[V1][Spec Decode] Update target_logits in place for rejection sampling (#15427)

Signed-off-by: Woosuk Kwon <woosuk.kwon@berkeley.edu>

.buildkite/check-wheel-size.py

@@ -1,12 +1,14 @@
+# SPDX-License-Identifier: Apache-2.0
+
 import os
 import sys
 import zipfile
 
-# Read the VLLM_MAX_SIZE_MB environment variable, defaulting to 300 MiB
+# 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', 300))
+VLLM_MAX_SIZE_MB = int(os.environ.get('VLLM_MAX_SIZE_MB', 400))
 
 
 def print_top_10_largest_files(zip_file):

.buildkite/generate_index.py

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

.buildkite/lm-eval-harness/configs/Minitron-4B-Base-FP8.yaml

@@ -4,8 +4,8 @@ tasks:
 - name: "gsm8k"
   metrics:
   - name: "exact_match,strict-match"
-    value: 0.233
+    value: 0.231
   - name: "exact_match,flexible-extract"
-    value: 0.236
+    value: 0.22
 limit: 1000
 num_fewshot: 5

.buildkite/lm-eval-harness/configs/SparseLlama3.1_2of4_fp8_compressed.yaml

@@ -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 

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

@@ -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
@@ -12,6 +13,7 @@ from pathlib import Path
 
 import lm_eval
 import numpy
+import pytest
 import yaml
 
 RTOL = 0.05
@@ -45,6 +47,10 @@ def test_lm_eval_correctness():
     eval_config = yaml.safe_load(
         Path(TEST_DATA_FILE).read_text(encoding="utf-8"))
 
+    if eval_config[
+            "model_name"] == "nm-testing/Meta-Llama-3-70B-Instruct-FBGEMM-nonuniform":  #noqa: E501
+        pytest.skip("FBGEMM is currently failing on main.")
+
     # Launch eval requests.
     results = launch_lm_eval(eval_config)
 

.buildkite/nightly-benchmarks/README.md

@@ -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).
-

.buildkite/nightly-benchmarks/benchmark-pipeline.yaml

@@ -10,12 +10,18 @@ steps:
           - image: badouralix/curl-jq
             command:
             - sh .buildkite/nightly-benchmarks/scripts/wait-for-image.sh
-
+  - 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:
@@ -50,6 +56,7 @@ steps:
     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:$BUILDKITE_COMMIT
@@ -75,6 +82,7 @@ steps:
     agents:
       queue: H100
     depends_on: wait-for-container-image
+    if: build.branch == "main"
     plugins:
     - docker#v5.12.0:
         image: public.ecr.aws/q9t5s3a7/vllm-ci-postmerge-repo:$BUILDKITE_COMMIT
@@ -90,3 +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

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

@@ -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`.
-

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

@@ -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.

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

@@ -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.
-

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

@@ -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

.buildkite/nightly-benchmarks/scripts/download-tokenizer.py

@@ -1,3 +1,5 @@
+# SPDX-License-Identifier: Apache-2.0
+
 import argparse
 
 from transformers import AutoTokenizer

.buildkite/nightly-benchmarks/scripts/generate-nightly-markdown.py

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

.buildkite/nightly-benchmarks/scripts/get-lmdeploy-modelname.py

@@ -1,3 +1,5 @@
+# SPDX-License-Identifier: Apache-2.0
+
 from lmdeploy.serve.openai.api_client import APIClient
 
 api_client = APIClient("http://localhost:8000")

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

@@ -426,7 +426,7 @@ main() {
 
   pip install -U transformers
 
-  pip install -r requirements-dev.txt
+  pip install -r requirements/dev.txt
   which genai-perf
 
   # check storage

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

@@ -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 [[ "${ENGINE_VERSION:-v0}" == "v1" ]]; then
+    export VLLM_USE_V1=1
+  fi
+
   # dependencies
   (which wget && which curl) || (apt-get update && apt-get install -y wget curl)
   (which jq) || (apt-get update && apt-get -y install jq)
@@ -353,7 +361,7 @@ main() {
   # get the current IP address, required by benchmark_serving.py
   export VLLM_HOST_IP=$(hostname -I | awk '{print $1}')
   # turn of the reporting of the status of each request, to clean up the terminal output
-  export VLLM_LOG_LEVEL="WARNING"
+  export VLLM_LOGGING_LEVEL="WARNING"
 
   # prepare for benchmarking
   cd benchmarks || exit 1

.buildkite/nightly-benchmarks/scripts/summary-nightly-results.py

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

.buildkite/nightly-benchmarks/scripts/wait-for-image.sh

@@ -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
 

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

@@ -29,4 +29,4 @@
             "num-iters": 15
         }
     }
-]
+]

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

@@ -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",

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

@@ -32,4 +32,4 @@
             "backend": "vllm"
         }
     }
-]
+]

.buildkite/release-pipeline.yaml

@@ -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"
@@ -56,6 +67,11 @@ steps:
     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: ~
@@ -66,7 +82,7 @@ steps:
       queue: cpu_queue_postmerge
     commands:
       - "aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin public.ecr.aws/q9t5s3a7"
-      - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg GIT_REPO_CHECK=1 --tag public.ecr.aws/q9t5s3a7/vllm-cpu-release-repo:$RELEASE_VERSION --progress plain -f Dockerfile.cpu ."
-      - "docker push public.ecr.aws/q9t5s3a7/vllm-cpu-release-repo:$RELEASE_VERSION"
+      - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg GIT_REPO_CHECK=1 --tag public.ecr.aws/q9t5s3a7/vllm-cpu-release-repo:$(buildkite-agent meta-data get release-version) --tag public.ecr.aws/q9t5s3a7/vllm-cpu-release-repo:latest --progress plain -f Dockerfile.cpu ."
+      - "docker push public.ecr.aws/q9t5s3a7/vllm-cpu-release-repo:$(buildkite-agent meta-data get release-version)"
     env:
       DOCKER_BUILDKIT: "1"

.buildkite/run-amd-test.sh

@@ -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 \
@@ -92,19 +91,40 @@ 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
+#ignore certain Entrypoints/openai tests
 if [[ $commands == *" entrypoints/openai "* ]]; then
   commands=${commands//" entrypoints/openai "/" entrypoints/openai \
-  --ignore=entrypoints/openai/test_accuracy.py \
   --ignore=entrypoints/openai/test_audio.py \
-  --ignore=entrypoints/openai/test_encoder_decoder.py \
-  --ignore=entrypoints/openai/test_embedding.py \
-  --ignore=entrypoints/openai/test_oot_registration.py "}
+  --ignore=entrypoints/openai/test_chat.py \
+  --ignore=entrypoints/openai/test_shutdown.py \
+  --ignore=entrypoints/openai/test_completion.py \
+  --ignore=entrypoints/openai/test_sleep.py \
+  --ignore=entrypoints/openai/test_models.py \
+  --ignore=entrypoints/openai/test_prompt_validation.py "}
 fi
 
+#ignore certain Entrypoints/llm tests
+if [[ $commands == *" && pytest -v -s entrypoints/llm/test_guided_generate.py"* ]]; then
+  commands=${commands//" && pytest -v -s entrypoints/llm/test_guided_generate.py"/" "}
+fi
+
+# --ignore=entrypoints/openai/test_encoder_decoder.py \
+# --ignore=entrypoints/openai/test_embedding.py \
+# --ignore=entrypoints/openai/test_oot_registration.py
+# --ignore=entrypoints/openai/test_accuracy.py \
+# --ignore=entrypoints/openai/test_models.py <= Fails on MI250 but passes on MI300 as of 2025-03-13
+
+
 PARALLEL_JOB_COUNT=8
 # check if the command contains shard flag, we will run all shards in parallel because the host have 8 GPUs. 
 if [[ $commands == *"--shard-id="* ]]; then
@@ -121,6 +141,8 @@ 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}_${GPU}" \
@@ -148,6 +170,8 @@ else
           --rm \
           -e HIP_VISIBLE_DEVICES=0 \
           -e HF_TOKEN \
+          -e AWS_ACCESS_KEY_ID \
+          -e AWS_SECRET_ACCESS_KEY \
           -v "${HF_CACHE}:${HF_MOUNT}" \
           -e "HF_HOME=${HF_MOUNT}" \
           --name "${container_name}" \

.buildkite/run-cpu-test.sh

@@ -19,23 +19,25 @@ 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-"$BUILDKITE_BUILD_NUMBER"-"$NUMA_NODE" cpu-test-"$BUILDKITE_BUILD_NUMBER"
+ --cpuset-mems="$NUMA_NODE" --privileged=true -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="$NUMA_NODE" --privileged=true --network host -e HF_TOKEN --env VLLM_CPU_KVCACHE_SPACE=4 --shm-size=4g --name cpu-test-"$BUILDKITE_BUILD_NUMBER"-avx2-"$NUMA_NODE" cpu-test-"$BUILDKITE_BUILD_NUMBER"-avx2
+ --cpuset-mems="$NUMA_NODE" --privileged=true -e HF_TOKEN --env VLLM_CPU_KVCACHE_SPACE=4 --shm-size=4g --name cpu-test-"$BUILDKITE_BUILD_NUMBER"-avx2-"$NUMA_NODE" cpu-test-"$BUILDKITE_BUILD_NUMBER"-avx2
 
 function cpu_tests() {
   set -e
   export NUMA_NODE=$2
+  export BUILDKITE_BUILD_NUMBER=$3
 
   # offline inference
   docker exec cpu-test-"$BUILDKITE_BUILD_NUMBER"-avx2-"$NUMA_NODE" bash -c "
     set -e
-    python3 examples/offline_inference/basic.py"
+    python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m"
 
   # Run basic model test
   docker exec cpu-test-"$BUILDKITE_BUILD_NUMBER"-"$NUMA_NODE" bash -c "
     set -e
-    pip install -r vllm/requirements-test.txt
+    pip install -r vllm/requirements/test.txt
+    pip install -r vllm/requirements/cpu.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
@@ -85,4 +87,4 @@ function cpu_tests() {
 
 # All of CPU tests are expected to be finished less than 40 mins.
 export -f cpu_tests
-timeout 40m bash -c "cpu_tests $CORE_RANGE $NUMA_NODE"
+timeout 40m bash -c "cpu_tests $CORE_RANGE $NUMA_NODE $BUILDKITE_BUILD_NUMBER"

.buildkite/run-gh200-test.sh

@@ -14,6 +14,7 @@ DOCKER_BUILDKIT=1 docker build . \
   -t gh200-test \
   --build-arg max_jobs=66 \
   --build-arg nvcc_threads=2 \
+  --build-arg RUN_WHEEL_CHECK=false \
   --build-arg torch_cuda_arch_list="9.0+PTX" \
   --build-arg vllm_fa_cmake_gpu_arches="90-real"
 
@@ -23,6 +24,6 @@ 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/basic.py
+docker run -e HF_TOKEN -e VLLM_WORKER_MULTIPROC_METHOD=spawn -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
 '

.buildkite/run-hpu-test.sh

@@ -20,5 +20,5 @@ 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/basic.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=$?

.buildkite/run-neuron-test.sh

@@ -29,9 +29,6 @@ if [ -f /tmp/neuron-docker-build-timestamp ]; then
         docker image prune -f
         # Remove unused volumes / force the system prune for old images as well.
         docker volume prune -f && docker system prune -f
-        # Remove huggingface model artifacts and compiler cache
-        rm -rf "${HF_MOUNT:?}/*"
-        rm -rf "${NEURON_COMPILE_CACHE_MOUNT:?}/*"
         echo "$current_time" > /tmp/neuron-docker-build-timestamp
     fi
 else
@@ -47,11 +44,11 @@ remove_docker_container() {
 trap remove_docker_container EXIT
 
 # Run the image
-docker run --rm -it --device=/dev/neuron0 --device=/dev/neuron1 --network host \
+docker run --rm -it --device=/dev/neuron0 --network bridge \
        -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"
+       /bin/bash -c "python3 /workspace/vllm/examples/offline_inference/neuron.py && python3 -m pytest /workspace/vllm/tests/neuron/1_core/ -v --capture=tee-sys && python3 -m pytest /workspace/vllm/tests/neuron/2_core/ -v --capture=tee-sys"

.buildkite/run-openvino-test.sh

@@ -1,16 +0,0 @@
-#!/bin/bash
-
-# This script build the OpenVINO docker image and run the offline inference inside the container.
-# It serves a sanity check for compilation and basic model usage.
-set -ex
-
-# Try building the docker image
-docker build -t openvino-test -f Dockerfile.openvino .
-
-# Setup cleanup
-remove_docker_container() { docker rm -f openvino-test || true; }
-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/basic.py

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

@@ -19,8 +19,20 @@ docker run --privileged --net host --shm-size=16G -it \
     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"
+    && export VLLM_USE_V1=1 \
+    && export VLLM_XLA_CHECK_RECOMPILATION=1 \
+    && echo TEST_1 \
+    && pytest /workspace/vllm/tests/tpu/test_compilation.py \
+    && echo TEST_2 \
+    && pytest -v -s /workspace/vllm/tests/v1/tpu/test_basic.py \
+    && echo TEST_3 \
+    && pytest -v -s /workspace/vllm/tests/entrypoints/llm/test_accuracy.py::test_lm_eval_accuracy_v1_engine \
+    && echo TEST_4 \
+    && pytest -s -v /workspace/vllm/tests/tpu/test_quantization_accuracy.py \
+    && echo TEST_5 \
+    && python3 /workspace/vllm/examples/offline_inference/tpu.py" \
+
+
+# TODO: This test fails because it uses RANDOM_SEED sampling
+# && VLLM_USE_V1=1 pytest -v -s /workspace/vllm/tests/tpu/test_custom_dispatcher.py \
+

.buildkite/run-xpu-test.sh

@@ -4,16 +4,28 @@
 # It serves a sanity check for compilation and basic model usage.
 set -ex
 
+image_name="xpu/vllm-ci:${BUILDKITE_COMMIT}"
+container_name="xpu_${BUILDKITE_COMMIT}_$(tr -dc A-Za-z0-9 < /dev/urandom | head -c 10; echo)"
+
 # Try building the docker image
-docker build -t xpu-test -f Dockerfile.xpu .
+docker build -t ${image_name} -f Dockerfile.xpu .
 
 # Setup cleanup
-remove_docker_container() { docker rm -f xpu-test || true; }
+remove_docker_container() { 
+  docker rm -f "${container_name}" || true; 
+  docker image rm -f "${image_name}" || true;
+  docker system prune -f || true;
+}
 trap remove_docker_container EXIT
-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/basic.py
-    python3 examples/offline_inference/cli.py -tp 2
+docker run \
+    --device /dev/dri \
+    -v /dev/dri/by-path:/dev/dri/by-path \
+    --entrypoint="" \
+    --name "${container_name}" \
+    "${image_name}" \
+    sh -c '
+    VLLM_USE_V1=0 python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m
+    VLLM_USE_V1=0 python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m -tp 2
 '

.buildkite/test-pipeline.yaml

@@ -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  #####
@@ -35,13 +35,12 @@ steps:
   fast_check: true
   no_gpu: True
   commands:
-  - pip install -r requirements-docs.txt
+  - 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/api/inference_params.html
 
 - label: Async Engine, Inputs, Utils, Worker Test # 24min
-  fast_check: true
   source_file_dependencies:
   - vllm/
   - tests/mq_llm_engine
@@ -50,9 +49,9 @@ 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
@@ -78,6 +77,7 @@ steps:
   - tests/basic_correctness/test_preemption
   - tests/basic_correctness/test_cumem.py
   commands:
+  - export VLLM_WORKER_MULTIPROC_METHOD=spawn
   - 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
@@ -107,45 +107,59 @@ steps:
   mirror_hardwares: [amd]
   source_file_dependencies:
   - vllm/
+  - tests/entrypoints/llm
+  - tests/entrypoints/openai
+  - tests/entrypoints/test_chat_utils
+  - tests/entrypoints/offline_mode
   commands:
+  - export VLLM_WORKER_MULTIPROC_METHOD=spawn
   - 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
+  - VLLM_USE_V1=0 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  --ignore=entrypoints/openai/test_chat_with_tool_reasoning.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
+  - VLLM_USE_V1=0 pytest -v -s entrypoints/offline_mode # Needs to avoid interference with other tests
 
 - label: Distributed Tests (4 GPUs) # 10min
   working_dir: "/vllm-workspace/tests"
   num_gpus: 4
-  fast_check: true
   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:
+  # test with tp=2 and external_dp=2
+  - VLLM_USE_V1=0 torchrun --nproc-per-node=4 distributed/test_torchrun_example.py
+  - torchrun --nproc-per-node=4 distributed/test_torchrun_example.py
+  # test with internal dp
+  - 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
+  - pushd ../examples/offline_inference
+  - VLLM_ENABLE_V1_MULTIPROCESSING=0 python3 rlhf.py
+  - VLLM_ENABLE_V1_MULTIPROCESSING=0 RAY_DEDUP_LOGS=0 python3 rlhf_colocate.py
+  - popd
 
 - label: Metrics, Tracing Test # 10min
-  num_gpus: 2 
-  fast_check: true
+  num_gpus: 2
   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' \
@@ -172,6 +186,9 @@ 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
@@ -183,7 +200,23 @@ steps:
     - vllm/
     - tests/v1
   commands:
-    - VLLM_USE_V1=1 pytest -v -s v1
+    # split the test to avoid interference
+    - pytest -v -s v1/core
+    - pytest -v -s v1/entrypoints
+    - pytest -v -s v1/engine
+    - pytest -v -s v1/entrypoints
+    - pytest -v -s v1/sample
+    - pytest -v -s v1/worker
+    - pytest -v -s v1/structured_output
+    - pytest -v -s v1/test_stats.py
+    - pytest -v -s v1/test_utils.py
+    - pytest -v -s v1/test_oracle.py
+    # TODO: accuracy does not match, whether setting
+    # VLLM_USE_FLASHINFER_SAMPLER or not on H100.
+    - pytest -v -s v1/e2e
+    # Integration test for streaming correctness (requires special branch).
+    - pip install -U git+https://github.com/robertgshaw2-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"
@@ -193,19 +226,22 @@ steps:
   - examples/
   commands:
     - pip install tensorizer # for tensorizer test
-    - python3 offline_inference/basic.py
-    - python3 offline_inference/cpu_offload.py
-    - python3 offline_inference/chat.py
+    - 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/audio_language.py --seed 0
+    - python3 offline_inference/vision_language.py --seed 0
+    - python3 offline_inference/vision_language_embedding.py --seed 0
+    - python3 offline_inference/vision_language_multi_image.py --seed 0
+    - VLLM_USE_V1=0 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/classification.py
-    - python3 offline_inference/embedding.py
-    - python3 offline_inference/scoring.py
-    - python3 offline_inference/profiling.py --model facebook/opt-125m run_num_steps --num-steps 2
+    - python3 offline_inference/encoder_decoder_multimodal.py --model-type whisper --seed 0
+    - python3 offline_inference/basic/classify.py
+    - python3 offline_inference/basic/embed.py
+    - python3 offline_inference/basic/score.py
+    - VLLM_USE_V1=0 python3 offline_inference/profiling.py --model facebook/opt-125m run_num_steps --num-steps 2
 
 - label: Prefix Caching Test # 9min
   mirror_hardwares: [amd]
@@ -232,7 +268,7 @@ steps:
   - vllm/model_executor/guided_decoding
   - tests/test_logits_processor
   - tests/model_executor/test_guided_processors
-  commands: 
+  commands:
     - pytest -v -s test_logits_processor.py
     - pytest -v -s model_executor/test_guided_processors.py
 
@@ -243,7 +279,7 @@ steps:
   - 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
@@ -251,11 +287,10 @@ steps:
   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 --ignore=lora/test_minicpmv_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
-  fast_check: true
+- label: PyTorch Fullgraph Smoke Test # 9min
   source_file_dependencies:
   - vllm/
   - tests/compile
@@ -264,8 +299,9 @@ steps:
   # these tests need to be separated, cannot combine
   - pytest -v -s compile/piecewise/test_simple.py
   - pytest -v -s compile/piecewise/test_toy_llama.py
+  - pytest -v -s compile/test_pass_manager.py
 
-- label: "PyTorch Fullgraph Test" # 18min
+- label: PyTorch Fullgraph Test # 18min
   source_file_dependencies:
   - vllm/
   - tests/compile
@@ -317,6 +353,14 @@ 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/
@@ -340,8 +384,10 @@ steps:
   - vllm/
   - tests/models
   commands:
+    - pytest -v -s models/test_transformers.py
     - pytest -v -s models/test_registry.py
-    - pytest -v -s models/test_initialization.py
+    # V1 Test: https://github.com/vllm-project/vllm/issues/14531
+    - VLLM_USE_V1=0 pytest -v -s models/test_initialization.py
 
 - label: Language Models Test (Standard) # 32min
   #mirror_hardwares: [amd]
@@ -469,26 +515,25 @@ steps:
   - vllm/worker/model_runner.py
   - entrypoints/llm/test_collective_rpc.py
   commands:
-  - pytest -v -s entrypoints/llm/test_collective_rpc.py
-  - torchrun --nproc-per-node=2 distributed/test_torchrun_example.py
+  - VLLM_ENABLE_V1_MULTIPROCESSING=0 pytest -v -s entrypoints/llm/test_collective_rpc.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)'
   # 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
+  - VLLM_USE_V1=0 CUDA_VISIBLE_DEVICES=0,1 pytest -v -s test_sharded_state_loader.py
+  - VLLM_USE_V1=0 CUDA_VISIBLE_DEVICES=0,1 pytest -v -s kv_transfer/test_disagg.py
 
 - label: Plugin Tests (2 GPUs) # 40min
   working_dir: "/vllm-workspace/tests"
   num_gpus: 2
-  fast_check: true
   source_file_dependencies:
   - vllm/plugins/
   - tests/plugins/
@@ -499,6 +544,7 @@ steps:
   - 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
@@ -548,11 +594,12 @@ 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
@@ -573,7 +620,7 @@ 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 #####
@@ -585,7 +632,7 @@ 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

.buildkite/upload-wheels.sh

@@ -50,8 +50,11 @@ 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 $normal_wheel matches cu121, do not upload the index.html
+    echo "Skipping index files for cu121 wheels"
 else
-    # only upload index.html for cu12 wheels (default wheels)
+    # only upload index.html for cu124 wheels (default wheels)
     aws s3 cp index.html "s3://vllm-wheels/$BUILDKITE_COMMIT/vllm/index.html"
     aws s3 cp "s3://vllm-wheels/nightly/index.html" "s3://vllm-wheels/$BUILDKITE_COMMIT/index.html"
 fi
@@ -63,8 +66,11 @@ aws s3 cp "$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 cu12 wheels (default wheels)
+    # only upload index.html for cu124 wheels (default wheels)
     aws s3 cp index.html "s3://vllm-wheels/nightly/vllm/index.html"
 fi
 

.github/CODEOWNERS

@@ -10,27 +10,32 @@
 /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/model_executor/guided_decoding @mgoin @russellb
 /vllm/multimodal @DarkLight1337 @ywang96
 CMakeLists.txt @tlrmchlsmth
 
 # vLLM V1
 /vllm/v1 @WoosukKwon @robertgshaw2-redhat @njhill @ywang96 @comaniac @alexm-redhat
+/vllm/v1/structured_output @mgoin @russellb
 
 # Test ownership
-/tests/async_engine @njhill @robertgshaw2-redhat @simon-mo
-/tests/test_inputs.py @DarkLight1337 @ywang96
-/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-redhat
 /.buildkite/lm-eval-harness @mgoin @simon-mo
+/tests/async_engine @njhill @robertgshaw2-redhat @simon-mo
+/tests/basic_correctness/test_chunked_prefill @rkooo567 @comaniac
 /tests/distributed/test_multi_node_assignment.py @youkaichao
 /tests/distributed/test_pipeline_parallel.py @youkaichao
 /tests/distributed/test_same_node.py @youkaichao
+/tests/entrypoints @DarkLight1337 @robertgshaw2-redhat @simon-mo
+/tests/entrypoints/llm/test_guided_generate.py @mgoin @russellb
+/tests/kernels @tlrmchlsmth @WoosukKwon
+/tests/model_executor/test_guided_processors.py @mgoin @russellb
+/tests/models @DarkLight1337 @ywang96
 /tests/multi_step @alexm-redhat @comaniac
+/tests/multimodal @DarkLight1337 @ywang96
+/tests/prefix_caching @comaniac @KuntaiDu
+/tests/quantization @mgoin @robertgshaw2-redhat
+/tests/spec_decode @njhill @LiuXiaoxuanPKU
+/tests/test_inputs.py @DarkLight1337 @ywang96
+/tests/v1/entrypoints/llm/test_struct_output_generate.py @mgoin @russellb
+/tests/v1/structured_output @mgoin @russellb
 /tests/weight_loading @mgoin @youkaichao
-/tests/basic_correctness/test_chunked_prefill @rkooo567 @comaniac

.github/ISSUE_TEMPLATE/400-bug-report.yml

@@ -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

.github/ISSUE_TEMPLATE/800-misc-discussion.yml

@@ -1,28 +0,0 @@
-name: 🎲 Misc/random discussions that do not fit into the above categories.
-description: Submit a discussion as you like. Note that developers are heavily overloaded and we mainly rely on community users to answer these issues.
-title: "[Misc]: "
-labels: ["misc"]
-
-body:
-- type: markdown
-  attributes:
-    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+).
-- type: textarea
-  attributes:
-    label: Anything you want to discuss about vllm.
-    description: >
-      Anything you want to discuss about vllm.
-  validations:
-    required: true
-- type: markdown
-  attributes:
-    value: >
-      Thanks for contributing 🎉!
-- type: checkboxes
-  id: askllm
-  attributes:
-    label: Before submitting a new issue...
-    options:
-      - label: Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the [documentation page](https://docs.vllm.ai/en/latest/), which can answer lots of frequently asked questions.
-        required: true

.github/ISSUE_TEMPLATE/config.yml

@@ -1 +1,5 @@
 blank_issues_enabled: false
+contact_links:
+  - name: Questions
+    url: https://discuss.vllm.ai
+    about: Ask questions and discuss with other vLLM community members

.github/PULL_REQUEST_TEMPLATE.md

@@ -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>**

.github/dependabot.yml

@@ -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:

.github/mergify.yml

@@ -5,6 +5,7 @@ pull_request_rules:
     - or:
       - files~=^[^/]+\.md$
       - files~=^docs/
+      - files~=^examples/
   actions:
     label:
       add:
@@ -35,6 +36,58 @@ pull_request_rules:
       add:
         - frontend
 
+- name: label-multi-modality
+  description: Automatically apply multi-modality label
+  conditions:
+    - or:
+      - files~=^vllm/multimodal/
+      - files~=^tests/multimodal/
+      - files~=^tests/models/multimodal/
+      - files~=^tests/models/*/audio_language/
+      - files~=^tests/models/*/vision_language/
+      - files=tests/models/test_vision.py
+  actions:
+    label:
+      add:
+        - multi-modality
+
+- name: label-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

.github/workflows/cleanup_pr_body.yml

@@ -16,7 +16,7 @@ jobs:
         uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
 
       - name: Set up Python
-        uses: actions/setup-python@0b93645e9fea7318ecaed2b359559ac225c90a2b # v5.3.0
+        uses: actions/setup-python@42375524e23c412d93fb67b49958b491fce71c38 # v5.4.0
         with:
           python-version: '3.12'
 

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

@@ -12,17 +12,17 @@ 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
 
@@ -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 .

.github/workflows/pre-commit.yml

@@ -10,10 +10,11 @@ jobs:
     runs-on: ubuntu-latest
     steps:
     - uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
-    - uses: actions/setup-python@0b93645e9fea7318ecaed2b359559ac225c90a2b # v5.3.0
+    - 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

.github/workflows/publish.yml

@@ -39,7 +39,7 @@ jobs:
             const script = require('.github/workflows/scripts/create_release.js')
             await script(github, context, core)
 
-  # NOTE(simon): No longer build wheel using Github Actions. See buildkite's release workflow. 
+  # NOTE(simon): No longer build wheel using GitHub Actions. See buildkite's release workflow. 
   # wheel:
   #   name: Build Wheel
   #   runs-on: ${{ matrix.os }}
@@ -50,7 +50,7 @@ jobs:
   #     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.
+  #         pytorch-version: ['2.4.0']  # Must be the most recent version that meets requirements/cuda.txt.
   #         cuda-version: ['11.8', '12.1']
 
   #   steps:

.github/workflows/reminder_comment.yml

@@ -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 }}

.github/workflows/scripts/build.sh

@@ -9,7 +9,7 @@ PATH=${cuda_home}/bin:$PATH
 LD_LIBRARY_PATH=${cuda_home}/lib64:$LD_LIBRARY_PATH
 
 # Install requirements
-$python_executable -m pip install -r requirements-build.txt -r requirements-cuda.txt
+$python_executable -m pip install -r requirements/build.txt -r requirements/cuda.txt
 
 # Limit the number of parallel jobs to avoid OOM
 export MAX_JOBS=1

.github/workflows/scripts/create_release.js

@@ -1,4 +1,4 @@
-// Uses Github's API to create the release and wait for result.
+// Uses GitHub's API to create the release and wait for result.
 // We use a JS script since github CLI doesn't provide a way to wait for the release's creation and returns immediately.
 
 module.exports = async (github, context, core) => {

.github/workflows/stale.yml

@@ -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

.gitignore

@@ -2,7 +2,8 @@
 /vllm/_version.py
 
 # vllm-flash-attn built from source
-vllm/vllm_flash_attn/
+vllm/vllm_flash_attn/*
+!vllm/vllm_flash_attn/fa_utils.py
 
 # Byte-compiled / optimized / DLL files
 __pycache__/
@@ -197,7 +198,7 @@ _build/
 hip_compat.h
 
 # Benchmark dataset
-benchmarks/*.json
+benchmarks/**/*.json
 
 # Linting
 actionlint

.pre-commit-config.yaml

@@ -1,43 +1,51 @@
 default_stages:
   - pre-commit # Run locally
   - manual # Run in CI
+exclude: 'vllm/third_party/.*'
 repos:
 - repo: https://github.com/google/yapf
-  rev: v0.32.0
+  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: v0.6.5
+  rev: v0.9.3
   hooks:
   - id: ruff
-    args: [--output-format, github]
+    args: [--output-format, github, --fix]
 - repo: https://github.com/codespell-project/codespell
-  rev: v2.3.0
+  rev: v2.4.0
   hooks:
   - id: codespell
-    exclude: 'benchmarks/sonnet.txt|(build|tests/(lora/data|models/fixtures|prompts))/.*'
+    additional_dependencies: ['tomli']
+    args: ['--toml', 'pyproject.toml']
 - repo: https://github.com/PyCQA/isort
-  rev: 5.13.2
+  rev: 0a0b7a830386ba6a31c2ec8316849ae4d1b8240d # 6.0.0
   hooks:
   - id: isort
 - repo: https://github.com/pre-commit/mirrors-clang-format
-  rev: v18.1.5
+  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))'
+    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
-    files: docs/.*
+    args: [fix]
 - repo: https://github.com/rhysd/actionlint
-  rev: v1.7.6
+  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
@@ -45,7 +53,7 @@ repos:
     entry: tools/mypy.sh 0 "local"
     language: python
     types: [python]
-    additional_dependencies: &mypy_deps [mypy==1.11.1, types-setuptools, types-PyYAML, types-requests]
+    additional_dependencies: &mypy_deps [mypy==1.11.1, types-cachetools, 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
@@ -85,8 +93,37 @@ repos:
     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

.readthedocs.yaml

@@ -18,4 +18,4 @@ formats: []
 # Optionally declare the Python requirements required to build your docs
 python:
   install:
-    - requirements: docs/requirements-docs.txt
+    - requirements: requirements/docs.txt

CMakeLists.txt

@@ -31,10 +31,10 @@ set(ignoreMe "${VLLM_PYTHON_PATH}")
 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.
@@ -46,8 +46,8 @@ set(HIP_SUPPORTED_ARCHS "gfx906;gfx908;gfx90a;gfx940;gfx941;gfx942;gfx1030;gfx11
 # requirements.txt files and should be kept consistent.  The ROCm torch
 # versions are derived from Dockerfile.rocm
 #
-set(TORCH_SUPPORTED_VERSION_CUDA "2.5.1")
-set(TORCH_SUPPORTED_VERSION_ROCM "2.5.1")
+set(TORCH_SUPPORTED_VERSION_CUDA "2.6.0")
+set(TORCH_SUPPORTED_VERSION_ROCM "2.6.0")
 
 #
 # Try to find python package with an executable that exactly matches
@@ -174,6 +174,25 @@ 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
 #
@@ -192,7 +211,7 @@ set_gencode_flags_for_srcs(
 if(VLLM_GPU_LANG STREQUAL "CUDA")
   message(STATUS "Enabling cumem allocator extension.")
   # link against cuda driver library
-  list(APPEND CUMEM_LIBS cuda)
+  list(APPEND CUMEM_LIBS CUDA::cuda_driver)
   define_gpu_extension_target(
     cumem_allocator
     DESTINATION vllm
@@ -228,7 +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 "v3.6.0" 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})
@@ -245,7 +265,8 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
     FetchContent_Declare(
         cutlass
         GIT_REPOSITORY https://github.com/nvidia/cutlass.git
-        GIT_TAG v3.6.0
+        # 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.
@@ -264,8 +285,9 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
     "csrc/custom_all_reduce.cu"
     "csrc/permute_cols.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/sparse/cutlass/sparse_compressor_entry.cu"
     "csrc/cutlass_extensions/common.cpp")
 
   set_gencode_flags_for_srcs(
@@ -275,7 +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"
@@ -295,38 +317,87 @@ 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" "${CUDA_ARCHS}")
+  if (ALLSPARK_ARCHS)
+    set(ALLSPARK_SRCS
+       "csrc/quantization/gptq_allspark/allspark_repack.cu"
+       "csrc/quantization/gptq_allspark/allspark_qgemm_w8a16.cu")
+    set_gencode_flags_for_srcs(
+      SRCS "${ALLSPARK_SRCS}"
+      CUDA_ARCHS "${ALLSPARK_ARCHS}")
+    list(APPEND VLLM_EXT_SRC "${ALLSPARK_SRCS}")
+    message(STATUS "Building AllSpark kernels for archs: ${ALLSPARK_ARCHS}")
+  else()
+    message(STATUS "Not building AllSpark kernels as no compatible archs found"
+                   " in CUDA target architectures")
+  endif()
+
+
+  set(SCALED_MM_3X_ARCHS)
   # The cutlass_scaled_mm kernels for Hopper (c3x, i.e. CUTLASS 3.x) require
-  # CUDA 12.0 or later (and only work on Hopper, 9.0/9.0a for now).
-  cuda_archs_loose_intersection(SCALED_MM_3X_ARCHS "9.0;9.0a" "${CUDA_ARCHS}")
-  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.0 AND SCALED_MM_3X_ARCHS)
-    set(SRCS "csrc/quantization/cutlass_w8a8/scaled_mm_c3x.cu")
+  # CUDA 12.0 or later
+  cuda_archs_loose_intersection(SCALED_MM_ARCHS "9.0a;" "${CUDA_ARCHS}")
+  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.0 AND SCALED_MM_ARCHS)
+    set(SRCS
+       "csrc/quantization/cutlass_w8a8/scaled_mm_c3x_sm90.cu"
+       "csrc/quantization/cutlass_w8a8/c3x/scaled_mm_sm90_fp8.cu"
+       "csrc/quantization/cutlass_w8a8/c3x/scaled_mm_sm90_int8.cu"
+       "csrc/quantization/cutlass_w8a8/c3x/scaled_mm_azp_sm90_int8.cu"
+       "csrc/quantization/cutlass_w8a8/c3x/scaled_mm_blockwise_sm90_fp8.cu")
     set_gencode_flags_for_srcs(
       SRCS "${SRCS}"
-      CUDA_ARCHS "${SCALED_MM_3X_ARCHS}")
+      CUDA_ARCHS "${SCALED_MM_ARCHS}")
     list(APPEND VLLM_EXT_SRC "${SRCS}")
-    list(APPEND VLLM_GPU_FLAGS "-DENABLE_SCALED_MM_C3X=1")
-    message(STATUS "Building scaled_mm_c3x for archs: ${SCALED_MM_3X_ARCHS}")
+    list(APPEND VLLM_GPU_FLAGS "-DENABLE_SCALED_MM_SM90=1")
+    # Let scaled_mm_c2x know it doesn't need to build these arches
+    list(APPEND SCALED_MM_3X_ARCHS "${SCALED_MM_ARCHS}")
+    message(STATUS "Building scaled_mm_c3x_sm90 for archs: ${SCALED_MM_ARCHS}")
   else()
-    if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.0 AND SCALED_MM_3X_ARCHS)
-      message(STATUS "Not building scaled_mm_c3x as CUDA Compiler version is "
+    if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.0 AND SCALED_MM_ARCHS)
+      message(STATUS "Not building scaled_mm_c3x_sm90 as CUDA Compiler version is "
                      "not >= 12.0, we recommend upgrading to CUDA 12.0 or "
                      "later if you intend on running FP8 quantized models on "
                      "Hopper.")
     else()
-      message(STATUS "Not building scaled_mm_c3x as no compatible archs found "
+      message(STATUS "Not building scaled_mm_c3x_sm90 as no compatible archs found "
                      "in CUDA target architectures")
     endif()
+  endif()
 
-    # clear SCALED_MM_3X_ARCHS so the scaled_mm_c2x kernels know we didn't
-    # build any 3x kernels
-    set(SCALED_MM_3X_ARCHS)
+  # The cutlass_scaled_mm kernels for Blackwell (c3x, i.e. CUTLASS 3.x) require
+  # CUDA 12.8 or later
+  cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0a;10.1a;12.0a" "${CUDA_ARCHS}")
+  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.8 AND SCALED_MM_ARCHS)
+    set(SRCS
+      "csrc/quantization/cutlass_w8a8/scaled_mm_c3x_sm100.cu"
+      "csrc/quantization/cutlass_w8a8/c3x/scaled_mm_sm100_fp8.cu"
+    )
+    set_gencode_flags_for_srcs(
+      SRCS "${SRCS}"
+      CUDA_ARCHS "${SCALED_MM_ARCHS}")
+    list(APPEND VLLM_EXT_SRC "${SRCS}")
+    list(APPEND VLLM_GPU_FLAGS "-DENABLE_SCALED_MM_SM100=1")
+    # Let scaled_mm_c2x know it doesn't need to build these arches
+    list(APPEND SCALED_MM_3X_ARCHS "${SCALED_MM_ARCHS}")
+    message(STATUS "Building scaled_mm_c3x_sm100 for archs: ${SCALED_MM_ARCHS}")
+  else()
+    if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.8 AND SCALED_MM_ARCHS)
+      message(STATUS "Not building scaled_mm_c3x_sm100 as CUDA Compiler version is "
+                     "not >= 12.8, we recommend upgrading to CUDA 12.8 or "
+                     "later if you intend on running FP8 quantized models on "
+                     "Blackwell.")
+    else()
+      message(STATUS "Not building scaled_mm_c3x_100 as no compatible archs found "
+                     "in CUDA target architectures")
+    endif()
   endif()
 
   #
   # 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)
@@ -351,18 +422,18 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
   # 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, 9.0/9.0a for now).
-  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.2 AND SCALED_MM_3X_ARCHS)
-    set(SRCS "csrc/sparse/cutlass/sparse_compressor_c3x.cu"
-             "csrc/sparse/cutlass/sparse_scaled_mm_c3x.cu")
+  # require CUDA 12.2 or later (and only work on Hopper).
+  cuda_archs_loose_intersection(SCALED_MM_ARCHS "9.0a;" "${CUDA_ARCHS}")
+  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.2 AND SCALED_MM_ARCHS)
+    set(SRCS "csrc/sparse/cutlass/sparse_scaled_mm_c3x.cu")
     set_gencode_flags_for_srcs(
       SRCS "${SRCS}"
-      CUDA_ARCHS "${SCALED_MM_3X_ARCHS}")
+      CUDA_ARCHS "${SCALED_MM_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}")
+    message(STATUS "Building sparse_scaled_mm_c3x for archs: ${SCALED_MM_ARCHS}")
   else()
-    if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.2 AND SCALED_MM_3X_ARCHS)
+    if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.2 AND SCALED_MM_ARCHS)
       message(STATUS "Not building sparse_scaled_mm_c3x kernels as CUDA Compiler version is "
                      "not >= 12.2, we recommend upgrading to CUDA 12.2 or later "
                      "if you intend on running FP8 sparse quantized models on Hopper.")
@@ -372,6 +443,23 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
     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()
 
   #
   # Machete kernels
@@ -446,9 +534,6 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
 endif()
 
 message(STATUS "Enabling C extension.")
-if(VLLM_GPU_LANG STREQUAL "CUDA")
-  list(APPEND VLLM_C_LIBS cuda)
-endif()
 define_gpu_extension_target(
   _C
   DESTINATION vllm
@@ -456,8 +541,8 @@ define_gpu_extension_target(
   SOURCES ${VLLM_EXT_SRC}
   COMPILE_FLAGS ${VLLM_GPU_FLAGS}
   ARCHITECTURES ${VLLM_GPU_ARCHES}
-  INCLUDE_DIRECTORIES ${CUTLASS_INCLUDE_DIR};${CUTLASS_TOOLS_UTIL_INCLUDE_DIR}
-  LIBRARIES ${VLLM_C_LIBS}
+  INCLUDE_DIRECTORIES ${CUTLASS_INCLUDE_DIR}
+  INCLUDE_DIRECTORIES ${CUTLASS_TOOLS_UTIL_INCLUDE_DIR}
   USE_SABI 3
   WITH_SOABI)
 
@@ -476,12 +561,24 @@ set(VLLM_MOE_EXT_SRC
   "csrc/moe/moe_align_sum_kernels.cu"
   "csrc/moe/topk_softmax_kernels.cu")
 
+if(VLLM_GPU_LANG STREQUAL "CUDA")
+  list(APPEND VLLM_MOE_EXT_SRC "csrc/moe/moe_wna16.cu")
+endif()
+
 set_gencode_flags_for_srcs(
   SRCS "${VLLM_MOE_EXT_SRC}"
   CUDA_ARCHS "${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" "${CUDA_ARCHS}")
+  set(VLLM_MOE_WNA16_SRC
+    "csrc/moe/moe_wna16.cu")
+
+  set_gencode_flags_for_srcs(
+    SRCS "${VLLM_MOE_WNA16_SRC}"
+    CUDA_ARCHS "${CUDA_ARCHS}")
+
+  list(APPEND VLLM_MOE_EXT_SRC "${VLLM_MOE_WNA16_SRC}")
+  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"
@@ -535,77 +632,8 @@ if(VLLM_GPU_LANG STREQUAL "HIP")
     WITH_SOABI)
 endif()
 
-# vllm-flash-attn currently only supported on CUDA
-if (NOT VLLM_GPU_LANG 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_fa2_C (for FA2) or --component _vllm_fa3_C (for FA3).
-# 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}
-          BINARY_DIR ${CMAKE_BINARY_DIR}/vllm-flash-attn
-  )
-else()
-  FetchContent_Declare(
-          vllm-flash-attn
-          GIT_REPOSITORY https://github.com/vllm-project/flash-attention.git
-          GIT_TAG d4e09037abf588af1ec47d0e966b237ee376876c
-          GIT_PROGRESS TRUE
-          # Don't share the vllm-flash-attn build between build types
-          BINARY_DIR ${CMAKE_BINARY_DIR}/vllm-flash-attn
-  )
-endif()
-
-
-# Fetch the vllm-flash-attn library
-FetchContent_MakeAvailable(vllm-flash-attn)
-message(STATUS "vllm-flash-attn is available at ${vllm-flash-attn_SOURCE_DIR}")
-
-# Copy over the vllm-flash-attn python files (duplicated for fa2 and fa3, in
-# case only one is built, in the case both are built redundant work is done)
-install(
-  DIRECTORY ${vllm-flash-attn_SOURCE_DIR}/vllm_flash_attn/
-  DESTINATION vllm_flash_attn
-  COMPONENT _vllm_fa2_C
-  FILES_MATCHING PATTERN "*.py"
-)
-
-install(
-  DIRECTORY ${vllm-flash-attn_SOURCE_DIR}/vllm_flash_attn/
-  DESTINATION vllm_flash_attn
-  COMPONENT _vllm_fa3_C
-  FILES_MATCHING PATTERN "*.py"
-)
-
-# Nothing after vllm-flash-attn, see comment about macros above

CODE_OF_CONDUCT.md

@@ -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).
-

Dockerfile

@@ -27,6 +27,13 @@ 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
+
+# This timeout (in seconds) is necessary when installing some dependencies via uv since it's likely to time out
+# Reference: https://github.com/astral-sh/uv/pull/1694
+ENV UV_HTTP_TIMEOUT=500
 
 # 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
@@ -50,15 +57,16 @@ WORKDIR /workspace
 # 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/pip \
+RUN --mount=type=cache,target=/root/.cache/uv \
     if [ "$TARGETPLATFORM" = "linux/arm64" ]; then \
-        python3 -m pip install --index-url https://download.pytorch.org/whl/nightly/cu126 "torch==2.7.0.dev20250121+cu126" "torchvision==0.22.0.dev20250121";  \
+        uv pip install --system --index-url https://download.pytorch.org/whl/nightly/cu128 "torch==2.8.0.dev20250318+cu128" "torchvision==0.22.0.dev20250319";  \
+        uv pip install --system --index-url https://download.pytorch.org/whl/nightly/cu128 --pre pytorch_triton==3.3.0+gitab727c40; \
     fi
 
-COPY requirements-common.txt requirements-common.txt
-COPY requirements-cuda.txt requirements-cuda.txt
-RUN --mount=type=cache,target=/root/.cache/pip \
-    python3 -m pip install -r requirements-cuda.txt
+COPY requirements/common.txt requirements/common.txt
+COPY requirements/cuda.txt requirements/cuda.txt
+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`
@@ -76,15 +84,19 @@ FROM base AS build
 ARG TARGETPLATFORM
 
 # install build dependencies
-COPY requirements-build.txt requirements-build.txt
+COPY requirements/build.txt requirements/build.txt
 
-RUN --mount=type=cache,target=/root/.cache/pip \
-    python3 -m pip install -r requirements-build.txt
+# This timeout (in seconds) is necessary when installing some dependencies via uv since it's likely to time out
+# Reference: https://github.com/astral-sh/uv/pull/1694
+ENV UV_HTTP_TIMEOUT=500
+
+RUN --mount=type=cache,target=/root/.cache/uv \
+    uv pip install --system -r requirements/build.txt
 
 COPY . .
 ARG GIT_REPO_CHECK=0
 RUN --mount=type=bind,source=.git,target=.git \
-    if [ "$GIT_REPO_CHECK" != 0 ]; then bash tools/check_repo.sh ; fi
+    if [ "$GIT_REPO_CHECK" != "0" ]; then bash tools/check_repo.sh ; fi
 
 # max jobs used by Ninja to build extensions
 ARG max_jobs=2
@@ -98,7 +110,7 @@ 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..." \
@@ -118,16 +130,19 @@ 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 \
+        # Clean any existing CMake artifacts
+        rm -rf .deps && \
+        mkdir -p .deps && \
         python3 setup.py bdist_wheel --dist-dir=dist --py-limited-api=cp38; \
     fi
 
 # Check the size of the wheel if RUN_WHEEL_CHECK is true
 COPY .buildkite/check-wheel-size.py check-wheel-size.py
 # sync the default value with .buildkite/check-wheel-size.py
-ARG VLLM_MAX_SIZE_MB=300
+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 \
@@ -140,16 +155,21 @@ RUN if [ "$RUN_WHEEL_CHECK" = "true" ]; then \
 #################### DEV IMAGE ####################
 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
+# This timeout (in seconds) is necessary when installing some dependencies via uv since it's likely to time out
+# Reference: https://github.com/astral-sh/uv/pull/1694
+ENV UV_HTTP_TIMEOUT=500
+
+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/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
@@ -173,6 +193,13 @@ 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
+
+# This timeout (in seconds) is necessary when installing some dependencies via uv since it's likely to time out
+# Reference: https://github.com/astral-sh/uv/pull/1694
+ENV UV_HTTP_TIMEOUT=500
 
 # Workaround for https://github.com/openai/triton/issues/2507 and
 # https://github.com/pytorch/pytorch/issues/107960 -- hopefully
@@ -184,22 +211,44 @@ RUN ldconfig /usr/local/cuda-$(echo $CUDA_VERSION | cut -d. -f1,2)/compat/
 # we need to install torch and torchvision from the nightly builds first,
 # pytorch will not appear as a vLLM dependency in all of the following steps
 # after this step
-RUN --mount=type=cache,target=/root/.cache/pip \
+RUN --mount=type=cache,target=/root/.cache/uv \
     if [ "$TARGETPLATFORM" = "linux/arm64" ]; then \
-        python3 -m pip install --index-url https://download.pytorch.org/whl/nightly/cu124 "torch==2.6.0.dev20241210+cu124" "torchvision==0.22.0.dev20241215";  \
+        uv pip install --system --index-url https://download.pytorch.org/whl/nightly/cu128 "torch==2.8.0.dev20250318+cu128" "torchvision==0.22.0.dev20250319";  \
+        uv pip install --system --index-url https://download.pytorch.org/whl/nightly/cu128 --pre pytorch_triton==3.3.0+gitab727c40; \
     fi
 
 # 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
+    --mount=type=cache,target=/root/.cache/uv \
+    uv pip install --system dist/*.whl --verbose
 
-RUN --mount=type=cache,target=/root/.cache/pip \
+# 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.post2/flashinfer_python-0.2.1.post2+cu124torch2.6-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 ####################
@@ -209,17 +258,21 @@ 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
+# This timeout (in seconds) is necessary when installing some dependencies via uv since it's likely to time out
+# Reference: https://github.com/astral-sh/uv/pull/1694
+ENV UV_HTTP_TIMEOUT=500
 
 # 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 -r requirements/dev.txt
+
+# install development dependencies (for testing)
+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)
@@ -237,12 +290,16 @@ RUN mv vllm test_docs/
 # base openai image with additional requirements, for any subsequent openai-style images
 FROM vllm-base AS vllm-openai-base
 
+# This timeout (in seconds) is necessary when installing some dependencies via uv since it's likely to time out
+# Reference: https://github.com/astral-sh/uv/pull/1694
+ENV UV_HTTP_TIMEOUT=500
+
 # 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' boto3 runai-model-streamer runai-model-streamer[s3]; \
+        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' boto3 runai-model-streamer runai-model-streamer[s3]; \
+        uv pip install --system accelerate hf_transfer 'modelscope!=1.15.0' 'bitsandbytes>=0.45.3' 'timm==0.9.10' boto3 runai-model-streamer runai-model-streamer[s3]; \
     fi
 
 ENV VLLM_USAGE_SOURCE production-docker-image

Dockerfile.arm

@@ -26,18 +26,18 @@ WORKDIR /workspace
 ARG PIP_EXTRA_INDEX_URL="https://download.pytorch.org/whl/cpu"
 ENV PIP_EXTRA_INDEX_URL=${PIP_EXTRA_INDEX_URL}
 RUN --mount=type=cache,target=/root/.cache/pip \
-    --mount=type=bind,src=requirements-build.txt,target=requirements-build.txt \
+    --mount=type=bind,src=requirements/build.txt,target=requirements/build.txt \
     pip install --upgrade pip && \
-    pip install -r requirements-build.txt
+    pip install -r requirements/build.txt
 
 FROM cpu-test-arm AS build
 
 WORKDIR /workspace/vllm
 
 RUN --mount=type=cache,target=/root/.cache/pip \
-    --mount=type=bind,src=requirements-common.txt,target=requirements-common.txt \
-    --mount=type=bind,src=requirements-cpu.txt,target=requirements-cpu.txt \
-    pip install -v -r requirements-cpu.txt
+    --mount=type=bind,src=requirements/common.txt,target=requirements/common.txt \
+    --mount=type=bind,src=requirements/cpu.txt,target=requirements/cpu.txt \
+    pip install -v -r requirements/cpu.txt
 
 COPY . .
 ARG GIT_REPO_CHECK=0

Dockerfile.cpu

@@ -22,25 +22,25 @@ ENV LD_PRELOAD="/usr/lib/x86_64-linux-gnu/libtcmalloc_minimal.so.4:/usr/local/li
 
 RUN echo 'ulimit -c 0' >> ~/.bashrc
 
-RUN pip install intel_extension_for_pytorch==2.5.0
+RUN pip install intel_extension_for_pytorch==2.6.0
 
 WORKDIR /workspace
 
 ARG PIP_EXTRA_INDEX_URL="https://download.pytorch.org/whl/cpu"
 ENV PIP_EXTRA_INDEX_URL=${PIP_EXTRA_INDEX_URL}
 RUN --mount=type=cache,target=/root/.cache/pip \
-    --mount=type=bind,src=requirements-build.txt,target=requirements-build.txt \
+    --mount=type=bind,src=requirements/build.txt,target=requirements/build.txt \
     pip install --upgrade pip && \
-    pip install -r requirements-build.txt
+    pip install -r requirements/build.txt
 
 FROM cpu-test-1 AS build
 
 WORKDIR /workspace/vllm
 
 RUN --mount=type=cache,target=/root/.cache/pip \
-    --mount=type=bind,src=requirements-common.txt,target=requirements-common.txt \
-    --mount=type=bind,src=requirements-cpu.txt,target=requirements-cpu.txt \
-    pip install -v -r requirements-cpu.txt
+    --mount=type=bind,src=requirements/common.txt,target=requirements/common.txt \
+    --mount=type=bind,src=requirements/cpu.txt,target=requirements/cpu.txt \
+    pip install -v -r requirements/cpu.txt
 
 COPY . .
 ARG GIT_REPO_CHECK=0

Dockerfile.hpu

@@ -4,7 +4,7 @@ COPY ./ /workspace/vllm
 
 WORKDIR /workspace/vllm
 
-RUN pip install -v -r requirements-hpu.txt
+RUN pip install -v -r requirements/hpu.txt
 
 ENV no_proxy=localhost,127.0.0.1
 ENV PT_HPU_ENABLE_LAZY_COLLECTIVES=true

Dockerfile.neuron

@@ -23,10 +23,12 @@ 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==4.45.2 -U
-RUN python3 -m pip install transformers-neuronx --extra-index-url=https://pip.repos.neuron.amazonaws.com -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
 RUN --mount=type=bind,source=.git,target=.git \
@@ -34,7 +36,7 @@ RUN --mount=type=bind,source=.git,target=.git \
 
 RUN python3 -m pip install -U \
         'cmake>=3.26' ninja packaging 'setuptools-scm>=8' wheel jinja2 \
-        -r requirements-neuron.txt
+        -r requirements/neuron.txt
 
 ENV VLLM_TARGET_DEVICE neuron
 RUN --mount=type=bind,source=.git,target=.git \
@@ -43,6 +45,10 @@ 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
 

Dockerfile.openvino

@@ -1,29 +0,0 @@
-# The vLLM Dockerfile is used to construct vLLM image that can be directly used
-# to run the OpenAI compatible server.
-
-FROM ubuntu:22.04 AS dev
-
-RUN apt-get update -y && \
-    apt-get install -y \
-        git python3-pip \
-        ffmpeg libsm6 libxext6 libgl1
-WORKDIR /workspace
-
-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 -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
-RUN PIP_EXTRA_INDEX_URL="https://download.pytorch.org/whl/cpu" VLLM_TARGET_DEVICE="openvino" python3 -m pip install /workspace
-
-COPY examples/ /workspace/examples
-COPY benchmarks/ /workspace/benchmarks
-
-# install development dependencies (for testing)
-RUN python3 -m pip install -e tests/vllm_test_utils
-
-CMD ["/bin/bash"]

Dockerfile.ppc64le

@@ -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 libssl-dev 
+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
+# 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
 
@@ -21,8 +21,7 @@ RUN --mount=type=bind,source=.git,target=.git \
 RUN --mount=type=cache,target=/root/.cache/pip  \
     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 \
+        -r requirements/cpu.txt \
         xformers uvloop==0.20.0
 
 RUN --mount=type=bind,source=.git,target=.git \

Dockerfile.rocm

@@ -38,14 +38,14 @@ FROM fetch_vllm AS build_vllm
 ARG USE_CYTHON
 # Build vLLM
 RUN cd vllm \
-    && python3 -m pip install -r requirements-rocm.txt \
+    && 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 \
+    && if [ ${USE_CYTHON} -eq "1" ]; then python3 tests/build_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/requirements /requirements
 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
@@ -60,7 +60,8 @@ RUN python3 -m pip install --upgrade pip && rm -rf /var/lib/apt/lists/*
 # Install vLLM
 RUN --mount=type=bind,from=export_vllm,src=/,target=/install \
     cd /install \
-    && pip install -U -r requirements-rocm.txt \
+    && pip install -U -r requirements/rocm.txt \
+    && pip install -U -r requirements/rocm-test.txt \
     && pip uninstall -y vllm \
     && pip install *.whl
 
@@ -99,7 +100,7 @@ RUN if [ ${BUILD_RPD} -eq "1" ]; then \
 # Install vLLM
 RUN --mount=type=bind,from=export_vllm,src=/,target=/install \
     cd /install \
-    && pip install -U -r requirements-rocm.txt \
+    && pip install -U -r requirements/rocm.txt \
     && pip uninstall -y vllm \
     && pip install *.whl
 

Dockerfile.rocm_base

@@ -6,12 +6,14 @@ 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="8d4926e"
+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"
+ARG AITER_BRANCH="21d47a9"
+ARG AITER_REPO="https://github.com/ROCm/aiter.git"
 
 FROM ${BASE_IMAGE} AS base
 
@@ -129,8 +131,18 @@ RUN --mount=type=bind,from=build_amdsmi,src=/app/install/,target=/install \
 RUN --mount=type=bind,from=build_pytorch,src=/app/install/,target=/install \
     pip install /install/*.whl
 
+ARG AITER_REPO
+ARG AITER_BRANCH
+RUN git clone --recursive ${AITER_REPO}
+RUN cd aiter \
+    && git checkout ${AITER_BRANCH} \
+    && git submodule update --init --recursive \
+    && pip install -r requirements.txt \
+    && PREBUILD_KERNELS=1 GPU_ARCHS=gfx942 python3 setup.py develop && pip show aiter 
+
 ARG BASE_IMAGE
 ARG HIPBLASLT_BRANCH
+ARG HIPBLAS_COMMON_BRANCH
 ARG LEGACY_HIPBLASLT_OPTION
 ARG RCCL_BRANCH
 ARG RCCL_REPO
@@ -155,4 +167,6 @@ RUN echo "BASE_IMAGE: ${BASE_IMAGE}" > /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
+    && echo "FA_REPO: ${FA_REPO}" >> /app/versions.txt \
+    && echo "AITER_BRANCH: ${AITER_BRANCH}" >> /app/versions.txt \
+    && echo "AITER_REPO: ${AITER_REPO}" >> /app/versions.txt

Dockerfile.s390x

@@ -0,0 +1,152 @@
+# Base UBI image for s390x architecture
+ARG BASE_UBI_IMAGE_TAG=9.5-1736404155
+ARG PYTHON_VERSION=3.12
+FROM registry.access.redhat.com/ubi9/ubi-minimal:${BASE_UBI_IMAGE_TAG} AS base
+
+# Install basic dependencies
+ARG PYTHON_VERSION
+ENV PYTHON_VERSION=${PYTHON_VERSION}
+
+WORKDIR /workspace
+
+ENV LANG=C.UTF-8 \
+    LC_ALL=C.UTF-8
+
+# Install development utilities
+RUN microdnf install -y \
+    which procps findutils tar vim git gcc gcc-gfortran g++ make patch zlib-devel \
+    libjpeg-turbo-devel libtiff-devel libpng-devel libwebp-devel freetype-devel harfbuzz-devel \
+    openssl-devel openblas openblas-devel autoconf automake libtool cmake && \
+    microdnf clean all
+
+# Python Installation
+FROM base AS python-install
+ARG PYTHON_VERSION
+
+ENV VIRTUAL_ENV=/opt/vllm
+ENV PATH="$VIRTUAL_ENV/bin:$PATH"
+ENV PYTHON_VERSION=${PYTHON_VERSION}
+RUN microdnf install -y \
+    python${PYTHON_VERSION}-devel python${PYTHON_VERSION}-pip python${PYTHON_VERSION}-wheel  && \
+    python${PYTHON_VERSION} -m venv $VIRTUAL_ENV && pip install --no-cache -U pip wheel uv && microdnf clean all
+
+FROM python-install AS pyarrow
+
+# Build Apache Arrow
+WORKDIR /tmp
+RUN --mount=type=cache,target=/root/.cache/uv \
+    git clone https://github.com/apache/arrow.git && \
+    cd arrow/cpp && \
+    mkdir release && cd release && \
+    cmake -DCMAKE_BUILD_TYPE=Release \
+          -DCMAKE_INSTALL_PREFIX=/usr/local \
+          -DARROW_PYTHON=ON \
+          -DARROW_PARQUET=ON \
+          -DARROW_ORC=ON \
+          -DARROW_FILESYSTEM=ON \
+          -DARROW_WITH_LZ4=ON \
+          -DARROW_WITH_ZSTD=ON \
+          -DARROW_WITH_SNAPPY=ON \
+          -DARROW_JSON=ON \
+          -DARROW_CSV=ON \
+          -DARROW_DATASET=ON \
+          -DPROTOBUF_PROTOC_EXECUTABLE=/usr/bin/protoc \
+          -DARROW_DEPENDENCY_SOURCE=BUNDLED \
+          .. && \
+    make -j$(nproc) && \
+    make install && \
+    cd ../../python && \
+    export PYARROW_PARALLEL=4 && \
+    export ARROW_BUILD_TYPE=release && \
+    uv pip install -r requirements/build.txt && \
+    python setup.py build_ext --build-type=$ARROW_BUILD_TYPE --bundle-arrow-cpp bdist_wheel
+
+FROM python-install AS numa-build
+# Install numactl (needed for numa.h dependency)
+WORKDIR /tmp
+RUN curl -LO https://github.com/numactl/numactl/archive/refs/tags/v2.0.16.tar.gz && \
+    tar -xvzf v2.0.16.tar.gz && \
+    cd numactl-2.0.16 && \
+    ./autogen.sh && \
+    ./configure && \
+    make
+
+# Set include path
+ENV C_INCLUDE_PATH="/usr/local/include:$C_INCLUDE_PATH"
+
+FROM python-install AS rust
+ENV CARGO_HOME=/root/.cargo
+ENV RUSTUP_HOME=/root/.rustup
+ENV PATH="$CARGO_HOME/bin:$RUSTUP_HOME/bin:$PATH"
+
+RUN curl https://sh.rustup.rs -sSf | sh -s -- -y && \
+    . "$CARGO_HOME/env" && \
+    rustup default stable && \
+    rustup show
+
+FROM python-install AS torch-vision
+# Install torchvision
+ARG TORCH_VERSION=2.7.0.dev20250304
+ARG TORCH_VISION_VERSION=v0.20.1
+WORKDIR /tmp
+RUN --mount=type=cache,target=/root/.cache/uv \
+    git clone https://github.com/pytorch/vision.git && \
+    cd vision && \
+    git checkout $TORCH_VISION_VERSION && \
+    uv pip install -v torch==${TORCH_VERSION} --extra-index-url https://download.pytorch.org/whl/nightly/cpu && \
+    python setup.py bdist_wheel
+
+# Final build stage
+FROM python-install AS vllm-cpu
+ARG PYTHON_VERSION
+
+# Set correct library path for torch and numactl
+ENV LD_LIBRARY_PATH="/opt/vllm/lib64/python${PYTHON_VERSION}/site-packages/torch/lib:/usr/local/lib:$LD_LIBRARY_PATH"
+ENV C_INCLUDE_PATH="/usr/local/include:$C_INCLUDE_PATH"
+ENV UV_LINK_MODE=copy
+ENV CARGO_HOME=/root/.cargo
+ENV RUSTUP_HOME=/root/.rustup
+ENV PATH="$CARGO_HOME/bin:$RUSTUP_HOME/bin:$PATH"
+
+COPY . /workspace/vllm
+WORKDIR /workspace/vllm
+
+RUN --mount=type=bind,from=numa-build,src=/tmp/numactl-2.0.16,target=/numactl \
+    make -C /numactl install
+
+# Install dependencies, including PyTorch and Apache Arrow
+RUN --mount=type=cache,target=/root/.cache/uv \
+    --mount=type=bind,from=rust,source=/root/.cargo,target=/root/.cargo,rw \
+    --mount=type=bind,from=rust,source=/root/.rustup,target=/root/.rustup,rw \
+    --mount=type=bind,from=pyarrow,source=/tmp/arrow/python/dist,target=/tmp/arrow-wheels \
+    --mount=type=bind,from=torch-vision,source=/tmp/vision/dist,target=/tmp/vision-wheels/ \
+     sed -i '/^torch/d' requirements/build.txt && \
+     ARROW_WHL_FILE=$(ls /tmp/arrow-wheels/pyarrow-*.whl | head -n 1) && \
+     VISION_WHL_FILE=$(ls /tmp/vision-wheels/*.whl | head -n 1) && \
+    uv pip install -v \    
+        $ARROW_WHL_FILE  \
+        $VISION_WHL_FILE \
+        --extra-index-url https://download.pytorch.org/whl/nightly/cpu \
+        --index-strategy unsafe-best-match \
+        -r requirements/build.txt \
+        -r requirements/cpu.txt 
+
+# Build and install vllm
+RUN --mount=type=cache,target=/root/.cache/uv \
+    VLLM_TARGET_DEVICE=cpu python setup.py bdist_wheel && \
+    uv pip install "$(echo dist/*.whl)[tensorizer]"
+
+# setup non-root user for vllm
+RUN umask 002 && \
+    useradd --uid 2000 --gid 0 vllm && \
+    mkdir -p /home/vllm && \
+    chmod g+rwx /home/vllm
+
+COPY LICENSE /licenses/vllm.md
+COPY examples/*.jinja /app/data/template/
+
+USER 2000
+WORKDIR /home/vllm
+
+# Set the default entrypoint
+ENTRYPOINT ["python", "-m", "vllm.entrypoints.openai.api_server"]

Dockerfile.tpu

@@ -15,11 +15,14 @@ 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
 
+# Remove existing versions of dependencies
+RUN pip uninstall -y torch torch_xla torchvision
+
 ENV VLLM_TARGET_DEVICE="tpu"
 RUN --mount=type=cache,target=/root/.cache/pip \
     --mount=type=bind,source=.git,target=.git \
     python3 -m pip install \
-        -r requirements-tpu.txt
+        -r requirements/tpu.txt
 RUN python3 setup.py develop
 
 # install development dependencies (for testing)

Dockerfile.xpu

@@ -1,11 +1,7 @@
-FROM intel/oneapi-basekit:2024.2.1-0-devel-ubuntu22.04 AS vllm-base
+# oneapi 2025.0.2 docker base image use rolling 2448 package. https://dgpu-docs.intel.com/releases/packages.html?release=Rolling+2448.13&os=Ubuntu+22.04, and we don't need install driver manually.
+FROM intel/deep-learning-essentials:2025.0.2-0-devel-ubuntu22.04 AS vllm-base
 
-RUN wget -O- https://apt.repos.intel.com/intel-gpg-keys/GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB | gpg --dearmor | tee /usr/share/keyrings/intel-oneapi-archive-keyring.gpg > /dev/null && \
-    echo "deb [signed-by=/usr/share/keyrings/intel-oneapi-archive-keyring.gpg] https://apt.repos.intel.com/oneapi all main " | tee /etc/apt/sources.list.d/oneAPI.list && \
-    chmod 644 /usr/share/keyrings/intel-oneapi-archive-keyring.gpg && \
-    wget -O- https://repositories.intel.com/graphics/intel-graphics.key | gpg --dearmor | tee /usr/share/keyrings/intel-graphics.gpg > /dev/null && \
-    echo "deb [arch=amd64,i386 signed-by=/usr/share/keyrings/intel-graphics.gpg] https://repositories.intel.com/graphics/ubuntu jammy arc" | tee /etc/apt/sources.list.d/intel.gpu.jammy.list && \
-    chmod 644 /usr/share/keyrings/intel-graphics.gpg
+RUN rm /etc/apt/sources.list.d/intel-graphics.list
 
 RUN apt-get update -y && \
     apt-get install -y --no-install-recommends --fix-missing \
@@ -21,30 +17,20 @@ RUN apt-get update -y && \
     python3 \
     python3-dev \
     python3-pip \
-    # vim \
     wget
 
 WORKDIR /workspace/vllm
-COPY requirements-xpu.txt /workspace/vllm/requirements-xpu.txt
-COPY requirements-common.txt /workspace/vllm/requirements-common.txt
+COPY requirements/xpu.txt /workspace/vllm/requirements/xpu.txt
+COPY requirements/common.txt /workspace/vllm/requirements/common.txt
 
 RUN --mount=type=cache,target=/root/.cache/pip \
     pip install --no-cache-dir \
-    -r requirements-xpu.txt
-
-RUN git clone https://github.com/intel/pti-gpu && \
-    cd pti-gpu/sdk && \
-    git checkout 6c491f07a777ed872c2654ca9942f1d0dde0a082 && \
-    mkdir build && \
-    cd build && \
-    cmake -DCMAKE_BUILD_TYPE=Release -DCMAKE_TOOLCHAIN_FILE=../cmake/toolchains/icpx_toolchain.cmake -DBUILD_TESTING=OFF .. && \
-    make -j && \
-    cmake --install . --config Release --prefix "/usr/local"
+    -r requirements/xpu.txt
 
 ENV LD_LIBRARY_PATH="$LD_LIBRARY_PATH:/usr/local/lib/"
 
 COPY . .
-ARG GIT_REPO_CHECK
+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
 
@@ -54,6 +40,12 @@ RUN --mount=type=cache,target=/root/.cache/pip \
     --mount=type=bind,source=.git,target=.git \
     python3 setup.py install
 
+# Please refer xpu doc, we need manually install intel-extension-for-pytorch 2.6.10+xpu due to there are some conflict dependencies with torch 2.6.0+xpu
+# FIXME: This will be fix in ipex 2.7. just leave this here for awareness.
+RUN --mount=type=cache,target=/root/.cache/pip \
+    pip install intel-extension-for-pytorch==2.6.10+xpu \
+    --extra-index-url=https://pytorch-extension.intel.com/release-whl/stable/xpu/us/
+
 CMD ["/bin/bash"]
 
 FROM vllm-base AS vllm-openai

MANIFEST.in

@@ -1,9 +1,9 @@
 include LICENSE
-include requirements-common.txt
-include requirements-cuda.txt
-include requirements-rocm.txt
-include requirements-neuron.txt
-include requirements-cpu.txt
+include requirements/common.txt
+include requirements/cuda.txt
+include requirements/rocm.txt
+include requirements/neuron.txt
+include requirements/cpu.txt
 include CMakeLists.txt
 
 recursive-include cmake *

README.md

@@ -10,14 +10,29 @@ 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://discuss.vllm.ai"><b>User Forum</b></a> | <a href="https://slack.vllm.ai"><b>Developer Slack</b></a> |
 </p>
 
 ---
 
+[2025/03] We are collaborating with Ollama to host an [Inference Night](https://lu.ma/vllm-ollama) at Y Combinator in San Francisco on Thursday, March 27, at 6 PM. Discuss all things inference local or data center!
+
+[2025/04] We're hosting our first-ever *vLLM Asia Developer Day* in Singapore on *April 3rd*! This is a full-day event (9 AM - 9 PM SGT) in partnership with SGInnovate, AMD, and Embedded LLM. Meet the vLLM team and learn about LLM inference for RL, MI300X, and more! [Register Now](https://www.sginnovate.com/event/limited-availability-morning-evening-slots-remaining-inaugural-vllm-asia-developer-day)
+
+---
+
 *Latest News* 🔥
-- [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).
+
+- [2025/03] We hosted [the first vLLM China Meetup](https://mp.weixin.qq.com/s/n77GibL2corAtQHtVEAzfg)! Please find the meetup slides from vLLM team [here](https://docs.google.com/presentation/d/1REHvfQMKGnvz6p3Fd23HhSO4c8j5WPGZV0bKYLwnHyQ/edit?usp=sharing).
+- [2025/03] We hosted [the East Coast vLLM Meetup](https://lu.ma/7mu4k4xx)! Please find the meetup slides [here](https://docs.google.com/presentation/d/1NHiv8EUFF1NLd3fEYODm56nDmL26lEeXCaDgyDlTsRs/edit#slide=id.g31441846c39_0_0).
+- [2025/02] We hosted [the ninth vLLM meetup](https://lu.ma/h7g3kuj9) with Meta! Please find the meetup slides from vLLM team [here](https://docs.google.com/presentation/d/1jzC_PZVXrVNSFVCW-V4cFXb6pn7zZ2CyP_Flwo05aqg/edit?usp=sharing) and AMD [here](https://drive.google.com/file/d/1Zk5qEJIkTmlQ2eQcXQZlljAx3m9s7nwn/view?usp=sharing). The slides from Meta will not be posted.
+- [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!
+
+<details>
+<summary>Previous News</summary>
+
 - [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!
 - [2024/10] Ray Summit 2024 held a special track for vLLM! Please find the opening talk slides from the vLLM team [here](https://docs.google.com/presentation/d/1B_KQxpHBTRa_mDF-tR6i8rWdOU5QoTZNcEg2MKZxEHM/edit?usp=sharing). Learn more from the [talks](https://www.youtube.com/playlist?list=PLzTswPQNepXl6AQwifuwUImLPFRVpksjR) from other vLLM contributors and users!
@@ -31,11 +46,14 @@ Easy, fast, and cheap LLM serving for everyone
 - [2023/08] We would like to express our sincere gratitude to [Andreessen Horowitz](https://a16z.com/2023/08/30/supporting-the-open-source-ai-community/) (a16z) for providing a generous grant to support the open-source development and research of vLLM.
 - [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).
 
+</details>
+
 ---
 ## 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 evloved into a community-driven project with contributions from both academia and industry.
+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:
 
@@ -78,7 +96,7 @@ pip install vllm
 ```
 
 Visit our [documentation](https://docs.vllm.ai/en/latest/) to learn more.
-- [Installation](https://docs.vllm.ai/en/latest/getting_started/installation/index.html)
+- [Installation](https://docs.vllm.ai/en/latest/getting_started/installation.html)
 - [Quickstart](https://docs.vllm.ai/en/latest/getting_started/quickstart.html)
 - [List of Supported Models](https://docs.vllm.ai/en/latest/models/supported_models.html)
 
@@ -126,6 +144,7 @@ We also have an official fundraising venue through [OpenCollective](https://open
 ## 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},
@@ -137,12 +156,12 @@ 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](https://github.com/vllm-project/vllm/issues) or [Discussions](https://github.com/vllm-project/vllm/discussions)
+- For discussing with fellow users, please use the [vLLM Forum](https://discuss.vllm.ai)
+- coordinating contributions and development, please use [Slack](https://slack.vllm.ai)
+- For security disclosures, please use GitHub's [Security Advisories](https://github.com/vllm-project/vllm/security/advisories) feature
+- For collaborations and partnerships, please contact us at [vllm-questions@lists.berkeley.edu](mailto:vllm-questions@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).

RELEASE.md

@@ -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.

benchmarks/README.md

@@ -1,19 +1,268 @@
 # Benchmarking vLLM
 
-## Downloading the ShareGPT dataset
+This README guides you through running benchmark tests with the extensive
+datasets supported on vLLM. It’s a living document, updated as new features and datasets
+become available.
+
+## Dataset Overview
+
+<table style="width:100%; border-collapse: collapse;">
+  <thead>
+    <tr>
+      <th style="width:15%; text-align: left;">Dataset</th>
+      <th style="width:10%; text-align: center;">Online</th>
+      <th style="width:10%; text-align: center;">Offline</th>
+      <th style="width:65%; text-align: left;">Data Path</th>
+    </tr>
+  </thead>
+  <tbody>
+    <tr>
+      <td><strong>ShareGPT</strong></td>
+      <td style="text-align: center;">✅</td>
+      <td style="text-align: center;">✅</td>
+      <td><code>wget https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/resolve/main/ShareGPT_V3_unfiltered_cleaned_split.json</code></td>
+    </tr>
+    <tr>
+      <td><strong>BurstGPT</strong></td>
+      <td style="text-align: center;">✅</td>
+      <td style="text-align: center;">✅</td>
+      <td><code>wget https://github.com/HPMLL/BurstGPT/releases/download/v1.1/BurstGPT_without_fails_2.csv</code></td>
+    </tr>
+    <tr>
+      <td><strong>Sonnet</strong></td>
+      <td style="text-align: center;">✅</td>
+      <td style="text-align: center;">✅</td>
+      <td>Local file: <code>benchmarks/sonnet.txt</code></td>
+    </tr>
+    <tr>
+      <td><strong>Random</strong></td>
+      <td style="text-align: center;">✅</td>
+      <td style="text-align: center;">✅</td>
+      <td><code>synthetic</code></td>
+    </tr>
+    <tr>
+      <td><strong>HuggingFace</strong></td>
+      <td style="text-align: center;">🟡</td>
+      <td style="text-align: center;">🟡</td>
+      <td>Specify your dataset path on HuggingFace</td>
+    </tr>
+    <tr>
+      <td><strong>VisionArena</strong></td>
+      <td style="text-align: center;">✅</td>
+      <td style="text-align: center;">✅</td>
+      <td><code>lmarena-ai/vision-arena-bench-v0.1</code> (a HuggingFace dataset)</td>
+    </tr>
+  </tbody>
+</table>
+
+✅: supported
+
+🚧: to be supported
+
+🟡: Partial support. Currently, HuggingFaceDataset only supports dataset formats
+similar to `lmms-lab/LLaVA-OneVision-Data` and `Aeala/ShareGPT_Vicuna_unfiltered`.
+If you need support for other dataset formats, please consider contributing.
+
+**Note**: VisionArena’s `dataset-name` should be set to `hf`
+
+---
+## Example - Online Benchmark
+
+First start serving your model
 
-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
+MODEL_NAME="NousResearch/Hermes-3-Llama-3.1-8B"
+vllm serve ${MODEL_NAME} --disable-log-requests
 ```
 
-## Downloading the ShareGPT4V dataset
+Then run the benchmarking script
 
-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/
+# download dataset
+# wget https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/resolve/main/ShareGPT_V3_unfiltered_cleaned_split.json
+MODEL_NAME="NousResearch/Hermes-3-Llama-3.1-8B"
+NUM_PROMPTS=10
+BACKEND="vllm"
+DATASET_NAME="sharegpt"
+DATASET_PATH="<your data path>/ShareGPT_V3_unfiltered_cleaned_split.json"
+python3 vllm/benchmarks/benchmark_serving.py --backend ${BACKEND} --model ${MODEL_NAME} --endpoint /v1/completions --dataset-name ${DATASET_NAME} --dataset-path ${DATASET_PATH} --num-prompts ${NUM_PROMPTS}
 ```
+
+If successful, you will see the following output
+
+```
+============ Serving Benchmark Result ============
+Successful requests:                     10        
+Benchmark duration (s):                  5.78      
+Total input tokens:                      1369      
+Total generated tokens:                  2212      
+Request throughput (req/s):              1.73      
+Output token throughput (tok/s):         382.89    
+Total Token throughput (tok/s):          619.85    
+---------------Time to First Token----------------
+Mean TTFT (ms):                          71.54     
+Median TTFT (ms):                        73.88     
+P99 TTFT (ms):                           79.49     
+-----Time per Output Token (excl. 1st token)------
+Mean TPOT (ms):                          7.91      
+Median TPOT (ms):                        7.96      
+P99 TPOT (ms):                           8.03      
+---------------Inter-token Latency----------------
+Mean ITL (ms):                           7.74      
+Median ITL (ms):                         7.70      
+P99 ITL (ms):                            8.39      
+==================================================
+```
+
+### VisionArena Benchmark for Vision Language Models
+
+```bash
+# need a model with vision capability here
+vllm serve Qwen/Qwen2-VL-7B-Instruct --disable-log-requests
+```
+
+```bash
+MODEL_NAME="Qwen/Qwen2-VL-7B-Instruct"
+NUM_PROMPTS=10
+BACKEND="openai-chat"
+DATASET_NAME="hf"
+DATASET_PATH="lmarena-ai/vision-arena-bench-v0.1"
+DATASET_SPLIT='train'
+
+python3 vllm/benchmarks/benchmark_serving.py \
+  --backend "${BACKEND}" \
+  --model "${MODEL_NAME}" \
+  --endpoint "/v1/chat/completions" \
+  --dataset-name "${DATASET_NAME}" \
+  --dataset-path "${DATASET_PATH}" \
+  --hf-split "${DATASET_SPLIT}" \
+  --num-prompts "${NUM_PROMPTS}"
+```
+
+### HuggingFaceDataset Examples
+
+Currently, HuggingFaceDataset only supports dataset formats
+similar to `lmms-lab/LLaVA-OneVision-Data` and `Aeala/ShareGPT_Vicuna_unfiltered`. If you need support for other dataset
+formats, please consider contributing.
+
+```bash
+# need a model with vision capability here
+vllm serve Qwen/Qwen2-VL-7B-Instruct --disable-log-requests
+```
+
+**`lmms-lab/LLaVA-OneVision-Data`**
+
+```bash
+MODEL_NAME="Qwen/Qwen2-VL-7B-Instruct"
+NUM_PROMPTS=10
+BACKEND="openai-chat"
+DATASET_NAME="hf"
+DATASET_PATH="lmms-lab/LLaVA-OneVision-Data"
+DATASET_SPLIT='train'
+DATASET_SUBSET='chart2text(cauldron)'
+python3 vllm/benchmarks/benchmark_serving.py \
+  --backend "${BACKEND}" \
+  --model "${MODEL_NAME}" \
+  --endpoint "/v1/chat/completions" \
+  --dataset-name "${DATASET_NAME}" \
+  --dataset-path "${DATASET_PATH}" \
+  --hf-split "${DATASET_SPLIT}" \
+  --num-prompts "${NUM_PROMPTS}" \
+  --hf-subset "${DATASET_SUBSET}"
+```
+
+**`Aeala/ShareGPT_Vicuna_unfiltered`**
+
+```bash
+MODEL_NAME="Qwen/Qwen2-VL-7B-Instruct"
+NUM_PROMPTS=10
+BACKEND="openai-chat"
+DATASET_NAME="hf"
+DATASET_PATH="Aeala/ShareGPT_Vicuna_unfiltered"
+DATASET_SPLIT='train'
+python3 vllm/benchmarks/benchmark_serving.py \
+  --backend "${BACKEND}" \
+  --model "${MODEL_NAME}" \
+  --endpoint "/v1/chat/completions" \
+  --dataset-name "${DATASET_NAME}" \
+  --dataset-path "${DATASET_PATH}" \
+  --hf-split "${DATASET_SPLIT}" \
+  --num-prompts "${NUM_PROMPTS}" \
+```
+
+---
+## Example - Offline Throughput Benchmark
+
+```bash
+MODEL_NAME="NousResearch/Hermes-3-Llama-3.1-8B"
+NUM_PROMPTS=10
+DATASET_NAME="sonnet"
+DATASET_PATH="vllm/benchmarks/sonnet.txt"
+
+python3 vllm/benchmarks/benchmark_throughput.py \
+  --model "${MODEL_NAME}" \
+  --dataset-name "${DATASET_NAME}" \
+  --dataset-path "${DATASET_PATH}" \
+  --num-prompts "${NUM_PROMPTS}"
+```
+
+If successful, you will see the following output
+
+```
+Throughput: 7.15 requests/s, 4656.00 total tokens/s, 1072.15 output tokens/s
+Total num prompt tokens:  5014
+Total num output tokens:  1500
+```
+
+### VisionArena Benchmark for Vision Language Models
+
+``` bash
+MODEL_NAME="Qwen/Qwen2-VL-7B-Instruct"
+NUM_PROMPTS=10
+DATASET_NAME="hf"
+DATASET_PATH="lmarena-ai/vision-arena-bench-v0.1"
+DATASET_SPLIT="train"
+
+python3 vllm/benchmarks/benchmark_throughput.py \
+  --model "${MODEL_NAME}" \
+  --backend "vllm-chat" \
+  --dataset-name "${DATASET_NAME}" \
+  --dataset-path "${DATASET_PATH}" \
+  --num-prompts "${NUM_PROMPTS}" \
+  --hf-split "${DATASET_SPLIT}"
+```
+
+The `num prompt tokens` now includes image token counts
+
+```
+Throughput: 2.55 requests/s, 4036.92 total tokens/s, 326.90 output tokens/s
+Total num prompt tokens:  14527
+Total num output tokens:  1280
+```
+
+### Benchmark with LoRA Adapters
+
+``` bash
+# download dataset
+# wget https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/resolve/main/ShareGPT_V3_unfiltered_cleaned_split.json
+MODEL_NAME="meta-llama/Llama-2-7b-hf"
+BACKEND="vllm"
+DATASET_NAME="sharegpt"
+DATASET_PATH="<your data path>/ShareGPT_V3_unfiltered_cleaned_split.json"
+NUM_PROMPTS=10
+MAX_LORAS=2
+MAX_LORA_RANK=8
+ENABLE_LORA="--enable-lora"
+LORA_PATH="yard1/llama-2-7b-sql-lora-test"
+
+python3 vllm/benchmarks/benchmark_throughput.py \
+  --model "${MODEL_NAME}" \
+  --backend "${BACKEND}" \
+  --dataset_path "${DATASET_PATH}" \
+  --dataset_name "${DATASET_NAME}" \
+  --num-prompts "${NUM_PROMPTS}" \
+  --max-loras "${MAX_LORAS}" \
+  --max-lora-rank "${MAX_LORA_RANK}" \
+  ${ENABLE_LORA} \
+  --lora-path "${LORA_PATH}"
+  ```

benchmarks/backend_request_func.py

@@ -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 +14,9 @@ from tqdm.asyncio import tqdm
 from transformers import (AutoTokenizer, PreTrainedTokenizer,
                           PreTrainedTokenizerFast)
 
+# NOTE(simon): do not import vLLM here so the benchmark script
+# can run without vLLM installed.
+
 AIOHTTP_TIMEOUT = aiohttp.ClientTimeout(total=6 * 60 * 60)
 
 
@@ -23,7 +28,6 @@ class RequestFuncInput:
     output_len: int
     model: str
     model_name: Optional[str] = None
-    best_of: int = 1
     logprobs: Optional[int] = None
     extra_body: Optional[dict] = None
     multi_modal_content: Optional[dict] = None
@@ -37,8 +41,8 @@ class RequestFuncOutput:
     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 = ""
@@ -54,13 +58,12 @@ async def async_request_tgi(
     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,
             "do_sample": True,
             "temperature": 0.01,  # TGI does not accept 0.0 temperature.
             "top_p": 0.99,  # TGI does not accept 1.0 top_p.
             "truncate": request_func_input.prompt_len,
-            # TGI does not accept ignore_eos flag.
+            "ignore_eos_token": request_func_input.ignore_eos,
         }
         payload = {
             "inputs": request_func_input.prompt,
@@ -68,6 +71,10 @@ async def async_request_tgi(
         }
         output = RequestFuncOutput()
         output.prompt_len = request_func_input.prompt_len
+        if request_func_input.ignore_eos:
+            output.output_tokens = request_func_input.output_len
+        else:
+            output.output_tokens = None
 
         ttft = 0.0
         st = time.perf_counter()
@@ -126,7 +133,6 @@ async def async_request_trt_llm(
 
     async with aiohttp.ClientSession(trust_env=True,
                                      timeout=AIOHTTP_TIMEOUT) as session:
-        assert request_func_input.best_of == 1
         payload = {
             "accumulate_tokens": True,
             "text_input": request_func_input.prompt,
@@ -191,7 +197,6 @@ async def async_request_deepspeed_mii(
 ) -> RequestFuncOutput:
     async with aiohttp.ClientSession(trust_env=True,
                                      timeout=AIOHTTP_TIMEOUT) as session:
-        assert request_func_input.best_of == 1
 
         payload = {
             "prompt": request_func_input.prompt,
@@ -245,7 +250,6 @@ async def async_request_openai_completions(
                 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,
@@ -334,7 +338,7 @@ async def async_request_openai_chat_completions(
 ) -> RequestFuncOutput:
     api_url = request_func_input.api_url
     assert api_url.endswith(
-        "chat/completions"
+        ("chat/completions", "profile")
     ), "OpenAI Chat Completions API URL must end with 'chat/completions'."
 
     async with aiohttp.ClientSession(trust_env=True,
@@ -428,12 +432,17 @@ 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"])
+        from vllm.model_executor.model_loader.weight_utils import get_lock
 
-        return model_path
+        # 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 pretrained_model_name_or_path
 
 

benchmarks/benchmark_dataset.py

@@ -0,0 +1,717 @@
+# SPDX-License-Identifier: Apache-2.0
+"""
+This module defines a framework for sampling benchmark requests from various
+datasets. Each dataset subclass of BenchmarkDataset must implement sample
+generation. Supported dataset types include:
+  - ShareGPT
+  - Random (synthetic)
+  - Sonnet
+  - BurstGPT
+  - HuggingFace
+  - VisionArena
+
+TODO: Implement CustomDataset to parse a JSON file and convert its contents into
+SampleRequest instances, similar to the approach used in ShareGPT.
+"""
+
+import base64
+import io
+import json
+import logging
+import random
+from abc import ABC, abstractmethod
+from collections.abc import Mapping
+from dataclasses import dataclass
+from functools import cache
+from typing import Any, Optional, Union
+
+import numpy as np
+import pandas as pd
+from datasets import load_dataset
+from PIL import Image
+from transformers import PreTrainedTokenizerBase
+
+from vllm.lora.request import LoRARequest
+from vllm.lora.utils import get_adapter_absolute_path
+from vllm.multimodal import MultiModalDataDict
+from vllm.transformers_utils.tokenizer import AnyTokenizer, get_lora_tokenizer
+
+logger = logging.getLogger(__name__)
+
+# -----------------------------------------------------------------------------
+# Data Classes
+# -----------------------------------------------------------------------------
+
+
+@dataclass
+class SampleRequest:
+    """
+    Represents a single inference request for benchmarking.
+    """
+
+    prompt: Union[str, Any]
+    prompt_len: int
+    expected_output_len: int
+    multi_modal_data: Optional[Union[MultiModalDataDict, dict]] = None
+    lora_request: Optional[LoRARequest] = None
+
+
+# -----------------------------------------------------------------------------
+# Benchmark Dataset Base Class
+# -----------------------------------------------------------------------------
+
+
+class BenchmarkDataset(ABC):
+    DEFAULT_SEED = 0
+
+    def __init__(
+        self,
+        dataset_path: Optional[str] = None,
+        random_seed: int = DEFAULT_SEED,
+    ) -> None:
+        """
+        Initialize the BenchmarkDataset with an optional dataset path and random
+        seed.  Args:
+            dataset_path (Optional[str]): Path to the dataset. If None, it
+            indicates that a default or random dataset might be used.
+            random_seed (int): Seed value for reproducible shuffling or
+            sampling. Defaults to DEFAULT_SEED.
+        """
+        self.dataset_path = dataset_path
+        # Set the random seed, ensuring that a None value is replaced with the
+        # default seed.
+        self.random_seed = (random_seed
+                            if random_seed is not None else self.DEFAULT_SEED)
+        self.data = None
+
+    def apply_multimodal_chat_transformation(
+            self,
+            prompt: str,
+            mm_content: Optional[MultiModalDataDict] = None) -> list[dict]:
+        """
+        Transform a prompt and optional multimodal content into a chat format.
+        This method is used for chat models that expect a specific conversation
+        format.
+        """
+        content = [{"text": prompt, "type": "text"}]
+        if mm_content is not None:
+            content.append(mm_content)
+        return [{"role": "user", "content": content}]
+
+    def load_data(self) -> None:
+        """
+        Load data from the dataset path into self.data.
+
+        This method must be overridden by subclasses since the method to load
+        data will vary depending on the dataset format and source.
+
+        Raises:
+            NotImplementedError: If a subclass does not implement this method.
+        """
+        # TODO (jenniferzhao): add support for downloading data
+        raise NotImplementedError(
+            "load_data must be implemented in subclasses.")
+
+    def get_random_lora_request(
+        self,
+        tokenizer: PreTrainedTokenizerBase,
+        max_loras: Optional[int] = None,
+        lora_path: Optional[str] = None,
+    ) -> tuple[Optional[LoRARequest], AnyTokenizer]:
+        """
+        Optionally select a random LoRA request and return its associated
+        tokenizer.
+
+        This method is used when LoRA parameters are provided.  It randomly
+        selects a LoRA based on max_loras and retrieves a cached tokenizer for
+        that LoRA if available. Otherwise, it returns the base tokenizer.
+
+        Args:
+            tokenizer (PreTrainedTokenizerBase): The base tokenizer to use if no
+            LoRA is selected.  max_loras (Optional[int]): The maximum number of
+            LoRAs available. If None, LoRA is not used.  lora_path
+            (Optional[str]): Path to the LoRA parameters on disk. If None, LoRA
+            is not used.
+
+        Returns:
+            tuple[Optional[LoRARequest], AnyTokenizer]: A tuple where the first
+            element is a LoRARequest (or None if not applicable) and the second
+            element is the tokenizer associated with the LoRA request (or the
+            base tokenizer).
+        """
+        if max_loras is None or lora_path is None:
+            return None, tokenizer
+
+        # Generate a random LoRA ID in the range [1, max_loras].
+        lora_id = random.randint(1, max_loras)
+        lora_request = LoRARequest(
+            lora_name=str(lora_id),
+            lora_int_id=lora_id,
+            lora_path=lora_path_on_disk(lora_path),
+        )
+        if lora_id not in lora_tokenizer_cache:
+            lora_tokenizer_cache[lora_id] = get_lora_tokenizer(lora_request)
+        # Return lora_request and the cached tokenizer if available; otherwise,
+        # return the base tokenizer
+        return lora_request, lora_tokenizer_cache[lora_id] or tokenizer
+
+    @abstractmethod
+    def sample(self, tokenizer: PreTrainedTokenizerBase,
+               num_requests: int) -> list[SampleRequest]:
+        """
+        Abstract method to generate sample requests from the dataset.
+
+        Subclasses must override this method to implement dataset-specific logic
+        for generating a list of SampleRequest objects.
+
+        Args:
+            tokenizer (PreTrainedTokenizerBase): The tokenizer to be used
+             for processing the dataset's text.
+            num_requests (int): The number of sample requests to generate.
+
+        Returns:
+            list[SampleRequest]: A list of sample requests generated from the
+            dataset.
+        """
+        raise NotImplementedError("sample must be implemented in subclasses.")
+
+    def maybe_oversample_requests(self, requests: list[SampleRequest],
+                                  num_requests: int) -> None:
+        """
+        Oversamples the list of requests if its size is less than the desired
+        number.
+
+        Args:
+            requests (List[SampleRequest]): The current list of sampled
+            requests.  num_requests (int): The target number of requests.
+        """
+        if len(requests) < num_requests:
+            random.seed(self.random_seed)
+            additional = random.choices(requests,
+                                        k=num_requests - len(requests))
+            requests.extend(additional)
+            logger.info("Oversampled requests to reach %d total samples.",
+                        num_requests)
+
+
+# -----------------------------------------------------------------------------
+# Utility Functions and Global Caches
+# -----------------------------------------------------------------------------
+
+
+def is_valid_sequence(
+    prompt_len: int,
+    output_len: int,
+    min_len: int = 4,
+    max_prompt_len: int = 1024,
+    max_total_len: int = 2048,
+    skip_min_output_len_check: bool = False,
+) -> bool:
+    """
+    Validate a sequence based on prompt and output lengths.
+
+    Default pruning criteria are copied from the original `sample_hf_requests`
+    and `sample_sharegpt_requests` functions in benchmark_serving.py, as well as
+    from `sample_requests` in benchmark_throughput.py.
+    """
+    # Check for invalid conditions
+    prompt_too_short = prompt_len < min_len
+    output_too_short = (not skip_min_output_len_check) and (output_len
+                                                            < min_len)
+    prompt_too_long = prompt_len > max_prompt_len
+    combined_too_long = (prompt_len + output_len) > max_total_len
+
+    # Return True if none of the invalid conditions are met
+    return not (prompt_too_short or output_too_short or prompt_too_long
+                or combined_too_long)
+
+
+@cache
+def lora_path_on_disk(lora_path: str) -> str:
+    return get_adapter_absolute_path(lora_path)
+
+
+# Global cache for LoRA tokenizers.
+lora_tokenizer_cache: dict[int, AnyTokenizer] = {}
+
+
+def process_image(image: Any) -> Mapping[str, Any]:
+    """
+    Process a single image input and return a multimedia content dictionary.
+
+    For a PIL.Image.Image input:
+      - Converts the image to RGB.
+      - Saves the image as a JPEG in-memory.
+      - Encodes the JPEG data as a base64 string.
+      - Returns a dictionary with the image as a base64 data URL.
+
+    For a string input:
+      - Treats the string as a URL or file path.
+      - Prepends "file://" if the string doesn't start with "http://" or
+        "file://".
+      - Returns a dictionary with the image URL.
+
+    Raises:
+      ValueError: If the input is neither a PIL.Image.Image nor a string.
+    """
+    if isinstance(image, Image.Image):
+        image = image.convert("RGB")
+        with io.BytesIO() as image_data:
+            image.save(image_data, format="JPEG")
+            image_base64 = base64.b64encode(
+                image_data.getvalue()).decode("utf-8")
+        return {
+            "type": "image_url",
+            "image_url": {
+                "url": f"data:image/jpeg;base64,{image_base64}"
+            },
+        }
+
+    if isinstance(image, str):
+        image_url = (image if image.startswith(
+            ("http://", "file://")) else f"file://{image}")
+        return {"type": "image_url", "image_url": {"url": image_url}}
+
+    raise ValueError(
+        f"Invalid image input {image}. Must be a PIL.Image.Image or str.")
+
+
+# -----------------------------------------------------------------------------
+# Random Dataset Implementation (Synthetic Data)
+# -----------------------------------------------------------------------------
+
+
+class RandomDataset(BenchmarkDataset):
+    # Default values copied from benchmark_serving.py for the random dataset.
+    DEFAULT_PREFIX_LEN = 0
+    DEFAULT_RANGE_RATIO = 1.0
+    DEFAULT_INPUT_LEN = 1024
+    DEFAULT_OUTPUT_LEN = 128
+
+    def __init__(
+        self,
+        **kwargs,
+    ) -> None:
+        super().__init__(**kwargs)
+
+    def sample(
+        self,
+        tokenizer: PreTrainedTokenizerBase,
+        num_requests: int,
+        prefix_len: int = DEFAULT_PREFIX_LEN,
+        range_ratio: float = DEFAULT_RANGE_RATIO,
+        input_len: int = DEFAULT_INPUT_LEN,
+        output_len: int = DEFAULT_OUTPUT_LEN,
+        **kwargs,
+    ) -> list[SampleRequest]:
+        vocab_size = tokenizer.vocab_size
+
+        prefix_token_ids = (np.random.randint(
+            0, vocab_size, size=prefix_len).tolist() if prefix_len > 0 else [])
+
+        input_low = int(input_len * range_ratio)
+        output_low = int(output_len * range_ratio)
+
+        input_lens = np.random.randint(input_low,
+                                       input_len + 1,
+                                       size=num_requests)
+        output_lens = np.random.randint(output_low,
+                                        output_len + 1,
+                                        size=num_requests)
+        offsets = np.random.randint(0, vocab_size, size=num_requests)
+
+        requests = []
+        for i in range(num_requests):
+            inner_seq = ((offsets[i] + i + np.arange(input_lens[i])) %
+                         vocab_size).tolist()
+            token_sequence = prefix_token_ids + inner_seq
+            prompt = tokenizer.decode(token_sequence)
+            total_input_len = prefix_len + int(input_lens[i])
+            requests.append(
+                SampleRequest(
+                    prompt=prompt,
+                    prompt_len=total_input_len,
+                    expected_output_len=int(output_lens[i]),
+                ))
+        return requests
+
+
+# -----------------------------------------------------------------------------
+# ShareGPT Dataset Implementation
+# -----------------------------------------------------------------------------
+
+
+class ShareGPTDataset(BenchmarkDataset):
+    """
+    Implements the ShareGPT dataset.  Loads data from a JSON file and generates
+    sample requests based on conversation turns.
+    """
+
+    def __init__(self, **kwargs) -> None:
+        super().__init__(**kwargs)
+        self.load_data()
+
+    def load_data(self) -> None:
+        if self.dataset_path is None:
+            raise ValueError("dataset_path must be provided for loading data.")
+
+        with open(self.dataset_path, encoding="utf-8") as f:
+            self.data = json.load(f)
+        # Filter entries with at least two conversation turns.
+        self.data = [
+            entry for entry in self.data
+            if "conversations" in entry and len(entry["conversations"]) >= 2
+        ]
+        random.seed(self.random_seed)
+        random.shuffle(self.data)
+
+    def sample(
+        self,
+        tokenizer: PreTrainedTokenizerBase,
+        num_requests: int,
+        lora_path: Optional[str] = None,
+        max_loras: Optional[int] = None,
+        output_len: Optional[int] = None,
+        enable_multimodal_chat: bool = False,
+        **kwargs,
+    ) -> list:
+        samples: list = []
+        for entry in self.data:
+            if len(samples) >= num_requests:
+                break
+            prompt, completion = (
+                entry["conversations"][0]["value"],
+                entry["conversations"][1]["value"],
+            )
+
+            lora_request, tokenizer = self.get_random_lora_request(
+                tokenizer=tokenizer, max_loras=max_loras, lora_path=lora_path)
+            prompt_ids = tokenizer(prompt).input_ids
+            completion_ids = tokenizer(completion).input_ids
+            prompt_len = len(prompt_ids)
+            new_output_len = (len(completion_ids)
+                              if output_len is None else output_len)
+            if not is_valid_sequence(prompt_len,
+                                     new_output_len,
+                                     skip_min_output_len_check=output_len
+                                     is not None):
+                continue
+            if enable_multimodal_chat:
+                prompt = self.apply_multimodal_chat_transformation(
+                    prompt, None)
+            samples.append(
+                SampleRequest(
+                    prompt=prompt,
+                    prompt_len=prompt_len,
+                    expected_output_len=new_output_len,
+                    lora_request=lora_request,
+                ))
+        self.maybe_oversample_requests(samples, num_requests)
+        return samples
+
+
+# -----------------------------------------------------------------------------
+# Sonnet Dataset Implementation
+# -----------------------------------------------------------------------------
+
+
+class SonnetDataset(BenchmarkDataset):
+    """
+    Simplified implementation of the Sonnet dataset.  Loads poem lines from a
+    text file and generates sample requests.  Default values here copied from
+    `benchmark_serving.py` for the sonnet dataset.
+    """
+
+    DEFAULT_PREFIX_LEN = 200
+    DEFAULT_INPUT_LEN = 550
+    DEFAULT_OUTPUT_LEN = 150
+
+    def __init__(
+        self,
+        **kwargs,
+    ) -> None:
+        super().__init__(**kwargs)
+        self.load_data()
+
+    def load_data(self) -> None:
+        if not self.dataset_path:
+            raise ValueError("dataset_path must be provided.")
+        with open(self.dataset_path, encoding="utf-8") as f:
+            self.data = f.readlines()
+
+    def sample(
+        self,
+        tokenizer,
+        num_requests: int,
+        prefix_len: int = DEFAULT_PREFIX_LEN,
+        input_len: int = DEFAULT_INPUT_LEN,
+        output_len: int = DEFAULT_OUTPUT_LEN,
+        return_prompt_formatted: bool = False,
+        **kwargs,
+    ) -> list:
+        # Calculate average token length for a poem line.
+        tokenized_lines = [tokenizer(line).input_ids for line in self.data]
+        avg_len = sum(len(tokens)
+                      for tokens in tokenized_lines) / len(tokenized_lines)
+
+        # Build the base prompt.
+        base_prompt = "Pick as many lines as you can from these poem lines:\n"
+        base_msg = [{"role": "user", "content": base_prompt}]
+        base_fmt = tokenizer.apply_chat_template(base_msg,
+                                                 add_generation_prompt=True,
+                                                 tokenize=False)
+        base_offset = len(tokenizer(base_fmt).input_ids)
+        if input_len <= base_offset:
+            raise ValueError(
+                f"'input_len' must be higher than the base prompt length "
+                f"({base_offset}).")
+
+        # Determine how many poem lines to use.
+        num_input_lines = round((input_len - base_offset) / avg_len)
+        num_prefix_lines = round((prefix_len - base_offset) / avg_len)
+        prefix_lines = self.data[:num_prefix_lines]
+
+        samples = []
+        for _ in range(num_requests):
+            extra_lines = random.choices(self.data,
+                                         k=num_input_lines - num_prefix_lines)
+            prompt = f"{base_prompt}{''.join(prefix_lines + extra_lines)}"
+            msg = [{"role": "user", "content": prompt}]
+            prompt_formatted = tokenizer.apply_chat_template(
+                msg, add_generation_prompt=True, tokenize=False)
+            prompt_len = len(tokenizer(prompt_formatted).input_ids)
+            samples.append(
+                SampleRequest(
+                    prompt=prompt_formatted
+                    if return_prompt_formatted else prompt,
+                    prompt_len=prompt_len,
+                    expected_output_len=output_len,
+                ))
+        return samples
+
+
+# -----------------------------------------------------------------------------
+# BurstGPT Dataset Implementation
+# -----------------------------------------------------------------------------
+
+
+class BurstGPTDataset(BenchmarkDataset):
+    """
+    Implements the BurstGPT dataset.  Loads data from a CSV file and generates
+    sample requests based on synthetic prompt generation. Only rows with Model
+    "GPT-4" and positive response tokens are used.
+    """
+
+    def __init__(self, **kwargs) -> None:
+        super().__init__(**kwargs)
+        self.load_data()
+
+    def load_data(self, ):
+        if self.dataset_path is None:
+            raise ValueError("dataset_path must be provided for loading data.")
+
+        df = pd.read_csv(self.dataset_path)
+        # Filter to keep only GPT-4 rows.
+        gpt4_df = df[df["Model"] == "GPT-4"]
+        # Remove failed requests (where Response tokens is 0 or less).
+        gpt4_df = gpt4_df[gpt4_df["Response tokens"] > 0]
+        # Sample the desired number of rows.
+        self.data = gpt4_df
+
+    def _sample_loaded_data(self, num_requests: int) -> list:
+        if num_requests <= len(self.data):
+            data = self.data.sample(n=num_requests,
+                                    random_state=self.random_seed)
+        else:
+            data = self.data.sample(
+                n=num_requests,
+                random_state=self.random_seed,
+                replace=True,
+            )
+        # Convert the dataframe to a list of lists.
+        return data.values.tolist()
+
+    def sample(
+        self,
+        tokenizer: PreTrainedTokenizerBase,
+        num_requests: int,
+        max_loras: Optional[int] = None,
+        lora_path: Optional[str] = None,
+        **kwargs,
+    ) -> list[SampleRequest]:
+        samples = []
+        data = self._sample_loaded_data(num_requests=num_requests)
+        for i in range(num_requests):
+            input_len = int(data[i][2])
+            output_len = int(data[i][3])
+            lora_req, tokenizer = self.get_random_lora_request(
+                tokenizer=tokenizer, max_loras=max_loras, lora_path=lora_path)
+            vocab_size = tokenizer.vocab_size
+            # Generate a synthetic prompt: a list of token IDs computed as (i +
+            # j) modulo vocab_size.
+            token_ids = [(i + j) % vocab_size for j in range(input_len)]
+            prompt = tokenizer.decode(token_ids)
+            samples.append(
+                SampleRequest(
+                    prompt=prompt,
+                    prompt_len=input_len,
+                    expected_output_len=output_len,
+                    lora_request=lora_req,
+                ))
+        return samples
+
+
+# -----------------------------------------------------------------------------
+# HuggingFace Dataset Implementation
+# -----------------------------------------------------------------------------
+
+
+class HuggingFaceDataset(BenchmarkDataset):
+    """
+    Dataset class for processing a HuggingFace dataset with conversation data
+    and optional images.
+    """
+
+    def __init__(
+        self,
+        dataset_split: str,
+        dataset_subset: Optional[str] = None,
+        **kwargs,
+    ) -> None:
+        super().__init__(**kwargs)
+        self.dataset_split = dataset_split
+        self.dataset_subset = dataset_subset
+
+        self.load_data()
+
+    def load_data(self) -> None:
+        if not self.dataset_path:
+            raise ValueError("dataset_path must be provided for loading data.")
+
+        self.data = load_dataset(
+            self.dataset_path,
+            name=self.dataset_subset,
+            split=self.dataset_split,
+            streaming=True,
+        )
+        if self.data.features is None or "conversations" \
+            not in self.data.features:
+            raise ValueError(
+                "HuggingFaceDataset currently only supports datasets with "
+                "a 'conversations' column like lmms-lab/LLaVA-OneVision-Data. "
+                "Please consider contributing if you would like to add "
+                "support for additional dataset formats.")
+        # Shuffle and filter examples with at least 2 conversations.
+        self.data = self.data.shuffle(seed=self.random_seed).filter(
+            lambda x: len(x["conversations"]) >= 2)
+
+    def sample(self,
+               tokenizer: PreTrainedTokenizerBase,
+               num_requests: int,
+               output_len: Optional[int] = None,
+               enable_multimodal_chat: bool = False,
+               **kwargs) -> list:
+        sampled_requests = []
+        dynamic_output = output_len is None
+
+        for item in self.data:
+            if len(sampled_requests) >= num_requests:
+                break
+            conv = item["conversations"]
+            prompt, completion = conv[0]["value"], conv[1]["value"]
+
+            prompt_ids = tokenizer(prompt).input_ids
+            completion_ids = tokenizer(completion).input_ids
+            prompt_len = len(prompt_ids)
+            completion_len = len(completion_ids)
+            output_len = completion_len if dynamic_output else output_len
+            assert isinstance(output_len, int) and output_len > 0
+            if dynamic_output and not is_valid_sequence(
+                    prompt_len, completion_len):
+                continue
+            mm_content = process_image(
+                item["image"]) if "image" in item else None
+            if enable_multimodal_chat:
+                # Note: when chat is enabled the request prompt_len is no longer
+                # accurate and we will be using request output to count the
+                # actual prompt len and output len
+                prompt = self.apply_multimodal_chat_transformation(
+                    prompt, mm_content)
+            sampled_requests.append(
+                SampleRequest(
+                    prompt=prompt,
+                    prompt_len=prompt_len,
+                    expected_output_len=output_len,
+                    multi_modal_data=mm_content,
+                ))
+        self.maybe_oversample_requests(sampled_requests, num_requests)
+        return sampled_requests
+
+
+# -----------------------------------------------------------------------------
+# Vision Arena Dataset Implementation
+# -----------------------------------------------------------------------------
+
+
+class VisionArenaDataset(HuggingFaceDataset):
+    """
+    Vision Arena Dataset.
+    """
+
+    DEFAULT_OUTPUT_LEN = 128
+    VISION_ARENA_DATASET_PATH = "lmarena-ai/vision-arena-bench-v0.1"
+
+    def __init__(
+        self,
+        **kwargs,
+    ) -> None:
+        super().__init__(**kwargs)
+        if self.dataset_path != self.VISION_ARENA_DATASET_PATH:
+            raise ValueError(f"Only support Vision Arena dataset.\
+                    This data path {self.dataset_path} is not valid.")
+        if self.dataset_subset is None and self.dataset_split != "train":
+            raise ValueError("Dataset split must be 'train'.")
+
+        self.load_data()
+
+    def load_data(self) -> None:
+        dataset = load_dataset(
+            self.dataset_path,
+            name=self.dataset_subset,
+            split=self.dataset_split,
+            streaming=True,
+        )
+        self.data = dataset.shuffle(seed=self.random_seed)
+
+    def sample(
+        self,
+        tokenizer: PreTrainedTokenizerBase,
+        num_requests: int,
+        output_len: Optional[int] = None,
+        enable_multimodal_chat: bool = False,
+        **kwargs,
+    ) -> list:
+        output_len = (output_len
+                      if output_len is not None else self.DEFAULT_OUTPUT_LEN)
+        sampled_requests = []
+        for item in self.data:
+            if len(sampled_requests) >= num_requests:
+                break
+            prompt = item["turns"][0][0]["content"]
+            mm_content = process_image(item["images"][0])
+            prompt_len = len(tokenizer(prompt).input_ids)
+            if enable_multimodal_chat:
+                # Note: when chat is enabled the request prompt_len is no longer
+                # accurate and we will be using request output to count the
+                # actual prompt len
+                prompt = self.apply_multimodal_chat_transformation(
+                    prompt, mm_content)
+            sampled_requests.append(
+                SampleRequest(
+                    prompt=prompt,
+                    prompt_len=prompt_len,
+                    expected_output_len=output_len,
+                    multi_modal_data=mm_content,
+                ))
+        self.maybe_oversample_requests(sampled_requests, num_requests)
+        return sampled_requests

benchmarks/benchmark_guided.py

@@ -1,494 +0,0 @@
-"""Benchmark guided decoding throughput."""
-import argparse
-import dataclasses
-import json
-import os
-import random
-import time
-from typing import List
-
-import datasets
-import pandas as pd
-import uvloop
-from transformers import AutoTokenizer, PreTrainedTokenizerBase
-
-from vllm.engine.arg_utils import AsyncEngineArgs, EngineArgs
-from vllm.entrypoints.openai.api_server import (
-    build_async_engine_client_from_engine_args)
-from vllm.sampling_params import GuidedDecodingParams
-from vllm.utils import FlexibleArgumentParser, merge_async_iterators
-
-
-@dataclasses.dataclass
-class SampleRequest:
-    """A class representing a single inference request for benchmarking.
-
-    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.
-    """
-    prompt: str
-    prompt_len: int
-    expected_output_len: int
-    schema: dict
-    structure_type: str = 'json'
-    completion: str = None
-
-
-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))
-
-    # Add the requests to the engine.
-    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))
-
-    if warmup:
-        print(">>>>> Running warmup prompt, for the first 5")
-        # We setup the first 5 requests to warmup FSM
-        # if using xgrammar dataset, we will skip warmup
-        warmup_requests = requests[:5]
-        for i, request in enumerate(warmup_requests):
-            prompts.append(request.prompt)
-            sampling_params.append(
-                SamplingParams(
-                    n=n,
-                    temperature=1.0,
-                    top_p=1.0,
-                    ignore_eos=True,
-                    max_tokens=request.expected_output_len,
-                    guided_decoding=GuidedDecodingParams(json=request.schema)
-                    if guided_decoding_rate > 0 else None,
-                ))
-        llm.generate(prompts, sampling_params, use_tqdm=False)
-
-    print(">>>>> Benchmark started...")
-    prompts = []
-    sampling_params = []
-    for i, request in enumerate(requests):
-        prompts.append(request.prompt)
-        sampling_params.append(
-            SamplingParams(
-                n=n,
-                temperature=1.0,
-                top_p=1.0,
-                ignore_eos=True,
-                max_tokens=request.expected_output_len,
-                guided_decoding=GuidedDecodingParams(
-                    **{request.structure_type: request.schema})
-                if i in guided_decoding_req_idx else None,
-            ))
-
-    start = time.perf_counter()
-    outputs = llm.generate(prompts, sampling_params, use_tqdm=False)
-    ret = []
-    for output, request in zip(outputs, requests):
-        generated_text = output.outputs[0].text
-        ret.append({
-            "generated": generated_text,
-            "expected": request.completion
-        })
-    end = time.perf_counter()
-    return end - start, ret
-
-
-async def run_vllm_async(
-        requests: List[SampleRequest],
-        engine_args: AsyncEngineArgs,
-        n: int,
-        guided_decoding_rate: float = 1.0,
-        warmup: bool = False,
-        disable_frontend_multiprocessing: bool = False) -> float:
-    from vllm import SamplingParams
-
-    async with build_async_engine_client_from_engine_args(
-            engine_args, disable_frontend_multiprocessing) as llm:
-
-        # Add the requests to the engine.
-        prompts: List[str] = []
-        sampling_params: List[SamplingParams] = []
-        guided_decoding_req_idx = random.sample(
-            range(len(requests)), int(len(requests) * guided_decoding_rate))
-
-        if warmup:
-            print(">>>>>> Running warmup prompt, for the first 5")
-            # We setup the first 5 requests to warmup FSM
-            # if using xgrammar dataset, we will skip warmup
-            warmup_requests = requests[:5]
-            for i, request in enumerate(warmup_requests):
-                prompts.append(request.prompt)
-                sampling_params.append(
-                    SamplingParams(
-                        n=n,
-                        temperature=1.0,
-                        top_p=1.0,
-                        ignore_eos=True,
-                        max_tokens=request.expected_output_len,
-                        guided_decoding=GuidedDecodingParams(
-                            json=request.schema)
-                        if guided_decoding_rate > 0 else None,
-                    ))
-            generators = []
-            for i, (prompt, sp) in enumerate(zip(prompts, sampling_params)):
-                generator = llm.generate(prompt, sp, request_id=f"test{i}")
-                generators.append(generator)
-            all_gens = merge_async_iterators(*generators)
-            async for i, res in all_gens:
-                pass
-
-        print(">>>>> Benchmark started...")
-        prompts = []
-        sampling_params = []
-        for i, request in enumerate(requests):
-            prompts.append(request.prompt)
-            sampling_params.append(
-                SamplingParams(
-                    n=n,
-                    temperature=1.0,
-                    top_p=1.0,
-                    ignore_eos=True,
-                    max_tokens=request.expected_output_len,
-                    guided_decoding=GuidedDecodingParams(json=request.schema)
-                    if i in guided_decoding_req_idx else None,
-                ))
-
-        generators = []
-        start_time = []
-        latencies = []
-        start = time.perf_counter()
-        for i, (prompt, sp) in enumerate(zip(prompts, sampling_params)):
-            generator = llm.generate(prompt, sp, request_id=f"test{i}")
-            generators.append(generator)
-            start_time.append(time.perf_counter())
-            latencies.append([])
-        all_gens = merge_async_iterators(*generators)
-        generated_texts = [''] * len(requests)
-        async for i, res in all_gens:
-            generated_texts[i] = res.outputs[0].text
-            lat = time.perf_counter() - start_time[i]
-            latencies[i].append(lat)
-        ret = [{
-            'generated': gt,
-            'expected': req.completion
-        } for gt, req in zip(generated_texts, requests)]
-        end = time.perf_counter()
-        first_latency = pd.Series([lat[0] * 1000 for lat in latencies])
-        next_latency = pd.Series([(lat[-1] - lat[0]) / len(lat[1:]) * 1000
-                                  for lat in latencies])
-        return end - start, ret, (first_latency, next_latency)
-
-
-def sample_requests(tokenizer: PreTrainedTokenizerBase,
-                    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__))
-            args.json_schema_path = os.path.join(dir_path,
-                                                 "structured_schemas",
-                                                 "structured_schema_1.json")
-        with open(args.json_schema_path) as f:
-            schema = json.load(f)
-        prompt = f"Generate an example of a user profile given the following schema: {json.dumps(schema)}"  # noqa: E501
-        input_len = len(tokenizer(prompt).input_ids)
-        print(f"Input length of the prompt: {input_len} tokens")
-        requests = [
-            SampleRequest(prompt=prompt,
-                          prompt_len=input_len,
-                          expected_output_len=args.output_len,
-                          schema=schema,
-                          structure_type=args.structure_type)
-            for _ in range(args.num_prompts)
-        ]
-
-    elif args.dataset == "grammar":
-        schema = """
-            ?start: select_statement
-
-            ?select_statement: "SELECT " column_list " FROM " table_name
-
-            ?column_list: column_name ("," column_name)*
-
-            ?table_name: identifier
-
-            ?column_name: identifier
-
-            ?identifier: /[a-zA-Z_][a-zA-Z0-9_]*/
-        """
-        prompt = "Generate an SQL query to show the 'username' \
-            and 'email' from the 'users' table."
-
-        input_len = len(tokenizer(prompt).input_ids)
-        print(f"Input length of the prompt: {input_len} tokens")
-        requests = [
-            SampleRequest(prompt=prompt,
-                          prompt_len=input_len,
-                          expected_output_len=args.output_len,
-                          schema=schema,
-                          structure_type=args.structure_type)
-            for _ in range(args.num_prompts)
-        ]
-
-    elif args.dataset == "regex":
-        regex = r"\w+@\w+\.com\n"
-        args.regex = regex
-        prompt = "Generate an email address for Alan Turing, \
-            who works in Enigma. End in .com and new line. \
-                Example result: alan.turing@enigma.com\n"
-
-        input_len = len(tokenizer(prompt).input_ids)
-        print(f"Input length of the prompt: {input_len} tokens")
-        requests = [
-            SampleRequest(prompt=prompt,
-                          prompt_len=input_len,
-                          expected_output_len=args.output_len,
-                          schema=regex,
-                          structure_type=args.structure_type)
-            for _ in range(args.num_prompts)
-        ]
-
-    elif args.dataset == "choice":
-        choice = ["Positive", "Negative"]
-        args.choice = choice
-        prompt = "Classify this sentiment: vLLM is wonderful!"
-        input_len = len(tokenizer(prompt).input_ids)
-        print(f"Input length of the prompt: {input_len} tokens")
-        requests = [
-            SampleRequest(prompt=prompt,
-                          prompt_len=input_len,
-                          expected_output_len=args.output_len,
-                          schema=choice,
-                          structure_type=args.structure_type)
-            for _ in range(args.num_prompts)
-        ]
-
-    elif args.dataset == "xgrammar_bench":
-        args.warmup = False
-        requests: List[SampleRequest] = []
-        dataset = datasets.load_dataset("NousResearch/json-mode-eval",
-                                        split="train")
-        print(f"dataset has {len(dataset)} entries")
-        len_dataset = len(dataset)
-        for data_point_idx in range(args.num_prompts):
-            idx = data_point_idx
-            while idx >= len_dataset:
-                idx -= len_dataset
-            schema = dataset["schema"][idx]
-            prompt = tokenizer.apply_chat_template(dataset["prompt"][idx],
-                                                   tokenize=False)
-            input_len = len(tokenizer(prompt).input_ids)
-            completion = dataset["completion"][idx]
-
-            requests.append(
-                SampleRequest(prompt=prompt,
-                              prompt_len=input_len,
-                              expected_output_len=args.output_len,
-                              schema=schema,
-                              completion=completion))
-
-    return requests
-
-
-def evaluate(ret, args):
-
-    def _eval_correctness_json(expected, actual):
-        # extract json string from string using regex
-        import re
-        actual = actual.replace('\n', '').replace(' ', '').strip()
-        try:
-            actual = re.search(r'\{.*\}', actual).group()
-            actual = json.loads(actual)
-        except Exception:
-            return False
-
-        return True
-
-    def _eval_correctness_choice(expected, actual):
-        return actual in args.choice
-
-    def _eval_correctness_regex(expected, actual):
-        import re
-        return re.match(args.regex, actual) is not None
-
-    def _eval_correctness(expected, actual):
-        if args.structure_type == 'json':
-            return _eval_correctness_json(expected, actual)
-        elif args.structure_type == 'regex':
-            return _eval_correctness_regex(expected, actual)
-        elif args.structure_type == 'choice':
-            return _eval_correctness_choice(expected, actual)
-        else:
-            return None
-
-    scores = []
-    for res in ret:
-        score = _eval_correctness(res['expected'], res['generated'])
-        res['correctness'] = score
-        scores.append(score)
-
-    not_none_scores = [score for score in scores if score is not None]
-
-    return (sum(not_none_scores) / len(not_none_scores) *
-            100) if len(not_none_scores) > 0 else None
-
-
-def main(args: argparse.Namespace):
-    print(args)
-    random.seed(args.seed)
-
-    # async engine is working for 'regex', 'choice' and 'grammar'
-    if args.dataset == 'grammar':
-        args.structure_type = 'grammar'
-        args.async_engine = False
-    elif args.dataset == 'regex':
-        args.structure_type = 'regex'
-        args.async_engine = False
-    elif args.dataset == 'choice':
-        args.structure_type = 'choice'
-        args.async_engine = False
-    else:
-        args.structure_type = 'json'
-
-    if args.no_guided_decoding:
-        args.guided_decoding_ratio = 0
-    if args.save_results:
-        result_file_name = f'{args.guided_decoding_ratio}guided'
-        result_file_name += f"_{args.model.split('/')[-1]}"
-        result_file_name += f"_{args.dataset}"
-        result_file_name += f"_{args.num_prompts}"
-        result_file_name += f"_out{args.output_len}"
-        result_file_name += f"_async{args.async_engine}"
-        result_file_name += f"_warmup{args.warmup}"
-        result_file_name += f"_chunkedprefill{args.enable_chunked_prefill}"
-        result_file_name += ".txt"
-    else:
-        result_file_name = None
-
-    # Synthesize a prompt with the given input length.
-    tokenizer = AutoTokenizer.from_pretrained(
-        args.tokenizer, trust_remote_code=args.trust_remote_code)
-    requests = sample_requests(tokenizer, args)
-
-    if args.async_engine:
-        engine_args = AsyncEngineArgs.from_cli_args(args)
-        elapsed_time, ret, (first_latency, next_latency) = uvloop.run(
-            run_vllm_async(requests, engine_args, args.n,
-                           args.guided_decoding_ratio, args.warmup,
-                           args.disable_frontend_multiprocessing))
-    else:
-        engine_args = EngineArgs.from_cli_args(args)
-        elapsed_time, ret = run_vllm(requests, engine_args, args.n,
-                                     args.guided_decoding_ratio, args.warmup)
-        first_latency, next_latency = None, None
-
-    score = evaluate(ret, args)
-    total_num_tokens = sum(request.prompt_len + request.expected_output_len
-                           for request in requests)
-    total_output_tokens = sum(request.expected_output_len
-                              for request in requests)
-    if first_latency is not None:
-        latency_breakdown = "\nFirst token latency(msecs):\n"
-        latency_breakdown += f"{first_latency.describe()}"
-        latency_breakdown += "\nNext token latency(msecs):\n"
-        latency_breakdown += f"{next_latency.describe()}"
-    print(
-        f"Throughput: {len(requests) / elapsed_time:.2f} requests/s, "
-        f"{total_num_tokens / elapsed_time:.2f} total tokens/s, "
-        f"{total_output_tokens / elapsed_time:.2f} output tokens/s",
-        f"Correct rate is {score} %",
-        f"{latency_breakdown if first_latency is not None else ''}")
-
-    # Output JSON results if specified
-    if args.output_json or result_file_name:
-        results = {
-            "elapsed_time": elapsed_time,
-            "num_requests": len(requests),
-            "total_num_tokens": total_num_tokens,
-            "total_output_tokens": total_output_tokens,
-            "requests_per_second": len(requests) / elapsed_time,
-            "tokens_per_second": f"{total_num_tokens / elapsed_time:.2f}",
-            "output_tokens_per_second":
-            f"{total_output_tokens / elapsed_time:.2f}",
-            "correct_rate(%)": score
-        }
-        results = {"outputs": ret, **results}
-        if first_latency is not None:
-            results["first_token_latency(msecs)"] = first_latency.describe(
-            ).to_dict()
-            results["next_token_latency(msecs)"] = next_latency.describe(
-            ).to_dict()
-        if args.output_json:
-            with open(args.output_json, "w") as f:
-                json.dump(results, f, indent=4)
-        elif result_file_name:
-            with open(result_file_name, "w") as f:
-                json.dump(results, f, indent=4)
-
-
-if __name__ == "__main__":
-    parser = FlexibleArgumentParser(description="Benchmark guided decoding.")
-    parser = AsyncEngineArgs.add_cli_args(parser)
-
-    parser.add_argument("--output-len",
-                        type=int,
-                        default=512,
-                        help="Output length for each request. Overrides the "
-                        "output length from the dataset.")
-    parser.add_argument(
-        "--dataset",
-        default='json',
-        choices=['json', 'grammar', 'regex', 'choice', 'xgrammar_bench'])
-    parser.add_argument("--json_schema_path",
-                        type=str,
-                        default=None,
-                        help="Path to json schema.")
-    parser.add_argument("--n",
-                        type=int,
-                        default=1,
-                        help="Number of generated sequences per prompt.")
-    parser.add_argument("--num-prompts",
-                        type=int,
-                        default=10,
-                        help="Number of prompts to process.")
-    parser.add_argument(
-        '--output-json',
-        type=str,
-        default=None,
-        help='Path to save the throughput results in JSON format.')
-    parser.add_argument("--async-engine",
-                        action='store_true',
-                        default=False,
-                        help="Use vLLM async engine rather than LLM class.")
-    parser.add_argument("--no-guided-decoding",
-                        action='store_true',
-                        default=False,
-                        help="Whether to disable JSON decoding or not.")
-    parser.add_argument("--guided-decoding-ratio",
-                        type=float,
-                        default=1.0,
-                        help="Ratio of Guided Decoding requests")
-    parser.add_argument("--disable-frontend-multiprocessing",
-                        action='store_true',
-                        default=False,
-                        help="Disable decoupled async engine frontend.")
-    parser.add_argument("--warmup",
-                        action="store_true",
-                        default=False,
-                        help="Run warmup prompts before benchmark.")
-    parser.add_argument("--save-results",
-                        action="store_true",
-                        default=False,
-                        help="save output results.")
-    args = parser.parse_args()
-    if args.tokenizer is None:
-        args.tokenizer = args.model
-    main(args)

benchmarks/benchmark_latency.py

@@ -1,13 +1,17 @@
+# 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
@@ -17,6 +21,18 @@ 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)
 
@@ -25,6 +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,
@@ -32,12 +52,13 @@ def main(args: argparse.Namespace):
         top_p=1.0,
         ignore_eos=True,
         max_tokens=args.output_len,
+        detokenize=not args.disable_detokenize,
     )
     print(sampling_params)
     dummy_prompt_token_ids = np.random.randint(10000,
                                                size=(args.batch_size,
                                                      args.input_len))
-    dummy_prompts: List[PromptType] = [{
+    dummy_prompts: list[PromptType] = [{
         "prompt_token_ids": batch
     } for batch in dummy_prompt_token_ids.tolist()]
 
@@ -53,7 +74,8 @@ def main(args: argparse.Namespace):
                     beam_width=args.n,
                     max_tokens=args.output_len,
                     ignore_eos=True,
-                ))
+                ),
+            )
 
     def run_to_completion(profile_dir: Optional[str] = None):
         if profile_dir:
@@ -63,7 +85,8 @@ def main(args: argparse.Namespace):
                         torch.profiler.ProfilerActivity.CUDA,
                     ],
                     on_trace_ready=torch.profiler.tensorboard_trace_handler(
-                        str(profile_dir))) as p:
+                        str(profile_dir)),
+            ) as p:
                 llm_generate()
             print(p.key_averages().table(sort_by="self_cuda_time_total"))
         else:
@@ -80,9 +103,8 @@ 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
@@ -94,9 +116,9 @@ 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:
@@ -107,43 +129,57 @@ 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.add_argument(
+        "--disable-detokenize",
+        action="store_true",
+        help=("Do not detokenize responses (i.e. do not include "
+              "detokenization time in the latency measurement)"),
+    )
 
     parser = EngineArgs.add_cli_args(parser)
     args = parser.parse_args()

benchmarks/benchmark_long_document_qa_throughput.py

@@ -1,3 +1,4 @@
+# SPDX-License-Identifier: Apache-2.0
 """
 Offline benchmark to test the long document QA throughput.
 

benchmarks/benchmark_prefix_caching.py

@@ -1,3 +1,4 @@
+# SPDX-License-Identifier: Apache-2.0
 """
 Benchmark the efficiency of prefix caching.
 
@@ -30,7 +31,7 @@ import dataclasses
 import json
 import random
 import time
-from typing import List, Optional, Tuple
+from typing import Optional
 
 from transformers import PreTrainedTokenizerBase
 
@@ -76,9 +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")
 
@@ -98,7 +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:
@@ -121,10 +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)
@@ -143,9 +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])
@@ -193,7 +194,9 @@ def main(args):
 
     llm = LLM(**dataclasses.asdict(engine_args))
 
-    sampling_params = SamplingParams(temperature=0, max_tokens=args.output_len)
+    sampling_params = SamplingParams(temperature=0,
+                                     max_tokens=args.output_len,
+                                     detokenize=not args.disable_detokenize)
 
     print("Testing filtered requests")
     prompts = repeat_and_sort_requests(filtered_requests,
@@ -242,6 +245,12 @@ if __name__ == "__main__":
         "subtract this length when filtering prompts. Only used "
         "when dataset-path is not provided.",
     )
+    parser.add_argument(
+        '--disable-detokenize',
+        action='store_true',
+        help=("Do not detokenize responses (i.e. do not include "
+              "detokenization time in the latency measurement)"),
+    )
 
     parser = EngineArgs.add_cli_args(parser)
     args = parser.parse_args()

benchmarks/benchmark_prioritization.py

@@ -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, 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 +68,20 @@ def sample_requests(
 
 
 def run_vllm(
-    requests: List[Tuple[str, int, int]],
+    requests: list[tuple[str, int, int]],
     n: int,
     engine_args: EngineArgs,
+    disable_detokenize: bool = False,
 ) -> 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 = []
@@ -84,6 +96,7 @@ def run_vllm(
                 top_p=1.0,
                 ignore_eos=True,
                 max_tokens=output_len,
+                detokenize=not disable_detokenize,
             ))
 
     start = time.perf_counter()
@@ -102,15 +115,16 @@ 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)
 
     if args.backend == "vllm":
         elapsed_time = run_vllm(requests, args.n,
-                                EngineArgs.from_cli_args(args))
+                                EngineArgs.from_cli_args(args),
+                                args.disable_detokenize)
     else:
         raise ValueError(f"Unknown backend: {args.backend}")
     total_num_tokens = sum(prompt_len + output_len
@@ -163,6 +177,12 @@ if __name__ == "__main__":
         type=str,
         default=None,
         help='Path to save the throughput results in JSON format.')
+    parser.add_argument(
+        '--disable-detokenize',
+        action='store_true',
+        help=("Do not detokenize responses (i.e. do not include "
+              "detokenization time in the latency measurement)"),
+    )
 
     parser = EngineArgs.add_cli_args(parser)
     args = parser.parse_args()

benchmarks/benchmark_serving.py

@@ -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:
@@ -24,23 +25,20 @@ 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, Iterable
 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
 from backend_request_func import (ASYNC_REQUEST_FUNCS, RequestFuncInput,
                                   RequestFuncOutput)
-from datasets import load_dataset
-from PIL.Image import Image
 from tqdm.asyncio import tqdm
 from transformers import PreTrainedTokenizerBase
 
@@ -54,6 +52,11 @@ try:
 except ImportError:
     from argparse import ArgumentParser as FlexibleArgumentParser
 
+from benchmark_dataset import (BurstGPTDataset, HuggingFaceDataset,
+                               RandomDataset, SampleRequest, ShareGPTDataset,
+                               SonnetDataset, VisionArenaDataset)
+from benchmark_utils import convert_to_pytorch_benchmark_format, write_to_json
+
 MILLISECONDS_TO_SECONDS_CONVERSION = 1000
 
 
@@ -69,326 +72,48 @@ 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]]
-
-
-def sample_sharegpt_requests(
-    dataset_path: str,
-    num_requests: int,
-    tokenizer: PreTrainedTokenizerBase,
-    fixed_output_len: Optional[int] = None,
-) -> List[Tuple[str, int, int, None]]:
-    # Load the dataset.
-    with open(dataset_path, encoding='utf-8') as f:
-        dataset = json.load(f)
-    # Filter out the conversations with less than 2 turns.
-    dataset = [data for data in dataset if len(data["conversations"]) >= 2]
-    # Only keep the first two turns of each conversation.
-    dataset = [(data["conversations"][0]["value"],
-                data["conversations"][1]["value"]) for data in dataset]
-
-    # Shuffle the dataset.
-    random.shuffle(dataset)
-
-    # Filter out sequences that are too long or too short
-    filtered_dataset: List[Tuple[str, int, int]] = []
-    for i in range(len(dataset)):
-        if len(filtered_dataset) == num_requests:
-            break
-
-        # Tokenize the prompts and completions.
-        prompt = dataset[i][0]
-        prompt_token_ids = tokenizer(prompt).input_ids
-        completion = dataset[i][1]
-        completion_token_ids = 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
-        if prompt_len < 4 or (fixed_output_len is None and output_len < 4):
-            # Prune too short sequences.
-            continue
-        if prompt_len > 1024 or prompt_len + output_len > 2048:
-            # Prune too long sequences.
-            continue
-        filtered_dataset.append((prompt, prompt_len, output_len, None))
-
-    return filtered_dataset
-
-
-def sample_sonnet_requests(
-    dataset_path: str,
-    num_requests: int,
-    input_len: int,
-    output_len: int,
-    prefix_len: int,
-    tokenizer: PreTrainedTokenizerBase,
-) -> List[Tuple[str, str, int, int, None]]:
-    assert (
-        input_len > prefix_len
-    ), "'args.sonnet-input-len' must be greater than 'args.prefix-input-len'."
-
-    # Load the dataset.
-    with open(dataset_path, encoding='utf-8') as f:
-        poem_lines = f.readlines()
-
-    # Tokenize the poem lines.
-    poem_token_ids = tokenizer(poem_lines).input_ids
-    average_poem_len = sum(
-        len(token_ids) for token_ids in poem_token_ids) / len(poem_token_ids)
-
-    # Base prefix for all requests.
-    base_prompt = "Pick as many lines as you can from these poem lines:\n"
-    base_message = [{
-        "role": "user",
-        "content": base_prompt,
-    }]
-    base_prompt_formatted = tokenizer.apply_chat_template(
-        base_message, add_generation_prompt=True, tokenize=False)
-    base_prompt_offset = len(tokenizer(base_prompt_formatted).input_ids)
-
-    assert (
-        input_len > base_prompt_offset
-    ), f"Please set 'args.sonnet-input-len' higher than {base_prompt_offset}."
-    num_input_lines = round(
-        (input_len - base_prompt_offset) / average_poem_len)
-
-    # First approximately `prefix_len` number of tokens in the
-    # prompt are fixed poem lines.
-    assert (
-        prefix_len > base_prompt_offset
-    ), f"Please set 'args.sonnet-prefix-len' higher than {base_prompt_offset}."
-
-    num_prefix_lines = round(
-        (prefix_len - base_prompt_offset) / average_poem_len)
-    prefix_lines = poem_lines[:num_prefix_lines]
-
-    # Sample the rest of lines per request.
-    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 +
-                                random.choices(poem_lines, k=num_lines_needed))
-
-        prompt = f"{base_prompt}{sampled_lines}"
-        message = [
-            {
-                "role": "user",
-                "content": prompt,
-            },
-        ]
-        prompt_formatted = tokenizer.apply_chat_template(
-            message, add_generation_prompt=True, tokenize=False)
-        prompt_len = len(tokenizer(prompt_formatted).input_ids)
-        sampled_requests.append(
-            (prompt, prompt_formatted, prompt_len, output_len, None))
-
-    return sampled_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,
-                                                     Collection[str]]]] = []
-    for data in dataset:
-        if len(sampled_requests) == num_requests:
-            break
-
-        prompt = data["turns"][0][0]['content']
-
-        prompt_token_ids = tokenizer(prompt).input_ids
-        if fixed_output_len is None:
-            # Default max output len is set to 128
-            print("--hf-output-len is not provided. Using default value 128.")
-            fixed_output_len = 128
-
-        prompt_len = len(prompt_token_ids)
-        output_len = fixed_output_len
-
-        assert isinstance(
-            data["images"][0],
-            Image), ("Input image format must be `PIL.Image.Image`, "
-                     f"given {type(data['image'])}.")
-        image: Image = data["images"][0]
-        image = image.convert("RGB")
-        image_data = io.BytesIO()
-        image.save(image_data, format='JPEG')
-        image_base64 = base64.b64encode(image_data.getvalue()).decode("utf-8")
-        mm_content = {
-            "type": "image_url",
-            "image_url": {
-                "url": f"data:image/jpeg;base64,{image_base64}"
-            },
-        }
-
-        sampled_requests.append((prompt, prompt_len, output_len, mm_content))
-
-    return sampled_requests
-
-
-def sample_hf_requests(
-    dataset_path: 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]]]]]:
-
-    # 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)
-        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,
-                           split=dataset_split,
-                           streaming=True)
-    assert "conversations" in dataset.features, (
-        "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,
-                                                     Collection[str]]]] = []
-    for data in filtered_dataset:
-        if len(sampled_requests) == num_requests:
-            break
-
-        # Tokenize the prompts and completions.
-        prompt = data["conversations"][0]["value"]
-        prompt_token_ids = tokenizer(prompt).input_ids
-        completion = data["conversations"][1]["value"]
-        completion_token_ids = 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
-        if fixed_output_len is None and (prompt_len < 4 or output_len < 4):
-            # Prune too short sequences.
-            continue
-        if fixed_output_len is None and \
-            (prompt_len > 1024 or prompt_len + output_len > 2048):
-            # Prune too long sequences.
-            continue
-
-        if "image" in data and isinstance(data["image"], Image):
-            image: Image = data["image"]
-            image = image.convert("RGB")
-            image_data = io.BytesIO()
-            image.save(image_data, format='JPEG')
-            image_base64 = base64.b64encode(
-                image_data.getvalue()).decode("utf-8")
-            mm_content = {
-                "type": "image_url",
-                "image_url": {
-                    "url": f"data:image/jpeg;base64,{image_base64}"
-                },
-            }
-        elif "image" in data and isinstance(data["image"], str):
-            if (data["image"].startswith("http://") or \
-                data["image"].startswith("file://")):
-                image_url = data["image"]
-            else:
-                image_url = f"file://{data['image']}"
-
-            mm_content = {
-                "type": "image_url",
-                "image_url": {
-                    "url": image_url
-                },
-            }
-        else:
-            mm_content = None
-
-        sampled_requests.append((prompt, prompt_len, output_len, mm_content))
-
-    return sampled_requests
-
-
-def sample_random_requests(
-    prefix_len: int,
-    input_len: int,
-    output_len: int,
-    num_prompts: int,
-    range_ratio: float,
-    tokenizer: PreTrainedTokenizerBase,
-) -> List[Tuple[str, int, int]]:
-    prefix_token_ids = np.random.randint(0,
-                                         tokenizer.vocab_size,
-                                         size=prefix_len).tolist()
-
-    input_lens = np.random.randint(
-        int(input_len * range_ratio),
-        input_len + 1,
-        size=num_prompts,
-    )
-    output_lens = np.random.randint(
-        int(output_len * range_ratio),
-        output_len + 1,
-        size=num_prompts,
-    )
-    offsets = np.random.randint(0, tokenizer.vocab_size, size=num_prompts)
-    input_requests = []
-    for i in range(num_prompts):
-        prompt = tokenizer.decode(prefix_token_ids +
-                                  [(offsets[i] + i + j) % tokenizer.vocab_size
-                                   for j in range(input_lens[i])])
-
-        input_requests.append((prompt, int(prefix_len + input_lens[i]),
-                               int(output_lens[i]), None))
-
-    return input_requests
+    percentiles_e2el_ms: list[tuple[float, float]]
 
 
 async def get_request(
-    input_requests: List[Tuple[str, int, int]],
+    input_requests: list[SampleRequest],
     request_rate: float,
     burstiness: float = 1.0,
-) -> AsyncGenerator[Tuple[str, int, int], None]:
+) -> AsyncGenerator[SampleRequest, None]:
     """
-    Asynchronously generates requests at a specified rate 
+    Asynchronously generates requests at a specified rate
     with OPTIONAL burstiness.
-    
+
     Args:
-        input_requests: 
-            A list of input requests, each represented as a tuple.
-        request_rate: 
+        input_requests:
+            A list of input requests, each represented as a SampleRequest.
+        request_rate:
             The rate at which requests are generated (requests/s).
-        burstiness (optional): 
-            The burstiness factor of the request generation. 
+        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)
+    input_requests: Iterable[SampleRequest] = iter(input_requests)
 
     # Calculate scale parameter theta to maintain the desired request_rate.
     assert burstiness > 0, (
@@ -410,23 +135,23 @@ async def get_request(
 
 
 def calculate_metrics(
-    input_requests: List[Tuple[str, int, int]],
-    outputs: List[RequestFuncOutput],
+    input_requests: list[SampleRequest],
+    outputs: list[RequestFuncOutput],
     dur_s: float,
     tokenizer: PreTrainedTokenizerBase,
-    selected_percentile_metrics: List[str],
-    selected_percentiles: List[float],
-    goodput_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:
             output_len = outputs[i].output_tokens
@@ -441,7 +166,7 @@ def calculate_metrics(
                     tokenizer(outputs[i].generated_text,
                               add_special_tokens=False).input_ids)
             actual_output_lens.append(output_len)
-            total_input += input_requests[i][1]
+            total_input += input_requests[i].prompt_len
             tpot = 0
             if output_len > 1:
                 latency_minus_ttft = outputs[i].latency - outputs[i].ttft
@@ -524,18 +249,18 @@ async def benchmark(
     model_id: str,
     model_name: str,
     tokenizer: PreTrainedTokenizerBase,
-    input_requests: List[Tuple[str, int, int]],
+    input_requests: list[SampleRequest],
     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[float],
     ignore_eos: bool,
-    goodput_config_dict: Dict[str, float],
+    goodput_config_dict: dict[str, float],
     max_concurrency: Optional[int],
+    lora_modules: Optional[Iterable[str]],
 ):
     if backend in ASYNC_REQUEST_FUNCS:
         request_func = ASYNC_REQUEST_FUNCS[backend]
@@ -543,12 +268,16 @@ async def benchmark(
         raise ValueError(f"Unknown backend: {backend}")
 
     print("Starting initial single prompt test run...")
-    test_prompt, test_prompt_len, test_output_len, test_mm_content = (
-        input_requests[0])
+    test_prompt, test_prompt_len, test_output_len, test_mm_content = \
+        input_requests[0].prompt, input_requests[0].prompt_len, \
+        input_requests[0].expected_output_len, \
+            input_requests[0].multi_modal_data
+
     if backend != "openai-chat" and test_mm_content is not None:
         # multi-modal benchmark is only available on OpenAI Chat backend.
         raise ValueError(
             "Multi-modal content is only supported on 'openai-chat' backend.")
+    assert test_mm_content is None or isinstance(test_mm_content, dict)
     test_input = RequestFuncInput(
         model=model_id,
         model_name=model_name,
@@ -557,10 +286,10 @@ async def benchmark(
         prompt_len=test_prompt_len,
         output_len=test_output_len,
         logprobs=logprobs,
-        best_of=best_of,
         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(
@@ -569,6 +298,12 @@ 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,
@@ -578,7 +313,6 @@ async def benchmark(
                                          prompt_len=test_prompt_len,
                                          output_len=test_output_len,
                                          logprobs=logprobs,
-                                         best_of=best_of,
                                          multi_modal_content=test_mm_content,
                                          ignore_eos=ignore_eos)
         profile_output = await request_func(request_func_input=profile_input)
@@ -612,24 +346,30 @@ 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,
-                                              model_name=model_name,
+        prompt, prompt_len, output_len, mm_content = request.prompt, \
+            request.prompt_len, request.expected_output_len, \
+                request.multi_modal_data
+        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,
                                               output_len=output_len,
                                               logprobs=logprobs,
-                                              best_of=best_of,
                                               multi_modal_content=mm_content,
                                               ignore_eos=ignore_eos)
         tasks.append(
             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...")
@@ -640,7 +380,6 @@ async def benchmark(
             prompt_len=test_prompt_len,
             output_len=test_output_len,
             logprobs=logprobs,
-            best_of=best_of,
         )
         profile_output = await request_func(request_func_input=profile_input)
         if profile_output.success:
@@ -774,6 +513,31 @@ def parse_goodput(slo_pairs):
     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):
     print(args)
     random.seed(args.seed)
@@ -796,81 +560,77 @@ 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(
-            dataset_path=args.dataset_path,
-            num_requests=args.num_prompts,
-            tokenizer=tokenizer,
-            fixed_output_len=args.sharegpt_output_len,
-        )
-
-    elif args.dataset_name == "sonnet":
-        # Do not format the prompt, pass to message directly
+    if args.dataset_name == "sonnet":
+        dataset = SonnetDataset(dataset_path=args.dataset_path)
+        # For the "sonnet" dataset, formatting depends on the backend.
         if args.backend == "openai-chat":
-            input_requests = sample_sonnet_requests(
-                dataset_path=args.dataset_path,
-                num_requests=args.num_prompts,
-                input_len=args.sonnet_input_len,
-                output_len=args.sonnet_output_len,
-                prefix_len=args.sonnet_prefix_len,
-                tokenizer=tokenizer,
-            )
-            input_requests = [(prompt, prompt_len, output_len, None)
-                              for prompt, prompt_formatted, prompt_len,
-                              output_len, _ in input_requests]
+            input_requests = dataset.sample(num_requests=args.num_prompts,
+                                            input_len=args.sonnet_input_len,
+                                            output_len=args.sonnet_output_len,
+                                            prefix_len=args.sonnet_prefix_len,
+                                            tokenizer=tokenizer,
+                                            return_prompt_formatted=False)
         else:
-            assert (
-                tokenizer.chat_template or tokenizer.default_chat_template
-            ), "Tokenizer/model must have chat template for sonnet dataset."
-            input_requests = sample_sonnet_requests(
-                dataset_path=args.dataset_path,
-                num_requests=args.num_prompts,
-                input_len=args.sonnet_input_len,
-                output_len=args.sonnet_output_len,
-                prefix_len=args.sonnet_prefix_len,
-                tokenizer=tokenizer,
-            )
-            input_requests = [(prompt_formatted, prompt_len, output_len, None)
-                              for prompt, prompt_formatted, prompt_len,
-                              output_len, _ in input_requests]
+            assert tokenizer.chat_template or tokenizer.default_chat_template, (
+                "Tokenizer/model must have chat template for sonnet dataset.")
+            input_requests = dataset.sample(num_requests=args.num_prompts,
+                                            input_len=args.sonnet_input_len,
+                                            output_len=args.sonnet_output_len,
+                                            prefix_len=args.sonnet_prefix_len,
+                                            tokenizer=tokenizer,
+                                            return_prompt_formatted=True)
 
     elif args.dataset_name == "hf":
-        input_requests = sample_hf_requests(
+        # Choose between VisionArenaDataset
+        # and HuggingFaceDataset based on provided parameters.
+        dataset_class = (VisionArenaDataset if args.dataset_path
+                         == VisionArenaDataset.VISION_ARENA_DATASET_PATH
+                         and args.hf_subset is None else HuggingFaceDataset)
+        input_requests = dataset_class(
             dataset_path=args.dataset_path,
             dataset_subset=args.hf_subset,
             dataset_split=args.hf_split,
+        ).sample(
             num_requests=args.num_prompts,
             tokenizer=tokenizer,
             random_seed=args.seed,
-            fixed_output_len=args.hf_output_len,
-        )
-
-    elif args.dataset_name == "random":
-        input_requests = sample_random_requests(
-            prefix_len=args.random_prefix_len,
-            input_len=args.random_input_len,
-            output_len=args.random_output_len,
-            num_prompts=args.num_prompts,
-            range_ratio=args.random_range_ratio,
-            tokenizer=tokenizer,
+            output_len=args.hf_output_len,
         )
 
     else:
-        raise ValueError(f"Unknown dataset: {args.dataset_name}")
+        # For datasets that follow a similar structure, use a mapping.
+        dataset_mapping = {
+            "sharegpt":
+            lambda: ShareGPTDataset(random_seed=args.seed,
+                                    dataset_path=args.dataset_path).sample(
+                                        tokenizer=tokenizer,
+                                        num_requests=args.num_prompts,
+                                        output_len=args.sharegpt_output_len,
+                                    ),
+            "burstgpt":
+            lambda: BurstGPTDataset(random_seed=args.seed,
+                                    dataset_path=args.dataset_path).
+            sample(tokenizer=tokenizer, num_requests=args.num_prompts),
+            "random":
+            lambda: RandomDataset(dataset_path=args.dataset_path).sample(
+                tokenizer=tokenizer,
+                num_requests=args.num_prompts,
+                prefix_len=args.random_prefix_len,
+                input_len=args.random_input_len,
+                output_len=args.random_output_len,
+                range_ratio=args.random_range_ratio,
+            )
+        }
 
+        try:
+            input_requests = dataset_mapping[args.dataset_name]()
+        except KeyError as err:
+            raise ValueError(f"Unknown dataset: {args.dataset_name}") from err
     goodput_config_dict = check_goodput_args(args)
 
     # Avoid GC processing "static" data - reduce pause times.
@@ -887,7 +647,6 @@ def main(args: argparse.Namespace):
             tokenizer=tokenizer,
             input_requests=input_requests,
             logprobs=args.logprobs,
-            best_of=args.best_of,
             request_rate=args.request_rate,
             burstiness=args.burstiness,
             disable_tqdm=args.disable_tqdm,
@@ -899,11 +658,12 @@ def main(args: argparse.Namespace):
             ignore_eos=args.ignore_eos,
             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")
@@ -911,7 +671,6 @@ def main(args: argparse.Namespace):
         result_json["backend"] = backend
         result_json["model_id"] = model_id
         result_json["tokenizer_id"] = tokenizer_id
-        result_json["best_of"] = args.best_of
         result_json["num_prompts"] = args.num_prompts
 
         # Metadata
@@ -925,9 +684,18 @@ def main(args: argparse.Namespace):
                         "Invalid metadata format. Please use KEY=VALUE format."
                     )
 
+        if not args.save_detailed:
+            # Remove fields with too many data points
+            for field in [
+                    "input_lens", "output_lens", "ttfts", "itls",
+                    "generated_texts", "errors"
+            ]:
+                if field in result_json:
+                    del result_json[field]
+
         # 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
 
@@ -945,6 +713,7 @@ 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__":
@@ -962,7 +731,8 @@ 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",
@@ -970,18 +740,11 @@ 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",
@@ -1014,13 +777,6 @@ if __name__ == "__main__":
         help=
         "Name or path of the tokenizer, if not using the default tokenizer.",  # noqa: E501
     )
-    parser.add_argument(
-        "--best-of",
-        type=int,
-        default=1,
-        help="Generates `best_of` sequences per prompt and "
-        "returns the best one.",
-    )
     parser.add_argument("--use-beam-search", action="store_true")
     parser.add_argument(
         "--num-prompts",
@@ -1081,6 +837,12 @@ if __name__ == "__main__":
         action="store_true",
         help="Specify to save benchmark results to a json file",
     )
+    parser.add_argument(
+        "--save-detailed",
+        action="store_true",
+        help="When saving the results, whether to include per request "
+        "information such as response, error, ttfs, tpots, etc.",
+    )
     parser.add_argument(
         "--metadata",
         metavar="KEY=VALUE",
@@ -1223,11 +985,12 @@ 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,
@@ -1236,5 +999,13 @@ if __name__ == "__main__":
                         "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)

benchmarks/benchmark_serving_structured_output.py

@@ -1,4 +1,5 @@
-r"""Benchmark online serving throughput with guided decoding.
+# SPDX-License-Identifier: Apache-2.0
+r"""Benchmark online serving throughput with structured outputs.
 
 On the server side, run one of the following commands:
     (vLLM OpenAI API server)
@@ -8,12 +9,12 @@ 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_structured_output.py \
         --backend <backend> \
         --model <your_model> \
         --dataset json \
-        --guided-decoding-ratio 1.0 \
-        --guided-decoding-backend xgrammar \
+        --structured-output-ratio 1.0 \
+        --structured-output-backend xgrammar \
         --request-rate 10 \
         --num-prompts 1000
 
@@ -23,14 +24,17 @@ On the client side, run:
 """
 import argparse
 import asyncio
+import copy
 import dataclasses
 import json
 import os
 import random
 import time
+import uuid
 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
@@ -50,6 +54,9 @@ try:
 except ImportError:
     from argparse import ArgumentParser as FlexibleArgumentParser
 
+from vllm.v1.structured_output.utils import (
+    has_xgrammar_unsupported_json_features)
+
 MILLISECONDS_TO_SECONDS_CONVERSION = 1000
 
 
@@ -65,22 +72,22 @@ 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
@@ -103,25 +110,44 @@ class SampleRequest:
 
 
 def sample_requests(tokenizer: PreTrainedTokenizerBase,
-                    args: argparse.Namespace) -> List[SampleRequest]:
-    if args.dataset == 'json':
+                    args: argparse.Namespace) -> list[SampleRequest]:
+    if args.dataset == 'json' or args.dataset == 'json-unique':
         if args.json_schema_path is None:
             dir_path = os.path.dirname(os.path.realpath(__file__))
             args.json_schema_path = os.path.join(dir_path,
                                                  "structured_schemas",
                                                  "structured_schema_1.json")
+        json_schemas = []
         with open(args.json_schema_path) as f:
             schema = json.load(f)
-        prompt = f"Generate an example of a user profile given the following schema: {json.dumps(schema)}"  # noqa: E501
-        input_len = len(tokenizer(prompt).input_ids)
-        print(f"Input length of the prompt: {input_len} tokens")
+
+        if args.dataset == 'json-unique':
+            json_schemas = [
+                copy.deepcopy(schema) for _ in range(args.num_prompts)
+            ]
+            for i in range(len(json_schemas)):
+                json_schemas[i]["properties"][
+                    f"__optional_field_{uuid.uuid4()}"] = {
+                        "type":
+                        "string",
+                        "description":
+                        "An unique optional field to avoid cached schemas"
+                    }
+
+        def gen_prompt(index: int):
+            schema = json_schemas[index % len(json_schemas)]
+            return f"Generate an example of a user profile given the following schema: {json.dumps(schema)}"  # noqa: E501
+
+        def get_schema(index: int):
+            return json_schemas[index % len(json_schemas)]
+
         requests = [
-            SampleRequest(prompt=prompt,
-                          prompt_len=input_len,
+            SampleRequest(prompt=gen_prompt(i),
+                          prompt_len=len(tokenizer(gen_prompt(i)).input_ids),
                           expected_output_len=args.output_len,
-                          schema=schema,
+                          schema=get_schema(i),
                           structure_type=args.structure_type)
-            for _ in range(args.num_prompts)
+            for i in range(args.num_prompts)
         ]
 
     elif args.dataset == "grammar":
@@ -186,10 +212,20 @@ 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")
+        full_dataset_len = len(dataset)
+
+        def _filter_func(item):
+            import json
+            schema = json.loads(item["schema"])
+            return not has_xgrammar_unsupported_json_features(schema)
+
+        dataset = dataset.filter(_filter_func)
+        num_filtered_out = full_dataset_len - len(dataset)
+        print(f"dataset has {len(dataset)} entries after filtering "
+              f"out {num_filtered_out} entries with unsupported features")
         len_dataset = len(dataset)
         for data_point_idx in range(args.num_prompts):
             idx = data_point_idx
@@ -213,26 +249,26 @@ 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 
+    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)
@@ -257,22 +293,23 @@ 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 +323,10 @@ 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 +336,28 @@ 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 +403,18 @@ 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,
+    structured_output_ratio: float,
+    structured_output_backend: str,
+    goodput_config_dict: Optional[dict[str, float]] = None,
 ):
     if backend in ASYNC_REQUEST_FUNCS:
         request_func = ASYNC_REQUEST_FUNCS[backend]
@@ -365,16 +425,18 @@ async def benchmark(
         extra_body = {}
         # Add the schema to the extra_body
         extra_body[request.structure_type] = request.schema
-        # Add the specific guided_decoding_backend
-        extra_body["guided_decoding_backend"] = guided_decoding_backend
+        # Add the specific structured_output_backend
+        extra_body["guided_decoding_backend"] = structured_output_backend
         return extra_body
 
     print("Starting initial single prompt test run...")
-    guided_decoding_req_idx = random.sample(
+    structured_output_req_idx = random.sample(
         range(len(input_requests)),
-        int(len(input_requests) * guided_decoding_ratio))
+        int(len(input_requests) * structured_output_ratio))
 
     test_request = input_requests[0]
+    test_req_extra_body = (prepare_extra_body(test_request)
+                           if 0 in structured_output_req_idx else None)
     test_input = RequestFuncInput(
         model=model_id,
         prompt=test_request.prompt,
@@ -382,7 +444,7 @@ async def benchmark(
         prompt_len=test_request.prompt_len,
         output_len=test_request.expected_output_len,
         ignore_eos=ignore_eos,
-        extra_body=prepare_extra_body(test_request),
+        extra_body=test_req_extra_body,
     )
     test_output = await request_func(request_func_input=test_input)
     if not test_output.success:
@@ -401,7 +463,7 @@ async def benchmark(
             prompt_len=test_request.prompt_len,
             output_len=test_request.expected_output_len,
             ignore_eos=ignore_eos,
-            extra_body=prepare_extra_body(test_request),
+            extra_body=test_req_extra_body,
         )
         profile_output = await request_func(request_func_input=profile_input)
         if profile_output.success:
@@ -434,12 +496,12 @@ 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(
-            request) if i in guided_decoding_req_idx else None
+            request) if i in structured_output_req_idx else None
         request_func_input = RequestFuncInput(
             model=model_id,
             prompt=request.prompt,
@@ -454,7 +516,7 @@ 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 +544,7 @@ 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 +556,9 @@ 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 +682,40 @@ 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)
@@ -632,8 +732,11 @@ def main(args: argparse.Namespace):
         api_url = f"http://{args.host}:{args.port}{args.endpoint}"
         base_url = f"http://{args.host}:{args.port}"
 
-    tokenizer = get_tokenizer(tokenizer_id,
-                              trust_remote_code=args.trust_remote_code)
+    tokenizer = get_tokenizer(
+        tokenizer_id,
+        trust_remote_code=args.trust_remote_code,
+        tokenizer_mode=args.tokenizer_mode,
+    )
 
     if args.dataset == 'grammar':
         args.structure_type = 'guided_grammar'
@@ -644,10 +747,10 @@ def main(args: argparse.Namespace):
     else:
         args.structure_type = 'guided_json'
 
-    if args.no_guided_decoding:
-        args.guided_decoding_ratio = 0
+    if args.no_structured_output:
+        args.structured_output_ratio = 0
     if args.save_results:
-        result_file_name = f'{args.guided_decoding_ratio}guided'
+        result_file_name = f'{args.structured_output_ratio}guided'
         result_file_name += f"_{backend}"
         result_file_name += f"_{args.request_rate}qps"
         result_file_name += f"_{args.model.split('/')[-1]}"
@@ -660,6 +763,8 @@ 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,
@@ -678,8 +783,9 @@ def main(args: argparse.Namespace):
             ],
             ignore_eos=args.ignore_eos,
             max_concurrency=args.max_concurrency,
-            guided_decoding_ratio=args.guided_decoding_ratio,
-            guided_decoding_backend=args.guided_decoding_backend,
+            structured_output_ratio=args.structured_output_ratio,
+            structured_output_backend=args.structured_output_backend,
+            goodput_config_dict=goodput_config_dict,
         ))
 
     # Save config and results to json
@@ -730,7 +836,8 @@ 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",
@@ -738,10 +845,12 @@ if __name__ == "__main__":
         default="/v1/completions",
         help="API endpoint.",
     )
-    parser.add_argument(
-        "--dataset",
-        default='json',
-        choices=['json', 'grammar', 'regex', 'choice', 'xgrammar_bench'])
+    parser.add_argument("--dataset",
+                        default='json',
+                        choices=[
+                            'json', 'json-unique', 'grammar', 'regex',
+                            'choice', 'xgrammar_bench'
+                        ])
     parser.add_argument("--json_schema_path",
                         type=str,
                         default=None,
@@ -770,6 +879,13 @@ if __name__ == "__main__":
         help=
         "Name or path of the tokenizer, if not using the default tokenizer.",  # noqa: E501
     )
+    parser.add_argument(
+        "--tokenizer-mode",
+        type=str,
+        default="auto",
+        help=
+        "Name or path of the tokenizer, if not using the default tokenizer.",  # noqa: E501
+    )
     parser.add_argument(
         "--num-prompts",
         type=int,
@@ -863,19 +979,32 @@ if __name__ == "__main__":
         "Default value is \"99\". "
         "Use \"--percentile-metrics\" to select metrics.",
     )
-    parser.add_argument("--no-guided-decoding",
+    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-structured-output",
                         action='store_true',
                         default=False,
                         help="Whether to disable JSON decoding or not.")
-    parser.add_argument("--guided-decoding-ratio",
+    parser.add_argument("--structured-output-ratio",
                         type=float,
                         default=1.0,
-                        help="Ratio of Guided Decoding requests")
-    parser.add_argument("--guided-decoding-backend",
-                        type=str,
-                        choices=["outlines", "lm-format-enforcer", "xgrammar"],
-                        default="xgrammar",
-                        help="Backend to use for guided decoding")
+                        help="Ratio of Structured Outputs requests")
+    parser.add_argument(
+        "--structured-output-backend",
+        type=str,
+        choices=["outlines", "lm-format-enforcer", "xgrammar", "guidance"],
+        default="xgrammar",
+        help="Backend to use for structured outputs")
 
     args = parser.parse_args()
     main(args)

benchmarks/benchmark_throughput.py

@@ -1,15 +1,20 @@
+# SPDX-License-Identifier: Apache-2.0
 """Benchmark offline inference throughput."""
 import argparse
 import dataclasses
 import json
+import os
 import random
 import time
-from functools import cache
-from typing import Dict, List, Optional, Tuple
+import warnings
+from typing import Any, Optional, Union
 
 import torch
 import uvloop
-from PIL import Image
+from benchmark_dataset import (BurstGPTDataset, HuggingFaceDataset,
+                               RandomDataset, SampleRequest, ShareGPTDataset,
+                               SonnetDataset, VisionArenaDataset)
+from benchmark_utils import convert_to_pytorch_benchmark_format, write_to_json
 from tqdm import tqdm
 from transformers import (AutoModelForCausalLM, AutoTokenizer,
                           PreTrainedTokenizerBase)
@@ -17,163 +22,35 @@ from transformers import (AutoModelForCausalLM, AutoTokenizer,
 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.inputs import TextPrompt, TokensPrompt
 from vllm.lora.request import LoRARequest
-from vllm.lora.utils import get_adapter_absolute_path
-from vllm.multimodal import MultiModalDataDict
+from vllm.outputs import RequestOutput
 from vllm.sampling_params import BeamSearchParams
-from vllm.transformers_utils.tokenizer import AnyTokenizer, get_lora_tokenizer
 from vllm.utils import FlexibleArgumentParser, merge_async_iterators
 
 
-@dataclasses.dataclass
-class SampleRequest:
-    """A class representing a single inference request for benchmarking.
-
-    Attributes:
-        prompt: The input text prompt for the model.
-        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:
-    """Prepend and append special tokens around the question to form a prompt.
-
-    Args:
-        question: The input question text to wrap with special tokens
-        model: The name of the model being used, to determine which special
-            tokens to add
-
-    Returns:
-        The formatted prompt string with appropriate special tokens for the
-            model
-
-    Raises:
-        ValueError: If an unsupported model name is provided
-    """
-    model = model.lower()
-    if "pixtral" in model:
-        return f"<s>[INST]{question}\n[IMG][/INST]"
-    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]:
-
-    dataset_path: str = args.dataset
-    num_requests: int = args.num_prompts
-    fixed_output_len: Optional[int] = args.output_len
-    model: str = args.model
-    if fixed_output_len is not None and fixed_output_len < 4:
-        raise ValueError("output_len too small")
-
-    # Load the dataset.
-    with open(dataset_path) as f:
-        dataset = json.load(f)
-    # Filter out the conversations with less than 2 turns.
-    dataset = [data for data in dataset if len(data["conversations"]) >= 2]
-    # Shuffle the dataset.
-    random.shuffle(dataset)
-
-    # Filter out sequences that are too long or too short
-    filtered_dataset: List[SampleRequest] = []
-    for data in tqdm(dataset,
-                     total=len(filtered_dataset),
-                     desc="sampling requests"):
-        if len(filtered_dataset) == num_requests:
-            break
-
-        # Only keep the first two turns of each conversation.
-        prompt = data["conversations"][0]["value"]
-        completion = data["conversations"][1]["value"]
-
-        multi_modal_data: Optional[MultiModalDataDict] = None
-        if "image" in data:
-            multi_modal_data = multi_modal_data or {}
-            image_path = data["image"]
-            # TODO(vllm-project/vllm/issues/9778): Support multiple images.
-            assert isinstance(image_path,
-                              str), "Only support single image input"
-            try:
-                multi_modal_data["image"] = Image.open(image_path).convert(
-                    "RGB")
-            except FileNotFoundError:
-                # Ignore datapoint where asset is missing
-                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 = 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
-        if prompt_len < 4 or output_len < 4:
-            # Prune too short sequences.
-            continue
-        if prompt_len > 1024 or prompt_len + output_len > 2048:
-            # Prune too long sequences.
-            continue
-        filtered_dataset.append(
-            SampleRequest(prompt=prompt,
-                          prompt_len=prompt_len,
-                          expected_output_len=output_len,
-                          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:
+    disable_detokenize: bool = False,
+) -> tuple[float, Optional[list[RequestOutput]]]:
     from vllm import LLM, SamplingParams
     llm = LLM(**dataclasses.asdict(engine_args))
-
+    assert all(
+        llm.llm_engine.model_config.max_model_len >= (
+            request.prompt_len + request.expected_output_len)
+        for request in requests), (
+            "Please ensure that max_model_len is greater than the sum of"
+            " prompt_len and expected_output_len for all requests.")
     # Add the requests to the engine.
-    prompts: List[TextPrompt] = []
-    sampling_params: List[SamplingParams] = []
+    prompts: list[Union[TextPrompt, TokensPrompt]] = []
+    sampling_params: list[SamplingParams] = []
     for request in requests:
         prompts.append(
+            TokensPrompt(prompt_token_ids=request.prompt["prompt_token_ids"],
+                       multi_modal_data=request.multi_modal_data)
+            if "prompt_token_ids" in request.prompt else \
             TextPrompt(prompt=request.prompt,
                        multi_modal_data=request.multi_modal_data))
         sampling_params.append(
@@ -183,19 +60,21 @@ def run_vllm(
                 top_p=1.0,
                 ignore_eos=True,
                 max_tokens=request.expected_output_len,
+                detokenize=not disable_detokenize,
             ))
-    lora_requests: Optional[List[LoRARequest]] = None
+    lora_requests: Optional[list[LoRARequest]] = None
     if engine_args.enable_lora:
         lora_requests = [request.lora_request for request in requests]
 
     use_beam_search = False
 
+    outputs = None
     if not use_beam_search:
         start = time.perf_counter()
-        llm.generate(prompts,
-                     sampling_params,
-                     lora_request=lora_requests,
-                     use_tqdm=True)
+        outputs = llm.generate(prompts,
+                               sampling_params,
+                               lora_request=lora_requests,
+                               use_tqdm=True)
         end = time.perf_counter()
     else:
         assert lora_requests is None, "BeamSearch API does not support LoRA"
@@ -213,26 +92,75 @@ def run_vllm(
                 ignore_eos=True,
             ))
         end = time.perf_counter()
-    return end - start
+    return end - start, outputs
+
+
+def run_vllm_chat(
+        requests: list[SampleRequest],
+        n: int,
+        engine_args: EngineArgs,
+        disable_detokenize: bool = False) -> tuple[float, list[RequestOutput]]:
+    """
+    Run vLLM chat benchmark. This function is recommended ONLY for benchmarking
+    multimodal models as it properly handles multimodal inputs and chat
+    formatting. For non-multimodal models, use run_vllm() instead.
+    """
+    from vllm import LLM, SamplingParams
+    llm = LLM(**dataclasses.asdict(engine_args))
+
+    assert all(
+        llm.llm_engine.model_config.max_model_len >= (
+            request.prompt_len + request.expected_output_len)
+        for request in requests), (
+            "Please ensure that max_model_len is greater than the sum of "
+            "prompt_len and expected_output_len for all requests.")
+
+    prompts = []
+    sampling_params: list[SamplingParams] = []
+    for request in requests:
+        prompts.append(request.prompt)
+        sampling_params.append(
+            SamplingParams(
+                n=n,
+                temperature=1.0,
+                top_p=1.0,
+                ignore_eos=True,
+                max_tokens=request.expected_output_len,
+                detokenize=not disable_detokenize,
+            ))
+    start = time.perf_counter()
+    outputs = llm.chat(prompts, sampling_params, use_tqdm=True)
+    end = time.perf_counter()
+    return end - start, outputs
 
 
 async def run_vllm_async(
-    requests: List[SampleRequest],
+    requests: list[SampleRequest],
     n: int,
     engine_args: AsyncEngineArgs,
     disable_frontend_multiprocessing: bool = False,
+    disable_detokenize: bool = False,
 ) -> float:
     from vllm import SamplingParams
 
     async with build_async_engine_client_from_engine_args(
             engine_args, disable_frontend_multiprocessing) 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] = []
-        lora_requests: List[Optional[LoRARequest]] = []
+        prompts: list[Union[TextPrompt, TokensPrompt]] = []
+        sampling_params: list[SamplingParams] = []
+        lora_requests: list[Optional[LoRARequest]] = []
         for request in requests:
             prompts.append(
+                TokensPrompt(prompt_token_ids=request.prompt["prompt_token_ids"],
+                        multi_modal_data=request.multi_modal_data)
+                if "prompt_token_ids" in request.prompt else \
                 TextPrompt(prompt=request.prompt,
                            multi_modal_data=request.multi_modal_data))
             sampling_params.append(
@@ -242,6 +170,7 @@ async def run_vllm_async(
                     top_p=1.0,
                     ignore_eos=True,
                     max_tokens=request.expected_output_len,
+                    detokenize=not disable_detokenize,
                 ))
             lora_requests.append(request.lora_request)
 
@@ -262,12 +191,13 @@ async def run_vllm_async(
 
 
 def run_hf(
-    requests: List[SampleRequest],
+    requests: list[SampleRequest],
     model: str,
     tokenizer: PreTrainedTokenizerBase,
     n: int,
     max_batch_size: int,
     trust_remote_code: bool,
+    disable_detokenize: bool = False,
 ) -> float:
     llm = AutoModelForCausalLM.from_pretrained(
         model, torch_dtype=torch.float16, trust_remote_code=trust_remote_code)
@@ -278,7 +208,7 @@ 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)):
@@ -307,8 +237,9 @@ def run_hf(
             use_cache=True,
             max_new_tokens=max_output_len,
         )
-        # Include the decoding time.
-        tokenizer.batch_decode(llm_outputs, skip_special_tokens=True)
+        if not disable_detokenize:
+            # Include the decoding time.
+            tokenizer.batch_decode(llm_outputs, skip_special_tokens=True)
         pbar.update(len(batch))
 
         # Clear the batch.
@@ -320,7 +251,7 @@ def run_hf(
 
 
 def run_mii(
-    requests: List[SampleRequest],
+    requests: list[SampleRequest],
     model: str,
     tensor_parallel_size: int,
     output_len: int,
@@ -337,58 +268,86 @@ 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 get_requests(args, tokenizer):
+    # Common parameters for all dataset types.
+    common_kwargs = {
+        "dataset_path": args.dataset_path,
+        "random_seed": args.seed,
+    }
+    sample_kwargs = {
+        "tokenizer": tokenizer,
+        "lora_path": args.lora_path,
+        "max_loras": args.max_loras,
+        "num_requests": args.num_prompts,
+        "input_len": args.input_len,
+        "output_len": args.output_len,
+    }
+    if args.dataset_path is None or args.dataset_name == "random":
+        sample_kwargs["range_ratio"] = args.random_range_ratio
+        sample_kwargs["prefix_len"] = args.prefix_len
+        dataset_cls = RandomDataset
+    elif args.dataset_name == "sharegpt":
+        dataset_cls = ShareGPTDataset
+        if args.backend == "vllm-chat":
+            sample_kwargs["enable_multimodal_chat"] = True
+    elif args.dataset_name == "sonnet":
+        assert tokenizer.chat_template or tokenizer.default_chat_template, (
+            "Tokenizer/model must have chat template for sonnet dataset.")
+        dataset_cls = SonnetDataset
+        sample_kwargs["prefix_len"] = args.prefix_len
+        sample_kwargs["return_prompt_formatted"] = True
+    elif args.dataset_name == "burstgpt":
+        dataset_cls = BurstGPTDataset
+    elif args.dataset_name == "hf":
+        if args.backend != "vllm-chat":
+            raise ValueError(
+                "hf datasets only are supported by vllm-chat backend")
+        # Choose between VisionArenaDataset and HuggingFaceDataset based on
+        # provided parameters.
+        dataset_cls = (VisionArenaDataset if args.dataset_path
+                       == VisionArenaDataset.VISION_ARENA_DATASET_PATH
+                       and args.hf_subset is None else HuggingFaceDataset)
+        common_kwargs['dataset_subset'] = args.hf_subset
+        common_kwargs['dataset_split'] = args.hf_split
+        sample_kwargs["enable_multimodal_chat"] = True
+
+    else:
+        raise ValueError(f"Unknown dataset name: {args.dataset_name}")
+    # Remove None values
+    sample_kwargs = {k: v for k, v in sample_kwargs.items() if v is not None}
+    return dataset_cls(**common_kwargs).sample(**sample_kwargs)
+
+
 def main(args: argparse.Namespace):
+    if args.seed is None:
+        args.seed = 0
     print(args)
     random.seed(args.seed)
-
     # Sample the requests.
     tokenizer = AutoTokenizer.from_pretrained(
         args.tokenizer, trust_remote_code=args.trust_remote_code)
-    if args.dataset is None:
-        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)
-                for _ in range(args.input_len)
-            ]
-            # 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 = request_tokenizer.decode(candidate_ids)
-                tokenized_len = len(request_tokenizer.encode(candidate_prompt))
-
-                if tokenized_len == args.input_len:
-                    break
-
-                # Adjust length based on difference
-                diff = args.input_len - tokenized_len
-                if diff > 0:
-                    candidate_ids.extend([
-                        random.randint(100, vocab_size - 100)
-                        for _ in range(diff)
-                    ])
-                else:
-                    candidate_ids = candidate_ids[:diff]
-            requests.append(
-                SampleRequest(prompt=candidate_prompt,
-                              prompt_len=args.input_len,
-                              expected_output_len=args.output_len,
-                              lora_request=lora_request))
-    else:
-        requests = sample_requests(tokenizer, args)
-
+    requests = get_requests(args, tokenizer)
     is_multi_modal = any(request.multi_modal_data is not None
                          for request in requests)
+    request_outputs: Optional[list[RequestOutput]] = None
     if args.backend == "vllm":
         if args.async_engine:
             elapsed_time = uvloop.run(
@@ -397,31 +356,59 @@ def main(args: argparse.Namespace):
                     args.n,
                     AsyncEngineArgs.from_cli_args(args),
                     args.disable_frontend_multiprocessing,
+                    args.disable_detokenize,
                 ))
         else:
-            elapsed_time = run_vllm(requests, args.n,
-                                    EngineArgs.from_cli_args(args))
+            elapsed_time, request_outputs = run_vllm(
+                requests, args.n, EngineArgs.from_cli_args(args),
+                args.disable_detokenize)
     elif args.backend == "hf":
         assert args.tensor_parallel_size == 1
         elapsed_time = run_hf(requests, args.model, tokenizer, args.n,
-                              args.hf_max_batch_size, args.trust_remote_code)
+                              args.hf_max_batch_size, args.trust_remote_code,
+                              args.disable_detokenize)
     elif args.backend == "mii":
         elapsed_time = run_mii(requests, args.model, args.tensor_parallel_size,
                                args.output_len)
+    elif args.backend == "vllm-chat":
+        elapsed_time, request_outputs = run_vllm_chat(
+            requests, args.n, EngineArgs.from_cli_args(args),
+            args.disable_detokenize)
     else:
         raise ValueError(f"Unknown backend: {args.backend}")
-    total_num_tokens = sum(request.prompt_len + request.expected_output_len
-                           for request in requests)
-    total_output_tokens = sum(request.expected_output_len
-                              for request in requests)
-    if is_multi_modal:
-        print("\033[91mWARNING\033[0m: Multi-modal request detected. The "
+
+    if request_outputs:
+        # Note: with the vllm and vllm-chat backends,
+        # we have request_outputs, which we use to count tokens.
+        total_prompt_tokens = 0
+        total_output_tokens = 0
+        for ro in request_outputs:
+            if not isinstance(ro, RequestOutput):
+                continue
+            total_prompt_tokens += len(
+                ro.prompt_token_ids) if ro.prompt_token_ids else 0
+            total_output_tokens += sum(
+                len(o.token_ids) for o in ro.outputs if o)
+        total_num_tokens = total_prompt_tokens + total_output_tokens
+    else:
+        total_num_tokens = sum(r.prompt_len + r.expected_output_len
+                               for r in requests)
+        total_output_tokens = sum(r.expected_output_len for r in requests)
+        total_prompt_tokens = total_num_tokens - total_output_tokens
+
+    if is_multi_modal and args.backend != "vllm-chat":
+        print("\033[91mWARNING\033[0m: Multi-modal request with "
+              f"{args.backend} backend detected. The "
               "following metrics are not accurate because image tokens are not"
               " counted. See vllm-project/vllm/issues/9778 for details.")
-        # TODO(vllm-project/vllm/issues/9778): Count molti-modal token length.
+        # TODO(vllm-project/vllm/issues/9778): Count multi-modal token length.
+        # vllm-chat backend counts the image tokens now
+
     print(f"Throughput: {len(requests) / elapsed_time:.2f} requests/s, "
           f"{total_num_tokens / elapsed_time:.2f} total tokens/s, "
           f"{total_output_tokens / elapsed_time:.2f} output tokens/s")
+    print(f"Total num prompt tokens:  {total_prompt_tokens}")
+    print(f"Total num output tokens:  {total_output_tokens}")
 
     # Output JSON results if specified
     if args.output_json:
@@ -434,20 +421,115 @@ 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)
+
+
+def validate_args(args):
+    """
+    Validate command-line arguments.
+    """
+
+    # === Deprecation and Defaulting ===
+    if args.dataset is not None:
+        warnings.warn(
+            "The '--dataset' argument will be deprecated in the next release. "
+            "Please use '--dataset-name' and '--dataset-path' instead.",
+            stacklevel=2)
+        args.dataset_path = args.dataset
+
+    if not getattr(args, "tokenizer", None):
+        args.tokenizer = args.model
+
+    # === Backend Validation ===
+    valid_backends = {"vllm", "hf", "mii", "vllm-chat"}
+    if args.backend not in valid_backends:
+        raise ValueError(f"Unsupported backend: {args.backend}")
+
+    # === Dataset Configuration ===
+    if not args.dataset and not args.dataset_path:
+        print(
+            "When dataset path is not set, it will default to random dataset")
+        args.dataset_name = 'random'
+        if args.input_len is None:
+            raise ValueError("input_len must be provided for a random dataset")
+
+    # === Dataset Name Specific Checks ===
+    # --hf-subset and --hf-split: only used
+    # when dataset_name is 'hf'
+    if args.dataset_name != "hf" and (
+            getattr(args, "hf_subset", None) is not None
+            or getattr(args, "hf_split", None) is not None):
+        warnings.warn("--hf-subset and --hf-split will be ignored \
+                since --dataset-name is not 'hf'.",
+                      stacklevel=2)
+    elif args.dataset_name == "hf" and args.backend != "vllm-chat":
+        raise ValueError(
+            "When --dataset-name is 'hf', backend must be 'vllm-chat'")
+
+    # --random-range-ratio: only used when dataset_name is 'random'
+    if args.dataset_name != 'random' and args.random_range_ratio is not None:
+        warnings.warn("--random-range-ratio will be ignored since \
+                --dataset-name is not 'random'.",
+                      stacklevel=2)
+
+    # --prefix-len: only used when dataset_name is 'random', 'sonnet', or not
+    # set.
+    if args.dataset_name not in {"random", "sonnet", None
+                                 } and args.prefix_len is not None:
+        warnings.warn("--prefix-len will be ignored since --dataset-name\
+                 is not 'random', 'sonnet', or not set.",
+                      stacklevel=2)
+
+    # === LoRA Settings ===
+    if getattr(args, "enable_lora", False) and args.backend != "vllm":
+        raise ValueError(
+            "LoRA benchmarking is only supported for vLLM backend")
+    if getattr(args, "enable_lora", False) and args.lora_path is None:
+        raise ValueError("LoRA path must be provided when enable_lora is True")
+
+    # === Backend-specific Validations ===
+    if args.backend == "hf" and args.hf_max_batch_size is None:
+        raise ValueError("HF max batch size is required for HF backend")
+    if args.backend != "hf" and args.hf_max_batch_size is not None:
+        raise ValueError("HF max batch size is only for HF backend.")
+
+    if args.backend in {"hf", "mii"} and getattr(args, "quantization",
+                                                 None) is not None:
+        raise ValueError("Quantization is only for vLLM backend.")
+
+    if args.backend == "mii" and args.dtype != "auto":
+        raise ValueError("dtype must be auto for MII backend.")
+    if args.backend == "mii" and args.n != 1:
+        raise ValueError("n must be 1 for MII backend.")
+    if args.backend == "mii" and args.tokenizer != args.model:
+        raise ValueError(
+            "Tokenizer must be the same as the model for MII backend.")
 
 
 if __name__ == "__main__":
     parser = FlexibleArgumentParser(description="Benchmark the throughput.")
     parser.add_argument("--backend",
                         type=str,
-                        choices=["vllm", "hf", "mii"],
+                        choices=["vllm", "hf", "mii", "vllm-chat"],
                         default="vllm")
-    parser.add_argument("--dataset",
+    parser.add_argument(
+        "--dataset-name",
+        type=str,
+        choices=["sharegpt", "random", "sonnet", "burstgpt", "hf"],
+        help="Name of the dataset to benchmark on.",
+        default="sharegpt")
+    parser.add_argument(
+        "--dataset",
+        type=str,
+        default=None,
+        help="Path to the ShareGPT dataset, will be deprecated in\
+            the next release. The dataset is expected to "
+        "be a json in form of list[dict[..., conversations: "
+        "list[dict[..., value: <prompt_or_response>]]]]")
+    parser.add_argument("--dataset-path",
                         type=str,
                         default=None,
-                        help="Path to the dataset. The dataset is expected to "
-                        "be a json in form of List[Dict[..., conversations: "
-                        "List[Dict[..., value: <prompt_or_response>]]]]")
+                        help="Path to the dataset")
     parser.add_argument("--input-len",
                         type=int,
                         default=None,
@@ -482,6 +564,11 @@ if __name__ == "__main__":
                         action='store_true',
                         default=False,
                         help="Disable decoupled async engine frontend.")
+    parser.add_argument(
+        "--disable-detokenize",
+        action="store_true",
+        help=("Do not detokenize the response (i.e. do not include "
+              "detokenization time in the measurement)"))
     # LoRA
     parser.add_argument(
         "--lora-path",
@@ -489,43 +576,33 @@ if __name__ == "__main__":
         default=None,
         help="Path to the lora adapters to use. This can be an absolute path, "
         "a relative path, or a Hugging Face model identifier.")
+    parser.add_argument("--prefix-len",
+                        type=int,
+                        default=None,
+                        help="Number of prefix tokens per request."
+                        "This is for the RandomDataset and SonnetDataset")
+    # random dataset
+    parser.add_argument(
+        "--random-range-ratio",
+        type=float,
+        default=None,
+        help="Range of sampled ratio of input/output length, "
+        "used only for RandomDataSet.",
+    )
+
+    # hf dtaset
+    parser.add_argument("--hf-subset",
+                        type=str,
+                        default=None,
+                        help="Subset of the HF dataset.")
+    parser.add_argument("--hf-split",
+                        type=str,
+                        default=None,
+                        help="Split of the HF dataset.")
 
     parser = AsyncEngineArgs.add_cli_args(parser)
     args = parser.parse_args()
     if args.tokenizer is None:
         args.tokenizer = args.model
-    if args.dataset is None:
-        assert args.input_len is not None
-        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:
-            raise ValueError("HF max batch size is only for HF backend.")
-    elif args.backend == "hf":
-        if args.hf_max_batch_size is None:
-            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.")
-        if args.n != 1:
-            raise ValueError("n must be 1 for MII backend.")
-        if args.quantization is not None:
-            raise ValueError("Quantization is only for vLLM backend.")
-        if args.hf_max_batch_size is not None:
-            raise ValueError("HF max batch size is only for HF backend.")
-        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")
+    validate_args(args)
     main(args)

benchmarks/benchmark_utils.py

@@ -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)

benchmarks/cutlass_benchmarks/sparse_benchmarks.py

@@ -1,9 +1,12 @@
+# 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
 
 import torch
 import torch.utils.benchmark as TBenchmark
@@ -226,7 +229,7 @@ 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]]) -> Iterable[TMeasurement]:
     results = []
     for m, k, n in MKNs:
         timers = bench(dtype, m, k, n, f"scaled-{dtype}-gemm",
@@ -239,7 +242,7 @@ def run(dtype: torch.dtype,
 
 # 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} ====")
@@ -280,7 +283,7 @@ 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

benchmarks/cutlass_benchmarks/utils.py

@@ -1,5 +1,7 @@
+# SPDX-License-Identifier: Apache-2.0
+
 # Cutlass bench utils
-from typing import Iterable, Tuple
+from collections.abc import Iterable
 
 import torch
 
@@ -25,7 +27,7 @@ def to_fp16(tensor: torch.Tensor) -> torch.Tensor:
 
 
 def make_rand_tensors(dtype: torch.dtype, m: int, n: int,
-                      k: int) -> Tuple[torch.Tensor, torch.Tensor]:
+                      k: int) -> tuple[torch.Tensor, torch.Tensor]:
     a = torch.randn((m, k), device='cuda') * 5
     b = torch.randn((n, k), device='cuda').t() * 5
 
@@ -61,7 +63,7 @@ def prune_to_2_4(tensor):
 
 
 def make_rand_sparse_tensors(dtype: torch.dtype, m: int, n: int,
-                             k: int) -> Tuple[torch.Tensor, torch.Tensor]:
+                             k: int) -> tuple[torch.Tensor, torch.Tensor]:
     a = torch.randn((m, k), device='cuda') * 5
     b = torch.randn((n, k), device='cuda').t() * 5
 
@@ -86,7 +88,7 @@ def make_rand_sparse_tensors(dtype: torch.dtype, m: int, n: int,
 
 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]]:
+                        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)

benchmarks/cutlass_benchmarks/w8a8_benchmarks.py

@@ -1,9 +1,12 @@
+# 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
@@ -12,6 +15,8 @@ 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())
@@ -38,8 +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)
@@ -48,155 +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")
 
 
@@ -207,20 +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} ====")
@@ -232,15 +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}")
 
 
@@ -251,8 +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}")
 
 
@@ -261,7 +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
@@ -278,7 +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
@@ -328,6 +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")
@@ -362,4 +359,4 @@ Benchmark Cutlass GEMM.
     model_parser.set_defaults(func=run_model_bench)
 
     args = parser.parse_args()
-    args.func(args)
+    args.func(args)

benchmarks/cutlass_benchmarks/weight_shapes.py

@@ -1,3 +1,5 @@
+# SPDX-License-Identifier: Apache-2.0
+
 # Weight Shapes are in the format
 # ([K, N], TP_SPLIT_DIM)
 # Example:

benchmarks/disagg_benchmarks/disagg_prefill_proxy_server.py

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

benchmarks/disagg_benchmarks/round_robin_proxy.py

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

benchmarks/disagg_benchmarks/visualize_benchmark_results.py

@@ -1,3 +1,5 @@
+# SPDX-License-Identifier: Apache-2.0
+
 import json
 
 import matplotlib.pyplot as plt

benchmarks/fused_kernels/layernorm_rms_benchmarks.py

@@ -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))

benchmarks/kernels/benchmark_aqlm.py

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

benchmarks/kernels/benchmark_layernorm.py

@@ -1,3 +1,5 @@
+# SPDX-License-Identifier: Apache-2.0
+
 import time
 
 import torch
@@ -38,7 +40,7 @@ def main(num_tokens: int,
 
         end_time = time.perf_counter()
         if profile:
-            torch.cuda.cudart().cudaProfilerStart()
+            torch.cuda.cudart().cudaProfilerStop()
         return (end_time - start_time) / num_iters
 
     # Warmup.

benchmarks/kernels/benchmark_lora.py

@@ -1,3 +1,5 @@
+# SPDX-License-Identifier: Apache-2.0
+
 import argparse
 import copy
 import json
@@ -7,7 +9,7 @@ from dataclasses import dataclass
 from enum import Enum, auto
 from itertools import product
 from pathlib import Path
-from typing import Any, Callable, Dict, List, Optional, Tuple
+from typing import Any, Callable, Optional
 
 import torch
 import torch.utils.benchmark as TBenchmark
@@ -15,11 +17,7 @@ from torch.utils.benchmark import Measurement as TMeasurement
 from utils import ArgPool, Bench, CudaGraphBenchParams
 from weight_shapes import WEIGHT_SHAPES
 
-from vllm.lora.ops.triton_ops.bgmv_expand import bgmv_expand
-from vllm.lora.ops.triton_ops.bgmv_expand_slice import bgmv_expand_slice
-from vllm.lora.ops.triton_ops.bgmv_shrink import bgmv_shrink
-from vllm.lora.ops.triton_ops.sgmv_expand import sgmv_expand
-from vllm.lora.ops.triton_ops.sgmv_shrink import sgmv_shrink
+from vllm.lora.ops.triton_ops import LoRAKernelMeta, lora_expand, lora_shrink
 from vllm.lora.ops.triton_ops.utils import _LORA_A_PTR_DICT, _LORA_B_PTR_DICT
 from vllm.utils import FlexibleArgumentParser
 
@@ -59,15 +57,15 @@ def make_rand_lora_weight_tensor(k: int,
 
 
 def make_rand_tensors(
-    a_shape: Tuple[int],
-    b_shape: Tuple[int],
-    c_shape: Tuple[int],
+    a_shape: tuple[int],
+    b_shape: tuple[int],
+    c_shape: tuple[int],
     a_dtype: torch.dtype,
     b_dtype: torch.dtype,
     c_dtype: torch.dtype,
     num_slices: int,
     device: str = "cuda",
-) -> Tuple[torch.Tensor, List[torch.Tensor], torch.Tensor]:
+) -> tuple[torch.Tensor, list[torch.Tensor], torch.Tensor]:
     """
     Make LoRA input/output matrices.
     """
@@ -87,7 +85,7 @@ def make_prompt_lora_mapping(num_prompts: int, num_active_loras: int,
                              sort_by_lora_id: bool,
                              device: str) -> torch.Tensor:
     """
-    All prompts are mapped to a Lora ID in range [0, num_active_loras).
+    All prompts are mapped to a LoRA ID in range [0, num_active_loras).
     where 0 refers to first lora, 1 refers to second lora and so on.
     """
     assert num_active_loras > 0
@@ -133,7 +131,7 @@ def make_token_lora_mapping(num_tokens: int, num_prompts: int,
 
 
 def ref_group_gemm(ref_out: torch.Tensor, input: torch.Tensor,
-                   lora_weights: List[torch.Tensor],
+                   lora_weights: list[torch.Tensor],
                    seq_lens_cpu: torch.Tensor,
                    prompt_lora_mapping_cpu: torch.Tensor, scaling: float,
                    add_inputs: Optional[bool]):
@@ -151,7 +149,6 @@ def ref_group_gemm(ref_out: torch.Tensor, input: torch.Tensor,
         result = torch.nn.functional.linear(x, w)
         result *= scaling
         out_list.append(result)
-    torch.cat(out_list, dim=0)
 
     cat_result = torch.cat(out_list, dim=0)
 
@@ -165,62 +162,35 @@ class OpType(Enum):
     """
     LoRA Ops to benchmark and its properties.
     """
-    SGMV_SHRINK = auto()
-    BGMV_SHRINK = auto()
-    SGMV_EXPAND = auto()
-    BGMV_EXPAND = auto()
-    BGMV_EXPAND_SLICE = auto()
+    LORA_SHRINK = auto()
+    LORA_EXPAND = auto()
 
     @staticmethod
     def from_str(s: str) -> "OpType":
-        if s.lower() == 'sgmv_shrink':
-            return OpType.SGMV_SHRINK
-        if s.lower() == 'sgmv_expand':
-            return OpType.SGMV_EXPAND
-        if s.lower() == 'bgmv_shrink':
-            return OpType.BGMV_SHRINK
-        if s.lower() == 'bgmv_expand':
-            return OpType.BGMV_EXPAND
-        if s.lower() == "bgmv_expand_slice":
-            return OpType.BGMV_EXPAND_SLICE
+        if s.lower() == "lora_shrink":
+            return OpType.LORA_SHRINK
+        if s.lower() == "lora_expand":
+            return OpType.LORA_EXPAND
         raise ValueError(f"Unrecognized str {s} to convert to OpType")
 
     def is_shrink_fn(self) -> bool:
-        return self in [OpType.SGMV_SHRINK, OpType.BGMV_SHRINK]
+        return self in [OpType.LORA_SHRINK]
 
     def is_expand_fn(self) -> bool:
-        return self in [OpType.SGMV_EXPAND, OpType.BGMV_EXPAND]
+        return self in [OpType.LORA_EXPAND]
 
-    def is_prefill_op(self) -> bool:
-        return self in [OpType.SGMV_SHRINK, OpType.SGMV_EXPAND]
-
-    def is_decode_op(self) -> bool:
-        return self in [
-            OpType.BGMV_SHRINK, OpType.BGMV_EXPAND, OpType.BGMV_EXPAND_SLICE
-        ]
-
-    def is_expand_slice_fn(self) -> bool:
-        return self in [OpType.BGMV_EXPAND_SLICE]
-
-    def num_slices(self) -> List[int]:
-        if self in [OpType.SGMV_EXPAND, OpType.SGMV_SHRINK]:
-            # SGMV kernels supports slices
-            return [1, 2, 3]
-        if self in [OpType.BGMV_SHRINK, OpType.BGMV_EXPAND]:
-            return [1]
-        if self in [OpType.BGMV_EXPAND_SLICE]:
-            return [2, 3]
-        raise ValueError(f"Unrecognized OpType {self}")
+    def num_slices(self) -> list[int]:
+        return [1, 2, 3]
 
     def mkn(self, batch_size: int, seq_length: int, hidden_size: int,
-            lora_rank: int) -> Tuple[int, int, int]:
+            lora_rank: int) -> tuple[int, int, int]:
         num_tokens = batch_size * seq_length
         if self.is_shrink_fn():
             m = num_tokens
             k = hidden_size
             n = lora_rank
         else:
-            assert self.is_expand_fn() or self.is_expand_slice_fn()
+            assert self.is_expand_fn()
             m = num_tokens
             k = lora_rank
             n = hidden_size
@@ -228,20 +198,20 @@ class OpType(Enum):
 
     def matmul_dtypes(
             self, op_dtype: torch.dtype
-    ) -> Tuple[torch.dtype, torch.dtype, torch.dtype]:
+    ) -> tuple[torch.dtype, torch.dtype, torch.dtype]:
         """
         return a type, b type and c type for A x B = C
         """
         if self.is_shrink_fn():
             return op_dtype, op_dtype, torch.float32
         else:
-            assert self.is_expand_fn() or self.is_expand_slice_fn()
+            assert self.is_expand_fn()
             return torch.float32, op_dtype, op_dtype
 
     def matmul_shapes(
             self, batch_size: int, seq_length: int, hidden_size: int,
             lora_rank: int, num_loras: int,
-            num_slices: int) -> Tuple[Tuple[int], Tuple[int], Tuple[int]]:
+            num_slices: int) -> tuple[tuple[int], tuple[int], tuple[int]]:
         """
         Given num_slices, return the shapes of the A, B, and C matrices
         in A x B = C, for the op_type
@@ -249,56 +219,39 @@ class OpType(Enum):
         m, k, n = self.mkn(batch_size, seq_length, hidden_size, lora_rank)
 
         b_shape = (num_loras, n, k)  # col-major
-        if self == OpType.SGMV_SHRINK:
-            # SGMV shrink supports num_slices inherently in the kernel
+        if self in [OpType.LORA_SHRINK]:
+            # LoRA shrink kernels support num_slices inherently in the kernel.
             return ((m, k), b_shape, (num_slices, m, n))
-        if self == OpType.SGMV_EXPAND:
-            # SGMV expand supports num_slices inherently in the kernel
+        if self in [OpType.LORA_EXPAND]:
+            # LoRA expand kernels support num_slices inherently in the kernel
             return ((num_slices, m, k), b_shape, (m, n * num_slices))
-        if self == OpType.BGMV_SHRINK:
-            return ((m, k), b_shape, (m, n))
-        if self == OpType.BGMV_EXPAND:
-            return ((m, k), b_shape, (m, n))
-        if self == OpType.BGMV_EXPAND_SLICE:
-            return ((num_slices, m, k), b_shape, (m, n * num_slices))
-
         raise ValueError(f"Unrecognized op_type {self}")
 
     def bench_fn(self) -> Callable:
+        if self == OpType.LORA_SHRINK:
+            return lora_shrink
+        if self == OpType.LORA_EXPAND:
+            return lora_expand
 
-        def emulate_bgmv_expand_slice(kwargs_list: List[Dict[str, Any]]):
-            for x in kwargs_list:
-                bgmv_expand_slice(**x)
-
-        if self == OpType.SGMV_SHRINK:
-            return sgmv_shrink
-        if self == OpType.SGMV_EXPAND:
-            return sgmv_expand
-        if self == OpType.BGMV_SHRINK:
-            return bgmv_shrink
-        if self == OpType.BGMV_EXPAND:
-            return bgmv_expand
-        if self == OpType.BGMV_EXPAND_SLICE:
-            return emulate_bgmv_expand_slice
         raise ValueError(f"Unrecognized optype {self}")
 
     def run_ref_group_gemm(self, output: torch.Tensor, input: torch.Tensor,
-                           lora_weights: List[torch.Tensor],
+                           lora_weights: list[torch.Tensor],
                            **kwargs) -> Callable:
-        """Each benchmark operation expected the input, lora_weights and outputs
+        """Each benchmark operation expects the input, lora_weights and outputs
            in a slightly different format. Refer to self.matmul_shapes().
            run_ref_group_gemm accounts for those differences in executing a
            reference group gemm for correctness testing.
         """
         w_dtype = lora_weights[0].dtype
         num_slices = len(lora_weights)
-        if self == OpType.SGMV_SHRINK:
+        if self in [OpType.LORA_SHRINK]:
             for slice_idx in range(num_slices):
                 ref_group_gemm(ref_out=output[slice_idx, :],
                                input=input,
                                lora_weights=lora_weights[slice_idx],
                                **kwargs)
-        if self == OpType.SGMV_EXPAND:
+        elif self in [OpType.LORA_EXPAND]:
             hidden_size = lora_weights[0].shape[1]
             for slice_idx in range(num_slices):
                 slice_offset = slice_idx * hidden_size
@@ -307,28 +260,8 @@ class OpType(Enum):
                     input=input[slice_idx].clone().to(dtype=w_dtype),
                     lora_weights=lora_weights[slice_idx],
                     **kwargs)
-        if self == OpType.BGMV_SHRINK:
-            assert num_slices == 1
-            ref_group_gemm(ref_out=output,
-                           input=input,
-                           lora_weights=lora_weights[0],
-                           **kwargs)
-        if self == OpType.BGMV_EXPAND:
-            assert num_slices == 1
-            ref_group_gemm(ref_out=output,
-                           input=input.clone().to(dtype=w_dtype),
-                           lora_weights=lora_weights[0],
-                           **kwargs)
-        if self == OpType.BGMV_EXPAND_SLICE:
-            hidden_size = lora_weights[0].shape[1]
-            for slice_idx in range(num_slices):
-                slice_offset = slice_idx * hidden_size
-                ref_group_gemm(
-                    ref_out=output[:, slice_offset:slice_offset + hidden_size],
-                    input=input[slice_idx].clone().to(dtype=w_dtype),
-                    lora_weights=lora_weights[slice_idx],
-                    **kwargs)
-        raise ValueError(f"Unrecognized optype {self}")
+        else:
+            raise ValueError(f"Unrecognized optype {self}")
 
 
 @dataclass
@@ -382,13 +315,13 @@ class BenchmarkTensors:
     """
     # matmul tensors
     input: torch.Tensor
-    lora_weights_lst: List[torch.Tensor]
+    lora_weights_lst: list[torch.Tensor]
     output: torch.Tensor
-    # metadata tensors
+    # LoRA kernel metadata
+    lora_kernel_meta: LoRAKernelMeta
+    # Metadata tensors used in testing correctness
     seq_lens: torch.Tensor
-    seq_start_loc: torch.Tensor
     prompt_lora_mapping: torch.Tensor
-    token_lora_mapping: torch.Tensor
 
     def io_types(self) -> str:
         return (f"{dtype_to_str(self.input.dtype)}x"
@@ -415,26 +348,29 @@ class BenchmarkTensors:
         assert ctx.num_active_loras <= ctx.num_loras
         total_tokens = ctx.batch_size * ctx.seq_length
 
+        # Make metadata tensors involved in correctness testing.
         # Prepare seq lens tensor
         seq_len_tensor = torch.randint(ctx.seq_length, ctx.seq_length + 1,
                                        (ctx.batch_size, ))
-        # Prepare seq_start_loc tensor
-        seq_start_loc_tensor = torch.cumsum(torch.tensor(
-            [0] + seq_len_tensor[:-1].tolist(), dtype=torch.long),
-                                            dim=0)
         assert total_tokens == seq_len_tensor.sum()
         # Prepare prompt lora indices tensor
         prompt_lora_indices_tensor = make_prompt_lora_mapping(
             ctx.batch_size, ctx.num_active_loras, ctx.sort_by_lora_id, "cpu")
-        # Prepare token lora indices tensor
+
+        # Make LoRAKernelMeta
         token_lora_indices_tensor = make_token_lora_mapping(
             total_tokens, ctx.batch_size, prompt_lora_indices_tensor,
             seq_len_tensor, "cpu")
+        lora_kernel_meta = LoRAKernelMeta.make(
+            max_loras=ctx.num_loras,
+            max_num_tokens=token_lora_indices_tensor.size(0),
+            device="cpu")
+        lora_kernel_meta.prepare_tensors(
+            token_lora_mapping=token_lora_indices_tensor)
 
         return BenchmarkTensors(input_tensor, lora_weights, output_tensor,
-                                seq_len_tensor, seq_start_loc_tensor,
-                                prompt_lora_indices_tensor,
-                                token_lora_indices_tensor)
+                                lora_kernel_meta, seq_len_tensor,
+                                prompt_lora_indices_tensor)
 
     def sanity_check(self) -> None:
         """
@@ -444,9 +380,9 @@ class BenchmarkTensors:
         # check metadata tensors
         assert torch.sum(self.seq_lens) == num_tokens
         num_seqs = self.seq_lens.shape[0]
-        assert self.seq_start_loc.shape[0] == num_seqs
+        #assert self.seq_start_loc.shape[0] == num_seqs
         assert self.prompt_lora_mapping.shape[0] == num_seqs
-        assert self.token_lora_mapping.shape[0] == num_tokens
+        assert self.lora_kernel_meta.token_lora_mapping.shape[0] == num_tokens
 
     def to_device(self, device: str):
         """
@@ -461,54 +397,31 @@ class BenchmarkTensors:
         self.input = to_device(self.input)
         self.output = to_device(self.output)
         self.seq_lens = to_device(self.seq_lens)
-        self.seq_start_loc = to_device(self.seq_start_loc)
         self.prompt_lora_mapping = to_device(self.prompt_lora_mapping)
-        self.token_lora_mapping = to_device(self.token_lora_mapping)
         for i in range(len(self.lora_weights_lst)):
             self.lora_weights_lst[i] = to_device(self.lora_weights_lst[i])
 
-    def metadata(self) -> Tuple[int, int, int]:
+        # LoRA meta
+        for field_name in LoRAKernelMeta.__dataclass_fields__:
+            field = getattr(self.lora_kernel_meta, field_name)
+            assert isinstance(field, torch.Tensor)
+            setattr(self.lora_kernel_meta, field_name, to_device(field))
+
+    def metadata(self) -> tuple[int, int, int]:
         """
         Return num_seqs, num_tokens and max_seq_len
         """
         num_seqs = self.seq_lens.shape[0]
-        num_tokens = self.token_lora_mapping.shape[0]
+        num_tokens = self.lora_kernel_meta.token_lora_mapping.shape[0]
         max_seq_len = torch.max(self.seq_lens).item()
         num_slices = len(self.lora_weights_lst)
         return num_seqs, num_tokens, max_seq_len, num_slices
 
-    def convert_to_sgmv_benchmark_tensors(self):
-        """
-        For sgmv punica kernels, when consecutive sequences have the
-        same LoRA ID, we just merge them together.
-        This happens in punica.py::compute_metadata
-        """
-
-        # Collapse seq_lens and seq_start_loc
-        _, seq_lens = torch.unique_consecutive(self.token_lora_mapping,
-                                               return_counts=True)
-        cum_result = torch.cumsum(seq_lens, dim=0)
-        seq_start_loc = torch.zeros_like(seq_lens)
-        seq_start_loc[1:].copy_(cum_result[:-1])
-
-        # Collapse prompt mapping
-        prompt_lora_mapping = torch.unique_consecutive(
-            self.prompt_lora_mapping)
-
-        assert torch.sum(seq_lens) == torch.sum(self.seq_lens), \
-         f"dont match - new {torch.sum(seq_lens)} vs {torch.sum(self.seq_lens)}"
-
-        self.prompt_lora_mapping = prompt_lora_mapping.to(
-            dtype=self.prompt_lora_mapping.dtype)
-        self.seq_lens = seq_lens.to(dtype=self.seq_lens.dtype)
-        self.seq_start_loc = seq_start_loc.to(dtype=self.seq_start_loc.dtype)
-
-    def as_sgmv_shrink_kwargs(self) -> Dict[str, Any]:
-        self.convert_to_sgmv_benchmark_tensors()
+    def as_lora_shrink_kwargs(self) -> dict[str, Any]:
         self.sanity_check()
         self.to_device(self.input.device)
 
-        num_seqs, num_tokens, max_seq_len, num_slices = self.metadata()
+        _, num_tokens, _, num_slices = self.metadata()
 
         # Sanity check matrix shapes.
         i_shape, lw_shape, o_shape = self.input.shape, self.lora_weights_lst[
@@ -529,22 +442,20 @@ class BenchmarkTensors:
             'inputs': self.input,
             'lora_a_weights': self.lora_weights_lst,
             'output_tensor': self.output,
-            'b_seq_start_loc': self.seq_start_loc,
-            'seq_len_tensor': self.seq_lens,
-            'lora_indices_tensor': self.prompt_lora_mapping,
-            'batches': num_seqs,
-            'max_seq_length': max_seq_len,
-            'token_nums': num_tokens,
+            'token_lora_mapping': self.lora_kernel_meta.token_lora_mapping,
+            'token_indices_sorted_by_lora_ids':
+            self.lora_kernel_meta.token_indices_sorted_by_lora_ids,
+            'num_tokens_per_lora': self.lora_kernel_meta.num_tokens_per_lora,
+            'lora_token_start_loc': self.lora_kernel_meta.lora_token_start_loc,
+            'lora_ids': self.lora_kernel_meta.active_lora_ids,
             'scaling': 1.0,
         }
 
-    def as_sgmv_expand_kwargs(self, add_inputs: bool) -> Dict[str, Any]:
-
-        self.convert_to_sgmv_benchmark_tensors()
+    def as_lora_expand_kwargs(self, add_inputs: bool) -> dict[str, Any]:
         self.sanity_check()
         self.to_device(self.input.device)
 
-        num_seqs, num_tokens, max_seq_len, num_slices = self.metadata()
+        _, num_tokens, _, num_slices = self.metadata()
 
         # Sanity check matrix shapes.
         i_shape, lw_shape, o_shape = self.input.shape, self.lora_weights_lst[
@@ -566,124 +477,28 @@ class BenchmarkTensors:
             'inputs': self.input,
             'lora_b_weights': self.lora_weights_lst,
             'output_tensor': self.output,
-            'b_seq_start_loc': self.seq_start_loc,
-            'seq_len_tensor': self.seq_lens,
-            'lora_indices_tensor': self.prompt_lora_mapping,
-            'batches': num_seqs,
-            'max_seq_length': max_seq_len,
-            'token_nums': num_tokens,
+            'token_lora_mapping': self.lora_kernel_meta.token_lora_mapping,
+            'token_indices_sorted_by_lora_ids':
+            self.lora_kernel_meta.token_indices_sorted_by_lora_ids,
+            'num_tokens_per_lora': self.lora_kernel_meta.num_tokens_per_lora,
+            'lora_token_start_loc': self.lora_kernel_meta.lora_token_start_loc,
+            'lora_ids': self.lora_kernel_meta.active_lora_ids,
             'offset_start': 0,
             'add_inputs': add_inputs,
         }
 
-    def as_bgmv_shrink_kwargs(self) -> Dict[str, Any]:
-        assert len(self.lora_weights_lst) == 1
-        self.to_device(self.input.device)
-
-        _, num_tokens, _, _ = self.metadata()
-        # Sanity check shapes
-        i_shape, lw_shape, o_shape = self.input.shape, self.lora_weights_lst[
-            0].shape, self.output.shape
-        # Expected input shape [num_tokens, hidden_size]
-        assert len(i_shape) == 2
-        assert i_shape[0] == num_tokens
-        hidden_size = i_shape[1]
-        # Expected lora weight shape [num_loras, lora_rank, hidden_size]
-        assert len(lw_shape) == 3
-        assert lw_shape[2] == hidden_size
-        lora_rank = lw_shape[1]
-        # Expected output shape [num_tokens, lora_rank]
-        assert len(o_shape) == 2
-        assert o_shape == (num_tokens, lora_rank)
-
-        return {
-            'inputs': self.input,
-            'lora_a_weights': self.lora_weights_lst[0],
-            'output_tensor': self.output,
-            'lora_indices_tensor': self.token_lora_mapping,
-            'scaling': 1.0
-        }
-
-    def as_bgmv_expand_kwargs(self, add_inputs: bool):
-        assert len(self.lora_weights_lst) == 1
-        self.to_device(self.input.device)
-
-        _, num_tokens, _, _ = self.metadata()
-        # Sanity check shapes
-        i_shape, lw_shape, o_shape = self.input.shape, self.lora_weights_lst[
-            0].shape, self.output.shape
-        # Expected input shape [num_tokens, lora_rank]
-        assert len(i_shape) == 2
-        assert i_shape[0] == num_tokens
-        lora_rank = i_shape[1]
-        # Expected lora weight shape [num_loras, hidden_size, lora_rank]
-        assert len(lw_shape) == 3
-        assert lw_shape[2] == lora_rank
-        hidden_size = lw_shape[1]
-        # Expected output shape [num_tokens, hidden_size]
-        assert len(o_shape) == 2
-        assert o_shape == (num_tokens, hidden_size)
-
-        return {
-            'inputs': self.input,
-            'lora_b_weights': self.lora_weights_lst[0],
-            'output_tensor': self.output,
-            'lora_indices_tensor': self.token_lora_mapping,
-            'add_inputs': add_inputs
-        }
-
-    def as_bgmv_expand_slice_kwargs(self, add_inputs: bool) -> Dict[str, Any]:
-
-        _, num_tokens, _, num_slices = self.metadata()
-        # Sanity check shapes
-        i_shape, lw_shape, o_shape = self.input.shape, self.lora_weights_lst[
-            0].shape, self.output.shape
-        # Expected input shape [num_slices, num_tokens, lora_rank]
-        assert len(i_shape) == 3
-        assert i_shape[0] == num_slices
-        assert i_shape[1] == num_tokens
-        lora_rank = i_shape[2]
-        # Expected lora weight shape [num_loras, hidden_size, lora_rank]
-        assert len(lw_shape) == 3
-        assert lw_shape[2] == lora_rank
-        hidden_size = lw_shape[1]
-        # Expected output shape [num_tokens, hidden_size * num_slices]
-        assert len(o_shape) == 2
-        assert o_shape == (num_tokens, hidden_size * num_slices)
-
-        self.to_device(self.input.device)
-
-        kwargs_list = []
-        for i in range(num_slices):
-            kwargs_list.append({
-                'inputs': self.input[i],
-                'lora_b_weights': self.lora_weights_lst[i],
-                'output_tensor': self.output,
-                'lora_indices_tensor': self.token_lora_mapping,
-                'slice_offset': i * hidden_size,
-                'slice_size': hidden_size,
-                'add_inputs': add_inputs,
-            })
-        return {'kwargs_list': kwargs_list}
-
     def bench_fn_kwargs(self,
                         op_type: OpType,
-                        add_inputs: Optional[bool] = None) -> Dict[str, Any]:
+                        add_inputs: Optional[bool] = None) -> dict[str, Any]:
         if op_type.is_shrink_fn():
             assert add_inputs is None
         else:
             assert add_inputs is not None
 
-        if op_type == OpType.SGMV_SHRINK:
-            return self.as_sgmv_shrink_kwargs()
-        if op_type == OpType.SGMV_EXPAND:
-            return self.as_sgmv_expand_kwargs(add_inputs)
-        if op_type == OpType.BGMV_SHRINK:
-            return self.as_bgmv_shrink_kwargs()
-        if op_type == OpType.BGMV_EXPAND:
-            return self.as_bgmv_expand_kwargs(add_inputs)
-        if op_type == OpType.BGMV_EXPAND_SLICE:
-            return self.as_bgmv_expand_slice_kwargs(add_inputs)
+        if op_type == OpType.LORA_SHRINK:
+            return self.as_lora_shrink_kwargs()
+        if op_type == OpType.LORA_EXPAND:
+            return self.as_lora_expand_kwargs(add_inputs)
         raise ValueError(f"Unrecognized optype {self}")
 
     def test_correctness(self, op_type: OpType,
@@ -732,7 +547,7 @@ def bench_optype(ctx: BenchmarkContext,
         assert expand_fn_add_inputs is not None
 
     # BenchmarkContext -> BenchmarkTensors
-    bench_tensors : List[BenchmarkTensors] = \
+    bench_tensors : list[BenchmarkTensors] = \
         [BenchmarkTensors.make(ctx, op_type) for _ in range(arg_pool_size)]
     for bt in bench_tensors:
         bt.sanity_check()
@@ -744,7 +559,7 @@ def bench_optype(ctx: BenchmarkContext,
             for bt in bench_tensors
         ])
 
-    # BenchmarkTensors -> Dict (kwargs)
+    # BenchmarkTensors -> dict (kwargs)
     kwargs_list = [
         bt.bench_fn_kwargs(op_type, add_inputs=expand_fn_add_inputs)
         for bt in bench_tensors
@@ -839,7 +654,7 @@ def use_cuda_graph_recommendation() -> str:
             """
 
 
-def print_timers(timers: List[TMeasurement],
+def print_timers(timers: list[TMeasurement],
                  args: Optional[argparse.Namespace] = None):
     compare = TBenchmark.Compare(timers)
     compare.print()
@@ -859,7 +674,7 @@ def print_timers(timers: List[TMeasurement],
           "small num_loras the goal should be to match the torch.mm numbers.")
 
 
-def run(args: argparse.Namespace, bench_ctxs: List[BenchmarkContext]):
+def run(args: argparse.Namespace, bench_ctxs: list[BenchmarkContext]):
 
     if args.cuda_graph_nops is not None:
         assert args.cuda_graph_nops > 0
@@ -871,14 +686,7 @@ def run(args: argparse.Namespace, bench_ctxs: List[BenchmarkContext]):
     timers = []
     for bench_ctx in bench_ctxs:
         for seq_len in args.seq_lengths:
-            bench_ops: List[OpType] = []
-            if seq_len == 1:
-                # bench all decode ops
-                bench_ops = [op for op in args.op_types if op.is_decode_op()]
-            else:
-                # bench all prefill ops
-                bench_ops = [op for op in args.op_types if op.is_prefill_op()]
-
+            bench_ops: list[OpType] = args.op_types
             seq_len_timers = []
             for bench_op in bench_ops:
                 for num_slices in bench_op.num_slices():
@@ -919,10 +727,10 @@ def run(args: argparse.Namespace, bench_ctxs: List[BenchmarkContext]):
             pickle.dump(timers, f)
 
 
-def as_benchmark_contexts(hidden_sizes: List[int], lora_ranks: List[int],
-                          args: argparse.Namespace) -> List[BenchmarkContext]:
+def as_benchmark_contexts(hidden_sizes: list[int], lora_ranks: list[int],
+                          args: argparse.Namespace) -> list[BenchmarkContext]:
 
-    ctxs: List[BenchmarkContext] = []
+    ctxs: list[BenchmarkContext] = []
     for batch_size, hidden_size, lora_rank, num_loras, sort_by_lora_id in product(  # noqa
             args.batch_sizes, list(hidden_sizes), lora_ranks, args.num_loras,
             args.sort_by_lora_id):
@@ -952,7 +760,7 @@ def run_list_bench(args: argparse.Namespace):
           f"  LoRA Ranks {args.lora_ranks}")
 
     # Get all benchmarking contexts
-    bench_contexts: List[BenchmarkContext] = as_benchmark_contexts(
+    bench_contexts: list[BenchmarkContext] = as_benchmark_contexts(
         hidden_sizes=args.hidden_sizes, lora_ranks=args.lora_ranks, args=args)
 
     run(args, bench_contexts)
@@ -973,7 +781,7 @@ def run_range_bench(args: argparse.Namespace):
           f" LoRA Ranks {lora_ranks}")
 
     # Get all benchmarking contexts
-    bench_contexts: List[BenchmarkContext] = as_benchmark_contexts(
+    bench_contexts: list[BenchmarkContext] = as_benchmark_contexts(
         hidden_sizes=hidden_sizes, lora_ranks=lora_ranks, args=args)
 
     run(args, bench_contexts)
@@ -1000,7 +808,7 @@ def run_model_bench(args: argparse.Namespace):
           f" LoRA Ranks {args.lora_ranks}")
 
     # Get all benchmarking contexts
-    bench_contexts: List[BenchmarkContext] = as_benchmark_contexts(
+    bench_contexts: list[BenchmarkContext] = as_benchmark_contexts(
         hidden_sizes=hidden_sizes, lora_ranks=args.lora_ranks, args=args)
 
     run(args, bench_contexts)
@@ -1088,13 +896,13 @@ Benchmark LoRA kernels:
     {use_cuda_graph_recommendation()}
 
     list_bench example:
-        python3 benchmarks/kernels/benchmark_lora.py list_bench --arg-pool-size 32 --batch-sizes 1 16 32 --dtype torch.float16 --hidden-sizes 2048 --lora-ranks 16 --num-loras 1 4 --op-types bgmv_shrink bgmv_expand sgmv_shrink sgmv_expand bgmv_expand_slice --seq-lengths 1 16 --sort-by-lora-id 1 --cuda-graph-nops 32
+        python3 benchmarks/kernels/benchmark_lora.py list_bench --arg-pool-size 32 --batch-sizes 1 16 32 --dtype torch.float16 --hidden-sizes 2048 --lora-ranks 16 --num-loras 1 4 --op-types lora_shrink lora_expand --seq-lengths 1 16 --sort-by-lora-id 1 --cuda-graph-nops 32
 
     model_bench example:
-        python3 benchmarks/kernels/benchmark_lora.py model_bench --models meta-llama/Llama-3-8b  --arg-pool-size 32 --batch-sizes 1 16 32 --dtype torch.float16  --lora-ranks 16 --num-loras 1 4 --op-types bgmv_shrink bgmv_expand sgmv_shrink sgmv_expand bgmv_expand_slice --seq-lengths 1 16 --sort-by-lora-id 1 --cuda-graph-nops 32 
+        python3 benchmarks/kernels/benchmark_lora.py model_bench --models meta-llama/Llama-3-8b  --arg-pool-size 32 --batch-sizes 1 16 32 --dtype torch.float16  --lora-ranks 16 --num-loras 1 4 --op-types lora_shrink lora_expand --seq-lengths 1 16 --sort-by-lora-id 1 --cuda-graph-nops 32 
 
     range_bench example:
-        python3 benchmarks/kernels/benchmark_lora.py range_bench  --arg-pool-size 32 --batch-sizes 1 16 32 --dtype torch.float16   --num-loras 1 4 --op-types bgmv_shrink bgmv_expand sgmv_shrink sgmv_expand bgmv_expand_slice --seq-lengths 1 16 --sort-by-lora-id 1 --cuda-graph-nops 32 --hidden-sizes-start 1024 --hidden-sizes-end 4096 --hidden-sizes-increment 1024 --lora-ranks-start 8 --lora-ranks-end 24 --lora-ranks-increment 8 
+        python3 benchmarks/kernels/benchmark_lora.py range_bench  --arg-pool-size 32 --batch-sizes 1 16 32 --dtype torch.float16   --num-loras 1 4 --op-types lora_shrink lora_expand --seq-lengths 1 16 --sort-by-lora-id 1 --cuda-graph-nops 32 --hidden-sizes-start 1024 --hidden-sizes-end 4096 --hidden-sizes-increment 1024 --lora-ranks-start 8 --lora-ranks-end 24 --lora-ranks-increment 8 
             """,  # noqa: E501
         formatter_class=argparse.RawTextHelpFormatter)
 

benchmarks/kernels/benchmark_machete.py

@@ -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
@@ -42,7 +45,6 @@ def terse_type_name(dt):
         torch.float16: "fp16",
         torch.int8: "int8",
         torch.float8_e4m3fn: "fp8",
-        torch.bfloat16: "bf16",
         torch.float: "float",
         torch.int: "int",
     }[dt]
@@ -100,8 +102,8 @@ 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 +114,7 @@ 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)
 
@@ -256,7 +258,7 @@ def machete_create_bench_fn(bt: BenchmarkTensors,
 
     return lambda: ops.machete_mm(
         a=bt.a,
-        b_q=bt.w_q,
+        b_q=w_q,
         b_type=bt.wtype,
         b_group_scales=bt.w_g_s,
         b_group_zeros=w_g_zp,
@@ -274,7 +276,7 @@ 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 +311,7 @@ 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 +414,12 @@ 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 +431,7 @@ 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 +449,8 @@ 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 +497,7 @@ 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

benchmarks/kernels/benchmark_marlin.py

@@ -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]:

benchmarks/kernels/benchmark_moe.py

@@ -1,8 +1,12 @@
+# SPDX-License-Identifier: Apache-2.0
+
 import argparse
+import json
 import time
+from contextlib import nullcontext
 from datetime import datetime
 from itertools import product
-from typing import Any, Dict, List, Tuple, TypedDict
+from typing import Any, TypedDict
 
 import ray
 import torch
@@ -14,8 +18,7 @@ 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
+FP8_DTYPE = current_platform.fp8_dtype()
 
 
 class BenchmarkConfig(TypedDict):
@@ -38,6 +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)
@@ -79,8 +83,24 @@ 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)
 
@@ -109,6 +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
@@ -130,7 +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()
@@ -173,8 +194,9 @@ def get_rocm_tuning_space(use_fp16):
     return param_ranges
 
 
-def get_configs_compute_bound(use_fp16) -> List[Dict[str, int]]:
-    configs: List[BenchmarkConfig] = []
+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)
@@ -202,17 +224,27 @@ def get_configs_compute_bound(use_fp16) -> List[Dict[str, int]]:
     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):
+                            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 * 2, N1, K1, search_space,
-                                        is_fp16)
-    pruned_space_2 = prune_rocm_configs(num_tokens * 2, N2, K2, search_space,
-                                        is_fp16)
+    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
 
@@ -333,7 +365,8 @@ class BenchmarkWorker:
         dtype: torch.dtype,
         use_fp8_w8a8: bool,
         use_int8_w8a16: bool,
-    ) -> Tuple[Dict[str, int], float]:
+        block_quant_shape: List[int] = None,
+    ) -> tuple[dict[str, int], float]:
         current_platform.seed_everything(self.seed)
         dtype_str = get_config_dtype_str(dtype,
                                          use_int8_w8a16=use_int8_w8a16,
@@ -343,16 +376,27 @@ 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))]
-        kernel_time = benchmark_config(config, num_tokens, num_experts,
-                                       shard_intermediate_size, hidden_size,
-                                       topk, dtype, use_fp8_w8a8,
-                                       use_int8_w8a16)
+        kernel_time = benchmark_config(config,
+                                       num_tokens,
+                                       num_experts,
+                                       shard_intermediate_size,
+                                       hidden_size,
+                                       topk,
+                                       dtype,
+                                       use_fp8_w8a8,
+                                       use_int8_w8a16,
+                                       num_iters=100,
+                                       block_quant_shape=block_quant_shape)
         return config, kernel_time
 
     def tune(
@@ -365,8 +409,9 @@ 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")
         if current_platform.is_rocm():
@@ -374,21 +419,24 @@ class BenchmarkWorker:
             search_space = prune_rocm_search_space(num_tokens,
                                                    shard_intermediate_size,
                                                    hidden_size, search_space,
-                                                   is_fp16)
+                                                   is_fp16, topk)
 
-        with torch.cuda.device(self.device_id):
+        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)
+                    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
@@ -428,10 +476,10 @@ def sort_config(config: BenchmarkConfig) -> BenchmarkConfig:
     }
 
 
-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)
@@ -439,7 +487,7 @@ 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:
@@ -447,10 +495,19 @@ def save_configs(configs: Dict[int, BenchmarkConfig], num_experts: int,
         f.write("\n")
 
 
+def get_weight_block_size_safety(config, default_value=None):
+
+    quantization_config = getattr(config, 'quantization_config', {})
+    if isinstance(quantization_config, dict):
+        return quantization_config.get('weight_block_size', default_value)
+    return default_value
+
+
 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
@@ -461,6 +518,18 @@ 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 = get_weight_block_size_safety(config)
+    elif config.architectures[0] == "Qwen2MoeForCausalLM":
+        E = config.num_experts
+        topk = config.num_experts_per_tok
+        intermediate_size = config.moe_intermediate_size
+        shard_intermediate_size = 2 * intermediate_size // args.tp_size
     else:
         # Default: Mixtral.
         E = config.num_local_experts
@@ -485,7 +554,7 @@ 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:
@@ -498,27 +567,30 @@ def main(args: argparse.Namespace):
 
     if args.tune:
         is_fp16 = not (use_fp8_w8a8 or use_int8_w8a16)
-        search_space = get_configs_compute_bound(is_fp16)
+        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}")
@@ -530,7 +602,11 @@ 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"],
@@ -538,6 +614,7 @@ 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)

benchmarks/kernels/benchmark_paged_attention.py

@@ -1,6 +1,8 @@
+# SPDX-License-Identifier: Apache-2.0
+
 import random
 import time
-from typing import List, Optional
+from typing import Optional
 
 import torch
 
@@ -9,8 +11,9 @@ from vllm.platforms import current_platform
 from vllm.utils import (STR_DTYPE_TO_TORCH_DTYPE, FlexibleArgumentParser,
                         create_kv_caches_with_random)
 
-NUM_BLOCKS = 1024
+NUM_BLOCKS = 128 * 1024
 PARTITION_SIZE = 512
+PARTITION_SIZE_ROCM = 256
 
 
 @torch.inference_mode()
@@ -52,7 +55,7 @@ def main(
 
     # Create the block tables.
     max_num_blocks_per_seq = (max_seq_len + block_size - 1) // block_size
-    block_tables_lst: List[List[int]] = []
+    block_tables_lst: list[list[int]] = []
     for _ in range(num_seqs):
         block_table = [
             random.randint(0, NUM_BLOCKS - 1)
@@ -78,6 +81,12 @@ def main(
     # Prepare for the paged attention kernel.
     output = torch.empty_like(query)
     if version == "v2":
+        if current_platform.is_rocm():
+            global PARTITION_SIZE
+            if not args.custom_paged_attn:
+                PARTITION_SIZE = 1024
+            else:
+                PARTITION_SIZE = PARTITION_SIZE_ROCM
         num_partitions = ((max_seq_len + PARTITION_SIZE - 1) // PARTITION_SIZE)
         tmp_output = torch.empty(
             size=(num_seqs, num_query_heads, num_partitions, head_size),
@@ -121,32 +130,53 @@ def main(
                     v_scale,
                 )
             elif version == "v2":
-                ops.paged_attention_v2(
-                    output,
-                    exp_sums,
-                    max_logits,
-                    tmp_output,
-                    query,
-                    key_cache,
-                    value_cache,
-                    num_kv_heads,
-                    scale,
-                    block_tables,
-                    seq_lens,
-                    block_size,
-                    max_seq_len,
-                    alibi_slopes,
-                    kv_cache_dtype,
-                    k_scale,
-                    v_scale,
-                )
+                if not args.custom_paged_attn:
+                    ops.paged_attention_v2(
+                        output,
+                        exp_sums,
+                        max_logits,
+                        tmp_output,
+                        query,
+                        key_cache,
+                        value_cache,
+                        num_kv_heads,
+                        scale,
+                        block_tables,
+                        seq_lens,
+                        block_size,
+                        max_seq_len,
+                        alibi_slopes,
+                        kv_cache_dtype,
+                        k_scale,
+                        v_scale,
+                    )
+                else:
+                    ops.paged_attention_rocm(
+                        output,
+                        exp_sums,
+                        max_logits,
+                        tmp_output,
+                        query,
+                        key_cache,
+                        value_cache,
+                        num_kv_heads,
+                        scale,
+                        block_tables,
+                        seq_lens,
+                        block_size,
+                        max_seq_len,
+                        alibi_slopes,
+                        kv_cache_dtype,
+                        k_scale,
+                        v_scale,
+                    )
             else:
                 raise ValueError(f"Invalid version: {version}")
         torch.cuda.synchronize()
 
         end_time = time.perf_counter()
         if profile:
-            torch.cuda.cudart().cudaProfilerStart()
+            torch.cuda.cudart().cudaProfilerStop()
         return (end_time - start_time) / num_iters
 
     # Warmup.
@@ -193,6 +223,9 @@ if __name__ == '__main__':
         help="Data type for kv cache storage. If 'auto', will use model "
         "data type. CUDA 11.8+ supports fp8 (=fp8_e4m3) and fp8_e5m2. "
         "ROCm (AMD GPU) supports fp8 (=fp8_e4m3)")
+    parser.add_argument("--custom-paged-attn",
+                        action="store_true",
+                        help="Use custom paged attention")
     args = parser.parse_args()
     print(args)
 

benchmarks/kernels/benchmark_quant.py

@@ -1,3 +1,5 @@
+# SPDX-License-Identifier: Apache-2.0
+
 import time
 
 import torch
@@ -38,7 +40,7 @@ def main(num_tokens: int,
 
         end_time = time.perf_counter()
         if profile:
-            torch.cuda.cudart().cudaProfilerStart()
+            torch.cuda.cudart().cudaProfilerStop()
         return (end_time - start_time) / num_iters
 
     # Warmup.

benchmarks/kernels/benchmark_rmsnorm.py

@@ -1,5 +1,7 @@
+# SPDX-License-Identifier: Apache-2.0
+
 import itertools
-from typing import Optional, Tuple, Union
+from typing import Optional, Union
 
 import torch
 import triton
@@ -20,7 +22,7 @@ class HuggingFaceRMSNorm(nn.Module):
         self,
         x: torch.Tensor,
         residual: Optional[torch.Tensor] = None,
-    ) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
+    ) -> Union[torch.Tensor, tuple[torch.Tensor, torch.Tensor]]:
         orig_dtype = x.dtype
         x = x.to(torch.float32)
         if residual is not None:
@@ -137,7 +139,7 @@ def calculate_diff(batch_size, seq_len, hidden_size, use_residual=True):
 
     print(f"Naive output={output_naive}")
     print(f"FlashInfer output={output_flashinfer}")
-    print(f"VLLM output={output_vllm}")
+    print(f"vLLM output={output_vllm}")
 
     if torch.allclose(output_naive, output_flashinfer, atol=1e-2,
                       rtol=1e-2) and torch.allclose(

benchmarks/kernels/benchmark_rope.py

@@ -1,5 +1,7 @@
+# SPDX-License-Identifier: Apache-2.0
+
 from itertools import accumulate
-from typing import List, Optional
+from typing import Optional
 
 import nvtx
 import torch
@@ -37,7 +39,7 @@ def benchmark_rope_kernels_multi_lora(
                             })
     # non-batched RoPE takes only one scaling factor, we create multiple
     # instances to simulate the same behavior
-    non_batched_ropes: List[RotaryEmbedding] = []
+    non_batched_ropes: list[RotaryEmbedding] = []
     for scaling_factor in scaling_factors:
         non_batched_ropes.append(
             get_rope(head_size, rotary_dim, max_position, base, is_neox_style,

benchmarks/kernels/benchmark_shapes.py

@@ -1,3 +1,5 @@
+# SPDX-License-Identifier: Apache-2.0
+
 WEIGHT_SHAPES = {
     "ideal": [[4 * 256 * 32, 256 * 32]],
     "mistralai/Mistral-7B-v0.1/TP1": [

benchmarks/kernels/benchmark_w8a8_block_fp8.py

@@ -0,0 +1,420 @@
+# SPDX-License-Identifier: Apache-2.0
+# Adapted from sglang quantization/tuning_block_wise_kernel.py
+
+import argparse
+import json
+import multiprocessing as mp
+import os
+import time
+from datetime import datetime
+from typing import Any
+
+import torch
+import tqdm
+import triton
+
+from vllm.model_executor.layers.quantization.utils.fp8_utils import (
+    _w8a8_block_fp8_matmul)
+from vllm.platforms import current_platform
+from vllm.utils import FlexibleArgumentParser
+
+mp.set_start_method("spawn", force=True)
+
+assert current_platform.is_cuda(
+), "Only support tune w8a8 block fp8 kernel on CUDA device."
+
+DTYPE_MAP = {
+    "float32": torch.float32,
+    "float16": torch.float16,
+    "half": torch.half,
+    "bfloat16": torch.bfloat16,
+}
+
+
+def w8a8_block_matmul(
+    A: torch.Tensor,
+    B: torch.Tensor,
+    As: torch.Tensor,
+    Bs: torch.Tensor,
+    block_size: list[int],
+    config: dict[str, Any],
+    output_dtype: torch.dtype = torch.float16,
+) -> torch.Tensor:
+    """This function performs matrix multiplication with 
+    block-wise quantization.
+
+    It takes two input tensors `A` and `B` with scales `As` and `Bs`.
+    The output is returned in the specified `output_dtype`.
+
+    Args:
+        A: The input tensor, e.g., activation.
+        B: The input tensor, e.g., weight.
+        As: The per-token-group quantization scale for `A`.
+        Bs: The per-block quantization scale for `B`.
+        block_size: The block size for per-block quantization. 
+                    It should be 2-dim, e.g., [128, 128].
+        output_dytpe: The dtype of the returned tensor.
+
+    Returns:
+        torch.Tensor: The result of matmul.
+    """
+    assert len(block_size) == 2
+    block_n, block_k = block_size[0], block_size[1]
+
+    assert A.shape[-1] == B.shape[-1]
+    assert A.shape[:-1] == As.shape[:-1] and A.is_contiguous()
+    assert triton.cdiv(A.shape[-1], block_k) == As.shape[-1]
+    M = A.numel() // A.shape[-1]
+
+    assert B.ndim == 2 and B.is_contiguous() and Bs.ndim == 2
+    N, K = B.shape
+    assert triton.cdiv(N, block_n) == Bs.shape[0]
+    assert triton.cdiv(K, block_k) == Bs.shape[1]
+
+    C_shape = A.shape[:-1] + (N, )
+    C = A.new_empty(C_shape, dtype=output_dtype)
+
+    def grid(META):
+        return (triton.cdiv(M, META["BLOCK_SIZE_M"]) *
+                triton.cdiv(N, META["BLOCK_SIZE_N"]), )
+
+    if A.dtype == torch.float8_e4m3fn:
+        kernel = _w8a8_block_fp8_matmul
+    else:
+        raise RuntimeError(
+            "Currently, only support tune w8a8 block fp8 kernel.")
+
+    kernel[grid](
+        A,
+        B,
+        C,
+        As,
+        Bs,
+        M,
+        N,
+        K,
+        block_n,
+        block_k,
+        A.stride(-2),
+        A.stride(-1),
+        B.stride(1),
+        B.stride(0),
+        C.stride(-2),
+        C.stride(-1),
+        As.stride(-2),
+        As.stride(-1),
+        Bs.stride(1),
+        Bs.stride(0),
+        **config,
+    )
+
+    return C
+
+
+def get_configs_compute_bound():
+    configs = []
+    for num_stages in [2, 3, 4, 5]:
+        for block_m in [16, 32, 64, 128, 256]:
+            for block_k in [64, 128]:
+                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,
+                            })
+    return configs
+
+
+def get_weight_shapes(tp_size):
+    # NOTE(HandH1998): The weight shapes only works for DeepSeek-V3.
+    # Modify them, if you tune for another different model.
+    # cannot TP
+    total = [
+        (512 + 64, 7168),
+        ((128 + 64) * 128, 7168),
+        (128 * (128 + 128), 512),
+        (7168, 16384),
+        (7168, 18432),
+    ]
+    # N can TP
+    n_tp = [
+        (18432 * 2, 7168),
+        ((128 + 64) * 128, 7168),
+        (128 * (128 + 128), 512),
+        (24576, 1536),
+        (12288, 7168),
+        (4096, 7168),
+    ]
+    # K can TP
+    k_tp = [(7168, 18432), (7168, 16384), (7168, 2048)]
+
+    weight_shapes = []
+    for t in total:
+        weight_shapes.append(t)
+    for n_t in n_tp:
+        new_t = (n_t[0] // tp_size, n_t[1])
+        weight_shapes.append(new_t)
+    for k_t in k_tp:
+        new_t = (k_t[0], k_t[1] // tp_size)
+        weight_shapes.append(new_t)
+    return weight_shapes
+
+
+def benchmark_config(A,
+                     B,
+                     As,
+                     Bs,
+                     block_size,
+                     config,
+                     out_dtype=torch.float16,
+                     num_iters=10):
+
+    def run():
+        w8a8_block_matmul(A, B, As, Bs, block_size, config, out_dtype)
+
+    torch.cuda.synchronize()
+    # JIT complication & warmup
+    for _ in range(5):
+        run()
+    torch.cuda.synchronize()
+
+    start_event = torch.cuda.Event(enable_timing=True)
+    end_event = torch.cuda.Event(enable_timing=True)
+
+    latencies: list[float] = []
+    for i in range(num_iters):
+        torch.cuda.synchronize()
+        start_event.record()
+        run()
+        end_event.record()
+        end_event.synchronize()
+        latencies.append(start_event.elapsed_time(end_event))
+    avg = sum(latencies) / (num_iters * 10) * 1000  # us
+    return avg
+
+
+def tune(M, N, K, block_size, out_dtype, search_space, input_type):
+    factor_for_scale = 1e-2
+
+    if input_type == "fp8":
+        fp8_info = torch.finfo(torch.float8_e4m3fn)
+        fp8_max, fp8_min = fp8_info.max, fp8_info.min
+
+        A_fp32 = (
+            (torch.rand(M, K, dtype=torch.float32, device="cuda") - 0.5) * 2 *
+            fp8_max)
+        A = A_fp32.clamp(min=fp8_min, max=fp8_max).to(torch.float8_e4m3fn)
+
+        B_fp32 = (
+            (torch.rand(N, K, dtype=torch.float32, device="cuda") - 0.5) * 2 *
+            fp8_max)
+        B = B_fp32.clamp(min=fp8_min, max=fp8_max).to(torch.float8_e4m3fn)
+    else:
+        raise RuntimeError(
+            "Currently, only support tune w8a8 block fp8 kernel.")
+
+    block_n, block_k = block_size[0], block_size[1]
+    n_tiles = (N + block_n - 1) // block_n
+    k_tiles = (K + block_k - 1) // block_k
+
+    As = torch.rand(M, k_tiles, dtype=torch.float32,
+                    device="cuda") * factor_for_scale
+    Bs = (torch.rand(n_tiles, k_tiles, dtype=torch.float32, device="cuda") *
+          factor_for_scale)
+
+    best_config = None
+    best_time = float("inf")
+    for config in tqdm(search_space):
+        try:
+            kernel_time = benchmark_config(
+                A,
+                B,
+                As,
+                Bs,
+                block_size,
+                config,
+                out_dtype,
+                num_iters=10,
+            )
+        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={M}")
+    assert best_config is not None
+    return best_config
+
+
+def save_configs(
+    N,
+    K,
+    block_n,
+    block_k,
+    configs,
+    save_path,
+    input_type="fp8",
+) -> None:
+    os.makedirs(save_path, exist_ok=True)
+    device_name = current_platform.get_device_name().replace(" ", "_")
+    json_file_name = (
+        f"N={N},K={K},device_name={device_name},dtype={input_type}_w8a8,"
+        f"block_shape=[{block_n},{block_k}].json")
+
+    config_file_path = os.path.join(save_path, json_file_name)
+    print(f"Writing best config to {config_file_path}...")
+
+    with open(config_file_path, "w") as f:
+        json.dump(configs, f, indent=4)
+        f.write("\n")
+
+
+def tune_on_gpu(args_dict):
+    """Run tuning on a specific GPU."""
+    gpu_id = args_dict["gpu_id"]
+    batch_sizes = args_dict["batch_sizes"]
+    weight_shapes = args_dict["weight_shapes"]
+    args = args_dict["args"]
+
+    torch.cuda.set_device(gpu_id)
+    print(f"Starting tuning on GPU {gpu_id} with batch sizes {batch_sizes}")
+
+    block_n = args.block_n
+    block_k = args.block_k
+    out_dtype = DTYPE_MAP[args.out_dtype]
+    save_path = args.save_path
+    input_type = args.input_type
+
+    search_space = get_configs_compute_bound()
+    search_space = [
+        config for config in search_space
+        if block_k % config["BLOCK_SIZE_K"] == 0
+    ]
+
+    start = time.time()
+    for shape in tqdm(weight_shapes, desc=f"GPU {gpu_id} - Shapes"):
+        N, K = shape[0], shape[1]
+        print(f"[GPU {gpu_id}] Tune for weight shape of `N: {N}, K: {K}`")
+        benchmark_results = [
+            tune(
+                batch_size,
+                N,
+                K,
+                [block_n, block_k],
+                out_dtype,
+                search_space,
+                input_type,
+            ) for batch_size in tqdm(batch_sizes,
+                                     desc=f"GPU {gpu_id} - Batch sizes")
+        ]
+        best_configs = {
+            M: config
+            for M, config in zip(batch_sizes, benchmark_results)
+        }
+        save_configs(N, K, block_n, block_k, best_configs, save_path,
+                     input_type)
+
+    end = time.time()
+    print(f"Tuning on GPU {gpu_id} took {end - start:.2f} seconds")
+
+
+def distribute_batch_sizes(batch_sizes, num_gpus):
+    """Distribute batch sizes across available GPUs."""
+    batches_per_gpu = []
+    for i in range(num_gpus):
+        start_idx = i * len(batch_sizes) // num_gpus
+        end_idx = (i + 1) * len(batch_sizes) // num_gpus
+        batches_per_gpu.append(batch_sizes[start_idx:end_idx])
+    return batches_per_gpu
+
+
+def main(args):
+    print(args)
+    num_gpus = torch.cuda.device_count()
+    if num_gpus == 0:
+        raise RuntimeError("No GPU available for tuning")
+    print(f"Found {num_gpus} GPUs for parallel tuning")
+
+    torch.cuda.init()
+
+    if args.batch_size is None:
+        batch_sizes = [
+            1,
+            2,
+            4,
+            8,
+            16,
+            24,
+            32,
+            48,
+            64,
+            96,
+            128,
+            256,
+            512,
+            1024,
+            1536,
+            2048,
+            3072,
+            4096,
+        ]
+    else:
+        batch_sizes = [args.batch_size]
+        num_gpus = 1  # If only one batch size, use only one GPU
+
+    weight_shapes = get_weight_shapes(args.tp_size)
+
+    batches_per_gpu = distribute_batch_sizes(batch_sizes, num_gpus)
+
+    process_args = []
+    for gpu_id in range(num_gpus):
+        process_args.append({
+            "gpu_id": gpu_id,
+            "batch_sizes": batches_per_gpu[gpu_id],
+            "weight_shapes":
+            weight_shapes,  # Each GPU processes all weight shapes
+            "args": args,
+        })
+
+    ctx = mp.get_context("spawn")
+    with ctx.Pool(num_gpus) as pool:
+        pool.map(tune_on_gpu, process_args)
+
+    print("Multi-GPU tuning completed")
+
+
+if __name__ == "__main__":
+    parser = FlexibleArgumentParser(
+        description="""
+Tune triton w8a8 block fp8 for DeepSeek-V3/DeepSeek-R1:
+    python3 benchmark_w8a8_block_fp8.py --tp-size 8 --input-type fp8
+Then copy to model_executor/layers/quantization/utils/configs
+        """,
+        formatter_class=argparse.RawTextHelpFormatter)
+
+    parser.add_argument("--tp-size", "-tp", type=int, default=8)
+    parser.add_argument("--input-type",
+                        type=str,
+                        choices=["fp8"],
+                        default="fp8")
+    parser.add_argument(
+        "--out-dtype",
+        type=str,
+        choices=["float32", "float16", "bfloat16", "half"],
+        default="float16",
+    )
+    parser.add_argument("--block-n", type=int, default=128)
+    parser.add_argument("--block-k", type=int, default=128)
+    parser.add_argument("--batch-size", type=int, required=False)
+    parser.add_argument("--save-path", type=str, default="./")
+    args = parser.parse_args()
+
+    main(args)

benchmarks/kernels/deepgemm/README.md

@@ -0,0 +1,129 @@
+# DeepSeek DeepGEMM Kernels Benchmark
+
+This directory includes benchmarks between DeepSeek's DeepGEMM block fp8 kernels against vLLM's existing triton and CUTLASS-based kernels.
+
+Currently this just includes dense GEMMs and only works on Hopper GPUs.
+
+## Setup
+
+You need to install vLLM in your usual fashion, then install DeepGEMM from source in its own directory:
+
+```
+git clone --recursive https://github.com/deepseek-ai/DeepGEMM
+cd DeepGEMM
+python setup.py install
+uv pip install -e .
+```
+
+## Usage
+
+```
+python benchmark_fp8_block_dense_gemm.py
+INFO 02-26 21:55:13 [__init__.py:207] Automatically detected platform cuda.
+===== STARTING FP8 GEMM BENCHMARK =====
+PyTorch version: 2.5.1+cu124
+CUDA version: 12.4
+Triton version: 3.1.0
+Using device: NVIDIA H100 80GB HBM3
+WARNING 02-26 21:55:15 [fp8_utils.py:458] Using default W8A8 Block FP8 kernel config. Performance might be sub-optimal! Config file not found at /home/mgoin/code/vllm/vllm/model_executor/layers/quantization/utils/configs/N=4096,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128,128].json
+INFO 02-26 21:55:15 [fp8_utils.py:449] Using configuration from /home/mgoin/code/vllm/vllm/model_executor/layers/quantization/utils/configs/N=7168,K=18432,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128,128].json for W8A8 Block FP8 kernel.
+WARNING 02-26 21:55:16 [fp8_utils.py:458] Using default W8A8 Block FP8 kernel config. Performance might be sub-optimal! Config file not found at /home/mgoin/code/vllm/vllm/model_executor/layers/quantization/utils/configs/N=18432,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128,128].json
+WARNING 02-26 21:55:17 [fp8_utils.py:458] Using default W8A8 Block FP8 kernel config. Performance might be sub-optimal! Config file not found at /home/mgoin/code/vllm/vllm/model_executor/layers/quantization/utils/configs/N=24576,K=1536,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128,128].json
+INFO 02-26 21:55:17 [fp8_utils.py:449] Using configuration from /home/mgoin/code/vllm/vllm/model_executor/layers/quantization/utils/configs/N=32768,K=512,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128,128].json for W8A8 Block FP8 kernel.
+INFO 02-26 21:55:17 [fp8_utils.py:449] Using configuration from /home/mgoin/code/vllm/vllm/model_executor/layers/quantization/utils/configs/N=7168,K=16384,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128,128].json for W8A8 Block FP8 kernel.
+
+===== PERFORMANCE COMPARISON =====
+
+DeepGEMM Implementation:
++------+-------+-------+-----------+--------+--------+
+| m    | n     | k     | Time (μs) | TFLOPS | GB/s   |
++------+-------+-------+-----------+--------+--------+
+|    8 |  4096 |  7168 | 102.9     | 4.6    | 286.4  |
+|    8 |  7168 | 18432 | 70.8      | 29.8   | 1868.8 |
+|    8 | 18432 |  7168 | 69.3      | 30.5   | 1911.8 |
+|   64 |  4096 |  7168 | 69.1      | 54.4   | 439.0  |
+|   64 |  7168 | 18432 | 69.4      | 243.6  | 1933.6 |
+|   64 | 18432 |  7168 | 70.4      | 240.3  | 1917.2 |
+|   64 | 24576 |  1536 | 70.1      | 68.9   | 584.6  |
+|   64 | 32768 |   512 | 68.4      | 31.4   | 307.1  |
+|   64 |  7168 | 16384 | 69.5      | 216.3  | 1718.5 |
+|  128 |  4096 |  7168 | 141.1     | 53.3   | 222.1  |
+|  128 |  7168 | 18432 | 71.9      | 470.5  | 1896.1 |
+|  128 | 18432 |  7168 | 69.3      | 488.2  | 1988.2 |
+| 1024 |  4096 |  7168 | 89.7      | 670.1  | 502.5  |
+| 1024 | 18432 |  7168 | 279.0     | 969.8  | 635.2  |
+| 2048 |  4096 |  7168 | 175.1     | 687.0  | 347.4  |
+| 4096 |  4096 |  7168 | 335.4     | 717.0  | 275.1  |
++------+-------+-------+-----------+--------+--------+
+
+vLLM Triton Implementation:
++------+-------+-------+-----------+--------+--------+--------------+
+| m    | n     | k     | Time (μs) | TFLOPS | GB/s   | vs DeepGEMM  |
++------+-------+-------+-----------+--------+--------+--------------+
+|    8 |  4096 |  7168 | 74.0      | 6.3    | 398.2  | 1.39x faster |
+|    8 |  7168 | 18432 | 89.6      | 23.6   | 1478.1 | 0.79x slower |
+|    8 | 18432 |  7168 | 113.2     | 18.7   | 1170.4 | 0.61x slower |
+|   64 |  4096 |  7168 | 79.4      | 47.3   | 382.2  | 0.87x slower |
+|   64 |  7168 | 18432 | 98.5      | 171.7  | 1363.0 | 0.70x slower |
+|   64 | 18432 |  7168 | 119.5     | 141.5  | 1129.4 | 0.59x slower |
+|   64 | 24576 |  1536 | 37.6      | 128.4  | 1089.7 | 1.86x faster |
+|   64 | 32768 |   512 | 38.7      | 55.5   | 542.6  | 1.77x faster |
+|   64 |  7168 | 16384 | 86.1      | 174.5  | 1386.4 | 0.81x slower |
+|  128 |  4096 |  7168 | 90.7      | 82.9   | 345.4  | 1.56x faster |
+|  128 |  7168 | 18432 | 144.0     | 234.9  | 946.9  | 0.50x slower |
+|  128 | 18432 |  7168 | 229.5     | 147.4  | 600.1  | 0.30x slower |
+| 1024 |  4096 |  7168 | 242.3     | 248.2  | 186.1  | 0.37x slower |
+| 1024 | 18432 |  7168 | 897.8     | 301.4  | 197.4  | 0.31x slower |
+| 2048 |  4096 |  7168 | 463.0     | 259.7  | 131.4  | 0.38x slower |
+| 4096 |  4096 |  7168 | 901.8     | 266.7  | 102.3  | 0.37x slower |
++------+-------+-------+-----------+--------+--------+--------------+
+
+vLLM CUTLASS Implementation:
++------+-------+-------+-----------+--------+--------+--------------+--------------+
+| m    | n     | k     | Time (μs) | TFLOPS | GB/s   | vs DeepGEMM  | vs Triton    |
++------+-------+-------+-----------+--------+--------+--------------+--------------+
+|    8 |  4096 |  7168 | 34.6      | 13.6   | 852.3  | 2.98x faster | 2.14x faster |
+|    8 |  7168 | 18432 | 78.9      | 26.8   | 1677.3 | 0.90x slower | 1.13x faster |
+|    8 | 18432 |  7168 | 81.2      | 26.0   | 1631.1 | 0.85x slower | 1.39x faster |
+|   64 |  4096 |  7168 | 36.9      | 101.9  | 822.9  | 1.87x faster | 2.15x faster |
+|   64 |  7168 | 18432 | 87.4      | 193.4  | 1535.2 | 0.79x slower | 1.13x faster |
+|   64 | 18432 |  7168 | 85.0      | 199.0  | 1587.6 | 0.83x slower | 1.41x faster |
+|   64 | 24576 |  1536 | 28.0      | 172.8  | 1465.8 | 2.51x faster | 1.35x faster |
+|   64 | 32768 |   512 | 28.8      | 74.5   | 728.5  | 2.37x faster | 1.34x faster |
+|   64 |  7168 | 16384 | 77.9      | 193.0  | 1532.8 | 0.89x slower | 1.11x faster |
+|  128 |  4096 |  7168 | 39.1      | 192.4  | 802.0  | 3.61x faster | 2.32x faster |
+|  128 |  7168 | 18432 | 93.7      | 360.8  | 1454.2 | 0.77x slower | 1.54x faster |
+|  128 | 18432 |  7168 | 85.7      | 394.8  | 1608.0 | 0.81x slower | 2.68x faster |
+| 1024 |  4096 |  7168 | 99.7      | 603.1  | 452.2  | 0.90x slower | 2.43x faster |
+| 1024 | 18432 |  7168 | 331.3     | 816.7  | 534.9  | 0.84x slower | 2.71x faster |
+| 2048 |  4096 |  7168 | 198.3     | 606.6  | 306.7  | 0.88x slower | 2.34x faster |
+| 4096 |  4096 |  7168 | 392.2     | 613.2  | 235.3  | 0.86x slower | 2.30x faster |
++------+-------+-------+-----------+--------+--------+--------------+--------------+
+
+===== AVERAGE PERFORMANCE =====
++----------------+------------+----------+---------------+
+| Implementation | Avg TFLOPS | Avg GB/s | Avg Time (ms) |
++----------------+------------+----------+---------------+
+| DeepGEMM       | 310.98     | 1052.10  | 0.11          |
+| vLLM Triton    | 144.30     | 715.60   | 0.23          |
+| vLLM CUTLASS   | 286.78     | 1076.67  | 0.11          |
++----------------+------------+----------+---------------+
+
+===== AVERAGE SPEEDUPS =====
++-----------------------------+--------------+
+| Comparison                  | Speedup      |
++-----------------------------+--------------+
+| DeepGEMM vs vLLM Triton     | 1.71x faster |
+| DeepGEMM vs vLLM CUTLASS    | 0.94x slower |
+| vLLM CUTLASS vs vLLM Triton | 1.84x faster |
++-----------------------------+--------------+
+
+===== ACCURACY COMPARISON =====
++----------------+-----------------------+
+| Implementation | Avg Diff vs Reference |
++----------------+-----------------------+
+| DeepGEMM       | 0.000684              |
+| vLLM Triton    | 0.000684              |
+| vLLM CUTLASS   | 0.000684              |
++----------------+-----------------------+
+```