Merge tag 'v0.8.2' into v0.8.2-ori

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

@@ -361,7 +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/release-pipeline.yaml

@@ -82,7 +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:$(buildkite-agent meta-data get release-version) --progress plain -f Dockerfile.cpu ."
+      - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg GIT_REPO_CHECK=1 --tag public.ecr.aws/q9t5s3a7/vllm-cpu-release-repo:$(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-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 -e HF_TOKEN -v /root/.cache/huggingface:/root/.cache/huggingface --name gh200-test --gpus=all --entrypoint="" gh200-test bash -c '
+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-openvino-test.sh

-#!/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/generate.py --model facebook/opt-125m

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

@@ -19,17 +19,19 @@ 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 \
+    && export VLLM_USE_V1=1 \
+    && export VLLM_XLA_CHECK_RECOMPILATION=1 \
     && echo TEST_1 \
-    && VLLM_USE_V1=1 python3 /workspace/vllm/tests/tpu/test_compilation.py \
+    && python3 /workspace/vllm/tests/tpu/test_compilation.py \
     && echo TEST_2 \
-    && VLLM_USE_V1=1 pytest -v -s /workspace/vllm/tests/v1/tpu/test_basic.py \
+    && pytest -v -s /workspace/vllm/tests/v1/tpu/test_basic.py \
     && echo TEST_3 \
-    && VLLM_USE_V1=1 pytest -v -s /workspace/vllm/tests/entrypoints/llm/test_accuracy.py::test_lm_eval_accuracy_v1_engine \
+    && pytest -v -s /workspace/vllm/tests/entrypoints/llm/test_accuracy.py::test_lm_eval_accuracy_v1_engine \
     && echo TEST_4 \
-    && VLLM_USE_V1=1 pytest -s -v /workspace/vllm/tests/tpu/test_quantization_accuracy.py \
+    && pytest -s -v /workspace/vllm/tests/tpu/test_quantization_accuracy.py \
     && echo TEST_5 \
-    && VLLM_USE_V1=1 python3 /workspace/vllm/examples/offline_inference/tpu.py" \
-    
+    && 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/test-pipeline.yaml

@@ -118,7 +118,7 @@ steps:
   - pytest -v -s entrypoints/llm/test_generate.py # it needs a clean process
   - pytest -v -s entrypoints/llm/test_generate_multiple_loras.py # it needs a clean process
   - VLLM_USE_V1=0 pytest -v -s entrypoints/llm/test_guided_generate.py # it needs a clean process
-  - pytest -v -s entrypoints/openai --ignore=entrypoints/openai/test_oot_registration.py --ignore=entrypoints/openai/correctness/
+  - 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
   - VLLM_USE_V1=0 pytest -v -s entrypoints/offline_mode # Needs to avoid interference with other tests
 
@@ -136,6 +136,10 @@ steps:
   - 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
@@ -144,8 +148,8 @@ steps:
   # TODO: create a dedicated test section for multi-GPU example tests
   # when we have multiple distributed example tests
   - pushd ../examples/offline_inference
-  - python3 rlhf.py
-  - RAY_DEDUP_LOGS=0 python3 rlhf_colocate.py
+  - 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
@@ -295,6 +299,7 @@ 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
   source_file_dependencies:
@@ -510,9 +515,7 @@ steps:
   - vllm/worker/model_runner.py
   - entrypoints/llm/test_collective_rpc.py
   commands:
-  - pytest -v -s entrypoints/llm/test_collective_rpc.py
-  - VLLM_USE_V1=1 torchrun --nproc-per-node=2 distributed/test_torchrun_example.py
-  - torchrun --nproc-per-node=2 distributed/test_torchrun_example.py
+  - 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'

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

-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

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

Dockerfile

@@ -52,7 +52,8 @@ WORKDIR /workspace
 # after this step
 RUN --mount=type=cache,target=/root/.cache/uv \
     if [ "$TARGETPLATFORM" = "linux/arm64" ]; then \
-        uv 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 --index-url https://download.pytorch.org/whl/nightly/cu128 "torch==2.8.0.dev20250318+cu128" "torchvision==0.22.0.dev20250319";  \
+        uv pip install --index-url https://download.pytorch.org/whl/nightly/cu128 --pre pytorch_triton==3.3.0+gitab727c40; \
     fi
 
 COPY requirements/common.txt requirements/common.txt
@@ -200,7 +201,8 @@ RUN ldconfig /usr/local/cuda-$(echo $CUDA_VERSION | cut -d. -f1,2)/compat/
 # after this step
 RUN --mount=type=cache,target=/root/.cache/uv \
     if [ "$TARGETPLATFORM" = "linux/arm64" ]; then \
-        uv 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 --index-url https://download.pytorch.org/whl/nightly/cu128 "torch==2.8.0.dev20250318+cu128" "torchvision==0.22.0.dev20250319";  \
+        uv pip install --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.

Dockerfile.openvino

-# 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.rocm_base

@@ -12,6 +12,8 @@ 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

README.md

@@ -10,9 +10,17 @@ 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://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* 🔥
 
 - [2025/03] 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).
@@ -21,6 +29,10 @@ Easy, fast, and cheap LLM serving for everyone
 - [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!
@@ -34,8 +46,9 @@ 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.
@@ -143,10 +156,11 @@ If you use vLLM for your research, please cite our [paper](https://arxiv.org/abs
 
 ## Contact Us
 
-- For technical questions and feature requests, please use GitHub issues or discussions.
-- For discussing with fellow users and coordinating contributions and development, please use Slack.
-- For security disclosures, please use GitHub's security advisory feature.
-- For collaborations and partnerships, please contact us at vllm-questions AT lists.berkeley.edu.
+- 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
 

benchmarks/README.md

@@ -42,7 +42,7 @@ become available.
     </tr>
     <tr>
       <td><strong>HuggingFace</strong></td>
-      <td style="text-align: center;">✅</td>
+      <td style="text-align: center;">🟡</td>
       <td style="text-align: center;">🟡</td>
       <td>Specify your dataset path on HuggingFace</td>
     </tr>
@@ -60,8 +60,8 @@ become available.
 🚧: to be supported
 
 🟡: Partial support. Currently, HuggingFaceDataset only supports dataset formats
-similar to `lmms-lab/LLaVA-OneVision-Data`. If you need support for other dataset
-formats, please consider contributing.
+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`
 
@@ -139,6 +139,57 @@ python3 vllm/benchmarks/benchmark_serving.py \
   --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
 

benchmarks/backend_request_func.py

@@ -63,7 +63,7 @@ async def async_request_tgi(
             "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,
@@ -71,6 +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()

benchmarks/benchmark_dataset.py

@@ -17,6 +17,7 @@ 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
@@ -35,6 +36,8 @@ 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
 # -----------------------------------------------------------------------------
@@ -61,9 +64,6 @@ class SampleRequest:
 class BenchmarkDataset(ABC):
     DEFAULT_SEED = 0
 
-    # num_requests has default 1000 in both the benchmark_serving.py and
-    # benchmark_throughput.py
-
     def __init__(
         self,
         dataset_path: Optional[str] = None,
@@ -90,8 +90,8 @@ class BenchmarkDataset(ABC):
             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.
+        This method is used for chat models that expect a specific conversation
+        format.
         """
         content = [{"text": prompt, "type": "text"}]
         if mm_content is not None:
@@ -101,10 +101,10 @@ class BenchmarkDataset(ABC):
     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.
         """
@@ -121,18 +121,18 @@ class BenchmarkDataset(ABC):
         """
         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
@@ -160,21 +160,39 @@ class BenchmarkDataset(ABC):
                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
@@ -276,15 +294,16 @@ class RandomDataset(BenchmarkDataset):
     ) -> 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]:
-
+    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(
@@ -346,20 +365,24 @@ class ShareGPTDataset(BenchmarkDataset):
         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:
+    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"]
+            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)
@@ -383,6 +406,7 @@ class ShareGPTDataset(BenchmarkDataset):
                     expected_output_len=new_output_len,
                     lora_request=lora_request,
                 ))
+        self.maybe_oversample_requests(samples, num_requests)
         return samples
 
 
@@ -415,19 +439,20 @@ class SonnetDataset(BenchmarkDataset):
         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:
+    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)
+                      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"
@@ -506,12 +531,14 @@ class BurstGPTDataset(BenchmarkDataset):
         # 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]:
+    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):
@@ -544,7 +571,6 @@ class HuggingFaceDataset(BenchmarkDataset):
     Dataset class for processing a HuggingFace dataset with conversation data
     and optional images.
     """
-    DEFAULT_NUM_REQUESTS = 1000
 
     def __init__(
         self,
@@ -618,6 +644,7 @@ class HuggingFaceDataset(BenchmarkDataset):
                     expected_output_len=output_len,
                     multi_modal_data=mm_content,
                 ))
+        self.maybe_oversample_requests(sampled_requests, num_requests)
         return sampled_requests
 
 
@@ -632,7 +659,6 @@ class VisionArenaDataset(HuggingFaceDataset):
     """
 
     DEFAULT_OUTPUT_LEN = 128
-    DEFAULT_NUM_REQUESTS = 1000
     VISION_ARENA_DATASET_PATH = "lmarena-ai/vision-arena-bench-v0.1"
 
     def __init__(
@@ -657,12 +683,14 @@ class VisionArenaDataset(HuggingFaceDataset):
         )
         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:
+    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 = []
@@ -685,4 +713,5 @@ class VisionArenaDataset(HuggingFaceDataset):
                     expected_output_len=output_len,
                     multi_modal_data=mm_content,
                 ))
+        self.maybe_oversample_requests(sampled_requests, num_requests)
         return sampled_requests

benchmarks/benchmark_serving_structured_output.py

@@ -999,11 +999,12 @@ if __name__ == "__main__":
                         type=float,
                         default=1.0,
                         help="Ratio of Structured Outputs requests")
-    parser.add_argument("--structured-output-backend",
-                        type=str,
-                        choices=["outlines", "lm-format-enforcer", "xgrammar"],
-                        default="xgrammar",
-                        help="Backend to use for structured outputs")
+    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/kernels/benchmark_w8a8_block_fp8.py

+# 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)

cmake/external_projects/vllm_flash_attn.cmake

@@ -38,7 +38,7 @@ else()
   FetchContent_Declare(
           vllm-flash-attn
           GIT_REPOSITORY https://github.com/vllm-project/flash-attention.git
-          GIT_TAG 9bfa9869829d8c593527eb34c5271d0090f7ccc9 
+          GIT_TAG dc9d410b3e2d6534a4c70724c2515f4def670a22
           GIT_PROGRESS TRUE
           # Don't share the vllm-flash-attn build between build types
           BINARY_DIR ${CMAKE_BINARY_DIR}/vllm-flash-attn

csrc/quantization/fused_kernels/layernorm_utils.cuh

@@ -24,7 +24,7 @@ __device__ void compute_rms(float* rms, scalar_t const* __restrict__ input,
   // sum of squares
   float ss = 0.0f;
 
-  for (int32_t i = threadIdx.x; i < hidden_size; i += blockDim.x) {
+  for (auto i = threadIdx.x; i < hidden_size; i += blockDim.x) {
     float x = static_cast<float>(input[token_offset + i]);
     if constexpr (has_residual) {
       x += static_cast<float>(residual[token_offset + i]);
@@ -58,7 +58,7 @@ __device__ void compute_dynamic_per_token_scales(
   constexpr scalar_out_t qmax{std::numeric_limits<scalar_out_t>::max()};
 
   float block_absmax_val_maybe = 0.0f;
-  for (int32_t i = threadIdx.x; i < hidden_size; i += blockDim.x) {
+  for (auto i = threadIdx.x; i < hidden_size; i += blockDim.x) {
     float x = static_cast<float>(input[token_offset + i]);
     if constexpr (has_residual) {
       x += static_cast<float>(residual[token_offset + i]);
@@ -103,7 +103,7 @@ __device__ void norm_and_quant(scalar_out_t* __restrict__ output,
   int64_t const token_offset = blockIdx.x * static_cast<int64_t>(hidden_size);
   ;
 
-  for (int32_t i = threadIdx.x; i < hidden_size; i += blockDim.x) {
+  for (auto i = threadIdx.x; i < hidden_size; i += blockDim.x) {
     float x = static_cast<float>(input[token_offset + i]);
     if constexpr (has_residual) {
       x += static_cast<float>(residual[token_offset + i]);
@@ -142,7 +142,7 @@ __device__ void compute_rms(float* rms, scalar_t const* __restrict__ input,
   int32_t const num_vec_elems = hidden_size >> 2;
 
 #pragma unroll 4
-  for (int32_t i = threadIdx.x; i < num_vec_elems; i += blockDim.x) {
+  for (auto i = threadIdx.x; i < num_vec_elems; i += blockDim.x) {
     vec4_t<scalar_t> in = vec_input[i];
 
     vec4_t<float> x;
@@ -206,7 +206,7 @@ __device__ void compute_dynamic_per_token_scales(
   float block_absmax_val_maybe = 0.0f;
 
 #pragma unroll 4
-  for (int32_t i = threadIdx.x; i < num_vec_elems; i += blockDim.x) {
+  for (auto i = threadIdx.x; i < num_vec_elems; i += blockDim.x) {
     vec4_t<scalar_t> in = vec_input[i];
     vec4_t<scalar_t> const w = vec_weight[i];
 
@@ -286,7 +286,7 @@ __device__ void norm_and_quant(scalar_out_t* __restrict__ output,
 // TODO(luka/varun) extract into type-agnostic vectorized quant function to
 //  replace scaled_fp8_conversion_vec
 #pragma unroll 4
-  for (int32_t i = threadIdx.x; i < num_vec_elems; i += blockDim.x) {
+  for (auto i = threadIdx.x; i < num_vec_elems; i += blockDim.x) {
     vec4_t<scalar_t> const in = vec_input[i];
     vec4_t<scalar_t> const w = vec_weight[i];
 

csrc/quantization/gguf/dequantize.cuh

@@ -101,10 +101,10 @@ static __global__ void dequantize_block(const void * __restrict__ vx, dst_t * __
 template<typename dst_t>
 static __global__ void dequantize_block_q2_K(const void * __restrict__ vx, dst_t * __restrict__ yy) {
 
-    const int i   = blockIdx.x;
+    const auto i   = blockIdx.x;
     const block_q2_K * x = (const block_q2_K *) vx;
 
-    const int tid = threadIdx.x;
+    const auto tid = threadIdx.x;
     const int n   = tid/32;
     const int l   = tid - 32*n;
     const int is  = 8*n + l/16;
@@ -123,10 +123,10 @@ static __global__ void dequantize_block_q2_K(const void * __restrict__ vx, dst_t
 template<typename dst_t>
 static __global__ void dequantize_block_q3_K(const void * __restrict__ vx, dst_t * __restrict__ yy) {
 
-    const int i = blockIdx.x;
+    const auto i = blockIdx.x;
     const block_q3_K * x = (const block_q3_K *) vx;
 
-    const int r = threadIdx.x/4;
+    const auto r = threadIdx.x/4;
     const int tid = r/2;
     const int is0 = r%2;
     const int l0 = 16*is0 + 4*(threadIdx.x%4);
@@ -164,10 +164,10 @@ template<typename dst_t>
 static __global__ void dequantize_block_q4_K(const void * __restrict__ vx, dst_t * __restrict__ yy) {
     const block_q4_K * x = (const block_q4_K *) vx;
 
-    const int i = blockIdx.x;
+    const auto i = blockIdx.x;
 
     // assume 32 threads
-    const int tid = threadIdx.x;
+    const auto tid = threadIdx.x;
     const int il  = tid/8;
     const int ir  = tid%8;
     const int is  = 2*il;
@@ -197,10 +197,10 @@ template<typename dst_t>
 static __global__ void dequantize_block_q5_K(const void * __restrict__ vx, dst_t * __restrict__ yy) {
     const block_q5_K * x = (const block_q5_K *) vx;
 
-    const int i = blockIdx.x;
+    const auto i = blockIdx.x;
 
     // assume 64 threads - this is very slightly better than the one below
-    const int tid = threadIdx.x;
+    const auto tid = threadIdx.x;
     const int il  = tid/16;   // il is in 0...3
     const int ir  = tid%16;   // ir is in 0...15
     const int is  = 2*il;     // is is in 0...6
@@ -231,10 +231,10 @@ template<typename dst_t>
 static __global__ void dequantize_block_q6_K(const void * __restrict__ vx, dst_t * __restrict__ yy) {
     const block_q6_K * x = (const block_q6_K *) vx;
 
-    const int i = blockIdx.x;
+    const auto i = blockIdx.x;
 
     // assume 64 threads - this is very slightly better than the one below
-    const int tid = threadIdx.x;
+    const auto tid = threadIdx.x;
     const int ip  = tid/32;   // ip is 0 or 1
     const int il  = tid - 32*ip; // 0...32
     const int is  = 8*ip + il/16;
@@ -256,10 +256,10 @@ static __global__ void dequantize_block_q6_K(const void * __restrict__ vx, dst_t
 template<typename dst_t>
 static __global__ void dequantize_block_iq2_xxs(const void * __restrict__ vx, dst_t * __restrict__ yy) {
 
-    const int i   = blockIdx.x;
+    const auto i   = blockIdx.x;
     const block_iq2_xxs * x = (const block_iq2_xxs  *) vx;
 
-    const int tid = threadIdx.x;
+    const auto tid = threadIdx.x;
     const int il = tid/8; // 0...3
     const int ib = tid%8; // 0...7
     dst_t * y = yy + i*QK_K + 32*ib + 8*il;
@@ -275,10 +275,10 @@ static __global__ void dequantize_block_iq2_xxs(const void * __restrict__ vx, ds
 template<typename dst_t>
 static __global__ void dequantize_block_iq2_xs(const void * __restrict__ vx, dst_t * __restrict__ yy) {
 
-    const int i   = blockIdx.x;
+    const auto i   = blockIdx.x;
     const block_iq2_xs * x = (const block_iq2_xs *) vx;
 
-    const int tid = threadIdx.x;
+    const auto tid = threadIdx.x;
     const int il = tid/8; // 0...3
     const int ib = tid%8; // 0...7
     dst_t * y = yy + i*QK_K + 32*ib + 8*il;
@@ -293,10 +293,10 @@ static __global__ void dequantize_block_iq2_xs(const void * __restrict__ vx, dst
 template<typename dst_t>
 static __global__ void dequantize_block_iq2_s(const void * __restrict__ vx, dst_t * __restrict__ yy) {
 
-    const int i   = blockIdx.x;
+    const auto i   = blockIdx.x;
     const block_iq2_s * x = (const block_iq2_s *) vx;
 
-    const int tid = threadIdx.x;
+    const auto tid = threadIdx.x;
     const int il = tid/8; // 0...3
     const int ib = tid%8; // 0...7
     dst_t * y = yy + i*QK_K + 32*ib + 8*il;
@@ -309,10 +309,10 @@ static __global__ void dequantize_block_iq2_s(const void * __restrict__ vx, dst_
 template<typename dst_t>
 static __global__ void dequantize_block_iq3_xxs(const void * __restrict__ vx, dst_t * __restrict__ yy) {
 
-    const int i   = blockIdx.x;
+    const auto i   = blockIdx.x;
     const block_iq3_xxs * x = (const block_iq3_xxs  *) vx;
 
-    const int tid = threadIdx.x;
+    const auto tid = threadIdx.x;
     const int il = tid/8; // 0...3
     const int ib = tid%8; // 0...7
     dst_t * y = yy + i*QK_K + 32*ib + 8*il;
@@ -332,10 +332,10 @@ static __global__ void dequantize_block_iq3_xxs(const void * __restrict__ vx, ds
 template<typename dst_t>
 static __global__ void dequantize_block_iq3_s(const void * __restrict__ vx, dst_t * __restrict__ yy) {
 
-    const int i   = blockIdx.x;
+    const auto i   = blockIdx.x;
     const block_iq3_s * x = (const block_iq3_s *) vx;
 
-    const int tid = threadIdx.x;
+    const auto tid = threadIdx.x;
     const int il = tid/8; // 0...3
     const int ib = tid%8; // 0...7
     dst_t * y = yy + i*QK_K + 32*ib + 8*il;
@@ -399,10 +399,10 @@ static __global__ void dequantize_block_iq1_m(const void * __restrict__ vx, dst_
 template<typename dst_t>
 static __global__ void dequantize_block_iq4_nl(const void * __restrict__ vx, dst_t * __restrict__ yy) {
 
-    const int i   = blockIdx.x;
+    const auto i   = blockIdx.x;
     const block_iq4_nl * x = (const block_iq4_nl *) vx + i*(QK_K/QK4_NL);
 
-    const int tid = threadIdx.x;
+    const auto tid = threadIdx.x;
     const int il = tid/8; // 0...3
     const int ib = tid%8; // 0...7
     dst_t * y = yy + i*QK_K + 32*ib + 4*il;
@@ -417,10 +417,10 @@ static __global__ void dequantize_block_iq4_nl(const void * __restrict__ vx, dst
 
 template<typename dst_t>
 static __global__ void dequantize_block_iq4_xs(const void * __restrict__ vx, dst_t * __restrict__ yy) {
-    const int i   = blockIdx.x;
+    const auto i   = blockIdx.x;
     const block_iq4_xs * x = (const block_iq4_xs *)vx;
 
-    const int tid = threadIdx.x;
+    const auto tid = threadIdx.x;
     const int il = tid/8; // 0...3
     const int ib = tid%8; // 0...7
     dst_t * y = yy + i*QK_K + 32*ib + 4*il;
@@ -565,4 +565,4 @@ static to_fp16_cuda_t ggml_get_to_fp16_cuda(int64_t type) {
         default:
             return nullptr;
     }
-}
\ No newline at end of file
+}