[V1] TPU - Remove self.kv_caches

[V1][Frontend] Add Testing For V1 Runtime Parameters (#14159 )
Signed-off-by: rshaw@neuralmagic.com <rshaw@neuralmagic.com>
2025-03-05 20:42:05 +00:00 · 2025-03-05 14:18:55 +00:00 · 2025-03-05 21:30:00 +08:00 · 2025-03-05 12:29:50 +00:00 · 2025-03-05 11:44:10 +00:00 · 2025-03-05 08:49:44 +00:00
1738 changed files with 149957 additions and 43764 deletions
@@ -1,9 +1,14 @@
+# SPDX-License-Identifier: Apache-2.0
+
 import os
 import sys
 import zipfile

-# Read the VLLM_MAX_SIZE_MB environment variable, defaulting to 250 MB
-VLLM_MAX_SIZE_MB = int(os.environ.get('VLLM_MAX_SIZE_MB', 250))
+# Read the VLLM_MAX_SIZE_MB environment variable, defaulting to 400 MiB
+# Note that we have 400 MiB quota, please use it wisely.
+# See https://github.com/pypi/support/issues/3792 .
+# Please also sync the value with the one in Dockerfile.
+VLLM_MAX_SIZE_MB = int(os.environ.get('VLLM_MAX_SIZE_MB', 400))


 def print_top_10_largest_files(zip_file):
@@ -0,0 +1,26 @@
+# SPDX-License-Identifier: Apache-2.0
+
+import argparse
+import os
+
+template = """<!DOCTYPE html>
+<html>
+    <body>
+    <h1>Links for vLLM</h1/>
+        <a href="../{wheel_html_escaped}">{wheel}</a><br/>
+    </body>
+</html>
+"""
+
+parser = argparse.ArgumentParser()
+parser.add_argument("--wheel", help="The wheel path.", required=True)
+args = parser.parse_args()
+
+filename = os.path.basename(args.wheel)
+
+with open("index.html", "w") as f:
+    print(f"Generated index.html for {args.wheel}")
+    # cloudfront requires escaping the '+' character
+    f.write(
+        template.format(wheel=filename,
+                        wheel_html_escaped=filename.replace("+", "%2B")))
@@ -0,0 +1,11 @@
+# bash ./run-lm-eval-gsm-vllm-baseline.sh -m nm-testing/SparseLlama-3.1-8B-gsm8k-pruned.2of4-chnl_wts_per_tok_dyn_act_fp8-BitM -b "auto" -t 2
+model_name: "nm-testing/SparseLlama-3.1-8B-gsm8k-pruned.2of4-chnl_wts_per_tok_dyn_act_fp8-BitM"
+tasks:
+- name: "gsm8k"
+  metrics:
+  - name: "exact_match,strict-match"
+    value: 0.6353
+  - name: "exact_match,flexible-extract"
+    value: 0.637
+limit: null
+num_fewshot: null 
@@ -1,3 +1,4 @@
+# SPDX-License-Identifier: Apache-2.0
 """
 LM eval harness on model to compare vs HF baseline computed offline.
 Configs are found in configs/$MODEL.yaml
@@ -1,15 +1,13 @@
 # vLLM benchmark suite

-
 ## Introduction

 This directory contains two sets of benchmark for vllm.
+
 - Performance benchmark: benchmark vllm's performance under various workload, for **developers** to gain clarity on whether their PR improves/degrades vllm's performance
 - Nightly benchmark: compare vllm's performance against alternatives (tgi, trt-llm and lmdeploy), for **the public** to know when to choose vllm.

-
-See  [vLLM performance dashboard](https://perf.vllm.ai) for the latest performance benchmark results and [vLLM GitHub README](https://github.com/vllm-project/vllm/blob/main/README.md) for latest nightly benchmark results.
-
+See [vLLM performance dashboard](https://perf.vllm.ai) for the latest performance benchmark results and [vLLM GitHub README](https://github.com/vllm-project/vllm/blob/main/README.md) for latest nightly benchmark results.

 ## Performance benchmark quick overview

@@ -19,17 +17,14 @@ See  [vLLM performance dashboard](https://perf.vllm.ai) for the latest performan

 **For benchmarking developers**: please try your best to constraint the duration of benchmarking to about 1 hr so that it won't take forever to run.

-
 ## Nightly benchmark quick overview

-**Benchmarking Coverage**: Fix-qps serving on A100 (the support for FP8 benchmark on H100 is coming!) on Llama-3 8B, 70B and Mixtral 8x7B. 
+**Benchmarking Coverage**: Fix-qps serving on A100 (the support for FP8 benchmark on H100 is coming!) on Llama-3 8B, 70B and Mixtral 8x7B.

 **Benchmarking engines**: vllm, TGI, trt-llm and lmdeploy.

 **Benchmarking Duration**: about 3.5hrs.

-
-
 ## Trigger the benchmark

 Performance benchmark will be triggered when:
@@ -39,16 +34,11 @@ Performance benchmark will be triggered when:
 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.

-
-#### Latency test
+### Latency test

 Here is an example of one test inside `latency-tests.json`:

@@ -68,23 +58,25 @@ 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 `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`

 Note that the performance numbers are highly sensitive to the value of the parameters. Please make sure the parameters are set correctly.

 WARNING: The benchmarking script will save json results by itself, so please do not configure `--output-json` parameter in the json file.

+### Throughput test

-#### 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 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
+### 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:

-```
+```json
 [
    {
        "test_name": "serving_llama8B_tp1_sharegpt",
@@ -109,6 +101,7 @@ We test the throughput by using `benchmark_serving.py` with request rate = inf t
 ```

 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`.
@@ -118,36 +111,33 @@ The number of this test is less stable compared to the delay and latency benchma

 WARNING: The benchmarking script will save json results by itself, so please do not configure `--save-results` or other results-saving-related parameters in `serving-tests.json`.

-#### Visualizing the results
+### Visualizing the results
+
 The `convert-results-json-to-markdown.py` helps you put the benchmarking results inside a markdown table, by formatting [descriptions.md](tests/descriptions.md) with real benchmarking results.
 You can find the result presented as a table inside the `buildkite/performance-benchmark` job page.
 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.

-
-
 ## 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.

+### Workflow

-#### Workflow
-
- The [nightly-pipeline.yaml](nightly-pipeline.yaml) specifies the docker containers for different LLM serving engines. 
+- The [nightly-pipeline.yaml](nightly-pipeline.yaml) specifies the docker containers for different LLM serving engines.
 - Inside each container, we run [run-nightly-suite.sh](run-nightly-suite.sh), which will probe the serving engine of the current container.
 - The `run-nightly-suite.sh` will redirect the request to `tests/run-[llm serving engine name]-nightly.sh`, which parses the workload described in [nightly-tests.json](tests/nightly-tests.json) and performs the benchmark.
 - At last, we run [scripts/plot-nightly-results.py](scripts/plot-nightly-results.py) to collect and plot the final benchmarking results, and update the results to buildkite.

-#### Nightly tests
+### Nightly tests

 In [nightly-tests.json](tests/nightly-tests.json), we include the command line arguments for benchmarking commands, together with the benchmarking test cases. The format is highly similar to performance benchmark.

-#### Docker containers
+### Docker containers

 The docker containers for benchmarking are specified in `nightly-pipeline.yaml`.

 WARNING: the docker versions are HARD-CODED and SHOULD BE ALIGNED WITH `nightly-descriptions.md`. The docker versions need to be hard-coded as there are several version-specific bug fixes inside `tests/run-[llm serving engine name]-nightly.sh`.

 WARNING: populating `trt-llm` to latest version is not easy, as it requires updating several protobuf files in [tensorrt-demo](https://github.com/neuralmagic/tensorrt-demo.git).
-
@@ -1,5 +1,6 @@
 steps:
  - label: "Wait for container to be ready"
+    key: wait-for-container-image
    agents:
      queue: A100
    plugins:
@@ -9,13 +10,18 @@ steps:
          - image: badouralix/curl-jq
            command:
            - sh .buildkite/nightly-benchmarks/scripts/wait-for-image.sh
-
-  - wait
-
+  - label: "Cleanup H100"
+    agents:
+      queue: H100
+    depends_on: ~
+    command: docker system prune -a --volumes --force
+  
  - label: "A100"
    # skip: "use this flag to conditionally skip the benchmark step, useful for PR testing"
    agents:
      queue: A100
+    depends_on: wait-for-container-image
+    if: build.branch == "main"
    plugins:
    - kubernetes:
        podSpec:
@@ -49,6 +55,8 @@ steps:
    # skip: "use this flag to conditionally skip the benchmark step, useful for PR testing"
    agents:
      queue: H200
+    depends_on: wait-for-container-image
+    if: build.branch == "main"
    plugins:
    - docker#v5.12.0:
        image: public.ecr.aws/q9t5s3a7/vllm-ci-postmerge-repo:$BUILDKITE_COMMIT
@@ -65,15 +73,16 @@ steps:
        - VLLM_USAGE_SOURCE
        - HF_TOKEN

-  - block: "Run H100 Benchmark"
-    key: block-h100
-    depends_on: ~
+  #- block: "Run H100 Benchmark"
+    #key: block-h100
+    #depends_on: ~

  - label: "H100"
    # skip: "use this flag to conditionally skip the benchmark step, useful for PR testing"
    agents:
      queue: H100
-    depends_on: block-h100
+    depends_on: wait-for-container-image
+    if: build.branch == "main"
    plugins:
    - docker#v5.12.0:
        image: public.ecr.aws/q9t5s3a7/vllm-ci-postmerge-repo:$BUILDKITE_COMMIT
@@ -89,3 +98,87 @@ steps:
        environment:
        - VLLM_USAGE_SOURCE
        - HF_TOKEN
+
+  # Premerge benchmark
+  - label: "A100"
+    # skip: "use this flag to conditionally skip the benchmark step, useful for PR testing"
+    agents:
+      queue: A100
+    depends_on: wait-for-container-image
+    if: build.branch != "main"
+    plugins:
+    - kubernetes:
+        podSpec:
+          priorityClassName: perf-benchmark
+          containers:
+          - image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
+            command:
+            - bash .buildkite/nightly-benchmarks/scripts/run-performance-benchmarks.sh
+            resources:
+              limits:
+                nvidia.com/gpu: 8
+            volumeMounts:
+            - name: devshm
+              mountPath: /dev/shm
+            env:
+            - name: VLLM_USAGE_SOURCE
+              value: ci-test
+            - name: HF_TOKEN
+              valueFrom:
+                secretKeyRef:
+                  name: hf-token-secret
+                  key: token
+          nodeSelector:
+            nvidia.com/gpu.product: NVIDIA-A100-SXM4-80GB
+          volumes:
+          - name: devshm
+            emptyDir:
+              medium: Memory
+
+  - label: "H200"
+    # skip: "use this flag to conditionally skip the benchmark step, useful for PR testing"
+    agents:
+      queue: H200
+    depends_on: wait-for-container-image
+    if: build.branch != "main"
+    plugins:
+    - docker#v5.12.0:
+        image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
+        command:
+        - bash
+        - .buildkite/nightly-benchmarks/scripts/run-performance-benchmarks.sh
+        mount-buildkite-agent: true
+        propagate-environment: true
+        ipc: host
+        gpus: 4,5,6,7
+        volumes:
+          - /data/benchmark-hf-cache:/root/.cache/huggingface
+        environment:
+        - VLLM_USAGE_SOURCE
+        - HF_TOKEN
+
+  #- block: "Run H100 Benchmark"
+    #key: block-h100
+    #depends_on: ~
+
+  - label: "H100"
+    # skip: "use this flag to conditionally skip the benchmark step, useful for PR testing"
+    agents:
+      queue: H100
+    depends_on: wait-for-container-image
+    if: build.branch != "main"
+    plugins:
+    - docker#v5.12.0:
+        image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
+        command:
+        - bash
+        - .buildkite/nightly-benchmarks/scripts/run-performance-benchmarks.sh
+        mount-buildkite-agent: true
+        propagate-environment: true
+        ipc: host
+        gpus: all # see CUDA_VISIBLE_DEVICES for actual GPUs used
+        volumes:
+          - /data/benchmark-hf-cache:/root/.cache/huggingface
+        environment:
+        - VLLM_USAGE_SOURCE
+        - HF_TOKEN
@@ -9,20 +9,19 @@ This file contains the downloading link for benchmarking results.

 Please download the visualization scripts in the post

-
 ## Results reproduction

 - Find the docker we use in `benchmarking pipeline`
 - Deploy the docker, and inside the docker:
-  - Download `nightly-benchmarks.zip`. 
-  - In the same folder, run the following code
-```
-export HF_TOKEN=<your HF token>
-apt update
-apt install -y git
-unzip nightly-benchmarks.zip
-VLLM_SOURCE_CODE_LOC=./ bash .buildkite/nightly-benchmarks/scripts/run-nightly-benchmarks.sh
-```
+  - Download `nightly-benchmarks.zip`.
+  - In the same folder, run the following code:
+
+  ```console
+  export HF_TOKEN=<your HF token>
+  apt update
+  apt install -y git
+  unzip nightly-benchmarks.zip
+  VLLM_SOURCE_CODE_LOC=./ bash .buildkite/nightly-benchmarks/scripts/run-nightly-benchmarks.sh
+  ```

 And the results will be inside `./benchmarks/results`.
-
@@ -2,6 +2,7 @@
 # Nightly benchmark

 This benchmark aims to:
+
 - Provide performance clarity: Provide clarity on which one (vllm, tensorrt-llm, lmdeploy and SGLang) leads in performance in what workload.
 - Be reproducible: one can run the exact same set of benchmarking commands inside the exact same docker by following reproducing instructions.

@@ -9,7 +10,6 @@ Latest results: [results link](https://blog.vllm.ai/2024/09/05/perf-update.html)

 Latest reproduction guilde: [github issue link](https://github.com/vllm-project/vllm/issues/8176)

-
 ## Setup

 - Docker images:
@@ -33,7 +33,7 @@ Latest reproduction guilde: [github issue link](https://github.com/vllm-project/
    - Queries are randomly sampled, and arrival patterns are determined via Poisson process, but all with fixed random seed.
  - Evaluation metrics: Throughput (higher the better), TTFT (time to the first token, lower the better), ITL (inter-token latency, lower the better).

-# Known issues
+## Known issues

 - TRT-LLM crashes with Llama 3.1 8B [issue](https://github.com/NVIDIA/TensorRT-LLM/issues/2105).
- TGI does not support `ignore-eos` flag.
+- TGI does not support `ignore-eos` flag.
@@ -7,10 +7,8 @@
 - Models: llama-3.1 8B, llama-3 70B, mixtral 8x7B.
 - Evaluation metrics: end-to-end latency (mean, median, p99).

-
 {latency_tests_markdown_table}

-
 ## Throughput tests

 - Input length: randomly sample 200 prompts from ShareGPT dataset (with fixed random seed).
@@ -19,10 +17,8 @@
 - Models: llama-3.1 8B, llama-3 70B, mixtral 8x7B.
 - Evaluation metrics: throughput.

-
 {throughput_tests_markdown_table}

-
 ## Serving tests

 - Input length: randomly sample 200 prompts from ShareGPT dataset (with fixed random seed).
@@ -33,13 +29,11 @@
 - We also added a speculative decoding test for llama-3 70B, under QPS 2
 - Evaluation metrics: throughput, TTFT (time to the first token, with mean, median and p99), ITL (inter-token latency, with mean, median and p99).

-
 {serving_tests_markdown_table}

-
 ## json version of the benchmarking tables

-This section contains the data of the markdown tables above in JSON format. 
+This section contains the data of the markdown tables above in JSON format.
 You can load the benchmarking tables into pandas dataframes as follows:

 ```python
@@ -54,9 +48,9 @@ serving_results = pd.DataFrame.from_dict(benchmarking_results["serving"])
 ```

 The json string for all benchmarking tables:
+
 ```json
 {benchmarking_results_in_json_string}
 ```

 You can also check the raw experiment data in the Artifact tab of the Buildkite page.
-
@@ -1,3 +1,5 @@
+# SPDX-License-Identifier: Apache-2.0
+
 import json
 import os
 from pathlib import Path
@@ -82,8 +84,13 @@ if __name__ == "__main__":
            # this result is generated via `benchmark_serving.py`

            # attach the benchmarking command to raw_result
-            with open(test_file.with_suffix(".commands")) as f:
-                command = json.loads(f.read())
+            try:
+                with open(test_file.with_suffix(".commands")) as f:
+                    command = json.loads(f.read())
+            except OSError as e:
+                print(e)
+                continue
+
            raw_result.update(command)

            # update the test name of this result
@@ -97,8 +104,13 @@ if __name__ == "__main__":
            # this result is generated via `benchmark_latency.py`

            # attach the benchmarking command to raw_result
-            with open(test_file.with_suffix(".commands")) as f:
-                command = json.loads(f.read())
+            try:
+                with open(test_file.with_suffix(".commands")) as f:
+                    command = json.loads(f.read())
+            except OSError as e:
+                print(e)
+                continue
+
            raw_result.update(command)

            # update the test name of this result
@@ -119,8 +131,13 @@ if __name__ == "__main__":
            # this result is generated via `benchmark_throughput.py`

            # attach the benchmarking command to raw_result
-            with open(test_file.with_suffix(".commands")) as f:
-                command = json.loads(f.read())
+            try:
+                with open(test_file.with_suffix(".commands")) as f:
+                    command = json.loads(f.read())
+            except OSError as e:
+                print(e)
+                continue
+
            raw_result.update(command)

            # update the test name of this result
@@ -1,3 +1,5 @@
+# SPDX-License-Identifier: Apache-2.0
+
 import argparse

 from transformers import AutoTokenizer
@@ -1,3 +1,5 @@
+# SPDX-License-Identifier: Apache-2.0
+
 import argparse
 import json
 from pathlib import Path
@@ -1,3 +1,5 @@
+# SPDX-License-Identifier: Apache-2.0
+
 from lmdeploy.serve.openai.api_client import APIClient

 api_client = APIClient("http://localhost:8000")
@@ -43,7 +43,7 @@ main() {
    


-    # The figures should be genereated by a separate process outside the CI/CD pipeline
+    # The figures should be generated by a separate process outside the CI/CD pipeline

    # # generate figures
    # python3 -m pip install tabulate pandas matplotlib
@@ -301,6 +301,104 @@ run_serving_tests() {
  kill_gpu_processes
 }

+run_genai_perf_tests() {
+  # run genai-perf tests 
+
+  # $1: a json file specifying genai-perf test cases
+  local genai_perf_test_file
+  genai_perf_test_file=$1
+
+  # Iterate over genai-perf tests
+  jq -c '.[]' "$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')    
+    
+    # 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}_${test_name}
+
+    # get common parameters
+    common_params=$(echo "$params" | jq -r '.common_parameters')
+    model=$(echo "$common_params" | jq -r '.model')
+    tp=$(echo "$common_params" | jq -r '.tp')
+    dataset_name=$(echo "$common_params" | jq -r '.dataset_name')
+    dataset_path=$(echo "$common_params" | jq -r '.dataset_path')
+    port=$(echo "$common_params" | jq -r '.port')
+    num_prompts=$(echo "$common_params" | jq -r '.num_prompts')
+    reuse_server=$(echo "$common_params" | jq -r '.reuse_server')
+
+    # get client and server arguments
+    server_params=$(echo "$params" | jq -r ".${CURRENT_LLM_SERVING_ENGINE}_server_parameters")
+    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
+    if [[ $gpu_count -lt $tp ]]; then
+      echo "Required num-shard $tp but only $gpu_count GPU found. Skip testcase $test_name."
+      continue
+    fi
+
+    if [[ $reuse_server == "true" ]]; then
+      echo "Reuse previous server for test case $test_name"
+    else
+      kill_gpu_processes
+      bash "$VLLM_SOURCE_CODE_LOC/.buildkite/nightly-benchmarks/scripts/launch-server.sh" \
+        "$server_params" "$common_params"
+    fi
+
+    if wait_for_server; then
+      echo ""
+      echo "$CURRENT_LLM_SERVING_ENGINE server is up and running."
+    else
+      echo ""
+      echo "$CURRENT_LLM_SERVING_ENGINE failed to start within the timeout period."
+      break
+    fi
+
+    # iterate over different QPS
+    for qps in $qps_list; do
+      # remove the surrounding single quote from qps
+      if [[ "$qps" == *"inf"* ]]; then
+        echo "qps was $qps"
+        qps=$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
+      #TODO: add output dir.
+      client_command="genai-perf profile \
+        -m $model \
+        --service-kind openai \
+        --backend vllm \
+        --endpoint-type chat \
+        --streaming \
+        --url localhost:$port \
+        --request-rate $qps \
+        --num-prompts $num_prompts \
+      "
+
+    echo "Client command: $client_command"
+
+    eval "$client_command"
+
+    #TODO: process/record outputs
+    done
+  done
+
+  kill_gpu_processes
+
+}

 prepare_dataset() {

@@ -328,12 +426,17 @@ main() {

  pip install -U transformers

+  pip install -r requirements-dev.txt
+  which genai-perf
+
  # check storage
  df -h

  ensure_installed wget
  ensure_installed curl
  ensure_installed jq
+  # genai-perf dependency
+  ensure_installed libb64-0d

  prepare_dataset

@@ -345,6 +448,10 @@ main() {
  # run the test
  run_serving_tests "$BENCHMARK_ROOT/tests/nightly-tests.json"

+  # run genai-perf tests
+  run_genai_perf_tests "$BENCHMARK_ROOT/tests/genai-perf-tests.json"
+  mv artifacts/ $RESULTS_FOLDER/
+
  # upload benchmark results to buildkite
  python3 -m pip install tabulate pandas
  python3 "$BENCHMARK_ROOT/scripts/summary-nightly-results.py"
@@ -309,11 +309,14 @@ run_serving_tests() {

      new_test_name=$test_name"_qps_"$qps

+      # pass the tensor parallel size to the client so that it can be displayed
+      # on the benchmark dashboard
      client_command="python3 benchmark_serving.py \
        --save-result \
        --result-dir $RESULTS_FOLDER \
        --result-filename ${new_test_name}.json \
        --request-rate $qps \
+        --metadata "tensor_parallel_size=$tp" \
        $client_args"

      echo "Running test case $test_name with qps $qps"
@@ -345,6 +348,11 @@ main() {
  check_gpus
  check_hf_token

+  # Set to v1 to run v1 benchmark
+  if [[ "${ENGINE_VERSION:-v0}" == "v1" ]]; then
+    export VLLM_USE_V1=1
+  fi
+
  # dependencies
  (which wget && which curl) || (apt-get update && apt-get install -y wget curl)
  (which jq) || (apt-get update && apt-get -y install jq)
@@ -1,3 +1,5 @@
+# SPDX-License-Identifier: Apache-2.0
+
 import datetime
 import json
 import os
@@ -1,6 +1,10 @@
 #!/bin/sh
 TOKEN=$(curl -s -L "https://public.ecr.aws/token?service=public.ecr.aws&scope=repository:q9t5s3a7/vllm-ci-postmerge-repo:pull" | jq -r .token)
-URL="https://public.ecr.aws/v2/q9t5s3a7/vllm-ci-postmerge-repo/manifests/$BUILDKITE_COMMIT"
+if [[ "$BUILDKITE_BRANCH" == "main" ]]; then
+    URL="https://public.ecr.aws/v2/q9t5s3a7/vllm-ci-postmerge-repo/manifests/$BUILDKITE_COMMIT"
+else
+    URL="https://public.ecr.aws/v2/q9t5s3a7/vllm-ci-test-repo/manifests/$BUILDKITE_COMMIT"
+fi

 TIMEOUT_SECONDS=10

@@ -0,0 +1,23 @@
+[
+    {
+        "test_name": "llama8B_tp1_genai_perf",
+        "qps_list": [4,8,16,32],
+        "common_parameters": {
+            "model": "meta-llama/Meta-Llama-3-8B-Instruct",
+            "tp": 1,
+            "port": 8000,
+            "num_prompts": 500,
+            "reuse_server": false
+        },
+        "vllm_server_parameters": {
+            "disable_log_stats": "",
+            "disable_log_requests": "",
+            "gpu_memory_utilization": 0.9,
+            "num_scheduler_steps": 10,
+            "max_num_seqs": 512,
+            "dtype": "bfloat16"
+        },
+        "genai_perf_input_parameters": {
+        }
+    }
+]
@@ -29,4 +29,4 @@
            "num-iters": 15
        }
    }
-]
+]
@@ -66,8 +66,7 @@
            "swap_space": 16, 
            "speculative_model": "turboderp/Qwama-0.5B-Instruct",
            "num_speculative_tokens": 4,
-            "speculative_draft_tensor_parallel_size": 1,
-            "use_v2_block_manager": ""
+            "speculative_draft_tensor_parallel_size": 1
        },
        "client_parameters": {
            "model": "meta-llama/Meta-Llama-3.1-70B-Instruct",
@@ -32,4 +32,4 @@
            "backend": "vllm"
        }
    }
-]
+]
@@ -1,4 +1,15 @@
 steps:
+  - label: "Build wheel - CUDA 12.4"
+    agents:
+      queue: cpu_queue_postmerge
+    commands:
+      - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=12.4.0 --tag vllm-ci:build-image --target build --progress plain ."
+      - "mkdir artifacts"
+      - "docker run --rm -v $(pwd)/artifacts:/artifacts_host vllm-ci:build-image bash -c 'cp -r dist /artifacts_host && chmod -R a+rw /artifacts_host'"
+      - "bash .buildkite/upload-wheels.sh"
+    env:
+      DOCKER_BUILDKIT: "1"
+
  - label: "Build wheel - CUDA 12.1"
    agents:
      queue: cpu_queue_postmerge
@@ -37,7 +48,7 @@ steps:
      queue: cpu_queue_postmerge
    commands:
      - "aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin public.ecr.aws/q9t5s3a7"
-      - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=12.1.0 --tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT --target vllm-openai --progress plain ."
+      - "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=12.4.0 --tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT --target vllm-openai --progress plain ."
      - "docker push public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT"

  - label: "Build and publish TPU release image"
@@ -55,3 +66,23 @@ steps:
          password-env: DOCKERHUB_TOKEN
    env:
      DOCKER_BUILDKIT: "1"
+
+  - input: "Provide Release version here"
+    fields:
+      - text: "What is the release version?"
+        key: "release-version"
+
+  - block: "Build CPU release image"
+    key: block-cpu-release-image-build
+    depends_on: ~
+
+  - label: "Build and publish CPU release image"
+    depends_on: block-cpu-release-image-build
+    agents:
+      queue: cpu_queue_postmerge
+    commands:
+      - "aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin public.ecr.aws/q9t5s3a7"
+      - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg GIT_REPO_CHECK=1 --tag public.ecr.aws/q9t5s3a7/vllm-cpu-release-repo:$(buildkite-agent meta-data get release-version) --progress plain -f Dockerfile.cpu ."
+      - "docker push public.ecr.aws/q9t5s3a7/vllm-cpu-release-repo:$(buildkite-agent meta-data get release-version)"
+    env:
+      DOCKER_BUILDKIT: "1"
@@ -77,7 +77,6 @@ echo "Commands:$commands"
 #ignore certain kernels tests
 if [[ $commands == *" kernels "* ]]; then
  commands="${commands} \
-  --ignore=kernels/test_attention.py \
  --ignore=kernels/test_attention_selector.py \
  --ignore=kernels/test_blocksparse_attention.py \
  --ignore=kernels/test_causal_conv1d.py \
@@ -92,7 +91,14 @@ if [[ $commands == *" kernels "* ]]; then
  --ignore=kernels/test_moe.py \
  --ignore=kernels/test_prefix_prefill.py \
  --ignore=kernels/test_rand.py \
-  --ignore=kernels/test_sampler.py"
+  --ignore=kernels/test_sampler.py \
+  --ignore=kernels/test_cascade_flash_attn.py \
+  --ignore=kernels/test_mamba_mixer2.py \
+  --ignore=kernels/test_aqlm.py \
+  --ignore=kernels/test_machete_mm.py \
+  --ignore=kernels/test_mha_attn.py \
+  --ignore=kernels/test_block_fp8.py \
+  --ignore=kernels/test_permute_cols.py"
 fi

 #ignore certain Entrypoints tests
@@ -121,6 +127,8 @@ if [[ $commands == *"--shard-id="* ]]; then
        --rm \
        -e HIP_VISIBLE_DEVICES="${GPU}" \
        -e HF_TOKEN \
+        -e AWS_ACCESS_KEY_ID \
+        -e AWS_SECRET_ACCESS_KEY \
        -v "${HF_CACHE}:${HF_MOUNT}" \
        -e "HF_HOME=${HF_MOUNT}" \
        --name "${container_name}_${GPU}" \
@@ -148,6 +156,8 @@ else
          --rm \
          -e HIP_VISIBLE_DEVICES=0 \
          -e HF_TOKEN \
+          -e AWS_ACCESS_KEY_ID \
+          -e AWS_SECRET_ACCESS_KEY \
          -v "${HF_CACHE}:${HF_MOUNT}" \
          -e "HF_HOME=${HF_MOUNT}" \
          --name "${container_name}" \
@@ -9,36 +9,33 @@ CORE_RANGE=${CORE_RANGE:-48-95}
 NUMA_NODE=${NUMA_NODE:-1}

 # Try building the docker image
-numactl -C "$CORE_RANGE" -N "$NUMA_NODE" docker build -t cpu-test -f Dockerfile.cpu .
-numactl -C "$CORE_RANGE" -N "$NUMA_NODE" docker build --build-arg VLLM_CPU_DISABLE_AVX512="true" -t cpu-test-avx2 -f Dockerfile.cpu .
+numactl -C "$CORE_RANGE" -N "$NUMA_NODE" docker build -t cpu-test-"$BUILDKITE_BUILD_NUMBER" -f Dockerfile.cpu .
+numactl -C "$CORE_RANGE" -N "$NUMA_NODE" docker build --build-arg VLLM_CPU_DISABLE_AVX512="true" -t cpu-test-"$BUILDKITE_BUILD_NUMBER"-avx2 -f Dockerfile.cpu .

 # Setup cleanup
-remove_docker_container() { docker rm -f cpu-test-"$NUMA_NODE" cpu-test-avx2-"$NUMA_NODE" || true; }
+remove_docker_container() { set -e; docker rm -f cpu-test-"$BUILDKITE_BUILD_NUMBER"-"$NUMA_NODE" cpu-test-"$BUILDKITE_BUILD_NUMBER"-avx2-"$NUMA_NODE" || true; }
 trap remove_docker_container EXIT
 remove_docker_container

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

 function cpu_tests() {
  set -e
  export NUMA_NODE=$2

  # offline inference
-  docker exec cpu-test-avx2-"$NUMA_NODE" bash -c "
+  docker exec cpu-test-"$BUILDKITE_BUILD_NUMBER"-avx2-"$NUMA_NODE" bash -c "
    set -e
-    python3 examples/offline_inference.py"
+    python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m"

  # Run basic model test
-  docker exec cpu-test-"$NUMA_NODE" bash -c "
+  docker exec cpu-test-"$BUILDKITE_BUILD_NUMBER"-"$NUMA_NODE" bash -c "
    set -e
-    pip install pytest pytest-asyncio \
-      decord einops librosa peft Pillow sentence-transformers soundfile \
-      transformers_stream_generator matplotlib datamodel_code_generator
-    pip install torchvision --index-url https://download.pytorch.org/whl/cpu
+    pip install -r vllm/requirements-test.txt
    pytest -v -s tests/models/decoder_only/language -m cpu_model
    pytest -v -s tests/models/embedding/language -m cpu_model
    pytest -v -s tests/models/encoder_decoder/language -m cpu_model
@@ -46,26 +43,26 @@ function cpu_tests() {
    pytest -v -s tests/models/decoder_only/vision_language -m cpu_model"

  # Run compressed-tensor test
-  docker exec cpu-test-"$NUMA_NODE" bash -c "
+  docker exec cpu-test-"$BUILDKITE_BUILD_NUMBER"-"$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"

  # Run AWQ test
-  docker exec cpu-test-"$NUMA_NODE" bash -c "
+  docker exec cpu-test-"$BUILDKITE_BUILD_NUMBER"-"$NUMA_NODE" bash -c "
    set -e
    pytest -s -v \
    tests/quantization/test_ipex_quant.py"

  # Run chunked-prefill and prefix-cache test
-  docker exec cpu-test-"$NUMA_NODE" bash -c "
+  docker exec cpu-test-"$BUILDKITE_BUILD_NUMBER"-"$NUMA_NODE" bash -c "
    set -e
    pytest -s -v -k cpu_model \
    tests/basic_correctness/test_chunked_prefill.py"  

-  # online inference
-  docker exec cpu-test-"$NUMA_NODE" bash -c "
+  # online serving
+  docker exec cpu-test-"$BUILDKITE_BUILD_NUMBER"-"$NUMA_NODE" bash -c "
    set -e
    export VLLM_CPU_KVCACHE_SPACE=10 
    export VLLM_CPU_OMP_THREADS_BIND=$1
@@ -78,8 +75,14 @@ function cpu_tests() {
      --num-prompts 20 \
      --endpoint /v1/completions \
      --tokenizer facebook/opt-125m"
+
+  # Run multi-lora tests
+  docker exec cpu-test-"$BUILDKITE_BUILD_NUMBER"-"$NUMA_NODE" bash -c "
+    set -e
+    pytest -s -v \
+    tests/lora/test_qwen2vl.py"
 }

-# All of CPU tests are expected to be finished less than 25 mins.
+# All of CPU tests are expected to be finished less than 40 mins.
 export -f cpu_tests
-timeout 30m bash -c "cpu_tests $CORE_RANGE $NUMA_NODE"
+timeout 40m bash -c "cpu_tests $CORE_RANGE $NUMA_NODE"
@@ -4,6 +4,9 @@
 # It serves a sanity check for compilation and basic model usage.
 set -ex

+# Skip the new torch installation during build since we are using the specified version for arm64 in the Dockerfile
+python3 use_existing_torch.py
+
 # Try building the docker image
 DOCKER_BUILDKIT=1 docker build . \
  --target vllm-openai \
@@ -20,6 +23,6 @@ trap remove_docker_container EXIT
 remove_docker_container

 # Run the image and test offline inference
-docker run --name gh200-test --gpus=all --entrypoint="" gh200-test bash -c '
-    python3 examples/offline_inference.py
+docker run -e HF_TOKEN -v /root/.cache/huggingface:/root/.cache/huggingface --name gh200-test --gpus=all --entrypoint="" gh200-test bash -c '
+    python3 examples/offline_inference/basic/generate.py --model meta-llama/Llama-3.2-1B
 '
@@ -8,9 +8,17 @@ set -ex
 docker build -t hpu-test-env -f Dockerfile.hpu .

 # Setup cleanup
+# certain versions of HPU software stack have a bug that can
+# override the exit code of the script, so we need to use
+# separate remove_docker_container and remove_docker_container_and_exit
+# functions, while other platforms only need one remove_docker_container
+# function.
+EXITCODE=1
 remove_docker_container() { docker rm -f hpu-test || true; }
-trap remove_docker_container EXIT
+remove_docker_container_and_exit() { remove_docker_container; exit $EXITCODE; }
+trap remove_docker_container_and_exit EXIT
 remove_docker_container

 # 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.py
+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
+EXITCODE=$?
@@ -3,6 +3,18 @@
 # This script build the Neuron docker image and run the API server inside the container.
 # It serves a sanity check for compilation and basic model usage.
 set -e
+set -v
+
+image_name="neuron/vllm-ci"
+container_name="neuron_$(tr -dc A-Za-z0-9 < /dev/urandom | head -c 10; echo)"
+
+HF_CACHE="$(realpath ~)/huggingface"
+mkdir -p "${HF_CACHE}"
+HF_MOUNT="/root/.cache/huggingface"
+
+NEURON_COMPILE_CACHE_URL="$(realpath ~)/neuron_compile_cache"
+mkdir -p "${NEURON_COMPILE_CACHE_URL}"
+NEURON_COMPILE_CACHE_MOUNT="/root/.cache/neuron_compile_cache"

 # Try building the docker image
 aws ecr get-login-password --region us-west-2 | docker login --username AWS --password-stdin 763104351884.dkr.ecr.us-west-2.amazonaws.com
@@ -13,41 +25,30 @@ if [ -f /tmp/neuron-docker-build-timestamp ]; then
    last_build=$(cat /tmp/neuron-docker-build-timestamp)
    current_time=$(date +%s)
    if [ $((current_time - last_build)) -gt 86400 ]; then
-        docker system prune -f
+        # Remove dangling images (those that are not tagged and not used by any container)
+        docker image prune -f
+        # Remove unused volumes / force the system prune for old images as well.
+        docker volume prune -f && docker system prune -f
        echo "$current_time" > /tmp/neuron-docker-build-timestamp
    fi
 else
    date "+%s" > /tmp/neuron-docker-build-timestamp
 fi

-docker build -t neuron -f Dockerfile.neuron .
+docker build -t "${image_name}" -f Dockerfile.neuron .

 # Setup cleanup
-remove_docker_container() { docker rm -f neuron || true; }
+remove_docker_container() {
+    docker image rm -f "${image_name}" || true;
+}
 trap remove_docker_container EXIT
-remove_docker_container

 # Run the image
-docker run --device=/dev/neuron0 --device=/dev/neuron1 --network host --name neuron neuron python3 -m vllm.entrypoints.api_server \
-       --model TinyLlama/TinyLlama-1.1B-Chat-v1.0 --max-num-seqs 8 --max-model-len 128 --block-size 128 --device neuron --tensor-parallel-size 2 &
-
-# Wait for the server to start
-wait_for_server_to_start() {
-    timeout=300
-    counter=0
-
-    while [ "$(curl -s -o /dev/null -w '%{http_code}' localhost:8000/health)" != "200" ]; do
-        sleep 1
-        counter=$((counter + 1))
-        if [ $counter -ge $timeout ]; then
-            echo "Timeout after $timeout seconds"
-            break
-        fi
-    done
-}
-wait_for_server_to_start
-
-# Test a simple prompt
-curl -X POST -H "Content-Type: application/json" \
-    localhost:8000/generate \
-    -d '{"prompt": "San Francisco is a"}'
+docker run --rm -it --device=/dev/neuron0 --device=/dev/neuron1 --network host \
+       -v "${HF_CACHE}:${HF_MOUNT}" \
+       -e "HF_HOME=${HF_MOUNT}" \
+       -v "${NEURON_COMPILE_CACHE_URL}:${NEURON_COMPILE_CACHE_MOUNT}" \
+       -e "NEURON_COMPILE_CACHE_URL=${NEURON_COMPILE_CACHE_MOUNT}" \
+       --name "${container_name}" \
+       ${image_name} \
+       /bin/bash -c "python3 /workspace/vllm/examples/offline_inference/neuron.py && python3 -m pytest /workspace/vllm/tests/neuron/ -v --capture=tee-sys"
@@ -13,4 +13,4 @@ trap remove_docker_container EXIT
 remove_docker_container

 # Run the image and launch offline inference
-docker run --network host --env VLLM_OPENVINO_KVCACHE_SPACE=1 --name openvino-test openvino-test python3 /workspace/examples/offline_inference.py
+docker run --network host --env VLLM_OPENVINO_KVCACHE_SPACE=1 --name openvino-test openvino-test python3 /workspace/examples/offline_inference/basic/generate.py --model facebook/opt-125m
@@ -14,4 +14,13 @@ remove_docker_container
 # For HF_TOKEN.
 source /etc/environment
 # Run a simple end-to-end example.
-docker run --privileged --net host --shm-size=16G -it -e "HF_TOKEN=$HF_TOKEN" --name tpu-test vllm-tpu /bin/bash -c "python3 -m pip install git+https://github.com/thuml/depyf.git && python3 -m pip install pytest && python3 -m pip install lm_eval[api]==0.4.4 && pytest -v -s /workspace/vllm/tests/entrypoints/openai/test_accuracy.py && pytest -v -s /workspace/vllm/tests/tpu/test_custom_dispatcher.py && python3 /workspace/vllm/tests/tpu/test_compilation.py && python3 /workspace/vllm/examples/offline_inference_tpu.py"
+docker run --privileged --net host --shm-size=16G -it \
+    -e "HF_TOKEN=$HF_TOKEN" --name tpu-test \
+    vllm-tpu /bin/bash -c "python3 -m pip install git+https://github.com/thuml/depyf.git \
+    && python3 -m pip install pytest \
+    && python3 -m pip install lm_eval[api]==0.4.4 \
+    && pytest -v -s /workspace/vllm/tests/entrypoints/openai/test_accuracy.py \
+    && pytest -v -s /workspace/vllm/tests/tpu/test_custom_dispatcher.py \
+    && python3 /workspace/vllm/tests/tpu/test_compilation.py \
+    && python3 /workspace/vllm/tests/tpu/test_quantization_accuracy.py \
+    && python3 /workspace/vllm/examples/offline_inference/tpu.py"
@@ -14,6 +14,6 @@ remove_docker_container

 # Run the image and test offline inference/tensor parallel
 docker run --name xpu-test --device /dev/dri -v /dev/dri/by-path:/dev/dri/by-path --entrypoint="" xpu-test sh -c '
-    python3 examples/offline_inference.py
-    python3 examples/offline_inference_cli.py -tp 2
+    python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m
+    python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m -tp 2
 '
@@ -2,7 +2,7 @@
 # adding a new command to an existing step. See different options here for examples.

 # This script will be feed into Jinja template in `test-template-aws.j2` at
-# https://github.com/vllm-project/buildkite-ci/blob/main/scripts/test-template-aws.j2 
+# https://github.com/vllm-project/buildkite-ci/blob/main/scripts/test-template-aws.j2
 # to generate the final pipeline yaml file.

 # Documentation
@@ -15,7 +15,7 @@
 # mirror_hardwares(list): the list of hardwares to run the test on as well. currently only supports [amd]
 # gpu(str): override the GPU selection for the test. default is on L4 GPUs. currently only supports a100
 # num_gpus(int): override the number of GPUs for the test. default to 1 GPU. currently support 2,4.
-# num_nodes(int): whether to simulate multi-node setup by launch multiple containers on one host, 
+# num_nodes(int): whether to simulate multi-node setup by launch multiple containers on one host,
 #     in this case, commands must be specified. the first command runs on first host, the second
 #     command runs on the second host.
 # working_dir(str): specify the place where command should execute, default to /vllm-workspace/tests
@@ -24,8 +24,8 @@
 # When adding a test
 # - If the test belong to an existing group, add it there
 # - If the test is short, add to any existing step
-# - If the test takes more than 10min, then it is okay to create a new step. 
-#   Note that all steps execute in parallel. 
+# - If the test takes more than 10min, then it is okay to create a new step.
+#   Note that all steps execute in parallel.

 steps:
 ##### fast check tests  #####
@@ -38,7 +38,7 @@ steps:
  - pip install -r requirements-docs.txt
  - SPHINXOPTS=\"-W\" make html
  # Check API reference (if it fails, you may have missing mock imports)
-  - grep \"sig sig-object py\" build/html/dev/sampling_params.html
+  - grep \"sig sig-object py\" build/html/api/inference_params.html

 - label: Async Engine, Inputs, Utils, Worker Test # 24min
  fast_check: true
@@ -50,9 +50,9 @@ steps:
  - tests/multimodal
  - tests/test_utils
  - tests/worker
-  - tests/standalone_tests/lazy_torch_compile.py
+  - tests/standalone_tests/lazy_imports.py
  commands:
-  - python3 standalone_tests/lazy_torch_compile.py
+  - python3 standalone_tests/lazy_imports.py
  - pytest -v -s mq_llm_engine # MQLLMEngine
  - pytest -v -s async_engine # AsyncLLMEngine
  - NUM_SCHEDULER_STEPS=4 pytest -v -s async_engine/test_async_llm_engine.py
@@ -76,7 +76,9 @@ steps:
  - tests/basic_correctness/test_basic_correctness
  - tests/basic_correctness/test_cpu_offload
  - tests/basic_correctness/test_preemption
+  - tests/basic_correctness/test_cumem.py
  commands:
+  - pytest -v -s basic_correctness/test_cumem.py
  - pytest -v -s basic_correctness/test_basic_correctness.py
  - pytest -v -s basic_correctness/test_cpu_offload.py
  - VLLM_TEST_ENABLE_ARTIFICIAL_PREEMPT=1 pytest -v -s basic_correctness/test_preemption.py
@@ -105,15 +107,17 @@ steps:
  mirror_hardwares: [amd]
  source_file_dependencies:
  - vllm/
+  - tests/entrypoints/llm
+  - tests/entrypoints/openai
+  - tests/entrypoints/test_chat_utils
+  - tests/entrypoints/offline_mode
  commands:
-  - pip install -e ./plugins/vllm_add_dummy_model
-  - pytest -v -s entrypoints/llm --ignore=entrypoints/llm/test_lazy_outlines.py --ignore=entrypoints/llm/test_generate.py --ignore=entrypoints/llm/test_generate_multiple_loras.py --ignore=entrypoints/llm/test_guided_generate.py
+  - pytest -v -s entrypoints/llm --ignore=entrypoints/llm/test_lazy_outlines.py --ignore=entrypoints/llm/test_generate.py --ignore=entrypoints/llm/test_generate_multiple_loras.py --ignore=entrypoints/llm/test_guided_generate.py --ignore=entrypoints/llm/test_collective_rpc.py
  - pytest -v -s entrypoints/llm/test_lazy_outlines.py # it needs a clean process
  - pytest -v -s entrypoints/llm/test_generate.py # it needs a clean process
  - pytest -v -s entrypoints/llm/test_generate_multiple_loras.py # it needs a clean process
  - pytest -v -s entrypoints/llm/test_guided_generate.py # it needs a clean process
-  - pytest -v -s entrypoints/openai --ignore=entrypoints/openai/test_oot_registration.py
-  - pytest -v -s entrypoints/openai/test_oot_registration.py # it needs a clean process
+  - pytest -v -s entrypoints/openai --ignore=entrypoints/openai/test_oot_registration.py --ignore=entrypoints/openai/correctness/
  - pytest -v -s entrypoints/test_chat_utils.py
  - pytest -v -s entrypoints/offline_mode # Needs to avoid interference with other tests

@@ -124,24 +128,33 @@ steps:
  source_file_dependencies:
  - vllm/distributed/
  - vllm/core/
-  - tests/distributed
+  - tests/distributed/test_utils
+  - tests/distributed/test_pynccl
  - tests/spec_decode/e2e/test_integration_dist_tp4
-  - tests/compile
+  - tests/compile/test_basic_correctness
+  - examples/offline_inference/rlhf.py
+  - examples/offline_inference/rlhf_colocate.py
+  - tests/examples/offline_inference/data_parallel.py
  commands:
+  - VLLM_USE_V1=1 python3 ../examples/offline_inference/data_parallel.py
  - 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 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
+  - python3 ../examples/offline_inference/rlhf.py
+  - RAY_DEDUP_LOGS=0 python3 ../examples/offline_inference/rlhf_colocate.py

 - label: Metrics, Tracing Test # 10min
-  num_gpus: 2 
+  num_gpus: 2
  fast_check: true
  source_file_dependencies:
  - vllm/
  - tests/metrics
  - tests/tracing
  commands:
-  - pytest -v -s metrics 
+  - pytest -v -s metrics
  - "pip install \
      'opentelemetry-sdk>=1.26.0,<1.27.0' \
      'opentelemetry-api>=1.26.0,<1.27.0' \
@@ -168,6 +181,9 @@ steps:
  - vllm/
  - tests/engine
  - tests/tokenization
+  - tests/test_sequence
+  - tests/test_config
+  - tests/test_logger
  commands:
  - pytest -v -s engine test_sequence.py test_config.py test_logger.py
  # OOM in the CI unless we run this separately
@@ -179,7 +195,19 @@ steps:
    - vllm/
    - tests/v1
  commands:
-    - VLLM_USE_V1=1 pytest -v -s v1
+    # split the test to avoid interference
+    - VLLM_USE_V1=1 pytest -v -s v1/core
+    - VLLM_USE_V1=1 pytest -v -s v1/engine
+    - VLLM_USE_V1=1 pytest -v -s v1/sample
+    - VLLM_USE_V1=1 pytest -v -s v1/worker
+    - VLLM_USE_V1=1 pytest -v -s v1/test_stats.py
+    - VLLM_USE_V1=1 pytest -v -s v1/test_utils.py
+    # TODO: accuracy does not match, whether setting
+    # VLLM_USE_FLASHINFER_SAMPLER or not on H100.
+    - VLLM_USE_V1=1 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
+    - pytest -v -s entrypoints/openai/correctness/test_lmeval.py::test_lm_eval_accuracy_v1_engine

 - label: Examples Test # 25min
  working_dir: "/vllm-workspace/examples"
@@ -189,19 +217,19 @@ steps:
  - examples/
  commands:
    - pip install tensorizer # for tensorizer test
-    - python3 offline_inference.py
-    - python3 cpu_offload.py
-    - python3 offline_inference_chat.py
-    - python3 offline_inference_with_prefix.py
-    - python3 llm_engine_example.py
-    - python3 offline_inference_vision_language.py
-    - python3 offline_inference_vision_language_multi_image.py
-    - python3 tensorize_vllm_model.py --model facebook/opt-125m serialize --serialized-directory /tmp/ --suffix v1 && python3 tensorize_vllm_model.py --model facebook/opt-125m deserialize --path-to-tensors /tmp/vllm/facebook/opt-125m/v1/model.tensors
-    - python3 offline_inference_encoder_decoder.py
-    - python3 offline_inference_classification.py
-    - python3 offline_inference_embedding.py
-    - python3 offline_inference_scoring.py
-    - python3 offline_profile.py --model facebook/opt-125m run_num_steps --num-steps 2
+    - python3 offline_inference/basic/generate.py --model facebook/opt-125m
+    - python3 offline_inference/basic/generate.py --model meta-llama/Llama-2-13b-chat-hf --cpu-offload-gb 10
+    - python3 offline_inference/basic/chat.py
+    - python3 offline_inference/prefix_caching.py
+    - python3 offline_inference/llm_engine_example.py
+    - python3 offline_inference/vision_language.py
+    - python3 offline_inference/vision_language_multi_image.py
+    - python3 other/tensorize_vllm_model.py --model facebook/opt-125m serialize --serialized-directory /tmp/ --suffix v1 && python3 other/tensorize_vllm_model.py --model facebook/opt-125m deserialize --path-to-tensors /tmp/vllm/facebook/opt-125m/v1/model.tensors
+    - python3 offline_inference/encoder_decoder.py
+    - python3 offline_inference/basic/classify.py
+    - python3 offline_inference/basic/embed.py
+    - python3 offline_inference/basic/score.py
+    - python3 offline_inference/profiling.py --model facebook/opt-125m run_num_steps --num-steps 2

 - label: Prefix Caching Test # 9min
  mirror_hardwares: [amd]
@@ -216,6 +244,7 @@ steps:
  - vllm/model_executor/layers
  - vllm/sampling_metadata.py
  - tests/samplers
+  - tests/conftest.py
  commands:
    - pytest -v -s samplers
    - VLLM_USE_FLASHINFER_SAMPLER=1 pytest -v -s samplers
@@ -224,26 +253,32 @@ steps:
  mirror_hardwares: [amd]
  source_file_dependencies:
  - vllm/model_executor/layers
+  - vllm/model_executor/guided_decoding
  - tests/test_logits_processor
-  command: pytest -v -s test_logits_processor.py
+  - tests/model_executor/test_guided_processors
+  commands:
+    - pytest -v -s test_logits_processor.py
+    - pytest -v -s model_executor/test_guided_processors.py

- label: Speculative decoding tests # 30min
+- label: Speculative decoding tests # 40min
  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
+    - 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: [amd]
  source_file_dependencies:
  - vllm/lora
  - tests/lora
-  command: pytest -v -s lora --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --ignore=lora/test_long_context.py --ignore=lora/test_chatglm3_tp.py --ignore=lora/test_llama_tp.py
+  command: pytest -v -s lora --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --ignore=lora/test_long_context.py --ignore=lora/test_chatglm3_tp.py --ignore=lora/test_llama_tp.py --ignore=lora/test_minicpmv_tp.py  --ignore=lora/test_transfomers_model.py
  parallelism: 4

- label: "PyTorch Fullgraph Smoke Test" # 9min
+- label: PyTorch Fullgraph Smoke Test # 9min
  fast_check: true
  source_file_dependencies:
  - vllm/
@@ -254,7 +289,7 @@ steps:
  - pytest -v -s compile/piecewise/test_simple.py
  - pytest -v -s compile/piecewise/test_toy_llama.py

- label: "PyTorch Fullgraph Test" # 18min
+- label: PyTorch Fullgraph Test # 18min
  source_file_dependencies:
  - vllm/
  - tests/compile
@@ -306,6 +341,14 @@ steps:
  - export VLLM_WORKER_MULTIPROC_METHOD=spawn
  - bash ./run-tests.sh -c configs/models-small.txt -t 1

+- label: OpenAI API correctness
+  source_file_dependencies:
+  - csrc/
+  - vllm/entrypoints/openai/
+  - vllm/model_executor/models/whisper.py
+  commands: # LMEval+Transcription WER check
+  - pytest -s entrypoints/openai/correctness/
+
 - label: Encoder Decoder tests # 5min
  source_file_dependencies:
  - vllm/
@@ -329,8 +372,7 @@ steps:
  - vllm/
  - tests/models
  commands:
-    - pip install -e ./plugins/vllm_add_dummy_model
-    - pytest -v -s models/test_oot_registration.py # it needs a clean process
+    - pytest -v -s models/test_transformers.py
    - pytest -v -s models/test_registry.py
    - pytest -v -s models/test_initialization.py

@@ -356,23 +398,26 @@ steps:
    - pytest -v -s models/decoder_only/language -m 'not core_model and not quant_model'
    - pytest -v -s models/embedding/language -m 'not core_model'

- label: Multi-Modal Models Test (Standard) # 28min
+- label: Multi-Modal Models Test (Standard) # 40min
  #mirror_hardwares: [amd]
  source_file_dependencies:
  - vllm/
  - tests/models/decoder_only/audio_language
  - tests/models/decoder_only/vision_language
  - tests/models/embedding/vision_language
+  - tests/models/encoder_decoder/audio_language
  - tests/models/encoder_decoder/vision_language
  commands:
    - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
+    - pytest -v -s models/multimodal
    - pytest -v -s models/decoder_only/audio_language -m 'core_model or quant_model'
    - pytest -v -s --ignore models/decoder_only/vision_language/test_phi3v.py models/decoder_only/vision_language -m 'core_model or quant_model'
    - pytest -v -s models/embedding/vision_language -m core_model
+    - pytest -v -s models/encoder_decoder/audio_language -m core_model
    - pytest -v -s models/encoder_decoder/language -m core_model
    - pytest -v -s models/encoder_decoder/vision_language -m core_model

- label: Multi-Modal Models Test (Extended) 1 # 1h16m
+- label: Multi-Modal Models Test (Extended) 1 # 48m
  optional: true
  source_file_dependencies:
  - vllm/
@@ -455,21 +500,46 @@ steps:
  - vllm/worker/worker_base.py
  - vllm/worker/worker.py
  - vllm/worker/model_runner.py
+  - entrypoints/llm/test_collective_rpc.py
  commands:
+  - pytest -v -s entrypoints/llm/test_collective_rpc.py
+  - VLLM_USE_V1=1 torchrun --nproc-per-node=2 distributed/test_torchrun_example.py
+  - torchrun --nproc-per-node=2 distributed/test_torchrun_example.py
  - pytest -v -s ./compile/test_basic_correctness.py
  - pytest -v -s ./compile/test_wrapper.py
  - VLLM_TEST_SAME_HOST=1 torchrun --nproc-per-node=4 distributed/test_same_node.py | grep 'Same node test passed'
  - TARGET_TEST_SUITE=L4 pytest basic_correctness/ -v -s -m 'distributed(num_gpus=2)'
  # Avoid importing model tests that cause CUDA reinitialization error
+  - pytest models/test_transformers.py -v -s -m 'distributed(num_gpus=2)'
  - pytest models/encoder_decoder/language/test_bart.py -v -s -m 'distributed(num_gpus=2)'
  - pytest models/encoder_decoder/vision_language/test_broadcast.py -v -s -m 'distributed(num_gpus=2)'
  - pytest models/decoder_only/vision_language/test_models.py -v -s -m 'distributed(num_gpus=2)'
-  - pytest -v -s spec_decode/e2e/test_integration_dist_tp2.py
-  - pip install -e ./plugins/vllm_add_dummy_model
-  - pytest -v -s distributed/test_distributed_oot.py
+  # this test fails consistently.
+  # TODO: investigate and fix
+  # - pytest -v -s spec_decode/e2e/test_integration_dist_tp2.py
  - CUDA_VISIBLE_DEVICES=0,1 pytest -v -s test_sharded_state_loader.py
  - CUDA_VISIBLE_DEVICES=0,1 pytest -v -s kv_transfer/disagg_test.py

+- label: Plugin Tests (2 GPUs) # 40min
+  working_dir: "/vllm-workspace/tests"
+  num_gpus: 2
+  fast_check: true
+  source_file_dependencies:
+  - vllm/plugins/
+  - tests/plugins/
+  commands:
+  # begin platform plugin tests, all the code in-between runs on dummy platform
+  - pip install -e ./plugins/vllm_add_dummy_platform
+  - pytest -v -s plugins_tests/test_platform_plugins.py
+  - pip uninstall vllm_add_dummy_platform -y
+  # end platform plugin tests
+  # other tests continue here:
+  - pytest -v -s plugins_tests/test_scheduler_plugins.py
+  - pip install -e ./plugins/vllm_add_dummy_model
+  - pytest -v -s distributed/test_distributed_oot.py
+  - pytest -v -s entrypoints/openai/test_oot_registration.py # it needs a clean process
+  - pytest -v -s models/test_oot_registration.py # it needs a clean process
+
 - label: Multi-step Tests (4 GPUs) # 36min
  working_dir: "/vllm-workspace/tests"
  num_gpus: 4
@@ -485,7 +555,9 @@ steps:
  - vllm/engine
  - tests/multi_step
  commands:
-  - pytest -v -s multi_step/test_correctness_async_llm.py
+  # this test is quite flaky
+  # TODO: investigate and fix.
+  # - pytest -v -s multi_step/test_correctness_async_llm.py
  - pytest -v -s multi_step/test_correctness_llm.py

 - label: Pipeline Parallelism Test # 45min
@@ -512,10 +584,12 @@ steps:
    - export VLLM_WORKER_MULTIPROC_METHOD=spawn
    # This test runs llama 13B, so it is required to run on 4 GPUs.
    - pytest -v -s -x lora/test_long_context.py
-    # There is some Tensor Parallelism related processing logic in LoRA that 
+    # There is some Tensor Parallelism related processing logic in LoRA that
    # requires multi-GPU testing for validation.
    - pytest -v -s -x lora/test_chatglm3_tp.py
    - pytest -v -s -x lora/test_llama_tp.py
+    - pytest -v -s -x lora/test_minicpmv_tp.py
+    - pytest -v -s -x lora/test_transfomers_model.py


 - label: Weight Loading Multiple GPU Test  # 33min
@@ -536,7 +610,7 @@ steps:
  - vllm/
  - tests/weight_loading
  commands:
-    - bash weight_loading/run_model_weight_loading_test.sh -c weight_loading/models-large.txt 
+    - bash weight_loading/run_model_weight_loading_test.sh -c weight_loading/models-large.txt


 ##### multi gpus test #####
@@ -548,7 +622,7 @@ steps:
  num_gpus: 4
  source_file_dependencies:
  - vllm/
-  commands: 
+  commands:
  # NOTE: don't test llama model here, it seems hf implementation is buggy
  # see https://github.com/vllm-project/vllm/pull/5689 for details
  - pytest -v -s distributed/test_custom_all_reduce.py
@@ -23,6 +23,8 @@ wheel="$new_wheel"
 version=$(unzip -p "$wheel" '**/METADATA' | grep '^Version: ' | cut -d' ' -f2)
 echo "Version: $version"

+normal_wheel="$wheel" # Save the original wheel filename
+
 # If the version contains "dev", rename it to v1.0.0.dev for consistency
 if [[ $version == *dev* ]]; then
    suffix="${version##*.}"
@@ -32,12 +34,44 @@ if [[ $version == *dev* ]]; then
        new_version="1.0.0.dev"
    fi
    new_wheel="${wheel/$version/$new_version}"
-    mv -- "$wheel" "$new_wheel"
+    # use cp to keep both files in the artifacts directory
+    cp -- "$wheel" "$new_wheel"
    wheel="$new_wheel"
    version="$new_version"
 fi

 # Upload the wheel to S3
+python3 .buildkite/generate_index.py --wheel "$normal_wheel"
+
+# generate index for this commit
 aws s3 cp "$wheel" "s3://vllm-wheels/$BUILDKITE_COMMIT/"
+aws s3 cp "$normal_wheel" "s3://vllm-wheels/$BUILDKITE_COMMIT/"
+
+if [[ $normal_wheel == *"cu118"* ]]; then
+    # if $normal_wheel matches cu118, do not upload the index.html
+    echo "Skipping index files for cu118 wheels"
+elif [[ $normal_wheel == *"cu121"* ]]; then
+    # if $normal_wheel matches cu121, do not upload the index.html
+    echo "Skipping index files for cu121 wheels"
+else
+    # only upload index.html for cu124 wheels (default wheels)
+    aws s3 cp index.html "s3://vllm-wheels/$BUILDKITE_COMMIT/vllm/index.html"
+    aws s3 cp "s3://vllm-wheels/nightly/index.html" "s3://vllm-wheels/$BUILDKITE_COMMIT/index.html"
+fi
+
+# generate index for nightly
 aws s3 cp "$wheel" "s3://vllm-wheels/nightly/"
+aws s3 cp "$normal_wheel" "s3://vllm-wheels/nightly/"
+
+if [[ $normal_wheel == *"cu118"* ]]; then
+    # if $normal_wheel matches cu118, do not upload the index.html
+    echo "Skipping index files for cu118 wheels"
+elif [[ $normal_wheel == *"cu121"* ]]; then
+    # if $normal_wheel matches cu121, do not upload the index.html
+    echo "Skipping index files for cu121 wheels"
+else
+    # only upload index.html for cu124 wheels (default wheels)
+    aws s3 cp index.html "s3://vllm-wheels/nightly/vllm/index.html"
+fi
+
 aws s3 cp "$wheel" "s3://vllm-wheels/$version/"
@@ -2,32 +2,35 @@
 # for more info about CODEOWNERS file

 # This lists cover the "core" components of vLLM that require careful review
-/vllm/attention/backends/abstract.py @WoosukKwon @zhuohan123 @youkaichao @alexm-neuralmagic @comaniac @njhill
-/vllm/core @zhuohan123 @youkaichao @alexm-neuralmagic @comaniac @njhill
-/vllm/engine/llm_engine.py @zhuohan123 @youkaichao @alexm-neuralmagic @comaniac @njhill
-/vllm/executor/executor_base.py @zhuohan123 @youkaichao @alexm-neuralmagic @comaniac @njhill
-/vllm/worker/worker_base.py @zhuohan123 @youkaichao @alexm-neuralmagic @comaniac @njhill
-/vllm/worker/worker.py @zhuohan123 @youkaichao @alexm-neuralmagic @comaniac @njhill
-/vllm/model_executor/layers/sampler.py @zhuohan123 @youkaichao @alexm-neuralmagic @comaniac @njhill
+/vllm/attention/backends/abstract.py @WoosukKwon @zhuohan123 @youkaichao @alexm-redhat @comaniac @njhill
+/vllm/core @zhuohan123 @youkaichao @alexm-redhat @comaniac @njhill
+/vllm/engine/llm_engine.py @zhuohan123 @youkaichao @alexm-redhat @comaniac @njhill
+/vllm/executor/executor_base.py @zhuohan123 @youkaichao @alexm-redhat @comaniac @njhill
+/vllm/worker/worker_base.py @zhuohan123 @youkaichao @alexm-redhat @comaniac @njhill
+/vllm/worker/worker.py @zhuohan123 @youkaichao @alexm-redhat @comaniac @njhill
+/vllm/model_executor/layers/sampler.py @zhuohan123 @youkaichao @alexm-redhat @comaniac @njhill
+/vllm/model_executor/layers/quantization @mgoin @robertgshaw2-redhat @tlrmchlsmth
+/vllm/model_executor/guided_decoding @mgoin
+/vllm/multimodal @DarkLight1337 @ywang96
 CMakeLists.txt @tlrmchlsmth

 # vLLM V1
-/vllm/v1 @WoosukKwon @robertgshaw2-neuralmagic @njhill @ywang96 @comaniac @alexm-neuralmagic
+/vllm/v1 @WoosukKwon @robertgshaw2-redhat @njhill @ywang96 @comaniac @alexm-redhat

 # Test ownership
-/tests/async_engine @njhill @robertgshaw2-neuralmagic @simon-mo
+/tests/async_engine @njhill @robertgshaw2-redhat @simon-mo
 /tests/test_inputs.py @DarkLight1337 @ywang96
-/tests/entrypoints @DarkLight1337 @robertgshaw2-neuralmagic @simon-mo
+/tests/entrypoints @DarkLight1337 @robertgshaw2-redhat @simon-mo
 /tests/models @DarkLight1337 @ywang96
 /tests/multimodal @DarkLight1337 @ywang96
 /tests/prefix_caching @comaniac @KuntaiDu
 /tests/spec_decode @njhill @LiuXiaoxuanPKU
 /tests/kernels @tlrmchlsmth @WoosukKwon
-/tests/quantization @mgoin @robertgshaw2-neuralmagic
+/tests/quantization @mgoin @robertgshaw2-redhat
 /.buildkite/lm-eval-harness @mgoin @simon-mo
 /tests/distributed/test_multi_node_assignment.py @youkaichao
 /tests/distributed/test_pipeline_parallel.py @youkaichao
 /tests/distributed/test_same_node.py @youkaichao
-/tests/multi_step @alexm-neuralmagic @comaniac
+/tests/multi_step @alexm-redhat @comaniac
 /tests/weight_loading @mgoin @youkaichao
 /tests/basic_correctness/test_chunked_prefill @rkooo567 @comaniac
@@ -30,15 +30,6 @@ body:
      </details>
  validations:
    required: true
- type: textarea
-  attributes:
-    label: Model Input Dumps
-    description: |
-      If you are facing crashing due to illegal memory access or other issues with model execution, vLLM may dump the problematic input of the model. In this case, you will see the message `Error in model execution (input dumped to /tmp/err_xxx.pkl)`. If you see this message, please zip the file (because GitHub doesn't support .pkl file format) and upload it here. This will help us to reproduce the issue and facilitate the debugging process.
-    placeholder: |
-      Upload the dumped input file.
-  validations:
-    required: false
 - type: textarea
  attributes:
    label: 🐛 Describe the bug
@@ -9,7 +9,7 @@ body:
    value: >
      #### Before submitting an issue, please make sure the issue hasn't been already addressed by searching through [the existing and past issues](https://github.com/vllm-project/vllm/issues?q=is%3Aissue+sort%3Acreated-desc+).

-      #### We also highly recommend you read https://docs.vllm.ai/en/latest/models/adding_model.html first to understand how to add a new model.
+      #### We also highly recommend you read https://docs.vllm.ai/en/latest/contributing/model/adding_model.html first to understand how to add a new model.
 - type: textarea
  attributes:
    label: The model to consider.
@@ -2,4 +2,5 @@ FILL IN THE PR DESCRIPTION HERE

 FIX #xxxx (*link existing issues this PR will resolve*)

-**BEFORE SUBMITTING, PLEASE READ https://docs.vllm.ai/en/latest/contributing/overview.html **
+<!--- pyml disable-next-line no-emphasis-as-heading -->
+**BEFORE SUBMITTING, PLEASE READ <https://docs.vllm.ai/en/latest/contributing/overview.html>**
@@ -23,7 +23,7 @@ updates:
      - dependency-name: "lm-format-enforcer"
      - dependency-name: "gguf"
      - dependency-name: "compressed-tensors"
-      - dependency-name: "ray[adag]"
+      - dependency-name: "ray[cgraph]" # Ray Compiled Graph
      - dependency-name: "lm-eval"
    groups:
      minor-update:
@@ -5,6 +5,7 @@ pull_request_rules:
    - or:
      - files~=^[^/]+\.md$
      - files~=^docs/
+      - files~=^examples/
  actions:
    label:
      add:
@@ -35,6 +36,43 @@ pull_request_rules:
      add:
        - frontend

+- name: label-structured-output
+  description: Automatically apply structured-output label
+  conditions:
+    - or:
+      - files~=^vllm/model_executor/guided_decoding/
+      - files=tests/model_executor/test_guided_processors.py
+      - files=tests/entrypoints/llm/test_guided_generate.py
+      - files=benchmarks/benchmark_serving_guided.py
+      - files=benchmarks/benchmark_guided.py
+  actions:
+    label:
+      add:
+        - structured-output
+
+- name: label-speculative-decoding
+  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/
+  actions:
+    label:
+      add:
+        - speculative-decoding
+
+- name: label-v1
+  description: Automatically apply v1 label
+  conditions:
+    - or:
+      - files~=^vllm/v1/
+      - files~=^tests/v1/
+  actions:
+    label:
+      add:
+        - v1
+
 - name: ping author on conflicts and add 'needs-rebase' label
  conditions:
      - conflict
@@ -1,40 +0,0 @@
-name: Lint GitHub Actions workflows
-on:
-  push:
-    branches:
-      - "main"
-    paths:
-      - '.github/workflows/*.ya?ml'
-      - '.github/workflows/actionlint.*'
-      - '.github/workflows/matchers/actionlint.json'
-  pull_request:
-    branches:
-      - "main"
-    paths:
-      - '.github/workflows/*.ya?ml'
-      - '.github/workflows/actionlint.*'
-      - '.github/workflows/matchers/actionlint.json'
-
-env:
-  LC_ALL: en_US.UTF-8
-
-defaults:
-  run:
-    shell: bash
-
-permissions:
-  contents: read
-
-jobs:
-  actionlint:
-    runs-on: ubuntu-latest
-    steps:
-      - name: "Checkout"
-        uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
-        with:
-          fetch-depth: 0
-
-      - name: "Run actionlint"
-        run: |
-          echo "::add-matcher::.github/workflows/matchers/actionlint.json"
-          tools/actionlint.sh -color
@@ -1,53 +0,0 @@
-name: clang-format
-
-on:
-  # Trigger the workflow on push or pull request,
-  # but only for the main branch
-  push:
-    branches:
-      - main
-    paths:
-      - '**/*.h'
-      - '**/*.cpp'
-      - '**/*.cu'
-      - '**/*.cuh'
-      - '.github/workflows/clang-format.yml'
-  pull_request:
-    branches:
-      - main
-    paths:
-      - '**/*.h'
-      - '**/*.cpp'
-      - '**/*.cu'
-      - '**/*.cuh'
-      - '.github/workflows/clang-format.yml'
-
-jobs:
-  clang-format:
-    runs-on: ubuntu-latest
-    strategy:
-      matrix:
-        python-version: ["3.11"]
-    steps:
-    - uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
-    - name: Set up Python ${{ matrix.python-version }}
-      uses: actions/setup-python@0b93645e9fea7318ecaed2b359559ac225c90a2b # v5.3.0
-      with:
-        python-version: ${{ matrix.python-version }}
-    - name: Install dependencies
-      run: |
-        python -m pip install --upgrade pip
-        pip install clang-format==18.1.5
-    - name: Running clang-format
-      run: |
-        EXCLUDES=(
-            'csrc/moe/topk_softmax_kernels.cu'
-            'csrc/quantization/gguf/ggml-common.h'
-            'csrc/quantization/gguf/dequantize.cuh'
-            'csrc/quantization/gguf/vecdotq.cuh'
-            'csrc/quantization/gguf/mmq.cuh'
-            'csrc/quantization/gguf/mmvq.cuh'
-        )
-        find csrc/ \( -name '*.h' -o -name '*.cpp' -o -name '*.cu' -o -name '*.cuh' \) -print \
-            | grep -vFf <(printf "%s\n" "${EXCLUDES[@]}") \
-            | xargs clang-format --dry-run --Werror
@@ -16,7 +16,7 @@ jobs:
        uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2

      - name: Set up Python
-        uses: actions/setup-python@0b93645e9fea7318ecaed2b359559ac225c90a2b # v5.3.0
+        uses: actions/setup-python@42375524e23c412d93fb67b49958b491fce71c38 # v5.4.0
        with:
          python-version: '3.12'

@@ -1,45 +0,0 @@
-name: codespell
-
-on:
-  # Trigger the workflow on push or pull request,
-  # but only for the main branch
-  push:
-    branches:
-      - main
-    paths:
-      - "**/*.py"
-      - "**/*.md"
-      - "**/*.rst"
-      - pyproject.toml
-      - requirements-lint.txt
-      - .github/workflows/codespell.yml
-  pull_request:
-    branches:
-      - main
-    paths:
-      - "**/*.py"
-      - "**/*.md"
-      - "**/*.rst"
-      - pyproject.toml
-      - requirements-lint.txt
-      - .github/workflows/codespell.yml
-
-jobs:
-  codespell:
-    runs-on: ubuntu-latest
-    strategy:
-      matrix:
-        python-version: ["3.12"]
-    steps:
-    - uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
-    - name: Set up Python ${{ matrix.python-version }}
-      uses: actions/setup-python@0b93645e9fea7318ecaed2b359559ac225c90a2b # v5.3.0
-      with:
-        python-version: ${{ matrix.python-version }}
-    - name: Install dependencies
-      run: |
-        python -m pip install --upgrade pip
-        pip install -r requirements-lint.txt
-    - name: Spelling check with codespell
-      run: |
-        codespell --toml pyproject.toml
@@ -12,22 +12,22 @@ jobs:
          fetch-depth: 0

      - name: Set up Helm
-        uses: azure/setup-helm@fe7b79cd5ee1e45176fcad797de68ecaf3ca4814 # v4.2.0
+        uses: azure/setup-helm@b9e51907a09c216f16ebe8536097933489208112 # v4.3.0
        with:
          version: v3.14.4

       #Python is required because ct lint runs Yamale and yamllint which require Python.
-      - uses: actions/setup-python@0b93645e9fea7318ecaed2b359559ac225c90a2b # v5.3.0
+      - uses: actions/setup-python@42375524e23c412d93fb67b49958b491fce71c38 # v5.4.0
        with:
          python-version: '3.13'

      - name: Set up chart-testing
-        uses: helm/chart-testing-action@e6669bcd63d7cb57cb4380c33043eebe5d111992 # v2.6.1
+        uses: helm/chart-testing-action@0d28d3144d3a25ea2cc349d6e59901c4ff469b3b # v2.7.0
        with:
          version: v3.10.1

      - name: Run chart-testing (lint)
-        run: ct lint --target-branch ${{ github.event.repository.default_branch }} --chart-dirs examples/chart-helm --charts examples/chart-helm
+        run: ct lint --target-branch ${{ github.event.repository.default_branch }} --chart-dirs examples/online_serving/chart-helm --charts examples/online_serving/chart-helm

      - name: Setup minio
        run: |
@@ -47,7 +47,7 @@ jobs:
          aws --endpoint-url http://127.0.0.1:9000/ s3 cp opt-125m/ s3://testbucket/opt-125m --recursive

      - name: Create kind cluster
-        uses: helm/kind-action@0025e74a8c7512023d06dc019c617aa3cf561fde # v1.10.0
+        uses: helm/kind-action@a1b0e391336a6ee6713a0583f8c6240d70863de3 # v1.12.0

      - name: Build the Docker image vllm cpu
        run: docker buildx build -f Dockerfile.cpu -t vllm-cpu-env .
@@ -64,7 +64,8 @@ jobs:
        run: |
          export AWS_ACCESS_KEY_ID=minioadmin
          export AWS_SECRET_ACCESS_KEY=minioadmin
-          helm install --wait --wait-for-jobs --timeout 5m0s --debug --create-namespace --namespace=ns-vllm test-vllm examples/chart-helm -f examples/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"
+          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"
    
      - name: curl test
        run: |
@@ -1,17 +0,0 @@
-{
-    "problemMatcher": [
-      {
-        "owner": "ruff",
-        "pattern": [
-          {
-            "regexp": "^(.+?):(\\d+):(\\d+): (\\w+): (.+)$",
-            "file": 1,
-            "line": 2,
-            "column": 3,
-            "code": 4,
-            "message": 5
-          }
-        ]
-      }
-    ]
-  }
@@ -1,51 +0,0 @@
-name: mypy
-
-on:
-  # Trigger the workflow on push or pull request,
-  # but only for the main branch
-  push:
-    branches:
-      - main
-    paths:
-      - '**/*.py'
-      - '.github/workflows/mypy.yaml'
-      - 'tools/mypy.sh'
-      - 'pyproject.toml'
-  pull_request:
-    branches:
-      - main
-    # This workflow is only relevant when one of the following files changes.
-    # However, we have github configured to expect and require this workflow
-    # to run and pass before github with auto-merge a pull request. Until github
-    # allows more flexible auto-merge policy, we can just run this on every PR.
-    # It doesn't take that long to run, anyway.
-    #paths:
-    #  - '**/*.py'
-    #  - '.github/workflows/mypy.yaml'
-    #  - 'tools/mypy.sh'
-    #  - 'pyproject.toml'
-
-jobs:
-  mypy:
-    runs-on: ubuntu-latest
-    strategy:
-      matrix:
-        python-version: ["3.9", "3.10", "3.11", "3.12"]
-    steps:
-    - uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
-    - name: Set up Python ${{ matrix.python-version }}
-      uses: actions/setup-python@0b93645e9fea7318ecaed2b359559ac225c90a2b # v5.3.0
-      with:
-        python-version: ${{ matrix.python-version }}
-    - name: Install dependencies
-      run: |
-        python -m pip install --upgrade pip
-        pip install mypy==1.11.1
-        pip install types-setuptools
-        pip install types-PyYAML
-        pip install types-requests
-        pip install types-setuptools
-    - name: Mypy
-      run: |
-        echo "::add-matcher::.github/workflows/matchers/mypy.json"
-        tools/mypy.sh 1 ${{ matrix.python-version }}
@@ -1,37 +0,0 @@
-name: Lint PNG exports from excalidraw
-on:
-  push:
-    branches:
-      - "main"
-    paths:
-      - '*.excalidraw.png'
-      - '.github/workflows/png-lint.yml'
-  pull_request:
-    branches:
-      - "main"
-    paths:
-      - '*.excalidraw.png'
-      - '.github/workflows/png-lint.yml'
-
-env:
-  LC_ALL: en_US.UTF-8
-
-defaults:
-  run:
-    shell: bash
-
-permissions:
-  contents: read
-
-jobs:
-  actionlint:
-    runs-on: ubuntu-latest
-    steps:
-      - name: "Checkout"
-        uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
-        with:
-          fetch-depth: 0
-
-      - name: "Run png-lint.sh to check excalidraw exported images"
-        run: |
-          tools/png-lint.sh
@@ -0,0 +1,20 @@
+name: pre-commit
+
+on:
+  pull_request:
+  push:
+    branches: [main]
+
+jobs:
+  pre-commit:
+    runs-on: ubuntu-latest
+    steps:
+    - uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
+    - uses: actions/setup-python@42375524e23c412d93fb67b49958b491fce71c38 # v5.4.0
+      with:
+        python-version: "3.12"
+    - run: echo "::add-matcher::.github/workflows/matchers/actionlint.json"
+    - run: echo "::add-matcher::.github/workflows/matchers/mypy.json"
+    - uses: pre-commit/action@2c7b3805fd2a0fd8c1884dcaebf91fc102a13ecd # v3.0.1
+      with:
+        extra_args: --all-files --hook-stage manual
@@ -39,67 +39,68 @@ jobs:
            const script = require('.github/workflows/scripts/create_release.js')
            await script(github, context, core)

-  wheel:
-    name: Build Wheel
-    runs-on: ${{ matrix.os }}
-    needs: release
+  # NOTE(simon): No longer build wheel using Github Actions. See buildkite's release workflow. 
+  # wheel:
+  #   name: Build Wheel
+  #   runs-on: ${{ matrix.os }}
+  #   needs: release

-    strategy:
-      fail-fast: false
-      matrix:
-          os: ['ubuntu-20.04']
-          python-version: ['3.9', '3.10', '3.11', '3.12']
-          pytorch-version: ['2.4.0']  # Must be the most recent version that meets requirements-cuda.txt.
-          cuda-version: ['11.8', '12.1']
+  #   strategy:
+  #     fail-fast: false
+  #     matrix:
+  #         os: ['ubuntu-20.04']
+  #         python-version: ['3.9', '3.10', '3.11', '3.12']
+  #         pytorch-version: ['2.4.0']  # Must be the most recent version that meets requirements-cuda.txt.
+  #         cuda-version: ['11.8', '12.1']

-    steps:
-      - name: Checkout
-        uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
+  #   steps:
+  #     - name: Checkout
+  #       uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2

-      - name: Setup ccache
-        uses: hendrikmuhs/ccache-action@ed74d11c0b343532753ecead8a951bb09bb34bc9 # v1.2.14
-        with:
-          create-symlink: true
-          key: ${{ github.job }}-${{ matrix.python-version }}-${{ matrix.cuda-version }}
+  #     - name: Setup ccache
+  #       uses: hendrikmuhs/ccache-action@ed74d11c0b343532753ecead8a951bb09bb34bc9 # v1.2.14
+  #       with:
+  #         create-symlink: true
+  #         key: ${{ github.job }}-${{ matrix.python-version }}-${{ matrix.cuda-version }}

-      - name: Set up Linux Env
-        if: ${{ runner.os == 'Linux' }}
-        run: |
-          bash -x .github/workflows/scripts/env.sh
+  #     - name: Set up Linux Env
+  #       if: ${{ runner.os == 'Linux' }}
+  #       run: |
+  #         bash -x .github/workflows/scripts/env.sh

-      - name: Set up Python
-        uses: actions/setup-python@0b93645e9fea7318ecaed2b359559ac225c90a2b # v5.3.0
-        with:
-            python-version: ${{ matrix.python-version }}
+  #     - name: Set up Python
+  #       uses: actions/setup-python@0b93645e9fea7318ecaed2b359559ac225c90a2b # v5.3.0
+  #       with:
+  #           python-version: ${{ matrix.python-version }}

-      - name: Install CUDA ${{ matrix.cuda-version }}
-        run: |
-          bash -x .github/workflows/scripts/cuda-install.sh ${{ matrix.cuda-version }} ${{ matrix.os }}
+  #     - name: Install CUDA ${{ matrix.cuda-version }}
+  #       run: |
+  #         bash -x .github/workflows/scripts/cuda-install.sh ${{ matrix.cuda-version }} ${{ matrix.os }}

-      - name: Install PyTorch ${{ matrix.pytorch-version }} with CUDA ${{ matrix.cuda-version }}
-        run: |
-          bash -x .github/workflows/scripts/pytorch-install.sh ${{ matrix.python-version }} ${{ matrix.pytorch-version }} ${{ matrix.cuda-version }}
+  #     - name: Install PyTorch ${{ matrix.pytorch-version }} with CUDA ${{ matrix.cuda-version }}
+  #       run: |
+  #         bash -x .github/workflows/scripts/pytorch-install.sh ${{ matrix.python-version }} ${{ matrix.pytorch-version }} ${{ matrix.cuda-version }}

-      - name: Build wheel
-        shell: bash
-        env:
-          CMAKE_BUILD_TYPE: Release # do not compile with debug symbol to reduce wheel size
-        run: |
-          bash -x .github/workflows/scripts/build.sh ${{ matrix.python-version }} ${{ matrix.cuda-version }}
-          wheel_name=$(find dist -name "*whl" -print0 | xargs -0 -n 1 basename)
-          asset_name=${wheel_name//"linux"/"manylinux1"}
-          echo "wheel_name=${wheel_name}" >> "$GITHUB_ENV"
-          echo "asset_name=${asset_name}" >> "$GITHUB_ENV"
+  #     - name: Build wheel
+  #       shell: bash
+  #       env:
+  #         CMAKE_BUILD_TYPE: Release # do not compile with debug symbol to reduce wheel size
+  #       run: |
+  #         bash -x .github/workflows/scripts/build.sh ${{ matrix.python-version }} ${{ matrix.cuda-version }}
+  #         wheel_name=$(find dist -name "*whl" -print0 | xargs -0 -n 1 basename)
+  #         asset_name=${wheel_name//"linux"/"manylinux1"}
+  #         echo "wheel_name=${wheel_name}" >> "$GITHUB_ENV"
+  #         echo "asset_name=${asset_name}" >> "$GITHUB_ENV"

-      - name: Upload Release Asset
-        uses: actions/upload-release-asset@e8f9f06c4b078e705bd2ea027f0926603fc9b4d5 # v1.0.2
-        env:
-          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-        with:
-          upload_url: ${{ needs.release.outputs.upload_url }}
-          asset_path: ./dist/${{ env.wheel_name }}
-          asset_name: ${{ env.asset_name }}
-          asset_content_type: application/*
+  #     - name: Upload Release Asset
+  #       uses: actions/upload-release-asset@e8f9f06c4b078e705bd2ea027f0926603fc9b4d5 # v1.0.2
+  #       env:
+  #         GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+  #       with:
+  #         upload_url: ${{ needs.release.outputs.upload_url }}
+  #         asset_path: ./dist/${{ env.wheel_name }}
+  #         asset_name: ${{ env.asset_name }}
+  #         asset_content_type: application/*

      # (Danielkinz): This last step will publish the .whl to pypi. Warning: untested
      # - name: Publish package
@@ -2,7 +2,6 @@ name: PR Reminder Comment Bot
 on:
  pull_request_target:
    types: [opened]
-
 jobs:
  pr_reminder:
    runs-on: ubuntu-latest
@@ -15,7 +14,12 @@ jobs:
              owner: context.repo.owner,
              repo: context.repo.repo,
              issue_number: context.issue.number,
-              body: '👋 Hi! Thank you for contributing to the vLLM project.\n Just a reminder: PRs would not trigger full CI run by default. Instead, it would only run `fastcheck` CI which starts running only a small and essential subset of CI tests to quickly catch errors. You can run other CI tests on top of those by going to your `fastcheck` build on Buildkite UI (linked in the PR checks section) and unblock them. If you do not have permission to unblock, ping `simon-mo` or `khluu` to add you in our Buildkite org. \n\nOnce the PR is approved and ready to go, your PR reviewer(s) can run CI to test the changes comprehensively before merging.\n\n To run CI, PR reviewers can do one of these:\n- Add `ready` label to the PR\n- Enable auto-merge.\n\n🚀'
+              body: '👋 Hi! Thank you for contributing to the vLLM project.\n\n' +
+                '💬 Join our developer Slack at https://slack.vllm.ai to discuss your PR in #pr-reviews, coordinate on features in #feat- channels, or join special interest groups in #sig- channels.\n\n' +
+                'Just a reminder: PRs would not trigger full CI run by default. Instead, it would only run `fastcheck` CI which starts running only a small and essential subset of CI tests to quickly catch errors. You can run other CI tests on top of those by going to your `fastcheck` build on Buildkite UI (linked in the PR checks section) and unblock them. If you do not have permission to unblock, ping `simon-mo` or `khluu` to add you in our Buildkite org.\n\n' +
+                'Once the PR is approved and ready to go, your PR reviewer(s) can run CI to test the changes comprehensively before merging.\n\n' +
+                'To run CI, PR reviewers can either: Add `ready` label to the PR or enable auto-merge.\n\n' +
+                '🚀'
            })
        env:
          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
@@ -1,52 +0,0 @@
-name: ruff
-
-on:
-  # Trigger the workflow on push or pull request,
-  # but only for the main branch
-  push:
-    branches:
-      - main
-    paths:
-      - "**/*.py"
-      - pyproject.toml
-      - requirements-lint.txt
-      - .github/workflows/matchers/ruff.json
-      - .github/workflows/ruff.yml
-  pull_request:
-    branches:
-      - main
-    # This workflow is only relevant when one of the following files changes.
-    # However, we have github configured to expect and require this workflow
-    # to run and pass before github with auto-merge a pull request. Until github
-    # allows more flexible auto-merge policy, we can just run this on every PR.
-    # It doesn't take that long to run, anyway.
-    #paths:
-    #  - "**/*.py"
-    #  - pyproject.toml
-    #  - requirements-lint.txt
-    #  - .github/workflows/matchers/ruff.json
-    #  - .github/workflows/ruff.yml
-
-jobs:
-  ruff:
-    runs-on: ubuntu-latest
-    strategy:
-      matrix:
-        python-version: ["3.12"]
-    steps:
-      - uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
-      - name: Set up Python ${{ matrix.python-version }}
-        uses: actions/setup-python@0b93645e9fea7318ecaed2b359559ac225c90a2b # v5.3.0
-        with:
-          python-version: ${{ matrix.python-version }}
-      - name: Install dependencies
-        run: |
-          python -m pip install --upgrade pip
-          pip install -r requirements-lint.txt
-      - name: Analysing the code with ruff
-        run: |
-          echo "::add-matcher::.github/workflows/matchers/ruff.json"
-          ruff check --output-format github .
-      - name: Run isort
-        run: |
-          isort . --check-only
@@ -1,37 +0,0 @@
-name: Lint shell scripts
-on:
-  push:
-    branches:
-      - "main"
-    paths:
-      - '**/*.sh'
-      - '.github/workflows/shellcheck.yml'
-  pull_request:
-    branches:
-      - "main"
-    paths:
-      - '**/*.sh'
-      - '.github/workflows/shellcheck.yml'
-
-env:
-  LC_ALL: en_US.UTF-8
-
-defaults:
-  run:
-    shell: bash
-
-permissions:
-  contents: read
-
-jobs:
-  shellcheck:
-    runs-on: ubuntu-latest
-    steps:
-      - name: "Checkout"
-        uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
-        with:
-          fetch-depth: 0
-
-      - name: "Check shell scripts"
-        run: |
-          tools/shellcheck.sh
@@ -1,32 +0,0 @@
-name: Lint documentation
-
-on:
-  push:
-    branches:
-      - main
-    paths:
-      - "docs/**"
-  pull_request:
-    branches:
-      - main
-    paths:
-      - "docs/**"
-
-jobs:
-  sphinx-lint:
-    runs-on: ubuntu-latest
-    strategy:
-      matrix:
-        python-version: ["3.12"]
-    steps:
-      - uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
-      - name: Set up Python ${{ matrix.python-version }}
-        uses: actions/setup-python@0b93645e9fea7318ecaed2b359559ac225c90a2b # v5.3.0
-        with:
-          python-version: ${{ matrix.python-version }}
-      - name: Install dependencies
-        run: |
-          python -m pip install --upgrade pip
-          pip install -r requirements-lint.txt
-      - name: Linting docs
-        run: tools/sphinx-lint.sh
@@ -13,7 +13,7 @@ jobs:
      actions: write
    runs-on: ubuntu-latest
    steps:
-      - uses: actions/stale@28ca1036281a5e5922ead5184a1bbf96e5fc984e # v9.0.0
+      - uses: actions/stale@5bef64f19d7facfb25b37b414482c7164d639639 # v9.1.0
        with:
          # Increasing this value ensures that changes to this workflow
          # propagate to all issues and PRs in days rather than months
@@ -1,38 +0,0 @@
-name: yapf
-
-on:
-  # Trigger the workflow on push or pull request,
-  # but only for the main branch
-  push:
-    branches:
-      - main
-    paths:
-      - "**/*.py"
-      - .github/workflows/yapf.yml
-  pull_request:
-    branches:
-      - main
-    paths:
-      - "**/*.py"
-      - .github/workflows/yapf.yml
-
-jobs:
-  yapf:
-    runs-on: ubuntu-latest
-    strategy:
-      matrix:
-        python-version: ["3.12"]
-    steps:
-      - uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
-      - name: Set up Python ${{ matrix.python-version }}
-        uses: actions/setup-python@0b93645e9fea7318ecaed2b359559ac225c90a2b # v5.3.0
-        with:
-          python-version: ${{ matrix.python-version }}
-      - name: Install dependencies
-        run: |
-          python -m pip install --upgrade pip
-          pip install yapf==0.32.0
-          pip install toml==0.10.2
-      - name: Running yapf
-        run: |
-          yapf --diff --recursive .
@@ -79,8 +79,7 @@ instance/

 # Sphinx documentation
 docs/_build/
-docs/source/getting_started/examples/*.rst
-!**/*.template.rst
+docs/source/getting_started/examples/

 # PyBuilder
 .pybuilder/
@@ -0,0 +1,129 @@
+default_stages:
+  - pre-commit # Run locally
+  - manual # Run in CI
+exclude: 'vllm/third_party/.*'
+repos:
+- repo: https://github.com/google/yapf
+  rev: v0.43.0
+  hooks:
+  - id: yapf
+    args: [--in-place, --verbose]
+    additional_dependencies: [toml] # TODO: Remove when yapf is upgraded
+- repo: https://github.com/astral-sh/ruff-pre-commit
+  rev: v0.9.3
+  hooks:
+  - id: ruff
+    args: [--output-format, github, --fix]
+- repo: https://github.com/codespell-project/codespell
+  rev: v2.4.0
+  hooks:
+  - id: codespell
+    additional_dependencies: ['tomli']
+    args: ['--toml', 'pyproject.toml']
+- repo: https://github.com/PyCQA/isort
+  rev: 0a0b7a830386ba6a31c2ec8316849ae4d1b8240d # 6.0.0
+  hooks:
+  - id: isort
+- repo: https://github.com/pre-commit/mirrors-clang-format
+  rev: v19.1.7
+  hooks:
+  - id: clang-format
+    exclude: 'csrc/(moe/topk_softmax_kernels.cu|quantization/gguf/(ggml-common.h|dequantize.cuh|vecdotq.cuh|mmq.cuh|mmvq.cuh))|vllm/third_party/.*'
+    types_or: [c++, cuda]
+    args: [--style=file, --verbose]
+- repo: https://github.com/jackdewinter/pymarkdown
+  rev: v0.9.27
+  hooks:
+  - id: pymarkdown
+    args: [fix]
+- repo: https://github.com/rhysd/actionlint
+  rev: v1.7.7
+  hooks:
+  - id: actionlint
+- repo: https://github.com/astral-sh/uv-pre-commit
+  rev: 0.6.2
+  hooks:
+    - id: pip-compile
+      args: [requirements-test.in, -o, requirements-test.txt]
+      files: ^requirements-test\.(in|txt)$
+- repo: local
+  hooks:
+  - id: mypy-local
+    name: Run mypy for local Python installation
+    entry: tools/mypy.sh 0 "local"
+    language: python
+    types: [python]
+    additional_dependencies: &mypy_deps [mypy==1.11.1, types-setuptools, types-PyYAML, types-requests]
+    stages: [pre-commit] # Don't run in CI
+  - id: mypy-3.9 # TODO: Use https://github.com/pre-commit/mirrors-mypy when mypy setup is less awkward
+    name: Run mypy for Python 3.9
+    entry: tools/mypy.sh 1 "3.9"
+    language: python
+    types: [python]
+    additional_dependencies: *mypy_deps
+    stages: [manual] # Only run in CI
+  - id: mypy-3.10 # TODO: Use https://github.com/pre-commit/mirrors-mypy when mypy setup is less awkward
+    name: Run mypy for Python 3.10
+    entry: tools/mypy.sh 1 "3.10"
+    language: python
+    types: [python]
+    additional_dependencies: *mypy_deps
+    stages: [manual] # Only run in CI
+  - id: mypy-3.11 # TODO: Use https://github.com/pre-commit/mirrors-mypy when mypy setup is less awkward
+    name: Run mypy for Python 3.11
+    entry: tools/mypy.sh 1 "3.11"
+    language: python
+    types: [python]
+    additional_dependencies: *mypy_deps
+    stages: [manual] # Only run in CI
+  - id: mypy-3.12 # TODO: Use https://github.com/pre-commit/mirrors-mypy when mypy setup is less awkward
+    name: Run mypy for Python 3.12
+    entry: tools/mypy.sh 1 "3.12"
+    language: python
+    types: [python]
+    additional_dependencies: *mypy_deps
+    stages: [manual] # Only run in CI
+  - id: shellcheck
+    name: Lint shell scripts
+    entry: tools/shellcheck.sh
+    language: script
+    types: [shell]
+  - id: png-lint
+    name: Lint PNG exports from excalidraw
+    entry: tools/png-lint.sh
+    language: script
+    types: [png]
+  - id: signoff-commit
+    name: Sign-off Commit
+    entry: bash
+    args:
+      - -c
+      - |
+        if ! grep -q "^Signed-off-by: $(git config user.name) <$(git config user.email)>" .git/COMMIT_EDITMSG; then
+          printf "\nSigned-off-by: $(git config user.name) <$(git config user.email)>\n" >> .git/COMMIT_EDITMSG
+        fi
+    language: system
+    verbose: true
+    stages: [commit-msg]
+  - id: check-spdx-header
+    name: Check SPDX headers
+    entry: python tools/check_spdx_header.py
+    language: python
+    types: [python]
+  - id: check-filenames
+    name: Check for spaces in all filenames
+    entry: bash
+    args:
+      - -c
+      - 'git ls-files | grep " " && echo "Filenames should not contain spaces!" && exit 1 || exit 0'
+    language: system
+    always_run: true
+    pass_filenames: false
+  # Keep `suggestion` last
+  - id: suggestion
+    name: Suggestion
+    entry: bash -c 'echo "To bypass pre-commit hooks, add --no-verify to git commit."'
+    language: system
+    verbose: true
+    pass_filenames: false
+  # Insert new entries above the `suggestion` entry
@@ -24,9 +24,6 @@ include(${CMAKE_CURRENT_LIST_DIR}/cmake/utils.cmake)
 # Suppress potential warnings about unused manually-specified variables
 set(ignoreMe "${VLLM_PYTHON_PATH}")

-# Prevent installation of dependencies (cutlass) by default.
-install(CODE "set(CMAKE_INSTALL_LOCAL_ONLY TRUE)" ALL_COMPONENTS)
-
 #
 # Supported python versions.  These versions will be searched in order, the
 # first match will be selected.  These should be kept in sync with setup.py.
@@ -34,10 +31,10 @@ install(CODE "set(CMAKE_INSTALL_LOCAL_ONLY TRUE)" ALL_COMPONENTS)
 set(PYTHON_SUPPORTED_VERSIONS "3.9" "3.10" "3.11" "3.12")

 # Supported NVIDIA architectures.
-set(CUDA_SUPPORTED_ARCHS "7.0;7.2;7.5;8.0;8.6;8.7;8.9;9.0")
+set(CUDA_SUPPORTED_ARCHS "7.0;7.2;7.5;8.0;8.6;8.7;8.9;9.0;10.0;10.1;12.0")

 # Supported AMD GPU architectures.
-set(HIP_SUPPORTED_ARCHS "gfx906;gfx908;gfx90a;gfx940;gfx941;gfx942;gfx1030;gfx1100;gfx1101")
+set(HIP_SUPPORTED_ARCHS "gfx906;gfx908;gfx90a;gfx942;gfx1030;gfx1100;gfx1101")

 #
 # Supported/expected torch versions for CUDA/ROCm.
@@ -177,10 +174,54 @@ include(FetchContent)
 file(MAKE_DIRECTORY ${FETCHCONTENT_BASE_DIR}) # Ensure the directory exists
 message(STATUS "FetchContent base directory: ${FETCHCONTENT_BASE_DIR}")

+#
+# Set rocm version dev int.
+#
+if(VLLM_GPU_LANG STREQUAL "HIP")
+  #
+  # Overriding the default -O set up by cmake, adding ggdb3 for the most verbose devug info
+  #
+  set(CMAKE_${VLLM_GPU_LANG}_FLAGS_DEBUG "${CMAKE_${VLLM_GPU_LANG}_FLAGS_DEBUG} -O0 -ggdb3")
+  set(CMAKE_CXX_FLAGS_DEBUG "${CMAKE_CXX_FLAGS_DEBUG} -O0 -ggdb3")
+
+
+  #
+  # Certain HIP functions are marked as [[nodiscard]], yet vllm ignores the result which generates
+  # a lot of warnings that always mask real issues. Suppressing until this is properly addressed.
+  #
+  set(CMAKE_${VLLM_GPU_LANG}_FLAGS "${CMAKE_${VLLM_GPU_LANG}_FLAGS} -Wno-unused-result")
+  set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wno-unused-result")
+endif()
+
 #
 # Define other extension targets
 #

+#
+# cumem_allocator extension
+#
+
+set(VLLM_CUMEM_EXT_SRC
+  "csrc/cumem_allocator.cpp")
+
+set_gencode_flags_for_srcs(
+  SRCS "${VLLM_CUMEM_EXT_SRC}"
+  CUDA_ARCHS "${CUDA_ARCHS}")
+
+if(VLLM_GPU_LANG STREQUAL "CUDA")
+  message(STATUS "Enabling cumem allocator extension.")
+  # link against cuda driver library
+  list(APPEND CUMEM_LIBS CUDA::cuda_driver)
+  define_gpu_extension_target(
+    cumem_allocator
+    DESTINATION vllm
+    LANGUAGE CXX
+    SOURCES ${VLLM_CUMEM_EXT_SRC}
+    LIBRARIES ${CUMEM_LIBS}
+    USE_SABI 3.8
+    WITH_SOABI)
+endif()
+
 #
 # _C extension
 #
@@ -206,7 +247,8 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
  SET(CUTLASS_ENABLE_HEADERS_ONLY ON CACHE BOOL "Enable only the header library")

  # Set CUTLASS_REVISION manually -- its revision detection doesn't work in this case.
-  set(CUTLASS_REVISION "v3.5.1" CACHE STRING "CUTLASS revision to use")
+  # Please keep this in sync with FetchContent_Declare line below.
+  set(CUTLASS_REVISION "v3.8.0" 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})
@@ -223,7 +265,8 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
    FetchContent_Declare(
        cutlass
        GIT_REPOSITORY https://github.com/nvidia/cutlass.git
-        GIT_TAG v3.5.1
+        # Please keep this in sync with CUTLASS_REVISION line above.
+        GIT_TAG v3.8.0
        GIT_PROGRESS TRUE

        # Speed up CUTLASS download by retrieving only the specified GIT_TAG instead of the history.
@@ -241,7 +284,11 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
    "csrc/quantization/awq/gemm_kernels.cu"
    "csrc/custom_all_reduce.cu"
    "csrc/permute_cols.cu"
-    "csrc/quantization/cutlass_w8a8/scaled_mm_entry.cu")
+    "csrc/quantization/cutlass_w8a8/scaled_mm_entry.cu"
+    "csrc/quantization/fp4/nvfp4_quant_entry.cu"
+    "csrc/quantization/fp4/nvfp4_scaled_mm_entry.cu"
+    "csrc/sparse/cutlass/sparse_scaled_mm_entry.cu"
+    "csrc/cutlass_extensions/common.cpp")

  set_gencode_flags_for_srcs(
    SRCS "${VLLM_EXT_SRC}"
@@ -250,7 +297,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
  # Only build Marlin kernels if we are building for at least some compatible archs.
  # Keep building Marlin for 9.0 as there are some group sizes and shapes that
  # are not supported by Machete yet.
-  cuda_archs_loose_intersection(MARLIN_ARCHS "8.0;8.6;8.7;8.9;9.0" ${CUDA_ARCHS})
+  cuda_archs_loose_intersection(MARLIN_ARCHS "8.0;8.6;8.7;8.9;9.0;10.0;10.1;12.0" "${CUDA_ARCHS}")
  if (MARLIN_ARCHS)
    set(MARLIN_SRCS
       "csrc/quantization/fp8/fp8_marlin.cu"
@@ -270,12 +317,32 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
                   " in CUDA target architectures")
  endif()

-  #
+  # Only build AllSpark kernels if we are building for at least some compatible archs.
+  cuda_archs_loose_intersection(ALLSPARK_ARCHS "8.0;8.6;8.7;8.9" "${CUDA_ARCHS}")
+  if (${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.0 AND ALLSPARK_ARCHS)
+    set(ALLSPARK_SRCS
+       "csrc/quantization/gptq_allspark/allspark_repack.cu"
+       "csrc/quantization/gptq_allspark/allspark_qgemm_w8a16.cu")
+    set_gencode_flags_for_srcs(
+      SRCS "${ALLSPARK_SRCS}"
+      CUDA_ARCHS "${ALLSPARK_ARCHS}")
+    list(APPEND VLLM_EXT_SRC "${ALLSPARK_SRCS}")
+    message(STATUS "Building AllSpark kernels for archs: ${ALLSPARK_ARCHS}")
+  else()
+    message(STATUS "Not building AllSpark kernels as no compatible archs found"
+                   " in CUDA target architectures, or CUDA not >= 12.0")
+  endif()
+
  # The cutlass_scaled_mm kernels for Hopper (c3x, i.e. CUTLASS 3.x) require
-  # CUDA 12.0 or later (and only work on Hopper, 9.0/9.0a for now).
-  cuda_archs_loose_intersection(SCALED_MM_3X_ARCHS "9.0;9.0a" "${CUDA_ARCHS}")
+  # CUDA 12.0 or later (and only work on Hopper, 9.0a for now).
+  cuda_archs_loose_intersection(SCALED_MM_3X_ARCHS "9.0a;10.0a;10.1a;12.0a" "${CUDA_ARCHS}")
  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.0 AND SCALED_MM_3X_ARCHS)
-    set(SRCS "csrc/quantization/cutlass_w8a8/scaled_mm_c3x.cu")
+    set(SRCS
+       "csrc/quantization/cutlass_w8a8/scaled_mm_c3x.cu"
+       "csrc/quantization/cutlass_w8a8/c3x/scaled_mm_sm90_fp8.cu"
+       "csrc/quantization/cutlass_w8a8/c3x/scaled_mm_sm90_int8.cu"
+       "csrc/quantization/cutlass_w8a8/c3x/scaled_mm_azp_sm90_int8.cu"
+       "csrc/quantization/cutlass_w8a8/c3x/scaled_mm_blockwise_sm90_fp8.cu")
    set_gencode_flags_for_srcs(
      SRCS "${SRCS}"
      CUDA_ARCHS "${SCALED_MM_3X_ARCHS}")
@@ -302,7 +369,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
  # For the cutlass_scaled_mm kernels we want to build the c2x (CUTLASS 2.x)
  # kernels for the remaining archs that are not already built for 3x.
  cuda_archs_loose_intersection(SCALED_MM_2X_ARCHS
-    "7.5;8.0;8.6;8.7;8.9;9.0" "${CUDA_ARCHS}")
+    "7.5;8.0;8.6;8.7;8.9;9.0;10.0;10.1;12.0" "${CUDA_ARCHS}")
  # subtract out the archs that are already built for 3x
  list(REMOVE_ITEM SCALED_MM_2X_ARCHS ${SCALED_MM_3X_ARCHS})
  if (SCALED_MM_2X_ARCHS)
@@ -323,7 +390,64 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
    endif()
  endif()

+  #
+  # 2:4 Sparse Kernels

+  # The 2:4 sparse kernels cutlass_scaled_sparse_mm and cutlass_compressor
+  # require CUDA 12.2 or later (and only work on Hopper and Blackwell).
+  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.2 AND SCALED_MM_3X_ARCHS)
+    set(SRCS "csrc/sparse/cutlass/sparse_scaled_mm_c3x.cu")
+    set_gencode_flags_for_srcs(
+      SRCS "${SRCS}"
+      CUDA_ARCHS "${SCALED_MM_3X_ARCHS}")
+    list(APPEND VLLM_EXT_SRC "${SRCS}")
+    list(APPEND VLLM_GPU_FLAGS "-DENABLE_SPARSE_SCALED_MM_C3X=1")
+    message(STATUS "Building sparse_scaled_mm_c3x for archs: ${SCALED_MM_3X_ARCHS}")
+  else()
+    if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.2 AND SCALED_MM_3X_ARCHS)
+      message(STATUS "Not building sparse_scaled_mm_c3x kernels as CUDA Compiler version is "
+                     "not >= 12.2, we recommend upgrading to CUDA 12.2 or later "
+                     "if you intend on running FP8 sparse quantized models on Hopper.")
+    else()
+      message(STATUS "Not building sparse_scaled_mm_c3x as no compatible archs found "
+                     "in CUDA target architectures")
+    endif()
+  endif()
+
+  # FP4 Archs and flags
+  cuda_archs_loose_intersection(FP4_ARCHS "10.0a" "${CUDA_ARCHS}")
+  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.8 AND FP4_ARCHS)
+    set(SRCS
+      "csrc/quantization/fp4/nvfp4_quant_kernels.cu"
+      "csrc/quantization/fp4/nvfp4_scaled_mm_kernels.cu")
+    set_gencode_flags_for_srcs(
+      SRCS "${SRCS}"
+      CUDA_ARCHS "${FP4_ARCHS}")
+    list(APPEND VLLM_EXT_SRC "${SRCS}")
+    list(APPEND VLLM_GPU_FLAGS "-DENABLE_NVFP4=1")
+    message(STATUS "Building NVFP4 for archs: ${FP4_ARCHS}")
+  else()
+    message(STATUS "Not building NVFP4 as no compatible archs were found.")
+    # clear FP4_ARCHS
+    set(FP4_ARCHS)
+  endif()
+
+  # FP8 Blackwell Archs 
+  cuda_archs_loose_intersection(BLACKWELL_ARCHS "10.0;10.1;12.0" "${CUDA_ARCHS}")
+  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.8 AND BLACKWELL_ARCHS)
+    set(SRCS 
+      "csrc/quantization/cutlass_w8a8/c3x/scaled_mm_sm100_fp8.cu"
+    )
+    set_gencode_flags_for_srcs(
+      SRCS "${SRCS}"
+      CUDA_ARCHS "${BLACKWELL_ARCHS}")
+    list(APPEND VLLM_EXT_SRC "${SRCS}")
+    message(STATUS "Building FP8 for archs: ${BLACKWELL_ARCHS}")
+  else()
+    # clear BLACKWELL_ARCHS
+    set(BLACKWELL_ARCHS)
+  endif()
+  
  #
  # Machete kernels

@@ -405,6 +529,7 @@ define_gpu_extension_target(
  COMPILE_FLAGS ${VLLM_GPU_FLAGS}
  ARCHITECTURES ${VLLM_GPU_ARCHES}
  INCLUDE_DIRECTORIES ${CUTLASS_INCLUDE_DIR}
+  INCLUDE_DIRECTORIES ${CUTLASS_TOOLS_UTIL_INCLUDE_DIR}
  USE_SABI 3
  WITH_SOABI)

@@ -428,7 +553,7 @@ set_gencode_flags_for_srcs(
  CUDA_ARCHS "${CUDA_ARCHS}")

 if(VLLM_GPU_LANG STREQUAL "CUDA")
-  cuda_archs_loose_intersection(MARLIN_MOE_ARCHS "8.0;8.6;8.7;8.9;9.0" "${CUDA_ARCHS}")
+  cuda_archs_loose_intersection(MARLIN_MOE_ARCHS "8.0;8.6;8.7;8.9;9.0;10.0;10.1;12.0" "${CUDA_ARCHS}")
  if (MARLIN_MOE_ARCHS)
    set(MARLIN_MOE_SRC
        "csrc/moe/marlin_kernels/marlin_moe_kernel.h"
@@ -482,79 +607,8 @@ if(VLLM_GPU_LANG STREQUAL "HIP")
    WITH_SOABI)
 endif()

-# vllm-flash-attn currently only supported on CUDA
-if (NOT VLLM_TARGET_DEVICE STREQUAL "cuda")
-  return()
+# For CUDA we also build and ship some external projects.
+if (VLLM_GPU_LANG STREQUAL "CUDA")
+    include(cmake/external_projects/flashmla.cmake)
+    include(cmake/external_projects/vllm_flash_attn.cmake)
 endif ()
-
-# vLLM flash attention requires VLLM_GPU_ARCHES to contain the set of target
-# arches in the CMake syntax (75-real, 89-virtual, etc), since we clear the
-# arches in the CUDA case (and instead set the gencodes on a per file basis)
-# we need to manually set VLLM_GPU_ARCHES here.
-if(VLLM_GPU_LANG STREQUAL "CUDA")
-  foreach(_ARCH ${CUDA_ARCHS})
-    string(REPLACE "." "" _ARCH "${_ARCH}")
-    list(APPEND VLLM_GPU_ARCHES "${_ARCH}-real")
-  endforeach()
-endif()
-
-#
-# Build vLLM flash attention from source
-#
-# IMPORTANT: This has to be the last thing we do, because vllm-flash-attn uses the same macros/functions as vLLM.
-# Because functions all belong to the global scope, vllm-flash-attn's functions overwrite vLLMs.
-# They should be identical but if they aren't, this is a massive footgun.
-#
-# The vllm-flash-attn install rules are nested under vllm to make sure the library gets installed in the correct place.
-# To only install vllm-flash-attn, use --component vllm_flash_attn_c.
-# If no component is specified, vllm-flash-attn is still installed.
-
-# If VLLM_FLASH_ATTN_SRC_DIR is set, vllm-flash-attn is installed from that directory instead of downloading.
-# This is to enable local development of vllm-flash-attn within vLLM.
-# It can be set as an environment variable or passed as a cmake argument.
-# The environment variable takes precedence.
-if (DEFINED ENV{VLLM_FLASH_ATTN_SRC_DIR})
-  set(VLLM_FLASH_ATTN_SRC_DIR $ENV{VLLM_FLASH_ATTN_SRC_DIR})
-endif()
-
-if(VLLM_FLASH_ATTN_SRC_DIR)
-  FetchContent_Declare(vllm-flash-attn SOURCE_DIR ${VLLM_FLASH_ATTN_SRC_DIR})
-else()
-  FetchContent_Declare(
-          vllm-flash-attn
-          GIT_REPOSITORY https://github.com/vllm-project/flash-attention.git
-          GIT_TAG 04325b6798bcc326c86fb35af62d05a9c8c8eceb
-          GIT_PROGRESS TRUE
-          # Don't share the vllm-flash-attn build between build types
-          BINARY_DIR ${CMAKE_BINARY_DIR}/vllm-flash-attn
-  )
-endif()
-
-# Set the parent build flag so that the vllm-flash-attn library does not redo compile flag and arch initialization.
-set(VLLM_PARENT_BUILD ON)
-
-# Ensure the vllm/vllm_flash_attn directory exists before installation
-install(CODE "file(MAKE_DIRECTORY \"\${CMAKE_INSTALL_PREFIX}/vllm/vllm_flash_attn\")" COMPONENT vllm_flash_attn_c)
-
-# Make sure vllm-flash-attn install rules are nested under vllm/
-install(CODE "set(CMAKE_INSTALL_LOCAL_ONLY FALSE)" COMPONENT vllm_flash_attn_c)
-install(CODE "set(OLD_CMAKE_INSTALL_PREFIX \"\${CMAKE_INSTALL_PREFIX}\")" COMPONENT vllm_flash_attn_c)
-install(CODE "set(CMAKE_INSTALL_PREFIX \"\${CMAKE_INSTALL_PREFIX}/vllm/\")" COMPONENT vllm_flash_attn_c)
-
-# Fetch the vllm-flash-attn library
-FetchContent_MakeAvailable(vllm-flash-attn)
-message(STATUS "vllm-flash-attn is available at ${vllm-flash-attn_SOURCE_DIR}")
-
-# Restore the install prefix
-install(CODE "set(CMAKE_INSTALL_PREFIX \"\${OLD_CMAKE_INSTALL_PREFIX}\")" COMPONENT vllm_flash_attn_c)
-install(CODE "set(CMAKE_INSTALL_LOCAL_ONLY TRUE)" COMPONENT vllm_flash_attn_c)
-
-# Copy over the vllm-flash-attn python files
-install(
-        DIRECTORY ${vllm-flash-attn_SOURCE_DIR}/vllm_flash_attn/
-        DESTINATION vllm/vllm_flash_attn
-        COMPONENT vllm_flash_attn_c
-        FILES_MATCHING PATTERN "*.py"
-)
-
-# Nothing after vllm-flash-attn, see comment about macros above
@@ -61,7 +61,7 @@ representative at an online or offline/IRL event.

 Instances of abusive, harassing, or otherwise unacceptable behavior may be
 reported to the community leaders responsible for enforcement in the #code-of-conduct
-channel in the [vLLM Discord](https://discord.com/invite/jz7wjKhh6g).
+channel in the [vLLM Slack](https://slack.vllm.ai).
 All complaints will be reviewed and investigated promptly and fairly.

 All community leaders are obligated to respect the privacy and security of the
@@ -125,4 +125,3 @@ Community Impact Guidelines were inspired by
 For answers to common questions about this code of conduct, see the
 [Contributor Covenant FAQ](https://www.contributor-covenant.org/faq). Translations are available at
 [Contributor Covenant translations](https://www.contributor-covenant.org/translations).
-
@@ -2,8 +2,8 @@
 # to run the OpenAI compatible server.

 # Please update any changes made here to
-# docs/source/dev/dockerfile/dockerfile.rst and
-# docs/source/assets/dev/dockerfile-stages-dependency.png
+# docs/source/contributing/dockerfile/dockerfile.md and
+# docs/source/assets/contributing/dockerfile-stages-dependency.png

 ARG CUDA_VERSION=12.4.1
 #################### BASE BUILD IMAGE ####################
@@ -27,6 +27,9 @@ RUN echo 'tzdata tzdata/Areas select America' | debconf-set-selections \
    && ln -sf /usr/bin/python${PYTHON_VERSION}-config /usr/bin/python3-config \
    && curl -sS https://bootstrap.pypa.io/get-pip.py | python${PYTHON_VERSION} \
    && python3 --version && python3 -m pip --version
+# Install uv for faster pip installs
+RUN --mount=type=cache,target=/root/.cache/uv \
+    python3 -m pip install uv

 # Upgrade to GCC 10 to avoid https://gcc.gnu.org/bugzilla/show_bug.cgi?id=92519
 # as it was causing spam when compiling the CUTLASS kernels
@@ -45,16 +48,20 @@ RUN ldconfig /usr/local/cuda-$(echo $CUDA_VERSION | cut -d. -f1,2)/compat/
 WORKDIR /workspace

 # install build and runtime dependencies
+
+# arm64 (GH200) build follows the practice of "use existing pytorch" build,
+# we need to install torch and torchvision from the nightly builds first,
+# pytorch will not appear as a vLLM dependency in all of the following steps
+# after this step
+RUN --mount=type=cache,target=/root/.cache/uv \
+    if [ "$TARGETPLATFORM" = "linux/arm64" ]; then \
+        uv pip install --system --index-url https://download.pytorch.org/whl/nightly/cu126 "torch==2.7.0.dev20250121+cu126" "torchvision==0.22.0.dev20250121";  \
+    fi
+
 COPY requirements-common.txt requirements-common.txt
 COPY requirements-cuda.txt requirements-cuda.txt
-COPY requirements-cuda-arm64.txt requirements-cuda-arm64.txt
-RUN --mount=type=cache,target=/root/.cache/pip \
-    python3 -m pip install -r requirements-cuda.txt
-
-RUN --mount=type=cache,target=/root/.cache/pip \
-    if [ "$TARGETPLATFORM" = "linux/arm64" ]; then \
-        python3 -m pip install -r requirements-cuda-arm64.txt; \
-    fi
+RUN --mount=type=cache,target=/root/.cache/uv \
+    uv pip install --system -r requirements-cuda.txt

 # cuda arch list used by torch
 # can be useful for both `dev` and `test`
@@ -74,13 +81,8 @@ ARG TARGETPLATFORM
 # install build dependencies
 COPY requirements-build.txt requirements-build.txt

-RUN --mount=type=cache,target=/root/.cache/pip \
-    python3 -m pip install -r requirements-build.txt
-
-RUN --mount=type=cache,target=/root/.cache/pip \
-    if [ "$TARGETPLATFORM" = "linux/arm64" ]; then \
-        python3 -m pip install -r requirements-cuda-arm64.txt; \
-    fi
+RUN --mount=type=cache,target=/root/.cache/uv \
+    uv pip install --system -r requirements-build.txt

 COPY . .
 ARG GIT_REPO_CHECK=0
@@ -99,7 +101,7 @@ ARG SCCACHE_BUCKET_NAME=vllm-build-sccache
 ARG SCCACHE_REGION_NAME=us-west-2
 ARG SCCACHE_S3_NO_CREDENTIALS=0
 # if USE_SCCACHE is set, use sccache to speed up compilation
-RUN --mount=type=cache,target=/root/.cache/pip \
+RUN --mount=type=cache,target=/root/.cache/uv \
    --mount=type=bind,source=.git,target=.git \
    if [ "$USE_SCCACHE" = "1" ]; then \
        echo "Installing sccache..." \
@@ -119,7 +121,7 @@ RUN --mount=type=cache,target=/root/.cache/pip \

 ENV CCACHE_DIR=/root/.cache/ccache
 RUN --mount=type=cache,target=/root/.cache/ccache \
-    --mount=type=cache,target=/root/.cache/pip \
+    --mount=type=cache,target=/root/.cache/uv \
    --mount=type=bind,source=.git,target=.git  \
    if [ "$USE_SCCACHE" != "1" ]; then \
        python3 setup.py bdist_wheel --dist-dir=dist --py-limited-api=cp38; \
@@ -127,8 +129,8 @@ RUN --mount=type=cache,target=/root/.cache/ccache \

 # Check the size of the wheel if RUN_WHEEL_CHECK is true
 COPY .buildkite/check-wheel-size.py check-wheel-size.py
-# Default max size of the wheel is 250MB
-ARG VLLM_MAX_SIZE_MB=250
+# sync the default value with .buildkite/check-wheel-size.py
+ARG VLLM_MAX_SIZE_MB=400
 ENV VLLM_MAX_SIZE_MB=$VLLM_MAX_SIZE_MB
 ARG RUN_WHEEL_CHECK=true
 RUN if [ "$RUN_WHEEL_CHECK" = "true" ]; then \
@@ -144,21 +146,20 @@ FROM base as dev
 COPY requirements-lint.txt requirements-lint.txt
 COPY requirements-test.txt requirements-test.txt
 COPY requirements-dev.txt requirements-dev.txt
-RUN --mount=type=cache,target=/root/.cache/pip \
-    python3 -m pip install -r requirements-dev.txt
+RUN --mount=type=cache,target=/root/.cache/uv \
+    uv pip install --system -r requirements-dev.txt
 #################### DEV IMAGE ####################

 #################### vLLM installation IMAGE ####################
 # image with vLLM installed
-FROM nvidia/cuda:${CUDA_VERSION}-base-ubuntu22.04 AS vllm-base
+# TODO: Restore to base image after FlashInfer AOT wheel fixed
+FROM nvidia/cuda:${CUDA_VERSION}-devel-ubuntu22.04 AS vllm-base
 ARG CUDA_VERSION=12.4.1
 ARG PYTHON_VERSION=3.12
 WORKDIR /vllm-workspace
 ENV DEBIAN_FRONTEND=noninteractive
 ARG TARGETPLATFORM

-COPY requirements-cuda-arm64.txt requirements-cuda-arm64.txt
-
 RUN PYTHON_VERSION_STR=$(echo ${PYTHON_VERSION} | sed 's/\.//g') && \
    echo "export PYTHON_VERSION_STR=${PYTHON_VERSION_STR}" >> /etc/environment

@@ -166,7 +167,7 @@ RUN PYTHON_VERSION_STR=$(echo ${PYTHON_VERSION} | sed 's/\.//g') && \
 RUN echo 'tzdata tzdata/Areas select America' | debconf-set-selections \
    && echo 'tzdata tzdata/Zones/America select Los_Angeles' | debconf-set-selections \
    && apt-get update -y \
-    && apt-get install -y ccache software-properties-common git curl sudo vim python3-pip \
+    && apt-get install -y ccache software-properties-common git curl wget sudo vim python3-pip \
    && apt-get install -y ffmpeg libsm6 libxext6 libgl1 \
    && add-apt-repository ppa:deadsnakes/ppa \
    && apt-get update -y \
@@ -176,6 +177,9 @@ RUN echo 'tzdata tzdata/Areas select America' | debconf-set-selections \
    && ln -sf /usr/bin/python${PYTHON_VERSION}-config /usr/bin/python3-config \
    && curl -sS https://bootstrap.pypa.io/get-pip.py | python${PYTHON_VERSION} \
    && python3 --version && python3 -m pip --version
+# Install uv for faster pip installs
+RUN --mount=type=cache,target=/root/.cache/uv \
+    python3 -m pip install uv

 # Workaround for https://github.com/openai/triton/issues/2507 and
 # https://github.com/pytorch/pytorch/issues/107960 -- hopefully
@@ -183,23 +187,47 @@ RUN echo 'tzdata tzdata/Areas select America' | debconf-set-selections \
 # or future versions of triton.
 RUN ldconfig /usr/local/cuda-$(echo $CUDA_VERSION | cut -d. -f1,2)/compat/

-# Install vllm wheel first, so that torch etc will be installed.
-RUN --mount=type=bind,from=build,src=/workspace/dist,target=/vllm-workspace/dist \
-    --mount=type=cache,target=/root/.cache/pip \
-    python3 -m pip install dist/*.whl --verbose
-
-RUN --mount=type=cache,target=/root/.cache/pip \
+# arm64 (GH200) build follows the practice of "use existing pytorch" build,
+# we need to install torch and torchvision from the nightly builds first,
+# pytorch will not appear as a vLLM dependency in all of the following steps
+# after this step
+RUN --mount=type=cache,target=/root/.cache/uv \
    if [ "$TARGETPLATFORM" = "linux/arm64" ]; then \
-        pip uninstall -y torch && \
-        python3 -m pip install -r requirements-cuda-arm64.txt; \
+        uv pip install --system --index-url https://download.pytorch.org/whl/nightly/cu124 "torch==2.6.0.dev20241210+cu124" "torchvision==0.22.0.dev20241215";  \
    fi

-RUN --mount=type=cache,target=/root/.cache/pip \
+# Install vllm wheel first, so that torch etc will be installed.
+RUN --mount=type=bind,from=build,src=/workspace/dist,target=/vllm-workspace/dist \
+    --mount=type=cache,target=/root/.cache/uv \
+    uv pip install --system dist/*.whl --verbose
+
+# If we need to build FlashInfer wheel before its release:
+# $ export FLASHINFER_ENABLE_AOT=1
+# $ # Note we remove 7.0 from the arch list compared to the list below, since FlashInfer only supports sm75+
+# $ export TORCH_CUDA_ARCH_LIST='7.5 8.0 8.6 8.9 9.0+PTX'
+# $ git clone https://github.com/flashinfer-ai/flashinfer.git --recursive
+# $ cd flashinfer
+# $ git checkout 524304395bd1d8cd7d07db083859523fcaa246a4
+# $ rm -rf build
+# $ python3 setup.py bdist_wheel --dist-dir=dist --verbose
+# $ ls dist
+# $ # upload the wheel to a public location, e.g. https://wheels.vllm.ai/flashinfer/524304395bd1d8cd7d07db083859523fcaa246a4/flashinfer_python-0.2.1.post1+cu124torch2.5-cp38-abi3-linux_x86_64.whl
+
+RUN --mount=type=cache,target=/root/.cache/uv \
 . /etc/environment && \
 if [ "$TARGETPLATFORM" != "linux/arm64" ]; then \
-    python3 -m pip install https://github.com/flashinfer-ai/flashinfer/releases/download/v0.1.6/flashinfer-0.1.6+cu121torch2.4-cp${PYTHON_VERSION_STR}-cp${PYTHON_VERSION_STR}-linux_x86_64.whl; \
+    uv pip install --system https://github.com/flashinfer-ai/flashinfer/releases/download/v0.2.1.post1/flashinfer_python-0.2.1.post1+cu124torch2.5-cp38-abi3-linux_x86_64.whl ; \
 fi
 COPY examples examples
+
+# Although we build Flashinfer with AOT mode, there's still
+# some issues w.r.t. JIT compilation. Therefore we need to
+# install build dependencies for JIT compilation.
+# TODO: Remove this once FlashInfer AOT wheel is fixed
+COPY requirements-build.txt requirements-build.txt
+RUN --mount=type=cache,target=/root/.cache/uv \
+    uv pip install --system -r requirements-build.txt
+
 #################### vLLM installation IMAGE ####################

 #################### TEST IMAGE ####################
@@ -210,16 +238,16 @@ FROM vllm-base AS test
 ADD . /vllm-workspace/

 # install development dependencies (for testing)
-RUN --mount=type=cache,target=/root/.cache/pip \
-    python3 -m pip install -r requirements-dev.txt
+RUN --mount=type=cache,target=/root/.cache/uv \
+    uv pip install --system -r requirements-dev.txt

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

 # enable fast downloads from hf (for testing)
-RUN --mount=type=cache,target=/root/.cache/pip \
-    python3 -m pip install hf_transfer
+RUN --mount=type=cache,target=/root/.cache/uv \
+    uv pip install --system hf_transfer
 ENV HF_HUB_ENABLE_HF_TRANSFER 1

 # Copy in the v1 package for testing (it isn't distributed yet)
@@ -234,17 +262,27 @@ RUN mv vllm test_docs/
 #################### TEST IMAGE ####################

 #################### OPENAI API SERVER ####################
-# openai api server alternative
-FROM vllm-base AS vllm-openai
+# base openai image with additional requirements, for any subsequent openai-style images
+FROM vllm-base AS vllm-openai-base

 # install additional dependencies for openai api server
-RUN --mount=type=cache,target=/root/.cache/pip \
+RUN --mount=type=cache,target=/root/.cache/uv \
    if [ "$TARGETPLATFORM" = "linux/arm64" ]; then \
-        pip install accelerate hf_transfer 'modelscope!=1.15.0' 'bitsandbytes>=0.42.0' 'timm==0.9.10'; \
+        uv pip install --system accelerate hf_transfer 'modelscope!=1.15.0' 'bitsandbytes>=0.42.0' 'timm==0.9.10' boto3 runai-model-streamer runai-model-streamer[s3]; \
    else \
-        pip install accelerate hf_transfer 'modelscope!=1.15.0' 'bitsandbytes>=0.45.0' 'timm==0.9.10'; \
+        uv pip install --system accelerate hf_transfer 'modelscope!=1.15.0' 'bitsandbytes>=0.45.0' 'timm==0.9.10' boto3 runai-model-streamer runai-model-streamer[s3]; \
    fi
+
 ENV VLLM_USAGE_SOURCE production-docker-image

+# define sagemaker first, so it is not default from `docker build`
+FROM vllm-openai-base AS vllm-sagemaker
+
+COPY examples/online_serving/sagemaker-entrypoint.sh .
+RUN chmod +x sagemaker-entrypoint.sh
+ENTRYPOINT ["./sagemaker-entrypoint.sh"]
+
+FROM vllm-openai-base AS vllm-openai
+
 ENTRYPOINT ["python3", "-m", "vllm.entrypoints.openai.api_server"]
 #################### OPENAI API SERVER ####################
@@ -1,4 +1,4 @@
-FROM vault.habana.ai/gaudi-docker/1.18.0/ubuntu22.04/habanalabs/pytorch-installer-2.4.0:latest
+FROM vault.habana.ai/gaudi-docker/1.19.1/ubuntu22.04/habanalabs/pytorch-installer-2.5.1:latest

 COPY ./ /workspace/vllm

@@ -1,6 +1,6 @@
 # default base image
 # https://gallery.ecr.aws/neuron/pytorch-inference-neuronx
-ARG BASE_IMAGE="public.ecr.aws/neuron/pytorch-inference-neuronx:2.1.2-neuronx-py310-sdk2.20.2-ubuntu20.04"
+ARG BASE_IMAGE="public.ecr.aws/neuron/pytorch-inference-neuronx:2.5.1-neuronx-py310-sdk2.21.0-ubuntu22.04"

 FROM $BASE_IMAGE

@@ -15,16 +15,19 @@ RUN apt-get update && \
        ffmpeg libsm6 libxext6 libgl1

 ### Mount Point ###
-# When launching the container, mount the code directory to /app
-ARG APP_MOUNT=/app
+# When launching the container, mount the code directory to /workspace
+ARG APP_MOUNT=/workspace
 VOLUME [ ${APP_MOUNT} ]
 WORKDIR ${APP_MOUNT}/vllm

 RUN python3 -m pip install --upgrade pip
 RUN python3 -m pip install --no-cache-dir fastapi ninja tokenizers pandas
-RUN python3 -m pip install sentencepiece transformers==4.36.2 -U
-RUN python3 -m pip install transformers-neuronx --extra-index-url=https://pip.repos.neuron.amazonaws.com -U
-RUN python3 -m pip install --pre neuronx-cc==2.15.* --extra-index-url=https://pip.repos.neuron.amazonaws.com -U
+RUN python3 -m pip install sentencepiece transformers==4.45.2 -U
+RUN python3 -m pip install neuronx-cc==2.16.345.0 --extra-index-url=https://pip.repos.neuron.amazonaws.com -U
+RUN python3 -m pip install pytest
+
+# uninstall transformers-neuronx package explicitly to avoid version conflict
+RUN python3 -m pip uninstall -y transformers-neuronx

 COPY . .
 ARG GIT_REPO_CHECK=0
@@ -42,4 +45,11 @@ RUN --mount=type=bind,source=.git,target=.git \
 # install development dependencies (for testing)
 RUN python3 -m pip install -e tests/vllm_test_utils

+# install transformers-neuronx package as an optional dependencies (for V0)
+# FIXME: `--no-deps` argument is temporarily added to resolve transformers package version conflict
+RUN python3 -m pip install transformers-neuronx==0.13.* --extra-index-url=https://pip.repos.neuron.amazonaws.com -U --no-deps
+
+# overwrite entrypoint to run bash script
+RUN echo "import subprocess; import sys; subprocess.check_call(sys.argv[1:])" > /usr/local/bin/dockerd-entrypoint.py
+
 CMD ["/bin/bash"]
@@ -14,6 +14,7 @@ ARG GIT_REPO_CHECK=0
 RUN --mount=type=bind,source=.git,target=.git \
    if [ "$GIT_REPO_CHECK" != 0 ]; then bash tools/check_repo.sh ; fi

+RUN python3 -m pip install -U pip
 # install build requirements
 RUN PIP_EXTRA_INDEX_URL="https://download.pytorch.org/whl/cpu" python3 -m pip install -r /workspace/requirements-build.txt
 # build vLLM with OpenVINO backend
@@ -4,12 +4,12 @@ USER root

 ENV PATH="/usr/local/cargo/bin:$PATH:/opt/conda/bin/"

-RUN apt-get update -y && apt-get install -y git wget curl vim libnuma-dev libsndfile-dev libprotobuf-dev build-essential ffmpeg libsm6 libxext6 libgl1 
+RUN apt-get update -y && apt-get install -y git wget kmod curl vim libnuma-dev libsndfile-dev libprotobuf-dev build-essential ffmpeg libsm6 libxext6 libgl1 libssl-dev 

 # Some packages in requirements-cpu are installed here
 # IBM provides optimized packages for ppc64le processors in the open-ce project for mamba
 # Currently these may not be available for venv or pip directly
-RUN micromamba install -y -n base -c https://ftp.osuosl.org/pub/open-ce/1.11.0-p10/ -c defaults python=3.10 torchvision-cpu=0.16.2 rust && micromamba clean --all --yes
+RUN micromamba install -y -n base -c https://ftp.osuosl.org/pub/open-ce/1.11.0-p10/ -c defaults python=3.10 rust && micromamba clean --all --yes

 COPY ./ /workspace/vllm

@@ -18,11 +18,9 @@ ARG GIT_REPO_CHECK=0
 RUN --mount=type=bind,source=.git,target=.git \
    if [ "$GIT_REPO_CHECK" != 0 ]; then bash tools/check_repo.sh; fi

-# These packages will be in rocketce eventually
 RUN --mount=type=cache,target=/root/.cache/pip  \
-    pip install -v --prefer-binary --extra-index-url https://repo.fury.io/mgiessing \
+    RUSTFLAGS='-L /opt/conda/lib' pip install -v --prefer-binary --extra-index-url https://repo.fury.io/mgiessing \
        'cmake>=3.26' ninja packaging 'setuptools-scm>=8' wheel jinja2 \
-        torch==2.3.1 \
        -r requirements-cpu.txt \
        xformers uvloop==0.20.0

@@ -1,174 +1,119 @@
-# Default ROCm 6.2 base image
-ARG BASE_IMAGE="rocm/pytorch:rocm6.2_ubuntu20.04_py3.9_pytorch_release_2.3.0"
+# default base image
+ARG REMOTE_VLLM="0"
+ARG USE_CYTHON="0"
+ARG BUILD_RPD="1"
+ARG COMMON_WORKDIR=/app
+ARG BASE_IMAGE=rocm/vllm-dev:base

-# Default ROCm ARCHes to build vLLM for.
-ARG PYTORCH_ROCM_ARCH="gfx908;gfx90a;gfx942;gfx1100"
+FROM ${BASE_IMAGE} AS base

-# Whether to install CK-based flash-attention
-# If 0, will not install flash-attention
-ARG BUILD_FA="1"
-ARG FA_GFX_ARCHS="gfx90a;gfx942"
-ARG FA_BRANCH="3cea2fb"
-
-# Whether to build triton on rocm
-ARG BUILD_TRITON="1"
-ARG TRITON_BRANCH="e192dba"
-
-### Base image build stage
-FROM $BASE_IMAGE AS base
-
-# Import arg(s) defined before this build stage
-ARG PYTORCH_ROCM_ARCH
+ARG ARG_PYTORCH_ROCM_ARCH
+ENV PYTORCH_ROCM_ARCH=${ARG_PYTORCH_ROCM_ARCH:-${PYTORCH_ROCM_ARCH}}

 # Install some basic utilities
-RUN apt-get update && apt-get install python3 python3-pip -y
-RUN apt-get update && apt-get install -y \
-    curl \
-    ca-certificates \
-    sudo \
-    git \
-    bzip2 \
-    libx11-6 \
-    build-essential \
-    wget \
-    unzip \
-    tmux \
-    ccache \
- && rm -rf /var/lib/apt/lists/*
-
-# When launching the container, mount the code directory to /vllm-workspace
-ARG APP_MOUNT=/vllm-workspace
-WORKDIR ${APP_MOUNT}
-
-RUN python3 -m pip install --upgrade pip
-# Remove sccache so it doesn't interfere with ccache
-# TODO: implement sccache support across components
+RUN apt-get update -q -y && apt-get install -q -y \
+    sqlite3 libsqlite3-dev libfmt-dev libmsgpack-dev libsuitesparse-dev
+# Remove sccache    
+RUN python3 -m pip install --upgrade pip && pip install setuptools_scm
 RUN apt-get purge -y sccache; python3 -m pip uninstall -y sccache; rm -f "$(which sccache)"
-
-# Install torch == 2.6.0 on ROCm
-RUN --mount=type=cache,target=/root/.cache/pip \
-    case "$(ls /opt | grep -Po 'rocm-[0-9]\.[0-9]')" in \
-        *"rocm-6.2"*) \
-            python3 -m pip uninstall -y torch torchvision \
-            && python3 -m pip install --pre \
-                torch==2.6.0.dev20241113+rocm6.2 \
-                'setuptools-scm>=8' \
-                torchvision==0.20.0.dev20241113+rocm6.2 \
-                --extra-index-url https://download.pytorch.org/whl/nightly/rocm6.2;; \
-        *) ;; esac
-
-ENV LLVM_SYMBOLIZER_PATH=/opt/rocm/llvm/bin/llvm-symbolizer
-ENV PATH=$PATH:/opt/rocm/bin:/libtorch/bin:
-ENV LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/opt/rocm/lib/:/libtorch/lib:
-ENV CPLUS_INCLUDE_PATH=$CPLUS_INCLUDE_PATH:/libtorch/include:/libtorch/include/torch/csrc/api/include/:/opt/rocm/include/:
-
-ENV PYTORCH_ROCM_ARCH=${PYTORCH_ROCM_ARCH}
-ENV CCACHE_DIR=/root/.cache/ccache
+ARG COMMON_WORKDIR
+WORKDIR ${COMMON_WORKDIR}


-### AMD-SMI build stage
-FROM base AS build_amdsmi
-# Build amdsmi wheel always
-RUN cd /opt/rocm/share/amd_smi \
-    && python3 -m pip wheel . --wheel-dir=/install
+# -----------------------
+# vLLM fetch stages
+FROM base AS fetch_vllm_0
+ONBUILD COPY ./ vllm/
+FROM base AS fetch_vllm_1
+ARG VLLM_REPO="https://github.com/vllm-project/vllm.git"
+ARG VLLM_BRANCH="main"
+ONBUILD RUN git clone ${VLLM_REPO} \
+	    && cd vllm \
+	    && git checkout ${VLLM_BRANCH}
+FROM fetch_vllm_${REMOTE_VLLM} AS fetch_vllm

+# -----------------------
+# vLLM build stages
+FROM fetch_vllm AS build_vllm
+ARG USE_CYTHON
+# Build vLLM
+RUN cd vllm \
+    && python3 -m pip install -r requirements-rocm.txt \
+    && python3 setup.py clean --all  \
+    && if [ ${USE_CYTHON} -eq "1" ]; then python3 setup_cython.py build_ext --inplace; fi \
+    && python3 setup.py bdist_wheel --dist-dir=dist
+FROM scratch AS export_vllm
+ARG COMMON_WORKDIR
+COPY --from=build_vllm ${COMMON_WORKDIR}/vllm/dist/*.whl /
+COPY --from=build_vllm ${COMMON_WORKDIR}/vllm/requirements*.txt /
+COPY --from=build_vllm ${COMMON_WORKDIR}/vllm/benchmarks /benchmarks
+COPY --from=build_vllm ${COMMON_WORKDIR}/vllm/tests /tests
+COPY --from=build_vllm ${COMMON_WORKDIR}/vllm/examples /examples
+COPY --from=build_vllm ${COMMON_WORKDIR}/vllm/.buildkite /.buildkite

-### Flash-Attention wheel build stage
-FROM base AS build_fa
-ARG BUILD_FA
-ARG FA_GFX_ARCHS
-ARG FA_BRANCH
-# Build ROCm flash-attention wheel if `BUILD_FA = 1`
-RUN --mount=type=cache,target=${CCACHE_DIR} \
-    if [ "$BUILD_FA" = "1" ]; then \
-        mkdir -p libs \
-        && cd libs \
-        && git clone https://github.com/ROCm/flash-attention.git \
-        && cd flash-attention \
-        && git checkout "${FA_BRANCH}" \
-        && git submodule update --init \
-        && GPU_ARCHS="${FA_GFX_ARCHS}" python3 setup.py bdist_wheel --dist-dir=/install; \
-    # Create an empty directory otherwise as later build stages expect one
-    else mkdir -p /install; \
-    fi
+# -----------------------
+# Test vLLM image
+FROM base AS test

+RUN python3 -m pip install --upgrade pip && rm -rf /var/lib/apt/lists/*

-### Triton wheel build stage
-FROM base AS build_triton
-ARG BUILD_TRITON
-ARG TRITON_BRANCH
-# Build triton wheel if `BUILD_TRITON = 1`
-RUN --mount=type=cache,target=${CCACHE_DIR} \
-    if [ "$BUILD_TRITON" = "1" ]; then \
-    mkdir -p libs \
-    && cd libs \
-    && python3 -m pip install ninja cmake wheel pybind11 \
-    && git clone https://github.com/OpenAI/triton.git \
-    && cd triton \
-    && git checkout "${TRITON_BRANCH}" \
-    && cd python \
-    && python3 setup.py bdist_wheel --dist-dir=/install; \
-    # Create an empty directory otherwise as later build stages expect one
-    else mkdir -p /install; \
-    fi
+# Install vLLM
+RUN --mount=type=bind,from=export_vllm,src=/,target=/install \
+    cd /install \
+    && pip install -U -r requirements-rocm.txt \
+    && pip uninstall -y vllm \
+    && pip install *.whl

-
-### Final vLLM build stage
-FROM base AS final
-# Import the vLLM development directory from the build context
-COPY . .
-ARG GIT_REPO_CHECK=0
-RUN --mount=type=bind,source=.git,target=.git \
-    if [ "$GIT_REPO_CHECK" != 0 ]; then bash tools/check_repo.sh ; fi
-
-RUN python3 -m pip install --upgrade pip
-
-# Package upgrades for useful functionality or to avoid dependency issues
-RUN --mount=type=cache,target=/root/.cache/pip \
-    python3 -m pip install --upgrade numba scipy huggingface-hub[cli] pytest-shard
-
-
-# Workaround for ray >= 2.10.0
-ENV RAY_EXPERIMENTAL_NOSET_ROCR_VISIBLE_DEVICES=1
-# Silences the HF Tokenizers warning
-ENV TOKENIZERS_PARALLELISM=false
-
-RUN --mount=type=cache,target=${CCACHE_DIR} \
-    --mount=type=bind,source=.git,target=.git \
-    --mount=type=cache,target=/root/.cache/pip \
-    python3 -m pip install -Ur requirements-rocm.txt \
-    && python3 setup.py clean --all \
-    && python3 setup.py develop
-
-# Copy amdsmi wheel into final image
-RUN --mount=type=bind,from=build_amdsmi,src=/install,target=/install \
-    mkdir -p libs \
-    && cp /install/*.whl libs \
-    # Preemptively uninstall to avoid same-version no-installs
-    && python3 -m pip uninstall -y amdsmi;
-
-# Copy triton wheel(s) into final image if they were built
-RUN --mount=type=bind,from=build_triton,src=/install,target=/install \
-    mkdir -p libs \
-    && if ls /install/*.whl; then \
-        cp /install/*.whl libs \
-        # Preemptively uninstall to avoid same-version no-installs
-        && python3 -m pip uninstall -y triton; fi
-
-# Copy flash-attn wheel(s) into final image if they were built
-RUN --mount=type=bind,from=build_fa,src=/install,target=/install \
-    mkdir -p libs \
-    && if ls /install/*.whl; then \
-        cp /install/*.whl libs \
-        # Preemptively uninstall to avoid same-version no-installs
-        && python3 -m pip uninstall -y flash-attn; fi
-
-# Install wheels that were built to the final image
-RUN --mount=type=cache,target=/root/.cache/pip \
-    if ls libs/*.whl; then \
-    python3 -m pip install libs/*.whl; fi
+WORKDIR /vllm-workspace
+ARG COMMON_WORKDIR
+COPY --from=build_vllm ${COMMON_WORKDIR}/vllm /vllm-workspace

 # install development dependencies (for testing)
-RUN python3 -m pip install -e tests/vllm_test_utils
+RUN cd /vllm-workspace \
+    && rm -rf vllm \
+    && python3 -m pip install -e tests/vllm_test_utils \
+    && python3 -m pip install lm-eval[api]==0.4.4 \
+    && python3 -m pip install pytest-shard
+
+# -----------------------
+# Final vLLM image
+FROM base AS final
+
+RUN python3 -m pip install --upgrade pip && rm -rf /var/lib/apt/lists/*
+# Error related to odd state for numpy 1.20.3 where there is no METADATA etc, but an extra LICENSES_bundled.txt.
+# Manually remove it so that later steps of numpy upgrade can continue
+RUN case "$(which python3)" in \
+        *"/opt/conda/envs/py_3.9"*) \
+            rm -rf /opt/conda/envs/py_3.9/lib/python3.9/site-packages/numpy-1.20.3.dist-info/;; \
+        *) ;; esac
+
+RUN python3 -m pip install --upgrade huggingface-hub[cli]
+ARG BUILD_RPD
+RUN if [ ${BUILD_RPD} -eq "1" ]; then \
+    git clone -b nvtx_enabled https://github.com/ROCm/rocmProfileData.git \
+    && cd rocmProfileData/rpd_tracer \
+    && pip install -r requirements.txt && cd ../ \
+    && make && make install \
+    && cd hipMarker && python3 setup.py install ; fi
+
+# Install vLLM
+RUN --mount=type=bind,from=export_vllm,src=/,target=/install \
+    cd /install \
+    && pip install -U -r requirements-rocm.txt \
+    && pip uninstall -y vllm \
+    && pip install *.whl
+
+ARG COMMON_WORKDIR
+
+# Copy over the benchmark scripts as well
+COPY --from=export_vllm /benchmarks ${COMMON_WORKDIR}/vllm/benchmarks
+COPY --from=export_vllm /examples ${COMMON_WORKDIR}/vllm/examples
+
+ENV RAY_EXPERIMENTAL_NOSET_ROCR_VISIBLE_DEVICES=1
+ENV TOKENIZERS_PARALLELISM=false
+
+# Performance environment variable.
+ENV HIP_FORCE_DEV_KERNARG=1

 CMD ["/bin/bash"]
+
@@ -0,0 +1,158 @@
+ARG BASE_IMAGE=rocm/dev-ubuntu-22.04:6.3.1-complete
+ARG HIPBLASLT_BRANCH="4d40e36"
+ARG HIPBLAS_COMMON_BRANCH="7c1566b"
+ARG LEGACY_HIPBLASLT_OPTION=
+ARG RCCL_BRANCH="648a58d"
+ARG RCCL_REPO="https://github.com/ROCm/rccl"
+ARG TRITON_BRANCH="e5be006"
+ARG TRITON_REPO="https://github.com/triton-lang/triton.git"
+ARG PYTORCH_BRANCH="3a585126"
+ARG PYTORCH_VISION_BRANCH="v0.19.1"
+ARG PYTORCH_REPO="https://github.com/pytorch/pytorch.git"
+ARG PYTORCH_VISION_REPO="https://github.com/pytorch/vision.git"
+ARG FA_BRANCH="b7d29fb"
+ARG FA_REPO="https://github.com/ROCm/flash-attention.git"
+
+FROM ${BASE_IMAGE} AS base
+
+ENV PATH=/opt/rocm/llvm/bin:$PATH
+ENV ROCM_PATH=/opt/rocm
+ENV LD_LIBRARY_PATH=/opt/rocm/lib:/usr/local/lib:
+ARG PYTORCH_ROCM_ARCH=gfx90a;gfx942
+ENV PYTORCH_ROCM_ARCH=${PYTORCH_ROCM_ARCH}
+
+ARG PYTHON_VERSION=3.12
+
+RUN mkdir -p /app
+WORKDIR /app
+ENV DEBIAN_FRONTEND=noninteractive
+
+# Install Python and other dependencies
+RUN apt-get update -y \
+    && apt-get install -y software-properties-common git curl sudo vim less \
+    && add-apt-repository ppa:deadsnakes/ppa \
+    && apt-get update -y \
+    && apt-get install -y python${PYTHON_VERSION} python${PYTHON_VERSION}-dev python${PYTHON_VERSION}-venv \
+       python${PYTHON_VERSION}-lib2to3 python-is-python3  \
+    && update-alternatives --install /usr/bin/python3 python3 /usr/bin/python${PYTHON_VERSION} 1 \
+    && update-alternatives --set python3 /usr/bin/python${PYTHON_VERSION} \
+    && ln -sf /usr/bin/python${PYTHON_VERSION}-config /usr/bin/python3-config \
+    && curl -sS https://bootstrap.pypa.io/get-pip.py | python${PYTHON_VERSION} \
+    && python3 --version && python3 -m pip --version
+
+RUN pip install -U packaging cmake ninja wheel setuptools pybind11 Cython
+
+FROM base AS build_hipblaslt
+ARG HIPBLASLT_BRANCH
+ARG HIPBLAS_COMMON_BRANCH
+# Set to "--legacy_hipblas_direct" for ROCm<=6.2
+ARG LEGACY_HIPBLASLT_OPTION
+RUN git clone https://github.com/ROCm/hipBLAS-common.git
+RUN cd hipBLAS-common \
+    && git checkout ${HIPBLAS_COMMON_BRANCH} \
+    && mkdir build \
+    && cd build \
+    && cmake .. \
+    && make package \
+    && dpkg -i ./*.deb
+RUN git clone https://github.com/ROCm/hipBLASLt
+RUN cd hipBLASLt \
+    && git checkout ${HIPBLASLT_BRANCH} \
+    && ./install.sh -d --architecture ${PYTORCH_ROCM_ARCH} ${LEGACY_HIPBLASLT_OPTION} \
+    && cd build/release \
+    && make package
+RUN mkdir -p /app/install && cp /app/hipBLASLt/build/release/*.deb /app/hipBLAS-common/build/*.deb /app/install
+
+FROM base AS build_rccl
+ARG RCCL_BRANCH
+ARG RCCL_REPO
+RUN git clone ${RCCL_REPO}
+RUN cd rccl \
+    && git checkout ${RCCL_BRANCH} \
+    && ./install.sh -p --amdgpu_targets ${PYTORCH_ROCM_ARCH}
+RUN mkdir -p /app/install && cp /app/rccl/build/release/*.deb /app/install
+
+FROM base AS build_triton
+ARG TRITON_BRANCH
+ARG TRITON_REPO
+RUN git clone ${TRITON_REPO}
+RUN cd triton \
+    && git checkout ${TRITON_BRANCH} \
+    && cd python \
+    && python3 setup.py bdist_wheel --dist-dir=dist
+RUN mkdir -p /app/install && cp /app/triton/python/dist/*.whl /app/install
+
+FROM base AS build_amdsmi
+RUN cd /opt/rocm/share/amd_smi \
+    && pip wheel . --wheel-dir=dist
+RUN mkdir -p /app/install && cp /opt/rocm/share/amd_smi/dist/*.whl /app/install
+
+FROM base AS build_pytorch
+ARG PYTORCH_BRANCH
+ARG PYTORCH_VISION_BRANCH
+ARG PYTORCH_REPO
+ARG PYTORCH_VISION_REPO
+ARG FA_BRANCH
+ARG FA_REPO
+RUN git clone ${PYTORCH_REPO} pytorch
+RUN cd pytorch && git checkout ${PYTORCH_BRANCH} && \
+    pip install -r requirements.txt && git submodule update --init --recursive \
+    && python3 tools/amd_build/build_amd.py \
+    && CMAKE_PREFIX_PATH=$(python3 -c 'import sys; print(sys.prefix)') python3 setup.py bdist_wheel --dist-dir=dist \
+    && pip install dist/*.whl
+RUN git clone ${PYTORCH_VISION_REPO} vision
+RUN cd vision && git checkout ${PYTORCH_VISION_BRANCH} \
+    && python3 setup.py bdist_wheel --dist-dir=dist \
+    && pip install dist/*.whl
+RUN git clone ${FA_REPO}
+RUN cd flash-attention \
+    && git checkout ${FA_BRANCH} \
+    && git submodule update --init \
+    && MAX_JOBS=64 GPU_ARCHS=${PYTORCH_ROCM_ARCH} python3 setup.py bdist_wheel --dist-dir=dist
+RUN mkdir -p /app/install && cp /app/pytorch/dist/*.whl /app/install \
+    && cp /app/vision/dist/*.whl /app/install \
+    && cp /app/flash-attention/dist/*.whl /app/install
+
+FROM base AS final
+RUN --mount=type=bind,from=build_hipblaslt,src=/app/install/,target=/install \
+    dpkg -i /install/*deb \
+    && sed -i 's/, hipblaslt-dev \(.*\), hipcub-dev/, hipcub-dev/g' /var/lib/dpkg/status \
+    && sed -i 's/, hipblaslt \(.*\), hipfft/, hipfft/g' /var/lib/dpkg/status
+RUN --mount=type=bind,from=build_rccl,src=/app/install/,target=/install \
+    dpkg -i /install/*deb \
+    && sed -i 's/, rccl-dev \(.*\), rocalution/, rocalution/g' /var/lib/dpkg/status \
+    && sed -i 's/, rccl \(.*\), rocalution/, rocalution/g' /var/lib/dpkg/status
+RUN --mount=type=bind,from=build_triton,src=/app/install/,target=/install \
+    pip install /install/*.whl
+RUN --mount=type=bind,from=build_amdsmi,src=/app/install/,target=/install \
+    pip install /install/*.whl
+RUN --mount=type=bind,from=build_pytorch,src=/app/install/,target=/install \
+    pip install /install/*.whl
+
+ARG BASE_IMAGE
+ARG HIPBLASLT_BRANCH
+ARG LEGACY_HIPBLASLT_OPTION
+ARG RCCL_BRANCH
+ARG RCCL_REPO
+ARG TRITON_BRANCH
+ARG TRITON_REPO
+ARG PYTORCH_BRANCH
+ARG PYTORCH_VISION_BRANCH
+ARG PYTORCH_REPO
+ARG PYTORCH_VISION_REPO
+ARG FA_BRANCH
+ARG FA_REPO
+RUN echo "BASE_IMAGE: ${BASE_IMAGE}" > /app/versions.txt \
+    && echo "HIPBLAS_COMMON_BRANCH: ${HIPBLAS_COMMON_BRANCH}" >> /app/versions.txt \
+    && echo "HIPBLASLT_BRANCH: ${HIPBLASLT_BRANCH}" >> /app/versions.txt \
+    && echo "LEGACY_HIPBLASLT_OPTION: ${LEGACY_HIPBLASLT_OPTION}" >> /app/versions.txt \
+    && echo "RCCL_BRANCH: ${RCCL_BRANCH}" >> /app/versions.txt \
+    && echo "RCCL_REPO: ${RCCL_REPO}" >> /app/versions.txt \
+    && echo "TRITON_BRANCH: ${TRITON_BRANCH}" >> /app/versions.txt \
+    && echo "TRITON_REPO: ${TRITON_REPO}" >> /app/versions.txt \
+    && echo "PYTORCH_BRANCH: ${PYTORCH_BRANCH}" >> /app/versions.txt \
+    && echo "PYTORCH_VISION_BRANCH: ${PYTORCH_VISION_BRANCH}" >> /app/versions.txt \
+    && echo "PYTORCH_REPO: ${PYTORCH_REPO}" >> /app/versions.txt \
+    && echo "PYTORCH_VISION_REPO: ${PYTORCH_VISION_REPO}" >> /app/versions.txt \
+    && echo "FA_BRANCH: ${FA_BRANCH}" >> /app/versions.txt \
+    && echo "FA_REPO: ${FA_REPO}" >> /app/versions.txt
@@ -1,4 +1,4 @@
-ARG NIGHTLY_DATE="20241017"
+ARG NIGHTLY_DATE="20250124"
 ARG BASE_IMAGE="us-central1-docker.pkg.dev/tpu-pytorch-releases/docker/xla:nightly_3.10_tpuvm_$NIGHTLY_DATE"

 FROM $BASE_IMAGE
@@ -10,12 +10,23 @@ Easy, fast, and cheap LLM serving for everyone
 </h3>

 <p align="center">
-| <a href="https://docs.vllm.ai"><b>Documentation</b></a> | <a href="https://vllm.ai"><b>Blog</b></a> | <a href="https://arxiv.org/abs/2309.06180"><b>Paper</b></a> | <a href="https://discord.gg/jz7wjKhh6g"><b>Discord</b></a> | <a href="https://x.com/vllm_project"><b>Twitter/X</b></a> | <a href="https://slack.vllm.ai"><b>Developer Slack</b></a> |
+| <a href="https://docs.vllm.ai"><b>Documentation</b></a> | <a href="https://vllm.ai"><b>Blog</b></a> | <a href="https://arxiv.org/abs/2309.06180"><b>Paper</b></a> | <a href="https://x.com/vllm_project"><b>Twitter/X</b></a> | <a href="https://slack.vllm.ai"><b>Developer Slack</b></a> |
 </p>

 ---

+We’re excited to invite you to the first **vLLM China Meetup** on **March 16** in **Beijing**!  
+
+Join us to connect with the **vLLM team** and explore how vLLM is leveraged in **post-training, fine-tuning, and deployment**, including [verl](https://github.com/volcengine/verl), [LLaMA-Factory](https://github.com/hiyouga/LLaMA-Factory), and [vllm-ascend](https://github.com/vllm-project/vllm-ascend).
+
+👉 **[Register Now](https://mp.weixin.qq.com/s/n77GibL2corAtQHtVEAzfg)** to be part of the discussion!  
+
+---
+
 *Latest News* 🔥
+
+- [2025/01] We are excited to announce the alpha release of vLLM V1: A major architectural upgrade with 1.7x speedup! Clean code, optimized execution loop, zero-overhead prefix caching, enhanced multimodal support, and more. Please check out our blog post [here](https://blog.vllm.ai/2025/01/27/v1-alpha-release.html).
+- [2025/01] We hosted [the eighth vLLM meetup](https://lu.ma/zep56hui) with Google Cloud! Please find the meetup slides from vLLM team [here](https://docs.google.com/presentation/d/1epVkt4Zu8Jz_S5OhEHPc798emsYh2BwYfRuDDVEF7u4/edit?usp=sharing), and Google Cloud team [here](https://drive.google.com/file/d/1h24pHewANyRL11xy5dXUbvRC9F9Kkjix/view?usp=sharing).
 - [2024/12] vLLM joins [pytorch ecosystem](https://pytorch.org/blog/vllm-joins-pytorch)! Easy, Fast, and Cheap LLM Serving for Everyone!
 - [2024/11] We hosted [the seventh vLLM meetup](https://lu.ma/h0qvrajz) with Snowflake! Please find the meetup slides from vLLM team [here](https://docs.google.com/presentation/d/1e3CxQBV3JsfGp30SwyvS3eM_tW-ghOhJ9PAJGK6KR54/edit?usp=sharing), and Snowflake team [here](https://docs.google.com/presentation/d/1qF3RkDAbOULwz9WK5TOltt2fE9t6uIc_hVNLFAaQX6A/edit?usp=sharing).
 - [2024/10] We have just created a developer slack ([slack.vllm.ai](https://slack.vllm.ai)) focusing on coordinating contributions and discussing features. Please feel free to join us there!
@@ -31,13 +42,17 @@ Easy, fast, and cheap LLM serving for everyone
 - [2023/06] We officially released vLLM! FastChat-vLLM integration has powered [LMSYS Vicuna and Chatbot Arena](https://chat.lmsys.org) since mid-April. Check out our [blog post](https://vllm.ai).

 ---
+
 ## About
+
 vLLM is a fast and easy-to-use library for LLM inference and serving.

+Originally developed in the [Sky Computing Lab](https://sky.cs.berkeley.edu) at UC Berkeley, vLLM has evolved into a community-driven project with contributions from both academia and industry.
+
 vLLM is fast with:

 - State-of-the-art serving throughput
- Efficient management of attention key and value memory with **PagedAttention**
+- Efficient management of attention key and value memory with [**PagedAttention**](https://blog.vllm.ai/2023/06/20/vllm.html)
 - Continuous batching of incoming requests
 - Fast model execution with CUDA/HIP graph
 - Quantizations: [GPTQ](https://arxiv.org/abs/2210.17323), [AWQ](https://arxiv.org/abs/2306.00978), INT4, INT8, and FP8.
@@ -60,7 +75,7 @@ vLLM is flexible and easy to use with:

 vLLM seamlessly supports most popular open-source models on HuggingFace, including:
 - Transformer-like LLMs (e.g., Llama)
- Mixture-of-Expert LLMs (e.g., Mixtral)
+- Mixture-of-Expert LLMs (e.g., Mixtral, Deepseek-V2 and V3)
 - Embedding Models (e.g. E5-Mistral)
 - Multi-modal LLMs (e.g., LLaVA)

@@ -68,16 +83,16 @@ Find the full list of supported models [here](https://docs.vllm.ai/en/latest/mod

 ## Getting Started

-Install vLLM with `pip` or [from source](https://vllm.readthedocs.io/en/latest/getting_started/installation.html#build-from-source):
+Install vLLM with `pip` or [from source](https://docs.vllm.ai/en/latest/getting_started/installation/gpu/index.html#build-wheel-from-source):

 ```bash
 pip install vllm
 ```

-Visit our [documentation](https://vllm.readthedocs.io/en/latest/) to learn more.
- [Installation](https://vllm.readthedocs.io/en/latest/getting_started/installation.html)
- [Quickstart](https://vllm.readthedocs.io/en/latest/getting_started/quickstart.html)
- [Supported Models](https://vllm.readthedocs.io/en/latest/models/supported_models.html)
+Visit our [documentation](https://docs.vllm.ai/en/latest/) to learn more.
+- [Installation](https://docs.vllm.ai/en/latest/getting_started/installation/index.html)
+- [Quickstart](https://docs.vllm.ai/en/latest/getting_started/quickstart.html)
+- [List of Supported Models](https://docs.vllm.ai/en/latest/models/supported_models.html)

 ## Contributing

@@ -90,34 +105,40 @@ vLLM is a community project. Our compute resources for development and testing a

 <!-- Note: Please sort them in alphabetical order. -->
 <!-- Note: Please keep these consistent with docs/source/community/sponsors.md -->
-
+Cash Donations:
 - a16z
+- Dropbox
+- Sequoia Capital
+- Skywork AI
+- ZhenFund
+
+Compute Resources:
 - AMD
 - Anyscale
 - AWS
 - Crusoe Cloud
 - Databricks
 - DeepInfra
- Dropbox
 - Google Cloud
 - Lambda Lab
 - Nebius
+- Novita AI
 - NVIDIA
 - Replicate
 - Roblox
 - RunPod
- Sequoia Capital
- Skywork AI
 - Trainy
 - UC Berkeley
 - UC San Diego
- ZhenFund
+
+Slack Sponsor: Anyscale

 We also have an official fundraising venue through [OpenCollective](https://opencollective.com/vllm). We plan to use the fund to support the development, maintenance, and adoption of vLLM.

 ## Citation

 If you use vLLM for your research, please cite our [paper](https://arxiv.org/abs/2309.06180):
+
 ```bibtex
@inproceedings{kwon2023efficient,
  title={Efficient Memory Management for Large Language Model Serving with PagedAttention},
@@ -129,12 +150,11 @@ If you use vLLM for your research, please cite our [paper](https://arxiv.org/abs

 ## Contact Us

-* For technical questions and feature requests, please use Github issues or discussions.
-* For discussing with fellow users, please use Discord.
-* For coordinating contributions and development, please use Slack.
-* For security disclosures, please use Github's security advisory feature.
-* For collaborations and partnerships, please contact us at vllm-questions AT lists.berkeley.edu.
+- For technical questions and feature requests, please use Github issues or discussions.
+- For discussing with fellow users and coordinating contributions and development, please use Slack.
+- For security disclosures, please use Github's security advisory feature.
+- For collaborations and partnerships, please contact us at vllm-questions AT lists.berkeley.edu.

 ## Media Kit

-* If you wish to use vLLM's logo, please refer to [our media kit repo](https://github.com/vllm-project/media-kit).
+- If you wish to use vLLM's logo, please refer to [our media kit repo](https://github.com/vllm-project/media-kit).
@@ -0,0 +1,54 @@
+# Releasing vLLM
+
+vLLM releases offer a reliable version of the code base, packaged into a binary format that can be conveniently accessed via PyPI. These releases also serve as key milestones for the development team to communicate with the community about newly available features, improvements, and upcoming changes that could affect users, including potential breaking changes.
+
+## Release Versioning
+
+vLLM uses a “right-shifted” versioning scheme where a new patch release is out every 2 weeks. And patch releases contain features and bug fixes (as opposed to semver where patch release contains only backwards-compatible bug fixes). When critical fixes need to be made, special release post1 is released.
+
+* _major_ major architectural milestone and when incompatible API changes are made, similar to PyTorch 2.0.
+* _minor_ major features
+* _patch_ features and backwards-compatible bug fixes
+* _post1_ or _patch-1_ backwards-compatible bug fixes, either explicit or implicit post release
+
+## Release Cadence
+
+Patch release is released on bi-weekly basis. Post release 1-3 days after patch release and uses same branch as patch release.
+Following is the release cadence for year 2025. All future release dates below are tentative. Please note: Post releases are optional.
+
+| Release Date | Patch release versions | Post Release versions |
+| --- | --- | --- |
+| Jan 2025 | 0.7.0 | --- |
+| Feb 2025 | 0.7.1, 0.7.2, 0.7.3  | --- |
+| Mar 2025 | 0.7.4, 0.7.5 | --- |
+| Apr 2025 | 0.7.6, 0.7.7 | --- |
+| May 2025 | 0.7.8, 0.7.9 | --- |
+| Jun 2025 | 0.7.10, 0.7.11 | --- |
+| Jul 2025 | 0.7.12, 0.7.13 | --- |
+| Aug 2025 | 0.7.14, 0.7.15 | --- |
+| Sep 2025 | 0.7.16, 0.7.17 | --- |
+| Oct 2025 | 0.7.18, 0.7.19 | --- |
+| Nov 2025 | 0.7.20, 0.7.21 | --- |
+| Dec 2025 | 0.7.22, 0.7.23 | --- |
+
+## Release branch
+
+Each release is built from a dedicated release branch.
+
+* For _major_, _minor_, _patch_ releases, the release branch cut is performed 1-2 days before release is live.
+* For post releases, previously cut release branch is reused
+* Release builds are triggered via push to RC tag like vX.Y.Z-rc1 . This enables us to build and test multiple RCs for each release.
+* Final tag : vX.Y.Z does not trigger the build but used for Release notes and assets.
+* After branch cut is created we monitor the main branch for any reverts and apply these reverts to a release branch.
+
+## Release Cherry-Pick Criteria
+
+After branch cut, we approach finalizing the release branch with clear criteria on what cherry picks are allowed in. Note: a cherry pick is a process to land a PR in the release branch after branch cut. These are typically limited to ensure that the team has sufficient time to complete a thorough round of testing on a stable code base.
+
+* Regression fixes - that address functional/performance regression against the most recent release (e.g. 0.7.0 for 0.7.1 release)
+* Critical fixes - critical fixes for severe issue such as silent incorrectness, backwards compatibility, crashes, deadlocks, (large) memory leaks
+* Fixes to new features introduced in the most recent release (e.g. 0.7.0 for 0.7.1 release)
+* Documentation improvements
+* 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.
@@ -4,7 +4,7 @@

 If you believe you have found a security vulnerability in vLLM, we encourage you to let us know right away. We will investigate all legitimate reports and do our best to quickly fix the problem.

-Please report security issues privately using [the vulnerability submission form](https://github.com/vllm-project/vllm/security/advisories/new).
+Please report security issues privately using [the vulnerability submission form](https://github.com/vllm-project/vllm/security/advisories/new). Reports will then be triaged by the [vulnerability management team](https://docs.vllm.ai/en/latest/contributing/vulnerability_management.html).

 ---

@@ -3,6 +3,7 @@
 ## Downloading the ShareGPT dataset

 You can download the dataset by running:
+
 ```bash
 wget https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/resolve/main/ShareGPT_V3_unfiltered_cleaned_split.json
 ```
@@ -11,9 +12,18 @@ wget https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/r

 The json file refers to several image datasets (coco, llava, etc.). The benchmark scripts
 will ignore a datapoint if the referred image is missing.
+
 ```bash
 wget https://huggingface.co/datasets/Lin-Chen/ShareGPT4V/resolve/main/sharegpt4v_instruct_gpt4-vision_cap100k.json
 mkdir coco -p
 wget http://images.cocodataset.org/zips/train2017.zip -O coco/train2017.zip
 unzip coco/train2017.zip -d coco/
 ```
+
+# Downloading the BurstGPT dataset
+
+You can download the BurstGPT v1.1 dataset by running:
+
+```bash
+wget https://github.com/HPMLL/BurstGPT/releases/download/v1.1/BurstGPT_without_fails_2.csv
+```
@@ -1,10 +1,12 @@
+# SPDX-License-Identifier: Apache-2.0
+
 import json
 import os
 import sys
 import time
 import traceback
 from dataclasses import dataclass, field
-from typing import List, Optional, Union
+from typing import Optional, Union

 import aiohttp
 import huggingface_hub.constants
@@ -12,6 +14,8 @@ from tqdm.asyncio import tqdm
 from transformers import (AutoTokenizer, PreTrainedTokenizer,
                          PreTrainedTokenizerFast)

+from vllm.model_executor.model_loader.weight_utils import get_lock
+
 AIOHTTP_TIMEOUT = aiohttp.ClientTimeout(total=6 * 60 * 60)


@@ -22,6 +26,7 @@ class RequestFuncInput:
    prompt_len: int
    output_len: int
    model: str
+    model_name: Optional[str] = None
    best_of: int = 1
    logprobs: Optional[int] = None
    extra_body: Optional[dict] = None
@@ -34,9 +39,10 @@ class RequestFuncOutput:
    generated_text: str = ""
    success: bool = False
    latency: float = 0.0
+    output_tokens: int = 0
    ttft: float = 0.0  # Time to first token
-    itl: List[float] = field(
-        default_factory=list)  # List of inter-token latencies
+    itl: list[float] = field(
+        default_factory=list)  # list of inter-token latencies
    tpot: float = 0.0  # avg next-token latencies
    prompt_len: int = 0
    error: str = ""
@@ -49,7 +55,8 @@ async def async_request_tgi(
    api_url = request_func_input.api_url
    assert api_url.endswith("generate_stream")

-    async with aiohttp.ClientSession(timeout=AIOHTTP_TIMEOUT) as session:
+    async with aiohttp.ClientSession(trust_env=True,
+                                     timeout=AIOHTTP_TIMEOUT) as session:
        params = {
            "best_of": request_func_input.best_of,
            "max_new_tokens": request_func_input.output_len,
@@ -78,7 +85,7 @@ async def async_request_tgi(
                            continue
                        chunk_bytes = chunk_bytes.decode("utf-8")

-                        #NOTE: Sometimes TGI returns a ping response without
+                        # NOTE: Sometimes TGI returns a ping response without
                        # any data, we should skip it.
                        if chunk_bytes.startswith(":"):
                            continue
@@ -121,7 +128,8 @@ async def async_request_trt_llm(
    api_url = request_func_input.api_url
    assert api_url.endswith("generate_stream")

-    async with aiohttp.ClientSession(timeout=AIOHTTP_TIMEOUT) as session:
+    async with aiohttp.ClientSession(trust_env=True,
+                                     timeout=AIOHTTP_TIMEOUT) as session:
        assert request_func_input.best_of == 1
        payload = {
            "accumulate_tokens": True,
@@ -155,7 +163,7 @@ async def async_request_trt_llm(
                        timestamp = time.perf_counter()
                        # First token
                        if ttft == 0.0:
-                            ttft = time.perf_counter() - st
+                            ttft = timestamp - st
                            output.ttft = ttft

                        # Decoding phase
@@ -185,7 +193,8 @@ async def async_request_deepspeed_mii(
    request_func_input: RequestFuncInput,
    pbar: Optional[tqdm] = None,
 ) -> RequestFuncOutput:
-    async with aiohttp.ClientSession(timeout=AIOHTTP_TIMEOUT) as session:
+    async with aiohttp.ClientSession(trust_env=True,
+                                     timeout=AIOHTTP_TIMEOUT) as session:
        assert request_func_input.best_of == 1

        payload = {
@@ -233,17 +242,23 @@ async def async_request_openai_completions(
        ("completions", "profile")
    ), "OpenAI Completions API URL must end with 'completions' or 'profile'."

-    async with aiohttp.ClientSession(timeout=AIOHTTP_TIMEOUT) as session:
+    async with aiohttp.ClientSession(trust_env=True,
+                                     timeout=AIOHTTP_TIMEOUT) as session:
        payload = {
-            "model": request_func_input.model,
+            "model": request_func_input.model_name \
+                if request_func_input.model_name else request_func_input.model,
            "prompt": request_func_input.prompt,
            "temperature": 0.0,
            "best_of": request_func_input.best_of,
            "max_tokens": request_func_input.output_len,
            "logprobs": request_func_input.logprobs,
            "stream": True,
-            "ignore_eos": request_func_input.ignore_eos,
+            "stream_options": {
+                "include_usage": True,
+            },
        }
+        if request_func_input.ignore_eos:
+            payload["ignore_eos"] = request_func_input.ignore_eos
        if request_func_input.extra_body:
            payload.update(request_func_input.extra_body)
        headers = {
@@ -254,7 +269,6 @@ async def async_request_openai_completions(
        output.prompt_len = request_func_input.prompt_len

        generated_text = ""
-        ttft = 0.0
        st = time.perf_counter()
        most_recent_timestamp = st
        try:
@@ -269,15 +283,16 @@ async def async_request_openai_completions(

                        chunk = chunk_bytes.decode("utf-8").removeprefix(
                            "data: ")
-                        if chunk == "[DONE]":
-                            latency = time.perf_counter() - st
-                        else:
+                        if chunk != "[DONE]":
                            data = json.loads(chunk)

                            # NOTE: Some completion API might have a last
                            # usage summary response without a token so we
                            # want to check a token was generated
-                            if data["choices"][0]["text"]:
+                            if choices := data.get("choices"):
+                                # Note that text could be empty here
+                                # e.g. for special tokens
+                                text = choices[0].get("text")
                                timestamp = time.perf_counter()
                                # First token
                                if not first_chunk_received:
@@ -291,7 +306,10 @@ async def async_request_openai_completions(
                                                      most_recent_timestamp)

                                most_recent_timestamp = timestamp
-                                generated_text += data["choices"][0]["text"]
+                                generated_text += text or ""
+                            elif usage := data.get("usage"):
+                                output.output_tokens = usage.get(
+                                    "completion_tokens")
                    if first_chunk_received:
                        output.success = True
                    else:
@@ -300,7 +318,7 @@ async def async_request_openai_completions(
                            "Never received a valid chunk to calculate TTFT."
                            "This response will be marked as failed!")
                    output.generated_text = generated_text
-                    output.latency = latency
+                    output.latency = most_recent_timestamp - st
                else:
                    output.error = response.reason or ""
                    output.success = False
@@ -323,12 +341,14 @@ async def async_request_openai_chat_completions(
        "chat/completions"
    ), "OpenAI Chat Completions API URL must end with 'chat/completions'."

-    async with aiohttp.ClientSession(timeout=AIOHTTP_TIMEOUT) as session:
+    async with aiohttp.ClientSession(trust_env=True,
+                                     timeout=AIOHTTP_TIMEOUT) as session:
        content = [{"type": "text", "text": request_func_input.prompt}]
        if request_func_input.multi_modal_content:
            content.append(request_func_input.multi_modal_content)
        payload = {
-            "model": request_func_input.model,
+            "model": request_func_input.model_name \
+                if request_func_input.model_name else request_func_input.model,
            "messages": [
                {
                    "role": "user",
@@ -338,8 +358,12 @@ async def async_request_openai_chat_completions(
            "temperature": 0.0,
            "max_completion_tokens": request_func_input.output_len,
            "stream": True,
-            "ignore_eos": request_func_input.ignore_eos,
+            "stream_options": {
+                "include_usage": True,
+            },
        }
+        if request_func_input.ignore_eos:
+            payload["ignore_eos"] = request_func_input.ignore_eos
        if request_func_input.extra_body:
            payload.update(request_func_input.extra_body)
        headers = {
@@ -365,17 +389,15 @@ async def async_request_openai_chat_completions(

                        chunk = chunk_bytes.decode("utf-8").removeprefix(
                            "data: ")
-                        if chunk == "[DONE]":
-                            latency = time.perf_counter() - st
-                        else:
+                        if chunk != "[DONE]":
                            timestamp = time.perf_counter()
                            data = json.loads(chunk)

-                            delta = data["choices"][0]["delta"]
-                            if delta.get("content", None):
+                            if choices := data.get("choices"):
+                                content = choices[0]["delta"].get("content")
                                # First token
                                if ttft == 0.0:
-                                    ttft = time.perf_counter() - st
+                                    ttft = timestamp - st
                                    output.ttft = ttft

                                # Decoding phase
@@ -383,13 +405,16 @@ async def async_request_openai_chat_completions(
                                    output.itl.append(timestamp -
                                                      most_recent_timestamp)

-                                generated_text += delta["content"]
+                                generated_text += content or ""
+                            elif usage := data.get("usage"):
+                                output.output_tokens = usage.get(
+                                    "completion_tokens")

                            most_recent_timestamp = timestamp

                    output.generated_text = generated_text
                    output.success = True
-                    output.latency = latency
+                    output.latency = most_recent_timestamp - st
                else:
                    output.error = response.reason or ""
                    output.success = False
@@ -407,24 +432,48 @@ def get_model(pretrained_model_name_or_path: str) -> str:
    if os.getenv('VLLM_USE_MODELSCOPE', 'False').lower() == 'true':
        from modelscope import snapshot_download

-        model_path = snapshot_download(
-            model_id=pretrained_model_name_or_path,
-            local_files_only=huggingface_hub.constants.HF_HUB_OFFLINE,
-            ignore_file_pattern=[".*.pt", ".*.safetensors", ".*.bin"])
+        # Use file lock to prevent multiple processes from
+        # downloading the same model weights at the same time.
+        with get_lock(pretrained_model_name_or_path):
+            model_path = snapshot_download(
+                model_id=pretrained_model_name_or_path,
+                local_files_only=huggingface_hub.constants.HF_HUB_OFFLINE,
+                ignore_file_pattern=[".*.pt", ".*.safetensors", ".*.bin"])

-        return model_path
+            return model_path
    return pretrained_model_name_or_path


 def get_tokenizer(
-    pretrained_model_name_or_path: str, trust_remote_code: bool
+    pretrained_model_name_or_path: str,
+    tokenizer_mode: str = "auto",
+    trust_remote_code: bool = False,
+    **kwargs,
 ) -> Union[PreTrainedTokenizer, PreTrainedTokenizerFast]:
    if pretrained_model_name_or_path is not None and not os.path.exists(
            pretrained_model_name_or_path):
        pretrained_model_name_or_path = get_model(
            pretrained_model_name_or_path)
-    return AutoTokenizer.from_pretrained(pretrained_model_name_or_path,
-                                         trust_remote_code=trust_remote_code)
+    if tokenizer_mode == "slow":
+        if kwargs.get("use_fast", False):
+            raise ValueError(
+                "Cannot use the fast tokenizer in slow tokenizer mode.")
+        kwargs["use_fast"] = False
+    if tokenizer_mode == "mistral":
+        try:
+            from vllm.transformers_utils.tokenizer import MistralTokenizer
+        except ImportError as e:
+            raise ImportError("MistralTokenizer requires vllm package.\n"
+                              "Please install it with `pip install vllm` "
+                              "to use mistral tokenizer mode.") from e
+        return MistralTokenizer.from_pretrained(
+            str(pretrained_model_name_or_path))
+    else:
+        return AutoTokenizer.from_pretrained(
+            pretrained_model_name_or_path,
+            trust_remote_code=trust_remote_code,
+            **kwargs,
+        )


 ASYNC_REQUEST_FUNCS = {
@@ -1,3 +1,4 @@
+# SPDX-License-Identifier: Apache-2.0
 """Benchmark guided decoding throughput."""
 import argparse
 import dataclasses
@@ -5,7 +6,6 @@ import json
 import os
 import random
 import time
-from typing import List

 import datasets
 import pandas as pd
@@ -38,17 +38,23 @@ class SampleRequest:
    completion: str = None


-def run_vllm(requests: List[SampleRequest],
+def run_vllm(requests: list[SampleRequest],
             engine_args: EngineArgs,
             n: int,
             guided_decoding_rate: float = 1.0,
             warmup: bool = False) -> float:
    from vllm import LLM, SamplingParams
    llm = LLM(**vars(engine_args))
+    assert all(
+        llm.llm_engine.model_config.max_model_len >= (
+            request.prompt_len + request.expected_output_len)
+        for request in requests), (
+            "Please ensure that max_model_len is greater than the sum of"
+            " prompt_len and expected_output_len for all requests.")

    # Add the requests to the engine.
-    prompts: List[str] = []
-    sampling_params: List[SamplingParams] = []
+    prompts: list[str] = []
+    sampling_params: list[SamplingParams] = []
    # create a list containing random selected true or false
    guided_decoding_req_idx = random.sample(
        range(len(requests)), int(len(requests) * guided_decoding_rate))
@@ -103,7 +109,7 @@ def run_vllm(requests: List[SampleRequest],


 async def run_vllm_async(
-        requests: List[SampleRequest],
+        requests: list[SampleRequest],
        engine_args: AsyncEngineArgs,
        n: int,
        guided_decoding_rate: float = 1.0,
@@ -114,9 +120,16 @@ async def run_vllm_async(
    async with build_async_engine_client_from_engine_args(
            engine_args, disable_frontend_multiprocessing) as llm:

+        assert all(
+            llm.model_config.max_model_len >= (request.prompt_len +
+                                               request.expected_output_len)
+            for request in requests), (
+                "Please ensure that max_model_len is greater than the sum of"
+                " prompt_len and expected_output_len for all requests.")
+
        # Add the requests to the engine.
-        prompts: List[str] = []
-        sampling_params: List[SamplingParams] = []
+        prompts: list[str] = []
+        sampling_params: list[SamplingParams] = []
        guided_decoding_req_idx = random.sample(
            range(len(requests)), int(len(requests) * guided_decoding_rate))

@@ -189,7 +202,7 @@ async def run_vllm_async(


 def sample_requests(tokenizer: PreTrainedTokenizerBase,
-                    args: argparse.Namespace) -> List[SampleRequest]:
+                    args: argparse.Namespace) -> list[SampleRequest]:
    if args.dataset == 'json':
        if args.json_schema_path is None:
            dir_path = os.path.dirname(os.path.realpath(__file__))
@@ -273,7 +286,7 @@ def sample_requests(tokenizer: PreTrainedTokenizerBase,

    elif args.dataset == "xgrammar_bench":
        args.warmup = False
-        requests: List[SampleRequest] = []
+        requests: list[SampleRequest] = []
        dataset = datasets.load_dataset("NousResearch/json-mode-eval",
                                        split="train")
        print(f"dataset has {len(dataset)} entries")
@@ -1,21 +1,38 @@
+# SPDX-License-Identifier: Apache-2.0
 """Benchmark the latency of processing a single batch of requests."""
+
 import argparse
 import dataclasses
 import json
+import os
 import time
 from pathlib import Path
-from typing import List, Optional
+from typing import Any, Optional

 import numpy as np
 import torch
+from benchmark_utils import convert_to_pytorch_benchmark_format, write_to_json
 from tqdm import tqdm

 from vllm import LLM, SamplingParams
 from vllm.engine.arg_utils import EngineArgs
 from vllm.inputs import PromptType
+from vllm.sampling_params import BeamSearchParams
 from vllm.utils import FlexibleArgumentParser


+def save_to_pytorch_benchmark_format(args: argparse.Namespace,
+                                     results: dict[str, Any]) -> None:
+    pt_records = convert_to_pytorch_benchmark_format(
+        args=args,
+        metrics={"latency": results["latencies"]},
+        extra_info={k: results[k]
+                    for k in ["avg_latency", "percentiles"]})
+    if pt_records:
+        pt_file = f"{os.path.splitext(args.output_json)[0]}.pytorch.json"
+        write_to_json(pt_file, pt_records)
+
+
 def main(args: argparse.Namespace):
    print(args)

@@ -24,6 +41,10 @@ def main(args: argparse.Namespace):
    # NOTE(woosuk): If the request cannot be processed in a single batch,
    # the engine will automatically process the request in multiple batches.
    llm = LLM(**dataclasses.asdict(engine_args))
+    assert llm.llm_engine.model_config.max_model_len >= (
+        args.input_len +
+        args.output_len), ("Please ensure that max_model_len is greater than"
+                           " the sum of input_len and output_len.")

    sampling_params = SamplingParams(
        n=args.n,
@@ -36,10 +57,25 @@ def main(args: argparse.Namespace):
    dummy_prompt_token_ids = np.random.randint(10000,
                                               size=(args.batch_size,
                                                     args.input_len))
-    dummy_prompts: List[PromptType] = [{
+    dummy_prompts: list[PromptType] = [{
        "prompt_token_ids": batch
    } for batch in dummy_prompt_token_ids.tolist()]

+    def llm_generate():
+        if not args.use_beam_search:
+            llm.generate(dummy_prompts,
+                         sampling_params=sampling_params,
+                         use_tqdm=False)
+        else:
+            llm.beam_search(
+                dummy_prompts,
+                BeamSearchParams(
+                    beam_width=args.n,
+                    max_tokens=args.output_len,
+                    ignore_eos=True,
+                ),
+            )
+
    def run_to_completion(profile_dir: Optional[str] = None):
        if profile_dir:
            with torch.profiler.profile(
@@ -48,16 +84,13 @@ def main(args: argparse.Namespace):
                        torch.profiler.ProfilerActivity.CUDA,
                    ],
                    on_trace_ready=torch.profiler.tensorboard_trace_handler(
-                        str(profile_dir))) as p:
-                llm.generate(dummy_prompts,
-                             sampling_params=sampling_params,
-                             use_tqdm=False)
-            print(p.key_averages())
+                        str(profile_dir)),
+            ) as p:
+                llm_generate()
+            print(p.key_averages().table(sort_by="self_cuda_time_total"))
        else:
            start_time = time.perf_counter()
-            llm.generate(dummy_prompts,
-                         sampling_params=sampling_params,
-                         use_tqdm=False)
+            llm_generate()
            end_time = time.perf_counter()
            latency = end_time - start_time
            return latency
@@ -69,9 +102,8 @@ def main(args: argparse.Namespace):
    if args.profile:
        profile_dir = args.profile_result_dir
        if not profile_dir:
-            profile_dir = Path(
-                "."
-            ) / "vllm_benchmark_result" / f"latency_result_{time.time()}"
+            profile_dir = (Path(".") / "vllm_benchmark_result" /
+                           f"latency_result_{time.time()}")
        print(f"Profiling (results will be saved to '{profile_dir}')...")
        run_to_completion(profile_dir=profile_dir)
        return
@@ -83,9 +115,9 @@ def main(args: argparse.Namespace):
    latencies = np.array(latencies)
    percentages = [10, 25, 50, 75, 90, 99]
    percentiles = np.percentile(latencies, percentages)
-    print(f'Avg latency: {np.mean(latencies)} seconds')
+    print(f"Avg latency: {np.mean(latencies)} seconds")
    for percentage, percentile in zip(percentages, percentiles):
-        print(f'{percentage}% percentile latency: {percentile} seconds')
+        print(f"{percentage}% percentile latency: {percentile} seconds")

    # Output JSON results if specified
    if args.output_json:
@@ -96,43 +128,51 @@ def main(args: argparse.Namespace):
        }
        with open(args.output_json, "w") as f:
            json.dump(results, f, indent=4)
+        save_to_pytorch_benchmark_format(args, results)


-if __name__ == '__main__':
+if __name__ == "__main__":
    parser = FlexibleArgumentParser(
-        description='Benchmark the latency of processing a single batch of '
-        'requests till completion.')
-    parser.add_argument('--input-len', type=int, default=32)
-    parser.add_argument('--output-len', type=int, default=128)
-    parser.add_argument('--batch-size', type=int, default=8)
-    parser.add_argument('--n',
-                        type=int,
-                        default=1,
-                        help='Number of generated sequences per prompt.')
-    parser.add_argument('--use-beam-search', action='store_true')
-    parser.add_argument('--num-iters-warmup',
-                        type=int,
-                        default=10,
-                        help='Number of iterations to run for warmup.')
-    parser.add_argument('--num-iters',
+        description="Benchmark the latency of processing a single batch of "
+        "requests till completion.")
+    parser.add_argument("--input-len", type=int, default=32)
+    parser.add_argument("--output-len", type=int, default=128)
+    parser.add_argument("--batch-size", type=int, default=8)
+    parser.add_argument(
+        "--n",
+        type=int,
+        default=1,
+        help="Number of generated sequences per prompt.",
+    )
+    parser.add_argument("--use-beam-search", action="store_true")
+    parser.add_argument(
+        "--num-iters-warmup",
+        type=int,
+        default=10,
+        help="Number of iterations to run for warmup.",
+    )
+    parser.add_argument("--num-iters",
                        type=int,
                        default=30,
-                        help='Number of iterations to run.')
+                        help="Number of iterations to run.")
    parser.add_argument(
-        '--profile',
-        action='store_true',
-        help='profile the generation process of a single batch')
+        "--profile",
+        action="store_true",
+        help="profile the generation process of a single batch",
+    )
    parser.add_argument(
-        '--profile-result-dir',
+        "--profile-result-dir",
        type=str,
        default=None,
-        help=('path to save the pytorch profiler output. Can be visualized '
-              'with ui.perfetto.dev or Tensorboard.'))
+        help=("path to save the pytorch profiler output. Can be visualized "
+              "with ui.perfetto.dev or Tensorboard."),
+    )
    parser.add_argument(
-        '--output-json',
+        "--output-json",
        type=str,
        default=None,
-        help='Path to save the latency results in JSON format.')
+        help="Path to save the latency results in JSON format.",
+    )

    parser = EngineArgs.add_cli_args(parser)
    args = parser.parse_args()
@@ -0,0 +1,184 @@
+# SPDX-License-Identifier: Apache-2.0
+"""
+Offline benchmark to test the long document QA throughput.
+
+Example usage:
+    # This workload samples 8 different prompts with a default input
+    # length of 20000 tokens, then replicates each prompt 2 times 
+    # in random order.
+    python benchmark_long_document_qa_throughput.py \
+        --model meta-llama/Llama-2-7b-chat-hf \
+        --enable-prefix-caching \
+        --num-documents 8 \
+        --repeat-count 2 
+
+Commandline arguments:
+    --num-documents: The number of documents to sample prompts from.
+
+    --document-length: The length of each document in tokens. 
+                       (Optional, default: 20000)
+
+    --output-len: The number of tokens to generate for each prompt.
+                  (Optional, default: 10)
+
+    --repeat-count: The number of times to repeat each prompt.
+                    (Optional, default: 2)
+
+    --repeat-mode: The mode to repeat prompts. The supported modes are:
+        - 'random': shuffle the prompts randomly. (Default)
+        - 'tile': the entire prompt list is repeated in sequence. (Potentially
+                  lowest cache hit)
+        - 'interleave': each prompt is repeated consecutively before 
+                        moving to the next element. (Highest cache hit)
+    
+    --shuffle-seed: Random seed when the repeat mode is "random".
+                    (Optional, default: 0)
+
+In the meantime, it also supports all the vLLM engine args to initialize the 
+LLM engine. You can refer to the `vllm.engine.arg_utils.EngineArgs` for more
+details.
+"""
+
+import dataclasses
+import random
+import time
+
+from vllm import LLM, SamplingParams
+from vllm.engine.arg_utils import EngineArgs
+from vllm.utils import FlexibleArgumentParser
+
+
+def test_long_document_qa(llm=None, sampling_params=None, prompts=None):
+    """
+    Test long document QA with the given prompts and sampling parameters.
+    Print the time spent in processing all the prompts.
+
+    Args:
+        llm: The language model used for generating responses.
+        sampling_params: Sampling parameter used to generate the response.
+        prompts: A list of prompt strings to be processed by the LLM.
+    """
+    start_time = time.time()
+    llm.generate(prompts, sampling_params=sampling_params)
+    end_time = time.time()
+    print(f"Time to execute all requests: {end_time - start_time:.4f} secs")
+
+
+def repeat_prompts(prompts, repeat_count, mode: str):
+    """
+    Repeat each prompt in the list for a specified number of times.
+    The order of prompts in the output list depends on the mode.
+
+    Args:
+        prompts: A list of prompts to be repeated.
+        repeat_count: The number of times each prompt is repeated.
+        mode: The mode of repetition. Supported modes are:
+            - 'random': Shuffle the prompts randomly after repetition.
+            - 'tile': Repeat the entire prompt list in sequence.
+              Example: [1, 2, 3] -> [1, 2, 3, 1, 2, 3].
+            - 'interleave': Repeat each prompt consecutively before moving to 
+              the next. Example: [1, 2, 3] -> [1, 1, 2, 2, 3, 3].
+
+    Returns:
+        A list of repeated prompts in the specified order.
+
+    Raises:
+        ValueError: If an invalid mode is provided.
+    """
+    print("Repeat mode: ", mode)
+    if mode == 'random':
+        repeated_prompts = prompts * repeat_count
+        random.shuffle(repeated_prompts)
+        return repeated_prompts
+    elif mode == 'tile':
+        return prompts * repeat_count
+    elif mode == 'interleave':
+        repeated_prompts = []
+        for prompt in prompts:
+            repeated_prompts.extend([prompt] * repeat_count)
+        return repeated_prompts
+    else:
+        raise ValueError(f"Invalid mode: {mode}, only support "
+                         "'random', 'tile', 'interleave'")
+
+
+def main(args):
+    random.seed(args.shuffle_seed)
+
+    # Prepare the prompts:
+    # we append the document id at the beginning to avoid any of the document
+    # being the prefix of other documents
+    prompts = [
+        str(i) + ' '.join(['hi'] * args.document_length)
+        for i in range(args.num_documents)
+    ]
+
+    prompts = repeat_prompts(prompts, args.repeat_count, mode=args.repeat_mode)
+
+    warmup_prompts = [
+        "This is warm up request " + str(i) + \
+                ' '.join(['hi'] * args.document_length)
+        for i in range(args.num_documents)]
+
+    # Create the LLM engine
+    engine_args = EngineArgs.from_cli_args(args)
+    llm = LLM(**dataclasses.asdict(engine_args))
+    sampling_params = SamplingParams(temperature=0, max_tokens=args.output_len)
+
+    print("------warm up------")
+    test_long_document_qa(
+        llm=llm,
+        prompts=warmup_prompts,
+        sampling_params=sampling_params,
+    )
+
+    print("------start generating------")
+    test_long_document_qa(
+        llm=llm,
+        prompts=prompts,
+        sampling_params=sampling_params,
+    )
+
+
+if __name__ == "__main__":
+    parser = FlexibleArgumentParser(
+        description=
+        'Benchmark the performance with or without automatic prefix caching.')
+
+    parser.add_argument(
+        '--document-length',
+        type=int,
+        # Roughly the number of tokens for a system paper,
+        # excluding images
+        default=20000,
+        help='Range of input lengths for sampling prompts,'
+        'specified as "min:max" (e.g., "128:256").')
+
+    parser.add_argument('--num-documents',
+                        type=int,
+                        default=8,
+                        help='Range of input lengths for sampling prompts,'
+                        'specified as "min:max" (e.g., "128:256").')
+
+    parser.add_argument('--output-len', type=int, default=10)
+
+    parser.add_argument('--repeat-count',
+                        type=int,
+                        default=2,
+                        help='Number of times to repeat each prompt')
+
+    parser.add_argument("--repeat-mode",
+                        type=str,
+                        default='random',
+                        help='The mode to repeat prompts. The supported '
+                        'modes are "random", "tile", and "interleave". '
+                        'See repeat_prompts() in the source code for details.')
+
+    parser.add_argument("--shuffle-seed",
+                        type=int,
+                        default=0,
+                        help='Random seed when the repeat mode is "random"')
+
+    parser = EngineArgs.add_cli_args(parser)
+    args = parser.parse_args()
+    main(args)
@@ -1,3 +1,4 @@
+# SPDX-License-Identifier: Apache-2.0
 """
 Benchmark the efficiency of prefix caching.

@@ -10,7 +11,8 @@ Fixed example usage:
        --model meta-llama/Llama-2-7b-chat-hf \
        --enable-prefix-caching \
        --num-prompts 1 \
-        --repeat-count 100
+        --repeat-count 100 \
+        --input-length-range 128:256

 ShareGPT example usage:
    # This command samples 20 prompts with input lengths
@@ -29,7 +31,7 @@ import dataclasses
 import json
 import random
 import time
-from typing import List, Optional, Tuple
+from typing import Optional

 from transformers import PreTrainedTokenizerBase

@@ -75,9 +77,9 @@ def sample_requests_from_dataset(
    dataset_path: str,
    num_requests: int,
    tokenizer: PreTrainedTokenizerBase,
-    input_length_range: Tuple[int, int],
+    input_length_range: tuple[int, int],
    fixed_output_len: Optional[int],
-) -> List[Request]:
+) -> list[Request]:
    if fixed_output_len is not None and fixed_output_len < 4:
        raise ValueError("output_len too small")

@@ -97,7 +99,7 @@ def sample_requests_from_dataset(
    assert min_len >= 0 and max_len >= min_len, "input_length_range too small"

    # Filter out sequences that are too long or too short
-    filtered_requests: List[Request] = []
+    filtered_requests: list[Request] = []

    for i in range(len(dataset)):
        if len(filtered_requests) == num_requests:
@@ -120,10 +122,10 @@ def sample_requests_from_dataset(
 def sample_requests_from_random(
    num_requests: int,
    tokenizer: PreTrainedTokenizerBase,
-    input_length_range: Tuple[int, int],
+    input_length_range: tuple[int, int],
    fixed_output_len: Optional[int],
    prefix_len: int,
-) -> List[Request]:
+) -> list[Request]:

    requests = []
    prefix_token_ids = sample_tokens(tokenizer, prefix_len)
@@ -142,9 +144,9 @@ def sample_requests_from_random(
    return requests


-def repeat_and_sort_requests(requests: List[Request],
+def repeat_and_sort_requests(requests: list[Request],
                             repeat_count: int,
-                             sort: bool = False) -> List[str]:
+                             sort: bool = False) -> list[str]:
    repeated_requests = requests * repeat_count
    if sort:
        repeated_requests.sort(key=lambda x: x[1])
@@ -1,10 +1,11 @@
+# SPDX-License-Identifier: Apache-2.0
 """Benchmark offline prioritization."""
 import argparse
 import dataclasses
 import json
 import random
 import time
-from typing import List, Optional, Tuple
+from typing import Optional

 from transformers import AutoTokenizer, PreTrainedTokenizerBase

@@ -12,12 +13,17 @@ from vllm.engine.arg_utils import EngineArgs
 from vllm.utils import FlexibleArgumentParser


+#Select a equi-probable random priority
+def get_random_flag():
+    return 0 if random.random() < 0.5 else 1
+
+
 def sample_requests(
    dataset_path: str,
    num_requests: int,
    tokenizer: PreTrainedTokenizerBase,
    fixed_output_len: Optional[int],
-) -> List[Tuple[str, int, int]]:
+) -> list[tuple[str, int, int]]:
    if fixed_output_len is not None and fixed_output_len < 4:
        raise ValueError("output_len too small")

@@ -34,7 +40,7 @@ def sample_requests(
    random.shuffle(dataset)

    # Filter out sequences that are too long or too short
-    filtered_dataset: List[Tuple[str, int, int]] = []
+    filtered_dataset: list[tuple[str, int, int]] = []
    for i in range(len(dataset)):
        if len(filtered_dataset) == num_requests:
            break
@@ -54,8 +60,7 @@ def sample_requests(
            # Prune too long sequences.
            continue

-        #Select a equi-probable random priority
-        priority = 0 if random.random() < 0.5 else 1
+        priority = get_random_flag()

        filtered_dataset.append((prompt, prompt_len, output_len, priority))

@@ -63,13 +68,19 @@ def sample_requests(


 def run_vllm(
-    requests: List[Tuple[str, int, int]],
+    requests: list[tuple[str, int, int]],
    n: int,
    engine_args: EngineArgs,
 ) -> float:
    from vllm import LLM, SamplingParams
    llm = LLM(**dataclasses.asdict(engine_args))

+    assert all(
+        llm.llm_engine.model_config.max_model_len >= (request[1] + request[2])
+        for request in requests), (
+            "Please ensure that max_model_len is greater than the sum of"
+            " input_len and output_len for all requests.")
+
    # Add the requests to the engine.
    prompts = []
    sampling_params = []
@@ -102,8 +113,8 @@ def main(args: argparse.Namespace):
    if args.dataset is None:
        # Synthesize a prompt with the given input length.
        prompt = "hi" * (args.input_len - 1)
-        requests = [(prompt, args.input_len, args.output_len)
-                    for _ in range(args.num_prompts)]
+        requests = [(prompt, args.input_len, args.output_len,
+                     get_random_flag()) for _ in range(args.num_prompts)]
    else:
        requests = sample_requests(args.dataset, args.num_prompts, tokenizer,
                                   args.output_len)
@@ -1,3 +1,4 @@
+# SPDX-License-Identifier: Apache-2.0
 r"""Benchmark online serving throughput.

 On the server side, run one of the following commands:
@@ -25,17 +26,20 @@ On the client side, run:
 import argparse
 import asyncio
 import base64
+import gc
 import io
 import json
 import os
 import random
 import time
 import warnings
+from collections.abc import AsyncGenerator, Collection
 from dataclasses import dataclass
 from datetime import datetime
-from typing import Any, AsyncGenerator, Collection, Dict, List, Optional, Tuple
+from typing import Any, Optional

 import numpy as np
+import pandas as pd
 from backend_request_func import (ASYNC_REQUEST_FUNCS, RequestFuncInput,
                                  RequestFuncOutput)
 from datasets import load_dataset
@@ -53,6 +57,8 @@ try:
 except ImportError:
    from argparse import ArgumentParser as FlexibleArgumentParser

+from benchmark_utils import convert_to_pytorch_benchmark_format, write_to_json
+
 MILLISECONDS_TO_SECONDS_CONVERSION = 1000


@@ -68,22 +74,22 @@ class BenchmarkMetrics:
    mean_ttft_ms: float
    median_ttft_ms: float
    std_ttft_ms: float
-    percentiles_ttft_ms: List[Tuple[float, float]]
+    percentiles_ttft_ms: list[tuple[float, float]]
    mean_tpot_ms: float
    median_tpot_ms: float
    std_tpot_ms: float
-    percentiles_tpot_ms: List[Tuple[float, float]]
+    percentiles_tpot_ms: list[tuple[float, float]]
    mean_itl_ms: float
    median_itl_ms: float
    std_itl_ms: float
-    percentiles_itl_ms: List[Tuple[float, float]]
+    percentiles_itl_ms: list[tuple[float, float]]
    # E2EL stands for end-to-end latency per request.
    # It is the time taken on the client side from sending
    # a request to receiving a complete response.
    mean_e2el_ms: float
    median_e2el_ms: float
    std_e2el_ms: float
-    percentiles_e2el_ms: List[Tuple[float, float]]
+    percentiles_e2el_ms: list[tuple[float, float]]


 def sample_sharegpt_requests(
@@ -91,7 +97,7 @@ def sample_sharegpt_requests(
    num_requests: int,
    tokenizer: PreTrainedTokenizerBase,
    fixed_output_len: Optional[int] = None,
-) -> List[Tuple[str, int, int, None]]:
+) -> list[tuple[str, int, int, None]]:
    # Load the dataset.
    with open(dataset_path, encoding='utf-8') as f:
        dataset = json.load(f)
@@ -105,7 +111,7 @@ def sample_sharegpt_requests(
    random.shuffle(dataset)

    # Filter out sequences that are too long or too short
-    filtered_dataset: List[Tuple[str, int, int]] = []
+    filtered_dataset: list[tuple[str, int, int]] = []
    for i in range(len(dataset)):
        if len(filtered_dataset) == num_requests:
            break
@@ -129,6 +135,35 @@ def sample_sharegpt_requests(
    return filtered_dataset


+def sample_burstgpt_requests(
+    dataset_path: str,
+    num_requests: int,
+    random_seed: int,
+    tokenizer: PreTrainedTokenizerBase,
+) -> list[tuple[str, int, int, None]]:
+    df = pd.read_csv(dataset_path)
+    gpt4_df = df[df["Model"] == "GPT-4"]
+    # Remove the failed requests (i.e., response length is 0)
+    gpt4_df = gpt4_df[gpt4_df["Response tokens"] > 0]
+    # Randomly sample num_requests from the dataset
+    if num_requests <= len(gpt4_df):
+        gpt4_df = gpt4_df.sample(n=num_requests, random_state=random_seed)
+    else:
+        gpt4_df = gpt4_df.sample(n=num_requests,
+                                 random_state=random_seed,
+                                 replace=True)
+    # Convert the dataframe to a list of tuples
+    dataset = gpt4_df.values.tolist()
+    input_requests = []
+    for i in range(num_requests):
+        input_len = int(dataset[i][2])
+        output_len = int(dataset[i][3])
+        prompt = tokenizer.decode([(i + j) % tokenizer.vocab_size
+                                   for j in range(input_len)])
+        input_requests.append((prompt, input_len, output_len, None))
+    return input_requests
+
+
 def sample_sonnet_requests(
    dataset_path: str,
    num_requests: int,
@@ -136,7 +171,7 @@ def sample_sonnet_requests(
    output_len: int,
    prefix_len: int,
    tokenizer: PreTrainedTokenizerBase,
-) -> List[Tuple[str, str, int, int, None]]:
+) -> list[tuple[str, str, int, int, None]]:
    assert (
        input_len > prefix_len
    ), "'args.sonnet-input-len' must be greater than 'args.prefix-input-len'."
@@ -177,7 +212,7 @@ def sample_sonnet_requests(
    prefix_lines = poem_lines[:num_prefix_lines]

    # Sample the rest of lines per request.
-    sampled_requests: List[Tuple[str, int, int]] = []
+    sampled_requests: list[tuple[str, int, int]] = []
    for _ in range(num_requests):
        num_lines_needed = num_input_lines - num_prefix_lines
        sampled_lines = "".join(prefix_lines +
@@ -199,25 +234,19 @@ def sample_sonnet_requests(
    return sampled_requests


-def sample_mmmu_pro_vision_requests(
+def sample_vision_arena_requests(
    dataset,
    num_requests: int,
    tokenizer: PreTrainedTokenizerBase,
    fixed_output_len: Optional[int] = None,
-) -> List[Tuple[str, str, int, Optional[Dict[str, Collection[str]]]]]:
-    sampled_requests: List[Tuple[str, int, int, Dict[str,
+) -> list[tuple[str, str, int, Optional[dict[str, Collection[str]]]]]:
+    sampled_requests: list[tuple[str, int, int, dict[str,
                                                     Collection[str]]]] = []
    for data in dataset:
        if len(sampled_requests) == num_requests:
            break

-        # MMMU-Pro vision direct prompt
-        # Ref: https://github.com/MMMU-Benchmark/MMMU/blob/6ce42f4d8f70c1841c67867152648974415b5cac/mmmu-pro/prompts.yaml#L5
-        prompt = (
-            "Answer with the option letter from the given choices directly. "
-            "The last line of your response should be of the following "
-            "format: 'Answer: $LETTER' (without quotes) where LETTER is one of "
-            "options.")
+        prompt = data["turns"][0][0]['content']

        prompt_token_ids = tokenizer(prompt).input_ids
        if fixed_output_len is None:
@@ -229,10 +258,10 @@ def sample_mmmu_pro_vision_requests(
        output_len = fixed_output_len

        assert isinstance(
-            data["image"],
+            data["images"][0],
            Image), ("Input image format must be `PIL.Image.Image`, "
                     f"given {type(data['image'])}.")
-        image: Image = data["image"]
+        image: Image = data["images"][0]
        image = image.convert("RGB")
        image_data = io.BytesIO()
        image.save(image_data, format='JPEG')
@@ -251,27 +280,25 @@ def sample_mmmu_pro_vision_requests(

 def sample_hf_requests(
    dataset_path: str,
-    dataset_subset: str,
+    dataset_subset: Optional[str],
    dataset_split: str,
    num_requests: int,
    tokenizer: PreTrainedTokenizerBase,
    random_seed: int,
    fixed_output_len: Optional[int] = None,
-) -> List[Tuple[str, str, int, Optional[Dict[str, Collection[str]]]]]:
+) -> list[tuple[str, str, int, Optional[dict[str, Collection[str]]]]]:

-    # Special case for MMMU-Pro vision dataset
-    if dataset_path == 'MMMU/MMMU_Pro' and dataset_subset == 'vision':
-        assert dataset_split == "test"
+    # Special case for vision_arena dataset
+    if dataset_path == 'lmarena-ai/vision-arena-bench-v0.1' \
+        and dataset_subset is None:
+        assert dataset_split == "train"
        dataset = load_dataset(dataset_path,
                               name=dataset_subset,
                               split=dataset_split,
                               streaming=True)
-        assert "image" in dataset.features, (
-            "MMMU/MMMU_Pro vision dataset must have 'image' column.")
-        filter_func = lambda x: isinstance(x["image"], Image)
-        dataset = dataset.shuffle(seed=random_seed).filter(filter_func)
-        return sample_mmmu_pro_vision_requests(dataset, num_requests,
-                                               tokenizer, fixed_output_len)
+        dataset = dataset.shuffle(seed=random_seed)
+        return sample_vision_arena_requests(dataset, num_requests, tokenizer,
+                                            fixed_output_len)

    dataset = load_dataset(dataset_path,
                           name=dataset_subset,
@@ -281,7 +308,7 @@ def sample_hf_requests(
        "HF Dataset must have 'conversations' column.")
    filter_func = lambda x: len(x["conversations"]) >= 2
    filtered_dataset = dataset.shuffle(seed=random_seed).filter(filter_func)
-    sampled_requests: List[Tuple[str, int, int, Dict[str,
+    sampled_requests: list[tuple[str, int, int, dict[str,
                                                     Collection[str]]]] = []
    for data in filtered_dataset:
        if len(sampled_requests) == num_requests:
@@ -344,7 +371,7 @@ def sample_random_requests(
    num_prompts: int,
    range_ratio: float,
    tokenizer: PreTrainedTokenizerBase,
-) -> List[Tuple[str, int, int]]:
+) -> list[tuple[str, int, int]]:
    prefix_token_ids = np.random.randint(0,
                                         tokenizer.vocab_size,
                                         size=prefix_len).tolist()
@@ -373,26 +400,26 @@ def sample_random_requests(


 async def get_request(
-    input_requests: List[Tuple[str, int, int]],
+    input_requests: list[tuple[str, int, int]],
    request_rate: float,
    burstiness: float = 1.0,
-) -> AsyncGenerator[Tuple[str, int, int], None]:
+) -> AsyncGenerator[tuple[str, int, int], None]:
    """
-    Asynchronously generates requests at a specified rate 
+    Asynchronously generates requests at a specified rate
    with OPTIONAL burstiness.
-    
+
    Args:
-        input_requests: 
+        input_requests:
            A list of input requests, each represented as a tuple.
-        request_rate: 
+        request_rate:
            The rate at which requests are generated (requests/s).
-        burstiness (optional): 
-            The burstiness factor of the request generation. 
+        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 
+            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.
    """
    input_requests = iter(input_requests)
@@ -417,38 +444,42 @@ async def get_request(


 def calculate_metrics(
-    input_requests: List[Tuple[str, int, int]],
-    outputs: List[RequestFuncOutput],
+    input_requests: list[tuple[str, int, int]],
+    outputs: list[RequestFuncOutput],
    dur_s: float,
    tokenizer: PreTrainedTokenizerBase,
-    selected_percentile_metrics: List[str],
-    selected_percentiles: List[float],
-    gootput_config_dict: Dict[str, float],
-) -> Tuple[BenchmarkMetrics, List[int]]:
-    actual_output_lens: List[int] = []
+    selected_percentile_metrics: list[str],
+    selected_percentiles: list[float],
+    goodput_config_dict: dict[str, float],
+) -> tuple[BenchmarkMetrics, list[int]]:
+    actual_output_lens: list[int] = []
    total_input = 0
    completed = 0
    good_completed = 0
-    itls: List[float] = []
-    tpots: List[float] = []
-    all_tpots: List[float] = []
-    ttfts: List[float] = []
-    e2els: List[float] = []
+    itls: list[float] = []
+    tpots: list[float] = []
+    all_tpots: list[float] = []
+    ttfts: list[float] = []
+    e2els: list[float] = []
    for i in range(len(outputs)):
        if outputs[i].success:
-            # We use the tokenizer to count the number of output tokens for all
-            # serving backends instead of looking at len(outputs[i].itl) since
-            # multiple output tokens may be bundled together
-            # Note : this may inflate the output token count slightly
-            output_len = len(
-                tokenizer(outputs[i].generated_text,
-                          add_special_tokens=False).input_ids)
+            output_len = outputs[i].output_tokens
+
+            if output_len is None:
+                # We use the tokenizer to count the number of output tokens
+                # for some serving backends instead of looking at
+                # len(outputs[i].itl) since multiple output tokens may be
+                # bundled together
+                # Note : this may inflate the output token count slightly
+                output_len = len(
+                    tokenizer(outputs[i].generated_text,
+                              add_special_tokens=False).input_ids)
            actual_output_lens.append(output_len)
            total_input += input_requests[i][1]
            tpot = 0
            if output_len > 1:
-                tpot = (outputs[i].latency - outputs[i].ttft) / (output_len -
-                                                                 1)
+                latency_minus_ttft = outputs[i].latency - outputs[i].ttft
+                tpot = latency_minus_ttft / (output_len - 1)
                tpots.append(tpot)
            # Note: if output_len <= 1, we regard tpot as 0 for goodput
            all_tpots.append(tpot)
@@ -459,21 +490,21 @@ def calculate_metrics(
        else:
            actual_output_lens.append(0)

-    if gootput_config_dict:
+    if goodput_config_dict:
        valid_metrics = []
        slo_values = []

-        if "ttft" in gootput_config_dict:
+        if "ttft" in goodput_config_dict:
            valid_metrics.append(ttfts)
-            slo_values.append(gootput_config_dict["ttft"] /
+            slo_values.append(goodput_config_dict["ttft"] /
                              MILLISECONDS_TO_SECONDS_CONVERSION)
-        if "tpot" in gootput_config_dict:
+        if "tpot" in goodput_config_dict:
            valid_metrics.append(all_tpots)
-            slo_values.append(gootput_config_dict["tpot"] /
+            slo_values.append(goodput_config_dict["tpot"] /
                              MILLISECONDS_TO_SECONDS_CONVERSION)
-        if "e2el" in gootput_config_dict:
+        if "e2el" in goodput_config_dict:
            valid_metrics.append(e2els)
-            slo_values.append(gootput_config_dict["e2el"] /
+            slo_values.append(goodput_config_dict["e2el"] /
                              MILLISECONDS_TO_SECONDS_CONVERSION)

        for req_metric in zip(*valid_metrics):
@@ -525,19 +556,21 @@ async def benchmark(
    api_url: str,
    base_url: str,
    model_id: str,
+    model_name: str,
    tokenizer: PreTrainedTokenizerBase,
-    input_requests: List[Tuple[str, int, int]],
+    input_requests: list[tuple[str, int, int]],
    logprobs: Optional[int],
    best_of: int,
    request_rate: float,
    burstiness: float,
    disable_tqdm: bool,
    profile: bool,
-    selected_percentile_metrics: List[str],
-    selected_percentiles: List[str],
+    selected_percentile_metrics: list[str],
+    selected_percentiles: list[str],
    ignore_eos: bool,
-    gootput_config_dict: Dict[str, float],
+    goodput_config_dict: dict[str, float],
    max_concurrency: Optional[int],
+    lora_modules: Optional[list[str]],
 ):
    if backend in ASYNC_REQUEST_FUNCS:
        request_func = ASYNC_REQUEST_FUNCS[backend]
@@ -553,6 +586,7 @@ async def benchmark(
            "Multi-modal content is only supported on 'openai-chat' backend.")
    test_input = RequestFuncInput(
        model=model_id,
+        model_name=model_name,
        prompt=test_prompt,
        api_url=api_url,
        prompt_len=test_prompt_len,
@@ -562,6 +596,7 @@ async def benchmark(
        multi_modal_content=test_mm_content,
        ignore_eos=ignore_eos,
    )
+
    test_output = await request_func(request_func_input=test_input)
    if not test_output.success:
        raise ValueError(
@@ -570,9 +605,15 @@ async def benchmark(
    else:
        print("Initial test run completed. Starting main benchmark run...")

+    if lora_modules:
+        # For each input request, choose a LoRA module at random.
+        lora_modules = iter(
+            [random.choice(lora_modules) for _ in range(len(input_requests))])
+
    if profile:
        print("Starting profiler...")
        profile_input = RequestFuncInput(model=model_id,
+                                         model_name=model_name,
                                         prompt=test_prompt,
                                         api_url=base_url + "/start_profile",
                                         prompt_len=test_prompt_len,
@@ -612,10 +653,16 @@ async def benchmark(
                                      pbar=pbar)

    benchmark_start_time = time.perf_counter()
-    tasks: List[asyncio.Task] = []
+    tasks: list[asyncio.Task] = []
    async for request in get_request(input_requests, request_rate, burstiness):
        prompt, prompt_len, output_len, mm_content = request
-        request_func_input = RequestFuncInput(model=model_id,
+        req_model_id, req_model_name = model_id, model_name
+        if lora_modules:
+            req_lora_module = next(lora_modules)
+            req_model_id, req_model_name = req_lora_module, req_lora_module
+
+        request_func_input = RequestFuncInput(model=req_model_id,
+                                              model_name=req_model_name,
                                              prompt=prompt,
                                              api_url=api_url,
                                              prompt_len=prompt_len,
@@ -628,7 +675,7 @@ async def benchmark(
            asyncio.create_task(
                limited_request_func(request_func_input=request_func_input,
                                     pbar=pbar)))
-    outputs: List[RequestFuncOutput] = await asyncio.gather(*tasks)
+    outputs: list[RequestFuncOutput] = await asyncio.gather(*tasks)

    if profile:
        print("Stopping profiler...")
@@ -657,7 +704,7 @@ async def benchmark(
        tokenizer=tokenizer,
        selected_percentile_metrics=selected_percentile_metrics,
        selected_percentiles=selected_percentiles,
-        gootput_config_dict=gootput_config_dict,
+        goodput_config_dict=goodput_config_dict,
    )

    print("{s:{c}^{n}}".format(s=' Serving Benchmark Result ', n=50, c='='))
@@ -669,7 +716,7 @@ async def benchmark(
                                 metrics.total_output))
    print("{:<40} {:<10.2f}".format("Request throughput (req/s):",
                                    metrics.request_throughput))
-    if gootput_config_dict:
+    if goodput_config_dict:
        print("{:<40} {:<10.2f}".format("Request goodput (req/s):",
                                        metrics.request_goodput))
    print("{:<40} {:<10.2f}".format("Output token throughput (tok/s):",
@@ -684,7 +731,7 @@ async def benchmark(
        "total_output_tokens": metrics.total_output,
        "request_throughput": metrics.request_throughput,
        "request_goodput:":
-        metrics.request_goodput if gootput_config_dict else None,
+        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],
@@ -740,11 +787,11 @@ async def benchmark(

 def check_goodput_args(args):
    # Check and parse goodput arguments
-    gootput_config_dict = {}
+    goodput_config_dict = {}
    VALID_NAMES = ["ttft", "tpot", "e2el"]
    if args.goodput:
-        gootput_config_dict = parse_goodput(args.goodput)
-        for slo_name, slo_val in gootput_config_dict.items():
+        goodput_config_dict = parse_goodput(args.goodput)
+        for slo_name, slo_val in goodput_config_dict.items():
            if slo_name not in VALID_NAMES:
                raise ValueError(
                    f"Invalid metric name found, {slo_name}: {slo_val}. "
@@ -755,22 +802,47 @@ def check_goodput_args(args):
                    f"Invalid value found, {slo_name}: {slo_val}. "
                    "The service level objective value should be "
                    "non-negative.")
-    return gootput_config_dict
+    return goodput_config_dict


 def parse_goodput(slo_pairs):
-    gootput_config_dict = {}
+    goodput_config_dict = {}
    try:
        for slo_pair in slo_pairs:
            slo_name, slo_val = slo_pair.split(":")
-            gootput_config_dict[slo_name] = float(slo_val)
+            goodput_config_dict[slo_name] = float(slo_val)
    except ValueError as err:
        raise argparse.ArgumentTypeError(
            "Invalid format found for service level objectives. "
            "Specify service level objectives for goodput as \"KEY:VALUE\" "
            "pairs, where the key is a metric name, and the value is a "
            "number in milliseconds.") from err
-    return gootput_config_dict
+    return goodput_config_dict
+
+
+def save_to_pytorch_benchmark_format(args: argparse.Namespace,
+                                     results: dict[str, Any],
+                                     file_name: str) -> None:
+    metrics = [
+        "median_ttft_ms", "mean_ttft_ms", "std_ttft_ms", "p99_ttft_ms",
+        "mean_tpot_ms", "median_tpot_ms", "std_tpot_ms", "p99_tpot_ms",
+        "median_itl_ms", "mean_itl_ms", "std_itl_ms", "p99_itl_ms"
+    ]
+    # These raw data might be useful, but they are rather big. They can be added
+    # later if needed
+    ignored_metrics = ["ttfts", "itls", "generated_texts", "errors"]
+    pt_records = convert_to_pytorch_benchmark_format(
+        args=args,
+        metrics={k: [results[k]]
+                 for k in metrics},
+        extra_info={
+            k: results[k]
+            for k in results if k not in metrics and k not in ignored_metrics
+        })
+    if pt_records:
+        # Don't use json suffix here as we don't want CI to pick it up
+        pt_file = f"{os.path.splitext(file_name)[0]}.pytorch.json"
+        write_to_json(pt_file, pt_records)


 def main(args: argparse.Namespace):
@@ -780,6 +852,7 @@ def main(args: argparse.Namespace):

    backend = args.backend
    model_id = args.model
+    model_name = args.served_model_name
    tokenizer_id = args.tokenizer if args.tokenizer is not None else args.model
    tokenizer_mode = args.tokenizer_mode

@@ -794,18 +867,10 @@ def main(args: argparse.Namespace):
                              tokenizer_mode=tokenizer_mode,
                              trust_remote_code=args.trust_remote_code)

-    if args.dataset is not None:
-        warnings.warn(
-            "The '--dataset' argument will be deprecated in the next "
-            "release. Please use '--dataset-name' and "
-            "'--dataset-path' in the future runs.",
-            stacklevel=2)
-        input_requests = sample_sharegpt_requests(
-            dataset_path=args.dataset,
-            num_requests=args.num_prompts,
-            tokenizer=tokenizer,
-            fixed_output_len=args.sharegpt_output_len,
-        )
+    if args.dataset_name is None:
+        raise ValueError(
+            "Please specify '--dataset-name' and the corresponding "
+            "'--dataset-path' if required.")

    elif args.dataset_name == "sharegpt":
        input_requests = sample_sharegpt_requests(
@@ -815,6 +880,14 @@ def main(args: argparse.Namespace):
            fixed_output_len=args.sharegpt_output_len,
        )

+    elif args.dataset_name == "burstgpt":
+        input_requests = sample_burstgpt_requests(
+            dataset_path=args.dataset_path,
+            num_requests=args.num_prompts,
+            random_seed=args.seed,
+            tokenizer=tokenizer,
+        )
+
    elif args.dataset_name == "sonnet":
        # Do not format the prompt, pass to message directly
        if args.backend == "openai-chat":
@@ -869,7 +942,11 @@ def main(args: argparse.Namespace):
    else:
        raise ValueError(f"Unknown dataset: {args.dataset_name}")

-    gootput_config_dict = check_goodput_args(args)
+    goodput_config_dict = check_goodput_args(args)
+
+    # Avoid GC processing "static" data - reduce pause times.
+    gc.collect()
+    gc.freeze()

    benchmark_result = asyncio.run(
        benchmark(
@@ -877,6 +954,7 @@ def main(args: argparse.Namespace):
            api_url=api_url,
            base_url=base_url,
            model_id=model_id,
+            model_name=model_name,
            tokenizer=tokenizer,
            input_requests=input_requests,
            logprobs=args.logprobs,
@@ -890,13 +968,14 @@ def main(args: argparse.Namespace):
                float(p) for p in args.metric_percentiles.split(",")
            ],
            ignore_eos=args.ignore_eos,
-            gootput_config_dict=gootput_config_dict,
+            goodput_config_dict=goodput_config_dict,
            max_concurrency=args.max_concurrency,
+            lora_modules=args.lora_modules,
        ))

    # Save config and results to json
    if args.save_result:
-        result_json: Dict[str, Any] = {}
+        result_json: dict[str, Any] = {}

        # Setup
        current_dt = datetime.now().strftime("%Y%m%d-%H%M%S")
@@ -919,8 +998,8 @@ def main(args: argparse.Namespace):
                    )

        # Traffic
-        result_json["request_rate"] = (
-            args.request_rate if args.request_rate < float("inf") else "inf")
+        result_json["request_rate"] = (args.request_rate if args.request_rate
+                                       < float("inf") else "inf")
        result_json["burstiness"] = args.burstiness
        result_json["max_concurrency"] = args.max_concurrency

@@ -938,6 +1017,7 @@ def main(args: argparse.Namespace):
            file_name = os.path.join(args.result_dir, file_name)
        with open(file_name, "w", encoding='utf-8') as outfile:
            json.dump(result_json, outfile)
+        save_to_pytorch_benchmark_format(args, result_json, file_name)


 if __name__ == "__main__":
@@ -955,7 +1035,8 @@ if __name__ == "__main__":
        default=None,
        help="Server or API base url if not using http host and port.",
    )
-    parser.add_argument("--host", type=str, default="localhost")
+    # Use 127.0.0.1 here instead of localhost to force the use of ipv4
+    parser.add_argument("--host", type=str, default="127.0.0.1")
    parser.add_argument("--port", type=int, default=8000)
    parser.add_argument(
        "--endpoint",
@@ -963,18 +1044,11 @@ if __name__ == "__main__":
        default="/v1/completions",
        help="API endpoint.",
    )
-    parser.add_argument(
-        "--dataset",
-        type=str,
-        default=None,
-        help="Path to the ShareGPT dataset, will be deprecated in the "
-        "next release.",
-    )
    parser.add_argument(
        "--dataset-name",
        type=str,
        default="sharegpt",
-        choices=["sharegpt", "sonnet", "random", "hf"],
+        choices=["sharegpt", "burstgpt", "sonnet", "random", "hf"],
        help="Name of the dataset to benchmark on.",
    )
    parser.add_argument("--dataset-path",
@@ -1216,11 +1290,26 @@ if __name__ == "__main__":
        '--tokenizer-mode',
        type=str,
        default="auto",
-        choices=['auto', 'slow', 'mistral'],
+        choices=['auto', 'slow', 'mistral', 'custom'],
        help='The tokenizer mode.\n\n* "auto" will use the '
        'fast tokenizer if available.\n* "slow" will '
        'always use the slow tokenizer. \n* '
-        '"mistral" will always use the `mistral_common` tokenizer.')
+        '"mistral" will always use the `mistral_common` tokenizer. \n*'
+        '"custom" will use --tokenizer to select the preregistered tokenizer.')
+
+    parser.add_argument("--served-model-name",
+                        type=str,
+                        default=None,
+                        help="The model name used in the API. "
+                        "If not specified, the model name will be the "
+                        "same as the ``--model`` argument. ")
+
+    parser.add_argument("--lora-modules",
+                        nargs='+',
+                        default=None,
+                        help="A subset of LoRA module names passed in when "
+                        "launching the server. For each request, the "
+                        "script chooses a LoRA module at random.")

    args = parser.parse_args()
    main(args)
@@ -1,3 +1,4 @@
+# SPDX-License-Identifier: Apache-2.0
 r"""Benchmark online serving throughput with guided decoding.

 On the server side, run one of the following commands:
@@ -8,7 +9,7 @@ On the server side, run one of the following commands:
    ./launch_tgi_server.sh <your_model> <max_batch_total_tokens>

 On the client side, run:
-    python benchmarks/benchmark_serving.py \
+    python benchmarks/benchmark_serving_guided.py \
        --backend <backend> \
        --model <your_model> \
        --dataset json \
@@ -29,8 +30,9 @@ import os
 import random
 import time
 import warnings
+from collections.abc import AsyncGenerator
 from dataclasses import dataclass
-from typing import AsyncGenerator, List, Optional, Tuple
+from typing import Optional

 import datasets
 import numpy as np
@@ -65,22 +67,22 @@ class BenchmarkMetrics:
    mean_ttft_ms: float
    median_ttft_ms: float
    std_ttft_ms: float
-    percentiles_ttft_ms: List[Tuple[float, float]]
+    percentiles_ttft_ms: list[tuple[float, float]]
    mean_tpot_ms: float
    median_tpot_ms: float
    std_tpot_ms: float
-    percentiles_tpot_ms: List[Tuple[float, float]]
+    percentiles_tpot_ms: list[tuple[float, float]]
    mean_itl_ms: float
    median_itl_ms: float
    std_itl_ms: float
-    percentiles_itl_ms: List[Tuple[float, float]]
+    percentiles_itl_ms: list[tuple[float, float]]
    # E2EL stands for end-to-end latency per request.
    # It is the time taken on the client side from sending
    # a request to receiving a complete response.
    mean_e2el_ms: float
    median_e2el_ms: float
    std_e2el_ms: float
-    percentiles_e2el_ms: List[Tuple[float, float]]
+    percentiles_e2el_ms: list[tuple[float, float]]


@dataclasses.dataclass
@@ -103,7 +105,7 @@ class SampleRequest:


 def sample_requests(tokenizer: PreTrainedTokenizerBase,
-                    args: argparse.Namespace) -> List[SampleRequest]:
+                    args: argparse.Namespace) -> list[SampleRequest]:
    if args.dataset == 'json':
        if args.json_schema_path is None:
            dir_path = os.path.dirname(os.path.realpath(__file__))
@@ -186,7 +188,7 @@ def sample_requests(tokenizer: PreTrainedTokenizerBase,
        ]

    elif args.dataset == "xgrammar_bench":
-        requests: List[SampleRequest] = []
+        requests: list[SampleRequest] = []
        dataset = datasets.load_dataset("NousResearch/json-mode-eval",
                                        split="train")
        print(f"dataset has {len(dataset)} entries")
@@ -213,10 +215,10 @@ def sample_requests(tokenizer: PreTrainedTokenizerBase,


 async def get_request(
-    input_requests: List[SampleRequest],
+    input_requests: list[SampleRequest],
    request_rate: float,
    burstiness: float = 1.0,
-) -> AsyncGenerator[Tuple[int, SampleRequest], None]:
+) -> AsyncGenerator[tuple[int, SampleRequest], None]:
    """
    Asynchronously generates requests at a specified rate 
    with OPTIONAL burstiness.
@@ -257,22 +259,23 @@ async def get_request(


 def calculate_metrics(
-    input_requests: List[Tuple[str, int, int]],
-    outputs: List[RequestFuncOutput],
+    input_requests: list[tuple[str, int, int]],
+    outputs: list[RequestFuncOutput],
    dur_s: float,
    tokenizer: PreTrainedTokenizerBase,
-    selected_percentile_metrics: List[str],
-    selected_percentiles: List[float],
-) -> Tuple[BenchmarkMetrics, List[int]]:
-    actual_output_lens: List[int] = []
+    selected_percentile_metrics: list[str],
+    selected_percentiles: list[float],
+    goodput_config_dict: Optional[dict[str, float]] = None,
+) -> tuple[BenchmarkMetrics, list[int]]:
+    actual_output_lens: list[int] = []
    total_input = 0
    completed = 0
    good_completed = 0
-    itls: List[float] = []
-    tpots: List[float] = []
-    all_tpots: List[float] = []
-    ttfts: List[float] = []
-    e2els: List[float] = []
+    itls: list[float] = []
+    tpots: list[float] = []
+    all_tpots: list[float] = []
+    ttfts: list[float] = []
+    e2els: list[float] = []
    for i in range(len(outputs)):
        if outputs[i].success:
            # We use the tokenizer to count the number of output tokens for all
@@ -286,10 +289,10 @@ def calculate_metrics(
            total_input += input_requests[i].prompt_len
            tpot = 0
            if output_len > 1:
-                tpot = (outputs[i].latency - outputs[i].ttft) / (output_len -
-                                                                 1)
+                latency_minus_ttft = outputs[i].latency - outputs[i].ttft
+                tpot = latency_minus_ttft / (output_len - 1)
                tpots.append(tpot)
-            outputs[i].tpot = sum(tpots) / len(tpots) if len(tpots) else 0
+            outputs[i].tpot = tpot
            # Note: if output_len <= 1, we regard tpot as 0 for goodput
            all_tpots.append(tpot)
            itls += outputs[i].itl
@@ -299,6 +302,28 @@ def calculate_metrics(
        else:
            actual_output_lens.append(0)

+    if goodput_config_dict:
+        valid_metrics = []
+        slo_values = []
+
+        if "ttft" in goodput_config_dict:
+            valid_metrics.append(ttfts)
+            slo_values.append(goodput_config_dict["ttft"] /
+                              MILLISECONDS_TO_SECONDS_CONVERSION)
+        if "tpot" in goodput_config_dict:
+            valid_metrics.append(all_tpots)
+            slo_values.append(goodput_config_dict["tpot"] /
+                              MILLISECONDS_TO_SECONDS_CONVERSION)
+        if "e2el" in goodput_config_dict:
+            valid_metrics.append(e2els)
+            slo_values.append(goodput_config_dict["e2el"] /
+                              MILLISECONDS_TO_SECONDS_CONVERSION)
+
+        for req_metric in zip(*valid_metrics):
+            is_good_req = all([s >= r for s, r in zip(slo_values, req_metric)])
+            if is_good_req:
+                good_completed += 1
+
    if completed == 0:
        warnings.warn(
            "All requests failed. This is likely due to a misconfiguration "
@@ -344,17 +369,18 @@ async def benchmark(
    base_url: str,
    model_id: str,
    tokenizer: PreTrainedTokenizerBase,
-    input_requests: List[SampleRequest],
+    input_requests: list[SampleRequest],
    request_rate: float,
    burstiness: float,
    disable_tqdm: bool,
    profile: bool,
-    selected_percentile_metrics: List[str],
-    selected_percentiles: List[str],
+    selected_percentile_metrics: list[str],
+    selected_percentiles: list[str],
    ignore_eos: bool,
    max_concurrency: Optional[int],
    guided_decoding_ratio: float,
    guided_decoding_backend: str,
+    goodput_config_dict: Optional[dict[str, float]] = None,
 ):
    if backend in ASYNC_REQUEST_FUNCS:
        request_func = ASYNC_REQUEST_FUNCS[backend]
@@ -434,8 +460,8 @@ async def benchmark(
                                      pbar=pbar)

    benchmark_start_time = time.perf_counter()
-    tasks: List[asyncio.Task] = []
-    expected: List[str] = []
+    tasks: list[asyncio.Task] = []
+    expected: list[str] = []
    async for i, request in get_request(input_requests, request_rate,
                                        burstiness):
        extra_body = prepare_extra_body(
@@ -454,7 +480,7 @@ async def benchmark(
            asyncio.create_task(
                limited_request_func(request_func_input=request_func_input,
                                     pbar=pbar)))
-    outputs: List[RequestFuncOutput] = await asyncio.gather(*tasks)
+    outputs: list[RequestFuncOutput] = await asyncio.gather(*tasks)

    if profile:
        print("Stopping profiler...")
@@ -482,6 +508,7 @@ async def benchmark(
        tokenizer=tokenizer,
        selected_percentile_metrics=selected_percentile_metrics,
        selected_percentiles=selected_percentiles,
+        goodput_config_dict=goodput_config_dict,
    )

    print("{s:{c}^{n}}".format(s=' Serving Benchmark Result ', n=50, c='='))
@@ -493,6 +520,9 @@ async def benchmark(
                                 metrics.total_output))
    print("{:<40} {:<10.2f}".format("Request throughput (req/s):",
                                    metrics.request_throughput))
+    if goodput_config_dict:
+        print("{:<40} {:<10.2f}".format("Request goodput (req/s):",
+                                        metrics.request_goodput))
    print("{:<40} {:<10.2f}".format("Output token throughput (tok/s):",
                                    metrics.output_throughput))
    print("{:<40} {:<10.2f}".format("Total Token throughput (tok/s):",
@@ -616,6 +646,40 @@ def evaluate(ret, args):
            100) if len(not_none_scores) > 0 else None


+def parse_goodput(slo_pairs):
+    goodput_config_dict = {}
+    try:
+        for slo_pair in slo_pairs:
+            slo_name, slo_val = slo_pair.split(":")
+            goodput_config_dict[slo_name] = float(slo_val)
+    except ValueError as err:
+        raise argparse.ArgumentTypeError(
+            "Invalid format found for service level objectives. "
+            "Specify service level objectives for goodput as \"KEY:VALUE\" "
+            "pairs, where the key is a metric name, and the value is a "
+            "number in milliseconds.") from err
+    return goodput_config_dict
+
+
+def check_goodput_args(args):
+    goodput_config_dict = {}
+    VALID_NAMES = ["ttft", "tpot", "e2el"]
+    if args.goodput:
+        goodput_config_dict = parse_goodput(args.goodput)
+        for slo_name, slo_val in goodput_config_dict.items():
+            if slo_name not in VALID_NAMES:
+                raise ValueError(
+                    f"Invalid metric name found, {slo_name}: {slo_val}. "
+                    "The service level objective name should be one of "
+                    f"{str(VALID_NAMES)}. ")
+            if slo_val < 0:
+                raise ValueError(
+                    f"Invalid value found, {slo_name}: {slo_val}. "
+                    "The service level objective value should be "
+                    "non-negative.")
+    return goodput_config_dict
+
+
 def main(args: argparse.Namespace):
    print(args)
    random.seed(args.seed)
@@ -660,6 +724,8 @@ def main(args: argparse.Namespace):

    input_requests = sample_requests(tokenizer, args)

+    goodput_config_dict = check_goodput_args(args)
+
    benchmark_result, ret = asyncio.run(
        benchmark(
            backend=backend,
@@ -680,6 +746,7 @@ def main(args: argparse.Namespace):
            max_concurrency=args.max_concurrency,
            guided_decoding_ratio=args.guided_decoding_ratio,
            guided_decoding_backend=args.guided_decoding_backend,
+            goodput_config_dict=goodput_config_dict,
        ))

    # Save config and results to json
@@ -730,7 +797,8 @@ if __name__ == "__main__":
        default=None,
        help="Server or API base url if not using http host and port.",
    )
-    parser.add_argument("--host", type=str, default="localhost")
+    # Use 127.0.0.1 here instead of localhost to force the use of ipv4
+    parser.add_argument("--host", type=str, default="127.0.0.1")
    parser.add_argument("--port", type=int, default=8000)
    parser.add_argument(
        "--endpoint",
@@ -863,6 +931,18 @@ if __name__ == "__main__":
        "Default value is \"99\". "
        "Use \"--percentile-metrics\" to select metrics.",
    )
+    parser.add_argument(
+        "--goodput",
+        nargs="+",
+        required=False,
+        help="Specify service level objectives for goodput as \"KEY:VALUE\" "
+        "pairs, where the key is a metric name, and the value is in "
+        "milliseconds. Multiple \"KEY:VALUE\" pairs can be provided, "
+        "separated by spaces. Allowed request level metric names are "
+        "\"ttft\", \"tpot\", \"e2el\". For more context on the definition of "
+        "goodput, refer to DistServe paper: https://arxiv.org/pdf/2401.09670 "
+        "and the blog: https://hao-ai-lab.github.io/blogs/distserve")
+
    parser.add_argument("--no-guided-decoding",
                        action='store_true',
                        default=False,
@@ -1,13 +1,17 @@
+# SPDX-License-Identifier: Apache-2.0
 """Benchmark offline inference throughput."""
 import argparse
 import dataclasses
 import json
+import os
 import random
 import time
-from typing import List, Optional
+from functools import cache
+from typing import Any, Optional

 import torch
 import uvloop
+from benchmark_utils import convert_to_pytorch_benchmark_format, write_to_json
 from PIL import Image
 from tqdm import tqdm
 from transformers import (AutoModelForCausalLM, AutoTokenizer,
@@ -17,8 +21,11 @@ from vllm.engine.arg_utils import AsyncEngineArgs, EngineArgs
 from vllm.entrypoints.openai.api_server import (
    build_async_engine_client_from_engine_args)
 from vllm.inputs import TextPrompt
+from vllm.lora.request import LoRARequest
+from vllm.lora.utils import get_adapter_absolute_path
 from vllm.multimodal import MultiModalDataDict
 from vllm.sampling_params import BeamSearchParams
+from vllm.transformers_utils.tokenizer import AnyTokenizer, get_lora_tokenizer
 from vllm.utils import FlexibleArgumentParser, merge_async_iterators


@@ -28,15 +35,17 @@ class SampleRequest:

    Attributes:
        prompt: The input text prompt for the model.
-        multi_modal_data: Optional dictionary containing multi-modal data (e.g.
-            images).
        prompt_len: The length of the prompt in tokens.
        expected_output_len: The expected length of the output in tokens.
+        multi_modal_data: Optional dictionary containing multi-modal data (e.g.
+            images).
+        lora_request: Optional LoRARequest specifying the LoRA to use. 
    """
    prompt: str
    prompt_len: int
    expected_output_len: int
    multi_modal_data: Optional[MultiModalDataDict] = None
+    lora_request: Optional[LoRARequest] = None


 def _get_prompt_for_image_model(question: str, *, model: str) -> str:
@@ -60,8 +69,30 @@ def _get_prompt_for_image_model(question: str, *, model: str) -> str:
    raise ValueError(f"Unsupported model {model}")


+@cache
+def lora_path_on_disk(lora_path: str) -> str:
+    return get_adapter_absolute_path(lora_path)
+
+
+lora_tokenizer_cache: dict[int, AnyTokenizer] = {}
+
+
+def get_random_lora_request(
+        args: argparse.Namespace
+) -> tuple[LoRARequest, Optional[AnyTokenizer]]:
+    global lora_tokenizer_cache
+    lora_id = random.randint(1, args.max_loras)
+    lora_request = LoRARequest(lora_name=str(lora_id),
+                               lora_int_id=lora_id,
+                               lora_path=lora_path_on_disk(args.lora_path))
+    if lora_id not in lora_tokenizer_cache:
+        lora_tokenizer_cache[lora_id] = get_lora_tokenizer(lora_request)
+    return lora_request, lora_tokenizer_cache[lora_id]
+
+
 def sample_requests(tokenizer: PreTrainedTokenizerBase,
-                    args: argparse.Namespace) -> List[SampleRequest]:
+                    args: argparse.Namespace) -> list[SampleRequest]:
+
    dataset_path: str = args.dataset
    num_requests: int = args.num_prompts
    fixed_output_len: Optional[int] = args.output_len
@@ -78,8 +109,10 @@ def sample_requests(tokenizer: PreTrainedTokenizerBase,
    random.shuffle(dataset)

    # Filter out sequences that are too long or too short
-    filtered_dataset: List[SampleRequest] = []
-    for data in dataset:
+    filtered_dataset: list[SampleRequest] = []
+    for data in tqdm(dataset,
+                     total=len(filtered_dataset),
+                     desc="sampling requests"):
        if len(filtered_dataset) == num_requests:
            break

@@ -102,9 +135,16 @@ def sample_requests(tokenizer: PreTrainedTokenizerBase,
                continue
            prompt = _get_prompt_for_image_model(question=prompt, model=model)

+        request_tokenizer = tokenizer
+        lora_request: Optional[LoRARequest] = None
+        if args.enable_lora:
+            lora_request, lora_tokenizer = get_random_lora_request(args)
+            if lora_tokenizer:
+                request_tokenizer = lora_tokenizer
+
        # Tokenize the prompts and completions.
-        prompt_token_ids = tokenizer(prompt).input_ids
-        completion_token_ids = tokenizer(completion).input_ids
+        prompt_token_ids = request_tokenizer(prompt).input_ids
+        completion_token_ids = request_tokenizer(completion).input_ids
        prompt_len = len(prompt_token_ids)
        output_len = len(completion_token_ids
                         ) if fixed_output_len is None else fixed_output_len
@@ -118,22 +158,28 @@ def sample_requests(tokenizer: PreTrainedTokenizerBase,
            SampleRequest(prompt=prompt,
                          prompt_len=prompt_len,
                          expected_output_len=output_len,
-                          multi_modal_data=multi_modal_data))
+                          multi_modal_data=multi_modal_data,
+                          lora_request=lora_request))

    return filtered_dataset


 def run_vllm(
-    requests: List[SampleRequest],
+    requests: list[SampleRequest],
    n: int,
    engine_args: EngineArgs,
 ) -> float:
    from vllm import LLM, SamplingParams
    llm = LLM(**dataclasses.asdict(engine_args))
-
+    assert all(
+        llm.llm_engine.model_config.max_model_len >= (
+            request.prompt_len + request.expected_output_len)
+        for request in requests), (
+            "Please ensure that max_model_len is greater than the sum of"
+            " prompt_len and expected_output_len for all requests.")
    # Add the requests to the engine.
-    prompts: List[TextPrompt] = []
-    sampling_params: List[SamplingParams] = []
+    prompts: list[TextPrompt] = []
+    sampling_params: list[SamplingParams] = []
    for request in requests:
        prompts.append(
            TextPrompt(prompt=request.prompt,
@@ -146,14 +192,21 @@ def run_vllm(
                ignore_eos=True,
                max_tokens=request.expected_output_len,
            ))
+    lora_requests: Optional[list[LoRARequest]] = None
+    if engine_args.enable_lora:
+        lora_requests = [request.lora_request for request in requests]

    use_beam_search = False

    if not use_beam_search:
        start = time.perf_counter()
-        llm.generate(prompts, sampling_params, use_tqdm=True)
+        llm.generate(prompts,
+                     sampling_params,
+                     lora_request=lora_requests,
+                     use_tqdm=True)
        end = time.perf_counter()
    else:
+        assert lora_requests is None, "BeamSearch API does not support LoRA"
        prompts = [request.prompt for request in requests]
        # output_len should be the same for all requests.
        output_len = requests[0][2]
@@ -172,7 +225,7 @@ def run_vllm(


 async def run_vllm_async(
-    requests: List[SampleRequest],
+    requests: list[SampleRequest],
    n: int,
    engine_args: AsyncEngineArgs,
    disable_frontend_multiprocessing: bool = False,
@@ -181,10 +234,17 @@ async def run_vllm_async(

    async with build_async_engine_client_from_engine_args(
            engine_args, disable_frontend_multiprocessing) as llm:
+        assert all(
+            llm.model_config.max_model_len >= (request.prompt_len +
+                                               request.expected_output_len)
+            for request in requests), (
+                "Please ensure that max_model_len is greater than the sum of"
+                " prompt_len and expected_output_len for all requests.")

        # Add the requests to the engine.
-        prompts: List[TextPrompt] = []
-        sampling_params: List[SamplingParams] = []
+        prompts: list[TextPrompt] = []
+        sampling_params: list[SamplingParams] = []
+        lora_requests: list[Optional[LoRARequest]] = []
        for request in requests:
            prompts.append(
                TextPrompt(prompt=request.prompt,
@@ -197,11 +257,16 @@ async def run_vllm_async(
                    ignore_eos=True,
                    max_tokens=request.expected_output_len,
                ))
+            lora_requests.append(request.lora_request)

        generators = []
        start = time.perf_counter()
-        for i, (prompt, sp) in enumerate(zip(prompts, sampling_params)):
-            generator = llm.generate(prompt, sp, request_id=f"test{i}")
+        for i, (prompt, sp,
+                lr) in enumerate(zip(prompts, sampling_params, lora_requests)):
+            generator = llm.generate(prompt,
+                                     sp,
+                                     lora_request=lr,
+                                     request_id=f"test{i}")
            generators.append(generator)
        all_gens = merge_async_iterators(*generators)
        async for i, res in all_gens:
@@ -211,7 +276,7 @@ async def run_vllm_async(


 def run_hf(
-    requests: List[SampleRequest],
+    requests: list[SampleRequest],
    model: str,
    tokenizer: PreTrainedTokenizerBase,
    n: int,
@@ -227,7 +292,7 @@ def run_hf(

    pbar = tqdm(total=len(requests))
    start = time.perf_counter()
-    batch: List[str] = []
+    batch: list[str] = []
    max_prompt_len = 0
    max_output_len = 0
    for i in range(len(requests)):
@@ -269,7 +334,7 @@ def run_hf(


 def run_mii(
-    requests: List[SampleRequest],
+    requests: list[SampleRequest],
    model: str,
    tensor_parallel_size: int,
    output_len: int,
@@ -286,6 +351,24 @@ def run_mii(
    return end - start


+def save_to_pytorch_benchmark_format(args: argparse.Namespace,
+                                     results: dict[str, Any]) -> None:
+    pt_records = convert_to_pytorch_benchmark_format(
+        args=args,
+        metrics={
+            "requests_per_second": [results["requests_per_second"]],
+            "tokens_per_second": [results["tokens_per_second"]],
+        },
+        extra_info={
+            k: results[k]
+            for k in ["elapsed_time", "num_requests", "total_num_tokens"]
+        })
+    if pt_records:
+        # Don't use json suffix here as we don't want CI to pick it up
+        pt_file = f"{os.path.splitext(args.output_json)[0]}.pytorch.json"
+        write_to_json(pt_file, pt_records)
+
+
 def main(args: argparse.Namespace):
    print(args)
    random.seed(args.seed)
@@ -297,6 +380,14 @@ def main(args: argparse.Namespace):
        vocab_size = tokenizer.vocab_size
        requests = []
        for _ in range(args.num_prompts):
+
+            request_tokenizer = tokenizer
+            lora_request: Optional[LoRARequest] = None
+            if args.enable_lora:
+                lora_request, lora_tokenizer = get_random_lora_request(args)
+                if lora_tokenizer:
+                    request_tokenizer = lora_tokenizer
+
            # Synthesize a prompt with the given input length.
            candidate_ids = [
                random.randint(0, vocab_size - 1)
@@ -305,8 +396,8 @@ def main(args: argparse.Namespace):
            # As tokenizer may add additional tokens like BOS, we need to try
            # different lengths to get the desired input length.
            for _ in range(5):  # Max attempts to correct
-                candidate_prompt = tokenizer.decode(candidate_ids)
-                tokenized_len = len(tokenizer.encode(candidate_prompt))
+                candidate_prompt = request_tokenizer.decode(candidate_ids)
+                tokenized_len = len(request_tokenizer.encode(candidate_prompt))

                if tokenized_len == args.input_len:
                    break
@@ -323,7 +414,8 @@ def main(args: argparse.Namespace):
            requests.append(
                SampleRequest(prompt=candidate_prompt,
                              prompt_len=args.input_len,
-                              expected_output_len=args.output_len))
+                              expected_output_len=args.output_len,
+                              lora_request=lora_request))
    else:
        requests = sample_requests(tokenizer, args)

@@ -374,6 +466,7 @@ def main(args: argparse.Namespace):
        }
        with open(args.output_json, "w") as f:
            json.dump(results, f, indent=4)
+        save_to_pytorch_benchmark_format(args, results)


 if __name__ == "__main__":
@@ -386,8 +479,8 @@ if __name__ == "__main__":
                        type=str,
                        default=None,
                        help="Path to the dataset. The dataset is expected to "
-                        "be a json in form of List[Dict[..., conversations: "
-                        "List[Dict[..., value: <prompt_or_response>]]]]")
+                        "be a json in form of list[dict[..., conversations: "
+                        "list[dict[..., value: <prompt_or_response>]]]]")
    parser.add_argument("--input-len",
                        type=int,
                        default=None,
@@ -422,6 +515,14 @@ if __name__ == "__main__":
                        action='store_true',
                        default=False,
                        help="Disable decoupled async engine frontend.")
+    # LoRA
+    parser.add_argument(
+        "--lora-path",
+        type=str,
+        default=None,
+        help="Path to the lora adapters to use. This can be an absolute path, "
+        "a relative path, or a Hugging Face model identifier.")
+
    parser = AsyncEngineArgs.add_cli_args(parser)
    args = parser.parse_args()
    if args.tokenizer is None:
@@ -431,6 +532,8 @@ if __name__ == "__main__":
        assert args.output_len is not None
    else:
        assert args.input_len is None
+    if args.enable_lora:
+        assert args.lora_path is not None

    if args.backend == "vllm":
        if args.hf_max_batch_size is not None:
@@ -440,6 +543,9 @@ if __name__ == "__main__":
            raise ValueError("HF max batch size is required for HF backend.")
        if args.quantization is not None:
            raise ValueError("Quantization is only for vLLM backend.")
+        if args.enable_lora is not None:
+            raise ValueError("LoRA benchmarking is only supported for vLLM"
+                             " backend")
    elif args.backend == "mii":
        if args.dtype != "auto":
            raise ValueError("dtype must be auto for MII backend.")
@@ -452,4 +558,7 @@ if __name__ == "__main__":
        if args.tokenizer != args.model:
            raise ValueError("Tokenizer must be the same as the model for MII "
                             "backend.")
+        if args.enable_lora is not None:
+            raise ValueError("LoRA benchmarking is only supported for vLLM"
+                             " backend")
    main(args)
@@ -0,0 +1,69 @@
+# SPDX-License-Identifier: Apache-2.0
+
+import argparse
+import json
+import math
+import os
+from typing import Any
+
+
+def convert_to_pytorch_benchmark_format(args: argparse.Namespace,
+                                        metrics: dict[str, list],
+                                        extra_info: dict[str, Any]) -> list:
+    """
+    Save the benchmark results in the format used by PyTorch OSS benchmark with
+    on metric per record
+    https://github.com/pytorch/pytorch/wiki/How-to-integrate-with-PyTorch-OSS-benchmark-database
+    """
+    records = []
+    if not os.environ.get("SAVE_TO_PYTORCH_BENCHMARK_FORMAT", False):
+        return records
+
+    for name, benchmark_values in metrics.items():
+        record = {
+            "benchmark": {
+                "name": "vLLM benchmark",
+                "extra_info": {
+                    "args": vars(args),
+                },
+            },
+            "model": {
+                "name": args.model,
+            },
+            "metric": {
+                "name": name,
+                "benchmark_values": benchmark_values,
+                "extra_info": extra_info,
+            },
+        }
+
+        tp = record["benchmark"]["extra_info"]["args"].get(
+            "tensor_parallel_size")
+        # Save tensor_parallel_size parameter if it's part of the metadata
+        if not tp and "tensor_parallel_size" in extra_info:
+            record["benchmark"]["extra_info"]["args"][
+                "tensor_parallel_size"] = extra_info["tensor_parallel_size"]
+
+        records.append(record)
+
+    return records
+
+
+class InfEncoder(json.JSONEncoder):
+
+    def clear_inf(self, o: Any):
+        if isinstance(o, dict):
+            return {k: self.clear_inf(v) for k, v in o.items()}
+        elif isinstance(o, list):
+            return [self.clear_inf(v) for v in o]
+        elif isinstance(o, float) and math.isinf(o):
+            return "inf"
+        return o
+
+    def iterencode(self, o: Any, *args, **kwargs) -> Any:
+        return super().iterencode(self.clear_inf(o), *args, **kwargs)
+
+
+def write_to_json(filename: str, records: list) -> None:
+    with open(filename, "w") as f:
+        json.dump(records, f, cls=InfEncoder)
@@ -0,0 +1,387 @@
+# SPDX-License-Identifier: Apache-2.0
+
+import argparse
+import copy
+import itertools
+import pickle as pkl
+import time
+from collections.abc import Iterable
+from typing import Callable
+
+import torch
+import torch.utils.benchmark as TBenchmark
+from torch.utils.benchmark import Measurement as TMeasurement
+from utils import make_rand_sparse_tensors
+from weight_shapes import WEIGHT_SHAPES
+
+from vllm import _custom_ops as ops
+from vllm.utils import FlexibleArgumentParser
+
+DEFAULT_MODELS = list(WEIGHT_SHAPES.keys())
+DEFAULT_BATCH_SIZES = [1, 16, 32, 64, 128, 256, 512]
+DEFAULT_TP_SIZES = [1]
+
+
+# bench
+def bench_fn(label: str, sub_label: str, description: str, fn: Callable, *args,
+             **kwargs) -> TMeasurement:
+    min_run_time = 1
+
+    globals = {
+        "args": args,
+        "kwargs": kwargs,
+        "fn": fn,
+    }
+    return TBenchmark.Timer(
+        stmt="fn(*args, **kwargs)",
+        globals=globals,
+        label=label,
+        sub_label=sub_label,
+        description=description,
+    ).blocked_autorange(min_run_time=min_run_time)
+
+
+def bench_int8(dtype: torch.dtype, m: int, k: int, n: int, label: str,
+               sub_label: str) -> Iterable[TMeasurement]:
+    assert dtype == torch.int8
+    b_compressed, e, a, b = make_rand_sparse_tensors(torch.int8, m, n, k)
+    scale_a = torch.tensor(1.0, device="cuda", dtype=torch.float32)
+    scale_b = torch.tensor(1.0, device="cuda", dtype=torch.float32)
+    bias = torch.zeros((n, ), device="cuda", dtype=torch.bfloat16)
+
+    out = ops.cutlass_scaled_sparse_mm(a, b_compressed, e, scale_a, scale_b,
+                                       torch.bfloat16)
+    out_ref = ops.cutlass_scaled_mm(a, b, scale_a, scale_b, torch.bfloat16)
+
+    if not torch.allclose(out, out_ref):
+        print("Incorrect results")
+        print(out)
+        print(out_ref)
+    else:
+        print("Correct results")
+
+    timers = []
+    # pytorch impl - bfloat16
+    timers.append(
+        bench_fn(label, sub_label, "pytorch_bf16_bf16_bf16_matmul-no-scales",
+                 torch.mm, a.to(dtype=torch.bfloat16),
+                 b.to(dtype=torch.bfloat16)))
+
+    # pytorch impl - float16
+    timers.append(
+        bench_fn(label, sub_label,
+                 "pytorch_fp16_fp16_fp16_matmul-no-scales", torch.mm,
+                 a.to(dtype=torch.float16), b.to(dtype=torch.float16)))
+
+    # cutlass impl
+    timers.append(
+        bench_fn(label, sub_label, "cutlass_i8_i8_bf16_scaled_mm",
+                 ops.cutlass_scaled_mm, a, b, scale_a, scale_b,
+                 torch.bfloat16))
+
+    # cutlass with bias
+    timers.append(
+        bench_fn(label, sub_label, "cutlass_i8_i8_bf16_scaled_mm_bias",
+                 ops.cutlass_scaled_mm, a, b, scale_a, scale_b, torch.bfloat16,
+                 bias))
+
+    # cutlass sparse impl
+    timers.append(
+        bench_fn(label, sub_label, "cutlass_i8_i8_bf16_scaled_sparse_mm",
+                 ops.cutlass_scaled_sparse_mm, a, b_compressed, e, scale_a,
+                 scale_b, torch.bfloat16))
+
+    # cutlass sparse with bias
+    timers.append(
+        bench_fn(label, sub_label, "cutlass_i8_i8_bf16_scaled_sparse_mm_bias",
+                 ops.cutlass_scaled_sparse_mm, a, b_compressed, e, scale_a,
+                 scale_b, torch.bfloat16, bias))
+
+    return timers
+
+
+def bench_fp8(dtype: torch.dtype, m: int, k: int, n: int, label: str,
+              sub_label: str) -> Iterable[TMeasurement]:
+    assert dtype == torch.float8_e4m3fn
+    b_compressed, e, a, b = make_rand_sparse_tensors(torch.float8_e4m3fn, m, n,
+                                                     k)
+    scale_a = torch.tensor(1.0, device="cuda", dtype=torch.float32)
+    scale_b = torch.tensor(1.0, device="cuda", dtype=torch.float32)
+    bias = torch.zeros((n, ), device="cuda", dtype=torch.bfloat16)
+
+    out = ops.cutlass_scaled_sparse_mm(a, b_compressed, e, scale_a, scale_b,
+                                       torch.bfloat16)
+    out_ref = ops.cutlass_scaled_mm(a, b, scale_a, scale_b, torch.bfloat16)
+
+    if not torch.allclose(out, out_ref):
+        print("Incorrect results")
+        print(out)
+        print(out_ref)
+    else:
+        print("Correct results")
+
+    timers = []
+
+    # pytorch impl w. bf16
+    timers.append(
+        bench_fn(label, sub_label, "pytorch_bf16_bf16_bf16_matmul-no-scales",
+                 torch.mm, a.to(dtype=torch.bfloat16, device="cuda"),
+                 b.to(dtype=torch.bfloat16, device="cuda")))
+
+    # pytorch impl: bf16 output, without fp8 fast accum
+    timers.append(
+        bench_fn(label,
+                 sub_label,
+                 "pytorch_fp8_fp8_bf16_scaled_mm",
+                 torch._scaled_mm,
+                 a,
+                 b,
+                 scale_a=scale_a,
+                 scale_b=scale_b,
+                 out_dtype=torch.bfloat16))
+
+    # pytorch impl: bf16 output, with fp8 fast accum
+    timers.append(
+        bench_fn(label,
+                 sub_label,
+                 "pytorch_fp8_fp8_bf16_scaled_mm_fast_accum",
+                 torch._scaled_mm,
+                 a,
+                 b,
+                 scale_a=scale_a,
+                 scale_b=scale_b,
+                 out_dtype=torch.bfloat16,
+                 use_fast_accum=True))
+
+    # pytorch impl: fp16 output, without fp8 fast accum
+    timers.append(
+        bench_fn(label,
+                 sub_label,
+                 "pytorch_fp8_fp8_fp16_scaled_mm",
+                 torch._scaled_mm,
+                 a,
+                 b,
+                 scale_a=scale_a,
+                 scale_b=scale_b,
+                 out_dtype=torch.float16))
+
+    # pytorch impl: fp16 output, with fp8 fast accum
+    timers.append(
+        bench_fn(label,
+                 sub_label,
+                 "pytorch_fp8_fp8_fp16_scaled_mm_fast_accum",
+                 torch._scaled_mm,
+                 a,
+                 b,
+                 scale_a=scale_a,
+                 scale_b=scale_b,
+                 out_dtype=torch.float16,
+                 use_fast_accum=True))
+
+    # cutlass impl: bf16 output
+    timers.append(
+        bench_fn(label, sub_label, "cutlass_fp8_fp8_bf16_scaled_mm",
+                 ops.cutlass_scaled_mm, a, b, scale_a, scale_b,
+                 torch.bfloat16))
+
+    # cutlass impl: bf16 output
+    timers.append(
+        bench_fn(label, sub_label, "cutlass_fp8_fp8_bf16_scaled_sparse_mm",
+                 ops.cutlass_scaled_sparse_mm, a, b_compressed, e, scale_a,
+                 scale_b, torch.bfloat16))
+
+    # cutlass impl: fp16 output
+    timers.append(
+        bench_fn(label, sub_label, "cutlass_fp8_fp8_fp16_scaled_sparse_mm",
+                 ops.cutlass_scaled_sparse_mm, a, b_compressed, e, scale_a,
+                 scale_b, torch.float16))
+
+    # cutlass impl: bf16 output, with bias
+    timers.append(
+        bench_fn(label, sub_label,
+                 "cutlass_fp8_fp8_bf16_scaled_sparse_mm_bias",
+                 ops.cutlass_scaled_sparse_mm, a, b_compressed, e, scale_a,
+                 scale_b, torch.bfloat16, bias))
+
+    # cutlass impl: fp16 output, with bias
+    timers.append(
+        bench_fn(label, sub_label,
+                 "cutlass_fp8_fp8_fp16_scaled_sparse_mm_bias",
+                 ops.cutlass_scaled_sparse_mm, a, b_compressed, e, scale_a,
+                 scale_b, torch.float16, bias.to(dtype=torch.float16)))
+
+    return timers
+
+
+def bench(dtype: torch.dtype, m: int, k: int, n: int, label: str,
+          sub_label: str) -> Iterable[TMeasurement]:
+    if dtype == torch.int8:
+        return bench_int8(dtype, m, k, n, label, sub_label)
+    if dtype == torch.float8_e4m3fn:
+        return bench_fp8(dtype, m, k, n, label, sub_label)
+    raise ValueError("unsupported type")
+
+
+# runner
+def print_timers(timers: Iterable[TMeasurement]):
+    compare = TBenchmark.Compare(timers)
+    compare.print()
+
+
+def run(dtype: torch.dtype,
+        MKNs: Iterable[tuple[int, int, int]]) -> Iterable[TMeasurement]:
+    results = []
+    for m, k, n in MKNs:
+        timers = bench(dtype, m, k, n, f"scaled-{dtype}-gemm",
+                       f"MKN=({m}x{k}x{n})")
+        print_timers(timers)
+        results.extend(timers)
+
+    return results
+
+
+# output makers
+def make_output(data: Iterable[TMeasurement],
+                MKNs: Iterable[tuple[int, int, int]],
+                base_description: str,
+                timestamp=None):
+    print(f"== All Results {base_description} ====")
+    print_timers(data)
+
+    # pickle all the results
+    timestamp = int(time.time()) if timestamp is None else timestamp
+    with open(f"{base_description}-{timestamp}.pkl", "wb") as f:
+        pkl.dump(data, f)
+
+
+# argparse runners
+
+
+def run_square_bench(args):
+    dim_sizes = list(
+        range(args.dim_start, args.dim_end + 1, args.dim_increment))
+    MKNs = list(zip(dim_sizes, dim_sizes, dim_sizes))
+    data = run(args.dtype, MKNs)
+
+    make_output(data, MKNs, f"square_bench-{args.dtype}")
+
+
+def run_range_bench(args):
+    dim_sizes = list(range(args.dim_start, args.dim_end, args.dim_increment))
+    n = len(dim_sizes)
+    Ms = [args.m_constant] * n if args.m_constant is not None else dim_sizes
+    Ks = [args.k_constant] * n if args.k_constant is not None else dim_sizes
+    Ns = [args.n_constant] * n if args.n_constant is not None else dim_sizes
+    MKNs = list(zip(Ms, Ks, Ns))
+    data = run(args.dtype, MKNs)
+
+    make_output(data, MKNs, f"range_bench-{args.dtype}")
+
+
+def run_model_bench(args):
+    print("Benchmarking models:")
+    for i, model in enumerate(args.models):
+        print(f"[{i}]  {model}")
+
+    def model_shapes(model_name: str, tp_size: int) -> list[tuple[int, int]]:
+        KNs = []
+        for KN, tp_split_dim in copy.deepcopy(WEIGHT_SHAPES[model_name]):
+            KN[tp_split_dim] = KN[tp_split_dim] // tp_size
+            KNs.append(KN)
+        return KNs
+
+    model_bench_data = []
+    models_tps = list(itertools.product(args.models, args.tp_sizes))
+    for model, tp_size in models_tps:
+        Ms = args.batch_sizes
+        KNs = model_shapes(model, tp_size)
+        MKNs = []
+        for m in Ms:
+            for k, n in KNs:
+                MKNs.append((m, k, n))
+
+        data = run(args.dtype, MKNs)
+        model_bench_data.append(data)
+
+    # Print all results
+    for data, model_tp in zip(model_bench_data, models_tps):
+        model, tp_size = model_tp
+        print(f"== Results {args.dtype} {model}-TP{tp_size} ====")
+        print_timers(data)
+
+    timestamp = int(time.time())
+
+    all_data = []
+    for d in model_bench_data:
+        all_data.extend(d)
+    # pickle all data
+    with open(f"model_bench-{args.dtype}-{timestamp}.pkl", "wb") as f:
+        pkl.dump(all_data, f)
+
+
+if __name__ == '__main__':
+
+    def to_torch_dtype(dt):
+        if dt == "int8":
+            return torch.int8
+        if dt == "fp8":
+            return torch.float8_e4m3fn
+        raise ValueError("unsupported dtype")
+
+    parser = FlexibleArgumentParser(
+        description="""
+Benchmark Cutlass GEMM.
+
+    To run square GEMMs:
+        python3 ./benchmarks/cutlass_benchmarks/sparse_benchmarks.py --dtype fp8 square_bench --dim-start 128 --dim-end 512 --dim-increment 64
+    
+    To run constant N and K and sweep M:
+        python3 ./benchmarks/cutlass_benchmarks/sparse_benchmarks.py --dtype fp8 range_bench --dim-start 128 --dim-end 512 --dim-increment 64 --n-constant 16384 --k-constant 16384
+    
+    To run dimensions from a model:
+        python3 ./benchmarks/cutlass_benchmarks/sparse_benchmarks.py --dtype fp8 model_bench --models meta-llama/Llama-2-7b-hf --batch-sizes 16 --tp-sizes 1
+    
+    Output:
+        - a .pkl file, that is a list of raw torch.benchmark.utils.Measurements for the pytorch and cutlass implementations for the various GEMMs.
+            """,  # noqa: E501
+        formatter_class=argparse.RawTextHelpFormatter)
+
+    parser.add_argument("--dtype",
+                        type=to_torch_dtype,
+                        required=True,
+                        help="Available options are ['int8', 'fp8']")
+    subparsers = parser.add_subparsers(dest="cmd")
+
+    square_parser = subparsers.add_parser("square_bench")
+    square_parser.add_argument("--dim-start", type=int, required=True)
+    square_parser.add_argument("--dim-end", type=int, required=True)
+    square_parser.add_argument("--dim-increment", type=int, required=True)
+    square_parser.set_defaults(func=run_square_bench)
+
+    range_parser = subparsers.add_parser("range_bench")
+    range_parser.add_argument("--dim-start", type=int, required=True)
+    range_parser.add_argument("--dim-end", type=int, required=True)
+    range_parser.add_argument("--dim-increment", type=int, required=True)
+    range_parser.add_argument("--m-constant", type=int, default=None)
+    range_parser.add_argument("--n-constant", type=int, default=None)
+    range_parser.add_argument("--k-constant", type=int, default=None)
+    range_parser.set_defaults(func=run_range_bench)
+
+    model_parser = subparsers.add_parser("model_bench")
+    model_parser.add_argument("--models",
+                              nargs="+",
+                              type=str,
+                              default=DEFAULT_MODELS,
+                              choices=WEIGHT_SHAPES.keys())
+    model_parser.add_argument("--tp-sizes",
+                              nargs="+",
+                              type=int,
+                              default=DEFAULT_TP_SIZES)
+    model_parser.add_argument("--batch-sizes",
+                              nargs="+",
+                              type=int,
+                              default=DEFAULT_BATCH_SIZES)
+    model_parser.set_defaults(func=run_model_bench)
+
+    args = parser.parse_args()
+    args.func(args)
@@ -0,0 +1,98 @@
+# SPDX-License-Identifier: Apache-2.0
+
+# Cutlass bench utils
+from collections.abc import Iterable
+
+import torch
+
+import vllm._custom_ops as ops
+
+
+def to_fp8(tensor: torch.Tensor) -> torch.Tensor:
+    finfo = torch.finfo(torch.float8_e4m3fn)
+    return torch.round(tensor.clamp(
+        min=finfo.min, max=finfo.max)).to(dtype=torch.float8_e4m3fn)
+
+
+def to_int8(tensor: torch.Tensor) -> torch.Tensor:
+    return torch.round(tensor.clamp(min=-128, max=127)).to(dtype=torch.int8)
+
+
+def to_bf16(tensor: torch.Tensor) -> torch.Tensor:
+    return tensor.to(dtype=torch.bfloat16)
+
+
+def to_fp16(tensor: torch.Tensor) -> torch.Tensor:
+    return tensor.to(dtype=torch.float16)
+
+
+def make_rand_tensors(dtype: torch.dtype, m: int, n: int,
+                      k: int) -> tuple[torch.Tensor, torch.Tensor]:
+    a = torch.randn((m, k), device='cuda') * 5
+    b = torch.randn((n, k), device='cuda').t() * 5
+
+    if dtype == torch.int8:
+        return to_int8(a), to_int8(b)
+    if dtype == torch.float8_e4m3fn:
+        return to_fp8(a), to_fp8(b)
+
+    raise ValueError("unsupported dtype")
+
+
+def prune_to_2_4(tensor):
+    # Reshape tensor to [N, 4] where N is number of groups of 4
+    original_shape = tensor.shape
+    reshaped = tensor.reshape(-1, 4)
+
+    # Get indices of top 2 absolute values in each group of 4
+    _, indices = torch.topk(torch.abs(reshaped), k=2, dim=1)
+
+    # Create binary mask
+    mask = torch.zeros_like(reshaped)
+    mask.scatter_(dim=1,
+                  index=indices,
+                  src=torch.ones_like(indices, dtype=mask.dtype))
+
+    # Apply mask and reshape back
+    pruned = reshaped * mask
+
+    # Turn all -0.0 to 0.0
+    pruned[pruned == -0.0] = 0.0
+
+    return pruned.reshape(original_shape)
+
+
+def make_rand_sparse_tensors(dtype: torch.dtype, m: int, n: int,
+                             k: int) -> tuple[torch.Tensor, torch.Tensor]:
+    a = torch.randn((m, k), device='cuda') * 5
+    b = torch.randn((n, k), device='cuda').t() * 5
+
+    b = prune_to_2_4(b.t()).t()
+
+    if dtype == torch.int8:
+        a, b = to_int8(a), to_int8(b)
+    elif dtype == torch.float8_e4m3fn:
+        a, b = to_fp8(a), to_fp8(b)
+    elif dtype == torch.float16:
+        a, b = to_fp16(a), to_fp16(b)
+    elif dtype == torch.bfloat16:
+        a, b = to_bf16(a), to_bf16(b)
+    else:
+        raise ValueError("unsupported dtype")
+
+    b_compressed, e = ops.cutlass_sparse_compress(b.t())
+
+    # Compressed B, Metadata, Original A, B
+    return b_compressed, e, a, b
+
+
+def make_n_rand_sparse_tensors(num_tensors: int, dtype: torch.dtype,
+                        m: int, n: int, k: int) -> \
+                        tuple[Iterable[torch.Tensor], Iterable[torch.Tensor]]:
+    ABs = []
+    for _ in range(num_tensors):
+        b_comp, e, a, b = make_rand_sparse_tensors(dtype, m, n, k)
+        if b_comp is not None:
+            ABs.append(make_rand_sparse_tensors(dtype, m, n, k))
+    BComps, Es, As, Bs = zip(*ABs)
+    return list(BComps), list(Es), list(As), list(Bs)
@@ -1,47 +1,28 @@
+# SPDX-License-Identifier: Apache-2.0
+
 import argparse
 import copy
 import itertools
 import pickle as pkl
 import time
-from typing import Callable, Iterable, List, Tuple
+from collections.abc import Iterable
+from typing import Callable, Optional

 import torch
 import torch.utils.benchmark as TBenchmark
 from torch.utils.benchmark import Measurement as TMeasurement
+from utils import make_rand_tensors
 from weight_shapes import WEIGHT_SHAPES

 from vllm import _custom_ops as ops
+from vllm.model_executor.layers.quantization.utils.fp8_utils import (
+    w8a8_block_fp8_matmul)
 from vllm.utils import FlexibleArgumentParser

 DEFAULT_MODELS = list(WEIGHT_SHAPES.keys())
 DEFAULT_BATCH_SIZES = [1, 16, 32, 64, 128, 256, 512]
 DEFAULT_TP_SIZES = [1]

-# helpers
-
-
-def to_fp8(tensor: torch.Tensor) -> torch.Tensor:
-    finfo = torch.finfo(torch.float8_e4m3fn)
-    return torch.round(tensor.clamp(
-        min=finfo.min, max=finfo.max)).to(dtype=torch.float8_e4m3fn)
-
-
-def to_int8(tensor: torch.Tensor) -> torch.Tensor:
-    return torch.round(tensor.clamp(min=-128, max=127)).to(dtype=torch.int8)
-
-
-def make_rand_tensors(dtype: torch.dtype, m: int, n: int,
-                      k: int) -> Tuple[torch.Tensor, torch.Tensor]:
-    a = torch.randn((m, k), device='cuda') * 5
-    b = torch.randn((n, k), device='cuda').t() * 5
-
-    if dtype == torch.int8:
-        return to_int8(a), to_int8(b)
-    if dtype == torch.float8_e4m3fn:
-        return to_fp8(a), to_fp8(b)
-
-    raise ValueError("unsupported dtype")
-

 # bench
 def bench_fn(label: str, sub_label: str, description: str, fn: Callable, *args,
@@ -62,8 +43,15 @@ def bench_fn(label: str, sub_label: str, description: str, fn: Callable, *args,
    ).blocked_autorange(min_run_time=min_run_time)


-def bench_int8(dtype: torch.dtype, m: int, k: int, n: int, label: str,
-               sub_label: str) -> Iterable[TMeasurement]:
+def bench_int8(
+        dtype: torch.dtype,
+        m: int,
+        k: int,
+        n: int,
+        label: str,
+        sub_label: str,
+        bench_kernels: Optional[list[str]] = None) -> Iterable[TMeasurement]:
+    """Benchmark INT8-based kernels."""
    assert dtype == torch.int8
    a, b = make_rand_tensors(torch.int8, m, n, k)
    scale_a = torch.tensor(1.0, device="cuda", dtype=torch.float32)
@@ -72,155 +60,132 @@ def bench_int8(dtype: torch.dtype, m: int, k: int, n: int, label: str,
    azp = torch.zeros((m, ), device="cuda", dtype=torch.int32)
    azp_adj = torch.zeros((n, ), device="cuda", dtype=torch.int32)

+    bench_fns = {
+        "pytorch_bf16_bf16_bf16_matmul-no-scales":
+        lambda: torch.mm(a.to(dtype=torch.bfloat16), b.to(dtype=torch.bfloat16)
+                         ),
+        "pytorch_fp16_fp16_fp16_matmul-no-scales":
+        lambda: torch.mm(a.to(dtype=torch.float16), b.to(dtype=torch.float16)),
+        "cutlass_i8_i8_bf16_scaled_mm":
+        lambda: ops.cutlass_scaled_mm(a, b, scale_a, scale_b, torch.bfloat16),
+        "cutlass_i8_i8_bf16_scaled_mm_bias":
+        lambda: ops.cutlass_scaled_mm(a, b, scale_a, scale_b, torch.bfloat16,
+                                      bias),
+        "cutlass_i8_i8_bf16_scaled_mm_azp":
+        lambda: ops.cutlass_scaled_mm_azp(a, b, scale_a, scale_b, torch.
+                                          bfloat16, azp_adj),
+        "cutlass_i8_i8_bf16_scaled_mm_azp_bias":
+        lambda: ops.cutlass_scaled_mm_azp(a, b, scale_a, scale_b, torch.
+                                          bfloat16, azp_adj, None, bias),
+        "cutlass_i8_i8_bf16_scaled_mm_azp_pt":
+        lambda: ops.cutlass_scaled_mm_azp(a, b, scale_a, scale_b, torch.
+                                          bfloat16, azp_adj, azp),
+        "cutlass_i8_i8_bf16_scaled_mm_azp_pt_bias":
+        lambda: ops.cutlass_scaled_mm_azp(a, b, scale_a, scale_b, torch.
+                                          bfloat16, azp_adj, azp, bias),
+    }
+
    timers = []
-    # pytorch impl - bfloat16
-    timers.append(
-        bench_fn(label, sub_label, "pytorch_bf16_bf16_bf16_matmul-no-scales",
-                 torch.mm, a.to(dtype=torch.bfloat16),
-                 b.to(dtype=torch.bfloat16)))
-
-    # pytorch impl - float16
-    timers.append(
-        bench_fn(label, sub_label,
-                 "pytorch_fp16_fp16_fp16_matmul-no-scales", torch.mm,
-                 a.to(dtype=torch.float16), b.to(dtype=torch.float16)))
-
-    # cutlass impl
-    timers.append(
-        bench_fn(label, sub_label, "cutlass_i8_i8_bf16_scaled_mm",
-                 ops.cutlass_scaled_mm, a, b, scale_a, scale_b,
-                 torch.bfloat16))
-
-    # cutlass with bias
-    timers.append(
-        bench_fn(label, sub_label, "cutlass_i8_i8_bf16_scaled_mm_bias",
-                 ops.cutlass_scaled_mm, a, b, scale_a, scale_b, torch.bfloat16,
-                 bias))
-
-    # cutlass with azp per-tensor
-    timers.append(
-        bench_fn(label, sub_label, "cutlass_i8_i8_bf16_scaled_mm_azp",
-                 ops.cutlass_scaled_mm_azp, a, b, scale_a, scale_b,
-                 torch.bfloat16, azp_adj))
-
-    # cutlass with azp per-tensor + bias
-    timers.append(
-        bench_fn(label, sub_label, "cutlass_i8_i8_bf16_scaled_mm_azp_bias",
-                 ops.cutlass_scaled_mm_azp, a, b, scale_a, scale_b,
-                 torch.bfloat16, azp_adj, None, bias))
-
-    # cutlass with azp per-token
-    timers.append(
-        bench_fn(label, sub_label, "cutlass_i8_i8_bf16_scaled_mm_azp_pt",
-                 ops.cutlass_scaled_mm_azp, a, b, scale_a, scale_b,
-                 torch.bfloat16, azp_adj, azp))
-
-    # cutlass with azp per-token + bias
-    timers.append(
-        bench_fn(label, sub_label, "cutlass_i8_i8_bf16_scaled_mm_azp_pt_bias",
-                 ops.cutlass_scaled_mm_azp, a, b, scale_a, scale_b,
-                 torch.bfloat16, azp_adj, azp, bias))
+    for name, fn in bench_fns.items():
+        # If bench_kernels is None, run all. Otherwise, run only exact matches.
+        if bench_kernels is None or name in bench_kernels:
+            print(f"Running {name}")
+            timers.append(bench_fn(label, sub_label, name, fn))

    return timers


-def bench_fp8(dtype: torch.dtype, m: int, k: int, n: int, label: str,
-              sub_label: str) -> Iterable[TMeasurement]:
+def bench_fp8(
+        dtype: torch.dtype,
+        m: int,
+        k: int,
+        n: int,
+        label: str,
+        sub_label: str,
+        bench_kernels: Optional[list[str]] = None) -> Iterable[TMeasurement]:
+    """Benchmark FP8-based kernels."""
    assert dtype == torch.float8_e4m3fn
    a, b = make_rand_tensors(torch.float8_e4m3fn, m, n, k)
+    a_cont = a.contiguous()
    scale_a = torch.tensor(1.0, device="cuda", dtype=torch.float32)
    scale_b = torch.tensor(1.0, device="cuda", dtype=torch.float32)
+    block_scale_a = torch.rand((m, k // 128),
+                               device="cuda",
+                               dtype=torch.float32)
+    block_scale_b = torch.rand((k // 128, n // 128),
+                               device="cuda",
+                               dtype=torch.float32)
+    block_scale_a_M_major = block_scale_a.t().contiguous().t()
+    block_scale_b_K_major = block_scale_b.t().contiguous().t()
    bias = torch.zeros((n, ), device="cuda", dtype=torch.bfloat16)

+    print(m, k, n)
+
+    bench_fns = {
+        "pytorch_bf16_bf16_bf16_matmul-no-scales":
+        lambda: torch.mm(a.to(dtype=torch.bfloat16), b.to(dtype=torch.bfloat16)
+                         ),
+        "pytorch_fp16_fp16_fp16_matmul-no-scales":
+        lambda: torch.mm(a.to(dtype=torch.float16), b.to(dtype=torch.float16)),
+        "pytorch_fp8_fp8_fp16_scaled_mm":
+        lambda: torch._scaled_mm(
+            a, b, scale_a, scale_b, out_dtype=torch.float16),
+        "pytorch_fp8_fp8_fp16_scaled_mm_fast_accum":
+        lambda: torch._scaled_mm(a,
+                                 b,
+                                 scale_a,
+                                 scale_b,
+                                 out_dtype=torch.float16,
+                                 use_fast_accum=True),
+        "pytorch_fp8_fp8_bf16_scaled_mm":
+        lambda: torch._scaled_mm(
+            a, b, scale_a, scale_b, out_dtype=torch.bfloat16),
+        "pytorch_fp8_fp8_bf16_scaled_mm_fast_accum":
+        lambda: torch._scaled_mm(a,
+                                 b,
+                                 scale_a,
+                                 scale_b,
+                                 out_dtype=torch.bfloat16,
+                                 use_fast_accum=True),
+        "cutlass_fp8_fp8_bf16_scaled_mm":
+        lambda: ops.cutlass_scaled_mm(a, b, scale_a, scale_b, torch.bfloat16),
+        "cutlass_fp8_fp8_fp16_scaled_mm":
+        lambda: ops.cutlass_scaled_mm(a, b, scale_a, scale_b, torch.float16),
+        "cutlass_fp8_fp8_bf16_scaled_mm_bias":
+        lambda: ops.cutlass_scaled_mm(a, b, scale_a, scale_b, torch.bfloat16,
+                                      bias),
+        "cutlass_fp8_fp8_fp16_scaled_mm_bias":
+        lambda: ops.cutlass_scaled_mm(a, b, scale_a, scale_b, torch.float16,
+                                      bias.to(dtype=torch.float16)),
+        "triton_fp8_fp8_fp16_scaled_mm_blockwise":
+        lambda: w8a8_block_fp8_matmul(a_cont, b.t(), block_scale_a,
+                                      block_scale_b.t(), (128, 128)),
+        "cutlass_fp8_fp8_fp16_scaled_mm_blockwise":
+        lambda: ops.cutlass_scaled_mm(a, b, block_scale_a_M_major,
+                                      block_scale_b_K_major, torch.float16),
+    }
+
    timers = []
-
-    # pytorch impl w. bf16
-    timers.append(
-        bench_fn(label, sub_label, "pytorch_bf16_bf16_bf16_matmul-no-scales",
-                 torch.mm, a.to(dtype=torch.bfloat16, device="cuda"),
-                 b.to(dtype=torch.bfloat16, device="cuda")))
-
-    # pytorch impl: bf16 output, without fp8 fast accum
-    timers.append(
-        bench_fn(label,
-                 sub_label,
-                 "pytorch_fp8_fp8_bf16_scaled_mm",
-                 torch._scaled_mm,
-                 a,
-                 b,
-                 scale_a=scale_a,
-                 scale_b=scale_b,
-                 out_dtype=torch.bfloat16))
-
-    # pytorch impl: bf16 output, with fp8 fast accum
-    timers.append(
-        bench_fn(label,
-                 sub_label,
-                 "pytorch_fp8_fp8_bf16_scaled_mm_fast_accum",
-                 torch._scaled_mm,
-                 a,
-                 b,
-                 scale_a=scale_a,
-                 scale_b=scale_b,
-                 out_dtype=torch.bfloat16,
-                 use_fast_accum=True))
-
-    # pytorch impl: fp16 output, without fp8 fast accum
-    timers.append(
-        bench_fn(label,
-                 sub_label,
-                 "pytorch_fp8_fp8_fp16_scaled_mm",
-                 torch._scaled_mm,
-                 a,
-                 b,
-                 scale_a=scale_a,
-                 scale_b=scale_b,
-                 out_dtype=torch.float16))
-
-    # pytorch impl: fp16 output, with fp8 fast accum
-    timers.append(
-        bench_fn(label,
-                 sub_label,
-                 "pytorch_fp8_fp8_fp16_scaled_mm_fast_accum",
-                 torch._scaled_mm,
-                 a,
-                 b,
-                 scale_a=scale_a,
-                 scale_b=scale_b,
-                 out_dtype=torch.float16,
-                 use_fast_accum=True))
-
-    # cutlass impl: bf16 output
-    timers.append(
-        bench_fn(label, sub_label, "cutlass_fp8_fp8_bf16_scaled_mm",
-                 ops.cutlass_scaled_mm, a, b, scale_a, scale_b,
-                 torch.bfloat16))
-    # cutlass impl: fp16 output
-    timers.append(
-        bench_fn(label, sub_label, "cutlass_fp8_fp8_fp16_scaled_mm",
-                 ops.cutlass_scaled_mm, a, b, scale_a, scale_b, torch.float16))
-
-    # cutlass impl: bf16 output, with bias
-    timers.append(
-        bench_fn(label, sub_label, "cutlass_fp8_fp8_bf16_scaled_mm_bias",
-                 ops.cutlass_scaled_mm, a, b, scale_a, scale_b, torch.bfloat16,
-                 bias))
-
-    # cutlass impl: fp16 output, with bias
-    timers.append(
-        bench_fn(label, sub_label, "cutlass_fp8_fp8_fp16_scaled_mm_bias",
-                 ops.cutlass_scaled_mm, a, b, scale_a, scale_b, torch.float16,
-                 bias.to(dtype=torch.float16)))
+    for name, fn in bench_fns.items():
+        # If bench_kernels is None, run all. Otherwise, run only exact matches.
+        if bench_kernels is None or name in bench_kernels:
+            print(f"Running {name}")
+            timers.append(bench_fn(label, sub_label, name, fn))

    return timers


-def bench(dtype: torch.dtype, m: int, k: int, n: int, label: str,
-          sub_label: str) -> Iterable[TMeasurement]:
+def bench(dtype: torch.dtype,
+          m: int,
+          k: int,
+          n: int,
+          label: str,
+          sub_label: str,
+          bench_kernels: Optional[list[str]] = None) -> Iterable[TMeasurement]:
    if dtype == torch.int8:
-        return bench_int8(dtype, m, k, n, label, sub_label)
+        return bench_int8(dtype, m, k, n, label, sub_label, bench_kernels)
    if dtype == torch.float8_e4m3fn:
-        return bench_fp8(dtype, m, k, n, label, sub_label)
+        return bench_fp8(dtype, m, k, n, label, sub_label, bench_kernels)
    raise ValueError("unsupported type")


@@ -231,20 +196,24 @@ def print_timers(timers: Iterable[TMeasurement]):


 def run(dtype: torch.dtype,
-        MKNs: Iterable[Tuple[int, int, int]]) -> Iterable[TMeasurement]:
+        MKNs: Iterable[tuple[int, int, int]],
+        bench_kernels: Optional[list[str]] = None) -> Iterable[TMeasurement]:
    results = []
    for m, k, n in MKNs:
-        timers = bench(dtype, m, k, n, f"scaled-{dtype}-gemm",
-                       f"MKN=({m}x{k}x{n})")
+        timers = bench(dtype,
+                       m,
+                       k,
+                       n,
+                       f"scaled-{dtype}-gemm",
+                       f"MKN=({m}x{k}x{n})",
+                       bench_kernels=bench_kernels)
        print_timers(timers)
        results.extend(timers)
-
    return results


-# output makers
 def make_output(data: Iterable[TMeasurement],
-                MKNs: Iterable[Tuple[int, int, int]],
+                MKNs: Iterable[tuple[int, int, int]],
                base_description: str,
                timestamp=None):
    print(f"== All Results {base_description} ====")
@@ -256,15 +225,11 @@ def make_output(data: Iterable[TMeasurement],
        pkl.dump(data, f)


-# argparse runners
-
-
 def run_square_bench(args):
    dim_sizes = list(
        range(args.dim_start, args.dim_end + 1, args.dim_increment))
    MKNs = list(zip(dim_sizes, dim_sizes, dim_sizes))
-    data = run(args.dtype, MKNs)
-
+    data = run(args.dtype, MKNs, bench_kernels=args.kernels)
    make_output(data, MKNs, f"square_bench-{args.dtype}")


@@ -275,8 +240,7 @@ def run_range_bench(args):
    Ks = [args.k_constant] * n if args.k_constant is not None else dim_sizes
    Ns = [args.n_constant] * n if args.n_constant is not None else dim_sizes
    MKNs = list(zip(Ms, Ks, Ns))
-    data = run(args.dtype, MKNs)
-
+    data = run(args.dtype, MKNs, bench_kernels=args.kernels)
    make_output(data, MKNs, f"range_bench-{args.dtype}")


@@ -285,7 +249,7 @@ def run_model_bench(args):
    for i, model in enumerate(args.models):
        print(f"[{i}]  {model}")

-    def model_shapes(model_name: str, tp_size: int) -> List[Tuple[int, int]]:
+    def model_shapes(model_name: str, tp_size: int) -> list[tuple[int, int]]:
        KNs = []
        for KN, tp_split_dim in copy.deepcopy(WEIGHT_SHAPES[model_name]):
            KN[tp_split_dim] = KN[tp_split_dim] // tp_size
@@ -302,7 +266,7 @@ def run_model_bench(args):
            for k, n in KNs:
                MKNs.append((m, k, n))

-        data = run(args.dtype, MKNs)
+        data = run(args.dtype, MKNs, bench_kernels=args.kernels)
        model_bench_data.append(data)

    # Print all results
@@ -352,6 +316,15 @@ Benchmark Cutlass GEMM.
                        type=to_torch_dtype,
                        required=True,
                        help="Available options are ['int8', 'fp8']")
+    parser.add_argument(
+        "--kernels",
+        nargs="+",
+        type=str,
+        default=None,
+        help=
+        "Exact names of the kernels to benchmark. If not set, runs all kernels."
+    )
+
    subparsers = parser.add_subparsers(dest="cmd")

    square_parser = subparsers.add_parser("square_bench")
@@ -1,3 +1,5 @@
+# SPDX-License-Identifier: Apache-2.0
+
 # Weight Shapes are in the format
 # ([K, N], TP_SPLIT_DIM)
 # Example:
@@ -40,4 +42,4 @@ WEIGHT_SHAPES = {
        ([8192, 57344], 1),
        ([28672, 8192], 0),
    ],
-}
+}
@@ -10,7 +10,8 @@ set -ex

 kill_gpu_processes() {
  # kill all processes on GPU.
-  pkill -f pt_main_thread
+  pgrep pt_main_thread | xargs -r kill -9
+  pgrep python3 | xargs -r kill -9
  sleep 10

  # remove vllm config file
@@ -54,7 +55,7 @@ benchmark() {

  CUDA_VISIBLE_DEVICES=0 python3 \
    -m vllm.entrypoints.openai.api_server \
-    --model meta-llama/Meta-Llama-3.1-8B-Instruct \
+    --model $model \
    --port 8100 \
    --max-model-len 10000 \
    --gpu-memory-utilization 0.6 \
@@ -64,7 +65,7 @@ benchmark() {

  CUDA_VISIBLE_DEVICES=1 python3 \
    -m vllm.entrypoints.openai.api_server \
-    --model meta-llama/Meta-Llama-3.1-8B-Instruct \
+    --model $model \
    --port 8200 \
    --max-model-len 10000 \
    --gpu-memory-utilization 0.6 \
@@ -87,7 +88,7 @@ benchmark() {
          --port 8100 \
          --save-result \
          --result-dir $results_folder \
-          --result-filename disagg_prefill_2xtp4.json \
+          --result-filename disagg_prefill_tp1.json \
          --request-rate "inf"


@@ -105,7 +106,7 @@ benchmark() {
          --port 8200 \
          --save-result \
          --result-dir $results_folder \
-          --result-filename disagg_prefill_2xtp4.json \
+          --result-filename disagg_prefill_tp1_overhead.json \
          --request-rate "$qps"
  kill_gpu_processes

@@ -118,7 +119,7 @@ main() {
  (which jq) || (apt-get -y install jq)
  (which socat) || (apt-get -y install socat)

-  pip install quart httpx
+  pip install quart httpx datasets

  cd "$(dirname "$0")"

@@ -1,13 +1,12 @@
 #!/bin/bash

-# Requirement: 8x H100 GPUs.
+# Requirement: 2x GPUs.


-# Model: neuralmagic/Meta-Llama-3-70B-Instruct-FP8-KV 
-# Query: 2048 input tokens, 11 output tokens, QPS 4, 500 requests
-# Resource: 8x H100
+# Model: meta-llama/Meta-Llama-3.1-8B-Instruct
+# Query: 1024 input tokens, 6 output tokens, QPS 2/4/6/8, 100 requests
+# Resource: 2x GPU
 # Approaches:
-# 1. Chunked prefill: 1 vllm instance with tp=8
 # 2. Chunked prefill: 2 vllm instance with tp=4, equivalent to 1 tp=4 instance with QPS 4
 # 3. Disaggregated prefill: 1 prefilling instance and 1 decoding instance
 # Prefilling instance: max_output_token=1
@@ -114,7 +113,6 @@ benchmark() {
          --request-rate "$qps"

  sleep 2
-
 }


@@ -123,8 +121,9 @@ main() {
  (which wget && which curl) || (apt-get update && apt-get install -y wget curl)
  (which jq) || (apt-get -y install jq)
  (which socat) || (apt-get -y install socat)
+  (which lsof) || (apt-get -y install lsof)

-  pip install quart httpx matplotlib aiohttp
+  pip install quart httpx matplotlib aiohttp datasets

  cd "$(dirname "$0")"

@@ -1,3 +1,5 @@
+# SPDX-License-Identifier: Apache-2.0
+
 import os

 import aiohttp
@@ -1,3 +1,5 @@
+# SPDX-License-Identifier: Apache-2.0
+
 import asyncio
 import itertools

@@ -1,3 +1,5 @@
+# SPDX-License-Identifier: Apache-2.0
+
 import json

 import matplotlib.pyplot as plt
@@ -1,8 +1,11 @@
+# SPDX-License-Identifier: Apache-2.0
+
 import pickle as pkl
 import time
+from collections.abc import Iterable
 from dataclasses import dataclass
 from itertools import product
-from typing import Callable, Iterable, List, Optional
+from typing import Callable, Optional

 import torch
 import torch.utils.benchmark as TBenchmark
@@ -27,7 +30,7 @@ class bench_params_t:
                f'x DT {self.dtype}')


-def get_bench_params() -> List[bench_params_t]:
+def get_bench_params() -> list[bench_params_t]:
    ## Test Fixtures
    NUM_TOKENS = [2**x for x in range(11)]
    HIDDEN_SIZES = list(range(1024, 8129, 1024))
@@ -1,3 +1,5 @@
+# SPDX-License-Identifier: Apache-2.0
+
 import os
 import sys
 from typing import Optional
@@ -1,3 +1,5 @@
+# SPDX-License-Identifier: Apache-2.0
+
 import time

 import torch
@@ -1,3 +1,5 @@
+# SPDX-License-Identifier: Apache-2.0
+
 import argparse
 import copy
 import itertools
@@ -5,9 +7,10 @@ import math
 import os
 import pickle as pkl
 import time
+from collections.abc import Iterable
 from dataclasses import dataclass
 from itertools import product
-from typing import Callable, Iterable, List, Optional, Tuple
+from typing import Callable, Optional

 import pandas as pd
 import torch
@@ -100,8 +103,8 @@ def quantize_and_pack(atype: torch.dtype,
    return w_ref, w_q, w_s, w_zp


-def create_bench_tensors(shape: Tuple[int, int, int], types: TypeConfig,
-                         group_size: Optional[int]) -> List[BenchmarkTensors]:
+def create_bench_tensors(shape: tuple[int, int, int], types: TypeConfig,
+                         group_size: Optional[int]) -> list[BenchmarkTensors]:
    m, n, k = shape

    # we want to make sure that weights don't fit into L2 cache between runs so
@@ -112,7 +115,7 @@ def create_bench_tensors(shape: Tuple[int, int, int], types: TypeConfig,

    a = rand_data((m, k), types.act_type, scale=5)

-    benchmark_tensors: List[BenchmarkTensors] = []
+    benchmark_tensors: list[BenchmarkTensors] = []
    for _ in range(num_weights):
        w = rand_data((k, n), types.act_type, scale=5)

@@ -274,7 +277,7 @@ def machete_create_bench_fn(bt: BenchmarkTensors,


 def bench_fns(label: str, sub_label: str, description: str,
-              fns: List[Callable]):
+              fns: list[Callable]):

    min_run_time = 1 if not NVTX_PROFILE else 0.1
    res = TBenchmark.Timer(
@@ -309,7 +312,7 @@ def bench(types: TypeConfig,
          n: int,
          label: str,
          sub_label: str,
-          sweep_schedules: bool = True) -> List[TMeasurement]:
+          sweep_schedules: bool = True) -> list[TMeasurement]:
    benchmark_tensors = create_bench_tensors((m, n, k), types, group_size)
    sub_label += f", L={len(benchmark_tensors)}"

@@ -412,12 +415,12 @@ def bench(types: TypeConfig,


 # runner
-def print_timers(timers: List[TMeasurement]):
+def print_timers(timers: list[TMeasurement]):
    compare = TBenchmark.Compare(timers)
    compare.print()


-def run(args, MKNs: Iterable[Tuple[int, int, int]]) -> Iterable[TMeasurement]:
+def run(args, MKNs: Iterable[tuple[int, int, int]]) -> Iterable[TMeasurement]:
    types = TypeConfig(
        act_type=args.act_type,
        weight_type=scalar_types.uint4b8 if args.group_zero_type is None \
@@ -429,7 +432,7 @@ def run(args, MKNs: Iterable[Tuple[int, int, int]]) -> Iterable[TMeasurement]:
        token_scale_type=args.token_scale_type,
    )

-    results: List[TMeasurement] = []
+    results: list[TMeasurement] = []
    for m, k, n in MKNs:
        timers = bench(types,
                       args.group_size,
@@ -447,8 +450,8 @@ def run(args, MKNs: Iterable[Tuple[int, int, int]]) -> Iterable[TMeasurement]:

 # output makers
 def make_output(
-    data: List[TMeasurement],
-    MKNs: Iterable[Tuple[int, int, int]],
+    data: list[TMeasurement],
+    MKNs: Iterable[tuple[int, int, int]],
    base_description: str,
    timestamp=None,
 ):
@@ -495,7 +498,7 @@ def run_model_bench(args):
    for i, model in enumerate(args.models):
        print(f"[{i}]  {model}")

-    def model_shapes(model_name: str, tp_size: int) -> List[Tuple[int, int]]:
+    def model_shapes(model_name: str, tp_size: int) -> list[tuple[int, int]]:
        KNs = []
        for KN, tp_split_dim in copy.deepcopy(WEIGHT_SHAPES[model_name]):
            KN[tp_split_dim] = KN[tp_split_dim] // tp_size
@@ -1,4 +1,4 @@
-from typing import List
+# SPDX-License-Identifier: Apache-2.0

 import torch
 import torch.utils.benchmark as benchmark
@@ -8,6 +8,8 @@ from vllm import _custom_ops as ops
 from vllm.model_executor.layers.quantization.gptq_marlin_24 import (
    GPTQ_MARLIN_24_MAX_PARALLEL, GPTQ_MARLIN_24_MIN_THREAD_N,
    GPTQ_MARLIN_24_SUPPORTED_GROUP_SIZES, GPTQ_MARLIN_24_SUPPORTED_QUANT_TYPES)
+from vllm.model_executor.layers.quantization.utils.allspark_utils import (
+    ALLSPARK_AMPERE_M_CUBLAS_THRESHOLD, ALLSPARK_SUPPORTED_QUANT_TYPES)
 from vllm.model_executor.layers.quantization.utils.marlin_utils import (
    GPTQ_MARLIN_MAX_PARALLEL, GPTQ_MARLIN_MIN_THREAD_N,
    MARLIN_SUPPORTED_GROUP_SIZES, query_marlin_supported_quant_types)
@@ -16,18 +18,18 @@ from vllm.model_executor.layers.quantization.utils.marlin_utils_test import (
 from vllm.model_executor.layers.quantization.utils.marlin_utils_test_24 import (
    marlin_24_quantize)
 from vllm.model_executor.layers.quantization.utils.quant_utils import (
-    gptq_pack, gptq_quantize_weights, sort_weights)
+    gptq_pack, gptq_quantize_weights, quantize_weights, sort_weights)
 from vllm.scalar_type import ScalarType
 from vllm.utils import FlexibleArgumentParser

 DEFAULT_MODELS = ["meta-llama/Llama-2-7b-hf/TP1"]
-DEFAULT_BATCH_SIZES = [1, 16, 32, 64, 128, 256, 512]
+DEFAULT_BATCH_SIZES = [1, 16, 32, 64, 128, 256, 512, 1024, 2048, 4096, 8192]

 ACT_ORDER_OPTS = [False, True]
 K_FULL_OPTS = [False, True]


-def bench_run(results: List[benchmark.Measurement], model: str,
+def bench_run(results: list[benchmark.Measurement], model: str,
              act_order: bool, is_k_full: bool, quant_type: ScalarType,
              group_size: int, size_m: int, size_k: int, size_n: int):
    label = "Quant Matmul"
@@ -79,6 +81,27 @@ def bench_run(results: List[benchmark.Measurement], model: str,
                                          GPTQ_MARLIN_24_MAX_PARALLEL)
    marlin_zp = torch.zeros_like(marlin_s, dtype=torch.int)

+    # AllSpark W8A16 quant
+    as_supported_case = (quant_type in ALLSPARK_SUPPORTED_QUANT_TYPES
+                         and group_size == -1 and not act_order and is_k_full)
+    if as_supported_case:
+        properties = torch.cuda.get_device_properties(b.device.index)
+        sm_count = properties.multi_processor_count
+        sm_version = properties.major * 10 + properties.minor
+
+        supported_arch = (sm_version >= 80 and sm_version < 90)
+        as_supported_case = as_supported_case and supported_arch
+        if supported_arch:
+            has_zp = False
+            w_ref, qw, s, zp = quantize_weights(b, quant_type, group_size,
+                                                has_zp)
+            qw = qw.to(torch.uint8)
+
+            qw_reorder, s_reorder, zp_reorder = \
+                ops.allspark_repack_weight(
+                qw, s, zp, has_zp)
+            CUBLAS_M_THRESHOLD = ALLSPARK_AMPERE_M_CUBLAS_THRESHOLD
+
    globals = {
        # Gen params
        "quant_type": quant_type,
@@ -107,10 +130,19 @@ def bench_run(results: List[benchmark.Measurement], model: str,
        # GPTQ params
        "q_w_gptq": q_w_gptq,
        "repack_sort_indices": repack_sort_indices,
+        # AllSpark W8A16 params
+        "qw_reorder": qw_reorder if as_supported_case else None,
+        "s_reorder": s_reorder if as_supported_case else None,
+        "zp_reorder": zp_reorder if as_supported_case else None,
+        "sm_count": sm_count if as_supported_case else None,
+        "sm_version": sm_version if as_supported_case else None,
+        "CUBLAS_M_THRESHOLD":
+        CUBLAS_M_THRESHOLD if as_supported_case else None,
        # Kernels
        "gptq_marlin_gemm": ops.gptq_marlin_gemm,
        "gptq_marlin_24_gemm": ops.gptq_marlin_24_gemm,
        "gptq_marlin_repack": ops.gptq_marlin_repack,
+        "allspark_w8a16_gemm": ops.allspark_w8a16_gemm,
    }

    min_run_time = 1
@@ -170,13 +202,24 @@ def bench_run(results: List[benchmark.Measurement], model: str,
            description="gptq_marlin_repack",
        ).blocked_autorange(min_run_time=min_run_time))

+    if as_supported_case:
+        results.append(
+            benchmark.Timer(
+                stmt=
+                "output = allspark_w8a16_gemm(a, qw_reorder, s_reorder, zp_reorder, size_n, group_size, sm_count, sm_version, CUBLAS_M_THRESHOLD, False, True)",  # noqa: E501
+                globals=globals,
+                label=label,
+                sub_label=sub_label,
+                description="allspark_w8a16_gemm_fp32",
+            ).blocked_autorange(min_run_time=min_run_time))
+

 def main(args):
    print("Benchmarking models:")
    for i, model in enumerate(args.models):
        print(f"[{i}]  {model}")

-    results: List[benchmark.Measurement] = []
+    results: list[benchmark.Measurement] = []

    for model in args.models:
        for layer in WEIGHT_SHAPES[model]:
@@ -1,7 +1,11 @@
+# SPDX-License-Identifier: Apache-2.0
+
 import argparse
 import time
+from contextlib import nullcontext
 from datetime import datetime
-from typing import Any, Dict, List, Tuple, TypedDict
+from itertools import product
+from typing import Any, TypedDict

 import ray
 import torch
@@ -13,6 +17,9 @@ from vllm.model_executor.layers.fused_moe.fused_moe import *
 from vllm.platforms import current_platform
 from vllm.utils import FlexibleArgumentParser

+FP8_DTYPE = torch.float8_e4m3fnuz if current_platform.is_rocm(
+) else torch.float8_e4m3fn
+

 class BenchmarkConfig(TypedDict):
    BLOCK_SIZE_M: int
@@ -34,6 +41,7 @@ def benchmark_config(
    use_fp8_w8a8: bool,
    use_int8_w8a16: bool,
    num_iters: int = 100,
+    block_quant_shape: List[int] = None,
 ) -> float:
    init_dtype = torch.float16 if use_fp8_w8a8 else dtype
    x = torch.randn(num_tokens, hidden_size, dtype=dtype)
@@ -75,13 +83,29 @@ def benchmark_config(
                               dtype=torch.float32)
        w2_scale = torch.randn((hidden_size, num_experts), dtype=torch.float32)
    if use_fp8_w8a8:
-        w1_scale = torch.randn(num_experts, dtype=torch.float32)
-        w2_scale = torch.randn(num_experts, dtype=torch.float32)
+        if block_quant_shape:
+            block_n, block_k = block_quant_shape[0], block_quant_shape[1]
+            E = num_experts
+            N = shard_intermediate_size // 2
+            K = hidden_size
+            factor_for_scale = 1e-2
+            n_tiles_w1 = (2 * N + block_n - 1) // block_n
+            n_tiles_w2 = (K + block_n - 1) // block_n
+            k_tiles_w1 = (K + block_k - 1) // block_k
+            k_tiles_w2 = (N + block_k - 1) // block_k
+            w1_scale = torch.rand((E, n_tiles_w1, k_tiles_w1),
+                                  dtype=torch.float32) * factor_for_scale
+            w2_scale = torch.rand((E, n_tiles_w2, k_tiles_w2),
+                                  dtype=torch.float32) * factor_for_scale
+        else:
+            w1_scale = torch.randn(num_experts, dtype=torch.float32)
+            w2_scale = torch.randn(num_experts, dtype=torch.float32)
+
        a1_scale = torch.randn(1, dtype=torch.float32)
        a2_scale = torch.randn(1, dtype=torch.float32)

-        w1 = w1.to(torch.float8_e4m3fn)
-        w2 = w2.to(torch.float8_e4m3fn)
+        w1 = w1.to(FP8_DTYPE)
+        w2 = w2.to(FP8_DTYPE)

    input_gating = torch.empty(num_tokens, num_experts, dtype=torch.float32)

@@ -105,6 +129,7 @@ def benchmark_config(
                w2_scale=w2_scale,
                a1_scale=a1_scale,
                a2_scale=a2_scale,
+                block_shape=block_quant_shape,
            )

    # JIT compilation & warmup
@@ -126,7 +151,7 @@ def benchmark_config(
    start_event = torch.cuda.Event(enable_timing=True)
    end_event = torch.cuda.Event(enable_timing=True)

-    latencies: List[float] = []
+    latencies: list[float] = []
    for i in range(num_iters):
        prepare(i)
        torch.cuda.synchronize()
@@ -141,28 +166,183 @@ def benchmark_config(
    return avg


-def get_configs_compute_bound() -> List[Dict[str, int]]:
-    # Reduced search space for faster tuning.
-    # TODO(woosuk): Increase the search space and use a performance model to
-    # prune the search space.
-    configs: List[BenchmarkConfig] = []
-    for num_stages in [2, 3, 4, 5]:
-        for block_m in [16, 32, 64, 128, 256]:
-            for block_k in [64, 128, 256]:
-                for block_n in [32, 64, 128, 256]:
-                    for num_warps in [4, 8]:
-                        for group_size in [1, 16, 32, 64]:
-                            configs.append({
-                                "BLOCK_SIZE_M": block_m,
-                                "BLOCK_SIZE_N": block_n,
-                                "BLOCK_SIZE_K": block_k,
-                                "GROUP_SIZE_M": group_size,
-                                "num_warps": num_warps,
-                                "num_stages": num_stages,
-                            })
+def get_rocm_tuning_space(use_fp16):
+    block_mn_range = [16, 32, 64, 128, 256]
+    block_k_range = [16, 32, 64, 128, 256]
+    if not use_fp16:
+        block_k_range.remove(16)  # BLOCK_K=16 not supported for fp8
+    num_warps_range = [1, 2, 4, 8]
+    group_m_range = [1, 4, 8, 16, 32]
+    num_stage_range = [2]
+    waves_per_eu_range = [0]
+    matrix_instr_nonkdim_range = [16, 32] if use_fp16 else []
+    kpack_range = [1, 2] if use_fp16 else []
+
+    param_ranges = {
+        "BLOCK_SIZE_M": block_mn_range,
+        "BLOCK_SIZE_N": block_mn_range,
+        "BLOCK_SIZE_K": block_k_range,
+        "GROUP_SIZE_M": group_m_range,
+        "num_warps": num_warps_range,
+        "num_stages": num_stage_range,
+        "waves_per_eu": waves_per_eu_range,
+    }
+    if use_fp16:
+        param_ranges["matrix_instr_nonkdim"] = matrix_instr_nonkdim_range
+        param_ranges["kpack"] = kpack_range
+
+    return param_ranges
+
+
+def get_configs_compute_bound(use_fp16,
+                              block_quant_shape) -> list[dict[str, int]]:
+    configs: list[BenchmarkConfig] = []
+
+    if current_platform.is_rocm():
+        param_ranges = get_rocm_tuning_space(use_fp16)
+    else:
+        # Reduced search space for faster tuning.
+        # TODO(woosuk): Increase the search space and use a performance model to
+        # prune the search space.
+        block_m_range = [16, 32, 64, 128, 256]
+        block_n_range = [32, 64, 128, 256]
+        block_k_range = [64, 128, 256]
+        num_warps_range = [4, 8]
+        group_m_range = [1, 16, 32, 64]
+        num_stage_range = [2, 3, 4, 5]
+
+        param_ranges = {
+            "BLOCK_SIZE_M": block_m_range,
+            "BLOCK_SIZE_N": block_n_range,
+            "BLOCK_SIZE_K": block_k_range,
+            "GROUP_SIZE_M": group_m_range,
+            "num_warps": num_warps_range,
+            "num_stages": num_stage_range,
+        }
+
+    keys, values = zip(*param_ranges.items())
+    for config_values in product(*values):
+        config = dict(zip(keys, config_values))
+        configs.append(config)
+
+    # Remove configs that are not compatible with fp8 block quantization
+    # BLOCK_SIZE_K must be a multiple of block_k
+    # BLOCK_SIZE_N must be a multiple of block_n
+    if block_quant_shape is not None and not use_fp16:
+        block_n, block_k = block_quant_shape[0], block_quant_shape[1]
+        for config in configs[:]:
+            if config["BLOCK_SIZE_K"] % block_k != 0 or config[
+                    "BLOCK_SIZE_N"] % block_n != 0:
+                configs.remove(config)
    return configs


+def prune_rocm_search_space(num_tokens, shard_intermediate_size, hidden_size,
+                            search_space, is_fp16, topk):
+    N1, K1 = shard_intermediate_size, hidden_size
+    N2, K2 = hidden_size, shard_intermediate_size // 2
+    pruned_space_1 = prune_rocm_configs(num_tokens * topk, N1, K1,
+                                        search_space, is_fp16)
+    pruned_space_2 = prune_rocm_configs(num_tokens * topk, N2, K2,
+                                        search_space, is_fp16)
+    search_space = merge_unique_dicts(pruned_space_1, pruned_space_2)
+    return search_space
+
+
+# The following code is inspired by ROCm/Triton GEMM tuning script:
+# https://github.com/ROCm/triton/blob/triton-mlir/scripts/amd/gemm/tune_gemm.py#L89
+def prune_rocm_configs(M, N, K, configs, is_fp16=True):
+    pruned_configs = []
+    elemBytes_a = 2 if is_fp16 else 1
+    elemBytes_b = 2 if is_fp16 else 1
+
+    mfma = 16 if M < 32 or N < 32 else 32
+
+    # TODO (zhanglx): figure out the boundary between large and small gemms
+    large_gemm = False
+    if M >= 2048 and N >= 2048:
+        large_gemm = True
+
+    for config in configs:
+        BLOCK_SIZE_M = config.get("BLOCK_SIZE_M")
+        BLOCK_SIZE_N = config.get("BLOCK_SIZE_N")
+        BLOCK_SIZE_K = config.get("BLOCK_SIZE_K")
+        num_warps = config.get("num_warps")
+
+        if is_fp16:
+            matrix_instr_nonkdim = config.get("matrix_instr_nonkdim")
+            if matrix_instr_nonkdim > mfma:
+                continue
+        if mfma == 4 and BLOCK_SIZE_K < 64:
+            continue
+        # some layouts could not work properly in case
+        # number elements per thread is less 1
+        if BLOCK_SIZE_M * BLOCK_SIZE_N < 64:
+            continue
+        SPLIT_K = config.get("SPLIT_K", 1)
+        GROUP_M = config.get("GROUP_SIZE_M")
+        if is_fp16:
+            if (matrix_instr_nonkdim > BLOCK_SIZE_M
+                    or matrix_instr_nonkdim > BLOCK_SIZE_N):
+                continue
+            if (matrix_instr_nonkdim >= M
+                    and matrix_instr_nonkdim != BLOCK_SIZE_M):
+                continue
+            if (matrix_instr_nonkdim >= N
+                    and matrix_instr_nonkdim != BLOCK_SIZE_N):
+                continue
+        # Skip BLOCK_SIZE that is too large compare to M/N
+        # unless BLOCK_SIZE is already small enough
+        if M * 2 < BLOCK_SIZE_M and BLOCK_SIZE_M != 16:
+            continue
+        if N * 2 < BLOCK_SIZE_N and BLOCK_SIZE_N != 16:
+            continue
+        # skip large split_k when not necessary
+        if SPLIT_K != 1 and not need_split_k(M, N, K):
+            continue
+        # skip split_k that leads to EVEN_K = false
+        leap = SPLIT_K * BLOCK_SIZE_K
+        modv = K % leap
+        if modv != 0:
+            continue
+        # skip large GROUP_M
+        if GROUP_M * BLOCK_SIZE_M > M and GROUP_M != 1:
+            continue
+        # out of shared memory resource
+        # TODO (zhanglx): This does not consider the LDS usage in the epilogue
+        LDS = (BLOCK_SIZE_K * BLOCK_SIZE_M * elemBytes_a +
+               BLOCK_SIZE_K * BLOCK_SIZE_N * elemBytes_b)
+        if LDS > 65536:
+            continue
+        # Skip small block sizes and num_warps for large gemm
+        # For fp16 and f8, we want to only use BLOCK_SIZE >= 64
+        if large_gemm:
+            if BLOCK_SIZE_M < 64 or BLOCK_SIZE_N < 64:
+                continue
+            if BLOCK_SIZE_K < 64:
+                continue
+            if num_warps < 4:
+                continue
+
+        pruned_configs.append(config)
+
+    return pruned_configs
+
+
+def need_split_k(SIZE_M, SIZE_N, SIZE_K):
+    return (SIZE_M < 64 or SIZE_N < 64) and SIZE_K > 1024
+
+
+def merge_unique_dicts(list1, list2):
+    result = []
+    combined_list = list1.copy()
+    combined_list.extend(list2)
+    for dictionary in combined_list:
+        if dictionary not in result:
+            result.append(dictionary)
+    return result
+
+
@ray.remote(num_gpus=1)
 class BenchmarkWorker:

@@ -170,6 +350,10 @@ class BenchmarkWorker:
        torch.set_default_device("cuda")
        current_platform.seed_everything(seed)
        self.seed = seed
+        # Get the device ID to allocate tensors and kernels
+        # on the respective GPU. This is required for Ray to work
+        # correctly with multi-GPU tuning on the ROCm platform.
+        self.device_id = int(ray.get_gpu_ids()[0])

    def benchmark(
        self,
@@ -181,7 +365,7 @@ class BenchmarkWorker:
        dtype: torch.dtype,
        use_fp8_w8a8: bool,
        use_int8_w8a16: bool,
-    ) -> Tuple[Dict[str, int], float]:
+    ) -> tuple[dict[str, int], float]:
        current_platform.seed_everything(self.seed)
        dtype_str = get_config_dtype_str(dtype,
                                         use_int8_w8a16=use_int8_w8a16,
@@ -191,9 +375,13 @@ class BenchmarkWorker:
        op_config = get_moe_configs(num_experts, shard_intermediate_size // 2,
                                    dtype_str)
        if op_config is None:
-            config = get_default_config(num_tokens, num_experts,
-                                        shard_intermediate_size, hidden_size,
-                                        topk, dtype_str)
+            config = get_default_config(num_tokens,
+                                        num_experts,
+                                        shard_intermediate_size,
+                                        hidden_size,
+                                        topk,
+                                        dtype_str,
+                                        is_marlin=False)
        else:
            config = op_config[min(op_config.keys(),
                                   key=lambda x: abs(x - num_tokens))]
@@ -213,29 +401,41 @@ class BenchmarkWorker:
        dtype: torch.dtype,
        use_fp8_w8a8: bool,
        use_int8_w8a16: bool,
-        search_space: List[Dict[str, int]],
-    ) -> Dict[str, int]:
+        search_space: list[dict[str, int]],
+        block_quant_shape: list[int],
+    ) -> dict[str, int]:
        best_config = None
        best_time = float("inf")
-        for config in tqdm(search_space):
-            try:
-                kernel_time = benchmark_config(config,
-                                               num_tokens,
-                                               num_experts,
-                                               shard_intermediate_size,
-                                               hidden_size,
-                                               topk,
-                                               dtype,
-                                               use_fp8_w8a8,
-                                               use_int8_w8a16,
-                                               num_iters=10)
-            except triton.runtime.autotuner.OutOfResources:
-                # Some configurations may be invalid and fail to compile.
-                continue
+        if current_platform.is_rocm():
+            is_fp16 = not (use_fp8_w8a8 or use_int8_w8a16)
+            search_space = prune_rocm_search_space(num_tokens,
+                                                   shard_intermediate_size,
+                                                   hidden_size, search_space,
+                                                   is_fp16, topk)

-            if kernel_time < best_time:
-                best_time = kernel_time
-                best_config = config
+        with torch.cuda.device(self.device_id) if current_platform.is_rocm(
+        ) else nullcontext():
+            for config in tqdm(search_space):
+                try:
+                    kernel_time = benchmark_config(
+                        config,
+                        num_tokens,
+                        num_experts,
+                        shard_intermediate_size,
+                        hidden_size,
+                        topk,
+                        dtype,
+                        use_fp8_w8a8,
+                        use_int8_w8a16,
+                        num_iters=20,
+                        block_quant_shape=block_quant_shape)
+                except triton.runtime.autotuner.OutOfResources:
+                    # Some configurations may be invalid and fail to compile.
+                    continue
+
+                if kernel_time < best_time:
+                    best_time = kernel_time
+                    best_config = config
        now = datetime.now()
        print(f"{now.ctime()}] Completed tuning for batch_size={num_tokens}")
        assert best_config is not None
@@ -244,19 +444,34 @@ class BenchmarkWorker:

 def sort_config(config: BenchmarkConfig) -> BenchmarkConfig:
    return {
-        "BLOCK_SIZE_M": config["BLOCK_SIZE_M"],
-        "BLOCK_SIZE_N": config["BLOCK_SIZE_N"],
-        "BLOCK_SIZE_K": config["BLOCK_SIZE_K"],
-        "GROUP_SIZE_M": config["GROUP_SIZE_M"],
-        "num_warps": config["num_warps"],
-        "num_stages": config["num_stages"],
+        "BLOCK_SIZE_M":
+        config["BLOCK_SIZE_M"],
+        "BLOCK_SIZE_N":
+        config["BLOCK_SIZE_N"],
+        "BLOCK_SIZE_K":
+        config["BLOCK_SIZE_K"],
+        "GROUP_SIZE_M":
+        config["GROUP_SIZE_M"],
+        "num_warps":
+        config["num_warps"],
+        "num_stages":
+        config["num_stages"],
+        **({
+            "waves_per_eu": config["waves_per_eu"]
+        } if "waves_per_eu" in config else {}),
+        **({
+            "matrix_instr_nonkdim": config["matrix_instr_nonkdim"]
+        } if "matrix_instr_nonkdim" in config else {}),
+        **({
+            "kpack": config["kpack"]
+        } if "kpack" in config else {}),
    }


-def save_configs(configs: Dict[int, BenchmarkConfig], num_experts: int,
+def save_configs(configs: dict[int, BenchmarkConfig], num_experts: int,
                 shard_intermediate_size: int, hidden_size: int, topk: int,
-                 dtype: torch.dtype, use_fp8_w8a8: bool,
-                 use_int8_w8a16: bool) -> None:
+                 dtype: torch.dtype, use_fp8_w8a8: bool, use_int8_w8a16: bool,
+                 block_quant_shape: List[int]) -> None:
    dtype_str = get_config_dtype_str(dtype,
                                     use_int8_w8a16=use_int8_w8a16,
                                     use_fp8_w8a8=use_fp8_w8a8)
@@ -264,7 +479,7 @@ def save_configs(configs: Dict[int, BenchmarkConfig], num_experts: int,
    # NOTE(woosuk): The current naming convention uses w2.shape[2], which
    # is the intermediate size after silu_and_mul.
    filename = get_config_file_name(num_experts, shard_intermediate_size // 2,
-                                    dtype_str)
+                                    dtype_str, block_quant_shape)

    print(f"Writing best config to {filename}...")
    with open(filename, "w") as f:
@@ -274,8 +489,9 @@ def save_configs(configs: Dict[int, BenchmarkConfig], num_experts: int,

 def main(args: argparse.Namespace):
    print(args)
-
-    config = AutoConfig.from_pretrained(args.model)
+    block_quant_shape = None
+    config = AutoConfig.from_pretrained(
+        args.model, trust_remote_code=args.trust_remote_code)
    if config.architectures[0] == "DbrxForCausalLM":
        E = config.ffn_config.moe_num_experts
        topk = config.ffn_config.moe_top_k
@@ -286,6 +502,13 @@ 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] == "DeepseekV3ForCausalLM"
+          or config.architectures[0] == "DeepseekV2ForCausalLM"):
+        E = config.n_routed_experts
+        topk = config.num_experts_per_tok
+        intermediate_size = config.moe_intermediate_size
+        shard_intermediate_size = 2 * intermediate_size // args.tp_size
+        block_quant_shape = config.quantization_config['weight_block_size']
    else:
        # Default: Mixtral.
        E = config.num_local_experts
@@ -294,7 +517,7 @@ def main(args: argparse.Namespace):
        shard_intermediate_size = 2 * intermediate_size // args.tp_size

    hidden_size = config.hidden_size
-    dtype = config.torch_dtype
+    dtype = torch.float16 if current_platform.is_rocm() else config.torch_dtype
    use_fp8_w8a8 = args.dtype == "fp8_w8a8"
    use_int8_w8a16 = args.dtype == "int8_w8a16"

@@ -310,7 +533,7 @@ def main(args: argparse.Namespace):
    num_gpus = int(ray.available_resources()["GPU"])
    workers = [BenchmarkWorker.remote(args.seed) for _ in range(num_gpus)]

-    def _distribute(method: str, inputs: List[Any]) -> List[Any]:
+    def _distribute(method: str, inputs: list[Any]) -> list[Any]:
        outputs = []
        worker_idx = 0
        for input_args in inputs:
@@ -322,27 +545,31 @@ def main(args: argparse.Namespace):
        return ray.get(outputs)

    if args.tune:
-        search_space = get_configs_compute_bound()
+        is_fp16 = not (use_fp8_w8a8 or use_int8_w8a16)
+        search_space = get_configs_compute_bound(is_fp16, block_quant_shape)
        print(f"Start tuning over {len(search_space)} configurations...")

        start = time.time()
        configs = _distribute(
-            "tune", [(batch_size, E, shard_intermediate_size, hidden_size,
-                      topk, dtype, use_fp8_w8a8, use_int8_w8a16, search_space)
-                     for batch_size in batch_sizes])
+            "tune",
+            [(batch_size, E, shard_intermediate_size, hidden_size, topk, dtype,
+              use_fp8_w8a8, use_int8_w8a16, search_space, block_quant_shape)
+             for batch_size in batch_sizes])
        best_configs = {
            M: sort_config(config)
            for M, config in zip(batch_sizes, configs)
        }
        save_configs(best_configs, E, shard_intermediate_size, hidden_size,
-                     topk, dtype, use_fp8_w8a8, use_int8_w8a16)
+                     topk, dtype, use_fp8_w8a8, use_int8_w8a16,
+                     block_quant_shape)
        end = time.time()
        print(f"Tuning took {end - start:.2f} seconds")
    else:
        outputs = _distribute(
-            "benchmark", [(batch_size, E, shard_intermediate_size, hidden_size,
-                           topk, dtype, use_fp8_w8a8, use_int8_w8a16)
-                          for batch_size in batch_sizes])
+            "benchmark",
+            [(batch_size, E, shard_intermediate_size, hidden_size, topk, dtype,
+              use_fp8_w8a8, use_int8_w8a16, block_quant_shape)
+             for batch_size in batch_sizes])

        for batch_size, (config, kernel_time) in zip(batch_sizes, outputs):
            print(f"Batch size: {batch_size}, config: {config}")
@@ -354,7 +581,11 @@ if __name__ == "__main__":
    parser.add_argument("--model",
                        type=str,
                        default="mistralai/Mixtral-8x7B-Instruct-v0.1")
-    parser.add_argument("--tp-size", "-tp", type=int, default=2)
+    parser.add_argument("--tp-size",
+                        "-tp",
+                        "--tensor-parallel-size",
+                        type=int,
+                        default=2)
    parser.add_argument("--dtype",
                        type=str,
                        choices=["auto", "fp8_w8a8", "int8_w8a16"],
@@ -362,6 +593,7 @@ if __name__ == "__main__":
    parser.add_argument("--seed", type=int, default=0)
    parser.add_argument("--batch-size", type=int, required=False)
    parser.add_argument("--tune", action="store_true")
+    parser.add_argument("--trust-remote-code", action="store_true")
    args = parser.parse_args()

    main(args)
--- a/Show More
+++ b/Show More