        "request_goodput:": metrics.request_goodput if goodput_config_dict else None,
+        "request_goodput": metrics.request_goodput if goodput_config_dict else None,
         "output_throughput": metrics.output_throughput,
         "total_token_throughput": metrics.total_token_throughput,
         "input_lens": [output.prompt_len for output in outputs],
@@ -687,6 +594,10 @@ def save_to_pytorch_benchmark_format(
         write_to_json(pt_file, pt_records)
 
 
+@deprecated(
+    "benchmark_serving.py is deprecated and will be removed in a future "
+    "version. Please use 'vllm bench serve' instead.",
+)
 def main(args: argparse.Namespace):
     print(args)
     random.seed(args.seed)
@@ -825,6 +736,7 @@ def main(args: argparse.Namespace):
             dataset_subset=args.hf_subset,
             dataset_split=args.hf_split,
             random_seed=args.seed,
+            no_stream=args.no_stream,
         ).sample(
             num_requests=args.num_prompts,
             tokenizer=tokenizer,
@@ -1033,6 +945,11 @@ def create_argument_parser():
         help="Path to the sharegpt/sonnet dataset. "
         "Or the huggingface dataset ID if using HF dataset.",
     )
+    parser.add_argument(
+        "--no-stream",
+        action="store_true",
+        help="Do not load the dataset in streaming mode.",
+    )
     parser.add_argument(
         "--max-concurrency",
         type=int,

benchmarks/benchmark_throughput.py

@@ -15,6 +15,7 @@ import torch
 import uvloop
 from tqdm import tqdm
 from transformers import AutoModelForCausalLM, AutoTokenizer, PreTrainedTokenizerBase
+from typing_extensions import deprecated
 
 from benchmark_dataset import (
     AIMODataset,
@@ -167,7 +168,8 @@ async def run_vllm_async(
     from vllm import SamplingParams
 
     async with build_async_engine_client_from_engine_args(
-        engine_args, disable_frontend_multiprocessing
+        engine_args,
+        disable_frontend_multiprocessing=disable_frontend_multiprocessing,
     ) as llm:
         model_config = await llm.get_model_config()
         assert all(
@@ -356,6 +358,7 @@ def get_requests(args, tokenizer):
     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
@@ -380,6 +383,10 @@ def get_requests(args, tokenizer):
     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
@@ -610,6 +617,11 @@ def create_argument_parser():
         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,

benchmarks/disagg_benchmarks/disagg_overhead_benchmark.sh

@@ -3,7 +3,7 @@
 # benchmark the overhead of disaggregated prefill.
 # methodology:
 # - send all request to prefill vLLM instance. It will buffer KV cache.
-# - then send all request to decode instance. 
+# - then send all request to decode instance.
 # - The TTFT of decode instance is the overhead.
 
 set -ex
@@ -12,6 +12,8 @@ kill_gpu_processes() {
   # kill all processes on GPU.
   pgrep pt_main_thread | xargs -r kill -9
   pgrep python3 | xargs -r kill -9
+  # vLLM now names the process with VLLM prefix after https://github.com/vllm-project/vllm/pull/21445
+  pgrep VLLM | xargs -r kill -9
   sleep 10
 
   # remove vllm config file
@@ -61,7 +63,7 @@ benchmark() {
     --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}' &
-    
+
 
   CUDA_VISIBLE_DEVICES=1 python3 \
     -m vllm.entrypoints.openai.api_server \
@@ -76,38 +78,38 @@ benchmark() {
   wait_for_server 8200
 
   # let the prefill instance finish prefill
-  python3 ../benchmark_serving.py \
-          --backend vllm \
-          --model $model \
-          --dataset-name $dataset_name \
-          --dataset-path $dataset_path \
-          --sonnet-input-len $input_len \
-          --sonnet-output-len "$output_len" \
-          --sonnet-prefix-len $prefix_len \
-          --num-prompts $num_prompts \
-          --port 8100 \
-          --save-result \
-          --result-dir $results_folder \
-          --result-filename disagg_prefill_tp1.json \
-          --request-rate "inf"
+  vllm bench serve \
+    --backend vllm \
+    --model $model \
+    --dataset-name $dataset_name \
+    --dataset-path $dataset_path \
+    --sonnet-input-len $input_len \
+    --sonnet-output-len "$output_len" \
+    --sonnet-prefix-len $prefix_len \
+    --num-prompts $num_prompts \
+    --port 8100 \
+    --save-result \
+    --result-dir $results_folder \
+    --result-filename disagg_prefill_tp1.json \
+    --request-rate "inf"
 
 
   # send the request to decode.
   # The TTFT of this command will be the overhead of disagg prefill impl.
-  python3 ../benchmark_serving.py \
-          --backend vllm \
-          --model $model \
-          --dataset-name $dataset_name \
-          --dataset-path $dataset_path \
-          --sonnet-input-len $input_len \
-          --sonnet-output-len "$output_len" \
-          --sonnet-prefix-len $prefix_len \
-          --num-prompts $num_prompts \
-          --port 8200 \
-          --save-result \
-          --result-dir $results_folder \
-          --result-filename disagg_prefill_tp1_overhead.json \
-          --request-rate "$qps"
+  vllm bench serve \
+    --backend vllm \
+    --model $model \
+    --dataset-name $dataset_name \
+    --dataset-path $dataset_path \
+    --sonnet-input-len $input_len \
+    --sonnet-output-len "$output_len" \
+    --sonnet-prefix-len $prefix_len \
+    --num-prompts $num_prompts \
+    --port 8200 \
+    --save-result \
+    --result-dir $results_folder \
+    --result-filename disagg_prefill_tp1_overhead.json \
+    --request-rate "$qps"
   kill_gpu_processes
 
 }

benchmarks/disagg_benchmarks/disagg_performance_benchmark.sh

@@ -18,6 +18,8 @@ kill_gpu_processes() {
   # kill all processes on GPU.
   pgrep pt_main_thread | xargs -r kill -9
   pgrep python3 | xargs -r kill -9
+  # vLLM now names the process with VLLM prefix after https://github.com/vllm-project/vllm/pull/21445
+  pgrep VLLM | xargs -r kill -9
   for port in 8000 8100 8200; do lsof -t -i:$port | xargs -r kill -9; done
   sleep 1
 }
@@ -58,7 +60,7 @@ launch_chunked_prefill() {
 
 
 launch_disagg_prefill() {
-  model="meta-llama/Meta-Llama-3.1-8B-Instruct" 
+  model="meta-llama/Meta-Llama-3.1-8B-Instruct"
   # disagg prefill
   CUDA_VISIBLE_DEVICES=0 python3 \
     -m vllm.entrypoints.openai.api_server \
@@ -97,20 +99,20 @@ benchmark() {
   output_len=$2
   tag=$3
 
-  python3 ../benchmark_serving.py \
-          --backend vllm \
-          --model $model \
-          --dataset-name $dataset_name \
-          --dataset-path $dataset_path \
-          --sonnet-input-len $input_len \
-          --sonnet-output-len "$output_len" \
-          --sonnet-prefix-len $prefix_len \
-          --num-prompts $num_prompts \
-          --port 8000 \
-          --save-result \
-          --result-dir $results_folder \
-          --result-filename "$tag"-qps-"$qps".json \
-          --request-rate "$qps"
+  vllm bench serve \
+    --backend vllm \
+    --model $model \
+    --dataset-name $dataset_name \
+    --dataset-path $dataset_path \
+    --sonnet-input-len $input_len \
+    --sonnet-output-len "$output_len" \
+    --sonnet-prefix-len $prefix_len \
+    --num-prompts $num_prompts \
+    --port 8000 \
+    --save-result \
+    --result-dir $results_folder \
+    --result-filename "$tag"-qps-"$qps".json \
+    --request-rate "$qps"
 
   sleep 2
 }

benchmarks/kernels/bench_fp8_gemm.py

@@ -1,4 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 import argparse
 import copy
 import itertools

benchmarks/kernels/bench_nvfp4_gemm.py

@@ -0,0 +1,141 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+import argparse
+import copy
+import itertools
+
+import torch
+from weight_shapes import WEIGHT_SHAPES
+
+from vllm import _custom_ops as ops
+from vllm.platforms import current_platform
+from vllm.scalar_type import scalar_types
+from vllm.triton_utils import triton
+
+if not current_platform.has_device_capability(100):
+    raise RuntimeError("NVFP4 requires compute capability of 10.0 (Blackwell)")
+
+
+FLOAT4_E2M1_MAX = scalar_types.float4_e2m1f.max()
+FLOAT8_E4M3_MAX = torch.finfo(torch.float8_e4m3fn).max
+
+PROVIDER_CFGS = {
+    "torch-bf16": dict(enabled=True),
+    "nvfp4": dict(no_a_quant=False, enabled=True),
+    "nvfp4-noquant": dict(no_a_quant=True, enabled=True),
+}
+
+_enabled = [k for k, v in PROVIDER_CFGS.items() if v["enabled"]]
+
+
+def _quant_weight_nvfp4(b: torch.Tensor, device: str):
+    # Compute global scale for weight
+    b_amax = torch.abs(b).max().to(torch.float32)
+    b_global_scale = FLOAT8_E4M3_MAX * FLOAT4_E2M1_MAX / b_amax
+    b_fp4, scale_b_fp4 = ops.scaled_fp4_quant(b, b_global_scale)
+    return b_fp4, scale_b_fp4, b_global_scale
+
+
+def build_nvfp4_runner(cfg, a, b, dtype, device):
+    b_fp4, scale_b_fp4, b_global_scale = _quant_weight_nvfp4(b, device)
+
+    # Compute global scale for activation
+    # NOTE: This is generally provided ahead-of-time by the model checkpoint.
+    a_amax = torch.abs(a).max().to(torch.float32)
+    a_global_scale = FLOAT8_E4M3_MAX * FLOAT4_E2M1_MAX / a_amax
+
+    # Alpha for the GEMM operation
+    alpha = 1.0 / (a_global_scale * b_global_scale)
+
+    if cfg["no_a_quant"]:
+        # Pre-quantize activation
+        a_fp4, scale_a_fp4 = ops.scaled_fp4_quant(a, a_global_scale)
+
+        def run():
+            return ops.cutlass_scaled_fp4_mm(
+                a_fp4, b_fp4, scale_a_fp4, scale_b_fp4, alpha, dtype
+            )
+
+        return run
+
+    # Quantize activation on-the-fly
+    def run():
+        a_fp4, scale_a_fp4 = ops.scaled_fp4_quant(a, a_global_scale)
+        return ops.cutlass_scaled_fp4_mm(
+            a_fp4, b_fp4, scale_a_fp4, scale_b_fp4, alpha, dtype
+        )
+
+    return run
+
+
+@triton.testing.perf_report(
+    triton.testing.Benchmark(
+        x_names=["batch_size"],
+        x_vals=[1, 16, 64, 128, 256, 512, 1024, 2048, 4096, 8192, 16384],
+        x_log=False,
+        line_arg="provider",
+        line_vals=_enabled,
+        line_names=_enabled,
+        ylabel="TFLOP/s (larger is better)",
+        plot_name="BF16 vs NVFP4 GEMMs",
+        args={},
+    )
+)
+def benchmark(batch_size, provider, N, K):
+    M = batch_size
+    device = "cuda"
+    dtype = torch.bfloat16
+
+    a = torch.randn((M, K), device=device, dtype=dtype)
+    b = torch.randn((N, K), device=device, dtype=dtype)
+
+    quantiles = [0.5, 0.2, 0.8]
+
+    if provider == "torch-bf16":
+        ms, min_ms, max_ms = triton.testing.do_bench_cudagraph(
+            lambda: torch.nn.functional.linear(a, b), quantiles=quantiles
+        )
+    else:
+        cfg = PROVIDER_CFGS[provider]
+        run_quant = build_nvfp4_runner(cfg, a, b, dtype, device)
+        ms, min_ms, max_ms = triton.testing.do_bench_cudagraph(
+            lambda: run_quant(), quantiles=quantiles
+        )
+
+    to_tflops = lambda t_ms: (2 * M * N * K) * 1e-12 / (t_ms * 1e-3)
+    return to_tflops(ms), to_tflops(max_ms), to_tflops(min_ms)
+
+
+def prepare_shapes(args):
+    out = []
+    for model, tp_size in itertools.product(args.models, args.tp_sizes):
+        for KN, tp_dim in copy.deepcopy(WEIGHT_SHAPES[model]):
+            KN[tp_dim] //= tp_size
+            KN.append(model)
+            out.append(KN)
+    return out
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--models",
+        nargs="+",
+        type=str,
+        default=["meta-llama/Llama-3.1-8B-Instruct"],
+        choices=list(WEIGHT_SHAPES.keys()),
+    )
+    parser.add_argument("--tp-sizes", nargs="+", type=int, default=[1])
+    args = parser.parse_args()
+
+    for K, N, model in prepare_shapes(args):
+        print(f"{model}, N={N} K={K}, BF16 vs NVFP4 GEMMs TFLOP/s:")
+        benchmark.run(
+            print_data=True,
+            show_plots=True,
+            save_path=f"bench_nvfp4_res_n{N}_k{K}",
+            N=N,
+            K=K,
+        )
+
+    print("Benchmark finished!")

benchmarks/kernels/bench_per_token_quant_fp8.py

@@ -0,0 +1,98 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+import itertools
+from typing import Callable
+
+import torch
+
+from vllm import _custom_ops as ops
+from vllm.config import CompilationConfig, VllmConfig, set_current_vllm_config
+from vllm.model_executor.layers.quantization.input_quant_fp8 import QuantFP8
+from vllm.model_executor.layers.quantization.utils.quant_utils import GroupShape
+from vllm.triton_utils import triton
+
+
+# TODO(luka): use standalone_compile utility
+def with_dyn_arg(fn: Callable, arg_index: int, dim_index: int):
+    def inner(*args):
+        torch._dynamo.mark_dynamic(args[arg_index], dim_index)
+        return fn(*args)
+
+    return inner
+
+
+torch._dynamo.config.recompile_limit = 8888
+compilation_config = CompilationConfig(custom_ops=["none"])
+with set_current_vllm_config(VllmConfig(compilation_config=compilation_config)):
+    torch_per_token_quant_fp8 = torch.compile(
+        QuantFP8(False, GroupShape.PER_TOKEN),
+        fullgraph=True,
+        dynamic=False,  # recompile for different shapes
+    )
+
+    # First dim is explicitly dynamic to simulate vLLM usage
+    torch_per_token_quant_fp8 = with_dyn_arg(torch_per_token_quant_fp8, 0, 0)
+
+
+def cuda_per_token_quant_fp8(
+    input: torch.Tensor,
+) -> tuple[torch.Tensor, torch.Tensor]:
+    return ops.scaled_fp8_quant(input)
+
+
+def calculate_diff(batch_size: int, seq_len: int):
+    """Calculate difference between Triton and CUDA implementations."""
+    device = torch.device("cuda")
+    x = torch.rand((batch_size * seq_len, 4096), dtype=torch.float16, device=device)
+
+    torch_out, torch_scale = torch_per_token_quant_fp8(x)
+    cuda_out, cuda_scale = cuda_per_token_quant_fp8(x)
+
+    if torch.allclose(
+        cuda_out.to(torch.float32), torch_out.to(torch.float32), rtol=1e-3, atol=1e-5
+    ) and torch.allclose(cuda_scale, torch_scale, rtol=1e-3, atol=1e-5):
+        print("✅ All implementations match")
+    else:
+        print("❌ Implementations differ")
+
+
+batch_size_range = [1, 16, 32, 64, 128]
+seq_len_range = [1, 16, 64, 128, 256, 512, 1024, 2048, 4096]
+
+configs = list(itertools.product(batch_size_range, seq_len_range))
+
+
+@triton.testing.perf_report(
+    triton.testing.Benchmark(
+        x_names=["batch_size", "seq_len"],
+        x_vals=configs,
+        line_arg="provider",
+        line_vals=["torch", "cuda"],
+        line_names=["Torch", "CUDA"],
+        styles=[("blue", "-"), ("green", "-")],
+        ylabel="us",
+        plot_name="per-token-dynamic-quant-fp8-performance",
+        args={},
+    )
+)
+def benchmark_quantization(batch_size, seq_len, provider):
+    dtype = torch.float16
+    device = torch.device("cuda")
+
+    x = torch.randn(batch_size * seq_len, 4096, device=device, dtype=dtype)
+
+    quantiles = [0.5, 0.2, 0.8]
+
+    if provider == "torch":
+        fn = lambda: torch_per_token_quant_fp8(x.clone())
+    elif provider == "cuda":
+        fn = lambda: cuda_per_token_quant_fp8(x.clone())
+
+    ms, min_ms, max_ms = triton.testing.do_bench_cudagraph(fn, quantiles=quantiles)
+
+    return 1000 * ms, 1000 * max_ms, 1000 * min_ms
+
+
+if __name__ == "__main__":
+    calculate_diff(batch_size=4, seq_len=4096)
+    benchmark_quantization.run(print_data=True)

benchmarks/kernels/benchmark_grouped_gemm_cutlass.py

@@ -113,6 +113,7 @@ def bench_run(
         w2_scale: torch.Tensor,
         topk_weights: torch.Tensor,
         topk_ids: torch.Tensor,
+        per_act_token: bool,
         num_repeats: int,
     ):
         for _ in range(num_repeats):
@@ -124,7 +125,8 @@ def bench_run(
                 topk_ids,
                 w1_scale,
                 w2_scale,
-                a1_scale=a_scale,
+                per_act_token,
+                a1_scale=None,
             )
 
     def run_cutlass_from_graph(
@@ -148,7 +150,8 @@ def bench_run(
                 topk_ids,
                 w1_scale,
                 w2_scale,
-                a1_scale=a_scale,
+                per_act_token,
+                a1_scale=None,
             )
 
     def run_triton_from_graph(
@@ -227,6 +230,7 @@ def bench_run(
         "w2_q": w2_q,
         "w1_scale": w1_scale,
         "w2_scale": w2_scale,
+        "per_act_token": per_act_token,
         # cuda graph params
         "cutlass_graph": cutlass_graph,
         "triton_graph": triton_graph,
@@ -287,12 +291,13 @@ def bench_run(
         w2_scale,
         topk_weights,
         topk_ids,
+        per_act_token,
         num_warmup,
     )
 
     results.append(
         benchmark.Timer(
-            stmt="run_cutlass_moe(a, a_scale, w1_q, w2_q, w1_scale, w2_scale, topk_weights, topk_ids, num_runs)",  # noqa: E501
+            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
             globals=globals,
             label=label,
             sub_label=sub_label,

benchmarks/kernels/benchmark_machete.py

@@ -234,8 +234,10 @@ def marlin_create_bench_fn(bt: BenchmarkTensors) -> Callable:
 
         fn = lambda: ops.gptq_marlin_gemm(
             a=bt.a,
+            c=None,
             b_q_weight=w_q,
             b_scales=w_s,
+            global_scale=None,
             b_zeros=w_zp,
             g_idx=g_idx,
             perm=sort_indices,

benchmarks/kernels/benchmark_moe.py

@@ -86,6 +86,9 @@ def benchmark_config(
             (num_experts, 2 * shard_intermediate_size), dtype=torch.float32
         )
         w2_scale = torch.randn((hidden_size, num_experts), dtype=torch.float32)
+    if use_deep_gemm:
+        # we use the default block shape for deepgemm
+        block_quant_shape = [128, 128]
     if use_fp8_w8a8:
         if block_quant_shape:
             block_n, block_k = block_quant_shape[0], block_quant_shape[1]
@@ -573,7 +576,11 @@ def main(args: argparse.Namespace):
         topk = config.num_experts_per_tok
         intermediate_size = config.intermediate_size
         shard_intermediate_size = 2 * intermediate_size // args.tp_size
-    elif config.architectures[0] in ("DeepseekV3ForCausalLM", "DeepseekV2ForCausalLM"):
+    elif config.architectures[0] in (
+        "DeepseekV3ForCausalLM",
+        "DeepseekV2ForCausalLM",
+        "Glm4MoeForCausalLM",
+    ):
         E = config.n_routed_experts
         topk = config.num_experts_per_tok
         intermediate_size = config.moe_intermediate_size
@@ -583,6 +590,11 @@ def main(args: argparse.Namespace):
         topk = config.num_experts_per_tok
         intermediate_size = config.moe_intermediate_size
         shard_intermediate_size = 2 * intermediate_size // args.tp_size
+    elif config.architectures[0] in ("HunYuanMoEV1ForCausalLM"):
+        E = config.num_experts
+        topk = config.moe_topk[0]
+        intermediate_size = config.moe_intermediate_size[0]
+        shard_intermediate_size = 2 * intermediate_size // args.tp_size
     else:
         # Support for llama4
         config = config.get_text_config()
@@ -620,7 +632,7 @@ def main(args: argparse.Namespace):
             4096,
         ]
     else:
-        batch_sizes = [args.batch_size]
+        batch_sizes = args.batch_size
 
     use_deep_gemm = bool(args.use_deep_gemm)
 
@@ -728,7 +740,7 @@ if __name__ == "__main__":
     )
     parser.add_argument("--use-deep-gemm", action="store_true")
     parser.add_argument("--seed", type=int, default=0)
-    parser.add_argument("--batch-size", type=int, required=False)
+    parser.add_argument("--batch-size", type=int, nargs="+", required=False)
     parser.add_argument("--tune", action="store_true")
     parser.add_argument("--trust-remote-code", action="store_true")
     parser.add_argument("--model-prefix", type=str, required=False)

benchmarks/kernels/benchmark_moe_align_block_size.py

@@ -5,9 +5,8 @@ import itertools
 
 import torch
 
-from vllm import _custom_ops as ops
 from vllm.model_executor.layers.fused_moe.moe_align_block_size import (
-    moe_align_block_size_triton,
+    moe_align_block_size,
 )
 from vllm.triton_utils import triton
 
@@ -21,62 +20,6 @@ def get_topk_ids(num_tokens: int, num_experts: int, topk: int) -> torch.Tensor:
     )
 
 
-def check_correctness(num_tokens, num_experts=256, block_size=256, topk=8):
-    """
-    Verifies vllm vs. Triton
-    """
-    topk_ids = get_topk_ids(num_tokens, num_experts, topk)
-
-    # 1. malloc space for triton and vllm
-    # malloc enough space (max_num_tokens_padded) for the sorted ids
-    max_num_tokens_padded = topk_ids.numel() + num_experts * (block_size - 1)
-    sorted_ids_triton = torch.empty(
-        (max_num_tokens_padded,), dtype=torch.int32, device="cuda"
-    )
-    sorted_ids_triton.fill_(topk_ids.numel())  # fill with sentinel value
-    expert_ids_triton = torch.zeros(
-        (max_num_tokens_padded // block_size,), dtype=torch.int32, device="cuda"
-    )
-    num_tokens_post_pad_triton = torch.empty((1,), dtype=torch.int32, device="cuda")
-
-    sorted_ids_vllm = torch.empty_like(sorted_ids_triton)
-    sorted_ids_vllm.fill_(topk_ids.numel())
-    expert_ids_vllm = torch.zeros_like(expert_ids_triton)
-    num_tokens_post_pad_vllm = torch.empty_like(num_tokens_post_pad_triton)
-
-    # 2. run implementations
-    moe_align_block_size_triton(
-        topk_ids,
-        num_experts,
-        block_size,
-        sorted_ids_triton,
-        expert_ids_triton,
-        num_tokens_post_pad_triton,
-    )
-
-    ops.moe_align_block_size(
-        topk_ids,
-        num_experts,
-        block_size,
-        sorted_ids_vllm,
-        expert_ids_vllm,
-        num_tokens_post_pad_vllm,
-    )
-    print(f"✅ VLLM implementation works with {num_experts} experts!")
-
-    # 3. compare results
-    if torch.allclose(expert_ids_triton, expert_ids_vllm) and torch.allclose(
-        num_tokens_post_pad_triton, num_tokens_post_pad_vllm
-    ):
-        print("✅ Triton and VLLM implementations match.")
-    else:
-        print("❌ Triton and VLLM implementations DO NOT match.")
-        print("Triton expert_ids:", expert_ids_triton)
-        print("VLLM expert_ids:", expert_ids_vllm)
-        print("Triton num_tokens_post_pad:", num_tokens_post_pad_triton)
-        print("VLLM num_tokens_post_pad:", num_tokens_post_pad_vllm)
-
-
 # test configurations
 num_tokens_range = [1, 16, 256, 4096]
 num_experts_range = [16, 64, 224, 256, 280, 512]
@@ -89,8 +32,8 @@ configs = list(itertools.product(num_tokens_range, num_experts_range, topk_range
         x_names=["num_tokens", "num_experts", "topk"],
         x_vals=configs,
         line_arg="provider",
-        line_vals=["vllm", "triton"],  # "triton"
-        line_names=["VLLM", "Triton"],  # "Triton"
+        line_vals=["vllm"],
+        line_names=["vLLM"],
         plot_name="moe-align-block-size-performance",
         args={},
     )
@@ -100,37 +43,11 @@ def benchmark(num_tokens, num_experts, topk, provider):
     block_size = 256
     topk_ids = get_topk_ids(num_tokens, num_experts, topk)
 
-    max_num_tokens_padded = topk_ids.numel() + num_experts * (block_size - 1)
-    sorted_ids = torch.empty((max_num_tokens_padded,), dtype=torch.int32, device="cuda")
-    sorted_ids.fill_(topk_ids.numel())
-    max_num_m_blocks = max_num_tokens_padded // block_size
-    expert_ids = torch.empty((max_num_m_blocks,), dtype=torch.int32, device="cuda")
-    num_tokens_post_pad = torch.empty((1,), dtype=torch.int32, device="cuda")
-
     quantiles = [0.5, 0.2, 0.8]
 
     if provider == "vllm":
         ms, min_ms, max_ms = triton.testing.do_bench(
-            lambda: ops.moe_align_block_size(
-                topk_ids,
-                num_experts,
-                block_size,
-                sorted_ids.clone(),
-                expert_ids.clone(),
-                num_tokens_post_pad.clone(),
-            ),
-            quantiles=quantiles,
-        )
-    elif provider == "triton":
-        ms, min_ms, max_ms = triton.testing.do_bench(
-            lambda: moe_align_block_size_triton(
-                topk_ids,
-                num_experts,
-                block_size,
-                sorted_ids.clone(),
-                expert_ids.clone(),
-                num_tokens_post_pad.clone(),
-            ),
+            lambda: moe_align_block_size(topk_ids, block_size, num_experts),
             quantiles=quantiles,
         )
 
@@ -154,6 +71,4 @@ if __name__ == "__main__":
     )
     args = parser.parse_args()
 
-    print("Running correctness check...")
-    check_correctness(num_tokens=1024, num_experts=args.num_experts, topk=args.topk)
     benchmark.run(print_data=True, show_plots=True)

benchmarks/kernels/benchmark_moe_permute_unpermute.py

@@ -8,12 +8,13 @@ import ray
 import torch
 from transformers import AutoConfig
 
-from vllm.model_executor.layers.fused_moe.deep_gemm_moe import (
+from vllm.model_executor.layers.fused_moe.fused_moe import *
+from vllm.model_executor.layers.fused_moe.moe_permute_unpermute import (
     _moe_permute,
     _moe_unpermute_and_reduce,
+    moe_permute,
+    moe_unpermute,
 )
-from vllm.model_executor.layers.fused_moe.fused_moe import *
-from vllm.model_executor.layers.fused_moe.moe_permute_unpermute import *
 from vllm.model_executor.layers.fused_moe.utils import _fp8_quantize
 from vllm.platforms import current_platform
 from vllm.utils import FlexibleArgumentParser
@@ -63,18 +64,19 @@ def benchmark_permute(
 
     def run():
         if use_customized_permute:
-            (permuted_hidden_states, first_token_off, inv_perm_idx, m_indices) = (
-                moe_permute(
-                    qhidden_states,
-                    topk_weights=topk_weights,
-                    topk_ids=topk_ids,
-                    token_expert_indices=token_expert_indices,
-                    topk=topk,
-                    n_expert=num_experts,
-                    n_local_expert=num_experts,
-                    expert_map=None,
-                    align_block_size=align_block_size,
-                )
+            (
+                permuted_hidden_states,
+                a1q_scale,
+                first_token_off,
+                inv_perm_idx,
+                m_indices,
+            ) = moe_permute(
+                qhidden_states,
+                a1q_scale=None,
+                topk_ids=topk_ids,
+                n_expert=num_experts,
+                expert_map=None,
+                align_block_size=align_block_size,
             )
         else:
             (
@@ -150,18 +152,19 @@ def benchmark_unpermute(
 
     def prepare():
         if use_customized_permute:
-            (permuted_hidden_states, first_token_off, inv_perm_idx, m_indices) = (
-                moe_permute(
-                    qhidden_states,
-                    topk_weights=topk_weights,
-                    topk_ids=topk_ids,
-                    token_expert_indices=token_expert_indices,
-                    topk=topk,
-                    n_expert=num_experts,
-                    n_local_expert=num_experts,
-                    expert_map=None,
-                    align_block_size=align_block_size,
-                )
+            (
+                permuted_hidden_states,
+                a1q_scale,
+                first_token_off,
+                inv_perm_idx,
+                m_indices,
+            ) = moe_permute(
+                qhidden_states,
+                a1q_scale=None,
+                topk_ids=topk_ids,
+                n_expert=num_experts,
+                expert_map=None,
+                align_block_size=align_block_size,
             )
             # convert to fp16/bf16 as gemm output
             return (
@@ -191,16 +194,19 @@ def benchmark_unpermute(
 
     def run(input: tuple):
         if use_customized_permute:
-            (permuted_hidden_states, first_token_off, inv_perm_idx, m_indices) = input
+            (
+                permuted_hidden_states,
+                first_token_off,
+                inv_perm_idx,
+                m_indices,
+            ) = input
+            output = torch.empty_like(hidden_states)
             moe_unpermute(
+                output,
                 permuted_hidden_states,
                 topk_weights,
-                topk_ids,
                 inv_perm_idx,
                 first_token_off,
-                topk,
-                num_experts,
-                num_experts,
             )
         else:
             (
@@ -211,7 +217,11 @@ def benchmark_unpermute(
                 inv_perm,
             ) = input
             _moe_unpermute_and_reduce(
-                output_hidden_states, permuted_hidden_states, inv_perm, topk_weights
+                output_hidden_states,
+                permuted_hidden_states,
+                inv_perm,
+                topk_weights,
+                True,
             )
 
     # JIT compilation & warmup
@@ -318,6 +328,7 @@ def main(args: argparse.Namespace):
     elif (
         config.architectures[0] == "DeepseekV3ForCausalLM"
         or config.architectures[0] == "DeepseekV2ForCausalLM"
+        or config.architectures[0] == "Glm4MoeForCausalLM"
     ):
         E = config.n_routed_experts
         topk = config.num_experts_per_tok

benchmarks/kernels/benchmark_trtllm_attention.py

@@ -0,0 +1,240 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+import csv
+import os
+import random
+from datetime import datetime
+
+import flashinfer
+import torch
+
+FLOAT32_BYTES = torch.finfo(torch.float).bits // 8
+
+# KV Cache Layout for TRT-LLM
+# kv_cache_shape = (num_blocks, 2, num_kv_heads, page_size, head_dim)
+
+
+def to_float8(x, dtype=torch.float8_e4m3fn):
+    finfo = torch.finfo(dtype)
+    min_val, max_val = x.aminmax()
+    amax = torch.maximum(min_val.abs(), max_val.abs()).clamp(min=1e-12)
+    scale = finfo.max / amax * 0.1
+    x_scl_sat = (x * scale).clamp(min=finfo.min, max=finfo.max)
+    return x_scl_sat.to(dtype), scale.float().reciprocal()
+
+
+@torch.no_grad()
+def benchmark_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,
+):
+    torch.set_default_device("cuda")
+    device = "cuda"
+    torch.manual_seed(0)
+
+    # Currently only HEAD_GRP_SIZE == 8 is supported
+    HEAD_GRP_SIZE = 8
+    MAX_SEQ_LEN = max_seq_len
+
+    # large number to reduce kv_cache reuse
+    NUM_BLOCKS = int(256000 / page_size)
+
+    workspace_buffer = torch.empty(1024 * 1024 * 1024, dtype=torch.int8, device=device)
+
+    # 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)]
+
+    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
+
+    block_tables = torch.randint(
+        0, NUM_BLOCKS, (num_seqs, max_num_blocks_per_seq), dtype=torch.int32
+    )
+
+    kv_cache_shape = (NUM_BLOCKS, 2, num_kv_heads, page_size, head_dim)
+    kv_cache = torch.randn(size=kv_cache_shape, device=device, dtype=dtype)
+    k_scale = v_scale = 1.0
+
+    if kv_cache_dtype.startswith("fp8"):
+        kv_cache, _ = to_float8(kv_cache)
+
+    # Benchmark TRT decode
+    def trt_decode():
+        return flashinfer.decode.trtllm_batch_decode_with_kv_cache(
+            q,
+            kv_cache,
+            workspace_buffer,
+            num_qo_heads,
+            num_kv_heads,
+            sm_scale,
+            block_tables,
+            kv_lens_tensor,
+            page_size,
+            max_kv_len,
+            kv_cache_dtype,
+            k_scale,
+            v_scale,
+        )
+
+    def time_fn(fn, warmup=10, trials=20):
+        torch.cuda.synchronize()
+        start = torch.cuda.Event(enable_timing=True)
+        end = torch.cuda.Event(enable_timing=True)
+        times = []
+        for i in range(warmup):
+            fn()
+        for i in range(trials):
+            start.record()
+            fn()
+            end.record()
+            torch.cuda.synchronize()
+            times.append(start.elapsed_time(end))  # ms
+        return sum(times) / len(times), torch.std(torch.tensor(times))
+
+    # 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)
+
+    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,
+    )
+
+    def baseline_decode():
+        return wrapper.run(q, kv_cache, sm_scale, k_scale, v_scale)
+
+    baseline_mean, baseline_std = time_fn(baseline_decode)
+
+    # Calculate percentage speedup (positive means TRT is faster)
+    speedup_percent = (baseline_mean - trt_mean) / baseline_mean
+
+    print(
+        f"\t{num_seqs}\t{max_seq_len}\t{trt_mean:.3f}\t{trt_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(),
+        "baseline_mean": baseline_mean,
+        "baseline_std": baseline_std.item(),
+        "speedup_percent": speedup_percent,
+        "q_dtype": str(dtype),
+        "kv_cache_dtype": kv_cache_dtype,
+        "page_size": page_size,
+        "num_kv_heads": num_kv_heads,
+        "head_dim": head_dim,
+        "max_seq_len": max_seq_len,
+    }
+
+
+def write_results_to_csv(results, filename=None):
+    """Write benchmark results to CSV file."""
+    if filename is None:
+        timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")
+        filename = f"flashinfer_trtllm_benchmark_{timestamp}.csv"
+
+    fieldnames = [
+        "num_seqs",
+        "trt_mean",
+        "trt_std",
+        "baseline_mean",
+        "baseline_std",
+        "speedup_percent",
+        "q_dtype",
+        "kv_cache_dtype",
+        "page_size",
+        "num_kv_heads",
+        "head_dim",
+        "max_seq_len",
+    ]
+
+    file_exists = os.path.exists(filename)
+
+    with open(filename, "a", newline="") as csvfile:
+        writer = csv.DictWriter(csvfile, fieldnames=fieldnames)
+
+        if not file_exists:
+            writer.writeheader()
+
+        for result in results:
+            writer.writerow(result)
+
+    print(f"Results written to {filename}")
+
+
+if __name__ == "__main__":
+    num_seqs = [1, 4, 8, 16, 32, 64, 128, 256]
+    max_seq_lens = [1024, 2048, 4096, 8192, 16384, 32768, 65536, 131072]
+    all_results = []
+
+    print("Running benchmark for kv_cache_dtype: bfloat16")
+    print(
+        "\tnum_seqs\tmax_seq_len\ttrt_mean\ttrt_std\tbaseline_mean\tbaseline_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)
+
+    print("Running benchmark for q_dtype = bfloat16, kv_cache_dtype: fp8")
+    print(
+        "\tnum_seqs\tmax_seq_len\ttrt_mean\ttrt_std\tbaseline_mean\tbaseline_std\tspeedup_percent"
+    )
+    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)
+
+    # Write all results to CSV
+    write_results_to_csv(all_results)

benchmarks/kv_cache/benchmark_block_pool.py

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

cmake/cpu_extension.cmake

@@ -12,9 +12,8 @@ endif()
 #
 # Define environment variables for special configurations
 #
-if(DEFINED ENV{VLLM_CPU_AVX512BF16})
-    set(ENABLE_AVX512BF16 ON)
-endif()
+set(ENABLE_AVX512BF16 $ENV{VLLM_CPU_AVX512BF16})
+set(ENABLE_AVX512VNNI $ENV{VLLM_CPU_AVX512VNNI})
 
 include_directories("${CMAKE_SOURCE_DIR}/csrc")
 
@@ -59,6 +58,22 @@ function (find_isa CPUINFO TARGET OUT)
     endif()
 endfunction()
 
+
+function(check_sysctl TARGET OUT)
+    execute_process(COMMAND sysctl -n "${TARGET}"
+                    RESULT_VARIABLE SYSCTL_RET
+                    OUTPUT_VARIABLE SYSCTL_INFO
+                    ERROR_QUIET
+                    OUTPUT_STRIP_TRAILING_WHITESPACE)
+    if(SYSCTL_RET EQUAL 0 AND
+      (SYSCTL_INFO STREQUAL "1" OR SYSCTL_INFO GREATER 0))
+        set(${OUT} ON PARENT_SCOPE)
+    else()
+        set(${OUT} OFF PARENT_SCOPE)
+    endif()
+endfunction()
+
+
 function (is_avx512_disabled OUT)
     set(DISABLE_AVX512 $ENV{VLLM_CPU_DISABLE_AVX512})
     if(DISABLE_AVX512 AND DISABLE_AVX512 STREQUAL "true")
@@ -71,7 +86,10 @@ endfunction()
 is_avx512_disabled(AVX512_DISABLED)
 
 if (MACOSX_FOUND AND CMAKE_SYSTEM_PROCESSOR STREQUAL "arm64")
-    set(APPLE_SILICON_FOUND TRUE)
+    message(STATUS "Apple Silicon Detected")
+    set(ENABLE_NUMA OFF)
+    check_sysctl(hw.optional.neon ASIMD_FOUND)
+    check_sysctl(hw.optional.arm.FEAT_BF16 ARM_BF16_FOUND)
 else()
     find_isa(${CPUINFO} "avx2" AVX2_FOUND)
     find_isa(${CPUINFO} "avx512f" AVX512_FOUND)
@@ -83,7 +101,6 @@ else()
     find_isa(${CPUINFO} "S390" S390_FOUND)
 endif()
 
-
 if (AVX512_FOUND AND NOT AVX512_DISABLED)
     list(APPEND CXX_COMPILE_FLAGS
         "-mavx512f"
@@ -96,12 +113,30 @@ if (AVX512_FOUND AND NOT AVX512_DISABLED)
         if (CMAKE_CXX_COMPILER_ID STREQUAL "GNU" AND
             CMAKE_CXX_COMPILER_VERSION VERSION_GREATER_EQUAL 12.3)
             list(APPEND CXX_COMPILE_FLAGS "-mavx512bf16")
+            set(ENABLE_AVX512BF16 ON)
         else()
+            set(ENABLE_AVX512BF16 OFF)
             message(WARNING "Disable AVX512-BF16 ISA support, requires gcc/g++ >= 12.3")
         endif()
     else()
+        set(ENABLE_AVX512BF16 OFF)
         message(WARNING "Disable AVX512-BF16 ISA support, no avx512_bf16 found in local CPU flags." " If cross-compilation is required, please set env VLLM_CPU_AVX512BF16=1.")
     endif()
+
+    find_isa(${CPUINFO} "avx512_vnni" AVX512VNNI_FOUND)
+    if (AVX512VNNI_FOUND OR ENABLE_AVX512VNNI)
+        if (CMAKE_CXX_COMPILER_ID STREQUAL "GNU" AND
+            CMAKE_CXX_COMPILER_VERSION VERSION_GREATER_EQUAL 12.3)
+            list(APPEND CXX_COMPILE_FLAGS "-mavx512vnni")
+            set(ENABLE_AVX512VNNI ON)
+        else()
+            set(ENABLE_AVX512VNNI OFF)
+            message(WARNING "Disable AVX512-VNNI ISA support, requires gcc/g++ >= 12.3")
+        endif()
+    else()
+        set(ENABLE_AVX512VNNI OFF)
+        message(WARNING "Disable AVX512-VNNI ISA support, no avx512_vnni found in local CPU flags." " If cross-compilation is required, please set env VLLM_CPU_AVX512VNNI=1.")
+    endif()
     
 elseif (AVX2_FOUND)
     list(APPEND CXX_COMPILE_FLAGS "-mavx2")
@@ -132,9 +167,6 @@ elseif (ASIMD_FOUND)
         set(MARCH_FLAGS "-march=armv8.2-a+dotprod+fp16")  
     endif()
     list(APPEND CXX_COMPILE_FLAGS ${MARCH_FLAGS})     
-elseif(APPLE_SILICON_FOUND)
-    message(STATUS "Apple Silicon Detected")
-    set(ENABLE_NUMA OFF)
 elseif (S390_FOUND)
     message(STATUS "S390 detected")
     # Check for S390 VXE support
@@ -148,17 +180,32 @@ else()
 endif()
 
 #
-# Build oneDNN for W8A8 GEMM kernels (only for x86-AVX512 platforms)
-#
-if (AVX512_FOUND AND NOT AVX512_DISABLED)
+# 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")
+    set(USE_ACL ON)
+else()
+    set(USE_ACL OFF)
+endif()
+
+if ((AVX512_FOUND AND NOT AVX512_DISABLED) OR ASIMD_FOUND)
     FetchContent_Declare(
         oneDNN
         GIT_REPOSITORY https://github.com/oneapi-src/oneDNN.git
-        GIT_TAG  v3.7.1
+        GIT_TAG  v3.8.1
         GIT_PROGRESS TRUE
         GIT_SHALLOW TRUE
     )
 
+    if(USE_ACL)
+        find_library(ARM_COMPUTE_LIBRARY NAMES arm_compute PATHS $ENV{ACL_ROOT_DIR}/build/)
+        if(NOT ARM_COMPUTE_LIBRARY)
+            message(FATAL_ERROR "Could not find ARM Compute Library: please set ACL_ROOT_DIR")
+        endif()
+        set(ONEDNN_AARCH64_USE_ACL "ON")
+        set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wl,-rpath,$ENV{ACL_ROOT_DIR}/build/")
+        endif()
+
     set(ONEDNN_LIBRARY_TYPE "STATIC")
     set(ONEDNN_BUILD_DOC "OFF")
     set(ONEDNN_BUILD_EXAMPLES "OFF")
@@ -231,11 +278,29 @@ if (AVX512_FOUND AND NOT AVX512_DISABLED)
         "csrc/cpu/quant.cpp"
         "csrc/cpu/shm.cpp"
         ${VLLM_EXT_SRC})
+    if (ENABLE_AVX512BF16 AND ENABLE_AVX512VNNI)
+        set(VLLM_EXT_SRC
+            "csrc/cpu/sgl-kernels/gemm.cpp"
+            "csrc/cpu/sgl-kernels/gemm_int8.cpp"
+            "csrc/cpu/sgl-kernels/gemm_fp8.cpp"
+            "csrc/cpu/sgl-kernels/moe.cpp"
+            "csrc/cpu/sgl-kernels/moe_int8.cpp"
+            "csrc/cpu/sgl-kernels/moe_fp8.cpp"
+            ${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)
+    set(VLLM_EXT_SRC
+        "csrc/cpu/quant.cpp"
+        ${VLLM_EXT_SRC})
+endif()
+
+message(STATUS "CPU extension source files: ${VLLM_EXT_SRC}")
 
 #
 # Define extension targets

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 5f3644181c7a15345ce20bfc65af117d3601b524
+          GIT_TAG 1c2624e53c078854e0637ee566c72fe2107e75f4
           GIT_PROGRESS TRUE
           # Don't share the vllm-flash-attn build between build types
           BINARY_DIR ${CMAKE_BINARY_DIR}/vllm-flash-attn

cmake/utils.cmake

@@ -265,8 +265,8 @@ macro(set_gencode_flags_for_srcs)
 endmacro()
 
 #
-# For the given `SRC_CUDA_ARCHS` list of gencode versions in the form 
-#  `<major>.<minor>[letter]` compute the "loose intersection" with the 
+# For the given `SRC_CUDA_ARCHS` list of gencode versions in the form
+#  `<major>.<minor>[letter]` compute the "loose intersection" with the
 #  `TGT_CUDA_ARCHS` list of gencodes. We also support the `+PTX` suffix in
 #  `SRC_CUDA_ARCHS` which indicates that the PTX code should be built when there
 #  is a CUDA_ARCH in `TGT_CUDA_ARCHS` that is equal to or larger than the
@@ -278,7 +278,7 @@ endmacro()
 #  in `SRC_CUDA_ARCHS` that is less or equal to the version in `TGT_CUDA_ARCHS`.
 # We have special handling for x.0a, if x.0a is in `SRC_CUDA_ARCHS` and x.0 is
 #  in `TGT_CUDA_ARCHS` then we should remove x.0a from `SRC_CUDA_ARCHS` and add
-#  x.0a to the result (and remove x.0 from TGT_CUDA_ARCHS). 
+#  x.0a to the result (and remove x.0 from TGT_CUDA_ARCHS).
 # The result is stored in `OUT_CUDA_ARCHS`.
 #
 # Example:
@@ -313,21 +313,16 @@ function(cuda_archs_loose_intersection OUT_CUDA_ARCHS SRC_CUDA_ARCHS TGT_CUDA_AR
   # if x.0a is in SRC_CUDA_ARCHS and x.0 is in CUDA_ARCHS then we should
   # remove x.0a from SRC_CUDA_ARCHS and add x.0a to _CUDA_ARCHS
   set(_CUDA_ARCHS)
-  if ("9.0a" IN_LIST _SRC_CUDA_ARCHS)
-    list(REMOVE_ITEM _SRC_CUDA_ARCHS "9.0a")
-    if ("9.0" IN_LIST TGT_CUDA_ARCHS)
-      list(REMOVE_ITEM _TGT_CUDA_ARCHS "9.0")
-      set(_CUDA_ARCHS "9.0a")
+  foreach(_arch ${_SRC_CUDA_ARCHS})
+    if(_arch MATCHES "\\a$")
+      list(REMOVE_ITEM _SRC_CUDA_ARCHS "${_arch}")
+      string(REPLACE "a" "" _base "${_arch}")
+      if ("${_base}" IN_LIST TGT_CUDA_ARCHS)
+        list(REMOVE_ITEM _TGT_CUDA_ARCHS "${_base}")
+        list(APPEND _CUDA_ARCHS "${_arch}")
+      endif()
     endif()
-  endif()
-
-  if ("10.0a" IN_LIST _SRC_CUDA_ARCHS)
-    list(REMOVE_ITEM _SRC_CUDA_ARCHS "10.0a")
-    if ("10.0" IN_LIST TGT_CUDA_ARCHS)
-      list(REMOVE_ITEM _TGT_CUDA_ARCHS "10.0")
-      set(_CUDA_ARCHS "10.0a")
-    endif()
-  endif()
+  endforeach()
 
   list(SORT _SRC_CUDA_ARCHS COMPARE NATURAL ORDER ASCENDING)
 
@@ -359,7 +354,7 @@ function(cuda_archs_loose_intersection OUT_CUDA_ARCHS SRC_CUDA_ARCHS TGT_CUDA_AR
   endforeach()
 
   list(REMOVE_DUPLICATES _CUDA_ARCHS)
-  
+
   # reapply +PTX suffix to architectures that requested PTX
   set(_FINAL_ARCHS)
   foreach(_arch ${_CUDA_ARCHS})
@@ -370,7 +365,7 @@ function(cuda_archs_loose_intersection OUT_CUDA_ARCHS SRC_CUDA_ARCHS TGT_CUDA_AR
     endif()
   endforeach()
   set(_CUDA_ARCHS ${_FINAL_ARCHS})
-  
+
   set(${OUT_CUDA_ARCHS} ${_CUDA_ARCHS} PARENT_SCOPE)
 endfunction()
 

csrc/attention/attention_kernels.cuh

@@ -24,6 +24,7 @@
 
 #include "attention_dtypes.h"
 #include "attention_utils.cuh"
+#include "../cuda_compat.h"
 
 #ifdef USE_ROCM
   #include <hip/hip_bf16.h>
@@ -33,12 +34,6 @@ typedef __hip_bfloat16 __nv_bfloat16;
   #include "../quantization/fp8/nvidia/quant_utils.cuh"
 #endif
 
-#ifndef USE_ROCM
-  #define WARP_SIZE 32
-#else
-  #define WARP_SIZE warpSize
-#endif
-
 #define MAX(a, b) ((a) > (b) ? (a) : (b))
 #define MIN(a, b) ((a) < (b) ? (a) : (b))
 #define DIVIDE_ROUND_UP(a, b) (((a) + (b) - 1) / (b))
@@ -670,7 +665,6 @@ __global__ void paged_attention_v2_reduce_kernel(
 
 }  // namespace vllm
 
-#undef WARP_SIZE
 #undef MAX
 #undef MIN
 #undef DIVIDE_ROUND_UP

csrc/attention/mla/cutlass_sm100_mla/device/sm100_mla.hpp

@@ -0,0 +1,372 @@
+/***************************************************************************************************
+ * Copyright (c) 2025 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.
+ *
+ **************************************************************************************************/
+/*
+ * Taken from SGLANG PR https://github.com/sgl-project/sglang/pull/6929
+ * by Alcanderian JieXin Liang
+ */
+
+/*!
+ \file
+ \brief An universal device layer for cutlass 3.x-style kernels.
+*/
+
+// clang-format off
+#pragma once
+
+// common
+#include "cutlass/cutlass.h"
+#include "cutlass/device_kernel.h"
+
+#if !defined(__CUDACC_RTC__)
+#include "cutlass/cluster_launch.hpp"
+#include "cutlass/trace.h"
+#endif // !defined(__CUDACC_RTC__)
+
+#include "../kernel/sm100_fmha_mla_tma_warpspecialized.hpp"
+#include "../kernel/sm100_fmha_mla_reduction.hpp"
+
+////////////////////////////////////////////////////////////////////////////////
+
+namespace cutlass::fmha::device {
+
+using namespace cute;
+using namespace cutlass::fmha::kernel;
+
+
+////////////////////////////////////////////////////////////////////////////////
+////////////////////////////// CUTLASS 3.x API /////////////////////////////////
+////////////////////////////////////////////////////////////////////////////////
+
+template<
+    class Kernel_
+>
+class MLA {
+public:
+
+  using Kernel = Kernel_;
+
+  using ReductionKernel = cutlass::fmha::kernel::Sm100FmhaMlaReductionKernel<
+      typename Kernel::ElementOut,
+      typename Kernel::ElementAcc,
+      typename Kernel::ElementAcc,
+      Kernel::TileShapeH::value,
+      Kernel::TileShapeL::value,
+      256 /*Max split*/
+  >;
+
+  /// Argument structure: User API
+  using KernelArguments = typename Kernel::Arguments;
+  using ReductionArguments = typename ReductionKernel::Arguments;
+
+  using Arguments = KernelArguments;
+
+  /// Argument structure: Kernel API
+  using KernelParams = typename Kernel::Params;
+  using ReductionParams = typename ReductionKernel::Params;
+  struct Params {
+    KernelParams fmha_params;
+    ReductionParams reduction_params;
+  };
+
+private:
+
+  /// Kernel API parameters object
+  Params params_;
+
+  bool is_initialized(bool set = false) {
+    static bool initialized = false;
+    if (set) initialized = true;
+    return initialized;
+  }
+
+  static ReductionArguments to_reduction_args(Arguments const& args) {
+    auto [H, K, D, B] = args.problem_shape;
+    return ReductionArguments{
+      nullptr, args.epilogue.ptr_o, nullptr, args.epilogue.ptr_lse,
+      args.mainloop.softmax_scale, B, args.split_kv, K, args.mainloop.ptr_seq,
+      args.ptr_split_kv, Kernel::TileShapeS::value
+    };
+  }
+
+public:
+
+  /// Access the Params structure
+  Params const& params() const {
+    return params_;
+  }
+
+  static void set_split_kv (KernelArguments& args) {
+    // printf("set_split_kv start");
+    if (args.split_kv >= 1) return;
+    auto [H, K, D, B] = args.problem_shape;
+    // std::cout << H << " " << K << " " << D << " " << B << "\n";      
+    int sm_count = args.hw_info.sm_count;
+    // printf("    sm_count = %d\n", sm_count);
+    int max_splits = ceil_div(K, 128);
+    max_splits = min(16, max_splits);
+    // printf("    max_splits = %d\n", max_splits);
+    int sms_per_batch = max(1, sm_count / B);
+    // printf("    sms_per_batch = %d\n", sms_per_batch);
+    int split_heur = min(max_splits, sms_per_batch);
+    int waves = ceil_div(B * split_heur, sm_count);
+    int k_waves = ceil_div(max_splits, split_heur);
+    int split_wave_aware = ceil_div(max_splits, k_waves);
+    args.split_kv = split_wave_aware;
+    // printf("    args.split_kv = %d\n", args.split_kv);
+
+  }
+
+  /// Determines whether the GEMM can execute the given problem.
+  static Status
+  can_implement(Arguments const& args) {
+    if (! Kernel::can_implement(args)) {
+      return Status::kInvalid;
+    }
+    if (! ReductionKernel::can_implement(to_reduction_args(args))) {
+      return Status::kInvalid;
+    }
+    return Status::kSuccess;
+  }
+
+  /// Gets the workspace size
+  static size_t
+  get_workspace_size(Arguments const& args) {
+    size_t workspace_bytes = 0;
+    workspace_bytes += Kernel::get_workspace_size(args);
+    workspace_bytes += ReductionKernel::get_workspace_size(to_reduction_args(args));
+    return workspace_bytes;
+  }
+
+  /// Computes the maximum number of active blocks per multiprocessor
+  static int maximum_active_blocks(int /* smem_capacity */ = -1) {
+    CUTLASS_TRACE_HOST("MLA::maximum_active_blocks()");
+    int max_active_blocks = -1;
+    int smem_size = Kernel::SharedStorageSize;
+
+    // first, account for dynamic smem capacity if needed
+    cudaError_t result;
+    if (smem_size >= (48 << 10)) {
+      CUTLASS_TRACE_HOST("  Setting smem size to " << smem_size);
+      result = cudaFuncSetAttribute(
+          device_kernel<Kernel>,
+          cudaFuncAttributeMaxDynamicSharedMemorySize,
+          smem_size);
+      if (cudaSuccess != result) {
+        result = cudaGetLastError(); // to clear the error bit
+        CUTLASS_TRACE_HOST(
+          "  cudaFuncSetAttribute() returned error: "
+          << cudaGetErrorString(result));
+        return -1;
+      }
+    }
+
+    // query occupancy after setting smem size
+    result = cudaOccupancyMaxActiveBlocksPerMultiprocessor(
+        &max_active_blocks,
+        device_kernel<Kernel>,
+        Kernel::MaxThreadsPerBlock,
+        smem_size);
+
+    if (cudaSuccess != result) {
+      result = cudaGetLastError(); // to clear the error bit
+      CUTLASS_TRACE_HOST(
+        "  cudaOccupancyMaxActiveBlocksPerMultiprocessor() returned error: "
+        << cudaGetErrorString(result));
+      return -1;
+    }
+
+    CUTLASS_TRACE_HOST("  max_active_blocks: " << max_active_blocks);
+    return max_active_blocks;
+  }
+
+  /// Initializes GEMM state from arguments.
+  Status
+  initialize(Arguments const& args, void* workspace = nullptr, cudaStream_t stream = nullptr) {
+    CUTLASS_TRACE_HOST("MLA::initialize() - workspace "
+      << workspace << ", stream: " << (stream ? "non-null" : "null"));
+
+    // Initialize the workspace
+    Status status = Kernel::initialize_workspace(args, workspace, stream);
+    if (status != Status::kSuccess) {
+      return status;
+    }
+    status = ReductionKernel::initialize_workspace(to_reduction_args(args), workspace, stream);
+    if (status != Status::kSuccess) {
+      return status;
+    }
+    KernelParams kernel_params = Kernel::to_underlying_arguments(args, workspace);
+
+    ReductionArguments reduction_args = to_reduction_args(args);
+    if (reduction_args.split_kv > 1) {
+      reduction_args.ptr_oaccum   = kernel_params.epilogue.ptr_o_acc;
+      reduction_args.ptr_lseaccum = kernel_params.epilogue.ptr_lse_acc;
+    }
+    ReductionParams reduction_params = ReductionKernel::to_underlying_arguments(reduction_args, workspace);
+    // Initialize the Params structure
+    params_ = Params {kernel_params, reduction_params};
+
+    if (is_initialized()) return Status::kSuccess;
+
+    // account for dynamic smem capacity if needed
+    // no dynamic smem is needed for reduction kernel
+    int smem_size = Kernel::SharedStorageSize;
+    if (smem_size >= (48 << 10)) {
+      CUTLASS_TRACE_HOST("  Setting smem size to " << smem_size);
+      cudaError_t result = cudaFuncSetAttribute(
+          device_kernel<Kernel>,
+          cudaFuncAttributeMaxDynamicSharedMemorySize,
+          smem_size);
+      if (cudaSuccess != result) {
+        result = cudaGetLastError(); // to clear the error bit
+        CUTLASS_TRACE_HOST("  cudaFuncSetAttribute() returned error: " << cudaGetErrorString(result));
+        return Status::kErrorInternal;
+      }
+    }
+
+    is_initialized(true);
+
+    return Status::kSuccess;
+  }
+
+  /// Update API is preserved in 3.0, but does not guarantee a lightweight update of params.
+  Status
+  update(Arguments const& args, void* workspace = nullptr) {
+    CUTLASS_TRACE_HOST("MLA()::update() - workspace: " << workspace);
+
+    size_t workspace_bytes = get_workspace_size(args);
+    if (workspace_bytes > 0 && nullptr == workspace) {
+      return Status::kErrorWorkspaceNull;
+    }
+
+    auto fmha_params = Kernel::to_underlying_arguments(args, workspace);
+
+    ReductionArguments reduction_args = to_reduction_args(args);
+    if (reduction_args.split_kv > 1) {
+      reduction_args.ptr_oaccum   = fmha_params.epilogue.ptr_o_acc;
+      reduction_args.ptr_lseaccum = fmha_params.epilogue.ptr_lse_acc;
+    }
+    ReductionParams reduction_params = ReductionKernel::to_underlying_arguments(reduction_args, workspace);
+    // Initialize the Params structure
+    params_ = Params {fmha_params, reduction_params};
+
+    return Status::kSuccess;
+  }
+
+  /// Primary run() entry point API that is static allowing users to create and manage their own params.
+  /// Supplied params struct must be construct by calling Kernel::to_underling_arguments()
+  static Status
+  run(Params& params, cudaStream_t stream = nullptr) {
+    CUTLASS_TRACE_HOST("MLA::run()");
+    dim3 const block = Kernel::get_block_shape();
+    dim3 const grid = Kernel::get_grid_shape(params.fmha_params);
+
+    // configure smem size and carveout
+    int smem_size = Kernel::SharedStorageSize;
+
+    Status launch_result;
+    // Use extended launch API only for mainloops that use it
+    if constexpr(Kernel::ArchTag::kMinComputeCapability >= 90) {
+      dim3 cluster(cute::size<0>(typename Kernel::ClusterShape{}),
+                   cute::size<1>(typename Kernel::ClusterShape{}),
+                   cute::size<2>(typename Kernel::ClusterShape{}));
+      void const* kernel = (void const*) device_kernel<Kernel>;
+      void* kernel_params[] = {&params.fmha_params};
+      launch_result = ClusterLauncher::launch(grid, cluster, block, smem_size, stream, kernel, kernel_params);
+    }
+    else {
+      launch_result = Status::kSuccess;
+      device_kernel<Kernel><<<grid, block, smem_size, stream>>>(params.fmha_params);
+    }
+
+    cudaError_t result = cudaGetLastError();
+    if (cudaSuccess != result or Status::kSuccess != launch_result) {
+      //return Status::kSuccess;
+      CUTLASS_TRACE_HOST("  Kernel launch failed. Reason: " << result);
+      return Status::kErrorInternal;
+    }
+    if (params.reduction_params.split_kv > 1) {
+      // launch reduction kernel
+      dim3 const block = ReductionKernel::get_block_shape();
+      dim3 const grid  = ReductionKernel::get_grid_shape(params.reduction_params);
+      device_kernel<ReductionKernel><<<grid, block, 0, stream>>>(params.reduction_params);
+      cudaError_t result = cudaGetLastError();
+      if (cudaSuccess == result) {
+        return Status::kSuccess;
+      }
+      else {
+        CUTLASS_TRACE_HOST("  Kernel launch failed. Reason: " << result);
+        return Status::kErrorInternal;
+      }
+    }
+    else {
+      return Status::kSuccess;
+    }
+  }
+
+  //
+  // Non-static launch overloads that first create and set the internal params struct of this kernel handle.
+  //
+
+  /// Launches the kernel after first constructing Params internal state from supplied arguments.
+  Status
+  run(Arguments const& args, void* workspace = nullptr, cudaStream_t stream = nullptr) {
+    Status status = initialize(args, workspace, stream);
+    if (Status::kSuccess == status) {
+      status = run(params_, stream);
+    }
+    return status;
+  }
+
+  /// Launches the kernel after first constructing Params internal state from supplied arguments.
+  Status
+  operator()(Arguments const& args, void* workspace = nullptr, cudaStream_t stream = nullptr) {
+    return run(args, workspace, stream);
+  }
+
+  /// Overload that allows a user to re-launch the same kernel without updating internal params struct.
+  Status
+  run(cudaStream_t stream = nullptr) {
+    return run(params_, stream);
+  }
+
+  /// Overload that allows a user to re-launch the same kernel without updating internal params struct.
+  Status
+  operator()(cudaStream_t stream = nullptr) {
+    return run(params_, stream);
+  }
+};
+
+////////////////////////////////////////////////////////////////////////////////
+
+} // namespace cutlass::fmha::device
+
+////////////////////////////////////////////////////////////////////////////////

csrc/attention/mla/cutlass_sm100_mla/kernel/sm100_fmha_mla_reduction.hpp

@@ -0,0 +1,203 @@
+/***************************************************************************************************
+ * Copyright (c) 2024 - 2025 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.
+ *
+ **************************************************************************************************/
+/*
+ * Taken from SGLANG PR https://github.com/sgl-project/sglang/pull/6929
+ * by Alcanderian JieXin Liang
+ */
+
+// clang-format off
+#pragma once
+
+#include "cutlass/cutlass.h"
+#include "cutlass/arch/arch.h"
+#include "cute/tensor.hpp"
+
+namespace cutlass::fmha::kernel {
+
+using namespace cute;
+template<
+    class ElementOut,
+    class ElementAcc,
+    class ElementScale,
+    size_t kNumHeads,
+    size_t kHeadDimLatent,
+    int kMaxSplits
+>
+struct Sm100FmhaMlaReductionKernel {
+
+  static const int SharedStorageSize = 0;
+  static const int MaxThreadsPerBlock = 128;
+  static const int MinBlocksPerMultiprocessor = 1;
+
+  using ArchTag = cutlass::arch::Sm100;
+
+  static_assert(kHeadDimLatent % MaxThreadsPerBlock == 0);
+  struct Arguments {
+    ElementAcc* ptr_oaccum = nullptr;
+    ElementOut* ptr_o = nullptr;
+    ElementAcc* ptr_lseaccum = nullptr;
+    ElementAcc* ptr_lse = nullptr;
+    ElementScale scale = 1.f;
+    int num_batches = 0;
+    int split_kv = -1;
+    int dim_k = -1;
+    int* ptr_seq = nullptr;
+    int* ptr_split_kv = nullptr;
+    int tile_shape_s = 128;
+  };
+  using Params = Arguments;
+
+  static Params to_underlying_arguments(Arguments const& args, void* workspace) {
+    return {args.ptr_oaccum, args.ptr_o, args.ptr_lseaccum, args.ptr_lse,
+	    args.scale, args.num_batches, args.split_kv, args.dim_k, args.ptr_seq,
+	    args.ptr_split_kv, args.tile_shape_s};
+  }
+
+  static size_t get_workspace_size(Arguments const& /*args*/) {
+    return 0;
+  }
+
+  static Status initialize_workspace(
+      Arguments const& /*args*/, void* /*ws*/, cudaStream_t /*stream*/) {
+    return Status::kSuccess;
+  }
+
+  static dim3 get_grid_shape(Params const& params) {
+    return dim3(kNumHeads, 1, params.num_batches);
+  }
+
+  static dim3 get_block_shape() {
+    return dim3(MaxThreadsPerBlock, 1, 1);
+  }
+
+  static bool can_implement(Arguments const& args) {
+    if (args.num_batches <= 0) return false;
+    if (args.split_kv <= 0) return false;
+    return true;
+  }
+
+  CUTLASS_DEVICE void operator() (Params const& params, char* smem_raw) {
+    if (params.split_kv <= 1) return;
+    auto blk_coord = make_coord(blockIdx.x, _0{}, blockIdx.z);
+
+    __shared__ ElementAcc sLseScale[kMaxSplits];
+    const size_t offset_lseaccum = get<0>(blk_coord) + kNumHeads * params.split_kv * get<2>(blk_coord);
+    const size_t offset_lse = get<0>(blk_coord) + kNumHeads * get<2>(blk_coord);
+
+    Tensor gLSEaccum = make_tensor(make_gmem_ptr(params.ptr_lseaccum + offset_lseaccum),
+                                   make_shape(params.split_kv), Stride<Int<kNumHeads>>{});
+
+    Tensor gLSE = make_tensor(make_gmem_ptr(params.ptr_lse + offset_lse),
+                              Shape<_1>{}, Stride<_1>{});
+
+    auto dim_k = params.ptr_seq == nullptr ?  params.dim_k : params.ptr_seq[get<2>(blk_coord)];
+    auto local_split_kv = params.ptr_split_kv == nullptr ? params.split_kv : params.ptr_split_kv[get<2>(blk_coord)];
+    auto k_tile_total = ceil_div(dim_k, params.tile_shape_s);
+    auto k_tile_per_cta = ceil_div(k_tile_total, local_split_kv);
+    local_split_kv = ceil_div(k_tile_total, k_tile_per_cta);
+
+    int warp_idx = cutlass::canonical_warp_idx_sync();
+    if (warp_idx == 0) {
+      constexpr int kNLsePerThread = cute::ceil_div(kMaxSplits, 32);
+
+      ElementAcc local_lse[kNLsePerThread];
+
+      CUTLASS_PRAGMA_UNROLL
+      for (int i = 0; i < kNLsePerThread; ++i) {
+        const int split = i * 32 + threadIdx.x;
+        local_lse[i] = split < local_split_kv ? gLSEaccum(split) : -std::numeric_limits<ElementAcc>::infinity();
+      }
+
+      ElementAcc lse_max = -std::numeric_limits<ElementAcc>::infinity();
+      CUTLASS_PRAGMA_UNROLL
+      for (int i = 0; i < kNLsePerThread; ++i) {
+        lse_max = max(lse_max, local_lse[i]);
+      }
+      CUTLASS_PRAGMA_UNROLL
+      for (int offset = 16; offset >= 1; offset /= 2) {
+        lse_max = max(lse_max, __shfl_xor_sync(0xffffffff, lse_max, offset));
+      }
+      lse_max = lse_max == -std::numeric_limits<ElementAcc>::infinity() ? 0.0f : lse_max;  // In case all local LSEs are -inf
+      lse_max = __shfl_sync(0xffffffff, lse_max, 0);
+
+      ElementAcc sum_lse = 0;
+      CUTLASS_PRAGMA_UNROLL
+      for (int i = 0; i < kNLsePerThread; ++i) {
+        sum_lse = sum_lse + expf(local_lse[i] - lse_max);
+      }
+
+      CUTLASS_PRAGMA_UNROLL
+      for (int offset = 16; offset >= 1; offset /= 2) {
+        sum_lse = sum_lse + __shfl_xor_sync(0xffffffff, sum_lse, offset);
+      }
+
+      sum_lse = __shfl_sync(0xffffffff, sum_lse, 0);
+
+      ElementAcc global_lse = (sum_lse == 0.f || sum_lse != sum_lse) ? std::numeric_limits<ElementAcc>::infinity() : logf(sum_lse) + lse_max;
+      if (threadIdx.x == 0 and params.ptr_lse != nullptr) {
+        gLSE(0) = global_lse;
+      }
+
+      CUTLASS_PRAGMA_UNROLL
+      for (int i = 0; i < kNLsePerThread; ++i) {
+        const int split = i * 32 + threadIdx.x;
+        if (split < local_split_kv) {
+          sLseScale[split] = expf(local_lse[i] - global_lse);
+        }
+      }
+    }
+    __syncthreads();
+
+    constexpr int Elements = kHeadDimLatent / MaxThreadsPerBlock;
+    const size_t offset_oaccum = kHeadDimLatent * params.split_kv * (get<0>(blk_coord) + kNumHeads * get<2>(blk_coord));
+    Tensor gOaccum = make_tensor(make_gmem_ptr(params.ptr_oaccum + offset_oaccum),
+                               Shape<Int<kHeadDimLatent>>{}, Stride<_1>{});
+    ElementAcc local_val[Elements] = {0};
+    for (int split = 0; split < local_split_kv; ++split) {
+      ElementAcc lse_scale = sLseScale[split];
+      CUTLASS_PRAGMA_UNROLL
+      for(int i = 0; i < Elements; ++i) {
+        local_val[i] += lse_scale * gOaccum(threadIdx.x + MaxThreadsPerBlock * i);
+      }
+      gOaccum.data() = gOaccum.data() + kHeadDimLatent;
+    }
+    auto ptr_o_local = params.ptr_o + (get<0>(blk_coord) + get<2>(blk_coord) * kNumHeads) * kHeadDimLatent;
+    Tensor gO = make_tensor(make_gmem_ptr(ptr_o_local), Shape<Int<kHeadDimLatent>>{}, Stride<_1>{});
+
+    CUTLASS_PRAGMA_UNROLL
+    for(int i = 0; i < Elements; ++i) {
+      gO(threadIdx.x + MaxThreadsPerBlock * i) = static_cast<ElementOut>(local_val[i]);
+    }
+  }
+};
+
+}  // namespace cutlass::fmha::kernel

csrc/attention/mla/cutlass_sm100_mla/kernel/sm100_mla_tile_scheduler.hpp

@@ -0,0 +1,165 @@
+/***************************************************************************************************
+ * Copyright (c) 2024 - 2025 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.
+ *
+ **************************************************************************************************/
+/*
+ * Taken from SGLANG PR https://github.com/sgl-project/sglang/pull/6929
+ * by Alcanderian JieXin Liang
+ */
+
+// clang-format off
+#pragma once
+
+#include "cutlass/cutlass.h"
+#include "cutlass/fast_math.h"
+#include "cutlass/kernel_hardware_info.h"
+
+namespace cutlass::fmha::kernel {
+
+////////////////////////////////////////////////////////////////////////////////
+
+struct Sm100MlaIndividualTileScheduler {
+
+  struct Params {
+    dim3 grid;
+  };
+
+  bool valid_ = true;
+
+  CUTLASS_DEVICE
+  Sm100MlaIndividualTileScheduler(Params const&) {}
+
+  template<class ProblemShape, class ClusterShape>
+  static Params to_underlying_arguments(
+      ProblemShape const& problem_shape, KernelHardwareInfo hw_info,
+      ClusterShape const& cluster_shape, int const& split_kv) {
+    using namespace cute;
+    dim3 grid(get<0>(cluster_shape), get<3>(problem_shape) /* Batch */, split_kv /*Maximum Split KV*/);
+    return Params{ grid };
+  }
+
+  static dim3 get_grid_shape(Params const& params) {
+    return params.grid;
+  }
+
+  CUTLASS_DEVICE
+  bool is_valid() {
+    return valid_;
+  }
+
+  CUTLASS_DEVICE
+  auto get_block_coord() {
+    using namespace cute;
+    return make_coord(blockIdx.x, _0{}, blockIdx.y, blockIdx.z);
+  }
+
+  CUTLASS_DEVICE
+  Sm100MlaIndividualTileScheduler& operator++() {
+    valid_ = false;
+    return *this;
+  }
+};
+
+////////////////////////////////////////////////////////////////////////////////
+
+struct Sm100MlaPersistentTileScheduler {
+
+  struct Params {
+    int num_blocks;
+    FastDivmod divmod_m_block;
+    FastDivmod divmod_b;
+    FastDivmod divmod_split_kv;
+    KernelHardwareInfo hw_info;
+  };
+
+  int block_idx = 0;
+  Params params;
+
+  CUTLASS_DEVICE
+  Sm100MlaPersistentTileScheduler(Params const& params) : block_idx(blockIdx.x), params(params) {}
+
+  template<class ProblemShape, class ClusterShape>
+  static Params to_underlying_arguments(
+      ProblemShape const& problem_shape, KernelHardwareInfo hw_info,
+      ClusterShape const& cluster_shape, int const& split_kv) {
+    using namespace cute;
+    // Get SM count if needed, otherwise use user supplied SM count
+    int sm_count = hw_info.sm_count;
+    if (sm_count <= 1 || sm_count % size<0>(cluster_shape) != 0) {
+      CUTLASS_TRACE_HOST("  WARNING: Arguments do not include a valid SM count.\n"
+          "  For optimal performance, populate the arguments KernelHardwareInfo struct with the SM count.");
+      sm_count = KernelHardwareInfo::query_device_multiprocessor_count(hw_info.device_id);
+    }
+
+    CUTLASS_TRACE_HOST("to_underlying_arguments(): Setting persistent grid SM count to " << sm_count);
+    hw_info.sm_count = sm_count;
+
+    int num_m_blocks = size<0>(cluster_shape);
+    int num_blocks = num_m_blocks * get<3>(problem_shape)  /* Batch */;
+    num_blocks *= split_kv; /* Maximum Split KV*/
+
+    return Params {
+      num_blocks,
+      { num_m_blocks}, { get<3>(problem_shape) }, {split_kv},
+      hw_info
+    };
+  }
+
+  static dim3 get_grid_shape(Params const& params) {
+    dim3 grid(std::min(params.num_blocks, params.hw_info.sm_count), 1, 1);
+    return grid;
+  }
+
+  CUTLASS_DEVICE
+  bool is_valid() {
+    return block_idx < params.num_blocks;
+  }
+
+  CUTLASS_DEVICE
+  auto get_block_coord() {
+    using namespace cute;
+    int block_decode = block_idx;
+    int m_block, bidb, n_split_kv;
+    params.divmod_m_block(block_decode, m_block, block_decode);
+    params.divmod_b(block_decode, bidb, block_decode);
+    params.divmod_split_kv(block_decode, n_split_kv, block_decode);
+    return make_coord(m_block, _0{}, bidb, n_split_kv);
+  }
+
+  CUTLASS_DEVICE
+  Sm100MlaPersistentTileScheduler& operator++() {
+    block_idx += gridDim.x;
+    return *this;
+  }
+};
+
+////////////////////////////////////////////////////////////////////////////////
+
+} // namespace cutlass::fmha::kernel

csrc/attention/mla/sm100_cutlass_mla_kernel.cu

@@ -0,0 +1,283 @@
+/*
+Copyright (c) 2025, NVIDIA CORPORATION.  All rights reserved.
+Copyright 2025 SGLang Team. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+/*
+ * Taken from SGLANG PR https://github.com/sgl-project/sglang/pull/6929
+ * by Alcanderian JieXin Liang
+ */
+#include "core/registration.h"
+
+#include <ATen/cuda/CUDAContext.h>
+#include <c10/cuda/CUDAGuard.h>
+#include <cutlass/cutlass.h>
+#include <cutlass/kernel_hardware_info.h>
+#include <torch/all.h>
+
+#include <cute/tensor.hpp>
+#include <iostream>
+
+#include "cutlass_sm100_mla/device/sm100_mla.hpp"
+#include "cutlass_sm100_mla/kernel/sm100_mla_tile_scheduler.hpp"
+
+// clang-format off
+#if !defined(CUDA_VERSION) || CUDA_VERSION < 12040
+void sm100_cutlass_mla_decode(
+    torch::Tensor const& out,
+    torch::Tensor const& q_nope,
+    torch::Tensor const& q_pe,
+    torch::Tensor const& kv_c_and_k_pe_cache,
+    torch::Tensor const& seq_lens,
+    torch::Tensor const& page_table,
+    torch::Tensor const& workspace,
+    int64_t num_kv_splits) {
+  TORCH_CHECK(false, "CUDA version must be >= 12.4 for 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) {
+  TORCH_CHECK(false, "CUDA version must be >= 12.4 for cutlass_mla_get_workspace_size");
+}
+#else
+
+#define CUTLASS_CHECK(status)                                                       \
+  {                                                                                 \
+    cutlass::Status error = status;                                                 \
+    TORCH_CHECK(error == cutlass::Status::kSuccess, cutlassGetStatusString(error)); \
+  }
+
+using namespace cute;
+using namespace cutlass::fmha::kernel;
+
+template <bool v>
+struct IsPersistent {
+  static const bool value = v;
+};
+
+template <typename T, bool IsPaged128, typename PersistenceOption = IsPersistent<true>>
+struct MlaSm100 {
+  using Element = T;
+  using ElementAcc = float;
+  using ElementOut = T;
+
+  using TileShape = Shape<_128, _128, Shape<_512, _64>>;
+  using TileShapeH = cute::tuple_element_t<0, TileShape>;
+  using TileShapeD = cute::tuple_element_t<2, TileShape>;
+
+  // H K (D_latent D_rope) B
+  using ProblemShape = cute::tuple<TileShapeH, int, TileShapeD, int>;
+
+  using StrideQ = cute::tuple<int64_t, _1, int64_t>;  // H D B
+  using StrideK = cute::tuple<int64_t, _1, int64_t>;  // K D B
+  using StrideO = StrideK;                            // H D B
+  using StrideLSE = cute::tuple<_1, int>;             // H B
+
+  using TileScheduler =
+      std::conditional_t<PersistenceOption::value, Sm100MlaPersistentTileScheduler, Sm100MlaIndividualTileScheduler>;
+
+  using FmhaKernel = cutlass::fmha::kernel::Sm100FmhaMlaKernelTmaWarpspecialized<
+      TileShape,
+      Element,
+      ElementAcc,
+      ElementOut,
+      ElementAcc,
+      TileScheduler,
+      /*kIsCpAsync=*/!IsPaged128>;
+  using Fmha = cutlass::fmha::device::MLA<FmhaKernel>;
+};
+
+template <typename T>
+typename T::Fmha::Arguments args_from_options(
+    at::Tensor const& out,
+    at::Tensor const& q_nope,
+    at::Tensor const& q_pe,
+    at::Tensor const& kv_c_and_k_pe_cache,
+    at::Tensor const& seq_lens,
+    at::Tensor const& page_table,
+    double sm_scale,
+    int64_t num_kv_splits) {
+  cutlass::KernelHardwareInfo hw_info;
+  hw_info.device_id = q_nope.device().index();
+  hw_info.sm_count = cutlass::KernelHardwareInfo::query_device_multiprocessor_count(hw_info.device_id);
+
+  int batches = q_nope.sizes()[0];
+  int page_count_per_seq = page_table.sizes()[1];
+  int page_count_total = kv_c_and_k_pe_cache.sizes()[0];
+  int page_size = kv_c_and_k_pe_cache.sizes()[1];
+  int max_seq_len = page_size * page_count_per_seq;
+  using TileShapeH = typename T::TileShapeH;
+  using TileShapeD = typename T::TileShapeD;
+  auto problem_shape = cute::make_tuple(TileShapeH{}, max_seq_len, TileShapeD{}, batches);
+
+  auto [H, K, D, B] = problem_shape;
+  auto [D_latent, D_rope] = D;
+
+  float scale = float(sm_scale);
+
+  using StrideQ = typename T::StrideQ;
+  using StrideK = typename T::StrideK;
+  using StrideO = typename T::StrideO;
+  using StrideLSE = typename T::StrideLSE;
+
+  StrideQ stride_Q_nope = cute::make_tuple(
+      static_cast<int64_t>(q_nope.stride(1)), _1{}, static_cast<int64_t>(q_nope.stride(0)));
+  StrideQ stride_Q_pe = cute::make_tuple(
+      static_cast<int64_t>(q_pe.stride(1)), _1{}, static_cast<int64_t>(q_pe.stride(0)));
+
+  StrideK stride_C = cute::make_tuple(
+      static_cast<int64_t>(0 + D_latent + D_rope), _1{}, static_cast<int64_t>(page_size * (D_latent + D_rope)));
+  StrideLSE stride_PT = cute::make_stride(_1{}, page_count_per_seq);
+  StrideLSE stride_LSE = cute::make_tuple(_1{}, 0 + H);
+  StrideO stride_O = cute::make_tuple(static_cast<int64_t>(0 + D_latent), _1{}, static_cast<int64_t>(0 + H * D_latent));
+
+  using Element = typename T::Element;
+  using ElementOut = typename T::ElementOut;
+  using ElementAcc = typename T::ElementAcc;
+  auto Q_nope_ptr = static_cast<Element*>(q_nope.data_ptr());
+  auto Q_pe_ptr = static_cast<Element*>(q_pe.data_ptr());
+  auto C_ptr = static_cast<Element*>(kv_c_and_k_pe_cache.data_ptr());
+  typename T::Fmha::Arguments arguments{
+      problem_shape,
+      {scale,
+       Q_nope_ptr,
+       stride_Q_nope,
+       Q_pe_ptr,
+       stride_Q_pe,
+       C_ptr,
+       stride_C,
+       C_ptr + D_latent,
+       stride_C,
+       static_cast<int*>(seq_lens.data_ptr()),
+       static_cast<int*>(page_table.data_ptr()),
+       stride_PT,
+       page_count_total,
+       page_size},
+      {static_cast<ElementOut*>(out.data_ptr()), stride_O, static_cast<ElementAcc*>(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
+      nullptr,       // is_var_split_kv
+  };
+  // TODO(kaixih@nvidia): When split_kv=-1 and is_var_split_kv=false, we compute
+  // split_kv automatically based on batch size and sequence length to balance
+  // workload across available SMs. Consider using var_split_kv for manual
+  // control if needed.
+  T::Fmha::set_split_kv(arguments);
+  return arguments;
+}
+
+template <typename Element, bool IsPaged128, typename PersistenceOption>
+void runMla(
+    at::Tensor const& out,
+    at::Tensor const& q_nope,
+    at::Tensor const& q_pe,
+    at::Tensor const& kv_c_and_k_pe_cache,
+    at::Tensor const& seq_lens,
+    at::Tensor const& page_table,
+    at::Tensor const& workspace,
+    double sm_scale,
+    int64_t num_kv_splits,
+    cudaStream_t stream) {
+  using MlaSm100Type = MlaSm100<Element, 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);
+
+  CUTLASS_CHECK(fmha.can_implement(arguments));
+
+  CUTLASS_CHECK(fmha.initialize(arguments, workspace.data_ptr(), stream));
+
+  CUTLASS_CHECK(fmha.run(arguments, workspace.data_ptr(), stream));
+}
+
+#define DISPATCH_BOOL(expr, const_expr, ...) \
+  [&]() -> bool {                            \
+    if (expr) {                              \
+      constexpr bool const_expr = true;      \
+      return __VA_ARGS__();                  \
+    } else {                                 \
+      constexpr bool const_expr = false;     \
+      return __VA_ARGS__();                  \
+    }                                        \
+  }()
+
+void sm100_cutlass_mla_decode(
+    torch::Tensor const& out,
+    torch::Tensor const& q_nope,
+    torch::Tensor const& q_pe,
+    torch::Tensor const& kv_c_and_k_pe_cache,
+    torch::Tensor const& seq_lens,
+    torch::Tensor const& page_table,
+    torch::Tensor const& workspace,
+    double sm_scale,
+    int64_t num_kv_splits) {
+  auto in_dtype = q_nope.dtype();
+  at::cuda::CUDAGuard device_guard{(char)q_nope.get_device()};
+  const cudaStream_t stream = at::cuda::getCurrentCUDAStream(q_nope.get_device());
+  const int page_size = kv_c_and_k_pe_cache.sizes()[1];
+  
+  // NOTE(alcanderian): IsPersistent has bug with manual split_kv.
+  // Kernel will hang if batch is too large with large num_kv_splits. (for example bs=8, num_kv_splits=8)
+  // Maybe per batch split kv will fix this.
+  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);
+      } 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);
+      } 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);
+      } else {
+        TORCH_CHECK(false, "Unsupported input data type of MLA");
+      }
+      return true;
+    });
+    return true;
+  });
+}
+
+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>;
+
+  // Get split kv. Requires problem shape and sm_count only.
+  typename MlaSm100Type::Fmha::Arguments arguments;
+  using TileShapeH = typename MlaSm100Type::TileShapeH;
+  using TileShapeD = typename MlaSm100Type::TileShapeD;
+  arguments.problem_shape =
+      cute::make_tuple(TileShapeH{}, static_cast<int>(max_seq_len), TileShapeD{}, static_cast<int>(num_batches));
+  // 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;
+  MlaSm100Type::Fmha::set_split_kv(arguments);
+
+  return MlaSm100Type::Fmha::get_workspace_size(arguments);
+}
+
+#endif
+
+TORCH_LIBRARY_IMPL_EXPAND(TORCH_EXTENSION_NAME, CUDA, m) {
+  m.impl("sm100_cutlass_mla_decode", &sm100_cutlass_mla_decode);
+}
+
+TORCH_LIBRARY_IMPL_EXPAND(TORCH_EXTENSION_NAME, CatchAll, m) {
+  m.impl("sm100_cutlass_mla_get_workspace_size", &sm100_cutlass_mla_get_workspace_size);
+}
+
+// clang-format on

csrc/attention/paged_attention_v1.cu

@@ -16,14 +16,8 @@
  * See the License for the specific language governing permissions and
  * limitations under the License.
  */
-
 #include "attention_kernels.cuh"
-
-#ifndef USE_ROCM
-  #define WARP_SIZE 32
-#else
-  #define WARP_SIZE warpSize
-#endif
+#include "../cuda_compat.h"
 
 #define MAX(a, b) ((a) > (b) ? (a) : (b))
 #define MIN(a, b) ((a) < (b) ? (a) : (b))
@@ -80,7 +74,7 @@ void paged_attention_v1_launcher(
   const float* k_scale_ptr = reinterpret_cast<const float*>(k_scale.data_ptr());
   const float* v_scale_ptr = reinterpret_cast<const float*>(v_scale.data_ptr());
 
-  constexpr int NUM_WARPS = NUM_THREADS / WARP_SIZE;
+  const int NUM_WARPS = NUM_THREADS / WARP_SIZE;
   int padded_max_seq_len =
       DIVIDE_ROUND_UP(max_seq_len, BLOCK_SIZE) * BLOCK_SIZE;
   int logits_size = padded_max_seq_len * sizeof(float);
@@ -187,7 +181,6 @@ void paged_attention_v1(
                              CALL_V1_LAUNCHER_BLOCK_SIZE)
 }
 
-#undef WARP_SIZE
 #undef MAX
 #undef MIN
 #undef DIVIDE_ROUND_UP

csrc/attention/paged_attention_v2.cu

@@ -16,14 +16,8 @@
  * See the License for the specific language governing permissions and
  * limitations under the License.
  */
-
 #include "attention_kernels.cuh"
-
-#ifndef USE_ROCM
-  #define WARP_SIZE 32
-#else
-  #define WARP_SIZE warpSize
-#endif
+#include "../cuda_compat.h"
 
 #define MAX(a, b) ((a) > (b) ? (a) : (b))
 #define MIN(a, b) ((a) < (b) ? (a) : (b))
@@ -84,7 +78,7 @@ void paged_attention_v2_launcher(
   const float* k_scale_ptr = reinterpret_cast<const float*>(k_scale.data_ptr());
   const float* v_scale_ptr = reinterpret_cast<const float*>(v_scale.data_ptr());
 
-  constexpr int NUM_WARPS = NUM_THREADS / WARP_SIZE;
+  const int NUM_WARPS = NUM_THREADS / WARP_SIZE;
   int max_num_partitions = DIVIDE_ROUND_UP(max_seq_len, PARTITION_SIZE);
   int logits_size = PARTITION_SIZE * sizeof(float);
   int outputs_size = (NUM_WARPS / 2) * head_size * sizeof(float);
@@ -197,7 +191,6 @@ void paged_attention_v2(
                              CALL_V2_LAUNCHER_BLOCK_SIZE)
 }
 
-#undef WARP_SIZE
 #undef MAX
 #undef MIN
 #undef DIVIDE_ROUND_UP

csrc/cpu/cpu_types_arm.hpp

@@ -33,6 +33,8 @@ namespace vec_op {
 #endif
 
 #define FORCE_INLINE __attribute__((always_inline)) inline
+// Number of elements in single ASIMD vector of given Datatype
+#define NUM_ELEMENTS_REG(vec) (sizeof(vec) / sizeof(vec[0]))
 
 namespace {
 template <typename T, T... indexes, typename F>
@@ -86,8 +88,8 @@ struct FP16Vec16 : public Vec<FP16Vec16> {
   }
 
   void save(void* ptr, const int elem_num) const {
-    int full_blocks = elem_num / 8;
-    int remainder = elem_num % 8;
+    int full_blocks = elem_num / NUM_ELEMENTS_REG(reg.val[0]);
+    int remainder = elem_num % NUM_ELEMENTS_REG(reg.val[0]);
 
     if (full_blocks > 0) {
       vst1q_f16(reinterpret_cast<__fp16*>(ptr), reg.val[0]);
@@ -197,6 +199,25 @@ struct BF16Vec16 : public Vec<BF16Vec16> {
              vcvtq_high_bf16_f32(vcvtq_low_bf16_f32(v.val[2]), v.val[3])}) {};
 
   void save(void* ptr) const { *reinterpret_cast<bfloat16x8x2_t*>(ptr) = reg; };
+  void save(void* ptr, const int elem_num) const {
+    int full_blocks = elem_num / NUM_ELEMENTS_REG(reg.val[0]);
+    int remainder = elem_num % NUM_ELEMENTS_REG(reg.val[0]);
+    for (int i = 0; i < full_blocks; i++)
+      vst1q_bf16(
+          reinterpret_cast<__bf16*>(ptr) + NUM_ELEMENTS_REG(reg.val[0]) * i,
+          reg.val[i]);
+    if (remainder > 0) {
+      bfloat16x8_t temp = reg.val[full_blocks];
+      bfloat16_t* base = reinterpret_cast<bfloat16_t*>(ptr) + full_blocks * 8;
+      if (remainder > 0) base[0] = vgetq_lane_bf16(temp, 0);
+      if (remainder > 1) base[1] = vgetq_lane_bf16(temp, 1);
+      if (remainder > 2) base[2] = vgetq_lane_bf16(temp, 2);
+      if (remainder > 3) base[3] = vgetq_lane_bf16(temp, 3);
+      if (remainder > 4) base[4] = vgetq_lane_bf16(temp, 4);
+      if (remainder > 5) base[5] = vgetq_lane_bf16(temp, 5);
+      if (remainder > 6) base[6] = vgetq_lane_bf16(temp, 6);
+    }
+  };
 };
 
 struct BF16Vec32 : public Vec<BF16Vec32> {
@@ -213,6 +234,25 @@ struct BF16Vec32 : public Vec<BF16Vec32> {
       : reg({vec8_data.reg, vec8_data.reg, vec8_data.reg, vec8_data.reg}) {};
 
   void save(void* ptr) const { *reinterpret_cast<bfloat16x8x4_t*>(ptr) = reg; };
+  void save(void* ptr, const int elem_num) const {
+    int full_blocks = elem_num / NUM_ELEMENTS_REG(reg.val[0]);
+    int remainder = elem_num % NUM_ELEMENTS_REG(reg.val[0]);
+    for (int i = 0; i < full_blocks; i++)
+      vst1q_bf16(
+          reinterpret_cast<__bf16*>(ptr) + NUM_ELEMENTS_REG(reg.val[0]) * i,
+          reg.val[i]);
+    if (remainder > 0) {
+      bfloat16x8_t temp = reg.val[full_blocks];
+      bfloat16_t* base = reinterpret_cast<bfloat16_t*>(ptr) + full_blocks * 8;
+      base[0] = vgetq_lane_bf16(temp, 0);
+      if (remainder > 1) base[1] = vgetq_lane_bf16(temp, 1);
+      if (remainder > 2) base[2] = vgetq_lane_bf16(temp, 2);
+      if (remainder > 3) base[3] = vgetq_lane_bf16(temp, 3);
+      if (remainder > 4) base[4] = vgetq_lane_bf16(temp, 4);
+      if (remainder > 5) base[5] = vgetq_lane_bf16(temp, 5);
+      if (remainder > 6) base[6] = vgetq_lane_bf16(temp, 6);
+    }
+  };
 };
 #endif
 
@@ -372,6 +412,48 @@ struct FP32Vec8 : public Vec<FP32Vec8> {
   }
 };
 
+struct INT32Vec16 : public Vec<INT32Vec16> {
+  constexpr static int VEC_ELEM_NUM = 16;
+  union AliasReg {
+    int32x4x4_t reg;
+    int32_t values[VEC_ELEM_NUM];
+  };
+  int32x4x4_t reg;
+
+  explicit INT32Vec16(const void* ptr) {
+    reg.val[0] = vld1q_s32(reinterpret_cast<const int32_t*>(ptr));
+    reg.val[1] = vld1q_s32(reinterpret_cast<const int32_t*>(ptr) + 4);
+    reg.val[2] = vld1q_s32(reinterpret_cast<const int32_t*>(ptr) + 8);
+    reg.val[3] = vld1q_s32(reinterpret_cast<const int32_t*>(ptr) + 12);
+  }
+
+  void save(int32_t* ptr) const {
+    vst1q_s32(ptr, reg.val[0]);
+    vst1q_s32(ptr + 4, reg.val[1]);
+    vst1q_s32(ptr + 8, reg.val[2]);
+    vst1q_s32(ptr + 12, reg.val[3]);
+  };
+
+  void save(int32_t* ptr, const int elem_num) const {
+    int full_blocks = elem_num / NUM_ELEMENTS_REG(reg.val[0]);
+    int remainder = elem_num % NUM_ELEMENTS_REG(reg.val[0]);
+
+    for (int i = 0; i < full_blocks; i++)
+      vst1q_s32(
+          reinterpret_cast<__int32_t*>(ptr) + NUM_ELEMENTS_REG(reg.val[0]) * i,
+          reg.val[i]);
+
+    if (remainder > 0) {
+      int32x4_t temp = reg.val[full_blocks];
+      int32_t* base = reinterpret_cast<int32_t*>(ptr) + full_blocks * 4;
+      if (remainder > 0) base[0] = vgetq_lane_s32(temp, 0);
+      if (remainder > 1) base[1] = vgetq_lane_s32(temp, 1);
+      if (remainder > 2) base[2] = vgetq_lane_s32(temp, 2);
+      if (remainder > 3) base[3] = vgetq_lane_s32(temp, 3);
+    }
+  }
+};
+
 struct FP32Vec16 : public Vec<FP32Vec16> {
   constexpr static int VEC_ELEM_NUM = 16;
   union AliasReg {
@@ -434,7 +516,12 @@ struct FP32Vec16 : public Vec<FP32Vec16> {
     reg.val[2] = vcvt_f32_f16(vget_low_f16(v.reg.val[1]));
     reg.val[3] = vcvt_f32_f16(vget_high_f16(v.reg.val[1]));
   };
-
+  explicit FP32Vec16(const INT32Vec16& v) {
+    reg.val[0] = vcvtq_f32_s32(v.reg.val[0]);
+    reg.val[1] = vcvtq_f32_s32(v.reg.val[1]);
+    reg.val[2] = vcvtq_f32_s32(v.reg.val[2]);
+    reg.val[3] = vcvtq_f32_s32(v.reg.val[3]);
+  };
   FP32Vec16 operator+(const FP32Vec16& b) const {
     return FP32Vec16(float32x4x4_t({vaddq_f32(reg.val[0], b.reg.val[0]),
                                     vaddq_f32(reg.val[1], b.reg.val[1]),
@@ -463,6 +550,85 @@ struct FP32Vec16 : public Vec<FP32Vec16> {
                                     vdivq_f32(reg.val[3], b.reg.val[3])}));
   };
 
+  FP32Vec16 clamp(const FP32Vec16& min, const FP32Vec16& max) const {
+    return FP32Vec16(float32x4x4_t(
+        {vminq_f32(max.reg.val[0], vmaxq_f32(min.reg.val[0], reg.val[0])),
+         vminq_f32(max.reg.val[1], vmaxq_f32(min.reg.val[1], reg.val[1])),
+         vminq_f32(max.reg.val[2], vmaxq_f32(min.reg.val[2], reg.val[2])),
+         vminq_f32(max.reg.val[3], vmaxq_f32(min.reg.val[3], reg.val[3]))}));
+  };
+
+  FP32Vec16 max(const FP32Vec16& b) const {
+    return FP32Vec16(float32x4x4_t({vmaxq_f32(b.reg.val[0], reg.val[0]),
+                                    vmaxq_f32(b.reg.val[1], reg.val[1]),
+                                    vmaxq_f32(b.reg.val[2], reg.val[2]),
+                                    vmaxq_f32(b.reg.val[3], reg.val[3])}));
+  };
+
+  FP32Vec16 max(const FP32Vec16& b, const int elem_num) const {
+    int full_blocks = elem_num / NUM_ELEMENTS_REG(reg.val[0]);
+    int remainder = elem_num % NUM_ELEMENTS_REG(reg.val[0]);
+    float32x4x4_t temp;
+
+    for (int i = 0; i < full_blocks; i++)
+      temp.val[i] = vmaxq_f32(b.reg.val[i], reg.val[i]);
+
+    if (remainder > 0) {
+      float max_v = std::max(vgetq_lane_f32(reg.val[full_blocks], 0),
+                             vgetq_lane_f32(b.reg.val[full_blocks], 0));
+      temp.val[full_blocks] = vsetq_lane_f32(max_v, temp.val[full_blocks], 0);
+    }
+    if (remainder > 1) {
+      float max_v = std::max(vgetq_lane_f32(reg.val[full_blocks], 1),
+                             vgetq_lane_f32(b.reg.val[full_blocks], 1));
+      temp.val[full_blocks] = vsetq_lane_f32(max_v, temp.val[full_blocks], 1);
+    }
+    if (remainder > 2) {
+      float max_v = std::max(vgetq_lane_f32(reg.val[full_blocks], 2),
+                             vgetq_lane_f32(b.reg.val[full_blocks], 2));
+      temp.val[full_blocks] = vsetq_lane_f32(max_v, temp.val[full_blocks], 2);
+    }
+    return FP32Vec16(temp);
+  };
+
+  FP32Vec16 min(const FP32Vec16& b) const {
+    return FP32Vec16(float32x4x4_t({
+        vminq_f32(b.reg.val[0], reg.val[0]),
+        vminq_f32(b.reg.val[1], reg.val[1]),
+        vminq_f32(b.reg.val[2], reg.val[2]),
+        vminq_f32(b.reg.val[3], reg.val[3]),
+    }));
+  };
+  FP32Vec16 min(const FP32Vec16& b, const int elem_num) const {
+    int full_blocks = elem_num / NUM_ELEMENTS_REG(reg.val[0]);
+    const int remainder = elem_num % NUM_ELEMENTS_REG(reg.val[0]);
+    float32x4x4_t temp;
+    for (int i = 0; i < full_blocks; i++)
+      temp.val[i] = vminq_f32(b.reg.val[i], reg.val[i]);
+
+    if (remainder > 0) {
+      float min_v = std::min(vgetq_lane_f32(reg.val[full_blocks], 0),
+                             vgetq_lane_f32(b.reg.val[full_blocks], 0));
+      temp.val[full_blocks] = vsetq_lane_f32(min_v, temp.val[full_blocks], 0);
+    }
+    if (remainder > 1) {
+      float min_v = std::min(vgetq_lane_f32(reg.val[full_blocks], 1),
+                             vgetq_lane_f32(b.reg.val[full_blocks], 1));
+      temp.val[full_blocks] = vsetq_lane_f32(min_v, temp.val[full_blocks], 1);
+    }
+    if (remainder > 2) {
+      float min_v = std::min(vgetq_lane_f32(reg.val[full_blocks], 2),
+                             vgetq_lane_f32(b.reg.val[full_blocks], 2));
+      temp.val[full_blocks] = vsetq_lane_f32(min_v, temp.val[full_blocks], 2);
+    }
+
+    return FP32Vec16(temp);
+  };
+  FP32Vec16 abs() const {
+    return FP32Vec16(
+        float32x4x4_t({vabsq_f32(reg.val[0]), vabsq_f32(reg.val[1]),
+                       vabsq_f32(reg.val[2]), vabsq_f32(reg.val[3])}));
+  }
   float reduce_sum() const {
     AliasReg ar;
     ar.reg = reg;
@@ -473,6 +639,24 @@ struct FP32Vec16 : public Vec<FP32Vec16> {
     return answer;
   };
 
+  float reduce_max() const {
+    AliasReg ar;
+    ar.reg = reg;
+    float max_v = std::numeric_limits<float>::lowest();
+    unroll_loop<int, VEC_ELEM_NUM>(
+        [&max_v, &ar](int i) { max_v = std::max(max_v, ar.values[i]); });
+    return max_v;
+  }
+
+  float reduce_min() const {
+    AliasReg ar;
+    ar.reg = reg;
+    float min_v = std::numeric_limits<float>::max();
+    unroll_loop<int, VEC_ELEM_NUM>(
+        [&min_v, &ar](int i) { min_v = std::min(min_v, ar.values[i]); });
+    return min_v;
+  }
+
   template <int group_size>
   float reduce_sub_sum(int idx) {
     static_assert(VEC_ELEM_NUM % group_size == 0);
@@ -493,6 +677,83 @@ struct FP32Vec16 : public Vec<FP32Vec16> {
     vst1q_f32(ptr + 8, reg.val[2]);
     vst1q_f32(ptr + 12, reg.val[3]);
   };
+
+  void save(float* ptr, const int elem_num) const {
+    int full_blocks = elem_num / NUM_ELEMENTS_REG(reg.val[0]);
+    int remainder = elem_num % NUM_ELEMENTS_REG(reg.val[0]);
+
+    for (int i = 0; i < full_blocks; i++)
+      vst1q_f32(
+          reinterpret_cast<float32_t*>(ptr) + NUM_ELEMENTS_REG(reg.val[0]) * i,
+          reg.val[i]);
+
+    if (remainder > 0) {
+      float32x4_t temp = reg.val[full_blocks];
+      float* base = reinterpret_cast<float32_t*>(ptr) +
+                    full_blocks * NUM_ELEMENTS_REG(reg.val[0]);
+      if (remainder > 0) base[0] = vgetq_lane_f32(temp, 0);
+      if (remainder > 1) base[1] = vgetq_lane_f32(temp, 1);
+      if (remainder > 2) base[2] = vgetq_lane_f32(temp, 2);
+    }
+  }
+};
+
+struct INT8Vec16 : public Vec<INT8Vec16> {
+  constexpr static int VEC_ELEM_NUM = 16;
+  union AliasReg {
+    int8x16_t reg;
+    int8_t values[VEC_ELEM_NUM];
+  };
+  int8x16_t reg;
+
+  explicit INT8Vec16(const FP32Vec16& vec) {
+    // Convert each 128-bit float32 vector to int32
+    int32x4_t part0 =
+        vcvtq_s32_f32(vec.reg.val[0]);  // Convert first 128-bit block
+    int32x4_t part1 =
+        vcvtq_s32_f32(vec.reg.val[1]);  // Convert second 128-bit block
+    int32x4_t part2 =
+        vcvtq_s32_f32(vec.reg.val[2]);  // Convert third 128-bit block
+    int32x4_t part3 =
+        vcvtq_s32_f32(vec.reg.val[3]);  // Convert fourth 128-bit block
+
+    // Narrow each 32-bit vector to 8 bits and combine
+    int8x8_t lower =
+        vqmovn_s16(vcombine_s16(vqmovn_s32(part0), vqmovn_s32(part1)));
+    int8x8_t upper =
+        vqmovn_s16(vcombine_s16(vqmovn_s32(part2), vqmovn_s32(part3)));
+    reg = vcombine_s8(lower, upper);  // Combine to form a single 128-bit vector
+  }
+
+  void save(int8_t* ptr) const { vst1q_s8(ptr, reg); };
+
+  void save(int8_t* ptr, const int elem_num) const {
+    int full_blocks = elem_num / NUM_ELEMENTS_REG(reg);
+    int remainder = elem_num % NUM_ELEMENTS_REG(reg);
+
+    for (int i = 0; i < full_blocks; i++)
+      vst1q_s8(reinterpret_cast<int8_t*>(ptr) + NUM_ELEMENTS_REG(reg) * i, reg);
+    if (remainder > 0) {
+      int8x16_t temp = reg;
+      int8_t* base =
+          reinterpret_cast<int8_t*>(ptr) + full_blocks * NUM_ELEMENTS_REG(reg);
+      if (remainder > 0) base[0] = vgetq_lane_s8(temp, 0);
+      if (remainder > 1) base[1] = vgetq_lane_s8(temp, 1);
+      if (remainder > 2) base[2] = vgetq_lane_s8(temp, 2);
+      if (remainder > 3) base[3] = vgetq_lane_s8(temp, 3);
+      if (remainder > 4) base[4] = vgetq_lane_s8(temp, 4);
+      if (remainder > 5) base[5] = vgetq_lane_s8(temp, 5);
+      if (remainder > 6) base[6] = vgetq_lane_s8(temp, 6);
+      if (remainder > 7) base[7] = vgetq_lane_s8(temp, 7);
+      if (remainder > 8) base[8] = vgetq_lane_s8(temp, 8);
+      if (remainder > 9) base[9] = vgetq_lane_s8(temp, 9);
+      if (remainder > 10) base[10] = vgetq_lane_s8(temp, 10);
+      if (remainder > 11) base[11] = vgetq_lane_s8(temp, 11);
+      if (remainder > 12) base[12] = vgetq_lane_s8(temp, 12);
+      if (remainder > 13) base[13] = vgetq_lane_s8(temp, 13);
+      if (remainder > 14) base[14] = vgetq_lane_s8(temp, 14);
+    }
+  };
 };
 
 template <typename T>

csrc/cpu/dnnl_helper.hpp

@@ -57,6 +57,7 @@ class DNNLPrimitiveHelper {
   // 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,
@@ -90,6 +91,27 @@ class DNNLPrimitiveHelper {
     }
 
     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,
@@ -98,6 +120,7 @@ class DNNLPrimitiveHelper {
       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();
@@ -111,24 +134,34 @@ class DNNLPrimitiveHelper {
                             (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, a_m},
-                        {DNNL_ARG_WEIGHTS, b_m},
+                        {DNNL_ARG_SRC, mat_src_mem},
+                        {DNNL_ARG_WEIGHTS, mat_weights_mem},
                         {DNNL_ARG_BIAS, bias_m},
-                        {DNNL_ARG_DST, c_m},
+                        {DNNL_ARG_DST, mat_dst_mem},
                         {DNNL_ARG_ATTR_SCALES | DNNL_ARG_WEIGHTS, b_scales_m},
                     });
       } else {
         matmul.execute(
             stream, {
-                        {DNNL_ARG_SRC, a_m},
-                        {DNNL_ARG_WEIGHTS, b_m},
-                        {DNNL_ARG_DST, c_m},
+                        {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},
                     });
       }
@@ -138,19 +171,19 @@ class DNNLPrimitiveHelper {
         dnnl::memory bias_m(bias_md, engine, (void*)bias);
         matmul.execute(
             stream, {
-                        {DNNL_ARG_SRC, a_m},
-                        {DNNL_ARG_WEIGHTS, b_m},
+                        {DNNL_ARG_SRC, mat_src_mem},
+                        {DNNL_ARG_WEIGHTS, mat_weights_mem},
                         {DNNL_ARG_BIAS, bias_m},
-                        {DNNL_ARG_DST, c_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, a_m},
-                        {DNNL_ARG_WEIGHTS, b_m},
-                        {DNNL_ARG_DST, c_m},
+                        {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},
                     });
@@ -170,5 +203,4 @@ class DNNLPrimitiveHelper {
     return stream;
   }
 };
-
 #endif

csrc/cpu/quant.cpp

@@ -36,7 +36,7 @@ struct KernelVecType<c10::Half> {
   using cvt_vec_type = vec_op::FP32Vec16;
 };
 
-#ifdef __AVX512F__
+#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,
@@ -598,8 +598,9 @@ 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 support.")
+  TORCH_CHECK(false,
+              "static_scaled_int8_quant_impl requires AVX512/powerpc64/AArch64 "
+              "support.")
 }
 
 template <typename scalar_t>
@@ -607,9 +608,9 @@ 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 support.")
+  TORCH_CHECK(false,
+              "dynamic_scaled_int8_quant_impl requires "
+              "AVX512/powerpc64/AArch64 support.")
 }
 
 template <bool PerChannel, typename scalar_t>
@@ -617,7 +618,8 @@ 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 support.")
+  TORCH_CHECK(
+      false, "static_quant_epilogue requires AVX512/powerpc64/AArch64 support.")
 }
 
 template <typename scalar_t>
@@ -626,8 +628,9 @@ void dynamic_quant_epilogue(const float* input, scalar_t* output,
                             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 support.")
+  TORCH_CHECK(
+      false,
+      "dynamic_quant_epilogue requires AVX512/powerpc64/AArch64 support.")
 }
 #endif
 }  // namespace

csrc/cpu/sgl-kernels/common.h

@@ -0,0 +1,238 @@
+// Adapted from
+// https://github.com/sgl-project/sglang/tree/main/sgl-kernel/csrc/cpu
+
+#pragma once
+
+#include <ATen/ATen.h>
+#include <ATen/Parallel.h>
+#include <ATen/record_function.h>
+
+// clang-format off
+
+#if defined(_OPENMP)
+#include <omp.h>
+#endif
+
+namespace {
+
+// dispatch bool
+#define AT_DISPATCH_BOOL(BOOL_V, BOOL_NAME, ...)                                 \
+  [&] {                                                                          \
+    if (BOOL_V) {                                                                \
+      constexpr bool BOOL_NAME = true;                                           \
+      return __VA_ARGS__();                                                      \
+    } else {                                                                     \
+      constexpr bool BOOL_NAME = false;                                          \
+      return __VA_ARGS__();                                                      \
+    }                                                                            \
+  }()
+
+// dispatch: bfloat16, float16, int8_t, fp8_e4m3
+#define CPU_DISPATCH_PACKED_TYPES(TYPE, ...)                                    \
+  [&] {                                                                         \
+    switch (TYPE) {                                                             \
+      case at::ScalarType::BFloat16 : {                                         \
+        using packed_t = at::BFloat16;                                          \
+        return __VA_ARGS__();                                                   \
+      }                                                                         \
+      case at::ScalarType::Half: {                                              \
+        using packed_t = at::Half;                                              \
+        return __VA_ARGS__();                                                   \
+      }                                                                         \
+      case at::ScalarType::Char : {                                             \
+        using packed_t = int8_t;                                                \
+        return __VA_ARGS__();                                                   \
+      }                                                                         \
+      case at::ScalarType::Float8_e4m3fn : {                                    \
+        using packed_t = at::Float8_e4m3fn;                                     \
+        return __VA_ARGS__();                                                   \
+      }                                                                         \
+      default:                                                                  \
+        TORCH_CHECK(false, "Unsupported floating data type.\n");                \
+    }                                                                           \
+  }()
+
+#define UNUSED(x) (void)(x)
+
+#define CHECK_CPU(x) TORCH_CHECK(x.device().type() == at::kCPU, #x " must be a CPU tensor")
+
+#define CHECK_CONTIGUOUS(x) TORCH_CHECK(x.is_contiguous(), #x " must be contiguous")
+#define CHECK_LAST_DIM_CONTIGUOUS(x) \
+  TORCH_CHECK(x.strides()[x.strides().size() - 1] == 1, #x "must be contiguous at last dimension")
+
+#define CHECK_INPUT(x) \
+  CHECK_CPU(x);        \
+  CHECK_CONTIGUOUS(x)
+#define CHECK_LAST_DIM_CONTIGUOUS_INPUT(x) \
+  CHECK_CPU(x);                            \
+  CHECK_LAST_DIM_CONTIGUOUS(x)
+
+#define CHECK_DIM(d, x) TORCH_CHECK(x.dim() == d, #x " must be a " #d "D tensor")
+
+#define CHECK_EQ(a, b) TORCH_CHECK((a) == (b), "CHECK_EQ(" #a ", " #b ") failed. ", a, " vs ", b)
+
+// parallel routines
+constexpr int GRAIN_SIZE = 1024;
+
+template <typename T, typename std::enable_if<std::is_integral<T>::value, int>::type = 0>
+inline T div_up(T x, T y) { return (x + y - 1) / y; }
+
+template <typename T>
+inline void balance211(T n, T nth, T ith, T& n_start, T& n_end) {
+#if 0
+    // onednn partition pattern
+    T& n_my = n_end;
+    if (nth <= 1 || n == 0) {
+        n_start = 0;
+        n_my = n;
+    } else {
+        T n1 = div_up(n, nth);
+        T n2 = n1 - 1;
+        T T1 = n - n2 * nth;
+        n_my = ith < T1 ? n1 : n2;
+        n_start = ith <= T1 ? ith*n1 : T1 * n1 + (ith - T1) * n2;
+    }
+    n_end += n_start;
+#else
+    // pytorch aten partition pattern
+    T n_my = div_up(n, nth);
+    n_start = ith * n_my;
+    n_end = std::min(n_start + n_my, n);
+#endif
+}
+
+template <typename func_t>
+inline void parallel_for(int n, const func_t& f) {
+#if defined(_OPENMP)
+#pragma omp parallel
+{
+    int nth = omp_get_num_threads();
+    int ith = omp_get_thread_num();
+    int tbegin, tend;
+    balance211(n, nth, ith, tbegin, tend);
+    f(tbegin, tend);
+}
+#else
+    f(0, n);
+#endif
+}
+
+// for 1d parallel, use `actual_nth`
+// for 2d parallel, use even nths, e.g. 43->42
+int inline adjust_num_threads(int m) {
+  int actual_nth = at::get_num_threads();
+  if (m == 1) {
+    return actual_nth;
+  }
+  return std::max(1, (actual_nth >> 1) * 2);
+}
+
+template <typename func_t>
+inline void parallel_2d(int m, int n, const func_t& f) {
+
+  // make sure we have even num_threads
+  int nth = adjust_num_threads(m);
+
+  // [NOTE] thread blocking:
+  //
+  //   1) prefer square block per thread
+  //   2) use even number of CPU cores
+  //   3) use all `num_threads` cores
+  //
+  //   we have:
+  //     TM * TN = T
+  //     BM / TM = BN / TN
+  //   then:
+  //     TM = ((BM / BN) * T) ^ 0.5
+  //
+  float r = float(m) / n;
+  int nth_m = std::ceil(std::sqrt(r * nth));
+  int nth_n = 1;
+  for (; nth_m > 0; --nth_m) {
+    nth_n = nth / nth_m;
+    if (nth_m * nth_n == nth) {
+      break;
+    }
+  }
+
+#if defined(_OPENMP)
+#pragma omp parallel num_threads(nth)
+{
+  int ith = omp_get_thread_num();
+  int ith_m = ith / nth_n;
+  int ith_n = ith % nth_n;
+
+  int thread_block_m = div_up(m, nth_m);
+  int thread_block_n = div_up(n, nth_n);
+
+  int begin_m = ith_m * thread_block_m;
+  int end_m = std::min(m, begin_m + thread_block_m);
+  int begin_n = ith_n * thread_block_n;
+  int end_n = std::min(n, begin_n + thread_block_n);
+
+  f(begin_m, end_m, begin_n, end_n);
+}
+#else
+  f(0, m, 0, n);
+#endif
+}
+
+template <typename T>
+int get_cache_blocks(int BLOCK_SIZE, int K) {
+  // L2 2MB and ratio of 50%
+  const int L2_size = 2048 * 1024 >> 1;
+  return std::max(1, int(L2_size / (BLOCK_SIZE * K * sizeof(T))));
+}
+
+// data indexing for dimension collapse
+template <typename T>
+inline T data_index_init(T offset) {
+  return offset;
+}
+
+template <typename T, typename... Args>
+inline T data_index_init(T offset, T& x, const T& X, Args&&... args) {
+  offset = data_index_init(offset, std::forward<Args>(args)...);
+  x = offset % X;
+  return offset / X;
+}
+
+inline bool data_index_step() {
+  return true;
+}
+
+template <typename T, typename... Args>
+inline bool data_index_step(T& x, const T& X, Args&&... args) {
+  if (data_index_step(std::forward<Args>(args)...)) {
+    x = ((x + 1) == X) ? 0 : (x + 1);
+    return x == 0;
+  }
+  return false;
+}
+
+// forced unroll for perf critical path
+
+#if __has_attribute(always_inline)
+#define ALWAYS_INLINE __attribute__((__always_inline__)) inline
+#else
+#define ALWAYS_INLINE inline
+#endif
+
+template <int n>
+struct Unroll {
+  template <typename Func, typename... Args>
+  ALWAYS_INLINE void operator()(const Func& f, Args... args) const {
+    Unroll<n - 1>{}(f, args...);
+    f(std::integral_constant<int, n - 1>{}, args...);
+  }
+};
+
+template <>
+struct Unroll<1> {
+  template <typename Func, typename... Args>
+  ALWAYS_INLINE void operator()(const Func& f, Args... args) const {
+    f(std::integral_constant<int, 0>{}, args...);
+  }
+};
+
+} // anonymous namespace

csrc/cpu/sgl-kernels/gemm.cpp

@@ -0,0 +1,464 @@
+// Adapted from
+// https://github.com/sgl-project/sglang/tree/main/sgl-kernel/csrc/cpu
+
+#include "common.h"
+#include "vec.h"
+#include "gemm.h"
+
+// clang-format off
+
+namespace {
+
+// packed   layout:
+//   quants {N, K}  int8_t
+//   comp   {N}     int32_t
+template <int BLOCK_N>
+inline void s8s8_compensation(int8_t* __restrict__ packed, int K) {
+#if defined(CPU_CAPABILITY_AVX512)
+  constexpr int COLS = BLOCK_N / 16;
+  __m512i vcomp[COLS];
+
+  for (int col = 0; col < COLS; ++col) {
+    vcomp[col] = _mm512_setzero_si512();
+  }
+
+  const int64_t offset = BLOCK_N * K;
+  const __m512i off = _mm512_set1_epi8(static_cast<char>(0x80));
+  for (int k = 0; k < K / 4; ++k) {
+    for (int col = 0; col < COLS; ++col) {
+      __m512i vb = _mm512_loadu_si512((const __m512i *)(packed + k * BLOCK_N * 4 + col * 64));
+      vcomp[col] = _mm512_dpbusd_epi32(vcomp[col], off, vb);
+    }
+  }
+
+  for (int col = 0; col < COLS; ++col) {
+    _mm512_storeu_si512((__m512i *)(packed + offset + col * 64), vcomp[col]);
+  }
+#else
+  TORCH_CHECK(false, "s8s8_compensation not implemented!");
+#endif
+}
+
+// convert to vnni format
+// from [N, K] to [K/2, N, 2] for bfloat16 and float16
+template <typename packed_t>
+inline void pack_vnni(packed_t* __restrict__ packed, const packed_t* __restrict__ weight, int N, int K) {
+  const int VNNI_BLK = 2;
+  for (int n = 0; n < N; ++n) {
+    for (int k = 0; k < K / VNNI_BLK; ++k) {
+      for (int d = 0; d < VNNI_BLK; ++d) {
+        packed[k * N * VNNI_BLK + n * VNNI_BLK + d] = weight[n * K + k * VNNI_BLK + d];
+      }
+    }
+  }
+}
+
+template <>
+inline void pack_vnni<int8_t>(int8_t* __restrict__ packed, const int8_t* __restrict__ weight, int N, int K) {
+  constexpr int BLOCK_N = block_size_n();
+  TORCH_CHECK(N == BLOCK_N);
+
+  const int VNNI_BLK = 4;
+  for (int n = 0; n < N; ++n) {
+    for (int k = 0; k < K / VNNI_BLK; ++k) {
+      for (int d = 0; d < VNNI_BLK; ++d) {
+        packed[k * N * VNNI_BLK + n * VNNI_BLK + d] = weight[n * K + k * VNNI_BLK + d];
+      }
+    }
+  }
+  s8s8_compensation<BLOCK_N>(packed, K);
+}
+
+template <typename scalar_t>
+inline void copy_stub(scalar_t* __restrict__ out, const float* __restrict__ input, int64_t size) {
+  using bVec = at::vec::Vectorized<scalar_t>;
+  using fVec = at::vec::Vectorized<float>;
+  constexpr int kVecSize = bVec::size();
+
+  int64_t d;
+  #pragma GCC unroll 4
+  for (d = 0; d <= size - kVecSize; d += kVecSize) {
+    fVec data0 = fVec::loadu(input + d);
+    fVec data1 = fVec::loadu(input + d + fVec::size());
+    bVec out_vec = convert_from_float_ext<scalar_t>(data0, data1);
+    out_vec.store(out + d);
+  }
+  for (; d < size; ++d) {
+    out[d] = static_cast<scalar_t>(input[d]);
+  }
+}
+
+template <typename scalar_t>
+inline void copy_add_stub(scalar_t* __restrict__ out, const float* __restrict__ input, const float* __restrict__ bias, int64_t size) {
+  using bVec = at::vec::Vectorized<scalar_t>;
+  using fVec = at::vec::Vectorized<float>;
+  constexpr int kVecSize = bVec::size();
+
+  int64_t d;
+  #pragma GCC unroll 4
+  for (d = 0; d <= size - kVecSize; d += kVecSize) {
+    fVec data0 = fVec::loadu(input + d) + fVec::loadu(bias + d);
+    fVec data1 = fVec::loadu(input + d + fVec::size()) + fVec::loadu(bias + d + fVec::size());
+    bVec out_vec = convert_from_float_ext<scalar_t>(data0, data1);
+    out_vec.store(out + d);
+  }
+  for (; d < size; ++d) {
+    out[d] = static_cast<scalar_t>(input[d] + bias[d]);
+  }
+}
+
+template <typename scalar_t, bool has_bias, int BLOCK_M, int BLOCK_N>
+struct tinygemm_kernel_nn {
+  static inline void apply(
+      const scalar_t* __restrict__ A, const scalar_t* __restrict__ B, scalar_t* __restrict__ C,
+      const float* __restrict__ bias, int64_t K, int64_t lda, int64_t ldb, int64_t ldc) {
+    TORCH_CHECK(false, "tinygemm_kernel_nn: scalar path not implemented!");
+  }
+};
+
+#if defined(CPU_CAPABILITY_AVX512)
+template <bool has_bias, int BLOCK_M, int BLOCK_N>
+struct tinygemm_kernel_nn<at::BFloat16, has_bias, BLOCK_M, BLOCK_N> {
+  static inline void apply(
+      const at::BFloat16* __restrict__ A, const at::BFloat16* __restrict__ B, at::BFloat16* __restrict__ C,
+      const float* __restrict__ bias, int64_t K, int64_t lda, int64_t ldb, int64_t ldc) {
+
+    constexpr int ROWS = BLOCK_M;
+    constexpr int COLS = BLOCK_N / 16;
+
+    // prefetch distance
+    constexpr int PREFETCH_SIZE_K = 0;
+
+    __m512bh va;
+    __m512bh vb[COLS];
+    __m512 vc[ROWS * COLS];
+
+    auto loadc = [&](auto i) {
+      constexpr int col = i % COLS;
+      if constexpr (has_bias) {
+        vc[i] = _mm512_loadu_ps(bias + col * 16);
+      } else {
+        vc[i] = _mm512_set1_ps(0.f);
+      }
+    };
+    Unroll<ROWS * COLS>{}(loadc);
+
+    const int64_t K2 = K >> 1;
+    const int64_t lda2 = lda >> 1;
+    const int64_t ldb2 = ldb; // ldb * 2 >> 1;
+    const float* a_ptr = reinterpret_cast<const float*>(A);
+    const float* b_ptr = reinterpret_cast<const float*>(B);
+
+    auto compute = [&](auto i, int64_t k) {
+      constexpr int row = i / COLS;
+      constexpr int col = i % COLS;
+
+      if constexpr (col == 0) {
+        va = (__m512bh)(_mm512_set1_ps(a_ptr[row * lda2 + k]));
+      }
+      if constexpr (row == 0) {
+        vb[col] = (__m512bh)(_mm512_loadu_si512(b_ptr + k * ldb2 + col * 16));
+        if constexpr (PREFETCH_SIZE_K > 0) {
+          _mm_prefetch(b_ptr + (k + PREFETCH_SIZE_K) * ldb2 + col * 16, _MM_HINT_T0);
+        }
+      }
+      vc[i] = _mm512_dpbf16_ps(vc[i], va, vb[col]);
+    };
+    for (int64_t k = 0; k < K2; ++k) {
+      Unroll<ROWS * COLS>{}(compute, k);
+    }
+
+    auto storec = [&](auto i) {
+      constexpr int row = i / COLS;
+      constexpr int col = i % COLS;
+      // for COLS = 2, 4 use 512bit store
+      // for COLS = 1, 3 use 256bit store
+      if constexpr (COLS % 2 == 0) {
+        if constexpr (col % 2 == 0) {
+          _mm512_storeu_si512(
+              reinterpret_cast<__m512i*>((C + row * ldc + col * 16)),
+              (__m512i)(_mm512_cvtne2ps_pbh(vc[row * COLS + col + 1], vc[row * COLS + col])));
+        }
+      } else {
+        _mm256_storeu_si256(
+            reinterpret_cast<__m256i*>(C + row * ldc + col * 16),
+            (__m256i)(_mm512_cvtneps_pbh(vc[i])));
+      }
+    };
+    Unroll<ROWS * COLS>{}(storec);
+  }
+};
+#endif
+
+#define LAUNCH_TINYGEMM_KERNEL_NN(MB_SIZE, NB_SIZE)                          \
+    tinygemm_kernel_nn<scalar_t, has_bias, MB_SIZE, NB_SIZE>::apply(         \
+        A + mb_start * lda, B + nb_start * 2, C + mb_start * ldc + nb_start, \
+        has_bias ? bias + nb_start : nullptr, K, lda, ldb, ldc);
+
+template <typename scalar_t, bool has_bias>
+struct brgemm {
+  static inline void apply(
+      const scalar_t* __restrict__ A, const scalar_t* __restrict__ B, scalar_t* __restrict__ C,
+      float* __restrict__ Ctmp, const float* __restrict__ bias,
+      int64_t M, int64_t N, int64_t K, int64_t lda, int64_t ldb, int64_t ldc) {
+
+    constexpr int BLOCK_N = block_size_n();
+    at::native::cpublas::brgemm(
+        M, N, K, lda, ldb, BLOCK_N, /* add_C */false,
+        A, B, Ctmp);
+
+    // copy from Ctmp to C
+    for (int64_t m = 0; m < M; ++m) {
+      if constexpr (has_bias) {
+        copy_add_stub(C + m * ldc, Ctmp + m * BLOCK_N, bias, N);
+      } else {
+        copy_stub(C + m * ldc, Ctmp + m * BLOCK_N, N);
+      }
+    }
+  }
+};
+
+template <typename scalar_t, bool has_bias>
+void tinygemm_kernel(
+    const scalar_t* __restrict__ A,
+    const scalar_t* __restrict__ B,
+    scalar_t* __restrict__ C,
+    float* __restrict__ Ctmp,
+    const float* __restrict__ bias,
+    int64_t M,
+    int64_t N,
+    int64_t K,
+    int64_t lda,
+    int64_t ldb,
+    int64_t ldc,
+    bool brg) {
+
+  if (brg) {
+    brgemm<scalar_t, has_bias>::apply(
+        A, B, C, Ctmp, bias,
+        M, N, K, lda, ldb, ldc);
+    return;
+  }
+
+  // pattern: 1-4-16
+  constexpr int64_t BLOCK_M = 4;
+  constexpr int64_t BLOCK_N = 64;
+  const int64_t MB = div_up(M, BLOCK_M);
+  const int64_t NB = div_up(N, BLOCK_N);
+  for (int mb = 0; mb < MB; ++mb) {
+    int64_t mb_start = mb * BLOCK_M;
+    int64_t mb_size = std::min(BLOCK_M, M - mb_start);
+    for (int64_t nb = 0; nb < NB; ++nb) {
+      int64_t nb_start = nb * BLOCK_N;
+      int64_t nb_size = std::min(BLOCK_N, N - nb_start);
+
+      switch(mb_size << 4 | nb_size >> 4) {
+        // mb_size = 1
+        case 0x12: LAUNCH_TINYGEMM_KERNEL_NN(1, 32); break;
+        case 0x14: LAUNCH_TINYGEMM_KERNEL_NN(1, 64); break;
+        // mb_size = 2
+        case 0x22: LAUNCH_TINYGEMM_KERNEL_NN(2, 32); break;
+        case 0x24: LAUNCH_TINYGEMM_KERNEL_NN(2, 64); break;
+        // mb_size = 3
+        case 0x32: LAUNCH_TINYGEMM_KERNEL_NN(3, 32); break;
+        case 0x34: LAUNCH_TINYGEMM_KERNEL_NN(3, 64); break;
+        // mb_size = 4
+        case 0x42: LAUNCH_TINYGEMM_KERNEL_NN(4, 32); break;
+        case 0x44: LAUNCH_TINYGEMM_KERNEL_NN(4, 64); break;
+        default: TORCH_CHECK(false, "Unexpected block size, ", mb_size, "x", "nb_size");
+      }
+    }
+  }
+}
+
+template <typename scalar_t>
+void weight_packed_linear_kernel_impl(
+    scalar_t* __restrict__ out,
+    const scalar_t* __restrict__ mat1,
+    const scalar_t* __restrict__ mat2,
+    const float* __restrict__ bias,
+    int64_t M,
+    int64_t N,
+    int64_t K,
+    int64_t mat1_strideM,
+    int64_t out_strideM) {
+
+  constexpr int64_t BLOCK_M = block_size_m();
+  constexpr int64_t BLOCK_N = block_size_n();
+  const int64_t MB = div_up(M, BLOCK_M);
+  const int64_t NB = div_up(N, BLOCK_N);
+
+  // use avx512-bf16 when a) M is small; b) dtype is bfloat16, otherwise use amx
+  const bool use_brgemm = (M > 4) || (!std::is_same_v<scalar_t, at::BFloat16>);
+
+  // l2 cache block for n
+  int64_t cache_blocks_nb = get_cache_blocks<scalar_t>(BLOCK_N, K);
+
+  // parallel on [MB, NB]
+  AT_DISPATCH_BOOL(bias != nullptr, has_bias, [&] {
+    parallel_2d(MB, NB, [&](int64_t begin_mb, int64_t end_mb, int64_t begin_nb, int64_t end_nb) {
+
+      // for brgemm, use float32 for accumulate
+      alignas(64) float Ctmp[BLOCK_M * BLOCK_N];
+
+      for (int64_t nbb = begin_nb; nbb < end_nb; nbb += cache_blocks_nb) {
+      for (int64_t mb = begin_mb; mb < end_mb; ++mb) {
+      for (int64_t nb = nbb; nb < std::min(nbb + cache_blocks_nb, end_nb); ++nb) {
+
+        int64_t mb_start = mb * BLOCK_M;
+        int64_t mb_size = std::min(M - mb_start, BLOCK_M);
+        int64_t nb_start = nb * BLOCK_N;
+        int64_t nb_size = std::min(N - nb_start, BLOCK_N);
+
+        tinygemm_kernel<scalar_t, has_bias>(
+            /*   A */ mat1 + mb_start * mat1_strideM,
+            /*   B */ mat2 + nb_start * K /* nb * BLOCK_N * K */,
+            /*   C */ out + mb_start * out_strideM + nb_start,
+            /* Ctmp*/ Ctmp,
+            /* bias*/ bias + nb_start,
+            /*   M */ mb_size,
+            /*   N */ nb_size,
+            /*   K */ K,
+            /* lda */ mat1_strideM,
+            /* ldb */ nb_size,
+            /* ldc */ out_strideM,
+            /* brg */ use_brgemm);
+      }}}
+
+      if (use_brgemm) {
+        at::native::cpublas::brgemm_release();
+      }
+    });
+  });
+}
+
+} // anonymous namespace
+
+// tinygemm interface
+template <typename scalar_t>
+void tinygemm_kernel(const scalar_t* __restrict__ A, const scalar_t* __restrict__ B, scalar_t* __restrict__ C,
+    float* __restrict__ Ctmp, int64_t M, int64_t N, int64_t K, int64_t lda, int64_t ldb, int64_t ldc, bool brg) {
+  tinygemm_kernel<scalar_t, false>(A, B, C, Ctmp, nullptr, M, N, K, lda, ldb, ldc, brg);
+}
+
+#define INSTANTIATE_TINYGEMM_TEMPLATE(TYPE)                                             \
+    template void tinygemm_kernel<TYPE>(                                                \
+        const TYPE* __restrict__ A, const TYPE* __restrict__ B, TYPE* __restrict__ C,   \
+        float* __restrict__ Ctmp, int64_t M, int64_t N, int64_t K, int64_t lda,         \
+        int64_t ldb, int64_t ldc, bool brg)
+
+INSTANTIATE_TINYGEMM_TEMPLATE(at::BFloat16);
+INSTANTIATE_TINYGEMM_TEMPLATE(at::Half);
+
+at::Tensor convert_weight_packed(at::Tensor& weight) {
+  // for 3d moe weights
+  // weight : [E, OC, IC]
+  //     w1 : [E, 2N,  K]
+  //     w2 : [E,  K,  N]
+  CHECK_INPUT(weight);
+
+  const int64_t ndim = weight.ndimension();
+  TORCH_CHECK(ndim == 2 || ndim == 3, "expect weight to be 2d or 3d, got ", ndim, "d tensor.");
+  const auto st = weight.scalar_type();
+  const int64_t E = ndim == 3 ? weight.size(0) : 1;
+  const int64_t OC = ndim == 3 ? weight.size(1) : weight.size(0);
+  const int64_t IC = ndim == 3 ? weight.size(2) : weight.size(1);
+
+  // we handle 2 TILE_N at a time.
+  TORCH_CHECK(OC % TILE_N == 0, "invalid weight out features ", OC);
+  TORCH_CHECK(IC % TILE_K == 0, "invalid weight input features ", IC);
+
+  constexpr int64_t BLOCK_N = block_size_n();
+  const int64_t NB = div_up(OC, BLOCK_N);
+
+  // use phony sizes here [E, OC, IC], for each [E], [OC, IC] -> [IC / 2, OC, 2]
+  auto packed_weight = at::empty({}, weight.options());
+  const int64_t stride = OC * IC;
+
+  TORCH_CHECK(st == at::kBFloat16 || st == at::kHalf || st == at::kChar || st == at::kFloat8_e4m3fn,
+      "expect weight to be bfloat16, float16, int8 or fp8_e4m3.");
+
+  CPU_DISPATCH_PACKED_TYPES(st, [&] {
+    // adjust most inner dimension size
+    const int packed_row_size = get_row_size<packed_t>(IC);
+    auto sizes = weight.sizes().vec();
+    sizes[ndim - 1] = packed_row_size;
+    packed_weight.resize_(sizes);
+
+    const packed_t* w_data = weight.data_ptr<packed_t>();
+    packed_t* packed_data = packed_weight.data_ptr<packed_t>();
+
+    // parallel on {E, NB}
+    at::parallel_for(0, E * NB, 0, [&](int64_t begin, int64_t end) {
+      int64_t e{0}, nb{0};
+      data_index_init(begin, e, E, nb, NB);
+
+      for (int64_t i = begin; i < end; ++i) {
+        UNUSED(i);
+
+        int64_t n = nb * BLOCK_N;
+        int64_t n_size = std::min(BLOCK_N, OC - n);
+        pack_vnni<packed_t>(
+            packed_data + e * OC * packed_row_size + n * packed_row_size,
+            w_data + e * stride + n * IC,
+            n_size,
+            IC);
+
+        // move to the next index
+        data_index_step(e, E, nb, NB);
+      }
+    });
+  });
+  return packed_weight;
+}
+
+// mat1 : [M, K]
+// mat2 : [N, K]
+// bias : [N]
+// out  : [M, N]
+//
+at::Tensor weight_packed_linear(at::Tensor& mat1, at::Tensor& mat2,
+    const std::optional<at::Tensor>& bias, bool is_vnni) {
+  RECORD_FUNCTION(
+    "sgl-kernel::weight_packed_linear", std::vector<c10::IValue>({mat1, mat2, bias}));
+
+  auto packed_w = is_vnni ? mat2 : convert_weight_packed(mat2);
+
+  CHECK_LAST_DIM_CONTIGUOUS_INPUT(mat1);
+  CHECK_INPUT(mat2);
+
+  int64_t M = mat1.size(0);
+  int64_t N = mat2.size(0);
+  int64_t K = mat2.size(1);
+  CHECK_EQ(mat1.size(1), K);
+  CHECK_DIM(2, mat1);
+  CHECK_DIM(2, mat2);
+
+  auto out = at::empty({M, N}, mat1.options());
+
+  // strides
+  int64_t mat1_strideM = mat1.stride(0);
+  int64_t out_strideM = out.stride(0);
+
+  const bool has_bias = bias.has_value();
+  const float* bias_data = nullptr;
+  if (has_bias) {
+    CHECK_EQ(bias.value().size(0), N);
+    bias_data = bias.value().data_ptr<float>();
+  }
+
+  AT_DISPATCH_REDUCED_FLOATING_TYPES(mat1.scalar_type(), "weight_packed_linear_kernel_impl", [&] {
+    weight_packed_linear_kernel_impl<scalar_t>(
+        out.data_ptr<scalar_t>(),
+        mat1.data_ptr<scalar_t>(),
+        packed_w.data_ptr<scalar_t>(),
+        bias_data,
+        M,
+        N,
+        K,
+        mat1_strideM,
+        out_strideM);
+  });
+
+  return out;
+}

csrc/cpu/sgl-kernels/gemm.h

@@ -0,0 +1,266 @@
+#pragma once
+
+#include <ATen/native/CPUBlas.h>
+
+// clang-format off
+
+// amx-bf16
+#define TILE_M 16
+#define TILE_N 16
+#define TILE_K 32
+
+// block size for AMX gemm
+constexpr int block_size_m() { return 2 * TILE_M; }
+constexpr int block_size_n() { return 2 * TILE_N; }
+
+// define threshold using brgemm (intel AMX)
+template <typename T> inline bool can_use_brgemm(int M);
+template <> inline bool can_use_brgemm<at::BFloat16>(int M) { return M > 4; }
+template <> inline bool can_use_brgemm<at::Half>(int M) { return true; }
+// TODO: add u8s8 brgemm, this requires PyTorch 2.7
+template <> inline bool can_use_brgemm<int8_t>(int M) { return false; }
+template <> inline bool can_use_brgemm<at::Float8_e4m3fn>(int M) { return M > 4; }
+template <> inline bool can_use_brgemm<at::quint4x2>(int M) { return M > 4; }
+
+// work around compiler internal error
+#define BLOCK_K 128 // 4 * TILE_K
+
+// adjust leading dimension size for K
+template <typename T>
+inline int64_t get_row_size(int64_t K) {
+  return K;
+}
+
+template <>
+inline int64_t get_row_size<int8_t>(int64_t K) {
+  return K + sizeof(int32_t);
+}
+
+inline int64_t get_row_size(int64_t K, bool use_int8_w8a8) {
+  return use_int8_w8a8 ? K + sizeof(int32_t) : K;
+}
+
+// pack weight to vnni format
+at::Tensor convert_weight_packed(at::Tensor& weight);
+
+// moe implementations for int8 w8a8
+template <typename scalar_t>
+void fused_experts_int8_kernel_impl(
+    scalar_t* __restrict__ output,
+    scalar_t* __restrict__ ic1,
+    scalar_t* __restrict__ ic2,
+    uint8_t* __restrict__ A_tmp,
+    float* __restrict__ C_tmp,
+    uint8_t* __restrict__ Aq_tmp,
+    float* __restrict__ As_tmp,
+    const scalar_t* __restrict__ input,
+    const int8_t* __restrict__ packed_w1,
+    const int8_t* __restrict__ packed_w2,
+    const float* __restrict__ w1s,
+    const float* __restrict__ w2s,
+    const float* __restrict__ topk_weights,
+    const int32_t* __restrict__ sorted_ids,
+    const int32_t* __restrict__ expert_ids,
+    const int32_t* __restrict__ offsets,
+    int64_t M,
+    int64_t N,
+    int64_t K,
+    int64_t E,
+    int64_t topk,
+    int64_t num_tokens_post_pad);
+
+// moe implementations for fp8 w8a16
+template <typename scalar_t>
+void fused_experts_fp8_kernel_impl(
+    scalar_t* __restrict__ output,
+    scalar_t* __restrict__ ic0,
+    scalar_t* __restrict__ ic1,
+    scalar_t* __restrict__ ic2,
+    scalar_t* __restrict__ A_tmp,
+    scalar_t* __restrict__ B_tmp,
+    float* __restrict__ C_tmp,
+    const scalar_t* __restrict__ input,
+    const at::Float8_e4m3fn* __restrict__ packed_w1,
+    const at::Float8_e4m3fn* __restrict__ packed_w2,
+    const float* __restrict__ w1s,
+    const float* __restrict__ w2s,
+    int64_t block_size_N,
+    int64_t block_size_K,
+    const float* __restrict__ topk_weights,
+    const int32_t* __restrict__ sorted_ids,
+    const int32_t* __restrict__ expert_ids,
+    const int32_t* __restrict__ offsets,
+    int64_t M,
+    int64_t N,
+    int64_t K,
+    int64_t E,
+    int64_t topk,
+    int64_t num_tokens_post_pad);
+
+// moe implementations for int4 w4a16
+template <typename scalar_t>
+void fused_experts_int4_w4a16_kernel_impl(
+    scalar_t* __restrict__ output,
+    scalar_t* __restrict__ ic0,
+    scalar_t* __restrict__ ic1,
+    scalar_t* __restrict__ ic2,
+    scalar_t* __restrict__ A_tmp,
+    scalar_t* __restrict__ B_tmp,
+    float* __restrict__ C_tmp,
+    const scalar_t* __restrict__ input,
+    const at::quint4x2* __restrict__ packed_w1,
+    const at::quint4x2* __restrict__ packed_w2,
+    const uint8_t* __restrict__ w1z,
+    const uint8_t* __restrict__ w2z,
+    const scalar_t* __restrict__ w1s,
+    const scalar_t* __restrict__ w2s,
+    int group_size,
+    const float* __restrict__ topk_weights,
+    const int32_t* __restrict__ sorted_ids,
+    const int32_t* __restrict__ expert_ids,
+    const int32_t* __restrict__ offsets,
+    int64_t M,
+    int64_t N,
+    int64_t K,
+    int64_t E,
+    int64_t topk,
+    int64_t num_tokens_post_pad);
+
+// shared expert implementation for int8 w8a8
+template <typename scalar_t>
+void shared_expert_int8_kernel_impl(
+    scalar_t* __restrict__ output,
+    scalar_t* __restrict__ ic1,
+    float* __restrict__ C_tmp,
+    uint8_t* __restrict__ Aq_tmp,
+    float* __restrict__ As_tmp,
+    const scalar_t* __restrict__ input,
+    const int8_t* __restrict__ packed_w1,
+    const int8_t* __restrict__ packed_w2,
+    const float* __restrict__ w1s,
+    const float* __restrict__ w2s,
+    const scalar_t* __restrict__ fused_experts_out,
+    float routed_scaling_factor,
+    int64_t M,
+    int64_t N,
+    int64_t K);
+
+template <typename scalar_t>
+void shared_expert_fp8_kernel_impl(
+    scalar_t* __restrict__ output,
+    scalar_t* __restrict__ ic0,
+    scalar_t* __restrict__ ic1,
+    scalar_t* __restrict__ B_tmp,
+    float* __restrict__ C_tmp,
+    const scalar_t* __restrict__ input,
+    const at::Float8_e4m3fn* __restrict__ packed_w1,
+    const at::Float8_e4m3fn* __restrict__ packed_w2,
+    const float* __restrict__ w1s,
+    const float* __restrict__ w2s,
+    int64_t block_size_N,
+    int64_t block_size_K,
+    const scalar_t* __restrict__ fused_experts_out,
+    float routed_scaling_factor,
+    int64_t M,
+    int64_t N,
+    int64_t K);
+
+// tinygemm interface
+template <typename scalar_t>
+void tinygemm_kernel(
+    const scalar_t* __restrict__ A,
+    const scalar_t* __restrict__ B,
+    scalar_t* __restrict__ C,
+    float* __restrict__ Ctmp,
+    int64_t M,
+    int64_t N,
+    int64_t K,
+    int64_t lda,
+    int64_t ldb,
+    int64_t ldc,
+    bool brg);
+
+template <typename scalar_t>
+void tinygemm_kernel(
+    const uint8_t* __restrict__ A,
+    const int8_t* __restrict__ B,
+    scalar_t* __restrict__ C,
+    int32_t* __restrict__ Ctmp,
+    const float* __restrict__ As,
+    const float* __restrict__ Bs,
+    int64_t M,
+    int64_t N,
+    int64_t K,
+    int64_t lda,
+    int64_t ldb,
+    int64_t ldc,
+    bool brg);
+
+template <typename scalar_t>
+void tinygemm_kernel(
+    const scalar_t* __restrict__ A,
+    const at::Float8_e4m3fn* __restrict__ B,
+    scalar_t* __restrict__ C,
+    scalar_t* __restrict__ Btmp,
+    float* __restrict__ Ctmp,
+    const float* __restrict__ scale,
+    int64_t M,
+    int64_t N,
+    int64_t K,
+    int64_t lda,
+    int64_t ldb,
+    int64_t ldc,
+    bool brg,
+    int64_t block_size_K);
+
+template <typename scalar_t>
+void tinygemm_kernel(
+    const scalar_t* __restrict__ A,
+    const at::quint4x2* __restrict__ B,
+    scalar_t* __restrict__ C,
+    const uint8_t* __restrict__ Bz,
+    const scalar_t* __restrict__ Bs,
+    scalar_t* __restrict__ Btmp,
+    float* __restrict__ Ctmp,
+    int64_t M,
+    int64_t N,
+    int64_t K,
+    int group_size,
+    int64_t lda,
+    int64_t ldb,
+    int64_t ldc,
+    int64_t strideBz,
+    int64_t strideBs,
+    bool brg);
+
+// TODO: debug print, remove me later
+inline void print_16x32i(const __m512i x) {
+  int32_t a[16];
+  _mm512_storeu_si512((__m512i *)a, x);
+
+  for (int i = 0; i < 16; i++){
+    std::cout << a[i] << " ";
+  }
+  std::cout << std::endl;
+}
+
+inline void print_16x32(const __m512 x) {
+  float a[16];
+  _mm512_storeu_ps((__m512 *)a, x);
+
+  for (int i = 0; i < 16; i++){
+    std::cout << a[i] << " ";
+  }
+  std::cout << std::endl;
+}
+
+
+inline void print_32x8u(const __m256i x) {
+  uint8_t a[32];
+  _mm256_storeu_si256((__m256i *)a, x);
+
+  for (int i = 0; i < 32; ++i) {
+    std::cout << int32_t(a[i]) << " ";
+  }
+  std::cout << std::endl;
+}

csrc/cpu/sgl-kernels/gemm_fp8.cpp

@@ -0,0 +1,530 @@
+// Adapted from
+// https://github.com/sgl-project/sglang/tree/main/sgl-kernel/csrc/cpu
+
+#include "common.h"
+#include "vec.h"
+#include "gemm.h"
+
+// clang-format off
+
+// we use 4x32 for BLOCK_M
+#define BLOCK_SIZE_M_SCALE 4
+
+namespace {
+
+template <typename scalar_t>
+inline void copy_stub(scalar_t* __restrict__ out, const float* __restrict__ input, int64_t size) {
+  using bVec = at::vec::Vectorized<scalar_t>;
+  using fVec = at::vec::Vectorized<float>;
+  constexpr int kVecSize = bVec::size();
+
+  int64_t d;
+  #pragma GCC unroll 4
+  for (d = 0; d <= size - kVecSize; d += kVecSize) {
+    fVec data0 = fVec::loadu(input + d);
+    fVec data1 = fVec::loadu(input + d + fVec::size());
+    bVec out_vec = convert_from_float_ext<scalar_t>(data0, data1);
+    out_vec.store(out + d);
+  }
+  for (; d < size; ++d) {
+    out[d] = static_cast<scalar_t>(input[d]);
+  }
+}
+
+template <typename scalar_t>
+inline void copy_add_stub(scalar_t* __restrict__ out, const float* __restrict__ input, const float* __restrict__ bias, int64_t size) {
+  using bVec = at::vec::Vectorized<scalar_t>;
+  using fVec = at::vec::Vectorized<float>;
+  constexpr int kVecSize = bVec::size();
+
+  int64_t d;
+  #pragma GCC unroll 4
+  for (d = 0; d <= size - kVecSize; d += kVecSize) {
+    fVec data0 = fVec::loadu(input + d) + fVec::loadu(bias + d);
+    fVec data1 = fVec::loadu(input + d + fVec::size()) + fVec::loadu(bias + d + fVec::size());
+    bVec out_vec = convert_from_float_ext<scalar_t>(data0, data1);
+    out_vec.store(out + d);
+  }
+  for (; d < size; ++d) {
+    out[d] = static_cast<scalar_t>(input[d] + bias[d]);
+  }
+}
+
+inline void unpack_B(
+    at::BFloat16* __restrict__ Btmp,
+    const at::Float8_e4m3fn* __restrict__ packed_B,
+    int N,
+    int K,
+    int ldb,
+    int ldb_tmp,
+    float scale) {
+#if defined(CPU_CAPABILITY_AVX512)
+  // [K/2, N, 2]
+  const int K2 = K >> 1;
+  const int ldb2 = ldb; // ldb * 2 >> 1;
+  const uint16_t* b_ptr = reinterpret_cast<const uint16_t*>(packed_B);
+  const __m512 vd = _mm512_set1_ps(scale);
+
+  constexpr int BLOCK_N = block_size_n();
+  static_assert(BLOCK_N == 32);
+
+  // prefetch distance
+  constexpr int PREFETCH_SIZE_K = 64;
+
+#pragma GCC unroll 4
+  for (int k = 0; k < K2; ++k) {
+    __m512i b8 = _mm512_loadu_si512(b_ptr + k * ldb2);
+    if constexpr (PREFETCH_SIZE_K > 0) {
+      _mm_prefetch(b_ptr + (k + PREFETCH_SIZE_K) * ldb2, _MM_HINT_T0);
+    }
+
+    __m256i b8_0 = _mm512_extracti32x8_epi32(b8, 0);
+    __m256i b8_1 = _mm512_extracti32x8_epi32(b8, 1);
+
+    __m512bh bf16_0 = CVT_FP8_TO_BF16(b8_0);
+    __m512bh bf16_1 = CVT_FP8_TO_BF16(b8_1);
+
+    // Apply scale
+    __m512 f0_lo = CVT_BF16_TO_FP32(_mm512_extracti32x8_epi32((__m512i)bf16_0, 0));
+    __m512 f0_hi = CVT_BF16_TO_FP32(_mm512_extracti32x8_epi32((__m512i)bf16_0, 1));
+    __m512 f1_lo = CVT_BF16_TO_FP32(_mm512_extracti32x8_epi32((__m512i)bf16_1, 0));
+    __m512 f1_hi = CVT_BF16_TO_FP32(_mm512_extracti32x8_epi32((__m512i)bf16_1, 1));
+
+    f0_lo = _mm512_mul_ps(f0_lo, vd);
+    f0_hi = _mm512_mul_ps(f0_hi, vd);
+    f1_lo = _mm512_mul_ps(f1_lo, vd);
+    f1_hi = _mm512_mul_ps(f1_hi, vd);
+
+    bf16_0 = _mm512_cvtne2ps_pbh(f0_hi, f0_lo);
+    bf16_1 = _mm512_cvtne2ps_pbh(f1_hi, f1_lo);
+
+    _mm512_storeu_si512(Btmp + k * ldb_tmp * 2 + 0, (__m512i)bf16_0);
+    _mm512_storeu_si512(Btmp + k * ldb_tmp * 2 + 32, (__m512i)bf16_1);
+  }
+#else
+  TORCH_CHECK(false, "unpack_B: scalar path not implemented!");
+#endif
+}
+
+template <typename scalar_t, typename packed_t, bool has_bias, int BLOCK_M, int BLOCK_N>
+struct tinygemm_kernel_nn {
+  static inline void apply(
+      const scalar_t* __restrict__ A, const packed_t* __restrict__ B, scalar_t* __restrict__ C,
+      const float* __restrict__ bias, const float* __restrict__ scale, int K, int lda, int ldb, int ldc, int64_t block_size_K) {
+    TORCH_CHECK(false, "tinygemm_kernel_nn: scalar path not implemented!");
+  }
+};
+
+#if defined(CPU_CAPABILITY_AVX512)
+template <bool has_bias, int BLOCK_M, int BLOCK_N>
+struct tinygemm_kernel_nn<at::BFloat16, at::Float8_e4m3fn, has_bias, BLOCK_M, BLOCK_N> {
+  static inline void apply(
+      const at::BFloat16* __restrict__ A, const at::Float8_e4m3fn* __restrict__ B, at::BFloat16* __restrict__ C,
+      const float* __restrict__ bias, const float* __restrict__ scale, int K, int lda, int ldb, int ldc, int64_t block_size_K) {
+
+    constexpr int ROWS = BLOCK_M;
+    constexpr int COLS = BLOCK_N / 16;
+
+    const int KB = div_up(K, BLOCK_K);
+
+    // prefetch distance
+    constexpr int PREFETCH_SIZE_K = 64;
+    constexpr int PREFETCH_SIZE_KB = 1;
+
+    __m512bh va;
+    __m512bh vb[COLS];
+    __m512 vc[ROWS * COLS];
+    __m512 vsum[ROWS * COLS];
+
+    // block quant scale
+    __m512 vscale;
+
+    auto loadc = [&](auto i) {
+      constexpr int col = i % COLS;
+      if constexpr (has_bias) {
+        vc[i] = _mm512_loadu_ps(bias + col * 16);
+      } else {
+        vc[i] = _mm512_setzero_ps();
+      }
+    };
+    Unroll<ROWS * COLS>{}(loadc);
+
+    const int lda2 = lda >> 1;
+    const int ldb2 = ldb; // ldb * 2 >> 1;
+    const float* a_ptr = reinterpret_cast<const float*>(A);
+    const uint16_t* b_ptr = reinterpret_cast<const uint16_t*>(B);
+
+    auto compute = [&](auto i, int k) {
+      constexpr int row = i / COLS;
+      constexpr int col = i % COLS;
+
+      if constexpr (col == 0) {
+        va = (__m512bh)(_mm512_set1_ps(a_ptr[row * lda2 + k]));
+        if constexpr (PREFETCH_SIZE_K > 0) {
+          _mm_prefetch(a_ptr + row * lda2 + k + PREFETCH_SIZE_K, _MM_HINT_T0);
+        }
+      }
+      if constexpr (row == 0) {
+        if constexpr (col % 2 == 0) {
+          __m512i b8 = _mm512_loadu_si512(b_ptr + k * ldb2 + col * 16);
+          if constexpr (PREFETCH_SIZE_K > 0) {
+            _mm_prefetch(b_ptr + (k + PREFETCH_SIZE_K) * ldb2 + col * 16, _MM_HINT_T0);
+          }
+          vb[col + 0] = CVT_FP8_TO_BF16(_mm512_extracti32x8_epi32(b8, 0));
+          vb[col + 1] = CVT_FP8_TO_BF16(_mm512_extracti32x8_epi32(b8, 1));
+        }
+      }
+      vsum[i] = _mm512_dpbf16_ps(vsum[i], va, vb[col]);
+    };
+
+    constexpr int BLOCK_K2 = BLOCK_K >> 1;
+    for (int kb = 0; kb < KB; ++kb) {
+      int kb_start = kb * BLOCK_K2;
+      int kb_end = std::min(K, kb_start + BLOCK_K2);
+      // 1. load scale vector
+      vscale = _mm512_set1_ps(scale[kb]);
+      if constexpr (PREFETCH_SIZE_KB > 0) {
+        _mm_prefetch(scale + kb + PREFETCH_SIZE_KB, _MM_HINT_T0);
+      }
+      // 2. zero vsum for each block
+      Unroll<ROWS * COLS>{}([&](auto i) {
+        vsum[i] = _mm512_setzero_ps();
+      });
+      // 3. accumulate across each block
+      for (int k = kb_start; k < kb_end; ++k) {
+        Unroll<ROWS * COLS>{}(compute, k);
+      }
+      // 4. apply scale
+      Unroll<ROWS * COLS>{}([&](auto i) {
+        vc[i] = _mm512_fmadd_ps(vsum[i], vscale, vc[i]);
+      });
+    }
+
+    auto storec = [&](auto i) {
+      constexpr int row = i / COLS;
+      constexpr int col = i % COLS;
+      // for COLS = 2,4 use 512bit store
+      if constexpr (col % 2 == 0) {
+        _mm512_storeu_si512(
+            reinterpret_cast<__m512i*>((C + row * ldc + col * 16)),
+            (__m512i)(_mm512_cvtne2ps_pbh(vc[row * COLS + col + 1], vc[row * COLS + col])));
+      }
+    };
+    Unroll<ROWS * COLS>{}(storec);
+  }
+};
+#endif
+
+#define LAUNCH_TINYGEMM_KERNEL_NN(MB_SIZE, NB_SIZE)                          \
+    tinygemm_kernel_nn<scalar_t, at::Float8_e4m3fn, has_bias, MB_SIZE, NB_SIZE>::apply(         \
+        A + mb_start * lda, B + nb_start * 2, C + mb_start * ldc + nb_start, \
+        has_bias ? bias + nb_start : nullptr, scale, K, lda, ldb, ldc, block_size_K);
+
+template <typename scalar_t, typename packed_t, bool has_bias>
+struct brgemm {
+  static inline void apply(
+      const scalar_t* __restrict__ A,
+      const packed_t* __restrict__ B,
+      scalar_t* __restrict__ C,
+      scalar_t* __restrict__ Btmp,
+      float* __restrict__ Ctmp,
+      const float* __restrict__ bias,
+      const float* __restrict__ scale,
+      int M,
+      int N,
+      int K,
+      int lda,
+      int ldb,
+      int ldc) {
+    TORCH_CHECK(false, "struct brgemm: primary template not implemented!");
+  }
+};
+
+template <bool has_bias>
+struct brgemm<at::BFloat16, at::Float8_e4m3fn, has_bias> {
+  static inline void apply(
+      const at::BFloat16* __restrict__ A,
+      const at::Float8_e4m3fn* __restrict__ B,
+      at::BFloat16* __restrict__ C,
+      at::BFloat16* __restrict__ Btmp,
+      float* __restrict__ Ctmp,
+      const float* __restrict__ bias,
+      const float* __restrict__ scale,
+      int M,
+      int N,
+      int K,
+      int lda,
+      int ldb,
+      int ldc) {
+
+    constexpr int BLOCK_N = block_size_n();
+
+    // [K, BLOCK_N] -> [K / 2, BLOCK_N * 2]
+    const int ldb_tmp = BLOCK_N;
+
+    for (int k = 0; k < K; k += BLOCK_K) {
+      int kb_size = std::min(BLOCK_K, K - k);
+
+      int idx = k >> 7; // k / BLOCK_K where BLOCK_K = 128
+      unpack_B(Btmp + k * ldb_tmp, B + k * ldb, N, kb_size, ldb, ldb_tmp, scale[idx]);
+    }
+
+    at::native::cpublas::brgemm(
+        M, N, K, lda, ldb_tmp, BLOCK_N, /* add_C */ false, A, Btmp, Ctmp);
+
+    // copy from Ctmp to C
+    for (int m = 0; m < M; ++m) {
+      if constexpr (has_bias) {
+        copy_add_stub(C + m * ldc, Ctmp + m * BLOCK_N, bias, N);
+      } else {
+        copy_stub(C + m * ldc, Ctmp + m * BLOCK_N, N);
+      }
+    }
+  }
+};
+
+template <typename scalar_t, bool has_bias>
+void tinygemm_kernel(
+    const scalar_t* __restrict__ A,
+    const at::Float8_e4m3fn* __restrict__ B,
+    scalar_t* __restrict__ C,
+    scalar_t* __restrict__ Btmp,
+    float* __restrict__ Ctmp,
+    const float* __restrict__ scale,
+    const float* __restrict__ bias,
+    int64_t M,
+    int64_t N,
+    int64_t K,
+    int64_t lda,
+    int64_t ldb,
+    int64_t ldc,
+    bool brg,
+    int64_t block_size_K) {
+
+  if (brg) {
+    brgemm<scalar_t, at::Float8_e4m3fn, has_bias>::apply(
+        A, B, C, Btmp, Ctmp, bias, scale, M, N, K, lda, ldb, ldc);
+    return;
+  }
+
+  // pattern: 1-4-16
+  constexpr int64_t BLOCK_M = 4;
+  constexpr int64_t BLOCK_N = 64;
+  const int64_t MB = div_up(M, BLOCK_M);
+  const int64_t NB = div_up(N, BLOCK_N);
+  for (int mb = 0; mb < MB; ++mb) {
+    int64_t mb_start = mb * BLOCK_M;
+    int64_t mb_size = std::min(BLOCK_M, M - mb_start);
+    for (int64_t nb = 0; nb < NB; ++nb) {
+      int64_t nb_start = nb * BLOCK_N;
+      int64_t nb_size = std::min(BLOCK_N, N - nb_start);
+
+      switch(mb_size << 4 | nb_size >> 4) {
+        case 0x12: LAUNCH_TINYGEMM_KERNEL_NN(1, 32); break;
+        case 0x22: LAUNCH_TINYGEMM_KERNEL_NN(2, 32); break;
+        case 0x32: LAUNCH_TINYGEMM_KERNEL_NN(3, 32); break;
+        case 0x42: LAUNCH_TINYGEMM_KERNEL_NN(4, 32); break;
+        default: TORCH_CHECK(false, "Unexpected block size, ", mb_size, "x", "nb_size");
+      }
+    }
+  }
+}
+
+template <typename scalar_t>
+void fp8_scaled_mm_kernel_impl(
+    scalar_t* __restrict__ out,
+    const scalar_t* __restrict__ mat1,
+    const at::Float8_e4m3fn* __restrict__ mat2,
+    const float* __restrict__ scales2,
+    const float* __restrict__ bias,
+    scalar_t* __restrict__ buffer,
+    int64_t M,
+    int64_t N,
+    int64_t K,
+    int64_t mat1_strideM,
+    int64_t out_strideM,
+    int64_t block_size_N,
+    int64_t block_size_K,
+    int64_t buffer_size_per_thread) {
+
+  constexpr int64_t BLOCK_M = block_size_m() * BLOCK_SIZE_M_SCALE;
+  constexpr int64_t BLOCK_N = block_size_n();
+  const int64_t MB = div_up(M, BLOCK_M);
+  const int64_t NB = div_up(N, BLOCK_N);
+
+  const int64_t scale_size_K = div_up(K, block_size_K);
+  const int64_t blocks_n_per_group = block_size_N / BLOCK_N;
+
+  const bool use_brgemm = can_use_brgemm<at::Float8_e4m3fn>(M);
+
+  // parallel on [MB, NB]
+  AT_DISPATCH_BOOL(bias != nullptr, has_bias, [&] {
+    at::parallel_for(0, MB * NB, 0, [&](int64_t begin, int64_t end) {
+      int64_t mb{0}, nb{0};
+      data_index_init(begin, mb, MB, nb, NB);
+
+      int tid = at::get_thread_num();
+      scalar_t* __restrict__ Btmp = buffer + tid * buffer_size_per_thread;
+      float* __restrict__ Ctmp = (float*)((void*)(Btmp + BLOCK_N * K));
+
+      for (int64_t i = begin; i < end; ++i) {
+        UNUSED(i);
+        const float* scale_ptr = scales2 + (nb / blocks_n_per_group) * scale_size_K;
+
+        int64_t mb_start = mb * BLOCK_M;
+        int64_t mb_size = std::min(M - mb_start, BLOCK_M);
+        int64_t nb_start = nb * BLOCK_N;
+        int64_t nb_size = std::min(N - nb_start, BLOCK_N);
+
+        tinygemm_kernel<scalar_t, has_bias>(
+            /*   A            */ mat1 + mb_start * mat1_strideM,
+            /*   B            */ mat2 + nb_start * K, // nb * BLOCK_N * K
+            /*   C            */ out + mb_start * out_strideM + nb_start,
+            /*   Btmp         */ Btmp,
+            /*   Ctmp         */ Ctmp,
+            /*   scale        */ scale_ptr,
+            /*   bias         */ bias + nb_start,
+            /*   M            */ mb_size,
+            /*   N            */ nb_size,
+            /*   K            */ K,
+            /*   lda          */ mat1_strideM,
+            /*   ldb          */ nb_size,
+            /*   ldc          */ out_strideM,
+            /*   brg          */ use_brgemm,
+            /*   block_size_K */ block_size_K);
+
+        // move to the next index
+        data_index_step(mb, MB, nb, NB);
+      }
+
+      if (use_brgemm) {
+        at::native::cpublas::brgemm_release();
+      }
+    });
+  });
+}
+
+} // anonymous namespace
+
+// tinygemm interface
+template <typename scalar_t>
+void tinygemm_kernel(
+    const scalar_t* __restrict__ A,
+    const at::Float8_e4m3fn* __restrict__ B,
+    scalar_t* __restrict__ C,
+    scalar_t* __restrict__ Btmp,
+    float* __restrict__ Ctmp,
+    const float* __restrict__ scale,
+    int64_t M,
+    int64_t N,
+    int64_t K,
+    int64_t lda,
+    int64_t ldb,
+    int64_t ldc,
+    bool brg,
+    int64_t block_size_K) {
+  tinygemm_kernel<scalar_t, false>(A, B, C, Btmp, Ctmp, scale, nullptr, M, N, K, lda, ldb, ldc, brg, block_size_K);
+}
+
+#define INSTANTIATE_TINYGEMM_TEMPLATE(TYPE)    \
+  template void tinygemm_kernel<TYPE>(         \
+      const TYPE* __restrict__ A,              \
+      const at::Float8_e4m3fn* __restrict__ B, \
+      TYPE* __restrict__ C,                    \
+      TYPE* __restrict__ Btmp,                 \
+      float* __restrict__ Ctmp,                \
+      const float* __restrict__ scale,         \
+      int64_t M,                               \
+      int64_t N,                               \
+      int64_t K,                               \
+      int64_t lda,                             \
+      int64_t ldb,                             \
+      int64_t ldc,                             \
+      bool brg,                                \
+      int64_t block_size_K)
+
+INSTANTIATE_TINYGEMM_TEMPLATE(at::BFloat16);
+INSTANTIATE_TINYGEMM_TEMPLATE(at::Half);
+
+at::Tensor fp8_scaled_mm_cpu(at::Tensor& mat1, at::Tensor& mat2, at::Tensor& scales2,
+    std::vector<int64_t> block_size, std::optional<at::Tensor>& bias,
+    at::ScalarType out_dtype, bool is_vnni) {
+  RECORD_FUNCTION("sgl-kernel::fp8_scaled_mm_cpu", std::vector<c10::IValue>({mat1, mat2, scales2, block_size, bias}));
+
+  auto packed_w = is_vnni ? mat2 : convert_weight_packed(mat2);
+
+  CHECK_LAST_DIM_CONTIGUOUS_INPUT(mat1);
+  CHECK_INPUT(mat2);
+  CHECK_INPUT(scales2);
+  TORCH_CHECK(scales2.scalar_type() == at::kFloat,
+      "fp8_scaled_mm_cpu: expect scales2 to be float32.");
+
+  int64_t M = mat1.size(0);
+  int64_t N = mat2.size(0);
+  int64_t K = mat2.size(1);
+
+  CHECK_EQ(mat1.size(1), K);
+  CHECK_DIM(2, mat1);
+  CHECK_DIM(2, mat2);
+
+  TORCH_CHECK(block_size.size() == 2,
+      "fp8_scaled_mm_cpu: expect block_size.size() to be 2.");
+
+  int64_t block_size_N = block_size[0];
+  int64_t block_size_K = block_size[1];
+
+  constexpr int64_t BLOCK_M = block_size_m() * BLOCK_SIZE_M_SCALE;
+  constexpr int64_t BLOCK_N = block_size_n();
+  TORCH_CHECK(block_size_N % BLOCK_N == 0, "fp8_scaled_mm_cpu: expect block_size_N to be multiples of BLOCK_N");
+  TORCH_CHECK(block_size_K == BLOCK_K, "fp8_scaled_mm_cpu: expect block_size_K equals to BLOCK_K");
+  CHECK_EQ(scales2.size(0), div_up(N, block_size_N));
+  CHECK_EQ(scales2.size(1), div_up(K, block_size_K));
+
+  const auto st = mat1.scalar_type();
+  TORCH_CHECK(st == at::kBFloat16 || st == at::kHalf,
+      "fp8_scaled_mm_cpu: expect A to be bfloat16 or half.");
+  TORCH_CHECK(st == out_dtype,
+      "fp8_scaled_mm_cpu: expect A has same dtype with out_dtype.");
+  TORCH_CHECK(mat2.scalar_type() == at::kFloat8_e4m3fn,
+      "fp8_scaled_mm_cpu: expect mat2 to be fp8_e4m3.");
+  TORCH_CHECK(scales2.scalar_type() == at::kFloat,
+      "fp8_scaled_mm_cpu: expect scales to be float32.");
+  auto out = at::empty({M, N}, mat1.options().dtype(out_dtype));
+
+  // strides
+  int64_t mat1_strideM = mat1.stride(0);
+  int64_t out_strideM = out.stride(0);
+
+  const bool has_bias = bias.has_value();
+  const float* bias_data = nullptr;
+  if (has_bias) {
+    CHECK_EQ(bias.value().size(0), N);
+    bias_data = bias.value().data_ptr<float>();
+  }
+
+  // Btmp : [T, BLOCK_N * K]
+  // Ctmp : [T, BLOCK_M * BLOCK_N]
+  int num_threads = at::get_num_threads();
+  int64_t size_per_thread = BLOCK_N * K + BLOCK_M * BLOCK_N * 2;
+  auto buffer = at::empty({num_threads, size_per_thread}, mat1.options());
+
+  AT_DISPATCH_REDUCED_FLOATING_TYPES(out_dtype, "fp8_scaled_mm_kernel_impl", [&] {
+    fp8_scaled_mm_kernel_impl<scalar_t>(
+        out.data_ptr<scalar_t>(),
+        mat1.data_ptr<scalar_t>(),
+        packed_w.data_ptr<at::Float8_e4m3fn>(),
+        scales2.data_ptr<float>(),
+        bias_data,
+        buffer.data_ptr<scalar_t>(),
+        M,
+        N,
+        K,
+        mat1_strideM,
+        out_strideM,
+        block_size_N,
+        block_size_K,
+        size_per_thread);
+  });
+
+  return out;
+}

csrc/cpu/sgl-kernels/gemm_int8.cpp

@@ -0,0 +1,440 @@
+// Adapted from
+// https://github.com/sgl-project/sglang/tree/main/sgl-kernel/csrc/cpu
+
+#include "common.h"
+#include "vec.h"
+#include "gemm.h"
+
+// clang-format off
+
+namespace {
+
+template <typename scalar_t, bool has_bias, int BLOCK_M, int BLOCK_N>
+struct tinygemm_kernel_nn {
+  static inline void apply(
+      const uint8_t* __restrict__ A, const int8_t* __restrict__ B, scalar_t* __restrict__ C,
+      const float* __restrict__ As, const float* __restrict__ Bs, const int32_t* __restrict__ Bcomp,
+      const float* __restrict__ bias, int64_t K, int64_t lda, int64_t ldb, int64_t ldc) {
+    TORCH_CHECK(false, "tinygemm_kernel_nn: scalar path not implemented!");
+  }
+};
+
+#if defined(CPU_CAPABILITY_AVX512)
+template <bool has_bias, int BLOCK_M, int BLOCK_N>
+struct tinygemm_kernel_nn<at::BFloat16, has_bias, BLOCK_M, BLOCK_N> {
+  static inline void apply(
+      const uint8_t* __restrict__ A, const int8_t* __restrict__ B, at::BFloat16* __restrict__ C,
+      const float* __restrict__ As, const float* __restrict__ Bs, const int32_t* __restrict__ Bcomp,
+      const float* __restrict__ bias, int64_t K, int64_t lda, int64_t ldb, int64_t ldc) {
+
+    constexpr int ROWS = BLOCK_M;
+    constexpr int COLS = BLOCK_N / 16;
+    static_assert(COLS % 2 == 0);
+
+    // prefetch distance
+    constexpr int PREFETCH_SIZE_K = 0;
+
+    __m512i va;
+    __m512i vb[COLS];
+    __m512i vc[ROWS * COLS];
+    __m512i vcomp[COLS];
+    __m512  vd0;
+    __m512  vd1[COLS];
+
+    // oops! 4x4 spills but luckily we use 4x2
+    __m512 vbias[COLS];
+
+    // [NOTE]: s8s8 igemm compensation in avx512-vnni
+    //
+    // avx512-vnni has no s8s8, so we need to change s8s8 to u8s8 with compensate:
+    //
+    //   a * b = (a + 128) * b - 128 * b
+    //   s   s       u       s    u    s
+    //
+    // 1) 128 * b is pre-computed when packing B to vnni formats
+    // 2) a + 128 is fused when dynamically quantize A
+    //
+    auto loadc = [&](auto i) {
+      vc[i] = _mm512_set1_epi32(0);
+    };
+    Unroll<ROWS * COLS>{}(loadc);
+
+    const int64_t K4 = K >> 2;
+    const int64_t lda4 = lda >> 2;
+    const int64_t ldb4 = ldb; // ldb * 4 >> 2;
+    const int32_t* a_ptr = reinterpret_cast<const int32_t*>(A);
+    const int32_t* b_ptr = reinterpret_cast<const int32_t*>(B);
+
+    auto compute = [&](auto i, int64_t k) {
+      constexpr int row = i / COLS;
+      constexpr int col = i % COLS;
+
+      if constexpr (col == 0) {
+        va = _mm512_set1_epi32(a_ptr[row * lda4 + k]);
+      }
+      if constexpr (row == 0) {
+        vb[col] = _mm512_loadu_si512(b_ptr + k * ldb4 + col * 16);
+        if constexpr (PREFETCH_SIZE_K > 0) {
+          _mm_prefetch(b_ptr + (k + PREFETCH_SIZE_K) * ldb4 + col * 16, _MM_HINT_T0);
+        }
+      }
+      vc[i] = _mm512_dpbusd_epi32(vc[i], va, vb[col]);
+    };
+    for (int64_t k = 0; k < K4; ++k) {
+      Unroll<ROWS * COLS>{}(compute, k);
+    }
+
+    auto storec = [&](auto i) {
+      constexpr int row = i / COLS;
+      constexpr int col = i % COLS;
+
+      // load a scale
+      if constexpr(col == 0) {
+        vd0 = _mm512_set1_ps(As[row]);
+      }
+      // load b scale and vcomp per 2 vectors
+      // also load bias if any
+      if constexpr (row == 0) {
+        if constexpr (col % 2 == 0) {
+          vd1[col + 0] = _mm512_loadu_ps(Bs + col * 16);
+          vd1[col + 1] = _mm512_loadu_ps(Bs + col * 16 + 16);
+          vcomp[col + 0] = _mm512_loadu_si512(Bcomp + col * 16);
+          vcomp[col + 1] = _mm512_loadu_si512(Bcomp + col * 16 + 16);
+          if constexpr (has_bias) {
+            vbias[col + 0] = _mm512_loadu_ps(bias + col * 16);
+            vbias[col + 1] = _mm512_loadu_ps(bias + col * 16 + 16);
+          }
+        }
+      }
+
+      // for COLS = 2, 4 use 512bit store
+      if constexpr (col % 2 == 0) {
+        __m512 vc0 = _mm512_cvtepi32_ps(_mm512_sub_epi32(vc[row * COLS + col + 0], vcomp[col + 0]));
+        __m512 vc1 = _mm512_cvtepi32_ps(_mm512_sub_epi32(vc[row * COLS + col + 1], vcomp[col + 1]));
+        if constexpr (has_bias) {
+          vc0 = _mm512_fmadd_ps(_mm512_mul_ps(vc0, vd0), vd1[col + 0], vbias[col + 0]);
+          vc1 = _mm512_fmadd_ps(_mm512_mul_ps(vc1, vd0), vd1[col + 1], vbias[col + 1]);
+        } else {
+          vc0 = _mm512_mul_ps(_mm512_mul_ps(vc0, vd0), vd1[col + 0]);
+          vc1 = _mm512_mul_ps(_mm512_mul_ps(vc1, vd0), vd1[col + 1]);
+        }
+
+        _mm512_storeu_si512(
+            reinterpret_cast<__m512i*>((C + row * ldc + col * 16)),
+            (__m512i)(_mm512_cvtne2ps_pbh(vc1, vc0)));
+      }
+    };
+    Unroll<ROWS * COLS>{}(storec);
+  }
+};
+#endif
+
+#define LAUNCH_TINYGEMM_KERNEL_NN(MB_SIZE, NB_SIZE)                          \
+    tinygemm_kernel_nn<scalar_t, has_bias, MB_SIZE, NB_SIZE>::apply(         \
+        A + mb_start * lda, B + nb_start * 4, C + mb_start * ldc + nb_start, \
+        As + mb_start, Bs + nb_start, Bcomp + nb_start,                      \
+        has_bias ? bias + nb_start : nullptr, K, lda, ldb, ldc);
+
+template <typename scalar_t, bool has_bias>
+void tinygemm_kernel(
+    const uint8_t* __restrict__ A,
+    const int8_t* __restrict__ B,
+    scalar_t* __restrict__ C,
+    int32_t* __restrict__ Ctmp,
+    const float* __restrict__ As,
+    const float* __restrict__ Bs,
+    const float* __restrict__ bias,
+    int64_t M,
+    int64_t N,
+    int64_t K,
+    int64_t lda,
+    int64_t ldb,
+    int64_t ldc,
+    bool brg) {
+
+  // B compensation
+  const int32_t* Bcomp = reinterpret_cast<const int32_t*>(B + block_size_n() * K);
+
+  // pattern: 1-4-16
+  constexpr int64_t BLOCK_M = 4;
+  constexpr int64_t BLOCK_N = 64;
+  const int64_t MB = div_up(M, BLOCK_M);
+  const int64_t NB = div_up(N, BLOCK_N);
+  for (int64_t mb = 0; mb < MB; ++mb) {
+    int64_t mb_start = mb * BLOCK_M;
+    int64_t mb_size = std::min(BLOCK_M, M - mb_start);
+    for (int64_t nb = 0; nb < NB; ++nb) {
+      int64_t nb_start = nb * BLOCK_N;
+      int64_t nb_size = std::min(BLOCK_N, N - nb_start);
+
+      switch(mb_size << 4 | nb_size >> 4) {
+        // mb_size = 1
+        case 0x12: LAUNCH_TINYGEMM_KERNEL_NN(1, 32); break;
+        case 0x14: LAUNCH_TINYGEMM_KERNEL_NN(1, 64); break;
+        // mb_size = 2
+        case 0x22: LAUNCH_TINYGEMM_KERNEL_NN(2, 32); break;
+        case 0x24: LAUNCH_TINYGEMM_KERNEL_NN(2, 64); break;
+        // mb_size = 3
+        case 0x32: LAUNCH_TINYGEMM_KERNEL_NN(3, 32); break;
+        case 0x34: LAUNCH_TINYGEMM_KERNEL_NN(3, 64); break;
+        // mb_size = 4
+        case 0x42: LAUNCH_TINYGEMM_KERNEL_NN(4, 32); break;
+        case 0x44: LAUNCH_TINYGEMM_KERNEL_NN(4, 64); break;
+        default: TORCH_CHECK(false, "Unexpected block size, ", mb_size, "x", "nb_size");
+      }
+    }
+  }
+}
+
+template<typename scalar_t>
+void int8_scaled_mm_kernel_impl(
+    scalar_t* __restrict__ out,
+    const uint8_t* __restrict__ mat1,
+    const int8_t* __restrict__ mat2,
+    const float* __restrict__ scales1,
+    const float* __restrict__ scales2,
+    const float* __restrict__ bias,
+    int64_t M,
+    int64_t N,
+    int64_t K) {
+
+  constexpr int64_t BLOCK_M = block_size_m();
+  constexpr int64_t BLOCK_N = block_size_n();
+  const int64_t MB = div_up(M, BLOCK_M);
+  const int64_t NB = div_up(N, BLOCK_N);
+
+  // TODO: brgemm u8s8 depends on PyTorch 2.7 release.
+  const bool use_brgemm = false;
+
+  // K + 4 after compensation
+  const int64_t packed_row_size = get_row_size<int8_t>(K);
+
+  AT_DISPATCH_BOOL(bias != nullptr, has_bias, [&] {
+    at::parallel_for(0, MB * NB, 0, [&](int64_t begin, int64_t end) {
+      int64_t mb{0}, nb{0};
+      data_index_init(begin, mb, MB, nb, NB);
+
+      // for brgemm, use int32_t for accumulate
+      alignas(64) int32_t Ctmp[BLOCK_M * BLOCK_N];
+
+      for (int i = begin; i < end; ++i) {
+        UNUSED(i);
+        int mb_start = mb * BLOCK_M;
+        int mb_size = std::min(M - mb_start, BLOCK_M);
+        int nb_start = nb * BLOCK_N;
+        int nb_size = std::min(N - nb_start, BLOCK_N);
+
+        tinygemm_kernel<scalar_t, has_bias>(
+            /*   A */ mat1 + mb_start * K,
+            /*   B */ mat2 + nb_start * packed_row_size /* nb * BLOCK_N * (K + 4) */,
+            /*   C */ out + mb_start * N + nb_start,
+            /* Ctmp*/ Ctmp,
+            /*  As */ scales1 + mb_start,
+            /*  Bs */ scales2 + nb_start,
+            /* bias*/ bias + nb_start,
+            /*   M */ mb_size,
+            /*   N */ nb_size,
+            /*   K */ K,
+            /* lda */ K,
+            /* ldb */ nb_size,
+            /* ldc */ N,
+            /* brg */ use_brgemm);
+
+        // move to the next index
+        data_index_step(mb, MB, nb, NB);
+      }
+
+      if (use_brgemm) {
+        at::native::cpublas::brgemm_release();
+      }
+    });
+  });
+}
+
+} // anonymous namespace
+
+// tinygemm interface
+template <typename scalar_t>
+void tinygemm_kernel(const uint8_t* __restrict__ A, const int8_t* __restrict__ B, scalar_t* __restrict__ C,
+    int32_t* __restrict__ Ctmp,  const float* __restrict__ As, const float* __restrict__ Bs,
+    int64_t M, int64_t N, int64_t K, int64_t lda, int64_t ldb, int64_t ldc, bool brg) {
+  tinygemm_kernel<scalar_t, false>(A, B, C, Ctmp, As, Bs, nullptr, M, N, K, lda, ldb, ldc, brg);
+}
+
+#define INSTANTIATE_TINYGEMM_TEMPLATE(TYPE)                                                     \
+    template void tinygemm_kernel<TYPE>(                                                        \
+        const uint8_t* __restrict__ A, const int8_t* __restrict__ B, TYPE* __restrict__ C,      \
+        int32_t* __restrict__ Ctmp, const float* __restrict__ As, const float* __restrict__ Bs, \
+        int64_t M, int64_t N, int64_t K, int64_t lda, int64_t ldb, int64_t ldc, bool brg)
+
+INSTANTIATE_TINYGEMM_TEMPLATE(at::BFloat16);
+INSTANTIATE_TINYGEMM_TEMPLATE(at::Half);
+
+std::tuple<at::Tensor, at::Tensor> per_token_quant_int8_cpu(at::Tensor& A) {
+  RECORD_FUNCTION("sgl-kernel::per_token_quant_int8_cpu", std::vector<c10::IValue>({A}));
+
+  CHECK_LAST_DIM_CONTIGUOUS_INPUT(A);
+  CHECK_DIM(2, A);
+
+  int64_t M = A.size(0);
+  int64_t K = A.size(1);
+  int64_t lda = A.stride(0);
+
+  const auto st = A.scalar_type();
+  TORCH_CHECK(st == at::kBFloat16 || st == at::kHalf,
+      "per_token_quant_int8: expect A to be bfloat16 or half.");
+
+  auto Aq = at::empty({M, K}, A.options().dtype(at::kByte));
+  auto As = at::empty({M}, A.options().dtype(at::kFloat));
+
+  AT_DISPATCH_REDUCED_FLOATING_TYPES(st, "per_token_quant_int8", [&] {
+    uint8_t* __restrict__ Aq_data = Aq.data_ptr<uint8_t>();
+    float* __restrict__ As_data = As.data_ptr<float>();
+    const scalar_t* __restrict__ A_data = A.data_ptr<scalar_t>();
+
+    at::parallel_for(0, M, 0, [&] (int64_t begin, int64_t end) {
+      for (int64_t m = begin; m < end; ++m) {
+        quantize_row_int8<scalar_t>(
+            Aq_data + m * K,
+            As_data[m],
+            A_data + m * lda,
+            K);
+      }
+    });
+  });
+  return std::make_tuple(Aq, As);
+}
+
+// weight     :  static, per-channel, symmetric
+// activation : dynamic,   per-token, symmetric
+//
+// mat1    : [M, K]
+// mat2    : [N, K]
+// scales1 : [M]
+// scales2 : [N]
+// bias    : [N]
+// out     : [M, N]
+//
+at::Tensor int8_scaled_mm_cpu(at::Tensor& mat1, at::Tensor& mat2,
+    at::Tensor& scales1, at::Tensor& scales2,
+    std::optional<at::Tensor>& bias, at::ScalarType out_dtype, bool is_vnni) {
+  RECORD_FUNCTION("sgl-kernel::int8_scaled_mm_cpu", std::vector<c10::IValue>({mat1, mat2, scales1, scales2, bias}));
+
+  auto packed_w = is_vnni ? mat2 : convert_weight_packed(mat2);
+
+  CHECK_INPUT(mat1);
+  CHECK_INPUT(mat2);
+  CHECK_INPUT(scales1);
+  CHECK_INPUT(scales2);
+  CHECK_DIM(2, mat1);
+  CHECK_DIM(2, mat2);
+
+  int64_t M = mat1.size(0);
+  int64_t N = mat2.size(0);
+  int64_t K = mat1.size(1);
+
+  // see [NOTE]: s8s8 igemm compensation in avx512-vnni
+  CHECK_EQ(mat2.size(1), (int64_t)(is_vnni ? K + sizeof(int32_t) : K));
+  CHECK_EQ(scales1.numel(), M);
+  CHECK_EQ(scales2.numel(), N);
+
+  TORCH_CHECK(mat1.scalar_type() == at::kByte, "int8_scaled_mm: expect mat1 to be uint8.");
+  TORCH_CHECK(mat2.scalar_type() == at::kChar, "int8_scaled_mm: expect mat2 to be int8.");
+  TORCH_CHECK(scales1.scalar_type() == at::kFloat && scales2.scalar_type() == at::kFloat,
+      "int8_scaled_mm: expect scales to be float32.");
+
+  auto out = at::empty({M, N}, mat1.options().dtype(out_dtype));
+
+  const bool has_bias = bias.has_value();
+  const float* bias_data = nullptr;
+  if (has_bias) {
+    CHECK_EQ(bias.value().size(0), N);
+    bias_data = bias.value().data_ptr<float>();
+  }
+
+  AT_DISPATCH_REDUCED_FLOATING_TYPES(out_dtype, "int8_scaled_mm_kernel_impl", [&] {
+    int8_scaled_mm_kernel_impl<scalar_t>(
+        out.data_ptr<scalar_t>(),
+        mat1.data_ptr<uint8_t>(),
+        packed_w.data_ptr<int8_t>(),
+        scales1.data_ptr<float>(),
+        scales2.data_ptr<float>(),
+        bias_data,
+        M,
+        N,
+        K);
+  });
+  return out;
+}
+
+// fused `per_token_quant_int8_cpu` and `int8_scaled_mm_cpu`
+at::Tensor int8_scaled_mm_with_quant(at::Tensor& mat1, at::Tensor& mat2, at::Tensor& scales2,
+    const std::optional<at::Tensor>& bias, at::ScalarType out_dtype, bool is_vnni) {
+  RECORD_FUNCTION("sgl-kernel::int8_scaled_mm_cpu", std::vector<c10::IValue>({mat1, mat2, scales2, bias}));
+
+  auto packed_w = is_vnni ? mat2 : convert_weight_packed(mat2);
+
+  CHECK_LAST_DIM_CONTIGUOUS_INPUT(mat1);
+  CHECK_INPUT(mat2);
+  CHECK_INPUT(scales2);
+  CHECK_DIM(2, mat1);
+  CHECK_DIM(2, mat2);
+
+  int64_t M = mat1.size(0);
+  int64_t N = mat2.size(0);
+  int64_t K = mat1.size(1);
+  int64_t lda = mat1.stride(0);
+
+  // see [NOTE]: s8s8 igemm compensation in avx512-vnni
+  CHECK_EQ(mat2.size(1), (int64_t)(is_vnni ? K + sizeof(int32_t) : K));
+  CHECK_EQ(scales2.numel(), N);
+
+  const auto st = mat1.scalar_type();
+  TORCH_CHECK(st == at::kBFloat16 || st == at::kHalf,
+      "int8_scaled_mm_with_quant: expect A to be bfloat16 or half.");
+  TORCH_CHECK(st == out_dtype,
+      "int8_scaled_mm_with_quant: expect A has same dtype with out_dtype.");
+  TORCH_CHECK(mat2.scalar_type() == at::kChar,
+      "int8_scaled_mm_with_quant: expect mat2 to be int8.");
+  TORCH_CHECK(scales2.scalar_type() == at::kFloat,
+      "int8_scaled_mm_with_quant: expect scales to be float32.");
+
+  const int64_t buffer_size = M * K + M * sizeof(float);
+  auto buffer = at::empty({buffer_size}, mat1.options().dtype(at::kByte));
+  auto out = at::empty({M, N}, mat1.options().dtype(out_dtype));
+
+  const bool has_bias = bias.has_value();
+  const float* bias_data = nullptr;
+  if (has_bias) {
+    CHECK_EQ(bias.value().size(0), N);
+    bias_data = bias.value().data_ptr<float>();
+  }
+
+  AT_DISPATCH_REDUCED_FLOATING_TYPES(out_dtype, "int8_scaled_mm_with_quant_kernel_impl", [&] {
+    uint8_t* __restrict__ Aq_data = buffer.data_ptr<uint8_t>();
+    float* __restrict__ As_data = (float*)((void*)(Aq_data + M * K));
+    const scalar_t* __restrict__ A_data = mat1.data_ptr<scalar_t>();
+
+    at::parallel_for(0, M, 0, [&] (int64_t begin, int64_t end) {
+      for (int64_t m = begin; m < end; ++m) {
+        quantize_row_int8<scalar_t>(
+            Aq_data + m * K,
+            As_data[m],
+            A_data + m * lda,
+            K);
+      }
+    });
+
+    int8_scaled_mm_kernel_impl<scalar_t>(
+        out.data_ptr<scalar_t>(),
+        Aq_data,
+        packed_w.data_ptr<int8_t>(),
+        As_data,
+        scales2.data_ptr<float>(),
+        bias_data,
+        M,
+        N,
+        K);
+  });
+  return out;
+}

csrc/cpu/sgl-kernels/moe_fp8.cpp

@@ -0,0 +1,502 @@
+// Adapted from
+// https://github.com/sgl-project/sglang/tree/main/sgl-kernel/csrc/cpu
+
+#include "common.h"
+#include "gemm.h"
+#include "vec.h"
+
+// clang-format off
+
+namespace {
+
+template <typename scalar_t>
+inline void copy_stub(scalar_t* __restrict__ out, const scalar_t* __restrict__ input, int64_t size) {
+  using Vec = at::vec::Vectorized<scalar_t>;
+  // no remainder
+  #pragma GCC unroll 4
+  for (int64_t d = 0; d < size; d += Vec::size()) {
+    Vec data = Vec::loadu(input + d);
+    data.store(out + d);
+  }
+}
+
+template <typename scalar_t>
+inline void copy_mul_stub(scalar_t* __restrict__ out, const scalar_t* __restrict__ input, float weight, int64_t size) {
+  using bVec = at::vec::Vectorized<scalar_t>;
+  using fVec = at::vec::Vectorized<float>;
+  constexpr int kVecSize = bVec::size();
+  const fVec weight_vec = fVec(weight);
+  int64_t d;
+  #pragma GCC unroll 4
+  for (d = 0; d <= size - kVecSize; d += kVecSize) {
+    bVec x = bVec::loadu(input + d);
+    fVec x0, x1;
+    std::tie(x0, x1) = at::vec::convert_to_float(x);
+    x0 = x0 * weight_vec;
+    x1 = x1 * weight_vec;
+    bVec out_vec = convert_from_float_ext<scalar_t>(x0, x1);
+    out_vec.store(out + d);
+  }
+  for (; d < size; ++d) {
+    out[d] = static_cast<scalar_t>(input[d] * weight);
+  }
+}
+
+// acc from [topk, K] to [K]
+template <typename scalar_t>
+inline void sum_stub(scalar_t* __restrict__ out, const scalar_t* __restrict__ input, int64_t topk, int64_t K) {
+  using bVec = at::vec::Vectorized<scalar_t>;
+  using fVec = at::vec::Vectorized<float>;
+  constexpr int kVecSize = bVec::size();
+  if (topk == 1) {
+    // do copy for topk = 1
+    copy_stub(out, input, K);
+  } else {
+    // do sum for topk != 1
+    int64_t d;
+    #pragma GCC unroll 4
+    for (d = 0; d <= K - kVecSize; d += kVecSize) {
+      fVec sum_fvec0 = fVec(0.f);
+      fVec sum_fvec1 = fVec(0.f);
+      for (int t = 0; t < topk; ++t) {
+        bVec x_bvec = bVec::loadu(input + t * K + d);
+        fVec x_fvec0, x_fvec1;
+        std::tie(x_fvec0, x_fvec1) = at::vec::convert_to_float(x_bvec);
+
+        sum_fvec0 += x_fvec0;
+        sum_fvec1 += x_fvec1;
+      }
+      bVec out_bvec = convert_from_float_ext<scalar_t>(sum_fvec0, sum_fvec1);
+      out_bvec.store(out + d);
+    }
+    for (; d < K; ++d) {
+      float sum_val = 0.f;
+      for (int t = 0; t < topk; ++t) {
+        sum_val += static_cast<float>(input[t * K + d]);
+      }
+      out[d] = static_cast<scalar_t>(sum_val);
+    }
+  }
+}
+
+// out = input + input2 * scale
+template <typename scalar_t>
+inline void add_mul_stub(
+    scalar_t* __restrict__ out,
+    const scalar_t* __restrict__ input,
+    const scalar_t* __restrict__ input2,
+    float scale,
+    int64_t size) {
+  using bVec = at::vec::Vectorized<scalar_t>;
+  using fVec = at::vec::Vectorized<float>;
+  constexpr int kVecSize = bVec::size();
+  const fVec s_vec = fVec(scale);
+
+  int64_t d;
+#pragma GCC unroll 4
+  for (d = 0; d <= size - kVecSize; d += kVecSize) {
+    bVec x_bvec = bVec::loadu(input + d);
+    fVec x0, x1;
+    std::tie(x0, x1) = at::vec::convert_to_float(x_bvec);
+
+    bVec y_bvec = bVec::loadu(input2 + d);
+    fVec y0, y1;
+    std::tie(y0, y1) = at::vec::convert_to_float(y_bvec);
+
+    x0 = x0 + y0 * s_vec;
+    x1 = x1 + y1 * s_vec;
+    bVec out_vec = convert_from_float_ext<scalar_t>(x0, x1);
+    out_vec.store(out + d);
+  }
+  for (; d < size; ++d) {
+    out[d] = static_cast<scalar_t>(input[d] + float(input2[d]) * scale);
+  }
+}
+
+template <typename scalar_t>
+inline void silu_and_mul_stub(
+    scalar_t* __restrict__ out,
+    const scalar_t* __restrict__ input,
+    const scalar_t* __restrict__ input2,
+    int64_t size) {
+  using bVec = at::vec::Vectorized<scalar_t>;
+  using fVec = at::vec::Vectorized<float>;
+  const fVec one = fVec(1.f);
+
+  // no remainder
+#pragma GCC unroll 4
+  for (int64_t d = 0; d < size; d += bVec::size()) {
+    bVec x = bVec::loadu(input + d);
+    fVec x0, x1;
+    std::tie(x0, x1) = at::vec::convert_to_float(x);
+    bVec y = bVec::loadu(input2 + d);
+    fVec y0, y1;
+    std::tie(y0, y1) = at::vec::convert_to_float(y);
+    x0 = x0 / (one + x0.neg().exp_u20());
+    x1 = x1 / (one + x1.neg().exp_u20());
+    x0 = x0 * y0;
+    x1 = x1 * y1;
+    bVec out_vec = convert_from_float_ext<scalar_t>(x0, x1);
+    out_vec.store(out + d);
+  }
+}
+
+} // anonymous namespace
+
+template <typename scalar_t>
+void fused_experts_fp8_kernel_impl(
+    scalar_t* __restrict__ output,
+    scalar_t* __restrict__ ic0,
+    scalar_t* __restrict__ ic1,
+    scalar_t* __restrict__ ic2,
+    scalar_t* __restrict__ A_tmp,
+    scalar_t* __restrict__ B_tmp,
+    float* __restrict__ C_tmp,
+    const scalar_t* __restrict__ input,
+    const at::Float8_e4m3fn* __restrict__ packed_w1,
+    const at::Float8_e4m3fn* __restrict__ packed_w2,
+    const float* __restrict__ w1s,
+    const float* __restrict__ w2s,
+    int64_t block_size_N,
+    int64_t block_size_K,
+    const float* __restrict__ topk_weights,
+    const int32_t* __restrict__ sorted_ids,
+    const int32_t* __restrict__ expert_ids,
+    const int32_t* __restrict__ offsets,
+    int64_t M,
+    int64_t N,
+    int64_t K,
+    int64_t E,
+    int64_t topk,
+    int64_t num_tokens_post_pad) {
+
+  constexpr int64_t BLOCK_M = block_size_m();
+  constexpr int64_t BLOCK_N = block_size_n();
+
+  // stage 1: intermediate_cache0 = hidden_states @ w1
+  const int64_t MB = div_up(num_tokens_post_pad, BLOCK_M);
+  const int64_t NB = div_up(2 * N, BLOCK_N);
+  int64_t scale_size_N = div_up(2 * N, block_size_N);
+  int64_t scale_size_K = div_up(K, block_size_K);
+  int64_t blocks_n_per_group = block_size_N / BLOCK_N;
+
+  const int64_t stride_e = 2 * N * K;
+  const int64_t stride_n = K;
+
+  // here we only parallel on half of 2N to fuse silu_and_mul with gemm
+  at::parallel_for(0, MB * NB, 0, [&](int64_t begin, int64_t end) {
+    // get local pointers
+    int tid = at::get_thread_num();
+    scalar_t* __restrict__ A = A_tmp + tid * BLOCK_M * K;
+
+    bool is_brgemm_used = false;
+
+    for (int64_t i = begin; i < end; ++i) {
+      int64_t mb = i / NB;
+      int64_t nb = i % NB;
+
+      int64_t n_size = std::min(2 * N - nb * BLOCK_N, BLOCK_N);
+
+      // B shape [K, n_size] in vnni format
+      int32_t expert_id = expert_ids[mb];
+      const at::Float8_e4m3fn* __restrict__ B = packed_w1 + expert_id * stride_e + nb * BLOCK_N * stride_n;
+      const float* __restrict__ Bs = w1s + expert_id * scale_size_N * scale_size_K + (nb / blocks_n_per_group) * scale_size_K;
+
+      // 1.a load A
+      const int32_t* A_ids = sorted_ids + mb * BLOCK_M;
+      int64_t m_size = offsets[mb + 1] - offsets[mb];
+
+      const bool use_brgemm = can_use_brgemm<at::Float8_e4m3fn>(m_size);
+      is_brgemm_used = is_brgemm_used || use_brgemm;
+
+      for (int64_t m = 0; m < m_size; ++m) {
+        int32_t index = A_ids[m] / topk;
+        copy_stub(A + m * K, input + index * K, K);
+      }
+
+      const int64_t offset = offsets[mb];
+      tinygemm_kernel<scalar_t>(
+          /*   A            */ A,
+          /*   B            */ B,
+          /*   C            */ ic0 + offset * 2 * N + nb * BLOCK_N,
+          /*   Btmp         */ B_tmp + tid * BLOCK_N * std::max(K, N),
+          /*   Ctmp         */ C_tmp + tid * 2 * BLOCK_M * BLOCK_N,
+          /*   scale        */ Bs,
+          /*   M            */ m_size,
+          /*   N            */ n_size,
+          /*   K            */ K,
+          /*   lda          */ K,
+          /*   ldb          */ n_size,
+          /*   ldc          */ 2 * N,
+          /*   brg          */ use_brgemm,
+          /*   block_size_K */ block_size_K);
+    }
+
+    if (is_brgemm_used) {
+      at::native::cpublas::brgemm_release();
+    }
+  });
+
+  // stage 1.5: intermediate_cache1 = silu(intermediate_cache0)
+  at::parallel_for(0, M * topk, 0, [&](int64_t begin, int64_t end) {
+    for (int64_t m = begin; m < end; ++m) {
+      silu_and_mul_stub(
+          ic1 + m * N,
+          ic0 + m * 2 * N,
+          ic0 + m * 2 * N + N,
+          N);
+    }
+  });
+
+  // stage 2: intermediate_cache2 = intermediate_cache1 @ w2
+  //   w2 : [E, K, N] as [E, OC, IC]
+  const int64_t OC = K;  // rename K as OC
+  const int64_t IC = N;  // rename N as IC
+  const int64_t MB2 = MB;
+  const int64_t NB2 = div_up(OC, BLOCK_N);
+  scale_size_N = div_up(K, block_size_N);
+  scale_size_K = div_up(N, block_size_K);
+  const int64_t stride_e2 = OC * IC;
+  const int64_t stride_oc = IC;
+
+  // parallel on [MB2, NB2]
+  at::parallel_for(0, MB2 * NB2, 0, [&](int64_t begin, int64_t end) {
+    int tid = at::get_thread_num();
+    alignas(64) scalar_t C[BLOCK_M * BLOCK_K];
+
+    bool is_brgemm_used = false;
+
+    for (int64_t i = begin; i < end; ++i) {
+      int64_t mb = i / NB2;
+      int64_t nb = i % NB2;
+
+      int64_t m_size = offsets[mb + 1] - offsets[mb];
+      int64_t n_size = std::min(OC - nb * BLOCK_N, BLOCK_N);
+
+      const bool use_brgemm = can_use_brgemm<at::Float8_e4m3fn>(m_size);
+      is_brgemm_used = is_brgemm_used || use_brgemm;
+
+      // A ptr from ic1 of [M * topk, N] in sorted order
+      // so as to avoid copy A to tmp buffer again
+      const scalar_t* __restrict__ A = ic1 + offsets[mb] * N;
+      const int32_t* A_ids = sorted_ids + mb * BLOCK_M;
+
+      // B shape [IC, n_size] in vnni format
+      int32_t expert_id = expert_ids[mb];
+      const at::Float8_e4m3fn* __restrict__ B = packed_w2 + expert_id * stride_e2 + nb * BLOCK_N * stride_oc;
+      const float* __restrict__ Bs = w2s + expert_id * scale_size_N * scale_size_K + (nb / blocks_n_per_group) * scale_size_K;
+
+      tinygemm_kernel<scalar_t>(
+          /*   A            */ A,
+          /*   B            */ B,
+          /*   C            */ C,
+          /*   Btmp         */ B_tmp + tid * BLOCK_N * std::max(K, N),
+          /*   Ctmp         */ C_tmp + tid * 2 * BLOCK_M * BLOCK_N,
+          /*   scale        */ Bs,
+          /*   M            */ m_size,
+          /*   N            */ n_size,
+          /*   K            */ IC,
+          /*   lda          */ IC,
+          /*   ldb          */ n_size,
+          /*   ldc          */ BLOCK_N,
+          /*   brg          */ use_brgemm,
+          /*   block_size_K */ block_size_K);
+
+      // 2.b copy from C to ic2 in original order
+      //   and also mul topk_weights in float32
+      for (int64_t m = 0; m < m_size; ++m) {
+        int32_t index = A_ids[m];
+        float weight = topk_weights[index];
+        copy_mul_stub(ic2 + index * K + nb * BLOCK_N, C + m * BLOCK_N, weight, n_size);
+      }
+    }
+
+    if (is_brgemm_used) {
+      at::native::cpublas::brgemm_release();
+    }
+  });
+
+  // stage 3: out = intermediate_cache2.sum(dim=1)
+  //   from [M, topk, K] to [M, K]
+  at::parallel_for(0, M, 0, [&](int64_t begin, int64_t end) {
+    for (int64_t m = begin; m < end; ++m) {
+      sum_stub(output + m * K, ic2 + m * topk * K, topk, K);
+    }
+  });
+}
+
+#define INSTANTIATE_MOE_FP8_TEMPLATE(TYPE)             \
+  template void fused_experts_fp8_kernel_impl<TYPE>(   \
+      TYPE* __restrict__ output,                       \
+      TYPE* __restrict__ ic0,                          \
+      TYPE* __restrict__ ic1,                          \
+      TYPE* __restrict__ ic2,                          \
+      TYPE* __restrict__ A_tmp,                        \
+      TYPE* __restrict__ B_tmp,                        \
+      float* __restrict__ C_tmp,                       \
+      const TYPE* __restrict__ input,                  \
+      const at::Float8_e4m3fn* __restrict__ packed_w1, \
+      const at::Float8_e4m3fn* __restrict__ packed_w2, \
+      const float* __restrict__ w1s,                   \
+      const float* __restrict__ w2s,                   \
+      int64_t block_size_N,                            \
+      int64_t block_size_K,                            \
+      const float* __restrict__ topk_weights,          \
+      const int32_t* __restrict__ sorted_ids,          \
+      const int32_t* __restrict__ expert_ids,          \
+      const int32_t* __restrict__ offsets,             \
+      int64_t M,                                       \
+      int64_t N,                                       \
+      int64_t K,                                       \
+      int64_t E,                                       \
+      int64_t topk,                                    \
+      int64_t num_tokens_post_pad)
+
+INSTANTIATE_MOE_FP8_TEMPLATE(at::BFloat16);
+INSTANTIATE_MOE_FP8_TEMPLATE(at::Half);
+
+template <typename scalar_t>
+void shared_expert_fp8_kernel_impl(
+    scalar_t* __restrict__ output,
+    scalar_t* __restrict__ ic0,
+    scalar_t* __restrict__ ic1,
+    scalar_t* __restrict__ B_tmp,
+    float* __restrict__ C_tmp,
+    const scalar_t* __restrict__ input,
+    const at::Float8_e4m3fn* __restrict__ packed_w1,
+    const at::Float8_e4m3fn* __restrict__ packed_w2,
+    const float* __restrict__ w1s,
+    const float* __restrict__ w2s,
+    int64_t block_size_N,
+    int64_t block_size_K,
+    const scalar_t* __restrict__ fused_experts_out,
+    float routed_scaling_factor,
+    int64_t M,
+    int64_t N,
+    int64_t K) {
+
+  constexpr int64_t BLOCK_M = block_size_m();
+  constexpr int64_t BLOCK_N = block_size_n();
+
+  // stage 1: intermediate_cache0 = hidden_states @ w1
+  const int64_t MB = div_up(M, BLOCK_M);
+  const int64_t NB = div_up(2 * N, BLOCK_N);
+  int64_t scale_size_K = div_up(K, block_size_K);
+  int64_t blocks_n_per_group = block_size_N / BLOCK_N;
+
+  const bool use_brgemm = can_use_brgemm<at::Float8_e4m3fn>(M);
+
+  at::parallel_for(0, MB * NB, 0, [&](int64_t begin, int64_t end) {
+    int tid = at::get_thread_num();
+
+    for (int64_t i = begin; i < end; ++i) {
+      int64_t mb = i / NB;
+      int64_t nb = i % NB;
+      int64_t m_size = std::min(M - mb * BLOCK_M, BLOCK_M);
+      int64_t n_size = std::min(2 * N - nb * BLOCK_N, BLOCK_N);
+
+      tinygemm_kernel<scalar_t>(
+          /*   A            */ input + mb * BLOCK_M * K,
+          /*   B            */ packed_w1 + nb * BLOCK_N * K,
+          /*   C            */ ic0 + mb * BLOCK_M * 2 * N + nb * BLOCK_N,
+          /*   Btmp         */ B_tmp + tid * BLOCK_N * std::max(K, N),
+          /*   Ctmp         */ C_tmp + tid * 2 * BLOCK_M * BLOCK_N,
+          /*   scale        */ w1s + (nb / blocks_n_per_group) * scale_size_K,
+          /*   M            */ m_size,
+          /*   N            */ n_size,
+          /*   K            */ K,
+          /*   lda          */ K,
+          /*   ldb          */ n_size,
+          /*   ldc          */ 2 * N,
+          /*   brg          */ use_brgemm,
+          /*   block_size_K */ block_size_K);
+    }
+
+    if (use_brgemm) {
+      at::native::cpublas::brgemm_release();
+    }
+  });
+
+  // stage 1.5: intermediate_cache1 = silu(intermediate_cache0)
+  at::parallel_for(0, M, 0, [&](int64_t begin, int64_t end) {
+    for (int64_t m = begin; m < end; ++m) {
+      silu_and_mul_stub(
+          ic1 + m * N,
+          ic0 + m * 2 * N,
+          ic0 + m * 2 * N + N,
+          N);
+    }
+  });
+
+  // stage 2: intermediate_cache2 = intermediate_cache1 @ w2
+  //   w2 : [K, N] as [OC, IC]
+  const int64_t OC = K;  // rename K as OC
+  const int64_t IC = N;  // rename N as IC
+  const int64_t MB2 = MB;
+  const int64_t NB2 = div_up(K, BLOCK_N);
+  scale_size_K = div_up(N, block_size_K);
+
+  // parallel on [MB2, NB2]
+  at::parallel_for(0, MB2 * NB2, 0, [&](int64_t begin, int64_t end) {
+    int tid = at::get_thread_num();
+    alignas(64) scalar_t C[BLOCK_M * BLOCK_K];
+
+    for (int64_t i = begin; i < end; ++i) {
+      int64_t mb = i / NB2;
+      int64_t nb = i % NB2;
+      int64_t m_size = std::min(M - mb * BLOCK_M, BLOCK_M);
+      int64_t n_size = std::min(OC - nb * BLOCK_N, BLOCK_N);
+
+      // 2.a gemm: C = A @ B
+      tinygemm_kernel<scalar_t>(
+          /*   A            */ ic1 + mb * BLOCK_M * N,
+          /*   B            */ packed_w2 + nb * BLOCK_N * N,
+          /*   C            */ C,
+          /*   Btmp         */ B_tmp + tid * BLOCK_N * std::max(K, N),
+          /*   Ctmp         */ C_tmp + tid * 2 * BLOCK_M * BLOCK_N,
+          /*   scale        */ w2s + (nb / blocks_n_per_group) * scale_size_K,
+          /*   M            */ m_size,
+          /*   N            */ n_size,
+          /*   K            */ IC,
+          /*   lda          */ IC,
+          /*   ldb          */ n_size,
+          /*   ldc          */ BLOCK_N,
+          /*   brg          */ use_brgemm,
+          /*   block_size_K */ block_size_K);
+
+      // 2.b copy from C to output and add fused_experts_out
+      scalar_t* __restrict__ out = output + mb * BLOCK_M * K + nb * BLOCK_N;
+      const scalar_t* __restrict__ fused_out = fused_experts_out + mb * BLOCK_M * K + nb * BLOCK_N;
+      for (int64_t m = 0; m < m_size; ++m) {
+        add_mul_stub(out + m * K, C + m * BLOCK_N, fused_out + m * K, routed_scaling_factor, n_size);
+      }
+    }
+  });
+
+  if (use_brgemm) {
+    at::native::cpublas::brgemm_release();
+  }
+}
+
+#define INSTANTIATE_SHARED_EXPERT_FP8_TEMPLATE(TYPE)   \
+  template void shared_expert_fp8_kernel_impl<TYPE>(   \
+      TYPE* __restrict__ output,                       \
+      TYPE* __restrict__ ic0,                          \
+      TYPE* __restrict__ ic1,                          \
+      TYPE* __restrict__ B_tmp,                        \
+      float* __restrict__ C_tmp,                       \
+      const TYPE* __restrict__ input,                  \
+      const at::Float8_e4m3fn* __restrict__ packed_w1, \
+      const at::Float8_e4m3fn* __restrict__ packed_w2, \
+      const float* __restrict__ w1s,                   \
+      const float* __restrict__ w2s,                   \
+      int64_t block_size_N,                            \
+      int64_t block_size_K,                            \
+      const TYPE* __restrict__ fused_experts_out,      \
+      float routed_scaling_factor,                     \
+      int64_t M,                                       \
+      int64_t N,                                       \
+      int64_t K)
+
+INSTANTIATE_SHARED_EXPERT_FP8_TEMPLATE(at::BFloat16);
+INSTANTIATE_SHARED_EXPERT_FP8_TEMPLATE(at::Half);

csrc/cpu/sgl-kernels/moe_int8.cpp

@@ -0,0 +1,769 @@
+// Adapted from
+// https://github.com/sgl-project/sglang/tree/main/sgl-kernel/csrc/cpu
+
+#include "common.h"
+#include "vec.h"
+#include "gemm.h"
+
+// clang-format off
+
+namespace {
+
+template <typename scalar_t>
+inline void copy_stub(scalar_t* __restrict__ out, const scalar_t* __restrict__ input, int64_t size) {
+  using Vec = at::vec::Vectorized<scalar_t>;
+  // no remainder
+  #pragma GCC unroll 4
+  for (int64_t d = 0; d < size; d += Vec::size()) {
+    Vec data = Vec::loadu(input + d);
+    data.store(out + d);
+  }
+}
+
+template <>
+inline void copy_stub<uint8_t>(uint8_t* __restrict__ out, const uint8_t* __restrict__ input, int64_t size) {
+  // size might be 64x + 32
+  std::memcpy(out, input, size * sizeof(uint8_t));
+}
+
+template <typename scalar_t>
+inline void copy_mul_stub(scalar_t* __restrict__ out, const float* __restrict__ input, float weight, int64_t size) {
+  using bVec = at::vec::Vectorized<scalar_t>;
+  using fVec = at::vec::Vectorized<float>;
+  constexpr int kVecSize = bVec::size();
+  const fVec weight_vec = fVec(weight);
+  int64_t d;
+  #pragma GCC unroll 4
+  for (d = 0; d <= size - kVecSize; d += kVecSize) {
+    fVec data0 = fVec::loadu(input + d) * weight_vec;
+    fVec data1 = fVec::loadu(input + d + fVec::size()) * weight_vec;
+    bVec out_vec = convert_from_float_ext<scalar_t>(data0, data1);
+    out_vec.store(out + d);
+  }
+  for (; d < size; ++d) {
+    out[d] = static_cast<scalar_t>(input[d] * weight);
+  }
+}
+
+// acc from [topk, K] to [K]
+template <typename scalar_t>
+inline void sum_stub(scalar_t* __restrict__ out, const scalar_t* __restrict__ input, int64_t topk, int64_t K) {
+  using bVec = at::vec::Vectorized<scalar_t>;
+  using fVec = at::vec::Vectorized<float>;
+  constexpr int kVecSize = bVec::size();
+  if (topk == 1) {
+    // do copy for topk = 1
+    copy_stub(out, input, K);
+  } else {
+    // do sum for topk != 1
+    int64_t d;
+    #pragma GCC unroll 4
+    for (d = 0; d <= K - kVecSize; d += kVecSize) {
+      fVec sum_fvec0 = fVec(0.f);
+      fVec sum_fvec1 = fVec(0.f);
+      for (int t = 0; t < topk; ++t) {
+        bVec x_bvec = bVec::loadu(input + t * K + d);
+        fVec x_fvec0, x_fvec1;
+        std::tie(x_fvec0, x_fvec1) = at::vec::convert_to_float(x_bvec);
+
+        sum_fvec0 += x_fvec0;
+        sum_fvec1 += x_fvec1;
+      }
+      bVec out_bvec = convert_from_float_ext<scalar_t>(sum_fvec0, sum_fvec1);
+      out_bvec.store(out + d);
+    }
+    for (; d < K; ++d) {
+      float sum_val = 0.f;
+      for (int t = 0; t < topk; ++t) {
+        sum_val += static_cast<float>(input[t * K + d]);
+      }
+      out[d] = static_cast<scalar_t>(sum_val);
+    }
+  }
+}
+
+// out = input + input2 * scale
+template <typename scalar_t>
+inline void add_mul_stub(scalar_t* __restrict__ out, const float* __restrict__ input,
+    const scalar_t* __restrict__ input2, float scale, int64_t size) {
+
+  using bVec = at::vec::Vectorized<scalar_t>;
+  using fVec = at::vec::Vectorized<float>;
+  constexpr int kVecSize = bVec::size();
+  const fVec s_vec = fVec(scale);
+  int64_t d;
+  #pragma GCC unroll 4
+  for (d = 0; d <= size - kVecSize; d += kVecSize) {
+    fVec x0 = fVec::loadu(input + d);
+    fVec x1 = fVec::loadu(input + d + fVec::size());
+
+    bVec y_bvec = bVec::loadu(input2 + d);
+    fVec y0, y1;
+    std::tie(y0, y1) = at::vec::convert_to_float(y_bvec);
+
+    x0 = x0 + y0 * s_vec;
+    x1 = x1 + y1 * s_vec;
+    bVec out_vec = convert_from_float_ext<scalar_t>(x0, x1);
+    out_vec.store(out + d);
+  }
+  for (; d < size; ++d) {
+    out[d] = static_cast<scalar_t>(input[d] + float(input2[d]) * scale);
+  }
+}
+
+/// gemm for w13
+template <typename scalar_t, int BLOCK_M, int BLOCK_N>
+struct tinygemm_kernel_vnni {
+  static inline void apply(
+      const uint8_t* __restrict__ A, const int8_t* __restrict__ B0, const int8_t* __restrict__ B1, scalar_t* __restrict__ C,
+      const float* __restrict__ As, const float* __restrict__ Bs0, const float* __restrict__ Bs1,
+      const int32_t* __restrict__ Bcomp0, const int32_t* __restrict__ Bcomp1,
+      int64_t K, int64_t lda, int64_t ldb, int64_t ldc) {
+    TORCH_CHECK(false, "tinygemm_kernel_nn: scalar path not implemented!");
+  }
+};
+
+#if defined(CPU_CAPABILITY_AVX512)
+template <int BLOCK_M, int BLOCK_N>
+struct tinygemm_kernel_vnni<at::BFloat16, BLOCK_M, BLOCK_N> {
+  static inline void apply(
+      const uint8_t* __restrict__ A, const int8_t* __restrict__ B0, const int8_t* __restrict__ B1, at::BFloat16* __restrict__ C,
+      const float* __restrict__ As, const float* __restrict__ Bs0, const float* __restrict__ Bs1,
+      const int32_t* __restrict__ Bcomp0, const int32_t* __restrict__ Bcomp1,
+      int64_t K, int64_t lda, int64_t ldb, int64_t ldc) {
+
+    constexpr int ROWS = BLOCK_M;
+    constexpr int COLS = BLOCK_N / 16;
+    static_assert(COLS % 2 == 0);
+
+    __m512i va;
+    __m512i vb0[COLS];
+    __m512i vb1[COLS];
+    __m512i vc0[ROWS * COLS];
+    __m512i vc1[ROWS * COLS];
+    __m512i vcomp0[COLS];
+    __m512i vcomp1[COLS];
+    __m512  was;
+    __m512  vbs0[COLS];
+    __m512  vbs1[COLS];
+
+    auto loadc = [&](auto i) {
+      vc0[i] = _mm512_set1_epi32(0);
+      vc1[i] = _mm512_set1_epi32(0);
+    };
+    Unroll<ROWS * COLS>{}(loadc);
+
+    const int64_t K4 = K >> 2;
+    const int64_t lda4 = lda >> 2;
+    const int64_t ldb4 = ldb; // ldb * 4 >> 2;
+    const int32_t* a_ptr = reinterpret_cast<const int32_t*>(A);
+    const int32_t* b0_ptr = reinterpret_cast<const int32_t*>(B0);
+    const int32_t* b1_ptr = reinterpret_cast<const int32_t*>(B1);
+
+    auto compute = [&](auto i, int64_t k) {
+      constexpr int row = i / COLS;
+      constexpr int col = i % COLS;
+
+      if constexpr (col == 0) {
+        va = _mm512_set1_epi32(a_ptr[row * lda4 + k]);
+      }
+      if constexpr (row == 0) {
+        vb0[col] = _mm512_loadu_si512(b0_ptr + k * ldb4 + col * 16);
+        vb1[col] = _mm512_loadu_si512(b1_ptr + k * ldb4 + col * 16);
+      }
+      vc0[i] = _mm512_dpbusd_epi32(vc0[i], va, vb0[col]);
+      vc1[i] = _mm512_dpbusd_epi32(vc1[i], va, vb1[col]);
+    };
+    for (int64_t k = 0; k < K4; ++k) {
+      Unroll<ROWS * COLS>{}(compute, k);
+    }
+
+    auto scalec = [&](auto i) {
+      constexpr int row = i / COLS;
+      constexpr int col = i % COLS;
+
+      // load a scale
+      if constexpr(col == 0) {
+        was = _mm512_set1_ps(As[row]);
+      }
+      // load b scale and vcomp
+      if constexpr (row == 0) {
+        vbs0[col] = _mm512_loadu_ps(Bs0 + col * 16);
+        vbs1[col] = _mm512_loadu_ps(Bs1 + col * 16);
+        vcomp0[col] = _mm512_loadu_si512(Bcomp0 + col * 16);
+        vcomp1[col] = _mm512_loadu_si512(Bcomp1 + col * 16);
+      }
+      __m512 c0 = _mm512_cvtepi32_ps(_mm512_sub_epi32(vc0[i], vcomp0[col]));
+      __m512 c1 = _mm512_cvtepi32_ps(_mm512_sub_epi32(vc1[i], vcomp1[col]));
+      vc0[i] = _mm512_castps_si512(_mm512_mul_ps(_mm512_mul_ps(c0, was), vbs0[col]));
+      vc1[i] = _mm512_castps_si512(_mm512_mul_ps(_mm512_mul_ps(c1, was), vbs1[col]));
+    };
+    Unroll<ROWS * COLS>{}(scalec);
+
+    using Vec = at::vec::Vectorized<float>;
+    const Vec one = Vec(1.f);
+    auto storec = [&](auto i) {
+      constexpr int row = i / COLS;
+      constexpr int col = i % COLS;
+      // for COLS = 2, 4 use 512bit store
+      if constexpr (col % 2 == 0) {
+        Vec x0 = _mm512_castsi512_ps(vc0[row * COLS + col + 0]);
+        Vec x1 = _mm512_castsi512_ps(vc0[row * COLS + col + 1]);
+        Vec y0 = _mm512_castsi512_ps(vc1[row * COLS + col + 0]);
+        Vec y1 = _mm512_castsi512_ps(vc1[row * COLS + col + 1]);
+        // silu
+        x0 = x0 / (one + x0.neg().exp_u20());
+        x1 = x1 / (one + x1.neg().exp_u20());
+        // mul
+        x0 = x0 * y0;
+        x1 = x1 * y1;
+
+        _mm512_storeu_si512(
+            reinterpret_cast<__m512i*>((C + row * ldc + col * 16)),
+            (__m512i)(_mm512_cvtne2ps_pbh(__m512(x1), __m512(x0))));
+        }
+    };
+    Unroll<ROWS * COLS>{}(storec);
+  }
+};
+#endif
+
+#define LAUNCH_TINYGEMM_KERNEL_VNNI(MB_SIZE, NB_SIZE)                        \
+    tinygemm_kernel_vnni<scalar_t, MB_SIZE, NB_SIZE>::apply(                 \
+        A + mb_start * lda, B0 + nb_start * 4, B1 + nb_start * 4,            \
+        C + mb_start * ldc + nb_start, As + mb_start,                        \
+        Bs0 + nb_start, Bs1 + nb_start, Bcomp0 + nb_start, Bcomp1 + nb_start,\
+        K, lda, ldb, ldc);
+
+template <typename scalar_t>
+void tinygemm_kernel(
+    const uint8_t* __restrict__ A,
+    const int8_t* __restrict__ B0,
+    const int8_t* __restrict__ B1,
+    scalar_t* __restrict__ C,
+    const float* __restrict__ As,
+    const float* __restrict__ Bs0,
+    const float* __restrict__ Bs1,
+    int64_t M,
+    int64_t N,
+    int64_t K,
+    int64_t lda,
+    int64_t ldb,
+    int64_t ldc) {
+
+  const int32_t* Bcomp0 = reinterpret_cast<const int32_t*>(B0 + block_size_n() * K);
+  const int32_t* Bcomp1 = reinterpret_cast<const int32_t*>(B1 + block_size_n() * K);
+
+  // pattern: 1-(2+2)-(8+8)
+  constexpr int64_t BLOCK_M = 4;
+  constexpr int64_t BLOCK_N = 32;
+  const int64_t MB = div_up(M, BLOCK_M);
+  const int64_t NB = div_up(N, BLOCK_N);
+  for (int mb = 0; mb < MB; ++mb) {
+    int64_t mb_start = mb * BLOCK_M;
+    int64_t mb_size = std::min(BLOCK_M, M - mb_start);
+    for (int64_t nb = 0; nb < NB; ++nb) {
+      int64_t nb_start = nb * BLOCK_N;
+      int64_t nb_size = std::min(BLOCK_N, N - nb_start);
+
+      switch(mb_size << 4 | nb_size >> 4) {
+        case 0x12: LAUNCH_TINYGEMM_KERNEL_VNNI(1, 32); break;
+        case 0x22: LAUNCH_TINYGEMM_KERNEL_VNNI(2, 32); break;
+        case 0x32: LAUNCH_TINYGEMM_KERNEL_VNNI(3, 32); break;
+        case 0x42: LAUNCH_TINYGEMM_KERNEL_VNNI(4, 32); break;
+        default: TORCH_CHECK(false, "Unexpected block size, ", mb_size, "x", "nb_size");
+      }
+    }
+  }
+}
+
+/// gemm for w2
+template <typename scalar_t, int BLOCK_M, int BLOCK_N>
+struct tinygemm_kernel_vnni2 {
+  static inline void apply(
+      const uint8_t* __restrict__ A, const int8_t* __restrict__ B, float* __restrict__ C,
+      const float* __restrict__ As, const float* __restrict__ Bs, const int32_t* __restrict__ Bcomp,
+      int64_t K, int64_t lda, int64_t ldb, int64_t ldc) {
+    TORCH_CHECK(false, "tinygemm_kernel_nn: scalar path not implemented!");
+  }
+};
+
+#if defined(CPU_CAPABILITY_AVX512)
+template <int BLOCK_M, int BLOCK_N>
+struct tinygemm_kernel_vnni2<at::BFloat16, BLOCK_M, BLOCK_N> {
+  static inline void apply(
+      const uint8_t* __restrict__ A, const int8_t* __restrict__ B, float* __restrict__ C,
+      const float* __restrict__ As, const float* __restrict__ Bs, const int32_t* __restrict__ Bcomp,
+      int64_t K, int64_t lda, int64_t ldb, int64_t ldc) {
+
+    constexpr int ROWS = BLOCK_M;
+    constexpr int COLS = BLOCK_N / 16;
+    static_assert(COLS % 2 == 0);
+
+    __m512i va;
+    __m512i vb[COLS];
+    __m512i vc[ROWS * COLS];
+    __m512i vcomp[COLS];
+    __m512  was;
+    __m512  vbs[COLS];
+
+    auto loadc = [&](auto i) {
+      vc[i] = _mm512_set1_epi32(0);
+    };
+    Unroll<ROWS * COLS>{}(loadc);
+
+    const int64_t K4 = K >> 2;
+    const int64_t lda4 = lda >> 2;
+    const int64_t ldb4 = ldb; // ldb * 4 >> 2;
+    const int32_t* a_ptr = reinterpret_cast<const int32_t*>(A);
+    const int32_t* b_ptr = reinterpret_cast<const int32_t*>(B);
+
+    auto compute = [&](auto i, int64_t k) {
+      constexpr int row = i / COLS;
+      constexpr int col = i % COLS;
+
+      if constexpr (col == 0) {
+        va = _mm512_set1_epi32(a_ptr[row * lda4 + k]);
+      }
+      if constexpr (row == 0) {
+        vb[col] = _mm512_loadu_si512(b_ptr + k * ldb4 + col * 16);
+      }
+      vc[i] = _mm512_dpbusd_epi32(vc[i], va, vb[col]);
+    };
+    for (int64_t k = 0; k < K4; ++k) {
+      Unroll<ROWS * COLS>{}(compute, k);
+    }
+
+    auto storec = [&](auto i) {
+      constexpr int row = i / COLS;
+      constexpr int col = i % COLS;
+
+      // load a scale
+      if constexpr(col == 0) {
+        was = _mm512_set1_ps(As[row]);
+      }
+      // load b scale and vcomp per 2 vectors
+      // also load bias if any
+      if constexpr (row == 0) {
+        if constexpr (col % 2 == 0) {
+          vbs[col + 0] = _mm512_loadu_ps(Bs + col * 16);
+          vbs[col + 1] = _mm512_loadu_ps(Bs + col * 16 + 16);
+          vcomp[col + 0] = _mm512_loadu_si512(Bcomp + col * 16);
+          vcomp[col + 1] = _mm512_loadu_si512(Bcomp + col * 16 + 16);
+        }
+      }
+      __m512 x = _mm512_cvtepi32_ps(_mm512_sub_epi32(vc[i], vcomp[col]));
+      x = _mm512_mul_ps(_mm512_mul_ps(x, was), vbs[col]);
+      _mm512_storeu_ps(reinterpret_cast<__m512*>(C + row * ldc + col * 16), x);
+    };
+    Unroll<ROWS * COLS>{}(storec);
+  }
+};
+#endif
+
+#define LAUNCH_TINYGEMM_KERNEL_VNNI2(MB_SIZE, NB_SIZE)                       \
+    tinygemm_kernel_vnni2<scalar_t, MB_SIZE, NB_SIZE>::apply(                \
+        A + mb_start * lda, B + nb_start * 4, C + mb_start * ldc + nb_start, \
+        As + mb_start, Bs + nb_start, Bcomp + nb_start,                      \
+        K, lda, ldb, ldc);
+
+template <typename scalar_t>
+void tinygemm_kernel(
+    const uint8_t* __restrict__ A,
+    const int8_t* __restrict__ B,
+    float* __restrict__ C,
+    const float* __restrict__ As,
+    const float* __restrict__ Bs,
+    int64_t M,
+    int64_t N,
+    int64_t K,
+    int64_t lda,
+    int64_t ldb,
+    int64_t ldc) {
+
+  // B compensation
+  const int32_t* Bcomp = reinterpret_cast<const int32_t*>(B + block_size_n() * K);
+
+  // pattern: 1-4-16
+  constexpr int64_t BLOCK_M = 4;
+  constexpr int64_t BLOCK_N = 64;
+  const int64_t MB = div_up(M, BLOCK_M);
+  const int64_t NB = div_up(N, BLOCK_N);
+  for (int64_t mb = 0; mb < MB; ++mb) {
+    int64_t mb_start = mb * BLOCK_M;
+    int64_t mb_size = std::min(BLOCK_M, M - mb_start);
+    for (int64_t nb = 0; nb < NB; ++nb) {
+      int64_t nb_start = nb * BLOCK_N;
+      int64_t nb_size = std::min(BLOCK_N, N - nb_start);
+
+      switch(mb_size << 4 | nb_size >> 4) {
+        case 0x12: LAUNCH_TINYGEMM_KERNEL_VNNI2(1, 32); break;
+        case 0x22: LAUNCH_TINYGEMM_KERNEL_VNNI2(2, 32); break;
+        case 0x32: LAUNCH_TINYGEMM_KERNEL_VNNI2(3, 32); break;
+        case 0x42: LAUNCH_TINYGEMM_KERNEL_VNNI2(4, 32); break;
+        default: TORCH_CHECK(false, "Unexpected block size, ", mb_size, "x", "nb_size");
+      }
+    }
+  }
+}
+
+} // anonymous namespace
+
+template <typename scalar_t>
+void fused_experts_int8_kernel_impl(
+    scalar_t* __restrict__ output,
+    scalar_t* __restrict__ ic1,
+    scalar_t* __restrict__ ic2,
+    uint8_t* __restrict__ A_tmp,
+    float* __restrict__ C_tmp,
+    uint8_t* __restrict__ Aq_tmp,
+    float* __restrict__ As_tmp,
+    const scalar_t* __restrict__ input,
+    const int8_t* __restrict__ packed_w1,
+    const int8_t* __restrict__ packed_w2,
+    const float* __restrict__ w1s,
+    const float* __restrict__ w2s,
+    const float* __restrict__ topk_weights,
+    const int32_t* __restrict__ sorted_ids,
+    const int32_t* __restrict__ expert_ids,
+    const int32_t* __restrict__ offsets,
+    int64_t M,
+    int64_t N,
+    int64_t K,
+    int64_t E,
+    int64_t topk,
+    int64_t num_tokens_post_pad) {
+
+  // handle 2 tiles per block
+  constexpr int64_t BLOCK_M = block_size_m();
+  constexpr int64_t BLOCK_N = block_size_n();
+
+  // stage 0: quantize input to uint8, [M, K]
+  at::parallel_for(0, M, 0, [&](int64_t begin, int64_t end) {
+    for (int64_t m = begin; m < end; ++m) {
+      quantize_row_int8<scalar_t>(
+          Aq_tmp + m * K,
+          As_tmp[m],
+          input + m * K,
+          K);
+    }
+  });
+
+  // stage 1: intermediate_cache1 = silu(hidden_states @ w1)
+  const int64_t MB = div_up(num_tokens_post_pad, BLOCK_M);
+  const int64_t NB = div_up(N, BLOCK_N);
+
+  // strides for w1: [E, 2N, K]
+  TORCH_CHECK(N % BLOCK_N == 0, "Fixme when N is not multiples of ", BLOCK_N);
+
+  // K and N are packed for int8
+  const int64_t packed_K = get_row_size<int8_t>(K);
+  const int64_t packed_N = get_row_size<int8_t>(N);
+
+  const int64_t stride_e = 2 * N * packed_K;
+  const int64_t stride_n = packed_K;
+  // here we only parallel on half of 2N to fuse silu_and_mul with gemm
+  at::parallel_for(0, MB * NB, 0, [&](int64_t begin, int64_t end) {
+    // get local pointers
+    int tid = at::get_thread_num();
+    uint8_t* __restrict__ A = A_tmp + tid * BLOCK_M * K;
+
+    alignas(64) float As[BLOCK_M];
+
+    for (int64_t i = begin; i < end; ++i) {
+      int64_t mb = i / NB;
+      int64_t nb = i % NB;
+
+      // nb0 from top half and nb1 from bottom half
+      int64_t nb0 = nb, nb1 = nb + NB;
+      int64_t n_size = std::min(N - nb0 * BLOCK_N, BLOCK_N);
+
+      // B shape [K, n_size] in vnni format
+      int32_t expert_id = expert_ids[mb];
+      const int8_t* __restrict__ B0 = packed_w1 + expert_id * stride_e + nb0 * BLOCK_N * stride_n;
+      const int8_t* __restrict__ B1 = packed_w1 + expert_id * stride_e + nb1 * BLOCK_N * stride_n;
+      const float* __restrict__ Bs0 = w1s + expert_id * 2 * N + nb0 * BLOCK_N;
+      const float* __restrict__ Bs1 = w1s + expert_id * 2 * N + nb1 * BLOCK_N;
+
+      // 1.a load A
+      const int32_t* A_ids = sorted_ids + mb * BLOCK_M;
+      int64_t m_size = offsets[mb + 1] - offsets[mb];
+
+      for (int64_t m = 0; m < m_size; ++m) {
+        int32_t index = A_ids[m] / topk;
+        copy_stub(A + m * K, Aq_tmp + index * K, K);
+        As[m] = As_tmp[index];
+      }
+
+      // fused 1.b: silu_and_mul(A @ B0, A @ B1)
+      const int64_t offset = offsets[mb];
+      tinygemm_kernel(
+          /* A     */ A,
+          /* B0    */ B0,
+          /* B1    */ B1,
+          /* C     */ ic1 + offset * N + nb * BLOCK_N,
+          /* As    */ As,
+          /* Bs0   */ Bs0,
+          /* Bs1   */ Bs1,
+          /* M     */ m_size,
+          /* N     */ n_size,
+          /* K     */ K,
+          /* lda   */ K,
+          /* ldb   */ n_size,
+          /* ldc   */ N);
+    }
+  });
+
+  // stage 1.5: quantize ic1 to uint8, [M * topk, N]
+  at::parallel_for(0, M * topk, 0, [&](int64_t begin, int64_t end) {
+    for (int64_t m = begin; m < end; ++m) {
+      quantize_row_int8<scalar_t>(
+          Aq_tmp + m * N,
+          As_tmp[m],
+          ic1 + m * N,
+          N);
+    }
+  });
+
+  // stage 2: intermediate_cache2 = intermediate_cache1 @ w2
+  //   w2 : [E, K, N] as [E, OC, IC]
+  const int64_t OC = K;  // rename K as OC
+  const int64_t IC = N;  // rename N as IC
+  const int64_t MB2 = MB;
+  const int64_t NB2 = div_up(OC, BLOCK_N);
+  const int64_t stride_e2 = OC * packed_N;
+  const int64_t stride_oc = packed_N;
+
+  // parallel on [MB2, NB2]
+  at::parallel_for(0, MB2 * NB2, 0, [&](int64_t begin, int64_t end) {
+    // get local pointers
+    int tid = at::get_thread_num();
+    // we won't be using C1 for gemm2
+    float* __restrict__ C = C_tmp + tid * 2 * BLOCK_M * BLOCK_N;
+
+    for (int64_t i = begin; i < end; ++i) {
+      int64_t mb = i / NB2;
+      int64_t nb = i % NB2;
+
+      int64_t m_size = offsets[mb + 1] - offsets[mb];
+      int64_t n_size = std::min(OC - nb * BLOCK_N, BLOCK_N);
+
+      // A ptr from ic1 of [M * topk, N] in sorted order
+      // so as to avoid copy A to tmp buffer again
+      const uint8_t* __restrict__ A = Aq_tmp + offsets[mb] * N;
+      const float* __restrict__ As = As_tmp + offsets[mb];
+      const int32_t* A_ids = sorted_ids + mb * BLOCK_M;
+
+      // B shape [IC, n_size] in vnni format
+      int32_t expert_id = expert_ids[mb];
+      const int8_t* __restrict__ B = packed_w2 + expert_id * stride_e2 + nb * BLOCK_N * stride_oc;
+      const float* __restrict__ Bs = w2s + expert_id * K + nb * BLOCK_N;
+
+      // 2.a gemm: C = A @ B
+      tinygemm_kernel<scalar_t>(
+          /* A     */ A,
+          /* B     */ B,
+          /* C     */ C,
+          /* As    */ As,
+          /* Bs    */ Bs,
+          /* M     */ m_size,
+          /* N     */ n_size,
+          /* K     */ IC,
+          /* lda   */ IC,
+          /* ldb   */ n_size,
+          /* ldc   */ BLOCK_N);
+
+      // 2.b copy from C to ic2 in original order
+      //   and also mul topk_weights in float32
+      for (int64_t m = 0; m < m_size; ++m) {
+        int32_t index = A_ids[m];
+        float weight = topk_weights[index];
+        copy_mul_stub(ic2 + index * K + nb * BLOCK_N, C + m * BLOCK_N, weight, n_size);
+      }
+    }
+  });
+
+  // stage 3: out = intermediate_cache2.sum(dim=1)
+  //   from [M, topk, K] to [M, K]
+  at::parallel_for(0, M, 0, [&](int64_t begin, int64_t end) {
+    for (int64_t m = begin; m < end; ++m) {
+      sum_stub(output + m * K, ic2 + m * topk * K, topk, K);
+    }
+  });
+}
+
+#define INSTANTIATE_MOE_INT8_TEMPLATE(TYPE)                                                  \
+  template void fused_experts_int8_kernel_impl<TYPE> (                                       \
+      TYPE* __restrict__ output, TYPE* __restrict__ ic1,                                     \
+      TYPE* __restrict__ ic2, uint8_t* __restrict__ A_tmp,                                   \
+      float* __restrict__ C_tmp, uint8_t* __restrict__ Aq_tmp,                               \
+      float* __restrict__ As_tmp, const TYPE* __restrict__ input,                            \
+      const int8_t* __restrict__ packed_w1, const int8_t* __restrict__ packed_w2,            \
+      const float* __restrict__ w1s, const float* __restrict__ w2s,                          \
+      const float* __restrict__ topk_weights, const int32_t* __restrict__ sorted_ids,        \
+      const int32_t* __restrict__ expert_ids, const int32_t* __restrict__ offsets,           \
+      int64_t M, int64_t N, int64_t K, int64_t E, int64_t topk, int64_t num_tokens_post_pad)
+
+INSTANTIATE_MOE_INT8_TEMPLATE(at::BFloat16);
+INSTANTIATE_MOE_INT8_TEMPLATE(at::Half);
+
+template <typename scalar_t>
+void shared_expert_int8_kernel_impl(
+    scalar_t* __restrict__ output,
+    scalar_t* __restrict__ ic1,
+    float* __restrict__ C_tmp,
+    uint8_t* __restrict__ Aq_tmp,
+    float* __restrict__ As_tmp,
+    const scalar_t* __restrict__ input,
+    const int8_t* __restrict__ packed_w1,
+    const int8_t* __restrict__ packed_w2,
+    const float* __restrict__ w1s,
+    const float* __restrict__ w2s,
+    const scalar_t* __restrict__ fused_experts_out,
+    float routed_scaling_factor,
+    int64_t M,
+    int64_t N,
+    int64_t K) {
+
+  // handle 2 tiles per block
+  constexpr int64_t BLOCK_M = block_size_m();
+  constexpr int64_t BLOCK_N = block_size_n();
+
+  // stage 0: quantize input to uint8, [M, K]
+  at::parallel_for(0, M, 0, [&](int64_t begin, int64_t end) {
+    for (int64_t m = begin; m < end; ++m) {
+      quantize_row_int8<scalar_t>(
+          Aq_tmp + m * K,
+          As_tmp[m],
+          input + m * K,
+          K);
+    }
+  });
+
+   // stage 1: intermediate_cache1 = silu(hidden_states @ w1)
+  const int64_t MB = div_up(M, BLOCK_M);
+  const int64_t NB = div_up(N, BLOCK_N);
+
+  TORCH_CHECK(N % BLOCK_N == 0, "Fixme when N is not multiples of ", BLOCK_N);
+
+  // K and N are packed for int8
+  const int64_t packed_K = get_row_size<int8_t>(K);
+  const int64_t packed_N = get_row_size<int8_t>(N);
+  const int64_t stride_n = packed_K;
+
+  // here we only parallel on half of 2N to fuse silu_and_mul with gemm
+  at::parallel_for(0, MB * NB, 0, [&](int64_t begin, int64_t end) {
+    for (int64_t i = begin; i < end; ++i) {
+      int64_t mb = i / NB;
+      int64_t nb = i % NB;
+
+      // nb0 from top half and nb1 from bottom half
+      int64_t nb0 = nb, nb1 = nb + NB;
+      int64_t n_size = std::min(N - nb0 * BLOCK_N, BLOCK_N);
+      int64_t m_size = std::min(M - mb * BLOCK_M, BLOCK_M);
+
+      // A shape [m_size, K]
+      const uint8_t* A = Aq_tmp + mb * BLOCK_M * K;
+      const float* As = As_tmp + mb * BLOCK_M;
+
+      // B shape [K, n_size] in vnni format
+      const int8_t* __restrict__ B0 = packed_w1 + nb0 * BLOCK_N * stride_n;
+      const int8_t* __restrict__ B1 = packed_w1 + nb1 * BLOCK_N * stride_n;
+      const float* __restrict__ Bs0 = w1s + nb0 * BLOCK_N;
+      const float* __restrict__ Bs1 = w1s + nb1 * BLOCK_N;
+
+      // fused 1.b: silu_and_mul(A @ B0, A @ B1)
+      tinygemm_kernel(
+          /* A     */ A,
+          /* B0    */ B0,
+          /* B1    */ B1,
+          /* C     */ ic1 + mb * BLOCK_M * N + nb * BLOCK_N,
+          /* As    */ As,
+          /* Bs0   */ Bs0,
+          /* Bs1   */ Bs1,
+          /* M     */ m_size,
+          /* N     */ n_size,
+          /* K     */ K,
+          /* lda   */ K,
+          /* ldb   */ n_size,
+          /* ldc   */ N);
+    }
+  });
+
+  // stage 1.5: quantize ic1 to uint8, [M * topk, N]
+  at::parallel_for(0, M, 0, [&](int64_t begin, int64_t end) {
+    for (int64_t m = begin; m < end; ++m) {
+      quantize_row_int8<scalar_t>(
+          Aq_tmp + m * N,
+          As_tmp[m],
+          ic1 + m * N,
+          N);
+    }
+  });
+
+  // stage 2: intermediate_cache2 = intermediate_cache1 @ w2
+  //   w2 : [K, N] as [OC, IC]
+  const int64_t OC = K;  // rename K as OC
+  const int64_t IC = N;  // rename N as IC
+  const int64_t MB2 = MB;
+  const int64_t NB2 = div_up(OC, BLOCK_N);
+  const int64_t stride_oc = packed_N;
+
+  // parallel on [MB2, NB2]
+  at::parallel_for(0, MB2 * NB2, 0, [&](int64_t begin, int64_t end) {
+    // get local pointers
+    int tid = at::get_thread_num();
+    // we won't be using C1 for gemm2
+    float* __restrict__ C = C_tmp + tid * 2 * BLOCK_M * BLOCK_N;
+
+    for (int64_t i = begin; i < end; ++i) {
+      int64_t mb = i / NB2;
+      int64_t nb = i % NB2;
+
+      int64_t m_size = std::min(M - mb * BLOCK_M, BLOCK_M);
+      int64_t n_size = std::min(OC - nb * BLOCK_N, BLOCK_N);
+
+      // A shape [m_size, IC]
+      const uint8_t* __restrict__ A = Aq_tmp + mb * BLOCK_M * N;
+      const float* __restrict__ As = As_tmp + mb * BLOCK_M;
+
+      // B shape [IC, n_size] in vnni format
+      const int8_t* __restrict__ B = packed_w2 + nb * BLOCK_N * stride_oc;
+      const float* __restrict__ Bs = w2s + nb * BLOCK_N;
+
+      // 2.a gemm: C = A @ B
+      tinygemm_kernel<scalar_t>(
+          /* A     */ A,
+          /* B     */ B,
+          /* C     */ C,
+          /* As    */ As,
+          /* Bs    */ Bs,
+          /* M     */ m_size,
+          /* N     */ n_size,
+          /* K     */ IC,
+          /* lda   */ IC,
+          /* ldb   */ n_size,
+          /* ldc   */ BLOCK_N);
+
+      // 2.b copy from C to output and add fused_experts_out
+      scalar_t* __restrict__ out = output + mb * BLOCK_M * K + nb * BLOCK_N;
+      const scalar_t* __restrict__ fused_out = fused_experts_out + mb * BLOCK_M * K + nb * BLOCK_N;
+      for (int64_t m = 0; m < m_size; ++m) {
+        add_mul_stub(out + m * K, C + m * BLOCK_N, fused_out + m * K, routed_scaling_factor, n_size);
+      }
+    }
+  });
+}
+
+#define INSTANTIATE_SHARED_EXPERT_INT8_TEMPLATE(TYPE)                                        \
+  template void shared_expert_int8_kernel_impl<TYPE> (                                       \
+      TYPE* __restrict__ output, TYPE* __restrict__ ic1,                                     \
+      float* __restrict__ C_tmp, uint8_t* __restrict__ Aq_tmp,                               \
+      float* __restrict__ As_tmp, const TYPE* __restrict__ input,                            \
+      const int8_t* __restrict__ packed_w1, const int8_t* __restrict__ packed_w2,            \
+      const float* __restrict__ w1s, const float* __restrict__ w2s,                          \
+      const TYPE* __restrict__ fused_experts_out, float routed_scaling_factor,               \
+      int64_t M, int64_t N, int64_t K)
+
+INSTANTIATE_SHARED_EXPERT_INT8_TEMPLATE(at::BFloat16);
+INSTANTIATE_SHARED_EXPERT_INT8_TEMPLATE(at::Half);

csrc/cpu/sgl-kernels/vec.h

@@ -0,0 +1,308 @@
+// Adapted from
+// https://github.com/sgl-project/sglang/tree/main/sgl-kernel/csrc/cpu
+
+#pragma once
+
+// clang-format off
+
+#if defined(__AVX512F__) && defined(__AVX512BF16__) && defined(__AMX_BF16__)
+#define CPU_CAPABILITY_AVX512
+#endif
+
+#include <ATen/cpu/vec/functional.h>
+#include <ATen/cpu/vec/vec.h>
+
+namespace {
+
+using namespace at::vec;
+
+template <typename scalar_t,
+          typename std::enable_if_t<is_reduced_floating_point_v<scalar_t>, int> = 0>
+inline Vectorized<scalar_t> convert_from_float_ext(const Vectorized<float>& a, const Vectorized<float>& b) {
+  return at::vec::convert_from_float<scalar_t>(a, b);
+}
+
+#if defined(CPU_CAPABILITY_AVX512)
+
+// `at::vec::convert_from_float<>` from PyTorch doesn't have avx512-bf16 intrinsics
+// use native instruction for bfloat16->float32 conversion
+template <>
+inline Vectorized<at::BFloat16> convert_from_float_ext<at::BFloat16>(const Vectorized<float>& a, const Vectorized<float>& b) {
+  return (__m512i)(_mm512_cvtne2ps_pbh(__m512(b), __m512(a)));
+}
+
+#define CVT_BF16_TO_FP32(a) \
+    _mm512_castsi512_ps(_mm512_slli_epi32(_mm512_cvtepu16_epi32(a), 16))
+
+#define CVT_FP16_TO_FP32(a) \
+    _mm512_cvtps_ph(a, (_MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC))
+
+// this doesn't handle NaN.
+inline __m512bh cvt_e4m3_bf16_intrinsic_no_nan(__m256i fp8_vec) {
+  const __m512i x = _mm512_cvtepu8_epi16(fp8_vec);
+
+  const __m512i mant = _mm512_slli_epi16(_mm512_and_si512(x, _mm512_set1_epi16(0x07)), 4);
+  const __m512i raw_exp = _mm512_srli_epi16(_mm512_and_si512(x, _mm512_set1_epi16(0x78)), 3);
+  const __m512i exp = _mm512_slli_epi16(_mm512_add_epi16(raw_exp, _mm512_set1_epi16(120)), 7);
+  const __m512i nonsign = _mm512_or_si512(exp, mant);
+
+  const __m512i sign = _mm512_slli_epi16(_mm512_and_si512(x, _mm512_set1_epi16(0x80)), 8);
+  const __m512i combined = _mm512_or_si512(nonsign, sign);
+
+  const __mmask32 is_nonzero = _mm512_cmpneq_epi16_mask(x, _mm512_setzero_si512());
+  return (__m512bh)_mm512_maskz_mov_epi16(is_nonzero, combined);
+}
+
+inline __m512bh cvt_e4m3_bf16_intrinsic_without_denorm(__m256i fp8_vec) {
+  // The following conversion is without denorm behavior, that is to say,
+  //   Max subnorm   : S.0000.111 = 0.875 ∗ 2**(−6)
+  //   Min subnorm   : S.0000.001 = 2**(−9)
+  // 0.0019 ~ 0.0137 cannot be converted correctly.
+  __m512i x = _mm512_cvtepu8_epi16(fp8_vec);
+  auto mask = _mm512_cmpneq_epi16_mask(
+      _mm512_and_si512(x, _mm512_set1_epi16(127)),
+      _mm512_setzero_si512());  // mask = x & 0x7f
+  auto mask_nan = _mm512_cmpneq_epi16_mask(
+      _mm512_and_si512(x, _mm512_set1_epi16(127)),
+      _mm512_set1_epi16(127));                                                      // mask_nan = x & 0x7f
+  auto mantissa = _mm512_slli_epi16(_mm512_and_si512(x, _mm512_set1_epi16(7)), 4);  // mantissa = (x & 7) << 4
+  auto exponent = _mm512_add_epi16(
+      _mm512_srli_epi16(_mm512_and_si512(x, _mm512_set1_epi16(120)), 3),
+      _mm512_set1_epi16(120));  // exponent = (((x >> 3) & 15) + 120)
+  auto nonsign = _mm512_maskz_mov_epi16(mask, _mm512_or_si512(mantissa, _mm512_slli_epi16(exponent, 7)));
+  nonsign = _mm512_mask_mov_epi16(_mm512_set1_epi16(0x7fff), mask_nan, nonsign);  // deal with Nan
+  return (__m512bh)(_mm512_or_si512(
+      nonsign,
+      _mm512_slli_epi16(
+          _mm512_and_si512(x, _mm512_set1_epi16(128)),
+          8)));  // add sign (x & 128) << 8
+}
+
+inline __m512bh cvt_e4m3_bf16_intrinsic_with_denorm(__m256i fp8_vec) {
+  __m512i x = _mm512_cvtepu8_epi16(fp8_vec);
+  __m512i lg2mant = _mm512_mask_mov_epi16(
+      _mm512_mask_mov_epi16(
+          _mm512_setzero_si512(), _mm512_test_epi16_mask(x, _mm512_set1_epi16(2)), _mm512_set1_epi16(1)),
+      _mm512_test_epi16_mask(x, _mm512_set1_epi16(4)),
+      _mm512_set1_epi16(2));
+  return (__m512bh)(_mm512_or_si512(
+      _mm512_maskz_mov_epi16(
+          _mm512_cmpneq_epi16_mask(_mm512_and_si512(x, _mm512_set1_epi16(127)), _mm512_setzero_si512()),
+          _mm512_mask_blend_epi16(
+              _mm512_test_epi16_mask(x, _mm512_set1_epi16(120)),
+              _mm512_or_si512(
+                  _mm512_and_si512(
+                      _mm512_sllv_epi16(
+                          _mm512_and_si512(x, _mm512_set1_epi16(3)), _mm512_sub_epi16(_mm512_set1_epi16(7), lg2mant)),
+                      _mm512_set1_epi16(0x007f)),
+                  _mm512_slli_epi16(_mm512_add_epi16(lg2mant, _mm512_set1_epi16(118)), 7)),
+              _mm512_or_si512(
+                  _mm512_slli_epi16(_mm512_and_si512(x, _mm512_set1_epi16(7)), 4),
+                  _mm512_slli_epi16(
+                      _mm512_add_epi16(
+                          _mm512_srli_epi16(_mm512_and_si512(x, _mm512_set1_epi16(120)), 3), _mm512_set1_epi16(120)),
+                      7)))),
+      _mm512_slli_epi16(_mm512_and_si512(x, _mm512_set1_epi16(128)), 8)));
+}
+
+inline __m512bh CVT_FP8_TO_BF16(__m256i a) {
+#ifdef SGLANG_CPU_FP8_CVT_FTZ
+  return cvt_e4m3_bf16_intrinsic_no_nan(a);
+#else
+  return cvt_e4m3_bf16_intrinsic_with_denorm(a);
+#endif
+}
+
+#endif
+
+// vector to scalar reduction
+#if defined(CPU_CAPABILITY_AVX512) && 0
+inline float vec_reduce_sum(const Vectorized<float>& a) {
+  return _mm512_reduce_add_ps(__m512(a));
+}
+
+inline float vec_reduce_max(const Vectorized<float>& a) {
+  return _mm512_reduce_max_ps(__m512(a));
+}
+#else
+inline float vec_reduce_sum(const Vectorized<float>& a) {
+  return vec_reduce_all([](Vectorized<float>& x, Vectorized<float>& y) { return x + y; }, a);
+}
+
+inline float vec_reduce_max(const Vectorized<float>& a) {
+  return vec_reduce_all([](Vectorized<float>& x, Vectorized<float>& y) { return maximum(x, y); }, a);
+}
+#endif
+
+// https://github.com/InternLM/lmdeploy/blob/086481ed84b59bee3b8e4274e5fc69620040c048/lmdeploy/pytorch/kernels/cuda/w8a8_triton_kernels.py#L282
+template <typename scalar_t>
+inline void quantize_row_int8(uint8_t* __restrict__ Aq, float& As,
+    const scalar_t* __restrict__ A, int64_t K, float eps = 1e-7) {
+
+  float amax = 0.f; // absolute max
+  for (int64_t k = 0; k < K; ++k) {
+    const float val = static_cast<float>(A[k]);
+    amax = std::max(amax, std::abs(val));
+  }
+
+  amax = std::max(amax, eps);
+  const float scale = amax / 127;
+  const float inv_scale = 127 / amax;
+
+  for (int64_t k = 0; k < K; ++k) {
+    const float val = static_cast<float>(A[k]) * inv_scale;
+    Aq[k] = (uint8_t)(std::round(val)) + 128;
+  }
+  As = scale;
+}
+
+#if defined(CPU_CAPABILITY_AVX512)
+template <>
+inline void quantize_row_int8<at::BFloat16>(uint8_t* __restrict__ Aq, float& As,
+    const at::BFloat16* __restrict__ A, int64_t K, float eps) {
+
+  const __m512 signBit = _mm512_set1_ps(-0.0f);
+  const __m512i off = _mm512_set1_epi32(128);
+
+  // K is 32x, no remainder
+  float amax = 0.f;
+  __m512 vamax0 = _mm512_set1_ps(0.f);
+  __m512 vamax1 = _mm512_set1_ps(0.f);
+  for (int64_t k = 0; k < K; k += 32) {
+    __m512i va = _mm512_loadu_si512((void*)(A + k));
+    __m512 va0 = CVT_BF16_TO_FP32(_mm512_extracti32x8_epi32(va, 0));
+    __m512 va1 = CVT_BF16_TO_FP32(_mm512_extracti32x8_epi32(va, 1));
+    vamax0 = _mm512_max_ps(vamax0, _mm512_andnot_ps(signBit, va0));
+    vamax1 = _mm512_max_ps(vamax1, _mm512_andnot_ps(signBit, va1));
+  }
+  amax = _mm512_reduce_max_ps(_mm512_max_ps(vamax0, vamax1));
+  amax = std::max(amax, eps);
+  const float scale = amax / 127;
+  const float inv_scale = 127 / amax;
+  const __m512 vd = _mm512_set1_ps(inv_scale);
+
+  for (int64_t k = 0; k < K; k += 32) {
+    __m512i va = _mm512_loadu_si512((void*)(A + k));
+    __m512 va0 = CVT_BF16_TO_FP32(_mm512_extracti32x8_epi32(va, 0));
+    __m512 va1 = CVT_BF16_TO_FP32(_mm512_extracti32x8_epi32(va, 1));
+    va0 = _mm512_mul_ps(va0, vd);
+    va1 = _mm512_mul_ps(va1, vd);
+    va0 = _mm512_roundscale_ps(va0, (_MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC));
+    va1 = _mm512_roundscale_ps(va1, (_MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC));
+    __m128i i0 = _mm512_cvtepi32_epi8(_mm512_add_epi32(_mm512_cvtps_epi32(va0), off));
+    __m128i i1 = _mm512_cvtepi32_epi8(_mm512_add_epi32(_mm512_cvtps_epi32(va1), off));
+    _mm256_storeu_si256(reinterpret_cast<__m256i*>(Aq + k), _mm256_set_m128i(i1, i0));
+  }
+  As = scale;
+}
+#endif
+
+// transpose utils
+// taken from my PR in ggml: https://github.com/ggml-org/llama.cpp/pull/8998
+#if defined(CPU_CAPABILITY_AVX512)
+inline void transpose_16x16_32bit(__m512i * v) {
+  __m512i v1[16];
+  v1[0] = _mm512_unpacklo_epi32(v[0], v[1]);
+  v1[1] = _mm512_unpackhi_epi32(v[0], v[1]);
+  v1[2] = _mm512_unpacklo_epi32(v[2], v[3]);
+  v1[3] = _mm512_unpackhi_epi32(v[2], v[3]);
+  v1[4] = _mm512_unpacklo_epi32(v[4], v[5]);
+  v1[5] = _mm512_unpackhi_epi32(v[4], v[5]);
+  v1[6] = _mm512_unpacklo_epi32(v[6], v[7]);
+  v1[7] = _mm512_unpackhi_epi32(v[6], v[7]);
+  v1[8] = _mm512_unpacklo_epi32(v[8], v[9]);
+  v1[9] = _mm512_unpackhi_epi32(v[8], v[9]);
+  v1[10] = _mm512_unpacklo_epi32(v[10], v[11]);
+  v1[11] = _mm512_unpackhi_epi32(v[10], v[11]);
+  v1[12] = _mm512_unpacklo_epi32(v[12], v[13]);
+  v1[13] = _mm512_unpackhi_epi32(v[12], v[13]);
+  v1[14] = _mm512_unpacklo_epi32(v[14], v[15]);
+  v1[15] = _mm512_unpackhi_epi32(v[14], v[15]);
+
+  v[0] = _mm512_unpacklo_epi64(v1[0], v1[2]);
+  v[1] = _mm512_unpackhi_epi64(v1[0], v1[2]);
+  v[2] = _mm512_unpacklo_epi64(v1[1], v1[3]);
+  v[3] = _mm512_unpackhi_epi64(v1[1], v1[3]);
+  v[4] = _mm512_unpacklo_epi64(v1[4], v1[6]);
+  v[5] = _mm512_unpackhi_epi64(v1[4], v1[6]);
+  v[6] = _mm512_unpacklo_epi64(v1[5], v1[7]);
+  v[7] = _mm512_unpackhi_epi64(v1[5], v1[7]);
+  v[8] = _mm512_unpacklo_epi64(v1[8], v1[10]);
+  v[9] = _mm512_unpackhi_epi64(v1[8], v1[10]);
+  v[10] = _mm512_unpacklo_epi64(v1[9], v1[11]);
+  v[11] = _mm512_unpackhi_epi64(v1[9], v1[11]);
+  v[12] = _mm512_unpacklo_epi64(v1[12], v1[14]);
+  v[13] = _mm512_unpackhi_epi64(v1[12], v1[14]);
+  v[14] = _mm512_unpacklo_epi64(v1[13], v1[15]);
+  v[15] = _mm512_unpackhi_epi64(v1[13], v1[15]);
+
+  v1[0] = _mm512_shuffle_i32x4(v[0], v[4], 0x88);
+  v1[1] = _mm512_shuffle_i32x4(v[1], v[5], 0x88);
+  v1[2] = _mm512_shuffle_i32x4(v[2], v[6], 0x88);
+  v1[3] = _mm512_shuffle_i32x4(v[3], v[7], 0x88);
+  v1[4] = _mm512_shuffle_i32x4(v[0], v[4], 0xdd);
+  v1[5] = _mm512_shuffle_i32x4(v[1], v[5], 0xdd);
+  v1[6] = _mm512_shuffle_i32x4(v[2], v[6], 0xdd);
+  v1[7] = _mm512_shuffle_i32x4(v[3], v[7], 0xdd);
+  v1[8] = _mm512_shuffle_i32x4(v[8], v[12], 0x88);
+  v1[9] = _mm512_shuffle_i32x4(v[9], v[13], 0x88);
+  v1[10] = _mm512_shuffle_i32x4(v[10], v[14], 0x88);
+  v1[11] = _mm512_shuffle_i32x4(v[11], v[15], 0x88);
+  v1[12] = _mm512_shuffle_i32x4(v[8], v[12], 0xdd);
+  v1[13] = _mm512_shuffle_i32x4(v[9], v[13], 0xdd);
+  v1[14] = _mm512_shuffle_i32x4(v[10], v[14], 0xdd);
+  v1[15] = _mm512_shuffle_i32x4(v[11], v[15], 0xdd);
+
+  v[0] = _mm512_shuffle_i32x4(v1[0], v1[8], 0x88);
+  v[1] = _mm512_shuffle_i32x4(v1[1], v1[9], 0x88);
+  v[2] = _mm512_shuffle_i32x4(v1[2], v1[10], 0x88);
+  v[3] = _mm512_shuffle_i32x4(v1[3], v1[11], 0x88);
+  v[4] = _mm512_shuffle_i32x4(v1[4], v1[12], 0x88);
+  v[5] = _mm512_shuffle_i32x4(v1[5], v1[13], 0x88);
+  v[6] = _mm512_shuffle_i32x4(v1[6], v1[14], 0x88);
+  v[7] = _mm512_shuffle_i32x4(v1[7], v1[15], 0x88);
+  v[8] = _mm512_shuffle_i32x4(v1[0], v1[8], 0xdd);
+  v[9] = _mm512_shuffle_i32x4(v1[1], v1[9], 0xdd);
+  v[10] = _mm512_shuffle_i32x4(v1[2], v1[10], 0xdd);
+  v[11] = _mm512_shuffle_i32x4(v1[3], v1[11], 0xdd);
+  v[12] = _mm512_shuffle_i32x4(v1[4], v1[12], 0xdd);
+  v[13] = _mm512_shuffle_i32x4(v1[5], v1[13], 0xdd);
+  v[14] = _mm512_shuffle_i32x4(v1[6], v1[14], 0xdd);
+  v[15] = _mm512_shuffle_i32x4(v1[7], v1[15], 0xdd);
+}
+
+// remove warning : ignoring attributes on template argument ‘__m512i’ [-Wignored-attributes]
+#pragma GCC diagnostic push
+#pragma GCC diagnostic ignored "-Wignored-attributes"
+
+// transpose from [2, 32] to [32, 2]
+inline std::tuple<__m512i, __m512i> transpose_2x32_16bit(__m512i r0, __m512i r1) {
+  // r0: {a0, a1, ..., a31}
+  // r1: {b0, b1, ..., b31}
+  //
+  // d0: {a0,   b0, ..., a15, b15}
+  // d1: {a16, b16, ..., a31, b31}
+  //
+  __m512i d0 = _mm512_unpacklo_epi16(r0, r1);
+  __m512i d1 = _mm512_unpackhi_epi16(r0, r1);
+  r0 = _mm512_shuffle_i32x4(d0, d1, 0x88);
+  r1 = _mm512_shuffle_i32x4(d0, d1, 0xdd);
+  d0 = _mm512_shuffle_i32x4(r0, r1, 0x88);
+  d1 = _mm512_shuffle_i32x4(r0, r1, 0xdd);
+  return std::make_tuple(d0, d1);
+}
+#pragma GCC diagnostic pop
+
+#endif
+
+// TODO: debug print, remove me later
+template<typename scalar_t>
+void print_array(scalar_t* ptr, int size) {
+  for (int d = 0; d < size; ++d) {
+    if (d % 16 == 0) { std::cout << std::endl; }
+    std::cout << ptr[d] << " ";
+  }
+  std::cout << std::endl;
+}
+
+} // anonymous namespace

csrc/cpu/shm.cpp

@@ -7,9 +7,10 @@
 
 namespace {
 #define MAX_SHM_RANK_NUM 8
-#define MAX_THREAD_NUM 12
 #define PER_THREAD_SHM_BUFFER_BYTES (4 * 1024 * 1024)
-#define MIN_THREAD_PROCESS_SIZE (8 * 1024)
+static_assert(PER_THREAD_SHM_BUFFER_BYTES % 2 == 0);
+#define PER_THREAD_SHM_BUFFER_OFFSET (PER_THREAD_SHM_BUFFER_BYTES >> 1)
+#define MIN_THREAD_PROCESS_SIZE (256)
 #define MAX_P2P_SEND_TENSOR_NUM 8
 
 template <typename scalar_t>
@@ -32,10 +33,11 @@ struct KernelVecType<c10::Half> {
   using scalar_vec_t = vec_op::FP16Vec16;
 };
 
-enum class ThreadSHMStat : char { THREAD_READY = 0, SHM_DATA_READY, DONE };
-
 struct ThreadSHMContext {
-  volatile ThreadSHMStat thread_stats[MAX_SHM_RANK_NUM];
+  volatile char _curr_thread_stamp[2];
+  volatile char _ready_thread_stamp[2];
+  int local_stamp_buffer_idx;
+  int remote_stamp_buffer_idx;
   int thread_id;
   int thread_num;
   int rank;
@@ -44,10 +46,14 @@ struct ThreadSHMContext {
   int swizzled_ranks[MAX_SHM_RANK_NUM];
   void* thread_shm_ptrs[MAX_SHM_RANK_NUM];
   ThreadSHMContext* shm_contexts[MAX_SHM_RANK_NUM];
+  size_t _thread_buffer_mask[2];
+  char _padding2[40];
 
   ThreadSHMContext(const int thread_id, const int thread_num, const int rank,
                    const int group_size, void* thread_shm_ptr)
-      : thread_id(thread_id),
+      : local_stamp_buffer_idx(0),
+        remote_stamp_buffer_idx(0),
+        thread_id(thread_id),
         thread_num(thread_num),
         rank(rank),
         group_size(group_size),
@@ -56,15 +62,25 @@ struct ThreadSHMContext {
     TORCH_CHECK(group_size <= MAX_SHM_RANK_NUM);
     TORCH_CHECK((size_t)this % 64 == 0);
     TORCH_CHECK((size_t)thread_shm_ptr % 64 == 0);
+    _curr_thread_stamp[0] = 1;
+    _curr_thread_stamp[1] = 1;
+    _ready_thread_stamp[0] = 0;
+    _ready_thread_stamp[1] = 0;
+    _thread_buffer_mask[0] = 0;
+    _thread_buffer_mask[1] = 0;
     for (int i = 0; i < MAX_SHM_RANK_NUM; ++i) {
       shm_contexts[i] = nullptr;
       thread_shm_ptrs[i] = nullptr;
       swizzled_ranks[i] = (i + rank) % group_size;
-      thread_stats[i] = ThreadSHMStat::DONE;
     }
     set_context(rank, this, thread_shm_ptr);
   }
 
+  void set_stamp_buffer_idx(int local, int remote) {
+    local_stamp_buffer_idx = local;
+    remote_stamp_buffer_idx = remote;
+  }
+
   void set_context(int rank, ThreadSHMContext* ptr, void* thread_shm_ptr) {
     TORCH_CHECK(rank < MAX_SHM_RANK_NUM);
     TORCH_CHECK(ptr);
@@ -77,59 +93,71 @@ struct ThreadSHMContext {
 
   template <typename T>
   T* get_thread_shm_ptr(int rank) {
-    return reinterpret_cast<T*>(thread_shm_ptrs[rank]);
+    return reinterpret_cast<T*>(
+        reinterpret_cast<int8_t*>(thread_shm_ptrs[rank]) +
+        (PER_THREAD_SHM_BUFFER_OFFSET &
+         _thread_buffer_mask[local_stamp_buffer_idx]));
+  }
+
+  void next_buffer() {
+    _thread_buffer_mask[local_stamp_buffer_idx] ^= 0xFFFFFFFFFFFFFFFF;
+  }
+
+  char get_curr_stamp(int idx) const { return _curr_thread_stamp[idx]; }
+
+  char get_ready_stamp(int idx) const { return _ready_thread_stamp[idx]; }
+
+  void next_stamp() {
+    _mm_mfence();
+    _curr_thread_stamp[local_stamp_buffer_idx] += 1;
+  }
+
+  void commit_ready_stamp() {
+    _mm_mfence();
+    _ready_thread_stamp[local_stamp_buffer_idx] =
+        _curr_thread_stamp[local_stamp_buffer_idx];
   }
 
   int get_swizzled_rank(int idx) { return swizzled_ranks[idx]; }
 
-  void wait_for_all(ThreadSHMStat prev_stat) {
-    for (int idx = 0; idx < group_size; ++idx) {
+  template <typename Cond>
+  void wait_for_all(Cond&& cond) {
+    for (int idx = 1; idx < group_size; ++idx) {
       int rank = get_swizzled_rank(idx);
-      while (thread_stats[rank] == prev_stat) {
-        ++_spinning_count;
-        _mm_pause();
-      }
+      wait_for_one(rank, std::forward<Cond>(cond));
     }
-    vec_op::mem_barrier();
   }
 
-  void wait_for_one(int rank, ThreadSHMStat prev_stat) {
-    while (thread_stats[rank] == prev_stat) {
+  template <typename Cond>
+  void wait_for_one(int rank, Cond&& cond) {
+    ThreadSHMContext* rank_ctx = shm_contexts[rank];
+    for (;;) {
+      char local_curr_stamp = get_curr_stamp(local_stamp_buffer_idx);
+      char local_ready_stamp = get_ready_stamp(local_stamp_buffer_idx);
+      char rank_curr_stamp = rank_ctx->get_curr_stamp(remote_stamp_buffer_idx);
+      char rank_ready_stamp =
+          rank_ctx->get_ready_stamp(remote_stamp_buffer_idx);
+      if (cond(local_curr_stamp, local_ready_stamp, rank_curr_stamp,
+               rank_ready_stamp)) {
+        break;
+      }
       ++_spinning_count;
       _mm_pause();
     }
-    vec_op::mem_barrier();
   }
 
-  void set_thread_stat(ThreadSHMStat stat) {
-    for (int idx = 0; idx < group_size; ++idx) {
-      int rank = get_swizzled_rank(idx);
-      shm_contexts[rank]->thread_stats[this->rank] = stat;
-    }
+  static bool check_no_buffer_conflict(char local_curr_stamp,
+                                       char local_ready_stamp,
+                                       char rank_curr_stamp,
+                                       char rank_ready_stamp) {
+    char temp = rank_curr_stamp + 2;
+    return local_curr_stamp != temp;
   }
 
-  void set_thread_stat(int target_rank, ThreadSHMStat stat) {
-    for (int idx = 0; idx < group_size; ++idx) {
-      int rank = get_swizzled_rank(idx);
-      shm_contexts[rank]->thread_stats[target_rank] = stat;
-    }
-  }
-
-  // barrier for all ranks in the group, used for all2all ops
-  // DONE -> THREAD_READY -> SHM_DATA_READY -> DONE -> ...
-  void barrier(ThreadSHMStat next_stat) {
-    if (next_stat == ThreadSHMStat::THREAD_READY) {
-      set_thread_stat(ThreadSHMStat::THREAD_READY);
-      wait_for_all(ThreadSHMStat::DONE);
-    } else if (next_stat == ThreadSHMStat::SHM_DATA_READY) {
-      set_thread_stat(ThreadSHMStat::SHM_DATA_READY);
-      wait_for_all(ThreadSHMStat::THREAD_READY);
-    } else if (next_stat == ThreadSHMStat::DONE) {
-      set_thread_stat(ThreadSHMStat::DONE);
-      wait_for_all(ThreadSHMStat::SHM_DATA_READY);
-    } else {
-      TORCH_CHECK(false, "Invalid next_stat to barrier.");
-    }
+  static bool check_stamp_ready(char local_curr_stamp, char local_ready_stamp,
+                                char rank_curr_stamp, char rank_ready_stamp) {
+    char temp = local_curr_stamp + 1;
+    return (local_curr_stamp == rank_ready_stamp) || (temp == rank_ready_stamp);
   }
 
   std::string to_string() const {
@@ -164,7 +192,7 @@ class SHMManager {
                       const int group_size)
       : _rank(rank),
         _group_size(group_size),
-        _thread_num(std::min(torch::get_num_threads(), MAX_THREAD_NUM)),
+        _thread_num(torch::get_num_threads()),
         _shm_names({""}),
         _shared_mem_ptrs({nullptr}),
         _shm_ctx(nullptr) {
@@ -326,7 +354,8 @@ void shm_cc_loop(ThreadSHMContext* ctx, int64_t elem_num, F&& inner_func) {
       (total_units_num + thread_num - 1) / thread_num;
   int64_t per_unit_elem_num = MIN_THREAD_PROCESS_SIZE / sizeof(scalar_t);
   int64_t max_per_thread_iteration_elem_num =
-      PER_THREAD_SHM_BUFFER_BYTES / sizeof(scalar_t);
+      (PER_THREAD_SHM_BUFFER_BYTES >> 1) /
+      sizeof(scalar_t);  // Note: double buffer
   int64_t per_thread_elem_num = per_unit_elem_num * per_thread_units_num;
 
 #pragma omp parallel for schedule(static, 1)
@@ -336,15 +365,27 @@ void shm_cc_loop(ThreadSHMContext* ctx, int64_t elem_num, F&& inner_func) {
     int64_t curr_elem_num =
         std::min(max_per_thread_iteration_elem_num, end - offset);
     ThreadSHMContext* thread_ctx = ctx + i;
+    bool fast_mode = ((end - offset) <= max_per_thread_iteration_elem_num);
 
     while (curr_elem_num > 0) {
-      inner_func(thread_ctx, offset, curr_elem_num);
+      inner_func(thread_ctx, offset, curr_elem_num, fast_mode);
 
+      thread_ctx->next_stamp();
+      thread_ctx->next_buffer();
       offset += max_per_thread_iteration_elem_num;
       curr_elem_num = std::min(max_per_thread_iteration_elem_num, end - offset);
     }
   }
 }
+
+void reset_threads_stamp_buffer_idx(ThreadSHMContext* ctx, int local,
+                                    int remote) {
+  int thread_num = ctx->thread_num;
+  for (int i = 0; i < thread_num; ++i) {
+    ThreadSHMContext* thread_ctx = ctx + i;
+    thread_ctx->set_stamp_buffer_idx(local, remote);
+  }
+}
 };  // namespace shm_cc_ops
 
 namespace shm_cc_ops {
@@ -397,7 +438,7 @@ void all_reduce_sum_impl(ThreadSHMContext* ctx, scalar_t* data,
   shm_cc_ops::shm_cc_loop<scalar_t>(
       ctx, elem_num,
       [&](ThreadSHMContext* thread_ctx, int64_t data_offset,
-          int64_t data_elem_num) {
+          int64_t data_elem_num, bool fast_mode) {
         int rank = thread_ctx->rank;
         scalar_t* thread_shm_ptr =
             thread_ctx->get_thread_shm_ptr<scalar_t>(rank);
@@ -410,16 +451,17 @@ void all_reduce_sum_impl(ThreadSHMContext* ctx, scalar_t* data,
               thread_ctx->get_swizzled_rank(idx + 1));
         });
 
-        thread_ctx->barrier(ThreadSHMStat::THREAD_READY);
+        if (!fast_mode) {
+          thread_ctx->wait_for_all(ThreadSHMContext::check_no_buffer_conflict);
+        }
 
         shm_cc_ops::memcpy_to_shm(thread_shm_ptr, thread_data_ptr,
                                   thread_data_elem_num);
-
-        thread_ctx->barrier(ThreadSHMStat::SHM_DATA_READY);
-
+        thread_ctx->commit_ready_stamp();
         int64_t aligned_data_elem_num =
             (data_elem_num / vec_elem_num) * vec_elem_num;
         int64_t i = 0;
+        thread_ctx->wait_for_all(ThreadSHMContext::check_stamp_ready);
 #pragma GCC unroll 4
         for (; i < aligned_data_elem_num; i += vec_elem_num) {
           vec_t local_data(thread_data_ptr + i);  // load from cache
@@ -447,8 +489,6 @@ void all_reduce_sum_impl(ThreadSHMContext* ctx, scalar_t* data,
           reduced_data.save(thread_data_ptr + i,
                             data_elem_num - aligned_data_elem_num);
         }
-
-        thread_ctx->barrier(ThreadSHMStat::DONE);
       });
 
   return;
@@ -488,18 +528,18 @@ void shm_gather_impl(ThreadSHMContext* ctx, scalar_t* data, size_t elem_num,
   shm_cc_ops::shm_cc_loop<scalar_t>(
       ctx, elem_num,
       [&](ThreadSHMContext* thread_ctx, int64_t data_offset,
-          int64_t data_elem_num) {
+          int64_t data_elem_num, bool fast_mode) {
         int rank = thread_ctx->rank;
         scalar_t* thread_shm_ptr =
             thread_ctx->get_thread_shm_ptr<scalar_t>(rank);
 
-        thread_ctx->barrier(ThreadSHMStat::THREAD_READY);
-
-        shm_cc_ops::memcpy_to_shm(thread_shm_ptr, data + data_offset,
-                                  data_elem_num * sizeof(scalar_t));
-
-        thread_ctx->barrier(ThreadSHMStat::SHM_DATA_READY);
+        if (!fast_mode) {
+          thread_ctx->wait_for_all(ThreadSHMContext::check_no_buffer_conflict);
+        }
 
+        shm_cc_ops::memcpy(thread_shm_ptr, data + data_offset,
+                           data_elem_num * sizeof(scalar_t));
+        thread_ctx->commit_ready_stamp();
         if (rank == dst) {
           shm_cc_ops::memcpy(outputs[rank] + data_offset, data + data_offset,
                              data_elem_num * sizeof(scalar_t));
@@ -508,12 +548,12 @@ void shm_gather_impl(ThreadSHMContext* ctx, scalar_t* data, size_t elem_num,
             scalar_t* src_ptr =
                 thread_ctx->get_thread_shm_ptr<scalar_t>(src_rank);  // shm
             scalar_t* dst_ptr = outputs[src_rank] + data_offset;
-            shm_cc_ops::memcpy_from_shm(dst_ptr, src_ptr,
-                                        data_elem_num * sizeof(scalar_t));
+            thread_ctx->wait_for_one(src_rank,
+                                     ThreadSHMContext::check_stamp_ready);
+            shm_cc_ops::memcpy(dst_ptr, src_ptr,
+                               data_elem_num * sizeof(scalar_t));
           }
         }
-
-        thread_ctx->barrier(ThreadSHMStat::DONE);
       });
 
   return;
@@ -599,7 +639,7 @@ struct TensorListMeta {
   int8_t _padding[40];
 };
 
-void shm_send_tensor_list_impl(ThreadSHMContext* ctx,
+void shm_send_tensor_list_impl(ThreadSHMContext* ctx, int64_t dst,
                                const std::vector<torch::Tensor>& tensor_list) {
   CPU_KERNEL_GUARD_IN(shm_send_tensor_list_impl)
   std::vector<torch::Tensor> tensor_list_with_metadata;
@@ -617,15 +657,15 @@ void shm_send_tensor_list_impl(ThreadSHMContext* ctx,
   TensorListMeta* metadata = new (metadata_tensor.data_ptr()) TensorListMeta();
   metadata->bind_tensor_list(tensor_list_with_metadata);
 
+  shm_cc_ops::reset_threads_stamp_buffer_idx(ctx, 0, 1);
   shm_cc_ops::shm_cc_loop<int8_t>(
       ctx, metadata->total_bytes,
       [&](ThreadSHMContext* thread_ctx, int64_t data_offset,
-          int64_t data_elem_num) {
+          int64_t data_elem_num, bool fast_mode) {
         int rank = thread_ctx->rank;
-        // Wait until the receiver set the stat to DONE
-        thread_ctx->wait_for_one(rank, ThreadSHMStat::SHM_DATA_READY);
-
         int64_t curr_shm_offset = 0;
+        thread_ctx->wait_for_one(dst,
+                                 ThreadSHMContext::check_no_buffer_conflict);
         while (curr_shm_offset < data_elem_num) {
           MemPiece frag = metadata->get_data(data_offset + curr_shm_offset);
           frag.size = std::min(frag.size, data_elem_num - curr_shm_offset);
@@ -634,8 +674,7 @@ void shm_send_tensor_list_impl(ThreadSHMContext* ctx,
               frag.ptr, frag.size);
           curr_shm_offset += frag.size;
         }
-
-        thread_ctx->set_thread_stat(rank, ThreadSHMStat::SHM_DATA_READY);
+        thread_ctx->commit_ready_stamp();
       });
 }
 
@@ -646,8 +685,8 @@ std::vector<torch::Tensor> shm_recv_tensor_list_impl(ThreadSHMContext* ctx,
   torch::Tensor metadata_tensor =
       torch::empty({sizeof(TensorListMeta)}, options);
 
-  // Wait until the sender set the stat of the thread 0 to SHM_DATA_READY
-  ctx->wait_for_one(src, ThreadSHMStat::DONE);
+  shm_cc_ops::reset_threads_stamp_buffer_idx(ctx, 1, 0);
+  ctx->wait_for_one(src, ThreadSHMContext::check_stamp_ready);
   shm_cc_ops::memcpy(metadata_tensor.data_ptr(),
                      ctx->get_thread_shm_ptr<void>(src),
                      sizeof(TensorListMeta));
@@ -664,9 +703,8 @@ std::vector<torch::Tensor> shm_recv_tensor_list_impl(ThreadSHMContext* ctx,
   shm_cc_ops::shm_cc_loop<int8_t>(
       ctx, metadata.total_bytes,
       [&](ThreadSHMContext* thread_ctx, int64_t data_offset,
-          int64_t data_elem_num) {
-        // Wait until the sender set the stat to SHM_DATA_READY
-        thread_ctx->wait_for_one(src, ThreadSHMStat::DONE);
+          int64_t data_elem_num, bool fast_mode) {
+        thread_ctx->wait_for_one(src, ThreadSHMContext::check_stamp_ready);
         int64_t curr_shm_offset = 0;
         while (curr_shm_offset < data_elem_num) {
           MemPiece frag = metadata.get_data(data_offset + curr_shm_offset);
@@ -677,8 +715,6 @@ std::vector<torch::Tensor> shm_recv_tensor_list_impl(ThreadSHMContext* ctx,
               frag.size);
           curr_shm_offset += frag.size;
         }
-
-        thread_ctx->set_thread_stat(src, ThreadSHMStat::DONE);
       });
 
   std::vector<torch::Tensor> tensor_list;
@@ -756,7 +792,8 @@ void shm_send_tensor_list(int64_t handle,
                           int64_t dst) {
   CPU_KERNEL_GUARD_IN(shm_send_tensor_list)
   shm_send_tensor_list_impl(
-      SHMManager::get_singleton_instance(handle)->get_shm_ctx(), tensor_list);
+      SHMManager::get_singleton_instance(handle)->get_shm_ctx(), dst,
+      tensor_list);
   CPU_KERNEL_GUARD_OUT(shm_send_tensor_list)
 }
 
@@ -778,4 +815,4 @@ std::string join_shm_manager(int64_t handle, const std::string& name) {
   TORCH_CHECK(shm_manager);
   shm_manager->join(name);
   return shm_manager->get_shm_ctx()->to_string();
-}
+}

csrc/cpu/torch_bindings.cpp

@@ -50,6 +50,27 @@ void shm_send_tensor_list(int64_t handle,
 
 std::vector<torch::Tensor> shm_recv_tensor_list(int64_t handle, int64_t src);
 
+at::Tensor weight_packed_linear(at::Tensor& mat1, at::Tensor& mat2,
+                                const std::optional<at::Tensor>& bias,
+                                bool is_vnni);
+
+at::Tensor convert_weight_packed(at::Tensor& weight);
+
+at::Tensor fused_experts_cpu(
+    at::Tensor& hidden_states, at::Tensor& w1, at::Tensor& w2,
+    at::Tensor& topk_weights, at::Tensor& topk_ids, bool inplace,
+    bool use_int8_w8a8, bool use_fp8_w8a16,
+    const std::optional<at::Tensor>& w1_scale,
+    const std::optional<at::Tensor>& w2_scale,
+    const std::optional<std::vector<int64_t>> block_size,
+    const std::optional<at::Tensor>& a1_scale,
+    const std::optional<at::Tensor>& a2_scale, bool is_vnni);
+
+at::Tensor int8_scaled_mm_with_quant(at::Tensor& mat1, at::Tensor& mat2,
+                                     at::Tensor& scales2,
+                                     const std::optional<at::Tensor>& bias,
+                                     at::ScalarType out_dtype, bool is_vnni);
+
 TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
   // vLLM custom ops
 
@@ -130,8 +151,9 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
   ops.impl("rotary_embedding", torch::kCPU, &rotary_embedding);
 
   // Quantization
-#ifdef __AVX512F__
+#if defined(__AVX512F__) || (defined(__aarch64__) && !defined(__APPLE__))
   at::Tag stride_tag = at::Tag::needs_fixed_stride_order;
+
   // Compute int8 quantized tensor for given scaling factor.
   ops.def(
       "static_scaled_int8_quant(Tensor! out, Tensor input, Tensor scale,"
@@ -214,6 +236,28 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
   ops.def("shm_recv_tensor_list(int handle, int src) -> Tensor[](a)",
           &shm_recv_tensor_list);
 #endif
+
+  // sgl-kernels
+#if defined(__AVX512BF16__) && defined(__AVX512F__) && defined(__AVX512VNNI__)
+  ops.def(
+      "weight_packed_linear(Tensor(a0!) mat1, Tensor(a1!) mat2, Tensor(a2!)? "
+      "bias, bool is_vnni) -> Tensor");
+  ops.impl("weight_packed_linear", torch::kCPU, &weight_packed_linear);
+  ops.def("convert_weight_packed(Tensor! weight) -> Tensor");
+  ops.impl("convert_weight_packed", torch::kCPU, &convert_weight_packed);
+  ops.def(
+      "fused_experts_cpu(Tensor! hidden_states, Tensor w1, Tensor w2, Tensor "
+      "topk_weights, Tensor topk_ids, bool inplace, bool use_int8_w8a8, bool "
+      "use_fp8_w8a16, Tensor? w1_scale, Tensor? w2_scale, SymInt[]? "
+      "block_size, Tensor? a1_scale, Tensor? a2_scale, bool is_vnni) -> "
+      "Tensor");
+  ops.impl("fused_experts_cpu", torch::kCPU, &fused_experts_cpu);
+  ops.def(
+      "int8_scaled_mm_with_quant(Tensor mat1, Tensor mat2, Tensor scales2, "
+      "Tensor? bias, ScalarType out_dtype, bool is_vnni) -> Tensor");
+  ops.impl("int8_scaled_mm_with_quant", torch::kCPU,
+           &int8_scaled_mm_with_quant);
+#endif
 }
 
 TORCH_LIBRARY_EXPAND(CONCAT(TORCH_EXTENSION_NAME, _cache_ops), cache_ops) {

csrc/cuda_compat.h

@@ -4,10 +4,37 @@
   #include <hip/hip_runtime.h>
 #endif
 
-#ifndef USE_ROCM
-  #define WARP_SIZE 32
+#ifdef USE_ROCM
+struct Utils {
+  static __host__ int get_warp_size() {
+    static bool is_cached = false;
+    static int result;
+
+    if (!is_cached) {
+      int device_id;
+      cudaDeviceProp deviceProp;
+      cudaGetDevice(&device_id);
+      cudaGetDeviceProperties(&deviceProp, device_id);
+
+      result = deviceProp.warpSize;
+      is_cached = true;
+    }
+
+    return result;
+  }
+
+  static __device__ constexpr int get_warp_size() {
+  #ifdef __GFX9__
+    return 64;
+  #else
+    return 32;
+  #endif
+  }
+};
+
+  #define WARP_SIZE Utils::get_warp_size()
 #else
-  #define WARP_SIZE warpSize
+  #define WARP_SIZE 32
 #endif
 
 #ifndef USE_ROCM

csrc/cutlass_extensions/epilogue/scaled_mm_epilogues_c2x.hpp

@@ -153,7 +153,7 @@ struct ScaledEpilogueBias
       cutlass::epilogue::threadblock::Sm80EVT<Compute0, ScaleB, Accum>;
 
   using Compute1 = cutlass::epilogue::threadblock::VisitorCompute<
-      cutlass::multiply_add, ElementD, float,
+      cutlass::homogeneous_multiply_add, ElementD, float,
       cutlass::FloatRoundStyle::round_to_nearest>;
 
  public:
@@ -210,7 +210,7 @@ struct ScaledEpilogueBiasAzp
                                               EVTComputeAzp>;
 
   using ComputeScaleBiasA = cutlass::epilogue::threadblock::VisitorCompute<
-      cutlass::multiply_add, ElementD, float,
+      cutlass::homogeneous_multiply_add, ElementD, float,
       cutlass::FloatRoundStyle::round_to_nearest>;
 
  public:
@@ -288,7 +288,7 @@ struct ScaledEpilogueBiasAzpToken
                                               EVTComputeAcc>;
 
   using ComputeScaleBiasA = cutlass::epilogue::threadblock::VisitorCompute<
-      cutlass::multiply_add, ElementD, float,
+      cutlass::homogeneous_multiply_add, ElementD, float,
       cutlass::FloatRoundStyle::round_to_nearest>;
 
  public:

csrc/cutlass_extensions/epilogue/scaled_mm_epilogues_c3x.hpp

@@ -195,7 +195,7 @@ struct ScaledEpilogueBias
       cutlass::epilogue::fusion::Sm90EVT<Compute0, ScaleB, Accum>;
 
   using Compute1 = cutlass::epilogue::fusion::Sm90Compute<
-      cutlass::multiply_add, ElementD, float,
+      cutlass::homogeneous_multiply_add, ElementD, float,
       cutlass::FloatRoundStyle::round_to_nearest>;
 
  public:
@@ -238,7 +238,7 @@ struct ScaledEpilogueColumnBias
       cutlass::epilogue::fusion::Sm90EVT<Compute0, ScaleB, Accum>;
 
   using Compute1 = cutlass::epilogue::fusion::Sm90Compute<
-      cutlass::multiply_add, ElementD, float,
+      cutlass::homogeneous_multiply_add, ElementD, float,
       cutlass::FloatRoundStyle::round_to_nearest>;
 
  public:
@@ -295,7 +295,7 @@ struct ScaledEpilogueBiasAzp
       cutlass::epilogue::fusion::Sm90EVT<ComputeScaleB, ScaleB, EVTComputeAzp>;
 
   using ComputeScaleBiasA = cutlass::epilogue::fusion::Sm90Compute<
-      cutlass::multiply_add, ElementD, float,
+      cutlass::homogeneous_multiply_add, ElementD, float,
       cutlass::FloatRoundStyle::round_to_nearest>;
 
  public:
@@ -371,7 +371,7 @@ struct ScaledEpilogueBiasAzpToken
       cutlass::epilogue::fusion::Sm90EVT<ComputeScaleB, ScaleB, EVTComputeAcc>;
 
   using ComputeScaleBiasA = cutlass::epilogue::fusion::Sm90Compute<
-      cutlass::multiply_add, ElementD, float,
+      cutlass::homogeneous_multiply_add, ElementD, float,
       cutlass::FloatRoundStyle::round_to_nearest>;
 
  public:

csrc/cutlass_extensions/gemm/collective/sm90_mma_tma_gmma_ss_warpspecialized_fp8_blockwise_scaling.hpp

@@ -45,7 +45,6 @@
 #include "cute/algorithm/functional.hpp"
 #include "cute/atom/mma_atom.hpp"
 #include "cute/algorithm/gemm.hpp"
-#include "cute/tensor_predicate.hpp"
 #include "cute/numeric/arithmetic_tuple.hpp"
 
 #include "cutlass_extensions/gemm/dispatch_policy.hpp"

csrc/layernorm_kernels.cu

@@ -15,15 +15,16 @@ namespace vllm {
 // TODO(woosuk): Further optimize this kernel.
 template <typename scalar_t>
 __global__ void rms_norm_kernel(
-    scalar_t* __restrict__ out,           // [..., hidden_size]
-    const scalar_t* __restrict__ input,   // [..., hidden_size]
+    scalar_t* __restrict__ out,          // [..., hidden_size]
+    const scalar_t* __restrict__ input,  // [..., hidden_size]
+    const int64_t input_stride,
     const scalar_t* __restrict__ weight,  // [hidden_size]
     const float epsilon, const int num_tokens, const int hidden_size) {
   __shared__ float s_variance;
   float variance = 0.0f;
 
   for (int idx = threadIdx.x; idx < hidden_size; idx += blockDim.x) {
-    const float x = (float)input[blockIdx.x * hidden_size + idx];
+    const float x = (float)input[blockIdx.x * input_stride + idx];
     variance += x * x;
   }
 
@@ -37,7 +38,7 @@ __global__ void rms_norm_kernel(
   __syncthreads();
 
   for (int idx = threadIdx.x; idx < hidden_size; idx += blockDim.x) {
-    float x = (float)input[blockIdx.x * hidden_size + idx];
+    float x = (float)input[blockIdx.x * input_stride + idx];
     out[blockIdx.x * hidden_size + idx] =
         ((scalar_t)(x * s_variance)) * weight[idx];
   }
@@ -50,7 +51,8 @@ __global__ void rms_norm_kernel(
 template <typename scalar_t, int width>
 __global__ std::enable_if_t<(width > 0) && _typeConvert<scalar_t>::exists>
 fused_add_rms_norm_kernel(
-    scalar_t* __restrict__ input,         // [..., hidden_size]
+    scalar_t* __restrict__ input,  // [..., hidden_size]
+    const int64_t input_stride,
     scalar_t* __restrict__ residual,      // [..., hidden_size]
     const scalar_t* __restrict__ weight,  // [hidden_size]
     const float epsilon, const int num_tokens, const int hidden_size) {
@@ -59,6 +61,7 @@ fused_add_rms_norm_kernel(
   static_assert(sizeof(_f16Vec<scalar_t, width>) == sizeof(scalar_t) * width);
 
   const int vec_hidden_size = hidden_size / width;
+  const int64_t vec_input_stride = input_stride / width;
   __shared__ float s_variance;
   float variance = 0.0f;
   /* These and the argument pointers are all declared `restrict` as they are
@@ -73,7 +76,8 @@ fused_add_rms_norm_kernel(
 
   for (int idx = threadIdx.x; idx < vec_hidden_size; idx += blockDim.x) {
     int id = blockIdx.x * vec_hidden_size + idx;
-    _f16Vec<scalar_t, width> temp = input_v[id];
+    int64_t strided_id = blockIdx.x * vec_input_stride + idx;
+    _f16Vec<scalar_t, width> temp = input_v[strided_id];
     temp += residual_v[id];
     variance += temp.sum_squares();
     residual_v[id] = temp;
@@ -90,10 +94,11 @@ fused_add_rms_norm_kernel(
 
   for (int idx = threadIdx.x; idx < vec_hidden_size; idx += blockDim.x) {
     int id = blockIdx.x * vec_hidden_size + idx;
+    int64_t strided_id = blockIdx.x * vec_input_stride + idx;
     _f16Vec<scalar_t, width> temp = residual_v[id];
     temp *= s_variance;
     temp *= weight_v[idx];
-    input_v[id] = temp;
+    input_v[strided_id] = temp;
   }
 }
 
@@ -103,7 +108,8 @@ fused_add_rms_norm_kernel(
 template <typename scalar_t, int width>
 __global__ std::enable_if_t<(width == 0) || !_typeConvert<scalar_t>::exists>
 fused_add_rms_norm_kernel(
-    scalar_t* __restrict__ input,         // [..., hidden_size]
+    scalar_t* __restrict__ input,  // [..., hidden_size]
+    const int64_t input_stride,
     scalar_t* __restrict__ residual,      // [..., hidden_size]
     const scalar_t* __restrict__ weight,  // [hidden_size]
     const float epsilon, const int num_tokens, const int hidden_size) {
@@ -111,7 +117,7 @@ fused_add_rms_norm_kernel(
   float variance = 0.0f;
 
   for (int idx = threadIdx.x; idx < hidden_size; idx += blockDim.x) {
-    scalar_t z = input[blockIdx.x * hidden_size + idx];
+    scalar_t z = input[blockIdx.x * input_stride + idx];
     z += residual[blockIdx.x * hidden_size + idx];
     float x = (float)z;
     variance += x * x;
@@ -129,7 +135,7 @@ fused_add_rms_norm_kernel(
 
   for (int idx = threadIdx.x; idx < hidden_size; idx += blockDim.x) {
     float x = (float)residual[blockIdx.x * hidden_size + idx];
-    input[blockIdx.x * hidden_size + idx] =
+    input[blockIdx.x * input_stride + idx] =
         ((scalar_t)(x * s_variance)) * weight[idx];
   }
 }
@@ -141,11 +147,12 @@ void rms_norm(torch::Tensor& out,     // [..., hidden_size]
               torch::Tensor& weight,  // [hidden_size]
               double epsilon) {
   TORCH_CHECK(out.is_contiguous());
-  TORCH_CHECK(input.is_contiguous());
+  TORCH_CHECK(input.stride(-1) == 1);
   TORCH_CHECK(weight.is_contiguous());
 
   int hidden_size = input.size(-1);
   int num_tokens = input.numel() / hidden_size;
+  int64_t input_stride = input.stride(-2);
 
   dim3 grid(num_tokens);
   dim3 block(std::min(hidden_size, 1024));
@@ -153,26 +160,29 @@ void rms_norm(torch::Tensor& out,     // [..., hidden_size]
   const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
   VLLM_DISPATCH_FLOATING_TYPES(input.scalar_type(), "rms_norm_kernel", [&] {
     vllm::rms_norm_kernel<scalar_t><<<grid, block, 0, stream>>>(
-        out.data_ptr<scalar_t>(), input.data_ptr<scalar_t>(),
+        out.data_ptr<scalar_t>(), input.data_ptr<scalar_t>(), input_stride,
         weight.data_ptr<scalar_t>(), epsilon, num_tokens, hidden_size);
   });
 }
 
-#define LAUNCH_FUSED_ADD_RMS_NORM(width)                                       \
-  VLLM_DISPATCH_FLOATING_TYPES(                                                \
-      input.scalar_type(), "fused_add_rms_norm_kernel", [&] {                  \
-        vllm::fused_add_rms_norm_kernel<scalar_t, width>                       \
-            <<<grid, block, 0, stream>>>(input.data_ptr<scalar_t>(),           \
-                                         residual.data_ptr<scalar_t>(),        \
-                                         weight.data_ptr<scalar_t>(), epsilon, \
-                                         num_tokens, hidden_size);             \
+#define LAUNCH_FUSED_ADD_RMS_NORM(width)                                    \
+  VLLM_DISPATCH_FLOATING_TYPES(                                             \
+      input.scalar_type(), "fused_add_rms_norm_kernel", [&] {               \
+        vllm::fused_add_rms_norm_kernel<scalar_t, width>                    \
+            <<<grid, block, 0, stream>>>(                                   \
+                input.data_ptr<scalar_t>(), input_stride,                   \
+                residual.data_ptr<scalar_t>(), weight.data_ptr<scalar_t>(), \
+                epsilon, num_tokens, hidden_size);                          \
       });
 
 void fused_add_rms_norm(torch::Tensor& input,     // [..., hidden_size]
                         torch::Tensor& residual,  // [..., hidden_size]
                         torch::Tensor& weight,    // [hidden_size]
                         double epsilon) {
+  TORCH_CHECK(residual.is_contiguous());
+  TORCH_CHECK(weight.is_contiguous());
   int hidden_size = input.size(-1);
+  int64_t input_stride = input.stride(-2);
   int num_tokens = input.numel() / hidden_size;
 
   dim3 grid(num_tokens);
@@ -194,9 +204,16 @@ void fused_add_rms_norm(torch::Tensor& input,     // [..., hidden_size]
   auto inp_ptr = reinterpret_cast<std::uintptr_t>(input.data_ptr());
   auto res_ptr = reinterpret_cast<std::uintptr_t>(residual.data_ptr());
   auto wt_ptr = reinterpret_cast<std::uintptr_t>(weight.data_ptr());
-  bool ptrs_are_aligned =
-      inp_ptr % 16 == 0 && res_ptr % 16 == 0 && wt_ptr % 16 == 0;
-  if (ptrs_are_aligned && hidden_size % 8 == 0) {
+  constexpr int vector_width = 8;
+  constexpr int req_alignment_bytes =
+      vector_width * 2;  // vector_width * sizeof(bfloat16 or float16) (float32
+                         // falls back to non-vectorized version anyway)
+  bool ptrs_are_aligned = inp_ptr % req_alignment_bytes == 0 &&
+                          res_ptr % req_alignment_bytes == 0 &&
+                          wt_ptr % req_alignment_bytes == 0;
+  bool offsets_are_multiple_of_vector_width =
+      hidden_size % vector_width == 0 && input_stride % vector_width == 0;
+  if (ptrs_are_aligned && offsets_are_multiple_of_vector_width) {
     LAUNCH_FUSED_ADD_RMS_NORM(8);
   } else {
     LAUNCH_FUSED_ADD_RMS_NORM(0);

csrc/layernorm_quant_kernels.cu

@@ -23,8 +23,9 @@ namespace vllm {
 // TODO(woosuk): Further optimize this kernel.
 template <typename scalar_t, typename fp8_type>
 __global__ void rms_norm_static_fp8_quant_kernel(
-    fp8_type* __restrict__ out,           // [..., hidden_size]
-    const scalar_t* __restrict__ input,   // [..., hidden_size]
+    fp8_type* __restrict__ out,          // [..., hidden_size]
+    const scalar_t* __restrict__ input,  // [..., hidden_size]
+    const int input_stride,
     const scalar_t* __restrict__ weight,  // [hidden_size]
     const float* __restrict__ scale,      // [1]
     const float epsilon, const int num_tokens, const int hidden_size) {
@@ -32,7 +33,7 @@ __global__ void rms_norm_static_fp8_quant_kernel(
   float variance = 0.0f;
 
   for (int idx = threadIdx.x; idx < hidden_size; idx += blockDim.x) {
-    const float x = (float)input[blockIdx.x * hidden_size + idx];
+    const float x = (float)input[blockIdx.x * input_stride + idx];
     variance += x * x;
   }
 
@@ -49,7 +50,7 @@ __global__ void rms_norm_static_fp8_quant_kernel(
   float const scale_inv = 1.0f / *scale;
 
   for (int idx = threadIdx.x; idx < hidden_size; idx += blockDim.x) {
-    float x = (float)input[blockIdx.x * hidden_size + idx];
+    float x = (float)input[blockIdx.x * input_stride + idx];
     float const out_norm = ((scalar_t)(x * s_variance)) * weight[idx];
     out[blockIdx.x * hidden_size + idx] =
         scaled_fp8_conversion<true, fp8_type>(out_norm, scale_inv);
@@ -63,8 +64,9 @@ __global__ void rms_norm_static_fp8_quant_kernel(
 template <typename scalar_t, int width, typename fp8_type>
 __global__ std::enable_if_t<(width > 0) && _typeConvert<scalar_t>::exists>
 fused_add_rms_norm_static_fp8_quant_kernel(
-    fp8_type* __restrict__ out,           // [..., hidden_size]
-    scalar_t* __restrict__ input,         // [..., hidden_size]
+    fp8_type* __restrict__ out,    // [..., hidden_size]
+    scalar_t* __restrict__ input,  // [..., hidden_size]
+    const int input_stride,
     scalar_t* __restrict__ residual,      // [..., hidden_size]
     const scalar_t* __restrict__ weight,  // [hidden_size]
     const float* __restrict__ scale,      // [1]
@@ -74,6 +76,7 @@ fused_add_rms_norm_static_fp8_quant_kernel(
   static_assert(sizeof(_f16Vec<scalar_t, width>) == sizeof(scalar_t) * width);
 
   const int vec_hidden_size = hidden_size / width;
+  const int vec_input_stride = input_stride / width;
   __shared__ float s_variance;
   float variance = 0.0f;
   /* These and the argument pointers are all declared `restrict` as they are
@@ -87,8 +90,9 @@ fused_add_rms_norm_static_fp8_quant_kernel(
       reinterpret_cast<const _f16Vec<scalar_t, width>*>(weight);
 
   for (int idx = threadIdx.x; idx < vec_hidden_size; idx += blockDim.x) {
+    int stride_id = blockIdx.x * vec_input_stride + idx;
     int id = blockIdx.x * vec_hidden_size + idx;
-    _f16Vec<scalar_t, width> temp = input_v[id];
+    _f16Vec<scalar_t, width> temp = input_v[stride_id];
     temp += residual_v[id];
     variance += temp.sum_squares();
     residual_v[id] = temp;
@@ -125,8 +129,9 @@ fused_add_rms_norm_static_fp8_quant_kernel(
 template <typename scalar_t, int width, typename fp8_type>
 __global__ std::enable_if_t<(width == 0) || !_typeConvert<scalar_t>::exists>
 fused_add_rms_norm_static_fp8_quant_kernel(
-    fp8_type* __restrict__ out,           // [..., hidden_size]
-    scalar_t* __restrict__ input,         // [..., hidden_size]
+    fp8_type* __restrict__ out,    // [..., hidden_size]
+    scalar_t* __restrict__ input,  // [..., hidden_size]
+    const int input_stride,
     scalar_t* __restrict__ residual,      // [..., hidden_size]
     const scalar_t* __restrict__ weight,  // [hidden_size]
     const float* __restrict__ scale,      // [1]
@@ -135,7 +140,7 @@ fused_add_rms_norm_static_fp8_quant_kernel(
   float variance = 0.0f;
 
   for (int idx = threadIdx.x; idx < hidden_size; idx += blockDim.x) {
-    scalar_t z = input[blockIdx.x * hidden_size + idx];
+    scalar_t z = input[blockIdx.x * input_stride + idx];
     z += residual[blockIdx.x * hidden_size + idx];
     float x = (float)z;
     variance += x * x;
@@ -169,7 +174,9 @@ void rms_norm_static_fp8_quant(torch::Tensor& out,     // [..., hidden_size]
                                torch::Tensor& weight,  // [hidden_size]
                                torch::Tensor& scale,   // [1]
                                double epsilon) {
+  TORCH_CHECK(out.is_contiguous());
   int hidden_size = input.size(-1);
+  int input_stride = input.stride(-2);
   int num_tokens = input.numel() / hidden_size;
 
   dim3 grid(num_tokens);
@@ -183,8 +190,9 @@ void rms_norm_static_fp8_quant(torch::Tensor& out,     // [..., hidden_size]
               vllm::rms_norm_static_fp8_quant_kernel<scalar_t, fp8_t>
                   <<<grid, block, 0, stream>>>(
                       out.data_ptr<fp8_t>(), input.data_ptr<scalar_t>(),
-                      weight.data_ptr<scalar_t>(), scale.data_ptr<float>(),
-                      epsilon, num_tokens, hidden_size);
+                      input_stride, weight.data_ptr<scalar_t>(),
+                      scale.data_ptr<float>(), epsilon, num_tokens,
+                      hidden_size);
             });
       });
 }
@@ -198,7 +206,7 @@ void rms_norm_static_fp8_quant(torch::Tensor& out,     // [..., hidden_size]
                                                                width, fp8_t> \
                   <<<grid, block, 0, stream>>>(                              \
                       out.data_ptr<fp8_t>(), input.data_ptr<scalar_t>(),     \
-                      residual.data_ptr<scalar_t>(),                         \
+                      input_stride, residual.data_ptr<scalar_t>(),           \
                       weight.data_ptr<scalar_t>(), scale.data_ptr<float>(),  \
                       epsilon, num_tokens, hidden_size);                     \
             });                                                              \
@@ -210,7 +218,10 @@ void fused_add_rms_norm_static_fp8_quant(
     torch::Tensor& weight,    // [hidden_size]
     torch::Tensor& scale,     // [1]
     double epsilon) {
+  TORCH_CHECK(out.is_contiguous());
+  TORCH_CHECK(residual.is_contiguous());
   int hidden_size = input.size(-1);
+  int input_stride = input.stride(-2);
   int num_tokens = input.numel() / hidden_size;
 
   dim3 grid(num_tokens);
@@ -234,7 +245,7 @@ void fused_add_rms_norm_static_fp8_quant(
   auto wt_ptr = reinterpret_cast<std::uintptr_t>(weight.data_ptr());
   bool ptrs_are_aligned =
       inp_ptr % 16 == 0 && res_ptr % 16 == 0 && wt_ptr % 16 == 0;
-  if (ptrs_are_aligned && hidden_size % 8 == 0) {
+  if (ptrs_are_aligned && hidden_size % 8 == 0 && input_stride % 8 == 0) {
     LAUNCH_FUSED_ADD_RMS_NORM(8);
   } else {
     LAUNCH_FUSED_ADD_RMS_NORM(0);

csrc/mamba/causal_conv1d/causal_conv1d.cu

@@ -1,656 +0,0 @@
-// clang-format off
-// adapted from https://github.com/Dao-AILab/causal-conv1d/blob/main/csrc/causal_conv1d_fwd.cu 
-// and https://github.com/Dao-AILab/causal-conv1d/blob/main/csrc/causal_conv1d_update.cu
-#include <torch/all.h>
-#include <ATen/cuda/CUDAContext.h>
-#include <c10/cuda/CUDAGuard.h>
-
-#include "causal_conv1d.h"
-#include <c10/util/BFloat16.h>
-#include <c10/util/Half.h>
-#include <c10/cuda/CUDAException.h>  // For C10_CUDA_CHECK and C10_CUDA_KERNEL_LAUNCH_CHECK
-
-#include <cub/block/block_load.cuh>
-#include <cub/block/block_store.cuh>
-
-#ifdef USE_ROCM
-    namespace cub = hipcub;
-#endif
-
-#include "static_switch.h"
-
-
-
-#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, ...)              \
-    if (ITYPE == at::ScalarType::Half) {                                            \
-        using input_t = at::Half;                                                   \
-        using weight_t = at::Half;                                                  \
-        __VA_ARGS__();                                                              \
-    } else if (ITYPE == at::ScalarType::BFloat16) {                                 \
-        using input_t = at::BFloat16;                                               \
-        using weight_t = at::BFloat16;                                              \
-        __VA_ARGS__();                                                              \
-    } else if (ITYPE == at::ScalarType::Float)  {                                   \
-        using input_t = float;                                                      \
-        using weight_t = float;                                                     \
-        __VA_ARGS__();                                                              \
-    } else {                                                                        \
-        AT_ERROR(#NAME, " not implemented for input type '", toString(ITYPE), "'"); \
-    }
-
-
-template<typename input_t, typename weight_t>
-void causal_conv1d_fwd_cuda(ConvParamsBase &params, cudaStream_t stream);
-
-template<typename input_t, typename weight_t>
-void causal_conv1d_update_cuda(ConvParamsBase &params, cudaStream_t stream);
-
-void set_conv_params_fwd(ConvParamsBase &params,
-                         // sizes
-                         const size_t batch,
-                         const size_t dim,
-                         const size_t seqlen,
-                         const size_t width,
-                         // device pointers
-                         const at::Tensor x,
-                         const at::Tensor weight,
-                         const at::Tensor out,
-                         const std::optional<at::Tensor>& bias,
-                         bool silu_activation,
-                         int64_t pad_slot_id,
-                         const std::optional<at::Tensor>& query_start_loc = std::nullopt,
-                         const std::optional<at::Tensor>& cache_indices = std::nullopt,
-                         const std::optional<at::Tensor>& has_initial_state = std::nullopt) {
-
-    // Reset the parameters
-    memset(&params, 0, sizeof(params));
-
-    params.batch = batch;
-    params.dim = dim;
-    params.seqlen = seqlen;
-    params.width = width;
-    params.pad_slot_id = pad_slot_id;
-
-    params.silu_activation = silu_activation;
-
-    // Set the pointers and strides.
-    params.x_ptr = x.data_ptr();
-    params.weight_ptr = weight.data_ptr();
-    params.bias_ptr = bias.has_value() ? bias.value().data_ptr() : nullptr;
-    params.out_ptr = out.data_ptr();
-    // All stride are in elements, not bytes.
-    params.query_start_loc_ptr = query_start_loc.has_value() ? query_start_loc.value().data_ptr() : nullptr;
-    params.cache_indices_ptr = cache_indices.has_value() ? cache_indices.value().data_ptr() : nullptr;
-    params.has_initial_state_ptr = has_initial_state.has_value() ? has_initial_state.value().data_ptr() : nullptr;
-    const bool varlen = params.query_start_loc_ptr != nullptr;
-    params.x_batch_stride = x.stride(varlen ? 1 : 0);
-    params.x_c_stride = x.stride(varlen ? 0 : 1);
-    params.x_l_stride = x.stride(varlen ? 1 : -1);
-    params.weight_c_stride = weight.stride(0);
-    params.weight_width_stride = weight.stride(1);
-    params.out_batch_stride = out.stride(varlen ? 1 : 0);
-    params.out_c_stride = out.stride(varlen ? 0 : 1);
-    params.out_l_stride = out.stride(varlen ? 1 : -1);
-}
-
-
-void causal_conv1d_fwd(const at::Tensor &x, const at::Tensor &weight,
-                  const std::optional<at::Tensor> &bias_,
-                  const std::optional<at::Tensor> &conv_states,
-                  const std::optional<at::Tensor> &query_start_loc,
-                  const std::optional<at::Tensor> &cache_indices,
-                  const std::optional<at::Tensor> &has_initial_state,
-                  bool silu_activation,
-                 // used to identify padding entries if cache_indices provided
-                 // in case of padding, the kernel will return early
-                  int64_t pad_slot_id) {
-    auto input_type = x.scalar_type();
-    auto weight_type = weight.scalar_type();
-    TORCH_CHECK(input_type == at::ScalarType::Float || input_type == at::ScalarType::Half || input_type == at::ScalarType::BFloat16);
-    TORCH_CHECK(weight_type == at::ScalarType::Float || weight_type == at::ScalarType::Half || weight_type == at::ScalarType::BFloat16);
-
-    TORCH_CHECK(x.is_cuda());
-    TORCH_CHECK(weight.is_cuda());
-    
-    const bool varlen = query_start_loc.has_value() ? true : false;
-    const auto sizes = x.sizes();
-    const int batch_size = varlen ? query_start_loc.value().sizes()[0] - 1 : sizes[0];
-    const int dim = varlen ? sizes[0] : sizes[1];
-    const int seqlen = varlen ? sizes[1] : sizes[2];
-    const int width = weight.size(-1);
-    if (varlen){
-        CHECK_SHAPE(x, dim, seqlen);
-    }
-    else {
-        CHECK_SHAPE(x, batch_size, dim, seqlen);
-    }
-    CHECK_SHAPE(weight, dim, width);
-
-
-
-    if (bias_.has_value()) {
-        auto bias = bias_.value();
-        TORCH_CHECK(bias.scalar_type() == weight_type);
-        TORCH_CHECK(bias.is_cuda());
-        TORCH_CHECK(bias.stride(-1) == 1);
-        CHECK_SHAPE(bias, dim);
-    }
-
-
-    if (has_initial_state.has_value()) {
-        auto has_initial_state_ = has_initial_state.value();
-        TORCH_CHECK(has_initial_state_.scalar_type() == at::ScalarType::Bool);
-        TORCH_CHECK(has_initial_state_.is_cuda());
-        CHECK_SHAPE(has_initial_state_, batch_size);
-    }
-
-
-    if (query_start_loc.has_value()) {
-        auto query_start_loc_ = query_start_loc.value();
-        TORCH_CHECK(query_start_loc_.scalar_type() == at::ScalarType::Int);
-        TORCH_CHECK(query_start_loc_.is_cuda());
-    }
-
-
-    if (cache_indices.has_value()) {
-        auto cache_indices_ = cache_indices.value();
-        TORCH_CHECK(cache_indices_.scalar_type() == at::ScalarType::Int);
-        TORCH_CHECK(cache_indices_.is_cuda());
-        CHECK_SHAPE(cache_indices_, batch_size);
-    }
-
-    at::Tensor out = x;
-
-    ConvParamsBase params;
-    set_conv_params_fwd(params, batch_size, dim, seqlen, width, x, weight, out,
-                        bias_,
-                        silu_activation, 
-                        pad_slot_id,
-                        query_start_loc,
-                        cache_indices,
-                        has_initial_state
-                        );
-
-    if (conv_states.has_value()) {
-        auto conv_states_ = conv_states.value();
-        TORCH_CHECK(conv_states_.scalar_type() == input_type);
-        TORCH_CHECK(conv_states_.is_cuda());
-        params.conv_states_ptr = conv_states_.data_ptr();
-        params.conv_states_batch_stride = conv_states_.stride(0);
-        params.conv_states_c_stride = conv_states_.stride(1);
-        params.conv_states_l_stride = conv_states_.stride(2);
-    } else {
-        params.conv_states_ptr = nullptr;
-    }
-
-    const at::cuda::OptionalCUDAGuard device_guard(device_of(x));
-    auto stream = at::cuda::getCurrentCUDAStream().stream();
-    DISPATCH_WTYPE_ITYPE_FLOAT_AND_HALF_AND_BF16(x.scalar_type(), "causal_conv1d_fwd", [&] {
-            causal_conv1d_fwd_cuda<input_t, weight_t>(params, stream);
-    });
-}
-
-
-void causal_conv1d_update(const at::Tensor &x,
-                     const at::Tensor &conv_state,
-                     const at::Tensor &weight,
-                     const std::optional<at::Tensor> &bias_,
-                     bool silu_activation,
-                     const std::optional<at::Tensor> &cache_seqlens_,
-                     const std::optional<at::Tensor> &conv_state_indices_,
-                     // used to identify padding entries if cache_indices provided
-                     // in case of padding, the kernel will return early
-                     int64_t pad_slot_id) {
-    auto input_type = x.scalar_type();
-    auto weight_type = weight.scalar_type();
-    TORCH_CHECK(input_type == at::ScalarType::Float || input_type == at::ScalarType::Half || input_type == at::ScalarType::BFloat16);
-    TORCH_CHECK(weight_type == at::ScalarType::Float || weight_type == at::ScalarType::Half || weight_type == at::ScalarType::BFloat16);
-    TORCH_CHECK(weight_type == input_type, "weight type must equal to input type, other variations are disabled due to binary size limitations");
-    TORCH_CHECK(conv_state.scalar_type() == input_type);
-
-    TORCH_CHECK(x.is_cuda());
-    TORCH_CHECK(conv_state.is_cuda());
-    TORCH_CHECK(weight.is_cuda());
-
-    const auto sizes = x.sizes();
-    const int batch_size = sizes[0];
-    const int dim = sizes[1];
-    const int seqlen = sizes[2];
-    const int width = weight.size(-1);
-    const int conv_state_len = conv_state.size(2);
-    TORCH_CHECK(conv_state_len >= width - 1);
-
-    CHECK_SHAPE(x, batch_size, dim, seqlen);
-    CHECK_SHAPE(weight, dim, width);
-
-    TORCH_CHECK(width >= 2 && width <= 4, "causal_conv1d only supports width between 2 and 4");
-
-    if (bias_.has_value()) {
-        auto bias = bias_.value();
-        TORCH_CHECK(bias.scalar_type() == weight_type);
-        TORCH_CHECK(bias.is_cuda());
-        TORCH_CHECK(bias.stride(-1) == 1);
-        CHECK_SHAPE(bias, dim);
-    }
-
-    at::Tensor out = x;
-
-    ConvParamsBase params;
-    set_conv_params_fwd(params, batch_size, dim, seqlen, width, x, weight, out,
-                        bias_,
-                        silu_activation,
-                        pad_slot_id);
-    params.conv_state_ptr = conv_state.data_ptr();
-    params.conv_state_len = conv_state_len;
-    // All stride are in elements, not bytes.
-    params.conv_state_batch_stride = conv_state.stride(0);
-    params.conv_state_c_stride = conv_state.stride(1);
-    params.conv_state_l_stride = conv_state.stride(2);
-
-    if (cache_seqlens_.has_value()) {
-        auto cache_seqlens = cache_seqlens_.value();
-        TORCH_CHECK(cache_seqlens.scalar_type() == torch::kInt32);
-        TORCH_CHECK(cache_seqlens.is_cuda());
-        TORCH_CHECK(cache_seqlens.stride(-1) == 1);
-        CHECK_SHAPE(cache_seqlens, batch_size);
-        params.cache_seqlens = cache_seqlens.data_ptr<int32_t>();
-    } else {
-        params.cache_seqlens = nullptr;
-    }
-
-    if (conv_state_indices_.has_value()) {
-        auto conv_state_indices = conv_state_indices_.value();
-        TORCH_CHECK(conv_state_indices.scalar_type() == torch::kInt32)
-        TORCH_CHECK(conv_state_indices.is_cuda());
-        TORCH_CHECK(conv_state_indices.stride(0) == 1)
-        CHECK_SHAPE(conv_state_indices, batch_size);
-
-        int conv_state_entries = conv_state.size(0);
-        CHECK_SHAPE(conv_state, conv_state_entries, dim, conv_state_len);
-
-        params.conv_state_indices_ptr = conv_state_indices.data_ptr<int32_t>();
-    } else {
-        CHECK_SHAPE(conv_state, batch_size, dim, conv_state_len);
-        params.conv_state_indices_ptr = nullptr;
-    }
-
-    const at::cuda::OptionalCUDAGuard device_guard(device_of(x));
-    auto stream = at::cuda::getCurrentCUDAStream().stream();
-    DISPATCH_WTYPE_ITYPE_FLOAT_AND_HALF_AND_BF16(x.scalar_type(), "causal_conv1d_update", [&] {
-            causal_conv1d_update_cuda<input_t, weight_t>(params, stream);
-    });
-}
-
-template<int kNThreads_, int kWidth_, bool kIsVecLoad_, typename input_t_, typename weight_t_>
-struct Causal_conv1d_fwd_kernel_traits {
-    using input_t = input_t_;
-    using weight_t = weight_t_;
-    static constexpr int kNThreads = kNThreads_;
-    static constexpr int kWidth = kWidth_;
-    static constexpr int kNBytes = sizeof(input_t);
-    static_assert(kNBytes == 2 || kNBytes == 4);
-    static constexpr int kNElts = kNBytes == 4 ? 4 : 8;
-    static_assert(kWidth <= kNElts);
-    static constexpr bool kIsVecLoad = kIsVecLoad_;
-    using vec_t = typename BytesToType<kNBytes * kNElts>::Type;
-    using BlockLoadT = cub::BlockLoad<input_t, kNThreads, kNElts, cub::BLOCK_LOAD_WARP_TRANSPOSE>;
-    using BlockLoadVecT = cub::BlockLoad<vec_t, kNThreads, 1, cub::BLOCK_LOAD_DIRECT>;
-    using BlockStoreT = cub::BlockStore<input_t, kNThreads, kNElts, cub::BLOCK_STORE_WARP_TRANSPOSE>;
-    using BlockStoreVecT = cub::BlockStore<vec_t, kNThreads, 1, cub::BLOCK_STORE_DIRECT>;
-    static constexpr int kSmemIOSize = kIsVecLoad
-        ? 0
-        : custom_max({sizeof(typename BlockLoadT::TempStorage), sizeof(typename BlockStoreT::TempStorage)});
-    static constexpr int kSmemExchangeSize = kNThreads * kNBytes * kNElts;
-    static constexpr int kSmemSize = kSmemIOSize + kSmemExchangeSize;
-};
-
-template<typename Ktraits>
-__global__ __launch_bounds__(Ktraits::kNThreads)
-void causal_conv1d_fwd_kernel(ConvParamsBase params) {
-    constexpr int kWidth = Ktraits::kWidth;
-    constexpr int kNThreads = Ktraits::kNThreads;
-    constexpr int kNElts = Ktraits::kNElts;
-    constexpr bool kIsVecLoad = Ktraits::kIsVecLoad;
-    using input_t = typename Ktraits::input_t;
-    using vec_t = typename Ktraits::vec_t;
-    using weight_t = typename Ktraits::weight_t;
-
-    // Shared memory.
-    extern __shared__ char smem_[];
-    auto& smem_load = reinterpret_cast<typename Ktraits::BlockLoadT::TempStorage&>(smem_);
-    auto& smem_load_vec = reinterpret_cast<typename Ktraits::BlockLoadVecT::TempStorage&>(smem_);
-    auto& smem_store = reinterpret_cast<typename Ktraits::BlockStoreT::TempStorage&>(smem_);
-    auto& smem_store_vec = reinterpret_cast<typename Ktraits::BlockStoreVecT::TempStorage&>(smem_);
-    vec_t *smem_exchange = reinterpret_cast<vec_t *>(smem_ + Ktraits::kSmemIOSize);
-
-    const bool kVarlen = params.query_start_loc_ptr != nullptr;
-    const int tidx = threadIdx.x;
-    const int batch_id = blockIdx.x;
-    const int channel_id = blockIdx.y;
-    const int *query_start_loc = kVarlen ? reinterpret_cast<int *>(params.query_start_loc_ptr) : nullptr;
-    const int sequence_start_index = kVarlen ? query_start_loc[batch_id] : batch_id;
-    const int seqlen = kVarlen ? query_start_loc[batch_id + 1] - sequence_start_index : params.seqlen;
-
-    input_t *x = reinterpret_cast<input_t *>(params.x_ptr) + sequence_start_index * params.x_batch_stride
-        + channel_id * params.x_c_stride;
-    weight_t *weight = reinterpret_cast<weight_t *>(params.weight_ptr) + channel_id * params.weight_c_stride;
-    input_t *out = reinterpret_cast<input_t *>(params.out_ptr) + sequence_start_index * params.out_batch_stride
-        + channel_id * params.out_c_stride;
-    float bias_val = params.bias_ptr == nullptr ? 0.f : float(reinterpret_cast<weight_t *>(params.bias_ptr)[channel_id]);
-
-    bool has_initial_state = params.has_initial_state_ptr == nullptr ? false
-        : reinterpret_cast<bool *>(params.has_initial_state_ptr)[batch_id];
-
-    int* cache_indices = params.cache_indices_ptr == nullptr ? nullptr
-        : reinterpret_cast<int *>(params.cache_indices_ptr);
-    int cache_index = cache_indices == nullptr ? batch_id : cache_indices[batch_id];
-    // cache_index == params.pad_slot_id is defined as padding, so we exit early
-    if (cache_index == params.pad_slot_id){
-        return;
-    }
-    input_t *conv_states = params.conv_states_ptr == nullptr ? nullptr
-        : reinterpret_cast<input_t *>(params.conv_states_ptr) + cache_index * params.conv_states_batch_stride + channel_id * params.conv_states_c_stride;
-
-    // Thread 0 will load the last elements of the previous chunk, so we initialize those to 0.
-    if (tidx == 0) {
-        input_t initial_state[kNElts] = {0};
-        if (has_initial_state) {
-            #pragma unroll
-            for (int w = 0; w < kWidth - 1; ++w){ initial_state[kNElts - 1 - (kWidth - 2) + w ] = conv_states[w]; }
-        }
-        smem_exchange[kNThreads - 1] = reinterpret_cast<vec_t *>(initial_state)[0];
-    }
-
-    float weight_vals[kWidth];
-    #pragma unroll
-    for (int i = 0; i < kWidth; ++i) { weight_vals[i] = float(weight[i * params.weight_width_stride]); }
-
-    constexpr int kChunkSize = kNThreads * kNElts;
-    const int n_chunks = (seqlen + kChunkSize - 1) / kChunkSize;
-    for (int chunk = 0; chunk < n_chunks; ++chunk) {
-        input_t x_vals_load[2 * kNElts] = {0};
-        if constexpr(kIsVecLoad) {
-            typename Ktraits::BlockLoadVecT(smem_load_vec).Load(reinterpret_cast<vec_t*>(x), *reinterpret_cast<vec_t (*)[1]>(&x_vals_load[kNElts]), (seqlen - chunk * kChunkSize) / kNElts);
-        } else {
-            __syncthreads();
-            typename Ktraits::BlockLoadT(smem_load).Load(x, *reinterpret_cast<input_t (*)[kNElts]>(&x_vals_load[kNElts]), seqlen - chunk * kChunkSize);
-        }
-        x += kChunkSize;
-        __syncthreads();
-        // Thread kNThreads - 1 don't write yet, so that thread 0 can read
-        // the last elements of the previous chunk.
-        if (tidx < kNThreads - 1) { smem_exchange[tidx] = reinterpret_cast<vec_t *>(x_vals_load)[1]; }
-        __syncthreads();
-        reinterpret_cast<vec_t *>(x_vals_load)[0] = smem_exchange[tidx > 0 ? tidx - 1 : kNThreads - 1];
-        __syncthreads();
-        // Now thread kNThreads - 1 can write the last elements of the current chunk.
-        if (tidx == kNThreads - 1) { smem_exchange[tidx] = reinterpret_cast<vec_t *>(x_vals_load)[1]; }
-
-        float x_vals[2 * kNElts];
-        #pragma unroll
-        for (int i = 0; i < 2 * kNElts; ++i) { x_vals[i] = float(x_vals_load[i]); }
-
-        float out_vals[kNElts];
-        #pragma unroll
-        for (int i = 0; i < kNElts; ++i) {
-            out_vals[i] = bias_val;
-            #pragma unroll
-            for (int w = 0; w < kWidth; ++w) {
-                out_vals[i] += weight_vals[w] * x_vals[kNElts + i - (kWidth - w - 1)];
-            }
-        }
-
-        if (params.silu_activation) {
-            #pragma unroll
-            for (int i = 0; i < kNElts; ++i) {
-                out_vals[i] = out_vals[i] / (1 + expf(-out_vals[i]));
-            }
-        }
-
-        input_t out_vals_store[kNElts];
-        #pragma unroll
-        for (int i = 0; i < kNElts; ++i) { out_vals_store[i] = out_vals[i]; }
-        if constexpr(kIsVecLoad) {
-            typename Ktraits::BlockStoreVecT(smem_store_vec).Store(reinterpret_cast<vec_t*>(out), reinterpret_cast<vec_t (&)[1]>(out_vals_store), (seqlen - chunk * kChunkSize) / kNElts);
-        } else {
-            typename Ktraits::BlockStoreT(smem_store).Store(out, out_vals_store, seqlen - chunk * kChunkSize);
-        }
-        out += kChunkSize;
-
-        int final_state_position =  ((seqlen - (kWidth - 1)) - (n_chunks - 1) * kChunkSize);
-        // in case the final state is separated between the last "smem_exchange" and 
-        // and the one before it (chunk = n_chunks - 1 and chunk = n_chunks - 2), 
-        // (which occurs when `final_state_position` is a non-positive index)
-        // we load the correct data from smem_exchange from both chunks, the last chunk iteration and the one before it
-        if (conv_states != nullptr && final_state_position < 0 && seqlen > kWidth){
-            input_t vals_load[kNElts] = {0};
-            if ((chunk == n_chunks - 2) && (tidx == kNThreads - 1)){
-                // chunk = n_chunks - 2, a segment of the final state sits in the last index
-                reinterpret_cast<vec_t *>(vals_load)[0] = smem_exchange[kNThreads - 1];
-                #pragma unroll
-                for (int w = 0; w < -final_state_position; ++w){
-                    conv_states[w] = vals_load[kNElts + final_state_position + w];
-                }
-            }
-            if ((chunk == n_chunks - 1) && tidx == 0){
-                // chunk = n_chunks - 1, the second segment of the final state first positions
-                reinterpret_cast<vec_t *>(vals_load)[0] = smem_exchange[0];
-                for (int w = -final_state_position; w < kWidth - 1; ++w){
-                    conv_states[w] = vals_load[w + final_state_position];
-                }
-                return;
-            }
-        }
-    }
-    // Final state is stored in the smem_exchange last token slot,
-    // in case seqlen < kWidth, we would need to take the final state from the 
-    // initial state which is stored in conv_states
-    // in case seqlen > kWidth, we would need to load the last kWidth - 1 data
-    // and load it into conv_state accordingly
-    int last_thread =  ((seqlen - (kWidth - 1)) - (n_chunks - 1) * kChunkSize) / kNElts;
-    if (conv_states != nullptr && tidx == last_thread) { 
-        input_t x_vals_load[kNElts * 2] = {0};
-        // in case we are on the first kWidth tokens
-        if (last_thread == 0 && seqlen < kWidth){
-            // Need to take the initial state
-            reinterpret_cast<vec_t *>(x_vals_load)[0] = smem_exchange[0];
-            const int offset = seqlen - (kWidth - 1);
-            #pragma unroll
-            for (int w = 0; w < kWidth - 1; ++w){
-                // pad the existing state
-                if ((w - seqlen) >= 0 && has_initial_state) { conv_states[w - seqlen] = conv_states[w]; }
-                else if ((w - seqlen) >= 0 && !has_initial_state) { conv_states[w - seqlen] = input_t(0.0f); }
-            }
-            #pragma unroll
-            for (int w = 0; w < kWidth - 1; ++w){
-                if (offset + w >= 0) 
-                    conv_states[w] = x_vals_load[offset + w ];
-            }
-        }
-        else {
-            // in case the final state is in between the threads data
-            const int offset = ((seqlen - (kWidth - 1)) % (kNElts));
-            if ((offset + kWidth - 2) >= kNElts && (last_thread + 1 < kNThreads)){
-                // In case last_thread == kNThreads - 1, accessing last_thread + 1 will result in a 
-                // illegal access error on H100.
-                // Therefore, we access last_thread + 1, only if the final state data sits there
-                reinterpret_cast<vec_t *>(x_vals_load)[1] = smem_exchange[last_thread + 1];
-            }
-            reinterpret_cast<vec_t *>(x_vals_load)[0] = smem_exchange[last_thread];
-            #pragma unroll
-            for (int w = 0; w < kWidth - 1; ++w){
-                conv_states[w] = x_vals_load[offset + w ];
-            }
-        }
-        
-    }
-}
-
-
-template<int kNThreads, int kWidth, typename input_t, typename weight_t>
-void causal_conv1d_fwd_launch(ConvParamsBase &params, cudaStream_t stream) {
-    static constexpr int kNElts = sizeof(input_t) == 4 ? 4 : 8;
-    const bool kVarlen = params.query_start_loc_ptr != nullptr;
-    BOOL_SWITCH(params.seqlen % kNElts == 0 && !kVarlen, kIsVecLoad, [&] {
-        using Ktraits = Causal_conv1d_fwd_kernel_traits<kNThreads, kWidth, kIsVecLoad, input_t, weight_t>;
-        constexpr int kSmemSize = Ktraits::kSmemSize;
-        dim3 grid(params.batch, params.dim);
-
-        auto kernel = &causal_conv1d_fwd_kernel<Ktraits>;
-
-        if (kSmemSize >= 48 * 1024) {
-            C10_CUDA_CHECK(cudaFuncSetAttribute(
-                (void *) kernel, cudaFuncAttributeMaxDynamicSharedMemorySize, kSmemSize));
-            std::cerr << "Warning (causal_conv1d fwd launch): attempting to set maxDynamicSharedMemorySize on an AMD GPU which is currently a non-op (in ROCm versions <= 6.1). This might lead to undefined behavior. \n" << std::endl;
-        }
-        kernel<<<grid, Ktraits::kNThreads, kSmemSize, stream>>>(params);
-
-        C10_CUDA_KERNEL_LAUNCH_CHECK();
-    });
-}
-
-template<typename input_t, typename weight_t>
-void causal_conv1d_fwd_cuda(ConvParamsBase &params, cudaStream_t stream) {
-    if (params.width == 2) {
-        causal_conv1d_fwd_launch<128, 2, input_t, weight_t>(params, stream);
-    } else if (params.width == 3) {
-        causal_conv1d_fwd_launch<128, 3, input_t, weight_t>(params, stream);
-    } else if (params.width == 4) {
-        causal_conv1d_fwd_launch<128, 4, input_t, weight_t>(params, stream);
-    }
-}
-
-
-template void causal_conv1d_fwd_cuda<float, float>(ConvParamsBase &params, cudaStream_t stream);
-template void causal_conv1d_fwd_cuda<at::Half, at::Half>(ConvParamsBase &params, cudaStream_t stream);
-template void causal_conv1d_fwd_cuda<at::BFloat16, at::BFloat16>(ConvParamsBase &params, cudaStream_t stream);
-
-
-
-
-template<int kNThreads_, int kWidth_, typename input_t_, typename weight_t_>
-struct Causal_conv1d_update_kernel_traits {
-    using input_t = input_t_;
-    using weight_t = weight_t_;
-    static constexpr int kNThreads = kNThreads_;
-    static constexpr int kWidth = kWidth_;
-    static constexpr int kNBytes = sizeof(input_t);
-    static_assert(kNBytes == 2 || kNBytes == 4);
-};
-
-template<typename Ktraits, bool kIsCircularBuffer>
-__global__ __launch_bounds__(Ktraits::kNThreads)
-void causal_conv1d_update_kernel(ConvParamsBase params) {
-    constexpr int kWidth = Ktraits::kWidth;
-    constexpr int kNThreads = Ktraits::kNThreads;
-    using input_t = typename Ktraits::input_t;
-    using weight_t = typename Ktraits::weight_t;
-
-    const int tidx = threadIdx.x;
-    const int batch_id = blockIdx.x;
-    const int channel_id = blockIdx.y * kNThreads + tidx;
-    if (channel_id >= params.dim) return;
-
-    input_t *x = reinterpret_cast<input_t *>(params.x_ptr) + batch_id * params.x_batch_stride
-        + channel_id * params.x_c_stride;
-
-    // If params.conv_state_batch_indices is set, then the conv state is gathered from the conv state tensor
-    // along the batch axis. Otherwise, the conv state coordinate is the same as the batch id.
-    const int conv_state_batch_coord = params.conv_state_indices_ptr == nullptr
-        ? batch_id
-        : params.conv_state_indices_ptr[batch_id];
-    // conv_state_batch_coord == params.pad_slot_id is defined as padding so we exit early
-    if (conv_state_batch_coord == params.pad_slot_id){
-        return;
-    }
-    input_t *conv_state = reinterpret_cast<input_t *>(params.conv_state_ptr) 
-        + conv_state_batch_coord * params.conv_state_batch_stride
-        + channel_id * params.conv_state_c_stride;
-
-    weight_t *weight = reinterpret_cast<weight_t *>(params.weight_ptr) + channel_id * params.weight_c_stride;
-    input_t *out = reinterpret_cast<input_t *>(params.out_ptr) + batch_id * params.out_batch_stride
-        + channel_id * params.out_c_stride;
-    float bias_val = params.bias_ptr == nullptr ? 0.f : float(reinterpret_cast<weight_t *>(params.bias_ptr)[channel_id]);
-
-    int state_len = params.conv_state_len;
-    int advance_len = params.seqlen;
-    int cache_seqlen = kIsCircularBuffer ? params.cache_seqlens[batch_id] % state_len : 0;
-    int update_idx = cache_seqlen - (kWidth - 1);
-    update_idx = update_idx < 0 ? update_idx + state_len : update_idx;
-
-    float weight_vals[kWidth] = {0};
-    #pragma unroll
-    for (int i = 0; i < kWidth; ++i) { weight_vals[i] = float(weight[i * params.weight_width_stride]); }
-
-    float x_vals[kWidth] = {0};
-    if constexpr (!kIsCircularBuffer) {
-        #pragma unroll 2
-        for (int i = 0; i < state_len - advance_len - (kWidth - 1); ++i) {
-            conv_state[i * params.conv_state_l_stride] = conv_state[(i + advance_len) * params.conv_state_l_stride];
-        }
-        #pragma unroll
-        for (int i = 0; i < kWidth - 1; ++i) {
-            input_t state_val = conv_state[(state_len - (kWidth - 1) + i) * params.conv_state_l_stride];
-            if (i < advance_len + (kWidth - 1) && state_len - advance_len - (kWidth - 1) + i >= 0) {
-                conv_state[(state_len - advance_len - (kWidth - 1) + i) * params.conv_state_l_stride] = state_val;
-            }
-            x_vals[i] = float(state_val);
-        }
-    } else {
-        #pragma unroll
-        for (int i = 0; i < kWidth - 1; ++i, update_idx = update_idx + 1 >= state_len ? update_idx + 1 - state_len : update_idx + 1) {
-            input_t state_val = conv_state[update_idx * params.conv_state_l_stride];
-            x_vals[i] = float(state_val);
-        }
-    }
-    #pragma unroll 2
-    for (int i = 0; i < params.seqlen; ++i) {
-        input_t x_val = x[i * params.x_l_stride];
-        if constexpr (!kIsCircularBuffer) {
-            if (i < advance_len && state_len - advance_len + i >= 0) {
-                conv_state[(state_len - advance_len + i) * params.conv_state_l_stride] = x_val;
-            }
-        } else {
-            conv_state[update_idx * params.conv_state_l_stride] = x_val;
-            ++update_idx;
-            update_idx = update_idx >= state_len ? update_idx - state_len : update_idx;
-        }
-        x_vals[kWidth - 1] = float(x_val);
-        float out_val = bias_val;
-        #pragma unroll
-        for (int j = 0; j < kWidth; ++j) { out_val += weight_vals[j] * x_vals[j]; }
-        if (params.silu_activation) { out_val = out_val / (1 + expf(-out_val)); }
-        out[i * params.out_l_stride] = input_t(out_val);
-        // Shift the input buffer by 1
-        #pragma unroll
-        for (int i = 0; i < kWidth - 1; ++i) { x_vals[i] = x_vals[i + 1]; }
-    }
-}
-
-template<int kNThreads, int kWidth, typename input_t, typename weight_t>
-void causal_conv1d_update_launch(ConvParamsBase &params, cudaStream_t stream) {
-    using Ktraits = Causal_conv1d_update_kernel_traits<kNThreads, kWidth, input_t, weight_t>;
-    dim3 grid(params.batch, (params.dim + kNThreads - 1) / kNThreads);
-    auto kernel = params.cache_seqlens == nullptr
-        ? &causal_conv1d_update_kernel<Ktraits, false>
-        : &causal_conv1d_update_kernel<Ktraits, true>;
-    kernel<<<grid, Ktraits::kNThreads, 0, stream>>>(params);
-    C10_CUDA_KERNEL_LAUNCH_CHECK();
-}
-
-template<typename input_t, typename weight_t>
-void causal_conv1d_update_cuda(ConvParamsBase &params, cudaStream_t stream) {
-    if (params.width == 2) {
-        causal_conv1d_update_launch<64, 2, input_t, weight_t>(params, stream);
-    } else if (params.width == 3) {
-        causal_conv1d_update_launch<64, 3, input_t, weight_t>(params, stream);
-    } else if (params.width == 4) {
-        causal_conv1d_update_launch<64, 4, input_t, weight_t>(params, stream);
-    }
-}
-
-template void causal_conv1d_update_cuda<float, float>(ConvParamsBase &params, cudaStream_t stream);
-template void causal_conv1d_update_cuda<at::Half, at::Half>(ConvParamsBase &params, cudaStream_t stream);
-template void causal_conv1d_update_cuda<at::BFloat16, at::BFloat16>(ConvParamsBase &params, cudaStream_t stream);

csrc/mamba/causal_conv1d/causal_conv1d.h

@@ -1,159 +0,0 @@
-/******************************************************************************
- * Copyright (c) 2024, Tri Dao.
- ******************************************************************************/
-// clang-format off
-// adapted from https://github.com/Dao-AILab/causal-conv1d/blob/main/csrc/causal_conv1d.h
-#pragma once
-
-#include <cuda_bf16.h>
-#include <cuda_fp16.h>
-////////////////////////////////////////////////////////////////////////////////////////////////////
-
-struct ConvParamsBase {
-    using index_t = uint32_t;
-
-    int batch, dim, seqlen, width;
-    int64_t pad_slot_id;
-    bool silu_activation;
-
-    index_t x_batch_stride;
-    index_t x_c_stride;
-    index_t x_l_stride;
-    index_t weight_c_stride;
-    index_t weight_width_stride;
-    index_t out_batch_stride;
-    index_t out_c_stride;
-    index_t out_l_stride;
-
-    int conv_state_len;
-    index_t conv_state_batch_stride;
-    index_t conv_state_c_stride;
-    index_t conv_state_l_stride;
-
-    // Common data pointers.
-    void *__restrict__ x_ptr;
-    void *__restrict__ weight_ptr;
-    void *__restrict__ bias_ptr;
-    void *__restrict__ out_ptr;
-
-    void *__restrict__ conv_state_ptr;
-    void *__restrict__ query_start_loc_ptr;
-    void *__restrict__ has_initial_state_ptr;
-    void *__restrict__ cache_indices_ptr;
-    int32_t *__restrict__ cache_seqlens;
-
-    // For the continuous batching case. Makes it so that the mamba state for 
-    // the current batch doesn't need to be a contiguous tensor.
-    int32_t *__restrict__ conv_state_indices_ptr;
-
-    void *__restrict__ seq_idx_ptr;
-
-    // No __restrict__ since initial_states could be the same as final_states.
-    void * initial_states_ptr;
-    index_t initial_states_batch_stride;
-    index_t initial_states_l_stride;
-    index_t initial_states_c_stride;
-
-    void * final_states_ptr;
-    index_t final_states_batch_stride;
-    index_t final_states_l_stride;
-    index_t final_states_c_stride;
-
-    void *  conv_states_ptr;
-    index_t conv_states_batch_stride;
-    index_t conv_states_l_stride;
-    index_t conv_states_c_stride;
-};
-
-
-#ifndef USE_ROCM
-    #include <cuda_bf16.h>
-
-    template<typename T>
-    __device__ inline T shuffle_xor(T val, int offset) {
-        return __shfl_xor_sync(uint32_t(-1), val, offset);
-    }
-
-    constexpr size_t custom_max(std::initializer_list<size_t> ilist) 
-    {
-        return std::max(ilist);
-    }
-
-    template<typename T>
-    constexpr T constexpr_min(T a, T b) {
-        return std::min(a, b);
-    }
-
-#else
-    #include <hip/hip_bf16.h>
-
-    template<typename T>
-    __device__ inline T shuffle_xor(T val, int offset) {
-        return __shfl_xor(val, offset);
-    }
-    constexpr size_t custom_max(std::initializer_list<size_t> ilist) 
-    {
-        return *std::max_element(ilist.begin(), ilist.end());
-    }
-
-    template<typename T>
-    constexpr T constexpr_min(T a, T b) {
-        return a < b ? a : b;
-    }
-#endif
-
-////////////////////////////////////////////////////////////////////////////////////////////////////
-
-template<int BYTES> struct BytesToType {};
-
-template<> struct BytesToType<16> {
-    using Type = uint4;
-    static_assert(sizeof(Type) == 16);
-};
-
-template<> struct BytesToType<8> {
-    using Type = uint64_t;
-    static_assert(sizeof(Type) == 8);
-};
-
-template<> struct BytesToType<4> {
-    using Type = uint32_t;
-    static_assert(sizeof(Type) == 4);
-};
-
-template<> struct BytesToType<2> {
-    using Type = uint16_t;
-    static_assert(sizeof(Type) == 2);
-};
-
-template<> struct BytesToType<1> {
-    using Type = uint8_t;
-    static_assert(sizeof(Type) == 1);
-};
-
-////////////////////////////////////////////////////////////////////////////////////////////////////
-
-template<typename T>
-struct SumOp {
-__device__ inline T operator()(T const & x, T const & y) { return x + y; }
-};
-
-template<int THREADS>
-struct Allreduce {
-    static_assert(THREADS == 32 || THREADS == 16 || THREADS == 8 || THREADS == 4);
-    template<typename T, typename Operator>
-    static __device__ inline T run(T x, Operator &op) {
-        constexpr int OFFSET = THREADS / 2;
-        x = op(x, __shfl_xor_sync(uint32_t(-1), x, OFFSET));
-        return Allreduce<OFFSET>::run(x, op);
-    }
-};
-
-template<>
-struct Allreduce<2> {
-template<typename T, typename Operator>
-static __device__ inline T run(T x, Operator &op) {
-    x = op(x, __shfl_xor_sync(uint32_t(-1), x, 1));
-    return x;
-}
-};

csrc/mamba/causal_conv1d/static_switch.h

@@ -1,28 +0,0 @@
-// Inspired by
-// https://github.com/NVIDIA/DALI/blob/main/include/dali/core/static_switch.h
-// and https://github.com/pytorch/pytorch/blob/master/aten/src/ATen/Dispatch.h
-// clang-format off
-// adapted from https://github.com/Dao-AILab/causal-conv1d/blob/main/csrc/static_switch.h
-
-#pragma once
-
-/// @param COND       - a boolean expression to switch by
-/// @param CONST_NAME - a name given for the constexpr bool variable.
-/// @param ...       - code to execute for true and false
-///
-/// Usage:
-/// ```
-/// BOOL_SWITCH(flag, BoolConst, [&] {
-///     some_function<BoolConst>(...);
-/// });
-/// ```
-#define BOOL_SWITCH(COND, CONST_NAME, ...)                                           \
-    [&] {                                                                            \
-        if (COND) {                                                                  \
-            static constexpr bool CONST_NAME = true;                                 \
-            return __VA_ARGS__();                                                    \
-        } else {                                                                     \
-            static constexpr bool CONST_NAME = false;                                \
-            return __VA_ARGS__();                                                    \
-        }                                                                            \
-    }()

csrc/mamba/mamba_ssm/selective_scan_fwd.cu

@@ -7,7 +7,11 @@
 
 #include <c10/util/BFloat16.h>
 #include <c10/util/Half.h>
-#include <c10/cuda/CUDAException.h>  // For C10_CUDA_CHECK and C10_CUDA_KERNEL_LAUNCH_CHECK
+#ifdef USE_ROCM
+    #include <c10/hip/HIPException.h>  // For C10_HIP_CHECK and C10_HIP_KERNEL_LAUNCH_CHECK
+#else
+    #include <c10/cuda/CUDAException.h>  // For C10_CUDA_CHECK and C10_CUDA_KERNEL_LAUNCH_CHECK
+#endif
 
 #ifndef USE_ROCM
     #include <cub/block/block_load.cuh>
@@ -312,19 +316,25 @@ void selective_scan_fwd_launch(SSMParamsBase &params, cudaStream_t stream) {
     // kIsVariableB, kIsVariableC and kHasZ are all set to True to reduce binary size
     constexpr bool kIsVariableB = true;
     constexpr bool kIsVariableC = true;
-    constexpr bool kHasZ = true;
     BOOL_SWITCH(params.seqlen % (kNThreads * kNItems) == 0, kIsEvenLen, [&] {
-        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>;
-            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>;
-            if (kSmemSize >= 48 * 1024) {
-                C10_CUDA_CHECK(cudaFuncSetAttribute(
-                    (void *) kernel, cudaFuncAttributeMaxDynamicSharedMemorySize, kSmemSize));
-            }
-            kernel<<<grid, Ktraits::kNThreads, kSmemSize, stream>>>(params);
-            C10_CUDA_KERNEL_LAUNCH_CHECK();
+        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>;
+                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>;
+                if (kSmemSize >= 48 * 1024) {
+#ifdef USE_ROCM
+                    C10_HIP_CHECK(hipFuncSetAttribute(
+                        reinterpret_cast<const void*>(kernel), hipFuncAttributeMaxDynamicSharedMemorySize, kSmemSize));
+#else
+                    C10_CUDA_CHECK(cudaFuncSetAttribute(
+                        kernel, cudaFuncAttributeMaxDynamicSharedMemorySize, kSmemSize));
+#endif
+                }
+                kernel<<<grid, Ktraits::kNThreads, kSmemSize, stream>>>(params);
+                C10_CUDA_KERNEL_LAUNCH_CHECK();
+            });
         });
     });
 }
@@ -612,19 +622,20 @@ void selective_scan_fwd(const torch::Tensor &u, const torch::Tensor &delta,
 
     at::Tensor z, out_z;
     const bool has_z = z_.has_value();
-    TORCH_CHECK(has_z, "has_z = False is disabled in favor of reduced binary size")
-    z = z_.value();
-    TORCH_CHECK(z.scalar_type() == input_type);
-    TORCH_CHECK(z.is_cuda());
-    TORCH_CHECK(z.stride(-1) == 1 || z.size(-1) == 1);
-    if (varlen){
-        CHECK_SHAPE(z, dim, seqlen);
-    } else {
-        CHECK_SHAPE(z, batch_size, dim, seqlen);
+    if (has_z) {
+        z = z_.value();
+        TORCH_CHECK(z.scalar_type() == input_type);
+        TORCH_CHECK(z.is_cuda());
+        TORCH_CHECK(z.stride(-1) == 1 || z.size(-1) == 1);
+        if (varlen){
+            CHECK_SHAPE(z, dim, seqlen);
+        } else {
+            CHECK_SHAPE(z, batch_size, dim, seqlen);
+        }
+        
+        out_z = z;
     }
 
-    out_z = z;
-
     // 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);
@@ -653,4 +664,3 @@ void selective_scan_fwd(const torch::Tensor &u, const torch::Tensor &delta,
         selective_scan_fwd_cuda<input_t, weight_t>(params, stream);
     });
 }
-

csrc/moe/marlin_moe_wna16/marlin_template.h

@@ -1255,8 +1255,6 @@ __global__ void Marlin(
     if constexpr (has_zp && !is_zp_float) {
       if (is_new_zp) {
         if constexpr (group_blocks == -1) is_first_matmul_in_slice = false;
-        FragB frag_zp_0;
-        FragB frag_zp_1;
         int zp_quant_0, zp_quant_1;
 
         if constexpr (w_type.size_bits() == 4) {

csrc/moe/moe_align_sum_kernels.cu

@@ -1,6 +1,7 @@
 #include <torch/all.h>
 #include <ATen/cuda/CUDAContext.h>
 #include <c10/cuda/CUDAGuard.h>
+#include <cub/cub.cuh>
 
 #include <ATen/ATen.h>
 #include <ATen/cuda/Atomic.cuh>
@@ -19,9 +20,14 @@ __global__ void moe_align_block_size_kernel(
     int32_t* __restrict__ sorted_token_ids, int32_t* __restrict__ expert_ids,
     int32_t* __restrict__ total_tokens_post_pad, int32_t num_experts,
     int32_t padded_num_experts, int32_t experts_per_warp, int32_t block_size,
-    size_t numel, int32_t* __restrict__ cumsum) {
+    size_t numel, int32_t* __restrict__ cumsum, int32_t max_num_tokens_padded) {
   extern __shared__ int32_t shared_counts[];
 
+  // Initialize sorted_token_ids with numel
+  for (size_t it = threadIdx.x; it < max_num_tokens_padded; it += blockDim.x) {
+    sorted_token_ids[it] = numel;
+  }
+
   const int warp_id = threadIdx.x / WARP_SIZE;
   const int my_expert_start = warp_id * experts_per_warp;
 
@@ -45,18 +51,27 @@ __global__ void moe_align_block_size_kernel(
 
   __syncthreads();
 
-  if (threadIdx.x == 0) {
-    cumsum[0] = 0;
-    for (int i = 1; i <= num_experts; ++i) {
-      int expert_count = 0;
-      int warp_idx = (i - 1) / experts_per_warp;
-      int expert_offset = (i - 1) % experts_per_warp;
-      expert_count = shared_counts[warp_idx * experts_per_warp + expert_offset];
+  // Compute prefix sum over token counts per expert
+  using BlockScan = cub::BlockScan<int32_t, 1024>;
+  __shared__ typename BlockScan::TempStorage temp_storage;
 
-      cumsum[i] =
-          cumsum[i - 1] + CEILDIV(expert_count, block_size) * block_size;
-    }
-    *total_tokens_post_pad = cumsum[num_experts];
+  int expert_count = 0;
+  int expert_id = threadIdx.x;
+  if (expert_id < num_experts) {
+    int warp_idx = expert_id / experts_per_warp;
+    int expert_offset = expert_id % experts_per_warp;
+    expert_count = shared_counts[warp_idx * experts_per_warp + expert_offset];
+    expert_count = CEILDIV(expert_count, block_size) * block_size;
+  }
+
+  int cumsum_val;
+  BlockScan(temp_storage).ExclusiveSum(expert_count, cumsum_val);
+  if (expert_id <= num_experts) {
+    cumsum[expert_id] = cumsum_val;
+  }
+
+  if (expert_id == num_experts) {
+    *total_tokens_post_pad = cumsum_val;
   }
 
   __syncthreads();
@@ -67,6 +82,13 @@ __global__ void moe_align_block_size_kernel(
       expert_ids[i / block_size] = threadIdx.x;
     }
   }
+
+  // Fill remaining expert_ids with 0
+  const size_t fill_start_idx = cumsum[num_experts] / block_size + threadIdx.x;
+  const size_t expert_ids_size = CEILDIV(max_num_tokens_padded, block_size);
+  for (size_t i = fill_start_idx; i < expert_ids_size; i += blockDim.x) {
+    expert_ids[i] = 0;
+  }
 }
 
 template <typename scalar_t>
@@ -105,7 +127,12 @@ __global__ void moe_align_block_size_small_batch_expert_kernel(
     const scalar_t* __restrict__ topk_ids,
     int32_t* __restrict__ sorted_token_ids, int32_t* __restrict__ expert_ids,
     int32_t* __restrict__ total_tokens_post_pad, int32_t num_experts,
-    int32_t block_size, size_t numel) {
+    int32_t block_size, size_t numel, int32_t max_num_tokens_padded) {
+  // Initialize sorted_token_ids with numel
+  for (size_t it = threadIdx.x; it < max_num_tokens_padded; it += blockDim.x) {
+    sorted_token_ids[it] = numel;
+  }
+
   const size_t tid = threadIdx.x;
   const size_t stride = blockDim.x;
 
@@ -153,6 +180,13 @@ __global__ void moe_align_block_size_small_batch_expert_kernel(
     }
   }
 
+  // Fill remaining expert_ids with 0
+  const size_t fill_start_idx = cumsum[num_experts] / block_size + threadIdx.x;
+  const size_t expert_ids_size = CEILDIV(max_num_tokens_padded, block_size);
+  for (size_t i = fill_start_idx; i < expert_ids_size; i += blockDim.x) {
+    expert_ids[i] = 0;
+  }
+
   for (size_t i = tid; i < numel; i += stride) {
     int32_t expert_id = topk_ids[i];
     int32_t rank_post_pad =
@@ -179,13 +213,17 @@ void moe_align_block_size(torch::Tensor topk_ids, int64_t num_experts,
   int threads = 1024;
   threads = ((threads + WARP_SIZE - 1) / WARP_SIZE) * WARP_SIZE;
 
+  // BlockScan uses 1024 threads and assigns one thread per expert.
+  TORCH_CHECK(padded_num_experts < 1024,
+              "padded_num_experts must be less than 1024");
+
   VLLM_DISPATCH_INTEGRAL_AND_UNSIGNED_TYPES(
       topk_ids.scalar_type(), "moe_align_block_size_kernel", [&] {
         // calc needed amount of shared mem for `cumsum` tensors
         auto options_int =
             torch::TensorOptions().dtype(torch::kInt).device(topk_ids.device());
         torch::Tensor cumsum_buffer =
-            torch::zeros({num_experts + 1}, options_int);
+            torch::empty({num_experts + 1}, options_int);
         bool small_batch_expert_mode =
             (topk_ids.numel() < 1024) && (num_experts <= 64);
 
@@ -203,7 +241,7 @@ void moe_align_block_size(torch::Tensor topk_ids, int64_t num_experts,
               sorted_token_ids.data_ptr<int32_t>(),
               experts_ids.data_ptr<int32_t>(),
               num_tokens_post_pad.data_ptr<int32_t>(), num_experts, block_size,
-              topk_ids.numel());
+              topk_ids.numel(), sorted_token_ids.size(0));
         } else {
           auto align_kernel = vllm::moe::moe_align_block_size_kernel<scalar_t>;
 
@@ -217,7 +255,8 @@ void moe_align_block_size(torch::Tensor topk_ids, int64_t num_experts,
               experts_ids.data_ptr<int32_t>(),
               num_tokens_post_pad.data_ptr<int32_t>(), num_experts,
               padded_num_experts, experts_per_warp, block_size,
-              topk_ids.numel(), cumsum_buffer.data_ptr<int32_t>());
+              topk_ids.numel(), cumsum_buffer.data_ptr<int32_t>(),
+              sorted_token_ids.size(0));
 
           const int block_threads = std::min(256, (int)threads);
           const int num_blocks =
@@ -239,7 +278,7 @@ void moe_sum(torch::Tensor& input,   // [num_tokens, topk, hidden_size]
              torch::Tensor& output)  // [num_tokens, hidden_size]
 {
   const int hidden_size = input.size(-1);
-  const int num_tokens = output.numel() / hidden_size;
+  const auto num_tokens = output.numel() / hidden_size;
   const int topk = input.size(1);
 
   dim3 grid(num_tokens);

csrc/moe/moe_permute_unpermute_op.cu

@@ -10,32 +10,28 @@
 
 void moe_permute(
     const torch::Tensor& input,                      // [n_token, hidden]
-    const torch::Tensor& topk_weights,               //[n_token, topk]
-    torch::Tensor& topk_ids,                         // [n_token, topk]
+    const torch::Tensor& topk_ids,                   // [n_token, topk]
     const torch::Tensor& token_expert_indices,       // [n_token, topk]
     const std::optional<torch::Tensor>& expert_map,  // [n_expert]
     int64_t n_expert, int64_t n_local_expert, int64_t topk,
     const std::optional<int64_t>& align_block_size,
-    torch::Tensor&
-        permuted_input,  // [topk * n_token/align_block_size_m, hidden]
+    torch::Tensor& permuted_input,             // [permuted_size, hidden]
     torch::Tensor& expert_first_token_offset,  // [n_local_expert + 1]
-    torch::Tensor& src_row_id2dst_row_id_map,  // [n_token, topk]
+    torch::Tensor& inv_permuted_idx,           // [n_token, topk]
+    torch::Tensor& permuted_idx,               // [permute_size]
     torch::Tensor& m_indices) {                // [align_expand_m]
-  TORCH_CHECK(topk_weights.scalar_type() == at::ScalarType::Float,
-              "topk_weights must be float32");
   TORCH_CHECK(expert_first_token_offset.scalar_type() == at::ScalarType::Long,
               "expert_first_token_offset must be int64");
   TORCH_CHECK(topk_ids.scalar_type() == at::ScalarType::Int,
               "topk_ids must be int32");
   TORCH_CHECK(token_expert_indices.scalar_type() == at::ScalarType::Int,
               "token_expert_indices must be int32");
-  TORCH_CHECK(src_row_id2dst_row_id_map.scalar_type() == at::ScalarType::Int,
-              "src_row_id2dst_row_id_map must be int32");
+  TORCH_CHECK(inv_permuted_idx.scalar_type() == at::ScalarType::Int,
+              "inv_permuted_idx must be int32");
   TORCH_CHECK(expert_first_token_offset.size(0) == n_local_expert + 1,
               "expert_first_token_offset shape != n_local_expert+1")
-  TORCH_CHECK(
-      src_row_id2dst_row_id_map.sizes() == token_expert_indices.sizes(),
-      "token_expert_indices shape must be same as src_row_id2dst_row_id_map");
+  TORCH_CHECK(inv_permuted_idx.sizes() == token_expert_indices.sizes(),
+              "token_expert_indices shape must be same as inv_permuted_idx");
   auto n_token = input.sizes()[0];
   auto n_hidden = input.sizes()[1];
   auto align_block_size_value =
@@ -46,8 +42,9 @@ void moe_permute(
   auto sort_workspace = torch::empty(
       {sorter_size},
       torch::dtype(torch::kInt8).device(torch::kCUDA).requires_grad(false));
+  auto copy_topk_ids = topk_ids.clone();  // copy topk_ids for preprocess
   auto permuted_experts_id = torch::empty_like(topk_ids);
-  auto dst_row_id2src_row_id_map = torch::empty_like(src_row_id2dst_row_id_map);
+  auto sorted_row_idx = torch::empty_like(inv_permuted_idx);
   auto align_expert_first_token_offset =
       torch::zeros_like(expert_first_token_offset);
 
@@ -67,24 +64,22 @@ void moe_permute(
     const int* expert_map_ptr = get_ptr<int>(expert_map.value());
     valid_num_ptr =
         get_ptr<int64_t>(expert_first_token_offset) + n_local_expert;
-    preprocessTopkIdLauncher(get_ptr<int>(topk_ids), n_token * topk,
+    preprocessTopkIdLauncher(get_ptr<int>(copy_topk_ids), n_token * topk,
                              expert_map_ptr, n_expert, stream);
   }
   // expert sort topk expert id and scan expert id get expert_first_token_offset
-  sortAndScanExpert(get_ptr<int>(topk_ids), get_ptr<int>(token_expert_indices),
-                    get_ptr<int>(permuted_experts_id),
-                    get_ptr<int>(dst_row_id2src_row_id_map),
-                    get_ptr<int64_t>(expert_first_token_offset), n_token,
-                    n_expert, n_local_expert, topk, sorter,
-                    get_ptr<int>(sort_workspace), stream);
+  sortAndScanExpert(
+      get_ptr<int>(copy_topk_ids), get_ptr<int>(token_expert_indices),
+      get_ptr<int>(permuted_experts_id), get_ptr<int>(sorted_row_idx),
+      get_ptr<int64_t>(expert_first_token_offset), n_token, n_expert,
+      n_local_expert, topk, sorter, get_ptr<int>(sort_workspace), stream);
 
   // dispatch expandInputRowsKernelLauncher
   MOE_DISPATCH(input.scalar_type(), [&] {
     expandInputRowsKernelLauncher<scalar_t>(
         get_ptr<scalar_t>(input), get_ptr<scalar_t>(permuted_input),
-        get_ptr<float>(topk_weights), get_ptr<int>(permuted_experts_id),
-        get_ptr<int>(dst_row_id2src_row_id_map),
-        get_ptr<int>(src_row_id2dst_row_id_map),
+        get_ptr<int>(permuted_experts_id), get_ptr<int>(sorted_row_idx),
+        get_ptr<int>(inv_permuted_idx), get_ptr<int>(permuted_idx),
         get_ptr<int64_t>(expert_first_token_offset), n_token, valid_num_ptr,
         n_hidden, topk, n_local_expert, align_block_size_value, stream);
   });
@@ -101,32 +96,34 @@ void moe_permute(
 }
 
 void moe_unpermute(
-    const torch::Tensor& permuted_hidden_states,     // [n_token * topk, hidden]
-    const torch::Tensor& topk_weights,               //[n_token, topk]
-    const torch::Tensor& topk_ids,                   // [n_token, topk]
-    const torch::Tensor& src_row_id2dst_row_id_map,  // [n_token, topk]
-    const torch::Tensor& expert_first_token_offset,  // [n_local_expert+1]
-    int64_t n_expert, int64_t n_local_expert, int64_t topk,
+    const torch::Tensor& permuted_hidden_states,  // [n_token * topk, hidden]
+    const torch::Tensor& topk_weights,            // [n_token, topk]
+    const torch::Tensor& inv_permuted_idx,        // [n_token, topk]
+    const std::optional<torch::Tensor>&
+        expert_first_token_offset,  // [n_local_expert+1]
+    int64_t topk,
     torch::Tensor& hidden_states  // [n_token, hidden]
 ) {
-  TORCH_CHECK(src_row_id2dst_row_id_map.sizes() == topk_ids.sizes(),
-              "topk_ids shape must be same as src_row_id2dst_row_id_map");
-  TORCH_CHECK(topk_ids.scalar_type() == at::ScalarType::Int,
-              "topk_ids must be int32");
   TORCH_CHECK(
       permuted_hidden_states.scalar_type() == hidden_states.scalar_type(),
-      "topk_ids dtype must be same as src_row_id2dst_row_id_map");
+      "permuted_hidden_states dtype must be same as hidden_states");
   auto n_token = hidden_states.size(0);
   auto n_hidden = hidden_states.size(1);
   auto stream = at::cuda::getCurrentCUDAStream().stream();
-  const int64_t* valid_ptr =
-      get_ptr<int64_t>(expert_first_token_offset) + n_local_expert;
+
+  int64_t const* valid_ptr = nullptr;
+  if (expert_first_token_offset.has_value()) {
+    int n_local_expert = expert_first_token_offset.value().size(0) - 1;
+    valid_ptr =
+        get_ptr<int64_t>(expert_first_token_offset.value()) + n_local_expert;
+  }
+
   MOE_DISPATCH(hidden_states.scalar_type(), [&] {
     finalizeMoeRoutingKernelLauncher<scalar_t, scalar_t>(
         get_ptr<scalar_t>(permuted_hidden_states),
         get_ptr<scalar_t>(hidden_states), get_ptr<float>(topk_weights),
-        get_ptr<int>(src_row_id2dst_row_id_map), get_ptr<int>(topk_ids),
-        n_token, n_hidden, topk, valid_ptr, stream);
+        get_ptr<int>(inv_permuted_idx), n_token, n_hidden, topk, valid_ptr,
+        stream);
   });
 }
 

csrc/moe/permute_unpermute_kernels/moe_permute_unpermute_kernel.cu

@@ -177,7 +177,7 @@ __global__ void getMIndicesKernel(int64_t* expert_first_token_offset,
   int tidx = threadIdx.x;
   extern __shared__ int64_t smem_expert_first_token_offset[];
   for (int i = tidx; i <= num_local_expert; i += blockDim.x) {
-    smem_expert_first_token_offset[tidx] = __ldg(expert_first_token_offset + i);
+    smem_expert_first_token_offset[i] = __ldg(expert_first_token_offset + i);
   }
   __syncthreads();
   auto last_token_offset = smem_expert_first_token_offset[eidx + 1];

csrc/moe/permute_unpermute_kernels/moe_permute_unpermute_kernel.h

@@ -57,31 +57,19 @@ void sortAndScanExpert(int* expert_for_source_row, const int* source_rows,
 
 template <typename T>
 void expandInputRowsKernelLauncher(
-    T const* unpermuted_input, T* permuted_output,
-    const float* unpermuted_scales, int* sorted_experts,
+    T const* unpermuted_input, T* permuted_output, int* sorted_experts,
     int const* expanded_dest_row_to_expanded_source_row,
-    int* expanded_source_row_to_expanded_dest_row,
+    int* expanded_source_row_to_expanded_dest_row, int* permuted_idx,
     int64_t* expert_first_token_offset, int64_t const num_rows,
     int64_t const* num_valid_tokens_ptr, int64_t const cols, int const k,
     int num_local_experts, const int& align_block_size, cudaStream_t stream);
 
-// Final kernel to unpermute and scale
-// This kernel unpermutes the original data, does the k-way reduction and
-// performs the final skip connection.
-template <typename T, typename OutputType, bool CHECK_SKIPPED>
-__global__ void finalizeMoeRoutingKernel(
-    T const* expanded_permuted_rows, OutputType* reduced_unpermuted_output,
-    float const* scales, int const* expanded_source_row_to_expanded_dest_row,
-    int const* expert_for_source_row, int64_t const orig_cols, int64_t const k,
-    int64_t const* num_valid_ptr);
-
 template <class T, class OutputType>
 void finalizeMoeRoutingKernelLauncher(
     T const* expanded_permuted_rows, OutputType* reduced_unpermuted_output,
     float const* scales, int const* expanded_source_row_to_expanded_dest_row,
-    int const* expert_for_source_row, int64_t const num_rows,
-    int64_t const cols, int64_t const k, int64_t const* num_valid_ptr,
-    cudaStream_t stream);
+    int64_t const num_rows, int64_t const cols, int64_t const k,
+    int64_t const* num_valid_ptr, cudaStream_t stream);
 
 void preprocessTopkIdLauncher(int* topk_id_ptr, int size,
                               const int* expert_map_ptr, int num_experts,

csrc/moe/permute_unpermute_kernels/moe_permute_unpermute_kernel.inl

@@ -2,10 +2,9 @@
 
 template <typename T, bool CHECK_SKIPPED, bool ALIGN_BLOCK_SIZE>
 __global__ void expandInputRowsKernel(
-    T const* unpermuted_input, T* permuted_output,
-    const float* unpermuted_scales, int* sorted_experts,
+    T const* unpermuted_input, T* permuted_output, int* sorted_experts,
     int const* expanded_dest_row_to_expanded_source_row,
-    int* expanded_source_row_to_expanded_dest_row,
+    int* expanded_source_row_to_expanded_dest_row, int* permuted_idx,
     int64_t* expert_first_token_offset, int64_t const num_rows,
     int64_t const* num_dest_rows, int64_t const cols, int64_t k,
     int num_local_experts, int align_block_size) {
@@ -54,6 +53,10 @@ __global__ void expandInputRowsKernel(
     assert(expanded_dest_row <= INT32_MAX);
     expanded_source_row_to_expanded_dest_row[expanded_source_row] =
         static_cast<int>(expanded_dest_row);
+    // skip non local expert token
+    if (!CHECK_SKIPPED || blockIdx.x < *num_dest_rows) {
+      permuted_idx[expanded_dest_row] = expanded_source_row;
+    }
   }
 
   if (!CHECK_SKIPPED || blockIdx.x < *num_dest_rows) {
@@ -62,7 +65,7 @@ __global__ void expandInputRowsKernel(
     using DataElem = cutlass::Array<T, ELEM_PER_THREAD>;
 
     // Duplicate and permute rows
-    int64_t const source_row = expanded_source_row % num_rows;
+    int64_t const source_row = expanded_source_row / k;
 
     auto const* source_row_ptr =
         reinterpret_cast<DataElem const*>(unpermuted_input + source_row * cols);
@@ -82,10 +85,9 @@ __global__ void expandInputRowsKernel(
 
 template <typename T>
 void expandInputRowsKernelLauncher(
-    T const* unpermuted_input, T* permuted_output,
-    const float* unpermuted_scales, int* sorted_experts,
+    T const* unpermuted_input, T* permuted_output, int* sorted_experts,
     int const* expanded_dest_row_to_expanded_source_row,
-    int* expanded_source_row_to_expanded_dest_row,
+    int* expanded_source_row_to_expanded_dest_row, int* permuted_idx,
     int64_t* expert_first_token_offset, int64_t const num_rows,
     int64_t const* num_valid_tokens_ptr, int64_t const cols, int const k,
     int num_local_experts, const int& align_block_size, cudaStream_t stream) {
@@ -105,11 +107,11 @@ void expandInputRowsKernelLauncher(
   int64_t smem_size = sizeof(int64_t) * (num_local_experts + 1);
 
   func<<<blocks, threads, smem_size, stream>>>(
-      unpermuted_input, permuted_output, unpermuted_scales, sorted_experts,
+      unpermuted_input, permuted_output, sorted_experts,
       expanded_dest_row_to_expanded_source_row,
-      expanded_source_row_to_expanded_dest_row, expert_first_token_offset,
-      num_rows, num_valid_tokens_ptr, cols, k, num_local_experts,
-      align_block_size);
+      expanded_source_row_to_expanded_dest_row, permuted_idx,
+      expert_first_token_offset, num_rows, num_valid_tokens_ptr, cols, k,
+      num_local_experts, align_block_size);
 }
 
 template <class T, class U>
@@ -128,11 +130,9 @@ template <typename T, typename OutputType, bool CHECK_SKIPPED>
 __global__ void finalizeMoeRoutingKernel(
     T const* expanded_permuted_rows, OutputType* reduced_unpermuted_output,
     float const* scales, int const* expanded_source_row_to_expanded_dest_row,
-    int const* expert_for_source_row, int64_t const orig_cols, int64_t const k,
-    int64_t const* num_valid_ptr) {
+    int64_t const orig_cols, int64_t const k, int64_t const* num_valid_ptr) {
   assert(orig_cols % 4 == 0);
   int64_t const original_row = blockIdx.x;
-  int64_t const num_rows = gridDim.x;
   auto const offset = original_row * orig_cols;
   OutputType* reduced_row_ptr = reduced_unpermuted_output + offset;
   int64_t const num_valid = *num_valid_ptr;
@@ -159,14 +159,13 @@ __global__ void finalizeMoeRoutingKernel(
     ComputeElem thread_output;
     thread_output.fill(0);
     for (int k_idx = 0; k_idx < k; ++k_idx) {
-      int64_t const expanded_original_row = original_row + k_idx * num_rows;
+      int64_t const expanded_original_row = original_row * k + k_idx;
       int64_t const expanded_permuted_row =
           expanded_source_row_to_expanded_dest_row[expanded_original_row];
 
       int64_t const k_offset = original_row * k + k_idx;
       float const row_scale = scales[k_offset];
 
-      // Check after row_rescale has accumulated
       if (CHECK_SKIPPED && expanded_permuted_row >= num_valid) {
         continue;
       }
@@ -189,9 +188,8 @@ template <class T, class OutputType>
 void finalizeMoeRoutingKernelLauncher(
     T const* expanded_permuted_rows, OutputType* reduced_unpermuted_output,
     float const* scales, int const* expanded_source_row_to_expanded_dest_row,
-    int const* expert_for_source_row, int64_t const num_rows,
-    int64_t const cols, int64_t const k, int64_t const* num_valid_ptr,
-    cudaStream_t stream) {
+    int64_t const num_rows, int64_t const cols, int64_t const k,
+    int64_t const* num_valid_ptr, cudaStream_t stream) {
   int64_t const blocks = num_rows;
   int64_t const threads = 256;
   bool const check_finished = num_valid_ptr != nullptr;
@@ -201,6 +199,5 @@ void finalizeMoeRoutingKernelLauncher(
   auto* const kernel = func_map[check_finished];
   kernel<<<blocks, threads, 0, stream>>>(
       expanded_permuted_rows, reduced_unpermuted_output, scales,
-      expanded_source_row_to_expanded_dest_row, expert_for_source_row, cols, k,
-      num_valid_ptr);
+      expanded_source_row_to_expanded_dest_row, cols, k, num_valid_ptr);
 }

csrc/moe/topk_softmax_kernels.cu

@@ -190,8 +190,8 @@ __launch_bounds__(TPB) __global__ void moeTopK(
   2) This implementation assumes k is small, but will work for any k.
 */
 
-template <int VPT, int NUM_EXPERTS, int WARPS_PER_CTA, int BYTES_PER_LDG, typename IndType>
-__launch_bounds__(WARPS_PER_CTA* WARP_SIZE) __global__
+template <int VPT, int NUM_EXPERTS, int WARPS_PER_CTA, int BYTES_PER_LDG, int WARP_SIZE_PARAM, typename IndType>
+__launch_bounds__(WARPS_PER_CTA* WARP_SIZE_PARAM) __global__
     void topkGatingSoftmax(const float* input, const bool* finished, float* output, const int num_rows, IndType* indices,
         int* source_rows, const int k, const int start_expert, const int end_expert)
 {
@@ -209,12 +209,12 @@ __launch_bounds__(WARPS_PER_CTA* WARP_SIZE) __global__
 
     // Restrictions based on previous section.
     static_assert(VPT % ELTS_PER_LDG == 0, "The elements per thread must be a multiple of the elements per ldg");
-    static_assert(WARP_SIZE % THREADS_PER_ROW == 0, "The threads per row must cleanly divide the threads per warp");
+    static_assert(WARP_SIZE_PARAM % THREADS_PER_ROW == 0, "The threads per row must cleanly divide the threads per warp");
     static_assert(THREADS_PER_ROW == (THREADS_PER_ROW & -THREADS_PER_ROW), "THREADS_PER_ROW must be power of 2");
-    static_assert(THREADS_PER_ROW <= WARP_SIZE, "THREADS_PER_ROW can be at most warp size");
+    static_assert(THREADS_PER_ROW <= WARP_SIZE_PARAM, "THREADS_PER_ROW can be at most warp size");
 
     // We have NUM_EXPERTS elements per row. We specialize for small #experts
-    static constexpr int ELTS_PER_WARP = WARP_SIZE * VPT;
+    static constexpr int ELTS_PER_WARP = WARP_SIZE_PARAM * VPT;
     static constexpr int ROWS_PER_WARP = ELTS_PER_WARP / ELTS_PER_ROW;
     static constexpr int ROWS_PER_CTA = WARPS_PER_CTA * ROWS_PER_WARP;
 
@@ -393,41 +393,51 @@ __launch_bounds__(WARPS_PER_CTA* WARP_SIZE) __global__
 namespace detail
 {
 // Constructs some constants needed to partition the work across threads at compile time.
-template <int EXPERTS, int BYTES_PER_LDG>
+template <int EXPERTS, int BYTES_PER_LDG, int WARP_SIZE_PARAM>
 struct TopkConstants
 {
     static constexpr int ELTS_PER_LDG = BYTES_PER_LDG / sizeof(float);
-    static_assert(EXPERTS / (ELTS_PER_LDG * WARP_SIZE) == 0 || EXPERTS % (ELTS_PER_LDG * WARP_SIZE) == 0, "");
-    static constexpr int VECs_PER_THREAD = MAX(1, EXPERTS / (ELTS_PER_LDG * WARP_SIZE));
+    static_assert(EXPERTS / (ELTS_PER_LDG * WARP_SIZE_PARAM) == 0 || EXPERTS % (ELTS_PER_LDG * WARP_SIZE_PARAM) == 0, "");
+    static constexpr int VECs_PER_THREAD = MAX(1, EXPERTS / (ELTS_PER_LDG * WARP_SIZE_PARAM));
     static constexpr int VPT = VECs_PER_THREAD * ELTS_PER_LDG;
     static constexpr int THREADS_PER_ROW = EXPERTS / VPT;
-    static constexpr int ROWS_PER_WARP = WARP_SIZE / THREADS_PER_ROW;
+    static const int ROWS_PER_WARP = WARP_SIZE_PARAM / THREADS_PER_ROW;
 };
 } // namespace detail
 
-template <int EXPERTS, int WARPS_PER_TB, typename IndType>
+template <int EXPERTS, int WARPS_PER_TB, int WARP_SIZE_PARAM, typename IndType>
 void topkGatingSoftmaxLauncherHelper(const float* input, const bool* finished, float* output, IndType* indices,
     int* source_row, const int num_rows, const int k, const int start_expert, const int end_expert, cudaStream_t stream)
 {
     static constexpr std::size_t MAX_BYTES_PER_LDG = 16;
 
     static constexpr int BYTES_PER_LDG = MIN(MAX_BYTES_PER_LDG, sizeof(float) * EXPERTS);
-    using Constants = detail::TopkConstants<EXPERTS, BYTES_PER_LDG>;
+    using Constants = detail::TopkConstants<EXPERTS, BYTES_PER_LDG, WARP_SIZE_PARAM>;
     static constexpr int VPT = Constants::VPT;
     static constexpr int ROWS_PER_WARP = Constants::ROWS_PER_WARP;
     const int num_warps = (num_rows + ROWS_PER_WARP - 1) / ROWS_PER_WARP;
     const int num_blocks = (num_warps + WARPS_PER_TB - 1) / WARPS_PER_TB;
 
-    dim3 block_dim(WARP_SIZE, WARPS_PER_TB);
-    topkGatingSoftmax<VPT, EXPERTS, WARPS_PER_TB, BYTES_PER_LDG><<<num_blocks, block_dim, 0, stream>>>(
+    dim3 block_dim(WARP_SIZE_PARAM, WARPS_PER_TB);
+    topkGatingSoftmax<VPT, EXPERTS, WARPS_PER_TB, BYTES_PER_LDG, WARP_SIZE_PARAM><<<num_blocks, block_dim, 0, stream>>>(
         input, finished, output, num_rows, indices, source_row, k, start_expert, end_expert);
 }
 
-#define LAUNCH_SOFTMAX(NUM_EXPERTS, WARPS_PER_TB)                       \
-    topkGatingSoftmaxLauncherHelper<NUM_EXPERTS, WARPS_PER_TB>(         \
-        gating_output, nullptr, topk_weights, topk_indices,            \
-        token_expert_indices, num_tokens, topk, 0, num_experts,         \
-        stream);
+#define LAUNCH_SOFTMAX(NUM_EXPERTS, WARPS_PER_TB)                                \
+    switch (warpSize) {                                                          \
+        case 32:                                                                 \
+            topkGatingSoftmaxLauncherHelper<NUM_EXPERTS, WARPS_PER_TB, 32>(      \
+                gating_output, nullptr, topk_weights, topk_indices,              \
+                token_expert_indices, num_tokens, topk, 0, num_experts, stream); \
+            break;                                                               \
+        case 64:                                                                 \
+            topkGatingSoftmaxLauncherHelper<NUM_EXPERTS, WARPS_PER_TB, 64>(      \
+                gating_output, nullptr, topk_weights, topk_indices,              \
+                token_expert_indices, num_tokens, topk, 0, num_experts, stream); \
+            break;                                                               \
+        default:                                                                 \
+            TORCH_CHECK(false, "Unsupported warp size: ", warpSize);             \
+    }
 
 template <typename IndType>
 void topkGatingSoftmaxKernelLauncher(
@@ -441,6 +451,7 @@ void topkGatingSoftmaxKernelLauncher(
     const int topk,
     cudaStream_t stream) {
     static constexpr int WARPS_PER_TB = 4;
+    auto warpSize = WARP_SIZE;
     switch (num_experts) {
         case 1:
             LAUNCH_SOFTMAX(1, WARPS_PER_TB);
@@ -492,7 +503,7 @@ void topk_softmax(
     torch::Tensor& gating_output)               // [num_tokens, num_experts]
 {
     const int num_experts = gating_output.size(-1);
-    const int num_tokens = gating_output.numel() / num_experts;
+    const auto num_tokens = gating_output.numel() / num_experts;
     const int topk = topk_weights.size(-1);
 
     const bool is_pow_2 = (num_experts != 0) && ((num_experts & (num_experts - 1)) == 0);

csrc/moe/torch_bindings.cpp

@@ -56,18 +56,17 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, m) {
       " -> Tensor");
 
   m.def(
-      "moe_permute(Tensor input, Tensor topk_weight, Tensor! topk_ids,"
+      "moe_permute(Tensor input, Tensor topk_ids,"
       "Tensor token_expert_indices, Tensor? expert_map, int n_expert,"
       "int n_local_expert,"
       "int topk, int? align_block_size,Tensor! permuted_input, Tensor! "
-      "expert_first_token_offset, Tensor! src_row_id2dst_row_id_map, Tensor! "
-      "m_indices)->()");
+      "expert_first_token_offset, Tensor! inv_permuted_idx, Tensor! "
+      "permuted_idx, Tensor! m_indices)->()");
 
   m.def(
       "moe_unpermute(Tensor permuted_hidden_states, Tensor topk_weights,"
-      "Tensor topk_ids,Tensor src_row_id2dst_row_id_map, Tensor "
-      "expert_first_token_offset, int n_expert, int n_local_expert,int "
-      "topk, Tensor! hidden_states)->()");
+      "Tensor inv_permuted_idx, Tensor? expert_first_token_offset, "
+      "int topk, Tensor! hidden_states)->()");
 
   m.def("moe_permute_unpermute_supported() -> bool");
   m.impl("moe_permute_unpermute_supported", &moe_permute_unpermute_supported);

csrc/ops.h

@@ -287,6 +287,16 @@ void scaled_fp4_experts_quant(
     torch::Tensor const& input, torch::Tensor const& input_global_scale,
     torch::Tensor const& input_offset_by_experts,
     torch::Tensor const& output_scale_offset_by_experts);
+
+void per_token_group_quant_fp8(const torch::Tensor& input,
+                               torch::Tensor& output_q, torch::Tensor& output_s,
+                               int64_t group_size, double eps, double fp8_min,
+                               double fp8_max, bool scale_ue8m0);
+
+void per_token_group_quant_int8(const torch::Tensor& input,
+                                torch::Tensor& output_q,
+                                torch::Tensor& output_s, int64_t group_size,
+                                double eps, double int8_min, double int8_max);
 #endif
 
 void static_scaled_int8_quant(torch::Tensor& out, torch::Tensor const& input,
@@ -326,22 +336,6 @@ void selective_scan_fwd(const torch::Tensor& u, const torch::Tensor& delta,
                         const std::optional<torch::Tensor>& has_initial_state,
                         const torch::Tensor& ssm_states, int64_t pad_slot_id);
 
-void causal_conv1d_update(const at::Tensor& x, const at::Tensor& conv_state,
-                          const at::Tensor& weight,
-                          const std::optional<at::Tensor>& bias_,
-                          bool silu_activation,
-                          const std::optional<at::Tensor>& cache_seqlens_,
-                          const std::optional<at::Tensor>& conv_state_indices_,
-                          int64_t pad_slot_id);
-
-void causal_conv1d_fwd(const at::Tensor& x, const at::Tensor& weight,
-                       const std::optional<at::Tensor>& bias_,
-                       const std::optional<at::Tensor>& conv_states,
-                       const std::optional<at::Tensor>& query_start_loc,
-                       const std::optional<at::Tensor>& cache_indices,
-                       const std::optional<at::Tensor>& has_initial_state,
-                       bool silu_activation, int64_t pad_slot_id);
-
 using fptr_t = int64_t;
 fptr_t init_custom_ar(const std::vector<int64_t>& fake_ipc_ptrs,
                       torch::Tensor& rank_data, int64_t rank,

csrc/quantization/activation_kernels.cu

@@ -4,7 +4,7 @@
 
 #include <cmath>
 #include "core/math.hpp"
-#include "cuda_compat.h"
+#include "../cuda_compat.h"
 #include "dispatch_utils.h"
 
 #include "quantization/fp8/common.cuh"

csrc/quantization/compressed_tensors/int8_quant_kernels.cu

@@ -1,6 +1,8 @@
 #include <ATen/cuda/CUDAContext.h>
 #include <torch/all.h>
 
+#include "../per_token_group_quant_8bit.h"
+
 #include <cmath>
 
 #include "../../dispatch_utils.h"
@@ -162,10 +164,11 @@ __global__ void dynamic_scaled_int8_quant_kernel(
 
   // calculate for absmax
   float thread_max = 0.f;
-  for (int i = tid; i < hidden_size; i += stride) {
-    const auto v = fabsf(static_cast<float>(row_in[i]));
-    thread_max = fmaxf(thread_max, v);
-  }
+  vectorize_read_with_alignment<16>(
+      row_in, hidden_size, tid, stride, [&] __device__(const scalar_t& src) {
+        const float v = fabsf(static_cast<float>(src));
+        thread_max = fmaxf(thread_max, v);
+      });
   using BlockReduce = cub::BlockReduce<float, 256>;
   __shared__ typename BlockReduce::TempStorage tmp;
   float block_max = BlockReduce(tmp).Reduce(thread_max, cub::Max{}, blockDim.x);
@@ -232,9 +235,10 @@ __global__ void dynamic_scaled_int8_azp_quant_kernel(
 
   // 1. calculate min & max
   MinMax thread_mm;
-  for (int i = tid; i < hidden_size; i += stride) {
-    thread_mm += static_cast<float>(row_in[i]);
-  }
+  vectorize_read_with_alignment<16>(row_in, hidden_size, tid, stride,
+                                    [&] __device__(const scalar_t& src) {
+                                      thread_mm += static_cast<float>(src);
+                                    });
 
   using BlockReduce = cub::BlockReduce<MinMax, 256>;
   __shared__ typename BlockReduce::TempStorage tmp;
@@ -334,3 +338,11 @@ void dynamic_scaled_int8_quant(
         }
       });
 }
+
+void per_token_group_quant_int8(const torch::Tensor& input,
+                                torch::Tensor& output_q,
+                                torch::Tensor& output_s, int64_t group_size,
+                                double eps, double int8_min, double int8_max) {
+  per_token_group_quant_8bit(input, output_q, output_s, group_size, eps,
+                             int8_min, int8_max);
+}

csrc/quantization/cutlass_w8a8/c3x/scaled_mm.cuh

@@ -51,7 +51,8 @@ struct cutlass_3x_gemm {
   // These are the minimum alignments needed for the kernels to compile
   static constexpr int AlignmentAB =
       128 / cutlass::sizeof_bits<ElementAB>::value;
-  static constexpr int AlignmentCD = 4;
+  static constexpr int AlignmentCD =
+      128 / cutlass::sizeof_bits<ElementD>::value;
 
   using CollectiveEpilogue =
       typename cutlass::epilogue::collective::CollectiveBuilder<
@@ -144,4 +145,65 @@ struct cutlass_3x_gemm_sm100 {
       Shape<int, int, int, int>, CollectiveMainloop, CollectiveEpilogue, void>;
 };
 
+template <typename ElementAB_, typename ElementD_,
+          template <typename, typename, typename> typename Epilogue_,
+          typename TileShape, typename ClusterShape, typename KernelSchedule,
+          typename EpilogueSchedule>
+struct cutlass_3x_gemm_sm120 {
+  using ElementAB = ElementAB_;
+  using LayoutA = cutlass::layout::RowMajor;
+  static constexpr int AlignmentA =
+      128 / cutlass::sizeof_bits<ElementAB>::value;
+
+  using LayoutB = cutlass::layout::ColumnMajor;
+  static constexpr int AlignmentB =
+      128 / cutlass::sizeof_bits<ElementAB>::value;
+
+  using ElementC = void;
+  using LayoutC = cutlass::layout::RowMajor;
+  static constexpr int AlignmentC =
+      128 / cutlass::sizeof_bits<ElementD_>::value;
+
+  using ElementD = ElementD_;
+  using LayoutD = cutlass::layout::RowMajor;
+  static constexpr int AlignmentD = AlignmentC;
+
+  using ElementAcc =
+      typename std::conditional<std::is_same_v<ElementAB, int8_t>, int32_t,
+                                float>::type;
+  using Epilogue = Epilogue_<ElementAcc, ElementD, TileShape>;
+
+  // MMA type
+  using ElementAccumulator = float;
+
+  // Epilogue types
+  using ElementBias = cutlass::half_t;
+  using ElementCompute = float;
+  using ElementAux = ElementD;
+  using LayoutAux = LayoutD;
+  using ElementAmax = float;
+
+  using EVTCompute = typename Epilogue::EVTCompute;
+
+  using CollectiveEpilogue =
+      typename cutlass::epilogue::collective::CollectiveBuilder<
+          cutlass::arch::Sm120, cutlass::arch::OpClassTensorOp, TileShape,
+          ClusterShape, cutlass::epilogue::collective::EpilogueTileAuto,
+          ElementAccumulator, ElementCompute, ElementC, LayoutC, AlignmentC,
+          ElementD, LayoutD, AlignmentD, EpilogueSchedule,
+          EVTCompute>::CollectiveOp;
+
+  using CollectiveMainloop =
+      typename cutlass::gemm::collective::CollectiveBuilder<
+          cutlass::arch::Sm120, cutlass::arch::OpClassTensorOp, ElementAB,
+          LayoutA, AlignmentA, ElementAB, LayoutB, AlignmentB,
+          ElementAccumulator, TileShape, ClusterShape,
+          cutlass::gemm::collective::StageCountAutoCarveout<static_cast<int>(
+              sizeof(typename CollectiveEpilogue::SharedStorage))>,
+          KernelSchedule>::CollectiveOp;
+
+  using GemmKernel = cutlass::gemm::kernel::GemmUniversal<
+      Shape<int, int, int, int>, CollectiveMainloop, CollectiveEpilogue, void>;
+};
+
 }  // namespace vllm

csrc/quantization/cutlass_w8a8/c3x/scaled_mm_kernels.hpp

@@ -36,6 +36,12 @@ void cutlass_scaled_mm_sm100_fp8(torch::Tensor& out, torch::Tensor const& a,
                                  torch::Tensor const& b_scales,
                                  std::optional<torch::Tensor> const& bias);
 
+void cutlass_scaled_mm_sm120_fp8(torch::Tensor& out, torch::Tensor const& a,
+                                 torch::Tensor const& b,
+                                 torch::Tensor const& a_scales,
+                                 torch::Tensor const& b_scales,
+                                 std::optional<torch::Tensor> const& bias);
+
 void cutlass_scaled_mm_blockwise_sm100_fp8(torch::Tensor& out,
                                            torch::Tensor const& a,
                                            torch::Tensor const& b,